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 pmap_qenter2(vm_offset_t va
, vm_page_t
*m
, int count
, cpumask_t
*mask
)
1062 cpumask_t cmask
= mycpu
->gd_cpumask
;
1064 end_va
= va
+ count
* PAGE_SIZE
;
1066 while (va
< end_va
) {
1071 * Install the new PTE. If the pte changed from the prior
1072 * mapping we must reset the cpu mask and invalidate the page.
1073 * If the pte is the same but we have not seen it on the
1074 * current cpu, invlpg the existing mapping. Otherwise the
1075 * entry is optimal and no invalidation is required.
1078 pteval
= VM_PAGE_TO_PHYS(*m
) | PG_A
| PG_RW
| PG_V
| pgeflag
;
1079 if (*pte
!= pteval
) {
1082 cpu_invlpg((void *)va
);
1083 } else if ((*mask
& cmask
) == 0) {
1084 cpu_invlpg((void *)va
);
1093 * This routine jerks page mappings from the
1094 * kernel -- it is meant only for temporary mappings.
1096 * MPSAFE, INTERRUPT SAFE (cluster callback)
1099 pmap_qremove(vm_offset_t va
, int count
)
1103 end_va
= va
+ count
* PAGE_SIZE
;
1105 while (va
< end_va
) {
1110 cpu_invlpg((void *)va
);
1119 * This routine works like vm_page_lookup() but also blocks as long as the
1120 * page is busy. This routine does not busy the page it returns.
1122 * Unless the caller is managing objects whos pages are in a known state,
1123 * the call should be made with a critical section held so the page's object
1124 * association remains valid on return.
1128 pmap_page_lookup(vm_object_t object
, vm_pindex_t pindex
)
1133 m
= vm_page_lookup(object
, pindex
);
1134 } while (m
&& vm_page_sleep_busy(m
, FALSE
, "pplookp"));
1140 * Create a new thread and optionally associate it with a (new) process.
1141 * NOTE! the new thread's cpu may not equal the current cpu.
1144 pmap_init_thread(thread_t td
)
1146 /* enforce pcb placement */
1147 td
->td_pcb
= (struct pcb
*)(td
->td_kstack
+ td
->td_kstack_size
) - 1;
1148 td
->td_savefpu
= &td
->td_pcb
->pcb_save
;
1149 td
->td_sp
= (char *)td
->td_pcb
- 16; /* JG is -16 needed on x86_64? */
1153 * This routine directly affects the fork perf for a process.
1156 pmap_init_proc(struct proc
*p
)
1161 * Dispose the UPAGES for a process that has exited.
1162 * This routine directly impacts the exit perf of a process.
1165 pmap_dispose_proc(struct proc
*p
)
1167 KASSERT(p
->p_lock
== 0, ("attempt to dispose referenced proc! %p", p
));
1170 /***************************************************
1171 * Page table page management routines.....
1172 ***************************************************/
1175 * This routine unholds page table pages, and if the hold count
1176 * drops to zero, then it decrements the wire count.
1180 pmap_unwire_pte_hold(pmap_t pmap
, vm_offset_t va
, vm_page_t m
,
1181 pmap_inval_info_t info
)
1183 KKASSERT(m
->hold_count
> 0);
1184 if (m
->hold_count
> 1) {
1188 return _pmap_unwire_pte_hold(pmap
, va
, m
, info
);
1194 _pmap_unwire_pte_hold(pmap_t pmap
, vm_offset_t va
, vm_page_t m
,
1195 pmap_inval_info_t info
)
1198 * Wait until we can busy the page ourselves. We cannot have
1199 * any active flushes if we block. We own one hold count on the
1200 * page so it cannot be freed out from under us.
1202 if (m
->flags
& PG_BUSY
) {
1203 pmap_inval_flush(info
);
1204 while (vm_page_sleep_busy(m
, FALSE
, "pmuwpt"))
1207 KASSERT(m
->queue
== PQ_NONE
,
1208 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m
));
1211 * This case can occur if new references were acquired while
1214 if (m
->hold_count
> 1) {
1215 KKASSERT(m
->hold_count
> 1);
1221 * Unmap the page table page
1223 KKASSERT(m
->hold_count
== 1);
1225 pmap_inval_interlock(info
, pmap
, -1);
1227 if (m
->pindex
>= (NUPDE
+ NUPDPE
)) {
1230 pml4
= pmap_pml4e(pmap
, va
);
1232 } else if (m
->pindex
>= NUPDE
) {
1235 pdp
= pmap_pdpe(pmap
, va
);
1240 pd
= pmap_pde(pmap
, va
);
1244 KKASSERT(pmap
->pm_stats
.resident_count
> 0);
1245 --pmap
->pm_stats
.resident_count
;
1247 if (pmap
->pm_ptphint
== m
)
1248 pmap
->pm_ptphint
= NULL
;
1249 pmap_inval_deinterlock(info
, pmap
);
1251 if (m
->pindex
< NUPDE
) {
1252 /* We just released a PT, unhold the matching PD */
1255 pdpg
= PHYS_TO_VM_PAGE(*pmap_pdpe(pmap
, va
) & PG_FRAME
);
1256 pmap_unwire_pte_hold(pmap
, va
, pdpg
, info
);
1258 if (m
->pindex
>= NUPDE
&& m
->pindex
< (NUPDE
+ NUPDPE
)) {
1259 /* We just released a PD, unhold the matching PDP */
1262 pdppg
= PHYS_TO_VM_PAGE(*pmap_pml4e(pmap
, va
) & PG_FRAME
);
1263 pmap_unwire_pte_hold(pmap
, va
, pdppg
, info
);
1267 * This was our last hold, the page had better be unwired
1268 * after we decrement wire_count.
1270 * FUTURE NOTE: shared page directory page could result in
1271 * multiple wire counts.
1275 KKASSERT(m
->wire_count
== 0);
1276 --vmstats
.v_wire_count
;
1277 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
1279 vm_page_free_zero(m
);
1285 * After removing a page table entry, this routine is used to
1286 * conditionally free the page, and manage the hold/wire counts.
1290 pmap_unuse_pt(pmap_t pmap
, vm_offset_t va
, vm_page_t mpte
,
1291 pmap_inval_info_t info
)
1293 vm_pindex_t ptepindex
;
1295 if (va
>= VM_MAX_USER_ADDRESS
)
1299 ptepindex
= pmap_pde_pindex(va
);
1301 if (pmap
->pm_ptphint
&&
1302 (pmap
->pm_ptphint
->pindex
== ptepindex
)) {
1303 mpte
= pmap
->pm_ptphint
;
1306 pmap_inval_flush(info
);
1307 mpte
= pmap_page_lookup(pmap
->pm_pteobj
, ptepindex
);
1308 pmap
->pm_ptphint
= mpte
;
1313 return pmap_unwire_pte_hold(pmap
, va
, mpte
, info
);
1317 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1318 * it, and IdlePTD, represents the template used to update all other pmaps.
1320 * On architectures where the kernel pmap is not integrated into the user
1321 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1322 * kernel_pmap should be used to directly access the kernel_pmap.
1325 pmap_pinit0(struct pmap
*pmap
)
1327 pmap
->pm_pml4
= (pml4_entry_t
*)(PTOV_OFFSET
+ KPML4phys
);
1329 pmap
->pm_active
= 0;
1330 pmap
->pm_ptphint
= NULL
;
1331 TAILQ_INIT(&pmap
->pm_pvlist
);
1332 bzero(&pmap
->pm_stats
, sizeof pmap
->pm_stats
);
1336 * Initialize a preallocated and zeroed pmap structure,
1337 * such as one in a vmspace structure.
1340 pmap_pinit(struct pmap
*pmap
)
1345 * No need to allocate page table space yet but we do need a valid
1346 * page directory table.
1348 if (pmap
->pm_pml4
== NULL
) {
1350 (pml4_entry_t
*)kmem_alloc_pageable(&kernel_map
, PAGE_SIZE
);
1354 * Allocate an object for the ptes
1356 if (pmap
->pm_pteobj
== NULL
)
1357 pmap
->pm_pteobj
= vm_object_allocate(OBJT_DEFAULT
, NUPDE
+ NUPDPE
+ PML4PML4I
+ 1);
1360 * Allocate the page directory page, unless we already have
1361 * one cached. If we used the cached page the wire_count will
1362 * already be set appropriately.
1364 if ((ptdpg
= pmap
->pm_pdirm
) == NULL
) {
1365 ptdpg
= vm_page_grab(pmap
->pm_pteobj
, NUPDE
+ NUPDPE
+ PML4PML4I
,
1366 VM_ALLOC_NORMAL
| VM_ALLOC_RETRY
);
1367 pmap
->pm_pdirm
= ptdpg
;
1368 vm_page_flag_clear(ptdpg
, PG_MAPPED
| PG_BUSY
);
1369 ptdpg
->valid
= VM_PAGE_BITS_ALL
;
1370 if (ptdpg
->wire_count
== 0)
1371 ++vmstats
.v_wire_count
;
1372 ptdpg
->wire_count
= 1;
1373 pmap_kenter((vm_offset_t
)pmap
->pm_pml4
, VM_PAGE_TO_PHYS(ptdpg
));
1375 if ((ptdpg
->flags
& PG_ZERO
) == 0)
1376 bzero(pmap
->pm_pml4
, PAGE_SIZE
);
1378 pmap
->pm_pml4
[KPML4I
] = KPDPphys
| PG_RW
| PG_V
| PG_U
;
1379 pmap
->pm_pml4
[DMPML4I
] = DMPDPphys
| PG_RW
| PG_V
| PG_U
;
1381 /* install self-referential address mapping entry */
1382 pmap
->pm_pml4
[PML4PML4I
] = VM_PAGE_TO_PHYS(ptdpg
) | PG_V
| PG_RW
| PG_A
| PG_M
;
1385 pmap
->pm_active
= 0;
1386 pmap
->pm_ptphint
= NULL
;
1387 TAILQ_INIT(&pmap
->pm_pvlist
);
1388 bzero(&pmap
->pm_stats
, sizeof pmap
->pm_stats
);
1389 pmap
->pm_stats
.resident_count
= 1;
1393 * Clean up a pmap structure so it can be physically freed. This routine
1394 * is called by the vmspace dtor function. A great deal of pmap data is
1395 * left passively mapped to improve vmspace management so we have a bit
1396 * of cleanup work to do here.
1399 pmap_puninit(pmap_t pmap
)
1403 KKASSERT(pmap
->pm_active
== 0);
1404 if ((p
= pmap
->pm_pdirm
) != NULL
) {
1405 KKASSERT(pmap
->pm_pml4
!= NULL
);
1406 KKASSERT(pmap
->pm_pml4
!= (void *)(PTOV_OFFSET
+ KPML4phys
));
1407 pmap_kremove((vm_offset_t
)pmap
->pm_pml4
);
1409 vmstats
.v_wire_count
--;
1410 KKASSERT((p
->flags
& PG_BUSY
) == 0);
1412 vm_page_free_zero(p
);
1413 pmap
->pm_pdirm
= NULL
;
1415 if (pmap
->pm_pml4
) {
1416 KKASSERT(pmap
->pm_pml4
!= (void *)(PTOV_OFFSET
+ KPML4phys
));
1417 kmem_free(&kernel_map
, (vm_offset_t
)pmap
->pm_pml4
, PAGE_SIZE
);
1418 pmap
->pm_pml4
= NULL
;
1420 if (pmap
->pm_pteobj
) {
1421 vm_object_deallocate(pmap
->pm_pteobj
);
1422 pmap
->pm_pteobj
= NULL
;
1427 * Wire in kernel global address entries. To avoid a race condition
1428 * between pmap initialization and pmap_growkernel, this procedure
1429 * adds the pmap to the master list (which growkernel scans to update),
1430 * then copies the template.
1433 pmap_pinit2(struct pmap
*pmap
)
1436 TAILQ_INSERT_TAIL(&pmap_list
, pmap
, pm_pmnode
);
1437 /* XXX copies current process, does not fill in MPPTDI */
1442 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1443 * 0 on failure (if the procedure had to sleep).
1445 * When asked to remove the page directory page itself, we actually just
1446 * leave it cached so we do not have to incur the SMP inval overhead of
1447 * removing the kernel mapping. pmap_puninit() will take care of it.
1451 pmap_release_free_page(struct pmap
*pmap
, vm_page_t p
)
1454 * This code optimizes the case of freeing non-busy
1455 * page-table pages. Those pages are zero now, and
1456 * might as well be placed directly into the zero queue.
1458 if (vm_page_sleep_busy(p
, FALSE
, "pmaprl"))
1464 * Remove the page table page from the processes address space.
1466 if (p
->pindex
== NUPDE
+ NUPDPE
+ PML4PML4I
) {
1468 * We are the pml4 table itself.
1470 /* XXX anything to do here? */
1471 } else if (p
->pindex
>= (NUPDE
+ NUPDPE
)) {
1473 * Remove a PDP page from the PML4. We do not maintain
1474 * hold counts on the PML4 page.
1480 m4
= vm_page_lookup(pmap
->pm_pteobj
, NUPDE
+ NUPDPE
+ PML4PML4I
);
1481 KKASSERT(m4
!= NULL
);
1482 pml4
= (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4
));
1483 idx
= (p
->pindex
- (NUPDE
+ NUPDPE
)) % NPML4EPG
;
1484 KKASSERT(pml4
[idx
] != 0);
1486 } else if (p
->pindex
>= NUPDE
) {
1488 * Remove a PD page from the PDP and drop the hold count
1489 * on the PDP. The PDP is left cached in the pmap if
1490 * the hold count drops to 0 so the wire count remains
1497 m3
= vm_page_lookup(pmap
->pm_pteobj
,
1498 NUPDE
+ NUPDPE
+ (p
->pindex
- NUPDE
) / NPDPEPG
);
1499 KKASSERT(m3
!= NULL
);
1500 pdp
= (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3
));
1501 idx
= (p
->pindex
- NUPDE
) % NPDPEPG
;
1502 KKASSERT(pdp
[idx
] != 0);
1507 * Remove a PT page from the PD and drop the hold count
1508 * on the PD. The PD is left cached in the pmap if
1509 * the hold count drops to 0 so the wire count remains
1516 m2
= vm_page_lookup(pmap
->pm_pteobj
,
1517 NUPDE
+ p
->pindex
/ NPDEPG
);
1518 KKASSERT(m2
!= NULL
);
1519 pd
= (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2
));
1520 idx
= p
->pindex
% NPDEPG
;
1526 * One fewer mappings in the pmap. p's hold count had better
1529 KKASSERT(pmap
->pm_stats
.resident_count
> 0);
1530 --pmap
->pm_stats
.resident_count
;
1532 panic("pmap_release: freeing held page table page");
1533 if (pmap
->pm_ptphint
&& (pmap
->pm_ptphint
->pindex
== p
->pindex
))
1534 pmap
->pm_ptphint
= NULL
;
1537 * We leave the top-level page table page cached, wired, and mapped in
1538 * the pmap until the dtor function (pmap_puninit()) gets called.
1539 * However, still clean it up so we can set PG_ZERO.
1541 if (p
->pindex
== NUPDE
+ NUPDPE
+ PML4PML4I
) {
1542 bzero(pmap
->pm_pml4
, PAGE_SIZE
);
1543 vm_page_flag_set(p
, PG_ZERO
);
1547 KKASSERT(p
->wire_count
== 0);
1548 vmstats
.v_wire_count
--;
1549 /* JG eventually revert to using vm_page_free_zero() */
1556 * This routine is called when various levels in the page table need to
1557 * be populated. This routine cannot fail.
1561 _pmap_allocpte(pmap_t pmap
, vm_pindex_t ptepindex
)
1566 * Find or fabricate a new pagetable page. This will busy the page.
1568 m
= vm_page_grab(pmap
->pm_pteobj
, ptepindex
,
1569 VM_ALLOC_NORMAL
| VM_ALLOC_ZERO
| VM_ALLOC_RETRY
);
1570 if ((m
->flags
& PG_ZERO
) == 0) {
1571 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
1574 KASSERT(m
->queue
== PQ_NONE
,
1575 ("_pmap_allocpte: %p->queue != PQ_NONE", m
));
1578 * Increment the hold count for the page we will be returning to
1582 if (m
->wire_count
++ == 0)
1583 vmstats
.v_wire_count
++;
1586 * Map the pagetable page into the process address space, if
1587 * it isn't already there.
1589 * It is possible that someone else got in and mapped the page
1590 * directory page while we were blocked, if so just unbusy and
1591 * return the held page.
1593 if (ptepindex
>= (NUPDE
+ NUPDPE
)) {
1595 * Wire up a new PDP page in the PML4
1597 vm_pindex_t pml4index
;
1600 pml4index
= ptepindex
- (NUPDE
+ NUPDPE
);
1601 pml4
= &pmap
->pm_pml4
[pml4index
];
1603 if (--m
->wire_count
== 0)
1604 --vmstats
.v_wire_count
;
1608 *pml4
= VM_PAGE_TO_PHYS(m
) | PG_U
| PG_RW
| PG_V
| PG_A
| PG_M
;
1609 } else if (ptepindex
>= NUPDE
) {
1611 * Wire up a new PD page in the PDP
1613 vm_pindex_t pml4index
;
1614 vm_pindex_t pdpindex
;
1619 pdpindex
= ptepindex
- NUPDE
;
1620 pml4index
= pdpindex
>> NPML4EPGSHIFT
;
1622 pml4
= &pmap
->pm_pml4
[pml4index
];
1623 if ((*pml4
& PG_V
) == 0) {
1625 * Have to allocate a new PDP page, recurse.
1626 * This always succeeds. Returned page will
1629 pdppg
= _pmap_allocpte(pmap
,
1630 NUPDE
+ NUPDPE
+ pml4index
);
1633 * Add a held reference to the PDP page.
1635 pdppg
= PHYS_TO_VM_PAGE(*pml4
& PG_FRAME
);
1636 pdppg
->hold_count
++;
1640 * Now find the pdp_entry and map the PDP. If the PDP
1641 * has already been mapped unwind and return the
1642 * already-mapped PDP held.
1644 * pdppg is left held (hold_count is incremented for
1645 * each PD in the PDP).
1647 pdp
= (pdp_entry_t
*)PHYS_TO_DMAP(*pml4
& PG_FRAME
);
1648 pdp
= &pdp
[pdpindex
& ((1ul << NPDPEPGSHIFT
) - 1)];
1650 vm_page_unhold(pdppg
);
1651 if (--m
->wire_count
== 0)
1652 --vmstats
.v_wire_count
;
1656 *pdp
= VM_PAGE_TO_PHYS(m
) | PG_U
| PG_RW
| PG_V
| PG_A
| PG_M
;
1659 * Wire up the new PT page in the PD
1661 vm_pindex_t pml4index
;
1662 vm_pindex_t pdpindex
;
1668 pdpindex
= ptepindex
>> NPDPEPGSHIFT
;
1669 pml4index
= pdpindex
>> NPML4EPGSHIFT
;
1672 * Locate the PDP page in the PML4, then the PD page in
1673 * the PDP. If either does not exist we simply recurse
1676 * We can just recurse on the PD page as it will recurse
1677 * on the PDP if necessary.
1679 pml4
= &pmap
->pm_pml4
[pml4index
];
1680 if ((*pml4
& PG_V
) == 0) {
1681 pdpg
= _pmap_allocpte(pmap
, NUPDE
+ pdpindex
);
1682 pdp
= (pdp_entry_t
*)PHYS_TO_DMAP(*pml4
& PG_FRAME
);
1683 pdp
= &pdp
[pdpindex
& ((1ul << NPDPEPGSHIFT
) - 1)];
1685 pdp
= (pdp_entry_t
*)PHYS_TO_DMAP(*pml4
& PG_FRAME
);
1686 pdp
= &pdp
[pdpindex
& ((1ul << NPDPEPGSHIFT
) - 1)];
1687 if ((*pdp
& PG_V
) == 0) {
1688 pdpg
= _pmap_allocpte(pmap
, NUPDE
+ pdpindex
);
1690 pdpg
= PHYS_TO_VM_PAGE(*pdp
& PG_FRAME
);
1696 * Now fill in the pte in the PD. If the pte already exists
1697 * (again, if we raced the grab), unhold pdpg and unwire
1698 * m, returning a held m.
1700 * pdpg is left held (hold_count is incremented for
1701 * each PT in the PD).
1703 pd
= (pd_entry_t
*)PHYS_TO_DMAP(*pdp
& PG_FRAME
);
1704 pd
= &pd
[ptepindex
& ((1ul << NPDEPGSHIFT
) - 1)];
1706 vm_page_unhold(pdpg
);
1707 if (--m
->wire_count
== 0)
1708 --vmstats
.v_wire_count
;
1712 *pd
= VM_PAGE_TO_PHYS(m
) | PG_U
| PG_RW
| PG_V
| PG_A
| PG_M
;
1716 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1717 * valid bits, mapped flag, unbusy, and we're done.
1719 pmap
->pm_ptphint
= m
;
1720 ++pmap
->pm_stats
.resident_count
;
1722 m
->valid
= VM_PAGE_BITS_ALL
;
1723 vm_page_flag_clear(m
, PG_ZERO
);
1724 vm_page_flag_set(m
, PG_MAPPED
);
1732 pmap_allocpte(pmap_t pmap
, vm_offset_t va
)
1734 vm_pindex_t ptepindex
;
1739 * Calculate pagetable page index
1741 ptepindex
= pmap_pde_pindex(va
);
1744 * Get the page directory entry
1746 pd
= pmap_pde(pmap
, va
);
1749 * This supports switching from a 2MB page to a
1752 if (pd
!= NULL
&& (*pd
& (PG_PS
| PG_V
)) == (PG_PS
| PG_V
)) {
1753 panic("no promotion/demotion yet");
1761 * If the page table page is mapped, we just increment the
1762 * hold count, and activate it.
1764 if (pd
!= NULL
&& (*pd
& PG_V
) != 0) {
1765 /* YYY hint is used here on i386 */
1766 m
= pmap_page_lookup( pmap
->pm_pteobj
, ptepindex
);
1767 pmap
->pm_ptphint
= m
;
1772 * Here if the pte page isn't mapped, or if it has been deallocated.
1774 return _pmap_allocpte(pmap
, ptepindex
);
1778 /***************************************************
1779 * Pmap allocation/deallocation routines.
1780 ***************************************************/
1783 * Release any resources held by the given physical map.
1784 * Called when a pmap initialized by pmap_pinit is being released.
1785 * Should only be called if the map contains no valid mappings.
1787 static int pmap_release_callback(struct vm_page
*p
, void *data
);
1790 pmap_release(struct pmap
*pmap
)
1792 vm_object_t object
= pmap
->pm_pteobj
;
1793 struct rb_vm_page_scan_info info
;
1795 KASSERT(pmap
->pm_active
== 0, ("pmap still active! %08x", pmap
->pm_active
));
1796 #if defined(DIAGNOSTIC)
1797 if (object
->ref_count
!= 1)
1798 panic("pmap_release: pteobj reference count != 1");
1802 info
.object
= object
;
1804 TAILQ_REMOVE(&pmap_list
, pmap
, pm_pmnode
);
1811 info
.limit
= object
->generation
;
1813 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, NULL
,
1814 pmap_release_callback
, &info
);
1815 if (info
.error
== 0 && info
.mpte
) {
1816 if (!pmap_release_free_page(pmap
, info
.mpte
))
1820 } while (info
.error
);
1825 pmap_release_callback(struct vm_page
*p
, void *data
)
1827 struct rb_vm_page_scan_info
*info
= data
;
1829 if (p
->pindex
== NUPDE
+ NUPDPE
+ PML4PML4I
) {
1833 if (!pmap_release_free_page(info
->pmap
, p
)) {
1837 if (info
->object
->generation
!= info
->limit
) {
1845 * Grow the number of kernel page table entries, if needed.
1848 pmap_growkernel(vm_offset_t addr
)
1851 vm_offset_t ptppaddr
;
1853 pd_entry_t
*pde
, newpdir
;
1857 if (kernel_vm_end
== 0) {
1858 kernel_vm_end
= VM_MIN_KERNEL_ADDRESS
;
1860 while ((*pmap_pde(&kernel_pmap
, kernel_vm_end
) & PG_V
) != 0) {
1861 kernel_vm_end
= (kernel_vm_end
+ PAGE_SIZE
* NPTEPG
) & ~(PAGE_SIZE
* NPTEPG
- 1);
1863 if (kernel_vm_end
- 1 >= kernel_map
.max_offset
) {
1864 kernel_vm_end
= kernel_map
.max_offset
;
1869 addr
= roundup2(addr
, PAGE_SIZE
* NPTEPG
);
1870 if (addr
- 1 >= kernel_map
.max_offset
)
1871 addr
= kernel_map
.max_offset
;
1872 while (kernel_vm_end
< addr
) {
1873 pde
= pmap_pde(&kernel_pmap
, kernel_vm_end
);
1875 /* We need a new PDP entry */
1876 nkpg
= vm_page_alloc(kptobj
, nkpt
,
1877 VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
1878 | VM_ALLOC_INTERRUPT
);
1880 panic("pmap_growkernel: no memory to grow kernel");
1881 paddr
= VM_PAGE_TO_PHYS(nkpg
);
1882 if ((nkpg
->flags
& PG_ZERO
) == 0)
1883 pmap_zero_page(paddr
);
1884 vm_page_flag_clear(nkpg
, PG_ZERO
);
1885 newpdp
= (pdp_entry_t
)
1886 (paddr
| PG_V
| PG_RW
| PG_A
| PG_M
);
1887 *pmap_pdpe(&kernel_pmap
, kernel_vm_end
) = newpdp
;
1889 continue; /* try again */
1891 if ((*pde
& PG_V
) != 0) {
1892 kernel_vm_end
= (kernel_vm_end
+ PAGE_SIZE
* NPTEPG
) & ~(PAGE_SIZE
* NPTEPG
- 1);
1893 if (kernel_vm_end
- 1 >= kernel_map
.max_offset
) {
1894 kernel_vm_end
= kernel_map
.max_offset
;
1901 * This index is bogus, but out of the way
1903 nkpg
= vm_page_alloc(kptobj
, nkpt
,
1904 VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
| VM_ALLOC_INTERRUPT
);
1906 panic("pmap_growkernel: no memory to grow kernel");
1909 ptppaddr
= VM_PAGE_TO_PHYS(nkpg
);
1910 pmap_zero_page(ptppaddr
);
1911 vm_page_flag_clear(nkpg
, PG_ZERO
);
1912 newpdir
= (pd_entry_t
) (ptppaddr
| PG_V
| PG_RW
| PG_A
| PG_M
);
1913 *pmap_pde(&kernel_pmap
, kernel_vm_end
) = newpdir
;
1916 kernel_vm_end
= (kernel_vm_end
+ PAGE_SIZE
* NPTEPG
) & ~(PAGE_SIZE
* NPTEPG
- 1);
1917 if (kernel_vm_end
- 1 >= kernel_map
.max_offset
) {
1918 kernel_vm_end
= kernel_map
.max_offset
;
1926 * Retire the given physical map from service.
1927 * Should only be called if the map contains
1928 * no valid mappings.
1931 pmap_destroy(pmap_t pmap
)
1938 count
= --pmap
->pm_count
;
1941 panic("destroying a pmap is not yet implemented");
1946 * Add a reference to the specified pmap.
1949 pmap_reference(pmap_t pmap
)
1956 /***************************************************
1957 * page management routines.
1958 ***************************************************/
1961 * free the pv_entry back to the free list. This function may be
1962 * called from an interrupt.
1966 free_pv_entry(pv_entry_t pv
)
1969 KKASSERT(pv_entry_count
>= 0);
1974 * get a new pv_entry, allocating a block from the system
1975 * when needed. This function may be called from an interrupt.
1982 if (pv_entry_high_water
&&
1983 (pv_entry_count
> pv_entry_high_water
) &&
1984 (pmap_pagedaemon_waken
== 0)) {
1985 pmap_pagedaemon_waken
= 1;
1986 wakeup(&vm_pages_needed
);
1988 return zalloc(pvzone
);
1992 * This routine is very drastic, but can save the system
2000 static int warningdone
=0;
2002 if (pmap_pagedaemon_waken
== 0)
2005 if (warningdone
< 5) {
2006 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
2010 for(i
= 0; i
< vm_page_array_size
; i
++) {
2011 m
= &vm_page_array
[i
];
2012 if (m
->wire_count
|| m
->hold_count
|| m
->busy
||
2013 (m
->flags
& PG_BUSY
))
2017 pmap_pagedaemon_waken
= 0;
2022 * If it is the first entry on the list, it is actually
2023 * in the header and we must copy the following entry up
2024 * to the header. Otherwise we must search the list for
2025 * the entry. In either case we free the now unused entry.
2029 pmap_remove_entry(struct pmap
*pmap
, vm_page_t m
,
2030 vm_offset_t va
, pmap_inval_info_t info
)
2036 if (m
->md
.pv_list_count
< pmap
->pm_stats
.resident_count
) {
2037 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
2038 if (pmap
== pv
->pv_pmap
&& va
== pv
->pv_va
)
2042 TAILQ_FOREACH(pv
, &pmap
->pm_pvlist
, pv_plist
) {
2043 if (va
== pv
->pv_va
)
2051 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
2052 m
->md
.pv_list_count
--;
2053 KKASSERT(m
->md
.pv_list_count
>= 0);
2054 if (TAILQ_EMPTY(&m
->md
.pv_list
))
2055 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
2056 TAILQ_REMOVE(&pmap
->pm_pvlist
, pv
, pv_plist
);
2057 ++pmap
->pm_generation
;
2058 rtval
= pmap_unuse_pt(pmap
, va
, pv
->pv_ptem
, info
);
2066 * Create a pv entry for page at pa for
2071 pmap_insert_entry(pmap_t pmap
, vm_offset_t va
, vm_page_t mpte
, vm_page_t m
)
2076 pv
= get_pv_entry();
2081 TAILQ_INSERT_TAIL(&pmap
->pm_pvlist
, pv
, pv_plist
);
2082 TAILQ_INSERT_TAIL(&m
->md
.pv_list
, pv
, pv_list
);
2083 ++pmap
->pm_generation
;
2084 m
->md
.pv_list_count
++;
2090 * pmap_remove_pte: do the things to unmap a page in a process
2094 pmap_remove_pte(struct pmap
*pmap
, pt_entry_t
*ptq
, vm_offset_t va
,
2095 pmap_inval_info_t info
)
2100 pmap_inval_interlock(info
, pmap
, va
);
2101 oldpte
= pte_load_clear(ptq
);
2102 pmap_inval_deinterlock(info
, pmap
);
2104 pmap
->pm_stats
.wired_count
-= 1;
2106 * Machines that don't support invlpg, also don't support
2107 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2111 cpu_invlpg((void *)va
);
2112 KKASSERT(pmap
->pm_stats
.resident_count
> 0);
2113 --pmap
->pm_stats
.resident_count
;
2114 if (oldpte
& PG_MANAGED
) {
2115 m
= PHYS_TO_VM_PAGE(oldpte
);
2116 if (oldpte
& PG_M
) {
2117 #if defined(PMAP_DIAGNOSTIC)
2118 if (pmap_nw_modified((pt_entry_t
) oldpte
)) {
2120 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2124 if (pmap_track_modified(va
))
2128 vm_page_flag_set(m
, PG_REFERENCED
);
2129 return pmap_remove_entry(pmap
, m
, va
, info
);
2131 return pmap_unuse_pt(pmap
, va
, NULL
, info
);
2140 * Remove a single page from a process address space.
2142 * This function may not be called from an interrupt if the pmap is
2147 pmap_remove_page(struct pmap
*pmap
, vm_offset_t va
, pmap_inval_info_t info
)
2151 pte
= pmap_pte(pmap
, va
);
2154 if ((*pte
& PG_V
) == 0)
2156 pmap_remove_pte(pmap
, pte
, va
, info
);
2162 * Remove the given range of addresses from the specified map.
2164 * It is assumed that the start and end are properly
2165 * rounded to the page size.
2167 * This function may not be called from an interrupt if the pmap is
2171 pmap_remove(struct pmap
*pmap
, vm_offset_t sva
, vm_offset_t eva
)
2173 vm_offset_t va_next
;
2174 pml4_entry_t
*pml4e
;
2176 pd_entry_t ptpaddr
, *pde
;
2178 struct pmap_inval_info info
;
2183 if (pmap
->pm_stats
.resident_count
== 0)
2186 pmap_inval_init(&info
);
2189 * special handling of removing one page. a very
2190 * common operation and easy to short circuit some
2193 if (sva
+ PAGE_SIZE
== eva
) {
2194 pde
= pmap_pde(pmap
, sva
);
2195 if (pde
&& (*pde
& PG_PS
) == 0) {
2196 pmap_remove_page(pmap
, sva
, &info
);
2197 pmap_inval_done(&info
);
2202 for (; sva
< eva
; sva
= va_next
) {
2203 pml4e
= pmap_pml4e(pmap
, sva
);
2204 if ((*pml4e
& PG_V
) == 0) {
2205 va_next
= (sva
+ NBPML4
) & ~PML4MASK
;
2211 pdpe
= pmap_pml4e_to_pdpe(pml4e
, sva
);
2212 if ((*pdpe
& PG_V
) == 0) {
2213 va_next
= (sva
+ NBPDP
) & ~PDPMASK
;
2220 * Calculate index for next page table.
2222 va_next
= (sva
+ NBPDR
) & ~PDRMASK
;
2226 pde
= pmap_pdpe_to_pde(pdpe
, sva
);
2230 * Weed out invalid mappings.
2236 * Check for large page.
2238 if ((ptpaddr
& PG_PS
) != 0) {
2239 /* JG FreeBSD has more complex treatment here */
2240 pmap_inval_interlock(&info
, pmap
, -1);
2242 pmap
->pm_stats
.resident_count
-= NBPDR
/ PAGE_SIZE
;
2243 pmap_inval_deinterlock(&info
, pmap
);
2248 * Limit our scan to either the end of the va represented
2249 * by the current page table page, or to the end of the
2250 * range being removed.
2256 * NOTE: pmap_remove_pte() can block.
2258 for (pte
= pmap_pde_to_pte(pde
, sva
); sva
!= va_next
; pte
++,
2262 if (pmap_remove_pte(pmap
, pte
, sva
, &info
))
2266 pmap_inval_done(&info
);
2272 * Removes this physical page from all physical maps in which it resides.
2273 * Reflects back modify bits to the pager.
2275 * This routine may not be called from an interrupt.
2280 pmap_remove_all(vm_page_t m
)
2282 struct pmap_inval_info info
;
2283 pt_entry_t
*pte
, tpte
;
2286 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
2289 pmap_inval_init(&info
);
2291 while ((pv
= TAILQ_FIRST(&m
->md
.pv_list
)) != NULL
) {
2292 KKASSERT(pv
->pv_pmap
->pm_stats
.resident_count
> 0);
2293 --pv
->pv_pmap
->pm_stats
.resident_count
;
2295 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_va
);
2296 pmap_inval_interlock(&info
, pv
->pv_pmap
, pv
->pv_va
);
2297 tpte
= pte_load_clear(pte
);
2299 pv
->pv_pmap
->pm_stats
.wired_count
--;
2300 pmap_inval_deinterlock(&info
, pv
->pv_pmap
);
2302 vm_page_flag_set(m
, PG_REFERENCED
);
2305 * Update the vm_page_t clean and reference bits.
2308 #if defined(PMAP_DIAGNOSTIC)
2309 if (pmap_nw_modified(tpte
)) {
2311 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2315 if (pmap_track_modified(pv
->pv_va
))
2318 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
2319 TAILQ_REMOVE(&pv
->pv_pmap
->pm_pvlist
, pv
, pv_plist
);
2320 ++pv
->pv_pmap
->pm_generation
;
2321 m
->md
.pv_list_count
--;
2322 KKASSERT(m
->md
.pv_list_count
>= 0);
2323 if (TAILQ_EMPTY(&m
->md
.pv_list
))
2324 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
2325 pmap_unuse_pt(pv
->pv_pmap
, pv
->pv_va
, pv
->pv_ptem
, &info
);
2329 KKASSERT((m
->flags
& (PG_MAPPED
|PG_WRITEABLE
)) == 0);
2330 pmap_inval_done(&info
);
2336 * Set the physical protection on the specified range of this map
2339 * This function may not be called from an interrupt if the map is
2340 * not the kernel_pmap.
2343 pmap_protect(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
, vm_prot_t prot
)
2345 vm_offset_t va_next
;
2346 pml4_entry_t
*pml4e
;
2348 pd_entry_t ptpaddr
, *pde
;
2350 pmap_inval_info info
;
2352 /* JG review for NX */
2357 if ((prot
& VM_PROT_READ
) == VM_PROT_NONE
) {
2358 pmap_remove(pmap
, sva
, eva
);
2362 if (prot
& VM_PROT_WRITE
)
2365 pmap_inval_init(&info
);
2367 for (; sva
< eva
; sva
= va_next
) {
2369 pml4e
= pmap_pml4e(pmap
, sva
);
2370 if ((*pml4e
& PG_V
) == 0) {
2371 va_next
= (sva
+ NBPML4
) & ~PML4MASK
;
2377 pdpe
= pmap_pml4e_to_pdpe(pml4e
, sva
);
2378 if ((*pdpe
& PG_V
) == 0) {
2379 va_next
= (sva
+ NBPDP
) & ~PDPMASK
;
2385 va_next
= (sva
+ NBPDR
) & ~PDRMASK
;
2389 pde
= pmap_pdpe_to_pde(pdpe
, sva
);
2393 * Check for large page.
2395 if ((ptpaddr
& PG_PS
) != 0) {
2396 pmap_inval_interlock(&info
, pmap
, -1);
2397 *pde
&= ~(PG_M
|PG_RW
);
2398 pmap
->pm_stats
.resident_count
-= NBPDR
/ PAGE_SIZE
;
2399 pmap_inval_deinterlock(&info
, pmap
);
2404 * Weed out invalid mappings. Note: we assume that the page
2405 * directory table is always allocated, and in kernel virtual.
2413 for (pte
= pmap_pde_to_pte(pde
, sva
); sva
!= va_next
; pte
++,
2420 * XXX non-optimal. Note also that there can be
2421 * no pmap_inval_flush() calls until after we modify
2422 * ptbase[sindex] (or otherwise we have to do another
2423 * pmap_inval_add() call).
2425 pmap_inval_interlock(&info
, pmap
, sva
);
2429 if ((pbits
& PG_V
) == 0) {
2430 pmap_inval_deinterlock(&info
, pmap
);
2433 if (pbits
& PG_MANAGED
) {
2436 m
= PHYS_TO_VM_PAGE(pbits
& PG_FRAME
);
2437 vm_page_flag_set(m
, PG_REFERENCED
);
2441 if (pmap_track_modified(sva
)) {
2443 m
= PHYS_TO_VM_PAGE(pbits
& PG_FRAME
);
2450 if (pbits
!= cbits
&&
2451 !atomic_cmpset_long(pte
, pbits
, cbits
)) {
2454 pmap_inval_deinterlock(&info
, pmap
);
2457 pmap_inval_done(&info
);
2461 * Insert the given physical page (p) at
2462 * the specified virtual address (v) in the
2463 * target physical map with the protection requested.
2465 * If specified, the page will be wired down, meaning
2466 * that the related pte can not be reclaimed.
2468 * NB: This is the only routine which MAY NOT lazy-evaluate
2469 * or lose information. That is, this routine must actually
2470 * insert this page into the given map NOW.
2473 pmap_enter(pmap_t pmap
, vm_offset_t va
, vm_page_t m
, vm_prot_t prot
,
2480 pt_entry_t origpte
, newpte
;
2482 pmap_inval_info info
;
2487 va
= trunc_page(va
);
2488 #ifdef PMAP_DIAGNOSTIC
2490 panic("pmap_enter: toobig");
2491 if ((va
>= UPT_MIN_ADDRESS
) && (va
< UPT_MAX_ADDRESS
))
2492 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va
);
2494 if (va
< UPT_MAX_ADDRESS
&& pmap
== &kernel_pmap
) {
2495 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2497 db_print_backtrace();
2500 if (va
>= UPT_MAX_ADDRESS
&& pmap
!= &kernel_pmap
) {
2501 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2503 db_print_backtrace();
2508 * In the case that a page table page is not
2509 * resident, we are creating it here.
2511 if (va
< VM_MAX_USER_ADDRESS
)
2512 mpte
= pmap_allocpte(pmap
, va
);
2516 pmap_inval_init(&info
);
2517 pde
= pmap_pde(pmap
, va
);
2518 if (pde
!= NULL
&& (*pde
& PG_V
) != 0) {
2519 if ((*pde
& PG_PS
) != 0)
2520 panic("pmap_enter: attempted pmap_enter on 2MB page");
2521 pte
= pmap_pde_to_pte(pde
, va
);
2523 panic("pmap_enter: invalid page directory va=%#lx", va
);
2525 KKASSERT(pte
!= NULL
);
2526 pa
= VM_PAGE_TO_PHYS(m
);
2528 opa
= origpte
& PG_FRAME
;
2531 * Mapping has not changed, must be protection or wiring change.
2533 if (origpte
&& (opa
== pa
)) {
2535 * Wiring change, just update stats. We don't worry about
2536 * wiring PT pages as they remain resident as long as there
2537 * are valid mappings in them. Hence, if a user page is wired,
2538 * the PT page will be also.
2540 if (wired
&& ((origpte
& PG_W
) == 0))
2541 pmap
->pm_stats
.wired_count
++;
2542 else if (!wired
&& (origpte
& PG_W
))
2543 pmap
->pm_stats
.wired_count
--;
2545 #if defined(PMAP_DIAGNOSTIC)
2546 if (pmap_nw_modified(origpte
)) {
2548 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2554 * Remove the extra pte reference. Note that we cannot
2555 * optimize the RO->RW case because we have adjusted the
2556 * wiring count above and may need to adjust the wiring
2563 * We might be turning off write access to the page,
2564 * so we go ahead and sense modify status.
2566 if (origpte
& PG_MANAGED
) {
2567 if ((origpte
& PG_M
) && pmap_track_modified(va
)) {
2569 om
= PHYS_TO_VM_PAGE(opa
);
2573 KKASSERT(m
->flags
& PG_MAPPED
);
2578 * Mapping has changed, invalidate old range and fall through to
2579 * handle validating new mapping.
2583 err
= pmap_remove_pte(pmap
, pte
, va
, &info
);
2585 panic("pmap_enter: pte vanished, va: 0x%lx", va
);
2587 opa
= origpte
& PG_FRAME
;
2589 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2595 * Enter on the PV list if part of our managed memory. Note that we
2596 * raise IPL while manipulating pv_table since pmap_enter can be
2597 * called at interrupt time.
2599 if (pmap_initialized
&&
2600 (m
->flags
& (PG_FICTITIOUS
|PG_UNMANAGED
)) == 0) {
2601 pmap_insert_entry(pmap
, va
, mpte
, m
);
2603 vm_page_flag_set(m
, PG_MAPPED
);
2607 * Increment counters
2609 ++pmap
->pm_stats
.resident_count
;
2611 pmap
->pm_stats
.wired_count
++;
2615 * Now validate mapping with desired protection/wiring.
2617 newpte
= (pt_entry_t
) (pa
| pte_prot(pmap
, prot
) | PG_V
);
2621 if (va
< VM_MAX_USER_ADDRESS
)
2623 if (pmap
== &kernel_pmap
)
2627 * if the mapping or permission bits are different, we need
2628 * to update the pte.
2630 if ((origpte
& ~(PG_M
|PG_A
)) != newpte
) {
2631 pmap_inval_interlock(&info
, pmap
, va
);
2632 *pte
= newpte
| PG_A
;
2633 pmap_inval_deinterlock(&info
, pmap
);
2635 vm_page_flag_set(m
, PG_WRITEABLE
);
2637 KKASSERT((newpte
& PG_MANAGED
) == 0 || (m
->flags
& PG_MAPPED
));
2638 pmap_inval_done(&info
);
2642 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2643 * This code also assumes that the pmap has no pre-existing entry for this
2646 * This code currently may only be used on user pmaps, not kernel_pmap.
2649 pmap_enter_quick(pmap_t pmap
, vm_offset_t va
, vm_page_t m
)
2654 vm_pindex_t ptepindex
;
2656 pmap_inval_info info
;
2658 pmap_inval_init(&info
);
2660 if (va
< UPT_MAX_ADDRESS
&& pmap
== &kernel_pmap
) {
2661 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2663 db_print_backtrace();
2666 if (va
>= UPT_MAX_ADDRESS
&& pmap
!= &kernel_pmap
) {
2667 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2669 db_print_backtrace();
2673 KKASSERT(va
< UPT_MIN_ADDRESS
); /* assert used on user pmaps only */
2676 * Calculate the page table page (mpte), allocating it if necessary.
2678 * A held page table page (mpte), or NULL, is passed onto the
2679 * section following.
2681 if (va
< VM_MAX_USER_ADDRESS
) {
2683 * Calculate pagetable page index
2685 ptepindex
= pmap_pde_pindex(va
);
2689 * Get the page directory entry
2691 ptepa
= pmap_pde(pmap
, va
);
2694 * If the page table page is mapped, we just increment
2695 * the hold count, and activate it.
2697 if (ptepa
&& (*ptepa
& PG_V
) != 0) {
2699 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2700 // if (pmap->pm_ptphint &&
2701 // (pmap->pm_ptphint->pindex == ptepindex)) {
2702 // mpte = pmap->pm_ptphint;
2704 mpte
= pmap_page_lookup( pmap
->pm_pteobj
, ptepindex
);
2705 pmap
->pm_ptphint
= mpte
;
2710 mpte
= _pmap_allocpte(pmap
, ptepindex
);
2712 } while (mpte
== NULL
);
2715 /* this code path is not yet used */
2719 * With a valid (and held) page directory page, we can just use
2720 * vtopte() to get to the pte. If the pte is already present
2721 * we do not disturb it.
2726 pmap_unwire_pte_hold(pmap
, va
, mpte
, &info
);
2727 pa
= VM_PAGE_TO_PHYS(m
);
2728 KKASSERT(((*pte
^ pa
) & PG_FRAME
) == 0);
2729 pmap_inval_done(&info
);
2734 * Enter on the PV list if part of our managed memory
2736 if ((m
->flags
& (PG_FICTITIOUS
|PG_UNMANAGED
)) == 0) {
2737 pmap_insert_entry(pmap
, va
, mpte
, m
);
2738 vm_page_flag_set(m
, PG_MAPPED
);
2742 * Increment counters
2744 ++pmap
->pm_stats
.resident_count
;
2746 pa
= VM_PAGE_TO_PHYS(m
);
2749 * Now validate mapping with RO protection
2751 if (m
->flags
& (PG_FICTITIOUS
|PG_UNMANAGED
))
2752 *pte
= pa
| PG_V
| PG_U
;
2754 *pte
= pa
| PG_V
| PG_U
| PG_MANAGED
;
2755 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2756 pmap_inval_done(&info
);
2760 * Make a temporary mapping for a physical address. This is only intended
2761 * to be used for panic dumps.
2763 /* JG Needed on x86_64? */
2765 pmap_kenter_temporary(vm_paddr_t pa
, int i
)
2767 pmap_kenter((vm_offset_t
)crashdumpmap
+ (i
* PAGE_SIZE
), pa
);
2768 return ((void *)crashdumpmap
);
2771 #define MAX_INIT_PT (96)
2774 * This routine preloads the ptes for a given object into the specified pmap.
2775 * This eliminates the blast of soft faults on process startup and
2776 * immediately after an mmap.
2778 static int pmap_object_init_pt_callback(vm_page_t p
, void *data
);
2781 pmap_object_init_pt(pmap_t pmap
, vm_offset_t addr
, vm_prot_t prot
,
2782 vm_object_t object
, vm_pindex_t pindex
,
2783 vm_size_t size
, int limit
)
2785 struct rb_vm_page_scan_info info
;
2790 * We can't preinit if read access isn't set or there is no pmap
2793 if ((prot
& VM_PROT_READ
) == 0 || pmap
== NULL
|| object
== NULL
)
2797 * We can't preinit if the pmap is not the current pmap
2799 lp
= curthread
->td_lwp
;
2800 if (lp
== NULL
|| pmap
!= vmspace_pmap(lp
->lwp_vmspace
))
2803 psize
= x86_64_btop(size
);
2805 if ((object
->type
!= OBJT_VNODE
) ||
2806 ((limit
& MAP_PREFAULT_PARTIAL
) && (psize
> MAX_INIT_PT
) &&
2807 (object
->resident_page_count
> MAX_INIT_PT
))) {
2811 if (psize
+ pindex
> object
->size
) {
2812 if (object
->size
< pindex
)
2814 psize
= object
->size
- pindex
;
2821 * Use a red-black scan to traverse the requested range and load
2822 * any valid pages found into the pmap.
2824 * We cannot safely scan the object's memq unless we are in a
2825 * critical section since interrupts can remove pages from objects.
2827 info
.start_pindex
= pindex
;
2828 info
.end_pindex
= pindex
+ psize
- 1;
2835 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, rb_vm_page_scancmp
,
2836 pmap_object_init_pt_callback
, &info
);
2842 pmap_object_init_pt_callback(vm_page_t p
, void *data
)
2844 struct rb_vm_page_scan_info
*info
= data
;
2845 vm_pindex_t rel_index
;
2847 * don't allow an madvise to blow away our really
2848 * free pages allocating pv entries.
2850 if ((info
->limit
& MAP_PREFAULT_MADVISE
) &&
2851 vmstats
.v_free_count
< vmstats
.v_free_reserved
) {
2854 if (((p
->valid
& VM_PAGE_BITS_ALL
) == VM_PAGE_BITS_ALL
) &&
2855 (p
->busy
== 0) && (p
->flags
& (PG_BUSY
| PG_FICTITIOUS
)) == 0) {
2856 if ((p
->queue
- p
->pc
) == PQ_CACHE
)
2857 vm_page_deactivate(p
);
2859 rel_index
= p
->pindex
- info
->start_pindex
;
2860 pmap_enter_quick(info
->pmap
,
2861 info
->addr
+ x86_64_ptob(rel_index
), p
);
2868 * Return TRUE if the pmap is in shape to trivially
2869 * pre-fault the specified address.
2871 * Returns FALSE if it would be non-trivial or if a
2872 * pte is already loaded into the slot.
2875 pmap_prefault_ok(pmap_t pmap
, vm_offset_t addr
)
2880 pde
= pmap_pde(pmap
, addr
);
2881 if (pde
== NULL
|| *pde
== 0)
2892 * Routine: pmap_change_wiring
2893 * Function: Change the wiring attribute for a map/virtual-address
2895 * In/out conditions:
2896 * The mapping must already exist in the pmap.
2899 pmap_change_wiring(pmap_t pmap
, vm_offset_t va
, boolean_t wired
)
2906 pte
= pmap_pte(pmap
, va
);
2908 if (wired
&& !pmap_pte_w(pte
))
2909 pmap
->pm_stats
.wired_count
++;
2910 else if (!wired
&& pmap_pte_w(pte
))
2911 pmap
->pm_stats
.wired_count
--;
2914 * Wiring is not a hardware characteristic so there is no need to
2915 * invalidate TLB. However, in an SMP environment we must use
2916 * a locked bus cycle to update the pte (if we are not using
2917 * the pmap_inval_*() API that is)... it's ok to do this for simple
2922 atomic_set_long(pte
, PG_W
);
2924 atomic_clear_long(pte
, PG_W
);
2927 atomic_set_long_nonlocked(pte
, PG_W
);
2929 atomic_clear_long_nonlocked(pte
, PG_W
);
2936 * Copy the range specified by src_addr/len
2937 * from the source map to the range dst_addr/len
2938 * in the destination map.
2940 * This routine is only advisory and need not do anything.
2943 pmap_copy(pmap_t dst_pmap
, pmap_t src_pmap
, vm_offset_t dst_addr
,
2944 vm_size_t len
, vm_offset_t src_addr
)
2948 pmap_inval_info info
;
2950 vm_offset_t end_addr
= src_addr
+ len
;
2952 pd_entry_t src_frame
, dst_frame
;
2955 if (dst_addr
!= src_addr
)
2958 src_frame
= src_pmap
->pm_pdir
[PTDPTDI
] & PG_FRAME
;
2959 if (src_frame
!= (PTDpde
& PG_FRAME
)) {
2963 dst_frame
= dst_pmap
->pm_pdir
[PTDPTDI
] & PG_FRAME
;
2964 if (dst_frame
!= (APTDpde
& PG_FRAME
)) {
2965 APTDpde
= (pd_entry_t
) (dst_frame
| PG_RW
| PG_V
);
2966 /* The page directory is not shared between CPUs */
2970 pmap_inval_init(&info
);
2971 pmap_inval_add(&info
, dst_pmap
, -1);
2972 pmap_inval_add(&info
, src_pmap
, -1);
2975 * critical section protection is required to maintain the page/object
2976 * association, interrupts can free pages and remove them from
2980 for (addr
= src_addr
; addr
< end_addr
; addr
= pdnxt
) {
2981 pt_entry_t
*src_pte
, *dst_pte
;
2982 vm_page_t dstmpte
, srcmpte
;
2983 vm_offset_t srcptepaddr
;
2984 vm_pindex_t ptepindex
;
2986 if (addr
>= UPT_MIN_ADDRESS
)
2987 panic("pmap_copy: invalid to pmap_copy page tables\n");
2990 * Don't let optional prefaulting of pages make us go
2991 * way below the low water mark of free pages or way
2992 * above high water mark of used pv entries.
2994 if (vmstats
.v_free_count
< vmstats
.v_free_reserved
||
2995 pv_entry_count
> pv_entry_high_water
)
2998 pdnxt
= ((addr
+ PAGE_SIZE
*NPTEPG
) & ~(PAGE_SIZE
*NPTEPG
- 1));
2999 ptepindex
= addr
>> PDRSHIFT
;
3002 srcptepaddr
= (vm_offset_t
) src_pmap
->pm_pdir
[ptepindex
];
3004 if (srcptepaddr
== 0)
3007 if (srcptepaddr
& PG_PS
) {
3009 if (dst_pmap
->pm_pdir
[ptepindex
] == 0) {
3010 dst_pmap
->pm_pdir
[ptepindex
] = (pd_entry_t
) srcptepaddr
;
3011 dst_pmap
->pm_stats
.resident_count
+= NBPDR
/ PAGE_SIZE
;
3017 srcmpte
= vm_page_lookup(src_pmap
->pm_pteobj
, ptepindex
);
3018 if ((srcmpte
== NULL
) || (srcmpte
->hold_count
== 0) ||
3019 (srcmpte
->flags
& PG_BUSY
)) {
3023 if (pdnxt
> end_addr
)
3026 src_pte
= vtopte(addr
);
3028 dst_pte
= avtopte(addr
);
3030 while (addr
< pdnxt
) {
3035 * we only virtual copy managed pages
3037 if ((ptetemp
& PG_MANAGED
) != 0) {
3039 * We have to check after allocpte for the
3040 * pte still being around... allocpte can
3043 * pmap_allocpte() can block. If we lose
3044 * our page directory mappings we stop.
3046 dstmpte
= pmap_allocpte(dst_pmap
, addr
);
3049 if (src_frame
!= (PTDpde
& PG_FRAME
) ||
3050 dst_frame
!= (APTDpde
& PG_FRAME
)
3052 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3053 pmap_unwire_pte_hold(dst_pmap
, dstmpte
, &info
);
3055 } else if ((*dst_pte
== 0) &&
3056 (ptetemp
= *src_pte
) != 0 &&
3057 (ptetemp
& PG_MANAGED
)) {
3059 * Clear the modified and
3060 * accessed (referenced) bits
3063 m
= PHYS_TO_VM_PAGE(ptetemp
);
3064 *dst_pte
= ptetemp
& ~(PG_M
| PG_A
);
3065 ++dst_pmap
->pm_stats
.resident_count
;
3066 pmap_insert_entry(dst_pmap
, addr
,
3068 KKASSERT(m
->flags
& PG_MAPPED
);
3070 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3071 pmap_unwire_pte_hold(dst_pmap
, dstmpte
, &info
);
3075 if (dstmpte
->hold_count
>= srcmpte
->hold_count
)
3085 pmap_inval_done(&info
);
3092 * Zero the specified physical page.
3094 * This function may be called from an interrupt and no locking is
3098 pmap_zero_page(vm_paddr_t phys
)
3100 vm_offset_t va
= PHYS_TO_DMAP(phys
);
3102 pagezero((void *)va
);
3106 * pmap_page_assertzero:
3108 * Assert that a page is empty, panic if it isn't.
3111 pmap_page_assertzero(vm_paddr_t phys
)
3113 vm_offset_t virt
= PHYS_TO_DMAP(phys
);
3116 for (i
= 0; i
< PAGE_SIZE
; i
+= sizeof(long)) {
3117 if (*(long *)((char *)virt
+ i
) != 0) {
3118 panic("pmap_page_assertzero() @ %p not zero!\n", (void *)virt
);
3126 * Zero part of a physical page by mapping it into memory and clearing
3127 * its contents with bzero.
3129 * off and size may not cover an area beyond a single hardware page.
3132 pmap_zero_page_area(vm_paddr_t phys
, int off
, int size
)
3134 vm_offset_t virt
= PHYS_TO_DMAP(phys
);
3136 bzero((char *)virt
+ off
, size
);
3142 * Copy the physical page from the source PA to the target PA.
3143 * This function may be called from an interrupt. No locking
3147 pmap_copy_page(vm_paddr_t src
, vm_paddr_t dst
)
3149 vm_offset_t src_virt
, dst_virt
;
3151 src_virt
= PHYS_TO_DMAP(src
);
3152 dst_virt
= PHYS_TO_DMAP(dst
);
3153 bcopy((void *)src_virt
, (void *)dst_virt
, PAGE_SIZE
);
3157 * pmap_copy_page_frag:
3159 * Copy the physical page from the source PA to the target PA.
3160 * This function may be called from an interrupt. No locking
3164 pmap_copy_page_frag(vm_paddr_t src
, vm_paddr_t dst
, size_t bytes
)
3166 vm_offset_t src_virt
, dst_virt
;
3168 src_virt
= PHYS_TO_DMAP(src
);
3169 dst_virt
= PHYS_TO_DMAP(dst
);
3171 bcopy((char *)src_virt
+ (src
& PAGE_MASK
),
3172 (char *)dst_virt
+ (dst
& PAGE_MASK
),
3177 * Returns true if the pmap's pv is one of the first
3178 * 16 pvs linked to from this page. This count may
3179 * be changed upwards or downwards in the future; it
3180 * is only necessary that true be returned for a small
3181 * subset of pmaps for proper page aging.
3184 pmap_page_exists_quick(pmap_t pmap
, vm_page_t m
)
3189 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
3194 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
3195 if (pv
->pv_pmap
== pmap
) {
3208 * Remove all pages from specified address space
3209 * this aids process exit speeds. Also, this code
3210 * is special cased for current process only, but
3211 * can have the more generic (and slightly slower)
3212 * mode enabled. This is much faster than pmap_remove
3213 * in the case of running down an entire address space.
3216 pmap_remove_pages(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
)
3219 pt_entry_t
*pte
, tpte
;
3222 pmap_inval_info info
;
3224 int save_generation
;
3226 lp
= curthread
->td_lwp
;
3227 if (lp
&& pmap
== vmspace_pmap(lp
->lwp_vmspace
))
3232 pmap_inval_init(&info
);
3233 for (pv
= TAILQ_FIRST(&pmap
->pm_pvlist
); pv
; pv
= npv
) {
3234 if (pv
->pv_va
>= eva
|| pv
->pv_va
< sva
) {
3235 npv
= TAILQ_NEXT(pv
, pv_plist
);
3239 KKASSERT(pmap
== pv
->pv_pmap
);
3242 pte
= vtopte(pv
->pv_va
);
3244 pte
= pmap_pte_quick(pmap
, pv
->pv_va
);
3245 pmap_inval_interlock(&info
, pmap
, pv
->pv_va
);
3248 * We cannot remove wired pages from a process' mapping
3252 pmap_inval_deinterlock(&info
, pmap
);
3253 npv
= TAILQ_NEXT(pv
, pv_plist
);
3256 tpte
= pte_load_clear(pte
);
3258 m
= PHYS_TO_VM_PAGE(tpte
& PG_FRAME
);
3260 KASSERT(m
< &vm_page_array
[vm_page_array_size
],
3261 ("pmap_remove_pages: bad tpte %lx", tpte
));
3263 KKASSERT(pmap
->pm_stats
.resident_count
> 0);
3264 --pmap
->pm_stats
.resident_count
;
3265 pmap_inval_deinterlock(&info
, pmap
);
3268 * Update the vm_page_t clean and reference bits.
3274 npv
= TAILQ_NEXT(pv
, pv_plist
);
3275 TAILQ_REMOVE(&pmap
->pm_pvlist
, pv
, pv_plist
);
3276 save_generation
= ++pmap
->pm_generation
;
3278 m
->md
.pv_list_count
--;
3279 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
3280 if (TAILQ_EMPTY(&m
->md
.pv_list
))
3281 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
3283 pmap_unuse_pt(pmap
, pv
->pv_va
, pv
->pv_ptem
, &info
);
3287 * Restart the scan if we blocked during the unuse or free
3288 * calls and other removals were made.
3290 if (save_generation
!= pmap
->pm_generation
) {
3291 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3292 npv
= TAILQ_FIRST(&pmap
->pm_pvlist
);
3295 pmap_inval_done(&info
);
3299 * pmap_testbit tests bits in pte's
3300 * note that the testbit/clearbit routines are inline,
3301 * and a lot of things compile-time evaluate.
3305 pmap_testbit(vm_page_t m
, int bit
)
3310 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
3313 if (TAILQ_FIRST(&m
->md
.pv_list
) == NULL
)
3318 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
3320 * if the bit being tested is the modified bit, then
3321 * mark clean_map and ptes as never
3324 if (bit
& (PG_A
|PG_M
)) {
3325 if (!pmap_track_modified(pv
->pv_va
))
3329 #if defined(PMAP_DIAGNOSTIC)
3330 if (pv
->pv_pmap
== NULL
) {
3331 kprintf("Null pmap (tb) at va: 0x%lx\n", pv
->pv_va
);
3335 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_va
);
3346 * this routine is used to modify bits in ptes
3350 pmap_clearbit(vm_page_t m
, int bit
)
3352 struct pmap_inval_info info
;
3357 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
3360 pmap_inval_init(&info
);
3363 * Loop over all current mappings setting/clearing as appropos If
3364 * setting RO do we need to clear the VAC?
3366 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
3368 * don't write protect pager mappings
3371 if (!pmap_track_modified(pv
->pv_va
))
3375 #if defined(PMAP_DIAGNOSTIC)
3376 if (pv
->pv_pmap
== NULL
) {
3377 kprintf("Null pmap (cb) at va: 0x%lx\n", pv
->pv_va
);
3383 * Careful here. We can use a locked bus instruction to
3384 * clear PG_A or PG_M safely but we need to synchronize
3385 * with the target cpus when we mess with PG_RW.
3387 * We do not have to force synchronization when clearing
3388 * PG_M even for PTEs generated via virtual memory maps,
3389 * because the virtual kernel will invalidate the pmap
3390 * entry when/if it needs to resynchronize the Modify bit.
3393 pmap_inval_interlock(&info
, pv
->pv_pmap
, pv
->pv_va
);
3394 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_va
);
3401 atomic_clear_long(pte
, PG_M
|PG_RW
);
3404 * The cpu may be trying to set PG_M
3405 * simultaniously with our clearing
3408 if (!atomic_cmpset_long(pte
, pbits
,
3412 } else if (bit
== PG_M
) {
3414 * We could also clear PG_RW here to force
3415 * a fault on write to redetect PG_M for
3416 * virtual kernels, but it isn't necessary
3417 * since virtual kernels invalidate the pte
3418 * when they clear the VPTE_M bit in their
3419 * virtual page tables.
3421 atomic_clear_long(pte
, PG_M
);
3423 atomic_clear_long(pte
, bit
);
3427 pmap_inval_deinterlock(&info
, pv
->pv_pmap
);
3429 pmap_inval_done(&info
);
3433 * pmap_page_protect:
3435 * Lower the permission for all mappings to a given page.
3438 pmap_page_protect(vm_page_t m
, vm_prot_t prot
)
3440 /* JG NX support? */
3441 if ((prot
& VM_PROT_WRITE
) == 0) {
3442 if (prot
& (VM_PROT_READ
| VM_PROT_EXECUTE
)) {
3443 pmap_clearbit(m
, PG_RW
);
3444 vm_page_flag_clear(m
, PG_WRITEABLE
);
3452 pmap_phys_address(vm_pindex_t ppn
)
3454 return (x86_64_ptob(ppn
));
3458 * pmap_ts_referenced:
3460 * Return a count of reference bits for a page, clearing those bits.
3461 * It is not necessary for every reference bit to be cleared, but it
3462 * is necessary that 0 only be returned when there are truly no
3463 * reference bits set.
3465 * XXX: The exact number of bits to check and clear is a matter that
3466 * should be tested and standardized at some point in the future for
3467 * optimal aging of shared pages.
3470 pmap_ts_referenced(vm_page_t m
)
3472 pv_entry_t pv
, pvf
, pvn
;
3476 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
3481 if ((pv
= TAILQ_FIRST(&m
->md
.pv_list
)) != NULL
) {
3486 pvn
= TAILQ_NEXT(pv
, pv_list
);
3489 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
3490 TAILQ_INSERT_TAIL(&m
->md
.pv_list
, pv
, pv_list
);
3493 if (!pmap_track_modified(pv
->pv_va
))
3496 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_va
);
3498 if (pte
&& (*pte
& PG_A
)) {
3500 atomic_clear_long(pte
, PG_A
);
3502 atomic_clear_long_nonlocked(pte
, PG_A
);
3509 } while ((pv
= pvn
) != NULL
&& pv
!= pvf
);
3519 * Return whether or not the specified physical page was modified
3520 * in any physical maps.
3523 pmap_is_modified(vm_page_t m
)
3525 return pmap_testbit(m
, PG_M
);
3529 * Clear the modify bits on the specified physical page.
3532 pmap_clear_modify(vm_page_t m
)
3534 pmap_clearbit(m
, PG_M
);
3538 * pmap_clear_reference:
3540 * Clear the reference bit on the specified physical page.
3543 pmap_clear_reference(vm_page_t m
)
3545 pmap_clearbit(m
, PG_A
);
3549 * Miscellaneous support routines follow
3554 i386_protection_init(void)
3558 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3559 kp
= protection_codes
;
3560 for (prot
= 0; prot
< 8; prot
++) {
3562 case VM_PROT_NONE
| VM_PROT_NONE
| VM_PROT_NONE
:
3564 * Read access is also 0. There isn't any execute bit,
3565 * so just make it readable.
3567 case VM_PROT_READ
| VM_PROT_NONE
| VM_PROT_NONE
:
3568 case VM_PROT_READ
| VM_PROT_NONE
| VM_PROT_EXECUTE
:
3569 case VM_PROT_NONE
| VM_PROT_NONE
| VM_PROT_EXECUTE
:
3572 case VM_PROT_NONE
| VM_PROT_WRITE
| VM_PROT_NONE
:
3573 case VM_PROT_NONE
| VM_PROT_WRITE
| VM_PROT_EXECUTE
:
3574 case VM_PROT_READ
| VM_PROT_WRITE
| VM_PROT_NONE
:
3575 case VM_PROT_READ
| VM_PROT_WRITE
| VM_PROT_EXECUTE
:
3583 * Map a set of physical memory pages into the kernel virtual
3584 * address space. Return a pointer to where it is mapped. This
3585 * routine is intended to be used for mapping device memory,
3588 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3592 pmap_mapdev(vm_paddr_t pa
, vm_size_t size
)
3594 vm_offset_t va
, tmpva
, offset
;
3597 offset
= pa
& PAGE_MASK
;
3598 size
= roundup(offset
+ size
, PAGE_SIZE
);
3600 va
= kmem_alloc_nofault(&kernel_map
, size
);
3602 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3604 pa
= pa
& ~PAGE_MASK
;
3605 for (tmpva
= va
; size
> 0;) {
3606 pte
= vtopte(tmpva
);
3607 *pte
= pa
| PG_RW
| PG_V
; /* | pgeflag; */
3615 return ((void *)(va
+ offset
));
3619 pmap_mapdev_uncacheable(vm_paddr_t pa
, vm_size_t size
)
3621 vm_offset_t va
, tmpva
, offset
;
3624 offset
= pa
& PAGE_MASK
;
3625 size
= roundup(offset
+ size
, PAGE_SIZE
);
3627 va
= kmem_alloc_nofault(&kernel_map
, size
);
3629 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3631 pa
= pa
& ~PAGE_MASK
;
3632 for (tmpva
= va
; size
> 0;) {
3633 pte
= vtopte(tmpva
);
3634 *pte
= pa
| PG_RW
| PG_V
| PG_N
; /* | pgeflag; */
3642 return ((void *)(va
+ offset
));
3646 pmap_unmapdev(vm_offset_t va
, vm_size_t size
)
3648 vm_offset_t base
, offset
;
3650 base
= va
& ~PAGE_MASK
;
3651 offset
= va
& PAGE_MASK
;
3652 size
= roundup(offset
+ size
, PAGE_SIZE
);
3653 pmap_qremove(va
, size
>> PAGE_SHIFT
);
3654 kmem_free(&kernel_map
, base
, size
);
3658 * perform the pmap work for mincore
3661 pmap_mincore(pmap_t pmap
, vm_offset_t addr
)
3663 pt_entry_t
*ptep
, pte
;
3667 ptep
= pmap_pte(pmap
, addr
);
3672 if ((pte
= *ptep
) != 0) {
3675 val
= MINCORE_INCORE
;
3676 if ((pte
& PG_MANAGED
) == 0)
3679 pa
= pte
& PG_FRAME
;
3681 m
= PHYS_TO_VM_PAGE(pa
);
3687 val
|= MINCORE_MODIFIED
|MINCORE_MODIFIED_OTHER
;
3689 * Modified by someone
3691 else if (m
->dirty
|| pmap_is_modified(m
))
3692 val
|= MINCORE_MODIFIED_OTHER
;
3697 val
|= MINCORE_REFERENCED
|MINCORE_REFERENCED_OTHER
;
3700 * Referenced by someone
3702 else if ((m
->flags
& PG_REFERENCED
) || pmap_ts_referenced(m
)) {
3703 val
|= MINCORE_REFERENCED_OTHER
;
3704 vm_page_flag_set(m
, PG_REFERENCED
);
3711 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3712 * vmspace will be ref'd and the old one will be deref'd.
3714 * The vmspace for all lwps associated with the process will be adjusted
3715 * and cr3 will be reloaded if any lwp is the current lwp.
3718 pmap_replacevm(struct proc
*p
, struct vmspace
*newvm
, int adjrefs
)
3720 struct vmspace
*oldvm
;
3724 oldvm
= p
->p_vmspace
;
3725 if (oldvm
!= newvm
) {
3726 p
->p_vmspace
= newvm
;
3727 KKASSERT(p
->p_nthreads
== 1);
3728 lp
= RB_ROOT(&p
->p_lwp_tree
);
3729 pmap_setlwpvm(lp
, newvm
);
3731 sysref_get(&newvm
->vm_sysref
);
3732 sysref_put(&oldvm
->vm_sysref
);
3739 * Set the vmspace for a LWP. The vmspace is almost universally set the
3740 * same as the process vmspace, but virtual kernels need to swap out contexts
3741 * on a per-lwp basis.
3744 pmap_setlwpvm(struct lwp
*lp
, struct vmspace
*newvm
)
3746 struct vmspace
*oldvm
;
3750 oldvm
= lp
->lwp_vmspace
;
3752 if (oldvm
!= newvm
) {
3753 lp
->lwp_vmspace
= newvm
;
3754 if (curthread
->td_lwp
== lp
) {
3755 pmap
= vmspace_pmap(newvm
);
3757 atomic_set_int(&pmap
->pm_active
, mycpu
->gd_cpumask
);
3758 if (pmap
->pm_active
& CPUMASK_LOCK
)
3759 pmap_interlock_wait(newvm
);
3761 pmap
->pm_active
|= 1;
3763 #if defined(SWTCH_OPTIM_STATS)
3766 curthread
->td_pcb
->pcb_cr3
= vtophys(pmap
->pm_pml4
);
3767 load_cr3(curthread
->td_pcb
->pcb_cr3
);
3768 pmap
= vmspace_pmap(oldvm
);
3770 atomic_clear_int(&pmap
->pm_active
, mycpu
->gd_cpumask
);
3772 pmap
->pm_active
&= ~1;
3782 * Called when switching to a locked pmap
3785 pmap_interlock_wait(struct vmspace
*vm
)
3787 struct pmap
*pmap
= &vm
->vm_pmap
;
3789 if (pmap
->pm_active
& CPUMASK_LOCK
) {
3790 kprintf("Warning: pmap_interlock %p %08x\n",
3791 pmap
, pmap
->pm_active
);
3792 while (pmap
->pm_active
& CPUMASK_LOCK
) {
3795 lwkt_process_ipiq();
3803 pmap_addr_hint(vm_object_t obj
, vm_offset_t addr
, vm_size_t size
)
3806 if ((obj
== NULL
) || (size
< NBPDR
) || (obj
->type
!= OBJT_DEVICE
)) {
3810 addr
= (addr
+ (NBPDR
- 1)) & ~(NBPDR
- 1);
3817 static void pads (pmap_t pm
);
3818 void pmap_pvdump (vm_paddr_t pa
);
3820 /* print address space of pmap*/
3829 if (pm
== &kernel_pmap
)
3832 for (i
= 0; i
< NPDEPG
; i
++) {
3840 pmap_pvdump(vm_paddr_t pa
)
3845 kprintf("pa %08llx", (long long)pa
);
3846 m
= PHYS_TO_VM_PAGE(pa
);
3847 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
3849 kprintf(" -> pmap %p, va %x, flags %x",
3850 (void *)pv
->pv_pmap
, pv
->pv_va
, pv
->pv_flags
);
3852 kprintf(" -> pmap %p, va %x", (void *)pv
->pv_pmap
, pv
->pv_va
);