2 * Copyright (c) 1991 Regents of the University of California.
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1994 David Greenman
5 * Copyright (c) 2003 Peter Wemm
6 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
7 * Copyright (c) 2008, 2009 The DragonFly Project.
8 * Copyright (c) 2008, 2009 Jordan Gordeev.
11 * This code is derived from software contributed to Berkeley by
12 * the Systems Programming Group of the University of Utah Computer
13 * Science Department and William Jolitz of UUNET Technologies Inc.
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
44 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
48 * Manages physical address maps.
50 * In addition to hardware address maps, this
51 * module is called upon to provide software-use-only
52 * maps which may or may not be stored in the same
53 * form as hardware maps. These pseudo-maps are
54 * used to store intermediate results from copy
55 * operations to and from address spaces.
57 * Since the information managed by this module is
58 * also stored by the logical address mapping module,
59 * this module may throw away valid virtual-to-physical
60 * mappings at almost any time. However, invalidations
61 * of virtual-to-physical mappings must be done as
64 * In order to cope with hardware architectures which
65 * make virtual-to-physical map invalidates expensive,
66 * this module may delay invalidate or reduced protection
67 * operations until such time as they are actually
68 * necessary. This module is given full information as
69 * to which processors are currently using which maps,
70 * and to when physical maps must be made correct.
74 #include "opt_disable_pse.h"
77 #include "opt_msgbuf.h"
79 #include <sys/param.h>
80 #include <sys/systm.h>
81 #include <sys/kernel.h>
83 #include <sys/msgbuf.h>
84 #include <sys/vmmeter.h>
88 #include <vm/vm_param.h>
89 #include <sys/sysctl.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_pageout.h>
97 #include <vm/vm_pager.h>
98 #include <vm/vm_zone.h>
100 #include <sys/user.h>
101 #include <sys/thread2.h>
102 #include <sys/sysref2.h>
104 #include <machine/cputypes.h>
105 #include <machine/md_var.h>
106 #include <machine/specialreg.h>
107 #include <machine/smp.h>
108 #include <machine_base/apic/apicreg.h>
109 #include <machine/globaldata.h>
110 #include <machine/pmap.h>
111 #include <machine/pmap_inval.h>
115 #define PMAP_KEEP_PDIRS
116 #ifndef PMAP_SHPGPERPROC
117 #define PMAP_SHPGPERPROC 200
120 #if defined(DIAGNOSTIC)
121 #define PMAP_DIAGNOSTIC
127 * Get PDEs and PTEs for user/kernel address space
129 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
130 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
132 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
133 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
134 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
135 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
136 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
140 * Given a map and a machine independent protection code,
141 * convert to a vax protection code.
143 #define pte_prot(m, p) \
144 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
145 static int protection_codes[8];
147 struct pmap kernel_pmap;
148 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
150 vm_paddr_t avail_start; /* PA of first available physical page */
151 vm_paddr_t avail_end; /* PA of last available physical page */
152 vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
153 vm_offset_t virtual2_end;
154 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
155 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
156 vm_offset_t KvaStart; /* VA start of KVA space */
157 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
158 vm_offset_t KvaSize; /* max size of kernel virtual address space */
159 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
160 static int pgeflag; /* PG_G or-in */
161 static int pseflag; /* PG_PS or-in */
163 static vm_object_t kptobj;
166 static vm_paddr_t dmaplimit;
168 vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
170 static uint64_t KPTbase;
171 static uint64_t KPTphys;
172 static uint64_t KPDphys; /* phys addr of kernel level 2 */
173 static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
174 uint64_t KPDPphys; /* phys addr of kernel level 3 */
175 uint64_t KPML4phys; /* phys addr of kernel level 4 */
177 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
178 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
181 * Data for the pv entry allocation mechanism
183 static vm_zone_t pvzone;
184 static struct vm_zone pvzone_store;
185 static struct vm_object pvzone_obj;
186 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
187 static int pmap_pagedaemon_waken = 0;
188 static struct pv_entry *pvinit;
191 * All those kernel PT submaps that BSD is so fond of
193 pt_entry_t *CMAP1 = 0, *ptmmap;
194 caddr_t CADDR1 = 0, ptvmmap = 0;
195 static pt_entry_t *msgbufmap;
196 struct msgbuf *msgbufp=0;
201 static pt_entry_t *pt_crashdumpmap;
202 static caddr_t crashdumpmap;
204 extern pt_entry_t *SMPpt;
205 extern uint64_t SMPptpa;
209 static pv_entry_t get_pv_entry (void);
210 static void i386_protection_init (void);
211 static void create_pagetables(vm_paddr_t *firstaddr);
212 static void pmap_remove_all (vm_page_t m);
213 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
214 vm_offset_t sva, pmap_inval_info_t info);
215 static void pmap_remove_page (struct pmap *pmap,
216 vm_offset_t va, pmap_inval_info_t info);
217 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
218 vm_offset_t va, pmap_inval_info_t info);
219 static boolean_t pmap_testbit (vm_page_t m, int bit);
220 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
221 vm_page_t mpte, vm_page_t m);
223 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
225 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
226 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
227 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
228 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
229 static int _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
230 pmap_inval_info_t info);
231 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
232 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
234 static unsigned pdir4mb;
237 * Move the kernel virtual free pointer to the next
238 * 2MB. This is used to help improve performance
239 * by using a large (2MB) page for much of the kernel
240 * (.text, .data, .bss)
244 pmap_kmem_choose(vm_offset_t addr)
246 vm_offset_t newaddr = addr;
248 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
255 * Super fast pmap_pte routine best used when scanning the pv lists.
256 * This eliminates many course-grained invltlb calls. Note that many of
257 * the pv list scans are across different pmaps and it is very wasteful
258 * to do an entire invltlb when checking a single mapping.
260 * Should only be called while in a critical section.
262 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
266 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
268 return pmap_pte(pmap, va);
271 /* Return a non-clipped PD index for a given VA */
274 pmap_pde_pindex(vm_offset_t va)
276 return va >> PDRSHIFT;
279 /* Return various clipped indexes for a given VA */
282 pmap_pte_index(vm_offset_t va)
285 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
290 pmap_pde_index(vm_offset_t va)
293 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
298 pmap_pdpe_index(vm_offset_t va)
301 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
306 pmap_pml4e_index(vm_offset_t va)
309 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
312 /* Return a pointer to the PML4 slot that corresponds to a VA */
315 pmap_pml4e(pmap_t pmap, vm_offset_t va)
318 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
321 /* Return a pointer to the PDP slot that corresponds to a VA */
324 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
328 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
329 return (&pdpe[pmap_pdpe_index(va)]);
332 /* Return a pointer to the PDP slot that corresponds to a VA */
335 pmap_pdpe(pmap_t pmap, vm_offset_t va)
339 pml4e = pmap_pml4e(pmap, va);
340 if ((*pml4e & PG_V) == 0)
342 return (pmap_pml4e_to_pdpe(pml4e, va));
345 /* Return a pointer to the PD slot that corresponds to a VA */
348 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
352 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
353 return (&pde[pmap_pde_index(va)]);
356 /* Return a pointer to the PD slot that corresponds to a VA */
359 pmap_pde(pmap_t pmap, vm_offset_t va)
363 pdpe = pmap_pdpe(pmap, va);
364 if (pdpe == NULL || (*pdpe & PG_V) == 0)
366 return (pmap_pdpe_to_pde(pdpe, va));
369 /* Return a pointer to the PT slot that corresponds to a VA */
372 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
376 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
377 return (&pte[pmap_pte_index(va)]);
380 /* Return a pointer to the PT slot that corresponds to a VA */
383 pmap_pte(pmap_t pmap, vm_offset_t va)
387 pde = pmap_pde(pmap, va);
388 if (pde == NULL || (*pde & PG_V) == 0)
390 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
391 return ((pt_entry_t *)pde);
392 return (pmap_pde_to_pte(pde, va));
397 vtopte(vm_offset_t va)
399 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
401 return (PTmap + ((va >> PAGE_SHIFT) & mask));
406 vtopde(vm_offset_t va)
408 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
410 return (PDmap + ((va >> PDRSHIFT) & mask));
414 allocpages(vm_paddr_t *firstaddr, int n)
419 bzero((void *)ret, n * PAGE_SIZE);
420 *firstaddr += n * PAGE_SIZE;
426 create_pagetables(vm_paddr_t *firstaddr)
430 /* we are running (mostly) V=P at this point */
433 KPTbase = allocpages(firstaddr, NKPT);
434 KPTphys = allocpages(firstaddr, NKPT);
435 KPML4phys = allocpages(firstaddr, 1);
436 KPDPphys = allocpages(firstaddr, NKPML4E);
439 * Calculate the page directory base for KERNBASE,
440 * that is where we start populating the page table pages.
441 * Basically this is the end - 2.
443 KPDphys = allocpages(firstaddr, NKPDPE);
444 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
446 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
447 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
449 DMPDPphys = allocpages(firstaddr, NDMPML4E);
450 if ((amd_feature & AMDID_PAGE1GB) == 0)
451 DMPDphys = allocpages(firstaddr, ndmpdp);
452 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
455 * Fill in the underlying page table pages for the area around
456 * KERNBASE. This remaps low physical memory to KERNBASE.
458 * Read-only from zero to physfree
459 * XXX not fully used, underneath 2M pages
461 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
462 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
463 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
467 * Now map the initial kernel page tables. One block of page
468 * tables is placed at the beginning of kernel virtual memory,
469 * and another block is placed at KERNBASE to map the kernel binary,
470 * data, bss, and initial pre-allocations.
472 for (i = 0; i < NKPT; i++) {
473 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
474 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
476 for (i = 0; i < NKPT; i++) {
477 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
478 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
482 * Map from zero to end of allocations using 2M pages as an
483 * optimization. This will bypass some of the KPTBase pages
484 * above in the KERNBASE area.
486 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
487 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
488 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
492 * And connect up the PD to the PDP. The kernel pmap is expected
493 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
495 for (i = 0; i < NKPDPE; i++) {
496 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
497 KPDphys + (i << PAGE_SHIFT);
498 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
502 /* Now set up the direct map space using either 2MB or 1GB pages */
503 /* Preset PG_M and PG_A because demotion expects it */
504 if ((amd_feature & AMDID_PAGE1GB) == 0) {
505 for (i = 0; i < NPDEPG * ndmpdp; i++) {
506 ((pd_entry_t *)DMPDphys)[i] = (vm_paddr_t)i << PDRSHIFT;
507 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
510 /* And the direct map space's PDP */
511 for (i = 0; i < ndmpdp; i++) {
512 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
514 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
517 for (i = 0; i < ndmpdp; i++) {
518 ((pdp_entry_t *)DMPDPphys)[i] =
519 (vm_paddr_t)i << PDPSHIFT;
520 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
525 /* And recursively map PML4 to itself in order to get PTmap */
526 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
527 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
529 /* Connect the Direct Map slot up to the PML4 */
530 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
531 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
533 /* Connect the KVA slot up to the PML4 */
534 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
535 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
539 init_paging(vm_paddr_t *firstaddr)
541 create_pagetables(firstaddr);
545 * Bootstrap the system enough to run with virtual memory.
547 * On the i386 this is called after mapping has already been enabled
548 * and just syncs the pmap module with what has already been done.
549 * [We can't call it easily with mapping off since the kernel is not
550 * mapped with PA == VA, hence we would have to relocate every address
551 * from the linked base (virtual) address "KERNBASE" to the actual
552 * (physical) address starting relative to 0]
555 pmap_bootstrap(vm_paddr_t *firstaddr)
559 struct mdglobaldata *gd;
562 KvaStart = VM_MIN_KERNEL_ADDRESS;
563 KvaEnd = VM_MAX_KERNEL_ADDRESS;
564 KvaSize = KvaEnd - KvaStart;
566 avail_start = *firstaddr;
569 * Create an initial set of page tables to run the kernel in.
571 create_pagetables(firstaddr);
573 virtual2_start = KvaStart;
574 virtual2_end = PTOV_OFFSET;
576 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
577 virtual_start = pmap_kmem_choose(virtual_start);
579 virtual_end = VM_MAX_KERNEL_ADDRESS;
581 /* XXX do %cr0 as well */
582 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
586 * Initialize protection array.
588 i386_protection_init();
591 * The kernel's pmap is statically allocated so we don't have to use
592 * pmap_create, which is unlikely to work correctly at this part of
593 * the boot sequence (XXX and which no longer exists).
595 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
596 kernel_pmap.pm_count = 1;
597 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
598 TAILQ_INIT(&kernel_pmap.pm_pvlist);
602 * Reserve some special page table entries/VA space for temporary
605 #define SYSMAP(c, p, v, n) \
606 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
610 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
616 * CMAP1/CMAP2 are used for zeroing and copying pages.
618 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
623 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
626 * ptvmmap is used for reading arbitrary physical pages via
629 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
632 * msgbufp is used to map the system message buffer.
633 * XXX msgbufmap is not used.
635 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
636 atop(round_page(MSGBUF_SIZE)))
643 * PG_G is terribly broken on SMP because we IPI invltlb's in some
644 * cases rather then invl1pg. Actually, I don't even know why it
645 * works under UP because self-referential page table mappings
650 if (cpu_feature & CPUID_PGE)
655 * Initialize the 4MB page size flag
659 * The 4MB page version of the initial
660 * kernel page mapping.
664 #if !defined(DISABLE_PSE)
665 if (cpu_feature & CPUID_PSE) {
668 * Note that we have enabled PSE mode
671 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
672 ptditmp &= ~(NBPDR - 1);
673 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
678 * Enable the PSE mode. If we are SMP we can't do this
679 * now because the APs will not be able to use it when
682 load_cr4(rcr4() | CR4_PSE);
685 * We can do the mapping here for the single processor
686 * case. We simply ignore the old page table page from
690 * For SMP, we still need 4K pages to bootstrap APs,
691 * PSE will be enabled as soon as all APs are up.
693 PTD[KPTDI] = (pd_entry_t)ptditmp;
699 if (cpu_apic_address == 0)
700 panic("pmap_bootstrap: no local apic!");
704 * We need to finish setting up the globaldata page for the BSP.
705 * locore has already populated the page table for the mdglobaldata
708 pg = MDGLOBALDATA_BASEALLOC_PAGES;
709 gd = &CPU_prvspace[0].mdglobaldata;
710 gd->gd_CMAP1 = &SMPpt[pg + 0];
711 gd->gd_CMAP2 = &SMPpt[pg + 1];
712 gd->gd_CMAP3 = &SMPpt[pg + 2];
713 gd->gd_PMAP1 = &SMPpt[pg + 3];
714 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
715 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
716 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
717 gd->gd_PADDR1 = (pt_entry_t *)CPU_prvspace[0].PPAGE1;
724 * Set 4mb pdir for mp startup
729 if (pseflag && (cpu_feature & CPUID_PSE)) {
730 load_cr4(rcr4() | CR4_PSE);
731 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
739 * Initialize the pmap module.
740 * Called by vm_init, to initialize any structures that the pmap
741 * system needs to map virtual memory.
742 * pmap_init has been enhanced to support in a fairly consistant
743 * way, discontiguous physical memory.
752 * object for kernel page table pages
754 /* JG I think the number can be arbitrary */
755 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
758 * Allocate memory for random pmap data structures. Includes the
762 for(i = 0; i < vm_page_array_size; i++) {
765 m = &vm_page_array[i];
766 TAILQ_INIT(&m->md.pv_list);
767 m->md.pv_list_count = 0;
771 * init the pv free list
773 initial_pvs = vm_page_array_size;
774 if (initial_pvs < MINPV)
776 pvzone = &pvzone_store;
777 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
778 initial_pvs * sizeof (struct pv_entry));
779 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
783 * Now it is safe to enable pv_table recording.
785 pmap_initialized = TRUE;
787 lapic = pmap_mapdev_uncacheable(cpu_apic_address, sizeof(struct LAPIC));
792 * Initialize the address space (zone) for the pv_entries. Set a
793 * high water mark so that the system can recover from excessive
794 * numbers of pv entries.
799 int shpgperproc = PMAP_SHPGPERPROC;
801 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
802 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
803 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
804 pv_entry_high_water = 9 * (pv_entry_max / 10);
805 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
809 /***************************************************
810 * Low level helper routines.....
811 ***************************************************/
813 #if defined(PMAP_DIAGNOSTIC)
816 * This code checks for non-writeable/modified pages.
817 * This should be an invalid condition.
821 pmap_nw_modified(pt_entry_t pte)
823 if ((pte & (PG_M|PG_RW)) == PG_M)
832 * this routine defines the region(s) of memory that should
833 * not be tested for the modified bit.
837 pmap_track_modified(vm_offset_t va)
839 if ((va < clean_sva) || (va >= clean_eva))
848 * Extract the physical page address associated with the map/VA pair.
850 * This function may not be called from an interrupt if the pmap is
854 pmap_extract(pmap_t pmap, vm_offset_t va)
858 pd_entry_t pde, *pdep;
861 pdep = pmap_pde(pmap, va);
865 if ((pde & PG_PS) != 0) {
866 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
868 pte = pmap_pde_to_pte(pdep, va);
869 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
877 * Routine: pmap_kextract
879 * Extract the physical page address associated
880 * kernel virtual address.
883 pmap_kextract(vm_offset_t va)
888 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
889 pa = DMAP_TO_PHYS(va);
893 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
896 * Beware of a concurrent promotion that changes the
897 * PDE at this point! For example, vtopte() must not
898 * be used to access the PTE because it would use the
899 * new PDE. It is, however, safe to use the old PDE
900 * because the page table page is preserved by the
903 pa = *pmap_pde_to_pte(&pde, va);
904 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
910 /***************************************************
911 * Low level mapping routines.....
912 ***************************************************/
915 * Routine: pmap_kenter
917 * Add a wired page to the KVA
918 * NOTE! note that in order for the mapping to take effect -- you
919 * should do an invltlb after doing the pmap_kenter().
922 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
926 pmap_inval_info info;
928 pmap_inval_init(&info);
929 npte = pa | PG_RW | PG_V | pgeflag;
931 pmap_inval_interlock(&info, &kernel_pmap, va);
933 pmap_inval_deinterlock(&info, &kernel_pmap);
934 pmap_inval_done(&info);
938 * Routine: pmap_kenter_quick
940 * Similar to pmap_kenter(), except we only invalidate the
941 * mapping on the current CPU.
944 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
949 npte = pa | PG_RW | PG_V | pgeflag;
952 cpu_invlpg((void *)va);
956 pmap_kenter_sync(vm_offset_t va)
958 pmap_inval_info info;
960 pmap_inval_init(&info);
961 pmap_inval_interlock(&info, &kernel_pmap, va);
962 pmap_inval_deinterlock(&info, &kernel_pmap);
963 pmap_inval_done(&info);
967 pmap_kenter_sync_quick(vm_offset_t va)
969 cpu_invlpg((void *)va);
973 * remove a page from the kernel pagetables
976 pmap_kremove(vm_offset_t va)
979 pmap_inval_info info;
981 pmap_inval_init(&info);
983 pmap_inval_interlock(&info, &kernel_pmap, va);
985 pmap_inval_deinterlock(&info, &kernel_pmap);
986 pmap_inval_done(&info);
990 pmap_kremove_quick(vm_offset_t va)
995 cpu_invlpg((void *)va);
999 * XXX these need to be recoded. They are not used in any critical path.
1002 pmap_kmodify_rw(vm_offset_t va)
1004 *vtopte(va) |= PG_RW;
1005 cpu_invlpg((void *)va);
1009 pmap_kmodify_nc(vm_offset_t va)
1011 *vtopte(va) |= PG_N;
1012 cpu_invlpg((void *)va);
1016 * Used to map a range of physical addresses into kernel
1017 * virtual address space.
1019 * For now, VM is already on, we only need to map the
1023 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
1025 return PHYS_TO_DMAP(start);
1030 * Add a list of wired pages to the kva
1031 * this routine is only used for temporary
1032 * kernel mappings that do not need to have
1033 * page modification or references recorded.
1034 * Note that old mappings are simply written
1035 * over. The page *must* be wired.
1038 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1042 end_va = va + count * PAGE_SIZE;
1044 while (va < end_va) {
1048 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1049 cpu_invlpg((void *)va);
1054 smp_invltlb(); /* XXX */
1059 * This routine jerks page mappings from the
1060 * kernel -- it is meant only for temporary mappings.
1062 * MPSAFE, INTERRUPT SAFE (cluster callback)
1065 pmap_qremove(vm_offset_t va, int count)
1069 end_va = va + count * PAGE_SIZE;
1071 while (va < end_va) {
1076 cpu_invlpg((void *)va);
1085 * This routine works like vm_page_lookup() but also blocks as long as the
1086 * page is busy. This routine does not busy the page it returns.
1088 * Unless the caller is managing objects whos pages are in a known state,
1089 * the call should be made with a critical section held so the page's object
1090 * association remains valid on return.
1094 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1099 m = vm_page_lookup(object, pindex);
1100 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
1106 * Create a new thread and optionally associate it with a (new) process.
1107 * NOTE! the new thread's cpu may not equal the current cpu.
1110 pmap_init_thread(thread_t td)
1112 /* enforce pcb placement */
1113 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1114 td->td_savefpu = &td->td_pcb->pcb_save;
1115 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on x86_64? */
1119 * This routine directly affects the fork perf for a process.
1122 pmap_init_proc(struct proc *p)
1127 * Dispose the UPAGES for a process that has exited.
1128 * This routine directly impacts the exit perf of a process.
1131 pmap_dispose_proc(struct proc *p)
1133 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1136 /***************************************************
1137 * Page table page management routines.....
1138 ***************************************************/
1141 * This routine unholds page table pages, and if the hold count
1142 * drops to zero, then it decrements the wire count.
1146 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1147 pmap_inval_info_t info)
1149 KKASSERT(m->hold_count > 0);
1150 if (m->hold_count > 1) {
1154 return _pmap_unwire_pte_hold(pmap, va, m, info);
1160 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1161 pmap_inval_info_t info)
1164 * Wait until we can busy the page ourselves. We cannot have
1165 * any active flushes if we block. We own one hold count on the
1166 * page so it cannot be freed out from under us.
1168 if (m->flags & PG_BUSY) {
1169 pmap_inval_flush(info);
1170 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1173 KASSERT(m->queue == PQ_NONE,
1174 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1177 * This case can occur if new references were acquired while
1180 if (m->hold_count > 1) {
1181 KKASSERT(m->hold_count > 1);
1187 * Unmap the page table page
1189 KKASSERT(m->hold_count == 1);
1191 pmap_inval_interlock(info, pmap, -1);
1193 if (m->pindex >= (NUPDE + NUPDPE)) {
1196 pml4 = pmap_pml4e(pmap, va);
1198 } else if (m->pindex >= NUPDE) {
1201 pdp = pmap_pdpe(pmap, va);
1206 pd = pmap_pde(pmap, va);
1210 KKASSERT(pmap->pm_stats.resident_count > 0);
1211 --pmap->pm_stats.resident_count;
1213 if (pmap->pm_ptphint == m)
1214 pmap->pm_ptphint = NULL;
1215 pmap_inval_deinterlock(info, pmap);
1217 if (m->pindex < NUPDE) {
1218 /* We just released a PT, unhold the matching PD */
1221 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
1222 pmap_unwire_pte_hold(pmap, va, pdpg, info);
1224 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
1225 /* We just released a PD, unhold the matching PDP */
1228 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
1229 pmap_unwire_pte_hold(pmap, va, pdppg, info);
1233 * This was our last hold, the page had better be unwired
1234 * after we decrement wire_count.
1236 * FUTURE NOTE: shared page directory page could result in
1237 * multiple wire counts.
1241 KKASSERT(m->wire_count == 0);
1242 --vmstats.v_wire_count;
1243 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1245 vm_page_free_zero(m);
1251 * After removing a page table entry, this routine is used to
1252 * conditionally free the page, and manage the hold/wire counts.
1256 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1257 pmap_inval_info_t info)
1259 vm_pindex_t ptepindex;
1261 if (va >= VM_MAX_USER_ADDRESS)
1265 ptepindex = pmap_pde_pindex(va);
1267 if (pmap->pm_ptphint &&
1268 (pmap->pm_ptphint->pindex == ptepindex)) {
1269 mpte = pmap->pm_ptphint;
1272 pmap_inval_flush(info);
1273 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1274 pmap->pm_ptphint = mpte;
1279 return pmap_unwire_pte_hold(pmap, va, mpte, info);
1283 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1284 * it, and IdlePTD, represents the template used to update all other pmaps.
1286 * On architectures where the kernel pmap is not integrated into the user
1287 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1288 * kernel_pmap should be used to directly access the kernel_pmap.
1291 pmap_pinit0(struct pmap *pmap)
1293 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1295 pmap->pm_active = 0;
1296 pmap->pm_ptphint = NULL;
1297 TAILQ_INIT(&pmap->pm_pvlist);
1298 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1302 * Initialize a preallocated and zeroed pmap structure,
1303 * such as one in a vmspace structure.
1306 pmap_pinit(struct pmap *pmap)
1311 * No need to allocate page table space yet but we do need a valid
1312 * page directory table.
1314 if (pmap->pm_pml4 == NULL) {
1316 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1320 * Allocate an object for the ptes
1322 if (pmap->pm_pteobj == NULL)
1323 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1);
1326 * Allocate the page directory page, unless we already have
1327 * one cached. If we used the cached page the wire_count will
1328 * already be set appropriately.
1330 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1331 ptdpg = vm_page_grab(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I,
1332 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1333 pmap->pm_pdirm = ptdpg;
1334 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1335 ptdpg->valid = VM_PAGE_BITS_ALL;
1336 if (ptdpg->wire_count == 0)
1337 ++vmstats.v_wire_count;
1338 ptdpg->wire_count = 1;
1339 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1341 if ((ptdpg->flags & PG_ZERO) == 0)
1342 bzero(pmap->pm_pml4, PAGE_SIZE);
1344 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1345 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1347 /* install self-referential address mapping entry */
1348 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1351 pmap->pm_active = 0;
1352 pmap->pm_ptphint = NULL;
1353 TAILQ_INIT(&pmap->pm_pvlist);
1354 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1355 pmap->pm_stats.resident_count = 1;
1359 * Clean up a pmap structure so it can be physically freed. This routine
1360 * is called by the vmspace dtor function. A great deal of pmap data is
1361 * left passively mapped to improve vmspace management so we have a bit
1362 * of cleanup work to do here.
1365 pmap_puninit(pmap_t pmap)
1369 KKASSERT(pmap->pm_active == 0);
1370 if ((p = pmap->pm_pdirm) != NULL) {
1371 KKASSERT(pmap->pm_pml4 != NULL);
1372 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1373 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1375 vmstats.v_wire_count--;
1376 KKASSERT((p->flags & PG_BUSY) == 0);
1378 vm_page_free_zero(p);
1379 pmap->pm_pdirm = NULL;
1381 if (pmap->pm_pml4) {
1382 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
1383 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1384 pmap->pm_pml4 = NULL;
1386 if (pmap->pm_pteobj) {
1387 vm_object_deallocate(pmap->pm_pteobj);
1388 pmap->pm_pteobj = NULL;
1393 * Wire in kernel global address entries. To avoid a race condition
1394 * between pmap initialization and pmap_growkernel, this procedure
1395 * adds the pmap to the master list (which growkernel scans to update),
1396 * then copies the template.
1399 pmap_pinit2(struct pmap *pmap)
1402 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1403 /* XXX copies current process, does not fill in MPPTDI */
1408 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1409 * 0 on failure (if the procedure had to sleep).
1411 * When asked to remove the page directory page itself, we actually just
1412 * leave it cached so we do not have to incur the SMP inval overhead of
1413 * removing the kernel mapping. pmap_puninit() will take care of it.
1417 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1420 * This code optimizes the case of freeing non-busy
1421 * page-table pages. Those pages are zero now, and
1422 * might as well be placed directly into the zero queue.
1424 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1430 * Remove the page table page from the processes address space.
1432 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1434 * We are the pml4 table itself.
1436 /* XXX anything to do here? */
1437 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1439 * Remove a PDP page from the PML4. We do not maintain
1440 * hold counts on the PML4 page.
1446 m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1447 KKASSERT(m4 != NULL);
1448 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1449 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1450 KKASSERT(pml4[idx] != 0);
1452 } else if (p->pindex >= NUPDE) {
1454 * Remove a PD page from the PDP and drop the hold count
1455 * on the PDP. The PDP is left cached in the pmap if
1456 * the hold count drops to 0 so the wire count remains
1463 m3 = vm_page_lookup(pmap->pm_pteobj,
1464 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1465 KKASSERT(m3 != NULL);
1466 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1467 idx = (p->pindex - NUPDE) % NPDPEPG;
1468 KKASSERT(pdp[idx] != 0);
1473 * Remove a PT page from the PD and drop the hold count
1474 * on the PD. The PD is left cached in the pmap if
1475 * the hold count drops to 0 so the wire count remains
1482 m2 = vm_page_lookup(pmap->pm_pteobj,
1483 NUPDE + p->pindex / NPDEPG);
1484 KKASSERT(m2 != NULL);
1485 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1486 idx = p->pindex % NPDEPG;
1492 * One fewer mappings in the pmap. p's hold count had better
1495 KKASSERT(pmap->pm_stats.resident_count > 0);
1496 --pmap->pm_stats.resident_count;
1498 panic("pmap_release: freeing held page table page");
1499 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1500 pmap->pm_ptphint = NULL;
1503 * We leave the top-level page table page cached, wired, and mapped in
1504 * the pmap until the dtor function (pmap_puninit()) gets called.
1505 * However, still clean it up so we can set PG_ZERO.
1507 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1508 bzero(pmap->pm_pml4, PAGE_SIZE);
1509 vm_page_flag_set(p, PG_ZERO);
1513 KKASSERT(p->wire_count == 0);
1514 vmstats.v_wire_count--;
1515 /* JG eventually revert to using vm_page_free_zero() */
1522 * This routine is called when various levels in the page table need to
1523 * be populated. This routine cannot fail.
1527 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1532 * Find or fabricate a new pagetable page. This will busy the page.
1534 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1535 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1536 if ((m->flags & PG_ZERO) == 0) {
1537 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1540 KASSERT(m->queue == PQ_NONE,
1541 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1544 * Increment the hold count for the page we will be returning to
1548 if (m->wire_count++ == 0)
1549 vmstats.v_wire_count++;
1552 * Map the pagetable page into the process address space, if
1553 * it isn't already there.
1555 * It is possible that someone else got in and mapped the page
1556 * directory page while we were blocked, if so just unbusy and
1557 * return the held page.
1559 if (ptepindex >= (NUPDE + NUPDPE)) {
1561 * Wire up a new PDP page in the PML4
1563 vm_pindex_t pml4index;
1566 pml4index = ptepindex - (NUPDE + NUPDPE);
1567 pml4 = &pmap->pm_pml4[pml4index];
1569 if (--m->wire_count == 0)
1570 --vmstats.v_wire_count;
1574 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1575 } else if (ptepindex >= NUPDE) {
1577 * Wire up a new PD page in the PDP
1579 vm_pindex_t pml4index;
1580 vm_pindex_t pdpindex;
1585 pdpindex = ptepindex - NUPDE;
1586 pml4index = pdpindex >> NPML4EPGSHIFT;
1588 pml4 = &pmap->pm_pml4[pml4index];
1589 if ((*pml4 & PG_V) == 0) {
1591 * Have to allocate a new PDP page, recurse.
1592 * This always succeeds. Returned page will
1595 pdppg = _pmap_allocpte(pmap,
1596 NUPDE + NUPDPE + pml4index);
1599 * Add a held reference to the PDP page.
1601 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1602 pdppg->hold_count++;
1606 * Now find the pdp_entry and map the PDP. If the PDP
1607 * has already been mapped unwind and return the
1608 * already-mapped PDP held.
1610 * pdppg is left held (hold_count is incremented for
1611 * each PD in the PDP).
1613 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1614 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1616 vm_page_unhold(pdppg);
1617 if (--m->wire_count == 0)
1618 --vmstats.v_wire_count;
1622 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1625 * Wire up the new PT page in the PD
1627 vm_pindex_t pml4index;
1628 vm_pindex_t pdpindex;
1634 pdpindex = ptepindex >> NPDPEPGSHIFT;
1635 pml4index = pdpindex >> NPML4EPGSHIFT;
1638 * Locate the PDP page in the PML4, then the PD page in
1639 * the PDP. If either does not exist we simply recurse
1642 * We can just recurse on the PD page as it will recurse
1643 * on the PDP if necessary.
1645 pml4 = &pmap->pm_pml4[pml4index];
1646 if ((*pml4 & PG_V) == 0) {
1647 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1648 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1649 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1651 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1652 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1653 if ((*pdp & PG_V) == 0) {
1654 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
1656 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1662 * Now fill in the pte in the PD. If the pte already exists
1663 * (again, if we raced the grab), unhold pdpg and unwire
1664 * m, returning a held m.
1666 * pdpg is left held (hold_count is incremented for
1667 * each PT in the PD).
1669 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
1670 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1672 vm_page_unhold(pdpg);
1673 if (--m->wire_count == 0)
1674 --vmstats.v_wire_count;
1678 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1682 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1683 * valid bits, mapped flag, unbusy, and we're done.
1685 pmap->pm_ptphint = m;
1686 ++pmap->pm_stats.resident_count;
1688 m->valid = VM_PAGE_BITS_ALL;
1689 vm_page_flag_clear(m, PG_ZERO);
1690 vm_page_flag_set(m, PG_MAPPED);
1698 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1700 vm_pindex_t ptepindex;
1705 * Calculate pagetable page index
1707 ptepindex = pmap_pde_pindex(va);
1710 * Get the page directory entry
1712 pd = pmap_pde(pmap, va);
1715 * This supports switching from a 2MB page to a
1718 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1719 panic("no promotion/demotion yet");
1727 * If the page table page is mapped, we just increment the
1728 * hold count, and activate it.
1730 if (pd != NULL && (*pd & PG_V) != 0) {
1731 /* YYY hint is used here on i386 */
1732 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1733 pmap->pm_ptphint = m;
1738 * Here if the pte page isn't mapped, or if it has been deallocated.
1740 return _pmap_allocpte(pmap, ptepindex);
1744 /***************************************************
1745 * Pmap allocation/deallocation routines.
1746 ***************************************************/
1749 * Release any resources held by the given physical map.
1750 * Called when a pmap initialized by pmap_pinit is being released.
1751 * Should only be called if the map contains no valid mappings.
1753 static int pmap_release_callback(struct vm_page *p, void *data);
1756 pmap_release(struct pmap *pmap)
1758 vm_object_t object = pmap->pm_pteobj;
1759 struct rb_vm_page_scan_info info;
1761 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1762 #if defined(DIAGNOSTIC)
1763 if (object->ref_count != 1)
1764 panic("pmap_release: pteobj reference count != 1");
1768 info.object = object;
1770 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1777 info.limit = object->generation;
1779 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1780 pmap_release_callback, &info);
1781 if (info.error == 0 && info.mpte) {
1782 if (!pmap_release_free_page(pmap, info.mpte))
1786 } while (info.error);
1791 pmap_release_callback(struct vm_page *p, void *data)
1793 struct rb_vm_page_scan_info *info = data;
1795 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1799 if (!pmap_release_free_page(info->pmap, p)) {
1803 if (info->object->generation != info->limit) {
1811 * Grow the number of kernel page table entries, if needed.
1814 pmap_growkernel(vm_offset_t addr)
1817 vm_offset_t ptppaddr;
1819 pd_entry_t *pde, newpdir;
1823 if (kernel_vm_end == 0) {
1824 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
1826 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1827 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1829 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1830 kernel_vm_end = kernel_map.max_offset;
1835 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1836 if (addr - 1 >= kernel_map.max_offset)
1837 addr = kernel_map.max_offset;
1838 while (kernel_vm_end < addr) {
1839 pde = pmap_pde(&kernel_pmap, kernel_vm_end);
1841 /* We need a new PDP entry */
1842 nkpg = vm_page_alloc(kptobj, nkpt,
1843 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM
1844 | VM_ALLOC_INTERRUPT);
1846 panic("pmap_growkernel: no memory to grow kernel");
1847 paddr = VM_PAGE_TO_PHYS(nkpg);
1848 if ((nkpg->flags & PG_ZERO) == 0)
1849 pmap_zero_page(paddr);
1850 vm_page_flag_clear(nkpg, PG_ZERO);
1851 newpdp = (pdp_entry_t)
1852 (paddr | PG_V | PG_RW | PG_A | PG_M);
1853 *pmap_pdpe(&kernel_pmap, kernel_vm_end) = newpdp;
1855 continue; /* try again */
1857 if ((*pde & PG_V) != 0) {
1858 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1859 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1860 kernel_vm_end = kernel_map.max_offset;
1867 * This index is bogus, but out of the way
1869 nkpg = vm_page_alloc(kptobj, nkpt,
1870 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1872 panic("pmap_growkernel: no memory to grow kernel");
1875 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1876 pmap_zero_page(ptppaddr);
1877 vm_page_flag_clear(nkpg, PG_ZERO);
1878 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1879 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1882 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1883 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1884 kernel_vm_end = kernel_map.max_offset;
1892 * Retire the given physical map from service.
1893 * Should only be called if the map contains
1894 * no valid mappings.
1897 pmap_destroy(pmap_t pmap)
1904 count = --pmap->pm_count;
1907 panic("destroying a pmap is not yet implemented");
1912 * Add a reference to the specified pmap.
1915 pmap_reference(pmap_t pmap)
1922 /***************************************************
1923 * page management routines.
1924 ***************************************************/
1927 * free the pv_entry back to the free list. This function may be
1928 * called from an interrupt.
1932 free_pv_entry(pv_entry_t pv)
1935 KKASSERT(pv_entry_count >= 0);
1940 * get a new pv_entry, allocating a block from the system
1941 * when needed. This function may be called from an interrupt.
1948 if (pv_entry_high_water &&
1949 (pv_entry_count > pv_entry_high_water) &&
1950 (pmap_pagedaemon_waken == 0)) {
1951 pmap_pagedaemon_waken = 1;
1952 wakeup(&vm_pages_needed);
1954 return zalloc(pvzone);
1958 * This routine is very drastic, but can save the system
1966 static int warningdone=0;
1968 if (pmap_pagedaemon_waken == 0)
1971 if (warningdone < 5) {
1972 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1976 for(i = 0; i < vm_page_array_size; i++) {
1977 m = &vm_page_array[i];
1978 if (m->wire_count || m->hold_count || m->busy ||
1979 (m->flags & PG_BUSY))
1983 pmap_pagedaemon_waken = 0;
1988 * If it is the first entry on the list, it is actually
1989 * in the header and we must copy the following entry up
1990 * to the header. Otherwise we must search the list for
1991 * the entry. In either case we free the now unused entry.
1995 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1996 vm_offset_t va, pmap_inval_info_t info)
2002 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
2003 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2004 if (pmap == pv->pv_pmap && va == pv->pv_va)
2008 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2009 if (va == pv->pv_va)
2017 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2018 m->md.pv_list_count--;
2019 KKASSERT(m->md.pv_list_count >= 0);
2020 if (TAILQ_EMPTY(&m->md.pv_list))
2021 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2022 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2023 ++pmap->pm_generation;
2024 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
2032 * Create a pv entry for page at pa for
2037 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2042 pv = get_pv_entry();
2047 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2048 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2049 ++pmap->pm_generation;
2050 m->md.pv_list_count++;
2056 * pmap_remove_pte: do the things to unmap a page in a process
2060 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2061 pmap_inval_info_t info)
2066 pmap_inval_interlock(info, pmap, va);
2067 oldpte = pte_load_clear(ptq);
2068 pmap_inval_deinterlock(info, pmap);
2070 pmap->pm_stats.wired_count -= 1;
2072 * Machines that don't support invlpg, also don't support
2073 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2077 cpu_invlpg((void *)va);
2078 KKASSERT(pmap->pm_stats.resident_count > 0);
2079 --pmap->pm_stats.resident_count;
2080 if (oldpte & PG_MANAGED) {
2081 m = PHYS_TO_VM_PAGE(oldpte);
2082 if (oldpte & PG_M) {
2083 #if defined(PMAP_DIAGNOSTIC)
2084 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2086 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2090 if (pmap_track_modified(va))
2094 vm_page_flag_set(m, PG_REFERENCED);
2095 return pmap_remove_entry(pmap, m, va, info);
2097 return pmap_unuse_pt(pmap, va, NULL, info);
2106 * Remove a single page from a process address space.
2108 * This function may not be called from an interrupt if the pmap is
2113 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2117 pte = pmap_pte(pmap, va);
2120 if ((*pte & PG_V) == 0)
2122 pmap_remove_pte(pmap, pte, va, info);
2128 * Remove the given range of addresses from the specified map.
2130 * It is assumed that the start and end are properly
2131 * rounded to the page size.
2133 * This function may not be called from an interrupt if the pmap is
2137 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2139 vm_offset_t va_next;
2140 pml4_entry_t *pml4e;
2142 pd_entry_t ptpaddr, *pde;
2144 struct pmap_inval_info info;
2149 if (pmap->pm_stats.resident_count == 0)
2152 pmap_inval_init(&info);
2155 * special handling of removing one page. a very
2156 * common operation and easy to short circuit some
2159 if (sva + PAGE_SIZE == eva) {
2160 pde = pmap_pde(pmap, sva);
2161 if (pde && (*pde & PG_PS) == 0) {
2162 pmap_remove_page(pmap, sva, &info);
2163 pmap_inval_done(&info);
2168 for (; sva < eva; sva = va_next) {
2169 pml4e = pmap_pml4e(pmap, sva);
2170 if ((*pml4e & PG_V) == 0) {
2171 va_next = (sva + NBPML4) & ~PML4MASK;
2177 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2178 if ((*pdpe & PG_V) == 0) {
2179 va_next = (sva + NBPDP) & ~PDPMASK;
2186 * Calculate index for next page table.
2188 va_next = (sva + NBPDR) & ~PDRMASK;
2192 pde = pmap_pdpe_to_pde(pdpe, sva);
2196 * Weed out invalid mappings.
2202 * Check for large page.
2204 if ((ptpaddr & PG_PS) != 0) {
2205 /* JG FreeBSD has more complex treatment here */
2206 pmap_inval_interlock(&info, pmap, -1);
2208 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2209 pmap_inval_deinterlock(&info, pmap);
2214 * Limit our scan to either the end of the va represented
2215 * by the current page table page, or to the end of the
2216 * range being removed.
2222 * NOTE: pmap_remove_pte() can block.
2224 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2228 if (pmap_remove_pte(pmap, pte, sva, &info))
2232 pmap_inval_done(&info);
2238 * Removes this physical page from all physical maps in which it resides.
2239 * Reflects back modify bits to the pager.
2241 * This routine may not be called from an interrupt.
2246 pmap_remove_all(vm_page_t m)
2248 struct pmap_inval_info info;
2249 pt_entry_t *pte, tpte;
2252 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2255 pmap_inval_init(&info);
2257 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2258 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2259 --pv->pv_pmap->pm_stats.resident_count;
2261 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2262 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
2263 tpte = pte_load_clear(pte);
2265 pv->pv_pmap->pm_stats.wired_count--;
2266 pmap_inval_deinterlock(&info, pv->pv_pmap);
2268 vm_page_flag_set(m, PG_REFERENCED);
2271 * Update the vm_page_t clean and reference bits.
2274 #if defined(PMAP_DIAGNOSTIC)
2275 if (pmap_nw_modified(tpte)) {
2277 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2281 if (pmap_track_modified(pv->pv_va))
2284 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2285 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2286 ++pv->pv_pmap->pm_generation;
2287 m->md.pv_list_count--;
2288 KKASSERT(m->md.pv_list_count >= 0);
2289 if (TAILQ_EMPTY(&m->md.pv_list))
2290 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2291 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2295 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2296 pmap_inval_done(&info);
2302 * Set the physical protection on the specified range of this map
2305 * This function may not be called from an interrupt if the map is
2306 * not the kernel_pmap.
2309 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2311 vm_offset_t va_next;
2312 pml4_entry_t *pml4e;
2314 pd_entry_t ptpaddr, *pde;
2316 pmap_inval_info info;
2318 /* JG review for NX */
2323 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2324 pmap_remove(pmap, sva, eva);
2328 if (prot & VM_PROT_WRITE)
2331 pmap_inval_init(&info);
2333 for (; sva < eva; sva = va_next) {
2335 pml4e = pmap_pml4e(pmap, sva);
2336 if ((*pml4e & PG_V) == 0) {
2337 va_next = (sva + NBPML4) & ~PML4MASK;
2343 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2344 if ((*pdpe & PG_V) == 0) {
2345 va_next = (sva + NBPDP) & ~PDPMASK;
2351 va_next = (sva + NBPDR) & ~PDRMASK;
2355 pde = pmap_pdpe_to_pde(pdpe, sva);
2359 * Check for large page.
2361 if ((ptpaddr & PG_PS) != 0) {
2362 pmap_inval_interlock(&info, pmap, -1);
2363 *pde &= ~(PG_M|PG_RW);
2364 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2365 pmap_inval_deinterlock(&info, pmap);
2370 * Weed out invalid mappings. Note: we assume that the page
2371 * directory table is always allocated, and in kernel virtual.
2379 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2386 * XXX non-optimal. Note also that there can be
2387 * no pmap_inval_flush() calls until after we modify
2388 * ptbase[sindex] (or otherwise we have to do another
2389 * pmap_inval_add() call).
2391 pmap_inval_interlock(&info, pmap, sva);
2395 if ((pbits & PG_V) == 0) {
2396 pmap_inval_deinterlock(&info, pmap);
2399 if (pbits & PG_MANAGED) {
2402 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2403 vm_page_flag_set(m, PG_REFERENCED);
2407 if (pmap_track_modified(sva)) {
2409 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2416 if (pbits != cbits &&
2417 !atomic_cmpset_long(pte, pbits, cbits)) {
2420 pmap_inval_deinterlock(&info, pmap);
2423 pmap_inval_done(&info);
2427 * Insert the given physical page (p) at
2428 * the specified virtual address (v) in the
2429 * target physical map with the protection requested.
2431 * If specified, the page will be wired down, meaning
2432 * that the related pte can not be reclaimed.
2434 * NB: This is the only routine which MAY NOT lazy-evaluate
2435 * or lose information. That is, this routine must actually
2436 * insert this page into the given map NOW.
2439 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2446 pt_entry_t origpte, newpte;
2448 pmap_inval_info info;
2453 va = trunc_page(va);
2454 #ifdef PMAP_DIAGNOSTIC
2456 panic("pmap_enter: toobig");
2457 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2458 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
2460 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2461 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2463 db_print_backtrace();
2466 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2467 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2469 db_print_backtrace();
2474 * In the case that a page table page is not
2475 * resident, we are creating it here.
2477 if (va < VM_MAX_USER_ADDRESS)
2478 mpte = pmap_allocpte(pmap, va);
2482 pmap_inval_init(&info);
2483 pde = pmap_pde(pmap, va);
2484 if (pde != NULL && (*pde & PG_V) != 0) {
2485 if ((*pde & PG_PS) != 0)
2486 panic("pmap_enter: attempted pmap_enter on 2MB page");
2487 pte = pmap_pde_to_pte(pde, va);
2489 panic("pmap_enter: invalid page directory va=%#lx", va);
2491 KKASSERT(pte != NULL);
2492 pa = VM_PAGE_TO_PHYS(m);
2494 opa = origpte & PG_FRAME;
2497 * Mapping has not changed, must be protection or wiring change.
2499 if (origpte && (opa == pa)) {
2501 * Wiring change, just update stats. We don't worry about
2502 * wiring PT pages as they remain resident as long as there
2503 * are valid mappings in them. Hence, if a user page is wired,
2504 * the PT page will be also.
2506 if (wired && ((origpte & PG_W) == 0))
2507 pmap->pm_stats.wired_count++;
2508 else if (!wired && (origpte & PG_W))
2509 pmap->pm_stats.wired_count--;
2511 #if defined(PMAP_DIAGNOSTIC)
2512 if (pmap_nw_modified(origpte)) {
2514 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2520 * Remove the extra pte reference. Note that we cannot
2521 * optimize the RO->RW case because we have adjusted the
2522 * wiring count above and may need to adjust the wiring
2529 * We might be turning off write access to the page,
2530 * so we go ahead and sense modify status.
2532 if (origpte & PG_MANAGED) {
2533 if ((origpte & PG_M) && pmap_track_modified(va)) {
2535 om = PHYS_TO_VM_PAGE(opa);
2539 KKASSERT(m->flags & PG_MAPPED);
2544 * Mapping has changed, invalidate old range and fall through to
2545 * handle validating new mapping.
2549 err = pmap_remove_pte(pmap, pte, va, &info);
2551 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2553 opa = origpte & PG_FRAME;
2555 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2561 * Enter on the PV list if part of our managed memory. Note that we
2562 * raise IPL while manipulating pv_table since pmap_enter can be
2563 * called at interrupt time.
2565 if (pmap_initialized &&
2566 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2567 pmap_insert_entry(pmap, va, mpte, m);
2569 vm_page_flag_set(m, PG_MAPPED);
2573 * Increment counters
2575 ++pmap->pm_stats.resident_count;
2577 pmap->pm_stats.wired_count++;
2581 * Now validate mapping with desired protection/wiring.
2583 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
2587 if (va < VM_MAX_USER_ADDRESS)
2589 if (pmap == &kernel_pmap)
2593 * if the mapping or permission bits are different, we need
2594 * to update the pte.
2596 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2597 pmap_inval_interlock(&info, pmap, va);
2598 *pte = newpte | PG_A;
2599 pmap_inval_deinterlock(&info, pmap);
2601 vm_page_flag_set(m, PG_WRITEABLE);
2603 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2604 pmap_inval_done(&info);
2608 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2609 * This code also assumes that the pmap has no pre-existing entry for this
2612 * This code currently may only be used on user pmaps, not kernel_pmap.
2615 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2620 vm_pindex_t ptepindex;
2622 pmap_inval_info info;
2624 pmap_inval_init(&info);
2626 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2627 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2629 db_print_backtrace();
2632 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2633 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2635 db_print_backtrace();
2639 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2642 * Calculate the page table page (mpte), allocating it if necessary.
2644 * A held page table page (mpte), or NULL, is passed onto the
2645 * section following.
2647 if (va < VM_MAX_USER_ADDRESS) {
2649 * Calculate pagetable page index
2651 ptepindex = pmap_pde_pindex(va);
2655 * Get the page directory entry
2657 ptepa = pmap_pde(pmap, va);
2660 * If the page table page is mapped, we just increment
2661 * the hold count, and activate it.
2663 if (ptepa && (*ptepa & PG_V) != 0) {
2665 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2666 // if (pmap->pm_ptphint &&
2667 // (pmap->pm_ptphint->pindex == ptepindex)) {
2668 // mpte = pmap->pm_ptphint;
2670 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2671 pmap->pm_ptphint = mpte;
2676 mpte = _pmap_allocpte(pmap, ptepindex);
2678 } while (mpte == NULL);
2681 /* this code path is not yet used */
2685 * With a valid (and held) page directory page, we can just use
2686 * vtopte() to get to the pte. If the pte is already present
2687 * we do not disturb it.
2692 pmap_unwire_pte_hold(pmap, va, mpte, &info);
2693 pa = VM_PAGE_TO_PHYS(m);
2694 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2695 pmap_inval_done(&info);
2700 * Enter on the PV list if part of our managed memory
2702 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2703 pmap_insert_entry(pmap, va, mpte, m);
2704 vm_page_flag_set(m, PG_MAPPED);
2708 * Increment counters
2710 ++pmap->pm_stats.resident_count;
2712 pa = VM_PAGE_TO_PHYS(m);
2715 * Now validate mapping with RO protection
2717 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2718 *pte = pa | PG_V | PG_U;
2720 *pte = pa | PG_V | PG_U | PG_MANAGED;
2721 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2722 pmap_inval_done(&info);
2726 * Make a temporary mapping for a physical address. This is only intended
2727 * to be used for panic dumps.
2729 /* JG Needed on x86_64? */
2731 pmap_kenter_temporary(vm_paddr_t pa, int i)
2733 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2734 return ((void *)crashdumpmap);
2737 #define MAX_INIT_PT (96)
2740 * This routine preloads the ptes for a given object into the specified pmap.
2741 * This eliminates the blast of soft faults on process startup and
2742 * immediately after an mmap.
2744 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2747 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2748 vm_object_t object, vm_pindex_t pindex,
2749 vm_size_t size, int limit)
2751 struct rb_vm_page_scan_info info;
2756 * We can't preinit if read access isn't set or there is no pmap
2759 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2763 * We can't preinit if the pmap is not the current pmap
2765 lp = curthread->td_lwp;
2766 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2769 psize = x86_64_btop(size);
2771 if ((object->type != OBJT_VNODE) ||
2772 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2773 (object->resident_page_count > MAX_INIT_PT))) {
2777 if (psize + pindex > object->size) {
2778 if (object->size < pindex)
2780 psize = object->size - pindex;
2787 * Use a red-black scan to traverse the requested range and load
2788 * any valid pages found into the pmap.
2790 * We cannot safely scan the object's memq unless we are in a
2791 * critical section since interrupts can remove pages from objects.
2793 info.start_pindex = pindex;
2794 info.end_pindex = pindex + psize - 1;
2801 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2802 pmap_object_init_pt_callback, &info);
2808 pmap_object_init_pt_callback(vm_page_t p, void *data)
2810 struct rb_vm_page_scan_info *info = data;
2811 vm_pindex_t rel_index;
2813 * don't allow an madvise to blow away our really
2814 * free pages allocating pv entries.
2816 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2817 vmstats.v_free_count < vmstats.v_free_reserved) {
2820 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2821 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2822 if ((p->queue - p->pc) == PQ_CACHE)
2823 vm_page_deactivate(p);
2825 rel_index = p->pindex - info->start_pindex;
2826 pmap_enter_quick(info->pmap,
2827 info->addr + x86_64_ptob(rel_index), p);
2834 * Return TRUE if the pmap is in shape to trivially
2835 * pre-fault the specified address.
2837 * Returns FALSE if it would be non-trivial or if a
2838 * pte is already loaded into the slot.
2841 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2846 pde = pmap_pde(pmap, addr);
2847 if (pde == NULL || *pde == 0)
2858 * Routine: pmap_change_wiring
2859 * Function: Change the wiring attribute for a map/virtual-address
2861 * In/out conditions:
2862 * The mapping must already exist in the pmap.
2865 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2872 pte = pmap_pte(pmap, va);
2874 if (wired && !pmap_pte_w(pte))
2875 pmap->pm_stats.wired_count++;
2876 else if (!wired && pmap_pte_w(pte))
2877 pmap->pm_stats.wired_count--;
2880 * Wiring is not a hardware characteristic so there is no need to
2881 * invalidate TLB. However, in an SMP environment we must use
2882 * a locked bus cycle to update the pte (if we are not using
2883 * the pmap_inval_*() API that is)... it's ok to do this for simple
2888 atomic_set_long(pte, PG_W);
2890 atomic_clear_long(pte, PG_W);
2893 atomic_set_long_nonlocked(pte, PG_W);
2895 atomic_clear_long_nonlocked(pte, PG_W);
2902 * Copy the range specified by src_addr/len
2903 * from the source map to the range dst_addr/len
2904 * in the destination map.
2906 * This routine is only advisory and need not do anything.
2909 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2910 vm_size_t len, vm_offset_t src_addr)
2914 pmap_inval_info info;
2916 vm_offset_t end_addr = src_addr + len;
2918 pd_entry_t src_frame, dst_frame;
2921 if (dst_addr != src_addr)
2924 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
2925 if (src_frame != (PTDpde & PG_FRAME)) {
2929 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
2930 if (dst_frame != (APTDpde & PG_FRAME)) {
2931 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2932 /* The page directory is not shared between CPUs */
2936 pmap_inval_init(&info);
2937 pmap_inval_add(&info, dst_pmap, -1);
2938 pmap_inval_add(&info, src_pmap, -1);
2941 * critical section protection is required to maintain the page/object
2942 * association, interrupts can free pages and remove them from
2946 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2947 pt_entry_t *src_pte, *dst_pte;
2948 vm_page_t dstmpte, srcmpte;
2949 vm_offset_t srcptepaddr;
2950 vm_pindex_t ptepindex;
2952 if (addr >= UPT_MIN_ADDRESS)
2953 panic("pmap_copy: invalid to pmap_copy page tables\n");
2956 * Don't let optional prefaulting of pages make us go
2957 * way below the low water mark of free pages or way
2958 * above high water mark of used pv entries.
2960 if (vmstats.v_free_count < vmstats.v_free_reserved ||
2961 pv_entry_count > pv_entry_high_water)
2964 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2965 ptepindex = addr >> PDRSHIFT;
2968 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2970 if (srcptepaddr == 0)
2973 if (srcptepaddr & PG_PS) {
2975 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2976 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2977 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2983 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2984 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2985 (srcmpte->flags & PG_BUSY)) {
2989 if (pdnxt > end_addr)
2992 src_pte = vtopte(addr);
2994 dst_pte = avtopte(addr);
2996 while (addr < pdnxt) {
3001 * we only virtual copy managed pages
3003 if ((ptetemp & PG_MANAGED) != 0) {
3005 * We have to check after allocpte for the
3006 * pte still being around... allocpte can
3009 * pmap_allocpte() can block. If we lose
3010 * our page directory mappings we stop.
3012 dstmpte = pmap_allocpte(dst_pmap, addr);
3015 if (src_frame != (PTDpde & PG_FRAME) ||
3016 dst_frame != (APTDpde & PG_FRAME)
3018 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3019 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3021 } else if ((*dst_pte == 0) &&
3022 (ptetemp = *src_pte) != 0 &&
3023 (ptetemp & PG_MANAGED)) {
3025 * Clear the modified and
3026 * accessed (referenced) bits
3029 m = PHYS_TO_VM_PAGE(ptetemp);
3030 *dst_pte = ptetemp & ~(PG_M | PG_A);
3031 ++dst_pmap->pm_stats.resident_count;
3032 pmap_insert_entry(dst_pmap, addr,
3034 KKASSERT(m->flags & PG_MAPPED);
3036 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3037 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3041 if (dstmpte->hold_count >= srcmpte->hold_count)
3051 pmap_inval_done(&info);
3058 * Zero the specified physical page.
3060 * This function may be called from an interrupt and no locking is
3064 pmap_zero_page(vm_paddr_t phys)
3066 vm_offset_t va = PHYS_TO_DMAP(phys);
3068 pagezero((void *)va);
3072 * pmap_page_assertzero:
3074 * Assert that a page is empty, panic if it isn't.
3077 pmap_page_assertzero(vm_paddr_t phys)
3079 vm_offset_t virt = PHYS_TO_DMAP(phys);
3082 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3083 if (*(long *)((char *)virt + i) != 0) {
3084 panic("pmap_page_assertzero() @ %p not zero!\n", (void *)virt);
3092 * Zero part of a physical page by mapping it into memory and clearing
3093 * its contents with bzero.
3095 * off and size may not cover an area beyond a single hardware page.
3098 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3100 vm_offset_t virt = PHYS_TO_DMAP(phys);
3102 bzero((char *)virt + off, size);
3108 * Copy the physical page from the source PA to the target PA.
3109 * This function may be called from an interrupt. No locking
3113 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3115 vm_offset_t src_virt, dst_virt;
3117 src_virt = PHYS_TO_DMAP(src);
3118 dst_virt = PHYS_TO_DMAP(dst);
3119 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
3123 * pmap_copy_page_frag:
3125 * Copy the physical page from the source PA to the target PA.
3126 * This function may be called from an interrupt. No locking
3130 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3132 vm_offset_t src_virt, dst_virt;
3134 src_virt = PHYS_TO_DMAP(src);
3135 dst_virt = PHYS_TO_DMAP(dst);
3137 bcopy((char *)src_virt + (src & PAGE_MASK),
3138 (char *)dst_virt + (dst & PAGE_MASK),
3143 * Returns true if the pmap's pv is one of the first
3144 * 16 pvs linked to from this page. This count may
3145 * be changed upwards or downwards in the future; it
3146 * is only necessary that true be returned for a small
3147 * subset of pmaps for proper page aging.
3150 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3155 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3160 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3161 if (pv->pv_pmap == pmap) {
3174 * Remove all pages from specified address space
3175 * this aids process exit speeds. Also, this code
3176 * is special cased for current process only, but
3177 * can have the more generic (and slightly slower)
3178 * mode enabled. This is much faster than pmap_remove
3179 * in the case of running down an entire address space.
3182 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3185 pt_entry_t *pte, tpte;
3188 pmap_inval_info info;
3190 int save_generation;
3192 lp = curthread->td_lwp;
3193 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3198 pmap_inval_init(&info);
3199 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3200 if (pv->pv_va >= eva || pv->pv_va < sva) {
3201 npv = TAILQ_NEXT(pv, pv_plist);
3205 KKASSERT(pmap == pv->pv_pmap);
3208 pte = vtopte(pv->pv_va);
3210 pte = pmap_pte_quick(pmap, pv->pv_va);
3211 pmap_inval_interlock(&info, pmap, pv->pv_va);
3214 * We cannot remove wired pages from a process' mapping
3218 pmap_inval_deinterlock(&info, pmap);
3219 npv = TAILQ_NEXT(pv, pv_plist);
3222 tpte = pte_load_clear(pte);
3224 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3226 KASSERT(m < &vm_page_array[vm_page_array_size],
3227 ("pmap_remove_pages: bad tpte %lx", tpte));
3229 KKASSERT(pmap->pm_stats.resident_count > 0);
3230 --pmap->pm_stats.resident_count;
3231 pmap_inval_deinterlock(&info, pmap);
3234 * Update the vm_page_t clean and reference bits.
3240 npv = TAILQ_NEXT(pv, pv_plist);
3241 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3242 save_generation = ++pmap->pm_generation;
3244 m->md.pv_list_count--;
3245 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3246 if (TAILQ_EMPTY(&m->md.pv_list))
3247 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3249 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3253 * Restart the scan if we blocked during the unuse or free
3254 * calls and other removals were made.
3256 if (save_generation != pmap->pm_generation) {
3257 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3258 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3261 pmap_inval_done(&info);
3265 * pmap_testbit tests bits in pte's
3266 * note that the testbit/clearbit routines are inline,
3267 * and a lot of things compile-time evaluate.
3271 pmap_testbit(vm_page_t m, int bit)
3276 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3279 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3284 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3286 * if the bit being tested is the modified bit, then
3287 * mark clean_map and ptes as never
3290 if (bit & (PG_A|PG_M)) {
3291 if (!pmap_track_modified(pv->pv_va))
3295 #if defined(PMAP_DIAGNOSTIC)
3296 if (pv->pv_pmap == NULL) {
3297 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3301 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3312 * this routine is used to modify bits in ptes
3316 pmap_clearbit(vm_page_t m, int bit)
3318 struct pmap_inval_info info;
3323 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3326 pmap_inval_init(&info);
3329 * Loop over all current mappings setting/clearing as appropos If
3330 * setting RO do we need to clear the VAC?
3332 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3334 * don't write protect pager mappings
3337 if (!pmap_track_modified(pv->pv_va))
3341 #if defined(PMAP_DIAGNOSTIC)
3342 if (pv->pv_pmap == NULL) {
3343 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3349 * Careful here. We can use a locked bus instruction to
3350 * clear PG_A or PG_M safely but we need to synchronize
3351 * with the target cpus when we mess with PG_RW.
3353 * We do not have to force synchronization when clearing
3354 * PG_M even for PTEs generated via virtual memory maps,
3355 * because the virtual kernel will invalidate the pmap
3356 * entry when/if it needs to resynchronize the Modify bit.
3359 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
3360 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3367 atomic_clear_long(pte, PG_M|PG_RW);
3370 * The cpu may be trying to set PG_M
3371 * simultaniously with our clearing
3374 if (!atomic_cmpset_long(pte, pbits,
3378 } else if (bit == PG_M) {
3380 * We could also clear PG_RW here to force
3381 * a fault on write to redetect PG_M for
3382 * virtual kernels, but it isn't necessary
3383 * since virtual kernels invalidate the pte
3384 * when they clear the VPTE_M bit in their
3385 * virtual page tables.
3387 atomic_clear_long(pte, PG_M);
3389 atomic_clear_long(pte, bit);
3393 pmap_inval_deinterlock(&info, pv->pv_pmap);
3395 pmap_inval_done(&info);
3399 * pmap_page_protect:
3401 * Lower the permission for all mappings to a given page.
3404 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3406 /* JG NX support? */
3407 if ((prot & VM_PROT_WRITE) == 0) {
3408 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3409 pmap_clearbit(m, PG_RW);
3410 vm_page_flag_clear(m, PG_WRITEABLE);
3418 pmap_phys_address(vm_pindex_t ppn)
3420 return (x86_64_ptob(ppn));
3424 * pmap_ts_referenced:
3426 * Return a count of reference bits for a page, clearing those bits.
3427 * It is not necessary for every reference bit to be cleared, but it
3428 * is necessary that 0 only be returned when there are truly no
3429 * reference bits set.
3431 * XXX: The exact number of bits to check and clear is a matter that
3432 * should be tested and standardized at some point in the future for
3433 * optimal aging of shared pages.
3436 pmap_ts_referenced(vm_page_t m)
3438 pv_entry_t pv, pvf, pvn;
3442 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3447 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3452 pvn = TAILQ_NEXT(pv, pv_list);
3455 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3456 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3459 if (!pmap_track_modified(pv->pv_va))
3462 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3464 if (pte && (*pte & PG_A)) {
3466 atomic_clear_long(pte, PG_A);
3468 atomic_clear_long_nonlocked(pte, PG_A);
3475 } while ((pv = pvn) != NULL && pv != pvf);
3485 * Return whether or not the specified physical page was modified
3486 * in any physical maps.
3489 pmap_is_modified(vm_page_t m)
3491 return pmap_testbit(m, PG_M);
3495 * Clear the modify bits on the specified physical page.
3498 pmap_clear_modify(vm_page_t m)
3500 pmap_clearbit(m, PG_M);
3504 * pmap_clear_reference:
3506 * Clear the reference bit on the specified physical page.
3509 pmap_clear_reference(vm_page_t m)
3511 pmap_clearbit(m, PG_A);
3515 * Miscellaneous support routines follow
3520 i386_protection_init(void)
3524 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3525 kp = protection_codes;
3526 for (prot = 0; prot < 8; prot++) {
3528 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3530 * Read access is also 0. There isn't any execute bit,
3531 * so just make it readable.
3533 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3534 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3535 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3538 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3539 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3540 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3541 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3549 * Map a set of physical memory pages into the kernel virtual
3550 * address space. Return a pointer to where it is mapped. This
3551 * routine is intended to be used for mapping device memory,
3554 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3558 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3560 vm_offset_t va, tmpva, offset;
3563 offset = pa & PAGE_MASK;
3564 size = roundup(offset + size, PAGE_SIZE);
3566 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3568 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3570 pa = pa & ~PAGE_MASK;
3571 for (tmpva = va; size > 0;) {
3572 pte = vtopte(tmpva);
3573 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3581 return ((void *)(va + offset));
3585 pmap_mapdev_uncacheable(vm_paddr_t pa, vm_size_t size)
3587 vm_offset_t va, tmpva, offset;
3590 offset = pa & PAGE_MASK;
3591 size = roundup(offset + size, PAGE_SIZE);
3593 va = kmem_alloc_nofault(&kernel_map, size, PAGE_SIZE);
3595 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3597 pa = pa & ~PAGE_MASK;
3598 for (tmpva = va; size > 0;) {
3599 pte = vtopte(tmpva);
3600 *pte = pa | PG_RW | PG_V | PG_N; /* | pgeflag; */
3608 return ((void *)(va + offset));
3612 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3614 vm_offset_t base, offset;
3616 base = va & ~PAGE_MASK;
3617 offset = va & PAGE_MASK;
3618 size = roundup(offset + size, PAGE_SIZE);
3619 pmap_qremove(va, size >> PAGE_SHIFT);
3620 kmem_free(&kernel_map, base, size);
3624 * perform the pmap work for mincore
3627 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3629 pt_entry_t *ptep, pte;
3633 ptep = pmap_pte(pmap, addr);
3638 if ((pte = *ptep) != 0) {
3641 val = MINCORE_INCORE;
3642 if ((pte & PG_MANAGED) == 0)
3645 pa = pte & PG_FRAME;
3647 m = PHYS_TO_VM_PAGE(pa);
3653 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3655 * Modified by someone
3657 else if (m->dirty || pmap_is_modified(m))
3658 val |= MINCORE_MODIFIED_OTHER;
3663 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3666 * Referenced by someone
3668 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3669 val |= MINCORE_REFERENCED_OTHER;
3670 vm_page_flag_set(m, PG_REFERENCED);
3677 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3678 * vmspace will be ref'd and the old one will be deref'd.
3680 * The vmspace for all lwps associated with the process will be adjusted
3681 * and cr3 will be reloaded if any lwp is the current lwp.
3684 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3686 struct vmspace *oldvm;
3690 oldvm = p->p_vmspace;
3691 if (oldvm != newvm) {
3692 p->p_vmspace = newvm;
3693 KKASSERT(p->p_nthreads == 1);
3694 lp = RB_ROOT(&p->p_lwp_tree);
3695 pmap_setlwpvm(lp, newvm);
3697 sysref_get(&newvm->vm_sysref);
3698 sysref_put(&oldvm->vm_sysref);
3705 * Set the vmspace for a LWP. The vmspace is almost universally set the
3706 * same as the process vmspace, but virtual kernels need to swap out contexts
3707 * on a per-lwp basis.
3710 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3712 struct vmspace *oldvm;
3716 oldvm = lp->lwp_vmspace;
3718 if (oldvm != newvm) {
3719 lp->lwp_vmspace = newvm;
3720 if (curthread->td_lwp == lp) {
3721 pmap = vmspace_pmap(newvm);
3723 atomic_set_int(&pmap->pm_active, mycpu->gd_cpumask);
3724 if (pmap->pm_active & CPUMASK_LOCK)
3725 pmap_interlock_wait(newvm);
3727 pmap->pm_active |= 1;
3729 #if defined(SWTCH_OPTIM_STATS)
3732 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
3733 curthread->td_pcb->pcb_cr3 |= PG_RW | PG_U | PG_V;
3734 load_cr3(curthread->td_pcb->pcb_cr3);
3735 pmap = vmspace_pmap(oldvm);
3737 atomic_clear_int(&pmap->pm_active, mycpu->gd_cpumask);
3739 pmap->pm_active &= ~1;
3749 * Called when switching to a locked pmap
3752 pmap_interlock_wait(struct vmspace *vm)
3754 struct pmap *pmap = &vm->vm_pmap;
3756 if (pmap->pm_active & CPUMASK_LOCK) {
3757 kprintf("Warning: pmap_interlock %p %08x\n",
3758 pmap, pmap->pm_active);
3759 while (pmap->pm_active & CPUMASK_LOCK) {
3762 lwkt_process_ipiq();
3770 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3773 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3777 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);