2 * Copyright (c) 1991 Regents of the University of California.
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1994 David Greenman
5 * Copyright (c) 2003 Peter Wemm
6 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
7 * Copyright (c) 2008, 2009 The DragonFly Project.
8 * Copyright (c) 2008, 2009 Jordan Gordeev.
11 * This code is derived from software contributed to Berkeley by
12 * the Systems Programming Group of the University of Utah Computer
13 * Science Department and William Jolitz of UUNET Technologies Inc.
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
44 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
48 * Manages physical address maps.
50 * In addition to hardware address maps, this
51 * module is called upon to provide software-use-only
52 * maps which may or may not be stored in the same
53 * form as hardware maps. These pseudo-maps are
54 * used to store intermediate results from copy
55 * operations to and from address spaces.
57 * Since the information managed by this module is
58 * also stored by the logical address mapping module,
59 * this module may throw away valid virtual-to-physical
60 * mappings at almost any time. However, invalidations
61 * of virtual-to-physical mappings must be done as
64 * In order to cope with hardware architectures which
65 * make virtual-to-physical map invalidates expensive,
66 * this module may delay invalidate or reduced protection
67 * operations until such time as they are actually
68 * necessary. This module is given full information as
69 * to which processors are currently using which maps,
70 * and to when physical maps must be made correct.
74 #include "opt_disable_pse.h"
77 #include "opt_msgbuf.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
83 #include <sys/msgbuf.h>
84 #include <sys/vmmeter.h>
88 #include <vm/vm_param.h>
89 #include <sys/sysctl.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_pageout.h>
97 #include <vm/vm_pager.h>
98 #include <vm/vm_zone.h>
100 #include <sys/user.h>
101 #include <sys/thread2.h>
102 #include <sys/sysref2.h>
104 #include <machine/cputypes.h>
105 #include <machine/md_var.h>
106 #include <machine/specialreg.h>
107 #include <machine/smp.h>
108 #include <machine_base/apic/apicreg.h>
109 #include <machine/globaldata.h>
110 #include <machine/pmap.h>
111 #include <machine/pmap_inval.h>
115 #define PMAP_KEEP_PDIRS
116 #ifndef PMAP_SHPGPERPROC
117 #define PMAP_SHPGPERPROC 200
120 #if defined(DIAGNOSTIC)
121 #define PMAP_DIAGNOSTIC
127 * Get PDEs and PTEs for user/kernel address space
129 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
130 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
132 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
133 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
134 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
135 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
136 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
140 * Given a map and a machine independent protection code,
141 * convert to a vax protection code.
143 #define pte_prot(m, p) \
144 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
145 static int protection_codes[8];
147 struct pmap kernel_pmap;
148 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
150 vm_paddr_t avail_start; /* PA of first available physical page */
151 vm_paddr_t avail_end; /* PA of last available physical page */
152 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
153 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
154 vm_offset_t KvaStart; /* VA start of KVA space */
155 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
156 vm_offset_t KvaSize; /* max size of kernel virtual address space */
157 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
158 static int pgeflag; /* PG_G or-in */
159 static int pseflag; /* PG_PS or-in */
161 static vm_object_t kptobj;
164 static vm_paddr_t dmaplimit;
166 vm_offset_t kernel_vm_end;
168 static uint64_t KPDphys; /* phys addr of kernel level 2 */
169 uint64_t KPDPphys; /* phys addr of kernel level 3 */
170 uint64_t KPML4phys; /* phys addr of kernel level 4 */
172 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
173 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
176 * Data for the pv entry allocation mechanism
178 static vm_zone_t pvzone;
179 static struct vm_zone pvzone_store;
180 static struct vm_object pvzone_obj;
181 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
182 static int pmap_pagedaemon_waken = 0;
183 static struct pv_entry *pvinit;
186 * All those kernel PT submaps that BSD is so fond of
188 pt_entry_t *CMAP1 = 0, *ptmmap;
189 caddr_t CADDR1 = 0, ptvmmap = 0;
190 static pt_entry_t *msgbufmap;
191 struct msgbuf *msgbufp=0;
196 static pt_entry_t *pt_crashdumpmap;
197 static caddr_t crashdumpmap;
199 extern uint64_t KPTphys;
200 extern pt_entry_t *SMPpt;
201 extern uint64_t SMPptpa;
205 static pv_entry_t get_pv_entry (void);
206 static void i386_protection_init (void);
207 static void create_pagetables(vm_paddr_t *firstaddr);
208 static void pmap_remove_all (vm_page_t m);
209 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
210 vm_offset_t sva, pmap_inval_info_t info);
211 static void pmap_remove_page (struct pmap *pmap,
212 vm_offset_t va, pmap_inval_info_t info);
213 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
214 vm_offset_t va, pmap_inval_info_t info);
215 static boolean_t pmap_testbit (vm_page_t m, int bit);
216 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
217 vm_page_t mpte, vm_page_t m);
219 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
221 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
222 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
223 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
224 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
225 static int _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
226 pmap_inval_info_t info);
227 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
228 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
230 static unsigned pdir4mb;
233 * Move the kernel virtual free pointer to the next
234 * 2MB. This is used to help improve performance
235 * by using a large (2MB) page for much of the kernel
236 * (.text, .data, .bss)
240 pmap_kmem_choose(vm_offset_t addr)
242 vm_offset_t newaddr = addr;
244 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
251 * Super fast pmap_pte routine best used when scanning the pv lists.
252 * This eliminates many course-grained invltlb calls. Note that many of
253 * the pv list scans are across different pmaps and it is very wasteful
254 * to do an entire invltlb when checking a single mapping.
256 * Should only be called while in a critical section.
258 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
262 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
264 return pmap_pte(pmap, va);
267 /* Return a non-clipped PD index for a given VA */
270 pmap_pde_pindex(vm_offset_t va)
272 return va >> PDRSHIFT;
275 /* Return various clipped indexes for a given VA */
278 pmap_pte_index(vm_offset_t va)
281 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
286 pmap_pde_index(vm_offset_t va)
289 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
294 pmap_pdpe_index(vm_offset_t va)
297 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
302 pmap_pml4e_index(vm_offset_t va)
305 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
308 /* Return a pointer to the PML4 slot that corresponds to a VA */
311 pmap_pml4e(pmap_t pmap, vm_offset_t va)
314 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
317 /* Return a pointer to the PDP slot that corresponds to a VA */
320 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
324 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
325 return (&pdpe[pmap_pdpe_index(va)]);
328 /* Return a pointer to the PDP slot that corresponds to a VA */
331 pmap_pdpe(pmap_t pmap, vm_offset_t va)
335 pml4e = pmap_pml4e(pmap, va);
336 if ((*pml4e & PG_V) == 0)
338 return (pmap_pml4e_to_pdpe(pml4e, va));
341 /* Return a pointer to the PD slot that corresponds to a VA */
344 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
348 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
349 return (&pde[pmap_pde_index(va)]);
352 /* Return a pointer to the PD slot that corresponds to a VA */
355 pmap_pde(pmap_t pmap, vm_offset_t va)
359 pdpe = pmap_pdpe(pmap, va);
360 if (pdpe == NULL || (*pdpe & PG_V) == 0)
362 return (pmap_pdpe_to_pde(pdpe, va));
365 /* Return a pointer to the PT slot that corresponds to a VA */
368 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
372 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
373 return (&pte[pmap_pte_index(va)]);
376 /* Return a pointer to the PT slot that corresponds to a VA */
379 pmap_pte(pmap_t pmap, vm_offset_t va)
383 pde = pmap_pde(pmap, va);
384 if (pde == NULL || (*pde & PG_V) == 0)
386 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
387 return ((pt_entry_t *)pde);
388 return (pmap_pde_to_pte(pde, va));
393 vtopte(vm_offset_t va)
395 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
397 return (PTmap + ((va >> PAGE_SHIFT) & mask));
402 vtopde(vm_offset_t va)
404 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
406 return (PDmap + ((va >> PDRSHIFT) & mask));
410 allocpages(vm_paddr_t *firstaddr, int n)
415 bzero((void *)ret, n * PAGE_SIZE);
416 *firstaddr += n * PAGE_SIZE;
422 create_pagetables(vm_paddr_t *firstaddr)
426 /* we are running (mostly) V=P at this point */
429 KPTphys = allocpages(firstaddr, NKPT);
430 KPML4phys = allocpages(firstaddr, 1);
431 KPDPphys = allocpages(firstaddr, NKPML4E);
432 KPDphys = allocpages(firstaddr, NKPDPE);
434 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
435 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
437 DMPDPphys = allocpages(firstaddr, NDMPML4E);
438 if ((amd_feature & AMDID_PAGE1GB) == 0)
439 DMPDphys = allocpages(firstaddr, ndmpdp);
440 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
442 /* Fill in the underlying page table pages */
443 /* Read-only from zero to physfree */
444 /* XXX not fully used, underneath 2M pages */
445 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
446 ((pt_entry_t *)KPTphys)[i] = i << PAGE_SHIFT;
447 ((pt_entry_t *)KPTphys)[i] |= PG_RW | PG_V | PG_G;
450 /* Now map the page tables at their location within PTmap */
451 for (i = 0; i < NKPT; i++) {
452 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
453 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
456 /* Map from zero to end of allocations under 2M pages */
457 /* This replaces some of the KPTphys entries above */
458 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
459 ((pd_entry_t *)KPDphys)[i] = i << PDRSHIFT;
460 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V | PG_PS | PG_G;
463 /* And connect up the PD to the PDP */
464 for (i = 0; i < NKPDPE; i++) {
465 ((pdp_entry_t *)KPDPphys)[i + KPDPI] = KPDphys +
467 ((pdp_entry_t *)KPDPphys)[i + KPDPI] |= PG_RW | PG_V | PG_U;
470 /* Now set up the direct map space using either 2MB or 1GB pages */
471 /* Preset PG_M and PG_A because demotion expects it */
472 if ((amd_feature & AMDID_PAGE1GB) == 0) {
473 for (i = 0; i < NPDEPG * ndmpdp; i++) {
474 ((pd_entry_t *)DMPDphys)[i] = (vm_paddr_t)i << PDRSHIFT;
475 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
478 /* And the direct map space's PDP */
479 for (i = 0; i < ndmpdp; i++) {
480 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
482 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
485 for (i = 0; i < ndmpdp; i++) {
486 ((pdp_entry_t *)DMPDPphys)[i] =
487 (vm_paddr_t)i << PDPSHIFT;
488 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
493 /* And recursively map PML4 to itself in order to get PTmap */
494 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
495 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
497 /* Connect the Direct Map slot up to the PML4 */
498 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
499 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
501 /* Connect the KVA slot up to the PML4 */
502 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
503 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
507 init_paging(vm_paddr_t *firstaddr)
509 create_pagetables(firstaddr);
513 * Bootstrap the system enough to run with virtual memory.
515 * On the i386 this is called after mapping has already been enabled
516 * and just syncs the pmap module with what has already been done.
517 * [We can't call it easily with mapping off since the kernel is not
518 * mapped with PA == VA, hence we would have to relocate every address
519 * from the linked base (virtual) address "KERNBASE" to the actual
520 * (physical) address starting relative to 0]
523 pmap_bootstrap(vm_paddr_t *firstaddr)
527 struct mdglobaldata *gd;
530 KvaStart = VM_MIN_KERNEL_ADDRESS;
531 KvaEnd = VM_MAX_KERNEL_ADDRESS;
532 KvaSize = KvaEnd - KvaStart;
534 avail_start = *firstaddr;
537 * Create an initial set of page tables to run the kernel in.
539 create_pagetables(firstaddr);
541 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
542 virtual_start = pmap_kmem_choose(virtual_start);
544 virtual_end = VM_MAX_KERNEL_ADDRESS;
546 /* XXX do %cr0 as well */
547 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
551 * Initialize protection array.
553 i386_protection_init();
556 * The kernel's pmap is statically allocated so we don't have to use
557 * pmap_create, which is unlikely to work correctly at this part of
558 * the boot sequence (XXX and which no longer exists).
560 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
561 kernel_pmap.pm_count = 1;
562 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
563 TAILQ_INIT(&kernel_pmap.pm_pvlist);
567 * Reserve some special page table entries/VA space for temporary
570 #define SYSMAP(c, p, v, n) \
571 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
575 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
581 * CMAP1/CMAP2 are used for zeroing and copying pages.
583 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
588 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
591 * ptvmmap is used for reading arbitrary physical pages via
594 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
597 * msgbufp is used to map the system message buffer.
598 * XXX msgbufmap is not used.
600 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
601 atop(round_page(MSGBUF_SIZE)))
608 * PG_G is terribly broken on SMP because we IPI invltlb's in some
609 * cases rather then invl1pg. Actually, I don't even know why it
610 * works under UP because self-referential page table mappings
615 if (cpu_feature & CPUID_PGE)
620 * Initialize the 4MB page size flag
624 * The 4MB page version of the initial
625 * kernel page mapping.
629 #if !defined(DISABLE_PSE)
630 if (cpu_feature & CPUID_PSE) {
633 * Note that we have enabled PSE mode
636 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
637 ptditmp &= ~(NBPDR - 1);
638 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
643 * Enable the PSE mode. If we are SMP we can't do this
644 * now because the APs will not be able to use it when
647 load_cr4(rcr4() | CR4_PSE);
650 * We can do the mapping here for the single processor
651 * case. We simply ignore the old page table page from
655 * For SMP, we still need 4K pages to bootstrap APs,
656 * PSE will be enabled as soon as all APs are up.
658 PTD[KPTDI] = (pd_entry_t)ptditmp;
664 if (cpu_apic_address == 0)
665 panic("pmap_bootstrap: no local apic!");
669 * We need to finish setting up the globaldata page for the BSP.
670 * locore has already populated the page table for the mdglobaldata
673 pg = MDGLOBALDATA_BASEALLOC_PAGES;
674 gd = &CPU_prvspace[0].mdglobaldata;
675 gd->gd_CMAP1 = &SMPpt[pg + 0];
676 gd->gd_CMAP2 = &SMPpt[pg + 1];
677 gd->gd_CMAP3 = &SMPpt[pg + 2];
678 gd->gd_PMAP1 = &SMPpt[pg + 3];
679 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
680 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
681 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
682 gd->gd_PADDR1 = (pt_entry_t *)CPU_prvspace[0].PPAGE1;
689 * Set 4mb pdir for mp startup
694 if (pseflag && (cpu_feature & CPUID_PSE)) {
695 load_cr4(rcr4() | CR4_PSE);
696 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
704 * Initialize the pmap module.
705 * Called by vm_init, to initialize any structures that the pmap
706 * system needs to map virtual memory.
707 * pmap_init has been enhanced to support in a fairly consistant
708 * way, discontiguous physical memory.
717 * object for kernel page table pages
719 /* JG I think the number can be arbitrary */
720 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
723 * Allocate memory for random pmap data structures. Includes the
727 for(i = 0; i < vm_page_array_size; i++) {
730 m = &vm_page_array[i];
731 TAILQ_INIT(&m->md.pv_list);
732 m->md.pv_list_count = 0;
736 * init the pv free list
738 initial_pvs = vm_page_array_size;
739 if (initial_pvs < MINPV)
741 pvzone = &pvzone_store;
742 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
743 initial_pvs * sizeof (struct pv_entry));
744 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
748 * Now it is safe to enable pv_table recording.
750 pmap_initialized = TRUE;
752 lapic = pmap_mapdev_uncacheable(cpu_apic_address, sizeof(struct LAPIC));
757 * Initialize the address space (zone) for the pv_entries. Set a
758 * high water mark so that the system can recover from excessive
759 * numbers of pv entries.
764 int shpgperproc = PMAP_SHPGPERPROC;
766 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
767 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
768 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
769 pv_entry_high_water = 9 * (pv_entry_max / 10);
770 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
774 /***************************************************
775 * Low level helper routines.....
776 ***************************************************/
778 #if defined(PMAP_DIAGNOSTIC)
781 * This code checks for non-writeable/modified pages.
782 * This should be an invalid condition.
786 pmap_nw_modified(pt_entry_t pte)
788 if ((pte & (PG_M|PG_RW)) == PG_M)
797 * this routine defines the region(s) of memory that should
798 * not be tested for the modified bit.
802 pmap_track_modified(vm_offset_t va)
804 if ((va < clean_sva) || (va >= clean_eva))
813 * Extract the physical page address associated with the map/VA pair.
815 * This function may not be called from an interrupt if the pmap is
819 pmap_extract(pmap_t pmap, vm_offset_t va)
823 pd_entry_t pde, *pdep;
826 pdep = pmap_pde(pmap, va);
830 if ((pde & PG_PS) != 0) {
831 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
833 pte = pmap_pde_to_pte(pdep, va);
834 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
842 * Routine: pmap_kextract
844 * Extract the physical page address associated
845 * kernel virtual address.
848 pmap_kextract(vm_offset_t va)
853 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
854 pa = DMAP_TO_PHYS(va);
858 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
861 * Beware of a concurrent promotion that changes the
862 * PDE at this point! For example, vtopte() must not
863 * be used to access the PTE because it would use the
864 * new PDE. It is, however, safe to use the old PDE
865 * because the page table page is preserved by the
868 pa = *pmap_pde_to_pte(&pde, va);
869 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
875 /***************************************************
876 * Low level mapping routines.....
877 ***************************************************/
880 * Routine: pmap_kenter
882 * Add a wired page to the KVA
883 * NOTE! note that in order for the mapping to take effect -- you
884 * should do an invltlb after doing the pmap_kenter().
887 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
891 pmap_inval_info info;
893 pmap_inval_init(&info);
894 npte = pa | PG_RW | PG_V | pgeflag;
896 pmap_inval_add(&info, &kernel_pmap, va);
898 pmap_inval_flush(&info);
902 * Routine: pmap_kenter_quick
904 * Similar to pmap_kenter(), except we only invalidate the
905 * mapping on the current CPU.
908 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
913 npte = pa | PG_RW | PG_V | pgeflag;
916 cpu_invlpg((void *)va);
920 pmap_kenter_sync(vm_offset_t va)
922 pmap_inval_info info;
924 pmap_inval_init(&info);
925 pmap_inval_add(&info, &kernel_pmap, va);
926 pmap_inval_flush(&info);
930 pmap_kenter_sync_quick(vm_offset_t va)
932 cpu_invlpg((void *)va);
936 * remove a page from the kernel pagetables
939 pmap_kremove(vm_offset_t va)
942 pmap_inval_info info;
944 pmap_inval_init(&info);
946 pmap_inval_add(&info, &kernel_pmap, va);
948 pmap_inval_flush(&info);
952 pmap_kremove_quick(vm_offset_t va)
957 cpu_invlpg((void *)va);
961 * XXX these need to be recoded. They are not used in any critical path.
964 pmap_kmodify_rw(vm_offset_t va)
966 *vtopte(va) |= PG_RW;
967 cpu_invlpg((void *)va);
971 pmap_kmodify_nc(vm_offset_t va)
974 cpu_invlpg((void *)va);
978 * Used to map a range of physical addresses into kernel
979 * virtual address space.
981 * For now, VM is already on, we only need to map the
985 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
987 return PHYS_TO_DMAP(start);
992 * Add a list of wired pages to the kva
993 * this routine is only used for temporary
994 * kernel mappings that do not need to have
995 * page modification or references recorded.
996 * Note that old mappings are simply written
997 * over. The page *must* be wired.
1000 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1004 end_va = va + count * PAGE_SIZE;
1006 while (va < end_va) {
1010 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1011 cpu_invlpg((void *)va);
1016 smp_invltlb(); /* XXX */
1021 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
1024 cpumask_t cmask = mycpu->gd_cpumask;
1026 end_va = va + count * PAGE_SIZE;
1028 while (va < end_va) {
1033 * Install the new PTE. If the pte changed from the prior
1034 * mapping we must reset the cpu mask and invalidate the page.
1035 * If the pte is the same but we have not seen it on the
1036 * current cpu, invlpg the existing mapping. Otherwise the
1037 * entry is optimal and no invalidation is required.
1040 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
1041 if (*pte != pteval) {
1044 cpu_invlpg((void *)va);
1045 } else if ((*mask & cmask) == 0) {
1046 cpu_invlpg((void *)va);
1055 * This routine jerks page mappings from the
1056 * kernel -- it is meant only for temporary mappings.
1058 * MPSAFE, INTERRUPT SAFE (cluster callback)
1061 pmap_qremove(vm_offset_t va, int count)
1065 end_va = va + count * PAGE_SIZE;
1067 while (va < end_va) {
1072 cpu_invlpg((void *)va);
1081 * This routine works like vm_page_lookup() but also blocks as long as the
1082 * page is busy. This routine does not busy the page it returns.
1084 * Unless the caller is managing objects whos pages are in a known state,
1085 * the call should be made with a critical section held so the page's object
1086 * association remains valid on return.
1090 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1095 m = vm_page_lookup(object, pindex);
1096 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
1102 * Create a new thread and optionally associate it with a (new) process.
1103 * NOTE! the new thread's cpu may not equal the current cpu.
1106 pmap_init_thread(thread_t td)
1108 /* enforce pcb placement */
1109 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1110 td->td_savefpu = &td->td_pcb->pcb_save;
1111 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on x86_64? */
1115 * This routine directly affects the fork perf for a process.
1118 pmap_init_proc(struct proc *p)
1123 * Dispose the UPAGES for a process that has exited.
1124 * This routine directly impacts the exit perf of a process.
1127 pmap_dispose_proc(struct proc *p)
1129 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1132 /***************************************************
1133 * Page table page management routines.....
1134 ***************************************************/
1137 * This routine unholds page table pages, and if the hold count
1138 * drops to zero, then it decrements the wire count.
1142 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1143 pmap_inval_info_t info)
1145 KKASSERT(m->hold_count > 0);
1146 if (m->hold_count > 1) {
1150 return _pmap_unwire_pte_hold(pmap, va, m, info);
1156 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1157 pmap_inval_info_t info)
1160 * Wait until we can busy the page ourselves. We cannot have
1161 * any active flushes if we block. We own one hold count on the
1162 * page so it cannot be freed out from under us.
1164 if (m->flags & PG_BUSY) {
1165 pmap_inval_flush(info);
1166 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1169 KASSERT(m->queue == PQ_NONE,
1170 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1173 * This case can occur if new references were acquired while
1176 if (m->hold_count > 1) {
1177 KKASSERT(m->hold_count > 1);
1183 * Unmap the page table page
1185 KKASSERT(m->hold_count == 1);
1187 pmap_inval_add(info, pmap, -1);
1189 if (m->pindex >= (NUPDE + NUPDPE)) {
1192 pml4 = pmap_pml4e(pmap, va);
1194 } else if (m->pindex >= NUPDE) {
1197 pdp = pmap_pdpe(pmap, va);
1202 pd = pmap_pde(pmap, va);
1206 KKASSERT(pmap->pm_stats.resident_count > 0);
1207 --pmap->pm_stats.resident_count;
1209 if (pmap->pm_ptphint == m)
1210 pmap->pm_ptphint = NULL;
1212 if (m->pindex < NUPDE) {
1213 /* We just released a PT, unhold the matching PD */
1216 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
1217 pmap_unwire_pte_hold(pmap, va, pdpg, info);
1219 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
1220 /* We just released a PD, unhold the matching PDP */
1223 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
1224 pmap_unwire_pte_hold(pmap, va, pdppg, info);
1228 * This was our last hold, the page had better be unwired
1229 * after we decrement wire_count.
1231 * FUTURE NOTE: shared page directory page could result in
1232 * multiple wire counts.
1236 KKASSERT(m->wire_count == 0);
1237 --vmstats.v_wire_count;
1238 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1240 vm_page_free_zero(m);
1246 * After removing a page table entry, this routine is used to
1247 * conditionally free the page, and manage the hold/wire counts.
1251 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1252 pmap_inval_info_t info)
1254 vm_pindex_t ptepindex;
1256 if (va >= VM_MAX_USER_ADDRESS)
1260 ptepindex = pmap_pde_pindex(va);
1262 if (pmap->pm_ptphint &&
1263 (pmap->pm_ptphint->pindex == ptepindex)) {
1264 mpte = pmap->pm_ptphint;
1267 pmap_inval_flush(info);
1268 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1269 pmap->pm_ptphint = mpte;
1274 return pmap_unwire_pte_hold(pmap, va, mpte, info);
1278 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1279 * it, and IdlePTD, represents the template used to update all other pmaps.
1281 * On architectures where the kernel pmap is not integrated into the user
1282 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1283 * kernel_pmap should be used to directly access the kernel_pmap.
1286 pmap_pinit0(struct pmap *pmap)
1288 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1290 pmap->pm_active = 0;
1291 pmap->pm_ptphint = NULL;
1292 TAILQ_INIT(&pmap->pm_pvlist);
1293 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1297 * Initialize a preallocated and zeroed pmap structure,
1298 * such as one in a vmspace structure.
1301 pmap_pinit(struct pmap *pmap)
1306 * No need to allocate page table space yet but we do need a valid
1307 * page directory table.
1309 if (pmap->pm_pml4 == NULL) {
1311 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1315 * Allocate an object for the ptes
1317 if (pmap->pm_pteobj == NULL)
1318 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1);
1321 * Allocate the page directory page, unless we already have
1322 * one cached. If we used the cached page the wire_count will
1323 * already be set appropriately.
1325 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1326 ptdpg = vm_page_grab(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I,
1327 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1328 pmap->pm_pdirm = ptdpg;
1329 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1330 ptdpg->valid = VM_PAGE_BITS_ALL;
1331 if (ptdpg->wire_count == 0)
1332 ++vmstats.v_wire_count;
1333 ptdpg->wire_count = 1;
1334 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1336 if ((ptdpg->flags & PG_ZERO) == 0)
1337 bzero(pmap->pm_pml4, PAGE_SIZE);
1339 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1340 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1342 /* install self-referential address mapping entry */
1343 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1346 pmap->pm_active = 0;
1347 pmap->pm_ptphint = NULL;
1348 TAILQ_INIT(&pmap->pm_pvlist);
1349 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1350 pmap->pm_stats.resident_count = 1;
1354 * Clean up a pmap structure so it can be physically freed. This routine
1355 * is called by the vmspace dtor function. A great deal of pmap data is
1356 * left passively mapped to improve vmspace management so we have a bit
1357 * of cleanup work to do here.
1360 pmap_puninit(pmap_t pmap)
1364 KKASSERT(pmap->pm_active == 0);
1365 if ((p = pmap->pm_pdirm) != NULL) {
1366 KKASSERT(pmap->pm_pml4 != NULL);
1367 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1368 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1370 vmstats.v_wire_count--;
1371 KKASSERT((p->flags & PG_BUSY) == 0);
1373 vm_page_free_zero(p);
1374 pmap->pm_pdirm = NULL;
1376 if (pmap->pm_pml4) {
1377 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1378 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1379 pmap->pm_pml4 = NULL;
1381 if (pmap->pm_pteobj) {
1382 vm_object_deallocate(pmap->pm_pteobj);
1383 pmap->pm_pteobj = NULL;
1388 * Wire in kernel global address entries. To avoid a race condition
1389 * between pmap initialization and pmap_growkernel, this procedure
1390 * adds the pmap to the master list (which growkernel scans to update),
1391 * then copies the template.
1394 pmap_pinit2(struct pmap *pmap)
1397 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1398 /* XXX copies current process, does not fill in MPPTDI */
1403 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1404 * 0 on failure (if the procedure had to sleep).
1406 * When asked to remove the page directory page itself, we actually just
1407 * leave it cached so we do not have to incur the SMP inval overhead of
1408 * removing the kernel mapping. pmap_puninit() will take care of it.
1412 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1415 * This code optimizes the case of freeing non-busy
1416 * page-table pages. Those pages are zero now, and
1417 * might as well be placed directly into the zero queue.
1419 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1425 * Remove the page table page from the processes address space.
1427 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1429 * We are the pml4 table itself.
1431 /* XXX anything to do here? */
1432 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1434 * Remove a PDP page from the PML4. We do not maintain
1435 * hold counts on the PML4 page.
1441 m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1442 KKASSERT(m4 != NULL);
1443 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1444 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1445 KKASSERT(pml4[idx] != 0);
1447 } else if (p->pindex >= NUPDE) {
1449 * Remove a PD page from the PDP and drop the hold count
1450 * on the PDP. The PDP is left cached in the pmap if
1451 * the hold count drops to 0 so the wire count remains
1458 m3 = vm_page_lookup(pmap->pm_pteobj,
1459 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1460 KKASSERT(m3 != NULL);
1461 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1462 idx = (p->pindex - NUPDE) % NPDPEPG;
1463 KKASSERT(pdp[idx] != 0);
1468 * Remove a PT page from the PD and drop the hold count
1469 * on the PD. The PD is left cached in the pmap if
1470 * the hold count drops to 0 so the wire count remains
1477 m2 = vm_page_lookup(pmap->pm_pteobj,
1478 NUPDE + p->pindex / NPDEPG);
1479 KKASSERT(m2 != NULL);
1480 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1481 idx = p->pindex % NPDEPG;
1487 * One fewer mappings in the pmap. p's hold count had better
1490 KKASSERT(pmap->pm_stats.resident_count > 0);
1491 --pmap->pm_stats.resident_count;
1493 panic("pmap_release: freeing held page table page");
1494 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1495 pmap->pm_ptphint = NULL;
1498 * We leave the top-level page table page cached, wired, and mapped in
1499 * the pmap until the dtor function (pmap_puninit()) gets called.
1500 * However, still clean it up so we can set PG_ZERO.
1502 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1503 bzero(pmap->pm_pml4, PAGE_SIZE);
1504 vm_page_flag_set(p, PG_ZERO);
1508 KKASSERT(p->wire_count == 0);
1509 vmstats.v_wire_count--;
1510 /* JG eventually revert to using vm_page_free_zero() */
1517 * This routine is called when various levels in the page table need to
1518 * be populated. This routine cannot fail.
1522 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1527 * Find or fabricate a new pagetable page. This will busy the page.
1529 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1530 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1531 if ((m->flags & PG_ZERO) == 0) {
1532 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1535 KASSERT(m->queue == PQ_NONE,
1536 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1539 * Increment the hold count for the page we will be returning to
1543 if (m->wire_count++ == 0)
1544 vmstats.v_wire_count++;
1547 * Map the pagetable page into the process address space, if
1548 * it isn't already there.
1550 * It is possible that someone else got in and mapped the page
1551 * directory page while we were blocked, if so just unbusy and
1552 * return the held page.
1554 if (ptepindex >= (NUPDE + NUPDPE)) {
1556 * Wire up a new PDP page in the PML4
1558 vm_pindex_t pml4index;
1561 pml4index = ptepindex - (NUPDE + NUPDPE);
1562 pml4 = &pmap->pm_pml4[pml4index];
1564 if (--m->wire_count == 0)
1565 --vmstats.v_wire_count;
1569 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1570 } else if (ptepindex >= NUPDE) {
1572 * Wire up a new PD page in the PDP
1574 vm_pindex_t pml4index;
1575 vm_pindex_t pdpindex;
1580 pdpindex = ptepindex - NUPDE;
1581 pml4index = pdpindex >> NPML4EPGSHIFT;
1583 pml4 = &pmap->pm_pml4[pml4index];
1584 if ((*pml4 & PG_V) == 0) {
1586 * Have to allocate a new PDP page, recurse.
1587 * This always succeeds. Returned page will
1590 pdppg = _pmap_allocpte(pmap,
1591 NUPDE + NUPDPE + pml4index);
1594 * Add a held reference to the PDP page.
1596 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1597 pdppg->hold_count++;
1601 * Now find the pdp_entry and map the PDP. If the PDP
1602 * has already been mapped unwind and return the
1603 * already-mapped PDP held.
1605 * pdppg is left held (hold_count is incremented for
1606 * each PD in the PDP).
1608 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1609 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1611 vm_page_unhold(pdppg);
1612 if (--m->wire_count == 0)
1613 --vmstats.v_wire_count;
1617 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1620 * Wire up the new PT page in the PD
1622 vm_pindex_t pml4index;
1623 vm_pindex_t pdpindex;
1629 pdpindex = ptepindex >> NPDPEPGSHIFT;
1630 pml4index = pdpindex >> NPML4EPGSHIFT;
1633 * Locate the PDP page in the PML4, then the PD page in
1634 * the PDP. If either does not exist we simply recurse
1637 * We can just recurse on the PD page as it will recurse
1638 * on the PDP if necessary.
1640 pml4 = &pmap->pm_pml4[pml4index];
1641 if ((*pml4 & PG_V) == 0) {
1642 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1643 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1644 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1646 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1647 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1648 if ((*pdp & PG_V) == 0) {
1649 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1651 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1657 * Now fill in the pte in the PD. If the pte already exists
1658 * (again, if we raced the grab), unhold pdpg and unwire
1659 * m, returning a held m.
1661 * pdpg is left held (hold_count is incremented for
1662 * each PT in the PD).
1664 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
1665 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1667 vm_page_unhold(pdpg);
1668 if (--m->wire_count == 0)
1669 --vmstats.v_wire_count;
1673 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1677 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1678 * valid bits, mapped flag, unbusy, and we're done.
1680 pmap->pm_ptphint = m;
1681 ++pmap->pm_stats.resident_count;
1683 m->valid = VM_PAGE_BITS_ALL;
1684 vm_page_flag_clear(m, PG_ZERO);
1685 vm_page_flag_set(m, PG_MAPPED);
1693 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1695 vm_pindex_t ptepindex;
1700 * Calculate pagetable page index
1702 ptepindex = pmap_pde_pindex(va);
1705 * Get the page directory entry
1707 pd = pmap_pde(pmap, va);
1710 * This supports switching from a 2MB page to a
1713 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1714 panic("no promotion/demotion yet");
1722 * If the page table page is mapped, we just increment the
1723 * hold count, and activate it.
1725 if (pd != NULL && (*pd & PG_V) != 0) {
1726 /* YYY hint is used here on i386 */
1727 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1728 pmap->pm_ptphint = m;
1733 * Here if the pte page isn't mapped, or if it has been deallocated.
1735 return _pmap_allocpte(pmap, ptepindex);
1739 /***************************************************
1740 * Pmap allocation/deallocation routines.
1741 ***************************************************/
1744 * Release any resources held by the given physical map.
1745 * Called when a pmap initialized by pmap_pinit is being released.
1746 * Should only be called if the map contains no valid mappings.
1748 static int pmap_release_callback(struct vm_page *p, void *data);
1751 pmap_release(struct pmap *pmap)
1753 vm_object_t object = pmap->pm_pteobj;
1754 struct rb_vm_page_scan_info info;
1756 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1757 #if defined(DIAGNOSTIC)
1758 if (object->ref_count != 1)
1759 panic("pmap_release: pteobj reference count != 1");
1763 info.object = object;
1765 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1772 info.limit = object->generation;
1774 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1775 pmap_release_callback, &info);
1776 if (info.error == 0 && info.mpte) {
1777 if (!pmap_release_free_page(pmap, info.mpte))
1781 } while (info.error);
1786 pmap_release_callback(struct vm_page *p, void *data)
1788 struct rb_vm_page_scan_info *info = data;
1790 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1794 if (!pmap_release_free_page(info->pmap, p)) {
1798 if (info->object->generation != info->limit) {
1806 * Grow the number of kernel page table entries, if needed.
1809 pmap_growkernel(vm_offset_t addr)
1812 vm_offset_t ptppaddr;
1814 pd_entry_t *pde, newpdir;
1818 if (kernel_vm_end == 0) {
1819 kernel_vm_end = KERNBASE;
1821 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1822 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1824 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1825 kernel_vm_end = kernel_map.max_offset;
1830 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1831 if (addr - 1 >= kernel_map.max_offset)
1832 addr = kernel_map.max_offset;
1833 while (kernel_vm_end < addr) {
1834 pde = pmap_pde(&kernel_pmap, kernel_vm_end);
1836 /* We need a new PDP entry */
1837 nkpg = vm_page_alloc(kptobj, nkpt,
1838 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM
1839 | VM_ALLOC_INTERRUPT);
1841 panic("pmap_growkernel: no memory to grow kernel");
1842 paddr = VM_PAGE_TO_PHYS(nkpg);
1843 if ((nkpg->flags & PG_ZERO) == 0)
1844 pmap_zero_page(paddr);
1845 vm_page_flag_clear(nkpg, PG_ZERO);
1846 newpdp = (pdp_entry_t)
1847 (paddr | PG_V | PG_RW | PG_A | PG_M);
1848 *pmap_pdpe(&kernel_pmap, kernel_vm_end) = newpdp;
1850 continue; /* try again */
1852 if ((*pde & PG_V) != 0) {
1853 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1854 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1855 kernel_vm_end = kernel_map.max_offset;
1862 * This index is bogus, but out of the way
1864 nkpg = vm_page_alloc(kptobj, nkpt,
1865 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1867 panic("pmap_growkernel: no memory to grow kernel");
1870 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1871 pmap_zero_page(ptppaddr);
1872 vm_page_flag_clear(nkpg, PG_ZERO);
1873 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1874 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1877 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1878 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1879 kernel_vm_end = kernel_map.max_offset;
1887 * Retire the given physical map from service.
1888 * Should only be called if the map contains
1889 * no valid mappings.
1892 pmap_destroy(pmap_t pmap)
1899 count = --pmap->pm_count;
1902 panic("destroying a pmap is not yet implemented");
1907 * Add a reference to the specified pmap.
1910 pmap_reference(pmap_t pmap)
1917 /***************************************************
1918 * page management routines.
1919 ***************************************************/
1922 * free the pv_entry back to the free list. This function may be
1923 * called from an interrupt.
1927 free_pv_entry(pv_entry_t pv)
1930 KKASSERT(pv_entry_count >= 0);
1935 * get a new pv_entry, allocating a block from the system
1936 * when needed. This function may be called from an interrupt.
1943 if (pv_entry_high_water &&
1944 (pv_entry_count > pv_entry_high_water) &&
1945 (pmap_pagedaemon_waken == 0)) {
1946 pmap_pagedaemon_waken = 1;
1947 wakeup(&vm_pages_needed);
1949 return zalloc(pvzone);
1953 * This routine is very drastic, but can save the system
1961 static int warningdone=0;
1963 if (pmap_pagedaemon_waken == 0)
1966 if (warningdone < 5) {
1967 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1971 for(i = 0; i < vm_page_array_size; i++) {
1972 m = &vm_page_array[i];
1973 if (m->wire_count || m->hold_count || m->busy ||
1974 (m->flags & PG_BUSY))
1978 pmap_pagedaemon_waken = 0;
1983 * If it is the first entry on the list, it is actually
1984 * in the header and we must copy the following entry up
1985 * to the header. Otherwise we must search the list for
1986 * the entry. In either case we free the now unused entry.
1990 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1991 vm_offset_t va, pmap_inval_info_t info)
1997 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1998 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1999 if (pmap == pv->pv_pmap && va == pv->pv_va)
2003 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2004 if (va == pv->pv_va)
2012 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2013 m->md.pv_list_count--;
2014 KKASSERT(m->md.pv_list_count >= 0);
2015 if (TAILQ_EMPTY(&m->md.pv_list))
2016 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2017 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2018 ++pmap->pm_generation;
2019 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
2027 * Create a pv entry for page at pa for
2032 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2037 pv = get_pv_entry();
2042 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2043 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2044 ++pmap->pm_generation;
2045 m->md.pv_list_count++;
2051 * pmap_remove_pte: do the things to unmap a page in a process
2055 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2056 pmap_inval_info_t info)
2061 pmap_inval_add(info, pmap, va);
2062 oldpte = pte_load_clear(ptq);
2064 pmap->pm_stats.wired_count -= 1;
2066 * Machines that don't support invlpg, also don't support
2067 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2071 cpu_invlpg((void *)va);
2072 KKASSERT(pmap->pm_stats.resident_count > 0);
2073 --pmap->pm_stats.resident_count;
2074 if (oldpte & PG_MANAGED) {
2075 m = PHYS_TO_VM_PAGE(oldpte);
2076 if (oldpte & PG_M) {
2077 #if defined(PMAP_DIAGNOSTIC)
2078 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2080 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2084 if (pmap_track_modified(va))
2088 vm_page_flag_set(m, PG_REFERENCED);
2089 return pmap_remove_entry(pmap, m, va, info);
2091 return pmap_unuse_pt(pmap, va, NULL, info);
2100 * Remove a single page from a process address space.
2102 * This function may not be called from an interrupt if the pmap is
2107 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2111 pte = pmap_pte(pmap, va);
2114 if ((*pte & PG_V) == 0)
2116 pmap_remove_pte(pmap, pte, va, info);
2122 * Remove the given range of addresses from the specified map.
2124 * It is assumed that the start and end are properly
2125 * rounded to the page size.
2127 * This function may not be called from an interrupt if the pmap is
2131 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2133 vm_offset_t va_next;
2134 pml4_entry_t *pml4e;
2136 pd_entry_t ptpaddr, *pde;
2138 struct pmap_inval_info info;
2143 if (pmap->pm_stats.resident_count == 0)
2146 pmap_inval_init(&info);
2149 * special handling of removing one page. a very
2150 * common operation and easy to short circuit some
2153 if (sva + PAGE_SIZE == eva) {
2154 pde = pmap_pde(pmap, sva);
2155 if (pde && (*pde & PG_PS) == 0) {
2156 pmap_remove_page(pmap, sva, &info);
2157 pmap_inval_flush(&info);
2162 for (; sva < eva; sva = va_next) {
2163 pml4e = pmap_pml4e(pmap, sva);
2164 if ((*pml4e & PG_V) == 0) {
2165 va_next = (sva + NBPML4) & ~PML4MASK;
2171 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2172 if ((*pdpe & PG_V) == 0) {
2173 va_next = (sva + NBPDP) & ~PDPMASK;
2180 * Calculate index for next page table.
2182 va_next = (sva + NBPDR) & ~PDRMASK;
2186 pde = pmap_pdpe_to_pde(pdpe, sva);
2190 * Weed out invalid mappings.
2196 * Check for large page.
2198 if ((ptpaddr & PG_PS) != 0) {
2199 /* JG FreeBSD has more complex treatment here */
2200 pmap_inval_add(&info, pmap, -1);
2202 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2207 * Limit our scan to either the end of the va represented
2208 * by the current page table page, or to the end of the
2209 * range being removed.
2215 * NOTE: pmap_remove_pte() can block.
2217 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2221 if (pmap_remove_pte(pmap, pte, sva, &info))
2225 pmap_inval_flush(&info);
2231 * Removes this physical page from all physical maps in which it resides.
2232 * Reflects back modify bits to the pager.
2234 * This routine may not be called from an interrupt.
2239 pmap_remove_all(vm_page_t m)
2241 struct pmap_inval_info info;
2242 pt_entry_t *pte, tpte;
2245 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2248 pmap_inval_init(&info);
2250 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2251 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2252 --pv->pv_pmap->pm_stats.resident_count;
2254 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2255 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2256 tpte = pte_load_clear(pte);
2259 pv->pv_pmap->pm_stats.wired_count--;
2262 vm_page_flag_set(m, PG_REFERENCED);
2265 * Update the vm_page_t clean and reference bits.
2268 #if defined(PMAP_DIAGNOSTIC)
2269 if (pmap_nw_modified(tpte)) {
2271 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2275 if (pmap_track_modified(pv->pv_va))
2278 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2279 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2280 ++pv->pv_pmap->pm_generation;
2281 m->md.pv_list_count--;
2282 KKASSERT(m->md.pv_list_count >= 0);
2283 if (TAILQ_EMPTY(&m->md.pv_list))
2284 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2285 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2289 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2290 pmap_inval_flush(&info);
2296 * Set the physical protection on the specified range of this map
2299 * This function may not be called from an interrupt if the map is
2300 * not the kernel_pmap.
2303 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2305 vm_offset_t va_next;
2306 pml4_entry_t *pml4e;
2308 pd_entry_t ptpaddr, *pde;
2310 pmap_inval_info info;
2312 /* JG review for NX */
2317 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2318 pmap_remove(pmap, sva, eva);
2322 if (prot & VM_PROT_WRITE)
2325 pmap_inval_init(&info);
2327 for (; sva < eva; sva = va_next) {
2329 pml4e = pmap_pml4e(pmap, sva);
2330 if ((*pml4e & PG_V) == 0) {
2331 va_next = (sva + NBPML4) & ~PML4MASK;
2337 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2338 if ((*pdpe & PG_V) == 0) {
2339 va_next = (sva + NBPDP) & ~PDPMASK;
2345 va_next = (sva + NBPDR) & ~PDRMASK;
2349 pde = pmap_pdpe_to_pde(pdpe, sva);
2353 * Check for large page.
2355 if ((ptpaddr & PG_PS) != 0) {
2356 pmap_inval_add(&info, pmap, -1);
2357 *pde &= ~(PG_M|PG_RW);
2358 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2363 * Weed out invalid mappings. Note: we assume that the page
2364 * directory table is always allocated, and in kernel virtual.
2372 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2374 pt_entry_t obits, pbits;
2378 * XXX non-optimal. Note also that there can be
2379 * no pmap_inval_flush() calls until after we modify
2380 * ptbase[sindex] (or otherwise we have to do another
2381 * pmap_inval_add() call).
2383 pmap_inval_add(&info, pmap, sva);
2384 obits = pbits = *pte;
2385 if ((pbits & PG_V) == 0)
2387 if (pbits & PG_MANAGED) {
2390 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2391 vm_page_flag_set(m, PG_REFERENCED);
2395 if (pmap_track_modified(sva)) {
2397 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2406 if (pbits != obits) {
2411 pmap_inval_flush(&info);
2415 * Insert the given physical page (p) at
2416 * the specified virtual address (v) in the
2417 * target physical map with the protection requested.
2419 * If specified, the page will be wired down, meaning
2420 * that the related pte can not be reclaimed.
2422 * NB: This is the only routine which MAY NOT lazy-evaluate
2423 * or lose information. That is, this routine must actually
2424 * insert this page into the given map NOW.
2427 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2434 pt_entry_t origpte, newpte;
2436 pmap_inval_info info;
2441 va = trunc_page(va);
2442 #ifdef PMAP_DIAGNOSTIC
2444 panic("pmap_enter: toobig");
2445 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2446 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
2448 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2449 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2451 db_print_backtrace();
2454 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2455 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2457 db_print_backtrace();
2462 * In the case that a page table page is not
2463 * resident, we are creating it here.
2465 if (va < VM_MAX_USER_ADDRESS)
2466 mpte = pmap_allocpte(pmap, va);
2470 pmap_inval_init(&info);
2471 pde = pmap_pde(pmap, va);
2472 if (pde != NULL && (*pde & PG_V) != 0) {
2473 if ((*pde & PG_PS) != 0)
2474 panic("pmap_enter: attempted pmap_enter on 2MB page");
2475 pte = pmap_pde_to_pte(pde, va);
2477 panic("pmap_enter: invalid page directory va=%#lx", va);
2479 KKASSERT(pte != NULL);
2480 pa = VM_PAGE_TO_PHYS(m);
2482 opa = origpte & PG_FRAME;
2485 * Mapping has not changed, must be protection or wiring change.
2487 if (origpte && (opa == pa)) {
2489 * Wiring change, just update stats. We don't worry about
2490 * wiring PT pages as they remain resident as long as there
2491 * are valid mappings in them. Hence, if a user page is wired,
2492 * the PT page will be also.
2494 if (wired && ((origpte & PG_W) == 0))
2495 pmap->pm_stats.wired_count++;
2496 else if (!wired && (origpte & PG_W))
2497 pmap->pm_stats.wired_count--;
2499 #if defined(PMAP_DIAGNOSTIC)
2500 if (pmap_nw_modified(origpte)) {
2502 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2508 * Remove the extra pte reference. Note that we cannot
2509 * optimize the RO->RW case because we have adjusted the
2510 * wiring count above and may need to adjust the wiring
2517 * We might be turning off write access to the page,
2518 * so we go ahead and sense modify status.
2520 if (origpte & PG_MANAGED) {
2521 if ((origpte & PG_M) && pmap_track_modified(va)) {
2523 om = PHYS_TO_VM_PAGE(opa);
2527 KKASSERT(m->flags & PG_MAPPED);
2532 * Mapping has changed, invalidate old range and fall through to
2533 * handle validating new mapping.
2537 err = pmap_remove_pte(pmap, pte, va, &info);
2539 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2541 opa = origpte & PG_FRAME;
2543 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2549 * Enter on the PV list if part of our managed memory. Note that we
2550 * raise IPL while manipulating pv_table since pmap_enter can be
2551 * called at interrupt time.
2553 if (pmap_initialized &&
2554 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2555 pmap_insert_entry(pmap, va, mpte, m);
2557 vm_page_flag_set(m, PG_MAPPED);
2561 * Increment counters
2563 ++pmap->pm_stats.resident_count;
2565 pmap->pm_stats.wired_count++;
2569 * Now validate mapping with desired protection/wiring.
2571 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
2575 if (va < VM_MAX_USER_ADDRESS)
2577 if (pmap == &kernel_pmap)
2581 * if the mapping or permission bits are different, we need
2582 * to update the pte.
2584 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2585 pmap_inval_add(&info, pmap, va);
2586 *pte = newpte | PG_A;
2588 vm_page_flag_set(m, PG_WRITEABLE);
2590 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2591 pmap_inval_flush(&info);
2595 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2596 * This code also assumes that the pmap has no pre-existing entry for this
2599 * This code currently may only be used on user pmaps, not kernel_pmap.
2602 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2607 vm_pindex_t ptepindex;
2609 pmap_inval_info info;
2611 pmap_inval_init(&info);
2613 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2614 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2616 db_print_backtrace();
2619 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2620 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2622 db_print_backtrace();
2626 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2629 * Calculate the page table page (mpte), allocating it if necessary.
2631 * A held page table page (mpte), or NULL, is passed onto the
2632 * section following.
2634 if (va < VM_MAX_USER_ADDRESS) {
2636 * Calculate pagetable page index
2638 ptepindex = pmap_pde_pindex(va);
2642 * Get the page directory entry
2644 ptepa = pmap_pde(pmap, va);
2647 * If the page table page is mapped, we just increment
2648 * the hold count, and activate it.
2650 if (ptepa && (*ptepa & PG_V) != 0) {
2652 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2653 // if (pmap->pm_ptphint &&
2654 // (pmap->pm_ptphint->pindex == ptepindex)) {
2655 // mpte = pmap->pm_ptphint;
2657 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2658 pmap->pm_ptphint = mpte;
2663 mpte = _pmap_allocpte(pmap, ptepindex);
2665 } while (mpte == NULL);
2668 /* this code path is not yet used */
2672 * With a valid (and held) page directory page, we can just use
2673 * vtopte() to get to the pte. If the pte is already present
2674 * we do not disturb it.
2679 pmap_unwire_pte_hold(pmap, va, mpte, &info);
2680 pa = VM_PAGE_TO_PHYS(m);
2681 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2686 * Enter on the PV list if part of our managed memory
2688 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2689 pmap_insert_entry(pmap, va, mpte, m);
2690 vm_page_flag_set(m, PG_MAPPED);
2694 * Increment counters
2696 ++pmap->pm_stats.resident_count;
2698 pa = VM_PAGE_TO_PHYS(m);
2701 * Now validate mapping with RO protection
2703 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2704 *pte = pa | PG_V | PG_U;
2706 *pte = pa | PG_V | PG_U | PG_MANAGED;
2707 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2708 pmap_inval_flush(&info);
2712 * Make a temporary mapping for a physical address. This is only intended
2713 * to be used for panic dumps.
2715 /* JG Needed on x86_64? */
2717 pmap_kenter_temporary(vm_paddr_t pa, int i)
2719 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2720 return ((void *)crashdumpmap);
2723 #define MAX_INIT_PT (96)
2726 * This routine preloads the ptes for a given object into the specified pmap.
2727 * This eliminates the blast of soft faults on process startup and
2728 * immediately after an mmap.
2730 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2733 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2734 vm_object_t object, vm_pindex_t pindex,
2735 vm_size_t size, int limit)
2737 struct rb_vm_page_scan_info info;
2742 * We can't preinit if read access isn't set or there is no pmap
2745 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2749 * We can't preinit if the pmap is not the current pmap
2751 lp = curthread->td_lwp;
2752 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2755 psize = x86_64_btop(size);
2757 if ((object->type != OBJT_VNODE) ||
2758 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2759 (object->resident_page_count > MAX_INIT_PT))) {
2763 if (psize + pindex > object->size) {
2764 if (object->size < pindex)
2766 psize = object->size - pindex;
2773 * Use a red-black scan to traverse the requested range and load
2774 * any valid pages found into the pmap.
2776 * We cannot safely scan the object's memq unless we are in a
2777 * critical section since interrupts can remove pages from objects.
2779 info.start_pindex = pindex;
2780 info.end_pindex = pindex + psize - 1;
2787 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2788 pmap_object_init_pt_callback, &info);
2794 pmap_object_init_pt_callback(vm_page_t p, void *data)
2796 struct rb_vm_page_scan_info *info = data;
2797 vm_pindex_t rel_index;
2799 * don't allow an madvise to blow away our really
2800 * free pages allocating pv entries.
2802 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2803 vmstats.v_free_count < vmstats.v_free_reserved) {
2806 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2807 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2808 if ((p->queue - p->pc) == PQ_CACHE)
2809 vm_page_deactivate(p);
2811 rel_index = p->pindex - info->start_pindex;
2812 pmap_enter_quick(info->pmap,
2813 info->addr + x86_64_ptob(rel_index), p);
2820 * Return TRUE if the pmap is in shape to trivially
2821 * pre-fault the specified address.
2823 * Returns FALSE if it would be non-trivial or if a
2824 * pte is already loaded into the slot.
2827 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2832 pde = pmap_pde(pmap, addr);
2833 if (pde == NULL || *pde == 0)
2844 * Routine: pmap_change_wiring
2845 * Function: Change the wiring attribute for a map/virtual-address
2847 * In/out conditions:
2848 * The mapping must already exist in the pmap.
2851 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2858 pte = pmap_pte(pmap, va);
2860 if (wired && !pmap_pte_w(pte))
2861 pmap->pm_stats.wired_count++;
2862 else if (!wired && pmap_pte_w(pte))
2863 pmap->pm_stats.wired_count--;
2866 * Wiring is not a hardware characteristic so there is no need to
2867 * invalidate TLB. However, in an SMP environment we must use
2868 * a locked bus cycle to update the pte (if we are not using
2869 * the pmap_inval_*() API that is)... it's ok to do this for simple
2874 atomic_set_long(pte, PG_W);
2876 atomic_clear_long(pte, PG_W);
2879 atomic_set_long_nonlocked(pte, PG_W);
2881 atomic_clear_long_nonlocked(pte, PG_W);
2888 * Copy the range specified by src_addr/len
2889 * from the source map to the range dst_addr/len
2890 * in the destination map.
2892 * This routine is only advisory and need not do anything.
2895 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2896 vm_size_t len, vm_offset_t src_addr)
2900 pmap_inval_info info;
2902 vm_offset_t end_addr = src_addr + len;
2904 pd_entry_t src_frame, dst_frame;
2907 if (dst_addr != src_addr)
2910 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
2911 if (src_frame != (PTDpde & PG_FRAME)) {
2915 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
2916 if (dst_frame != (APTDpde & PG_FRAME)) {
2917 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2918 /* The page directory is not shared between CPUs */
2922 pmap_inval_init(&info);
2923 pmap_inval_add(&info, dst_pmap, -1);
2924 pmap_inval_add(&info, src_pmap, -1);
2927 * critical section protection is required to maintain the page/object
2928 * association, interrupts can free pages and remove them from
2932 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2933 pt_entry_t *src_pte, *dst_pte;
2934 vm_page_t dstmpte, srcmpte;
2935 vm_offset_t srcptepaddr;
2936 vm_pindex_t ptepindex;
2938 if (addr >= UPT_MIN_ADDRESS)
2939 panic("pmap_copy: invalid to pmap_copy page tables\n");
2942 * Don't let optional prefaulting of pages make us go
2943 * way below the low water mark of free pages or way
2944 * above high water mark of used pv entries.
2946 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2947 pv_entry_count > pv_entry_high_water)
2950 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2951 ptepindex = addr >> PDRSHIFT;
2954 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2956 if (srcptepaddr == 0)
2959 if (srcptepaddr & PG_PS) {
2961 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2962 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2963 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2969 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2970 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2971 (srcmpte->flags & PG_BUSY)) {
2975 if (pdnxt > end_addr)
2978 src_pte = vtopte(addr);
2980 dst_pte = avtopte(addr);
2982 while (addr < pdnxt) {
2987 * we only virtual copy managed pages
2989 if ((ptetemp & PG_MANAGED) != 0) {
2991 * We have to check after allocpte for the
2992 * pte still being around... allocpte can
2995 * pmap_allocpte() can block. If we lose
2996 * our page directory mappings we stop.
2998 dstmpte = pmap_allocpte(dst_pmap, addr);
3001 if (src_frame != (PTDpde & PG_FRAME) ||
3002 dst_frame != (APTDpde & PG_FRAME)
3004 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3005 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3007 } else if ((*dst_pte == 0) &&
3008 (ptetemp = *src_pte) != 0 &&
3009 (ptetemp & PG_MANAGED)) {
3011 * Clear the modified and
3012 * accessed (referenced) bits
3015 m = PHYS_TO_VM_PAGE(ptetemp);
3016 *dst_pte = ptetemp & ~(PG_M | PG_A);
3017 ++dst_pmap->pm_stats.resident_count;
3018 pmap_insert_entry(dst_pmap, addr,
3020 KKASSERT(m->flags & PG_MAPPED);
3022 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3023 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3027 if (dstmpte->hold_count >= srcmpte->hold_count)
3037 pmap_inval_flush(&info);
3044 * Zero the specified physical page.
3046 * This function may be called from an interrupt and no locking is
3050 pmap_zero_page(vm_paddr_t phys)
3052 vm_offset_t va = PHYS_TO_DMAP(phys);
3054 pagezero((void *)va);
3058 * pmap_page_assertzero:
3060 * Assert that a page is empty, panic if it isn't.
3063 pmap_page_assertzero(vm_paddr_t phys)
3065 vm_offset_t virt = PHYS_TO_DMAP(phys);
3068 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3069 if (*(long *)((char *)virt + i) != 0) {
3070 panic("pmap_page_assertzero() @ %p not zero!\n", (void *)virt);
3078 * Zero part of a physical page by mapping it into memory and clearing
3079 * its contents with bzero.
3081 * off and size may not cover an area beyond a single hardware page.
3084 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3086 vm_offset_t virt = PHYS_TO_DMAP(phys);
3088 bzero((char *)virt + off, size);
3094 * Copy the physical page from the source PA to the target PA.
3095 * This function may be called from an interrupt. No locking
3099 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3101 vm_offset_t src_virt, dst_virt;
3103 src_virt = PHYS_TO_DMAP(src);
3104 dst_virt = PHYS_TO_DMAP(dst);
3105 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3109 * pmap_copy_page_frag:
3111 * Copy the physical page from the source PA to the target PA.
3112 * This function may be called from an interrupt. No locking
3116 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3118 vm_offset_t src_virt, dst_virt;
3120 src_virt = PHYS_TO_DMAP(src);
3121 dst_virt = PHYS_TO_DMAP(dst);
3123 bcopy((char *)src_virt + (src & PAGE_MASK),
3124 (char *)dst_virt + (dst & PAGE_MASK),
3129 * Returns true if the pmap's pv is one of the first
3130 * 16 pvs linked to from this page. This count may
3131 * be changed upwards or downwards in the future; it
3132 * is only necessary that true be returned for a small
3133 * subset of pmaps for proper page aging.
3136 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3141 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3146 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3147 if (pv->pv_pmap == pmap) {
3160 * Remove all pages from specified address space
3161 * this aids process exit speeds. Also, this code
3162 * is special cased for current process only, but
3163 * can have the more generic (and slightly slower)
3164 * mode enabled. This is much faster than pmap_remove
3165 * in the case of running down an entire address space.
3168 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3171 pt_entry_t *pte, tpte;
3174 pmap_inval_info info;
3176 int save_generation;
3178 lp = curthread->td_lwp;
3179 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3184 pmap_inval_init(&info);
3186 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3187 if (pv->pv_va >= eva || pv->pv_va < sva) {
3188 npv = TAILQ_NEXT(pv, pv_plist);
3192 KKASSERT(pmap == pv->pv_pmap);
3195 pte = vtopte(pv->pv_va);
3197 pte = pmap_pte_quick(pmap, pv->pv_va);
3198 if (pmap->pm_active)
3199 pmap_inval_add(&info, pmap, pv->pv_va);
3202 * We cannot remove wired pages from a process' mapping
3206 npv = TAILQ_NEXT(pv, pv_plist);
3209 tpte = pte_load_clear(pte);
3211 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3213 KASSERT(m < &vm_page_array[vm_page_array_size],
3214 ("pmap_remove_pages: bad tpte %lx", tpte));
3216 KKASSERT(pmap->pm_stats.resident_count > 0);
3217 --pmap->pm_stats.resident_count;
3220 * Update the vm_page_t clean and reference bits.
3226 npv = TAILQ_NEXT(pv, pv_plist);
3227 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3228 save_generation = ++pmap->pm_generation;
3230 m->md.pv_list_count--;
3231 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3232 if (TAILQ_EMPTY(&m->md.pv_list))
3233 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3235 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3239 * Restart the scan if we blocked during the unuse or free
3240 * calls and other removals were made.
3242 if (save_generation != pmap->pm_generation) {
3243 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3244 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3247 pmap_inval_flush(&info);
3252 * pmap_testbit tests bits in pte's
3253 * note that the testbit/clearbit routines are inline,
3254 * and a lot of things compile-time evaluate.
3258 pmap_testbit(vm_page_t m, int bit)
3263 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3266 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3271 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3273 * if the bit being tested is the modified bit, then
3274 * mark clean_map and ptes as never
3277 if (bit & (PG_A|PG_M)) {
3278 if (!pmap_track_modified(pv->pv_va))
3282 #if defined(PMAP_DIAGNOSTIC)
3283 if (pv->pv_pmap == NULL) {
3284 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3288 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3299 * this routine is used to modify bits in ptes
3303 pmap_clearbit(vm_page_t m, int bit)
3305 struct pmap_inval_info info;
3310 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3313 pmap_inval_init(&info);
3317 * Loop over all current mappings setting/clearing as appropos If
3318 * setting RO do we need to clear the VAC?
3320 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3322 * don't write protect pager mappings
3325 if (!pmap_track_modified(pv->pv_va))
3329 #if defined(PMAP_DIAGNOSTIC)
3330 if (pv->pv_pmap == NULL) {
3331 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3337 * Careful here. We can use a locked bus instruction to
3338 * clear PG_A or PG_M safely but we need to synchronize
3339 * with the target cpus when we mess with PG_RW.
3341 * We do not have to force synchronization when clearing
3342 * PG_M even for PTEs generated via virtual memory maps,
3343 * because the virtual kernel will invalidate the pmap
3344 * entry when/if it needs to resynchronize the Modify bit.
3347 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
3348 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3355 atomic_clear_long(pte, PG_M|PG_RW);
3358 * The cpu may be trying to set PG_M
3359 * simultaniously with our clearing
3362 if (!atomic_cmpset_long(pte, pbits,
3366 } else if (bit == PG_M) {
3368 * We could also clear PG_RW here to force
3369 * a fault on write to redetect PG_M for
3370 * virtual kernels, but it isn't necessary
3371 * since virtual kernels invalidate the pte
3372 * when they clear the VPTE_M bit in their
3373 * virtual page tables.
3375 atomic_clear_long(pte, PG_M);
3377 atomic_clear_long(pte, bit);
3381 pmap_inval_flush(&info);
3386 * pmap_page_protect:
3388 * Lower the permission for all mappings to a given page.
3391 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3393 /* JG NX support? */
3394 if ((prot & VM_PROT_WRITE) == 0) {
3395 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3396 pmap_clearbit(m, PG_RW);
3397 vm_page_flag_clear(m, PG_WRITEABLE);
3405 pmap_phys_address(vm_pindex_t ppn)
3407 return (x86_64_ptob(ppn));
3411 * pmap_ts_referenced:
3413 * Return a count of reference bits for a page, clearing those bits.
3414 * It is not necessary for every reference bit to be cleared, but it
3415 * is necessary that 0 only be returned when there are truly no
3416 * reference bits set.
3418 * XXX: The exact number of bits to check and clear is a matter that
3419 * should be tested and standardized at some point in the future for
3420 * optimal aging of shared pages.
3423 pmap_ts_referenced(vm_page_t m)
3425 pv_entry_t pv, pvf, pvn;
3429 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3434 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3439 pvn = TAILQ_NEXT(pv, pv_list);
3442 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3443 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3446 if (!pmap_track_modified(pv->pv_va))
3449 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3451 if (pte && (*pte & PG_A)) {
3453 atomic_clear_long(pte, PG_A);
3455 atomic_clear_long_nonlocked(pte, PG_A);
3462 } while ((pv = pvn) != NULL && pv != pvf);
3472 * Return whether or not the specified physical page was modified
3473 * in any physical maps.
3476 pmap_is_modified(vm_page_t m)
3478 return pmap_testbit(m, PG_M);
3482 * Clear the modify bits on the specified physical page.
3485 pmap_clear_modify(vm_page_t m)
3487 pmap_clearbit(m, PG_M);
3491 * pmap_clear_reference:
3493 * Clear the reference bit on the specified physical page.
3496 pmap_clear_reference(vm_page_t m)
3498 pmap_clearbit(m, PG_A);
3502 * Miscellaneous support routines follow
3507 i386_protection_init(void)
3511 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3512 kp = protection_codes;
3513 for (prot = 0; prot < 8; prot++) {
3515 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3517 * Read access is also 0. There isn't any execute bit,
3518 * so just make it readable.
3520 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3521 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3522 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3525 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3526 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3527 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3528 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3536 * Map a set of physical memory pages into the kernel virtual
3537 * address space. Return a pointer to where it is mapped. This
3538 * routine is intended to be used for mapping device memory,
3541 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3545 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3547 vm_offset_t va, tmpva, offset;
3550 offset = pa & PAGE_MASK;
3551 size = roundup(offset + size, PAGE_SIZE);
3553 va = kmem_alloc_nofault(&kernel_map, size);
3555 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3557 pa = pa & ~PAGE_MASK;
3558 for (tmpva = va; size > 0;) {
3559 pte = vtopte(tmpva);
3560 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3568 return ((void *)(va + offset));
3572 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3574 vm_offset_t va, tmpva, offset;
3577 offset = pa & PAGE_MASK;
3578 size = roundup(offset + size, PAGE_SIZE);
3580 va = kmem_alloc_nofault(&kernel_map, size);
3582 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3584 pa = pa & ~PAGE_MASK;
3585 for (tmpva = va; size > 0;) {
3586 pte = vtopte(tmpva);
3587 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3595 return ((void *)(va + offset));
3599 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3601 vm_offset_t base, offset;
3603 base = va & ~PAGE_MASK;
3604 offset = va & PAGE_MASK;
3605 size = roundup(offset + size, PAGE_SIZE);
3606 pmap_qremove(va, size >> PAGE_SHIFT);
3607 kmem_free(&kernel_map, base, size);
3611 * perform the pmap work for mincore
3614 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3616 pt_entry_t *ptep, pte;
3620 ptep = pmap_pte(pmap, addr);
3625 if ((pte = *ptep) != 0) {
3628 val = MINCORE_INCORE;
3629 if ((pte & PG_MANAGED) == 0)
3632 pa = pte & PG_FRAME;
3634 m = PHYS_TO_VM_PAGE(pa);
3640 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3642 * Modified by someone
3644 else if (m->dirty || pmap_is_modified(m))
3645 val |= MINCORE_MODIFIED_OTHER;
3650 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3653 * Referenced by someone
3655 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3656 val |= MINCORE_REFERENCED_OTHER;
3657 vm_page_flag_set(m, PG_REFERENCED);
3664 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3665 * vmspace will be ref'd and the old one will be deref'd.
3667 * The vmspace for all lwps associated with the process will be adjusted
3668 * and cr3 will be reloaded if any lwp is the current lwp.
3671 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3673 struct vmspace *oldvm;
3677 oldvm = p->p_vmspace;
3678 if (oldvm != newvm) {
3679 p->p_vmspace = newvm;
3680 KKASSERT(p->p_nthreads == 1);
3681 lp = RB_ROOT(&p->p_lwp_tree);
3682 pmap_setlwpvm(lp, newvm);
3684 sysref_get(&newvm->vm_sysref);
3685 sysref_put(&oldvm->vm_sysref);
3692 * Set the vmspace for a LWP. The vmspace is almost universally set the
3693 * same as the process vmspace, but virtual kernels need to swap out contexts
3694 * on a per-lwp basis.
3697 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3699 struct vmspace *oldvm;
3703 oldvm = lp->lwp_vmspace;
3705 if (oldvm != newvm) {
3706 lp->lwp_vmspace = newvm;
3707 if (curthread->td_lwp == lp) {
3708 pmap = vmspace_pmap(newvm);
3710 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3712 pmap->pm_active |= 1;
3714 #if defined(SWTCH_OPTIM_STATS)
3717 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
3718 load_cr3(curthread->td_pcb->pcb_cr3);
3719 pmap = vmspace_pmap(oldvm);
3721 atomic_clear_int(&pmap->pm_active,
3722 1 << mycpu->gd_cpuid);
3724 pmap->pm_active &= ~1;
3732 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3735 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3739 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3746 static void pads (pmap_t pm);
3747 void pmap_pvdump (vm_paddr_t pa);
3749 /* print address space of pmap*/
3758 if (pm == &kernel_pmap)
3761 for (i = 0; i < NPDEPG; i++) {
3769 pmap_pvdump(vm_paddr_t pa)
3774 kprintf("pa %08llx", (long long)pa);
3775 m = PHYS_TO_VM_PAGE(pa);
3776 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3778 kprintf(" -> pmap %p, va %x, flags %x",
3779 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3781 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);