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 1.73 2007/02/03 17:05:58 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>
100 #include <machine/cputypes.h>
101 #include <machine/md_var.h>
102 #include <machine/specialreg.h>
103 #include <machine/smp.h>
104 #include <machine_base/apic/apicreg.h>
105 #include <machine/globaldata.h>
106 #include <machine/pmap.h>
107 #include <machine/pmap_inval.h>
111 #define PMAP_KEEP_PDIRS
112 #ifndef PMAP_SHPGPERPROC
113 #define PMAP_SHPGPERPROC 200
116 #if defined(DIAGNOSTIC)
117 #define PMAP_DIAGNOSTIC
122 #if !defined(PMAP_DIAGNOSTIC)
123 #define PMAP_INLINE __inline
129 * Get PDEs and PTEs for user/kernel address space
131 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
132 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
134 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
135 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
136 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
137 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
138 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
142 * Given a map and a machine independent protection code,
143 * convert to a vax protection code.
145 #define pte_prot(m, p) \
146 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
147 static int protection_codes[8];
149 struct pmap kernel_pmap;
150 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
152 vm_paddr_t avail_start; /* PA of first available physical page */
153 vm_paddr_t avail_end; /* PA of last available physical page */
154 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
155 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
156 vm_offset_t KvaStart; /* VA start of KVA space */
157 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
158 vm_offset_t KvaSize; /* max size of kernel virtual address space */
159 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
160 static int pgeflag; /* PG_G or-in */
161 static int pseflag; /* PG_PS or-in */
163 static vm_object_t kptobj;
166 vm_offset_t kernel_vm_end;
169 * Data for the pv entry allocation mechanism
171 static vm_zone_t pvzone;
172 static struct vm_zone pvzone_store;
173 static struct vm_object pvzone_obj;
174 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
175 static int pmap_pagedaemon_waken = 0;
176 static struct pv_entry *pvinit;
179 * All those kernel PT submaps that BSD is so fond of
181 pt_entry_t *CMAP1 = 0, *ptmmap;
182 caddr_t CADDR1 = 0, ptvmmap = 0;
183 static pt_entry_t *msgbufmap;
184 struct msgbuf *msgbufp=0;
189 static pt_entry_t *pt_crashdumpmap;
190 static caddr_t crashdumpmap;
192 extern pt_entry_t *SMPpt;
194 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
195 static unsigned * get_ptbase (pmap_t pmap);
196 static pv_entry_t get_pv_entry (void);
197 static void i386_protection_init (void);
198 static __inline void pmap_clearbit (vm_page_t m, int bit);
200 static void pmap_remove_all (vm_page_t m);
201 static vm_page_t pmap_enter_quick (pmap_t pmap, vm_offset_t va,
202 vm_page_t m, vm_page_t mpte);
203 static int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
204 vm_offset_t sva, pmap_inval_info_t info);
205 static void pmap_remove_page (struct pmap *pmap,
206 vm_offset_t va, pmap_inval_info_t info);
207 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
208 vm_offset_t va, pmap_inval_info_t info);
209 static boolean_t pmap_testbit (vm_page_t m, int bit);
210 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
211 vm_page_t mpte, vm_page_t m);
213 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
215 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
216 static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
217 static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
218 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
219 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
220 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
222 static unsigned pdir4mb;
225 * Move the kernel virtual free pointer to the next
226 * 4MB. This is used to help improve performance
227 * by using a large (4MB) page for much of the kernel
228 * (.text, .data, .bss)
231 pmap_kmem_choose(vm_offset_t addr)
233 vm_offset_t newaddr = addr;
235 if (cpu_feature & CPUID_PSE) {
236 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
245 * Extract the page table entry associated with the given map/virtual
248 * This function may NOT be called from an interrupt.
250 PMAP_INLINE unsigned *
251 pmap_pte(pmap_t pmap, vm_offset_t va)
256 pdeaddr = (unsigned *) pmap_pde(pmap, va);
257 if (*pdeaddr & PG_PS)
260 return get_ptbase(pmap) + i386_btop(va);
269 * Super fast pmap_pte routine best used when scanning the pv lists.
270 * This eliminates many course-grained invltlb calls. Note that many of
271 * the pv list scans are across different pmaps and it is very wasteful
272 * to do an entire invltlb when checking a single mapping.
274 * Should only be called while in a critical section.
277 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
279 struct mdglobaldata *gd = mdcpu;
282 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
283 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
284 unsigned index = i386_btop(va);
285 /* are we current address space or kernel? */
286 if ((pmap == &kernel_pmap) ||
287 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
288 return (unsigned *) PTmap + index;
290 newpf = pde & PG_FRAME;
291 if ( ((* (unsigned *) gd->gd_PMAP1) & PG_FRAME) != newpf) {
292 * (unsigned *) gd->gd_PMAP1 = newpf | PG_RW | PG_V;
293 cpu_invlpg(gd->gd_PADDR1);
295 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
302 * Bootstrap the system enough to run with virtual memory.
304 * On the i386 this is called after mapping has already been enabled
305 * and just syncs the pmap module with what has already been done.
306 * [We can't call it easily with mapping off since the kernel is not
307 * mapped with PA == VA, hence we would have to relocate every address
308 * from the linked base (virtual) address "KERNBASE" to the actual
309 * (physical) address starting relative to 0]
312 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
316 struct mdglobaldata *gd;
320 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
321 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
322 KvaEnd = KvaStart + KvaSize;
324 avail_start = firstaddr;
327 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
328 * too large. It should instead be correctly calculated in locore.s and
329 * not based on 'first' (which is a physical address, not a virtual
330 * address, for the start of unused physical memory). The kernel
331 * page tables are NOT double mapped and thus should not be included
332 * in this calculation.
334 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
335 virtual_start = pmap_kmem_choose(virtual_start);
336 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
339 * Initialize protection array.
341 i386_protection_init();
344 * The kernel's pmap is statically allocated so we don't have to use
345 * pmap_create, which is unlikely to work correctly at this part of
346 * the boot sequence (XXX and which no longer exists).
348 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
349 kernel_pmap.pm_count = 1;
350 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
351 TAILQ_INIT(&kernel_pmap.pm_pvlist);
355 * Reserve some special page table entries/VA space for temporary
358 #define SYSMAP(c, p, v, n) \
359 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
362 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
365 * CMAP1/CMAP2 are used for zeroing and copying pages.
367 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
372 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
375 * ptvmmap is used for reading arbitrary physical pages via
378 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
381 * msgbufp is used to map the system message buffer.
382 * XXX msgbufmap is not used.
384 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
385 atop(round_page(MSGBUF_SIZE)))
390 for (i = 0; i < NKPT; i++)
394 * PG_G is terribly broken on SMP because we IPI invltlb's in some
395 * cases rather then invl1pg. Actually, I don't even know why it
396 * works under UP because self-referential page table mappings
401 if (cpu_feature & CPUID_PGE)
406 * Initialize the 4MB page size flag
410 * The 4MB page version of the initial
411 * kernel page mapping.
415 #if !defined(DISABLE_PSE)
416 if (cpu_feature & CPUID_PSE) {
419 * Note that we have enabled PSE mode
422 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
423 ptditmp &= ~(NBPDR - 1);
424 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
429 * Enable the PSE mode. If we are SMP we can't do this
430 * now because the APs will not be able to use it when
433 load_cr4(rcr4() | CR4_PSE);
436 * We can do the mapping here for the single processor
437 * case. We simply ignore the old page table page from
441 * For SMP, we still need 4K pages to bootstrap APs,
442 * PSE will be enabled as soon as all APs are up.
444 PTD[KPTDI] = (pd_entry_t)ptditmp;
445 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
451 if (cpu_apic_address == 0)
452 panic("pmap_bootstrap: no local apic!");
454 /* local apic is mapped on last page */
455 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag |
456 (cpu_apic_address & PG_FRAME));
460 * We need to finish setting up the globaldata page for the BSP.
461 * locore has already populated the page table for the mdglobaldata
464 pg = MDGLOBALDATA_BASEALLOC_PAGES;
465 gd = &CPU_prvspace[0].mdglobaldata;
466 gd->gd_CMAP1 = &SMPpt[pg + 0];
467 gd->gd_CMAP2 = &SMPpt[pg + 1];
468 gd->gd_CMAP3 = &SMPpt[pg + 2];
469 gd->gd_PMAP1 = &SMPpt[pg + 3];
470 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
471 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
472 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
473 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
480 * Set 4mb pdir for mp startup
485 if (pseflag && (cpu_feature & CPUID_PSE)) {
486 load_cr4(rcr4() | CR4_PSE);
487 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
488 kernel_pmap.pm_pdir[KPTDI] =
489 PTD[KPTDI] = (pd_entry_t)pdir4mb;
497 * Initialize the pmap module.
498 * Called by vm_init, to initialize any structures that the pmap
499 * system needs to map virtual memory.
500 * pmap_init has been enhanced to support in a fairly consistant
501 * way, discontiguous physical memory.
510 * object for kernel page table pages
512 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
515 * Allocate memory for random pmap data structures. Includes the
519 for(i = 0; i < vm_page_array_size; i++) {
522 m = &vm_page_array[i];
523 TAILQ_INIT(&m->md.pv_list);
524 m->md.pv_list_count = 0;
528 * init the pv free list
530 initial_pvs = vm_page_array_size;
531 if (initial_pvs < MINPV)
533 pvzone = &pvzone_store;
534 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
535 initial_pvs * sizeof (struct pv_entry));
536 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
540 * Now it is safe to enable pv_table recording.
542 pmap_initialized = TRUE;
546 * Initialize the address space (zone) for the pv_entries. Set a
547 * high water mark so that the system can recover from excessive
548 * numbers of pv entries.
553 int shpgperproc = PMAP_SHPGPERPROC;
555 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
556 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
557 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
558 pv_entry_high_water = 9 * (pv_entry_max / 10);
559 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
563 /***************************************************
564 * Low level helper routines.....
565 ***************************************************/
567 #if defined(PMAP_DIAGNOSTIC)
570 * This code checks for non-writeable/modified pages.
571 * This should be an invalid condition.
574 pmap_nw_modified(pt_entry_t ptea)
580 if ((pte & (PG_M|PG_RW)) == PG_M)
589 * this routine defines the region(s) of memory that should
590 * not be tested for the modified bit.
592 static PMAP_INLINE int
593 pmap_track_modified(vm_offset_t va)
595 if ((va < clean_sva) || (va >= clean_eva))
602 get_ptbase(pmap_t pmap)
604 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
605 struct globaldata *gd = mycpu;
607 /* are we current address space or kernel? */
608 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
609 return (unsigned *) PTmap;
612 /* otherwise, we are alternate address space */
613 KKASSERT(gd->gd_intr_nesting_level == 0 &&
614 (gd->gd_curthread->td_flags & TDF_INTTHREAD) == 0);
616 if (frame != (((unsigned) APTDpde) & PG_FRAME)) {
617 APTDpde = (pd_entry_t)(frame | PG_RW | PG_V);
618 /* The page directory is not shared between CPUs */
621 return (unsigned *) APTmap;
627 * Extract the physical page address associated with the map/VA pair.
629 * This function may not be called from an interrupt if the pmap is
633 pmap_extract(pmap_t pmap, vm_offset_t va)
636 vm_offset_t pdirindex;
638 pdirindex = va >> PDRSHIFT;
639 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
641 if ((rtval & PG_PS) != 0) {
642 rtval &= ~(NBPDR - 1);
643 rtval |= va & (NBPDR - 1);
646 pte = get_ptbase(pmap) + i386_btop(va);
647 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
653 /***************************************************
654 * Low level mapping routines.....
655 ***************************************************/
658 * Routine: pmap_kenter
660 * Add a wired page to the KVA
661 * NOTE! note that in order for the mapping to take effect -- you
662 * should do an invltlb after doing the pmap_kenter().
665 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
669 pmap_inval_info info;
671 pmap_inval_init(&info);
672 pmap_inval_add(&info, &kernel_pmap, va);
673 npte = pa | PG_RW | PG_V | pgeflag;
674 pte = (unsigned *)vtopte(va);
676 pmap_inval_flush(&info);
680 * Routine: pmap_kenter_quick
682 * Similar to pmap_kenter(), except we only invalidate the
683 * mapping on the current CPU.
686 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
691 npte = pa | PG_RW | PG_V | pgeflag;
692 pte = (unsigned *)vtopte(va);
694 cpu_invlpg((void *)va);
698 pmap_kenter_sync(vm_offset_t va)
700 pmap_inval_info info;
702 pmap_inval_init(&info);
703 pmap_inval_add(&info, &kernel_pmap, va);
704 pmap_inval_flush(&info);
708 pmap_kenter_sync_quick(vm_offset_t va)
710 cpu_invlpg((void *)va);
714 * remove a page from the kernel pagetables
717 pmap_kremove(vm_offset_t va)
720 pmap_inval_info info;
722 pmap_inval_init(&info);
723 pmap_inval_add(&info, &kernel_pmap, va);
724 pte = (unsigned *)vtopte(va);
726 pmap_inval_flush(&info);
730 pmap_kremove_quick(vm_offset_t va)
733 pte = (unsigned *)vtopte(va);
735 cpu_invlpg((void *)va);
739 * XXX these need to be recoded. They are not used in any critical path.
742 pmap_kmodify_rw(vm_offset_t va)
744 *vtopte(va) |= PG_RW;
745 cpu_invlpg((void *)va);
749 pmap_kmodify_nc(vm_offset_t va)
752 cpu_invlpg((void *)va);
756 * Used to map a range of physical addresses into kernel
757 * virtual address space.
759 * For now, VM is already on, we only need to map the
763 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
765 while (start < end) {
766 pmap_kenter(virt, start);
775 * Add a list of wired pages to the kva
776 * this routine is only used for temporary
777 * kernel mappings that do not need to have
778 * page modification or references recorded.
779 * Note that old mappings are simply written
780 * over. The page *must* be wired.
783 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
787 end_va = va + count * PAGE_SIZE;
789 while (va < end_va) {
792 pte = (unsigned *)vtopte(va);
793 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
794 cpu_invlpg((void *)va);
799 smp_invltlb(); /* XXX */
804 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
807 cpumask_t cmask = mycpu->gd_cpumask;
809 end_va = va + count * PAGE_SIZE;
811 while (va < end_va) {
816 * Install the new PTE. If the pte changed from the prior
817 * mapping we must reset the cpu mask and invalidate the page.
818 * If the pte is the same but we have not seen it on the
819 * current cpu, invlpg the existing mapping. Otherwise the
820 * entry is optimal and no invalidation is required.
822 pte = (unsigned *)vtopte(va);
823 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
824 if (*pte != pteval) {
827 cpu_invlpg((void *)va);
828 } else if ((*mask & cmask) == 0) {
829 cpu_invlpg((void *)va);
838 * this routine jerks page mappings from the
839 * kernel -- it is meant only for temporary mappings.
842 pmap_qremove(vm_offset_t va, int count)
846 end_va = va + count*PAGE_SIZE;
848 while (va < end_va) {
851 pte = (unsigned *)vtopte(va);
853 cpu_invlpg((void *)va);
862 * This routine works like vm_page_lookup() but also blocks as long as the
863 * page is busy. This routine does not busy the page it returns.
865 * Unless the caller is managing objects whos pages are in a known state,
866 * the call should be made with a critical section held so the page's object
867 * association remains valid on return.
870 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
875 m = vm_page_lookup(object, pindex);
876 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
882 * Create a new thread and optionally associate it with a (new) process.
883 * NOTE! the new thread's cpu may not equal the current cpu.
886 pmap_init_thread(thread_t td)
888 /* enforce pcb placement */
889 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
890 td->td_savefpu = &td->td_pcb->pcb_save;
891 td->td_sp = (char *)td->td_pcb - 16;
895 * Create the UPAGES for a new process.
896 * This routine directly affects the fork perf for a process.
899 pmap_init_proc(struct proc *p, struct thread *td)
901 struct lwp *lp = ONLY_LWP_IN_PROC(p);
903 p->p_addr = (void *)td->td_kstack;
907 td->td_switch = cpu_heavy_switch;
909 KKASSERT(td->td_mpcount == 1);
911 bzero(p->p_addr, sizeof(*p->p_addr));
915 * Dispose the UPAGES for a process that has exited.
916 * This routine directly impacts the exit perf of a process.
919 pmap_dispose_proc(struct proc *p)
921 struct thread *td = NULL;
924 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
926 lp = ONLY_LWP_IN_PROC(p);
927 if (lp != NULL && (td = lp->lwp_thread) != NULL) {
930 lp->lwp_thread = NULL;
936 /***************************************************
937 * Page table page management routines.....
938 ***************************************************/
941 * This routine unholds page table pages, and if the hold count
942 * drops to zero, then it decrements the wire count.
945 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
947 pmap_inval_flush(info);
948 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
951 if (m->hold_count == 0) {
953 * unmap the page table page
955 pmap_inval_add(info, pmap, -1);
956 pmap->pm_pdir[m->pindex] = 0;
957 --pmap->pm_stats.resident_count;
959 if (pmap->pm_ptphint == m)
960 pmap->pm_ptphint = NULL;
963 * If the page is finally unwired, simply free it.
966 if (m->wire_count == 0) {
969 vm_page_free_zero(m);
970 --vmstats.v_wire_count;
977 static PMAP_INLINE int
978 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
981 if (m->hold_count == 0)
982 return _pmap_unwire_pte_hold(pmap, m, info);
988 * After removing a page table entry, this routine is used to
989 * conditionally free the page, and manage the hold/wire counts.
992 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
993 pmap_inval_info_t info)
996 if (va >= UPT_MIN_ADDRESS)
1000 ptepindex = (va >> PDRSHIFT);
1001 if (pmap->pm_ptphint &&
1002 (pmap->pm_ptphint->pindex == ptepindex)) {
1003 mpte = pmap->pm_ptphint;
1005 pmap_inval_flush(info);
1006 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1007 pmap->pm_ptphint = mpte;
1011 return pmap_unwire_pte_hold(pmap, mpte, info);
1015 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1016 * it, and IdlePTD, represents the template used to update all other pmaps.
1018 * On architectures where the kernel pmap is not integrated into the user
1019 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1020 * kernel_pmap should be used to directly access the kernel_pmap.
1023 pmap_pinit0(struct pmap *pmap)
1026 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1027 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1029 pmap->pm_active = 0;
1030 pmap->pm_ptphint = NULL;
1031 TAILQ_INIT(&pmap->pm_pvlist);
1032 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1036 * Initialize a preallocated and zeroed pmap structure,
1037 * such as one in a vmspace structure.
1040 pmap_pinit(struct pmap *pmap)
1045 * No need to allocate page table space yet but we do need a valid
1046 * page directory table.
1048 if (pmap->pm_pdir == NULL) {
1050 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1054 * allocate object for the ptes
1056 if (pmap->pm_pteobj == NULL)
1057 pmap->pm_pteobj = vm_object_allocate( OBJT_DEFAULT, PTDPTDI + 1);
1060 * allocate the page directory page
1062 ptdpg = vm_page_grab( pmap->pm_pteobj, PTDPTDI,
1063 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1065 ptdpg->wire_count = 1;
1066 ++vmstats.v_wire_count;
1069 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY); /* not usually mapped*/
1070 ptdpg->valid = VM_PAGE_BITS_ALL;
1072 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1073 if ((ptdpg->flags & PG_ZERO) == 0)
1074 bzero(pmap->pm_pdir, PAGE_SIZE);
1076 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1078 /* install self-referential address mapping entry */
1079 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1080 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1083 pmap->pm_active = 0;
1084 pmap->pm_ptphint = NULL;
1085 TAILQ_INIT(&pmap->pm_pvlist);
1086 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1090 * Wire in kernel global address entries. To avoid a race condition
1091 * between pmap initialization and pmap_growkernel, this procedure
1092 * adds the pmap to the master list (which growkernel scans to update),
1093 * then copies the template.
1096 pmap_pinit2(struct pmap *pmap)
1099 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1100 /* XXX copies current process, does not fill in MPPTDI */
1101 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1106 * Attempt to release and free and vm_page in a pmap. Returns 1 on success,
1107 * 0 on failure (if the procedure had to sleep).
1110 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1112 unsigned *pde = (unsigned *) pmap->pm_pdir;
1114 * This code optimizes the case of freeing non-busy
1115 * page-table pages. Those pages are zero now, and
1116 * might as well be placed directly into the zero queue.
1118 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1124 * Remove the page table page from the processes address space.
1127 pmap->pm_stats.resident_count--;
1129 if (p->hold_count) {
1130 panic("pmap_release: freeing held page table page");
1133 * Page directory pages need to have the kernel
1134 * stuff cleared, so they can go into the zero queue also.
1136 if (p->pindex == PTDPTDI) {
1137 bzero(pde + KPTDI, nkpt * PTESIZE);
1140 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1143 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1144 pmap->pm_ptphint = NULL;
1147 vmstats.v_wire_count--;
1148 vm_page_free_zero(p);
1153 * this routine is called if the page table page is not
1157 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1159 vm_offset_t pteva, ptepa;
1163 * Find or fabricate a new pagetable page
1165 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1166 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1168 KASSERT(m->queue == PQ_NONE,
1169 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1171 if (m->wire_count == 0)
1172 vmstats.v_wire_count++;
1176 * Increment the hold count for the page table page
1177 * (denoting a new mapping.)
1182 * Map the pagetable page into the process address space, if
1183 * it isn't already there.
1186 pmap->pm_stats.resident_count++;
1188 ptepa = VM_PAGE_TO_PHYS(m);
1189 pmap->pm_pdir[ptepindex] =
1190 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1193 * Set the page table hint
1195 pmap->pm_ptphint = m;
1198 * Try to use the new mapping, but if we cannot, then
1199 * do it with the routine that maps the page explicitly.
1201 if ((m->flags & PG_ZERO) == 0) {
1202 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1203 (((unsigned) PTDpde) & PG_FRAME)) {
1204 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1205 bzero((caddr_t) pteva, PAGE_SIZE);
1207 pmap_zero_page(ptepa);
1211 m->valid = VM_PAGE_BITS_ALL;
1212 vm_page_flag_clear(m, PG_ZERO);
1213 vm_page_flag_set(m, PG_MAPPED);
1220 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1227 * Calculate pagetable page index
1229 ptepindex = va >> PDRSHIFT;
1232 * Get the page directory entry
1234 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1237 * This supports switching from a 4MB page to a
1240 if (ptepa & PG_PS) {
1241 pmap->pm_pdir[ptepindex] = 0;
1248 * If the page table page is mapped, we just increment the
1249 * hold count, and activate it.
1253 * In order to get the page table page, try the
1256 if (pmap->pm_ptphint &&
1257 (pmap->pm_ptphint->pindex == ptepindex)) {
1258 m = pmap->pm_ptphint;
1260 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1261 pmap->pm_ptphint = m;
1267 * Here if the pte page isn't mapped, or if it has been deallocated.
1269 return _pmap_allocpte(pmap, ptepindex);
1273 /***************************************************
1274 * Pmap allocation/deallocation routines.
1275 ***************************************************/
1278 * Release any resources held by the given physical map.
1279 * Called when a pmap initialized by pmap_pinit is being released.
1280 * Should only be called if the map contains no valid mappings.
1282 static int pmap_release_callback(struct vm_page *p, void *data);
1285 pmap_release(struct pmap *pmap)
1287 vm_object_t object = pmap->pm_pteobj;
1288 struct rb_vm_page_scan_info info;
1290 #if defined(DIAGNOSTIC)
1291 if (object->ref_count != 1)
1292 panic("pmap_release: pteobj reference count != 1");
1296 info.object = object;
1298 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1305 info.limit = object->generation;
1307 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1308 pmap_release_callback, &info);
1309 if (info.error == 0 && info.mpte) {
1310 if (!pmap_release_free_page(pmap, info.mpte))
1314 } while (info.error);
1318 pmap_release_callback(struct vm_page *p, void *data)
1320 struct rb_vm_page_scan_info *info = data;
1322 if (p->pindex == PTDPTDI) {
1326 if (!pmap_release_free_page(info->pmap, p)) {
1330 if (info->object->generation != info->limit) {
1338 * Grow the number of kernel page table entries, if needed.
1342 pmap_growkernel(vm_offset_t addr)
1345 vm_offset_t ptppaddr;
1350 if (kernel_vm_end == 0) {
1351 kernel_vm_end = KERNBASE;
1353 while (pdir_pde(PTD, kernel_vm_end)) {
1354 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1358 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1359 while (kernel_vm_end < addr) {
1360 if (pdir_pde(PTD, kernel_vm_end)) {
1361 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1366 * This index is bogus, but out of the way
1368 nkpg = vm_page_alloc(kptobj, nkpt,
1369 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1371 panic("pmap_growkernel: no memory to grow kernel");
1374 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1375 pmap_zero_page(ptppaddr);
1376 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1377 pdir_pde(PTD, kernel_vm_end) = newpdir;
1378 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1382 * This update must be interlocked with pmap_pinit2.
1384 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1385 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1387 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1388 ~(PAGE_SIZE * NPTEPG - 1);
1394 * Retire the given physical map from service.
1395 * Should only be called if the map contains
1396 * no valid mappings.
1399 pmap_destroy(pmap_t pmap)
1406 count = --pmap->pm_count;
1409 panic("destroying a pmap is not yet implemented");
1414 * Add a reference to the specified pmap.
1417 pmap_reference(pmap_t pmap)
1424 /***************************************************
1425 * page management routines.
1426 ***************************************************/
1429 * free the pv_entry back to the free list. This function may be
1430 * called from an interrupt.
1432 static PMAP_INLINE void
1433 free_pv_entry(pv_entry_t pv)
1440 * get a new pv_entry, allocating a block from the system
1441 * when needed. This function may be called from an interrupt.
1447 if (pv_entry_high_water &&
1448 (pv_entry_count > pv_entry_high_water) &&
1449 (pmap_pagedaemon_waken == 0)) {
1450 pmap_pagedaemon_waken = 1;
1451 wakeup (&vm_pages_needed);
1453 return zalloc(pvzone);
1457 * This routine is very drastic, but can save the system
1465 static int warningdone=0;
1467 if (pmap_pagedaemon_waken == 0)
1470 if (warningdone < 5) {
1471 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1475 for(i = 0; i < vm_page_array_size; i++) {
1476 m = &vm_page_array[i];
1477 if (m->wire_count || m->hold_count || m->busy ||
1478 (m->flags & PG_BUSY))
1482 pmap_pagedaemon_waken = 0;
1487 * If it is the first entry on the list, it is actually
1488 * in the header and we must copy the following entry up
1489 * to the header. Otherwise we must search the list for
1490 * the entry. In either case we free the now unused entry.
1493 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1494 vm_offset_t va, pmap_inval_info_t info)
1500 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1501 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1502 if (pmap == pv->pv_pmap && va == pv->pv_va)
1506 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1507 if (va == pv->pv_va)
1514 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1515 m->md.pv_list_count--;
1516 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1517 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1518 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1519 ++pmap->pm_generation;
1520 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1528 * Create a pv entry for page at pa for
1532 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1537 pv = get_pv_entry();
1542 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1543 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1544 m->md.pv_list_count++;
1550 * pmap_remove_pte: do the things to unmap a page in a process
1553 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1554 pmap_inval_info_t info)
1559 pmap_inval_add(info, pmap, va);
1560 oldpte = loadandclear(ptq);
1562 pmap->pm_stats.wired_count -= 1;
1564 * Machines that don't support invlpg, also don't support
1565 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1569 cpu_invlpg((void *)va);
1570 pmap->pm_stats.resident_count -= 1;
1571 if (oldpte & PG_MANAGED) {
1572 m = PHYS_TO_VM_PAGE(oldpte);
1573 if (oldpte & PG_M) {
1574 #if defined(PMAP_DIAGNOSTIC)
1575 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1577 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1581 if (pmap_track_modified(va))
1585 vm_page_flag_set(m, PG_REFERENCED);
1586 return pmap_remove_entry(pmap, m, va, info);
1588 return pmap_unuse_pt(pmap, va, NULL, info);
1597 * Remove a single page from a process address space.
1599 * This function may not be called from an interrupt if the pmap is
1603 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1608 * if there is no pte for this address, just skip it!!! Otherwise
1609 * get a local va for mappings for this pmap and remove the entry.
1611 if (*pmap_pde(pmap, va) != 0) {
1612 ptq = get_ptbase(pmap) + i386_btop(va);
1614 pmap_remove_pte(pmap, ptq, va, info);
1622 * Remove the given range of addresses from the specified map.
1624 * It is assumed that the start and end are properly
1625 * rounded to the page size.
1627 * This function may not be called from an interrupt if the pmap is
1631 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1635 vm_offset_t ptpaddr;
1636 vm_offset_t sindex, eindex;
1637 struct pmap_inval_info info;
1642 if (pmap->pm_stats.resident_count == 0)
1645 pmap_inval_init(&info);
1648 * special handling of removing one page. a very
1649 * common operation and easy to short circuit some
1652 if (((sva + PAGE_SIZE) == eva) &&
1653 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1654 pmap_remove_page(pmap, sva, &info);
1655 pmap_inval_flush(&info);
1660 * Get a local virtual address for the mappings that are being
1663 sindex = i386_btop(sva);
1664 eindex = i386_btop(eva);
1666 for (; sindex < eindex; sindex = pdnxt) {
1670 * Calculate index for next page table.
1672 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1673 if (pmap->pm_stats.resident_count == 0)
1676 pdirindex = sindex / NPDEPG;
1677 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1678 pmap_inval_add(&info, pmap, -1);
1679 pmap->pm_pdir[pdirindex] = 0;
1680 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1685 * Weed out invalid mappings. Note: we assume that the page
1686 * directory table is always allocated, and in kernel virtual.
1692 * Limit our scan to either the end of the va represented
1693 * by the current page table page, or to the end of the
1694 * range being removed.
1696 if (pdnxt > eindex) {
1701 * NOTE: pmap_remove_pte() can block.
1703 for (; sindex != pdnxt; sindex++) {
1706 ptbase = get_ptbase(pmap);
1707 if (ptbase[sindex] == 0)
1709 va = i386_ptob(sindex);
1710 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1714 pmap_inval_flush(&info);
1720 * Removes this physical page from all physical maps in which it resides.
1721 * Reflects back modify bits to the pager.
1723 * This routine may not be called from an interrupt.
1727 pmap_remove_all(vm_page_t m)
1729 struct pmap_inval_info info;
1730 unsigned *pte, tpte;
1733 #if defined(PMAP_DIAGNOSTIC)
1735 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1738 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1739 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
1743 pmap_inval_init(&info);
1745 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1746 pv->pv_pmap->pm_stats.resident_count--;
1748 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1749 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1751 tpte = loadandclear(pte);
1753 pv->pv_pmap->pm_stats.wired_count--;
1756 vm_page_flag_set(m, PG_REFERENCED);
1759 * Update the vm_page_t clean and reference bits.
1762 #if defined(PMAP_DIAGNOSTIC)
1763 if (pmap_nw_modified((pt_entry_t) tpte)) {
1765 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1769 if (pmap_track_modified(pv->pv_va))
1772 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1773 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1774 ++pv->pv_pmap->pm_generation;
1775 m->md.pv_list_count--;
1776 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1780 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1782 pmap_inval_flush(&info);
1788 * Set the physical protection on the specified range of this map
1791 * This function may not be called from an interrupt if the map is
1792 * not the kernel_pmap.
1795 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1798 vm_offset_t pdnxt, ptpaddr;
1799 vm_pindex_t sindex, eindex;
1800 pmap_inval_info info;
1805 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1806 pmap_remove(pmap, sva, eva);
1810 if (prot & VM_PROT_WRITE)
1813 pmap_inval_init(&info);
1815 ptbase = get_ptbase(pmap);
1817 sindex = i386_btop(sva);
1818 eindex = i386_btop(eva);
1820 for (; sindex < eindex; sindex = pdnxt) {
1824 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1826 pdirindex = sindex / NPDEPG;
1827 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1828 pmap_inval_add(&info, pmap, -1);
1829 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1830 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1835 * Weed out invalid mappings. Note: we assume that the page
1836 * directory table is always allocated, and in kernel virtual.
1841 if (pdnxt > eindex) {
1845 for (; sindex != pdnxt; sindex++) {
1850 /* XXX this isn't optimal */
1851 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1852 pbits = ptbase[sindex];
1854 if (pbits & PG_MANAGED) {
1857 m = PHYS_TO_VM_PAGE(pbits);
1858 vm_page_flag_set(m, PG_REFERENCED);
1862 if (pmap_track_modified(i386_ptob(sindex))) {
1864 m = PHYS_TO_VM_PAGE(pbits);
1873 if (pbits != ptbase[sindex]) {
1874 ptbase[sindex] = pbits;
1878 pmap_inval_flush(&info);
1882 * Insert the given physical page (p) at
1883 * the specified virtual address (v) in the
1884 * target physical map with the protection requested.
1886 * If specified, the page will be wired down, meaning
1887 * that the related pte can not be reclaimed.
1889 * NB: This is the only routine which MAY NOT lazy-evaluate
1890 * or lose information. That is, this routine must actually
1891 * insert this page into the given map NOW.
1894 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1900 vm_offset_t origpte, newpte;
1902 pmap_inval_info info;
1908 #ifdef PMAP_DIAGNOSTIC
1910 panic("pmap_enter: toobig");
1911 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1912 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1914 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
1915 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
1917 db_print_backtrace();
1920 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
1921 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
1923 db_print_backtrace();
1929 * In the case that a page table page is not
1930 * resident, we are creating it here.
1932 if (va < UPT_MIN_ADDRESS) {
1933 mpte = pmap_allocpte(pmap, va);
1936 pmap_inval_init(&info);
1937 pte = pmap_pte(pmap, va);
1940 * Page Directory table entry not valid, we need a new PT page
1943 panic("pmap_enter: invalid page directory pdir=%x, va=0x%x\n",
1944 (unsigned) pmap->pm_pdir[PTDPTDI], va);
1947 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
1948 pmap_inval_add(&info, pmap, va); /* XXX non-optimal */
1949 origpte = *(vm_offset_t *)pte;
1950 opa = origpte & PG_FRAME;
1952 if (origpte & PG_PS)
1953 panic("pmap_enter: attempted pmap_enter on 4MB page");
1956 * Mapping has not changed, must be protection or wiring change.
1958 if (origpte && (opa == pa)) {
1960 * Wiring change, just update stats. We don't worry about
1961 * wiring PT pages as they remain resident as long as there
1962 * are valid mappings in them. Hence, if a user page is wired,
1963 * the PT page will be also.
1965 if (wired && ((origpte & PG_W) == 0))
1966 pmap->pm_stats.wired_count++;
1967 else if (!wired && (origpte & PG_W))
1968 pmap->pm_stats.wired_count--;
1970 #if defined(PMAP_DIAGNOSTIC)
1971 if (pmap_nw_modified((pt_entry_t) origpte)) {
1973 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1979 * Remove the extra pte reference. Note that we cannot
1980 * optimize the RO->RW case because we have adjusted the
1981 * wiring count above and may need to adjust the wiring
1988 * We might be turning off write access to the page,
1989 * so we go ahead and sense modify status.
1991 if (origpte & PG_MANAGED) {
1992 if ((origpte & PG_M) && pmap_track_modified(va)) {
1994 om = PHYS_TO_VM_PAGE(opa);
2002 * Mapping has changed, invalidate old range and fall through to
2003 * handle validating new mapping.
2007 err = pmap_remove_pte(pmap, pte, va, &info);
2009 panic("pmap_enter: pte vanished, va: 0x%x", va);
2013 * Enter on the PV list if part of our managed memory. Note that we
2014 * raise IPL while manipulating pv_table since pmap_enter can be
2015 * called at interrupt time.
2017 if (pmap_initialized &&
2018 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2019 pmap_insert_entry(pmap, va, mpte, m);
2024 * Increment counters
2026 pmap->pm_stats.resident_count++;
2028 pmap->pm_stats.wired_count++;
2032 * Now validate mapping with desired protection/wiring.
2034 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2038 if (va < UPT_MIN_ADDRESS)
2040 if (pmap == &kernel_pmap)
2044 * if the mapping or permission bits are different, we need
2045 * to update the pte.
2047 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2048 *pte = newpte | PG_A;
2050 pmap_inval_flush(&info);
2054 * this code makes some *MAJOR* assumptions:
2055 * 1. Current pmap & pmap exists.
2058 * 4. No page table pages.
2059 * 5. Tlbflush is deferred to calling procedure.
2060 * 6. Page IS managed.
2061 * but is *MUCH* faster than pmap_enter...
2065 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
2069 pmap_inval_info info;
2071 pmap_inval_init(&info);
2073 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2074 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2076 db_print_backtrace();
2079 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2080 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2082 db_print_backtrace();
2087 * In the case that a page table page is not
2088 * resident, we are creating it here.
2090 if (va < UPT_MIN_ADDRESS) {
2095 * Calculate pagetable page index
2097 ptepindex = va >> PDRSHIFT;
2098 if (mpte && (mpte->pindex == ptepindex)) {
2103 * Get the page directory entry
2105 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2108 * If the page table page is mapped, we just increment
2109 * the hold count, and activate it.
2113 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2114 if (pmap->pm_ptphint &&
2115 (pmap->pm_ptphint->pindex == ptepindex)) {
2116 mpte = pmap->pm_ptphint;
2118 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2119 pmap->pm_ptphint = mpte;
2125 mpte = _pmap_allocpte(pmap, ptepindex);
2133 * This call to vtopte makes the assumption that we are
2134 * entering the page into the current pmap. In order to support
2135 * quick entry into any pmap, one would likely use pmap_pte_quick.
2136 * But that isn't as quick as vtopte.
2138 pte = (unsigned *)vtopte(va);
2141 pmap_unwire_pte_hold(pmap, mpte, &info);
2146 * Enter on the PV list if part of our managed memory. Note that we
2147 * raise IPL while manipulating pv_table since pmap_enter can be
2148 * called at interrupt time.
2150 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
2151 pmap_insert_entry(pmap, va, mpte, m);
2154 * Increment counters
2156 pmap->pm_stats.resident_count++;
2158 pa = VM_PAGE_TO_PHYS(m);
2161 * Now validate mapping with RO protection
2163 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2164 *pte = pa | PG_V | PG_U;
2166 *pte = pa | PG_V | PG_U | PG_MANAGED;
2172 * Make a temporary mapping for a physical address. This is only intended
2173 * to be used for panic dumps.
2176 pmap_kenter_temporary(vm_paddr_t pa, int i)
2178 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2179 return ((void *)crashdumpmap);
2182 #define MAX_INIT_PT (96)
2185 * This routine preloads the ptes for a given object into the specified pmap.
2186 * This eliminates the blast of soft faults on process startup and
2187 * immediately after an mmap.
2189 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2192 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2193 vm_object_t object, vm_pindex_t pindex,
2194 vm_size_t size, int limit)
2196 struct rb_vm_page_scan_info info;
2200 * We can't preinit if read access isn't set or there is no pmap
2203 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2207 * We can't preinit if the pmap is not the current pmap
2209 if (curproc == NULL || pmap != vmspace_pmap(curproc->p_vmspace))
2212 psize = i386_btop(size);
2214 if ((object->type != OBJT_VNODE) ||
2215 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2216 (object->resident_page_count > MAX_INIT_PT))) {
2220 if (psize + pindex > object->size) {
2221 if (object->size < pindex)
2223 psize = object->size - pindex;
2230 * Use a red-black scan to traverse the requested range and load
2231 * any valid pages found into the pmap.
2233 * We cannot safely scan the object's memq unless we are in a
2234 * critical section since interrupts can remove pages from objects.
2236 info.start_pindex = pindex;
2237 info.end_pindex = pindex + psize - 1;
2244 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2245 pmap_object_init_pt_callback, &info);
2251 pmap_object_init_pt_callback(vm_page_t p, void *data)
2253 struct rb_vm_page_scan_info *info = data;
2254 vm_pindex_t rel_index;
2256 * don't allow an madvise to blow away our really
2257 * free pages allocating pv entries.
2259 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2260 vmstats.v_free_count < vmstats.v_free_reserved) {
2263 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2264 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2265 if ((p->queue - p->pc) == PQ_CACHE)
2266 vm_page_deactivate(p);
2268 rel_index = p->pindex - info->start_pindex;
2269 info->mpte = pmap_enter_quick(info->pmap,
2270 info->addr + i386_ptob(rel_index),
2272 vm_page_flag_set(p, PG_MAPPED);
2279 * pmap_prefault provides a quick way of clustering pagefaults into a
2280 * processes address space. It is a "cousin" of pmap_object_init_pt,
2281 * except it runs at page fault time instead of mmap time.
2285 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2287 static int pmap_prefault_pageorder[] = {
2288 -PAGE_SIZE, PAGE_SIZE,
2289 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2290 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2291 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2295 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2305 * We do not currently prefault mappings that use virtual page
2306 * tables. We do not prefault foreign pmaps.
2308 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2310 if (curproc == NULL || (pmap != vmspace_pmap(curproc->p_vmspace)))
2313 object = entry->object.vm_object;
2315 starta = addra - PFBAK * PAGE_SIZE;
2316 if (starta < entry->start)
2317 starta = entry->start;
2318 else if (starta > addra)
2322 * critical section protection is required to maintain the
2323 * page/object association, interrupts can free pages and remove
2324 * them from their objects.
2328 for (i = 0; i < PAGEORDER_SIZE; i++) {
2329 vm_object_t lobject;
2332 addr = addra + pmap_prefault_pageorder[i];
2333 if (addr > addra + (PFFOR * PAGE_SIZE))
2336 if (addr < starta || addr >= entry->end)
2339 if ((*pmap_pde(pmap, addr)) == NULL)
2342 pte = (unsigned *) vtopte(addr);
2346 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2349 for (m = vm_page_lookup(lobject, pindex);
2350 (!m && (lobject->type == OBJT_DEFAULT) &&
2351 (lobject->backing_object));
2352 lobject = lobject->backing_object
2354 if (lobject->backing_object_offset & PAGE_MASK)
2356 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2357 m = vm_page_lookup(lobject->backing_object, pindex);
2361 * give-up when a page is not in memory
2366 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2368 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2370 if ((m->queue - m->pc) == PQ_CACHE) {
2371 vm_page_deactivate(m);
2374 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2375 vm_page_flag_set(m, PG_MAPPED);
2383 * Routine: pmap_change_wiring
2384 * Function: Change the wiring attribute for a map/virtual-address
2386 * In/out conditions:
2387 * The mapping must already exist in the pmap.
2390 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2397 pte = pmap_pte(pmap, va);
2399 if (wired && !pmap_pte_w(pte))
2400 pmap->pm_stats.wired_count++;
2401 else if (!wired && pmap_pte_w(pte))
2402 pmap->pm_stats.wired_count--;
2405 * Wiring is not a hardware characteristic so there is no need to
2406 * invalidate TLB. However, in an SMP environment we must use
2407 * a locked bus cycle to update the pte (if we are not using
2408 * the pmap_inval_*() API that is)... it's ok to do this for simple
2413 atomic_set_int(pte, PG_W);
2415 atomic_clear_int(pte, PG_W);
2418 atomic_set_int_nonlocked(pte, PG_W);
2420 atomic_clear_int_nonlocked(pte, PG_W);
2427 * Copy the range specified by src_addr/len
2428 * from the source map to the range dst_addr/len
2429 * in the destination map.
2431 * This routine is only advisory and need not do anything.
2434 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2435 vm_size_t len, vm_offset_t src_addr)
2437 pmap_inval_info info;
2439 vm_offset_t end_addr = src_addr + len;
2441 unsigned src_frame, dst_frame;
2444 if (dst_addr != src_addr)
2447 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2448 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2452 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2453 if (dst_frame != (((unsigned) APTDpde) & PG_FRAME)) {
2454 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2455 /* The page directory is not shared between CPUs */
2458 pmap_inval_init(&info);
2459 pmap_inval_add(&info, dst_pmap, -1);
2460 pmap_inval_add(&info, src_pmap, -1);
2463 * critical section protection is required to maintain the page/object
2464 * association, interrupts can free pages and remove them from
2468 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2469 unsigned *src_pte, *dst_pte;
2470 vm_page_t dstmpte, srcmpte;
2471 vm_offset_t srcptepaddr;
2474 if (addr >= UPT_MIN_ADDRESS)
2475 panic("pmap_copy: invalid to pmap_copy page tables\n");
2478 * Don't let optional prefaulting of pages make us go
2479 * way below the low water mark of free pages or way
2480 * above high water mark of used pv entries.
2482 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2483 pv_entry_count > pv_entry_high_water)
2486 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2487 ptepindex = addr >> PDRSHIFT;
2489 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2490 if (srcptepaddr == 0)
2493 if (srcptepaddr & PG_PS) {
2494 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2495 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2496 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2501 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2502 if ((srcmpte == NULL) ||
2503 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2506 if (pdnxt > end_addr)
2509 src_pte = (unsigned *) vtopte(addr);
2510 dst_pte = (unsigned *) avtopte(addr);
2511 while (addr < pdnxt) {
2516 * we only virtual copy managed pages
2518 if ((ptetemp & PG_MANAGED) != 0) {
2520 * We have to check after allocpte for the
2521 * pte still being around... allocpte can
2524 dstmpte = pmap_allocpte(dst_pmap, addr);
2525 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2527 * Clear the modified and
2528 * accessed (referenced) bits
2531 m = PHYS_TO_VM_PAGE(ptetemp);
2532 *dst_pte = ptetemp & ~(PG_M | PG_A);
2533 dst_pmap->pm_stats.resident_count++;
2534 pmap_insert_entry(dst_pmap, addr,
2537 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2539 if (dstmpte->hold_count >= srcmpte->hold_count)
2548 pmap_inval_flush(&info);
2554 * Zero the specified PA by mapping the page into KVM and clearing its
2557 * This function may be called from an interrupt and no locking is
2561 pmap_zero_page(vm_paddr_t phys)
2563 struct mdglobaldata *gd = mdcpu;
2566 if (*(int *)gd->gd_CMAP3)
2567 panic("pmap_zero_page: CMAP3 busy");
2568 *(int *)gd->gd_CMAP3 =
2569 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2570 cpu_invlpg(gd->gd_CADDR3);
2572 #if defined(I686_CPU)
2573 if (cpu_class == CPUCLASS_686)
2574 i686_pagezero(gd->gd_CADDR3);
2577 bzero(gd->gd_CADDR3, PAGE_SIZE);
2578 *(int *) gd->gd_CMAP3 = 0;
2583 * pmap_page_assertzero:
2585 * Assert that a page is empty, panic if it isn't.
2588 pmap_page_assertzero(vm_paddr_t phys)
2590 struct mdglobaldata *gd = mdcpu;
2594 if (*(int *)gd->gd_CMAP3)
2595 panic("pmap_zero_page: CMAP3 busy");
2596 *(int *)gd->gd_CMAP3 =
2597 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2598 cpu_invlpg(gd->gd_CADDR3);
2599 for (i = 0; i < PAGE_SIZE; i += 4) {
2600 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2601 panic("pmap_page_assertzero() @ %p not zero!\n",
2602 (void *)gd->gd_CADDR3);
2605 *(int *) gd->gd_CMAP3 = 0;
2612 * Zero part of a physical page by mapping it into memory and clearing
2613 * its contents with bzero.
2615 * off and size may not cover an area beyond a single hardware page.
2618 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2620 struct mdglobaldata *gd = mdcpu;
2623 if (*(int *) gd->gd_CMAP3)
2624 panic("pmap_zero_page: CMAP3 busy");
2625 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2626 cpu_invlpg(gd->gd_CADDR3);
2628 #if defined(I686_CPU)
2629 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2630 i686_pagezero(gd->gd_CADDR3);
2633 bzero((char *)gd->gd_CADDR3 + off, size);
2634 *(int *) gd->gd_CMAP3 = 0;
2641 * Copy the physical page from the source PA to the target PA.
2642 * This function may be called from an interrupt. No locking
2646 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2648 struct mdglobaldata *gd = mdcpu;
2651 if (*(int *) gd->gd_CMAP1)
2652 panic("pmap_copy_page: CMAP1 busy");
2653 if (*(int *) gd->gd_CMAP2)
2654 panic("pmap_copy_page: CMAP2 busy");
2656 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2657 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2659 cpu_invlpg(gd->gd_CADDR1);
2660 cpu_invlpg(gd->gd_CADDR2);
2662 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2664 *(int *) gd->gd_CMAP1 = 0;
2665 *(int *) gd->gd_CMAP2 = 0;
2670 * pmap_copy_page_frag:
2672 * Copy the physical page from the source PA to the target PA.
2673 * This function may be called from an interrupt. No locking
2677 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2679 struct mdglobaldata *gd = mdcpu;
2682 if (*(int *) gd->gd_CMAP1)
2683 panic("pmap_copy_page: CMAP1 busy");
2684 if (*(int *) gd->gd_CMAP2)
2685 panic("pmap_copy_page: CMAP2 busy");
2687 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2688 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2690 cpu_invlpg(gd->gd_CADDR1);
2691 cpu_invlpg(gd->gd_CADDR2);
2693 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2694 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2697 *(int *) gd->gd_CMAP1 = 0;
2698 *(int *) gd->gd_CMAP2 = 0;
2703 * Returns true if the pmap's pv is one of the first
2704 * 16 pvs linked to from this page. This count may
2705 * be changed upwards or downwards in the future; it
2706 * is only necessary that true be returned for a small
2707 * subset of pmaps for proper page aging.
2710 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2715 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2720 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2721 if (pv->pv_pmap == pmap) {
2734 * Remove all pages from specified address space
2735 * this aids process exit speeds. Also, this code
2736 * is special cased for current process only, but
2737 * can have the more generic (and slightly slower)
2738 * mode enabled. This is much faster than pmap_remove
2739 * in the case of running down an entire address space.
2742 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2744 unsigned *pte, tpte;
2747 pmap_inval_info info;
2749 int32_t save_generation;
2751 if (curproc && pmap == vmspace_pmap(curproc->p_vmspace))
2756 pmap_inval_init(&info);
2758 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2759 if (pv->pv_va >= eva || pv->pv_va < sva) {
2760 npv = TAILQ_NEXT(pv, pv_plist);
2764 KKASSERT(pmap == pv->pv_pmap);
2767 pte = (unsigned *)vtopte(pv->pv_va);
2769 pte = pmap_pte_quick(pmap, pv->pv_va);
2770 if (pmap->pm_active)
2771 pmap_inval_add(&info, pmap, pv->pv_va);
2775 * We cannot remove wired pages from a process' mapping
2779 npv = TAILQ_NEXT(pv, pv_plist);
2784 m = PHYS_TO_VM_PAGE(tpte);
2786 KASSERT(m < &vm_page_array[vm_page_array_size],
2787 ("pmap_remove_pages: bad tpte %x", tpte));
2789 pmap->pm_stats.resident_count--;
2792 * Update the vm_page_t clean and reference bits.
2798 npv = TAILQ_NEXT(pv, pv_plist);
2799 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2800 save_generation = ++pmap->pm_generation;
2802 m->md.pv_list_count--;
2803 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2804 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2805 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2808 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2812 * Restart the scan if we blocked during the unuse or free
2813 * calls and other removals were made.
2815 if (save_generation != pmap->pm_generation) {
2816 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2817 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2820 pmap_inval_flush(&info);
2825 * pmap_testbit tests bits in pte's
2826 * note that the testbit/clearbit routines are inline,
2827 * and a lot of things compile-time evaluate.
2830 pmap_testbit(vm_page_t m, int bit)
2835 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2838 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2843 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2845 * if the bit being tested is the modified bit, then
2846 * mark clean_map and ptes as never
2849 if (bit & (PG_A|PG_M)) {
2850 if (!pmap_track_modified(pv->pv_va))
2854 #if defined(PMAP_DIAGNOSTIC)
2856 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2860 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2871 * this routine is used to modify bits in ptes
2873 static __inline void
2874 pmap_clearbit(vm_page_t m, int bit)
2876 struct pmap_inval_info info;
2881 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2884 pmap_inval_init(&info);
2888 * Loop over all current mappings setting/clearing as appropos If
2889 * setting RO do we need to clear the VAC?
2891 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2893 * don't write protect pager mappings
2896 if (!pmap_track_modified(pv->pv_va))
2900 #if defined(PMAP_DIAGNOSTIC)
2902 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2908 * Careful here. We can use a locked bus instruction to
2909 * clear PG_A or PG_M safely but we need to synchronize
2910 * with the target cpus when we mess with PG_RW.
2912 * We do not have to force synchronization when clearing
2913 * PG_M even for PTEs generated via virtual memory maps,
2914 * because the virtual kernel will invalidate the pmap
2915 * entry when/if it needs to resynchronize the Modify bit.
2917 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2919 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2926 atomic_clear_int(pte, PG_M|PG_RW);
2927 } else if (bit == PG_M) {
2929 * We could also clear PG_RW here to force
2930 * a fault on write to redetect PG_M for
2931 * virtual kernels, but it isn't necessary
2932 * since virtual kernels invalidate the pte
2933 * when they clear the VPTE_M bit in their
2934 * virtual page tables.
2936 atomic_clear_int(pte, PG_M);
2938 atomic_clear_int(pte, bit);
2942 pmap_inval_flush(&info);
2947 * pmap_page_protect:
2949 * Lower the permission for all mappings to a given page.
2952 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2954 if ((prot & VM_PROT_WRITE) == 0) {
2955 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2956 pmap_clearbit(m, PG_RW);
2964 pmap_phys_address(int ppn)
2966 return (i386_ptob(ppn));
2970 * pmap_ts_referenced:
2972 * Return a count of reference bits for a page, clearing those bits.
2973 * It is not necessary for every reference bit to be cleared, but it
2974 * is necessary that 0 only be returned when there are truly no
2975 * reference bits set.
2977 * XXX: The exact number of bits to check and clear is a matter that
2978 * should be tested and standardized at some point in the future for
2979 * optimal aging of shared pages.
2982 pmap_ts_referenced(vm_page_t m)
2984 pv_entry_t pv, pvf, pvn;
2988 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2993 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2998 pvn = TAILQ_NEXT(pv, pv_list);
3000 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3002 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3004 if (!pmap_track_modified(pv->pv_va))
3007 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3009 if (pte && (*pte & PG_A)) {
3011 atomic_clear_int(pte, PG_A);
3013 atomic_clear_int_nonlocked(pte, PG_A);
3020 } while ((pv = pvn) != NULL && pv != pvf);
3030 * Return whether or not the specified physical page was modified
3031 * in any physical maps.
3034 pmap_is_modified(vm_page_t m)
3036 return pmap_testbit(m, PG_M);
3040 * Clear the modify bits on the specified physical page.
3043 pmap_clear_modify(vm_page_t m)
3045 pmap_clearbit(m, PG_M);
3049 * pmap_clear_reference:
3051 * Clear the reference bit on the specified physical page.
3054 pmap_clear_reference(vm_page_t m)
3056 pmap_clearbit(m, PG_A);
3060 * Miscellaneous support routines follow
3064 i386_protection_init(void)
3068 kp = protection_codes;
3069 for (prot = 0; prot < 8; prot++) {
3071 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3073 * Read access is also 0. There isn't any execute bit,
3074 * so just make it readable.
3076 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3077 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3078 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3081 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3082 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3083 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3084 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3092 * Map a set of physical memory pages into the kernel virtual
3093 * address space. Return a pointer to where it is mapped. This
3094 * routine is intended to be used for mapping device memory,
3097 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3101 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3103 vm_offset_t va, tmpva, offset;
3106 offset = pa & PAGE_MASK;
3107 size = roundup(offset + size, PAGE_SIZE);
3109 va = kmem_alloc_nofault(&kernel_map, size);
3111 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3114 for (tmpva = va; size > 0;) {
3115 pte = (unsigned *)vtopte(tmpva);
3116 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3124 return ((void *)(va + offset));
3128 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3130 vm_offset_t base, offset;
3132 base = va & PG_FRAME;
3133 offset = va & PAGE_MASK;
3134 size = roundup(offset + size, PAGE_SIZE);
3135 pmap_qremove(va, size >> PAGE_SHIFT);
3136 kmem_free(&kernel_map, base, size);
3140 * perform the pmap work for mincore
3143 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3145 unsigned *ptep, pte;
3149 ptep = pmap_pte(pmap, addr);
3154 if ((pte = *ptep) != 0) {
3157 val = MINCORE_INCORE;
3158 if ((pte & PG_MANAGED) == 0)
3161 pa = pte & PG_FRAME;
3163 m = PHYS_TO_VM_PAGE(pa);
3169 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3171 * Modified by someone
3173 else if (m->dirty || pmap_is_modified(m))
3174 val |= MINCORE_MODIFIED_OTHER;
3179 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3182 * Referenced by someone
3184 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3185 val |= MINCORE_REFERENCED_OTHER;
3186 vm_page_flag_set(m, PG_REFERENCED);
3193 pmap_activate(struct proc *p)
3197 KKASSERT((p == curproc));
3199 pmap = vmspace_pmap(p->p_vmspace);
3201 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3203 pmap->pm_active |= 1;
3205 #if defined(SWTCH_OPTIM_STATS)
3208 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3209 load_cr3(curthread->td_pcb->pcb_cr3);
3213 pmap_deactivate(struct proc *p)
3217 pmap = vmspace_pmap(p->p_vmspace);
3219 atomic_clear_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3221 pmap->pm_active &= ~1;
3224 * XXX - note we do not adjust %cr3. The caller is expected to
3225 * activate a new pmap or do a thread-exit.
3230 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3233 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3237 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3244 static void pads (pmap_t pm);
3245 void pmap_pvdump (vm_paddr_t pa);
3247 /* print address space of pmap*/
3254 if (pm == &kernel_pmap)
3256 for (i = 0; i < 1024; i++)
3258 for (j = 0; j < 1024; j++) {
3259 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3260 if (pm == &kernel_pmap && va < KERNBASE)
3262 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3264 ptep = pmap_pte_quick(pm, va);
3265 if (pmap_pte_v(ptep))
3266 kprintf("%x:%x ", va, *(int *) ptep);
3272 pmap_pvdump(vm_paddr_t pa)
3277 kprintf("pa %08llx", (long long)pa);
3278 m = PHYS_TO_VM_PAGE(pa);
3279 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3281 kprintf(" -> pmap %p, va %x, flags %x",
3282 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3284 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);