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 1.83 2008/04/28 07:05:06 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 #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"))
927 KASSERT(m->queue == PQ_NONE,
928 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
930 if (m->hold_count == 1) {
932 * Unmap the page table page
935 pmap_inval_add(info, pmap, -1);
936 pmap->pm_pdir[m->pindex] = 0;
938 KKASSERT(pmap->pm_stats.resident_count > 0);
939 --pmap->pm_stats.resident_count;
941 if (pmap->pm_ptphint == m)
942 pmap->pm_ptphint = NULL;
945 * This was our last hold, the page had better be unwired
946 * after we decrement wire_count.
948 * FUTURE NOTE: shared page directory page could result in
949 * multiple wire counts.
953 KKASSERT(m->wire_count == 0);
954 --vmstats.v_wire_count;
956 vm_page_free_zero(m);
962 static PMAP_INLINE int
963 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
965 KKASSERT(m->hold_count > 0);
966 if (m->hold_count > 1) {
970 return _pmap_unwire_pte_hold(pmap, m, info);
975 * After removing a page table entry, this routine is used to
976 * conditionally free the page, and manage the hold/wire counts.
979 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
980 pmap_inval_info_t info)
983 if (va >= UPT_MIN_ADDRESS)
987 ptepindex = (va >> PDRSHIFT);
988 if (pmap->pm_ptphint &&
989 (pmap->pm_ptphint->pindex == ptepindex)) {
990 mpte = pmap->pm_ptphint;
992 pmap_inval_flush(info);
993 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
994 pmap->pm_ptphint = mpte;
998 return pmap_unwire_pte_hold(pmap, mpte, info);
1002 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1003 * it, and IdlePTD, represents the template used to update all other pmaps.
1005 * On architectures where the kernel pmap is not integrated into the user
1006 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1007 * kernel_pmap should be used to directly access the kernel_pmap.
1010 pmap_pinit0(struct pmap *pmap)
1013 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1014 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1016 pmap->pm_active = 0;
1017 pmap->pm_ptphint = NULL;
1018 TAILQ_INIT(&pmap->pm_pvlist);
1019 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1023 * Initialize a preallocated and zeroed pmap structure,
1024 * such as one in a vmspace structure.
1027 pmap_pinit(struct pmap *pmap)
1032 * No need to allocate page table space yet but we do need a valid
1033 * page directory table.
1035 if (pmap->pm_pdir == NULL) {
1037 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1041 * Allocate an object for the ptes
1043 if (pmap->pm_pteobj == NULL)
1044 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1047 * Allocate the page directory page, unless we already have
1048 * one cached. If we used the cached page the wire_count will
1049 * already be set appropriately.
1051 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1052 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1053 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1054 pmap->pm_pdirm = ptdpg;
1055 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1056 ptdpg->valid = VM_PAGE_BITS_ALL;
1057 ptdpg->wire_count = 1;
1058 ++vmstats.v_wire_count;
1059 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1061 if ((ptdpg->flags & PG_ZERO) == 0)
1062 bzero(pmap->pm_pdir, PAGE_SIZE);
1064 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1066 /* install self-referential address mapping entry */
1067 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1068 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1071 pmap->pm_active = 0;
1072 pmap->pm_ptphint = NULL;
1073 TAILQ_INIT(&pmap->pm_pvlist);
1074 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1075 pmap->pm_stats.resident_count = 1;
1079 * Clean up a pmap structure so it can be physically freed. This routine
1080 * is called by the vmspace dtor function. A great deal of pmap data is
1081 * left passively mapped to improve vmspace management so we have a bit
1082 * of cleanup work to do here.
1085 pmap_puninit(pmap_t pmap)
1089 KKASSERT(pmap->pm_active == 0);
1090 if ((p = pmap->pm_pdirm) != NULL) {
1091 KKASSERT(pmap->pm_pdir != NULL);
1092 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1094 vmstats.v_wire_count--;
1095 KKASSERT((p->flags & PG_BUSY) == 0);
1097 vm_page_free_zero(p);
1098 pmap->pm_pdirm = NULL;
1100 if (pmap->pm_pdir) {
1101 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1102 pmap->pm_pdir = NULL;
1104 if (pmap->pm_pteobj) {
1105 vm_object_deallocate(pmap->pm_pteobj);
1106 pmap->pm_pteobj = NULL;
1111 * Wire in kernel global address entries. To avoid a race condition
1112 * between pmap initialization and pmap_growkernel, this procedure
1113 * adds the pmap to the master list (which growkernel scans to update),
1114 * then copies the template.
1117 pmap_pinit2(struct pmap *pmap)
1120 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1121 /* XXX copies current process, does not fill in MPPTDI */
1122 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1127 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1128 * 0 on failure (if the procedure had to sleep).
1130 * When asked to remove the page directory page itself, we actually just
1131 * leave it cached so we do not have to incur the SMP inval overhead of
1132 * removing the kernel mapping. pmap_puninit() will take care of it.
1135 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1137 unsigned *pde = (unsigned *) pmap->pm_pdir;
1139 * This code optimizes the case of freeing non-busy
1140 * page-table pages. Those pages are zero now, and
1141 * might as well be placed directly into the zero queue.
1143 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1149 * Remove the page table page from the processes address space.
1152 KKASSERT(pmap->pm_stats.resident_count > 0);
1153 --pmap->pm_stats.resident_count;
1155 if (p->hold_count) {
1156 panic("pmap_release: freeing held page table page");
1158 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1159 pmap->pm_ptphint = NULL;
1162 * We leave the page directory page cached, wired, and mapped in
1163 * the pmap until the dtor function (pmap_puninit()) gets called.
1164 * However, still clean it up so we can set PG_ZERO.
1166 if (p->pindex == PTDPTDI) {
1167 bzero(pde + KPTDI, nkpt * PTESIZE);
1170 vm_page_flag_set(p, PG_ZERO);
1174 vmstats.v_wire_count--;
1175 vm_page_free_zero(p);
1181 * this routine is called if the page table page is not
1185 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1187 vm_offset_t pteva, ptepa;
1191 * Find or fabricate a new pagetable page
1193 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1194 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1196 KASSERT(m->queue == PQ_NONE,
1197 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1200 * Increment the hold count for the page we will be returning to
1206 * It is possible that someone else got in and mapped by the page
1207 * directory page while we were blocked, if so just unbusy and
1208 * return the held page.
1210 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1211 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1216 if (m->wire_count == 0)
1217 vmstats.v_wire_count++;
1222 * Map the pagetable page into the process address space, if
1223 * it isn't already there.
1226 ++pmap->pm_stats.resident_count;
1228 ptepa = VM_PAGE_TO_PHYS(m);
1229 pmap->pm_pdir[ptepindex] =
1230 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1233 * Set the page table hint
1235 pmap->pm_ptphint = m;
1238 * Try to use the new mapping, but if we cannot, then
1239 * do it with the routine that maps the page explicitly.
1241 if ((m->flags & PG_ZERO) == 0) {
1242 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1243 (((unsigned) PTDpde) & PG_FRAME)) {
1244 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1245 bzero((caddr_t) pteva, PAGE_SIZE);
1247 pmap_zero_page(ptepa);
1251 m->valid = VM_PAGE_BITS_ALL;
1252 vm_page_flag_clear(m, PG_ZERO);
1253 vm_page_flag_set(m, PG_MAPPED);
1260 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1267 * Calculate pagetable page index
1269 ptepindex = va >> PDRSHIFT;
1272 * Get the page directory entry
1274 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1277 * This supports switching from a 4MB page to a
1280 if (ptepa & PG_PS) {
1281 pmap->pm_pdir[ptepindex] = 0;
1288 * If the page table page is mapped, we just increment the
1289 * hold count, and activate it.
1293 * In order to get the page table page, try the
1296 if (pmap->pm_ptphint &&
1297 (pmap->pm_ptphint->pindex == ptepindex)) {
1298 m = pmap->pm_ptphint;
1300 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1301 pmap->pm_ptphint = m;
1307 * Here if the pte page isn't mapped, or if it has been deallocated.
1309 return _pmap_allocpte(pmap, ptepindex);
1313 /***************************************************
1314 * Pmap allocation/deallocation routines.
1315 ***************************************************/
1318 * Release any resources held by the given physical map.
1319 * Called when a pmap initialized by pmap_pinit is being released.
1320 * Should only be called if the map contains no valid mappings.
1322 static int pmap_release_callback(struct vm_page *p, void *data);
1325 pmap_release(struct pmap *pmap)
1327 vm_object_t object = pmap->pm_pteobj;
1328 struct rb_vm_page_scan_info info;
1330 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1331 #if defined(DIAGNOSTIC)
1332 if (object->ref_count != 1)
1333 panic("pmap_release: pteobj reference count != 1");
1337 info.object = object;
1339 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1346 info.limit = object->generation;
1348 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1349 pmap_release_callback, &info);
1350 if (info.error == 0 && info.mpte) {
1351 if (!pmap_release_free_page(pmap, info.mpte))
1355 } while (info.error);
1359 pmap_release_callback(struct vm_page *p, void *data)
1361 struct rb_vm_page_scan_info *info = data;
1363 if (p->pindex == PTDPTDI) {
1367 if (!pmap_release_free_page(info->pmap, p)) {
1371 if (info->object->generation != info->limit) {
1379 * Grow the number of kernel page table entries, if needed.
1383 pmap_growkernel(vm_offset_t addr)
1386 vm_offset_t ptppaddr;
1391 if (kernel_vm_end == 0) {
1392 kernel_vm_end = KERNBASE;
1394 while (pdir_pde(PTD, kernel_vm_end)) {
1395 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1399 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1400 while (kernel_vm_end < addr) {
1401 if (pdir_pde(PTD, kernel_vm_end)) {
1402 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1407 * This index is bogus, but out of the way
1409 nkpg = vm_page_alloc(kptobj, nkpt,
1410 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1412 panic("pmap_growkernel: no memory to grow kernel");
1415 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1416 pmap_zero_page(ptppaddr);
1417 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1418 pdir_pde(PTD, kernel_vm_end) = newpdir;
1419 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1423 * This update must be interlocked with pmap_pinit2.
1425 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1426 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1428 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1429 ~(PAGE_SIZE * NPTEPG - 1);
1435 * Retire the given physical map from service.
1436 * Should only be called if the map contains
1437 * no valid mappings.
1440 pmap_destroy(pmap_t pmap)
1447 count = --pmap->pm_count;
1450 panic("destroying a pmap is not yet implemented");
1455 * Add a reference to the specified pmap.
1458 pmap_reference(pmap_t pmap)
1465 /***************************************************
1466 * page management routines.
1467 ***************************************************/
1470 * free the pv_entry back to the free list. This function may be
1471 * called from an interrupt.
1473 static PMAP_INLINE void
1474 free_pv_entry(pv_entry_t pv)
1481 * get a new pv_entry, allocating a block from the system
1482 * when needed. This function may be called from an interrupt.
1488 if (pv_entry_high_water &&
1489 (pv_entry_count > pv_entry_high_water) &&
1490 (pmap_pagedaemon_waken == 0)) {
1491 pmap_pagedaemon_waken = 1;
1492 wakeup (&vm_pages_needed);
1494 return zalloc(pvzone);
1498 * This routine is very drastic, but can save the system
1506 static int warningdone=0;
1508 if (pmap_pagedaemon_waken == 0)
1511 if (warningdone < 5) {
1512 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1516 for(i = 0; i < vm_page_array_size; i++) {
1517 m = &vm_page_array[i];
1518 if (m->wire_count || m->hold_count || m->busy ||
1519 (m->flags & PG_BUSY))
1523 pmap_pagedaemon_waken = 0;
1528 * If it is the first entry on the list, it is actually
1529 * in the header and we must copy the following entry up
1530 * to the header. Otherwise we must search the list for
1531 * the entry. In either case we free the now unused entry.
1534 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1535 vm_offset_t va, pmap_inval_info_t info)
1541 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1542 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1543 if (pmap == pv->pv_pmap && va == pv->pv_va)
1547 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1548 if (va == pv->pv_va)
1555 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1556 m->md.pv_list_count--;
1557 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1558 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1559 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1560 ++pmap->pm_generation;
1561 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1569 * Create a pv entry for page at pa for
1573 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1578 pv = get_pv_entry();
1583 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1584 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1585 m->md.pv_list_count++;
1591 * pmap_remove_pte: do the things to unmap a page in a process
1594 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1595 pmap_inval_info_t info)
1600 pmap_inval_add(info, pmap, va);
1601 oldpte = loadandclear(ptq);
1603 pmap->pm_stats.wired_count -= 1;
1605 * Machines that don't support invlpg, also don't support
1606 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1610 cpu_invlpg((void *)va);
1611 KKASSERT(pmap->pm_stats.resident_count > 0);
1612 --pmap->pm_stats.resident_count;
1613 if (oldpte & PG_MANAGED) {
1614 m = PHYS_TO_VM_PAGE(oldpte);
1615 if (oldpte & PG_M) {
1616 #if defined(PMAP_DIAGNOSTIC)
1617 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1619 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1623 if (pmap_track_modified(va))
1627 vm_page_flag_set(m, PG_REFERENCED);
1628 return pmap_remove_entry(pmap, m, va, info);
1630 return pmap_unuse_pt(pmap, va, NULL, info);
1639 * Remove a single page from a process address space.
1641 * This function may not be called from an interrupt if the pmap is
1645 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1650 * if there is no pte for this address, just skip it!!! Otherwise
1651 * get a local va for mappings for this pmap and remove the entry.
1653 if (*pmap_pde(pmap, va) != 0) {
1654 ptq = get_ptbase(pmap) + i386_btop(va);
1656 pmap_remove_pte(pmap, ptq, va, info);
1664 * Remove the given range of addresses from the specified map.
1666 * It is assumed that the start and end are properly
1667 * rounded to the page size.
1669 * This function may not be called from an interrupt if the pmap is
1673 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1677 vm_offset_t ptpaddr;
1678 vm_offset_t sindex, eindex;
1679 struct pmap_inval_info info;
1684 if (pmap->pm_stats.resident_count == 0)
1687 pmap_inval_init(&info);
1690 * special handling of removing one page. a very
1691 * common operation and easy to short circuit some
1694 if (((sva + PAGE_SIZE) == eva) &&
1695 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1696 pmap_remove_page(pmap, sva, &info);
1697 pmap_inval_flush(&info);
1702 * Get a local virtual address for the mappings that are being
1705 sindex = i386_btop(sva);
1706 eindex = i386_btop(eva);
1708 for (; sindex < eindex; sindex = pdnxt) {
1712 * Calculate index for next page table.
1714 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1715 if (pmap->pm_stats.resident_count == 0)
1718 pdirindex = sindex / NPDEPG;
1719 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1720 pmap_inval_add(&info, pmap, -1);
1721 pmap->pm_pdir[pdirindex] = 0;
1722 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1727 * Weed out invalid mappings. Note: we assume that the page
1728 * directory table is always allocated, and in kernel virtual.
1734 * Limit our scan to either the end of the va represented
1735 * by the current page table page, or to the end of the
1736 * range being removed.
1738 if (pdnxt > eindex) {
1743 * NOTE: pmap_remove_pte() can block.
1745 for (; sindex != pdnxt; sindex++) {
1748 ptbase = get_ptbase(pmap);
1749 if (ptbase[sindex] == 0)
1751 va = i386_ptob(sindex);
1752 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1756 pmap_inval_flush(&info);
1762 * Removes this physical page from all physical maps in which it resides.
1763 * Reflects back modify bits to the pager.
1765 * This routine may not be called from an interrupt.
1769 pmap_remove_all(vm_page_t m)
1771 struct pmap_inval_info info;
1772 unsigned *pte, tpte;
1775 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
1778 pmap_inval_init(&info);
1780 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1781 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1782 --pv->pv_pmap->pm_stats.resident_count;
1784 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1785 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1787 tpte = loadandclear(pte);
1789 pv->pv_pmap->pm_stats.wired_count--;
1792 vm_page_flag_set(m, PG_REFERENCED);
1795 * Update the vm_page_t clean and reference bits.
1798 #if defined(PMAP_DIAGNOSTIC)
1799 if (pmap_nw_modified((pt_entry_t) tpte)) {
1801 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1805 if (pmap_track_modified(pv->pv_va))
1808 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1809 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1810 ++pv->pv_pmap->pm_generation;
1811 m->md.pv_list_count--;
1812 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1816 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1818 pmap_inval_flush(&info);
1824 * Set the physical protection on the specified range of this map
1827 * This function may not be called from an interrupt if the map is
1828 * not the kernel_pmap.
1831 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1834 vm_offset_t pdnxt, ptpaddr;
1835 vm_pindex_t sindex, eindex;
1836 pmap_inval_info info;
1841 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1842 pmap_remove(pmap, sva, eva);
1846 if (prot & VM_PROT_WRITE)
1849 pmap_inval_init(&info);
1851 ptbase = get_ptbase(pmap);
1853 sindex = i386_btop(sva);
1854 eindex = i386_btop(eva);
1856 for (; sindex < eindex; sindex = pdnxt) {
1860 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1862 pdirindex = sindex / NPDEPG;
1863 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1864 pmap_inval_add(&info, pmap, -1);
1865 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1866 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1871 * Weed out invalid mappings. Note: we assume that the page
1872 * directory table is always allocated, and in kernel virtual.
1877 if (pdnxt > eindex) {
1881 for (; sindex != pdnxt; sindex++) {
1886 /* XXX this isn't optimal */
1887 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1888 pbits = ptbase[sindex];
1890 if (pbits & PG_MANAGED) {
1893 m = PHYS_TO_VM_PAGE(pbits);
1894 vm_page_flag_set(m, PG_REFERENCED);
1898 if (pmap_track_modified(i386_ptob(sindex))) {
1900 m = PHYS_TO_VM_PAGE(pbits);
1909 if (pbits != ptbase[sindex]) {
1910 ptbase[sindex] = pbits;
1914 pmap_inval_flush(&info);
1918 * Insert the given physical page (p) at
1919 * the specified virtual address (v) in the
1920 * target physical map with the protection requested.
1922 * If specified, the page will be wired down, meaning
1923 * that the related pte can not be reclaimed.
1925 * NB: This is the only routine which MAY NOT lazy-evaluate
1926 * or lose information. That is, this routine must actually
1927 * insert this page into the given map NOW.
1930 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1936 vm_offset_t origpte, newpte;
1938 pmap_inval_info info;
1944 #ifdef PMAP_DIAGNOSTIC
1946 panic("pmap_enter: toobig");
1947 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1948 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1950 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
1951 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
1953 db_print_backtrace();
1956 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
1957 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
1959 db_print_backtrace();
1965 * In the case that a page table page is not
1966 * resident, we are creating it here.
1968 if (va < UPT_MIN_ADDRESS) {
1969 mpte = pmap_allocpte(pmap, va);
1972 pmap_inval_init(&info);
1973 pte = pmap_pte(pmap, va);
1976 * Page Directory table entry not valid, we need a new PT page
1979 panic("pmap_enter: invalid page directory pdir=%x, va=0x%x\n",
1980 (unsigned) pmap->pm_pdir[PTDPTDI], va);
1983 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
1984 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1985 origpte = *(vm_offset_t *)pte;
1986 opa = origpte & PG_FRAME;
1988 if (origpte & PG_PS)
1989 panic("pmap_enter: attempted pmap_enter on 4MB page");
1992 * Mapping has not changed, must be protection or wiring change.
1994 if (origpte && (opa == pa)) {
1996 * Wiring change, just update stats. We don't worry about
1997 * wiring PT pages as they remain resident as long as there
1998 * are valid mappings in them. Hence, if a user page is wired,
1999 * the PT page will be also.
2001 if (wired && ((origpte & PG_W) == 0))
2002 pmap->pm_stats.wired_count++;
2003 else if (!wired && (origpte & PG_W))
2004 pmap->pm_stats.wired_count--;
2006 #if defined(PMAP_DIAGNOSTIC)
2007 if (pmap_nw_modified((pt_entry_t) origpte)) {
2009 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
2015 * Remove the extra pte reference. Note that we cannot
2016 * optimize the RO->RW case because we have adjusted the
2017 * wiring count above and may need to adjust the wiring
2024 * We might be turning off write access to the page,
2025 * so we go ahead and sense modify status.
2027 if (origpte & PG_MANAGED) {
2028 if ((origpte & PG_M) && pmap_track_modified(va)) {
2030 om = PHYS_TO_VM_PAGE(opa);
2038 * Mapping has changed, invalidate old range and fall through to
2039 * handle validating new mapping.
2043 err = pmap_remove_pte(pmap, pte, va, &info);
2045 panic("pmap_enter: pte vanished, va: 0x%x", va);
2049 * Enter on the PV list if part of our managed memory. Note that we
2050 * raise IPL while manipulating pv_table since pmap_enter can be
2051 * called at interrupt time.
2053 if (pmap_initialized &&
2054 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2055 pmap_insert_entry(pmap, va, mpte, m);
2060 * Increment counters
2062 ++pmap->pm_stats.resident_count;
2064 pmap->pm_stats.wired_count++;
2068 * Now validate mapping with desired protection/wiring.
2070 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2074 if (va < UPT_MIN_ADDRESS)
2076 if (pmap == &kernel_pmap)
2080 * if the mapping or permission bits are different, we need
2081 * to update the pte.
2083 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2084 *pte = newpte | PG_A;
2086 pmap_inval_flush(&info);
2090 * this code makes some *MAJOR* assumptions:
2091 * 1. Current pmap & pmap exists.
2094 * 4. No page table pages.
2095 * 5. Tlbflush is deferred to calling procedure.
2096 * 6. Page IS managed.
2097 * but is *MUCH* faster than pmap_enter...
2101 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
2105 pmap_inval_info info;
2107 pmap_inval_init(&info);
2109 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2110 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2112 db_print_backtrace();
2115 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2116 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2118 db_print_backtrace();
2123 * In the case that a page table page is not
2124 * resident, we are creating it here.
2126 if (va < UPT_MIN_ADDRESS) {
2131 * Calculate pagetable page index
2133 ptepindex = va >> PDRSHIFT;
2134 if (mpte && (mpte->pindex == ptepindex)) {
2139 * Get the page directory entry
2141 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2144 * If the page table page is mapped, we just increment
2145 * the hold count, and activate it.
2149 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2150 if (pmap->pm_ptphint &&
2151 (pmap->pm_ptphint->pindex == ptepindex)) {
2152 mpte = pmap->pm_ptphint;
2154 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2155 pmap->pm_ptphint = mpte;
2161 mpte = _pmap_allocpte(pmap, ptepindex);
2169 * This call to vtopte makes the assumption that we are
2170 * entering the page into the current pmap. In order to support
2171 * quick entry into any pmap, one would likely use pmap_pte_quick.
2172 * But that isn't as quick as vtopte.
2174 pte = (unsigned *)vtopte(va);
2177 pmap_unwire_pte_hold(pmap, mpte, &info);
2182 * Enter on the PV list if part of our managed memory. Note that we
2183 * raise IPL while manipulating pv_table since pmap_enter can be
2184 * called at interrupt time.
2186 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
2187 pmap_insert_entry(pmap, va, mpte, m);
2190 * Increment counters
2192 ++pmap->pm_stats.resident_count;
2194 pa = VM_PAGE_TO_PHYS(m);
2197 * Now validate mapping with RO protection
2199 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2200 *pte = pa | PG_V | PG_U;
2202 *pte = pa | PG_V | PG_U | PG_MANAGED;
2208 * Make a temporary mapping for a physical address. This is only intended
2209 * to be used for panic dumps.
2212 pmap_kenter_temporary(vm_paddr_t pa, int i)
2214 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2215 return ((void *)crashdumpmap);
2218 #define MAX_INIT_PT (96)
2221 * This routine preloads the ptes for a given object into the specified pmap.
2222 * This eliminates the blast of soft faults on process startup and
2223 * immediately after an mmap.
2225 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2228 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2229 vm_object_t object, vm_pindex_t pindex,
2230 vm_size_t size, int limit)
2232 struct rb_vm_page_scan_info info;
2237 * We can't preinit if read access isn't set or there is no pmap
2240 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2244 * We can't preinit if the pmap is not the current pmap
2246 lp = curthread->td_lwp;
2247 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2250 psize = i386_btop(size);
2252 if ((object->type != OBJT_VNODE) ||
2253 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2254 (object->resident_page_count > MAX_INIT_PT))) {
2258 if (psize + pindex > object->size) {
2259 if (object->size < pindex)
2261 psize = object->size - pindex;
2268 * Use a red-black scan to traverse the requested range and load
2269 * any valid pages found into the pmap.
2271 * We cannot safely scan the object's memq unless we are in a
2272 * critical section since interrupts can remove pages from objects.
2274 info.start_pindex = pindex;
2275 info.end_pindex = pindex + psize - 1;
2282 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2283 pmap_object_init_pt_callback, &info);
2289 pmap_object_init_pt_callback(vm_page_t p, void *data)
2291 struct rb_vm_page_scan_info *info = data;
2292 vm_pindex_t rel_index;
2294 * don't allow an madvise to blow away our really
2295 * free pages allocating pv entries.
2297 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2298 vmstats.v_free_count < vmstats.v_free_reserved) {
2301 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2302 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2303 if ((p->queue - p->pc) == PQ_CACHE)
2304 vm_page_deactivate(p);
2306 rel_index = p->pindex - info->start_pindex;
2307 info->mpte = pmap_enter_quick(info->pmap,
2308 info->addr + i386_ptob(rel_index),
2310 vm_page_flag_set(p, PG_MAPPED);
2317 * pmap_prefault provides a quick way of clustering pagefaults into a
2318 * processes address space. It is a "cousin" of pmap_object_init_pt,
2319 * except it runs at page fault time instead of mmap time.
2323 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2325 static int pmap_prefault_pageorder[] = {
2326 -PAGE_SIZE, PAGE_SIZE,
2327 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2328 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2329 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2333 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2344 * We do not currently prefault mappings that use virtual page
2345 * tables. We do not prefault foreign pmaps.
2347 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2349 lp = curthread->td_lwp;
2350 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2353 object = entry->object.vm_object;
2355 starta = addra - PFBAK * PAGE_SIZE;
2356 if (starta < entry->start)
2357 starta = entry->start;
2358 else if (starta > addra)
2362 * critical section protection is required to maintain the
2363 * page/object association, interrupts can free pages and remove
2364 * them from their objects.
2368 for (i = 0; i < PAGEORDER_SIZE; i++) {
2369 vm_object_t lobject;
2372 addr = addra + pmap_prefault_pageorder[i];
2373 if (addr > addra + (PFFOR * PAGE_SIZE))
2376 if (addr < starta || addr >= entry->end)
2379 if ((*pmap_pde(pmap, addr)) == NULL)
2382 pte = (unsigned *) vtopte(addr);
2386 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2389 for (m = vm_page_lookup(lobject, pindex);
2390 (!m && (lobject->type == OBJT_DEFAULT) &&
2391 (lobject->backing_object));
2392 lobject = lobject->backing_object
2394 if (lobject->backing_object_offset & PAGE_MASK)
2396 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2397 m = vm_page_lookup(lobject->backing_object, pindex);
2401 * give-up when a page is not in memory
2406 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2408 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2410 if ((m->queue - m->pc) == PQ_CACHE) {
2411 vm_page_deactivate(m);
2414 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2415 vm_page_flag_set(m, PG_MAPPED);
2423 * Routine: pmap_change_wiring
2424 * Function: Change the wiring attribute for a map/virtual-address
2426 * In/out conditions:
2427 * The mapping must already exist in the pmap.
2430 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2437 pte = pmap_pte(pmap, va);
2439 if (wired && !pmap_pte_w(pte))
2440 pmap->pm_stats.wired_count++;
2441 else if (!wired && pmap_pte_w(pte))
2442 pmap->pm_stats.wired_count--;
2445 * Wiring is not a hardware characteristic so there is no need to
2446 * invalidate TLB. However, in an SMP environment we must use
2447 * a locked bus cycle to update the pte (if we are not using
2448 * the pmap_inval_*() API that is)... it's ok to do this for simple
2453 atomic_set_int(pte, PG_W);
2455 atomic_clear_int(pte, PG_W);
2458 atomic_set_int_nonlocked(pte, PG_W);
2460 atomic_clear_int_nonlocked(pte, PG_W);
2467 * Copy the range specified by src_addr/len
2468 * from the source map to the range dst_addr/len
2469 * in the destination map.
2471 * This routine is only advisory and need not do anything.
2474 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2475 vm_size_t len, vm_offset_t src_addr)
2477 pmap_inval_info info;
2479 vm_offset_t end_addr = src_addr + len;
2481 unsigned src_frame, dst_frame;
2484 if (dst_addr != src_addr)
2487 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2488 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2492 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2493 if (dst_frame != (((unsigned) APTDpde) & PG_FRAME)) {
2494 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2495 /* The page directory is not shared between CPUs */
2498 pmap_inval_init(&info);
2499 pmap_inval_add(&info, dst_pmap, -1);
2500 pmap_inval_add(&info, src_pmap, -1);
2503 * critical section protection is required to maintain the page/object
2504 * association, interrupts can free pages and remove them from
2508 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2509 unsigned *src_pte, *dst_pte;
2510 vm_page_t dstmpte, srcmpte;
2511 vm_offset_t srcptepaddr;
2514 if (addr >= UPT_MIN_ADDRESS)
2515 panic("pmap_copy: invalid to pmap_copy page tables\n");
2518 * Don't let optional prefaulting of pages make us go
2519 * way below the low water mark of free pages or way
2520 * above high water mark of used pv entries.
2522 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2523 pv_entry_count > pv_entry_high_water)
2526 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2527 ptepindex = addr >> PDRSHIFT;
2529 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2530 if (srcptepaddr == 0)
2533 if (srcptepaddr & PG_PS) {
2534 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2535 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2536 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2541 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2542 if ((srcmpte == NULL) ||
2543 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2546 if (pdnxt > end_addr)
2549 src_pte = (unsigned *) vtopte(addr);
2550 dst_pte = (unsigned *) avtopte(addr);
2551 while (addr < pdnxt) {
2556 * we only virtual copy managed pages
2558 if ((ptetemp & PG_MANAGED) != 0) {
2560 * We have to check after allocpte for the
2561 * pte still being around... allocpte can
2564 * pmap_allocpte() can block. If we lose
2565 * our page directory mappings we stop.
2567 dstmpte = pmap_allocpte(dst_pmap, addr);
2569 if (src_frame != (((unsigned) PTDpde) & PG_FRAME) ||
2570 dst_frame != (((unsigned) APTDpde) & PG_FRAME)
2572 kprintf("WARNING: pmap_copy: detected and corrected race\n");
2573 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2575 } else if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2577 * Clear the modified and
2578 * accessed (referenced) bits
2581 m = PHYS_TO_VM_PAGE(ptetemp);
2582 *dst_pte = ptetemp & ~(PG_M | PG_A);
2583 ++dst_pmap->pm_stats.resident_count;
2584 pmap_insert_entry(dst_pmap, addr,
2587 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2589 if (dstmpte->hold_count >= srcmpte->hold_count)
2599 pmap_inval_flush(&info);
2605 * Zero the specified PA by mapping the page into KVM and clearing its
2608 * This function may be called from an interrupt and no locking is
2612 pmap_zero_page(vm_paddr_t phys)
2614 struct mdglobaldata *gd = mdcpu;
2617 if (*(int *)gd->gd_CMAP3)
2618 panic("pmap_zero_page: CMAP3 busy");
2619 *(int *)gd->gd_CMAP3 =
2620 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2621 cpu_invlpg(gd->gd_CADDR3);
2623 #if defined(I686_CPU)
2624 if (cpu_class == CPUCLASS_686)
2625 i686_pagezero(gd->gd_CADDR3);
2628 bzero(gd->gd_CADDR3, PAGE_SIZE);
2629 *(int *) gd->gd_CMAP3 = 0;
2634 * pmap_page_assertzero:
2636 * Assert that a page is empty, panic if it isn't.
2639 pmap_page_assertzero(vm_paddr_t phys)
2641 struct mdglobaldata *gd = mdcpu;
2645 if (*(int *)gd->gd_CMAP3)
2646 panic("pmap_zero_page: CMAP3 busy");
2647 *(int *)gd->gd_CMAP3 =
2648 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2649 cpu_invlpg(gd->gd_CADDR3);
2650 for (i = 0; i < PAGE_SIZE; i += 4) {
2651 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2652 panic("pmap_page_assertzero() @ %p not zero!\n",
2653 (void *)gd->gd_CADDR3);
2656 *(int *) gd->gd_CMAP3 = 0;
2663 * Zero part of a physical page by mapping it into memory and clearing
2664 * its contents with bzero.
2666 * off and size may not cover an area beyond a single hardware page.
2669 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2671 struct mdglobaldata *gd = mdcpu;
2674 if (*(int *) gd->gd_CMAP3)
2675 panic("pmap_zero_page: CMAP3 busy");
2676 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2677 cpu_invlpg(gd->gd_CADDR3);
2679 #if defined(I686_CPU)
2680 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2681 i686_pagezero(gd->gd_CADDR3);
2684 bzero((char *)gd->gd_CADDR3 + off, size);
2685 *(int *) gd->gd_CMAP3 = 0;
2692 * Copy the physical page from the source PA to the target PA.
2693 * This function may be called from an interrupt. No locking
2697 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2699 struct mdglobaldata *gd = mdcpu;
2702 if (*(int *) gd->gd_CMAP1)
2703 panic("pmap_copy_page: CMAP1 busy");
2704 if (*(int *) gd->gd_CMAP2)
2705 panic("pmap_copy_page: CMAP2 busy");
2707 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2708 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2710 cpu_invlpg(gd->gd_CADDR1);
2711 cpu_invlpg(gd->gd_CADDR2);
2713 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2715 *(int *) gd->gd_CMAP1 = 0;
2716 *(int *) gd->gd_CMAP2 = 0;
2721 * pmap_copy_page_frag:
2723 * Copy the physical page from the source PA to the target PA.
2724 * This function may be called from an interrupt. No locking
2728 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2730 struct mdglobaldata *gd = mdcpu;
2733 if (*(int *) gd->gd_CMAP1)
2734 panic("pmap_copy_page: CMAP1 busy");
2735 if (*(int *) gd->gd_CMAP2)
2736 panic("pmap_copy_page: CMAP2 busy");
2738 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2739 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2741 cpu_invlpg(gd->gd_CADDR1);
2742 cpu_invlpg(gd->gd_CADDR2);
2744 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2745 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2748 *(int *) gd->gd_CMAP1 = 0;
2749 *(int *) gd->gd_CMAP2 = 0;
2754 * Returns true if the pmap's pv is one of the first
2755 * 16 pvs linked to from this page. This count may
2756 * be changed upwards or downwards in the future; it
2757 * is only necessary that true be returned for a small
2758 * subset of pmaps for proper page aging.
2761 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2766 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2771 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2772 if (pv->pv_pmap == pmap) {
2785 * Remove all pages from specified address space
2786 * this aids process exit speeds. Also, this code
2787 * is special cased for current process only, but
2788 * can have the more generic (and slightly slower)
2789 * mode enabled. This is much faster than pmap_remove
2790 * in the case of running down an entire address space.
2793 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2796 unsigned *pte, tpte;
2799 pmap_inval_info info;
2801 int32_t save_generation;
2803 lp = curthread->td_lwp;
2804 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2809 pmap_inval_init(&info);
2811 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2812 if (pv->pv_va >= eva || pv->pv_va < sva) {
2813 npv = TAILQ_NEXT(pv, pv_plist);
2817 KKASSERT(pmap == pv->pv_pmap);
2820 pte = (unsigned *)vtopte(pv->pv_va);
2822 pte = pmap_pte_quick(pmap, pv->pv_va);
2823 if (pmap->pm_active)
2824 pmap_inval_add(&info, pmap, pv->pv_va);
2828 * We cannot remove wired pages from a process' mapping
2832 npv = TAILQ_NEXT(pv, pv_plist);
2837 m = PHYS_TO_VM_PAGE(tpte);
2839 KASSERT(m < &vm_page_array[vm_page_array_size],
2840 ("pmap_remove_pages: bad tpte %x", tpte));
2842 KKASSERT(pmap->pm_stats.resident_count > 0);
2843 --pmap->pm_stats.resident_count;
2846 * Update the vm_page_t clean and reference bits.
2852 npv = TAILQ_NEXT(pv, pv_plist);
2853 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2854 save_generation = ++pmap->pm_generation;
2856 m->md.pv_list_count--;
2857 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2858 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2859 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2862 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2866 * Restart the scan if we blocked during the unuse or free
2867 * calls and other removals were made.
2869 if (save_generation != pmap->pm_generation) {
2870 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2871 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2874 pmap_inval_flush(&info);
2879 * pmap_testbit tests bits in pte's
2880 * note that the testbit/clearbit routines are inline,
2881 * and a lot of things compile-time evaluate.
2884 pmap_testbit(vm_page_t m, int bit)
2889 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2892 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2897 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2899 * if the bit being tested is the modified bit, then
2900 * mark clean_map and ptes as never
2903 if (bit & (PG_A|PG_M)) {
2904 if (!pmap_track_modified(pv->pv_va))
2908 #if defined(PMAP_DIAGNOSTIC)
2910 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2914 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2925 * this routine is used to modify bits in ptes
2927 static __inline void
2928 pmap_clearbit(vm_page_t m, int bit)
2930 struct pmap_inval_info info;
2935 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2938 pmap_inval_init(&info);
2942 * Loop over all current mappings setting/clearing as appropos If
2943 * setting RO do we need to clear the VAC?
2945 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2947 * don't write protect pager mappings
2950 if (!pmap_track_modified(pv->pv_va))
2954 #if defined(PMAP_DIAGNOSTIC)
2956 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2962 * Careful here. We can use a locked bus instruction to
2963 * clear PG_A or PG_M safely but we need to synchronize
2964 * with the target cpus when we mess with PG_RW.
2966 * We do not have to force synchronization when clearing
2967 * PG_M even for PTEs generated via virtual memory maps,
2968 * because the virtual kernel will invalidate the pmap
2969 * entry when/if it needs to resynchronize the Modify bit.
2971 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2973 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2980 atomic_clear_int(pte, PG_M|PG_RW);
2981 } else if (bit == PG_M) {
2983 * We could also clear PG_RW here to force
2984 * a fault on write to redetect PG_M for
2985 * virtual kernels, but it isn't necessary
2986 * since virtual kernels invalidate the pte
2987 * when they clear the VPTE_M bit in their
2988 * virtual page tables.
2990 atomic_clear_int(pte, PG_M);
2992 atomic_clear_int(pte, bit);
2996 pmap_inval_flush(&info);
3001 * pmap_page_protect:
3003 * Lower the permission for all mappings to a given page.
3006 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3008 if ((prot & VM_PROT_WRITE) == 0) {
3009 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3010 pmap_clearbit(m, PG_RW);
3018 pmap_phys_address(int ppn)
3020 return (i386_ptob(ppn));
3024 * pmap_ts_referenced:
3026 * Return a count of reference bits for a page, clearing those bits.
3027 * It is not necessary for every reference bit to be cleared, but it
3028 * is necessary that 0 only be returned when there are truly no
3029 * reference bits set.
3031 * XXX: The exact number of bits to check and clear is a matter that
3032 * should be tested and standardized at some point in the future for
3033 * optimal aging of shared pages.
3036 pmap_ts_referenced(vm_page_t m)
3038 pv_entry_t pv, pvf, pvn;
3042 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3047 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3052 pvn = TAILQ_NEXT(pv, pv_list);
3054 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3056 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3058 if (!pmap_track_modified(pv->pv_va))
3061 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3063 if (pte && (*pte & PG_A)) {
3065 atomic_clear_int(pte, PG_A);
3067 atomic_clear_int_nonlocked(pte, PG_A);
3074 } while ((pv = pvn) != NULL && pv != pvf);
3084 * Return whether or not the specified physical page was modified
3085 * in any physical maps.
3088 pmap_is_modified(vm_page_t m)
3090 return pmap_testbit(m, PG_M);
3094 * Clear the modify bits on the specified physical page.
3097 pmap_clear_modify(vm_page_t m)
3099 pmap_clearbit(m, PG_M);
3103 * pmap_clear_reference:
3105 * Clear the reference bit on the specified physical page.
3108 pmap_clear_reference(vm_page_t m)
3110 pmap_clearbit(m, PG_A);
3114 * Miscellaneous support routines follow
3118 i386_protection_init(void)
3122 kp = protection_codes;
3123 for (prot = 0; prot < 8; prot++) {
3125 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3127 * Read access is also 0. There isn't any execute bit,
3128 * so just make it readable.
3130 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3131 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3132 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3135 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3136 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3137 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3138 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3146 * Map a set of physical memory pages into the kernel virtual
3147 * address space. Return a pointer to where it is mapped. This
3148 * routine is intended to be used for mapping device memory,
3151 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3155 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3157 vm_offset_t va, tmpva, offset;
3160 offset = pa & PAGE_MASK;
3161 size = roundup(offset + size, PAGE_SIZE);
3163 va = kmem_alloc_nofault(&kernel_map, size);
3165 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3168 for (tmpva = va; size > 0;) {
3169 pte = (unsigned *)vtopte(tmpva);
3170 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3178 return ((void *)(va + offset));
3182 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3184 vm_offset_t base, offset;
3186 base = va & PG_FRAME;
3187 offset = va & PAGE_MASK;
3188 size = roundup(offset + size, PAGE_SIZE);
3189 pmap_qremove(va, size >> PAGE_SHIFT);
3190 kmem_free(&kernel_map, base, size);
3194 * perform the pmap work for mincore
3197 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3199 unsigned *ptep, pte;
3203 ptep = pmap_pte(pmap, addr);
3208 if ((pte = *ptep) != 0) {
3211 val = MINCORE_INCORE;
3212 if ((pte & PG_MANAGED) == 0)
3215 pa = pte & PG_FRAME;
3217 m = PHYS_TO_VM_PAGE(pa);
3223 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3225 * Modified by someone
3227 else if (m->dirty || pmap_is_modified(m))
3228 val |= MINCORE_MODIFIED_OTHER;
3233 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3236 * Referenced by someone
3238 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3239 val |= MINCORE_REFERENCED_OTHER;
3240 vm_page_flag_set(m, PG_REFERENCED);
3247 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3248 * vmspace will be ref'd and the old one will be deref'd.
3250 * The vmspace for all lwps associated with the process will be adjusted
3251 * and cr3 will be reloaded if any lwp is the current lwp.
3254 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3256 struct vmspace *oldvm;
3260 oldvm = p->p_vmspace;
3261 if (oldvm != newvm) {
3262 p->p_vmspace = newvm;
3263 KKASSERT(p->p_nthreads == 1);
3264 lp = RB_ROOT(&p->p_lwp_tree);
3265 pmap_setlwpvm(lp, newvm);
3267 sysref_get(&newvm->vm_sysref);
3268 sysref_put(&oldvm->vm_sysref);
3275 * Set the vmspace for a LWP. The vmspace is almost universally set the
3276 * same as the process vmspace, but virtual kernels need to swap out contexts
3277 * on a per-lwp basis.
3280 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3282 struct vmspace *oldvm;
3286 oldvm = lp->lwp_vmspace;
3288 if (oldvm != newvm) {
3289 lp->lwp_vmspace = newvm;
3290 if (curthread->td_lwp == lp) {
3291 pmap = vmspace_pmap(newvm);
3293 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3295 pmap->pm_active |= 1;
3297 #if defined(SWTCH_OPTIM_STATS)
3300 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3301 load_cr3(curthread->td_pcb->pcb_cr3);
3302 pmap = vmspace_pmap(oldvm);
3304 atomic_clear_int(&pmap->pm_active,
3305 1 << mycpu->gd_cpuid);
3307 pmap->pm_active &= ~1;
3315 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3318 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3322 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3329 static void pads (pmap_t pm);
3330 void pmap_pvdump (vm_paddr_t pa);
3332 /* print address space of pmap*/
3339 if (pm == &kernel_pmap)
3342 for (i = 0; i < 1024; i++) {
3343 if (pm->pm_pdir[i]) {
3344 for (j = 0; j < 1024; j++) {
3345 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3346 if (pm == &kernel_pmap && va < KERNBASE)
3348 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3350 ptep = pmap_pte_quick(pm, va);
3351 if (pmap_pte_v(ptep))
3352 kprintf("%x:%x ", va, *(int *) ptep);
3361 pmap_pvdump(vm_paddr_t pa)
3366 kprintf("pa %08llx", (long long)pa);
3367 m = PHYS_TO_VM_PAGE(pa);
3368 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3370 kprintf(" -> pmap %p, va %x, flags %x",
3371 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3373 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);