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) 2011-2017 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, pmarkp) _pv_get(pmap, pindex, pmarkp \
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 \
122 #define pv_free(pv, pvp) _pv_free(pv, pvp PMAP_DEBUG_ARGS)
126 #define PMAP_DEBUG_DECL
127 #define PMAP_DEBUG_ARGS
128 #define PMAP_DEBUG_COPY
130 #define pv_get(pmap, pindex, pmarkp) _pv_get(pmap, pindex, pmarkp)
131 #define pv_lock(pv) _pv_lock(pv)
132 #define pv_hold_try(pv) _pv_hold_try(pv)
133 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp)
134 #define pv_free(pv, pvp) _pv_free(pv, pvp)
139 * Get PDEs and PTEs for user/kernel address space
141 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
143 #define pmap_pde_v(pmap, pte) ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
144 #define pmap_pte_w(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
145 #define pmap_pte_m(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
146 #define pmap_pte_u(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
147 #define pmap_pte_v(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
150 * Given a map and a machine independent protection code,
151 * convert to a vax protection code.
153 #define pte_prot(m, p) \
154 (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
155 static uint64_t protection_codes
[PROTECTION_CODES_SIZE
];
157 struct pmap kernel_pmap
;
159 MALLOC_DEFINE(M_OBJPMAP
, "objpmap", "pmaps associated with VM objects");
161 vm_paddr_t avail_start
; /* PA of first available physical page */
162 vm_paddr_t avail_end
; /* PA of last available physical page */
163 vm_offset_t virtual2_start
; /* cutout free area prior to kernel start */
164 vm_offset_t virtual2_end
;
165 vm_offset_t virtual_start
; /* VA of first avail page (after kernel bss) */
166 vm_offset_t virtual_end
; /* VA of last avail page (end of kernel AS) */
167 vm_offset_t KvaStart
; /* VA start of KVA space */
168 vm_offset_t KvaEnd
; /* VA end of KVA space (non-inclusive) */
169 vm_offset_t KvaSize
; /* max size of kernel virtual address space */
170 static boolean_t pmap_initialized
= FALSE
; /* Has pmap_init completed? */
171 //static int pgeflag; /* PG_G or-in */
172 //static int pseflag; /* PG_PS or-in */
176 static vm_paddr_t dmaplimit
;
178 vm_offset_t kernel_vm_end
= VM_MIN_KERNEL_ADDRESS
;
180 static pt_entry_t pat_pte_index
[PAT_INDEX_SIZE
]; /* PAT -> PG_ bits */
181 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/ /* PAT -> PG_ bits */
183 static uint64_t KPTbase
;
184 static uint64_t KPTphys
;
185 static uint64_t KPDphys
; /* phys addr of kernel level 2 */
186 static uint64_t KPDbase
; /* phys addr of kernel level 2 @ KERNBASE */
187 uint64_t KPDPphys
; /* phys addr of kernel level 3 */
188 uint64_t KPML4phys
; /* phys addr of kernel level 4 */
190 static uint64_t DMPDphys
; /* phys addr of direct mapped level 2 */
191 static uint64_t DMPDPphys
; /* phys addr of direct mapped level 3 */
194 * Data for the pv entry allocation mechanism
196 static vm_zone_t pvzone
;
197 static struct vm_zone pvzone_store
;
198 static int pv_entry_max
=0, pv_entry_high_water
=0;
199 static int pmap_pagedaemon_waken
= 0;
200 static struct pv_entry
*pvinit
;
203 * All those kernel PT submaps that BSD is so fond of
205 pt_entry_t
*CMAP1
= NULL
, *ptmmap
;
206 caddr_t CADDR1
= NULL
, ptvmmap
= NULL
;
207 static pt_entry_t
*msgbufmap
;
208 struct msgbuf
*msgbufp
=NULL
;
211 * PMAP default PG_* bits. Needed to be able to add
212 * EPT/NPT pagetable pmap_bits for the VMM module
214 uint64_t pmap_bits_default
[] = {
215 REGULAR_PMAP
, /* TYPE_IDX 0 */
216 X86_PG_V
, /* PG_V_IDX 1 */
217 X86_PG_RW
, /* PG_RW_IDX 2 */
218 X86_PG_U
, /* PG_U_IDX 3 */
219 X86_PG_A
, /* PG_A_IDX 4 */
220 X86_PG_M
, /* PG_M_IDX 5 */
221 X86_PG_PS
, /* PG_PS_IDX3 6 */
222 X86_PG_G
, /* PG_G_IDX 7 */
223 X86_PG_AVAIL1
, /* PG_AVAIL1_IDX 8 */
224 X86_PG_AVAIL2
, /* PG_AVAIL2_IDX 9 */
225 X86_PG_AVAIL3
, /* PG_AVAIL3_IDX 10 */
226 X86_PG_NC_PWT
| X86_PG_NC_PCD
, /* PG_N_IDX 11 */
227 X86_PG_NX
, /* PG_NX_IDX 12 */
232 static pt_entry_t
*pt_crashdumpmap
;
233 static caddr_t crashdumpmap
;
235 static int pmap_debug
= 0;
236 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_debug
, CTLFLAG_RW
,
237 &pmap_debug
, 0, "Debug pmap's");
239 static int pmap_enter_debug
= 0;
240 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_enter_debug
, CTLFLAG_RW
,
241 &pmap_enter_debug
, 0, "Debug pmap_enter's");
243 static int pmap_yield_count
= 64;
244 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_yield_count
, CTLFLAG_RW
,
245 &pmap_yield_count
, 0, "Yield during init_pt/release");
246 static int pmap_mmu_optimize
= 0;
247 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_mmu_optimize
, CTLFLAG_RW
,
248 &pmap_mmu_optimize
, 0, "Share page table pages when possible");
249 int pmap_fast_kernel_cpusync
= 0;
250 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_fast_kernel_cpusync
, CTLFLAG_RW
,
251 &pmap_fast_kernel_cpusync
, 0, "Share page table pages when possible");
252 int pmap_dynamic_delete
= 0;
253 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_dynamic_delete
, CTLFLAG_RW
,
254 &pmap_dynamic_delete
, 0, "Dynamically delete PT/PD/PDPs");
256 static int pmap_nx_enable
= 0;
257 /* needs manual TUNABLE in early probe, see below */
261 /* Standard user access funtions */
262 extern int std_copyinstr (const void *udaddr
, void *kaddr
, size_t len
,
264 extern int std_copyin (const void *udaddr
, void *kaddr
, size_t len
);
265 extern int std_copyout (const void *kaddr
, void *udaddr
, size_t len
);
266 extern int std_fubyte (const uint8_t *base
);
267 extern int std_subyte (uint8_t *base
, uint8_t byte
);
268 extern int32_t std_fuword32 (const uint32_t *base
);
269 extern int64_t std_fuword64 (const uint64_t *base
);
270 extern int std_suword64 (uint64_t *base
, uint64_t word
);
271 extern int std_suword32 (uint32_t *base
, int word
);
272 extern uint32_t std_swapu32 (volatile uint32_t *base
, uint32_t v
);
273 extern uint64_t std_swapu64 (volatile uint64_t *base
, uint64_t v
);
275 static void pv_hold(pv_entry_t pv
);
276 static int _pv_hold_try(pv_entry_t pv
278 static void pv_drop(pv_entry_t pv
);
279 static void _pv_lock(pv_entry_t pv
281 static void pv_unlock(pv_entry_t pv
);
282 static pv_entry_t
_pv_alloc(pmap_t pmap
, vm_pindex_t pindex
, int *isnew
284 static pv_entry_t
_pv_get(pmap_t pmap
, vm_pindex_t pindex
, vm_pindex_t
**pmarkp
286 static void _pv_free(pv_entry_t pv
, pv_entry_t pvp PMAP_DEBUG_DECL
);
287 static pv_entry_t
pv_get_try(pmap_t pmap
, vm_pindex_t pindex
,
288 vm_pindex_t
**pmarkp
, int *errorp
);
289 static void pv_put(pv_entry_t pv
);
290 static void *pv_pte_lookup(pv_entry_t pv
, vm_pindex_t pindex
);
291 static pv_entry_t
pmap_allocpte(pmap_t pmap
, vm_pindex_t ptepindex
,
293 static pv_entry_t
pmap_allocpte_seg(pmap_t pmap
, vm_pindex_t ptepindex
,
294 pv_entry_t
*pvpp
, vm_map_entry_t entry
, vm_offset_t va
);
295 static void pmap_remove_pv_pte(pv_entry_t pv
, pv_entry_t pvp
,
296 pmap_inval_bulk_t
*bulk
, int destroy
);
297 static vm_page_t
pmap_remove_pv_page(pv_entry_t pv
);
298 static int pmap_release_pv(pv_entry_t pv
, pv_entry_t pvp
,
299 pmap_inval_bulk_t
*bulk
);
301 struct pmap_scan_info
;
302 static void pmap_remove_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
303 pv_entry_t pte_pv
, vm_pindex_t
*pte_placemark
,
304 pv_entry_t pt_pv
, int sharept
,
305 vm_offset_t va
, pt_entry_t
*ptep
, void *arg __unused
);
306 static void pmap_protect_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
307 pv_entry_t pte_pv
, vm_pindex_t
*pte_placemark
,
308 pv_entry_t pt_pv
, int sharept
,
309 vm_offset_t va
, pt_entry_t
*ptep
, void *arg __unused
);
311 static void i386_protection_init (void);
312 static void create_pagetables(vm_paddr_t
*firstaddr
);
313 static void pmap_remove_all (vm_page_t m
);
314 static boolean_t
pmap_testbit (vm_page_t m
, int bit
);
316 static pt_entry_t
* pmap_pte_quick (pmap_t pmap
, vm_offset_t va
);
317 static vm_offset_t
pmap_kmem_choose(vm_offset_t addr
);
319 static void pmap_pinit_defaults(struct pmap
*pmap
);
320 static void pv_placemarker_wait(pmap_t pmap
, vm_pindex_t
*pmark
);
321 static void pv_placemarker_wakeup(pmap_t pmap
, vm_pindex_t
*pmark
);
323 static unsigned pdir4mb
;
326 pv_entry_compare(pv_entry_t pv1
, pv_entry_t pv2
)
328 if (pv1
->pv_pindex
< pv2
->pv_pindex
)
330 if (pv1
->pv_pindex
> pv2
->pv_pindex
)
335 RB_GENERATE2(pv_entry_rb_tree
, pv_entry
, pv_entry
,
336 pv_entry_compare
, vm_pindex_t
, pv_pindex
);
340 pmap_page_stats_adding(vm_page_t m
)
342 globaldata_t gd
= mycpu
;
344 if (TAILQ_EMPTY(&m
->md
.pv_list
)) {
345 ++gd
->gd_vmtotal
.t_arm
;
346 } else if (TAILQ_FIRST(&m
->md
.pv_list
) ==
347 TAILQ_LAST(&m
->md
.pv_list
, md_page_pv_list
)) {
348 ++gd
->gd_vmtotal
.t_armshr
;
349 ++gd
->gd_vmtotal
.t_avmshr
;
351 ++gd
->gd_vmtotal
.t_avmshr
;
357 pmap_page_stats_deleting(vm_page_t m
)
359 globaldata_t gd
= mycpu
;
361 if (TAILQ_EMPTY(&m
->md
.pv_list
)) {
362 --gd
->gd_vmtotal
.t_arm
;
363 } else if (TAILQ_FIRST(&m
->md
.pv_list
) ==
364 TAILQ_LAST(&m
->md
.pv_list
, md_page_pv_list
)) {
365 --gd
->gd_vmtotal
.t_armshr
;
366 --gd
->gd_vmtotal
.t_avmshr
;
368 --gd
->gd_vmtotal
.t_avmshr
;
373 * Move the kernel virtual free pointer to the next
374 * 2MB. This is used to help improve performance
375 * by using a large (2MB) page for much of the kernel
376 * (.text, .data, .bss)
380 pmap_kmem_choose(vm_offset_t addr
)
382 vm_offset_t newaddr
= addr
;
384 newaddr
= roundup2(addr
, NBPDR
);
391 * Super fast pmap_pte routine best used when scanning the pv lists.
392 * This eliminates many course-grained invltlb calls. Note that many of
393 * the pv list scans are across different pmaps and it is very wasteful
394 * to do an entire invltlb when checking a single mapping.
396 static __inline pt_entry_t
*pmap_pte(pmap_t pmap
, vm_offset_t va
);
400 pmap_pte_quick(pmap_t pmap
, vm_offset_t va
)
402 return pmap_pte(pmap
, va
);
406 * Returns the pindex of a page table entry (representing a terminal page).
407 * There are NUPTE_TOTAL page table entries possible (a huge number)
409 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
410 * We want to properly translate negative KVAs.
414 pmap_pte_pindex(vm_offset_t va
)
416 return ((va
>> PAGE_SHIFT
) & (NUPTE_TOTAL
- 1));
420 * Returns the pindex of a page table.
424 pmap_pt_pindex(vm_offset_t va
)
426 return (NUPTE_TOTAL
+ ((va
>> PDRSHIFT
) & (NUPT_TOTAL
- 1)));
430 * Returns the pindex of a page directory.
434 pmap_pd_pindex(vm_offset_t va
)
436 return (NUPTE_TOTAL
+ NUPT_TOTAL
+
437 ((va
>> PDPSHIFT
) & (NUPD_TOTAL
- 1)));
442 pmap_pdp_pindex(vm_offset_t va
)
444 return (NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
+
445 ((va
>> PML4SHIFT
) & (NUPDP_TOTAL
- 1)));
450 pmap_pml4_pindex(void)
452 return (NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
+ NUPDP_TOTAL
);
456 * Return various clipped indexes for a given VA
458 * Returns the index of a pt in a page directory, representing a page
463 pmap_pt_index(vm_offset_t va
)
465 return ((va
>> PDRSHIFT
) & ((1ul << NPDEPGSHIFT
) - 1));
469 * Returns the index of a pd in a page directory page, representing a page
474 pmap_pd_index(vm_offset_t va
)
476 return ((va
>> PDPSHIFT
) & ((1ul << NPDPEPGSHIFT
) - 1));
480 * Returns the index of a pdp in the pml4 table, representing a page
485 pmap_pdp_index(vm_offset_t va
)
487 return ((va
>> PML4SHIFT
) & ((1ul << NPML4EPGSHIFT
) - 1));
491 * The placemarker hash must be broken up into four zones so lock
492 * ordering semantics continue to work (e.g. pte, pt, pd, then pdp).
494 * Placemarkers are used to 'lock' page table indices that do not have
495 * a pv_entry. This allows the pmap to support managed and unmanaged
496 * pages and shared page tables.
498 #define PM_PLACE_BASE (PM_PLACEMARKS >> 2)
502 pmap_placemarker_hash(pmap_t pmap
, vm_pindex_t pindex
)
506 if (pindex
< pmap_pt_pindex(0)) /* zone 0 - PTE */
508 else if (pindex
< pmap_pd_pindex(0)) /* zone 1 - PT */
510 else if (pindex
< pmap_pdp_pindex(0)) /* zone 2 - PD */
511 hi
= PM_PLACE_BASE
<< 1;
512 else /* zone 3 - PDP (and PML4E) */
513 hi
= PM_PLACE_BASE
| (PM_PLACE_BASE
<< 1);
514 hi
+= pindex
& (PM_PLACE_BASE
- 1);
516 return (&pmap
->pm_placemarks
[hi
]);
521 * Generic procedure to index a pte from a pt, pd, or pdp.
523 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
524 * a page table page index but is instead of PV lookup index.
528 pv_pte_lookup(pv_entry_t pv
, vm_pindex_t pindex
)
532 pte
= (pt_entry_t
*)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv
->pv_m
));
533 return(&pte
[pindex
]);
537 * Return pointer to PDP slot in the PML4
541 pmap_pdp(pmap_t pmap
, vm_offset_t va
)
543 return (&pmap
->pm_pml4
[pmap_pdp_index(va
)]);
547 * Return pointer to PD slot in the PDP given a pointer to the PDP
551 pmap_pdp_to_pd(pml4_entry_t pdp_pte
, vm_offset_t va
)
555 pd
= (pdp_entry_t
*)PHYS_TO_DMAP(pdp_pte
& PG_FRAME
);
556 return (&pd
[pmap_pd_index(va
)]);
560 * Return pointer to PD slot in the PDP.
564 pmap_pd(pmap_t pmap
, vm_offset_t va
)
568 pdp
= pmap_pdp(pmap
, va
);
569 if ((*pdp
& pmap
->pmap_bits
[PG_V_IDX
]) == 0)
571 return (pmap_pdp_to_pd(*pdp
, va
));
575 * Return pointer to PT slot in the PD given a pointer to the PD
579 pmap_pd_to_pt(pdp_entry_t pd_pte
, vm_offset_t va
)
583 pt
= (pd_entry_t
*)PHYS_TO_DMAP(pd_pte
& PG_FRAME
);
584 return (&pt
[pmap_pt_index(va
)]);
588 * Return pointer to PT slot in the PD
590 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
591 * so we cannot lookup the PD via the PDP. Instead we
592 * must look it up via the pmap.
596 pmap_pt(pmap_t pmap
, vm_offset_t va
)
600 vm_pindex_t pd_pindex
;
602 if (pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) {
603 pd_pindex
= pmap_pd_pindex(va
);
604 spin_lock(&pmap
->pm_spin
);
605 pv
= pv_entry_rb_tree_RB_LOOKUP(&pmap
->pm_pvroot
, pd_pindex
);
606 spin_unlock(&pmap
->pm_spin
);
607 if (pv
== NULL
|| pv
->pv_m
== NULL
)
609 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv
->pv_m
), va
));
611 pd
= pmap_pd(pmap
, va
);
612 if (pd
== NULL
|| (*pd
& pmap
->pmap_bits
[PG_V_IDX
]) == 0)
614 return (pmap_pd_to_pt(*pd
, va
));
619 * Return pointer to PTE slot in the PT given a pointer to the PT
623 pmap_pt_to_pte(pd_entry_t pt_pte
, vm_offset_t va
)
627 pte
= (pt_entry_t
*)PHYS_TO_DMAP(pt_pte
& PG_FRAME
);
628 return (&pte
[pmap_pte_index(va
)]);
632 * Return pointer to PTE slot in the PT
636 pmap_pte(pmap_t pmap
, vm_offset_t va
)
640 pt
= pmap_pt(pmap
, va
);
641 if (pt
== NULL
|| (*pt
& pmap
->pmap_bits
[PG_V_IDX
]) == 0)
643 if ((*pt
& pmap
->pmap_bits
[PG_PS_IDX
]) != 0)
644 return ((pt_entry_t
*)pt
);
645 return (pmap_pt_to_pte(*pt
, va
));
649 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
650 * the PT layer. This will speed up core pmap operations considerably.
652 * NOTE: The pmap spinlock does not need to be held but the passed-in pv
653 * must be in a known associated state (typically by being locked when
654 * the pmap spinlock isn't held). We allow the race for that case.
656 * NOTE: pm_pvhint is only accessed (read) with the spin-lock held, using
657 * cpu_ccfence() to prevent compiler optimizations from reloading the
662 pv_cache(pv_entry_t pv
, vm_pindex_t pindex
)
664 if (pindex
>= pmap_pt_pindex(0) && pindex
< pmap_pd_pindex(0)) {
666 pv
->pv_pmap
->pm_pvhint
= pv
;
672 * Return address of PT slot in PD (KVM only)
674 * Cannot be used for user page tables because it might interfere with
675 * the shared page-table-page optimization (pmap_mmu_optimize).
679 vtopt(vm_offset_t va
)
681 uint64_t mask
= ((1ul << (NPDEPGSHIFT
+ NPDPEPGSHIFT
+
682 NPML4EPGSHIFT
)) - 1);
684 return (PDmap
+ ((va
>> PDRSHIFT
) & mask
));
688 * KVM - return address of PTE slot in PT
692 vtopte(vm_offset_t va
)
694 uint64_t mask
= ((1ul << (NPTEPGSHIFT
+ NPDEPGSHIFT
+
695 NPDPEPGSHIFT
+ NPML4EPGSHIFT
)) - 1);
697 return (PTmap
+ ((va
>> PAGE_SHIFT
) & mask
));
701 allocpages(vm_paddr_t
*firstaddr
, long n
)
706 bzero((void *)ret
, n
* PAGE_SIZE
);
707 *firstaddr
+= n
* PAGE_SIZE
;
713 create_pagetables(vm_paddr_t
*firstaddr
)
715 long i
; /* must be 64 bits */
721 * We are running (mostly) V=P at this point
723 * Calculate NKPT - number of kernel page tables. We have to
724 * accomodoate prealloction of the vm_page_array, dump bitmap,
725 * MSGBUF_SIZE, and other stuff. Be generous.
727 * Maxmem is in pages.
729 * ndmpdp is the number of 1GB pages we wish to map.
731 ndmpdp
= (ptoa(Maxmem
) + NBPDP
- 1) >> PDPSHIFT
;
732 if (ndmpdp
< 4) /* Minimum 4GB of dirmap */
734 KKASSERT(ndmpdp
<= NKPDPE
* NPDEPG
);
737 * Starting at the beginning of kvm (not KERNBASE).
739 nkpt_phys
= (Maxmem
* sizeof(struct vm_page
) + NBPDR
- 1) / NBPDR
;
740 nkpt_phys
+= (Maxmem
* sizeof(struct pv_entry
) + NBPDR
- 1) / NBPDR
;
741 nkpt_phys
+= ((nkpt
+ nkpt
+ 1 + NKPML4E
+ NKPDPE
+ NDMPML4E
+
742 ndmpdp
) + 511) / 512;
746 * Starting at KERNBASE - map 2G worth of page table pages.
747 * KERNBASE is offset -2G from the end of kvm.
749 nkpt_base
= (NPDPEPG
- KPDPI
) * NPTEPG
; /* typically 2 x 512 */
754 KPTbase
= allocpages(firstaddr
, nkpt_base
);
755 KPTphys
= allocpages(firstaddr
, nkpt_phys
);
756 KPML4phys
= allocpages(firstaddr
, 1);
757 KPDPphys
= allocpages(firstaddr
, NKPML4E
);
758 KPDphys
= allocpages(firstaddr
, NKPDPE
);
761 * Calculate the page directory base for KERNBASE,
762 * that is where we start populating the page table pages.
763 * Basically this is the end - 2.
765 KPDbase
= KPDphys
+ ((NKPDPE
- (NPDPEPG
- KPDPI
)) << PAGE_SHIFT
);
767 DMPDPphys
= allocpages(firstaddr
, NDMPML4E
);
768 if ((amd_feature
& AMDID_PAGE1GB
) == 0)
769 DMPDphys
= allocpages(firstaddr
, ndmpdp
);
770 dmaplimit
= (vm_paddr_t
)ndmpdp
<< PDPSHIFT
;
773 * Fill in the underlying page table pages for the area around
774 * KERNBASE. This remaps low physical memory to KERNBASE.
776 * Read-only from zero to physfree
777 * XXX not fully used, underneath 2M pages
779 for (i
= 0; (i
<< PAGE_SHIFT
) < *firstaddr
; i
++) {
780 ((pt_entry_t
*)KPTbase
)[i
] = i
<< PAGE_SHIFT
;
781 ((pt_entry_t
*)KPTbase
)[i
] |=
782 pmap_bits_default
[PG_RW_IDX
] |
783 pmap_bits_default
[PG_V_IDX
] |
784 pmap_bits_default
[PG_G_IDX
];
788 * Now map the initial kernel page tables. One block of page
789 * tables is placed at the beginning of kernel virtual memory,
790 * and another block is placed at KERNBASE to map the kernel binary,
791 * data, bss, and initial pre-allocations.
793 for (i
= 0; i
< nkpt_base
; i
++) {
794 ((pd_entry_t
*)KPDbase
)[i
] = KPTbase
+ (i
<< PAGE_SHIFT
);
795 ((pd_entry_t
*)KPDbase
)[i
] |=
796 pmap_bits_default
[PG_RW_IDX
] |
797 pmap_bits_default
[PG_V_IDX
];
799 for (i
= 0; i
< nkpt_phys
; i
++) {
800 ((pd_entry_t
*)KPDphys
)[i
] = KPTphys
+ (i
<< PAGE_SHIFT
);
801 ((pd_entry_t
*)KPDphys
)[i
] |=
802 pmap_bits_default
[PG_RW_IDX
] |
803 pmap_bits_default
[PG_V_IDX
];
807 * Map from zero to end of allocations using 2M pages as an
808 * optimization. This will bypass some of the KPTBase pages
809 * above in the KERNBASE area.
811 for (i
= 0; (i
<< PDRSHIFT
) < *firstaddr
; i
++) {
812 ((pd_entry_t
*)KPDbase
)[i
] = i
<< PDRSHIFT
;
813 ((pd_entry_t
*)KPDbase
)[i
] |=
814 pmap_bits_default
[PG_RW_IDX
] |
815 pmap_bits_default
[PG_V_IDX
] |
816 pmap_bits_default
[PG_PS_IDX
] |
817 pmap_bits_default
[PG_G_IDX
];
821 * And connect up the PD to the PDP. The kernel pmap is expected
822 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
824 for (i
= 0; i
< NKPDPE
; i
++) {
825 ((pdp_entry_t
*)KPDPphys
)[NPDPEPG
- NKPDPE
+ i
] =
826 KPDphys
+ (i
<< PAGE_SHIFT
);
827 ((pdp_entry_t
*)KPDPphys
)[NPDPEPG
- NKPDPE
+ i
] |=
828 pmap_bits_default
[PG_RW_IDX
] |
829 pmap_bits_default
[PG_V_IDX
] |
830 pmap_bits_default
[PG_U_IDX
];
834 * Now set up the direct map space using either 2MB or 1GB pages
835 * Preset PG_M and PG_A because demotion expects it.
837 * When filling in entries in the PD pages make sure any excess
838 * entries are set to zero as we allocated enough PD pages
840 if ((amd_feature
& AMDID_PAGE1GB
) == 0) {
841 for (i
= 0; i
< NPDEPG
* ndmpdp
; i
++) {
842 ((pd_entry_t
*)DMPDphys
)[i
] = i
<< PDRSHIFT
;
843 ((pd_entry_t
*)DMPDphys
)[i
] |=
844 pmap_bits_default
[PG_RW_IDX
] |
845 pmap_bits_default
[PG_V_IDX
] |
846 pmap_bits_default
[PG_PS_IDX
] |
847 pmap_bits_default
[PG_G_IDX
] |
848 pmap_bits_default
[PG_M_IDX
] |
849 pmap_bits_default
[PG_A_IDX
];
853 * And the direct map space's PDP
855 for (i
= 0; i
< ndmpdp
; i
++) {
856 ((pdp_entry_t
*)DMPDPphys
)[i
] = DMPDphys
+
858 ((pdp_entry_t
*)DMPDPphys
)[i
] |=
859 pmap_bits_default
[PG_RW_IDX
] |
860 pmap_bits_default
[PG_V_IDX
] |
861 pmap_bits_default
[PG_U_IDX
];
864 for (i
= 0; i
< ndmpdp
; i
++) {
865 ((pdp_entry_t
*)DMPDPphys
)[i
] =
866 (vm_paddr_t
)i
<< PDPSHIFT
;
867 ((pdp_entry_t
*)DMPDPphys
)[i
] |=
868 pmap_bits_default
[PG_RW_IDX
] |
869 pmap_bits_default
[PG_V_IDX
] |
870 pmap_bits_default
[PG_PS_IDX
] |
871 pmap_bits_default
[PG_G_IDX
] |
872 pmap_bits_default
[PG_M_IDX
] |
873 pmap_bits_default
[PG_A_IDX
];
877 /* And recursively map PML4 to itself in order to get PTmap */
878 ((pdp_entry_t
*)KPML4phys
)[PML4PML4I
] = KPML4phys
;
879 ((pdp_entry_t
*)KPML4phys
)[PML4PML4I
] |=
880 pmap_bits_default
[PG_RW_IDX
] |
881 pmap_bits_default
[PG_V_IDX
] |
882 pmap_bits_default
[PG_U_IDX
];
885 * Connect the Direct Map slots up to the PML4
887 for (j
= 0; j
< NDMPML4E
; ++j
) {
888 ((pdp_entry_t
*)KPML4phys
)[DMPML4I
+ j
] =
889 (DMPDPphys
+ ((vm_paddr_t
)j
<< PML4SHIFT
)) |
890 pmap_bits_default
[PG_RW_IDX
] |
891 pmap_bits_default
[PG_V_IDX
] |
892 pmap_bits_default
[PG_U_IDX
];
896 * Connect the KVA slot up to the PML4
898 ((pdp_entry_t
*)KPML4phys
)[KPML4I
] = KPDPphys
;
899 ((pdp_entry_t
*)KPML4phys
)[KPML4I
] |=
900 pmap_bits_default
[PG_RW_IDX
] |
901 pmap_bits_default
[PG_V_IDX
] |
902 pmap_bits_default
[PG_U_IDX
];
906 * Bootstrap the system enough to run with virtual memory.
908 * On the i386 this is called after mapping has already been enabled
909 * and just syncs the pmap module with what has already been done.
910 * [We can't call it easily with mapping off since the kernel is not
911 * mapped with PA == VA, hence we would have to relocate every address
912 * from the linked base (virtual) address "KERNBASE" to the actual
913 * (physical) address starting relative to 0]
916 pmap_bootstrap(vm_paddr_t
*firstaddr
)
922 KvaStart
= VM_MIN_KERNEL_ADDRESS
;
923 KvaEnd
= VM_MAX_KERNEL_ADDRESS
;
924 KvaSize
= KvaEnd
- KvaStart
;
926 avail_start
= *firstaddr
;
929 * Create an initial set of page tables to run the kernel in.
931 create_pagetables(firstaddr
);
933 virtual2_start
= KvaStart
;
934 virtual2_end
= PTOV_OFFSET
;
936 virtual_start
= (vm_offset_t
) PTOV_OFFSET
+ *firstaddr
;
937 virtual_start
= pmap_kmem_choose(virtual_start
);
939 virtual_end
= VM_MAX_KERNEL_ADDRESS
;
941 /* XXX do %cr0 as well */
942 load_cr4(rcr4() | CR4_PGE
| CR4_PSE
);
946 * Initialize protection array.
948 i386_protection_init();
951 * The kernel's pmap is statically allocated so we don't have to use
952 * pmap_create, which is unlikely to work correctly at this part of
953 * the boot sequence (XXX and which no longer exists).
955 kernel_pmap
.pm_pml4
= (pdp_entry_t
*) (PTOV_OFFSET
+ KPML4phys
);
956 kernel_pmap
.pm_count
= 1;
957 CPUMASK_ASSALLONES(kernel_pmap
.pm_active
);
958 RB_INIT(&kernel_pmap
.pm_pvroot
);
959 spin_init(&kernel_pmap
.pm_spin
, "pmapbootstrap");
960 for (i
= 0; i
< PM_PLACEMARKS
; ++i
)
961 kernel_pmap
.pm_placemarks
[i
] = PM_NOPLACEMARK
;
964 * Reserve some special page table entries/VA space for temporary
967 #define SYSMAP(c, p, v, n) \
968 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
974 * CMAP1/CMAP2 are used for zeroing and copying pages.
976 SYSMAP(caddr_t
, CMAP1
, CADDR1
, 1)
981 SYSMAP(caddr_t
, pt_crashdumpmap
, crashdumpmap
, MAXDUMPPGS
);
984 * ptvmmap is used for reading arbitrary physical pages via
987 SYSMAP(caddr_t
, ptmmap
, ptvmmap
, 1)
990 * msgbufp is used to map the system message buffer.
991 * XXX msgbufmap is not used.
993 SYSMAP(struct msgbuf
*, msgbufmap
, msgbufp
,
994 atop(round_page(MSGBUF_SIZE
)))
997 virtual_start
= pmap_kmem_choose(virtual_start
);
1002 * PG_G is terribly broken on SMP because we IPI invltlb's in some
1003 * cases rather then invl1pg. Actually, I don't even know why it
1004 * works under UP because self-referential page table mappings
1009 * Initialize the 4MB page size flag
1013 * The 4MB page version of the initial
1014 * kernel page mapping.
1018 #if !defined(DISABLE_PSE)
1019 if (cpu_feature
& CPUID_PSE
) {
1022 * Note that we have enabled PSE mode
1024 // pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
1025 ptditmp
= *(PTmap
+ x86_64_btop(KERNBASE
));
1026 ptditmp
&= ~(NBPDR
- 1);
1027 ptditmp
|= pmap_bits_default
[PG_V_IDX
] |
1028 pmap_bits_default
[PG_RW_IDX
] |
1029 pmap_bits_default
[PG_PS_IDX
] |
1030 pmap_bits_default
[PG_U_IDX
];
1037 /* Initialize the PAT MSR */
1039 pmap_pinit_defaults(&kernel_pmap
);
1041 TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
1042 &pmap_fast_kernel_cpusync
);
1047 * Setup the PAT MSR.
1056 * Default values mapping PATi,PCD,PWT bits at system reset.
1057 * The default values effectively ignore the PATi bit by
1058 * repeating the encodings for 0-3 in 4-7, and map the PCD
1059 * and PWT bit combinations to the expected PAT types.
1061 pat_msr
= PAT_VALUE(0, PAT_WRITE_BACK
) | /* 000 */
1062 PAT_VALUE(1, PAT_WRITE_THROUGH
) | /* 001 */
1063 PAT_VALUE(2, PAT_UNCACHED
) | /* 010 */
1064 PAT_VALUE(3, PAT_UNCACHEABLE
) | /* 011 */
1065 PAT_VALUE(4, PAT_WRITE_BACK
) | /* 100 */
1066 PAT_VALUE(5, PAT_WRITE_THROUGH
) | /* 101 */
1067 PAT_VALUE(6, PAT_UNCACHED
) | /* 110 */
1068 PAT_VALUE(7, PAT_UNCACHEABLE
); /* 111 */
1069 pat_pte_index
[PAT_WRITE_BACK
] = 0;
1070 pat_pte_index
[PAT_WRITE_THROUGH
]= 0 | X86_PG_NC_PWT
;
1071 pat_pte_index
[PAT_UNCACHED
] = X86_PG_NC_PCD
;
1072 pat_pte_index
[PAT_UNCACHEABLE
] = X86_PG_NC_PCD
| X86_PG_NC_PWT
;
1073 pat_pte_index
[PAT_WRITE_PROTECTED
] = pat_pte_index
[PAT_UNCACHEABLE
];
1074 pat_pte_index
[PAT_WRITE_COMBINING
] = pat_pte_index
[PAT_UNCACHEABLE
];
1076 if (cpu_feature
& CPUID_PAT
) {
1078 * If we support the PAT then set-up entries for
1079 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1082 pat_msr
= (pat_msr
& ~PAT_MASK(5)) |
1083 PAT_VALUE(5, PAT_WRITE_PROTECTED
);
1084 pat_msr
= (pat_msr
& ~PAT_MASK(6)) |
1085 PAT_VALUE(6, PAT_WRITE_COMBINING
);
1086 pat_pte_index
[PAT_WRITE_PROTECTED
] = X86_PG_PTE_PAT
| X86_PG_NC_PWT
;
1087 pat_pte_index
[PAT_WRITE_COMBINING
] = X86_PG_PTE_PAT
| X86_PG_NC_PCD
;
1090 * Then enable the PAT
1095 load_cr4(cr4
& ~CR4_PGE
);
1097 /* Disable caches (CD = 1, NW = 0). */
1099 load_cr0((cr0
& ~CR0_NW
) | CR0_CD
);
1101 /* Flushes caches and TLBs. */
1105 /* Update PAT and index table. */
1106 wrmsr(MSR_PAT
, pat_msr
);
1108 /* Flush caches and TLBs again. */
1112 /* Restore caches and PGE. */
1120 * Set 4mb pdir for mp startup
1125 if (cpu_feature
& CPUID_PSE
) {
1126 load_cr4(rcr4() | CR4_PSE
);
1127 if (pdir4mb
&& mycpu
->gd_cpuid
== 0) { /* only on BSP */
1134 * Initialize the pmap module.
1135 * Called by vm_init, to initialize any structures that the pmap
1136 * system needs to map virtual memory.
1137 * pmap_init has been enhanced to support in a fairly consistant
1138 * way, discontiguous physical memory.
1147 * Allocate memory for random pmap data structures. Includes the
1151 for (i
= 0; i
< vm_page_array_size
; i
++) {
1154 m
= &vm_page_array
[i
];
1155 TAILQ_INIT(&m
->md
.pv_list
);
1159 * init the pv free list
1161 initial_pvs
= vm_page_array_size
;
1162 if (initial_pvs
< MINPV
)
1163 initial_pvs
= MINPV
;
1164 pvzone
= &pvzone_store
;
1165 pvinit
= (void *)kmem_alloc(&kernel_map
,
1166 initial_pvs
* sizeof (struct pv_entry
),
1168 zbootinit(pvzone
, "PV ENTRY", sizeof (struct pv_entry
),
1169 pvinit
, initial_pvs
);
1172 * Now it is safe to enable pv_table recording.
1174 pmap_initialized
= TRUE
;
1178 * Initialize the address space (zone) for the pv_entries. Set a
1179 * high water mark so that the system can recover from excessive
1180 * numbers of pv entries.
1185 int shpgperproc
= PMAP_SHPGPERPROC
;
1188 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc
);
1189 pv_entry_max
= shpgperproc
* maxproc
+ vm_page_array_size
;
1190 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max
);
1191 pv_entry_high_water
= 9 * (pv_entry_max
/ 10);
1194 * Subtract out pages already installed in the zone (hack)
1196 entry_max
= pv_entry_max
- vm_page_array_size
;
1200 zinitna(pvzone
, NULL
, 0, entry_max
, ZONE_INTERRUPT
);
1203 * Enable dynamic deletion of empty higher-level page table pages
1204 * by default only if system memory is < 8GB (use 7GB for slop).
1205 * This can save a little memory, but imposes significant
1206 * performance overhead for things like bulk builds, and for programs
1207 * which do a lot of memory mapping and memory unmapping.
1209 if (pmap_dynamic_delete
< 0) {
1210 if (vmstats
.v_page_count
< 7LL * 1024 * 1024 * 1024 / PAGE_SIZE
)
1211 pmap_dynamic_delete
= 1;
1213 pmap_dynamic_delete
= 0;
1218 * Typically used to initialize a fictitious page by vm/device_pager.c
1221 pmap_page_init(struct vm_page
*m
)
1224 TAILQ_INIT(&m
->md
.pv_list
);
1227 /***************************************************
1228 * Low level helper routines.....
1229 ***************************************************/
1232 * this routine defines the region(s) of memory that should
1233 * not be tested for the modified bit.
1237 pmap_track_modified(vm_pindex_t pindex
)
1239 vm_offset_t va
= (vm_offset_t
)pindex
<< PAGE_SHIFT
;
1240 if ((va
< clean_sva
) || (va
>= clean_eva
))
1247 * Extract the physical page address associated with the map/VA pair.
1248 * The page must be wired for this to work reliably.
1251 pmap_extract(pmap_t pmap
, vm_offset_t va
, void **handlep
)
1258 if (va
>= VM_MAX_USER_ADDRESS
) {
1260 * Kernel page directories might be direct-mapped and
1261 * there is typically no PV tracking of pte's
1265 pt
= pmap_pt(pmap
, va
);
1266 if (pt
&& (*pt
& pmap
->pmap_bits
[PG_V_IDX
])) {
1267 if (*pt
& pmap
->pmap_bits
[PG_PS_IDX
]) {
1268 rtval
= *pt
& PG_PS_FRAME
;
1269 rtval
|= va
& PDRMASK
;
1271 ptep
= pmap_pt_to_pte(*pt
, va
);
1272 if (*pt
& pmap
->pmap_bits
[PG_V_IDX
]) {
1273 rtval
= *ptep
& PG_FRAME
;
1274 rtval
|= va
& PAGE_MASK
;
1282 * User pages currently do not direct-map the page directory
1283 * and some pages might not used managed PVs. But all PT's
1286 pt_pv
= pv_get(pmap
, pmap_pt_pindex(va
), NULL
);
1288 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
1289 if (*ptep
& pmap
->pmap_bits
[PG_V_IDX
]) {
1290 rtval
= *ptep
& PG_FRAME
;
1291 rtval
|= va
& PAGE_MASK
;
1294 *handlep
= pt_pv
; /* locked until done */
1297 } else if (handlep
) {
1305 pmap_extract_done(void *handle
)
1308 pv_put((pv_entry_t
)handle
);
1312 * Similar to extract but checks protections, SMP-friendly short-cut for
1313 * vm_fault_page[_quick](). Can return NULL to cause the caller to
1314 * fall-through to the real fault code. Does not work with HVM page
1317 * The returned page, if not NULL, is held (and not busied).
1319 * WARNING! THE RETURNED PAGE IS ONLY HELD AND NOT SUITABLE FOR READING
1323 pmap_fault_page_quick(pmap_t pmap
, vm_offset_t va
, vm_prot_t prot
, int *busyp
)
1326 va
< VM_MAX_USER_ADDRESS
&&
1327 (pmap
->pm_flags
& PMAP_HVM
) == 0) {
1335 req
= pmap
->pmap_bits
[PG_V_IDX
] |
1336 pmap
->pmap_bits
[PG_U_IDX
];
1337 if (prot
& VM_PROT_WRITE
)
1338 req
|= pmap
->pmap_bits
[PG_RW_IDX
];
1340 pt_pv
= pv_get(pmap
, pmap_pt_pindex(va
), NULL
);
1343 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
1344 if ((*ptep
& req
) != req
) {
1348 pte_pv
= pv_get_try(pmap
, pmap_pte_pindex(va
), NULL
, &error
);
1349 if (pte_pv
&& error
== 0) {
1351 if (prot
& VM_PROT_WRITE
) {
1352 /* interlocked by presence of pv_entry */
1356 if (prot
& VM_PROT_WRITE
) {
1357 if (vm_page_busy_try(m
, TRUE
))
1368 } else if (pte_pv
) {
1372 /* error, since we didn't request a placemarker */
1383 * Extract the physical page address associated kernel virtual address.
1386 pmap_kextract(vm_offset_t va
)
1388 pd_entry_t pt
; /* pt entry in pd */
1391 if (va
>= DMAP_MIN_ADDRESS
&& va
< DMAP_MAX_ADDRESS
) {
1392 pa
= DMAP_TO_PHYS(va
);
1395 if (pt
& kernel_pmap
.pmap_bits
[PG_PS_IDX
]) {
1396 pa
= (pt
& PG_PS_FRAME
) | (va
& PDRMASK
);
1399 * Beware of a concurrent promotion that changes the
1400 * PDE at this point! For example, vtopte() must not
1401 * be used to access the PTE because it would use the
1402 * new PDE. It is, however, safe to use the old PDE
1403 * because the page table page is preserved by the
1406 pa
= *pmap_pt_to_pte(pt
, va
);
1407 pa
= (pa
& PG_FRAME
) | (va
& PAGE_MASK
);
1413 /***************************************************
1414 * Low level mapping routines.....
1415 ***************************************************/
1418 * Routine: pmap_kenter
1420 * Add a wired page to the KVA
1421 * NOTE! note that in order for the mapping to take effect -- you
1422 * should do an invltlb after doing the pmap_kenter().
1425 pmap_kenter(vm_offset_t va
, vm_paddr_t pa
)
1431 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
1432 kernel_pmap
.pmap_bits
[PG_V_IDX
];
1436 pmap_inval_smp(&kernel_pmap
, va
, 1, ptep
, npte
);
1440 pmap_inval_smp(&kernel_pmap
, va
, ptep
, npte
);
1447 * Similar to pmap_kenter(), except we only invalidate the mapping on the
1448 * current CPU. Returns 0 if the previous pte was 0, 1 if it wasn't
1449 * (caller can conditionalize calling smp_invltlb()).
1452 pmap_kenter_quick(vm_offset_t va
, vm_paddr_t pa
)
1458 npte
= pa
| kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
1459 kernel_pmap
.pmap_bits
[PG_V_IDX
];
1468 atomic_swap_long(ptep
, npte
);
1469 cpu_invlpg((void *)va
);
1475 * Enter addresses into the kernel pmap but don't bother
1476 * doing any tlb invalidations. Caller will do a rollup
1477 * invalidation via pmap_rollup_inval().
1480 pmap_kenter_noinval(vm_offset_t va
, vm_paddr_t pa
)
1487 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
1488 kernel_pmap
.pmap_bits
[PG_V_IDX
];
1497 atomic_swap_long(ptep
, npte
);
1498 cpu_invlpg((void *)va
);
1504 * remove a page from the kernel pagetables
1507 pmap_kremove(vm_offset_t va
)
1512 pmap_inval_smp(&kernel_pmap
, va
, 1, ptep
, 0);
1516 pmap_kremove_quick(vm_offset_t va
)
1521 (void)pte_load_clear(ptep
);
1522 cpu_invlpg((void *)va
);
1526 * Remove addresses from the kernel pmap but don't bother
1527 * doing any tlb invalidations. Caller will do a rollup
1528 * invalidation via pmap_rollup_inval().
1531 pmap_kremove_noinval(vm_offset_t va
)
1536 (void)pte_load_clear(ptep
);
1540 * XXX these need to be recoded. They are not used in any critical path.
1543 pmap_kmodify_rw(vm_offset_t va
)
1545 atomic_set_long(vtopte(va
), kernel_pmap
.pmap_bits
[PG_RW_IDX
]);
1546 cpu_invlpg((void *)va
);
1551 pmap_kmodify_nc(vm_offset_t va)
1553 atomic_set_long(vtopte(va), PG_N);
1554 cpu_invlpg((void *)va);
1559 * Used to map a range of physical addresses into kernel virtual
1560 * address space during the low level boot, typically to map the
1561 * dump bitmap, message buffer, and vm_page_array.
1563 * These mappings are typically made at some pointer after the end of the
1566 * We could return PHYS_TO_DMAP(start) here and not allocate any
1567 * via (*virtp), but then kmem from userland and kernel dumps won't
1568 * have access to the related pointers.
1571 pmap_map(vm_offset_t
*virtp
, vm_paddr_t start
, vm_paddr_t end
, int prot
)
1574 vm_offset_t va_start
;
1576 /*return PHYS_TO_DMAP(start);*/
1581 while (start
< end
) {
1582 pmap_kenter_quick(va
, start
);
1590 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1593 * Remove the specified set of pages from the data and instruction caches.
1595 * In contrast to pmap_invalidate_cache_range(), this function does not
1596 * rely on the CPU's self-snoop feature, because it is intended for use
1597 * when moving pages into a different cache domain.
1600 pmap_invalidate_cache_pages(vm_page_t
*pages
, int count
)
1602 vm_offset_t daddr
, eva
;
1605 if (count
>= PMAP_CLFLUSH_THRESHOLD
/ PAGE_SIZE
||
1606 (cpu_feature
& CPUID_CLFSH
) == 0)
1610 for (i
= 0; i
< count
; i
++) {
1611 daddr
= PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages
[i
]));
1612 eva
= daddr
+ PAGE_SIZE
;
1613 for (; daddr
< eva
; daddr
+= cpu_clflush_line_size
)
1621 pmap_invalidate_cache_range(vm_offset_t sva
, vm_offset_t eva
)
1623 KASSERT((sva
& PAGE_MASK
) == 0,
1624 ("pmap_invalidate_cache_range: sva not page-aligned"));
1625 KASSERT((eva
& PAGE_MASK
) == 0,
1626 ("pmap_invalidate_cache_range: eva not page-aligned"));
1628 if (cpu_feature
& CPUID_SS
) {
1629 ; /* If "Self Snoop" is supported, do nothing. */
1631 /* Globally invalidate caches */
1632 cpu_wbinvd_on_all_cpus();
1637 * Invalidate the specified range of virtual memory on all cpus associated
1641 pmap_invalidate_range(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
)
1643 pmap_inval_smp(pmap
, sva
, (eva
- sva
) >> PAGE_SHIFT
, NULL
, 0);
1647 * Add a list of wired pages to the kva. This routine is used for temporary
1648 * kernel mappings such as those found in buffer cache buffer. Page
1649 * modifications and accesses are not tracked or recorded.
1651 * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1652 * semantics as previous mappings may have been zerod without any
1655 * The page *must* be wired.
1657 static __inline
void
1658 _pmap_qenter(vm_offset_t beg_va
, vm_page_t
*m
, int count
, int doinval
)
1663 end_va
= beg_va
+ count
* PAGE_SIZE
;
1665 for (va
= beg_va
; va
< end_va
; va
+= PAGE_SIZE
) {
1670 pte
= VM_PAGE_TO_PHYS(*m
) |
1671 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
1672 kernel_pmap
.pmap_bits
[PG_V_IDX
] |
1673 kernel_pmap
.pmap_cache_bits
[(*m
)->pat_mode
];
1675 atomic_swap_long(ptep
, pte
);
1679 pmap_invalidate_range(&kernel_pmap
, beg_va
, end_va
);
1683 pmap_qenter(vm_offset_t beg_va
, vm_page_t
*m
, int count
)
1685 _pmap_qenter(beg_va
, m
, count
, 1);
1689 pmap_qenter_noinval(vm_offset_t beg_va
, vm_page_t
*m
, int count
)
1691 _pmap_qenter(beg_va
, m
, count
, 0);
1695 * This routine jerks page mappings from the kernel -- it is meant only
1696 * for temporary mappings such as those found in buffer cache buffers.
1697 * No recording modified or access status occurs.
1699 * MPSAFE, INTERRUPT SAFE (cluster callback)
1702 pmap_qremove(vm_offset_t beg_va
, int count
)
1707 end_va
= beg_va
+ count
* PAGE_SIZE
;
1709 for (va
= beg_va
; va
< end_va
; va
+= PAGE_SIZE
) {
1713 (void)pte_load_clear(pte
);
1714 cpu_invlpg((void *)va
);
1716 pmap_invalidate_range(&kernel_pmap
, beg_va
, end_va
);
1720 * This routine removes temporary kernel mappings, only invalidating them
1721 * on the current cpu. It should only be used under carefully controlled
1725 pmap_qremove_quick(vm_offset_t beg_va
, int count
)
1730 end_va
= beg_va
+ count
* PAGE_SIZE
;
1732 for (va
= beg_va
; va
< end_va
; va
+= PAGE_SIZE
) {
1736 (void)pte_load_clear(pte
);
1737 cpu_invlpg((void *)va
);
1742 * This routine removes temporary kernel mappings *without* invalidating
1743 * the TLB. It can only be used on permanent kva reservations such as those
1744 * found in buffer cache buffers, under carefully controlled circumstances.
1746 * NOTE: Repopulating these KVAs requires unconditional invalidation.
1747 * (pmap_qenter() does unconditional invalidation).
1750 pmap_qremove_noinval(vm_offset_t beg_va
, int count
)
1755 end_va
= beg_va
+ count
* PAGE_SIZE
;
1757 for (va
= beg_va
; va
< end_va
; va
+= PAGE_SIZE
) {
1761 (void)pte_load_clear(pte
);
1766 * Create a new thread and optionally associate it with a (new) process.
1767 * NOTE! the new thread's cpu may not equal the current cpu.
1770 pmap_init_thread(thread_t td
)
1772 /* enforce pcb placement & alignment */
1773 td
->td_pcb
= (struct pcb
*)(td
->td_kstack
+ td
->td_kstack_size
) - 1;
1774 td
->td_pcb
= (struct pcb
*)((intptr_t)td
->td_pcb
& ~(intptr_t)0xF);
1775 td
->td_savefpu
= &td
->td_pcb
->pcb_save
;
1776 td
->td_sp
= (char *)td
->td_pcb
; /* no -16 */
1780 * This routine directly affects the fork perf for a process.
1783 pmap_init_proc(struct proc
*p
)
1788 pmap_pinit_defaults(struct pmap
*pmap
)
1790 bcopy(pmap_bits_default
, pmap
->pmap_bits
,
1791 sizeof(pmap_bits_default
));
1792 bcopy(protection_codes
, pmap
->protection_codes
,
1793 sizeof(protection_codes
));
1794 bcopy(pat_pte_index
, pmap
->pmap_cache_bits
,
1795 sizeof(pat_pte_index
));
1796 pmap
->pmap_cache_mask
= X86_PG_NC_PWT
| X86_PG_NC_PCD
| X86_PG_PTE_PAT
;
1797 pmap
->copyinstr
= std_copyinstr
;
1798 pmap
->copyin
= std_copyin
;
1799 pmap
->copyout
= std_copyout
;
1800 pmap
->fubyte
= std_fubyte
;
1801 pmap
->subyte
= std_subyte
;
1802 pmap
->fuword32
= std_fuword32
;
1803 pmap
->fuword64
= std_fuword64
;
1804 pmap
->suword32
= std_suword32
;
1805 pmap
->suword64
= std_suword64
;
1806 pmap
->swapu32
= std_swapu32
;
1807 pmap
->swapu64
= std_swapu64
;
1810 * Initialize pmap0/vmspace0.
1812 * On architectures where the kernel pmap is not integrated into the user
1813 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1814 * kernel_pmap should be used to directly access the kernel_pmap.
1817 pmap_pinit0(struct pmap
*pmap
)
1821 pmap
->pm_pml4
= (pml4_entry_t
*)(PTOV_OFFSET
+ KPML4phys
);
1823 CPUMASK_ASSZERO(pmap
->pm_active
);
1824 pmap
->pm_pvhint
= NULL
;
1825 RB_INIT(&pmap
->pm_pvroot
);
1826 spin_init(&pmap
->pm_spin
, "pmapinit0");
1827 for (i
= 0; i
< PM_PLACEMARKS
; ++i
)
1828 pmap
->pm_placemarks
[i
] = PM_NOPLACEMARK
;
1829 bzero(&pmap
->pm_stats
, sizeof pmap
->pm_stats
);
1830 pmap_pinit_defaults(pmap
);
1834 * Initialize a preallocated and zeroed pmap structure,
1835 * such as one in a vmspace structure.
1838 pmap_pinit_simple(struct pmap
*pmap
)
1843 * Misc initialization
1846 CPUMASK_ASSZERO(pmap
->pm_active
);
1847 pmap
->pm_pvhint
= NULL
;
1848 pmap
->pm_flags
= PMAP_FLAG_SIMPLE
;
1850 pmap_pinit_defaults(pmap
);
1853 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1856 if (pmap
->pm_pmlpv
== NULL
) {
1857 RB_INIT(&pmap
->pm_pvroot
);
1858 bzero(&pmap
->pm_stats
, sizeof pmap
->pm_stats
);
1859 spin_init(&pmap
->pm_spin
, "pmapinitsimple");
1860 for (i
= 0; i
< PM_PLACEMARKS
; ++i
)
1861 pmap
->pm_placemarks
[i
] = PM_NOPLACEMARK
;
1866 pmap_pinit(struct pmap
*pmap
)
1871 if (pmap
->pm_pmlpv
) {
1872 if (pmap
->pmap_bits
[TYPE_IDX
] != REGULAR_PMAP
) {
1877 pmap_pinit_simple(pmap
);
1878 pmap
->pm_flags
&= ~PMAP_FLAG_SIMPLE
;
1881 * No need to allocate page table space yet but we do need a valid
1882 * page directory table.
1884 if (pmap
->pm_pml4
== NULL
) {
1886 (pml4_entry_t
*)kmem_alloc_pageable(&kernel_map
,
1892 * Allocate the page directory page, which wires it even though
1893 * it isn't being entered into some higher level page table (it
1894 * being the highest level). If one is already cached we don't
1895 * have to do anything.
1897 if ((pv
= pmap
->pm_pmlpv
) == NULL
) {
1898 pv
= pmap_allocpte(pmap
, pmap_pml4_pindex(), NULL
);
1899 pmap
->pm_pmlpv
= pv
;
1900 pmap_kenter((vm_offset_t
)pmap
->pm_pml4
,
1901 VM_PAGE_TO_PHYS(pv
->pv_m
));
1905 * Install DMAP and KMAP.
1907 for (j
= 0; j
< NDMPML4E
; ++j
) {
1908 pmap
->pm_pml4
[DMPML4I
+ j
] =
1909 (DMPDPphys
+ ((vm_paddr_t
)j
<< PML4SHIFT
)) |
1910 pmap
->pmap_bits
[PG_RW_IDX
] |
1911 pmap
->pmap_bits
[PG_V_IDX
] |
1912 pmap
->pmap_bits
[PG_U_IDX
];
1914 pmap
->pm_pml4
[KPML4I
] = KPDPphys
|
1915 pmap
->pmap_bits
[PG_RW_IDX
] |
1916 pmap
->pmap_bits
[PG_V_IDX
] |
1917 pmap
->pmap_bits
[PG_U_IDX
];
1920 * install self-referential address mapping entry
1922 pmap
->pm_pml4
[PML4PML4I
] = VM_PAGE_TO_PHYS(pv
->pv_m
) |
1923 pmap
->pmap_bits
[PG_V_IDX
] |
1924 pmap
->pmap_bits
[PG_RW_IDX
] |
1925 pmap
->pmap_bits
[PG_A_IDX
] |
1926 pmap
->pmap_bits
[PG_M_IDX
];
1928 KKASSERT(pv
->pv_m
->flags
& PG_MAPPED
);
1929 KKASSERT(pv
->pv_m
->flags
& PG_WRITEABLE
);
1931 KKASSERT(pmap
->pm_pml4
[255] == 0);
1932 KKASSERT(RB_ROOT(&pmap
->pm_pvroot
) == pv
);
1933 KKASSERT(pv
->pv_entry
.rbe_left
== NULL
);
1934 KKASSERT(pv
->pv_entry
.rbe_right
== NULL
);
1938 * Clean up a pmap structure so it can be physically freed. This routine
1939 * is called by the vmspace dtor function. A great deal of pmap data is
1940 * left passively mapped to improve vmspace management so we have a bit
1941 * of cleanup work to do here.
1944 pmap_puninit(pmap_t pmap
)
1949 KKASSERT(CPUMASK_TESTZERO(pmap
->pm_active
));
1950 if ((pv
= pmap
->pm_pmlpv
) != NULL
) {
1951 if (pv_hold_try(pv
) == 0)
1953 KKASSERT(pv
== pmap
->pm_pmlpv
);
1954 p
= pmap_remove_pv_page(pv
);
1956 pv
= NULL
; /* safety */
1957 pmap_kremove((vm_offset_t
)pmap
->pm_pml4
);
1958 vm_page_busy_wait(p
, FALSE
, "pgpun");
1959 KKASSERT(p
->flags
& (PG_FICTITIOUS
|PG_UNMANAGED
));
1960 vm_page_unwire(p
, 0);
1961 vm_page_flag_clear(p
, PG_MAPPED
| PG_WRITEABLE
);
1964 * XXX eventually clean out PML4 static entries and
1965 * use vm_page_free_zero()
1968 pmap
->pm_pmlpv
= NULL
;
1970 if (pmap
->pm_pml4
) {
1971 KKASSERT(pmap
->pm_pml4
!= (void *)(PTOV_OFFSET
+ KPML4phys
));
1972 kmem_free(&kernel_map
, (vm_offset_t
)pmap
->pm_pml4
, PAGE_SIZE
);
1973 pmap
->pm_pml4
= NULL
;
1975 KKASSERT(pmap
->pm_stats
.resident_count
== 0);
1976 KKASSERT(pmap
->pm_stats
.wired_count
== 0);
1980 * This function is now unused (used to add the pmap to the pmap_list)
1983 pmap_pinit2(struct pmap
*pmap
)
1988 * This routine is called when various levels in the page table need to
1989 * be populated. This routine cannot fail.
1991 * This function returns two locked pv_entry's, one representing the
1992 * requested pv and one representing the requested pv's parent pv. If
1993 * an intermediate page table does not exist it will be created, mapped,
1994 * wired, and the parent page table will be given an additional hold
1995 * count representing the presence of the child pv_entry.
1999 pmap_allocpte(pmap_t pmap
, vm_pindex_t ptepindex
, pv_entry_t
*pvpp
)
2005 vm_pindex_t pt_pindex
;
2011 * If the pv already exists and we aren't being asked for the
2012 * parent page table page we can just return it. A locked+held pv
2013 * is returned. The pv will also have a second hold related to the
2014 * pmap association that we don't have to worry about.
2017 pv
= pv_alloc(pmap
, ptepindex
, &isnew
);
2018 if (isnew
== 0 && pvpp
== NULL
)
2022 * Special case terminal PVs. These are not page table pages so
2023 * no vm_page is allocated (the caller supplied the vm_page). If
2024 * pvpp is non-NULL we are being asked to also removed the pt_pv
2027 * Note that pt_pv's are only returned for user VAs. We assert that
2028 * a pt_pv is not being requested for kernel VAs. The kernel
2029 * pre-wires all higher-level page tables so don't overload managed
2030 * higher-level page tables on top of it!
2032 if (ptepindex
< pmap_pt_pindex(0)) {
2033 if (ptepindex
>= NUPTE_USER
) {
2034 /* kernel manages this manually for KVM */
2035 KKASSERT(pvpp
== NULL
);
2037 KKASSERT(pvpp
!= NULL
);
2038 pt_pindex
= NUPTE_TOTAL
+ (ptepindex
>> NPTEPGSHIFT
);
2039 pvp
= pmap_allocpte(pmap
, pt_pindex
, NULL
);
2041 vm_page_wire_quick(pvp
->pv_m
);
2048 * The kernel never uses managed PT/PD/PDP pages.
2050 KKASSERT(pmap
!= &kernel_pmap
);
2053 * Non-terminal PVs allocate a VM page to represent the page table,
2054 * so we have to resolve pvp and calculate ptepindex for the pvp
2055 * and then for the page table entry index in the pvp for
2058 if (ptepindex
< pmap_pd_pindex(0)) {
2060 * pv is PT, pvp is PD
2062 ptepindex
= (ptepindex
- pmap_pt_pindex(0)) >> NPDEPGSHIFT
;
2063 ptepindex
+= NUPTE_TOTAL
+ NUPT_TOTAL
;
2064 pvp
= pmap_allocpte(pmap
, ptepindex
, NULL
);
2069 ptepindex
= pv
->pv_pindex
- pmap_pt_pindex(0);
2070 ptepindex
&= ((1ul << NPDEPGSHIFT
) - 1);
2072 } else if (ptepindex
< pmap_pdp_pindex(0)) {
2074 * pv is PD, pvp is PDP
2076 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
2079 ptepindex
= (ptepindex
- pmap_pd_pindex(0)) >> NPDPEPGSHIFT
;
2080 ptepindex
+= NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
;
2082 if (pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) {
2083 KKASSERT(pvpp
== NULL
);
2086 pvp
= pmap_allocpte(pmap
, ptepindex
, NULL
);
2092 ptepindex
= pv
->pv_pindex
- pmap_pd_pindex(0);
2093 ptepindex
&= ((1ul << NPDPEPGSHIFT
) - 1);
2094 } else if (ptepindex
< pmap_pml4_pindex()) {
2096 * pv is PDP, pvp is the root pml4 table
2098 pvp
= pmap_allocpte(pmap
, pmap_pml4_pindex(), NULL
);
2103 ptepindex
= pv
->pv_pindex
- pmap_pdp_pindex(0);
2104 ptepindex
&= ((1ul << NPML4EPGSHIFT
) - 1);
2107 * pv represents the top-level PML4, there is no parent.
2116 * (isnew) is TRUE, pv is not terminal.
2118 * (1) Add a wire count to the parent page table (pvp).
2119 * (2) Allocate a VM page for the page table.
2120 * (3) Enter the VM page into the parent page table.
2122 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2125 vm_page_wire_quick(pvp
->pv_m
);
2128 m
= vm_page_alloc(NULL
, pv
->pv_pindex
,
2129 VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
|
2130 VM_ALLOC_INTERRUPT
);
2135 vm_page_wire(m
); /* wire for mapping in parent */
2136 vm_page_unmanage(m
); /* m must be spinunlocked */
2137 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
2138 m
->valid
= VM_PAGE_BITS_ALL
;
2140 vm_page_spin_lock(m
);
2141 pmap_page_stats_adding(m
);
2142 TAILQ_INSERT_TAIL(&m
->md
.pv_list
, pv
, pv_list
);
2144 vm_page_flag_set(m
, PG_MAPPED
| PG_WRITEABLE
);
2145 vm_page_spin_unlock(m
);
2148 * (isnew) is TRUE, pv is not terminal.
2150 * Wire the page into pvp. Bump the resident_count for the pmap.
2151 * There is no pvp for the top level, address the pm_pml4[] array
2154 * If the caller wants the parent we return it, otherwise
2155 * we just put it away.
2157 * No interlock is needed for pte 0 -> non-zero.
2159 * In the situation where *ptep is valid we might have an unmanaged
2160 * page table page shared from another page table which we need to
2161 * unshare before installing our private page table page.
2164 v
= VM_PAGE_TO_PHYS(m
) |
2165 (pmap
->pmap_bits
[PG_U_IDX
] |
2166 pmap
->pmap_bits
[PG_RW_IDX
] |
2167 pmap
->pmap_bits
[PG_V_IDX
] |
2168 pmap
->pmap_bits
[PG_A_IDX
] |
2169 pmap
->pmap_bits
[PG_M_IDX
]);
2170 ptep
= pv_pte_lookup(pvp
, ptepindex
);
2171 if (*ptep
& pmap
->pmap_bits
[PG_V_IDX
]) {
2175 panic("pmap_allocpte: unexpected pte %p/%d",
2176 pvp
, (int)ptepindex
);
2178 pte
= pmap_inval_smp(pmap
, (vm_offset_t
)-1, 1, ptep
, v
);
2179 if (vm_page_unwire_quick(
2180 PHYS_TO_VM_PAGE(pte
& PG_FRAME
))) {
2181 panic("pmap_allocpte: shared pgtable "
2182 "pg bad wirecount");
2187 pte
= atomic_swap_long(ptep
, v
);
2189 kprintf("install pgtbl mixup 0x%016jx "
2190 "old/new 0x%016jx/0x%016jx\n",
2191 (intmax_t)ptepindex
, pte
, v
);
2198 * (isnew) may be TRUE or FALSE, pv may or may not be terminal.
2202 KKASSERT(pvp
->pv_m
!= NULL
);
2203 ptep
= pv_pte_lookup(pvp
, ptepindex
);
2204 v
= VM_PAGE_TO_PHYS(pv
->pv_m
) |
2205 (pmap
->pmap_bits
[PG_U_IDX
] |
2206 pmap
->pmap_bits
[PG_RW_IDX
] |
2207 pmap
->pmap_bits
[PG_V_IDX
] |
2208 pmap
->pmap_bits
[PG_A_IDX
] |
2209 pmap
->pmap_bits
[PG_M_IDX
]);
2211 kprintf("mismatched upper level pt %016jx/%016jx\n",
2223 * This version of pmap_allocpte() checks for possible segment optimizations
2224 * that would allow page-table sharing. It can be called for terminal
2225 * page or page table page ptepindex's.
2227 * The function is called with page table page ptepindex's for fictitious
2228 * and unmanaged terminal pages. That is, we don't want to allocate a
2229 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
2232 * This function can return a pv and *pvpp associated with the passed in pmap
2233 * OR a pv and *pvpp associated with the shared pmap. In the latter case
2234 * an unmanaged page table page will be entered into the pass in pmap.
2238 pmap_allocpte_seg(pmap_t pmap
, vm_pindex_t ptepindex
, pv_entry_t
*pvpp
,
2239 vm_map_entry_t entry
, vm_offset_t va
)
2245 pv_entry_t pte_pv
; /* in original or shared pmap */
2246 pv_entry_t pt_pv
; /* in original or shared pmap */
2247 pv_entry_t proc_pd_pv
; /* in original pmap */
2248 pv_entry_t proc_pt_pv
; /* in original pmap */
2249 pv_entry_t xpv
; /* PT in shared pmap */
2250 pd_entry_t
*pt
; /* PT entry in PD of original pmap */
2251 pd_entry_t opte
; /* contents of *pt */
2252 pd_entry_t npte
; /* contents of *pt */
2256 * Basic tests, require a non-NULL vm_map_entry, require proper
2257 * alignment and type for the vm_map_entry, require that the
2258 * underlying object already be allocated.
2260 * We allow almost any type of object to use this optimization.
2261 * The object itself does NOT have to be sized to a multiple of the
2262 * segment size, but the memory mapping does.
2264 * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2265 * won't work as expected.
2267 if (entry
== NULL
||
2268 pmap_mmu_optimize
== 0 || /* not enabled */
2269 (pmap
->pm_flags
& PMAP_HVM
) || /* special pmap */
2270 ptepindex
>= pmap_pd_pindex(0) || /* not terminal or pt */
2271 entry
->inheritance
!= VM_INHERIT_SHARE
|| /* not shared */
2272 entry
->maptype
!= VM_MAPTYPE_NORMAL
|| /* weird map type */
2273 entry
->object
.vm_object
== NULL
|| /* needs VM object */
2274 entry
->object
.vm_object
->type
== OBJT_DEVICE
|| /* ick */
2275 entry
->object
.vm_object
->type
== OBJT_MGTDEVICE
|| /* ick */
2276 (entry
->offset
& SEG_MASK
) || /* must be aligned */
2277 (entry
->start
& SEG_MASK
)) {
2278 return(pmap_allocpte(pmap
, ptepindex
, pvpp
));
2282 * Make sure the full segment can be represented.
2284 b
= va
& ~(vm_offset_t
)SEG_MASK
;
2285 if (b
< entry
->start
|| b
+ SEG_SIZE
> entry
->end
)
2286 return(pmap_allocpte(pmap
, ptepindex
, pvpp
));
2289 * If the full segment can be represented dive the VM object's
2290 * shared pmap, allocating as required.
2292 object
= entry
->object
.vm_object
;
2294 if (entry
->protection
& VM_PROT_WRITE
)
2295 obpmapp
= &object
->md
.pmap_rw
;
2297 obpmapp
= &object
->md
.pmap_ro
;
2300 if (pmap_enter_debug
> 0) {
2302 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2304 va
, entry
->protection
, object
,
2306 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2307 entry
, entry
->start
, entry
->end
);
2312 * We allocate what appears to be a normal pmap but because portions
2313 * of this pmap are shared with other unrelated pmaps we have to
2314 * set pm_active to point to all cpus.
2316 * XXX Currently using pmap_spin to interlock the update, can't use
2317 * vm_object_hold/drop because the token might already be held
2318 * shared OR exclusive and we don't know.
2320 while ((obpmap
= *obpmapp
) == NULL
) {
2321 obpmap
= kmalloc(sizeof(*obpmap
), M_OBJPMAP
, M_WAITOK
|M_ZERO
);
2322 pmap_pinit_simple(obpmap
);
2323 pmap_pinit2(obpmap
);
2324 spin_lock(&pmap_spin
);
2325 if (*obpmapp
!= NULL
) {
2329 spin_unlock(&pmap_spin
);
2330 pmap_release(obpmap
);
2331 pmap_puninit(obpmap
);
2332 kfree(obpmap
, M_OBJPMAP
);
2333 obpmap
= *obpmapp
; /* safety */
2335 obpmap
->pm_active
= smp_active_mask
;
2336 obpmap
->pm_flags
|= PMAP_SEGSHARED
;
2338 spin_unlock(&pmap_spin
);
2343 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
2344 * pte/pt using the shared pmap from the object but also adjust
2345 * the process pmap's page table page as a side effect.
2349 * Resolve the terminal PTE and PT in the shared pmap. This is what
2350 * we will return. This is true if ptepindex represents a terminal
2351 * page, otherwise pte_pv is actually the PT and pt_pv is actually
2355 pte_pv
= pmap_allocpte(obpmap
, ptepindex
, &pt_pv
);
2357 if (ptepindex
>= pmap_pt_pindex(0))
2363 * Resolve the PD in the process pmap so we can properly share the
2364 * page table page. Lock order is bottom-up (leaf first)!
2366 * NOTE: proc_pt_pv can be NULL.
2368 proc_pt_pv
= pv_get(pmap
, pmap_pt_pindex(b
), NULL
);
2369 proc_pd_pv
= pmap_allocpte(pmap
, pmap_pd_pindex(b
), NULL
);
2371 if (pmap_enter_debug
> 0) {
2373 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2375 (proc_pt_pv
? proc_pt_pv
->pv_m
->wire_count
: -1),
2382 * xpv is the page table page pv from the shared object
2383 * (for convenience), from above.
2385 * Calculate the pte value for the PT to load into the process PD.
2386 * If we have to change it we must properly dispose of the previous
2389 pt
= pv_pte_lookup(proc_pd_pv
, pmap_pt_index(b
));
2390 npte
= VM_PAGE_TO_PHYS(xpv
->pv_m
) |
2391 (pmap
->pmap_bits
[PG_U_IDX
] |
2392 pmap
->pmap_bits
[PG_RW_IDX
] |
2393 pmap
->pmap_bits
[PG_V_IDX
] |
2394 pmap
->pmap_bits
[PG_A_IDX
] |
2395 pmap
->pmap_bits
[PG_M_IDX
]);
2398 * Dispose of previous page table page if it was local to the
2399 * process pmap. If the old pt is not empty we cannot dispose of it
2400 * until we clean it out. This case should not arise very often so
2401 * it is not optimized.
2403 * Leave pt_pv and pte_pv (in our object pmap) locked and intact
2407 pmap_inval_bulk_t bulk
;
2409 if (proc_pt_pv
->pv_m
->wire_count
!= 1) {
2413 va
& ~(vm_offset_t
)SEG_MASK
,
2414 (va
+ SEG_SIZE
) & ~(vm_offset_t
)SEG_MASK
);
2419 * The release call will indirectly clean out *pt
2421 pmap_inval_bulk_init(&bulk
, proc_pt_pv
->pv_pmap
);
2422 pmap_release_pv(proc_pt_pv
, proc_pd_pv
, &bulk
);
2423 pmap_inval_bulk_flush(&bulk
);
2426 pt
= pv_pte_lookup(proc_pd_pv
, pmap_pt_index(b
));
2430 * Handle remaining cases.
2433 atomic_swap_long(pt
, npte
);
2434 vm_page_wire_quick(xpv
->pv_m
); /* shared pt -> proc */
2435 vm_page_wire_quick(proc_pd_pv
->pv_m
); /* proc pd for sh pt */
2436 atomic_add_long(&pmap
->pm_stats
.resident_count
, 1);
2437 } else if (*pt
!= npte
) {
2438 opte
= pmap_inval_smp(pmap
, (vm_offset_t
)-1, 1, pt
, npte
);
2441 opte
= pte_load_clear(pt
);
2442 KKASSERT(opte
&& opte
!= npte
);
2446 vm_page_wire_quick(xpv
->pv_m
); /* shared pt -> proc */
2449 * Clean up opte, bump the wire_count for the process
2450 * PD page representing the new entry if it was
2453 * If the entry was not previously empty and we have
2454 * a PT in the proc pmap then opte must match that
2455 * pt. The proc pt must be retired (this is done
2456 * later on in this procedure).
2458 * NOTE: replacing valid pte, wire_count on proc_pd_pv
2461 KKASSERT(opte
& pmap
->pmap_bits
[PG_V_IDX
]);
2462 m
= PHYS_TO_VM_PAGE(opte
& PG_FRAME
);
2463 if (vm_page_unwire_quick(m
)) {
2464 panic("pmap_allocpte_seg: "
2465 "bad wire count %p",
2471 * The existing process page table was replaced and must be destroyed
2485 * Release any resources held by the given physical map.
2487 * Called when a pmap initialized by pmap_pinit is being released. Should
2488 * only be called if the map contains no valid mappings.
2490 struct pmap_release_info
{
2496 static int pmap_release_callback(pv_entry_t pv
, void *data
);
2499 pmap_release(struct pmap
*pmap
)
2501 struct pmap_release_info info
;
2503 KASSERT(CPUMASK_TESTZERO(pmap
->pm_active
),
2504 ("pmap still active! %016jx",
2505 (uintmax_t)CPUMASK_LOWMASK(pmap
->pm_active
)));
2508 * There is no longer a pmap_list, if there were we would remove the
2509 * pmap from it here.
2513 * Pull pv's off the RB tree in order from low to high and release
2521 spin_lock(&pmap
->pm_spin
);
2522 RB_SCAN(pv_entry_rb_tree
, &pmap
->pm_pvroot
, NULL
,
2523 pmap_release_callback
, &info
);
2524 spin_unlock(&pmap
->pm_spin
);
2528 } while (info
.retry
);
2532 * One resident page (the pml4 page) should remain.
2533 * No wired pages should remain.
2536 if (pmap
->pm_stats
.resident_count
!=
2537 ((pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) ? 0 : 1) ||
2538 pmap
->pm_stats
.wired_count
!= 0) {
2539 kprintf("fatal pmap problem - pmap %p flags %08x "
2540 "rescnt=%jd wirecnt=%jd\n",
2543 pmap
->pm_stats
.resident_count
,
2544 pmap
->pm_stats
.wired_count
);
2545 tsleep(pmap
, 0, "DEAD", 0);
2548 KKASSERT(pmap
->pm_stats
.resident_count
==
2549 ((pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) ? 0 : 1));
2550 KKASSERT(pmap
->pm_stats
.wired_count
== 0);
2555 * Called from low to high. We must cache the proper parent pv so we
2556 * can adjust its wired count.
2559 pmap_release_callback(pv_entry_t pv
, void *data
)
2561 struct pmap_release_info
*info
= data
;
2562 pmap_t pmap
= info
->pmap
;
2567 * Acquire a held and locked pv, check for release race
2569 pindex
= pv
->pv_pindex
;
2570 if (info
->pvp
== pv
) {
2571 spin_unlock(&pmap
->pm_spin
);
2573 } else if (pv_hold_try(pv
)) {
2574 spin_unlock(&pmap
->pm_spin
);
2576 spin_unlock(&pmap
->pm_spin
);
2580 spin_lock(&pmap
->pm_spin
);
2584 KKASSERT(pv
->pv_pmap
== pmap
&& pindex
== pv
->pv_pindex
);
2586 if (pv
->pv_pindex
< pmap_pt_pindex(0)) {
2588 * I am PTE, parent is PT
2590 pindex
= pv
->pv_pindex
>> NPTEPGSHIFT
;
2591 pindex
+= NUPTE_TOTAL
;
2592 } else if (pv
->pv_pindex
< pmap_pd_pindex(0)) {
2594 * I am PT, parent is PD
2596 pindex
= (pv
->pv_pindex
- NUPTE_TOTAL
) >> NPDEPGSHIFT
;
2597 pindex
+= NUPTE_TOTAL
+ NUPT_TOTAL
;
2598 } else if (pv
->pv_pindex
< pmap_pdp_pindex(0)) {
2600 * I am PD, parent is PDP
2602 pindex
= (pv
->pv_pindex
- NUPTE_TOTAL
- NUPT_TOTAL
) >>
2604 pindex
+= NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
;
2605 } else if (pv
->pv_pindex
< pmap_pml4_pindex()) {
2607 * I am PDP, parent is PML4 (there's only one)
2610 pindex
= (pv
->pv_pindex
- NUPTE_TOTAL
- NUPT_TOTAL
-
2611 NUPD_TOTAL
) >> NPML4EPGSHIFT
;
2612 pindex
+= NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
+ NUPDP_TOTAL
;
2614 pindex
= pmap_pml4_pindex();
2626 if (info
->pvp
&& info
->pvp
->pv_pindex
!= pindex
) {
2630 if (info
->pvp
== NULL
)
2631 info
->pvp
= pv_get(pmap
, pindex
, NULL
);
2638 r
= pmap_release_pv(pv
, info
->pvp
, NULL
);
2639 spin_lock(&pmap
->pm_spin
);
2645 * Called with held (i.e. also locked) pv. This function will dispose of
2646 * the lock along with the pv.
2648 * If the caller already holds the locked parent page table for pv it
2649 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2650 * pass NULL for pvp.
2653 pmap_release_pv(pv_entry_t pv
, pv_entry_t pvp
, pmap_inval_bulk_t
*bulk
)
2658 * The pmap is currently not spinlocked, pv is held+locked.
2659 * Remove the pv's page from its parent's page table. The
2660 * parent's page table page's wire_count will be decremented.
2662 * This will clean out the pte at any level of the page table.
2663 * If smp != 0 all cpus are affected.
2665 * Do not tear-down recursively, its faster to just let the
2666 * release run its course.
2668 pmap_remove_pv_pte(pv
, pvp
, bulk
, 0);
2671 * Terminal pvs are unhooked from their vm_pages. Because
2672 * terminal pages aren't page table pages they aren't wired
2673 * by us, so we have to be sure not to unwire them either.
2675 if (pv
->pv_pindex
< pmap_pt_pindex(0)) {
2676 pmap_remove_pv_page(pv
);
2681 * We leave the top-level page table page cached, wired, and
2682 * mapped in the pmap until the dtor function (pmap_puninit())
2685 * Since we are leaving the top-level pv intact we need
2686 * to break out of what would otherwise be an infinite loop.
2688 if (pv
->pv_pindex
== pmap_pml4_pindex()) {
2694 * For page table pages (other than the top-level page),
2695 * remove and free the vm_page. The representitive mapping
2696 * removed above by pmap_remove_pv_pte() did not undo the
2697 * last wire_count so we have to do that as well.
2699 p
= pmap_remove_pv_page(pv
);
2700 vm_page_busy_wait(p
, FALSE
, "pmaprl");
2701 if (p
->wire_count
!= 1) {
2702 kprintf("p->wire_count was %016lx %d\n",
2703 pv
->pv_pindex
, p
->wire_count
);
2705 KKASSERT(p
->wire_count
== 1);
2706 KKASSERT(p
->flags
& PG_UNMANAGED
);
2708 vm_page_unwire(p
, 0);
2709 KKASSERT(p
->wire_count
== 0);
2719 * This function will remove the pte associated with a pv from its parent.
2720 * Terminal pv's are supported. All cpus specified by (bulk) are properly
2723 * The wire count will be dropped on the parent page table. The wire
2724 * count on the page being removed (pv->pv_m) from the parent page table
2725 * is NOT touched. Note that terminal pages will not have any additional
2726 * wire counts while page table pages will have at least one representing
2727 * the mapping, plus others representing sub-mappings.
2729 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2730 * pages and user page table and terminal pages.
2732 * NOTE: The pte being removed might be unmanaged, and the pv supplied might
2733 * be freshly allocated and not imply that the pte is managed. In this
2734 * case pv->pv_m should be NULL.
2736 * The pv must be locked. The pvp, if supplied, must be locked. All
2737 * supplied pv's will remain locked on return.
2739 * XXX must lock parent pv's if they exist to remove pte XXX
2743 pmap_remove_pv_pte(pv_entry_t pv
, pv_entry_t pvp
, pmap_inval_bulk_t
*bulk
,
2746 vm_pindex_t ptepindex
= pv
->pv_pindex
;
2747 pmap_t pmap
= pv
->pv_pmap
;
2753 if (ptepindex
== pmap_pml4_pindex()) {
2755 * We are the top level PML4E table, there is no parent.
2757 p
= pmap
->pm_pmlpv
->pv_m
;
2758 KKASSERT(pv
->pv_m
== p
); /* debugging */
2759 } else if (ptepindex
>= pmap_pdp_pindex(0)) {
2761 * Remove a PDP page from the PML4E. This can only occur
2762 * with user page tables. We do not have to lock the
2763 * pml4 PV so just ignore pvp.
2765 vm_pindex_t pml4_pindex
;
2766 vm_pindex_t pdp_index
;
2769 pdp_index
= ptepindex
- pmap_pdp_pindex(0);
2771 pml4_pindex
= pmap_pml4_pindex();
2772 pvp
= pv_get(pv
->pv_pmap
, pml4_pindex
, NULL
);
2777 pdp
= &pmap
->pm_pml4
[pdp_index
& ((1ul << NPML4EPGSHIFT
) - 1)];
2778 KKASSERT((*pdp
& pmap
->pmap_bits
[PG_V_IDX
]) != 0);
2779 p
= PHYS_TO_VM_PAGE(*pdp
& PG_FRAME
);
2780 pmap_inval_bulk(bulk
, (vm_offset_t
)-1, pdp
, 0);
2781 KKASSERT(pv
->pv_m
== p
); /* debugging */
2782 } else if (ptepindex
>= pmap_pd_pindex(0)) {
2784 * Remove a PD page from the PDP
2786 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2787 * of a simple pmap because it stops at
2790 vm_pindex_t pdp_pindex
;
2791 vm_pindex_t pd_index
;
2794 pd_index
= ptepindex
- pmap_pd_pindex(0);
2797 pdp_pindex
= NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
+
2798 (pd_index
>> NPML4EPGSHIFT
);
2799 pvp
= pv_get(pv
->pv_pmap
, pdp_pindex
, NULL
);
2804 pd
= pv_pte_lookup(pvp
, pd_index
&
2805 ((1ul << NPDPEPGSHIFT
) - 1));
2806 KKASSERT((*pd
& pmap
->pmap_bits
[PG_V_IDX
]) != 0);
2807 p
= PHYS_TO_VM_PAGE(*pd
& PG_FRAME
);
2808 pmap_inval_bulk(bulk
, (vm_offset_t
)-1, pd
, 0);
2810 KKASSERT(pmap
->pm_flags
& PMAP_FLAG_SIMPLE
);
2811 p
= pv
->pv_m
; /* degenerate test later */
2813 KKASSERT(pv
->pv_m
== p
); /* debugging */
2814 } else if (ptepindex
>= pmap_pt_pindex(0)) {
2816 * Remove a PT page from the PD
2818 vm_pindex_t pd_pindex
;
2819 vm_pindex_t pt_index
;
2822 pt_index
= ptepindex
- pmap_pt_pindex(0);
2825 pd_pindex
= NUPTE_TOTAL
+ NUPT_TOTAL
+
2826 (pt_index
>> NPDPEPGSHIFT
);
2827 pvp
= pv_get(pv
->pv_pmap
, pd_pindex
, NULL
);
2832 pt
= pv_pte_lookup(pvp
, pt_index
& ((1ul << NPDPEPGSHIFT
) - 1));
2834 KASSERT((*pt
& pmap
->pmap_bits
[PG_V_IDX
]) != 0,
2835 ("*pt unexpectedly invalid %016jx "
2836 "gotpvp=%d ptepindex=%ld ptindex=%ld pv=%p pvp=%p",
2837 *pt
, gotpvp
, ptepindex
, pt_index
, pv
, pvp
));
2838 p
= PHYS_TO_VM_PAGE(*pt
& PG_FRAME
);
2840 if ((*pt
& pmap
->pmap_bits
[PG_V_IDX
]) == 0) {
2841 kprintf("*pt unexpectedly invalid %016jx "
2842 "gotpvp=%d ptepindex=%ld ptindex=%ld "
2844 *pt
, gotpvp
, ptepindex
, pt_index
, pv
, pvp
);
2845 tsleep(pt
, 0, "DEAD", 0);
2848 p
= PHYS_TO_VM_PAGE(*pt
& PG_FRAME
);
2851 pmap_inval_bulk(bulk
, (vm_offset_t
)-1, pt
, 0);
2852 KKASSERT(pv
->pv_m
== p
); /* debugging */
2855 * Remove a PTE from the PT page. The PV might exist even if
2856 * the PTE is not managed, in whichcase pv->pv_m should be
2859 * NOTE: Userland pmaps manage the parent PT/PD/PDP page
2860 * table pages but the kernel_pmap does not.
2862 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2863 * pv is a pte_pv so we can safely lock pt_pv.
2865 * NOTE: FICTITIOUS pages may have multiple physical mappings
2866 * so PHYS_TO_VM_PAGE() will not necessarily work for
2869 vm_pindex_t pt_pindex
;
2874 pt_pindex
= ptepindex
>> NPTEPGSHIFT
;
2875 va
= (vm_offset_t
)ptepindex
<< PAGE_SHIFT
;
2877 if (ptepindex
>= NUPTE_USER
) {
2878 ptep
= vtopte(ptepindex
<< PAGE_SHIFT
);
2879 KKASSERT(pvp
== NULL
);
2880 /* pvp remains NULL */
2883 pt_pindex
= NUPTE_TOTAL
+
2884 (ptepindex
>> NPDPEPGSHIFT
);
2885 pvp
= pv_get(pv
->pv_pmap
, pt_pindex
, NULL
);
2889 ptep
= pv_pte_lookup(pvp
, ptepindex
&
2890 ((1ul << NPDPEPGSHIFT
) - 1));
2892 pte
= pmap_inval_bulk(bulk
, va
, ptep
, 0);
2893 if (bulk
== NULL
) /* XXX */
2894 cpu_invlpg((void *)va
); /* XXX */
2897 * Now update the vm_page_t
2899 if ((pte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) &&
2900 (pte
& pmap
->pmap_bits
[PG_V_IDX
])) {
2902 * Valid managed page, adjust (p).
2904 if (pte
& pmap
->pmap_bits
[PG_DEVICE_IDX
]) {
2907 p
= PHYS_TO_VM_PAGE(pte
& PG_FRAME
);
2908 KKASSERT(pv
->pv_m
== p
);
2910 if (pte
& pmap
->pmap_bits
[PG_M_IDX
]) {
2911 if (pmap_track_modified(ptepindex
))
2914 if (pte
& pmap
->pmap_bits
[PG_A_IDX
]) {
2915 vm_page_flag_set(p
, PG_REFERENCED
);
2919 * Unmanaged page, do not try to adjust the vm_page_t.
2920 * pv could be freshly allocated for a pmap_enter(),
2921 * replacing an unmanaged page with a managed one.
2923 * pv->pv_m might reflect the new page and not the
2926 * We could extract p from the physical address and
2927 * adjust it but we explicitly do not for unmanaged
2932 if (pte
& pmap
->pmap_bits
[PG_W_IDX
])
2933 atomic_add_long(&pmap
->pm_stats
.wired_count
, -1);
2934 if (pte
& pmap
->pmap_bits
[PG_G_IDX
])
2935 cpu_invlpg((void *)va
);
2939 * If requested, scrap the underlying pv->pv_m and the underlying
2940 * pv. If this is a page-table-page we must also free the page.
2942 * pvp must be returned locked.
2946 * page table page (PT, PD, PDP, PML4), caller was responsible
2947 * for testing wired_count.
2949 KKASSERT(pv
->pv_m
->wire_count
== 1);
2950 p
= pmap_remove_pv_page(pv
);
2954 vm_page_busy_wait(p
, FALSE
, "pgpun");
2955 vm_page_unwire(p
, 0);
2956 vm_page_flag_clear(p
, PG_MAPPED
| PG_WRITEABLE
);
2958 } else if (destroy
== 2) {
2960 * Normal page, remove from pmap and leave the underlying
2963 pmap_remove_pv_page(pv
);
2965 pv
= NULL
; /* safety */
2969 * If we acquired pvp ourselves then we are responsible for
2970 * recursively deleting it.
2972 if (pvp
&& gotpvp
) {
2974 * Recursively destroy higher-level page tables.
2976 * This is optional. If we do not, they will still
2977 * be destroyed when the process exits.
2979 * NOTE: Do not destroy pv_entry's with extra hold refs,
2980 * a caller may have unlocked it and intends to
2981 * continue to use it.
2983 if (pmap_dynamic_delete
&&
2985 pvp
->pv_m
->wire_count
== 1 &&
2986 (pvp
->pv_hold
& PV_HOLD_MASK
) == 2 &&
2987 pvp
->pv_pindex
!= pmap_pml4_pindex()) {
2988 if (pmap_dynamic_delete
== 2)
2989 kprintf("A %jd %08x\n", pvp
->pv_pindex
, pvp
->pv_hold
);
2990 if (pmap
!= &kernel_pmap
) {
2991 pmap_remove_pv_pte(pvp
, NULL
, bulk
, 1);
2992 pvp
= NULL
; /* safety */
2994 kprintf("Attempt to remove kernel_pmap pindex "
2995 "%jd\n", pvp
->pv_pindex
);
3005 * Remove the vm_page association to a pv. The pv must be locked.
3009 pmap_remove_pv_page(pv_entry_t pv
)
3014 vm_page_spin_lock(m
);
3015 KKASSERT(m
&& m
== pv
->pv_m
);
3017 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
3018 pmap_page_stats_deleting(m
);
3019 if (TAILQ_EMPTY(&m
->md
.pv_list
))
3020 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
3021 vm_page_spin_unlock(m
);
3027 * Grow the number of kernel page table entries, if needed.
3029 * This routine is always called to validate any address space
3030 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
3031 * space below KERNBASE.
3033 * kernel_map must be locked exclusively by the caller.
3036 pmap_growkernel(vm_offset_t kstart
, vm_offset_t kend
)
3039 vm_offset_t ptppaddr
;
3041 pd_entry_t
*pt
, newpt
;
3043 int update_kernel_vm_end
;
3046 * bootstrap kernel_vm_end on first real VM use
3048 if (kernel_vm_end
== 0) {
3049 kernel_vm_end
= VM_MIN_KERNEL_ADDRESS
;
3051 while ((*pmap_pt(&kernel_pmap
, kernel_vm_end
) & kernel_pmap
.pmap_bits
[PG_V_IDX
]) != 0) {
3052 kernel_vm_end
= (kernel_vm_end
+ PAGE_SIZE
* NPTEPG
) &
3053 ~(PAGE_SIZE
* NPTEPG
- 1);
3055 if (kernel_vm_end
- 1 >= kernel_map
.max_offset
) {
3056 kernel_vm_end
= kernel_map
.max_offset
;
3063 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
3064 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
3065 * do not want to force-fill 128G worth of page tables.
3067 if (kstart
< KERNBASE
) {
3068 if (kstart
> kernel_vm_end
)
3069 kstart
= kernel_vm_end
;
3070 KKASSERT(kend
<= KERNBASE
);
3071 update_kernel_vm_end
= 1;
3073 update_kernel_vm_end
= 0;
3076 kstart
= rounddown2(kstart
, PAGE_SIZE
* NPTEPG
);
3077 kend
= roundup2(kend
, PAGE_SIZE
* NPTEPG
);
3079 if (kend
- 1 >= kernel_map
.max_offset
)
3080 kend
= kernel_map
.max_offset
;
3082 while (kstart
< kend
) {
3083 pt
= pmap_pt(&kernel_pmap
, kstart
);
3085 /* We need a new PD entry */
3086 nkpg
= vm_page_alloc(NULL
, mycpu
->gd_rand_incr
++,
3089 VM_ALLOC_INTERRUPT
);
3091 panic("pmap_growkernel: no memory to grow "
3094 paddr
= VM_PAGE_TO_PHYS(nkpg
);
3095 pmap_zero_page(paddr
);
3096 newpd
= (pdp_entry_t
)
3098 kernel_pmap
.pmap_bits
[PG_V_IDX
] |
3099 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
3100 kernel_pmap
.pmap_bits
[PG_A_IDX
] |
3101 kernel_pmap
.pmap_bits
[PG_M_IDX
]);
3102 *pmap_pd(&kernel_pmap
, kstart
) = newpd
;
3103 continue; /* try again */
3105 if ((*pt
& kernel_pmap
.pmap_bits
[PG_V_IDX
]) != 0) {
3106 kstart
= (kstart
+ PAGE_SIZE
* NPTEPG
) &
3107 ~(PAGE_SIZE
* NPTEPG
- 1);
3108 if (kstart
- 1 >= kernel_map
.max_offset
) {
3109 kstart
= kernel_map
.max_offset
;
3118 * This index is bogus, but out of the way
3120 nkpg
= vm_page_alloc(NULL
, mycpu
->gd_rand_incr
++,
3123 VM_ALLOC_INTERRUPT
);
3125 panic("pmap_growkernel: no memory to grow kernel");
3128 ptppaddr
= VM_PAGE_TO_PHYS(nkpg
);
3129 pmap_zero_page(ptppaddr
);
3130 newpt
= (pd_entry_t
)(ptppaddr
|
3131 kernel_pmap
.pmap_bits
[PG_V_IDX
] |
3132 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
3133 kernel_pmap
.pmap_bits
[PG_A_IDX
] |
3134 kernel_pmap
.pmap_bits
[PG_M_IDX
]);
3135 atomic_swap_long(pmap_pt(&kernel_pmap
, kstart
), newpt
);
3137 kstart
= (kstart
+ PAGE_SIZE
* NPTEPG
) &
3138 ~(PAGE_SIZE
* NPTEPG
- 1);
3140 if (kstart
- 1 >= kernel_map
.max_offset
) {
3141 kstart
= kernel_map
.max_offset
;
3147 * Only update kernel_vm_end for areas below KERNBASE.
3149 if (update_kernel_vm_end
&& kernel_vm_end
< kstart
)
3150 kernel_vm_end
= kstart
;
3154 * Add a reference to the specified pmap.
3157 pmap_reference(pmap_t pmap
)
3160 atomic_add_int(&pmap
->pm_count
, 1);
3163 /***************************************************
3164 * page management routines.
3165 ***************************************************/
3168 * Hold a pv without locking it
3171 pv_hold(pv_entry_t pv
)
3173 atomic_add_int(&pv
->pv_hold
, 1);
3177 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
3178 * was successfully locked, FALSE if it wasn't. The caller must dispose of
3181 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3182 * pv list via its page) must be held by the caller in order to stabilize
3186 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL
)
3191 * Critical path shortcut expects pv to already have one ref
3192 * (for the pv->pv_pmap).
3194 if (atomic_cmpset_int(&pv
->pv_hold
, 1, PV_HOLD_LOCKED
| 2)) {
3197 pv
->pv_line
= lineno
;
3203 count
= pv
->pv_hold
;
3205 if ((count
& PV_HOLD_LOCKED
) == 0) {
3206 if (atomic_cmpset_int(&pv
->pv_hold
, count
,
3207 (count
+ 1) | PV_HOLD_LOCKED
)) {
3210 pv
->pv_line
= lineno
;
3215 if (atomic_cmpset_int(&pv
->pv_hold
, count
, count
+ 1))
3223 * Drop a previously held pv_entry which could not be locked, allowing its
3226 * Must not be called with a spinlock held as we might zfree() the pv if it
3227 * is no longer associated with a pmap and this was the last hold count.
3230 pv_drop(pv_entry_t pv
)
3235 count
= pv
->pv_hold
;
3237 KKASSERT((count
& PV_HOLD_MASK
) > 0);
3238 KKASSERT((count
& (PV_HOLD_LOCKED
| PV_HOLD_MASK
)) !=
3239 (PV_HOLD_LOCKED
| 1));
3240 if (atomic_cmpset_int(&pv
->pv_hold
, count
, count
- 1)) {
3241 if ((count
& PV_HOLD_MASK
) == 1) {
3243 if (pmap_enter_debug
> 0) {
3245 kprintf("pv_drop: free pv %p\n", pv
);
3248 KKASSERT(count
== 1);
3249 KKASSERT(pv
->pv_pmap
== NULL
);
3259 * Find or allocate the requested PV entry, returning a locked, held pv.
3261 * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3262 * for the caller and one representing the pmap and vm_page association.
3264 * If (*isnew) is zero, the returned pv will have only one hold count.
3266 * Since both associations can only be adjusted while the pv is locked,
3267 * together they represent just one additional hold.
3271 _pv_alloc(pmap_t pmap
, vm_pindex_t pindex
, int *isnew PMAP_DEBUG_DECL
)
3275 struct mdglobaldata
*md
= mdcpu
;
3280 pnew
= atomic_swap_ptr((void *)&md
->gd_newpv
, NULL
);
3283 pnew
= md
->gd_newpv
; /* might race NULL */
3284 md
->gd_newpv
= NULL
;
3289 pnew
= zalloc(pvzone
);
3291 spin_lock(&pmap
->pm_spin
);
3296 pv
= pmap
->pm_pvhint
;
3299 pv
->pv_pmap
!= pmap
||
3300 pv
->pv_pindex
!= pindex
) {
3301 pv
= pv_entry_rb_tree_RB_LOOKUP(&pmap
->pm_pvroot
,
3308 * We need to block if someone is holding our
3309 * placemarker. As long as we determine the
3310 * placemarker has not been aquired we do not
3311 * need to get it as acquision also requires
3312 * the pmap spin lock.
3314 * However, we can race the wakeup.
3316 pmark
= pmap_placemarker_hash(pmap
, pindex
);
3318 if (((*pmark
^ pindex
) & ~PM_PLACEMARK_WAKEUP
) == 0) {
3319 atomic_set_long(pmark
, PM_PLACEMARK_WAKEUP
);
3320 tsleep_interlock(pmark
, 0);
3321 if (((*pmark
^ pindex
) &
3322 ~PM_PLACEMARK_WAKEUP
) == 0) {
3323 spin_unlock(&pmap
->pm_spin
);
3324 tsleep(pmark
, PINTERLOCKED
, "pvplc", 0);
3325 spin_lock(&pmap
->pm_spin
);
3331 * Setup the new entry
3333 pnew
->pv_pmap
= pmap
;
3334 pnew
->pv_pindex
= pindex
;
3335 pnew
->pv_hold
= PV_HOLD_LOCKED
| 2;
3337 pnew
->pv_func
= func
;
3338 pnew
->pv_line
= lineno
;
3339 if (pnew
->pv_line_lastfree
> 0) {
3340 pnew
->pv_line_lastfree
=
3341 -pnew
->pv_line_lastfree
;
3344 pv
= pv_entry_rb_tree_RB_INSERT(&pmap
->pm_pvroot
, pnew
);
3345 atomic_add_long(&pmap
->pm_stats
.resident_count
, 1);
3346 spin_unlock(&pmap
->pm_spin
);
3349 KKASSERT(pv
== NULL
);
3354 * We already have an entry, cleanup the staged pnew if
3355 * we can get the lock, otherwise block and retry.
3357 if (__predict_true(_pv_hold_try(pv PMAP_DEBUG_COPY
))) {
3358 spin_unlock(&pmap
->pm_spin
);
3360 pnew
= atomic_swap_ptr((void *)&md
->gd_newpv
, pnew
);
3362 zfree(pvzone
, pnew
);
3365 if (md
->gd_newpv
== NULL
)
3366 md
->gd_newpv
= pnew
;
3368 zfree(pvzone
, pnew
);
3371 KKASSERT(pv
->pv_pmap
== pmap
&&
3372 pv
->pv_pindex
== pindex
);
3376 spin_unlock(&pmap
->pm_spin
);
3377 _pv_lock(pv PMAP_DEBUG_COPY
);
3379 spin_lock(&pmap
->pm_spin
);
3385 * Find the requested PV entry, returning a locked+held pv or NULL
3389 _pv_get(pmap_t pmap
, vm_pindex_t pindex
, vm_pindex_t
**pmarkp PMAP_DEBUG_DECL
)
3393 spin_lock(&pmap
->pm_spin
);
3398 pv
= pmap
->pm_pvhint
;
3401 pv
->pv_pmap
!= pmap
||
3402 pv
->pv_pindex
!= pindex
) {
3403 pv
= pv_entry_rb_tree_RB_LOOKUP(&pmap
->pm_pvroot
,
3408 * Block if there is ANY placemarker. If we are to
3409 * return it, we must also aquire the spot, so we
3410 * have to block even if the placemarker is held on
3411 * a different address.
3413 * OPTIMIZATION: If pmarkp is passed as NULL the
3414 * caller is just probing (or looking for a real
3415 * pv_entry), and in this case we only need to check
3416 * to see if the placemarker matches pindex.
3420 pmark
= pmap_placemarker_hash(pmap
, pindex
);
3422 if ((pmarkp
&& *pmark
!= PM_NOPLACEMARK
) ||
3423 ((*pmark
^ pindex
) & ~PM_PLACEMARK_WAKEUP
) == 0) {
3424 atomic_set_long(pmark
, PM_PLACEMARK_WAKEUP
);
3425 tsleep_interlock(pmark
, 0);
3426 if ((pmarkp
&& *pmark
!= PM_NOPLACEMARK
) ||
3427 ((*pmark
^ pindex
) &
3428 ~PM_PLACEMARK_WAKEUP
) == 0) {
3429 spin_unlock(&pmap
->pm_spin
);
3430 tsleep(pmark
, PINTERLOCKED
, "pvpld", 0);
3431 spin_lock(&pmap
->pm_spin
);
3436 if (atomic_swap_long(pmark
, pindex
) !=
3438 panic("_pv_get: pmark race");
3442 spin_unlock(&pmap
->pm_spin
);
3445 if (_pv_hold_try(pv PMAP_DEBUG_COPY
)) {
3446 pv_cache(pv
, pindex
);
3447 spin_unlock(&pmap
->pm_spin
);
3448 KKASSERT(pv
->pv_pmap
== pmap
&&
3449 pv
->pv_pindex
== pindex
);
3452 spin_unlock(&pmap
->pm_spin
);
3453 _pv_lock(pv PMAP_DEBUG_COPY
);
3455 spin_lock(&pmap
->pm_spin
);
3460 * Lookup, hold, and attempt to lock (pmap,pindex).
3462 * If the entry does not exist NULL is returned and *errorp is set to 0
3464 * If the entry exists and could be successfully locked it is returned and
3465 * errorp is set to 0.
3467 * If the entry exists but could NOT be successfully locked it is returned
3468 * held and *errorp is set to 1.
3470 * If the entry is placemarked by someone else NULL is returned and *errorp
3475 pv_get_try(pmap_t pmap
, vm_pindex_t pindex
, vm_pindex_t
**pmarkp
, int *errorp
)
3479 spin_lock_shared(&pmap
->pm_spin
);
3481 pv
= pmap
->pm_pvhint
;
3484 pv
->pv_pmap
!= pmap
||
3485 pv
->pv_pindex
!= pindex
) {
3486 pv
= pv_entry_rb_tree_RB_LOOKUP(&pmap
->pm_pvroot
, pindex
);
3492 pmark
= pmap_placemarker_hash(pmap
, pindex
);
3494 if (((*pmark
^ pindex
) & ~PM_PLACEMARK_WAKEUP
) == 0) {
3496 } else if (pmarkp
&&
3497 atomic_cmpset_long(pmark
, PM_NOPLACEMARK
, pindex
)) {
3501 * Can't set a placemark with a NULL pmarkp, or if
3502 * pmarkp is non-NULL but we failed to set our
3509 spin_unlock_shared(&pmap
->pm_spin
);
3515 * XXX This has problems if the lock is shared, why?
3517 if (pv_hold_try(pv
)) {
3518 pv_cache(pv
, pindex
); /* overwrite ok (shared lock) */
3519 spin_unlock_shared(&pmap
->pm_spin
);
3521 KKASSERT(pv
->pv_pmap
== pmap
&& pv
->pv_pindex
== pindex
);
3522 return(pv
); /* lock succeeded */
3524 spin_unlock_shared(&pmap
->pm_spin
);
3527 return (pv
); /* lock failed */
3531 * Lock a held pv, keeping the hold count
3535 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL
)
3540 count
= pv
->pv_hold
;
3542 if ((count
& PV_HOLD_LOCKED
) == 0) {
3543 if (atomic_cmpset_int(&pv
->pv_hold
, count
,
3544 count
| PV_HOLD_LOCKED
)) {
3547 pv
->pv_line
= lineno
;
3553 tsleep_interlock(pv
, 0);
3554 if (atomic_cmpset_int(&pv
->pv_hold
, count
,
3555 count
| PV_HOLD_WAITING
)) {
3557 if (pmap_enter_debug
> 0) {
3559 kprintf("pv waiting on %s:%d\n",
3560 pv
->pv_func
, pv
->pv_line
);
3563 tsleep(pv
, PINTERLOCKED
, "pvwait", hz
);
3570 * Unlock a held and locked pv, keeping the hold count.
3574 pv_unlock(pv_entry_t pv
)
3579 count
= pv
->pv_hold
;
3581 KKASSERT((count
& (PV_HOLD_LOCKED
| PV_HOLD_MASK
)) >=
3582 (PV_HOLD_LOCKED
| 1));
3583 if (atomic_cmpset_int(&pv
->pv_hold
, count
,
3585 ~(PV_HOLD_LOCKED
| PV_HOLD_WAITING
))) {
3586 if (count
& PV_HOLD_WAITING
)
3594 * Unlock and drop a pv. If the pv is no longer associated with a pmap
3595 * and the hold count drops to zero we will free it.
3597 * Caller should not hold any spin locks. We are protected from hold races
3598 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3599 * lock held. A pv cannot be located otherwise.
3603 pv_put(pv_entry_t pv
)
3606 if (pmap_enter_debug
> 0) {
3608 kprintf("pv_put pv=%p hold=%08x\n", pv
, pv
->pv_hold
);
3613 * Normal put-aways must have a pv_m associated with the pv,
3614 * but allow the case where the pv has been destructed due
3615 * to pmap_dynamic_delete.
3617 KKASSERT(pv
->pv_pmap
== NULL
|| pv
->pv_m
!= NULL
);
3620 * Fast - shortcut most common condition
3622 if (atomic_cmpset_int(&pv
->pv_hold
, PV_HOLD_LOCKED
| 2, 1))
3633 * Remove the pmap association from a pv, require that pv_m already be removed,
3634 * then unlock and drop the pv. Any pte operations must have already been
3635 * completed. This call may result in a last-drop which will physically free
3638 * Removing the pmap association entails an additional drop.
3640 * pv must be exclusively locked on call and will be disposed of on return.
3644 _pv_free(pv_entry_t pv
, pv_entry_t pvp PMAP_DEBUG_DECL
)
3649 pv
->pv_func_lastfree
= func
;
3650 pv
->pv_line_lastfree
= lineno
;
3652 KKASSERT(pv
->pv_m
== NULL
);
3653 KKASSERT((pv
->pv_hold
& (PV_HOLD_LOCKED
|PV_HOLD_MASK
)) >=
3654 (PV_HOLD_LOCKED
|1));
3655 if ((pmap
= pv
->pv_pmap
) != NULL
) {
3656 spin_lock(&pmap
->pm_spin
);
3657 KKASSERT(pv
->pv_pmap
== pmap
);
3658 if (pmap
->pm_pvhint
== pv
)
3659 pmap
->pm_pvhint
= NULL
;
3660 pv_entry_rb_tree_RB_REMOVE(&pmap
->pm_pvroot
, pv
);
3661 atomic_add_long(&pmap
->pm_stats
.resident_count
, -1);
3664 spin_unlock(&pmap
->pm_spin
);
3667 * Try to shortcut three atomic ops, otherwise fall through
3668 * and do it normally. Drop two refs and the lock all in
3672 vm_page_unwire_quick(pvp
->pv_m
);
3673 if (atomic_cmpset_int(&pv
->pv_hold
, PV_HOLD_LOCKED
| 2, 0)) {
3675 if (pmap_enter_debug
> 0) {
3677 kprintf("pv_free: free pv %p\n", pv
);
3683 pv_drop(pv
); /* ref for pv_pmap */
3690 * This routine is very drastic, but can save the system
3698 static int warningdone
=0;
3700 if (pmap_pagedaemon_waken
== 0)
3702 pmap_pagedaemon_waken
= 0;
3703 if (warningdone
< 5) {
3704 kprintf("pmap_collect: collecting pv entries -- "
3705 "suggest increasing PMAP_SHPGPERPROC\n");
3709 for (i
= 0; i
< vm_page_array_size
; i
++) {
3710 m
= &vm_page_array
[i
];
3711 if (m
->wire_count
|| m
->hold_count
)
3713 if (vm_page_busy_try(m
, TRUE
) == 0) {
3714 if (m
->wire_count
== 0 && m
->hold_count
== 0) {
3723 * Scan the pmap for active page table entries and issue a callback.
3724 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3725 * its parent page table.
3727 * pte_pv will be NULL if the page or page table is unmanaged.
3728 * pt_pv will point to the page table page containing the pte for the page.
3730 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3731 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3732 * process pmap's PD and page to the callback function. This can be
3733 * confusing because the pt_pv is really a pd_pv, and the target page
3734 * table page is simply aliased by the pmap and not owned by it.
3736 * It is assumed that the start and end are properly rounded to the page size.
3738 * It is assumed that PD pages and above are managed and thus in the RB tree,
3739 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3741 struct pmap_scan_info
{
3745 vm_pindex_t sva_pd_pindex
;
3746 vm_pindex_t eva_pd_pindex
;
3747 void (*func
)(pmap_t
, struct pmap_scan_info
*,
3748 pv_entry_t
, vm_pindex_t
*, pv_entry_t
,
3750 pt_entry_t
*, void *);
3752 pmap_inval_bulk_t bulk_core
;
3753 pmap_inval_bulk_t
*bulk
;
3758 static int pmap_scan_cmp(pv_entry_t pv
, void *data
);
3759 static int pmap_scan_callback(pv_entry_t pv
, void *data
);
3762 pmap_scan(struct pmap_scan_info
*info
, int smp_inval
)
3764 struct pmap
*pmap
= info
->pmap
;
3765 pv_entry_t pd_pv
; /* A page directory PV */
3766 pv_entry_t pt_pv
; /* A page table PV */
3767 pv_entry_t pte_pv
; /* A page table entry PV */
3768 vm_pindex_t
*pte_placemark
;
3769 vm_pindex_t
*pt_placemark
;
3772 struct pv_entry dummy_pv
;
3777 if (info
->sva
== info
->eva
)
3780 info
->bulk
= &info
->bulk_core
;
3781 pmap_inval_bulk_init(&info
->bulk_core
, pmap
);
3787 * Hold the token for stability; if the pmap is empty we have nothing
3791 if (pmap
->pm_stats
.resident_count
== 0) {
3799 * Special handling for scanning one page, which is a very common
3800 * operation (it is?).
3802 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3804 if (info
->sva
+ PAGE_SIZE
== info
->eva
) {
3805 if (info
->sva
>= VM_MAX_USER_ADDRESS
) {
3807 * Kernel mappings do not track wire counts on
3808 * page table pages and only maintain pd_pv and
3809 * pte_pv levels so pmap_scan() works.
3812 pte_pv
= pv_get(pmap
, pmap_pte_pindex(info
->sva
),
3814 ptep
= vtopte(info
->sva
);
3817 * User pages which are unmanaged will not have a
3818 * pte_pv. User page table pages which are unmanaged
3819 * (shared from elsewhere) will also not have a pt_pv.
3820 * The func() callback will pass both pte_pv and pt_pv
3821 * as NULL in that case.
3823 * We hold pte_placemark across the operation for
3826 * WARNING! We must hold pt_placemark across the
3827 * *ptep test to prevent misintepreting
3828 * a non-zero *ptep as a shared page
3829 * table page. Hold it across the function
3830 * callback as well for SMP safety.
3832 pte_pv
= pv_get(pmap
, pmap_pte_pindex(info
->sva
),
3834 pt_pv
= pv_get(pmap
, pmap_pt_pindex(info
->sva
),
3836 if (pt_pv
== NULL
) {
3837 KKASSERT(pte_pv
== NULL
);
3838 pd_pv
= pv_get(pmap
,
3839 pmap_pd_pindex(info
->sva
),
3842 ptep
= pv_pte_lookup(pd_pv
,
3843 pmap_pt_index(info
->sva
));
3845 info
->func(pmap
, info
,
3851 pv_placemarker_wakeup(pmap
,
3856 pv_placemarker_wakeup(pmap
,
3859 pv_placemarker_wakeup(pmap
, pte_placemark
);
3862 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(info
->sva
));
3866 * NOTE: *ptep can't be ripped out from under us if we hold
3867 * pte_pv (or pte_placemark) locked, but bits can
3873 KKASSERT(pte_pv
== NULL
);
3874 pv_placemarker_wakeup(pmap
, pte_placemark
);
3875 } else if (pte_pv
) {
3876 KASSERT((oldpte
& (pmap
->pmap_bits
[PG_MANAGED_IDX
] |
3877 pmap
->pmap_bits
[PG_V_IDX
])) ==
3878 (pmap
->pmap_bits
[PG_MANAGED_IDX
] |
3879 pmap
->pmap_bits
[PG_V_IDX
]),
3880 ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p",
3881 *ptep
, oldpte
, info
->sva
, pte_pv
));
3882 info
->func(pmap
, info
, pte_pv
, NULL
, pt_pv
, 0,
3883 info
->sva
, ptep
, info
->arg
);
3885 KASSERT((oldpte
& (pmap
->pmap_bits
[PG_MANAGED_IDX
] |
3886 pmap
->pmap_bits
[PG_V_IDX
])) ==
3887 pmap
->pmap_bits
[PG_V_IDX
],
3888 ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL",
3889 *ptep
, oldpte
, info
->sva
));
3890 info
->func(pmap
, info
, NULL
, pte_placemark
, pt_pv
, 0,
3891 info
->sva
, ptep
, info
->arg
);
3896 pmap_inval_bulk_flush(info
->bulk
);
3901 * Nominal scan case, RB_SCAN() for PD pages and iterate from
3904 * WARNING! eva can overflow our standard ((N + mask) >> bits)
3905 * bounds, resulting in a pd_pindex of 0. To solve the
3906 * problem we use an inclusive range.
3908 info
->sva_pd_pindex
= pmap_pd_pindex(info
->sva
);
3909 info
->eva_pd_pindex
= pmap_pd_pindex(info
->eva
- PAGE_SIZE
);
3911 if (info
->sva
>= VM_MAX_USER_ADDRESS
) {
3913 * The kernel does not currently maintain any pv_entry's for
3914 * higher-level page tables.
3916 bzero(&dummy_pv
, sizeof(dummy_pv
));
3917 dummy_pv
.pv_pindex
= info
->sva_pd_pindex
;
3918 spin_lock(&pmap
->pm_spin
);
3919 while (dummy_pv
.pv_pindex
<= info
->eva_pd_pindex
) {
3920 pmap_scan_callback(&dummy_pv
, info
);
3921 ++dummy_pv
.pv_pindex
;
3922 if (dummy_pv
.pv_pindex
< info
->sva_pd_pindex
) /*wrap*/
3925 spin_unlock(&pmap
->pm_spin
);
3928 * User page tables maintain local PML4, PDP, and PD
3929 * pv_entry's at the very least. PT pv's might be
3930 * unmanaged and thus not exist. PTE pv's might be
3931 * unmanaged and thus not exist.
3933 spin_lock(&pmap
->pm_spin
);
3934 pv_entry_rb_tree_RB_SCAN(&pmap
->pm_pvroot
, pmap_scan_cmp
,
3935 pmap_scan_callback
, info
);
3936 spin_unlock(&pmap
->pm_spin
);
3938 pmap_inval_bulk_flush(info
->bulk
);
3942 * WARNING! pmap->pm_spin held
3944 * WARNING! eva can overflow our standard ((N + mask) >> bits)
3945 * bounds, resulting in a pd_pindex of 0. To solve the
3946 * problem we use an inclusive range.
3949 pmap_scan_cmp(pv_entry_t pv
, void *data
)
3951 struct pmap_scan_info
*info
= data
;
3952 if (pv
->pv_pindex
< info
->sva_pd_pindex
)
3954 if (pv
->pv_pindex
> info
->eva_pd_pindex
)
3960 * pmap_scan() by PDs
3962 * WARNING! pmap->pm_spin held
3965 pmap_scan_callback(pv_entry_t pv
, void *data
)
3967 struct pmap_scan_info
*info
= data
;
3968 struct pmap
*pmap
= info
->pmap
;
3969 pv_entry_t pd_pv
; /* A page directory PV */
3970 pv_entry_t pt_pv
; /* A page table PV */
3971 vm_pindex_t
*pt_placemark
;
3976 vm_offset_t va_next
;
3977 vm_pindex_t pd_pindex
;
3987 * Pull the PD pindex from the pv before releasing the spinlock.
3989 * WARNING: pv is faked for kernel pmap scans.
3991 pd_pindex
= pv
->pv_pindex
;
3992 spin_unlock(&pmap
->pm_spin
);
3993 pv
= NULL
; /* invalid after spinlock unlocked */
3996 * Calculate the page range within the PD. SIMPLE pmaps are
3997 * direct-mapped for the entire 2^64 address space. Normal pmaps
3998 * reflect the user and kernel address space which requires
3999 * cannonicalization w/regards to converting pd_pindex's back
4002 sva
= (pd_pindex
- pmap_pd_pindex(0)) << PDPSHIFT
;
4003 if ((pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) == 0 &&
4004 (sva
& PML4_SIGNMASK
)) {
4005 sva
|= PML4_SIGNMASK
;
4007 eva
= sva
+ NBPDP
; /* can overflow */
4008 if (sva
< info
->sva
)
4010 if (eva
< info
->sva
|| eva
> info
->eva
)
4014 * NOTE: kernel mappings do not track page table pages, only
4017 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
4018 * However, for the scan to be efficient we try to
4019 * cache items top-down.
4024 for (; sva
< eva
; sva
= va_next
) {
4027 if (sva
>= VM_MAX_USER_ADDRESS
) {
4036 * PD cache, scan shortcut if it doesn't exist.
4038 if (pd_pv
== NULL
) {
4039 pd_pv
= pv_get(pmap
, pmap_pd_pindex(sva
), NULL
);
4040 } else if (pd_pv
->pv_pmap
!= pmap
||
4041 pd_pv
->pv_pindex
!= pmap_pd_pindex(sva
)) {
4043 pd_pv
= pv_get(pmap
, pmap_pd_pindex(sva
), NULL
);
4045 if (pd_pv
== NULL
) {
4046 va_next
= (sva
+ NBPDP
) & ~PDPMASK
;
4055 * NOTE: The cached pt_pv can be removed from the pmap when
4056 * pmap_dynamic_delete is enabled.
4058 if (pt_pv
&& (pt_pv
->pv_pmap
!= pmap
||
4059 pt_pv
->pv_pindex
!= pmap_pt_pindex(sva
))) {
4063 if (pt_pv
== NULL
) {
4064 pt_pv
= pv_get_try(pmap
, pmap_pt_pindex(sva
),
4065 &pt_placemark
, &error
);
4067 pv_put(pd_pv
); /* lock order */
4074 pv_placemarker_wait(pmap
, pt_placemark
);
4079 /* may have to re-check later if pt_pv is NULL here */
4083 * If pt_pv is NULL we either have an shared page table
4084 * page and must issue a callback specific to that case,
4085 * or there is no page table page.
4087 * Either way we can skip the page table page.
4089 * WARNING! pt_pv can also be NULL due to a pv creation
4090 * race where we find it to be NULL and then
4091 * later see a pte_pv. But its possible the pt_pv
4092 * got created inbetween the two operations, so
4095 if (pt_pv
== NULL
) {
4097 * Possible unmanaged (shared from another pmap)
4100 * WARNING! We must hold pt_placemark across the
4101 * *ptep test to prevent misintepreting
4102 * a non-zero *ptep as a shared page
4103 * table page. Hold it across the function
4104 * callback as well for SMP safety.
4106 ptep
= pv_pte_lookup(pd_pv
, pmap_pt_index(sva
));
4107 if (*ptep
& pmap
->pmap_bits
[PG_V_IDX
]) {
4108 info
->func(pmap
, info
, NULL
, pt_placemark
,
4110 sva
, ptep
, info
->arg
);
4112 pv_placemarker_wakeup(pmap
, pt_placemark
);
4116 * Done, move to next page table page.
4118 va_next
= (sva
+ NBPDR
) & ~PDRMASK
;
4125 * From this point in the loop testing pt_pv for non-NULL
4126 * means we are in UVM, else if it is NULL we are in KVM.
4128 * Limit our scan to either the end of the va represented
4129 * by the current page table page, or to the end of the
4130 * range being removed.
4133 va_next
= (sva
+ NBPDR
) & ~PDRMASK
;
4140 * Scan the page table for pages. Some pages may not be
4141 * managed (might not have a pv_entry).
4143 * There is no page table management for kernel pages so
4144 * pt_pv will be NULL in that case, but otherwise pt_pv
4145 * is non-NULL, locked, and referenced.
4149 * At this point a non-NULL pt_pv means a UVA, and a NULL
4150 * pt_pv means a KVA.
4153 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(sva
));
4157 while (sva
< va_next
) {
4159 vm_pindex_t
*pte_placemark
;
4162 * Yield every 64 pages, stop if requested.
4164 if ((++info
->count
& 63) == 0)
4170 * We can shortcut our scan if *ptep == 0. This is
4171 * an unlocked check.
4181 * Acquire the related pte_pv, if any. If *ptep == 0
4182 * the related pte_pv should not exist, but if *ptep
4183 * is not zero the pte_pv may or may not exist (e.g.
4184 * will not exist for an unmanaged page).
4186 * However a multitude of races are possible here
4187 * so if we cannot lock definite state we clean out
4188 * our cache and break the inner while() loop to
4189 * force a loop up to the top of the for().
4191 * XXX unlock/relock pd_pv, pt_pv, and re-test their
4192 * validity instead of looping up?
4194 pte_pv
= pv_get_try(pmap
, pmap_pte_pindex(sva
),
4195 &pte_placemark
, &error
);
4197 pv_put(pd_pv
); /* lock order */
4200 pv_put(pt_pv
); /* lock order */
4203 if (pte_pv
) { /* block */
4208 pv_placemarker_wait(pmap
,
4211 va_next
= sva
; /* retry */
4216 * Reload *ptep after successfully locking the
4217 * pindex. If *ptep == 0 we had better NOT have a
4224 kprintf("Unexpected non-NULL pte_pv "
4226 "*ptep = %016lx/%016lx\n",
4227 pte_pv
, pt_pv
, *ptep
, oldpte
);
4228 panic("Unexpected non-NULL pte_pv");
4230 pv_placemarker_wakeup(pmap
, pte_placemark
);
4238 * We can't hold pd_pv across the callback (because
4239 * we don't pass it to the callback and the callback
4243 vm_page_wire_quick(pd_pv
->pv_m
);
4248 * Ready for the callback. The locked pte_pv (if any)
4249 * is consumed by the callback. pte_pv will exist if
4250 * the page is managed, and will not exist if it
4253 if (oldpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) {
4258 (oldpte
& pmap
->pmap_bits
[PG_V_IDX
]),
4259 ("badC *ptep %016lx/%016lx sva %016lx "
4261 *ptep
, oldpte
, sva
, pte_pv
));
4263 * We must unlock pd_pv across the callback
4264 * to avoid deadlocks on any recursive
4265 * disposal. Re-check that it still exists
4268 * Call target disposes of pte_pv and may
4269 * destroy but will not dispose of pt_pv.
4271 info
->func(pmap
, info
, pte_pv
, NULL
,
4273 sva
, ptep
, info
->arg
);
4278 * We must unlock pd_pv across the callback
4279 * to avoid deadlocks on any recursive
4280 * disposal. Re-check that it still exists
4283 * Call target disposes of pte_pv or
4284 * pte_placemark and may destroy but will
4285 * not dispose of pt_pv.
4287 KASSERT(pte_pv
== NULL
&&
4288 (oldpte
& pmap
->pmap_bits
[PG_V_IDX
]),
4289 ("badD *ptep %016lx/%016lx sva %016lx "
4290 "pte_pv %p pte_pv->pv_m %p ",
4292 pte_pv
, (pte_pv
? pte_pv
->pv_m
: NULL
)));
4296 info
->func(pmap
, info
,
4299 sva
, ptep
, info
->arg
);
4301 info
->func(pmap
, info
,
4302 NULL
, pte_placemark
,
4304 sva
, ptep
, info
->arg
);
4309 vm_page_unwire_quick(pd_pv
->pv_m
);
4310 if (pd_pv
->pv_pmap
== NULL
) {
4311 va_next
= sva
; /* retry */
4317 * NOTE: The cached pt_pv can be removed from the
4318 * pmap when pmap_dynamic_delete is enabled,
4319 * which will cause ptep to become stale.
4321 * This also means that no pages remain under
4322 * the PT, so we can just break out of the inner
4323 * loop and let the outer loop clean everything
4326 if (pt_pv
&& pt_pv
->pv_pmap
!= pmap
)
4341 if ((++info
->count
& 7) == 0)
4345 * Relock before returning.
4347 spin_lock(&pmap
->pm_spin
);
4352 pmap_remove(struct pmap
*pmap
, vm_offset_t sva
, vm_offset_t eva
)
4354 struct pmap_scan_info info
;
4359 info
.func
= pmap_remove_callback
;
4361 pmap_scan(&info
, 1);
4364 if (eva
- sva
< 1024*1024) {
4366 cpu_invlpg((void *)sva
);
4374 pmap_remove_noinval(struct pmap
*pmap
, vm_offset_t sva
, vm_offset_t eva
)
4376 struct pmap_scan_info info
;
4381 info
.func
= pmap_remove_callback
;
4383 pmap_scan(&info
, 0);
4387 pmap_remove_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
4388 pv_entry_t pte_pv
, vm_pindex_t
*pte_placemark
,
4389 pv_entry_t pt_pv
, int sharept
,
4390 vm_offset_t va
, pt_entry_t
*ptep
, void *arg __unused
)
4398 * This will also drop pt_pv's wire_count. Note that
4399 * terminal pages are not wired based on mmu presence.
4401 * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4403 KKASSERT(pte_pv
->pv_m
!= NULL
);
4404 pmap_remove_pv_pte(pte_pv
, pt_pv
, info
->bulk
, 2);
4405 pte_pv
= NULL
; /* safety */
4408 * Recursively destroy higher-level page tables.
4410 * This is optional. If we do not, they will still
4411 * be destroyed when the process exits.
4413 * NOTE: Do not destroy pv_entry's with extra hold refs,
4414 * a caller may have unlocked it and intends to
4415 * continue to use it.
4417 if (pmap_dynamic_delete
&&
4420 pt_pv
->pv_m
->wire_count
== 1 &&
4421 (pt_pv
->pv_hold
& PV_HOLD_MASK
) == 2 &&
4422 pt_pv
->pv_pindex
!= pmap_pml4_pindex()) {
4423 if (pmap_dynamic_delete
== 2)
4424 kprintf("B %jd %08x\n", pt_pv
->pv_pindex
, pt_pv
->pv_hold
);
4425 pv_hold(pt_pv
); /* extra hold */
4426 pmap_remove_pv_pte(pt_pv
, NULL
, info
->bulk
, 1);
4427 pv_lock(pt_pv
); /* prior extra hold + relock */
4429 } else if (sharept
== 0) {
4431 * Unmanaged pte (pte_placemark is non-NULL)
4433 * pt_pv's wire_count is still bumped by unmanaged pages
4434 * so we must decrement it manually.
4436 * We have to unwire the target page table page.
4438 pte
= pmap_inval_bulk(info
->bulk
, va
, ptep
, 0);
4439 if (pte
& pmap
->pmap_bits
[PG_W_IDX
])
4440 atomic_add_long(&pmap
->pm_stats
.wired_count
, -1);
4441 atomic_add_long(&pmap
->pm_stats
.resident_count
, -1);
4442 if (vm_page_unwire_quick(pt_pv
->pv_m
))
4443 panic("pmap_remove: insufficient wirecount");
4444 pv_placemarker_wakeup(pmap
, pte_placemark
);
4447 * Unmanaged page table (pt, pd, or pdp. Not pte) for
4448 * a shared page table.
4450 * pt_pv is actually the pd_pv for our pmap (not the shared
4453 * We have to unwire the target page table page and we
4454 * have to unwire our page directory page.
4456 * It is unclear how we can invalidate a segment so we
4457 * invalidate -1 which invlidates the tlb.
4459 pte
= pmap_inval_bulk(info
->bulk
, (vm_offset_t
)-1, ptep
, 0);
4460 atomic_add_long(&pmap
->pm_stats
.resident_count
, -1);
4461 KKASSERT((pte
& pmap
->pmap_bits
[PG_DEVICE_IDX
]) == 0);
4462 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte
& PG_FRAME
)))
4463 panic("pmap_remove: shared pgtable1 bad wirecount");
4464 if (vm_page_unwire_quick(pt_pv
->pv_m
))
4465 panic("pmap_remove: shared pgtable2 bad wirecount");
4466 pv_placemarker_wakeup(pmap
, pte_placemark
);
4471 * Removes this physical page from all physical maps in which it resides.
4472 * Reflects back modify bits to the pager.
4474 * This routine may not be called from an interrupt.
4478 pmap_remove_all(vm_page_t m
)
4481 pmap_inval_bulk_t bulk
;
4483 if (!pmap_initialized
/* || (m->flags & PG_FICTITIOUS)*/)
4486 vm_page_spin_lock(m
);
4487 while ((pv
= TAILQ_FIRST(&m
->md
.pv_list
)) != NULL
) {
4488 KKASSERT(pv
->pv_m
== m
);
4489 if (pv_hold_try(pv
)) {
4490 vm_page_spin_unlock(m
);
4492 vm_page_spin_unlock(m
);
4495 vm_page_spin_lock(m
);
4498 KKASSERT(pv
->pv_pmap
&& pv
->pv_m
== m
);
4501 * Holding no spinlocks, pv is locked. Once we scrap
4502 * pv we can no longer use it as a list iterator (but
4503 * we are doing a TAILQ_FIRST() so we are ok).
4505 pmap_inval_bulk_init(&bulk
, pv
->pv_pmap
);
4506 pmap_remove_pv_pte(pv
, NULL
, &bulk
, 2);
4507 pv
= NULL
; /* safety */
4508 pmap_inval_bulk_flush(&bulk
);
4509 vm_page_spin_lock(m
);
4511 KKASSERT((m
->flags
& (PG_MAPPED
|PG_WRITEABLE
)) == 0);
4512 vm_page_spin_unlock(m
);
4516 * Removes the page from a particular pmap
4519 pmap_remove_specific(pmap_t pmap
, vm_page_t m
)
4522 pmap_inval_bulk_t bulk
;
4524 if (!pmap_initialized
)
4528 vm_page_spin_lock(m
);
4529 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
4530 if (pv
->pv_pmap
!= pmap
)
4532 KKASSERT(pv
->pv_m
== m
);
4533 if (pv_hold_try(pv
)) {
4534 vm_page_spin_unlock(m
);
4536 vm_page_spin_unlock(m
);
4541 KKASSERT(pv
->pv_pmap
== pmap
&& pv
->pv_m
== m
);
4544 * Holding no spinlocks, pv is locked. Once gone it can't
4545 * be used as an iterator. In fact, because we couldn't
4546 * necessarily lock it atomically it may have moved within
4547 * the list and ALSO cannot be used as an iterator.
4549 pmap_inval_bulk_init(&bulk
, pv
->pv_pmap
);
4550 pmap_remove_pv_pte(pv
, NULL
, &bulk
, 2);
4551 pv
= NULL
; /* safety */
4552 pmap_inval_bulk_flush(&bulk
);
4555 vm_page_spin_unlock(m
);
4559 * Set the physical protection on the specified range of this map
4560 * as requested. This function is typically only used for debug watchpoints
4563 * This function may not be called from an interrupt if the map is
4564 * not the kernel_pmap.
4566 * NOTE! For shared page table pages we just unmap the page.
4569 pmap_protect(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
, vm_prot_t prot
)
4571 struct pmap_scan_info info
;
4572 /* JG review for NX */
4576 if ((prot
& (VM_PROT_READ
| VM_PROT_EXECUTE
)) == VM_PROT_NONE
) {
4577 pmap_remove(pmap
, sva
, eva
);
4580 if (prot
& VM_PROT_WRITE
)
4585 info
.func
= pmap_protect_callback
;
4587 pmap_scan(&info
, 1);
4592 pmap_protect_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
4593 pv_entry_t pte_pv
, vm_pindex_t
*pte_placemark
,
4594 pv_entry_t pt_pv
, int sharept
,
4595 vm_offset_t va
, pt_entry_t
*ptep
, void *arg __unused
)
4606 KKASSERT(pte_pv
->pv_m
!= NULL
);
4608 if (pbits
& pmap
->pmap_bits
[PG_A_IDX
]) {
4609 if ((pbits
& pmap
->pmap_bits
[PG_DEVICE_IDX
]) == 0) {
4610 m
= PHYS_TO_VM_PAGE(pbits
& PG_FRAME
);
4611 KKASSERT(m
== pte_pv
->pv_m
);
4612 vm_page_flag_set(m
, PG_REFERENCED
);
4614 cbits
&= ~pmap
->pmap_bits
[PG_A_IDX
];
4616 if (pbits
& pmap
->pmap_bits
[PG_M_IDX
]) {
4617 if (pmap_track_modified(pte_pv
->pv_pindex
)) {
4618 if ((pbits
& pmap
->pmap_bits
[PG_DEVICE_IDX
]) == 0) {
4620 m
= PHYS_TO_VM_PAGE(pbits
&
4625 cbits
&= ~pmap
->pmap_bits
[PG_M_IDX
];
4628 } else if (sharept
) {
4630 * Unmanaged page table, pt_pv is actually the pd_pv
4631 * for our pmap (not the object's shared pmap).
4633 * When asked to protect something in a shared page table
4634 * page we just unmap the page table page. We have to
4635 * invalidate the tlb in this situation.
4637 * XXX Warning, shared page tables will not be used for
4638 * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4639 * so PHYS_TO_VM_PAGE() should be safe here.
4641 pte
= pmap_inval_smp(pmap
, (vm_offset_t
)-1, 1, ptep
, 0);
4642 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte
& PG_FRAME
)))
4643 panic("pmap_protect: pgtable1 pg bad wirecount");
4644 if (vm_page_unwire_quick(pt_pv
->pv_m
))
4645 panic("pmap_protect: pgtable2 pg bad wirecount");
4648 /* else unmanaged page, adjust bits, no wire changes */
4651 cbits
&= ~pmap
->pmap_bits
[PG_RW_IDX
];
4653 if (pmap_enter_debug
> 0) {
4655 kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4656 "pt_pv=%p cbits=%08lx\n",
4662 if (pbits
!= cbits
) {
4665 xva
= (sharept
) ? (vm_offset_t
)-1 : va
;
4666 if (!pmap_inval_smp_cmpset(pmap
, xva
,
4667 ptep
, pbits
, cbits
)) {
4675 pv_placemarker_wakeup(pmap
, pte_placemark
);
4679 * Insert the vm_page (m) at the virtual address (va), replacing any prior
4680 * mapping at that address. Set protection and wiring as requested.
4682 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4683 * possible. If it is we enter the page into the appropriate shared pmap
4684 * hanging off the related VM object instead of the passed pmap, then we
4685 * share the page table page from the VM object's pmap into the current pmap.
4687 * NOTE: This routine MUST insert the page into the pmap now, it cannot
4690 * NOTE: If (m) is PG_UNMANAGED it may also be a temporary fake vm_page_t.
4694 pmap_enter(pmap_t pmap
, vm_offset_t va
, vm_page_t m
, vm_prot_t prot
,
4695 boolean_t wired
, vm_map_entry_t entry
)
4697 pv_entry_t pt_pv
; /* page table */
4698 pv_entry_t pte_pv
; /* page table entry */
4699 vm_pindex_t
*pte_placemark
;
4702 pt_entry_t origpte
, newpte
;
4707 va
= trunc_page(va
);
4708 #ifdef PMAP_DIAGNOSTIC
4710 panic("pmap_enter: toobig");
4711 if ((va
>= UPT_MIN_ADDRESS
) && (va
< UPT_MAX_ADDRESS
))
4712 panic("pmap_enter: invalid to pmap_enter page table "
4713 "pages (va: 0x%lx)", va
);
4715 if (va
< UPT_MAX_ADDRESS
&& pmap
== &kernel_pmap
) {
4716 kprintf("Warning: pmap_enter called on UVA with "
4719 db_print_backtrace();
4722 if (va
>= UPT_MAX_ADDRESS
&& pmap
!= &kernel_pmap
) {
4723 kprintf("Warning: pmap_enter called on KVA without"
4726 db_print_backtrace();
4731 * Get locked PV entries for our new page table entry (pte_pv or
4732 * pte_placemark) and for its parent page table (pt_pv). We need
4733 * the parent so we can resolve the location of the ptep.
4735 * Only hardware MMU actions can modify the ptep out from
4738 * if (m) is fictitious or unmanaged we do not create a managing
4739 * pte_pv for it. Any pre-existing page's management state must
4740 * match (avoiding code complexity).
4742 * If the pmap is still being initialized we assume existing
4745 * Kernel mapppings do not track page table pages (i.e. pt_pv).
4747 * WARNING! If replacing a managed mapping with an unmanaged mapping
4748 * pte_pv will wind up being non-NULL and must be handled
4751 if (pmap_initialized
== FALSE
) {
4754 pte_placemark
= NULL
;
4757 } else if (m
->flags
& (/*PG_FICTITIOUS |*/ PG_UNMANAGED
)) { /* XXX */
4758 pte_pv
= pv_get(pmap
, pmap_pte_pindex(va
), &pte_placemark
);
4759 KKASSERT(pte_pv
== NULL
);
4760 if (va
>= VM_MAX_USER_ADDRESS
) {
4764 pt_pv
= pmap_allocpte_seg(pmap
, pmap_pt_pindex(va
),
4766 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
4770 KASSERT(origpte
== 0 ||
4771 (origpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0,
4772 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte
, va
));
4774 if (va
>= VM_MAX_USER_ADDRESS
) {
4776 * Kernel map, pv_entry-tracked.
4779 pte_pv
= pmap_allocpte(pmap
, pmap_pte_pindex(va
), NULL
);
4785 pte_pv
= pmap_allocpte_seg(pmap
, pmap_pte_pindex(va
),
4787 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
4789 pte_placemark
= NULL
; /* safety */
4792 KASSERT(origpte
== 0 ||
4793 (origpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]),
4794 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte
, va
));
4797 pa
= VM_PAGE_TO_PHYS(m
);
4798 opa
= origpte
& PG_FRAME
;
4801 * Calculate the new PTE. Note that pte_pv alone does not mean
4802 * the new pte_pv is managed, it could exist because the old pte
4803 * was managed even if the new one is not.
4805 newpte
= (pt_entry_t
)(pa
| pte_prot(pmap
, prot
) |
4806 pmap
->pmap_bits
[PG_V_IDX
] | pmap
->pmap_bits
[PG_A_IDX
]);
4808 newpte
|= pmap
->pmap_bits
[PG_W_IDX
];
4809 if (va
< VM_MAX_USER_ADDRESS
)
4810 newpte
|= pmap
->pmap_bits
[PG_U_IDX
];
4811 if (pte_pv
&& (m
->flags
& (/*PG_FICTITIOUS |*/ PG_UNMANAGED
)) == 0)
4812 newpte
|= pmap
->pmap_bits
[PG_MANAGED_IDX
];
4813 // if (pmap == &kernel_pmap)
4814 // newpte |= pgeflag;
4815 newpte
|= pmap
->pmap_cache_bits
[m
->pat_mode
];
4816 if (m
->flags
& PG_FICTITIOUS
)
4817 newpte
|= pmap
->pmap_bits
[PG_DEVICE_IDX
];
4820 * It is possible for multiple faults to occur in threaded
4821 * environments, the existing pte might be correct.
4823 if (((origpte
^ newpte
) &
4824 ~(pt_entry_t
)(pmap
->pmap_bits
[PG_M_IDX
] |
4825 pmap
->pmap_bits
[PG_A_IDX
])) == 0) {
4830 * Ok, either the address changed or the protection or wiring
4833 * Clear the current entry, interlocking the removal. For managed
4834 * pte's this will also flush the modified state to the vm_page.
4835 * Atomic ops are mandatory in order to ensure that PG_M events are
4836 * not lost during any transition.
4838 * WARNING: The caller has busied the new page but not the original
4839 * vm_page which we are trying to replace. Because we hold
4840 * the pte_pv lock, but have not busied the page, PG bits
4841 * can be cleared out from under us.
4844 if (origpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) {
4846 * Old page was managed. Expect pte_pv to exist.
4847 * (it might also exist if the old page was unmanaged).
4849 * NOTE: pt_pv won't exist for a kernel page
4850 * (managed or otherwise).
4852 * NOTE: We may be reusing the pte_pv so we do not
4853 * destroy it in pmap_remove_pv_pte().
4855 KKASSERT(pte_pv
&& pte_pv
->pv_m
);
4856 if (prot
& VM_PROT_NOSYNC
) {
4857 pmap_remove_pv_pte(pte_pv
, pt_pv
, NULL
, 0);
4859 pmap_inval_bulk_t bulk
;
4861 pmap_inval_bulk_init(&bulk
, pmap
);
4862 pmap_remove_pv_pte(pte_pv
, pt_pv
, &bulk
, 0);
4863 pmap_inval_bulk_flush(&bulk
);
4865 pmap_remove_pv_page(pte_pv
);
4866 /* will either set pte_pv->pv_m or pv_free() later */
4869 * Old page was not managed. If we have a pte_pv
4870 * it better not have a pv_m assigned to it. If the
4871 * new page is managed the pte_pv will be destroyed
4872 * near the end (we need its interlock).
4874 * NOTE: We leave the wire count on the PT page
4875 * intact for the followup enter, but adjust
4876 * the wired-pages count on the pmap.
4878 KKASSERT(pte_pv
== NULL
);
4879 if (prot
& VM_PROT_NOSYNC
) {
4881 * NOSYNC (no mmu sync) requested.
4883 (void)pte_load_clear(ptep
);
4884 cpu_invlpg((void *)va
);
4889 pmap_inval_smp(pmap
, va
, 1, ptep
, 0);
4893 * We must adjust pm_stats manually for unmanaged
4897 atomic_add_long(&pmap
->pm_stats
.
4898 resident_count
, -1);
4900 if (origpte
& pmap
->pmap_bits
[PG_W_IDX
]) {
4901 atomic_add_long(&pmap
->pm_stats
.
4905 KKASSERT(*ptep
== 0);
4909 if (pmap_enter_debug
> 0) {
4911 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
4912 " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
4914 origpte
, newpte
, ptep
,
4915 pte_pv
, pt_pv
, opa
, prot
);
4919 if ((newpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0) {
4921 * Entering an unmanaged page. We must wire the pt_pv unless
4922 * we retained the wiring from an unmanaged page we had
4923 * removed (if we retained it via pte_pv that will go away
4926 if (pt_pv
&& (opa
== 0 ||
4927 (origpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]))) {
4928 vm_page_wire_quick(pt_pv
->pv_m
);
4931 atomic_add_long(&pmap
->pm_stats
.wired_count
, 1);
4934 * Unmanaged pages need manual resident_count tracking.
4937 atomic_add_long(&pt_pv
->pv_pmap
->pm_stats
.
4940 if (newpte
& pmap
->pmap_bits
[PG_RW_IDX
])
4941 vm_page_flag_set(m
, PG_WRITEABLE
);
4944 * Entering a managed page. Our pte_pv takes care of the
4945 * PT wiring, so if we had removed an unmanaged page before
4948 * We have to take care of the pmap wired count ourselves.
4950 * Enter on the PV list if part of our managed memory.
4952 KKASSERT(pte_pv
&& (pte_pv
->pv_m
== NULL
|| pte_pv
->pv_m
== m
));
4953 vm_page_spin_lock(m
);
4955 pmap_page_stats_adding(m
);
4956 TAILQ_INSERT_TAIL(&m
->md
.pv_list
, pte_pv
, pv_list
);
4957 vm_page_flag_set(m
, PG_MAPPED
);
4958 if (newpte
& pmap
->pmap_bits
[PG_RW_IDX
])
4959 vm_page_flag_set(m
, PG_WRITEABLE
);
4960 vm_page_spin_unlock(m
);
4963 (origpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0) {
4964 vm_page_unwire_quick(pt_pv
->pv_m
);
4968 * Adjust pmap wired pages count for new entry.
4971 atomic_add_long(&pte_pv
->pv_pmap
->pm_stats
.
4977 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
4979 * User VMAs do not because those will be zero->non-zero, so no
4980 * stale entries to worry about at this point.
4982 * For KVM there appear to still be issues. Theoretically we
4983 * should be able to scrap the interlocks entirely but we
4986 if ((prot
& VM_PROT_NOSYNC
) == 0 && pt_pv
== NULL
) {
4987 pmap_inval_smp(pmap
, va
, 1, ptep
, newpte
);
4989 origpte
= atomic_swap_long(ptep
, newpte
);
4990 if (origpte
& pmap
->pmap_bits
[PG_M_IDX
]) {
4991 kprintf("pmap [M] race @ %016jx\n", va
);
4992 atomic_set_long(ptep
, pmap
->pmap_bits
[PG_M_IDX
]);
4995 cpu_invlpg((void *)va
);
5002 KKASSERT((newpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0 ||
5003 (m
->flags
& PG_MAPPED
));
5006 * Cleanup the pv entry, allowing other accessors. If the new page
5007 * is not managed but we have a pte_pv (which was locking our
5008 * operation), we can free it now. pte_pv->pv_m should be NULL.
5010 if (pte_pv
&& (newpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0) {
5011 pv_free(pte_pv
, pt_pv
);
5012 } else if (pte_pv
) {
5014 } else if (pte_placemark
) {
5015 pv_placemarker_wakeup(pmap
, pte_placemark
);
5022 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
5023 * This code also assumes that the pmap has no pre-existing entry for this
5026 * This code currently may only be used on user pmaps, not kernel_pmap.
5029 pmap_enter_quick(pmap_t pmap
, vm_offset_t va
, vm_page_t m
)
5031 pmap_enter(pmap
, va
, m
, VM_PROT_READ
, FALSE
, NULL
);
5035 * Make a temporary mapping for a physical address. This is only intended
5036 * to be used for panic dumps.
5038 * The caller is responsible for calling smp_invltlb().
5041 pmap_kenter_temporary(vm_paddr_t pa
, long i
)
5043 pmap_kenter_quick((vm_offset_t
)crashdumpmap
+ (i
* PAGE_SIZE
), pa
);
5044 return ((void *)crashdumpmap
);
5047 #define MAX_INIT_PT (96)
5050 * This routine preloads the ptes for a given object into the specified pmap.
5051 * This eliminates the blast of soft faults on process startup and
5052 * immediately after an mmap.
5054 static int pmap_object_init_pt_callback(vm_page_t p
, void *data
);
5057 pmap_object_init_pt(pmap_t pmap
, vm_offset_t addr
, vm_prot_t prot
,
5058 vm_object_t object
, vm_pindex_t pindex
,
5059 vm_size_t size
, int limit
)
5061 struct rb_vm_page_scan_info info
;
5066 * We can't preinit if read access isn't set or there is no pmap
5069 if ((prot
& VM_PROT_READ
) == 0 || pmap
== NULL
|| object
== NULL
)
5073 * We can't preinit if the pmap is not the current pmap
5075 lp
= curthread
->td_lwp
;
5076 if (lp
== NULL
|| pmap
!= vmspace_pmap(lp
->lwp_vmspace
))
5080 * Misc additional checks
5082 psize
= x86_64_btop(size
);
5084 if ((object
->type
!= OBJT_VNODE
) ||
5085 ((limit
& MAP_PREFAULT_PARTIAL
) && (psize
> MAX_INIT_PT
) &&
5086 (object
->resident_page_count
> MAX_INIT_PT
))) {
5090 if (pindex
+ psize
> object
->size
) {
5091 if (object
->size
< pindex
)
5093 psize
= object
->size
- pindex
;
5100 * If everything is segment-aligned do not pre-init here. Instead
5101 * allow the normal vm_fault path to pass a segment hint to
5102 * pmap_enter() which will then use an object-referenced shared
5105 if ((addr
& SEG_MASK
) == 0 &&
5106 (ctob(psize
) & SEG_MASK
) == 0 &&
5107 (ctob(pindex
) & SEG_MASK
) == 0) {
5112 * Use a red-black scan to traverse the requested range and load
5113 * any valid pages found into the pmap.
5115 * We cannot safely scan the object's memq without holding the
5118 info
.start_pindex
= pindex
;
5119 info
.end_pindex
= pindex
+ psize
- 1;
5125 vm_object_hold_shared(object
);
5126 vm_page_rb_tree_RB_SCAN(&object
->rb_memq
, rb_vm_page_scancmp
,
5127 pmap_object_init_pt_callback
, &info
);
5128 vm_object_drop(object
);
5133 pmap_object_init_pt_callback(vm_page_t p
, void *data
)
5135 struct rb_vm_page_scan_info
*info
= data
;
5136 vm_pindex_t rel_index
;
5139 * don't allow an madvise to blow away our really
5140 * free pages allocating pv entries.
5142 if ((info
->limit
& MAP_PREFAULT_MADVISE
) &&
5143 vmstats
.v_free_count
< vmstats
.v_free_reserved
) {
5148 * Ignore list markers and ignore pages we cannot instantly
5149 * busy (while holding the object token).
5151 if (p
->flags
& PG_MARKER
)
5153 if (vm_page_busy_try(p
, TRUE
))
5155 if (((p
->valid
& VM_PAGE_BITS_ALL
) == VM_PAGE_BITS_ALL
) &&
5156 (p
->flags
& PG_FICTITIOUS
) == 0) {
5157 if ((p
->queue
- p
->pc
) == PQ_CACHE
)
5158 vm_page_deactivate(p
);
5159 rel_index
= p
->pindex
- info
->start_pindex
;
5160 pmap_enter_quick(info
->pmap
,
5161 info
->addr
+ x86_64_ptob(rel_index
), p
);
5169 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
5172 * Returns FALSE if it would be non-trivial or if a pte is already loaded
5175 * XXX This is safe only because page table pages are not freed.
5178 pmap_prefault_ok(pmap_t pmap
, vm_offset_t addr
)
5182 /*spin_lock(&pmap->pm_spin);*/
5183 if ((pte
= pmap_pte(pmap
, addr
)) != NULL
) {
5184 if (*pte
& pmap
->pmap_bits
[PG_V_IDX
]) {
5185 /*spin_unlock(&pmap->pm_spin);*/
5189 /*spin_unlock(&pmap->pm_spin);*/
5194 * Change the wiring attribute for a pmap/va pair. The mapping must already
5195 * exist in the pmap. The mapping may or may not be managed. The wiring in
5196 * the page is not changed, the page is returned so the caller can adjust
5197 * its wiring (the page is not locked in any way).
5199 * Wiring is not a hardware characteristic so there is no need to invalidate
5200 * TLB. However, in an SMP environment we must use a locked bus cycle to
5201 * update the pte (if we are not using the pmap_inval_*() API that is)...
5202 * it's ok to do this for simple wiring changes.
5205 pmap_unwire(pmap_t pmap
, vm_offset_t va
)
5216 * Assume elements in the kernel pmap are stable
5218 if (pmap
== &kernel_pmap
) {
5219 if (pmap_pt(pmap
, va
) == 0)
5221 ptep
= pmap_pte_quick(pmap
, va
);
5222 if (pmap_pte_v(pmap
, ptep
)) {
5223 if (pmap_pte_w(pmap
, ptep
))
5224 atomic_add_long(&pmap
->pm_stats
.wired_count
,-1);
5225 atomic_clear_long(ptep
, pmap
->pmap_bits
[PG_W_IDX
]);
5226 pa
= *ptep
& PG_FRAME
;
5227 m
= PHYS_TO_VM_PAGE(pa
);
5233 * We can only [un]wire pmap-local pages (we cannot wire
5236 pt_pv
= pv_get(pmap
, pmap_pt_pindex(va
), NULL
);
5240 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
5241 if ((*ptep
& pmap
->pmap_bits
[PG_V_IDX
]) == 0) {
5246 if (pmap_pte_w(pmap
, ptep
)) {
5247 atomic_add_long(&pt_pv
->pv_pmap
->pm_stats
.wired_count
,
5250 /* XXX else return NULL so caller doesn't unwire m ? */
5252 atomic_clear_long(ptep
, pmap
->pmap_bits
[PG_W_IDX
]);
5254 pa
= *ptep
& PG_FRAME
;
5255 m
= PHYS_TO_VM_PAGE(pa
); /* held by wired count */
5262 * Copy the range specified by src_addr/len from the source map to
5263 * the range dst_addr/len in the destination map.
5265 * This routine is only advisory and need not do anything.
5268 pmap_copy(pmap_t dst_pmap
, pmap_t src_pmap
, vm_offset_t dst_addr
,
5269 vm_size_t len
, vm_offset_t src_addr
)
5276 * Zero the specified physical page.
5278 * This function may be called from an interrupt and no locking is
5282 pmap_zero_page(vm_paddr_t phys
)
5284 vm_offset_t va
= PHYS_TO_DMAP(phys
);
5286 pagezero((void *)va
);
5292 * Zero part of a physical page by mapping it into memory and clearing
5293 * its contents with bzero.
5295 * off and size may not cover an area beyond a single hardware page.
5298 pmap_zero_page_area(vm_paddr_t phys
, int off
, int size
)
5300 vm_offset_t virt
= PHYS_TO_DMAP(phys
);
5302 bzero((char *)virt
+ off
, size
);
5308 * Copy the physical page from the source PA to the target PA.
5309 * This function may be called from an interrupt. No locking
5313 pmap_copy_page(vm_paddr_t src
, vm_paddr_t dst
)
5315 vm_offset_t src_virt
, dst_virt
;
5317 src_virt
= PHYS_TO_DMAP(src
);
5318 dst_virt
= PHYS_TO_DMAP(dst
);
5319 bcopy((void *)src_virt
, (void *)dst_virt
, PAGE_SIZE
);
5323 * pmap_copy_page_frag:
5325 * Copy the physical page from the source PA to the target PA.
5326 * This function may be called from an interrupt. No locking
5330 pmap_copy_page_frag(vm_paddr_t src
, vm_paddr_t dst
, size_t bytes
)
5332 vm_offset_t src_virt
, dst_virt
;
5334 src_virt
= PHYS_TO_DMAP(src
);
5335 dst_virt
= PHYS_TO_DMAP(dst
);
5337 bcopy((char *)src_virt
+ (src
& PAGE_MASK
),
5338 (char *)dst_virt
+ (dst
& PAGE_MASK
),
5343 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
5344 * this page. This count may be changed upwards or downwards in the future;
5345 * it is only necessary that true be returned for a small subset of pmaps
5346 * for proper page aging.
5349 pmap_page_exists_quick(pmap_t pmap
, vm_page_t m
)
5354 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
5357 vm_page_spin_lock(m
);
5358 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
5359 if (pv
->pv_pmap
== pmap
) {
5360 vm_page_spin_unlock(m
);
5367 vm_page_spin_unlock(m
);
5372 * Remove all pages from specified address space this aids process exit
5373 * speeds. Also, this code may be special cased for the current process
5377 pmap_remove_pages(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
)
5379 pmap_remove_noinval(pmap
, sva
, eva
);
5384 * pmap_testbit tests bits in pte's note that the testbit/clearbit
5385 * routines are inline, and a lot of things compile-time evaluate.
5389 pmap_testbit(vm_page_t m
, int bit
)
5395 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
5398 if (TAILQ_FIRST(&m
->md
.pv_list
) == NULL
)
5400 vm_page_spin_lock(m
);
5401 if (TAILQ_FIRST(&m
->md
.pv_list
) == NULL
) {
5402 vm_page_spin_unlock(m
);
5406 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
5408 #if defined(PMAP_DIAGNOSTIC)
5409 if (pv
->pv_pmap
== NULL
) {
5410 kprintf("Null pmap (tb) at pindex: %"PRIu64
"\n",
5418 * If the bit being tested is the modified bit, then
5419 * mark clean_map and ptes as never
5422 * WARNING! Because we do not lock the pv, *pte can be in a
5423 * state of flux. Despite this the value of *pte
5424 * will still be related to the vm_page in some way
5425 * because the pv cannot be destroyed as long as we
5426 * hold the vm_page spin lock.
5428 if (bit
== PG_A_IDX
|| bit
== PG_M_IDX
) {
5429 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5430 if (!pmap_track_modified(pv
->pv_pindex
))
5434 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_pindex
<< PAGE_SHIFT
);
5435 if (*pte
& pmap
->pmap_bits
[bit
]) {
5436 vm_page_spin_unlock(m
);
5440 vm_page_spin_unlock(m
);
5445 * This routine is used to modify bits in ptes. Only one bit should be
5446 * specified. PG_RW requires special handling.
5448 * Caller must NOT hold any spin locks
5452 pmap_clearbit(vm_page_t m
, int bit_index
)
5459 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
)) {
5460 if (bit_index
== PG_RW_IDX
)
5461 vm_page_flag_clear(m
, PG_WRITEABLE
);
5468 * Loop over all current mappings setting/clearing as appropos If
5469 * setting RO do we need to clear the VAC?
5471 * NOTE: When clearing PG_M we could also (not implemented) drop
5472 * through to the PG_RW code and clear PG_RW too, forcing
5473 * a fault on write to redetect PG_M for virtual kernels, but
5474 * it isn't necessary since virtual kernels invalidate the
5475 * pte when they clear the VPTE_M bit in their virtual page
5478 * NOTE: Does not re-dirty the page when clearing only PG_M.
5480 * NOTE: Because we do not lock the pv, *pte can be in a state of
5481 * flux. Despite this the value of *pte is still somewhat
5482 * related while we hold the vm_page spin lock.
5484 * *pte can be zero due to this race. Since we are clearing
5485 * bits we basically do no harm when this race occurs.
5487 if (bit_index
!= PG_RW_IDX
) {
5488 vm_page_spin_lock(m
);
5489 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
5490 #if defined(PMAP_DIAGNOSTIC)
5491 if (pv
->pv_pmap
== NULL
) {
5492 kprintf("Null pmap (cb) at pindex: %"PRIu64
"\n",
5498 pte
= pmap_pte_quick(pv
->pv_pmap
,
5499 pv
->pv_pindex
<< PAGE_SHIFT
);
5501 if (pbits
& pmap
->pmap_bits
[bit_index
])
5502 atomic_clear_long(pte
, pmap
->pmap_bits
[bit_index
]);
5504 vm_page_spin_unlock(m
);
5509 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
5513 vm_page_spin_lock(m
);
5514 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
5516 * don't write protect pager mappings
5518 if (!pmap_track_modified(pv
->pv_pindex
))
5521 #if defined(PMAP_DIAGNOSTIC)
5522 if (pv
->pv_pmap
== NULL
) {
5523 kprintf("Null pmap (cb) at pindex: %"PRIu64
"\n",
5531 * Skip pages which do not have PG_RW set.
5533 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_pindex
<< PAGE_SHIFT
);
5534 if ((*pte
& pmap
->pmap_bits
[PG_RW_IDX
]) == 0)
5538 * We must lock the PV to be able to safely test the pte.
5540 if (pv_hold_try(pv
)) {
5541 vm_page_spin_unlock(m
);
5543 vm_page_spin_unlock(m
);
5544 pv_lock(pv
); /* held, now do a blocking lock */
5550 * Reload pte after acquiring pv.
5552 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_pindex
<< PAGE_SHIFT
);
5554 if ((*pte
& pmap
->pmap_bits
[PG_RW_IDX
]) == 0) {
5560 KKASSERT(pv
->pv_pmap
== pmap
&& pv
->pv_m
== m
);
5566 nbits
= pbits
& ~(pmap
->pmap_bits
[PG_RW_IDX
] |
5567 pmap
->pmap_bits
[PG_M_IDX
]);
5568 if (pmap_inval_smp_cmpset(pmap
,
5569 ((vm_offset_t
)pv
->pv_pindex
<< PAGE_SHIFT
),
5570 pte
, pbits
, nbits
)) {
5577 * If PG_M was found to be set while we were clearing PG_RW
5578 * we also clear PG_M (done above) and mark the page dirty.
5579 * Callers expect this behavior.
5581 * we lost pv so it cannot be used as an iterator. In fact,
5582 * because we couldn't necessarily lock it atomically it may
5583 * have moved within the list and ALSO cannot be used as an
5586 vm_page_spin_lock(m
);
5587 if (pbits
& pmap
->pmap_bits
[PG_M_IDX
])
5589 vm_page_spin_unlock(m
);
5593 if (bit_index
== PG_RW_IDX
)
5594 vm_page_flag_clear(m
, PG_WRITEABLE
);
5595 vm_page_spin_unlock(m
);
5599 * Lower the permission for all mappings to a given page.
5601 * Page must be busied by caller. Because page is busied by caller this
5602 * should not be able to race a pmap_enter().
5605 pmap_page_protect(vm_page_t m
, vm_prot_t prot
)
5607 /* JG NX support? */
5608 if ((prot
& VM_PROT_WRITE
) == 0) {
5609 if (prot
& (VM_PROT_READ
| VM_PROT_EXECUTE
)) {
5611 * NOTE: pmap_clearbit(.. PG_RW) also clears
5612 * the PG_WRITEABLE flag in (m).
5614 pmap_clearbit(m
, PG_RW_IDX
);
5622 pmap_phys_address(vm_pindex_t ppn
)
5624 return (x86_64_ptob(ppn
));
5628 * Return a count of reference bits for a page, clearing those bits.
5629 * It is not necessary for every reference bit to be cleared, but it
5630 * is necessary that 0 only be returned when there are truly no
5631 * reference bits set.
5633 * XXX: The exact number of bits to check and clear is a matter that
5634 * should be tested and standardized at some point in the future for
5635 * optimal aging of shared pages.
5637 * This routine may not block.
5640 pmap_ts_referenced(vm_page_t m
)
5647 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
5650 vm_page_spin_lock(m
);
5651 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
5652 if (!pmap_track_modified(pv
->pv_pindex
))
5655 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_pindex
<< PAGE_SHIFT
);
5656 if (pte
&& (*pte
& pmap
->pmap_bits
[PG_A_IDX
])) {
5657 atomic_clear_long(pte
, pmap
->pmap_bits
[PG_A_IDX
]);
5663 vm_page_spin_unlock(m
);
5670 * Return whether or not the specified physical page was modified
5671 * in any physical maps.
5674 pmap_is_modified(vm_page_t m
)
5678 res
= pmap_testbit(m
, PG_M_IDX
);
5683 * Clear the modify bits on the specified physical page.
5686 pmap_clear_modify(vm_page_t m
)
5688 pmap_clearbit(m
, PG_M_IDX
);
5692 * pmap_clear_reference:
5694 * Clear the reference bit on the specified physical page.
5697 pmap_clear_reference(vm_page_t m
)
5699 pmap_clearbit(m
, PG_A_IDX
);
5703 * Miscellaneous support routines follow
5708 i386_protection_init(void)
5714 * NX supported? (boot time loader.conf override only)
5716 TUNABLE_INT_FETCH("machdep.pmap_nx_enable", &pmap_nx_enable
);
5717 if (pmap_nx_enable
== 0 || (amd_feature
& AMDID_NX
) == 0)
5718 pmap_bits_default
[PG_NX_IDX
] = 0;
5721 * 0 is basically read-only access, but also set the NX (no-execute)
5722 * bit when VM_PROT_EXECUTE is not specified.
5724 kp
= protection_codes
;
5725 for (prot
= 0; prot
< PROTECTION_CODES_SIZE
; prot
++) {
5727 case VM_PROT_NONE
| VM_PROT_NONE
| VM_PROT_NONE
:
5729 * This case handled elsewhere
5733 case VM_PROT_READ
| VM_PROT_NONE
| VM_PROT_NONE
:
5737 *kp
++ = pmap_bits_default
[PG_NX_IDX
];
5739 case VM_PROT_READ
| VM_PROT_NONE
| VM_PROT_EXECUTE
:
5740 case VM_PROT_NONE
| VM_PROT_NONE
| VM_PROT_EXECUTE
:
5742 * Execute requires read access
5746 case VM_PROT_NONE
| VM_PROT_WRITE
| VM_PROT_NONE
:
5747 case VM_PROT_READ
| VM_PROT_WRITE
| VM_PROT_NONE
:
5749 * Write without execute is RW|NX
5751 *kp
++ = pmap_bits_default
[PG_RW_IDX
] |
5752 pmap_bits_default
[PG_NX_IDX
];
5754 case VM_PROT_READ
| VM_PROT_WRITE
| VM_PROT_EXECUTE
:
5755 case VM_PROT_NONE
| VM_PROT_WRITE
| VM_PROT_EXECUTE
:
5757 * Write with execute is RW
5759 *kp
++ = pmap_bits_default
[PG_RW_IDX
];
5766 * Map a set of physical memory pages into the kernel virtual
5767 * address space. Return a pointer to where it is mapped. This
5768 * routine is intended to be used for mapping device memory,
5771 * NOTE: We can't use pgeflag unless we invalidate the pages one at
5774 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5775 * work whether the cpu supports PAT or not. The remaining PAT
5776 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5780 pmap_mapdev(vm_paddr_t pa
, vm_size_t size
)
5782 return(pmap_mapdev_attr(pa
, size
, PAT_WRITE_BACK
));
5786 pmap_mapdev_uncacheable(vm_paddr_t pa
, vm_size_t size
)
5788 return(pmap_mapdev_attr(pa
, size
, PAT_UNCACHEABLE
));
5792 pmap_mapbios(vm_paddr_t pa
, vm_size_t size
)
5794 return (pmap_mapdev_attr(pa
, size
, PAT_WRITE_BACK
));
5798 * Map a set of physical memory pages into the kernel virtual
5799 * address space. Return a pointer to where it is mapped. This
5800 * routine is intended to be used for mapping device memory,
5804 pmap_mapdev_attr(vm_paddr_t pa
, vm_size_t size
, int mode
)
5806 vm_offset_t va
, tmpva
, offset
;
5810 offset
= pa
& PAGE_MASK
;
5811 size
= roundup(offset
+ size
, PAGE_SIZE
);
5813 va
= kmem_alloc_nofault(&kernel_map
, size
, VM_SUBSYS_MAPDEV
, PAGE_SIZE
);
5815 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
5817 pa
= pa
& ~PAGE_MASK
;
5818 for (tmpva
= va
, tmpsize
= size
; tmpsize
> 0;) {
5819 pte
= vtopte(tmpva
);
5821 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
5822 kernel_pmap
.pmap_bits
[PG_V_IDX
] | /* pgeflag | */
5823 kernel_pmap
.pmap_cache_bits
[mode
];
5824 tmpsize
-= PAGE_SIZE
;
5828 pmap_invalidate_range(&kernel_pmap
, va
, va
+ size
);
5829 pmap_invalidate_cache_range(va
, va
+ size
);
5831 return ((void *)(va
+ offset
));
5835 pmap_unmapdev(vm_offset_t va
, vm_size_t size
)
5837 vm_offset_t base
, offset
;
5839 base
= va
& ~PAGE_MASK
;
5840 offset
= va
& PAGE_MASK
;
5841 size
= roundup(offset
+ size
, PAGE_SIZE
);
5842 pmap_qremove(va
, size
>> PAGE_SHIFT
);
5843 kmem_free(&kernel_map
, base
, size
);
5847 * Sets the memory attribute for the specified page.
5850 pmap_page_set_memattr(vm_page_t m
, vm_memattr_t ma
)
5856 * If "m" is a normal page, update its direct mapping. This update
5857 * can be relied upon to perform any cache operations that are
5858 * required for data coherence.
5860 if ((m
->flags
& PG_FICTITIOUS
) == 0)
5861 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m
)), 1, m
->pat_mode
);
5865 * Change the PAT attribute on an existing kernel memory map. Caller
5866 * must ensure that the virtual memory in question is not accessed
5867 * during the adjustment.
5870 pmap_change_attr(vm_offset_t va
, vm_size_t count
, int mode
)
5877 panic("pmap_change_attr: va is NULL");
5878 base
= trunc_page(va
);
5882 *pte
= (*pte
& ~(pt_entry_t
)(kernel_pmap
.pmap_cache_mask
)) |
5883 kernel_pmap
.pmap_cache_bits
[mode
];
5888 changed
= 1; /* XXX: not optimal */
5891 * Flush CPU caches if required to make sure any data isn't cached that
5892 * shouldn't be, etc.
5895 pmap_invalidate_range(&kernel_pmap
, base
, va
);
5896 pmap_invalidate_cache_range(base
, va
);
5901 * perform the pmap work for mincore
5904 pmap_mincore(pmap_t pmap
, vm_offset_t addr
)
5906 pt_entry_t
*ptep
, pte
;
5910 ptep
= pmap_pte(pmap
, addr
);
5912 if (ptep
&& (pte
= *ptep
) != 0) {
5915 val
= MINCORE_INCORE
;
5916 if ((pte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0)
5919 pa
= pte
& PG_FRAME
;
5921 if (pte
& pmap
->pmap_bits
[PG_DEVICE_IDX
])
5924 m
= PHYS_TO_VM_PAGE(pa
);
5929 if (pte
& pmap
->pmap_bits
[PG_M_IDX
])
5930 val
|= MINCORE_MODIFIED
|MINCORE_MODIFIED_OTHER
;
5932 * Modified by someone
5934 else if (m
&& (m
->dirty
|| pmap_is_modified(m
)))
5935 val
|= MINCORE_MODIFIED_OTHER
;
5939 if (pte
& pmap
->pmap_bits
[PG_A_IDX
])
5940 val
|= MINCORE_REFERENCED
|MINCORE_REFERENCED_OTHER
;
5943 * Referenced by someone
5945 else if (m
&& ((m
->flags
& PG_REFERENCED
) ||
5946 pmap_ts_referenced(m
))) {
5947 val
|= MINCORE_REFERENCED_OTHER
;
5948 vm_page_flag_set(m
, PG_REFERENCED
);
5957 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
5958 * vmspace will be ref'd and the old one will be deref'd.
5960 * The vmspace for all lwps associated with the process will be adjusted
5961 * and cr3 will be reloaded if any lwp is the current lwp.
5963 * The process must hold the vmspace->vm_map.token for oldvm and newvm
5966 pmap_replacevm(struct proc
*p
, struct vmspace
*newvm
, int adjrefs
)
5968 struct vmspace
*oldvm
;
5971 oldvm
= p
->p_vmspace
;
5972 if (oldvm
!= newvm
) {
5975 p
->p_vmspace
= newvm
;
5976 KKASSERT(p
->p_nthreads
== 1);
5977 lp
= RB_ROOT(&p
->p_lwp_tree
);
5978 pmap_setlwpvm(lp
, newvm
);
5985 * Set the vmspace for a LWP. The vmspace is almost universally set the
5986 * same as the process vmspace, but virtual kernels need to swap out contexts
5987 * on a per-lwp basis.
5989 * Caller does not necessarily hold any vmspace tokens. Caller must control
5990 * the lwp (typically be in the context of the lwp). We use a critical
5991 * section to protect against statclock and hardclock (statistics collection).
5994 pmap_setlwpvm(struct lwp
*lp
, struct vmspace
*newvm
)
5996 struct vmspace
*oldvm
;
5999 oldvm
= lp
->lwp_vmspace
;
6001 if (oldvm
!= newvm
) {
6003 KKASSERT((newvm
->vm_refcnt
& VM_REF_DELETED
) == 0);
6004 lp
->lwp_vmspace
= newvm
;
6005 if (curthread
->td_lwp
== lp
) {
6006 pmap
= vmspace_pmap(newvm
);
6007 ATOMIC_CPUMASK_ORBIT(pmap
->pm_active
, mycpu
->gd_cpuid
);
6008 if (pmap
->pm_active_lock
& CPULOCK_EXCL
)
6009 pmap_interlock_wait(newvm
);
6010 #if defined(SWTCH_OPTIM_STATS)
6013 if (pmap
->pmap_bits
[TYPE_IDX
] == REGULAR_PMAP
) {
6014 curthread
->td_pcb
->pcb_cr3
= vtophys(pmap
->pm_pml4
);
6015 } else if (pmap
->pmap_bits
[TYPE_IDX
] == EPT_PMAP
) {
6016 curthread
->td_pcb
->pcb_cr3
= KPML4phys
;
6018 panic("pmap_setlwpvm: unknown pmap type\n");
6020 load_cr3(curthread
->td_pcb
->pcb_cr3
);
6021 pmap
= vmspace_pmap(oldvm
);
6022 ATOMIC_CPUMASK_NANDBIT(pmap
->pm_active
,
6030 * Called when switching to a locked pmap, used to interlock against pmaps
6031 * undergoing modifications to prevent us from activating the MMU for the
6032 * target pmap until all such modifications have completed. We have to do
6033 * this because the thread making the modifications has already set up its
6034 * SMP synchronization mask.
6036 * This function cannot sleep!
6041 pmap_interlock_wait(struct vmspace
*vm
)
6043 struct pmap
*pmap
= &vm
->vm_pmap
;
6045 if (pmap
->pm_active_lock
& CPULOCK_EXCL
) {
6047 KKASSERT(curthread
->td_critcount
>= 2);
6048 DEBUG_PUSH_INFO("pmap_interlock_wait");
6049 while (pmap
->pm_active_lock
& CPULOCK_EXCL
) {
6051 lwkt_process_ipiq();
6059 pmap_addr_hint(vm_object_t obj
, vm_offset_t addr
, vm_size_t size
)
6062 if ((obj
== NULL
) || (size
< NBPDR
) ||
6063 ((obj
->type
!= OBJT_DEVICE
) && (obj
->type
!= OBJT_MGTDEVICE
))) {
6067 addr
= roundup2(addr
, NBPDR
);
6072 * Used by kmalloc/kfree, page already exists at va
6075 pmap_kvtom(vm_offset_t va
)
6077 pt_entry_t
*ptep
= vtopte(va
);
6079 KKASSERT((*ptep
& kernel_pmap
.pmap_bits
[PG_DEVICE_IDX
]) == 0);
6080 return(PHYS_TO_VM_PAGE(*ptep
& PG_FRAME
));
6084 * Initialize machine-specific shared page directory support. This
6085 * is executed when a VM object is created.
6088 pmap_object_init(vm_object_t object
)
6090 object
->md
.pmap_rw
= NULL
;
6091 object
->md
.pmap_ro
= NULL
;
6095 * Clean up machine-specific shared page directory support. This
6096 * is executed when a VM object is destroyed.
6099 pmap_object_free(vm_object_t object
)
6103 if ((pmap
= object
->md
.pmap_rw
) != NULL
) {
6104 object
->md
.pmap_rw
= NULL
;
6105 pmap_remove_noinval(pmap
,
6106 VM_MIN_USER_ADDRESS
, VM_MAX_USER_ADDRESS
);
6107 CPUMASK_ASSZERO(pmap
->pm_active
);
6110 kfree(pmap
, M_OBJPMAP
);
6112 if ((pmap
= object
->md
.pmap_ro
) != NULL
) {
6113 object
->md
.pmap_ro
= NULL
;
6114 pmap_remove_noinval(pmap
,
6115 VM_MIN_USER_ADDRESS
, VM_MAX_USER_ADDRESS
);
6116 CPUMASK_ASSZERO(pmap
->pm_active
);
6119 kfree(pmap
, M_OBJPMAP
);
6124 * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
6125 * VM page and issue a pginfo->callback.
6127 * We are expected to dispose of any non-NULL pte_pv.
6131 pmap_pgscan_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
6132 pv_entry_t pte_pv
, vm_pindex_t
*pte_placemark
,
6133 pv_entry_t pt_pv
, int sharept
,
6134 vm_offset_t va
, pt_entry_t
*ptep
, void *arg
)
6136 struct pmap_pgscan_info
*pginfo
= arg
;
6141 * Try to busy the page while we hold the pte_pv locked.
6143 KKASSERT(pte_pv
->pv_m
);
6144 m
= PHYS_TO_VM_PAGE(*ptep
& PG_FRAME
);
6145 if (vm_page_busy_try(m
, TRUE
) == 0) {
6146 if (m
== PHYS_TO_VM_PAGE(*ptep
& PG_FRAME
)) {
6148 * The callback is issued with the pte_pv
6149 * unlocked and put away, and the pt_pv
6154 vm_page_wire_quick(pt_pv
->pv_m
);
6157 if (pginfo
->callback(pginfo
, va
, m
) < 0)
6161 vm_page_unwire_quick(pt_pv
->pv_m
);
6168 ++pginfo
->busycount
;
6173 * Shared page table or unmanaged page (sharept or !sharept)
6175 pv_placemarker_wakeup(pmap
, pte_placemark
);
6180 pmap_pgscan(struct pmap_pgscan_info
*pginfo
)
6182 struct pmap_scan_info info
;
6184 pginfo
->offset
= pginfo
->beg_addr
;
6185 info
.pmap
= pginfo
->pmap
;
6186 info
.sva
= pginfo
->beg_addr
;
6187 info
.eva
= pginfo
->end_addr
;
6188 info
.func
= pmap_pgscan_callback
;
6190 pmap_scan(&info
, 0);
6192 pginfo
->offset
= pginfo
->end_addr
;
6196 * Wait for a placemarker that we do not own to clear. The placemarker
6197 * in question is not necessarily set to the pindex we want, we may have
6198 * to wait on the element because we want to reserve it ourselves.
6200 * NOTE: PM_PLACEMARK_WAKEUP sets a bit which is already set in
6201 * PM_NOPLACEMARK, so it does not interfere with placemarks
6202 * which have already been woken up.
6206 pv_placemarker_wait(pmap_t pmap
, vm_pindex_t
*pmark
)
6208 if (*pmark
!= PM_NOPLACEMARK
) {
6209 atomic_set_long(pmark
, PM_PLACEMARK_WAKEUP
);
6210 tsleep_interlock(pmark
, 0);
6211 if (*pmark
!= PM_NOPLACEMARK
)
6212 tsleep(pmark
, PINTERLOCKED
, "pvplw", 0);
6217 * Wakeup a placemarker that we own. Replace the entry with
6218 * PM_NOPLACEMARK and issue a wakeup() if necessary.
6222 pv_placemarker_wakeup(pmap_t pmap
, vm_pindex_t
*pmark
)
6226 pindex
= atomic_swap_long(pmark
, PM_NOPLACEMARK
);
6227 KKASSERT(pindex
!= PM_NOPLACEMARK
);
6228 if (pindex
& PM_PLACEMARK_WAKEUP
)