4 * Copyright (c) 1991 Regents of the University of California.
5 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
7 * Copyright (c) 2003 Peter Wemm
8 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
9 * Copyright (c) 2008, 2009 The DragonFly Project.
10 * Copyright (c) 2008, 2009 Jordan Gordeev.
11 * All rights reserved.
13 * This code is derived from software contributed to Berkeley by
14 * the Systems Programming Group of the University of Utah Computer
15 * Science Department and William Jolitz of UUNET Technologies Inc.
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions
20 * 1. Redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer.
22 * 2. Redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution.
25 * 3. All advertising materials mentioning features or use of this software
26 * must display the following acknowledgement:
27 * This product includes software developed by the University of
28 * California, Berkeley and its contributors.
29 * 4. Neither the name of the University nor the names of its contributors
30 * may be used to endorse or promote products derived from this software
31 * without specific prior written permission.
33 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
34 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
35 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
36 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
37 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
38 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
39 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
40 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
41 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
42 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
45 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
46 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
50 * Manages physical address maps.
52 * In addition to hardware address maps, this
53 * module is called upon to provide software-use-only
54 * maps which may or may not be stored in the same
55 * form as hardware maps. These pseudo-maps are
56 * used to store intermediate results from copy
57 * operations to and from address spaces.
59 * Since the information managed by this module is
60 * also stored by the logical address mapping module,
61 * this module may throw away valid virtual-to-physical
62 * mappings at almost any time. However, invalidations
63 * of virtual-to-physical mappings must be done as
66 * In order to cope with hardware architectures which
67 * make virtual-to-physical map invalidates expensive,
68 * this module may delay invalidate or reduced protection
69 * operations until such time as they are actually
70 * necessary. This module is given full information as
71 * to which processors are currently using which maps,
72 * and to when physical maps must be made correct.
76 #include "opt_disable_pse.h"
79 #include "opt_msgbuf.h"
81 #include <sys/param.h>
82 #include <sys/systm.h>
83 #include <sys/kernel.h>
85 #include <sys/msgbuf.h>
86 #include <sys/vmmeter.h>
90 #include <vm/vm_param.h>
91 #include <sys/sysctl.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_extern.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pager.h>
100 #include <vm/vm_zone.h>
102 #include <sys/user.h>
103 #include <sys/thread2.h>
104 #include <sys/sysref2.h>
106 #include <machine/cputypes.h>
107 #include <machine/md_var.h>
108 #include <machine/specialreg.h>
109 #include <machine/smp.h>
110 #include <machine_base/apic/apicreg.h>
111 #include <machine/globaldata.h>
112 #include <machine/pmap.h>
113 #include <machine/pmap_inval.h>
117 #define PMAP_KEEP_PDIRS
118 #ifndef PMAP_SHPGPERPROC
119 #define PMAP_SHPGPERPROC 200
122 #if defined(DIAGNOSTIC)
123 #define PMAP_DIAGNOSTIC
129 * Get PDEs and PTEs for user/kernel address space
131 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
132 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
134 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
135 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
136 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
137 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
138 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
142 * Given a map and a machine independent protection code,
143 * convert to a vax protection code.
145 #define pte_prot(m, p) \
146 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
147 static int protection_codes[8];
149 struct pmap kernel_pmap;
150 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
152 vm_paddr_t avail_start; /* PA of first available physical page */
153 vm_paddr_t avail_end; /* PA of last available physical page */
154 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
155 vm_offset_t virtual2_end;
156 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
157 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
158 vm_offset_t KvaStart; /* VA start of KVA space */
159 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
160 vm_offset_t KvaSize; /* max size of kernel virtual address space */
161 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
162 static int pgeflag; /* PG_G or-in */
163 static int pseflag; /* PG_PS or-in */
165 static vm_object_t kptobj;
168 static vm_paddr_t dmaplimit;
170 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
172 static uint64_t KPTbase;
173 static uint64_t KPTphys;
174 static uint64_t KPDphys; /* phys addr of kernel level 2 */
175 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
176 uint64_t KPDPphys; /* phys addr of kernel level 3 */
177 uint64_t KPML4phys; /* phys addr of kernel level 4 */
179 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
180 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
183 * Data for the pv entry allocation mechanism
185 static vm_zone_t pvzone;
186 static struct vm_zone pvzone_store;
187 static struct vm_object pvzone_obj;
188 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
189 static int pmap_pagedaemon_waken = 0;
190 static struct pv_entry *pvinit;
193 * All those kernel PT submaps that BSD is so fond of
195 pt_entry_t *CMAP1 = 0, *ptmmap;
196 caddr_t CADDR1 = 0, ptvmmap = 0;
197 static pt_entry_t *msgbufmap;
198 struct msgbuf *msgbufp=0;
203 static pt_entry_t *pt_crashdumpmap;
204 static caddr_t crashdumpmap;
206 extern pt_entry_t *SMPpt;
207 extern uint64_t SMPptpa;
211 static pv_entry_t get_pv_entry (void);
212 static void i386_protection_init (void);
213 static void create_pagetables(vm_paddr_t *firstaddr);
214 static void pmap_remove_all (vm_page_t m);
215 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
216 vm_offset_t sva, pmap_inval_info_t info);
217 static void pmap_remove_page (struct pmap *pmap,
218 vm_offset_t va, pmap_inval_info_t info);
219 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
220 vm_offset_t va, pmap_inval_info_t info);
221 static boolean_t pmap_testbit (vm_page_t m, int bit);
222 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
223 vm_page_t mpte, vm_page_t m);
225 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
227 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
228 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
229 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
230 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
231 static int _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
232 pmap_inval_info_t info);
233 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
234 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
236 static unsigned pdir4mb;
239 * Move the kernel virtual free pointer to the next
240 * 2MB. This is used to help improve performance
241 * by using a large (2MB) page for much of the kernel
242 * (.text, .data, .bss)
246 pmap_kmem_choose(vm_offset_t addr)
248 vm_offset_t newaddr = addr;
250 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
257 * Super fast pmap_pte routine best used when scanning the pv lists.
258 * This eliminates many course-grained invltlb calls. Note that many of
259 * the pv list scans are across different pmaps and it is very wasteful
260 * to do an entire invltlb when checking a single mapping.
262 * Should only be called while in a critical section.
264 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
268 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
270 return pmap_pte(pmap, va);
273 /* Return a non-clipped PD index for a given VA */
276 pmap_pde_pindex(vm_offset_t va)
278 return va >> PDRSHIFT;
281 /* Return various clipped indexes for a given VA */
284 pmap_pte_index(vm_offset_t va)
287 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
292 pmap_pde_index(vm_offset_t va)
295 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
300 pmap_pdpe_index(vm_offset_t va)
303 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
308 pmap_pml4e_index(vm_offset_t va)
311 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
314 /* Return a pointer to the PML4 slot that corresponds to a VA */
317 pmap_pml4e(pmap_t pmap, vm_offset_t va)
320 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
323 /* Return a pointer to the PDP slot that corresponds to a VA */
326 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
330 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
331 return (&pdpe[pmap_pdpe_index(va)]);
334 /* Return a pointer to the PDP slot that corresponds to a VA */
337 pmap_pdpe(pmap_t pmap, vm_offset_t va)
341 pml4e = pmap_pml4e(pmap, va);
342 if ((*pml4e & PG_V) == 0)
344 return (pmap_pml4e_to_pdpe(pml4e, va));
347 /* Return a pointer to the PD slot that corresponds to a VA */
350 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
354 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
355 return (&pde[pmap_pde_index(va)]);
358 /* Return a pointer to the PD slot that corresponds to a VA */
361 pmap_pde(pmap_t pmap, vm_offset_t va)
365 pdpe = pmap_pdpe(pmap, va);
366 if (pdpe == NULL || (*pdpe & PG_V) == 0)
368 return (pmap_pdpe_to_pde(pdpe, va));
371 /* Return a pointer to the PT slot that corresponds to a VA */
374 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
378 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
379 return (&pte[pmap_pte_index(va)]);
382 /* Return a pointer to the PT slot that corresponds to a VA */
385 pmap_pte(pmap_t pmap, vm_offset_t va)
389 pde = pmap_pde(pmap, va);
390 if (pde == NULL || (*pde & PG_V) == 0)
392 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
393 return ((pt_entry_t *)pde);
394 return (pmap_pde_to_pte(pde, va));
399 vtopte(vm_offset_t va)
401 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
403 return (PTmap + ((va >> PAGE_SHIFT) & mask));
408 vtopde(vm_offset_t va)
410 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
412 return (PDmap + ((va >> PDRSHIFT) & mask));
416 allocpages(vm_paddr_t *firstaddr, int n)
421 bzero((void *)ret, n * PAGE_SIZE);
422 *firstaddr += n * PAGE_SIZE;
428 create_pagetables(vm_paddr_t *firstaddr)
433 * We are running (mostly) V=P at this point
435 * Calculate NKPT - number of kernel page tables. We have to
436 * accomodoate prealloction of the vm_page_array, dump bitmap,
437 * MSGBUF_SIZE, and other stuff. Be generous.
439 * Maxmem is in pages.
441 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
442 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
445 nkpt = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
446 nkpt += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E + ndmpdp) +
453 KPTbase = allocpages(firstaddr, nkpt);
454 KPTphys = allocpages(firstaddr, nkpt);
455 KPML4phys = allocpages(firstaddr, 1);
456 KPDPphys = allocpages(firstaddr, NKPML4E);
459 * Calculate the page directory base for KERNBASE,
460 * that is where we start populating the page table pages.
461 * Basically this is the end - 2.
463 KPDphys = allocpages(firstaddr, NKPDPE);
464 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
466 DMPDPphys = allocpages(firstaddr, NDMPML4E);
467 if ((amd_feature & AMDID_PAGE1GB) == 0)
468 DMPDphys = allocpages(firstaddr, ndmpdp);
469 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
472 * Fill in the underlying page table pages for the area around
473 * KERNBASE. This remaps low physical memory to KERNBASE.
475 * Read-only from zero to physfree
476 * XXX not fully used, underneath 2M pages
478 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
479 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
480 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
484 * Now map the initial kernel page tables. One block of page
485 * tables is placed at the beginning of kernel virtual memory,
486 * and another block is placed at KERNBASE to map the kernel binary,
487 * data, bss, and initial pre-allocations.
489 for (i = 0; i < nkpt; i++) {
490 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
491 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
493 for (i = 0; i < nkpt; i++) {
494 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
495 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
499 * Map from zero to end of allocations using 2M pages as an
500 * optimization. This will bypass some of the KPTBase pages
501 * above in the KERNBASE area.
503 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
504 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
505 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
509 * And connect up the PD to the PDP. The kernel pmap is expected
510 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
512 for (i = 0; i < NKPDPE; i++) {
513 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
514 KPDphys + (i << PAGE_SHIFT);
515 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
519 /* Now set up the direct map space using either 2MB or 1GB pages */
520 /* Preset PG_M and PG_A because demotion expects it */
521 if ((amd_feature & AMDID_PAGE1GB) == 0) {
522 for (i = 0; i < NPDEPG * ndmpdp; i++) {
523 ((pd_entry_t *)DMPDphys)[i] = (vm_paddr_t)i << PDRSHIFT;
524 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
527 /* And the direct map space's PDP */
528 for (i = 0; i < ndmpdp; i++) {
529 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
531 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
534 for (i = 0; i < ndmpdp; i++) {
535 ((pdp_entry_t *)DMPDPphys)[i] =
536 (vm_paddr_t)i << PDPSHIFT;
537 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
542 /* And recursively map PML4 to itself in order to get PTmap */
543 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
544 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
546 /* Connect the Direct Map slot up to the PML4 */
547 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
548 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
550 /* Connect the KVA slot up to the PML4 */
551 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
552 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
556 * Bootstrap the system enough to run with virtual memory.
558 * On the i386 this is called after mapping has already been enabled
559 * and just syncs the pmap module with what has already been done.
560 * [We can't call it easily with mapping off since the kernel is not
561 * mapped with PA == VA, hence we would have to relocate every address
562 * from the linked base (virtual) address "KERNBASE" to the actual
563 * (physical) address starting relative to 0]
566 pmap_bootstrap(vm_paddr_t *firstaddr)
570 struct mdglobaldata *gd;
573 KvaStart = VM_MIN_KERNEL_ADDRESS;
574 KvaEnd = VM_MAX_KERNEL_ADDRESS;
575 KvaSize = KvaEnd - KvaStart;
577 avail_start = *firstaddr;
580 * Create an initial set of page tables to run the kernel in.
582 create_pagetables(firstaddr);
584 virtual2_start = KvaStart;
585 virtual2_end = PTOV_OFFSET;
587 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
588 virtual_start = pmap_kmem_choose(virtual_start);
590 virtual_end = VM_MAX_KERNEL_ADDRESS;
592 /* XXX do %cr0 as well */
593 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
597 * Initialize protection array.
599 i386_protection_init();
602 * The kernel's pmap is statically allocated so we don't have to use
603 * pmap_create, which is unlikely to work correctly at this part of
604 * the boot sequence (XXX and which no longer exists).
606 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
607 kernel_pmap.pm_count = 1;
608 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
609 TAILQ_INIT(&kernel_pmap.pm_pvlist);
612 * Reserve some special page table entries/VA space for temporary
615 #define SYSMAP(c, p, v, n) \
616 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
622 * CMAP1/CMAP2 are used for zeroing and copying pages.
624 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
629 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
632 * ptvmmap is used for reading arbitrary physical pages via
635 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
638 * msgbufp is used to map the system message buffer.
639 * XXX msgbufmap is not used.
641 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
642 atop(round_page(MSGBUF_SIZE)))
649 * PG_G is terribly broken on SMP because we IPI invltlb's in some
650 * cases rather then invl1pg. Actually, I don't even know why it
651 * works under UP because self-referential page table mappings
656 if (cpu_feature & CPUID_PGE)
661 * Initialize the 4MB page size flag
665 * The 4MB page version of the initial
666 * kernel page mapping.
670 #if !defined(DISABLE_PSE)
671 if (cpu_feature & CPUID_PSE) {
674 * Note that we have enabled PSE mode
677 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
678 ptditmp &= ~(NBPDR - 1);
679 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
684 * Enable the PSE mode. If we are SMP we can't do this
685 * now because the APs will not be able to use it when
688 load_cr4(rcr4() | CR4_PSE);
691 * We can do the mapping here for the single processor
692 * case. We simply ignore the old page table page from
696 * For SMP, we still need 4K pages to bootstrap APs,
697 * PSE will be enabled as soon as all APs are up.
699 PTD[KPTDI] = (pd_entry_t)ptditmp;
706 * We need to finish setting up the globaldata page for the BSP.
707 * locore has already populated the page table for the mdglobaldata
710 pg = MDGLOBALDATA_BASEALLOC_PAGES;
711 gd = &CPU_prvspace[0].mdglobaldata;
712 gd->gd_CMAP1 = &SMPpt[pg + 0];
713 gd->gd_CMAP2 = &SMPpt[pg + 1];
714 gd->gd_CMAP3 = &SMPpt[pg + 2];
715 gd->gd_PMAP1 = &SMPpt[pg + 3];
716 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
717 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
718 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
719 gd->gd_PADDR1 = (pt_entry_t *)CPU_prvspace[0].PPAGE1;
726 * Set 4mb pdir for mp startup
731 if (pseflag && (cpu_feature & CPUID_PSE)) {
732 load_cr4(rcr4() | CR4_PSE);
733 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
741 * XXX: Hack. Required by pmap_init()
743 extern vm_offset_t cpu_apic_addr;
746 * Initialize the pmap module.
747 * Called by vm_init, to initialize any structures that the pmap
748 * system needs to map virtual memory.
749 * pmap_init has been enhanced to support in a fairly consistant
750 * way, discontiguous physical memory.
759 * object for kernel page table pages
761 /* JG I think the number can be arbitrary */
762 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
765 * Allocate memory for random pmap data structures. Includes the
769 for(i = 0; i < vm_page_array_size; i++) {
772 m = &vm_page_array[i];
773 TAILQ_INIT(&m->md.pv_list);
774 m->md.pv_list_count = 0;
778 * init the pv free list
780 initial_pvs = vm_page_array_size;
781 if (initial_pvs < MINPV)
783 pvzone = &pvzone_store;
784 pvinit = (void *)kmem_alloc(&kernel_map,
785 initial_pvs * sizeof (struct pv_entry));
786 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
787 pvinit, initial_pvs);
790 * Now it is safe to enable pv_table recording.
792 pmap_initialized = TRUE;
797 lapic = pmap_mapdev_uncacheable(cpu_apic_addr, sizeof(struct LAPIC));
802 * Initialize the address space (zone) for the pv_entries. Set a
803 * high water mark so that the system can recover from excessive
804 * numbers of pv entries.
809 int shpgperproc = PMAP_SHPGPERPROC;
812 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
813 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
814 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
815 pv_entry_high_water = 9 * (pv_entry_max / 10);
818 * Subtract out pages already installed in the zone (hack)
820 entry_max = pv_entry_max - vm_page_array_size;
824 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
828 /***************************************************
829 * Low level helper routines.....
830 ***************************************************/
832 #if defined(PMAP_DIAGNOSTIC)
835 * This code checks for non-writeable/modified pages.
836 * This should be an invalid condition.
840 pmap_nw_modified(pt_entry_t pte)
842 if ((pte & (PG_M|PG_RW)) == PG_M)
851 * this routine defines the region(s) of memory that should
852 * not be tested for the modified bit.
856 pmap_track_modified(vm_offset_t va)
858 if ((va < clean_sva) || (va >= clean_eva))
865 * Extract the physical page address associated with the map/VA pair.
867 * The caller must hold vm_token if non-blocking operation is desired.
870 pmap_extract(pmap_t pmap, vm_offset_t va)
874 pd_entry_t pde, *pdep;
876 lwkt_gettoken(&vm_token);
878 pdep = pmap_pde(pmap, va);
882 if ((pde & PG_PS) != 0) {
883 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
885 pte = pmap_pde_to_pte(pdep, va);
886 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
890 lwkt_reltoken(&vm_token);
895 * Extract the physical page address associated kernel virtual address.
898 pmap_kextract(vm_offset_t va)
903 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
904 pa = DMAP_TO_PHYS(va);
908 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
911 * Beware of a concurrent promotion that changes the
912 * PDE at this point! For example, vtopte() must not
913 * be used to access the PTE because it would use the
914 * new PDE. It is, however, safe to use the old PDE
915 * because the page table page is preserved by the
918 pa = *pmap_pde_to_pte(&pde, va);
919 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
925 /***************************************************
926 * Low level mapping routines.....
927 ***************************************************/
930 * Routine: pmap_kenter
932 * Add a wired page to the KVA
933 * NOTE! note that in order for the mapping to take effect -- you
934 * should do an invltlb after doing the pmap_kenter().
937 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
941 pmap_inval_info info;
943 pmap_inval_init(&info);
944 npte = pa | PG_RW | PG_V | pgeflag;
946 pmap_inval_interlock(&info, &kernel_pmap, va);
948 pmap_inval_deinterlock(&info, &kernel_pmap);
949 pmap_inval_done(&info);
953 * Routine: pmap_kenter_quick
955 * Similar to pmap_kenter(), except we only invalidate the
956 * mapping on the current CPU.
959 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
964 npte = pa | PG_RW | PG_V | pgeflag;
967 cpu_invlpg((void *)va);
971 pmap_kenter_sync(vm_offset_t va)
973 pmap_inval_info info;
975 pmap_inval_init(&info);
976 pmap_inval_interlock(&info, &kernel_pmap, va);
977 pmap_inval_deinterlock(&info, &kernel_pmap);
978 pmap_inval_done(&info);
982 pmap_kenter_sync_quick(vm_offset_t va)
984 cpu_invlpg((void *)va);
988 * remove a page from the kernel pagetables
991 pmap_kremove(vm_offset_t va)
994 pmap_inval_info info;
996 pmap_inval_init(&info);
998 pmap_inval_interlock(&info, &kernel_pmap, va);
1000 pmap_inval_deinterlock(&info, &kernel_pmap);
1001 pmap_inval_done(&info);
1005 pmap_kremove_quick(vm_offset_t va)
1010 cpu_invlpg((void *)va);
1014 * XXX these need to be recoded. They are not used in any critical path.
1017 pmap_kmodify_rw(vm_offset_t va)
1019 *vtopte(va) |= PG_RW;
1020 cpu_invlpg((void *)va);
1024 pmap_kmodify_nc(vm_offset_t va)
1026 *vtopte(va) |= PG_N;
1027 cpu_invlpg((void *)va);
1031 * Used to map a range of physical addresses into kernel virtual
1032 * address space during the low level boot, typically to map the
1033 * dump bitmap, message buffer, and vm_page_array.
1035 * These mappings are typically made at some pointer after the end of the
1038 * We could return PHYS_TO_DMAP(start) here and not allocate any
1039 * via (*virtp), but then kmem from userland and kernel dumps won't
1040 * have access to the related pointers.
1043 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1046 vm_offset_t va_start;
1048 /*return PHYS_TO_DMAP(start);*/
1053 while (start < end) {
1054 pmap_kenter_quick(va, start);
1064 * Add a list of wired pages to the kva
1065 * this routine is only used for temporary
1066 * kernel mappings that do not need to have
1067 * page modification or references recorded.
1068 * Note that old mappings are simply written
1069 * over. The page *must* be wired.
1072 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1076 end_va = va + count * PAGE_SIZE;
1078 while (va < end_va) {
1082 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1083 cpu_invlpg((void *)va);
1091 * This routine jerks page mappings from the
1092 * kernel -- it is meant only for temporary mappings.
1094 * MPSAFE, INTERRUPT SAFE (cluster callback)
1097 pmap_qremove(vm_offset_t va, int count)
1101 end_va = va + count * PAGE_SIZE;
1103 while (va < end_va) {
1108 cpu_invlpg((void *)va);
1115 * This routine works like vm_page_lookup() but also blocks as long as the
1116 * page is busy. This routine does not busy the page it returns.
1118 * Unless the caller is managing objects whos pages are in a known state,
1119 * the call should be made with a critical section held so the page's object
1120 * association remains valid on return.
1124 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1129 m = vm_page_lookup(object, pindex);
1130 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
1136 * Create a new thread and optionally associate it with a (new) process.
1137 * NOTE! the new thread's cpu may not equal the current cpu.
1140 pmap_init_thread(thread_t td)
1142 /* enforce pcb placement & alignment */
1143 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1144 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1145 td->td_savefpu = &td->td_pcb->pcb_save;
1146 td->td_sp = (char *)td->td_pcb; /* no -16 */
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_interlock(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;
1246 pmap_inval_deinterlock(info, pmap);
1248 if (m->pindex < NUPDE) {
1249 /* We just released a PT, unhold the matching PD */
1252 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
1253 pmap_unwire_pte_hold(pmap, va, pdpg, info);
1255 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
1256 /* We just released a PD, unhold the matching PDP */
1259 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
1260 pmap_unwire_pte_hold(pmap, va, pdppg, info);
1264 * This was our last hold, the page had better be unwired
1265 * after we decrement wire_count.
1267 * FUTURE NOTE: shared page directory page could result in
1268 * multiple wire counts.
1272 KKASSERT(m->wire_count == 0);
1273 --vmstats.v_wire_count;
1274 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1276 vm_page_free_zero(m);
1282 * After removing a page table entry, this routine is used to
1283 * conditionally free the page, and manage the hold/wire counts.
1287 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1288 pmap_inval_info_t info)
1290 vm_pindex_t ptepindex;
1292 if (va >= VM_MAX_USER_ADDRESS)
1296 ptepindex = pmap_pde_pindex(va);
1298 if (pmap->pm_ptphint &&
1299 (pmap->pm_ptphint->pindex == ptepindex)) {
1300 mpte = pmap->pm_ptphint;
1303 pmap_inval_flush(info);
1304 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1305 pmap->pm_ptphint = mpte;
1310 return pmap_unwire_pte_hold(pmap, va, mpte, info);
1314 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1315 * it, and IdlePTD, represents the template used to update all other pmaps.
1317 * On architectures where the kernel pmap is not integrated into the user
1318 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1319 * kernel_pmap should be used to directly access the kernel_pmap.
1322 pmap_pinit0(struct pmap *pmap)
1324 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1326 pmap->pm_active = 0;
1327 pmap->pm_ptphint = NULL;
1328 TAILQ_INIT(&pmap->pm_pvlist);
1329 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1333 * Initialize a preallocated and zeroed pmap structure,
1334 * such as one in a vmspace structure.
1337 pmap_pinit(struct pmap *pmap)
1342 * No need to allocate page table space yet but we do need a valid
1343 * page directory table.
1345 if (pmap->pm_pml4 == NULL) {
1347 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1351 * Allocate an object for the ptes
1353 if (pmap->pm_pteobj == NULL)
1354 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1);
1357 * Allocate the page directory page, unless we already have
1358 * one cached. If we used the cached page the wire_count will
1359 * already be set appropriately.
1361 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1362 ptdpg = vm_page_grab(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I,
1363 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1364 pmap->pm_pdirm = ptdpg;
1365 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1366 ptdpg->valid = VM_PAGE_BITS_ALL;
1367 if (ptdpg->wire_count == 0)
1368 ++vmstats.v_wire_count;
1369 ptdpg->wire_count = 1;
1370 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1372 if ((ptdpg->flags & PG_ZERO) == 0)
1373 bzero(pmap->pm_pml4, PAGE_SIZE);
1375 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1376 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1378 /* install self-referential address mapping entry */
1379 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1382 pmap->pm_active = 0;
1383 pmap->pm_ptphint = NULL;
1384 TAILQ_INIT(&pmap->pm_pvlist);
1385 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1386 pmap->pm_stats.resident_count = 1;
1390 * Clean up a pmap structure so it can be physically freed. This routine
1391 * is called by the vmspace dtor function. A great deal of pmap data is
1392 * left passively mapped to improve vmspace management so we have a bit
1393 * of cleanup work to do here.
1396 pmap_puninit(pmap_t pmap)
1400 KKASSERT(pmap->pm_active == 0);
1401 lwkt_gettoken(&vm_token);
1402 if ((p = pmap->pm_pdirm) != NULL) {
1403 KKASSERT(pmap->pm_pml4 != NULL);
1404 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1405 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1407 vmstats.v_wire_count--;
1408 KKASSERT((p->flags & PG_BUSY) == 0);
1410 vm_page_free_zero(p);
1411 pmap->pm_pdirm = NULL;
1413 if (pmap->pm_pml4) {
1414 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1415 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1416 pmap->pm_pml4 = NULL;
1418 if (pmap->pm_pteobj) {
1419 vm_object_deallocate(pmap->pm_pteobj);
1420 pmap->pm_pteobj = NULL;
1422 lwkt_reltoken(&vm_token);
1426 * Wire in kernel global address entries. To avoid a race condition
1427 * between pmap initialization and pmap_growkernel, this procedure
1428 * adds the pmap to the master list (which growkernel scans to update),
1429 * then copies the template.
1432 pmap_pinit2(struct pmap *pmap)
1435 lwkt_gettoken(&vm_token);
1436 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1437 /* XXX copies current process, does not fill in MPPTDI */
1438 lwkt_reltoken(&vm_token);
1443 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1444 * 0 on failure (if the procedure had to sleep).
1446 * When asked to remove the page directory page itself, we actually just
1447 * leave it cached so we do not have to incur the SMP inval overhead of
1448 * removing the kernel mapping. pmap_puninit() will take care of it.
1452 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1455 * This code optimizes the case of freeing non-busy
1456 * page-table pages. Those pages are zero now, and
1457 * might as well be placed directly into the zero queue.
1459 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1465 * Remove the page table page from the processes address space.
1467 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1469 * We are the pml4 table itself.
1471 /* XXX anything to do here? */
1472 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1474 * Remove a PDP page from the PML4. We do not maintain
1475 * hold counts on the PML4 page.
1481 m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1482 KKASSERT(m4 != NULL);
1483 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1484 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1485 KKASSERT(pml4[idx] != 0);
1487 } else if (p->pindex >= NUPDE) {
1489 * Remove a PD page from the PDP and drop the hold count
1490 * on the PDP. The PDP is left cached in the pmap if
1491 * the hold count drops to 0 so the wire count remains
1498 m3 = vm_page_lookup(pmap->pm_pteobj,
1499 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1500 KKASSERT(m3 != NULL);
1501 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1502 idx = (p->pindex - NUPDE) % NPDPEPG;
1503 KKASSERT(pdp[idx] != 0);
1508 * Remove a PT page from the PD and drop the hold count
1509 * on the PD. The PD is left cached in the pmap if
1510 * the hold count drops to 0 so the wire count remains
1517 m2 = vm_page_lookup(pmap->pm_pteobj,
1518 NUPDE + p->pindex / NPDEPG);
1519 KKASSERT(m2 != NULL);
1520 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1521 idx = p->pindex % NPDEPG;
1527 * One fewer mappings in the pmap. p's hold count had better
1530 KKASSERT(pmap->pm_stats.resident_count > 0);
1531 --pmap->pm_stats.resident_count;
1533 panic("pmap_release: freeing held page table page");
1534 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1535 pmap->pm_ptphint = NULL;
1538 * We leave the top-level page table page cached, wired, and mapped in
1539 * the pmap until the dtor function (pmap_puninit()) gets called.
1540 * However, still clean it up so we can set PG_ZERO.
1542 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1543 bzero(pmap->pm_pml4, PAGE_SIZE);
1544 vm_page_flag_set(p, PG_ZERO);
1548 KKASSERT(p->wire_count == 0);
1549 vmstats.v_wire_count--;
1550 /* JG eventually revert to using vm_page_free_zero() */
1557 * This routine is called when various levels in the page table need to
1558 * be populated. This routine cannot fail.
1562 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1567 * Find or fabricate a new pagetable page. This will busy the page.
1569 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1570 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1571 if ((m->flags & PG_ZERO) == 0) {
1572 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1575 KASSERT(m->queue == PQ_NONE,
1576 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1579 * Increment the hold count for the page we will be returning to
1583 if (m->wire_count++ == 0)
1584 vmstats.v_wire_count++;
1587 * Map the pagetable page into the process address space, if
1588 * it isn't already there.
1590 * It is possible that someone else got in and mapped the page
1591 * directory page while we were blocked, if so just unbusy and
1592 * return the held page.
1594 if (ptepindex >= (NUPDE + NUPDPE)) {
1596 * Wire up a new PDP page in the PML4
1598 vm_pindex_t pml4index;
1601 pml4index = ptepindex - (NUPDE + NUPDPE);
1602 pml4 = &pmap->pm_pml4[pml4index];
1604 if (--m->wire_count == 0)
1605 --vmstats.v_wire_count;
1609 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1610 } else if (ptepindex >= NUPDE) {
1612 * Wire up a new PD page in the PDP
1614 vm_pindex_t pml4index;
1615 vm_pindex_t pdpindex;
1620 pdpindex = ptepindex - NUPDE;
1621 pml4index = pdpindex >> NPML4EPGSHIFT;
1623 pml4 = &pmap->pm_pml4[pml4index];
1624 if ((*pml4 & PG_V) == 0) {
1626 * Have to allocate a new PDP page, recurse.
1627 * This always succeeds. Returned page will
1630 pdppg = _pmap_allocpte(pmap,
1631 NUPDE + NUPDPE + pml4index);
1634 * Add a held reference to the PDP page.
1636 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1637 pdppg->hold_count++;
1641 * Now find the pdp_entry and map the PDP. If the PDP
1642 * has already been mapped unwind and return the
1643 * already-mapped PDP held.
1645 * pdppg is left held (hold_count is incremented for
1646 * each PD in the PDP).
1648 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1649 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1651 vm_page_unhold(pdppg);
1652 if (--m->wire_count == 0)
1653 --vmstats.v_wire_count;
1657 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1660 * Wire up the new PT page in the PD
1662 vm_pindex_t pml4index;
1663 vm_pindex_t pdpindex;
1669 pdpindex = ptepindex >> NPDPEPGSHIFT;
1670 pml4index = pdpindex >> NPML4EPGSHIFT;
1673 * Locate the PDP page in the PML4, then the PD page in
1674 * the PDP. If either does not exist we simply recurse
1677 * We can just recurse on the PD page as it will recurse
1678 * on the PDP if necessary.
1680 pml4 = &pmap->pm_pml4[pml4index];
1681 if ((*pml4 & PG_V) == 0) {
1682 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1683 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1684 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1686 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1687 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1688 if ((*pdp & PG_V) == 0) {
1689 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1691 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1697 * Now fill in the pte in the PD. If the pte already exists
1698 * (again, if we raced the grab), unhold pdpg and unwire
1699 * m, returning a held m.
1701 * pdpg is left held (hold_count is incremented for
1702 * each PT in the PD).
1704 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
1705 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1707 vm_page_unhold(pdpg);
1708 if (--m->wire_count == 0)
1709 --vmstats.v_wire_count;
1713 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1717 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1718 * valid bits, mapped flag, unbusy, and we're done.
1720 pmap->pm_ptphint = m;
1721 ++pmap->pm_stats.resident_count;
1723 m->valid = VM_PAGE_BITS_ALL;
1724 vm_page_flag_clear(m, PG_ZERO);
1725 vm_page_flag_set(m, PG_MAPPED);
1733 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1735 vm_pindex_t ptepindex;
1740 * Calculate pagetable page index
1742 ptepindex = pmap_pde_pindex(va);
1745 * Get the page directory entry
1747 pd = pmap_pde(pmap, va);
1750 * This supports switching from a 2MB page to a
1753 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1754 panic("no promotion/demotion yet");
1762 * If the page table page is mapped, we just increment the
1763 * hold count, and activate it.
1765 if (pd != NULL && (*pd & PG_V) != 0) {
1766 /* YYY hint is used here on i386 */
1767 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1768 pmap->pm_ptphint = m;
1773 * Here if the pte page isn't mapped, or if it has been deallocated.
1775 return _pmap_allocpte(pmap, ptepindex);
1779 /***************************************************
1780 * Pmap allocation/deallocation routines.
1781 ***************************************************/
1784 * Release any resources held by the given physical map.
1785 * Called when a pmap initialized by pmap_pinit is being released.
1786 * Should only be called if the map contains no valid mappings.
1788 static int pmap_release_callback(struct vm_page *p, void *data);
1791 pmap_release(struct pmap *pmap)
1793 vm_object_t object = pmap->pm_pteobj;
1794 struct rb_vm_page_scan_info info;
1796 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1797 #if defined(DIAGNOSTIC)
1798 if (object->ref_count != 1)
1799 panic("pmap_release: pteobj reference count != 1");
1803 info.object = object;
1805 lwkt_gettoken(&vm_token);
1806 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1813 info.limit = object->generation;
1815 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1816 pmap_release_callback, &info);
1817 if (info.error == 0 && info.mpte) {
1818 if (!pmap_release_free_page(pmap, info.mpte))
1822 } while (info.error);
1823 lwkt_reltoken(&vm_token);
1828 pmap_release_callback(struct vm_page *p, void *data)
1830 struct rb_vm_page_scan_info *info = data;
1832 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1836 if (!pmap_release_free_page(info->pmap, p)) {
1840 if (info->object->generation != info->limit) {
1848 * Grow the number of kernel page table entries, if needed.
1850 * This routine is always called to validate any address space
1851 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
1852 * space below KERNBASE.
1855 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1858 vm_offset_t ptppaddr;
1860 pd_entry_t *pde, newpdir;
1862 int update_kernel_vm_end;
1865 lwkt_gettoken(&vm_token);
1868 * bootstrap kernel_vm_end on first real VM use
1870 if (kernel_vm_end == 0) {
1871 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1873 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1874 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1875 ~(PAGE_SIZE * NPTEPG - 1);
1877 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1878 kernel_vm_end = kernel_map.max_offset;
1885 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
1886 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
1887 * do not want to force-fill 128G worth of page tables.
1889 if (kstart < KERNBASE) {
1890 if (kstart > kernel_vm_end)
1891 kstart = kernel_vm_end;
1892 KKASSERT(kend <= KERNBASE);
1893 update_kernel_vm_end = 1;
1895 update_kernel_vm_end = 0;
1898 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
1899 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
1901 if (kend - 1 >= kernel_map.max_offset)
1902 kend = kernel_map.max_offset;
1904 while (kstart < kend) {
1905 pde = pmap_pde(&kernel_pmap, kstart);
1907 /* We need a new PDP entry */
1908 nkpg = vm_page_alloc(kptobj, nkpt,
1911 VM_ALLOC_INTERRUPT);
1913 panic("pmap_growkernel: no memory to grow "
1916 paddr = VM_PAGE_TO_PHYS(nkpg);
1917 if ((nkpg->flags & PG_ZERO) == 0)
1918 pmap_zero_page(paddr);
1919 vm_page_flag_clear(nkpg, PG_ZERO);
1920 newpdp = (pdp_entry_t)
1921 (paddr | PG_V | PG_RW | PG_A | PG_M);
1922 *pmap_pdpe(&kernel_pmap, kstart) = newpdp;
1924 continue; /* try again */
1926 if ((*pde & PG_V) != 0) {
1927 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1928 ~(PAGE_SIZE * NPTEPG - 1);
1929 if (kstart - 1 >= kernel_map.max_offset) {
1930 kstart = kernel_map.max_offset;
1937 * This index is bogus, but out of the way
1939 nkpg = vm_page_alloc(kptobj, nkpt,
1942 VM_ALLOC_INTERRUPT);
1944 panic("pmap_growkernel: no memory to grow kernel");
1947 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1948 pmap_zero_page(ptppaddr);
1949 vm_page_flag_clear(nkpg, PG_ZERO);
1950 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1951 *pmap_pde(&kernel_pmap, kstart) = newpdir;
1954 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1955 ~(PAGE_SIZE * NPTEPG - 1);
1957 if (kstart - 1 >= kernel_map.max_offset) {
1958 kstart = kernel_map.max_offset;
1964 * Only update kernel_vm_end for areas below KERNBASE.
1966 if (update_kernel_vm_end && kernel_vm_end < kstart)
1967 kernel_vm_end = kstart;
1969 lwkt_reltoken(&vm_token);
1974 * Retire the given physical map from service.
1975 * Should only be called if the map contains
1976 * no valid mappings.
1979 pmap_destroy(pmap_t pmap)
1986 lwkt_gettoken(&vm_token);
1987 count = --pmap->pm_count;
1990 panic("destroying a pmap is not yet implemented");
1992 lwkt_reltoken(&vm_token);
1996 * Add a reference to the specified pmap.
1999 pmap_reference(pmap_t pmap)
2002 lwkt_gettoken(&vm_token);
2004 lwkt_reltoken(&vm_token);
2008 /***************************************************
2009 * page management routines.
2010 ***************************************************/
2013 * free the pv_entry back to the free list. This function may be
2014 * called from an interrupt.
2018 free_pv_entry(pv_entry_t pv)
2021 KKASSERT(pv_entry_count >= 0);
2026 * get a new pv_entry, allocating a block from the system
2027 * when needed. This function may be called from an interrupt.
2034 if (pv_entry_high_water &&
2035 (pv_entry_count > pv_entry_high_water) &&
2036 (pmap_pagedaemon_waken == 0)) {
2037 pmap_pagedaemon_waken = 1;
2038 wakeup(&vm_pages_needed);
2040 return zalloc(pvzone);
2044 * This routine is very drastic, but can save the system
2052 static int warningdone=0;
2054 if (pmap_pagedaemon_waken == 0)
2056 lwkt_gettoken(&vm_token);
2057 if (warningdone < 5) {
2058 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
2062 for(i = 0; i < vm_page_array_size; i++) {
2063 m = &vm_page_array[i];
2064 if (m->wire_count || m->hold_count || m->busy ||
2065 (m->flags & PG_BUSY))
2069 pmap_pagedaemon_waken = 0;
2070 lwkt_reltoken(&vm_token);
2075 * If it is the first entry on the list, it is actually
2076 * in the header and we must copy the following entry up
2077 * to the header. Otherwise we must search the list for
2078 * the entry. In either case we free the now unused entry.
2082 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
2083 vm_offset_t va, pmap_inval_info_t info)
2089 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
2090 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2091 if (pmap == pv->pv_pmap && va == pv->pv_va)
2095 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2096 if (va == pv->pv_va)
2104 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2105 m->md.pv_list_count--;
2106 m->object->agg_pv_list_count--;
2107 KKASSERT(m->md.pv_list_count >= 0);
2108 if (TAILQ_EMPTY(&m->md.pv_list))
2109 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2110 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2111 ++pmap->pm_generation;
2112 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
2120 * Create a pv entry for page at pa for
2125 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2130 pv = get_pv_entry();
2135 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2136 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2137 ++pmap->pm_generation;
2138 m->md.pv_list_count++;
2139 m->object->agg_pv_list_count++;
2145 * pmap_remove_pte: do the things to unmap a page in a process
2149 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2150 pmap_inval_info_t info)
2155 pmap_inval_interlock(info, pmap, va);
2156 oldpte = pte_load_clear(ptq);
2157 pmap_inval_deinterlock(info, pmap);
2159 pmap->pm_stats.wired_count -= 1;
2161 * Machines that don't support invlpg, also don't support
2162 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2166 cpu_invlpg((void *)va);
2167 KKASSERT(pmap->pm_stats.resident_count > 0);
2168 --pmap->pm_stats.resident_count;
2169 if (oldpte & PG_MANAGED) {
2170 m = PHYS_TO_VM_PAGE(oldpte);
2171 if (oldpte & PG_M) {
2172 #if defined(PMAP_DIAGNOSTIC)
2173 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2175 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2179 if (pmap_track_modified(va))
2183 vm_page_flag_set(m, PG_REFERENCED);
2184 return pmap_remove_entry(pmap, m, va, info);
2186 return pmap_unuse_pt(pmap, va, NULL, info);
2195 * Remove a single page from a process address space.
2197 * This function may not be called from an interrupt if the pmap is
2202 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2206 pte = pmap_pte(pmap, va);
2209 if ((*pte & PG_V) == 0)
2211 pmap_remove_pte(pmap, pte, va, info);
2217 * Remove the given range of addresses from the specified map.
2219 * It is assumed that the start and end are properly
2220 * rounded to the page size.
2222 * This function may not be called from an interrupt if the pmap is
2226 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2228 vm_offset_t va_next;
2229 pml4_entry_t *pml4e;
2231 pd_entry_t ptpaddr, *pde;
2233 struct pmap_inval_info info;
2238 lwkt_gettoken(&vm_token);
2239 if (pmap->pm_stats.resident_count == 0) {
2240 lwkt_reltoken(&vm_token);
2244 pmap_inval_init(&info);
2247 * special handling of removing one page. a very
2248 * common operation and easy to short circuit some
2251 if (sva + PAGE_SIZE == eva) {
2252 pde = pmap_pde(pmap, sva);
2253 if (pde && (*pde & PG_PS) == 0) {
2254 pmap_remove_page(pmap, sva, &info);
2255 pmap_inval_done(&info);
2256 lwkt_reltoken(&vm_token);
2261 for (; sva < eva; sva = va_next) {
2262 pml4e = pmap_pml4e(pmap, sva);
2263 if ((*pml4e & PG_V) == 0) {
2264 va_next = (sva + NBPML4) & ~PML4MASK;
2270 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2271 if ((*pdpe & PG_V) == 0) {
2272 va_next = (sva + NBPDP) & ~PDPMASK;
2279 * Calculate index for next page table.
2281 va_next = (sva + NBPDR) & ~PDRMASK;
2285 pde = pmap_pdpe_to_pde(pdpe, sva);
2289 * Weed out invalid mappings.
2295 * Check for large page.
2297 if ((ptpaddr & PG_PS) != 0) {
2298 /* JG FreeBSD has more complex treatment here */
2299 pmap_inval_interlock(&info, pmap, -1);
2301 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2302 pmap_inval_deinterlock(&info, pmap);
2307 * Limit our scan to either the end of the va represented
2308 * by the current page table page, or to the end of the
2309 * range being removed.
2315 * NOTE: pmap_remove_pte() can block.
2317 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2321 if (pmap_remove_pte(pmap, pte, sva, &info))
2325 pmap_inval_done(&info);
2326 lwkt_reltoken(&vm_token);
2332 * Removes this physical page from all physical maps in which it resides.
2333 * Reflects back modify bits to the pager.
2335 * This routine may not be called from an interrupt.
2340 pmap_remove_all(vm_page_t m)
2342 struct pmap_inval_info info;
2343 pt_entry_t *pte, tpte;
2346 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2349 lwkt_gettoken(&vm_token);
2350 pmap_inval_init(&info);
2352 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2353 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2354 --pv->pv_pmap->pm_stats.resident_count;
2356 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2357 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
2358 tpte = pte_load_clear(pte);
2360 pv->pv_pmap->pm_stats.wired_count--;
2361 pmap_inval_deinterlock(&info, pv->pv_pmap);
2363 vm_page_flag_set(m, PG_REFERENCED);
2366 * Update the vm_page_t clean and reference bits.
2369 #if defined(PMAP_DIAGNOSTIC)
2370 if (pmap_nw_modified(tpte)) {
2372 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2376 if (pmap_track_modified(pv->pv_va))
2379 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2380 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2381 ++pv->pv_pmap->pm_generation;
2382 m->md.pv_list_count--;
2383 m->object->agg_pv_list_count--;
2384 KKASSERT(m->md.pv_list_count >= 0);
2385 if (TAILQ_EMPTY(&m->md.pv_list))
2386 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2387 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2391 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2392 pmap_inval_done(&info);
2393 lwkt_reltoken(&vm_token);
2399 * Set the physical protection on the specified range of this map
2402 * This function may not be called from an interrupt if the map is
2403 * not the kernel_pmap.
2406 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2408 vm_offset_t va_next;
2409 pml4_entry_t *pml4e;
2411 pd_entry_t ptpaddr, *pde;
2413 pmap_inval_info info;
2415 /* JG review for NX */
2420 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2421 pmap_remove(pmap, sva, eva);
2425 if (prot & VM_PROT_WRITE)
2428 lwkt_gettoken(&vm_token);
2429 pmap_inval_init(&info);
2431 for (; sva < eva; sva = va_next) {
2433 pml4e = pmap_pml4e(pmap, sva);
2434 if ((*pml4e & PG_V) == 0) {
2435 va_next = (sva + NBPML4) & ~PML4MASK;
2441 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2442 if ((*pdpe & PG_V) == 0) {
2443 va_next = (sva + NBPDP) & ~PDPMASK;
2449 va_next = (sva + NBPDR) & ~PDRMASK;
2453 pde = pmap_pdpe_to_pde(pdpe, sva);
2457 * Check for large page.
2459 if ((ptpaddr & PG_PS) != 0) {
2460 pmap_inval_interlock(&info, pmap, -1);
2461 *pde &= ~(PG_M|PG_RW);
2462 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2463 pmap_inval_deinterlock(&info, pmap);
2468 * Weed out invalid mappings. Note: we assume that the page
2469 * directory table is always allocated, and in kernel virtual.
2477 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2484 * XXX non-optimal. Note also that there can be
2485 * no pmap_inval_flush() calls until after we modify
2486 * ptbase[sindex] (or otherwise we have to do another
2487 * pmap_inval_add() call).
2489 pmap_inval_interlock(&info, pmap, sva);
2493 if ((pbits & PG_V) == 0) {
2494 pmap_inval_deinterlock(&info, pmap);
2497 if (pbits & PG_MANAGED) {
2500 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2501 vm_page_flag_set(m, PG_REFERENCED);
2505 if (pmap_track_modified(sva)) {
2507 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2514 if (pbits != cbits &&
2515 !atomic_cmpset_long(pte, pbits, cbits)) {
2518 pmap_inval_deinterlock(&info, pmap);
2521 pmap_inval_done(&info);
2522 lwkt_reltoken(&vm_token);
2526 * Insert the given physical page (p) at
2527 * the specified virtual address (v) in the
2528 * target physical map with the protection requested.
2530 * If specified, the page will be wired down, meaning
2531 * that the related pte can not be reclaimed.
2533 * NB: This is the only routine which MAY NOT lazy-evaluate
2534 * or lose information. That is, this routine must actually
2535 * insert this page into the given map NOW.
2538 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2545 pt_entry_t origpte, newpte;
2547 pmap_inval_info info;
2552 va = trunc_page(va);
2553 #ifdef PMAP_DIAGNOSTIC
2555 panic("pmap_enter: toobig");
2556 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2557 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
2559 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2560 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2562 db_print_backtrace();
2565 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2566 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2568 db_print_backtrace();
2572 lwkt_gettoken(&vm_token);
2575 * In the case that a page table page is not
2576 * resident, we are creating it here.
2578 if (va < VM_MAX_USER_ADDRESS)
2579 mpte = pmap_allocpte(pmap, va);
2583 pmap_inval_init(&info);
2584 pde = pmap_pde(pmap, va);
2585 if (pde != NULL && (*pde & PG_V) != 0) {
2586 if ((*pde & PG_PS) != 0)
2587 panic("pmap_enter: attempted pmap_enter on 2MB page");
2588 pte = pmap_pde_to_pte(pde, va);
2590 panic("pmap_enter: invalid page directory va=%#lx", va);
2592 KKASSERT(pte != NULL);
2593 pa = VM_PAGE_TO_PHYS(m);
2595 opa = origpte & PG_FRAME;
2598 * Mapping has not changed, must be protection or wiring change.
2600 if (origpte && (opa == pa)) {
2602 * Wiring change, just update stats. We don't worry about
2603 * wiring PT pages as they remain resident as long as there
2604 * are valid mappings in them. Hence, if a user page is wired,
2605 * the PT page will be also.
2607 if (wired && ((origpte & PG_W) == 0))
2608 pmap->pm_stats.wired_count++;
2609 else if (!wired && (origpte & PG_W))
2610 pmap->pm_stats.wired_count--;
2612 #if defined(PMAP_DIAGNOSTIC)
2613 if (pmap_nw_modified(origpte)) {
2615 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2621 * Remove the extra pte reference. Note that we cannot
2622 * optimize the RO->RW case because we have adjusted the
2623 * wiring count above and may need to adjust the wiring
2630 * We might be turning off write access to the page,
2631 * so we go ahead and sense modify status.
2633 if (origpte & PG_MANAGED) {
2634 if ((origpte & PG_M) && pmap_track_modified(va)) {
2636 om = PHYS_TO_VM_PAGE(opa);
2640 KKASSERT(m->flags & PG_MAPPED);
2645 * Mapping has changed, invalidate old range and fall through to
2646 * handle validating new mapping.
2650 err = pmap_remove_pte(pmap, pte, va, &info);
2652 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2654 opa = origpte & PG_FRAME;
2656 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2662 * Enter on the PV list if part of our managed memory. Note that we
2663 * raise IPL while manipulating pv_table since pmap_enter can be
2664 * called at interrupt time.
2666 if (pmap_initialized &&
2667 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2668 pmap_insert_entry(pmap, va, mpte, m);
2670 vm_page_flag_set(m, PG_MAPPED);
2674 * Increment counters
2676 ++pmap->pm_stats.resident_count;
2678 pmap->pm_stats.wired_count++;
2682 * Now validate mapping with desired protection/wiring.
2684 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
2688 if (va < VM_MAX_USER_ADDRESS)
2690 if (pmap == &kernel_pmap)
2694 * if the mapping or permission bits are different, we need
2695 * to update the pte.
2697 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2698 pmap_inval_interlock(&info, pmap, va);
2699 *pte = newpte | PG_A;
2700 pmap_inval_deinterlock(&info, pmap);
2702 vm_page_flag_set(m, PG_WRITEABLE);
2704 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2705 pmap_inval_done(&info);
2706 lwkt_reltoken(&vm_token);
2710 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2711 * This code also assumes that the pmap has no pre-existing entry for this
2714 * This code currently may only be used on user pmaps, not kernel_pmap.
2717 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2722 vm_pindex_t ptepindex;
2724 pmap_inval_info info;
2726 lwkt_gettoken(&vm_token);
2727 pmap_inval_init(&info);
2729 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2730 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2732 db_print_backtrace();
2735 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2736 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2738 db_print_backtrace();
2742 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2745 * Calculate the page table page (mpte), allocating it if necessary.
2747 * A held page table page (mpte), or NULL, is passed onto the
2748 * section following.
2750 if (va < VM_MAX_USER_ADDRESS) {
2752 * Calculate pagetable page index
2754 ptepindex = pmap_pde_pindex(va);
2758 * Get the page directory entry
2760 ptepa = pmap_pde(pmap, va);
2763 * If the page table page is mapped, we just increment
2764 * the hold count, and activate it.
2766 if (ptepa && (*ptepa & PG_V) != 0) {
2768 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2769 // if (pmap->pm_ptphint &&
2770 // (pmap->pm_ptphint->pindex == ptepindex)) {
2771 // mpte = pmap->pm_ptphint;
2773 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2774 pmap->pm_ptphint = mpte;
2779 mpte = _pmap_allocpte(pmap, ptepindex);
2781 } while (mpte == NULL);
2784 /* this code path is not yet used */
2788 * With a valid (and held) page directory page, we can just use
2789 * vtopte() to get to the pte. If the pte is already present
2790 * we do not disturb it.
2795 pmap_unwire_pte_hold(pmap, va, mpte, &info);
2796 pa = VM_PAGE_TO_PHYS(m);
2797 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2798 pmap_inval_done(&info);
2799 lwkt_reltoken(&vm_token);
2804 * Enter on the PV list if part of our managed memory
2806 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2807 pmap_insert_entry(pmap, va, mpte, m);
2808 vm_page_flag_set(m, PG_MAPPED);
2812 * Increment counters
2814 ++pmap->pm_stats.resident_count;
2816 pa = VM_PAGE_TO_PHYS(m);
2819 * Now validate mapping with RO protection
2821 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2822 *pte = pa | PG_V | PG_U;
2824 *pte = pa | PG_V | PG_U | PG_MANAGED;
2825 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2826 pmap_inval_done(&info);
2827 lwkt_reltoken(&vm_token);
2831 * Make a temporary mapping for a physical address. This is only intended
2832 * to be used for panic dumps.
2834 /* JG Needed on x86_64? */
2836 pmap_kenter_temporary(vm_paddr_t pa, int i)
2838 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2839 return ((void *)crashdumpmap);
2842 #define MAX_INIT_PT (96)
2845 * This routine preloads the ptes for a given object into the specified pmap.
2846 * This eliminates the blast of soft faults on process startup and
2847 * immediately after an mmap.
2849 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2852 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2853 vm_object_t object, vm_pindex_t pindex,
2854 vm_size_t size, int limit)
2856 struct rb_vm_page_scan_info info;
2861 * We can't preinit if read access isn't set or there is no pmap
2864 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2868 * We can't preinit if the pmap is not the current pmap
2870 lp = curthread->td_lwp;
2871 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2874 psize = x86_64_btop(size);
2876 if ((object->type != OBJT_VNODE) ||
2877 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2878 (object->resident_page_count > MAX_INIT_PT))) {
2882 if (psize + pindex > object->size) {
2883 if (object->size < pindex)
2885 psize = object->size - pindex;
2892 * Use a red-black scan to traverse the requested range and load
2893 * any valid pages found into the pmap.
2895 * We cannot safely scan the object's memq unless we are in a
2896 * critical section since interrupts can remove pages from objects.
2898 info.start_pindex = pindex;
2899 info.end_pindex = pindex + psize - 1;
2906 lwkt_gettoken(&vm_token);
2907 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2908 pmap_object_init_pt_callback, &info);
2909 lwkt_reltoken(&vm_token);
2915 pmap_object_init_pt_callback(vm_page_t p, void *data)
2917 struct rb_vm_page_scan_info *info = data;
2918 vm_pindex_t rel_index;
2920 * don't allow an madvise to blow away our really
2921 * free pages allocating pv entries.
2923 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2924 vmstats.v_free_count < vmstats.v_free_reserved) {
2927 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2928 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2929 if ((p->queue - p->pc) == PQ_CACHE)
2930 vm_page_deactivate(p);
2932 rel_index = p->pindex - info->start_pindex;
2933 pmap_enter_quick(info->pmap,
2934 info->addr + x86_64_ptob(rel_index), p);
2941 * Return TRUE if the pmap is in shape to trivially
2942 * pre-fault the specified address.
2944 * Returns FALSE if it would be non-trivial or if a
2945 * pte is already loaded into the slot.
2948 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2954 lwkt_gettoken(&vm_token);
2955 pde = pmap_pde(pmap, addr);
2956 if (pde == NULL || *pde == 0) {
2960 ret = (*pte) ? 0 : 1;
2962 lwkt_reltoken(&vm_token);
2967 * Routine: pmap_change_wiring
2968 * Function: Change the wiring attribute for a map/virtual-address
2970 * In/out conditions:
2971 * The mapping must already exist in the pmap.
2974 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2981 lwkt_gettoken(&vm_token);
2982 pte = pmap_pte(pmap, va);
2984 if (wired && !pmap_pte_w(pte))
2985 pmap->pm_stats.wired_count++;
2986 else if (!wired && pmap_pte_w(pte))
2987 pmap->pm_stats.wired_count--;
2990 * Wiring is not a hardware characteristic so there is no need to
2991 * invalidate TLB. However, in an SMP environment we must use
2992 * a locked bus cycle to update the pte (if we are not using
2993 * the pmap_inval_*() API that is)... it's ok to do this for simple
2998 atomic_set_long(pte, PG_W);
3000 atomic_clear_long(pte, PG_W);
3003 atomic_set_long_nonlocked(pte, PG_W);
3005 atomic_clear_long_nonlocked(pte, PG_W);
3007 lwkt_reltoken(&vm_token);
3013 * Copy the range specified by src_addr/len
3014 * from the source map to the range dst_addr/len
3015 * in the destination map.
3017 * This routine is only advisory and need not do anything.
3020 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3021 vm_size_t len, vm_offset_t src_addr)
3025 pmap_inval_info info;
3027 vm_offset_t end_addr = src_addr + len;
3029 pd_entry_t src_frame, dst_frame;
3032 if (dst_addr != src_addr)
3035 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3036 if (src_frame != (PTDpde & PG_FRAME)) {
3040 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3041 if (dst_frame != (APTDpde & PG_FRAME)) {
3042 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
3043 /* The page directory is not shared between CPUs */
3047 pmap_inval_init(&info);
3048 pmap_inval_add(&info, dst_pmap, -1);
3049 pmap_inval_add(&info, src_pmap, -1);
3052 * critical section protection is required to maintain the page/object
3053 * association, interrupts can free pages and remove them from
3057 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
3058 pt_entry_t *src_pte, *dst_pte;
3059 vm_page_t dstmpte, srcmpte;
3060 vm_offset_t srcptepaddr;
3061 vm_pindex_t ptepindex;
3063 if (addr >= UPT_MIN_ADDRESS)
3064 panic("pmap_copy: invalid to pmap_copy page tables\n");
3067 * Don't let optional prefaulting of pages make us go
3068 * way below the low water mark of free pages or way
3069 * above high water mark of used pv entries.
3071 if (vmstats.v_free_count < vmstats.v_free_reserved ||
3072 pv_entry_count > pv_entry_high_water)
3075 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
3076 ptepindex = addr >> PDRSHIFT;
3079 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
3081 if (srcptepaddr == 0)
3084 if (srcptepaddr & PG_PS) {
3086 if (dst_pmap->pm_pdir[ptepindex] == 0) {
3087 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
3088 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
3094 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
3095 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
3096 (srcmpte->flags & PG_BUSY)) {
3100 if (pdnxt > end_addr)
3103 src_pte = vtopte(addr);
3105 dst_pte = avtopte(addr);
3107 while (addr < pdnxt) {
3112 * we only virtual copy managed pages
3114 if ((ptetemp & PG_MANAGED) != 0) {
3116 * We have to check after allocpte for the
3117 * pte still being around... allocpte can
3120 * pmap_allocpte() can block. If we lose
3121 * our page directory mappings we stop.
3123 dstmpte = pmap_allocpte(dst_pmap, addr);
3126 if (src_frame != (PTDpde & PG_FRAME) ||
3127 dst_frame != (APTDpde & PG_FRAME)
3129 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3130 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3132 } else if ((*dst_pte == 0) &&
3133 (ptetemp = *src_pte) != 0 &&
3134 (ptetemp & PG_MANAGED)) {
3136 * Clear the modified and
3137 * accessed (referenced) bits
3140 m = PHYS_TO_VM_PAGE(ptetemp);
3141 *dst_pte = ptetemp & ~(PG_M | PG_A);
3142 ++dst_pmap->pm_stats.resident_count;
3143 pmap_insert_entry(dst_pmap, addr,
3145 KKASSERT(m->flags & PG_MAPPED);
3147 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3148 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3152 if (dstmpte->hold_count >= srcmpte->hold_count)
3162 pmap_inval_done(&info);
3169 * Zero the specified physical page.
3171 * This function may be called from an interrupt and no locking is
3175 pmap_zero_page(vm_paddr_t phys)
3177 vm_offset_t va = PHYS_TO_DMAP(phys);
3179 pagezero((void *)va);
3183 * pmap_page_assertzero:
3185 * Assert that a page is empty, panic if it isn't.
3188 pmap_page_assertzero(vm_paddr_t phys)
3190 vm_offset_t virt = PHYS_TO_DMAP(phys);
3193 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3194 if (*(long *)((char *)virt + i) != 0) {
3195 panic("pmap_page_assertzero() @ %p not zero!\n", (void *)virt);
3203 * Zero part of a physical page by mapping it into memory and clearing
3204 * its contents with bzero.
3206 * off and size may not cover an area beyond a single hardware page.
3209 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3211 vm_offset_t virt = PHYS_TO_DMAP(phys);
3213 bzero((char *)virt + off, size);
3219 * Copy the physical page from the source PA to the target PA.
3220 * This function may be called from an interrupt. No locking
3224 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3226 vm_offset_t src_virt, dst_virt;
3228 src_virt = PHYS_TO_DMAP(src);
3229 dst_virt = PHYS_TO_DMAP(dst);
3230 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3234 * pmap_copy_page_frag:
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_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3243 vm_offset_t src_virt, dst_virt;
3245 src_virt = PHYS_TO_DMAP(src);
3246 dst_virt = PHYS_TO_DMAP(dst);
3248 bcopy((char *)src_virt + (src & PAGE_MASK),
3249 (char *)dst_virt + (dst & PAGE_MASK),
3254 * Returns true if the pmap's pv is one of the first
3255 * 16 pvs linked to from this page. This count may
3256 * be changed upwards or downwards in the future; it
3257 * is only necessary that true be returned for a small
3258 * subset of pmaps for proper page aging.
3261 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3266 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3270 lwkt_gettoken(&vm_token);
3272 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3273 if (pv->pv_pmap == pmap) {
3274 lwkt_reltoken(&vm_token);
3282 lwkt_reltoken(&vm_token);
3288 * Remove all pages from specified address space
3289 * this aids process exit speeds. Also, this code
3290 * is special cased for current process only, but
3291 * can have the more generic (and slightly slower)
3292 * mode enabled. This is much faster than pmap_remove
3293 * in the case of running down an entire address space.
3296 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3299 pt_entry_t *pte, tpte;
3302 pmap_inval_info info;
3304 int save_generation;
3306 lp = curthread->td_lwp;
3307 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3312 lwkt_gettoken(&vm_token);
3313 pmap_inval_init(&info);
3314 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3315 if (pv->pv_va >= eva || pv->pv_va < sva) {
3316 npv = TAILQ_NEXT(pv, pv_plist);
3320 KKASSERT(pmap == pv->pv_pmap);
3323 pte = vtopte(pv->pv_va);
3325 pte = pmap_pte_quick(pmap, pv->pv_va);
3326 pmap_inval_interlock(&info, pmap, pv->pv_va);
3329 * We cannot remove wired pages from a process' mapping
3333 pmap_inval_deinterlock(&info, pmap);
3334 npv = TAILQ_NEXT(pv, pv_plist);
3337 tpte = pte_load_clear(pte);
3339 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3341 KASSERT(m < &vm_page_array[vm_page_array_size],
3342 ("pmap_remove_pages: bad tpte %lx", tpte));
3344 KKASSERT(pmap->pm_stats.resident_count > 0);
3345 --pmap->pm_stats.resident_count;
3346 pmap_inval_deinterlock(&info, pmap);
3349 * Update the vm_page_t clean and reference bits.
3355 npv = TAILQ_NEXT(pv, pv_plist);
3356 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3357 save_generation = ++pmap->pm_generation;
3359 m->md.pv_list_count--;
3360 m->object->agg_pv_list_count--;
3361 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3362 if (TAILQ_EMPTY(&m->md.pv_list))
3363 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3365 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3369 * Restart the scan if we blocked during the unuse or free
3370 * calls and other removals were made.
3372 if (save_generation != pmap->pm_generation) {
3373 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3374 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3377 pmap_inval_done(&info);
3378 lwkt_reltoken(&vm_token);
3382 * pmap_testbit tests bits in pte's
3383 * note that the testbit/clearbit routines are inline,
3384 * and a lot of things compile-time evaluate.
3388 pmap_testbit(vm_page_t m, int bit)
3393 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3396 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3401 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3403 * if the bit being tested is the modified bit, then
3404 * mark clean_map and ptes as never
3407 if (bit & (PG_A|PG_M)) {
3408 if (!pmap_track_modified(pv->pv_va))
3412 #if defined(PMAP_DIAGNOSTIC)
3413 if (pv->pv_pmap == NULL) {
3414 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3418 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3429 * this routine is used to modify bits in ptes
3433 pmap_clearbit(vm_page_t m, int bit)
3435 struct pmap_inval_info info;
3440 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3443 pmap_inval_init(&info);
3446 * Loop over all current mappings setting/clearing as appropos If
3447 * setting RO do we need to clear the VAC?
3449 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3451 * don't write protect pager mappings
3454 if (!pmap_track_modified(pv->pv_va))
3458 #if defined(PMAP_DIAGNOSTIC)
3459 if (pv->pv_pmap == NULL) {
3460 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3466 * Careful here. We can use a locked bus instruction to
3467 * clear PG_A or PG_M safely but we need to synchronize
3468 * with the target cpus when we mess with PG_RW.
3470 * We do not have to force synchronization when clearing
3471 * PG_M even for PTEs generated via virtual memory maps,
3472 * because the virtual kernel will invalidate the pmap
3473 * entry when/if it needs to resynchronize the Modify bit.
3476 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
3477 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3484 atomic_clear_long(pte, PG_M|PG_RW);
3487 * The cpu may be trying to set PG_M
3488 * simultaniously with our clearing
3491 if (!atomic_cmpset_long(pte, pbits,
3495 } else if (bit == PG_M) {
3497 * We could also clear PG_RW here to force
3498 * a fault on write to redetect PG_M for
3499 * virtual kernels, but it isn't necessary
3500 * since virtual kernels invalidate the pte
3501 * when they clear the VPTE_M bit in their
3502 * virtual page tables.
3504 atomic_clear_long(pte, PG_M);
3506 atomic_clear_long(pte, bit);
3510 pmap_inval_deinterlock(&info, pv->pv_pmap);
3512 pmap_inval_done(&info);
3516 * pmap_page_protect:
3518 * Lower the permission for all mappings to a given page.
3521 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3523 /* JG NX support? */
3524 if ((prot & VM_PROT_WRITE) == 0) {
3525 lwkt_gettoken(&vm_token);
3526 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3527 pmap_clearbit(m, PG_RW);
3528 vm_page_flag_clear(m, PG_WRITEABLE);
3532 lwkt_reltoken(&vm_token);
3537 pmap_phys_address(vm_pindex_t ppn)
3539 return (x86_64_ptob(ppn));
3543 * pmap_ts_referenced:
3545 * Return a count of reference bits for a page, clearing those bits.
3546 * It is not necessary for every reference bit to be cleared, but it
3547 * is necessary that 0 only be returned when there are truly no
3548 * reference bits set.
3550 * XXX: The exact number of bits to check and clear is a matter that
3551 * should be tested and standardized at some point in the future for
3552 * optimal aging of shared pages.
3555 pmap_ts_referenced(vm_page_t m)
3557 pv_entry_t pv, pvf, pvn;
3561 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3565 lwkt_gettoken(&vm_token);
3567 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3572 pvn = TAILQ_NEXT(pv, pv_list);
3575 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3576 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3579 if (!pmap_track_modified(pv->pv_va))
3582 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3584 if (pte && (*pte & PG_A)) {
3586 atomic_clear_long(pte, PG_A);
3588 atomic_clear_long_nonlocked(pte, PG_A);
3595 } while ((pv = pvn) != NULL && pv != pvf);
3597 lwkt_reltoken(&vm_token);
3606 * Return whether or not the specified physical page was modified
3607 * in any physical maps.
3610 pmap_is_modified(vm_page_t m)
3614 lwkt_gettoken(&vm_token);
3615 res = pmap_testbit(m, PG_M);
3616 lwkt_reltoken(&vm_token);
3621 * Clear the modify bits on the specified physical page.
3624 pmap_clear_modify(vm_page_t m)
3626 lwkt_gettoken(&vm_token);
3627 pmap_clearbit(m, PG_M);
3628 lwkt_reltoken(&vm_token);
3632 * pmap_clear_reference:
3634 * Clear the reference bit on the specified physical page.
3637 pmap_clear_reference(vm_page_t m)
3639 lwkt_gettoken(&vm_token);
3640 pmap_clearbit(m, PG_A);
3641 lwkt_reltoken(&vm_token);
3645 * Miscellaneous support routines follow
3650 i386_protection_init(void)
3654 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3655 kp = protection_codes;
3656 for (prot = 0; prot < 8; prot++) {
3658 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3660 * Read access is also 0. There isn't any execute bit,
3661 * so just make it readable.
3663 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3664 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3665 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3668 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3669 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3670 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3671 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3679 * Map a set of physical memory pages into the kernel virtual
3680 * address space. Return a pointer to where it is mapped. This
3681 * routine is intended to be used for mapping device memory,
3684 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3688 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3690 vm_offset_t va, tmpva, offset;
3693 offset = pa & PAGE_MASK;
3694 size = roundup(offset + size, PAGE_SIZE);
3696 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3698 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3700 pa = pa & ~PAGE_MASK;
3701 for (tmpva = va; size > 0;) {
3702 pte = vtopte(tmpva);
3703 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3711 return ((void *)(va + offset));
3715 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3717 vm_offset_t va, tmpva, offset;
3720 offset = pa & PAGE_MASK;
3721 size = roundup(offset + size, PAGE_SIZE);
3723 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3725 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3727 pa = pa & ~PAGE_MASK;
3728 for (tmpva = va; size > 0;) {
3729 pte = vtopte(tmpva);
3730 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3738 return ((void *)(va + offset));
3742 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3744 vm_offset_t base, offset;
3746 base = va & ~PAGE_MASK;
3747 offset = va & PAGE_MASK;
3748 size = roundup(offset + size, PAGE_SIZE);
3749 pmap_qremove(va, size >> PAGE_SHIFT);
3750 kmem_free(&kernel_map, base, size);
3754 * perform the pmap work for mincore
3757 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3759 pt_entry_t *ptep, pte;
3763 lwkt_gettoken(&vm_token);
3764 ptep = pmap_pte(pmap, addr);
3766 if (ptep && (pte = *ptep) != 0) {
3769 val = MINCORE_INCORE;
3770 if ((pte & PG_MANAGED) == 0)
3773 pa = pte & PG_FRAME;
3775 m = PHYS_TO_VM_PAGE(pa);
3781 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3783 * Modified by someone
3785 else if (m->dirty || pmap_is_modified(m))
3786 val |= MINCORE_MODIFIED_OTHER;
3791 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3794 * Referenced by someone
3796 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3797 val |= MINCORE_REFERENCED_OTHER;
3798 vm_page_flag_set(m, PG_REFERENCED);
3802 lwkt_reltoken(&vm_token);
3807 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3808 * vmspace will be ref'd and the old one will be deref'd.
3810 * The vmspace for all lwps associated with the process will be adjusted
3811 * and cr3 will be reloaded if any lwp is the current lwp.
3814 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3816 struct vmspace *oldvm;
3820 oldvm = p->p_vmspace;
3821 if (oldvm != newvm) {
3822 p->p_vmspace = newvm;
3823 KKASSERT(p->p_nthreads == 1);
3824 lp = RB_ROOT(&p->p_lwp_tree);
3825 pmap_setlwpvm(lp, newvm);
3827 sysref_get(&newvm->vm_sysref);
3828 sysref_put(&oldvm->vm_sysref);
3835 * Set the vmspace for a LWP. The vmspace is almost universally set the
3836 * same as the process vmspace, but virtual kernels need to swap out contexts
3837 * on a per-lwp basis.
3840 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3842 struct vmspace *oldvm;
3846 oldvm = lp->lwp_vmspace;
3848 if (oldvm != newvm) {
3849 lp->lwp_vmspace = newvm;
3850 if (curthread->td_lwp == lp) {
3851 pmap = vmspace_pmap(newvm);
3853 atomic_set_int(&pmap->pm_active, mycpu->gd_cpumask);
3854 if (pmap->pm_active & CPUMASK_LOCK)
3855 pmap_interlock_wait(newvm);
3857 pmap->pm_active |= 1;
3859 #if defined(SWTCH_OPTIM_STATS)
3862 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
3863 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
3864 load_cr3(curthread->td_pcb->pcb_cr3);
3865 pmap = vmspace_pmap(oldvm);
3867 atomic_clear_int(&pmap->pm_active, mycpu->gd_cpumask);
3869 pmap->pm_active &= ~1;
3879 * Called when switching to a locked pmap
3882 pmap_interlock_wait(struct vmspace *vm)
3884 struct pmap *pmap = &vm->vm_pmap;
3886 if (pmap->pm_active & CPUMASK_LOCK) {
3887 while (pmap->pm_active & CPUMASK_LOCK) {
3890 lwkt_process_ipiq();
3898 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3901 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3905 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3910 * Used by kmalloc/kfree, page already exists at va
3913 pmap_kvtom(vm_offset_t va)
3915 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));