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, long n)
421 bzero((void *)ret, n * PAGE_SIZE);
422 *firstaddr += n * PAGE_SIZE;
428 create_pagetables(vm_paddr_t *firstaddr)
430 long i; /* must be 64 bits */
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 += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
447 nkpt += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E + ndmpdp) +
454 KPTbase = allocpages(firstaddr, nkpt);
455 KPTphys = allocpages(firstaddr, nkpt);
456 KPML4phys = allocpages(firstaddr, 1);
457 KPDPphys = allocpages(firstaddr, NKPML4E);
460 * Calculate the page directory base for KERNBASE,
461 * that is where we start populating the page table pages.
462 * Basically this is the end - 2.
464 KPDphys = allocpages(firstaddr, NKPDPE);
465 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
467 DMPDPphys = allocpages(firstaddr, NDMPML4E);
468 if ((amd_feature & AMDID_PAGE1GB) == 0)
469 DMPDphys = allocpages(firstaddr, ndmpdp);
470 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
473 * Fill in the underlying page table pages for the area around
474 * KERNBASE. This remaps low physical memory to KERNBASE.
476 * Read-only from zero to physfree
477 * XXX not fully used, underneath 2M pages
479 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
480 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
481 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
485 * Now map the initial kernel page tables. One block of page
486 * tables is placed at the beginning of kernel virtual memory,
487 * and another block is placed at KERNBASE to map the kernel binary,
488 * data, bss, and initial pre-allocations.
490 for (i = 0; i < nkpt; i++) {
491 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
492 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
494 for (i = 0; i < nkpt; i++) {
495 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
496 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
500 * Map from zero to end of allocations using 2M pages as an
501 * optimization. This will bypass some of the KPTBase pages
502 * above in the KERNBASE area.
504 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
505 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
506 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
510 * And connect up the PD to the PDP. The kernel pmap is expected
511 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
513 for (i = 0; i < NKPDPE; i++) {
514 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
515 KPDphys + (i << PAGE_SHIFT);
516 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
520 /* Now set up the direct map space using either 2MB or 1GB pages */
521 /* Preset PG_M and PG_A because demotion expects it */
522 if ((amd_feature & AMDID_PAGE1GB) == 0) {
523 for (i = 0; i < NPDEPG * ndmpdp; i++) {
524 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
525 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
528 /* And the direct map space's PDP */
529 for (i = 0; i < ndmpdp; i++) {
530 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
532 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
535 for (i = 0; i < ndmpdp; i++) {
536 ((pdp_entry_t *)DMPDPphys)[i] =
537 (vm_paddr_t)i << PDPSHIFT;
538 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
543 /* And recursively map PML4 to itself in order to get PTmap */
544 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
545 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
547 /* Connect the Direct Map slot up to the PML4 */
548 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
549 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
551 /* Connect the KVA slot up to the PML4 */
552 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
553 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
557 * Bootstrap the system enough to run with virtual memory.
559 * On the i386 this is called after mapping has already been enabled
560 * and just syncs the pmap module with what has already been done.
561 * [We can't call it easily with mapping off since the kernel is not
562 * mapped with PA == VA, hence we would have to relocate every address
563 * from the linked base (virtual) address "KERNBASE" to the actual
564 * (physical) address starting relative to 0]
567 pmap_bootstrap(vm_paddr_t *firstaddr)
571 struct mdglobaldata *gd;
574 KvaStart = VM_MIN_KERNEL_ADDRESS;
575 KvaEnd = VM_MAX_KERNEL_ADDRESS;
576 KvaSize = KvaEnd - KvaStart;
578 avail_start = *firstaddr;
581 * Create an initial set of page tables to run the kernel in.
583 create_pagetables(firstaddr);
585 virtual2_start = KvaStart;
586 virtual2_end = PTOV_OFFSET;
588 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
589 virtual_start = pmap_kmem_choose(virtual_start);
591 virtual_end = VM_MAX_KERNEL_ADDRESS;
593 /* XXX do %cr0 as well */
594 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
598 * Initialize protection array.
600 i386_protection_init();
603 * The kernel's pmap is statically allocated so we don't have to use
604 * pmap_create, which is unlikely to work correctly at this part of
605 * the boot sequence (XXX and which no longer exists).
607 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
608 kernel_pmap.pm_count = 1;
609 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
610 TAILQ_INIT(&kernel_pmap.pm_pvlist);
613 * Reserve some special page table entries/VA space for temporary
616 #define SYSMAP(c, p, v, n) \
617 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
623 * CMAP1/CMAP2 are used for zeroing and copying pages.
625 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
630 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
633 * ptvmmap is used for reading arbitrary physical pages via
636 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
639 * msgbufp is used to map the system message buffer.
640 * XXX msgbufmap is not used.
642 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
643 atop(round_page(MSGBUF_SIZE)))
650 * PG_G is terribly broken on SMP because we IPI invltlb's in some
651 * cases rather then invl1pg. Actually, I don't even know why it
652 * works under UP because self-referential page table mappings
657 if (cpu_feature & CPUID_PGE)
662 * Initialize the 4MB page size flag
666 * The 4MB page version of the initial
667 * kernel page mapping.
671 #if !defined(DISABLE_PSE)
672 if (cpu_feature & CPUID_PSE) {
675 * Note that we have enabled PSE mode
678 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
679 ptditmp &= ~(NBPDR - 1);
680 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
685 * Enable the PSE mode. If we are SMP we can't do this
686 * now because the APs will not be able to use it when
689 load_cr4(rcr4() | CR4_PSE);
692 * We can do the mapping here for the single processor
693 * case. We simply ignore the old page table page from
697 * For SMP, we still need 4K pages to bootstrap APs,
698 * PSE will be enabled as soon as all APs are up.
700 PTD[KPTDI] = (pd_entry_t)ptditmp;
707 * We need to finish setting up the globaldata page for the BSP.
708 * locore has already populated the page table for the mdglobaldata
711 pg = MDGLOBALDATA_BASEALLOC_PAGES;
712 gd = &CPU_prvspace[0].mdglobaldata;
713 gd->gd_CMAP1 = &SMPpt[pg + 0];
714 gd->gd_CMAP2 = &SMPpt[pg + 1];
715 gd->gd_CMAP3 = &SMPpt[pg + 2];
716 gd->gd_PMAP1 = &SMPpt[pg + 3];
717 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
718 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
719 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
720 gd->gd_PADDR1 = (pt_entry_t *)CPU_prvspace[0].PPAGE1;
727 * Set 4mb pdir for mp startup
732 if (pseflag && (cpu_feature & CPUID_PSE)) {
733 load_cr4(rcr4() | CR4_PSE);
734 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
742 * XXX: Hack. Required by pmap_init()
744 extern vm_offset_t cpu_apic_addr;
747 * Initialize the pmap module.
748 * Called by vm_init, to initialize any structures that the pmap
749 * system needs to map virtual memory.
750 * pmap_init has been enhanced to support in a fairly consistant
751 * way, discontiguous physical memory.
760 * object for kernel page table pages
762 /* JG I think the number can be arbitrary */
763 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
766 * Allocate memory for random pmap data structures. Includes the
770 for(i = 0; i < vm_page_array_size; i++) {
773 m = &vm_page_array[i];
774 TAILQ_INIT(&m->md.pv_list);
775 m->md.pv_list_count = 0;
779 * init the pv free list
781 initial_pvs = vm_page_array_size;
782 if (initial_pvs < MINPV)
784 pvzone = &pvzone_store;
785 pvinit = (void *)kmem_alloc(&kernel_map,
786 initial_pvs * sizeof (struct pv_entry));
787 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
788 pvinit, initial_pvs);
791 * Now it is safe to enable pv_table recording.
793 pmap_initialized = TRUE;
798 lapic = pmap_mapdev_uncacheable(cpu_apic_addr, sizeof(struct LAPIC));
803 * Initialize the address space (zone) for the pv_entries. Set a
804 * high water mark so that the system can recover from excessive
805 * numbers of pv entries.
810 int shpgperproc = PMAP_SHPGPERPROC;
813 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
814 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
815 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
816 pv_entry_high_water = 9 * (pv_entry_max / 10);
819 * Subtract out pages already installed in the zone (hack)
821 entry_max = pv_entry_max - vm_page_array_size;
825 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
829 /***************************************************
830 * Low level helper routines.....
831 ***************************************************/
833 #if defined(PMAP_DIAGNOSTIC)
836 * This code checks for non-writeable/modified pages.
837 * This should be an invalid condition.
841 pmap_nw_modified(pt_entry_t pte)
843 if ((pte & (PG_M|PG_RW)) == PG_M)
852 * this routine defines the region(s) of memory that should
853 * not be tested for the modified bit.
857 pmap_track_modified(vm_offset_t va)
859 if ((va < clean_sva) || (va >= clean_eva))
866 * Extract the physical page address associated with the map/VA pair.
868 * The caller must hold vm_token if non-blocking operation is desired.
871 pmap_extract(pmap_t pmap, vm_offset_t va)
875 pd_entry_t pde, *pdep;
877 lwkt_gettoken(&vm_token);
879 pdep = pmap_pde(pmap, va);
883 if ((pde & PG_PS) != 0) {
884 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
886 pte = pmap_pde_to_pte(pdep, va);
887 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
891 lwkt_reltoken(&vm_token);
896 * Extract the physical page address associated kernel virtual address.
899 pmap_kextract(vm_offset_t va)
904 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
905 pa = DMAP_TO_PHYS(va);
909 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
912 * Beware of a concurrent promotion that changes the
913 * PDE at this point! For example, vtopte() must not
914 * be used to access the PTE because it would use the
915 * new PDE. It is, however, safe to use the old PDE
916 * because the page table page is preserved by the
919 pa = *pmap_pde_to_pte(&pde, va);
920 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
926 /***************************************************
927 * Low level mapping routines.....
928 ***************************************************/
931 * Routine: pmap_kenter
933 * Add a wired page to the KVA
934 * NOTE! note that in order for the mapping to take effect -- you
935 * should do an invltlb after doing the pmap_kenter().
938 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
942 pmap_inval_info info;
944 pmap_inval_init(&info);
945 npte = pa | PG_RW | PG_V | pgeflag;
947 pmap_inval_interlock(&info, &kernel_pmap, va);
949 pmap_inval_deinterlock(&info, &kernel_pmap);
950 pmap_inval_done(&info);
954 * Routine: pmap_kenter_quick
956 * Similar to pmap_kenter(), except we only invalidate the
957 * mapping on the current CPU.
960 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
965 npte = pa | PG_RW | PG_V | pgeflag;
968 cpu_invlpg((void *)va);
972 pmap_kenter_sync(vm_offset_t va)
974 pmap_inval_info info;
976 pmap_inval_init(&info);
977 pmap_inval_interlock(&info, &kernel_pmap, va);
978 pmap_inval_deinterlock(&info, &kernel_pmap);
979 pmap_inval_done(&info);
983 pmap_kenter_sync_quick(vm_offset_t va)
985 cpu_invlpg((void *)va);
989 * remove a page from the kernel pagetables
992 pmap_kremove(vm_offset_t va)
995 pmap_inval_info info;
997 pmap_inval_init(&info);
999 pmap_inval_interlock(&info, &kernel_pmap, va);
1001 pmap_inval_deinterlock(&info, &kernel_pmap);
1002 pmap_inval_done(&info);
1006 pmap_kremove_quick(vm_offset_t va)
1011 cpu_invlpg((void *)va);
1015 * XXX these need to be recoded. They are not used in any critical path.
1018 pmap_kmodify_rw(vm_offset_t va)
1020 *vtopte(va) |= PG_RW;
1021 cpu_invlpg((void *)va);
1025 pmap_kmodify_nc(vm_offset_t va)
1027 *vtopte(va) |= PG_N;
1028 cpu_invlpg((void *)va);
1032 * Used to map a range of physical addresses into kernel virtual
1033 * address space during the low level boot, typically to map the
1034 * dump bitmap, message buffer, and vm_page_array.
1036 * These mappings are typically made at some pointer after the end of the
1039 * We could return PHYS_TO_DMAP(start) here and not allocate any
1040 * via (*virtp), but then kmem from userland and kernel dumps won't
1041 * have access to the related pointers.
1044 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1047 vm_offset_t va_start;
1049 /*return PHYS_TO_DMAP(start);*/
1054 while (start < end) {
1055 if ((start / PAGE_SIZE & 15) == 0)
1056 kprintf("%p %p\n", (void *)va, (void *)start);
1057 pmap_kenter_quick(va, start);
1067 * Add a list of wired pages to the kva
1068 * this routine is only used for temporary
1069 * kernel mappings that do not need to have
1070 * page modification or references recorded.
1071 * Note that old mappings are simply written
1072 * over. The page *must* be wired.
1075 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1079 end_va = va + count * PAGE_SIZE;
1081 while (va < end_va) {
1085 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1086 cpu_invlpg((void *)va);
1094 * This routine jerks page mappings from the
1095 * kernel -- it is meant only for temporary mappings.
1097 * MPSAFE, INTERRUPT SAFE (cluster callback)
1100 pmap_qremove(vm_offset_t va, int count)
1104 end_va = va + count * PAGE_SIZE;
1106 while (va < end_va) {
1111 cpu_invlpg((void *)va);
1118 * This routine works like vm_page_lookup() but also blocks as long as the
1119 * page is busy. This routine does not busy the page it returns.
1121 * Unless the caller is managing objects whos pages are in a known state,
1122 * the call should be made with a critical section held so the page's object
1123 * association remains valid on return.
1127 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1132 m = vm_page_lookup(object, pindex);
1133 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
1139 * Create a new thread and optionally associate it with a (new) process.
1140 * NOTE! the new thread's cpu may not equal the current cpu.
1143 pmap_init_thread(thread_t td)
1145 /* enforce pcb placement & alignment */
1146 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1147 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1148 td->td_savefpu = &td->td_pcb->pcb_save;
1149 td->td_sp = (char *)td->td_pcb; /* no -16 */
1153 * This routine directly affects the fork perf for a process.
1156 pmap_init_proc(struct proc *p)
1161 * Dispose the UPAGES for a process that has exited.
1162 * This routine directly impacts the exit perf of a process.
1165 pmap_dispose_proc(struct proc *p)
1167 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1170 /***************************************************
1171 * Page table page management routines.....
1172 ***************************************************/
1175 * This routine unholds page table pages, and if the hold count
1176 * drops to zero, then it decrements the wire count.
1180 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1181 pmap_inval_info_t info)
1183 KKASSERT(m->hold_count > 0);
1184 if (m->hold_count > 1) {
1188 return _pmap_unwire_pte_hold(pmap, va, m, info);
1194 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1195 pmap_inval_info_t info)
1198 * Wait until we can busy the page ourselves. We cannot have
1199 * any active flushes if we block. We own one hold count on the
1200 * page so it cannot be freed out from under us.
1202 if (m->flags & PG_BUSY) {
1203 pmap_inval_flush(info);
1204 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1207 KASSERT(m->queue == PQ_NONE,
1208 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1211 * This case can occur if new references were acquired while
1214 if (m->hold_count > 1) {
1215 KKASSERT(m->hold_count > 1);
1221 * Unmap the page table page
1223 KKASSERT(m->hold_count == 1);
1225 pmap_inval_interlock(info, pmap, -1);
1227 if (m->pindex >= (NUPDE + NUPDPE)) {
1230 pml4 = pmap_pml4e(pmap, va);
1232 } else if (m->pindex >= NUPDE) {
1235 pdp = pmap_pdpe(pmap, va);
1240 pd = pmap_pde(pmap, va);
1244 KKASSERT(pmap->pm_stats.resident_count > 0);
1245 --pmap->pm_stats.resident_count;
1247 if (pmap->pm_ptphint == m)
1248 pmap->pm_ptphint = NULL;
1249 pmap_inval_deinterlock(info, pmap);
1251 if (m->pindex < NUPDE) {
1252 /* We just released a PT, unhold the matching PD */
1255 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
1256 pmap_unwire_pte_hold(pmap, va, pdpg, info);
1258 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
1259 /* We just released a PD, unhold the matching PDP */
1262 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
1263 pmap_unwire_pte_hold(pmap, va, pdppg, info);
1267 * This was our last hold, the page had better be unwired
1268 * after we decrement wire_count.
1270 * FUTURE NOTE: shared page directory page could result in
1271 * multiple wire counts.
1275 KKASSERT(m->wire_count == 0);
1276 --vmstats.v_wire_count;
1277 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1279 vm_page_free_zero(m);
1285 * After removing a page table entry, this routine is used to
1286 * conditionally free the page, and manage the hold/wire counts.
1290 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1291 pmap_inval_info_t info)
1293 vm_pindex_t ptepindex;
1295 if (va >= VM_MAX_USER_ADDRESS)
1299 ptepindex = pmap_pde_pindex(va);
1301 if (pmap->pm_ptphint &&
1302 (pmap->pm_ptphint->pindex == ptepindex)) {
1303 mpte = pmap->pm_ptphint;
1306 pmap_inval_flush(info);
1307 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1308 pmap->pm_ptphint = mpte;
1313 return pmap_unwire_pte_hold(pmap, va, mpte, info);
1317 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1318 * it, and IdlePTD, represents the template used to update all other pmaps.
1320 * On architectures where the kernel pmap is not integrated into the user
1321 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1322 * kernel_pmap should be used to directly access the kernel_pmap.
1325 pmap_pinit0(struct pmap *pmap)
1327 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1329 pmap->pm_active = 0;
1330 pmap->pm_ptphint = NULL;
1331 TAILQ_INIT(&pmap->pm_pvlist);
1332 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1336 * Initialize a preallocated and zeroed pmap structure,
1337 * such as one in a vmspace structure.
1340 pmap_pinit(struct pmap *pmap)
1345 * No need to allocate page table space yet but we do need a valid
1346 * page directory table.
1348 if (pmap->pm_pml4 == NULL) {
1350 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1354 * Allocate an object for the ptes
1356 if (pmap->pm_pteobj == NULL)
1357 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1);
1360 * Allocate the page directory page, unless we already have
1361 * one cached. If we used the cached page the wire_count will
1362 * already be set appropriately.
1364 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1365 ptdpg = vm_page_grab(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I,
1366 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1367 pmap->pm_pdirm = ptdpg;
1368 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1369 ptdpg->valid = VM_PAGE_BITS_ALL;
1370 if (ptdpg->wire_count == 0)
1371 ++vmstats.v_wire_count;
1372 ptdpg->wire_count = 1;
1373 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1375 if ((ptdpg->flags & PG_ZERO) == 0)
1376 bzero(pmap->pm_pml4, PAGE_SIZE);
1379 pmap_page_assertzero(VM_PAGE_TO_PHYS(ptdpg));
1382 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1383 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1385 /* install self-referential address mapping entry */
1386 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1389 pmap->pm_active = 0;
1390 pmap->pm_ptphint = NULL;
1391 TAILQ_INIT(&pmap->pm_pvlist);
1392 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1393 pmap->pm_stats.resident_count = 1;
1397 * Clean up a pmap structure so it can be physically freed. This routine
1398 * is called by the vmspace dtor function. A great deal of pmap data is
1399 * left passively mapped to improve vmspace management so we have a bit
1400 * of cleanup work to do here.
1403 pmap_puninit(pmap_t pmap)
1407 KKASSERT(pmap->pm_active == 0);
1408 lwkt_gettoken(&vm_token);
1409 if ((p = pmap->pm_pdirm) != NULL) {
1410 KKASSERT(pmap->pm_pml4 != NULL);
1411 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1412 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1414 vmstats.v_wire_count--;
1415 KKASSERT((p->flags & PG_BUSY) == 0);
1417 vm_page_free_zero(p);
1418 pmap->pm_pdirm = NULL;
1420 if (pmap->pm_pml4) {
1421 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1422 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1423 pmap->pm_pml4 = NULL;
1425 if (pmap->pm_pteobj) {
1426 vm_object_deallocate(pmap->pm_pteobj);
1427 pmap->pm_pteobj = NULL;
1429 lwkt_reltoken(&vm_token);
1433 * Wire in kernel global address entries. To avoid a race condition
1434 * between pmap initialization and pmap_growkernel, this procedure
1435 * adds the pmap to the master list (which growkernel scans to update),
1436 * then copies the template.
1439 pmap_pinit2(struct pmap *pmap)
1442 lwkt_gettoken(&vm_token);
1443 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1444 /* XXX copies current process, does not fill in MPPTDI */
1445 lwkt_reltoken(&vm_token);
1450 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1451 * 0 on failure (if the procedure had to sleep).
1453 * When asked to remove the page directory page itself, we actually just
1454 * leave it cached so we do not have to incur the SMP inval overhead of
1455 * removing the kernel mapping. pmap_puninit() will take care of it.
1459 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1462 * This code optimizes the case of freeing non-busy
1463 * page-table pages. Those pages are zero now, and
1464 * might as well be placed directly into the zero queue.
1466 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1472 * Remove the page table page from the processes address space.
1474 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1476 * We are the pml4 table itself.
1478 /* XXX anything to do here? */
1479 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1481 * Remove a PDP page from the PML4. We do not maintain
1482 * hold counts on the PML4 page.
1488 m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1489 KKASSERT(m4 != NULL);
1490 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1491 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1492 KKASSERT(pml4[idx] != 0);
1494 } else if (p->pindex >= NUPDE) {
1496 * Remove a PD page from the PDP and drop the hold count
1497 * on the PDP. The PDP is left cached in the pmap if
1498 * the hold count drops to 0 so the wire count remains
1505 m3 = vm_page_lookup(pmap->pm_pteobj,
1506 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1507 KKASSERT(m3 != NULL);
1508 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1509 idx = (p->pindex - NUPDE) % NPDPEPG;
1510 KKASSERT(pdp[idx] != 0);
1515 * Remove a PT page from the PD and drop the hold count
1516 * on the PD. The PD is left cached in the pmap if
1517 * the hold count drops to 0 so the wire count remains
1524 m2 = vm_page_lookup(pmap->pm_pteobj,
1525 NUPDE + p->pindex / NPDEPG);
1526 KKASSERT(m2 != NULL);
1527 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1528 idx = p->pindex % NPDEPG;
1534 * One fewer mappings in the pmap. p's hold count had better
1537 KKASSERT(pmap->pm_stats.resident_count > 0);
1538 --pmap->pm_stats.resident_count;
1540 panic("pmap_release: freeing held page table page");
1541 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1542 pmap->pm_ptphint = NULL;
1545 * We leave the top-level page table page cached, wired, and mapped in
1546 * the pmap until the dtor function (pmap_puninit()) gets called.
1547 * However, still clean it up so we can set PG_ZERO.
1549 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1550 bzero(pmap->pm_pml4, PAGE_SIZE);
1551 vm_page_flag_set(p, PG_ZERO);
1555 KKASSERT(p->wire_count == 0);
1556 vmstats.v_wire_count--;
1557 /* JG eventually revert to using vm_page_free_zero() */
1564 * This routine is called when various levels in the page table need to
1565 * be populated. This routine cannot fail.
1569 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1574 * Find or fabricate a new pagetable page. This will busy the page.
1576 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1577 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1578 if ((m->flags & PG_ZERO) == 0) {
1579 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1583 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1587 KASSERT(m->queue == PQ_NONE,
1588 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1591 * Increment the hold count for the page we will be returning to
1595 if (m->wire_count++ == 0)
1596 vmstats.v_wire_count++;
1597 m->valid = VM_PAGE_BITS_ALL;
1598 vm_page_flag_clear(m, PG_ZERO);
1601 * Map the pagetable page into the process address space, if
1602 * it isn't already there.
1604 * It is possible that someone else got in and mapped the page
1605 * directory page while we were blocked, if so just unbusy and
1606 * return the held page.
1608 if (ptepindex >= (NUPDE + NUPDPE)) {
1610 * Wire up a new PDP page in the PML4
1612 vm_pindex_t pml4index;
1615 pml4index = ptepindex - (NUPDE + NUPDPE);
1616 pml4 = &pmap->pm_pml4[pml4index];
1618 if (--m->wire_count == 0)
1619 --vmstats.v_wire_count;
1623 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1624 } else if (ptepindex >= NUPDE) {
1626 * Wire up a new PD page in the PDP
1628 vm_pindex_t pml4index;
1629 vm_pindex_t pdpindex;
1634 pdpindex = ptepindex - NUPDE;
1635 pml4index = pdpindex >> NPML4EPGSHIFT;
1637 pml4 = &pmap->pm_pml4[pml4index];
1638 if ((*pml4 & PG_V) == 0) {
1640 * Have to allocate a new PDP page, recurse.
1641 * This always succeeds. Returned page will
1644 pdppg = _pmap_allocpte(pmap,
1645 NUPDE + NUPDPE + pml4index);
1648 * Add a held reference to the PDP page.
1650 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1651 pdppg->hold_count++;
1655 * Now find the pdp_entry and map the PDP. If the PDP
1656 * has already been mapped unwind and return the
1657 * already-mapped PDP held.
1659 * pdppg is left held (hold_count is incremented for
1660 * each PD in the PDP).
1662 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1663 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1665 vm_page_unhold(pdppg);
1666 if (--m->wire_count == 0)
1667 --vmstats.v_wire_count;
1671 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1674 * Wire up the new PT page in the PD
1676 vm_pindex_t pml4index;
1677 vm_pindex_t pdpindex;
1683 pdpindex = ptepindex >> NPDPEPGSHIFT;
1684 pml4index = pdpindex >> NPML4EPGSHIFT;
1687 * Locate the PDP page in the PML4, then the PD page in
1688 * the PDP. If either does not exist we simply recurse
1691 * We can just recurse on the PD page as it will recurse
1692 * on the PDP if necessary.
1694 pml4 = &pmap->pm_pml4[pml4index];
1695 if ((*pml4 & PG_V) == 0) {
1696 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1697 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1698 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1700 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1701 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1702 if ((*pdp & PG_V) == 0) {
1703 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1705 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1711 * Now fill in the pte in the PD. If the pte already exists
1712 * (again, if we raced the grab), unhold pdpg and unwire
1713 * m, returning a held m.
1715 * pdpg is left held (hold_count is incremented for
1716 * each PT in the PD).
1718 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
1719 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1721 vm_page_unhold(pdpg);
1722 if (--m->wire_count == 0)
1723 --vmstats.v_wire_count;
1727 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1731 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1732 * valid bits, mapped flag, unbusy, and we're done.
1734 pmap->pm_ptphint = m;
1735 ++pmap->pm_stats.resident_count;
1738 m->valid = VM_PAGE_BITS_ALL;
1739 vm_page_flag_clear(m, PG_ZERO);
1741 vm_page_flag_set(m, PG_MAPPED);
1749 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1751 vm_pindex_t ptepindex;
1756 * Calculate pagetable page index
1758 ptepindex = pmap_pde_pindex(va);
1761 * Get the page directory entry
1763 pd = pmap_pde(pmap, va);
1766 * This supports switching from a 2MB page to a
1769 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1770 panic("no promotion/demotion yet");
1778 * If the page table page is mapped, we just increment the
1779 * hold count, and activate it.
1781 if (pd != NULL && (*pd & PG_V) != 0) {
1782 /* YYY hint is used here on i386 */
1783 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1784 pmap->pm_ptphint = m;
1789 * Here if the pte page isn't mapped, or if it has been deallocated.
1791 return _pmap_allocpte(pmap, ptepindex);
1795 /***************************************************
1796 * Pmap allocation/deallocation routines.
1797 ***************************************************/
1800 * Release any resources held by the given physical map.
1801 * Called when a pmap initialized by pmap_pinit is being released.
1802 * Should only be called if the map contains no valid mappings.
1804 static int pmap_release_callback(struct vm_page *p, void *data);
1807 pmap_release(struct pmap *pmap)
1809 vm_object_t object = pmap->pm_pteobj;
1810 struct rb_vm_page_scan_info info;
1812 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1813 #if defined(DIAGNOSTIC)
1814 if (object->ref_count != 1)
1815 panic("pmap_release: pteobj reference count != 1");
1819 info.object = object;
1821 lwkt_gettoken(&vm_token);
1822 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1829 info.limit = object->generation;
1831 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1832 pmap_release_callback, &info);
1833 if (info.error == 0 && info.mpte) {
1834 if (!pmap_release_free_page(pmap, info.mpte))
1838 } while (info.error);
1839 lwkt_reltoken(&vm_token);
1844 pmap_release_callback(struct vm_page *p, void *data)
1846 struct rb_vm_page_scan_info *info = data;
1848 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1852 if (!pmap_release_free_page(info->pmap, p)) {
1856 if (info->object->generation != info->limit) {
1864 * Grow the number of kernel page table entries, if needed.
1866 * This routine is always called to validate any address space
1867 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
1868 * space below KERNBASE.
1871 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1874 vm_offset_t ptppaddr;
1876 pd_entry_t *pde, newpdir;
1878 int update_kernel_vm_end;
1881 lwkt_gettoken(&vm_token);
1884 * bootstrap kernel_vm_end on first real VM use
1886 if (kernel_vm_end == 0) {
1887 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1889 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1890 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1891 ~(PAGE_SIZE * NPTEPG - 1);
1893 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1894 kernel_vm_end = kernel_map.max_offset;
1901 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
1902 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
1903 * do not want to force-fill 128G worth of page tables.
1905 if (kstart < KERNBASE) {
1906 if (kstart > kernel_vm_end)
1907 kstart = kernel_vm_end;
1908 KKASSERT(kend <= KERNBASE);
1909 update_kernel_vm_end = 1;
1911 update_kernel_vm_end = 0;
1914 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
1915 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
1917 if (kend - 1 >= kernel_map.max_offset)
1918 kend = kernel_map.max_offset;
1920 while (kstart < kend) {
1921 pde = pmap_pde(&kernel_pmap, kstart);
1923 /* We need a new PDP entry */
1924 nkpg = vm_page_alloc(kptobj, nkpt,
1927 VM_ALLOC_INTERRUPT);
1929 panic("pmap_growkernel: no memory to grow "
1932 paddr = VM_PAGE_TO_PHYS(nkpg);
1933 if ((nkpg->flags & PG_ZERO) == 0)
1934 pmap_zero_page(paddr);
1935 vm_page_flag_clear(nkpg, PG_ZERO);
1936 newpdp = (pdp_entry_t)
1937 (paddr | PG_V | PG_RW | PG_A | PG_M);
1938 *pmap_pdpe(&kernel_pmap, kstart) = newpdp;
1940 continue; /* try again */
1942 if ((*pde & PG_V) != 0) {
1943 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1944 ~(PAGE_SIZE * NPTEPG - 1);
1945 if (kstart - 1 >= kernel_map.max_offset) {
1946 kstart = kernel_map.max_offset;
1953 * This index is bogus, but out of the way
1955 nkpg = vm_page_alloc(kptobj, nkpt,
1958 VM_ALLOC_INTERRUPT);
1960 panic("pmap_growkernel: no memory to grow kernel");
1963 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1964 pmap_zero_page(ptppaddr);
1965 vm_page_flag_clear(nkpg, PG_ZERO);
1966 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1967 *pmap_pde(&kernel_pmap, kstart) = newpdir;
1970 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1971 ~(PAGE_SIZE * NPTEPG - 1);
1973 if (kstart - 1 >= kernel_map.max_offset) {
1974 kstart = kernel_map.max_offset;
1980 * Only update kernel_vm_end for areas below KERNBASE.
1982 if (update_kernel_vm_end && kernel_vm_end < kstart)
1983 kernel_vm_end = kstart;
1985 lwkt_reltoken(&vm_token);
1990 * Retire the given physical map from service.
1991 * Should only be called if the map contains
1992 * no valid mappings.
1995 pmap_destroy(pmap_t pmap)
2002 lwkt_gettoken(&vm_token);
2003 count = --pmap->pm_count;
2006 panic("destroying a pmap is not yet implemented");
2008 lwkt_reltoken(&vm_token);
2012 * Add a reference to the specified pmap.
2015 pmap_reference(pmap_t pmap)
2018 lwkt_gettoken(&vm_token);
2020 lwkt_reltoken(&vm_token);
2024 /***************************************************
2025 * page management routines.
2026 ***************************************************/
2029 * free the pv_entry back to the free list. This function may be
2030 * called from an interrupt.
2034 free_pv_entry(pv_entry_t pv)
2037 KKASSERT(pv_entry_count >= 0);
2042 * get a new pv_entry, allocating a block from the system
2043 * when needed. This function may be called from an interrupt.
2050 if (pv_entry_high_water &&
2051 (pv_entry_count > pv_entry_high_water) &&
2052 (pmap_pagedaemon_waken == 0)) {
2053 pmap_pagedaemon_waken = 1;
2054 wakeup(&vm_pages_needed);
2056 return zalloc(pvzone);
2060 * This routine is very drastic, but can save the system
2068 static int warningdone=0;
2070 if (pmap_pagedaemon_waken == 0)
2072 lwkt_gettoken(&vm_token);
2073 if (warningdone < 5) {
2074 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
2078 for(i = 0; i < vm_page_array_size; i++) {
2079 m = &vm_page_array[i];
2080 if (m->wire_count || m->hold_count || m->busy ||
2081 (m->flags & PG_BUSY))
2085 pmap_pagedaemon_waken = 0;
2086 lwkt_reltoken(&vm_token);
2091 * If it is the first entry on the list, it is actually
2092 * in the header and we must copy the following entry up
2093 * to the header. Otherwise we must search the list for
2094 * the entry. In either case we free the now unused entry.
2098 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
2099 vm_offset_t va, pmap_inval_info_t info)
2105 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
2106 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2107 if (pmap == pv->pv_pmap && va == pv->pv_va)
2111 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2112 if (va == pv->pv_va)
2120 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2121 m->md.pv_list_count--;
2122 m->object->agg_pv_list_count--;
2123 KKASSERT(m->md.pv_list_count >= 0);
2124 if (TAILQ_EMPTY(&m->md.pv_list))
2125 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2126 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2127 ++pmap->pm_generation;
2128 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
2136 * Create a pv entry for page at pa for
2141 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2146 pv = get_pv_entry();
2151 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2152 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2153 ++pmap->pm_generation;
2154 m->md.pv_list_count++;
2155 m->object->agg_pv_list_count++;
2161 * pmap_remove_pte: do the things to unmap a page in a process
2165 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2166 pmap_inval_info_t info)
2171 pmap_inval_interlock(info, pmap, va);
2172 oldpte = pte_load_clear(ptq);
2173 pmap_inval_deinterlock(info, pmap);
2175 pmap->pm_stats.wired_count -= 1;
2177 * Machines that don't support invlpg, also don't support
2178 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2182 cpu_invlpg((void *)va);
2183 KKASSERT(pmap->pm_stats.resident_count > 0);
2184 --pmap->pm_stats.resident_count;
2185 if (oldpte & PG_MANAGED) {
2186 m = PHYS_TO_VM_PAGE(oldpte);
2187 if (oldpte & PG_M) {
2188 #if defined(PMAP_DIAGNOSTIC)
2189 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2191 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2195 if (pmap_track_modified(va))
2199 vm_page_flag_set(m, PG_REFERENCED);
2200 return pmap_remove_entry(pmap, m, va, info);
2202 return pmap_unuse_pt(pmap, va, NULL, info);
2211 * Remove a single page from a process address space.
2213 * This function may not be called from an interrupt if the pmap is
2218 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2222 pte = pmap_pte(pmap, va);
2225 if ((*pte & PG_V) == 0)
2227 pmap_remove_pte(pmap, pte, va, info);
2233 * Remove the given range of addresses from the specified map.
2235 * It is assumed that the start and end are properly
2236 * rounded to the page size.
2238 * This function may not be called from an interrupt if the pmap is
2242 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2244 vm_offset_t va_next;
2245 pml4_entry_t *pml4e;
2247 pd_entry_t ptpaddr, *pde;
2249 struct pmap_inval_info info;
2254 lwkt_gettoken(&vm_token);
2255 if (pmap->pm_stats.resident_count == 0) {
2256 lwkt_reltoken(&vm_token);
2260 pmap_inval_init(&info);
2263 * special handling of removing one page. a very
2264 * common operation and easy to short circuit some
2267 if (sva + PAGE_SIZE == eva) {
2268 pde = pmap_pde(pmap, sva);
2269 if (pde && (*pde & PG_PS) == 0) {
2270 pmap_remove_page(pmap, sva, &info);
2271 pmap_inval_done(&info);
2272 lwkt_reltoken(&vm_token);
2277 for (; sva < eva; sva = va_next) {
2278 pml4e = pmap_pml4e(pmap, sva);
2279 if ((*pml4e & PG_V) == 0) {
2280 va_next = (sva + NBPML4) & ~PML4MASK;
2286 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2287 if ((*pdpe & PG_V) == 0) {
2288 va_next = (sva + NBPDP) & ~PDPMASK;
2295 * Calculate index for next page table.
2297 va_next = (sva + NBPDR) & ~PDRMASK;
2301 pde = pmap_pdpe_to_pde(pdpe, sva);
2305 * Weed out invalid mappings.
2311 * Check for large page.
2313 if ((ptpaddr & PG_PS) != 0) {
2314 /* JG FreeBSD has more complex treatment here */
2315 pmap_inval_interlock(&info, pmap, -1);
2317 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2318 pmap_inval_deinterlock(&info, pmap);
2323 * Limit our scan to either the end of the va represented
2324 * by the current page table page, or to the end of the
2325 * range being removed.
2331 * NOTE: pmap_remove_pte() can block.
2333 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2337 if (pmap_remove_pte(pmap, pte, sva, &info))
2341 pmap_inval_done(&info);
2342 lwkt_reltoken(&vm_token);
2348 * Removes this physical page from all physical maps in which it resides.
2349 * Reflects back modify bits to the pager.
2351 * This routine may not be called from an interrupt.
2356 pmap_remove_all(vm_page_t m)
2358 struct pmap_inval_info info;
2359 pt_entry_t *pte, tpte;
2362 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2365 lwkt_gettoken(&vm_token);
2366 pmap_inval_init(&info);
2368 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2369 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2370 --pv->pv_pmap->pm_stats.resident_count;
2372 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2373 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
2374 tpte = pte_load_clear(pte);
2376 pv->pv_pmap->pm_stats.wired_count--;
2377 pmap_inval_deinterlock(&info, pv->pv_pmap);
2379 vm_page_flag_set(m, PG_REFERENCED);
2382 * Update the vm_page_t clean and reference bits.
2385 #if defined(PMAP_DIAGNOSTIC)
2386 if (pmap_nw_modified(tpte)) {
2388 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2392 if (pmap_track_modified(pv->pv_va))
2395 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2396 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2397 ++pv->pv_pmap->pm_generation;
2398 m->md.pv_list_count--;
2399 m->object->agg_pv_list_count--;
2400 KKASSERT(m->md.pv_list_count >= 0);
2401 if (TAILQ_EMPTY(&m->md.pv_list))
2402 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2403 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2407 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2408 pmap_inval_done(&info);
2409 lwkt_reltoken(&vm_token);
2415 * Set the physical protection on the specified range of this map
2418 * This function may not be called from an interrupt if the map is
2419 * not the kernel_pmap.
2422 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2424 vm_offset_t va_next;
2425 pml4_entry_t *pml4e;
2427 pd_entry_t ptpaddr, *pde;
2429 pmap_inval_info info;
2431 /* JG review for NX */
2436 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2437 pmap_remove(pmap, sva, eva);
2441 if (prot & VM_PROT_WRITE)
2444 lwkt_gettoken(&vm_token);
2445 pmap_inval_init(&info);
2447 for (; sva < eva; sva = va_next) {
2449 pml4e = pmap_pml4e(pmap, sva);
2450 if ((*pml4e & PG_V) == 0) {
2451 va_next = (sva + NBPML4) & ~PML4MASK;
2457 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2458 if ((*pdpe & PG_V) == 0) {
2459 va_next = (sva + NBPDP) & ~PDPMASK;
2465 va_next = (sva + NBPDR) & ~PDRMASK;
2469 pde = pmap_pdpe_to_pde(pdpe, sva);
2473 * Check for large page.
2475 if ((ptpaddr & PG_PS) != 0) {
2476 pmap_inval_interlock(&info, pmap, -1);
2477 *pde &= ~(PG_M|PG_RW);
2478 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2479 pmap_inval_deinterlock(&info, pmap);
2484 * Weed out invalid mappings. Note: we assume that the page
2485 * directory table is always allocated, and in kernel virtual.
2493 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2500 * XXX non-optimal. Note also that there can be
2501 * no pmap_inval_flush() calls until after we modify
2502 * ptbase[sindex] (or otherwise we have to do another
2503 * pmap_inval_add() call).
2505 pmap_inval_interlock(&info, pmap, sva);
2509 if ((pbits & PG_V) == 0) {
2510 pmap_inval_deinterlock(&info, pmap);
2513 if (pbits & PG_MANAGED) {
2516 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2517 vm_page_flag_set(m, PG_REFERENCED);
2521 if (pmap_track_modified(sva)) {
2523 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2530 if (pbits != cbits &&
2531 !atomic_cmpset_long(pte, pbits, cbits)) {
2534 pmap_inval_deinterlock(&info, pmap);
2537 pmap_inval_done(&info);
2538 lwkt_reltoken(&vm_token);
2542 * Insert the given physical page (p) at
2543 * the specified virtual address (v) in the
2544 * target physical map with the protection requested.
2546 * If specified, the page will be wired down, meaning
2547 * that the related pte can not be reclaimed.
2549 * NB: This is the only routine which MAY NOT lazy-evaluate
2550 * or lose information. That is, this routine must actually
2551 * insert this page into the given map NOW.
2554 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2561 pt_entry_t origpte, newpte;
2563 pmap_inval_info info;
2568 va = trunc_page(va);
2569 #ifdef PMAP_DIAGNOSTIC
2571 panic("pmap_enter: toobig");
2572 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2573 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
2575 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2576 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2578 db_print_backtrace();
2581 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2582 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2584 db_print_backtrace();
2588 lwkt_gettoken(&vm_token);
2591 * In the case that a page table page is not
2592 * resident, we are creating it here.
2594 if (va < VM_MAX_USER_ADDRESS)
2595 mpte = pmap_allocpte(pmap, va);
2599 pmap_inval_init(&info);
2600 pde = pmap_pde(pmap, va);
2601 if (pde != NULL && (*pde & PG_V) != 0) {
2602 if ((*pde & PG_PS) != 0)
2603 panic("pmap_enter: attempted pmap_enter on 2MB page");
2604 pte = pmap_pde_to_pte(pde, va);
2606 panic("pmap_enter: invalid page directory va=%#lx", va);
2608 KKASSERT(pte != NULL);
2609 pa = VM_PAGE_TO_PHYS(m);
2611 opa = origpte & PG_FRAME;
2614 * Mapping has not changed, must be protection or wiring change.
2616 if (origpte && (opa == pa)) {
2618 * Wiring change, just update stats. We don't worry about
2619 * wiring PT pages as they remain resident as long as there
2620 * are valid mappings in them. Hence, if a user page is wired,
2621 * the PT page will be also.
2623 if (wired && ((origpte & PG_W) == 0))
2624 pmap->pm_stats.wired_count++;
2625 else if (!wired && (origpte & PG_W))
2626 pmap->pm_stats.wired_count--;
2628 #if defined(PMAP_DIAGNOSTIC)
2629 if (pmap_nw_modified(origpte)) {
2631 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2637 * Remove the extra pte reference. Note that we cannot
2638 * optimize the RO->RW case because we have adjusted the
2639 * wiring count above and may need to adjust the wiring
2646 * We might be turning off write access to the page,
2647 * so we go ahead and sense modify status.
2649 if (origpte & PG_MANAGED) {
2650 if ((origpte & PG_M) && pmap_track_modified(va)) {
2652 om = PHYS_TO_VM_PAGE(opa);
2656 KKASSERT(m->flags & PG_MAPPED);
2661 * Mapping has changed, invalidate old range and fall through to
2662 * handle validating new mapping.
2666 err = pmap_remove_pte(pmap, pte, va, &info);
2668 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2670 opa = origpte & PG_FRAME;
2672 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2678 * Enter on the PV list if part of our managed memory. Note that we
2679 * raise IPL while manipulating pv_table since pmap_enter can be
2680 * called at interrupt time.
2682 if (pmap_initialized &&
2683 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2684 pmap_insert_entry(pmap, va, mpte, m);
2686 vm_page_flag_set(m, PG_MAPPED);
2690 * Increment counters
2692 ++pmap->pm_stats.resident_count;
2694 pmap->pm_stats.wired_count++;
2698 * Now validate mapping with desired protection/wiring.
2700 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
2704 if (va < VM_MAX_USER_ADDRESS)
2706 if (pmap == &kernel_pmap)
2710 * if the mapping or permission bits are different, we need
2711 * to update the pte.
2713 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2714 pmap_inval_interlock(&info, pmap, va);
2715 *pte = newpte | PG_A;
2716 pmap_inval_deinterlock(&info, pmap);
2718 vm_page_flag_set(m, PG_WRITEABLE);
2720 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2721 pmap_inval_done(&info);
2722 lwkt_reltoken(&vm_token);
2726 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2727 * This code also assumes that the pmap has no pre-existing entry for this
2730 * This code currently may only be used on user pmaps, not kernel_pmap.
2733 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2738 vm_pindex_t ptepindex;
2740 pmap_inval_info info;
2742 lwkt_gettoken(&vm_token);
2743 pmap_inval_init(&info);
2745 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2746 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2748 db_print_backtrace();
2751 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2752 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2754 db_print_backtrace();
2758 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2761 * Calculate the page table page (mpte), allocating it if necessary.
2763 * A held page table page (mpte), or NULL, is passed onto the
2764 * section following.
2766 if (va < VM_MAX_USER_ADDRESS) {
2768 * Calculate pagetable page index
2770 ptepindex = pmap_pde_pindex(va);
2774 * Get the page directory entry
2776 ptepa = pmap_pde(pmap, va);
2779 * If the page table page is mapped, we just increment
2780 * the hold count, and activate it.
2782 if (ptepa && (*ptepa & PG_V) != 0) {
2784 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2785 // if (pmap->pm_ptphint &&
2786 // (pmap->pm_ptphint->pindex == ptepindex)) {
2787 // mpte = pmap->pm_ptphint;
2789 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2790 pmap->pm_ptphint = mpte;
2795 mpte = _pmap_allocpte(pmap, ptepindex);
2797 } while (mpte == NULL);
2800 /* this code path is not yet used */
2804 * With a valid (and held) page directory page, we can just use
2805 * vtopte() to get to the pte. If the pte is already present
2806 * we do not disturb it.
2811 pmap_unwire_pte_hold(pmap, va, mpte, &info);
2812 pa = VM_PAGE_TO_PHYS(m);
2813 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2814 pmap_inval_done(&info);
2815 lwkt_reltoken(&vm_token);
2820 * Enter on the PV list if part of our managed memory
2822 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2823 pmap_insert_entry(pmap, va, mpte, m);
2824 vm_page_flag_set(m, PG_MAPPED);
2828 * Increment counters
2830 ++pmap->pm_stats.resident_count;
2832 pa = VM_PAGE_TO_PHYS(m);
2835 * Now validate mapping with RO protection
2837 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2838 *pte = pa | PG_V | PG_U;
2840 *pte = pa | PG_V | PG_U | PG_MANAGED;
2841 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2842 pmap_inval_done(&info);
2843 lwkt_reltoken(&vm_token);
2847 * Make a temporary mapping for a physical address. This is only intended
2848 * to be used for panic dumps.
2850 /* JG Needed on x86_64? */
2852 pmap_kenter_temporary(vm_paddr_t pa, long i)
2854 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2855 return ((void *)crashdumpmap);
2858 #define MAX_INIT_PT (96)
2861 * This routine preloads the ptes for a given object into the specified pmap.
2862 * This eliminates the blast of soft faults on process startup and
2863 * immediately after an mmap.
2865 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2868 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2869 vm_object_t object, vm_pindex_t pindex,
2870 vm_size_t size, int limit)
2872 struct rb_vm_page_scan_info info;
2877 * We can't preinit if read access isn't set or there is no pmap
2880 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2884 * We can't preinit if the pmap is not the current pmap
2886 lp = curthread->td_lwp;
2887 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2890 psize = x86_64_btop(size);
2892 if ((object->type != OBJT_VNODE) ||
2893 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2894 (object->resident_page_count > MAX_INIT_PT))) {
2898 if (psize + pindex > object->size) {
2899 if (object->size < pindex)
2901 psize = object->size - pindex;
2908 * Use a red-black scan to traverse the requested range and load
2909 * any valid pages found into the pmap.
2911 * We cannot safely scan the object's memq unless we are in a
2912 * critical section since interrupts can remove pages from objects.
2914 info.start_pindex = pindex;
2915 info.end_pindex = pindex + psize - 1;
2922 lwkt_gettoken(&vm_token);
2923 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2924 pmap_object_init_pt_callback, &info);
2925 lwkt_reltoken(&vm_token);
2931 pmap_object_init_pt_callback(vm_page_t p, void *data)
2933 struct rb_vm_page_scan_info *info = data;
2934 vm_pindex_t rel_index;
2936 * don't allow an madvise to blow away our really
2937 * free pages allocating pv entries.
2939 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2940 vmstats.v_free_count < vmstats.v_free_reserved) {
2943 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2944 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2945 if ((p->queue - p->pc) == PQ_CACHE)
2946 vm_page_deactivate(p);
2948 rel_index = p->pindex - info->start_pindex;
2949 pmap_enter_quick(info->pmap,
2950 info->addr + x86_64_ptob(rel_index), p);
2957 * Return TRUE if the pmap is in shape to trivially
2958 * pre-fault the specified address.
2960 * Returns FALSE if it would be non-trivial or if a
2961 * pte is already loaded into the slot.
2964 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2970 lwkt_gettoken(&vm_token);
2971 pde = pmap_pde(pmap, addr);
2972 if (pde == NULL || *pde == 0) {
2976 ret = (*pte) ? 0 : 1;
2978 lwkt_reltoken(&vm_token);
2983 * Routine: pmap_change_wiring
2984 * Function: Change the wiring attribute for a map/virtual-address
2986 * In/out conditions:
2987 * The mapping must already exist in the pmap.
2990 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2997 lwkt_gettoken(&vm_token);
2998 pte = pmap_pte(pmap, va);
3000 if (wired && !pmap_pte_w(pte))
3001 pmap->pm_stats.wired_count++;
3002 else if (!wired && pmap_pte_w(pte))
3003 pmap->pm_stats.wired_count--;
3006 * Wiring is not a hardware characteristic so there is no need to
3007 * invalidate TLB. However, in an SMP environment we must use
3008 * a locked bus cycle to update the pte (if we are not using
3009 * the pmap_inval_*() API that is)... it's ok to do this for simple
3014 atomic_set_long(pte, PG_W);
3016 atomic_clear_long(pte, PG_W);
3019 atomic_set_long_nonlocked(pte, PG_W);
3021 atomic_clear_long_nonlocked(pte, PG_W);
3023 lwkt_reltoken(&vm_token);
3029 * Copy the range specified by src_addr/len
3030 * from the source map to the range dst_addr/len
3031 * in the destination map.
3033 * This routine is only advisory and need not do anything.
3036 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3037 vm_size_t len, vm_offset_t src_addr)
3041 pmap_inval_info info;
3043 vm_offset_t end_addr = src_addr + len;
3045 pd_entry_t src_frame, dst_frame;
3048 if (dst_addr != src_addr)
3051 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3052 if (src_frame != (PTDpde & PG_FRAME)) {
3056 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3057 if (dst_frame != (APTDpde & PG_FRAME)) {
3058 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
3059 /* The page directory is not shared between CPUs */
3063 pmap_inval_init(&info);
3064 pmap_inval_add(&info, dst_pmap, -1);
3065 pmap_inval_add(&info, src_pmap, -1);
3068 * critical section protection is required to maintain the page/object
3069 * association, interrupts can free pages and remove them from
3073 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
3074 pt_entry_t *src_pte, *dst_pte;
3075 vm_page_t dstmpte, srcmpte;
3076 vm_offset_t srcptepaddr;
3077 vm_pindex_t ptepindex;
3079 if (addr >= UPT_MIN_ADDRESS)
3080 panic("pmap_copy: invalid to pmap_copy page tables\n");
3083 * Don't let optional prefaulting of pages make us go
3084 * way below the low water mark of free pages or way
3085 * above high water mark of used pv entries.
3087 if (vmstats.v_free_count < vmstats.v_free_reserved ||
3088 pv_entry_count > pv_entry_high_water)
3091 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
3092 ptepindex = addr >> PDRSHIFT;
3095 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
3097 if (srcptepaddr == 0)
3100 if (srcptepaddr & PG_PS) {
3102 if (dst_pmap->pm_pdir[ptepindex] == 0) {
3103 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
3104 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
3110 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
3111 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
3112 (srcmpte->flags & PG_BUSY)) {
3116 if (pdnxt > end_addr)
3119 src_pte = vtopte(addr);
3121 dst_pte = avtopte(addr);
3123 while (addr < pdnxt) {
3128 * we only virtual copy managed pages
3130 if ((ptetemp & PG_MANAGED) != 0) {
3132 * We have to check after allocpte for the
3133 * pte still being around... allocpte can
3136 * pmap_allocpte() can block. If we lose
3137 * our page directory mappings we stop.
3139 dstmpte = pmap_allocpte(dst_pmap, addr);
3142 if (src_frame != (PTDpde & PG_FRAME) ||
3143 dst_frame != (APTDpde & PG_FRAME)
3145 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3146 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3148 } else if ((*dst_pte == 0) &&
3149 (ptetemp = *src_pte) != 0 &&
3150 (ptetemp & PG_MANAGED)) {
3152 * Clear the modified and
3153 * accessed (referenced) bits
3156 m = PHYS_TO_VM_PAGE(ptetemp);
3157 *dst_pte = ptetemp & ~(PG_M | PG_A);
3158 ++dst_pmap->pm_stats.resident_count;
3159 pmap_insert_entry(dst_pmap, addr,
3161 KKASSERT(m->flags & PG_MAPPED);
3163 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3164 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3168 if (dstmpte->hold_count >= srcmpte->hold_count)
3178 pmap_inval_done(&info);
3185 * Zero the specified physical page.
3187 * This function may be called from an interrupt and no locking is
3191 pmap_zero_page(vm_paddr_t phys)
3193 vm_offset_t va = PHYS_TO_DMAP(phys);
3195 pagezero((void *)va);
3199 * pmap_page_assertzero:
3201 * Assert that a page is empty, panic if it isn't.
3204 pmap_page_assertzero(vm_paddr_t phys)
3206 vm_offset_t va = PHYS_TO_DMAP(phys);
3209 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3210 if (*(long *)((char *)va + i) != 0) {
3211 panic("pmap_page_assertzero() @ %p not zero!\n",
3212 (void *)(intptr_t)va);
3220 * Zero part of a physical page by mapping it into memory and clearing
3221 * its contents with bzero.
3223 * off and size may not cover an area beyond a single hardware page.
3226 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3228 vm_offset_t virt = PHYS_TO_DMAP(phys);
3230 bzero((char *)virt + off, size);
3236 * Copy the physical page from the source PA to the target PA.
3237 * This function may be called from an interrupt. No locking
3241 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3243 vm_offset_t src_virt, dst_virt;
3245 src_virt = PHYS_TO_DMAP(src);
3246 dst_virt = PHYS_TO_DMAP(dst);
3247 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3251 * pmap_copy_page_frag:
3253 * Copy the physical page from the source PA to the target PA.
3254 * This function may be called from an interrupt. No locking
3258 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3260 vm_offset_t src_virt, dst_virt;
3262 src_virt = PHYS_TO_DMAP(src);
3263 dst_virt = PHYS_TO_DMAP(dst);
3265 bcopy((char *)src_virt + (src & PAGE_MASK),
3266 (char *)dst_virt + (dst & PAGE_MASK),
3271 * Returns true if the pmap's pv is one of the first
3272 * 16 pvs linked to from this page. This count may
3273 * be changed upwards or downwards in the future; it
3274 * is only necessary that true be returned for a small
3275 * subset of pmaps for proper page aging.
3278 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3283 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3287 lwkt_gettoken(&vm_token);
3289 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3290 if (pv->pv_pmap == pmap) {
3291 lwkt_reltoken(&vm_token);
3299 lwkt_reltoken(&vm_token);
3305 * Remove all pages from specified address space
3306 * this aids process exit speeds. Also, this code
3307 * is special cased for current process only, but
3308 * can have the more generic (and slightly slower)
3309 * mode enabled. This is much faster than pmap_remove
3310 * in the case of running down an entire address space.
3313 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3316 pt_entry_t *pte, tpte;
3319 pmap_inval_info info;
3321 int save_generation;
3323 lp = curthread->td_lwp;
3324 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3329 lwkt_gettoken(&vm_token);
3330 pmap_inval_init(&info);
3331 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3332 if (pv->pv_va >= eva || pv->pv_va < sva) {
3333 npv = TAILQ_NEXT(pv, pv_plist);
3337 KKASSERT(pmap == pv->pv_pmap);
3340 pte = vtopte(pv->pv_va);
3342 pte = pmap_pte_quick(pmap, pv->pv_va);
3343 pmap_inval_interlock(&info, pmap, pv->pv_va);
3346 * We cannot remove wired pages from a process' mapping
3350 pmap_inval_deinterlock(&info, pmap);
3351 npv = TAILQ_NEXT(pv, pv_plist);
3354 tpte = pte_load_clear(pte);
3356 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3358 KASSERT(m < &vm_page_array[vm_page_array_size],
3359 ("pmap_remove_pages: bad tpte %lx", tpte));
3361 KKASSERT(pmap->pm_stats.resident_count > 0);
3362 --pmap->pm_stats.resident_count;
3363 pmap_inval_deinterlock(&info, pmap);
3366 * Update the vm_page_t clean and reference bits.
3372 npv = TAILQ_NEXT(pv, pv_plist);
3373 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3374 save_generation = ++pmap->pm_generation;
3376 m->md.pv_list_count--;
3377 m->object->agg_pv_list_count--;
3378 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3379 if (TAILQ_EMPTY(&m->md.pv_list))
3380 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3382 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3386 * Restart the scan if we blocked during the unuse or free
3387 * calls and other removals were made.
3389 if (save_generation != pmap->pm_generation) {
3390 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3391 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3394 pmap_inval_done(&info);
3395 lwkt_reltoken(&vm_token);
3399 * pmap_testbit tests bits in pte's
3400 * note that the testbit/clearbit routines are inline,
3401 * and a lot of things compile-time evaluate.
3405 pmap_testbit(vm_page_t m, int bit)
3410 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3413 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3418 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3420 * if the bit being tested is the modified bit, then
3421 * mark clean_map and ptes as never
3424 if (bit & (PG_A|PG_M)) {
3425 if (!pmap_track_modified(pv->pv_va))
3429 #if defined(PMAP_DIAGNOSTIC)
3430 if (pv->pv_pmap == NULL) {
3431 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3435 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3446 * this routine is used to modify bits in ptes
3450 pmap_clearbit(vm_page_t m, int bit)
3452 struct pmap_inval_info info;
3457 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3460 pmap_inval_init(&info);
3463 * Loop over all current mappings setting/clearing as appropos If
3464 * setting RO do we need to clear the VAC?
3466 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3468 * don't write protect pager mappings
3471 if (!pmap_track_modified(pv->pv_va))
3475 #if defined(PMAP_DIAGNOSTIC)
3476 if (pv->pv_pmap == NULL) {
3477 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3483 * Careful here. We can use a locked bus instruction to
3484 * clear PG_A or PG_M safely but we need to synchronize
3485 * with the target cpus when we mess with PG_RW.
3487 * We do not have to force synchronization when clearing
3488 * PG_M even for PTEs generated via virtual memory maps,
3489 * because the virtual kernel will invalidate the pmap
3490 * entry when/if it needs to resynchronize the Modify bit.
3493 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
3494 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3501 atomic_clear_long(pte, PG_M|PG_RW);
3504 * The cpu may be trying to set PG_M
3505 * simultaniously with our clearing
3508 if (!atomic_cmpset_long(pte, pbits,
3512 } else if (bit == PG_M) {
3514 * We could also clear PG_RW here to force
3515 * a fault on write to redetect PG_M for
3516 * virtual kernels, but it isn't necessary
3517 * since virtual kernels invalidate the pte
3518 * when they clear the VPTE_M bit in their
3519 * virtual page tables.
3521 atomic_clear_long(pte, PG_M);
3523 atomic_clear_long(pte, bit);
3527 pmap_inval_deinterlock(&info, pv->pv_pmap);
3529 pmap_inval_done(&info);
3533 * pmap_page_protect:
3535 * Lower the permission for all mappings to a given page.
3538 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3540 /* JG NX support? */
3541 if ((prot & VM_PROT_WRITE) == 0) {
3542 lwkt_gettoken(&vm_token);
3543 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3544 pmap_clearbit(m, PG_RW);
3545 vm_page_flag_clear(m, PG_WRITEABLE);
3549 lwkt_reltoken(&vm_token);
3554 pmap_phys_address(vm_pindex_t ppn)
3556 return (x86_64_ptob(ppn));
3560 * pmap_ts_referenced:
3562 * Return a count of reference bits for a page, clearing those bits.
3563 * It is not necessary for every reference bit to be cleared, but it
3564 * is necessary that 0 only be returned when there are truly no
3565 * reference bits set.
3567 * XXX: The exact number of bits to check and clear is a matter that
3568 * should be tested and standardized at some point in the future for
3569 * optimal aging of shared pages.
3572 pmap_ts_referenced(vm_page_t m)
3574 pv_entry_t pv, pvf, pvn;
3578 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3582 lwkt_gettoken(&vm_token);
3584 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3589 pvn = TAILQ_NEXT(pv, pv_list);
3592 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3593 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3596 if (!pmap_track_modified(pv->pv_va))
3599 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3601 if (pte && (*pte & PG_A)) {
3603 atomic_clear_long(pte, PG_A);
3605 atomic_clear_long_nonlocked(pte, PG_A);
3612 } while ((pv = pvn) != NULL && pv != pvf);
3614 lwkt_reltoken(&vm_token);
3623 * Return whether or not the specified physical page was modified
3624 * in any physical maps.
3627 pmap_is_modified(vm_page_t m)
3631 lwkt_gettoken(&vm_token);
3632 res = pmap_testbit(m, PG_M);
3633 lwkt_reltoken(&vm_token);
3638 * Clear the modify bits on the specified physical page.
3641 pmap_clear_modify(vm_page_t m)
3643 lwkt_gettoken(&vm_token);
3644 pmap_clearbit(m, PG_M);
3645 lwkt_reltoken(&vm_token);
3649 * pmap_clear_reference:
3651 * Clear the reference bit on the specified physical page.
3654 pmap_clear_reference(vm_page_t m)
3656 lwkt_gettoken(&vm_token);
3657 pmap_clearbit(m, PG_A);
3658 lwkt_reltoken(&vm_token);
3662 * Miscellaneous support routines follow
3667 i386_protection_init(void)
3671 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3672 kp = protection_codes;
3673 for (prot = 0; prot < 8; prot++) {
3675 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3677 * Read access is also 0. There isn't any execute bit,
3678 * so just make it readable.
3680 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3681 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3682 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3685 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3686 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3687 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3688 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3696 * Map a set of physical memory pages into the kernel virtual
3697 * address space. Return a pointer to where it is mapped. This
3698 * routine is intended to be used for mapping device memory,
3701 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3705 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3707 vm_offset_t va, tmpva, offset;
3710 offset = pa & PAGE_MASK;
3711 size = roundup(offset + size, PAGE_SIZE);
3713 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3715 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3717 pa = pa & ~PAGE_MASK;
3718 for (tmpva = va; size > 0;) {
3719 pte = vtopte(tmpva);
3720 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3728 return ((void *)(va + offset));
3732 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3734 vm_offset_t va, tmpva, offset;
3737 offset = pa & PAGE_MASK;
3738 size = roundup(offset + size, PAGE_SIZE);
3740 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3742 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3744 pa = pa & ~PAGE_MASK;
3745 for (tmpva = va; size > 0;) {
3746 pte = vtopte(tmpva);
3747 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3755 return ((void *)(va + offset));
3759 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3761 vm_offset_t base, offset;
3763 base = va & ~PAGE_MASK;
3764 offset = va & PAGE_MASK;
3765 size = roundup(offset + size, PAGE_SIZE);
3766 pmap_qremove(va, size >> PAGE_SHIFT);
3767 kmem_free(&kernel_map, base, size);
3771 * perform the pmap work for mincore
3774 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3776 pt_entry_t *ptep, pte;
3780 lwkt_gettoken(&vm_token);
3781 ptep = pmap_pte(pmap, addr);
3783 if (ptep && (pte = *ptep) != 0) {
3786 val = MINCORE_INCORE;
3787 if ((pte & PG_MANAGED) == 0)
3790 pa = pte & PG_FRAME;
3792 m = PHYS_TO_VM_PAGE(pa);
3798 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3800 * Modified by someone
3802 else if (m->dirty || pmap_is_modified(m))
3803 val |= MINCORE_MODIFIED_OTHER;
3808 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3811 * Referenced by someone
3813 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3814 val |= MINCORE_REFERENCED_OTHER;
3815 vm_page_flag_set(m, PG_REFERENCED);
3819 lwkt_reltoken(&vm_token);
3824 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3825 * vmspace will be ref'd and the old one will be deref'd.
3827 * The vmspace for all lwps associated with the process will be adjusted
3828 * and cr3 will be reloaded if any lwp is the current lwp.
3831 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3833 struct vmspace *oldvm;
3837 oldvm = p->p_vmspace;
3838 if (oldvm != newvm) {
3839 p->p_vmspace = newvm;
3840 KKASSERT(p->p_nthreads == 1);
3841 lp = RB_ROOT(&p->p_lwp_tree);
3842 pmap_setlwpvm(lp, newvm);
3844 sysref_get(&newvm->vm_sysref);
3845 sysref_put(&oldvm->vm_sysref);
3852 * Set the vmspace for a LWP. The vmspace is almost universally set the
3853 * same as the process vmspace, but virtual kernels need to swap out contexts
3854 * on a per-lwp basis.
3857 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3859 struct vmspace *oldvm;
3863 oldvm = lp->lwp_vmspace;
3865 if (oldvm != newvm) {
3866 lp->lwp_vmspace = newvm;
3867 if (curthread->td_lwp == lp) {
3868 pmap = vmspace_pmap(newvm);
3870 atomic_set_int(&pmap->pm_active, mycpu->gd_cpumask);
3871 if (pmap->pm_active & CPUMASK_LOCK)
3872 pmap_interlock_wait(newvm);
3874 pmap->pm_active |= 1;
3876 #if defined(SWTCH_OPTIM_STATS)
3879 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
3880 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
3881 load_cr3(curthread->td_pcb->pcb_cr3);
3882 pmap = vmspace_pmap(oldvm);
3884 atomic_clear_int(&pmap->pm_active, mycpu->gd_cpumask);
3886 pmap->pm_active &= ~1;
3896 * Called when switching to a locked pmap
3899 pmap_interlock_wait(struct vmspace *vm)
3901 struct pmap *pmap = &vm->vm_pmap;
3903 if (pmap->pm_active & CPUMASK_LOCK) {
3904 while (pmap->pm_active & CPUMASK_LOCK) {
3907 lwkt_process_ipiq();
3915 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3918 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3922 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3927 * Used by kmalloc/kfree, page already exists at va
3930 pmap_kvtom(vm_offset_t va)
3932 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));