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 $
45 * $DragonFly: src/sys/platform/pc64/amd64/pmap.c,v 1.3 2008/08/29 17:07:10 dillon Exp $
49 * Manages physical address maps.
51 * In addition to hardware address maps, this
52 * module is called upon to provide software-use-only
53 * maps which may or may not be stored in the same
54 * form as hardware maps. These pseudo-maps are
55 * used to store intermediate results from copy
56 * operations to and from address spaces.
58 * Since the information managed by this module is
59 * also stored by the logical address mapping module,
60 * this module may throw away valid virtual-to-physical
61 * mappings at almost any time. However, invalidations
62 * of virtual-to-physical mappings must be done as
65 * In order to cope with hardware architectures which
66 * make virtual-to-physical map invalidates expensive,
67 * this module may delay invalidate or reduced protection
68 * operations until such time as they are actually
69 * necessary. This module is given full information as
70 * to which processors are currently using which maps,
71 * and to when physical maps must be made correct.
75 #include "opt_disable_pse.h"
78 #include "opt_msgbuf.h"
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/kernel.h>
84 #include <sys/msgbuf.h>
85 #include <sys/vmmeter.h>
89 #include <vm/vm_param.h>
90 #include <sys/sysctl.h>
92 #include <vm/vm_kern.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_extern.h>
97 #include <vm/vm_pageout.h>
98 #include <vm/vm_pager.h>
99 #include <vm/vm_zone.h>
101 #include <sys/user.h>
102 #include <sys/thread2.h>
103 #include <sys/sysref2.h>
105 #include <machine/cputypes.h>
106 #include <machine/md_var.h>
107 #include <machine/specialreg.h>
108 #include <machine/smp.h>
109 #include <machine_base/apic/apicreg.h>
110 #include <machine/globaldata.h>
111 #include <machine/pmap.h>
112 #include <machine/pmap_inval.h>
116 #define PMAP_KEEP_PDIRS
117 #ifndef PMAP_SHPGPERPROC
118 #define PMAP_SHPGPERPROC 200
121 #if defined(DIAGNOSTIC)
122 #define PMAP_DIAGNOSTIC
128 * Get PDEs and PTEs for user/kernel address space
130 static pd_entry_t
*pmap_pde(pmap_t pmap
, vm_offset_t va
);
131 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
133 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
134 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
135 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
136 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
137 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
141 * Given a map and a machine independent protection code,
142 * convert to a vax protection code.
144 #define pte_prot(m, p) \
145 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
146 static int protection_codes
[8];
148 struct pmap kernel_pmap
;
149 static TAILQ_HEAD(,pmap
) pmap_list
= TAILQ_HEAD_INITIALIZER(pmap_list
);
151 vm_paddr_t avail_start
; /* PA of first available physical page */
152 vm_paddr_t avail_end
; /* PA of last available physical page */
153 vm_offset_t virtual_start
; /* VA of first avail page (after kernel bss) */
154 vm_offset_t virtual_end
; /* VA of last avail page (end of kernel AS) */
155 vm_offset_t KvaStart
; /* VA start of KVA space */
156 vm_offset_t KvaEnd
; /* VA end of KVA space (non-inclusive) */
157 vm_offset_t KvaSize
; /* max size of kernel virtual address space */
158 static boolean_t pmap_initialized
= FALSE
; /* Has pmap_init completed? */
159 static int pgeflag
; /* PG_G or-in */
160 static int pseflag
; /* PG_PS or-in */
162 static vm_object_t kptobj
;
165 static vm_paddr_t dmaplimit
;
167 vm_offset_t kernel_vm_end
;
169 static uint64_t KPDphys
; /* phys addr of kernel level 2 */
170 uint64_t KPDPphys
; /* phys addr of kernel level 3 */
171 uint64_t KPML4phys
; /* phys addr of kernel level 4 */
173 static uint64_t DMPDphys
; /* phys addr of direct mapped level 2 */
174 static uint64_t DMPDPphys
; /* phys addr of direct mapped level 3 */
177 * Data for the pv entry allocation mechanism
179 static vm_zone_t pvzone
;
180 static struct vm_zone pvzone_store
;
181 static struct vm_object pvzone_obj
;
182 static int pv_entry_count
=0, pv_entry_max
=0, pv_entry_high_water
=0;
183 static int pmap_pagedaemon_waken
= 0;
184 static struct pv_entry
*pvinit
;
187 * All those kernel PT submaps that BSD is so fond of
189 pt_entry_t
*CMAP1
= 0, *ptmmap
;
190 caddr_t CADDR1
= 0, ptvmmap
= 0;
191 static pt_entry_t
*msgbufmap
;
192 struct msgbuf
*msgbufp
=0;
197 static pt_entry_t
*pt_crashdumpmap
;
198 static caddr_t crashdumpmap
;
200 extern uint64_t KPTphys
;
201 extern pt_entry_t
*SMPpt
;
202 extern uint64_t SMPptpa
;
206 static pv_entry_t
get_pv_entry (void);
207 static void i386_protection_init (void);
208 static void create_pagetables(vm_paddr_t
*firstaddr
);
209 static void pmap_remove_all (vm_page_t m
);
210 static void pmap_enter_quick (pmap_t pmap
, vm_offset_t va
, vm_page_t m
);
211 static int pmap_remove_pte (struct pmap
*pmap
, pt_entry_t
*ptq
,
212 vm_offset_t sva
, pmap_inval_info_t info
);
213 static void pmap_remove_page (struct pmap
*pmap
,
214 vm_offset_t va
, pmap_inval_info_t info
);
215 static int pmap_remove_entry (struct pmap
*pmap
, vm_page_t m
,
216 vm_offset_t va
, pmap_inval_info_t info
);
217 static boolean_t
pmap_testbit (vm_page_t m
, int bit
);
218 static void pmap_insert_entry (pmap_t pmap
, vm_offset_t va
,
219 vm_page_t mpte
, vm_page_t m
);
221 static vm_page_t
pmap_allocpte (pmap_t pmap
, vm_offset_t va
);
223 static int pmap_release_free_page (pmap_t pmap
, vm_page_t p
);
224 static vm_page_t
_pmap_allocpte (pmap_t pmap
, vm_pindex_t ptepindex
);
225 static pt_entry_t
* pmap_pte_quick (pmap_t pmap
, vm_offset_t va
);
226 static vm_page_t
pmap_page_lookup (vm_object_t object
, vm_pindex_t pindex
);
227 static int _pmap_unwire_pte_hold(pmap_t pmap
, vm_offset_t va
, vm_page_t m
,
228 pmap_inval_info_t info
);
229 static int pmap_unuse_pt (pmap_t
, vm_offset_t
, vm_page_t
, pmap_inval_info_t
);
230 static vm_offset_t
pmap_kmem_choose(vm_offset_t addr
);
232 static unsigned pdir4mb
;
235 * Move the kernel virtual free pointer to the next
236 * 2MB. This is used to help improve performance
237 * by using a large (2MB) page for much of the kernel
238 * (.text, .data, .bss)
242 pmap_kmem_choose(vm_offset_t addr
)
244 vm_offset_t newaddr
= addr
;
246 newaddr
= (addr
+ (NBPDR
- 1)) & ~(NBPDR
- 1);
253 * Super fast pmap_pte routine best used when scanning the pv lists.
254 * This eliminates many course-grained invltlb calls. Note that many of
255 * the pv list scans are across different pmaps and it is very wasteful
256 * to do an entire invltlb when checking a single mapping.
258 * Should only be called while in a critical section.
260 static __inline pt_entry_t
*pmap_pte(pmap_t pmap
, vm_offset_t va
);
264 pmap_pte_quick(pmap_t pmap
, vm_offset_t va
)
266 return pmap_pte(pmap
, va
);
269 /* Return a non-clipped PD index for a given VA */
272 pmap_pde_pindex(vm_offset_t va
)
274 return va
>> PDRSHIFT
;
277 /* Return various clipped indexes for a given VA */
280 pmap_pte_index(vm_offset_t va
)
283 return ((va
>> PAGE_SHIFT
) & ((1ul << NPTEPGSHIFT
) - 1));
288 pmap_pde_index(vm_offset_t va
)
291 return ((va
>> PDRSHIFT
) & ((1ul << NPDEPGSHIFT
) - 1));
296 pmap_pdpe_index(vm_offset_t va
)
299 return ((va
>> PDPSHIFT
) & ((1ul << NPDPEPGSHIFT
) - 1));
304 pmap_pml4e_index(vm_offset_t va
)
307 return ((va
>> PML4SHIFT
) & ((1ul << NPML4EPGSHIFT
) - 1));
310 /* Return a pointer to the PML4 slot that corresponds to a VA */
313 pmap_pml4e(pmap_t pmap
, vm_offset_t va
)
316 return (&pmap
->pm_pml4
[pmap_pml4e_index(va
)]);
319 /* Return a pointer to the PDP slot that corresponds to a VA */
322 pmap_pml4e_to_pdpe(pml4_entry_t
*pml4e
, vm_offset_t va
)
326 pdpe
= (pdp_entry_t
*)PHYS_TO_DMAP(*pml4e
& PG_FRAME
);
327 return (&pdpe
[pmap_pdpe_index(va
)]);
330 /* Return a pointer to the PDP slot that corresponds to a VA */
333 pmap_pdpe(pmap_t pmap
, vm_offset_t va
)
337 pml4e
= pmap_pml4e(pmap
, va
);
338 if ((*pml4e
& PG_V
) == 0)
340 return (pmap_pml4e_to_pdpe(pml4e
, va
));
343 /* Return a pointer to the PD slot that corresponds to a VA */
346 pmap_pdpe_to_pde(pdp_entry_t
*pdpe
, vm_offset_t va
)
350 pde
= (pd_entry_t
*)PHYS_TO_DMAP(*pdpe
& PG_FRAME
);
351 return (&pde
[pmap_pde_index(va
)]);
354 /* Return a pointer to the PD slot that corresponds to a VA */
357 pmap_pde(pmap_t pmap
, vm_offset_t va
)
361 pdpe
= pmap_pdpe(pmap
, va
);
362 if (pdpe
== NULL
|| (*pdpe
& PG_V
) == 0)
364 return (pmap_pdpe_to_pde(pdpe
, va
));
367 /* Return a pointer to the PT slot that corresponds to a VA */
370 pmap_pde_to_pte(pd_entry_t
*pde
, vm_offset_t va
)
374 pte
= (pt_entry_t
*)PHYS_TO_DMAP(*pde
& PG_FRAME
);
375 return (&pte
[pmap_pte_index(va
)]);
378 /* Return a pointer to the PT slot that corresponds to a VA */
381 pmap_pte(pmap_t pmap
, vm_offset_t va
)
385 pde
= pmap_pde(pmap
, va
);
386 if (pde
== NULL
|| (*pde
& PG_V
) == 0)
388 if ((*pde
& PG_PS
) != 0) /* compat with i386 pmap_pte() */
389 return ((pt_entry_t
*)pde
);
390 return (pmap_pde_to_pte(pde
, va
));
395 vtopte(vm_offset_t va
)
397 uint64_t mask
= ((1ul << (NPTEPGSHIFT
+ NPDEPGSHIFT
+ NPDPEPGSHIFT
+ NPML4EPGSHIFT
)) - 1);
399 return (PTmap
+ ((va
>> PAGE_SHIFT
) & mask
));
404 vtopde(vm_offset_t va
)
406 uint64_t mask
= ((1ul << (NPDEPGSHIFT
+ NPDPEPGSHIFT
+ NPML4EPGSHIFT
)) - 1);
408 return (PDmap
+ ((va
>> PDRSHIFT
) & mask
));
412 allocpages(vm_paddr_t
*firstaddr
, int n
)
417 bzero((void *)ret
, n
* PAGE_SIZE
);
418 *firstaddr
+= n
* PAGE_SIZE
;
424 create_pagetables(vm_paddr_t
*firstaddr
)
428 /* we are running (mostly) V=P at this point */
431 KPTphys
= allocpages(firstaddr
, NKPT
);
432 KPML4phys
= allocpages(firstaddr
, 1);
433 KPDPphys
= allocpages(firstaddr
, NKPML4E
);
434 KPDphys
= allocpages(firstaddr
, NKPDPE
);
436 ndmpdp
= (ptoa(Maxmem
) + NBPDP
- 1) >> PDPSHIFT
;
437 if (ndmpdp
< 4) /* Minimum 4GB of dirmap */
439 DMPDPphys
= allocpages(firstaddr
, NDMPML4E
);
440 if ((amd_feature
& AMDID_PAGE1GB
) == 0)
441 DMPDphys
= allocpages(firstaddr
, ndmpdp
);
442 dmaplimit
= (vm_paddr_t
)ndmpdp
<< PDPSHIFT
;
444 /* Fill in the underlying page table pages */
445 /* Read-only from zero to physfree */
446 /* XXX not fully used, underneath 2M pages */
447 for (i
= 0; (i
<< PAGE_SHIFT
) < *firstaddr
; i
++) {
448 ((pt_entry_t
*)KPTphys
)[i
] = i
<< PAGE_SHIFT
;
449 ((pt_entry_t
*)KPTphys
)[i
] |= PG_RW
| PG_V
| PG_G
;
452 /* Now map the page tables at their location within PTmap */
453 for (i
= 0; i
< NKPT
; i
++) {
454 ((pd_entry_t
*)KPDphys
)[i
] = KPTphys
+ (i
<< PAGE_SHIFT
);
455 ((pd_entry_t
*)KPDphys
)[i
] |= PG_RW
| PG_V
;
458 /* Map from zero to end of allocations under 2M pages */
459 /* This replaces some of the KPTphys entries above */
460 for (i
= 0; (i
<< PDRSHIFT
) < *firstaddr
; i
++) {
461 ((pd_entry_t
*)KPDphys
)[i
] = i
<< PDRSHIFT
;
462 ((pd_entry_t
*)KPDphys
)[i
] |= PG_RW
| PG_V
| PG_PS
| PG_G
;
465 /* And connect up the PD to the PDP */
466 for (i
= 0; i
< NKPDPE
; i
++) {
467 ((pdp_entry_t
*)KPDPphys
)[i
+ KPDPI
] = KPDphys
+
469 ((pdp_entry_t
*)KPDPphys
)[i
+ KPDPI
] |= PG_RW
| PG_V
| PG_U
;
472 /* Now set up the direct map space using either 2MB or 1GB pages */
473 /* Preset PG_M and PG_A because demotion expects it */
474 if ((amd_feature
& AMDID_PAGE1GB
) == 0) {
475 for (i
= 0; i
< NPDEPG
* ndmpdp
; i
++) {
476 ((pd_entry_t
*)DMPDphys
)[i
] = (vm_paddr_t
)i
<< PDRSHIFT
;
477 ((pd_entry_t
*)DMPDphys
)[i
] |= PG_RW
| PG_V
| PG_PS
|
480 /* And the direct map space's PDP */
481 for (i
= 0; i
< ndmpdp
; i
++) {
482 ((pdp_entry_t
*)DMPDPphys
)[i
] = DMPDphys
+
484 ((pdp_entry_t
*)DMPDPphys
)[i
] |= PG_RW
| PG_V
| PG_U
;
487 for (i
= 0; i
< ndmpdp
; i
++) {
488 ((pdp_entry_t
*)DMPDPphys
)[i
] =
489 (vm_paddr_t
)i
<< PDPSHIFT
;
490 ((pdp_entry_t
*)DMPDPphys
)[i
] |= PG_RW
| PG_V
| PG_PS
|
495 /* And recursively map PML4 to itself in order to get PTmap */
496 ((pdp_entry_t
*)KPML4phys
)[PML4PML4I
] = KPML4phys
;
497 ((pdp_entry_t
*)KPML4phys
)[PML4PML4I
] |= PG_RW
| PG_V
| PG_U
;
499 /* Connect the Direct Map slot up to the PML4 */
500 ((pdp_entry_t
*)KPML4phys
)[DMPML4I
] = DMPDPphys
;
501 ((pdp_entry_t
*)KPML4phys
)[DMPML4I
] |= PG_RW
| PG_V
| PG_U
;
503 /* Connect the KVA slot up to the PML4 */
504 ((pdp_entry_t
*)KPML4phys
)[KPML4I
] = KPDPphys
;
505 ((pdp_entry_t
*)KPML4phys
)[KPML4I
] |= PG_RW
| PG_V
| PG_U
;
509 init_paging(vm_paddr_t
*firstaddr
)
511 create_pagetables(firstaddr
);
515 * Bootstrap the system enough to run with virtual memory.
517 * On the i386 this is called after mapping has already been enabled
518 * and just syncs the pmap module with what has already been done.
519 * [We can't call it easily with mapping off since the kernel is not
520 * mapped with PA == VA, hence we would have to relocate every address
521 * from the linked base (virtual) address "KERNBASE" to the actual
522 * (physical) address starting relative to 0]
525 pmap_bootstrap(vm_paddr_t
*firstaddr
)
529 struct mdglobaldata
*gd
;
532 KvaStart
= VM_MIN_KERNEL_ADDRESS
;
533 KvaEnd
= VM_MAX_KERNEL_ADDRESS
;
534 KvaSize
= KvaEnd
- KvaStart
;
536 avail_start
= *firstaddr
;
539 * Create an initial set of page tables to run the kernel in.
541 create_pagetables(firstaddr
);
543 virtual_start
= (vm_offset_t
) PTOV_OFFSET
+ *firstaddr
;
544 virtual_start
= pmap_kmem_choose(virtual_start
);
546 virtual_end
= VM_MAX_KERNEL_ADDRESS
;
548 /* XXX do %cr0 as well */
549 load_cr4(rcr4() | CR4_PGE
| CR4_PSE
);
553 * Initialize protection array.
555 i386_protection_init();
558 * The kernel's pmap is statically allocated so we don't have to use
559 * pmap_create, which is unlikely to work correctly at this part of
560 * the boot sequence (XXX and which no longer exists).
562 kernel_pmap
.pm_pml4
= (pdp_entry_t
*) (PTOV_OFFSET
+ KPML4phys
);
563 kernel_pmap
.pm_count
= 1;
564 kernel_pmap
.pm_active
= (cpumask_t
)-1; /* don't allow deactivation */
565 TAILQ_INIT(&kernel_pmap
.pm_pvlist
);
569 * Reserve some special page table entries/VA space for temporary
572 #define SYSMAP(c, p, v, n) \
573 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
577 pte
= (pt_entry_t
*) pmap_pte(&kernel_pmap
, va
);
583 * CMAP1/CMAP2 are used for zeroing and copying pages.
585 SYSMAP(caddr_t
, CMAP1
, CADDR1
, 1)
590 SYSMAP(caddr_t
, pt_crashdumpmap
, crashdumpmap
, MAXDUMPPGS
);
593 * ptvmmap is used for reading arbitrary physical pages via
596 SYSMAP(caddr_t
, ptmmap
, ptvmmap
, 1)
599 * msgbufp is used to map the system message buffer.
600 * XXX msgbufmap is not used.
602 SYSMAP(struct msgbuf
*, msgbufmap
, msgbufp
,
603 atop(round_page(MSGBUF_SIZE
)))
610 * PG_G is terribly broken on SMP because we IPI invltlb's in some
611 * cases rather then invl1pg. Actually, I don't even know why it
612 * works under UP because self-referential page table mappings
617 if (cpu_feature
& CPUID_PGE
)
622 * Initialize the 4MB page size flag
626 * The 4MB page version of the initial
627 * kernel page mapping.
631 #if !defined(DISABLE_PSE)
632 if (cpu_feature
& CPUID_PSE
) {
635 * Note that we have enabled PSE mode
638 ptditmp
= *(PTmap
+ amd64_btop(KERNBASE
));
639 ptditmp
&= ~(NBPDR
- 1);
640 ptditmp
|= PG_V
| PG_RW
| PG_PS
| PG_U
| pgeflag
;
645 * Enable the PSE mode. If we are SMP we can't do this
646 * now because the APs will not be able to use it when
649 load_cr4(rcr4() | CR4_PSE
);
652 * We can do the mapping here for the single processor
653 * case. We simply ignore the old page table page from
657 * For SMP, we still need 4K pages to bootstrap APs,
658 * PSE will be enabled as soon as all APs are up.
660 PTD
[KPTDI
] = (pd_entry_t
)ptditmp
;
666 if (cpu_apic_address
== 0)
667 panic("pmap_bootstrap: no local apic!");
671 * We need to finish setting up the globaldata page for the BSP.
672 * locore has already populated the page table for the mdglobaldata
675 pg
= MDGLOBALDATA_BASEALLOC_PAGES
;
676 gd
= &CPU_prvspace
[0].mdglobaldata
;
677 gd
->gd_CMAP1
= &SMPpt
[pg
+ 0];
678 gd
->gd_CMAP2
= &SMPpt
[pg
+ 1];
679 gd
->gd_CMAP3
= &SMPpt
[pg
+ 2];
680 gd
->gd_PMAP1
= &SMPpt
[pg
+ 3];
681 gd
->gd_CADDR1
= CPU_prvspace
[0].CPAGE1
;
682 gd
->gd_CADDR2
= CPU_prvspace
[0].CPAGE2
;
683 gd
->gd_CADDR3
= CPU_prvspace
[0].CPAGE3
;
684 gd
->gd_PADDR1
= (pt_entry_t
*)CPU_prvspace
[0].PPAGE1
;
691 * Set 4mb pdir for mp startup
696 if (pseflag
&& (cpu_feature
& CPUID_PSE
)) {
697 load_cr4(rcr4() | CR4_PSE
);
698 if (pdir4mb
&& mycpu
->gd_cpuid
== 0) { /* only on BSP */
706 * Initialize the pmap module.
707 * Called by vm_init, to initialize any structures that the pmap
708 * system needs to map virtual memory.
709 * pmap_init has been enhanced to support in a fairly consistant
710 * way, discontiguous physical memory.
719 * object for kernel page table pages
721 /* JG I think the number can be arbitrary */
722 kptobj
= vm_object_allocate(OBJT_DEFAULT
, 5);
725 * Allocate memory for random pmap data structures. Includes the
729 for(i
= 0; i
< vm_page_array_size
; i
++) {
732 m
= &vm_page_array
[i
];
733 TAILQ_INIT(&m
->md
.pv_list
);
734 m
->md
.pv_list_count
= 0;
738 * init the pv free list
740 initial_pvs
= vm_page_array_size
;
741 if (initial_pvs
< MINPV
)
743 pvzone
= &pvzone_store
;
744 pvinit
= (struct pv_entry
*) kmem_alloc(&kernel_map
,
745 initial_pvs
* sizeof (struct pv_entry
));
746 zbootinit(pvzone
, "PV ENTRY", sizeof (struct pv_entry
), pvinit
,
750 * Now it is safe to enable pv_table recording.
752 pmap_initialized
= TRUE
;
754 lapic
= pmap_mapdev_uncacheable(cpu_apic_address
, sizeof(struct LAPIC
));
759 * Initialize the address space (zone) for the pv_entries. Set a
760 * high water mark so that the system can recover from excessive
761 * numbers of pv entries.
766 int shpgperproc
= PMAP_SHPGPERPROC
;
768 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc
);
769 pv_entry_max
= shpgperproc
* maxproc
+ vm_page_array_size
;
770 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max
);
771 pv_entry_high_water
= 9 * (pv_entry_max
/ 10);
772 zinitna(pvzone
, &pvzone_obj
, NULL
, 0, pv_entry_max
, ZONE_INTERRUPT
, 1);
776 /***************************************************
777 * Low level helper routines.....
778 ***************************************************/
780 #if defined(PMAP_DIAGNOSTIC)
783 * This code checks for non-writeable/modified pages.
784 * This should be an invalid condition.
788 pmap_nw_modified(pt_entry_t pte
)
790 if ((pte
& (PG_M
|PG_RW
)) == PG_M
)
799 * this routine defines the region(s) of memory that should
800 * not be tested for the modified bit.
804 pmap_track_modified(vm_offset_t va
)
806 if ((va
< clean_sva
) || (va
>= clean_eva
))
815 * Extract the physical page address associated with the map/VA pair.
817 * This function may not be called from an interrupt if the pmap is
821 pmap_extract(pmap_t pmap
, vm_offset_t va
)
825 pd_entry_t pde
, *pdep
;
828 pdep
= pmap_pde(pmap
, va
);
832 if ((pde
& PG_PS
) != 0) {
833 rtval
= (pde
& PG_PS_FRAME
) | (va
& PDRMASK
);
835 pte
= pmap_pde_to_pte(pdep
, va
);
836 rtval
= (*pte
& PG_FRAME
) | (va
& PAGE_MASK
);
844 * Routine: pmap_kextract
846 * Extract the physical page address associated
847 * kernel virtual address.
850 pmap_kextract(vm_offset_t va
)
855 if (va
>= DMAP_MIN_ADDRESS
&& va
< DMAP_MAX_ADDRESS
) {
856 pa
= DMAP_TO_PHYS(va
);
860 pa
= (pde
& PG_PS_FRAME
) | (va
& PDRMASK
);
863 * Beware of a concurrent promotion that changes the
864 * PDE at this point! For example, vtopte() must not
865 * be used to access the PTE because it would use the
866 * new PDE. It is, however, safe to use the old PDE
867 * because the page table page is preserved by the
870 pa
= *pmap_pde_to_pte(&pde
, va
);
871 pa
= (pa
& PG_FRAME
) | (va
& PAGE_MASK
);
877 /***************************************************
878 * Low level mapping routines.....
879 ***************************************************/
882 * Routine: pmap_kenter
884 * Add a wired page to the KVA
885 * NOTE! note that in order for the mapping to take effect -- you
886 * should do an invltlb after doing the pmap_kenter().
889 pmap_kenter(vm_offset_t va
, vm_paddr_t pa
)
893 pmap_inval_info info
;
895 pmap_inval_init(&info
);
896 npte
= pa
| PG_RW
| PG_V
| pgeflag
;
898 pmap_inval_add(&info
, &kernel_pmap
, va
);
900 pmap_inval_flush(&info
);
904 * Routine: pmap_kenter_quick
906 * Similar to pmap_kenter(), except we only invalidate the
907 * mapping on the current CPU.
910 pmap_kenter_quick(vm_offset_t va
, vm_paddr_t pa
)
915 npte
= pa
| PG_RW
| PG_V
| pgeflag
;
918 cpu_invlpg((void *)va
);
922 pmap_kenter_sync(vm_offset_t va
)
924 pmap_inval_info info
;
926 pmap_inval_init(&info
);
927 pmap_inval_add(&info
, &kernel_pmap
, va
);
928 pmap_inval_flush(&info
);
932 pmap_kenter_sync_quick(vm_offset_t va
)
934 cpu_invlpg((void *)va
);
938 * remove a page from the kernel pagetables
941 pmap_kremove(vm_offset_t va
)
944 pmap_inval_info info
;
946 pmap_inval_init(&info
);
948 pmap_inval_add(&info
, &kernel_pmap
, va
);
950 pmap_inval_flush(&info
);
954 pmap_kremove_quick(vm_offset_t va
)
959 cpu_invlpg((void *)va
);
963 * XXX these need to be recoded. They are not used in any critical path.
966 pmap_kmodify_rw(vm_offset_t va
)
968 *vtopte(va
) |= PG_RW
;
969 cpu_invlpg((void *)va
);
973 pmap_kmodify_nc(vm_offset_t va
)
976 cpu_invlpg((void *)va
);
980 * Used to map a range of physical addresses into kernel
981 * virtual address space.
983 * For now, VM is already on, we only need to map the
987 pmap_map(vm_offset_t virt
, vm_paddr_t start
, vm_paddr_t end
, int prot
)
989 return PHYS_TO_DMAP(start
);
994 * Add a list of wired pages to the kva
995 * this routine is only used for temporary
996 * kernel mappings that do not need to have
997 * page modification or references recorded.
998 * Note that old mappings are simply written
999 * over. The page *must* be wired.
1002 pmap_qenter(vm_offset_t va
, vm_page_t
*m
, int count
)
1006 end_va
= va
+ count
* PAGE_SIZE
;
1008 while (va
< end_va
) {
1012 *pte
= VM_PAGE_TO_PHYS(*m
) | PG_RW
| PG_V
| pgeflag
;
1013 cpu_invlpg((void *)va
);
1018 smp_invltlb(); /* XXX */
1023 pmap_qenter2(vm_offset_t va
, vm_page_t
*m
, int count
, cpumask_t
*mask
)
1026 cpumask_t cmask
= mycpu
->gd_cpumask
;
1028 end_va
= va
+ count
* PAGE_SIZE
;
1030 while (va
< end_va
) {
1035 * Install the new PTE. If the pte changed from the prior
1036 * mapping we must reset the cpu mask and invalidate the page.
1037 * If the pte is the same but we have not seen it on the
1038 * current cpu, invlpg the existing mapping. Otherwise the
1039 * entry is optimal and no invalidation is required.
1042 pteval
= VM_PAGE_TO_PHYS(*m
) | PG_A
| PG_RW
| PG_V
| pgeflag
;
1043 if (*pte
!= pteval
) {
1046 cpu_invlpg((void *)va
);
1047 } else if ((*mask
& cmask
) == 0) {
1048 cpu_invlpg((void *)va
);
1057 * This routine jerks page mappings from the
1058 * kernel -- it is meant only for temporary mappings.
1060 * MPSAFE, INTERRUPT SAFE (cluster callback)
1063 pmap_qremove(vm_offset_t va
, int count
)
1067 end_va
= va
+ count
* PAGE_SIZE
;
1069 while (va
< end_va
) {
1074 cpu_invlpg((void *)va
);
1083 * This routine works like vm_page_lookup() but also blocks as long as the
1084 * page is busy. This routine does not busy the page it returns.
1086 * Unless the caller is managing objects whos pages are in a known state,
1087 * the call should be made with a critical section held so the page's object
1088 * association remains valid on return.
1092 pmap_page_lookup(vm_object_t object
, vm_pindex_t pindex
)
1097 m
= vm_page_lookup(object
, pindex
);
1098 } while (m
&& vm_page_sleep_busy(m
, FALSE
, "pplookp"));
1104 * Create a new thread and optionally associate it with a (new) process.
1105 * NOTE! the new thread's cpu may not equal the current cpu.
1108 pmap_init_thread(thread_t td
)
1110 /* enforce pcb placement */
1111 td
->td_pcb
= (struct pcb
*)(td
->td_kstack
+ td
->td_kstack_size
) - 1;
1112 td
->td_savefpu
= &td
->td_pcb
->pcb_save
;
1113 td
->td_sp
= (char *)td
->td_pcb
- 16; /* JG is -16 needed on amd64? */
1117 * This routine directly affects the fork perf for a process.
1120 pmap_init_proc(struct proc
*p
)
1125 * Dispose the UPAGES for a process that has exited.
1126 * This routine directly impacts the exit perf of a process.
1129 pmap_dispose_proc(struct proc
*p
)
1131 KASSERT(p
->p_lock
== 0, ("attempt to dispose referenced proc! %p", p
));
1134 /***************************************************
1135 * Page table page management routines.....
1136 ***************************************************/
1139 * This routine unholds page table pages, and if the hold count
1140 * drops to zero, then it decrements the wire count.
1144 pmap_unwire_pte_hold(pmap_t pmap
, vm_offset_t va
, vm_page_t m
,
1145 pmap_inval_info_t info
)
1147 KKASSERT(m
->hold_count
> 0);
1148 if (m
->hold_count
> 1) {
1152 return _pmap_unwire_pte_hold(pmap
, va
, m
, info
);
1158 _pmap_unwire_pte_hold(pmap_t pmap
, vm_offset_t va
, vm_page_t m
,
1159 pmap_inval_info_t info
)
1162 * Wait until we can busy the page ourselves. We cannot have
1163 * any active flushes if we block. We own one hold count on the
1164 * page so it cannot be freed out from under us.
1166 if (m
->flags
& PG_BUSY
) {
1167 pmap_inval_flush(info
);
1168 while (vm_page_sleep_busy(m
, FALSE
, "pmuwpt"))
1171 KASSERT(m
->queue
== PQ_NONE
,
1172 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m
));
1175 * This case can occur if new references were acquired while
1178 if (m
->hold_count
> 1) {
1179 KKASSERT(m
->hold_count
> 1);
1185 * Unmap the page table page
1187 KKASSERT(m
->hold_count
== 1);
1189 pmap_inval_add(info
, pmap
, -1);
1191 if (m
->pindex
>= (NUPDE
+ NUPDPE
)) {
1194 pml4
= pmap_pml4e(pmap
, va
);
1196 } else if (m
->pindex
>= NUPDE
) {
1199 pdp
= pmap_pdpe(pmap
, va
);
1204 pd
= pmap_pde(pmap
, va
);
1208 KKASSERT(pmap
->pm_stats
.resident_count
> 0);
1209 --pmap
->pm_stats
.resident_count
;
1211 if (pmap
->pm_ptphint
== m
)
1212 pmap
->pm_ptphint
= NULL
;
1214 if (m
->pindex
< NUPDE
) {
1215 /* We just released a PT, unhold the matching PD */
1218 pdpg
= PHYS_TO_VM_PAGE(*pmap_pdpe(pmap
, va
) & PG_FRAME
);
1219 pmap_unwire_pte_hold(pmap
, va
, pdpg
, info
);
1221 if (m
->pindex
>= NUPDE
&& m
->pindex
< (NUPDE
+ NUPDPE
)) {
1222 /* We just released a PD, unhold the matching PDP */
1225 pdppg
= PHYS_TO_VM_PAGE(*pmap_pml4e(pmap
, va
) & PG_FRAME
);
1226 pmap_unwire_pte_hold(pmap
, va
, pdppg
, info
);
1230 * This was our last hold, the page had better be unwired
1231 * after we decrement wire_count.
1233 * FUTURE NOTE: shared page directory page could result in
1234 * multiple wire counts.
1238 KKASSERT(m
->wire_count
== 0);
1239 --vmstats
.v_wire_count
;
1240 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
1242 vm_page_free_zero(m
);
1248 * After removing a page table entry, this routine is used to
1249 * conditionally free the page, and manage the hold/wire counts.
1253 pmap_unuse_pt(pmap_t pmap
, vm_offset_t va
, vm_page_t mpte
,
1254 pmap_inval_info_t info
)
1256 vm_pindex_t ptepindex
;
1258 if (va
>= VM_MAX_USER_ADDRESS
)
1262 ptepindex
= pmap_pde_pindex(va
);
1264 if (pmap
->pm_ptphint
&&
1265 (pmap
->pm_ptphint
->pindex
== ptepindex
)) {
1266 mpte
= pmap
->pm_ptphint
;
1269 pmap_inval_flush(info
);
1270 mpte
= pmap_page_lookup(pmap
->pm_pteobj
, ptepindex
);
1271 pmap
->pm_ptphint
= mpte
;
1276 return pmap_unwire_pte_hold(pmap
, va
, mpte
, info
);
1280 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1281 * it, and IdlePTD, represents the template used to update all other pmaps.
1283 * On architectures where the kernel pmap is not integrated into the user
1284 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1285 * kernel_pmap should be used to directly access the kernel_pmap.
1288 pmap_pinit0(struct pmap
*pmap
)
1290 pmap
->pm_pml4
= (pml4_entry_t
*)(PTOV_OFFSET
+ KPML4phys
);
1292 pmap
->pm_active
= 0;
1293 pmap
->pm_ptphint
= NULL
;
1294 TAILQ_INIT(&pmap
->pm_pvlist
);
1295 bzero(&pmap
->pm_stats
, sizeof pmap
->pm_stats
);
1299 * Initialize a preallocated and zeroed pmap structure,
1300 * such as one in a vmspace structure.
1303 pmap_pinit(struct pmap
*pmap
)
1308 * No need to allocate page table space yet but we do need a valid
1309 * page directory table.
1311 if (pmap
->pm_pml4
== NULL
) {
1313 (pml4_entry_t
*)kmem_alloc_pageable(&kernel_map
, PAGE_SIZE
);
1317 * Allocate an object for the ptes
1319 if (pmap
->pm_pteobj
== NULL
)
1320 pmap
->pm_pteobj
= vm_object_allocate(OBJT_DEFAULT
, NUPDE
+ NUPDPE
+ PML4PML4I
+ 1);
1323 * Allocate the page directory page, unless we already have
1324 * one cached. If we used the cached page the wire_count will
1325 * already be set appropriately.
1327 if ((ptdpg
= pmap
->pm_pdirm
) == NULL
) {
1328 ptdpg
= vm_page_grab(pmap
->pm_pteobj
, NUPDE
+ NUPDPE
+ PML4PML4I
,
1329 VM_ALLOC_NORMAL
| VM_ALLOC_RETRY
);
1330 pmap
->pm_pdirm
= ptdpg
;
1331 vm_page_flag_clear(ptdpg
, PG_MAPPED
| PG_BUSY
);
1332 ptdpg
->valid
= VM_PAGE_BITS_ALL
;
1333 if (ptdpg
->wire_count
== 0)
1334 ++vmstats
.v_wire_count
;
1335 ptdpg
->wire_count
= 1;
1336 pmap_kenter((vm_offset_t
)pmap
->pm_pml4
, VM_PAGE_TO_PHYS(ptdpg
));
1338 if ((ptdpg
->flags
& PG_ZERO
) == 0)
1339 bzero(pmap
->pm_pml4
, PAGE_SIZE
);
1341 pmap
->pm_pml4
[KPML4I
] = KPDPphys
| PG_RW
| PG_V
| PG_U
;
1342 pmap
->pm_pml4
[DMPML4I
] = DMPDPphys
| PG_RW
| PG_V
| PG_U
;
1344 /* install self-referential address mapping entry */
1345 pmap
->pm_pml4
[PML4PML4I
] = VM_PAGE_TO_PHYS(ptdpg
) | PG_V
| PG_RW
| PG_A
| PG_M
;
1348 pmap
->pm_active
= 0;
1349 pmap
->pm_ptphint
= NULL
;
1350 TAILQ_INIT(&pmap
->pm_pvlist
);
1351 bzero(&pmap
->pm_stats
, sizeof pmap
->pm_stats
);
1352 pmap
->pm_stats
.resident_count
= 1;
1356 * Clean up a pmap structure so it can be physically freed. This routine
1357 * is called by the vmspace dtor function. A great deal of pmap data is
1358 * left passively mapped to improve vmspace management so we have a bit
1359 * of cleanup work to do here.
1362 pmap_puninit(pmap_t pmap
)
1366 KKASSERT(pmap
->pm_active
== 0);
1367 if ((p
= pmap
->pm_pdirm
) != NULL
) {
1368 KKASSERT(pmap
->pm_pml4
!= NULL
);
1369 KKASSERT(pmap
->pm_pml4
!= (void *)(PTOV_OFFSET
+ KPML4phys
));
1370 pmap_kremove((vm_offset_t
)pmap
->pm_pml4
);
1372 vmstats
.v_wire_count
--;
1373 KKASSERT((p
->flags
& PG_BUSY
) == 0);
1375 vm_page_free_zero(p
);
1376 pmap
->pm_pdirm
= NULL
;
1378 if (pmap
->pm_pml4
) {
1379 KKASSERT(pmap
->pm_pml4
!= (void *)(PTOV_OFFSET
+ KPML4phys
));
1380 kmem_free(&kernel_map
, (vm_offset_t
)pmap
->pm_pml4
, PAGE_SIZE
);
1381 pmap
->pm_pml4
= NULL
;
1383 if (pmap
->pm_pteobj
) {
1384 vm_object_deallocate(pmap
->pm_pteobj
);
1385 pmap
->pm_pteobj
= NULL
;
1390 * Wire in kernel global address entries. To avoid a race condition
1391 * between pmap initialization and pmap_growkernel, this procedure
1392 * adds the pmap to the master list (which growkernel scans to update),
1393 * then copies the template.
1396 pmap_pinit2(struct pmap
*pmap
)
1399 TAILQ_INSERT_TAIL(&pmap_list
, pmap
, pm_pmnode
);
1400 /* XXX copies current process, does not fill in MPPTDI */
1405 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1406 * 0 on failure (if the procedure had to sleep).
1408 * When asked to remove the page directory page itself, we actually just
1409 * leave it cached so we do not have to incur the SMP inval overhead of
1410 * removing the kernel mapping. pmap_puninit() will take care of it.
1414 pmap_release_free_page(struct pmap
*pmap
, vm_page_t p
)
1417 * This code optimizes the case of freeing non-busy
1418 * page-table pages. Those pages are zero now, and
1419 * might as well be placed directly into the zero queue.
1421 if (vm_page_sleep_busy(p
, FALSE
, "pmaprl"))
1427 * Remove the page table page from the processes address space.
1429 if (p
->pindex
== NUPDE
+ NUPDPE
+ PML4PML4I
) {
1431 * We are the pml4 table itself.
1433 /* XXX anything to do here? */
1434 } else if (p
->pindex
>= (NUPDE
+ NUPDPE
)) {
1436 * Remove a PDP page from the PML4. We do not maintain
1437 * hold counts on the PML4 page.
1443 m4
= vm_page_lookup(pmap
->pm_pteobj
, NUPDE
+ NUPDPE
+ PML4PML4I
);
1444 KKASSERT(m4
!= NULL
);
1445 pml4
= (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4
));
1446 idx
= (p
->pindex
- (NUPDE
+ NUPDPE
)) % NPML4EPG
;
1447 KKASSERT(pml4
[idx
] != 0);
1449 } else if (p
->pindex
>= NUPDE
) {
1451 * Remove a PD page from the PDP and drop the hold count
1452 * on the PDP. The PDP is left cached in the pmap if
1453 * the hold count drops to 0 so the wire count remains
1460 m3
= vm_page_lookup(pmap
->pm_pteobj
,
1461 NUPDE
+ NUPDPE
+ (p
->pindex
- NUPDE
) / NPDPEPG
);
1462 KKASSERT(m3
!= NULL
);
1463 pdp
= (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3
));
1464 idx
= (p
->pindex
- NUPDE
) % NPDPEPG
;
1465 KKASSERT(pdp
[idx
] != 0);
1470 * Remove a PT page from the PD and drop the hold count
1471 * on the PD. The PD is left cached in the pmap if
1472 * the hold count drops to 0 so the wire count remains
1479 m2
= vm_page_lookup(pmap
->pm_pteobj
,
1480 NUPDE
+ p
->pindex
/ NPDEPG
);
1481 KKASSERT(m2
!= NULL
);
1482 pd
= (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2
));
1483 idx
= p
->pindex
% NPDEPG
;
1489 * One fewer mappings in the pmap. p's hold count had better
1492 KKASSERT(pmap
->pm_stats
.resident_count
> 0);
1493 --pmap
->pm_stats
.resident_count
;
1495 panic("pmap_release: freeing held page table page");
1496 if (pmap
->pm_ptphint
&& (pmap
->pm_ptphint
->pindex
== p
->pindex
))
1497 pmap
->pm_ptphint
= NULL
;
1500 * We leave the top-level page table page cached, wired, and mapped in
1501 * the pmap until the dtor function (pmap_puninit()) gets called.
1502 * However, still clean it up so we can set PG_ZERO.
1504 if (p
->pindex
== NUPDE
+ NUPDPE
+ PML4PML4I
) {
1505 bzero(pmap
->pm_pml4
, PAGE_SIZE
);
1506 vm_page_flag_set(p
, PG_ZERO
);
1510 KKASSERT(p
->wire_count
== 0);
1511 vmstats
.v_wire_count
--;
1512 /* JG eventually revert to using vm_page_free_zero() */
1519 * This routine is called when various levels in the page table need to
1520 * be populated. This routine cannot fail.
1524 _pmap_allocpte(pmap_t pmap
, vm_pindex_t ptepindex
)
1529 * Find or fabricate a new pagetable page. This will busy the page.
1531 m
= vm_page_grab(pmap
->pm_pteobj
, ptepindex
,
1532 VM_ALLOC_NORMAL
| VM_ALLOC_ZERO
| VM_ALLOC_RETRY
);
1533 if ((m
->flags
& PG_ZERO
) == 0) {
1534 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
1537 KASSERT(m
->queue
== PQ_NONE
,
1538 ("_pmap_allocpte: %p->queue != PQ_NONE", m
));
1541 * Increment the hold count for the page we will be returning to
1545 if (m
->wire_count
++ == 0)
1546 vmstats
.v_wire_count
++;
1549 * Map the pagetable page into the process address space, if
1550 * it isn't already there.
1552 * It is possible that someone else got in and mapped the page
1553 * directory page while we were blocked, if so just unbusy and
1554 * return the held page.
1556 if (ptepindex
>= (NUPDE
+ NUPDPE
)) {
1558 * Wire up a new PDP page in the PML4
1560 vm_pindex_t pml4index
;
1563 pml4index
= ptepindex
- (NUPDE
+ NUPDPE
);
1564 pml4
= &pmap
->pm_pml4
[pml4index
];
1566 if (--m
->wire_count
== 0)
1567 --vmstats
.v_wire_count
;
1571 *pml4
= VM_PAGE_TO_PHYS(m
) | PG_U
| PG_RW
| PG_V
| PG_A
| PG_M
;
1572 } else if (ptepindex
>= NUPDE
) {
1574 * Wire up a new PD page in the PDP
1576 vm_pindex_t pml4index
;
1577 vm_pindex_t pdpindex
;
1582 pdpindex
= ptepindex
- NUPDE
;
1583 pml4index
= pdpindex
>> NPML4EPGSHIFT
;
1585 pml4
= &pmap
->pm_pml4
[pml4index
];
1586 if ((*pml4
& PG_V
) == 0) {
1588 * Have to allocate a new PDP page, recurse.
1589 * This always succeeds. Returned page will
1592 pdppg
= _pmap_allocpte(pmap
,
1593 NUPDE
+ NUPDPE
+ pml4index
);
1596 * Add a held reference to the PDP page.
1598 pdppg
= PHYS_TO_VM_PAGE(*pml4
& PG_FRAME
);
1599 pdppg
->hold_count
++;
1603 * Now find the pdp_entry and map the PDP. If the PDP
1604 * has already been mapped unwind and return the
1605 * already-mapped PDP held.
1607 * pdppg is left held (hold_count is incremented for
1608 * each PD in the PDP).
1610 pdp
= (pdp_entry_t
*)PHYS_TO_DMAP(*pml4
& PG_FRAME
);
1611 pdp
= &pdp
[pdpindex
& ((1ul << NPDPEPGSHIFT
) - 1)];
1613 vm_page_unhold(pdppg
);
1614 if (--m
->wire_count
== 0)
1615 --vmstats
.v_wire_count
;
1619 *pdp
= VM_PAGE_TO_PHYS(m
) | PG_U
| PG_RW
| PG_V
| PG_A
| PG_M
;
1622 * Wire up the new PT page in the PD
1624 vm_pindex_t pml4index
;
1625 vm_pindex_t pdpindex
;
1631 pdpindex
= ptepindex
>> NPDPEPGSHIFT
;
1632 pml4index
= pdpindex
>> NPML4EPGSHIFT
;
1635 * Locate the PDP page in the PML4, then the PD page in
1636 * the PDP. If either does not exist we simply recurse
1639 * We can just recurse on the PD page as it will recurse
1640 * on the PDP if necessary.
1642 pml4
= &pmap
->pm_pml4
[pml4index
];
1643 if ((*pml4
& PG_V
) == 0) {
1644 pdpg
= _pmap_allocpte(pmap
, NUPDE
+ pdpindex
);
1645 pdp
= (pdp_entry_t
*)PHYS_TO_DMAP(*pml4
& PG_FRAME
);
1646 pdp
= &pdp
[pdpindex
& ((1ul << NPDPEPGSHIFT
) - 1)];
1648 pdp
= (pdp_entry_t
*)PHYS_TO_DMAP(*pml4
& PG_FRAME
);
1649 pdp
= &pdp
[pdpindex
& ((1ul << NPDPEPGSHIFT
) - 1)];
1650 if ((*pdp
& PG_V
) == 0) {
1651 pdpg
= _pmap_allocpte(pmap
, NUPDE
+ pdpindex
);
1653 pdpg
= PHYS_TO_VM_PAGE(*pdp
& PG_FRAME
);
1659 * Now fill in the pte in the PD. If the pte already exists
1660 * (again, if we raced the grab), unhold pdpg and unwire
1661 * m, returning a held m.
1663 * pdpg is left held (hold_count is incremented for
1664 * each PT in the PD).
1666 pd
= (pd_entry_t
*)PHYS_TO_DMAP(*pdp
& PG_FRAME
);
1667 pd
= &pd
[ptepindex
& ((1ul << NPDEPGSHIFT
) - 1)];
1669 vm_page_unhold(pdpg
);
1670 if (--m
->wire_count
== 0)
1671 --vmstats
.v_wire_count
;
1675 *pd
= VM_PAGE_TO_PHYS(m
) | PG_U
| PG_RW
| PG_V
| PG_A
| PG_M
;
1679 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1680 * valid bits, mapped flag, unbusy, and we're done.
1682 pmap
->pm_ptphint
= m
;
1683 ++pmap
->pm_stats
.resident_count
;
1685 m
->valid
= VM_PAGE_BITS_ALL
;
1686 vm_page_flag_clear(m
, PG_ZERO
);
1687 vm_page_flag_set(m
, PG_MAPPED
);
1695 pmap_allocpte(pmap_t pmap
, vm_offset_t va
)
1697 vm_pindex_t ptepindex
;
1702 * Calculate pagetable page index
1704 ptepindex
= pmap_pde_pindex(va
);
1707 * Get the page directory entry
1709 pd
= pmap_pde(pmap
, va
);
1712 * This supports switching from a 2MB page to a
1715 if (pd
!= NULL
&& (*pd
& (PG_PS
| PG_V
)) == (PG_PS
| PG_V
)) {
1716 panic("no promotion/demotion yet");
1724 * If the page table page is mapped, we just increment the
1725 * hold count, and activate it.
1727 if (pd
!= NULL
&& (*pd
& PG_V
) != 0) {
1728 /* YYY hint is used here on i386 */
1729 m
= pmap_page_lookup( pmap
->pm_pteobj
, ptepindex
);
1730 pmap
->pm_ptphint
= m
;
1735 * Here if the pte page isn't mapped, or if it has been deallocated.
1737 return _pmap_allocpte(pmap
, ptepindex
);
1741 /***************************************************
1742 * Pmap allocation/deallocation routines.
1743 ***************************************************/
1746 * Release any resources held by the given physical map.
1747 * Called when a pmap initialized by pmap_pinit is being released.
1748 * Should only be called if the map contains no valid mappings.
1750 static int pmap_release_callback(struct vm_page
*p
, void *data
);
1753 pmap_release(struct pmap
*pmap
)
1755 vm_object_t object
= pmap
->pm_pteobj
;
1756 struct rb_vm_page_scan_info info
;
1758 KASSERT(pmap
->pm_active
== 0, ("pmap still active! %08x", pmap
->pm_active
));
1759 #if defined(DIAGNOSTIC)
1760 if (object
->ref_count
!= 1)
1761 panic("pmap_release: pteobj reference count != 1");
1765 info
.object
= object
;
1767 TAILQ_REMOVE(&pmap_list
, pmap
, pm_pmnode
);
1774 info
.limit
= object
->generation
;
1776 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, NULL
,
1777 pmap_release_callback
, &info
);
1778 if (info
.error
== 0 && info
.mpte
) {
1779 if (!pmap_release_free_page(pmap
, info
.mpte
))
1783 } while (info
.error
);
1788 pmap_release_callback(struct vm_page
*p
, void *data
)
1790 struct rb_vm_page_scan_info
*info
= data
;
1792 if (p
->pindex
== NUPDE
+ NUPDPE
+ PML4PML4I
) {
1796 if (!pmap_release_free_page(info
->pmap
, p
)) {
1800 if (info
->object
->generation
!= info
->limit
) {
1808 * Grow the number of kernel page table entries, if needed.
1811 pmap_growkernel(vm_offset_t addr
)
1814 vm_offset_t ptppaddr
;
1816 pd_entry_t
*pde
, newpdir
;
1820 if (kernel_vm_end
== 0) {
1821 kernel_vm_end
= KERNBASE
;
1823 while ((*pmap_pde(&kernel_pmap
, kernel_vm_end
) & PG_V
) != 0) {
1824 kernel_vm_end
= (kernel_vm_end
+ PAGE_SIZE
* NPTEPG
) & ~(PAGE_SIZE
* NPTEPG
- 1);
1826 if (kernel_vm_end
- 1 >= kernel_map
.max_offset
) {
1827 kernel_vm_end
= kernel_map
.max_offset
;
1832 addr
= roundup2(addr
, PAGE_SIZE
* NPTEPG
);
1833 if (addr
- 1 >= kernel_map
.max_offset
)
1834 addr
= kernel_map
.max_offset
;
1835 while (kernel_vm_end
< addr
) {
1836 pde
= pmap_pde(&kernel_pmap
, kernel_vm_end
);
1838 /* We need a new PDP entry */
1839 nkpg
= vm_page_alloc(kptobj
, nkpt
,
1840 VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
1841 | VM_ALLOC_INTERRUPT
);
1843 panic("pmap_growkernel: no memory to grow kernel");
1844 paddr
= VM_PAGE_TO_PHYS(nkpg
);
1845 if ((nkpg
->flags
& PG_ZERO
) == 0)
1846 pmap_zero_page(paddr
);
1847 vm_page_flag_clear(nkpg
, PG_ZERO
);
1848 newpdp
= (pdp_entry_t
)
1849 (paddr
| PG_V
| PG_RW
| PG_A
| PG_M
);
1850 *pmap_pdpe(&kernel_pmap
, kernel_vm_end
) = newpdp
;
1852 continue; /* try again */
1854 if ((*pde
& PG_V
) != 0) {
1855 kernel_vm_end
= (kernel_vm_end
+ PAGE_SIZE
* NPTEPG
) & ~(PAGE_SIZE
* NPTEPG
- 1);
1856 if (kernel_vm_end
- 1 >= kernel_map
.max_offset
) {
1857 kernel_vm_end
= kernel_map
.max_offset
;
1864 * This index is bogus, but out of the way
1866 nkpg
= vm_page_alloc(kptobj
, nkpt
,
1867 VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
| VM_ALLOC_INTERRUPT
);
1869 panic("pmap_growkernel: no memory to grow kernel");
1872 ptppaddr
= VM_PAGE_TO_PHYS(nkpg
);
1873 pmap_zero_page(ptppaddr
);
1874 vm_page_flag_clear(nkpg
, PG_ZERO
);
1875 newpdir
= (pd_entry_t
) (ptppaddr
| PG_V
| PG_RW
| PG_A
| PG_M
);
1876 *pmap_pde(&kernel_pmap
, kernel_vm_end
) = newpdir
;
1879 kernel_vm_end
= (kernel_vm_end
+ PAGE_SIZE
* NPTEPG
) & ~(PAGE_SIZE
* NPTEPG
- 1);
1880 if (kernel_vm_end
- 1 >= kernel_map
.max_offset
) {
1881 kernel_vm_end
= kernel_map
.max_offset
;
1889 * Retire the given physical map from service.
1890 * Should only be called if the map contains
1891 * no valid mappings.
1894 pmap_destroy(pmap_t pmap
)
1901 count
= --pmap
->pm_count
;
1904 panic("destroying a pmap is not yet implemented");
1909 * Add a reference to the specified pmap.
1912 pmap_reference(pmap_t pmap
)
1919 /***************************************************
1920 * page management routines.
1921 ***************************************************/
1924 * free the pv_entry back to the free list. This function may be
1925 * called from an interrupt.
1929 free_pv_entry(pv_entry_t pv
)
1932 KKASSERT(pv_entry_count
>= 0);
1937 * get a new pv_entry, allocating a block from the system
1938 * when needed. This function may be called from an interrupt.
1945 if (pv_entry_high_water
&&
1946 (pv_entry_count
> pv_entry_high_water
) &&
1947 (pmap_pagedaemon_waken
== 0)) {
1948 pmap_pagedaemon_waken
= 1;
1949 wakeup(&vm_pages_needed
);
1951 return zalloc(pvzone
);
1955 * This routine is very drastic, but can save the system
1963 static int warningdone
=0;
1965 if (pmap_pagedaemon_waken
== 0)
1968 if (warningdone
< 5) {
1969 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1973 for(i
= 0; i
< vm_page_array_size
; i
++) {
1974 m
= &vm_page_array
[i
];
1975 if (m
->wire_count
|| m
->hold_count
|| m
->busy
||
1976 (m
->flags
& PG_BUSY
))
1980 pmap_pagedaemon_waken
= 0;
1985 * If it is the first entry on the list, it is actually
1986 * in the header and we must copy the following entry up
1987 * to the header. Otherwise we must search the list for
1988 * the entry. In either case we free the now unused entry.
1992 pmap_remove_entry(struct pmap
*pmap
, vm_page_t m
,
1993 vm_offset_t va
, pmap_inval_info_t info
)
1999 if (m
->md
.pv_list_count
< pmap
->pm_stats
.resident_count
) {
2000 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
2001 if (pmap
== pv
->pv_pmap
&& va
== pv
->pv_va
)
2005 TAILQ_FOREACH(pv
, &pmap
->pm_pvlist
, pv_plist
) {
2006 if (va
== pv
->pv_va
)
2012 /* JGXXX When can 'pv' be NULL? */
2014 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
2015 m
->md
.pv_list_count
--;
2016 KKASSERT(m
->md
.pv_list_count
>= 0);
2017 if (TAILQ_EMPTY(&m
->md
.pv_list
))
2018 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
2019 TAILQ_REMOVE(&pmap
->pm_pvlist
, pv
, pv_plist
);
2020 ++pmap
->pm_generation
;
2021 rtval
= pmap_unuse_pt(pmap
, va
, pv
->pv_ptem
, info
);
2029 * Create a pv entry for page at pa for
2034 pmap_insert_entry(pmap_t pmap
, vm_offset_t va
, vm_page_t mpte
, vm_page_t m
)
2039 pv
= get_pv_entry();
2044 TAILQ_INSERT_TAIL(&pmap
->pm_pvlist
, pv
, pv_plist
);
2045 TAILQ_INSERT_TAIL(&m
->md
.pv_list
, pv
, pv_list
);
2046 m
->md
.pv_list_count
++;
2052 * pmap_remove_pte: do the things to unmap a page in a process
2056 pmap_remove_pte(struct pmap
*pmap
, pt_entry_t
*ptq
, vm_offset_t va
,
2057 pmap_inval_info_t info
)
2062 pmap_inval_add(info
, pmap
, va
);
2063 oldpte
= pte_load_clear(ptq
);
2065 pmap
->pm_stats
.wired_count
-= 1;
2067 * Machines that don't support invlpg, also don't support
2068 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2072 cpu_invlpg((void *)va
);
2073 KKASSERT(pmap
->pm_stats
.resident_count
> 0);
2074 --pmap
->pm_stats
.resident_count
;
2075 if (oldpte
& PG_MANAGED
) {
2076 m
= PHYS_TO_VM_PAGE(oldpte
);
2077 if (oldpte
& PG_M
) {
2078 #if defined(PMAP_DIAGNOSTIC)
2079 if (pmap_nw_modified((pt_entry_t
) oldpte
)) {
2081 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2085 if (pmap_track_modified(va
))
2089 vm_page_flag_set(m
, PG_REFERENCED
);
2090 return pmap_remove_entry(pmap
, m
, va
, info
);
2092 return pmap_unuse_pt(pmap
, va
, NULL
, info
);
2101 * Remove a single page from a process address space.
2103 * This function may not be called from an interrupt if the pmap is
2108 pmap_remove_page(struct pmap
*pmap
, vm_offset_t va
, pmap_inval_info_t info
)
2112 pte
= pmap_pte(pmap
, va
);
2115 if ((*pte
& PG_V
) == 0)
2117 pmap_remove_pte(pmap
, pte
, va
, info
);
2123 * Remove the given range of addresses from the specified map.
2125 * It is assumed that the start and end are properly
2126 * rounded to the page size.
2128 * This function may not be called from an interrupt if the pmap is
2132 pmap_remove(struct pmap
*pmap
, vm_offset_t sva
, vm_offset_t eva
)
2134 vm_offset_t va_next
;
2135 pml4_entry_t
*pml4e
;
2137 pd_entry_t ptpaddr
, *pde
;
2139 struct pmap_inval_info info
;
2144 if (pmap
->pm_stats
.resident_count
== 0)
2147 pmap_inval_init(&info
);
2150 * special handling of removing one page. a very
2151 * common operation and easy to short circuit some
2154 if (sva
+ PAGE_SIZE
== eva
) {
2155 pde
= pmap_pde(pmap
, sva
);
2156 if (pde
&& (*pde
& PG_PS
) == 0) {
2157 pmap_remove_page(pmap
, sva
, &info
);
2158 pmap_inval_flush(&info
);
2163 for (; sva
< eva
; sva
= va_next
) {
2164 pml4e
= pmap_pml4e(pmap
, sva
);
2165 if ((*pml4e
& PG_V
) == 0) {
2166 va_next
= (sva
+ NBPML4
) & ~PML4MASK
;
2172 pdpe
= pmap_pml4e_to_pdpe(pml4e
, sva
);
2173 if ((*pdpe
& PG_V
) == 0) {
2174 va_next
= (sva
+ NBPDP
) & ~PDPMASK
;
2181 * Calculate index for next page table.
2183 va_next
= (sva
+ NBPDR
) & ~PDRMASK
;
2187 pde
= pmap_pdpe_to_pde(pdpe
, sva
);
2191 * Weed out invalid mappings.
2197 * Check for large page.
2199 if ((ptpaddr
& PG_PS
) != 0) {
2200 /* JG FreeBSD has more complex treatment here */
2201 pmap_inval_add(&info
, pmap
, -1);
2203 pmap
->pm_stats
.resident_count
-= NBPDR
/ PAGE_SIZE
;
2208 * Limit our scan to either the end of the va represented
2209 * by the current page table page, or to the end of the
2210 * range being removed.
2216 * NOTE: pmap_remove_pte() can block.
2218 for (pte
= pmap_pde_to_pte(pde
, sva
); sva
!= va_next
; pte
++,
2222 if (pmap_remove_pte(pmap
, pte
, sva
, &info
))
2226 pmap_inval_flush(&info
);
2232 * Removes this physical page from all physical maps in which it resides.
2233 * Reflects back modify bits to the pager.
2235 * This routine may not be called from an interrupt.
2240 pmap_remove_all(vm_page_t m
)
2242 struct pmap_inval_info info
;
2243 pt_entry_t
*pte
, tpte
;
2246 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
2249 pmap_inval_init(&info
);
2251 while ((pv
= TAILQ_FIRST(&m
->md
.pv_list
)) != NULL
) {
2252 KKASSERT(pv
->pv_pmap
->pm_stats
.resident_count
> 0);
2253 --pv
->pv_pmap
->pm_stats
.resident_count
;
2255 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_va
);
2256 pmap_inval_add(&info
, pv
->pv_pmap
, pv
->pv_va
);
2257 tpte
= pte_load_clear(pte
);
2260 pv
->pv_pmap
->pm_stats
.wired_count
--;
2263 vm_page_flag_set(m
, PG_REFERENCED
);
2266 * Update the vm_page_t clean and reference bits.
2269 #if defined(PMAP_DIAGNOSTIC)
2270 if (pmap_nw_modified(tpte
)) {
2272 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2276 if (pmap_track_modified(pv
->pv_va
))
2279 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
2280 TAILQ_REMOVE(&pv
->pv_pmap
->pm_pvlist
, pv
, pv_plist
);
2281 ++pv
->pv_pmap
->pm_generation
;
2282 m
->md
.pv_list_count
--;
2283 KKASSERT(m
->md
.pv_list_count
>= 0);
2284 if (TAILQ_EMPTY(&m
->md
.pv_list
))
2285 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
2286 pmap_unuse_pt(pv
->pv_pmap
, pv
->pv_va
, pv
->pv_ptem
, &info
);
2290 KKASSERT((m
->flags
& (PG_MAPPED
|PG_WRITEABLE
)) == 0);
2291 pmap_inval_flush(&info
);
2297 * Set the physical protection on the specified range of this map
2300 * This function may not be called from an interrupt if the map is
2301 * not the kernel_pmap.
2304 pmap_protect(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
, vm_prot_t prot
)
2306 vm_offset_t va_next
;
2307 pml4_entry_t
*pml4e
;
2309 pd_entry_t ptpaddr
, *pde
;
2311 pmap_inval_info info
;
2313 /* JG review for NX */
2318 if ((prot
& VM_PROT_READ
) == VM_PROT_NONE
) {
2319 pmap_remove(pmap
, sva
, eva
);
2323 if (prot
& VM_PROT_WRITE
)
2326 pmap_inval_init(&info
);
2328 for (; sva
< eva
; sva
= va_next
) {
2330 pml4e
= pmap_pml4e(pmap
, sva
);
2331 if ((*pml4e
& PG_V
) == 0) {
2332 va_next
= (sva
+ NBPML4
) & ~PML4MASK
;
2338 pdpe
= pmap_pml4e_to_pdpe(pml4e
, sva
);
2339 if ((*pdpe
& PG_V
) == 0) {
2340 va_next
= (sva
+ NBPDP
) & ~PDPMASK
;
2346 va_next
= (sva
+ NBPDR
) & ~PDRMASK
;
2350 pde
= pmap_pdpe_to_pde(pdpe
, sva
);
2354 * Check for large page.
2356 if ((ptpaddr
& PG_PS
) != 0) {
2357 pmap_inval_add(&info
, pmap
, -1);
2358 *pde
&= ~(PG_M
|PG_RW
);
2359 pmap
->pm_stats
.resident_count
-= NBPDR
/ PAGE_SIZE
;
2364 * Weed out invalid mappings. Note: we assume that the page
2365 * directory table is always allocated, and in kernel virtual.
2373 for (pte
= pmap_pde_to_pte(pde
, sva
); sva
!= va_next
; pte
++,
2375 pt_entry_t obits
, pbits
;
2379 * XXX non-optimal. Note also that there can be
2380 * no pmap_inval_flush() calls until after we modify
2381 * ptbase[sindex] (or otherwise we have to do another
2382 * pmap_inval_add() call).
2384 pmap_inval_add(&info
, pmap
, sva
);
2385 obits
= pbits
= *pte
;
2386 if ((pbits
& PG_V
) == 0)
2388 if (pbits
& PG_MANAGED
) {
2391 m
= PHYS_TO_VM_PAGE(pbits
& PG_FRAME
);
2392 vm_page_flag_set(m
, PG_REFERENCED
);
2396 if (pmap_track_modified(sva
)) {
2398 m
= PHYS_TO_VM_PAGE(pbits
& PG_FRAME
);
2407 if (pbits
!= obits
) {
2412 pmap_inval_flush(&info
);
2416 * Insert the given physical page (p) at
2417 * the specified virtual address (v) in the
2418 * target physical map with the protection requested.
2420 * If specified, the page will be wired down, meaning
2421 * that the related pte can not be reclaimed.
2423 * NB: This is the only routine which MAY NOT lazy-evaluate
2424 * or lose information. That is, this routine must actually
2425 * insert this page into the given map NOW.
2428 pmap_enter(pmap_t pmap
, vm_offset_t va
, vm_page_t m
, vm_prot_t prot
,
2435 pt_entry_t origpte
, newpte
;
2437 pmap_inval_info info
;
2442 va
= trunc_page(va
);
2443 #ifdef PMAP_DIAGNOSTIC
2445 panic("pmap_enter: toobig");
2446 if ((va
>= UPT_MIN_ADDRESS
) && (va
< UPT_MAX_ADDRESS
))
2447 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va
);
2449 if (va
< UPT_MAX_ADDRESS
&& pmap
== &kernel_pmap
) {
2450 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2452 db_print_backtrace();
2455 if (va
>= UPT_MAX_ADDRESS
&& pmap
!= &kernel_pmap
) {
2456 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2458 db_print_backtrace();
2463 * In the case that a page table page is not
2464 * resident, we are creating it here.
2466 if (va
< VM_MAX_USER_ADDRESS
)
2467 mpte
= pmap_allocpte(pmap
, va
);
2471 pmap_inval_init(&info
);
2472 pde
= pmap_pde(pmap
, va
);
2473 if (pde
!= NULL
&& (*pde
& PG_V
) != 0) {
2474 if ((*pde
& PG_PS
) != 0)
2475 panic("pmap_enter: attempted pmap_enter on 2MB page");
2476 pte
= pmap_pde_to_pte(pde
, va
);
2478 panic("pmap_enter: invalid page directory va=%#lx", va
);
2480 KKASSERT(pte
!= NULL
);
2481 pa
= VM_PAGE_TO_PHYS(m
);
2483 opa
= origpte
& PG_FRAME
;
2486 * Mapping has not changed, must be protection or wiring change.
2488 if (origpte
&& (opa
== pa
)) {
2490 * Wiring change, just update stats. We don't worry about
2491 * wiring PT pages as they remain resident as long as there
2492 * are valid mappings in them. Hence, if a user page is wired,
2493 * the PT page will be also.
2495 if (wired
&& ((origpte
& PG_W
) == 0))
2496 pmap
->pm_stats
.wired_count
++;
2497 else if (!wired
&& (origpte
& PG_W
))
2498 pmap
->pm_stats
.wired_count
--;
2500 #if defined(PMAP_DIAGNOSTIC)
2501 if (pmap_nw_modified(origpte
)) {
2503 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2509 * Remove the extra pte reference. Note that we cannot
2510 * optimize the RO->RW case because we have adjusted the
2511 * wiring count above and may need to adjust the wiring
2518 * We might be turning off write access to the page,
2519 * so we go ahead and sense modify status.
2521 if (origpte
& PG_MANAGED
) {
2522 if ((origpte
& PG_M
) && pmap_track_modified(va
)) {
2524 om
= PHYS_TO_VM_PAGE(opa
);
2528 KKASSERT(m
->flags
& PG_MAPPED
);
2533 * Mapping has changed, invalidate old range and fall through to
2534 * handle validating new mapping.
2538 err
= pmap_remove_pte(pmap
, pte
, va
, &info
);
2540 panic("pmap_enter: pte vanished, va: 0x%lx", va
);
2544 * Enter on the PV list if part of our managed memory. Note that we
2545 * raise IPL while manipulating pv_table since pmap_enter can be
2546 * called at interrupt time.
2548 if (pmap_initialized
&&
2549 (m
->flags
& (PG_FICTITIOUS
|PG_UNMANAGED
)) == 0) {
2550 pmap_insert_entry(pmap
, va
, mpte
, m
);
2552 vm_page_flag_set(m
, PG_MAPPED
);
2556 * Increment counters
2558 ++pmap
->pm_stats
.resident_count
;
2560 pmap
->pm_stats
.wired_count
++;
2564 * Now validate mapping with desired protection/wiring.
2566 newpte
= (pt_entry_t
) (pa
| pte_prot(pmap
, prot
) | PG_V
);
2570 if (va
< VM_MAX_USER_ADDRESS
)
2572 if (pmap
== &kernel_pmap
)
2576 * if the mapping or permission bits are different, we need
2577 * to update the pte.
2579 if ((origpte
& ~(PG_M
|PG_A
)) != newpte
) {
2580 pmap_inval_add(&info
, pmap
, va
);
2581 *pte
= newpte
| PG_A
;
2583 vm_page_flag_set(m
, PG_WRITEABLE
);
2585 KKASSERT((newpte
& PG_MANAGED
) == 0 || (m
->flags
& PG_MAPPED
));
2586 pmap_inval_flush(&info
);
2590 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2591 * This code also assumes that the pmap has no pre-existing entry for this
2594 * This code currently may only be used on user pmaps, not kernel_pmap.
2598 pmap_enter_quick(pmap_t pmap
, vm_offset_t va
, vm_page_t m
)
2603 vm_pindex_t ptepindex
;
2605 pmap_inval_info info
;
2607 pmap_inval_init(&info
);
2609 if (va
< UPT_MAX_ADDRESS
&& pmap
== &kernel_pmap
) {
2610 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2612 db_print_backtrace();
2615 if (va
>= UPT_MAX_ADDRESS
&& pmap
!= &kernel_pmap
) {
2616 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2618 db_print_backtrace();
2622 KKASSERT(va
< UPT_MIN_ADDRESS
); /* assert used on user pmaps only */
2625 * Calculate the page table page (mpte), allocating it if necessary.
2627 * A held page table page (mpte), or NULL, is passed onto the
2628 * section following.
2630 if (va
< VM_MAX_USER_ADDRESS
) {
2632 * Calculate pagetable page index
2634 ptepindex
= pmap_pde_pindex(va
);
2638 * Get the page directory entry
2640 ptepa
= pmap_pde(pmap
, va
);
2643 * If the page table page is mapped, we just increment
2644 * the hold count, and activate it.
2646 if (ptepa
&& (*ptepa
& PG_V
) != 0) {
2648 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2649 // if (pmap->pm_ptphint &&
2650 // (pmap->pm_ptphint->pindex == ptepindex)) {
2651 // mpte = pmap->pm_ptphint;
2653 mpte
= pmap_page_lookup( pmap
->pm_pteobj
, ptepindex
);
2654 pmap
->pm_ptphint
= mpte
;
2659 mpte
= _pmap_allocpte(pmap
, ptepindex
);
2661 } while (mpte
== NULL
);
2664 /* this code path is not yet used */
2668 * With a valid (and held) page directory page, we can just use
2669 * vtopte() to get to the pte. If the pte is already present
2670 * we do not disturb it.
2675 pmap_unwire_pte_hold(pmap
, va
, mpte
, &info
);
2676 pa
= VM_PAGE_TO_PHYS(m
);
2677 KKASSERT(((*pte
^ pa
) & PG_FRAME
) == 0);
2682 * Enter on the PV list if part of our managed memory
2684 if ((m
->flags
& (PG_FICTITIOUS
|PG_UNMANAGED
)) == 0) {
2685 pmap_insert_entry(pmap
, va
, mpte
, m
);
2686 vm_page_flag_set(m
, PG_MAPPED
);
2690 * Increment counters
2692 ++pmap
->pm_stats
.resident_count
;
2694 pa
= VM_PAGE_TO_PHYS(m
);
2697 * Now validate mapping with RO protection
2699 if (m
->flags
& (PG_FICTITIOUS
|PG_UNMANAGED
))
2700 *pte
= pa
| PG_V
| PG_U
;
2702 *pte
= pa
| PG_V
| PG_U
| PG_MANAGED
;
2703 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2704 pmap_inval_flush(&info
);
2708 * Make a temporary mapping for a physical address. This is only intended
2709 * to be used for panic dumps.
2711 /* JG Needed on amd64? */
2713 pmap_kenter_temporary(vm_paddr_t pa
, int i
)
2715 pmap_kenter((vm_offset_t
)crashdumpmap
+ (i
* PAGE_SIZE
), pa
);
2716 return ((void *)crashdumpmap
);
2719 #define MAX_INIT_PT (96)
2722 * This routine preloads the ptes for a given object into the specified pmap.
2723 * This eliminates the blast of soft faults on process startup and
2724 * immediately after an mmap.
2726 static int pmap_object_init_pt_callback(vm_page_t p
, void *data
);
2729 pmap_object_init_pt(pmap_t pmap
, vm_offset_t addr
, vm_prot_t prot
,
2730 vm_object_t object
, vm_pindex_t pindex
,
2731 vm_size_t size
, int limit
)
2733 struct rb_vm_page_scan_info info
;
2738 * We can't preinit if read access isn't set or there is no pmap
2741 if ((prot
& VM_PROT_READ
) == 0 || pmap
== NULL
|| object
== NULL
)
2745 * We can't preinit if the pmap is not the current pmap
2747 lp
= curthread
->td_lwp
;
2748 if (lp
== NULL
|| pmap
!= vmspace_pmap(lp
->lwp_vmspace
))
2751 psize
= amd64_btop(size
);
2753 if ((object
->type
!= OBJT_VNODE
) ||
2754 ((limit
& MAP_PREFAULT_PARTIAL
) && (psize
> MAX_INIT_PT
) &&
2755 (object
->resident_page_count
> MAX_INIT_PT
))) {
2759 if (psize
+ pindex
> object
->size
) {
2760 if (object
->size
< pindex
)
2762 psize
= object
->size
- pindex
;
2769 * Use a red-black scan to traverse the requested range and load
2770 * any valid pages found into the pmap.
2772 * We cannot safely scan the object's memq unless we are in a
2773 * critical section since interrupts can remove pages from objects.
2775 info
.start_pindex
= pindex
;
2776 info
.end_pindex
= pindex
+ psize
- 1;
2783 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, rb_vm_page_scancmp
,
2784 pmap_object_init_pt_callback
, &info
);
2790 pmap_object_init_pt_callback(vm_page_t p
, void *data
)
2792 struct rb_vm_page_scan_info
*info
= data
;
2793 vm_pindex_t rel_index
;
2795 * don't allow an madvise to blow away our really
2796 * free pages allocating pv entries.
2798 if ((info
->limit
& MAP_PREFAULT_MADVISE
) &&
2799 vmstats
.v_free_count
< vmstats
.v_free_reserved
) {
2802 if (((p
->valid
& VM_PAGE_BITS_ALL
) == VM_PAGE_BITS_ALL
) &&
2803 (p
->busy
== 0) && (p
->flags
& (PG_BUSY
| PG_FICTITIOUS
)) == 0) {
2804 if ((p
->queue
- p
->pc
) == PQ_CACHE
)
2805 vm_page_deactivate(p
);
2807 rel_index
= p
->pindex
- info
->start_pindex
;
2808 pmap_enter_quick(info
->pmap
,
2809 info
->addr
+ amd64_ptob(rel_index
), p
);
2816 * pmap_prefault provides a quick way of clustering pagefaults into a
2817 * processes address space. It is a "cousin" of pmap_object_init_pt,
2818 * except it runs at page fault time instead of mmap time.
2822 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2824 static int pmap_prefault_pageorder
[] = {
2825 -PAGE_SIZE
, PAGE_SIZE
,
2826 -2 * PAGE_SIZE
, 2 * PAGE_SIZE
,
2827 -3 * PAGE_SIZE
, 3 * PAGE_SIZE
,
2828 -4 * PAGE_SIZE
, 4 * PAGE_SIZE
2832 pmap_prefault(pmap_t pmap
, vm_offset_t addra
, vm_map_entry_t entry
)
2843 * We do not currently prefault mappings that use virtual page
2844 * tables. We do not prefault foreign pmaps.
2846 if (entry
->maptype
== VM_MAPTYPE_VPAGETABLE
)
2848 lp
= curthread
->td_lwp
;
2849 if (lp
== NULL
|| (pmap
!= vmspace_pmap(lp
->lwp_vmspace
)))
2852 object
= entry
->object
.vm_object
;
2854 starta
= addra
- PFBAK
* PAGE_SIZE
;
2855 if (starta
< entry
->start
)
2856 starta
= entry
->start
;
2857 else if (starta
> addra
)
2861 * critical section protection is required to maintain the
2862 * page/object association, interrupts can free pages and remove
2863 * them from their objects.
2866 for (i
= 0; i
< PAGEORDER_SIZE
; i
++) {
2867 vm_object_t lobject
;
2871 addr
= addra
+ pmap_prefault_pageorder
[i
];
2872 if (addr
> addra
+ (PFFOR
* PAGE_SIZE
))
2875 if (addr
< starta
|| addr
>= entry
->end
)
2878 pde
= pmap_pde(pmap
, addr
);
2879 if (pde
== NULL
|| *pde
== 0)
2886 pindex
= ((addr
- entry
->start
) + entry
->offset
) >> PAGE_SHIFT
;
2889 for (m
= vm_page_lookup(lobject
, pindex
);
2890 (!m
&& (lobject
->type
== OBJT_DEFAULT
) &&
2891 (lobject
->backing_object
));
2892 lobject
= lobject
->backing_object
2894 if (lobject
->backing_object_offset
& PAGE_MASK
)
2896 pindex
+= (lobject
->backing_object_offset
>> PAGE_SHIFT
);
2897 m
= vm_page_lookup(lobject
->backing_object
, pindex
);
2901 * give-up when a page is not in memory
2906 if (((m
->valid
& VM_PAGE_BITS_ALL
) == VM_PAGE_BITS_ALL
) &&
2908 (m
->flags
& (PG_BUSY
| PG_FICTITIOUS
)) == 0) {
2910 if ((m
->queue
- m
->pc
) == PQ_CACHE
) {
2911 vm_page_deactivate(m
);
2914 pmap_enter_quick(pmap
, addr
, m
);
2922 * Routine: pmap_change_wiring
2923 * Function: Change the wiring attribute for a map/virtual-address
2925 * In/out conditions:
2926 * The mapping must already exist in the pmap.
2929 pmap_change_wiring(pmap_t pmap
, vm_offset_t va
, boolean_t wired
)
2936 pte
= pmap_pte(pmap
, va
);
2938 if (wired
&& !pmap_pte_w(pte
))
2939 pmap
->pm_stats
.wired_count
++;
2940 else if (!wired
&& pmap_pte_w(pte
))
2941 pmap
->pm_stats
.wired_count
--;
2944 * Wiring is not a hardware characteristic so there is no need to
2945 * invalidate TLB. However, in an SMP environment we must use
2946 * a locked bus cycle to update the pte (if we are not using
2947 * the pmap_inval_*() API that is)... it's ok to do this for simple
2952 atomic_set_long(pte
, PG_W
);
2954 atomic_clear_long(pte
, PG_W
);
2957 atomic_set_long_nonlocked(pte
, PG_W
);
2959 atomic_clear_long_nonlocked(pte
, PG_W
);
2966 * Copy the range specified by src_addr/len
2967 * from the source map to the range dst_addr/len
2968 * in the destination map.
2970 * This routine is only advisory and need not do anything.
2973 pmap_copy(pmap_t dst_pmap
, pmap_t src_pmap
, vm_offset_t dst_addr
,
2974 vm_size_t len
, vm_offset_t src_addr
)
2978 pmap_inval_info info
;
2980 vm_offset_t end_addr
= src_addr
+ len
;
2982 pd_entry_t src_frame
, dst_frame
;
2985 if (dst_addr
!= src_addr
)
2988 src_frame
= src_pmap
->pm_pdir
[PTDPTDI
] & PG_FRAME
;
2989 if (src_frame
!= (PTDpde
& PG_FRAME
)) {
2993 dst_frame
= dst_pmap
->pm_pdir
[PTDPTDI
] & PG_FRAME
;
2994 if (dst_frame
!= (APTDpde
& PG_FRAME
)) {
2995 APTDpde
= (pd_entry_t
) (dst_frame
| PG_RW
| PG_V
);
2996 /* The page directory is not shared between CPUs */
3000 pmap_inval_init(&info
);
3001 pmap_inval_add(&info
, dst_pmap
, -1);
3002 pmap_inval_add(&info
, src_pmap
, -1);
3005 * critical section protection is required to maintain the page/object
3006 * association, interrupts can free pages and remove them from
3010 for (addr
= src_addr
; addr
< end_addr
; addr
= pdnxt
) {
3011 pt_entry_t
*src_pte
, *dst_pte
;
3012 vm_page_t dstmpte
, srcmpte
;
3013 vm_offset_t srcptepaddr
;
3014 vm_pindex_t ptepindex
;
3016 if (addr
>= UPT_MIN_ADDRESS
)
3017 panic("pmap_copy: invalid to pmap_copy page tables\n");
3020 * Don't let optional prefaulting of pages make us go
3021 * way below the low water mark of free pages or way
3022 * above high water mark of used pv entries.
3024 if (vmstats
.v_free_count
< vmstats
.v_free_reserved
||
3025 pv_entry_count
> pv_entry_high_water
)
3028 pdnxt
= ((addr
+ PAGE_SIZE
*NPTEPG
) & ~(PAGE_SIZE
*NPTEPG
- 1));
3029 ptepindex
= addr
>> PDRSHIFT
;
3032 srcptepaddr
= (vm_offset_t
) src_pmap
->pm_pdir
[ptepindex
];
3034 if (srcptepaddr
== 0)
3037 if (srcptepaddr
& PG_PS
) {
3039 if (dst_pmap
->pm_pdir
[ptepindex
] == 0) {
3040 dst_pmap
->pm_pdir
[ptepindex
] = (pd_entry_t
) srcptepaddr
;
3041 dst_pmap
->pm_stats
.resident_count
+= NBPDR
/ PAGE_SIZE
;
3047 srcmpte
= vm_page_lookup(src_pmap
->pm_pteobj
, ptepindex
);
3048 if ((srcmpte
== NULL
) || (srcmpte
->hold_count
== 0) ||
3049 (srcmpte
->flags
& PG_BUSY
)) {
3053 if (pdnxt
> end_addr
)
3056 src_pte
= vtopte(addr
);
3058 dst_pte
= avtopte(addr
);
3060 while (addr
< pdnxt
) {
3065 * we only virtual copy managed pages
3067 if ((ptetemp
& PG_MANAGED
) != 0) {
3069 * We have to check after allocpte for the
3070 * pte still being around... allocpte can
3073 * pmap_allocpte() can block. If we lose
3074 * our page directory mappings we stop.
3076 dstmpte
= pmap_allocpte(dst_pmap
, addr
);
3079 if (src_frame
!= (PTDpde
& PG_FRAME
) ||
3080 dst_frame
!= (APTDpde
& PG_FRAME
)
3082 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3083 pmap_unwire_pte_hold(dst_pmap
, dstmpte
, &info
);
3085 } else if ((*dst_pte
== 0) &&
3086 (ptetemp
= *src_pte
) != 0 &&
3087 (ptetemp
& PG_MANAGED
)) {
3089 * Clear the modified and
3090 * accessed (referenced) bits
3093 m
= PHYS_TO_VM_PAGE(ptetemp
);
3094 *dst_pte
= ptetemp
& ~(PG_M
| PG_A
);
3095 ++dst_pmap
->pm_stats
.resident_count
;
3096 pmap_insert_entry(dst_pmap
, addr
,
3098 KKASSERT(m
->flags
& PG_MAPPED
);
3100 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3101 pmap_unwire_pte_hold(dst_pmap
, dstmpte
, &info
);
3105 if (dstmpte
->hold_count
>= srcmpte
->hold_count
)
3115 pmap_inval_flush(&info
);
3122 * Zero the specified physical page.
3124 * This function may be called from an interrupt and no locking is
3128 pmap_zero_page(vm_paddr_t phys
)
3130 vm_offset_t va
= PHYS_TO_DMAP(phys
);
3132 pagezero((void *)va
);
3136 * pmap_page_assertzero:
3138 * Assert that a page is empty, panic if it isn't.
3141 pmap_page_assertzero(vm_paddr_t phys
)
3143 vm_offset_t virt
= PHYS_TO_DMAP(phys
);
3146 for (i
= 0; i
< PAGE_SIZE
; i
+= sizeof(long)) {
3147 if (*(long *)((char *)virt
+ i
) != 0) {
3148 panic("pmap_page_assertzero() @ %p not zero!\n", (void *)virt
);
3156 * Zero part of a physical page by mapping it into memory and clearing
3157 * its contents with bzero.
3159 * off and size may not cover an area beyond a single hardware page.
3162 pmap_zero_page_area(vm_paddr_t phys
, int off
, int size
)
3164 vm_offset_t virt
= PHYS_TO_DMAP(phys
);
3166 bzero((char *)virt
+ off
, size
);
3172 * Copy the physical page from the source PA to the target PA.
3173 * This function may be called from an interrupt. No locking
3177 pmap_copy_page(vm_paddr_t src
, vm_paddr_t dst
)
3179 vm_offset_t src_virt
, dst_virt
;
3181 src_virt
= PHYS_TO_DMAP(src
);
3182 dst_virt
= PHYS_TO_DMAP(dst
);
3183 bcopy((void *)src_virt
, (void *)dst_virt
, PAGE_SIZE
);
3187 * pmap_copy_page_frag:
3189 * Copy the physical page from the source PA to the target PA.
3190 * This function may be called from an interrupt. No locking
3194 pmap_copy_page_frag(vm_paddr_t src
, vm_paddr_t dst
, size_t bytes
)
3196 vm_offset_t src_virt
, dst_virt
;
3198 src_virt
= PHYS_TO_DMAP(src
);
3199 dst_virt
= PHYS_TO_DMAP(dst
);
3201 bcopy((char *)src_virt
+ (src
& PAGE_MASK
),
3202 (char *)dst_virt
+ (dst
& PAGE_MASK
),
3207 * Returns true if the pmap's pv is one of the first
3208 * 16 pvs linked to from this page. This count may
3209 * be changed upwards or downwards in the future; it
3210 * is only necessary that true be returned for a small
3211 * subset of pmaps for proper page aging.
3214 pmap_page_exists_quick(pmap_t pmap
, vm_page_t m
)
3219 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
3224 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
3225 if (pv
->pv_pmap
== pmap
) {
3238 * Remove all pages from specified address space
3239 * this aids process exit speeds. Also, this code
3240 * is special cased for current process only, but
3241 * can have the more generic (and slightly slower)
3242 * mode enabled. This is much faster than pmap_remove
3243 * in the case of running down an entire address space.
3246 pmap_remove_pages(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
)
3249 pt_entry_t
*pte
, tpte
;
3252 pmap_inval_info info
;
3254 int save_generation
;
3256 lp
= curthread
->td_lwp
;
3257 if (lp
&& pmap
== vmspace_pmap(lp
->lwp_vmspace
))
3262 pmap_inval_init(&info
);
3264 for (pv
= TAILQ_FIRST(&pmap
->pm_pvlist
); pv
; pv
= npv
) {
3265 if (pv
->pv_va
>= eva
|| pv
->pv_va
< sva
) {
3266 npv
= TAILQ_NEXT(pv
, pv_plist
);
3270 KKASSERT(pmap
== pv
->pv_pmap
);
3273 pte
= vtopte(pv
->pv_va
);
3275 pte
= pmap_pte_quick(pmap
, pv
->pv_va
);
3276 if (pmap
->pm_active
)
3277 pmap_inval_add(&info
, pmap
, pv
->pv_va
);
3280 * We cannot remove wired pages from a process' mapping
3284 npv
= TAILQ_NEXT(pv
, pv_plist
);
3287 tpte
= pte_load_clear(pte
);
3289 m
= PHYS_TO_VM_PAGE(tpte
& PG_FRAME
);
3291 KASSERT(m
< &vm_page_array
[vm_page_array_size
],
3292 ("pmap_remove_pages: bad tpte %lx", tpte
));
3294 KKASSERT(pmap
->pm_stats
.resident_count
> 0);
3295 --pmap
->pm_stats
.resident_count
;
3298 * Update the vm_page_t clean and reference bits.
3304 npv
= TAILQ_NEXT(pv
, pv_plist
);
3305 TAILQ_REMOVE(&pmap
->pm_pvlist
, pv
, pv_plist
);
3306 save_generation
= ++pmap
->pm_generation
;
3308 m
->md
.pv_list_count
--;
3309 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
3310 if (TAILQ_EMPTY(&m
->md
.pv_list
))
3311 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
3313 pmap_unuse_pt(pmap
, pv
->pv_va
, pv
->pv_ptem
, &info
);
3317 * Restart the scan if we blocked during the unuse or free
3318 * calls and other removals were made.
3320 if (save_generation
!= pmap
->pm_generation
) {
3321 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3322 pv
= TAILQ_FIRST(&pmap
->pm_pvlist
);
3325 pmap_inval_flush(&info
);
3330 * pmap_testbit tests bits in pte's
3331 * note that the testbit/clearbit routines are inline,
3332 * and a lot of things compile-time evaluate.
3336 pmap_testbit(vm_page_t m
, int bit
)
3341 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
3344 if (TAILQ_FIRST(&m
->md
.pv_list
) == NULL
)
3349 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
3351 * if the bit being tested is the modified bit, then
3352 * mark clean_map and ptes as never
3355 if (bit
& (PG_A
|PG_M
)) {
3356 if (!pmap_track_modified(pv
->pv_va
))
3360 #if defined(PMAP_DIAGNOSTIC)
3361 if (pv
->pv_pmap
== NULL
) {
3362 kprintf("Null pmap (tb) at va: 0x%lx\n", pv
->pv_va
);
3366 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_va
);
3377 * this routine is used to modify bits in ptes
3381 pmap_clearbit(vm_page_t m
, int bit
)
3383 struct pmap_inval_info info
;
3388 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
3391 pmap_inval_init(&info
);
3395 * Loop over all current mappings setting/clearing as appropos If
3396 * setting RO do we need to clear the VAC?
3398 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
3400 * don't write protect pager mappings
3403 if (!pmap_track_modified(pv
->pv_va
))
3407 #if defined(PMAP_DIAGNOSTIC)
3408 if (pv
->pv_pmap
== NULL
) {
3409 kprintf("Null pmap (cb) at va: 0x%lx\n", pv
->pv_va
);
3415 * Careful here. We can use a locked bus instruction to
3416 * clear PG_A or PG_M safely but we need to synchronize
3417 * with the target cpus when we mess with PG_RW.
3419 * We do not have to force synchronization when clearing
3420 * PG_M even for PTEs generated via virtual memory maps,
3421 * because the virtual kernel will invalidate the pmap
3422 * entry when/if it needs to resynchronize the Modify bit.
3425 pmap_inval_add(&info
, pv
->pv_pmap
, pv
->pv_va
);
3426 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_va
);
3433 atomic_clear_long(pte
, PG_M
|PG_RW
);
3436 * The cpu may be trying to set PG_M
3437 * simultaniously with our clearing
3440 if (!atomic_cmpset_long(pte
, pbits
,
3444 } else if (bit
== PG_M
) {
3446 * We could also clear PG_RW here to force
3447 * a fault on write to redetect PG_M for
3448 * virtual kernels, but it isn't necessary
3449 * since virtual kernels invalidate the pte
3450 * when they clear the VPTE_M bit in their
3451 * virtual page tables.
3453 atomic_clear_long(pte
, PG_M
);
3455 atomic_clear_long(pte
, bit
);
3459 pmap_inval_flush(&info
);
3464 * pmap_page_protect:
3466 * Lower the permission for all mappings to a given page.
3469 pmap_page_protect(vm_page_t m
, vm_prot_t prot
)
3471 /* JG NX support? */
3472 if ((prot
& VM_PROT_WRITE
) == 0) {
3473 if (prot
& (VM_PROT_READ
| VM_PROT_EXECUTE
)) {
3474 pmap_clearbit(m
, PG_RW
);
3475 vm_page_flag_clear(m
, PG_WRITEABLE
);
3483 pmap_phys_address(vm_pindex_t ppn
)
3485 return (amd64_ptob(ppn
));
3489 * pmap_ts_referenced:
3491 * Return a count of reference bits for a page, clearing those bits.
3492 * It is not necessary for every reference bit to be cleared, but it
3493 * is necessary that 0 only be returned when there are truly no
3494 * reference bits set.
3496 * XXX: The exact number of bits to check and clear is a matter that
3497 * should be tested and standardized at some point in the future for
3498 * optimal aging of shared pages.
3501 pmap_ts_referenced(vm_page_t m
)
3503 pv_entry_t pv
, pvf
, pvn
;
3507 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
3512 if ((pv
= TAILQ_FIRST(&m
->md
.pv_list
)) != NULL
) {
3517 pvn
= TAILQ_NEXT(pv
, pv_list
);
3519 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
3521 TAILQ_INSERT_TAIL(&m
->md
.pv_list
, pv
, pv_list
);
3523 if (!pmap_track_modified(pv
->pv_va
))
3526 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_va
);
3528 if (pte
&& (*pte
& PG_A
)) {
3530 atomic_clear_long(pte
, PG_A
);
3532 atomic_clear_long_nonlocked(pte
, PG_A
);
3539 } while ((pv
= pvn
) != NULL
&& pv
!= pvf
);
3549 * Return whether or not the specified physical page was modified
3550 * in any physical maps.
3553 pmap_is_modified(vm_page_t m
)
3555 return pmap_testbit(m
, PG_M
);
3559 * Clear the modify bits on the specified physical page.
3562 pmap_clear_modify(vm_page_t m
)
3564 pmap_clearbit(m
, PG_M
);
3568 * pmap_clear_reference:
3570 * Clear the reference bit on the specified physical page.
3573 pmap_clear_reference(vm_page_t m
)
3575 pmap_clearbit(m
, PG_A
);
3579 * Miscellaneous support routines follow
3584 i386_protection_init(void)
3588 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3589 kp
= protection_codes
;
3590 for (prot
= 0; prot
< 8; prot
++) {
3592 case VM_PROT_NONE
| VM_PROT_NONE
| VM_PROT_NONE
:
3594 * Read access is also 0. There isn't any execute bit,
3595 * so just make it readable.
3597 case VM_PROT_READ
| VM_PROT_NONE
| VM_PROT_NONE
:
3598 case VM_PROT_READ
| VM_PROT_NONE
| VM_PROT_EXECUTE
:
3599 case VM_PROT_NONE
| VM_PROT_NONE
| VM_PROT_EXECUTE
:
3602 case VM_PROT_NONE
| VM_PROT_WRITE
| VM_PROT_NONE
:
3603 case VM_PROT_NONE
| VM_PROT_WRITE
| VM_PROT_EXECUTE
:
3604 case VM_PROT_READ
| VM_PROT_WRITE
| VM_PROT_NONE
:
3605 case VM_PROT_READ
| VM_PROT_WRITE
| VM_PROT_EXECUTE
:
3613 * Map a set of physical memory pages into the kernel virtual
3614 * address space. Return a pointer to where it is mapped. This
3615 * routine is intended to be used for mapping device memory,
3618 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3622 pmap_mapdev(vm_paddr_t pa
, vm_size_t size
)
3624 vm_offset_t va
, tmpva
, offset
;
3627 offset
= pa
& PAGE_MASK
;
3628 size
= roundup(offset
+ size
, PAGE_SIZE
);
3630 va
= kmem_alloc_nofault(&kernel_map
, size
);
3632 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3634 pa
= pa
& ~PAGE_MASK
;
3635 for (tmpva
= va
; size
> 0;) {
3636 pte
= vtopte(tmpva
);
3637 *pte
= pa
| PG_RW
| PG_V
; /* | pgeflag; */
3645 return ((void *)(va
+ offset
));
3649 pmap_mapdev_uncacheable(vm_paddr_t pa
, vm_size_t size
)
3651 vm_offset_t va
, tmpva
, offset
;
3654 offset
= pa
& PAGE_MASK
;
3655 size
= roundup(offset
+ size
, PAGE_SIZE
);
3657 va
= kmem_alloc_nofault(&kernel_map
, size
);
3659 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3661 pa
= pa
& ~PAGE_MASK
;
3662 for (tmpva
= va
; size
> 0;) {
3663 pte
= vtopte(tmpva
);
3664 *pte
= pa
| PG_RW
| PG_V
| PG_N
; /* | pgeflag; */
3672 return ((void *)(va
+ offset
));
3676 pmap_unmapdev(vm_offset_t va
, vm_size_t size
)
3678 vm_offset_t base
, offset
;
3680 base
= va
& ~PAGE_MASK
;
3681 offset
= va
& PAGE_MASK
;
3682 size
= roundup(offset
+ size
, PAGE_SIZE
);
3683 pmap_qremove(va
, size
>> PAGE_SHIFT
);
3684 kmem_free(&kernel_map
, base
, size
);
3688 * perform the pmap work for mincore
3691 pmap_mincore(pmap_t pmap
, vm_offset_t addr
)
3693 pt_entry_t
*ptep
, pte
;
3697 ptep
= pmap_pte(pmap
, addr
);
3702 if ((pte
= *ptep
) != 0) {
3705 val
= MINCORE_INCORE
;
3706 if ((pte
& PG_MANAGED
) == 0)
3709 pa
= pte
& PG_FRAME
;
3711 m
= PHYS_TO_VM_PAGE(pa
);
3717 val
|= MINCORE_MODIFIED
|MINCORE_MODIFIED_OTHER
;
3719 * Modified by someone
3721 else if (m
->dirty
|| pmap_is_modified(m
))
3722 val
|= MINCORE_MODIFIED_OTHER
;
3727 val
|= MINCORE_REFERENCED
|MINCORE_REFERENCED_OTHER
;
3730 * Referenced by someone
3732 else if ((m
->flags
& PG_REFERENCED
) || pmap_ts_referenced(m
)) {
3733 val
|= MINCORE_REFERENCED_OTHER
;
3734 vm_page_flag_set(m
, PG_REFERENCED
);
3741 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3742 * vmspace will be ref'd and the old one will be deref'd.
3744 * The vmspace for all lwps associated with the process will be adjusted
3745 * and cr3 will be reloaded if any lwp is the current lwp.
3748 pmap_replacevm(struct proc
*p
, struct vmspace
*newvm
, int adjrefs
)
3750 struct vmspace
*oldvm
;
3754 oldvm
= p
->p_vmspace
;
3755 if (oldvm
!= newvm
) {
3756 p
->p_vmspace
= newvm
;
3757 KKASSERT(p
->p_nthreads
== 1);
3758 lp
= RB_ROOT(&p
->p_lwp_tree
);
3759 pmap_setlwpvm(lp
, newvm
);
3761 sysref_get(&newvm
->vm_sysref
);
3762 sysref_put(&oldvm
->vm_sysref
);
3769 * Set the vmspace for a LWP. The vmspace is almost universally set the
3770 * same as the process vmspace, but virtual kernels need to swap out contexts
3771 * on a per-lwp basis.
3774 pmap_setlwpvm(struct lwp
*lp
, struct vmspace
*newvm
)
3776 struct vmspace
*oldvm
;
3780 oldvm
= lp
->lwp_vmspace
;
3782 if (oldvm
!= newvm
) {
3783 lp
->lwp_vmspace
= newvm
;
3784 if (curthread
->td_lwp
== lp
) {
3785 pmap
= vmspace_pmap(newvm
);
3787 atomic_set_int(&pmap
->pm_active
, 1 << mycpu
->gd_cpuid
);
3789 pmap
->pm_active
|= 1;
3791 #if defined(SWTCH_OPTIM_STATS)
3794 curthread
->td_pcb
->pcb_cr3
= vtophys(pmap
->pm_pml4
);
3795 load_cr3(curthread
->td_pcb
->pcb_cr3
);
3796 pmap
= vmspace_pmap(oldvm
);
3798 atomic_clear_int(&pmap
->pm_active
,
3799 1 << mycpu
->gd_cpuid
);
3801 pmap
->pm_active
&= ~1;
3809 pmap_addr_hint(vm_object_t obj
, vm_offset_t addr
, vm_size_t size
)
3812 if ((obj
== NULL
) || (size
< NBPDR
) || (obj
->type
!= OBJT_DEVICE
)) {
3816 addr
= (addr
+ (NBPDR
- 1)) & ~(NBPDR
- 1);
3823 static void pads (pmap_t pm
);
3824 void pmap_pvdump (vm_paddr_t pa
);
3826 /* print address space of pmap*/
3835 if (pm
== &kernel_pmap
)
3838 for (i
= 0; i
< NPDEPG
; i
++) {
3846 pmap_pvdump(vm_paddr_t pa
)
3851 kprintf("pa %08llx", (long long)pa
);
3852 m
= PHYS_TO_VM_PAGE(pa
);
3853 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
3855 kprintf(" -> pmap %p, va %x, flags %x",
3856 (void *)pv
->pv_pmap
, pv
->pv_va
, pv
->pv_flags
);
3858 kprintf(" -> pmap %p, va %x", (void *)pv
->pv_pmap
, pv
->pv_va
);