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 1.87 2008/08/25 17:01:38 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 * PMAP_DEBUG - see platform/pc32/include/pmap.h
75 #include "opt_disable_pse.h"
77 #include "opt_msgbuf.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
83 #include <sys/msgbuf.h>
84 #include <sys/vmmeter.h>
88 #include <vm/vm_param.h>
89 #include <sys/sysctl.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_pageout.h>
97 #include <vm/vm_pager.h>
98 #include <vm/vm_zone.h>
100 #include <sys/user.h>
101 #include <sys/thread2.h>
102 #include <sys/sysref2.h>
104 #include <machine/cputypes.h>
105 #include <machine/md_var.h>
106 #include <machine/specialreg.h>
107 #include <machine/smp.h>
108 #include <machine_base/apic/apicreg.h>
109 #include <machine/globaldata.h>
110 #include <machine/pmap.h>
111 #include <machine/pmap_inval.h>
113 #define PMAP_KEEP_PDIRS
114 #ifndef PMAP_SHPGPERPROC
115 #define PMAP_SHPGPERPROC 200
118 #if defined(DIAGNOSTIC)
119 #define PMAP_DIAGNOSTIC
124 #if !defined(PMAP_DIAGNOSTIC)
125 #define PMAP_INLINE __inline
131 * Get PDEs and PTEs for user/kernel address space
133 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
134 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
136 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
137 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
138 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
139 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
140 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
144 * Given a map and a machine independent protection code,
145 * convert to a vax protection code.
147 #define pte_prot(m, p) \
148 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
149 static int protection_codes[8];
151 struct pmap kernel_pmap;
152 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
154 vm_paddr_t avail_start; /* PA of first available physical page */
155 vm_paddr_t avail_end; /* PA of last available physical page */
156 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
157 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
158 vm_offset_t virtual2_start;
159 vm_offset_t virtual2_end;
160 vm_offset_t KvaStart; /* VA start of KVA space */
161 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
162 vm_offset_t KvaSize; /* max size of kernel virtual address space */
163 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
164 static int pgeflag; /* PG_G or-in */
165 static int pseflag; /* PG_PS or-in */
167 static vm_object_t kptobj;
170 vm_offset_t kernel_vm_end;
173 * Data for the pv entry allocation mechanism
175 static vm_zone_t pvzone;
176 static struct vm_zone pvzone_store;
177 static struct vm_object pvzone_obj;
178 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
179 static int pmap_pagedaemon_waken = 0;
180 static struct pv_entry *pvinit;
183 * Considering all the issues I'm having with pmap caching, if breakage
184 * continues to occur, and for debugging, I've added a sysctl that will
185 * just do an unconditional invltlb.
187 static int dreadful_invltlb;
189 SYSCTL_INT(_vm, OID_AUTO, dreadful_invltlb,
190 CTLFLAG_RW, &dreadful_invltlb, 0, "");
193 * All those kernel PT submaps that BSD is so fond of
195 pt_entry_t *CMAP1 = 0, *ptmmap;
196 caddr_t CADDR1 = 0, ptvmmap = 0;
197 static pt_entry_t *msgbufmap;
198 struct msgbuf *msgbufp=0;
203 static pt_entry_t *pt_crashdumpmap;
204 static caddr_t crashdumpmap;
206 extern pt_entry_t *SMPpt;
208 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
209 static unsigned * get_ptbase (pmap_t pmap);
210 static pv_entry_t get_pv_entry (void);
211 static void i386_protection_init (void);
212 static __inline void pmap_clearbit (vm_page_t m, int bit);
214 static void pmap_remove_all (vm_page_t m);
215 static int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
216 vm_offset_t sva, pmap_inval_info_t info);
217 static void pmap_remove_page (struct pmap *pmap,
218 vm_offset_t va, pmap_inval_info_t info);
219 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
220 vm_offset_t va, pmap_inval_info_t info);
221 static boolean_t pmap_testbit (vm_page_t m, int bit);
222 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
223 vm_page_t mpte, vm_page_t m);
225 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
227 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
228 static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
229 static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
230 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
231 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
232 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
234 static unsigned pdir4mb;
237 * Move the kernel virtual free pointer to the next
238 * 4MB. This is used to help improve performance
239 * by using a large (4MB) page for much of the kernel
240 * (.text, .data, .bss)
243 pmap_kmem_choose(vm_offset_t addr)
245 vm_offset_t newaddr = addr;
247 if (cpu_feature & CPUID_PSE) {
248 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
257 * Extract the page table entry associated with the given map/virtual
260 * This function may NOT be called from an interrupt.
262 PMAP_INLINE unsigned *
263 pmap_pte(pmap_t pmap, vm_offset_t va)
268 pdeaddr = (unsigned *) pmap_pde(pmap, va);
269 if (*pdeaddr & PG_PS)
272 return get_ptbase(pmap) + i386_btop(va);
281 * Super fast pmap_pte routine best used when scanning the pv lists.
282 * This eliminates many course-grained invltlb calls. Note that many of
283 * the pv list scans are across different pmaps and it is very wasteful
284 * to do an entire invltlb when checking a single mapping.
286 * Should only be called while in a critical section.
288 * Unlike get_ptbase(), this function MAY be called from an interrupt or
292 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
294 struct mdglobaldata *gd = mdcpu;
297 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
298 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
299 unsigned index = i386_btop(va);
300 /* are we current address space or kernel? */
301 if ((pmap == &kernel_pmap) ||
302 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
303 return (unsigned *) PTmap + index;
305 newpf = pde & PG_FRAME;
306 if ( ((* (unsigned *) gd->gd_PMAP1) & PG_FRAME) != newpf) {
307 * (unsigned *) gd->gd_PMAP1 = newpf | PG_RW | PG_V;
308 cpu_invlpg(gd->gd_PADDR1);
310 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
317 * Bootstrap the system enough to run with virtual memory.
319 * On the i386 this is called after mapping has already been enabled
320 * and just syncs the pmap module with what has already been done.
321 * [We can't call it easily with mapping off since the kernel is not
322 * mapped with PA == VA, hence we would have to relocate every address
323 * from the linked base (virtual) address "KERNBASE" to the actual
324 * (physical) address starting relative to 0]
327 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
331 struct mdglobaldata *gd;
335 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
336 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
337 KvaEnd = KvaStart + KvaSize;
339 avail_start = firstaddr;
342 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
343 * too large. It should instead be correctly calculated in locore.s and
344 * not based on 'first' (which is a physical address, not a virtual
345 * address, for the start of unused physical memory). The kernel
346 * page tables are NOT double mapped and thus should not be included
347 * in this calculation.
349 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
350 virtual_start = pmap_kmem_choose(virtual_start);
351 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
354 * Initialize protection array.
356 i386_protection_init();
359 * The kernel's pmap is statically allocated so we don't have to use
360 * pmap_create, which is unlikely to work correctly at this part of
361 * the boot sequence (XXX and which no longer exists).
363 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
364 kernel_pmap.pm_count = 1;
365 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
366 TAILQ_INIT(&kernel_pmap.pm_pvlist);
370 * Reserve some special page table entries/VA space for temporary
373 #define SYSMAP(c, p, v, n) \
374 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
377 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
380 * CMAP1/CMAP2 are used for zeroing and copying pages.
382 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
387 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
390 * ptvmmap is used for reading arbitrary physical pages via
393 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
396 * msgbufp is used to map the system message buffer.
397 * XXX msgbufmap is not used.
399 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
400 atop(round_page(MSGBUF_SIZE)))
405 for (i = 0; i < NKPT; i++)
409 * PG_G is terribly broken on SMP because we IPI invltlb's in some
410 * cases rather then invl1pg. Actually, I don't even know why it
411 * works under UP because self-referential page table mappings
416 if (cpu_feature & CPUID_PGE)
421 * Initialize the 4MB page size flag
425 * The 4MB page version of the initial
426 * kernel page mapping.
430 #if !defined(DISABLE_PSE)
431 if (cpu_feature & CPUID_PSE) {
434 * Note that we have enabled PSE mode
437 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
438 ptditmp &= ~(NBPDR - 1);
439 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
444 * Enable the PSE mode. If we are SMP we can't do this
445 * now because the APs will not be able to use it when
448 load_cr4(rcr4() | CR4_PSE);
451 * We can do the mapping here for the single processor
452 * case. We simply ignore the old page table page from
456 * For SMP, we still need 4K pages to bootstrap APs,
457 * PSE will be enabled as soon as all APs are up.
459 PTD[KPTDI] = (pd_entry_t)ptditmp;
460 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
467 * We need to finish setting up the globaldata page for the BSP.
468 * locore has already populated the page table for the mdglobaldata
471 pg = MDGLOBALDATA_BASEALLOC_PAGES;
472 gd = &CPU_prvspace[0].mdglobaldata;
473 gd->gd_CMAP1 = &SMPpt[pg + 0];
474 gd->gd_CMAP2 = &SMPpt[pg + 1];
475 gd->gd_CMAP3 = &SMPpt[pg + 2];
476 gd->gd_PMAP1 = &SMPpt[pg + 3];
477 gd->gd_GDMAP1 = &PTD[KGDTDI];
478 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
479 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
480 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
481 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
482 gd->gd_GDADDR1= (unsigned *)VADDR(KGDTDI, 0);
489 * Set 4mb pdir for mp startup
494 if (pseflag && (cpu_feature & CPUID_PSE)) {
495 load_cr4(rcr4() | CR4_PSE);
496 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
497 kernel_pmap.pm_pdir[KPTDI] =
498 PTD[KPTDI] = (pd_entry_t)pdir4mb;
506 * Initialize the pmap module.
507 * Called by vm_init, to initialize any structures that the pmap
508 * system needs to map virtual memory.
509 * pmap_init has been enhanced to support in a fairly consistant
510 * way, discontiguous physical memory.
519 * object for kernel page table pages
521 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
524 * Allocate memory for random pmap data structures. Includes the
528 for(i = 0; i < vm_page_array_size; i++) {
531 m = &vm_page_array[i];
532 TAILQ_INIT(&m->md.pv_list);
533 m->md.pv_list_count = 0;
537 * init the pv free list
539 initial_pvs = vm_page_array_size;
540 if (initial_pvs < MINPV)
542 pvzone = &pvzone_store;
543 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
544 initial_pvs * sizeof (struct pv_entry));
545 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
549 * Now it is safe to enable pv_table recording.
551 pmap_initialized = TRUE;
555 * Initialize the address space (zone) for the pv_entries. Set a
556 * high water mark so that the system can recover from excessive
557 * numbers of pv entries.
562 int shpgperproc = PMAP_SHPGPERPROC;
564 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
565 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
566 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
567 pv_entry_high_water = 9 * (pv_entry_max / 10);
568 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
572 /***************************************************
573 * Low level helper routines.....
574 ***************************************************/
579 test_m_maps_pv(vm_page_t m, pv_entry_t pv)
585 KKASSERT(pv->pv_m == m);
587 TAILQ_FOREACH(spv, &m->md.pv_list, pv_list) {
594 panic("test_m_maps_pv: failed m %p pv %p\n", m, pv);
598 ptbase_assert(struct pmap *pmap)
600 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
602 /* are we current address space or kernel? */
603 if (pmap == &kernel_pmap || frame == (((unsigned)PTDpde) & PG_FRAME)) {
606 KKASSERT(frame == (*mycpu->gd_GDMAP1 & PG_FRAME));
611 #define test_m_maps_pv(m, pv)
612 #define ptbase_assert(pmap)
616 #if defined(PMAP_DIAGNOSTIC)
619 * This code checks for non-writeable/modified pages.
620 * This should be an invalid condition.
623 pmap_nw_modified(pt_entry_t ptea)
629 if ((pte & (PG_M|PG_RW)) == PG_M)
638 * this routine defines the region(s) of memory that should
639 * not be tested for the modified bit.
641 static PMAP_INLINE int
642 pmap_track_modified(vm_offset_t va)
644 if ((va < clean_sva) || (va >= clean_eva))
651 * Retrieve the mapped page table base for a particular pmap. Use our self
652 * mapping for the kernel_pmap or our current pmap.
654 * For foreign pmaps we use the per-cpu page table map. Since this involves
655 * installing a ptd it's actually (per-process x per-cpu). However, we
656 * still cannot depend on our mapping to survive thread switches because
657 * the process might be threaded and switching to another thread for the
658 * same process on the same cpu will allow that other thread to make its
661 * This could be a bit confusing but the jist is for something like the
662 * vkernel which uses foreign pmaps all the time this represents a pretty
663 * good cache that avoids unnecessary invltlb()s.
666 get_ptbase(pmap_t pmap)
668 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
669 struct mdglobaldata *gd = mdcpu;
672 * We can use PTmap if the pmap is our current address space or
673 * the kernel address space.
675 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
676 return (unsigned *) PTmap;
680 * Otherwise we use the per-cpu alternative page table map. Each
681 * cpu gets its own map. Because of this we cannot use this map
682 * from interrupts or threads which can preempt.
684 * Even if we already have the map cached we may still have to
685 * invalidate the TLB if another cpu modified a PDE in the map.
687 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
688 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
690 if ((*gd->gd_GDMAP1 & PG_FRAME) != frame) {
691 *gd->gd_GDMAP1 = frame | PG_RW | PG_V;
692 pmap->pm_cached |= gd->mi.gd_cpumask;
694 } else if ((pmap->pm_cached & gd->mi.gd_cpumask) == 0) {
695 pmap->pm_cached |= gd->mi.gd_cpumask;
697 } else if (dreadful_invltlb) {
700 return ((unsigned *)gd->gd_GDADDR1);
706 * Extract the physical page address associated with the map/VA pair.
708 * This function may not be called from an interrupt if the pmap is
712 pmap_extract(pmap_t pmap, vm_offset_t va)
715 vm_offset_t pdirindex;
717 pdirindex = va >> PDRSHIFT;
718 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
720 if ((rtval & PG_PS) != 0) {
721 rtval &= ~(NBPDR - 1);
722 rtval |= va & (NBPDR - 1);
725 pte = get_ptbase(pmap) + i386_btop(va);
726 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
732 /***************************************************
733 * Low level mapping routines.....
734 ***************************************************/
737 * Routine: pmap_kenter
739 * Add a wired page to the KVA
740 * NOTE! note that in order for the mapping to take effect -- you
741 * should do an invltlb after doing the pmap_kenter().
744 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
748 pmap_inval_info info;
750 pmap_inval_init(&info);
751 npte = pa | PG_RW | PG_V | pgeflag;
752 pte = (unsigned *)vtopte(va);
753 pmap_inval_add(&info, &kernel_pmap, va);
755 pmap_inval_flush(&info);
759 * Routine: pmap_kenter_quick
761 * Similar to pmap_kenter(), except we only invalidate the
762 * mapping on the current CPU.
765 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
770 npte = pa | PG_RW | PG_V | pgeflag;
771 pte = (unsigned *)vtopte(va);
773 cpu_invlpg((void *)va);
777 pmap_kenter_sync(vm_offset_t va)
779 pmap_inval_info info;
781 pmap_inval_init(&info);
782 pmap_inval_add(&info, &kernel_pmap, va);
783 pmap_inval_flush(&info);
787 pmap_kenter_sync_quick(vm_offset_t va)
789 cpu_invlpg((void *)va);
793 * remove a page from the kernel pagetables
796 pmap_kremove(vm_offset_t va)
799 pmap_inval_info info;
801 pmap_inval_init(&info);
802 pte = (unsigned *)vtopte(va);
803 pmap_inval_add(&info, &kernel_pmap, va);
805 pmap_inval_flush(&info);
809 pmap_kremove_quick(vm_offset_t va)
812 pte = (unsigned *)vtopte(va);
814 cpu_invlpg((void *)va);
818 * XXX these need to be recoded. They are not used in any critical path.
821 pmap_kmodify_rw(vm_offset_t va)
823 *vtopte(va) |= PG_RW;
824 cpu_invlpg((void *)va);
828 pmap_kmodify_nc(vm_offset_t va)
831 cpu_invlpg((void *)va);
835 * Used to map a range of physical addresses into kernel
836 * virtual address space.
838 * For now, VM is already on, we only need to map the
842 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
844 vm_offset_t sva, virt;
847 while (start < end) {
848 pmap_kenter(virt, start);
858 * Add a list of wired pages to the kva
859 * this routine is only used for temporary
860 * kernel mappings that do not need to have
861 * page modification or references recorded.
862 * Note that old mappings are simply written
863 * over. The page *must* be wired.
866 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
870 end_va = va + count * PAGE_SIZE;
872 while (va < end_va) {
875 pte = (unsigned *)vtopte(va);
876 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
877 cpu_invlpg((void *)va);
882 smp_invltlb(); /* XXX */
887 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
890 cpumask_t cmask = mycpu->gd_cpumask;
892 end_va = va + count * PAGE_SIZE;
894 while (va < end_va) {
899 * Install the new PTE. If the pte changed from the prior
900 * mapping we must reset the cpu mask and invalidate the page.
901 * If the pte is the same but we have not seen it on the
902 * current cpu, invlpg the existing mapping. Otherwise the
903 * entry is optimal and no invalidation is required.
905 pte = (unsigned *)vtopte(va);
906 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
907 if (*pte != pteval) {
910 cpu_invlpg((void *)va);
911 } else if ((*mask & cmask) == 0) {
912 cpu_invlpg((void *)va);
921 * This routine jerks page mappings from the
922 * kernel -- it is meant only for temporary mappings.
924 * MPSAFE, INTERRUPT SAFE (cluster callback)
927 pmap_qremove(vm_offset_t va, int count)
931 end_va = va + count*PAGE_SIZE;
933 while (va < end_va) {
936 pte = (unsigned *)vtopte(va);
938 cpu_invlpg((void *)va);
947 * This routine works like vm_page_lookup() but also blocks as long as the
948 * page is busy. This routine does not busy the page it returns.
950 * Unless the caller is managing objects whos pages are in a known state,
951 * the call should be made with a critical section held so the page's object
952 * association remains valid on return.
955 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
960 m = vm_page_lookup(object, pindex);
961 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
967 * Create a new thread and optionally associate it with a (new) process.
968 * NOTE! the new thread's cpu may not equal the current cpu.
971 pmap_init_thread(thread_t td)
973 /* enforce pcb placement */
974 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
975 td->td_savefpu = &td->td_pcb->pcb_save;
976 td->td_sp = (char *)td->td_pcb - 16;
980 * This routine directly affects the fork perf for a process.
983 pmap_init_proc(struct proc *p)
988 * Dispose the UPAGES for a process that has exited.
989 * This routine directly impacts the exit perf of a process.
992 pmap_dispose_proc(struct proc *p)
994 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
997 /***************************************************
998 * Page table page management routines.....
999 ***************************************************/
1002 * This routine unholds page table pages, and if the hold count
1003 * drops to zero, then it decrements the wire count.
1006 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1009 * Wait until we can busy the page ourselves. We cannot have
1010 * any active flushes if we block.
1012 if (m->flags & PG_BUSY) {
1013 pmap_inval_flush(info);
1014 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1017 KASSERT(m->queue == PQ_NONE,
1018 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1020 if (m->hold_count == 1) {
1022 * Unmap the page table page.
1024 * NOTE: We must clear pm_cached for all cpus, including
1025 * the current one, when clearing a page directory
1029 pmap_inval_add(info, pmap, -1);
1030 KKASSERT(pmap->pm_pdir[m->pindex]);
1031 pmap->pm_pdir[m->pindex] = 0;
1032 pmap->pm_cached = 0;
1034 KKASSERT(pmap->pm_stats.resident_count > 0);
1035 --pmap->pm_stats.resident_count;
1037 if (pmap->pm_ptphint == m)
1038 pmap->pm_ptphint = NULL;
1041 * This was our last hold, the page had better be unwired
1042 * after we decrement wire_count.
1044 * FUTURE NOTE: shared page directory page could result in
1045 * multiple wire counts.
1049 KKASSERT(m->wire_count == 0);
1050 --vmstats.v_wire_count;
1051 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1053 vm_page_free_zero(m);
1056 KKASSERT(m->hold_count > 1);
1062 static PMAP_INLINE int
1063 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1065 KKASSERT(m->hold_count > 0);
1066 if (m->hold_count > 1) {
1070 return _pmap_unwire_pte_hold(pmap, m, info);
1075 * After removing a page table entry, this routine is used to
1076 * conditionally free the page, and manage the hold/wire counts.
1078 * WARNING: This function can block
1081 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1082 pmap_inval_info_t info)
1085 if (va >= UPT_MIN_ADDRESS)
1089 ptepindex = (va >> PDRSHIFT);
1090 if (pmap->pm_ptphint &&
1091 (pmap->pm_ptphint->pindex == ptepindex)) {
1092 mpte = pmap->pm_ptphint;
1094 pmap_inval_flush(info);
1095 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1096 pmap->pm_ptphint = mpte;
1100 return pmap_unwire_pte_hold(pmap, mpte, info);
1104 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1105 * it, and IdlePTD, represents the template used to update all other pmaps.
1107 * On architectures where the kernel pmap is not integrated into the user
1108 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1109 * kernel_pmap should be used to directly access the kernel_pmap.
1112 pmap_pinit0(struct pmap *pmap)
1115 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1116 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1118 pmap->pm_active = 0;
1119 pmap->pm_cached = 0;
1120 pmap->pm_ptphint = NULL;
1121 TAILQ_INIT(&pmap->pm_pvlist);
1122 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1126 * Initialize a preallocated and zeroed pmap structure,
1127 * such as one in a vmspace structure.
1130 pmap_pinit(struct pmap *pmap)
1135 * No need to allocate page table space yet but we do need a valid
1136 * page directory table.
1138 if (pmap->pm_pdir == NULL) {
1140 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1144 * Allocate an object for the ptes
1146 if (pmap->pm_pteobj == NULL)
1147 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1150 * Allocate the page directory page, unless we already have
1151 * one cached. If we used the cached page the wire_count will
1152 * already be set appropriately.
1154 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1155 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1156 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1157 pmap->pm_pdirm = ptdpg;
1158 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1159 ptdpg->valid = VM_PAGE_BITS_ALL;
1160 ptdpg->wire_count = 1;
1161 ++vmstats.v_wire_count;
1162 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1164 if ((ptdpg->flags & PG_ZERO) == 0)
1165 bzero(pmap->pm_pdir, PAGE_SIZE);
1168 pmap_page_assertzero(VM_PAGE_TO_PHYS(ptdpg));
1171 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1173 /* install self-referential address mapping entry */
1174 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1175 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1178 pmap->pm_active = 0;
1179 pmap->pm_cached = 0;
1180 pmap->pm_ptphint = NULL;
1181 TAILQ_INIT(&pmap->pm_pvlist);
1182 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1183 pmap->pm_stats.resident_count = 1;
1187 * Clean up a pmap structure so it can be physically freed. This routine
1188 * is called by the vmspace dtor function. A great deal of pmap data is
1189 * left passively mapped to improve vmspace management so we have a bit
1190 * of cleanup work to do here.
1193 pmap_puninit(pmap_t pmap)
1197 KKASSERT(pmap->pm_active == 0);
1198 if ((p = pmap->pm_pdirm) != NULL) {
1199 KKASSERT(pmap->pm_pdir != NULL);
1200 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1202 vmstats.v_wire_count--;
1203 KKASSERT((p->flags & PG_BUSY) == 0);
1205 vm_page_free_zero(p);
1206 pmap->pm_pdirm = NULL;
1208 if (pmap->pm_pdir) {
1209 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1210 pmap->pm_pdir = NULL;
1212 if (pmap->pm_pteobj) {
1213 vm_object_deallocate(pmap->pm_pteobj);
1214 pmap->pm_pteobj = NULL;
1219 * Wire in kernel global address entries. To avoid a race condition
1220 * between pmap initialization and pmap_growkernel, this procedure
1221 * adds the pmap to the master list (which growkernel scans to update),
1222 * then copies the template.
1225 pmap_pinit2(struct pmap *pmap)
1228 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1229 /* XXX copies current process, does not fill in MPPTDI */
1230 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1235 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1236 * 0 on failure (if the procedure had to sleep).
1238 * When asked to remove the page directory page itself, we actually just
1239 * leave it cached so we do not have to incur the SMP inval overhead of
1240 * removing the kernel mapping. pmap_puninit() will take care of it.
1243 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1245 unsigned *pde = (unsigned *) pmap->pm_pdir;
1247 * This code optimizes the case of freeing non-busy
1248 * page-table pages. Those pages are zero now, and
1249 * might as well be placed directly into the zero queue.
1251 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1257 * Remove the page table page from the processes address space.
1259 KKASSERT(pmap->pm_stats.resident_count > 0);
1260 KKASSERT(pde[p->pindex]);
1262 --pmap->pm_stats.resident_count;
1263 pmap->pm_cached = 0;
1265 if (p->hold_count) {
1266 panic("pmap_release: freeing held page table page");
1268 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1269 pmap->pm_ptphint = NULL;
1272 * We leave the page directory page cached, wired, and mapped in
1273 * the pmap until the dtor function (pmap_puninit()) gets called.
1274 * However, still clean it up so we can set PG_ZERO.
1276 * The pmap has already been removed from the pmap_list in the
1279 if (p->pindex == PTDPTDI) {
1280 bzero(pde + KPTDI, nkpt * PTESIZE);
1281 bzero(pde + KGDTDI, (NPDEPG - KGDTDI) * PTESIZE);
1282 vm_page_flag_set(p, PG_ZERO);
1286 vmstats.v_wire_count--;
1287 vm_page_free_zero(p);
1293 * this routine is called if the page table page is not
1297 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1299 vm_offset_t pteva, ptepa;
1303 * Find or fabricate a new pagetable page
1305 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1306 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1308 KASSERT(m->queue == PQ_NONE,
1309 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1312 * Increment the hold count for the page we will be returning to
1318 * It is possible that someone else got in and mapped by the page
1319 * directory page while we were blocked, if so just unbusy and
1320 * return the held page.
1322 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1323 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1328 if (m->wire_count == 0)
1329 vmstats.v_wire_count++;
1334 * Map the pagetable page into the process address space, if
1335 * it isn't already there.
1337 * NOTE: For safety clear pm_cached for all cpus including the
1338 * current one when adding a PDE to the map.
1340 ++pmap->pm_stats.resident_count;
1342 ptepa = VM_PAGE_TO_PHYS(m);
1343 pmap->pm_pdir[ptepindex] =
1344 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1345 pmap->pm_cached = 0;
1348 * Set the page table hint
1350 pmap->pm_ptphint = m;
1353 * Try to use the new mapping, but if we cannot, then
1354 * do it with the routine that maps the page explicitly.
1356 if ((m->flags & PG_ZERO) == 0) {
1357 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1358 (((unsigned) PTDpde) & PG_FRAME)) {
1359 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1360 bzero((caddr_t) pteva, PAGE_SIZE);
1362 pmap_zero_page(ptepa);
1367 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1371 m->valid = VM_PAGE_BITS_ALL;
1372 vm_page_flag_clear(m, PG_ZERO);
1373 vm_page_flag_set(m, PG_MAPPED);
1380 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1387 * Calculate pagetable page index
1389 ptepindex = va >> PDRSHIFT;
1392 * Get the page directory entry
1394 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1397 * This supports switching from a 4MB page to a
1400 if (ptepa & PG_PS) {
1401 pmap->pm_pdir[ptepindex] = 0;
1408 * If the page table page is mapped, we just increment the
1409 * hold count, and activate it.
1413 * In order to get the page table page, try the
1416 if (pmap->pm_ptphint &&
1417 (pmap->pm_ptphint->pindex == ptepindex)) {
1418 m = pmap->pm_ptphint;
1420 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1421 pmap->pm_ptphint = m;
1427 * Here if the pte page isn't mapped, or if it has been deallocated.
1429 return _pmap_allocpte(pmap, ptepindex);
1433 /***************************************************
1434 * Pmap allocation/deallocation routines.
1435 ***************************************************/
1438 * Release any resources held by the given physical map.
1439 * Called when a pmap initialized by pmap_pinit is being released.
1440 * Should only be called if the map contains no valid mappings.
1442 static int pmap_release_callback(struct vm_page *p, void *data);
1445 pmap_release(struct pmap *pmap)
1447 vm_object_t object = pmap->pm_pteobj;
1448 struct rb_vm_page_scan_info info;
1450 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1451 #if defined(DIAGNOSTIC)
1452 if (object->ref_count != 1)
1453 panic("pmap_release: pteobj reference count != 1");
1457 info.object = object;
1459 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1466 info.limit = object->generation;
1468 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1469 pmap_release_callback, &info);
1470 if (info.error == 0 && info.mpte) {
1471 if (!pmap_release_free_page(pmap, info.mpte))
1475 } while (info.error);
1476 pmap->pm_cached = 0;
1480 pmap_release_callback(struct vm_page *p, void *data)
1482 struct rb_vm_page_scan_info *info = data;
1484 if (p->pindex == PTDPTDI) {
1488 if (!pmap_release_free_page(info->pmap, p)) {
1492 if (info->object->generation != info->limit) {
1500 * Grow the number of kernel page table entries, if needed.
1504 pmap_growkernel(vm_offset_t addr)
1507 vm_offset_t ptppaddr;
1512 if (kernel_vm_end == 0) {
1513 kernel_vm_end = KERNBASE;
1515 while (pdir_pde(PTD, kernel_vm_end)) {
1516 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1520 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1521 while (kernel_vm_end < addr) {
1522 if (pdir_pde(PTD, kernel_vm_end)) {
1523 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1528 * This index is bogus, but out of the way
1530 nkpg = vm_page_alloc(kptobj, nkpt,
1531 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1533 panic("pmap_growkernel: no memory to grow kernel");
1536 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1537 pmap_zero_page(ptppaddr);
1538 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1539 pdir_pde(PTD, kernel_vm_end) = newpdir;
1540 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1544 * This update must be interlocked with pmap_pinit2.
1546 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1547 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1549 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1550 ~(PAGE_SIZE * NPTEPG - 1);
1556 * Retire the given physical map from service.
1557 * Should only be called if the map contains
1558 * no valid mappings.
1561 pmap_destroy(pmap_t pmap)
1568 count = --pmap->pm_count;
1571 panic("destroying a pmap is not yet implemented");
1576 * Add a reference to the specified pmap.
1579 pmap_reference(pmap_t pmap)
1586 /***************************************************
1587 * page management routines.
1588 ***************************************************/
1591 * free the pv_entry back to the free list. This function may be
1592 * called from an interrupt.
1594 static PMAP_INLINE void
1595 free_pv_entry(pv_entry_t pv)
1598 KKASSERT(pv->pv_m != NULL);
1606 * get a new pv_entry, allocating a block from the system
1607 * when needed. This function may be called from an interrupt.
1613 if (pv_entry_high_water &&
1614 (pv_entry_count > pv_entry_high_water) &&
1615 (pmap_pagedaemon_waken == 0)) {
1616 pmap_pagedaemon_waken = 1;
1617 wakeup (&vm_pages_needed);
1619 return zalloc(pvzone);
1623 * This routine is very drastic, but can save the system
1631 static int warningdone=0;
1633 if (pmap_pagedaemon_waken == 0)
1635 pmap_pagedaemon_waken = 0;
1637 if (warningdone < 5) {
1638 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1642 for(i = 0; i < vm_page_array_size; i++) {
1643 m = &vm_page_array[i];
1644 if (m->wire_count || m->hold_count || m->busy ||
1645 (m->flags & PG_BUSY))
1653 * If it is the first entry on the list, it is actually
1654 * in the header and we must copy the following entry up
1655 * to the header. Otherwise we must search the list for
1656 * the entry. In either case we free the now unused entry.
1659 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1660 vm_offset_t va, pmap_inval_info_t info)
1666 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1667 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1668 if (pmap == pv->pv_pmap && va == pv->pv_va)
1672 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1674 KKASSERT(pv->pv_pmap == pmap);
1676 if (va == pv->pv_va)
1683 test_m_maps_pv(m, pv);
1684 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1685 m->md.pv_list_count--;
1686 if (TAILQ_EMPTY(&m->md.pv_list))
1687 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1688 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1689 ++pmap->pm_generation;
1690 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1697 * Create a pv entry for page at pa for
1701 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1706 pv = get_pv_entry();
1708 KKASSERT(pv->pv_m == NULL);
1715 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1716 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1717 ++pmap->pm_generation;
1718 m->md.pv_list_count++;
1724 * pmap_remove_pte: do the things to unmap a page in a process.
1726 * WARNING: This function may block (via pmap_remove_entry/pmap_unuse_pt),
1727 * callers using temporary pmaps must reload them.
1730 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1731 pmap_inval_info_t info)
1736 ptbase_assert(pmap);
1737 pmap_inval_add(info, pmap, va);
1738 ptbase_assert(pmap);
1739 oldpte = loadandclear(ptq);
1742 pmap->pm_stats.wired_count -= 1;
1744 * Machines that don't support invlpg, also don't support
1745 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1749 cpu_invlpg((void *)va);
1750 KKASSERT(pmap->pm_stats.resident_count > 0);
1751 --pmap->pm_stats.resident_count;
1752 if (oldpte & PG_MANAGED) {
1753 m = PHYS_TO_VM_PAGE(oldpte);
1754 if (oldpte & PG_M) {
1755 #if defined(PMAP_DIAGNOSTIC)
1756 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1757 kprintf("pmap_remove: modified page not "
1758 "writable: va: %p, pte: 0x%lx\n",
1759 (void *)va, (long)oldpte);
1762 if (pmap_track_modified(va))
1766 vm_page_flag_set(m, PG_REFERENCED);
1767 return pmap_remove_entry(pmap, m, va, info);
1769 return pmap_unuse_pt(pmap, va, NULL, info);
1778 * Remove a single page from a process address space.
1780 * This function may not be called from an interrupt if the pmap is
1784 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1789 * if there is no pte for this address, just skip it!!! Otherwise
1790 * get a local va for mappings for this pmap and remove the entry.
1792 if (*pmap_pde(pmap, va) != 0) {
1793 ptq = get_ptbase(pmap) + i386_btop(va);
1795 pmap_remove_pte(pmap, ptq, va, info);
1804 * Remove the given range of addresses from the specified map.
1806 * It is assumed that the start and end are properly
1807 * rounded to the page size.
1809 * This function may not be called from an interrupt if the pmap is
1813 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1817 vm_offset_t ptpaddr;
1818 vm_offset_t sindex, eindex;
1819 struct pmap_inval_info info;
1824 if (pmap->pm_stats.resident_count == 0)
1827 pmap_inval_init(&info);
1830 * special handling of removing one page. a very
1831 * common operation and easy to short circuit some
1834 if (((sva + PAGE_SIZE) == eva) &&
1835 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1836 pmap_remove_page(pmap, sva, &info);
1837 pmap_inval_flush(&info);
1842 * Get a local virtual address for the mappings that are being
1845 sindex = i386_btop(sva);
1846 eindex = i386_btop(eva);
1848 for (; sindex < eindex; sindex = pdnxt) {
1852 * Calculate index for next page table.
1854 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1855 if (pmap->pm_stats.resident_count == 0)
1858 pdirindex = sindex / NPDEPG;
1859 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1860 pmap_inval_add(&info, pmap, -1);
1861 pmap->pm_pdir[pdirindex] = 0;
1862 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1863 pmap->pm_cached = 0;
1868 * Weed out invalid mappings. Note: we assume that the page
1869 * directory table is always allocated, and in kernel virtual.
1875 * Limit our scan to either the end of the va represented
1876 * by the current page table page, or to the end of the
1877 * range being removed.
1879 if (pdnxt > eindex) {
1884 * NOTE: pmap_remove_pte() can block and wipe the temporary
1887 for (; sindex != pdnxt; sindex++) {
1890 ptbase = get_ptbase(pmap);
1891 if (ptbase[sindex] == 0)
1893 va = i386_ptob(sindex);
1894 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1898 pmap_inval_flush(&info);
1904 * Removes this physical page from all physical maps in which it resides.
1905 * Reflects back modify bits to the pager.
1907 * This routine may not be called from an interrupt.
1911 pmap_remove_all(vm_page_t m)
1913 struct pmap_inval_info info;
1914 unsigned *pte, tpte;
1917 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
1920 pmap_inval_init(&info);
1922 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1923 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1924 --pv->pv_pmap->pm_stats.resident_count;
1926 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1927 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1928 tpte = loadandclear(pte);
1930 KKASSERT(PHYS_TO_VM_PAGE(tpte) == m);
1933 pv->pv_pmap->pm_stats.wired_count--;
1936 vm_page_flag_set(m, PG_REFERENCED);
1939 * Update the vm_page_t clean and reference bits.
1942 #if defined(PMAP_DIAGNOSTIC)
1943 if (pmap_nw_modified((pt_entry_t) tpte)) {
1944 kprintf("pmap_remove_all: modified page "
1945 "not writable: va: %p, pte: 0x%lx\n",
1946 (void *)pv->pv_va, (long)tpte);
1949 if (pmap_track_modified(pv->pv_va))
1953 KKASSERT(pv->pv_m == m);
1955 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1956 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1957 ++pv->pv_pmap->pm_generation;
1958 m->md.pv_list_count--;
1959 if (TAILQ_EMPTY(&m->md.pv_list))
1960 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1961 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1965 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
1966 pmap_inval_flush(&info);
1972 * Set the physical protection on the specified range of this map
1975 * This function may not be called from an interrupt if the map is
1976 * not the kernel_pmap.
1979 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1982 vm_offset_t pdnxt, ptpaddr;
1983 vm_pindex_t sindex, eindex;
1984 pmap_inval_info info;
1989 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1990 pmap_remove(pmap, sva, eva);
1994 if (prot & VM_PROT_WRITE)
1997 pmap_inval_init(&info);
1999 ptbase = get_ptbase(pmap);
2001 sindex = i386_btop(sva);
2002 eindex = i386_btop(eva);
2004 for (; sindex < eindex; sindex = pdnxt) {
2008 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
2010 pdirindex = sindex / NPDEPG;
2011 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
2012 pmap_inval_add(&info, pmap, -1);
2013 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
2014 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2019 * Weed out invalid mappings. Note: we assume that the page
2020 * directory table is always allocated, and in kernel virtual.
2025 if (pdnxt > eindex) {
2029 for (; sindex != pdnxt; sindex++) {
2035 * XXX non-optimal. Note also that there can be
2036 * no pmap_inval_flush() calls until after we modify
2037 * ptbase[sindex] (or otherwise we have to do another
2038 * pmap_inval_add() call).
2040 pmap_inval_add(&info, pmap, i386_ptob(sindex));
2041 pbits = ptbase[sindex];
2043 if (pbits & PG_MANAGED) {
2046 m = PHYS_TO_VM_PAGE(pbits);
2047 vm_page_flag_set(m, PG_REFERENCED);
2051 if (pmap_track_modified(i386_ptob(sindex))) {
2053 m = PHYS_TO_VM_PAGE(pbits);
2062 if (pbits != ptbase[sindex]) {
2063 ptbase[sindex] = pbits;
2067 pmap_inval_flush(&info);
2071 * Insert the given physical page (p) at
2072 * the specified virtual address (v) in the
2073 * target physical map with the protection requested.
2075 * If specified, the page will be wired down, meaning
2076 * that the related pte can not be reclaimed.
2078 * NB: This is the only routine which MAY NOT lazy-evaluate
2079 * or lose information. That is, this routine must actually
2080 * insert this page into the given map NOW.
2083 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2089 vm_offset_t origpte, newpte;
2091 pmap_inval_info info;
2097 #ifdef PMAP_DIAGNOSTIC
2099 panic("pmap_enter: toobig");
2100 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) {
2101 panic("pmap_enter: invalid to pmap_enter page "
2102 "table pages (va: %p)", (void *)va);
2105 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2106 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2109 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2110 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2115 * In the case that a page table page is not
2116 * resident, we are creating it here.
2118 if (va < UPT_MIN_ADDRESS)
2119 mpte = pmap_allocpte(pmap, va);
2123 pmap_inval_init(&info);
2124 pte = pmap_pte(pmap, va);
2127 * Page Directory table entry not valid, we need a new PT page
2130 panic("pmap_enter: invalid page directory pdir=0x%lx, va=%p\n",
2131 (long)pmap->pm_pdir[PTDPTDI], (void *)va);
2134 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
2135 origpte = *(vm_offset_t *)pte;
2136 opa = origpte & PG_FRAME;
2138 if (origpte & PG_PS)
2139 panic("pmap_enter: attempted pmap_enter on 4MB page");
2142 * Mapping has not changed, must be protection or wiring change.
2144 if (origpte && (opa == pa)) {
2146 * Wiring change, just update stats. We don't worry about
2147 * wiring PT pages as they remain resident as long as there
2148 * are valid mappings in them. Hence, if a user page is wired,
2149 * the PT page will be also.
2151 if (wired && ((origpte & PG_W) == 0))
2152 pmap->pm_stats.wired_count++;
2153 else if (!wired && (origpte & PG_W))
2154 pmap->pm_stats.wired_count--;
2156 #if defined(PMAP_DIAGNOSTIC)
2157 if (pmap_nw_modified((pt_entry_t) origpte)) {
2158 kprintf("pmap_enter: modified page not "
2159 "writable: va: %p, pte: 0x%lx\n",
2160 (void *)va, (long )origpte);
2165 * Remove the extra pte reference. Note that we cannot
2166 * optimize the RO->RW case because we have adjusted the
2167 * wiring count above and may need to adjust the wiring
2174 * We might be turning off write access to the page,
2175 * so we go ahead and sense modify status.
2177 if (origpte & PG_MANAGED) {
2178 if ((origpte & PG_M) && pmap_track_modified(va)) {
2180 om = PHYS_TO_VM_PAGE(opa);
2184 KKASSERT(m->flags & PG_MAPPED);
2189 * Mapping has changed, invalidate old range and fall through to
2190 * handle validating new mapping.
2192 * Since we have a ref on the page directory page pmap_pte()
2193 * will always return non-NULL.
2195 * NOTE: pmap_remove_pte() can block and cause the temporary ptbase
2196 * to get wiped. reload the ptbase. I'm not sure if it is
2197 * also possible to race another pmap_enter() but check for
2203 KKASSERT((origpte & PG_FRAME) ==
2204 (*(vm_offset_t *)pte & PG_FRAME));
2205 err = pmap_remove_pte(pmap, pte, va, &info);
2207 panic("pmap_enter: pte vanished, va: %p", (void *)va);
2208 pte = pmap_pte(pmap, va);
2209 origpte = *(vm_offset_t *)pte;
2210 opa = origpte & PG_FRAME;
2212 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2218 * Enter on the PV list if part of our managed memory. Note that we
2219 * raise IPL while manipulating pv_table since pmap_enter can be
2220 * called at interrupt time.
2222 if (pmap_initialized &&
2223 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2224 pmap_insert_entry(pmap, va, mpte, m);
2225 ptbase_assert(pmap);
2227 vm_page_flag_set(m, PG_MAPPED);
2231 * Increment counters
2233 ++pmap->pm_stats.resident_count;
2235 pmap->pm_stats.wired_count++;
2236 KKASSERT(*pte == 0);
2240 * Now validate mapping with desired protection/wiring.
2242 ptbase_assert(pmap);
2243 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2247 if (va < UPT_MIN_ADDRESS)
2249 if (pmap == &kernel_pmap)
2253 * if the mapping or permission bits are different, we need
2254 * to update the pte.
2256 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2257 pmap_inval_add(&info, pmap, va);
2258 ptbase_assert(pmap);
2259 KKASSERT(*pte == 0 ||
2260 (*pte & PG_FRAME) == (newpte & PG_FRAME));
2261 *pte = newpte | PG_A;
2263 vm_page_flag_set(m, PG_WRITEABLE);
2265 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2266 pmap_inval_flush(&info);
2270 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2271 * This code also assumes that the pmap has no pre-existing entry for this
2274 * This code currently may only be used on user pmaps, not kernel_pmap.
2277 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2284 pmap_inval_info info;
2286 pmap_inval_init(&info);
2288 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2289 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2292 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2293 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2297 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2300 * Calculate the page table page (mpte), allocating it if necessary.
2302 * A held page table page (mpte), or NULL, is passed onto the
2303 * section following.
2305 if (va < UPT_MIN_ADDRESS) {
2307 * Calculate pagetable page index
2309 ptepindex = va >> PDRSHIFT;
2313 * Get the page directory entry
2315 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2318 * If the page table page is mapped, we just increment
2319 * the hold count, and activate it.
2323 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2324 if (pmap->pm_ptphint &&
2325 (pmap->pm_ptphint->pindex == ptepindex)) {
2326 mpte = pmap->pm_ptphint;
2328 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2329 pmap->pm_ptphint = mpte;
2334 mpte = _pmap_allocpte(pmap, ptepindex);
2336 } while (mpte == NULL);
2339 /* this code path is not yet used */
2343 * With a valid (and held) page directory page, we can just use
2344 * vtopte() to get to the pte. If the pte is already present
2345 * we do not disturb it.
2347 pte = (unsigned *)vtopte(va);
2350 pmap_unwire_pte_hold(pmap, mpte, &info);
2351 pa = VM_PAGE_TO_PHYS(m);
2352 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2357 * Enter on the PV list if part of our managed memory
2359 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2360 pmap_insert_entry(pmap, va, mpte, m);
2361 vm_page_flag_set(m, PG_MAPPED);
2365 * Increment counters
2367 ++pmap->pm_stats.resident_count;
2369 pa = VM_PAGE_TO_PHYS(m);
2372 * Now validate mapping with RO protection
2374 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2375 *pte = pa | PG_V | PG_U;
2377 *pte = pa | PG_V | PG_U | PG_MANAGED;
2378 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2379 pmap_inval_flush(&info);
2383 * Make a temporary mapping for a physical address. This is only intended
2384 * to be used for panic dumps.
2387 pmap_kenter_temporary(vm_paddr_t pa, int i)
2389 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2390 return ((void *)crashdumpmap);
2393 #define MAX_INIT_PT (96)
2396 * This routine preloads the ptes for a given object into the specified pmap.
2397 * This eliminates the blast of soft faults on process startup and
2398 * immediately after an mmap.
2400 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2403 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2404 vm_object_t object, vm_pindex_t pindex,
2405 vm_size_t size, int limit)
2407 struct rb_vm_page_scan_info info;
2412 * We can't preinit if read access isn't set or there is no pmap
2415 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2419 * We can't preinit if the pmap is not the current pmap
2421 lp = curthread->td_lwp;
2422 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2425 psize = i386_btop(size);
2427 if ((object->type != OBJT_VNODE) ||
2428 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2429 (object->resident_page_count > MAX_INIT_PT))) {
2433 if (psize + pindex > object->size) {
2434 if (object->size < pindex)
2436 psize = object->size - pindex;
2443 * Use a red-black scan to traverse the requested range and load
2444 * any valid pages found into the pmap.
2446 * We cannot safely scan the object's memq unless we are in a
2447 * critical section since interrupts can remove pages from objects.
2449 info.start_pindex = pindex;
2450 info.end_pindex = pindex + psize - 1;
2457 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2458 pmap_object_init_pt_callback, &info);
2464 pmap_object_init_pt_callback(vm_page_t p, void *data)
2466 struct rb_vm_page_scan_info *info = data;
2467 vm_pindex_t rel_index;
2469 * don't allow an madvise to blow away our really
2470 * free pages allocating pv entries.
2472 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2473 vmstats.v_free_count < vmstats.v_free_reserved) {
2476 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2477 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2478 if ((p->queue - p->pc) == PQ_CACHE)
2479 vm_page_deactivate(p);
2481 rel_index = p->pindex - info->start_pindex;
2482 pmap_enter_quick(info->pmap,
2483 info->addr + i386_ptob(rel_index), p);
2490 * Return TRUE if the pmap is in shape to trivially
2491 * pre-fault the specified address.
2493 * Returns FALSE if it would be non-trivial or if a
2494 * pte is already loaded into the slot.
2497 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2501 if ((*pmap_pde(pmap, addr)) == 0)
2503 pte = (unsigned *) vtopte(addr);
2510 * Routine: pmap_change_wiring
2511 * Function: Change the wiring attribute for a map/virtual-address
2513 * In/out conditions:
2514 * The mapping must already exist in the pmap.
2517 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2524 pte = pmap_pte(pmap, va);
2526 if (wired && !pmap_pte_w(pte))
2527 pmap->pm_stats.wired_count++;
2528 else if (!wired && pmap_pte_w(pte))
2529 pmap->pm_stats.wired_count--;
2532 * Wiring is not a hardware characteristic so there is no need to
2533 * invalidate TLB. However, in an SMP environment we must use
2534 * a locked bus cycle to update the pte (if we are not using
2535 * the pmap_inval_*() API that is)... it's ok to do this for simple
2540 atomic_set_int(pte, PG_W);
2542 atomic_clear_int(pte, PG_W);
2545 atomic_set_int_nonlocked(pte, PG_W);
2547 atomic_clear_int_nonlocked(pte, PG_W);
2554 * Copy the range specified by src_addr/len
2555 * from the source map to the range dst_addr/len
2556 * in the destination map.
2558 * This routine is only advisory and need not do anything.
2561 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2562 vm_size_t len, vm_offset_t src_addr)
2565 pmap_inval_info info;
2567 vm_offset_t end_addr = src_addr + len;
2569 unsigned src_frame, dst_frame;
2573 if (dst_addr != src_addr)
2576 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2577 * valid through blocking calls, and that's just not going to
2584 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2585 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2589 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2590 if ((*gd->gd_GDMAP1 & PG_FRAME) != dst_frame) {
2591 *gd->gd_GDMAP1 = dst_frame | PG_RW | PG_V;
2595 pmap_inval_init(&info);
2596 pmap_inval_add(&info, dst_pmap, -1);
2597 pmap_inval_add(&info, src_pmap, -1);
2600 * critical section protection is required to maintain the page/object
2601 * association, interrupts can free pages and remove them from
2605 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2606 unsigned *src_pte, *dst_pte;
2607 vm_page_t dstmpte, srcmpte;
2608 vm_offset_t srcptepaddr;
2611 if (addr >= UPT_MIN_ADDRESS)
2612 panic("pmap_copy: invalid to pmap_copy page tables\n");
2615 * Don't let optional prefaulting of pages make us go
2616 * way below the low water mark of free pages or way
2617 * above high water mark of used pv entries.
2619 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2620 pv_entry_count > pv_entry_high_water)
2623 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2624 ptepindex = addr >> PDRSHIFT;
2626 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2627 if (srcptepaddr == 0)
2630 if (srcptepaddr & PG_PS) {
2631 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2632 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2633 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2639 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2640 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2641 (srcmpte->flags & PG_BUSY)) {
2645 if (pdnxt > end_addr)
2648 src_pte = (unsigned *) vtopte(addr);
2649 dst_pte = (unsigned *) avtopte(addr);
2650 while (addr < pdnxt) {
2655 * we only virtual copy managed pages
2657 if ((ptetemp & PG_MANAGED) != 0) {
2659 * We have to check after allocpte for the
2660 * pte still being around... allocpte can
2663 * pmap_allocpte() can block. If we lose
2664 * our page directory mappings we stop.
2666 dstmpte = pmap_allocpte(dst_pmap, addr);
2668 if (src_frame != (((unsigned) PTDpde) & PG_FRAME) ||
2669 XXX dst_frame != (((unsigned) xxx) & PG_FRAME)
2671 kprintf("WARNING: pmap_copy: detected and corrected race\n");
2672 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2674 } else if ((*dst_pte == 0) &&
2675 (ptetemp = *src_pte) != 0 &&
2676 (ptetemp & PG_MANAGED)) {
2678 * Clear the modified and
2679 * accessed (referenced) bits
2682 m = PHYS_TO_VM_PAGE(ptetemp);
2683 *dst_pte = ptetemp & ~(PG_M | PG_A);
2684 ++dst_pmap->pm_stats.resident_count;
2685 pmap_insert_entry(dst_pmap, addr,
2687 KKASSERT(m->flags & PG_MAPPED);
2689 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
2690 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2693 if (dstmpte->hold_count >= srcmpte->hold_count)
2703 pmap_inval_flush(&info);
2710 * Zero the specified PA by mapping the page into KVM and clearing its
2713 * This function may be called from an interrupt and no locking is
2717 pmap_zero_page(vm_paddr_t phys)
2719 struct mdglobaldata *gd = mdcpu;
2722 if (*(int *)gd->gd_CMAP3)
2723 panic("pmap_zero_page: CMAP3 busy");
2724 *(int *)gd->gd_CMAP3 =
2725 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2726 cpu_invlpg(gd->gd_CADDR3);
2728 #if defined(I686_CPU)
2729 if (cpu_class == CPUCLASS_686)
2730 i686_pagezero(gd->gd_CADDR3);
2733 bzero(gd->gd_CADDR3, PAGE_SIZE);
2734 *(int *) gd->gd_CMAP3 = 0;
2739 * pmap_page_assertzero:
2741 * Assert that a page is empty, panic if it isn't.
2744 pmap_page_assertzero(vm_paddr_t phys)
2746 struct mdglobaldata *gd = mdcpu;
2750 if (*(int *)gd->gd_CMAP3)
2751 panic("pmap_zero_page: CMAP3 busy");
2752 *(int *)gd->gd_CMAP3 =
2753 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2754 cpu_invlpg(gd->gd_CADDR3);
2755 for (i = 0; i < PAGE_SIZE; i += 4) {
2756 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2757 panic("pmap_page_assertzero() @ %p not zero!\n",
2758 (void *)gd->gd_CADDR3);
2761 *(int *) gd->gd_CMAP3 = 0;
2768 * Zero part of a physical page by mapping it into memory and clearing
2769 * its contents with bzero.
2771 * off and size may not cover an area beyond a single hardware page.
2774 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2776 struct mdglobaldata *gd = mdcpu;
2779 if (*(int *) gd->gd_CMAP3)
2780 panic("pmap_zero_page: CMAP3 busy");
2781 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2782 cpu_invlpg(gd->gd_CADDR3);
2784 #if defined(I686_CPU)
2785 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2786 i686_pagezero(gd->gd_CADDR3);
2789 bzero((char *)gd->gd_CADDR3 + off, size);
2790 *(int *) gd->gd_CMAP3 = 0;
2797 * Copy the physical page from the source PA to the target PA.
2798 * This function may be called from an interrupt. No locking
2802 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2804 struct mdglobaldata *gd = mdcpu;
2807 if (*(int *) gd->gd_CMAP1)
2808 panic("pmap_copy_page: CMAP1 busy");
2809 if (*(int *) gd->gd_CMAP2)
2810 panic("pmap_copy_page: CMAP2 busy");
2812 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2813 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2815 cpu_invlpg(gd->gd_CADDR1);
2816 cpu_invlpg(gd->gd_CADDR2);
2818 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2820 *(int *) gd->gd_CMAP1 = 0;
2821 *(int *) gd->gd_CMAP2 = 0;
2826 * pmap_copy_page_frag:
2828 * Copy the physical page from the source PA to the target PA.
2829 * This function may be called from an interrupt. No locking
2833 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2835 struct mdglobaldata *gd = mdcpu;
2838 if (*(int *) gd->gd_CMAP1)
2839 panic("pmap_copy_page: CMAP1 busy");
2840 if (*(int *) gd->gd_CMAP2)
2841 panic("pmap_copy_page: CMAP2 busy");
2843 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2844 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2846 cpu_invlpg(gd->gd_CADDR1);
2847 cpu_invlpg(gd->gd_CADDR2);
2849 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2850 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2853 *(int *) gd->gd_CMAP1 = 0;
2854 *(int *) gd->gd_CMAP2 = 0;
2859 * Returns true if the pmap's pv is one of the first
2860 * 16 pvs linked to from this page. This count may
2861 * be changed upwards or downwards in the future; it
2862 * is only necessary that true be returned for a small
2863 * subset of pmaps for proper page aging.
2866 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2871 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2876 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2877 if (pv->pv_pmap == pmap) {
2890 * Remove all pages from specified address space
2891 * this aids process exit speeds. Also, this code
2892 * is special cased for current process only, but
2893 * can have the more generic (and slightly slower)
2894 * mode enabled. This is much faster than pmap_remove
2895 * in the case of running down an entire address space.
2898 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2901 unsigned *pte, tpte;
2904 pmap_inval_info info;
2906 int32_t save_generation;
2908 lp = curthread->td_lwp;
2909 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2914 pmap_inval_init(&info);
2916 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2917 if (pv->pv_va >= eva || pv->pv_va < sva) {
2918 npv = TAILQ_NEXT(pv, pv_plist);
2922 KKASSERT(pmap == pv->pv_pmap);
2925 pte = (unsigned *)vtopte(pv->pv_va);
2927 pte = pmap_pte_quick(pmap, pv->pv_va);
2929 if (pmap->pm_active)
2930 pmap_inval_add(&info, pmap, pv->pv_va);
2933 * We cannot remove wired pages from a process' mapping
2937 npv = TAILQ_NEXT(pv, pv_plist);
2941 tpte = loadandclear(pte);
2943 m = PHYS_TO_VM_PAGE(tpte);
2944 test_m_maps_pv(m, pv);
2946 KASSERT(m < &vm_page_array[vm_page_array_size],
2947 ("pmap_remove_pages: bad tpte %x", tpte));
2949 KKASSERT(pmap->pm_stats.resident_count > 0);
2950 --pmap->pm_stats.resident_count;
2953 * Update the vm_page_t clean and reference bits.
2959 npv = TAILQ_NEXT(pv, pv_plist);
2961 KKASSERT(pv->pv_m == m);
2962 KKASSERT(pv->pv_pmap == pmap);
2964 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2965 save_generation = ++pmap->pm_generation;
2967 m->md.pv_list_count--;
2968 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2969 if (TAILQ_EMPTY(&m->md.pv_list))
2970 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2972 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2976 * Restart the scan if we blocked during the unuse or free
2977 * calls and other removals were made.
2979 if (save_generation != pmap->pm_generation) {
2980 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2981 npv = TAILQ_FIRST(&pmap->pm_pvlist);
2984 pmap_inval_flush(&info);
2989 * pmap_testbit tests bits in pte's
2990 * note that the testbit/clearbit routines are inline,
2991 * and a lot of things compile-time evaluate.
2994 pmap_testbit(vm_page_t m, int bit)
2999 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3002 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3007 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3009 * if the bit being tested is the modified bit, then
3010 * mark clean_map and ptes as never
3013 if (bit & (PG_A|PG_M)) {
3014 if (!pmap_track_modified(pv->pv_va))
3018 #if defined(PMAP_DIAGNOSTIC)
3020 kprintf("Null pmap (tb) at va: %p\n",
3025 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3036 * this routine is used to modify bits in ptes
3038 static __inline void
3039 pmap_clearbit(vm_page_t m, int bit)
3041 struct pmap_inval_info info;
3046 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3049 pmap_inval_init(&info);
3053 * Loop over all current mappings setting/clearing as appropos If
3054 * setting RO do we need to clear the VAC?
3056 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3058 * don't write protect pager mappings
3061 if (!pmap_track_modified(pv->pv_va))
3065 #if defined(PMAP_DIAGNOSTIC)
3067 kprintf("Null pmap (cb) at va: %p\n",
3074 * Careful here. We can use a locked bus instruction to
3075 * clear PG_A or PG_M safely but we need to synchronize
3076 * with the target cpus when we mess with PG_RW.
3078 * We do not have to force synchronization when clearing
3079 * PG_M even for PTEs generated via virtual memory maps,
3080 * because the virtual kernel will invalidate the pmap
3081 * entry when/if it needs to resynchronize the Modify bit.
3084 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
3085 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3092 atomic_clear_int(pte, PG_M|PG_RW);
3095 * The cpu may be trying to set PG_M
3096 * simultaniously with our clearing
3099 if (!atomic_cmpset_int(pte, pbits,
3103 } else if (bit == PG_M) {
3105 * We could also clear PG_RW here to force
3106 * a fault on write to redetect PG_M for
3107 * virtual kernels, but it isn't necessary
3108 * since virtual kernels invalidate the pte
3109 * when they clear the VPTE_M bit in their
3110 * virtual page tables.
3112 atomic_clear_int(pte, PG_M);
3114 atomic_clear_int(pte, bit);
3118 pmap_inval_flush(&info);
3123 * pmap_page_protect:
3125 * Lower the permission for all mappings to a given page.
3128 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3130 if ((prot & VM_PROT_WRITE) == 0) {
3131 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3132 pmap_clearbit(m, PG_RW);
3133 vm_page_flag_clear(m, PG_WRITEABLE);
3141 pmap_phys_address(vm_pindex_t ppn)
3143 return (i386_ptob(ppn));
3147 * pmap_ts_referenced:
3149 * Return a count of reference bits for a page, clearing those bits.
3150 * It is not necessary for every reference bit to be cleared, but it
3151 * is necessary that 0 only be returned when there are truly no
3152 * reference bits set.
3154 * XXX: The exact number of bits to check and clear is a matter that
3155 * should be tested and standardized at some point in the future for
3156 * optimal aging of shared pages.
3159 pmap_ts_referenced(vm_page_t m)
3161 pv_entry_t pv, pvf, pvn;
3165 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3170 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3175 pvn = TAILQ_NEXT(pv, pv_list);
3178 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3179 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3182 if (!pmap_track_modified(pv->pv_va))
3185 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3187 if (pte && (*pte & PG_A)) {
3189 atomic_clear_int(pte, PG_A);
3191 atomic_clear_int_nonlocked(pte, PG_A);
3198 } while ((pv = pvn) != NULL && pv != pvf);
3208 * Return whether or not the specified physical page was modified
3209 * in any physical maps.
3212 pmap_is_modified(vm_page_t m)
3214 return pmap_testbit(m, PG_M);
3218 * Clear the modify bits on the specified physical page.
3221 pmap_clear_modify(vm_page_t m)
3223 pmap_clearbit(m, PG_M);
3227 * pmap_clear_reference:
3229 * Clear the reference bit on the specified physical page.
3232 pmap_clear_reference(vm_page_t m)
3234 pmap_clearbit(m, PG_A);
3238 * Miscellaneous support routines follow
3242 i386_protection_init(void)
3246 kp = protection_codes;
3247 for (prot = 0; prot < 8; prot++) {
3249 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3251 * Read access is also 0. There isn't any execute bit,
3252 * so just make it readable.
3254 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3255 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3256 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3259 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3260 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3261 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3262 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3270 * Map a set of physical memory pages into the kernel virtual
3271 * address space. Return a pointer to where it is mapped. This
3272 * routine is intended to be used for mapping device memory,
3275 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3279 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3281 vm_offset_t va, tmpva, offset;
3284 offset = pa & PAGE_MASK;
3285 size = roundup(offset + size, PAGE_SIZE);
3287 va = kmem_alloc_nofault(&kernel_map, size);
3289 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3292 for (tmpva = va; size > 0;) {
3293 pte = (unsigned *)vtopte(tmpva);
3294 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3302 return ((void *)(va + offset));
3306 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3308 vm_offset_t base, offset;
3310 base = va & PG_FRAME;
3311 offset = va & PAGE_MASK;
3312 size = roundup(offset + size, PAGE_SIZE);
3313 pmap_qremove(va, size >> PAGE_SHIFT);
3314 kmem_free(&kernel_map, base, size);
3318 * perform the pmap work for mincore
3321 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3323 unsigned *ptep, pte;
3327 ptep = pmap_pte(pmap, addr);
3332 if ((pte = *ptep) != 0) {
3335 val = MINCORE_INCORE;
3336 if ((pte & PG_MANAGED) == 0)
3339 pa = pte & PG_FRAME;
3341 m = PHYS_TO_VM_PAGE(pa);
3347 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3349 * Modified by someone
3351 else if (m->dirty || pmap_is_modified(m))
3352 val |= MINCORE_MODIFIED_OTHER;
3357 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3360 * Referenced by someone
3362 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3363 val |= MINCORE_REFERENCED_OTHER;
3364 vm_page_flag_set(m, PG_REFERENCED);
3371 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3372 * vmspace will be ref'd and the old one will be deref'd.
3374 * The vmspace for all lwps associated with the process will be adjusted
3375 * and cr3 will be reloaded if any lwp is the current lwp.
3378 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3380 struct vmspace *oldvm;
3384 oldvm = p->p_vmspace;
3385 if (oldvm != newvm) {
3386 p->p_vmspace = newvm;
3387 KKASSERT(p->p_nthreads == 1);
3388 lp = RB_ROOT(&p->p_lwp_tree);
3389 pmap_setlwpvm(lp, newvm);
3391 sysref_get(&newvm->vm_sysref);
3392 sysref_put(&oldvm->vm_sysref);
3399 * Set the vmspace for a LWP. The vmspace is almost universally set the
3400 * same as the process vmspace, but virtual kernels need to swap out contexts
3401 * on a per-lwp basis.
3404 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3406 struct vmspace *oldvm;
3410 oldvm = lp->lwp_vmspace;
3412 if (oldvm != newvm) {
3413 lp->lwp_vmspace = newvm;
3414 if (curthread->td_lwp == lp) {
3415 pmap = vmspace_pmap(newvm);
3417 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3419 pmap->pm_active |= 1;
3421 #if defined(SWTCH_OPTIM_STATS)
3424 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3425 load_cr3(curthread->td_pcb->pcb_cr3);
3426 pmap = vmspace_pmap(oldvm);
3428 atomic_clear_int(&pmap->pm_active,
3429 1 << mycpu->gd_cpuid);
3431 pmap->pm_active &= ~1;
3439 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3442 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3446 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3451 pmap_get_pgeflag(void)
3458 static void pads (pmap_t pm);
3459 void pmap_pvdump (vm_paddr_t pa);
3461 /* print address space of pmap*/
3468 if (pm == &kernel_pmap)
3471 for (i = 0; i < 1024; i++) {
3472 if (pm->pm_pdir[i]) {
3473 for (j = 0; j < 1024; j++) {
3474 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3475 if (pm == &kernel_pmap && va < KERNBASE)
3477 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3479 ptep = pmap_pte_quick(pm, va);
3480 if (pmap_pte_v(ptep))
3481 kprintf("%x:%x ", va, *(int *) ptep);
3490 pmap_pvdump(vm_paddr_t pa)
3495 kprintf("pa %08llx", (long long)pa);
3496 m = PHYS_TO_VM_PAGE(pa);
3497 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3499 kprintf(" -> pmap %p, va %p, flags %x",
3500 (void *)pv->pv_pmap, (long)pv->pv_va, pv->pv_flags);
3502 kprintf(" -> pmap %p, va %p",
3503 (void *)pv->pv_pmap, (void *)pv->pv_va);