2 * Copyright (c) 1991 Regents of the University of California.
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1994 David Greenman
5 * Copyright (c) 2003 Peter Wemm
6 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
7 * Copyright (c) 2008, 2009 The DragonFly Project.
8 * Copyright (c) 2008, 2009 Jordan Gordeev.
11 * This code is derived from software contributed to Berkeley by
12 * the Systems Programming Group of the University of Utah Computer
13 * Science Department and William Jolitz of UUNET Technologies Inc.
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
44 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
48 * Manages physical address maps.
50 * In addition to hardware address maps, this
51 * module is called upon to provide software-use-only
52 * maps which may or may not be stored in the same
53 * form as hardware maps. These pseudo-maps are
54 * used to store intermediate results from copy
55 * operations to and from address spaces.
57 * Since the information managed by this module is
58 * also stored by the logical address mapping module,
59 * this module may throw away valid virtual-to-physical
60 * mappings at almost any time. However, invalidations
61 * of virtual-to-physical mappings must be done as
64 * In order to cope with hardware architectures which
65 * make virtual-to-physical map invalidates expensive,
66 * this module may delay invalidate or reduced protection
67 * operations until such time as they are actually
68 * necessary. This module is given full information as
69 * to which processors are currently using which maps,
70 * and to when physical maps must be made correct.
74 #include "opt_disable_pse.h"
77 #include "opt_msgbuf.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
83 #include <sys/msgbuf.h>
84 #include <sys/vmmeter.h>
88 #include <vm/vm_param.h>
89 #include <sys/sysctl.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_pageout.h>
97 #include <vm/vm_pager.h>
98 #include <vm/vm_zone.h>
100 #include <sys/user.h>
101 #include <sys/thread2.h>
102 #include <sys/sysref2.h>
104 #include <machine/cputypes.h>
105 #include <machine/md_var.h>
106 #include <machine/specialreg.h>
107 #include <machine/smp.h>
108 #include <machine_base/apic/apicreg.h>
109 #include <machine/globaldata.h>
110 #include <machine/pmap.h>
111 #include <machine/pmap_inval.h>
115 #define PMAP_KEEP_PDIRS
116 #ifndef PMAP_SHPGPERPROC
117 #define PMAP_SHPGPERPROC 200
120 #if defined(DIAGNOSTIC)
121 #define PMAP_DIAGNOSTIC
127 * Get PDEs and PTEs for user/kernel address space
129 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
130 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
132 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
133 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
134 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
135 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
136 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
140 * Given a map and a machine independent protection code,
141 * convert to a vax protection code.
143 #define pte_prot(m, p) \
144 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
145 static int protection_codes[8];
147 struct pmap kernel_pmap;
148 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
150 vm_paddr_t avail_start; /* PA of first available physical page */
151 vm_paddr_t avail_end; /* PA of last available physical page */
152 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
153 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
154 vm_offset_t KvaStart; /* VA start of KVA space */
155 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
156 vm_offset_t KvaSize; /* max size of kernel virtual address space */
157 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
158 static int pgeflag; /* PG_G or-in */
159 static int pseflag; /* PG_PS or-in */
161 static vm_object_t kptobj;
164 static vm_paddr_t dmaplimit;
166 vm_offset_t kernel_vm_end;
168 static uint64_t KPDphys; /* phys addr of kernel level 2 */
169 uint64_t KPDPphys; /* phys addr of kernel level 3 */
170 uint64_t KPML4phys; /* phys addr of kernel level 4 */
172 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
173 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
176 * Data for the pv entry allocation mechanism
178 static vm_zone_t pvzone;
179 static struct vm_zone pvzone_store;
180 static struct vm_object pvzone_obj;
181 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
182 static int pmap_pagedaemon_waken = 0;
183 static struct pv_entry *pvinit;
186 * All those kernel PT submaps that BSD is so fond of
188 pt_entry_t *CMAP1 = 0, *ptmmap;
189 caddr_t CADDR1 = 0, ptvmmap = 0;
190 static pt_entry_t *msgbufmap;
191 struct msgbuf *msgbufp=0;
196 static pt_entry_t *pt_crashdumpmap;
197 static caddr_t crashdumpmap;
199 extern uint64_t KPTphys;
200 extern pt_entry_t *SMPpt;
201 extern uint64_t SMPptpa;
205 static pv_entry_t get_pv_entry (void);
206 static void i386_protection_init (void);
207 static void create_pagetables(vm_paddr_t *firstaddr);
208 static void pmap_remove_all (vm_page_t m);
209 static void pmap_enter_quick (pmap_t pmap, vm_offset_t va, vm_page_t m);
210 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
211 vm_offset_t sva, pmap_inval_info_t info);
212 static void pmap_remove_page (struct pmap *pmap,
213 vm_offset_t va, pmap_inval_info_t info);
214 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
215 vm_offset_t va, pmap_inval_info_t info);
216 static boolean_t pmap_testbit (vm_page_t m, int bit);
217 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
218 vm_page_t mpte, vm_page_t m);
220 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
222 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
223 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
224 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
225 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
226 static int _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
227 pmap_inval_info_t info);
228 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
229 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
231 static unsigned pdir4mb;
234 * Move the kernel virtual free pointer to the next
235 * 2MB. This is used to help improve performance
236 * by using a large (2MB) page for much of the kernel
237 * (.text, .data, .bss)
241 pmap_kmem_choose(vm_offset_t addr)
243 vm_offset_t newaddr = addr;
245 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
252 * Super fast pmap_pte routine best used when scanning the pv lists.
253 * This eliminates many course-grained invltlb calls. Note that many of
254 * the pv list scans are across different pmaps and it is very wasteful
255 * to do an entire invltlb when checking a single mapping.
257 * Should only be called while in a critical section.
259 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
263 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
265 return pmap_pte(pmap, va);
268 /* Return a non-clipped PD index for a given VA */
271 pmap_pde_pindex(vm_offset_t va)
273 return va >> PDRSHIFT;
276 /* Return various clipped indexes for a given VA */
279 pmap_pte_index(vm_offset_t va)
282 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
287 pmap_pde_index(vm_offset_t va)
290 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
295 pmap_pdpe_index(vm_offset_t va)
298 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
303 pmap_pml4e_index(vm_offset_t va)
306 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
309 /* Return a pointer to the PML4 slot that corresponds to a VA */
312 pmap_pml4e(pmap_t pmap, vm_offset_t va)
315 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
318 /* Return a pointer to the PDP slot that corresponds to a VA */
321 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
325 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
326 return (&pdpe[pmap_pdpe_index(va)]);
329 /* Return a pointer to the PDP slot that corresponds to a VA */
332 pmap_pdpe(pmap_t pmap, vm_offset_t va)
336 pml4e = pmap_pml4e(pmap, va);
337 if ((*pml4e & PG_V) == 0)
339 return (pmap_pml4e_to_pdpe(pml4e, va));
342 /* Return a pointer to the PD slot that corresponds to a VA */
345 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
349 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
350 return (&pde[pmap_pde_index(va)]);
353 /* Return a pointer to the PD slot that corresponds to a VA */
356 pmap_pde(pmap_t pmap, vm_offset_t va)
360 pdpe = pmap_pdpe(pmap, va);
361 if (pdpe == NULL || (*pdpe & PG_V) == 0)
363 return (pmap_pdpe_to_pde(pdpe, va));
366 /* Return a pointer to the PT slot that corresponds to a VA */
369 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
373 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
374 return (&pte[pmap_pte_index(va)]);
377 /* Return a pointer to the PT slot that corresponds to a VA */
380 pmap_pte(pmap_t pmap, vm_offset_t va)
384 pde = pmap_pde(pmap, va);
385 if (pde == NULL || (*pde & PG_V) == 0)
387 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
388 return ((pt_entry_t *)pde);
389 return (pmap_pde_to_pte(pde, va));
394 vtopte(vm_offset_t va)
396 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
398 return (PTmap + ((va >> PAGE_SHIFT) & mask));
403 vtopde(vm_offset_t va)
405 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
407 return (PDmap + ((va >> PDRSHIFT) & mask));
411 allocpages(vm_paddr_t *firstaddr, int n)
416 bzero((void *)ret, n * PAGE_SIZE);
417 *firstaddr += n * PAGE_SIZE;
423 create_pagetables(vm_paddr_t *firstaddr)
427 /* we are running (mostly) V=P at this point */
430 KPTphys = allocpages(firstaddr, NKPT);
431 KPML4phys = allocpages(firstaddr, 1);
432 KPDPphys = allocpages(firstaddr, NKPML4E);
433 KPDphys = allocpages(firstaddr, NKPDPE);
435 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
436 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
438 DMPDPphys = allocpages(firstaddr, NDMPML4E);
439 if ((amd_feature & AMDID_PAGE1GB) == 0)
440 DMPDphys = allocpages(firstaddr, ndmpdp);
441 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
443 /* Fill in the underlying page table pages */
444 /* Read-only from zero to physfree */
445 /* XXX not fully used, underneath 2M pages */
446 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
447 ((pt_entry_t *)KPTphys)[i] = i << PAGE_SHIFT;
448 ((pt_entry_t *)KPTphys)[i] |= PG_RW | PG_V | PG_G;
451 /* Now map the page tables at their location within PTmap */
452 for (i = 0; i < NKPT; i++) {
453 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
454 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
457 /* Map from zero to end of allocations under 2M pages */
458 /* This replaces some of the KPTphys entries above */
459 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
460 ((pd_entry_t *)KPDphys)[i] = i << PDRSHIFT;
461 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V | PG_PS | PG_G;
464 /* And connect up the PD to the PDP */
465 for (i = 0; i < NKPDPE; i++) {
466 ((pdp_entry_t *)KPDPphys)[i + KPDPI] = KPDphys +
468 ((pdp_entry_t *)KPDPphys)[i + KPDPI] |= PG_RW | PG_V | PG_U;
471 /* Now set up the direct map space using either 2MB or 1GB pages */
472 /* Preset PG_M and PG_A because demotion expects it */
473 if ((amd_feature & AMDID_PAGE1GB) == 0) {
474 for (i = 0; i < NPDEPG * ndmpdp; i++) {
475 ((pd_entry_t *)DMPDphys)[i] = (vm_paddr_t)i << PDRSHIFT;
476 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
479 /* And the direct map space's PDP */
480 for (i = 0; i < ndmpdp; i++) {
481 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
483 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
486 for (i = 0; i < ndmpdp; i++) {
487 ((pdp_entry_t *)DMPDPphys)[i] =
488 (vm_paddr_t)i << PDPSHIFT;
489 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
494 /* And recursively map PML4 to itself in order to get PTmap */
495 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
496 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
498 /* Connect the Direct Map slot up to the PML4 */
499 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
500 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
502 /* Connect the KVA slot up to the PML4 */
503 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
504 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
508 init_paging(vm_paddr_t *firstaddr)
510 create_pagetables(firstaddr);
514 * Bootstrap the system enough to run with virtual memory.
516 * On the i386 this is called after mapping has already been enabled
517 * and just syncs the pmap module with what has already been done.
518 * [We can't call it easily with mapping off since the kernel is not
519 * mapped with PA == VA, hence we would have to relocate every address
520 * from the linked base (virtual) address "KERNBASE" to the actual
521 * (physical) address starting relative to 0]
524 pmap_bootstrap(vm_paddr_t *firstaddr)
528 struct mdglobaldata *gd;
531 KvaStart = VM_MIN_KERNEL_ADDRESS;
532 KvaEnd = VM_MAX_KERNEL_ADDRESS;
533 KvaSize = KvaEnd - KvaStart;
535 avail_start = *firstaddr;
538 * Create an initial set of page tables to run the kernel in.
540 create_pagetables(firstaddr);
542 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
543 virtual_start = pmap_kmem_choose(virtual_start);
545 virtual_end = VM_MAX_KERNEL_ADDRESS;
547 /* XXX do %cr0 as well */
548 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
552 * Initialize protection array.
554 i386_protection_init();
557 * The kernel's pmap is statically allocated so we don't have to use
558 * pmap_create, which is unlikely to work correctly at this part of
559 * the boot sequence (XXX and which no longer exists).
561 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
562 kernel_pmap.pm_count = 1;
563 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
564 TAILQ_INIT(&kernel_pmap.pm_pvlist);
568 * Reserve some special page table entries/VA space for temporary
571 #define SYSMAP(c, p, v, n) \
572 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
576 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
582 * CMAP1/CMAP2 are used for zeroing and copying pages.
584 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
589 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
592 * ptvmmap is used for reading arbitrary physical pages via
595 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
598 * msgbufp is used to map the system message buffer.
599 * XXX msgbufmap is not used.
601 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
602 atop(round_page(MSGBUF_SIZE)))
609 * PG_G is terribly broken on SMP because we IPI invltlb's in some
610 * cases rather then invl1pg. Actually, I don't even know why it
611 * works under UP because self-referential page table mappings
616 if (cpu_feature & CPUID_PGE)
621 * Initialize the 4MB page size flag
625 * The 4MB page version of the initial
626 * kernel page mapping.
630 #if !defined(DISABLE_PSE)
631 if (cpu_feature & CPUID_PSE) {
634 * Note that we have enabled PSE mode
637 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
638 ptditmp &= ~(NBPDR - 1);
639 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
644 * Enable the PSE mode. If we are SMP we can't do this
645 * now because the APs will not be able to use it when
648 load_cr4(rcr4() | CR4_PSE);
651 * We can do the mapping here for the single processor
652 * case. We simply ignore the old page table page from
656 * For SMP, we still need 4K pages to bootstrap APs,
657 * PSE will be enabled as soon as all APs are up.
659 PTD[KPTDI] = (pd_entry_t)ptditmp;
665 if (cpu_apic_address == 0)
666 panic("pmap_bootstrap: no local apic!");
670 * We need to finish setting up the globaldata page for the BSP.
671 * locore has already populated the page table for the mdglobaldata
674 pg = MDGLOBALDATA_BASEALLOC_PAGES;
675 gd = &CPU_prvspace[0].mdglobaldata;
676 gd->gd_CMAP1 = &SMPpt[pg + 0];
677 gd->gd_CMAP2 = &SMPpt[pg + 1];
678 gd->gd_CMAP3 = &SMPpt[pg + 2];
679 gd->gd_PMAP1 = &SMPpt[pg + 3];
680 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
681 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
682 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
683 gd->gd_PADDR1 = (pt_entry_t *)CPU_prvspace[0].PPAGE1;
690 * Set 4mb pdir for mp startup
695 if (pseflag && (cpu_feature & CPUID_PSE)) {
696 load_cr4(rcr4() | CR4_PSE);
697 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
705 * Initialize the pmap module.
706 * Called by vm_init, to initialize any structures that the pmap
707 * system needs to map virtual memory.
708 * pmap_init has been enhanced to support in a fairly consistant
709 * way, discontiguous physical memory.
718 * object for kernel page table pages
720 /* JG I think the number can be arbitrary */
721 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
724 * Allocate memory for random pmap data structures. Includes the
728 for(i = 0; i < vm_page_array_size; i++) {
731 m = &vm_page_array[i];
732 TAILQ_INIT(&m->md.pv_list);
733 m->md.pv_list_count = 0;
737 * init the pv free list
739 initial_pvs = vm_page_array_size;
740 if (initial_pvs < MINPV)
742 pvzone = &pvzone_store;
743 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
744 initial_pvs * sizeof (struct pv_entry));
745 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
749 * Now it is safe to enable pv_table recording.
751 pmap_initialized = TRUE;
753 lapic = pmap_mapdev_uncacheable(cpu_apic_address, sizeof(struct LAPIC));
758 * Initialize the address space (zone) for the pv_entries. Set a
759 * high water mark so that the system can recover from excessive
760 * numbers of pv entries.
765 int shpgperproc = PMAP_SHPGPERPROC;
767 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
768 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
769 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
770 pv_entry_high_water = 9 * (pv_entry_max / 10);
771 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
775 /***************************************************
776 * Low level helper routines.....
777 ***************************************************/
779 #if defined(PMAP_DIAGNOSTIC)
782 * This code checks for non-writeable/modified pages.
783 * This should be an invalid condition.
787 pmap_nw_modified(pt_entry_t pte)
789 if ((pte & (PG_M|PG_RW)) == PG_M)
798 * this routine defines the region(s) of memory that should
799 * not be tested for the modified bit.
803 pmap_track_modified(vm_offset_t va)
805 if ((va < clean_sva) || (va >= clean_eva))
814 * Extract the physical page address associated with the map/VA pair.
816 * This function may not be called from an interrupt if the pmap is
820 pmap_extract(pmap_t pmap, vm_offset_t va)
824 pd_entry_t pde, *pdep;
827 pdep = pmap_pde(pmap, va);
831 if ((pde & PG_PS) != 0) {
832 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
834 pte = pmap_pde_to_pte(pdep, va);
835 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
843 * Routine: pmap_kextract
845 * Extract the physical page address associated
846 * kernel virtual address.
849 pmap_kextract(vm_offset_t va)
854 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
855 pa = DMAP_TO_PHYS(va);
859 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
862 * Beware of a concurrent promotion that changes the
863 * PDE at this point! For example, vtopte() must not
864 * be used to access the PTE because it would use the
865 * new PDE. It is, however, safe to use the old PDE
866 * because the page table page is preserved by the
869 pa = *pmap_pde_to_pte(&pde, va);
870 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
876 /***************************************************
877 * Low level mapping routines.....
878 ***************************************************/
881 * Routine: pmap_kenter
883 * Add a wired page to the KVA
884 * NOTE! note that in order for the mapping to take effect -- you
885 * should do an invltlb after doing the pmap_kenter().
888 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
892 pmap_inval_info info;
894 pmap_inval_init(&info);
895 npte = pa | PG_RW | PG_V | pgeflag;
897 pmap_inval_add(&info, &kernel_pmap, va);
899 pmap_inval_flush(&info);
903 * Routine: pmap_kenter_quick
905 * Similar to pmap_kenter(), except we only invalidate the
906 * mapping on the current CPU.
909 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
914 npte = pa | PG_RW | PG_V | pgeflag;
917 cpu_invlpg((void *)va);
921 pmap_kenter_sync(vm_offset_t va)
923 pmap_inval_info info;
925 pmap_inval_init(&info);
926 pmap_inval_add(&info, &kernel_pmap, va);
927 pmap_inval_flush(&info);
931 pmap_kenter_sync_quick(vm_offset_t va)
933 cpu_invlpg((void *)va);
937 * remove a page from the kernel pagetables
940 pmap_kremove(vm_offset_t va)
943 pmap_inval_info info;
945 pmap_inval_init(&info);
947 pmap_inval_add(&info, &kernel_pmap, va);
949 pmap_inval_flush(&info);
953 pmap_kremove_quick(vm_offset_t va)
958 cpu_invlpg((void *)va);
962 * XXX these need to be recoded. They are not used in any critical path.
965 pmap_kmodify_rw(vm_offset_t va)
967 *vtopte(va) |= PG_RW;
968 cpu_invlpg((void *)va);
972 pmap_kmodify_nc(vm_offset_t va)
975 cpu_invlpg((void *)va);
979 * Used to map a range of physical addresses into kernel
980 * virtual address space.
982 * For now, VM is already on, we only need to map the
986 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
988 return PHYS_TO_DMAP(start);
993 * Add a list of wired pages to the kva
994 * this routine is only used for temporary
995 * kernel mappings that do not need to have
996 * page modification or references recorded.
997 * Note that old mappings are simply written
998 * over. The page *must* be wired.
1001 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1005 end_va = va + count * PAGE_SIZE;
1007 while (va < end_va) {
1011 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1012 cpu_invlpg((void *)va);
1017 smp_invltlb(); /* XXX */
1022 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
1025 cpumask_t cmask = mycpu->gd_cpumask;
1027 end_va = va + count * PAGE_SIZE;
1029 while (va < end_va) {
1034 * Install the new PTE. If the pte changed from the prior
1035 * mapping we must reset the cpu mask and invalidate the page.
1036 * If the pte is the same but we have not seen it on the
1037 * current cpu, invlpg the existing mapping. Otherwise the
1038 * entry is optimal and no invalidation is required.
1041 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
1042 if (*pte != pteval) {
1045 cpu_invlpg((void *)va);
1046 } else if ((*mask & cmask) == 0) {
1047 cpu_invlpg((void *)va);
1056 * This routine jerks page mappings from the
1057 * kernel -- it is meant only for temporary mappings.
1059 * MPSAFE, INTERRUPT SAFE (cluster callback)
1062 pmap_qremove(vm_offset_t va, int count)
1066 end_va = va + count * PAGE_SIZE;
1068 while (va < end_va) {
1073 cpu_invlpg((void *)va);
1082 * This routine works like vm_page_lookup() but also blocks as long as the
1083 * page is busy. This routine does not busy the page it returns.
1085 * Unless the caller is managing objects whos pages are in a known state,
1086 * the call should be made with a critical section held so the page's object
1087 * association remains valid on return.
1091 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1096 m = vm_page_lookup(object, pindex);
1097 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
1103 * Create a new thread and optionally associate it with a (new) process.
1104 * NOTE! the new thread's cpu may not equal the current cpu.
1107 pmap_init_thread(thread_t td)
1109 /* enforce pcb placement */
1110 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1111 td->td_savefpu = &td->td_pcb->pcb_save;
1112 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on x86_64? */
1116 * This routine directly affects the fork perf for a process.
1119 pmap_init_proc(struct proc *p)
1124 * Dispose the UPAGES for a process that has exited.
1125 * This routine directly impacts the exit perf of a process.
1128 pmap_dispose_proc(struct proc *p)
1130 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1133 /***************************************************
1134 * Page table page management routines.....
1135 ***************************************************/
1138 * This routine unholds page table pages, and if the hold count
1139 * drops to zero, then it decrements the wire count.
1143 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1144 pmap_inval_info_t info)
1146 KKASSERT(m->hold_count > 0);
1147 if (m->hold_count > 1) {
1151 return _pmap_unwire_pte_hold(pmap, va, m, info);
1157 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1158 pmap_inval_info_t info)
1161 * Wait until we can busy the page ourselves. We cannot have
1162 * any active flushes if we block. We own one hold count on the
1163 * page so it cannot be freed out from under us.
1165 if (m->flags & PG_BUSY) {
1166 pmap_inval_flush(info);
1167 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1170 KASSERT(m->queue == PQ_NONE,
1171 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1174 * This case can occur if new references were acquired while
1177 if (m->hold_count > 1) {
1178 KKASSERT(m->hold_count > 1);
1184 * Unmap the page table page
1186 KKASSERT(m->hold_count == 1);
1188 pmap_inval_add(info, pmap, -1);
1190 if (m->pindex >= (NUPDE + NUPDPE)) {
1193 pml4 = pmap_pml4e(pmap, va);
1195 } else if (m->pindex >= NUPDE) {
1198 pdp = pmap_pdpe(pmap, va);
1203 pd = pmap_pde(pmap, va);
1207 KKASSERT(pmap->pm_stats.resident_count > 0);
1208 --pmap->pm_stats.resident_count;
1210 if (pmap->pm_ptphint == m)
1211 pmap->pm_ptphint = NULL;
1213 if (m->pindex < NUPDE) {
1214 /* We just released a PT, unhold the matching PD */
1217 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
1218 pmap_unwire_pte_hold(pmap, va, pdpg, info);
1220 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
1221 /* We just released a PD, unhold the matching PDP */
1224 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
1225 pmap_unwire_pte_hold(pmap, va, pdppg, info);
1229 * This was our last hold, the page had better be unwired
1230 * after we decrement wire_count.
1232 * FUTURE NOTE: shared page directory page could result in
1233 * multiple wire counts.
1237 KKASSERT(m->wire_count == 0);
1238 --vmstats.v_wire_count;
1239 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1241 vm_page_free_zero(m);
1247 * After removing a page table entry, this routine is used to
1248 * conditionally free the page, and manage the hold/wire counts.
1252 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1253 pmap_inval_info_t info)
1255 vm_pindex_t ptepindex;
1257 if (va >= VM_MAX_USER_ADDRESS)
1261 ptepindex = pmap_pde_pindex(va);
1263 if (pmap->pm_ptphint &&
1264 (pmap->pm_ptphint->pindex == ptepindex)) {
1265 mpte = pmap->pm_ptphint;
1268 pmap_inval_flush(info);
1269 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1270 pmap->pm_ptphint = mpte;
1275 return pmap_unwire_pte_hold(pmap, va, mpte, info);
1279 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1280 * it, and IdlePTD, represents the template used to update all other pmaps.
1282 * On architectures where the kernel pmap is not integrated into the user
1283 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1284 * kernel_pmap should be used to directly access the kernel_pmap.
1287 pmap_pinit0(struct pmap *pmap)
1289 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1291 pmap->pm_active = 0;
1292 pmap->pm_ptphint = NULL;
1293 TAILQ_INIT(&pmap->pm_pvlist);
1294 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1298 * Initialize a preallocated and zeroed pmap structure,
1299 * such as one in a vmspace structure.
1302 pmap_pinit(struct pmap *pmap)
1307 * No need to allocate page table space yet but we do need a valid
1308 * page directory table.
1310 if (pmap->pm_pml4 == NULL) {
1312 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1316 * Allocate an object for the ptes
1318 if (pmap->pm_pteobj == NULL)
1319 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1);
1322 * Allocate the page directory page, unless we already have
1323 * one cached. If we used the cached page the wire_count will
1324 * already be set appropriately.
1326 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1327 ptdpg = vm_page_grab(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I,
1328 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1329 pmap->pm_pdirm = ptdpg;
1330 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1331 ptdpg->valid = VM_PAGE_BITS_ALL;
1332 if (ptdpg->wire_count == 0)
1333 ++vmstats.v_wire_count;
1334 ptdpg->wire_count = 1;
1335 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1337 if ((ptdpg->flags & PG_ZERO) == 0)
1338 bzero(pmap->pm_pml4, PAGE_SIZE);
1340 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1341 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1343 /* install self-referential address mapping entry */
1344 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1347 pmap->pm_active = 0;
1348 pmap->pm_ptphint = NULL;
1349 TAILQ_INIT(&pmap->pm_pvlist);
1350 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1351 pmap->pm_stats.resident_count = 1;
1355 * Clean up a pmap structure so it can be physically freed. This routine
1356 * is called by the vmspace dtor function. A great deal of pmap data is
1357 * left passively mapped to improve vmspace management so we have a bit
1358 * of cleanup work to do here.
1361 pmap_puninit(pmap_t pmap)
1365 KKASSERT(pmap->pm_active == 0);
1366 if ((p = pmap->pm_pdirm) != NULL) {
1367 KKASSERT(pmap->pm_pml4 != NULL);
1368 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1369 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1371 vmstats.v_wire_count--;
1372 KKASSERT((p->flags & PG_BUSY) == 0);
1374 vm_page_free_zero(p);
1375 pmap->pm_pdirm = NULL;
1377 if (pmap->pm_pml4) {
1378 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1379 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1380 pmap->pm_pml4 = NULL;
1382 if (pmap->pm_pteobj) {
1383 vm_object_deallocate(pmap->pm_pteobj);
1384 pmap->pm_pteobj = NULL;
1389 * Wire in kernel global address entries. To avoid a race condition
1390 * between pmap initialization and pmap_growkernel, this procedure
1391 * adds the pmap to the master list (which growkernel scans to update),
1392 * then copies the template.
1395 pmap_pinit2(struct pmap *pmap)
1398 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1399 /* XXX copies current process, does not fill in MPPTDI */
1404 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1405 * 0 on failure (if the procedure had to sleep).
1407 * When asked to remove the page directory page itself, we actually just
1408 * leave it cached so we do not have to incur the SMP inval overhead of
1409 * removing the kernel mapping. pmap_puninit() will take care of it.
1413 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1416 * This code optimizes the case of freeing non-busy
1417 * page-table pages. Those pages are zero now, and
1418 * might as well be placed directly into the zero queue.
1420 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1426 * Remove the page table page from the processes address space.
1428 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1430 * We are the pml4 table itself.
1432 /* XXX anything to do here? */
1433 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1435 * Remove a PDP page from the PML4. We do not maintain
1436 * hold counts on the PML4 page.
1442 m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1443 KKASSERT(m4 != NULL);
1444 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1445 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1446 KKASSERT(pml4[idx] != 0);
1448 } else if (p->pindex >= NUPDE) {
1450 * Remove a PD page from the PDP and drop the hold count
1451 * on the PDP. The PDP is left cached in the pmap if
1452 * the hold count drops to 0 so the wire count remains
1459 m3 = vm_page_lookup(pmap->pm_pteobj,
1460 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1461 KKASSERT(m3 != NULL);
1462 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1463 idx = (p->pindex - NUPDE) % NPDPEPG;
1464 KKASSERT(pdp[idx] != 0);
1469 * Remove a PT page from the PD and drop the hold count
1470 * on the PD. The PD is left cached in the pmap if
1471 * the hold count drops to 0 so the wire count remains
1478 m2 = vm_page_lookup(pmap->pm_pteobj,
1479 NUPDE + p->pindex / NPDEPG);
1480 KKASSERT(m2 != NULL);
1481 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1482 idx = p->pindex % NPDEPG;
1488 * One fewer mappings in the pmap. p's hold count had better
1491 KKASSERT(pmap->pm_stats.resident_count > 0);
1492 --pmap->pm_stats.resident_count;
1494 panic("pmap_release: freeing held page table page");
1495 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1496 pmap->pm_ptphint = NULL;
1499 * We leave the top-level page table page cached, wired, and mapped in
1500 * the pmap until the dtor function (pmap_puninit()) gets called.
1501 * However, still clean it up so we can set PG_ZERO.
1503 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1504 bzero(pmap->pm_pml4, PAGE_SIZE);
1505 vm_page_flag_set(p, PG_ZERO);
1509 KKASSERT(p->wire_count == 0);
1510 vmstats.v_wire_count--;
1511 /* JG eventually revert to using vm_page_free_zero() */
1518 * This routine is called when various levels in the page table need to
1519 * be populated. This routine cannot fail.
1523 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1528 * Find or fabricate a new pagetable page. This will busy the page.
1530 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1531 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1532 if ((m->flags & PG_ZERO) == 0) {
1533 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1536 KASSERT(m->queue == PQ_NONE,
1537 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1540 * Increment the hold count for the page we will be returning to
1544 if (m->wire_count++ == 0)
1545 vmstats.v_wire_count++;
1548 * Map the pagetable page into the process address space, if
1549 * it isn't already there.
1551 * It is possible that someone else got in and mapped the page
1552 * directory page while we were blocked, if so just unbusy and
1553 * return the held page.
1555 if (ptepindex >= (NUPDE + NUPDPE)) {
1557 * Wire up a new PDP page in the PML4
1559 vm_pindex_t pml4index;
1562 pml4index = ptepindex - (NUPDE + NUPDPE);
1563 pml4 = &pmap->pm_pml4[pml4index];
1565 if (--m->wire_count == 0)
1566 --vmstats.v_wire_count;
1570 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1571 } else if (ptepindex >= NUPDE) {
1573 * Wire up a new PD page in the PDP
1575 vm_pindex_t pml4index;
1576 vm_pindex_t pdpindex;
1581 pdpindex = ptepindex - NUPDE;
1582 pml4index = pdpindex >> NPML4EPGSHIFT;
1584 pml4 = &pmap->pm_pml4[pml4index];
1585 if ((*pml4 & PG_V) == 0) {
1587 * Have to allocate a new PDP page, recurse.
1588 * This always succeeds. Returned page will
1591 pdppg = _pmap_allocpte(pmap,
1592 NUPDE + NUPDPE + pml4index);
1595 * Add a held reference to the PDP page.
1597 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1598 pdppg->hold_count++;
1602 * Now find the pdp_entry and map the PDP. If the PDP
1603 * has already been mapped unwind and return the
1604 * already-mapped PDP held.
1606 * pdppg is left held (hold_count is incremented for
1607 * each PD in the PDP).
1609 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1610 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1612 vm_page_unhold(pdppg);
1613 if (--m->wire_count == 0)
1614 --vmstats.v_wire_count;
1618 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1621 * Wire up the new PT page in the PD
1623 vm_pindex_t pml4index;
1624 vm_pindex_t pdpindex;
1630 pdpindex = ptepindex >> NPDPEPGSHIFT;
1631 pml4index = pdpindex >> NPML4EPGSHIFT;
1634 * Locate the PDP page in the PML4, then the PD page in
1635 * the PDP. If either does not exist we simply recurse
1638 * We can just recurse on the PD page as it will recurse
1639 * on the PDP if necessary.
1641 pml4 = &pmap->pm_pml4[pml4index];
1642 if ((*pml4 & PG_V) == 0) {
1643 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1644 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1645 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1647 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1648 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1649 if ((*pdp & PG_V) == 0) {
1650 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1652 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1658 * Now fill in the pte in the PD. If the pte already exists
1659 * (again, if we raced the grab), unhold pdpg and unwire
1660 * m, returning a held m.
1662 * pdpg is left held (hold_count is incremented for
1663 * each PT in the PD).
1665 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
1666 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1668 vm_page_unhold(pdpg);
1669 if (--m->wire_count == 0)
1670 --vmstats.v_wire_count;
1674 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1678 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1679 * valid bits, mapped flag, unbusy, and we're done.
1681 pmap->pm_ptphint = m;
1682 ++pmap->pm_stats.resident_count;
1684 m->valid = VM_PAGE_BITS_ALL;
1685 vm_page_flag_clear(m, PG_ZERO);
1686 vm_page_flag_set(m, PG_MAPPED);
1694 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1696 vm_pindex_t ptepindex;
1701 * Calculate pagetable page index
1703 ptepindex = pmap_pde_pindex(va);
1706 * Get the page directory entry
1708 pd = pmap_pde(pmap, va);
1711 * This supports switching from a 2MB page to a
1714 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1715 panic("no promotion/demotion yet");
1723 * If the page table page is mapped, we just increment the
1724 * hold count, and activate it.
1726 if (pd != NULL && (*pd & PG_V) != 0) {
1727 /* YYY hint is used here on i386 */
1728 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1729 pmap->pm_ptphint = m;
1734 * Here if the pte page isn't mapped, or if it has been deallocated.
1736 return _pmap_allocpte(pmap, ptepindex);
1740 /***************************************************
1741 * Pmap allocation/deallocation routines.
1742 ***************************************************/
1745 * Release any resources held by the given physical map.
1746 * Called when a pmap initialized by pmap_pinit is being released.
1747 * Should only be called if the map contains no valid mappings.
1749 static int pmap_release_callback(struct vm_page *p, void *data);
1752 pmap_release(struct pmap *pmap)
1754 vm_object_t object = pmap->pm_pteobj;
1755 struct rb_vm_page_scan_info info;
1757 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1758 #if defined(DIAGNOSTIC)
1759 if (object->ref_count != 1)
1760 panic("pmap_release: pteobj reference count != 1");
1764 info.object = object;
1766 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1773 info.limit = object->generation;
1775 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1776 pmap_release_callback, &info);
1777 if (info.error == 0 && info.mpte) {
1778 if (!pmap_release_free_page(pmap, info.mpte))
1782 } while (info.error);
1787 pmap_release_callback(struct vm_page *p, void *data)
1789 struct rb_vm_page_scan_info *info = data;
1791 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1795 if (!pmap_release_free_page(info->pmap, p)) {
1799 if (info->object->generation != info->limit) {
1807 * Grow the number of kernel page table entries, if needed.
1810 pmap_growkernel(vm_offset_t addr)
1813 vm_offset_t ptppaddr;
1815 pd_entry_t *pde, newpdir;
1819 if (kernel_vm_end == 0) {
1820 kernel_vm_end = KERNBASE;
1822 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1823 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1825 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1826 kernel_vm_end = kernel_map.max_offset;
1831 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1832 if (addr - 1 >= kernel_map.max_offset)
1833 addr = kernel_map.max_offset;
1834 while (kernel_vm_end < addr) {
1835 pde = pmap_pde(&kernel_pmap, kernel_vm_end);
1837 /* We need a new PDP entry */
1838 nkpg = vm_page_alloc(kptobj, nkpt,
1839 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM
1840 | VM_ALLOC_INTERRUPT);
1842 panic("pmap_growkernel: no memory to grow kernel");
1843 paddr = VM_PAGE_TO_PHYS(nkpg);
1844 if ((nkpg->flags & PG_ZERO) == 0)
1845 pmap_zero_page(paddr);
1846 vm_page_flag_clear(nkpg, PG_ZERO);
1847 newpdp = (pdp_entry_t)
1848 (paddr | PG_V | PG_RW | PG_A | PG_M);
1849 *pmap_pdpe(&kernel_pmap, kernel_vm_end) = newpdp;
1851 continue; /* try again */
1853 if ((*pde & PG_V) != 0) {
1854 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1855 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1856 kernel_vm_end = kernel_map.max_offset;
1863 * This index is bogus, but out of the way
1865 nkpg = vm_page_alloc(kptobj, nkpt,
1866 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1868 panic("pmap_growkernel: no memory to grow kernel");
1871 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1872 pmap_zero_page(ptppaddr);
1873 vm_page_flag_clear(nkpg, PG_ZERO);
1874 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1875 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1878 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1879 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1880 kernel_vm_end = kernel_map.max_offset;
1888 * Retire the given physical map from service.
1889 * Should only be called if the map contains
1890 * no valid mappings.
1893 pmap_destroy(pmap_t pmap)
1900 count = --pmap->pm_count;
1903 panic("destroying a pmap is not yet implemented");
1908 * Add a reference to the specified pmap.
1911 pmap_reference(pmap_t pmap)
1918 /***************************************************
1919 * page management routines.
1920 ***************************************************/
1923 * free the pv_entry back to the free list. This function may be
1924 * called from an interrupt.
1928 free_pv_entry(pv_entry_t pv)
1931 KKASSERT(pv_entry_count >= 0);
1936 * get a new pv_entry, allocating a block from the system
1937 * when needed. This function may be called from an interrupt.
1944 if (pv_entry_high_water &&
1945 (pv_entry_count > pv_entry_high_water) &&
1946 (pmap_pagedaemon_waken == 0)) {
1947 pmap_pagedaemon_waken = 1;
1948 wakeup(&vm_pages_needed);
1950 return zalloc(pvzone);
1954 * This routine is very drastic, but can save the system
1962 static int warningdone=0;
1964 if (pmap_pagedaemon_waken == 0)
1967 if (warningdone < 5) {
1968 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1972 for(i = 0; i < vm_page_array_size; i++) {
1973 m = &vm_page_array[i];
1974 if (m->wire_count || m->hold_count || m->busy ||
1975 (m->flags & PG_BUSY))
1979 pmap_pagedaemon_waken = 0;
1984 * If it is the first entry on the list, it is actually
1985 * in the header and we must copy the following entry up
1986 * to the header. Otherwise we must search the list for
1987 * the entry. In either case we free the now unused entry.
1991 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1992 vm_offset_t va, pmap_inval_info_t info)
1998 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1999 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2000 if (pmap == pv->pv_pmap && va == pv->pv_va)
2004 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2005 if (va == pv->pv_va)
2013 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2014 m->md.pv_list_count--;
2015 KKASSERT(m->md.pv_list_count >= 0);
2016 if (TAILQ_EMPTY(&m->md.pv_list))
2017 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2018 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2019 ++pmap->pm_generation;
2020 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
2028 * Create a pv entry for page at pa for
2033 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2038 pv = get_pv_entry();
2043 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2044 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2045 ++pmap->pm_generation;
2046 m->md.pv_list_count++;
2052 * pmap_remove_pte: do the things to unmap a page in a process
2056 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2057 pmap_inval_info_t info)
2062 pmap_inval_add(info, pmap, va);
2063 oldpte = pte_load_clear(ptq);
2065 pmap->pm_stats.wired_count -= 1;
2067 * Machines that don't support invlpg, also don't support
2068 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2072 cpu_invlpg((void *)va);
2073 KKASSERT(pmap->pm_stats.resident_count > 0);
2074 --pmap->pm_stats.resident_count;
2075 if (oldpte & PG_MANAGED) {
2076 m = PHYS_TO_VM_PAGE(oldpte);
2077 if (oldpte & PG_M) {
2078 #if defined(PMAP_DIAGNOSTIC)
2079 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2081 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2085 if (pmap_track_modified(va))
2089 vm_page_flag_set(m, PG_REFERENCED);
2090 return pmap_remove_entry(pmap, m, va, info);
2092 return pmap_unuse_pt(pmap, va, NULL, info);
2101 * Remove a single page from a process address space.
2103 * This function may not be called from an interrupt if the pmap is
2108 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2112 pte = pmap_pte(pmap, va);
2115 if ((*pte & PG_V) == 0)
2117 pmap_remove_pte(pmap, pte, va, info);
2123 * Remove the given range of addresses from the specified map.
2125 * It is assumed that the start and end are properly
2126 * rounded to the page size.
2128 * This function may not be called from an interrupt if the pmap is
2132 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2134 vm_offset_t va_next;
2135 pml4_entry_t *pml4e;
2137 pd_entry_t ptpaddr, *pde;
2139 struct pmap_inval_info info;
2144 if (pmap->pm_stats.resident_count == 0)
2147 pmap_inval_init(&info);
2150 * special handling of removing one page. a very
2151 * common operation and easy to short circuit some
2154 if (sva + PAGE_SIZE == eva) {
2155 pde = pmap_pde(pmap, sva);
2156 if (pde && (*pde & PG_PS) == 0) {
2157 pmap_remove_page(pmap, sva, &info);
2158 pmap_inval_flush(&info);
2163 for (; sva < eva; sva = va_next) {
2164 pml4e = pmap_pml4e(pmap, sva);
2165 if ((*pml4e & PG_V) == 0) {
2166 va_next = (sva + NBPML4) & ~PML4MASK;
2172 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2173 if ((*pdpe & PG_V) == 0) {
2174 va_next = (sva + NBPDP) & ~PDPMASK;
2181 * Calculate index for next page table.
2183 va_next = (sva + NBPDR) & ~PDRMASK;
2187 pde = pmap_pdpe_to_pde(pdpe, sva);
2191 * Weed out invalid mappings.
2197 * Check for large page.
2199 if ((ptpaddr & PG_PS) != 0) {
2200 /* JG FreeBSD has more complex treatment here */
2201 pmap_inval_add(&info, pmap, -1);
2203 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2208 * Limit our scan to either the end of the va represented
2209 * by the current page table page, or to the end of the
2210 * range being removed.
2216 * NOTE: pmap_remove_pte() can block.
2218 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2222 if (pmap_remove_pte(pmap, pte, sva, &info))
2226 pmap_inval_flush(&info);
2232 * Removes this physical page from all physical maps in which it resides.
2233 * Reflects back modify bits to the pager.
2235 * This routine may not be called from an interrupt.
2240 pmap_remove_all(vm_page_t m)
2242 struct pmap_inval_info info;
2243 pt_entry_t *pte, tpte;
2246 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2249 pmap_inval_init(&info);
2251 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2252 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2253 --pv->pv_pmap->pm_stats.resident_count;
2255 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2256 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2257 tpte = pte_load_clear(pte);
2260 pv->pv_pmap->pm_stats.wired_count--;
2263 vm_page_flag_set(m, PG_REFERENCED);
2266 * Update the vm_page_t clean and reference bits.
2269 #if defined(PMAP_DIAGNOSTIC)
2270 if (pmap_nw_modified(tpte)) {
2272 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2276 if (pmap_track_modified(pv->pv_va))
2279 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2280 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2281 ++pv->pv_pmap->pm_generation;
2282 m->md.pv_list_count--;
2283 KKASSERT(m->md.pv_list_count >= 0);
2284 if (TAILQ_EMPTY(&m->md.pv_list))
2285 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2286 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2290 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2291 pmap_inval_flush(&info);
2297 * Set the physical protection on the specified range of this map
2300 * This function may not be called from an interrupt if the map is
2301 * not the kernel_pmap.
2304 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2306 vm_offset_t va_next;
2307 pml4_entry_t *pml4e;
2309 pd_entry_t ptpaddr, *pde;
2311 pmap_inval_info info;
2313 /* JG review for NX */
2318 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2319 pmap_remove(pmap, sva, eva);
2323 if (prot & VM_PROT_WRITE)
2326 pmap_inval_init(&info);
2328 for (; sva < eva; sva = va_next) {
2330 pml4e = pmap_pml4e(pmap, sva);
2331 if ((*pml4e & PG_V) == 0) {
2332 va_next = (sva + NBPML4) & ~PML4MASK;
2338 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2339 if ((*pdpe & PG_V) == 0) {
2340 va_next = (sva + NBPDP) & ~PDPMASK;
2346 va_next = (sva + NBPDR) & ~PDRMASK;
2350 pde = pmap_pdpe_to_pde(pdpe, sva);
2354 * Check for large page.
2356 if ((ptpaddr & PG_PS) != 0) {
2357 pmap_inval_add(&info, pmap, -1);
2358 *pde &= ~(PG_M|PG_RW);
2359 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2364 * Weed out invalid mappings. Note: we assume that the page
2365 * directory table is always allocated, and in kernel virtual.
2373 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2375 pt_entry_t obits, pbits;
2379 * XXX non-optimal. Note also that there can be
2380 * no pmap_inval_flush() calls until after we modify
2381 * ptbase[sindex] (or otherwise we have to do another
2382 * pmap_inval_add() call).
2384 pmap_inval_add(&info, pmap, sva);
2385 obits = pbits = *pte;
2386 if ((pbits & PG_V) == 0)
2388 if (pbits & PG_MANAGED) {
2391 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2392 vm_page_flag_set(m, PG_REFERENCED);
2396 if (pmap_track_modified(sva)) {
2398 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2407 if (pbits != obits) {
2412 pmap_inval_flush(&info);
2416 * Insert the given physical page (p) at
2417 * the specified virtual address (v) in the
2418 * target physical map with the protection requested.
2420 * If specified, the page will be wired down, meaning
2421 * that the related pte can not be reclaimed.
2423 * NB: This is the only routine which MAY NOT lazy-evaluate
2424 * or lose information. That is, this routine must actually
2425 * insert this page into the given map NOW.
2428 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2435 pt_entry_t origpte, newpte;
2437 pmap_inval_info info;
2442 va = trunc_page(va);
2443 #ifdef PMAP_DIAGNOSTIC
2445 panic("pmap_enter: toobig");
2446 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2447 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
2449 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2450 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2452 db_print_backtrace();
2455 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2456 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2458 db_print_backtrace();
2463 * In the case that a page table page is not
2464 * resident, we are creating it here.
2466 if (va < VM_MAX_USER_ADDRESS)
2467 mpte = pmap_allocpte(pmap, va);
2471 pmap_inval_init(&info);
2472 pde = pmap_pde(pmap, va);
2473 if (pde != NULL && (*pde & PG_V) != 0) {
2474 if ((*pde & PG_PS) != 0)
2475 panic("pmap_enter: attempted pmap_enter on 2MB page");
2476 pte = pmap_pde_to_pte(pde, va);
2478 panic("pmap_enter: invalid page directory va=%#lx", va);
2480 KKASSERT(pte != NULL);
2481 pa = VM_PAGE_TO_PHYS(m);
2483 opa = origpte & PG_FRAME;
2486 * Mapping has not changed, must be protection or wiring change.
2488 if (origpte && (opa == pa)) {
2490 * Wiring change, just update stats. We don't worry about
2491 * wiring PT pages as they remain resident as long as there
2492 * are valid mappings in them. Hence, if a user page is wired,
2493 * the PT page will be also.
2495 if (wired && ((origpte & PG_W) == 0))
2496 pmap->pm_stats.wired_count++;
2497 else if (!wired && (origpte & PG_W))
2498 pmap->pm_stats.wired_count--;
2500 #if defined(PMAP_DIAGNOSTIC)
2501 if (pmap_nw_modified(origpte)) {
2503 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2509 * Remove the extra pte reference. Note that we cannot
2510 * optimize the RO->RW case because we have adjusted the
2511 * wiring count above and may need to adjust the wiring
2518 * We might be turning off write access to the page,
2519 * so we go ahead and sense modify status.
2521 if (origpte & PG_MANAGED) {
2522 if ((origpte & PG_M) && pmap_track_modified(va)) {
2524 om = PHYS_TO_VM_PAGE(opa);
2528 KKASSERT(m->flags & PG_MAPPED);
2533 * Mapping has changed, invalidate old range and fall through to
2534 * handle validating new mapping.
2538 err = pmap_remove_pte(pmap, pte, va, &info);
2540 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2542 opa = origpte & PG_FRAME;
2544 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2550 * Enter on the PV list if part of our managed memory. Note that we
2551 * raise IPL while manipulating pv_table since pmap_enter can be
2552 * called at interrupt time.
2554 if (pmap_initialized &&
2555 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2556 pmap_insert_entry(pmap, va, mpte, m);
2558 vm_page_flag_set(m, PG_MAPPED);
2562 * Increment counters
2564 ++pmap->pm_stats.resident_count;
2566 pmap->pm_stats.wired_count++;
2570 * Now validate mapping with desired protection/wiring.
2572 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
2576 if (va < VM_MAX_USER_ADDRESS)
2578 if (pmap == &kernel_pmap)
2582 * if the mapping or permission bits are different, we need
2583 * to update the pte.
2585 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2586 pmap_inval_add(&info, pmap, va);
2587 *pte = newpte | PG_A;
2589 vm_page_flag_set(m, PG_WRITEABLE);
2591 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2592 pmap_inval_flush(&info);
2596 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2597 * This code also assumes that the pmap has no pre-existing entry for this
2600 * This code currently may only be used on user pmaps, not kernel_pmap.
2604 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2609 vm_pindex_t ptepindex;
2611 pmap_inval_info info;
2613 pmap_inval_init(&info);
2615 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2616 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2618 db_print_backtrace();
2621 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2622 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2624 db_print_backtrace();
2628 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2631 * Calculate the page table page (mpte), allocating it if necessary.
2633 * A held page table page (mpte), or NULL, is passed onto the
2634 * section following.
2636 if (va < VM_MAX_USER_ADDRESS) {
2638 * Calculate pagetable page index
2640 ptepindex = pmap_pde_pindex(va);
2644 * Get the page directory entry
2646 ptepa = pmap_pde(pmap, va);
2649 * If the page table page is mapped, we just increment
2650 * the hold count, and activate it.
2652 if (ptepa && (*ptepa & PG_V) != 0) {
2654 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2655 // if (pmap->pm_ptphint &&
2656 // (pmap->pm_ptphint->pindex == ptepindex)) {
2657 // mpte = pmap->pm_ptphint;
2659 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2660 pmap->pm_ptphint = mpte;
2665 mpte = _pmap_allocpte(pmap, ptepindex);
2667 } while (mpte == NULL);
2670 /* this code path is not yet used */
2674 * With a valid (and held) page directory page, we can just use
2675 * vtopte() to get to the pte. If the pte is already present
2676 * we do not disturb it.
2681 pmap_unwire_pte_hold(pmap, va, mpte, &info);
2682 pa = VM_PAGE_TO_PHYS(m);
2683 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2688 * Enter on the PV list if part of our managed memory
2690 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2691 pmap_insert_entry(pmap, va, mpte, m);
2692 vm_page_flag_set(m, PG_MAPPED);
2696 * Increment counters
2698 ++pmap->pm_stats.resident_count;
2700 pa = VM_PAGE_TO_PHYS(m);
2703 * Now validate mapping with RO protection
2705 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2706 *pte = pa | PG_V | PG_U;
2708 *pte = pa | PG_V | PG_U | PG_MANAGED;
2709 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2710 pmap_inval_flush(&info);
2714 * Make a temporary mapping for a physical address. This is only intended
2715 * to be used for panic dumps.
2717 /* JG Needed on x86_64? */
2719 pmap_kenter_temporary(vm_paddr_t pa, int i)
2721 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2722 return ((void *)crashdumpmap);
2725 #define MAX_INIT_PT (96)
2728 * This routine preloads the ptes for a given object into the specified pmap.
2729 * This eliminates the blast of soft faults on process startup and
2730 * immediately after an mmap.
2732 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2735 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2736 vm_object_t object, vm_pindex_t pindex,
2737 vm_size_t size, int limit)
2739 struct rb_vm_page_scan_info info;
2744 * We can't preinit if read access isn't set or there is no pmap
2747 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2751 * We can't preinit if the pmap is not the current pmap
2753 lp = curthread->td_lwp;
2754 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2757 psize = x86_64_btop(size);
2759 if ((object->type != OBJT_VNODE) ||
2760 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2761 (object->resident_page_count > MAX_INIT_PT))) {
2765 if (psize + pindex > object->size) {
2766 if (object->size < pindex)
2768 psize = object->size - pindex;
2775 * Use a red-black scan to traverse the requested range and load
2776 * any valid pages found into the pmap.
2778 * We cannot safely scan the object's memq unless we are in a
2779 * critical section since interrupts can remove pages from objects.
2781 info.start_pindex = pindex;
2782 info.end_pindex = pindex + psize - 1;
2789 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2790 pmap_object_init_pt_callback, &info);
2796 pmap_object_init_pt_callback(vm_page_t p, void *data)
2798 struct rb_vm_page_scan_info *info = data;
2799 vm_pindex_t rel_index;
2801 * don't allow an madvise to blow away our really
2802 * free pages allocating pv entries.
2804 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2805 vmstats.v_free_count < vmstats.v_free_reserved) {
2808 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2809 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2810 if ((p->queue - p->pc) == PQ_CACHE)
2811 vm_page_deactivate(p);
2813 rel_index = p->pindex - info->start_pindex;
2814 pmap_enter_quick(info->pmap,
2815 info->addr + x86_64_ptob(rel_index), p);
2822 * pmap_prefault provides a quick way of clustering pagefaults into a
2823 * processes address space. It is a "cousin" of pmap_object_init_pt,
2824 * except it runs at page fault time instead of mmap time.
2828 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2830 static int pmap_prefault_pageorder[] = {
2831 -PAGE_SIZE, PAGE_SIZE,
2832 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2833 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2834 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2838 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2849 * We do not currently prefault mappings that use virtual page
2850 * tables. We do not prefault foreign pmaps.
2852 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2854 lp = curthread->td_lwp;
2855 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2858 object = entry->object.vm_object;
2860 starta = addra - PFBAK * PAGE_SIZE;
2861 if (starta < entry->start)
2862 starta = entry->start;
2863 else if (starta > addra)
2867 * critical section protection is required to maintain the
2868 * page/object association, interrupts can free pages and remove
2869 * them from their objects.
2872 for (i = 0; i < PAGEORDER_SIZE; i++) {
2873 vm_object_t lobject;
2877 addr = addra + pmap_prefault_pageorder[i];
2878 if (addr > addra + (PFFOR * PAGE_SIZE))
2881 if (addr < starta || addr >= entry->end)
2884 pde = pmap_pde(pmap, addr);
2885 if (pde == NULL || *pde == 0)
2892 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2895 for (m = vm_page_lookup(lobject, pindex);
2896 (!m && (lobject->type == OBJT_DEFAULT) &&
2897 (lobject->backing_object));
2898 lobject = lobject->backing_object
2900 if (lobject->backing_object_offset & PAGE_MASK)
2902 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2903 m = vm_page_lookup(lobject->backing_object, pindex);
2907 * give-up when a page is not in memory
2912 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2914 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2916 if ((m->queue - m->pc) == PQ_CACHE) {
2917 vm_page_deactivate(m);
2920 pmap_enter_quick(pmap, addr, m);
2928 * Routine: pmap_change_wiring
2929 * Function: Change the wiring attribute for a map/virtual-address
2931 * In/out conditions:
2932 * The mapping must already exist in the pmap.
2935 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2942 pte = pmap_pte(pmap, va);
2944 if (wired && !pmap_pte_w(pte))
2945 pmap->pm_stats.wired_count++;
2946 else if (!wired && pmap_pte_w(pte))
2947 pmap->pm_stats.wired_count--;
2950 * Wiring is not a hardware characteristic so there is no need to
2951 * invalidate TLB. However, in an SMP environment we must use
2952 * a locked bus cycle to update the pte (if we are not using
2953 * the pmap_inval_*() API that is)... it's ok to do this for simple
2958 atomic_set_long(pte, PG_W);
2960 atomic_clear_long(pte, PG_W);
2963 atomic_set_long_nonlocked(pte, PG_W);
2965 atomic_clear_long_nonlocked(pte, PG_W);
2972 * Copy the range specified by src_addr/len
2973 * from the source map to the range dst_addr/len
2974 * in the destination map.
2976 * This routine is only advisory and need not do anything.
2979 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2980 vm_size_t len, vm_offset_t src_addr)
2984 pmap_inval_info info;
2986 vm_offset_t end_addr = src_addr + len;
2988 pd_entry_t src_frame, dst_frame;
2991 if (dst_addr != src_addr)
2994 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
2995 if (src_frame != (PTDpde & PG_FRAME)) {
2999 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3000 if (dst_frame != (APTDpde & PG_FRAME)) {
3001 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
3002 /* The page directory is not shared between CPUs */
3006 pmap_inval_init(&info);
3007 pmap_inval_add(&info, dst_pmap, -1);
3008 pmap_inval_add(&info, src_pmap, -1);
3011 * critical section protection is required to maintain the page/object
3012 * association, interrupts can free pages and remove them from
3016 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
3017 pt_entry_t *src_pte, *dst_pte;
3018 vm_page_t dstmpte, srcmpte;
3019 vm_offset_t srcptepaddr;
3020 vm_pindex_t ptepindex;
3022 if (addr >= UPT_MIN_ADDRESS)
3023 panic("pmap_copy: invalid to pmap_copy page tables\n");
3026 * Don't let optional prefaulting of pages make us go
3027 * way below the low water mark of free pages or way
3028 * above high water mark of used pv entries.
3030 if (vmstats.v_free_count < vmstats.v_free_reserved ||
3031 pv_entry_count > pv_entry_high_water)
3034 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
3035 ptepindex = addr >> PDRSHIFT;
3038 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
3040 if (srcptepaddr == 0)
3043 if (srcptepaddr & PG_PS) {
3045 if (dst_pmap->pm_pdir[ptepindex] == 0) {
3046 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
3047 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
3053 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
3054 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
3055 (srcmpte->flags & PG_BUSY)) {
3059 if (pdnxt > end_addr)
3062 src_pte = vtopte(addr);
3064 dst_pte = avtopte(addr);
3066 while (addr < pdnxt) {
3071 * we only virtual copy managed pages
3073 if ((ptetemp & PG_MANAGED) != 0) {
3075 * We have to check after allocpte for the
3076 * pte still being around... allocpte can
3079 * pmap_allocpte() can block. If we lose
3080 * our page directory mappings we stop.
3082 dstmpte = pmap_allocpte(dst_pmap, addr);
3085 if (src_frame != (PTDpde & PG_FRAME) ||
3086 dst_frame != (APTDpde & PG_FRAME)
3088 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3089 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3091 } else if ((*dst_pte == 0) &&
3092 (ptetemp = *src_pte) != 0 &&
3093 (ptetemp & PG_MANAGED)) {
3095 * Clear the modified and
3096 * accessed (referenced) bits
3099 m = PHYS_TO_VM_PAGE(ptetemp);
3100 *dst_pte = ptetemp & ~(PG_M | PG_A);
3101 ++dst_pmap->pm_stats.resident_count;
3102 pmap_insert_entry(dst_pmap, addr,
3104 KKASSERT(m->flags & PG_MAPPED);
3106 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3107 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3111 if (dstmpte->hold_count >= srcmpte->hold_count)
3121 pmap_inval_flush(&info);
3128 * Zero the specified physical page.
3130 * This function may be called from an interrupt and no locking is
3134 pmap_zero_page(vm_paddr_t phys)
3136 vm_offset_t va = PHYS_TO_DMAP(phys);
3138 pagezero((void *)va);
3142 * pmap_page_assertzero:
3144 * Assert that a page is empty, panic if it isn't.
3147 pmap_page_assertzero(vm_paddr_t phys)
3149 vm_offset_t virt = PHYS_TO_DMAP(phys);
3152 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3153 if (*(long *)((char *)virt + i) != 0) {
3154 panic("pmap_page_assertzero() @ %p not zero!\n", (void *)virt);
3162 * Zero part of a physical page by mapping it into memory and clearing
3163 * its contents with bzero.
3165 * off and size may not cover an area beyond a single hardware page.
3168 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3170 vm_offset_t virt = PHYS_TO_DMAP(phys);
3172 bzero((char *)virt + off, size);
3178 * Copy the physical page from the source PA to the target PA.
3179 * This function may be called from an interrupt. No locking
3183 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3185 vm_offset_t src_virt, dst_virt;
3187 src_virt = PHYS_TO_DMAP(src);
3188 dst_virt = PHYS_TO_DMAP(dst);
3189 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3193 * pmap_copy_page_frag:
3195 * Copy the physical page from the source PA to the target PA.
3196 * This function may be called from an interrupt. No locking
3200 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3202 vm_offset_t src_virt, dst_virt;
3204 src_virt = PHYS_TO_DMAP(src);
3205 dst_virt = PHYS_TO_DMAP(dst);
3207 bcopy((char *)src_virt + (src & PAGE_MASK),
3208 (char *)dst_virt + (dst & PAGE_MASK),
3213 * Returns true if the pmap's pv is one of the first
3214 * 16 pvs linked to from this page. This count may
3215 * be changed upwards or downwards in the future; it
3216 * is only necessary that true be returned for a small
3217 * subset of pmaps for proper page aging.
3220 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3225 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3230 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3231 if (pv->pv_pmap == pmap) {
3244 * Remove all pages from specified address space
3245 * this aids process exit speeds. Also, this code
3246 * is special cased for current process only, but
3247 * can have the more generic (and slightly slower)
3248 * mode enabled. This is much faster than pmap_remove
3249 * in the case of running down an entire address space.
3252 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3255 pt_entry_t *pte, tpte;
3258 pmap_inval_info info;
3260 int save_generation;
3262 lp = curthread->td_lwp;
3263 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3268 pmap_inval_init(&info);
3270 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3271 if (pv->pv_va >= eva || pv->pv_va < sva) {
3272 npv = TAILQ_NEXT(pv, pv_plist);
3276 KKASSERT(pmap == pv->pv_pmap);
3279 pte = vtopte(pv->pv_va);
3281 pte = pmap_pte_quick(pmap, pv->pv_va);
3282 if (pmap->pm_active)
3283 pmap_inval_add(&info, pmap, pv->pv_va);
3286 * We cannot remove wired pages from a process' mapping
3290 npv = TAILQ_NEXT(pv, pv_plist);
3293 tpte = pte_load_clear(pte);
3295 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3297 KASSERT(m < &vm_page_array[vm_page_array_size],
3298 ("pmap_remove_pages: bad tpte %lx", tpte));
3300 KKASSERT(pmap->pm_stats.resident_count > 0);
3301 --pmap->pm_stats.resident_count;
3304 * Update the vm_page_t clean and reference bits.
3310 npv = TAILQ_NEXT(pv, pv_plist);
3311 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3312 save_generation = ++pmap->pm_generation;
3314 m->md.pv_list_count--;
3315 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3316 if (TAILQ_EMPTY(&m->md.pv_list))
3317 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3319 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3323 * Restart the scan if we blocked during the unuse or free
3324 * calls and other removals were made.
3326 if (save_generation != pmap->pm_generation) {
3327 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3328 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3331 pmap_inval_flush(&info);
3336 * pmap_testbit tests bits in pte's
3337 * note that the testbit/clearbit routines are inline,
3338 * and a lot of things compile-time evaluate.
3342 pmap_testbit(vm_page_t m, int bit)
3347 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3350 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3355 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3357 * if the bit being tested is the modified bit, then
3358 * mark clean_map and ptes as never
3361 if (bit & (PG_A|PG_M)) {
3362 if (!pmap_track_modified(pv->pv_va))
3366 #if defined(PMAP_DIAGNOSTIC)
3367 if (pv->pv_pmap == NULL) {
3368 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3372 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3383 * this routine is used to modify bits in ptes
3387 pmap_clearbit(vm_page_t m, int bit)
3389 struct pmap_inval_info info;
3394 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3397 pmap_inval_init(&info);
3401 * Loop over all current mappings setting/clearing as appropos If
3402 * setting RO do we need to clear the VAC?
3404 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3406 * don't write protect pager mappings
3409 if (!pmap_track_modified(pv->pv_va))
3413 #if defined(PMAP_DIAGNOSTIC)
3414 if (pv->pv_pmap == NULL) {
3415 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3421 * Careful here. We can use a locked bus instruction to
3422 * clear PG_A or PG_M safely but we need to synchronize
3423 * with the target cpus when we mess with PG_RW.
3425 * We do not have to force synchronization when clearing
3426 * PG_M even for PTEs generated via virtual memory maps,
3427 * because the virtual kernel will invalidate the pmap
3428 * entry when/if it needs to resynchronize the Modify bit.
3431 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
3432 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3439 atomic_clear_long(pte, PG_M|PG_RW);
3442 * The cpu may be trying to set PG_M
3443 * simultaniously with our clearing
3446 if (!atomic_cmpset_long(pte, pbits,
3450 } else if (bit == PG_M) {
3452 * We could also clear PG_RW here to force
3453 * a fault on write to redetect PG_M for
3454 * virtual kernels, but it isn't necessary
3455 * since virtual kernels invalidate the pte
3456 * when they clear the VPTE_M bit in their
3457 * virtual page tables.
3459 atomic_clear_long(pte, PG_M);
3461 atomic_clear_long(pte, bit);
3465 pmap_inval_flush(&info);
3470 * pmap_page_protect:
3472 * Lower the permission for all mappings to a given page.
3475 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3477 /* JG NX support? */
3478 if ((prot & VM_PROT_WRITE) == 0) {
3479 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3480 pmap_clearbit(m, PG_RW);
3481 vm_page_flag_clear(m, PG_WRITEABLE);
3489 pmap_phys_address(vm_pindex_t ppn)
3491 return (x86_64_ptob(ppn));
3495 * pmap_ts_referenced:
3497 * Return a count of reference bits for a page, clearing those bits.
3498 * It is not necessary for every reference bit to be cleared, but it
3499 * is necessary that 0 only be returned when there are truly no
3500 * reference bits set.
3502 * XXX: The exact number of bits to check and clear is a matter that
3503 * should be tested and standardized at some point in the future for
3504 * optimal aging of shared pages.
3507 pmap_ts_referenced(vm_page_t m)
3509 pv_entry_t pv, pvf, pvn;
3513 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3518 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3523 pvn = TAILQ_NEXT(pv, pv_list);
3526 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3527 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3530 if (!pmap_track_modified(pv->pv_va))
3533 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3535 if (pte && (*pte & PG_A)) {
3537 atomic_clear_long(pte, PG_A);
3539 atomic_clear_long_nonlocked(pte, PG_A);
3546 } while ((pv = pvn) != NULL && pv != pvf);
3556 * Return whether or not the specified physical page was modified
3557 * in any physical maps.
3560 pmap_is_modified(vm_page_t m)
3562 return pmap_testbit(m, PG_M);
3566 * Clear the modify bits on the specified physical page.
3569 pmap_clear_modify(vm_page_t m)
3571 pmap_clearbit(m, PG_M);
3575 * pmap_clear_reference:
3577 * Clear the reference bit on the specified physical page.
3580 pmap_clear_reference(vm_page_t m)
3582 pmap_clearbit(m, PG_A);
3586 * Miscellaneous support routines follow
3591 i386_protection_init(void)
3595 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3596 kp = protection_codes;
3597 for (prot = 0; prot < 8; prot++) {
3599 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3601 * Read access is also 0. There isn't any execute bit,
3602 * so just make it readable.
3604 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3605 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3606 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3609 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3610 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3611 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3612 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3620 * Map a set of physical memory pages into the kernel virtual
3621 * address space. Return a pointer to where it is mapped. This
3622 * routine is intended to be used for mapping device memory,
3625 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3629 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3631 vm_offset_t va, tmpva, offset;
3634 offset = pa & PAGE_MASK;
3635 size = roundup(offset + size, PAGE_SIZE);
3637 va = kmem_alloc_nofault(&kernel_map, size);
3639 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3641 pa = pa & ~PAGE_MASK;
3642 for (tmpva = va; size > 0;) {
3643 pte = vtopte(tmpva);
3644 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3652 return ((void *)(va + offset));
3656 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3658 vm_offset_t va, tmpva, offset;
3661 offset = pa & PAGE_MASK;
3662 size = roundup(offset + size, PAGE_SIZE);
3664 va = kmem_alloc_nofault(&kernel_map, size);
3666 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3668 pa = pa & ~PAGE_MASK;
3669 for (tmpva = va; size > 0;) {
3670 pte = vtopte(tmpva);
3671 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3679 return ((void *)(va + offset));
3683 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3685 vm_offset_t base, offset;
3687 base = va & ~PAGE_MASK;
3688 offset = va & PAGE_MASK;
3689 size = roundup(offset + size, PAGE_SIZE);
3690 pmap_qremove(va, size >> PAGE_SHIFT);
3691 kmem_free(&kernel_map, base, size);
3695 * perform the pmap work for mincore
3698 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3700 pt_entry_t *ptep, pte;
3704 ptep = pmap_pte(pmap, addr);
3709 if ((pte = *ptep) != 0) {
3712 val = MINCORE_INCORE;
3713 if ((pte & PG_MANAGED) == 0)
3716 pa = pte & PG_FRAME;
3718 m = PHYS_TO_VM_PAGE(pa);
3724 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3726 * Modified by someone
3728 else if (m->dirty || pmap_is_modified(m))
3729 val |= MINCORE_MODIFIED_OTHER;
3734 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3737 * Referenced by someone
3739 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3740 val |= MINCORE_REFERENCED_OTHER;
3741 vm_page_flag_set(m, PG_REFERENCED);
3748 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3749 * vmspace will be ref'd and the old one will be deref'd.
3751 * The vmspace for all lwps associated with the process will be adjusted
3752 * and cr3 will be reloaded if any lwp is the current lwp.
3755 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3757 struct vmspace *oldvm;
3761 oldvm = p->p_vmspace;
3762 if (oldvm != newvm) {
3763 p->p_vmspace = newvm;
3764 KKASSERT(p->p_nthreads == 1);
3765 lp = RB_ROOT(&p->p_lwp_tree);
3766 pmap_setlwpvm(lp, newvm);
3768 sysref_get(&newvm->vm_sysref);
3769 sysref_put(&oldvm->vm_sysref);
3776 * Set the vmspace for a LWP. The vmspace is almost universally set the
3777 * same as the process vmspace, but virtual kernels need to swap out contexts
3778 * on a per-lwp basis.
3781 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3783 struct vmspace *oldvm;
3787 oldvm = lp->lwp_vmspace;
3789 if (oldvm != newvm) {
3790 lp->lwp_vmspace = newvm;
3791 if (curthread->td_lwp == lp) {
3792 pmap = vmspace_pmap(newvm);
3794 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3796 pmap->pm_active |= 1;
3798 #if defined(SWTCH_OPTIM_STATS)
3801 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
3802 load_cr3(curthread->td_pcb->pcb_cr3);
3803 pmap = vmspace_pmap(oldvm);
3805 atomic_clear_int(&pmap->pm_active,
3806 1 << mycpu->gd_cpuid);
3808 pmap->pm_active &= ~1;
3816 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3819 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3823 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3830 static void pads (pmap_t pm);
3831 void pmap_pvdump (vm_paddr_t pa);
3833 /* print address space of pmap*/
3842 if (pm == &kernel_pmap)
3845 for (i = 0; i < NPDEPG; i++) {
3853 pmap_pvdump(vm_paddr_t pa)
3858 kprintf("pa %08llx", (long long)pa);
3859 m = PHYS_TO_VM_PAGE(pa);
3860 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3862 kprintf(" -> pmap %p, va %x, flags %x",
3863 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3865 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);