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;
208 static pv_entry_t get_pv_entry (void);
209 static void i386_protection_init (void);
210 static void create_pagetables(vm_paddr_t *firstaddr);
211 static void pmap_remove_all (vm_page_t m);
212 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
213 vm_offset_t sva, pmap_inval_info_t info);
214 static void pmap_remove_page (struct pmap *pmap,
215 vm_offset_t va, pmap_inval_info_t info);
216 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
217 vm_offset_t va, pmap_inval_info_t info);
218 static boolean_t pmap_testbit (vm_page_t m, int bit);
219 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
220 vm_page_t mpte, vm_page_t m);
222 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
224 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
225 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
226 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
227 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
228 static int _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
229 pmap_inval_info_t info);
230 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
231 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
233 static unsigned pdir4mb;
236 * Move the kernel virtual free pointer to the next
237 * 2MB. This is used to help improve performance
238 * by using a large (2MB) page for much of the kernel
239 * (.text, .data, .bss)
243 pmap_kmem_choose(vm_offset_t addr)
245 vm_offset_t newaddr = addr;
247 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
254 * Super fast pmap_pte routine best used when scanning the pv lists.
255 * This eliminates many course-grained invltlb calls. Note that many of
256 * the pv list scans are across different pmaps and it is very wasteful
257 * to do an entire invltlb when checking a single mapping.
259 * Should only be called while in a critical section.
261 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
265 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
267 return pmap_pte(pmap, va);
270 /* Return a non-clipped PD index for a given VA */
273 pmap_pde_pindex(vm_offset_t va)
275 return va >> PDRSHIFT;
278 /* Return various clipped indexes for a given VA */
281 pmap_pte_index(vm_offset_t va)
284 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
289 pmap_pde_index(vm_offset_t va)
292 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
297 pmap_pdpe_index(vm_offset_t va)
300 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
305 pmap_pml4e_index(vm_offset_t va)
308 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
311 /* Return a pointer to the PML4 slot that corresponds to a VA */
314 pmap_pml4e(pmap_t pmap, vm_offset_t va)
317 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
320 /* Return a pointer to the PDP slot that corresponds to a VA */
323 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
327 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
328 return (&pdpe[pmap_pdpe_index(va)]);
331 /* Return a pointer to the PDP slot that corresponds to a VA */
334 pmap_pdpe(pmap_t pmap, vm_offset_t va)
338 pml4e = pmap_pml4e(pmap, va);
339 if ((*pml4e & PG_V) == 0)
341 return (pmap_pml4e_to_pdpe(pml4e, va));
344 /* Return a pointer to the PD slot that corresponds to a VA */
347 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
351 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
352 return (&pde[pmap_pde_index(va)]);
355 /* Return a pointer to the PD slot that corresponds to a VA */
358 pmap_pde(pmap_t pmap, vm_offset_t va)
362 pdpe = pmap_pdpe(pmap, va);
363 if (pdpe == NULL || (*pdpe & PG_V) == 0)
365 return (pmap_pdpe_to_pde(pdpe, va));
368 /* Return a pointer to the PT slot that corresponds to a VA */
371 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
375 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
376 return (&pte[pmap_pte_index(va)]);
379 /* Return a pointer to the PT slot that corresponds to a VA */
382 pmap_pte(pmap_t pmap, vm_offset_t va)
386 pde = pmap_pde(pmap, va);
387 if (pde == NULL || (*pde & PG_V) == 0)
389 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
390 return ((pt_entry_t *)pde);
391 return (pmap_pde_to_pte(pde, va));
396 vtopte(vm_offset_t va)
398 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
400 return (PTmap + ((va >> PAGE_SHIFT) & mask));
405 vtopde(vm_offset_t va)
407 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
409 return (PDmap + ((va >> PDRSHIFT) & mask));
413 allocpages(vm_paddr_t *firstaddr, long n)
418 bzero((void *)ret, n * PAGE_SIZE);
419 *firstaddr += n * PAGE_SIZE;
425 create_pagetables(vm_paddr_t *firstaddr)
427 long i; /* must be 64 bits */
432 * We are running (mostly) V=P at this point
434 * Calculate NKPT - number of kernel page tables. We have to
435 * accomodoate prealloction of the vm_page_array, dump bitmap,
436 * MSGBUF_SIZE, and other stuff. Be generous.
438 * Maxmem is in pages.
440 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
441 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
445 * Starting at the beginning of kvm (not KERNBASE).
447 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
448 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
449 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E + ndmpdp) +
454 * Starting at KERNBASE - map 2G worth of page table pages.
455 * KERNBASE is offset -2G from the end of kvm.
457 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
462 KPTbase = allocpages(firstaddr, nkpt_base);
463 KPTphys = allocpages(firstaddr, nkpt_phys);
464 KPML4phys = allocpages(firstaddr, 1);
465 KPDPphys = allocpages(firstaddr, NKPML4E);
466 KPDphys = allocpages(firstaddr, NKPDPE);
469 * Calculate the page directory base for KERNBASE,
470 * that is where we start populating the page table pages.
471 * Basically this is the end - 2.
473 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
475 DMPDPphys = allocpages(firstaddr, NDMPML4E);
476 if ((amd_feature & AMDID_PAGE1GB) == 0)
477 DMPDphys = allocpages(firstaddr, ndmpdp);
478 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
481 * Fill in the underlying page table pages for the area around
482 * KERNBASE. This remaps low physical memory to KERNBASE.
484 * Read-only from zero to physfree
485 * XXX not fully used, underneath 2M pages
487 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
488 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
489 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
493 * Now map the initial kernel page tables. One block of page
494 * tables is placed at the beginning of kernel virtual memory,
495 * and another block is placed at KERNBASE to map the kernel binary,
496 * data, bss, and initial pre-allocations.
498 for (i = 0; i < nkpt_base; i++) {
499 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
500 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
502 for (i = 0; i < nkpt_phys; i++) {
503 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
504 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
508 * Map from zero to end of allocations using 2M pages as an
509 * optimization. This will bypass some of the KPTBase pages
510 * above in the KERNBASE area.
512 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
513 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
514 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
518 * And connect up the PD to the PDP. The kernel pmap is expected
519 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
521 for (i = 0; i < NKPDPE; i++) {
522 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
523 KPDphys + (i << PAGE_SHIFT);
524 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
528 /* Now set up the direct map space using either 2MB or 1GB pages */
529 /* Preset PG_M and PG_A because demotion expects it */
530 if ((amd_feature & AMDID_PAGE1GB) == 0) {
531 for (i = 0; i < NPDEPG * ndmpdp; i++) {
532 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
533 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
536 /* And the direct map space's PDP */
537 for (i = 0; i < ndmpdp; i++) {
538 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
540 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
543 for (i = 0; i < ndmpdp; i++) {
544 ((pdp_entry_t *)DMPDPphys)[i] =
545 (vm_paddr_t)i << PDPSHIFT;
546 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
551 /* And recursively map PML4 to itself in order to get PTmap */
552 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
553 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
555 /* Connect the Direct Map slot up to the PML4 */
556 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
557 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
559 /* Connect the KVA slot up to the PML4 */
560 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
561 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
565 * Bootstrap the system enough to run with virtual memory.
567 * On the i386 this is called after mapping has already been enabled
568 * and just syncs the pmap module with what has already been done.
569 * [We can't call it easily with mapping off since the kernel is not
570 * mapped with PA == VA, hence we would have to relocate every address
571 * from the linked base (virtual) address "KERNBASE" to the actual
572 * (physical) address starting relative to 0]
575 pmap_bootstrap(vm_paddr_t *firstaddr)
579 struct mdglobaldata *gd;
582 KvaStart = VM_MIN_KERNEL_ADDRESS;
583 KvaEnd = VM_MAX_KERNEL_ADDRESS;
584 KvaSize = KvaEnd - KvaStart;
586 avail_start = *firstaddr;
589 * Create an initial set of page tables to run the kernel in.
591 create_pagetables(firstaddr);
593 virtual2_start = KvaStart;
594 virtual2_end = PTOV_OFFSET;
596 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
597 virtual_start = pmap_kmem_choose(virtual_start);
599 virtual_end = VM_MAX_KERNEL_ADDRESS;
601 /* XXX do %cr0 as well */
602 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
606 * Initialize protection array.
608 i386_protection_init();
611 * The kernel's pmap is statically allocated so we don't have to use
612 * pmap_create, which is unlikely to work correctly at this part of
613 * the boot sequence (XXX and which no longer exists).
615 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
616 kernel_pmap.pm_count = 1;
617 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
618 TAILQ_INIT(&kernel_pmap.pm_pvlist);
621 * Reserve some special page table entries/VA space for temporary
624 #define SYSMAP(c, p, v, n) \
625 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
631 * CMAP1/CMAP2 are used for zeroing and copying pages.
633 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
638 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
641 * ptvmmap is used for reading arbitrary physical pages via
644 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
647 * msgbufp is used to map the system message buffer.
648 * XXX msgbufmap is not used.
650 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
651 atop(round_page(MSGBUF_SIZE)))
658 * PG_G is terribly broken on SMP because we IPI invltlb's in some
659 * cases rather then invl1pg. Actually, I don't even know why it
660 * works under UP because self-referential page table mappings
665 if (cpu_feature & CPUID_PGE)
670 * Initialize the 4MB page size flag
674 * The 4MB page version of the initial
675 * kernel page mapping.
679 #if !defined(DISABLE_PSE)
680 if (cpu_feature & CPUID_PSE) {
683 * Note that we have enabled PSE mode
686 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
687 ptditmp &= ~(NBPDR - 1);
688 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
693 * Enable the PSE mode. If we are SMP we can't do this
694 * now because the APs will not be able to use it when
697 load_cr4(rcr4() | CR4_PSE);
700 * We can do the mapping here for the single processor
701 * case. We simply ignore the old page table page from
705 * For SMP, we still need 4K pages to bootstrap APs,
706 * PSE will be enabled as soon as all APs are up.
708 PTD[KPTDI] = (pd_entry_t)ptditmp;
715 * We need to finish setting up the globaldata page for the BSP.
716 * locore has already populated the page table for the mdglobaldata
719 pg = MDGLOBALDATA_BASEALLOC_PAGES;
720 gd = &CPU_prvspace[0].mdglobaldata;
727 * Set 4mb pdir for mp startup
732 if (pseflag && (cpu_feature & CPUID_PSE)) {
733 load_cr4(rcr4() | CR4_PSE);
734 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
742 * XXX: Hack. Required by pmap_init()
744 extern vm_offset_t cpu_apic_addr;
747 * Initialize the pmap module.
748 * Called by vm_init, to initialize any structures that the pmap
749 * system needs to map virtual memory.
750 * pmap_init has been enhanced to support in a fairly consistant
751 * way, discontiguous physical memory.
760 * object for kernel page table pages
762 /* JG I think the number can be arbitrary */
763 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
766 * Allocate memory for random pmap data structures. Includes the
770 for(i = 0; i < vm_page_array_size; i++) {
773 m = &vm_page_array[i];
774 TAILQ_INIT(&m->md.pv_list);
775 m->md.pv_list_count = 0;
779 * init the pv free list
781 initial_pvs = vm_page_array_size;
782 if (initial_pvs < MINPV)
784 pvzone = &pvzone_store;
785 pvinit = (void *)kmem_alloc(&kernel_map,
786 initial_pvs * sizeof (struct pv_entry));
787 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
788 pvinit, initial_pvs);
791 * Now it is safe to enable pv_table recording.
793 pmap_initialized = TRUE;
798 lapic = pmap_mapdev_uncacheable(cpu_apic_addr, sizeof(struct LAPIC));
803 * Initialize the address space (zone) for the pv_entries. Set a
804 * high water mark so that the system can recover from excessive
805 * numbers of pv entries.
810 int shpgperproc = PMAP_SHPGPERPROC;
813 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
814 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
815 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
816 pv_entry_high_water = 9 * (pv_entry_max / 10);
819 * Subtract out pages already installed in the zone (hack)
821 entry_max = pv_entry_max - vm_page_array_size;
825 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
829 /***************************************************
830 * Low level helper routines.....
831 ***************************************************/
833 #if defined(PMAP_DIAGNOSTIC)
836 * This code checks for non-writeable/modified pages.
837 * This should be an invalid condition.
841 pmap_nw_modified(pt_entry_t pte)
843 if ((pte & (PG_M|PG_RW)) == PG_M)
852 * this routine defines the region(s) of memory that should
853 * not be tested for the modified bit.
857 pmap_track_modified(vm_offset_t va)
859 if ((va < clean_sva) || (va >= clean_eva))
866 * Extract the physical page address associated with the map/VA pair.
868 * The caller must hold vm_token if non-blocking operation is desired.
871 pmap_extract(pmap_t pmap, vm_offset_t va)
875 pd_entry_t pde, *pdep;
877 lwkt_gettoken(&vm_token);
879 pdep = pmap_pde(pmap, va);
883 if ((pde & PG_PS) != 0) {
884 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
886 pte = pmap_pde_to_pte(pdep, va);
887 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
891 lwkt_reltoken(&vm_token);
896 * Extract the physical page address associated kernel virtual address.
899 pmap_kextract(vm_offset_t va)
904 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
905 pa = DMAP_TO_PHYS(va);
909 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
912 * Beware of a concurrent promotion that changes the
913 * PDE at this point! For example, vtopte() must not
914 * be used to access the PTE because it would use the
915 * new PDE. It is, however, safe to use the old PDE
916 * because the page table page is preserved by the
919 pa = *pmap_pde_to_pte(&pde, va);
920 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
926 /***************************************************
927 * Low level mapping routines.....
928 ***************************************************/
931 * Routine: pmap_kenter
933 * Add a wired page to the KVA
934 * NOTE! note that in order for the mapping to take effect -- you
935 * should do an invltlb after doing the pmap_kenter().
938 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
942 pmap_inval_info info;
944 pmap_inval_init(&info);
945 npte = pa | PG_RW | PG_V | pgeflag;
947 pmap_inval_interlock(&info, &kernel_pmap, va);
949 pmap_inval_deinterlock(&info, &kernel_pmap);
950 pmap_inval_done(&info);
954 * Routine: pmap_kenter_quick
956 * Similar to pmap_kenter(), except we only invalidate the
957 * mapping on the current CPU.
960 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
965 npte = pa | PG_RW | PG_V | pgeflag;
968 cpu_invlpg((void *)va);
972 pmap_kenter_sync(vm_offset_t va)
974 pmap_inval_info info;
976 pmap_inval_init(&info);
977 pmap_inval_interlock(&info, &kernel_pmap, va);
978 pmap_inval_deinterlock(&info, &kernel_pmap);
979 pmap_inval_done(&info);
983 pmap_kenter_sync_quick(vm_offset_t va)
985 cpu_invlpg((void *)va);
989 * remove a page from the kernel pagetables
992 pmap_kremove(vm_offset_t va)
995 pmap_inval_info info;
997 pmap_inval_init(&info);
999 pmap_inval_interlock(&info, &kernel_pmap, va);
1001 pmap_inval_deinterlock(&info, &kernel_pmap);
1002 pmap_inval_done(&info);
1006 pmap_kremove_quick(vm_offset_t va)
1011 cpu_invlpg((void *)va);
1015 * XXX these need to be recoded. They are not used in any critical path.
1018 pmap_kmodify_rw(vm_offset_t va)
1020 *vtopte(va) |= PG_RW;
1021 cpu_invlpg((void *)va);
1025 pmap_kmodify_nc(vm_offset_t va)
1027 *vtopte(va) |= PG_N;
1028 cpu_invlpg((void *)va);
1032 * Used to map a range of physical addresses into kernel virtual
1033 * address space during the low level boot, typically to map the
1034 * dump bitmap, message buffer, and vm_page_array.
1036 * These mappings are typically made at some pointer after the end of the
1039 * We could return PHYS_TO_DMAP(start) here and not allocate any
1040 * via (*virtp), but then kmem from userland and kernel dumps won't
1041 * have access to the related pointers.
1044 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1047 vm_offset_t va_start;
1049 /*return PHYS_TO_DMAP(start);*/
1054 while (start < end) {
1055 pmap_kenter_quick(va, start);
1065 * Add a list of wired pages to the kva
1066 * this routine is only used for temporary
1067 * kernel mappings that do not need to have
1068 * page modification or references recorded.
1069 * Note that old mappings are simply written
1070 * over. The page *must* be wired.
1073 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1077 end_va = va + count * PAGE_SIZE;
1079 while (va < end_va) {
1083 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1084 cpu_invlpg((void *)va);
1092 * This routine jerks page mappings from the
1093 * kernel -- it is meant only for temporary mappings.
1095 * MPSAFE, INTERRUPT SAFE (cluster callback)
1098 pmap_qremove(vm_offset_t va, int count)
1102 end_va = va + count * PAGE_SIZE;
1104 while (va < end_va) {
1109 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 & alignment */
1144 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1145 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
1146 td->td_savefpu = &td->td_pcb->pcb_save;
1147 td->td_sp = (char *)td->td_pcb; /* no -16 */
1151 * This routine directly affects the fork perf for a process.
1154 pmap_init_proc(struct proc *p)
1159 * Dispose the UPAGES for a process that has exited.
1160 * This routine directly impacts the exit perf of a process.
1163 pmap_dispose_proc(struct proc *p)
1165 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1168 /***************************************************
1169 * Page table page management routines.....
1170 ***************************************************/
1173 * This routine unholds page table pages, and if the hold count
1174 * drops to zero, then it decrements the wire count.
1178 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1179 pmap_inval_info_t info)
1181 KKASSERT(m->hold_count > 0);
1182 if (m->hold_count > 1) {
1186 return _pmap_unwire_pte_hold(pmap, va, m, info);
1192 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1193 pmap_inval_info_t info)
1196 * Wait until we can busy the page ourselves. We cannot have
1197 * any active flushes if we block. We own one hold count on the
1198 * page so it cannot be freed out from under us.
1200 if (m->flags & PG_BUSY) {
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 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);
1375 pmap_page_assertzero(VM_PAGE_TO_PHYS(ptdpg));
1378 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1379 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1381 /* install self-referential address mapping entry */
1382 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1385 pmap->pm_active = 0;
1386 pmap->pm_ptphint = NULL;
1387 TAILQ_INIT(&pmap->pm_pvlist);
1388 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1389 pmap->pm_stats.resident_count = 1;
1393 * Clean up a pmap structure so it can be physically freed. This routine
1394 * is called by the vmspace dtor function. A great deal of pmap data is
1395 * left passively mapped to improve vmspace management so we have a bit
1396 * of cleanup work to do here.
1399 pmap_puninit(pmap_t pmap)
1403 KKASSERT(pmap->pm_active == 0);
1404 lwkt_gettoken(&vm_token);
1405 if ((p = pmap->pm_pdirm) != NULL) {
1406 KKASSERT(pmap->pm_pml4 != NULL);
1407 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1408 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1410 vmstats.v_wire_count--;
1411 KKASSERT((p->flags & PG_BUSY) == 0);
1413 vm_page_free_zero(p);
1414 pmap->pm_pdirm = NULL;
1416 if (pmap->pm_pml4) {
1417 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1418 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1419 pmap->pm_pml4 = NULL;
1421 if (pmap->pm_pteobj) {
1422 vm_object_deallocate(pmap->pm_pteobj);
1423 pmap->pm_pteobj = NULL;
1425 lwkt_reltoken(&vm_token);
1429 * Wire in kernel global address entries. To avoid a race condition
1430 * between pmap initialization and pmap_growkernel, this procedure
1431 * adds the pmap to the master list (which growkernel scans to update),
1432 * then copies the template.
1435 pmap_pinit2(struct pmap *pmap)
1438 lwkt_gettoken(&vm_token);
1439 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1440 /* XXX copies current process, does not fill in MPPTDI */
1441 lwkt_reltoken(&vm_token);
1446 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1447 * 0 on failure (if the procedure had to sleep).
1449 * When asked to remove the page directory page itself, we actually just
1450 * leave it cached so we do not have to incur the SMP inval overhead of
1451 * removing the kernel mapping. pmap_puninit() will take care of it.
1455 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1458 * This code optimizes the case of freeing non-busy
1459 * page-table pages. Those pages are zero now, and
1460 * might as well be placed directly into the zero queue.
1462 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1468 * Remove the page table page from the processes address space.
1470 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1472 * We are the pml4 table itself.
1474 /* XXX anything to do here? */
1475 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1477 * Remove a PDP page from the PML4. We do not maintain
1478 * hold counts on the PML4 page.
1484 m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1485 KKASSERT(m4 != NULL);
1486 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1487 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1488 KKASSERT(pml4[idx] != 0);
1490 } else if (p->pindex >= NUPDE) {
1492 * Remove a PD page from the PDP and drop the hold count
1493 * on the PDP. The PDP is left cached in the pmap if
1494 * the hold count drops to 0 so the wire count remains
1501 m3 = vm_page_lookup(pmap->pm_pteobj,
1502 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1503 KKASSERT(m3 != NULL);
1504 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1505 idx = (p->pindex - NUPDE) % NPDPEPG;
1506 KKASSERT(pdp[idx] != 0);
1511 * Remove a PT page from the PD and drop the hold count
1512 * on the PD. The PD is left cached in the pmap if
1513 * the hold count drops to 0 so the wire count remains
1520 m2 = vm_page_lookup(pmap->pm_pteobj,
1521 NUPDE + p->pindex / NPDEPG);
1522 KKASSERT(m2 != NULL);
1523 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1524 idx = p->pindex % NPDEPG;
1530 * One fewer mappings in the pmap. p's hold count had better
1533 KKASSERT(pmap->pm_stats.resident_count > 0);
1534 --pmap->pm_stats.resident_count;
1536 panic("pmap_release: freeing held page table page");
1537 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1538 pmap->pm_ptphint = NULL;
1541 * We leave the top-level page table page cached, wired, and mapped in
1542 * the pmap until the dtor function (pmap_puninit()) gets called.
1543 * However, still clean it up so we can set PG_ZERO.
1545 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1546 bzero(pmap->pm_pml4, PAGE_SIZE);
1547 vm_page_flag_set(p, PG_ZERO);
1551 KKASSERT(p->wire_count == 0);
1552 vmstats.v_wire_count--;
1553 /* JG eventually revert to using vm_page_free_zero() */
1560 * This routine is called when various levels in the page table need to
1561 * be populated. This routine cannot fail.
1565 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1570 * Find or fabricate a new pagetable page. This will busy the page.
1572 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1573 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1574 if ((m->flags & PG_ZERO) == 0) {
1575 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1579 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1583 KASSERT(m->queue == PQ_NONE,
1584 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1587 * Increment the hold count for the page we will be returning to
1591 if (m->wire_count++ == 0)
1592 vmstats.v_wire_count++;
1593 m->valid = VM_PAGE_BITS_ALL;
1594 vm_page_flag_clear(m, PG_ZERO);
1597 * Map the pagetable page into the process address space, if
1598 * it isn't already there.
1600 * It is possible that someone else got in and mapped the page
1601 * directory page while we were blocked, if so just unbusy and
1602 * return the held page.
1604 if (ptepindex >= (NUPDE + NUPDPE)) {
1606 * Wire up a new PDP page in the PML4
1608 vm_pindex_t pml4index;
1611 pml4index = ptepindex - (NUPDE + NUPDPE);
1612 pml4 = &pmap->pm_pml4[pml4index];
1614 if (--m->wire_count == 0)
1615 --vmstats.v_wire_count;
1619 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1620 } else if (ptepindex >= NUPDE) {
1622 * Wire up a new PD page in the PDP
1624 vm_pindex_t pml4index;
1625 vm_pindex_t pdpindex;
1630 pdpindex = ptepindex - NUPDE;
1631 pml4index = pdpindex >> NPML4EPGSHIFT;
1633 pml4 = &pmap->pm_pml4[pml4index];
1634 if ((*pml4 & PG_V) == 0) {
1636 * Have to allocate a new PDP page, recurse.
1637 * This always succeeds. Returned page will
1640 pdppg = _pmap_allocpte(pmap,
1641 NUPDE + NUPDPE + pml4index);
1644 * Add a held reference to the PDP page.
1646 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1647 pdppg->hold_count++;
1651 * Now find the pdp_entry and map the PDP. If the PDP
1652 * has already been mapped unwind and return the
1653 * already-mapped PDP held.
1655 * pdppg is left held (hold_count is incremented for
1656 * each PD in the PDP).
1658 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1659 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1661 vm_page_unhold(pdppg);
1662 if (--m->wire_count == 0)
1663 --vmstats.v_wire_count;
1667 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1670 * Wire up the new PT page in the PD
1672 vm_pindex_t pml4index;
1673 vm_pindex_t pdpindex;
1679 pdpindex = ptepindex >> NPDPEPGSHIFT;
1680 pml4index = pdpindex >> NPML4EPGSHIFT;
1683 * Locate the PDP page in the PML4, then the PD page in
1684 * the PDP. If either does not exist we simply recurse
1687 * We can just recurse on the PD page as it will recurse
1688 * on the PDP if necessary.
1690 pml4 = &pmap->pm_pml4[pml4index];
1691 if ((*pml4 & PG_V) == 0) {
1692 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1693 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1694 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1696 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1697 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1698 if ((*pdp & PG_V) == 0) {
1699 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1701 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1707 * Now fill in the pte in the PD. If the pte already exists
1708 * (again, if we raced the grab), unhold pdpg and unwire
1709 * m, returning a held m.
1711 * pdpg is left held (hold_count is incremented for
1712 * each PT in the PD).
1714 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
1715 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1717 vm_page_unhold(pdpg);
1718 if (--m->wire_count == 0)
1719 --vmstats.v_wire_count;
1723 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1727 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1728 * valid bits, mapped flag, unbusy, and we're done.
1730 pmap->pm_ptphint = m;
1731 ++pmap->pm_stats.resident_count;
1734 m->valid = VM_PAGE_BITS_ALL;
1735 vm_page_flag_clear(m, PG_ZERO);
1737 vm_page_flag_set(m, PG_MAPPED);
1745 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1747 vm_pindex_t ptepindex;
1752 * Calculate pagetable page index
1754 ptepindex = pmap_pde_pindex(va);
1757 * Get the page directory entry
1759 pd = pmap_pde(pmap, va);
1762 * This supports switching from a 2MB page to a
1765 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1766 panic("no promotion/demotion yet");
1774 * If the page table page is mapped, we just increment the
1775 * hold count, and activate it.
1777 if (pd != NULL && (*pd & PG_V) != 0) {
1778 /* YYY hint is used here on i386 */
1779 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1780 pmap->pm_ptphint = m;
1785 * Here if the pte page isn't mapped, or if it has been deallocated.
1787 return _pmap_allocpte(pmap, ptepindex);
1791 /***************************************************
1792 * Pmap allocation/deallocation routines.
1793 ***************************************************/
1796 * Release any resources held by the given physical map.
1797 * Called when a pmap initialized by pmap_pinit is being released.
1798 * Should only be called if the map contains no valid mappings.
1800 static int pmap_release_callback(struct vm_page *p, void *data);
1803 pmap_release(struct pmap *pmap)
1805 vm_object_t object = pmap->pm_pteobj;
1806 struct rb_vm_page_scan_info info;
1808 KASSERT(pmap->pm_active == 0,
1809 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
1810 #if defined(DIAGNOSTIC)
1811 if (object->ref_count != 1)
1812 panic("pmap_release: pteobj reference count != 1");
1816 info.object = object;
1818 lwkt_gettoken(&vm_token);
1819 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1826 info.limit = object->generation;
1828 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1829 pmap_release_callback, &info);
1830 if (info.error == 0 && info.mpte) {
1831 if (!pmap_release_free_page(pmap, info.mpte))
1835 } while (info.error);
1836 lwkt_reltoken(&vm_token);
1841 pmap_release_callback(struct vm_page *p, void *data)
1843 struct rb_vm_page_scan_info *info = data;
1845 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1849 if (!pmap_release_free_page(info->pmap, p)) {
1853 if (info->object->generation != info->limit) {
1861 * Grow the number of kernel page table entries, if needed.
1863 * This routine is always called to validate any address space
1864 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
1865 * space below KERNBASE.
1868 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1871 vm_offset_t ptppaddr;
1873 pd_entry_t *pde, newpdir;
1875 int update_kernel_vm_end;
1878 lwkt_gettoken(&vm_token);
1881 * bootstrap kernel_vm_end on first real VM use
1883 if (kernel_vm_end == 0) {
1884 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1886 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1887 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1888 ~(PAGE_SIZE * NPTEPG - 1);
1890 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1891 kernel_vm_end = kernel_map.max_offset;
1898 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
1899 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
1900 * do not want to force-fill 128G worth of page tables.
1902 if (kstart < KERNBASE) {
1903 if (kstart > kernel_vm_end)
1904 kstart = kernel_vm_end;
1905 KKASSERT(kend <= KERNBASE);
1906 update_kernel_vm_end = 1;
1908 update_kernel_vm_end = 0;
1911 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
1912 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
1914 if (kend - 1 >= kernel_map.max_offset)
1915 kend = kernel_map.max_offset;
1917 while (kstart < kend) {
1918 pde = pmap_pde(&kernel_pmap, kstart);
1920 /* We need a new PDP entry */
1921 nkpg = vm_page_alloc(kptobj, nkpt,
1924 VM_ALLOC_INTERRUPT);
1926 panic("pmap_growkernel: no memory to grow "
1929 paddr = VM_PAGE_TO_PHYS(nkpg);
1930 if ((nkpg->flags & PG_ZERO) == 0)
1931 pmap_zero_page(paddr);
1932 vm_page_flag_clear(nkpg, PG_ZERO);
1933 newpdp = (pdp_entry_t)
1934 (paddr | PG_V | PG_RW | PG_A | PG_M);
1935 *pmap_pdpe(&kernel_pmap, kstart) = newpdp;
1937 continue; /* try again */
1939 if ((*pde & PG_V) != 0) {
1940 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1941 ~(PAGE_SIZE * NPTEPG - 1);
1942 if (kstart - 1 >= kernel_map.max_offset) {
1943 kstart = kernel_map.max_offset;
1950 * This index is bogus, but out of the way
1952 nkpg = vm_page_alloc(kptobj, nkpt,
1955 VM_ALLOC_INTERRUPT);
1957 panic("pmap_growkernel: no memory to grow kernel");
1960 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1961 pmap_zero_page(ptppaddr);
1962 vm_page_flag_clear(nkpg, PG_ZERO);
1963 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1964 *pmap_pde(&kernel_pmap, kstart) = newpdir;
1967 kstart = (kstart + PAGE_SIZE * NPTEPG) &
1968 ~(PAGE_SIZE * NPTEPG - 1);
1970 if (kstart - 1 >= kernel_map.max_offset) {
1971 kstart = kernel_map.max_offset;
1977 * Only update kernel_vm_end for areas below KERNBASE.
1979 if (update_kernel_vm_end && kernel_vm_end < kstart)
1980 kernel_vm_end = kstart;
1982 lwkt_reltoken(&vm_token);
1987 * Retire the given physical map from service.
1988 * Should only be called if the map contains
1989 * no valid mappings.
1992 pmap_destroy(pmap_t pmap)
1999 lwkt_gettoken(&vm_token);
2000 count = --pmap->pm_count;
2003 panic("destroying a pmap is not yet implemented");
2005 lwkt_reltoken(&vm_token);
2009 * Add a reference to the specified pmap.
2012 pmap_reference(pmap_t pmap)
2015 lwkt_gettoken(&vm_token);
2017 lwkt_reltoken(&vm_token);
2021 /***************************************************
2022 * page management routines.
2023 ***************************************************/
2026 * free the pv_entry back to the free list. This function may be
2027 * called from an interrupt.
2031 free_pv_entry(pv_entry_t pv)
2034 KKASSERT(pv_entry_count >= 0);
2039 * get a new pv_entry, allocating a block from the system
2040 * when needed. This function may be called from an interrupt.
2047 if (pv_entry_high_water &&
2048 (pv_entry_count > pv_entry_high_water) &&
2049 (pmap_pagedaemon_waken == 0)) {
2050 pmap_pagedaemon_waken = 1;
2051 wakeup(&vm_pages_needed);
2053 return zalloc(pvzone);
2057 * This routine is very drastic, but can save the system
2065 static int warningdone=0;
2067 if (pmap_pagedaemon_waken == 0)
2069 lwkt_gettoken(&vm_token);
2070 if (warningdone < 5) {
2071 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
2075 for(i = 0; i < vm_page_array_size; i++) {
2076 m = &vm_page_array[i];
2077 if (m->wire_count || m->hold_count || m->busy ||
2078 (m->flags & PG_BUSY))
2082 pmap_pagedaemon_waken = 0;
2083 lwkt_reltoken(&vm_token);
2088 * If it is the first entry on the list, it is actually
2089 * in the header and we must copy the following entry up
2090 * to the header. Otherwise we must search the list for
2091 * the entry. In either case we free the now unused entry.
2095 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
2096 vm_offset_t va, pmap_inval_info_t info)
2102 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
2103 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2104 if (pmap == pv->pv_pmap && va == pv->pv_va)
2108 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2109 if (va == pv->pv_va)
2117 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2118 m->md.pv_list_count--;
2119 m->object->agg_pv_list_count--;
2120 KKASSERT(m->md.pv_list_count >= 0);
2121 if (TAILQ_EMPTY(&m->md.pv_list))
2122 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2123 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2124 ++pmap->pm_generation;
2125 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
2133 * Create a pv entry for page at pa for
2138 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2143 pv = get_pv_entry();
2148 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2149 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2150 ++pmap->pm_generation;
2151 m->md.pv_list_count++;
2152 m->object->agg_pv_list_count++;
2158 * pmap_remove_pte: do the things to unmap a page in a process
2162 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2163 pmap_inval_info_t info)
2168 pmap_inval_interlock(info, pmap, va);
2169 oldpte = pte_load_clear(ptq);
2170 pmap_inval_deinterlock(info, pmap);
2172 pmap->pm_stats.wired_count -= 1;
2174 * Machines that don't support invlpg, also don't support
2175 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2179 cpu_invlpg((void *)va);
2180 KKASSERT(pmap->pm_stats.resident_count > 0);
2181 --pmap->pm_stats.resident_count;
2182 if (oldpte & PG_MANAGED) {
2183 m = PHYS_TO_VM_PAGE(oldpte);
2184 if (oldpte & PG_M) {
2185 #if defined(PMAP_DIAGNOSTIC)
2186 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2188 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2192 if (pmap_track_modified(va))
2196 vm_page_flag_set(m, PG_REFERENCED);
2197 return pmap_remove_entry(pmap, m, va, info);
2199 return pmap_unuse_pt(pmap, va, NULL, info);
2208 * Remove a single page from a process address space.
2210 * This function may not be called from an interrupt if the pmap is
2215 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2219 pte = pmap_pte(pmap, va);
2222 if ((*pte & PG_V) == 0)
2224 pmap_remove_pte(pmap, pte, va, info);
2230 * Remove the given range of addresses from the specified map.
2232 * It is assumed that the start and end are properly
2233 * rounded to the page size.
2235 * This function may not be called from an interrupt if the pmap is
2239 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2241 vm_offset_t va_next;
2242 pml4_entry_t *pml4e;
2244 pd_entry_t ptpaddr, *pde;
2246 struct pmap_inval_info info;
2251 lwkt_gettoken(&vm_token);
2252 if (pmap->pm_stats.resident_count == 0) {
2253 lwkt_reltoken(&vm_token);
2257 pmap_inval_init(&info);
2260 * special handling of removing one page. a very
2261 * common operation and easy to short circuit some
2264 if (sva + PAGE_SIZE == eva) {
2265 pde = pmap_pde(pmap, sva);
2266 if (pde && (*pde & PG_PS) == 0) {
2267 pmap_remove_page(pmap, sva, &info);
2268 pmap_inval_done(&info);
2269 lwkt_reltoken(&vm_token);
2274 for (; sva < eva; sva = va_next) {
2275 pml4e = pmap_pml4e(pmap, sva);
2276 if ((*pml4e & PG_V) == 0) {
2277 va_next = (sva + NBPML4) & ~PML4MASK;
2283 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2284 if ((*pdpe & PG_V) == 0) {
2285 va_next = (sva + NBPDP) & ~PDPMASK;
2292 * Calculate index for next page table.
2294 va_next = (sva + NBPDR) & ~PDRMASK;
2298 pde = pmap_pdpe_to_pde(pdpe, sva);
2302 * Weed out invalid mappings.
2308 * Check for large page.
2310 if ((ptpaddr & PG_PS) != 0) {
2311 /* JG FreeBSD has more complex treatment here */
2312 pmap_inval_interlock(&info, pmap, -1);
2314 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2315 pmap_inval_deinterlock(&info, pmap);
2320 * Limit our scan to either the end of the va represented
2321 * by the current page table page, or to the end of the
2322 * range being removed.
2328 * NOTE: pmap_remove_pte() can block.
2330 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2334 if (pmap_remove_pte(pmap, pte, sva, &info))
2338 pmap_inval_done(&info);
2339 lwkt_reltoken(&vm_token);
2345 * Removes this physical page from all physical maps in which it resides.
2346 * Reflects back modify bits to the pager.
2348 * This routine may not be called from an interrupt.
2353 pmap_remove_all(vm_page_t m)
2355 struct pmap_inval_info info;
2356 pt_entry_t *pte, tpte;
2359 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2362 lwkt_gettoken(&vm_token);
2363 pmap_inval_init(&info);
2365 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2366 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2367 --pv->pv_pmap->pm_stats.resident_count;
2369 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2370 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
2371 tpte = pte_load_clear(pte);
2373 pv->pv_pmap->pm_stats.wired_count--;
2374 pmap_inval_deinterlock(&info, pv->pv_pmap);
2376 vm_page_flag_set(m, PG_REFERENCED);
2379 * Update the vm_page_t clean and reference bits.
2382 #if defined(PMAP_DIAGNOSTIC)
2383 if (pmap_nw_modified(tpte)) {
2385 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2389 if (pmap_track_modified(pv->pv_va))
2392 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2393 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2394 ++pv->pv_pmap->pm_generation;
2395 m->md.pv_list_count--;
2396 m->object->agg_pv_list_count--;
2397 KKASSERT(m->md.pv_list_count >= 0);
2398 if (TAILQ_EMPTY(&m->md.pv_list))
2399 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2400 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2404 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2405 pmap_inval_done(&info);
2406 lwkt_reltoken(&vm_token);
2412 * Set the physical protection on the specified range of this map
2415 * This function may not be called from an interrupt if the map is
2416 * not the kernel_pmap.
2419 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2421 vm_offset_t va_next;
2422 pml4_entry_t *pml4e;
2424 pd_entry_t ptpaddr, *pde;
2426 pmap_inval_info info;
2428 /* JG review for NX */
2433 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2434 pmap_remove(pmap, sva, eva);
2438 if (prot & VM_PROT_WRITE)
2441 lwkt_gettoken(&vm_token);
2442 pmap_inval_init(&info);
2444 for (; sva < eva; sva = va_next) {
2446 pml4e = pmap_pml4e(pmap, sva);
2447 if ((*pml4e & PG_V) == 0) {
2448 va_next = (sva + NBPML4) & ~PML4MASK;
2454 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2455 if ((*pdpe & PG_V) == 0) {
2456 va_next = (sva + NBPDP) & ~PDPMASK;
2462 va_next = (sva + NBPDR) & ~PDRMASK;
2466 pde = pmap_pdpe_to_pde(pdpe, sva);
2470 * Check for large page.
2472 if ((ptpaddr & PG_PS) != 0) {
2473 pmap_inval_interlock(&info, pmap, -1);
2474 *pde &= ~(PG_M|PG_RW);
2475 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2476 pmap_inval_deinterlock(&info, pmap);
2481 * Weed out invalid mappings. Note: we assume that the page
2482 * directory table is always allocated, and in kernel virtual.
2490 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2499 pmap_inval_interlock(&info, pmap, sva);
2503 if ((pbits & PG_V) == 0) {
2504 pmap_inval_deinterlock(&info, pmap);
2507 if (pbits & PG_MANAGED) {
2510 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2511 vm_page_flag_set(m, PG_REFERENCED);
2515 if (pmap_track_modified(sva)) {
2517 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2524 if (pbits != cbits &&
2525 !atomic_cmpset_long(pte, pbits, cbits)) {
2528 pmap_inval_deinterlock(&info, pmap);
2531 pmap_inval_done(&info);
2532 lwkt_reltoken(&vm_token);
2536 * Insert the given physical page (p) at
2537 * the specified virtual address (v) in the
2538 * target physical map with the protection requested.
2540 * If specified, the page will be wired down, meaning
2541 * that the related pte can not be reclaimed.
2543 * NB: This is the only routine which MAY NOT lazy-evaluate
2544 * or lose information. That is, this routine must actually
2545 * insert this page into the given map NOW.
2548 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2555 pt_entry_t origpte, newpte;
2557 pmap_inval_info info;
2562 va = trunc_page(va);
2563 #ifdef PMAP_DIAGNOSTIC
2565 panic("pmap_enter: toobig");
2566 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2567 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
2569 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2570 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2572 db_print_backtrace();
2575 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2576 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2578 db_print_backtrace();
2582 lwkt_gettoken(&vm_token);
2585 * In the case that a page table page is not
2586 * resident, we are creating it here.
2588 if (va < VM_MAX_USER_ADDRESS)
2589 mpte = pmap_allocpte(pmap, va);
2593 pmap_inval_init(&info);
2594 pde = pmap_pde(pmap, va);
2595 if (pde != NULL && (*pde & PG_V) != 0) {
2596 if ((*pde & PG_PS) != 0)
2597 panic("pmap_enter: attempted pmap_enter on 2MB page");
2598 pte = pmap_pde_to_pte(pde, va);
2600 panic("pmap_enter: invalid page directory va=%#lx", va);
2602 KKASSERT(pte != NULL);
2603 pa = VM_PAGE_TO_PHYS(m);
2605 opa = origpte & PG_FRAME;
2608 * Mapping has not changed, must be protection or wiring change.
2610 if (origpte && (opa == pa)) {
2612 * Wiring change, just update stats. We don't worry about
2613 * wiring PT pages as they remain resident as long as there
2614 * are valid mappings in them. Hence, if a user page is wired,
2615 * the PT page will be also.
2617 if (wired && ((origpte & PG_W) == 0))
2618 pmap->pm_stats.wired_count++;
2619 else if (!wired && (origpte & PG_W))
2620 pmap->pm_stats.wired_count--;
2622 #if defined(PMAP_DIAGNOSTIC)
2623 if (pmap_nw_modified(origpte)) {
2625 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2631 * Remove the extra pte reference. Note that we cannot
2632 * optimize the RO->RW case because we have adjusted the
2633 * wiring count above and may need to adjust the wiring
2640 * We might be turning off write access to the page,
2641 * so we go ahead and sense modify status.
2643 if (origpte & PG_MANAGED) {
2644 if ((origpte & PG_M) && pmap_track_modified(va)) {
2646 om = PHYS_TO_VM_PAGE(opa);
2650 KKASSERT(m->flags & PG_MAPPED);
2655 * Mapping has changed, invalidate old range and fall through to
2656 * handle validating new mapping.
2660 err = pmap_remove_pte(pmap, pte, va, &info);
2662 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2664 opa = origpte & PG_FRAME;
2666 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2672 * Enter on the PV list if part of our managed memory. Note that we
2673 * raise IPL while manipulating pv_table since pmap_enter can be
2674 * called at interrupt time.
2676 if (pmap_initialized &&
2677 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2678 pmap_insert_entry(pmap, va, mpte, m);
2680 vm_page_flag_set(m, PG_MAPPED);
2684 * Increment counters
2686 ++pmap->pm_stats.resident_count;
2688 pmap->pm_stats.wired_count++;
2692 * Now validate mapping with desired protection/wiring.
2694 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
2698 if (va < VM_MAX_USER_ADDRESS)
2700 if (pmap == &kernel_pmap)
2704 * if the mapping or permission bits are different, we need
2705 * to update the pte.
2707 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2708 pmap_inval_interlock(&info, pmap, va);
2709 *pte = newpte | PG_A;
2710 pmap_inval_deinterlock(&info, pmap);
2712 vm_page_flag_set(m, PG_WRITEABLE);
2714 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2715 pmap_inval_done(&info);
2716 lwkt_reltoken(&vm_token);
2720 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2721 * This code also assumes that the pmap has no pre-existing entry for this
2724 * This code currently may only be used on user pmaps, not kernel_pmap.
2727 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2732 vm_pindex_t ptepindex;
2734 pmap_inval_info info;
2736 lwkt_gettoken(&vm_token);
2737 pmap_inval_init(&info);
2739 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2740 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2742 db_print_backtrace();
2745 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2746 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2748 db_print_backtrace();
2752 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2755 * Calculate the page table page (mpte), allocating it if necessary.
2757 * A held page table page (mpte), or NULL, is passed onto the
2758 * section following.
2760 if (va < VM_MAX_USER_ADDRESS) {
2762 * Calculate pagetable page index
2764 ptepindex = pmap_pde_pindex(va);
2768 * Get the page directory entry
2770 ptepa = pmap_pde(pmap, va);
2773 * If the page table page is mapped, we just increment
2774 * the hold count, and activate it.
2776 if (ptepa && (*ptepa & PG_V) != 0) {
2778 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2779 // if (pmap->pm_ptphint &&
2780 // (pmap->pm_ptphint->pindex == ptepindex)) {
2781 // mpte = pmap->pm_ptphint;
2783 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2784 pmap->pm_ptphint = mpte;
2789 mpte = _pmap_allocpte(pmap, ptepindex);
2791 } while (mpte == NULL);
2794 /* this code path is not yet used */
2798 * With a valid (and held) page directory page, we can just use
2799 * vtopte() to get to the pte. If the pte is already present
2800 * we do not disturb it.
2805 pmap_unwire_pte_hold(pmap, va, mpte, &info);
2806 pa = VM_PAGE_TO_PHYS(m);
2807 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2808 pmap_inval_done(&info);
2809 lwkt_reltoken(&vm_token);
2814 * Enter on the PV list if part of our managed memory
2816 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2817 pmap_insert_entry(pmap, va, mpte, m);
2818 vm_page_flag_set(m, PG_MAPPED);
2822 * Increment counters
2824 ++pmap->pm_stats.resident_count;
2826 pa = VM_PAGE_TO_PHYS(m);
2829 * Now validate mapping with RO protection
2831 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2832 *pte = pa | PG_V | PG_U;
2834 *pte = pa | PG_V | PG_U | PG_MANAGED;
2835 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2836 pmap_inval_done(&info);
2837 lwkt_reltoken(&vm_token);
2841 * Make a temporary mapping for a physical address. This is only intended
2842 * to be used for panic dumps.
2844 * The caller is responsible for calling smp_invltlb().
2847 pmap_kenter_temporary(vm_paddr_t pa, long i)
2849 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2850 return ((void *)crashdumpmap);
2853 #define MAX_INIT_PT (96)
2856 * This routine preloads the ptes for a given object into the specified pmap.
2857 * This eliminates the blast of soft faults on process startup and
2858 * immediately after an mmap.
2860 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2863 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2864 vm_object_t object, vm_pindex_t pindex,
2865 vm_size_t size, int limit)
2867 struct rb_vm_page_scan_info info;
2872 * We can't preinit if read access isn't set or there is no pmap
2875 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2879 * We can't preinit if the pmap is not the current pmap
2881 lp = curthread->td_lwp;
2882 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2885 psize = x86_64_btop(size);
2887 if ((object->type != OBJT_VNODE) ||
2888 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2889 (object->resident_page_count > MAX_INIT_PT))) {
2893 if (psize + pindex > object->size) {
2894 if (object->size < pindex)
2896 psize = object->size - pindex;
2903 * Use a red-black scan to traverse the requested range and load
2904 * any valid pages found into the pmap.
2906 * We cannot safely scan the object's memq unless we are in a
2907 * critical section since interrupts can remove pages from objects.
2909 info.start_pindex = pindex;
2910 info.end_pindex = pindex + psize - 1;
2917 lwkt_gettoken(&vm_token);
2918 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2919 pmap_object_init_pt_callback, &info);
2920 lwkt_reltoken(&vm_token);
2926 pmap_object_init_pt_callback(vm_page_t p, void *data)
2928 struct rb_vm_page_scan_info *info = data;
2929 vm_pindex_t rel_index;
2931 * don't allow an madvise to blow away our really
2932 * free pages allocating pv entries.
2934 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2935 vmstats.v_free_count < vmstats.v_free_reserved) {
2938 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2939 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2941 if ((p->queue - p->pc) == PQ_CACHE)
2942 vm_page_deactivate(p);
2943 rel_index = p->pindex - info->start_pindex;
2944 pmap_enter_quick(info->pmap,
2945 info->addr + x86_64_ptob(rel_index), p);
2952 * Return TRUE if the pmap is in shape to trivially
2953 * pre-fault the specified address.
2955 * Returns FALSE if it would be non-trivial or if a
2956 * pte is already loaded into the slot.
2959 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2965 lwkt_gettoken(&vm_token);
2966 pde = pmap_pde(pmap, addr);
2967 if (pde == NULL || *pde == 0) {
2971 ret = (*pte) ? 0 : 1;
2973 lwkt_reltoken(&vm_token);
2978 * Routine: pmap_change_wiring
2979 * Function: Change the wiring attribute for a map/virtual-address
2981 * In/out conditions:
2982 * The mapping must already exist in the pmap.
2985 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2992 lwkt_gettoken(&vm_token);
2993 pte = pmap_pte(pmap, va);
2995 if (wired && !pmap_pte_w(pte))
2996 pmap->pm_stats.wired_count++;
2997 else if (!wired && pmap_pte_w(pte))
2998 pmap->pm_stats.wired_count--;
3001 * Wiring is not a hardware characteristic so there is no need to
3002 * invalidate TLB. However, in an SMP environment we must use
3003 * a locked bus cycle to update the pte (if we are not using
3004 * the pmap_inval_*() API that is)... it's ok to do this for simple
3009 atomic_set_long(pte, PG_W);
3011 atomic_clear_long(pte, PG_W);
3014 atomic_set_long_nonlocked(pte, PG_W);
3016 atomic_clear_long_nonlocked(pte, PG_W);
3018 lwkt_reltoken(&vm_token);
3024 * Copy the range specified by src_addr/len
3025 * from the source map to the range dst_addr/len
3026 * in the destination map.
3028 * This routine is only advisory and need not do anything.
3031 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3032 vm_size_t len, vm_offset_t src_addr)
3036 pmap_inval_info info;
3038 vm_offset_t end_addr = src_addr + len;
3040 pd_entry_t src_frame, dst_frame;
3043 if (dst_addr != src_addr)
3046 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3047 if (src_frame != (PTDpde & PG_FRAME)) {
3051 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3052 if (dst_frame != (APTDpde & PG_FRAME)) {
3053 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
3054 /* The page directory is not shared between CPUs */
3058 pmap_inval_init(&info);
3059 pmap_inval_add(&info, dst_pmap, -1);
3060 pmap_inval_add(&info, src_pmap, -1);
3063 * critical section protection is required to maintain the page/object
3064 * association, interrupts can free pages and remove them from
3068 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
3069 pt_entry_t *src_pte, *dst_pte;
3070 vm_page_t dstmpte, srcmpte;
3071 vm_offset_t srcptepaddr;
3072 vm_pindex_t ptepindex;
3074 if (addr >= UPT_MIN_ADDRESS)
3075 panic("pmap_copy: invalid to pmap_copy page tables\n");
3078 * Don't let optional prefaulting of pages make us go
3079 * way below the low water mark of free pages or way
3080 * above high water mark of used pv entries.
3082 if (vmstats.v_free_count < vmstats.v_free_reserved ||
3083 pv_entry_count > pv_entry_high_water)
3086 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
3087 ptepindex = addr >> PDRSHIFT;
3090 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
3092 if (srcptepaddr == 0)
3095 if (srcptepaddr & PG_PS) {
3097 if (dst_pmap->pm_pdir[ptepindex] == 0) {
3098 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
3099 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
3105 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
3106 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
3107 (srcmpte->flags & PG_BUSY)) {
3111 if (pdnxt > end_addr)
3114 src_pte = vtopte(addr);
3116 dst_pte = avtopte(addr);
3118 while (addr < pdnxt) {
3123 * we only virtual copy managed pages
3125 if ((ptetemp & PG_MANAGED) != 0) {
3127 * We have to check after allocpte for the
3128 * pte still being around... allocpte can
3131 * pmap_allocpte() can block. If we lose
3132 * our page directory mappings we stop.
3134 dstmpte = pmap_allocpte(dst_pmap, addr);
3137 if (src_frame != (PTDpde & PG_FRAME) ||
3138 dst_frame != (APTDpde & PG_FRAME)
3140 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3141 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3143 } else if ((*dst_pte == 0) &&
3144 (ptetemp = *src_pte) != 0 &&
3145 (ptetemp & PG_MANAGED)) {
3147 * Clear the modified and
3148 * accessed (referenced) bits
3151 m = PHYS_TO_VM_PAGE(ptetemp);
3152 *dst_pte = ptetemp & ~(PG_M | PG_A);
3153 ++dst_pmap->pm_stats.resident_count;
3154 pmap_insert_entry(dst_pmap, addr,
3156 KKASSERT(m->flags & PG_MAPPED);
3158 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3159 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3163 if (dstmpte->hold_count >= srcmpte->hold_count)
3173 pmap_inval_done(&info);
3180 * Zero the specified physical page.
3182 * This function may be called from an interrupt and no locking is
3186 pmap_zero_page(vm_paddr_t phys)
3188 vm_offset_t va = PHYS_TO_DMAP(phys);
3190 pagezero((void *)va);
3194 * pmap_page_assertzero:
3196 * Assert that a page is empty, panic if it isn't.
3199 pmap_page_assertzero(vm_paddr_t phys)
3201 vm_offset_t va = PHYS_TO_DMAP(phys);
3204 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3205 if (*(long *)((char *)va + i) != 0) {
3206 panic("pmap_page_assertzero() @ %p not zero!\n",
3207 (void *)(intptr_t)va);
3215 * Zero part of a physical page by mapping it into memory and clearing
3216 * its contents with bzero.
3218 * off and size may not cover an area beyond a single hardware page.
3221 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3223 vm_offset_t virt = PHYS_TO_DMAP(phys);
3225 bzero((char *)virt + off, size);
3231 * Copy the physical page from the source PA to the target PA.
3232 * This function may be called from an interrupt. No locking
3236 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3238 vm_offset_t src_virt, dst_virt;
3240 src_virt = PHYS_TO_DMAP(src);
3241 dst_virt = PHYS_TO_DMAP(dst);
3242 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3246 * pmap_copy_page_frag:
3248 * Copy the physical page from the source PA to the target PA.
3249 * This function may be called from an interrupt. No locking
3253 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3255 vm_offset_t src_virt, dst_virt;
3257 src_virt = PHYS_TO_DMAP(src);
3258 dst_virt = PHYS_TO_DMAP(dst);
3260 bcopy((char *)src_virt + (src & PAGE_MASK),
3261 (char *)dst_virt + (dst & PAGE_MASK),
3266 * Returns true if the pmap's pv is one of the first
3267 * 16 pvs linked to from this page. This count may
3268 * be changed upwards or downwards in the future; it
3269 * is only necessary that true be returned for a small
3270 * subset of pmaps for proper page aging.
3273 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3278 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3282 lwkt_gettoken(&vm_token);
3284 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3285 if (pv->pv_pmap == pmap) {
3286 lwkt_reltoken(&vm_token);
3294 lwkt_reltoken(&vm_token);
3300 * Remove all pages from specified address space
3301 * this aids process exit speeds. Also, this code
3302 * is special cased for current process only, but
3303 * can have the more generic (and slightly slower)
3304 * mode enabled. This is much faster than pmap_remove
3305 * in the case of running down an entire address space.
3308 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3311 pt_entry_t *pte, tpte;
3314 pmap_inval_info info;
3316 int save_generation;
3318 lp = curthread->td_lwp;
3319 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3324 lwkt_gettoken(&vm_token);
3325 pmap_inval_init(&info);
3326 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3327 if (pv->pv_va >= eva || pv->pv_va < sva) {
3328 npv = TAILQ_NEXT(pv, pv_plist);
3332 KKASSERT(pmap == pv->pv_pmap);
3335 pte = vtopte(pv->pv_va);
3337 pte = pmap_pte_quick(pmap, pv->pv_va);
3338 pmap_inval_interlock(&info, pmap, pv->pv_va);
3341 * We cannot remove wired pages from a process' mapping
3345 pmap_inval_deinterlock(&info, pmap);
3346 npv = TAILQ_NEXT(pv, pv_plist);
3349 tpte = pte_load_clear(pte);
3351 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3353 KASSERT(m < &vm_page_array[vm_page_array_size],
3354 ("pmap_remove_pages: bad tpte %lx", tpte));
3356 KKASSERT(pmap->pm_stats.resident_count > 0);
3357 --pmap->pm_stats.resident_count;
3358 pmap_inval_deinterlock(&info, pmap);
3361 * Update the vm_page_t clean and reference bits.
3367 npv = TAILQ_NEXT(pv, pv_plist);
3368 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3369 save_generation = ++pmap->pm_generation;
3371 m->md.pv_list_count--;
3372 m->object->agg_pv_list_count--;
3373 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3374 if (TAILQ_EMPTY(&m->md.pv_list))
3375 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3377 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3381 * Restart the scan if we blocked during the unuse or free
3382 * calls and other removals were made.
3384 if (save_generation != pmap->pm_generation) {
3385 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3386 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3389 pmap_inval_done(&info);
3390 lwkt_reltoken(&vm_token);
3394 * pmap_testbit tests bits in pte's
3395 * note that the testbit/clearbit routines are inline,
3396 * and a lot of things compile-time evaluate.
3400 pmap_testbit(vm_page_t m, int bit)
3405 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3408 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3413 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3415 * if the bit being tested is the modified bit, then
3416 * mark clean_map and ptes as never
3419 if (bit & (PG_A|PG_M)) {
3420 if (!pmap_track_modified(pv->pv_va))
3424 #if defined(PMAP_DIAGNOSTIC)
3425 if (pv->pv_pmap == NULL) {
3426 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3430 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3441 * this routine is used to modify bits in ptes
3445 pmap_clearbit(vm_page_t m, int bit)
3447 struct pmap_inval_info info;
3452 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3455 pmap_inval_init(&info);
3458 * Loop over all current mappings setting/clearing as appropos If
3459 * setting RO do we need to clear the VAC?
3461 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3463 * don't write protect pager mappings
3466 if (!pmap_track_modified(pv->pv_va))
3470 #if defined(PMAP_DIAGNOSTIC)
3471 if (pv->pv_pmap == NULL) {
3472 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3478 * Careful here. We can use a locked bus instruction to
3479 * clear PG_A or PG_M safely but we need to synchronize
3480 * with the target cpus when we mess with PG_RW.
3482 * We do not have to force synchronization when clearing
3483 * PG_M even for PTEs generated via virtual memory maps,
3484 * because the virtual kernel will invalidate the pmap
3485 * entry when/if it needs to resynchronize the Modify bit.
3488 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
3489 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3496 atomic_clear_long(pte, PG_M|PG_RW);
3499 * The cpu may be trying to set PG_M
3500 * simultaniously with our clearing
3503 if (!atomic_cmpset_long(pte, pbits,
3507 } else if (bit == PG_M) {
3509 * We could also clear PG_RW here to force
3510 * a fault on write to redetect PG_M for
3511 * virtual kernels, but it isn't necessary
3512 * since virtual kernels invalidate the pte
3513 * when they clear the VPTE_M bit in their
3514 * virtual page tables.
3516 atomic_clear_long(pte, PG_M);
3518 atomic_clear_long(pte, bit);
3522 pmap_inval_deinterlock(&info, pv->pv_pmap);
3524 pmap_inval_done(&info);
3528 * pmap_page_protect:
3530 * Lower the permission for all mappings to a given page.
3533 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3535 /* JG NX support? */
3536 if ((prot & VM_PROT_WRITE) == 0) {
3537 lwkt_gettoken(&vm_token);
3538 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3539 pmap_clearbit(m, PG_RW);
3540 vm_page_flag_clear(m, PG_WRITEABLE);
3544 lwkt_reltoken(&vm_token);
3549 pmap_phys_address(vm_pindex_t ppn)
3551 return (x86_64_ptob(ppn));
3555 * pmap_ts_referenced:
3557 * Return a count of reference bits for a page, clearing those bits.
3558 * It is not necessary for every reference bit to be cleared, but it
3559 * is necessary that 0 only be returned when there are truly no
3560 * reference bits set.
3562 * XXX: The exact number of bits to check and clear is a matter that
3563 * should be tested and standardized at some point in the future for
3564 * optimal aging of shared pages.
3567 pmap_ts_referenced(vm_page_t m)
3569 pv_entry_t pv, pvf, pvn;
3573 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3577 lwkt_gettoken(&vm_token);
3579 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3584 pvn = TAILQ_NEXT(pv, pv_list);
3587 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3588 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3591 if (!pmap_track_modified(pv->pv_va))
3594 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3596 if (pte && (*pte & PG_A)) {
3598 atomic_clear_long(pte, PG_A);
3600 atomic_clear_long_nonlocked(pte, PG_A);
3607 } while ((pv = pvn) != NULL && pv != pvf);
3609 lwkt_reltoken(&vm_token);
3618 * Return whether or not the specified physical page was modified
3619 * in any physical maps.
3622 pmap_is_modified(vm_page_t m)
3626 lwkt_gettoken(&vm_token);
3627 res = pmap_testbit(m, PG_M);
3628 lwkt_reltoken(&vm_token);
3633 * Clear the modify bits on the specified physical page.
3636 pmap_clear_modify(vm_page_t m)
3638 lwkt_gettoken(&vm_token);
3639 pmap_clearbit(m, PG_M);
3640 lwkt_reltoken(&vm_token);
3644 * pmap_clear_reference:
3646 * Clear the reference bit on the specified physical page.
3649 pmap_clear_reference(vm_page_t m)
3651 lwkt_gettoken(&vm_token);
3652 pmap_clearbit(m, PG_A);
3653 lwkt_reltoken(&vm_token);
3657 * Miscellaneous support routines follow
3662 i386_protection_init(void)
3666 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3667 kp = protection_codes;
3668 for (prot = 0; prot < 8; prot++) {
3670 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3672 * Read access is also 0. There isn't any execute bit,
3673 * so just make it readable.
3675 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3676 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3677 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3680 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3681 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3682 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3683 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3691 * Map a set of physical memory pages into the kernel virtual
3692 * address space. Return a pointer to where it is mapped. This
3693 * routine is intended to be used for mapping device memory,
3696 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3700 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3702 vm_offset_t va, tmpva, offset;
3705 offset = pa & PAGE_MASK;
3706 size = roundup(offset + size, PAGE_SIZE);
3708 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3710 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3712 pa = pa & ~PAGE_MASK;
3713 for (tmpva = va; size > 0;) {
3714 pte = vtopte(tmpva);
3715 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3723 return ((void *)(va + offset));
3727 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3729 vm_offset_t va, tmpva, offset;
3732 offset = pa & PAGE_MASK;
3733 size = roundup(offset + size, PAGE_SIZE);
3735 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3737 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3739 pa = pa & ~PAGE_MASK;
3740 for (tmpva = va; size > 0;) {
3741 pte = vtopte(tmpva);
3742 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3750 return ((void *)(va + offset));
3754 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3756 vm_offset_t base, offset;
3758 base = va & ~PAGE_MASK;
3759 offset = va & PAGE_MASK;
3760 size = roundup(offset + size, PAGE_SIZE);
3761 pmap_qremove(va, size >> PAGE_SHIFT);
3762 kmem_free(&kernel_map, base, size);
3766 * perform the pmap work for mincore
3769 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3771 pt_entry_t *ptep, pte;
3775 lwkt_gettoken(&vm_token);
3776 ptep = pmap_pte(pmap, addr);
3778 if (ptep && (pte = *ptep) != 0) {
3781 val = MINCORE_INCORE;
3782 if ((pte & PG_MANAGED) == 0)
3785 pa = pte & PG_FRAME;
3787 m = PHYS_TO_VM_PAGE(pa);
3793 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3795 * Modified by someone
3797 else if (m->dirty || pmap_is_modified(m))
3798 val |= MINCORE_MODIFIED_OTHER;
3803 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3806 * Referenced by someone
3808 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3809 val |= MINCORE_REFERENCED_OTHER;
3810 vm_page_flag_set(m, PG_REFERENCED);
3814 lwkt_reltoken(&vm_token);
3819 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3820 * vmspace will be ref'd and the old one will be deref'd.
3822 * The vmspace for all lwps associated with the process will be adjusted
3823 * and cr3 will be reloaded if any lwp is the current lwp.
3826 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3828 struct vmspace *oldvm;
3832 oldvm = p->p_vmspace;
3833 if (oldvm != newvm) {
3834 p->p_vmspace = newvm;
3835 KKASSERT(p->p_nthreads == 1);
3836 lp = RB_ROOT(&p->p_lwp_tree);
3837 pmap_setlwpvm(lp, newvm);
3839 sysref_get(&newvm->vm_sysref);
3840 sysref_put(&oldvm->vm_sysref);
3847 * Set the vmspace for a LWP. The vmspace is almost universally set the
3848 * same as the process vmspace, but virtual kernels need to swap out contexts
3849 * on a per-lwp basis.
3852 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3854 struct vmspace *oldvm;
3858 oldvm = lp->lwp_vmspace;
3860 if (oldvm != newvm) {
3861 lp->lwp_vmspace = newvm;
3862 if (curthread->td_lwp == lp) {
3863 pmap = vmspace_pmap(newvm);
3865 atomic_set_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3866 if (pmap->pm_active & CPUMASK_LOCK)
3867 pmap_interlock_wait(newvm);
3869 pmap->pm_active |= 1;
3871 #if defined(SWTCH_OPTIM_STATS)
3874 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
3875 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
3876 load_cr3(curthread->td_pcb->pcb_cr3);
3877 pmap = vmspace_pmap(oldvm);
3879 atomic_clear_cpumask(&pmap->pm_active, mycpu->gd_cpumask);
3881 pmap->pm_active &= ~(cpumask_t)1;
3891 * Called when switching to a locked pmap
3894 pmap_interlock_wait(struct vmspace *vm)
3896 struct pmap *pmap = &vm->vm_pmap;
3898 if (pmap->pm_active & CPUMASK_LOCK) {
3899 DEBUG_PUSH_INFO("pmap_interlock_wait");
3900 while (pmap->pm_active & CPUMASK_LOCK) {
3903 lwkt_process_ipiq();
3912 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3915 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3919 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3924 * Used by kmalloc/kfree, page already exists at va
3927 pmap_kvtom(vm_offset_t va)
3929 return(PHYS_TO_VM_PAGE(*vtopte(va) & PG_FRAME));