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/spinlock2.h>
78 #include <vm/vm_page2.h>
80 #include <machine/cputypes.h>
81 #include <machine/md_var.h>
82 #include <machine/specialreg.h>
83 #include <machine/smp.h>
84 #include <machine_base/apic/apicreg.h>
85 #include <machine/globaldata.h>
86 #include <machine/pmap.h>
87 #include <machine/pmap_inval.h>
88 #include <machine/inttypes.h>
92 #define PMAP_KEEP_PDIRS
93 #ifndef PMAP_SHPGPERPROC
94 #define PMAP_SHPGPERPROC 2000
97 #if defined(DIAGNOSTIC)
98 #define PMAP_DIAGNOSTIC
104 * pmap debugging will report who owns a pv lock when blocking.
108 #define PMAP_DEBUG_DECL ,const char *func, int lineno
109 #define PMAP_DEBUG_ARGS , __func__, __LINE__
110 #define PMAP_DEBUG_COPY , func, lineno
112 #define pv_get(pmap, pindex, pmarkp) _pv_get(pmap, pindex, pmarkp \
114 #define pv_lock(pv) _pv_lock(pv \
116 #define pv_hold_try(pv) _pv_hold_try(pv \
118 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp \
121 #define pv_free(pv, pvp) _pv_free(pv, pvp PMAP_DEBUG_ARGS)
125 #define PMAP_DEBUG_DECL
126 #define PMAP_DEBUG_ARGS
127 #define PMAP_DEBUG_COPY
129 #define pv_get(pmap, pindex, pmarkp) _pv_get(pmap, pindex, pmarkp)
130 #define pv_lock(pv) _pv_lock(pv)
131 #define pv_hold_try(pv) _pv_hold_try(pv)
132 #define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp)
133 #define pv_free(pv, pvp) _pv_free(pv, pvp)
138 * Get PDEs and PTEs for user/kernel address space
140 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
142 #define pmap_pde_v(pmap, pte) ((*(pd_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
143 #define pmap_pte_w(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_W_IDX]) != 0)
144 #define pmap_pte_m(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_M_IDX]) != 0)
145 #define pmap_pte_u(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_U_IDX]) != 0)
146 #define pmap_pte_v(pmap, pte) ((*(pt_entry_t *)pte & pmap->pmap_bits[PG_V_IDX]) != 0)
149 * Given a map and a machine independent protection code,
150 * convert to a vax protection code.
152 #define pte_prot(m, p) \
153 (m->protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
154 static uint64_t protection_codes
[PROTECTION_CODES_SIZE
];
156 struct pmap kernel_pmap
;
158 MALLOC_DEFINE(M_OBJPMAP
, "objpmap", "pmaps associated with VM objects");
160 vm_paddr_t avail_start
; /* PA of first available physical page */
161 vm_paddr_t avail_end
; /* PA of last available physical page */
162 vm_offset_t virtual2_start
; /* cutout free area prior to kernel start */
163 vm_offset_t virtual2_end
;
164 vm_offset_t virtual_start
; /* VA of first avail page (after kernel bss) */
165 vm_offset_t virtual_end
; /* VA of last avail page (end of kernel AS) */
166 vm_offset_t KvaStart
; /* VA start of KVA space */
167 vm_offset_t KvaEnd
; /* VA end of KVA space (non-inclusive) */
168 vm_offset_t KvaSize
; /* max size of kernel virtual address space */
169 static boolean_t pmap_initialized
= FALSE
; /* Has pmap_init completed? */
170 //static int pgeflag; /* PG_G or-in */
171 //static int pseflag; /* PG_PS or-in */
175 static vm_paddr_t dmaplimit
;
177 vm_offset_t kernel_vm_end
= VM_MIN_KERNEL_ADDRESS
;
179 static pt_entry_t pat_pte_index
[PAT_INDEX_SIZE
]; /* PAT -> PG_ bits */
180 /*static pt_entry_t pat_pde_index[PAT_INDEX_SIZE];*/ /* PAT -> PG_ bits */
182 static uint64_t KPTbase
;
183 static uint64_t KPTphys
;
184 static uint64_t KPDphys
; /* phys addr of kernel level 2 */
185 static uint64_t KPDbase
; /* phys addr of kernel level 2 @ KERNBASE */
186 uint64_t KPDPphys
; /* phys addr of kernel level 3 */
187 uint64_t KPML4phys
; /* phys addr of kernel level 4 */
189 static uint64_t DMPDphys
; /* phys addr of direct mapped level 2 */
190 static uint64_t DMPDPphys
; /* phys addr of direct mapped level 3 */
193 * Data for the pv entry allocation mechanism
195 static vm_zone_t pvzone
;
196 static struct vm_zone pvzone_store
;
197 static int pv_entry_max
=0, pv_entry_high_water
=0;
198 static int pmap_pagedaemon_waken
= 0;
199 static struct pv_entry
*pvinit
;
202 * All those kernel PT submaps that BSD is so fond of
204 pt_entry_t
*CMAP1
= NULL
, *ptmmap
;
205 caddr_t CADDR1
= NULL
, ptvmmap
= NULL
;
206 static pt_entry_t
*msgbufmap
;
207 struct msgbuf
*msgbufp
=NULL
;
210 * PMAP default PG_* bits. Needed to be able to add
211 * EPT/NPT pagetable pmap_bits for the VMM module
213 uint64_t pmap_bits_default
[] = {
214 REGULAR_PMAP
, /* TYPE_IDX 0 */
215 X86_PG_V
, /* PG_V_IDX 1 */
216 X86_PG_RW
, /* PG_RW_IDX 2 */
217 X86_PG_U
, /* PG_U_IDX 3 */
218 X86_PG_A
, /* PG_A_IDX 4 */
219 X86_PG_M
, /* PG_M_IDX 5 */
220 X86_PG_PS
, /* PG_PS_IDX3 6 */
221 X86_PG_G
, /* PG_G_IDX 7 */
222 X86_PG_AVAIL1
, /* PG_AVAIL1_IDX 8 */
223 X86_PG_AVAIL2
, /* PG_AVAIL2_IDX 9 */
224 X86_PG_AVAIL3
, /* PG_AVAIL3_IDX 10 */
225 X86_PG_NC_PWT
| X86_PG_NC_PCD
, /* PG_N_IDX 11 */
226 X86_PG_NX
, /* PG_NX_IDX 12 */
231 static pt_entry_t
*pt_crashdumpmap
;
232 static caddr_t crashdumpmap
;
234 static int pmap_debug
= 0;
235 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_debug
, CTLFLAG_RW
,
236 &pmap_debug
, 0, "Debug pmap's");
238 static int pmap_enter_debug
= 0;
239 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_enter_debug
, CTLFLAG_RW
,
240 &pmap_enter_debug
, 0, "Debug pmap_enter's");
242 static int pmap_yield_count
= 64;
243 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_yield_count
, CTLFLAG_RW
,
244 &pmap_yield_count
, 0, "Yield during init_pt/release");
245 static int pmap_mmu_optimize
= 0;
246 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_mmu_optimize
, CTLFLAG_RW
,
247 &pmap_mmu_optimize
, 0, "Share page table pages when possible");
248 int pmap_fast_kernel_cpusync
= 0;
249 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_fast_kernel_cpusync
, CTLFLAG_RW
,
250 &pmap_fast_kernel_cpusync
, 0, "Share page table pages when possible");
251 int pmap_dynamic_delete
= 0;
252 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_dynamic_delete
, CTLFLAG_RW
,
253 &pmap_dynamic_delete
, 0, "Dynamically delete PT/PD/PDPs");
254 int pmap_lock_delay
= 100;
255 SYSCTL_INT(_machdep
, OID_AUTO
, pmap_lock_delay
, CTLFLAG_RW
,
256 &pmap_lock_delay
, 0, "Spin loops");
258 static int pmap_nx_enable
= 0;
259 /* needs manual TUNABLE in early probe, see below */
263 /* Standard user access funtions */
264 extern int std_copyinstr (const void *udaddr
, void *kaddr
, size_t len
,
266 extern int std_copyin (const void *udaddr
, void *kaddr
, size_t len
);
267 extern int std_copyout (const void *kaddr
, void *udaddr
, size_t len
);
268 extern int std_fubyte (const uint8_t *base
);
269 extern int std_subyte (uint8_t *base
, uint8_t byte
);
270 extern int32_t std_fuword32 (const uint32_t *base
);
271 extern int64_t std_fuword64 (const uint64_t *base
);
272 extern int std_suword64 (uint64_t *base
, uint64_t word
);
273 extern int std_suword32 (uint32_t *base
, int word
);
274 extern uint32_t std_swapu32 (volatile uint32_t *base
, uint32_t v
);
275 extern uint64_t std_swapu64 (volatile uint64_t *base
, uint64_t v
);
277 static void pv_hold(pv_entry_t pv
);
278 static int _pv_hold_try(pv_entry_t pv
280 static void pv_drop(pv_entry_t pv
);
281 static void _pv_lock(pv_entry_t pv
283 static void pv_unlock(pv_entry_t pv
);
284 static pv_entry_t
_pv_alloc(pmap_t pmap
, vm_pindex_t pindex
, int *isnew
286 static pv_entry_t
_pv_get(pmap_t pmap
, vm_pindex_t pindex
, vm_pindex_t
**pmarkp
288 static void _pv_free(pv_entry_t pv
, pv_entry_t pvp PMAP_DEBUG_DECL
);
289 static pv_entry_t
pv_get_try(pmap_t pmap
, vm_pindex_t pindex
,
290 vm_pindex_t
**pmarkp
, int *errorp
);
291 static void pv_put(pv_entry_t pv
);
292 static void *pv_pte_lookup(pv_entry_t pv
, vm_pindex_t pindex
);
293 static pv_entry_t
pmap_allocpte(pmap_t pmap
, vm_pindex_t ptepindex
,
295 static pv_entry_t
pmap_allocpte_seg(pmap_t pmap
, vm_pindex_t ptepindex
,
296 pv_entry_t
*pvpp
, vm_map_entry_t entry
, vm_offset_t va
);
297 static void pmap_remove_pv_pte(pv_entry_t pv
, pv_entry_t pvp
,
298 pmap_inval_bulk_t
*bulk
, int destroy
);
299 static vm_page_t
pmap_remove_pv_page(pv_entry_t pv
);
300 static int pmap_release_pv(pv_entry_t pv
, pv_entry_t pvp
,
301 pmap_inval_bulk_t
*bulk
);
303 struct pmap_scan_info
;
304 static void pmap_remove_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
305 pv_entry_t pte_pv
, vm_pindex_t
*pte_placemark
,
306 pv_entry_t pt_pv
, int sharept
,
307 vm_offset_t va
, pt_entry_t
*ptep
, void *arg __unused
);
308 static void pmap_protect_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
309 pv_entry_t pte_pv
, vm_pindex_t
*pte_placemark
,
310 pv_entry_t pt_pv
, int sharept
,
311 vm_offset_t va
, pt_entry_t
*ptep
, void *arg __unused
);
313 static void i386_protection_init (void);
314 static void create_pagetables(vm_paddr_t
*firstaddr
);
315 static void pmap_remove_all (vm_page_t m
);
316 static boolean_t
pmap_testbit (vm_page_t m
, int bit
);
318 static pt_entry_t
* pmap_pte_quick (pmap_t pmap
, vm_offset_t va
);
319 static vm_offset_t
pmap_kmem_choose(vm_offset_t addr
);
321 static void pmap_pinit_defaults(struct pmap
*pmap
);
322 static void pv_placemarker_wait(pmap_t pmap
, vm_pindex_t
*pmark
);
323 static void pv_placemarker_wakeup(pmap_t pmap
, vm_pindex_t
*pmark
);
325 static unsigned pdir4mb
;
328 pv_entry_compare(pv_entry_t pv1
, pv_entry_t pv2
)
330 if (pv1
->pv_pindex
< pv2
->pv_pindex
)
332 if (pv1
->pv_pindex
> pv2
->pv_pindex
)
337 RB_GENERATE2(pv_entry_rb_tree
, pv_entry
, pv_entry
,
338 pv_entry_compare
, vm_pindex_t
, pv_pindex
);
342 pmap_page_stats_adding(vm_page_t m
)
344 globaldata_t gd
= mycpu
;
346 if (TAILQ_EMPTY(&m
->md
.pv_list
)) {
347 ++gd
->gd_vmtotal
.t_arm
;
348 } else if (TAILQ_FIRST(&m
->md
.pv_list
) ==
349 TAILQ_LAST(&m
->md
.pv_list
, md_page_pv_list
)) {
350 ++gd
->gd_vmtotal
.t_armshr
;
351 ++gd
->gd_vmtotal
.t_avmshr
;
353 ++gd
->gd_vmtotal
.t_avmshr
;
359 pmap_page_stats_deleting(vm_page_t m
)
361 globaldata_t gd
= mycpu
;
363 if (TAILQ_EMPTY(&m
->md
.pv_list
)) {
364 --gd
->gd_vmtotal
.t_arm
;
365 } else if (TAILQ_FIRST(&m
->md
.pv_list
) ==
366 TAILQ_LAST(&m
->md
.pv_list
, md_page_pv_list
)) {
367 --gd
->gd_vmtotal
.t_armshr
;
368 --gd
->gd_vmtotal
.t_avmshr
;
370 --gd
->gd_vmtotal
.t_avmshr
;
375 * This is an ineligent crowbar to prevent heavily threaded programs
376 * from creating long live-locks in the pmap code when pmap_mmu_optimize
377 * is enabled. Without it a pmap-local page table page can wind up being
378 * constantly created and destroyed (without injury, but also without
379 * progress) as the optimization tries to switch to the object's shared page
383 pmap_softwait(pmap_t pmap
)
385 while (pmap
->pm_softhold
) {
386 tsleep_interlock(&pmap
->pm_softhold
, 0);
387 if (pmap
->pm_softhold
)
388 tsleep(&pmap
->pm_softhold
, PINTERLOCKED
, "mmopt", 0);
393 pmap_softhold(pmap_t pmap
)
395 while (atomic_swap_int(&pmap
->pm_softhold
, 1) == 1) {
396 tsleep_interlock(&pmap
->pm_softhold
, 0);
397 if (atomic_swap_int(&pmap
->pm_softhold
, 1) == 1)
398 tsleep(&pmap
->pm_softhold
, PINTERLOCKED
, "mmopt", 0);
403 pmap_softdone(pmap_t pmap
)
405 atomic_swap_int(&pmap
->pm_softhold
, 0);
406 wakeup(&pmap
->pm_softhold
);
410 * Move the kernel virtual free pointer to the next
411 * 2MB. This is used to help improve performance
412 * by using a large (2MB) page for much of the kernel
413 * (.text, .data, .bss)
417 pmap_kmem_choose(vm_offset_t addr
)
419 vm_offset_t newaddr
= addr
;
421 newaddr
= roundup2(addr
, NBPDR
);
426 * Returns the pindex of a page table entry (representing a terminal page).
427 * There are NUPTE_TOTAL page table entries possible (a huge number)
429 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
430 * We want to properly translate negative KVAs.
434 pmap_pte_pindex(vm_offset_t va
)
436 return ((va
>> PAGE_SHIFT
) & (NUPTE_TOTAL
- 1));
440 * Returns the pindex of a page table.
444 pmap_pt_pindex(vm_offset_t va
)
446 return (NUPTE_TOTAL
+ ((va
>> PDRSHIFT
) & (NUPT_TOTAL
- 1)));
450 * Returns the pindex of a page directory.
454 pmap_pd_pindex(vm_offset_t va
)
456 return (NUPTE_TOTAL
+ NUPT_TOTAL
+
457 ((va
>> PDPSHIFT
) & (NUPD_TOTAL
- 1)));
462 pmap_pdp_pindex(vm_offset_t va
)
464 return (NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
+
465 ((va
>> PML4SHIFT
) & (NUPDP_TOTAL
- 1)));
470 pmap_pml4_pindex(void)
472 return (NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
+ NUPDP_TOTAL
);
476 * Return various clipped indexes for a given VA
478 * Returns the index of a pt in a page directory, representing a page
483 pmap_pt_index(vm_offset_t va
)
485 return ((va
>> PDRSHIFT
) & ((1ul << NPDEPGSHIFT
) - 1));
489 * Returns the index of a pd in a page directory page, representing a page
494 pmap_pd_index(vm_offset_t va
)
496 return ((va
>> PDPSHIFT
) & ((1ul << NPDPEPGSHIFT
) - 1));
500 * Returns the index of a pdp in the pml4 table, representing a page
505 pmap_pdp_index(vm_offset_t va
)
507 return ((va
>> PML4SHIFT
) & ((1ul << NPML4EPGSHIFT
) - 1));
511 * Locate the requested pt_entry
515 pv_entry_lookup(pmap_t pmap
, vm_pindex_t pindex
)
519 if (pindex
< pmap_pt_pindex(0))
520 pv
= pmap
->pm_pvhint_pte
;
521 else if (pindex
< pmap_pd_pindex(0))
522 pv
= pmap
->pm_pvhint_pt
;
526 if (pv
== NULL
|| pv
->pv_pmap
!= pmap
) {
527 pv
= pv_entry_rb_tree_RB_LOOKUP(&pmap
->pm_pvroot
,
529 } else if (pv
->pv_pindex
!= pindex
) {
530 pv
= pv_entry_rb_tree_RB_LOOKUP_REL(&pmap
->pm_pvroot
,
539 * Super fast pmap_pte routine best used when scanning the pv lists.
540 * This eliminates many course-grained invltlb calls. Note that many of
541 * the pv list scans are across different pmaps and it is very wasteful
542 * to do an entire invltlb when checking a single mapping.
544 static __inline pt_entry_t
*pmap_pte(pmap_t pmap
, vm_offset_t va
);
548 pmap_pte_quick(pmap_t pmap
, vm_offset_t va
)
550 return pmap_pte(pmap
, va
);
554 * The placemarker hash must be broken up into four zones so lock
555 * ordering semantics continue to work (e.g. pte, pt, pd, then pdp).
557 * Placemarkers are used to 'lock' page table indices that do not have
558 * a pv_entry. This allows the pmap to support managed and unmanaged
559 * pages and shared page tables.
561 #define PM_PLACE_BASE (PM_PLACEMARKS >> 2)
565 pmap_placemarker_hash(pmap_t pmap
, vm_pindex_t pindex
)
569 if (pindex
< pmap_pt_pindex(0)) /* zone 0 - PTE */
571 else if (pindex
< pmap_pd_pindex(0)) /* zone 1 - PT */
573 else if (pindex
< pmap_pdp_pindex(0)) /* zone 2 - PD */
574 hi
= PM_PLACE_BASE
<< 1;
575 else /* zone 3 - PDP (and PML4E) */
576 hi
= PM_PLACE_BASE
| (PM_PLACE_BASE
<< 1);
577 hi
+= pindex
& (PM_PLACE_BASE
- 1);
579 return (&pmap
->pm_placemarks
[hi
]);
584 * Generic procedure to index a pte from a pt, pd, or pdp.
586 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
587 * a page table page index but is instead of PV lookup index.
591 pv_pte_lookup(pv_entry_t pv
, vm_pindex_t pindex
)
595 pte
= (pt_entry_t
*)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv
->pv_m
));
596 return(&pte
[pindex
]);
600 * Return pointer to PDP slot in the PML4
604 pmap_pdp(pmap_t pmap
, vm_offset_t va
)
606 return (&pmap
->pm_pml4
[pmap_pdp_index(va
)]);
610 * Return pointer to PD slot in the PDP given a pointer to the PDP
614 pmap_pdp_to_pd(pml4_entry_t pdp_pte
, vm_offset_t va
)
618 pd
= (pdp_entry_t
*)PHYS_TO_DMAP(pdp_pte
& PG_FRAME
);
619 return (&pd
[pmap_pd_index(va
)]);
623 * Return pointer to PD slot in the PDP.
627 pmap_pd(pmap_t pmap
, vm_offset_t va
)
631 pdp
= pmap_pdp(pmap
, va
);
632 if ((*pdp
& pmap
->pmap_bits
[PG_V_IDX
]) == 0)
634 return (pmap_pdp_to_pd(*pdp
, va
));
638 * Return pointer to PT slot in the PD given a pointer to the PD
642 pmap_pd_to_pt(pdp_entry_t pd_pte
, vm_offset_t va
)
646 pt
= (pd_entry_t
*)PHYS_TO_DMAP(pd_pte
& PG_FRAME
);
647 return (&pt
[pmap_pt_index(va
)]);
651 * Return pointer to PT slot in the PD
653 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
654 * so we cannot lookup the PD via the PDP. Instead we
655 * must look it up via the pmap.
659 pmap_pt(pmap_t pmap
, vm_offset_t va
)
663 vm_pindex_t pd_pindex
;
666 if (pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) {
667 pd_pindex
= pmap_pd_pindex(va
);
668 spin_lock_shared(&pmap
->pm_spin
);
669 pv
= pv_entry_rb_tree_RB_LOOKUP(&pmap
->pm_pvroot
, pd_pindex
);
670 if (pv
== NULL
|| pv
->pv_m
== NULL
) {
671 spin_unlock_shared(&pmap
->pm_spin
);
674 phys
= VM_PAGE_TO_PHYS(pv
->pv_m
);
675 spin_unlock_shared(&pmap
->pm_spin
);
676 return (pmap_pd_to_pt(phys
, va
));
678 pd
= pmap_pd(pmap
, va
);
679 if (pd
== NULL
|| (*pd
& pmap
->pmap_bits
[PG_V_IDX
]) == 0)
681 return (pmap_pd_to_pt(*pd
, va
));
686 * Return pointer to PTE slot in the PT given a pointer to the PT
690 pmap_pt_to_pte(pd_entry_t pt_pte
, vm_offset_t va
)
694 pte
= (pt_entry_t
*)PHYS_TO_DMAP(pt_pte
& PG_FRAME
);
695 return (&pte
[pmap_pte_index(va
)]);
699 * Return pointer to PTE slot in the PT
703 pmap_pte(pmap_t pmap
, vm_offset_t va
)
707 pt
= pmap_pt(pmap
, va
);
708 if (pt
== NULL
|| (*pt
& pmap
->pmap_bits
[PG_V_IDX
]) == 0)
710 if ((*pt
& pmap
->pmap_bits
[PG_PS_IDX
]) != 0)
711 return ((pt_entry_t
*)pt
);
712 return (pmap_pt_to_pte(*pt
, va
));
716 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
717 * the PT layer. This will speed up core pmap operations considerably.
719 * NOTE: The pmap spinlock does not need to be held but the passed-in pv
720 * must be in a known associated state (typically by being locked when
721 * the pmap spinlock isn't held). We allow the race for that case.
723 * NOTE: pm_pvhint* is only accessed (read) with the spin-lock held, using
724 * cpu_ccfence() to prevent compiler optimizations from reloading the
729 pv_cache(pv_entry_t pv
, vm_pindex_t pindex
)
731 if (pindex
< pmap_pt_pindex(0)) {
733 pv
->pv_pmap
->pm_pvhint_pte
= pv
;
734 } else if (pindex
< pmap_pd_pindex(0)) {
736 pv
->pv_pmap
->pm_pvhint_pt
= pv
;
742 * Return address of PT slot in PD (KVM only)
744 * Cannot be used for user page tables because it might interfere with
745 * the shared page-table-page optimization (pmap_mmu_optimize).
749 vtopt(vm_offset_t va
)
751 uint64_t mask
= ((1ul << (NPDEPGSHIFT
+ NPDPEPGSHIFT
+
752 NPML4EPGSHIFT
)) - 1);
754 return (PDmap
+ ((va
>> PDRSHIFT
) & mask
));
758 * KVM - return address of PTE slot in PT
762 vtopte(vm_offset_t va
)
764 uint64_t mask
= ((1ul << (NPTEPGSHIFT
+ NPDEPGSHIFT
+
765 NPDPEPGSHIFT
+ NPML4EPGSHIFT
)) - 1);
767 return (PTmap
+ ((va
>> PAGE_SHIFT
) & mask
));
771 * Returns the physical address translation from va for a user address.
772 * (vm_paddr_t)-1 is returned on failure.
775 uservtophys(vm_offset_t va
)
777 uint64_t mask
= ((1ul << (NPTEPGSHIFT
+ NPDEPGSHIFT
+
778 NPDPEPGSHIFT
+ NPML4EPGSHIFT
)) - 1);
783 pmap
= vmspace_pmap(mycpu
->gd_curthread
->td_lwp
->lwp_vmspace
);
785 if (va
< VM_MAX_USER_ADDRESS
) {
786 pte
= kreadmem64(PTmap
+ ((va
>> PAGE_SHIFT
) & mask
));
787 if (pte
& pmap
->pmap_bits
[PG_V_IDX
])
788 pa
= (pte
& PG_FRAME
) | (va
& PAGE_MASK
);
794 allocpages(vm_paddr_t
*firstaddr
, long n
)
799 bzero((void *)ret
, n
* PAGE_SIZE
);
800 *firstaddr
+= n
* PAGE_SIZE
;
806 create_pagetables(vm_paddr_t
*firstaddr
)
808 long i
; /* must be 64 bits */
814 * We are running (mostly) V=P at this point
816 * Calculate NKPT - number of kernel page tables. We have to
817 * accomodoate prealloction of the vm_page_array, dump bitmap,
818 * MSGBUF_SIZE, and other stuff. Be generous.
820 * Maxmem is in pages.
822 * ndmpdp is the number of 1GB pages we wish to map.
824 ndmpdp
= (ptoa(Maxmem
) + NBPDP
- 1) >> PDPSHIFT
;
825 if (ndmpdp
< 4) /* Minimum 4GB of dirmap */
827 KKASSERT(ndmpdp
<= NKPDPE
* NPDEPG
);
830 * Starting at the beginning of kvm (not KERNBASE).
832 nkpt_phys
= (Maxmem
* sizeof(struct vm_page
) + NBPDR
- 1) / NBPDR
;
833 nkpt_phys
+= (Maxmem
* sizeof(struct pv_entry
) + NBPDR
- 1) / NBPDR
;
834 nkpt_phys
+= ((nkpt
+ nkpt
+ 1 + NKPML4E
+ NKPDPE
+ NDMPML4E
+
835 ndmpdp
) + 511) / 512;
839 * Starting at KERNBASE - map 2G worth of page table pages.
840 * KERNBASE is offset -2G from the end of kvm.
842 nkpt_base
= (NPDPEPG
- KPDPI
) * NPTEPG
; /* typically 2 x 512 */
847 KPTbase
= allocpages(firstaddr
, nkpt_base
);
848 KPTphys
= allocpages(firstaddr
, nkpt_phys
);
849 KPML4phys
= allocpages(firstaddr
, 1);
850 KPDPphys
= allocpages(firstaddr
, NKPML4E
);
851 KPDphys
= allocpages(firstaddr
, NKPDPE
);
854 * Calculate the page directory base for KERNBASE,
855 * that is where we start populating the page table pages.
856 * Basically this is the end - 2.
858 KPDbase
= KPDphys
+ ((NKPDPE
- (NPDPEPG
- KPDPI
)) << PAGE_SHIFT
);
860 DMPDPphys
= allocpages(firstaddr
, NDMPML4E
);
861 if ((amd_feature
& AMDID_PAGE1GB
) == 0)
862 DMPDphys
= allocpages(firstaddr
, ndmpdp
);
863 dmaplimit
= (vm_paddr_t
)ndmpdp
<< PDPSHIFT
;
866 * Fill in the underlying page table pages for the area around
867 * KERNBASE. This remaps low physical memory to KERNBASE.
869 * Read-only from zero to physfree
870 * XXX not fully used, underneath 2M pages
872 for (i
= 0; (i
<< PAGE_SHIFT
) < *firstaddr
; i
++) {
873 ((pt_entry_t
*)KPTbase
)[i
] = i
<< PAGE_SHIFT
;
874 ((pt_entry_t
*)KPTbase
)[i
] |=
875 pmap_bits_default
[PG_RW_IDX
] |
876 pmap_bits_default
[PG_V_IDX
] |
877 pmap_bits_default
[PG_G_IDX
];
881 * Now map the initial kernel page tables. One block of page
882 * tables is placed at the beginning of kernel virtual memory,
883 * and another block is placed at KERNBASE to map the kernel binary,
884 * data, bss, and initial pre-allocations.
886 for (i
= 0; i
< nkpt_base
; i
++) {
887 ((pd_entry_t
*)KPDbase
)[i
] = KPTbase
+ (i
<< PAGE_SHIFT
);
888 ((pd_entry_t
*)KPDbase
)[i
] |=
889 pmap_bits_default
[PG_RW_IDX
] |
890 pmap_bits_default
[PG_V_IDX
];
892 for (i
= 0; i
< nkpt_phys
; i
++) {
893 ((pd_entry_t
*)KPDphys
)[i
] = KPTphys
+ (i
<< PAGE_SHIFT
);
894 ((pd_entry_t
*)KPDphys
)[i
] |=
895 pmap_bits_default
[PG_RW_IDX
] |
896 pmap_bits_default
[PG_V_IDX
];
900 * Map from zero to end of allocations using 2M pages as an
901 * optimization. This will bypass some of the KPTBase pages
902 * above in the KERNBASE area.
904 for (i
= 0; (i
<< PDRSHIFT
) < *firstaddr
; i
++) {
905 ((pd_entry_t
*)KPDbase
)[i
] = i
<< PDRSHIFT
;
906 ((pd_entry_t
*)KPDbase
)[i
] |=
907 pmap_bits_default
[PG_RW_IDX
] |
908 pmap_bits_default
[PG_V_IDX
] |
909 pmap_bits_default
[PG_PS_IDX
] |
910 pmap_bits_default
[PG_G_IDX
];
914 * And connect up the PD to the PDP. The kernel pmap is expected
915 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
917 for (i
= 0; i
< NKPDPE
; i
++) {
918 ((pdp_entry_t
*)KPDPphys
)[NPDPEPG
- NKPDPE
+ i
] =
919 KPDphys
+ (i
<< PAGE_SHIFT
);
920 ((pdp_entry_t
*)KPDPphys
)[NPDPEPG
- NKPDPE
+ i
] |=
921 pmap_bits_default
[PG_RW_IDX
] |
922 pmap_bits_default
[PG_V_IDX
] |
923 pmap_bits_default
[PG_U_IDX
];
927 * Now set up the direct map space using either 2MB or 1GB pages
928 * Preset PG_M and PG_A because demotion expects it.
930 * When filling in entries in the PD pages make sure any excess
931 * entries are set to zero as we allocated enough PD pages
933 if ((amd_feature
& AMDID_PAGE1GB
) == 0) {
934 for (i
= 0; i
< NPDEPG
* ndmpdp
; i
++) {
935 ((pd_entry_t
*)DMPDphys
)[i
] = i
<< PDRSHIFT
;
936 ((pd_entry_t
*)DMPDphys
)[i
] |=
937 pmap_bits_default
[PG_RW_IDX
] |
938 pmap_bits_default
[PG_V_IDX
] |
939 pmap_bits_default
[PG_PS_IDX
] |
940 pmap_bits_default
[PG_G_IDX
] |
941 pmap_bits_default
[PG_M_IDX
] |
942 pmap_bits_default
[PG_A_IDX
];
946 * And the direct map space's PDP
948 for (i
= 0; i
< ndmpdp
; i
++) {
949 ((pdp_entry_t
*)DMPDPphys
)[i
] = DMPDphys
+
951 ((pdp_entry_t
*)DMPDPphys
)[i
] |=
952 pmap_bits_default
[PG_RW_IDX
] |
953 pmap_bits_default
[PG_V_IDX
] |
954 pmap_bits_default
[PG_U_IDX
];
957 for (i
= 0; i
< ndmpdp
; i
++) {
958 ((pdp_entry_t
*)DMPDPphys
)[i
] =
959 (vm_paddr_t
)i
<< PDPSHIFT
;
960 ((pdp_entry_t
*)DMPDPphys
)[i
] |=
961 pmap_bits_default
[PG_RW_IDX
] |
962 pmap_bits_default
[PG_V_IDX
] |
963 pmap_bits_default
[PG_PS_IDX
] |
964 pmap_bits_default
[PG_G_IDX
] |
965 pmap_bits_default
[PG_M_IDX
] |
966 pmap_bits_default
[PG_A_IDX
];
970 /* And recursively map PML4 to itself in order to get PTmap */
971 ((pdp_entry_t
*)KPML4phys
)[PML4PML4I
] = KPML4phys
;
972 ((pdp_entry_t
*)KPML4phys
)[PML4PML4I
] |=
973 pmap_bits_default
[PG_RW_IDX
] |
974 pmap_bits_default
[PG_V_IDX
] |
975 pmap_bits_default
[PG_U_IDX
];
978 * Connect the Direct Map slots up to the PML4
980 for (j
= 0; j
< NDMPML4E
; ++j
) {
981 ((pdp_entry_t
*)KPML4phys
)[DMPML4I
+ j
] =
982 (DMPDPphys
+ ((vm_paddr_t
)j
<< PAGE_SHIFT
)) |
983 pmap_bits_default
[PG_RW_IDX
] |
984 pmap_bits_default
[PG_V_IDX
] |
985 pmap_bits_default
[PG_U_IDX
];
989 * Connect the KVA slot up to the PML4
991 ((pdp_entry_t
*)KPML4phys
)[KPML4I
] = KPDPphys
;
992 ((pdp_entry_t
*)KPML4phys
)[KPML4I
] |=
993 pmap_bits_default
[PG_RW_IDX
] |
994 pmap_bits_default
[PG_V_IDX
] |
995 pmap_bits_default
[PG_U_IDX
];
999 * Bootstrap the system enough to run with virtual memory.
1001 * On the i386 this is called after mapping has already been enabled
1002 * and just syncs the pmap module with what has already been done.
1003 * [We can't call it easily with mapping off since the kernel is not
1004 * mapped with PA == VA, hence we would have to relocate every address
1005 * from the linked base (virtual) address "KERNBASE" to the actual
1006 * (physical) address starting relative to 0]
1009 pmap_bootstrap(vm_paddr_t
*firstaddr
)
1015 KvaStart
= VM_MIN_KERNEL_ADDRESS
;
1016 KvaEnd
= VM_MAX_KERNEL_ADDRESS
;
1017 KvaSize
= KvaEnd
- KvaStart
;
1019 avail_start
= *firstaddr
;
1022 * Create an initial set of page tables to run the kernel in.
1024 create_pagetables(firstaddr
);
1026 virtual2_start
= KvaStart
;
1027 virtual2_end
= PTOV_OFFSET
;
1029 virtual_start
= (vm_offset_t
) PTOV_OFFSET
+ *firstaddr
;
1030 virtual_start
= pmap_kmem_choose(virtual_start
);
1032 virtual_end
= VM_MAX_KERNEL_ADDRESS
;
1034 /* XXX do %cr0 as well */
1035 load_cr4(rcr4() | CR4_PGE
| CR4_PSE
);
1036 load_cr3(KPML4phys
);
1039 * Initialize protection array.
1041 i386_protection_init();
1044 * The kernel's pmap is statically allocated so we don't have to use
1045 * pmap_create, which is unlikely to work correctly at this part of
1046 * the boot sequence (XXX and which no longer exists).
1048 kernel_pmap
.pm_pml4
= (pdp_entry_t
*) (PTOV_OFFSET
+ KPML4phys
);
1049 kernel_pmap
.pm_count
= 1;
1050 CPUMASK_ASSALLONES(kernel_pmap
.pm_active
);
1051 RB_INIT(&kernel_pmap
.pm_pvroot
);
1052 spin_init(&kernel_pmap
.pm_spin
, "pmapbootstrap");
1053 for (i
= 0; i
< PM_PLACEMARKS
; ++i
)
1054 kernel_pmap
.pm_placemarks
[i
] = PM_NOPLACEMARK
;
1057 * Reserve some special page table entries/VA space for temporary
1060 #define SYSMAP(c, p, v, n) \
1061 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
1067 * CMAP1/CMAP2 are used for zeroing and copying pages.
1069 SYSMAP(caddr_t
, CMAP1
, CADDR1
, 1)
1074 SYSMAP(caddr_t
, pt_crashdumpmap
, crashdumpmap
, MAXDUMPPGS
);
1077 * ptvmmap is used for reading arbitrary physical pages via
1080 SYSMAP(caddr_t
, ptmmap
, ptvmmap
, 1)
1083 * msgbufp is used to map the system message buffer.
1084 * XXX msgbufmap is not used.
1086 SYSMAP(struct msgbuf
*, msgbufmap
, msgbufp
,
1087 atop(round_page(MSGBUF_SIZE
)))
1090 virtual_start
= pmap_kmem_choose(virtual_start
);
1095 * PG_G is terribly broken on SMP because we IPI invltlb's in some
1096 * cases rather then invl1pg. Actually, I don't even know why it
1097 * works under UP because self-referential page table mappings
1102 * Initialize the 4MB page size flag
1106 * The 4MB page version of the initial
1107 * kernel page mapping.
1111 #if !defined(DISABLE_PSE)
1112 if (cpu_feature
& CPUID_PSE
) {
1115 * Note that we have enabled PSE mode
1117 // pseflag = kernel_pmap.pmap_bits[PG_PS_IDX];
1118 ptditmp
= *(PTmap
+ x86_64_btop(KERNBASE
));
1119 ptditmp
&= ~(NBPDR
- 1);
1120 ptditmp
|= pmap_bits_default
[PG_V_IDX
] |
1121 pmap_bits_default
[PG_RW_IDX
] |
1122 pmap_bits_default
[PG_PS_IDX
] |
1123 pmap_bits_default
[PG_U_IDX
];
1130 /* Initialize the PAT MSR */
1132 pmap_pinit_defaults(&kernel_pmap
);
1134 TUNABLE_INT_FETCH("machdep.pmap_fast_kernel_cpusync",
1135 &pmap_fast_kernel_cpusync
);
1140 * Setup the PAT MSR.
1149 * Default values mapping PATi,PCD,PWT bits at system reset.
1150 * The default values effectively ignore the PATi bit by
1151 * repeating the encodings for 0-3 in 4-7, and map the PCD
1152 * and PWT bit combinations to the expected PAT types.
1154 pat_msr
= PAT_VALUE(0, PAT_WRITE_BACK
) | /* 000 */
1155 PAT_VALUE(1, PAT_WRITE_THROUGH
) | /* 001 */
1156 PAT_VALUE(2, PAT_UNCACHED
) | /* 010 */
1157 PAT_VALUE(3, PAT_UNCACHEABLE
) | /* 011 */
1158 PAT_VALUE(4, PAT_WRITE_BACK
) | /* 100 */
1159 PAT_VALUE(5, PAT_WRITE_THROUGH
) | /* 101 */
1160 PAT_VALUE(6, PAT_UNCACHED
) | /* 110 */
1161 PAT_VALUE(7, PAT_UNCACHEABLE
); /* 111 */
1162 pat_pte_index
[PAT_WRITE_BACK
] = 0;
1163 pat_pte_index
[PAT_WRITE_THROUGH
]= 0 | X86_PG_NC_PWT
;
1164 pat_pte_index
[PAT_UNCACHED
] = X86_PG_NC_PCD
;
1165 pat_pte_index
[PAT_UNCACHEABLE
] = X86_PG_NC_PCD
| X86_PG_NC_PWT
;
1166 pat_pte_index
[PAT_WRITE_PROTECTED
] = pat_pte_index
[PAT_UNCACHEABLE
];
1167 pat_pte_index
[PAT_WRITE_COMBINING
] = pat_pte_index
[PAT_UNCACHEABLE
];
1169 if (cpu_feature
& CPUID_PAT
) {
1171 * If we support the PAT then set-up entries for
1172 * WRITE_PROTECTED and WRITE_COMBINING using bit patterns
1175 pat_msr
= (pat_msr
& ~PAT_MASK(5)) |
1176 PAT_VALUE(5, PAT_WRITE_PROTECTED
);
1177 pat_msr
= (pat_msr
& ~PAT_MASK(6)) |
1178 PAT_VALUE(6, PAT_WRITE_COMBINING
);
1179 pat_pte_index
[PAT_WRITE_PROTECTED
] = X86_PG_PTE_PAT
| X86_PG_NC_PWT
;
1180 pat_pte_index
[PAT_WRITE_COMBINING
] = X86_PG_PTE_PAT
| X86_PG_NC_PCD
;
1183 * Then enable the PAT
1188 load_cr4(cr4
& ~CR4_PGE
);
1190 /* Disable caches (CD = 1, NW = 0). */
1192 load_cr0((cr0
& ~CR0_NW
) | CR0_CD
);
1194 /* Flushes caches and TLBs. */
1198 /* Update PAT and index table. */
1199 wrmsr(MSR_PAT
, pat_msr
);
1201 /* Flush caches and TLBs again. */
1205 /* Restore caches and PGE. */
1213 * Set 4mb pdir for mp startup
1218 if (cpu_feature
& CPUID_PSE
) {
1219 load_cr4(rcr4() | CR4_PSE
);
1220 if (pdir4mb
&& mycpu
->gd_cpuid
== 0) { /* only on BSP */
1227 * Initialize the pmap module.
1228 * Called by vm_init, to initialize any structures that the pmap
1229 * system needs to map virtual memory.
1230 * pmap_init has been enhanced to support in a fairly consistant
1231 * way, discontiguous physical memory.
1240 * Allocate memory for random pmap data structures. Includes the
1244 for (i
= 0; i
< vm_page_array_size
; i
++) {
1247 m
= &vm_page_array
[i
];
1248 TAILQ_INIT(&m
->md
.pv_list
);
1252 * init the pv free list
1254 initial_pvs
= vm_page_array_size
;
1255 if (initial_pvs
< MINPV
)
1256 initial_pvs
= MINPV
;
1257 pvzone
= &pvzone_store
;
1258 pvinit
= (void *)kmem_alloc(&kernel_map
,
1259 initial_pvs
* sizeof (struct pv_entry
),
1261 zbootinit(pvzone
, "PV ENTRY", sizeof (struct pv_entry
),
1262 pvinit
, initial_pvs
);
1265 * Now it is safe to enable pv_table recording.
1267 pmap_initialized
= TRUE
;
1271 * Initialize the address space (zone) for the pv_entries. Set a
1272 * high water mark so that the system can recover from excessive
1273 * numbers of pv entries.
1278 int shpgperproc
= PMAP_SHPGPERPROC
;
1281 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc
);
1282 pv_entry_max
= shpgperproc
* maxproc
+ vm_page_array_size
;
1283 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max
);
1284 pv_entry_high_water
= 9 * (pv_entry_max
/ 10);
1287 * Subtract out pages already installed in the zone (hack)
1289 entry_max
= pv_entry_max
- vm_page_array_size
;
1293 zinitna(pvzone
, NULL
, 0, entry_max
, ZONE_INTERRUPT
);
1296 * Enable dynamic deletion of empty higher-level page table pages
1297 * by default only if system memory is < 8GB (use 7GB for slop).
1298 * This can save a little memory, but imposes significant
1299 * performance overhead for things like bulk builds, and for programs
1300 * which do a lot of memory mapping and memory unmapping.
1302 if (pmap_dynamic_delete
< 0) {
1303 if (vmstats
.v_page_count
< 7LL * 1024 * 1024 * 1024 / PAGE_SIZE
)
1304 pmap_dynamic_delete
= 1;
1306 pmap_dynamic_delete
= 0;
1311 * Typically used to initialize a fictitious page by vm/device_pager.c
1314 pmap_page_init(struct vm_page
*m
)
1317 TAILQ_INIT(&m
->md
.pv_list
);
1320 /***************************************************
1321 * Low level helper routines.....
1322 ***************************************************/
1325 * this routine defines the region(s) of memory that should
1326 * not be tested for the modified bit.
1330 pmap_track_modified(vm_pindex_t pindex
)
1332 vm_offset_t va
= (vm_offset_t
)pindex
<< PAGE_SHIFT
;
1333 if ((va
< clean_sva
) || (va
>= clean_eva
))
1340 * Extract the physical page address associated with the map/VA pair.
1341 * The page must be wired for this to work reliably.
1344 pmap_extract(pmap_t pmap
, vm_offset_t va
, void **handlep
)
1351 if (va
>= VM_MAX_USER_ADDRESS
) {
1353 * Kernel page directories might be direct-mapped and
1354 * there is typically no PV tracking of pte's
1358 pt
= pmap_pt(pmap
, va
);
1359 if (pt
&& (*pt
& pmap
->pmap_bits
[PG_V_IDX
])) {
1360 if (*pt
& pmap
->pmap_bits
[PG_PS_IDX
]) {
1361 rtval
= *pt
& PG_PS_FRAME
;
1362 rtval
|= va
& PDRMASK
;
1364 ptep
= pmap_pt_to_pte(*pt
, va
);
1365 if (*pt
& pmap
->pmap_bits
[PG_V_IDX
]) {
1366 rtval
= *ptep
& PG_FRAME
;
1367 rtval
|= va
& PAGE_MASK
;
1375 * User pages currently do not direct-map the page directory
1376 * and some pages might not used managed PVs. But all PT's
1379 pt_pv
= pv_get(pmap
, pmap_pt_pindex(va
), NULL
);
1381 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
1382 if (*ptep
& pmap
->pmap_bits
[PG_V_IDX
]) {
1383 rtval
= *ptep
& PG_FRAME
;
1384 rtval
|= va
& PAGE_MASK
;
1387 *handlep
= pt_pv
; /* locked until done */
1390 } else if (handlep
) {
1398 pmap_extract_done(void *handle
)
1401 pv_put((pv_entry_t
)handle
);
1405 * Similar to extract but checks protections, SMP-friendly short-cut for
1406 * vm_fault_page[_quick](). Can return NULL to cause the caller to
1407 * fall-through to the real fault code. Does not work with HVM page
1410 * if busyp is NULL the returned page, if not NULL, is held (and not busied).
1412 * If busyp is not NULL and this function sets *busyp non-zero, the returned
1413 * page is busied (and not held).
1415 * If busyp is not NULL and this function sets *busyp to zero, the returned
1416 * page is held (and not busied).
1418 * If VM_PROT_WRITE or VM_PROT_OVERRIDE_WRITE is set in prot, and the pte
1419 * is already writable, the returned page will be dirtied. If the pte
1420 * is not already writable NULL is returned. In otherwords, if either
1421 * bit is set and a vm_page_t is returned, any COW will already have happened
1422 * and that page can be written by the caller.
1424 * WARNING! THE RETURNED PAGE IS ONLY HELD AND NOT SUITABLE FOR READING
1428 pmap_fault_page_quick(pmap_t pmap
, vm_offset_t va
, vm_prot_t prot
, int *busyp
)
1431 va
< VM_MAX_USER_ADDRESS
&&
1432 (pmap
->pm_flags
& PMAP_HVM
) == 0) {
1440 req
= pmap
->pmap_bits
[PG_V_IDX
] |
1441 pmap
->pmap_bits
[PG_U_IDX
];
1442 if (prot
& (VM_PROT_WRITE
| VM_PROT_OVERRIDE_WRITE
))
1443 req
|= pmap
->pmap_bits
[PG_RW_IDX
];
1445 pt_pv
= pv_get(pmap
, pmap_pt_pindex(va
), NULL
);
1448 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
1449 if ((*ptep
& req
) != req
) {
1453 pte_pv
= pv_get_try(pmap
, pmap_pte_pindex(va
), NULL
, &error
);
1454 if (pte_pv
&& error
== 0) {
1456 if (prot
& (VM_PROT_WRITE
| VM_PROT_OVERRIDE_WRITE
)) {
1457 /* interlocked by presence of pv_entry */
1461 if (prot
& VM_PROT_WRITE
) {
1462 if (vm_page_busy_try(m
, TRUE
))
1473 } else if (pte_pv
) {
1477 /* error, since we didn't request a placemarker */
1488 * Extract the physical page address associated kernel virtual address.
1491 pmap_kextract(vm_offset_t va
)
1493 pd_entry_t pt
; /* pt entry in pd */
1496 if (va
>= DMAP_MIN_ADDRESS
&& va
< DMAP_MAX_ADDRESS
) {
1497 pa
= DMAP_TO_PHYS(va
);
1500 if (pt
& kernel_pmap
.pmap_bits
[PG_PS_IDX
]) {
1501 pa
= (pt
& PG_PS_FRAME
) | (va
& PDRMASK
);
1504 * Beware of a concurrent promotion that changes the
1505 * PDE at this point! For example, vtopte() must not
1506 * be used to access the PTE because it would use the
1507 * new PDE. It is, however, safe to use the old PDE
1508 * because the page table page is preserved by the
1511 pa
= *pmap_pt_to_pte(pt
, va
);
1512 pa
= (pa
& PG_FRAME
) | (va
& PAGE_MASK
);
1518 /***************************************************
1519 * Low level mapping routines.....
1520 ***************************************************/
1523 * Routine: pmap_kenter
1525 * Add a wired page to the KVA
1526 * NOTE! note that in order for the mapping to take effect -- you
1527 * should do an invltlb after doing the pmap_kenter().
1530 pmap_kenter(vm_offset_t va
, vm_paddr_t pa
)
1536 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
1537 kernel_pmap
.pmap_bits
[PG_V_IDX
];
1541 pmap_inval_smp(&kernel_pmap
, va
, 1, ptep
, npte
);
1545 pmap_inval_smp(&kernel_pmap
, va
, ptep
, npte
);
1552 * Similar to pmap_kenter(), except we only invalidate the mapping on the
1553 * current CPU. Returns 0 if the previous pte was 0, 1 if it wasn't
1554 * (caller can conditionalize calling smp_invltlb()).
1557 pmap_kenter_quick(vm_offset_t va
, vm_paddr_t pa
)
1563 npte
= pa
| kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
1564 kernel_pmap
.pmap_bits
[PG_V_IDX
];
1573 atomic_swap_long(ptep
, npte
);
1574 cpu_invlpg((void *)va
);
1580 * Enter addresses into the kernel pmap but don't bother
1581 * doing any tlb invalidations. Caller will do a rollup
1582 * invalidation via pmap_rollup_inval().
1585 pmap_kenter_noinval(vm_offset_t va
, vm_paddr_t pa
)
1592 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
1593 kernel_pmap
.pmap_bits
[PG_V_IDX
];
1602 atomic_swap_long(ptep
, npte
);
1603 cpu_invlpg((void *)va
);
1609 * remove a page from the kernel pagetables
1612 pmap_kremove(vm_offset_t va
)
1617 pmap_inval_smp(&kernel_pmap
, va
, 1, ptep
, 0);
1621 pmap_kremove_quick(vm_offset_t va
)
1626 (void)pte_load_clear(ptep
);
1627 cpu_invlpg((void *)va
);
1631 * Remove addresses from the kernel pmap but don't bother
1632 * doing any tlb invalidations. Caller will do a rollup
1633 * invalidation via pmap_rollup_inval().
1636 pmap_kremove_noinval(vm_offset_t va
)
1641 (void)pte_load_clear(ptep
);
1645 * XXX these need to be recoded. They are not used in any critical path.
1648 pmap_kmodify_rw(vm_offset_t va
)
1650 atomic_set_long(vtopte(va
), kernel_pmap
.pmap_bits
[PG_RW_IDX
]);
1651 cpu_invlpg((void *)va
);
1656 pmap_kmodify_nc(vm_offset_t va)
1658 atomic_set_long(vtopte(va), PG_N);
1659 cpu_invlpg((void *)va);
1664 * Used to map a range of physical addresses into kernel virtual
1665 * address space during the low level boot, typically to map the
1666 * dump bitmap, message buffer, and vm_page_array.
1668 * These mappings are typically made at some pointer after the end of the
1671 * We could return PHYS_TO_DMAP(start) here and not allocate any
1672 * via (*virtp), but then kmem from userland and kernel dumps won't
1673 * have access to the related pointers.
1676 pmap_map(vm_offset_t
*virtp
, vm_paddr_t start
, vm_paddr_t end
, int prot
)
1679 vm_offset_t va_start
;
1681 /*return PHYS_TO_DMAP(start);*/
1686 while (start
< end
) {
1687 pmap_kenter_quick(va
, start
);
1695 #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024)
1698 * Remove the specified set of pages from the data and instruction caches.
1700 * In contrast to pmap_invalidate_cache_range(), this function does not
1701 * rely on the CPU's self-snoop feature, because it is intended for use
1702 * when moving pages into a different cache domain.
1705 pmap_invalidate_cache_pages(vm_page_t
*pages
, int count
)
1707 vm_offset_t daddr
, eva
;
1710 if (count
>= PMAP_CLFLUSH_THRESHOLD
/ PAGE_SIZE
||
1711 (cpu_feature
& CPUID_CLFSH
) == 0)
1715 for (i
= 0; i
< count
; i
++) {
1716 daddr
= PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages
[i
]));
1717 eva
= daddr
+ PAGE_SIZE
;
1718 for (; daddr
< eva
; daddr
+= cpu_clflush_line_size
)
1726 pmap_invalidate_cache_range(vm_offset_t sva
, vm_offset_t eva
)
1728 KASSERT((sva
& PAGE_MASK
) == 0,
1729 ("pmap_invalidate_cache_range: sva not page-aligned"));
1730 KASSERT((eva
& PAGE_MASK
) == 0,
1731 ("pmap_invalidate_cache_range: eva not page-aligned"));
1733 if (cpu_feature
& CPUID_SS
) {
1734 ; /* If "Self Snoop" is supported, do nothing. */
1736 /* Globally invalidate caches */
1737 cpu_wbinvd_on_all_cpus();
1742 * Invalidate the specified range of virtual memory on all cpus associated
1746 pmap_invalidate_range(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
)
1748 pmap_inval_smp(pmap
, sva
, (eva
- sva
) >> PAGE_SHIFT
, NULL
, 0);
1752 * Add a list of wired pages to the kva. This routine is used for temporary
1753 * kernel mappings such as those found in buffer cache buffer. Page
1754 * modifications and accesses are not tracked or recorded.
1756 * NOTE! Old mappings are simply overwritten, and we cannot assume relaxed
1757 * semantics as previous mappings may have been zerod without any
1760 * The page *must* be wired.
1762 static __inline
void
1763 _pmap_qenter(vm_offset_t beg_va
, vm_page_t
*m
, int count
, int doinval
)
1768 end_va
= beg_va
+ count
* PAGE_SIZE
;
1770 for (va
= beg_va
; va
< end_va
; va
+= PAGE_SIZE
) {
1775 pte
= VM_PAGE_TO_PHYS(*m
) |
1776 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
1777 kernel_pmap
.pmap_bits
[PG_V_IDX
] |
1778 kernel_pmap
.pmap_cache_bits
[(*m
)->pat_mode
];
1780 atomic_swap_long(ptep
, pte
);
1784 pmap_invalidate_range(&kernel_pmap
, beg_va
, end_va
);
1788 pmap_qenter(vm_offset_t beg_va
, vm_page_t
*m
, int count
)
1790 _pmap_qenter(beg_va
, m
, count
, 1);
1794 pmap_qenter_noinval(vm_offset_t beg_va
, vm_page_t
*m
, int count
)
1796 _pmap_qenter(beg_va
, m
, count
, 0);
1800 * This routine jerks page mappings from the kernel -- it is meant only
1801 * for temporary mappings such as those found in buffer cache buffers.
1802 * No recording modified or access status occurs.
1804 * MPSAFE, INTERRUPT SAFE (cluster callback)
1807 pmap_qremove(vm_offset_t beg_va
, int count
)
1812 end_va
= beg_va
+ count
* PAGE_SIZE
;
1814 for (va
= beg_va
; va
< end_va
; va
+= PAGE_SIZE
) {
1818 (void)pte_load_clear(pte
);
1819 cpu_invlpg((void *)va
);
1821 pmap_invalidate_range(&kernel_pmap
, beg_va
, end_va
);
1825 * This routine removes temporary kernel mappings, only invalidating them
1826 * on the current cpu. It should only be used under carefully controlled
1830 pmap_qremove_quick(vm_offset_t beg_va
, int count
)
1835 end_va
= beg_va
+ count
* PAGE_SIZE
;
1837 for (va
= beg_va
; va
< end_va
; va
+= PAGE_SIZE
) {
1841 (void)pte_load_clear(pte
);
1842 cpu_invlpg((void *)va
);
1847 * This routine removes temporary kernel mappings *without* invalidating
1848 * the TLB. It can only be used on permanent kva reservations such as those
1849 * found in buffer cache buffers, under carefully controlled circumstances.
1851 * NOTE: Repopulating these KVAs requires unconditional invalidation.
1852 * (pmap_qenter() does unconditional invalidation).
1855 pmap_qremove_noinval(vm_offset_t beg_va
, int count
)
1860 end_va
= beg_va
+ count
* PAGE_SIZE
;
1862 for (va
= beg_va
; va
< end_va
; va
+= PAGE_SIZE
) {
1866 (void)pte_load_clear(pte
);
1871 * Create a new thread and optionally associate it with a (new) process.
1872 * NOTE! the new thread's cpu may not equal the current cpu.
1875 pmap_init_thread(thread_t td
)
1877 /* enforce pcb placement & alignment */
1878 td
->td_pcb
= (struct pcb
*)(td
->td_kstack
+ td
->td_kstack_size
) - 1;
1879 td
->td_pcb
= (struct pcb
*)((intptr_t)td
->td_pcb
& ~(intptr_t)0xF);
1880 td
->td_savefpu
= &td
->td_pcb
->pcb_save
;
1881 td
->td_sp
= (char *)td
->td_pcb
; /* no -16 */
1885 * This routine directly affects the fork perf for a process.
1888 pmap_init_proc(struct proc
*p
)
1893 pmap_pinit_defaults(struct pmap
*pmap
)
1895 bcopy(pmap_bits_default
, pmap
->pmap_bits
,
1896 sizeof(pmap_bits_default
));
1897 bcopy(protection_codes
, pmap
->protection_codes
,
1898 sizeof(protection_codes
));
1899 bcopy(pat_pte_index
, pmap
->pmap_cache_bits
,
1900 sizeof(pat_pte_index
));
1901 pmap
->pmap_cache_mask
= X86_PG_NC_PWT
| X86_PG_NC_PCD
| X86_PG_PTE_PAT
;
1902 pmap
->copyinstr
= std_copyinstr
;
1903 pmap
->copyin
= std_copyin
;
1904 pmap
->copyout
= std_copyout
;
1905 pmap
->fubyte
= std_fubyte
;
1906 pmap
->subyte
= std_subyte
;
1907 pmap
->fuword32
= std_fuword32
;
1908 pmap
->fuword64
= std_fuword64
;
1909 pmap
->suword32
= std_suword32
;
1910 pmap
->suword64
= std_suword64
;
1911 pmap
->swapu32
= std_swapu32
;
1912 pmap
->swapu64
= std_swapu64
;
1915 * Initialize pmap0/vmspace0.
1917 * On architectures where the kernel pmap is not integrated into the user
1918 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1919 * kernel_pmap should be used to directly access the kernel_pmap.
1922 pmap_pinit0(struct pmap
*pmap
)
1926 pmap
->pm_pml4
= (pml4_entry_t
*)(PTOV_OFFSET
+ KPML4phys
);
1928 CPUMASK_ASSZERO(pmap
->pm_active
);
1929 pmap
->pm_pvhint_pt
= NULL
;
1930 pmap
->pm_pvhint_pte
= NULL
;
1931 RB_INIT(&pmap
->pm_pvroot
);
1932 spin_init(&pmap
->pm_spin
, "pmapinit0");
1933 for (i
= 0; i
< PM_PLACEMARKS
; ++i
)
1934 pmap
->pm_placemarks
[i
] = PM_NOPLACEMARK
;
1935 bzero(&pmap
->pm_stats
, sizeof pmap
->pm_stats
);
1936 pmap_pinit_defaults(pmap
);
1940 * Initialize a preallocated and zeroed pmap structure,
1941 * such as one in a vmspace structure.
1944 pmap_pinit_simple(struct pmap
*pmap
)
1949 * Misc initialization
1952 CPUMASK_ASSZERO(pmap
->pm_active
);
1953 pmap
->pm_pvhint_pt
= NULL
;
1954 pmap
->pm_pvhint_pte
= NULL
;
1955 pmap
->pm_flags
= PMAP_FLAG_SIMPLE
;
1957 pmap_pinit_defaults(pmap
);
1960 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1963 if (pmap
->pm_pmlpv
== NULL
) {
1964 RB_INIT(&pmap
->pm_pvroot
);
1965 bzero(&pmap
->pm_stats
, sizeof pmap
->pm_stats
);
1966 spin_init(&pmap
->pm_spin
, "pmapinitsimple");
1967 for (i
= 0; i
< PM_PLACEMARKS
; ++i
)
1968 pmap
->pm_placemarks
[i
] = PM_NOPLACEMARK
;
1973 pmap_pinit(struct pmap
*pmap
)
1978 if (pmap
->pm_pmlpv
) {
1979 if (pmap
->pmap_bits
[TYPE_IDX
] != REGULAR_PMAP
) {
1984 pmap_pinit_simple(pmap
);
1985 pmap
->pm_flags
&= ~PMAP_FLAG_SIMPLE
;
1988 * No need to allocate page table space yet but we do need a valid
1989 * page directory table.
1991 if (pmap
->pm_pml4
== NULL
) {
1993 (pml4_entry_t
*)kmem_alloc_pageable(&kernel_map
,
1999 * Allocate the page directory page, which wires it even though
2000 * it isn't being entered into some higher level page table (it
2001 * being the highest level). If one is already cached we don't
2002 * have to do anything.
2004 if ((pv
= pmap
->pm_pmlpv
) == NULL
) {
2005 pv
= pmap_allocpte(pmap
, pmap_pml4_pindex(), NULL
);
2006 pmap
->pm_pmlpv
= pv
;
2007 pmap_kenter((vm_offset_t
)pmap
->pm_pml4
,
2008 VM_PAGE_TO_PHYS(pv
->pv_m
));
2012 * Install DMAP and KMAP.
2014 for (j
= 0; j
< NDMPML4E
; ++j
) {
2015 pmap
->pm_pml4
[DMPML4I
+ j
] =
2016 (DMPDPphys
+ ((vm_paddr_t
)j
<< PAGE_SHIFT
)) |
2017 pmap
->pmap_bits
[PG_RW_IDX
] |
2018 pmap
->pmap_bits
[PG_V_IDX
] |
2019 pmap
->pmap_bits
[PG_U_IDX
];
2021 pmap
->pm_pml4
[KPML4I
] = KPDPphys
|
2022 pmap
->pmap_bits
[PG_RW_IDX
] |
2023 pmap
->pmap_bits
[PG_V_IDX
] |
2024 pmap
->pmap_bits
[PG_U_IDX
];
2027 * install self-referential address mapping entry
2029 pmap
->pm_pml4
[PML4PML4I
] = VM_PAGE_TO_PHYS(pv
->pv_m
) |
2030 pmap
->pmap_bits
[PG_V_IDX
] |
2031 pmap
->pmap_bits
[PG_RW_IDX
] |
2032 pmap
->pmap_bits
[PG_A_IDX
] |
2033 pmap
->pmap_bits
[PG_M_IDX
];
2035 KKASSERT(pv
->pv_m
->flags
& PG_MAPPED
);
2036 KKASSERT(pv
->pv_m
->flags
& PG_WRITEABLE
);
2038 KKASSERT(pmap
->pm_pml4
[255] == 0);
2039 KKASSERT(RB_ROOT(&pmap
->pm_pvroot
) == pv
);
2040 KKASSERT(pv
->pv_entry
.rbe_left
== NULL
);
2041 KKASSERT(pv
->pv_entry
.rbe_right
== NULL
);
2045 * Clean up a pmap structure so it can be physically freed. This routine
2046 * is called by the vmspace dtor function. A great deal of pmap data is
2047 * left passively mapped to improve vmspace management so we have a bit
2048 * of cleanup work to do here.
2051 pmap_puninit(pmap_t pmap
)
2056 KKASSERT(CPUMASK_TESTZERO(pmap
->pm_active
));
2057 if ((pv
= pmap
->pm_pmlpv
) != NULL
) {
2058 if (pv_hold_try(pv
) == 0)
2060 KKASSERT(pv
== pmap
->pm_pmlpv
);
2061 p
= pmap_remove_pv_page(pv
);
2063 pv
= NULL
; /* safety */
2064 pmap_kremove((vm_offset_t
)pmap
->pm_pml4
);
2065 vm_page_busy_wait(p
, FALSE
, "pgpun");
2066 KKASSERT(p
->flags
& (PG_FICTITIOUS
|PG_UNMANAGED
));
2067 vm_page_unwire(p
, 0);
2068 vm_page_flag_clear(p
, PG_MAPPED
| PG_WRITEABLE
);
2071 * XXX eventually clean out PML4 static entries and
2072 * use vm_page_free_zero()
2075 pmap
->pm_pmlpv
= NULL
;
2077 if (pmap
->pm_pml4
) {
2078 KKASSERT(pmap
->pm_pml4
!= (void *)(PTOV_OFFSET
+ KPML4phys
));
2079 kmem_free(&kernel_map
, (vm_offset_t
)pmap
->pm_pml4
, PAGE_SIZE
);
2080 pmap
->pm_pml4
= NULL
;
2082 KKASSERT(pmap
->pm_stats
.resident_count
== 0);
2083 KKASSERT(pmap
->pm_stats
.wired_count
== 0);
2087 * This function is now unused (used to add the pmap to the pmap_list)
2090 pmap_pinit2(struct pmap
*pmap
)
2095 * This routine is called when various levels in the page table need to
2096 * be populated. This routine cannot fail.
2098 * This function returns two locked pv_entry's, one representing the
2099 * requested pv and one representing the requested pv's parent pv. If
2100 * an intermediate page table does not exist it will be created, mapped,
2101 * wired, and the parent page table will be given an additional hold
2102 * count representing the presence of the child pv_entry.
2106 pmap_allocpte(pmap_t pmap
, vm_pindex_t ptepindex
, pv_entry_t
*pvpp
)
2112 vm_pindex_t pt_pindex
;
2118 * If the pv already exists and we aren't being asked for the
2119 * parent page table page we can just return it. A locked+held pv
2120 * is returned. The pv will also have a second hold related to the
2121 * pmap association that we don't have to worry about.
2124 pv
= pv_alloc(pmap
, ptepindex
, &isnew
);
2125 if (isnew
== 0 && pvpp
== NULL
)
2129 * Special case terminal PVs. These are not page table pages so
2130 * no vm_page is allocated (the caller supplied the vm_page). If
2131 * pvpp is non-NULL we are being asked to also removed the pt_pv
2134 * Note that pt_pv's are only returned for user VAs. We assert that
2135 * a pt_pv is not being requested for kernel VAs. The kernel
2136 * pre-wires all higher-level page tables so don't overload managed
2137 * higher-level page tables on top of it!
2139 if (ptepindex
< pmap_pt_pindex(0)) {
2140 if (ptepindex
>= NUPTE_USER
) {
2141 /* kernel manages this manually for KVM */
2142 KKASSERT(pvpp
== NULL
);
2144 KKASSERT(pvpp
!= NULL
);
2145 pt_pindex
= NUPTE_TOTAL
+ (ptepindex
>> NPTEPGSHIFT
);
2146 pvp
= pmap_allocpte(pmap
, pt_pindex
, NULL
);
2148 vm_page_wire_quick(pvp
->pv_m
);
2155 * The kernel never uses managed PT/PD/PDP pages.
2157 KKASSERT(pmap
!= &kernel_pmap
);
2160 * Non-terminal PVs allocate a VM page to represent the page table,
2161 * so we have to resolve pvp and calculate ptepindex for the pvp
2162 * and then for the page table entry index in the pvp for
2165 if (ptepindex
< pmap_pd_pindex(0)) {
2167 * pv is PT, pvp is PD
2169 ptepindex
= (ptepindex
- pmap_pt_pindex(0)) >> NPDEPGSHIFT
;
2170 ptepindex
+= NUPTE_TOTAL
+ NUPT_TOTAL
;
2171 pvp
= pmap_allocpte(pmap
, ptepindex
, NULL
);
2176 ptepindex
= pv
->pv_pindex
- pmap_pt_pindex(0);
2177 ptepindex
&= ((1ul << NPDEPGSHIFT
) - 1);
2179 } else if (ptepindex
< pmap_pdp_pindex(0)) {
2181 * pv is PD, pvp is PDP
2183 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
2186 ptepindex
= (ptepindex
- pmap_pd_pindex(0)) >> NPDPEPGSHIFT
;
2187 ptepindex
+= NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
;
2189 if (pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) {
2190 KKASSERT(pvpp
== NULL
);
2193 pvp
= pmap_allocpte(pmap
, ptepindex
, NULL
);
2199 ptepindex
= pv
->pv_pindex
- pmap_pd_pindex(0);
2200 ptepindex
&= ((1ul << NPDPEPGSHIFT
) - 1);
2201 } else if (ptepindex
< pmap_pml4_pindex()) {
2203 * pv is PDP, pvp is the root pml4 table
2205 pvp
= pmap_allocpte(pmap
, pmap_pml4_pindex(), NULL
);
2210 ptepindex
= pv
->pv_pindex
- pmap_pdp_pindex(0);
2211 ptepindex
&= ((1ul << NPML4EPGSHIFT
) - 1);
2214 * pv represents the top-level PML4, there is no parent.
2223 * (isnew) is TRUE, pv is not terminal.
2225 * (1) Add a wire count to the parent page table (pvp).
2226 * (2) Allocate a VM page for the page table.
2227 * (3) Enter the VM page into the parent page table.
2229 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
2232 vm_page_wire_quick(pvp
->pv_m
);
2235 m
= vm_page_alloc(NULL
, pv
->pv_pindex
,
2236 VM_ALLOC_NORMAL
| VM_ALLOC_SYSTEM
|
2237 VM_ALLOC_INTERRUPT
);
2242 vm_page_wire(m
); /* wire for mapping in parent */
2243 vm_page_unmanage(m
); /* m must be spinunlocked */
2244 pmap_zero_page(VM_PAGE_TO_PHYS(m
));
2245 m
->valid
= VM_PAGE_BITS_ALL
;
2247 vm_page_spin_lock(m
);
2248 pmap_page_stats_adding(m
);
2249 TAILQ_INSERT_TAIL(&m
->md
.pv_list
, pv
, pv_list
);
2251 vm_page_flag_set(m
, PG_MAPPED
| PG_WRITEABLE
);
2252 vm_page_spin_unlock(m
);
2255 * (isnew) is TRUE, pv is not terminal.
2257 * Wire the page into pvp. Bump the resident_count for the pmap.
2258 * There is no pvp for the top level, address the pm_pml4[] array
2261 * If the caller wants the parent we return it, otherwise
2262 * we just put it away.
2264 * No interlock is needed for pte 0 -> non-zero.
2266 * In the situation where *ptep is valid we might have an unmanaged
2267 * page table page shared from another page table which we need to
2268 * unshare before installing our private page table page.
2271 v
= VM_PAGE_TO_PHYS(m
) |
2272 (pmap
->pmap_bits
[PG_U_IDX
] |
2273 pmap
->pmap_bits
[PG_RW_IDX
] |
2274 pmap
->pmap_bits
[PG_V_IDX
] |
2275 pmap
->pmap_bits
[PG_A_IDX
] |
2276 pmap
->pmap_bits
[PG_M_IDX
]);
2277 ptep
= pv_pte_lookup(pvp
, ptepindex
);
2278 if (*ptep
& pmap
->pmap_bits
[PG_V_IDX
]) {
2282 panic("pmap_allocpte: unexpected pte %p/%d",
2283 pvp
, (int)ptepindex
);
2285 pte
= pmap_inval_smp(pmap
, (vm_offset_t
)-1, 1, ptep
, v
);
2286 if (vm_page_unwire_quick(
2287 PHYS_TO_VM_PAGE(pte
& PG_FRAME
))) {
2288 panic("pmap_allocpte: shared pgtable "
2289 "pg bad wirecount");
2294 pte
= atomic_swap_long(ptep
, v
);
2296 kprintf("install pgtbl mixup 0x%016jx "
2297 "old/new 0x%016jx/0x%016jx\n",
2298 (intmax_t)ptepindex
, pte
, v
);
2305 * (isnew) may be TRUE or FALSE, pv may or may not be terminal.
2309 KKASSERT(pvp
->pv_m
!= NULL
);
2310 ptep
= pv_pte_lookup(pvp
, ptepindex
);
2311 v
= VM_PAGE_TO_PHYS(pv
->pv_m
) |
2312 (pmap
->pmap_bits
[PG_U_IDX
] |
2313 pmap
->pmap_bits
[PG_RW_IDX
] |
2314 pmap
->pmap_bits
[PG_V_IDX
] |
2315 pmap
->pmap_bits
[PG_A_IDX
] |
2316 pmap
->pmap_bits
[PG_M_IDX
]);
2318 kprintf("mismatched upper level pt %016jx/%016jx\n",
2330 * This version of pmap_allocpte() checks for possible segment optimizations
2331 * that would allow page-table sharing. It can be called for terminal
2332 * page or page table page ptepindex's.
2334 * The function is called with page table page ptepindex's for fictitious
2335 * and unmanaged terminal pages. That is, we don't want to allocate a
2336 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
2339 * This function can return a pv and *pvpp associated with the passed in pmap
2340 * OR a pv and *pvpp associated with the shared pmap. In the latter case
2341 * an unmanaged page table page will be entered into the pass in pmap.
2345 pmap_allocpte_seg(pmap_t pmap
, vm_pindex_t ptepindex
, pv_entry_t
*pvpp
,
2346 vm_map_entry_t entry
, vm_offset_t va
)
2351 vm_pindex_t
*pt_placemark
;
2353 pv_entry_t pte_pv
; /* in original or shared pmap */
2354 pv_entry_t pt_pv
; /* in original or shared pmap */
2355 pv_entry_t proc_pd_pv
; /* in original pmap */
2356 pv_entry_t proc_pt_pv
; /* in original pmap */
2357 pv_entry_t xpv
; /* PT in shared pmap */
2358 pd_entry_t
*pt
; /* PT entry in PD of original pmap */
2359 pd_entry_t opte
; /* contents of *pt */
2360 pd_entry_t npte
; /* contents of *pt */
2365 * Basic tests, require a non-NULL vm_map_entry, require proper
2366 * alignment and type for the vm_map_entry, require that the
2367 * underlying object already be allocated.
2369 * We allow almost any type of object to use this optimization.
2370 * The object itself does NOT have to be sized to a multiple of the
2371 * segment size, but the memory mapping does.
2373 * XXX don't handle devices currently, because VM_PAGE_TO_PHYS()
2374 * won't work as expected.
2376 if (entry
== NULL
||
2377 pmap_mmu_optimize
== 0 || /* not enabled */
2378 (pmap
->pm_flags
& PMAP_HVM
) || /* special pmap */
2379 ptepindex
>= pmap_pd_pindex(0) || /* not terminal or pt */
2380 entry
->inheritance
!= VM_INHERIT_SHARE
|| /* not shared */
2381 entry
->maptype
!= VM_MAPTYPE_NORMAL
|| /* weird map type */
2382 entry
->object
.vm_object
== NULL
|| /* needs VM object */
2383 entry
->object
.vm_object
->type
== OBJT_DEVICE
|| /* ick */
2384 entry
->object
.vm_object
->type
== OBJT_MGTDEVICE
|| /* ick */
2385 (entry
->offset
& SEG_MASK
) || /* must be aligned */
2386 (entry
->start
& SEG_MASK
)) {
2387 return(pmap_allocpte(pmap
, ptepindex
, pvpp
));
2391 * Make sure the full segment can be represented.
2393 b
= va
& ~(vm_offset_t
)SEG_MASK
;
2394 if (b
< entry
->start
|| b
+ SEG_SIZE
> entry
->end
)
2395 return(pmap_allocpte(pmap
, ptepindex
, pvpp
));
2398 * If the full segment can be represented dive the VM object's
2399 * shared pmap, allocating as required.
2401 object
= entry
->object
.vm_object
;
2403 if (entry
->protection
& VM_PROT_WRITE
)
2404 obpmapp
= &object
->md
.pmap_rw
;
2406 obpmapp
= &object
->md
.pmap_ro
;
2409 if (pmap_enter_debug
> 0) {
2411 kprintf("pmap_allocpte_seg: va=%jx prot %08x o=%p "
2413 va
, entry
->protection
, object
,
2415 kprintf("pmap_allocpte_seg: entry %p %jx-%jx\n",
2416 entry
, entry
->start
, entry
->end
);
2421 * We allocate what appears to be a normal pmap but because portions
2422 * of this pmap are shared with other unrelated pmaps we have to
2423 * set pm_active to point to all cpus.
2425 * XXX Currently using pmap_spin to interlock the update, can't use
2426 * vm_object_hold/drop because the token might already be held
2427 * shared OR exclusive and we don't know.
2429 while ((obpmap
= *obpmapp
) == NULL
) {
2430 obpmap
= kmalloc(sizeof(*obpmap
), M_OBJPMAP
, M_WAITOK
|M_ZERO
);
2431 pmap_pinit_simple(obpmap
);
2432 pmap_pinit2(obpmap
);
2433 spin_lock(&pmap_spin
);
2434 if (*obpmapp
!= NULL
) {
2438 spin_unlock(&pmap_spin
);
2439 pmap_release(obpmap
);
2440 pmap_puninit(obpmap
);
2441 kfree(obpmap
, M_OBJPMAP
);
2442 obpmap
= *obpmapp
; /* safety */
2444 obpmap
->pm_active
= smp_active_mask
;
2445 obpmap
->pm_flags
|= PMAP_SEGSHARED
;
2447 spin_unlock(&pmap_spin
);
2452 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
2453 * pte/pt using the shared pmap from the object but also adjust
2454 * the process pmap's page table page as a side effect.
2458 * Resolve the terminal PTE and PT in the shared pmap. This is what
2459 * we will return. This is true if ptepindex represents a terminal
2460 * page, otherwise pte_pv is actually the PT and pt_pv is actually
2464 pte_pv
= pmap_allocpte(obpmap
, ptepindex
, &pt_pv
);
2467 if (ptepindex
>= pmap_pt_pindex(0))
2473 * Resolve the PD in the process pmap so we can properly share the
2474 * page table page. Lock order is bottom-up (leaf first)!
2476 * NOTE: proc_pt_pv can be NULL.
2478 proc_pt_pv
= pv_get(pmap
, pmap_pt_pindex(b
), &pt_placemark
);
2479 proc_pd_pv
= pmap_allocpte(pmap
, pmap_pd_pindex(b
), NULL
);
2481 if (pmap_enter_debug
> 0) {
2483 kprintf("proc_pt_pv %p (wc %d) pd_pv %p va=%jx\n",
2485 (proc_pt_pv
? proc_pt_pv
->pv_m
->wire_count
: -1),
2492 * xpv is the page table page pv from the shared object
2493 * (for convenience), from above.
2495 * Calculate the pte value for the PT to load into the process PD.
2496 * If we have to change it we must properly dispose of the previous
2499 pt
= pv_pte_lookup(proc_pd_pv
, pmap_pt_index(b
));
2500 npte
= VM_PAGE_TO_PHYS(xpv
->pv_m
) |
2501 (pmap
->pmap_bits
[PG_U_IDX
] |
2502 pmap
->pmap_bits
[PG_RW_IDX
] |
2503 pmap
->pmap_bits
[PG_V_IDX
] |
2504 pmap
->pmap_bits
[PG_A_IDX
] |
2505 pmap
->pmap_bits
[PG_M_IDX
]);
2508 * Dispose of previous page table page if it was local to the
2509 * process pmap. If the old pt is not empty we cannot dispose of it
2510 * until we clean it out. This case should not arise very often so
2511 * it is not optimized.
2513 * Leave pt_pv and pte_pv (in our object pmap) locked and intact
2517 pmap_inval_bulk_t bulk
;
2519 if (proc_pt_pv
->pv_m
->wire_count
!= 1) {
2521 * The page table has a bunch of stuff in it
2522 * which we have to scrap.
2524 if (softhold
== 0) {
2526 pmap_softhold(pmap
);
2531 va
& ~(vm_offset_t
)SEG_MASK
,
2532 (va
+ SEG_SIZE
) & ~(vm_offset_t
)SEG_MASK
);
2535 * The page table is empty and can be destroyed.
2536 * However, doing so leaves the pt slot unlocked,
2537 * so we have to loop-up to handle any races until
2538 * we get a NULL proc_pt_pv and a proper pt_placemark.
2540 pmap_inval_bulk_init(&bulk
, proc_pt_pv
->pv_pmap
);
2541 pmap_release_pv(proc_pt_pv
, proc_pd_pv
, &bulk
);
2542 pmap_inval_bulk_flush(&bulk
);
2549 * Handle remaining cases. We are holding pt_placemark to lock
2550 * the page table page in the primary pmap while we manipulate
2554 atomic_swap_long(pt
, npte
);
2555 vm_page_wire_quick(xpv
->pv_m
); /* shared pt -> proc */
2556 vm_page_wire_quick(proc_pd_pv
->pv_m
); /* proc pd for sh pt */
2557 atomic_add_long(&pmap
->pm_stats
.resident_count
, 1);
2558 } else if (*pt
!= npte
) {
2559 opte
= pmap_inval_smp(pmap
, (vm_offset_t
)-1, 1, pt
, npte
);
2562 opte
= pte_load_clear(pt
);
2563 KKASSERT(opte
&& opte
!= npte
);
2567 vm_page_wire_quick(xpv
->pv_m
); /* shared pt -> proc */
2570 * Clean up opte, bump the wire_count for the process
2571 * PD page representing the new entry if it was
2574 * If the entry was not previously empty and we have
2575 * a PT in the proc pmap then opte must match that
2576 * pt. The proc pt must be retired (this is done
2577 * later on in this procedure).
2579 * NOTE: replacing valid pte, wire_count on proc_pd_pv
2582 KKASSERT(opte
& pmap
->pmap_bits
[PG_V_IDX
]);
2583 m
= PHYS_TO_VM_PAGE(opte
& PG_FRAME
);
2584 if (vm_page_unwire_quick(m
)) {
2585 panic("pmap_allocpte_seg: "
2586 "bad wire count %p",
2592 pmap_softdone(pmap
);
2595 * Remove our earmark on the page table page.
2597 pv_placemarker_wakeup(pmap
, pt_placemark
);
2600 * The existing process page table was replaced and must be destroyed
2613 * Release any resources held by the given physical map.
2615 * Called when a pmap initialized by pmap_pinit is being released. Should
2616 * only be called if the map contains no valid mappings.
2618 struct pmap_release_info
{
2624 static int pmap_release_callback(pv_entry_t pv
, void *data
);
2627 pmap_release(struct pmap
*pmap
)
2629 struct pmap_release_info info
;
2631 KASSERT(CPUMASK_TESTZERO(pmap
->pm_active
),
2632 ("pmap still active! %016jx",
2633 (uintmax_t)CPUMASK_LOWMASK(pmap
->pm_active
)));
2636 * There is no longer a pmap_list, if there were we would remove the
2637 * pmap from it here.
2641 * Pull pv's off the RB tree in order from low to high and release
2649 spin_lock(&pmap
->pm_spin
);
2650 RB_SCAN(pv_entry_rb_tree
, &pmap
->pm_pvroot
, NULL
,
2651 pmap_release_callback
, &info
);
2652 spin_unlock(&pmap
->pm_spin
);
2656 } while (info
.retry
);
2660 * One resident page (the pml4 page) should remain.
2661 * No wired pages should remain.
2664 if (pmap
->pm_stats
.resident_count
!=
2665 ((pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) ? 0 : 1) ||
2666 pmap
->pm_stats
.wired_count
!= 0) {
2667 kprintf("fatal pmap problem - pmap %p flags %08x "
2668 "rescnt=%jd wirecnt=%jd\n",
2671 pmap
->pm_stats
.resident_count
,
2672 pmap
->pm_stats
.wired_count
);
2673 tsleep(pmap
, 0, "DEAD", 0);
2676 KKASSERT(pmap
->pm_stats
.resident_count
==
2677 ((pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) ? 0 : 1));
2678 KKASSERT(pmap
->pm_stats
.wired_count
== 0);
2683 * Called from low to high. We must cache the proper parent pv so we
2684 * can adjust its wired count.
2687 pmap_release_callback(pv_entry_t pv
, void *data
)
2689 struct pmap_release_info
*info
= data
;
2690 pmap_t pmap
= info
->pmap
;
2695 * Acquire a held and locked pv, check for release race
2697 pindex
= pv
->pv_pindex
;
2698 if (info
->pvp
== pv
) {
2699 spin_unlock(&pmap
->pm_spin
);
2701 } else if (pv_hold_try(pv
)) {
2702 spin_unlock(&pmap
->pm_spin
);
2704 spin_unlock(&pmap
->pm_spin
);
2708 spin_lock(&pmap
->pm_spin
);
2712 KKASSERT(pv
->pv_pmap
== pmap
&& pindex
== pv
->pv_pindex
);
2714 if (pv
->pv_pindex
< pmap_pt_pindex(0)) {
2716 * I am PTE, parent is PT
2718 pindex
= pv
->pv_pindex
>> NPTEPGSHIFT
;
2719 pindex
+= NUPTE_TOTAL
;
2720 } else if (pv
->pv_pindex
< pmap_pd_pindex(0)) {
2722 * I am PT, parent is PD
2724 pindex
= (pv
->pv_pindex
- NUPTE_TOTAL
) >> NPDEPGSHIFT
;
2725 pindex
+= NUPTE_TOTAL
+ NUPT_TOTAL
;
2726 } else if (pv
->pv_pindex
< pmap_pdp_pindex(0)) {
2728 * I am PD, parent is PDP
2730 pindex
= (pv
->pv_pindex
- NUPTE_TOTAL
- NUPT_TOTAL
) >>
2732 pindex
+= NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
;
2733 } else if (pv
->pv_pindex
< pmap_pml4_pindex()) {
2735 * I am PDP, parent is PML4 (there's only one)
2738 pindex
= (pv
->pv_pindex
- NUPTE_TOTAL
- NUPT_TOTAL
-
2739 NUPD_TOTAL
) >> NPML4EPGSHIFT
;
2740 pindex
+= NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
+ NUPDP_TOTAL
;
2742 pindex
= pmap_pml4_pindex();
2754 if (info
->pvp
&& info
->pvp
->pv_pindex
!= pindex
) {
2758 if (info
->pvp
== NULL
)
2759 info
->pvp
= pv_get(pmap
, pindex
, NULL
);
2766 r
= pmap_release_pv(pv
, info
->pvp
, NULL
);
2767 spin_lock(&pmap
->pm_spin
);
2773 * Called with held (i.e. also locked) pv. This function will dispose of
2774 * the lock along with the pv.
2776 * If the caller already holds the locked parent page table for pv it
2777 * must pass it as pvp, allowing us to avoid a deadlock, else it can
2778 * pass NULL for pvp.
2781 pmap_release_pv(pv_entry_t pv
, pv_entry_t pvp
, pmap_inval_bulk_t
*bulk
)
2786 * The pmap is currently not spinlocked, pv is held+locked.
2787 * Remove the pv's page from its parent's page table. The
2788 * parent's page table page's wire_count will be decremented.
2790 * This will clean out the pte at any level of the page table.
2791 * If smp != 0 all cpus are affected.
2793 * Do not tear-down recursively, its faster to just let the
2794 * release run its course.
2796 pmap_remove_pv_pte(pv
, pvp
, bulk
, 0);
2799 * Terminal pvs are unhooked from their vm_pages. Because
2800 * terminal pages aren't page table pages they aren't wired
2801 * by us, so we have to be sure not to unwire them either.
2803 if (pv
->pv_pindex
< pmap_pt_pindex(0)) {
2804 pmap_remove_pv_page(pv
);
2809 * We leave the top-level page table page cached, wired, and
2810 * mapped in the pmap until the dtor function (pmap_puninit())
2813 * Since we are leaving the top-level pv intact we need
2814 * to break out of what would otherwise be an infinite loop.
2816 if (pv
->pv_pindex
== pmap_pml4_pindex()) {
2822 * For page table pages (other than the top-level page),
2823 * remove and free the vm_page. The representitive mapping
2824 * removed above by pmap_remove_pv_pte() did not undo the
2825 * last wire_count so we have to do that as well.
2827 p
= pmap_remove_pv_page(pv
);
2828 vm_page_busy_wait(p
, FALSE
, "pmaprl");
2829 if (p
->wire_count
!= 1) {
2830 kprintf("p->wire_count was %016lx %d\n",
2831 pv
->pv_pindex
, p
->wire_count
);
2833 KKASSERT(p
->wire_count
== 1);
2834 KKASSERT(p
->flags
& PG_UNMANAGED
);
2836 vm_page_unwire(p
, 0);
2837 KKASSERT(p
->wire_count
== 0);
2847 * This function will remove the pte associated with a pv from its parent.
2848 * Terminal pv's are supported. All cpus specified by (bulk) are properly
2851 * The wire count will be dropped on the parent page table. The wire
2852 * count on the page being removed (pv->pv_m) from the parent page table
2853 * is NOT touched. Note that terminal pages will not have any additional
2854 * wire counts while page table pages will have at least one representing
2855 * the mapping, plus others representing sub-mappings.
2857 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2858 * pages and user page table and terminal pages.
2860 * NOTE: The pte being removed might be unmanaged, and the pv supplied might
2861 * be freshly allocated and not imply that the pte is managed. In this
2862 * case pv->pv_m should be NULL.
2864 * The pv must be locked. The pvp, if supplied, must be locked. All
2865 * supplied pv's will remain locked on return.
2867 * XXX must lock parent pv's if they exist to remove pte XXX
2871 pmap_remove_pv_pte(pv_entry_t pv
, pv_entry_t pvp
, pmap_inval_bulk_t
*bulk
,
2874 vm_pindex_t ptepindex
= pv
->pv_pindex
;
2875 pmap_t pmap
= pv
->pv_pmap
;
2881 if (ptepindex
== pmap_pml4_pindex()) {
2883 * We are the top level PML4E table, there is no parent.
2885 p
= pmap
->pm_pmlpv
->pv_m
;
2886 KKASSERT(pv
->pv_m
== p
); /* debugging */
2887 } else if (ptepindex
>= pmap_pdp_pindex(0)) {
2889 * Remove a PDP page from the PML4E. This can only occur
2890 * with user page tables. We do not have to lock the
2891 * pml4 PV so just ignore pvp.
2893 vm_pindex_t pml4_pindex
;
2894 vm_pindex_t pdp_index
;
2897 pdp_index
= ptepindex
- pmap_pdp_pindex(0);
2899 pml4_pindex
= pmap_pml4_pindex();
2900 pvp
= pv_get(pv
->pv_pmap
, pml4_pindex
, NULL
);
2905 pdp
= &pmap
->pm_pml4
[pdp_index
& ((1ul << NPML4EPGSHIFT
) - 1)];
2906 KKASSERT((*pdp
& pmap
->pmap_bits
[PG_V_IDX
]) != 0);
2907 p
= PHYS_TO_VM_PAGE(*pdp
& PG_FRAME
);
2908 pmap_inval_bulk(bulk
, (vm_offset_t
)-1, pdp
, 0);
2909 KKASSERT(pv
->pv_m
== p
); /* debugging */
2910 } else if (ptepindex
>= pmap_pd_pindex(0)) {
2912 * Remove a PD page from the PDP
2914 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2915 * of a simple pmap because it stops at
2918 vm_pindex_t pdp_pindex
;
2919 vm_pindex_t pd_index
;
2922 pd_index
= ptepindex
- pmap_pd_pindex(0);
2925 pdp_pindex
= NUPTE_TOTAL
+ NUPT_TOTAL
+ NUPD_TOTAL
+
2926 (pd_index
>> NPML4EPGSHIFT
);
2927 pvp
= pv_get(pv
->pv_pmap
, pdp_pindex
, NULL
);
2932 pd
= pv_pte_lookup(pvp
, pd_index
&
2933 ((1ul << NPDPEPGSHIFT
) - 1));
2934 KKASSERT((*pd
& pmap
->pmap_bits
[PG_V_IDX
]) != 0);
2935 p
= PHYS_TO_VM_PAGE(*pd
& PG_FRAME
);
2936 pmap_inval_bulk(bulk
, (vm_offset_t
)-1, pd
, 0);
2938 KKASSERT(pmap
->pm_flags
& PMAP_FLAG_SIMPLE
);
2939 p
= pv
->pv_m
; /* degenerate test later */
2941 KKASSERT(pv
->pv_m
== p
); /* debugging */
2942 } else if (ptepindex
>= pmap_pt_pindex(0)) {
2944 * Remove a PT page from the PD
2946 vm_pindex_t pd_pindex
;
2947 vm_pindex_t pt_index
;
2950 pt_index
= ptepindex
- pmap_pt_pindex(0);
2953 pd_pindex
= NUPTE_TOTAL
+ NUPT_TOTAL
+
2954 (pt_index
>> NPDPEPGSHIFT
);
2955 pvp
= pv_get(pv
->pv_pmap
, pd_pindex
, NULL
);
2960 pt
= pv_pte_lookup(pvp
, pt_index
& ((1ul << NPDPEPGSHIFT
) - 1));
2962 KASSERT((*pt
& pmap
->pmap_bits
[PG_V_IDX
]) != 0,
2963 ("*pt unexpectedly invalid %016jx "
2964 "gotpvp=%d ptepindex=%ld ptindex=%ld pv=%p pvp=%p",
2965 *pt
, gotpvp
, ptepindex
, pt_index
, pv
, pvp
));
2966 p
= PHYS_TO_VM_PAGE(*pt
& PG_FRAME
);
2968 if ((*pt
& pmap
->pmap_bits
[PG_V_IDX
]) == 0) {
2969 kprintf("*pt unexpectedly invalid %016jx "
2970 "gotpvp=%d ptepindex=%ld ptindex=%ld "
2972 *pt
, gotpvp
, ptepindex
, pt_index
, pv
, pvp
);
2973 tsleep(pt
, 0, "DEAD", 0);
2976 p
= PHYS_TO_VM_PAGE(*pt
& PG_FRAME
);
2979 pmap_inval_bulk(bulk
, (vm_offset_t
)-1, pt
, 0);
2980 KKASSERT(pv
->pv_m
== p
); /* debugging */
2983 * Remove a PTE from the PT page. The PV might exist even if
2984 * the PTE is not managed, in whichcase pv->pv_m should be
2987 * NOTE: Userland pmaps manage the parent PT/PD/PDP page
2988 * table pages but the kernel_pmap does not.
2990 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2991 * pv is a pte_pv so we can safely lock pt_pv.
2993 * NOTE: FICTITIOUS pages may have multiple physical mappings
2994 * so PHYS_TO_VM_PAGE() will not necessarily work for
2997 vm_pindex_t pt_pindex
;
3002 pt_pindex
= ptepindex
>> NPTEPGSHIFT
;
3003 va
= (vm_offset_t
)ptepindex
<< PAGE_SHIFT
;
3005 if (ptepindex
>= NUPTE_USER
) {
3006 ptep
= vtopte(ptepindex
<< PAGE_SHIFT
);
3007 KKASSERT(pvp
== NULL
);
3008 /* pvp remains NULL */
3011 pt_pindex
= NUPTE_TOTAL
+
3012 (ptepindex
>> NPDPEPGSHIFT
);
3013 pvp
= pv_get(pv
->pv_pmap
, pt_pindex
, NULL
);
3017 ptep
= pv_pte_lookup(pvp
, ptepindex
&
3018 ((1ul << NPDPEPGSHIFT
) - 1));
3020 pte
= pmap_inval_bulk(bulk
, va
, ptep
, 0);
3021 if (bulk
== NULL
) /* XXX */
3022 cpu_invlpg((void *)va
); /* XXX */
3025 * Now update the vm_page_t
3027 if ((pte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) &&
3028 (pte
& pmap
->pmap_bits
[PG_V_IDX
])) {
3030 * Valid managed page, adjust (p).
3032 if (pte
& pmap
->pmap_bits
[PG_DEVICE_IDX
]) {
3035 p
= PHYS_TO_VM_PAGE(pte
& PG_FRAME
);
3036 KKASSERT(pv
->pv_m
== p
);
3038 if (pte
& pmap
->pmap_bits
[PG_M_IDX
]) {
3039 if (pmap_track_modified(ptepindex
))
3042 if (pte
& pmap
->pmap_bits
[PG_A_IDX
]) {
3043 vm_page_flag_set(p
, PG_REFERENCED
);
3047 * Unmanaged page, do not try to adjust the vm_page_t.
3048 * pv could be freshly allocated for a pmap_enter(),
3049 * replacing an unmanaged page with a managed one.
3051 * pv->pv_m might reflect the new page and not the
3054 * We could extract p from the physical address and
3055 * adjust it but we explicitly do not for unmanaged
3060 if (pte
& pmap
->pmap_bits
[PG_W_IDX
])
3061 atomic_add_long(&pmap
->pm_stats
.wired_count
, -1);
3062 if (pte
& pmap
->pmap_bits
[PG_G_IDX
])
3063 cpu_invlpg((void *)va
);
3067 * If requested, scrap the underlying pv->pv_m and the underlying
3068 * pv. If this is a page-table-page we must also free the page.
3070 * pvp must be returned locked.
3074 * page table page (PT, PD, PDP, PML4), caller was responsible
3075 * for testing wired_count.
3077 KKASSERT(pv
->pv_m
->wire_count
== 1);
3078 p
= pmap_remove_pv_page(pv
);
3082 vm_page_busy_wait(p
, FALSE
, "pgpun");
3083 vm_page_unwire(p
, 0);
3084 vm_page_flag_clear(p
, PG_MAPPED
| PG_WRITEABLE
);
3086 } else if (destroy
== 2) {
3088 * Normal page, remove from pmap and leave the underlying
3091 pmap_remove_pv_page(pv
);
3093 pv
= NULL
; /* safety */
3097 * If we acquired pvp ourselves then we are responsible for
3098 * recursively deleting it.
3100 if (pvp
&& gotpvp
) {
3102 * Recursively destroy higher-level page tables.
3104 * This is optional. If we do not, they will still
3105 * be destroyed when the process exits.
3107 * NOTE: Do not destroy pv_entry's with extra hold refs,
3108 * a caller may have unlocked it and intends to
3109 * continue to use it.
3111 if (pmap_dynamic_delete
&&
3113 pvp
->pv_m
->wire_count
== 1 &&
3114 (pvp
->pv_hold
& PV_HOLD_MASK
) == 2 &&
3115 pvp
->pv_pindex
!= pmap_pml4_pindex()) {
3116 if (pmap_dynamic_delete
== 2)
3117 kprintf("A %jd %08x\n", pvp
->pv_pindex
, pvp
->pv_hold
);
3118 if (pmap
!= &kernel_pmap
) {
3119 pmap_remove_pv_pte(pvp
, NULL
, bulk
, 1);
3120 pvp
= NULL
; /* safety */
3122 kprintf("Attempt to remove kernel_pmap pindex "
3123 "%jd\n", pvp
->pv_pindex
);
3133 * Remove the vm_page association to a pv. The pv must be locked.
3137 pmap_remove_pv_page(pv_entry_t pv
)
3142 vm_page_spin_lock(m
);
3143 KKASSERT(m
&& m
== pv
->pv_m
);
3145 TAILQ_REMOVE(&m
->md
.pv_list
, pv
, pv_list
);
3146 pmap_page_stats_deleting(m
);
3147 if (TAILQ_EMPTY(&m
->md
.pv_list
))
3148 vm_page_flag_clear(m
, PG_MAPPED
| PG_WRITEABLE
);
3149 vm_page_spin_unlock(m
);
3155 * Grow the number of kernel page table entries, if needed.
3157 * This routine is always called to validate any address space
3158 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
3159 * space below KERNBASE.
3161 * kernel_map must be locked exclusively by the caller.
3164 pmap_growkernel(vm_offset_t kstart
, vm_offset_t kend
)
3167 vm_offset_t ptppaddr
;
3169 pd_entry_t
*pt
, newpt
;
3171 int update_kernel_vm_end
;
3174 * bootstrap kernel_vm_end on first real VM use
3176 if (kernel_vm_end
== 0) {
3177 kernel_vm_end
= VM_MIN_KERNEL_ADDRESS
;
3179 while ((*pmap_pt(&kernel_pmap
, kernel_vm_end
) & kernel_pmap
.pmap_bits
[PG_V_IDX
]) != 0) {
3180 kernel_vm_end
= (kernel_vm_end
+ PAGE_SIZE
* NPTEPG
) &
3181 ~(PAGE_SIZE
* NPTEPG
- 1);
3183 if (kernel_vm_end
- 1 >= kernel_map
.max_offset
) {
3184 kernel_vm_end
= kernel_map
.max_offset
;
3191 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
3192 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
3193 * do not want to force-fill 128G worth of page tables.
3195 if (kstart
< KERNBASE
) {
3196 if (kstart
> kernel_vm_end
)
3197 kstart
= kernel_vm_end
;
3198 KKASSERT(kend
<= KERNBASE
);
3199 update_kernel_vm_end
= 1;
3201 update_kernel_vm_end
= 0;
3204 kstart
= rounddown2(kstart
, PAGE_SIZE
* NPTEPG
);
3205 kend
= roundup2(kend
, PAGE_SIZE
* NPTEPG
);
3207 if (kend
- 1 >= kernel_map
.max_offset
)
3208 kend
= kernel_map
.max_offset
;
3210 while (kstart
< kend
) {
3211 pt
= pmap_pt(&kernel_pmap
, kstart
);
3213 /* We need a new PD entry */
3214 nkpg
= vm_page_alloc(NULL
, mycpu
->gd_rand_incr
++,
3217 VM_ALLOC_INTERRUPT
);
3219 panic("pmap_growkernel: no memory to grow "
3222 paddr
= VM_PAGE_TO_PHYS(nkpg
);
3223 pmap_zero_page(paddr
);
3224 newpd
= (pdp_entry_t
)
3226 kernel_pmap
.pmap_bits
[PG_V_IDX
] |
3227 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
3228 kernel_pmap
.pmap_bits
[PG_A_IDX
] |
3229 kernel_pmap
.pmap_bits
[PG_M_IDX
]);
3230 *pmap_pd(&kernel_pmap
, kstart
) = newpd
;
3231 continue; /* try again */
3233 if ((*pt
& kernel_pmap
.pmap_bits
[PG_V_IDX
]) != 0) {
3234 kstart
= (kstart
+ PAGE_SIZE
* NPTEPG
) &
3235 ~(PAGE_SIZE
* NPTEPG
- 1);
3236 if (kstart
- 1 >= kernel_map
.max_offset
) {
3237 kstart
= kernel_map
.max_offset
;
3246 * This index is bogus, but out of the way
3248 nkpg
= vm_page_alloc(NULL
, mycpu
->gd_rand_incr
++,
3251 VM_ALLOC_INTERRUPT
);
3253 panic("pmap_growkernel: no memory to grow kernel");
3256 ptppaddr
= VM_PAGE_TO_PHYS(nkpg
);
3257 pmap_zero_page(ptppaddr
);
3258 newpt
= (pd_entry_t
)(ptppaddr
|
3259 kernel_pmap
.pmap_bits
[PG_V_IDX
] |
3260 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
3261 kernel_pmap
.pmap_bits
[PG_A_IDX
] |
3262 kernel_pmap
.pmap_bits
[PG_M_IDX
]);
3263 atomic_swap_long(pmap_pt(&kernel_pmap
, kstart
), newpt
);
3265 kstart
= (kstart
+ PAGE_SIZE
* NPTEPG
) &
3266 ~(PAGE_SIZE
* NPTEPG
- 1);
3268 if (kstart
- 1 >= kernel_map
.max_offset
) {
3269 kstart
= kernel_map
.max_offset
;
3275 * Only update kernel_vm_end for areas below KERNBASE.
3277 if (update_kernel_vm_end
&& kernel_vm_end
< kstart
)
3278 kernel_vm_end
= kstart
;
3282 * Add a reference to the specified pmap.
3285 pmap_reference(pmap_t pmap
)
3288 atomic_add_int(&pmap
->pm_count
, 1);
3291 /***************************************************
3292 * page management routines.
3293 ***************************************************/
3296 * Hold a pv without locking it
3299 pv_hold(pv_entry_t pv
)
3301 atomic_add_int(&pv
->pv_hold
, 1);
3305 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
3306 * was successfully locked, FALSE if it wasn't. The caller must dispose of
3309 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
3310 * pv list via its page) must be held by the caller in order to stabilize
3314 _pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL
)
3319 * Critical path shortcut expects pv to already have one ref
3320 * (for the pv->pv_pmap).
3322 if (atomic_cmpset_int(&pv
->pv_hold
, 1, PV_HOLD_LOCKED
| 2)) {
3325 pv
->pv_line
= lineno
;
3331 count
= pv
->pv_hold
;
3333 if ((count
& PV_HOLD_LOCKED
) == 0) {
3334 if (atomic_cmpset_int(&pv
->pv_hold
, count
,
3335 (count
+ 1) | PV_HOLD_LOCKED
)) {
3338 pv
->pv_line
= lineno
;
3343 if (atomic_cmpset_int(&pv
->pv_hold
, count
, count
+ 1))
3351 * Drop a previously held pv_entry which could not be locked, allowing its
3354 * Must not be called with a spinlock held as we might zfree() the pv if it
3355 * is no longer associated with a pmap and this was the last hold count.
3358 pv_drop(pv_entry_t pv
)
3363 count
= pv
->pv_hold
;
3365 KKASSERT((count
& PV_HOLD_MASK
) > 0);
3366 KKASSERT((count
& (PV_HOLD_LOCKED
| PV_HOLD_MASK
)) !=
3367 (PV_HOLD_LOCKED
| 1));
3368 if (atomic_cmpset_int(&pv
->pv_hold
, count
, count
- 1)) {
3369 if ((count
& PV_HOLD_MASK
) == 1) {
3371 if (pmap_enter_debug
> 0) {
3373 kprintf("pv_drop: free pv %p\n", pv
);
3376 KKASSERT(count
== 1);
3377 KKASSERT(pv
->pv_pmap
== NULL
);
3387 * Find or allocate the requested PV entry, returning a locked, held pv.
3389 * If (*isnew) is non-zero, the returned pv will have two hold counts, one
3390 * for the caller and one representing the pmap and vm_page association.
3392 * If (*isnew) is zero, the returned pv will have only one hold count.
3394 * Since both associations can only be adjusted while the pv is locked,
3395 * together they represent just one additional hold.
3399 _pv_alloc(pmap_t pmap
, vm_pindex_t pindex
, int *isnew PMAP_DEBUG_DECL
)
3401 struct mdglobaldata
*md
= mdcpu
;
3409 pnew
= atomic_swap_ptr((void *)&md
->gd_newpv
, NULL
);
3412 pnew
= md
->gd_newpv
; /* might race NULL */
3413 md
->gd_newpv
= NULL
;
3418 pnew
= zalloc(pvzone
);
3420 spin_lock_shared(&pmap
->pm_spin
);
3425 pv
= pv_entry_lookup(pmap
, pindex
);
3430 * Requires exclusive pmap spinlock
3432 if (pmap_excl
== 0) {
3434 if (!spin_lock_upgrade_try(&pmap
->pm_spin
)) {
3435 spin_unlock_shared(&pmap
->pm_spin
);
3436 spin_lock(&pmap
->pm_spin
);
3442 * We need to block if someone is holding our
3443 * placemarker. As long as we determine the
3444 * placemarker has not been aquired we do not
3445 * need to get it as acquision also requires
3446 * the pmap spin lock.
3448 * However, we can race the wakeup.
3450 pmark
= pmap_placemarker_hash(pmap
, pindex
);
3452 if (((*pmark
^ pindex
) & ~PM_PLACEMARK_WAKEUP
) == 0) {
3453 atomic_set_long(pmark
, PM_PLACEMARK_WAKEUP
);
3454 tsleep_interlock(pmark
, 0);
3455 if (((*pmark
^ pindex
) &
3456 ~PM_PLACEMARK_WAKEUP
) == 0) {
3457 spin_unlock(&pmap
->pm_spin
);
3458 tsleep(pmark
, PINTERLOCKED
, "pvplc", 0);
3459 spin_lock(&pmap
->pm_spin
);
3465 * Setup the new entry
3467 pnew
->pv_pmap
= pmap
;
3468 pnew
->pv_pindex
= pindex
;
3469 pnew
->pv_hold
= PV_HOLD_LOCKED
| 2;
3471 pnew
->pv_func
= func
;
3472 pnew
->pv_line
= lineno
;
3473 if (pnew
->pv_line_lastfree
> 0) {
3474 pnew
->pv_line_lastfree
=
3475 -pnew
->pv_line_lastfree
;
3478 pv
= pv_entry_rb_tree_RB_INSERT(&pmap
->pm_pvroot
, pnew
);
3479 atomic_add_long(&pmap
->pm_stats
.resident_count
, 1);
3480 spin_unlock(&pmap
->pm_spin
);
3483 KKASSERT(pv
== NULL
);
3488 * We already have an entry, cleanup the staged pnew if
3489 * we can get the lock, otherwise block and retry.
3491 if (__predict_true(_pv_hold_try(pv PMAP_DEBUG_COPY
))) {
3493 spin_unlock(&pmap
->pm_spin
);
3495 spin_unlock_shared(&pmap
->pm_spin
);
3497 pnew
= atomic_swap_ptr((void *)&md
->gd_newpv
, pnew
);
3499 zfree(pvzone
, pnew
);
3502 if (md
->gd_newpv
== NULL
)
3503 md
->gd_newpv
= pnew
;
3505 zfree(pvzone
, pnew
);
3508 KKASSERT(pv
->pv_pmap
== pmap
&&
3509 pv
->pv_pindex
== pindex
);
3514 spin_unlock(&pmap
->pm_spin
);
3515 _pv_lock(pv PMAP_DEBUG_COPY
);
3517 spin_lock(&pmap
->pm_spin
);
3519 spin_unlock_shared(&pmap
->pm_spin
);
3520 _pv_lock(pv PMAP_DEBUG_COPY
);
3522 spin_lock_shared(&pmap
->pm_spin
);
3529 * Find the requested PV entry, returning a locked+held pv or NULL
3533 _pv_get(pmap_t pmap
, vm_pindex_t pindex
, vm_pindex_t
**pmarkp PMAP_DEBUG_DECL
)
3538 spin_lock_shared(&pmap
->pm_spin
);
3543 pv
= pv_entry_lookup(pmap
, pindex
);
3546 * Block if there is ANY placemarker. If we are to
3547 * return it, we must also aquire the spot, so we
3548 * have to block even if the placemarker is held on
3549 * a different address.
3551 * OPTIMIZATION: If pmarkp is passed as NULL the
3552 * caller is just probing (or looking for a real
3553 * pv_entry), and in this case we only need to check
3554 * to see if the placemarker matches pindex.
3559 * Requires exclusive pmap spinlock
3561 if (pmap_excl
== 0) {
3563 if (!spin_lock_upgrade_try(&pmap
->pm_spin
)) {
3564 spin_unlock_shared(&pmap
->pm_spin
);
3565 spin_lock(&pmap
->pm_spin
);
3570 pmark
= pmap_placemarker_hash(pmap
, pindex
);
3572 if ((pmarkp
&& *pmark
!= PM_NOPLACEMARK
) ||
3573 ((*pmark
^ pindex
) & ~PM_PLACEMARK_WAKEUP
) == 0) {
3574 atomic_set_long(pmark
, PM_PLACEMARK_WAKEUP
);
3575 tsleep_interlock(pmark
, 0);
3576 if ((pmarkp
&& *pmark
!= PM_NOPLACEMARK
) ||
3577 ((*pmark
^ pindex
) &
3578 ~PM_PLACEMARK_WAKEUP
) == 0) {
3579 spin_unlock(&pmap
->pm_spin
);
3580 tsleep(pmark
, PINTERLOCKED
, "pvpld", 0);
3581 spin_lock(&pmap
->pm_spin
);
3586 if (atomic_swap_long(pmark
, pindex
) !=
3588 panic("_pv_get: pmark race");
3592 spin_unlock(&pmap
->pm_spin
);
3595 if (_pv_hold_try(pv PMAP_DEBUG_COPY
)) {
3596 pv_cache(pv
, pindex
);
3598 spin_unlock(&pmap
->pm_spin
);
3600 spin_unlock_shared(&pmap
->pm_spin
);
3601 KKASSERT(pv
->pv_pmap
== pmap
&&
3602 pv
->pv_pindex
== pindex
);
3606 spin_unlock(&pmap
->pm_spin
);
3607 _pv_lock(pv PMAP_DEBUG_COPY
);
3609 spin_lock(&pmap
->pm_spin
);
3611 spin_unlock_shared(&pmap
->pm_spin
);
3612 _pv_lock(pv PMAP_DEBUG_COPY
);
3614 spin_lock_shared(&pmap
->pm_spin
);
3620 * Lookup, hold, and attempt to lock (pmap,pindex).
3622 * If the entry does not exist NULL is returned and *errorp is set to 0
3624 * If the entry exists and could be successfully locked it is returned and
3625 * errorp is set to 0.
3627 * If the entry exists but could NOT be successfully locked it is returned
3628 * held and *errorp is set to 1.
3630 * If the entry is placemarked by someone else NULL is returned and *errorp
3635 pv_get_try(pmap_t pmap
, vm_pindex_t pindex
, vm_pindex_t
**pmarkp
, int *errorp
)
3639 spin_lock_shared(&pmap
->pm_spin
);
3641 pv
= pv_entry_lookup(pmap
, pindex
);
3645 pmark
= pmap_placemarker_hash(pmap
, pindex
);
3647 if (((*pmark
^ pindex
) & ~PM_PLACEMARK_WAKEUP
) == 0) {
3649 } else if (pmarkp
&&
3650 atomic_cmpset_long(pmark
, PM_NOPLACEMARK
, pindex
)) {
3654 * Can't set a placemark with a NULL pmarkp, or if
3655 * pmarkp is non-NULL but we failed to set our
3662 spin_unlock_shared(&pmap
->pm_spin
);
3668 * XXX This has problems if the lock is shared, why?
3670 if (pv_hold_try(pv
)) {
3671 pv_cache(pv
, pindex
); /* overwrite ok (shared lock) */
3672 spin_unlock_shared(&pmap
->pm_spin
);
3674 KKASSERT(pv
->pv_pmap
== pmap
&& pv
->pv_pindex
== pindex
);
3675 return(pv
); /* lock succeeded */
3677 spin_unlock_shared(&pmap
->pm_spin
);
3680 return (pv
); /* lock failed */
3684 * Lock a held pv, keeping the hold count
3688 _pv_lock(pv_entry_t pv PMAP_DEBUG_DECL
)
3693 count
= pv
->pv_hold
;
3695 if ((count
& PV_HOLD_LOCKED
) == 0) {
3696 if (atomic_cmpset_int(&pv
->pv_hold
, count
,
3697 count
| PV_HOLD_LOCKED
)) {
3700 pv
->pv_line
= lineno
;
3706 tsleep_interlock(pv
, 0);
3707 if (atomic_cmpset_int(&pv
->pv_hold
, count
,
3708 count
| PV_HOLD_WAITING
)) {
3710 if (pmap_enter_debug
> 0) {
3712 kprintf("pv waiting on %s:%d\n",
3713 pv
->pv_func
, pv
->pv_line
);
3716 tsleep(pv
, PINTERLOCKED
, "pvwait", hz
);
3723 * Unlock a held and locked pv, keeping the hold count.
3727 pv_unlock(pv_entry_t pv
)
3732 count
= pv
->pv_hold
;
3734 KKASSERT((count
& (PV_HOLD_LOCKED
| PV_HOLD_MASK
)) >=
3735 (PV_HOLD_LOCKED
| 1));
3736 if (atomic_cmpset_int(&pv
->pv_hold
, count
,
3738 ~(PV_HOLD_LOCKED
| PV_HOLD_WAITING
))) {
3739 if (count
& PV_HOLD_WAITING
)
3747 * Unlock and drop a pv. If the pv is no longer associated with a pmap
3748 * and the hold count drops to zero we will free it.
3750 * Caller should not hold any spin locks. We are protected from hold races
3751 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
3752 * lock held. A pv cannot be located otherwise.
3756 pv_put(pv_entry_t pv
)
3759 if (pmap_enter_debug
> 0) {
3761 kprintf("pv_put pv=%p hold=%08x\n", pv
, pv
->pv_hold
);
3766 * Normal put-aways must have a pv_m associated with the pv,
3767 * but allow the case where the pv has been destructed due
3768 * to pmap_dynamic_delete.
3770 KKASSERT(pv
->pv_pmap
== NULL
|| pv
->pv_m
!= NULL
);
3773 * Fast - shortcut most common condition
3775 if (atomic_cmpset_int(&pv
->pv_hold
, PV_HOLD_LOCKED
| 2, 1))
3786 * Remove the pmap association from a pv, require that pv_m already be removed,
3787 * then unlock and drop the pv. Any pte operations must have already been
3788 * completed. This call may result in a last-drop which will physically free
3791 * Removing the pmap association entails an additional drop.
3793 * pv must be exclusively locked on call and will be disposed of on return.
3797 _pv_free(pv_entry_t pv
, pv_entry_t pvp PMAP_DEBUG_DECL
)
3802 pv
->pv_func_lastfree
= func
;
3803 pv
->pv_line_lastfree
= lineno
;
3805 KKASSERT(pv
->pv_m
== NULL
);
3806 KKASSERT((pv
->pv_hold
& (PV_HOLD_LOCKED
|PV_HOLD_MASK
)) >=
3807 (PV_HOLD_LOCKED
|1));
3808 if ((pmap
= pv
->pv_pmap
) != NULL
) {
3809 spin_lock(&pmap
->pm_spin
);
3810 KKASSERT(pv
->pv_pmap
== pmap
);
3811 if (pmap
->pm_pvhint_pt
== pv
)
3812 pmap
->pm_pvhint_pt
= NULL
;
3813 if (pmap
->pm_pvhint_pte
== pv
)
3814 pmap
->pm_pvhint_pte
= NULL
;
3815 pv_entry_rb_tree_RB_REMOVE(&pmap
->pm_pvroot
, pv
);
3816 atomic_add_long(&pmap
->pm_stats
.resident_count
, -1);
3819 spin_unlock(&pmap
->pm_spin
);
3822 * Try to shortcut three atomic ops, otherwise fall through
3823 * and do it normally. Drop two refs and the lock all in
3827 vm_page_unwire_quick(pvp
->pv_m
);
3828 if (atomic_cmpset_int(&pv
->pv_hold
, PV_HOLD_LOCKED
| 2, 0)) {
3830 if (pmap_enter_debug
> 0) {
3832 kprintf("pv_free: free pv %p\n", pv
);
3838 pv_drop(pv
); /* ref for pv_pmap */
3845 * This routine is very drastic, but can save the system
3853 static int warningdone
=0;
3855 if (pmap_pagedaemon_waken
== 0)
3857 pmap_pagedaemon_waken
= 0;
3858 if (warningdone
< 5) {
3859 kprintf("pmap_collect: collecting pv entries -- "
3860 "suggest increasing PMAP_SHPGPERPROC\n");
3864 for (i
= 0; i
< vm_page_array_size
; i
++) {
3865 m
= &vm_page_array
[i
];
3866 if (m
->wire_count
|| m
->hold_count
)
3868 if (vm_page_busy_try(m
, TRUE
) == 0) {
3869 if (m
->wire_count
== 0 && m
->hold_count
== 0) {
3878 * Scan the pmap for active page table entries and issue a callback.
3879 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
3880 * its parent page table.
3882 * pte_pv will be NULL if the page or page table is unmanaged.
3883 * pt_pv will point to the page table page containing the pte for the page.
3885 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
3886 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
3887 * process pmap's PD and page to the callback function. This can be
3888 * confusing because the pt_pv is really a pd_pv, and the target page
3889 * table page is simply aliased by the pmap and not owned by it.
3891 * It is assumed that the start and end are properly rounded to the page size.
3893 * It is assumed that PD pages and above are managed and thus in the RB tree,
3894 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
3896 struct pmap_scan_info
{
3900 vm_pindex_t sva_pd_pindex
;
3901 vm_pindex_t eva_pd_pindex
;
3902 void (*func
)(pmap_t
, struct pmap_scan_info
*,
3903 pv_entry_t
, vm_pindex_t
*, pv_entry_t
,
3905 pt_entry_t
*, void *);
3907 pmap_inval_bulk_t bulk_core
;
3908 pmap_inval_bulk_t
*bulk
;
3913 static int pmap_scan_cmp(pv_entry_t pv
, void *data
);
3914 static int pmap_scan_callback(pv_entry_t pv
, void *data
);
3917 pmap_scan(struct pmap_scan_info
*info
, int smp_inval
)
3919 struct pmap
*pmap
= info
->pmap
;
3920 pv_entry_t pd_pv
; /* A page directory PV */
3921 pv_entry_t pt_pv
; /* A page table PV */
3922 pv_entry_t pte_pv
; /* A page table entry PV */
3923 vm_pindex_t
*pte_placemark
;
3924 vm_pindex_t
*pt_placemark
;
3927 struct pv_entry dummy_pv
;
3932 if (info
->sva
== info
->eva
)
3935 info
->bulk
= &info
->bulk_core
;
3936 pmap_inval_bulk_init(&info
->bulk_core
, pmap
);
3942 * Hold the token for stability; if the pmap is empty we have nothing
3946 if (pmap
->pm_stats
.resident_count
== 0) {
3954 * Special handling for scanning one page, which is a very common
3955 * operation (it is?).
3957 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
3959 if (info
->sva
+ PAGE_SIZE
== info
->eva
) {
3960 if (info
->sva
>= VM_MAX_USER_ADDRESS
) {
3962 * Kernel mappings do not track wire counts on
3963 * page table pages and only maintain pd_pv and
3964 * pte_pv levels so pmap_scan() works.
3967 pte_pv
= pv_get(pmap
, pmap_pte_pindex(info
->sva
),
3969 ptep
= vtopte(info
->sva
);
3972 * User pages which are unmanaged will not have a
3973 * pte_pv. User page table pages which are unmanaged
3974 * (shared from elsewhere) will also not have a pt_pv.
3975 * The func() callback will pass both pte_pv and pt_pv
3976 * as NULL in that case.
3978 * We hold pte_placemark across the operation for
3981 * WARNING! We must hold pt_placemark across the
3982 * *ptep test to prevent misintepreting
3983 * a non-zero *ptep as a shared page
3984 * table page. Hold it across the function
3985 * callback as well for SMP safety.
3987 pte_pv
= pv_get(pmap
, pmap_pte_pindex(info
->sva
),
3989 pt_pv
= pv_get(pmap
, pmap_pt_pindex(info
->sva
),
3991 if (pt_pv
== NULL
) {
3992 KKASSERT(pte_pv
== NULL
);
3993 pd_pv
= pv_get(pmap
,
3994 pmap_pd_pindex(info
->sva
),
3997 ptep
= pv_pte_lookup(pd_pv
,
3998 pmap_pt_index(info
->sva
));
4000 info
->func(pmap
, info
,
4006 pv_placemarker_wakeup(pmap
,
4011 pv_placemarker_wakeup(pmap
,
4014 pv_placemarker_wakeup(pmap
, pte_placemark
);
4017 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(info
->sva
));
4021 * NOTE: *ptep can't be ripped out from under us if we hold
4022 * pte_pv (or pte_placemark) locked, but bits can
4028 KKASSERT(pte_pv
== NULL
);
4029 pv_placemarker_wakeup(pmap
, pte_placemark
);
4030 } else if (pte_pv
) {
4031 KASSERT((oldpte
& (pmap
->pmap_bits
[PG_MANAGED_IDX
] |
4032 pmap
->pmap_bits
[PG_V_IDX
])) ==
4033 (pmap
->pmap_bits
[PG_MANAGED_IDX
] |
4034 pmap
->pmap_bits
[PG_V_IDX
]),
4035 ("badA *ptep %016lx/%016lx sva %016lx pte_pv %p",
4036 *ptep
, oldpte
, info
->sva
, pte_pv
));
4037 info
->func(pmap
, info
, pte_pv
, NULL
, pt_pv
, 0,
4038 info
->sva
, ptep
, info
->arg
);
4040 KASSERT((oldpte
& (pmap
->pmap_bits
[PG_MANAGED_IDX
] |
4041 pmap
->pmap_bits
[PG_V_IDX
])) ==
4042 pmap
->pmap_bits
[PG_V_IDX
],
4043 ("badB *ptep %016lx/%016lx sva %016lx pte_pv NULL",
4044 *ptep
, oldpte
, info
->sva
));
4045 info
->func(pmap
, info
, NULL
, pte_placemark
, pt_pv
, 0,
4046 info
->sva
, ptep
, info
->arg
);
4051 pmap_inval_bulk_flush(info
->bulk
);
4056 * Nominal scan case, RB_SCAN() for PD pages and iterate from
4059 * WARNING! eva can overflow our standard ((N + mask) >> bits)
4060 * bounds, resulting in a pd_pindex of 0. To solve the
4061 * problem we use an inclusive range.
4063 info
->sva_pd_pindex
= pmap_pd_pindex(info
->sva
);
4064 info
->eva_pd_pindex
= pmap_pd_pindex(info
->eva
- PAGE_SIZE
);
4066 if (info
->sva
>= VM_MAX_USER_ADDRESS
) {
4068 * The kernel does not currently maintain any pv_entry's for
4069 * higher-level page tables.
4071 bzero(&dummy_pv
, sizeof(dummy_pv
));
4072 dummy_pv
.pv_pindex
= info
->sva_pd_pindex
;
4073 spin_lock(&pmap
->pm_spin
);
4074 while (dummy_pv
.pv_pindex
<= info
->eva_pd_pindex
) {
4075 pmap_scan_callback(&dummy_pv
, info
);
4076 ++dummy_pv
.pv_pindex
;
4077 if (dummy_pv
.pv_pindex
< info
->sva_pd_pindex
) /*wrap*/
4080 spin_unlock(&pmap
->pm_spin
);
4083 * User page tables maintain local PML4, PDP, and PD
4084 * pv_entry's at the very least. PT pv's might be
4085 * unmanaged and thus not exist. PTE pv's might be
4086 * unmanaged and thus not exist.
4088 spin_lock(&pmap
->pm_spin
);
4089 pv_entry_rb_tree_RB_SCAN(&pmap
->pm_pvroot
, pmap_scan_cmp
,
4090 pmap_scan_callback
, info
);
4091 spin_unlock(&pmap
->pm_spin
);
4093 pmap_inval_bulk_flush(info
->bulk
);
4097 * WARNING! pmap->pm_spin held
4099 * WARNING! eva can overflow our standard ((N + mask) >> bits)
4100 * bounds, resulting in a pd_pindex of 0. To solve the
4101 * problem we use an inclusive range.
4104 pmap_scan_cmp(pv_entry_t pv
, void *data
)
4106 struct pmap_scan_info
*info
= data
;
4107 if (pv
->pv_pindex
< info
->sva_pd_pindex
)
4109 if (pv
->pv_pindex
> info
->eva_pd_pindex
)
4115 * pmap_scan() by PDs
4117 * WARNING! pmap->pm_spin held
4120 pmap_scan_callback(pv_entry_t pv
, void *data
)
4122 struct pmap_scan_info
*info
= data
;
4123 struct pmap
*pmap
= info
->pmap
;
4124 pv_entry_t pd_pv
; /* A page directory PV */
4125 pv_entry_t pt_pv
; /* A page table PV */
4126 vm_pindex_t
*pt_placemark
;
4131 vm_offset_t va_next
;
4132 vm_pindex_t pd_pindex
;
4142 * Pull the PD pindex from the pv before releasing the spinlock.
4144 * WARNING: pv is faked for kernel pmap scans.
4146 pd_pindex
= pv
->pv_pindex
;
4147 spin_unlock(&pmap
->pm_spin
);
4148 pv
= NULL
; /* invalid after spinlock unlocked */
4151 * Calculate the page range within the PD. SIMPLE pmaps are
4152 * direct-mapped for the entire 2^64 address space. Normal pmaps
4153 * reflect the user and kernel address space which requires
4154 * cannonicalization w/regards to converting pd_pindex's back
4157 sva
= (pd_pindex
- pmap_pd_pindex(0)) << PDPSHIFT
;
4158 if ((pmap
->pm_flags
& PMAP_FLAG_SIMPLE
) == 0 &&
4159 (sva
& PML4_SIGNMASK
)) {
4160 sva
|= PML4_SIGNMASK
;
4162 eva
= sva
+ NBPDP
; /* can overflow */
4163 if (sva
< info
->sva
)
4165 if (eva
< info
->sva
|| eva
> info
->eva
)
4169 * NOTE: kernel mappings do not track page table pages, only
4172 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
4173 * However, for the scan to be efficient we try to
4174 * cache items top-down.
4179 for (; sva
< eva
; sva
= va_next
) {
4182 if (sva
>= VM_MAX_USER_ADDRESS
) {
4191 * PD cache, scan shortcut if it doesn't exist.
4193 if (pd_pv
== NULL
) {
4194 pd_pv
= pv_get(pmap
, pmap_pd_pindex(sva
), NULL
);
4195 } else if (pd_pv
->pv_pmap
!= pmap
||
4196 pd_pv
->pv_pindex
!= pmap_pd_pindex(sva
)) {
4198 pd_pv
= pv_get(pmap
, pmap_pd_pindex(sva
), NULL
);
4200 if (pd_pv
== NULL
) {
4201 va_next
= (sva
+ NBPDP
) & ~PDPMASK
;
4210 * NOTE: The cached pt_pv can be removed from the pmap when
4211 * pmap_dynamic_delete is enabled.
4213 if (pt_pv
&& (pt_pv
->pv_pmap
!= pmap
||
4214 pt_pv
->pv_pindex
!= pmap_pt_pindex(sva
))) {
4218 if (pt_pv
== NULL
) {
4219 pt_pv
= pv_get_try(pmap
, pmap_pt_pindex(sva
),
4220 &pt_placemark
, &error
);
4222 pv_put(pd_pv
); /* lock order */
4229 pv_placemarker_wait(pmap
, pt_placemark
);
4234 /* may have to re-check later if pt_pv is NULL here */
4238 * If pt_pv is NULL we either have an shared page table
4239 * page and must issue a callback specific to that case,
4240 * or there is no page table page.
4242 * Either way we can skip the page table page.
4244 * WARNING! pt_pv can also be NULL due to a pv creation
4245 * race where we find it to be NULL and then
4246 * later see a pte_pv. But its possible the pt_pv
4247 * got created inbetween the two operations, so
4250 if (pt_pv
== NULL
) {
4252 * Possible unmanaged (shared from another pmap)
4255 * WARNING! We must hold pt_placemark across the
4256 * *ptep test to prevent misintepreting
4257 * a non-zero *ptep as a shared page
4258 * table page. Hold it across the function
4259 * callback as well for SMP safety.
4261 ptep
= pv_pte_lookup(pd_pv
, pmap_pt_index(sva
));
4262 if (*ptep
& pmap
->pmap_bits
[PG_V_IDX
]) {
4263 info
->func(pmap
, info
, NULL
, pt_placemark
,
4265 sva
, ptep
, info
->arg
);
4267 pv_placemarker_wakeup(pmap
, pt_placemark
);
4271 * Done, move to next page table page.
4273 va_next
= (sva
+ NBPDR
) & ~PDRMASK
;
4280 * From this point in the loop testing pt_pv for non-NULL
4281 * means we are in UVM, else if it is NULL we are in KVM.
4283 * Limit our scan to either the end of the va represented
4284 * by the current page table page, or to the end of the
4285 * range being removed.
4288 va_next
= (sva
+ NBPDR
) & ~PDRMASK
;
4295 * Scan the page table for pages. Some pages may not be
4296 * managed (might not have a pv_entry).
4298 * There is no page table management for kernel pages so
4299 * pt_pv will be NULL in that case, but otherwise pt_pv
4300 * is non-NULL, locked, and referenced.
4304 * At this point a non-NULL pt_pv means a UVA, and a NULL
4305 * pt_pv means a KVA.
4308 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(sva
));
4312 while (sva
< va_next
) {
4314 vm_pindex_t
*pte_placemark
;
4317 * Yield every 64 pages, stop if requested.
4319 if ((++info
->count
& 63) == 0)
4325 * We can shortcut our scan if *ptep == 0. This is
4326 * an unlocked check.
4336 * Acquire the related pte_pv, if any. If *ptep == 0
4337 * the related pte_pv should not exist, but if *ptep
4338 * is not zero the pte_pv may or may not exist (e.g.
4339 * will not exist for an unmanaged page).
4341 * However a multitude of races are possible here
4342 * so if we cannot lock definite state we clean out
4343 * our cache and break the inner while() loop to
4344 * force a loop up to the top of the for().
4346 * XXX unlock/relock pd_pv, pt_pv, and re-test their
4347 * validity instead of looping up?
4349 pte_pv
= pv_get_try(pmap
, pmap_pte_pindex(sva
),
4350 &pte_placemark
, &error
);
4352 pv_put(pd_pv
); /* lock order */
4355 pv_put(pt_pv
); /* lock order */
4358 if (pte_pv
) { /* block */
4363 pv_placemarker_wait(pmap
,
4366 va_next
= sva
; /* retry */
4371 * Reload *ptep after successfully locking the
4372 * pindex. If *ptep == 0 we had better NOT have a
4379 kprintf("Unexpected non-NULL pte_pv "
4381 "*ptep = %016lx/%016lx\n",
4382 pte_pv
, pt_pv
, *ptep
, oldpte
);
4383 panic("Unexpected non-NULL pte_pv");
4385 pv_placemarker_wakeup(pmap
, pte_placemark
);
4393 * We can't hold pd_pv across the callback (because
4394 * we don't pass it to the callback and the callback
4398 vm_page_wire_quick(pd_pv
->pv_m
);
4403 * Ready for the callback. The locked pte_pv (if any)
4404 * is consumed by the callback. pte_pv will exist if
4405 * the page is managed, and will not exist if it
4408 if (oldpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) {
4413 (oldpte
& pmap
->pmap_bits
[PG_V_IDX
]),
4414 ("badC *ptep %016lx/%016lx sva %016lx "
4416 *ptep
, oldpte
, sva
, pte_pv
));
4418 * We must unlock pd_pv across the callback
4419 * to avoid deadlocks on any recursive
4420 * disposal. Re-check that it still exists
4423 * Call target disposes of pte_pv and may
4424 * destroy but will not dispose of pt_pv.
4426 info
->func(pmap
, info
, pte_pv
, NULL
,
4428 sva
, ptep
, info
->arg
);
4433 * We must unlock pd_pv across the callback
4434 * to avoid deadlocks on any recursive
4435 * disposal. Re-check that it still exists
4438 * Call target disposes of pte_pv or
4439 * pte_placemark and may destroy but will
4440 * not dispose of pt_pv.
4442 KASSERT(pte_pv
== NULL
&&
4443 (oldpte
& pmap
->pmap_bits
[PG_V_IDX
]),
4444 ("badD *ptep %016lx/%016lx sva %016lx "
4445 "pte_pv %p pte_pv->pv_m %p ",
4447 pte_pv
, (pte_pv
? pte_pv
->pv_m
: NULL
)));
4451 info
->func(pmap
, info
,
4454 sva
, ptep
, info
->arg
);
4456 info
->func(pmap
, info
,
4457 NULL
, pte_placemark
,
4459 sva
, ptep
, info
->arg
);
4464 vm_page_unwire_quick(pd_pv
->pv_m
);
4465 if (pd_pv
->pv_pmap
== NULL
) {
4466 va_next
= sva
; /* retry */
4472 * NOTE: The cached pt_pv can be removed from the
4473 * pmap when pmap_dynamic_delete is enabled,
4474 * which will cause ptep to become stale.
4476 * This also means that no pages remain under
4477 * the PT, so we can just break out of the inner
4478 * loop and let the outer loop clean everything
4481 if (pt_pv
&& pt_pv
->pv_pmap
!= pmap
)
4496 if ((++info
->count
& 7) == 0)
4500 * Relock before returning.
4502 spin_lock(&pmap
->pm_spin
);
4507 pmap_remove(struct pmap
*pmap
, vm_offset_t sva
, vm_offset_t eva
)
4509 struct pmap_scan_info info
;
4514 info
.func
= pmap_remove_callback
;
4516 pmap_scan(&info
, 1);
4519 if (eva
- sva
< 1024*1024) {
4521 cpu_invlpg((void *)sva
);
4529 pmap_remove_noinval(struct pmap
*pmap
, vm_offset_t sva
, vm_offset_t eva
)
4531 struct pmap_scan_info info
;
4536 info
.func
= pmap_remove_callback
;
4538 pmap_scan(&info
, 0);
4542 pmap_remove_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
4543 pv_entry_t pte_pv
, vm_pindex_t
*pte_placemark
,
4544 pv_entry_t pt_pv
, int sharept
,
4545 vm_offset_t va
, pt_entry_t
*ptep
, void *arg __unused
)
4553 * This will also drop pt_pv's wire_count. Note that
4554 * terminal pages are not wired based on mmu presence.
4556 * NOTE: If this is the kernel_pmap, pt_pv can be NULL.
4558 KKASSERT(pte_pv
->pv_m
!= NULL
);
4559 pmap_remove_pv_pte(pte_pv
, pt_pv
, info
->bulk
, 2);
4560 pte_pv
= NULL
; /* safety */
4563 * Recursively destroy higher-level page tables.
4565 * This is optional. If we do not, they will still
4566 * be destroyed when the process exits.
4568 * NOTE: Do not destroy pv_entry's with extra hold refs,
4569 * a caller may have unlocked it and intends to
4570 * continue to use it.
4572 if (pmap_dynamic_delete
&&
4575 pt_pv
->pv_m
->wire_count
== 1 &&
4576 (pt_pv
->pv_hold
& PV_HOLD_MASK
) == 2 &&
4577 pt_pv
->pv_pindex
!= pmap_pml4_pindex()) {
4578 if (pmap_dynamic_delete
== 2)
4579 kprintf("B %jd %08x\n", pt_pv
->pv_pindex
, pt_pv
->pv_hold
);
4580 pv_hold(pt_pv
); /* extra hold */
4581 pmap_remove_pv_pte(pt_pv
, NULL
, info
->bulk
, 1);
4582 pv_lock(pt_pv
); /* prior extra hold + relock */
4584 } else if (sharept
== 0) {
4586 * Unmanaged pte (pte_placemark is non-NULL)
4588 * pt_pv's wire_count is still bumped by unmanaged pages
4589 * so we must decrement it manually.
4591 * We have to unwire the target page table page.
4593 pte
= pmap_inval_bulk(info
->bulk
, va
, ptep
, 0);
4594 if (pte
& pmap
->pmap_bits
[PG_W_IDX
])
4595 atomic_add_long(&pmap
->pm_stats
.wired_count
, -1);
4596 atomic_add_long(&pmap
->pm_stats
.resident_count
, -1);
4597 if (vm_page_unwire_quick(pt_pv
->pv_m
))
4598 panic("pmap_remove: insufficient wirecount");
4599 pv_placemarker_wakeup(pmap
, pte_placemark
);
4602 * Unmanaged page table (pt, pd, or pdp. Not pte) for
4603 * a shared page table.
4605 * pt_pv is actually the pd_pv for our pmap (not the shared
4608 * We have to unwire the target page table page and we
4609 * have to unwire our page directory page.
4611 * It is unclear how we can invalidate a segment so we
4612 * invalidate -1 which invlidates the tlb.
4614 pte
= pmap_inval_bulk(info
->bulk
, (vm_offset_t
)-1, ptep
, 0);
4615 atomic_add_long(&pmap
->pm_stats
.resident_count
, -1);
4616 KKASSERT((pte
& pmap
->pmap_bits
[PG_DEVICE_IDX
]) == 0);
4617 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte
& PG_FRAME
)))
4618 panic("pmap_remove: shared pgtable1 bad wirecount");
4619 if (vm_page_unwire_quick(pt_pv
->pv_m
))
4620 panic("pmap_remove: shared pgtable2 bad wirecount");
4621 pv_placemarker_wakeup(pmap
, pte_placemark
);
4626 * Removes this physical page from all physical maps in which it resides.
4627 * Reflects back modify bits to the pager.
4629 * This routine may not be called from an interrupt.
4633 pmap_remove_all(vm_page_t m
)
4636 pmap_inval_bulk_t bulk
;
4638 if (!pmap_initialized
/* || (m->flags & PG_FICTITIOUS)*/)
4641 vm_page_spin_lock(m
);
4642 while ((pv
= TAILQ_FIRST(&m
->md
.pv_list
)) != NULL
) {
4643 KKASSERT(pv
->pv_m
== m
);
4644 if (pv_hold_try(pv
)) {
4645 vm_page_spin_unlock(m
);
4647 vm_page_spin_unlock(m
);
4650 vm_page_spin_lock(m
);
4653 KKASSERT(pv
->pv_pmap
&& pv
->pv_m
== m
);
4656 * Holding no spinlocks, pv is locked. Once we scrap
4657 * pv we can no longer use it as a list iterator (but
4658 * we are doing a TAILQ_FIRST() so we are ok).
4660 pmap_inval_bulk_init(&bulk
, pv
->pv_pmap
);
4661 pmap_remove_pv_pte(pv
, NULL
, &bulk
, 2);
4662 pv
= NULL
; /* safety */
4663 pmap_inval_bulk_flush(&bulk
);
4664 vm_page_spin_lock(m
);
4666 KKASSERT((m
->flags
& (PG_MAPPED
|PG_WRITEABLE
)) == 0);
4667 vm_page_spin_unlock(m
);
4671 * Removes the page from a particular pmap
4674 pmap_remove_specific(pmap_t pmap
, vm_page_t m
)
4677 pmap_inval_bulk_t bulk
;
4679 if (!pmap_initialized
)
4683 vm_page_spin_lock(m
);
4684 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
4685 if (pv
->pv_pmap
!= pmap
)
4687 KKASSERT(pv
->pv_m
== m
);
4688 if (pv_hold_try(pv
)) {
4689 vm_page_spin_unlock(m
);
4691 vm_page_spin_unlock(m
);
4696 KKASSERT(pv
->pv_pmap
== pmap
&& pv
->pv_m
== m
);
4699 * Holding no spinlocks, pv is locked. Once gone it can't
4700 * be used as an iterator. In fact, because we couldn't
4701 * necessarily lock it atomically it may have moved within
4702 * the list and ALSO cannot be used as an iterator.
4704 pmap_inval_bulk_init(&bulk
, pv
->pv_pmap
);
4705 pmap_remove_pv_pte(pv
, NULL
, &bulk
, 2);
4706 pv
= NULL
; /* safety */
4707 pmap_inval_bulk_flush(&bulk
);
4710 vm_page_spin_unlock(m
);
4714 * Set the physical protection on the specified range of this map
4715 * as requested. This function is typically only used for debug watchpoints
4718 * This function may not be called from an interrupt if the map is
4719 * not the kernel_pmap.
4721 * NOTE! For shared page table pages we just unmap the page.
4724 pmap_protect(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
, vm_prot_t prot
)
4726 struct pmap_scan_info info
;
4727 /* JG review for NX */
4731 if ((prot
& (VM_PROT_READ
| VM_PROT_EXECUTE
)) == VM_PROT_NONE
) {
4732 pmap_remove(pmap
, sva
, eva
);
4735 if (prot
& VM_PROT_WRITE
)
4740 info
.func
= pmap_protect_callback
;
4742 pmap_scan(&info
, 1);
4747 pmap_protect_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
4748 pv_entry_t pte_pv
, vm_pindex_t
*pte_placemark
,
4749 pv_entry_t pt_pv
, int sharept
,
4750 vm_offset_t va
, pt_entry_t
*ptep
, void *arg __unused
)
4761 KKASSERT(pte_pv
->pv_m
!= NULL
);
4763 if (pbits
& pmap
->pmap_bits
[PG_A_IDX
]) {
4764 if ((pbits
& pmap
->pmap_bits
[PG_DEVICE_IDX
]) == 0) {
4765 m
= PHYS_TO_VM_PAGE(pbits
& PG_FRAME
);
4766 KKASSERT(m
== pte_pv
->pv_m
);
4767 vm_page_flag_set(m
, PG_REFERENCED
);
4769 cbits
&= ~pmap
->pmap_bits
[PG_A_IDX
];
4771 if (pbits
& pmap
->pmap_bits
[PG_M_IDX
]) {
4772 if (pmap_track_modified(pte_pv
->pv_pindex
)) {
4773 if ((pbits
& pmap
->pmap_bits
[PG_DEVICE_IDX
]) == 0) {
4775 m
= PHYS_TO_VM_PAGE(pbits
&
4780 cbits
&= ~pmap
->pmap_bits
[PG_M_IDX
];
4783 } else if (sharept
) {
4785 * Unmanaged page table, pt_pv is actually the pd_pv
4786 * for our pmap (not the object's shared pmap).
4788 * When asked to protect something in a shared page table
4789 * page we just unmap the page table page. We have to
4790 * invalidate the tlb in this situation.
4792 * XXX Warning, shared page tables will not be used for
4793 * OBJT_DEVICE or OBJT_MGTDEVICE (PG_FICTITIOUS) mappings
4794 * so PHYS_TO_VM_PAGE() should be safe here.
4796 pte
= pmap_inval_smp(pmap
, (vm_offset_t
)-1, 1, ptep
, 0);
4797 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte
& PG_FRAME
)))
4798 panic("pmap_protect: pgtable1 pg bad wirecount");
4799 if (vm_page_unwire_quick(pt_pv
->pv_m
))
4800 panic("pmap_protect: pgtable2 pg bad wirecount");
4803 /* else unmanaged page, adjust bits, no wire changes */
4806 cbits
&= ~pmap
->pmap_bits
[PG_RW_IDX
];
4808 if (pmap_enter_debug
> 0) {
4810 kprintf("pmap_protect va=%lx ptep=%p pte_pv=%p "
4811 "pt_pv=%p cbits=%08lx\n",
4817 if (pbits
!= cbits
) {
4820 xva
= (sharept
) ? (vm_offset_t
)-1 : va
;
4821 if (!pmap_inval_smp_cmpset(pmap
, xva
,
4822 ptep
, pbits
, cbits
)) {
4830 pv_placemarker_wakeup(pmap
, pte_placemark
);
4834 * Insert the vm_page (m) at the virtual address (va), replacing any prior
4835 * mapping at that address. Set protection and wiring as requested.
4837 * If entry is non-NULL we check to see if the SEG_SIZE optimization is
4838 * possible. If it is we enter the page into the appropriate shared pmap
4839 * hanging off the related VM object instead of the passed pmap, then we
4840 * share the page table page from the VM object's pmap into the current pmap.
4842 * NOTE: This routine MUST insert the page into the pmap now, it cannot
4845 * NOTE: If (m) is PG_UNMANAGED it may also be a temporary fake vm_page_t.
4849 pmap_enter(pmap_t pmap
, vm_offset_t va
, vm_page_t m
, vm_prot_t prot
,
4850 boolean_t wired
, vm_map_entry_t entry
)
4852 pv_entry_t pt_pv
; /* page table */
4853 pv_entry_t pte_pv
; /* page table entry */
4854 vm_pindex_t
*pte_placemark
;
4857 pt_entry_t origpte
, newpte
;
4862 va
= trunc_page(va
);
4863 #ifdef PMAP_DIAGNOSTIC
4865 panic("pmap_enter: toobig");
4866 if ((va
>= UPT_MIN_ADDRESS
) && (va
< UPT_MAX_ADDRESS
))
4867 panic("pmap_enter: invalid to pmap_enter page table "
4868 "pages (va: 0x%lx)", va
);
4870 if (va
< UPT_MAX_ADDRESS
&& pmap
== &kernel_pmap
) {
4871 kprintf("Warning: pmap_enter called on UVA with "
4874 db_print_backtrace();
4877 if (va
>= UPT_MAX_ADDRESS
&& pmap
!= &kernel_pmap
) {
4878 kprintf("Warning: pmap_enter called on KVA without"
4881 db_print_backtrace();
4886 * Get locked PV entries for our new page table entry (pte_pv or
4887 * pte_placemark) and for its parent page table (pt_pv). We need
4888 * the parent so we can resolve the location of the ptep.
4890 * Only hardware MMU actions can modify the ptep out from
4893 * if (m) is fictitious or unmanaged we do not create a managing
4894 * pte_pv for it. Any pre-existing page's management state must
4895 * match (avoiding code complexity).
4897 * If the pmap is still being initialized we assume existing
4900 * Kernel mapppings do not track page table pages (i.e. pt_pv).
4902 * WARNING! If replacing a managed mapping with an unmanaged mapping
4903 * pte_pv will wind up being non-NULL and must be handled
4906 if (pmap_initialized
== FALSE
) {
4909 pte_placemark
= NULL
;
4912 } else if (m
->flags
& (/*PG_FICTITIOUS |*/ PG_UNMANAGED
)) { /* XXX */
4913 pmap_softwait(pmap
);
4914 pte_pv
= pv_get(pmap
, pmap_pte_pindex(va
), &pte_placemark
);
4915 KKASSERT(pte_pv
== NULL
);
4916 if (va
>= VM_MAX_USER_ADDRESS
) {
4920 pt_pv
= pmap_allocpte_seg(pmap
, pmap_pt_pindex(va
),
4922 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
4926 KASSERT(origpte
== 0 ||
4927 (origpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0,
4928 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte
, va
));
4930 pmap_softwait(pmap
);
4931 if (va
>= VM_MAX_USER_ADDRESS
) {
4933 * Kernel map, pv_entry-tracked.
4936 pte_pv
= pmap_allocpte(pmap
, pmap_pte_pindex(va
), NULL
);
4942 pte_pv
= pmap_allocpte_seg(pmap
, pmap_pte_pindex(va
),
4944 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
4946 pte_placemark
= NULL
; /* safety */
4949 KASSERT(origpte
== 0 ||
4950 (origpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]),
4951 ("Invalid PTE 0x%016jx @ 0x%016jx\n", origpte
, va
));
4954 pa
= VM_PAGE_TO_PHYS(m
);
4955 opa
= origpte
& PG_FRAME
;
4958 * Calculate the new PTE. Note that pte_pv alone does not mean
4959 * the new pte_pv is managed, it could exist because the old pte
4960 * was managed even if the new one is not.
4962 newpte
= (pt_entry_t
)(pa
| pte_prot(pmap
, prot
) |
4963 pmap
->pmap_bits
[PG_V_IDX
] | pmap
->pmap_bits
[PG_A_IDX
]);
4965 newpte
|= pmap
->pmap_bits
[PG_W_IDX
];
4966 if (va
< VM_MAX_USER_ADDRESS
)
4967 newpte
|= pmap
->pmap_bits
[PG_U_IDX
];
4968 if (pte_pv
&& (m
->flags
& (/*PG_FICTITIOUS |*/ PG_UNMANAGED
)) == 0)
4969 newpte
|= pmap
->pmap_bits
[PG_MANAGED_IDX
];
4970 // if (pmap == &kernel_pmap)
4971 // newpte |= pgeflag;
4972 newpte
|= pmap
->pmap_cache_bits
[m
->pat_mode
];
4973 if (m
->flags
& PG_FICTITIOUS
)
4974 newpte
|= pmap
->pmap_bits
[PG_DEVICE_IDX
];
4977 * It is possible for multiple faults to occur in threaded
4978 * environments, the existing pte might be correct.
4980 if (((origpte
^ newpte
) &
4981 ~(pt_entry_t
)(pmap
->pmap_bits
[PG_M_IDX
] |
4982 pmap
->pmap_bits
[PG_A_IDX
])) == 0) {
4987 * Ok, either the address changed or the protection or wiring
4990 * Clear the current entry, interlocking the removal. For managed
4991 * pte's this will also flush the modified state to the vm_page.
4992 * Atomic ops are mandatory in order to ensure that PG_M events are
4993 * not lost during any transition.
4995 * WARNING: The caller has busied the new page but not the original
4996 * vm_page which we are trying to replace. Because we hold
4997 * the pte_pv lock, but have not busied the page, PG bits
4998 * can be cleared out from under us.
5001 if (origpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) {
5003 * Old page was managed. Expect pte_pv to exist.
5004 * (it might also exist if the old page was unmanaged).
5006 * NOTE: pt_pv won't exist for a kernel page
5007 * (managed or otherwise).
5009 * NOTE: We may be reusing the pte_pv so we do not
5010 * destroy it in pmap_remove_pv_pte().
5012 KKASSERT(pte_pv
&& pte_pv
->pv_m
);
5013 if (prot
& VM_PROT_NOSYNC
) {
5014 pmap_remove_pv_pte(pte_pv
, pt_pv
, NULL
, 0);
5016 pmap_inval_bulk_t bulk
;
5018 pmap_inval_bulk_init(&bulk
, pmap
);
5019 pmap_remove_pv_pte(pte_pv
, pt_pv
, &bulk
, 0);
5020 pmap_inval_bulk_flush(&bulk
);
5022 pmap_remove_pv_page(pte_pv
);
5023 /* will either set pte_pv->pv_m or pv_free() later */
5026 * Old page was not managed. If we have a pte_pv
5027 * it better not have a pv_m assigned to it. If the
5028 * new page is managed the pte_pv will be destroyed
5029 * near the end (we need its interlock).
5031 * NOTE: We leave the wire count on the PT page
5032 * intact for the followup enter, but adjust
5033 * the wired-pages count on the pmap.
5035 KKASSERT(pte_pv
== NULL
);
5036 if (prot
& VM_PROT_NOSYNC
) {
5038 * NOSYNC (no mmu sync) requested.
5040 (void)pte_load_clear(ptep
);
5041 cpu_invlpg((void *)va
);
5046 pmap_inval_smp(pmap
, va
, 1, ptep
, 0);
5050 * We must adjust pm_stats manually for unmanaged
5054 atomic_add_long(&pmap
->pm_stats
.
5055 resident_count
, -1);
5057 if (origpte
& pmap
->pmap_bits
[PG_W_IDX
]) {
5058 atomic_add_long(&pmap
->pm_stats
.
5062 KKASSERT(*ptep
== 0);
5066 if (pmap_enter_debug
> 0) {
5068 kprintf("pmap_enter: va=%lx m=%p origpte=%lx newpte=%lx ptep=%p"
5069 " pte_pv=%p pt_pv=%p opa=%lx prot=%02x\n",
5071 origpte
, newpte
, ptep
,
5072 pte_pv
, pt_pv
, opa
, prot
);
5076 if ((newpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0) {
5078 * Entering an unmanaged page. We must wire the pt_pv unless
5079 * we retained the wiring from an unmanaged page we had
5080 * removed (if we retained it via pte_pv that will go away
5083 if (pt_pv
&& (opa
== 0 ||
5084 (origpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]))) {
5085 vm_page_wire_quick(pt_pv
->pv_m
);
5088 atomic_add_long(&pmap
->pm_stats
.wired_count
, 1);
5091 * Unmanaged pages need manual resident_count tracking.
5094 atomic_add_long(&pt_pv
->pv_pmap
->pm_stats
.
5097 if (newpte
& pmap
->pmap_bits
[PG_RW_IDX
])
5098 vm_page_flag_set(m
, PG_WRITEABLE
);
5101 * Entering a managed page. Our pte_pv takes care of the
5102 * PT wiring, so if we had removed an unmanaged page before
5105 * We have to take care of the pmap wired count ourselves.
5107 * Enter on the PV list if part of our managed memory.
5109 KKASSERT(pte_pv
&& (pte_pv
->pv_m
== NULL
|| pte_pv
->pv_m
== m
));
5110 vm_page_spin_lock(m
);
5112 pmap_page_stats_adding(m
);
5113 TAILQ_INSERT_TAIL(&m
->md
.pv_list
, pte_pv
, pv_list
);
5114 vm_page_flag_set(m
, PG_MAPPED
);
5115 if (newpte
& pmap
->pmap_bits
[PG_RW_IDX
])
5116 vm_page_flag_set(m
, PG_WRITEABLE
);
5117 vm_page_spin_unlock(m
);
5120 (origpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0) {
5121 vm_page_unwire_quick(pt_pv
->pv_m
);
5125 * Adjust pmap wired pages count for new entry.
5128 atomic_add_long(&pte_pv
->pv_pmap
->pm_stats
.
5134 * Kernel VMAs (pt_pv == NULL) require pmap invalidation interlocks.
5136 * User VMAs do not because those will be zero->non-zero, so no
5137 * stale entries to worry about at this point.
5139 * For KVM there appear to still be issues. Theoretically we
5140 * should be able to scrap the interlocks entirely but we
5143 if ((prot
& VM_PROT_NOSYNC
) == 0 && pt_pv
== NULL
) {
5144 pmap_inval_smp(pmap
, va
, 1, ptep
, newpte
);
5146 origpte
= atomic_swap_long(ptep
, newpte
);
5147 if (origpte
& pmap
->pmap_bits
[PG_M_IDX
]) {
5148 kprintf("pmap [M] race @ %016jx\n", va
);
5149 atomic_set_long(ptep
, pmap
->pmap_bits
[PG_M_IDX
]);
5152 cpu_invlpg((void *)va
);
5159 KKASSERT((newpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0 ||
5160 (m
->flags
& PG_MAPPED
));
5163 * Cleanup the pv entry, allowing other accessors. If the new page
5164 * is not managed but we have a pte_pv (which was locking our
5165 * operation), we can free it now. pte_pv->pv_m should be NULL.
5167 if (pte_pv
&& (newpte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0) {
5168 pv_free(pte_pv
, pt_pv
);
5169 } else if (pte_pv
) {
5171 } else if (pte_placemark
) {
5172 pv_placemarker_wakeup(pmap
, pte_placemark
);
5179 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
5180 * This code also assumes that the pmap has no pre-existing entry for this
5183 * This code currently may only be used on user pmaps, not kernel_pmap.
5186 pmap_enter_quick(pmap_t pmap
, vm_offset_t va
, vm_page_t m
)
5188 pmap_enter(pmap
, va
, m
, VM_PROT_READ
, FALSE
, NULL
);
5192 * Make a temporary mapping for a physical address. This is only intended
5193 * to be used for panic dumps.
5195 * The caller is responsible for calling smp_invltlb().
5198 pmap_kenter_temporary(vm_paddr_t pa
, long i
)
5200 pmap_kenter_quick((vm_offset_t
)crashdumpmap
+ (i
* PAGE_SIZE
), pa
);
5201 return ((void *)crashdumpmap
);
5204 #define MAX_INIT_PT (96)
5207 * This routine preloads the ptes for a given object into the specified pmap.
5208 * This eliminates the blast of soft faults on process startup and
5209 * immediately after an mmap.
5211 static int pmap_object_init_pt_callback(vm_page_t p
, void *data
);
5214 pmap_object_init_pt(pmap_t pmap
, vm_offset_t addr
, vm_prot_t prot
,
5215 vm_object_t object
, vm_pindex_t pindex
,
5216 vm_size_t size
, int limit
)
5218 struct rb_vm_page_scan_info info
;
5223 * We can't preinit if read access isn't set or there is no pmap
5226 if ((prot
& VM_PROT_READ
) == 0 || pmap
== NULL
|| object
== NULL
)
5230 * We can't preinit if the pmap is not the current pmap
5232 lp
= curthread
->td_lwp
;
5233 if (lp
== NULL
|| pmap
!= vmspace_pmap(lp
->lwp_vmspace
))
5237 * Misc additional checks
5239 psize
= x86_64_btop(size
);
5241 if ((object
->type
!= OBJT_VNODE
) ||
5242 ((limit
& MAP_PREFAULT_PARTIAL
) && (psize
> MAX_INIT_PT
) &&
5243 (object
->resident_page_count
> MAX_INIT_PT
))) {
5247 if (pindex
+ psize
> object
->size
) {
5248 if (object
->size
< pindex
)
5250 psize
= object
->size
- pindex
;
5257 * If everything is segment-aligned do not pre-init here. Instead
5258 * allow the normal vm_fault path to pass a segment hint to
5259 * pmap_enter() which will then use an object-referenced shared
5262 if ((addr
& SEG_MASK
) == 0 &&
5263 (ctob(psize
) & SEG_MASK
) == 0 &&
5264 (ctob(pindex
) & SEG_MASK
) == 0) {
5269 * Use a red-black scan to traverse the requested range and load
5270 * any valid pages found into the pmap.
5272 * We cannot safely scan the object's memq without holding the
5275 info
.start_pindex
= pindex
;
5276 info
.end_pindex
= pindex
+ psize
- 1;
5281 info
.object
= object
;
5284 * By using the NOLK scan, the callback function must be sure
5285 * to return -1 if the VM page falls out of the object.
5287 vm_object_hold_shared(object
);
5288 vm_page_rb_tree_RB_SCAN_NOLK(&object
->rb_memq
, rb_vm_page_scancmp
,
5289 pmap_object_init_pt_callback
, &info
);
5290 vm_object_drop(object
);
5295 pmap_object_init_pt_callback(vm_page_t p
, void *data
)
5297 struct rb_vm_page_scan_info
*info
= data
;
5298 vm_pindex_t rel_index
;
5302 * don't allow an madvise to blow away our really
5303 * free pages allocating pv entries.
5305 if ((info
->limit
& MAP_PREFAULT_MADVISE
) &&
5306 vmstats
.v_free_count
< vmstats
.v_free_reserved
) {
5311 * Ignore list markers and ignore pages we cannot instantly
5312 * busy (while holding the object token).
5314 if (p
->flags
& PG_MARKER
)
5319 if (vm_page_busy_try(p
, TRUE
))
5322 if (vm_page_sbusy_try(p
))
5325 if (((p
->valid
& VM_PAGE_BITS_ALL
) == VM_PAGE_BITS_ALL
) &&
5326 (p
->flags
& PG_FICTITIOUS
) == 0) {
5327 if ((p
->queue
- p
->pc
) == PQ_CACHE
) {
5328 if (hard_busy
== 0) {
5329 vm_page_sbusy_drop(p
);
5333 vm_page_deactivate(p
);
5335 rel_index
= p
->pindex
- info
->start_pindex
;
5336 pmap_enter_quick(info
->pmap
,
5337 info
->addr
+ x86_64_ptob(rel_index
), p
);
5342 vm_page_sbusy_drop(p
);
5345 * We are using an unlocked scan (that is, the scan expects its
5346 * current element to remain in the tree on return). So we have
5347 * to check here and abort the scan if it isn't.
5349 if (p
->object
!= info
->object
)
5356 * Return TRUE if the pmap is in shape to trivially pre-fault the specified
5359 * Returns FALSE if it would be non-trivial or if a pte is already loaded
5362 * XXX This is safe only because page table pages are not freed.
5365 pmap_prefault_ok(pmap_t pmap
, vm_offset_t addr
)
5369 /*spin_lock(&pmap->pm_spin);*/
5370 if ((pte
= pmap_pte(pmap
, addr
)) != NULL
) {
5371 if (*pte
& pmap
->pmap_bits
[PG_V_IDX
]) {
5372 /*spin_unlock(&pmap->pm_spin);*/
5376 /*spin_unlock(&pmap->pm_spin);*/
5381 * Change the wiring attribute for a pmap/va pair. The mapping must already
5382 * exist in the pmap. The mapping may or may not be managed. The wiring in
5383 * the page is not changed, the page is returned so the caller can adjust
5384 * its wiring (the page is not locked in any way).
5386 * Wiring is not a hardware characteristic so there is no need to invalidate
5387 * TLB. However, in an SMP environment we must use a locked bus cycle to
5388 * update the pte (if we are not using the pmap_inval_*() API that is)...
5389 * it's ok to do this for simple wiring changes.
5392 pmap_unwire(pmap_t pmap
, vm_offset_t va
)
5403 * Assume elements in the kernel pmap are stable
5405 if (pmap
== &kernel_pmap
) {
5406 if (pmap_pt(pmap
, va
) == 0)
5408 ptep
= pmap_pte_quick(pmap
, va
);
5409 if (pmap_pte_v(pmap
, ptep
)) {
5410 if (pmap_pte_w(pmap
, ptep
))
5411 atomic_add_long(&pmap
->pm_stats
.wired_count
,-1);
5412 atomic_clear_long(ptep
, pmap
->pmap_bits
[PG_W_IDX
]);
5413 pa
= *ptep
& PG_FRAME
;
5414 m
= PHYS_TO_VM_PAGE(pa
);
5420 * We can only [un]wire pmap-local pages (we cannot wire
5423 pt_pv
= pv_get(pmap
, pmap_pt_pindex(va
), NULL
);
5427 ptep
= pv_pte_lookup(pt_pv
, pmap_pte_index(va
));
5428 if ((*ptep
& pmap
->pmap_bits
[PG_V_IDX
]) == 0) {
5433 if (pmap_pte_w(pmap
, ptep
)) {
5434 atomic_add_long(&pt_pv
->pv_pmap
->pm_stats
.wired_count
,
5437 /* XXX else return NULL so caller doesn't unwire m ? */
5439 atomic_clear_long(ptep
, pmap
->pmap_bits
[PG_W_IDX
]);
5441 pa
= *ptep
& PG_FRAME
;
5442 m
= PHYS_TO_VM_PAGE(pa
); /* held by wired count */
5449 * Copy the range specified by src_addr/len from the source map to
5450 * the range dst_addr/len in the destination map.
5452 * This routine is only advisory and need not do anything.
5455 pmap_copy(pmap_t dst_pmap
, pmap_t src_pmap
, vm_offset_t dst_addr
,
5456 vm_size_t len
, vm_offset_t src_addr
)
5463 * Zero the specified physical page.
5465 * This function may be called from an interrupt and no locking is
5469 pmap_zero_page(vm_paddr_t phys
)
5471 vm_offset_t va
= PHYS_TO_DMAP(phys
);
5473 pagezero((void *)va
);
5479 * Zero part of a physical page by mapping it into memory and clearing
5480 * its contents with bzero.
5482 * off and size may not cover an area beyond a single hardware page.
5485 pmap_zero_page_area(vm_paddr_t phys
, int off
, int size
)
5487 vm_offset_t virt
= PHYS_TO_DMAP(phys
);
5489 bzero((char *)virt
+ off
, size
);
5495 * Copy the physical page from the source PA to the target PA.
5496 * This function may be called from an interrupt. No locking
5500 pmap_copy_page(vm_paddr_t src
, vm_paddr_t dst
)
5502 vm_offset_t src_virt
, dst_virt
;
5504 src_virt
= PHYS_TO_DMAP(src
);
5505 dst_virt
= PHYS_TO_DMAP(dst
);
5506 bcopy((void *)src_virt
, (void *)dst_virt
, PAGE_SIZE
);
5510 * pmap_copy_page_frag:
5512 * Copy the physical page from the source PA to the target PA.
5513 * This function may be called from an interrupt. No locking
5517 pmap_copy_page_frag(vm_paddr_t src
, vm_paddr_t dst
, size_t bytes
)
5519 vm_offset_t src_virt
, dst_virt
;
5521 src_virt
= PHYS_TO_DMAP(src
);
5522 dst_virt
= PHYS_TO_DMAP(dst
);
5524 bcopy((char *)src_virt
+ (src
& PAGE_MASK
),
5525 (char *)dst_virt
+ (dst
& PAGE_MASK
),
5530 * Returns true if the pmap's pv is one of the first 16 pvs linked to from
5531 * this page. This count may be changed upwards or downwards in the future;
5532 * it is only necessary that true be returned for a small subset of pmaps
5533 * for proper page aging.
5536 pmap_page_exists_quick(pmap_t pmap
, vm_page_t m
)
5541 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
5544 vm_page_spin_lock(m
);
5545 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
5546 if (pv
->pv_pmap
== pmap
) {
5547 vm_page_spin_unlock(m
);
5554 vm_page_spin_unlock(m
);
5559 * Remove all pages from specified address space this aids process exit
5560 * speeds. Also, this code may be special cased for the current process
5564 pmap_remove_pages(pmap_t pmap
, vm_offset_t sva
, vm_offset_t eva
)
5566 pmap_remove_noinval(pmap
, sva
, eva
);
5571 * pmap_testbit tests bits in pte's note that the testbit/clearbit
5572 * routines are inline, and a lot of things compile-time evaluate.
5577 pmap_testbit(vm_page_t m
, int bit
)
5583 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
5586 if (TAILQ_FIRST(&m
->md
.pv_list
) == NULL
)
5588 vm_page_spin_lock(m
);
5589 if (TAILQ_FIRST(&m
->md
.pv_list
) == NULL
) {
5590 vm_page_spin_unlock(m
);
5594 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
5595 #if defined(PMAP_DIAGNOSTIC)
5596 if (pv
->pv_pmap
== NULL
) {
5597 kprintf("Null pmap (tb) at pindex: %"PRIu64
"\n",
5605 * If the bit being tested is the modified bit, then
5606 * mark clean_map and ptes as never
5609 * WARNING! Because we do not lock the pv, *pte can be in a
5610 * state of flux. Despite this the value of *pte
5611 * will still be related to the vm_page in some way
5612 * because the pv cannot be destroyed as long as we
5613 * hold the vm_page spin lock.
5615 if (bit
== PG_A_IDX
|| bit
== PG_M_IDX
) {
5616 //& (pmap->pmap_bits[PG_A_IDX] | pmap->pmap_bits[PG_M_IDX])) {
5617 if (!pmap_track_modified(pv
->pv_pindex
))
5621 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_pindex
<< PAGE_SHIFT
);
5622 if (*pte
& pmap
->pmap_bits
[bit
]) {
5623 vm_page_spin_unlock(m
);
5627 vm_page_spin_unlock(m
);
5632 * This routine is used to modify bits in ptes. Only one bit should be
5633 * specified. PG_RW requires special handling.
5635 * Caller must NOT hold any spin locks
5639 pmap_clearbit(vm_page_t m
, int bit_index
)
5646 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
)) {
5647 if (bit_index
== PG_RW_IDX
)
5648 vm_page_flag_clear(m
, PG_WRITEABLE
);
5655 * Loop over all current mappings setting/clearing as appropos If
5656 * setting RO do we need to clear the VAC?
5658 * NOTE: When clearing PG_M we could also (not implemented) drop
5659 * through to the PG_RW code and clear PG_RW too, forcing
5660 * a fault on write to redetect PG_M for virtual kernels, but
5661 * it isn't necessary since virtual kernels invalidate the
5662 * pte when they clear the VPTE_M bit in their virtual page
5665 * NOTE: Does not re-dirty the page when clearing only PG_M.
5667 * NOTE: Because we do not lock the pv, *pte can be in a state of
5668 * flux. Despite this the value of *pte is still somewhat
5669 * related while we hold the vm_page spin lock.
5671 * *pte can be zero due to this race. Since we are clearing
5672 * bits we basically do no harm when this race occurs.
5674 if (bit_index
!= PG_RW_IDX
) {
5675 vm_page_spin_lock(m
);
5676 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
5677 #if defined(PMAP_DIAGNOSTIC)
5678 if (pv
->pv_pmap
== NULL
) {
5679 kprintf("Null pmap (cb) at pindex: %"PRIu64
"\n",
5685 pte
= pmap_pte_quick(pv
->pv_pmap
,
5686 pv
->pv_pindex
<< PAGE_SHIFT
);
5688 if (pbits
& pmap
->pmap_bits
[bit_index
])
5689 atomic_clear_long(pte
, pmap
->pmap_bits
[bit_index
]);
5691 vm_page_spin_unlock(m
);
5696 * Clear PG_RW. Also clears PG_M and marks the page dirty if PG_M
5700 vm_page_spin_lock(m
);
5701 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
5703 * don't write protect pager mappings
5705 if (!pmap_track_modified(pv
->pv_pindex
))
5708 #if defined(PMAP_DIAGNOSTIC)
5709 if (pv
->pv_pmap
== NULL
) {
5710 kprintf("Null pmap (cb) at pindex: %"PRIu64
"\n",
5718 * Skip pages which do not have PG_RW set.
5720 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_pindex
<< PAGE_SHIFT
);
5721 if ((*pte
& pmap
->pmap_bits
[PG_RW_IDX
]) == 0)
5725 * We must lock the PV to be able to safely test the pte.
5727 if (pv_hold_try(pv
)) {
5728 vm_page_spin_unlock(m
);
5730 vm_page_spin_unlock(m
);
5731 pv_lock(pv
); /* held, now do a blocking lock */
5737 * Reload pte after acquiring pv.
5739 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_pindex
<< PAGE_SHIFT
);
5741 if ((*pte
& pmap
->pmap_bits
[PG_RW_IDX
]) == 0) {
5747 KKASSERT(pv
->pv_pmap
== pmap
&& pv
->pv_m
== m
);
5753 nbits
= pbits
& ~(pmap
->pmap_bits
[PG_RW_IDX
] |
5754 pmap
->pmap_bits
[PG_M_IDX
]);
5755 if (pmap_inval_smp_cmpset(pmap
,
5756 ((vm_offset_t
)pv
->pv_pindex
<< PAGE_SHIFT
),
5757 pte
, pbits
, nbits
)) {
5764 * If PG_M was found to be set while we were clearing PG_RW
5765 * we also clear PG_M (done above) and mark the page dirty.
5766 * Callers expect this behavior.
5768 * we lost pv so it cannot be used as an iterator. In fact,
5769 * because we couldn't necessarily lock it atomically it may
5770 * have moved within the list and ALSO cannot be used as an
5773 vm_page_spin_lock(m
);
5774 if (pbits
& pmap
->pmap_bits
[PG_M_IDX
])
5776 vm_page_spin_unlock(m
);
5780 if (bit_index
== PG_RW_IDX
)
5781 vm_page_flag_clear(m
, PG_WRITEABLE
);
5782 vm_page_spin_unlock(m
);
5786 * Lower the permission for all mappings to a given page.
5788 * Page must be busied by caller. Because page is busied by caller this
5789 * should not be able to race a pmap_enter().
5792 pmap_page_protect(vm_page_t m
, vm_prot_t prot
)
5794 /* JG NX support? */
5795 if ((prot
& VM_PROT_WRITE
) == 0) {
5796 if (prot
& (VM_PROT_READ
| VM_PROT_EXECUTE
)) {
5798 * NOTE: pmap_clearbit(.. PG_RW) also clears
5799 * the PG_WRITEABLE flag in (m).
5801 pmap_clearbit(m
, PG_RW_IDX
);
5809 pmap_phys_address(vm_pindex_t ppn
)
5811 return (x86_64_ptob(ppn
));
5815 * Return a count of reference bits for a page, clearing those bits.
5816 * It is not necessary for every reference bit to be cleared, but it
5817 * is necessary that 0 only be returned when there are truly no
5818 * reference bits set.
5820 * XXX: The exact number of bits to check and clear is a matter that
5821 * should be tested and standardized at some point in the future for
5822 * optimal aging of shared pages.
5824 * This routine may not block.
5827 pmap_ts_referenced(vm_page_t m
)
5834 if (!pmap_initialized
|| (m
->flags
& PG_FICTITIOUS
))
5837 vm_page_spin_lock(m
);
5838 TAILQ_FOREACH(pv
, &m
->md
.pv_list
, pv_list
) {
5839 if (!pmap_track_modified(pv
->pv_pindex
))
5842 pte
= pmap_pte_quick(pv
->pv_pmap
, pv
->pv_pindex
<< PAGE_SHIFT
);
5843 if (pte
&& (*pte
& pmap
->pmap_bits
[PG_A_IDX
])) {
5844 atomic_clear_long(pte
, pmap
->pmap_bits
[PG_A_IDX
]);
5850 vm_page_spin_unlock(m
);
5857 * Return whether or not the specified physical page was modified
5858 * in any physical maps.
5861 pmap_is_modified(vm_page_t m
)
5865 res
= pmap_testbit(m
, PG_M_IDX
);
5870 * Clear the modify bits on the specified physical page.
5873 pmap_clear_modify(vm_page_t m
)
5875 pmap_clearbit(m
, PG_M_IDX
);
5879 * pmap_clear_reference:
5881 * Clear the reference bit on the specified physical page.
5884 pmap_clear_reference(vm_page_t m
)
5886 pmap_clearbit(m
, PG_A_IDX
);
5890 * Miscellaneous support routines follow
5895 i386_protection_init(void)
5901 * NX supported? (boot time loader.conf override only)
5903 TUNABLE_INT_FETCH("machdep.pmap_nx_enable", &pmap_nx_enable
);
5904 if (pmap_nx_enable
== 0 || (amd_feature
& AMDID_NX
) == 0)
5905 pmap_bits_default
[PG_NX_IDX
] = 0;
5908 * 0 is basically read-only access, but also set the NX (no-execute)
5909 * bit when VM_PROT_EXECUTE is not specified.
5911 kp
= protection_codes
;
5912 for (prot
= 0; prot
< PROTECTION_CODES_SIZE
; prot
++) {
5914 case VM_PROT_NONE
| VM_PROT_NONE
| VM_PROT_NONE
:
5916 * This case handled elsewhere
5920 case VM_PROT_READ
| VM_PROT_NONE
| VM_PROT_NONE
:
5924 *kp
++ = pmap_bits_default
[PG_NX_IDX
];
5926 case VM_PROT_READ
| VM_PROT_NONE
| VM_PROT_EXECUTE
:
5927 case VM_PROT_NONE
| VM_PROT_NONE
| VM_PROT_EXECUTE
:
5929 * Execute requires read access
5933 case VM_PROT_NONE
| VM_PROT_WRITE
| VM_PROT_NONE
:
5934 case VM_PROT_READ
| VM_PROT_WRITE
| VM_PROT_NONE
:
5936 * Write without execute is RW|NX
5938 *kp
++ = pmap_bits_default
[PG_RW_IDX
] |
5939 pmap_bits_default
[PG_NX_IDX
];
5941 case VM_PROT_READ
| VM_PROT_WRITE
| VM_PROT_EXECUTE
:
5942 case VM_PROT_NONE
| VM_PROT_WRITE
| VM_PROT_EXECUTE
:
5944 * Write with execute is RW
5946 *kp
++ = pmap_bits_default
[PG_RW_IDX
];
5953 * Map a set of physical memory pages into the kernel virtual
5954 * address space. Return a pointer to where it is mapped. This
5955 * routine is intended to be used for mapping device memory,
5958 * NOTE: We can't use pgeflag unless we invalidate the pages one at
5961 * NOTE: The PAT attributes {WRITE_BACK, WRITE_THROUGH, UNCACHED, UNCACHEABLE}
5962 * work whether the cpu supports PAT or not. The remaining PAT
5963 * attributes {WRITE_PROTECTED, WRITE_COMBINING} only work if the cpu
5967 pmap_mapdev(vm_paddr_t pa
, vm_size_t size
)
5969 return(pmap_mapdev_attr(pa
, size
, PAT_WRITE_BACK
));
5973 pmap_mapdev_uncacheable(vm_paddr_t pa
, vm_size_t size
)
5975 return(pmap_mapdev_attr(pa
, size
, PAT_UNCACHEABLE
));
5979 pmap_mapbios(vm_paddr_t pa
, vm_size_t size
)
5981 return (pmap_mapdev_attr(pa
, size
, PAT_WRITE_BACK
));
5985 * Map a set of physical memory pages into the kernel virtual
5986 * address space. Return a pointer to where it is mapped. This
5987 * routine is intended to be used for mapping device memory,
5991 pmap_mapdev_attr(vm_paddr_t pa
, vm_size_t size
, int mode
)
5993 vm_offset_t va
, tmpva
, offset
;
5997 offset
= pa
& PAGE_MASK
;
5998 size
= roundup(offset
+ size
, PAGE_SIZE
);
6000 va
= kmem_alloc_nofault(&kernel_map
, size
, VM_SUBSYS_MAPDEV
, PAGE_SIZE
);
6002 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
6004 pa
= pa
& ~PAGE_MASK
;
6005 for (tmpva
= va
, tmpsize
= size
; tmpsize
> 0;) {
6006 pte
= vtopte(tmpva
);
6008 kernel_pmap
.pmap_bits
[PG_RW_IDX
] |
6009 kernel_pmap
.pmap_bits
[PG_V_IDX
] | /* pgeflag | */
6010 kernel_pmap
.pmap_cache_bits
[mode
];
6011 tmpsize
-= PAGE_SIZE
;
6015 pmap_invalidate_range(&kernel_pmap
, va
, va
+ size
);
6016 pmap_invalidate_cache_range(va
, va
+ size
);
6018 return ((void *)(va
+ offset
));
6022 pmap_unmapdev(vm_offset_t va
, vm_size_t size
)
6024 vm_offset_t base
, offset
;
6026 base
= va
& ~PAGE_MASK
;
6027 offset
= va
& PAGE_MASK
;
6028 size
= roundup(offset
+ size
, PAGE_SIZE
);
6029 pmap_qremove(va
, size
>> PAGE_SHIFT
);
6030 kmem_free(&kernel_map
, base
, size
);
6034 * Sets the memory attribute for the specified page.
6037 pmap_page_set_memattr(vm_page_t m
, vm_memattr_t ma
)
6043 * If "m" is a normal page, update its direct mapping. This update
6044 * can be relied upon to perform any cache operations that are
6045 * required for data coherence.
6047 if ((m
->flags
& PG_FICTITIOUS
) == 0)
6048 pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m
)), 1, m
->pat_mode
);
6052 * Change the PAT attribute on an existing kernel memory map. Caller
6053 * must ensure that the virtual memory in question is not accessed
6054 * during the adjustment.
6057 pmap_change_attr(vm_offset_t va
, vm_size_t count
, int mode
)
6064 panic("pmap_change_attr: va is NULL");
6065 base
= trunc_page(va
);
6069 *pte
= (*pte
& ~(pt_entry_t
)(kernel_pmap
.pmap_cache_mask
)) |
6070 kernel_pmap
.pmap_cache_bits
[mode
];
6075 changed
= 1; /* XXX: not optimal */
6078 * Flush CPU caches if required to make sure any data isn't cached that
6079 * shouldn't be, etc.
6082 pmap_invalidate_range(&kernel_pmap
, base
, va
);
6083 pmap_invalidate_cache_range(base
, va
);
6088 * perform the pmap work for mincore
6091 pmap_mincore(pmap_t pmap
, vm_offset_t addr
)
6093 pt_entry_t
*ptep
, pte
;
6097 ptep
= pmap_pte(pmap
, addr
);
6099 if (ptep
&& (pte
= *ptep
) != 0) {
6102 val
= MINCORE_INCORE
;
6103 if ((pte
& pmap
->pmap_bits
[PG_MANAGED_IDX
]) == 0)
6106 pa
= pte
& PG_FRAME
;
6108 if (pte
& pmap
->pmap_bits
[PG_DEVICE_IDX
])
6111 m
= PHYS_TO_VM_PAGE(pa
);
6116 if (pte
& pmap
->pmap_bits
[PG_M_IDX
])
6117 val
|= MINCORE_MODIFIED
|MINCORE_MODIFIED_OTHER
;
6119 * Modified by someone
6121 else if (m
&& (m
->dirty
|| pmap_is_modified(m
)))
6122 val
|= MINCORE_MODIFIED_OTHER
;
6126 if (pte
& pmap
->pmap_bits
[PG_A_IDX
])
6127 val
|= MINCORE_REFERENCED
|MINCORE_REFERENCED_OTHER
;
6130 * Referenced by someone
6132 else if (m
&& ((m
->flags
& PG_REFERENCED
) ||
6133 pmap_ts_referenced(m
))) {
6134 val
|= MINCORE_REFERENCED_OTHER
;
6135 vm_page_flag_set(m
, PG_REFERENCED
);
6144 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
6145 * vmspace will be ref'd and the old one will be deref'd.
6147 * The vmspace for all lwps associated with the process will be adjusted
6148 * and cr3 will be reloaded if any lwp is the current lwp.
6150 * The process must hold the vmspace->vm_map.token for oldvm and newvm
6153 pmap_replacevm(struct proc
*p
, struct vmspace
*newvm
, int adjrefs
)
6155 struct vmspace
*oldvm
;
6158 oldvm
= p
->p_vmspace
;
6159 if (oldvm
!= newvm
) {
6162 p
->p_vmspace
= newvm
;
6163 KKASSERT(p
->p_nthreads
== 1);
6164 lp
= RB_ROOT(&p
->p_lwp_tree
);
6165 pmap_setlwpvm(lp
, newvm
);
6172 * Set the vmspace for a LWP. The vmspace is almost universally set the
6173 * same as the process vmspace, but virtual kernels need to swap out contexts
6174 * on a per-lwp basis.
6176 * Caller does not necessarily hold any vmspace tokens. Caller must control
6177 * the lwp (typically be in the context of the lwp). We use a critical
6178 * section to protect against statclock and hardclock (statistics collection).
6181 pmap_setlwpvm(struct lwp
*lp
, struct vmspace
*newvm
)
6183 struct vmspace
*oldvm
;
6186 oldvm
= lp
->lwp_vmspace
;
6188 if (oldvm
!= newvm
) {
6190 KKASSERT((newvm
->vm_refcnt
& VM_REF_DELETED
) == 0);
6191 lp
->lwp_vmspace
= newvm
;
6192 if (curthread
->td_lwp
== lp
) {
6193 pmap
= vmspace_pmap(newvm
);
6194 ATOMIC_CPUMASK_ORBIT(pmap
->pm_active
, mycpu
->gd_cpuid
);
6195 if (pmap
->pm_active_lock
& CPULOCK_EXCL
)
6196 pmap_interlock_wait(newvm
);
6197 #if defined(SWTCH_OPTIM_STATS)
6200 if (pmap
->pmap_bits
[TYPE_IDX
] == REGULAR_PMAP
) {
6201 curthread
->td_pcb
->pcb_cr3
= vtophys(pmap
->pm_pml4
);
6202 } else if (pmap
->pmap_bits
[TYPE_IDX
] == EPT_PMAP
) {
6203 curthread
->td_pcb
->pcb_cr3
= KPML4phys
;
6205 panic("pmap_setlwpvm: unknown pmap type\n");
6207 load_cr3(curthread
->td_pcb
->pcb_cr3
);
6208 pmap
= vmspace_pmap(oldvm
);
6209 ATOMIC_CPUMASK_NANDBIT(pmap
->pm_active
,
6217 * Called when switching to a locked pmap, used to interlock against pmaps
6218 * undergoing modifications to prevent us from activating the MMU for the
6219 * target pmap until all such modifications have completed. We have to do
6220 * this because the thread making the modifications has already set up its
6221 * SMP synchronization mask.
6223 * This function cannot sleep!
6228 pmap_interlock_wait(struct vmspace
*vm
)
6230 struct pmap
*pmap
= &vm
->vm_pmap
;
6232 if (pmap
->pm_active_lock
& CPULOCK_EXCL
) {
6234 KKASSERT(curthread
->td_critcount
>= 2);
6235 DEBUG_PUSH_INFO("pmap_interlock_wait");
6236 while (pmap
->pm_active_lock
& CPULOCK_EXCL
) {
6238 lwkt_process_ipiq();
6246 pmap_addr_hint(vm_object_t obj
, vm_offset_t addr
, vm_size_t size
)
6249 if ((obj
== NULL
) || (size
< NBPDR
) ||
6250 ((obj
->type
!= OBJT_DEVICE
) && (obj
->type
!= OBJT_MGTDEVICE
))) {
6254 addr
= roundup2(addr
, NBPDR
);
6259 * Used by kmalloc/kfree, page already exists at va
6262 pmap_kvtom(vm_offset_t va
)
6264 pt_entry_t
*ptep
= vtopte(va
);
6266 KKASSERT((*ptep
& kernel_pmap
.pmap_bits
[PG_DEVICE_IDX
]) == 0);
6267 return(PHYS_TO_VM_PAGE(*ptep
& PG_FRAME
));
6271 * Initialize machine-specific shared page directory support. This
6272 * is executed when a VM object is created.
6275 pmap_object_init(vm_object_t object
)
6277 object
->md
.pmap_rw
= NULL
;
6278 object
->md
.pmap_ro
= NULL
;
6282 * Clean up machine-specific shared page directory support. This
6283 * is executed when a VM object is destroyed.
6286 pmap_object_free(vm_object_t object
)
6290 if ((pmap
= object
->md
.pmap_rw
) != NULL
) {
6291 object
->md
.pmap_rw
= NULL
;
6292 pmap_remove_noinval(pmap
,
6293 VM_MIN_USER_ADDRESS
, VM_MAX_USER_ADDRESS
);
6294 CPUMASK_ASSZERO(pmap
->pm_active
);
6297 kfree(pmap
, M_OBJPMAP
);
6299 if ((pmap
= object
->md
.pmap_ro
) != NULL
) {
6300 object
->md
.pmap_ro
= NULL
;
6301 pmap_remove_noinval(pmap
,
6302 VM_MIN_USER_ADDRESS
, VM_MAX_USER_ADDRESS
);
6303 CPUMASK_ASSZERO(pmap
->pm_active
);
6306 kfree(pmap
, M_OBJPMAP
);
6311 * pmap_pgscan_callback - Used by pmap_pgscan to acquire the related
6312 * VM page and issue a pginfo->callback.
6314 * We are expected to dispose of any non-NULL pte_pv.
6318 pmap_pgscan_callback(pmap_t pmap
, struct pmap_scan_info
*info
,
6319 pv_entry_t pte_pv
, vm_pindex_t
*pte_placemark
,
6320 pv_entry_t pt_pv
, int sharept
,
6321 vm_offset_t va
, pt_entry_t
*ptep
, void *arg
)
6323 struct pmap_pgscan_info
*pginfo
= arg
;
6328 * Try to busy the page while we hold the pte_pv locked.
6330 KKASSERT(pte_pv
->pv_m
);
6331 m
= PHYS_TO_VM_PAGE(*ptep
& PG_FRAME
);
6332 if (vm_page_busy_try(m
, TRUE
) == 0) {
6333 if (m
== PHYS_TO_VM_PAGE(*ptep
& PG_FRAME
)) {
6335 * The callback is issued with the pte_pv
6336 * unlocked and put away, and the pt_pv
6341 vm_page_wire_quick(pt_pv
->pv_m
);
6344 if (pginfo
->callback(pginfo
, va
, m
) < 0)
6348 vm_page_unwire_quick(pt_pv
->pv_m
);
6355 ++pginfo
->busycount
;
6360 * Shared page table or unmanaged page (sharept or !sharept)
6362 pv_placemarker_wakeup(pmap
, pte_placemark
);
6367 pmap_pgscan(struct pmap_pgscan_info
*pginfo
)
6369 struct pmap_scan_info info
;
6371 pginfo
->offset
= pginfo
->beg_addr
;
6372 info
.pmap
= pginfo
->pmap
;
6373 info
.sva
= pginfo
->beg_addr
;
6374 info
.eva
= pginfo
->end_addr
;
6375 info
.func
= pmap_pgscan_callback
;
6377 pmap_scan(&info
, 0);
6379 pginfo
->offset
= pginfo
->end_addr
;
6383 * Wait for a placemarker that we do not own to clear. The placemarker
6384 * in question is not necessarily set to the pindex we want, we may have
6385 * to wait on the element because we want to reserve it ourselves.
6387 * NOTE: PM_PLACEMARK_WAKEUP sets a bit which is already set in
6388 * PM_NOPLACEMARK, so it does not interfere with placemarks
6389 * which have already been woken up.
6393 pv_placemarker_wait(pmap_t pmap
, vm_pindex_t
*pmark
)
6395 if (*pmark
!= PM_NOPLACEMARK
) {
6396 atomic_set_long(pmark
, PM_PLACEMARK_WAKEUP
);
6397 tsleep_interlock(pmark
, 0);
6398 if (*pmark
!= PM_NOPLACEMARK
)
6399 tsleep(pmark
, PINTERLOCKED
, "pvplw", 0);
6404 * Wakeup a placemarker that we own. Replace the entry with
6405 * PM_NOPLACEMARK and issue a wakeup() if necessary.
6409 pv_placemarker_wakeup(pmap_t pmap
, vm_pindex_t
*pmark
)
6413 pindex
= atomic_swap_long(pmark
, PM_NOPLACEMARK
);
6414 KKASSERT(pindex
!= PM_NOPLACEMARK
);
6415 if (pindex
& PM_PLACEMARK_WAKEUP
)