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 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
442 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
445 nkpt = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
446 nkpt += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E + ndmpdp) +
453 KPTbase = allocpages(firstaddr, nkpt);
454 KPTphys = allocpages(firstaddr, nkpt);
455 KPML4phys = allocpages(firstaddr, 1);
456 KPDPphys = allocpages(firstaddr, NKPML4E);
459 * Calculate the page directory base for KERNBASE,
460 * that is where we start populating the page table pages.
461 * Basically this is the end - 2.
463 KPDphys = allocpages(firstaddr, NKPDPE);
464 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
466 DMPDPphys = allocpages(firstaddr, NDMPML4E);
467 if ((amd_feature & AMDID_PAGE1GB) == 0)
468 DMPDphys = allocpages(firstaddr, ndmpdp);
469 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
472 * Fill in the underlying page table pages for the area around
473 * KERNBASE. This remaps low physical memory to KERNBASE.
475 * Read-only from zero to physfree
476 * XXX not fully used, underneath 2M pages
478 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
479 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
480 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
484 * Now map the initial kernel page tables. One block of page
485 * tables is placed at the beginning of kernel virtual memory,
486 * and another block is placed at KERNBASE to map the kernel binary,
487 * data, bss, and initial pre-allocations.
489 for (i = 0; i < nkpt; i++) {
490 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
491 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
493 for (i = 0; i < nkpt; i++) {
494 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
495 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
499 * Map from zero to end of allocations using 2M pages as an
500 * optimization. This will bypass some of the KPTBase pages
501 * above in the KERNBASE area.
503 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
504 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
505 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
509 * And connect up the PD to the PDP. The kernel pmap is expected
510 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
512 for (i = 0; i < NKPDPE; i++) {
513 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
514 KPDphys + (i << PAGE_SHIFT);
515 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
519 /* Now set up the direct map space using either 2MB or 1GB pages */
520 /* Preset PG_M and PG_A because demotion expects it */
521 if ((amd_feature & AMDID_PAGE1GB) == 0) {
522 for (i = 0; i < NPDEPG * ndmpdp; i++) {
523 ((pd_entry_t *)DMPDphys)[i] = (vm_paddr_t)i << PDRSHIFT;
524 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
527 /* And the direct map space's PDP */
528 for (i = 0; i < ndmpdp; i++) {
529 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
531 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
534 for (i = 0; i < ndmpdp; i++) {
535 ((pdp_entry_t *)DMPDPphys)[i] =
536 (vm_paddr_t)i << PDPSHIFT;
537 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
542 /* And recursively map PML4 to itself in order to get PTmap */
543 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
544 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
546 /* Connect the Direct Map slot up to the PML4 */
547 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
548 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
550 /* Connect the KVA slot up to the PML4 */
551 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
552 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
556 * Bootstrap the system enough to run with virtual memory.
558 * On the i386 this is called after mapping has already been enabled
559 * and just syncs the pmap module with what has already been done.
560 * [We can't call it easily with mapping off since the kernel is not
561 * mapped with PA == VA, hence we would have to relocate every address
562 * from the linked base (virtual) address "KERNBASE" to the actual
563 * (physical) address starting relative to 0]
566 pmap_bootstrap(vm_paddr_t *firstaddr)
570 struct mdglobaldata *gd;
573 KvaStart = VM_MIN_KERNEL_ADDRESS;
574 KvaEnd = VM_MAX_KERNEL_ADDRESS;
575 KvaSize = KvaEnd - KvaStart;
577 avail_start = *firstaddr;
580 * Create an initial set of page tables to run the kernel in.
582 create_pagetables(firstaddr);
584 virtual2_start = KvaStart;
585 virtual2_end = PTOV_OFFSET;
587 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
588 virtual_start = pmap_kmem_choose(virtual_start);
590 virtual_end = VM_MAX_KERNEL_ADDRESS;
592 /* XXX do %cr0 as well */
593 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
597 * Initialize protection array.
599 i386_protection_init();
602 * The kernel's pmap is statically allocated so we don't have to use
603 * pmap_create, which is unlikely to work correctly at this part of
604 * the boot sequence (XXX and which no longer exists).
606 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
607 kernel_pmap.pm_count = 1;
608 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
609 TAILQ_INIT(&kernel_pmap.pm_pvlist);
612 * Reserve some special page table entries/VA space for temporary
615 #define SYSMAP(c, p, v, n) \
616 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
622 * CMAP1/CMAP2 are used for zeroing and copying pages.
624 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
629 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
632 * ptvmmap is used for reading arbitrary physical pages via
635 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
638 * msgbufp is used to map the system message buffer.
639 * XXX msgbufmap is not used.
641 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
642 atop(round_page(MSGBUF_SIZE)))
649 * PG_G is terribly broken on SMP because we IPI invltlb's in some
650 * cases rather then invl1pg. Actually, I don't even know why it
651 * works under UP because self-referential page table mappings
656 if (cpu_feature & CPUID_PGE)
661 * Initialize the 4MB page size flag
665 * The 4MB page version of the initial
666 * kernel page mapping.
670 #if !defined(DISABLE_PSE)
671 if (cpu_feature & CPUID_PSE) {
674 * Note that we have enabled PSE mode
677 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
678 ptditmp &= ~(NBPDR - 1);
679 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
684 * Enable the PSE mode. If we are SMP we can't do this
685 * now because the APs will not be able to use it when
688 load_cr4(rcr4() | CR4_PSE);
691 * We can do the mapping here for the single processor
692 * case. We simply ignore the old page table page from
696 * For SMP, we still need 4K pages to bootstrap APs,
697 * PSE will be enabled as soon as all APs are up.
699 PTD[KPTDI] = (pd_entry_t)ptditmp;
706 * We need to finish setting up the globaldata page for the BSP.
707 * locore has already populated the page table for the mdglobaldata
710 pg = MDGLOBALDATA_BASEALLOC_PAGES;
711 gd = &CPU_prvspace[0].mdglobaldata;
712 gd->gd_CMAP1 = &SMPpt[pg + 0];
713 gd->gd_CMAP2 = &SMPpt[pg + 1];
714 gd->gd_CMAP3 = &SMPpt[pg + 2];
715 gd->gd_PMAP1 = &SMPpt[pg + 3];
716 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
717 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
718 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
719 gd->gd_PADDR1 = (pt_entry_t *)CPU_prvspace[0].PPAGE1;
726 * Set 4mb pdir for mp startup
731 if (pseflag && (cpu_feature & CPUID_PSE)) {
732 load_cr4(rcr4() | CR4_PSE);
733 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
741 * Initialize the pmap module.
742 * Called by vm_init, to initialize any structures that the pmap
743 * system needs to map virtual memory.
744 * pmap_init has been enhanced to support in a fairly consistant
745 * way, discontiguous physical memory.
754 * object for kernel page table pages
756 /* JG I think the number can be arbitrary */
757 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
760 * Allocate memory for random pmap data structures. Includes the
764 for(i = 0; i < vm_page_array_size; i++) {
767 m = &vm_page_array[i];
768 TAILQ_INIT(&m->md.pv_list);
769 m->md.pv_list_count = 0;
773 * init the pv free list
775 initial_pvs = vm_page_array_size;
776 if (initial_pvs < MINPV)
778 pvzone = &pvzone_store;
779 pvinit = (void *)kmem_alloc(&kernel_map,
780 initial_pvs * sizeof (struct pv_entry));
781 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
782 pvinit, initial_pvs);
785 * Now it is safe to enable pv_table recording.
787 pmap_initialized = TRUE;
791 * Initialize the address space (zone) for the pv_entries. Set a
792 * high water mark so that the system can recover from excessive
793 * numbers of pv entries.
798 int shpgperproc = PMAP_SHPGPERPROC;
801 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
802 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
803 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
804 pv_entry_high_water = 9 * (pv_entry_max / 10);
807 * Subtract out pages already installed in the zone (hack)
809 entry_max = pv_entry_max - vm_page_array_size;
813 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
817 /***************************************************
818 * Low level helper routines.....
819 ***************************************************/
821 #if defined(PMAP_DIAGNOSTIC)
824 * This code checks for non-writeable/modified pages.
825 * This should be an invalid condition.
829 pmap_nw_modified(pt_entry_t pte)
831 if ((pte & (PG_M|PG_RW)) == PG_M)
840 * this routine defines the region(s) of memory that should
841 * not be tested for the modified bit.
845 pmap_track_modified(vm_offset_t va)
847 if ((va < clean_sva) || (va >= clean_eva))
854 * Extract the physical page address associated with the map/VA pair.
856 * The caller must hold vm_token if non-blocking operation is desired.
859 pmap_extract(pmap_t pmap, vm_offset_t va)
863 pd_entry_t pde, *pdep;
865 lwkt_gettoken(&vm_token);
867 pdep = pmap_pde(pmap, va);
871 if ((pde & PG_PS) != 0) {
872 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
874 pte = pmap_pde_to_pte(pdep, va);
875 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
879 lwkt_reltoken(&vm_token);
884 * Extract the physical page address associated kernel virtual address.
887 pmap_kextract(vm_offset_t va)
892 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
893 pa = DMAP_TO_PHYS(va);
897 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
900 * Beware of a concurrent promotion that changes the
901 * PDE at this point! For example, vtopte() must not
902 * be used to access the PTE because it would use the
903 * new PDE. It is, however, safe to use the old PDE
904 * because the page table page is preserved by the
907 pa = *pmap_pde_to_pte(&pde, va);
908 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
914 /***************************************************
915 * Low level mapping routines.....
916 ***************************************************/
919 * Routine: pmap_kenter
921 * Add a wired page to the KVA
922 * NOTE! note that in order for the mapping to take effect -- you
923 * should do an invltlb after doing the pmap_kenter().
926 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
930 pmap_inval_info info;
932 pmap_inval_init(&info);
933 npte = pa | PG_RW | PG_V | pgeflag;
935 pmap_inval_interlock(&info, &kernel_pmap, va);
937 pmap_inval_deinterlock(&info, &kernel_pmap);
938 pmap_inval_done(&info);
942 * Routine: pmap_kenter_quick
944 * Similar to pmap_kenter(), except we only invalidate the
945 * mapping on the current CPU.
948 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
953 npte = pa | PG_RW | PG_V | pgeflag;
956 cpu_invlpg((void *)va);
960 pmap_kenter_sync(vm_offset_t va)
962 pmap_inval_info info;
964 pmap_inval_init(&info);
965 pmap_inval_interlock(&info, &kernel_pmap, va);
966 pmap_inval_deinterlock(&info, &kernel_pmap);
967 pmap_inval_done(&info);
971 pmap_kenter_sync_quick(vm_offset_t va)
973 cpu_invlpg((void *)va);
977 * remove a page from the kernel pagetables
980 pmap_kremove(vm_offset_t va)
983 pmap_inval_info info;
985 pmap_inval_init(&info);
987 pmap_inval_interlock(&info, &kernel_pmap, va);
989 pmap_inval_deinterlock(&info, &kernel_pmap);
990 pmap_inval_done(&info);
994 pmap_kremove_quick(vm_offset_t va)
999 cpu_invlpg((void *)va);
1003 * XXX these need to be recoded. They are not used in any critical path.
1006 pmap_kmodify_rw(vm_offset_t va)
1008 *vtopte(va) |= PG_RW;
1009 cpu_invlpg((void *)va);
1013 pmap_kmodify_nc(vm_offset_t va)
1015 *vtopte(va) |= PG_N;
1016 cpu_invlpg((void *)va);
1020 * Used to map a range of physical addresses into kernel virtual
1021 * address space during the low level boot, typically to map the
1022 * dump bitmap, message buffer, and vm_page_array.
1024 * These mappings are typically made at some pointer after the end of the
1027 * We could return PHYS_TO_DMAP(start) here and not allocate any
1028 * via (*virtp), but then kmem from userland and kernel dumps won't
1029 * have access to the related pointers.
1032 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1035 vm_offset_t va_start;
1037 /*return PHYS_TO_DMAP(start);*/
1042 while (start < end) {
1043 pmap_kenter_quick(va, start);
1053 * Add a list of wired pages to the kva
1054 * this routine is only used for temporary
1055 * kernel mappings that do not need to have
1056 * page modification or references recorded.
1057 * Note that old mappings are simply written
1058 * over. The page *must* be wired.
1061 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1065 end_va = va + count * PAGE_SIZE;
1067 while (va < end_va) {
1071 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1072 cpu_invlpg((void *)va);
1077 smp_invltlb(); /* XXX */
1082 * This routine jerks page mappings from the
1083 * kernel -- it is meant only for temporary mappings.
1085 * MPSAFE, INTERRUPT SAFE (cluster callback)
1088 pmap_qremove(vm_offset_t va, int count)
1092 end_va = va + count * PAGE_SIZE;
1094 while (va < end_va) {
1099 cpu_invlpg((void *)va);
1108 * This routine works like vm_page_lookup() but also blocks as long as the
1109 * page is busy. This routine does not busy the page it returns.
1111 * Unless the caller is managing objects whos pages are in a known state,
1112 * the call should be made with a critical section held so the page's object
1113 * association remains valid on return.
1117 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1122 m = vm_page_lookup(object, pindex);
1123 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
1129 * Create a new thread and optionally associate it with a (new) process.
1130 * NOTE! the new thread's cpu may not equal the current cpu.
1133 pmap_init_thread(thread_t td)
1135 /* enforce pcb placement */
1136 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1137 td->td_savefpu = &td->td_pcb->pcb_save;
1138 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on x86_64? */
1142 * This routine directly affects the fork perf for a process.
1145 pmap_init_proc(struct proc *p)
1150 * Dispose the UPAGES for a process that has exited.
1151 * This routine directly impacts the exit perf of a process.
1154 pmap_dispose_proc(struct proc *p)
1156 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1159 /***************************************************
1160 * Page table page management routines.....
1161 ***************************************************/
1164 * This routine unholds page table pages, and if the hold count
1165 * drops to zero, then it decrements the wire count.
1169 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1170 pmap_inval_info_t info)
1172 KKASSERT(m->hold_count > 0);
1173 if (m->hold_count > 1) {
1177 return _pmap_unwire_pte_hold(pmap, va, m, info);
1183 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1184 pmap_inval_info_t info)
1187 * Wait until we can busy the page ourselves. We cannot have
1188 * any active flushes if we block. We own one hold count on the
1189 * page so it cannot be freed out from under us.
1191 if (m->flags & PG_BUSY) {
1192 pmap_inval_flush(info);
1193 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1196 KASSERT(m->queue == PQ_NONE,
1197 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1200 * This case can occur if new references were acquired while
1203 if (m->hold_count > 1) {
1204 KKASSERT(m->hold_count > 1);
1210 * Unmap the page table page
1212 KKASSERT(m->hold_count == 1);
1214 pmap_inval_interlock(info, pmap, -1);
1216 if (m->pindex >= (NUPDE + NUPDPE)) {
1219 pml4 = pmap_pml4e(pmap, va);
1221 } else if (m->pindex >= NUPDE) {
1224 pdp = pmap_pdpe(pmap, va);
1229 pd = pmap_pde(pmap, va);
1233 KKASSERT(pmap->pm_stats.resident_count > 0);
1234 --pmap->pm_stats.resident_count;
1236 if (pmap->pm_ptphint == m)
1237 pmap->pm_ptphint = NULL;
1238 pmap_inval_deinterlock(info, pmap);
1240 if (m->pindex < NUPDE) {
1241 /* We just released a PT, unhold the matching PD */
1244 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
1245 pmap_unwire_pte_hold(pmap, va, pdpg, info);
1247 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
1248 /* We just released a PD, unhold the matching PDP */
1251 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
1252 pmap_unwire_pte_hold(pmap, va, pdppg, info);
1256 * This was our last hold, the page had better be unwired
1257 * after we decrement wire_count.
1259 * FUTURE NOTE: shared page directory page could result in
1260 * multiple wire counts.
1264 KKASSERT(m->wire_count == 0);
1265 --vmstats.v_wire_count;
1266 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1268 vm_page_free_zero(m);
1274 * After removing a page table entry, this routine is used to
1275 * conditionally free the page, and manage the hold/wire counts.
1279 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1280 pmap_inval_info_t info)
1282 vm_pindex_t ptepindex;
1284 if (va >= VM_MAX_USER_ADDRESS)
1288 ptepindex = pmap_pde_pindex(va);
1290 if (pmap->pm_ptphint &&
1291 (pmap->pm_ptphint->pindex == ptepindex)) {
1292 mpte = pmap->pm_ptphint;
1295 pmap_inval_flush(info);
1296 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1297 pmap->pm_ptphint = mpte;
1302 return pmap_unwire_pte_hold(pmap, va, mpte, info);
1306 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1307 * it, and IdlePTD, represents the template used to update all other pmaps.
1309 * On architectures where the kernel pmap is not integrated into the user
1310 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1311 * kernel_pmap should be used to directly access the kernel_pmap.
1314 pmap_pinit0(struct pmap *pmap)
1316 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1318 pmap->pm_active = 0;
1319 pmap->pm_ptphint = NULL;
1320 TAILQ_INIT(&pmap->pm_pvlist);
1321 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1325 * Initialize a preallocated and zeroed pmap structure,
1326 * such as one in a vmspace structure.
1329 pmap_pinit(struct pmap *pmap)
1334 * No need to allocate page table space yet but we do need a valid
1335 * page directory table.
1337 if (pmap->pm_pml4 == NULL) {
1339 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1343 * Allocate an object for the ptes
1345 if (pmap->pm_pteobj == NULL)
1346 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1);
1349 * Allocate the page directory page, unless we already have
1350 * one cached. If we used the cached page the wire_count will
1351 * already be set appropriately.
1353 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1354 ptdpg = vm_page_grab(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I,
1355 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1356 pmap->pm_pdirm = ptdpg;
1357 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1358 ptdpg->valid = VM_PAGE_BITS_ALL;
1359 if (ptdpg->wire_count == 0)
1360 ++vmstats.v_wire_count;
1361 ptdpg->wire_count = 1;
1362 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1364 if ((ptdpg->flags & PG_ZERO) == 0)
1365 bzero(pmap->pm_pml4, PAGE_SIZE);
1367 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1368 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1370 /* install self-referential address mapping entry */
1371 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1374 pmap->pm_active = 0;
1375 pmap->pm_ptphint = NULL;
1376 TAILQ_INIT(&pmap->pm_pvlist);
1377 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1378 pmap->pm_stats.resident_count = 1;
1382 * Clean up a pmap structure so it can be physically freed. This routine
1383 * is called by the vmspace dtor function. A great deal of pmap data is
1384 * left passively mapped to improve vmspace management so we have a bit
1385 * of cleanup work to do here.
1388 pmap_puninit(pmap_t pmap)
1392 KKASSERT(pmap->pm_active == 0);
1393 lwkt_gettoken(&vm_token);
1394 if ((p = pmap->pm_pdirm) != NULL) {
1395 KKASSERT(pmap->pm_pml4 != NULL);
1396 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1397 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1399 vmstats.v_wire_count--;
1400 KKASSERT((p->flags & PG_BUSY) == 0);
1402 vm_page_free_zero(p);
1403 pmap->pm_pdirm = NULL;
1405 if (pmap->pm_pml4) {
1406 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1407 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1408 pmap->pm_pml4 = NULL;
1410 if (pmap->pm_pteobj) {
1411 vm_object_deallocate(pmap->pm_pteobj);
1412 pmap->pm_pteobj = NULL;
1414 lwkt_reltoken(&vm_token);
1418 * Wire in kernel global address entries. To avoid a race condition
1419 * between pmap initialization and pmap_growkernel, this procedure
1420 * adds the pmap to the master list (which growkernel scans to update),
1421 * then copies the template.
1424 pmap_pinit2(struct pmap *pmap)
1427 lwkt_gettoken(&vm_token);
1428 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1429 /* XXX copies current process, does not fill in MPPTDI */
1430 lwkt_reltoken(&vm_token);
1435 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1436 * 0 on failure (if the procedure had to sleep).
1438 * When asked to remove the page directory page itself, we actually just
1439 * leave it cached so we do not have to incur the SMP inval overhead of
1440 * removing the kernel mapping. pmap_puninit() will take care of it.
1444 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1447 * This code optimizes the case of freeing non-busy
1448 * page-table pages. Those pages are zero now, and
1449 * might as well be placed directly into the zero queue.
1451 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1457 * Remove the page table page from the processes address space.
1459 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1461 * We are the pml4 table itself.
1463 /* XXX anything to do here? */
1464 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1466 * Remove a PDP page from the PML4. We do not maintain
1467 * hold counts on the PML4 page.
1473 m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1474 KKASSERT(m4 != NULL);
1475 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1476 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1477 KKASSERT(pml4[idx] != 0);
1479 } else if (p->pindex >= NUPDE) {
1481 * Remove a PD page from the PDP and drop the hold count
1482 * on the PDP. The PDP is left cached in the pmap if
1483 * the hold count drops to 0 so the wire count remains
1490 m3 = vm_page_lookup(pmap->pm_pteobj,
1491 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1492 KKASSERT(m3 != NULL);
1493 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1494 idx = (p->pindex - NUPDE) % NPDPEPG;
1495 KKASSERT(pdp[idx] != 0);
1500 * Remove a PT page from the PD and drop the hold count
1501 * on the PD. The PD is left cached in the pmap if
1502 * the hold count drops to 0 so the wire count remains
1509 m2 = vm_page_lookup(pmap->pm_pteobj,
1510 NUPDE + p->pindex / NPDEPG);
1511 KKASSERT(m2 != NULL);
1512 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1513 idx = p->pindex % NPDEPG;
1519 * One fewer mappings in the pmap. p's hold count had better
1522 KKASSERT(pmap->pm_stats.resident_count > 0);
1523 --pmap->pm_stats.resident_count;
1525 panic("pmap_release: freeing held page table page");
1526 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1527 pmap->pm_ptphint = NULL;
1530 * We leave the top-level page table page cached, wired, and mapped in
1531 * the pmap until the dtor function (pmap_puninit()) gets called.
1532 * However, still clean it up so we can set PG_ZERO.
1534 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1535 bzero(pmap->pm_pml4, PAGE_SIZE);
1536 vm_page_flag_set(p, PG_ZERO);
1540 KKASSERT(p->wire_count == 0);
1541 vmstats.v_wire_count--;
1542 /* JG eventually revert to using vm_page_free_zero() */
1549 * This routine is called when various levels in the page table need to
1550 * be populated. This routine cannot fail.
1554 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1559 * Find or fabricate a new pagetable page. This will busy the page.
1561 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1562 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1563 if ((m->flags & PG_ZERO) == 0) {
1564 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1567 KASSERT(m->queue == PQ_NONE,
1568 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1571 * Increment the hold count for the page we will be returning to
1575 if (m->wire_count++ == 0)
1576 vmstats.v_wire_count++;
1579 * Map the pagetable page into the process address space, if
1580 * it isn't already there.
1582 * It is possible that someone else got in and mapped the page
1583 * directory page while we were blocked, if so just unbusy and
1584 * return the held page.
1586 if (ptepindex >= (NUPDE + NUPDPE)) {
1588 * Wire up a new PDP page in the PML4
1590 vm_pindex_t pml4index;
1593 pml4index = ptepindex - (NUPDE + NUPDPE);
1594 pml4 = &pmap->pm_pml4[pml4index];
1596 if (--m->wire_count == 0)
1597 --vmstats.v_wire_count;
1601 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1602 } else if (ptepindex >= NUPDE) {
1604 * Wire up a new PD page in the PDP
1606 vm_pindex_t pml4index;
1607 vm_pindex_t pdpindex;
1612 pdpindex = ptepindex - NUPDE;
1613 pml4index = pdpindex >> NPML4EPGSHIFT;
1615 pml4 = &pmap->pm_pml4[pml4index];
1616 if ((*pml4 & PG_V) == 0) {
1618 * Have to allocate a new PDP page, recurse.
1619 * This always succeeds. Returned page will
1622 pdppg = _pmap_allocpte(pmap,
1623 NUPDE + NUPDPE + pml4index);
1626 * Add a held reference to the PDP page.
1628 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1629 pdppg->hold_count++;
1633 * Now find the pdp_entry and map the PDP. If the PDP
1634 * has already been mapped unwind and return the
1635 * already-mapped PDP held.
1637 * pdppg is left held (hold_count is incremented for
1638 * each PD in the PDP).
1640 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1641 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1643 vm_page_unhold(pdppg);
1644 if (--m->wire_count == 0)
1645 --vmstats.v_wire_count;
1649 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1652 * Wire up the new PT page in the PD
1654 vm_pindex_t pml4index;
1655 vm_pindex_t pdpindex;
1661 pdpindex = ptepindex >> NPDPEPGSHIFT;
1662 pml4index = pdpindex >> NPML4EPGSHIFT;
1665 * Locate the PDP page in the PML4, then the PD page in
1666 * the PDP. If either does not exist we simply recurse
1669 * We can just recurse on the PD page as it will recurse
1670 * on the PDP if necessary.
1672 pml4 = &pmap->pm_pml4[pml4index];
1673 if ((*pml4 & PG_V) == 0) {
1674 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1675 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1676 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1678 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1679 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1680 if ((*pdp & PG_V) == 0) {
1681 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1683 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1689 * Now fill in the pte in the PD. If the pte already exists
1690 * (again, if we raced the grab), unhold pdpg and unwire
1691 * m, returning a held m.
1693 * pdpg is left held (hold_count is incremented for
1694 * each PT in the PD).
1696 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
1697 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1699 vm_page_unhold(pdpg);
1700 if (--m->wire_count == 0)
1701 --vmstats.v_wire_count;
1705 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1709 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1710 * valid bits, mapped flag, unbusy, and we're done.
1712 pmap->pm_ptphint = m;
1713 ++pmap->pm_stats.resident_count;
1715 m->valid = VM_PAGE_BITS_ALL;
1716 vm_page_flag_clear(m, PG_ZERO);
1717 vm_page_flag_set(m, PG_MAPPED);
1725 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1727 vm_pindex_t ptepindex;
1732 * Calculate pagetable page index
1734 ptepindex = pmap_pde_pindex(va);
1737 * Get the page directory entry
1739 pd = pmap_pde(pmap, va);
1742 * This supports switching from a 2MB page to a
1745 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1746 panic("no promotion/demotion yet");
1754 * If the page table page is mapped, we just increment the
1755 * hold count, and activate it.
1757 if (pd != NULL && (*pd & PG_V) != 0) {
1758 /* YYY hint is used here on i386 */
1759 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1760 pmap->pm_ptphint = m;
1765 * Here if the pte page isn't mapped, or if it has been deallocated.
1767 return _pmap_allocpte(pmap, ptepindex);
1771 /***************************************************
1772 * Pmap allocation/deallocation routines.
1773 ***************************************************/
1776 * Release any resources held by the given physical map.
1777 * Called when a pmap initialized by pmap_pinit is being released.
1778 * Should only be called if the map contains no valid mappings.
1780 static int pmap_release_callback(struct vm_page *p, void *data);
1783 pmap_release(struct pmap *pmap)
1785 vm_object_t object = pmap->pm_pteobj;
1786 struct rb_vm_page_scan_info info;
1788 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1789 #if defined(DIAGNOSTIC)
1790 if (object->ref_count != 1)
1791 panic("pmap_release: pteobj reference count != 1");
1795 info.object = object;
1797 lwkt_gettoken(&vm_token);
1798 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1805 info.limit = object->generation;
1807 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1808 pmap_release_callback, &info);
1809 if (info.error == 0 && info.mpte) {
1810 if (!pmap_release_free_page(pmap, info.mpte))
1814 } while (info.error);
1815 lwkt_reltoken(&vm_token);
1820 pmap_release_callback(struct vm_page *p, void *data)
1822 struct rb_vm_page_scan_info *info = data;
1824 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1828 if (!pmap_release_free_page(info->pmap, p)) {
1832 if (info->object->generation != info->limit) {
1840 * Grow the number of kernel page table entries, if needed.
1842 * This routine is always called to validate any address space
1843 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
1844 * space below KERNBASE.
1847 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1850 vm_offset_t ptppaddr;
1852 pd_entry_t *pde, newpdir;
1854 int update_kernel_vm_end;
1857 lwkt_gettoken(&vm_token);
1860 * bootstrap kernel_vm_end on first real VM use
1862 if (kernel_vm_end == 0) {
1863 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1865 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1866 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1867 ~(PAGE_SIZE * NPTEPG - 1);
1869 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1870 kernel_vm_end = kernel_map.max_offset;
1877 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
1878 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
1879 * do not want to force-fill 128G worth of page tables.
1881 if (kstart < KERNBASE) {
1882 if (kstart > kernel_vm_end)
1883 kstart = kernel_vm_end;
1884 KKASSERT(kend <= KERNBASE);
1885 update_kernel_vm_end = 1;
1887 update_kernel_vm_end = 0;
1890 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
1891 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
1893 if (kend - 1 >= kernel_map.max_offset)
1894 kend = kernel_map.max_offset;
1896 while (kstart < kend) {
1897 pde = pmap_pde(&kernel_pmap, kstart);
1899 /* We need a new PDP entry */
1900 nkpg = vm_page_alloc(kptobj, nkpt,
1903 VM_ALLOC_INTERRUPT);
1905 panic("pmap_growkernel: no memory to grow "
1908 paddr = VM_PAGE_TO_PHYS(nkpg);
1909 if ((nkpg->flags & PG_ZERO) == 0)
1910 pmap_zero_page(paddr);
1911 vm_page_flag_clear(nkpg, PG_ZERO);
1912 newpdp = (pdp_entry_t)
1913 (paddr | PG_V | PG_RW | PG_A | PG_M);
1914 *pmap_pdpe(&kernel_pmap, kstart) = newpdp;
1916 continue; /* try again */
1918 if ((*pde & PG_V) != 0) {
1919 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1920 ~(PAGE_SIZE * NPTEPG - 1);
1921 if (kstart - 1 >= kernel_map.max_offset) {
1922 kstart = kernel_map.max_offset;
1929 * This index is bogus, but out of the way
1931 nkpg = vm_page_alloc(kptobj, nkpt,
1934 VM_ALLOC_INTERRUPT);
1936 panic("pmap_growkernel: no memory to grow kernel");
1939 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1940 pmap_zero_page(ptppaddr);
1941 vm_page_flag_clear(nkpg, PG_ZERO);
1942 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1943 *pmap_pde(&kernel_pmap, kstart) = newpdir;
1946 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1947 ~(PAGE_SIZE * NPTEPG - 1);
1949 if (kstart - 1 >= kernel_map.max_offset) {
1950 kstart = kernel_map.max_offset;
1956 * Only update kernel_vm_end for areas below KERNBASE.
1958 if (update_kernel_vm_end && kernel_vm_end < kstart)
1959 kernel_vm_end = kstart;
1961 lwkt_reltoken(&vm_token);
1966 * Retire the given physical map from service.
1967 * Should only be called if the map contains
1968 * no valid mappings.
1971 pmap_destroy(pmap_t pmap)
1978 lwkt_gettoken(&vm_token);
1979 count = --pmap->pm_count;
1982 panic("destroying a pmap is not yet implemented");
1984 lwkt_reltoken(&vm_token);
1988 * Add a reference to the specified pmap.
1991 pmap_reference(pmap_t pmap)
1994 lwkt_gettoken(&vm_token);
1996 lwkt_reltoken(&vm_token);
2000 /***************************************************
2001 * page management routines.
2002 ***************************************************/
2005 * free the pv_entry back to the free list. This function may be
2006 * called from an interrupt.
2010 free_pv_entry(pv_entry_t pv)
2013 KKASSERT(pv_entry_count >= 0);
2018 * get a new pv_entry, allocating a block from the system
2019 * when needed. This function may be called from an interrupt.
2026 if (pv_entry_high_water &&
2027 (pv_entry_count > pv_entry_high_water) &&
2028 (pmap_pagedaemon_waken == 0)) {
2029 pmap_pagedaemon_waken = 1;
2030 wakeup(&vm_pages_needed);
2032 return zalloc(pvzone);
2036 * This routine is very drastic, but can save the system
2044 static int warningdone=0;
2046 if (pmap_pagedaemon_waken == 0)
2048 lwkt_gettoken(&vm_token);
2049 if (warningdone < 5) {
2050 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
2054 for(i = 0; i < vm_page_array_size; i++) {
2055 m = &vm_page_array[i];
2056 if (m->wire_count || m->hold_count || m->busy ||
2057 (m->flags & PG_BUSY))
2061 pmap_pagedaemon_waken = 0;
2062 lwkt_reltoken(&vm_token);
2067 * If it is the first entry on the list, it is actually
2068 * in the header and we must copy the following entry up
2069 * to the header. Otherwise we must search the list for
2070 * the entry. In either case we free the now unused entry.
2074 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
2075 vm_offset_t va, pmap_inval_info_t info)
2081 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
2082 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2083 if (pmap == pv->pv_pmap && va == pv->pv_va)
2087 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2088 if (va == pv->pv_va)
2096 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2097 m->md.pv_list_count--;
2098 KKASSERT(m->md.pv_list_count >= 0);
2099 if (TAILQ_EMPTY(&m->md.pv_list))
2100 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2101 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2102 ++pmap->pm_generation;
2103 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
2111 * Create a pv entry for page at pa for
2116 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2121 pv = get_pv_entry();
2126 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2127 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2128 ++pmap->pm_generation;
2129 m->md.pv_list_count++;
2135 * pmap_remove_pte: do the things to unmap a page in a process
2139 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2140 pmap_inval_info_t info)
2145 pmap_inval_interlock(info, pmap, va);
2146 oldpte = pte_load_clear(ptq);
2147 pmap_inval_deinterlock(info, pmap);
2149 pmap->pm_stats.wired_count -= 1;
2151 * Machines that don't support invlpg, also don't support
2152 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2156 cpu_invlpg((void *)va);
2157 KKASSERT(pmap->pm_stats.resident_count > 0);
2158 --pmap->pm_stats.resident_count;
2159 if (oldpte & PG_MANAGED) {
2160 m = PHYS_TO_VM_PAGE(oldpte);
2161 if (oldpte & PG_M) {
2162 #if defined(PMAP_DIAGNOSTIC)
2163 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2165 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2169 if (pmap_track_modified(va))
2173 vm_page_flag_set(m, PG_REFERENCED);
2174 return pmap_remove_entry(pmap, m, va, info);
2176 return pmap_unuse_pt(pmap, va, NULL, info);
2185 * Remove a single page from a process address space.
2187 * This function may not be called from an interrupt if the pmap is
2192 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2196 pte = pmap_pte(pmap, va);
2199 if ((*pte & PG_V) == 0)
2201 pmap_remove_pte(pmap, pte, va, info);
2207 * Remove the given range of addresses from the specified map.
2209 * It is assumed that the start and end are properly
2210 * rounded to the page size.
2212 * This function may not be called from an interrupt if the pmap is
2216 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2218 vm_offset_t va_next;
2219 pml4_entry_t *pml4e;
2221 pd_entry_t ptpaddr, *pde;
2223 struct pmap_inval_info info;
2228 lwkt_gettoken(&vm_token);
2229 if (pmap->pm_stats.resident_count == 0) {
2230 lwkt_reltoken(&vm_token);
2234 pmap_inval_init(&info);
2237 * special handling of removing one page. a very
2238 * common operation and easy to short circuit some
2241 if (sva + PAGE_SIZE == eva) {
2242 pde = pmap_pde(pmap, sva);
2243 if (pde && (*pde & PG_PS) == 0) {
2244 pmap_remove_page(pmap, sva, &info);
2245 pmap_inval_done(&info);
2246 lwkt_reltoken(&vm_token);
2251 for (; sva < eva; sva = va_next) {
2252 pml4e = pmap_pml4e(pmap, sva);
2253 if ((*pml4e & PG_V) == 0) {
2254 va_next = (sva + NBPML4) & ~PML4MASK;
2260 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2261 if ((*pdpe & PG_V) == 0) {
2262 va_next = (sva + NBPDP) & ~PDPMASK;
2269 * Calculate index for next page table.
2271 va_next = (sva + NBPDR) & ~PDRMASK;
2275 pde = pmap_pdpe_to_pde(pdpe, sva);
2279 * Weed out invalid mappings.
2285 * Check for large page.
2287 if ((ptpaddr & PG_PS) != 0) {
2288 /* JG FreeBSD has more complex treatment here */
2289 pmap_inval_interlock(&info, pmap, -1);
2291 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2292 pmap_inval_deinterlock(&info, pmap);
2297 * Limit our scan to either the end of the va represented
2298 * by the current page table page, or to the end of the
2299 * range being removed.
2305 * NOTE: pmap_remove_pte() can block.
2307 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2311 if (pmap_remove_pte(pmap, pte, sva, &info))
2315 pmap_inval_done(&info);
2316 lwkt_reltoken(&vm_token);
2322 * Removes this physical page from all physical maps in which it resides.
2323 * Reflects back modify bits to the pager.
2325 * This routine may not be called from an interrupt.
2330 pmap_remove_all(vm_page_t m)
2332 struct pmap_inval_info info;
2333 pt_entry_t *pte, tpte;
2336 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2339 lwkt_gettoken(&vm_token);
2340 pmap_inval_init(&info);
2342 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2343 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2344 --pv->pv_pmap->pm_stats.resident_count;
2346 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2347 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
2348 tpte = pte_load_clear(pte);
2350 pv->pv_pmap->pm_stats.wired_count--;
2351 pmap_inval_deinterlock(&info, pv->pv_pmap);
2353 vm_page_flag_set(m, PG_REFERENCED);
2356 * Update the vm_page_t clean and reference bits.
2359 #if defined(PMAP_DIAGNOSTIC)
2360 if (pmap_nw_modified(tpte)) {
2362 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2366 if (pmap_track_modified(pv->pv_va))
2369 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2370 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2371 ++pv->pv_pmap->pm_generation;
2372 m->md.pv_list_count--;
2373 KKASSERT(m->md.pv_list_count >= 0);
2374 if (TAILQ_EMPTY(&m->md.pv_list))
2375 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2376 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2380 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2381 pmap_inval_done(&info);
2382 lwkt_reltoken(&vm_token);
2388 * Set the physical protection on the specified range of this map
2391 * This function may not be called from an interrupt if the map is
2392 * not the kernel_pmap.
2395 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2397 vm_offset_t va_next;
2398 pml4_entry_t *pml4e;
2400 pd_entry_t ptpaddr, *pde;
2402 pmap_inval_info info;
2404 /* JG review for NX */
2409 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2410 pmap_remove(pmap, sva, eva);
2414 if (prot & VM_PROT_WRITE)
2417 lwkt_gettoken(&vm_token);
2418 pmap_inval_init(&info);
2420 for (; sva < eva; sva = va_next) {
2422 pml4e = pmap_pml4e(pmap, sva);
2423 if ((*pml4e & PG_V) == 0) {
2424 va_next = (sva + NBPML4) & ~PML4MASK;
2430 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2431 if ((*pdpe & PG_V) == 0) {
2432 va_next = (sva + NBPDP) & ~PDPMASK;
2438 va_next = (sva + NBPDR) & ~PDRMASK;
2442 pde = pmap_pdpe_to_pde(pdpe, sva);
2446 * Check for large page.
2448 if ((ptpaddr & PG_PS) != 0) {
2449 pmap_inval_interlock(&info, pmap, -1);
2450 *pde &= ~(PG_M|PG_RW);
2451 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2452 pmap_inval_deinterlock(&info, pmap);
2457 * Weed out invalid mappings. Note: we assume that the page
2458 * directory table is always allocated, and in kernel virtual.
2466 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2473 * XXX non-optimal. Note also that there can be
2474 * no pmap_inval_flush() calls until after we modify
2475 * ptbase[sindex] (or otherwise we have to do another
2476 * pmap_inval_add() call).
2478 pmap_inval_interlock(&info, pmap, sva);
2482 if ((pbits & PG_V) == 0) {
2483 pmap_inval_deinterlock(&info, pmap);
2486 if (pbits & PG_MANAGED) {
2489 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2490 vm_page_flag_set(m, PG_REFERENCED);
2494 if (pmap_track_modified(sva)) {
2496 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2503 if (pbits != cbits &&
2504 !atomic_cmpset_long(pte, pbits, cbits)) {
2507 pmap_inval_deinterlock(&info, pmap);
2510 pmap_inval_done(&info);
2511 lwkt_reltoken(&vm_token);
2515 * Insert the given physical page (p) at
2516 * the specified virtual address (v) in the
2517 * target physical map with the protection requested.
2519 * If specified, the page will be wired down, meaning
2520 * that the related pte can not be reclaimed.
2522 * NB: This is the only routine which MAY NOT lazy-evaluate
2523 * or lose information. That is, this routine must actually
2524 * insert this page into the given map NOW.
2527 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2534 pt_entry_t origpte, newpte;
2536 pmap_inval_info info;
2541 va = trunc_page(va);
2542 #ifdef PMAP_DIAGNOSTIC
2544 panic("pmap_enter: toobig");
2545 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2546 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
2548 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2549 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2551 db_print_backtrace();
2554 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2555 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2557 db_print_backtrace();
2561 lwkt_gettoken(&vm_token);
2564 * In the case that a page table page is not
2565 * resident, we are creating it here.
2567 if (va < VM_MAX_USER_ADDRESS)
2568 mpte = pmap_allocpte(pmap, va);
2572 pmap_inval_init(&info);
2573 pde = pmap_pde(pmap, va);
2574 if (pde != NULL && (*pde & PG_V) != 0) {
2575 if ((*pde & PG_PS) != 0)
2576 panic("pmap_enter: attempted pmap_enter on 2MB page");
2577 pte = pmap_pde_to_pte(pde, va);
2579 panic("pmap_enter: invalid page directory va=%#lx", va);
2581 KKASSERT(pte != NULL);
2582 pa = VM_PAGE_TO_PHYS(m);
2584 opa = origpte & PG_FRAME;
2587 * Mapping has not changed, must be protection or wiring change.
2589 if (origpte && (opa == pa)) {
2591 * Wiring change, just update stats. We don't worry about
2592 * wiring PT pages as they remain resident as long as there
2593 * are valid mappings in them. Hence, if a user page is wired,
2594 * the PT page will be also.
2596 if (wired && ((origpte & PG_W) == 0))
2597 pmap->pm_stats.wired_count++;
2598 else if (!wired && (origpte & PG_W))
2599 pmap->pm_stats.wired_count--;
2601 #if defined(PMAP_DIAGNOSTIC)
2602 if (pmap_nw_modified(origpte)) {
2604 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2610 * Remove the extra pte reference. Note that we cannot
2611 * optimize the RO->RW case because we have adjusted the
2612 * wiring count above and may need to adjust the wiring
2619 * We might be turning off write access to the page,
2620 * so we go ahead and sense modify status.
2622 if (origpte & PG_MANAGED) {
2623 if ((origpte & PG_M) && pmap_track_modified(va)) {
2625 om = PHYS_TO_VM_PAGE(opa);
2629 KKASSERT(m->flags & PG_MAPPED);
2634 * Mapping has changed, invalidate old range and fall through to
2635 * handle validating new mapping.
2639 err = pmap_remove_pte(pmap, pte, va, &info);
2641 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2643 opa = origpte & PG_FRAME;
2645 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2651 * Enter on the PV list if part of our managed memory. Note that we
2652 * raise IPL while manipulating pv_table since pmap_enter can be
2653 * called at interrupt time.
2655 if (pmap_initialized &&
2656 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2657 pmap_insert_entry(pmap, va, mpte, m);
2659 vm_page_flag_set(m, PG_MAPPED);
2663 * Increment counters
2665 ++pmap->pm_stats.resident_count;
2667 pmap->pm_stats.wired_count++;
2671 * Now validate mapping with desired protection/wiring.
2673 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
2677 if (va < VM_MAX_USER_ADDRESS)
2679 if (pmap == &kernel_pmap)
2683 * if the mapping or permission bits are different, we need
2684 * to update the pte.
2686 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2687 pmap_inval_interlock(&info, pmap, va);
2688 *pte = newpte | PG_A;
2689 pmap_inval_deinterlock(&info, pmap);
2691 vm_page_flag_set(m, PG_WRITEABLE);
2693 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2694 pmap_inval_done(&info);
2695 lwkt_reltoken(&vm_token);
2699 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2700 * This code also assumes that the pmap has no pre-existing entry for this
2703 * This code currently may only be used on user pmaps, not kernel_pmap.
2706 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2711 vm_pindex_t ptepindex;
2713 pmap_inval_info info;
2715 lwkt_gettoken(&vm_token);
2716 pmap_inval_init(&info);
2718 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2719 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2721 db_print_backtrace();
2724 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2725 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2727 db_print_backtrace();
2731 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2734 * Calculate the page table page (mpte), allocating it if necessary.
2736 * A held page table page (mpte), or NULL, is passed onto the
2737 * section following.
2739 if (va < VM_MAX_USER_ADDRESS) {
2741 * Calculate pagetable page index
2743 ptepindex = pmap_pde_pindex(va);
2747 * Get the page directory entry
2749 ptepa = pmap_pde(pmap, va);
2752 * If the page table page is mapped, we just increment
2753 * the hold count, and activate it.
2755 if (ptepa && (*ptepa & PG_V) != 0) {
2757 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2758 // if (pmap->pm_ptphint &&
2759 // (pmap->pm_ptphint->pindex == ptepindex)) {
2760 // mpte = pmap->pm_ptphint;
2762 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2763 pmap->pm_ptphint = mpte;
2768 mpte = _pmap_allocpte(pmap, ptepindex);
2770 } while (mpte == NULL);
2773 /* this code path is not yet used */
2777 * With a valid (and held) page directory page, we can just use
2778 * vtopte() to get to the pte. If the pte is already present
2779 * we do not disturb it.
2784 pmap_unwire_pte_hold(pmap, va, mpte, &info);
2785 pa = VM_PAGE_TO_PHYS(m);
2786 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2787 pmap_inval_done(&info);
2788 lwkt_reltoken(&vm_token);
2793 * Enter on the PV list if part of our managed memory
2795 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2796 pmap_insert_entry(pmap, va, mpte, m);
2797 vm_page_flag_set(m, PG_MAPPED);
2801 * Increment counters
2803 ++pmap->pm_stats.resident_count;
2805 pa = VM_PAGE_TO_PHYS(m);
2808 * Now validate mapping with RO protection
2810 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2811 *pte = pa | PG_V | PG_U;
2813 *pte = pa | PG_V | PG_U | PG_MANAGED;
2814 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2815 pmap_inval_done(&info);
2816 lwkt_reltoken(&vm_token);
2820 * Make a temporary mapping for a physical address. This is only intended
2821 * to be used for panic dumps.
2823 /* JG Needed on x86_64? */
2825 pmap_kenter_temporary(vm_paddr_t pa, int i)
2827 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2828 return ((void *)crashdumpmap);
2831 #define MAX_INIT_PT (96)
2834 * This routine preloads the ptes for a given object into the specified pmap.
2835 * This eliminates the blast of soft faults on process startup and
2836 * immediately after an mmap.
2838 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2841 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2842 vm_object_t object, vm_pindex_t pindex,
2843 vm_size_t size, int limit)
2845 struct rb_vm_page_scan_info info;
2850 * We can't preinit if read access isn't set or there is no pmap
2853 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2857 * We can't preinit if the pmap is not the current pmap
2859 lp = curthread->td_lwp;
2860 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2863 psize = x86_64_btop(size);
2865 if ((object->type != OBJT_VNODE) ||
2866 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2867 (object->resident_page_count > MAX_INIT_PT))) {
2871 if (psize + pindex > object->size) {
2872 if (object->size < pindex)
2874 psize = object->size - pindex;
2881 * Use a red-black scan to traverse the requested range and load
2882 * any valid pages found into the pmap.
2884 * We cannot safely scan the object's memq unless we are in a
2885 * critical section since interrupts can remove pages from objects.
2887 info.start_pindex = pindex;
2888 info.end_pindex = pindex + psize - 1;
2895 lwkt_gettoken(&vm_token);
2896 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2897 pmap_object_init_pt_callback, &info);
2898 lwkt_reltoken(&vm_token);
2904 pmap_object_init_pt_callback(vm_page_t p, void *data)
2906 struct rb_vm_page_scan_info *info = data;
2907 vm_pindex_t rel_index;
2909 * don't allow an madvise to blow away our really
2910 * free pages allocating pv entries.
2912 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2913 vmstats.v_free_count < vmstats.v_free_reserved) {
2916 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2917 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2918 if ((p->queue - p->pc) == PQ_CACHE)
2919 vm_page_deactivate(p);
2921 rel_index = p->pindex - info->start_pindex;
2922 pmap_enter_quick(info->pmap,
2923 info->addr + x86_64_ptob(rel_index), p);
2930 * Return TRUE if the pmap is in shape to trivially
2931 * pre-fault the specified address.
2933 * Returns FALSE if it would be non-trivial or if a
2934 * pte is already loaded into the slot.
2937 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2943 lwkt_gettoken(&vm_token);
2944 pde = pmap_pde(pmap, addr);
2945 if (pde == NULL || *pde == 0) {
2949 ret = (*pte) ? 0 : 1;
2951 lwkt_reltoken(&vm_token);
2956 * Routine: pmap_change_wiring
2957 * Function: Change the wiring attribute for a map/virtual-address
2959 * In/out conditions:
2960 * The mapping must already exist in the pmap.
2963 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2970 lwkt_gettoken(&vm_token);
2971 pte = pmap_pte(pmap, va);
2973 if (wired && !pmap_pte_w(pte))
2974 pmap->pm_stats.wired_count++;
2975 else if (!wired && pmap_pte_w(pte))
2976 pmap->pm_stats.wired_count--;
2979 * Wiring is not a hardware characteristic so there is no need to
2980 * invalidate TLB. However, in an SMP environment we must use
2981 * a locked bus cycle to update the pte (if we are not using
2982 * the pmap_inval_*() API that is)... it's ok to do this for simple
2987 atomic_set_long(pte, PG_W);
2989 atomic_clear_long(pte, PG_W);
2992 atomic_set_long_nonlocked(pte, PG_W);
2994 atomic_clear_long_nonlocked(pte, PG_W);
2996 lwkt_reltoken(&vm_token);
3002 * Copy the range specified by src_addr/len
3003 * from the source map to the range dst_addr/len
3004 * in the destination map.
3006 * This routine is only advisory and need not do anything.
3009 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3010 vm_size_t len, vm_offset_t src_addr)
3014 pmap_inval_info info;
3016 vm_offset_t end_addr = src_addr + len;
3018 pd_entry_t src_frame, dst_frame;
3021 if (dst_addr != src_addr)
3024 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3025 if (src_frame != (PTDpde & PG_FRAME)) {
3029 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3030 if (dst_frame != (APTDpde & PG_FRAME)) {
3031 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
3032 /* The page directory is not shared between CPUs */
3036 pmap_inval_init(&info);
3037 pmap_inval_add(&info, dst_pmap, -1);
3038 pmap_inval_add(&info, src_pmap, -1);
3041 * critical section protection is required to maintain the page/object
3042 * association, interrupts can free pages and remove them from
3046 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
3047 pt_entry_t *src_pte, *dst_pte;
3048 vm_page_t dstmpte, srcmpte;
3049 vm_offset_t srcptepaddr;
3050 vm_pindex_t ptepindex;
3052 if (addr >= UPT_MIN_ADDRESS)
3053 panic("pmap_copy: invalid to pmap_copy page tables\n");
3056 * Don't let optional prefaulting of pages make us go
3057 * way below the low water mark of free pages or way
3058 * above high water mark of used pv entries.
3060 if (vmstats.v_free_count < vmstats.v_free_reserved ||
3061 pv_entry_count > pv_entry_high_water)
3064 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
3065 ptepindex = addr >> PDRSHIFT;
3068 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
3070 if (srcptepaddr == 0)
3073 if (srcptepaddr & PG_PS) {
3075 if (dst_pmap->pm_pdir[ptepindex] == 0) {
3076 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
3077 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
3083 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
3084 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
3085 (srcmpte->flags & PG_BUSY)) {
3089 if (pdnxt > end_addr)
3092 src_pte = vtopte(addr);
3094 dst_pte = avtopte(addr);
3096 while (addr < pdnxt) {
3101 * we only virtual copy managed pages
3103 if ((ptetemp & PG_MANAGED) != 0) {
3105 * We have to check after allocpte for the
3106 * pte still being around... allocpte can
3109 * pmap_allocpte() can block. If we lose
3110 * our page directory mappings we stop.
3112 dstmpte = pmap_allocpte(dst_pmap, addr);
3115 if (src_frame != (PTDpde & PG_FRAME) ||
3116 dst_frame != (APTDpde & PG_FRAME)
3118 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3119 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3121 } else if ((*dst_pte == 0) &&
3122 (ptetemp = *src_pte) != 0 &&
3123 (ptetemp & PG_MANAGED)) {
3125 * Clear the modified and
3126 * accessed (referenced) bits
3129 m = PHYS_TO_VM_PAGE(ptetemp);
3130 *dst_pte = ptetemp & ~(PG_M | PG_A);
3131 ++dst_pmap->pm_stats.resident_count;
3132 pmap_insert_entry(dst_pmap, addr,
3134 KKASSERT(m->flags & PG_MAPPED);
3136 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3137 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3141 if (dstmpte->hold_count >= srcmpte->hold_count)
3151 pmap_inval_done(&info);
3158 * Zero the specified physical page.
3160 * This function may be called from an interrupt and no locking is
3164 pmap_zero_page(vm_paddr_t phys)
3166 vm_offset_t va = PHYS_TO_DMAP(phys);
3168 pagezero((void *)va);
3172 * pmap_page_assertzero:
3174 * Assert that a page is empty, panic if it isn't.
3177 pmap_page_assertzero(vm_paddr_t phys)
3179 vm_offset_t virt = PHYS_TO_DMAP(phys);
3182 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3183 if (*(long *)((char *)virt + i) != 0) {
3184 panic("pmap_page_assertzero() @ %p not zero!\n", (void *)virt);
3192 * Zero part of a physical page by mapping it into memory and clearing
3193 * its contents with bzero.
3195 * off and size may not cover an area beyond a single hardware page.
3198 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3200 vm_offset_t virt = PHYS_TO_DMAP(phys);
3202 bzero((char *)virt + off, size);
3208 * Copy the physical page from the source PA to the target PA.
3209 * This function may be called from an interrupt. No locking
3213 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3215 vm_offset_t src_virt, dst_virt;
3217 src_virt = PHYS_TO_DMAP(src);
3218 dst_virt = PHYS_TO_DMAP(dst);
3219 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3223 * pmap_copy_page_frag:
3225 * Copy the physical page from the source PA to the target PA.
3226 * This function may be called from an interrupt. No locking
3230 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3232 vm_offset_t src_virt, dst_virt;
3234 src_virt = PHYS_TO_DMAP(src);
3235 dst_virt = PHYS_TO_DMAP(dst);
3237 bcopy((char *)src_virt + (src & PAGE_MASK),
3238 (char *)dst_virt + (dst & PAGE_MASK),
3243 * Returns true if the pmap's pv is one of the first
3244 * 16 pvs linked to from this page. This count may
3245 * be changed upwards or downwards in the future; it
3246 * is only necessary that true be returned for a small
3247 * subset of pmaps for proper page aging.
3250 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3255 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3259 lwkt_gettoken(&vm_token);
3261 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3262 if (pv->pv_pmap == pmap) {
3263 lwkt_reltoken(&vm_token);
3271 lwkt_reltoken(&vm_token);
3277 * Remove all pages from specified address space
3278 * this aids process exit speeds. Also, this code
3279 * is special cased for current process only, but
3280 * can have the more generic (and slightly slower)
3281 * mode enabled. This is much faster than pmap_remove
3282 * in the case of running down an entire address space.
3285 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3288 pt_entry_t *pte, tpte;
3291 pmap_inval_info info;
3293 int save_generation;
3295 lp = curthread->td_lwp;
3296 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3301 lwkt_gettoken(&vm_token);
3302 pmap_inval_init(&info);
3303 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3304 if (pv->pv_va >= eva || pv->pv_va < sva) {
3305 npv = TAILQ_NEXT(pv, pv_plist);
3309 KKASSERT(pmap == pv->pv_pmap);
3312 pte = vtopte(pv->pv_va);
3314 pte = pmap_pte_quick(pmap, pv->pv_va);
3315 pmap_inval_interlock(&info, pmap, pv->pv_va);
3318 * We cannot remove wired pages from a process' mapping
3322 pmap_inval_deinterlock(&info, pmap);
3323 npv = TAILQ_NEXT(pv, pv_plist);
3326 tpte = pte_load_clear(pte);
3328 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3330 KASSERT(m < &vm_page_array[vm_page_array_size],
3331 ("pmap_remove_pages: bad tpte %lx", tpte));
3333 KKASSERT(pmap->pm_stats.resident_count > 0);
3334 --pmap->pm_stats.resident_count;
3335 pmap_inval_deinterlock(&info, pmap);
3338 * Update the vm_page_t clean and reference bits.
3344 npv = TAILQ_NEXT(pv, pv_plist);
3345 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3346 save_generation = ++pmap->pm_generation;
3348 m->md.pv_list_count--;
3349 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3350 if (TAILQ_EMPTY(&m->md.pv_list))
3351 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3353 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3357 * Restart the scan if we blocked during the unuse or free
3358 * calls and other removals were made.
3360 if (save_generation != pmap->pm_generation) {
3361 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3362 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3365 pmap_inval_done(&info);
3366 lwkt_reltoken(&vm_token);
3370 * pmap_testbit tests bits in pte's
3371 * note that the testbit/clearbit routines are inline,
3372 * and a lot of things compile-time evaluate.
3376 pmap_testbit(vm_page_t m, int bit)
3381 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3384 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3389 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3391 * if the bit being tested is the modified bit, then
3392 * mark clean_map and ptes as never
3395 if (bit & (PG_A|PG_M)) {
3396 if (!pmap_track_modified(pv->pv_va))
3400 #if defined(PMAP_DIAGNOSTIC)
3401 if (pv->pv_pmap == NULL) {
3402 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3406 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3417 * this routine is used to modify bits in ptes
3421 pmap_clearbit(vm_page_t m, int bit)
3423 struct pmap_inval_info info;
3428 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3431 pmap_inval_init(&info);
3434 * Loop over all current mappings setting/clearing as appropos If
3435 * setting RO do we need to clear the VAC?
3437 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3439 * don't write protect pager mappings
3442 if (!pmap_track_modified(pv->pv_va))
3446 #if defined(PMAP_DIAGNOSTIC)
3447 if (pv->pv_pmap == NULL) {
3448 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3454 * Careful here. We can use a locked bus instruction to
3455 * clear PG_A or PG_M safely but we need to synchronize
3456 * with the target cpus when we mess with PG_RW.
3458 * We do not have to force synchronization when clearing
3459 * PG_M even for PTEs generated via virtual memory maps,
3460 * because the virtual kernel will invalidate the pmap
3461 * entry when/if it needs to resynchronize the Modify bit.
3464 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
3465 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3472 atomic_clear_long(pte, PG_M|PG_RW);
3475 * The cpu may be trying to set PG_M
3476 * simultaniously with our clearing
3479 if (!atomic_cmpset_long(pte, pbits,
3483 } else if (bit == PG_M) {
3485 * We could also clear PG_RW here to force
3486 * a fault on write to redetect PG_M for
3487 * virtual kernels, but it isn't necessary
3488 * since virtual kernels invalidate the pte
3489 * when they clear the VPTE_M bit in their
3490 * virtual page tables.
3492 atomic_clear_long(pte, PG_M);
3494 atomic_clear_long(pte, bit);
3498 pmap_inval_deinterlock(&info, pv->pv_pmap);
3500 pmap_inval_done(&info);
3504 * pmap_page_protect:
3506 * Lower the permission for all mappings to a given page.
3509 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3511 /* JG NX support? */
3512 if ((prot & VM_PROT_WRITE) == 0) {
3513 lwkt_gettoken(&vm_token);
3514 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3515 pmap_clearbit(m, PG_RW);
3516 vm_page_flag_clear(m, PG_WRITEABLE);
3520 lwkt_reltoken(&vm_token);
3525 pmap_phys_address(vm_pindex_t ppn)
3527 return (x86_64_ptob(ppn));
3531 * pmap_ts_referenced:
3533 * Return a count of reference bits for a page, clearing those bits.
3534 * It is not necessary for every reference bit to be cleared, but it
3535 * is necessary that 0 only be returned when there are truly no
3536 * reference bits set.
3538 * XXX: The exact number of bits to check and clear is a matter that
3539 * should be tested and standardized at some point in the future for
3540 * optimal aging of shared pages.
3543 pmap_ts_referenced(vm_page_t m)
3545 pv_entry_t pv, pvf, pvn;
3549 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3553 lwkt_gettoken(&vm_token);
3555 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3560 pvn = TAILQ_NEXT(pv, pv_list);
3563 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3564 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3567 if (!pmap_track_modified(pv->pv_va))
3570 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3572 if (pte && (*pte & PG_A)) {
3574 atomic_clear_long(pte, PG_A);
3576 atomic_clear_long_nonlocked(pte, PG_A);
3583 } while ((pv = pvn) != NULL && pv != pvf);
3585 lwkt_reltoken(&vm_token);
3594 * Return whether or not the specified physical page was modified
3595 * in any physical maps.
3598 pmap_is_modified(vm_page_t m)
3602 lwkt_gettoken(&vm_token);
3603 res = pmap_testbit(m, PG_M);
3604 lwkt_reltoken(&vm_token);
3609 * Clear the modify bits on the specified physical page.
3612 pmap_clear_modify(vm_page_t m)
3614 lwkt_gettoken(&vm_token);
3615 pmap_clearbit(m, PG_M);
3616 lwkt_reltoken(&vm_token);
3620 * pmap_clear_reference:
3622 * Clear the reference bit on the specified physical page.
3625 pmap_clear_reference(vm_page_t m)
3627 lwkt_gettoken(&vm_token);
3628 pmap_clearbit(m, PG_A);
3629 lwkt_reltoken(&vm_token);
3633 * Miscellaneous support routines follow
3638 i386_protection_init(void)
3642 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3643 kp = protection_codes;
3644 for (prot = 0; prot < 8; prot++) {
3646 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3648 * Read access is also 0. There isn't any execute bit,
3649 * so just make it readable.
3651 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3652 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3653 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3656 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3657 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3658 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3659 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3667 * Map a set of physical memory pages into the kernel virtual
3668 * address space. Return a pointer to where it is mapped. This
3669 * routine is intended to be used for mapping device memory,
3672 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3676 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3678 vm_offset_t va, tmpva, offset;
3681 offset = pa & PAGE_MASK;
3682 size = roundup(offset + size, PAGE_SIZE);
3684 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3686 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3688 pa = pa & ~PAGE_MASK;
3689 for (tmpva = va; size > 0;) {
3690 pte = vtopte(tmpva);
3691 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3699 return ((void *)(va + offset));
3703 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3705 vm_offset_t va, tmpva, offset;
3708 offset = pa & PAGE_MASK;
3709 size = roundup(offset + size, PAGE_SIZE);
3711 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3713 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3715 pa = pa & ~PAGE_MASK;
3716 for (tmpva = va; size > 0;) {
3717 pte = vtopte(tmpva);
3718 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3726 return ((void *)(va + offset));
3730 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3732 vm_offset_t base, offset;
3734 base = va & ~PAGE_MASK;
3735 offset = va & PAGE_MASK;
3736 size = roundup(offset + size, PAGE_SIZE);
3737 pmap_qremove(va, size >> PAGE_SHIFT);
3738 kmem_free(&kernel_map, base, size);
3742 * perform the pmap work for mincore
3745 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3747 pt_entry_t *ptep, pte;
3751 lwkt_gettoken(&vm_token);
3752 ptep = pmap_pte(pmap, addr);
3754 if (ptep && (pte = *ptep) != 0) {
3757 val = MINCORE_INCORE;
3758 if ((pte & PG_MANAGED) == 0)
3761 pa = pte & PG_FRAME;
3763 m = PHYS_TO_VM_PAGE(pa);
3769 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3771 * Modified by someone
3773 else if (m->dirty || pmap_is_modified(m))
3774 val |= MINCORE_MODIFIED_OTHER;
3779 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3782 * Referenced by someone
3784 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3785 val |= MINCORE_REFERENCED_OTHER;
3786 vm_page_flag_set(m, PG_REFERENCED);
3790 lwkt_reltoken(&vm_token);
3795 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3796 * vmspace will be ref'd and the old one will be deref'd.
3798 * The vmspace for all lwps associated with the process will be adjusted
3799 * and cr3 will be reloaded if any lwp is the current lwp.
3802 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3804 struct vmspace *oldvm;
3808 oldvm = p->p_vmspace;
3809 if (oldvm != newvm) {
3810 p->p_vmspace = newvm;
3811 KKASSERT(p->p_nthreads == 1);
3812 lp = RB_ROOT(&p->p_lwp_tree);
3813 pmap_setlwpvm(lp, newvm);
3815 sysref_get(&newvm->vm_sysref);
3816 sysref_put(&oldvm->vm_sysref);
3823 * Set the vmspace for a LWP. The vmspace is almost universally set the
3824 * same as the process vmspace, but virtual kernels need to swap out contexts
3825 * on a per-lwp basis.
3828 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3830 struct vmspace *oldvm;
3834 oldvm = lp->lwp_vmspace;
3836 if (oldvm != newvm) {
3837 lp->lwp_vmspace = newvm;
3838 if (curthread->td_lwp == lp) {
3839 pmap = vmspace_pmap(newvm);
3841 atomic_set_int(&pmap->pm_active, mycpu->gd_cpumask);
3842 if (pmap->pm_active & CPUMASK_LOCK)
3843 pmap_interlock_wait(newvm);
3845 pmap->pm_active |= 1;
3847 #if defined(SWTCH_OPTIM_STATS)
3850 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
3851 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
3852 load_cr3(curthread->td_pcb->pcb_cr3);
3853 pmap = vmspace_pmap(oldvm);
3855 atomic_clear_int(&pmap->pm_active, mycpu->gd_cpumask);
3857 pmap->pm_active &= ~1;
3867 * Called when switching to a locked pmap
3870 pmap_interlock_wait(struct vmspace *vm)
3872 struct pmap *pmap = &vm->vm_pmap;
3874 if (pmap->pm_active & CPUMASK_LOCK) {
3875 while (pmap->pm_active & CPUMASK_LOCK) {
3878 lwkt_process_ipiq();
3886 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3889 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3893 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3898 * Used by kmalloc/kfree, page already exists at va
3901 pmap_kvtom(vm_offset_t va)
3903 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));