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 1.82 2008/01/21 10:28:04 corecode Exp $
47 * Manages physical address maps.
49 * In addition to hardware address maps, this
50 * module is called upon to provide software-use-only
51 * maps which may or may not be stored in the same
52 * form as hardware maps. These pseudo-maps are
53 * used to store intermediate results from copy
54 * operations to and from address spaces.
56 * Since the information managed by this module is
57 * also stored by the logical address mapping module,
58 * this module may throw away valid virtual-to-physical
59 * mappings at almost any time. However, invalidations
60 * of virtual-to-physical mappings must be done as
63 * In order to cope with hardware architectures which
64 * make virtual-to-physical map invalidates expensive,
65 * this module may delay invalidate or reduced protection
66 * operations until such time as they are actually
67 * necessary. This module is given full information as
68 * to which processors are currently using which maps,
69 * and to when physical maps must be made correct.
72 #include "opt_disable_pse.h"
74 #include "opt_msgbuf.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
80 #include <sys/msgbuf.h>
81 #include <sys/vmmeter.h>
85 #include <vm/vm_param.h>
86 #include <sys/sysctl.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_extern.h>
93 #include <vm/vm_pageout.h>
94 #include <vm/vm_pager.h>
95 #include <vm/vm_zone.h>
98 #include <sys/thread2.h>
99 #include <sys/sysref2.h>
101 #include <machine/cputypes.h>
102 #include <machine/md_var.h>
103 #include <machine/specialreg.h>
104 #include <machine/smp.h>
105 #include <machine_base/apic/apicreg.h>
106 #include <machine/globaldata.h>
107 #include <machine/pmap.h>
108 #include <machine/pmap_inval.h>
112 #define PMAP_KEEP_PDIRS
113 #ifndef PMAP_SHPGPERPROC
114 #define PMAP_SHPGPERPROC 200
117 #if defined(DIAGNOSTIC)
118 #define PMAP_DIAGNOSTIC
123 #if !defined(PMAP_DIAGNOSTIC)
124 #define PMAP_INLINE __inline
130 * Get PDEs and PTEs for user/kernel address space
132 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
133 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
135 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
136 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
137 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
138 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
139 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
143 * Given a map and a machine independent protection code,
144 * convert to a vax protection code.
146 #define pte_prot(m, p) \
147 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
148 static int protection_codes[8];
150 struct pmap kernel_pmap;
151 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
153 vm_paddr_t avail_start; /* PA of first available physical page */
154 vm_paddr_t avail_end; /* PA of last available physical page */
155 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
156 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
157 vm_offset_t KvaStart; /* VA start of KVA space */
158 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
159 vm_offset_t KvaSize; /* max size of kernel virtual address space */
160 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
161 static int pgeflag; /* PG_G or-in */
162 static int pseflag; /* PG_PS or-in */
164 static vm_object_t kptobj;
167 vm_offset_t kernel_vm_end;
170 * Data for the pv entry allocation mechanism
172 static vm_zone_t pvzone;
173 static struct vm_zone pvzone_store;
174 static struct vm_object pvzone_obj;
175 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
176 static int pmap_pagedaemon_waken = 0;
177 static struct pv_entry *pvinit;
180 * All those kernel PT submaps that BSD is so fond of
182 pt_entry_t *CMAP1 = 0, *ptmmap;
183 caddr_t CADDR1 = 0, ptvmmap = 0;
184 static pt_entry_t *msgbufmap;
185 struct msgbuf *msgbufp=0;
190 static pt_entry_t *pt_crashdumpmap;
191 static caddr_t crashdumpmap;
193 extern pt_entry_t *SMPpt;
195 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
196 static unsigned * get_ptbase (pmap_t pmap);
197 static pv_entry_t get_pv_entry (void);
198 static void i386_protection_init (void);
199 static __inline void pmap_clearbit (vm_page_t m, int bit);
201 static void pmap_remove_all (vm_page_t m);
202 static vm_page_t pmap_enter_quick (pmap_t pmap, vm_offset_t va,
203 vm_page_t m, vm_page_t mpte);
204 static int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
205 vm_offset_t sva, pmap_inval_info_t info);
206 static void pmap_remove_page (struct pmap *pmap,
207 vm_offset_t va, pmap_inval_info_t info);
208 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
209 vm_offset_t va, pmap_inval_info_t info);
210 static boolean_t pmap_testbit (vm_page_t m, int bit);
211 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
212 vm_page_t mpte, vm_page_t m);
214 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
216 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
217 static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
218 static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
219 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
220 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
221 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
223 static unsigned pdir4mb;
226 * Move the kernel virtual free pointer to the next
227 * 4MB. This is used to help improve performance
228 * by using a large (4MB) page for much of the kernel
229 * (.text, .data, .bss)
232 pmap_kmem_choose(vm_offset_t addr)
234 vm_offset_t newaddr = addr;
236 if (cpu_feature & CPUID_PSE) {
237 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
246 * Extract the page table entry associated with the given map/virtual
249 * This function may NOT be called from an interrupt.
251 PMAP_INLINE unsigned *
252 pmap_pte(pmap_t pmap, vm_offset_t va)
257 pdeaddr = (unsigned *) pmap_pde(pmap, va);
258 if (*pdeaddr & PG_PS)
261 return get_ptbase(pmap) + i386_btop(va);
270 * Super fast pmap_pte routine best used when scanning the pv lists.
271 * This eliminates many course-grained invltlb calls. Note that many of
272 * the pv list scans are across different pmaps and it is very wasteful
273 * to do an entire invltlb when checking a single mapping.
275 * Should only be called while in a critical section.
278 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
280 struct mdglobaldata *gd = mdcpu;
283 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
284 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
285 unsigned index = i386_btop(va);
286 /* are we current address space or kernel? */
287 if ((pmap == &kernel_pmap) ||
288 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
289 return (unsigned *) PTmap + index;
291 newpf = pde & PG_FRAME;
292 if ( ((* (unsigned *) gd->gd_PMAP1) & PG_FRAME) != newpf) {
293 * (unsigned *) gd->gd_PMAP1 = newpf | PG_RW | PG_V;
294 cpu_invlpg(gd->gd_PADDR1);
296 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
303 * Bootstrap the system enough to run with virtual memory.
305 * On the i386 this is called after mapping has already been enabled
306 * and just syncs the pmap module with what has already been done.
307 * [We can't call it easily with mapping off since the kernel is not
308 * mapped with PA == VA, hence we would have to relocate every address
309 * from the linked base (virtual) address "KERNBASE" to the actual
310 * (physical) address starting relative to 0]
313 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
317 struct mdglobaldata *gd;
321 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
322 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
323 KvaEnd = KvaStart + KvaSize;
325 avail_start = firstaddr;
328 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
329 * too large. It should instead be correctly calculated in locore.s and
330 * not based on 'first' (which is a physical address, not a virtual
331 * address, for the start of unused physical memory). The kernel
332 * page tables are NOT double mapped and thus should not be included
333 * in this calculation.
335 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
336 virtual_start = pmap_kmem_choose(virtual_start);
337 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
340 * Initialize protection array.
342 i386_protection_init();
345 * The kernel's pmap is statically allocated so we don't have to use
346 * pmap_create, which is unlikely to work correctly at this part of
347 * the boot sequence (XXX and which no longer exists).
349 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
350 kernel_pmap.pm_count = 1;
351 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
352 TAILQ_INIT(&kernel_pmap.pm_pvlist);
356 * Reserve some special page table entries/VA space for temporary
359 #define SYSMAP(c, p, v, n) \
360 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
363 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
366 * CMAP1/CMAP2 are used for zeroing and copying pages.
368 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
373 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
376 * ptvmmap is used for reading arbitrary physical pages via
379 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
382 * msgbufp is used to map the system message buffer.
383 * XXX msgbufmap is not used.
385 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
386 atop(round_page(MSGBUF_SIZE)))
391 for (i = 0; i < NKPT; i++)
395 * PG_G is terribly broken on SMP because we IPI invltlb's in some
396 * cases rather then invl1pg. Actually, I don't even know why it
397 * works under UP because self-referential page table mappings
402 if (cpu_feature & CPUID_PGE)
407 * Initialize the 4MB page size flag
411 * The 4MB page version of the initial
412 * kernel page mapping.
416 #if !defined(DISABLE_PSE)
417 if (cpu_feature & CPUID_PSE) {
420 * Note that we have enabled PSE mode
423 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
424 ptditmp &= ~(NBPDR - 1);
425 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
430 * Enable the PSE mode. If we are SMP we can't do this
431 * now because the APs will not be able to use it when
434 load_cr4(rcr4() | CR4_PSE);
437 * We can do the mapping here for the single processor
438 * case. We simply ignore the old page table page from
442 * For SMP, we still need 4K pages to bootstrap APs,
443 * PSE will be enabled as soon as all APs are up.
445 PTD[KPTDI] = (pd_entry_t)ptditmp;
446 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
452 if (cpu_apic_address == 0)
453 panic("pmap_bootstrap: no local apic!");
455 /* local apic is mapped on last page */
456 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag |
457 (cpu_apic_address & PG_FRAME));
461 * We need to finish setting up the globaldata page for the BSP.
462 * locore has already populated the page table for the mdglobaldata
465 pg = MDGLOBALDATA_BASEALLOC_PAGES;
466 gd = &CPU_prvspace[0].mdglobaldata;
467 gd->gd_CMAP1 = &SMPpt[pg + 0];
468 gd->gd_CMAP2 = &SMPpt[pg + 1];
469 gd->gd_CMAP3 = &SMPpt[pg + 2];
470 gd->gd_PMAP1 = &SMPpt[pg + 3];
471 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
472 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
473 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
474 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
481 * Set 4mb pdir for mp startup
486 if (pseflag && (cpu_feature & CPUID_PSE)) {
487 load_cr4(rcr4() | CR4_PSE);
488 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
489 kernel_pmap.pm_pdir[KPTDI] =
490 PTD[KPTDI] = (pd_entry_t)pdir4mb;
498 * Initialize the pmap module.
499 * Called by vm_init, to initialize any structures that the pmap
500 * system needs to map virtual memory.
501 * pmap_init has been enhanced to support in a fairly consistant
502 * way, discontiguous physical memory.
511 * object for kernel page table pages
513 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
516 * Allocate memory for random pmap data structures. Includes the
520 for(i = 0; i < vm_page_array_size; i++) {
523 m = &vm_page_array[i];
524 TAILQ_INIT(&m->md.pv_list);
525 m->md.pv_list_count = 0;
529 * init the pv free list
531 initial_pvs = vm_page_array_size;
532 if (initial_pvs < MINPV)
534 pvzone = &pvzone_store;
535 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
536 initial_pvs * sizeof (struct pv_entry));
537 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
541 * Now it is safe to enable pv_table recording.
543 pmap_initialized = TRUE;
547 * Initialize the address space (zone) for the pv_entries. Set a
548 * high water mark so that the system can recover from excessive
549 * numbers of pv entries.
554 int shpgperproc = PMAP_SHPGPERPROC;
556 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
557 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
558 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
559 pv_entry_high_water = 9 * (pv_entry_max / 10);
560 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
564 /***************************************************
565 * Low level helper routines.....
566 ***************************************************/
568 #if defined(PMAP_DIAGNOSTIC)
571 * This code checks for non-writeable/modified pages.
572 * This should be an invalid condition.
575 pmap_nw_modified(pt_entry_t ptea)
581 if ((pte & (PG_M|PG_RW)) == PG_M)
590 * this routine defines the region(s) of memory that should
591 * not be tested for the modified bit.
593 static PMAP_INLINE int
594 pmap_track_modified(vm_offset_t va)
596 if ((va < clean_sva) || (va >= clean_eva))
603 get_ptbase(pmap_t pmap)
605 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
606 struct globaldata *gd = mycpu;
608 /* are we current address space or kernel? */
609 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
610 return (unsigned *) PTmap;
613 /* otherwise, we are alternate address space */
614 KKASSERT(gd->gd_intr_nesting_level == 0 &&
615 (gd->gd_curthread->td_flags & TDF_INTTHREAD) == 0);
617 if (frame != (((unsigned) APTDpde) & PG_FRAME)) {
618 APTDpde = (pd_entry_t)(frame | PG_RW | PG_V);
619 /* The page directory is not shared between CPUs */
622 return (unsigned *) APTmap;
628 * Extract the physical page address associated with the map/VA pair.
630 * This function may not be called from an interrupt if the pmap is
634 pmap_extract(pmap_t pmap, vm_offset_t va)
637 vm_offset_t pdirindex;
639 pdirindex = va >> PDRSHIFT;
640 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
642 if ((rtval & PG_PS) != 0) {
643 rtval &= ~(NBPDR - 1);
644 rtval |= va & (NBPDR - 1);
647 pte = get_ptbase(pmap) + i386_btop(va);
648 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
654 /***************************************************
655 * Low level mapping routines.....
656 ***************************************************/
659 * Routine: pmap_kenter
661 * Add a wired page to the KVA
662 * NOTE! note that in order for the mapping to take effect -- you
663 * should do an invltlb after doing the pmap_kenter().
666 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
670 pmap_inval_info info;
672 pmap_inval_init(&info);
673 pmap_inval_add(&info, &kernel_pmap, va);
674 npte = pa | PG_RW | PG_V | pgeflag;
675 pte = (unsigned *)vtopte(va);
677 pmap_inval_flush(&info);
681 * Routine: pmap_kenter_quick
683 * Similar to pmap_kenter(), except we only invalidate the
684 * mapping on the current CPU.
687 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
692 npte = pa | PG_RW | PG_V | pgeflag;
693 pte = (unsigned *)vtopte(va);
695 cpu_invlpg((void *)va);
699 pmap_kenter_sync(vm_offset_t va)
701 pmap_inval_info info;
703 pmap_inval_init(&info);
704 pmap_inval_add(&info, &kernel_pmap, va);
705 pmap_inval_flush(&info);
709 pmap_kenter_sync_quick(vm_offset_t va)
711 cpu_invlpg((void *)va);
715 * remove a page from the kernel pagetables
718 pmap_kremove(vm_offset_t va)
721 pmap_inval_info info;
723 pmap_inval_init(&info);
724 pmap_inval_add(&info, &kernel_pmap, va);
725 pte = (unsigned *)vtopte(va);
727 pmap_inval_flush(&info);
731 pmap_kremove_quick(vm_offset_t va)
734 pte = (unsigned *)vtopte(va);
736 cpu_invlpg((void *)va);
740 * XXX these need to be recoded. They are not used in any critical path.
743 pmap_kmodify_rw(vm_offset_t va)
745 *vtopte(va) |= PG_RW;
746 cpu_invlpg((void *)va);
750 pmap_kmodify_nc(vm_offset_t va)
753 cpu_invlpg((void *)va);
757 * Used to map a range of physical addresses into kernel
758 * virtual address space.
760 * For now, VM is already on, we only need to map the
764 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
766 while (start < end) {
767 pmap_kenter(virt, start);
776 * Add a list of wired pages to the kva
777 * this routine is only used for temporary
778 * kernel mappings that do not need to have
779 * page modification or references recorded.
780 * Note that old mappings are simply written
781 * over. The page *must* be wired.
784 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
788 end_va = va + count * PAGE_SIZE;
790 while (va < end_va) {
793 pte = (unsigned *)vtopte(va);
794 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
795 cpu_invlpg((void *)va);
800 smp_invltlb(); /* XXX */
805 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
808 cpumask_t cmask = mycpu->gd_cpumask;
810 end_va = va + count * PAGE_SIZE;
812 while (va < end_va) {
817 * Install the new PTE. If the pte changed from the prior
818 * mapping we must reset the cpu mask and invalidate the page.
819 * If the pte is the same but we have not seen it on the
820 * current cpu, invlpg the existing mapping. Otherwise the
821 * entry is optimal and no invalidation is required.
823 pte = (unsigned *)vtopte(va);
824 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
825 if (*pte != pteval) {
828 cpu_invlpg((void *)va);
829 } else if ((*mask & cmask) == 0) {
830 cpu_invlpg((void *)va);
839 * this routine jerks page mappings from the
840 * kernel -- it is meant only for temporary mappings.
843 pmap_qremove(vm_offset_t va, int count)
847 end_va = va + count*PAGE_SIZE;
849 while (va < end_va) {
852 pte = (unsigned *)vtopte(va);
854 cpu_invlpg((void *)va);
863 * This routine works like vm_page_lookup() but also blocks as long as the
864 * page is busy. This routine does not busy the page it returns.
866 * Unless the caller is managing objects whos pages are in a known state,
867 * the call should be made with a critical section held so the page's object
868 * association remains valid on return.
871 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
876 m = vm_page_lookup(object, pindex);
877 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
883 * Create a new thread and optionally associate it with a (new) process.
884 * NOTE! the new thread's cpu may not equal the current cpu.
887 pmap_init_thread(thread_t td)
889 /* enforce pcb placement */
890 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
891 td->td_savefpu = &td->td_pcb->pcb_save;
892 td->td_sp = (char *)td->td_pcb - 16;
896 * This routine directly affects the fork perf for a process.
899 pmap_init_proc(struct proc *p)
904 * Dispose the UPAGES for a process that has exited.
905 * This routine directly impacts the exit perf of a process.
908 pmap_dispose_proc(struct proc *p)
910 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
913 /***************************************************
914 * Page table page management routines.....
915 ***************************************************/
918 * This routine unholds page table pages, and if the hold count
919 * drops to zero, then it decrements the wire count.
922 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
924 pmap_inval_flush(info);
925 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
928 if (m->hold_count == 0) {
930 * unmap the page table page
932 pmap_inval_add(info, pmap, -1);
933 pmap->pm_pdir[m->pindex] = 0;
934 --pmap->pm_stats.resident_count;
936 if (pmap->pm_ptphint == m)
937 pmap->pm_ptphint = NULL;
940 * If the page is finally unwired, simply free it.
943 if (m->wire_count == 0) {
946 vm_page_free_zero(m);
947 --vmstats.v_wire_count;
954 static PMAP_INLINE int
955 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
958 if (m->hold_count == 0)
959 return _pmap_unwire_pte_hold(pmap, m, info);
965 * After removing a page table entry, this routine is used to
966 * conditionally free the page, and manage the hold/wire counts.
969 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
970 pmap_inval_info_t info)
973 if (va >= UPT_MIN_ADDRESS)
977 ptepindex = (va >> PDRSHIFT);
978 if (pmap->pm_ptphint &&
979 (pmap->pm_ptphint->pindex == ptepindex)) {
980 mpte = pmap->pm_ptphint;
982 pmap_inval_flush(info);
983 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
984 pmap->pm_ptphint = mpte;
988 return pmap_unwire_pte_hold(pmap, mpte, info);
992 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
993 * it, and IdlePTD, represents the template used to update all other pmaps.
995 * On architectures where the kernel pmap is not integrated into the user
996 * process pmap, this pmap represents the process pmap, not the kernel pmap.
997 * kernel_pmap should be used to directly access the kernel_pmap.
1000 pmap_pinit0(struct pmap *pmap)
1003 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1004 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1006 pmap->pm_active = 0;
1007 pmap->pm_ptphint = NULL;
1008 TAILQ_INIT(&pmap->pm_pvlist);
1009 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1013 * Initialize a preallocated and zeroed pmap structure,
1014 * such as one in a vmspace structure.
1017 pmap_pinit(struct pmap *pmap)
1022 * No need to allocate page table space yet but we do need a valid
1023 * page directory table.
1025 if (pmap->pm_pdir == NULL) {
1027 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1031 * Allocate an object for the ptes
1033 if (pmap->pm_pteobj == NULL)
1034 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1037 * Allocate the page directory page, unless we already have
1038 * one cached. If we used the cached page the wire_count will
1039 * already be set appropriately.
1041 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1042 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1043 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1044 pmap->pm_pdirm = ptdpg;
1045 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1046 ptdpg->valid = VM_PAGE_BITS_ALL;
1047 ptdpg->wire_count = 1;
1048 ++vmstats.v_wire_count;
1049 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1051 if ((ptdpg->flags & PG_ZERO) == 0)
1052 bzero(pmap->pm_pdir, PAGE_SIZE);
1054 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1056 /* install self-referential address mapping entry */
1057 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1058 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1061 pmap->pm_active = 0;
1062 pmap->pm_ptphint = NULL;
1063 TAILQ_INIT(&pmap->pm_pvlist);
1064 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1068 * Clean up a pmap structure so it can be physically freed. This routine
1069 * is called by the vmspace dtor function. A great deal of pmap data is
1070 * left passively mapped to improve vmspace management so we have a bit
1071 * of cleanup work to do here.
1074 pmap_puninit(pmap_t pmap)
1078 KKASSERT(pmap->pm_active == 0);
1079 if ((p = pmap->pm_pdirm) != NULL) {
1080 KKASSERT(pmap->pm_pdir != NULL);
1081 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1083 vmstats.v_wire_count--;
1084 KKASSERT((p->flags & PG_BUSY) == 0);
1086 vm_page_free_zero(p);
1087 pmap->pm_pdirm = NULL;
1089 if (pmap->pm_pdir) {
1090 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1091 pmap->pm_pdir = NULL;
1093 if (pmap->pm_pteobj) {
1094 vm_object_deallocate(pmap->pm_pteobj);
1095 pmap->pm_pteobj = NULL;
1100 * Wire in kernel global address entries. To avoid a race condition
1101 * between pmap initialization and pmap_growkernel, this procedure
1102 * adds the pmap to the master list (which growkernel scans to update),
1103 * then copies the template.
1106 pmap_pinit2(struct pmap *pmap)
1109 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1110 /* XXX copies current process, does not fill in MPPTDI */
1111 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1116 * Attempt to release and free and vm_page in a pmap. Returns 1 on success,
1117 * 0 on failure (if the procedure had to sleep).
1119 * When asked to remove the page directory page itself, we actually just
1120 * leave it cached so we do not have to incur the SMP inval overhead of
1121 * removing the kernel mapping. pmap_puninit() will take care of it.
1124 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1126 unsigned *pde = (unsigned *) pmap->pm_pdir;
1128 * This code optimizes the case of freeing non-busy
1129 * page-table pages. Those pages are zero now, and
1130 * might as well be placed directly into the zero queue.
1132 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1138 * Remove the page table page from the processes address space.
1141 pmap->pm_stats.resident_count--;
1143 if (p->hold_count) {
1144 panic("pmap_release: freeing held page table page");
1146 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1147 pmap->pm_ptphint = NULL;
1150 * We leave the page directory page cached, wired, and mapped in
1151 * the pmap until the dtor function (pmap_puninit()) gets called.
1152 * However, still clean it up so we can set PG_ZERO.
1154 if (p->pindex == PTDPTDI) {
1155 bzero(pde + KPTDI, nkpt * PTESIZE);
1158 vm_page_flag_set(p, PG_ZERO);
1162 vmstats.v_wire_count--;
1163 vm_page_free_zero(p);
1169 * this routine is called if the page table page is not
1173 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1175 vm_offset_t pteva, ptepa;
1179 * Find or fabricate a new pagetable page
1181 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1182 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1184 KASSERT(m->queue == PQ_NONE,
1185 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1187 if (m->wire_count == 0)
1188 vmstats.v_wire_count++;
1192 * Increment the hold count for the page table page
1193 * (denoting a new mapping.)
1198 * Map the pagetable page into the process address space, if
1199 * it isn't already there.
1202 pmap->pm_stats.resident_count++;
1204 ptepa = VM_PAGE_TO_PHYS(m);
1205 pmap->pm_pdir[ptepindex] =
1206 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1209 * Set the page table hint
1211 pmap->pm_ptphint = m;
1214 * Try to use the new mapping, but if we cannot, then
1215 * do it with the routine that maps the page explicitly.
1217 if ((m->flags & PG_ZERO) == 0) {
1218 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1219 (((unsigned) PTDpde) & PG_FRAME)) {
1220 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1221 bzero((caddr_t) pteva, PAGE_SIZE);
1223 pmap_zero_page(ptepa);
1227 m->valid = VM_PAGE_BITS_ALL;
1228 vm_page_flag_clear(m, PG_ZERO);
1229 vm_page_flag_set(m, PG_MAPPED);
1236 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1243 * Calculate pagetable page index
1245 ptepindex = va >> PDRSHIFT;
1248 * Get the page directory entry
1250 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1253 * This supports switching from a 4MB page to a
1256 if (ptepa & PG_PS) {
1257 pmap->pm_pdir[ptepindex] = 0;
1264 * If the page table page is mapped, we just increment the
1265 * hold count, and activate it.
1269 * In order to get the page table page, try the
1272 if (pmap->pm_ptphint &&
1273 (pmap->pm_ptphint->pindex == ptepindex)) {
1274 m = pmap->pm_ptphint;
1276 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1277 pmap->pm_ptphint = m;
1283 * Here if the pte page isn't mapped, or if it has been deallocated.
1285 return _pmap_allocpte(pmap, ptepindex);
1289 /***************************************************
1290 * Pmap allocation/deallocation routines.
1291 ***************************************************/
1294 * Release any resources held by the given physical map.
1295 * Called when a pmap initialized by pmap_pinit is being released.
1296 * Should only be called if the map contains no valid mappings.
1298 static int pmap_release_callback(struct vm_page *p, void *data);
1301 pmap_release(struct pmap *pmap)
1303 vm_object_t object = pmap->pm_pteobj;
1304 struct rb_vm_page_scan_info info;
1306 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1307 #if defined(DIAGNOSTIC)
1308 if (object->ref_count != 1)
1309 panic("pmap_release: pteobj reference count != 1");
1313 info.object = object;
1315 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1322 info.limit = object->generation;
1324 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1325 pmap_release_callback, &info);
1326 if (info.error == 0 && info.mpte) {
1327 if (!pmap_release_free_page(pmap, info.mpte))
1331 } while (info.error);
1335 pmap_release_callback(struct vm_page *p, void *data)
1337 struct rb_vm_page_scan_info *info = data;
1339 if (p->pindex == PTDPTDI) {
1343 if (!pmap_release_free_page(info->pmap, p)) {
1347 if (info->object->generation != info->limit) {
1355 * Grow the number of kernel page table entries, if needed.
1359 pmap_growkernel(vm_offset_t addr)
1362 vm_offset_t ptppaddr;
1367 if (kernel_vm_end == 0) {
1368 kernel_vm_end = KERNBASE;
1370 while (pdir_pde(PTD, kernel_vm_end)) {
1371 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1375 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1376 while (kernel_vm_end < addr) {
1377 if (pdir_pde(PTD, kernel_vm_end)) {
1378 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1383 * This index is bogus, but out of the way
1385 nkpg = vm_page_alloc(kptobj, nkpt,
1386 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1388 panic("pmap_growkernel: no memory to grow kernel");
1391 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1392 pmap_zero_page(ptppaddr);
1393 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1394 pdir_pde(PTD, kernel_vm_end) = newpdir;
1395 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1399 * This update must be interlocked with pmap_pinit2.
1401 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1402 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1404 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1405 ~(PAGE_SIZE * NPTEPG - 1);
1411 * Retire the given physical map from service.
1412 * Should only be called if the map contains
1413 * no valid mappings.
1416 pmap_destroy(pmap_t pmap)
1423 count = --pmap->pm_count;
1426 panic("destroying a pmap is not yet implemented");
1431 * Add a reference to the specified pmap.
1434 pmap_reference(pmap_t pmap)
1441 /***************************************************
1442 * page management routines.
1443 ***************************************************/
1446 * free the pv_entry back to the free list. This function may be
1447 * called from an interrupt.
1449 static PMAP_INLINE void
1450 free_pv_entry(pv_entry_t pv)
1457 * get a new pv_entry, allocating a block from the system
1458 * when needed. This function may be called from an interrupt.
1464 if (pv_entry_high_water &&
1465 (pv_entry_count > pv_entry_high_water) &&
1466 (pmap_pagedaemon_waken == 0)) {
1467 pmap_pagedaemon_waken = 1;
1468 wakeup (&vm_pages_needed);
1470 return zalloc(pvzone);
1474 * This routine is very drastic, but can save the system
1482 static int warningdone=0;
1484 if (pmap_pagedaemon_waken == 0)
1487 if (warningdone < 5) {
1488 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1492 for(i = 0; i < vm_page_array_size; i++) {
1493 m = &vm_page_array[i];
1494 if (m->wire_count || m->hold_count || m->busy ||
1495 (m->flags & PG_BUSY))
1499 pmap_pagedaemon_waken = 0;
1504 * If it is the first entry on the list, it is actually
1505 * in the header and we must copy the following entry up
1506 * to the header. Otherwise we must search the list for
1507 * the entry. In either case we free the now unused entry.
1510 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1511 vm_offset_t va, pmap_inval_info_t info)
1517 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1518 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1519 if (pmap == pv->pv_pmap && va == pv->pv_va)
1523 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1524 if (va == pv->pv_va)
1531 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1532 m->md.pv_list_count--;
1533 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1534 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1535 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1536 ++pmap->pm_generation;
1537 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1545 * Create a pv entry for page at pa for
1549 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1554 pv = get_pv_entry();
1559 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1560 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1561 m->md.pv_list_count++;
1567 * pmap_remove_pte: do the things to unmap a page in a process
1570 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1571 pmap_inval_info_t info)
1576 pmap_inval_add(info, pmap, va);
1577 oldpte = loadandclear(ptq);
1579 pmap->pm_stats.wired_count -= 1;
1581 * Machines that don't support invlpg, also don't support
1582 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1586 cpu_invlpg((void *)va);
1587 pmap->pm_stats.resident_count -= 1;
1588 if (oldpte & PG_MANAGED) {
1589 m = PHYS_TO_VM_PAGE(oldpte);
1590 if (oldpte & PG_M) {
1591 #if defined(PMAP_DIAGNOSTIC)
1592 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1594 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1598 if (pmap_track_modified(va))
1602 vm_page_flag_set(m, PG_REFERENCED);
1603 return pmap_remove_entry(pmap, m, va, info);
1605 return pmap_unuse_pt(pmap, va, NULL, info);
1614 * Remove a single page from a process address space.
1616 * This function may not be called from an interrupt if the pmap is
1620 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1625 * if there is no pte for this address, just skip it!!! Otherwise
1626 * get a local va for mappings for this pmap and remove the entry.
1628 if (*pmap_pde(pmap, va) != 0) {
1629 ptq = get_ptbase(pmap) + i386_btop(va);
1631 pmap_remove_pte(pmap, ptq, va, info);
1639 * Remove the given range of addresses from the specified map.
1641 * It is assumed that the start and end are properly
1642 * rounded to the page size.
1644 * This function may not be called from an interrupt if the pmap is
1648 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1652 vm_offset_t ptpaddr;
1653 vm_offset_t sindex, eindex;
1654 struct pmap_inval_info info;
1659 if (pmap->pm_stats.resident_count == 0)
1662 pmap_inval_init(&info);
1665 * special handling of removing one page. a very
1666 * common operation and easy to short circuit some
1669 if (((sva + PAGE_SIZE) == eva) &&
1670 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1671 pmap_remove_page(pmap, sva, &info);
1672 pmap_inval_flush(&info);
1677 * Get a local virtual address for the mappings that are being
1680 sindex = i386_btop(sva);
1681 eindex = i386_btop(eva);
1683 for (; sindex < eindex; sindex = pdnxt) {
1687 * Calculate index for next page table.
1689 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1690 if (pmap->pm_stats.resident_count == 0)
1693 pdirindex = sindex / NPDEPG;
1694 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1695 pmap_inval_add(&info, pmap, -1);
1696 pmap->pm_pdir[pdirindex] = 0;
1697 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1702 * Weed out invalid mappings. Note: we assume that the page
1703 * directory table is always allocated, and in kernel virtual.
1709 * Limit our scan to either the end of the va represented
1710 * by the current page table page, or to the end of the
1711 * range being removed.
1713 if (pdnxt > eindex) {
1718 * NOTE: pmap_remove_pte() can block.
1720 for (; sindex != pdnxt; sindex++) {
1723 ptbase = get_ptbase(pmap);
1724 if (ptbase[sindex] == 0)
1726 va = i386_ptob(sindex);
1727 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1731 pmap_inval_flush(&info);
1737 * Removes this physical page from all physical maps in which it resides.
1738 * Reflects back modify bits to the pager.
1740 * This routine may not be called from an interrupt.
1744 pmap_remove_all(vm_page_t m)
1746 struct pmap_inval_info info;
1747 unsigned *pte, tpte;
1750 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
1753 pmap_inval_init(&info);
1755 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1756 pv->pv_pmap->pm_stats.resident_count--;
1758 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1759 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1761 tpte = loadandclear(pte);
1763 pv->pv_pmap->pm_stats.wired_count--;
1766 vm_page_flag_set(m, PG_REFERENCED);
1769 * Update the vm_page_t clean and reference bits.
1772 #if defined(PMAP_DIAGNOSTIC)
1773 if (pmap_nw_modified((pt_entry_t) tpte)) {
1775 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1779 if (pmap_track_modified(pv->pv_va))
1782 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1783 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1784 ++pv->pv_pmap->pm_generation;
1785 m->md.pv_list_count--;
1786 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1790 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1792 pmap_inval_flush(&info);
1798 * Set the physical protection on the specified range of this map
1801 * This function may not be called from an interrupt if the map is
1802 * not the kernel_pmap.
1805 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1808 vm_offset_t pdnxt, ptpaddr;
1809 vm_pindex_t sindex, eindex;
1810 pmap_inval_info info;
1815 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1816 pmap_remove(pmap, sva, eva);
1820 if (prot & VM_PROT_WRITE)
1823 pmap_inval_init(&info);
1825 ptbase = get_ptbase(pmap);
1827 sindex = i386_btop(sva);
1828 eindex = i386_btop(eva);
1830 for (; sindex < eindex; sindex = pdnxt) {
1834 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1836 pdirindex = sindex / NPDEPG;
1837 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1838 pmap_inval_add(&info, pmap, -1);
1839 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1840 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1845 * Weed out invalid mappings. Note: we assume that the page
1846 * directory table is always allocated, and in kernel virtual.
1851 if (pdnxt > eindex) {
1855 for (; sindex != pdnxt; sindex++) {
1860 /* XXX this isn't optimal */
1861 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1862 pbits = ptbase[sindex];
1864 if (pbits & PG_MANAGED) {
1867 m = PHYS_TO_VM_PAGE(pbits);
1868 vm_page_flag_set(m, PG_REFERENCED);
1872 if (pmap_track_modified(i386_ptob(sindex))) {
1874 m = PHYS_TO_VM_PAGE(pbits);
1883 if (pbits != ptbase[sindex]) {
1884 ptbase[sindex] = pbits;
1888 pmap_inval_flush(&info);
1892 * Insert the given physical page (p) at
1893 * the specified virtual address (v) in the
1894 * target physical map with the protection requested.
1896 * If specified, the page will be wired down, meaning
1897 * that the related pte can not be reclaimed.
1899 * NB: This is the only routine which MAY NOT lazy-evaluate
1900 * or lose information. That is, this routine must actually
1901 * insert this page into the given map NOW.
1904 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1910 vm_offset_t origpte, newpte;
1912 pmap_inval_info info;
1918 #ifdef PMAP_DIAGNOSTIC
1920 panic("pmap_enter: toobig");
1921 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1922 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1924 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
1925 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
1927 db_print_backtrace();
1930 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
1931 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
1933 db_print_backtrace();
1939 * In the case that a page table page is not
1940 * resident, we are creating it here.
1942 if (va < UPT_MIN_ADDRESS) {
1943 mpte = pmap_allocpte(pmap, va);
1946 pmap_inval_init(&info);
1947 pte = pmap_pte(pmap, va);
1950 * Page Directory table entry not valid, we need a new PT page
1953 panic("pmap_enter: invalid page directory pdir=%x, va=0x%x\n",
1954 (unsigned) pmap->pm_pdir[PTDPTDI], va);
1957 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
1958 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1959 origpte = *(vm_offset_t *)pte;
1960 opa = origpte & PG_FRAME;
1962 if (origpte & PG_PS)
1963 panic("pmap_enter: attempted pmap_enter on 4MB page");
1966 * Mapping has not changed, must be protection or wiring change.
1968 if (origpte && (opa == pa)) {
1970 * Wiring change, just update stats. We don't worry about
1971 * wiring PT pages as they remain resident as long as there
1972 * are valid mappings in them. Hence, if a user page is wired,
1973 * the PT page will be also.
1975 if (wired && ((origpte & PG_W) == 0))
1976 pmap->pm_stats.wired_count++;
1977 else if (!wired && (origpte & PG_W))
1978 pmap->pm_stats.wired_count--;
1980 #if defined(PMAP_DIAGNOSTIC)
1981 if (pmap_nw_modified((pt_entry_t) origpte)) {
1983 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1989 * Remove the extra pte reference. Note that we cannot
1990 * optimize the RO->RW case because we have adjusted the
1991 * wiring count above and may need to adjust the wiring
1998 * We might be turning off write access to the page,
1999 * so we go ahead and sense modify status.
2001 if (origpte & PG_MANAGED) {
2002 if ((origpte & PG_M) && pmap_track_modified(va)) {
2004 om = PHYS_TO_VM_PAGE(opa);
2012 * Mapping has changed, invalidate old range and fall through to
2013 * handle validating new mapping.
2017 err = pmap_remove_pte(pmap, pte, va, &info);
2019 panic("pmap_enter: pte vanished, va: 0x%x", va);
2023 * Enter on the PV list if part of our managed memory. Note that we
2024 * raise IPL while manipulating pv_table since pmap_enter can be
2025 * called at interrupt time.
2027 if (pmap_initialized &&
2028 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2029 pmap_insert_entry(pmap, va, mpte, m);
2034 * Increment counters
2036 pmap->pm_stats.resident_count++;
2038 pmap->pm_stats.wired_count++;
2042 * Now validate mapping with desired protection/wiring.
2044 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2048 if (va < UPT_MIN_ADDRESS)
2050 if (pmap == &kernel_pmap)
2054 * if the mapping or permission bits are different, we need
2055 * to update the pte.
2057 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2058 *pte = newpte | PG_A;
2060 pmap_inval_flush(&info);
2064 * this code makes some *MAJOR* assumptions:
2065 * 1. Current pmap & pmap exists.
2068 * 4. No page table pages.
2069 * 5. Tlbflush is deferred to calling procedure.
2070 * 6. Page IS managed.
2071 * but is *MUCH* faster than pmap_enter...
2075 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
2079 pmap_inval_info info;
2081 pmap_inval_init(&info);
2083 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2084 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2086 db_print_backtrace();
2089 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2090 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2092 db_print_backtrace();
2097 * In the case that a page table page is not
2098 * resident, we are creating it here.
2100 if (va < UPT_MIN_ADDRESS) {
2105 * Calculate pagetable page index
2107 ptepindex = va >> PDRSHIFT;
2108 if (mpte && (mpte->pindex == ptepindex)) {
2113 * Get the page directory entry
2115 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2118 * If the page table page is mapped, we just increment
2119 * the hold count, and activate it.
2123 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2124 if (pmap->pm_ptphint &&
2125 (pmap->pm_ptphint->pindex == ptepindex)) {
2126 mpte = pmap->pm_ptphint;
2128 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2129 pmap->pm_ptphint = mpte;
2135 mpte = _pmap_allocpte(pmap, ptepindex);
2143 * This call to vtopte makes the assumption that we are
2144 * entering the page into the current pmap. In order to support
2145 * quick entry into any pmap, one would likely use pmap_pte_quick.
2146 * But that isn't as quick as vtopte.
2148 pte = (unsigned *)vtopte(va);
2151 pmap_unwire_pte_hold(pmap, mpte, &info);
2156 * Enter on the PV list if part of our managed memory. Note that we
2157 * raise IPL while manipulating pv_table since pmap_enter can be
2158 * called at interrupt time.
2160 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
2161 pmap_insert_entry(pmap, va, mpte, m);
2164 * Increment counters
2166 pmap->pm_stats.resident_count++;
2168 pa = VM_PAGE_TO_PHYS(m);
2171 * Now validate mapping with RO protection
2173 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2174 *pte = pa | PG_V | PG_U;
2176 *pte = pa | PG_V | PG_U | PG_MANAGED;
2182 * Make a temporary mapping for a physical address. This is only intended
2183 * to be used for panic dumps.
2186 pmap_kenter_temporary(vm_paddr_t pa, int i)
2188 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2189 return ((void *)crashdumpmap);
2192 #define MAX_INIT_PT (96)
2195 * This routine preloads the ptes for a given object into the specified pmap.
2196 * This eliminates the blast of soft faults on process startup and
2197 * immediately after an mmap.
2199 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2202 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2203 vm_object_t object, vm_pindex_t pindex,
2204 vm_size_t size, int limit)
2206 struct rb_vm_page_scan_info info;
2211 * We can't preinit if read access isn't set or there is no pmap
2214 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2218 * We can't preinit if the pmap is not the current pmap
2220 lp = curthread->td_lwp;
2221 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2224 psize = i386_btop(size);
2226 if ((object->type != OBJT_VNODE) ||
2227 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2228 (object->resident_page_count > MAX_INIT_PT))) {
2232 if (psize + pindex > object->size) {
2233 if (object->size < pindex)
2235 psize = object->size - pindex;
2242 * Use a red-black scan to traverse the requested range and load
2243 * any valid pages found into the pmap.
2245 * We cannot safely scan the object's memq unless we are in a
2246 * critical section since interrupts can remove pages from objects.
2248 info.start_pindex = pindex;
2249 info.end_pindex = pindex + psize - 1;
2256 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2257 pmap_object_init_pt_callback, &info);
2263 pmap_object_init_pt_callback(vm_page_t p, void *data)
2265 struct rb_vm_page_scan_info *info = data;
2266 vm_pindex_t rel_index;
2268 * don't allow an madvise to blow away our really
2269 * free pages allocating pv entries.
2271 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2272 vmstats.v_free_count < vmstats.v_free_reserved) {
2275 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2276 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2277 if ((p->queue - p->pc) == PQ_CACHE)
2278 vm_page_deactivate(p);
2280 rel_index = p->pindex - info->start_pindex;
2281 info->mpte = pmap_enter_quick(info->pmap,
2282 info->addr + i386_ptob(rel_index),
2284 vm_page_flag_set(p, PG_MAPPED);
2291 * pmap_prefault provides a quick way of clustering pagefaults into a
2292 * processes address space. It is a "cousin" of pmap_object_init_pt,
2293 * except it runs at page fault time instead of mmap time.
2297 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2299 static int pmap_prefault_pageorder[] = {
2300 -PAGE_SIZE, PAGE_SIZE,
2301 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2302 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2303 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2307 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2318 * We do not currently prefault mappings that use virtual page
2319 * tables. We do not prefault foreign pmaps.
2321 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2323 lp = curthread->td_lwp;
2324 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2327 object = entry->object.vm_object;
2329 starta = addra - PFBAK * PAGE_SIZE;
2330 if (starta < entry->start)
2331 starta = entry->start;
2332 else if (starta > addra)
2336 * critical section protection is required to maintain the
2337 * page/object association, interrupts can free pages and remove
2338 * them from their objects.
2342 for (i = 0; i < PAGEORDER_SIZE; i++) {
2343 vm_object_t lobject;
2346 addr = addra + pmap_prefault_pageorder[i];
2347 if (addr > addra + (PFFOR * PAGE_SIZE))
2350 if (addr < starta || addr >= entry->end)
2353 if ((*pmap_pde(pmap, addr)) == NULL)
2356 pte = (unsigned *) vtopte(addr);
2360 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2363 for (m = vm_page_lookup(lobject, pindex);
2364 (!m && (lobject->type == OBJT_DEFAULT) &&
2365 (lobject->backing_object));
2366 lobject = lobject->backing_object
2368 if (lobject->backing_object_offset & PAGE_MASK)
2370 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2371 m = vm_page_lookup(lobject->backing_object, pindex);
2375 * give-up when a page is not in memory
2380 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2382 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2384 if ((m->queue - m->pc) == PQ_CACHE) {
2385 vm_page_deactivate(m);
2388 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2389 vm_page_flag_set(m, PG_MAPPED);
2397 * Routine: pmap_change_wiring
2398 * Function: Change the wiring attribute for a map/virtual-address
2400 * In/out conditions:
2401 * The mapping must already exist in the pmap.
2404 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2411 pte = pmap_pte(pmap, va);
2413 if (wired && !pmap_pte_w(pte))
2414 pmap->pm_stats.wired_count++;
2415 else if (!wired && pmap_pte_w(pte))
2416 pmap->pm_stats.wired_count--;
2419 * Wiring is not a hardware characteristic so there is no need to
2420 * invalidate TLB. However, in an SMP environment we must use
2421 * a locked bus cycle to update the pte (if we are not using
2422 * the pmap_inval_*() API that is)... it's ok to do this for simple
2427 atomic_set_int(pte, PG_W);
2429 atomic_clear_int(pte, PG_W);
2432 atomic_set_int_nonlocked(pte, PG_W);
2434 atomic_clear_int_nonlocked(pte, PG_W);
2441 * Copy the range specified by src_addr/len
2442 * from the source map to the range dst_addr/len
2443 * in the destination map.
2445 * This routine is only advisory and need not do anything.
2448 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2449 vm_size_t len, vm_offset_t src_addr)
2451 pmap_inval_info info;
2453 vm_offset_t end_addr = src_addr + len;
2455 unsigned src_frame, dst_frame;
2458 if (dst_addr != src_addr)
2461 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2462 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2466 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2467 if (dst_frame != (((unsigned) APTDpde) & PG_FRAME)) {
2468 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2469 /* The page directory is not shared between CPUs */
2472 pmap_inval_init(&info);
2473 pmap_inval_add(&info, dst_pmap, -1);
2474 pmap_inval_add(&info, src_pmap, -1);
2477 * critical section protection is required to maintain the page/object
2478 * association, interrupts can free pages and remove them from
2482 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2483 unsigned *src_pte, *dst_pte;
2484 vm_page_t dstmpte, srcmpte;
2485 vm_offset_t srcptepaddr;
2488 if (addr >= UPT_MIN_ADDRESS)
2489 panic("pmap_copy: invalid to pmap_copy page tables\n");
2492 * Don't let optional prefaulting of pages make us go
2493 * way below the low water mark of free pages or way
2494 * above high water mark of used pv entries.
2496 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2497 pv_entry_count > pv_entry_high_water)
2500 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2501 ptepindex = addr >> PDRSHIFT;
2503 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2504 if (srcptepaddr == 0)
2507 if (srcptepaddr & PG_PS) {
2508 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2509 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2510 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2515 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2516 if ((srcmpte == NULL) ||
2517 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2520 if (pdnxt > end_addr)
2523 src_pte = (unsigned *) vtopte(addr);
2524 dst_pte = (unsigned *) avtopte(addr);
2525 while (addr < pdnxt) {
2530 * we only virtual copy managed pages
2532 if ((ptetemp & PG_MANAGED) != 0) {
2534 * We have to check after allocpte for the
2535 * pte still being around... allocpte can
2538 dstmpte = pmap_allocpte(dst_pmap, addr);
2539 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2541 * Clear the modified and
2542 * accessed (referenced) bits
2545 m = PHYS_TO_VM_PAGE(ptetemp);
2546 *dst_pte = ptetemp & ~(PG_M | PG_A);
2547 dst_pmap->pm_stats.resident_count++;
2548 pmap_insert_entry(dst_pmap, addr,
2551 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2553 if (dstmpte->hold_count >= srcmpte->hold_count)
2562 pmap_inval_flush(&info);
2568 * Zero the specified PA by mapping the page into KVM and clearing its
2571 * This function may be called from an interrupt and no locking is
2575 pmap_zero_page(vm_paddr_t phys)
2577 struct mdglobaldata *gd = mdcpu;
2580 if (*(int *)gd->gd_CMAP3)
2581 panic("pmap_zero_page: CMAP3 busy");
2582 *(int *)gd->gd_CMAP3 =
2583 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2584 cpu_invlpg(gd->gd_CADDR3);
2586 #if defined(I686_CPU)
2587 if (cpu_class == CPUCLASS_686)
2588 i686_pagezero(gd->gd_CADDR3);
2591 bzero(gd->gd_CADDR3, PAGE_SIZE);
2592 *(int *) gd->gd_CMAP3 = 0;
2597 * pmap_page_assertzero:
2599 * Assert that a page is empty, panic if it isn't.
2602 pmap_page_assertzero(vm_paddr_t phys)
2604 struct mdglobaldata *gd = mdcpu;
2608 if (*(int *)gd->gd_CMAP3)
2609 panic("pmap_zero_page: CMAP3 busy");
2610 *(int *)gd->gd_CMAP3 =
2611 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2612 cpu_invlpg(gd->gd_CADDR3);
2613 for (i = 0; i < PAGE_SIZE; i += 4) {
2614 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2615 panic("pmap_page_assertzero() @ %p not zero!\n",
2616 (void *)gd->gd_CADDR3);
2619 *(int *) gd->gd_CMAP3 = 0;
2626 * Zero part of a physical page by mapping it into memory and clearing
2627 * its contents with bzero.
2629 * off and size may not cover an area beyond a single hardware page.
2632 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2634 struct mdglobaldata *gd = mdcpu;
2637 if (*(int *) gd->gd_CMAP3)
2638 panic("pmap_zero_page: CMAP3 busy");
2639 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2640 cpu_invlpg(gd->gd_CADDR3);
2642 #if defined(I686_CPU)
2643 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2644 i686_pagezero(gd->gd_CADDR3);
2647 bzero((char *)gd->gd_CADDR3 + off, size);
2648 *(int *) gd->gd_CMAP3 = 0;
2655 * Copy the physical page from the source PA to the target PA.
2656 * This function may be called from an interrupt. No locking
2660 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2662 struct mdglobaldata *gd = mdcpu;
2665 if (*(int *) gd->gd_CMAP1)
2666 panic("pmap_copy_page: CMAP1 busy");
2667 if (*(int *) gd->gd_CMAP2)
2668 panic("pmap_copy_page: CMAP2 busy");
2670 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2671 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2673 cpu_invlpg(gd->gd_CADDR1);
2674 cpu_invlpg(gd->gd_CADDR2);
2676 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2678 *(int *) gd->gd_CMAP1 = 0;
2679 *(int *) gd->gd_CMAP2 = 0;
2684 * pmap_copy_page_frag:
2686 * Copy the physical page from the source PA to the target PA.
2687 * This function may be called from an interrupt. No locking
2691 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2693 struct mdglobaldata *gd = mdcpu;
2696 if (*(int *) gd->gd_CMAP1)
2697 panic("pmap_copy_page: CMAP1 busy");
2698 if (*(int *) gd->gd_CMAP2)
2699 panic("pmap_copy_page: CMAP2 busy");
2701 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2702 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2704 cpu_invlpg(gd->gd_CADDR1);
2705 cpu_invlpg(gd->gd_CADDR2);
2707 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2708 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2711 *(int *) gd->gd_CMAP1 = 0;
2712 *(int *) gd->gd_CMAP2 = 0;
2717 * Returns true if the pmap's pv is one of the first
2718 * 16 pvs linked to from this page. This count may
2719 * be changed upwards or downwards in the future; it
2720 * is only necessary that true be returned for a small
2721 * subset of pmaps for proper page aging.
2724 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2729 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2734 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2735 if (pv->pv_pmap == pmap) {
2748 * Remove all pages from specified address space
2749 * this aids process exit speeds. Also, this code
2750 * is special cased for current process only, but
2751 * can have the more generic (and slightly slower)
2752 * mode enabled. This is much faster than pmap_remove
2753 * in the case of running down an entire address space.
2756 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2759 unsigned *pte, tpte;
2762 pmap_inval_info info;
2764 int32_t save_generation;
2766 lp = curthread->td_lwp;
2767 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2772 pmap_inval_init(&info);
2774 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2775 if (pv->pv_va >= eva || pv->pv_va < sva) {
2776 npv = TAILQ_NEXT(pv, pv_plist);
2780 KKASSERT(pmap == pv->pv_pmap);
2783 pte = (unsigned *)vtopte(pv->pv_va);
2785 pte = pmap_pte_quick(pmap, pv->pv_va);
2786 if (pmap->pm_active)
2787 pmap_inval_add(&info, pmap, pv->pv_va);
2791 * We cannot remove wired pages from a process' mapping
2795 npv = TAILQ_NEXT(pv, pv_plist);
2800 m = PHYS_TO_VM_PAGE(tpte);
2802 KASSERT(m < &vm_page_array[vm_page_array_size],
2803 ("pmap_remove_pages: bad tpte %x", tpte));
2805 pmap->pm_stats.resident_count--;
2808 * Update the vm_page_t clean and reference bits.
2814 npv = TAILQ_NEXT(pv, pv_plist);
2815 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2816 save_generation = ++pmap->pm_generation;
2818 m->md.pv_list_count--;
2819 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2820 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2821 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2824 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2828 * Restart the scan if we blocked during the unuse or free
2829 * calls and other removals were made.
2831 if (save_generation != pmap->pm_generation) {
2832 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2833 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2836 pmap_inval_flush(&info);
2841 * pmap_testbit tests bits in pte's
2842 * note that the testbit/clearbit routines are inline,
2843 * and a lot of things compile-time evaluate.
2846 pmap_testbit(vm_page_t m, int bit)
2851 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2854 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2859 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2861 * if the bit being tested is the modified bit, then
2862 * mark clean_map and ptes as never
2865 if (bit & (PG_A|PG_M)) {
2866 if (!pmap_track_modified(pv->pv_va))
2870 #if defined(PMAP_DIAGNOSTIC)
2872 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2876 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2887 * this routine is used to modify bits in ptes
2889 static __inline void
2890 pmap_clearbit(vm_page_t m, int bit)
2892 struct pmap_inval_info info;
2897 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2900 pmap_inval_init(&info);
2904 * Loop over all current mappings setting/clearing as appropos If
2905 * setting RO do we need to clear the VAC?
2907 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2909 * don't write protect pager mappings
2912 if (!pmap_track_modified(pv->pv_va))
2916 #if defined(PMAP_DIAGNOSTIC)
2918 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2924 * Careful here. We can use a locked bus instruction to
2925 * clear PG_A or PG_M safely but we need to synchronize
2926 * with the target cpus when we mess with PG_RW.
2928 * We do not have to force synchronization when clearing
2929 * PG_M even for PTEs generated via virtual memory maps,
2930 * because the virtual kernel will invalidate the pmap
2931 * entry when/if it needs to resynchronize the Modify bit.
2933 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2935 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2942 atomic_clear_int(pte, PG_M|PG_RW);
2943 } else if (bit == PG_M) {
2945 * We could also clear PG_RW here to force
2946 * a fault on write to redetect PG_M for
2947 * virtual kernels, but it isn't necessary
2948 * since virtual kernels invalidate the pte
2949 * when they clear the VPTE_M bit in their
2950 * virtual page tables.
2952 atomic_clear_int(pte, PG_M);
2954 atomic_clear_int(pte, bit);
2958 pmap_inval_flush(&info);
2963 * pmap_page_protect:
2965 * Lower the permission for all mappings to a given page.
2968 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2970 if ((prot & VM_PROT_WRITE) == 0) {
2971 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2972 pmap_clearbit(m, PG_RW);
2980 pmap_phys_address(int ppn)
2982 return (i386_ptob(ppn));
2986 * pmap_ts_referenced:
2988 * Return a count of reference bits for a page, clearing those bits.
2989 * It is not necessary for every reference bit to be cleared, but it
2990 * is necessary that 0 only be returned when there are truly no
2991 * reference bits set.
2993 * XXX: The exact number of bits to check and clear is a matter that
2994 * should be tested and standardized at some point in the future for
2995 * optimal aging of shared pages.
2998 pmap_ts_referenced(vm_page_t m)
3000 pv_entry_t pv, pvf, pvn;
3004 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3009 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3014 pvn = TAILQ_NEXT(pv, pv_list);
3016 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3018 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3020 if (!pmap_track_modified(pv->pv_va))
3023 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3025 if (pte && (*pte & PG_A)) {
3027 atomic_clear_int(pte, PG_A);
3029 atomic_clear_int_nonlocked(pte, PG_A);
3036 } while ((pv = pvn) != NULL && pv != pvf);
3046 * Return whether or not the specified physical page was modified
3047 * in any physical maps.
3050 pmap_is_modified(vm_page_t m)
3052 return pmap_testbit(m, PG_M);
3056 * Clear the modify bits on the specified physical page.
3059 pmap_clear_modify(vm_page_t m)
3061 pmap_clearbit(m, PG_M);
3065 * pmap_clear_reference:
3067 * Clear the reference bit on the specified physical page.
3070 pmap_clear_reference(vm_page_t m)
3072 pmap_clearbit(m, PG_A);
3076 * Miscellaneous support routines follow
3080 i386_protection_init(void)
3084 kp = protection_codes;
3085 for (prot = 0; prot < 8; prot++) {
3087 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3089 * Read access is also 0. There isn't any execute bit,
3090 * so just make it readable.
3092 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3093 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3094 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3097 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3098 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3099 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3100 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3108 * Map a set of physical memory pages into the kernel virtual
3109 * address space. Return a pointer to where it is mapped. This
3110 * routine is intended to be used for mapping device memory,
3113 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3117 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3119 vm_offset_t va, tmpva, offset;
3122 offset = pa & PAGE_MASK;
3123 size = roundup(offset + size, PAGE_SIZE);
3125 va = kmem_alloc_nofault(&kernel_map, size);
3127 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3130 for (tmpva = va; size > 0;) {
3131 pte = (unsigned *)vtopte(tmpva);
3132 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3140 return ((void *)(va + offset));
3144 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3146 vm_offset_t base, offset;
3148 base = va & PG_FRAME;
3149 offset = va & PAGE_MASK;
3150 size = roundup(offset + size, PAGE_SIZE);
3151 pmap_qremove(va, size >> PAGE_SHIFT);
3152 kmem_free(&kernel_map, base, size);
3156 * perform the pmap work for mincore
3159 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3161 unsigned *ptep, pte;
3165 ptep = pmap_pte(pmap, addr);
3170 if ((pte = *ptep) != 0) {
3173 val = MINCORE_INCORE;
3174 if ((pte & PG_MANAGED) == 0)
3177 pa = pte & PG_FRAME;
3179 m = PHYS_TO_VM_PAGE(pa);
3185 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3187 * Modified by someone
3189 else if (m->dirty || pmap_is_modified(m))
3190 val |= MINCORE_MODIFIED_OTHER;
3195 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3198 * Referenced by someone
3200 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3201 val |= MINCORE_REFERENCED_OTHER;
3202 vm_page_flag_set(m, PG_REFERENCED);
3209 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3210 * vmspace will be ref'd and the old one will be deref'd.
3212 * The vmspace for all lwps associated with the process will be adjusted
3213 * and cr3 will be reloaded if any lwp is the current lwp.
3216 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3218 struct vmspace *oldvm;
3222 oldvm = p->p_vmspace;
3223 if (oldvm != newvm) {
3224 p->p_vmspace = newvm;
3225 KKASSERT(p->p_nthreads == 1);
3226 lp = RB_ROOT(&p->p_lwp_tree);
3227 pmap_setlwpvm(lp, newvm);
3229 sysref_get(&newvm->vm_sysref);
3230 sysref_put(&oldvm->vm_sysref);
3237 * Set the vmspace for a LWP. The vmspace is almost universally set the
3238 * same as the process vmspace, but virtual kernels need to swap out contexts
3239 * on a per-lwp basis.
3242 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3244 struct vmspace *oldvm;
3248 oldvm = lp->lwp_vmspace;
3250 if (oldvm != newvm) {
3251 lp->lwp_vmspace = newvm;
3252 if (curthread->td_lwp == lp) {
3253 pmap = vmspace_pmap(newvm);
3255 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3257 pmap->pm_active |= 1;
3259 #if defined(SWTCH_OPTIM_STATS)
3262 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3263 load_cr3(curthread->td_pcb->pcb_cr3);
3264 pmap = vmspace_pmap(oldvm);
3266 atomic_clear_int(&pmap->pm_active,
3267 1 << mycpu->gd_cpuid);
3269 pmap->pm_active &= ~1;
3277 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3280 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3284 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3291 static void pads (pmap_t pm);
3292 void pmap_pvdump (vm_paddr_t pa);
3294 /* print address space of pmap*/
3301 if (pm == &kernel_pmap)
3303 for (i = 0; i < 1024; i++)
3305 for (j = 0; j < 1024; j++) {
3306 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3307 if (pm == &kernel_pmap && va < KERNBASE)
3309 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3311 ptep = pmap_pte_quick(pm, va);
3312 if (pmap_pte_v(ptep))
3313 kprintf("%x:%x ", va, *(int *) ptep);
3319 pmap_pvdump(vm_paddr_t pa)
3324 kprintf("pa %08llx", (long long)pa);
3325 m = PHYS_TO_VM_PAGE(pa);
3326 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3328 kprintf(" -> pmap %p, va %x, flags %x",
3329 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3331 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);