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 1.63 2006/12/23 00:27:03 swildner 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 <machine_base/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;
148 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
152 vm_paddr_t avail_start; /* PA of first available physical page */
153 vm_paddr_t avail_end; /* PA of last available physical page */
154 vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */
155 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
156 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
157 static int pgeflag; /* PG_G or-in */
158 static int pseflag; /* PG_PS or-in */
160 static vm_object_t kptobj;
163 vm_offset_t kernel_vm_end;
166 * Data for the pv entry allocation mechanism
168 static vm_zone_t pvzone;
169 static struct vm_zone pvzone_store;
170 static struct vm_object pvzone_obj;
171 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
172 static int pmap_pagedaemon_waken = 0;
173 static struct pv_entry *pvinit;
176 * All those kernel PT submaps that BSD is so fond of
178 pt_entry_t *CMAP1 = 0, *ptmmap;
179 caddr_t CADDR1 = 0, ptvmmap = 0;
180 static pt_entry_t *msgbufmap;
181 struct msgbuf *msgbufp=0;
186 static pt_entry_t *pt_crashdumpmap;
187 static caddr_t crashdumpmap;
189 extern pt_entry_t *SMPpt;
191 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
192 static unsigned * get_ptbase (pmap_t pmap);
193 static pv_entry_t get_pv_entry (void);
194 static void i386_protection_init (void);
195 static __inline void pmap_changebit (vm_page_t m, int bit, boolean_t setem);
197 static void pmap_remove_all (vm_page_t m);
198 static vm_page_t pmap_enter_quick (pmap_t pmap, vm_offset_t va,
199 vm_page_t m, vm_page_t mpte);
200 static int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
201 vm_offset_t sva, pmap_inval_info_t info);
202 static void pmap_remove_page (struct pmap *pmap,
203 vm_offset_t va, pmap_inval_info_t info);
204 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
205 vm_offset_t va, pmap_inval_info_t info);
206 static boolean_t pmap_testbit (vm_page_t m, int bit);
207 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
208 vm_page_t mpte, vm_page_t m);
210 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
212 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
213 static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
214 static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
215 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
216 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
217 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
219 static unsigned pdir4mb;
222 * Move the kernel virtual free pointer to the next
223 * 4MB. This is used to help improve performance
224 * by using a large (4MB) page for much of the kernel
225 * (.text, .data, .bss)
228 pmap_kmem_choose(vm_offset_t addr)
230 vm_offset_t newaddr = addr;
232 if (cpu_feature & CPUID_PSE) {
233 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
242 * Extract the page table entry associated with the given map/virtual
245 * This function may NOT be called from an interrupt.
247 PMAP_INLINE unsigned *
248 pmap_pte(pmap_t pmap, vm_offset_t va)
253 pdeaddr = (unsigned *) pmap_pde(pmap, va);
254 if (*pdeaddr & PG_PS)
257 return get_ptbase(pmap) + i386_btop(va);
266 * Super fast pmap_pte routine best used when scanning the pv lists.
267 * This eliminates many course-grained invltlb calls. Note that many of
268 * the pv list scans are across different pmaps and it is very wasteful
269 * to do an entire invltlb when checking a single mapping.
271 * Should only be called while in a critical section.
274 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
276 struct mdglobaldata *gd = mdcpu;
279 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
280 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
281 unsigned index = i386_btop(va);
282 /* are we current address space or kernel? */
283 if ((pmap == kernel_pmap) ||
284 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
285 return (unsigned *) PTmap + index;
287 newpf = pde & PG_FRAME;
288 if ( ((* (unsigned *) gd->gd_PMAP1) & PG_FRAME) != newpf) {
289 * (unsigned *) gd->gd_PMAP1 = newpf | PG_RW | PG_V;
290 cpu_invlpg(gd->gd_PADDR1);
292 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
299 * Bootstrap the system enough to run with virtual memory.
301 * On the i386 this is called after mapping has already been enabled
302 * and just syncs the pmap module with what has already been done.
303 * [We can't call it easily with mapping off since the kernel is not
304 * mapped with PA == VA, hence we would have to relocate every address
305 * from the linked base (virtual) address "KERNBASE" to the actual
306 * (physical) address starting relative to 0]
309 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
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;
923 KKASSERT(td->td_mpcount == 1);
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) {
947 /***************************************************
948 * Page table page management routines.....
949 ***************************************************/
952 * This routine unholds page table pages, and if the hold count
953 * drops to zero, then it decrements the wire count.
956 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
958 pmap_inval_flush(info);
959 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
962 if (m->hold_count == 0) {
964 * unmap the page table page
966 pmap_inval_add(info, pmap, -1);
967 pmap->pm_pdir[m->pindex] = 0;
968 --pmap->pm_stats.resident_count;
970 if (pmap->pm_ptphint == m)
971 pmap->pm_ptphint = NULL;
974 * If the page is finally unwired, simply free it.
977 if (m->wire_count == 0) {
980 vm_page_free_zero(m);
981 --vmstats.v_wire_count;
988 static PMAP_INLINE int
989 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
992 if (m->hold_count == 0)
993 return _pmap_unwire_pte_hold(pmap, m, info);
999 * After removing a page table entry, this routine is used to
1000 * conditionally free the page, and manage the hold/wire counts.
1003 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1004 pmap_inval_info_t info)
1007 if (va >= UPT_MIN_ADDRESS)
1011 ptepindex = (va >> PDRSHIFT);
1012 if (pmap->pm_ptphint &&
1013 (pmap->pm_ptphint->pindex == ptepindex)) {
1014 mpte = pmap->pm_ptphint;
1016 pmap_inval_flush(info);
1017 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1018 pmap->pm_ptphint = mpte;
1022 return pmap_unwire_pte_hold(pmap, mpte, info);
1026 * Initialize pmap0/vmspace0 - the kernel pmap. This pmap is not added
1027 * to pmap_list because it, and IdlePTD, represents the template used
1028 * to update all other pmaps.
1031 pmap_pinit0(struct pmap *pmap)
1034 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE);
1035 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1037 pmap->pm_active = 0;
1038 pmap->pm_ptphint = NULL;
1039 TAILQ_INIT(&pmap->pm_pvlist);
1040 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1044 * Initialize a preallocated and zeroed pmap structure,
1045 * such as one in a vmspace structure.
1048 pmap_pinit(struct pmap *pmap)
1053 * No need to allocate page table space yet but we do need a valid
1054 * page directory table.
1056 if (pmap->pm_pdir == NULL) {
1058 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE);
1062 * allocate object for the ptes
1064 if (pmap->pm_pteobj == NULL)
1065 pmap->pm_pteobj = vm_object_allocate( OBJT_DEFAULT, PTDPTDI + 1);
1068 * allocate the page directory page
1070 ptdpg = vm_page_grab( pmap->pm_pteobj, PTDPTDI,
1071 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1073 ptdpg->wire_count = 1;
1074 ++vmstats.v_wire_count;
1077 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY); /* not usually mapped*/
1078 ptdpg->valid = VM_PAGE_BITS_ALL;
1080 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1081 if ((ptdpg->flags & PG_ZERO) == 0)
1082 bzero(pmap->pm_pdir, PAGE_SIZE);
1084 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1086 /* install self-referential address mapping entry */
1087 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1088 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1091 pmap->pm_active = 0;
1092 pmap->pm_ptphint = NULL;
1093 TAILQ_INIT(&pmap->pm_pvlist);
1094 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1098 * Wire in kernel global address entries. To avoid a race condition
1099 * between pmap initialization and pmap_growkernel, this procedure
1100 * adds the pmap to the master list (which growkernel scans to update),
1101 * then copies the template.
1104 pmap_pinit2(struct pmap *pmap)
1107 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1108 /* XXX copies current process, does not fill in MPPTDI */
1109 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1114 * Attempt to release and free and vm_page in a pmap. Returns 1 on success,
1115 * 0 on failure (if the procedure had to sleep).
1118 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1120 unsigned *pde = (unsigned *) pmap->pm_pdir;
1122 * This code optimizes the case of freeing non-busy
1123 * page-table pages. Those pages are zero now, and
1124 * might as well be placed directly into the zero queue.
1126 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1132 * Remove the page table page from the processes address space.
1135 pmap->pm_stats.resident_count--;
1137 if (p->hold_count) {
1138 panic("pmap_release: freeing held page table page");
1141 * Page directory pages need to have the kernel
1142 * stuff cleared, so they can go into the zero queue also.
1144 if (p->pindex == PTDPTDI) {
1145 bzero(pde + KPTDI, nkpt * PTESIZE);
1148 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1151 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1152 pmap->pm_ptphint = NULL;
1155 vmstats.v_wire_count--;
1156 vm_page_free_zero(p);
1161 * this routine is called if the page table page is not
1165 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1167 vm_offset_t pteva, ptepa;
1171 * Find or fabricate a new pagetable page
1173 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1174 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1176 KASSERT(m->queue == PQ_NONE,
1177 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1179 if (m->wire_count == 0)
1180 vmstats.v_wire_count++;
1184 * Increment the hold count for the page table page
1185 * (denoting a new mapping.)
1190 * Map the pagetable page into the process address space, if
1191 * it isn't already there.
1194 pmap->pm_stats.resident_count++;
1196 ptepa = VM_PAGE_TO_PHYS(m);
1197 pmap->pm_pdir[ptepindex] =
1198 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1201 * Set the page table hint
1203 pmap->pm_ptphint = m;
1206 * Try to use the new mapping, but if we cannot, then
1207 * do it with the routine that maps the page explicitly.
1209 if ((m->flags & PG_ZERO) == 0) {
1210 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1211 (((unsigned) PTDpde) & PG_FRAME)) {
1212 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1213 bzero((caddr_t) pteva, PAGE_SIZE);
1215 pmap_zero_page(ptepa);
1219 m->valid = VM_PAGE_BITS_ALL;
1220 vm_page_flag_clear(m, PG_ZERO);
1221 vm_page_flag_set(m, PG_MAPPED);
1228 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1235 * Calculate pagetable page index
1237 ptepindex = va >> PDRSHIFT;
1240 * Get the page directory entry
1242 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1245 * This supports switching from a 4MB page to a
1248 if (ptepa & PG_PS) {
1249 pmap->pm_pdir[ptepindex] = 0;
1256 * If the page table page is mapped, we just increment the
1257 * hold count, and activate it.
1261 * In order to get the page table page, try the
1264 if (pmap->pm_ptphint &&
1265 (pmap->pm_ptphint->pindex == ptepindex)) {
1266 m = pmap->pm_ptphint;
1268 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1269 pmap->pm_ptphint = m;
1275 * Here if the pte page isn't mapped, or if it has been deallocated.
1277 return _pmap_allocpte(pmap, ptepindex);
1281 /***************************************************
1282 * Pmap allocation/deallocation routines.
1283 ***************************************************/
1286 * Release any resources held by the given physical map.
1287 * Called when a pmap initialized by pmap_pinit is being released.
1288 * Should only be called if the map contains no valid mappings.
1290 static int pmap_release_callback(struct vm_page *p, void *data);
1293 pmap_release(struct pmap *pmap)
1295 vm_object_t object = pmap->pm_pteobj;
1296 struct rb_vm_page_scan_info info;
1298 #if defined(DIAGNOSTIC)
1299 if (object->ref_count != 1)
1300 panic("pmap_release: pteobj reference count != 1");
1304 info.object = object;
1306 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1313 info.limit = object->generation;
1315 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1316 pmap_release_callback, &info);
1317 if (info.error == 0 && info.mpte) {
1318 if (!pmap_release_free_page(pmap, info.mpte))
1322 } while (info.error);
1326 pmap_release_callback(struct vm_page *p, void *data)
1328 struct rb_vm_page_scan_info *info = data;
1330 if (p->pindex == PTDPTDI) {
1334 if (!pmap_release_free_page(info->pmap, p)) {
1338 if (info->object->generation != info->limit) {
1346 kvm_size(SYSCTL_HANDLER_ARGS)
1348 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE;
1350 return sysctl_handle_long(oidp, &ksize, 0, req);
1352 SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD,
1353 0, 0, kvm_size, "IU", "Size of KVM");
1356 kvm_free(SYSCTL_HANDLER_ARGS)
1358 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end;
1360 return sysctl_handle_long(oidp, &kfree, 0, req);
1362 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD,
1363 0, 0, kvm_free, "IU", "Amount of KVM free");
1366 * Grow the number of kernel page table entries, if needed.
1370 pmap_growkernel(vm_offset_t addr)
1373 vm_offset_t ptppaddr;
1378 if (kernel_vm_end == 0) {
1379 kernel_vm_end = KERNBASE;
1381 while (pdir_pde(PTD, kernel_vm_end)) {
1382 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1386 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1387 while (kernel_vm_end < addr) {
1388 if (pdir_pde(PTD, kernel_vm_end)) {
1389 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1394 * This index is bogus, but out of the way
1396 nkpg = vm_page_alloc(kptobj, nkpt,
1397 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1399 panic("pmap_growkernel: no memory to grow kernel");
1402 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1403 pmap_zero_page(ptppaddr);
1404 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1405 pdir_pde(PTD, kernel_vm_end) = newpdir;
1406 *pmap_pde(kernel_pmap, kernel_vm_end) = newpdir;
1410 * This update must be interlocked with pmap_pinit2.
1412 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1413 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1415 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1416 ~(PAGE_SIZE * NPTEPG - 1);
1422 * Retire the given physical map from service.
1423 * Should only be called if the map contains
1424 * no valid mappings.
1427 pmap_destroy(pmap_t pmap)
1434 count = --pmap->pm_count;
1437 panic("destroying a pmap is not yet implemented");
1442 * Add a reference to the specified pmap.
1445 pmap_reference(pmap_t pmap)
1452 /***************************************************
1453 * page management routines.
1454 ***************************************************/
1457 * free the pv_entry back to the free list. This function may be
1458 * called from an interrupt.
1460 static PMAP_INLINE void
1461 free_pv_entry(pv_entry_t pv)
1468 * get a new pv_entry, allocating a block from the system
1469 * when needed. This function may be called from an interrupt.
1475 if (pv_entry_high_water &&
1476 (pv_entry_count > pv_entry_high_water) &&
1477 (pmap_pagedaemon_waken == 0)) {
1478 pmap_pagedaemon_waken = 1;
1479 wakeup (&vm_pages_needed);
1481 return zalloc(pvzone);
1485 * This routine is very drastic, but can save the system
1493 static int warningdone=0;
1495 if (pmap_pagedaemon_waken == 0)
1498 if (warningdone < 5) {
1499 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1503 for(i = 0; i < vm_page_array_size; i++) {
1504 m = &vm_page_array[i];
1505 if (m->wire_count || m->hold_count || m->busy ||
1506 (m->flags & PG_BUSY))
1510 pmap_pagedaemon_waken = 0;
1515 * If it is the first entry on the list, it is actually
1516 * in the header and we must copy the following entry up
1517 * to the header. Otherwise we must search the list for
1518 * the entry. In either case we free the now unused entry.
1521 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1522 vm_offset_t va, pmap_inval_info_t info)
1528 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1529 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1530 if (pmap == pv->pv_pmap && va == pv->pv_va)
1534 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1535 if (va == pv->pv_va)
1542 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1543 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1544 m->md.pv_list_count--;
1545 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1546 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1547 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1555 * Create a pv entry for page at pa for
1559 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1564 pv = get_pv_entry();
1569 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1570 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1571 m->md.pv_list_count++;
1577 * pmap_remove_pte: do the things to unmap a page in a process
1580 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1581 pmap_inval_info_t info)
1586 pmap_inval_add(info, pmap, va);
1587 oldpte = loadandclear(ptq);
1589 pmap->pm_stats.wired_count -= 1;
1591 * Machines that don't support invlpg, also don't support
1592 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1596 cpu_invlpg((void *)va);
1597 pmap->pm_stats.resident_count -= 1;
1598 if (oldpte & PG_MANAGED) {
1599 m = PHYS_TO_VM_PAGE(oldpte);
1600 if (oldpte & PG_M) {
1601 #if defined(PMAP_DIAGNOSTIC)
1602 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1604 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1608 if (pmap_track_modified(va))
1612 vm_page_flag_set(m, PG_REFERENCED);
1613 return pmap_remove_entry(pmap, m, va, info);
1615 return pmap_unuse_pt(pmap, va, NULL, info);
1624 * Remove a single page from a process address space.
1626 * This function may not be called from an interrupt if the pmap is
1630 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1635 * if there is no pte for this address, just skip it!!! Otherwise
1636 * get a local va for mappings for this pmap and remove the entry.
1638 if (*pmap_pde(pmap, va) != 0) {
1639 ptq = get_ptbase(pmap) + i386_btop(va);
1641 pmap_remove_pte(pmap, ptq, va, info);
1649 * Remove the given range of addresses from the specified map.
1651 * It is assumed that the start and end are properly
1652 * rounded to the page size.
1654 * This function may not be called from an interrupt if the pmap is
1658 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1662 vm_offset_t ptpaddr;
1663 vm_offset_t sindex, eindex;
1664 struct pmap_inval_info info;
1669 if (pmap->pm_stats.resident_count == 0)
1672 pmap_inval_init(&info);
1675 * special handling of removing one page. a very
1676 * common operation and easy to short circuit some
1679 if (((sva + PAGE_SIZE) == eva) &&
1680 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1681 pmap_remove_page(pmap, sva, &info);
1682 pmap_inval_flush(&info);
1687 * Get a local virtual address for the mappings that are being
1690 ptbase = get_ptbase(pmap);
1692 sindex = i386_btop(sva);
1693 eindex = i386_btop(eva);
1695 for (; sindex < eindex; sindex = pdnxt) {
1699 * Calculate index for next page table.
1701 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1702 if (pmap->pm_stats.resident_count == 0)
1705 pdirindex = sindex / NPDEPG;
1706 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1707 pmap_inval_add(&info, pmap, -1);
1708 pmap->pm_pdir[pdirindex] = 0;
1709 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1714 * Weed out invalid mappings. Note: we assume that the page
1715 * directory table is always allocated, and in kernel virtual.
1721 * Limit our scan to either the end of the va represented
1722 * by the current page table page, or to the end of the
1723 * range being removed.
1725 if (pdnxt > eindex) {
1729 for (; sindex != pdnxt; sindex++) {
1731 if (ptbase[sindex] == 0)
1733 va = i386_ptob(sindex);
1734 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1738 pmap_inval_flush(&info);
1744 * Removes this physical page from all physical maps in which it resides.
1745 * Reflects back modify bits to the pager.
1747 * This routine may not be called from an interrupt.
1751 pmap_remove_all(vm_page_t m)
1753 struct pmap_inval_info info;
1754 unsigned *pte, tpte;
1757 #if defined(PMAP_DIAGNOSTIC)
1759 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1762 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1763 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1767 pmap_inval_init(&info);
1769 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1770 pv->pv_pmap->pm_stats.resident_count--;
1772 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1773 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1775 tpte = loadandclear(pte);
1777 pv->pv_pmap->pm_stats.wired_count--;
1780 vm_page_flag_set(m, PG_REFERENCED);
1783 * Update the vm_page_t clean and reference bits.
1786 #if defined(PMAP_DIAGNOSTIC)
1787 if (pmap_nw_modified((pt_entry_t) tpte)) {
1789 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1793 if (pmap_track_modified(pv->pv_va))
1796 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1797 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1798 m->md.pv_list_count--;
1799 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1803 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1805 pmap_inval_flush(&info);
1811 * Set the physical protection on the specified range of this map
1814 * This function may not be called from an interrupt if the map is
1815 * not the kernel_pmap.
1818 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1821 vm_offset_t pdnxt, ptpaddr;
1822 vm_pindex_t sindex, eindex;
1823 pmap_inval_info info;
1828 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1829 pmap_remove(pmap, sva, eva);
1833 if (prot & VM_PROT_WRITE)
1836 pmap_inval_init(&info);
1838 ptbase = get_ptbase(pmap);
1840 sindex = i386_btop(sva);
1841 eindex = i386_btop(eva);
1843 for (; sindex < eindex; sindex = pdnxt) {
1847 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1849 pdirindex = sindex / NPDEPG;
1850 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1851 pmap_inval_add(&info, pmap, -1);
1852 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1853 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1858 * Weed out invalid mappings. Note: we assume that the page
1859 * directory table is always allocated, and in kernel virtual.
1864 if (pdnxt > eindex) {
1868 for (; sindex != pdnxt; sindex++) {
1873 /* XXX this isn't optimal */
1874 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1875 pbits = ptbase[sindex];
1877 if (pbits & PG_MANAGED) {
1880 m = PHYS_TO_VM_PAGE(pbits);
1881 vm_page_flag_set(m, PG_REFERENCED);
1885 if (pmap_track_modified(i386_ptob(sindex))) {
1887 m = PHYS_TO_VM_PAGE(pbits);
1896 if (pbits != ptbase[sindex]) {
1897 ptbase[sindex] = pbits;
1901 pmap_inval_flush(&info);
1905 * Insert the given physical page (p) at
1906 * the specified virtual address (v) in the
1907 * target physical map with the protection requested.
1909 * If specified, the page will be wired down, meaning
1910 * that the related pte can not be reclaimed.
1912 * NB: This is the only routine which MAY NOT lazy-evaluate
1913 * or lose information. That is, this routine must actually
1914 * insert this page into the given map NOW.
1917 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1923 vm_offset_t origpte, newpte;
1925 pmap_inval_info info;
1931 #ifdef PMAP_DIAGNOSTIC
1932 if (va > VM_MAX_KERNEL_ADDRESS)
1933 panic("pmap_enter: toobig");
1934 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1935 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1940 * In the case that a page table page is not
1941 * resident, we are creating it here.
1943 if (va < UPT_MIN_ADDRESS) {
1944 mpte = pmap_allocpte(pmap, va);
1947 pmap_inval_init(&info);
1948 pte = pmap_pte(pmap, va);
1951 * Page Directory table entry not valid, we need a new PT page
1954 panic("pmap_enter: invalid page directory pdir=%x, va=0x%x\n",
1955 (unsigned) pmap->pm_pdir[PTDPTDI], va);
1958 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
1959 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1960 origpte = *(vm_offset_t *)pte;
1961 opa = origpte & PG_FRAME;
1963 if (origpte & PG_PS)
1964 panic("pmap_enter: attempted pmap_enter on 4MB page");
1967 * Mapping has not changed, must be protection or wiring change.
1969 if (origpte && (opa == pa)) {
1971 * Wiring change, just update stats. We don't worry about
1972 * wiring PT pages as they remain resident as long as there
1973 * are valid mappings in them. Hence, if a user page is wired,
1974 * the PT page will be also.
1976 if (wired && ((origpte & PG_W) == 0))
1977 pmap->pm_stats.wired_count++;
1978 else if (!wired && (origpte & PG_W))
1979 pmap->pm_stats.wired_count--;
1981 #if defined(PMAP_DIAGNOSTIC)
1982 if (pmap_nw_modified((pt_entry_t) origpte)) {
1984 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1990 * Remove the extra pte reference. Note that we cannot
1991 * optimize the RO->RW case because we have adjusted the
1992 * wiring count above and may need to adjust the wiring
1999 * We might be turning off write access to the page,
2000 * so we go ahead and sense modify status.
2002 if (origpte & PG_MANAGED) {
2003 if ((origpte & PG_M) && pmap_track_modified(va)) {
2005 om = PHYS_TO_VM_PAGE(opa);
2013 * Mapping has changed, invalidate old range and fall through to
2014 * handle validating new mapping.
2018 err = pmap_remove_pte(pmap, pte, va, &info);
2020 panic("pmap_enter: pte vanished, va: 0x%x", va);
2024 * Enter on the PV list if part of our managed memory. Note that we
2025 * raise IPL while manipulating pv_table since pmap_enter can be
2026 * called at interrupt time.
2028 if (pmap_initialized &&
2029 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2030 pmap_insert_entry(pmap, va, mpte, m);
2035 * Increment counters
2037 pmap->pm_stats.resident_count++;
2039 pmap->pm_stats.wired_count++;
2043 * Now validate mapping with desired protection/wiring.
2045 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2049 if (va < UPT_MIN_ADDRESS)
2051 if (pmap == kernel_pmap)
2055 * if the mapping or permission bits are different, we need
2056 * to update the pte.
2058 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2059 *pte = newpte | PG_A;
2061 pmap_inval_flush(&info);
2065 * this code makes some *MAJOR* assumptions:
2066 * 1. Current pmap & pmap exists.
2069 * 4. No page table pages.
2070 * 5. Tlbflush is deferred to calling procedure.
2071 * 6. Page IS managed.
2072 * but is *MUCH* faster than pmap_enter...
2076 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
2080 pmap_inval_info info;
2082 pmap_inval_init(&info);
2085 * In the case that a page table page is not
2086 * resident, we are creating it here.
2088 if (va < UPT_MIN_ADDRESS) {
2093 * Calculate pagetable page index
2095 ptepindex = va >> PDRSHIFT;
2096 if (mpte && (mpte->pindex == ptepindex)) {
2101 * Get the page directory entry
2103 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2106 * If the page table page is mapped, we just increment
2107 * the hold count, and activate it.
2111 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2112 if (pmap->pm_ptphint &&
2113 (pmap->pm_ptphint->pindex == ptepindex)) {
2114 mpte = pmap->pm_ptphint;
2116 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2117 pmap->pm_ptphint = mpte;
2123 mpte = _pmap_allocpte(pmap, ptepindex);
2131 * This call to vtopte makes the assumption that we are
2132 * entering the page into the current pmap. In order to support
2133 * quick entry into any pmap, one would likely use pmap_pte_quick.
2134 * But that isn't as quick as vtopte.
2136 pte = (unsigned *)vtopte(va);
2139 pmap_unwire_pte_hold(pmap, mpte, &info);
2144 * Enter on the PV list if part of our managed memory. Note that we
2145 * raise IPL while manipulating pv_table since pmap_enter can be
2146 * called at interrupt time.
2148 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
2149 pmap_insert_entry(pmap, va, mpte, m);
2152 * Increment counters
2154 pmap->pm_stats.resident_count++;
2156 pa = VM_PAGE_TO_PHYS(m);
2159 * Now validate mapping with RO protection
2161 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2162 *pte = pa | PG_V | PG_U;
2164 *pte = pa | PG_V | PG_U | PG_MANAGED;
2170 * Make a temporary mapping for a physical address. This is only intended
2171 * to be used for panic dumps.
2174 pmap_kenter_temporary(vm_paddr_t pa, int i)
2176 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2177 return ((void *)crashdumpmap);
2180 #define MAX_INIT_PT (96)
2183 * This routine preloads the ptes for a given object into the specified pmap.
2184 * This eliminates the blast of soft faults on process startup and
2185 * immediately after an mmap.
2187 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2190 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2191 vm_object_t object, vm_pindex_t pindex,
2192 vm_size_t size, int limit)
2194 struct rb_vm_page_scan_info info;
2198 * We can't preinit if read access isn't set or there is no pmap
2201 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2205 * We can't preinit if the pmap is not the current pmap
2207 if (curproc == NULL || pmap != vmspace_pmap(curproc->p_vmspace))
2210 psize = i386_btop(size);
2212 if ((object->type != OBJT_VNODE) ||
2213 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2214 (object->resident_page_count > MAX_INIT_PT))) {
2218 if (psize + pindex > object->size) {
2219 if (object->size < pindex)
2221 psize = object->size - pindex;
2228 * Use a red-black scan to traverse the requested range and load
2229 * any valid pages found into the pmap.
2231 * We cannot safely scan the object's memq unless we are in a
2232 * critical section since interrupts can remove pages from objects.
2234 info.start_pindex = pindex;
2235 info.end_pindex = pindex + psize - 1;
2242 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2243 pmap_object_init_pt_callback, &info);
2249 pmap_object_init_pt_callback(vm_page_t p, void *data)
2251 struct rb_vm_page_scan_info *info = data;
2252 vm_pindex_t rel_index;
2254 * don't allow an madvise to blow away our really
2255 * free pages allocating pv entries.
2257 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2258 vmstats.v_free_count < vmstats.v_free_reserved) {
2261 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2262 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2263 if ((p->queue - p->pc) == PQ_CACHE)
2264 vm_page_deactivate(p);
2266 rel_index = p->pindex - info->start_pindex;
2267 info->mpte = pmap_enter_quick(info->pmap,
2268 info->addr + i386_ptob(rel_index),
2270 vm_page_flag_set(p, PG_MAPPED);
2277 * pmap_prefault provides a quick way of clustering pagefaults into a
2278 * processes address space. It is a "cousin" of pmap_object_init_pt,
2279 * except it runs at page fault time instead of mmap time.
2283 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2285 static int pmap_prefault_pageorder[] = {
2286 -PAGE_SIZE, PAGE_SIZE,
2287 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2288 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2289 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2293 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2303 * We do not currently prefault mappings that use virtual page
2304 * tables. We do not prefault foreign pmaps.
2306 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2308 if (curproc == NULL || (pmap != vmspace_pmap(curproc->p_vmspace)))
2311 object = entry->object.vm_object;
2313 starta = addra - PFBAK * PAGE_SIZE;
2314 if (starta < entry->start)
2315 starta = entry->start;
2316 else if (starta > addra)
2320 * critical section protection is required to maintain the
2321 * page/object association, interrupts can free pages and remove
2322 * them from their objects.
2326 for (i = 0; i < PAGEORDER_SIZE; i++) {
2327 vm_object_t lobject;
2330 addr = addra + pmap_prefault_pageorder[i];
2331 if (addr > addra + (PFFOR * PAGE_SIZE))
2334 if (addr < starta || addr >= entry->end)
2337 if ((*pmap_pde(pmap, addr)) == NULL)
2340 pte = (unsigned *) vtopte(addr);
2344 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2347 for (m = vm_page_lookup(lobject, pindex);
2348 (!m && (lobject->type == OBJT_DEFAULT) &&
2349 (lobject->backing_object));
2350 lobject = lobject->backing_object
2352 if (lobject->backing_object_offset & PAGE_MASK)
2354 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2355 m = vm_page_lookup(lobject->backing_object, pindex);
2359 * give-up when a page is not in memory
2364 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2366 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2368 if ((m->queue - m->pc) == PQ_CACHE) {
2369 vm_page_deactivate(m);
2372 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2373 vm_page_flag_set(m, PG_MAPPED);
2381 * Routine: pmap_change_wiring
2382 * Function: Change the wiring attribute for a map/virtual-address
2384 * In/out conditions:
2385 * The mapping must already exist in the pmap.
2388 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2395 pte = pmap_pte(pmap, va);
2397 if (wired && !pmap_pte_w(pte))
2398 pmap->pm_stats.wired_count++;
2399 else if (!wired && pmap_pte_w(pte))
2400 pmap->pm_stats.wired_count--;
2403 * Wiring is not a hardware characteristic so there is no need to
2404 * invalidate TLB. However, in an SMP environment we must use
2405 * a locked bus cycle to update the pte (if we are not using
2406 * the pmap_inval_*() API that is)... it's ok to do this for simple
2411 atomic_set_int(pte, PG_W);
2413 atomic_clear_int(pte, PG_W);
2416 atomic_set_int_nonlocked(pte, PG_W);
2418 atomic_clear_int_nonlocked(pte, PG_W);
2425 * Copy the range specified by src_addr/len
2426 * from the source map to the range dst_addr/len
2427 * in the destination map.
2429 * This routine is only advisory and need not do anything.
2432 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2433 vm_size_t len, vm_offset_t src_addr)
2435 pmap_inval_info info;
2437 vm_offset_t end_addr = src_addr + len;
2439 unsigned src_frame, dst_frame;
2442 if (dst_addr != src_addr)
2445 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2446 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2450 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2451 if (dst_frame != (((unsigned) APTDpde) & PG_FRAME)) {
2452 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2453 /* The page directory is not shared between CPUs */
2456 pmap_inval_init(&info);
2457 pmap_inval_add(&info, dst_pmap, -1);
2458 pmap_inval_add(&info, src_pmap, -1);
2461 * critical section protection is required to maintain the page/object
2462 * association, interrupts can free pages and remove them from
2466 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2467 unsigned *src_pte, *dst_pte;
2468 vm_page_t dstmpte, srcmpte;
2469 vm_offset_t srcptepaddr;
2472 if (addr >= UPT_MIN_ADDRESS)
2473 panic("pmap_copy: invalid to pmap_copy page tables\n");
2476 * Don't let optional prefaulting of pages make us go
2477 * way below the low water mark of free pages or way
2478 * above high water mark of used pv entries.
2480 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2481 pv_entry_count > pv_entry_high_water)
2484 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2485 ptepindex = addr >> PDRSHIFT;
2487 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2488 if (srcptepaddr == 0)
2491 if (srcptepaddr & PG_PS) {
2492 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2493 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2494 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2499 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2500 if ((srcmpte == NULL) ||
2501 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2504 if (pdnxt > end_addr)
2507 src_pte = (unsigned *) vtopte(addr);
2508 dst_pte = (unsigned *) avtopte(addr);
2509 while (addr < pdnxt) {
2513 * we only virtual copy managed pages
2515 if ((ptetemp & PG_MANAGED) != 0) {
2517 * We have to check after allocpte for the
2518 * pte still being around... allocpte can
2521 dstmpte = pmap_allocpte(dst_pmap, addr);
2522 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2524 * Clear the modified and
2525 * accessed (referenced) bits
2528 m = PHYS_TO_VM_PAGE(ptetemp);
2529 *dst_pte = ptetemp & ~(PG_M | PG_A);
2530 dst_pmap->pm_stats.resident_count++;
2531 pmap_insert_entry(dst_pmap, addr,
2534 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2536 if (dstmpte->hold_count >= srcmpte->hold_count)
2545 pmap_inval_flush(&info);
2549 * Routine: pmap_kernel
2551 * Returns the physical map handle for the kernel.
2556 return (kernel_pmap);
2562 * Zero the specified PA by mapping the page into KVM and clearing its
2565 * This function may be called from an interrupt and no locking is
2569 pmap_zero_page(vm_paddr_t phys)
2571 struct mdglobaldata *gd = mdcpu;
2574 if (*(int *)gd->gd_CMAP3)
2575 panic("pmap_zero_page: CMAP3 busy");
2576 *(int *)gd->gd_CMAP3 =
2577 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2578 cpu_invlpg(gd->gd_CADDR3);
2580 #if defined(I686_CPU)
2581 if (cpu_class == CPUCLASS_686)
2582 i686_pagezero(gd->gd_CADDR3);
2585 bzero(gd->gd_CADDR3, PAGE_SIZE);
2586 *(int *) gd->gd_CMAP3 = 0;
2591 * pmap_page_assertzero:
2593 * Assert that a page is empty, panic if it isn't.
2596 pmap_page_assertzero(vm_paddr_t phys)
2598 struct mdglobaldata *gd = mdcpu;
2602 if (*(int *)gd->gd_CMAP3)
2603 panic("pmap_zero_page: CMAP3 busy");
2604 *(int *)gd->gd_CMAP3 =
2605 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2606 cpu_invlpg(gd->gd_CADDR3);
2607 for (i = 0; i < PAGE_SIZE; i += 4) {
2608 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2609 panic("pmap_page_assertzero() @ %p not zero!\n",
2610 (void *)gd->gd_CADDR3);
2613 *(int *) gd->gd_CMAP3 = 0;
2620 * Zero part of a physical page by mapping it into memory and clearing
2621 * its contents with bzero.
2623 * off and size may not cover an area beyond a single hardware page.
2626 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2628 struct mdglobaldata *gd = mdcpu;
2631 if (*(int *) gd->gd_CMAP3)
2632 panic("pmap_zero_page: CMAP3 busy");
2633 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2634 cpu_invlpg(gd->gd_CADDR3);
2636 #if defined(I686_CPU)
2637 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2638 i686_pagezero(gd->gd_CADDR3);
2641 bzero((char *)gd->gd_CADDR3 + off, size);
2642 *(int *) gd->gd_CMAP3 = 0;
2649 * Copy the physical page from the source PA to the target PA.
2650 * This function may be called from an interrupt. No locking
2654 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2656 struct mdglobaldata *gd = mdcpu;
2659 if (*(int *) gd->gd_CMAP1)
2660 panic("pmap_copy_page: CMAP1 busy");
2661 if (*(int *) gd->gd_CMAP2)
2662 panic("pmap_copy_page: CMAP2 busy");
2664 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2665 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2667 cpu_invlpg(gd->gd_CADDR1);
2668 cpu_invlpg(gd->gd_CADDR2);
2670 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2672 *(int *) gd->gd_CMAP1 = 0;
2673 *(int *) gd->gd_CMAP2 = 0;
2678 * pmap_copy_page_frag:
2680 * Copy the physical page from the source PA to the target PA.
2681 * This function may be called from an interrupt. No locking
2685 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2687 struct mdglobaldata *gd = mdcpu;
2690 if (*(int *) gd->gd_CMAP1)
2691 panic("pmap_copy_page: CMAP1 busy");
2692 if (*(int *) gd->gd_CMAP2)
2693 panic("pmap_copy_page: CMAP2 busy");
2695 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2696 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2698 cpu_invlpg(gd->gd_CADDR1);
2699 cpu_invlpg(gd->gd_CADDR2);
2701 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2702 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2705 *(int *) gd->gd_CMAP1 = 0;
2706 *(int *) gd->gd_CMAP2 = 0;
2711 * Returns true if the pmap's pv is one of the first
2712 * 16 pvs linked to from this page. This count may
2713 * be changed upwards or downwards in the future; it
2714 * is only necessary that true be returned for a small
2715 * subset of pmaps for proper page aging.
2718 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2723 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2728 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2729 if (pv->pv_pmap == pmap) {
2742 * Remove all pages from specified address space
2743 * this aids process exit speeds. Also, this code
2744 * is special cased for current process only, but
2745 * can have the more generic (and slightly slower)
2746 * mode enabled. This is much faster than pmap_remove
2747 * in the case of running down an entire address space.
2750 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2752 unsigned *pte, tpte;
2755 pmap_inval_info info;
2758 if (curproc && pmap == vmspace_pmap(curproc->p_vmspace))
2763 pmap_inval_init(&info);
2765 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2767 if (pv->pv_va >= eva || pv->pv_va < sva) {
2768 npv = TAILQ_NEXT(pv, pv_plist);
2773 pte = (unsigned *)vtopte(pv->pv_va);
2775 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2776 if (pmap->pm_active)
2777 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2781 * We cannot remove wired pages from a process' mapping
2785 npv = TAILQ_NEXT(pv, pv_plist);
2790 m = PHYS_TO_VM_PAGE(tpte);
2792 KASSERT(m < &vm_page_array[vm_page_array_size],
2793 ("pmap_remove_pages: bad tpte %x", tpte));
2795 pv->pv_pmap->pm_stats.resident_count--;
2798 * Update the vm_page_t clean and reference bits.
2805 npv = TAILQ_NEXT(pv, pv_plist);
2806 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2808 m->md.pv_list_count--;
2809 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2810 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2811 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2814 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2817 pmap_inval_flush(&info);
2822 * pmap_testbit tests bits in pte's
2823 * note that the testbit/changebit routines are inline,
2824 * and a lot of things compile-time evaluate.
2827 pmap_testbit(vm_page_t m, int bit)
2832 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2835 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2840 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2842 * if the bit being tested is the modified bit, then
2843 * mark clean_map and ptes as never
2846 if (bit & (PG_A|PG_M)) {
2847 if (!pmap_track_modified(pv->pv_va))
2851 #if defined(PMAP_DIAGNOSTIC)
2853 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2857 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2868 * this routine is used to modify bits in ptes
2870 static __inline void
2871 pmap_changebit(vm_page_t m, int bit, boolean_t setem)
2873 struct pmap_inval_info info;
2877 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2880 pmap_inval_init(&info);
2884 * Loop over all current mappings setting/clearing as appropos If
2885 * setting RO do we need to clear the VAC?
2887 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2889 * don't write protect pager mappings
2891 if (!setem && (bit == PG_RW)) {
2892 if (!pmap_track_modified(pv->pv_va))
2896 #if defined(PMAP_DIAGNOSTIC)
2898 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2904 * Careful here. We can use a locked bus instruction to
2905 * clear PG_A or PG_M safely but we need to synchronize
2906 * with the target cpus when we mess with PG_RW.
2908 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2910 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2914 atomic_set_int(pte, bit);
2916 atomic_set_int_nonlocked(pte, bit);
2919 vm_offset_t pbits = *(vm_offset_t *)pte;
2926 atomic_clear_int(pte, PG_M|PG_RW);
2928 atomic_clear_int_nonlocked(pte, PG_M|PG_RW);
2932 atomic_clear_int(pte, bit);
2934 atomic_clear_int_nonlocked(pte, bit);
2940 pmap_inval_flush(&info);
2945 * pmap_page_protect:
2947 * Lower the permission for all mappings to a given page.
2950 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2952 if ((prot & VM_PROT_WRITE) == 0) {
2953 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2954 pmap_changebit(m, PG_RW, FALSE);
2962 pmap_phys_address(int ppn)
2964 return (i386_ptob(ppn));
2968 * pmap_ts_referenced:
2970 * Return a count of reference bits for a page, clearing those bits.
2971 * It is not necessary for every reference bit to be cleared, but it
2972 * is necessary that 0 only be returned when there are truly no
2973 * reference bits set.
2975 * XXX: The exact number of bits to check and clear is a matter that
2976 * should be tested and standardized at some point in the future for
2977 * optimal aging of shared pages.
2980 pmap_ts_referenced(vm_page_t m)
2982 pv_entry_t pv, pvf, pvn;
2986 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2991 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2996 pvn = TAILQ_NEXT(pv, pv_list);
2998 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3000 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3002 if (!pmap_track_modified(pv->pv_va))
3005 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3007 if (pte && (*pte & PG_A)) {
3009 atomic_clear_int(pte, PG_A);
3011 atomic_clear_int_nonlocked(pte, PG_A);
3018 } while ((pv = pvn) != NULL && pv != pvf);
3028 * Return whether or not the specified physical page was modified
3029 * in any physical maps.
3032 pmap_is_modified(vm_page_t m)
3034 return pmap_testbit(m, PG_M);
3038 * Clear the modify bits on the specified physical page.
3041 pmap_clear_modify(vm_page_t m)
3043 pmap_changebit(m, PG_M, FALSE);
3047 * pmap_clear_reference:
3049 * Clear the reference bit on the specified physical page.
3052 pmap_clear_reference(vm_page_t m)
3054 pmap_changebit(m, PG_A, FALSE);
3058 * Miscellaneous support routines follow
3062 i386_protection_init(void)
3066 kp = protection_codes;
3067 for (prot = 0; prot < 8; prot++) {
3069 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3071 * Read access is also 0. There isn't any execute bit,
3072 * so just make it readable.
3074 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3075 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3076 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3079 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3080 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3081 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3082 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3090 * Map a set of physical memory pages into the kernel virtual
3091 * address space. Return a pointer to where it is mapped. This
3092 * routine is intended to be used for mapping device memory,
3095 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3099 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3101 vm_offset_t va, tmpva, offset;
3104 offset = pa & PAGE_MASK;
3105 size = roundup(offset + size, PAGE_SIZE);
3107 va = kmem_alloc_nofault(kernel_map, size);
3109 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3112 for (tmpva = va; size > 0;) {
3113 pte = (unsigned *)vtopte(tmpva);
3114 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3122 return ((void *)(va + offset));
3126 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3128 vm_offset_t base, offset;
3130 base = va & PG_FRAME;
3131 offset = va & PAGE_MASK;
3132 size = roundup(offset + size, PAGE_SIZE);
3133 pmap_qremove(va, size >> PAGE_SHIFT);
3134 kmem_free(kernel_map, base, size);
3138 * perform the pmap work for mincore
3141 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3143 unsigned *ptep, pte;
3147 ptep = pmap_pte(pmap, addr);
3152 if ((pte = *ptep) != 0) {
3155 val = MINCORE_INCORE;
3156 if ((pte & PG_MANAGED) == 0)
3159 pa = pte & PG_FRAME;
3161 m = PHYS_TO_VM_PAGE(pa);
3167 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3169 * Modified by someone
3171 else if (m->dirty || pmap_is_modified(m))
3172 val |= MINCORE_MODIFIED_OTHER;
3177 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3180 * Referenced by someone
3182 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3183 val |= MINCORE_REFERENCED_OTHER;
3184 vm_page_flag_set(m, PG_REFERENCED);
3191 pmap_activate(struct proc *p)
3195 pmap = vmspace_pmap(p->p_vmspace);
3197 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3199 pmap->pm_active |= 1;
3201 #if defined(SWTCH_OPTIM_STATS)
3204 p->p_thread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3205 load_cr3(p->p_thread->td_pcb->pcb_cr3);
3209 pmap_deactivate(struct proc *p)
3213 pmap = vmspace_pmap(p->p_vmspace);
3215 atomic_clear_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3217 pmap->pm_active &= ~1;
3220 * XXX - note we do not adjust %cr3. The caller is expected to
3221 * activate a new pmap or do a thread-exit.
3226 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3229 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3233 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3240 static void pads (pmap_t pm);
3241 void pmap_pvdump (vm_paddr_t pa);
3243 /* print address space of pmap*/
3250 if (pm == kernel_pmap)
3252 for (i = 0; i < 1024; i++)
3254 for (j = 0; j < 1024; j++) {
3255 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3256 if (pm == kernel_pmap && va < KERNBASE)
3258 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS)
3260 ptep = pmap_pte_quick(pm, va);
3261 if (pmap_pte_v(ptep))
3262 kprintf("%x:%x ", va, *(int *) ptep);
3268 pmap_pvdump(vm_paddr_t pa)
3273 kprintf("pa %08llx", (long long)pa);
3274 m = PHYS_TO_VM_PAGE(pa);
3275 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3277 kprintf(" -> pmap %p, va %x, flags %x",
3278 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3280 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);