4 * Copyright (c) 1991 Regents of the University of California.
5 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
7 * Copyright (c) 2003 Peter Wemm
8 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
9 * Copyright (c) 2008, 2009 The DragonFly Project.
10 * Copyright (c) 2008, 2009 Jordan Gordeev.
11 * All rights reserved.
13 * This code is derived from software contributed to Berkeley by
14 * the Systems Programming Group of the University of Utah Computer
15 * Science Department and William Jolitz of UUNET Technologies Inc.
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions
20 * 1. Redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer.
22 * 2. Redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution.
25 * 3. All advertising materials mentioning features or use of this software
26 * must display the following acknowledgement:
27 * This product includes software developed by the University of
28 * California, Berkeley and its contributors.
29 * 4. Neither the name of the University nor the names of its contributors
30 * may be used to endorse or promote products derived from this software
31 * without specific prior written permission.
33 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
34 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
35 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
36 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
37 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
38 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
39 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
40 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
41 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
42 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
45 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
46 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
50 * Manages physical address maps.
52 * In addition to hardware address maps, this
53 * module is called upon to provide software-use-only
54 * maps which may or may not be stored in the same
55 * form as hardware maps. These pseudo-maps are
56 * used to store intermediate results from copy
57 * operations to and from address spaces.
59 * Since the information managed by this module is
60 * also stored by the logical address mapping module,
61 * this module may throw away valid virtual-to-physical
62 * mappings at almost any time. However, invalidations
63 * of virtual-to-physical mappings must be done as
66 * In order to cope with hardware architectures which
67 * make virtual-to-physical map invalidates expensive,
68 * this module may delay invalidate or reduced protection
69 * operations until such time as they are actually
70 * necessary. This module is given full information as
71 * to which processors are currently using which maps,
72 * and to when physical maps must be made correct.
76 #include "opt_disable_pse.h"
79 #include "opt_msgbuf.h"
81 #include <sys/param.h>
82 #include <sys/systm.h>
83 #include <sys/kernel.h>
85 #include <sys/msgbuf.h>
86 #include <sys/vmmeter.h>
90 #include <vm/vm_param.h>
91 #include <sys/sysctl.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_extern.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pager.h>
100 #include <vm/vm_zone.h>
102 #include <sys/user.h>
103 #include <sys/thread2.h>
104 #include <sys/sysref2.h>
106 #include <machine/cputypes.h>
107 #include <machine/md_var.h>
108 #include <machine/specialreg.h>
109 #include <machine/smp.h>
110 #include <machine_base/apic/apicreg.h>
111 #include <machine/globaldata.h>
112 #include <machine/pmap.h>
113 #include <machine/pmap_inval.h>
117 #define PMAP_KEEP_PDIRS
118 #ifndef PMAP_SHPGPERPROC
119 #define PMAP_SHPGPERPROC 200
122 #if defined(DIAGNOSTIC)
123 #define PMAP_DIAGNOSTIC
129 * Get PDEs and PTEs for user/kernel address space
131 static pd_entry_t
*pmap_pde(pmap_t pmap
, vm_offset_t va
);
132 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
134 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
135 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
136 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
137 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
138 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
142 * Given a map and a machine independent protection code,
143 * convert to a vax protection code.
145 #define pte_prot(m, p) \
146 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
147 static int protection_codes
[8];
149 struct pmap kernel_pmap
;
150 static TAILQ_HEAD(,pmap
) pmap_list
= TAILQ_HEAD_INITIALIZER(pmap_list
);
152 vm_paddr_t avail_start
; /* PA of first available physical page */
153 vm_paddr_t avail_end
; /* PA of last available physical page */
154 vm_offset_t virtual2_start
; /* cutout free area prior to kernel start */
155 vm_offset_t virtual2_end
;
156 vm_offset_t virtual_start
; /* VA of first avail page (after kernel bss) */
157 vm_offset_t virtual_end
; /* VA of last avail page (end of kernel AS) */
158 vm_offset_t KvaStart
; /* VA start of KVA space */
159 vm_offset_t KvaEnd
; /* VA end of KVA space (non-inclusive) */
160 vm_offset_t KvaSize
; /* max size of kernel virtual address space */
161 static boolean_t pmap_initialized
= FALSE
; /* Has pmap_init completed? */
162 static int pgeflag
; /* PG_G or-in */
163 static int pseflag
; /* PG_PS or-in */
165 static vm_object_t kptobj
;
168 static vm_paddr_t dmaplimit
;
170 vm_offset_t kernel_vm_end
= VM_MIN_KERNEL_ADDRESS
;
172 static uint64_t KPTbase
;
173 static uint64_t KPTphys
;
174 static uint64_t KPDphys
; /* phys addr of kernel level 2 */
175 static uint64_t KPDbase
; /* phys addr of kernel level 2 @ KERNBASE */
176 uint64_t KPDPphys
; /* phys addr of kernel level 3 */
177 uint64_t KPML4phys
; /* phys addr of kernel level 4 */
179 static uint64_t DMPDphys
; /* phys addr of direct mapped level 2 */
180 static uint64_t DMPDPphys
; /* phys addr of direct mapped level 3 */
183 * Data for the pv entry allocation mechanism
185 static vm_zone_t pvzone
;
186 static struct vm_zone pvzone_store
;
187 static struct vm_object pvzone_obj
;
188 static int pv_entry_count
=0, pv_entry_max
=0, pv_entry_high_water
=0;
189 static int pmap_pagedaemon_waken
= 0;
190 static struct pv_entry
*pvinit
;
193 * All those kernel PT submaps that BSD is so fond of
195 pt_entry_t
*CMAP1
= 0, *ptmmap
;
196 caddr_t CADDR1
= 0, ptvmmap
= 0;
197 static pt_entry_t
*msgbufmap
;
198 struct msgbuf
*msgbufp
=0;
203 static pt_entry_t
*pt_crashdumpmap
;
204 static caddr_t crashdumpmap
;
206 extern pt_entry_t
*SMPpt
;
207 extern uint64_t SMPptpa
;
211 static pv_entry_t
get_pv_entry (void);
212 static void i386_protection_init (void);
213 static void create_pagetables(vm_paddr_t
*firstaddr
);
214 static void pmap_remove_all (vm_page_t m
);
215 static int pmap_remove_pte (struct pmap
*pmap
, pt_entry_t
*ptq
,
216 vm_offset_t sva
, pmap_inval_info_t info
);
217 static void pmap_remove_page (struct pmap
*pmap
,
218 vm_offset_t va
, pmap_inval_info_t info
);
219 static int pmap_remove_entry (struct pmap
*pmap
, vm_page_t m
,
220 vm_offset_t va
, pmap_inval_info_t info
);
221 static boolean_t
pmap_testbit (vm_page_t m
, int bit
);
222 static void pmap_insert_entry (pmap_t pmap
, vm_offset_t va
,
223 vm_page_t mpte
, vm_page_t m
);
225 static vm_page_t
pmap_allocpte (pmap_t pmap
, vm_offset_t va
);
227 static int pmap_release_free_page (pmap_t pmap
, vm_page_t p
);
228 static vm_page_t
_pmap_allocpte (pmap_t pmap
, vm_pindex_t ptepindex
);
229 static pt_entry_t
* pmap_pte_quick (pmap_t pmap
, vm_offset_t va
);
230 static vm_page_t
pmap_page_lookup (vm_object_t object
, vm_pindex_t pindex
);
231 static int _pmap_unwire_pte_hold(pmap_t pmap
, vm_offset_t va
, vm_page_t m
,
232 pmap_inval_info_t info
);
233 static int pmap_unuse_pt (pmap_t
, vm_offset_t
, vm_page_t
, pmap_inval_info_t
);
234 static vm_offset_t
pmap_kmem_choose(vm_offset_t addr
);
236 static unsigned pdir4mb
;
239 * Move the kernel virtual free pointer to the next
240 * 2MB. This is used to help improve performance
241 * by using a large (2MB) page for much of the kernel
242 * (.text, .data, .bss)
246 pmap_kmem_choose(vm_offset_t addr
)
248 vm_offset_t newaddr
= addr
;
250 newaddr
= (addr
+ (NBPDR
- 1)) & ~(NBPDR
- 1);
257 * Super fast pmap_pte routine best used when scanning the pv lists.
258 * This eliminates many course-grained invltlb calls. Note that many of
259 * the pv list scans are across different pmaps and it is very wasteful
260 * to do an entire invltlb when checking a single mapping.
262 * Should only be called while in a critical section.
264 static __inline pt_entry_t
*pmap_pte(pmap_t pmap
, vm_offset_t va
);
268 pmap_pte_quick(pmap_t pmap
, vm_offset_t va
)
270 return pmap_pte(pmap
, va
);
273 /* Return a non-clipped PD index for a given VA */
276 pmap_pde_pindex(vm_offset_t va
)
278 return va
>> PDRSHIFT
;
281 /* Return various clipped indexes for a given VA */
284 pmap_pte_index(vm_offset_t va
)
287 return ((va
>> PAGE_SHIFT
) & ((1ul << NPTEPGSHIFT
) - 1));
292 pmap_pde_index(vm_offset_t va
)
295 return ((va
>> PDRSHIFT
) & ((1ul << NPDEPGSHIFT
) - 1));
300 pmap_pdpe_index(vm_offset_t va
)
303 return ((va
>> PDPSHIFT
) & ((1ul << NPDPEPGSHIFT
) - 1));
308 pmap_pml4e_index(vm_offset_t va
)
311 return ((va
>> PML4SHIFT
) & ((1ul << NPML4EPGSHIFT
) - 1));
314 /* Return a pointer to the PML4 slot that corresponds to a VA */
317 pmap_pml4e(pmap_t pmap
, vm_offset_t va
)
320 return (&pmap
->pm_pml4
[pmap_pml4e_index(va
)]);
323 /* Return a pointer to the PDP slot that corresponds to a VA */
326 pmap_pml4e_to_pdpe(pml4_entry_t
*pml4e
, vm_offset_t va
)
330 pdpe
= (pdp_entry_t
*)PHYS_TO_DMAP(*pml4e
& PG_FRAME
);
331 return (&pdpe
[pmap_pdpe_index(va
)]);
334 /* Return a pointer to the PDP slot that corresponds to a VA */
337 pmap_pdpe(pmap_t pmap
, vm_offset_t va
)
341 pml4e
= pmap_pml4e(pmap
, va
);
342 if ((*pml4e
& PG_V
) == 0)
344 return (pmap_pml4e_to_pdpe(pml4e
, va
));
347 /* Return a pointer to the PD slot that corresponds to a VA */
350 pmap_pdpe_to_pde(pdp_entry_t
*pdpe
, vm_offset_t va
)
354 pde
= (pd_entry_t
*)PHYS_TO_DMAP(*pdpe
& PG_FRAME
);
355 return (&pde
[pmap_pde_index(va
)]);
358 /* Return a pointer to the PD slot that corresponds to a VA */
361 pmap_pde(pmap_t pmap
, vm_offset_t va
)
365 pdpe
= pmap_pdpe(pmap
, va
);
366 if (pdpe
== NULL
|| (*pdpe
& PG_V
) == 0)
368 return (pmap_pdpe_to_pde(pdpe
, va
));
371 /* Return a pointer to the PT slot that corresponds to a VA */
374 pmap_pde_to_pte(pd_entry_t
*pde
, vm_offset_t va
)
378 pte
= (pt_entry_t
*)PHYS_TO_DMAP(*pde
& PG_FRAME
);
379 return (&pte
[pmap_pte_index(va
)]);
382 /* Return a pointer to the PT slot that corresponds to a VA */
385 pmap_pte(pmap_t pmap
, vm_offset_t va
)
389 pde
= pmap_pde(pmap
, va
);
390 if (pde
== NULL
|| (*pde
& PG_V
) == 0)
392 if ((*pde
& PG_PS
) != 0) /* compat with i386 pmap_pte() */
393 return ((pt_entry_t
*)pde
);
394 return (pmap_pde_to_pte(pde
, va
));
399 vtopte(vm_offset_t va
)
401 uint64_t mask
= ((1ul << (NPTEPGSHIFT
+ NPDEPGSHIFT
+ NPDPEPGSHIFT
+ NPML4EPGSHIFT
)) - 1);
403 return (PTmap
+ ((va
>> PAGE_SHIFT
) & mask
));
408 vtopde(vm_offset_t va
)
410 uint64_t mask
= ((1ul << (NPDEPGSHIFT
+ NPDPEPGSHIFT
+ NPML4EPGSHIFT
)) - 1);
412 return (PDmap
+ ((va
>> PDRSHIFT
) & mask
));
416 allocpages(vm_paddr_t
*firstaddr
, int n
)
421 bzero((void *)ret
, n
* PAGE_SIZE
);
422 *firstaddr
+= n
* PAGE_SIZE
;
428 create_pagetables(vm_paddr_t
*firstaddr
)
433 * We are running (mostly) V=P at this point
435 * Calculate NKPT - number of kernel page tables. We have to
436 * accomodoate prealloction of the vm_page_array, dump bitmap,
437 * MSGBUF_SIZE, and other stuff. Be generous.
439 * Maxmem is in pages.
441 ndmpdp
= (ptoa(Maxmem
) + NBPDP
- 1) >> PDPSHIFT
;
442 if (ndmpdp
< 4) /* Minimum 4GB of dirmap */
445 nkpt
= (Maxmem
* sizeof(struct vm_page
) + NBPDR
- 1) / NBPDR
;
446 nkpt
+= ((nkpt
+ nkpt
+ 1 + NKPML4E
+ NKPDPE
+ NDMPML4E
+ ndmpdp
) +
453 KPTbase
= allocpages(firstaddr
, nkpt
);
454 KPTphys
= allocpages(firstaddr
, nkpt
);
455 KPML4phys
= allocpages(firstaddr
, 1);
456 KPDPphys
= allocpages(firstaddr
, NKPML4E
);
459 * Calculate the page directory base for KERNBASE,
460 * that is where we start populating the page table pages.
461 * Basically this is the end - 2.
463 KPDphys
= allocpages(firstaddr
, NKPDPE
);
464 KPDbase
= KPDphys
+ ((NKPDPE
- (NPDPEPG
- KPDPI
)) << PAGE_SHIFT
);
466 DMPDPphys
= allocpages(firstaddr
, NDMPML4E
);
467 if ((amd_feature
& AMDID_PAGE1GB
) == 0)
468 DMPDphys
= allocpages(firstaddr
, ndmpdp
);
469 dmaplimit
= (vm_paddr_t
)ndmpdp
<< PDPSHIFT
;
472 * Fill in the underlying page table pages for the area around
473 * KERNBASE. This remaps low physical memory to KERNBASE.
475 * Read-only from zero to physfree
476 * XXX not fully used, underneath 2M pages
478 for (i
= 0; (i
<< PAGE_SHIFT
) < *firstaddr
; i
++) {
479 ((pt_entry_t
*)KPTbase
)[i
] = i
<< PAGE_SHIFT
;
480 ((pt_entry_t
*)KPTbase
)[i
] |= PG_RW
| PG_V
| PG_G
;
484 * Now map the initial kernel page tables. One block of page
485 * tables is placed at the beginning of kernel virtual memory,
486 * and another block is placed at KERNBASE to map the kernel binary,
487 * data, bss, and initial pre-allocations.
489 for (i
= 0; i
< nkpt
; i
++) {
490 ((pd_entry_t
*)KPDbase
)[i
] = KPTbase
+ (i
<< PAGE_SHIFT
);
491 ((pd_entry_t
*)KPDbase
)[i
] |= PG_RW
| PG_V
;
493 for (i
= 0; i
< nkpt
; i
++) {
494 ((pd_entry_t
*)KPDphys
)[i
] = KPTphys
+ (i
<< PAGE_SHIFT
);
495 ((pd_entry_t
*)KPDphys
)[i
] |= PG_RW
| PG_V
;
499 * Map from zero to end of allocations using 2M pages as an
500 * optimization. This will bypass some of the KPTBase pages
501 * above in the KERNBASE area.
503 for (i
= 0; (i
<< PDRSHIFT
) < *firstaddr
; i
++) {
504 ((pd_entry_t
*)KPDbase
)[i
] = i
<< PDRSHIFT
;
505 ((pd_entry_t
*)KPDbase
)[i
] |= PG_RW
| PG_V
| PG_PS
| PG_G
;
509 * And connect up the PD to the PDP. The kernel pmap is expected
510 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
512 for (i
= 0; i
< NKPDPE
; i
++) {
513 ((pdp_entry_t
*)KPDPphys
)[NPDPEPG
- NKPDPE
+ i
] =
514 KPDphys
+ (i
<< PAGE_SHIFT
);
515 ((pdp_entry_t
*)KPDPphys
)[NPDPEPG
- NKPDPE
+ i
] |=
519 /* Now set up the direct map space using either 2MB or 1GB pages */
520 /* Preset PG_M and PG_A because demotion expects it */
521 if ((amd_feature
& AMDID_PAGE1GB
) == 0) {
522 for (i
= 0; i
< NPDEPG
* ndmpdp
; i
++) {
523 ((pd_entry_t
*)DMPDphys
)[i
] = (vm_paddr_t
)i
<< PDRSHIFT
;
524 ((pd_entry_t
*)DMPDphys
)[i
] |= PG_RW
| PG_V
| PG_PS
|
527 /* And the direct map space's PDP */
528 for (i
= 0; i
< ndmpdp
; i
++) {
529 ((pdp_entry_t
*)DMPDPphys
)[i
] = DMPDphys
+
531 ((pdp_entry_t
*)DMPDPphys
)[i
] |= PG_RW
| PG_V
| PG_U
;
534 for (i
= 0; i
< ndmpdp
; i
++) {
535 ((pdp_entry_t
*)DMPDPphys
)[i
] =
536 (vm_paddr_t
)i
<< PDPSHIFT
;
537 ((pdp_entry_t
*)DMPDPphys
)[i
] |= PG_RW
| PG_V
| PG_PS
|
542 /* And recursively map PML4 to itself in order to get PTmap */
543 ((pdp_entry_t
*)KPML4phys
)[PML4PML4I
] = KPML4phys
;
544 ((pdp_entry_t
*)KPML4phys
)[PML4PML4I
] |= PG_RW
| PG_V
| PG_U
;
546 /* Connect the Direct Map slot up to the PML4 */
547 ((pdp_entry_t
*)KPML4phys
)[DMPML4I
] = DMPDPphys
;
548 ((pdp_entry_t
*)KPML4phys
)[DMPML4I
] |= PG_RW
| PG_V
| PG_U
;
550 /* Connect the KVA slot up to the PML4 */
551 ((pdp_entry_t
*)KPML4phys
)[KPML4I
] = KPDPphys
;
552 ((pdp_entry_t
*)KPML4phys
)[KPML4I
] |= PG_RW
| PG_V
| PG_U
;
556 * Bootstrap the system enough to run with virtual memory.
558 * On the i386 this is called after mapping has already been enabled
559 * and just syncs the pmap module with what has already been done.
560 * [We can't call it easily with mapping off since the kernel is not
561 * mapped with PA == VA, hence we would have to relocate every address
562 * from the linked base (virtual) address "KERNBASE" to the actual
563 * (physical) address starting relative to 0]
566 pmap_bootstrap(vm_paddr_t
*firstaddr
)
570 struct mdglobaldata
*gd
;
573 KvaStart
= VM_MIN_KERNEL_ADDRESS
;
574 KvaEnd
= VM_MAX_KERNEL_ADDRESS
;
575 KvaSize
= KvaEnd
- KvaStart
;
577 avail_start
= *firstaddr
;
580 * Create an initial set of page tables to run the kernel in.
582 create_pagetables(firstaddr
);
584 virtual2_start
= KvaStart
;
585 virtual2_end
= PTOV_OFFSET
;
587 virtual_start
= (vm_offset_t
) PTOV_OFFSET
+ *firstaddr
;
588 virtual_start
= pmap_kmem_choose(virtual_start
);
590 virtual_end
= VM_MAX_KERNEL_ADDRESS
;
592 /* XXX do %cr0 as well */
593 load_cr4(rcr4() | CR4_PGE
| CR4_PSE
);
597 * Initialize protection array.
599 i386_protection_init();
602 * The kernel's pmap is statically allocated so we don't have to use
603 * pmap_create, which is unlikely to work correctly at this part of
604 * the boot sequence (XXX and which no longer exists).
606 kernel_pmap
.pm_pml4
= (pdp_entry_t
*) (PTOV_OFFSET
+ KPML4phys
);
607 kernel_pmap
.pm_count
= 1;
608 kernel_pmap
.pm_active
= (cpumask_t
)-1 & ~CPUMASK_LOCK
;
609 TAILQ_INIT(&kernel_pmap
.pm_pvlist
);
612 * Reserve some special page table entries/VA space for temporary
615 #define SYSMAP(c, p, v, n) \
616 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
622 * CMAP1/CMAP2 are used for zeroing and copying pages.
624 SYSMAP(caddr_t
, CMAP1
, CADDR1
, 1)
629 SYSMAP(caddr_t
, pt_crashdumpmap
, crashdumpmap
, MAXDUMPPGS
);
632 * ptvmmap is used for reading arbitrary physical pages via
635 SYSMAP(caddr_t
, ptmmap
, ptvmmap
, 1)
638 * msgbufp is used to map the system message buffer.
639 * XXX msgbufmap is not used.
641 SYSMAP(struct msgbuf
*, msgbufmap
, msgbufp
,
642 atop(round_page(MSGBUF_SIZE
)))
649 * PG_G is terribly broken on SMP because we IPI invltlb's in some
650 * cases rather then invl1pg. Actually, I don't even know why it
651 * works under UP because self-referential page table mappings
656 if (cpu_feature
& CPUID_PGE
)
661 * Initialize the 4MB page size flag
665 * The 4MB page version of the initial
666 * kernel page mapping.
670 #if !defined(DISABLE_PSE)
671 if (cpu_feature
& CPUID_PSE
) {
674 * Note that we have enabled PSE mode
677 ptditmp
= *(PTmap
+ x86_64_btop(KERNBASE
));
678 ptditmp
&= ~(NBPDR
- 1);
679 ptditmp
|= PG_V
| PG_RW
| PG_PS
| PG_U
| pgeflag
;
684 * Enable the PSE mode. If we are SMP we can't do this
685 * now because the APs will not be able to use it when
688 load_cr4(rcr4() | CR4_PSE
);
691 * We can do the mapping here for the single processor
692 * case. We simply ignore the old page table page from
696 * For SMP, we still need 4K pages to bootstrap APs,
697 * PSE will be enabled as soon as all APs are up.
699 PTD
[KPTDI
] = (pd_entry_t
)ptditmp
;
706 * We need to finish setting up the globaldata page for the BSP.
707 * locore has already populated the page table for the mdglobaldata
710 pg
= MDGLOBALDATA_BASEALLOC_PAGES
;
711 gd
= &CPU_prvspace
[0].mdglobaldata
;
712 gd
->gd_CMAP1
= &SMPpt
[pg
+ 0];
713 gd
->gd_CMAP2
= &SMPpt
[pg
+ 1];
714 gd
->gd_CMAP3
= &SMPpt
[pg
+ 2];
715 gd
->gd_PMAP1
= &SMPpt
[pg
+ 3];
716 gd
->gd_CADDR1
= CPU_prvspace
[0].CPAGE1
;
717 gd
->gd_CADDR2
= CPU_prvspace
[0].CPAGE2
;
718 gd
->gd_CADDR3
= CPU_prvspace
[0].CPAGE3
;
719 gd
->gd_PADDR1
= (pt_entry_t
*)CPU_prvspace
[0].PPAGE1
;
726 * Set 4mb pdir for mp startup
731 if (pseflag
&& (cpu_feature
& CPUID_PSE
)) {
732 load_cr4(rcr4() | CR4_PSE
);
733 if (pdir4mb
&& mycpu
->gd_cpuid
== 0) { /* only on BSP */
741 * Initialize the pmap module.
742 * Called by vm_init, to initialize any structures that the pmap
743 * system needs to map virtual memory.
744 * pmap_init has been enhanced to support in a fairly consistant
745 * way, discontiguous physical memory.
754 * object for kernel page table pages
756 /* JG I think the number can be arbitrary */
757 kptobj
= vm_object_allocate(OBJT_DEFAULT
, 5);
760 * Allocate memory for random pmap data structures. Includes the
764 for(i
= 0; i
< vm_page_array_size
; i
++) {
767 m
= &vm_page_array
[i
];
768 TAILQ_INIT(&m
->md
.pv_list
);
769 m
->md
.pv_list_count
= 0;
773 * init the pv free list
775 initial_pvs
= vm_page_array_size
;
776 if (initial_pvs
< MINPV
)
778 pvzone
= &pvzone_store
;
779 pvinit
= (void *)kmem_alloc(&kernel_map
,
780 initial_pvs
* sizeof (struct pv_entry
));
781 zbootinit(pvzone
, "PV ENTRY", sizeof (struct pv_entry
),
782 pvinit
, initial_pvs
);
785 * Now it is safe to enable pv_table recording.
787 pmap_initialized
= TRUE
;
789 lapic
= pmap_mapdev_uncacheable(cpu_apic_address
, sizeof(struct LAPIC
));
794 * Initialize the address space (zone) for the pv_entries. Set a
795 * high water mark so that the system can recover from excessive
796 * numbers of pv entries.
801 int shpgperproc
= PMAP_SHPGPERPROC
;
804 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc
);
805 pv_entry_max
= shpgperproc
* maxproc
+ vm_page_array_size
;
806 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max
);
807 pv_entry_high_water
= 9 * (pv_entry_max
/ 10);
810 * Subtract out pages already installed in the zone (hack)
812 entry_max
= pv_entry_max
- vm_page_array_size
;
816 zinitna(pvzone
, &pvzone_obj
, NULL
, 0, entry_max
, ZONE_INTERRUPT
, 1);
820 /***************************************************
821 * Low level helper routines.....
822 ***************************************************/
824 #if defined(PMAP_DIAGNOSTIC)
827 * This code checks for non-writeable/modified pages.
828 * This should be an invalid condition.
832 pmap_nw_modified(pt_entry_t pte
)
834 if ((pte
& (PG_M
|PG_RW
)) == PG_M
)
843 * this routine defines the region(s) of memory that should
844 * not be tested for the modified bit.
848 pmap_track_modified(vm_offset_t va
)
850 if ((va
< clean_sva
) || (va
>= clean_eva
))
857 * Extract the physical page address associated with the map/VA pair.
859 * The caller must hold vm_token if non-blocking operation is desired.
862 pmap_extract(pmap_t pmap
, vm_offset_t va
)
866 pd_entry_t pde
, *pdep
;
868 lwkt_gettoken(&vm_token
);
870 pdep
= pmap_pde(pmap
, va
);
874 if ((pde
& PG_PS
) != 0) {
875 rtval
= (pde
& PG_PS_FRAME
) | (va
& PDRMASK
);
877 pte
= pmap_pde_to_pte(pdep
, va
);
878 rtval
= (*pte
& PG_FRAME
) | (va
& PAGE_MASK
);
882 lwkt_reltoken(&vm_token
);
887 * Extract the physical page address associated kernel virtual address.
890 pmap_kextract(vm_offset_t va
)
895 if (va
>= DMAP_MIN_ADDRESS
&& va
< DMAP_MAX_ADDRESS
) {
896 pa
= DMAP_TO_PHYS(va
);
900 pa
= (pde
& PG_PS_FRAME
) | (va
& PDRMASK
);
903 * Beware of a concurrent promotion that changes the
904 * PDE at this point! For example, vtopte() must not
905 * be used to access the PTE because it would use the
906 * new PDE. It is, however, safe to use the old PDE
907 * because the page table page is preserved by the
910 pa
= *pmap_pde_to_pte(&pde
, va
);
911 pa
= (pa
& PG_FRAME
) | (va
& PAGE_MASK
);
917 /***************************************************
918 * Low level mapping routines.....
919 ***************************************************/
922 * Routine: pmap_kenter
924 * Add a wired page to the KVA
925 * NOTE! note that in order for the mapping to take effect -- you
926 * should do an invltlb after doing the pmap_kenter().
929 pmap_kenter(vm_offset_t va
, vm_paddr_t pa
)
933 pmap_inval_info info
;
935 pmap_inval_init(&info
);
936 npte
= pa
| PG_RW
| PG_V
| pgeflag
;
938 pmap_inval_interlock(&info
, &kernel_pmap
, va
);
940 pmap_inval_deinterlock(&info
, &kernel_pmap
);
941 pmap_inval_done(&info
);
945 * Routine: pmap_kenter_quick
947 * Similar to pmap_kenter(), except we only invalidate the
948 * mapping on the current CPU.
951 pmap_kenter_quick(vm_offset_t va
, vm_paddr_t pa
)
956 npte
= pa
| PG_RW
| PG_V
| pgeflag
;
959 cpu_invlpg((void *)va
);
963 pmap_kenter_sync(vm_offset_t va
)
965 pmap_inval_info info
;
967 pmap_inval_init(&info
);
968 pmap_inval_interlock(&info
, &kernel_pmap
, va
);
969 pmap_inval_deinterlock(&info
, &kernel_pmap
);
970 pmap_inval_done(&info
);
974 pmap_kenter_sync_quick(vm_offset_t va
)
976 cpu_invlpg((void *)va
);
980 * remove a page from the kernel pagetables
983 pmap_kremove(vm_offset_t va
)
986 pmap_inval_info info
;
988 pmap_inval_init(&info
);
990 pmap_inval_interlock(&info
, &kernel_pmap
, va
);
992 pmap_inval_deinterlock(&info
, &kernel_pmap
);
993 pmap_inval_done(&info
);
997 pmap_kremove_quick(vm_offset_t va
)
1002 cpu_invlpg((void *)va
);
1006 * XXX these need to be recoded. They are not used in any critical path.
1009 pmap_kmodify_rw(vm_offset_t va
)
1011 *vtopte(va
) |= PG_RW
;
1012 cpu_invlpg((void *)va
);
1016 pmap_kmodify_nc(vm_offset_t va
)
1018 *vtopte(va
) |= PG_N
;
1019 cpu_invlpg((void *)va
);
1023 * Used to map a range of physical addresses into kernel virtual
1024 * address space during the low level boot, typically to map the
1025 * dump bitmap, message buffer, and vm_page_array.
1027 * These mappings are typically made at some pointer after the end of the
1030 * We could return PHYS_TO_DMAP(start) here and not allocate any
1031 * via (*virtp), but then kmem from userland and kernel dumps won't
1032 * have access to the related pointers.
1035 pmap_map(vm_offset_t
*virtp
, vm_paddr_t start
, vm_paddr_t end
, int prot
)
1038 vm_offset_t va_start
;
1040 /*return PHYS_TO_DMAP(start);*/
1045 while (start
< end
) {
1046 pmap_kenter_quick(va
, start
);
1056 * Add a list of wired pages to the kva
1057 * this routine is only used for temporary
1058 * kernel mappings that do not need to have
1059 * page modification or references recorded.
1060 * Note that old mappings are simply written
1061 * over. The page *must* be wired.
1064 pmap_qenter(vm_offset_t va
, vm_page_t
*m
, int count
)
1068 end_va
= va
+ count
* PAGE_SIZE
;
1070 while (va
< end_va
) {
1074 *pte
= VM_PAGE_TO_PHYS(*m
) | PG_RW
| PG_V
| pgeflag
;
1075 cpu_invlpg((void *)va
);
1080 smp_invltlb(); /* XXX */
1085 * This routine jerks page mappings from the
1086 * kernel -- it is meant only for temporary mappings.
1088 * MPSAFE, INTERRUPT SAFE (cluster callback)
1091 pmap_qremove(vm_offset_t va
, int count
)
1095 end_va
= va
+ count
* PAGE_SIZE
;
1097 while (va
< end_va
) {
1102 cpu_invlpg((void *)va
);
1111 * This routine works like vm_page_lookup() but also blocks as long as the
1112 * page is busy. This routine does not busy the page it returns.
1114 * Unless the caller is managing objects whos pages are in a known state,
1115 * the call should be made with a critical section held so the page's object
1116 * association remains valid on return.
1120 pmap_page_lookup(vm_object_t object
, vm_pindex_t pindex
)
1125 m
= vm_page_lookup(object
, pindex
);
1126 } while (m
&& vm_page_sleep_busy(m
, FALSE
, "pplookp"));
1132 * Create a new thread and optionally associate it with a (new) process.
1133 * NOTE! the new thread's cpu may not equal the current cpu.
1136 pmap_init_thread(thread_t td
)
1138 /* enforce pcb placement */
1139 td
->td_pcb
= (struct pcb
*)(td
->td_kstack
+ td
->td_kstack_size
) - 1;
1140 td
->td_savefpu
= &td
->td_pcb
->pcb_save
;
1141 td
->td_sp
= (char *)td
->td_pcb
- 16; /* JG is -16 needed on x86_64? */
1145 * This routine directly affects the fork perf for a process.
1148 pmap_init_proc(struct proc
*p
)
1153 * Dispose the UPAGES for a process that has exited.
1154 * This routine directly impacts the exit perf of a process.
1157 pmap_dispose_proc(struct proc
*p
)
1159 KASSERT(p
->p_lock
== 0, ("attempt to dispose referenced proc! %p", p
));
1162 /***************************************************
1163 * Page table page management routines.....
1164 ***************************************************/
1167 * This routine unholds page table pages, and if the hold count
1168 * drops to zero, then it decrements the wire count.
1172 pmap_unwire_pte_hold(pmap_t pmap
, vm_offset_t va
, vm_page_t m
,
1173 pmap_inval_info_t info
)
1175 KKASSERT(m
->hold_count
> 0);
1176 if (m
->hold_count
> 1) {
1180 return _pmap_unwire_pte_hold(pmap
, va
, m
, info
);
1186 _pmap_unwire_pte_hold(pmap_t pmap
, vm_offset_t va
, vm_page_t m
,
1187 pmap_inval_info_t info
)
1190 * Wait until we can busy the page ourselves. We cannot have
1191 * any active flushes if we block. We own one hold count on the
1192 * page so it cannot be freed out from under us.
1194 if (m
->flags
& PG_BUSY
) {
1195 pmap_inval_flush(info
);
1196 while (vm_page_sleep_busy(m
, FALSE
, "pmuwpt"))
1199 KASSERT(m
->queue
== PQ_NONE
,
1200 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m
));
1203 * This case can occur if new references were acquired while
1206 if (m
->hold_count
> 1) {
1207 KKASSERT(m
->hold_count
> 1);
1213 * Unmap the page table page
1215 KKASSERT(m
->hold_count
== 1);
1217 pmap_inval_interlock(info
, pmap
, -1);
1219 if (m
->pindex
>= (NUPDE
+ NUPDPE
)) {
1222 pml4
= pmap_pml4e(pmap
, va
);
1224 } else if (m
->pindex
>= NUPDE
) {
1227 pdp
= pmap_pdpe(pmap
, va
);
1232 pd
= pmap_pde(pmap
, va
);
1236 KKASSERT(pmap
->pm_stats
.resident_count
> 0);
1237 --pmap
->pm_stats
.resident_count
;
1239 if (pmap
->pm_ptphint
== m
)
1240 pmap
->pm_ptphint
= NULL
;
1241 pmap_inval_deinterlock(info
, pmap
);
1243 if (m
->pindex
< NUPDE
) {
1244 /* We just released a PT, unhold the matching PD */
1247 pdpg
= PHYS_TO_VM_PAGE(*pmap_pdpe(pmap
, va
) & PG_FRAME
);
1248 pmap_unwire_pte_hold(pmap
, va
, pdpg
, info
);
1250 if (m
->pindex
>= NUPDE
&& m
->pindex
< (NUPDE
+ NUPDPE
)) {
1251 /* We just released a PD, unhold the matching PDP */
1254 pdppg
= PHYS_TO_VM_PAGE(*pmap_pml4e(pmap
, va
) & PG_FRAME
);
1255 pmap_unwire_pte_hold(pmap
, va
, pdppg
, info
);
1259 * This was our last hold, the page had better be unwired
1260 * after we decrement wire_count.
1262 * FUTURE NOTE: shared page directory page could result in
1263 * multiple wire counts.
1267 KKASSERT(m
->wire_count
== 0);
1268 --vmstats
.v_wire_count
;
1269 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
1271 vm_page_free_zero(m
);
1277 * After removing a page table entry, this routine is used to
1278 * conditionally free the page, and manage the hold/wire counts.
1282 pmap_unuse_pt(pmap_t pmap
, vm_offset_t va
, vm_page_t mpte
,
1283 pmap_inval_info_t info
)
1285 vm_pindex_t ptepindex
;
1287 if (va
>= VM_MAX_USER_ADDRESS
)
1291 ptepindex
= pmap_pde_pindex(va
);
1293 if (pmap
->pm_ptphint
&&
1294 (pmap
->pm_ptphint
->pindex
== ptepindex
)) {
1295 mpte
= pmap
->pm_ptphint
;
1298 pmap_inval_flush(info
);
1299 mpte
= pmap_page_lookup(pmap
->pm_pteobj
, ptepindex
);
1300 pmap
->pm_ptphint
= mpte
;
1305 return pmap_unwire_pte_hold(pmap
, va
, mpte
, info
);
1309 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1310 * it, and IdlePTD, represents the template used to update all other pmaps.
1312 * On architectures where the kernel pmap is not integrated into the user
1313 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1314 * kernel_pmap should be used to directly access the kernel_pmap.
1317 pmap_pinit0(struct pmap
*pmap
)
1319 pmap
->pm_pml4
= (pml4_entry_t
*)(PTOV_OFFSET
+ KPML4phys
);
1321 pmap
->pm_active
= 0;
1322 pmap
->pm_ptphint
= NULL
;
1323 TAILQ_INIT(&pmap
->pm_pvlist
);
1324 bzero(&pmap
->pm_stats
, sizeof pmap
->pm_stats
);
1328 * Initialize a preallocated and zeroed pmap structure,
1329 * such as one in a vmspace structure.
1332 pmap_pinit(struct pmap
*pmap
)
1337 * No need to allocate page table space yet but we do need a valid
1338 * page directory table.
1340 if (pmap
->pm_pml4
== NULL
) {
1342 (pml4_entry_t
*)kmem_alloc_pageable(&kernel_map
, PAGE_SIZE
);
1346 * Allocate an object for the ptes
1348 if (pmap
->pm_pteobj
== NULL
)
1349 pmap
->pm_pteobj
= vm_object_allocate(OBJT_DEFAULT
, NUPDE
+ NUPDPE
+ PML4PML4I
+ 1);
1352 * Allocate the page directory page, unless we already have
1353 * one cached. If we used the cached page the wire_count will
1354 * already be set appropriately.
1356 if ((ptdpg
= pmap
->pm_pdirm
) == NULL
) {
1357 ptdpg
= vm_page_grab(pmap
->pm_pteobj
, NUPDE
+ NUPDPE
+ PML4PML4I
,
1358 VM_ALLOC_NORMAL
| VM_ALLOC_RETRY
);
1359 pmap
->pm_pdirm
= ptdpg
;
1360 vm_page_flag_clear(ptdpg
, PG_MAPPED
| PG_BUSY
);
1361 ptdpg
->valid
= VM_PAGE_BITS_ALL
;
1362 if (ptdpg
->wire_count
== 0)
1363 ++vmstats
.v_wire_count
;
1364 ptdpg
->wire_count
= 1;
1365 pmap_kenter((vm_offset_t
)pmap
->pm_pml4
, VM_PAGE_TO_PHYS(ptdpg
));
1367 if ((ptdpg
->flags
& PG_ZERO
) == 0)
1368 bzero(pmap
->pm_pml4
, PAGE_SIZE
);
1370 pmap
->pm_pml4
[KPML4I
] = KPDPphys
| PG_RW
| PG_V
| PG_U
;
1371 pmap
->pm_pml4
[DMPML4I
] = DMPDPphys
| PG_RW
| PG_V
| PG_U
;
1373 /* install self-referential address mapping entry */
1374 pmap
->pm_pml4
[PML4PML4I
] = VM_PAGE_TO_PHYS(ptdpg
) | PG_V
| PG_RW
| PG_A
| PG_M
;
1377 pmap
->pm_active
= 0;
1378 pmap
->pm_ptphint
= NULL
;
1379 TAILQ_INIT(&pmap
->pm_pvlist
);
1380 bzero(&pmap
->pm_stats
, sizeof pmap
->pm_stats
);
1381 pmap
->pm_stats
.resident_count
= 1;
1385 * Clean up a pmap structure so it can be physically freed. This routine
1386 * is called by the vmspace dtor function. A great deal of pmap data is
1387 * left passively mapped to improve vmspace management so we have a bit
1388 * of cleanup work to do here.
1391 pmap_puninit(pmap_t pmap
)
1395 KKASSERT(pmap
->pm_active
== 0);
1396 lwkt_gettoken(&vm_token
);
1397 if ((p
= pmap
->pm_pdirm
) != NULL
) {
1398 KKASSERT(pmap
->pm_pml4
!= NULL
);
1399 KKASSERT(pmap
->pm_pml4
!= (void *)(PTOV_OFFSET
+ KPML4phys
));
1400 pmap_kremove((vm_offset_t
)pmap
->pm_pml4
);
1402 vmstats
.v_wire_count
--;
1403 KKASSERT((p
->flags
& PG_BUSY
) == 0);
1405 vm_page_free_zero(p
);
1406 pmap
->pm_pdirm
= NULL
;
1408 if (pmap
->pm_pml4
) {
1409 KKASSERT(pmap
->pm_pml4
!= (void *)(PTOV_OFFSET
+ KPML4phys
));
1410 kmem_free(&kernel_map
, (vm_offset_t
)pmap
->pm_pml4
, PAGE_SIZE
);
1411 pmap
->pm_pml4
= NULL
;
1413 if (pmap
->pm_pteobj
) {
1414 vm_object_deallocate(pmap
->pm_pteobj
);
1415 pmap
->pm_pteobj
= NULL
;
1417 lwkt_reltoken(&vm_token
);
1421 * Wire in kernel global address entries. To avoid a race condition
1422 * between pmap initialization and pmap_growkernel, this procedure
1423 * adds the pmap to the master list (which growkernel scans to update),
1424 * then copies the template.
1427 pmap_pinit2(struct pmap
*pmap
)
1430 lwkt_gettoken(&vm_token
);
1431 TAILQ_INSERT_TAIL(&pmap_list
, pmap
, pm_pmnode
);
1432 /* XXX copies current process, does not fill in MPPTDI */
1433 lwkt_reltoken(&vm_token
);
1438 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1439 * 0 on failure (if the procedure had to sleep).
1441 * When asked to remove the page directory page itself, we actually just
1442 * leave it cached so we do not have to incur the SMP inval overhead of
1443 * removing the kernel mapping. pmap_puninit() will take care of it.
1447 pmap_release_free_page(struct pmap
*pmap
, vm_page_t p
)
1450 * This code optimizes the case of freeing non-busy
1451 * page-table pages. Those pages are zero now, and
1452 * might as well be placed directly into the zero queue.
1454 if (vm_page_sleep_busy(p
, FALSE
, "pmaprl"))
1460 * Remove the page table page from the processes address space.
1462 if (p
->pindex
== NUPDE
+ NUPDPE
+ PML4PML4I
) {
1464 * We are the pml4 table itself.
1466 /* XXX anything to do here? */
1467 } else if (p
->pindex
>= (NUPDE
+ NUPDPE
)) {
1469 * Remove a PDP page from the PML4. We do not maintain
1470 * hold counts on the PML4 page.
1476 m4
= vm_page_lookup(pmap
->pm_pteobj
, NUPDE
+ NUPDPE
+ PML4PML4I
);
1477 KKASSERT(m4
!= NULL
);
1478 pml4
= (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4
));
1479 idx
= (p
->pindex
- (NUPDE
+ NUPDPE
)) % NPML4EPG
;
1480 KKASSERT(pml4
[idx
] != 0);
1482 } else if (p
->pindex
>= NUPDE
) {
1484 * Remove a PD page from the PDP and drop the hold count
1485 * on the PDP. The PDP is left cached in the pmap if
1486 * the hold count drops to 0 so the wire count remains
1493 m3
= vm_page_lookup(pmap
->pm_pteobj
,
1494 NUPDE
+ NUPDPE
+ (p
->pindex
- NUPDE
) / NPDPEPG
);
1495 KKASSERT(m3
!= NULL
);
1496 pdp
= (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3
));
1497 idx
= (p
->pindex
- NUPDE
) % NPDPEPG
;
1498 KKASSERT(pdp
[idx
] != 0);
1503 * Remove a PT page from the PD and drop the hold count
1504 * on the PD. The PD is left cached in the pmap if
1505 * the hold count drops to 0 so the wire count remains
1512 m2
= vm_page_lookup(pmap
->pm_pteobj
,
1513 NUPDE
+ p
->pindex
/ NPDEPG
);
1514 KKASSERT(m2
!= NULL
);
1515 pd
= (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2
));
1516 idx
= p
->pindex
% NPDEPG
;
1522 * One fewer mappings in the pmap. p's hold count had better
1525 KKASSERT(pmap
->pm_stats
.resident_count
> 0);
1526 --pmap
->pm_stats
.resident_count
;
1528 panic("pmap_release: freeing held page table page");
1529 if (pmap
->pm_ptphint
&& (pmap
->pm_ptphint
->pindex
== p
->pindex
))
1530 pmap
->pm_ptphint
= NULL
;
1533 * We leave the top-level page table page cached, wired, and mapped in
1534 * the pmap until the dtor function (pmap_puninit()) gets called.
1535 * However, still clean it up so we can set PG_ZERO.
1537 if (p
->pindex
== NUPDE
+ NUPDPE
+ PML4PML4I
) {
1538 bzero(pmap
->pm_pml4
, PAGE_SIZE
);
1539 vm_page_flag_set(p
, PG_ZERO
);
1543 KKASSERT(p
->wire_count
== 0);
1544 vmstats
.v_wire_count
--;
1545 /* JG eventually revert to using vm_page_free_zero() */
1552 * This routine is called when various levels in the page table need to
1553 * be populated. This routine cannot fail.
1557 _pmap_allocpte(pmap_t pmap
, vm_pindex_t ptepindex
)
1562 * Find or fabricate a new pagetable page. This will busy the page.
1564 m
= vm_page_grab(pmap
->pm_pteobj
, ptepindex
,
1565 VM_ALLOC_NORMAL
| VM_ALLOC_ZERO
| VM_ALLOC_RETRY
);
1566 if ((m
->flags
& PG_ZERO
) == 0) {
1567 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
1570 KASSERT(m
->queue
== PQ_NONE
,
1571 ("_pmap_allocpte: %p->queue != PQ_NONE", m
));
1574 * Increment the hold count for the page we will be returning to
1578 if (m
->wire_count
++ == 0)
1579 vmstats
.v_wire_count
++;
1582 * Map the pagetable page into the process address space, if
1583 * it isn't already there.
1585 * It is possible that someone else got in and mapped the page
1586 * directory page while we were blocked, if so just unbusy and
1587 * return the held page.
1589 if (ptepindex
>= (NUPDE
+ NUPDPE
)) {
1591 * Wire up a new PDP page in the PML4
1593 vm_pindex_t pml4index
;
1596 pml4index
= ptepindex
- (NUPDE
+ NUPDPE
);
1597 pml4
= &pmap
->pm_pml4
[pml4index
];
1599 if (--m
->wire_count
== 0)
1600 --vmstats
.v_wire_count
;
1604 *pml4
= VM_PAGE_TO_PHYS(m
) | PG_U
| PG_RW
| PG_V
| PG_A
| PG_M
;
1605 } else if (ptepindex
>= NUPDE
) {
1607 * Wire up a new PD page in the PDP
1609 vm_pindex_t pml4index
;
1610 vm_pindex_t pdpindex
;
1615 pdpindex
= ptepindex
- NUPDE
;
1616 pml4index
= pdpindex
>> NPML4EPGSHIFT
;
1618 pml4
= &pmap
->pm_pml4
[pml4index
];
1619 if ((*pml4
& PG_V
) == 0) {
1621 * Have to allocate a new PDP page, recurse.
1622 * This always succeeds. Returned page will
1625 pdppg
= _pmap_allocpte(pmap
,
1626 NUPDE
+ NUPDPE
+ pml4index
);
1629 * Add a held reference to the PDP page.
1631 pdppg
= PHYS_TO_VM_PAGE(*pml4
& PG_FRAME
);
1632 pdppg
->hold_count
++;
1636 * Now find the pdp_entry and map the PDP. If the PDP
1637 * has already been mapped unwind and return the
1638 * already-mapped PDP held.
1640 * pdppg is left held (hold_count is incremented for
1641 * each PD in the PDP).
1643 pdp
= (pdp_entry_t
*)PHYS_TO_DMAP(*pml4
& PG_FRAME
);
1644 pdp
= &pdp
[pdpindex
& ((1ul << NPDPEPGSHIFT
) - 1)];
1646 vm_page_unhold(pdppg
);
1647 if (--m
->wire_count
== 0)
1648 --vmstats
.v_wire_count
;
1652 *pdp
= VM_PAGE_TO_PHYS(m
) | PG_U
| PG_RW
| PG_V
| PG_A
| PG_M
;
1655 * Wire up the new PT page in the PD
1657 vm_pindex_t pml4index
;
1658 vm_pindex_t pdpindex
;
1664 pdpindex
= ptepindex
>> NPDPEPGSHIFT
;
1665 pml4index
= pdpindex
>> NPML4EPGSHIFT
;
1668 * Locate the PDP page in the PML4, then the PD page in
1669 * the PDP. If either does not exist we simply recurse
1672 * We can just recurse on the PD page as it will recurse
1673 * on the PDP if necessary.
1675 pml4
= &pmap
->pm_pml4
[pml4index
];
1676 if ((*pml4
& PG_V
) == 0) {
1677 pdpg
= _pmap_allocpte(pmap
, NUPDE
+ pdpindex
);
1678 pdp
= (pdp_entry_t
*)PHYS_TO_DMAP(*pml4
& PG_FRAME
);
1679 pdp
= &pdp
[pdpindex
& ((1ul << NPDPEPGSHIFT
) - 1)];
1681 pdp
= (pdp_entry_t
*)PHYS_TO_DMAP(*pml4
& PG_FRAME
);
1682 pdp
= &pdp
[pdpindex
& ((1ul << NPDPEPGSHIFT
) - 1)];
1683 if ((*pdp
& PG_V
) == 0) {
1684 pdpg
= _pmap_allocpte(pmap
, NUPDE
+ pdpindex
);
1686 pdpg
= PHYS_TO_VM_PAGE(*pdp
& PG_FRAME
);
1692 * Now fill in the pte in the PD. If the pte already exists
1693 * (again, if we raced the grab), unhold pdpg and unwire
1694 * m, returning a held m.
1696 * pdpg is left held (hold_count is incremented for
1697 * each PT in the PD).
1699 pd
= (pd_entry_t
*)PHYS_TO_DMAP(*pdp
& PG_FRAME
);
1700 pd
= &pd
[ptepindex
& ((1ul << NPDEPGSHIFT
) - 1)];
1702 vm_page_unhold(pdpg
);
1703 if (--m
->wire_count
== 0)
1704 --vmstats
.v_wire_count
;
1708 *pd
= VM_PAGE_TO_PHYS(m
) | PG_U
| PG_RW
| PG_V
| PG_A
| PG_M
;
1712 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1713 * valid bits, mapped flag, unbusy, and we're done.
1715 pmap
->pm_ptphint
= m
;
1716 ++pmap
->pm_stats
.resident_count
;
1718 m
->valid
= VM_PAGE_BITS_ALL
;
1719 vm_page_flag_clear(m
, PG_ZERO
);
1720 vm_page_flag_set(m
, PG_MAPPED
);
1728 pmap_allocpte(pmap_t pmap
, vm_offset_t va
)
1730 vm_pindex_t ptepindex
;
1735 * Calculate pagetable page index
1737 ptepindex
= pmap_pde_pindex(va
);
1740 * Get the page directory entry
1742 pd
= pmap_pde(pmap
, va
);
1745 * This supports switching from a 2MB page to a
1748 if (pd
!= NULL
&& (*pd
& (PG_PS
| PG_V
)) == (PG_PS
| PG_V
)) {
1749 panic("no promotion/demotion yet");
1757 * If the page table page is mapped, we just increment the
1758 * hold count, and activate it.
1760 if (pd
!= NULL
&& (*pd
& PG_V
) != 0) {
1761 /* YYY hint is used here on i386 */
1762 m
= pmap_page_lookup( pmap
->pm_pteobj
, ptepindex
);
1763 pmap
->pm_ptphint
= m
;
1768 * Here if the pte page isn't mapped, or if it has been deallocated.
1770 return _pmap_allocpte(pmap
, ptepindex
);
1774 /***************************************************
1775 * Pmap allocation/deallocation routines.
1776 ***************************************************/
1779 * Release any resources held by the given physical map.
1780 * Called when a pmap initialized by pmap_pinit is being released.
1781 * Should only be called if the map contains no valid mappings.
1783 static int pmap_release_callback(struct vm_page
*p
, void *data
);
1786 pmap_release(struct pmap
*pmap
)
1788 vm_object_t object
= pmap
->pm_pteobj
;
1789 struct rb_vm_page_scan_info info
;
1791 KASSERT(pmap
->pm_active
== 0, ("pmap still active! %08x", pmap
->pm_active
));
1792 #if defined(DIAGNOSTIC)
1793 if (object
->ref_count
!= 1)
1794 panic("pmap_release: pteobj reference count != 1");
1798 info
.object
= object
;
1800 lwkt_gettoken(&vm_token
);
1801 TAILQ_REMOVE(&pmap_list
, pmap
, pm_pmnode
);
1808 info
.limit
= object
->generation
;
1810 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, NULL
,
1811 pmap_release_callback
, &info
);
1812 if (info
.error
== 0 && info
.mpte
) {
1813 if (!pmap_release_free_page(pmap
, info
.mpte
))
1817 } while (info
.error
);
1818 lwkt_reltoken(&vm_token
);
1823 pmap_release_callback(struct vm_page
*p
, void *data
)
1825 struct rb_vm_page_scan_info
*info
= data
;
1827 if (p
->pindex
== NUPDE
+ NUPDPE
+ PML4PML4I
) {
1831 if (!pmap_release_free_page(info
->pmap
, p
)) {
1835 if (info
->object
->generation
!= info
->limit
) {
1843 * Grow the number of kernel page table entries, if needed.
1845 * This routine is always called to validate any address space
1846 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
1847 * space below KERNBASE.
1850 pmap_growkernel(vm_offset_t kstart
, vm_offset_t kend
)
1853 vm_offset_t ptppaddr
;
1855 pd_entry_t
*pde
, newpdir
;
1857 int update_kernel_vm_end
;
1860 lwkt_gettoken(&vm_token
);
1863 * bootstrap kernel_vm_end on first real VM use
1865 if (kernel_vm_end
== 0) {
1866 kernel_vm_end
= VM_MIN_KERNEL_ADDRESS
;
1868 while ((*pmap_pde(&kernel_pmap
, kernel_vm_end
) & PG_V
) != 0) {
1869 kernel_vm_end
= (kernel_vm_end
+ PAGE_SIZE
* NPTEPG
) &
1870 ~(PAGE_SIZE
* NPTEPG
- 1);
1872 if (kernel_vm_end
- 1 >= kernel_map
.max_offset
) {
1873 kernel_vm_end
= kernel_map
.max_offset
;
1880 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
1881 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
1882 * do not want to force-fill 128G worth of page tables.
1884 if (kstart
< KERNBASE
) {
1885 if (kstart
> kernel_vm_end
)
1886 kstart
= kernel_vm_end
;
1887 KKASSERT(kend
<= KERNBASE
);
1888 update_kernel_vm_end
= 1;
1890 update_kernel_vm_end
= 0;
1893 kstart
= rounddown2(kstart
, PAGE_SIZE
* NPTEPG
);
1894 kend
= roundup2(kend
, PAGE_SIZE
* NPTEPG
);
1896 if (kend
- 1 >= kernel_map
.max_offset
)
1897 kend
= kernel_map
.max_offset
;
1899 while (kstart
< kend
) {
1900 pde
= pmap_pde(&kernel_pmap
, kstart
);
1902 /* We need a new PDP entry */
1903 nkpg
= vm_page_alloc(kptobj
, nkpt
,
1906 VM_ALLOC_INTERRUPT
);
1908 panic("pmap_growkernel: no memory to grow "
1911 paddr
= VM_PAGE_TO_PHYS(nkpg
);
1912 if ((nkpg
->flags
& PG_ZERO
) == 0)
1913 pmap_zero_page(paddr
);
1914 vm_page_flag_clear(nkpg
, PG_ZERO
);
1915 newpdp
= (pdp_entry_t
)
1916 (paddr
| PG_V
| PG_RW
| PG_A
| PG_M
);
1917 *pmap_pdpe(&kernel_pmap
, kstart
) = newpdp
;
1919 continue; /* try again */
1921 if ((*pde
& PG_V
) != 0) {
1922 kstart
= (kstart
+ PAGE_SIZE
* NPTEPG
) &
1923 ~(PAGE_SIZE
* NPTEPG
- 1);
1924 if (kstart
- 1 >= kernel_map
.max_offset
) {
1925 kstart
= kernel_map
.max_offset
;
1932 * This index is bogus, but out of the way
1934 nkpg
= vm_page_alloc(kptobj
, nkpt
,
1937 VM_ALLOC_INTERRUPT
);
1939 panic("pmap_growkernel: no memory to grow kernel");
1942 ptppaddr
= VM_PAGE_TO_PHYS(nkpg
);
1943 pmap_zero_page(ptppaddr
);
1944 vm_page_flag_clear(nkpg
, PG_ZERO
);
1945 newpdir
= (pd_entry_t
) (ptppaddr
| PG_V
| PG_RW
| PG_A
| PG_M
);
1946 *pmap_pde(&kernel_pmap
, kstart
) = newpdir
;
1949 kstart
= (kstart
+ PAGE_SIZE
* NPTEPG
) &
1950 ~(PAGE_SIZE
* NPTEPG
- 1);
1952 if (kstart
- 1 >= kernel_map
.max_offset
) {
1953 kstart
= kernel_map
.max_offset
;
1959 * Only update kernel_vm_end for areas below KERNBASE.
1961 if (update_kernel_vm_end
&& kernel_vm_end
< kstart
)
1962 kernel_vm_end
= kstart
;
1964 lwkt_reltoken(&vm_token
);
1969 * Retire the given physical map from service.
1970 * Should only be called if the map contains
1971 * no valid mappings.
1974 pmap_destroy(pmap_t pmap
)
1981 lwkt_gettoken(&vm_token
);
1982 count
= --pmap
->pm_count
;
1985 panic("destroying a pmap is not yet implemented");
1987 lwkt_reltoken(&vm_token
);
1991 * Add a reference to the specified pmap.
1994 pmap_reference(pmap_t pmap
)
1997 lwkt_gettoken(&vm_token
);
1999 lwkt_reltoken(&vm_token
);
2003 /***************************************************
2004 * page management routines.
2005 ***************************************************/
2008 * free the pv_entry back to the free list. This function may be
2009 * called from an interrupt.
2013 free_pv_entry(pv_entry_t pv
)
2016 KKASSERT(pv_entry_count
>= 0);
2021 * get a new pv_entry, allocating a block from the system
2022 * when needed. This function may be called from an interrupt.
2029 if (pv_entry_high_water
&&
2030 (pv_entry_count
> pv_entry_high_water
) &&
2031 (pmap_pagedaemon_waken
== 0)) {
2032 pmap_pagedaemon_waken
= 1;
2033 wakeup(&vm_pages_needed
);
2035 return zalloc(pvzone
);
2039 * This routine is very drastic, but can save the system
2047 static int warningdone
=0;
2049 if (pmap_pagedaemon_waken
== 0)
2051 lwkt_gettoken(&vm_token
);
2052 if (warningdone
< 5) {
2053 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
2057 for(i
= 0; i
< vm_page_array_size
; i
++) {
2058 m
= &vm_page_array
[i
];
2059 if (m
->wire_count
|| m
->hold_count
|| m
->busy
||
2060 (m
->flags
& PG_BUSY
))
2064 pmap_pagedaemon_waken
= 0;
2065 lwkt_reltoken(&vm_token
);
2070 * If it is the first entry on the list, it is actually
2071 * in the header and we must copy the following entry up
2072 * to the header. Otherwise we must search the list for
2073 * the entry. In either case we free the now unused entry.
2077 pmap_remove_entry(struct pmap
*pmap
, vm_page_t m
,
2078 vm_offset_t va
, pmap_inval_info_t info
)
2084 if (m
->md
.pv_list_count
< pmap
->pm_stats
.resident_count
) {
2085 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
2086 if (pmap
== pv
->pv_pmap
&& va
== pv
->pv_va
)
2090 TAILQ_FOREACH(pv
, &pmap
->pm_pvlist
, pv_plist
) {
2091 if (va
== pv
->pv_va
)
2099 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
2100 m
->md
.pv_list_count
--;
2101 m
->object
->agg_pv_list_count
--;
2102 KKASSERT(m
->md
.pv_list_count
>= 0);
2103 if (TAILQ_EMPTY(&m
->md
.pv_list
))
2104 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
2105 TAILQ_REMOVE(&pmap
->pm_pvlist
, pv
, pv_plist
);
2106 ++pmap
->pm_generation
;
2107 rtval
= pmap_unuse_pt(pmap
, va
, pv
->pv_ptem
, info
);
2115 * Create a pv entry for page at pa for
2120 pmap_insert_entry(pmap_t pmap
, vm_offset_t va
, vm_page_t mpte
, vm_page_t m
)
2125 pv
= get_pv_entry();
2130 TAILQ_INSERT_TAIL(&pmap
->pm_pvlist
, pv
, pv_plist
);
2131 TAILQ_INSERT_TAIL(&m
->md
.pv_list
, pv
, pv_list
);
2132 ++pmap
->pm_generation
;
2133 m
->md
.pv_list_count
++;
2134 m
->object
->agg_pv_list_count
++;
2140 * pmap_remove_pte: do the things to unmap a page in a process
2144 pmap_remove_pte(struct pmap
*pmap
, pt_entry_t
*ptq
, vm_offset_t va
,
2145 pmap_inval_info_t info
)
2150 pmap_inval_interlock(info
, pmap
, va
);
2151 oldpte
= pte_load_clear(ptq
);
2152 pmap_inval_deinterlock(info
, pmap
);
2154 pmap
->pm_stats
.wired_count
-= 1;
2156 * Machines that don't support invlpg, also don't support
2157 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2161 cpu_invlpg((void *)va
);
2162 KKASSERT(pmap
->pm_stats
.resident_count
> 0);
2163 --pmap
->pm_stats
.resident_count
;
2164 if (oldpte
& PG_MANAGED
) {
2165 m
= PHYS_TO_VM_PAGE(oldpte
);
2166 if (oldpte
& PG_M
) {
2167 #if defined(PMAP_DIAGNOSTIC)
2168 if (pmap_nw_modified((pt_entry_t
) oldpte
)) {
2170 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2174 if (pmap_track_modified(va
))
2178 vm_page_flag_set(m
, PG_REFERENCED
);
2179 return pmap_remove_entry(pmap
, m
, va
, info
);
2181 return pmap_unuse_pt(pmap
, va
, NULL
, info
);
2190 * Remove a single page from a process address space.
2192 * This function may not be called from an interrupt if the pmap is
2197 pmap_remove_page(struct pmap
*pmap
, vm_offset_t va
, pmap_inval_info_t info
)
2201 pte
= pmap_pte(pmap
, va
);
2204 if ((*pte
& PG_V
) == 0)
2206 pmap_remove_pte(pmap
, pte
, va
, info
);
2212 * Remove the given range of addresses from the specified map.
2214 * It is assumed that the start and end are properly
2215 * rounded to the page size.
2217 * This function may not be called from an interrupt if the pmap is
2221 pmap_remove(struct pmap
*pmap
, vm_offset_t sva
, vm_offset_t eva
)
2223 vm_offset_t va_next
;
2224 pml4_entry_t
*pml4e
;
2226 pd_entry_t ptpaddr
, *pde
;
2228 struct pmap_inval_info info
;
2233 lwkt_gettoken(&vm_token
);
2234 if (pmap
->pm_stats
.resident_count
== 0) {
2235 lwkt_reltoken(&vm_token
);
2239 pmap_inval_init(&info
);
2242 * special handling of removing one page. a very
2243 * common operation and easy to short circuit some
2246 if (sva
+ PAGE_SIZE
== eva
) {
2247 pde
= pmap_pde(pmap
, sva
);
2248 if (pde
&& (*pde
& PG_PS
) == 0) {
2249 pmap_remove_page(pmap
, sva
, &info
);
2250 pmap_inval_done(&info
);
2251 lwkt_reltoken(&vm_token
);
2256 for (; sva
< eva
; sva
= va_next
) {
2257 pml4e
= pmap_pml4e(pmap
, sva
);
2258 if ((*pml4e
& PG_V
) == 0) {
2259 va_next
= (sva
+ NBPML4
) & ~PML4MASK
;
2265 pdpe
= pmap_pml4e_to_pdpe(pml4e
, sva
);
2266 if ((*pdpe
& PG_V
) == 0) {
2267 va_next
= (sva
+ NBPDP
) & ~PDPMASK
;
2274 * Calculate index for next page table.
2276 va_next
= (sva
+ NBPDR
) & ~PDRMASK
;
2280 pde
= pmap_pdpe_to_pde(pdpe
, sva
);
2284 * Weed out invalid mappings.
2290 * Check for large page.
2292 if ((ptpaddr
& PG_PS
) != 0) {
2293 /* JG FreeBSD has more complex treatment here */
2294 pmap_inval_interlock(&info
, pmap
, -1);
2296 pmap
->pm_stats
.resident_count
-= NBPDR
/ PAGE_SIZE
;
2297 pmap_inval_deinterlock(&info
, pmap
);
2302 * Limit our scan to either the end of the va represented
2303 * by the current page table page, or to the end of the
2304 * range being removed.
2310 * NOTE: pmap_remove_pte() can block.
2312 for (pte
= pmap_pde_to_pte(pde
, sva
); sva
!= va_next
; pte
++,
2316 if (pmap_remove_pte(pmap
, pte
, sva
, &info
))
2320 pmap_inval_done(&info
);
2321 lwkt_reltoken(&vm_token
);
2327 * Removes this physical page from all physical maps in which it resides.
2328 * Reflects back modify bits to the pager.
2330 * This routine may not be called from an interrupt.
2335 pmap_remove_all(vm_page_t m
)
2337 struct pmap_inval_info info
;
2338 pt_entry_t
*pte
, tpte
;
2341 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
2344 lwkt_gettoken(&vm_token
);
2345 pmap_inval_init(&info
);
2347 while ((pv
= TAILQ_FIRST(&m
->md
.pv_list
)) != NULL
) {
2348 KKASSERT(pv
->pv_pmap
->pm_stats
.resident_count
> 0);
2349 --pv
->pv_pmap
->pm_stats
.resident_count
;
2351 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_va
);
2352 pmap_inval_interlock(&info
, pv
->pv_pmap
, pv
->pv_va
);
2353 tpte
= pte_load_clear(pte
);
2355 pv
->pv_pmap
->pm_stats
.wired_count
--;
2356 pmap_inval_deinterlock(&info
, pv
->pv_pmap
);
2358 vm_page_flag_set(m
, PG_REFERENCED
);
2361 * Update the vm_page_t clean and reference bits.
2364 #if defined(PMAP_DIAGNOSTIC)
2365 if (pmap_nw_modified(tpte
)) {
2367 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2371 if (pmap_track_modified(pv
->pv_va
))
2374 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
2375 TAILQ_REMOVE(&pv
->pv_pmap
->pm_pvlist
, pv
, pv_plist
);
2376 ++pv
->pv_pmap
->pm_generation
;
2377 m
->md
.pv_list_count
--;
2378 m
->object
->agg_pv_list_count
--;
2379 KKASSERT(m
->md
.pv_list_count
>= 0);
2380 if (TAILQ_EMPTY(&m
->md
.pv_list
))
2381 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
2382 pmap_unuse_pt(pv
->pv_pmap
, pv
->pv_va
, pv
->pv_ptem
, &info
);
2386 KKASSERT((m
->flags
& (PG_MAPPED
|PG_WRITEABLE
)) == 0);
2387 pmap_inval_done(&info
);
2388 lwkt_reltoken(&vm_token
);
2394 * Set the physical protection on the specified range of this map
2397 * This function may not be called from an interrupt if the map is
2398 * not the kernel_pmap.
2401 pmap_protect(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
, vm_prot_t prot
)
2403 vm_offset_t va_next
;
2404 pml4_entry_t
*pml4e
;
2406 pd_entry_t ptpaddr
, *pde
;
2408 pmap_inval_info info
;
2410 /* JG review for NX */
2415 if ((prot
& VM_PROT_READ
) == VM_PROT_NONE
) {
2416 pmap_remove(pmap
, sva
, eva
);
2420 if (prot
& VM_PROT_WRITE
)
2423 lwkt_gettoken(&vm_token
);
2424 pmap_inval_init(&info
);
2426 for (; sva
< eva
; sva
= va_next
) {
2428 pml4e
= pmap_pml4e(pmap
, sva
);
2429 if ((*pml4e
& PG_V
) == 0) {
2430 va_next
= (sva
+ NBPML4
) & ~PML4MASK
;
2436 pdpe
= pmap_pml4e_to_pdpe(pml4e
, sva
);
2437 if ((*pdpe
& PG_V
) == 0) {
2438 va_next
= (sva
+ NBPDP
) & ~PDPMASK
;
2444 va_next
= (sva
+ NBPDR
) & ~PDRMASK
;
2448 pde
= pmap_pdpe_to_pde(pdpe
, sva
);
2452 * Check for large page.
2454 if ((ptpaddr
& PG_PS
) != 0) {
2455 pmap_inval_interlock(&info
, pmap
, -1);
2456 *pde
&= ~(PG_M
|PG_RW
);
2457 pmap
->pm_stats
.resident_count
-= NBPDR
/ PAGE_SIZE
;
2458 pmap_inval_deinterlock(&info
, pmap
);
2463 * Weed out invalid mappings. Note: we assume that the page
2464 * directory table is always allocated, and in kernel virtual.
2472 for (pte
= pmap_pde_to_pte(pde
, sva
); sva
!= va_next
; pte
++,
2479 * XXX non-optimal. Note also that there can be
2480 * no pmap_inval_flush() calls until after we modify
2481 * ptbase[sindex] (or otherwise we have to do another
2482 * pmap_inval_add() call).
2484 pmap_inval_interlock(&info
, pmap
, sva
);
2488 if ((pbits
& PG_V
) == 0) {
2489 pmap_inval_deinterlock(&info
, pmap
);
2492 if (pbits
& PG_MANAGED
) {
2495 m
= PHYS_TO_VM_PAGE(pbits
& PG_FRAME
);
2496 vm_page_flag_set(m
, PG_REFERENCED
);
2500 if (pmap_track_modified(sva
)) {
2502 m
= PHYS_TO_VM_PAGE(pbits
& PG_FRAME
);
2509 if (pbits
!= cbits
&&
2510 !atomic_cmpset_long(pte
, pbits
, cbits
)) {
2513 pmap_inval_deinterlock(&info
, pmap
);
2516 pmap_inval_done(&info
);
2517 lwkt_reltoken(&vm_token
);
2521 * Insert the given physical page (p) at
2522 * the specified virtual address (v) in the
2523 * target physical map with the protection requested.
2525 * If specified, the page will be wired down, meaning
2526 * that the related pte can not be reclaimed.
2528 * NB: This is the only routine which MAY NOT lazy-evaluate
2529 * or lose information. That is, this routine must actually
2530 * insert this page into the given map NOW.
2533 pmap_enter(pmap_t pmap
, vm_offset_t va
, vm_page_t m
, vm_prot_t prot
,
2540 pt_entry_t origpte
, newpte
;
2542 pmap_inval_info info
;
2547 va
= trunc_page(va
);
2548 #ifdef PMAP_DIAGNOSTIC
2550 panic("pmap_enter: toobig");
2551 if ((va
>= UPT_MIN_ADDRESS
) && (va
< UPT_MAX_ADDRESS
))
2552 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va
);
2554 if (va
< UPT_MAX_ADDRESS
&& pmap
== &kernel_pmap
) {
2555 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2557 db_print_backtrace();
2560 if (va
>= UPT_MAX_ADDRESS
&& pmap
!= &kernel_pmap
) {
2561 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2563 db_print_backtrace();
2567 lwkt_gettoken(&vm_token
);
2570 * In the case that a page table page is not
2571 * resident, we are creating it here.
2573 if (va
< VM_MAX_USER_ADDRESS
)
2574 mpte
= pmap_allocpte(pmap
, va
);
2578 pmap_inval_init(&info
);
2579 pde
= pmap_pde(pmap
, va
);
2580 if (pde
!= NULL
&& (*pde
& PG_V
) != 0) {
2581 if ((*pde
& PG_PS
) != 0)
2582 panic("pmap_enter: attempted pmap_enter on 2MB page");
2583 pte
= pmap_pde_to_pte(pde
, va
);
2585 panic("pmap_enter: invalid page directory va=%#lx", va
);
2587 KKASSERT(pte
!= NULL
);
2588 pa
= VM_PAGE_TO_PHYS(m
);
2590 opa
= origpte
& PG_FRAME
;
2593 * Mapping has not changed, must be protection or wiring change.
2595 if (origpte
&& (opa
== pa
)) {
2597 * Wiring change, just update stats. We don't worry about
2598 * wiring PT pages as they remain resident as long as there
2599 * are valid mappings in them. Hence, if a user page is wired,
2600 * the PT page will be also.
2602 if (wired
&& ((origpte
& PG_W
) == 0))
2603 pmap
->pm_stats
.wired_count
++;
2604 else if (!wired
&& (origpte
& PG_W
))
2605 pmap
->pm_stats
.wired_count
--;
2607 #if defined(PMAP_DIAGNOSTIC)
2608 if (pmap_nw_modified(origpte
)) {
2610 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2616 * Remove the extra pte reference. Note that we cannot
2617 * optimize the RO->RW case because we have adjusted the
2618 * wiring count above and may need to adjust the wiring
2625 * We might be turning off write access to the page,
2626 * so we go ahead and sense modify status.
2628 if (origpte
& PG_MANAGED
) {
2629 if ((origpte
& PG_M
) && pmap_track_modified(va
)) {
2631 om
= PHYS_TO_VM_PAGE(opa
);
2635 KKASSERT(m
->flags
& PG_MAPPED
);
2640 * Mapping has changed, invalidate old range and fall through to
2641 * handle validating new mapping.
2645 err
= pmap_remove_pte(pmap
, pte
, va
, &info
);
2647 panic("pmap_enter: pte vanished, va: 0x%lx", va
);
2649 opa
= origpte
& PG_FRAME
;
2651 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2657 * Enter on the PV list if part of our managed memory. Note that we
2658 * raise IPL while manipulating pv_table since pmap_enter can be
2659 * called at interrupt time.
2661 if (pmap_initialized
&&
2662 (m
->flags
& (PG_FICTITIOUS
|PG_UNMANAGED
)) == 0) {
2663 pmap_insert_entry(pmap
, va
, mpte
, m
);
2665 vm_page_flag_set(m
, PG_MAPPED
);
2669 * Increment counters
2671 ++pmap
->pm_stats
.resident_count
;
2673 pmap
->pm_stats
.wired_count
++;
2677 * Now validate mapping with desired protection/wiring.
2679 newpte
= (pt_entry_t
) (pa
| pte_prot(pmap
, prot
) | PG_V
);
2683 if (va
< VM_MAX_USER_ADDRESS
)
2685 if (pmap
== &kernel_pmap
)
2689 * if the mapping or permission bits are different, we need
2690 * to update the pte.
2692 if ((origpte
& ~(PG_M
|PG_A
)) != newpte
) {
2693 pmap_inval_interlock(&info
, pmap
, va
);
2694 *pte
= newpte
| PG_A
;
2695 pmap_inval_deinterlock(&info
, pmap
);
2697 vm_page_flag_set(m
, PG_WRITEABLE
);
2699 KKASSERT((newpte
& PG_MANAGED
) == 0 || (m
->flags
& PG_MAPPED
));
2700 pmap_inval_done(&info
);
2701 lwkt_reltoken(&vm_token
);
2705 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2706 * This code also assumes that the pmap has no pre-existing entry for this
2709 * This code currently may only be used on user pmaps, not kernel_pmap.
2712 pmap_enter_quick(pmap_t pmap
, vm_offset_t va
, vm_page_t m
)
2717 vm_pindex_t ptepindex
;
2719 pmap_inval_info info
;
2721 lwkt_gettoken(&vm_token
);
2722 pmap_inval_init(&info
);
2724 if (va
< UPT_MAX_ADDRESS
&& pmap
== &kernel_pmap
) {
2725 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2727 db_print_backtrace();
2730 if (va
>= UPT_MAX_ADDRESS
&& pmap
!= &kernel_pmap
) {
2731 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2733 db_print_backtrace();
2737 KKASSERT(va
< UPT_MIN_ADDRESS
); /* assert used on user pmaps only */
2740 * Calculate the page table page (mpte), allocating it if necessary.
2742 * A held page table page (mpte), or NULL, is passed onto the
2743 * section following.
2745 if (va
< VM_MAX_USER_ADDRESS
) {
2747 * Calculate pagetable page index
2749 ptepindex
= pmap_pde_pindex(va
);
2753 * Get the page directory entry
2755 ptepa
= pmap_pde(pmap
, va
);
2758 * If the page table page is mapped, we just increment
2759 * the hold count, and activate it.
2761 if (ptepa
&& (*ptepa
& PG_V
) != 0) {
2763 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2764 // if (pmap->pm_ptphint &&
2765 // (pmap->pm_ptphint->pindex == ptepindex)) {
2766 // mpte = pmap->pm_ptphint;
2768 mpte
= pmap_page_lookup( pmap
->pm_pteobj
, ptepindex
);
2769 pmap
->pm_ptphint
= mpte
;
2774 mpte
= _pmap_allocpte(pmap
, ptepindex
);
2776 } while (mpte
== NULL
);
2779 /* this code path is not yet used */
2783 * With a valid (and held) page directory page, we can just use
2784 * vtopte() to get to the pte. If the pte is already present
2785 * we do not disturb it.
2790 pmap_unwire_pte_hold(pmap
, va
, mpte
, &info
);
2791 pa
= VM_PAGE_TO_PHYS(m
);
2792 KKASSERT(((*pte
^ pa
) & PG_FRAME
) == 0);
2793 pmap_inval_done(&info
);
2794 lwkt_reltoken(&vm_token
);
2799 * Enter on the PV list if part of our managed memory
2801 if ((m
->flags
& (PG_FICTITIOUS
|PG_UNMANAGED
)) == 0) {
2802 pmap_insert_entry(pmap
, va
, mpte
, m
);
2803 vm_page_flag_set(m
, PG_MAPPED
);
2807 * Increment counters
2809 ++pmap
->pm_stats
.resident_count
;
2811 pa
= VM_PAGE_TO_PHYS(m
);
2814 * Now validate mapping with RO protection
2816 if (m
->flags
& (PG_FICTITIOUS
|PG_UNMANAGED
))
2817 *pte
= pa
| PG_V
| PG_U
;
2819 *pte
= pa
| PG_V
| PG_U
| PG_MANAGED
;
2820 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2821 pmap_inval_done(&info
);
2822 lwkt_reltoken(&vm_token
);
2826 * Make a temporary mapping for a physical address. This is only intended
2827 * to be used for panic dumps.
2829 /* JG Needed on x86_64? */
2831 pmap_kenter_temporary(vm_paddr_t pa
, int i
)
2833 pmap_kenter((vm_offset_t
)crashdumpmap
+ (i
* PAGE_SIZE
), pa
);
2834 return ((void *)crashdumpmap
);
2837 #define MAX_INIT_PT (96)
2840 * This routine preloads the ptes for a given object into the specified pmap.
2841 * This eliminates the blast of soft faults on process startup and
2842 * immediately after an mmap.
2844 static int pmap_object_init_pt_callback(vm_page_t p
, void *data
);
2847 pmap_object_init_pt(pmap_t pmap
, vm_offset_t addr
, vm_prot_t prot
,
2848 vm_object_t object
, vm_pindex_t pindex
,
2849 vm_size_t size
, int limit
)
2851 struct rb_vm_page_scan_info info
;
2856 * We can't preinit if read access isn't set or there is no pmap
2859 if ((prot
& VM_PROT_READ
) == 0 || pmap
== NULL
|| object
== NULL
)
2863 * We can't preinit if the pmap is not the current pmap
2865 lp
= curthread
->td_lwp
;
2866 if (lp
== NULL
|| pmap
!= vmspace_pmap(lp
->lwp_vmspace
))
2869 psize
= x86_64_btop(size
);
2871 if ((object
->type
!= OBJT_VNODE
) ||
2872 ((limit
& MAP_PREFAULT_PARTIAL
) && (psize
> MAX_INIT_PT
) &&
2873 (object
->resident_page_count
> MAX_INIT_PT
))) {
2877 if (psize
+ pindex
> object
->size
) {
2878 if (object
->size
< pindex
)
2880 psize
= object
->size
- pindex
;
2887 * Use a red-black scan to traverse the requested range and load
2888 * any valid pages found into the pmap.
2890 * We cannot safely scan the object's memq unless we are in a
2891 * critical section since interrupts can remove pages from objects.
2893 info
.start_pindex
= pindex
;
2894 info
.end_pindex
= pindex
+ psize
- 1;
2901 lwkt_gettoken(&vm_token
);
2902 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, rb_vm_page_scancmp
,
2903 pmap_object_init_pt_callback
, &info
);
2904 lwkt_reltoken(&vm_token
);
2910 pmap_object_init_pt_callback(vm_page_t p
, void *data
)
2912 struct rb_vm_page_scan_info
*info
= data
;
2913 vm_pindex_t rel_index
;
2915 * don't allow an madvise to blow away our really
2916 * free pages allocating pv entries.
2918 if ((info
->limit
& MAP_PREFAULT_MADVISE
) &&
2919 vmstats
.v_free_count
< vmstats
.v_free_reserved
) {
2922 if (((p
->valid
& VM_PAGE_BITS_ALL
) == VM_PAGE_BITS_ALL
) &&
2923 (p
->busy
== 0) && (p
->flags
& (PG_BUSY
| PG_FICTITIOUS
)) == 0) {
2924 if ((p
->queue
- p
->pc
) == PQ_CACHE
)
2925 vm_page_deactivate(p
);
2927 rel_index
= p
->pindex
- info
->start_pindex
;
2928 pmap_enter_quick(info
->pmap
,
2929 info
->addr
+ x86_64_ptob(rel_index
), p
);
2936 * Return TRUE if the pmap is in shape to trivially
2937 * pre-fault the specified address.
2939 * Returns FALSE if it would be non-trivial or if a
2940 * pte is already loaded into the slot.
2943 pmap_prefault_ok(pmap_t pmap
, vm_offset_t addr
)
2949 lwkt_gettoken(&vm_token
);
2950 pde
= pmap_pde(pmap
, addr
);
2951 if (pde
== NULL
|| *pde
== 0) {
2955 ret
= (*pte
) ? 0 : 1;
2957 lwkt_reltoken(&vm_token
);
2962 * Routine: pmap_change_wiring
2963 * Function: Change the wiring attribute for a map/virtual-address
2965 * In/out conditions:
2966 * The mapping must already exist in the pmap.
2969 pmap_change_wiring(pmap_t pmap
, vm_offset_t va
, boolean_t wired
)
2976 lwkt_gettoken(&vm_token
);
2977 pte
= pmap_pte(pmap
, va
);
2979 if (wired
&& !pmap_pte_w(pte
))
2980 pmap
->pm_stats
.wired_count
++;
2981 else if (!wired
&& pmap_pte_w(pte
))
2982 pmap
->pm_stats
.wired_count
--;
2985 * Wiring is not a hardware characteristic so there is no need to
2986 * invalidate TLB. However, in an SMP environment we must use
2987 * a locked bus cycle to update the pte (if we are not using
2988 * the pmap_inval_*() API that is)... it's ok to do this for simple
2993 atomic_set_long(pte
, PG_W
);
2995 atomic_clear_long(pte
, PG_W
);
2998 atomic_set_long_nonlocked(pte
, PG_W
);
3000 atomic_clear_long_nonlocked(pte
, PG_W
);
3002 lwkt_reltoken(&vm_token
);
3008 * Copy the range specified by src_addr/len
3009 * from the source map to the range dst_addr/len
3010 * in the destination map.
3012 * This routine is only advisory and need not do anything.
3015 pmap_copy(pmap_t dst_pmap
, pmap_t src_pmap
, vm_offset_t dst_addr
,
3016 vm_size_t len
, vm_offset_t src_addr
)
3020 pmap_inval_info info
;
3022 vm_offset_t end_addr
= src_addr
+ len
;
3024 pd_entry_t src_frame
, dst_frame
;
3027 if (dst_addr
!= src_addr
)
3030 src_frame
= src_pmap
->pm_pdir
[PTDPTDI
] & PG_FRAME
;
3031 if (src_frame
!= (PTDpde
& PG_FRAME
)) {
3035 dst_frame
= dst_pmap
->pm_pdir
[PTDPTDI
] & PG_FRAME
;
3036 if (dst_frame
!= (APTDpde
& PG_FRAME
)) {
3037 APTDpde
= (pd_entry_t
) (dst_frame
| PG_RW
| PG_V
);
3038 /* The page directory is not shared between CPUs */
3042 pmap_inval_init(&info
);
3043 pmap_inval_add(&info
, dst_pmap
, -1);
3044 pmap_inval_add(&info
, src_pmap
, -1);
3047 * critical section protection is required to maintain the page/object
3048 * association, interrupts can free pages and remove them from
3052 for (addr
= src_addr
; addr
< end_addr
; addr
= pdnxt
) {
3053 pt_entry_t
*src_pte
, *dst_pte
;
3054 vm_page_t dstmpte
, srcmpte
;
3055 vm_offset_t srcptepaddr
;
3056 vm_pindex_t ptepindex
;
3058 if (addr
>= UPT_MIN_ADDRESS
)
3059 panic("pmap_copy: invalid to pmap_copy page tables\n");
3062 * Don't let optional prefaulting of pages make us go
3063 * way below the low water mark of free pages or way
3064 * above high water mark of used pv entries.
3066 if (vmstats
.v_free_count
< vmstats
.v_free_reserved
||
3067 pv_entry_count
> pv_entry_high_water
)
3070 pdnxt
= ((addr
+ PAGE_SIZE
*NPTEPG
) & ~(PAGE_SIZE
*NPTEPG
- 1));
3071 ptepindex
= addr
>> PDRSHIFT
;
3074 srcptepaddr
= (vm_offset_t
) src_pmap
->pm_pdir
[ptepindex
];
3076 if (srcptepaddr
== 0)
3079 if (srcptepaddr
& PG_PS
) {
3081 if (dst_pmap
->pm_pdir
[ptepindex
] == 0) {
3082 dst_pmap
->pm_pdir
[ptepindex
] = (pd_entry_t
) srcptepaddr
;
3083 dst_pmap
->pm_stats
.resident_count
+= NBPDR
/ PAGE_SIZE
;
3089 srcmpte
= vm_page_lookup(src_pmap
->pm_pteobj
, ptepindex
);
3090 if ((srcmpte
== NULL
) || (srcmpte
->hold_count
== 0) ||
3091 (srcmpte
->flags
& PG_BUSY
)) {
3095 if (pdnxt
> end_addr
)
3098 src_pte
= vtopte(addr
);
3100 dst_pte
= avtopte(addr
);
3102 while (addr
< pdnxt
) {
3107 * we only virtual copy managed pages
3109 if ((ptetemp
& PG_MANAGED
) != 0) {
3111 * We have to check after allocpte for the
3112 * pte still being around... allocpte can
3115 * pmap_allocpte() can block. If we lose
3116 * our page directory mappings we stop.
3118 dstmpte
= pmap_allocpte(dst_pmap
, addr
);
3121 if (src_frame
!= (PTDpde
& PG_FRAME
) ||
3122 dst_frame
!= (APTDpde
& PG_FRAME
)
3124 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3125 pmap_unwire_pte_hold(dst_pmap
, dstmpte
, &info
);
3127 } else if ((*dst_pte
== 0) &&
3128 (ptetemp
= *src_pte
) != 0 &&
3129 (ptetemp
& PG_MANAGED
)) {
3131 * Clear the modified and
3132 * accessed (referenced) bits
3135 m
= PHYS_TO_VM_PAGE(ptetemp
);
3136 *dst_pte
= ptetemp
& ~(PG_M
| PG_A
);
3137 ++dst_pmap
->pm_stats
.resident_count
;
3138 pmap_insert_entry(dst_pmap
, addr
,
3140 KKASSERT(m
->flags
& PG_MAPPED
);
3142 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3143 pmap_unwire_pte_hold(dst_pmap
, dstmpte
, &info
);
3147 if (dstmpte
->hold_count
>= srcmpte
->hold_count
)
3157 pmap_inval_done(&info
);
3164 * Zero the specified physical page.
3166 * This function may be called from an interrupt and no locking is
3170 pmap_zero_page(vm_paddr_t phys
)
3172 vm_offset_t va
= PHYS_TO_DMAP(phys
);
3174 pagezero((void *)va
);
3178 * pmap_page_assertzero:
3180 * Assert that a page is empty, panic if it isn't.
3183 pmap_page_assertzero(vm_paddr_t phys
)
3185 vm_offset_t virt
= PHYS_TO_DMAP(phys
);
3188 for (i
= 0; i
< PAGE_SIZE
; i
+= sizeof(long)) {
3189 if (*(long *)((char *)virt
+ i
) != 0) {
3190 panic("pmap_page_assertzero() @ %p not zero!\n", (void *)virt
);
3198 * Zero part of a physical page by mapping it into memory and clearing
3199 * its contents with bzero.
3201 * off and size may not cover an area beyond a single hardware page.
3204 pmap_zero_page_area(vm_paddr_t phys
, int off
, int size
)
3206 vm_offset_t virt
= PHYS_TO_DMAP(phys
);
3208 bzero((char *)virt
+ off
, size
);
3214 * Copy the physical page from the source PA to the target PA.
3215 * This function may be called from an interrupt. No locking
3219 pmap_copy_page(vm_paddr_t src
, vm_paddr_t dst
)
3221 vm_offset_t src_virt
, dst_virt
;
3223 src_virt
= PHYS_TO_DMAP(src
);
3224 dst_virt
= PHYS_TO_DMAP(dst
);
3225 bcopy((void *)src_virt
, (void *)dst_virt
, PAGE_SIZE
);
3229 * pmap_copy_page_frag:
3231 * Copy the physical page from the source PA to the target PA.
3232 * This function may be called from an interrupt. No locking
3236 pmap_copy_page_frag(vm_paddr_t src
, vm_paddr_t dst
, size_t bytes
)
3238 vm_offset_t src_virt
, dst_virt
;
3240 src_virt
= PHYS_TO_DMAP(src
);
3241 dst_virt
= PHYS_TO_DMAP(dst
);
3243 bcopy((char *)src_virt
+ (src
& PAGE_MASK
),
3244 (char *)dst_virt
+ (dst
& PAGE_MASK
),
3249 * Returns true if the pmap's pv is one of the first
3250 * 16 pvs linked to from this page. This count may
3251 * be changed upwards or downwards in the future; it
3252 * is only necessary that true be returned for a small
3253 * subset of pmaps for proper page aging.
3256 pmap_page_exists_quick(pmap_t pmap
, vm_page_t m
)
3261 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
3265 lwkt_gettoken(&vm_token
);
3267 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
3268 if (pv
->pv_pmap
== pmap
) {
3269 lwkt_reltoken(&vm_token
);
3277 lwkt_reltoken(&vm_token
);
3283 * Remove all pages from specified address space
3284 * this aids process exit speeds. Also, this code
3285 * is special cased for current process only, but
3286 * can have the more generic (and slightly slower)
3287 * mode enabled. This is much faster than pmap_remove
3288 * in the case of running down an entire address space.
3291 pmap_remove_pages(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
)
3294 pt_entry_t
*pte
, tpte
;
3297 pmap_inval_info info
;
3299 int save_generation
;
3301 lp
= curthread
->td_lwp
;
3302 if (lp
&& pmap
== vmspace_pmap(lp
->lwp_vmspace
))
3307 lwkt_gettoken(&vm_token
);
3308 pmap_inval_init(&info
);
3309 for (pv
= TAILQ_FIRST(&pmap
->pm_pvlist
); pv
; pv
= npv
) {
3310 if (pv
->pv_va
>= eva
|| pv
->pv_va
< sva
) {
3311 npv
= TAILQ_NEXT(pv
, pv_plist
);
3315 KKASSERT(pmap
== pv
->pv_pmap
);
3318 pte
= vtopte(pv
->pv_va
);
3320 pte
= pmap_pte_quick(pmap
, pv
->pv_va
);
3321 pmap_inval_interlock(&info
, pmap
, pv
->pv_va
);
3324 * We cannot remove wired pages from a process' mapping
3328 pmap_inval_deinterlock(&info
, pmap
);
3329 npv
= TAILQ_NEXT(pv
, pv_plist
);
3332 tpte
= pte_load_clear(pte
);
3334 m
= PHYS_TO_VM_PAGE(tpte
& PG_FRAME
);
3336 KASSERT(m
< &vm_page_array
[vm_page_array_size
],
3337 ("pmap_remove_pages: bad tpte %lx", tpte
));
3339 KKASSERT(pmap
->pm_stats
.resident_count
> 0);
3340 --pmap
->pm_stats
.resident_count
;
3341 pmap_inval_deinterlock(&info
, pmap
);
3344 * Update the vm_page_t clean and reference bits.
3350 npv
= TAILQ_NEXT(pv
, pv_plist
);
3351 TAILQ_REMOVE(&pmap
->pm_pvlist
, pv
, pv_plist
);
3352 save_generation
= ++pmap
->pm_generation
;
3354 m
->md
.pv_list_count
--;
3355 m
->object
->agg_pv_list_count
--;
3356 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
3357 if (TAILQ_EMPTY(&m
->md
.pv_list
))
3358 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
3360 pmap_unuse_pt(pmap
, pv
->pv_va
, pv
->pv_ptem
, &info
);
3364 * Restart the scan if we blocked during the unuse or free
3365 * calls and other removals were made.
3367 if (save_generation
!= pmap
->pm_generation
) {
3368 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3369 npv
= TAILQ_FIRST(&pmap
->pm_pvlist
);
3372 pmap_inval_done(&info
);
3373 lwkt_reltoken(&vm_token
);
3377 * pmap_testbit tests bits in pte's
3378 * note that the testbit/clearbit routines are inline,
3379 * and a lot of things compile-time evaluate.
3383 pmap_testbit(vm_page_t m
, int bit
)
3388 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
3391 if (TAILQ_FIRST(&m
->md
.pv_list
) == NULL
)
3396 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
3398 * if the bit being tested is the modified bit, then
3399 * mark clean_map and ptes as never
3402 if (bit
& (PG_A
|PG_M
)) {
3403 if (!pmap_track_modified(pv
->pv_va
))
3407 #if defined(PMAP_DIAGNOSTIC)
3408 if (pv
->pv_pmap
== NULL
) {
3409 kprintf("Null pmap (tb) at va: 0x%lx\n", pv
->pv_va
);
3413 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_va
);
3424 * this routine is used to modify bits in ptes
3428 pmap_clearbit(vm_page_t m
, int bit
)
3430 struct pmap_inval_info info
;
3435 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
3438 pmap_inval_init(&info
);
3441 * Loop over all current mappings setting/clearing as appropos If
3442 * setting RO do we need to clear the VAC?
3444 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
3446 * don't write protect pager mappings
3449 if (!pmap_track_modified(pv
->pv_va
))
3453 #if defined(PMAP_DIAGNOSTIC)
3454 if (pv
->pv_pmap
== NULL
) {
3455 kprintf("Null pmap (cb) at va: 0x%lx\n", pv
->pv_va
);
3461 * Careful here. We can use a locked bus instruction to
3462 * clear PG_A or PG_M safely but we need to synchronize
3463 * with the target cpus when we mess with PG_RW.
3465 * We do not have to force synchronization when clearing
3466 * PG_M even for PTEs generated via virtual memory maps,
3467 * because the virtual kernel will invalidate the pmap
3468 * entry when/if it needs to resynchronize the Modify bit.
3471 pmap_inval_interlock(&info
, pv
->pv_pmap
, pv
->pv_va
);
3472 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_va
);
3479 atomic_clear_long(pte
, PG_M
|PG_RW
);
3482 * The cpu may be trying to set PG_M
3483 * simultaniously with our clearing
3486 if (!atomic_cmpset_long(pte
, pbits
,
3490 } else if (bit
== PG_M
) {
3492 * We could also clear PG_RW here to force
3493 * a fault on write to redetect PG_M for
3494 * virtual kernels, but it isn't necessary
3495 * since virtual kernels invalidate the pte
3496 * when they clear the VPTE_M bit in their
3497 * virtual page tables.
3499 atomic_clear_long(pte
, PG_M
);
3501 atomic_clear_long(pte
, bit
);
3505 pmap_inval_deinterlock(&info
, pv
->pv_pmap
);
3507 pmap_inval_done(&info
);
3511 * pmap_page_protect:
3513 * Lower the permission for all mappings to a given page.
3516 pmap_page_protect(vm_page_t m
, vm_prot_t prot
)
3518 /* JG NX support? */
3519 if ((prot
& VM_PROT_WRITE
) == 0) {
3520 lwkt_gettoken(&vm_token
);
3521 if (prot
& (VM_PROT_READ
| VM_PROT_EXECUTE
)) {
3522 pmap_clearbit(m
, PG_RW
);
3523 vm_page_flag_clear(m
, PG_WRITEABLE
);
3527 lwkt_reltoken(&vm_token
);
3532 pmap_phys_address(vm_pindex_t ppn
)
3534 return (x86_64_ptob(ppn
));
3538 * pmap_ts_referenced:
3540 * Return a count of reference bits for a page, clearing those bits.
3541 * It is not necessary for every reference bit to be cleared, but it
3542 * is necessary that 0 only be returned when there are truly no
3543 * reference bits set.
3545 * XXX: The exact number of bits to check and clear is a matter that
3546 * should be tested and standardized at some point in the future for
3547 * optimal aging of shared pages.
3550 pmap_ts_referenced(vm_page_t m
)
3552 pv_entry_t pv
, pvf
, pvn
;
3556 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
3560 lwkt_gettoken(&vm_token
);
3562 if ((pv
= TAILQ_FIRST(&m
->md
.pv_list
)) != NULL
) {
3567 pvn
= TAILQ_NEXT(pv
, pv_list
);
3570 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
3571 TAILQ_INSERT_TAIL(&m
->md
.pv_list
, pv
, pv_list
);
3574 if (!pmap_track_modified(pv
->pv_va
))
3577 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_va
);
3579 if (pte
&& (*pte
& PG_A
)) {
3581 atomic_clear_long(pte
, PG_A
);
3583 atomic_clear_long_nonlocked(pte
, PG_A
);
3590 } while ((pv
= pvn
) != NULL
&& pv
!= pvf
);
3592 lwkt_reltoken(&vm_token
);
3601 * Return whether or not the specified physical page was modified
3602 * in any physical maps.
3605 pmap_is_modified(vm_page_t m
)
3609 lwkt_gettoken(&vm_token
);
3610 res
= pmap_testbit(m
, PG_M
);
3611 lwkt_reltoken(&vm_token
);
3616 * Clear the modify bits on the specified physical page.
3619 pmap_clear_modify(vm_page_t m
)
3621 lwkt_gettoken(&vm_token
);
3622 pmap_clearbit(m
, PG_M
);
3623 lwkt_reltoken(&vm_token
);
3627 * pmap_clear_reference:
3629 * Clear the reference bit on the specified physical page.
3632 pmap_clear_reference(vm_page_t m
)
3634 lwkt_gettoken(&vm_token
);
3635 pmap_clearbit(m
, PG_A
);
3636 lwkt_reltoken(&vm_token
);
3640 * Miscellaneous support routines follow
3645 i386_protection_init(void)
3649 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3650 kp
= protection_codes
;
3651 for (prot
= 0; prot
< 8; prot
++) {
3653 case VM_PROT_NONE
| VM_PROT_NONE
| VM_PROT_NONE
:
3655 * Read access is also 0. There isn't any execute bit,
3656 * so just make it readable.
3658 case VM_PROT_READ
| VM_PROT_NONE
| VM_PROT_NONE
:
3659 case VM_PROT_READ
| VM_PROT_NONE
| VM_PROT_EXECUTE
:
3660 case VM_PROT_NONE
| VM_PROT_NONE
| VM_PROT_EXECUTE
:
3663 case VM_PROT_NONE
| VM_PROT_WRITE
| VM_PROT_NONE
:
3664 case VM_PROT_NONE
| VM_PROT_WRITE
| VM_PROT_EXECUTE
:
3665 case VM_PROT_READ
| VM_PROT_WRITE
| VM_PROT_NONE
:
3666 case VM_PROT_READ
| VM_PROT_WRITE
| VM_PROT_EXECUTE
:
3674 * Map a set of physical memory pages into the kernel virtual
3675 * address space. Return a pointer to where it is mapped. This
3676 * routine is intended to be used for mapping device memory,
3679 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3683 pmap_mapdev(vm_paddr_t pa
, vm_size_t size
)
3685 vm_offset_t va
, tmpva
, offset
;
3688 offset
= pa
& PAGE_MASK
;
3689 size
= roundup(offset
+ size
, PAGE_SIZE
);
3691 va
= kmem_alloc_nofault(&kernel_map
, size
, PAGE_SIZE
);
3693 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3695 pa
= pa
& ~PAGE_MASK
;
3696 for (tmpva
= va
; size
> 0;) {
3697 pte
= vtopte(tmpva
);
3698 *pte
= pa
| PG_RW
| PG_V
; /* | pgeflag; */
3706 return ((void *)(va
+ offset
));
3710 pmap_mapdev_uncacheable(vm_paddr_t pa
, vm_size_t size
)
3712 vm_offset_t va
, tmpva
, offset
;
3715 offset
= pa
& PAGE_MASK
;
3716 size
= roundup(offset
+ size
, PAGE_SIZE
);
3718 va
= kmem_alloc_nofault(&kernel_map
, size
, PAGE_SIZE
);
3720 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3722 pa
= pa
& ~PAGE_MASK
;
3723 for (tmpva
= va
; size
> 0;) {
3724 pte
= vtopte(tmpva
);
3725 *pte
= pa
| PG_RW
| PG_V
| PG_N
; /* | pgeflag; */
3733 return ((void *)(va
+ offset
));
3737 pmap_unmapdev(vm_offset_t va
, vm_size_t size
)
3739 vm_offset_t base
, offset
;
3741 base
= va
& ~PAGE_MASK
;
3742 offset
= va
& PAGE_MASK
;
3743 size
= roundup(offset
+ size
, PAGE_SIZE
);
3744 pmap_qremove(va
, size
>> PAGE_SHIFT
);
3745 kmem_free(&kernel_map
, base
, size
);
3749 * perform the pmap work for mincore
3752 pmap_mincore(pmap_t pmap
, vm_offset_t addr
)
3754 pt_entry_t
*ptep
, pte
;
3758 lwkt_gettoken(&vm_token
);
3759 ptep
= pmap_pte(pmap
, addr
);
3761 if (ptep
&& (pte
= *ptep
) != 0) {
3764 val
= MINCORE_INCORE
;
3765 if ((pte
& PG_MANAGED
) == 0)
3768 pa
= pte
& PG_FRAME
;
3770 m
= PHYS_TO_VM_PAGE(pa
);
3776 val
|= MINCORE_MODIFIED
|MINCORE_MODIFIED_OTHER
;
3778 * Modified by someone
3780 else if (m
->dirty
|| pmap_is_modified(m
))
3781 val
|= MINCORE_MODIFIED_OTHER
;
3786 val
|= MINCORE_REFERENCED
|MINCORE_REFERENCED_OTHER
;
3789 * Referenced by someone
3791 else if ((m
->flags
& PG_REFERENCED
) || pmap_ts_referenced(m
)) {
3792 val
|= MINCORE_REFERENCED_OTHER
;
3793 vm_page_flag_set(m
, PG_REFERENCED
);
3797 lwkt_reltoken(&vm_token
);
3802 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3803 * vmspace will be ref'd and the old one will be deref'd.
3805 * The vmspace for all lwps associated with the process will be adjusted
3806 * and cr3 will be reloaded if any lwp is the current lwp.
3809 pmap_replacevm(struct proc
*p
, struct vmspace
*newvm
, int adjrefs
)
3811 struct vmspace
*oldvm
;
3815 oldvm
= p
->p_vmspace
;
3816 if (oldvm
!= newvm
) {
3817 p
->p_vmspace
= newvm
;
3818 KKASSERT(p
->p_nthreads
== 1);
3819 lp
= RB_ROOT(&p
->p_lwp_tree
);
3820 pmap_setlwpvm(lp
, newvm
);
3822 sysref_get(&newvm
->vm_sysref
);
3823 sysref_put(&oldvm
->vm_sysref
);
3830 * Set the vmspace for a LWP. The vmspace is almost universally set the
3831 * same as the process vmspace, but virtual kernels need to swap out contexts
3832 * on a per-lwp basis.
3835 pmap_setlwpvm(struct lwp
*lp
, struct vmspace
*newvm
)
3837 struct vmspace
*oldvm
;
3841 oldvm
= lp
->lwp_vmspace
;
3843 if (oldvm
!= newvm
) {
3844 lp
->lwp_vmspace
= newvm
;
3845 if (curthread
->td_lwp
== lp
) {
3846 pmap
= vmspace_pmap(newvm
);
3848 atomic_set_int(&pmap
->pm_active
, mycpu
->gd_cpumask
);
3849 if (pmap
->pm_active
& CPUMASK_LOCK
)
3850 pmap_interlock_wait(newvm
);
3852 pmap
->pm_active
|= 1;
3854 #if defined(SWTCH_OPTIM_STATS)
3857 curthread
->td_pcb
->pcb_cr3
= vtophys(pmap
->pm_pml4
);
3858 curthread
->td_pcb
->pcb_cr3
|= PG_RW
| PG_U
| PG_V
;
3859 load_cr3(curthread
->td_pcb
->pcb_cr3
);
3860 pmap
= vmspace_pmap(oldvm
);
3862 atomic_clear_int(&pmap
->pm_active
, mycpu
->gd_cpumask
);
3864 pmap
->pm_active
&= ~1;
3874 * Called when switching to a locked pmap
3877 pmap_interlock_wait(struct vmspace
*vm
)
3879 struct pmap
*pmap
= &vm
->vm_pmap
;
3881 if (pmap
->pm_active
& CPUMASK_LOCK
) {
3882 while (pmap
->pm_active
& CPUMASK_LOCK
) {
3885 lwkt_process_ipiq();
3893 pmap_addr_hint(vm_object_t obj
, vm_offset_t addr
, vm_size_t size
)
3896 if ((obj
== NULL
) || (size
< NBPDR
) || (obj
->type
!= OBJT_DEVICE
)) {
3900 addr
= (addr
+ (NBPDR
- 1)) & ~(NBPDR
- 1);
3905 * Used by kmalloc/kfree, page already exists at va
3908 pmap_kvtom(vm_offset_t va
)
3910 return(PHYS_TO_VM_PAGE(*vtopte(va
) & PG_FRAME
));