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.
9 * Copyright (c) 2011-2012 Matthew Dillon
10 * All rights reserved.
12 * This code is derived from software contributed to Berkeley by
13 * the Systems Programming Group of the University of Utah Computer
14 * Science Department and William Jolitz of UUNET Technologies Inc.
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
19 * 1. Redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer.
21 * 2. Redistributions in binary form must reproduce the above copyright
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution.
24 * 3. All advertising materials mentioning features or use of this software
25 * must display the following acknowledgement:
26 * This product includes software developed by the University of
27 * California, Berkeley and its contributors.
28 * 4. Neither the name of the University nor the names of its contributors
29 * may be used to endorse or promote products derived from this software
30 * without specific prior written permission.
32 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
45 * Manage physical address maps for x86-64 systems.
49 #include "opt_disable_pse.h"
52 #include "opt_msgbuf.h"
54 #include <sys/param.h>
55 #include <sys/kernel.h>
57 #include <sys/msgbuf.h>
58 #include <sys/vmmeter.h>
60 #include <sys/systm.h>
63 #include <vm/vm_param.h>
64 #include <sys/sysctl.h>
66 #include <vm/vm_kern.h>
67 #include <vm/vm_page.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_object.h>
70 #include <vm/vm_extern.h>
71 #include <vm/vm_pageout.h>
72 #include <vm/vm_pager.h>
73 #include <vm/vm_zone.h>
76 #include <sys/thread2.h>
77 #include <sys/sysref2.h>
78 #include <sys/spinlock2.h>
79 #include <vm/vm_page2.h>
81 #include <machine/cputypes.h>
82 #include <machine/md_var.h>
83 #include <machine/specialreg.h>
84 #include <machine/smp.h>
85 #include <machine_base/apic/apicreg.h>
86 #include <machine/globaldata.h>
87 #include <machine/pmap.h>
88 #include <machine/pmap_inval.h>
89 #include <machine/inttypes.h>
93 #define PMAP_KEEP_PDIRS
94 #ifndef PMAP_SHPGPERPROC
95 #define PMAP_SHPGPERPROC 2000
98 #if defined(DIAGNOSTIC)
99 #define PMAP_DIAGNOSTIC
105 * pmap debugging will report who owns a pv lock when blocking.
109 #define PMAP_DEBUG_DECL ,const char *func, int lineno
110 #define PMAP_DEBUG_ARGS , __func__, __LINE__
111 #define PMAP_DEBUG_COPY , func, lineno
113 #define pv_get(pmap, pindex) _pv_get(pmap, pindex \
115 #define pv_lock(pv) _pv_lock(pv \
117 #define pv_hold_try(pv) _pv_hold_try(pv \
119 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp \
124 #define PMAP_DEBUG_DECL
125 #define PMAP_DEBUG_ARGS
126 #define PMAP_DEBUG_COPY
128 #define pv_get(pmap, pindex) _pv_get(pmap, pindex)
129 #define pv_lock(pv) _pv_lock(pv)
130 #define pv_hold_try(pv) _pv_hold_try(pv)
131 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp)
136 * Get PDEs and PTEs for user/kernel address space
138 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
140 #define pmap_pde_v(pmap, pte) ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
141 #define pmap_pte_w(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
142 #define pmap_pte_m(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
143 #define pmap_pte_u(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
144 #define pmap_pte_v(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
147 * Given a map and a machine independent protection code,
148 * convert to a vax protection code.
150 #define pte_prot(m, p) \
151 (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
152 static int protection_codes
[PROTECTION_CODES_SIZE
];
154 struct pmap kernel_pmap
;
155 static TAILQ_HEAD(,pmap
) pmap_list
= TAILQ_HEAD_INITIALIZER(pmap_list
);
157 MALLOC_DEFINE(M_OBJPMAP
, "objpmap", "pmaps associated with VM objects");
159 vm_paddr_t avail_start
; /* PA of first available physical page */
160 vm_paddr_t avail_end
; /* PA of last available physical page */
161 vm_offset_t virtual2_start
; /* cutout free area prior to kernel start */
162 vm_offset_t virtual2_end
;
163 vm_offset_t virtual_start
; /* VA of first avail page (after kernel bss) */
164 vm_offset_t virtual_end
; /* VA of last avail page (end of kernel AS) */
165 vm_offset_t KvaStart
; /* VA start of KVA space */
166 vm_offset_t KvaEnd
; /* VA end of KVA space (non-inclusive) */
167 vm_offset_t KvaSize
; /* max size of kernel virtual address space */
168 static boolean_t pmap_initialized
= FALSE
; /* Has pmap_init completed? */
169 //static int pgeflag; /* PG_G or-in */
170 //static int pseflag; /* PG_PS or-in */
174 static vm_paddr_t dmaplimit
;
176 vm_offset_t kernel_vm_end
= VM_MIN_KERNEL_ADDRESS
;
178 static pt_entry_t pat_pte_index
[PAT_INDEX_SIZE
]; /* PAT -> PG_ bits */
179 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/ /* PAT -> PG_ bits */
181 static uint64_t KPTbase
;
182 static uint64_t KPTphys
;
183 static uint64_t KPDphys
; /* phys addr of kernel level 2 */
184 static uint64_t KPDbase
; /* phys addr of kernel level 2 @ KERNBASE */
185 uint64_t KPDPphys
; /* phys addr of kernel level 3 */
186 uint64_t KPML4phys
; /* phys addr of kernel level 4 */
188 static uint64_t DMPDphys
; /* phys addr of direct mapped level 2 */
189 static uint64_t DMPDPphys
; /* phys addr of direct mapped level 3 */
192 * Data for the pv entry allocation mechanism
194 static vm_zone_t pvzone
;
195 static struct vm_zone pvzone_store
;
196 static struct vm_object pvzone_obj
;
197 static int pv_entry_max
=0, pv_entry_high_water
=0;
198 static int pmap_pagedaemon_waken
= 0;
199 static struct pv_entry
*pvinit
;
202 * All those kernel PT submaps that BSD is so fond of
204 pt_entry_t
*CMAP1
= NULL
, *ptmmap
;
205 caddr_t CADDR1
= NULL
, ptvmmap
= NULL
;
206 static pt_entry_t
*msgbufmap
;
207 struct msgbuf
*msgbufp
=NULL
;
210 * PMAP default PG_* bits. Needed to be able to add
211 * EPT/NPT pagetable pmap_bits for the VMM module
213 uint64_t pmap_bits_default
[] = {
214 REGULAR_PMAP
, /* TYPE_IDX 0 */
215 X86_PG_V
, /* PG_V_IDX 1 */
216 X86_PG_RW
, /* PG_RW_IDX 2 */
217 X86_PG_U
, /* PG_U_IDX 3 */
218 X86_PG_A
, /* PG_A_IDX 4 */
219 X86_PG_M
, /* PG_M_IDX 5 */
220 X86_PG_PS
, /* PG_PS_IDX3 6 */
221 X86_PG_G
, /* PG_G_IDX 7 */
222 X86_PG_AVAIL1
, /* PG_AVAIL1_IDX 8 */
223 X86_PG_AVAIL2
, /* PG_AVAIL2_IDX 9 */
224 X86_PG_AVAIL3
, /* PG_AVAIL3_IDX 10 */
225 X86_PG_NC_PWT
| X86_PG_NC_PCD
, /* PG_N_IDX 11 */
230 static pt_entry_t
*pt_crashdumpmap
;
231 static caddr_t crashdumpmap
;
234 static int pmap_enter_debug
= 0;
235 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_enter_debug
, CTLFLAG_RW
,
236 &pmap_enter_debug
, 0, "Debug pmap_enter's");
238 static int pmap_yield_count
= 64;
239 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_yield_count
, CTLFLAG_RW
,
240 &pmap_yield_count
, 0, "Yield during init_pt/release");
241 static int pmap_mmu_optimize
= 0;
242 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_mmu_optimize
, CTLFLAG_RW
,
243 &pmap_mmu_optimize
, 0, "Share page table pages when possible");
244 int pmap_fast_kernel_cpusync
= 0;
245 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_fast_kernel_cpusync
, CTLFLAG_RW
,
246 &pmap_fast_kernel_cpusync
, 0, "Share page table pages when possible");
250 /* Standard user access funtions */
251 extern int std_copyinstr (const void *udaddr
, void *kaddr
, size_t len
,
253 extern int std_copyin (const void *udaddr
, void *kaddr
, size_t len
);
254 extern int std_copyout (const void *kaddr
, void *udaddr
, size_t len
);
255 extern int std_fubyte (const void *base
);
256 extern int std_subyte (void *base
, int byte
);
257 extern long std_fuword (const void *base
);
258 extern int std_suword (void *base
, long word
);
259 extern int std_suword32 (void *base
, int word
);
261 static void pv_hold(pv_entry_t pv
);
262 static int _pv_hold_try(pv_entry_t pv
264 static void pv_drop(pv_entry_t pv
);
265 static void _pv_lock(pv_entry_t pv
267 static void pv_unlock(pv_entry_t pv
);
268 static pv_entry_t
_pv_alloc(pmap_t pmap
, vm_pindex_t pindex
, int *isnew
270 static pv_entry_t
_pv_get(pmap_t pmap
, vm_pindex_t pindex
272 static pv_entry_t
pv_get_try(pmap_t pmap
, vm_pindex_t pindex
, int *errorp
);
273 static pv_entry_t
pv_find(pmap_t pmap
, vm_pindex_t pindex
);
274 static void pv_put(pv_entry_t pv
);
275 static void pv_free(pv_entry_t pv
);
276 static void *pv_pte_lookup(pv_entry_t pv
, vm_pindex_t pindex
);
277 static pv_entry_t
pmap_allocpte(pmap_t pmap
, vm_pindex_t ptepindex
,
279 static pv_entry_t
pmap_allocpte_seg(pmap_t pmap
, vm_pindex_t ptepindex
,
280 pv_entry_t
*pvpp
, vm_map_entry_t entry
, vm_offset_t va
);
281 static void pmap_remove_pv_pte(pv_entry_t pv
, pv_entry_t pvp
,
282 pmap_inval_bulk_t
*bulk
);
283 static vm_page_t
pmap_remove_pv_page(pv_entry_t pv
);
284 static int pmap_release_pv(pv_entry_t pv
, pv_entry_t pvp
,
285 pmap_inval_bulk_t
*bulk
);
287 struct pmap_scan_info
;
288 static void pmap_remove_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
289 pv_entry_t pte_pv
, pv_entry_t pt_pv
, int sharept
,
290 vm_offset_t va
, pt_entry_t
*ptep
, void *arg __unused
);
291 static void pmap_protect_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
292 pv_entry_t pte_pv
, pv_entry_t pt_pv
, int sharept
,
293 vm_offset_t va
, pt_entry_t
*ptep
, void *arg __unused
);
295 static void i386_protection_init (void);
296 static void create_pagetables(vm_paddr_t
*firstaddr
);
297 static void pmap_remove_all (vm_page_t m
);
298 static boolean_t
pmap_testbit (vm_page_t m
, int bit
);
300 static pt_entry_t
* pmap_pte_quick (pmap_t pmap
, vm_offset_t va
);
301 static vm_offset_t
pmap_kmem_choose(vm_offset_t addr
);
303 static void pmap_pinit_defaults(struct pmap
*pmap
);
305 static unsigned pdir4mb
;
308 pv_entry_compare(pv_entry_t pv1
, pv_entry_t pv2
)
310 if (pv1
->pv_pindex
< pv2
->pv_pindex
)
312 if (pv1
->pv_pindex
> pv2
->pv_pindex
)
317 RB_GENERATE2(pv_entry_rb_tree
, pv_entry
, pv_entry
,
318 pv_entry_compare
, vm_pindex_t
, pv_pindex
);
322 pmap_page_stats_adding(vm_page_t m
)
324 globaldata_t gd
= mycpu
;
326 if (TAILQ_EMPTY(&m
->md
.pv_list
)) {
327 ++gd
->gd_vmtotal
.t_arm
;
328 } else if (TAILQ_FIRST(&m
->md
.pv_list
) ==
329 TAILQ_LAST(&m
->md
.pv_list
, md_page_pv_list
)) {
330 ++gd
->gd_vmtotal
.t_armshr
;
331 ++gd
->gd_vmtotal
.t_avmshr
;
333 ++gd
->gd_vmtotal
.t_avmshr
;
339 pmap_page_stats_deleting(vm_page_t m
)
341 globaldata_t gd
= mycpu
;
343 if (TAILQ_EMPTY(&m
->md
.pv_list
)) {
344 --gd
->gd_vmtotal
.t_arm
;
345 } else if (TAILQ_FIRST(&m
->md
.pv_list
) ==
346 TAILQ_LAST(&m
->md
.pv_list
, md_page_pv_list
)) {
347 --gd
->gd_vmtotal
.t_armshr
;
348 --gd
->gd_vmtotal
.t_avmshr
;
350 --gd
->gd_vmtotal
.t_avmshr
;
355 * Move the kernel virtual free pointer to the next
356 * 2MB. This is used to help improve performance
357 * by using a large (2MB) page for much of the kernel
358 * (.text, .data, .bss)
362 pmap_kmem_choose(vm_offset_t addr
)
364 vm_offset_t newaddr
= addr
;
366 newaddr
= roundup2(addr
, NBPDR
);
373 * Super fast pmap_pte routine best used when scanning the pv lists.
374 * This eliminates many course-grained invltlb calls. Note that many of
375 * the pv list scans are across different pmaps and it is very wasteful
376 * to do an entire invltlb when checking a single mapping.
378 static __inline pt_entry_t
*pmap_pte(pmap_t pmap
, vm_offset_t va
);
382 pmap_pte_quick(pmap_t pmap
, vm_offset_t va
)
384 return pmap_pte(pmap
, va
);
388 * Returns the pindex of a page table entry (representing a terminal page).
389 * There are NUPTE_TOTAL page table entries possible (a huge number)
391 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
392 * We want to properly translate negative KVAs.
396 pmap_pte_pindex(vm_offset_t va
)
398 return ((va
>> PAGE_SHIFT
) & (NUPTE_TOTAL
- 1));
402 * Returns the pindex of a page table.
406 pmap_pt_pindex(vm_offset_t va
)
408 return (NUPTE_TOTAL
+ ((va
>> PDRSHIFT
) & (NUPT_TOTAL
- 1)));
412 * Returns the pindex of a page directory.
416 pmap_pd_pindex(vm_offset_t va
)
418 return (NUPTE_TOTAL
+ NUPT_TOTAL
+
419 ((va
>> PDPSHIFT
) & (NUPD_TOTAL
- 1)));
424 pmap_pdp_pindex(vm_offset_t va
)
426 return (NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
+
427 ((va
>> PML4SHIFT
) & (NUPDP_TOTAL
- 1)));
432 pmap_pml4_pindex(void)
434 return (NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
+ NUPDP_TOTAL
);
438 * Return various clipped indexes for a given VA
440 * Returns the index of a pte in a page table, representing a terminal
445 pmap_pte_index(vm_offset_t va
)
447 return ((va
>> PAGE_SHIFT
) & ((1ul << NPTEPGSHIFT
) - 1));
451 * Returns the index of a pt in a page directory, representing a page
456 pmap_pt_index(vm_offset_t va
)
458 return ((va
>> PDRSHIFT
) & ((1ul << NPDEPGSHIFT
) - 1));
462 * Returns the index of a pd in a page directory page, representing a page
467 pmap_pd_index(vm_offset_t va
)
469 return ((va
>> PDPSHIFT
) & ((1ul << NPDPEPGSHIFT
) - 1));
473 * Returns the index of a pdp in the pml4 table, representing a page
478 pmap_pdp_index(vm_offset_t va
)
480 return ((va
>> PML4SHIFT
) & ((1ul << NPML4EPGSHIFT
) - 1));
484 * Generic procedure to index a pte from a pt, pd, or pdp.
486 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
487 * a page table page index but is instead of PV lookup index.
491 pv_pte_lookup(pv_entry_t pv
, vm_pindex_t pindex
)
495 pte
= (pt_entry_t
*)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv
->pv_m
));
496 return(&pte
[pindex
]);
500 * Return pointer to PDP slot in the PML4
504 pmap_pdp(pmap_t pmap
, vm_offset_t va
)
506 return (&pmap
->pm_pml4
[pmap_pdp_index(va
)]);
510 * Return pointer to PD slot in the PDP given a pointer to the PDP
514 pmap_pdp_to_pd(pml4_entry_t pdp_pte
, vm_offset_t va
)
518 pd
= (pdp_entry_t
*)PHYS_TO_DMAP(pdp_pte
& PG_FRAME
);
519 return (&pd
[pmap_pd_index(va
)]);
523 * Return pointer to PD slot in the PDP.
527 pmap_pd(pmap_t pmap
, vm_offset_t va
)
531 pdp
= pmap_pdp(pmap
, va
);
532 if ((*pdp
& pmap
->pmap_bits
[PG_V_IDX
]) == 0)
534 return (pmap_pdp_to_pd(*pdp
, va
));
538 * Return pointer to PT slot in the PD given a pointer to the PD
542 pmap_pd_to_pt(pdp_entry_t pd_pte
, vm_offset_t va
)
546 pt
= (pd_entry_t
*)PHYS_TO_DMAP(pd_pte
& PG_FRAME
);
547 return (&pt
[pmap_pt_index(va
)]);
551 * Return pointer to PT slot in the PD
553 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
554 * so we cannot lookup the PD via the PDP. Instead we
555 * must look it up via the pmap.
559 pmap_pt(pmap_t pmap
, vm_offset_t va
)
563 vm_pindex_t pd_pindex
;
565 if (pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) {
566 pd_pindex
= pmap_pd_pindex(va
);
567 spin_lock(&pmap
->pm_spin
);
568 pv
= pv_entry_rb_tree_RB_LOOKUP(&pmap
->pm_pvroot
, pd_pindex
);
569 spin_unlock(&pmap
->pm_spin
);
570 if (pv
== NULL
|| pv
->pv_m
== NULL
)
572 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv
->pv_m
), va
));
574 pd
= pmap_pd(pmap
, va
);
575 if (pd
== NULL
|| (*pd
& pmap
->pmap_bits
[PG_V_IDX
]) == 0)
577 return (pmap_pd_to_pt(*pd
, va
));
582 * Return pointer to PTE slot in the PT given a pointer to the PT
586 pmap_pt_to_pte(pd_entry_t pt_pte
, vm_offset_t va
)
590 pte
= (pt_entry_t
*)PHYS_TO_DMAP(pt_pte
& PG_FRAME
);
591 return (&pte
[pmap_pte_index(va
)]);
595 * Return pointer to PTE slot in the PT
599 pmap_pte(pmap_t pmap
, vm_offset_t va
)
603 pt
= pmap_pt(pmap
, va
);
604 if (pt
== NULL
|| (*pt
& pmap
->pmap_bits
[PG_V_IDX
]) == 0)
606 if ((*pt
& pmap
->pmap_bits
[PG_PS_IDX
]) != 0)
607 return ((pt_entry_t
*)pt
);
608 return (pmap_pt_to_pte(*pt
, va
));
612 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
613 * the PT layer. This will speed up core pmap operations considerably.
615 * NOTE: The pmap spinlock does not need to be held but the passed-in pv
616 * must be in a known associated state (typically by being locked when
617 * the pmap spinlock isn't held). We allow the race for that case.
621 pv_cache(pv_entry_t pv
, vm_pindex_t pindex
)
623 if (pindex
>= pmap_pt_pindex(0) && pindex
<= pmap_pd_pindex(0))
624 pv
->pv_pmap
->pm_pvhint
= pv
;
629 * Return address of PT slot in PD (KVM only)
631 * Cannot be used for user page tables because it might interfere with
632 * the shared page-table-page optimization (pmap_mmu_optimize).
636 vtopt(vm_offset_t va
)
638 uint64_t mask
= ((1ul << (NPDEPGSHIFT
+ NPDPEPGSHIFT
+
639 NPML4EPGSHIFT
)) - 1);
641 return (PDmap
+ ((va
>> PDRSHIFT
) & mask
));
645 * KVM - return address of PTE slot in PT
649 vtopte(vm_offset_t va
)
651 uint64_t mask
= ((1ul << (NPTEPGSHIFT
+ NPDEPGSHIFT
+
652 NPDPEPGSHIFT
+ NPML4EPGSHIFT
)) - 1);
654 return (PTmap
+ ((va
>> PAGE_SHIFT
) & mask
));
658 allocpages(vm_paddr_t
*firstaddr
, long n
)
663 bzero((void *)ret
, n
* PAGE_SIZE
);
664 *firstaddr
+= n
* PAGE_SIZE
;
670 create_pagetables(vm_paddr_t
*firstaddr
)
672 long i
; /* must be 64 bits */
678 * We are running (mostly) V=P at this point
680 * Calculate NKPT - number of kernel page tables. We have to
681 * accomodoate prealloction of the vm_page_array, dump bitmap,
682 * MSGBUF_SIZE, and other stuff. Be generous.
684 * Maxmem is in pages.
686 * ndmpdp is the number of 1GB pages we wish to map.
688 ndmpdp
= (ptoa(Maxmem
) + NBPDP
- 1) >> PDPSHIFT
;
689 if (ndmpdp
< 4) /* Minimum 4GB of dirmap */
691 KKASSERT(ndmpdp
<= NKPDPE
* NPDEPG
);
694 * Starting at the beginning of kvm (not KERNBASE).
696 nkpt_phys
= (Maxmem
* sizeof(struct vm_page
) + NBPDR
- 1) / NBPDR
;
697 nkpt_phys
+= (Maxmem
* sizeof(struct pv_entry
) + NBPDR
- 1) / NBPDR
;
698 nkpt_phys
+= ((nkpt
+ nkpt
+ 1 + NKPML4E
+ NKPDPE
+ NDMPML4E
+
699 ndmpdp
) + 511) / 512;
703 * Starting at KERNBASE - map 2G worth of page table pages.
704 * KERNBASE is offset -2G from the end of kvm.
706 nkpt_base
= (NPDPEPG
- KPDPI
) * NPTEPG
; /* typically 2 x 512 */
711 KPTbase
= allocpages(firstaddr
, nkpt_base
);
712 KPTphys
= allocpages(firstaddr
, nkpt_phys
);
713 KPML4phys
= allocpages(firstaddr
, 1);
714 KPDPphys
= allocpages(firstaddr
, NKPML4E
);
715 KPDphys
= allocpages(firstaddr
, NKPDPE
);
718 * Calculate the page directory base for KERNBASE,
719 * that is where we start populating the page table pages.
720 * Basically this is the end - 2.
722 KPDbase
= KPDphys
+ ((NKPDPE
- (NPDPEPG
- KPDPI
)) << PAGE_SHIFT
);
724 DMPDPphys
= allocpages(firstaddr
, NDMPML4E
);
725 if ((amd_feature
& AMDID_PAGE1GB
) == 0)
726 DMPDphys
= allocpages(firstaddr
, ndmpdp
);
727 dmaplimit
= (vm_paddr_t
)ndmpdp
<< PDPSHIFT
;
730 * Fill in the underlying page table pages for the area around
731 * KERNBASE. This remaps low physical memory to KERNBASE.
733 * Read-only from zero to physfree
734 * XXX not fully used, underneath 2M pages
736 for (i
= 0; (i
<< PAGE_SHIFT
) < *firstaddr
; i
++) {
737 ((pt_entry_t
*)KPTbase
)[i
] = i
<< PAGE_SHIFT
;
738 ((pt_entry_t
*)KPTbase
)[i
] |=
739 pmap_bits_default
[PG_RW_IDX
] |
740 pmap_bits_default
[PG_V_IDX
] |
741 pmap_bits_default
[PG_G_IDX
];
745 * Now map the initial kernel page tables. One block of page
746 * tables is placed at the beginning of kernel virtual memory,
747 * and another block is placed at KERNBASE to map the kernel binary,
748 * data, bss, and initial pre-allocations.
750 for (i
= 0; i
< nkpt_base
; i
++) {
751 ((pd_entry_t
*)KPDbase
)[i
] = KPTbase
+ (i
<< PAGE_SHIFT
);
752 ((pd_entry_t
*)KPDbase
)[i
] |=
753 pmap_bits_default
[PG_RW_IDX
] |
754 pmap_bits_default
[PG_V_IDX
];
756 for (i
= 0; i
< nkpt_phys
; i
++) {
757 ((pd_entry_t
*)KPDphys
)[i
] = KPTphys
+ (i
<< PAGE_SHIFT
);
758 ((pd_entry_t
*)KPDphys
)[i
] |=
759 pmap_bits_default
[PG_RW_IDX
] |
760 pmap_bits_default
[PG_V_IDX
];
764 * Map from zero to end of allocations using 2M pages as an
765 * optimization. This will bypass some of the KPTBase pages
766 * above in the KERNBASE area.
768 for (i
= 0; (i
<< PDRSHIFT
) < *firstaddr
; i
++) {
769 ((pd_entry_t
*)KPDbase
)[i
] = i
<< PDRSHIFT
;
770 ((pd_entry_t
*)KPDbase
)[i
] |=
771 pmap_bits_default
[PG_RW_IDX
] |
772 pmap_bits_default
[PG_V_IDX
] |
773 pmap_bits_default
[PG_PS_IDX
] |
774 pmap_bits_default
[PG_G_IDX
];
778 * And connect up the PD to the PDP. The kernel pmap is expected
779 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
781 for (i
= 0; i
< NKPDPE
; i
++) {
782 ((pdp_entry_t
*)KPDPphys
)[NPDPEPG
- NKPDPE
+ i
] =
783 KPDphys
+ (i
<< PAGE_SHIFT
);
784 ((pdp_entry_t
*)KPDPphys
)[NPDPEPG
- NKPDPE
+ i
] |=
785 pmap_bits_default
[PG_RW_IDX
] |
786 pmap_bits_default
[PG_V_IDX
] |
787 pmap_bits_default
[PG_U_IDX
];
791 * Now set up the direct map space using either 2MB or 1GB pages
792 * Preset PG_M and PG_A because demotion expects it.
794 * When filling in entries in the PD pages make sure any excess
795 * entries are set to zero as we allocated enough PD pages
797 if ((amd_feature
& AMDID_PAGE1GB
) == 0) {
798 for (i
= 0; i
< NPDEPG
* ndmpdp
; i
++) {
799 ((pd_entry_t
*)DMPDphys
)[i
] = i
<< PDRSHIFT
;
800 ((pd_entry_t
*)DMPDphys
)[i
] |=
801 pmap_bits_default
[PG_RW_IDX
] |
802 pmap_bits_default
[PG_V_IDX
] |
803 pmap_bits_default
[PG_PS_IDX
] |
804 pmap_bits_default
[PG_G_IDX
] |
805 pmap_bits_default
[PG_M_IDX
] |
806 pmap_bits_default
[PG_A_IDX
];
810 * And the direct map space's PDP
812 for (i
= 0; i
< ndmpdp
; i
++) {
813 ((pdp_entry_t
*)DMPDPphys
)[i
] = DMPDphys
+
815 ((pdp_entry_t
*)DMPDPphys
)[i
] |=
816 pmap_bits_default
[PG_RW_IDX
] |
817 pmap_bits_default
[PG_V_IDX
] |
818 pmap_bits_default
[PG_U_IDX
];
821 for (i
= 0; i
< ndmpdp
; i
++) {
822 ((pdp_entry_t
*)DMPDPphys
)[i
] =
823 (vm_paddr_t
)i
<< PDPSHIFT
;
824 ((pdp_entry_t
*)DMPDPphys
)[i
] |=
825 pmap_bits_default
[PG_RW_IDX
] |
826 pmap_bits_default
[PG_V_IDX
] |
827 pmap_bits_default
[PG_PS_IDX
] |
828 pmap_bits_default
[PG_G_IDX
] |
829 pmap_bits_default
[PG_M_IDX
] |
830 pmap_bits_default
[PG_A_IDX
];
834 /* And recursively map PML4 to itself in order to get PTmap */
835 ((pdp_entry_t
*)KPML4phys
)[PML4PML4I
] = KPML4phys
;
836 ((pdp_entry_t
*)KPML4phys
)[PML4PML4I
] |=
837 pmap_bits_default
[PG_RW_IDX
] |
838 pmap_bits_default
[PG_V_IDX
] |
839 pmap_bits_default
[PG_U_IDX
];
842 * Connect the Direct Map slots up to the PML4
844 for (j
= 0; j
< NDMPML4E
; ++j
) {
845 ((pdp_entry_t
*)KPML4phys
)[DMPML4I
+ j
] =
846 (DMPDPphys
+ ((vm_paddr_t
)j
<< PML4SHIFT
)) |
847 pmap_bits_default
[PG_RW_IDX
] |
848 pmap_bits_default
[PG_V_IDX
] |
849 pmap_bits_default
[PG_U_IDX
];
853 * Connect the KVA slot up to the PML4
855 ((pdp_entry_t
*)KPML4phys
)[KPML4I
] = KPDPphys
;
856 ((pdp_entry_t
*)KPML4phys
)[KPML4I
] |=
857 pmap_bits_default
[PG_RW_IDX
] |
858 pmap_bits_default
[PG_V_IDX
] |
859 pmap_bits_default
[PG_U_IDX
];
863 * Bootstrap the system enough to run with virtual memory.
865 * On the i386 this is called after mapping has already been enabled
866 * and just syncs the pmap module with what has already been done.
867 * [We can't call it easily with mapping off since the kernel is not
868 * mapped with PA == VA, hence we would have to relocate every address
869 * from the linked base (virtual) address "KERNBASE" to the actual
870 * (physical) address starting relative to 0]
873 pmap_bootstrap(vm_paddr_t
*firstaddr
)
878 KvaStart
= VM_MIN_KERNEL_ADDRESS
;
879 KvaEnd
= VM_MAX_KERNEL_ADDRESS
;
880 KvaSize
= KvaEnd
- KvaStart
;
882 avail_start
= *firstaddr
;
885 * Create an initial set of page tables to run the kernel in.
887 create_pagetables(firstaddr
);
889 virtual2_start
= KvaStart
;
890 virtual2_end
= PTOV_OFFSET
;
892 virtual_start
= (vm_offset_t
) PTOV_OFFSET
+ *firstaddr
;
893 virtual_start
= pmap_kmem_choose(virtual_start
);
895 virtual_end
= VM_MAX_KERNEL_ADDRESS
;
897 /* XXX do %cr0 as well */
898 load_cr4(rcr4() | CR4_PGE
| CR4_PSE
);
902 * Initialize protection array.
904 i386_protection_init();
907 * The kernel's pmap is statically allocated so we don't have to use
908 * pmap_create, which is unlikely to work correctly at this part of
909 * the boot sequence (XXX and which no longer exists).
911 kernel_pmap
.pm_pml4
= (pdp_entry_t
*) (PTOV_OFFSET
+ KPML4phys
);
912 kernel_pmap
.pm_count
= 1;
913 CPUMASK_ASSALLONES(kernel_pmap
.pm_active
);
914 RB_INIT(&kernel_pmap
.pm_pvroot
);
915 spin_init(&kernel_pmap
.pm_spin
, "pmapbootstrap");
916 lwkt_token_init(&kernel_pmap
.pm_token
, "kpmap_tok");
919 * Reserve some special page table entries/VA space for temporary
922 #define SYSMAP(c, p, v, n) \
923 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
929 * CMAP1/CMAP2 are used for zeroing and copying pages.
931 SYSMAP(caddr_t
, CMAP1
, CADDR1
, 1)
936 SYSMAP(caddr_t
, pt_crashdumpmap
, crashdumpmap
, MAXDUMPPGS
);
939 * ptvmmap is used for reading arbitrary physical pages via
942 SYSMAP(caddr_t
, ptmmap
, ptvmmap
, 1)
945 * msgbufp is used to map the system message buffer.
946 * XXX msgbufmap is not used.
948 SYSMAP(struct msgbuf
*, msgbufmap
, msgbufp
,
949 atop(round_page(MSGBUF_SIZE
)))
952 virtual_start
= pmap_kmem_choose(virtual_start
);
957 * PG_G is terribly broken on SMP because we IPI invltlb's in some
958 * cases rather then invl1pg. Actually, I don't even know why it
959 * works under UP because self-referential page table mappings
964 * Initialize the 4MB page size flag
968 * The 4MB page version of the initial
969 * kernel page mapping.
973 #if !defined(DISABLE_PSE)
974 if (cpu_feature
& CPUID_PSE
) {
977 * Note that we have enabled PSE mode
979 // pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
980 ptditmp
= *(PTmap
+ x86_64_btop(KERNBASE
));
981 ptditmp
&= ~(NBPDR
- 1);
982 ptditmp
|= pmap_bits_default
[PG_V_IDX
] |
983 pmap_bits_default
[PG_RW_IDX
] |
984 pmap_bits_default
[PG_PS_IDX
] |
985 pmap_bits_default
[PG_U_IDX
];
992 /* Initialize the PAT MSR */
994 pmap_pinit_defaults(&kernel_pmap
);
996 TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
997 &pmap_fast_kernel_cpusync
);
1002 * Setup the PAT MSR.
1011 * Default values mapping PATi,PCD,PWT bits at system reset.
1012 * The default values effectively ignore the PATi bit by
1013 * repeating the encodings for 0-3 in 4-7, and map the PCD
1014 * and PWT bit combinations to the expected PAT types.
1016 pat_msr
= PAT_VALUE(0, PAT_WRITE_BACK
) | /* 000 */
1017 PAT_VALUE(1, PAT_WRITE_THROUGH
) | /* 001 */
1018 PAT_VALUE(2, PAT_UNCACHED
) | /* 010 */
1019 PAT_VALUE(3, PAT_UNCACHEABLE
) | /* 011 */
1020 PAT_VALUE(4, PAT_WRITE_BACK
) | /* 100 */
1021 PAT_VALUE(5, PAT_WRITE_THROUGH
) | /* 101 */
1022 PAT_VALUE(6, PAT_UNCACHED
) | /* 110 */
1023 PAT_VALUE(7, PAT_UNCACHEABLE
); /* 111 */
1024 pat_pte_index
[PAT_WRITE_BACK
] = 0;
1025 pat_pte_index
[PAT_WRITE_THROUGH
]= 0 | X86_PG_NC_PWT
;
1026 pat_pte_index
[PAT_UNCACHED
] = X86_PG_NC_PCD
;
1027 pat_pte_index
[PAT_UNCACHEABLE
] = X86_PG_NC_PCD
| X86_PG_NC_PWT
;
1028 pat_pte_index
[PAT_WRITE_PROTECTED
] = pat_pte_index
[PAT_UNCACHEABLE
];
1029 pat_pte_index
[PAT_WRITE_COMBINING
] = pat_pte_index
[PAT_UNCACHEABLE
];
1031 if (cpu_feature
& CPUID_PAT
) {
1033 * If we support the PAT then set-up entries for
1034 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1037 pat_msr
= (pat_msr
& ~PAT_MASK(4)) |
1038 PAT_VALUE(4, PAT_WRITE_PROTECTED
);
1039 pat_msr
= (pat_msr
& ~PAT_MASK(5)) |
1040 PAT_VALUE(5, PAT_WRITE_COMBINING
);
1041 pat_pte_index
[PAT_WRITE_PROTECTED
] = X86_PG_PTE_PAT
| 0;
1042 pat_pte_index
[PAT_WRITE_COMBINING
] = X86_PG_PTE_PAT
| X86_PG_NC_PWT
;
1045 * Then enable the PAT
1050 load_cr4(cr4
& ~CR4_PGE
);
1052 /* Disable caches (CD = 1, NW = 0). */
1054 load_cr0((cr0
& ~CR0_NW
) | CR0_CD
);
1056 /* Flushes caches and TLBs. */
1060 /* Update PAT and index table. */
1061 wrmsr(MSR_PAT
, pat_msr
);
1063 /* Flush caches and TLBs again. */
1067 /* Restore caches and PGE. */
1075 * Set 4mb pdir for mp startup
1080 if (cpu_feature
& CPUID_PSE
) {
1081 load_cr4(rcr4() | CR4_PSE
);
1082 if (pdir4mb
&& mycpu
->gd_cpuid
== 0) { /* only on BSP */
1089 * Initialize the pmap module.
1090 * Called by vm_init, to initialize any structures that the pmap
1091 * system needs to map virtual memory.
1092 * pmap_init has been enhanced to support in a fairly consistant
1093 * way, discontiguous physical memory.
1102 * Allocate memory for random pmap data structures. Includes the
1106 for (i
= 0; i
< vm_page_array_size
; i
++) {
1109 m
= &vm_page_array
[i
];
1110 TAILQ_INIT(&m
->md
.pv_list
);
1114 * init the pv free list
1116 initial_pvs
= vm_page_array_size
;
1117 if (initial_pvs
< MINPV
)
1118 initial_pvs
= MINPV
;
1119 pvzone
= &pvzone_store
;
1120 pvinit
= (void *)kmem_alloc(&kernel_map
,
1121 initial_pvs
* sizeof (struct pv_entry
));
1122 zbootinit(pvzone
, "PV ENTRY", sizeof (struct pv_entry
),
1123 pvinit
, initial_pvs
);
1126 * Now it is safe to enable pv_table recording.
1128 pmap_initialized
= TRUE
;
1132 * Initialize the address space (zone) for the pv_entries. Set a
1133 * high water mark so that the system can recover from excessive
1134 * numbers of pv entries.
1139 int shpgperproc
= PMAP_SHPGPERPROC
;
1142 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc
);
1143 pv_entry_max
= shpgperproc
* maxproc
+ vm_page_array_size
;
1144 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max
);
1145 pv_entry_high_water
= 9 * (pv_entry_max
/ 10);
1148 * Subtract out pages already installed in the zone (hack)
1150 entry_max
= pv_entry_max
- vm_page_array_size
;
1154 zinitna(pvzone
, &pvzone_obj
, NULL
, 0, entry_max
, ZONE_INTERRUPT
, 1);
1158 * Typically used to initialize a fictitious page by vm/device_pager.c
1161 pmap_page_init(struct vm_page
*m
)
1164 TAILQ_INIT(&m
->md
.pv_list
);
1167 /***************************************************
1168 * Low level helper routines.....
1169 ***************************************************/
1172 * this routine defines the region(s) of memory that should
1173 * not be tested for the modified bit.
1177 pmap_track_modified(vm_pindex_t pindex
)
1179 vm_offset_t va
= (vm_offset_t
)pindex
<< PAGE_SHIFT
;
1180 if ((va
< clean_sva
) || (va
>= clean_eva
))
1187 * Extract the physical page address associated with the map/VA pair.
1188 * The page must be wired for this to work reliably.
1190 * XXX for the moment we're using pv_find() instead of pv_get(), as
1191 * callers might be expecting non-blocking operation.
1194 pmap_extract(pmap_t pmap
, vm_offset_t va
)
1201 if (va
>= VM_MAX_USER_ADDRESS
) {
1203 * Kernel page directories might be direct-mapped and
1204 * there is typically no PV tracking of pte's
1208 pt
= pmap_pt(pmap
, va
);
1209 if (pt
&& (*pt
& pmap
->pmap_bits
[PG_V_IDX
])) {
1210 if (*pt
& pmap
->pmap_bits
[PG_PS_IDX
]) {
1211 rtval
= *pt
& PG_PS_FRAME
;
1212 rtval
|= va
& PDRMASK
;
1214 ptep
= pmap_pt_to_pte(*pt
, va
);
1215 if (*pt
& pmap
->pmap_bits
[PG_V_IDX
]) {
1216 rtval
= *ptep
& PG_FRAME
;
1217 rtval
|= va
& PAGE_MASK
;
1223 * User pages currently do not direct-map the page directory
1224 * and some pages might not used managed PVs. But all PT's
1227 pt_pv
= pv_find(pmap
, pmap_pt_pindex(va
));
1229 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
1230 if (*ptep
& pmap
->pmap_bits
[PG_V_IDX
]) {
1231 rtval
= *ptep
& PG_FRAME
;
1232 rtval
|= va
& PAGE_MASK
;
1241 * Similar to extract but checks protections, SMP-friendly short-cut for
1242 * vm_fault_page[_quick](). Can return NULL to cause the caller to
1243 * fall-through to the real fault code.
1245 * The returned page, if not NULL, is held (and not busied).
1248 pmap_fault_page_quick(pmap_t pmap
, vm_offset_t va
, vm_prot_t prot
)
1250 if (pmap
&& va
< VM_MAX_USER_ADDRESS
) {
1258 req
= pmap
->pmap_bits
[PG_V_IDX
] |
1259 pmap
->pmap_bits
[PG_U_IDX
];
1260 if (prot
& VM_PROT_WRITE
)
1261 req
|= pmap
->pmap_bits
[PG_RW_IDX
];
1263 pt_pv
= pv_find(pmap
, pmap_pt_pindex(va
));
1266 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
1267 if ((*ptep
& req
) != req
) {
1271 pte_pv
= pv_get_try(pmap
, pmap_pte_pindex(va
), &error
);
1272 if (pte_pv
&& error
== 0) {
1275 if (prot
& VM_PROT_WRITE
)
1278 } else if (pte_pv
) {
1292 * Extract the physical page address associated kernel virtual address.
1295 pmap_kextract(vm_offset_t va
)
1297 pd_entry_t pt
; /* pt entry in pd */
1300 if (va
>= DMAP_MIN_ADDRESS
&& va
< DMAP_MAX_ADDRESS
) {
1301 pa
= DMAP_TO_PHYS(va
);
1304 if (pt
& kernel_pmap
.pmap_bits
[PG_PS_IDX
]) {
1305 pa
= (pt
& PG_PS_FRAME
) | (va
& PDRMASK
);
1308 * Beware of a concurrent promotion that changes the
1309 * PDE at this point! For example, vtopte() must not
1310 * be used to access the PTE because it would use the
1311 * new PDE. It is, however, safe to use the old PDE
1312 * because the page table page is preserved by the
1315 pa
= *pmap_pt_to_pte(pt
, va
);
1316 pa
= (pa
& PG_FRAME
) | (va
& PAGE_MASK
);
1322 /***************************************************
1323 * Low level mapping routines.....
1324 ***************************************************/
1327 * Routine: pmap_kenter
1329 * Add a wired page to the KVA
1330 * NOTE! note that in order for the mapping to take effect -- you
1331 * should do an invltlb after doing the pmap_kenter().
1334 pmap_kenter(vm_offset_t va
, vm_paddr_t pa
)
1340 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
1341 kernel_pmap
.pmap_bits
[PG_V_IDX
];
1345 pmap_inval_smp(&kernel_pmap
, va
, 1, ptep
, npte
);
1349 pmap_inval_smp(&kernel_pmap
, va
, ptep
, npte
);
1356 * Similar to pmap_kenter(), except we only invalidate the mapping on the
1357 * current CPU. Returns 0 if the previous pte was 0, 1 if it wasn't
1358 * (caller can conditionalize calling smp_invltlb()).
1361 pmap_kenter_quick(vm_offset_t va
, vm_paddr_t pa
)
1368 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
1369 kernel_pmap
.pmap_bits
[PG_V_IDX
];
1379 cpu_invlpg((void *)va
);
1385 * Enter addresses into the kernel pmap but don't bother
1386 * doing any tlb invalidations. Caller will do a rollup
1387 * invalidation via pmap_rollup_inval().
1390 pmap_kenter_noinval(vm_offset_t va
, vm_paddr_t pa
)
1397 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
1398 kernel_pmap
.pmap_bits
[PG_V_IDX
];
1408 cpu_invlpg((void *)va
);
1414 * remove a page from the kernel pagetables
1417 pmap_kremove(vm_offset_t va
)
1422 pmap_inval_smp(&kernel_pmap
, va
, 1, ptep
, 0);
1426 pmap_kremove_quick(vm_offset_t va
)
1431 (void)pte_load_clear(ptep
);
1432 cpu_invlpg((void *)va
);
1436 * Remove addresses from the kernel pmap but don't bother
1437 * doing any tlb invalidations. Caller will do a rollup
1438 * invalidation via pmap_rollup_inval().
1441 pmap_kremove_noinval(vm_offset_t va
)
1446 (void)pte_load_clear(ptep
);
1450 * XXX these need to be recoded. They are not used in any critical path.
1453 pmap_kmodify_rw(vm_offset_t va
)
1455 atomic_set_long(vtopte(va
), kernel_pmap
.pmap_bits
[PG_RW_IDX
]);
1456 cpu_invlpg((void *)va
);
1461 pmap_kmodify_nc(vm_offset_t va)
1463 atomic_set_long(vtopte(va), PG_N);
1464 cpu_invlpg((void *)va);
1469 * Used to map a range of physical addresses into kernel virtual
1470 * address space during the low level boot, typically to map the
1471 * dump bitmap, message buffer, and vm_page_array.
1473 * These mappings are typically made at some pointer after the end of the
1476 * We could return PHYS_TO_DMAP(start) here and not allocate any
1477 * via (*virtp), but then kmem from userland and kernel dumps won't
1478 * have access to the related pointers.
1481 pmap_map(vm_offset_t
*virtp
, vm_paddr_t start
, vm_paddr_t end
, int prot
)
1484 vm_offset_t va_start
;
1486 /*return PHYS_TO_DMAP(start);*/
1491 while (start
< end
) {
1492 pmap_kenter_quick(va
, start
);
1500 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1503 * Remove the specified set of pages from the data and instruction caches.
1505 * In contrast to pmap_invalidate_cache_range(), this function does not
1506 * rely on the CPU's self-snoop feature, because it is intended for use
1507 * when moving pages into a different cache domain.
1510 pmap_invalidate_cache_pages(vm_page_t
*pages
, int count
)
1512 vm_offset_t daddr
, eva
;
1515 if (count
>= PMAP_CLFLUSH_THRESHOLD
/ PAGE_SIZE
||
1516 (cpu_feature
& CPUID_CLFSH
) == 0)
1520 for (i
= 0; i
< count
; i
++) {
1521 daddr
= PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages
[i
]));
1522 eva
= daddr
+ PAGE_SIZE
;
1523 for (; daddr
< eva
; daddr
+= cpu_clflush_line_size
)
1531 pmap_invalidate_cache_range(vm_offset_t sva
, vm_offset_t eva
)
1533 KASSERT((sva
& PAGE_MASK
) == 0,
1534 ("pmap_invalidate_cache_range: sva not page-aligned"));
1535 KASSERT((eva
& PAGE_MASK
) == 0,
1536 ("pmap_invalidate_cache_range: eva not page-aligned"));
1538 if (cpu_feature
& CPUID_SS
) {
1539 ; /* If "Self Snoop" is supported, do nothing. */
1541 /* Globally invalidate caches */
1542 cpu_wbinvd_on_all_cpus();
1547 * Invalidate the specified range of virtual memory on all cpus associated
1551 pmap_invalidate_range(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
)
1553 pmap_inval_smp(pmap
, sva
, (eva
- sva
) >> PAGE_SHIFT
, NULL
, 0);
1557 * Add a list of wired pages to the kva. This routine is used for temporary
1558 * kernel mappings such as those found in buffer cache buffer. Page
1559 * modifications and accesses are not tracked or recorded.
1561 * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1562 * semantics as previous mappings may have been zerod without any
1565 * The page *must* be wired.
1568 pmap_qenter(vm_offset_t beg_va
, vm_page_t
*m
, int count
)
1573 end_va
= beg_va
+ count
* PAGE_SIZE
;
1575 for (va
= beg_va
; va
< end_va
; va
+= PAGE_SIZE
) {
1579 *pte
= VM_PAGE_TO_PHYS(*m
) |
1580 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
1581 kernel_pmap
.pmap_bits
[PG_V_IDX
] |
1582 kernel_pmap
.pmap_cache_bits
[(*m
)->pat_mode
];
1586 pmap_invalidate_range(&kernel_pmap
, beg_va
, end_va
);
1590 * This routine jerks page mappings from the kernel -- it is meant only
1591 * for temporary mappings such as those found in buffer cache buffers.
1592 * No recording modified or access status occurs.
1594 * MPSAFE, INTERRUPT SAFE (cluster callback)
1597 pmap_qremove(vm_offset_t beg_va
, int count
)
1602 end_va
= beg_va
+ count
* PAGE_SIZE
;
1604 for (va
= beg_va
; va
< end_va
; va
+= PAGE_SIZE
) {
1608 (void)pte_load_clear(pte
);
1609 cpu_invlpg((void *)va
);
1611 pmap_invalidate_range(&kernel_pmap
, beg_va
, end_va
);
1615 * This routine removes temporary kernel mappings, only invalidating them
1616 * on the current cpu. It should only be used under carefully controlled
1620 pmap_qremove_quick(vm_offset_t beg_va
, int count
)
1625 end_va
= beg_va
+ count
* PAGE_SIZE
;
1627 for (va
= beg_va
; va
< end_va
; va
+= PAGE_SIZE
) {
1631 (void)pte_load_clear(pte
);
1632 cpu_invlpg((void *)va
);
1637 * This routine removes temporary kernel mappings *without* invalidating
1638 * the TLB. It can only be used on permanent kva reservations such as those
1639 * found in buffer cache buffers, under carefully controlled circumstances.
1641 * NOTE: Repopulating these KVAs requires unconditional invalidation.
1642 * (pmap_qenter() does unconditional invalidation).
1645 pmap_qremove_noinval(vm_offset_t beg_va
, int count
)
1650 end_va
= beg_va
+ count
* PAGE_SIZE
;
1652 for (va
= beg_va
; va
< end_va
; va
+= PAGE_SIZE
) {
1656 (void)pte_load_clear(pte
);
1661 * Create a new thread and optionally associate it with a (new) process.
1662 * NOTE! the new thread's cpu may not equal the current cpu.
1665 pmap_init_thread(thread_t td
)
1667 /* enforce pcb placement & alignment */
1668 td
->td_pcb
= (struct pcb
*)(td
->td_kstack
+ td
->td_kstack_size
) - 1;
1669 td
->td_pcb
= (struct pcb
*)((intptr_t)td
->td_pcb
& ~(intptr_t)0xF);
1670 td
->td_savefpu
= &td
->td_pcb
->pcb_save
;
1671 td
->td_sp
= (char *)td
->td_pcb
; /* no -16 */
1675 * This routine directly affects the fork perf for a process.
1678 pmap_init_proc(struct proc
*p
)
1683 pmap_pinit_defaults(struct pmap
*pmap
)
1685 bcopy(pmap_bits_default
, pmap
->pmap_bits
,
1686 sizeof(pmap_bits_default
));
1687 bcopy(protection_codes
, pmap
->protection_codes
,
1688 sizeof(protection_codes
));
1689 bcopy(pat_pte_index
, pmap
->pmap_cache_bits
,
1690 sizeof(pat_pte_index
));
1691 pmap
->pmap_cache_mask
= X86_PG_NC_PWT
| X86_PG_NC_PCD
| X86_PG_PTE_PAT
;
1692 pmap
->copyinstr
= std_copyinstr
;
1693 pmap
->copyin
= std_copyin
;
1694 pmap
->copyout
= std_copyout
;
1695 pmap
->fubyte
= std_fubyte
;
1696 pmap
->subyte
= std_subyte
;
1697 pmap
->fuword
= std_fuword
;
1698 pmap
->suword
= std_suword
;
1699 pmap
->suword32
= std_suword32
;
1702 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1703 * it, and IdlePTD, represents the template used to update all other pmaps.
1705 * On architectures where the kernel pmap is not integrated into the user
1706 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1707 * kernel_pmap should be used to directly access the kernel_pmap.
1710 pmap_pinit0(struct pmap
*pmap
)
1712 pmap
->pm_pml4
= (pml4_entry_t
*)(PTOV_OFFSET
+ KPML4phys
);
1714 CPUMASK_ASSZERO(pmap
->pm_active
);
1715 pmap
->pm_pvhint
= NULL
;
1716 RB_INIT(&pmap
->pm_pvroot
);
1717 spin_init(&pmap
->pm_spin
, "pmapinit0");
1718 lwkt_token_init(&pmap
->pm_token
, "pmap_tok");
1719 bzero(&pmap
->pm_stats
, sizeof pmap
->pm_stats
);
1720 pmap_pinit_defaults(pmap
);
1724 * Initialize a preallocated and zeroed pmap structure,
1725 * such as one in a vmspace structure.
1728 pmap_pinit_simple(struct pmap
*pmap
)
1731 * Misc initialization
1734 CPUMASK_ASSZERO(pmap
->pm_active
);
1735 pmap
->pm_pvhint
= NULL
;
1736 pmap
->pm_flags
= PMAP_FLAG_SIMPLE
;
1738 pmap_pinit_defaults(pmap
);
1741 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1744 if (pmap
->pm_pmlpv
== NULL
) {
1745 RB_INIT(&pmap
->pm_pvroot
);
1746 bzero(&pmap
->pm_stats
, sizeof pmap
->pm_stats
);
1747 spin_init(&pmap
->pm_spin
, "pmapinitsimple");
1748 lwkt_token_init(&pmap
->pm_token
, "pmap_tok");
1753 pmap_pinit(struct pmap
*pmap
)
1758 if (pmap
->pm_pmlpv
) {
1759 if (pmap
->pmap_bits
[TYPE_IDX
] != REGULAR_PMAP
) {
1764 pmap_pinit_simple(pmap
);
1765 pmap
->pm_flags
&= ~PMAP_FLAG_SIMPLE
;
1768 * No need to allocate page table space yet but we do need a valid
1769 * page directory table.
1771 if (pmap
->pm_pml4
== NULL
) {
1773 (pml4_entry_t
*)kmem_alloc_pageable(&kernel_map
, PAGE_SIZE
);
1777 * Allocate the page directory page, which wires it even though
1778 * it isn't being entered into some higher level page table (it
1779 * being the highest level). If one is already cached we don't
1780 * have to do anything.
1782 if ((pv
= pmap
->pm_pmlpv
) == NULL
) {
1783 pv
= pmap_allocpte(pmap
, pmap_pml4_pindex(), NULL
);
1784 pmap
->pm_pmlpv
= pv
;
1785 pmap_kenter((vm_offset_t
)pmap
->pm_pml4
,
1786 VM_PAGE_TO_PHYS(pv
->pv_m
));
1790 * Install DMAP and KMAP.
1792 for (j
= 0; j
< NDMPML4E
; ++j
) {
1793 pmap
->pm_pml4
[DMPML4I
+ j
] =
1794 (DMPDPphys
+ ((vm_paddr_t
)j
<< PML4SHIFT
)) |
1795 pmap
->pmap_bits
[PG_RW_IDX
] |
1796 pmap
->pmap_bits
[PG_V_IDX
] |
1797 pmap
->pmap_bits
[PG_U_IDX
];
1799 pmap
->pm_pml4
[KPML4I
] = KPDPphys
|
1800 pmap
->pmap_bits
[PG_RW_IDX
] |
1801 pmap
->pmap_bits
[PG_V_IDX
] |
1802 pmap
->pmap_bits
[PG_U_IDX
];
1805 * install self-referential address mapping entry
1807 pmap
->pm_pml4
[PML4PML4I
] = VM_PAGE_TO_PHYS(pv
->pv_m
) |
1808 pmap
->pmap_bits
[PG_V_IDX
] |
1809 pmap
->pmap_bits
[PG_RW_IDX
] |
1810 pmap
->pmap_bits
[PG_A_IDX
] |
1811 pmap
->pmap_bits
[PG_M_IDX
];
1813 KKASSERT(pv
->pv_m
->flags
& PG_MAPPED
);
1814 KKASSERT(pv
->pv_m
->flags
& PG_WRITEABLE
);
1816 KKASSERT(pmap
->pm_pml4
[255] == 0);
1817 KKASSERT(RB_ROOT(&pmap
->pm_pvroot
) == pv
);
1818 KKASSERT(pv
->pv_entry
.rbe_left
== NULL
);
1819 KKASSERT(pv
->pv_entry
.rbe_right
== NULL
);
1823 * Clean up a pmap structure so it can be physically freed. This routine
1824 * is called by the vmspace dtor function. A great deal of pmap data is
1825 * left passively mapped to improve vmspace management so we have a bit
1826 * of cleanup work to do here.
1829 pmap_puninit(pmap_t pmap
)
1834 KKASSERT(CPUMASK_TESTZERO(pmap
->pm_active
));
1835 if ((pv
= pmap
->pm_pmlpv
) != NULL
) {
1836 if (pv_hold_try(pv
) == 0)
1838 KKASSERT(pv
== pmap
->pm_pmlpv
);
1839 p
= pmap_remove_pv_page(pv
);
1841 pmap_kremove((vm_offset_t
)pmap
->pm_pml4
);
1842 vm_page_busy_wait(p
, FALSE
, "pgpun");
1843 KKASSERT(p
->flags
& (PG_FICTITIOUS
|PG_UNMANAGED
));
1844 vm_page_unwire(p
, 0);
1845 vm_page_flag_clear(p
, PG_MAPPED
| PG_WRITEABLE
);
1848 * XXX eventually clean out PML4 static entries and
1849 * use vm_page_free_zero()
1852 pmap
->pm_pmlpv
= NULL
;
1854 if (pmap
->pm_pml4
) {
1855 KKASSERT(pmap
->pm_pml4
!= (void *)(PTOV_OFFSET
+ KPML4phys
));
1856 kmem_free(&kernel_map
, (vm_offset_t
)pmap
->pm_pml4
, PAGE_SIZE
);
1857 pmap
->pm_pml4
= NULL
;
1859 KKASSERT(pmap
->pm_stats
.resident_count
== 0);
1860 KKASSERT(pmap
->pm_stats
.wired_count
== 0);
1864 * Wire in kernel global address entries. To avoid a race condition
1865 * between pmap initialization and pmap_growkernel, this procedure
1866 * adds the pmap to the master list (which growkernel scans to update),
1867 * then copies the template.
1870 pmap_pinit2(struct pmap
*pmap
)
1872 spin_lock(&pmap_spin
);
1873 TAILQ_INSERT_TAIL(&pmap_list
, pmap
, pm_pmnode
);
1874 spin_unlock(&pmap_spin
);
1878 * This routine is called when various levels in the page table need to
1879 * be populated. This routine cannot fail.
1881 * This function returns two locked pv_entry's, one representing the
1882 * requested pv and one representing the requested pv's parent pv. If
1883 * the pv did not previously exist it will be mapped into its parent
1884 * and wired, otherwise no additional wire count will be added.
1888 pmap_allocpte(pmap_t pmap
, vm_pindex_t ptepindex
, pv_entry_t
*pvpp
)
1893 vm_pindex_t pt_pindex
;
1899 * If the pv already exists and we aren't being asked for the
1900 * parent page table page we can just return it. A locked+held pv
1901 * is returned. The pv will also have a second hold related to the
1902 * pmap association that we don't have to worry about.
1905 pv
= pv_alloc(pmap
, ptepindex
, &isnew
);
1906 if (isnew
== 0 && pvpp
== NULL
)
1910 * Special case terminal PVs. These are not page table pages so
1911 * no vm_page is allocated (the caller supplied the vm_page). If
1912 * pvpp is non-NULL we are being asked to also removed the pt_pv
1915 * Note that pt_pv's are only returned for user VAs. We assert that
1916 * a pt_pv is not being requested for kernel VAs.
1918 if (ptepindex
< pmap_pt_pindex(0)) {
1919 if (ptepindex
>= NUPTE_USER
)
1920 KKASSERT(pvpp
== NULL
);
1922 KKASSERT(pvpp
!= NULL
);
1924 pt_pindex
= NUPTE_TOTAL
+ (ptepindex
>> NPTEPGSHIFT
);
1925 pvp
= pmap_allocpte(pmap
, pt_pindex
, NULL
);
1927 vm_page_wire_quick(pvp
->pv_m
);
1936 * Non-terminal PVs allocate a VM page to represent the page table,
1937 * so we have to resolve pvp and calculate ptepindex for the pvp
1938 * and then for the page table entry index in the pvp for
1941 if (ptepindex
< pmap_pd_pindex(0)) {
1943 * pv is PT, pvp is PD
1945 ptepindex
= (ptepindex
- pmap_pt_pindex(0)) >> NPDEPGSHIFT
;
1946 ptepindex
+= NUPTE_TOTAL
+ NUPT_TOTAL
;
1947 pvp
= pmap_allocpte(pmap
, ptepindex
, NULL
);
1954 ptepindex
= pv
->pv_pindex
- pmap_pt_pindex(0);
1955 ptepindex
&= ((1ul << NPDEPGSHIFT
) - 1);
1957 } else if (ptepindex
< pmap_pdp_pindex(0)) {
1959 * pv is PD, pvp is PDP
1961 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1964 ptepindex
= (ptepindex
- pmap_pd_pindex(0)) >> NPDPEPGSHIFT
;
1965 ptepindex
+= NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
;
1967 if (pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) {
1968 KKASSERT(pvpp
== NULL
);
1971 pvp
= pmap_allocpte(pmap
, ptepindex
, NULL
);
1979 ptepindex
= pv
->pv_pindex
- pmap_pd_pindex(0);
1980 ptepindex
&= ((1ul << NPDPEPGSHIFT
) - 1);
1981 } else if (ptepindex
< pmap_pml4_pindex()) {
1983 * pv is PDP, pvp is the root pml4 table
1985 pvp
= pmap_allocpte(pmap
, pmap_pml4_pindex(), NULL
);
1992 ptepindex
= pv
->pv_pindex
- pmap_pdp_pindex(0);
1993 ptepindex
&= ((1ul << NPML4EPGSHIFT
) - 1);
1996 * pv represents the top-level PML4, there is no parent.
2004 * This code is only reached if isnew is TRUE and this is not a
2005 * terminal PV. We need to allocate a vm_page for the page table
2006 * at this level and enter it into the parent page table.
2008 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2011 m
= vm_page_alloc(NULL
, pv
->pv_pindex
,
2012 VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
|
2013 VM_ALLOC_INTERRUPT
);
2018 vm_page_spin_lock(m
);
2019 pmap_page_stats_adding(m
);
2020 TAILQ_INSERT_TAIL(&m
->md
.pv_list
, pv
, pv_list
);
2022 vm_page_flag_set(m
, PG_MAPPED
| PG_WRITEABLE
);
2023 vm_page_spin_unlock(m
);
2024 vm_page_unmanage(m
); /* m must be spinunlocked */
2026 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
2027 m
->valid
= VM_PAGE_BITS_ALL
;
2028 vm_page_wire(m
); /* wire for mapping in parent */
2031 * Wire the page into pvp, bump the wire-count for pvp's page table
2032 * page. Bump the resident_count for the pmap. There is no pvp
2033 * for the top level, address the pm_pml4[] array directly.
2035 * If the caller wants the parent we return it, otherwise
2036 * we just put it away.
2038 * No interlock is needed for pte 0 -> non-zero.
2040 * In the situation where *ptep is valid we might have an unmanaged
2041 * page table page shared from another page table which we need to
2042 * unshare before installing our private page table page.
2045 ptep
= pv_pte_lookup(pvp
, ptepindex
);
2046 if (*ptep
& pmap
->pmap_bits
[PG_V_IDX
]) {
2050 panic("pmap_allocpte: unexpected pte %p/%d",
2051 pvp
, (int)ptepindex
);
2053 pte
= pmap_inval_smp(pmap
, (vm_offset_t
)-1, 1, ptep
, 0);
2054 if (vm_page_unwire_quick(
2055 PHYS_TO_VM_PAGE(pte
& PG_FRAME
))) {
2056 panic("pmap_allocpte: shared pgtable "
2057 "pg bad wirecount");
2059 atomic_add_long(&pmap
->pm_stats
.resident_count
, -1);
2061 vm_page_wire_quick(pvp
->pv_m
);
2063 *ptep
= VM_PAGE_TO_PHYS(m
) |
2064 (pmap
->pmap_bits
[PG_U_IDX
] |
2065 pmap
->pmap_bits
[PG_RW_IDX
] |
2066 pmap
->pmap_bits
[PG_V_IDX
] |
2067 pmap
->pmap_bits
[PG_A_IDX
] |
2068 pmap
->pmap_bits
[PG_M_IDX
]);
2080 * This version of pmap_allocpte() checks for possible segment optimizations
2081 * that would allow page-table sharing. It can be called for terminal
2082 * page or page table page ptepindex's.
2084 * The function is called with page table page ptepindex's for fictitious
2085 * and unmanaged terminal pages. That is, we don't want to allocate a
2086 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
2089 * This function can return a pv and *pvpp associated with the passed in pmap
2090 * OR a pv and *pvpp associated with the shared pmap. In the latter case
2091 * an unmanaged page table page will be entered into the pass in pmap.
2095 pmap_allocpte_seg(pmap_t pmap
, vm_pindex_t ptepindex
, pv_entry_t
*pvpp
,
2096 vm_map_entry_t entry
, vm_offset_t va
)
2102 pv_entry_t pte_pv
; /* in original or shared pmap */
2103 pv_entry_t pt_pv
; /* in original or shared pmap */
2104 pv_entry_t proc_pd_pv
; /* in original pmap */
2105 pv_entry_t proc_pt_pv
; /* in original pmap */
2106 pv_entry_t xpv
; /* PT in shared pmap */
2107 pd_entry_t
*pt
; /* PT entry in PD of original pmap */
2108 pd_entry_t opte
; /* contents of *pt */
2109 pd_entry_t npte
; /* contents of *pt */
2114 * Basic tests, require a non-NULL vm_map_entry, require proper
2115 * alignment and type for the vm_map_entry, require that the
2116 * underlying object already be allocated.
2118 * We allow almost any type of object to use this optimization.
2119 * The object itself does NOT have to be sized to a multiple of the
2120 * segment size, but the memory mapping does.
2122 * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2123 * won't work as expected.
2125 if (entry
== NULL
||
2126 pmap_mmu_optimize
== 0 || /* not enabled */
2127 (pmap
->pm_flags
& PMAP_HVM
) || /* special pmap */
2128 ptepindex
>= pmap_pd_pindex(0) || /* not terminal or pt */
2129 entry
->inheritance
!= VM_INHERIT_SHARE
|| /* not shared */
2130 entry
->maptype
!= VM_MAPTYPE_NORMAL
|| /* weird map type */
2131 entry
->object
.vm_object
== NULL
|| /* needs VM object */
2132 entry
->object
.vm_object
->type
== OBJT_DEVICE
|| /* ick */
2133 entry
->object
.vm_object
->type
== OBJT_MGTDEVICE
|| /* ick */
2134 (entry
->offset
& SEG_MASK
) || /* must be aligned */
2135 (entry
->start
& SEG_MASK
)) {
2136 return(pmap_allocpte(pmap
, ptepindex
, pvpp
));
2140 * Make sure the full segment can be represented.
2142 b
= va
& ~(vm_offset_t
)SEG_MASK
;
2143 if (b
< entry
->start
|| b
+ SEG_SIZE
> entry
->end
)
2144 return(pmap_allocpte(pmap
, ptepindex
, pvpp
));
2147 * If the full segment can be represented dive the VM object's
2148 * shared pmap, allocating as required.
2150 object
= entry
->object
.vm_object
;
2152 if (entry
->protection
& VM_PROT_WRITE
)
2153 obpmapp
= &object
->md
.pmap_rw
;
2155 obpmapp
= &object
->md
.pmap_ro
;
2158 if (pmap_enter_debug
> 0) {
2160 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2162 va
, entry
->protection
, object
,
2164 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2165 entry
, entry
->start
, entry
->end
);
2170 * We allocate what appears to be a normal pmap but because portions
2171 * of this pmap are shared with other unrelated pmaps we have to
2172 * set pm_active to point to all cpus.
2174 * XXX Currently using pmap_spin to interlock the update, can't use
2175 * vm_object_hold/drop because the token might already be held
2176 * shared OR exclusive and we don't know.
2178 while ((obpmap
= *obpmapp
) == NULL
) {
2179 obpmap
= kmalloc(sizeof(*obpmap
), M_OBJPMAP
, M_WAITOK
|M_ZERO
);
2180 pmap_pinit_simple(obpmap
);
2181 pmap_pinit2(obpmap
);
2182 spin_lock(&pmap_spin
);
2183 if (*obpmapp
!= NULL
) {
2187 spin_unlock(&pmap_spin
);
2188 pmap_release(obpmap
);
2189 pmap_puninit(obpmap
);
2190 kfree(obpmap
, M_OBJPMAP
);
2191 obpmap
= *obpmapp
; /* safety */
2193 obpmap
->pm_active
= smp_active_mask
;
2194 obpmap
->pm_flags
|= PMAP_SEGSHARED
;
2196 spin_unlock(&pmap_spin
);
2201 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
2202 * pte/pt using the shared pmap from the object but also adjust
2203 * the process pmap's page table page as a side effect.
2207 * Resolve the terminal PTE and PT in the shared pmap. This is what
2208 * we will return. This is true if ptepindex represents a terminal
2209 * page, otherwise pte_pv is actually the PT and pt_pv is actually
2213 pte_pv
= pmap_allocpte(obpmap
, ptepindex
, &pt_pv
);
2214 if (ptepindex
>= pmap_pt_pindex(0))
2220 * Resolve the PD in the process pmap so we can properly share the
2221 * page table page. Lock order is bottom-up (leaf first)!
2223 * NOTE: proc_pt_pv can be NULL.
2225 proc_pt_pv
= pv_get(pmap
, pmap_pt_pindex(b
));
2226 proc_pd_pv
= pmap_allocpte(pmap
, pmap_pd_pindex(b
), NULL
);
2228 if (pmap_enter_debug
> 0) {
2230 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2232 (proc_pt_pv
? proc_pt_pv
->pv_m
->wire_count
: -1),
2239 * xpv is the page table page pv from the shared object
2240 * (for convenience), from above.
2242 * Calculate the pte value for the PT to load into the process PD.
2243 * If we have to change it we must properly dispose of the previous
2246 pt
= pv_pte_lookup(proc_pd_pv
, pmap_pt_index(b
));
2247 npte
= VM_PAGE_TO_PHYS(xpv
->pv_m
) |
2248 (pmap
->pmap_bits
[PG_U_IDX
] |
2249 pmap
->pmap_bits
[PG_RW_IDX
] |
2250 pmap
->pmap_bits
[PG_V_IDX
] |
2251 pmap
->pmap_bits
[PG_A_IDX
] |
2252 pmap
->pmap_bits
[PG_M_IDX
]);
2255 * Dispose of previous page table page if it was local to the
2256 * process pmap. If the old pt is not empty we cannot dispose of it
2257 * until we clean it out. This case should not arise very often so
2258 * it is not optimized.
2261 pmap_inval_bulk_t bulk
;
2263 if (proc_pt_pv
->pv_m
->wire_count
!= 1) {
2269 va
& ~(vm_offset_t
)SEG_MASK
,
2270 (va
+ SEG_SIZE
) & ~(vm_offset_t
)SEG_MASK
);
2275 * The release call will indirectly clean out *pt
2277 pmap_inval_bulk_init(&bulk
, proc_pt_pv
->pv_pmap
);
2278 pmap_release_pv(proc_pt_pv
, proc_pd_pv
, &bulk
);
2279 pmap_inval_bulk_flush(&bulk
);
2282 pt
= pv_pte_lookup(proc_pd_pv
, pmap_pt_index(b
));
2286 * Handle remaining cases.
2290 vm_page_wire_quick(xpv
->pv_m
);
2291 vm_page_wire_quick(proc_pd_pv
->pv_m
);
2292 atomic_add_long(&pmap
->pm_stats
.resident_count
, 1);
2293 } else if (*pt
!= npte
) {
2294 opte
= pmap_inval_smp(pmap
, (vm_offset_t
)-1, 1, pt
, npte
);
2297 opte
= pte_load_clear(pt
);
2298 KKASSERT(opte
&& opte
!= npte
);
2302 vm_page_wire_quick(xpv
->pv_m
); /* pgtable pg that is npte */
2305 * Clean up opte, bump the wire_count for the process
2306 * PD page representing the new entry if it was
2309 * If the entry was not previously empty and we have
2310 * a PT in the proc pmap then opte must match that
2311 * pt. The proc pt must be retired (this is done
2312 * later on in this procedure).
2314 * NOTE: replacing valid pte, wire_count on proc_pd_pv
2317 KKASSERT(opte
& pmap
->pmap_bits
[PG_V_IDX
]);
2318 m
= PHYS_TO_VM_PAGE(opte
& PG_FRAME
);
2319 if (vm_page_unwire_quick(m
)) {
2320 panic("pmap_allocpte_seg: "
2321 "bad wire count %p",
2327 * The existing process page table was replaced and must be destroyed
2341 * Release any resources held by the given physical map.
2343 * Called when a pmap initialized by pmap_pinit is being released. Should
2344 * only be called if the map contains no valid mappings.
2346 * Caller must hold pmap->pm_token
2348 struct pmap_release_info
{
2353 static int pmap_release_callback(pv_entry_t pv
, void *data
);
2356 pmap_release(struct pmap
*pmap
)
2358 struct pmap_release_info info
;
2360 KASSERT(CPUMASK_TESTZERO(pmap
->pm_active
),
2361 ("pmap still active! %016jx",
2362 (uintmax_t)CPUMASK_LOWMASK(pmap
->pm_active
)));
2364 spin_lock(&pmap_spin
);
2365 TAILQ_REMOVE(&pmap_list
, pmap
, pm_pmnode
);
2366 spin_unlock(&pmap_spin
);
2369 * Pull pv's off the RB tree in order from low to high and release
2375 spin_lock(&pmap
->pm_spin
);
2376 RB_SCAN(pv_entry_rb_tree
, &pmap
->pm_pvroot
, NULL
,
2377 pmap_release_callback
, &info
);
2378 spin_unlock(&pmap
->pm_spin
);
2379 } while (info
.retry
);
2383 * One resident page (the pml4 page) should remain.
2384 * No wired pages should remain.
2386 KKASSERT(pmap
->pm_stats
.resident_count
==
2387 ((pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) ? 0 : 1));
2389 KKASSERT(pmap
->pm_stats
.wired_count
== 0);
2393 pmap_release_callback(pv_entry_t pv
, void *data
)
2395 struct pmap_release_info
*info
= data
;
2396 pmap_t pmap
= info
->pmap
;
2399 if (pv_hold_try(pv
)) {
2400 spin_unlock(&pmap
->pm_spin
);
2402 spin_unlock(&pmap
->pm_spin
);
2405 if (pv
->pv_pmap
!= pmap
) {
2407 spin_lock(&pmap
->pm_spin
);
2411 r
= pmap_release_pv(pv
, NULL
, NULL
);
2412 spin_lock(&pmap
->pm_spin
);
2417 * Called with held (i.e. also locked) pv. This function will dispose of
2418 * the lock along with the pv.
2420 * If the caller already holds the locked parent page table for pv it
2421 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2422 * pass NULL for pvp.
2425 pmap_release_pv(pv_entry_t pv
, pv_entry_t pvp
, pmap_inval_bulk_t
*bulk
)
2430 * The pmap is currently not spinlocked, pv is held+locked.
2431 * Remove the pv's page from its parent's page table. The
2432 * parent's page table page's wire_count will be decremented.
2434 * This will clean out the pte at any level of the page table.
2435 * If smp != 0 all cpus are affected.
2437 pmap_remove_pv_pte(pv
, pvp
, bulk
);
2440 * Terminal pvs are unhooked from their vm_pages. Because
2441 * terminal pages aren't page table pages they aren't wired
2442 * by us, so we have to be sure not to unwire them either.
2444 if (pv
->pv_pindex
< pmap_pt_pindex(0)) {
2445 pmap_remove_pv_page(pv
);
2450 * We leave the top-level page table page cached, wired, and
2451 * mapped in the pmap until the dtor function (pmap_puninit())
2454 * Since we are leaving the top-level pv intact we need
2455 * to break out of what would otherwise be an infinite loop.
2457 if (pv
->pv_pindex
== pmap_pml4_pindex()) {
2463 * For page table pages (other than the top-level page),
2464 * remove and free the vm_page. The representitive mapping
2465 * removed above by pmap_remove_pv_pte() did not undo the
2466 * last wire_count so we have to do that as well.
2468 p
= pmap_remove_pv_page(pv
);
2469 vm_page_busy_wait(p
, FALSE
, "pmaprl");
2470 if (p
->wire_count
!= 1) {
2471 kprintf("p->wire_count was %016lx %d\n",
2472 pv
->pv_pindex
, p
->wire_count
);
2474 KKASSERT(p
->wire_count
== 1);
2475 KKASSERT(p
->flags
& PG_UNMANAGED
);
2477 vm_page_unwire(p
, 0);
2478 KKASSERT(p
->wire_count
== 0);
2487 * This function will remove the pte associated with a pv from its parent.
2488 * Terminal pv's are supported. All cpus are affected if smp != 0.
2490 * The wire count will be dropped on the parent page table. The wire
2491 * count on the page being removed (pv->pv_m) from the parent page table
2492 * is NOT touched. Note that terminal pages will not have any additional
2493 * wire counts while page table pages will have at least one representing
2494 * the mapping, plus others representing sub-mappings.
2496 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2497 * pages and user page table and terminal pages.
2499 * The pv must be locked.
2501 * XXX must lock parent pv's if they exist to remove pte XXX
2505 pmap_remove_pv_pte(pv_entry_t pv
, pv_entry_t pvp
, pmap_inval_bulk_t
*bulk
)
2507 vm_pindex_t ptepindex
= pv
->pv_pindex
;
2508 pmap_t pmap
= pv
->pv_pmap
;
2514 if (ptepindex
== pmap_pml4_pindex()) {
2516 * We are the top level pml4 table, there is no parent.
2518 p
= pmap
->pm_pmlpv
->pv_m
;
2519 } else if (ptepindex
>= pmap_pdp_pindex(0)) {
2521 * Remove a PDP page from the pml4e. This can only occur
2522 * with user page tables. We do not have to lock the
2523 * pml4 PV so just ignore pvp.
2525 vm_pindex_t pml4_pindex
;
2526 vm_pindex_t pdp_index
;
2529 pdp_index
= ptepindex
- pmap_pdp_pindex(0);
2531 pml4_pindex
= pmap_pml4_pindex();
2532 pvp
= pv_get(pv
->pv_pmap
, pml4_pindex
);
2536 pdp
= &pmap
->pm_pml4
[pdp_index
& ((1ul << NPML4EPGSHIFT
) - 1)];
2537 KKASSERT((*pdp
& pmap
->pmap_bits
[PG_V_IDX
]) != 0);
2538 p
= PHYS_TO_VM_PAGE(*pdp
& PG_FRAME
);
2539 pmap_inval_bulk(bulk
, (vm_offset_t
)-1, pdp
, 0);
2540 } else if (ptepindex
>= pmap_pd_pindex(0)) {
2542 * Remove a PD page from the pdp
2544 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2545 * of a simple pmap because it stops at
2548 vm_pindex_t pdp_pindex
;
2549 vm_pindex_t pd_index
;
2552 pd_index
= ptepindex
- pmap_pd_pindex(0);
2555 pdp_pindex
= NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
+
2556 (pd_index
>> NPML4EPGSHIFT
);
2557 pvp
= pv_get(pv
->pv_pmap
, pdp_pindex
);
2562 pd
= pv_pte_lookup(pvp
, pd_index
&
2563 ((1ul << NPDPEPGSHIFT
) - 1));
2564 KKASSERT((*pd
& pmap
->pmap_bits
[PG_V_IDX
]) != 0);
2565 p
= PHYS_TO_VM_PAGE(*pd
& PG_FRAME
);
2566 pmap_inval_bulk(bulk
, (vm_offset_t
)-1, pd
, 0);
2568 KKASSERT(pmap
->pm_flags
& PMAP_FLAG_SIMPLE
);
2569 p
= pv
->pv_m
; /* degenerate test later */
2571 } else if (ptepindex
>= pmap_pt_pindex(0)) {
2573 * Remove a PT page from the pd
2575 vm_pindex_t pd_pindex
;
2576 vm_pindex_t pt_index
;
2579 pt_index
= ptepindex
- pmap_pt_pindex(0);
2582 pd_pindex
= NUPTE_TOTAL
+ NUPT_TOTAL
+
2583 (pt_index
>> NPDPEPGSHIFT
);
2584 pvp
= pv_get(pv
->pv_pmap
, pd_pindex
);
2588 pt
= pv_pte_lookup(pvp
, pt_index
& ((1ul << NPDPEPGSHIFT
) - 1));
2589 KKASSERT((*pt
& pmap
->pmap_bits
[PG_V_IDX
]) != 0);
2590 p
= PHYS_TO_VM_PAGE(*pt
& PG_FRAME
);
2591 pmap_inval_bulk(bulk
, (vm_offset_t
)-1, pt
, 0);
2594 * Remove a PTE from the PT page
2596 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2597 * pv is a pte_pv so we can safely lock pt_pv.
2599 * NOTE: FICTITIOUS pages may have multiple physical mappings
2600 * so PHYS_TO_VM_PAGE() will not necessarily work for
2603 vm_pindex_t pt_pindex
;
2608 pt_pindex
= ptepindex
>> NPTEPGSHIFT
;
2609 va
= (vm_offset_t
)ptepindex
<< PAGE_SHIFT
;
2611 if (ptepindex
>= NUPTE_USER
) {
2612 ptep
= vtopte(ptepindex
<< PAGE_SHIFT
);
2613 KKASSERT(pvp
== NULL
);
2616 pt_pindex
= NUPTE_TOTAL
+
2617 (ptepindex
>> NPDPEPGSHIFT
);
2618 pvp
= pv_get(pv
->pv_pmap
, pt_pindex
);
2622 ptep
= pv_pte_lookup(pvp
, ptepindex
&
2623 ((1ul << NPDPEPGSHIFT
) - 1));
2625 pte
= pmap_inval_bulk(bulk
, va
, ptep
, 0);
2626 if (bulk
== NULL
) /* XXX */
2627 cpu_invlpg((void *)va
); /* XXX */
2630 * Now update the vm_page_t
2632 if ((pte
& (pmap
->pmap_bits
[PG_MANAGED_IDX
] | pmap
->pmap_bits
[PG_V_IDX
])) !=
2633 (pmap
->pmap_bits
[PG_MANAGED_IDX
]|pmap
->pmap_bits
[PG_V_IDX
])) {
2634 kprintf("remove_pte badpte %016lx %016lx %d\n",
2636 pv
->pv_pindex
< pmap_pt_pindex(0));
2638 /* PHYS_TO_VM_PAGE() will not work for FICTITIOUS pages */
2639 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2640 if (pte
& pmap
->pmap_bits
[PG_DEVICE_IDX
])
2643 p
= PHYS_TO_VM_PAGE(pte
& PG_FRAME
);
2646 if (pte
& pmap
->pmap_bits
[PG_M_IDX
]) {
2647 if (pmap_track_modified(ptepindex
))
2650 if (pte
& pmap
->pmap_bits
[PG_A_IDX
]) {
2651 vm_page_flag_set(p
, PG_REFERENCED
);
2653 if (pte
& pmap
->pmap_bits
[PG_W_IDX
])
2654 atomic_add_long(&pmap
->pm_stats
.wired_count
, -1);
2655 if (pte
& pmap
->pmap_bits
[PG_G_IDX
])
2656 cpu_invlpg((void *)va
);
2660 * Unwire the parent page table page. The wire_count cannot go below
2661 * 1 here because the parent page table page is itself still mapped.
2663 * XXX remove the assertions later.
2665 KKASSERT(pv
->pv_m
== p
);
2666 if (pvp
&& vm_page_unwire_quick(pvp
->pv_m
))
2667 panic("pmap_remove_pv_pte: Insufficient wire_count");
2674 * Remove the vm_page association to a pv. The pv must be locked.
2678 pmap_remove_pv_page(pv_entry_t pv
)
2684 vm_page_spin_lock(m
);
2686 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
2687 pmap_page_stats_deleting(m
);
2690 atomic_add_int(&m->object->agg_pv_list_count, -1);
2692 if (TAILQ_EMPTY(&m
->md
.pv_list
))
2693 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
2694 vm_page_spin_unlock(m
);
2699 * Grow the number of kernel page table entries, if needed.
2701 * This routine is always called to validate any address space
2702 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2703 * space below KERNBASE.
2706 pmap_growkernel(vm_offset_t kstart
, vm_offset_t kend
)
2709 vm_offset_t ptppaddr
;
2711 pd_entry_t
*pt
, newpt
;
2713 int update_kernel_vm_end
;
2716 * bootstrap kernel_vm_end on first real VM use
2718 if (kernel_vm_end
== 0) {
2719 kernel_vm_end
= VM_MIN_KERNEL_ADDRESS
;
2721 while ((*pmap_pt(&kernel_pmap
, kernel_vm_end
) & kernel_pmap
.pmap_bits
[PG_V_IDX
]) != 0) {
2722 kernel_vm_end
= (kernel_vm_end
+ PAGE_SIZE
* NPTEPG
) &
2723 ~(PAGE_SIZE
* NPTEPG
- 1);
2725 if (kernel_vm_end
- 1 >= kernel_map
.max_offset
) {
2726 kernel_vm_end
= kernel_map
.max_offset
;
2733 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
2734 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
2735 * do not want to force-fill 128G worth of page tables.
2737 if (kstart
< KERNBASE
) {
2738 if (kstart
> kernel_vm_end
)
2739 kstart
= kernel_vm_end
;
2740 KKASSERT(kend
<= KERNBASE
);
2741 update_kernel_vm_end
= 1;
2743 update_kernel_vm_end
= 0;
2746 kstart
= rounddown2(kstart
, PAGE_SIZE
* NPTEPG
);
2747 kend
= roundup2(kend
, PAGE_SIZE
* NPTEPG
);
2749 if (kend
- 1 >= kernel_map
.max_offset
)
2750 kend
= kernel_map
.max_offset
;
2752 while (kstart
< kend
) {
2753 pt
= pmap_pt(&kernel_pmap
, kstart
);
2755 /* We need a new PDP entry */
2756 nkpg
= vm_page_alloc(NULL
, nkpt
,
2759 VM_ALLOC_INTERRUPT
);
2761 panic("pmap_growkernel: no memory to grow "
2764 paddr
= VM_PAGE_TO_PHYS(nkpg
);
2765 pmap_zero_page(paddr
);
2766 newpd
= (pdp_entry_t
)
2768 kernel_pmap
.pmap_bits
[PG_V_IDX
] |
2769 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
2770 kernel_pmap
.pmap_bits
[PG_A_IDX
] |
2771 kernel_pmap
.pmap_bits
[PG_M_IDX
]);
2772 *pmap_pd(&kernel_pmap
, kstart
) = newpd
;
2774 continue; /* try again */
2776 if ((*pt
& kernel_pmap
.pmap_bits
[PG_V_IDX
]) != 0) {
2777 kstart
= (kstart
+ PAGE_SIZE
* NPTEPG
) &
2778 ~(PAGE_SIZE
* NPTEPG
- 1);
2779 if (kstart
- 1 >= kernel_map
.max_offset
) {
2780 kstart
= kernel_map
.max_offset
;
2787 * This index is bogus, but out of the way
2789 nkpg
= vm_page_alloc(NULL
, nkpt
,
2792 VM_ALLOC_INTERRUPT
);
2794 panic("pmap_growkernel: no memory to grow kernel");
2797 ptppaddr
= VM_PAGE_TO_PHYS(nkpg
);
2798 pmap_zero_page(ptppaddr
);
2799 newpt
= (pd_entry_t
) (ptppaddr
|
2800 kernel_pmap
.pmap_bits
[PG_V_IDX
] |
2801 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
2802 kernel_pmap
.pmap_bits
[PG_A_IDX
] |
2803 kernel_pmap
.pmap_bits
[PG_M_IDX
]);
2804 *pmap_pt(&kernel_pmap
, kstart
) = newpt
;
2807 kstart
= (kstart
+ PAGE_SIZE
* NPTEPG
) &
2808 ~(PAGE_SIZE
* NPTEPG
- 1);
2810 if (kstart
- 1 >= kernel_map
.max_offset
) {
2811 kstart
= kernel_map
.max_offset
;
2817 * Only update kernel_vm_end for areas below KERNBASE.
2819 if (update_kernel_vm_end
&& kernel_vm_end
< kstart
)
2820 kernel_vm_end
= kstart
;
2824 * Add a reference to the specified pmap.
2827 pmap_reference(pmap_t pmap
)
2830 lwkt_gettoken(&pmap
->pm_token
);
2832 lwkt_reltoken(&pmap
->pm_token
);
2836 /***************************************************
2837 * page management routines.
2838 ***************************************************/
2841 * Hold a pv without locking it
2844 pv_hold(pv_entry_t pv
)
2846 atomic_add_int(&pv
->pv_hold
, 1);
2850 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2851 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2854 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2855 * pv list via its page) must be held by the caller.
2858 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL
)
2863 * Critical path shortcut expects pv to already have one ref
2864 * (for the pv->pv_pmap).
2866 if (atomic_cmpset_int(&pv
->pv_hold
, 1, PV_HOLD_LOCKED
| 2)) {
2869 pv
->pv_line
= lineno
;
2875 count
= pv
->pv_hold
;
2877 if ((count
& PV_HOLD_LOCKED
) == 0) {
2878 if (atomic_cmpset_int(&pv
->pv_hold
, count
,
2879 (count
+ 1) | PV_HOLD_LOCKED
)) {
2882 pv
->pv_line
= lineno
;
2887 if (atomic_cmpset_int(&pv
->pv_hold
, count
, count
+ 1))
2895 * Drop a previously held pv_entry which could not be locked, allowing its
2898 * Must not be called with a spinlock held as we might zfree() the pv if it
2899 * is no longer associated with a pmap and this was the last hold count.
2902 pv_drop(pv_entry_t pv
)
2907 count
= pv
->pv_hold
;
2909 KKASSERT((count
& PV_HOLD_MASK
) > 0);
2910 KKASSERT((count
& (PV_HOLD_LOCKED
| PV_HOLD_MASK
)) !=
2911 (PV_HOLD_LOCKED
| 1));
2912 if (atomic_cmpset_int(&pv
->pv_hold
, count
, count
- 1)) {
2913 if ((count
& PV_HOLD_MASK
) == 1) {
2915 if (pmap_enter_debug
> 0) {
2917 kprintf("pv_drop: free pv %p\n", pv
);
2920 KKASSERT(count
== 1);
2921 KKASSERT(pv
->pv_pmap
== NULL
);
2931 * Find or allocate the requested PV entry, returning a locked, held pv.
2933 * If (*isnew) is non-zero, the returned pv will have two hold counts, one
2934 * for the caller and one representing the pmap and vm_page association.
2936 * If (*isnew) is zero, the returned pv will have only one hold count.
2938 * Since both associations can only be adjusted while the pv is locked,
2939 * together they represent just one additional hold.
2943 _pv_alloc(pmap_t pmap
, vm_pindex_t pindex
, int *isnew PMAP_DEBUG_DECL
)
2946 pv_entry_t pnew
= NULL
;
2948 spin_lock(&pmap
->pm_spin
);
2950 if ((pv
= pmap
->pm_pvhint
) == NULL
|| pv
->pv_pindex
!= pindex
) {
2951 pv
= pv_entry_rb_tree_RB_LOOKUP(&pmap
->pm_pvroot
,
2956 spin_unlock(&pmap
->pm_spin
);
2957 pnew
= zalloc(pvzone
);
2958 spin_lock(&pmap
->pm_spin
);
2961 pnew
->pv_pmap
= pmap
;
2962 pnew
->pv_pindex
= pindex
;
2963 pnew
->pv_hold
= PV_HOLD_LOCKED
| 2;
2965 pnew
->pv_func
= func
;
2966 pnew
->pv_line
= lineno
;
2968 pv_entry_rb_tree_RB_INSERT(&pmap
->pm_pvroot
, pnew
);
2969 ++pmap
->pm_generation
;
2970 atomic_add_long(&pmap
->pm_stats
.resident_count
, 1);
2971 spin_unlock(&pmap
->pm_spin
);
2976 spin_unlock(&pmap
->pm_spin
);
2977 zfree(pvzone
, pnew
);
2979 spin_lock(&pmap
->pm_spin
);
2982 if (_pv_hold_try(pv PMAP_DEBUG_COPY
)) {
2983 spin_unlock(&pmap
->pm_spin
);
2985 spin_unlock(&pmap
->pm_spin
);
2986 _pv_lock(pv PMAP_DEBUG_COPY
);
2988 if (pv
->pv_pmap
== pmap
&& pv
->pv_pindex
== pindex
) {
2993 spin_lock(&pmap
->pm_spin
);
2998 * Find the requested PV entry, returning a locked+held pv or NULL
3002 _pv_get(pmap_t pmap
, vm_pindex_t pindex PMAP_DEBUG_DECL
)
3006 spin_lock(&pmap
->pm_spin
);
3011 if ((pv
= pmap
->pm_pvhint
) == NULL
|| pv
->pv_pindex
!= pindex
) {
3012 pv
= pv_entry_rb_tree_RB_LOOKUP(&pmap
->pm_pvroot
,
3016 spin_unlock(&pmap
->pm_spin
);
3019 if (_pv_hold_try(pv PMAP_DEBUG_COPY
)) {
3020 spin_unlock(&pmap
->pm_spin
);
3022 spin_unlock(&pmap
->pm_spin
);
3023 _pv_lock(pv PMAP_DEBUG_COPY
);
3025 if (pv
->pv_pmap
== pmap
&& pv
->pv_pindex
== pindex
) {
3026 pv_cache(pv
, pindex
);
3030 spin_lock(&pmap
->pm_spin
);
3035 * Lookup, hold, and attempt to lock (pmap,pindex).
3037 * If the entry does not exist NULL is returned and *errorp is set to 0
3039 * If the entry exists and could be successfully locked it is returned and
3040 * errorp is set to 0.
3042 * If the entry exists but could NOT be successfully locked it is returned
3043 * held and *errorp is set to 1.
3047 pv_get_try(pmap_t pmap
, vm_pindex_t pindex
, int *errorp
)
3051 spin_lock_shared(&pmap
->pm_spin
);
3052 if ((pv
= pmap
->pm_pvhint
) == NULL
|| pv
->pv_pindex
!= pindex
)
3053 pv
= pv_entry_rb_tree_RB_LOOKUP(&pmap
->pm_pvroot
, pindex
);
3055 spin_unlock_shared(&pmap
->pm_spin
);
3059 if (pv_hold_try(pv
)) {
3060 pv_cache(pv
, pindex
);
3061 spin_unlock_shared(&pmap
->pm_spin
);
3063 KKASSERT(pv
->pv_pmap
== pmap
&& pv
->pv_pindex
== pindex
);
3064 return(pv
); /* lock succeeded */
3066 spin_unlock_shared(&pmap
->pm_spin
);
3068 return (pv
); /* lock failed */
3072 * Find the requested PV entry, returning a held pv or NULL
3076 pv_find(pmap_t pmap
, vm_pindex_t pindex
)
3080 spin_lock_shared(&pmap
->pm_spin
);
3082 if ((pv
= pmap
->pm_pvhint
) == NULL
|| pv
->pv_pindex
!= pindex
)
3083 pv
= pv_entry_rb_tree_RB_LOOKUP(&pmap
->pm_pvroot
, pindex
);
3085 spin_unlock_shared(&pmap
->pm_spin
);
3089 pv_cache(pv
, pindex
);
3090 spin_unlock_shared(&pmap
->pm_spin
);
3095 * Lock a held pv, keeping the hold count
3099 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL
)
3104 count
= pv
->pv_hold
;
3106 if ((count
& PV_HOLD_LOCKED
) == 0) {
3107 if (atomic_cmpset_int(&pv
->pv_hold
, count
,
3108 count
| PV_HOLD_LOCKED
)) {
3111 pv
->pv_line
= lineno
;
3117 tsleep_interlock(pv
, 0);
3118 if (atomic_cmpset_int(&pv
->pv_hold
, count
,
3119 count
| PV_HOLD_WAITING
)) {
3121 kprintf("pv waiting on %s:%d\n",
3122 pv
->pv_func
, pv
->pv_line
);
3124 tsleep(pv
, PINTERLOCKED
, "pvwait", hz
);
3131 * Unlock a held and locked pv, keeping the hold count.
3135 pv_unlock(pv_entry_t pv
)
3140 count
= pv
->pv_hold
;
3142 KKASSERT((count
& (PV_HOLD_LOCKED
| PV_HOLD_MASK
)) >=
3143 (PV_HOLD_LOCKED
| 1));
3144 if (atomic_cmpset_int(&pv
->pv_hold
, count
,
3146 ~(PV_HOLD_LOCKED
| PV_HOLD_WAITING
))) {
3147 if (count
& PV_HOLD_WAITING
)
3155 * Unlock and drop a pv. If the pv is no longer associated with a pmap
3156 * and the hold count drops to zero we will free it.
3158 * Caller should not hold any spin locks. We are protected from hold races
3159 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3160 * lock held. A pv cannot be located otherwise.
3164 pv_put(pv_entry_t pv
)
3167 if (pmap_enter_debug
> 0) {
3169 kprintf("pv_put pv=%p hold=%08x\n", pv
, pv
->pv_hold
);
3174 * Fast - shortcut most common condition
3176 if (atomic_cmpset_int(&pv
->pv_hold
, PV_HOLD_LOCKED
| 2, 1))
3187 * Remove the pmap association from a pv, require that pv_m already be removed,
3188 * then unlock and drop the pv. Any pte operations must have already been
3189 * completed. This call may result in a last-drop which will physically free
3192 * Removing the pmap association entails an additional drop.
3194 * pv must be exclusively locked on call and will be disposed of on return.
3198 pv_free(pv_entry_t pv
)
3202 KKASSERT(pv
->pv_m
== NULL
);
3203 KKASSERT((pv
->pv_hold
& PV_HOLD_MASK
) >= 2);
3204 if ((pmap
= pv
->pv_pmap
) != NULL
) {
3205 spin_lock(&pmap
->pm_spin
);
3206 pv_entry_rb_tree_RB_REMOVE(&pmap
->pm_pvroot
, pv
);
3207 ++pmap
->pm_generation
;
3208 if (pmap
->pm_pvhint
== pv
)
3209 pmap
->pm_pvhint
= NULL
;
3210 atomic_add_long(&pmap
->pm_stats
.resident_count
, -1);
3213 spin_unlock(&pmap
->pm_spin
);
3216 * Try to shortcut three atomic ops, otherwise fall through
3217 * and do it normally. Drop two refs and the lock all in
3220 if (atomic_cmpset_int(&pv
->pv_hold
, PV_HOLD_LOCKED
| 2, 0)) {
3222 if (pmap_enter_debug
> 0) {
3224 kprintf("pv_free: free pv %p\n", pv
);
3230 pv_drop(pv
); /* ref for pv_pmap */
3236 * This routine is very drastic, but can save the system
3244 static int warningdone
=0;
3246 if (pmap_pagedaemon_waken
== 0)
3248 pmap_pagedaemon_waken
= 0;
3249 if (warningdone
< 5) {
3250 kprintf("pmap_collect: collecting pv entries -- "
3251 "suggest increasing PMAP_SHPGPERPROC\n");
3255 for (i
= 0; i
< vm_page_array_size
; i
++) {
3256 m
= &vm_page_array
[i
];
3257 if (m
->wire_count
|| m
->hold_count
)
3259 if (vm_page_busy_try(m
, TRUE
) == 0) {
3260 if (m
->wire_count
== 0 && m
->hold_count
== 0) {
3269 * Scan the pmap for active page table entries and issue a callback.
3270 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3271 * its parent page table.
3273 * pte_pv will be NULL if the page or page table is unmanaged.
3274 * pt_pv will point to the page table page containing the pte for the page.
3276 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3277 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3278 * process pmap's PD and page to the callback function. This can be
3279 * confusing because the pt_pv is really a pd_pv, and the target page
3280 * table page is simply aliased by the pmap and not owned by it.
3282 * It is assumed that the start and end are properly rounded to the page size.
3284 * It is assumed that PD pages and above are managed and thus in the RB tree,
3285 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3287 struct pmap_scan_info
{
3291 vm_pindex_t sva_pd_pindex
;
3292 vm_pindex_t eva_pd_pindex
;
3293 void (*func
)(pmap_t
, struct pmap_scan_info
*,
3294 pv_entry_t
, pv_entry_t
, int, vm_offset_t
,
3295 pt_entry_t
*, void *);
3297 pmap_inval_bulk_t bulk_core
;
3298 pmap_inval_bulk_t
*bulk
;
3302 static int pmap_scan_cmp(pv_entry_t pv
, void *data
);
3303 static int pmap_scan_callback(pv_entry_t pv
, void *data
);
3306 pmap_scan(struct pmap_scan_info
*info
, int smp_inval
)
3308 struct pmap
*pmap
= info
->pmap
;
3309 pv_entry_t pd_pv
; /* A page directory PV */
3310 pv_entry_t pt_pv
; /* A page table PV */
3311 pv_entry_t pte_pv
; /* A page table entry PV */
3314 struct pv_entry dummy_pv
;
3320 info
->bulk
= &info
->bulk_core
;
3321 pmap_inval_bulk_init(&info
->bulk_core
, pmap
);
3327 * Hold the token for stability; if the pmap is empty we have nothing
3330 lwkt_gettoken(&pmap
->pm_token
);
3332 if (pmap
->pm_stats
.resident_count
== 0) {
3333 lwkt_reltoken(&pmap
->pm_token
);
3342 * Special handling for scanning one page, which is a very common
3343 * operation (it is?).
3345 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3347 if (info
->sva
+ PAGE_SIZE
== info
->eva
) {
3348 generation
= pmap
->pm_generation
;
3349 if (info
->sva
>= VM_MAX_USER_ADDRESS
) {
3351 * Kernel mappings do not track wire counts on
3352 * page table pages and only maintain pd_pv and
3353 * pte_pv levels so pmap_scan() works.
3356 pte_pv
= pv_get(pmap
, pmap_pte_pindex(info
->sva
));
3357 ptep
= vtopte(info
->sva
);
3360 * User pages which are unmanaged will not have a
3361 * pte_pv. User page table pages which are unmanaged
3362 * (shared from elsewhere) will also not have a pt_pv.
3363 * The func() callback will pass both pte_pv and pt_pv
3364 * as NULL in that case.
3366 pte_pv
= pv_get(pmap
, pmap_pte_pindex(info
->sva
));
3367 pt_pv
= pv_get(pmap
, pmap_pt_pindex(info
->sva
));
3368 if (pt_pv
== NULL
) {
3369 KKASSERT(pte_pv
== NULL
);
3370 pd_pv
= pv_get(pmap
, pmap_pd_pindex(info
->sva
));
3372 ptep
= pv_pte_lookup(pd_pv
,
3373 pmap_pt_index(info
->sva
));
3375 info
->func(pmap
, info
,
3384 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(info
->sva
));
3388 * NOTE: *ptep can't be ripped out from under us if we hold
3389 * pte_pv locked, but bits can change. However, there is
3390 * a race where another thread may be inserting pte_pv
3391 * and setting *ptep just after our pte_pv lookup fails.
3393 * In this situation we can end up with a NULL pte_pv
3394 * but find that we have a managed *ptep. We explicitly
3395 * check for this race.
3401 * Unlike the pv_find() case below we actually
3402 * acquired a locked pv in this case so any
3403 * race should have been resolved. It is expected
3406 KKASSERT(pte_pv
== NULL
);
3407 } else if (pte_pv
) {
3408 KASSERT((oldpte
& (pmap
->pmap_bits
[PG_MANAGED_IDX
] |
3409 pmap
->pmap_bits
[PG_V_IDX
])) ==
3410 (pmap
->pmap_bits
[PG_MANAGED_IDX
] |
3411 pmap
->pmap_bits
[PG_V_IDX
]),
3412 ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p"
3414 *ptep
, oldpte
, info
->sva
, pte_pv
,
3415 generation
, pmap
->pm_generation
));
3416 info
->func(pmap
, info
, pte_pv
, pt_pv
, 0,
3417 info
->sva
, ptep
, info
->arg
);
3420 * Check for insertion race
3422 if ((oldpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) &&
3424 pte_pv
= pv_find(pmap
,
3425 pmap_pte_pindex(info
->sva
));
3429 kprintf("pmap_scan: RACE1 "
3439 KASSERT((oldpte
& (pmap
->pmap_bits
[PG_MANAGED_IDX
] |
3440 pmap
->pmap_bits
[PG_V_IDX
])) ==
3441 pmap
->pmap_bits
[PG_V_IDX
],
3442 ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL"
3444 *ptep
, oldpte
, info
->sva
,
3445 generation
, pmap
->pm_generation
));
3446 info
->func(pmap
, info
, NULL
, pt_pv
, 0,
3447 info
->sva
, ptep
, info
->arg
);
3452 pmap_inval_bulk_flush(info
->bulk
);
3453 lwkt_reltoken(&pmap
->pm_token
);
3458 * Nominal scan case, RB_SCAN() for PD pages and iterate from
3461 info
->sva_pd_pindex
= pmap_pd_pindex(info
->sva
);
3462 info
->eva_pd_pindex
= pmap_pd_pindex(info
->eva
+ NBPDP
- 1);
3464 if (info
->sva
>= VM_MAX_USER_ADDRESS
) {
3466 * The kernel does not currently maintain any pv_entry's for
3467 * higher-level page tables.
3469 bzero(&dummy_pv
, sizeof(dummy_pv
));
3470 dummy_pv
.pv_pindex
= info
->sva_pd_pindex
;
3471 spin_lock(&pmap
->pm_spin
);
3472 while (dummy_pv
.pv_pindex
< info
->eva_pd_pindex
) {
3473 pmap_scan_callback(&dummy_pv
, info
);
3474 ++dummy_pv
.pv_pindex
;
3476 spin_unlock(&pmap
->pm_spin
);
3479 * User page tables maintain local PML4, PDP, and PD
3480 * pv_entry's at the very least. PT pv's might be
3481 * unmanaged and thus not exist. PTE pv's might be
3482 * unmanaged and thus not exist.
3484 spin_lock(&pmap
->pm_spin
);
3485 pv_entry_rb_tree_RB_SCAN(&pmap
->pm_pvroot
,
3486 pmap_scan_cmp
, pmap_scan_callback
, info
);
3487 spin_unlock(&pmap
->pm_spin
);
3489 pmap_inval_bulk_flush(info
->bulk
);
3490 lwkt_reltoken(&pmap
->pm_token
);
3494 * WARNING! pmap->pm_spin held
3497 pmap_scan_cmp(pv_entry_t pv
, void *data
)
3499 struct pmap_scan_info
*info
= data
;
3500 if (pv
->pv_pindex
< info
->sva_pd_pindex
)
3502 if (pv
->pv_pindex
>= info
->eva_pd_pindex
)
3508 * WARNING! pmap->pm_spin held
3511 pmap_scan_callback(pv_entry_t pv
, void *data
)
3513 struct pmap_scan_info
*info
= data
;
3514 struct pmap
*pmap
= info
->pmap
;
3515 pv_entry_t pd_pv
; /* A page directory PV */
3516 pv_entry_t pt_pv
; /* A page table PV */
3517 pv_entry_t pte_pv
; /* A page table entry PV */
3522 vm_offset_t va_next
;
3523 vm_pindex_t pd_pindex
;
3528 * Pull the PD pindex from the pv before releasing the spinlock.
3530 * WARNING: pv is faked for kernel pmap scans.
3532 pd_pindex
= pv
->pv_pindex
;
3533 spin_unlock(&pmap
->pm_spin
);
3534 pv
= NULL
; /* invalid after spinlock unlocked */
3537 * Calculate the page range within the PD. SIMPLE pmaps are
3538 * direct-mapped for the entire 2^64 address space. Normal pmaps
3539 * reflect the user and kernel address space which requires
3540 * cannonicalization w/regards to converting pd_pindex's back
3543 sva
= (pd_pindex
- NUPTE_TOTAL
- NUPT_TOTAL
) << PDPSHIFT
;
3544 if ((pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) == 0 &&
3545 (sva
& PML4_SIGNMASK
)) {
3546 sva
|= PML4_SIGNMASK
;
3548 eva
= sva
+ NBPDP
; /* can overflow */
3549 if (sva
< info
->sva
)
3551 if (eva
< info
->sva
|| eva
> info
->eva
)
3555 * NOTE: kernel mappings do not track page table pages, only
3558 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3559 * However, for the scan to be efficient we try to
3560 * cache items top-down.
3565 for (; sva
< eva
; sva
= va_next
) {
3566 if (sva
>= VM_MAX_USER_ADDRESS
) {
3575 * PD cache (degenerate case if we skip). It is possible
3576 * for the PD to not exist due to races. This is ok.
3578 if (pd_pv
== NULL
) {
3579 pd_pv
= pv_get(pmap
, pmap_pd_pindex(sva
));
3580 } else if (pd_pv
->pv_pindex
!= pmap_pd_pindex(sva
)) {
3582 pd_pv
= pv_get(pmap
, pmap_pd_pindex(sva
));
3584 if (pd_pv
== NULL
) {
3585 va_next
= (sva
+ NBPDP
) & ~PDPMASK
;
3594 if (pt_pv
== NULL
) {
3599 pt_pv
= pv_get(pmap
, pmap_pt_pindex(sva
));
3600 } else if (pt_pv
->pv_pindex
!= pmap_pt_pindex(sva
)) {
3606 pt_pv
= pv_get(pmap
, pmap_pt_pindex(sva
));
3610 * If pt_pv is NULL we either have an shared page table
3611 * page and must issue a callback specific to that case,
3612 * or there is no page table page.
3614 * Either way we can skip the page table page.
3616 if (pt_pv
== NULL
) {
3618 * Possible unmanaged (shared from another pmap)
3622 pd_pv
= pv_get(pmap
, pmap_pd_pindex(sva
));
3623 KKASSERT(pd_pv
!= NULL
);
3624 ptep
= pv_pte_lookup(pd_pv
, pmap_pt_index(sva
));
3625 if (*ptep
& pmap
->pmap_bits
[PG_V_IDX
]) {
3626 info
->func(pmap
, info
, NULL
, pd_pv
, 1,
3627 sva
, ptep
, info
->arg
);
3631 * Done, move to next page table page.
3633 va_next
= (sva
+ NBPDR
) & ~PDRMASK
;
3640 * From this point in the loop testing pt_pv for non-NULL
3641 * means we are in UVM, else if it is NULL we are in KVM.
3643 * Limit our scan to either the end of the va represented
3644 * by the current page table page, or to the end of the
3645 * range being removed.
3648 va_next
= (sva
+ NBPDR
) & ~PDRMASK
;
3655 * Scan the page table for pages. Some pages may not be
3656 * managed (might not have a pv_entry).
3658 * There is no page table management for kernel pages so
3659 * pt_pv will be NULL in that case, but otherwise pt_pv
3660 * is non-NULL, locked, and referenced.
3664 * At this point a non-NULL pt_pv means a UVA, and a NULL
3665 * pt_pv means a KVA.
3668 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(sva
));
3672 while (sva
< va_next
) {
3674 * Yield every 64 pages.
3676 if ((++info
->count
& 63) == 0)
3680 * Acquire the related pte_pv, if any. If *ptep == 0
3681 * the related pte_pv should not exist, but if *ptep
3682 * is not zero the pte_pv may or may not exist (e.g.
3683 * will not exist for an unmanaged page).
3685 * However a multitude of races are possible here.
3687 * In addition, the (pt_pv, pte_pv) lock order is
3688 * backwards, so we have to be careful in aquiring
3689 * a properly locked pte_pv.
3691 generation
= pmap
->pm_generation
;
3693 pte_pv
= pv_get_try(pmap
, pmap_pte_pindex(sva
),
3700 pv_put(pt_pv
); /* must be non-NULL */
3702 pv_lock(pte_pv
); /* safe to block now */
3705 pt_pv
= pv_get(pmap
,
3706 pmap_pt_pindex(sva
));
3708 * pt_pv reloaded, need new ptep
3710 KKASSERT(pt_pv
!= NULL
);
3711 ptep
= pv_pte_lookup(pt_pv
,
3712 pmap_pte_index(sva
));
3716 pte_pv
= pv_get(pmap
, pmap_pte_pindex(sva
));
3720 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
3725 kprintf("Unexpected non-NULL pte_pv "
3727 "*ptep = %016lx/%016lx\n",
3728 pte_pv
, pt_pv
, *ptep
, oldpte
);
3729 panic("Unexpected non-NULL pte_pv");
3737 * Ready for the callback. The locked pte_pv (if any)
3738 * is consumed by the callback. pte_pv will exist if
3739 * the page is managed, and will not exist if it
3743 KASSERT((oldpte
& (pmap
->pmap_bits
[PG_MANAGED_IDX
] | pmap
->pmap_bits
[PG_V_IDX
])) ==
3744 (pmap
->pmap_bits
[PG_MANAGED_IDX
] | pmap
->pmap_bits
[PG_V_IDX
]),
3745 ("badC *ptep %016lx/%016lx sva %016lx "
3746 "pte_pv %p pm_generation %d/%d",
3747 *ptep
, oldpte
, sva
, pte_pv
,
3748 generation
, pmap
->pm_generation
));
3749 info
->func(pmap
, info
, pte_pv
, pt_pv
, 0,
3750 sva
, ptep
, info
->arg
);
3753 * Check for insertion race. Since there is no
3754 * pte_pv to guard us it is possible for us
3755 * to race another thread doing an insertion.
3756 * Our lookup misses the pte_pv but our *ptep
3757 * check sees the inserted pte.
3759 * XXX panic case seems to occur within a
3760 * vm_fork() of /bin/sh, which frankly
3761 * shouldn't happen since no other threads
3762 * should be inserting to our pmap in that
3763 * situation. Removing, possibly. Inserting,
3766 if ((oldpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) &&
3768 pte_pv
= pv_find(pmap
,
3769 pmap_pte_pindex(sva
));
3772 kprintf("pmap_scan: RACE2 "
3782 KASSERT((oldpte
& (pmap
->pmap_bits
[PG_MANAGED_IDX
] | pmap
->pmap_bits
[PG_V_IDX
])) ==
3783 pmap
->pmap_bits
[PG_V_IDX
],
3784 ("badD *ptep %016lx/%016lx sva %016lx "
3785 "pte_pv NULL pm_generation %d/%d",
3787 generation
, pmap
->pm_generation
));
3788 info
->func(pmap
, info
, NULL
, pt_pv
, 0,
3789 sva
, ptep
, info
->arg
);
3804 if ((++info
->count
& 7) == 0)
3808 * Relock before returning.
3810 spin_lock(&pmap
->pm_spin
);
3815 pmap_remove(struct pmap
*pmap
, vm_offset_t sva
, vm_offset_t eva
)
3817 struct pmap_scan_info info
;
3822 info
.func
= pmap_remove_callback
;
3824 pmap_scan(&info
, 1);
3828 pmap_remove_noinval(struct pmap
*pmap
, vm_offset_t sva
, vm_offset_t eva
)
3830 struct pmap_scan_info info
;
3835 info
.func
= pmap_remove_callback
;
3837 pmap_scan(&info
, 0);
3841 pmap_remove_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
3842 pv_entry_t pte_pv
, pv_entry_t pt_pv
, int sharept
,
3843 vm_offset_t va
, pt_entry_t
*ptep
, void *arg __unused
)
3849 * This will also drop pt_pv's wire_count. Note that
3850 * terminal pages are not wired based on mmu presence.
3852 pmap_remove_pv_pte(pte_pv
, pt_pv
, info
->bulk
);
3853 pmap_remove_pv_page(pte_pv
);
3855 } else if (sharept
== 0) {
3857 * Unmanaged page table (pt, pd, or pdp. Not pte).
3859 * pt_pv's wire_count is still bumped by unmanaged pages
3860 * so we must decrement it manually.
3862 * We have to unwire the target page table page.
3864 * It is unclear how we can invalidate a segment so we
3865 * invalidate -1 which invlidates the tlb.
3867 pte
= pmap_inval_bulk(info
->bulk
, (vm_offset_t
)-1, ptep
, 0);
3868 if (pte
& pmap
->pmap_bits
[PG_W_IDX
])
3869 atomic_add_long(&pmap
->pm_stats
.wired_count
, -1);
3870 atomic_add_long(&pmap
->pm_stats
.resident_count
, -1);
3871 if (vm_page_unwire_quick(pt_pv
->pv_m
))
3872 panic("pmap_remove: insufficient wirecount");
3875 * Unmanaged page table (pt, pd, or pdp. Not pte) for
3876 * a shared page table.
3878 * pt_pv is actually the pd_pv for our pmap (not the shared
3881 * We have to unwire the target page table page and we
3882 * have to unwire our page directory page.
3884 * It is unclear how we can invalidate a segment so we
3885 * invalidate -1 which invlidates the tlb.
3887 pte
= pmap_inval_bulk(info
->bulk
, (vm_offset_t
)-1, ptep
, 0);
3888 atomic_add_long(&pmap
->pm_stats
.resident_count
, -1);
3889 KKASSERT((pte
& pmap
->pmap_bits
[PG_DEVICE_IDX
]) == 0);
3890 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte
& PG_FRAME
)))
3891 panic("pmap_remove: shared pgtable1 bad wirecount");
3892 if (vm_page_unwire_quick(pt_pv
->pv_m
))
3893 panic("pmap_remove: shared pgtable2 bad wirecount");
3898 * Removes this physical page from all physical maps in which it resides.
3899 * Reflects back modify bits to the pager.
3901 * This routine may not be called from an interrupt.
3905 pmap_remove_all(vm_page_t m
)
3908 pmap_inval_bulk_t bulk
;
3910 if (!pmap_initialized
/* || (m->flags & PG_FICTITIOUS)*/)
3913 vm_page_spin_lock(m
);
3914 while ((pv
= TAILQ_FIRST(&m
->md
.pv_list
)) != NULL
) {
3915 KKASSERT(pv
->pv_m
== m
);
3916 if (pv_hold_try(pv
)) {
3917 vm_page_spin_unlock(m
);
3919 vm_page_spin_unlock(m
);
3922 if (pv
->pv_m
!= m
) {
3924 vm_page_spin_lock(m
);
3929 * Holding no spinlocks, pv is locked.
3931 pmap_inval_bulk_init(&bulk
, pv
->pv_pmap
);
3932 pmap_remove_pv_pte(pv
, NULL
, &bulk
);
3933 pmap_inval_bulk_flush(&bulk
);
3934 pmap_remove_pv_page(pv
);
3936 vm_page_spin_lock(m
);
3938 KKASSERT((m
->flags
& (PG_MAPPED
|PG_WRITEABLE
)) == 0);
3939 vm_page_spin_unlock(m
);
3943 * Set the physical protection on the specified range of this map
3944 * as requested. This function is typically only used for debug watchpoints
3947 * This function may not be called from an interrupt if the map is
3948 * not the kernel_pmap.
3950 * NOTE! For shared page table pages we just unmap the page.
3953 pmap_protect(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
, vm_prot_t prot
)
3955 struct pmap_scan_info info
;
3956 /* JG review for NX */
3960 if ((prot
& VM_PROT_READ
) == VM_PROT_NONE
) {
3961 pmap_remove(pmap
, sva
, eva
);
3964 if (prot
& VM_PROT_WRITE
)
3969 info
.func
= pmap_protect_callback
;
3971 pmap_scan(&info
, 1);
3976 pmap_protect_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
3977 pv_entry_t pte_pv
, pv_entry_t pt_pv
, int sharept
,
3978 vm_offset_t va
, pt_entry_t
*ptep
, void *arg __unused
)
3990 if (pbits
& pmap
->pmap_bits
[PG_A_IDX
]) {
3991 if ((pbits
& pmap
->pmap_bits
[PG_DEVICE_IDX
]) == 0) {
3992 m
= PHYS_TO_VM_PAGE(pbits
& PG_FRAME
);
3993 KKASSERT(m
== pte_pv
->pv_m
);
3994 vm_page_flag_set(m
, PG_REFERENCED
);
3996 cbits
&= ~pmap
->pmap_bits
[PG_A_IDX
];
3998 if (pbits
& pmap
->pmap_bits
[PG_M_IDX
]) {
3999 if (pmap_track_modified(pte_pv
->pv_pindex
)) {
4000 if ((pbits
& pmap
->pmap_bits
[PG_DEVICE_IDX
]) == 0) {
4002 m
= PHYS_TO_VM_PAGE(pbits
&
4007 cbits
&= ~pmap
->pmap_bits
[PG_M_IDX
];
4010 } else if (sharept
) {
4012 * Unmanaged page table, pt_pv is actually the pd_pv
4013 * for our pmap (not the object's shared pmap).
4015 * When asked to protect something in a shared page table
4016 * page we just unmap the page table page. We have to
4017 * invalidate the tlb in this situation.
4019 * XXX Warning, shared page tables will not be used for
4020 * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4021 * so PHYS_TO_VM_PAGE() should be safe here.
4023 pte
= pmap_inval_smp(pmap
, (vm_offset_t
)-1, 1, ptep
, 0);
4024 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte
& PG_FRAME
)))
4025 panic("pmap_protect: pgtable1 pg bad wirecount");
4026 if (vm_page_unwire_quick(pt_pv
->pv_m
))
4027 panic("pmap_protect: pgtable2 pg bad wirecount");
4030 /* else unmanaged page, adjust bits, no wire changes */
4033 cbits
&= ~pmap
->pmap_bits
[PG_RW_IDX
];
4035 if (pmap_enter_debug
> 0) {
4037 kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4038 "pt_pv=%p cbits=%08lx\n",
4044 if (pbits
!= cbits
) {
4045 if (!pmap_inval_smp_cmpset(pmap
, (vm_offset_t
)-1,
4046 ptep
, pbits
, cbits
)) {
4056 * Insert the vm_page (m) at the virtual address (va), replacing any prior
4057 * mapping at that address. Set protection and wiring as requested.
4059 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4060 * possible. If it is we enter the page into the appropriate shared pmap
4061 * hanging off the related VM object instead of the passed pmap, then we
4062 * share the page table page from the VM object's pmap into the current pmap.
4064 * NOTE: This routine MUST insert the page into the pmap now, it cannot
4068 pmap_enter(pmap_t pmap
, vm_offset_t va
, vm_page_t m
, vm_prot_t prot
,
4069 boolean_t wired
, vm_map_entry_t entry
)
4071 pv_entry_t pt_pv
; /* page table */
4072 pv_entry_t pte_pv
; /* page table entry */
4075 pt_entry_t origpte
, newpte
;
4080 va
= trunc_page(va
);
4081 #ifdef PMAP_DIAGNOSTIC
4083 panic("pmap_enter: toobig");
4084 if ((va
>= UPT_MIN_ADDRESS
) && (va
< UPT_MAX_ADDRESS
))
4085 panic("pmap_enter: invalid to pmap_enter page table "
4086 "pages (va: 0x%lx)", va
);
4088 if (va
< UPT_MAX_ADDRESS
&& pmap
== &kernel_pmap
) {
4089 kprintf("Warning: pmap_enter called on UVA with "
4092 db_print_backtrace();
4095 if (va
>= UPT_MAX_ADDRESS
&& pmap
!= &kernel_pmap
) {
4096 kprintf("Warning: pmap_enter called on KVA without"
4099 db_print_backtrace();
4104 * Get locked PV entries for our new page table entry (pte_pv)
4105 * and for its parent page table (pt_pv). We need the parent
4106 * so we can resolve the location of the ptep.
4108 * Only hardware MMU actions can modify the ptep out from
4111 * if (m) is fictitious or unmanaged we do not create a managing
4112 * pte_pv for it. Any pre-existing page's management state must
4113 * match (avoiding code complexity).
4115 * If the pmap is still being initialized we assume existing
4118 * Kernel mapppings do not track page table pages (i.e. pt_pv).
4120 if (pmap_initialized
== FALSE
) {
4125 } else if (m
->flags
& (/*PG_FICTITIOUS |*/ PG_UNMANAGED
)) { /* XXX */
4127 if (va
>= VM_MAX_USER_ADDRESS
) {
4131 pt_pv
= pmap_allocpte_seg(pmap
, pmap_pt_pindex(va
),
4133 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
4137 KASSERT(origpte
== 0 ||
4138 (origpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0,
4139 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte
, va
));
4141 if (va
>= VM_MAX_USER_ADDRESS
) {
4143 * Kernel map, pv_entry-tracked.
4146 pte_pv
= pmap_allocpte(pmap
, pmap_pte_pindex(va
), NULL
);
4152 pte_pv
= pmap_allocpte_seg(pmap
, pmap_pte_pindex(va
),
4154 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
4158 KASSERT(origpte
== 0 ||
4159 (origpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]),
4160 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte
, va
));
4163 pa
= VM_PAGE_TO_PHYS(m
);
4164 opa
= origpte
& PG_FRAME
;
4166 newpte
= (pt_entry_t
)(pa
| pte_prot(pmap
, prot
) |
4167 pmap
->pmap_bits
[PG_V_IDX
] | pmap
->pmap_bits
[PG_A_IDX
]);
4169 newpte
|= pmap
->pmap_bits
[PG_W_IDX
];
4170 if (va
< VM_MAX_USER_ADDRESS
)
4171 newpte
|= pmap
->pmap_bits
[PG_U_IDX
];
4173 newpte
|= pmap
->pmap_bits
[PG_MANAGED_IDX
];
4174 // if (pmap == &kernel_pmap)
4175 // newpte |= pgeflag;
4176 newpte
|= pmap
->pmap_cache_bits
[m
->pat_mode
];
4177 if (m
->flags
& PG_FICTITIOUS
)
4178 newpte
|= pmap
->pmap_bits
[PG_DEVICE_IDX
];
4181 * It is possible for multiple faults to occur in threaded
4182 * environments, the existing pte might be correct.
4184 if (((origpte
^ newpte
) & ~(pt_entry_t
)(pmap
->pmap_bits
[PG_M_IDX
] |
4185 pmap
->pmap_bits
[PG_A_IDX
])) == 0)
4189 * Ok, either the address changed or the protection or wiring
4192 * Clear the current entry, interlocking the removal. For managed
4193 * pte's this will also flush the modified state to the vm_page.
4194 * Atomic ops are mandatory in order to ensure that PG_M events are
4195 * not lost during any transition.
4197 * WARNING: The caller has busied the new page but not the original
4198 * vm_page which we are trying to replace. Because we hold
4199 * the pte_pv lock, but have not busied the page, PG bits
4200 * can be cleared out from under us.
4205 * pmap_remove_pv_pte() unwires pt_pv and assumes
4206 * we will free pte_pv, but since we are reusing
4207 * pte_pv we want to retain the wire count.
4209 * pt_pv won't exist for a kernel page (managed or
4213 vm_page_wire_quick(pt_pv
->pv_m
);
4214 if (prot
& VM_PROT_NOSYNC
) {
4215 pmap_remove_pv_pte(pte_pv
, pt_pv
, NULL
);
4217 pmap_inval_bulk_t bulk
;
4219 pmap_inval_bulk_init(&bulk
, pmap
);
4220 pmap_remove_pv_pte(pte_pv
, pt_pv
, &bulk
);
4221 pmap_inval_bulk_flush(&bulk
);
4224 pmap_remove_pv_page(pte_pv
);
4225 } else if (prot
& VM_PROT_NOSYNC
) {
4227 * Unmanaged page, NOSYNC (no mmu sync) requested.
4229 * Leave wire count on PT page intact.
4231 (void)pte_load_clear(ptep
);
4232 cpu_invlpg((void *)va
);
4233 atomic_add_long(&pmap
->pm_stats
.resident_count
, -1);
4236 * Unmanaged page, normal enter.
4238 * Leave wire count on PT page intact.
4240 pmap_inval_smp(pmap
, va
, 1, ptep
, 0);
4241 atomic_add_long(&pmap
->pm_stats
.resident_count
, -1);
4243 KKASSERT(*ptep
== 0);
4247 if (pmap_enter_debug
> 0) {
4249 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4250 " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4252 origpte
, newpte
, ptep
,
4253 pte_pv
, pt_pv
, opa
, prot
);
4259 * Enter on the PV list if part of our managed memory.
4260 * Wiring of the PT page is already handled.
4262 KKASSERT(pte_pv
->pv_m
== NULL
);
4263 vm_page_spin_lock(m
);
4265 pmap_page_stats_adding(m
);
4266 TAILQ_INSERT_TAIL(&m
->md
.pv_list
, pte_pv
, pv_list
);
4267 vm_page_flag_set(m
, PG_MAPPED
);
4268 vm_page_spin_unlock(m
);
4269 } else if (pt_pv
&& opa
== 0) {
4271 * We have to adjust the wire count on the PT page ourselves
4272 * for unmanaged entries. If opa was non-zero we retained
4273 * the existing wire count from the removal.
4275 vm_page_wire_quick(pt_pv
->pv_m
);
4279 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4281 * User VMAs do not because those will be zero->non-zero, so no
4282 * stale entries to worry about at this point.
4284 * For KVM there appear to still be issues. Theoretically we
4285 * should be able to scrap the interlocks entirely but we
4288 if ((prot
& VM_PROT_NOSYNC
) == 0 && pt_pv
== NULL
) {
4289 pmap_inval_smp(pmap
, va
, 1, ptep
, newpte
);
4291 *(volatile pt_entry_t
*)ptep
= newpte
;
4293 cpu_invlpg((void *)va
);
4298 atomic_add_long(&pte_pv
->pv_pmap
->pm_stats
.wired_count
,
4301 atomic_add_long(&pmap
->pm_stats
.wired_count
, 1);
4304 if (newpte
& pmap
->pmap_bits
[PG_RW_IDX
])
4305 vm_page_flag_set(m
, PG_WRITEABLE
);
4308 * Unmanaged pages need manual resident_count tracking.
4310 if (pte_pv
== NULL
&& pt_pv
)
4311 atomic_add_long(&pt_pv
->pv_pmap
->pm_stats
.resident_count
, 1);
4317 KKASSERT((newpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0 ||
4318 (m
->flags
& PG_MAPPED
));
4321 * Cleanup the pv entry, allowing other accessors.
4330 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
4331 * This code also assumes that the pmap has no pre-existing entry for this
4334 * This code currently may only be used on user pmaps, not kernel_pmap.
4337 pmap_enter_quick(pmap_t pmap
, vm_offset_t va
, vm_page_t m
)
4339 pmap_enter(pmap
, va
, m
, VM_PROT_READ
, FALSE
, NULL
);
4343 * Make a temporary mapping for a physical address. This is only intended
4344 * to be used for panic dumps.
4346 * The caller is responsible for calling smp_invltlb().
4349 pmap_kenter_temporary(vm_paddr_t pa
, long i
)
4351 pmap_kenter_quick((vm_offset_t
)crashdumpmap
+ (i
* PAGE_SIZE
), pa
);
4352 return ((void *)crashdumpmap
);
4355 #define MAX_INIT_PT (96)
4358 * This routine preloads the ptes for a given object into the specified pmap.
4359 * This eliminates the blast of soft faults on process startup and
4360 * immediately after an mmap.
4362 static int pmap_object_init_pt_callback(vm_page_t p
, void *data
);
4365 pmap_object_init_pt(pmap_t pmap
, vm_offset_t addr
, vm_prot_t prot
,
4366 vm_object_t object
, vm_pindex_t pindex
,
4367 vm_size_t size
, int limit
)
4369 struct rb_vm_page_scan_info info
;
4374 * We can't preinit if read access isn't set or there is no pmap
4377 if ((prot
& VM_PROT_READ
) == 0 || pmap
== NULL
|| object
== NULL
)
4381 * We can't preinit if the pmap is not the current pmap
4383 lp
= curthread
->td_lwp
;
4384 if (lp
== NULL
|| pmap
!= vmspace_pmap(lp
->lwp_vmspace
))
4388 * Misc additional checks
4390 psize
= x86_64_btop(size
);
4392 if ((object
->type
!= OBJT_VNODE
) ||
4393 ((limit
& MAP_PREFAULT_PARTIAL
) && (psize
> MAX_INIT_PT
) &&
4394 (object
->resident_page_count
> MAX_INIT_PT
))) {
4398 if (pindex
+ psize
> object
->size
) {
4399 if (object
->size
< pindex
)
4401 psize
= object
->size
- pindex
;
4408 * If everything is segment-aligned do not pre-init here. Instead
4409 * allow the normal vm_fault path to pass a segment hint to
4410 * pmap_enter() which will then use an object-referenced shared
4413 if ((addr
& SEG_MASK
) == 0 &&
4414 (ctob(psize
) & SEG_MASK
) == 0 &&
4415 (ctob(pindex
) & SEG_MASK
) == 0) {
4420 * Use a red-black scan to traverse the requested range and load
4421 * any valid pages found into the pmap.
4423 * We cannot safely scan the object's memq without holding the
4426 info
.start_pindex
= pindex
;
4427 info
.end_pindex
= pindex
+ psize
- 1;
4433 vm_object_hold_shared(object
);
4434 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, rb_vm_page_scancmp
,
4435 pmap_object_init_pt_callback
, &info
);
4436 vm_object_drop(object
);
4441 pmap_object_init_pt_callback(vm_page_t p
, void *data
)
4443 struct rb_vm_page_scan_info
*info
= data
;
4444 vm_pindex_t rel_index
;
4447 * don't allow an madvise to blow away our really
4448 * free pages allocating pv entries.
4450 if ((info
->limit
& MAP_PREFAULT_MADVISE
) &&
4451 vmstats
.v_free_count
< vmstats
.v_free_reserved
) {
4456 * Ignore list markers and ignore pages we cannot instantly
4457 * busy (while holding the object token).
4459 if (p
->flags
& PG_MARKER
)
4461 if (vm_page_busy_try(p
, TRUE
))
4463 if (((p
->valid
& VM_PAGE_BITS_ALL
) == VM_PAGE_BITS_ALL
) &&
4464 (p
->flags
& PG_FICTITIOUS
) == 0) {
4465 if ((p
->queue
- p
->pc
) == PQ_CACHE
)
4466 vm_page_deactivate(p
);
4467 rel_index
= p
->pindex
- info
->start_pindex
;
4468 pmap_enter_quick(info
->pmap
,
4469 info
->addr
+ x86_64_ptob(rel_index
), p
);
4477 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
4480 * Returns FALSE if it would be non-trivial or if a pte is already loaded
4483 * XXX This is safe only because page table pages are not freed.
4486 pmap_prefault_ok(pmap_t pmap
, vm_offset_t addr
)
4490 /*spin_lock(&pmap->pm_spin);*/
4491 if ((pte
= pmap_pte(pmap
, addr
)) != NULL
) {
4492 if (*pte
& pmap
->pmap_bits
[PG_V_IDX
]) {
4493 /*spin_unlock(&pmap->pm_spin);*/
4497 /*spin_unlock(&pmap->pm_spin);*/
4502 * Change the wiring attribute for a pmap/va pair. The mapping must already
4503 * exist in the pmap. The mapping may or may not be managed.
4506 pmap_change_wiring(pmap_t pmap
, vm_offset_t va
, boolean_t wired
,
4507 vm_map_entry_t entry
)
4514 lwkt_gettoken(&pmap
->pm_token
);
4515 pv
= pmap_allocpte_seg(pmap
, pmap_pt_pindex(va
), NULL
, entry
, va
);
4516 ptep
= pv_pte_lookup(pv
, pmap_pte_index(va
));
4518 if (wired
&& !pmap_pte_w(pmap
, ptep
))
4519 atomic_add_long(&pv
->pv_pmap
->pm_stats
.wired_count
, 1);
4520 else if (!wired
&& pmap_pte_w(pmap
, ptep
))
4521 atomic_add_long(&pv
->pv_pmap
->pm_stats
.wired_count
, -1);
4524 * Wiring is not a hardware characteristic so there is no need to
4525 * invalidate TLB. However, in an SMP environment we must use
4526 * a locked bus cycle to update the pte (if we are not using
4527 * the pmap_inval_*() API that is)... it's ok to do this for simple
4531 atomic_set_long(ptep
, pmap
->pmap_bits
[PG_W_IDX
]);
4533 atomic_clear_long(ptep
, pmap
->pmap_bits
[PG_W_IDX
]);
4535 lwkt_reltoken(&pmap
->pm_token
);
4541 * Copy the range specified by src_addr/len from the source map to
4542 * the range dst_addr/len in the destination map.
4544 * This routine is only advisory and need not do anything.
4547 pmap_copy(pmap_t dst_pmap
, pmap_t src_pmap
, vm_offset_t dst_addr
,
4548 vm_size_t len
, vm_offset_t src_addr
)
4555 * Zero the specified physical page.
4557 * This function may be called from an interrupt and no locking is
4561 pmap_zero_page(vm_paddr_t phys
)
4563 vm_offset_t va
= PHYS_TO_DMAP(phys
);
4565 pagezero((void *)va
);
4571 * Zero part of a physical page by mapping it into memory and clearing
4572 * its contents with bzero.
4574 * off and size may not cover an area beyond a single hardware page.
4577 pmap_zero_page_area(vm_paddr_t phys
, int off
, int size
)
4579 vm_offset_t virt
= PHYS_TO_DMAP(phys
);
4581 bzero((char *)virt
+ off
, size
);
4587 * Copy the physical page from the source PA to the target PA.
4588 * This function may be called from an interrupt. No locking
4592 pmap_copy_page(vm_paddr_t src
, vm_paddr_t dst
)
4594 vm_offset_t src_virt
, dst_virt
;
4596 src_virt
= PHYS_TO_DMAP(src
);
4597 dst_virt
= PHYS_TO_DMAP(dst
);
4598 bcopy((void *)src_virt
, (void *)dst_virt
, PAGE_SIZE
);
4602 * pmap_copy_page_frag:
4604 * Copy the physical page from the source PA to the target PA.
4605 * This function may be called from an interrupt. No locking
4609 pmap_copy_page_frag(vm_paddr_t src
, vm_paddr_t dst
, size_t bytes
)
4611 vm_offset_t src_virt
, dst_virt
;
4613 src_virt
= PHYS_TO_DMAP(src
);
4614 dst_virt
= PHYS_TO_DMAP(dst
);
4616 bcopy((char *)src_virt
+ (src
& PAGE_MASK
),
4617 (char *)dst_virt
+ (dst
& PAGE_MASK
),
4622 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
4623 * this page. This count may be changed upwards or downwards in the future;
4624 * it is only necessary that true be returned for a small subset of pmaps
4625 * for proper page aging.
4628 pmap_page_exists_quick(pmap_t pmap
, vm_page_t m
)
4633 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
4636 vm_page_spin_lock(m
);
4637 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
4638 if (pv
->pv_pmap
== pmap
) {
4639 vm_page_spin_unlock(m
);
4646 vm_page_spin_unlock(m
);
4651 * Remove all pages from specified address space this aids process exit
4652 * speeds. Also, this code may be special cased for the current process
4656 pmap_remove_pages(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
)
4658 pmap_remove_noinval(pmap
, sva
, eva
);
4663 * pmap_testbit tests bits in pte's note that the testbit/clearbit
4664 * routines are inline, and a lot of things compile-time evaluate.
4668 pmap_testbit(vm_page_t m
, int bit
)
4674 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
4677 if (TAILQ_FIRST(&m
->md
.pv_list
) == NULL
)
4679 vm_page_spin_lock(m
);
4680 if (TAILQ_FIRST(&m
->md
.pv_list
) == NULL
) {
4681 vm_page_spin_unlock(m
);
4685 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
4687 #if defined(PMAP_DIAGNOSTIC)
4688 if (pv
->pv_pmap
== NULL
) {
4689 kprintf("Null pmap (tb) at pindex: %"PRIu64
"\n",
4697 * If the bit being tested is the modified bit, then
4698 * mark clean_map and ptes as never
4701 * WARNING! Because we do not lock the pv, *pte can be in a
4702 * state of flux. Despite this the value of *pte
4703 * will still be related to the vm_page in some way
4704 * because the pv cannot be destroyed as long as we
4705 * hold the vm_page spin lock.
4707 if (bit
== PG_A_IDX
|| bit
== PG_M_IDX
) {
4708 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
4709 if (!pmap_track_modified(pv
->pv_pindex
))
4713 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_pindex
<< PAGE_SHIFT
);
4714 if (*pte
& pmap
->pmap_bits
[bit
]) {
4715 vm_page_spin_unlock(m
);
4719 vm_page_spin_unlock(m
);
4724 * This routine is used to modify bits in ptes. Only one bit should be
4725 * specified. PG_RW requires special handling.
4727 * Caller must NOT hold any spin locks
4731 pmap_clearbit(vm_page_t m
, int bit_index
)
4738 if (bit_index
== PG_RW_IDX
)
4739 vm_page_flag_clear(m
, PG_WRITEABLE
);
4740 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
)) {
4747 * Loop over all current mappings setting/clearing as appropos If
4748 * setting RO do we need to clear the VAC?
4750 * NOTE: When clearing PG_M we could also (not implemented) drop
4751 * through to the PG_RW code and clear PG_RW too, forcing
4752 * a fault on write to redetect PG_M for virtual kernels, but
4753 * it isn't necessary since virtual kernels invalidate the
4754 * pte when they clear the VPTE_M bit in their virtual page
4757 * NOTE: Does not re-dirty the page when clearing only PG_M.
4759 * NOTE: Because we do not lock the pv, *pte can be in a state of
4760 * flux. Despite this the value of *pte is still somewhat
4761 * related while we hold the vm_page spin lock.
4763 * *pte can be zero due to this race. Since we are clearing
4764 * bits we basically do no harm when this race ccurs.
4766 if (bit_index
!= PG_RW_IDX
) {
4767 vm_page_spin_lock(m
);
4768 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
4769 #if defined(PMAP_DIAGNOSTIC)
4770 if (pv
->pv_pmap
== NULL
) {
4771 kprintf("Null pmap (cb) at pindex: %"PRIu64
"\n",
4777 pte
= pmap_pte_quick(pv
->pv_pmap
,
4778 pv
->pv_pindex
<< PAGE_SHIFT
);
4780 if (pbits
& pmap
->pmap_bits
[bit_index
])
4781 atomic_clear_long(pte
, pmap
->pmap_bits
[bit_index
]);
4783 vm_page_spin_unlock(m
);
4788 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
4792 vm_page_spin_lock(m
);
4793 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
4795 * don't write protect pager mappings
4797 if (!pmap_track_modified(pv
->pv_pindex
))
4800 #if defined(PMAP_DIAGNOSTIC)
4801 if (pv
->pv_pmap
== NULL
) {
4802 kprintf("Null pmap (cb) at pindex: %"PRIu64
"\n",
4809 * Skip pages which do not have PG_RW set.
4811 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_pindex
<< PAGE_SHIFT
);
4812 if ((*pte
& pmap
->pmap_bits
[PG_RW_IDX
]) == 0)
4818 if (pv_hold_try(pv
)) {
4819 vm_page_spin_unlock(m
);
4821 vm_page_spin_unlock(m
);
4822 pv_lock(pv
); /* held, now do a blocking lock */
4824 if (pv
->pv_pmap
!= pmap
|| pv
->pv_m
!= m
) {
4825 pv_put(pv
); /* and release */
4826 goto restart
; /* anything could have happened */
4828 KKASSERT(pv
->pv_pmap
== pmap
);
4834 nbits
= pbits
& ~(pmap
->pmap_bits
[PG_RW_IDX
] |
4835 pmap
->pmap_bits
[PG_M_IDX
]);
4836 if (pmap_inval_smp_cmpset(pmap
,
4837 ((vm_offset_t
)pv
->pv_pindex
<< PAGE_SHIFT
),
4838 pte
, pbits
, nbits
)) {
4843 vm_page_spin_lock(m
);
4846 * If PG_M was found to be set while we were clearing PG_RW
4847 * we also clear PG_M (done above) and mark the page dirty.
4848 * Callers expect this behavior.
4850 if (pbits
& pmap
->pmap_bits
[PG_M_IDX
])
4854 vm_page_spin_unlock(m
);
4858 * Lower the permission for all mappings to a given page.
4860 * Page must be busied by caller. Because page is busied by caller this
4861 * should not be able to race a pmap_enter().
4864 pmap_page_protect(vm_page_t m
, vm_prot_t prot
)
4866 /* JG NX support? */
4867 if ((prot
& VM_PROT_WRITE
) == 0) {
4868 if (prot
& (VM_PROT_READ
| VM_PROT_EXECUTE
)) {
4870 * NOTE: pmap_clearbit(.. PG_RW) also clears
4871 * the PG_WRITEABLE flag in (m).
4873 pmap_clearbit(m
, PG_RW_IDX
);
4881 pmap_phys_address(vm_pindex_t ppn
)
4883 return (x86_64_ptob(ppn
));
4887 * Return a count of reference bits for a page, clearing those bits.
4888 * It is not necessary for every reference bit to be cleared, but it
4889 * is necessary that 0 only be returned when there are truly no
4890 * reference bits set.
4892 * XXX: The exact number of bits to check and clear is a matter that
4893 * should be tested and standardized at some point in the future for
4894 * optimal aging of shared pages.
4896 * This routine may not block.
4899 pmap_ts_referenced(vm_page_t m
)
4906 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
4909 vm_page_spin_lock(m
);
4910 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
4911 if (!pmap_track_modified(pv
->pv_pindex
))
4914 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_pindex
<< PAGE_SHIFT
);
4915 if (pte
&& (*pte
& pmap
->pmap_bits
[PG_A_IDX
])) {
4916 atomic_clear_long(pte
, pmap
->pmap_bits
[PG_A_IDX
]);
4922 vm_page_spin_unlock(m
);
4929 * Return whether or not the specified physical page was modified
4930 * in any physical maps.
4933 pmap_is_modified(vm_page_t m
)
4937 res
= pmap_testbit(m
, PG_M_IDX
);
4942 * Clear the modify bits on the specified physical page.
4945 pmap_clear_modify(vm_page_t m
)
4947 pmap_clearbit(m
, PG_M_IDX
);
4951 * pmap_clear_reference:
4953 * Clear the reference bit on the specified physical page.
4956 pmap_clear_reference(vm_page_t m
)
4958 pmap_clearbit(m
, PG_A_IDX
);
4962 * Miscellaneous support routines follow
4967 i386_protection_init(void)
4971 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
4972 kp
= protection_codes
;
4973 for (prot
= 0; prot
< PROTECTION_CODES_SIZE
; prot
++) {
4975 case VM_PROT_NONE
| VM_PROT_NONE
| VM_PROT_NONE
:
4977 * Read access is also 0. There isn't any execute bit,
4978 * so just make it readable.
4980 case VM_PROT_READ
| VM_PROT_NONE
| VM_PROT_NONE
:
4981 case VM_PROT_READ
| VM_PROT_NONE
| VM_PROT_EXECUTE
:
4982 case VM_PROT_NONE
| VM_PROT_NONE
| VM_PROT_EXECUTE
:
4985 case VM_PROT_NONE
| VM_PROT_WRITE
| VM_PROT_NONE
:
4986 case VM_PROT_NONE
| VM_PROT_WRITE
| VM_PROT_EXECUTE
:
4987 case VM_PROT_READ
| VM_PROT_WRITE
| VM_PROT_NONE
:
4988 case VM_PROT_READ
| VM_PROT_WRITE
| VM_PROT_EXECUTE
:
4989 *kp
++ = pmap_bits_default
[PG_RW_IDX
];
4996 * Map a set of physical memory pages into the kernel virtual
4997 * address space. Return a pointer to where it is mapped. This
4998 * routine is intended to be used for mapping device memory,
5001 * NOTE: We can't use pgeflag unless we invalidate the pages one at
5004 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5005 * work whether the cpu supports PAT or not. The remaining PAT
5006 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5010 pmap_mapdev(vm_paddr_t pa
, vm_size_t size
)
5012 return(pmap_mapdev_attr(pa
, size
, PAT_WRITE_BACK
));
5016 pmap_mapdev_uncacheable(vm_paddr_t pa
, vm_size_t size
)
5018 return(pmap_mapdev_attr(pa
, size
, PAT_UNCACHEABLE
));
5022 pmap_mapbios(vm_paddr_t pa
, vm_size_t size
)
5024 return (pmap_mapdev_attr(pa
, size
, PAT_WRITE_BACK
));
5028 * Map a set of physical memory pages into the kernel virtual
5029 * address space. Return a pointer to where it is mapped. This
5030 * routine is intended to be used for mapping device memory,
5034 pmap_mapdev_attr(vm_paddr_t pa
, vm_size_t size
, int mode
)
5036 vm_offset_t va
, tmpva
, offset
;
5040 offset
= pa
& PAGE_MASK
;
5041 size
= roundup(offset
+ size
, PAGE_SIZE
);
5043 va
= kmem_alloc_nofault(&kernel_map
, size
, PAGE_SIZE
);
5045 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5047 pa
= pa
& ~PAGE_MASK
;
5048 for (tmpva
= va
, tmpsize
= size
; tmpsize
> 0;) {
5049 pte
= vtopte(tmpva
);
5051 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
5052 kernel_pmap
.pmap_bits
[PG_V_IDX
] | /* pgeflag | */
5053 kernel_pmap
.pmap_cache_bits
[mode
];
5054 tmpsize
-= PAGE_SIZE
;
5058 pmap_invalidate_range(&kernel_pmap
, va
, va
+ size
);
5059 pmap_invalidate_cache_range(va
, va
+ size
);
5061 return ((void *)(va
+ offset
));
5065 pmap_unmapdev(vm_offset_t va
, vm_size_t size
)
5067 vm_offset_t base
, offset
;
5069 base
= va
& ~PAGE_MASK
;
5070 offset
= va
& PAGE_MASK
;
5071 size
= roundup(offset
+ size
, PAGE_SIZE
);
5072 pmap_qremove(va
, size
>> PAGE_SHIFT
);
5073 kmem_free(&kernel_map
, base
, size
);
5077 * Sets the memory attribute for the specified page.
5080 pmap_page_set_memattr(vm_page_t m
, vm_memattr_t ma
)
5086 * If "m" is a normal page, update its direct mapping. This update
5087 * can be relied upon to perform any cache operations that are
5088 * required for data coherence.
5090 if ((m
->flags
& PG_FICTITIOUS
) == 0)
5091 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m
)), 1, m
->pat_mode
);
5095 * Change the PAT attribute on an existing kernel memory map. Caller
5096 * must ensure that the virtual memory in question is not accessed
5097 * during the adjustment.
5100 pmap_change_attr(vm_offset_t va
, vm_size_t count
, int mode
)
5107 panic("pmap_change_attr: va is NULL");
5108 base
= trunc_page(va
);
5112 *pte
= (*pte
& ~(pt_entry_t
)(kernel_pmap
.pmap_cache_mask
)) |
5113 kernel_pmap
.pmap_cache_bits
[mode
];
5118 changed
= 1; /* XXX: not optimal */
5121 * Flush CPU caches if required to make sure any data isn't cached that
5122 * shouldn't be, etc.
5125 pmap_invalidate_range(&kernel_pmap
, base
, va
);
5126 pmap_invalidate_cache_range(base
, va
);
5131 * perform the pmap work for mincore
5134 pmap_mincore(pmap_t pmap
, vm_offset_t addr
)
5136 pt_entry_t
*ptep
, pte
;
5140 lwkt_gettoken(&pmap
->pm_token
);
5141 ptep
= pmap_pte(pmap
, addr
);
5143 if (ptep
&& (pte
= *ptep
) != 0) {
5146 val
= MINCORE_INCORE
;
5147 if ((pte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0)
5150 pa
= pte
& PG_FRAME
;
5152 if (pte
& pmap
->pmap_bits
[PG_DEVICE_IDX
])
5155 m
= PHYS_TO_VM_PAGE(pa
);
5160 if (pte
& pmap
->pmap_bits
[PG_M_IDX
])
5161 val
|= MINCORE_MODIFIED
|MINCORE_MODIFIED_OTHER
;
5163 * Modified by someone
5165 else if (m
&& (m
->dirty
|| pmap_is_modified(m
)))
5166 val
|= MINCORE_MODIFIED_OTHER
;
5170 if (pte
& pmap
->pmap_bits
[PG_A_IDX
])
5171 val
|= MINCORE_REFERENCED
|MINCORE_REFERENCED_OTHER
;
5174 * Referenced by someone
5176 else if (m
&& ((m
->flags
& PG_REFERENCED
) ||
5177 pmap_ts_referenced(m
))) {
5178 val
|= MINCORE_REFERENCED_OTHER
;
5179 vm_page_flag_set(m
, PG_REFERENCED
);
5183 lwkt_reltoken(&pmap
->pm_token
);
5189 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
5190 * vmspace will be ref'd and the old one will be deref'd.
5192 * The vmspace for all lwps associated with the process will be adjusted
5193 * and cr3 will be reloaded if any lwp is the current lwp.
5195 * The process must hold the vmspace->vm_map.token for oldvm and newvm
5198 pmap_replacevm(struct proc
*p
, struct vmspace
*newvm
, int adjrefs
)
5200 struct vmspace
*oldvm
;
5203 oldvm
= p
->p_vmspace
;
5204 if (oldvm
!= newvm
) {
5207 p
->p_vmspace
= newvm
;
5208 KKASSERT(p
->p_nthreads
== 1);
5209 lp
= RB_ROOT(&p
->p_lwp_tree
);
5210 pmap_setlwpvm(lp
, newvm
);
5217 * Set the vmspace for a LWP. The vmspace is almost universally set the
5218 * same as the process vmspace, but virtual kernels need to swap out contexts
5219 * on a per-lwp basis.
5221 * Caller does not necessarily hold any vmspace tokens. Caller must control
5222 * the lwp (typically be in the context of the lwp). We use a critical
5223 * section to protect against statclock and hardclock (statistics collection).
5226 pmap_setlwpvm(struct lwp
*lp
, struct vmspace
*newvm
)
5228 struct vmspace
*oldvm
;
5231 oldvm
= lp
->lwp_vmspace
;
5233 if (oldvm
!= newvm
) {
5235 lp
->lwp_vmspace
= newvm
;
5236 if (curthread
->td_lwp
== lp
) {
5237 pmap
= vmspace_pmap(newvm
);
5238 ATOMIC_CPUMASK_ORBIT(pmap
->pm_active
, mycpu
->gd_cpuid
);
5239 if (pmap
->pm_active_lock
& CPULOCK_EXCL
)
5240 pmap_interlock_wait(newvm
);
5241 #if defined(SWTCH_OPTIM_STATS)
5244 if (pmap
->pmap_bits
[TYPE_IDX
] == REGULAR_PMAP
) {
5245 curthread
->td_pcb
->pcb_cr3
= vtophys(pmap
->pm_pml4
);
5246 } else if (pmap
->pmap_bits
[TYPE_IDX
] == EPT_PMAP
) {
5247 curthread
->td_pcb
->pcb_cr3
= KPML4phys
;
5249 panic("pmap_setlwpvm: unknown pmap type\n");
5251 load_cr3(curthread
->td_pcb
->pcb_cr3
);
5252 pmap
= vmspace_pmap(oldvm
);
5253 ATOMIC_CPUMASK_NANDBIT(pmap
->pm_active
,
5261 * Called when switching to a locked pmap, used to interlock against pmaps
5262 * undergoing modifications to prevent us from activating the MMU for the
5263 * target pmap until all such modifications have completed. We have to do
5264 * this because the thread making the modifications has already set up its
5265 * SMP synchronization mask.
5267 * This function cannot sleep!
5272 pmap_interlock_wait(struct vmspace
*vm
)
5274 struct pmap
*pmap
= &vm
->vm_pmap
;
5276 if (pmap
->pm_active_lock
& CPULOCK_EXCL
) {
5278 KKASSERT(curthread
->td_critcount
>= 2);
5279 DEBUG_PUSH_INFO("pmap_interlock_wait");
5280 while (pmap
->pm_active_lock
& CPULOCK_EXCL
) {
5282 lwkt_process_ipiq();
5290 pmap_addr_hint(vm_object_t obj
, vm_offset_t addr
, vm_size_t size
)
5293 if ((obj
== NULL
) || (size
< NBPDR
) ||
5294 ((obj
->type
!= OBJT_DEVICE
) && (obj
->type
!= OBJT_MGTDEVICE
))) {
5298 addr
= roundup2(addr
, NBPDR
);
5303 * Used by kmalloc/kfree, page already exists at va
5306 pmap_kvtom(vm_offset_t va
)
5308 pt_entry_t
*ptep
= vtopte(va
);
5310 KKASSERT((*ptep
& kernel_pmap
.pmap_bits
[PG_DEVICE_IDX
]) == 0);
5311 return(PHYS_TO_VM_PAGE(*ptep
& PG_FRAME
));
5315 * Initialize machine-specific shared page directory support. This
5316 * is executed when a VM object is created.
5319 pmap_object_init(vm_object_t object
)
5321 object
->md
.pmap_rw
= NULL
;
5322 object
->md
.pmap_ro
= NULL
;
5326 * Clean up machine-specific shared page directory support. This
5327 * is executed when a VM object is destroyed.
5330 pmap_object_free(vm_object_t object
)
5334 if ((pmap
= object
->md
.pmap_rw
) != NULL
) {
5335 object
->md
.pmap_rw
= NULL
;
5336 pmap_remove_noinval(pmap
,
5337 VM_MIN_USER_ADDRESS
, VM_MAX_USER_ADDRESS
);
5338 CPUMASK_ASSZERO(pmap
->pm_active
);
5341 kfree(pmap
, M_OBJPMAP
);
5343 if ((pmap
= object
->md
.pmap_ro
) != NULL
) {
5344 object
->md
.pmap_ro
= NULL
;
5345 pmap_remove_noinval(pmap
,
5346 VM_MIN_USER_ADDRESS
, VM_MAX_USER_ADDRESS
);
5347 CPUMASK_ASSZERO(pmap
->pm_active
);
5350 kfree(pmap
, M_OBJPMAP
);