2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
9 * This code is derived from software contributed to Berkeley by
10 * the Systems Programming Group of the University of Utah Computer
11 * Science Department and William Jolitz of UUNET Technologies Inc.
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. All advertising materials mentioning features or use of this software
22 * must display the following acknowledgement:
23 * This product includes software developed by the University of
24 * California, Berkeley and its contributors.
25 * 4. Neither the name of the University nor the names of its contributors
26 * may be used to endorse or promote products derived from this software
27 * without specific prior written permission.
29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
42 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
43 * $DragonFly: src/sys/platform/pc32/i386/pmap.c,v 1.87 2008/08/25 17:01:38 dillon Exp $
47 * Manages physical address maps.
49 * In addition to hardware address maps, this
50 * module is called upon to provide software-use-only
51 * maps which may or may not be stored in the same
52 * form as hardware maps. These pseudo-maps are
53 * used to store intermediate results from copy
54 * operations to and from address spaces.
56 * Since the information managed by this module is
57 * also stored by the logical address mapping module,
58 * this module may throw away valid virtual-to-physical
59 * mappings at almost any time. However, invalidations
60 * of virtual-to-physical mappings must be done as
63 * In order to cope with hardware architectures which
64 * make virtual-to-physical map invalidates expensive,
65 * this module may delay invalidate or reduced protection
66 * operations until such time as they are actually
67 * necessary. This module is given full information as
68 * to which processors are currently using which maps,
69 * and to when physical maps must be made correct.
72 #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>
110 #define PMAP_KEEP_PDIRS
111 #ifndef PMAP_SHPGPERPROC
112 #define PMAP_SHPGPERPROC 200
115 #if defined(DIAGNOSTIC)
116 #define PMAP_DIAGNOSTIC
121 #if !defined(PMAP_DIAGNOSTIC)
122 #define PMAP_INLINE __inline
128 * Get PDEs and PTEs for user/kernel address space
130 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
131 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
133 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
134 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
135 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
136 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
137 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
141 * Given a map and a machine independent protection code,
142 * convert to a vax protection code.
144 #define pte_prot(m, p) \
145 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
146 static int protection_codes[8];
148 struct pmap kernel_pmap;
149 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
151 vm_paddr_t avail_start; /* PA of first available physical page */
152 vm_paddr_t avail_end; /* PA of last available physical page */
153 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
154 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
155 vm_offset_t KvaStart; /* VA start of KVA space */
156 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
157 vm_offset_t KvaSize; /* max size of kernel virtual address space */
158 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
159 static int pgeflag; /* PG_G or-in */
160 static int pseflag; /* PG_PS or-in */
162 static vm_object_t kptobj;
165 vm_offset_t kernel_vm_end;
168 * Data for the pv entry allocation mechanism
170 static vm_zone_t pvzone;
171 static struct vm_zone pvzone_store;
172 static struct vm_object pvzone_obj;
173 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
174 static int pmap_pagedaemon_waken = 0;
175 static struct pv_entry *pvinit;
178 * All those kernel PT submaps that BSD is so fond of
180 pt_entry_t *CMAP1 = 0, *ptmmap;
181 caddr_t CADDR1 = 0, ptvmmap = 0;
182 static pt_entry_t *msgbufmap;
183 struct msgbuf *msgbufp=0;
188 static pt_entry_t *pt_crashdumpmap;
189 static caddr_t crashdumpmap;
191 extern pt_entry_t *SMPpt;
193 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
194 static unsigned * get_ptbase (pmap_t pmap);
195 static pv_entry_t get_pv_entry (void);
196 static void i386_protection_init (void);
197 static __inline void pmap_clearbit (vm_page_t m, int bit);
199 static void pmap_remove_all (vm_page_t m);
200 static void pmap_enter_quick (pmap_t pmap, vm_offset_t va, vm_page_t m);
201 static int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
202 vm_offset_t sva, pmap_inval_info_t info);
203 static void pmap_remove_page (struct pmap *pmap,
204 vm_offset_t va, pmap_inval_info_t info);
205 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
206 vm_offset_t va, pmap_inval_info_t info);
207 static boolean_t pmap_testbit (vm_page_t m, int bit);
208 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
209 vm_page_t mpte, vm_page_t m);
211 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
213 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
214 static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
215 static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
216 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
217 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
218 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
220 static unsigned pdir4mb;
223 * Move the kernel virtual free pointer to the next
224 * 4MB. This is used to help improve performance
225 * by using a large (4MB) page for much of the kernel
226 * (.text, .data, .bss)
229 pmap_kmem_choose(vm_offset_t addr)
231 vm_offset_t newaddr = addr;
233 if (cpu_feature & CPUID_PSE) {
234 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
243 * Extract the page table entry associated with the given map/virtual
246 * This function may NOT be called from an interrupt.
248 PMAP_INLINE unsigned *
249 pmap_pte(pmap_t pmap, vm_offset_t va)
254 pdeaddr = (unsigned *) pmap_pde(pmap, va);
255 if (*pdeaddr & PG_PS)
258 return get_ptbase(pmap) + i386_btop(va);
267 * Super fast pmap_pte routine best used when scanning the pv lists.
268 * This eliminates many course-grained invltlb calls. Note that many of
269 * the pv list scans are across different pmaps and it is very wasteful
270 * to do an entire invltlb when checking a single mapping.
272 * Should only be called while in a critical section.
275 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
277 struct mdglobaldata *gd = mdcpu;
280 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
281 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
282 unsigned index = i386_btop(va);
283 /* are we current address space or kernel? */
284 if ((pmap == &kernel_pmap) ||
285 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
286 return (unsigned *) PTmap + index;
288 newpf = pde & PG_FRAME;
289 if ( ((* (unsigned *) gd->gd_PMAP1) & PG_FRAME) != newpf) {
290 * (unsigned *) gd->gd_PMAP1 = newpf | PG_RW | PG_V;
291 cpu_invlpg(gd->gd_PADDR1);
293 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
300 * Bootstrap the system enough to run with virtual memory.
302 * On the i386 this is called after mapping has already been enabled
303 * and just syncs the pmap module with what has already been done.
304 * [We can't call it easily with mapping off since the kernel is not
305 * mapped with PA == VA, hence we would have to relocate every address
306 * from the linked base (virtual) address "KERNBASE" to the actual
307 * (physical) address starting relative to 0]
310 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
314 struct mdglobaldata *gd;
318 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
319 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
320 KvaEnd = KvaStart + KvaSize;
322 avail_start = firstaddr;
325 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
326 * too large. It should instead be correctly calculated in locore.s and
327 * not based on 'first' (which is a physical address, not a virtual
328 * address, for the start of unused physical memory). The kernel
329 * page tables are NOT double mapped and thus should not be included
330 * in this calculation.
332 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
333 virtual_start = pmap_kmem_choose(virtual_start);
334 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
337 * Initialize protection array.
339 i386_protection_init();
342 * The kernel's pmap is statically allocated so we don't have to use
343 * pmap_create, which is unlikely to work correctly at this part of
344 * the boot sequence (XXX and which no longer exists).
346 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
347 kernel_pmap.pm_count = 1;
348 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
349 TAILQ_INIT(&kernel_pmap.pm_pvlist);
353 * Reserve some special page table entries/VA space for temporary
356 #define SYSMAP(c, p, v, n) \
357 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
360 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
363 * CMAP1/CMAP2 are used for zeroing and copying pages.
365 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
370 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
373 * ptvmmap is used for reading arbitrary physical pages via
376 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
379 * msgbufp is used to map the system message buffer.
380 * XXX msgbufmap is not used.
382 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
383 atop(round_page(MSGBUF_SIZE)))
388 for (i = 0; i < NKPT; i++)
392 * PG_G is terribly broken on SMP because we IPI invltlb's in some
393 * cases rather then invl1pg. Actually, I don't even know why it
394 * works under UP because self-referential page table mappings
399 if (cpu_feature & CPUID_PGE)
404 * Initialize the 4MB page size flag
408 * The 4MB page version of the initial
409 * kernel page mapping.
413 #if !defined(DISABLE_PSE)
414 if (cpu_feature & CPUID_PSE) {
417 * Note that we have enabled PSE mode
420 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
421 ptditmp &= ~(NBPDR - 1);
422 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
427 * Enable the PSE mode. If we are SMP we can't do this
428 * now because the APs will not be able to use it when
431 load_cr4(rcr4() | CR4_PSE);
434 * We can do the mapping here for the single processor
435 * case. We simply ignore the old page table page from
439 * For SMP, we still need 4K pages to bootstrap APs,
440 * PSE will be enabled as soon as all APs are up.
442 PTD[KPTDI] = (pd_entry_t)ptditmp;
443 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
450 * We need to finish setting up the globaldata page for the BSP.
451 * locore has already populated the page table for the mdglobaldata
454 pg = MDGLOBALDATA_BASEALLOC_PAGES;
455 gd = &CPU_prvspace[0].mdglobaldata;
456 gd->gd_CMAP1 = &SMPpt[pg + 0];
457 gd->gd_CMAP2 = &SMPpt[pg + 1];
458 gd->gd_CMAP3 = &SMPpt[pg + 2];
459 gd->gd_PMAP1 = &SMPpt[pg + 3];
460 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
461 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
462 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
463 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
470 * Set 4mb pdir for mp startup
475 if (pseflag && (cpu_feature & CPUID_PSE)) {
476 load_cr4(rcr4() | CR4_PSE);
477 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
478 kernel_pmap.pm_pdir[KPTDI] =
479 PTD[KPTDI] = (pd_entry_t)pdir4mb;
487 * Initialize the pmap module.
488 * Called by vm_init, to initialize any structures that the pmap
489 * system needs to map virtual memory.
490 * pmap_init has been enhanced to support in a fairly consistant
491 * way, discontiguous physical memory.
500 * object for kernel page table pages
502 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
505 * Allocate memory for random pmap data structures. Includes the
509 for(i = 0; i < vm_page_array_size; i++) {
512 m = &vm_page_array[i];
513 TAILQ_INIT(&m->md.pv_list);
514 m->md.pv_list_count = 0;
518 * init the pv free list
520 initial_pvs = vm_page_array_size;
521 if (initial_pvs < MINPV)
523 pvzone = &pvzone_store;
524 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
525 initial_pvs * sizeof (struct pv_entry));
526 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
530 * Now it is safe to enable pv_table recording.
532 pmap_initialized = TRUE;
536 * Initialize the address space (zone) for the pv_entries. Set a
537 * high water mark so that the system can recover from excessive
538 * numbers of pv entries.
543 int shpgperproc = PMAP_SHPGPERPROC;
545 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
546 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
547 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
548 pv_entry_high_water = 9 * (pv_entry_max / 10);
549 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
553 /***************************************************
554 * Low level helper routines.....
555 ***************************************************/
557 #if defined(PMAP_DIAGNOSTIC)
560 * This code checks for non-writeable/modified pages.
561 * This should be an invalid condition.
564 pmap_nw_modified(pt_entry_t ptea)
570 if ((pte & (PG_M|PG_RW)) == PG_M)
579 * this routine defines the region(s) of memory that should
580 * not be tested for the modified bit.
582 static PMAP_INLINE int
583 pmap_track_modified(vm_offset_t va)
585 if ((va < clean_sva) || (va >= clean_eva))
592 get_ptbase(pmap_t pmap)
594 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
595 struct globaldata *gd = mycpu;
597 /* are we current address space or kernel? */
598 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
599 return (unsigned *) PTmap;
602 /* otherwise, we are alternate address space */
603 KKASSERT(gd->gd_intr_nesting_level == 0 &&
604 (gd->gd_curthread->td_flags & TDF_INTTHREAD) == 0);
606 if (frame != (((unsigned) APTDpde) & PG_FRAME)) {
607 APTDpde = (pd_entry_t)(frame | PG_RW | PG_V);
608 /* The page directory is not shared between CPUs */
611 return (unsigned *) APTmap;
617 * Extract the physical page address associated with the map/VA pair.
619 * This function may not be called from an interrupt if the pmap is
623 pmap_extract(pmap_t pmap, vm_offset_t va)
626 vm_offset_t pdirindex;
628 pdirindex = va >> PDRSHIFT;
629 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
631 if ((rtval & PG_PS) != 0) {
632 rtval &= ~(NBPDR - 1);
633 rtval |= va & (NBPDR - 1);
636 pte = get_ptbase(pmap) + i386_btop(va);
637 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
643 /***************************************************
644 * Low level mapping routines.....
645 ***************************************************/
648 * Routine: pmap_kenter
650 * Add a wired page to the KVA
651 * NOTE! note that in order for the mapping to take effect -- you
652 * should do an invltlb after doing the pmap_kenter().
655 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
659 pmap_inval_info info;
661 pmap_inval_init(&info);
662 npte = pa | PG_RW | PG_V | pgeflag;
663 pte = (unsigned *)vtopte(va);
664 pmap_inval_add(&info, &kernel_pmap, va);
666 pmap_inval_flush(&info);
670 * Routine: pmap_kenter_quick
672 * Similar to pmap_kenter(), except we only invalidate the
673 * mapping on the current CPU.
676 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
681 npte = pa | PG_RW | PG_V | pgeflag;
682 pte = (unsigned *)vtopte(va);
684 cpu_invlpg((void *)va);
688 pmap_kenter_sync(vm_offset_t va)
690 pmap_inval_info info;
692 pmap_inval_init(&info);
693 pmap_inval_add(&info, &kernel_pmap, va);
694 pmap_inval_flush(&info);
698 pmap_kenter_sync_quick(vm_offset_t va)
700 cpu_invlpg((void *)va);
704 * remove a page from the kernel pagetables
707 pmap_kremove(vm_offset_t va)
710 pmap_inval_info info;
712 pmap_inval_init(&info);
713 pte = (unsigned *)vtopte(va);
714 pmap_inval_add(&info, &kernel_pmap, va);
716 pmap_inval_flush(&info);
720 pmap_kremove_quick(vm_offset_t va)
723 pte = (unsigned *)vtopte(va);
725 cpu_invlpg((void *)va);
729 * XXX these need to be recoded. They are not used in any critical path.
732 pmap_kmodify_rw(vm_offset_t va)
734 *vtopte(va) |= PG_RW;
735 cpu_invlpg((void *)va);
739 pmap_kmodify_nc(vm_offset_t va)
742 cpu_invlpg((void *)va);
746 * Used to map a range of physical addresses into kernel
747 * virtual address space.
749 * For now, VM is already on, we only need to map the
753 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
755 while (start < end) {
756 pmap_kenter(virt, start);
765 * Add a list of wired pages to the kva
766 * this routine is only used for temporary
767 * kernel mappings that do not need to have
768 * page modification or references recorded.
769 * Note that old mappings are simply written
770 * over. The page *must* be wired.
773 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
777 end_va = va + count * PAGE_SIZE;
779 while (va < end_va) {
782 pte = (unsigned *)vtopte(va);
783 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
784 cpu_invlpg((void *)va);
789 smp_invltlb(); /* XXX */
794 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
797 cpumask_t cmask = mycpu->gd_cpumask;
799 end_va = va + count * PAGE_SIZE;
801 while (va < end_va) {
806 * Install the new PTE. If the pte changed from the prior
807 * mapping we must reset the cpu mask and invalidate the page.
808 * If the pte is the same but we have not seen it on the
809 * current cpu, invlpg the existing mapping. Otherwise the
810 * entry is optimal and no invalidation is required.
812 pte = (unsigned *)vtopte(va);
813 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
814 if (*pte != pteval) {
817 cpu_invlpg((void *)va);
818 } else if ((*mask & cmask) == 0) {
819 cpu_invlpg((void *)va);
828 * this routine jerks page mappings from the
829 * kernel -- it is meant only for temporary mappings.
832 pmap_qremove(vm_offset_t va, int count)
836 end_va = va + count*PAGE_SIZE;
838 while (va < end_va) {
841 pte = (unsigned *)vtopte(va);
843 cpu_invlpg((void *)va);
852 * This routine works like vm_page_lookup() but also blocks as long as the
853 * page is busy. This routine does not busy the page it returns.
855 * Unless the caller is managing objects whos pages are in a known state,
856 * the call should be made with a critical section held so the page's object
857 * association remains valid on return.
860 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
865 m = vm_page_lookup(object, pindex);
866 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
872 * Create a new thread and optionally associate it with a (new) process.
873 * NOTE! the new thread's cpu may not equal the current cpu.
876 pmap_init_thread(thread_t td)
878 /* enforce pcb placement */
879 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
880 td->td_savefpu = &td->td_pcb->pcb_save;
881 td->td_sp = (char *)td->td_pcb - 16;
885 * This routine directly affects the fork perf for a process.
888 pmap_init_proc(struct proc *p)
893 * Dispose the UPAGES for a process that has exited.
894 * This routine directly impacts the exit perf of a process.
897 pmap_dispose_proc(struct proc *p)
899 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
902 /***************************************************
903 * Page table page management routines.....
904 ***************************************************/
907 * This routine unholds page table pages, and if the hold count
908 * drops to zero, then it decrements the wire count.
911 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
914 * Wait until we can busy the page ourselves. We cannot have
915 * any active flushes if we block.
917 if (m->flags & PG_BUSY) {
918 pmap_inval_flush(info);
919 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
922 KASSERT(m->queue == PQ_NONE,
923 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
925 if (m->hold_count == 1) {
927 * Unmap the page table page
930 pmap_inval_add(info, pmap, -1);
931 pmap->pm_pdir[m->pindex] = 0;
933 KKASSERT(pmap->pm_stats.resident_count > 0);
934 --pmap->pm_stats.resident_count;
936 if (pmap->pm_ptphint == m)
937 pmap->pm_ptphint = NULL;
940 * This was our last hold, the page had better be unwired
941 * after we decrement wire_count.
943 * FUTURE NOTE: shared page directory page could result in
944 * multiple wire counts.
948 KKASSERT(m->wire_count == 0);
949 --vmstats.v_wire_count;
950 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
952 vm_page_free_zero(m);
955 KKASSERT(m->hold_count > 1);
961 static PMAP_INLINE int
962 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
964 KKASSERT(m->hold_count > 0);
965 if (m->hold_count > 1) {
969 return _pmap_unwire_pte_hold(pmap, m, info);
974 * After removing a page table entry, this routine is used to
975 * conditionally free the page, and manage the hold/wire counts.
978 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
979 pmap_inval_info_t info)
982 if (va >= UPT_MIN_ADDRESS)
986 ptepindex = (va >> PDRSHIFT);
987 if (pmap->pm_ptphint &&
988 (pmap->pm_ptphint->pindex == ptepindex)) {
989 mpte = pmap->pm_ptphint;
991 pmap_inval_flush(info);
992 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
993 pmap->pm_ptphint = mpte;
997 return pmap_unwire_pte_hold(pmap, mpte, info);
1001 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1002 * it, and IdlePTD, represents the template used to update all other pmaps.
1004 * On architectures where the kernel pmap is not integrated into the user
1005 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1006 * kernel_pmap should be used to directly access the kernel_pmap.
1009 pmap_pinit0(struct pmap *pmap)
1012 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1013 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1015 pmap->pm_active = 0;
1016 pmap->pm_ptphint = NULL;
1017 TAILQ_INIT(&pmap->pm_pvlist);
1018 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1022 * Initialize a preallocated and zeroed pmap structure,
1023 * such as one in a vmspace structure.
1026 pmap_pinit(struct pmap *pmap)
1031 * No need to allocate page table space yet but we do need a valid
1032 * page directory table.
1034 if (pmap->pm_pdir == NULL) {
1036 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1040 * Allocate an object for the ptes
1042 if (pmap->pm_pteobj == NULL)
1043 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1046 * Allocate the page directory page, unless we already have
1047 * one cached. If we used the cached page the wire_count will
1048 * already be set appropriately.
1050 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1051 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1052 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1053 pmap->pm_pdirm = ptdpg;
1054 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1055 ptdpg->valid = VM_PAGE_BITS_ALL;
1056 ptdpg->wire_count = 1;
1057 ++vmstats.v_wire_count;
1058 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1060 if ((ptdpg->flags & PG_ZERO) == 0)
1061 bzero(pmap->pm_pdir, PAGE_SIZE);
1063 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1065 /* install self-referential address mapping entry */
1066 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1067 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1070 pmap->pm_active = 0;
1071 pmap->pm_ptphint = NULL;
1072 TAILQ_INIT(&pmap->pm_pvlist);
1073 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1074 pmap->pm_stats.resident_count = 1;
1078 * Clean up a pmap structure so it can be physically freed. This routine
1079 * is called by the vmspace dtor function. A great deal of pmap data is
1080 * left passively mapped to improve vmspace management so we have a bit
1081 * of cleanup work to do here.
1084 pmap_puninit(pmap_t pmap)
1088 KKASSERT(pmap->pm_active == 0);
1089 if ((p = pmap->pm_pdirm) != NULL) {
1090 KKASSERT(pmap->pm_pdir != NULL);
1091 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1093 vmstats.v_wire_count--;
1094 KKASSERT((p->flags & PG_BUSY) == 0);
1096 vm_page_free_zero(p);
1097 pmap->pm_pdirm = NULL;
1099 if (pmap->pm_pdir) {
1100 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1101 pmap->pm_pdir = NULL;
1103 if (pmap->pm_pteobj) {
1104 vm_object_deallocate(pmap->pm_pteobj);
1105 pmap->pm_pteobj = NULL;
1110 * Wire in kernel global address entries. To avoid a race condition
1111 * between pmap initialization and pmap_growkernel, this procedure
1112 * adds the pmap to the master list (which growkernel scans to update),
1113 * then copies the template.
1116 pmap_pinit2(struct pmap *pmap)
1119 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1120 /* XXX copies current process, does not fill in MPPTDI */
1121 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1126 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1127 * 0 on failure (if the procedure had to sleep).
1129 * When asked to remove the page directory page itself, we actually just
1130 * leave it cached so we do not have to incur the SMP inval overhead of
1131 * removing the kernel mapping. pmap_puninit() will take care of it.
1134 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1136 unsigned *pde = (unsigned *) pmap->pm_pdir;
1138 * This code optimizes the case of freeing non-busy
1139 * page-table pages. Those pages are zero now, and
1140 * might as well be placed directly into the zero queue.
1142 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1148 * Remove the page table page from the processes address space.
1151 KKASSERT(pmap->pm_stats.resident_count > 0);
1152 --pmap->pm_stats.resident_count;
1154 if (p->hold_count) {
1155 panic("pmap_release: freeing held page table page");
1157 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1158 pmap->pm_ptphint = NULL;
1161 * We leave the page directory page cached, wired, and mapped in
1162 * the pmap until the dtor function (pmap_puninit()) gets called.
1163 * However, still clean it up so we can set PG_ZERO.
1165 if (p->pindex == PTDPTDI) {
1166 bzero(pde + KPTDI, nkpt * PTESIZE);
1169 vm_page_flag_set(p, PG_ZERO);
1173 vmstats.v_wire_count--;
1174 vm_page_free_zero(p);
1180 * this routine is called if the page table page is not
1184 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1186 vm_offset_t pteva, ptepa;
1190 * Find or fabricate a new pagetable page
1192 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1193 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1195 KASSERT(m->queue == PQ_NONE,
1196 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1199 * Increment the hold count for the page we will be returning to
1205 * It is possible that someone else got in and mapped by the page
1206 * directory page while we were blocked, if so just unbusy and
1207 * return the held page.
1209 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1210 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1215 if (m->wire_count == 0)
1216 vmstats.v_wire_count++;
1221 * Map the pagetable page into the process address space, if
1222 * it isn't already there.
1225 ++pmap->pm_stats.resident_count;
1227 ptepa = VM_PAGE_TO_PHYS(m);
1228 pmap->pm_pdir[ptepindex] =
1229 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1232 * Set the page table hint
1234 pmap->pm_ptphint = m;
1237 * Try to use the new mapping, but if we cannot, then
1238 * do it with the routine that maps the page explicitly.
1240 if ((m->flags & PG_ZERO) == 0) {
1241 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1242 (((unsigned) PTDpde) & PG_FRAME)) {
1243 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1244 bzero((caddr_t) pteva, PAGE_SIZE);
1246 pmap_zero_page(ptepa);
1250 m->valid = VM_PAGE_BITS_ALL;
1251 vm_page_flag_clear(m, PG_ZERO);
1252 vm_page_flag_set(m, PG_MAPPED);
1259 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1266 * Calculate pagetable page index
1268 ptepindex = va >> PDRSHIFT;
1271 * Get the page directory entry
1273 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1276 * This supports switching from a 4MB page to a
1279 if (ptepa & PG_PS) {
1280 pmap->pm_pdir[ptepindex] = 0;
1287 * If the page table page is mapped, we just increment the
1288 * hold count, and activate it.
1292 * In order to get the page table page, try the
1295 if (pmap->pm_ptphint &&
1296 (pmap->pm_ptphint->pindex == ptepindex)) {
1297 m = pmap->pm_ptphint;
1299 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1300 pmap->pm_ptphint = m;
1306 * Here if the pte page isn't mapped, or if it has been deallocated.
1308 return _pmap_allocpte(pmap, ptepindex);
1312 /***************************************************
1313 * Pmap allocation/deallocation routines.
1314 ***************************************************/
1317 * Release any resources held by the given physical map.
1318 * Called when a pmap initialized by pmap_pinit is being released.
1319 * Should only be called if the map contains no valid mappings.
1321 static int pmap_release_callback(struct vm_page *p, void *data);
1324 pmap_release(struct pmap *pmap)
1326 vm_object_t object = pmap->pm_pteobj;
1327 struct rb_vm_page_scan_info info;
1329 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1330 #if defined(DIAGNOSTIC)
1331 if (object->ref_count != 1)
1332 panic("pmap_release: pteobj reference count != 1");
1336 info.object = object;
1338 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1345 info.limit = object->generation;
1347 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1348 pmap_release_callback, &info);
1349 if (info.error == 0 && info.mpte) {
1350 if (!pmap_release_free_page(pmap, info.mpte))
1354 } while (info.error);
1358 pmap_release_callback(struct vm_page *p, void *data)
1360 struct rb_vm_page_scan_info *info = data;
1362 if (p->pindex == PTDPTDI) {
1366 if (!pmap_release_free_page(info->pmap, p)) {
1370 if (info->object->generation != info->limit) {
1378 * Grow the number of kernel page table entries, if needed.
1382 pmap_growkernel(vm_offset_t addr)
1385 vm_offset_t ptppaddr;
1390 if (kernel_vm_end == 0) {
1391 kernel_vm_end = KERNBASE;
1393 while (pdir_pde(PTD, kernel_vm_end)) {
1394 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1398 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1399 while (kernel_vm_end < addr) {
1400 if (pdir_pde(PTD, kernel_vm_end)) {
1401 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1406 * This index is bogus, but out of the way
1408 nkpg = vm_page_alloc(kptobj, nkpt,
1409 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1411 panic("pmap_growkernel: no memory to grow kernel");
1414 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1415 pmap_zero_page(ptppaddr);
1416 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1417 pdir_pde(PTD, kernel_vm_end) = newpdir;
1418 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1422 * This update must be interlocked with pmap_pinit2.
1424 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1425 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1427 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1428 ~(PAGE_SIZE * NPTEPG - 1);
1434 * Retire the given physical map from service.
1435 * Should only be called if the map contains
1436 * no valid mappings.
1439 pmap_destroy(pmap_t pmap)
1446 count = --pmap->pm_count;
1449 panic("destroying a pmap is not yet implemented");
1454 * Add a reference to the specified pmap.
1457 pmap_reference(pmap_t pmap)
1464 /***************************************************
1465 * page management routines.
1466 ***************************************************/
1469 * free the pv_entry back to the free list. This function may be
1470 * called from an interrupt.
1472 static PMAP_INLINE void
1473 free_pv_entry(pv_entry_t pv)
1480 * get a new pv_entry, allocating a block from the system
1481 * when needed. This function may be called from an interrupt.
1487 if (pv_entry_high_water &&
1488 (pv_entry_count > pv_entry_high_water) &&
1489 (pmap_pagedaemon_waken == 0)) {
1490 pmap_pagedaemon_waken = 1;
1491 wakeup (&vm_pages_needed);
1493 return zalloc(pvzone);
1497 * This routine is very drastic, but can save the system
1505 static int warningdone=0;
1507 if (pmap_pagedaemon_waken == 0)
1509 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))
1527 * If it is the first entry on the list, it is actually
1528 * in the header and we must copy the following entry up
1529 * to the header. Otherwise we must search the list for
1530 * the entry. In either case we free the now unused entry.
1533 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1534 vm_offset_t va, pmap_inval_info_t info)
1540 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1541 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1542 if (pmap == pv->pv_pmap && va == pv->pv_va)
1546 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1547 if (va == pv->pv_va)
1554 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1555 m->md.pv_list_count--;
1556 if (TAILQ_EMPTY(&m->md.pv_list))
1557 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1558 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1559 ++pmap->pm_generation;
1560 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1568 * Create a pv entry for page at pa for
1572 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1577 pv = get_pv_entry();
1582 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1583 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1584 m->md.pv_list_count++;
1590 * pmap_remove_pte: do the things to unmap a page in a process
1593 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1594 pmap_inval_info_t info)
1599 pmap_inval_add(info, pmap, va);
1600 oldpte = loadandclear(ptq);
1602 pmap->pm_stats.wired_count -= 1;
1604 * Machines that don't support invlpg, also don't support
1605 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1609 cpu_invlpg((void *)va);
1610 KKASSERT(pmap->pm_stats.resident_count > 0);
1611 --pmap->pm_stats.resident_count;
1612 if (oldpte & PG_MANAGED) {
1613 m = PHYS_TO_VM_PAGE(oldpte);
1614 if (oldpte & PG_M) {
1615 #if defined(PMAP_DIAGNOSTIC)
1616 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1618 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1622 if (pmap_track_modified(va))
1626 vm_page_flag_set(m, PG_REFERENCED);
1627 return pmap_remove_entry(pmap, m, va, info);
1629 return pmap_unuse_pt(pmap, va, NULL, info);
1638 * Remove a single page from a process address space.
1640 * This function may not be called from an interrupt if the pmap is
1644 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1649 * if there is no pte for this address, just skip it!!! Otherwise
1650 * get a local va for mappings for this pmap and remove the entry.
1652 if (*pmap_pde(pmap, va) != 0) {
1653 ptq = get_ptbase(pmap) + i386_btop(va);
1655 pmap_remove_pte(pmap, ptq, va, info);
1663 * Remove the given range of addresses from the specified map.
1665 * It is assumed that the start and end are properly
1666 * rounded to the page size.
1668 * This function may not be called from an interrupt if the pmap is
1672 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1676 vm_offset_t ptpaddr;
1677 vm_offset_t sindex, eindex;
1678 struct pmap_inval_info info;
1683 if (pmap->pm_stats.resident_count == 0)
1686 pmap_inval_init(&info);
1689 * special handling of removing one page. a very
1690 * common operation and easy to short circuit some
1693 if (((sva + PAGE_SIZE) == eva) &&
1694 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1695 pmap_remove_page(pmap, sva, &info);
1696 pmap_inval_flush(&info);
1701 * Get a local virtual address for the mappings that are being
1704 sindex = i386_btop(sva);
1705 eindex = i386_btop(eva);
1707 for (; sindex < eindex; sindex = pdnxt) {
1711 * Calculate index for next page table.
1713 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1714 if (pmap->pm_stats.resident_count == 0)
1717 pdirindex = sindex / NPDEPG;
1718 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1719 pmap_inval_add(&info, pmap, -1);
1720 pmap->pm_pdir[pdirindex] = 0;
1721 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1726 * Weed out invalid mappings. Note: we assume that the page
1727 * directory table is always allocated, and in kernel virtual.
1733 * Limit our scan to either the end of the va represented
1734 * by the current page table page, or to the end of the
1735 * range being removed.
1737 if (pdnxt > eindex) {
1742 * NOTE: pmap_remove_pte() can block.
1744 for (; sindex != pdnxt; sindex++) {
1747 ptbase = get_ptbase(pmap);
1748 if (ptbase[sindex] == 0)
1750 va = i386_ptob(sindex);
1751 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1755 pmap_inval_flush(&info);
1761 * Removes this physical page from all physical maps in which it resides.
1762 * Reflects back modify bits to the pager.
1764 * This routine may not be called from an interrupt.
1768 pmap_remove_all(vm_page_t m)
1770 struct pmap_inval_info info;
1771 unsigned *pte, tpte;
1774 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
1777 pmap_inval_init(&info);
1779 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1780 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1781 --pv->pv_pmap->pm_stats.resident_count;
1783 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1784 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1785 tpte = loadandclear(pte);
1788 pv->pv_pmap->pm_stats.wired_count--;
1791 vm_page_flag_set(m, PG_REFERENCED);
1794 * Update the vm_page_t clean and reference bits.
1797 #if defined(PMAP_DIAGNOSTIC)
1798 if (pmap_nw_modified((pt_entry_t) tpte)) {
1800 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1804 if (pmap_track_modified(pv->pv_va))
1807 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1808 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1809 ++pv->pv_pmap->pm_generation;
1810 m->md.pv_list_count--;
1811 if (TAILQ_EMPTY(&m->md.pv_list))
1812 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1813 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1817 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
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++) {
1887 * XXX non-optimal. Note also that there can be
1888 * no pmap_inval_flush() calls until after we modify
1889 * ptbase[sindex] (or otherwise we have to do another
1890 * pmap_inval_add() call).
1892 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1893 pbits = ptbase[sindex];
1895 if (pbits & PG_MANAGED) {
1898 m = PHYS_TO_VM_PAGE(pbits);
1899 vm_page_flag_set(m, PG_REFERENCED);
1903 if (pmap_track_modified(i386_ptob(sindex))) {
1905 m = PHYS_TO_VM_PAGE(pbits);
1914 if (pbits != ptbase[sindex]) {
1915 ptbase[sindex] = pbits;
1919 pmap_inval_flush(&info);
1923 * Insert the given physical page (p) at
1924 * the specified virtual address (v) in the
1925 * target physical map with the protection requested.
1927 * If specified, the page will be wired down, meaning
1928 * that the related pte can not be reclaimed.
1930 * NB: This is the only routine which MAY NOT lazy-evaluate
1931 * or lose information. That is, this routine must actually
1932 * insert this page into the given map NOW.
1935 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1941 vm_offset_t origpte, newpte;
1943 pmap_inval_info info;
1949 #ifdef PMAP_DIAGNOSTIC
1951 panic("pmap_enter: toobig");
1952 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1953 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1955 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
1956 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
1959 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
1960 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
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);
1973 pmap_inval_init(&info);
1974 pte = pmap_pte(pmap, va);
1977 * Page Directory table entry not valid, we need a new PT page
1980 panic("pmap_enter: invalid page directory pdir=%x, va=0x%x\n",
1981 (unsigned) pmap->pm_pdir[PTDPTDI], va);
1984 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
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);
2034 KKASSERT(m->flags & PG_MAPPED);
2039 * Mapping has changed, invalidate old range and fall through to
2040 * handle validating new mapping.
2044 err = pmap_remove_pte(pmap, pte, va, &info);
2046 panic("pmap_enter: pte vanished, va: 0x%x", va);
2050 * Enter on the PV list if part of our managed memory. Note that we
2051 * raise IPL while manipulating pv_table since pmap_enter can be
2052 * called at interrupt time.
2054 if (pmap_initialized &&
2055 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2056 pmap_insert_entry(pmap, va, mpte, m);
2058 vm_page_flag_set(m, PG_MAPPED);
2062 * Increment counters
2064 ++pmap->pm_stats.resident_count;
2066 pmap->pm_stats.wired_count++;
2070 * Now validate mapping with desired protection/wiring.
2072 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2076 if (va < UPT_MIN_ADDRESS)
2078 if (pmap == &kernel_pmap)
2082 * if the mapping or permission bits are different, we need
2083 * to update the pte.
2085 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2086 pmap_inval_add(&info, pmap, va);
2087 *pte = newpte | PG_A;
2089 vm_page_flag_set(m, PG_WRITEABLE);
2091 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2092 pmap_inval_flush(&info);
2096 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2097 * This code also assumes that the pmap has no pre-existing entry for this
2100 * This code currently may only be used on user pmaps, not kernel_pmap.
2103 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2110 pmap_inval_info info;
2112 pmap_inval_init(&info);
2114 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2115 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2118 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2119 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2123 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2126 * Calculate the page table page (mpte), allocating it if necessary.
2128 * A held page table page (mpte), or NULL, is passed onto the
2129 * section following.
2131 if (va < UPT_MIN_ADDRESS) {
2133 * Calculate pagetable page index
2135 ptepindex = va >> PDRSHIFT;
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;
2160 mpte = _pmap_allocpte(pmap, ptepindex);
2162 } while (mpte == NULL);
2165 /* this code path is not yet used */
2169 * With a valid (and held) page directory page, we can just use
2170 * vtopte() to get to the pte. If the pte is already present
2171 * we do not disturb it.
2173 pte = (unsigned *)vtopte(va);
2176 pmap_unwire_pte_hold(pmap, mpte, &info);
2177 pa = VM_PAGE_TO_PHYS(m);
2178 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2183 * Enter on the PV list if part of our managed memory
2185 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2186 pmap_insert_entry(pmap, va, mpte, m);
2187 vm_page_flag_set(m, PG_MAPPED);
2191 * Increment counters
2193 ++pmap->pm_stats.resident_count;
2195 pa = VM_PAGE_TO_PHYS(m);
2198 * Now validate mapping with RO protection
2200 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2201 *pte = pa | PG_V | PG_U;
2203 *pte = pa | PG_V | PG_U | PG_MANAGED;
2204 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2205 pmap_inval_flush(&info);
2209 * Make a temporary mapping for a physical address. This is only intended
2210 * to be used for panic dumps.
2213 pmap_kenter_temporary(vm_paddr_t pa, int i)
2215 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2216 return ((void *)crashdumpmap);
2219 #define MAX_INIT_PT (96)
2222 * This routine preloads the ptes for a given object into the specified pmap.
2223 * This eliminates the blast of soft faults on process startup and
2224 * immediately after an mmap.
2226 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2229 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2230 vm_object_t object, vm_pindex_t pindex,
2231 vm_size_t size, int limit)
2233 struct rb_vm_page_scan_info info;
2238 * We can't preinit if read access isn't set or there is no pmap
2241 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2245 * We can't preinit if the pmap is not the current pmap
2247 lp = curthread->td_lwp;
2248 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2251 psize = i386_btop(size);
2253 if ((object->type != OBJT_VNODE) ||
2254 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2255 (object->resident_page_count > MAX_INIT_PT))) {
2259 if (psize + pindex > object->size) {
2260 if (object->size < pindex)
2262 psize = object->size - pindex;
2269 * Use a red-black scan to traverse the requested range and load
2270 * any valid pages found into the pmap.
2272 * We cannot safely scan the object's memq unless we are in a
2273 * critical section since interrupts can remove pages from objects.
2275 info.start_pindex = pindex;
2276 info.end_pindex = pindex + psize - 1;
2283 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2284 pmap_object_init_pt_callback, &info);
2290 pmap_object_init_pt_callback(vm_page_t p, void *data)
2292 struct rb_vm_page_scan_info *info = data;
2293 vm_pindex_t rel_index;
2295 * don't allow an madvise to blow away our really
2296 * free pages allocating pv entries.
2298 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2299 vmstats.v_free_count < vmstats.v_free_reserved) {
2302 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2303 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2304 if ((p->queue - p->pc) == PQ_CACHE)
2305 vm_page_deactivate(p);
2307 rel_index = p->pindex - info->start_pindex;
2308 pmap_enter_quick(info->pmap,
2309 info->addr + i386_ptob(rel_index), p);
2316 * pmap_prefault provides a quick way of clustering pagefaults into a
2317 * processes address space. It is a "cousin" of pmap_object_init_pt,
2318 * except it runs at page fault time instead of mmap time.
2322 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2324 static int pmap_prefault_pageorder[] = {
2325 -PAGE_SIZE, PAGE_SIZE,
2326 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2327 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2328 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2332 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2343 * We do not currently prefault mappings that use virtual page
2344 * tables. We do not prefault foreign pmaps.
2346 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2348 lp = curthread->td_lwp;
2349 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2352 object = entry->object.vm_object;
2354 starta = addra - PFBAK * PAGE_SIZE;
2355 if (starta < entry->start)
2356 starta = entry->start;
2357 else if (starta > addra)
2361 * critical section protection is required to maintain the
2362 * page/object association, interrupts can free pages and remove
2363 * them from their objects.
2366 for (i = 0; i < PAGEORDER_SIZE; i++) {
2367 vm_object_t lobject;
2370 addr = addra + pmap_prefault_pageorder[i];
2371 if (addr > addra + (PFFOR * PAGE_SIZE))
2374 if (addr < starta || addr >= entry->end)
2377 if ((*pmap_pde(pmap, addr)) == 0)
2380 pte = (unsigned *) vtopte(addr);
2384 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2387 for (m = vm_page_lookup(lobject, pindex);
2388 (!m && (lobject->type == OBJT_DEFAULT) &&
2389 (lobject->backing_object));
2390 lobject = lobject->backing_object
2392 if (lobject->backing_object_offset & PAGE_MASK)
2394 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2395 m = vm_page_lookup(lobject->backing_object, pindex);
2399 * give-up when a page is not in memory
2404 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2406 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2408 if ((m->queue - m->pc) == PQ_CACHE) {
2409 vm_page_deactivate(m);
2412 pmap_enter_quick(pmap, addr, m);
2420 * Routine: pmap_change_wiring
2421 * Function: Change the wiring attribute for a map/virtual-address
2423 * In/out conditions:
2424 * The mapping must already exist in the pmap.
2427 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2434 pte = pmap_pte(pmap, va);
2436 if (wired && !pmap_pte_w(pte))
2437 pmap->pm_stats.wired_count++;
2438 else if (!wired && pmap_pte_w(pte))
2439 pmap->pm_stats.wired_count--;
2442 * Wiring is not a hardware characteristic so there is no need to
2443 * invalidate TLB. However, in an SMP environment we must use
2444 * a locked bus cycle to update the pte (if we are not using
2445 * the pmap_inval_*() API that is)... it's ok to do this for simple
2450 atomic_set_int(pte, PG_W);
2452 atomic_clear_int(pte, PG_W);
2455 atomic_set_int_nonlocked(pte, PG_W);
2457 atomic_clear_int_nonlocked(pte, PG_W);
2464 * Copy the range specified by src_addr/len
2465 * from the source map to the range dst_addr/len
2466 * in the destination map.
2468 * This routine is only advisory and need not do anything.
2471 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2472 vm_size_t len, vm_offset_t src_addr)
2474 pmap_inval_info info;
2476 vm_offset_t end_addr = src_addr + len;
2478 unsigned src_frame, dst_frame;
2481 if (dst_addr != src_addr)
2484 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2485 * valid through blocking calls, and that's just not going to
2492 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2493 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2497 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2498 if (dst_frame != (((unsigned) APTDpde) & PG_FRAME)) {
2499 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2500 /* The page directory is not shared between CPUs */
2503 pmap_inval_init(&info);
2504 pmap_inval_add(&info, dst_pmap, -1);
2505 pmap_inval_add(&info, src_pmap, -1);
2508 * critical section protection is required to maintain the page/object
2509 * association, interrupts can free pages and remove them from
2513 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2514 unsigned *src_pte, *dst_pte;
2515 vm_page_t dstmpte, srcmpte;
2516 vm_offset_t srcptepaddr;
2519 if (addr >= UPT_MIN_ADDRESS)
2520 panic("pmap_copy: invalid to pmap_copy page tables\n");
2523 * Don't let optional prefaulting of pages make us go
2524 * way below the low water mark of free pages or way
2525 * above high water mark of used pv entries.
2527 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2528 pv_entry_count > pv_entry_high_water)
2531 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2532 ptepindex = addr >> PDRSHIFT;
2534 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2535 if (srcptepaddr == 0)
2538 if (srcptepaddr & PG_PS) {
2539 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2540 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2541 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2546 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2547 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2548 (srcmpte->flags & PG_BUSY)) {
2552 if (pdnxt > end_addr)
2555 src_pte = (unsigned *) vtopte(addr);
2556 dst_pte = (unsigned *) avtopte(addr);
2557 while (addr < pdnxt) {
2562 * we only virtual copy managed pages
2564 if ((ptetemp & PG_MANAGED) != 0) {
2566 * We have to check after allocpte for the
2567 * pte still being around... allocpte can
2570 * pmap_allocpte() can block. If we lose
2571 * our page directory mappings we stop.
2573 dstmpte = pmap_allocpte(dst_pmap, addr);
2575 if (src_frame != (((unsigned) PTDpde) & PG_FRAME) ||
2576 dst_frame != (((unsigned) APTDpde) & PG_FRAME)
2578 kprintf("WARNING: pmap_copy: detected and corrected race\n");
2579 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2581 } else if ((*dst_pte == 0) &&
2582 (ptetemp = *src_pte) != 0 &&
2583 (ptetemp & PG_MANAGED)) {
2585 * Clear the modified and
2586 * accessed (referenced) bits
2589 m = PHYS_TO_VM_PAGE(ptetemp);
2590 *dst_pte = ptetemp & ~(PG_M | PG_A);
2591 ++dst_pmap->pm_stats.resident_count;
2592 pmap_insert_entry(dst_pmap, addr,
2594 KKASSERT(m->flags & PG_MAPPED);
2596 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
2597 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2600 if (dstmpte->hold_count >= srcmpte->hold_count)
2610 pmap_inval_flush(&info);
2616 * Zero the specified PA by mapping the page into KVM and clearing its
2619 * This function may be called from an interrupt and no locking is
2623 pmap_zero_page(vm_paddr_t phys)
2625 struct mdglobaldata *gd = mdcpu;
2628 if (*(int *)gd->gd_CMAP3)
2629 panic("pmap_zero_page: CMAP3 busy");
2630 *(int *)gd->gd_CMAP3 =
2631 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2632 cpu_invlpg(gd->gd_CADDR3);
2634 #if defined(I686_CPU)
2635 if (cpu_class == CPUCLASS_686)
2636 i686_pagezero(gd->gd_CADDR3);
2639 bzero(gd->gd_CADDR3, PAGE_SIZE);
2640 *(int *) gd->gd_CMAP3 = 0;
2645 * pmap_page_assertzero:
2647 * Assert that a page is empty, panic if it isn't.
2650 pmap_page_assertzero(vm_paddr_t phys)
2652 struct mdglobaldata *gd = mdcpu;
2656 if (*(int *)gd->gd_CMAP3)
2657 panic("pmap_zero_page: CMAP3 busy");
2658 *(int *)gd->gd_CMAP3 =
2659 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2660 cpu_invlpg(gd->gd_CADDR3);
2661 for (i = 0; i < PAGE_SIZE; i += 4) {
2662 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2663 panic("pmap_page_assertzero() @ %p not zero!\n",
2664 (void *)gd->gd_CADDR3);
2667 *(int *) gd->gd_CMAP3 = 0;
2674 * Zero part of a physical page by mapping it into memory and clearing
2675 * its contents with bzero.
2677 * off and size may not cover an area beyond a single hardware page.
2680 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2682 struct mdglobaldata *gd = mdcpu;
2685 if (*(int *) gd->gd_CMAP3)
2686 panic("pmap_zero_page: CMAP3 busy");
2687 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2688 cpu_invlpg(gd->gd_CADDR3);
2690 #if defined(I686_CPU)
2691 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2692 i686_pagezero(gd->gd_CADDR3);
2695 bzero((char *)gd->gd_CADDR3 + off, size);
2696 *(int *) gd->gd_CMAP3 = 0;
2703 * Copy the physical page from the source PA to the target PA.
2704 * This function may be called from an interrupt. No locking
2708 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2710 struct mdglobaldata *gd = mdcpu;
2713 if (*(int *) gd->gd_CMAP1)
2714 panic("pmap_copy_page: CMAP1 busy");
2715 if (*(int *) gd->gd_CMAP2)
2716 panic("pmap_copy_page: CMAP2 busy");
2718 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2719 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2721 cpu_invlpg(gd->gd_CADDR1);
2722 cpu_invlpg(gd->gd_CADDR2);
2724 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2726 *(int *) gd->gd_CMAP1 = 0;
2727 *(int *) gd->gd_CMAP2 = 0;
2732 * pmap_copy_page_frag:
2734 * Copy the physical page from the source PA to the target PA.
2735 * This function may be called from an interrupt. No locking
2739 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2741 struct mdglobaldata *gd = mdcpu;
2744 if (*(int *) gd->gd_CMAP1)
2745 panic("pmap_copy_page: CMAP1 busy");
2746 if (*(int *) gd->gd_CMAP2)
2747 panic("pmap_copy_page: CMAP2 busy");
2749 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2750 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2752 cpu_invlpg(gd->gd_CADDR1);
2753 cpu_invlpg(gd->gd_CADDR2);
2755 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2756 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2759 *(int *) gd->gd_CMAP1 = 0;
2760 *(int *) gd->gd_CMAP2 = 0;
2765 * Returns true if the pmap's pv is one of the first
2766 * 16 pvs linked to from this page. This count may
2767 * be changed upwards or downwards in the future; it
2768 * is only necessary that true be returned for a small
2769 * subset of pmaps for proper page aging.
2772 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2777 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2782 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2783 if (pv->pv_pmap == pmap) {
2796 * Remove all pages from specified address space
2797 * this aids process exit speeds. Also, this code
2798 * is special cased for current process only, but
2799 * can have the more generic (and slightly slower)
2800 * mode enabled. This is much faster than pmap_remove
2801 * in the case of running down an entire address space.
2804 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2807 unsigned *pte, tpte;
2810 pmap_inval_info info;
2812 int32_t save_generation;
2814 lp = curthread->td_lwp;
2815 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2820 pmap_inval_init(&info);
2822 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2823 if (pv->pv_va >= eva || pv->pv_va < sva) {
2824 npv = TAILQ_NEXT(pv, pv_plist);
2828 KKASSERT(pmap == pv->pv_pmap);
2831 pte = (unsigned *)vtopte(pv->pv_va);
2833 pte = pmap_pte_quick(pmap, pv->pv_va);
2834 if (pmap->pm_active)
2835 pmap_inval_add(&info, pmap, pv->pv_va);
2838 * We cannot remove wired pages from a process' mapping
2842 npv = TAILQ_NEXT(pv, pv_plist);
2845 tpte = loadandclear(pte);
2847 m = PHYS_TO_VM_PAGE(tpte);
2849 KASSERT(m < &vm_page_array[vm_page_array_size],
2850 ("pmap_remove_pages: bad tpte %x", tpte));
2852 KKASSERT(pmap->pm_stats.resident_count > 0);
2853 --pmap->pm_stats.resident_count;
2856 * Update the vm_page_t clean and reference bits.
2862 npv = TAILQ_NEXT(pv, pv_plist);
2863 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2864 save_generation = ++pmap->pm_generation;
2866 m->md.pv_list_count--;
2867 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2868 if (TAILQ_EMPTY(&m->md.pv_list))
2869 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2871 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2875 * Restart the scan if we blocked during the unuse or free
2876 * calls and other removals were made.
2878 if (save_generation != pmap->pm_generation) {
2879 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2880 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2883 pmap_inval_flush(&info);
2888 * pmap_testbit tests bits in pte's
2889 * note that the testbit/clearbit routines are inline,
2890 * and a lot of things compile-time evaluate.
2893 pmap_testbit(vm_page_t m, int bit)
2898 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2901 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2906 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2908 * if the bit being tested is the modified bit, then
2909 * mark clean_map and ptes as never
2912 if (bit & (PG_A|PG_M)) {
2913 if (!pmap_track_modified(pv->pv_va))
2917 #if defined(PMAP_DIAGNOSTIC)
2919 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2923 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2934 * this routine is used to modify bits in ptes
2936 static __inline void
2937 pmap_clearbit(vm_page_t m, int bit)
2939 struct pmap_inval_info info;
2944 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2947 pmap_inval_init(&info);
2951 * Loop over all current mappings setting/clearing as appropos If
2952 * setting RO do we need to clear the VAC?
2954 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2956 * don't write protect pager mappings
2959 if (!pmap_track_modified(pv->pv_va))
2963 #if defined(PMAP_DIAGNOSTIC)
2965 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2971 * Careful here. We can use a locked bus instruction to
2972 * clear PG_A or PG_M safely but we need to synchronize
2973 * with the target cpus when we mess with PG_RW.
2975 * We do not have to force synchronization when clearing
2976 * PG_M even for PTEs generated via virtual memory maps,
2977 * because the virtual kernel will invalidate the pmap
2978 * entry when/if it needs to resynchronize the Modify bit.
2981 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2982 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2989 atomic_clear_int(pte, PG_M|PG_RW);
2992 * The cpu may be trying to set PG_M
2993 * simultaniously with our clearing
2996 if (!atomic_cmpset_int(pte, pbits,
3000 } else if (bit == PG_M) {
3002 * We could also clear PG_RW here to force
3003 * a fault on write to redetect PG_M for
3004 * virtual kernels, but it isn't necessary
3005 * since virtual kernels invalidate the pte
3006 * when they clear the VPTE_M bit in their
3007 * virtual page tables.
3009 atomic_clear_int(pte, PG_M);
3011 atomic_clear_int(pte, bit);
3015 pmap_inval_flush(&info);
3020 * pmap_page_protect:
3022 * Lower the permission for all mappings to a given page.
3025 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3027 if ((prot & VM_PROT_WRITE) == 0) {
3028 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3029 pmap_clearbit(m, PG_RW);
3030 vm_page_flag_clear(m, PG_WRITEABLE);
3038 pmap_phys_address(vm_pindex_t ppn)
3040 return (i386_ptob(ppn));
3044 * pmap_ts_referenced:
3046 * Return a count of reference bits for a page, clearing those bits.
3047 * It is not necessary for every reference bit to be cleared, but it
3048 * is necessary that 0 only be returned when there are truly no
3049 * reference bits set.
3051 * XXX: The exact number of bits to check and clear is a matter that
3052 * should be tested and standardized at some point in the future for
3053 * optimal aging of shared pages.
3056 pmap_ts_referenced(vm_page_t m)
3058 pv_entry_t pv, pvf, pvn;
3062 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3067 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3072 pvn = TAILQ_NEXT(pv, pv_list);
3074 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3076 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3078 if (!pmap_track_modified(pv->pv_va))
3081 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3083 if (pte && (*pte & PG_A)) {
3085 atomic_clear_int(pte, PG_A);
3087 atomic_clear_int_nonlocked(pte, PG_A);
3094 } while ((pv = pvn) != NULL && pv != pvf);
3104 * Return whether or not the specified physical page was modified
3105 * in any physical maps.
3108 pmap_is_modified(vm_page_t m)
3110 return pmap_testbit(m, PG_M);
3114 * Clear the modify bits on the specified physical page.
3117 pmap_clear_modify(vm_page_t m)
3119 pmap_clearbit(m, PG_M);
3123 * pmap_clear_reference:
3125 * Clear the reference bit on the specified physical page.
3128 pmap_clear_reference(vm_page_t m)
3130 pmap_clearbit(m, PG_A);
3134 * Miscellaneous support routines follow
3138 i386_protection_init(void)
3142 kp = protection_codes;
3143 for (prot = 0; prot < 8; prot++) {
3145 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3147 * Read access is also 0. There isn't any execute bit,
3148 * so just make it readable.
3150 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3151 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3152 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3155 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3156 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3157 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3158 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3166 * Map a set of physical memory pages into the kernel virtual
3167 * address space. Return a pointer to where it is mapped. This
3168 * routine is intended to be used for mapping device memory,
3171 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3175 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3177 vm_offset_t va, tmpva, offset;
3180 offset = pa & PAGE_MASK;
3181 size = roundup(offset + size, PAGE_SIZE);
3183 va = kmem_alloc_nofault(&kernel_map, size);
3185 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3188 for (tmpva = va; size > 0;) {
3189 pte = (unsigned *)vtopte(tmpva);
3190 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3198 return ((void *)(va + offset));
3202 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3204 vm_offset_t base, offset;
3206 base = va & PG_FRAME;
3207 offset = va & PAGE_MASK;
3208 size = roundup(offset + size, PAGE_SIZE);
3209 pmap_qremove(va, size >> PAGE_SHIFT);
3210 kmem_free(&kernel_map, base, size);
3214 * perform the pmap work for mincore
3217 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3219 unsigned *ptep, pte;
3223 ptep = pmap_pte(pmap, addr);
3228 if ((pte = *ptep) != 0) {
3231 val = MINCORE_INCORE;
3232 if ((pte & PG_MANAGED) == 0)
3235 pa = pte & PG_FRAME;
3237 m = PHYS_TO_VM_PAGE(pa);
3243 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3245 * Modified by someone
3247 else if (m->dirty || pmap_is_modified(m))
3248 val |= MINCORE_MODIFIED_OTHER;
3253 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3256 * Referenced by someone
3258 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3259 val |= MINCORE_REFERENCED_OTHER;
3260 vm_page_flag_set(m, PG_REFERENCED);
3267 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3268 * vmspace will be ref'd and the old one will be deref'd.
3270 * The vmspace for all lwps associated with the process will be adjusted
3271 * and cr3 will be reloaded if any lwp is the current lwp.
3274 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3276 struct vmspace *oldvm;
3280 oldvm = p->p_vmspace;
3281 if (oldvm != newvm) {
3282 p->p_vmspace = newvm;
3283 KKASSERT(p->p_nthreads == 1);
3284 lp = RB_ROOT(&p->p_lwp_tree);
3285 pmap_setlwpvm(lp, newvm);
3287 sysref_get(&newvm->vm_sysref);
3288 sysref_put(&oldvm->vm_sysref);
3295 * Set the vmspace for a LWP. The vmspace is almost universally set the
3296 * same as the process vmspace, but virtual kernels need to swap out contexts
3297 * on a per-lwp basis.
3300 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3302 struct vmspace *oldvm;
3306 oldvm = lp->lwp_vmspace;
3308 if (oldvm != newvm) {
3309 lp->lwp_vmspace = newvm;
3310 if (curthread->td_lwp == lp) {
3311 pmap = vmspace_pmap(newvm);
3313 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3315 pmap->pm_active |= 1;
3317 #if defined(SWTCH_OPTIM_STATS)
3320 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3321 load_cr3(curthread->td_pcb->pcb_cr3);
3322 pmap = vmspace_pmap(oldvm);
3324 atomic_clear_int(&pmap->pm_active,
3325 1 << mycpu->gd_cpuid);
3327 pmap->pm_active &= ~1;
3335 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3338 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3342 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3347 pmap_get_pgeflag(void)
3354 static void pads (pmap_t pm);
3355 void pmap_pvdump (vm_paddr_t pa);
3357 /* print address space of pmap*/
3364 if (pm == &kernel_pmap)
3367 for (i = 0; i < 1024; i++) {
3368 if (pm->pm_pdir[i]) {
3369 for (j = 0; j < 1024; j++) {
3370 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3371 if (pm == &kernel_pmap && va < KERNBASE)
3373 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3375 ptep = pmap_pte_quick(pm, va);
3376 if (pmap_pte_v(ptep))
3377 kprintf("%x:%x ", va, *(int *) ptep);
3386 pmap_pvdump(vm_paddr_t pa)
3391 kprintf("pa %08llx", (long long)pa);
3392 m = PHYS_TO_VM_PAGE(pa);
3393 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3395 kprintf(" -> pmap %p, va %x, flags %x",
3396 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3398 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);