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 1.86 2008/06/05 18:06:32 swildner Exp $
47 * Manages physical address maps.
49 * In addition to hardware address maps, this
50 * module is called upon to provide software-use-only
51 * maps which may or may not be stored in the same
52 * form as hardware maps. These pseudo-maps are
53 * used to store intermediate results from copy
54 * operations to and from address spaces.
56 * Since the information managed by this module is
57 * also stored by the logical address mapping module,
58 * this module may throw away valid virtual-to-physical
59 * mappings at almost any time. However, invalidations
60 * of virtual-to-physical mappings must be done as
63 * In order to cope with hardware architectures which
64 * make virtual-to-physical map invalidates expensive,
65 * this module may delay invalidate or reduced protection
66 * operations until such time as they are actually
67 * necessary. This module is given full information as
68 * to which processors are currently using which maps,
69 * and to when physical maps must be made correct.
72 #include "opt_disable_pse.h"
74 #include "opt_msgbuf.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
80 #include <sys/msgbuf.h>
81 #include <sys/vmmeter.h>
85 #include <vm/vm_param.h>
86 #include <sys/sysctl.h>
88 #include <vm/vm_kern.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_map.h>
91 #include <vm/vm_object.h>
92 #include <vm/vm_extern.h>
93 #include <vm/vm_pageout.h>
94 #include <vm/vm_pager.h>
95 #include <vm/vm_zone.h>
98 #include <sys/thread2.h>
99 #include <sys/sysref2.h>
101 #include <machine/cputypes.h>
102 #include <machine/md_var.h>
103 #include <machine/specialreg.h>
104 #include <machine/smp.h>
105 #include <machine_base/apic/apicreg.h>
106 #include <machine/globaldata.h>
107 #include <machine/pmap.h>
108 #include <machine/pmap_inval.h>
112 #define PMAP_KEEP_PDIRS
113 #ifndef PMAP_SHPGPERPROC
114 #define PMAP_SHPGPERPROC 200
117 #if defined(DIAGNOSTIC)
118 #define PMAP_DIAGNOSTIC
123 #if !defined(PMAP_DIAGNOSTIC)
124 #define PMAP_INLINE __inline
130 * Get PDEs and PTEs for user/kernel address space
132 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
133 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
135 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
136 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
137 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
138 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
139 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
143 * Given a map and a machine independent protection code,
144 * convert to a vax protection code.
146 #define pte_prot(m, p) \
147 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
148 static int protection_codes[8];
150 struct pmap kernel_pmap;
151 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
153 vm_paddr_t avail_start; /* PA of first available physical page */
154 vm_paddr_t avail_end; /* PA of last available physical page */
155 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
156 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
157 vm_offset_t KvaStart; /* VA start of KVA space */
158 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
159 vm_offset_t KvaSize; /* max size of kernel virtual address space */
160 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
161 static int pgeflag; /* PG_G or-in */
162 static int pseflag; /* PG_PS or-in */
164 static vm_object_t kptobj;
167 vm_offset_t kernel_vm_end;
170 * Data for the pv entry allocation mechanism
172 static vm_zone_t pvzone;
173 static struct vm_zone pvzone_store;
174 static struct vm_object pvzone_obj;
175 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
176 static int pmap_pagedaemon_waken = 0;
177 static struct pv_entry *pvinit;
180 * All those kernel PT submaps that BSD is so fond of
182 pt_entry_t *CMAP1 = 0, *ptmmap;
183 caddr_t CADDR1 = 0, ptvmmap = 0;
184 static pt_entry_t *msgbufmap;
185 struct msgbuf *msgbufp=0;
190 static pt_entry_t *pt_crashdumpmap;
191 static caddr_t crashdumpmap;
193 extern pt_entry_t *SMPpt;
195 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
196 static unsigned * get_ptbase (pmap_t pmap);
197 static pv_entry_t get_pv_entry (void);
198 static void i386_protection_init (void);
199 static __inline void pmap_clearbit (vm_page_t m, int bit);
201 static void pmap_remove_all (vm_page_t m);
202 static void pmap_enter_quick (pmap_t pmap, vm_offset_t va, vm_page_t m);
203 static int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
204 vm_offset_t sva, pmap_inval_info_t info);
205 static void pmap_remove_page (struct pmap *pmap,
206 vm_offset_t va, pmap_inval_info_t info);
207 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
208 vm_offset_t va, pmap_inval_info_t info);
209 static boolean_t pmap_testbit (vm_page_t m, int bit);
210 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
211 vm_page_t mpte, vm_page_t m);
213 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
215 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
216 static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
217 static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
218 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
219 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
220 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
222 static unsigned pdir4mb;
225 * Move the kernel virtual free pointer to the next
226 * 4MB. This is used to help improve performance
227 * by using a large (4MB) page for much of the kernel
228 * (.text, .data, .bss)
231 pmap_kmem_choose(vm_offset_t addr)
233 vm_offset_t newaddr = addr;
235 if (cpu_feature & CPUID_PSE) {
236 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
245 * Extract the page table entry associated with the given map/virtual
248 * This function may NOT be called from an interrupt.
250 PMAP_INLINE unsigned *
251 pmap_pte(pmap_t pmap, vm_offset_t va)
256 pdeaddr = (unsigned *) pmap_pde(pmap, va);
257 if (*pdeaddr & PG_PS)
260 return get_ptbase(pmap) + i386_btop(va);
269 * Super fast pmap_pte routine best used when scanning the pv lists.
270 * This eliminates many course-grained invltlb calls. Note that many of
271 * the pv list scans are across different pmaps and it is very wasteful
272 * to do an entire invltlb when checking a single mapping.
274 * Should only be called while in a critical section.
277 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
279 struct mdglobaldata *gd = mdcpu;
282 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
283 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
284 unsigned index = i386_btop(va);
285 /* are we current address space or kernel? */
286 if ((pmap == &kernel_pmap) ||
287 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
288 return (unsigned *) PTmap + index;
290 newpf = pde & PG_FRAME;
291 if ( ((* (unsigned *) gd->gd_PMAP1) & PG_FRAME) != newpf) {
292 * (unsigned *) gd->gd_PMAP1 = newpf | PG_RW | PG_V;
293 cpu_invlpg(gd->gd_PADDR1);
295 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
302 * Bootstrap the system enough to run with virtual memory.
304 * On the i386 this is called after mapping has already been enabled
305 * and just syncs the pmap module with what has already been done.
306 * [We can't call it easily with mapping off since the kernel is not
307 * mapped with PA == VA, hence we would have to relocate every address
308 * from the linked base (virtual) address "KERNBASE" to the actual
309 * (physical) address starting relative to 0]
312 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
316 struct mdglobaldata *gd;
320 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
321 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
322 KvaEnd = KvaStart + KvaSize;
324 avail_start = firstaddr;
327 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
328 * too large. It should instead be correctly calculated in locore.s and
329 * not based on 'first' (which is a physical address, not a virtual
330 * address, for the start of unused physical memory). The kernel
331 * page tables are NOT double mapped and thus should not be included
332 * in this calculation.
334 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
335 virtual_start = pmap_kmem_choose(virtual_start);
336 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
339 * Initialize protection array.
341 i386_protection_init();
344 * The kernel's pmap is statically allocated so we don't have to use
345 * pmap_create, which is unlikely to work correctly at this part of
346 * the boot sequence (XXX and which no longer exists).
348 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
349 kernel_pmap.pm_count = 1;
350 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
351 TAILQ_INIT(&kernel_pmap.pm_pvlist);
355 * Reserve some special page table entries/VA space for temporary
358 #define SYSMAP(c, p, v, n) \
359 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
362 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
365 * CMAP1/CMAP2 are used for zeroing and copying pages.
367 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
372 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
375 * ptvmmap is used for reading arbitrary physical pages via
378 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
381 * msgbufp is used to map the system message buffer.
382 * XXX msgbufmap is not used.
384 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
385 atop(round_page(MSGBUF_SIZE)))
390 for (i = 0; i < NKPT; i++)
394 * PG_G is terribly broken on SMP because we IPI invltlb's in some
395 * cases rather then invl1pg. Actually, I don't even know why it
396 * works under UP because self-referential page table mappings
401 if (cpu_feature & CPUID_PGE)
406 * Initialize the 4MB page size flag
410 * The 4MB page version of the initial
411 * kernel page mapping.
415 #if !defined(DISABLE_PSE)
416 if (cpu_feature & CPUID_PSE) {
419 * Note that we have enabled PSE mode
422 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
423 ptditmp &= ~(NBPDR - 1);
424 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
429 * Enable the PSE mode. If we are SMP we can't do this
430 * now because the APs will not be able to use it when
433 load_cr4(rcr4() | CR4_PSE);
436 * We can do the mapping here for the single processor
437 * case. We simply ignore the old page table page from
441 * For SMP, we still need 4K pages to bootstrap APs,
442 * PSE will be enabled as soon as all APs are up.
444 PTD[KPTDI] = (pd_entry_t)ptditmp;
445 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
451 if (cpu_apic_address == 0)
452 panic("pmap_bootstrap: no local apic!");
454 /* local apic is mapped on last page */
455 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag |
456 (cpu_apic_address & PG_FRAME));
460 * We need to finish setting up the globaldata page for the BSP.
461 * locore has already populated the page table for the mdglobaldata
464 pg = MDGLOBALDATA_BASEALLOC_PAGES;
465 gd = &CPU_prvspace[0].mdglobaldata;
466 gd->gd_CMAP1 = &SMPpt[pg + 0];
467 gd->gd_CMAP2 = &SMPpt[pg + 1];
468 gd->gd_CMAP3 = &SMPpt[pg + 2];
469 gd->gd_PMAP1 = &SMPpt[pg + 3];
470 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
471 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
472 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
473 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
480 * Set 4mb pdir for mp startup
485 if (pseflag && (cpu_feature & CPUID_PSE)) {
486 load_cr4(rcr4() | CR4_PSE);
487 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
488 kernel_pmap.pm_pdir[KPTDI] =
489 PTD[KPTDI] = (pd_entry_t)pdir4mb;
497 * Initialize the pmap module.
498 * Called by vm_init, to initialize any structures that the pmap
499 * system needs to map virtual memory.
500 * pmap_init has been enhanced to support in a fairly consistant
501 * way, discontiguous physical memory.
510 * object for kernel page table pages
512 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
515 * Allocate memory for random pmap data structures. Includes the
519 for(i = 0; i < vm_page_array_size; i++) {
522 m = &vm_page_array[i];
523 TAILQ_INIT(&m->md.pv_list);
524 m->md.pv_list_count = 0;
528 * init the pv free list
530 initial_pvs = vm_page_array_size;
531 if (initial_pvs < MINPV)
533 pvzone = &pvzone_store;
534 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
535 initial_pvs * sizeof (struct pv_entry));
536 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
540 * Now it is safe to enable pv_table recording.
542 pmap_initialized = TRUE;
546 * Initialize the address space (zone) for the pv_entries. Set a
547 * high water mark so that the system can recover from excessive
548 * numbers of pv entries.
553 int shpgperproc = PMAP_SHPGPERPROC;
555 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
556 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
557 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
558 pv_entry_high_water = 9 * (pv_entry_max / 10);
559 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
563 /***************************************************
564 * Low level helper routines.....
565 ***************************************************/
567 #if defined(PMAP_DIAGNOSTIC)
570 * This code checks for non-writeable/modified pages.
571 * This should be an invalid condition.
574 pmap_nw_modified(pt_entry_t ptea)
580 if ((pte & (PG_M|PG_RW)) == PG_M)
589 * this routine defines the region(s) of memory that should
590 * not be tested for the modified bit.
592 static PMAP_INLINE int
593 pmap_track_modified(vm_offset_t va)
595 if ((va < clean_sva) || (va >= clean_eva))
602 get_ptbase(pmap_t pmap)
604 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
605 struct globaldata *gd = mycpu;
607 /* are we current address space or kernel? */
608 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
609 return (unsigned *) PTmap;
612 /* otherwise, we are alternate address space */
613 KKASSERT(gd->gd_intr_nesting_level == 0 &&
614 (gd->gd_curthread->td_flags & TDF_INTTHREAD) == 0);
616 if (frame != (((unsigned) APTDpde) & PG_FRAME)) {
617 APTDpde = (pd_entry_t)(frame | PG_RW | PG_V);
618 /* The page directory is not shared between CPUs */
621 return (unsigned *) APTmap;
627 * Extract the physical page address associated with the map/VA pair.
629 * This function may not be called from an interrupt if the pmap is
633 pmap_extract(pmap_t pmap, vm_offset_t va)
636 vm_offset_t pdirindex;
638 pdirindex = va >> PDRSHIFT;
639 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
641 if ((rtval & PG_PS) != 0) {
642 rtval &= ~(NBPDR - 1);
643 rtval |= va & (NBPDR - 1);
646 pte = get_ptbase(pmap) + i386_btop(va);
647 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
653 /***************************************************
654 * Low level mapping routines.....
655 ***************************************************/
658 * Routine: pmap_kenter
660 * Add a wired page to the KVA
661 * NOTE! note that in order for the mapping to take effect -- you
662 * should do an invltlb after doing the pmap_kenter().
665 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
669 pmap_inval_info info;
671 pmap_inval_init(&info);
672 npte = pa | PG_RW | PG_V | pgeflag;
673 pte = (unsigned *)vtopte(va);
674 pmap_inval_add(&info, &kernel_pmap, va);
676 pmap_inval_flush(&info);
680 * Routine: pmap_kenter_quick
682 * Similar to pmap_kenter(), except we only invalidate the
683 * mapping on the current CPU.
686 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
691 npte = pa | PG_RW | PG_V | pgeflag;
692 pte = (unsigned *)vtopte(va);
694 cpu_invlpg((void *)va);
698 pmap_kenter_sync(vm_offset_t va)
700 pmap_inval_info info;
702 pmap_inval_init(&info);
703 pmap_inval_add(&info, &kernel_pmap, va);
704 pmap_inval_flush(&info);
708 pmap_kenter_sync_quick(vm_offset_t va)
710 cpu_invlpg((void *)va);
714 * remove a page from the kernel pagetables
717 pmap_kremove(vm_offset_t va)
720 pmap_inval_info info;
722 pmap_inval_init(&info);
723 pte = (unsigned *)vtopte(va);
724 pmap_inval_add(&info, &kernel_pmap, va);
726 pmap_inval_flush(&info);
730 pmap_kremove_quick(vm_offset_t va)
733 pte = (unsigned *)vtopte(va);
735 cpu_invlpg((void *)va);
739 * XXX these need to be recoded. They are not used in any critical path.
742 pmap_kmodify_rw(vm_offset_t va)
744 *vtopte(va) |= PG_RW;
745 cpu_invlpg((void *)va);
749 pmap_kmodify_nc(vm_offset_t va)
752 cpu_invlpg((void *)va);
756 * Used to map a range of physical addresses into kernel
757 * virtual address space.
759 * For now, VM is already on, we only need to map the
763 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
765 while (start < end) {
766 pmap_kenter(virt, start);
775 * Add a list of wired pages to the kva
776 * this routine is only used for temporary
777 * kernel mappings that do not need to have
778 * page modification or references recorded.
779 * Note that old mappings are simply written
780 * over. The page *must* be wired.
783 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
787 end_va = va + count * PAGE_SIZE;
789 while (va < end_va) {
792 pte = (unsigned *)vtopte(va);
793 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
794 cpu_invlpg((void *)va);
799 smp_invltlb(); /* XXX */
804 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
807 cpumask_t cmask = mycpu->gd_cpumask;
809 end_va = va + count * PAGE_SIZE;
811 while (va < end_va) {
816 * Install the new PTE. If the pte changed from the prior
817 * mapping we must reset the cpu mask and invalidate the page.
818 * If the pte is the same but we have not seen it on the
819 * current cpu, invlpg the existing mapping. Otherwise the
820 * entry is optimal and no invalidation is required.
822 pte = (unsigned *)vtopte(va);
823 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
824 if (*pte != pteval) {
827 cpu_invlpg((void *)va);
828 } else if ((*mask & cmask) == 0) {
829 cpu_invlpg((void *)va);
838 * this routine jerks page mappings from the
839 * kernel -- it is meant only for temporary mappings.
842 pmap_qremove(vm_offset_t va, int count)
846 end_va = va + count*PAGE_SIZE;
848 while (va < end_va) {
851 pte = (unsigned *)vtopte(va);
853 cpu_invlpg((void *)va);
862 * This routine works like vm_page_lookup() but also blocks as long as the
863 * page is busy. This routine does not busy the page it returns.
865 * Unless the caller is managing objects whos pages are in a known state,
866 * the call should be made with a critical section held so the page's object
867 * association remains valid on return.
870 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
875 m = vm_page_lookup(object, pindex);
876 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
882 * Create a new thread and optionally associate it with a (new) process.
883 * NOTE! the new thread's cpu may not equal the current cpu.
886 pmap_init_thread(thread_t td)
888 /* enforce pcb placement */
889 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
890 td->td_savefpu = &td->td_pcb->pcb_save;
891 td->td_sp = (char *)td->td_pcb - 16;
895 * This routine directly affects the fork perf for a process.
898 pmap_init_proc(struct proc *p)
903 * Dispose the UPAGES for a process that has exited.
904 * This routine directly impacts the exit perf of a process.
907 pmap_dispose_proc(struct proc *p)
909 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
912 /***************************************************
913 * Page table page management routines.....
914 ***************************************************/
917 * This routine unholds page table pages, and if the hold count
918 * drops to zero, then it decrements the wire count.
921 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
924 * Wait until we can busy the page ourselves. We cannot have
925 * any active flushes if we block.
927 if (m->flags & PG_BUSY) {
928 pmap_inval_flush(info);
929 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
932 KASSERT(m->queue == PQ_NONE,
933 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
935 if (m->hold_count == 1) {
937 * Unmap the page table page
940 pmap_inval_add(info, pmap, -1);
941 pmap->pm_pdir[m->pindex] = 0;
943 KKASSERT(pmap->pm_stats.resident_count > 0);
944 --pmap->pm_stats.resident_count;
946 if (pmap->pm_ptphint == m)
947 pmap->pm_ptphint = NULL;
950 * This was our last hold, the page had better be unwired
951 * after we decrement wire_count.
953 * FUTURE NOTE: shared page directory page could result in
954 * multiple wire counts.
958 KKASSERT(m->wire_count == 0);
959 --vmstats.v_wire_count;
960 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
962 vm_page_free_zero(m);
965 KKASSERT(m->hold_count > 1);
971 static PMAP_INLINE int
972 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
974 KKASSERT(m->hold_count > 0);
975 if (m->hold_count > 1) {
979 return _pmap_unwire_pte_hold(pmap, m, info);
984 * After removing a page table entry, this routine is used to
985 * conditionally free the page, and manage the hold/wire counts.
988 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
989 pmap_inval_info_t info)
992 if (va >= UPT_MIN_ADDRESS)
996 ptepindex = (va >> PDRSHIFT);
997 if (pmap->pm_ptphint &&
998 (pmap->pm_ptphint->pindex == ptepindex)) {
999 mpte = pmap->pm_ptphint;
1001 pmap_inval_flush(info);
1002 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1003 pmap->pm_ptphint = mpte;
1007 return pmap_unwire_pte_hold(pmap, mpte, info);
1011 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1012 * it, and IdlePTD, represents the template used to update all other pmaps.
1014 * On architectures where the kernel pmap is not integrated into the user
1015 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1016 * kernel_pmap should be used to directly access the kernel_pmap.
1019 pmap_pinit0(struct pmap *pmap)
1022 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1023 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1025 pmap->pm_active = 0;
1026 pmap->pm_ptphint = NULL;
1027 TAILQ_INIT(&pmap->pm_pvlist);
1028 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1032 * Initialize a preallocated and zeroed pmap structure,
1033 * such as one in a vmspace structure.
1036 pmap_pinit(struct pmap *pmap)
1041 * No need to allocate page table space yet but we do need a valid
1042 * page directory table.
1044 if (pmap->pm_pdir == NULL) {
1046 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1050 * Allocate an object for the ptes
1052 if (pmap->pm_pteobj == NULL)
1053 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1056 * Allocate the page directory page, unless we already have
1057 * one cached. If we used the cached page the wire_count will
1058 * already be set appropriately.
1060 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1061 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1062 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1063 pmap->pm_pdirm = ptdpg;
1064 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1065 ptdpg->valid = VM_PAGE_BITS_ALL;
1066 ptdpg->wire_count = 1;
1067 ++vmstats.v_wire_count;
1068 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1070 if ((ptdpg->flags & PG_ZERO) == 0)
1071 bzero(pmap->pm_pdir, PAGE_SIZE);
1073 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1075 /* install self-referential address mapping entry */
1076 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1077 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1080 pmap->pm_active = 0;
1081 pmap->pm_ptphint = NULL;
1082 TAILQ_INIT(&pmap->pm_pvlist);
1083 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1084 pmap->pm_stats.resident_count = 1;
1088 * Clean up a pmap structure so it can be physically freed. This routine
1089 * is called by the vmspace dtor function. A great deal of pmap data is
1090 * left passively mapped to improve vmspace management so we have a bit
1091 * of cleanup work to do here.
1094 pmap_puninit(pmap_t pmap)
1098 KKASSERT(pmap->pm_active == 0);
1099 if ((p = pmap->pm_pdirm) != NULL) {
1100 KKASSERT(pmap->pm_pdir != NULL);
1101 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1103 vmstats.v_wire_count--;
1104 KKASSERT((p->flags & PG_BUSY) == 0);
1106 vm_page_free_zero(p);
1107 pmap->pm_pdirm = NULL;
1109 if (pmap->pm_pdir) {
1110 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1111 pmap->pm_pdir = NULL;
1113 if (pmap->pm_pteobj) {
1114 vm_object_deallocate(pmap->pm_pteobj);
1115 pmap->pm_pteobj = NULL;
1120 * Wire in kernel global address entries. To avoid a race condition
1121 * between pmap initialization and pmap_growkernel, this procedure
1122 * adds the pmap to the master list (which growkernel scans to update),
1123 * then copies the template.
1126 pmap_pinit2(struct pmap *pmap)
1129 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1130 /* XXX copies current process, does not fill in MPPTDI */
1131 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1136 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1137 * 0 on failure (if the procedure had to sleep).
1139 * When asked to remove the page directory page itself, we actually just
1140 * leave it cached so we do not have to incur the SMP inval overhead of
1141 * removing the kernel mapping. pmap_puninit() will take care of it.
1144 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1146 unsigned *pde = (unsigned *) pmap->pm_pdir;
1148 * This code optimizes the case of freeing non-busy
1149 * page-table pages. Those pages are zero now, and
1150 * might as well be placed directly into the zero queue.
1152 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1158 * Remove the page table page from the processes address space.
1161 KKASSERT(pmap->pm_stats.resident_count > 0);
1162 --pmap->pm_stats.resident_count;
1164 if (p->hold_count) {
1165 panic("pmap_release: freeing held page table page");
1167 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1168 pmap->pm_ptphint = NULL;
1171 * We leave the page directory page cached, wired, and mapped in
1172 * the pmap until the dtor function (pmap_puninit()) gets called.
1173 * However, still clean it up so we can set PG_ZERO.
1175 if (p->pindex == PTDPTDI) {
1176 bzero(pde + KPTDI, nkpt * PTESIZE);
1179 vm_page_flag_set(p, PG_ZERO);
1183 vmstats.v_wire_count--;
1184 vm_page_free_zero(p);
1190 * this routine is called if the page table page is not
1194 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1196 vm_offset_t pteva, ptepa;
1200 * Find or fabricate a new pagetable page
1202 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1203 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1205 KASSERT(m->queue == PQ_NONE,
1206 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1209 * Increment the hold count for the page we will be returning to
1215 * It is possible that someone else got in and mapped by the page
1216 * directory page while we were blocked, if so just unbusy and
1217 * return the held page.
1219 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1220 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1225 if (m->wire_count == 0)
1226 vmstats.v_wire_count++;
1231 * Map the pagetable page into the process address space, if
1232 * it isn't already there.
1235 ++pmap->pm_stats.resident_count;
1237 ptepa = VM_PAGE_TO_PHYS(m);
1238 pmap->pm_pdir[ptepindex] =
1239 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1242 * Set the page table hint
1244 pmap->pm_ptphint = m;
1247 * Try to use the new mapping, but if we cannot, then
1248 * do it with the routine that maps the page explicitly.
1250 if ((m->flags & PG_ZERO) == 0) {
1251 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1252 (((unsigned) PTDpde) & PG_FRAME)) {
1253 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1254 bzero((caddr_t) pteva, PAGE_SIZE);
1256 pmap_zero_page(ptepa);
1260 m->valid = VM_PAGE_BITS_ALL;
1261 vm_page_flag_clear(m, PG_ZERO);
1262 vm_page_flag_set(m, PG_MAPPED);
1269 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1276 * Calculate pagetable page index
1278 ptepindex = va >> PDRSHIFT;
1281 * Get the page directory entry
1283 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1286 * This supports switching from a 4MB page to a
1289 if (ptepa & PG_PS) {
1290 pmap->pm_pdir[ptepindex] = 0;
1297 * If the page table page is mapped, we just increment the
1298 * hold count, and activate it.
1302 * In order to get the page table page, try the
1305 if (pmap->pm_ptphint &&
1306 (pmap->pm_ptphint->pindex == ptepindex)) {
1307 m = pmap->pm_ptphint;
1309 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1310 pmap->pm_ptphint = m;
1316 * Here if the pte page isn't mapped, or if it has been deallocated.
1318 return _pmap_allocpte(pmap, ptepindex);
1322 /***************************************************
1323 * Pmap allocation/deallocation routines.
1324 ***************************************************/
1327 * Release any resources held by the given physical map.
1328 * Called when a pmap initialized by pmap_pinit is being released.
1329 * Should only be called if the map contains no valid mappings.
1331 static int pmap_release_callback(struct vm_page *p, void *data);
1334 pmap_release(struct pmap *pmap)
1336 vm_object_t object = pmap->pm_pteobj;
1337 struct rb_vm_page_scan_info info;
1339 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1340 #if defined(DIAGNOSTIC)
1341 if (object->ref_count != 1)
1342 panic("pmap_release: pteobj reference count != 1");
1346 info.object = object;
1348 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1355 info.limit = object->generation;
1357 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1358 pmap_release_callback, &info);
1359 if (info.error == 0 && info.mpte) {
1360 if (!pmap_release_free_page(pmap, info.mpte))
1364 } while (info.error);
1368 pmap_release_callback(struct vm_page *p, void *data)
1370 struct rb_vm_page_scan_info *info = data;
1372 if (p->pindex == PTDPTDI) {
1376 if (!pmap_release_free_page(info->pmap, p)) {
1380 if (info->object->generation != info->limit) {
1388 * Grow the number of kernel page table entries, if needed.
1392 pmap_growkernel(vm_offset_t addr)
1395 vm_offset_t ptppaddr;
1400 if (kernel_vm_end == 0) {
1401 kernel_vm_end = KERNBASE;
1403 while (pdir_pde(PTD, kernel_vm_end)) {
1404 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1408 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1409 while (kernel_vm_end < addr) {
1410 if (pdir_pde(PTD, kernel_vm_end)) {
1411 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1416 * This index is bogus, but out of the way
1418 nkpg = vm_page_alloc(kptobj, nkpt,
1419 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1421 panic("pmap_growkernel: no memory to grow kernel");
1424 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1425 pmap_zero_page(ptppaddr);
1426 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1427 pdir_pde(PTD, kernel_vm_end) = newpdir;
1428 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1432 * This update must be interlocked with pmap_pinit2.
1434 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1435 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1437 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1438 ~(PAGE_SIZE * NPTEPG - 1);
1444 * Retire the given physical map from service.
1445 * Should only be called if the map contains
1446 * no valid mappings.
1449 pmap_destroy(pmap_t pmap)
1456 count = --pmap->pm_count;
1459 panic("destroying a pmap is not yet implemented");
1464 * Add a reference to the specified pmap.
1467 pmap_reference(pmap_t pmap)
1474 /***************************************************
1475 * page management routines.
1476 ***************************************************/
1479 * free the pv_entry back to the free list. This function may be
1480 * called from an interrupt.
1482 static PMAP_INLINE void
1483 free_pv_entry(pv_entry_t pv)
1490 * get a new pv_entry, allocating a block from the system
1491 * when needed. This function may be called from an interrupt.
1497 if (pv_entry_high_water &&
1498 (pv_entry_count > pv_entry_high_water) &&
1499 (pmap_pagedaemon_waken == 0)) {
1500 pmap_pagedaemon_waken = 1;
1501 wakeup (&vm_pages_needed);
1503 return zalloc(pvzone);
1507 * This routine is very drastic, but can save the system
1515 static int warningdone=0;
1517 if (pmap_pagedaemon_waken == 0)
1520 if (warningdone < 5) {
1521 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1525 for(i = 0; i < vm_page_array_size; i++) {
1526 m = &vm_page_array[i];
1527 if (m->wire_count || m->hold_count || m->busy ||
1528 (m->flags & PG_BUSY))
1532 pmap_pagedaemon_waken = 0;
1537 * If it is the first entry on the list, it is actually
1538 * in the header and we must copy the following entry up
1539 * to the header. Otherwise we must search the list for
1540 * the entry. In either case we free the now unused entry.
1543 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1544 vm_offset_t va, pmap_inval_info_t info)
1550 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1551 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1552 if (pmap == pv->pv_pmap && va == pv->pv_va)
1556 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1557 if (va == pv->pv_va)
1564 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1565 m->md.pv_list_count--;
1566 if (TAILQ_EMPTY(&m->md.pv_list))
1567 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1568 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1569 ++pmap->pm_generation;
1570 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1578 * Create a pv entry for page at pa for
1582 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1587 pv = get_pv_entry();
1592 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1593 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1594 m->md.pv_list_count++;
1600 * pmap_remove_pte: do the things to unmap a page in a process
1603 pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
1604 pmap_inval_info_t info)
1609 pmap_inval_add(info, pmap, va);
1610 oldpte = loadandclear(ptq);
1612 pmap->pm_stats.wired_count -= 1;
1614 * Machines that don't support invlpg, also don't support
1615 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1619 cpu_invlpg((void *)va);
1620 KKASSERT(pmap->pm_stats.resident_count > 0);
1621 --pmap->pm_stats.resident_count;
1622 if (oldpte & PG_MANAGED) {
1623 m = PHYS_TO_VM_PAGE(oldpte);
1624 if (oldpte & PG_M) {
1625 #if defined(PMAP_DIAGNOSTIC)
1626 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1628 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1632 if (pmap_track_modified(va))
1636 vm_page_flag_set(m, PG_REFERENCED);
1637 return pmap_remove_entry(pmap, m, va, info);
1639 return pmap_unuse_pt(pmap, va, NULL, info);
1648 * Remove a single page from a process address space.
1650 * This function may not be called from an interrupt if the pmap is
1654 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
1659 * if there is no pte for this address, just skip it!!! Otherwise
1660 * get a local va for mappings for this pmap and remove the entry.
1662 if (*pmap_pde(pmap, va) != 0) {
1663 ptq = get_ptbase(pmap) + i386_btop(va);
1665 pmap_remove_pte(pmap, ptq, va, info);
1673 * Remove the given range of addresses from the specified map.
1675 * It is assumed that the start and end are properly
1676 * rounded to the page size.
1678 * This function may not be called from an interrupt if the pmap is
1682 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1686 vm_offset_t ptpaddr;
1687 vm_offset_t sindex, eindex;
1688 struct pmap_inval_info info;
1693 if (pmap->pm_stats.resident_count == 0)
1696 pmap_inval_init(&info);
1699 * special handling of removing one page. a very
1700 * common operation and easy to short circuit some
1703 if (((sva + PAGE_SIZE) == eva) &&
1704 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1705 pmap_remove_page(pmap, sva, &info);
1706 pmap_inval_flush(&info);
1711 * Get a local virtual address for the mappings that are being
1714 sindex = i386_btop(sva);
1715 eindex = i386_btop(eva);
1717 for (; sindex < eindex; sindex = pdnxt) {
1721 * Calculate index for next page table.
1723 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1724 if (pmap->pm_stats.resident_count == 0)
1727 pdirindex = sindex / NPDEPG;
1728 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1729 pmap_inval_add(&info, pmap, -1);
1730 pmap->pm_pdir[pdirindex] = 0;
1731 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1736 * Weed out invalid mappings. Note: we assume that the page
1737 * directory table is always allocated, and in kernel virtual.
1743 * Limit our scan to either the end of the va represented
1744 * by the current page table page, or to the end of the
1745 * range being removed.
1747 if (pdnxt > eindex) {
1752 * NOTE: pmap_remove_pte() can block.
1754 for (; sindex != pdnxt; sindex++) {
1757 ptbase = get_ptbase(pmap);
1758 if (ptbase[sindex] == 0)
1760 va = i386_ptob(sindex);
1761 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
1765 pmap_inval_flush(&info);
1771 * Removes this physical page from all physical maps in which it resides.
1772 * Reflects back modify bits to the pager.
1774 * This routine may not be called from an interrupt.
1778 pmap_remove_all(vm_page_t m)
1780 struct pmap_inval_info info;
1781 unsigned *pte, tpte;
1784 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
1787 pmap_inval_init(&info);
1789 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1790 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1791 --pv->pv_pmap->pm_stats.resident_count;
1793 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1794 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
1795 tpte = loadandclear(pte);
1798 pv->pv_pmap->pm_stats.wired_count--;
1801 vm_page_flag_set(m, PG_REFERENCED);
1804 * Update the vm_page_t clean and reference bits.
1807 #if defined(PMAP_DIAGNOSTIC)
1808 if (pmap_nw_modified((pt_entry_t) tpte)) {
1810 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1814 if (pmap_track_modified(pv->pv_va))
1817 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1818 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1819 ++pv->pv_pmap->pm_generation;
1820 m->md.pv_list_count--;
1821 if (TAILQ_EMPTY(&m->md.pv_list))
1822 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1823 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
1827 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
1828 pmap_inval_flush(&info);
1834 * Set the physical protection on the specified range of this map
1837 * This function may not be called from an interrupt if the map is
1838 * not the kernel_pmap.
1841 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1844 vm_offset_t pdnxt, ptpaddr;
1845 vm_pindex_t sindex, eindex;
1846 pmap_inval_info info;
1851 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1852 pmap_remove(pmap, sva, eva);
1856 if (prot & VM_PROT_WRITE)
1859 pmap_inval_init(&info);
1861 ptbase = get_ptbase(pmap);
1863 sindex = i386_btop(sva);
1864 eindex = i386_btop(eva);
1866 for (; sindex < eindex; sindex = pdnxt) {
1870 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1872 pdirindex = sindex / NPDEPG;
1873 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1874 pmap_inval_add(&info, pmap, -1);
1875 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1876 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1881 * Weed out invalid mappings. Note: we assume that the page
1882 * directory table is always allocated, and in kernel virtual.
1887 if (pdnxt > eindex) {
1891 for (; sindex != pdnxt; sindex++) {
1897 * XXX non-optimal. Note also that there can be
1898 * no pmap_inval_flush() calls until after we modify
1899 * ptbase[sindex] (or otherwise we have to do another
1900 * pmap_inval_add() call).
1902 pmap_inval_add(&info, pmap, i386_ptob(sindex));
1903 pbits = ptbase[sindex];
1905 if (pbits & PG_MANAGED) {
1908 m = PHYS_TO_VM_PAGE(pbits);
1909 vm_page_flag_set(m, PG_REFERENCED);
1913 if (pmap_track_modified(i386_ptob(sindex))) {
1915 m = PHYS_TO_VM_PAGE(pbits);
1924 if (pbits != ptbase[sindex]) {
1925 ptbase[sindex] = pbits;
1929 pmap_inval_flush(&info);
1933 * Insert the given physical page (p) at
1934 * the specified virtual address (v) in the
1935 * target physical map with the protection requested.
1937 * If specified, the page will be wired down, meaning
1938 * that the related pte can not be reclaimed.
1940 * NB: This is the only routine which MAY NOT lazy-evaluate
1941 * or lose information. That is, this routine must actually
1942 * insert this page into the given map NOW.
1945 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1951 vm_offset_t origpte, newpte;
1953 pmap_inval_info info;
1959 #ifdef PMAP_DIAGNOSTIC
1961 panic("pmap_enter: toobig");
1962 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1963 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1965 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
1966 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
1968 db_print_backtrace();
1971 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
1972 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
1974 db_print_backtrace();
1979 * In the case that a page table page is not
1980 * resident, we are creating it here.
1982 if (va < UPT_MIN_ADDRESS)
1983 mpte = pmap_allocpte(pmap, va);
1987 pmap_inval_init(&info);
1988 pte = pmap_pte(pmap, va);
1991 * Page Directory table entry not valid, we need a new PT page
1994 panic("pmap_enter: invalid page directory pdir=%x, va=0x%x\n",
1995 (unsigned) pmap->pm_pdir[PTDPTDI], va);
1998 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
1999 origpte = *(vm_offset_t *)pte;
2000 opa = origpte & PG_FRAME;
2002 if (origpte & PG_PS)
2003 panic("pmap_enter: attempted pmap_enter on 4MB page");
2006 * Mapping has not changed, must be protection or wiring change.
2008 if (origpte && (opa == pa)) {
2010 * Wiring change, just update stats. We don't worry about
2011 * wiring PT pages as they remain resident as long as there
2012 * are valid mappings in them. Hence, if a user page is wired,
2013 * the PT page will be also.
2015 if (wired && ((origpte & PG_W) == 0))
2016 pmap->pm_stats.wired_count++;
2017 else if (!wired && (origpte & PG_W))
2018 pmap->pm_stats.wired_count--;
2020 #if defined(PMAP_DIAGNOSTIC)
2021 if (pmap_nw_modified((pt_entry_t) origpte)) {
2023 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
2029 * Remove the extra pte reference. Note that we cannot
2030 * optimize the RO->RW case because we have adjusted the
2031 * wiring count above and may need to adjust the wiring
2038 * We might be turning off write access to the page,
2039 * so we go ahead and sense modify status.
2041 if (origpte & PG_MANAGED) {
2042 if ((origpte & PG_M) && pmap_track_modified(va)) {
2044 om = PHYS_TO_VM_PAGE(opa);
2048 KKASSERT(m->flags & PG_MAPPED);
2053 * Mapping has changed, invalidate old range and fall through to
2054 * handle validating new mapping.
2058 err = pmap_remove_pte(pmap, pte, va, &info);
2060 panic("pmap_enter: pte vanished, va: 0x%x", va);
2064 * Enter on the PV list if part of our managed memory. Note that we
2065 * raise IPL while manipulating pv_table since pmap_enter can be
2066 * called at interrupt time.
2068 if (pmap_initialized &&
2069 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2070 pmap_insert_entry(pmap, va, mpte, m);
2072 vm_page_flag_set(m, PG_MAPPED);
2076 * Increment counters
2078 ++pmap->pm_stats.resident_count;
2080 pmap->pm_stats.wired_count++;
2084 * Now validate mapping with desired protection/wiring.
2086 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2090 if (va < UPT_MIN_ADDRESS)
2092 if (pmap == &kernel_pmap)
2096 * if the mapping or permission bits are different, we need
2097 * to update the pte.
2099 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2100 pmap_inval_add(&info, pmap, va);
2101 *pte = newpte | PG_A;
2103 vm_page_flag_set(m, PG_WRITEABLE);
2105 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2106 pmap_inval_flush(&info);
2110 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2111 * This code also assumes that the pmap has no pre-existing entry for this
2114 * This code currently may only be used on user pmaps, not kernel_pmap.
2117 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2124 pmap_inval_info info;
2126 pmap_inval_init(&info);
2128 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2129 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2131 db_print_backtrace();
2134 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2135 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2137 db_print_backtrace();
2141 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2144 * Calculate the page table page (mpte), allocating it if necessary.
2146 * A held page table page (mpte), or NULL, is passed onto the
2147 * section following.
2149 if (va < UPT_MIN_ADDRESS) {
2151 * Calculate pagetable page index
2153 ptepindex = va >> PDRSHIFT;
2157 * Get the page directory entry
2159 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2162 * If the page table page is mapped, we just increment
2163 * the hold count, and activate it.
2167 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2168 if (pmap->pm_ptphint &&
2169 (pmap->pm_ptphint->pindex == ptepindex)) {
2170 mpte = pmap->pm_ptphint;
2172 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2173 pmap->pm_ptphint = mpte;
2178 mpte = _pmap_allocpte(pmap, ptepindex);
2180 } while (mpte == NULL);
2183 /* this code path is not yet used */
2187 * With a valid (and held) page directory page, we can just use
2188 * vtopte() to get to the pte. If the pte is already present
2189 * we do not disturb it.
2191 pte = (unsigned *)vtopte(va);
2194 pmap_unwire_pte_hold(pmap, mpte, &info);
2195 pa = VM_PAGE_TO_PHYS(m);
2196 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2201 * Enter on the PV list if part of our managed memory
2203 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2204 pmap_insert_entry(pmap, va, mpte, m);
2205 vm_page_flag_set(m, PG_MAPPED);
2209 * Increment counters
2211 ++pmap->pm_stats.resident_count;
2213 pa = VM_PAGE_TO_PHYS(m);
2216 * Now validate mapping with RO protection
2218 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2219 *pte = pa | PG_V | PG_U;
2221 *pte = pa | PG_V | PG_U | PG_MANAGED;
2222 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2223 pmap_inval_flush(&info);
2227 * Make a temporary mapping for a physical address. This is only intended
2228 * to be used for panic dumps.
2231 pmap_kenter_temporary(vm_paddr_t pa, int i)
2233 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2234 return ((void *)crashdumpmap);
2237 #define MAX_INIT_PT (96)
2240 * This routine preloads the ptes for a given object into the specified pmap.
2241 * This eliminates the blast of soft faults on process startup and
2242 * immediately after an mmap.
2244 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2247 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2248 vm_object_t object, vm_pindex_t pindex,
2249 vm_size_t size, int limit)
2251 struct rb_vm_page_scan_info info;
2256 * We can't preinit if read access isn't set or there is no pmap
2259 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2263 * We can't preinit if the pmap is not the current pmap
2265 lp = curthread->td_lwp;
2266 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2269 psize = i386_btop(size);
2271 if ((object->type != OBJT_VNODE) ||
2272 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2273 (object->resident_page_count > MAX_INIT_PT))) {
2277 if (psize + pindex > object->size) {
2278 if (object->size < pindex)
2280 psize = object->size - pindex;
2287 * Use a red-black scan to traverse the requested range and load
2288 * any valid pages found into the pmap.
2290 * We cannot safely scan the object's memq unless we are in a
2291 * critical section since interrupts can remove pages from objects.
2293 info.start_pindex = pindex;
2294 info.end_pindex = pindex + psize - 1;
2301 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2302 pmap_object_init_pt_callback, &info);
2308 pmap_object_init_pt_callback(vm_page_t p, void *data)
2310 struct rb_vm_page_scan_info *info = data;
2311 vm_pindex_t rel_index;
2313 * don't allow an madvise to blow away our really
2314 * free pages allocating pv entries.
2316 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2317 vmstats.v_free_count < vmstats.v_free_reserved) {
2320 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2321 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2322 if ((p->queue - p->pc) == PQ_CACHE)
2323 vm_page_deactivate(p);
2325 rel_index = p->pindex - info->start_pindex;
2326 pmap_enter_quick(info->pmap,
2327 info->addr + i386_ptob(rel_index), p);
2334 * pmap_prefault provides a quick way of clustering pagefaults into a
2335 * processes address space. It is a "cousin" of pmap_object_init_pt,
2336 * except it runs at page fault time instead of mmap time.
2340 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2342 static int pmap_prefault_pageorder[] = {
2343 -PAGE_SIZE, PAGE_SIZE,
2344 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2345 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2346 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2350 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2361 * We do not currently prefault mappings that use virtual page
2362 * tables. We do not prefault foreign pmaps.
2364 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2366 lp = curthread->td_lwp;
2367 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2370 object = entry->object.vm_object;
2372 starta = addra - PFBAK * PAGE_SIZE;
2373 if (starta < entry->start)
2374 starta = entry->start;
2375 else if (starta > addra)
2379 * critical section protection is required to maintain the
2380 * page/object association, interrupts can free pages and remove
2381 * them from their objects.
2384 for (i = 0; i < PAGEORDER_SIZE; i++) {
2385 vm_object_t lobject;
2388 addr = addra + pmap_prefault_pageorder[i];
2389 if (addr > addra + (PFFOR * PAGE_SIZE))
2392 if (addr < starta || addr >= entry->end)
2395 if ((*pmap_pde(pmap, addr)) == 0)
2398 pte = (unsigned *) vtopte(addr);
2402 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2405 for (m = vm_page_lookup(lobject, pindex);
2406 (!m && (lobject->type == OBJT_DEFAULT) &&
2407 (lobject->backing_object));
2408 lobject = lobject->backing_object
2410 if (lobject->backing_object_offset & PAGE_MASK)
2412 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2413 m = vm_page_lookup(lobject->backing_object, pindex);
2417 * give-up when a page is not in memory
2422 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2424 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2426 if ((m->queue - m->pc) == PQ_CACHE) {
2427 vm_page_deactivate(m);
2430 pmap_enter_quick(pmap, addr, m);
2438 * Routine: pmap_change_wiring
2439 * Function: Change the wiring attribute for a map/virtual-address
2441 * In/out conditions:
2442 * The mapping must already exist in the pmap.
2445 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2452 pte = pmap_pte(pmap, va);
2454 if (wired && !pmap_pte_w(pte))
2455 pmap->pm_stats.wired_count++;
2456 else if (!wired && pmap_pte_w(pte))
2457 pmap->pm_stats.wired_count--;
2460 * Wiring is not a hardware characteristic so there is no need to
2461 * invalidate TLB. However, in an SMP environment we must use
2462 * a locked bus cycle to update the pte (if we are not using
2463 * the pmap_inval_*() API that is)... it's ok to do this for simple
2468 atomic_set_int(pte, PG_W);
2470 atomic_clear_int(pte, PG_W);
2473 atomic_set_int_nonlocked(pte, PG_W);
2475 atomic_clear_int_nonlocked(pte, PG_W);
2482 * Copy the range specified by src_addr/len
2483 * from the source map to the range dst_addr/len
2484 * in the destination map.
2486 * This routine is only advisory and need not do anything.
2489 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2490 vm_size_t len, vm_offset_t src_addr)
2492 pmap_inval_info info;
2494 vm_offset_t end_addr = src_addr + len;
2496 unsigned src_frame, dst_frame;
2499 if (dst_addr != src_addr)
2502 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2503 * valid through blocking calls, and that's just not going to
2510 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2511 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2515 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2516 if (dst_frame != (((unsigned) APTDpde) & PG_FRAME)) {
2517 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2518 /* The page directory is not shared between CPUs */
2521 pmap_inval_init(&info);
2522 pmap_inval_add(&info, dst_pmap, -1);
2523 pmap_inval_add(&info, src_pmap, -1);
2526 * critical section protection is required to maintain the page/object
2527 * association, interrupts can free pages and remove them from
2531 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2532 unsigned *src_pte, *dst_pte;
2533 vm_page_t dstmpte, srcmpte;
2534 vm_offset_t srcptepaddr;
2537 if (addr >= UPT_MIN_ADDRESS)
2538 panic("pmap_copy: invalid to pmap_copy page tables\n");
2541 * Don't let optional prefaulting of pages make us go
2542 * way below the low water mark of free pages or way
2543 * above high water mark of used pv entries.
2545 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2546 pv_entry_count > pv_entry_high_water)
2549 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2550 ptepindex = addr >> PDRSHIFT;
2552 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2553 if (srcptepaddr == 0)
2556 if (srcptepaddr & PG_PS) {
2557 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2558 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2559 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2564 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2565 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2566 (srcmpte->flags & PG_BUSY)) {
2570 if (pdnxt > end_addr)
2573 src_pte = (unsigned *) vtopte(addr);
2574 dst_pte = (unsigned *) avtopte(addr);
2575 while (addr < pdnxt) {
2580 * we only virtual copy managed pages
2582 if ((ptetemp & PG_MANAGED) != 0) {
2584 * We have to check after allocpte for the
2585 * pte still being around... allocpte can
2588 * pmap_allocpte() can block. If we lose
2589 * our page directory mappings we stop.
2591 dstmpte = pmap_allocpte(dst_pmap, addr);
2593 if (src_frame != (((unsigned) PTDpde) & PG_FRAME) ||
2594 dst_frame != (((unsigned) APTDpde) & PG_FRAME)
2596 kprintf("WARNING: pmap_copy: detected and corrected race\n");
2597 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2599 } else if ((*dst_pte == 0) &&
2600 (ptetemp = *src_pte) != 0 &&
2601 (ptetemp & PG_MANAGED)) {
2603 * Clear the modified and
2604 * accessed (referenced) bits
2607 m = PHYS_TO_VM_PAGE(ptetemp);
2608 *dst_pte = ptetemp & ~(PG_M | PG_A);
2609 ++dst_pmap->pm_stats.resident_count;
2610 pmap_insert_entry(dst_pmap, addr,
2612 KKASSERT(m->flags & PG_MAPPED);
2614 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
2615 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2618 if (dstmpte->hold_count >= srcmpte->hold_count)
2628 pmap_inval_flush(&info);
2634 * Zero the specified PA by mapping the page into KVM and clearing its
2637 * This function may be called from an interrupt and no locking is
2641 pmap_zero_page(vm_paddr_t phys)
2643 struct mdglobaldata *gd = mdcpu;
2646 if (*(int *)gd->gd_CMAP3)
2647 panic("pmap_zero_page: CMAP3 busy");
2648 *(int *)gd->gd_CMAP3 =
2649 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2650 cpu_invlpg(gd->gd_CADDR3);
2652 #if defined(I686_CPU)
2653 if (cpu_class == CPUCLASS_686)
2654 i686_pagezero(gd->gd_CADDR3);
2657 bzero(gd->gd_CADDR3, PAGE_SIZE);
2658 *(int *) gd->gd_CMAP3 = 0;
2663 * pmap_page_assertzero:
2665 * Assert that a page is empty, panic if it isn't.
2668 pmap_page_assertzero(vm_paddr_t phys)
2670 struct mdglobaldata *gd = mdcpu;
2674 if (*(int *)gd->gd_CMAP3)
2675 panic("pmap_zero_page: CMAP3 busy");
2676 *(int *)gd->gd_CMAP3 =
2677 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2678 cpu_invlpg(gd->gd_CADDR3);
2679 for (i = 0; i < PAGE_SIZE; i += 4) {
2680 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2681 panic("pmap_page_assertzero() @ %p not zero!\n",
2682 (void *)gd->gd_CADDR3);
2685 *(int *) gd->gd_CMAP3 = 0;
2692 * Zero part of a physical page by mapping it into memory and clearing
2693 * its contents with bzero.
2695 * off and size may not cover an area beyond a single hardware page.
2698 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2700 struct mdglobaldata *gd = mdcpu;
2703 if (*(int *) gd->gd_CMAP3)
2704 panic("pmap_zero_page: CMAP3 busy");
2705 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2706 cpu_invlpg(gd->gd_CADDR3);
2708 #if defined(I686_CPU)
2709 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2710 i686_pagezero(gd->gd_CADDR3);
2713 bzero((char *)gd->gd_CADDR3 + off, size);
2714 *(int *) gd->gd_CMAP3 = 0;
2721 * Copy the physical page from the source PA to the target PA.
2722 * This function may be called from an interrupt. No locking
2726 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2728 struct mdglobaldata *gd = mdcpu;
2731 if (*(int *) gd->gd_CMAP1)
2732 panic("pmap_copy_page: CMAP1 busy");
2733 if (*(int *) gd->gd_CMAP2)
2734 panic("pmap_copy_page: CMAP2 busy");
2736 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2737 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2739 cpu_invlpg(gd->gd_CADDR1);
2740 cpu_invlpg(gd->gd_CADDR2);
2742 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
2744 *(int *) gd->gd_CMAP1 = 0;
2745 *(int *) gd->gd_CMAP2 = 0;
2750 * pmap_copy_page_frag:
2752 * Copy the physical page from the source PA to the target PA.
2753 * This function may be called from an interrupt. No locking
2757 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2759 struct mdglobaldata *gd = mdcpu;
2762 if (*(int *) gd->gd_CMAP1)
2763 panic("pmap_copy_page: CMAP1 busy");
2764 if (*(int *) gd->gd_CMAP2)
2765 panic("pmap_copy_page: CMAP2 busy");
2767 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2768 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2770 cpu_invlpg(gd->gd_CADDR1);
2771 cpu_invlpg(gd->gd_CADDR2);
2773 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2774 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2777 *(int *) gd->gd_CMAP1 = 0;
2778 *(int *) gd->gd_CMAP2 = 0;
2783 * Returns true if the pmap's pv is one of the first
2784 * 16 pvs linked to from this page. This count may
2785 * be changed upwards or downwards in the future; it
2786 * is only necessary that true be returned for a small
2787 * subset of pmaps for proper page aging.
2790 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2795 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2800 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2801 if (pv->pv_pmap == pmap) {
2814 * Remove all pages from specified address space
2815 * this aids process exit speeds. Also, this code
2816 * is special cased for current process only, but
2817 * can have the more generic (and slightly slower)
2818 * mode enabled. This is much faster than pmap_remove
2819 * in the case of running down an entire address space.
2822 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2825 unsigned *pte, tpte;
2828 pmap_inval_info info;
2830 int32_t save_generation;
2832 lp = curthread->td_lwp;
2833 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
2838 pmap_inval_init(&info);
2840 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2841 if (pv->pv_va >= eva || pv->pv_va < sva) {
2842 npv = TAILQ_NEXT(pv, pv_plist);
2846 KKASSERT(pmap == pv->pv_pmap);
2849 pte = (unsigned *)vtopte(pv->pv_va);
2851 pte = pmap_pte_quick(pmap, pv->pv_va);
2852 if (pmap->pm_active)
2853 pmap_inval_add(&info, pmap, pv->pv_va);
2856 * We cannot remove wired pages from a process' mapping
2860 npv = TAILQ_NEXT(pv, pv_plist);
2863 tpte = loadandclear(pte);
2865 m = PHYS_TO_VM_PAGE(tpte);
2867 KASSERT(m < &vm_page_array[vm_page_array_size],
2868 ("pmap_remove_pages: bad tpte %x", tpte));
2870 KKASSERT(pmap->pm_stats.resident_count > 0);
2871 --pmap->pm_stats.resident_count;
2874 * Update the vm_page_t clean and reference bits.
2880 npv = TAILQ_NEXT(pv, pv_plist);
2881 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2882 save_generation = ++pmap->pm_generation;
2884 m->md.pv_list_count--;
2885 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2886 if (TAILQ_EMPTY(&m->md.pv_list))
2887 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2889 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2893 * Restart the scan if we blocked during the unuse or free
2894 * calls and other removals were made.
2896 if (save_generation != pmap->pm_generation) {
2897 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2898 pv = TAILQ_FIRST(&pmap->pm_pvlist);
2901 pmap_inval_flush(&info);
2906 * pmap_testbit tests bits in pte's
2907 * note that the testbit/clearbit routines are inline,
2908 * and a lot of things compile-time evaluate.
2911 pmap_testbit(vm_page_t m, int bit)
2916 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2919 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2924 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2926 * if the bit being tested is the modified bit, then
2927 * mark clean_map and ptes as never
2930 if (bit & (PG_A|PG_M)) {
2931 if (!pmap_track_modified(pv->pv_va))
2935 #if defined(PMAP_DIAGNOSTIC)
2937 kprintf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2941 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2952 * this routine is used to modify bits in ptes
2954 static __inline void
2955 pmap_clearbit(vm_page_t m, int bit)
2957 struct pmap_inval_info info;
2962 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2965 pmap_inval_init(&info);
2969 * Loop over all current mappings setting/clearing as appropos If
2970 * setting RO do we need to clear the VAC?
2972 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2974 * don't write protect pager mappings
2977 if (!pmap_track_modified(pv->pv_va))
2981 #if defined(PMAP_DIAGNOSTIC)
2983 kprintf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2989 * Careful here. We can use a locked bus instruction to
2990 * clear PG_A or PG_M safely but we need to synchronize
2991 * with the target cpus when we mess with PG_RW.
2993 * We do not have to force synchronization when clearing
2994 * PG_M even for PTEs generated via virtual memory maps,
2995 * because the virtual kernel will invalidate the pmap
2996 * entry when/if it needs to resynchronize the Modify bit.
2999 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
3000 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3007 atomic_clear_int(pte, PG_M|PG_RW);
3010 * The cpu may be trying to set PG_M
3011 * simultaniously with our clearing
3014 if (!atomic_cmpset_int(pte, pbits,
3018 } else if (bit == PG_M) {
3020 * We could also clear PG_RW here to force
3021 * a fault on write to redetect PG_M for
3022 * virtual kernels, but it isn't necessary
3023 * since virtual kernels invalidate the pte
3024 * when they clear the VPTE_M bit in their
3025 * virtual page tables.
3027 atomic_clear_int(pte, PG_M);
3029 atomic_clear_int(pte, bit);
3033 pmap_inval_flush(&info);
3038 * pmap_page_protect:
3040 * Lower the permission for all mappings to a given page.
3043 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3045 if ((prot & VM_PROT_WRITE) == 0) {
3046 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3047 pmap_clearbit(m, PG_RW);
3048 vm_page_flag_clear(m, PG_WRITEABLE);
3056 pmap_phys_address(int ppn)
3058 return (i386_ptob(ppn));
3062 * pmap_ts_referenced:
3064 * Return a count of reference bits for a page, clearing those bits.
3065 * It is not necessary for every reference bit to be cleared, but it
3066 * is necessary that 0 only be returned when there are truly no
3067 * reference bits set.
3069 * XXX: The exact number of bits to check and clear is a matter that
3070 * should be tested and standardized at some point in the future for
3071 * optimal aging of shared pages.
3074 pmap_ts_referenced(vm_page_t m)
3076 pv_entry_t pv, pvf, pvn;
3080 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3085 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3090 pvn = TAILQ_NEXT(pv, pv_list);
3092 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3094 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3096 if (!pmap_track_modified(pv->pv_va))
3099 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3101 if (pte && (*pte & PG_A)) {
3103 atomic_clear_int(pte, PG_A);
3105 atomic_clear_int_nonlocked(pte, PG_A);
3112 } while ((pv = pvn) != NULL && pv != pvf);
3122 * Return whether or not the specified physical page was modified
3123 * in any physical maps.
3126 pmap_is_modified(vm_page_t m)
3128 return pmap_testbit(m, PG_M);
3132 * Clear the modify bits on the specified physical page.
3135 pmap_clear_modify(vm_page_t m)
3137 pmap_clearbit(m, PG_M);
3141 * pmap_clear_reference:
3143 * Clear the reference bit on the specified physical page.
3146 pmap_clear_reference(vm_page_t m)
3148 pmap_clearbit(m, PG_A);
3152 * Miscellaneous support routines follow
3156 i386_protection_init(void)
3160 kp = protection_codes;
3161 for (prot = 0; prot < 8; prot++) {
3163 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3165 * Read access is also 0. There isn't any execute bit,
3166 * so just make it readable.
3168 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3169 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3170 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3173 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3174 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3175 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3176 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3184 * Map a set of physical memory pages into the kernel virtual
3185 * address space. Return a pointer to where it is mapped. This
3186 * routine is intended to be used for mapping device memory,
3189 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3193 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3195 vm_offset_t va, tmpva, offset;
3198 offset = pa & PAGE_MASK;
3199 size = roundup(offset + size, PAGE_SIZE);
3201 va = kmem_alloc_nofault(&kernel_map, size);
3203 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3206 for (tmpva = va; size > 0;) {
3207 pte = (unsigned *)vtopte(tmpva);
3208 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3216 return ((void *)(va + offset));
3220 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3222 vm_offset_t base, offset;
3224 base = va & PG_FRAME;
3225 offset = va & PAGE_MASK;
3226 size = roundup(offset + size, PAGE_SIZE);
3227 pmap_qremove(va, size >> PAGE_SHIFT);
3228 kmem_free(&kernel_map, base, size);
3232 * perform the pmap work for mincore
3235 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3237 unsigned *ptep, pte;
3241 ptep = pmap_pte(pmap, addr);
3246 if ((pte = *ptep) != 0) {
3249 val = MINCORE_INCORE;
3250 if ((pte & PG_MANAGED) == 0)
3253 pa = pte & PG_FRAME;
3255 m = PHYS_TO_VM_PAGE(pa);
3261 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3263 * Modified by someone
3265 else if (m->dirty || pmap_is_modified(m))
3266 val |= MINCORE_MODIFIED_OTHER;
3271 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3274 * Referenced by someone
3276 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3277 val |= MINCORE_REFERENCED_OTHER;
3278 vm_page_flag_set(m, PG_REFERENCED);
3285 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3286 * vmspace will be ref'd and the old one will be deref'd.
3288 * The vmspace for all lwps associated with the process will be adjusted
3289 * and cr3 will be reloaded if any lwp is the current lwp.
3292 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3294 struct vmspace *oldvm;
3298 oldvm = p->p_vmspace;
3299 if (oldvm != newvm) {
3300 p->p_vmspace = newvm;
3301 KKASSERT(p->p_nthreads == 1);
3302 lp = RB_ROOT(&p->p_lwp_tree);
3303 pmap_setlwpvm(lp, newvm);
3305 sysref_get(&newvm->vm_sysref);
3306 sysref_put(&oldvm->vm_sysref);
3313 * Set the vmspace for a LWP. The vmspace is almost universally set the
3314 * same as the process vmspace, but virtual kernels need to swap out contexts
3315 * on a per-lwp basis.
3318 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3320 struct vmspace *oldvm;
3324 oldvm = lp->lwp_vmspace;
3326 if (oldvm != newvm) {
3327 lp->lwp_vmspace = newvm;
3328 if (curthread->td_lwp == lp) {
3329 pmap = vmspace_pmap(newvm);
3331 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3333 pmap->pm_active |= 1;
3335 #if defined(SWTCH_OPTIM_STATS)
3338 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pdir);
3339 load_cr3(curthread->td_pcb->pcb_cr3);
3340 pmap = vmspace_pmap(oldvm);
3342 atomic_clear_int(&pmap->pm_active,
3343 1 << mycpu->gd_cpuid);
3345 pmap->pm_active &= ~1;
3353 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3356 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3360 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3367 static void pads (pmap_t pm);
3368 void pmap_pvdump (vm_paddr_t pa);
3370 /* print address space of pmap*/
3377 if (pm == &kernel_pmap)
3380 for (i = 0; i < 1024; i++) {
3381 if (pm->pm_pdir[i]) {
3382 for (j = 0; j < 1024; j++) {
3383 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3384 if (pm == &kernel_pmap && va < KERNBASE)
3386 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3388 ptep = pmap_pte_quick(pm, va);
3389 if (pmap_pte_v(ptep))
3390 kprintf("%x:%x ", va, *(int *) ptep);
3399 pmap_pvdump(vm_paddr_t pa)
3404 kprintf("pa %08llx", (long long)pa);
3405 m = PHYS_TO_VM_PAGE(pa);
3406 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3408 kprintf(" -> pmap %p, va %x, flags %x",
3409 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3411 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);