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 1.52 2005/11/07 20:05:51 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(firstaddr, loadaddr)
308 vm_paddr_t firstaddr;
313 struct mdglobaldata *gd;
317 avail_start = firstaddr;
320 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too
321 * large. It should instead be correctly calculated in locore.s and
322 * not based on 'first' (which is a physical address, not a virtual
323 * address, for the start of unused physical memory). The kernel
324 * page tables are NOT double mapped and thus should not be included
325 * in this calculation.
327 virtual_avail = (vm_offset_t) KERNBASE + firstaddr;
328 virtual_avail = pmap_kmem_choose(virtual_avail);
330 virtual_end = VM_MAX_KERNEL_ADDRESS;
333 * Initialize protection array.
335 i386_protection_init();
338 * The kernel's pmap is statically allocated so we don't have to use
339 * pmap_create, which is unlikely to work correctly at this part of
340 * the boot sequence (XXX and which no longer exists).
342 kernel_pmap = &kernel_pmap_store;
344 kernel_pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
345 kernel_pmap->pm_count = 1;
346 kernel_pmap->pm_active = (cpumask_t)-1; /* don't allow deactivation */
347 TAILQ_INIT(&kernel_pmap->pm_pvlist);
351 * Reserve some special page table entries/VA space for temporary
354 #define SYSMAP(c, p, v, n) \
355 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
358 pte = (pt_entry_t *) pmap_pte(kernel_pmap, va);
361 * CMAP1/CMAP2 are used for zeroing and copying pages.
363 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
368 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
371 * ptvmmap is used for reading arbitrary physical pages via
374 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
377 * msgbufp is used to map the system message buffer.
378 * XXX msgbufmap is not used.
380 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
381 atop(round_page(MSGBUF_SIZE)))
386 for (i = 0; i < NKPT; i++)
390 * PG_G is terribly broken on SMP because we IPI invltlb's in some
391 * cases rather then invl1pg. Actually, I don't even know why it
392 * works under UP because self-referential page table mappings
397 if (cpu_feature & CPUID_PGE)
402 * Initialize the 4MB page size flag
406 * The 4MB page version of the initial
407 * kernel page mapping.
411 #if !defined(DISABLE_PSE)
412 if (cpu_feature & CPUID_PSE) {
415 * Note that we have enabled PSE mode
418 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
419 ptditmp &= ~(NBPDR - 1);
420 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
425 * Enable the PSE mode. If we are SMP we can't do this
426 * now because the APs will not be able to use it when
429 load_cr4(rcr4() | CR4_PSE);
432 * We can do the mapping here for the single processor
433 * case. We simply ignore the old page table page from
437 * For SMP, we still need 4K pages to bootstrap APs,
438 * PSE will be enabled as soon as all APs are up.
440 PTD[KPTDI] = (pd_entry_t)ptditmp;
441 kernel_pmap->pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
447 if (cpu_apic_address == 0)
448 panic("pmap_bootstrap: no local apic!");
450 /* local apic is mapped on last page */
451 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag |
452 (cpu_apic_address & PG_FRAME));
456 * We need to finish setting up the globaldata page for the BSP.
457 * locore has already populated the page table for the mdglobaldata
460 pg = MDGLOBALDATA_BASEALLOC_PAGES;
461 gd = &CPU_prvspace[0].mdglobaldata;
462 gd->gd_CMAP1 = &SMPpt[pg + 0];
463 gd->gd_CMAP2 = &SMPpt[pg + 1];
464 gd->gd_CMAP3 = &SMPpt[pg + 2];
465 gd->gd_PMAP1 = &SMPpt[pg + 3];
466 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
467 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
468 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
469 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
476 * Set 4mb pdir for mp startup
481 if (pseflag && (cpu_feature & CPUID_PSE)) {
482 load_cr4(rcr4() | CR4_PSE);
483 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
484 kernel_pmap->pm_pdir[KPTDI] =
485 PTD[KPTDI] = (pd_entry_t)pdir4mb;
493 * Initialize the pmap module.
494 * Called by vm_init, to initialize any structures that the pmap
495 * system needs to map virtual memory.
496 * pmap_init has been enhanced to support in a fairly consistant
497 * way, discontiguous physical memory.
506 * object for kernel page table pages
508 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
511 * Allocate memory for random pmap data structures. Includes the
515 for(i = 0; i < vm_page_array_size; i++) {
518 m = &vm_page_array[i];
519 TAILQ_INIT(&m->md.pv_list);
520 m->md.pv_list_count = 0;
524 * init the pv free list
526 initial_pvs = vm_page_array_size;
527 if (initial_pvs < MINPV)
529 pvzone = &pvzone_store;
530 pvinit = (struct pv_entry *) kmem_alloc(kernel_map,
531 initial_pvs * sizeof (struct pv_entry));
532 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
536 * Now it is safe to enable pv_table recording.
538 pmap_initialized = TRUE;
542 * Initialize the address space (zone) for the pv_entries. Set a
543 * high water mark so that the system can recover from excessive
544 * numbers of pv entries.
549 int shpgperproc = PMAP_SHPGPERPROC;
551 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
552 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
553 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
554 pv_entry_high_water = 9 * (pv_entry_max / 10);
555 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
559 /***************************************************
560 * Low level helper routines.....
561 ***************************************************/
563 #if defined(PMAP_DIAGNOSTIC)
566 * This code checks for non-writeable/modified pages.
567 * This should be an invalid condition.
570 pmap_nw_modified(pt_entry_t ptea)
576 if ((pte & (PG_M|PG_RW)) == PG_M)
585 * this routine defines the region(s) of memory that should
586 * not be tested for the modified bit.
588 static PMAP_INLINE int
589 pmap_track_modified(vm_offset_t va)
591 if ((va < clean_sva) || (va >= clean_eva))
598 get_ptbase(pmap_t pmap)
600 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
601 struct globaldata *gd = mycpu;
603 /* are we current address space or kernel? */
604 if (pmap == kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
605 return (unsigned *) PTmap;
608 /* otherwise, we are alternate address space */
609 KKASSERT(gd->gd_intr_nesting_level == 0 && (gd->gd_curthread->td_flags & TDF_INTTHREAD) == 0);
611 if (frame != (((unsigned) APTDpde) & PG_FRAME)) {
612 APTDpde = (pd_entry_t)(frame | PG_RW | PG_V);
613 /* The page directory is not shared between CPUs */
616 return (unsigned *) APTmap;
622 * Extract the physical page address associated with the map/VA pair.
624 * This function may not be called from an interrupt if the pmap is
628 pmap_extract(pmap_t pmap, vm_offset_t va)
631 vm_offset_t pdirindex;
633 pdirindex = va >> PDRSHIFT;
634 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
636 if ((rtval & PG_PS) != 0) {
637 rtval &= ~(NBPDR - 1);
638 rtval |= va & (NBPDR - 1);
641 pte = get_ptbase(pmap) + i386_btop(va);
642 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
649 * Extract user accessible page only, return NULL if the page is not
650 * present or if it's current state is not sufficient. Caller will
651 * generally call vm_fault() on failure and try again.
654 pmap_extract_vmpage(pmap_t pmap, vm_offset_t va, int prot)
657 vm_offset_t pdirindex;
659 pdirindex = va >> PDRSHIFT;
660 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
664 if ((rtval & PG_PS) != 0) {
665 if ((rtval & (PG_V|PG_U)) != (PG_V|PG_U))
667 if ((prot & VM_PROT_WRITE) && (rtval & PG_RW) == 0)
669 rtval &= ~(NBPDR - 1);
670 rtval |= va & (NBPDR - 1);
671 m = PHYS_TO_VM_PAGE(rtval);
673 pte = get_ptbase(pmap) + i386_btop(va);
674 if ((*pte & (PG_V|PG_U)) != (PG_V|PG_U))
676 if ((prot & VM_PROT_WRITE) && (*pte & PG_RW) == 0)
678 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
679 m = PHYS_TO_VM_PAGE(rtval);
686 /***************************************************
687 * Low level mapping routines.....
688 ***************************************************/
691 * Routine: pmap_kenter
693 * Add a wired page to the KVA
694 * NOTE! note that in order for the mapping to take effect -- you
695 * should do an invltlb after doing the pmap_kenter().
698 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
702 pmap_inval_info info;
704 pmap_inval_init(&info);
705 pmap_inval_add(&info, kernel_pmap, va);
706 npte = pa | PG_RW | PG_V | pgeflag;
707 pte = (unsigned *)vtopte(va);
709 pmap_inval_flush(&info);
713 * Routine: pmap_kenter_quick
715 * Similar to pmap_kenter(), except we only invalidate the
716 * mapping on the current CPU.
719 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
724 npte = pa | PG_RW | PG_V | pgeflag;
725 pte = (unsigned *)vtopte(va);
727 cpu_invlpg((void *)va);
731 pmap_kenter_sync(vm_offset_t va)
733 pmap_inval_info info;
735 pmap_inval_init(&info);
736 pmap_inval_add(&info, kernel_pmap, va);
737 pmap_inval_flush(&info);
741 pmap_kenter_sync_quick(vm_offset_t va)
743 cpu_invlpg((void *)va);
747 * remove a page from the kernel pagetables
750 pmap_kremove(vm_offset_t va)
753 pmap_inval_info info;
755 pmap_inval_init(&info);
756 pmap_inval_add(&info, kernel_pmap, va);
757 pte = (unsigned *)vtopte(va);
759 pmap_inval_flush(&info);
763 pmap_kremove_quick(vm_offset_t va)
766 pte = (unsigned *)vtopte(va);
768 cpu_invlpg((void *)va);
772 * Used to map a range of physical addresses into kernel
773 * virtual address space.
775 * For now, VM is already on, we only need to map the
779 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
781 while (start < end) {
782 pmap_kenter(virt, start);
791 * Add a list of wired pages to the kva
792 * this routine is only used for temporary
793 * kernel mappings that do not need to have
794 * page modification or references recorded.
795 * Note that old mappings are simply written
796 * over. The page *must* be wired.
799 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
803 end_va = va + count * PAGE_SIZE;
805 while (va < end_va) {
808 pte = (unsigned *)vtopte(va);
809 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
810 cpu_invlpg((void *)va);
815 smp_invltlb(); /* XXX */
820 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
823 cpumask_t cmask = mycpu->gd_cpumask;
825 end_va = va + count * PAGE_SIZE;
827 while (va < end_va) {
832 * Install the new PTE. If the pte changed from the prior
833 * mapping we must reset the cpu mask and invalidate the page.
834 * If the pte is the same but we have not seen it on the
835 * current cpu, invlpg the existing mapping. Otherwise the
836 * entry is optimal and no invalidation is required.
838 pte = (unsigned *)vtopte(va);
839 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
840 if (*pte != pteval) {
843 cpu_invlpg((void *)va);
844 } else if ((*mask & cmask) == 0) {
845 cpu_invlpg((void *)va);
854 * this routine jerks page mappings from the
855 * kernel -- it is meant only for temporary mappings.
858 pmap_qremove(vm_offset_t va, int count)
862 end_va = va + count*PAGE_SIZE;
864 while (va < end_va) {
867 pte = (unsigned *)vtopte(va);
869 cpu_invlpg((void *)va);
878 * This routine works like vm_page_lookup() but also blocks as long as the
879 * page is busy. This routine does not busy the page it returns.
881 * Unless the caller is managing objects whos pages are in a known state,
882 * the call should be made with a critical section held so the page's object
883 * association remains valid on return.
886 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
891 m = vm_page_lookup(object, pindex);
892 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
898 * Create a new thread and optionally associate it with a (new) process.
899 * NOTE! the new thread's cpu may not equal the current cpu.
902 pmap_init_thread(thread_t td)
904 /* enforce pcb placement */
905 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
906 td->td_savefpu = &td->td_pcb->pcb_save;
907 td->td_sp = (char *)td->td_pcb - 16;
911 * Create the UPAGES for a new process.
912 * This routine directly affects the fork perf for a process.
915 pmap_init_proc(struct proc *p, struct thread *td)
917 p->p_addr = (void *)td->td_kstack;
920 td->td_lwp = &p->p_lwp;
921 td->td_switch = cpu_heavy_switch;
925 bzero(p->p_addr, sizeof(*p->p_addr));
929 * Dispose the UPAGES for a process that has exited.
930 * This routine directly impacts the exit perf of a process.
933 pmap_dispose_proc(struct proc *p)
937 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
939 if ((td = p->p_thread) != NULL) {
948 * Allow the UPAGES for a process to be prejudicially paged out.
951 pmap_swapout_proc(struct proc *p)
959 upobj = p->p_upages_obj;
962 * Unwiring the pages allow them to be paged to their backing store
966 for (i = 0; i < UPAGES; i++) {
967 if ((m = vm_page_lookup(upobj, i)) == NULL)
968 panic("pmap_swapout_proc: upage already missing???");
970 vm_page_unwire(m, 0);
971 pmap_kremove((vm_offset_t)p->p_addr + (PAGE_SIZE * i));
978 * Bring the UPAGES for a specified process back in.
981 pmap_swapin_proc(struct proc *p)
989 upobj = p->p_upages_obj;
990 for (i = 0; i < UPAGES; i++) {
991 m = vm_page_grab(upobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
993 pmap_kenter((vm_offset_t)p->p_addr + (i * PAGE_SIZE),
996 if (m->valid != VM_PAGE_BITS_ALL) {
997 rv = vm_pager_get_pages(upobj, &m, 1, 0);
998 if (rv != VM_PAGER_OK)
999 panic("pmap_swapin_proc: cannot get upages for proc: %d\n", p->p_pid);
1000 m = vm_page_lookup(upobj, i);
1001 m->valid = VM_PAGE_BITS_ALL;
1005 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1011 /***************************************************
1012 * Page table page management routines.....
1013 ***************************************************/
1016 * This routine unholds page table pages, and if the hold count
1017 * drops to zero, then it decrements the wire count.
1020 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1022 pmap_inval_flush(info);
1023 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1026 if (m->hold_count == 0) {
1028 * unmap the page table page
1030 pmap_inval_add(info, pmap, -1);
1031 pmap->pm_pdir[m->pindex] = 0;
1032 --pmap->pm_stats.resident_count;
1034 if (pmap->pm_ptphint == m)
1035 pmap->pm_ptphint = NULL;
1038 * If the page is finally unwired, simply free it.
1041 if (m->wire_count == 0) {
1044 vm_page_free_zero(m);
1045 --vmstats.v_wire_count;
1052 static PMAP_INLINE int
1053 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1056 if (m->hold_count == 0)
1057 return _pmap_unwire_pte_hold(pmap, m, info);
1063 * After removing a page table entry, this routine is used to
1064 * conditionally free the page, and manage the hold/wire counts.
1067 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1068 pmap_inval_info_t info)
1071 if (va >= UPT_MIN_ADDRESS)
1075 ptepindex = (va >> PDRSHIFT);
1076 if (pmap->pm_ptphint &&
1077 (pmap->pm_ptphint->pindex == ptepindex)) {
1078 mpte = pmap->pm_ptphint;
1080 pmap_inval_flush(info);
1081 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1082 pmap->pm_ptphint = mpte;
1086 return pmap_unwire_pte_hold(pmap, mpte, info);
1090 pmap_pinit0(struct pmap *pmap)
1093 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE);
1094 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1096 pmap->pm_active = 0;
1097 pmap->pm_ptphint = NULL;
1098 TAILQ_INIT(&pmap->pm_pvlist);
1099 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1103 * Initialize a preallocated and zeroed pmap structure,
1104 * such as one in a vmspace structure.
1107 pmap_pinit(struct pmap *pmap)
1112 * No need to allocate page table space yet but we do need a valid
1113 * page directory table.
1115 if (pmap->pm_pdir == NULL) {
1117 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE);
1121 * allocate object for the ptes
1123 if (pmap->pm_pteobj == NULL)
1124 pmap->pm_pteobj = vm_object_allocate( OBJT_DEFAULT, PTDPTDI + 1);
1127 * allocate the page directory page
1129 ptdpg = vm_page_grab( pmap->pm_pteobj, PTDPTDI,
1130 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1132 ptdpg->wire_count = 1;
1133 ++vmstats.v_wire_count;
1136 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY); /* not usually mapped*/
1137 ptdpg->valid = VM_PAGE_BITS_ALL;
1139 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1140 if ((ptdpg->flags & PG_ZERO) == 0)
1141 bzero(pmap->pm_pdir, PAGE_SIZE);
1143 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1145 /* install self-referential address mapping entry */
1146 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1147 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1150 pmap->pm_active = 0;
1151 pmap->pm_ptphint = NULL;
1152 TAILQ_INIT(&pmap->pm_pvlist);
1153 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1157 * Wire in kernel global address entries. To avoid a race condition
1158 * between pmap initialization and pmap_growkernel, this procedure
1159 * should be called after the vmspace is attached to the process
1160 * but before this pmap is activated.
1163 pmap_pinit2(struct pmap *pmap)
1165 /* XXX copies current process, does not fill in MPPTDI */
1166 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1170 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1172 unsigned *pde = (unsigned *) pmap->pm_pdir;
1174 * This code optimizes the case of freeing non-busy
1175 * page-table pages. Those pages are zero now, and
1176 * might as well be placed directly into the zero queue.
1178 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1184 * Remove the page table page from the processes address space.
1187 pmap->pm_stats.resident_count--;
1189 if (p->hold_count) {
1190 panic("pmap_release: freeing held page table page");
1193 * Page directory pages need to have the kernel
1194 * stuff cleared, so they can go into the zero queue also.
1196 if (p->pindex == PTDPTDI) {
1197 bzero(pde + KPTDI, nkpt * PTESIZE);
1200 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1203 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1204 pmap->pm_ptphint = NULL;
1207 vmstats.v_wire_count--;
1208 vm_page_free_zero(p);
1213 * this routine is called if the page table page is not
1217 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1219 vm_offset_t pteva, ptepa;
1223 * Find or fabricate a new pagetable page
1225 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1226 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1228 KASSERT(m->queue == PQ_NONE,
1229 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1231 if (m->wire_count == 0)
1232 vmstats.v_wire_count++;
1236 * Increment the hold count for the page table page
1237 * (denoting a new mapping.)
1242 * Map the pagetable page into the process address space, if
1243 * it isn't already there.
1246 pmap->pm_stats.resident_count++;
1248 ptepa = VM_PAGE_TO_PHYS(m);
1249 pmap->pm_pdir[ptepindex] =
1250 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1253 * Set the page table hint
1255 pmap->pm_ptphint = m;
1258 * Try to use the new mapping, but if we cannot, then
1259 * do it with the routine that maps the page explicitly.
1261 if ((m->flags & PG_ZERO) == 0) {
1262 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1263 (((unsigned) PTDpde) & PG_FRAME)) {
1264 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1265 bzero((caddr_t) pteva, PAGE_SIZE);
1267 pmap_zero_page(ptepa);
1271 m->valid = VM_PAGE_BITS_ALL;
1272 vm_page_flag_clear(m, PG_ZERO);
1273 vm_page_flag_set(m, PG_MAPPED);
1280 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1287 * Calculate pagetable page index
1289 ptepindex = va >> PDRSHIFT;
1292 * Get the page directory entry
1294 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1297 * This supports switching from a 4MB page to a
1300 if (ptepa & PG_PS) {
1301 pmap->pm_pdir[ptepindex] = 0;
1308 * If the page table page is mapped, we just increment the
1309 * hold count, and activate it.
1313 * In order to get the page table page, try the
1316 if (pmap->pm_ptphint &&
1317 (pmap->pm_ptphint->pindex == ptepindex)) {
1318 m = pmap->pm_ptphint;
1320 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1321 pmap->pm_ptphint = m;
1327 * Here if the pte page isn't mapped, or if it has been deallocated.
1329 return _pmap_allocpte(pmap, ptepindex);
1333 /***************************************************
1334 * Pmap allocation/deallocation routines.
1335 ***************************************************/
1338 * Release any resources held by the given physical map.
1339 * Called when a pmap initialized by pmap_pinit is being released.
1340 * Should only be called if the map contains no valid mappings.
1343 pmap_release(struct pmap *pmap)
1345 vm_page_t p,n,ptdpg;
1346 vm_object_t object = pmap->pm_pteobj;
1349 #if defined(DIAGNOSTIC)
1350 if (object->ref_count != 1)
1351 panic("pmap_release: pteobj reference count != 1");
1357 curgeneration = object->generation;
1358 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = n) {
1359 n = TAILQ_NEXT(p, listq);
1360 if (p->pindex == PTDPTDI) {
1365 if (!pmap_release_free_page(pmap, p) &&
1366 (object->generation != curgeneration)) {
1374 if (ptdpg && !pmap_release_free_page(pmap, ptdpg))
1379 kvm_size(SYSCTL_HANDLER_ARGS)
1381 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE;
1383 return sysctl_handle_long(oidp, &ksize, 0, req);
1385 SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD,
1386 0, 0, kvm_size, "IU", "Size of KVM");
1389 kvm_free(SYSCTL_HANDLER_ARGS)
1391 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end;
1393 return sysctl_handle_long(oidp, &kfree, 0, req);
1395 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD,
1396 0, 0, kvm_free, "IU", "Amount of KVM free");
1399 * grow the number of kernel page table entries, if needed
1402 pmap_growkernel(vm_offset_t addr)
1406 vm_offset_t ptppaddr;
1411 if (kernel_vm_end == 0) {
1412 kernel_vm_end = KERNBASE;
1414 while (pdir_pde(PTD, kernel_vm_end)) {
1415 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1419 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1420 while (kernel_vm_end < addr) {
1421 if (pdir_pde(PTD, kernel_vm_end)) {
1422 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1427 * This index is bogus, but out of the way
1429 nkpg = vm_page_alloc(kptobj, nkpt,
1430 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1432 panic("pmap_growkernel: no memory to grow kernel");
1437 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1438 pmap_zero_page(ptppaddr);
1439 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1440 pdir_pde(PTD, kernel_vm_end) = newpdir;
1442 FOREACH_PROC_IN_SYSTEM(p) {
1444 pmap = vmspace_pmap(p->p_vmspace);
1445 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1448 *pmap_pde(kernel_pmap, kernel_vm_end) = newpdir;
1449 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1455 * Retire the given physical map from service.
1456 * Should only be called if the map contains
1457 * no valid mappings.
1460 pmap_destroy(pmap_t pmap)
1467 count = --pmap->pm_count;
1470 panic("destroying a pmap is not yet implemented");
1475 * Add a reference to the specified pmap.
1478 pmap_reference(pmap_t pmap)
1485 /***************************************************
1486 * page management routines.
1487 ***************************************************/
1490 * free the pv_entry back to the free list. This function may be
1491 * called from an interrupt.
1493 static PMAP_INLINE void
1494 free_pv_entry(pv_entry_t pv)
1501 * get a new pv_entry, allocating a block from the system
1502 * when needed. This function may be called from an interrupt.
1508 if (pv_entry_high_water &&
1509 (pv_entry_count > pv_entry_high_water) &&
1510 (pmap_pagedaemon_waken == 0)) {
1511 pmap_pagedaemon_waken = 1;
1512 wakeup (&vm_pages_needed);
1514 return zalloc(pvzone);
1518 * This routine is very drastic, but can save the system
1526 static int warningdone=0;
1528 if (pmap_pagedaemon_waken == 0)
1531 if (warningdone < 5) {
1532 printf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1536 for(i = 0; i < vm_page_array_size; i++) {
1537 m = &vm_page_array[i];
1538 if (m->wire_count || m->hold_count || m->busy ||
1539 (m->flags & PG_BUSY))
1543 pmap_pagedaemon_waken = 0;
1548 * If it is the first entry on the list, it is actually
1549 * in the header and we must copy the following entry up
1550 * to the header. Otherwise we must search the list for
1551 * the entry. In either case we free the now unused entry.
1554 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1555 vm_offset_t va, pmap_inval_info_t info)
1561 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1562 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1563 if (pmap == pv->pv_pmap && va == pv->pv_va)
1567 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1568 if (va == pv->pv_va)
1575 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1576 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1577 m->md.pv_list_count--;
1578 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1579 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1580 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1588 * Create a pv entry for page at pa for
1592 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1597 pv = get_pv_entry();
1602 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1603 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1604 m->md.pv_list_count++;
1610 * pmap_remove_pte: do the things to unmap a page in a process
1613 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1614 pmap_inval_info_t info)
1619 pmap_inval_add(info, pmap, va);
1620 oldpte = loadandclear(ptq);
1622 pmap->pm_stats.wired_count -= 1;
1624 * Machines that don't support invlpg, also don't support
1625 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1629 cpu_invlpg((void *)va);
1630 pmap->pm_stats.resident_count -= 1;
1631 if (oldpte & PG_MANAGED) {
1632 m = PHYS_TO_VM_PAGE(oldpte);
1633 if (oldpte & PG_M) {
1634 #if defined(PMAP_DIAGNOSTIC)
1635 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1637 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1641 if (pmap_track_modified(va))
1645 vm_page_flag_set(m, PG_REFERENCED);
1646 return pmap_remove_entry(pmap, m, va, info);
1648 return pmap_unuse_pt(pmap, va, NULL, info);
1657 * Remove a single page from a process address space.
1659 * This function may not be called from an interrupt if the pmap is
1663 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1668 * if there is no pte for this address, just skip it!!! Otherwise
1669 * get a local va for mappings for this pmap and remove the entry.
1671 if (*pmap_pde(pmap, va) != 0) {
1672 ptq = get_ptbase(pmap) + i386_btop(va);
1674 pmap_remove_pte(pmap, ptq, va, info);
1682 * Remove the given range of addresses from the specified map.
1684 * It is assumed that the start and end are properly
1685 * rounded to the page size.
1687 * This function may not be called from an interrupt if the pmap is
1691 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1695 vm_offset_t ptpaddr;
1696 vm_offset_t sindex, eindex;
1697 struct pmap_inval_info info;
1702 if (pmap->pm_stats.resident_count == 0)
1705 pmap_inval_init(&info);
1708 * special handling of removing one page. a very
1709 * common operation and easy to short circuit some
1712 if (((sva + PAGE_SIZE) == eva) &&
1713 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1714 pmap_remove_page(pmap, sva, &info);
1715 pmap_inval_flush(&info);
1720 * Get a local virtual address for the mappings that are being
1723 ptbase = get_ptbase(pmap);
1725 sindex = i386_btop(sva);
1726 eindex = i386_btop(eva);
1728 for (; sindex < eindex; sindex = pdnxt) {
1732 * Calculate index for next page table.
1734 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1735 if (pmap->pm_stats.resident_count == 0)
1738 pdirindex = sindex / NPDEPG;
1739 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1740 pmap_inval_add(&info, pmap, -1);
1741 pmap->pm_pdir[pdirindex] = 0;
1742 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1747 * Weed out invalid mappings. Note: we assume that the page
1748 * directory table is always allocated, and in kernel virtual.
1754 * Limit our scan to either the end of the va represented
1755 * by the current page table page, or to the end of the
1756 * range being removed.
1758 if (pdnxt > eindex) {
1762 for (; sindex != pdnxt; sindex++) {
1764 if (ptbase[sindex] == 0)
1766 va = i386_ptob(sindex);
1767 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1771 pmap_inval_flush(&info);
1777 * Removes this physical page from all physical maps in which it resides.
1778 * Reflects back modify bits to the pager.
1780 * This routine may not be called from an interrupt.
1784 pmap_remove_all(vm_page_t m)
1786 struct pmap_inval_info info;
1787 unsigned *pte, tpte;
1790 #if defined(PMAP_DIAGNOSTIC)
1792 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1795 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1796 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1800 pmap_inval_init(&info);
1802 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1803 pv->pv_pmap->pm_stats.resident_count--;
1805 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1806 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1808 tpte = loadandclear(pte);
1810 pv->pv_pmap->pm_stats.wired_count--;
1813 vm_page_flag_set(m, PG_REFERENCED);
1816 * Update the vm_page_t clean and reference bits.
1819 #if defined(PMAP_DIAGNOSTIC)
1820 if (pmap_nw_modified((pt_entry_t) tpte)) {
1822 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1826 if (pmap_track_modified(pv->pv_va))
1829 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1830 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1831 m->md.pv_list_count--;
1832 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1836 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1838 pmap_inval_flush(&info);
1844 * Set the physical protection on the specified range of this map
1847 * This function may not be called from an interrupt if the map is
1848 * not the kernel_pmap.
1851 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1854 vm_offset_t pdnxt, ptpaddr;
1855 vm_pindex_t sindex, eindex;
1856 pmap_inval_info info;
1861 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1862 pmap_remove(pmap, sva, eva);
1866 if (prot & VM_PROT_WRITE)
1869 pmap_inval_init(&info);
1871 ptbase = get_ptbase(pmap);
1873 sindex = i386_btop(sva);
1874 eindex = i386_btop(eva);
1876 for (; sindex < eindex; sindex = pdnxt) {
1880 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1882 pdirindex = sindex / NPDEPG;
1883 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1884 pmap_inval_add(&info, pmap, -1);
1885 (unsigned) pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1886 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1891 * Weed out invalid mappings. Note: we assume that the page
1892 * directory table is always allocated, and in kernel virtual.
1897 if (pdnxt > eindex) {
1901 for (; sindex != pdnxt; sindex++) {
1906 /* XXX this isn't optimal */
1907 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1908 pbits = ptbase[sindex];
1910 if (pbits & PG_MANAGED) {
1913 m = PHYS_TO_VM_PAGE(pbits);
1914 vm_page_flag_set(m, PG_REFERENCED);
1918 if (pmap_track_modified(i386_ptob(sindex))) {
1920 m = PHYS_TO_VM_PAGE(pbits);
1929 if (pbits != ptbase[sindex]) {
1930 ptbase[sindex] = pbits;
1934 pmap_inval_flush(&info);
1938 * Insert the given physical page (p) at
1939 * the specified virtual address (v) in the
1940 * target physical map with the protection requested.
1942 * If specified, the page will be wired down, meaning
1943 * that the related pte can not be reclaimed.
1945 * NB: This is the only routine which MAY NOT lazy-evaluate
1946 * or lose information. That is, this routine must actually
1947 * insert this page into the given map NOW.
1950 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1956 vm_offset_t origpte, newpte;
1958 pmap_inval_info info;
1964 #ifdef PMAP_DIAGNOSTIC
1965 if (va > VM_MAX_KERNEL_ADDRESS)
1966 panic("pmap_enter: toobig");
1967 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1968 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1973 * In the case that a page table page is not
1974 * resident, we are creating it here.
1976 if (va < UPT_MIN_ADDRESS) {
1977 mpte = pmap_allocpte(pmap, va);
1980 pmap_inval_init(&info);
1981 pte = pmap_pte(pmap, va);
1984 * Page Directory table entry not valid, we need a new PT page
1987 panic("pmap_enter: invalid page directory pdir=%x, va=0x%x\n",
1988 (unsigned) pmap->pm_pdir[PTDPTDI], va);
1991 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
1992 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1993 origpte = *(vm_offset_t *)pte;
1994 opa = origpte & PG_FRAME;
1996 if (origpte & PG_PS)
1997 panic("pmap_enter: attempted pmap_enter on 4MB page");
2000 * Mapping has not changed, must be protection or wiring change.
2002 if (origpte && (opa == pa)) {
2004 * Wiring change, just update stats. We don't worry about
2005 * wiring PT pages as they remain resident as long as there
2006 * are valid mappings in them. Hence, if a user page is wired,
2007 * the PT page will be also.
2009 if (wired && ((origpte & PG_W) == 0))
2010 pmap->pm_stats.wired_count++;
2011 else if (!wired && (origpte & PG_W))
2012 pmap->pm_stats.wired_count--;
2014 #if defined(PMAP_DIAGNOSTIC)
2015 if (pmap_nw_modified((pt_entry_t) origpte)) {
2017 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
2023 * Remove the extra pte reference. Note that we cannot
2024 * optimize the RO->RW case because we have adjusted the
2025 * wiring count above and may need to adjust the wiring
2032 * We might be turning off write access to the page,
2033 * so we go ahead and sense modify status.
2035 if (origpte & PG_MANAGED) {
2036 if ((origpte & PG_M) && pmap_track_modified(va)) {
2038 om = PHYS_TO_VM_PAGE(opa);
2046 * Mapping has changed, invalidate old range and fall through to
2047 * handle validating new mapping.
2051 err = pmap_remove_pte(pmap, pte, va, &info);
2053 panic("pmap_enter: pte vanished, va: 0x%x", va);
2057 * Enter on the PV list if part of our managed memory. Note that we
2058 * raise IPL while manipulating pv_table since pmap_enter can be
2059 * called at interrupt time.
2061 if (pmap_initialized &&
2062 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2063 pmap_insert_entry(pmap, va, mpte, m);
2068 * Increment counters
2070 pmap->pm_stats.resident_count++;
2072 pmap->pm_stats.wired_count++;
2076 * Now validate mapping with desired protection/wiring.
2078 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2082 if (va < UPT_MIN_ADDRESS)
2084 if (pmap == kernel_pmap)
2088 * if the mapping or permission bits are different, we need
2089 * to update the pte.
2091 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2092 *pte = newpte | PG_A;
2094 pmap_inval_flush(&info);
2098 * this code makes some *MAJOR* assumptions:
2099 * 1. Current pmap & pmap exists.
2102 * 4. No page table pages.
2103 * 5. Tlbflush is deferred to calling procedure.
2104 * 6. Page IS managed.
2105 * but is *MUCH* faster than pmap_enter...
2109 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
2113 pmap_inval_info info;
2115 pmap_inval_init(&info);
2118 * In the case that a page table page is not
2119 * resident, we are creating it here.
2121 if (va < UPT_MIN_ADDRESS) {
2126 * Calculate pagetable page index
2128 ptepindex = va >> PDRSHIFT;
2129 if (mpte && (mpte->pindex == ptepindex)) {
2134 * Get the page directory entry
2136 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2139 * If the page table page is mapped, we just increment
2140 * the hold count, and activate it.
2144 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2145 if (pmap->pm_ptphint &&
2146 (pmap->pm_ptphint->pindex == ptepindex)) {
2147 mpte = pmap->pm_ptphint;
2149 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2150 pmap->pm_ptphint = mpte;
2156 mpte = _pmap_allocpte(pmap, ptepindex);
2164 * This call to vtopte makes the assumption that we are
2165 * entering the page into the current pmap. In order to support
2166 * quick entry into any pmap, one would likely use pmap_pte_quick.
2167 * But that isn't as quick as vtopte.
2169 pte = (unsigned *)vtopte(va);
2172 pmap_unwire_pte_hold(pmap, mpte, &info);
2177 * Enter on the PV list if part of our managed memory. Note that we
2178 * raise IPL while manipulating pv_table since pmap_enter can be
2179 * called at interrupt time.
2181 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
2182 pmap_insert_entry(pmap, va, mpte, m);
2185 * Increment counters
2187 pmap->pm_stats.resident_count++;
2189 pa = VM_PAGE_TO_PHYS(m);
2192 * Now validate mapping with RO protection
2194 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2195 *pte = pa | PG_V | PG_U;
2197 *pte = pa | PG_V | PG_U | PG_MANAGED;
2203 * Make a temporary mapping for a physical address. This is only intended
2204 * to be used for panic dumps.
2207 pmap_kenter_temporary(vm_paddr_t pa, int i)
2209 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2210 return ((void *)crashdumpmap);
2213 #define MAX_INIT_PT (96)
2216 * This routine preloads the ptes for a given object into the specified pmap.
2217 * This eliminates the blast of soft faults on process startup and
2218 * immediately after an mmap.
2221 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2222 vm_object_t object, vm_pindex_t pindex,
2223 vm_size_t size, int limit)
2230 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2235 * XXX you must be joking, entering PTE's into a user page table
2236 * without any accounting? This could result in the page table
2237 * being freed while it still contains mappings (free with PG_ZERO
2238 * assumption leading to a non-zero page being marked PG_ZERO).
2241 * This code maps large physical mmap regions into the
2242 * processor address space. Note that some shortcuts
2243 * are taken, but the code works.
2246 (object->type == OBJT_DEVICE) &&
2247 ((addr & (NBPDR - 1)) == 0) &&
2248 ((size & (NBPDR - 1)) == 0) ) {
2251 unsigned int ptepindex;
2255 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)])
2259 p = vm_page_lookup(object, pindex);
2260 if (p && vm_page_sleep_busy(p, FALSE, "init4p"))
2264 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
2269 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) {
2274 p = vm_page_lookup(object, pindex);
2278 ptepa = (vm_offset_t) VM_PAGE_TO_PHYS(p);
2279 if (ptepa & (NBPDR - 1)) {
2283 p->valid = VM_PAGE_BITS_ALL;
2285 pmap->pm_stats.resident_count += size >> PAGE_SHIFT;
2286 npdes = size >> PDRSHIFT;
2287 for (i = 0; i < npdes; i++) {
2288 pmap->pm_pdir[ptepindex] =
2289 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_PS);
2293 vm_page_flag_set(p, PG_MAPPED);
2300 psize = i386_btop(size);
2302 if ((object->type != OBJT_VNODE) ||
2303 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2304 (object->resident_page_count > MAX_INIT_PT))) {
2308 if (psize + pindex > object->size) {
2309 if (object->size < pindex)
2311 psize = object->size - pindex;
2316 * If we are processing a major portion of the object, then scan the
2319 * We cannot safely scan the object's memq unless we are in a
2320 * critical section since interrupts can remove pages from objects.
2324 if (psize > (object->resident_page_count >> 2)) {
2327 for (p = TAILQ_FIRST(&object->memq);
2328 objpgs > 0 && p != NULL;
2329 p = TAILQ_NEXT(p, listq)
2332 if (tmpidx < pindex)
2335 if (tmpidx >= psize)
2339 * don't allow an madvise to blow away our really
2340 * free pages allocating pv entries.
2342 if ((limit & MAP_PREFAULT_MADVISE) &&
2343 vmstats.v_free_count < vmstats.v_free_reserved) {
2346 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2348 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2349 if ((p->queue - p->pc) == PQ_CACHE)
2350 vm_page_deactivate(p);
2352 mpte = pmap_enter_quick(pmap,
2353 addr + i386_ptob(tmpidx), p, mpte);
2354 vm_page_flag_set(p, PG_MAPPED);
2361 * else lookup the pages one-by-one.
2363 for (tmpidx = 0; tmpidx < psize; tmpidx += 1) {
2365 * don't allow an madvise to blow away our really
2366 * free pages allocating pv entries.
2368 if ((limit & MAP_PREFAULT_MADVISE) &&
2369 vmstats.v_free_count < vmstats.v_free_reserved) {
2372 p = vm_page_lookup(object, tmpidx + pindex);
2374 ((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2376 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2377 if ((p->queue - p->pc) == PQ_CACHE)
2378 vm_page_deactivate(p);
2380 mpte = pmap_enter_quick(pmap,
2381 addr + i386_ptob(tmpidx), p, mpte);
2382 vm_page_flag_set(p, PG_MAPPED);
2391 * pmap_prefault provides a quick way of clustering pagefaults into a
2392 * processes address space. It is a "cousin" of pmap_object_init_pt,
2393 * except it runs at page fault time instead of mmap time.
2397 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2399 static int pmap_prefault_pageorder[] = {
2400 -PAGE_SIZE, PAGE_SIZE,
2401 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2402 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2403 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2407 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2416 if (!curproc || (pmap != vmspace_pmap(curproc->p_vmspace)))
2419 object = entry->object.vm_object;
2421 starta = addra - PFBAK * PAGE_SIZE;
2422 if (starta < entry->start)
2423 starta = entry->start;
2424 else if (starta > addra)
2428 * critical section protection is required to maintain the
2429 * page/object association, interrupts can free pages and remove
2430 * them from their objects.
2434 for (i = 0; i < PAGEORDER_SIZE; i++) {
2435 vm_object_t lobject;
2438 addr = addra + pmap_prefault_pageorder[i];
2439 if (addr > addra + (PFFOR * PAGE_SIZE))
2442 if (addr < starta || addr >= entry->end)
2445 if ((*pmap_pde(pmap, addr)) == NULL)
2448 pte = (unsigned *) vtopte(addr);
2452 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2455 for (m = vm_page_lookup(lobject, pindex);
2456 (!m && (lobject->type == OBJT_DEFAULT) &&
2457 (lobject->backing_object));
2458 lobject = lobject->backing_object
2460 if (lobject->backing_object_offset & PAGE_MASK)
2462 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2463 m = vm_page_lookup(lobject->backing_object, pindex);
2467 * give-up when a page is not in memory
2472 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2474 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2476 if ((m->queue - m->pc) == PQ_CACHE) {
2477 vm_page_deactivate(m);
2480 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2481 vm_page_flag_set(m, PG_MAPPED);
2489 * Routine: pmap_change_wiring
2490 * Function: Change the wiring attribute for a map/virtual-address
2492 * In/out conditions:
2493 * The mapping must already exist in the pmap.
2496 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2503 pte = pmap_pte(pmap, va);
2505 if (wired && !pmap_pte_w(pte))
2506 pmap->pm_stats.wired_count++;
2507 else if (!wired && pmap_pte_w(pte))
2508 pmap->pm_stats.wired_count--;
2511 * Wiring is not a hardware characteristic so there is no need to
2512 * invalidate TLB. However, in an SMP environment we must use
2513 * a locked bus cycle to update the pte (if we are not using
2514 * the pmap_inval_*() API that is)... it's ok to do this for simple
2519 atomic_set_int(pte, PG_W);
2521 atomic_clear_int(pte, PG_W);
2524 atomic_set_int_nonlocked(pte, PG_W);
2526 atomic_clear_int_nonlocked(pte, PG_W);
2533 * Copy the range specified by src_addr/len
2534 * from the source map to the range dst_addr/len
2535 * in the destination map.
2537 * This routine is only advisory and need not do anything.
2540 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2541 vm_size_t len, vm_offset_t src_addr)
2543 pmap_inval_info info;
2545 vm_offset_t end_addr = src_addr + len;
2547 unsigned src_frame, dst_frame;
2550 if (dst_addr != src_addr)
2553 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2554 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2558 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2559 if (dst_frame != (((unsigned) APTDpde) & PG_FRAME)) {
2560 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2561 /* The page directory is not shared between CPUs */
2564 pmap_inval_init(&info);
2565 pmap_inval_add(&info, dst_pmap, -1);
2566 pmap_inval_add(&info, src_pmap, -1);
2569 * critical section protection is required to maintain the page/object
2570 * association, interrupts can free pages and remove them from
2574 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2575 unsigned *src_pte, *dst_pte;
2576 vm_page_t dstmpte, srcmpte;
2577 vm_offset_t srcptepaddr;
2580 if (addr >= UPT_MIN_ADDRESS)
2581 panic("pmap_copy: invalid to pmap_copy page tables\n");
2584 * Don't let optional prefaulting of pages make us go
2585 * way below the low water mark of free pages or way
2586 * above high water mark of used pv entries.
2588 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2589 pv_entry_count > pv_entry_high_water)
2592 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2593 ptepindex = addr >> PDRSHIFT;
2595 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2596 if (srcptepaddr == 0)
2599 if (srcptepaddr & PG_PS) {
2600 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2601 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2602 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2607 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2608 if ((srcmpte == NULL) ||
2609 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2612 if (pdnxt > end_addr)
2615 src_pte = (unsigned *) vtopte(addr);
2616 dst_pte = (unsigned *) avtopte(addr);
2617 while (addr < pdnxt) {
2621 * we only virtual copy managed pages
2623 if ((ptetemp & PG_MANAGED) != 0) {
2625 * We have to check after allocpte for the
2626 * pte still being around... allocpte can
2629 dstmpte = pmap_allocpte(dst_pmap, addr);
2630 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2632 * Clear the modified and
2633 * accessed (referenced) bits
2636 m = PHYS_TO_VM_PAGE(ptetemp);
2637 *dst_pte = ptetemp & ~(PG_M | PG_A);
2638 dst_pmap->pm_stats.resident_count++;
2639 pmap_insert_entry(dst_pmap, addr,
2642 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2644 if (dstmpte->hold_count >= srcmpte->hold_count)
2653 pmap_inval_flush(&info);
2657 * Routine: pmap_kernel
2659 * Returns the physical map handle for the kernel.
2664 return (kernel_pmap);
2670 * Zero the specified PA by mapping the page into KVM and clearing its
2673 * This function may be called from an interrupt and no locking is
2677 pmap_zero_page(vm_paddr_t phys)
2679 struct mdglobaldata *gd = mdcpu;
2682 if (*(int *)gd->gd_CMAP3)
2683 panic("pmap_zero_page: CMAP3 busy");
2684 *(int *)gd->gd_CMAP3 =
2685 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2686 cpu_invlpg(gd->gd_CADDR3);
2688 #if defined(I686_CPU)
2689 if (cpu_class == CPUCLASS_686)
2690 i686_pagezero(gd->gd_CADDR3);
2693 bzero(gd->gd_CADDR3, PAGE_SIZE);
2694 *(int *) gd->gd_CMAP3 = 0;
2699 * pmap_page_assertzero:
2701 * Assert that a page is empty, panic if it isn't.
2704 pmap_page_assertzero(vm_paddr_t phys)
2706 struct mdglobaldata *gd = mdcpu;
2710 if (*(int *)gd->gd_CMAP3)
2711 panic("pmap_zero_page: CMAP3 busy");
2712 *(int *)gd->gd_CMAP3 =
2713 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2714 cpu_invlpg(gd->gd_CADDR3);
2715 for (i = 0; i < PAGE_SIZE; i += 4) {
2716 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2717 panic("pmap_page_assertzero() @ %p not zero!\n",
2718 (void *)gd->gd_CADDR3);
2721 *(int *) gd->gd_CMAP3 = 0;
2728 * Zero part of a physical page by mapping it into memory and clearing
2729 * its contents with bzero.
2731 * off and size may not cover an area beyond a single hardware page.
2734 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2736 struct mdglobaldata *gd = mdcpu;
2739 if (*(int *) gd->gd_CMAP3)
2740 panic("pmap_zero_page: CMAP3 busy");
2741 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2742 cpu_invlpg(gd->gd_CADDR3);
2744 #if defined(I686_CPU)
2745 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2746 i686_pagezero(gd->gd_CADDR3);
2749 bzero((char *)gd->gd_CADDR3 + off, size);
2750 *(int *) gd->gd_CMAP3 = 0;
2757 * Copy the physical page from the source PA to the target PA.
2758 * This function may be called from an interrupt. No locking
2762 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2764 struct mdglobaldata *gd = mdcpu;
2767 if (*(int *) gd->gd_CMAP1)
2768 panic("pmap_copy_page: CMAP1 busy");
2769 if (*(int *) gd->gd_CMAP2)
2770 panic("pmap_copy_page: CMAP2 busy");
2772 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2773 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2775 cpu_invlpg(gd->gd_CADDR1);
2776 cpu_invlpg(gd->gd_CADDR2);
2778 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2780 *(int *) gd->gd_CMAP1 = 0;
2781 *(int *) gd->gd_CMAP2 = 0;
2786 * pmap_copy_page_frag:
2788 * Copy the physical page from the source PA to the target PA.
2789 * This function may be called from an interrupt. No locking
2793 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2795 struct mdglobaldata *gd = mdcpu;
2798 if (*(int *) gd->gd_CMAP1)
2799 panic("pmap_copy_page: CMAP1 busy");
2800 if (*(int *) gd->gd_CMAP2)
2801 panic("pmap_copy_page: CMAP2 busy");
2803 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2804 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2806 cpu_invlpg(gd->gd_CADDR1);
2807 cpu_invlpg(gd->gd_CADDR2);
2809 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2810 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2813 *(int *) gd->gd_CMAP1 = 0;
2814 *(int *) gd->gd_CMAP2 = 0;
2819 * Returns true if the pmap's pv is one of the first
2820 * 16 pvs linked to from this page. This count may
2821 * be changed upwards or downwards in the future; it
2822 * is only necessary that true be returned for a small
2823 * subset of pmaps for proper page aging.
2826 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2831 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2836 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2837 if (pv->pv_pmap == pmap) {
2849 #define PMAP_REMOVE_PAGES_CURPROC_ONLY
2851 * Remove all pages from specified address space
2852 * this aids process exit speeds. Also, this code
2853 * is special cased for current process only, but
2854 * can have the more generic (and slightly slower)
2855 * mode enabled. This is much faster than pmap_remove
2856 * in the case of running down an entire address space.
2859 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2861 unsigned *pte, tpte;
2864 pmap_inval_info info;
2866 #ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY
2867 if (!curproc || (pmap != vmspace_pmap(curproc->p_vmspace))) {
2868 printf("warning: pmap_remove_pages called with non-current pmap\n");
2873 pmap_inval_init(&info);
2875 for(pv = TAILQ_FIRST(&pmap->pm_pvlist);
2879 if (pv->pv_va >= eva || pv->pv_va < sva) {
2880 npv = TAILQ_NEXT(pv, pv_plist);
2884 #ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY
2885 pte = (unsigned *)vtopte(pv->pv_va);
2887 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2889 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2893 * We cannot remove wired pages from a process' mapping at this time
2896 npv = TAILQ_NEXT(pv, pv_plist);
2901 m = PHYS_TO_VM_PAGE(tpte);
2903 KASSERT(m < &vm_page_array[vm_page_array_size],
2904 ("pmap_remove_pages: bad tpte %x", tpte));
2906 pv->pv_pmap->pm_stats.resident_count--;
2909 * Update the vm_page_t clean and reference bits.
2916 npv = TAILQ_NEXT(pv, pv_plist);
2917 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2919 m->md.pv_list_count--;
2920 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2921 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2922 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2925 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2928 pmap_inval_flush(&info);
2933 * pmap_testbit tests bits in pte's
2934 * note that the testbit/changebit routines are inline,
2935 * and a lot of things compile-time evaluate.
2938 pmap_testbit(vm_page_t m, int bit)
2943 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2946 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2951 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2953 * if the bit being tested is the modified bit, then
2954 * mark clean_map and ptes as never
2957 if (bit & (PG_A|PG_M)) {
2958 if (!pmap_track_modified(pv->pv_va))
2962 #if defined(PMAP_DIAGNOSTIC)
2964 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2968 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2979 * this routine is used to modify bits in ptes
2981 static __inline void
2982 pmap_changebit(vm_page_t m, int bit, boolean_t setem)
2984 struct pmap_inval_info info;
2988 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2991 pmap_inval_init(&info);
2995 * Loop over all current mappings setting/clearing as appropos If
2996 * setting RO do we need to clear the VAC?
2998 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3000 * don't write protect pager mappings
3002 if (!setem && (bit == PG_RW)) {
3003 if (!pmap_track_modified(pv->pv_va))
3007 #if defined(PMAP_DIAGNOSTIC)
3009 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
3015 * Careful here. We can use a locked bus instruction to
3016 * clear PG_A or PG_M safely but we need to synchronize
3017 * with the target cpus when we mess with PG_RW.
3019 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3021 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
3025 atomic_set_int(pte, bit);
3027 atomic_set_int_nonlocked(pte, bit);
3030 vm_offset_t pbits = *(vm_offset_t *)pte;
3037 atomic_clear_int(pte, PG_M|PG_RW);
3039 atomic_clear_int_nonlocked(pte, PG_M|PG_RW);
3043 atomic_clear_int(pte, bit);
3045 atomic_clear_int_nonlocked(pte, bit);
3051 pmap_inval_flush(&info);
3056 * pmap_page_protect:
3058 * Lower the permission for all mappings to a given page.
3061 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3063 if ((prot & VM_PROT_WRITE) == 0) {
3064 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3065 pmap_changebit(m, PG_RW, FALSE);
3073 pmap_phys_address(int ppn)
3075 return (i386_ptob(ppn));
3079 * pmap_ts_referenced:
3081 * Return a count of reference bits for a page, clearing those bits.
3082 * It is not necessary for every reference bit to be cleared, but it
3083 * is necessary that 0 only be returned when there are truly no
3084 * reference bits set.
3086 * XXX: The exact number of bits to check and clear is a matter that
3087 * should be tested and standardized at some point in the future for
3088 * optimal aging of shared pages.
3091 pmap_ts_referenced(vm_page_t m)
3093 pv_entry_t pv, pvf, pvn;
3097 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3102 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3107 pvn = TAILQ_NEXT(pv, pv_list);
3109 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3111 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3113 if (!pmap_track_modified(pv->pv_va))
3116 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3118 if (pte && (*pte & PG_A)) {
3120 atomic_clear_int(pte, PG_A);
3122 atomic_clear_int_nonlocked(pte, PG_A);
3129 } while ((pv = pvn) != NULL && pv != pvf);
3139 * Return whether or not the specified physical page was modified
3140 * in any physical maps.
3143 pmap_is_modified(vm_page_t m)
3145 return pmap_testbit(m, PG_M);
3149 * Clear the modify bits on the specified physical page.
3152 pmap_clear_modify(vm_page_t m)
3154 pmap_changebit(m, PG_M, FALSE);
3158 * pmap_clear_reference:
3160 * Clear the reference bit on the specified physical page.
3163 pmap_clear_reference(vm_page_t m)
3165 pmap_changebit(m, PG_A, FALSE);
3169 * Miscellaneous support routines follow
3173 i386_protection_init(void)
3177 kp = protection_codes;
3178 for (prot = 0; prot < 8; prot++) {
3180 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3182 * Read access is also 0. There isn't any execute bit,
3183 * so just make it readable.
3185 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3186 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3187 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3190 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3191 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3192 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3193 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3201 * Map a set of physical memory pages into the kernel virtual
3202 * address space. Return a pointer to where it is mapped. This
3203 * routine is intended to be used for mapping device memory,
3206 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3210 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3212 vm_offset_t va, tmpva, offset;
3215 offset = pa & PAGE_MASK;
3216 size = roundup(offset + size, PAGE_SIZE);
3218 va = kmem_alloc_nofault(kernel_map, size);
3220 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3223 for (tmpva = va; size > 0;) {
3224 pte = (unsigned *)vtopte(tmpva);
3225 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3233 return ((void *)(va + offset));
3237 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3239 vm_offset_t base, offset;
3241 base = va & PG_FRAME;
3242 offset = va & PAGE_MASK;
3243 size = roundup(offset + size, PAGE_SIZE);
3244 pmap_qremove(va, size >> PAGE_SHIFT);
3245 kmem_free(kernel_map, base, size);
3249 * perform the pmap work for mincore
3252 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3254 unsigned *ptep, pte;
3258 ptep = pmap_pte(pmap, addr);
3263 if ((pte = *ptep) != 0) {
3266 val = MINCORE_INCORE;
3267 if ((pte & PG_MANAGED) == 0)
3270 pa = pte & PG_FRAME;
3272 m = PHYS_TO_VM_PAGE(pa);
3278 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3280 * Modified by someone
3282 else if (m->dirty || pmap_is_modified(m))
3283 val |= MINCORE_MODIFIED_OTHER;
3288 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3291 * Referenced by someone
3293 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3294 val |= MINCORE_REFERENCED_OTHER;
3295 vm_page_flag_set(m, PG_REFERENCED);
3302 pmap_activate(struct proc *p)
3306 pmap = vmspace_pmap(p->p_vmspace);
3308 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3310 pmap->pm_active |= 1;
3312 #if defined(SWTCH_OPTIM_STATS)
3315 p->p_thread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3316 load_cr3(p->p_thread->td_pcb->pcb_cr3);
3320 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3323 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3327 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3332 #if defined(PMAP_DEBUG)
3334 pmap_pid_dump(int pid)
3340 FOREACH_PROC_IN_SYSTEM(p) {
3341 if (p->p_pid != pid)
3347 pmap = vmspace_pmap(p->p_vmspace);
3348 for(i=0;i<1024;i++) {
3351 unsigned base = i << PDRSHIFT;
3353 pde = &pmap->pm_pdir[i];
3354 if (pde && pmap_pde_v(pde)) {
3355 for(j=0;j<1024;j++) {
3356 unsigned va = base + (j << PAGE_SHIFT);
3357 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) {
3364 pte = pmap_pte_quick( pmap, va);
3365 if (pte && pmap_pte_v(pte)) {
3369 m = PHYS_TO_VM_PAGE(pa);
3370 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x",
3371 va, pa, m->hold_count, m->wire_count, m->flags);
3392 static void pads (pmap_t pm);
3393 void pmap_pvdump (vm_paddr_t pa);
3395 /* print address space of pmap*/
3402 if (pm == kernel_pmap)
3404 for (i = 0; i < 1024; i++)
3406 for (j = 0; j < 1024; j++) {
3407 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3408 if (pm == kernel_pmap && va < KERNBASE)
3410 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS)
3412 ptep = pmap_pte_quick(pm, va);
3413 if (pmap_pte_v(ptep))
3414 printf("%x:%x ", va, *(int *) ptep);
3420 pmap_pvdump(vm_paddr_t pa)
3425 printf("pa %08llx", (long long)pa);
3426 m = PHYS_TO_VM_PAGE(pa);
3427 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3429 printf(" -> pmap %p, va %x, flags %x",
3430 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3432 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);