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.
831 * MPSAFE, INTERRUPT SAFE (cluster callback)
834 pmap_qremove(vm_offset_t va, int count)
838 end_va = va + count*PAGE_SIZE;
840 while (va < end_va) {
843 pte = (unsigned *)vtopte(va);
845 cpu_invlpg((void *)va);
854 * This routine works like vm_page_lookup() but also blocks as long as the
855 * page is busy. This routine does not busy the page it returns.
857 * Unless the caller is managing objects whos pages are in a known state,
858 * the call should be made with a critical section held so the page's object
859 * association remains valid on return.
862 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
867 m = vm_page_lookup(object, pindex);
868 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
874 * Create a new thread and optionally associate it with a (new) process.
875 * NOTE! the new thread's cpu may not equal the current cpu.
878 pmap_init_thread(thread_t td)
880 /* enforce pcb placement */
881 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
882 td->td_savefpu = &td->td_pcb->pcb_save;
883 td->td_sp = (char *)td->td_pcb - 16;
887 * This routine directly affects the fork perf for a process.
890 pmap_init_proc(struct proc *p)
895 * Dispose the UPAGES for a process that has exited.
896 * This routine directly impacts the exit perf of a process.
899 pmap_dispose_proc(struct proc *p)
901 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
904 /***************************************************
905 * Page table page management routines.....
906 ***************************************************/
909 * This routine unholds page table pages, and if the hold count
910 * drops to zero, then it decrements the wire count.
913 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
916 * Wait until we can busy the page ourselves. We cannot have
917 * any active flushes if we block.
919 if (m->flags & PG_BUSY) {
920 pmap_inval_flush(info);
921 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
924 KASSERT(m->queue == PQ_NONE,
925 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
927 if (m->hold_count == 1) {
929 * Unmap the page table page
932 pmap_inval_add(info, pmap, -1);
933 pmap->pm_pdir[m->pindex] = 0;
935 KKASSERT(pmap->pm_stats.resident_count > 0);
936 --pmap->pm_stats.resident_count;
938 if (pmap->pm_ptphint == m)
939 pmap->pm_ptphint = NULL;
942 * This was our last hold, the page had better be unwired
943 * after we decrement wire_count.
945 * FUTURE NOTE: shared page directory page could result in
946 * multiple wire counts.
950 KKASSERT(m->wire_count == 0);
951 --vmstats.v_wire_count;
952 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
954 vm_page_free_zero(m);
957 KKASSERT(m->hold_count > 1);
963 static PMAP_INLINE int
964 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
966 KKASSERT(m->hold_count > 0);
967 if (m->hold_count > 1) {
971 return _pmap_unwire_pte_hold(pmap, m, info);
976 * After removing a page table entry, this routine is used to
977 * conditionally free the page, and manage the hold/wire counts.
980 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
981 pmap_inval_info_t info)
984 if (va >= UPT_MIN_ADDRESS)
988 ptepindex = (va >> PDRSHIFT);
989 if (pmap->pm_ptphint &&
990 (pmap->pm_ptphint->pindex == ptepindex)) {
991 mpte = pmap->pm_ptphint;
993 pmap_inval_flush(info);
994 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
995 pmap->pm_ptphint = mpte;
999 return pmap_unwire_pte_hold(pmap, mpte, info);
1003 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1004 * it, and IdlePTD, represents the template used to update all other pmaps.
1006 * On architectures where the kernel pmap is not integrated into the user
1007 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1008 * kernel_pmap should be used to directly access the kernel_pmap.
1011 pmap_pinit0(struct pmap *pmap)
1014 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1015 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1017 pmap->pm_active = 0;
1018 pmap->pm_ptphint = NULL;
1019 TAILQ_INIT(&pmap->pm_pvlist);
1020 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1024 * Initialize a preallocated and zeroed pmap structure,
1025 * such as one in a vmspace structure.
1028 pmap_pinit(struct pmap *pmap)
1033 * No need to allocate page table space yet but we do need a valid
1034 * page directory table.
1036 if (pmap->pm_pdir == NULL) {
1038 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1042 * Allocate an object for the ptes
1044 if (pmap->pm_pteobj == NULL)
1045 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1048 * Allocate the page directory page, unless we already have
1049 * one cached. If we used the cached page the wire_count will
1050 * already be set appropriately.
1052 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1053 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1054 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1055 pmap->pm_pdirm = ptdpg;
1056 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1057 ptdpg->valid = VM_PAGE_BITS_ALL;
1058 ptdpg->wire_count = 1;
1059 ++vmstats.v_wire_count;
1060 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1062 if ((ptdpg->flags & PG_ZERO) == 0)
1063 bzero(pmap->pm_pdir, PAGE_SIZE);
1065 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1067 /* install self-referential address mapping entry */
1068 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1069 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1072 pmap->pm_active = 0;
1073 pmap->pm_ptphint = NULL;
1074 TAILQ_INIT(&pmap->pm_pvlist);
1075 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1076 pmap->pm_stats.resident_count = 1;
1080 * Clean up a pmap structure so it can be physically freed. This routine
1081 * is called by the vmspace dtor function. A great deal of pmap data is
1082 * left passively mapped to improve vmspace management so we have a bit
1083 * of cleanup work to do here.
1086 pmap_puninit(pmap_t pmap)
1090 KKASSERT(pmap->pm_active == 0);
1091 if ((p = pmap->pm_pdirm) != NULL) {
1092 KKASSERT(pmap->pm_pdir != NULL);
1093 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1095 vmstats.v_wire_count--;
1096 KKASSERT((p->flags & PG_BUSY) == 0);
1098 vm_page_free_zero(p);
1099 pmap->pm_pdirm = NULL;
1101 if (pmap->pm_pdir) {
1102 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1103 pmap->pm_pdir = NULL;
1105 if (pmap->pm_pteobj) {
1106 vm_object_deallocate(pmap->pm_pteobj);
1107 pmap->pm_pteobj = NULL;
1112 * Wire in kernel global address entries. To avoid a race condition
1113 * between pmap initialization and pmap_growkernel, this procedure
1114 * adds the pmap to the master list (which growkernel scans to update),
1115 * then copies the template.
1118 pmap_pinit2(struct pmap *pmap)
1121 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1122 /* XXX copies current process, does not fill in MPPTDI */
1123 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1128 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1129 * 0 on failure (if the procedure had to sleep).
1131 * When asked to remove the page directory page itself, we actually just
1132 * leave it cached so we do not have to incur the SMP inval overhead of
1133 * removing the kernel mapping. pmap_puninit() will take care of it.
1136 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1138 unsigned *pde = (unsigned *) pmap->pm_pdir;
1140 * This code optimizes the case of freeing non-busy
1141 * page-table pages. Those pages are zero now, and
1142 * might as well be placed directly into the zero queue.
1144 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1150 * Remove the page table page from the processes address space.
1153 KKASSERT(pmap->pm_stats.resident_count > 0);
1154 --pmap->pm_stats.resident_count;
1156 if (p->hold_count) {
1157 panic("pmap_release: freeing held page table page");
1159 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1160 pmap->pm_ptphint = NULL;
1163 * We leave the page directory page cached, wired, and mapped in
1164 * the pmap until the dtor function (pmap_puninit()) gets called.
1165 * However, still clean it up so we can set PG_ZERO.
1167 if (p->pindex == PTDPTDI) {
1168 bzero(pde + KPTDI, nkpt * PTESIZE);
1171 vm_page_flag_set(p, PG_ZERO);
1175 vmstats.v_wire_count--;
1176 vm_page_free_zero(p);
1182 * this routine is called if the page table page is not
1186 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1188 vm_offset_t pteva, ptepa;
1192 * Find or fabricate a new pagetable page
1194 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1195 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1197 KASSERT(m->queue == PQ_NONE,
1198 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1201 * Increment the hold count for the page we will be returning to
1207 * It is possible that someone else got in and mapped by the page
1208 * directory page while we were blocked, if so just unbusy and
1209 * return the held page.
1211 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1212 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1217 if (m->wire_count == 0)
1218 vmstats.v_wire_count++;
1223 * Map the pagetable page into the process address space, if
1224 * it isn't already there.
1227 ++pmap->pm_stats.resident_count;
1229 ptepa = VM_PAGE_TO_PHYS(m);
1230 pmap->pm_pdir[ptepindex] =
1231 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1234 * Set the page table hint
1236 pmap->pm_ptphint = m;
1239 * Try to use the new mapping, but if we cannot, then
1240 * do it with the routine that maps the page explicitly.
1242 if ((m->flags & PG_ZERO) == 0) {
1243 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1244 (((unsigned) PTDpde) & PG_FRAME)) {
1245 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1246 bzero((caddr_t) pteva, PAGE_SIZE);
1248 pmap_zero_page(ptepa);
1252 m->valid = VM_PAGE_BITS_ALL;
1253 vm_page_flag_clear(m, PG_ZERO);
1254 vm_page_flag_set(m, PG_MAPPED);
1261 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1268 * Calculate pagetable page index
1270 ptepindex = va >> PDRSHIFT;
1273 * Get the page directory entry
1275 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1278 * This supports switching from a 4MB page to a
1281 if (ptepa & PG_PS) {
1282 pmap->pm_pdir[ptepindex] = 0;
1289 * If the page table page is mapped, we just increment the
1290 * hold count, and activate it.
1294 * In order to get the page table page, try the
1297 if (pmap->pm_ptphint &&
1298 (pmap->pm_ptphint->pindex == ptepindex)) {
1299 m = pmap->pm_ptphint;
1301 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1302 pmap->pm_ptphint = m;
1308 * Here if the pte page isn't mapped, or if it has been deallocated.
1310 return _pmap_allocpte(pmap, ptepindex);
1314 /***************************************************
1315 * Pmap allocation/deallocation routines.
1316 ***************************************************/
1319 * Release any resources held by the given physical map.
1320 * Called when a pmap initialized by pmap_pinit is being released.
1321 * Should only be called if the map contains no valid mappings.
1323 static int pmap_release_callback(struct vm_page *p, void *data);
1326 pmap_release(struct pmap *pmap)
1328 vm_object_t object = pmap->pm_pteobj;
1329 struct rb_vm_page_scan_info info;
1331 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1332 #if defined(DIAGNOSTIC)
1333 if (object->ref_count != 1)
1334 panic("pmap_release: pteobj reference count != 1");
1338 info.object = object;
1340 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1347 info.limit = object->generation;
1349 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1350 pmap_release_callback, &info);
1351 if (info.error == 0 && info.mpte) {
1352 if (!pmap_release_free_page(pmap, info.mpte))
1356 } while (info.error);
1360 pmap_release_callback(struct vm_page *p, void *data)
1362 struct rb_vm_page_scan_info *info = data;
1364 if (p->pindex == PTDPTDI) {
1368 if (!pmap_release_free_page(info->pmap, p)) {
1372 if (info->object->generation != info->limit) {
1380 * Grow the number of kernel page table entries, if needed.
1384 pmap_growkernel(vm_offset_t addr)
1387 vm_offset_t ptppaddr;
1392 if (kernel_vm_end == 0) {
1393 kernel_vm_end = KERNBASE;
1395 while (pdir_pde(PTD, kernel_vm_end)) {
1396 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1400 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1401 while (kernel_vm_end < addr) {
1402 if (pdir_pde(PTD, kernel_vm_end)) {
1403 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1408 * This index is bogus, but out of the way
1410 nkpg = vm_page_alloc(kptobj, nkpt,
1411 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1413 panic("pmap_growkernel: no memory to grow kernel");
1416 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1417 pmap_zero_page(ptppaddr);
1418 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1419 pdir_pde(PTD, kernel_vm_end) = newpdir;
1420 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1424 * This update must be interlocked with pmap_pinit2.
1426 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1427 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1429 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1430 ~(PAGE_SIZE * NPTEPG - 1);
1436 * Retire the given physical map from service.
1437 * Should only be called if the map contains
1438 * no valid mappings.
1441 pmap_destroy(pmap_t pmap)
1448 count = --pmap->pm_count;
1451 panic("destroying a pmap is not yet implemented");
1456 * Add a reference to the specified pmap.
1459 pmap_reference(pmap_t pmap)
1466 /***************************************************
1467 * page management routines.
1468 ***************************************************/
1471 * free the pv_entry back to the free list. This function may be
1472 * called from an interrupt.
1474 static PMAP_INLINE void
1475 free_pv_entry(pv_entry_t pv)
1482 * get a new pv_entry, allocating a block from the system
1483 * when needed. This function may be called from an interrupt.
1489 if (pv_entry_high_water &&
1490 (pv_entry_count > pv_entry_high_water) &&
1491 (pmap_pagedaemon_waken == 0)) {
1492 pmap_pagedaemon_waken = 1;
1493 wakeup (&vm_pages_needed);
1495 return zalloc(pvzone);
1499 * This routine is very drastic, but can save the system
1507 static int warningdone=0;
1509 if (pmap_pagedaemon_waken == 0)
1511 pmap_pagedaemon_waken = 0;
1513 if (warningdone < 5) {
1514 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1518 for(i = 0; i < vm_page_array_size; i++) {
1519 m = &vm_page_array[i];
1520 if (m->wire_count || m->hold_count || m->busy ||
1521 (m->flags & PG_BUSY))
1529 * If it is the first entry on the list, it is actually
1530 * in the header and we must copy the following entry up
1531 * to the header. Otherwise we must search the list for
1532 * the entry. In either case we free the now unused entry.
1535 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1536 vm_offset_t va, pmap_inval_info_t info)
1542 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1543 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1544 if (pmap == pv->pv_pmap && va == pv->pv_va)
1548 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1549 if (va == pv->pv_va)
1556 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1557 m->md.pv_list_count--;
1558 if (TAILQ_EMPTY(&m->md.pv_list))
1559 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1560 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1561 ++pmap->pm_generation;
1562 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1570 * Create a pv entry for page at pa for
1574 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1579 pv = get_pv_entry();
1584 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1585 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1586 m->md.pv_list_count++;
1592 * pmap_remove_pte: do the things to unmap a page in a process
1595 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1596 pmap_inval_info_t info)
1601 pmap_inval_add(info, pmap, va);
1602 oldpte = loadandclear(ptq);
1604 pmap->pm_stats.wired_count -= 1;
1606 * Machines that don't support invlpg, also don't support
1607 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1611 cpu_invlpg((void *)va);
1612 KKASSERT(pmap->pm_stats.resident_count > 0);
1613 --pmap->pm_stats.resident_count;
1614 if (oldpte & PG_MANAGED) {
1615 m = PHYS_TO_VM_PAGE(oldpte);
1616 if (oldpte & PG_M) {
1617 #if defined(PMAP_DIAGNOSTIC)
1618 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1620 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1624 if (pmap_track_modified(va))
1628 vm_page_flag_set(m, PG_REFERENCED);
1629 return pmap_remove_entry(pmap, m, va, info);
1631 return pmap_unuse_pt(pmap, va, NULL, info);
1640 * Remove a single page from a process address space.
1642 * This function may not be called from an interrupt if the pmap is
1646 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1651 * if there is no pte for this address, just skip it!!! Otherwise
1652 * get a local va for mappings for this pmap and remove the entry.
1654 if (*pmap_pde(pmap, va) != 0) {
1655 ptq = get_ptbase(pmap) + i386_btop(va);
1657 pmap_remove_pte(pmap, ptq, va, info);
1665 * Remove the given range of addresses from the specified map.
1667 * It is assumed that the start and end are properly
1668 * rounded to the page size.
1670 * This function may not be called from an interrupt if the pmap is
1674 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1678 vm_offset_t ptpaddr;
1679 vm_offset_t sindex, eindex;
1680 struct pmap_inval_info info;
1685 if (pmap->pm_stats.resident_count == 0)
1688 pmap_inval_init(&info);
1691 * special handling of removing one page. a very
1692 * common operation and easy to short circuit some
1695 if (((sva + PAGE_SIZE) == eva) &&
1696 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1697 pmap_remove_page(pmap, sva, &info);
1698 pmap_inval_flush(&info);
1703 * Get a local virtual address for the mappings that are being
1706 sindex = i386_btop(sva);
1707 eindex = i386_btop(eva);
1709 for (; sindex < eindex; sindex = pdnxt) {
1713 * Calculate index for next page table.
1715 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1716 if (pmap->pm_stats.resident_count == 0)
1719 pdirindex = sindex / NPDEPG;
1720 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1721 pmap_inval_add(&info, pmap, -1);
1722 pmap->pm_pdir[pdirindex] = 0;
1723 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1728 * Weed out invalid mappings. Note: we assume that the page
1729 * directory table is always allocated, and in kernel virtual.
1735 * Limit our scan to either the end of the va represented
1736 * by the current page table page, or to the end of the
1737 * range being removed.
1739 if (pdnxt > eindex) {
1744 * NOTE: pmap_remove_pte() can block.
1746 for (; sindex != pdnxt; sindex++) {
1749 ptbase = get_ptbase(pmap);
1750 if (ptbase[sindex] == 0)
1752 va = i386_ptob(sindex);
1753 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1757 pmap_inval_flush(&info);
1763 * Removes this physical page from all physical maps in which it resides.
1764 * Reflects back modify bits to the pager.
1766 * This routine may not be called from an interrupt.
1770 pmap_remove_all(vm_page_t m)
1772 struct pmap_inval_info info;
1773 unsigned *pte, tpte;
1776 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
1779 pmap_inval_init(&info);
1781 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1782 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1783 --pv->pv_pmap->pm_stats.resident_count;
1785 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1786 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1787 tpte = loadandclear(pte);
1790 pv->pv_pmap->pm_stats.wired_count--;
1793 vm_page_flag_set(m, PG_REFERENCED);
1796 * Update the vm_page_t clean and reference bits.
1799 #if defined(PMAP_DIAGNOSTIC)
1800 if (pmap_nw_modified((pt_entry_t) tpte)) {
1802 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1806 if (pmap_track_modified(pv->pv_va))
1809 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1810 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1811 ++pv->pv_pmap->pm_generation;
1812 m->md.pv_list_count--;
1813 if (TAILQ_EMPTY(&m->md.pv_list))
1814 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1815 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1819 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
1820 pmap_inval_flush(&info);
1826 * Set the physical protection on the specified range of this map
1829 * This function may not be called from an interrupt if the map is
1830 * not the kernel_pmap.
1833 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1836 vm_offset_t pdnxt, ptpaddr;
1837 vm_pindex_t sindex, eindex;
1838 pmap_inval_info info;
1843 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1844 pmap_remove(pmap, sva, eva);
1848 if (prot & VM_PROT_WRITE)
1851 pmap_inval_init(&info);
1853 ptbase = get_ptbase(pmap);
1855 sindex = i386_btop(sva);
1856 eindex = i386_btop(eva);
1858 for (; sindex < eindex; sindex = pdnxt) {
1862 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1864 pdirindex = sindex / NPDEPG;
1865 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1866 pmap_inval_add(&info, pmap, -1);
1867 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1868 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1873 * Weed out invalid mappings. Note: we assume that the page
1874 * directory table is always allocated, and in kernel virtual.
1879 if (pdnxt > eindex) {
1883 for (; sindex != pdnxt; sindex++) {
1889 * XXX non-optimal. Note also that there can be
1890 * no pmap_inval_flush() calls until after we modify
1891 * ptbase[sindex] (or otherwise we have to do another
1892 * pmap_inval_add() call).
1894 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1895 pbits = ptbase[sindex];
1897 if (pbits & PG_MANAGED) {
1900 m = PHYS_TO_VM_PAGE(pbits);
1901 vm_page_flag_set(m, PG_REFERENCED);
1905 if (pmap_track_modified(i386_ptob(sindex))) {
1907 m = PHYS_TO_VM_PAGE(pbits);
1916 if (pbits != ptbase[sindex]) {
1917 ptbase[sindex] = pbits;
1921 pmap_inval_flush(&info);
1925 * Insert the given physical page (p) at
1926 * the specified virtual address (v) in the
1927 * target physical map with the protection requested.
1929 * If specified, the page will be wired down, meaning
1930 * that the related pte can not be reclaimed.
1932 * NB: This is the only routine which MAY NOT lazy-evaluate
1933 * or lose information. That is, this routine must actually
1934 * insert this page into the given map NOW.
1937 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1943 vm_offset_t origpte, newpte;
1945 pmap_inval_info info;
1951 #ifdef PMAP_DIAGNOSTIC
1953 panic("pmap_enter: toobig");
1954 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1955 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1957 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
1958 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
1961 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
1962 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
1967 * In the case that a page table page is not
1968 * resident, we are creating it here.
1970 if (va < UPT_MIN_ADDRESS)
1971 mpte = pmap_allocpte(pmap, va);
1975 pmap_inval_init(&info);
1976 pte = pmap_pte(pmap, va);
1979 * Page Directory table entry not valid, we need a new PT page
1982 panic("pmap_enter: invalid page directory pdir=%x, va=0x%x\n",
1983 (unsigned) pmap->pm_pdir[PTDPTDI], va);
1986 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
1987 origpte = *(vm_offset_t *)pte;
1988 opa = origpte & PG_FRAME;
1990 if (origpte & PG_PS)
1991 panic("pmap_enter: attempted pmap_enter on 4MB page");
1994 * Mapping has not changed, must be protection or wiring change.
1996 if (origpte && (opa == pa)) {
1998 * Wiring change, just update stats. We don't worry about
1999 * wiring PT pages as they remain resident as long as there
2000 * are valid mappings in them. Hence, if a user page is wired,
2001 * the PT page will be also.
2003 if (wired && ((origpte & PG_W) == 0))
2004 pmap->pm_stats.wired_count++;
2005 else if (!wired && (origpte & PG_W))
2006 pmap->pm_stats.wired_count--;
2008 #if defined(PMAP_DIAGNOSTIC)
2009 if (pmap_nw_modified((pt_entry_t) origpte)) {
2011 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
2017 * Remove the extra pte reference. Note that we cannot
2018 * optimize the RO->RW case because we have adjusted the
2019 * wiring count above and may need to adjust the wiring
2026 * We might be turning off write access to the page,
2027 * so we go ahead and sense modify status.
2029 if (origpte & PG_MANAGED) {
2030 if ((origpte & PG_M) && pmap_track_modified(va)) {
2032 om = PHYS_TO_VM_PAGE(opa);
2036 KKASSERT(m->flags & PG_MAPPED);
2041 * Mapping has changed, invalidate old range and fall through to
2042 * handle validating new mapping.
2046 err = pmap_remove_pte(pmap, pte, va, &info);
2048 panic("pmap_enter: pte vanished, va: 0x%x", va);
2052 * Enter on the PV list if part of our managed memory. Note that we
2053 * raise IPL while manipulating pv_table since pmap_enter can be
2054 * called at interrupt time.
2056 if (pmap_initialized &&
2057 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2058 pmap_insert_entry(pmap, va, mpte, m);
2060 vm_page_flag_set(m, PG_MAPPED);
2064 * Increment counters
2066 ++pmap->pm_stats.resident_count;
2068 pmap->pm_stats.wired_count++;
2072 * Now validate mapping with desired protection/wiring.
2074 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2078 if (va < UPT_MIN_ADDRESS)
2080 if (pmap == &kernel_pmap)
2084 * if the mapping or permission bits are different, we need
2085 * to update the pte.
2087 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2088 pmap_inval_add(&info, pmap, va);
2089 *pte = newpte | PG_A;
2091 vm_page_flag_set(m, PG_WRITEABLE);
2093 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2094 pmap_inval_flush(&info);
2098 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2099 * This code also assumes that the pmap has no pre-existing entry for this
2102 * This code currently may only be used on user pmaps, not kernel_pmap.
2105 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2112 pmap_inval_info info;
2114 pmap_inval_init(&info);
2116 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2117 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2120 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2121 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2125 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2128 * Calculate the page table page (mpte), allocating it if necessary.
2130 * A held page table page (mpte), or NULL, is passed onto the
2131 * section following.
2133 if (va < UPT_MIN_ADDRESS) {
2135 * Calculate pagetable page index
2137 ptepindex = va >> PDRSHIFT;
2141 * Get the page directory entry
2143 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2146 * If the page table page is mapped, we just increment
2147 * the hold count, and activate it.
2151 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2152 if (pmap->pm_ptphint &&
2153 (pmap->pm_ptphint->pindex == ptepindex)) {
2154 mpte = pmap->pm_ptphint;
2156 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2157 pmap->pm_ptphint = mpte;
2162 mpte = _pmap_allocpte(pmap, ptepindex);
2164 } while (mpte == NULL);
2167 /* this code path is not yet used */
2171 * With a valid (and held) page directory page, we can just use
2172 * vtopte() to get to the pte. If the pte is already present
2173 * we do not disturb it.
2175 pte = (unsigned *)vtopte(va);
2178 pmap_unwire_pte_hold(pmap, mpte, &info);
2179 pa = VM_PAGE_TO_PHYS(m);
2180 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2185 * Enter on the PV list if part of our managed memory
2187 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2188 pmap_insert_entry(pmap, va, mpte, m);
2189 vm_page_flag_set(m, PG_MAPPED);
2193 * Increment counters
2195 ++pmap->pm_stats.resident_count;
2197 pa = VM_PAGE_TO_PHYS(m);
2200 * Now validate mapping with RO protection
2202 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2203 *pte = pa | PG_V | PG_U;
2205 *pte = pa | PG_V | PG_U | PG_MANAGED;
2206 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2207 pmap_inval_flush(&info);
2211 * Make a temporary mapping for a physical address. This is only intended
2212 * to be used for panic dumps.
2215 pmap_kenter_temporary(vm_paddr_t pa, int i)
2217 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2218 return ((void *)crashdumpmap);
2221 #define MAX_INIT_PT (96)
2224 * This routine preloads the ptes for a given object into the specified pmap.
2225 * This eliminates the blast of soft faults on process startup and
2226 * immediately after an mmap.
2228 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2231 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2232 vm_object_t object, vm_pindex_t pindex,
2233 vm_size_t size, int limit)
2235 struct rb_vm_page_scan_info info;
2240 * We can't preinit if read access isn't set or there is no pmap
2243 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2247 * We can't preinit if the pmap is not the current pmap
2249 lp = curthread->td_lwp;
2250 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2253 psize = i386_btop(size);
2255 if ((object->type != OBJT_VNODE) ||
2256 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2257 (object->resident_page_count > MAX_INIT_PT))) {
2261 if (psize + pindex > object->size) {
2262 if (object->size < pindex)
2264 psize = object->size - pindex;
2271 * Use a red-black scan to traverse the requested range and load
2272 * any valid pages found into the pmap.
2274 * We cannot safely scan the object's memq unless we are in a
2275 * critical section since interrupts can remove pages from objects.
2277 info.start_pindex = pindex;
2278 info.end_pindex = pindex + psize - 1;
2285 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2286 pmap_object_init_pt_callback, &info);
2292 pmap_object_init_pt_callback(vm_page_t p, void *data)
2294 struct rb_vm_page_scan_info *info = data;
2295 vm_pindex_t rel_index;
2297 * don't allow an madvise to blow away our really
2298 * free pages allocating pv entries.
2300 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2301 vmstats.v_free_count < vmstats.v_free_reserved) {
2304 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2305 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2306 if ((p->queue - p->pc) == PQ_CACHE)
2307 vm_page_deactivate(p);
2309 rel_index = p->pindex - info->start_pindex;
2310 pmap_enter_quick(info->pmap,
2311 info->addr + i386_ptob(rel_index), p);
2318 * pmap_prefault provides a quick way of clustering pagefaults into a
2319 * processes address space. It is a "cousin" of pmap_object_init_pt,
2320 * except it runs at page fault time instead of mmap time.
2324 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2326 static int pmap_prefault_pageorder[] = {
2327 -PAGE_SIZE, PAGE_SIZE,
2328 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2329 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2330 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2334 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2345 * We do not currently prefault mappings that use virtual page
2346 * tables. We do not prefault foreign pmaps.
2348 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2350 lp = curthread->td_lwp;
2351 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2354 object = entry->object.vm_object;
2356 starta = addra - PFBAK * PAGE_SIZE;
2357 if (starta < entry->start)
2358 starta = entry->start;
2359 else if (starta > addra)
2363 * critical section protection is required to maintain the
2364 * page/object association, interrupts can free pages and remove
2365 * them from their objects.
2368 for (i = 0; i < PAGEORDER_SIZE; i++) {
2369 vm_object_t lobject;
2372 addr = addra + pmap_prefault_pageorder[i];
2373 if (addr > addra + (PFFOR * PAGE_SIZE))
2376 if (addr < starta || addr >= entry->end)
2379 if ((*pmap_pde(pmap, addr)) == 0)
2382 pte = (unsigned *) vtopte(addr);
2386 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2389 for (m = vm_page_lookup(lobject, pindex);
2390 (!m && (lobject->type == OBJT_DEFAULT) &&
2391 (lobject->backing_object));
2392 lobject = lobject->backing_object
2394 if (lobject->backing_object_offset & PAGE_MASK)
2396 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2397 m = vm_page_lookup(lobject->backing_object, pindex);
2401 * give-up when a page is not in memory
2406 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2408 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2410 if ((m->queue - m->pc) == PQ_CACHE) {
2411 vm_page_deactivate(m);
2414 pmap_enter_quick(pmap, addr, m);
2422 * Routine: pmap_change_wiring
2423 * Function: Change the wiring attribute for a map/virtual-address
2425 * In/out conditions:
2426 * The mapping must already exist in the pmap.
2429 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2436 pte = pmap_pte(pmap, va);
2438 if (wired && !pmap_pte_w(pte))
2439 pmap->pm_stats.wired_count++;
2440 else if (!wired && pmap_pte_w(pte))
2441 pmap->pm_stats.wired_count--;
2444 * Wiring is not a hardware characteristic so there is no need to
2445 * invalidate TLB. However, in an SMP environment we must use
2446 * a locked bus cycle to update the pte (if we are not using
2447 * the pmap_inval_*() API that is)... it's ok to do this for simple
2452 atomic_set_int(pte, PG_W);
2454 atomic_clear_int(pte, PG_W);
2457 atomic_set_int_nonlocked(pte, PG_W);
2459 atomic_clear_int_nonlocked(pte, PG_W);
2466 * Copy the range specified by src_addr/len
2467 * from the source map to the range dst_addr/len
2468 * in the destination map.
2470 * This routine is only advisory and need not do anything.
2473 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2474 vm_size_t len, vm_offset_t src_addr)
2476 pmap_inval_info info;
2478 vm_offset_t end_addr = src_addr + len;
2480 unsigned src_frame, dst_frame;
2483 if (dst_addr != src_addr)
2486 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2487 * valid through blocking calls, and that's just not going to
2494 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2495 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2499 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2500 if (dst_frame != (((unsigned) APTDpde) & PG_FRAME)) {
2501 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2502 /* The page directory is not shared between CPUs */
2505 pmap_inval_init(&info);
2506 pmap_inval_add(&info, dst_pmap, -1);
2507 pmap_inval_add(&info, src_pmap, -1);
2510 * critical section protection is required to maintain the page/object
2511 * association, interrupts can free pages and remove them from
2515 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2516 unsigned *src_pte, *dst_pte;
2517 vm_page_t dstmpte, srcmpte;
2518 vm_offset_t srcptepaddr;
2521 if (addr >= UPT_MIN_ADDRESS)
2522 panic("pmap_copy: invalid to pmap_copy page tables\n");
2525 * Don't let optional prefaulting of pages make us go
2526 * way below the low water mark of free pages or way
2527 * above high water mark of used pv entries.
2529 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2530 pv_entry_count > pv_entry_high_water)
2533 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2534 ptepindex = addr >> PDRSHIFT;
2536 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2537 if (srcptepaddr == 0)
2540 if (srcptepaddr & PG_PS) {
2541 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2542 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2543 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2548 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2549 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2550 (srcmpte->flags & PG_BUSY)) {
2554 if (pdnxt > end_addr)
2557 src_pte = (unsigned *) vtopte(addr);
2558 dst_pte = (unsigned *) avtopte(addr);
2559 while (addr < pdnxt) {
2564 * we only virtual copy managed pages
2566 if ((ptetemp & PG_MANAGED) != 0) {
2568 * We have to check after allocpte for the
2569 * pte still being around... allocpte can
2572 * pmap_allocpte() can block. If we lose
2573 * our page directory mappings we stop.
2575 dstmpte = pmap_allocpte(dst_pmap, addr);
2577 if (src_frame != (((unsigned) PTDpde) & PG_FRAME) ||
2578 dst_frame != (((unsigned) APTDpde) & PG_FRAME)
2580 kprintf("WARNING: pmap_copy: detected and corrected race\n");
2581 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2583 } else if ((*dst_pte == 0) &&
2584 (ptetemp = *src_pte) != 0 &&
2585 (ptetemp & PG_MANAGED)) {
2587 * Clear the modified and
2588 * accessed (referenced) bits
2591 m = PHYS_TO_VM_PAGE(ptetemp);
2592 *dst_pte = ptetemp & ~(PG_M | PG_A);
2593 ++dst_pmap->pm_stats.resident_count;
2594 pmap_insert_entry(dst_pmap, addr,
2596 KKASSERT(m->flags & PG_MAPPED);
2598 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
2599 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2602 if (dstmpte->hold_count >= srcmpte->hold_count)
2612 pmap_inval_flush(&info);
2618 * Zero the specified PA by mapping the page into KVM and clearing its
2621 * This function may be called from an interrupt and no locking is
2625 pmap_zero_page(vm_paddr_t phys)
2627 struct mdglobaldata *gd = mdcpu;
2630 if (*(int *)gd->gd_CMAP3)
2631 panic("pmap_zero_page: CMAP3 busy");
2632 *(int *)gd->gd_CMAP3 =
2633 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2634 cpu_invlpg(gd->gd_CADDR3);
2636 #if defined(I686_CPU)
2637 if (cpu_class == CPUCLASS_686)
2638 i686_pagezero(gd->gd_CADDR3);
2641 bzero(gd->gd_CADDR3, PAGE_SIZE);
2642 *(int *) gd->gd_CMAP3 = 0;
2647 * pmap_page_assertzero:
2649 * Assert that a page is empty, panic if it isn't.
2652 pmap_page_assertzero(vm_paddr_t phys)
2654 struct mdglobaldata *gd = mdcpu;
2658 if (*(int *)gd->gd_CMAP3)
2659 panic("pmap_zero_page: CMAP3 busy");
2660 *(int *)gd->gd_CMAP3 =
2661 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2662 cpu_invlpg(gd->gd_CADDR3);
2663 for (i = 0; i < PAGE_SIZE; i += 4) {
2664 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2665 panic("pmap_page_assertzero() @ %p not zero!\n",
2666 (void *)gd->gd_CADDR3);
2669 *(int *) gd->gd_CMAP3 = 0;
2676 * Zero part of a physical page by mapping it into memory and clearing
2677 * its contents with bzero.
2679 * off and size may not cover an area beyond a single hardware page.
2682 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2684 struct mdglobaldata *gd = mdcpu;
2687 if (*(int *) gd->gd_CMAP3)
2688 panic("pmap_zero_page: CMAP3 busy");
2689 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2690 cpu_invlpg(gd->gd_CADDR3);
2692 #if defined(I686_CPU)
2693 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2694 i686_pagezero(gd->gd_CADDR3);
2697 bzero((char *)gd->gd_CADDR3 + off, size);
2698 *(int *) gd->gd_CMAP3 = 0;
2705 * Copy the physical page from the source PA to the target PA.
2706 * This function may be called from an interrupt. No locking
2710 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2712 struct mdglobaldata *gd = mdcpu;
2715 if (*(int *) gd->gd_CMAP1)
2716 panic("pmap_copy_page: CMAP1 busy");
2717 if (*(int *) gd->gd_CMAP2)
2718 panic("pmap_copy_page: CMAP2 busy");
2720 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2721 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2723 cpu_invlpg(gd->gd_CADDR1);
2724 cpu_invlpg(gd->gd_CADDR2);
2726 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2728 *(int *) gd->gd_CMAP1 = 0;
2729 *(int *) gd->gd_CMAP2 = 0;
2734 * pmap_copy_page_frag:
2736 * Copy the physical page from the source PA to the target PA.
2737 * This function may be called from an interrupt. No locking
2741 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2743 struct mdglobaldata *gd = mdcpu;
2746 if (*(int *) gd->gd_CMAP1)
2747 panic("pmap_copy_page: CMAP1 busy");
2748 if (*(int *) gd->gd_CMAP2)
2749 panic("pmap_copy_page: CMAP2 busy");
2751 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2752 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2754 cpu_invlpg(gd->gd_CADDR1);
2755 cpu_invlpg(gd->gd_CADDR2);
2757 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2758 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2761 *(int *) gd->gd_CMAP1 = 0;
2762 *(int *) gd->gd_CMAP2 = 0;
2767 * Returns true if the pmap's pv is one of the first
2768 * 16 pvs linked to from this page. This count may
2769 * be changed upwards or downwards in the future; it
2770 * is only necessary that true be returned for a small
2771 * subset of pmaps for proper page aging.
2774 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2779 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2784 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2785 if (pv->pv_pmap == pmap) {
2798 * Remove all pages from specified address space
2799 * this aids process exit speeds. Also, this code
2800 * is special cased for current process only, but
2801 * can have the more generic (and slightly slower)
2802 * mode enabled. This is much faster than pmap_remove
2803 * in the case of running down an entire address space.
2806 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2809 unsigned *pte, tpte;
2812 pmap_inval_info info;
2814 int32_t save_generation;
2816 lp = curthread->td_lwp;
2817 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2822 pmap_inval_init(&info);
2824 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2825 if (pv->pv_va >= eva || pv->pv_va < sva) {
2826 npv = TAILQ_NEXT(pv, pv_plist);
2830 KKASSERT(pmap == pv->pv_pmap);
2833 pte = (unsigned *)vtopte(pv->pv_va);
2835 pte = pmap_pte_quick(pmap, pv->pv_va);
2836 if (pmap->pm_active)
2837 pmap_inval_add(&info, pmap, pv->pv_va);
2840 * We cannot remove wired pages from a process' mapping
2844 npv = TAILQ_NEXT(pv, pv_plist);
2847 tpte = loadandclear(pte);
2849 m = PHYS_TO_VM_PAGE(tpte);
2851 KASSERT(m < &vm_page_array[vm_page_array_size],
2852 ("pmap_remove_pages: bad tpte %x", tpte));
2854 KKASSERT(pmap->pm_stats.resident_count > 0);
2855 --pmap->pm_stats.resident_count;
2858 * Update the vm_page_t clean and reference bits.
2864 npv = TAILQ_NEXT(pv, pv_plist);
2865 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2866 save_generation = ++pmap->pm_generation;
2868 m->md.pv_list_count--;
2869 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2870 if (TAILQ_EMPTY(&m->md.pv_list))
2871 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2873 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2877 * Restart the scan if we blocked during the unuse or free
2878 * calls and other removals were made.
2880 if (save_generation != pmap->pm_generation) {
2881 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2882 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2885 pmap_inval_flush(&info);
2890 * pmap_testbit tests bits in pte's
2891 * note that the testbit/clearbit routines are inline,
2892 * and a lot of things compile-time evaluate.
2895 pmap_testbit(vm_page_t m, int bit)
2900 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2903 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2908 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2910 * if the bit being tested is the modified bit, then
2911 * mark clean_map and ptes as never
2914 if (bit & (PG_A|PG_M)) {
2915 if (!pmap_track_modified(pv->pv_va))
2919 #if defined(PMAP_DIAGNOSTIC)
2921 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2925 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2936 * this routine is used to modify bits in ptes
2938 static __inline void
2939 pmap_clearbit(vm_page_t m, int bit)
2941 struct pmap_inval_info info;
2946 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2949 pmap_inval_init(&info);
2953 * Loop over all current mappings setting/clearing as appropos If
2954 * setting RO do we need to clear the VAC?
2956 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2958 * don't write protect pager mappings
2961 if (!pmap_track_modified(pv->pv_va))
2965 #if defined(PMAP_DIAGNOSTIC)
2967 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2973 * Careful here. We can use a locked bus instruction to
2974 * clear PG_A or PG_M safely but we need to synchronize
2975 * with the target cpus when we mess with PG_RW.
2977 * We do not have to force synchronization when clearing
2978 * PG_M even for PTEs generated via virtual memory maps,
2979 * because the virtual kernel will invalidate the pmap
2980 * entry when/if it needs to resynchronize the Modify bit.
2983 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2984 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2991 atomic_clear_int(pte, PG_M|PG_RW);
2994 * The cpu may be trying to set PG_M
2995 * simultaniously with our clearing
2998 if (!atomic_cmpset_int(pte, pbits,
3002 } else if (bit == PG_M) {
3004 * We could also clear PG_RW here to force
3005 * a fault on write to redetect PG_M for
3006 * virtual kernels, but it isn't necessary
3007 * since virtual kernels invalidate the pte
3008 * when they clear the VPTE_M bit in their
3009 * virtual page tables.
3011 atomic_clear_int(pte, PG_M);
3013 atomic_clear_int(pte, bit);
3017 pmap_inval_flush(&info);
3022 * pmap_page_protect:
3024 * Lower the permission for all mappings to a given page.
3027 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3029 if ((prot & VM_PROT_WRITE) == 0) {
3030 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3031 pmap_clearbit(m, PG_RW);
3032 vm_page_flag_clear(m, PG_WRITEABLE);
3040 pmap_phys_address(vm_pindex_t ppn)
3042 return (i386_ptob(ppn));
3046 * pmap_ts_referenced:
3048 * Return a count of reference bits for a page, clearing those bits.
3049 * It is not necessary for every reference bit to be cleared, but it
3050 * is necessary that 0 only be returned when there are truly no
3051 * reference bits set.
3053 * XXX: The exact number of bits to check and clear is a matter that
3054 * should be tested and standardized at some point in the future for
3055 * optimal aging of shared pages.
3058 pmap_ts_referenced(vm_page_t m)
3060 pv_entry_t pv, pvf, pvn;
3064 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3069 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3074 pvn = TAILQ_NEXT(pv, pv_list);
3076 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3078 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3080 if (!pmap_track_modified(pv->pv_va))
3083 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3085 if (pte && (*pte & PG_A)) {
3087 atomic_clear_int(pte, PG_A);
3089 atomic_clear_int_nonlocked(pte, PG_A);
3096 } while ((pv = pvn) != NULL && pv != pvf);
3106 * Return whether or not the specified physical page was modified
3107 * in any physical maps.
3110 pmap_is_modified(vm_page_t m)
3112 return pmap_testbit(m, PG_M);
3116 * Clear the modify bits on the specified physical page.
3119 pmap_clear_modify(vm_page_t m)
3121 pmap_clearbit(m, PG_M);
3125 * pmap_clear_reference:
3127 * Clear the reference bit on the specified physical page.
3130 pmap_clear_reference(vm_page_t m)
3132 pmap_clearbit(m, PG_A);
3136 * Miscellaneous support routines follow
3140 i386_protection_init(void)
3144 kp = protection_codes;
3145 for (prot = 0; prot < 8; prot++) {
3147 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3149 * Read access is also 0. There isn't any execute bit,
3150 * so just make it readable.
3152 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3153 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3154 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3157 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3158 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3159 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3160 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3168 * Map a set of physical memory pages into the kernel virtual
3169 * address space. Return a pointer to where it is mapped. This
3170 * routine is intended to be used for mapping device memory,
3173 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3177 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3179 vm_offset_t va, tmpva, offset;
3182 offset = pa & PAGE_MASK;
3183 size = roundup(offset + size, PAGE_SIZE);
3185 va = kmem_alloc_nofault(&kernel_map, size);
3187 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3190 for (tmpva = va; size > 0;) {
3191 pte = (unsigned *)vtopte(tmpva);
3192 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3200 return ((void *)(va + offset));
3204 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3206 vm_offset_t base, offset;
3208 base = va & PG_FRAME;
3209 offset = va & PAGE_MASK;
3210 size = roundup(offset + size, PAGE_SIZE);
3211 pmap_qremove(va, size >> PAGE_SHIFT);
3212 kmem_free(&kernel_map, base, size);
3216 * perform the pmap work for mincore
3219 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3221 unsigned *ptep, pte;
3225 ptep = pmap_pte(pmap, addr);
3230 if ((pte = *ptep) != 0) {
3233 val = MINCORE_INCORE;
3234 if ((pte & PG_MANAGED) == 0)
3237 pa = pte & PG_FRAME;
3239 m = PHYS_TO_VM_PAGE(pa);
3245 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3247 * Modified by someone
3249 else if (m->dirty || pmap_is_modified(m))
3250 val |= MINCORE_MODIFIED_OTHER;
3255 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3258 * Referenced by someone
3260 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3261 val |= MINCORE_REFERENCED_OTHER;
3262 vm_page_flag_set(m, PG_REFERENCED);
3269 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3270 * vmspace will be ref'd and the old one will be deref'd.
3272 * The vmspace for all lwps associated with the process will be adjusted
3273 * and cr3 will be reloaded if any lwp is the current lwp.
3276 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3278 struct vmspace *oldvm;
3282 oldvm = p->p_vmspace;
3283 if (oldvm != newvm) {
3284 p->p_vmspace = newvm;
3285 KKASSERT(p->p_nthreads == 1);
3286 lp = RB_ROOT(&p->p_lwp_tree);
3287 pmap_setlwpvm(lp, newvm);
3289 sysref_get(&newvm->vm_sysref);
3290 sysref_put(&oldvm->vm_sysref);
3297 * Set the vmspace for a LWP. The vmspace is almost universally set the
3298 * same as the process vmspace, but virtual kernels need to swap out contexts
3299 * on a per-lwp basis.
3302 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3304 struct vmspace *oldvm;
3308 oldvm = lp->lwp_vmspace;
3310 if (oldvm != newvm) {
3311 lp->lwp_vmspace = newvm;
3312 if (curthread->td_lwp == lp) {
3313 pmap = vmspace_pmap(newvm);
3315 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3317 pmap->pm_active |= 1;
3319 #if defined(SWTCH_OPTIM_STATS)
3322 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3323 load_cr3(curthread->td_pcb->pcb_cr3);
3324 pmap = vmspace_pmap(oldvm);
3326 atomic_clear_int(&pmap->pm_active,
3327 1 << mycpu->gd_cpuid);
3329 pmap->pm_active &= ~1;
3337 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3340 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3344 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3349 pmap_get_pgeflag(void)
3356 static void pads (pmap_t pm);
3357 void pmap_pvdump (vm_paddr_t pa);
3359 /* print address space of pmap*/
3366 if (pm == &kernel_pmap)
3369 for (i = 0; i < 1024; i++) {
3370 if (pm->pm_pdir[i]) {
3371 for (j = 0; j < 1024; j++) {
3372 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3373 if (pm == &kernel_pmap && va < KERNBASE)
3375 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3377 ptep = pmap_pte_quick(pm, va);
3378 if (pmap_pte_v(ptep))
3379 kprintf("%x:%x ", va, *(int *) ptep);
3388 pmap_pvdump(vm_paddr_t pa)
3393 kprintf("pa %08llx", (long long)pa);
3394 m = PHYS_TO_VM_PAGE(pa);
3395 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3397 kprintf(" -> pmap %p, va %x, flags %x",
3398 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3400 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);