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 int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
214 vm_offset_t sva, pmap_inval_info_t info);
215 static void pmap_remove_page (struct pmap *pmap,
216 vm_offset_t va, pmap_inval_info_t info);
217 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
218 vm_offset_t va, pmap_inval_info_t info);
219 static boolean_t pmap_testbit (vm_page_t m, int bit);
220 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
221 vm_page_t mpte, vm_page_t m);
223 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
225 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
226 static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
227 static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
228 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
229 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
230 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
232 static unsigned pdir4mb;
235 * Move the kernel virtual free pointer to the next
236 * 4MB. This is used to help improve performance
237 * by using a large (4MB) page for much of the kernel
238 * (.text, .data, .bss)
241 pmap_kmem_choose(vm_offset_t addr)
243 vm_offset_t newaddr = addr;
245 if (cpu_feature & CPUID_PSE) {
246 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
255 * Extract the page table entry associated with the given map/virtual
258 * This function may NOT be called from an interrupt.
260 PMAP_INLINE unsigned *
261 pmap_pte(pmap_t pmap, vm_offset_t va)
266 pdeaddr = (unsigned *) pmap_pde(pmap, va);
267 if (*pdeaddr & PG_PS)
270 return get_ptbase(pmap) + i386_btop(va);
279 * Super fast pmap_pte routine best used when scanning the pv lists.
280 * This eliminates many course-grained invltlb calls. Note that many of
281 * the pv list scans are across different pmaps and it is very wasteful
282 * to do an entire invltlb when checking a single mapping.
284 * Should only be called while in a critical section.
286 * Unlike get_ptbase(), this function MAY be called from an interrupt or
290 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
292 struct mdglobaldata *gd = mdcpu;
295 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
296 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
297 unsigned index = i386_btop(va);
298 /* are we current address space or kernel? */
299 if ((pmap == &kernel_pmap) ||
300 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
301 return (unsigned *) PTmap + index;
303 newpf = pde & PG_FRAME;
304 if ( ((* (unsigned *) gd->gd_PMAP1) & PG_FRAME) != newpf) {
305 * (unsigned *) gd->gd_PMAP1 = newpf | PG_RW | PG_V;
306 cpu_invlpg(gd->gd_PADDR1);
308 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
315 * Bootstrap the system enough to run with virtual memory.
317 * On the i386 this is called after mapping has already been enabled
318 * and just syncs the pmap module with what has already been done.
319 * [We can't call it easily with mapping off since the kernel is not
320 * mapped with PA == VA, hence we would have to relocate every address
321 * from the linked base (virtual) address "KERNBASE" to the actual
322 * (physical) address starting relative to 0]
325 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
329 struct mdglobaldata *gd;
333 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
334 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
335 KvaEnd = KvaStart + KvaSize;
337 avail_start = firstaddr;
340 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
341 * too large. It should instead be correctly calculated in locore.s and
342 * not based on 'first' (which is a physical address, not a virtual
343 * address, for the start of unused physical memory). The kernel
344 * page tables are NOT double mapped and thus should not be included
345 * in this calculation.
347 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
348 virtual_start = pmap_kmem_choose(virtual_start);
349 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
352 * Initialize protection array.
354 i386_protection_init();
357 * The kernel's pmap is statically allocated so we don't have to use
358 * pmap_create, which is unlikely to work correctly at this part of
359 * the boot sequence (XXX and which no longer exists).
361 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
362 kernel_pmap.pm_count = 1;
363 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
364 TAILQ_INIT(&kernel_pmap.pm_pvlist);
368 * Reserve some special page table entries/VA space for temporary
371 #define SYSMAP(c, p, v, n) \
372 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
375 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
378 * CMAP1/CMAP2 are used for zeroing and copying pages.
380 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
385 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
388 * ptvmmap is used for reading arbitrary physical pages via
391 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
394 * msgbufp is used to map the system message buffer.
395 * XXX msgbufmap is not used.
397 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
398 atop(round_page(MSGBUF_SIZE)))
403 for (i = 0; i < NKPT; i++)
407 * PG_G is terribly broken on SMP because we IPI invltlb's in some
408 * cases rather then invl1pg. Actually, I don't even know why it
409 * works under UP because self-referential page table mappings
414 if (cpu_feature & CPUID_PGE)
419 * Initialize the 4MB page size flag
423 * The 4MB page version of the initial
424 * kernel page mapping.
428 #if !defined(DISABLE_PSE)
429 if (cpu_feature & CPUID_PSE) {
432 * Note that we have enabled PSE mode
435 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
436 ptditmp &= ~(NBPDR - 1);
437 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
442 * Enable the PSE mode. If we are SMP we can't do this
443 * now because the APs will not be able to use it when
446 load_cr4(rcr4() | CR4_PSE);
449 * We can do the mapping here for the single processor
450 * case. We simply ignore the old page table page from
454 * For SMP, we still need 4K pages to bootstrap APs,
455 * PSE will be enabled as soon as all APs are up.
457 PTD[KPTDI] = (pd_entry_t)ptditmp;
458 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
465 * We need to finish setting up the globaldata page for the BSP.
466 * locore has already populated the page table for the mdglobaldata
469 pg = MDGLOBALDATA_BASEALLOC_PAGES;
470 gd = &CPU_prvspace[0].mdglobaldata;
471 gd->gd_CMAP1 = &SMPpt[pg + 0];
472 gd->gd_CMAP2 = &SMPpt[pg + 1];
473 gd->gd_CMAP3 = &SMPpt[pg + 2];
474 gd->gd_PMAP1 = &SMPpt[pg + 3];
475 gd->gd_GDMAP1 = &PTD[KGDTDI];
476 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
477 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
478 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
479 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
480 gd->gd_GDADDR1= (unsigned *)VADDR(KGDTDI, 0);
487 * Set 4mb pdir for mp startup
492 if (pseflag && (cpu_feature & CPUID_PSE)) {
493 load_cr4(rcr4() | CR4_PSE);
494 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
495 kernel_pmap.pm_pdir[KPTDI] =
496 PTD[KPTDI] = (pd_entry_t)pdir4mb;
504 * Initialize the pmap module.
505 * Called by vm_init, to initialize any structures that the pmap
506 * system needs to map virtual memory.
507 * pmap_init has been enhanced to support in a fairly consistant
508 * way, discontiguous physical memory.
517 * object for kernel page table pages
519 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
522 * Allocate memory for random pmap data structures. Includes the
526 for(i = 0; i < vm_page_array_size; i++) {
529 m = &vm_page_array[i];
530 TAILQ_INIT(&m->md.pv_list);
531 m->md.pv_list_count = 0;
535 * init the pv free list
537 initial_pvs = vm_page_array_size;
538 if (initial_pvs < MINPV)
540 pvzone = &pvzone_store;
541 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
542 initial_pvs * sizeof (struct pv_entry));
543 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
547 * Now it is safe to enable pv_table recording.
549 pmap_initialized = TRUE;
553 * Initialize the address space (zone) for the pv_entries. Set a
554 * high water mark so that the system can recover from excessive
555 * numbers of pv entries.
560 int shpgperproc = PMAP_SHPGPERPROC;
562 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
563 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
564 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
565 pv_entry_high_water = 9 * (pv_entry_max / 10);
566 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
570 /***************************************************
571 * Low level helper routines.....
572 ***************************************************/
577 test_m_maps_pv(vm_page_t m, pv_entry_t pv)
583 KKASSERT(pv->pv_m == m);
585 TAILQ_FOREACH(spv, &m->md.pv_list, pv_list) {
592 panic("test_m_maps_pv: failed m %p pv %p\n", m, pv);
596 ptbase_assert(struct pmap *pmap)
598 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
600 /* are we current address space or kernel? */
601 if (pmap == &kernel_pmap || frame == (((unsigned)PTDpde) & PG_FRAME)) {
604 KKASSERT(frame == (*mycpu->gd_GDMAP1 & PG_FRAME));
609 #define test_m_maps_pv(m, pv)
610 #define ptbase_assert(pmap)
614 #if defined(PMAP_DIAGNOSTIC)
617 * This code checks for non-writeable/modified pages.
618 * This should be an invalid condition.
621 pmap_nw_modified(pt_entry_t ptea)
627 if ((pte & (PG_M|PG_RW)) == PG_M)
636 * this routine defines the region(s) of memory that should
637 * not be tested for the modified bit.
639 static PMAP_INLINE int
640 pmap_track_modified(vm_offset_t va)
642 if ((va < clean_sva) || (va >= clean_eva))
649 * Retrieve the mapped page table base for a particular pmap. Use our self
650 * mapping for the kernel_pmap or our current pmap.
652 * For foreign pmaps we use the per-cpu page table map. Since this involves
653 * installing a ptd it's actually (per-process x per-cpu). However, we
654 * still cannot depend on our mapping to survive thread switches because
655 * the process might be threaded and switching to another thread for the
656 * same process on the same cpu will allow that other thread to make its
659 * This could be a bit confusing but the jist is for something like the
660 * vkernel which uses foreign pmaps all the time this represents a pretty
661 * good cache that avoids unnecessary invltlb()s.
664 get_ptbase(pmap_t pmap)
666 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
667 struct mdglobaldata *gd = mdcpu;
670 * We can use PTmap if the pmap is our current address space or
671 * the kernel address space.
673 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
674 return (unsigned *) PTmap;
678 * Otherwise we use the per-cpu alternative page table map. Each
679 * cpu gets its own map. Because of this we cannot use this map
680 * from interrupts or threads which can preempt.
682 * Even if we already have the map cached we may still have to
683 * invalidate the TLB if another cpu modified a PDE in the map.
685 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
686 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
688 if ((*gd->gd_GDMAP1 & PG_FRAME) != frame) {
689 *gd->gd_GDMAP1 = frame | PG_RW | PG_V;
690 pmap->pm_cached |= gd->mi.gd_cpumask;
692 } else if ((pmap->pm_cached & gd->mi.gd_cpumask) == 0) {
693 pmap->pm_cached |= gd->mi.gd_cpumask;
695 } else if (dreadful_invltlb) {
698 return ((unsigned *)gd->gd_GDADDR1);
704 * Extract the physical page address associated with the map/VA pair.
706 * This function may not be called from an interrupt if the pmap is
710 pmap_extract(pmap_t pmap, vm_offset_t va)
713 vm_offset_t pdirindex;
715 pdirindex = va >> PDRSHIFT;
716 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
718 if ((rtval & PG_PS) != 0) {
719 rtval &= ~(NBPDR - 1);
720 rtval |= va & (NBPDR - 1);
723 pte = get_ptbase(pmap) + i386_btop(va);
724 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
730 /***************************************************
731 * Low level mapping routines.....
732 ***************************************************/
735 * Routine: pmap_kenter
737 * Add a wired page to the KVA
738 * NOTE! note that in order for the mapping to take effect -- you
739 * should do an invltlb after doing the pmap_kenter().
742 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
746 pmap_inval_info info;
748 pmap_inval_init(&info);
749 npte = pa | PG_RW | PG_V | pgeflag;
750 pte = (unsigned *)vtopte(va);
751 pmap_inval_add(&info, &kernel_pmap, va);
753 pmap_inval_flush(&info);
757 * Routine: pmap_kenter_quick
759 * Similar to pmap_kenter(), except we only invalidate the
760 * mapping on the current CPU.
763 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
768 npte = pa | PG_RW | PG_V | pgeflag;
769 pte = (unsigned *)vtopte(va);
771 cpu_invlpg((void *)va);
775 pmap_kenter_sync(vm_offset_t va)
777 pmap_inval_info info;
779 pmap_inval_init(&info);
780 pmap_inval_add(&info, &kernel_pmap, va);
781 pmap_inval_flush(&info);
785 pmap_kenter_sync_quick(vm_offset_t va)
787 cpu_invlpg((void *)va);
791 * remove a page from the kernel pagetables
794 pmap_kremove(vm_offset_t va)
797 pmap_inval_info info;
799 pmap_inval_init(&info);
800 pte = (unsigned *)vtopte(va);
801 pmap_inval_add(&info, &kernel_pmap, va);
803 pmap_inval_flush(&info);
807 pmap_kremove_quick(vm_offset_t va)
810 pte = (unsigned *)vtopte(va);
812 cpu_invlpg((void *)va);
816 * XXX these need to be recoded. They are not used in any critical path.
819 pmap_kmodify_rw(vm_offset_t va)
821 *vtopte(va) |= PG_RW;
822 cpu_invlpg((void *)va);
826 pmap_kmodify_nc(vm_offset_t va)
829 cpu_invlpg((void *)va);
833 * Used to map a range of physical addresses into kernel
834 * virtual address space.
836 * For now, VM is already on, we only need to map the
840 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
842 vm_offset_t sva, virt;
845 while (start < end) {
846 pmap_kenter(virt, start);
856 * Add a list of wired pages to the kva
857 * this routine is only used for temporary
858 * kernel mappings that do not need to have
859 * page modification or references recorded.
860 * Note that old mappings are simply written
861 * over. The page *must* be wired.
864 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
868 end_va = va + count * PAGE_SIZE;
870 while (va < end_va) {
873 pte = (unsigned *)vtopte(va);
874 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
875 cpu_invlpg((void *)va);
880 smp_invltlb(); /* XXX */
885 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
888 cpumask_t cmask = mycpu->gd_cpumask;
890 end_va = va + count * PAGE_SIZE;
892 while (va < end_va) {
897 * Install the new PTE. If the pte changed from the prior
898 * mapping we must reset the cpu mask and invalidate the page.
899 * If the pte is the same but we have not seen it on the
900 * current cpu, invlpg the existing mapping. Otherwise the
901 * entry is optimal and no invalidation is required.
903 pte = (unsigned *)vtopte(va);
904 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
905 if (*pte != pteval) {
908 cpu_invlpg((void *)va);
909 } else if ((*mask & cmask) == 0) {
910 cpu_invlpg((void *)va);
919 * This routine jerks page mappings from the
920 * kernel -- it is meant only for temporary mappings.
922 * MPSAFE, INTERRUPT SAFE (cluster callback)
925 pmap_qremove(vm_offset_t va, int count)
929 end_va = va + count*PAGE_SIZE;
931 while (va < end_va) {
934 pte = (unsigned *)vtopte(va);
936 cpu_invlpg((void *)va);
945 * This routine works like vm_page_lookup() but also blocks as long as the
946 * page is busy. This routine does not busy the page it returns.
948 * Unless the caller is managing objects whos pages are in a known state,
949 * the call should be made with a critical section held so the page's object
950 * association remains valid on return.
953 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
958 m = vm_page_lookup(object, pindex);
959 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
965 * Create a new thread and optionally associate it with a (new) process.
966 * NOTE! the new thread's cpu may not equal the current cpu.
969 pmap_init_thread(thread_t td)
971 /* enforce pcb placement */
972 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
973 td->td_savefpu = &td->td_pcb->pcb_save;
974 td->td_sp = (char *)td->td_pcb - 16;
978 * This routine directly affects the fork perf for a process.
981 pmap_init_proc(struct proc *p)
986 * Dispose the UPAGES for a process that has exited.
987 * This routine directly impacts the exit perf of a process.
990 pmap_dispose_proc(struct proc *p)
992 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
995 /***************************************************
996 * Page table page management routines.....
997 ***************************************************/
1000 * This routine unholds page table pages, and if the hold count
1001 * drops to zero, then it decrements the wire count.
1004 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1007 * Wait until we can busy the page ourselves. We cannot have
1008 * any active flushes if we block.
1010 if (m->flags & PG_BUSY) {
1011 pmap_inval_flush(info);
1012 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1015 KASSERT(m->queue == PQ_NONE,
1016 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1018 if (m->hold_count == 1) {
1020 * Unmap the page table page.
1022 * NOTE: We must clear pm_cached for all cpus, including
1023 * the current one, when clearing a page directory
1027 pmap_inval_add(info, pmap, -1);
1028 KKASSERT(pmap->pm_pdir[m->pindex]);
1029 pmap->pm_pdir[m->pindex] = 0;
1030 pmap->pm_cached = 0;
1032 KKASSERT(pmap->pm_stats.resident_count > 0);
1033 --pmap->pm_stats.resident_count;
1035 if (pmap->pm_ptphint == m)
1036 pmap->pm_ptphint = NULL;
1039 * This was our last hold, the page had better be unwired
1040 * after we decrement wire_count.
1042 * FUTURE NOTE: shared page directory page could result in
1043 * multiple wire counts.
1047 KKASSERT(m->wire_count == 0);
1048 --vmstats.v_wire_count;
1049 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1051 vm_page_free_zero(m);
1054 KKASSERT(m->hold_count > 1);
1060 static PMAP_INLINE int
1061 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
1063 KKASSERT(m->hold_count > 0);
1064 if (m->hold_count > 1) {
1068 return _pmap_unwire_pte_hold(pmap, m, info);
1073 * After removing a page table entry, this routine is used to
1074 * conditionally free the page, and manage the hold/wire counts.
1076 * WARNING: This function can block
1079 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1080 pmap_inval_info_t info)
1083 if (va >= UPT_MIN_ADDRESS)
1087 ptepindex = (va >> PDRSHIFT);
1088 if (pmap->pm_ptphint &&
1089 (pmap->pm_ptphint->pindex == ptepindex)) {
1090 mpte = pmap->pm_ptphint;
1092 pmap_inval_flush(info);
1093 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1094 pmap->pm_ptphint = mpte;
1098 return pmap_unwire_pte_hold(pmap, mpte, info);
1102 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1103 * it, and IdlePTD, represents the template used to update all other pmaps.
1105 * On architectures where the kernel pmap is not integrated into the user
1106 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1107 * kernel_pmap should be used to directly access the kernel_pmap.
1110 pmap_pinit0(struct pmap *pmap)
1113 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1114 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1116 pmap->pm_active = 0;
1117 pmap->pm_cached = 0;
1118 pmap->pm_ptphint = NULL;
1119 TAILQ_INIT(&pmap->pm_pvlist);
1120 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1124 * Initialize a preallocated and zeroed pmap structure,
1125 * such as one in a vmspace structure.
1128 pmap_pinit(struct pmap *pmap)
1133 * No need to allocate page table space yet but we do need a valid
1134 * page directory table.
1136 if (pmap->pm_pdir == NULL) {
1138 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1142 * Allocate an object for the ptes
1144 if (pmap->pm_pteobj == NULL)
1145 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1148 * Allocate the page directory page, unless we already have
1149 * one cached. If we used the cached page the wire_count will
1150 * already be set appropriately.
1152 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1153 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1154 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1155 pmap->pm_pdirm = ptdpg;
1156 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1157 ptdpg->valid = VM_PAGE_BITS_ALL;
1158 ptdpg->wire_count = 1;
1159 ++vmstats.v_wire_count;
1160 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1162 if ((ptdpg->flags & PG_ZERO) == 0)
1163 bzero(pmap->pm_pdir, PAGE_SIZE);
1166 pmap_page_assertzero(VM_PAGE_TO_PHYS(ptdpg));
1169 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1171 /* install self-referential address mapping entry */
1172 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1173 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1176 pmap->pm_active = 0;
1177 pmap->pm_cached = 0;
1178 pmap->pm_ptphint = NULL;
1179 TAILQ_INIT(&pmap->pm_pvlist);
1180 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1181 pmap->pm_stats.resident_count = 1;
1185 * Clean up a pmap structure so it can be physically freed. This routine
1186 * is called by the vmspace dtor function. A great deal of pmap data is
1187 * left passively mapped to improve vmspace management so we have a bit
1188 * of cleanup work to do here.
1191 pmap_puninit(pmap_t pmap)
1195 KKASSERT(pmap->pm_active == 0);
1196 if ((p = pmap->pm_pdirm) != NULL) {
1197 KKASSERT(pmap->pm_pdir != NULL);
1198 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1200 vmstats.v_wire_count--;
1201 KKASSERT((p->flags & PG_BUSY) == 0);
1203 vm_page_free_zero(p);
1204 pmap->pm_pdirm = NULL;
1206 if (pmap->pm_pdir) {
1207 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1208 pmap->pm_pdir = NULL;
1210 if (pmap->pm_pteobj) {
1211 vm_object_deallocate(pmap->pm_pteobj);
1212 pmap->pm_pteobj = NULL;
1217 * Wire in kernel global address entries. To avoid a race condition
1218 * between pmap initialization and pmap_growkernel, this procedure
1219 * adds the pmap to the master list (which growkernel scans to update),
1220 * then copies the template.
1223 pmap_pinit2(struct pmap *pmap)
1226 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1227 /* XXX copies current process, does not fill in MPPTDI */
1228 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1233 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1234 * 0 on failure (if the procedure had to sleep).
1236 * When asked to remove the page directory page itself, we actually just
1237 * leave it cached so we do not have to incur the SMP inval overhead of
1238 * removing the kernel mapping. pmap_puninit() will take care of it.
1241 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1243 unsigned *pde = (unsigned *) pmap->pm_pdir;
1245 * This code optimizes the case of freeing non-busy
1246 * page-table pages. Those pages are zero now, and
1247 * might as well be placed directly into the zero queue.
1249 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1255 * Remove the page table page from the processes address space.
1257 KKASSERT(pmap->pm_stats.resident_count > 0);
1258 KKASSERT(pde[p->pindex]);
1260 --pmap->pm_stats.resident_count;
1261 pmap->pm_cached = 0;
1263 if (p->hold_count) {
1264 panic("pmap_release: freeing held page table page");
1266 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1267 pmap->pm_ptphint = NULL;
1270 * We leave the page directory page cached, wired, and mapped in
1271 * the pmap until the dtor function (pmap_puninit()) gets called.
1272 * However, still clean it up so we can set PG_ZERO.
1274 * The pmap has already been removed from the pmap_list in the
1277 if (p->pindex == PTDPTDI) {
1278 bzero(pde + KPTDI, nkpt * PTESIZE);
1279 bzero(pde + KGDTDI, (NPDEPG - KGDTDI) * PTESIZE);
1280 vm_page_flag_set(p, PG_ZERO);
1284 vmstats.v_wire_count--;
1285 vm_page_free_zero(p);
1291 * this routine is called if the page table page is not
1295 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1297 vm_offset_t pteva, ptepa;
1301 * Find or fabricate a new pagetable page
1303 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1304 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1306 KASSERT(m->queue == PQ_NONE,
1307 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1310 * Increment the hold count for the page we will be returning to
1316 * It is possible that someone else got in and mapped by the page
1317 * directory page while we were blocked, if so just unbusy and
1318 * return the held page.
1320 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1321 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1326 if (m->wire_count == 0)
1327 vmstats.v_wire_count++;
1332 * Map the pagetable page into the process address space, if
1333 * it isn't already there.
1335 * NOTE: For safety clear pm_cached for all cpus including the
1336 * current one when adding a PDE to the map.
1338 ++pmap->pm_stats.resident_count;
1340 ptepa = VM_PAGE_TO_PHYS(m);
1341 pmap->pm_pdir[ptepindex] =
1342 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1343 pmap->pm_cached = 0;
1346 * Set the page table hint
1348 pmap->pm_ptphint = m;
1351 * Try to use the new mapping, but if we cannot, then
1352 * do it with the routine that maps the page explicitly.
1354 if ((m->flags & PG_ZERO) == 0) {
1355 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1356 (((unsigned) PTDpde) & PG_FRAME)) {
1357 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1358 bzero((caddr_t) pteva, PAGE_SIZE);
1360 pmap_zero_page(ptepa);
1365 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1369 m->valid = VM_PAGE_BITS_ALL;
1370 vm_page_flag_clear(m, PG_ZERO);
1371 vm_page_flag_set(m, PG_MAPPED);
1378 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1385 * Calculate pagetable page index
1387 ptepindex = va >> PDRSHIFT;
1390 * Get the page directory entry
1392 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1395 * This supports switching from a 4MB page to a
1398 if (ptepa & PG_PS) {
1399 pmap->pm_pdir[ptepindex] = 0;
1406 * If the page table page is mapped, we just increment the
1407 * hold count, and activate it.
1411 * In order to get the page table page, try the
1414 if (pmap->pm_ptphint &&
1415 (pmap->pm_ptphint->pindex == ptepindex)) {
1416 m = pmap->pm_ptphint;
1418 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1419 pmap->pm_ptphint = m;
1425 * Here if the pte page isn't mapped, or if it has been deallocated.
1427 return _pmap_allocpte(pmap, ptepindex);
1431 /***************************************************
1432 * Pmap allocation/deallocation routines.
1433 ***************************************************/
1436 * Release any resources held by the given physical map.
1437 * Called when a pmap initialized by pmap_pinit is being released.
1438 * Should only be called if the map contains no valid mappings.
1440 static int pmap_release_callback(struct vm_page *p, void *data);
1443 pmap_release(struct pmap *pmap)
1445 vm_object_t object = pmap->pm_pteobj;
1446 struct rb_vm_page_scan_info info;
1448 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1449 #if defined(DIAGNOSTIC)
1450 if (object->ref_count != 1)
1451 panic("pmap_release: pteobj reference count != 1");
1455 info.object = object;
1457 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1464 info.limit = object->generation;
1466 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1467 pmap_release_callback, &info);
1468 if (info.error == 0 && info.mpte) {
1469 if (!pmap_release_free_page(pmap, info.mpte))
1473 } while (info.error);
1474 pmap->pm_cached = 0;
1478 pmap_release_callback(struct vm_page *p, void *data)
1480 struct rb_vm_page_scan_info *info = data;
1482 if (p->pindex == PTDPTDI) {
1486 if (!pmap_release_free_page(info->pmap, p)) {
1490 if (info->object->generation != info->limit) {
1498 * Grow the number of kernel page table entries, if needed.
1502 pmap_growkernel(vm_offset_t addr)
1505 vm_offset_t ptppaddr;
1510 if (kernel_vm_end == 0) {
1511 kernel_vm_end = KERNBASE;
1513 while (pdir_pde(PTD, kernel_vm_end)) {
1514 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1518 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1519 while (kernel_vm_end < addr) {
1520 if (pdir_pde(PTD, kernel_vm_end)) {
1521 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1526 * This index is bogus, but out of the way
1528 nkpg = vm_page_alloc(kptobj, nkpt,
1529 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1531 panic("pmap_growkernel: no memory to grow kernel");
1534 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1535 pmap_zero_page(ptppaddr);
1536 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1537 pdir_pde(PTD, kernel_vm_end) = newpdir;
1538 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1542 * This update must be interlocked with pmap_pinit2.
1544 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1545 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1547 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1548 ~(PAGE_SIZE * NPTEPG - 1);
1554 * Retire the given physical map from service.
1555 * Should only be called if the map contains
1556 * no valid mappings.
1559 pmap_destroy(pmap_t pmap)
1566 count = --pmap->pm_count;
1569 panic("destroying a pmap is not yet implemented");
1574 * Add a reference to the specified pmap.
1577 pmap_reference(pmap_t pmap)
1584 /***************************************************
1585 * page management routines.
1586 ***************************************************/
1589 * free the pv_entry back to the free list. This function may be
1590 * called from an interrupt.
1592 static PMAP_INLINE void
1593 free_pv_entry(pv_entry_t pv)
1596 KKASSERT(pv->pv_m != NULL);
1604 * get a new pv_entry, allocating a block from the system
1605 * when needed. This function may be called from an interrupt.
1611 if (pv_entry_high_water &&
1612 (pv_entry_count > pv_entry_high_water) &&
1613 (pmap_pagedaemon_waken == 0)) {
1614 pmap_pagedaemon_waken = 1;
1615 wakeup (&vm_pages_needed);
1617 return zalloc(pvzone);
1621 * This routine is very drastic, but can save the system
1629 static int warningdone=0;
1631 if (pmap_pagedaemon_waken == 0)
1633 pmap_pagedaemon_waken = 0;
1635 if (warningdone < 5) {
1636 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1640 for(i = 0; i < vm_page_array_size; i++) {
1641 m = &vm_page_array[i];
1642 if (m->wire_count || m->hold_count || m->busy ||
1643 (m->flags & PG_BUSY))
1651 * If it is the first entry on the list, it is actually
1652 * in the header and we must copy the following entry up
1653 * to the header. Otherwise we must search the list for
1654 * the entry. In either case we free the now unused entry.
1657 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1658 vm_offset_t va, pmap_inval_info_t info)
1664 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1665 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1666 if (pmap == pv->pv_pmap && va == pv->pv_va)
1670 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1672 KKASSERT(pv->pv_pmap == pmap);
1674 if (va == pv->pv_va)
1681 test_m_maps_pv(m, pv);
1682 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1683 m->md.pv_list_count--;
1684 if (TAILQ_EMPTY(&m->md.pv_list))
1685 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1686 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1687 ++pmap->pm_generation;
1688 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1695 * Create a pv entry for page at pa for
1699 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1704 pv = get_pv_entry();
1706 KKASSERT(pv->pv_m == NULL);
1713 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1714 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1715 ++pmap->pm_generation;
1716 m->md.pv_list_count++;
1722 * pmap_remove_pte: do the things to unmap a page in a process.
1724 * WARNING: This function may block (via pmap_remove_entry/pmap_unuse_pt),
1725 * callers using temporary pmaps must reload them.
1728 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1729 pmap_inval_info_t info)
1734 ptbase_assert(pmap);
1735 pmap_inval_add(info, pmap, va);
1736 ptbase_assert(pmap);
1737 oldpte = loadandclear(ptq);
1740 pmap->pm_stats.wired_count -= 1;
1742 * Machines that don't support invlpg, also don't support
1743 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1747 cpu_invlpg((void *)va);
1748 KKASSERT(pmap->pm_stats.resident_count > 0);
1749 --pmap->pm_stats.resident_count;
1750 if (oldpte & PG_MANAGED) {
1751 m = PHYS_TO_VM_PAGE(oldpte);
1752 if (oldpte & PG_M) {
1753 #if defined(PMAP_DIAGNOSTIC)
1754 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1755 kprintf("pmap_remove: modified page not "
1756 "writable: va: %p, pte: 0x%lx\n",
1757 (void *)va, (long)oldpte);
1760 if (pmap_track_modified(va))
1764 vm_page_flag_set(m, PG_REFERENCED);
1765 return pmap_remove_entry(pmap, m, va, info);
1767 return pmap_unuse_pt(pmap, va, NULL, info);
1776 * Remove a single page from a process address space.
1778 * This function may not be called from an interrupt if the pmap is
1782 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1787 * if there is no pte for this address, just skip it!!! Otherwise
1788 * get a local va for mappings for this pmap and remove the entry.
1790 if (*pmap_pde(pmap, va) != 0) {
1791 ptq = get_ptbase(pmap) + i386_btop(va);
1793 pmap_remove_pte(pmap, ptq, va, info);
1802 * Remove the given range of addresses from the specified map.
1804 * It is assumed that the start and end are properly
1805 * rounded to the page size.
1807 * This function may not be called from an interrupt if the pmap is
1811 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1815 vm_offset_t ptpaddr;
1816 vm_offset_t sindex, eindex;
1817 struct pmap_inval_info info;
1822 if (pmap->pm_stats.resident_count == 0)
1825 pmap_inval_init(&info);
1828 * special handling of removing one page. a very
1829 * common operation and easy to short circuit some
1832 if (((sva + PAGE_SIZE) == eva) &&
1833 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1834 pmap_remove_page(pmap, sva, &info);
1835 pmap_inval_flush(&info);
1840 * Get a local virtual address for the mappings that are being
1843 sindex = i386_btop(sva);
1844 eindex = i386_btop(eva);
1846 for (; sindex < eindex; sindex = pdnxt) {
1850 * Calculate index for next page table.
1852 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1853 if (pmap->pm_stats.resident_count == 0)
1856 pdirindex = sindex / NPDEPG;
1857 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1858 pmap_inval_add(&info, pmap, -1);
1859 pmap->pm_pdir[pdirindex] = 0;
1860 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1861 pmap->pm_cached = 0;
1866 * Weed out invalid mappings. Note: we assume that the page
1867 * directory table is always allocated, and in kernel virtual.
1873 * Limit our scan to either the end of the va represented
1874 * by the current page table page, or to the end of the
1875 * range being removed.
1877 if (pdnxt > eindex) {
1882 * NOTE: pmap_remove_pte() can block and wipe the temporary
1885 for (; sindex != pdnxt; sindex++) {
1888 ptbase = get_ptbase(pmap);
1889 if (ptbase[sindex] == 0)
1891 va = i386_ptob(sindex);
1892 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1896 pmap_inval_flush(&info);
1902 * Removes this physical page from all physical maps in which it resides.
1903 * Reflects back modify bits to the pager.
1905 * This routine may not be called from an interrupt.
1909 pmap_remove_all(vm_page_t m)
1911 struct pmap_inval_info info;
1912 unsigned *pte, tpte;
1915 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
1918 pmap_inval_init(&info);
1920 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1921 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1922 --pv->pv_pmap->pm_stats.resident_count;
1924 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1925 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1926 tpte = loadandclear(pte);
1928 KKASSERT(PHYS_TO_VM_PAGE(tpte) == m);
1931 pv->pv_pmap->pm_stats.wired_count--;
1934 vm_page_flag_set(m, PG_REFERENCED);
1937 * Update the vm_page_t clean and reference bits.
1940 #if defined(PMAP_DIAGNOSTIC)
1941 if (pmap_nw_modified((pt_entry_t) tpte)) {
1942 kprintf("pmap_remove_all: modified page "
1943 "not writable: va: %p, pte: 0x%lx\n",
1944 (void *)pv->pv_va, (long)tpte);
1947 if (pmap_track_modified(pv->pv_va))
1951 KKASSERT(pv->pv_m == m);
1953 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1954 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1955 ++pv->pv_pmap->pm_generation;
1956 m->md.pv_list_count--;
1957 if (TAILQ_EMPTY(&m->md.pv_list))
1958 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1959 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1963 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
1964 pmap_inval_flush(&info);
1970 * Set the physical protection on the specified range of this map
1973 * This function may not be called from an interrupt if the map is
1974 * not the kernel_pmap.
1977 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1980 vm_offset_t pdnxt, ptpaddr;
1981 vm_pindex_t sindex, eindex;
1982 pmap_inval_info info;
1987 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1988 pmap_remove(pmap, sva, eva);
1992 if (prot & VM_PROT_WRITE)
1995 pmap_inval_init(&info);
1997 ptbase = get_ptbase(pmap);
1999 sindex = i386_btop(sva);
2000 eindex = i386_btop(eva);
2002 for (; sindex < eindex; sindex = pdnxt) {
2006 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
2008 pdirindex = sindex / NPDEPG;
2009 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
2010 pmap_inval_add(&info, pmap, -1);
2011 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
2012 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2017 * Weed out invalid mappings. Note: we assume that the page
2018 * directory table is always allocated, and in kernel virtual.
2023 if (pdnxt > eindex) {
2027 for (; sindex != pdnxt; sindex++) {
2033 * XXX non-optimal. Note also that there can be
2034 * no pmap_inval_flush() calls until after we modify
2035 * ptbase[sindex] (or otherwise we have to do another
2036 * pmap_inval_add() call).
2038 pmap_inval_add(&info, pmap, i386_ptob(sindex));
2039 pbits = ptbase[sindex];
2041 if (pbits & PG_MANAGED) {
2044 m = PHYS_TO_VM_PAGE(pbits);
2045 vm_page_flag_set(m, PG_REFERENCED);
2049 if (pmap_track_modified(i386_ptob(sindex))) {
2051 m = PHYS_TO_VM_PAGE(pbits);
2060 if (pbits != ptbase[sindex]) {
2061 ptbase[sindex] = pbits;
2065 pmap_inval_flush(&info);
2069 * Insert the given physical page (p) at
2070 * the specified virtual address (v) in the
2071 * target physical map with the protection requested.
2073 * If specified, the page will be wired down, meaning
2074 * that the related pte can not be reclaimed.
2076 * NB: This is the only routine which MAY NOT lazy-evaluate
2077 * or lose information. That is, this routine must actually
2078 * insert this page into the given map NOW.
2081 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2087 vm_offset_t origpte, newpte;
2089 pmap_inval_info info;
2095 #ifdef PMAP_DIAGNOSTIC
2097 panic("pmap_enter: toobig");
2098 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) {
2099 panic("pmap_enter: invalid to pmap_enter page "
2100 "table pages (va: %p)", (void *)va);
2103 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2104 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2107 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2108 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2113 * In the case that a page table page is not
2114 * resident, we are creating it here.
2116 if (va < UPT_MIN_ADDRESS)
2117 mpte = pmap_allocpte(pmap, va);
2121 pmap_inval_init(&info);
2122 pte = pmap_pte(pmap, va);
2125 * Page Directory table entry not valid, we need a new PT page
2128 panic("pmap_enter: invalid page directory pdir=0x%lx, va=%p\n",
2129 (long)pmap->pm_pdir[PTDPTDI], (void *)va);
2132 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
2133 origpte = *(vm_offset_t *)pte;
2134 opa = origpte & PG_FRAME;
2136 if (origpte & PG_PS)
2137 panic("pmap_enter: attempted pmap_enter on 4MB page");
2140 * Mapping has not changed, must be protection or wiring change.
2142 if (origpte && (opa == pa)) {
2144 * Wiring change, just update stats. We don't worry about
2145 * wiring PT pages as they remain resident as long as there
2146 * are valid mappings in them. Hence, if a user page is wired,
2147 * the PT page will be also.
2149 if (wired && ((origpte & PG_W) == 0))
2150 pmap->pm_stats.wired_count++;
2151 else if (!wired && (origpte & PG_W))
2152 pmap->pm_stats.wired_count--;
2154 #if defined(PMAP_DIAGNOSTIC)
2155 if (pmap_nw_modified((pt_entry_t) origpte)) {
2156 kprintf("pmap_enter: modified page not "
2157 "writable: va: %p, pte: 0x%lx\n",
2158 (void *)va, (long )origpte);
2163 * Remove the extra pte reference. Note that we cannot
2164 * optimize the RO->RW case because we have adjusted the
2165 * wiring count above and may need to adjust the wiring
2172 * We might be turning off write access to the page,
2173 * so we go ahead and sense modify status.
2175 if (origpte & PG_MANAGED) {
2176 if ((origpte & PG_M) && pmap_track_modified(va)) {
2178 om = PHYS_TO_VM_PAGE(opa);
2182 KKASSERT(m->flags & PG_MAPPED);
2187 * Mapping has changed, invalidate old range and fall through to
2188 * handle validating new mapping.
2190 * Since we have a ref on the page directory page pmap_pte()
2191 * will always return non-NULL.
2193 * NOTE: pmap_remove_pte() can block and cause the temporary ptbase
2194 * to get wiped. reload the ptbase. I'm not sure if it is
2195 * also possible to race another pmap_enter() but check for
2201 KKASSERT((origpte & PG_FRAME) ==
2202 (*(vm_offset_t *)pte & PG_FRAME));
2203 err = pmap_remove_pte(pmap, pte, va, &info);
2205 panic("pmap_enter: pte vanished, va: %p", (void *)va);
2206 pte = pmap_pte(pmap, va);
2207 origpte = *(vm_offset_t *)pte;
2208 opa = origpte & PG_FRAME;
2210 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2216 * Enter on the PV list if part of our managed memory. Note that we
2217 * raise IPL while manipulating pv_table since pmap_enter can be
2218 * called at interrupt time.
2220 if (pmap_initialized &&
2221 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2222 pmap_insert_entry(pmap, va, mpte, m);
2223 ptbase_assert(pmap);
2225 vm_page_flag_set(m, PG_MAPPED);
2229 * Increment counters
2231 ++pmap->pm_stats.resident_count;
2233 pmap->pm_stats.wired_count++;
2234 KKASSERT(*pte == 0);
2238 * Now validate mapping with desired protection/wiring.
2240 ptbase_assert(pmap);
2241 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2245 if (va < UPT_MIN_ADDRESS)
2247 if (pmap == &kernel_pmap)
2251 * if the mapping or permission bits are different, we need
2252 * to update the pte.
2254 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2255 pmap_inval_add(&info, pmap, va);
2256 ptbase_assert(pmap);
2257 KKASSERT(*pte == 0 ||
2258 (*pte & PG_FRAME) == (newpte & PG_FRAME));
2259 *pte = newpte | PG_A;
2261 vm_page_flag_set(m, PG_WRITEABLE);
2263 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2264 pmap_inval_flush(&info);
2268 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2269 * This code also assumes that the pmap has no pre-existing entry for this
2272 * This code currently may only be used on user pmaps, not kernel_pmap.
2275 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2282 pmap_inval_info info;
2284 pmap_inval_init(&info);
2286 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2287 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2290 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2291 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2295 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2298 * Calculate the page table page (mpte), allocating it if necessary.
2300 * A held page table page (mpte), or NULL, is passed onto the
2301 * section following.
2303 if (va < UPT_MIN_ADDRESS) {
2305 * Calculate pagetable page index
2307 ptepindex = va >> PDRSHIFT;
2311 * Get the page directory entry
2313 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2316 * If the page table page is mapped, we just increment
2317 * the hold count, and activate it.
2321 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2322 if (pmap->pm_ptphint &&
2323 (pmap->pm_ptphint->pindex == ptepindex)) {
2324 mpte = pmap->pm_ptphint;
2326 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2327 pmap->pm_ptphint = mpte;
2332 mpte = _pmap_allocpte(pmap, ptepindex);
2334 } while (mpte == NULL);
2337 /* this code path is not yet used */
2341 * With a valid (and held) page directory page, we can just use
2342 * vtopte() to get to the pte. If the pte is already present
2343 * we do not disturb it.
2345 pte = (unsigned *)vtopte(va);
2348 pmap_unwire_pte_hold(pmap, mpte, &info);
2349 pa = VM_PAGE_TO_PHYS(m);
2350 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2355 * Enter on the PV list if part of our managed memory
2357 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2358 pmap_insert_entry(pmap, va, mpte, m);
2359 vm_page_flag_set(m, PG_MAPPED);
2363 * Increment counters
2365 ++pmap->pm_stats.resident_count;
2367 pa = VM_PAGE_TO_PHYS(m);
2370 * Now validate mapping with RO protection
2372 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2373 *pte = pa | PG_V | PG_U;
2375 *pte = pa | PG_V | PG_U | PG_MANAGED;
2376 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2377 pmap_inval_flush(&info);
2381 * Make a temporary mapping for a physical address. This is only intended
2382 * to be used for panic dumps.
2385 pmap_kenter_temporary(vm_paddr_t pa, int i)
2387 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2388 return ((void *)crashdumpmap);
2391 #define MAX_INIT_PT (96)
2394 * This routine preloads the ptes for a given object into the specified pmap.
2395 * This eliminates the blast of soft faults on process startup and
2396 * immediately after an mmap.
2398 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2401 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2402 vm_object_t object, vm_pindex_t pindex,
2403 vm_size_t size, int limit)
2405 struct rb_vm_page_scan_info info;
2410 * We can't preinit if read access isn't set or there is no pmap
2413 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2417 * We can't preinit if the pmap is not the current pmap
2419 lp = curthread->td_lwp;
2420 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2423 psize = i386_btop(size);
2425 if ((object->type != OBJT_VNODE) ||
2426 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2427 (object->resident_page_count > MAX_INIT_PT))) {
2431 if (psize + pindex > object->size) {
2432 if (object->size < pindex)
2434 psize = object->size - pindex;
2441 * Use a red-black scan to traverse the requested range and load
2442 * any valid pages found into the pmap.
2444 * We cannot safely scan the object's memq unless we are in a
2445 * critical section since interrupts can remove pages from objects.
2447 info.start_pindex = pindex;
2448 info.end_pindex = pindex + psize - 1;
2455 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2456 pmap_object_init_pt_callback, &info);
2462 pmap_object_init_pt_callback(vm_page_t p, void *data)
2464 struct rb_vm_page_scan_info *info = data;
2465 vm_pindex_t rel_index;
2467 * don't allow an madvise to blow away our really
2468 * free pages allocating pv entries.
2470 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2471 vmstats.v_free_count < vmstats.v_free_reserved) {
2474 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2475 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2476 if ((p->queue - p->pc) == PQ_CACHE)
2477 vm_page_deactivate(p);
2479 rel_index = p->pindex - info->start_pindex;
2480 pmap_enter_quick(info->pmap,
2481 info->addr + i386_ptob(rel_index), p);
2488 * Return TRUE if the pmap is in shape to trivially
2489 * pre-fault the specified address.
2491 * Returns FALSE if it would be non-trivial or if a
2492 * pte is already loaded into the slot.
2495 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2499 if ((*pmap_pde(pmap, addr)) == 0)
2501 pte = (unsigned *) vtopte(addr);
2508 * Routine: pmap_change_wiring
2509 * Function: Change the wiring attribute for a map/virtual-address
2511 * In/out conditions:
2512 * The mapping must already exist in the pmap.
2515 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2522 pte = pmap_pte(pmap, va);
2524 if (wired && !pmap_pte_w(pte))
2525 pmap->pm_stats.wired_count++;
2526 else if (!wired && pmap_pte_w(pte))
2527 pmap->pm_stats.wired_count--;
2530 * Wiring is not a hardware characteristic so there is no need to
2531 * invalidate TLB. However, in an SMP environment we must use
2532 * a locked bus cycle to update the pte (if we are not using
2533 * the pmap_inval_*() API that is)... it's ok to do this for simple
2538 atomic_set_int(pte, PG_W);
2540 atomic_clear_int(pte, PG_W);
2543 atomic_set_int_nonlocked(pte, PG_W);
2545 atomic_clear_int_nonlocked(pte, PG_W);
2552 * Copy the range specified by src_addr/len
2553 * from the source map to the range dst_addr/len
2554 * in the destination map.
2556 * This routine is only advisory and need not do anything.
2559 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2560 vm_size_t len, vm_offset_t src_addr)
2563 pmap_inval_info info;
2565 vm_offset_t end_addr = src_addr + len;
2567 unsigned src_frame, dst_frame;
2571 if (dst_addr != src_addr)
2574 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2575 * valid through blocking calls, and that's just not going to
2582 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2583 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2587 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2588 if ((*gd->gd_GDMAP1 & PG_FRAME) != dst_frame) {
2589 *gd->gd_GDMAP1 = dst_frame | PG_RW | PG_V;
2593 pmap_inval_init(&info);
2594 pmap_inval_add(&info, dst_pmap, -1);
2595 pmap_inval_add(&info, src_pmap, -1);
2598 * critical section protection is required to maintain the page/object
2599 * association, interrupts can free pages and remove them from
2603 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2604 unsigned *src_pte, *dst_pte;
2605 vm_page_t dstmpte, srcmpte;
2606 vm_offset_t srcptepaddr;
2609 if (addr >= UPT_MIN_ADDRESS)
2610 panic("pmap_copy: invalid to pmap_copy page tables\n");
2613 * Don't let optional prefaulting of pages make us go
2614 * way below the low water mark of free pages or way
2615 * above high water mark of used pv entries.
2617 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2618 pv_entry_count > pv_entry_high_water)
2621 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2622 ptepindex = addr >> PDRSHIFT;
2624 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2625 if (srcptepaddr == 0)
2628 if (srcptepaddr & PG_PS) {
2629 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2630 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2631 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2637 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2638 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2639 (srcmpte->flags & PG_BUSY)) {
2643 if (pdnxt > end_addr)
2646 src_pte = (unsigned *) vtopte(addr);
2647 dst_pte = (unsigned *) avtopte(addr);
2648 while (addr < pdnxt) {
2653 * we only virtual copy managed pages
2655 if ((ptetemp & PG_MANAGED) != 0) {
2657 * We have to check after allocpte for the
2658 * pte still being around... allocpte can
2661 * pmap_allocpte() can block. If we lose
2662 * our page directory mappings we stop.
2664 dstmpte = pmap_allocpte(dst_pmap, addr);
2666 if (src_frame != (((unsigned) PTDpde) & PG_FRAME) ||
2667 XXX dst_frame != (((unsigned) xxx) & PG_FRAME)
2669 kprintf("WARNING: pmap_copy: detected and corrected race\n");
2670 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2672 } else if ((*dst_pte == 0) &&
2673 (ptetemp = *src_pte) != 0 &&
2674 (ptetemp & PG_MANAGED)) {
2676 * Clear the modified and
2677 * accessed (referenced) bits
2680 m = PHYS_TO_VM_PAGE(ptetemp);
2681 *dst_pte = ptetemp & ~(PG_M | PG_A);
2682 ++dst_pmap->pm_stats.resident_count;
2683 pmap_insert_entry(dst_pmap, addr,
2685 KKASSERT(m->flags & PG_MAPPED);
2687 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
2688 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2691 if (dstmpte->hold_count >= srcmpte->hold_count)
2701 pmap_inval_flush(&info);
2708 * Zero the specified PA by mapping the page into KVM and clearing its
2711 * This function may be called from an interrupt and no locking is
2715 pmap_zero_page(vm_paddr_t phys)
2717 struct mdglobaldata *gd = mdcpu;
2720 if (*(int *)gd->gd_CMAP3)
2721 panic("pmap_zero_page: CMAP3 busy");
2722 *(int *)gd->gd_CMAP3 =
2723 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2724 cpu_invlpg(gd->gd_CADDR3);
2726 #if defined(I686_CPU)
2727 if (cpu_class == CPUCLASS_686)
2728 i686_pagezero(gd->gd_CADDR3);
2731 bzero(gd->gd_CADDR3, PAGE_SIZE);
2732 *(int *) gd->gd_CMAP3 = 0;
2737 * pmap_page_assertzero:
2739 * Assert that a page is empty, panic if it isn't.
2742 pmap_page_assertzero(vm_paddr_t phys)
2744 struct mdglobaldata *gd = mdcpu;
2748 if (*(int *)gd->gd_CMAP3)
2749 panic("pmap_zero_page: CMAP3 busy");
2750 *(int *)gd->gd_CMAP3 =
2751 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2752 cpu_invlpg(gd->gd_CADDR3);
2753 for (i = 0; i < PAGE_SIZE; i += 4) {
2754 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2755 panic("pmap_page_assertzero() @ %p not zero!\n",
2756 (void *)gd->gd_CADDR3);
2759 *(int *) gd->gd_CMAP3 = 0;
2766 * Zero part of a physical page by mapping it into memory and clearing
2767 * its contents with bzero.
2769 * off and size may not cover an area beyond a single hardware page.
2772 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2774 struct mdglobaldata *gd = mdcpu;
2777 if (*(int *) gd->gd_CMAP3)
2778 panic("pmap_zero_page: CMAP3 busy");
2779 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2780 cpu_invlpg(gd->gd_CADDR3);
2782 #if defined(I686_CPU)
2783 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2784 i686_pagezero(gd->gd_CADDR3);
2787 bzero((char *)gd->gd_CADDR3 + off, size);
2788 *(int *) gd->gd_CMAP3 = 0;
2795 * Copy the physical page from the source PA to the target PA.
2796 * This function may be called from an interrupt. No locking
2800 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2802 struct mdglobaldata *gd = mdcpu;
2805 if (*(int *) gd->gd_CMAP1)
2806 panic("pmap_copy_page: CMAP1 busy");
2807 if (*(int *) gd->gd_CMAP2)
2808 panic("pmap_copy_page: CMAP2 busy");
2810 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2811 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2813 cpu_invlpg(gd->gd_CADDR1);
2814 cpu_invlpg(gd->gd_CADDR2);
2816 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2818 *(int *) gd->gd_CMAP1 = 0;
2819 *(int *) gd->gd_CMAP2 = 0;
2824 * pmap_copy_page_frag:
2826 * Copy the physical page from the source PA to the target PA.
2827 * This function may be called from an interrupt. No locking
2831 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2833 struct mdglobaldata *gd = mdcpu;
2836 if (*(int *) gd->gd_CMAP1)
2837 panic("pmap_copy_page: CMAP1 busy");
2838 if (*(int *) gd->gd_CMAP2)
2839 panic("pmap_copy_page: CMAP2 busy");
2841 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2842 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2844 cpu_invlpg(gd->gd_CADDR1);
2845 cpu_invlpg(gd->gd_CADDR2);
2847 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2848 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2851 *(int *) gd->gd_CMAP1 = 0;
2852 *(int *) gd->gd_CMAP2 = 0;
2857 * Returns true if the pmap's pv is one of the first
2858 * 16 pvs linked to from this page. This count may
2859 * be changed upwards or downwards in the future; it
2860 * is only necessary that true be returned for a small
2861 * subset of pmaps for proper page aging.
2864 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2869 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2874 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2875 if (pv->pv_pmap == pmap) {
2888 * Remove all pages from specified address space
2889 * this aids process exit speeds. Also, this code
2890 * is special cased for current process only, but
2891 * can have the more generic (and slightly slower)
2892 * mode enabled. This is much faster than pmap_remove
2893 * in the case of running down an entire address space.
2896 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2899 unsigned *pte, tpte;
2902 pmap_inval_info info;
2904 int32_t save_generation;
2906 lp = curthread->td_lwp;
2907 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2912 pmap_inval_init(&info);
2914 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2915 if (pv->pv_va >= eva || pv->pv_va < sva) {
2916 npv = TAILQ_NEXT(pv, pv_plist);
2920 KKASSERT(pmap == pv->pv_pmap);
2923 pte = (unsigned *)vtopte(pv->pv_va);
2925 pte = pmap_pte_quick(pmap, pv->pv_va);
2927 if (pmap->pm_active)
2928 pmap_inval_add(&info, pmap, pv->pv_va);
2931 * We cannot remove wired pages from a process' mapping
2935 npv = TAILQ_NEXT(pv, pv_plist);
2939 tpte = loadandclear(pte);
2941 m = PHYS_TO_VM_PAGE(tpte);
2942 test_m_maps_pv(m, pv);
2944 KASSERT(m < &vm_page_array[vm_page_array_size],
2945 ("pmap_remove_pages: bad tpte %x", tpte));
2947 KKASSERT(pmap->pm_stats.resident_count > 0);
2948 --pmap->pm_stats.resident_count;
2951 * Update the vm_page_t clean and reference bits.
2957 npv = TAILQ_NEXT(pv, pv_plist);
2959 KKASSERT(pv->pv_m == m);
2960 KKASSERT(pv->pv_pmap == pmap);
2962 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2963 save_generation = ++pmap->pm_generation;
2965 m->md.pv_list_count--;
2966 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2967 if (TAILQ_EMPTY(&m->md.pv_list))
2968 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2970 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2974 * Restart the scan if we blocked during the unuse or free
2975 * calls and other removals were made.
2977 if (save_generation != pmap->pm_generation) {
2978 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2979 npv = TAILQ_FIRST(&pmap->pm_pvlist);
2982 pmap_inval_flush(&info);
2987 * pmap_testbit tests bits in pte's
2988 * note that the testbit/clearbit routines are inline,
2989 * and a lot of things compile-time evaluate.
2992 pmap_testbit(vm_page_t m, int bit)
2997 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3000 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3005 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3007 * if the bit being tested is the modified bit, then
3008 * mark clean_map and ptes as never
3011 if (bit & (PG_A|PG_M)) {
3012 if (!pmap_track_modified(pv->pv_va))
3016 #if defined(PMAP_DIAGNOSTIC)
3018 kprintf("Null pmap (tb) at va: %p\n",
3023 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3034 * this routine is used to modify bits in ptes
3036 static __inline void
3037 pmap_clearbit(vm_page_t m, int bit)
3039 struct pmap_inval_info info;
3044 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3047 pmap_inval_init(&info);
3051 * Loop over all current mappings setting/clearing as appropos If
3052 * setting RO do we need to clear the VAC?
3054 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3056 * don't write protect pager mappings
3059 if (!pmap_track_modified(pv->pv_va))
3063 #if defined(PMAP_DIAGNOSTIC)
3065 kprintf("Null pmap (cb) at va: %p\n",
3072 * Careful here. We can use a locked bus instruction to
3073 * clear PG_A or PG_M safely but we need to synchronize
3074 * with the target cpus when we mess with PG_RW.
3076 * We do not have to force synchronization when clearing
3077 * PG_M even for PTEs generated via virtual memory maps,
3078 * because the virtual kernel will invalidate the pmap
3079 * entry when/if it needs to resynchronize the Modify bit.
3082 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
3083 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3090 atomic_clear_int(pte, PG_M|PG_RW);
3093 * The cpu may be trying to set PG_M
3094 * simultaniously with our clearing
3097 if (!atomic_cmpset_int(pte, pbits,
3101 } else if (bit == PG_M) {
3103 * We could also clear PG_RW here to force
3104 * a fault on write to redetect PG_M for
3105 * virtual kernels, but it isn't necessary
3106 * since virtual kernels invalidate the pte
3107 * when they clear the VPTE_M bit in their
3108 * virtual page tables.
3110 atomic_clear_int(pte, PG_M);
3112 atomic_clear_int(pte, bit);
3116 pmap_inval_flush(&info);
3121 * pmap_page_protect:
3123 * Lower the permission for all mappings to a given page.
3126 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3128 if ((prot & VM_PROT_WRITE) == 0) {
3129 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3130 pmap_clearbit(m, PG_RW);
3131 vm_page_flag_clear(m, PG_WRITEABLE);
3139 pmap_phys_address(vm_pindex_t ppn)
3141 return (i386_ptob(ppn));
3145 * pmap_ts_referenced:
3147 * Return a count of reference bits for a page, clearing those bits.
3148 * It is not necessary for every reference bit to be cleared, but it
3149 * is necessary that 0 only be returned when there are truly no
3150 * reference bits set.
3152 * XXX: The exact number of bits to check and clear is a matter that
3153 * should be tested and standardized at some point in the future for
3154 * optimal aging of shared pages.
3157 pmap_ts_referenced(vm_page_t m)
3159 pv_entry_t pv, pvf, pvn;
3163 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3168 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3173 pvn = TAILQ_NEXT(pv, pv_list);
3176 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3177 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3180 if (!pmap_track_modified(pv->pv_va))
3183 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3185 if (pte && (*pte & PG_A)) {
3187 atomic_clear_int(pte, PG_A);
3189 atomic_clear_int_nonlocked(pte, PG_A);
3196 } while ((pv = pvn) != NULL && pv != pvf);
3206 * Return whether or not the specified physical page was modified
3207 * in any physical maps.
3210 pmap_is_modified(vm_page_t m)
3212 return pmap_testbit(m, PG_M);
3216 * Clear the modify bits on the specified physical page.
3219 pmap_clear_modify(vm_page_t m)
3221 pmap_clearbit(m, PG_M);
3225 * pmap_clear_reference:
3227 * Clear the reference bit on the specified physical page.
3230 pmap_clear_reference(vm_page_t m)
3232 pmap_clearbit(m, PG_A);
3236 * Miscellaneous support routines follow
3240 i386_protection_init(void)
3244 kp = protection_codes;
3245 for (prot = 0; prot < 8; prot++) {
3247 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3249 * Read access is also 0. There isn't any execute bit,
3250 * so just make it readable.
3252 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3253 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3254 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3257 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3258 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3259 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3260 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3268 * Map a set of physical memory pages into the kernel virtual
3269 * address space. Return a pointer to where it is mapped. This
3270 * routine is intended to be used for mapping device memory,
3273 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3277 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3279 vm_offset_t va, tmpva, offset;
3282 offset = pa & PAGE_MASK;
3283 size = roundup(offset + size, PAGE_SIZE);
3285 va = kmem_alloc_nofault(&kernel_map, size);
3287 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3290 for (tmpva = va; size > 0;) {
3291 pte = (unsigned *)vtopte(tmpva);
3292 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3300 return ((void *)(va + offset));
3304 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3306 vm_offset_t base, offset;
3308 base = va & PG_FRAME;
3309 offset = va & PAGE_MASK;
3310 size = roundup(offset + size, PAGE_SIZE);
3311 pmap_qremove(va, size >> PAGE_SHIFT);
3312 kmem_free(&kernel_map, base, size);
3316 * perform the pmap work for mincore
3319 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3321 unsigned *ptep, pte;
3325 ptep = pmap_pte(pmap, addr);
3330 if ((pte = *ptep) != 0) {
3333 val = MINCORE_INCORE;
3334 if ((pte & PG_MANAGED) == 0)
3337 pa = pte & PG_FRAME;
3339 m = PHYS_TO_VM_PAGE(pa);
3345 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3347 * Modified by someone
3349 else if (m->dirty || pmap_is_modified(m))
3350 val |= MINCORE_MODIFIED_OTHER;
3355 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3358 * Referenced by someone
3360 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3361 val |= MINCORE_REFERENCED_OTHER;
3362 vm_page_flag_set(m, PG_REFERENCED);
3369 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3370 * vmspace will be ref'd and the old one will be deref'd.
3372 * The vmspace for all lwps associated with the process will be adjusted
3373 * and cr3 will be reloaded if any lwp is the current lwp.
3376 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3378 struct vmspace *oldvm;
3382 oldvm = p->p_vmspace;
3383 if (oldvm != newvm) {
3384 p->p_vmspace = newvm;
3385 KKASSERT(p->p_nthreads == 1);
3386 lp = RB_ROOT(&p->p_lwp_tree);
3387 pmap_setlwpvm(lp, newvm);
3389 sysref_get(&newvm->vm_sysref);
3390 sysref_put(&oldvm->vm_sysref);
3397 * Set the vmspace for a LWP. The vmspace is almost universally set the
3398 * same as the process vmspace, but virtual kernels need to swap out contexts
3399 * on a per-lwp basis.
3402 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3404 struct vmspace *oldvm;
3408 oldvm = lp->lwp_vmspace;
3410 if (oldvm != newvm) {
3411 lp->lwp_vmspace = newvm;
3412 if (curthread->td_lwp == lp) {
3413 pmap = vmspace_pmap(newvm);
3415 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3417 pmap->pm_active |= 1;
3419 #if defined(SWTCH_OPTIM_STATS)
3422 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3423 load_cr3(curthread->td_pcb->pcb_cr3);
3424 pmap = vmspace_pmap(oldvm);
3426 atomic_clear_int(&pmap->pm_active,
3427 1 << mycpu->gd_cpuid);
3429 pmap->pm_active &= ~1;
3437 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3440 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3444 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3449 pmap_get_pgeflag(void)
3456 static void pads (pmap_t pm);
3457 void pmap_pvdump (vm_paddr_t pa);
3459 /* print address space of pmap*/
3466 if (pm == &kernel_pmap)
3469 for (i = 0; i < 1024; i++) {
3470 if (pm->pm_pdir[i]) {
3471 for (j = 0; j < 1024; j++) {
3472 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3473 if (pm == &kernel_pmap && va < KERNBASE)
3475 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3477 ptep = pmap_pte_quick(pm, va);
3478 if (pmap_pte_v(ptep))
3479 kprintf("%x:%x ", va, *(int *) ptep);
3488 pmap_pvdump(vm_paddr_t pa)
3493 kprintf("pa %08llx", (long long)pa);
3494 m = PHYS_TO_VM_PAGE(pa);
3495 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3497 kprintf(" -> pmap %p, va %p, flags %x",
3498 (void *)pv->pv_pmap, (long)pv->pv_va, pv->pv_flags);
3500 kprintf(" -> pmap %p, va %p",
3501 (void *)pv->pv_pmap, (void *)pv->pv_va);