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 KvaStart; /* VA start of KVA space */
159 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
160 vm_offset_t KvaSize; /* max size of kernel virtual address space */
161 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
162 static int pgeflag; /* PG_G or-in */
163 static int pseflag; /* PG_PS or-in */
165 static vm_object_t kptobj;
168 vm_offset_t kernel_vm_end;
171 * Data for the pv entry allocation mechanism
173 static vm_zone_t pvzone;
174 static struct vm_zone pvzone_store;
175 static struct vm_object pvzone_obj;
176 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
177 static int pmap_pagedaemon_waken = 0;
178 static struct pv_entry *pvinit;
181 * Considering all the issues I'm having with pmap caching, if breakage
182 * continues to occur, and for debugging, I've added a sysctl that will
183 * just do an unconditional invltlb.
185 static int dreadful_invltlb;
187 SYSCTL_INT(_vm, OID_AUTO, dreadful_invltlb,
188 CTLFLAG_RW, &dreadful_invltlb, 0, "");
191 * All those kernel PT submaps that BSD is so fond of
193 pt_entry_t *CMAP1 = 0, *ptmmap;
194 caddr_t CADDR1 = 0, ptvmmap = 0;
195 static pt_entry_t *msgbufmap;
196 struct msgbuf *msgbufp=0;
201 static pt_entry_t *pt_crashdumpmap;
202 static caddr_t crashdumpmap;
204 extern pt_entry_t *SMPpt;
206 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
207 static unsigned * get_ptbase (pmap_t pmap);
208 static pv_entry_t get_pv_entry (void);
209 static void i386_protection_init (void);
210 static __inline void pmap_clearbit (vm_page_t m, int bit);
212 static void pmap_remove_all (vm_page_t m);
213 static void pmap_enter_quick (pmap_t pmap, vm_offset_t va, vm_page_t m);
214 static int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
215 vm_offset_t sva, pmap_inval_info_t info);
216 static void pmap_remove_page (struct pmap *pmap,
217 vm_offset_t va, pmap_inval_info_t info);
218 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
219 vm_offset_t va, pmap_inval_info_t info);
220 static boolean_t pmap_testbit (vm_page_t m, int bit);
221 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
222 vm_page_t mpte, vm_page_t m);
224 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
226 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
227 static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
228 static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
229 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
230 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
231 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
233 static unsigned pdir4mb;
236 * Move the kernel virtual free pointer to the next
237 * 4MB. This is used to help improve performance
238 * by using a large (4MB) page for much of the kernel
239 * (.text, .data, .bss)
242 pmap_kmem_choose(vm_offset_t addr)
244 vm_offset_t newaddr = addr;
246 if (cpu_feature & CPUID_PSE) {
247 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
256 * Extract the page table entry associated with the given map/virtual
259 * This function may NOT be called from an interrupt.
261 PMAP_INLINE unsigned *
262 pmap_pte(pmap_t pmap, vm_offset_t va)
267 pdeaddr = (unsigned *) pmap_pde(pmap, va);
268 if (*pdeaddr & PG_PS)
271 return get_ptbase(pmap) + i386_btop(va);
280 * Super fast pmap_pte routine best used when scanning the pv lists.
281 * This eliminates many course-grained invltlb calls. Note that many of
282 * the pv list scans are across different pmaps and it is very wasteful
283 * to do an entire invltlb when checking a single mapping.
285 * Should only be called while in a critical section.
287 * Unlike get_ptbase(), this function MAY be called from an interrupt or
291 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
293 struct mdglobaldata *gd = mdcpu;
296 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
297 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
298 unsigned index = i386_btop(va);
299 /* are we current address space or kernel? */
300 if ((pmap == &kernel_pmap) ||
301 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
302 return (unsigned *) PTmap + index;
304 newpf = pde & PG_FRAME;
305 if ( ((* (unsigned *) gd->gd_PMAP1) & PG_FRAME) != newpf) {
306 * (unsigned *) gd->gd_PMAP1 = newpf | PG_RW | PG_V;
307 cpu_invlpg(gd->gd_PADDR1);
309 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
316 * Bootstrap the system enough to run with virtual memory.
318 * On the i386 this is called after mapping has already been enabled
319 * and just syncs the pmap module with what has already been done.
320 * [We can't call it easily with mapping off since the kernel is not
321 * mapped with PA == VA, hence we would have to relocate every address
322 * from the linked base (virtual) address "KERNBASE" to the actual
323 * (physical) address starting relative to 0]
326 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
330 struct mdglobaldata *gd;
334 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
335 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
336 KvaEnd = KvaStart + KvaSize;
338 avail_start = firstaddr;
341 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
342 * too large. It should instead be correctly calculated in locore.s and
343 * not based on 'first' (which is a physical address, not a virtual
344 * address, for the start of unused physical memory). The kernel
345 * page tables are NOT double mapped and thus should not be included
346 * in this calculation.
348 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
349 virtual_start = pmap_kmem_choose(virtual_start);
350 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
353 * Initialize protection array.
355 i386_protection_init();
358 * The kernel's pmap is statically allocated so we don't have to use
359 * pmap_create, which is unlikely to work correctly at this part of
360 * the boot sequence (XXX and which no longer exists).
362 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
363 kernel_pmap.pm_count = 1;
364 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
365 TAILQ_INIT(&kernel_pmap.pm_pvlist);
369 * Reserve some special page table entries/VA space for temporary
372 #define SYSMAP(c, p, v, n) \
373 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
376 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
379 * CMAP1/CMAP2 are used for zeroing and copying pages.
381 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
386 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
389 * ptvmmap is used for reading arbitrary physical pages via
392 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
395 * msgbufp is used to map the system message buffer.
396 * XXX msgbufmap is not used.
398 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
399 atop(round_page(MSGBUF_SIZE)))
404 for (i = 0; i < NKPT; i++)
408 * PG_G is terribly broken on SMP because we IPI invltlb's in some
409 * cases rather then invl1pg. Actually, I don't even know why it
410 * works under UP because self-referential page table mappings
415 if (cpu_feature & CPUID_PGE)
420 * Initialize the 4MB page size flag
424 * The 4MB page version of the initial
425 * kernel page mapping.
429 #if !defined(DISABLE_PSE)
430 if (cpu_feature & CPUID_PSE) {
433 * Note that we have enabled PSE mode
436 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
437 ptditmp &= ~(NBPDR - 1);
438 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
443 * Enable the PSE mode. If we are SMP we can't do this
444 * now because the APs will not be able to use it when
447 load_cr4(rcr4() | CR4_PSE);
450 * We can do the mapping here for the single processor
451 * case. We simply ignore the old page table page from
455 * For SMP, we still need 4K pages to bootstrap APs,
456 * PSE will be enabled as soon as all APs are up.
458 PTD[KPTDI] = (pd_entry_t)ptditmp;
459 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
466 * We need to finish setting up the globaldata page for the BSP.
467 * locore has already populated the page table for the mdglobaldata
470 pg = MDGLOBALDATA_BASEALLOC_PAGES;
471 gd = &CPU_prvspace[0].mdglobaldata;
472 gd->gd_CMAP1 = &SMPpt[pg + 0];
473 gd->gd_CMAP2 = &SMPpt[pg + 1];
474 gd->gd_CMAP3 = &SMPpt[pg + 2];
475 gd->gd_PMAP1 = &SMPpt[pg + 3];
476 gd->gd_GDMAP1 = &PTD[KGDTDI];
477 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
478 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
479 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
480 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
481 gd->gd_GDADDR1= (unsigned *)VADDR(KGDTDI, 0);
488 * Set 4mb pdir for mp startup
493 if (pseflag && (cpu_feature & CPUID_PSE)) {
494 load_cr4(rcr4() | CR4_PSE);
495 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
496 kernel_pmap.pm_pdir[KPTDI] =
497 PTD[KPTDI] = (pd_entry_t)pdir4mb;
505 * Initialize the pmap module.
506 * Called by vm_init, to initialize any structures that the pmap
507 * system needs to map virtual memory.
508 * pmap_init has been enhanced to support in a fairly consistant
509 * way, discontiguous physical memory.
518 * object for kernel page table pages
520 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
523 * Allocate memory for random pmap data structures. Includes the
527 for(i = 0; i < vm_page_array_size; i++) {
530 m = &vm_page_array[i];
531 TAILQ_INIT(&m->md.pv_list);
532 m->md.pv_list_count = 0;
536 * init the pv free list
538 initial_pvs = vm_page_array_size;
539 if (initial_pvs < MINPV)
541 pvzone = &pvzone_store;
542 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
543 initial_pvs * sizeof (struct pv_entry));
544 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
548 * Now it is safe to enable pv_table recording.
550 pmap_initialized = TRUE;
554 * Initialize the address space (zone) for the pv_entries. Set a
555 * high water mark so that the system can recover from excessive
556 * numbers of pv entries.
561 int shpgperproc = PMAP_SHPGPERPROC;
563 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
564 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
565 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
566 pv_entry_high_water = 9 * (pv_entry_max / 10);
567 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
571 /***************************************************
572 * Low level helper routines.....
573 ***************************************************/
578 test_m_maps_pv(vm_page_t m, pv_entry_t pv)
584 KKASSERT(pv->pv_m == m);
586 TAILQ_FOREACH(spv, &m->md.pv_list, pv_list) {
593 panic("test_m_maps_pv: failed m %p pv %p\n", m, pv);
597 ptbase_assert(struct pmap *pmap)
599 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
601 /* are we current address space or kernel? */
602 if (pmap == &kernel_pmap || frame == (((unsigned)PTDpde) & PG_FRAME)) {
605 KKASSERT(frame == (*mycpu->gd_GDMAP1 & PG_FRAME));
610 #define test_m_maps_pv(m, pv)
611 #define ptbase_assert(pmap)
615 #if defined(PMAP_DIAGNOSTIC)
618 * This code checks for non-writeable/modified pages.
619 * This should be an invalid condition.
622 pmap_nw_modified(pt_entry_t ptea)
628 if ((pte & (PG_M|PG_RW)) == PG_M)
637 * this routine defines the region(s) of memory that should
638 * not be tested for the modified bit.
640 static PMAP_INLINE int
641 pmap_track_modified(vm_offset_t va)
643 if ((va < clean_sva) || (va >= clean_eva))
650 * Retrieve the mapped page table base for a particular pmap. Use our self
651 * mapping for the kernel_pmap or our current pmap.
653 * For foreign pmaps we use the per-cpu page table map. Since this involves
654 * installing a ptd it's actually (per-process x per-cpu). However, we
655 * still cannot depend on our mapping to survive thread switches because
656 * the process might be threaded and switching to another thread for the
657 * same process on the same cpu will allow that other thread to make its
660 * This could be a bit confusing but the jist is for something like the
661 * vkernel which uses foreign pmaps all the time this represents a pretty
662 * good cache that avoids unnecessary invltlb()s.
665 get_ptbase(pmap_t pmap)
667 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
668 struct mdglobaldata *gd = mdcpu;
671 * We can use PTmap if the pmap is our current address space or
672 * the kernel address space.
674 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
675 return (unsigned *) PTmap;
679 * Otherwise we use the per-cpu alternative page table map. Each
680 * cpu gets its own map. Because of this we cannot use this map
681 * from interrupts or threads which can preempt.
683 * Even if we already have the map cached we may still have to
684 * invalidate the TLB if another cpu modified a PDE in the map.
686 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
687 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
689 if ((*gd->gd_GDMAP1 & PG_FRAME) != frame) {
690 *gd->gd_GDMAP1 = frame | PG_RW | PG_V;
691 pmap->pm_cached |= gd->mi.gd_cpumask;
693 } else if ((pmap->pm_cached & gd->mi.gd_cpumask) == 0) {
694 pmap->pm_cached |= gd->mi.gd_cpumask;
696 } else if (dreadful_invltlb) {
699 return ((unsigned *)gd->gd_GDADDR1);
705 * Extract the physical page address associated with the map/VA pair.
707 * This function may not be called from an interrupt if the pmap is
711 pmap_extract(pmap_t pmap, vm_offset_t va)
714 vm_offset_t pdirindex;
716 pdirindex = va >> PDRSHIFT;
717 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
719 if ((rtval & PG_PS) != 0) {
720 rtval &= ~(NBPDR - 1);
721 rtval |= va & (NBPDR - 1);
724 pte = get_ptbase(pmap) + i386_btop(va);
725 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
731 /***************************************************
732 * Low level mapping routines.....
733 ***************************************************/
736 * Routine: pmap_kenter
738 * Add a wired page to the KVA
739 * NOTE! note that in order for the mapping to take effect -- you
740 * should do an invltlb after doing the pmap_kenter().
743 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
747 pmap_inval_info info;
749 pmap_inval_init(&info);
750 npte = pa | PG_RW | PG_V | pgeflag;
751 pte = (unsigned *)vtopte(va);
752 pmap_inval_add(&info, &kernel_pmap, va);
754 pmap_inval_flush(&info);
758 * Routine: pmap_kenter_quick
760 * Similar to pmap_kenter(), except we only invalidate the
761 * mapping on the current CPU.
764 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
769 npte = pa | PG_RW | PG_V | pgeflag;
770 pte = (unsigned *)vtopte(va);
772 cpu_invlpg((void *)va);
776 pmap_kenter_sync(vm_offset_t va)
778 pmap_inval_info info;
780 pmap_inval_init(&info);
781 pmap_inval_add(&info, &kernel_pmap, va);
782 pmap_inval_flush(&info);
786 pmap_kenter_sync_quick(vm_offset_t va)
788 cpu_invlpg((void *)va);
792 * remove a page from the kernel pagetables
795 pmap_kremove(vm_offset_t va)
798 pmap_inval_info info;
800 pmap_inval_init(&info);
801 pte = (unsigned *)vtopte(va);
802 pmap_inval_add(&info, &kernel_pmap, va);
804 pmap_inval_flush(&info);
808 pmap_kremove_quick(vm_offset_t va)
811 pte = (unsigned *)vtopte(va);
813 cpu_invlpg((void *)va);
817 * XXX these need to be recoded. They are not used in any critical path.
820 pmap_kmodify_rw(vm_offset_t va)
822 *vtopte(va) |= PG_RW;
823 cpu_invlpg((void *)va);
827 pmap_kmodify_nc(vm_offset_t va)
830 cpu_invlpg((void *)va);
834 * Used to map a range of physical addresses into kernel
835 * virtual address space.
837 * For now, VM is already on, we only need to map the
841 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
843 vm_offset_t sva, virt;
846 while (start < end) {
847 pmap_kenter(virt, start);
857 * Add a list of wired pages to the kva
858 * this routine is only used for temporary
859 * kernel mappings that do not need to have
860 * page modification or references recorded.
861 * Note that old mappings are simply written
862 * over. The page *must* be wired.
865 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
869 end_va = va + count * PAGE_SIZE;
871 while (va < end_va) {
874 pte = (unsigned *)vtopte(va);
875 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
876 cpu_invlpg((void *)va);
881 smp_invltlb(); /* XXX */
886 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
889 cpumask_t cmask = mycpu->gd_cpumask;
891 end_va = va + count * PAGE_SIZE;
893 while (va < end_va) {
898 * Install the new PTE. If the pte changed from the prior
899 * mapping we must reset the cpu mask and invalidate the page.
900 * If the pte is the same but we have not seen it on the
901 * current cpu, invlpg the existing mapping. Otherwise the
902 * entry is optimal and no invalidation is required.
904 pte = (unsigned *)vtopte(va);
905 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
906 if (*pte != pteval) {
909 cpu_invlpg((void *)va);
910 } else if ((*mask & cmask) == 0) {
911 cpu_invlpg((void *)va);
920 * This routine jerks page mappings from the
921 * kernel -- it is meant only for temporary mappings.
923 * MPSAFE, INTERRUPT SAFE (cluster callback)
926 pmap_qremove(vm_offset_t va, int count)
930 end_va = va + count*PAGE_SIZE;
932 while (va < end_va) {
935 pte = (unsigned *)vtopte(va);
937 cpu_invlpg((void *)va);
946 * This routine works like vm_page_lookup() but also blocks as long as the
947 * page is busy. This routine does not busy the page it returns.
949 * Unless the caller is managing objects whos pages are in a known state,
950 * the call should be made with a critical section held so the page's object
951 * association remains valid on return.
954 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
959 m = vm_page_lookup(object, pindex);
960 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
966 * Create a new thread and optionally associate it with a (new) process.
967 * NOTE! the new thread's cpu may not equal the current cpu.
970 pmap_init_thread(thread_t td)
972 /* enforce pcb placement */
973 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
974 td->td_savefpu = &td->td_pcb->pcb_save;
975 td->td_sp = (char *)td->td_pcb - 16;
979 * This routine directly affects the fork perf for a process.
982 pmap_init_proc(struct proc *p)
987 * Dispose the UPAGES for a process that has exited.
988 * This routine directly impacts the exit perf of a process.
991 pmap_dispose_proc(struct proc *p)
993 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
996 /***************************************************
997 * Page table page management routines.....
998 ***************************************************/
1001 * This routine unholds page table pages, and if the hold count
1002 * drops to zero, then it decrements the wire count.
1005 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1008 * Wait until we can busy the page ourselves. We cannot have
1009 * any active flushes if we block.
1011 if (m->flags & PG_BUSY) {
1012 pmap_inval_flush(info);
1013 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1016 KASSERT(m->queue == PQ_NONE,
1017 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1019 if (m->hold_count == 1) {
1021 * Unmap the page table page.
1023 * NOTE: We must clear pm_cached for all cpus, including
1024 * the current one, when clearing a page directory
1028 pmap_inval_add(info, pmap, -1);
1029 KKASSERT(pmap->pm_pdir[m->pindex]);
1030 pmap->pm_pdir[m->pindex] = 0;
1031 pmap->pm_cached = 0;
1033 KKASSERT(pmap->pm_stats.resident_count > 0);
1034 --pmap->pm_stats.resident_count;
1036 if (pmap->pm_ptphint == m)
1037 pmap->pm_ptphint = NULL;
1040 * This was our last hold, the page had better be unwired
1041 * after we decrement wire_count.
1043 * FUTURE NOTE: shared page directory page could result in
1044 * multiple wire counts.
1048 KKASSERT(m->wire_count == 0);
1049 --vmstats.v_wire_count;
1050 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1052 vm_page_free_zero(m);
1055 KKASSERT(m->hold_count > 1);
1061 static PMAP_INLINE int
1062 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1064 KKASSERT(m->hold_count > 0);
1065 if (m->hold_count > 1) {
1069 return _pmap_unwire_pte_hold(pmap, m, info);
1074 * After removing a page table entry, this routine is used to
1075 * conditionally free the page, and manage the hold/wire counts.
1077 * WARNING: This function can block
1080 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1081 pmap_inval_info_t info)
1084 if (va >= UPT_MIN_ADDRESS)
1088 ptepindex = (va >> PDRSHIFT);
1089 if (pmap->pm_ptphint &&
1090 (pmap->pm_ptphint->pindex == ptepindex)) {
1091 mpte = pmap->pm_ptphint;
1093 pmap_inval_flush(info);
1094 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1095 pmap->pm_ptphint = mpte;
1099 return pmap_unwire_pte_hold(pmap, mpte, info);
1103 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1104 * it, and IdlePTD, represents the template used to update all other pmaps.
1106 * On architectures where the kernel pmap is not integrated into the user
1107 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1108 * kernel_pmap should be used to directly access the kernel_pmap.
1111 pmap_pinit0(struct pmap *pmap)
1114 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1115 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1117 pmap->pm_active = 0;
1118 pmap->pm_cached = 0;
1119 pmap->pm_ptphint = NULL;
1120 TAILQ_INIT(&pmap->pm_pvlist);
1121 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1125 * Initialize a preallocated and zeroed pmap structure,
1126 * such as one in a vmspace structure.
1129 pmap_pinit(struct pmap *pmap)
1134 * No need to allocate page table space yet but we do need a valid
1135 * page directory table.
1137 if (pmap->pm_pdir == NULL) {
1139 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1143 * Allocate an object for the ptes
1145 if (pmap->pm_pteobj == NULL)
1146 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1149 * Allocate the page directory page, unless we already have
1150 * one cached. If we used the cached page the wire_count will
1151 * already be set appropriately.
1153 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1154 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1155 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1156 pmap->pm_pdirm = ptdpg;
1157 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1158 ptdpg->valid = VM_PAGE_BITS_ALL;
1159 ptdpg->wire_count = 1;
1160 ++vmstats.v_wire_count;
1161 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1163 if ((ptdpg->flags & PG_ZERO) == 0)
1164 bzero(pmap->pm_pdir, PAGE_SIZE);
1167 pmap_page_assertzero(VM_PAGE_TO_PHYS(ptdpg));
1170 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1172 /* install self-referential address mapping entry */
1173 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1174 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1177 pmap->pm_active = 0;
1178 pmap->pm_cached = 0;
1179 pmap->pm_ptphint = NULL;
1180 TAILQ_INIT(&pmap->pm_pvlist);
1181 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1182 pmap->pm_stats.resident_count = 1;
1186 * Clean up a pmap structure so it can be physically freed. This routine
1187 * is called by the vmspace dtor function. A great deal of pmap data is
1188 * left passively mapped to improve vmspace management so we have a bit
1189 * of cleanup work to do here.
1192 pmap_puninit(pmap_t pmap)
1196 KKASSERT(pmap->pm_active == 0);
1197 if ((p = pmap->pm_pdirm) != NULL) {
1198 KKASSERT(pmap->pm_pdir != NULL);
1199 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1201 vmstats.v_wire_count--;
1202 KKASSERT((p->flags & PG_BUSY) == 0);
1204 vm_page_free_zero(p);
1205 pmap->pm_pdirm = NULL;
1207 if (pmap->pm_pdir) {
1208 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1209 pmap->pm_pdir = NULL;
1211 if (pmap->pm_pteobj) {
1212 vm_object_deallocate(pmap->pm_pteobj);
1213 pmap->pm_pteobj = NULL;
1218 * Wire in kernel global address entries. To avoid a race condition
1219 * between pmap initialization and pmap_growkernel, this procedure
1220 * adds the pmap to the master list (which growkernel scans to update),
1221 * then copies the template.
1224 pmap_pinit2(struct pmap *pmap)
1227 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1228 /* XXX copies current process, does not fill in MPPTDI */
1229 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1234 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1235 * 0 on failure (if the procedure had to sleep).
1237 * When asked to remove the page directory page itself, we actually just
1238 * leave it cached so we do not have to incur the SMP inval overhead of
1239 * removing the kernel mapping. pmap_puninit() will take care of it.
1242 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1244 unsigned *pde = (unsigned *) pmap->pm_pdir;
1246 * This code optimizes the case of freeing non-busy
1247 * page-table pages. Those pages are zero now, and
1248 * might as well be placed directly into the zero queue.
1250 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1256 * Remove the page table page from the processes address space.
1258 KKASSERT(pmap->pm_stats.resident_count > 0);
1259 KKASSERT(pde[p->pindex]);
1261 --pmap->pm_stats.resident_count;
1262 pmap->pm_cached = 0;
1264 if (p->hold_count) {
1265 panic("pmap_release: freeing held page table page");
1267 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1268 pmap->pm_ptphint = NULL;
1271 * We leave the page directory page cached, wired, and mapped in
1272 * the pmap until the dtor function (pmap_puninit()) gets called.
1273 * However, still clean it up so we can set PG_ZERO.
1275 * The pmap has already been removed from the pmap_list in the
1278 if (p->pindex == PTDPTDI) {
1279 bzero(pde + KPTDI, nkpt * PTESIZE);
1280 bzero(pde + KGDTDI, (NPDEPG - KGDTDI) * PTESIZE);
1281 vm_page_flag_set(p, PG_ZERO);
1285 vmstats.v_wire_count--;
1286 vm_page_free_zero(p);
1292 * this routine is called if the page table page is not
1296 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1298 vm_offset_t pteva, ptepa;
1302 * Find or fabricate a new pagetable page
1304 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1305 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1307 KASSERT(m->queue == PQ_NONE,
1308 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1311 * Increment the hold count for the page we will be returning to
1317 * It is possible that someone else got in and mapped by the page
1318 * directory page while we were blocked, if so just unbusy and
1319 * return the held page.
1321 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1322 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1327 if (m->wire_count == 0)
1328 vmstats.v_wire_count++;
1333 * Map the pagetable page into the process address space, if
1334 * it isn't already there.
1336 * NOTE: For safety clear pm_cached for all cpus including the
1337 * current one when adding a PDE to the map.
1339 ++pmap->pm_stats.resident_count;
1341 ptepa = VM_PAGE_TO_PHYS(m);
1342 pmap->pm_pdir[ptepindex] =
1343 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1344 pmap->pm_cached = 0;
1347 * Set the page table hint
1349 pmap->pm_ptphint = m;
1352 * Try to use the new mapping, but if we cannot, then
1353 * do it with the routine that maps the page explicitly.
1355 if ((m->flags & PG_ZERO) == 0) {
1356 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1357 (((unsigned) PTDpde) & PG_FRAME)) {
1358 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1359 bzero((caddr_t) pteva, PAGE_SIZE);
1361 pmap_zero_page(ptepa);
1366 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1370 m->valid = VM_PAGE_BITS_ALL;
1371 vm_page_flag_clear(m, PG_ZERO);
1372 vm_page_flag_set(m, PG_MAPPED);
1379 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1386 * Calculate pagetable page index
1388 ptepindex = va >> PDRSHIFT;
1391 * Get the page directory entry
1393 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1396 * This supports switching from a 4MB page to a
1399 if (ptepa & PG_PS) {
1400 pmap->pm_pdir[ptepindex] = 0;
1407 * If the page table page is mapped, we just increment the
1408 * hold count, and activate it.
1412 * In order to get the page table page, try the
1415 if (pmap->pm_ptphint &&
1416 (pmap->pm_ptphint->pindex == ptepindex)) {
1417 m = pmap->pm_ptphint;
1419 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1420 pmap->pm_ptphint = m;
1426 * Here if the pte page isn't mapped, or if it has been deallocated.
1428 return _pmap_allocpte(pmap, ptepindex);
1432 /***************************************************
1433 * Pmap allocation/deallocation routines.
1434 ***************************************************/
1437 * Release any resources held by the given physical map.
1438 * Called when a pmap initialized by pmap_pinit is being released.
1439 * Should only be called if the map contains no valid mappings.
1441 static int pmap_release_callback(struct vm_page *p, void *data);
1444 pmap_release(struct pmap *pmap)
1446 vm_object_t object = pmap->pm_pteobj;
1447 struct rb_vm_page_scan_info info;
1449 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1450 #if defined(DIAGNOSTIC)
1451 if (object->ref_count != 1)
1452 panic("pmap_release: pteobj reference count != 1");
1456 info.object = object;
1458 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1465 info.limit = object->generation;
1467 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1468 pmap_release_callback, &info);
1469 if (info.error == 0 && info.mpte) {
1470 if (!pmap_release_free_page(pmap, info.mpte))
1474 } while (info.error);
1475 pmap->pm_cached = 0;
1479 pmap_release_callback(struct vm_page *p, void *data)
1481 struct rb_vm_page_scan_info *info = data;
1483 if (p->pindex == PTDPTDI) {
1487 if (!pmap_release_free_page(info->pmap, p)) {
1491 if (info->object->generation != info->limit) {
1499 * Grow the number of kernel page table entries, if needed.
1503 pmap_growkernel(vm_offset_t addr)
1506 vm_offset_t ptppaddr;
1511 if (kernel_vm_end == 0) {
1512 kernel_vm_end = KERNBASE;
1514 while (pdir_pde(PTD, kernel_vm_end)) {
1515 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1519 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1520 while (kernel_vm_end < addr) {
1521 if (pdir_pde(PTD, kernel_vm_end)) {
1522 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1527 * This index is bogus, but out of the way
1529 nkpg = vm_page_alloc(kptobj, nkpt,
1530 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1532 panic("pmap_growkernel: no memory to grow kernel");
1535 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1536 pmap_zero_page(ptppaddr);
1537 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1538 pdir_pde(PTD, kernel_vm_end) = newpdir;
1539 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1543 * This update must be interlocked with pmap_pinit2.
1545 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1546 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1548 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1549 ~(PAGE_SIZE * NPTEPG - 1);
1555 * Retire the given physical map from service.
1556 * Should only be called if the map contains
1557 * no valid mappings.
1560 pmap_destroy(pmap_t pmap)
1567 count = --pmap->pm_count;
1570 panic("destroying a pmap is not yet implemented");
1575 * Add a reference to the specified pmap.
1578 pmap_reference(pmap_t pmap)
1585 /***************************************************
1586 * page management routines.
1587 ***************************************************/
1590 * free the pv_entry back to the free list. This function may be
1591 * called from an interrupt.
1593 static PMAP_INLINE void
1594 free_pv_entry(pv_entry_t pv)
1597 KKASSERT(pv->pv_m != NULL);
1605 * get a new pv_entry, allocating a block from the system
1606 * when needed. This function may be called from an interrupt.
1612 if (pv_entry_high_water &&
1613 (pv_entry_count > pv_entry_high_water) &&
1614 (pmap_pagedaemon_waken == 0)) {
1615 pmap_pagedaemon_waken = 1;
1616 wakeup (&vm_pages_needed);
1618 return zalloc(pvzone);
1622 * This routine is very drastic, but can save the system
1630 static int warningdone=0;
1632 if (pmap_pagedaemon_waken == 0)
1634 pmap_pagedaemon_waken = 0;
1636 if (warningdone < 5) {
1637 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1641 for(i = 0; i < vm_page_array_size; i++) {
1642 m = &vm_page_array[i];
1643 if (m->wire_count || m->hold_count || m->busy ||
1644 (m->flags & PG_BUSY))
1652 * If it is the first entry on the list, it is actually
1653 * in the header and we must copy the following entry up
1654 * to the header. Otherwise we must search the list for
1655 * the entry. In either case we free the now unused entry.
1658 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1659 vm_offset_t va, pmap_inval_info_t info)
1665 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1666 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1667 if (pmap == pv->pv_pmap && va == pv->pv_va)
1671 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1673 KKASSERT(pv->pv_pmap == pmap);
1675 if (va == pv->pv_va)
1682 test_m_maps_pv(m, pv);
1683 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1684 m->md.pv_list_count--;
1685 if (TAILQ_EMPTY(&m->md.pv_list))
1686 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1687 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1688 ++pmap->pm_generation;
1689 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1696 * Create a pv entry for page at pa for
1700 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1705 pv = get_pv_entry();
1707 KKASSERT(pv->pv_m == NULL);
1714 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1715 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1716 ++pmap->pm_generation;
1717 m->md.pv_list_count++;
1723 * pmap_remove_pte: do the things to unmap a page in a process.
1725 * WARNING: This function may block (via pmap_remove_entry/pmap_unuse_pt),
1726 * callers using temporary pmaps must reload them.
1729 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1730 pmap_inval_info_t info)
1735 ptbase_assert(pmap);
1736 pmap_inval_add(info, pmap, va);
1737 ptbase_assert(pmap);
1738 oldpte = loadandclear(ptq);
1741 pmap->pm_stats.wired_count -= 1;
1743 * Machines that don't support invlpg, also don't support
1744 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1748 cpu_invlpg((void *)va);
1749 KKASSERT(pmap->pm_stats.resident_count > 0);
1750 --pmap->pm_stats.resident_count;
1751 if (oldpte & PG_MANAGED) {
1752 m = PHYS_TO_VM_PAGE(oldpte);
1753 if (oldpte & PG_M) {
1754 #if defined(PMAP_DIAGNOSTIC)
1755 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1756 kprintf("pmap_remove: modified page not "
1757 "writable: va: %p, pte: 0x%lx\n",
1758 (void *)va, (long)oldpte);
1761 if (pmap_track_modified(va))
1765 vm_page_flag_set(m, PG_REFERENCED);
1766 return pmap_remove_entry(pmap, m, va, info);
1768 return pmap_unuse_pt(pmap, va, NULL, info);
1777 * Remove a single page from a process address space.
1779 * This function may not be called from an interrupt if the pmap is
1783 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1788 * if there is no pte for this address, just skip it!!! Otherwise
1789 * get a local va for mappings for this pmap and remove the entry.
1791 if (*pmap_pde(pmap, va) != 0) {
1792 ptq = get_ptbase(pmap) + i386_btop(va);
1794 pmap_remove_pte(pmap, ptq, va, info);
1803 * Remove the given range of addresses from the specified map.
1805 * It is assumed that the start and end are properly
1806 * rounded to the page size.
1808 * This function may not be called from an interrupt if the pmap is
1812 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1816 vm_offset_t ptpaddr;
1817 vm_offset_t sindex, eindex;
1818 struct pmap_inval_info info;
1823 if (pmap->pm_stats.resident_count == 0)
1826 pmap_inval_init(&info);
1829 * special handling of removing one page. a very
1830 * common operation and easy to short circuit some
1833 if (((sva + PAGE_SIZE) == eva) &&
1834 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1835 pmap_remove_page(pmap, sva, &info);
1836 pmap_inval_flush(&info);
1841 * Get a local virtual address for the mappings that are being
1844 sindex = i386_btop(sva);
1845 eindex = i386_btop(eva);
1847 for (; sindex < eindex; sindex = pdnxt) {
1851 * Calculate index for next page table.
1853 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1854 if (pmap->pm_stats.resident_count == 0)
1857 pdirindex = sindex / NPDEPG;
1858 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1859 pmap_inval_add(&info, pmap, -1);
1860 pmap->pm_pdir[pdirindex] = 0;
1861 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1862 pmap->pm_cached = 0;
1867 * Weed out invalid mappings. Note: we assume that the page
1868 * directory table is always allocated, and in kernel virtual.
1874 * Limit our scan to either the end of the va represented
1875 * by the current page table page, or to the end of the
1876 * range being removed.
1878 if (pdnxt > eindex) {
1883 * NOTE: pmap_remove_pte() can block and wipe the temporary
1886 for (; sindex != pdnxt; sindex++) {
1889 ptbase = get_ptbase(pmap);
1890 if (ptbase[sindex] == 0)
1892 va = i386_ptob(sindex);
1893 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1897 pmap_inval_flush(&info);
1903 * Removes this physical page from all physical maps in which it resides.
1904 * Reflects back modify bits to the pager.
1906 * This routine may not be called from an interrupt.
1910 pmap_remove_all(vm_page_t m)
1912 struct pmap_inval_info info;
1913 unsigned *pte, tpte;
1916 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
1919 pmap_inval_init(&info);
1921 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1922 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1923 --pv->pv_pmap->pm_stats.resident_count;
1925 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1926 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1927 tpte = loadandclear(pte);
1929 KKASSERT(PHYS_TO_VM_PAGE(tpte) == m);
1932 pv->pv_pmap->pm_stats.wired_count--;
1935 vm_page_flag_set(m, PG_REFERENCED);
1938 * Update the vm_page_t clean and reference bits.
1941 #if defined(PMAP_DIAGNOSTIC)
1942 if (pmap_nw_modified((pt_entry_t) tpte)) {
1943 kprintf("pmap_remove_all: modified page "
1944 "not writable: va: %p, pte: 0x%lx\n",
1945 (void *)pv->pv_va, (long)tpte);
1948 if (pmap_track_modified(pv->pv_va))
1952 KKASSERT(pv->pv_m == m);
1954 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1955 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1956 ++pv->pv_pmap->pm_generation;
1957 m->md.pv_list_count--;
1958 if (TAILQ_EMPTY(&m->md.pv_list))
1959 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1960 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1964 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
1965 pmap_inval_flush(&info);
1971 * Set the physical protection on the specified range of this map
1974 * This function may not be called from an interrupt if the map is
1975 * not the kernel_pmap.
1978 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1981 vm_offset_t pdnxt, ptpaddr;
1982 vm_pindex_t sindex, eindex;
1983 pmap_inval_info info;
1988 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1989 pmap_remove(pmap, sva, eva);
1993 if (prot & VM_PROT_WRITE)
1996 pmap_inval_init(&info);
1998 ptbase = get_ptbase(pmap);
2000 sindex = i386_btop(sva);
2001 eindex = i386_btop(eva);
2003 for (; sindex < eindex; sindex = pdnxt) {
2007 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
2009 pdirindex = sindex / NPDEPG;
2010 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
2011 pmap_inval_add(&info, pmap, -1);
2012 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
2013 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2018 * Weed out invalid mappings. Note: we assume that the page
2019 * directory table is always allocated, and in kernel virtual.
2024 if (pdnxt > eindex) {
2028 for (; sindex != pdnxt; sindex++) {
2034 * XXX non-optimal. Note also that there can be
2035 * no pmap_inval_flush() calls until after we modify
2036 * ptbase[sindex] (or otherwise we have to do another
2037 * pmap_inval_add() call).
2039 pmap_inval_add(&info, pmap, i386_ptob(sindex));
2040 pbits = ptbase[sindex];
2042 if (pbits & PG_MANAGED) {
2045 m = PHYS_TO_VM_PAGE(pbits);
2046 vm_page_flag_set(m, PG_REFERENCED);
2050 if (pmap_track_modified(i386_ptob(sindex))) {
2052 m = PHYS_TO_VM_PAGE(pbits);
2061 if (pbits != ptbase[sindex]) {
2062 ptbase[sindex] = pbits;
2066 pmap_inval_flush(&info);
2070 * Insert the given physical page (p) at
2071 * the specified virtual address (v) in the
2072 * target physical map with the protection requested.
2074 * If specified, the page will be wired down, meaning
2075 * that the related pte can not be reclaimed.
2077 * NB: This is the only routine which MAY NOT lazy-evaluate
2078 * or lose information. That is, this routine must actually
2079 * insert this page into the given map NOW.
2082 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2088 vm_offset_t origpte, newpte;
2090 pmap_inval_info info;
2096 #ifdef PMAP_DIAGNOSTIC
2098 panic("pmap_enter: toobig");
2099 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) {
2100 panic("pmap_enter: invalid to pmap_enter page "
2101 "table pages (va: %p)", (void *)va);
2104 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2105 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2108 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2109 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2114 * In the case that a page table page is not
2115 * resident, we are creating it here.
2117 if (va < UPT_MIN_ADDRESS)
2118 mpte = pmap_allocpte(pmap, va);
2122 pmap_inval_init(&info);
2123 pte = pmap_pte(pmap, va);
2126 * Page Directory table entry not valid, we need a new PT page
2129 panic("pmap_enter: invalid page directory pdir=0x%lx, va=%p\n",
2130 (long)pmap->pm_pdir[PTDPTDI], (void *)va);
2133 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
2134 origpte = *(vm_offset_t *)pte;
2135 opa = origpte & PG_FRAME;
2137 if (origpte & PG_PS)
2138 panic("pmap_enter: attempted pmap_enter on 4MB page");
2141 * Mapping has not changed, must be protection or wiring change.
2143 if (origpte && (opa == pa)) {
2145 * Wiring change, just update stats. We don't worry about
2146 * wiring PT pages as they remain resident as long as there
2147 * are valid mappings in them. Hence, if a user page is wired,
2148 * the PT page will be also.
2150 if (wired && ((origpte & PG_W) == 0))
2151 pmap->pm_stats.wired_count++;
2152 else if (!wired && (origpte & PG_W))
2153 pmap->pm_stats.wired_count--;
2155 #if defined(PMAP_DIAGNOSTIC)
2156 if (pmap_nw_modified((pt_entry_t) origpte)) {
2157 kprintf("pmap_enter: modified page not "
2158 "writable: va: %p, pte: 0x%lx\n",
2159 (void *)va, (long )origpte);
2164 * Remove the extra pte reference. Note that we cannot
2165 * optimize the RO->RW case because we have adjusted the
2166 * wiring count above and may need to adjust the wiring
2173 * We might be turning off write access to the page,
2174 * so we go ahead and sense modify status.
2176 if (origpte & PG_MANAGED) {
2177 if ((origpte & PG_M) && pmap_track_modified(va)) {
2179 om = PHYS_TO_VM_PAGE(opa);
2183 KKASSERT(m->flags & PG_MAPPED);
2188 * Mapping has changed, invalidate old range and fall through to
2189 * handle validating new mapping.
2191 * Since we have a ref on the page directory page pmap_pte()
2192 * will always return non-NULL.
2194 * NOTE: pmap_remove_pte() can block and cause the temporary ptbase
2195 * to get wiped. reload the ptbase. I'm not sure if it is
2196 * also possible to race another pmap_enter() but check for
2202 KKASSERT((origpte & PG_FRAME) ==
2203 (*(vm_offset_t *)pte & PG_FRAME));
2204 err = pmap_remove_pte(pmap, pte, va, &info);
2206 panic("pmap_enter: pte vanished, va: %p", (void *)va);
2207 pte = pmap_pte(pmap, va);
2208 origpte = *(vm_offset_t *)pte;
2209 opa = origpte & PG_FRAME;
2211 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2217 * Enter on the PV list if part of our managed memory. Note that we
2218 * raise IPL while manipulating pv_table since pmap_enter can be
2219 * called at interrupt time.
2221 if (pmap_initialized &&
2222 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2223 pmap_insert_entry(pmap, va, mpte, m);
2224 ptbase_assert(pmap);
2226 vm_page_flag_set(m, PG_MAPPED);
2230 * Increment counters
2232 ++pmap->pm_stats.resident_count;
2234 pmap->pm_stats.wired_count++;
2235 KKASSERT(*pte == 0);
2239 * Now validate mapping with desired protection/wiring.
2241 ptbase_assert(pmap);
2242 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2246 if (va < UPT_MIN_ADDRESS)
2248 if (pmap == &kernel_pmap)
2252 * if the mapping or permission bits are different, we need
2253 * to update the pte.
2255 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2256 pmap_inval_add(&info, pmap, va);
2257 ptbase_assert(pmap);
2258 KKASSERT(*pte == 0 ||
2259 (*pte & PG_FRAME) == (newpte & PG_FRAME));
2260 *pte = newpte | PG_A;
2262 vm_page_flag_set(m, PG_WRITEABLE);
2264 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2265 pmap_inval_flush(&info);
2269 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2270 * This code also assumes that the pmap has no pre-existing entry for this
2273 * This code currently may only be used on user pmaps, not kernel_pmap.
2276 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2283 pmap_inval_info info;
2285 pmap_inval_init(&info);
2287 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2288 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2291 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2292 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2296 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2299 * Calculate the page table page (mpte), allocating it if necessary.
2301 * A held page table page (mpte), or NULL, is passed onto the
2302 * section following.
2304 if (va < UPT_MIN_ADDRESS) {
2306 * Calculate pagetable page index
2308 ptepindex = va >> PDRSHIFT;
2312 * Get the page directory entry
2314 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2317 * If the page table page is mapped, we just increment
2318 * the hold count, and activate it.
2322 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2323 if (pmap->pm_ptphint &&
2324 (pmap->pm_ptphint->pindex == ptepindex)) {
2325 mpte = pmap->pm_ptphint;
2327 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2328 pmap->pm_ptphint = mpte;
2333 mpte = _pmap_allocpte(pmap, ptepindex);
2335 } while (mpte == NULL);
2338 /* this code path is not yet used */
2342 * With a valid (and held) page directory page, we can just use
2343 * vtopte() to get to the pte. If the pte is already present
2344 * we do not disturb it.
2346 pte = (unsigned *)vtopte(va);
2349 pmap_unwire_pte_hold(pmap, mpte, &info);
2350 pa = VM_PAGE_TO_PHYS(m);
2351 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2356 * Enter on the PV list if part of our managed memory
2358 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2359 pmap_insert_entry(pmap, va, mpte, m);
2360 vm_page_flag_set(m, PG_MAPPED);
2364 * Increment counters
2366 ++pmap->pm_stats.resident_count;
2368 pa = VM_PAGE_TO_PHYS(m);
2371 * Now validate mapping with RO protection
2373 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2374 *pte = pa | PG_V | PG_U;
2376 *pte = pa | PG_V | PG_U | PG_MANAGED;
2377 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2378 pmap_inval_flush(&info);
2382 * Make a temporary mapping for a physical address. This is only intended
2383 * to be used for panic dumps.
2386 pmap_kenter_temporary(vm_paddr_t pa, int i)
2388 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2389 return ((void *)crashdumpmap);
2392 #define MAX_INIT_PT (96)
2395 * This routine preloads the ptes for a given object into the specified pmap.
2396 * This eliminates the blast of soft faults on process startup and
2397 * immediately after an mmap.
2399 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2402 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2403 vm_object_t object, vm_pindex_t pindex,
2404 vm_size_t size, int limit)
2406 struct rb_vm_page_scan_info info;
2411 * We can't preinit if read access isn't set or there is no pmap
2414 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2418 * We can't preinit if the pmap is not the current pmap
2420 lp = curthread->td_lwp;
2421 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2424 psize = i386_btop(size);
2426 if ((object->type != OBJT_VNODE) ||
2427 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2428 (object->resident_page_count > MAX_INIT_PT))) {
2432 if (psize + pindex > object->size) {
2433 if (object->size < pindex)
2435 psize = object->size - pindex;
2442 * Use a red-black scan to traverse the requested range and load
2443 * any valid pages found into the pmap.
2445 * We cannot safely scan the object's memq unless we are in a
2446 * critical section since interrupts can remove pages from objects.
2448 info.start_pindex = pindex;
2449 info.end_pindex = pindex + psize - 1;
2456 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2457 pmap_object_init_pt_callback, &info);
2463 pmap_object_init_pt_callback(vm_page_t p, void *data)
2465 struct rb_vm_page_scan_info *info = data;
2466 vm_pindex_t rel_index;
2468 * don't allow an madvise to blow away our really
2469 * free pages allocating pv entries.
2471 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2472 vmstats.v_free_count < vmstats.v_free_reserved) {
2475 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2476 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2477 if ((p->queue - p->pc) == PQ_CACHE)
2478 vm_page_deactivate(p);
2480 rel_index = p->pindex - info->start_pindex;
2481 pmap_enter_quick(info->pmap,
2482 info->addr + i386_ptob(rel_index), p);
2489 * pmap_prefault provides a quick way of clustering pagefaults into a
2490 * processes address space. It is a "cousin" of pmap_object_init_pt,
2491 * except it runs at page fault time instead of mmap time.
2495 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2497 static int pmap_prefault_pageorder[] = {
2498 -PAGE_SIZE, PAGE_SIZE,
2499 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2500 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2501 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2505 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2516 * We do not currently prefault mappings that use virtual page
2517 * tables. We do not prefault foreign pmaps.
2519 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2521 lp = curthread->td_lwp;
2522 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2525 object = entry->object.vm_object;
2527 starta = addra - PFBAK * PAGE_SIZE;
2528 if (starta < entry->start)
2529 starta = entry->start;
2530 else if (starta > addra)
2534 * critical section protection is required to maintain the
2535 * page/object association, interrupts can free pages and remove
2536 * them from their objects.
2539 for (i = 0; i < PAGEORDER_SIZE; i++) {
2540 vm_object_t lobject;
2543 addr = addra + pmap_prefault_pageorder[i];
2544 if (addr > addra + (PFFOR * PAGE_SIZE))
2547 if (addr < starta || addr >= entry->end)
2550 if ((*pmap_pde(pmap, addr)) == 0)
2553 pte = (unsigned *) vtopte(addr);
2557 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2560 for (m = vm_page_lookup(lobject, pindex);
2561 (!m && (lobject->type == OBJT_DEFAULT) &&
2562 (lobject->backing_object));
2563 lobject = lobject->backing_object
2565 if (lobject->backing_object_offset & PAGE_MASK)
2567 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2568 m = vm_page_lookup(lobject->backing_object, pindex);
2572 * give-up when a page is not in memory
2577 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2579 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2581 if ((m->queue - m->pc) == PQ_CACHE) {
2582 vm_page_deactivate(m);
2585 pmap_enter_quick(pmap, addr, m);
2593 * Routine: pmap_change_wiring
2594 * Function: Change the wiring attribute for a map/virtual-address
2596 * In/out conditions:
2597 * The mapping must already exist in the pmap.
2600 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2607 pte = pmap_pte(pmap, va);
2609 if (wired && !pmap_pte_w(pte))
2610 pmap->pm_stats.wired_count++;
2611 else if (!wired && pmap_pte_w(pte))
2612 pmap->pm_stats.wired_count--;
2615 * Wiring is not a hardware characteristic so there is no need to
2616 * invalidate TLB. However, in an SMP environment we must use
2617 * a locked bus cycle to update the pte (if we are not using
2618 * the pmap_inval_*() API that is)... it's ok to do this for simple
2623 atomic_set_int(pte, PG_W);
2625 atomic_clear_int(pte, PG_W);
2628 atomic_set_int_nonlocked(pte, PG_W);
2630 atomic_clear_int_nonlocked(pte, PG_W);
2637 * Copy the range specified by src_addr/len
2638 * from the source map to the range dst_addr/len
2639 * in the destination map.
2641 * This routine is only advisory and need not do anything.
2644 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2645 vm_size_t len, vm_offset_t src_addr)
2648 pmap_inval_info info;
2650 vm_offset_t end_addr = src_addr + len;
2652 unsigned src_frame, dst_frame;
2656 if (dst_addr != src_addr)
2659 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2660 * valid through blocking calls, and that's just not going to
2667 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2668 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2672 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2673 if ((*gd->gd_GDMAP1 & PG_FRAME) != dst_frame) {
2674 *gd->gd_GDMAP1 = dst_frame | PG_RW | PG_V;
2678 pmap_inval_init(&info);
2679 pmap_inval_add(&info, dst_pmap, -1);
2680 pmap_inval_add(&info, src_pmap, -1);
2683 * critical section protection is required to maintain the page/object
2684 * association, interrupts can free pages and remove them from
2688 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2689 unsigned *src_pte, *dst_pte;
2690 vm_page_t dstmpte, srcmpte;
2691 vm_offset_t srcptepaddr;
2694 if (addr >= UPT_MIN_ADDRESS)
2695 panic("pmap_copy: invalid to pmap_copy page tables\n");
2698 * Don't let optional prefaulting of pages make us go
2699 * way below the low water mark of free pages or way
2700 * above high water mark of used pv entries.
2702 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2703 pv_entry_count > pv_entry_high_water)
2706 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2707 ptepindex = addr >> PDRSHIFT;
2709 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2710 if (srcptepaddr == 0)
2713 if (srcptepaddr & PG_PS) {
2714 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2715 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2716 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2722 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2723 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2724 (srcmpte->flags & PG_BUSY)) {
2728 if (pdnxt > end_addr)
2731 src_pte = (unsigned *) vtopte(addr);
2732 dst_pte = (unsigned *) avtopte(addr);
2733 while (addr < pdnxt) {
2738 * we only virtual copy managed pages
2740 if ((ptetemp & PG_MANAGED) != 0) {
2742 * We have to check after allocpte for the
2743 * pte still being around... allocpte can
2746 * pmap_allocpte() can block. If we lose
2747 * our page directory mappings we stop.
2749 dstmpte = pmap_allocpte(dst_pmap, addr);
2751 if (src_frame != (((unsigned) PTDpde) & PG_FRAME) ||
2752 XXX dst_frame != (((unsigned) xxx) & PG_FRAME)
2754 kprintf("WARNING: pmap_copy: detected and corrected race\n");
2755 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2757 } else if ((*dst_pte == 0) &&
2758 (ptetemp = *src_pte) != 0 &&
2759 (ptetemp & PG_MANAGED)) {
2761 * Clear the modified and
2762 * accessed (referenced) bits
2765 m = PHYS_TO_VM_PAGE(ptetemp);
2766 *dst_pte = ptetemp & ~(PG_M | PG_A);
2767 ++dst_pmap->pm_stats.resident_count;
2768 pmap_insert_entry(dst_pmap, addr,
2770 KKASSERT(m->flags & PG_MAPPED);
2772 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
2773 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2776 if (dstmpte->hold_count >= srcmpte->hold_count)
2786 pmap_inval_flush(&info);
2793 * Zero the specified PA by mapping the page into KVM and clearing its
2796 * This function may be called from an interrupt and no locking is
2800 pmap_zero_page(vm_paddr_t phys)
2802 struct mdglobaldata *gd = mdcpu;
2805 if (*(int *)gd->gd_CMAP3)
2806 panic("pmap_zero_page: CMAP3 busy");
2807 *(int *)gd->gd_CMAP3 =
2808 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2809 cpu_invlpg(gd->gd_CADDR3);
2811 #if defined(I686_CPU)
2812 if (cpu_class == CPUCLASS_686)
2813 i686_pagezero(gd->gd_CADDR3);
2816 bzero(gd->gd_CADDR3, PAGE_SIZE);
2817 *(int *) gd->gd_CMAP3 = 0;
2822 * pmap_page_assertzero:
2824 * Assert that a page is empty, panic if it isn't.
2827 pmap_page_assertzero(vm_paddr_t phys)
2829 struct mdglobaldata *gd = mdcpu;
2833 if (*(int *)gd->gd_CMAP3)
2834 panic("pmap_zero_page: CMAP3 busy");
2835 *(int *)gd->gd_CMAP3 =
2836 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2837 cpu_invlpg(gd->gd_CADDR3);
2838 for (i = 0; i < PAGE_SIZE; i += 4) {
2839 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2840 panic("pmap_page_assertzero() @ %p not zero!\n",
2841 (void *)gd->gd_CADDR3);
2844 *(int *) gd->gd_CMAP3 = 0;
2851 * Zero part of a physical page by mapping it into memory and clearing
2852 * its contents with bzero.
2854 * off and size may not cover an area beyond a single hardware page.
2857 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2859 struct mdglobaldata *gd = mdcpu;
2862 if (*(int *) gd->gd_CMAP3)
2863 panic("pmap_zero_page: CMAP3 busy");
2864 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2865 cpu_invlpg(gd->gd_CADDR3);
2867 #if defined(I686_CPU)
2868 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2869 i686_pagezero(gd->gd_CADDR3);
2872 bzero((char *)gd->gd_CADDR3 + off, size);
2873 *(int *) gd->gd_CMAP3 = 0;
2880 * Copy the physical page from the source PA to the target PA.
2881 * This function may be called from an interrupt. No locking
2885 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2887 struct mdglobaldata *gd = mdcpu;
2890 if (*(int *) gd->gd_CMAP1)
2891 panic("pmap_copy_page: CMAP1 busy");
2892 if (*(int *) gd->gd_CMAP2)
2893 panic("pmap_copy_page: CMAP2 busy");
2895 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2896 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2898 cpu_invlpg(gd->gd_CADDR1);
2899 cpu_invlpg(gd->gd_CADDR2);
2901 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2903 *(int *) gd->gd_CMAP1 = 0;
2904 *(int *) gd->gd_CMAP2 = 0;
2909 * pmap_copy_page_frag:
2911 * Copy the physical page from the source PA to the target PA.
2912 * This function may be called from an interrupt. No locking
2916 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2918 struct mdglobaldata *gd = mdcpu;
2921 if (*(int *) gd->gd_CMAP1)
2922 panic("pmap_copy_page: CMAP1 busy");
2923 if (*(int *) gd->gd_CMAP2)
2924 panic("pmap_copy_page: CMAP2 busy");
2926 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2927 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2929 cpu_invlpg(gd->gd_CADDR1);
2930 cpu_invlpg(gd->gd_CADDR2);
2932 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2933 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2936 *(int *) gd->gd_CMAP1 = 0;
2937 *(int *) gd->gd_CMAP2 = 0;
2942 * Returns true if the pmap's pv is one of the first
2943 * 16 pvs linked to from this page. This count may
2944 * be changed upwards or downwards in the future; it
2945 * is only necessary that true be returned for a small
2946 * subset of pmaps for proper page aging.
2949 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2954 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2959 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2960 if (pv->pv_pmap == pmap) {
2973 * Remove all pages from specified address space
2974 * this aids process exit speeds. Also, this code
2975 * is special cased for current process only, but
2976 * can have the more generic (and slightly slower)
2977 * mode enabled. This is much faster than pmap_remove
2978 * in the case of running down an entire address space.
2981 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2984 unsigned *pte, tpte;
2987 pmap_inval_info info;
2989 int32_t save_generation;
2991 lp = curthread->td_lwp;
2992 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2997 pmap_inval_init(&info);
2999 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3000 if (pv->pv_va >= eva || pv->pv_va < sva) {
3001 npv = TAILQ_NEXT(pv, pv_plist);
3005 KKASSERT(pmap == pv->pv_pmap);
3008 pte = (unsigned *)vtopte(pv->pv_va);
3010 pte = pmap_pte_quick(pmap, pv->pv_va);
3012 if (pmap->pm_active)
3013 pmap_inval_add(&info, pmap, pv->pv_va);
3016 * We cannot remove wired pages from a process' mapping
3020 npv = TAILQ_NEXT(pv, pv_plist);
3024 tpte = loadandclear(pte);
3026 m = PHYS_TO_VM_PAGE(tpte);
3027 test_m_maps_pv(m, pv);
3029 KASSERT(m < &vm_page_array[vm_page_array_size],
3030 ("pmap_remove_pages: bad tpte %x", tpte));
3032 KKASSERT(pmap->pm_stats.resident_count > 0);
3033 --pmap->pm_stats.resident_count;
3036 * Update the vm_page_t clean and reference bits.
3042 npv = TAILQ_NEXT(pv, pv_plist);
3044 KKASSERT(pv->pv_m == m);
3045 KKASSERT(pv->pv_pmap == pmap);
3047 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3048 save_generation = ++pmap->pm_generation;
3050 m->md.pv_list_count--;
3051 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3052 if (TAILQ_EMPTY(&m->md.pv_list))
3053 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3055 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3059 * Restart the scan if we blocked during the unuse or free
3060 * calls and other removals were made.
3062 if (save_generation != pmap->pm_generation) {
3063 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3064 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3067 pmap_inval_flush(&info);
3072 * pmap_testbit tests bits in pte's
3073 * note that the testbit/clearbit routines are inline,
3074 * and a lot of things compile-time evaluate.
3077 pmap_testbit(vm_page_t m, int bit)
3082 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3085 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3090 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3092 * if the bit being tested is the modified bit, then
3093 * mark clean_map and ptes as never
3096 if (bit & (PG_A|PG_M)) {
3097 if (!pmap_track_modified(pv->pv_va))
3101 #if defined(PMAP_DIAGNOSTIC)
3103 kprintf("Null pmap (tb) at va: %p\n",
3108 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3119 * this routine is used to modify bits in ptes
3121 static __inline void
3122 pmap_clearbit(vm_page_t m, int bit)
3124 struct pmap_inval_info info;
3129 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3132 pmap_inval_init(&info);
3136 * Loop over all current mappings setting/clearing as appropos If
3137 * setting RO do we need to clear the VAC?
3139 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3141 * don't write protect pager mappings
3144 if (!pmap_track_modified(pv->pv_va))
3148 #if defined(PMAP_DIAGNOSTIC)
3150 kprintf("Null pmap (cb) at va: %p\n",
3157 * Careful here. We can use a locked bus instruction to
3158 * clear PG_A or PG_M safely but we need to synchronize
3159 * with the target cpus when we mess with PG_RW.
3161 * We do not have to force synchronization when clearing
3162 * PG_M even for PTEs generated via virtual memory maps,
3163 * because the virtual kernel will invalidate the pmap
3164 * entry when/if it needs to resynchronize the Modify bit.
3167 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
3168 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3175 atomic_clear_int(pte, PG_M|PG_RW);
3178 * The cpu may be trying to set PG_M
3179 * simultaniously with our clearing
3182 if (!atomic_cmpset_int(pte, pbits,
3186 } else if (bit == PG_M) {
3188 * We could also clear PG_RW here to force
3189 * a fault on write to redetect PG_M for
3190 * virtual kernels, but it isn't necessary
3191 * since virtual kernels invalidate the pte
3192 * when they clear the VPTE_M bit in their
3193 * virtual page tables.
3195 atomic_clear_int(pte, PG_M);
3197 atomic_clear_int(pte, bit);
3201 pmap_inval_flush(&info);
3206 * pmap_page_protect:
3208 * Lower the permission for all mappings to a given page.
3211 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3213 if ((prot & VM_PROT_WRITE) == 0) {
3214 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3215 pmap_clearbit(m, PG_RW);
3216 vm_page_flag_clear(m, PG_WRITEABLE);
3224 pmap_phys_address(vm_pindex_t ppn)
3226 return (i386_ptob(ppn));
3230 * pmap_ts_referenced:
3232 * Return a count of reference bits for a page, clearing those bits.
3233 * It is not necessary for every reference bit to be cleared, but it
3234 * is necessary that 0 only be returned when there are truly no
3235 * reference bits set.
3237 * XXX: The exact number of bits to check and clear is a matter that
3238 * should be tested and standardized at some point in the future for
3239 * optimal aging of shared pages.
3242 pmap_ts_referenced(vm_page_t m)
3244 pv_entry_t pv, pvf, pvn;
3248 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3253 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3258 pvn = TAILQ_NEXT(pv, pv_list);
3261 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3262 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3265 if (!pmap_track_modified(pv->pv_va))
3268 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3270 if (pte && (*pte & PG_A)) {
3272 atomic_clear_int(pte, PG_A);
3274 atomic_clear_int_nonlocked(pte, PG_A);
3281 } while ((pv = pvn) != NULL && pv != pvf);
3291 * Return whether or not the specified physical page was modified
3292 * in any physical maps.
3295 pmap_is_modified(vm_page_t m)
3297 return pmap_testbit(m, PG_M);
3301 * Clear the modify bits on the specified physical page.
3304 pmap_clear_modify(vm_page_t m)
3306 pmap_clearbit(m, PG_M);
3310 * pmap_clear_reference:
3312 * Clear the reference bit on the specified physical page.
3315 pmap_clear_reference(vm_page_t m)
3317 pmap_clearbit(m, PG_A);
3321 * Miscellaneous support routines follow
3325 i386_protection_init(void)
3329 kp = protection_codes;
3330 for (prot = 0; prot < 8; prot++) {
3332 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3334 * Read access is also 0. There isn't any execute bit,
3335 * so just make it readable.
3337 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3338 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3339 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3342 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3343 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3344 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3345 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3353 * Map a set of physical memory pages into the kernel virtual
3354 * address space. Return a pointer to where it is mapped. This
3355 * routine is intended to be used for mapping device memory,
3358 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3362 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3364 vm_offset_t va, tmpva, offset;
3367 offset = pa & PAGE_MASK;
3368 size = roundup(offset + size, PAGE_SIZE);
3370 va = kmem_alloc_nofault(&kernel_map, size);
3372 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3375 for (tmpva = va; size > 0;) {
3376 pte = (unsigned *)vtopte(tmpva);
3377 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3385 return ((void *)(va + offset));
3389 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3391 vm_offset_t base, offset;
3393 base = va & PG_FRAME;
3394 offset = va & PAGE_MASK;
3395 size = roundup(offset + size, PAGE_SIZE);
3396 pmap_qremove(va, size >> PAGE_SHIFT);
3397 kmem_free(&kernel_map, base, size);
3401 * perform the pmap work for mincore
3404 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3406 unsigned *ptep, pte;
3410 ptep = pmap_pte(pmap, addr);
3415 if ((pte = *ptep) != 0) {
3418 val = MINCORE_INCORE;
3419 if ((pte & PG_MANAGED) == 0)
3422 pa = pte & PG_FRAME;
3424 m = PHYS_TO_VM_PAGE(pa);
3430 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3432 * Modified by someone
3434 else if (m->dirty || pmap_is_modified(m))
3435 val |= MINCORE_MODIFIED_OTHER;
3440 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3443 * Referenced by someone
3445 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3446 val |= MINCORE_REFERENCED_OTHER;
3447 vm_page_flag_set(m, PG_REFERENCED);
3454 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3455 * vmspace will be ref'd and the old one will be deref'd.
3457 * The vmspace for all lwps associated with the process will be adjusted
3458 * and cr3 will be reloaded if any lwp is the current lwp.
3461 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3463 struct vmspace *oldvm;
3467 oldvm = p->p_vmspace;
3468 if (oldvm != newvm) {
3469 p->p_vmspace = newvm;
3470 KKASSERT(p->p_nthreads == 1);
3471 lp = RB_ROOT(&p->p_lwp_tree);
3472 pmap_setlwpvm(lp, newvm);
3474 sysref_get(&newvm->vm_sysref);
3475 sysref_put(&oldvm->vm_sysref);
3482 * Set the vmspace for a LWP. The vmspace is almost universally set the
3483 * same as the process vmspace, but virtual kernels need to swap out contexts
3484 * on a per-lwp basis.
3487 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3489 struct vmspace *oldvm;
3493 oldvm = lp->lwp_vmspace;
3495 if (oldvm != newvm) {
3496 lp->lwp_vmspace = newvm;
3497 if (curthread->td_lwp == lp) {
3498 pmap = vmspace_pmap(newvm);
3500 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3502 pmap->pm_active |= 1;
3504 #if defined(SWTCH_OPTIM_STATS)
3507 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3508 load_cr3(curthread->td_pcb->pcb_cr3);
3509 pmap = vmspace_pmap(oldvm);
3511 atomic_clear_int(&pmap->pm_active,
3512 1 << mycpu->gd_cpuid);
3514 pmap->pm_active &= ~1;
3522 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3525 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3529 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3534 pmap_get_pgeflag(void)
3541 static void pads (pmap_t pm);
3542 void pmap_pvdump (vm_paddr_t pa);
3544 /* print address space of pmap*/
3551 if (pm == &kernel_pmap)
3554 for (i = 0; i < 1024; i++) {
3555 if (pm->pm_pdir[i]) {
3556 for (j = 0; j < 1024; j++) {
3557 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3558 if (pm == &kernel_pmap && va < KERNBASE)
3560 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3562 ptep = pmap_pte_quick(pm, va);
3563 if (pmap_pte_v(ptep))
3564 kprintf("%x:%x ", va, *(int *) ptep);
3573 pmap_pvdump(vm_paddr_t pa)
3578 kprintf("pa %08llx", (long long)pa);
3579 m = PHYS_TO_VM_PAGE(pa);
3580 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3582 kprintf(" -> pmap %p, va %p, flags %x",
3583 (void *)pv->pv_pmap, (long)pv->pv_va, pv->pv_flags);
3585 kprintf(" -> pmap %p, va %p",
3586 (void *)pv->pv_pmap, (void *)pv->pv_va);