2 * Copyright (c) 1991 Regents of the University of California.
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1994 David Greenman
5 * Copyright (c) 2003 Peter Wemm
6 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
7 * Copyright (c) 2008, 2009 The DragonFly Project.
8 * Copyright (c) 2008, 2009 Jordan Gordeev.
11 * This code is derived from software contributed to Berkeley by
12 * the Systems Programming Group of the University of Utah Computer
13 * Science Department and William Jolitz of UUNET Technologies Inc.
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
44 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
48 * Manages physical address maps.
50 * In addition to hardware address maps, this
51 * module is called upon to provide software-use-only
52 * maps which may or may not be stored in the same
53 * form as hardware maps. These pseudo-maps are
54 * used to store intermediate results from copy
55 * operations to and from address spaces.
57 * Since the information managed by this module is
58 * also stored by the logical address mapping module,
59 * this module may throw away valid virtual-to-physical
60 * mappings at almost any time. However, invalidations
61 * of virtual-to-physical mappings must be done as
64 * In order to cope with hardware architectures which
65 * make virtual-to-physical map invalidates expensive,
66 * this module may delay invalidate or reduced protection
67 * operations until such time as they are actually
68 * necessary. This module is given full information as
69 * to which processors are currently using which maps,
70 * and to when physical maps must be made correct.
74 #include "opt_disable_pse.h"
77 #include "opt_msgbuf.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
83 #include <sys/msgbuf.h>
84 #include <sys/vmmeter.h>
88 #include <vm/vm_param.h>
89 #include <sys/sysctl.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_pageout.h>
97 #include <vm/vm_pager.h>
98 #include <vm/vm_zone.h>
100 #include <sys/user.h>
101 #include <sys/thread2.h>
102 #include <sys/sysref2.h>
104 #include <machine/cputypes.h>
105 #include <machine/md_var.h>
106 #include <machine/specialreg.h>
107 #include <machine/smp.h>
108 #include <machine_base/apic/apicreg.h>
109 #include <machine/globaldata.h>
110 #include <machine/pmap.h>
111 #include <machine/pmap_inval.h>
115 #define PMAP_KEEP_PDIRS
116 #ifndef PMAP_SHPGPERPROC
117 #define PMAP_SHPGPERPROC 200
120 #if defined(DIAGNOSTIC)
121 #define PMAP_DIAGNOSTIC
127 * Get PDEs and PTEs for user/kernel address space
129 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
130 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
132 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
133 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
134 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
135 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
136 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
140 * Given a map and a machine independent protection code,
141 * convert to a vax protection code.
143 #define pte_prot(m, p) \
144 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
145 static int protection_codes[8];
147 struct pmap kernel_pmap;
148 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
150 vm_paddr_t avail_start; /* PA of first available physical page */
151 vm_paddr_t avail_end; /* PA of last available physical page */
152 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
153 vm_offset_t virtual2_end;
154 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
155 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
156 vm_offset_t KvaStart; /* VA start of KVA space */
157 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
158 vm_offset_t KvaSize; /* max size of kernel virtual address space */
159 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
160 static int pgeflag; /* PG_G or-in */
161 static int pseflag; /* PG_PS or-in */
163 static vm_object_t kptobj;
166 static vm_paddr_t dmaplimit;
168 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
170 static uint64_t KPTbase;
171 static uint64_t KPTphys;
172 static uint64_t KPDphys; /* phys addr of kernel level 2 */
173 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
174 uint64_t KPDPphys; /* phys addr of kernel level 3 */
175 uint64_t KPML4phys; /* phys addr of kernel level 4 */
177 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
178 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
181 * Data for the pv entry allocation mechanism
183 static vm_zone_t pvzone;
184 static struct vm_zone pvzone_store;
185 static struct vm_object pvzone_obj;
186 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
187 static int pmap_pagedaemon_waken = 0;
188 static struct pv_entry *pvinit;
191 * All those kernel PT submaps that BSD is so fond of
193 pt_entry_t *CMAP1 = 0, *ptmmap;
194 caddr_t CADDR1 = 0, ptvmmap = 0;
195 static pt_entry_t *msgbufmap;
196 struct msgbuf *msgbufp=0;
201 static pt_entry_t *pt_crashdumpmap;
202 static caddr_t crashdumpmap;
204 extern pt_entry_t *SMPpt;
205 extern uint64_t SMPptpa;
209 static pv_entry_t get_pv_entry (void);
210 static void i386_protection_init (void);
211 static void create_pagetables(vm_paddr_t *firstaddr);
212 static void pmap_remove_all (vm_page_t m);
213 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
214 vm_offset_t sva, pmap_inval_info_t info);
215 static void pmap_remove_page (struct pmap *pmap,
216 vm_offset_t va, pmap_inval_info_t info);
217 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
218 vm_offset_t va, pmap_inval_info_t info);
219 static boolean_t pmap_testbit (vm_page_t m, int bit);
220 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
221 vm_page_t mpte, vm_page_t m);
223 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
225 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
226 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
227 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
228 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
229 static int _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
230 pmap_inval_info_t info);
231 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
232 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
234 static unsigned pdir4mb;
237 * Move the kernel virtual free pointer to the next
238 * 2MB. This is used to help improve performance
239 * by using a large (2MB) page for much of the kernel
240 * (.text, .data, .bss)
244 pmap_kmem_choose(vm_offset_t addr)
246 vm_offset_t newaddr = addr;
248 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
255 * Super fast pmap_pte routine best used when scanning the pv lists.
256 * This eliminates many course-grained invltlb calls. Note that many of
257 * the pv list scans are across different pmaps and it is very wasteful
258 * to do an entire invltlb when checking a single mapping.
260 * Should only be called while in a critical section.
262 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
266 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
268 return pmap_pte(pmap, va);
271 /* Return a non-clipped PD index for a given VA */
274 pmap_pde_pindex(vm_offset_t va)
276 return va >> PDRSHIFT;
279 /* Return various clipped indexes for a given VA */
282 pmap_pte_index(vm_offset_t va)
285 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
290 pmap_pde_index(vm_offset_t va)
293 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
298 pmap_pdpe_index(vm_offset_t va)
301 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
306 pmap_pml4e_index(vm_offset_t va)
309 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
312 /* Return a pointer to the PML4 slot that corresponds to a VA */
315 pmap_pml4e(pmap_t pmap, vm_offset_t va)
318 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
321 /* Return a pointer to the PDP slot that corresponds to a VA */
324 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
328 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
329 return (&pdpe[pmap_pdpe_index(va)]);
332 /* Return a pointer to the PDP slot that corresponds to a VA */
335 pmap_pdpe(pmap_t pmap, vm_offset_t va)
339 pml4e = pmap_pml4e(pmap, va);
340 if ((*pml4e & PG_V) == 0)
342 return (pmap_pml4e_to_pdpe(pml4e, va));
345 /* Return a pointer to the PD slot that corresponds to a VA */
348 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
352 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
353 return (&pde[pmap_pde_index(va)]);
356 /* Return a pointer to the PD slot that corresponds to a VA */
359 pmap_pde(pmap_t pmap, vm_offset_t va)
363 pdpe = pmap_pdpe(pmap, va);
364 if (pdpe == NULL || (*pdpe & PG_V) == 0)
366 return (pmap_pdpe_to_pde(pdpe, va));
369 /* Return a pointer to the PT slot that corresponds to a VA */
372 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
376 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
377 return (&pte[pmap_pte_index(va)]);
380 /* Return a pointer to the PT slot that corresponds to a VA */
383 pmap_pte(pmap_t pmap, vm_offset_t va)
387 pde = pmap_pde(pmap, va);
388 if (pde == NULL || (*pde & PG_V) == 0)
390 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
391 return ((pt_entry_t *)pde);
392 return (pmap_pde_to_pte(pde, va));
397 vtopte(vm_offset_t va)
399 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
401 return (PTmap + ((va >> PAGE_SHIFT) & mask));
406 vtopde(vm_offset_t va)
408 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
410 return (PDmap + ((va >> PDRSHIFT) & mask));
414 allocpages(vm_paddr_t *firstaddr, int n)
419 bzero((void *)ret, n * PAGE_SIZE);
420 *firstaddr += n * PAGE_SIZE;
426 create_pagetables(vm_paddr_t *firstaddr)
430 /* we are running (mostly) V=P at this point */
433 KPTbase = allocpages(firstaddr, NKPT);
434 KPTphys = allocpages(firstaddr, NKPT);
435 KPML4phys = allocpages(firstaddr, 1);
436 KPDPphys = allocpages(firstaddr, NKPML4E);
439 * Calculate the page directory base for KERNBASE,
440 * that is where we start populating the page table pages.
441 * Basically this is the end - 2.
443 KPDphys = allocpages(firstaddr, NKPDPE);
444 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
446 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
447 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
449 DMPDPphys = allocpages(firstaddr, NDMPML4E);
450 if ((amd_feature & AMDID_PAGE1GB) == 0)
451 DMPDphys = allocpages(firstaddr, ndmpdp);
452 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
455 * Fill in the underlying page table pages for the area around
456 * KERNBASE. This remaps low physical memory to KERNBASE.
458 * Read-only from zero to physfree
459 * XXX not fully used, underneath 2M pages
461 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
462 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
463 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
467 * Now map the initial kernel page tables. One block of page
468 * tables is placed at the beginning of kernel virtual memory,
469 * and another block is placed at KERNBASE to map the kernel binary,
470 * data, bss, and initial pre-allocations.
472 for (i = 0; i < NKPT; i++) {
473 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
474 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
476 for (i = 0; i < NKPT; i++) {
477 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
478 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
482 * Map from zero to end of allocations using 2M pages as an
483 * optimization. This will bypass some of the KPTBase pages
484 * above in the KERNBASE area.
486 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
487 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
488 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
492 * And connect up the PD to the PDP. The kernel pmap is expected
493 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
495 for (i = 0; i < NKPDPE; i++) {
496 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
497 KPDphys + (i << PAGE_SHIFT);
498 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
502 /* Now set up the direct map space using either 2MB or 1GB pages */
503 /* Preset PG_M and PG_A because demotion expects it */
504 if ((amd_feature & AMDID_PAGE1GB) == 0) {
505 for (i = 0; i < NPDEPG * ndmpdp; i++) {
506 ((pd_entry_t *)DMPDphys)[i] = (vm_paddr_t)i << PDRSHIFT;
507 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
510 /* And the direct map space's PDP */
511 for (i = 0; i < ndmpdp; i++) {
512 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
514 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
517 for (i = 0; i < ndmpdp; i++) {
518 ((pdp_entry_t *)DMPDPphys)[i] =
519 (vm_paddr_t)i << PDPSHIFT;
520 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
525 /* And recursively map PML4 to itself in order to get PTmap */
526 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
527 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
529 /* Connect the Direct Map slot up to the PML4 */
530 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
531 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
533 /* Connect the KVA slot up to the PML4 */
534 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
535 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
539 init_paging(vm_paddr_t *firstaddr)
541 create_pagetables(firstaddr);
545 * Bootstrap the system enough to run with virtual memory.
547 * On the i386 this is called after mapping has already been enabled
548 * and just syncs the pmap module with what has already been done.
549 * [We can't call it easily with mapping off since the kernel is not
550 * mapped with PA == VA, hence we would have to relocate every address
551 * from the linked base (virtual) address "KERNBASE" to the actual
552 * (physical) address starting relative to 0]
555 pmap_bootstrap(vm_paddr_t *firstaddr)
559 struct mdglobaldata *gd;
562 KvaStart = VM_MIN_KERNEL_ADDRESS;
563 KvaEnd = VM_MAX_KERNEL_ADDRESS;
564 KvaSize = KvaEnd - KvaStart;
566 avail_start = *firstaddr;
569 * Create an initial set of page tables to run the kernel in.
571 create_pagetables(firstaddr);
573 virtual2_start = KvaStart;
574 virtual2_end = PTOV_OFFSET;
576 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
577 virtual_start = pmap_kmem_choose(virtual_start);
579 virtual_end = VM_MAX_KERNEL_ADDRESS;
581 /* XXX do %cr0 as well */
582 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
586 * Initialize protection array.
588 i386_protection_init();
591 * The kernel's pmap is statically allocated so we don't have to use
592 * pmap_create, which is unlikely to work correctly at this part of
593 * the boot sequence (XXX and which no longer exists).
595 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
596 kernel_pmap.pm_count = 1;
597 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
598 TAILQ_INIT(&kernel_pmap.pm_pvlist);
602 * Reserve some special page table entries/VA space for temporary
605 #define SYSMAP(c, p, v, n) \
606 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
610 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
616 * CMAP1/CMAP2 are used for zeroing and copying pages.
618 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
623 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
626 * ptvmmap is used for reading arbitrary physical pages via
629 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
632 * msgbufp is used to map the system message buffer.
633 * XXX msgbufmap is not used.
635 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
636 atop(round_page(MSGBUF_SIZE)))
643 * PG_G is terribly broken on SMP because we IPI invltlb's in some
644 * cases rather then invl1pg. Actually, I don't even know why it
645 * works under UP because self-referential page table mappings
650 if (cpu_feature & CPUID_PGE)
655 * Initialize the 4MB page size flag
659 * The 4MB page version of the initial
660 * kernel page mapping.
664 #if !defined(DISABLE_PSE)
665 if (cpu_feature & CPUID_PSE) {
668 * Note that we have enabled PSE mode
671 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
672 ptditmp &= ~(NBPDR - 1);
673 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
678 * Enable the PSE mode. If we are SMP we can't do this
679 * now because the APs will not be able to use it when
682 load_cr4(rcr4() | CR4_PSE);
685 * We can do the mapping here for the single processor
686 * case. We simply ignore the old page table page from
690 * For SMP, we still need 4K pages to bootstrap APs,
691 * PSE will be enabled as soon as all APs are up.
693 PTD[KPTDI] = (pd_entry_t)ptditmp;
699 if (cpu_apic_address == 0)
700 panic("pmap_bootstrap: no local apic!");
704 * We need to finish setting up the globaldata page for the BSP.
705 * locore has already populated the page table for the mdglobaldata
708 pg = MDGLOBALDATA_BASEALLOC_PAGES;
709 gd = &CPU_prvspace[0].mdglobaldata;
710 gd->gd_CMAP1 = &SMPpt[pg + 0];
711 gd->gd_CMAP2 = &SMPpt[pg + 1];
712 gd->gd_CMAP3 = &SMPpt[pg + 2];
713 gd->gd_PMAP1 = &SMPpt[pg + 3];
714 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
715 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
716 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
717 gd->gd_PADDR1 = (pt_entry_t *)CPU_prvspace[0].PPAGE1;
724 * Set 4mb pdir for mp startup
729 if (pseflag && (cpu_feature & CPUID_PSE)) {
730 load_cr4(rcr4() | CR4_PSE);
731 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
739 * Initialize the pmap module.
740 * Called by vm_init, to initialize any structures that the pmap
741 * system needs to map virtual memory.
742 * pmap_init has been enhanced to support in a fairly consistant
743 * way, discontiguous physical memory.
752 * object for kernel page table pages
754 /* JG I think the number can be arbitrary */
755 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
758 * Allocate memory for random pmap data structures. Includes the
762 for(i = 0; i < vm_page_array_size; i++) {
765 m = &vm_page_array[i];
766 TAILQ_INIT(&m->md.pv_list);
767 m->md.pv_list_count = 0;
771 * init the pv free list
773 initial_pvs = vm_page_array_size;
774 if (initial_pvs < MINPV)
776 pvzone = &pvzone_store;
777 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
778 initial_pvs * sizeof (struct pv_entry));
779 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
783 * Now it is safe to enable pv_table recording.
785 pmap_initialized = TRUE;
787 lapic = pmap_mapdev_uncacheable(cpu_apic_address, sizeof(struct LAPIC));
792 * Initialize the address space (zone) for the pv_entries. Set a
793 * high water mark so that the system can recover from excessive
794 * numbers of pv entries.
799 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);
805 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
809 /***************************************************
810 * Low level helper routines.....
811 ***************************************************/
813 #if defined(PMAP_DIAGNOSTIC)
816 * This code checks for non-writeable/modified pages.
817 * This should be an invalid condition.
821 pmap_nw_modified(pt_entry_t pte)
823 if ((pte & (PG_M|PG_RW)) == PG_M)
832 * this routine defines the region(s) of memory that should
833 * not be tested for the modified bit.
837 pmap_track_modified(vm_offset_t va)
839 if ((va < clean_sva) || (va >= clean_eva))
848 * Extract the physical page address associated with the map/VA pair.
850 * This function may not be called from an interrupt if the pmap is
854 pmap_extract(pmap_t pmap, vm_offset_t va)
858 pd_entry_t pde, *pdep;
861 pdep = pmap_pde(pmap, va);
865 if ((pde & PG_PS) != 0) {
866 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
868 pte = pmap_pde_to_pte(pdep, va);
869 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
877 * Routine: pmap_kextract
879 * Extract the physical page address associated
880 * kernel virtual address.
883 pmap_kextract(vm_offset_t va)
888 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
889 pa = DMAP_TO_PHYS(va);
893 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
896 * Beware of a concurrent promotion that changes the
897 * PDE at this point! For example, vtopte() must not
898 * be used to access the PTE because it would use the
899 * new PDE. It is, however, safe to use the old PDE
900 * because the page table page is preserved by the
903 pa = *pmap_pde_to_pte(&pde, va);
904 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
910 /***************************************************
911 * Low level mapping routines.....
912 ***************************************************/
915 * Routine: pmap_kenter
917 * Add a wired page to the KVA
918 * NOTE! note that in order for the mapping to take effect -- you
919 * should do an invltlb after doing the pmap_kenter().
922 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
926 pmap_inval_info info;
928 pmap_inval_init(&info);
929 npte = pa | PG_RW | PG_V | pgeflag;
931 pmap_inval_add(&info, &kernel_pmap, va);
933 pmap_inval_flush(&info);
937 * Routine: pmap_kenter_quick
939 * Similar to pmap_kenter(), except we only invalidate the
940 * mapping on the current CPU.
943 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
948 npte = pa | PG_RW | PG_V | pgeflag;
951 cpu_invlpg((void *)va);
955 pmap_kenter_sync(vm_offset_t va)
957 pmap_inval_info info;
959 pmap_inval_init(&info);
960 pmap_inval_add(&info, &kernel_pmap, va);
961 pmap_inval_flush(&info);
965 pmap_kenter_sync_quick(vm_offset_t va)
967 cpu_invlpg((void *)va);
971 * remove a page from the kernel pagetables
974 pmap_kremove(vm_offset_t va)
977 pmap_inval_info info;
979 pmap_inval_init(&info);
981 pmap_inval_add(&info, &kernel_pmap, va);
983 pmap_inval_flush(&info);
987 pmap_kremove_quick(vm_offset_t va)
992 cpu_invlpg((void *)va);
996 * XXX these need to be recoded. They are not used in any critical path.
999 pmap_kmodify_rw(vm_offset_t va)
1001 *vtopte(va) |= PG_RW;
1002 cpu_invlpg((void *)va);
1006 pmap_kmodify_nc(vm_offset_t va)
1008 *vtopte(va) |= PG_N;
1009 cpu_invlpg((void *)va);
1013 * Used to map a range of physical addresses into kernel
1014 * virtual address space.
1016 * For now, VM is already on, we only need to map the
1020 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1022 return PHYS_TO_DMAP(start);
1027 * Add a list of wired pages to the kva
1028 * this routine is only used for temporary
1029 * kernel mappings that do not need to have
1030 * page modification or references recorded.
1031 * Note that old mappings are simply written
1032 * over. The page *must* be wired.
1035 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1039 end_va = va + count * PAGE_SIZE;
1041 while (va < end_va) {
1045 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1046 cpu_invlpg((void *)va);
1051 smp_invltlb(); /* XXX */
1056 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
1059 cpumask_t cmask = mycpu->gd_cpumask;
1061 end_va = va + count * PAGE_SIZE;
1063 while (va < end_va) {
1068 * Install the new PTE. If the pte changed from the prior
1069 * mapping we must reset the cpu mask and invalidate the page.
1070 * If the pte is the same but we have not seen it on the
1071 * current cpu, invlpg the existing mapping. Otherwise the
1072 * entry is optimal and no invalidation is required.
1075 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
1076 if (*pte != pteval) {
1079 cpu_invlpg((void *)va);
1080 } else if ((*mask & cmask) == 0) {
1081 cpu_invlpg((void *)va);
1090 * This routine jerks page mappings from the
1091 * kernel -- it is meant only for temporary mappings.
1093 * MPSAFE, INTERRUPT SAFE (cluster callback)
1096 pmap_qremove(vm_offset_t va, int count)
1100 end_va = va + count * PAGE_SIZE;
1102 while (va < end_va) {
1107 cpu_invlpg((void *)va);
1116 * This routine works like vm_page_lookup() but also blocks as long as the
1117 * page is busy. This routine does not busy the page it returns.
1119 * Unless the caller is managing objects whos pages are in a known state,
1120 * the call should be made with a critical section held so the page's object
1121 * association remains valid on return.
1125 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1130 m = vm_page_lookup(object, pindex);
1131 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
1137 * Create a new thread and optionally associate it with a (new) process.
1138 * NOTE! the new thread's cpu may not equal the current cpu.
1141 pmap_init_thread(thread_t td)
1143 /* enforce pcb placement */
1144 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1145 td->td_savefpu = &td->td_pcb->pcb_save;
1146 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on x86_64? */
1150 * This routine directly affects the fork perf for a process.
1153 pmap_init_proc(struct proc *p)
1158 * Dispose the UPAGES for a process that has exited.
1159 * This routine directly impacts the exit perf of a process.
1162 pmap_dispose_proc(struct proc *p)
1164 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1167 /***************************************************
1168 * Page table page management routines.....
1169 ***************************************************/
1172 * This routine unholds page table pages, and if the hold count
1173 * drops to zero, then it decrements the wire count.
1177 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1178 pmap_inval_info_t info)
1180 KKASSERT(m->hold_count > 0);
1181 if (m->hold_count > 1) {
1185 return _pmap_unwire_pte_hold(pmap, va, m, info);
1191 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1192 pmap_inval_info_t info)
1195 * Wait until we can busy the page ourselves. We cannot have
1196 * any active flushes if we block. We own one hold count on the
1197 * page so it cannot be freed out from under us.
1199 if (m->flags & PG_BUSY) {
1200 pmap_inval_flush(info);
1201 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1204 KASSERT(m->queue == PQ_NONE,
1205 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1208 * This case can occur if new references were acquired while
1211 if (m->hold_count > 1) {
1212 KKASSERT(m->hold_count > 1);
1218 * Unmap the page table page
1220 KKASSERT(m->hold_count == 1);
1222 pmap_inval_add(info, pmap, -1);
1224 if (m->pindex >= (NUPDE + NUPDPE)) {
1227 pml4 = pmap_pml4e(pmap, va);
1229 } else if (m->pindex >= NUPDE) {
1232 pdp = pmap_pdpe(pmap, va);
1237 pd = pmap_pde(pmap, va);
1241 KKASSERT(pmap->pm_stats.resident_count > 0);
1242 --pmap->pm_stats.resident_count;
1244 if (pmap->pm_ptphint == m)
1245 pmap->pm_ptphint = NULL;
1247 if (m->pindex < NUPDE) {
1248 /* We just released a PT, unhold the matching PD */
1251 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
1252 pmap_unwire_pte_hold(pmap, va, pdpg, info);
1254 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
1255 /* We just released a PD, unhold the matching PDP */
1258 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
1259 pmap_unwire_pte_hold(pmap, va, pdppg, info);
1263 * This was our last hold, the page had better be unwired
1264 * after we decrement wire_count.
1266 * FUTURE NOTE: shared page directory page could result in
1267 * multiple wire counts.
1271 KKASSERT(m->wire_count == 0);
1272 --vmstats.v_wire_count;
1273 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1275 vm_page_free_zero(m);
1281 * After removing a page table entry, this routine is used to
1282 * conditionally free the page, and manage the hold/wire counts.
1286 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1287 pmap_inval_info_t info)
1289 vm_pindex_t ptepindex;
1291 if (va >= VM_MAX_USER_ADDRESS)
1295 ptepindex = pmap_pde_pindex(va);
1297 if (pmap->pm_ptphint &&
1298 (pmap->pm_ptphint->pindex == ptepindex)) {
1299 mpte = pmap->pm_ptphint;
1302 pmap_inval_flush(info);
1303 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1304 pmap->pm_ptphint = mpte;
1309 return pmap_unwire_pte_hold(pmap, va, mpte, info);
1313 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1314 * it, and IdlePTD, represents the template used to update all other pmaps.
1316 * On architectures where the kernel pmap is not integrated into the user
1317 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1318 * kernel_pmap should be used to directly access the kernel_pmap.
1321 pmap_pinit0(struct pmap *pmap)
1323 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1325 pmap->pm_active = 0;
1326 pmap->pm_ptphint = NULL;
1327 TAILQ_INIT(&pmap->pm_pvlist);
1328 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1332 * Initialize a preallocated and zeroed pmap structure,
1333 * such as one in a vmspace structure.
1336 pmap_pinit(struct pmap *pmap)
1341 * No need to allocate page table space yet but we do need a valid
1342 * page directory table.
1344 if (pmap->pm_pml4 == NULL) {
1346 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1350 * Allocate an object for the ptes
1352 if (pmap->pm_pteobj == NULL)
1353 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1);
1356 * Allocate the page directory page, unless we already have
1357 * one cached. If we used the cached page the wire_count will
1358 * already be set appropriately.
1360 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1361 ptdpg = vm_page_grab(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I,
1362 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1363 pmap->pm_pdirm = ptdpg;
1364 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1365 ptdpg->valid = VM_PAGE_BITS_ALL;
1366 if (ptdpg->wire_count == 0)
1367 ++vmstats.v_wire_count;
1368 ptdpg->wire_count = 1;
1369 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1371 if ((ptdpg->flags & PG_ZERO) == 0)
1372 bzero(pmap->pm_pml4, PAGE_SIZE);
1374 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1375 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1377 /* install self-referential address mapping entry */
1378 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1381 pmap->pm_active = 0;
1382 pmap->pm_ptphint = NULL;
1383 TAILQ_INIT(&pmap->pm_pvlist);
1384 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1385 pmap->pm_stats.resident_count = 1;
1389 * Clean up a pmap structure so it can be physically freed. This routine
1390 * is called by the vmspace dtor function. A great deal of pmap data is
1391 * left passively mapped to improve vmspace management so we have a bit
1392 * of cleanup work to do here.
1395 pmap_puninit(pmap_t pmap)
1399 KKASSERT(pmap->pm_active == 0);
1400 if ((p = pmap->pm_pdirm) != NULL) {
1401 KKASSERT(pmap->pm_pml4 != NULL);
1402 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1403 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1405 vmstats.v_wire_count--;
1406 KKASSERT((p->flags & PG_BUSY) == 0);
1408 vm_page_free_zero(p);
1409 pmap->pm_pdirm = NULL;
1411 if (pmap->pm_pml4) {
1412 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1413 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1414 pmap->pm_pml4 = NULL;
1416 if (pmap->pm_pteobj) {
1417 vm_object_deallocate(pmap->pm_pteobj);
1418 pmap->pm_pteobj = NULL;
1423 * Wire in kernel global address entries. To avoid a race condition
1424 * between pmap initialization and pmap_growkernel, this procedure
1425 * adds the pmap to the master list (which growkernel scans to update),
1426 * then copies the template.
1429 pmap_pinit2(struct pmap *pmap)
1432 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1433 /* XXX copies current process, does not fill in MPPTDI */
1438 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1439 * 0 on failure (if the procedure had to sleep).
1441 * When asked to remove the page directory page itself, we actually just
1442 * leave it cached so we do not have to incur the SMP inval overhead of
1443 * removing the kernel mapping. pmap_puninit() will take care of it.
1447 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1450 * This code optimizes the case of freeing non-busy
1451 * page-table pages. Those pages are zero now, and
1452 * might as well be placed directly into the zero queue.
1454 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1460 * Remove the page table page from the processes address space.
1462 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1464 * We are the pml4 table itself.
1466 /* XXX anything to do here? */
1467 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1469 * Remove a PDP page from the PML4. We do not maintain
1470 * hold counts on the PML4 page.
1476 m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1477 KKASSERT(m4 != NULL);
1478 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1479 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1480 KKASSERT(pml4[idx] != 0);
1482 } else if (p->pindex >= NUPDE) {
1484 * Remove a PD page from the PDP and drop the hold count
1485 * on the PDP. The PDP is left cached in the pmap if
1486 * the hold count drops to 0 so the wire count remains
1493 m3 = vm_page_lookup(pmap->pm_pteobj,
1494 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1495 KKASSERT(m3 != NULL);
1496 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1497 idx = (p->pindex - NUPDE) % NPDPEPG;
1498 KKASSERT(pdp[idx] != 0);
1503 * Remove a PT page from the PD and drop the hold count
1504 * on the PD. The PD is left cached in the pmap if
1505 * the hold count drops to 0 so the wire count remains
1512 m2 = vm_page_lookup(pmap->pm_pteobj,
1513 NUPDE + p->pindex / NPDEPG);
1514 KKASSERT(m2 != NULL);
1515 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1516 idx = p->pindex % NPDEPG;
1522 * One fewer mappings in the pmap. p's hold count had better
1525 KKASSERT(pmap->pm_stats.resident_count > 0);
1526 --pmap->pm_stats.resident_count;
1528 panic("pmap_release: freeing held page table page");
1529 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1530 pmap->pm_ptphint = NULL;
1533 * We leave the top-level page table page cached, wired, and mapped in
1534 * the pmap until the dtor function (pmap_puninit()) gets called.
1535 * However, still clean it up so we can set PG_ZERO.
1537 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1538 bzero(pmap->pm_pml4, PAGE_SIZE);
1539 vm_page_flag_set(p, PG_ZERO);
1543 KKASSERT(p->wire_count == 0);
1544 vmstats.v_wire_count--;
1545 /* JG eventually revert to using vm_page_free_zero() */
1552 * This routine is called when various levels in the page table need to
1553 * be populated. This routine cannot fail.
1557 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1562 * Find or fabricate a new pagetable page. This will busy the page.
1564 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1565 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1566 if ((m->flags & PG_ZERO) == 0) {
1567 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1570 KASSERT(m->queue == PQ_NONE,
1571 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1574 * Increment the hold count for the page we will be returning to
1578 if (m->wire_count++ == 0)
1579 vmstats.v_wire_count++;
1582 * Map the pagetable page into the process address space, if
1583 * it isn't already there.
1585 * It is possible that someone else got in and mapped the page
1586 * directory page while we were blocked, if so just unbusy and
1587 * return the held page.
1589 if (ptepindex >= (NUPDE + NUPDPE)) {
1591 * Wire up a new PDP page in the PML4
1593 vm_pindex_t pml4index;
1596 pml4index = ptepindex - (NUPDE + NUPDPE);
1597 pml4 = &pmap->pm_pml4[pml4index];
1599 if (--m->wire_count == 0)
1600 --vmstats.v_wire_count;
1604 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1605 } else if (ptepindex >= NUPDE) {
1607 * Wire up a new PD page in the PDP
1609 vm_pindex_t pml4index;
1610 vm_pindex_t pdpindex;
1615 pdpindex = ptepindex - NUPDE;
1616 pml4index = pdpindex >> NPML4EPGSHIFT;
1618 pml4 = &pmap->pm_pml4[pml4index];
1619 if ((*pml4 & PG_V) == 0) {
1621 * Have to allocate a new PDP page, recurse.
1622 * This always succeeds. Returned page will
1625 pdppg = _pmap_allocpte(pmap,
1626 NUPDE + NUPDPE + pml4index);
1629 * Add a held reference to the PDP page.
1631 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1632 pdppg->hold_count++;
1636 * Now find the pdp_entry and map the PDP. If the PDP
1637 * has already been mapped unwind and return the
1638 * already-mapped PDP held.
1640 * pdppg is left held (hold_count is incremented for
1641 * each PD in the PDP).
1643 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1644 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1646 vm_page_unhold(pdppg);
1647 if (--m->wire_count == 0)
1648 --vmstats.v_wire_count;
1652 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1655 * Wire up the new PT page in the PD
1657 vm_pindex_t pml4index;
1658 vm_pindex_t pdpindex;
1664 pdpindex = ptepindex >> NPDPEPGSHIFT;
1665 pml4index = pdpindex >> NPML4EPGSHIFT;
1668 * Locate the PDP page in the PML4, then the PD page in
1669 * the PDP. If either does not exist we simply recurse
1672 * We can just recurse on the PD page as it will recurse
1673 * on the PDP if necessary.
1675 pml4 = &pmap->pm_pml4[pml4index];
1676 if ((*pml4 & PG_V) == 0) {
1677 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1678 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1679 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1681 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1682 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1683 if ((*pdp & PG_V) == 0) {
1684 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1686 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1692 * Now fill in the pte in the PD. If the pte already exists
1693 * (again, if we raced the grab), unhold pdpg and unwire
1694 * m, returning a held m.
1696 * pdpg is left held (hold_count is incremented for
1697 * each PT in the PD).
1699 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
1700 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1702 vm_page_unhold(pdpg);
1703 if (--m->wire_count == 0)
1704 --vmstats.v_wire_count;
1708 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1712 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1713 * valid bits, mapped flag, unbusy, and we're done.
1715 pmap->pm_ptphint = m;
1716 ++pmap->pm_stats.resident_count;
1718 m->valid = VM_PAGE_BITS_ALL;
1719 vm_page_flag_clear(m, PG_ZERO);
1720 vm_page_flag_set(m, PG_MAPPED);
1728 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1730 vm_pindex_t ptepindex;
1735 * Calculate pagetable page index
1737 ptepindex = pmap_pde_pindex(va);
1740 * Get the page directory entry
1742 pd = pmap_pde(pmap, va);
1745 * This supports switching from a 2MB page to a
1748 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1749 panic("no promotion/demotion yet");
1757 * If the page table page is mapped, we just increment the
1758 * hold count, and activate it.
1760 if (pd != NULL && (*pd & PG_V) != 0) {
1761 /* YYY hint is used here on i386 */
1762 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1763 pmap->pm_ptphint = m;
1768 * Here if the pte page isn't mapped, or if it has been deallocated.
1770 return _pmap_allocpte(pmap, ptepindex);
1774 /***************************************************
1775 * Pmap allocation/deallocation routines.
1776 ***************************************************/
1779 * Release any resources held by the given physical map.
1780 * Called when a pmap initialized by pmap_pinit is being released.
1781 * Should only be called if the map contains no valid mappings.
1783 static int pmap_release_callback(struct vm_page *p, void *data);
1786 pmap_release(struct pmap *pmap)
1788 vm_object_t object = pmap->pm_pteobj;
1789 struct rb_vm_page_scan_info info;
1791 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1792 #if defined(DIAGNOSTIC)
1793 if (object->ref_count != 1)
1794 panic("pmap_release: pteobj reference count != 1");
1798 info.object = object;
1800 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1807 info.limit = object->generation;
1809 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1810 pmap_release_callback, &info);
1811 if (info.error == 0 && info.mpte) {
1812 if (!pmap_release_free_page(pmap, info.mpte))
1816 } while (info.error);
1821 pmap_release_callback(struct vm_page *p, void *data)
1823 struct rb_vm_page_scan_info *info = data;
1825 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1829 if (!pmap_release_free_page(info->pmap, p)) {
1833 if (info->object->generation != info->limit) {
1841 * Grow the number of kernel page table entries, if needed.
1844 pmap_growkernel(vm_offset_t addr)
1847 vm_offset_t ptppaddr;
1849 pd_entry_t *pde, newpdir;
1853 if (kernel_vm_end == 0) {
1854 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1856 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1857 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1859 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1860 kernel_vm_end = kernel_map.max_offset;
1865 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1866 if (addr - 1 >= kernel_map.max_offset)
1867 addr = kernel_map.max_offset;
1868 while (kernel_vm_end < addr) {
1869 pde = pmap_pde(&kernel_pmap, kernel_vm_end);
1871 /* We need a new PDP entry */
1872 nkpg = vm_page_alloc(kptobj, nkpt,
1873 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM
1874 | VM_ALLOC_INTERRUPT);
1876 panic("pmap_growkernel: no memory to grow kernel");
1877 paddr = VM_PAGE_TO_PHYS(nkpg);
1878 if ((nkpg->flags & PG_ZERO) == 0)
1879 pmap_zero_page(paddr);
1880 vm_page_flag_clear(nkpg, PG_ZERO);
1881 newpdp = (pdp_entry_t)
1882 (paddr | PG_V | PG_RW | PG_A | PG_M);
1883 *pmap_pdpe(&kernel_pmap, kernel_vm_end) = newpdp;
1885 continue; /* try again */
1887 if ((*pde & PG_V) != 0) {
1888 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1889 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1890 kernel_vm_end = kernel_map.max_offset;
1897 * This index is bogus, but out of the way
1899 nkpg = vm_page_alloc(kptobj, nkpt,
1900 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1902 panic("pmap_growkernel: no memory to grow kernel");
1905 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1906 pmap_zero_page(ptppaddr);
1907 vm_page_flag_clear(nkpg, PG_ZERO);
1908 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1909 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1912 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1913 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1914 kernel_vm_end = kernel_map.max_offset;
1922 * Retire the given physical map from service.
1923 * Should only be called if the map contains
1924 * no valid mappings.
1927 pmap_destroy(pmap_t pmap)
1934 count = --pmap->pm_count;
1937 panic("destroying a pmap is not yet implemented");
1942 * Add a reference to the specified pmap.
1945 pmap_reference(pmap_t pmap)
1952 /***************************************************
1953 * page management routines.
1954 ***************************************************/
1957 * free the pv_entry back to the free list. This function may be
1958 * called from an interrupt.
1962 free_pv_entry(pv_entry_t pv)
1965 KKASSERT(pv_entry_count >= 0);
1970 * get a new pv_entry, allocating a block from the system
1971 * when needed. This function may be called from an interrupt.
1978 if (pv_entry_high_water &&
1979 (pv_entry_count > pv_entry_high_water) &&
1980 (pmap_pagedaemon_waken == 0)) {
1981 pmap_pagedaemon_waken = 1;
1982 wakeup(&vm_pages_needed);
1984 return zalloc(pvzone);
1988 * This routine is very drastic, but can save the system
1996 static int warningdone=0;
1998 if (pmap_pagedaemon_waken == 0)
2001 if (warningdone < 5) {
2002 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
2006 for(i = 0; i < vm_page_array_size; i++) {
2007 m = &vm_page_array[i];
2008 if (m->wire_count || m->hold_count || m->busy ||
2009 (m->flags & PG_BUSY))
2013 pmap_pagedaemon_waken = 0;
2018 * If it is the first entry on the list, it is actually
2019 * in the header and we must copy the following entry up
2020 * to the header. Otherwise we must search the list for
2021 * the entry. In either case we free the now unused entry.
2025 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
2026 vm_offset_t va, pmap_inval_info_t info)
2032 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
2033 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2034 if (pmap == pv->pv_pmap && va == pv->pv_va)
2038 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2039 if (va == pv->pv_va)
2047 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2048 m->md.pv_list_count--;
2049 KKASSERT(m->md.pv_list_count >= 0);
2050 if (TAILQ_EMPTY(&m->md.pv_list))
2051 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2052 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2053 ++pmap->pm_generation;
2054 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
2062 * Create a pv entry for page at pa for
2067 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2072 pv = get_pv_entry();
2077 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2078 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2079 ++pmap->pm_generation;
2080 m->md.pv_list_count++;
2086 * pmap_remove_pte: do the things to unmap a page in a process
2090 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2091 pmap_inval_info_t info)
2096 pmap_inval_add(info, pmap, va);
2097 oldpte = pte_load_clear(ptq);
2099 pmap->pm_stats.wired_count -= 1;
2101 * Machines that don't support invlpg, also don't support
2102 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2106 cpu_invlpg((void *)va);
2107 KKASSERT(pmap->pm_stats.resident_count > 0);
2108 --pmap->pm_stats.resident_count;
2109 if (oldpte & PG_MANAGED) {
2110 m = PHYS_TO_VM_PAGE(oldpte);
2111 if (oldpte & PG_M) {
2112 #if defined(PMAP_DIAGNOSTIC)
2113 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2115 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2119 if (pmap_track_modified(va))
2123 vm_page_flag_set(m, PG_REFERENCED);
2124 return pmap_remove_entry(pmap, m, va, info);
2126 return pmap_unuse_pt(pmap, va, NULL, info);
2135 * Remove a single page from a process address space.
2137 * This function may not be called from an interrupt if the pmap is
2142 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2146 pte = pmap_pte(pmap, va);
2149 if ((*pte & PG_V) == 0)
2151 pmap_remove_pte(pmap, pte, va, info);
2157 * Remove the given range of addresses from the specified map.
2159 * It is assumed that the start and end are properly
2160 * rounded to the page size.
2162 * This function may not be called from an interrupt if the pmap is
2166 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2168 vm_offset_t va_next;
2169 pml4_entry_t *pml4e;
2171 pd_entry_t ptpaddr, *pde;
2173 struct pmap_inval_info info;
2178 if (pmap->pm_stats.resident_count == 0)
2181 pmap_inval_init(&info);
2184 * special handling of removing one page. a very
2185 * common operation and easy to short circuit some
2188 if (sva + PAGE_SIZE == eva) {
2189 pde = pmap_pde(pmap, sva);
2190 if (pde && (*pde & PG_PS) == 0) {
2191 pmap_remove_page(pmap, sva, &info);
2192 pmap_inval_flush(&info);
2197 for (; sva < eva; sva = va_next) {
2198 pml4e = pmap_pml4e(pmap, sva);
2199 if ((*pml4e & PG_V) == 0) {
2200 va_next = (sva + NBPML4) & ~PML4MASK;
2206 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2207 if ((*pdpe & PG_V) == 0) {
2208 va_next = (sva + NBPDP) & ~PDPMASK;
2215 * Calculate index for next page table.
2217 va_next = (sva + NBPDR) & ~PDRMASK;
2221 pde = pmap_pdpe_to_pde(pdpe, sva);
2225 * Weed out invalid mappings.
2231 * Check for large page.
2233 if ((ptpaddr & PG_PS) != 0) {
2234 /* JG FreeBSD has more complex treatment here */
2235 pmap_inval_add(&info, pmap, -1);
2237 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2242 * Limit our scan to either the end of the va represented
2243 * by the current page table page, or to the end of the
2244 * range being removed.
2250 * NOTE: pmap_remove_pte() can block.
2252 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2256 if (pmap_remove_pte(pmap, pte, sva, &info))
2260 pmap_inval_flush(&info);
2266 * Removes this physical page from all physical maps in which it resides.
2267 * Reflects back modify bits to the pager.
2269 * This routine may not be called from an interrupt.
2274 pmap_remove_all(vm_page_t m)
2276 struct pmap_inval_info info;
2277 pt_entry_t *pte, tpte;
2280 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2283 pmap_inval_init(&info);
2285 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2286 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2287 --pv->pv_pmap->pm_stats.resident_count;
2289 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2290 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2291 tpte = pte_load_clear(pte);
2294 pv->pv_pmap->pm_stats.wired_count--;
2297 vm_page_flag_set(m, PG_REFERENCED);
2300 * Update the vm_page_t clean and reference bits.
2303 #if defined(PMAP_DIAGNOSTIC)
2304 if (pmap_nw_modified(tpte)) {
2306 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2310 if (pmap_track_modified(pv->pv_va))
2313 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2314 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2315 ++pv->pv_pmap->pm_generation;
2316 m->md.pv_list_count--;
2317 KKASSERT(m->md.pv_list_count >= 0);
2318 if (TAILQ_EMPTY(&m->md.pv_list))
2319 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2320 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2324 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2325 pmap_inval_flush(&info);
2331 * Set the physical protection on the specified range of this map
2334 * This function may not be called from an interrupt if the map is
2335 * not the kernel_pmap.
2338 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2340 vm_offset_t va_next;
2341 pml4_entry_t *pml4e;
2343 pd_entry_t ptpaddr, *pde;
2345 pmap_inval_info info;
2347 /* JG review for NX */
2352 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2353 pmap_remove(pmap, sva, eva);
2357 if (prot & VM_PROT_WRITE)
2360 pmap_inval_init(&info);
2362 for (; sva < eva; sva = va_next) {
2364 pml4e = pmap_pml4e(pmap, sva);
2365 if ((*pml4e & PG_V) == 0) {
2366 va_next = (sva + NBPML4) & ~PML4MASK;
2372 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2373 if ((*pdpe & PG_V) == 0) {
2374 va_next = (sva + NBPDP) & ~PDPMASK;
2380 va_next = (sva + NBPDR) & ~PDRMASK;
2384 pde = pmap_pdpe_to_pde(pdpe, sva);
2388 * Check for large page.
2390 if ((ptpaddr & PG_PS) != 0) {
2391 pmap_inval_add(&info, pmap, -1);
2392 *pde &= ~(PG_M|PG_RW);
2393 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2398 * Weed out invalid mappings. Note: we assume that the page
2399 * directory table is always allocated, and in kernel virtual.
2407 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2409 pt_entry_t obits, pbits;
2413 * XXX non-optimal. Note also that there can be
2414 * no pmap_inval_flush() calls until after we modify
2415 * ptbase[sindex] (or otherwise we have to do another
2416 * pmap_inval_add() call).
2418 pmap_inval_add(&info, pmap, sva);
2419 obits = pbits = *pte;
2420 if ((pbits & PG_V) == 0)
2422 if (pbits & PG_MANAGED) {
2425 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2426 vm_page_flag_set(m, PG_REFERENCED);
2430 if (pmap_track_modified(sva)) {
2432 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2441 if (pbits != obits) {
2446 pmap_inval_flush(&info);
2450 * Insert the given physical page (p) at
2451 * the specified virtual address (v) in the
2452 * target physical map with the protection requested.
2454 * If specified, the page will be wired down, meaning
2455 * that the related pte can not be reclaimed.
2457 * NB: This is the only routine which MAY NOT lazy-evaluate
2458 * or lose information. That is, this routine must actually
2459 * insert this page into the given map NOW.
2462 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2469 pt_entry_t origpte, newpte;
2471 pmap_inval_info info;
2476 va = trunc_page(va);
2477 #ifdef PMAP_DIAGNOSTIC
2479 panic("pmap_enter: toobig");
2480 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2481 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
2483 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2484 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2486 db_print_backtrace();
2489 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2490 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2492 db_print_backtrace();
2497 * In the case that a page table page is not
2498 * resident, we are creating it here.
2500 if (va < VM_MAX_USER_ADDRESS)
2501 mpte = pmap_allocpte(pmap, va);
2505 pmap_inval_init(&info);
2506 pde = pmap_pde(pmap, va);
2507 if (pde != NULL && (*pde & PG_V) != 0) {
2508 if ((*pde & PG_PS) != 0)
2509 panic("pmap_enter: attempted pmap_enter on 2MB page");
2510 pte = pmap_pde_to_pte(pde, va);
2512 panic("pmap_enter: invalid page directory va=%#lx", va);
2514 KKASSERT(pte != NULL);
2515 pa = VM_PAGE_TO_PHYS(m);
2517 opa = origpte & PG_FRAME;
2520 * Mapping has not changed, must be protection or wiring change.
2522 if (origpte && (opa == pa)) {
2524 * Wiring change, just update stats. We don't worry about
2525 * wiring PT pages as they remain resident as long as there
2526 * are valid mappings in them. Hence, if a user page is wired,
2527 * the PT page will be also.
2529 if (wired && ((origpte & PG_W) == 0))
2530 pmap->pm_stats.wired_count++;
2531 else if (!wired && (origpte & PG_W))
2532 pmap->pm_stats.wired_count--;
2534 #if defined(PMAP_DIAGNOSTIC)
2535 if (pmap_nw_modified(origpte)) {
2537 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2543 * Remove the extra pte reference. Note that we cannot
2544 * optimize the RO->RW case because we have adjusted the
2545 * wiring count above and may need to adjust the wiring
2552 * We might be turning off write access to the page,
2553 * so we go ahead and sense modify status.
2555 if (origpte & PG_MANAGED) {
2556 if ((origpte & PG_M) && pmap_track_modified(va)) {
2558 om = PHYS_TO_VM_PAGE(opa);
2562 KKASSERT(m->flags & PG_MAPPED);
2567 * Mapping has changed, invalidate old range and fall through to
2568 * handle validating new mapping.
2572 err = pmap_remove_pte(pmap, pte, va, &info);
2574 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2576 opa = origpte & PG_FRAME;
2578 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2584 * Enter on the PV list if part of our managed memory. Note that we
2585 * raise IPL while manipulating pv_table since pmap_enter can be
2586 * called at interrupt time.
2588 if (pmap_initialized &&
2589 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2590 pmap_insert_entry(pmap, va, mpte, m);
2592 vm_page_flag_set(m, PG_MAPPED);
2596 * Increment counters
2598 ++pmap->pm_stats.resident_count;
2600 pmap->pm_stats.wired_count++;
2604 * Now validate mapping with desired protection/wiring.
2606 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
2610 if (va < VM_MAX_USER_ADDRESS)
2612 if (pmap == &kernel_pmap)
2616 * if the mapping or permission bits are different, we need
2617 * to update the pte.
2619 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2620 pmap_inval_add(&info, pmap, va);
2621 *pte = newpte | PG_A;
2623 vm_page_flag_set(m, PG_WRITEABLE);
2625 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2626 pmap_inval_flush(&info);
2630 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2631 * This code also assumes that the pmap has no pre-existing entry for this
2634 * This code currently may only be used on user pmaps, not kernel_pmap.
2637 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2642 vm_pindex_t ptepindex;
2644 pmap_inval_info info;
2646 pmap_inval_init(&info);
2648 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2649 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2651 db_print_backtrace();
2654 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2655 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2657 db_print_backtrace();
2661 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2664 * Calculate the page table page (mpte), allocating it if necessary.
2666 * A held page table page (mpte), or NULL, is passed onto the
2667 * section following.
2669 if (va < VM_MAX_USER_ADDRESS) {
2671 * Calculate pagetable page index
2673 ptepindex = pmap_pde_pindex(va);
2677 * Get the page directory entry
2679 ptepa = pmap_pde(pmap, va);
2682 * If the page table page is mapped, we just increment
2683 * the hold count, and activate it.
2685 if (ptepa && (*ptepa & PG_V) != 0) {
2687 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2688 // if (pmap->pm_ptphint &&
2689 // (pmap->pm_ptphint->pindex == ptepindex)) {
2690 // mpte = pmap->pm_ptphint;
2692 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2693 pmap->pm_ptphint = mpte;
2698 mpte = _pmap_allocpte(pmap, ptepindex);
2700 } while (mpte == NULL);
2703 /* this code path is not yet used */
2707 * With a valid (and held) page directory page, we can just use
2708 * vtopte() to get to the pte. If the pte is already present
2709 * we do not disturb it.
2714 pmap_unwire_pte_hold(pmap, va, mpte, &info);
2715 pa = VM_PAGE_TO_PHYS(m);
2716 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2721 * Enter on the PV list if part of our managed memory
2723 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2724 pmap_insert_entry(pmap, va, mpte, m);
2725 vm_page_flag_set(m, PG_MAPPED);
2729 * Increment counters
2731 ++pmap->pm_stats.resident_count;
2733 pa = VM_PAGE_TO_PHYS(m);
2736 * Now validate mapping with RO protection
2738 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2739 *pte = pa | PG_V | PG_U;
2741 *pte = pa | PG_V | PG_U | PG_MANAGED;
2742 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2743 pmap_inval_flush(&info);
2747 * Make a temporary mapping for a physical address. This is only intended
2748 * to be used for panic dumps.
2750 /* JG Needed on x86_64? */
2752 pmap_kenter_temporary(vm_paddr_t pa, int i)
2754 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2755 return ((void *)crashdumpmap);
2758 #define MAX_INIT_PT (96)
2761 * This routine preloads the ptes for a given object into the specified pmap.
2762 * This eliminates the blast of soft faults on process startup and
2763 * immediately after an mmap.
2765 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2768 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2769 vm_object_t object, vm_pindex_t pindex,
2770 vm_size_t size, int limit)
2772 struct rb_vm_page_scan_info info;
2777 * We can't preinit if read access isn't set or there is no pmap
2780 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2784 * We can't preinit if the pmap is not the current pmap
2786 lp = curthread->td_lwp;
2787 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2790 psize = x86_64_btop(size);
2792 if ((object->type != OBJT_VNODE) ||
2793 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2794 (object->resident_page_count > MAX_INIT_PT))) {
2798 if (psize + pindex > object->size) {
2799 if (object->size < pindex)
2801 psize = object->size - pindex;
2808 * Use a red-black scan to traverse the requested range and load
2809 * any valid pages found into the pmap.
2811 * We cannot safely scan the object's memq unless we are in a
2812 * critical section since interrupts can remove pages from objects.
2814 info.start_pindex = pindex;
2815 info.end_pindex = pindex + psize - 1;
2822 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2823 pmap_object_init_pt_callback, &info);
2829 pmap_object_init_pt_callback(vm_page_t p, void *data)
2831 struct rb_vm_page_scan_info *info = data;
2832 vm_pindex_t rel_index;
2834 * don't allow an madvise to blow away our really
2835 * free pages allocating pv entries.
2837 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2838 vmstats.v_free_count < vmstats.v_free_reserved) {
2841 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2842 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2843 if ((p->queue - p->pc) == PQ_CACHE)
2844 vm_page_deactivate(p);
2846 rel_index = p->pindex - info->start_pindex;
2847 pmap_enter_quick(info->pmap,
2848 info->addr + x86_64_ptob(rel_index), p);
2855 * Return TRUE if the pmap is in shape to trivially
2856 * pre-fault the specified address.
2858 * Returns FALSE if it would be non-trivial or if a
2859 * pte is already loaded into the slot.
2862 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2867 pde = pmap_pde(pmap, addr);
2868 if (pde == NULL || *pde == 0)
2879 * Routine: pmap_change_wiring
2880 * Function: Change the wiring attribute for a map/virtual-address
2882 * In/out conditions:
2883 * The mapping must already exist in the pmap.
2886 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2893 pte = pmap_pte(pmap, va);
2895 if (wired && !pmap_pte_w(pte))
2896 pmap->pm_stats.wired_count++;
2897 else if (!wired && pmap_pte_w(pte))
2898 pmap->pm_stats.wired_count--;
2901 * Wiring is not a hardware characteristic so there is no need to
2902 * invalidate TLB. However, in an SMP environment we must use
2903 * a locked bus cycle to update the pte (if we are not using
2904 * the pmap_inval_*() API that is)... it's ok to do this for simple
2909 atomic_set_long(pte, PG_W);
2911 atomic_clear_long(pte, PG_W);
2914 atomic_set_long_nonlocked(pte, PG_W);
2916 atomic_clear_long_nonlocked(pte, PG_W);
2923 * Copy the range specified by src_addr/len
2924 * from the source map to the range dst_addr/len
2925 * in the destination map.
2927 * This routine is only advisory and need not do anything.
2930 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2931 vm_size_t len, vm_offset_t src_addr)
2935 pmap_inval_info info;
2937 vm_offset_t end_addr = src_addr + len;
2939 pd_entry_t src_frame, dst_frame;
2942 if (dst_addr != src_addr)
2945 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
2946 if (src_frame != (PTDpde & PG_FRAME)) {
2950 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
2951 if (dst_frame != (APTDpde & PG_FRAME)) {
2952 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2953 /* The page directory is not shared between CPUs */
2957 pmap_inval_init(&info);
2958 pmap_inval_add(&info, dst_pmap, -1);
2959 pmap_inval_add(&info, src_pmap, -1);
2962 * critical section protection is required to maintain the page/object
2963 * association, interrupts can free pages and remove them from
2967 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2968 pt_entry_t *src_pte, *dst_pte;
2969 vm_page_t dstmpte, srcmpte;
2970 vm_offset_t srcptepaddr;
2971 vm_pindex_t ptepindex;
2973 if (addr >= UPT_MIN_ADDRESS)
2974 panic("pmap_copy: invalid to pmap_copy page tables\n");
2977 * Don't let optional prefaulting of pages make us go
2978 * way below the low water mark of free pages or way
2979 * above high water mark of used pv entries.
2981 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2982 pv_entry_count > pv_entry_high_water)
2985 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2986 ptepindex = addr >> PDRSHIFT;
2989 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2991 if (srcptepaddr == 0)
2994 if (srcptepaddr & PG_PS) {
2996 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2997 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2998 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
3004 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
3005 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
3006 (srcmpte->flags & PG_BUSY)) {
3010 if (pdnxt > end_addr)
3013 src_pte = vtopte(addr);
3015 dst_pte = avtopte(addr);
3017 while (addr < pdnxt) {
3022 * we only virtual copy managed pages
3024 if ((ptetemp & PG_MANAGED) != 0) {
3026 * We have to check after allocpte for the
3027 * pte still being around... allocpte can
3030 * pmap_allocpte() can block. If we lose
3031 * our page directory mappings we stop.
3033 dstmpte = pmap_allocpte(dst_pmap, addr);
3036 if (src_frame != (PTDpde & PG_FRAME) ||
3037 dst_frame != (APTDpde & PG_FRAME)
3039 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3040 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3042 } else if ((*dst_pte == 0) &&
3043 (ptetemp = *src_pte) != 0 &&
3044 (ptetemp & PG_MANAGED)) {
3046 * Clear the modified and
3047 * accessed (referenced) bits
3050 m = PHYS_TO_VM_PAGE(ptetemp);
3051 *dst_pte = ptetemp & ~(PG_M | PG_A);
3052 ++dst_pmap->pm_stats.resident_count;
3053 pmap_insert_entry(dst_pmap, addr,
3055 KKASSERT(m->flags & PG_MAPPED);
3057 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3058 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3062 if (dstmpte->hold_count >= srcmpte->hold_count)
3072 pmap_inval_flush(&info);
3079 * Zero the specified physical page.
3081 * This function may be called from an interrupt and no locking is
3085 pmap_zero_page(vm_paddr_t phys)
3087 vm_offset_t va = PHYS_TO_DMAP(phys);
3089 pagezero((void *)va);
3093 * pmap_page_assertzero:
3095 * Assert that a page is empty, panic if it isn't.
3098 pmap_page_assertzero(vm_paddr_t phys)
3100 vm_offset_t virt = PHYS_TO_DMAP(phys);
3103 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3104 if (*(long *)((char *)virt + i) != 0) {
3105 panic("pmap_page_assertzero() @ %p not zero!\n", (void *)virt);
3113 * Zero part of a physical page by mapping it into memory and clearing
3114 * its contents with bzero.
3116 * off and size may not cover an area beyond a single hardware page.
3119 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3121 vm_offset_t virt = PHYS_TO_DMAP(phys);
3123 bzero((char *)virt + off, size);
3129 * Copy the physical page from the source PA to the target PA.
3130 * This function may be called from an interrupt. No locking
3134 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3136 vm_offset_t src_virt, dst_virt;
3138 src_virt = PHYS_TO_DMAP(src);
3139 dst_virt = PHYS_TO_DMAP(dst);
3140 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3144 * pmap_copy_page_frag:
3146 * Copy the physical page from the source PA to the target PA.
3147 * This function may be called from an interrupt. No locking
3151 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3153 vm_offset_t src_virt, dst_virt;
3155 src_virt = PHYS_TO_DMAP(src);
3156 dst_virt = PHYS_TO_DMAP(dst);
3158 bcopy((char *)src_virt + (src & PAGE_MASK),
3159 (char *)dst_virt + (dst & PAGE_MASK),
3164 * Returns true if the pmap's pv is one of the first
3165 * 16 pvs linked to from this page. This count may
3166 * be changed upwards or downwards in the future; it
3167 * is only necessary that true be returned for a small
3168 * subset of pmaps for proper page aging.
3171 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3176 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3181 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3182 if (pv->pv_pmap == pmap) {
3195 * Remove all pages from specified address space
3196 * this aids process exit speeds. Also, this code
3197 * is special cased for current process only, but
3198 * can have the more generic (and slightly slower)
3199 * mode enabled. This is much faster than pmap_remove
3200 * in the case of running down an entire address space.
3203 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3206 pt_entry_t *pte, tpte;
3209 pmap_inval_info info;
3211 int save_generation;
3213 lp = curthread->td_lwp;
3214 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3219 pmap_inval_init(&info);
3221 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3222 if (pv->pv_va >= eva || pv->pv_va < sva) {
3223 npv = TAILQ_NEXT(pv, pv_plist);
3227 KKASSERT(pmap == pv->pv_pmap);
3230 pte = vtopte(pv->pv_va);
3232 pte = pmap_pte_quick(pmap, pv->pv_va);
3233 if (pmap->pm_active)
3234 pmap_inval_add(&info, pmap, pv->pv_va);
3237 * We cannot remove wired pages from a process' mapping
3241 npv = TAILQ_NEXT(pv, pv_plist);
3244 tpte = pte_load_clear(pte);
3246 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3248 KASSERT(m < &vm_page_array[vm_page_array_size],
3249 ("pmap_remove_pages: bad tpte %lx", tpte));
3251 KKASSERT(pmap->pm_stats.resident_count > 0);
3252 --pmap->pm_stats.resident_count;
3255 * Update the vm_page_t clean and reference bits.
3261 npv = TAILQ_NEXT(pv, pv_plist);
3262 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3263 save_generation = ++pmap->pm_generation;
3265 m->md.pv_list_count--;
3266 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3267 if (TAILQ_EMPTY(&m->md.pv_list))
3268 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3270 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3274 * Restart the scan if we blocked during the unuse or free
3275 * calls and other removals were made.
3277 if (save_generation != pmap->pm_generation) {
3278 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3279 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3282 pmap_inval_flush(&info);
3287 * pmap_testbit tests bits in pte's
3288 * note that the testbit/clearbit routines are inline,
3289 * and a lot of things compile-time evaluate.
3293 pmap_testbit(vm_page_t m, int bit)
3298 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3301 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3306 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3308 * if the bit being tested is the modified bit, then
3309 * mark clean_map and ptes as never
3312 if (bit & (PG_A|PG_M)) {
3313 if (!pmap_track_modified(pv->pv_va))
3317 #if defined(PMAP_DIAGNOSTIC)
3318 if (pv->pv_pmap == NULL) {
3319 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3323 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3334 * this routine is used to modify bits in ptes
3338 pmap_clearbit(vm_page_t m, int bit)
3340 struct pmap_inval_info info;
3345 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3348 pmap_inval_init(&info);
3352 * Loop over all current mappings setting/clearing as appropos If
3353 * setting RO do we need to clear the VAC?
3355 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3357 * don't write protect pager mappings
3360 if (!pmap_track_modified(pv->pv_va))
3364 #if defined(PMAP_DIAGNOSTIC)
3365 if (pv->pv_pmap == NULL) {
3366 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3372 * Careful here. We can use a locked bus instruction to
3373 * clear PG_A or PG_M safely but we need to synchronize
3374 * with the target cpus when we mess with PG_RW.
3376 * We do not have to force synchronization when clearing
3377 * PG_M even for PTEs generated via virtual memory maps,
3378 * because the virtual kernel will invalidate the pmap
3379 * entry when/if it needs to resynchronize the Modify bit.
3382 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
3383 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3390 atomic_clear_long(pte, PG_M|PG_RW);
3393 * The cpu may be trying to set PG_M
3394 * simultaniously with our clearing
3397 if (!atomic_cmpset_long(pte, pbits,
3401 } else if (bit == PG_M) {
3403 * We could also clear PG_RW here to force
3404 * a fault on write to redetect PG_M for
3405 * virtual kernels, but it isn't necessary
3406 * since virtual kernels invalidate the pte
3407 * when they clear the VPTE_M bit in their
3408 * virtual page tables.
3410 atomic_clear_long(pte, PG_M);
3412 atomic_clear_long(pte, bit);
3416 pmap_inval_flush(&info);
3421 * pmap_page_protect:
3423 * Lower the permission for all mappings to a given page.
3426 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3428 /* JG NX support? */
3429 if ((prot & VM_PROT_WRITE) == 0) {
3430 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3431 pmap_clearbit(m, PG_RW);
3432 vm_page_flag_clear(m, PG_WRITEABLE);
3440 pmap_phys_address(vm_pindex_t ppn)
3442 return (x86_64_ptob(ppn));
3446 * pmap_ts_referenced:
3448 * Return a count of reference bits for a page, clearing those bits.
3449 * It is not necessary for every reference bit to be cleared, but it
3450 * is necessary that 0 only be returned when there are truly no
3451 * reference bits set.
3453 * XXX: The exact number of bits to check and clear is a matter that
3454 * should be tested and standardized at some point in the future for
3455 * optimal aging of shared pages.
3458 pmap_ts_referenced(vm_page_t m)
3460 pv_entry_t pv, pvf, pvn;
3464 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3469 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3474 pvn = TAILQ_NEXT(pv, pv_list);
3477 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3478 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3481 if (!pmap_track_modified(pv->pv_va))
3484 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3486 if (pte && (*pte & PG_A)) {
3488 atomic_clear_long(pte, PG_A);
3490 atomic_clear_long_nonlocked(pte, PG_A);
3497 } while ((pv = pvn) != NULL && pv != pvf);
3507 * Return whether or not the specified physical page was modified
3508 * in any physical maps.
3511 pmap_is_modified(vm_page_t m)
3513 return pmap_testbit(m, PG_M);
3517 * Clear the modify bits on the specified physical page.
3520 pmap_clear_modify(vm_page_t m)
3522 pmap_clearbit(m, PG_M);
3526 * pmap_clear_reference:
3528 * Clear the reference bit on the specified physical page.
3531 pmap_clear_reference(vm_page_t m)
3533 pmap_clearbit(m, PG_A);
3537 * Miscellaneous support routines follow
3542 i386_protection_init(void)
3546 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3547 kp = protection_codes;
3548 for (prot = 0; prot < 8; prot++) {
3550 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3552 * Read access is also 0. There isn't any execute bit,
3553 * so just make it readable.
3555 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3556 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3557 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3560 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3561 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3562 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3563 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3571 * Map a set of physical memory pages into the kernel virtual
3572 * address space. Return a pointer to where it is mapped. This
3573 * routine is intended to be used for mapping device memory,
3576 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3580 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3582 vm_offset_t va, tmpva, offset;
3585 offset = pa & PAGE_MASK;
3586 size = roundup(offset + size, PAGE_SIZE);
3588 va = kmem_alloc_nofault(&kernel_map, size);
3590 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3592 pa = pa & ~PAGE_MASK;
3593 for (tmpva = va; size > 0;) {
3594 pte = vtopte(tmpva);
3595 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3603 return ((void *)(va + offset));
3607 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3609 vm_offset_t va, tmpva, offset;
3612 offset = pa & PAGE_MASK;
3613 size = roundup(offset + size, PAGE_SIZE);
3615 va = kmem_alloc_nofault(&kernel_map, size);
3617 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3619 pa = pa & ~PAGE_MASK;
3620 for (tmpva = va; size > 0;) {
3621 pte = vtopte(tmpva);
3622 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3630 return ((void *)(va + offset));
3634 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3636 vm_offset_t base, offset;
3638 base = va & ~PAGE_MASK;
3639 offset = va & PAGE_MASK;
3640 size = roundup(offset + size, PAGE_SIZE);
3641 pmap_qremove(va, size >> PAGE_SHIFT);
3642 kmem_free(&kernel_map, base, size);
3646 * perform the pmap work for mincore
3649 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3651 pt_entry_t *ptep, pte;
3655 ptep = pmap_pte(pmap, addr);
3660 if ((pte = *ptep) != 0) {
3663 val = MINCORE_INCORE;
3664 if ((pte & PG_MANAGED) == 0)
3667 pa = pte & PG_FRAME;
3669 m = PHYS_TO_VM_PAGE(pa);
3675 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3677 * Modified by someone
3679 else if (m->dirty || pmap_is_modified(m))
3680 val |= MINCORE_MODIFIED_OTHER;
3685 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3688 * Referenced by someone
3690 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3691 val |= MINCORE_REFERENCED_OTHER;
3692 vm_page_flag_set(m, PG_REFERENCED);
3699 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3700 * vmspace will be ref'd and the old one will be deref'd.
3702 * The vmspace for all lwps associated with the process will be adjusted
3703 * and cr3 will be reloaded if any lwp is the current lwp.
3706 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3708 struct vmspace *oldvm;
3712 oldvm = p->p_vmspace;
3713 if (oldvm != newvm) {
3714 p->p_vmspace = newvm;
3715 KKASSERT(p->p_nthreads == 1);
3716 lp = RB_ROOT(&p->p_lwp_tree);
3717 pmap_setlwpvm(lp, newvm);
3719 sysref_get(&newvm->vm_sysref);
3720 sysref_put(&oldvm->vm_sysref);
3727 * Set the vmspace for a LWP. The vmspace is almost universally set the
3728 * same as the process vmspace, but virtual kernels need to swap out contexts
3729 * on a per-lwp basis.
3732 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3734 struct vmspace *oldvm;
3738 oldvm = lp->lwp_vmspace;
3740 if (oldvm != newvm) {
3741 lp->lwp_vmspace = newvm;
3742 if (curthread->td_lwp == lp) {
3743 pmap = vmspace_pmap(newvm);
3745 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3747 pmap->pm_active |= 1;
3749 #if defined(SWTCH_OPTIM_STATS)
3752 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
3753 load_cr3(curthread->td_pcb->pcb_cr3);
3754 pmap = vmspace_pmap(oldvm);
3756 atomic_clear_int(&pmap->pm_active,
3757 1 << mycpu->gd_cpuid);
3759 pmap->pm_active &= ~1;
3767 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3770 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3774 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3781 static void pads (pmap_t pm);
3782 void pmap_pvdump (vm_paddr_t pa);
3784 /* print address space of pmap*/
3793 if (pm == &kernel_pmap)
3796 for (i = 0; i < NPDEPG; i++) {
3804 pmap_pvdump(vm_paddr_t pa)
3809 kprintf("pa %08llx", (long long)pa);
3810 m = PHYS_TO_VM_PAGE(pa);
3811 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3813 kprintf(" -> pmap %p, va %x, flags %x",
3814 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3816 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);