4 * Copyright (c) 1991 Regents of the University of California.
5 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
7 * Copyright (c) 2003 Peter Wemm
8 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
9 * Copyright (c) 2008, 2009 The DragonFly Project.
10 * Copyright (c) 2008, 2009 Jordan Gordeev.
11 * All rights reserved.
13 * This code is derived from software contributed to Berkeley by
14 * the Systems Programming Group of the University of Utah Computer
15 * Science Department and William Jolitz of UUNET Technologies Inc.
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions
20 * 1. Redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer.
22 * 2. Redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution.
25 * 3. All advertising materials mentioning features or use of this software
26 * must display the following acknowledgement:
27 * This product includes software developed by the University of
28 * California, Berkeley and its contributors.
29 * 4. Neither the name of the University nor the names of its contributors
30 * may be used to endorse or promote products derived from this software
31 * without specific prior written permission.
33 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
34 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
35 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
36 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
37 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
38 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
39 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
40 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
41 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
42 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
45 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
46 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
50 * Manages physical address maps.
52 * In addition to hardware address maps, this
53 * module is called upon to provide software-use-only
54 * maps which may or may not be stored in the same
55 * form as hardware maps. These pseudo-maps are
56 * used to store intermediate results from copy
57 * operations to and from address spaces.
59 * Since the information managed by this module is
60 * also stored by the logical address mapping module,
61 * this module may throw away valid virtual-to-physical
62 * mappings at almost any time. However, invalidations
63 * of virtual-to-physical mappings must be done as
66 * In order to cope with hardware architectures which
67 * make virtual-to-physical map invalidates expensive,
68 * this module may delay invalidate or reduced protection
69 * operations until such time as they are actually
70 * necessary. This module is given full information as
71 * to which processors are currently using which maps,
72 * and to when physical maps must be made correct.
76 #include "opt_disable_pse.h"
79 #include "opt_msgbuf.h"
81 #include <sys/param.h>
82 #include <sys/systm.h>
83 #include <sys/kernel.h>
85 #include <sys/msgbuf.h>
86 #include <sys/vmmeter.h>
90 #include <vm/vm_param.h>
91 #include <sys/sysctl.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_extern.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pager.h>
100 #include <vm/vm_zone.h>
102 #include <sys/user.h>
103 #include <sys/thread2.h>
104 #include <sys/sysref2.h>
106 #include <machine/cputypes.h>
107 #include <machine/md_var.h>
108 #include <machine/specialreg.h>
109 #include <machine/smp.h>
110 #include <machine_base/apic/apicreg.h>
111 #include <machine/globaldata.h>
112 #include <machine/pmap.h>
113 #include <machine/pmap_inval.h>
117 #define PMAP_KEEP_PDIRS
118 #ifndef PMAP_SHPGPERPROC
119 #define PMAP_SHPGPERPROC 200
122 #if defined(DIAGNOSTIC)
123 #define PMAP_DIAGNOSTIC
129 * Get PDEs and PTEs for user/kernel address space
131 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
132 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
134 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
135 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
136 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
137 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
138 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
142 * Given a map and a machine independent protection code,
143 * convert to a vax protection code.
145 #define pte_prot(m, p) \
146 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
147 static int protection_codes[8];
149 struct pmap kernel_pmap;
150 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
152 vm_paddr_t avail_start; /* PA of first available physical page */
153 vm_paddr_t avail_end; /* PA of last available physical page */
154 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
155 vm_offset_t virtual2_end;
156 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
157 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
158 vm_offset_t KvaStart; /* VA start of KVA space */
159 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
160 vm_offset_t KvaSize; /* max size of kernel virtual address space */
161 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
162 static int pgeflag; /* PG_G or-in */
163 static int pseflag; /* PG_PS or-in */
165 static vm_object_t kptobj;
168 static vm_paddr_t dmaplimit;
170 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
172 static uint64_t KPTbase;
173 static uint64_t KPTphys;
174 static uint64_t KPDphys; /* phys addr of kernel level 2 */
175 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
176 uint64_t KPDPphys; /* phys addr of kernel level 3 */
177 uint64_t KPML4phys; /* phys addr of kernel level 4 */
179 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
180 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
183 * Data for the pv entry allocation mechanism
185 static vm_zone_t pvzone;
186 static struct vm_zone pvzone_store;
187 static struct vm_object pvzone_obj;
188 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
189 static int pmap_pagedaemon_waken = 0;
190 static struct pv_entry *pvinit;
193 * All those kernel PT submaps that BSD is so fond of
195 pt_entry_t *CMAP1 = 0, *ptmmap;
196 caddr_t CADDR1 = 0, ptvmmap = 0;
197 static pt_entry_t *msgbufmap;
198 struct msgbuf *msgbufp=0;
203 static pt_entry_t *pt_crashdumpmap;
204 static caddr_t crashdumpmap;
206 extern pt_entry_t *SMPpt;
207 extern uint64_t SMPptpa;
211 static pv_entry_t get_pv_entry (void);
212 static void i386_protection_init (void);
213 static void create_pagetables(vm_paddr_t *firstaddr);
214 static void pmap_remove_all (vm_page_t m);
215 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
216 vm_offset_t sva, pmap_inval_info_t info);
217 static void pmap_remove_page (struct pmap *pmap,
218 vm_offset_t va, pmap_inval_info_t info);
219 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
220 vm_offset_t va, pmap_inval_info_t info);
221 static boolean_t pmap_testbit (vm_page_t m, int bit);
222 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
223 vm_page_t mpte, vm_page_t m);
225 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
227 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
228 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
229 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
230 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
231 static int _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
232 pmap_inval_info_t info);
233 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
234 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
236 static unsigned pdir4mb;
239 * Move the kernel virtual free pointer to the next
240 * 2MB. This is used to help improve performance
241 * by using a large (2MB) page for much of the kernel
242 * (.text, .data, .bss)
246 pmap_kmem_choose(vm_offset_t addr)
248 vm_offset_t newaddr = addr;
250 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
257 * Super fast pmap_pte routine best used when scanning the pv lists.
258 * This eliminates many course-grained invltlb calls. Note that many of
259 * the pv list scans are across different pmaps and it is very wasteful
260 * to do an entire invltlb when checking a single mapping.
262 * Should only be called while in a critical section.
264 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
268 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
270 return pmap_pte(pmap, va);
273 /* Return a non-clipped PD index for a given VA */
276 pmap_pde_pindex(vm_offset_t va)
278 return va >> PDRSHIFT;
281 /* Return various clipped indexes for a given VA */
284 pmap_pte_index(vm_offset_t va)
287 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
292 pmap_pde_index(vm_offset_t va)
295 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
300 pmap_pdpe_index(vm_offset_t va)
303 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
308 pmap_pml4e_index(vm_offset_t va)
311 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
314 /* Return a pointer to the PML4 slot that corresponds to a VA */
317 pmap_pml4e(pmap_t pmap, vm_offset_t va)
320 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
323 /* Return a pointer to the PDP slot that corresponds to a VA */
326 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
330 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
331 return (&pdpe[pmap_pdpe_index(va)]);
334 /* Return a pointer to the PDP slot that corresponds to a VA */
337 pmap_pdpe(pmap_t pmap, vm_offset_t va)
341 pml4e = pmap_pml4e(pmap, va);
342 if ((*pml4e & PG_V) == 0)
344 return (pmap_pml4e_to_pdpe(pml4e, va));
347 /* Return a pointer to the PD slot that corresponds to a VA */
350 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
354 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
355 return (&pde[pmap_pde_index(va)]);
358 /* Return a pointer to the PD slot that corresponds to a VA */
361 pmap_pde(pmap_t pmap, vm_offset_t va)
365 pdpe = pmap_pdpe(pmap, va);
366 if (pdpe == NULL || (*pdpe & PG_V) == 0)
368 return (pmap_pdpe_to_pde(pdpe, va));
371 /* Return a pointer to the PT slot that corresponds to a VA */
374 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
378 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
379 return (&pte[pmap_pte_index(va)]);
382 /* Return a pointer to the PT slot that corresponds to a VA */
385 pmap_pte(pmap_t pmap, vm_offset_t va)
389 pde = pmap_pde(pmap, va);
390 if (pde == NULL || (*pde & PG_V) == 0)
392 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
393 return ((pt_entry_t *)pde);
394 return (pmap_pde_to_pte(pde, va));
399 vtopte(vm_offset_t va)
401 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
403 return (PTmap + ((va >> PAGE_SHIFT) & mask));
408 vtopde(vm_offset_t va)
410 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
412 return (PDmap + ((va >> PDRSHIFT) & mask));
416 allocpages(vm_paddr_t *firstaddr, int n)
421 bzero((void *)ret, n * PAGE_SIZE);
422 *firstaddr += n * PAGE_SIZE;
428 create_pagetables(vm_paddr_t *firstaddr)
433 * We are running (mostly) V=P at this point
435 * Calculate NKPT - number of kernel page tables. We have to
436 * accomodoate prealloction of the vm_page_array, dump bitmap,
437 * MSGBUF_SIZE, and other stuff. Be generous.
439 * Maxmem is in pages.
441 nkpt = (Maxmem * (sizeof(struct vm_page) * 2) + MSGBUF_SIZE) / NBPDR;
446 KPTbase = allocpages(firstaddr, nkpt);
447 KPTphys = allocpages(firstaddr, nkpt);
448 KPML4phys = allocpages(firstaddr, 1);
449 KPDPphys = allocpages(firstaddr, NKPML4E);
452 * Calculate the page directory base for KERNBASE,
453 * that is where we start populating the page table pages.
454 * Basically this is the end - 2.
456 KPDphys = allocpages(firstaddr, NKPDPE);
457 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
459 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
460 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
462 DMPDPphys = allocpages(firstaddr, NDMPML4E);
463 if ((amd_feature & AMDID_PAGE1GB) == 0)
464 DMPDphys = allocpages(firstaddr, ndmpdp);
465 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
468 * Fill in the underlying page table pages for the area around
469 * KERNBASE. This remaps low physical memory to KERNBASE.
471 * Read-only from zero to physfree
472 * XXX not fully used, underneath 2M pages
474 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
475 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
476 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
480 * Now map the initial kernel page tables. One block of page
481 * tables is placed at the beginning of kernel virtual memory,
482 * and another block is placed at KERNBASE to map the kernel binary,
483 * data, bss, and initial pre-allocations.
485 for (i = 0; i < nkpt; i++) {
486 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
487 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
489 for (i = 0; i < nkpt; i++) {
490 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
491 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
495 * Map from zero to end of allocations using 2M pages as an
496 * optimization. This will bypass some of the KPTBase pages
497 * above in the KERNBASE area.
499 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
500 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
501 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
505 * And connect up the PD to the PDP. The kernel pmap is expected
506 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
508 for (i = 0; i < NKPDPE; i++) {
509 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
510 KPDphys + (i << PAGE_SHIFT);
511 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
515 /* Now set up the direct map space using either 2MB or 1GB pages */
516 /* Preset PG_M and PG_A because demotion expects it */
517 if ((amd_feature & AMDID_PAGE1GB) == 0) {
518 for (i = 0; i < NPDEPG * ndmpdp; i++) {
519 ((pd_entry_t *)DMPDphys)[i] = (vm_paddr_t)i << PDRSHIFT;
520 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
523 /* And the direct map space's PDP */
524 for (i = 0; i < ndmpdp; i++) {
525 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
527 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
530 for (i = 0; i < ndmpdp; i++) {
531 ((pdp_entry_t *)DMPDPphys)[i] =
532 (vm_paddr_t)i << PDPSHIFT;
533 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
538 /* And recursively map PML4 to itself in order to get PTmap */
539 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
540 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
542 /* Connect the Direct Map slot up to the PML4 */
543 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
544 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
546 /* Connect the KVA slot up to the PML4 */
547 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
548 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
552 * Bootstrap the system enough to run with virtual memory.
554 * On the i386 this is called after mapping has already been enabled
555 * and just syncs the pmap module with what has already been done.
556 * [We can't call it easily with mapping off since the kernel is not
557 * mapped with PA == VA, hence we would have to relocate every address
558 * from the linked base (virtual) address "KERNBASE" to the actual
559 * (physical) address starting relative to 0]
562 pmap_bootstrap(vm_paddr_t *firstaddr)
566 struct mdglobaldata *gd;
569 KvaStart = VM_MIN_KERNEL_ADDRESS;
570 KvaEnd = VM_MAX_KERNEL_ADDRESS;
571 KvaSize = KvaEnd - KvaStart;
573 avail_start = *firstaddr;
576 * Create an initial set of page tables to run the kernel in.
578 create_pagetables(firstaddr);
580 virtual2_start = KvaStart;
581 virtual2_end = PTOV_OFFSET;
583 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
584 virtual_start = pmap_kmem_choose(virtual_start);
586 virtual_end = VM_MAX_KERNEL_ADDRESS;
588 /* XXX do %cr0 as well */
589 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
593 * Initialize protection array.
595 i386_protection_init();
598 * The kernel's pmap is statically allocated so we don't have to use
599 * pmap_create, which is unlikely to work correctly at this part of
600 * the boot sequence (XXX and which no longer exists).
602 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
603 kernel_pmap.pm_count = 1;
604 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
605 TAILQ_INIT(&kernel_pmap.pm_pvlist);
608 * Reserve some special page table entries/VA space for temporary
611 #define SYSMAP(c, p, v, n) \
612 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
618 * CMAP1/CMAP2 are used for zeroing and copying pages.
620 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
625 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
628 * ptvmmap is used for reading arbitrary physical pages via
631 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
634 * msgbufp is used to map the system message buffer.
635 * XXX msgbufmap is not used.
637 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
638 atop(round_page(MSGBUF_SIZE)))
645 * PG_G is terribly broken on SMP because we IPI invltlb's in some
646 * cases rather then invl1pg. Actually, I don't even know why it
647 * works under UP because self-referential page table mappings
652 if (cpu_feature & CPUID_PGE)
657 * Initialize the 4MB page size flag
661 * The 4MB page version of the initial
662 * kernel page mapping.
666 #if !defined(DISABLE_PSE)
667 if (cpu_feature & CPUID_PSE) {
670 * Note that we have enabled PSE mode
673 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
674 ptditmp &= ~(NBPDR - 1);
675 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
680 * Enable the PSE mode. If we are SMP we can't do this
681 * now because the APs will not be able to use it when
684 load_cr4(rcr4() | CR4_PSE);
687 * We can do the mapping here for the single processor
688 * case. We simply ignore the old page table page from
692 * For SMP, we still need 4K pages to bootstrap APs,
693 * PSE will be enabled as soon as all APs are up.
695 PTD[KPTDI] = (pd_entry_t)ptditmp;
702 * We need to finish setting up the globaldata page for the BSP.
703 * locore has already populated the page table for the mdglobaldata
706 pg = MDGLOBALDATA_BASEALLOC_PAGES;
707 gd = &CPU_prvspace[0].mdglobaldata;
708 gd->gd_CMAP1 = &SMPpt[pg + 0];
709 gd->gd_CMAP2 = &SMPpt[pg + 1];
710 gd->gd_CMAP3 = &SMPpt[pg + 2];
711 gd->gd_PMAP1 = &SMPpt[pg + 3];
712 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
713 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
714 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
715 gd->gd_PADDR1 = (pt_entry_t *)CPU_prvspace[0].PPAGE1;
722 * Set 4mb pdir for mp startup
727 if (pseflag && (cpu_feature & CPUID_PSE)) {
728 load_cr4(rcr4() | CR4_PSE);
729 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
737 * Initialize the pmap module.
738 * Called by vm_init, to initialize any structures that the pmap
739 * system needs to map virtual memory.
740 * pmap_init has been enhanced to support in a fairly consistant
741 * way, discontiguous physical memory.
750 * object for kernel page table pages
752 /* JG I think the number can be arbitrary */
753 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
756 * Allocate memory for random pmap data structures. Includes the
760 for(i = 0; i < vm_page_array_size; i++) {
763 m = &vm_page_array[i];
764 TAILQ_INIT(&m->md.pv_list);
765 m->md.pv_list_count = 0;
769 * init the pv free list
771 initial_pvs = vm_page_array_size;
772 if (initial_pvs < MINPV)
774 pvzone = &pvzone_store;
775 pvinit = (void *)kmem_alloc(&kernel_map,
776 initial_pvs * sizeof (struct pv_entry));
777 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
778 pvinit, initial_pvs);
781 * Now it is safe to enable pv_table recording.
783 pmap_initialized = TRUE;
785 lapic = pmap_mapdev_uncacheable(cpu_apic_address, sizeof(struct LAPIC));
790 * Initialize the address space (zone) for the pv_entries. Set a
791 * high water mark so that the system can recover from excessive
792 * numbers of pv entries.
797 int shpgperproc = PMAP_SHPGPERPROC;
800 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
801 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
802 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
803 pv_entry_high_water = 9 * (pv_entry_max / 10);
806 * Subtract out pages already installed in the zone (hack)
808 entry_max = pv_entry_max - vm_page_array_size;
812 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
816 /***************************************************
817 * Low level helper routines.....
818 ***************************************************/
820 #if defined(PMAP_DIAGNOSTIC)
823 * This code checks for non-writeable/modified pages.
824 * This should be an invalid condition.
828 pmap_nw_modified(pt_entry_t pte)
830 if ((pte & (PG_M|PG_RW)) == PG_M)
839 * this routine defines the region(s) of memory that should
840 * not be tested for the modified bit.
844 pmap_track_modified(vm_offset_t va)
846 if ((va < clean_sva) || (va >= clean_eva))
853 * Extract the physical page address associated with the map/VA pair.
855 * The caller must hold vm_token if non-blocking operation is desired.
858 pmap_extract(pmap_t pmap, vm_offset_t va)
862 pd_entry_t pde, *pdep;
864 lwkt_gettoken(&vm_token);
866 pdep = pmap_pde(pmap, va);
870 if ((pde & PG_PS) != 0) {
871 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
873 pte = pmap_pde_to_pte(pdep, va);
874 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
878 lwkt_reltoken(&vm_token);
883 * Extract the physical page address associated kernel virtual address.
886 pmap_kextract(vm_offset_t va)
891 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
892 pa = DMAP_TO_PHYS(va);
896 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
899 * Beware of a concurrent promotion that changes the
900 * PDE at this point! For example, vtopte() must not
901 * be used to access the PTE because it would use the
902 * new PDE. It is, however, safe to use the old PDE
903 * because the page table page is preserved by the
906 pa = *pmap_pde_to_pte(&pde, va);
907 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
913 /***************************************************
914 * Low level mapping routines.....
915 ***************************************************/
918 * Routine: pmap_kenter
920 * Add a wired page to the KVA
921 * NOTE! note that in order for the mapping to take effect -- you
922 * should do an invltlb after doing the pmap_kenter().
925 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
929 pmap_inval_info info;
931 pmap_inval_init(&info);
932 npte = pa | PG_RW | PG_V | pgeflag;
934 pmap_inval_interlock(&info, &kernel_pmap, va);
936 pmap_inval_deinterlock(&info, &kernel_pmap);
937 pmap_inval_done(&info);
941 * Routine: pmap_kenter_quick
943 * Similar to pmap_kenter(), except we only invalidate the
944 * mapping on the current CPU.
947 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
952 npte = pa | PG_RW | PG_V | pgeflag;
955 cpu_invlpg((void *)va);
959 pmap_kenter_sync(vm_offset_t va)
961 pmap_inval_info info;
963 pmap_inval_init(&info);
964 pmap_inval_interlock(&info, &kernel_pmap, va);
965 pmap_inval_deinterlock(&info, &kernel_pmap);
966 pmap_inval_done(&info);
970 pmap_kenter_sync_quick(vm_offset_t va)
972 cpu_invlpg((void *)va);
976 * remove a page from the kernel pagetables
979 pmap_kremove(vm_offset_t va)
982 pmap_inval_info info;
984 pmap_inval_init(&info);
986 pmap_inval_interlock(&info, &kernel_pmap, va);
988 pmap_inval_deinterlock(&info, &kernel_pmap);
989 pmap_inval_done(&info);
993 pmap_kremove_quick(vm_offset_t va)
998 cpu_invlpg((void *)va);
1002 * XXX these need to be recoded. They are not used in any critical path.
1005 pmap_kmodify_rw(vm_offset_t va)
1007 *vtopte(va) |= PG_RW;
1008 cpu_invlpg((void *)va);
1012 pmap_kmodify_nc(vm_offset_t va)
1014 *vtopte(va) |= PG_N;
1015 cpu_invlpg((void *)va);
1019 * Used to map a range of physical addresses into kernel virtual
1020 * address space during the low level boot, typically to map the
1021 * dump bitmap, message buffer, and vm_page_array.
1023 * These mappings are typically made at some pointer after the end of the
1026 * We could return PHYS_TO_DMAP(start) here and not allocate any
1027 * via (*virtp), but then kmem from userland and kernel dumps won't
1028 * have access to the related pointers.
1031 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1034 vm_offset_t va_start;
1036 /*return PHYS_TO_DMAP(start);*/
1041 while (start < end) {
1042 pmap_kenter_quick(va, start);
1052 * Add a list of wired pages to the kva
1053 * this routine is only used for temporary
1054 * kernel mappings that do not need to have
1055 * page modification or references recorded.
1056 * Note that old mappings are simply written
1057 * over. The page *must* be wired.
1060 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1064 end_va = va + count * PAGE_SIZE;
1066 while (va < end_va) {
1070 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1071 cpu_invlpg((void *)va);
1076 smp_invltlb(); /* XXX */
1081 * This routine jerks page mappings from the
1082 * kernel -- it is meant only for temporary mappings.
1084 * MPSAFE, INTERRUPT SAFE (cluster callback)
1087 pmap_qremove(vm_offset_t va, int count)
1091 end_va = va + count * PAGE_SIZE;
1093 while (va < end_va) {
1098 cpu_invlpg((void *)va);
1107 * This routine works like vm_page_lookup() but also blocks as long as the
1108 * page is busy. This routine does not busy the page it returns.
1110 * Unless the caller is managing objects whos pages are in a known state,
1111 * the call should be made with a critical section held so the page's object
1112 * association remains valid on return.
1116 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1121 m = vm_page_lookup(object, pindex);
1122 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
1128 * Create a new thread and optionally associate it with a (new) process.
1129 * NOTE! the new thread's cpu may not equal the current cpu.
1132 pmap_init_thread(thread_t td)
1134 /* enforce pcb placement */
1135 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1136 td->td_savefpu = &td->td_pcb->pcb_save;
1137 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on x86_64? */
1141 * This routine directly affects the fork perf for a process.
1144 pmap_init_proc(struct proc *p)
1149 * Dispose the UPAGES for a process that has exited.
1150 * This routine directly impacts the exit perf of a process.
1153 pmap_dispose_proc(struct proc *p)
1155 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1158 /***************************************************
1159 * Page table page management routines.....
1160 ***************************************************/
1163 * This routine unholds page table pages, and if the hold count
1164 * drops to zero, then it decrements the wire count.
1168 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1169 pmap_inval_info_t info)
1171 KKASSERT(m->hold_count > 0);
1172 if (m->hold_count > 1) {
1176 return _pmap_unwire_pte_hold(pmap, va, m, info);
1182 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1183 pmap_inval_info_t info)
1186 * Wait until we can busy the page ourselves. We cannot have
1187 * any active flushes if we block. We own one hold count on the
1188 * page so it cannot be freed out from under us.
1190 if (m->flags & PG_BUSY) {
1191 pmap_inval_flush(info);
1192 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1195 KASSERT(m->queue == PQ_NONE,
1196 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1199 * This case can occur if new references were acquired while
1202 if (m->hold_count > 1) {
1203 KKASSERT(m->hold_count > 1);
1209 * Unmap the page table page
1211 KKASSERT(m->hold_count == 1);
1213 pmap_inval_interlock(info, pmap, -1);
1215 if (m->pindex >= (NUPDE + NUPDPE)) {
1218 pml4 = pmap_pml4e(pmap, va);
1220 } else if (m->pindex >= NUPDE) {
1223 pdp = pmap_pdpe(pmap, va);
1228 pd = pmap_pde(pmap, va);
1232 KKASSERT(pmap->pm_stats.resident_count > 0);
1233 --pmap->pm_stats.resident_count;
1235 if (pmap->pm_ptphint == m)
1236 pmap->pm_ptphint = NULL;
1237 pmap_inval_deinterlock(info, pmap);
1239 if (m->pindex < NUPDE) {
1240 /* We just released a PT, unhold the matching PD */
1243 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
1244 pmap_unwire_pte_hold(pmap, va, pdpg, info);
1246 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
1247 /* We just released a PD, unhold the matching PDP */
1250 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
1251 pmap_unwire_pte_hold(pmap, va, pdppg, info);
1255 * This was our last hold, the page had better be unwired
1256 * after we decrement wire_count.
1258 * FUTURE NOTE: shared page directory page could result in
1259 * multiple wire counts.
1263 KKASSERT(m->wire_count == 0);
1264 --vmstats.v_wire_count;
1265 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1267 vm_page_free_zero(m);
1273 * After removing a page table entry, this routine is used to
1274 * conditionally free the page, and manage the hold/wire counts.
1278 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1279 pmap_inval_info_t info)
1281 vm_pindex_t ptepindex;
1283 if (va >= VM_MAX_USER_ADDRESS)
1287 ptepindex = pmap_pde_pindex(va);
1289 if (pmap->pm_ptphint &&
1290 (pmap->pm_ptphint->pindex == ptepindex)) {
1291 mpte = pmap->pm_ptphint;
1294 pmap_inval_flush(info);
1295 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1296 pmap->pm_ptphint = mpte;
1301 return pmap_unwire_pte_hold(pmap, va, mpte, info);
1305 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1306 * it, and IdlePTD, represents the template used to update all other pmaps.
1308 * On architectures where the kernel pmap is not integrated into the user
1309 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1310 * kernel_pmap should be used to directly access the kernel_pmap.
1313 pmap_pinit0(struct pmap *pmap)
1315 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1317 pmap->pm_active = 0;
1318 pmap->pm_ptphint = NULL;
1319 TAILQ_INIT(&pmap->pm_pvlist);
1320 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1324 * Initialize a preallocated and zeroed pmap structure,
1325 * such as one in a vmspace structure.
1328 pmap_pinit(struct pmap *pmap)
1333 * No need to allocate page table space yet but we do need a valid
1334 * page directory table.
1336 if (pmap->pm_pml4 == NULL) {
1338 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1342 * Allocate an object for the ptes
1344 if (pmap->pm_pteobj == NULL)
1345 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1);
1348 * Allocate the page directory page, unless we already have
1349 * one cached. If we used the cached page the wire_count will
1350 * already be set appropriately.
1352 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1353 ptdpg = vm_page_grab(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I,
1354 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1355 pmap->pm_pdirm = ptdpg;
1356 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1357 ptdpg->valid = VM_PAGE_BITS_ALL;
1358 if (ptdpg->wire_count == 0)
1359 ++vmstats.v_wire_count;
1360 ptdpg->wire_count = 1;
1361 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1363 if ((ptdpg->flags & PG_ZERO) == 0)
1364 bzero(pmap->pm_pml4, PAGE_SIZE);
1366 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1367 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1369 /* install self-referential address mapping entry */
1370 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1373 pmap->pm_active = 0;
1374 pmap->pm_ptphint = NULL;
1375 TAILQ_INIT(&pmap->pm_pvlist);
1376 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1377 pmap->pm_stats.resident_count = 1;
1381 * Clean up a pmap structure so it can be physically freed. This routine
1382 * is called by the vmspace dtor function. A great deal of pmap data is
1383 * left passively mapped to improve vmspace management so we have a bit
1384 * of cleanup work to do here.
1387 pmap_puninit(pmap_t pmap)
1391 KKASSERT(pmap->pm_active == 0);
1392 lwkt_gettoken(&vm_token);
1393 if ((p = pmap->pm_pdirm) != NULL) {
1394 KKASSERT(pmap->pm_pml4 != NULL);
1395 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1396 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1398 vmstats.v_wire_count--;
1399 KKASSERT((p->flags & PG_BUSY) == 0);
1401 vm_page_free_zero(p);
1402 pmap->pm_pdirm = NULL;
1404 if (pmap->pm_pml4) {
1405 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1406 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1407 pmap->pm_pml4 = NULL;
1409 if (pmap->pm_pteobj) {
1410 vm_object_deallocate(pmap->pm_pteobj);
1411 pmap->pm_pteobj = NULL;
1413 lwkt_reltoken(&vm_token);
1417 * Wire in kernel global address entries. To avoid a race condition
1418 * between pmap initialization and pmap_growkernel, this procedure
1419 * adds the pmap to the master list (which growkernel scans to update),
1420 * then copies the template.
1423 pmap_pinit2(struct pmap *pmap)
1426 lwkt_gettoken(&vm_token);
1427 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1428 /* XXX copies current process, does not fill in MPPTDI */
1429 lwkt_reltoken(&vm_token);
1434 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1435 * 0 on failure (if the procedure had to sleep).
1437 * When asked to remove the page directory page itself, we actually just
1438 * leave it cached so we do not have to incur the SMP inval overhead of
1439 * removing the kernel mapping. pmap_puninit() will take care of it.
1443 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1446 * This code optimizes the case of freeing non-busy
1447 * page-table pages. Those pages are zero now, and
1448 * might as well be placed directly into the zero queue.
1450 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1456 * Remove the page table page from the processes address space.
1458 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1460 * We are the pml4 table itself.
1462 /* XXX anything to do here? */
1463 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1465 * Remove a PDP page from the PML4. We do not maintain
1466 * hold counts on the PML4 page.
1472 m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1473 KKASSERT(m4 != NULL);
1474 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1475 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1476 KKASSERT(pml4[idx] != 0);
1478 } else if (p->pindex >= NUPDE) {
1480 * Remove a PD page from the PDP and drop the hold count
1481 * on the PDP. The PDP is left cached in the pmap if
1482 * the hold count drops to 0 so the wire count remains
1489 m3 = vm_page_lookup(pmap->pm_pteobj,
1490 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1491 KKASSERT(m3 != NULL);
1492 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1493 idx = (p->pindex - NUPDE) % NPDPEPG;
1494 KKASSERT(pdp[idx] != 0);
1499 * Remove a PT page from the PD and drop the hold count
1500 * on the PD. The PD is left cached in the pmap if
1501 * the hold count drops to 0 so the wire count remains
1508 m2 = vm_page_lookup(pmap->pm_pteobj,
1509 NUPDE + p->pindex / NPDEPG);
1510 KKASSERT(m2 != NULL);
1511 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1512 idx = p->pindex % NPDEPG;
1518 * One fewer mappings in the pmap. p's hold count had better
1521 KKASSERT(pmap->pm_stats.resident_count > 0);
1522 --pmap->pm_stats.resident_count;
1524 panic("pmap_release: freeing held page table page");
1525 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1526 pmap->pm_ptphint = NULL;
1529 * We leave the top-level page table page cached, wired, and mapped in
1530 * the pmap until the dtor function (pmap_puninit()) gets called.
1531 * However, still clean it up so we can set PG_ZERO.
1533 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1534 bzero(pmap->pm_pml4, PAGE_SIZE);
1535 vm_page_flag_set(p, PG_ZERO);
1539 KKASSERT(p->wire_count == 0);
1540 vmstats.v_wire_count--;
1541 /* JG eventually revert to using vm_page_free_zero() */
1548 * This routine is called when various levels in the page table need to
1549 * be populated. This routine cannot fail.
1553 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1558 * Find or fabricate a new pagetable page. This will busy the page.
1560 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1561 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1562 if ((m->flags & PG_ZERO) == 0) {
1563 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1566 KASSERT(m->queue == PQ_NONE,
1567 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1570 * Increment the hold count for the page we will be returning to
1574 if (m->wire_count++ == 0)
1575 vmstats.v_wire_count++;
1578 * Map the pagetable page into the process address space, if
1579 * it isn't already there.
1581 * It is possible that someone else got in and mapped the page
1582 * directory page while we were blocked, if so just unbusy and
1583 * return the held page.
1585 if (ptepindex >= (NUPDE + NUPDPE)) {
1587 * Wire up a new PDP page in the PML4
1589 vm_pindex_t pml4index;
1592 pml4index = ptepindex - (NUPDE + NUPDPE);
1593 pml4 = &pmap->pm_pml4[pml4index];
1595 if (--m->wire_count == 0)
1596 --vmstats.v_wire_count;
1600 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1601 } else if (ptepindex >= NUPDE) {
1603 * Wire up a new PD page in the PDP
1605 vm_pindex_t pml4index;
1606 vm_pindex_t pdpindex;
1611 pdpindex = ptepindex - NUPDE;
1612 pml4index = pdpindex >> NPML4EPGSHIFT;
1614 pml4 = &pmap->pm_pml4[pml4index];
1615 if ((*pml4 & PG_V) == 0) {
1617 * Have to allocate a new PDP page, recurse.
1618 * This always succeeds. Returned page will
1621 pdppg = _pmap_allocpte(pmap,
1622 NUPDE + NUPDPE + pml4index);
1625 * Add a held reference to the PDP page.
1627 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1628 pdppg->hold_count++;
1632 * Now find the pdp_entry and map the PDP. If the PDP
1633 * has already been mapped unwind and return the
1634 * already-mapped PDP held.
1636 * pdppg is left held (hold_count is incremented for
1637 * each PD in the PDP).
1639 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1640 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1642 vm_page_unhold(pdppg);
1643 if (--m->wire_count == 0)
1644 --vmstats.v_wire_count;
1648 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1651 * Wire up the new PT page in the PD
1653 vm_pindex_t pml4index;
1654 vm_pindex_t pdpindex;
1660 pdpindex = ptepindex >> NPDPEPGSHIFT;
1661 pml4index = pdpindex >> NPML4EPGSHIFT;
1664 * Locate the PDP page in the PML4, then the PD page in
1665 * the PDP. If either does not exist we simply recurse
1668 * We can just recurse on the PD page as it will recurse
1669 * on the PDP if necessary.
1671 pml4 = &pmap->pm_pml4[pml4index];
1672 if ((*pml4 & PG_V) == 0) {
1673 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1674 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1675 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1677 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1678 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1679 if ((*pdp & PG_V) == 0) {
1680 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1682 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1688 * Now fill in the pte in the PD. If the pte already exists
1689 * (again, if we raced the grab), unhold pdpg and unwire
1690 * m, returning a held m.
1692 * pdpg is left held (hold_count is incremented for
1693 * each PT in the PD).
1695 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
1696 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1698 vm_page_unhold(pdpg);
1699 if (--m->wire_count == 0)
1700 --vmstats.v_wire_count;
1704 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1708 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1709 * valid bits, mapped flag, unbusy, and we're done.
1711 pmap->pm_ptphint = m;
1712 ++pmap->pm_stats.resident_count;
1714 m->valid = VM_PAGE_BITS_ALL;
1715 vm_page_flag_clear(m, PG_ZERO);
1716 vm_page_flag_set(m, PG_MAPPED);
1724 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1726 vm_pindex_t ptepindex;
1731 * Calculate pagetable page index
1733 ptepindex = pmap_pde_pindex(va);
1736 * Get the page directory entry
1738 pd = pmap_pde(pmap, va);
1741 * This supports switching from a 2MB page to a
1744 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1745 panic("no promotion/demotion yet");
1753 * If the page table page is mapped, we just increment the
1754 * hold count, and activate it.
1756 if (pd != NULL && (*pd & PG_V) != 0) {
1757 /* YYY hint is used here on i386 */
1758 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1759 pmap->pm_ptphint = m;
1764 * Here if the pte page isn't mapped, or if it has been deallocated.
1766 return _pmap_allocpte(pmap, ptepindex);
1770 /***************************************************
1771 * Pmap allocation/deallocation routines.
1772 ***************************************************/
1775 * Release any resources held by the given physical map.
1776 * Called when a pmap initialized by pmap_pinit is being released.
1777 * Should only be called if the map contains no valid mappings.
1779 static int pmap_release_callback(struct vm_page *p, void *data);
1782 pmap_release(struct pmap *pmap)
1784 vm_object_t object = pmap->pm_pteobj;
1785 struct rb_vm_page_scan_info info;
1787 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1788 #if defined(DIAGNOSTIC)
1789 if (object->ref_count != 1)
1790 panic("pmap_release: pteobj reference count != 1");
1794 info.object = object;
1796 lwkt_gettoken(&vm_token);
1797 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1804 info.limit = object->generation;
1806 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1807 pmap_release_callback, &info);
1808 if (info.error == 0 && info.mpte) {
1809 if (!pmap_release_free_page(pmap, info.mpte))
1813 } while (info.error);
1814 lwkt_reltoken(&vm_token);
1819 pmap_release_callback(struct vm_page *p, void *data)
1821 struct rb_vm_page_scan_info *info = data;
1823 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1827 if (!pmap_release_free_page(info->pmap, p)) {
1831 if (info->object->generation != info->limit) {
1839 * Grow the number of kernel page table entries, if needed.
1841 * This routine is always called to validate any address space
1842 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
1843 * space below KERNBASE.
1846 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1849 vm_offset_t ptppaddr;
1851 pd_entry_t *pde, newpdir;
1853 int update_kernel_vm_end;
1856 lwkt_gettoken(&vm_token);
1859 * bootstrap kernel_vm_end on first real VM use
1861 if (kernel_vm_end == 0) {
1862 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1864 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1865 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1866 ~(PAGE_SIZE * NPTEPG - 1);
1868 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1869 kernel_vm_end = kernel_map.max_offset;
1876 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
1877 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
1878 * do not want to force-fill 128G worth of page tables.
1880 if (kstart < KERNBASE) {
1881 if (kstart > kernel_vm_end)
1882 kstart = kernel_vm_end;
1883 KKASSERT(kend <= KERNBASE);
1884 update_kernel_vm_end = 1;
1886 update_kernel_vm_end = 0;
1889 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
1890 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
1892 if (kend - 1 >= kernel_map.max_offset)
1893 kend = kernel_map.max_offset;
1895 while (kstart < kend) {
1896 pde = pmap_pde(&kernel_pmap, kstart);
1898 /* We need a new PDP entry */
1899 nkpg = vm_page_alloc(kptobj, nkpt,
1902 VM_ALLOC_INTERRUPT);
1904 panic("pmap_growkernel: no memory to grow "
1907 paddr = VM_PAGE_TO_PHYS(nkpg);
1908 if ((nkpg->flags & PG_ZERO) == 0)
1909 pmap_zero_page(paddr);
1910 vm_page_flag_clear(nkpg, PG_ZERO);
1911 newpdp = (pdp_entry_t)
1912 (paddr | PG_V | PG_RW | PG_A | PG_M);
1913 *pmap_pdpe(&kernel_pmap, kstart) = newpdp;
1915 continue; /* try again */
1917 if ((*pde & PG_V) != 0) {
1918 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1919 ~(PAGE_SIZE * NPTEPG - 1);
1920 if (kstart - 1 >= kernel_map.max_offset) {
1921 kstart = kernel_map.max_offset;
1928 * This index is bogus, but out of the way
1930 nkpg = vm_page_alloc(kptobj, nkpt,
1933 VM_ALLOC_INTERRUPT);
1935 panic("pmap_growkernel: no memory to grow kernel");
1938 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1939 pmap_zero_page(ptppaddr);
1940 vm_page_flag_clear(nkpg, PG_ZERO);
1941 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1942 *pmap_pde(&kernel_pmap, kstart) = newpdir;
1945 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1946 ~(PAGE_SIZE * NPTEPG - 1);
1948 if (kstart - 1 >= kernel_map.max_offset) {
1949 kstart = kernel_map.max_offset;
1955 * Only update kernel_vm_end for areas below KERNBASE.
1957 if (update_kernel_vm_end && kernel_vm_end < kstart)
1958 kernel_vm_end = kstart;
1960 lwkt_reltoken(&vm_token);
1965 * Retire the given physical map from service.
1966 * Should only be called if the map contains
1967 * no valid mappings.
1970 pmap_destroy(pmap_t pmap)
1977 lwkt_gettoken(&vm_token);
1978 count = --pmap->pm_count;
1981 panic("destroying a pmap is not yet implemented");
1983 lwkt_reltoken(&vm_token);
1987 * Add a reference to the specified pmap.
1990 pmap_reference(pmap_t pmap)
1993 lwkt_gettoken(&vm_token);
1995 lwkt_reltoken(&vm_token);
1999 /***************************************************
2000 * page management routines.
2001 ***************************************************/
2004 * free the pv_entry back to the free list. This function may be
2005 * called from an interrupt.
2009 free_pv_entry(pv_entry_t pv)
2012 KKASSERT(pv_entry_count >= 0);
2017 * get a new pv_entry, allocating a block from the system
2018 * when needed. This function may be called from an interrupt.
2025 if (pv_entry_high_water &&
2026 (pv_entry_count > pv_entry_high_water) &&
2027 (pmap_pagedaemon_waken == 0)) {
2028 pmap_pagedaemon_waken = 1;
2029 wakeup(&vm_pages_needed);
2031 return zalloc(pvzone);
2035 * This routine is very drastic, but can save the system
2043 static int warningdone=0;
2045 if (pmap_pagedaemon_waken == 0)
2047 lwkt_gettoken(&vm_token);
2048 if (warningdone < 5) {
2049 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
2053 for(i = 0; i < vm_page_array_size; i++) {
2054 m = &vm_page_array[i];
2055 if (m->wire_count || m->hold_count || m->busy ||
2056 (m->flags & PG_BUSY))
2060 pmap_pagedaemon_waken = 0;
2061 lwkt_reltoken(&vm_token);
2066 * If it is the first entry on the list, it is actually
2067 * in the header and we must copy the following entry up
2068 * to the header. Otherwise we must search the list for
2069 * the entry. In either case we free the now unused entry.
2073 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
2074 vm_offset_t va, pmap_inval_info_t info)
2080 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
2081 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2082 if (pmap == pv->pv_pmap && va == pv->pv_va)
2086 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2087 if (va == pv->pv_va)
2095 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2096 m->md.pv_list_count--;
2097 KKASSERT(m->md.pv_list_count >= 0);
2098 if (TAILQ_EMPTY(&m->md.pv_list))
2099 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2100 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2101 ++pmap->pm_generation;
2102 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
2110 * Create a pv entry for page at pa for
2115 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2120 pv = get_pv_entry();
2125 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2126 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2127 ++pmap->pm_generation;
2128 m->md.pv_list_count++;
2134 * pmap_remove_pte: do the things to unmap a page in a process
2138 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2139 pmap_inval_info_t info)
2144 pmap_inval_interlock(info, pmap, va);
2145 oldpte = pte_load_clear(ptq);
2146 pmap_inval_deinterlock(info, pmap);
2148 pmap->pm_stats.wired_count -= 1;
2150 * Machines that don't support invlpg, also don't support
2151 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2155 cpu_invlpg((void *)va);
2156 KKASSERT(pmap->pm_stats.resident_count > 0);
2157 --pmap->pm_stats.resident_count;
2158 if (oldpte & PG_MANAGED) {
2159 m = PHYS_TO_VM_PAGE(oldpte);
2160 if (oldpte & PG_M) {
2161 #if defined(PMAP_DIAGNOSTIC)
2162 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2164 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2168 if (pmap_track_modified(va))
2172 vm_page_flag_set(m, PG_REFERENCED);
2173 return pmap_remove_entry(pmap, m, va, info);
2175 return pmap_unuse_pt(pmap, va, NULL, info);
2184 * Remove a single page from a process address space.
2186 * This function may not be called from an interrupt if the pmap is
2191 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2195 pte = pmap_pte(pmap, va);
2198 if ((*pte & PG_V) == 0)
2200 pmap_remove_pte(pmap, pte, va, info);
2206 * Remove the given range of addresses from the specified map.
2208 * It is assumed that the start and end are properly
2209 * rounded to the page size.
2211 * This function may not be called from an interrupt if the pmap is
2215 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2217 vm_offset_t va_next;
2218 pml4_entry_t *pml4e;
2220 pd_entry_t ptpaddr, *pde;
2222 struct pmap_inval_info info;
2227 lwkt_gettoken(&vm_token);
2228 if (pmap->pm_stats.resident_count == 0) {
2229 lwkt_reltoken(&vm_token);
2233 pmap_inval_init(&info);
2236 * special handling of removing one page. a very
2237 * common operation and easy to short circuit some
2240 if (sva + PAGE_SIZE == eva) {
2241 pde = pmap_pde(pmap, sva);
2242 if (pde && (*pde & PG_PS) == 0) {
2243 pmap_remove_page(pmap, sva, &info);
2244 pmap_inval_done(&info);
2245 lwkt_reltoken(&vm_token);
2250 for (; sva < eva; sva = va_next) {
2251 pml4e = pmap_pml4e(pmap, sva);
2252 if ((*pml4e & PG_V) == 0) {
2253 va_next = (sva + NBPML4) & ~PML4MASK;
2259 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2260 if ((*pdpe & PG_V) == 0) {
2261 va_next = (sva + NBPDP) & ~PDPMASK;
2268 * Calculate index for next page table.
2270 va_next = (sva + NBPDR) & ~PDRMASK;
2274 pde = pmap_pdpe_to_pde(pdpe, sva);
2278 * Weed out invalid mappings.
2284 * Check for large page.
2286 if ((ptpaddr & PG_PS) != 0) {
2287 /* JG FreeBSD has more complex treatment here */
2288 pmap_inval_interlock(&info, pmap, -1);
2290 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2291 pmap_inval_deinterlock(&info, pmap);
2296 * Limit our scan to either the end of the va represented
2297 * by the current page table page, or to the end of the
2298 * range being removed.
2304 * NOTE: pmap_remove_pte() can block.
2306 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2310 if (pmap_remove_pte(pmap, pte, sva, &info))
2314 pmap_inval_done(&info);
2315 lwkt_reltoken(&vm_token);
2321 * Removes this physical page from all physical maps in which it resides.
2322 * Reflects back modify bits to the pager.
2324 * This routine may not be called from an interrupt.
2329 pmap_remove_all(vm_page_t m)
2331 struct pmap_inval_info info;
2332 pt_entry_t *pte, tpte;
2335 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2338 lwkt_gettoken(&vm_token);
2339 pmap_inval_init(&info);
2341 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2342 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2343 --pv->pv_pmap->pm_stats.resident_count;
2345 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2346 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
2347 tpte = pte_load_clear(pte);
2349 pv->pv_pmap->pm_stats.wired_count--;
2350 pmap_inval_deinterlock(&info, pv->pv_pmap);
2352 vm_page_flag_set(m, PG_REFERENCED);
2355 * Update the vm_page_t clean and reference bits.
2358 #if defined(PMAP_DIAGNOSTIC)
2359 if (pmap_nw_modified(tpte)) {
2361 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2365 if (pmap_track_modified(pv->pv_va))
2368 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2369 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2370 ++pv->pv_pmap->pm_generation;
2371 m->md.pv_list_count--;
2372 KKASSERT(m->md.pv_list_count >= 0);
2373 if (TAILQ_EMPTY(&m->md.pv_list))
2374 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2375 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2379 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2380 pmap_inval_done(&info);
2381 lwkt_reltoken(&vm_token);
2387 * Set the physical protection on the specified range of this map
2390 * This function may not be called from an interrupt if the map is
2391 * not the kernel_pmap.
2394 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2396 vm_offset_t va_next;
2397 pml4_entry_t *pml4e;
2399 pd_entry_t ptpaddr, *pde;
2401 pmap_inval_info info;
2403 /* JG review for NX */
2408 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2409 pmap_remove(pmap, sva, eva);
2413 if (prot & VM_PROT_WRITE)
2416 lwkt_gettoken(&vm_token);
2417 pmap_inval_init(&info);
2419 for (; sva < eva; sva = va_next) {
2421 pml4e = pmap_pml4e(pmap, sva);
2422 if ((*pml4e & PG_V) == 0) {
2423 va_next = (sva + NBPML4) & ~PML4MASK;
2429 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2430 if ((*pdpe & PG_V) == 0) {
2431 va_next = (sva + NBPDP) & ~PDPMASK;
2437 va_next = (sva + NBPDR) & ~PDRMASK;
2441 pde = pmap_pdpe_to_pde(pdpe, sva);
2445 * Check for large page.
2447 if ((ptpaddr & PG_PS) != 0) {
2448 pmap_inval_interlock(&info, pmap, -1);
2449 *pde &= ~(PG_M|PG_RW);
2450 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2451 pmap_inval_deinterlock(&info, pmap);
2456 * Weed out invalid mappings. Note: we assume that the page
2457 * directory table is always allocated, and in kernel virtual.
2465 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2472 * XXX non-optimal. Note also that there can be
2473 * no pmap_inval_flush() calls until after we modify
2474 * ptbase[sindex] (or otherwise we have to do another
2475 * pmap_inval_add() call).
2477 pmap_inval_interlock(&info, pmap, sva);
2481 if ((pbits & PG_V) == 0) {
2482 pmap_inval_deinterlock(&info, pmap);
2485 if (pbits & PG_MANAGED) {
2488 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2489 vm_page_flag_set(m, PG_REFERENCED);
2493 if (pmap_track_modified(sva)) {
2495 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2502 if (pbits != cbits &&
2503 !atomic_cmpset_long(pte, pbits, cbits)) {
2506 pmap_inval_deinterlock(&info, pmap);
2509 pmap_inval_done(&info);
2510 lwkt_reltoken(&vm_token);
2514 * Insert the given physical page (p) at
2515 * the specified virtual address (v) in the
2516 * target physical map with the protection requested.
2518 * If specified, the page will be wired down, meaning
2519 * that the related pte can not be reclaimed.
2521 * NB: This is the only routine which MAY NOT lazy-evaluate
2522 * or lose information. That is, this routine must actually
2523 * insert this page into the given map NOW.
2526 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2533 pt_entry_t origpte, newpte;
2535 pmap_inval_info info;
2540 va = trunc_page(va);
2541 #ifdef PMAP_DIAGNOSTIC
2543 panic("pmap_enter: toobig");
2544 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2545 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
2547 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2548 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2550 db_print_backtrace();
2553 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2554 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2556 db_print_backtrace();
2560 lwkt_gettoken(&vm_token);
2563 * In the case that a page table page is not
2564 * resident, we are creating it here.
2566 if (va < VM_MAX_USER_ADDRESS)
2567 mpte = pmap_allocpte(pmap, va);
2571 pmap_inval_init(&info);
2572 pde = pmap_pde(pmap, va);
2573 if (pde != NULL && (*pde & PG_V) != 0) {
2574 if ((*pde & PG_PS) != 0)
2575 panic("pmap_enter: attempted pmap_enter on 2MB page");
2576 pte = pmap_pde_to_pte(pde, va);
2578 panic("pmap_enter: invalid page directory va=%#lx", va);
2580 KKASSERT(pte != NULL);
2581 pa = VM_PAGE_TO_PHYS(m);
2583 opa = origpte & PG_FRAME;
2586 * Mapping has not changed, must be protection or wiring change.
2588 if (origpte && (opa == pa)) {
2590 * Wiring change, just update stats. We don't worry about
2591 * wiring PT pages as they remain resident as long as there
2592 * are valid mappings in them. Hence, if a user page is wired,
2593 * the PT page will be also.
2595 if (wired && ((origpte & PG_W) == 0))
2596 pmap->pm_stats.wired_count++;
2597 else if (!wired && (origpte & PG_W))
2598 pmap->pm_stats.wired_count--;
2600 #if defined(PMAP_DIAGNOSTIC)
2601 if (pmap_nw_modified(origpte)) {
2603 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2609 * Remove the extra pte reference. Note that we cannot
2610 * optimize the RO->RW case because we have adjusted the
2611 * wiring count above and may need to adjust the wiring
2618 * We might be turning off write access to the page,
2619 * so we go ahead and sense modify status.
2621 if (origpte & PG_MANAGED) {
2622 if ((origpte & PG_M) && pmap_track_modified(va)) {
2624 om = PHYS_TO_VM_PAGE(opa);
2628 KKASSERT(m->flags & PG_MAPPED);
2633 * Mapping has changed, invalidate old range and fall through to
2634 * handle validating new mapping.
2638 err = pmap_remove_pte(pmap, pte, va, &info);
2640 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2642 opa = origpte & PG_FRAME;
2644 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2650 * Enter on the PV list if part of our managed memory. Note that we
2651 * raise IPL while manipulating pv_table since pmap_enter can be
2652 * called at interrupt time.
2654 if (pmap_initialized &&
2655 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2656 pmap_insert_entry(pmap, va, mpte, m);
2658 vm_page_flag_set(m, PG_MAPPED);
2662 * Increment counters
2664 ++pmap->pm_stats.resident_count;
2666 pmap->pm_stats.wired_count++;
2670 * Now validate mapping with desired protection/wiring.
2672 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
2676 if (va < VM_MAX_USER_ADDRESS)
2678 if (pmap == &kernel_pmap)
2682 * if the mapping or permission bits are different, we need
2683 * to update the pte.
2685 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2686 pmap_inval_interlock(&info, pmap, va);
2687 *pte = newpte | PG_A;
2688 pmap_inval_deinterlock(&info, pmap);
2690 vm_page_flag_set(m, PG_WRITEABLE);
2692 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2693 pmap_inval_done(&info);
2694 lwkt_reltoken(&vm_token);
2698 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2699 * This code also assumes that the pmap has no pre-existing entry for this
2702 * This code currently may only be used on user pmaps, not kernel_pmap.
2705 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2710 vm_pindex_t ptepindex;
2712 pmap_inval_info info;
2714 lwkt_gettoken(&vm_token);
2715 pmap_inval_init(&info);
2717 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2718 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2720 db_print_backtrace();
2723 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2724 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2726 db_print_backtrace();
2730 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2733 * Calculate the page table page (mpte), allocating it if necessary.
2735 * A held page table page (mpte), or NULL, is passed onto the
2736 * section following.
2738 if (va < VM_MAX_USER_ADDRESS) {
2740 * Calculate pagetable page index
2742 ptepindex = pmap_pde_pindex(va);
2746 * Get the page directory entry
2748 ptepa = pmap_pde(pmap, va);
2751 * If the page table page is mapped, we just increment
2752 * the hold count, and activate it.
2754 if (ptepa && (*ptepa & PG_V) != 0) {
2756 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2757 // if (pmap->pm_ptphint &&
2758 // (pmap->pm_ptphint->pindex == ptepindex)) {
2759 // mpte = pmap->pm_ptphint;
2761 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2762 pmap->pm_ptphint = mpte;
2767 mpte = _pmap_allocpte(pmap, ptepindex);
2769 } while (mpte == NULL);
2772 /* this code path is not yet used */
2776 * With a valid (and held) page directory page, we can just use
2777 * vtopte() to get to the pte. If the pte is already present
2778 * we do not disturb it.
2783 pmap_unwire_pte_hold(pmap, va, mpte, &info);
2784 pa = VM_PAGE_TO_PHYS(m);
2785 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2786 pmap_inval_done(&info);
2787 lwkt_reltoken(&vm_token);
2792 * Enter on the PV list if part of our managed memory
2794 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2795 pmap_insert_entry(pmap, va, mpte, m);
2796 vm_page_flag_set(m, PG_MAPPED);
2800 * Increment counters
2802 ++pmap->pm_stats.resident_count;
2804 pa = VM_PAGE_TO_PHYS(m);
2807 * Now validate mapping with RO protection
2809 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2810 *pte = pa | PG_V | PG_U;
2812 *pte = pa | PG_V | PG_U | PG_MANAGED;
2813 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2814 pmap_inval_done(&info);
2815 lwkt_reltoken(&vm_token);
2819 * Make a temporary mapping for a physical address. This is only intended
2820 * to be used for panic dumps.
2822 /* JG Needed on x86_64? */
2824 pmap_kenter_temporary(vm_paddr_t pa, int i)
2826 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2827 return ((void *)crashdumpmap);
2830 #define MAX_INIT_PT (96)
2833 * This routine preloads the ptes for a given object into the specified pmap.
2834 * This eliminates the blast of soft faults on process startup and
2835 * immediately after an mmap.
2837 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2840 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2841 vm_object_t object, vm_pindex_t pindex,
2842 vm_size_t size, int limit)
2844 struct rb_vm_page_scan_info info;
2849 * We can't preinit if read access isn't set or there is no pmap
2852 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2856 * We can't preinit if the pmap is not the current pmap
2858 lp = curthread->td_lwp;
2859 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2862 psize = x86_64_btop(size);
2864 if ((object->type != OBJT_VNODE) ||
2865 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2866 (object->resident_page_count > MAX_INIT_PT))) {
2870 if (psize + pindex > object->size) {
2871 if (object->size < pindex)
2873 psize = object->size - pindex;
2880 * Use a red-black scan to traverse the requested range and load
2881 * any valid pages found into the pmap.
2883 * We cannot safely scan the object's memq unless we are in a
2884 * critical section since interrupts can remove pages from objects.
2886 info.start_pindex = pindex;
2887 info.end_pindex = pindex + psize - 1;
2894 lwkt_gettoken(&vm_token);
2895 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2896 pmap_object_init_pt_callback, &info);
2897 lwkt_reltoken(&vm_token);
2903 pmap_object_init_pt_callback(vm_page_t p, void *data)
2905 struct rb_vm_page_scan_info *info = data;
2906 vm_pindex_t rel_index;
2908 * don't allow an madvise to blow away our really
2909 * free pages allocating pv entries.
2911 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2912 vmstats.v_free_count < vmstats.v_free_reserved) {
2915 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2916 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2917 if ((p->queue - p->pc) == PQ_CACHE)
2918 vm_page_deactivate(p);
2920 rel_index = p->pindex - info->start_pindex;
2921 pmap_enter_quick(info->pmap,
2922 info->addr + x86_64_ptob(rel_index), p);
2929 * Return TRUE if the pmap is in shape to trivially
2930 * pre-fault the specified address.
2932 * Returns FALSE if it would be non-trivial or if a
2933 * pte is already loaded into the slot.
2936 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2942 lwkt_gettoken(&vm_token);
2943 pde = pmap_pde(pmap, addr);
2944 if (pde == NULL || *pde == 0) {
2948 ret = (*pte) ? 0 : 1;
2950 lwkt_reltoken(&vm_token);
2955 * Routine: pmap_change_wiring
2956 * Function: Change the wiring attribute for a map/virtual-address
2958 * In/out conditions:
2959 * The mapping must already exist in the pmap.
2962 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2969 lwkt_gettoken(&vm_token);
2970 pte = pmap_pte(pmap, va);
2972 if (wired && !pmap_pte_w(pte))
2973 pmap->pm_stats.wired_count++;
2974 else if (!wired && pmap_pte_w(pte))
2975 pmap->pm_stats.wired_count--;
2978 * Wiring is not a hardware characteristic so there is no need to
2979 * invalidate TLB. However, in an SMP environment we must use
2980 * a locked bus cycle to update the pte (if we are not using
2981 * the pmap_inval_*() API that is)... it's ok to do this for simple
2986 atomic_set_long(pte, PG_W);
2988 atomic_clear_long(pte, PG_W);
2991 atomic_set_long_nonlocked(pte, PG_W);
2993 atomic_clear_long_nonlocked(pte, PG_W);
2995 lwkt_reltoken(&vm_token);
3001 * Copy the range specified by src_addr/len
3002 * from the source map to the range dst_addr/len
3003 * in the destination map.
3005 * This routine is only advisory and need not do anything.
3008 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3009 vm_size_t len, vm_offset_t src_addr)
3013 pmap_inval_info info;
3015 vm_offset_t end_addr = src_addr + len;
3017 pd_entry_t src_frame, dst_frame;
3020 if (dst_addr != src_addr)
3023 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3024 if (src_frame != (PTDpde & PG_FRAME)) {
3028 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3029 if (dst_frame != (APTDpde & PG_FRAME)) {
3030 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
3031 /* The page directory is not shared between CPUs */
3035 pmap_inval_init(&info);
3036 pmap_inval_add(&info, dst_pmap, -1);
3037 pmap_inval_add(&info, src_pmap, -1);
3040 * critical section protection is required to maintain the page/object
3041 * association, interrupts can free pages and remove them from
3045 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
3046 pt_entry_t *src_pte, *dst_pte;
3047 vm_page_t dstmpte, srcmpte;
3048 vm_offset_t srcptepaddr;
3049 vm_pindex_t ptepindex;
3051 if (addr >= UPT_MIN_ADDRESS)
3052 panic("pmap_copy: invalid to pmap_copy page tables\n");
3055 * Don't let optional prefaulting of pages make us go
3056 * way below the low water mark of free pages or way
3057 * above high water mark of used pv entries.
3059 if (vmstats.v_free_count < vmstats.v_free_reserved ||
3060 pv_entry_count > pv_entry_high_water)
3063 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
3064 ptepindex = addr >> PDRSHIFT;
3067 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
3069 if (srcptepaddr == 0)
3072 if (srcptepaddr & PG_PS) {
3074 if (dst_pmap->pm_pdir[ptepindex] == 0) {
3075 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
3076 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
3082 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
3083 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
3084 (srcmpte->flags & PG_BUSY)) {
3088 if (pdnxt > end_addr)
3091 src_pte = vtopte(addr);
3093 dst_pte = avtopte(addr);
3095 while (addr < pdnxt) {
3100 * we only virtual copy managed pages
3102 if ((ptetemp & PG_MANAGED) != 0) {
3104 * We have to check after allocpte for the
3105 * pte still being around... allocpte can
3108 * pmap_allocpte() can block. If we lose
3109 * our page directory mappings we stop.
3111 dstmpte = pmap_allocpte(dst_pmap, addr);
3114 if (src_frame != (PTDpde & PG_FRAME) ||
3115 dst_frame != (APTDpde & PG_FRAME)
3117 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3118 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3120 } else if ((*dst_pte == 0) &&
3121 (ptetemp = *src_pte) != 0 &&
3122 (ptetemp & PG_MANAGED)) {
3124 * Clear the modified and
3125 * accessed (referenced) bits
3128 m = PHYS_TO_VM_PAGE(ptetemp);
3129 *dst_pte = ptetemp & ~(PG_M | PG_A);
3130 ++dst_pmap->pm_stats.resident_count;
3131 pmap_insert_entry(dst_pmap, addr,
3133 KKASSERT(m->flags & PG_MAPPED);
3135 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3136 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3140 if (dstmpte->hold_count >= srcmpte->hold_count)
3150 pmap_inval_done(&info);
3157 * Zero the specified physical page.
3159 * This function may be called from an interrupt and no locking is
3163 pmap_zero_page(vm_paddr_t phys)
3165 vm_offset_t va = PHYS_TO_DMAP(phys);
3167 pagezero((void *)va);
3171 * pmap_page_assertzero:
3173 * Assert that a page is empty, panic if it isn't.
3176 pmap_page_assertzero(vm_paddr_t phys)
3178 vm_offset_t virt = PHYS_TO_DMAP(phys);
3181 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3182 if (*(long *)((char *)virt + i) != 0) {
3183 panic("pmap_page_assertzero() @ %p not zero!\n", (void *)virt);
3191 * Zero part of a physical page by mapping it into memory and clearing
3192 * its contents with bzero.
3194 * off and size may not cover an area beyond a single hardware page.
3197 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3199 vm_offset_t virt = PHYS_TO_DMAP(phys);
3201 bzero((char *)virt + off, size);
3207 * Copy the physical page from the source PA to the target PA.
3208 * This function may be called from an interrupt. No locking
3212 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3214 vm_offset_t src_virt, dst_virt;
3216 src_virt = PHYS_TO_DMAP(src);
3217 dst_virt = PHYS_TO_DMAP(dst);
3218 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3222 * pmap_copy_page_frag:
3224 * Copy the physical page from the source PA to the target PA.
3225 * This function may be called from an interrupt. No locking
3229 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3231 vm_offset_t src_virt, dst_virt;
3233 src_virt = PHYS_TO_DMAP(src);
3234 dst_virt = PHYS_TO_DMAP(dst);
3236 bcopy((char *)src_virt + (src & PAGE_MASK),
3237 (char *)dst_virt + (dst & PAGE_MASK),
3242 * Returns true if the pmap's pv is one of the first
3243 * 16 pvs linked to from this page. This count may
3244 * be changed upwards or downwards in the future; it
3245 * is only necessary that true be returned for a small
3246 * subset of pmaps for proper page aging.
3249 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3254 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3258 lwkt_gettoken(&vm_token);
3260 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3261 if (pv->pv_pmap == pmap) {
3262 lwkt_reltoken(&vm_token);
3270 lwkt_reltoken(&vm_token);
3276 * Remove all pages from specified address space
3277 * this aids process exit speeds. Also, this code
3278 * is special cased for current process only, but
3279 * can have the more generic (and slightly slower)
3280 * mode enabled. This is much faster than pmap_remove
3281 * in the case of running down an entire address space.
3284 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3287 pt_entry_t *pte, tpte;
3290 pmap_inval_info info;
3292 int save_generation;
3294 lp = curthread->td_lwp;
3295 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3300 lwkt_gettoken(&vm_token);
3301 pmap_inval_init(&info);
3302 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3303 if (pv->pv_va >= eva || pv->pv_va < sva) {
3304 npv = TAILQ_NEXT(pv, pv_plist);
3308 KKASSERT(pmap == pv->pv_pmap);
3311 pte = vtopte(pv->pv_va);
3313 pte = pmap_pte_quick(pmap, pv->pv_va);
3314 pmap_inval_interlock(&info, pmap, pv->pv_va);
3317 * We cannot remove wired pages from a process' mapping
3321 pmap_inval_deinterlock(&info, pmap);
3322 npv = TAILQ_NEXT(pv, pv_plist);
3325 tpte = pte_load_clear(pte);
3327 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3329 KASSERT(m < &vm_page_array[vm_page_array_size],
3330 ("pmap_remove_pages: bad tpte %lx", tpte));
3332 KKASSERT(pmap->pm_stats.resident_count > 0);
3333 --pmap->pm_stats.resident_count;
3334 pmap_inval_deinterlock(&info, pmap);
3337 * Update the vm_page_t clean and reference bits.
3343 npv = TAILQ_NEXT(pv, pv_plist);
3344 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3345 save_generation = ++pmap->pm_generation;
3347 m->md.pv_list_count--;
3348 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3349 if (TAILQ_EMPTY(&m->md.pv_list))
3350 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3352 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3356 * Restart the scan if we blocked during the unuse or free
3357 * calls and other removals were made.
3359 if (save_generation != pmap->pm_generation) {
3360 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3361 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3364 pmap_inval_done(&info);
3365 lwkt_reltoken(&vm_token);
3369 * pmap_testbit tests bits in pte's
3370 * note that the testbit/clearbit routines are inline,
3371 * and a lot of things compile-time evaluate.
3375 pmap_testbit(vm_page_t m, int bit)
3380 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3383 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3388 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3390 * if the bit being tested is the modified bit, then
3391 * mark clean_map and ptes as never
3394 if (bit & (PG_A|PG_M)) {
3395 if (!pmap_track_modified(pv->pv_va))
3399 #if defined(PMAP_DIAGNOSTIC)
3400 if (pv->pv_pmap == NULL) {
3401 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3405 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3416 * this routine is used to modify bits in ptes
3420 pmap_clearbit(vm_page_t m, int bit)
3422 struct pmap_inval_info info;
3427 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3430 pmap_inval_init(&info);
3433 * Loop over all current mappings setting/clearing as appropos If
3434 * setting RO do we need to clear the VAC?
3436 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3438 * don't write protect pager mappings
3441 if (!pmap_track_modified(pv->pv_va))
3445 #if defined(PMAP_DIAGNOSTIC)
3446 if (pv->pv_pmap == NULL) {
3447 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3453 * Careful here. We can use a locked bus instruction to
3454 * clear PG_A or PG_M safely but we need to synchronize
3455 * with the target cpus when we mess with PG_RW.
3457 * We do not have to force synchronization when clearing
3458 * PG_M even for PTEs generated via virtual memory maps,
3459 * because the virtual kernel will invalidate the pmap
3460 * entry when/if it needs to resynchronize the Modify bit.
3463 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
3464 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3471 atomic_clear_long(pte, PG_M|PG_RW);
3474 * The cpu may be trying to set PG_M
3475 * simultaniously with our clearing
3478 if (!atomic_cmpset_long(pte, pbits,
3482 } else if (bit == PG_M) {
3484 * We could also clear PG_RW here to force
3485 * a fault on write to redetect PG_M for
3486 * virtual kernels, but it isn't necessary
3487 * since virtual kernels invalidate the pte
3488 * when they clear the VPTE_M bit in their
3489 * virtual page tables.
3491 atomic_clear_long(pte, PG_M);
3493 atomic_clear_long(pte, bit);
3497 pmap_inval_deinterlock(&info, pv->pv_pmap);
3499 pmap_inval_done(&info);
3503 * pmap_page_protect:
3505 * Lower the permission for all mappings to a given page.
3508 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3510 /* JG NX support? */
3511 if ((prot & VM_PROT_WRITE) == 0) {
3512 lwkt_gettoken(&vm_token);
3513 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3514 pmap_clearbit(m, PG_RW);
3515 vm_page_flag_clear(m, PG_WRITEABLE);
3519 lwkt_reltoken(&vm_token);
3524 pmap_phys_address(vm_pindex_t ppn)
3526 return (x86_64_ptob(ppn));
3530 * pmap_ts_referenced:
3532 * Return a count of reference bits for a page, clearing those bits.
3533 * It is not necessary for every reference bit to be cleared, but it
3534 * is necessary that 0 only be returned when there are truly no
3535 * reference bits set.
3537 * XXX: The exact number of bits to check and clear is a matter that
3538 * should be tested and standardized at some point in the future for
3539 * optimal aging of shared pages.
3542 pmap_ts_referenced(vm_page_t m)
3544 pv_entry_t pv, pvf, pvn;
3548 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3552 lwkt_gettoken(&vm_token);
3554 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3559 pvn = TAILQ_NEXT(pv, pv_list);
3562 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3563 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3566 if (!pmap_track_modified(pv->pv_va))
3569 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3571 if (pte && (*pte & PG_A)) {
3573 atomic_clear_long(pte, PG_A);
3575 atomic_clear_long_nonlocked(pte, PG_A);
3582 } while ((pv = pvn) != NULL && pv != pvf);
3584 lwkt_reltoken(&vm_token);
3593 * Return whether or not the specified physical page was modified
3594 * in any physical maps.
3597 pmap_is_modified(vm_page_t m)
3601 lwkt_gettoken(&vm_token);
3602 res = pmap_testbit(m, PG_M);
3603 lwkt_reltoken(&vm_token);
3608 * Clear the modify bits on the specified physical page.
3611 pmap_clear_modify(vm_page_t m)
3613 lwkt_gettoken(&vm_token);
3614 pmap_clearbit(m, PG_M);
3615 lwkt_reltoken(&vm_token);
3619 * pmap_clear_reference:
3621 * Clear the reference bit on the specified physical page.
3624 pmap_clear_reference(vm_page_t m)
3626 lwkt_gettoken(&vm_token);
3627 pmap_clearbit(m, PG_A);
3628 lwkt_reltoken(&vm_token);
3632 * Miscellaneous support routines follow
3637 i386_protection_init(void)
3641 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3642 kp = protection_codes;
3643 for (prot = 0; prot < 8; prot++) {
3645 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3647 * Read access is also 0. There isn't any execute bit,
3648 * so just make it readable.
3650 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3651 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3652 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3655 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3656 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3657 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3658 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3666 * Map a set of physical memory pages into the kernel virtual
3667 * address space. Return a pointer to where it is mapped. This
3668 * routine is intended to be used for mapping device memory,
3671 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3675 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3677 vm_offset_t va, tmpva, offset;
3680 offset = pa & PAGE_MASK;
3681 size = roundup(offset + size, PAGE_SIZE);
3683 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3685 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3687 pa = pa & ~PAGE_MASK;
3688 for (tmpva = va; size > 0;) {
3689 pte = vtopte(tmpva);
3690 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3698 return ((void *)(va + offset));
3702 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3704 vm_offset_t va, tmpva, offset;
3707 offset = pa & PAGE_MASK;
3708 size = roundup(offset + size, PAGE_SIZE);
3710 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3712 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3714 pa = pa & ~PAGE_MASK;
3715 for (tmpva = va; size > 0;) {
3716 pte = vtopte(tmpva);
3717 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3725 return ((void *)(va + offset));
3729 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3731 vm_offset_t base, offset;
3733 base = va & ~PAGE_MASK;
3734 offset = va & PAGE_MASK;
3735 size = roundup(offset + size, PAGE_SIZE);
3736 pmap_qremove(va, size >> PAGE_SHIFT);
3737 kmem_free(&kernel_map, base, size);
3741 * perform the pmap work for mincore
3744 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3746 pt_entry_t *ptep, pte;
3750 lwkt_gettoken(&vm_token);
3751 ptep = pmap_pte(pmap, addr);
3753 if (ptep && (pte = *ptep) != 0) {
3756 val = MINCORE_INCORE;
3757 if ((pte & PG_MANAGED) == 0)
3760 pa = pte & PG_FRAME;
3762 m = PHYS_TO_VM_PAGE(pa);
3768 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3770 * Modified by someone
3772 else if (m->dirty || pmap_is_modified(m))
3773 val |= MINCORE_MODIFIED_OTHER;
3778 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3781 * Referenced by someone
3783 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3784 val |= MINCORE_REFERENCED_OTHER;
3785 vm_page_flag_set(m, PG_REFERENCED);
3789 lwkt_reltoken(&vm_token);
3794 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3795 * vmspace will be ref'd and the old one will be deref'd.
3797 * The vmspace for all lwps associated with the process will be adjusted
3798 * and cr3 will be reloaded if any lwp is the current lwp.
3801 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3803 struct vmspace *oldvm;
3807 oldvm = p->p_vmspace;
3808 if (oldvm != newvm) {
3809 p->p_vmspace = newvm;
3810 KKASSERT(p->p_nthreads == 1);
3811 lp = RB_ROOT(&p->p_lwp_tree);
3812 pmap_setlwpvm(lp, newvm);
3814 sysref_get(&newvm->vm_sysref);
3815 sysref_put(&oldvm->vm_sysref);
3822 * Set the vmspace for a LWP. The vmspace is almost universally set the
3823 * same as the process vmspace, but virtual kernels need to swap out contexts
3824 * on a per-lwp basis.
3827 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3829 struct vmspace *oldvm;
3833 oldvm = lp->lwp_vmspace;
3835 if (oldvm != newvm) {
3836 lp->lwp_vmspace = newvm;
3837 if (curthread->td_lwp == lp) {
3838 pmap = vmspace_pmap(newvm);
3840 atomic_set_int(&pmap->pm_active, mycpu->gd_cpumask);
3841 if (pmap->pm_active & CPUMASK_LOCK)
3842 pmap_interlock_wait(newvm);
3844 pmap->pm_active |= 1;
3846 #if defined(SWTCH_OPTIM_STATS)
3849 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
3850 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
3851 load_cr3(curthread->td_pcb->pcb_cr3);
3852 pmap = vmspace_pmap(oldvm);
3854 atomic_clear_int(&pmap->pm_active, mycpu->gd_cpumask);
3856 pmap->pm_active &= ~1;
3866 * Called when switching to a locked pmap
3869 pmap_interlock_wait(struct vmspace *vm)
3871 struct pmap *pmap = &vm->vm_pmap;
3873 if (pmap->pm_active & CPUMASK_LOCK) {
3874 while (pmap->pm_active & CPUMASK_LOCK) {
3877 lwkt_process_ipiq();
3885 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3888 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3892 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3897 * Used by kmalloc/kfree, page already exists at va
3900 pmap_kvtom(vm_offset_t va)
3902 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));