4 * Copyright (c) 1991 Regents of the University of California.
5 * Copyright (c) 1994 John S. Dyson
6 * Copyright (c) 1994 David Greenman
7 * Copyright (c) 2003 Peter Wemm
8 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
9 * Copyright (c) 2008, 2009 The DragonFly Project.
10 * Copyright (c) 2008, 2009 Jordan Gordeev.
11 * All rights reserved.
13 * This code is derived from software contributed to Berkeley by
14 * the Systems Programming Group of the University of Utah Computer
15 * Science Department and William Jolitz of UUNET Technologies Inc.
17 * Redistribution and use in source and binary forms, with or without
18 * modification, are permitted provided that the following conditions
20 * 1. Redistributions of source code must retain the above copyright
21 * notice, this list of conditions and the following disclaimer.
22 * 2. Redistributions in binary form must reproduce the above copyright
23 * notice, this list of conditions and the following disclaimer in the
24 * documentation and/or other materials provided with the distribution.
25 * 3. All advertising materials mentioning features or use of this software
26 * must display the following acknowledgement:
27 * This product includes software developed by the University of
28 * California, Berkeley and its contributors.
29 * 4. Neither the name of the University nor the names of its contributors
30 * may be used to endorse or promote products derived from this software
31 * without specific prior written permission.
33 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
34 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
35 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
36 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
37 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
38 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
39 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
40 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
41 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
42 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
45 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
46 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
50 * Manages physical address maps.
56 #include "opt_msgbuf.h"
58 #include <sys/param.h>
59 #include <sys/systm.h>
60 #include <sys/kernel.h>
62 #include <sys/msgbuf.h>
63 #include <sys/vmmeter.h>
65 #include <sys/vmspace.h>
68 #include <vm/vm_param.h>
69 #include <sys/sysctl.h>
71 #include <vm/vm_kern.h>
72 #include <vm/vm_page.h>
73 #include <vm/vm_map.h>
74 #include <vm/vm_object.h>
75 #include <vm/vm_extern.h>
76 #include <vm/vm_pageout.h>
77 #include <vm/vm_pager.h>
78 #include <vm/vm_zone.h>
81 #include <sys/thread2.h>
82 #include <sys/sysref2.h>
83 #include <sys/spinlock2.h>
84 #include <vm/vm_page2.h>
86 #include <machine/cputypes.h>
87 #include <machine/md_var.h>
88 #include <machine/specialreg.h>
89 #include <machine/smp.h>
90 #include <machine/globaldata.h>
91 #include <machine/pmap.h>
92 #include <machine/pmap_inval.h>
101 #define PMAP_KEEP_PDIRS
102 #ifndef PMAP_SHPGPERPROC
103 #define PMAP_SHPGPERPROC 1000
106 #if defined(DIAGNOSTIC)
107 #define PMAP_DIAGNOSTIC
112 #if !defined(PMAP_DIAGNOSTIC)
113 #define PMAP_INLINE __inline
119 * Get PDEs and PTEs for user/kernel address space
121 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
122 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
124 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & VPTE_V) != 0)
125 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & VPTE_WIRED) != 0)
126 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & VPTE_M) != 0)
127 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & VPTE_A) != 0)
128 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & VPTE_V) != 0)
131 * Given a map and a machine independent protection code,
132 * convert to a vax protection code.
134 #define pte_prot(m, p) \
135 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
136 static int protection_codes[8];
138 struct pmap kernel_pmap;
140 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
142 static struct vm_object kptobj;
146 static uint64_t KPDphys; /* phys addr of kernel level 2 */
147 uint64_t KPDPphys; /* phys addr of kernel level 3 */
148 uint64_t KPML4phys; /* phys addr of kernel level 4 */
150 extern int vmm_enabled;
151 extern void *vkernel_stack;
154 * Data for the pv entry allocation mechanism
156 static vm_zone_t pvzone;
157 static struct vm_zone pvzone_store;
158 static struct vm_object pvzone_obj;
159 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
160 static int pmap_pagedaemon_waken = 0;
161 static struct pv_entry *pvinit;
164 * All those kernel PT submaps that BSD is so fond of
166 pt_entry_t *CMAP1 = NULL, *ptmmap;
167 caddr_t CADDR1 = NULL;
168 static pt_entry_t *msgbufmap;
172 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
173 static pv_entry_t get_pv_entry (void);
174 static void i386_protection_init (void);
175 static __inline void pmap_clearbit (vm_page_t m, int bit);
177 static void pmap_remove_all (vm_page_t m);
178 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
180 static void pmap_remove_page (struct pmap *pmap, vm_offset_t va);
181 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
183 static boolean_t pmap_testbit (vm_page_t m, int bit);
184 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
185 vm_page_t mpte, vm_page_t m);
187 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
189 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
190 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
192 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
194 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
195 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t);
200 * Super fast pmap_pte routine best used when scanning the pv lists.
201 * This eliminates many course-grained invltlb calls. Note that many of
202 * the pv list scans are across different pmaps and it is very wasteful
203 * to do an entire invltlb when checking a single mapping.
205 * Should only be called while in a critical section.
208 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
211 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
213 return pmap_pte(pmap, va);
217 /* Return a non-clipped PD index for a given VA */
218 static __inline vm_pindex_t
219 pmap_pde_pindex(vm_offset_t va)
221 return va >> PDRSHIFT;
224 /* Return various clipped indexes for a given VA */
225 static __inline vm_pindex_t
226 pmap_pte_index(vm_offset_t va)
228 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
231 static __inline vm_pindex_t
232 pmap_pde_index(vm_offset_t va)
234 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
237 static __inline vm_pindex_t
238 pmap_pdpe_index(vm_offset_t va)
240 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
243 static __inline vm_pindex_t
244 pmap_pml4e_index(vm_offset_t va)
246 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
249 /* Return a pointer to the PML4 slot that corresponds to a VA */
250 static __inline pml4_entry_t *
251 pmap_pml4e(pmap_t pmap, vm_offset_t va)
253 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
256 /* Return a pointer to the PDP slot that corresponds to a VA */
257 static __inline pdp_entry_t *
258 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
262 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & VPTE_FRAME);
263 return (&pdpe[pmap_pdpe_index(va)]);
266 /* Return a pointer to the PDP slot that corresponds to a VA */
267 static __inline pdp_entry_t *
268 pmap_pdpe(pmap_t pmap, vm_offset_t va)
272 pml4e = pmap_pml4e(pmap, va);
273 if ((*pml4e & VPTE_V) == 0)
275 return (pmap_pml4e_to_pdpe(pml4e, va));
278 /* Return a pointer to the PD slot that corresponds to a VA */
279 static __inline pd_entry_t *
280 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
284 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & VPTE_FRAME);
285 return (&pde[pmap_pde_index(va)]);
288 /* Return a pointer to the PD slot that corresponds to a VA */
289 static __inline pd_entry_t *
290 pmap_pde(pmap_t pmap, vm_offset_t va)
294 pdpe = pmap_pdpe(pmap, va);
295 if (pdpe == NULL || (*pdpe & VPTE_V) == 0)
297 return (pmap_pdpe_to_pde(pdpe, va));
300 /* Return a pointer to the PT slot that corresponds to a VA */
301 static __inline pt_entry_t *
302 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
306 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & VPTE_FRAME);
307 return (&pte[pmap_pte_index(va)]);
310 /* Return a pointer to the PT slot that corresponds to a VA */
311 static __inline pt_entry_t *
312 pmap_pte(pmap_t pmap, vm_offset_t va)
316 pde = pmap_pde(pmap, va);
317 if (pde == NULL || (*pde & VPTE_V) == 0)
319 if ((*pde & VPTE_PS) != 0) /* compat with i386 pmap_pte() */
320 return ((pt_entry_t *)pde);
321 return (pmap_pde_to_pte(pde, va));
326 PMAP_INLINE pt_entry_t *
327 vtopte(vm_offset_t va)
329 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
330 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
332 return (PTmap + ((va >> PAGE_SHIFT) & mask));
335 static __inline pd_entry_t *
336 vtopde(vm_offset_t va)
338 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
339 NPML4EPGSHIFT)) - 1);
341 return (PDmap + ((va >> PDRSHIFT) & mask));
344 static PMAP_INLINE pt_entry_t *
345 vtopte(vm_offset_t va)
348 x = pmap_pte(&kernel_pmap, va);
353 static __inline pd_entry_t *
354 vtopde(vm_offset_t va)
357 x = pmap_pde(&kernel_pmap, va);
364 allocpages(vm_paddr_t *firstaddr, int n)
370 bzero((void *)ret, n * PAGE_SIZE);
372 *firstaddr += n * PAGE_SIZE;
377 create_dmap_vmm(vm_paddr_t *firstaddr)
380 int pml4_stack_index;
387 uint64_t KPDP_DMAP_phys = allocpages(firstaddr, NDMPML4E);
388 uint64_t KPDP_VSTACK_phys = allocpages(firstaddr, 1);
389 uint64_t KPD_VSTACK_phys = allocpages(firstaddr, 1);
391 pml4_entry_t *KPML4virt = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
392 pdp_entry_t *KPDP_DMAP_virt = (pdp_entry_t *)PHYS_TO_DMAP(KPDP_DMAP_phys);
393 pdp_entry_t *KPDP_VSTACK_virt = (pdp_entry_t *)PHYS_TO_DMAP(KPDP_VSTACK_phys);
394 pd_entry_t *KPD_VSTACK_virt = (pd_entry_t *)PHYS_TO_DMAP(KPD_VSTACK_phys);
396 bzero(KPDP_DMAP_virt, NDMPML4E * PAGE_SIZE);
397 bzero(KPDP_VSTACK_virt, 1 * PAGE_SIZE);
398 bzero(KPD_VSTACK_virt, 1 * PAGE_SIZE);
400 do_cpuid(0x80000001, regs);
401 amd_feature = regs[3];
403 /* Build the mappings for the first 512GB */
404 if (amd_feature & AMDID_PAGE1GB) {
405 /* In pages of 1 GB, if supported */
406 for (i = 0; i < NPDPEPG; i++) {
407 KPDP_DMAP_virt[i] = ((uint64_t)i << PDPSHIFT);
408 KPDP_DMAP_virt[i] |= VPTE_RW | VPTE_V | VPTE_PS | VPTE_U;
411 /* In page of 2MB, otherwise */
412 for (i = 0; i < NPDPEPG; i++) {
413 uint64_t KPD_DMAP_phys = allocpages(firstaddr, 1);
414 pd_entry_t *KPD_DMAP_virt = (pd_entry_t *)PHYS_TO_DMAP(KPD_DMAP_phys);
416 bzero(KPD_DMAP_virt, PAGE_SIZE);
418 KPDP_DMAP_virt[i] = KPD_DMAP_phys;
419 KPDP_DMAP_virt[i] |= VPTE_RW | VPTE_V | VPTE_U;
421 /* For each PD, we have to allocate NPTEPG PT */
422 for (j = 0; j < NPTEPG; j++) {
423 KPD_DMAP_virt[j] = (i << PDPSHIFT) | (j << PDRSHIFT);
424 KPD_DMAP_virt[j] |= VPTE_RW | VPTE_V | VPTE_PS | VPTE_U;
429 /* DMAP for the first 512G */
430 KPML4virt[0] = KPDP_DMAP_phys;
431 KPML4virt[0] |= VPTE_RW | VPTE_V | VPTE_U;
433 /* create a 2 MB map of the new stack */
434 pml4_stack_index = (uint64_t)&stack_addr >> PML4SHIFT;
435 KPML4virt[pml4_stack_index] = KPDP_VSTACK_phys;
436 KPML4virt[pml4_stack_index] |= VPTE_RW | VPTE_V | VPTE_U;
438 pdp_stack_index = ((uint64_t)&stack_addr & PML4MASK) >> PDPSHIFT;
439 KPDP_VSTACK_virt[pdp_stack_index] = KPD_VSTACK_phys;
440 KPDP_VSTACK_virt[pdp_stack_index] |= VPTE_RW | VPTE_V | VPTE_U;
442 pd_stack_index = ((uint64_t)&stack_addr & PDPMASK) >> PDRSHIFT;
443 KPD_VSTACK_virt[pd_stack_index] = (uint64_t) vkernel_stack;
444 KPD_VSTACK_virt[pd_stack_index] |= VPTE_RW | VPTE_V | VPTE_U | VPTE_PS;
448 create_pagetables(vm_paddr_t *firstaddr, int64_t ptov_offset)
451 pml4_entry_t *KPML4virt;
452 pdp_entry_t *KPDPvirt;
455 int kpml4i = pmap_pml4e_index(ptov_offset);
456 int kpdpi = pmap_pdpe_index(ptov_offset);
457 int kpdi = pmap_pde_index(ptov_offset);
460 * Calculate NKPT - number of kernel page tables. We have to
461 * accomodoate prealloction of the vm_page_array, dump bitmap,
462 * MSGBUF_SIZE, and other stuff. Be generous.
464 * Maxmem is in pages.
466 nkpt = (Maxmem * (sizeof(struct vm_page) * 2) + MSGBUF_SIZE) / NBPDR;
470 KPML4phys = allocpages(firstaddr, 1);
471 KPDPphys = allocpages(firstaddr, NKPML4E);
472 KPDphys = allocpages(firstaddr, NKPDPE);
473 KPTphys = allocpages(firstaddr, nkpt);
475 KPML4virt = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
476 KPDPvirt = (pdp_entry_t *)PHYS_TO_DMAP(KPDPphys);
477 KPDvirt = (pd_entry_t *)PHYS_TO_DMAP(KPDphys);
478 KPTvirt = (pt_entry_t *)PHYS_TO_DMAP(KPTphys);
480 bzero(KPML4virt, 1 * PAGE_SIZE);
481 bzero(KPDPvirt, NKPML4E * PAGE_SIZE);
482 bzero(KPDvirt, NKPDPE * PAGE_SIZE);
483 bzero(KPTvirt, nkpt * PAGE_SIZE);
485 /* Now map the page tables at their location within PTmap */
486 for (i = 0; i < nkpt; i++) {
487 KPDvirt[i + kpdi] = KPTphys + (i << PAGE_SHIFT);
488 KPDvirt[i + kpdi] |= VPTE_RW | VPTE_V | VPTE_U;
491 /* And connect up the PD to the PDP */
492 for (i = 0; i < NKPDPE; i++) {
493 KPDPvirt[i + kpdpi] = KPDphys + (i << PAGE_SHIFT);
494 KPDPvirt[i + kpdpi] |= VPTE_RW | VPTE_V | VPTE_U;
497 /* And recursively map PML4 to itself in order to get PTmap */
498 KPML4virt[PML4PML4I] = KPML4phys;
499 KPML4virt[PML4PML4I] |= VPTE_RW | VPTE_V | VPTE_U;
501 /* Connect the KVA slot up to the PML4 */
502 KPML4virt[kpml4i] = KPDPphys;
503 KPML4virt[kpml4i] |= VPTE_RW | VPTE_V | VPTE_U;
507 * Typically used to initialize a fictitious page by vm/device_pager.c
510 pmap_page_init(struct vm_page *m)
513 TAILQ_INIT(&m->md.pv_list);
517 * Bootstrap the system enough to run with virtual memory.
519 * On the i386 this is called after mapping has already been enabled
520 * and just syncs the pmap module with what has already been done.
521 * [We can't call it easily with mapping off since the kernel is not
522 * mapped with PA == VA, hence we would have to relocate every address
523 * from the linked base (virtual) address "KERNBASE" to the actual
524 * (physical) address starting relative to 0]
527 pmap_bootstrap(vm_paddr_t *firstaddr, int64_t ptov_offset)
533 * Create an initial set of page tables to run the kernel in.
535 create_pagetables(firstaddr, ptov_offset);
537 /* Create the DMAP for the VMM */
539 create_dmap_vmm(firstaddr);
542 virtual_start = KvaStart;
543 virtual_end = KvaEnd;
546 * Initialize protection array.
548 i386_protection_init();
551 * The kernel's pmap is statically allocated so we don't have to use
552 * pmap_create, which is unlikely to work correctly at this part of
553 * the boot sequence (XXX and which no longer exists).
555 * The kernel_pmap's pm_pteobj is used only for locking and not
558 kernel_pmap.pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
559 kernel_pmap.pm_count = 1;
560 /* don't allow deactivation */
561 CPUMASK_ASSALLONES(kernel_pmap.pm_active);
562 kernel_pmap.pm_pteobj = NULL; /* see pmap_init */
563 TAILQ_INIT(&kernel_pmap.pm_pvlist);
564 TAILQ_INIT(&kernel_pmap.pm_pvlist_free);
565 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
566 spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
569 * Reserve some special page table entries/VA space for temporary
572 #define SYSMAP(c, p, v, n) \
573 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
576 pte = pmap_pte(&kernel_pmap, va);
578 * CMAP1/CMAP2 are used for zeroing and copying pages.
580 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
586 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
590 * ptvmmap is used for reading arbitrary physical pages via
593 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
596 * msgbufp is used to map the system message buffer.
597 * XXX msgbufmap is not used.
599 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
600 atop(round_page(MSGBUF_SIZE)))
605 /* Not ready to do an invltlb yet for VMM*/
612 * Initialize the pmap module.
613 * Called by vm_init, to initialize any structures that the pmap
614 * system needs to map virtual memory.
615 * pmap_init has been enhanced to support in a fairly consistant
616 * way, discontiguous physical memory.
625 * object for kernel page table pages
627 /* JG I think the number can be arbitrary */
628 vm_object_init(&kptobj, 5);
629 kernel_pmap.pm_pteobj = &kptobj;
632 * Allocate memory for random pmap data structures. Includes the
635 for(i = 0; i < vm_page_array_size; i++) {
638 m = &vm_page_array[i];
639 TAILQ_INIT(&m->md.pv_list);
640 m->md.pv_list_count = 0;
644 * init the pv free list
646 initial_pvs = vm_page_array_size;
647 if (initial_pvs < MINPV)
649 pvzone = &pvzone_store;
650 pvinit = (struct pv_entry *)
651 kmem_alloc(&kernel_map,
652 initial_pvs * sizeof (struct pv_entry),
654 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
658 * Now it is safe to enable pv_table recording.
660 pmap_initialized = TRUE;
664 * Initialize the address space (zone) for the pv_entries. Set a
665 * high water mark so that the system can recover from excessive
666 * numbers of pv entries.
671 int shpgperproc = PMAP_SHPGPERPROC;
673 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
674 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
675 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
676 pv_entry_high_water = 9 * (pv_entry_max / 10);
677 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT);
681 /***************************************************
682 * Low level helper routines.....
683 ***************************************************/
686 * The modification bit is not tracked for any pages in this range. XXX
687 * such pages in this maps should always use pmap_k*() functions and not
690 * XXX User and kernel address spaces are independant for virtual kernels,
691 * this function only applies to the kernel pmap.
694 pmap_track_modified(pmap_t pmap, vm_offset_t va)
696 if (pmap != &kernel_pmap)
698 if ((va < clean_sva) || (va >= clean_eva))
705 * Extract the physical page address associated with the map/VA pair.
710 pmap_extract(pmap_t pmap, vm_offset_t va)
714 pd_entry_t pde, *pdep;
716 vm_object_hold(pmap->pm_pteobj);
718 pdep = pmap_pde(pmap, va);
722 if ((pde & VPTE_PS) != 0) {
724 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
726 pte = pmap_pde_to_pte(pdep, va);
727 rtval = (*pte & VPTE_FRAME) | (va & PAGE_MASK);
731 vm_object_drop(pmap->pm_pteobj);
737 * Similar to extract but checks protections, SMP-friendly short-cut for
738 * vm_fault_page[_quick]().
741 pmap_fault_page_quick(pmap_t pmap __unused, vm_offset_t vaddr __unused,
742 vm_prot_t prot __unused)
748 * Routine: pmap_kextract
750 * Extract the physical page address associated
751 * kernel virtual address.
754 pmap_kextract(vm_offset_t va)
759 KKASSERT(va >= KvaStart && va < KvaEnd);
762 * The DMAP region is not included in [KvaStart, KvaEnd)
765 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
766 pa = DMAP_TO_PHYS(va);
772 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
775 * Beware of a concurrent promotion that changes the
776 * PDE at this point! For example, vtopte() must not
777 * be used to access the PTE because it would use the
778 * new PDE. It is, however, safe to use the old PDE
779 * because the page table page is preserved by the
782 pa = *pmap_pde_to_pte(&pde, va);
783 pa = (pa & VPTE_FRAME) | (va & PAGE_MASK);
791 /***************************************************
792 * Low level mapping routines.....
793 ***************************************************/
796 * Enter a mapping into kernel_pmap. Mappings created in this fashion
797 * are not managed. Mappings must be immediately accessible on all cpus.
799 * Call pmap_inval_pte() to invalidate the virtual pte and clean out the
800 * real pmap and handle related races before storing the new vpte. The
801 * new semantics for kenter require use to do an UNCONDITIONAL invalidation,
802 * because the entry may have previously been cleared without an invalidation.
805 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
810 KKASSERT(va >= KvaStart && va < KvaEnd);
811 npte = pa | VPTE_RW | VPTE_V | VPTE_U;
816 pmap_inval_pte(pte, &kernel_pmap, va);
819 pmap_inval_pte(pte, &kernel_pmap, va);
825 * Enter an unmanaged KVA mapping for the private use of the current
828 * It is illegal for the mapping to be accessed by other cpus without
829 * proper invalidation.
832 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
838 KKASSERT(va >= KvaStart && va < KvaEnd);
840 npte = (vpte_t)pa | VPTE_RW | VPTE_V | VPTE_U;
850 pmap_inval_pte_quick(ptep, &kernel_pmap, va);
857 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
863 KKASSERT(va >= KvaStart && va < KvaEnd);
865 npte = (vpte_t)pa | VPTE_RW | VPTE_V | VPTE_U;
880 * Remove an unmanaged mapping created with pmap_kenter*().
883 pmap_kremove(vm_offset_t va)
887 KKASSERT(va >= KvaStart && va < KvaEnd);
891 pmap_inval_pte(pte, &kernel_pmap, va);
895 * Remove an unmanaged mapping created with pmap_kenter*() but synchronize
896 * only with this cpu.
898 * Unfortunately because we optimize new entries by testing VPTE_V later
899 * on, we actually still have to synchronize with all the cpus. XXX maybe
900 * store a junk value and test against 0 in the other places instead?
903 pmap_kremove_quick(vm_offset_t va)
907 KKASSERT(va >= KvaStart && va < KvaEnd);
911 pmap_inval_pte(pte, &kernel_pmap, va); /* NOT _quick */
915 pmap_kremove_noinval(vm_offset_t va)
919 KKASSERT(va >= KvaStart && va < KvaEnd);
926 * Used to map a range of physical addresses into kernel
927 * virtual address space.
929 * For now, VM is already on, we only need to map the
933 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
935 return PHYS_TO_DMAP(start);
939 * Map a set of unmanaged VM pages into KVM.
942 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
946 end_va = va + count * PAGE_SIZE;
947 KKASSERT(va >= KvaStart && end_va < KvaEnd);
949 while (va < end_va) {
954 pmap_inval_pte(pte, &kernel_pmap, va);
955 *pte = VM_PAGE_TO_PHYS(*m) | VPTE_RW | VPTE_V | VPTE_U;
962 * Undo the effects of pmap_qenter*().
965 pmap_qremove(vm_offset_t va, int count)
969 end_va = va + count * PAGE_SIZE;
970 KKASSERT(va >= KvaStart && end_va < KvaEnd);
972 while (va < end_va) {
976 atomic_swap_long(pte, 0);
977 pmap_inval_pte(pte, &kernel_pmap, va);
983 pmap_qremove_quick(vm_offset_t va, int count)
987 end_va = va + count * PAGE_SIZE;
988 KKASSERT(va >= KvaStart && end_va < KvaEnd);
990 while (va < end_va) {
994 atomic_swap_long(pte, 0);
995 cpu_invlpg((void *)va);
1001 pmap_qremove_noinval(vm_offset_t va, int count)
1005 end_va = va + count * PAGE_SIZE;
1006 KKASSERT(va >= KvaStart && end_va < KvaEnd);
1008 while (va < end_va) {
1012 atomic_swap_long(pte, 0);
1018 * This routine works like vm_page_lookup() but also blocks as long as the
1019 * page is busy. This routine does not busy the page it returns.
1021 * Unless the caller is managing objects whos pages are in a known state,
1022 * the call should be made with a critical section held so the page's object
1023 * association remains valid on return.
1026 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1030 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1031 m = vm_page_lookup_busy_wait(object, pindex, FALSE, "pplookp");
1037 * Create a new thread and optionally associate it with a (new) process.
1038 * NOTE! the new thread's cpu may not equal the current cpu.
1041 pmap_init_thread(thread_t td)
1043 /* enforce pcb placement */
1044 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1045 td->td_savefpu = &td->td_pcb->pcb_save;
1046 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on x86_64? */
1050 * This routine directly affects the fork perf for a process.
1053 pmap_init_proc(struct proc *p)
1057 /***************************************************
1058 * Page table page management routines.....
1059 ***************************************************/
1061 static __inline int pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va,
1065 * This routine unholds page table pages, and if the hold count
1066 * drops to zero, then it decrements the wire count.
1068 * We must recheck that this is the last hold reference after busy-sleeping
1072 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m)
1074 vm_page_busy_wait(m, FALSE, "pmuwpt");
1075 KASSERT(m->queue == PQ_NONE,
1076 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1078 if (m->hold_count == 1) {
1080 * Unmap the page table page.
1083 /* pmap_inval_add(info, pmap, -1); */
1085 if (m->pindex >= (NUPDE + NUPDPE)) {
1088 pml4 = pmap_pml4e(pmap, va);
1090 } else if (m->pindex >= NUPDE) {
1093 pdp = pmap_pdpe(pmap, va);
1098 pd = pmap_pde(pmap, va);
1102 KKASSERT(pmap->pm_stats.resident_count > 0);
1103 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1105 if (pmap->pm_ptphint == m)
1106 pmap->pm_ptphint = NULL;
1108 if (m->pindex < NUPDE) {
1109 /* We just released a PT, unhold the matching PD */
1112 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & VPTE_FRAME);
1113 pmap_unwire_pte_hold(pmap, va, pdpg);
1115 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
1116 /* We just released a PD, unhold the matching PDP */
1119 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & VPTE_FRAME);
1120 pmap_unwire_pte_hold(pmap, va, pdppg);
1124 * This was our last hold, the page had better be unwired
1125 * after we decrement wire_count.
1127 * FUTURE NOTE: shared page directory page could result in
1128 * multiple wire counts.
1132 KKASSERT(m->wire_count == 0);
1133 atomic_add_int(&vmstats.v_wire_count, -1);
1134 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1136 vm_page_free_zero(m);
1139 KKASSERT(m->hold_count > 1);
1147 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m)
1149 KKASSERT(m->hold_count > 0);
1150 if (m->hold_count > 1) {
1154 return _pmap_unwire_pte_hold(pmap, va, m);
1159 * After removing a page table entry, this routine is used to
1160 * conditionally free the page, and manage the hold/wire counts.
1163 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
1165 /* JG Use FreeBSD/amd64 or FreeBSD/i386 ptepde approaches? */
1166 vm_pindex_t ptepindex;
1168 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1172 * page table pages in the kernel_pmap are not managed.
1174 if (pmap == &kernel_pmap)
1176 ptepindex = pmap_pde_pindex(va);
1177 if (pmap->pm_ptphint &&
1178 (pmap->pm_ptphint->pindex == ptepindex)) {
1179 mpte = pmap->pm_ptphint;
1181 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1182 pmap->pm_ptphint = mpte;
1183 vm_page_wakeup(mpte);
1187 return pmap_unwire_pte_hold(pmap, va, mpte);
1191 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
1192 * just dummy it up so it works well enough for fork().
1194 * In DragonFly, process pmaps may only be used to manipulate user address
1195 * space, never kernel address space.
1198 pmap_pinit0(struct pmap *pmap)
1204 * Initialize a preallocated and zeroed pmap structure,
1205 * such as one in a vmspace structure.
1208 pmap_pinit(struct pmap *pmap)
1213 * No need to allocate page table space yet but we do need a valid
1214 * page directory table.
1216 if (pmap->pm_pml4 == NULL) {
1217 pmap->pm_pml4 = (pml4_entry_t *)
1218 kmem_alloc_pageable(&kernel_map, PAGE_SIZE,
1223 * Allocate an object for the ptes
1225 if (pmap->pm_pteobj == NULL)
1226 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1);
1229 * Allocate the page directory page, unless we already have
1230 * one cached. If we used the cached page the wire_count will
1231 * already be set appropriately.
1233 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1234 ptdpg = vm_page_grab(pmap->pm_pteobj,
1235 NUPDE + NUPDPE + PML4PML4I,
1236 VM_ALLOC_NORMAL | VM_ALLOC_RETRY |
1238 pmap->pm_pdirm = ptdpg;
1239 vm_page_flag_clear(ptdpg, PG_MAPPED);
1240 vm_page_wire(ptdpg);
1241 vm_page_wakeup(ptdpg);
1242 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1245 CPUMASK_ASSZERO(pmap->pm_active);
1246 pmap->pm_ptphint = NULL;
1247 TAILQ_INIT(&pmap->pm_pvlist);
1248 TAILQ_INIT(&pmap->pm_pvlist_free);
1249 spin_init(&pmap->pm_spin, "pmapinit");
1250 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1251 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1252 pmap->pm_stats.resident_count = 1;
1256 * Clean up a pmap structure so it can be physically freed. This routine
1257 * is called by the vmspace dtor function. A great deal of pmap data is
1258 * left passively mapped to improve vmspace management so we have a bit
1259 * of cleanup work to do here.
1264 pmap_puninit(pmap_t pmap)
1268 KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1269 if ((p = pmap->pm_pdirm) != NULL) {
1270 KKASSERT(pmap->pm_pml4 != NULL);
1271 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1272 vm_page_busy_wait(p, FALSE, "pgpun");
1274 atomic_add_int(&vmstats.v_wire_count, -1);
1275 vm_page_free_zero(p);
1276 pmap->pm_pdirm = NULL;
1278 if (pmap->pm_pml4) {
1279 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1280 pmap->pm_pml4 = NULL;
1282 if (pmap->pm_pteobj) {
1283 vm_object_deallocate(pmap->pm_pteobj);
1284 pmap->pm_pteobj = NULL;
1289 * This function is now unused (used to add the pmap to the pmap_list)
1292 pmap_pinit2(struct pmap *pmap)
1297 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1298 * 0 on failure (if the procedure had to sleep).
1300 * When asked to remove the page directory page itself, we actually just
1301 * leave it cached so we do not have to incur the SMP inval overhead of
1302 * removing the kernel mapping. pmap_puninit() will take care of it.
1305 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1308 * This code optimizes the case of freeing non-busy
1309 * page-table pages. Those pages are zero now, and
1310 * might as well be placed directly into the zero queue.
1312 if (vm_page_busy_try(p, FALSE)) {
1313 vm_page_sleep_busy(p, FALSE, "pmaprl");
1318 * Remove the page table page from the processes address space.
1320 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1322 * We are the pml4 table itself.
1324 /* XXX anything to do here? */
1325 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1327 * We are a PDP page.
1328 * We look for the PML4 entry that points to us.
1330 vm_page_t m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1331 KKASSERT(m4 != NULL);
1332 pml4_entry_t *pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1333 int idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1334 KKASSERT(pml4[idx] != 0);
1337 /* JG What about wire_count? */
1338 } else if (p->pindex >= NUPDE) {
1341 * We look for the PDP entry that points to us.
1343 vm_page_t m3 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1344 KKASSERT(m3 != NULL);
1345 pdp_entry_t *pdp = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1346 int idx = (p->pindex - NUPDE) % NPDPEPG;
1347 KKASSERT(pdp[idx] != 0);
1350 /* JG What about wire_count? */
1352 /* We are a PT page.
1353 * We look for the PD entry that points to us.
1355 vm_page_t m2 = vm_page_lookup(pmap->pm_pteobj, NUPDE + p->pindex / NPDEPG);
1356 KKASSERT(m2 != NULL);
1357 pd_entry_t *pd = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1358 int idx = p->pindex % NPDEPG;
1361 /* JG What about wire_count? */
1363 KKASSERT(pmap->pm_stats.resident_count > 0);
1364 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1366 if (p->hold_count) {
1367 panic("pmap_release: freeing held pt page "
1368 "pmap=%p pg=%p dmap=%p pi=%ld {%ld,%ld,%ld}",
1369 pmap, p, (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(p)),
1370 p->pindex, NUPDE, NUPDPE, PML4PML4I);
1372 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1373 pmap->pm_ptphint = NULL;
1376 * We leave the top-level page table page cached, wired, and mapped in
1377 * the pmap until the dtor function (pmap_puninit()) gets called.
1378 * However, still clean it up.
1380 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1381 bzero(pmap->pm_pml4, PAGE_SIZE);
1386 atomic_add_int(&vmstats.v_wire_count, -1);
1387 /* JG eventually revert to using vm_page_free_zero() */
1394 * this routine is called if the page table page is not
1398 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1400 vm_page_t m, pdppg, pdpg;
1403 * Find or fabricate a new pagetable page. Handle allocation
1404 * races by checking m->valid.
1406 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1407 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1409 KASSERT(m->queue == PQ_NONE,
1410 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1413 * Increment the hold count for the page we will be returning to
1420 * Map the pagetable page into the process address space, if
1421 * it isn't already there.
1423 atomic_add_long(&pmap->pm_stats.resident_count, 1);
1425 if (ptepindex >= (NUPDE + NUPDPE)) {
1427 vm_pindex_t pml4index;
1429 /* Wire up a new PDP page */
1430 pml4index = ptepindex - (NUPDE + NUPDPE);
1431 pml4 = &pmap->pm_pml4[pml4index];
1432 *pml4 = VM_PAGE_TO_PHYS(m) |
1433 VPTE_RW | VPTE_V | VPTE_U |
1435 } else if (ptepindex >= NUPDE) {
1436 vm_pindex_t pml4index;
1437 vm_pindex_t pdpindex;
1441 /* Wire up a new PD page */
1442 pdpindex = ptepindex - NUPDE;
1443 pml4index = pdpindex >> NPML4EPGSHIFT;
1445 pml4 = &pmap->pm_pml4[pml4index];
1446 if ((*pml4 & VPTE_V) == 0) {
1447 /* Have to allocate a new PDP page, recurse */
1448 if (_pmap_allocpte(pmap, NUPDE + NUPDPE + pml4index)
1455 /* Add reference to the PDP page */
1456 pdppg = PHYS_TO_VM_PAGE(*pml4 & VPTE_FRAME);
1457 pdppg->hold_count++;
1459 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & VPTE_FRAME);
1461 /* Now find the pdp page */
1462 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1463 KKASSERT(*pdp == 0); /* JG DEBUG64 */
1464 *pdp = VM_PAGE_TO_PHYS(m) | VPTE_RW | VPTE_V | VPTE_U |
1467 vm_pindex_t pml4index;
1468 vm_pindex_t pdpindex;
1473 /* Wire up a new PT page */
1474 pdpindex = ptepindex >> NPDPEPGSHIFT;
1475 pml4index = pdpindex >> NPML4EPGSHIFT;
1477 /* First, find the pdp and check that its valid. */
1478 pml4 = &pmap->pm_pml4[pml4index];
1479 if ((*pml4 & VPTE_V) == 0) {
1480 /* We miss a PDP page. We ultimately need a PD page.
1481 * Recursively allocating a PD page will allocate
1482 * the missing PDP page and will also allocate
1483 * the PD page we need.
1485 /* Have to allocate a new PD page, recurse */
1486 if (_pmap_allocpte(pmap, NUPDE + pdpindex)
1492 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & VPTE_FRAME);
1493 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1495 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & VPTE_FRAME);
1496 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1497 if ((*pdp & VPTE_V) == 0) {
1498 /* Have to allocate a new PD page, recurse */
1499 if (_pmap_allocpte(pmap, NUPDE + pdpindex)
1506 /* Add reference to the PD page */
1507 pdpg = PHYS_TO_VM_PAGE(*pdp & VPTE_FRAME);
1511 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & VPTE_FRAME);
1513 /* Now we know where the page directory page is */
1514 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1515 KKASSERT(*pd == 0); /* JG DEBUG64 */
1516 *pd = VM_PAGE_TO_PHYS(m) | VPTE_RW | VPTE_V | VPTE_U |
1521 * Set the page table hint
1523 pmap->pm_ptphint = m;
1524 vm_page_flag_set(m, PG_MAPPED);
1531 * Determine the page table page required to access the VA in the pmap
1532 * and allocate it if necessary. Return a held vm_page_t for the page.
1534 * Only used with user pmaps.
1537 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1539 vm_pindex_t ptepindex;
1543 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1546 * Calculate pagetable page index
1548 ptepindex = pmap_pde_pindex(va);
1551 * Get the page directory entry
1553 pd = pmap_pde(pmap, va);
1556 * This supports switching from a 2MB page to a
1559 if (pd != NULL && (*pd & (VPTE_PS | VPTE_V)) == (VPTE_PS | VPTE_V)) {
1560 panic("no promotion/demotion yet");
1568 * If the page table page is mapped, we just increment the
1569 * hold count, and activate it.
1571 if (pd != NULL && (*pd & VPTE_V) != 0) {
1572 /* YYY hint is used here on i386 */
1573 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1574 pmap->pm_ptphint = m;
1580 * Here if the pte page isn't mapped, or if it has been deallocated.
1582 return _pmap_allocpte(pmap, ptepindex);
1586 /***************************************************
1587 * Pmap allocation/deallocation routines.
1588 ***************************************************/
1591 * Release any resources held by the given physical map.
1592 * Called when a pmap initialized by pmap_pinit is being released.
1593 * Should only be called if the map contains no valid mappings.
1595 * Caller must hold pmap->pm_token
1597 static int pmap_release_callback(struct vm_page *p, void *data);
1600 pmap_release(struct pmap *pmap)
1602 vm_object_t object = pmap->pm_pteobj;
1603 struct rb_vm_page_scan_info info;
1605 KKASSERT(pmap != &kernel_pmap);
1607 #if defined(DIAGNOSTIC)
1608 if (object->ref_count != 1)
1609 panic("pmap_release: pteobj reference count != 1");
1613 info.object = object;
1615 KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
1616 ("pmap %p still active! %016jx",
1618 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
1620 vm_object_hold(object);
1624 info.limit = object->generation;
1626 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1627 pmap_release_callback, &info);
1628 if (info.error == 0 && info.mpte) {
1629 if (!pmap_release_free_page(pmap, info.mpte))
1632 } while (info.error);
1633 vm_object_drop(object);
1637 pmap_release_callback(struct vm_page *p, void *data)
1639 struct rb_vm_page_scan_info *info = data;
1641 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1645 if (!pmap_release_free_page(info->pmap, p)) {
1649 if (info->object->generation != info->limit) {
1657 * Grow the number of kernel page table entries, if needed.
1662 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1666 vm_offset_t ptppaddr;
1668 pd_entry_t *pde, newpdir;
1673 vm_object_hold(&kptobj);
1674 if (kernel_vm_end == 0) {
1675 kernel_vm_end = KvaStart;
1677 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & VPTE_V) != 0) {
1678 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1680 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1681 kernel_vm_end = kernel_map.max_offset;
1686 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1687 if (addr - 1 >= kernel_map.max_offset)
1688 addr = kernel_map.max_offset;
1689 while (kernel_vm_end < addr) {
1690 pde = pmap_pde(&kernel_pmap, kernel_vm_end);
1692 /* We need a new PDP entry */
1693 nkpg = vm_page_alloc(&kptobj, nkpt,
1694 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM
1695 | VM_ALLOC_INTERRUPT);
1697 panic("pmap_growkernel: no memory to "
1700 paddr = VM_PAGE_TO_PHYS(nkpg);
1701 pmap_zero_page(paddr);
1702 newpdp = (pdp_entry_t)(paddr |
1703 VPTE_V | VPTE_RW | VPTE_U |
1705 *pmap_pdpe(&kernel_pmap, kernel_vm_end) = newpdp;
1707 continue; /* try again */
1709 if ((*pde & VPTE_V) != 0) {
1710 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1711 ~(PAGE_SIZE * NPTEPG - 1);
1712 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1713 kernel_vm_end = kernel_map.max_offset;
1720 * This index is bogus, but out of the way
1722 nkpg = vm_page_alloc(&kptobj, nkpt,
1725 VM_ALLOC_INTERRUPT);
1727 panic("pmap_growkernel: no memory to grow kernel");
1730 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1731 pmap_zero_page(ptppaddr);
1732 newpdir = (pd_entry_t)(ptppaddr |
1733 VPTE_V | VPTE_RW | VPTE_U |
1735 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1738 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1739 ~(PAGE_SIZE * NPTEPG - 1);
1740 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1741 kernel_vm_end = kernel_map.max_offset;
1745 vm_object_drop(&kptobj);
1749 * Add a reference to the specified pmap.
1754 pmap_reference(pmap_t pmap)
1757 atomic_add_int(&pmap->pm_count, 1);
1760 /************************************************************************
1761 * VMSPACE MANAGEMENT *
1762 ************************************************************************
1764 * The VMSPACE management we do in our virtual kernel must be reflected
1765 * in the real kernel. This is accomplished by making vmspace system
1766 * calls to the real kernel.
1769 cpu_vmspace_alloc(struct vmspace *vm)
1776 * If VMM enable, don't do nothing, we
1777 * are able to use real page tables
1782 #define USER_SIZE (VM_MAX_USER_ADDRESS - VM_MIN_USER_ADDRESS)
1784 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
1785 panic("vmspace_create() failed");
1787 rp = vmspace_mmap(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE,
1788 PROT_READ|PROT_WRITE,
1789 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
1791 if (rp == MAP_FAILED)
1792 panic("vmspace_mmap: failed");
1793 vmspace_mcontrol(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE,
1795 vpte = VM_PAGE_TO_PHYS(vmspace_pmap(vm)->pm_pdirm) | VPTE_RW | VPTE_V | VPTE_U;
1796 r = vmspace_mcontrol(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE,
1799 panic("vmspace_mcontrol: failed");
1803 cpu_vmspace_free(struct vmspace *vm)
1806 * If VMM enable, don't do nothing, we
1807 * are able to use real page tables
1812 if (vmspace_destroy(&vm->vm_pmap) < 0)
1813 panic("vmspace_destroy() failed");
1816 /***************************************************
1817 * page management routines.
1818 ***************************************************/
1821 * free the pv_entry back to the free list. This function may be
1822 * called from an interrupt.
1824 static __inline void
1825 free_pv_entry(pv_entry_t pv)
1828 KKASSERT(pv_entry_count >= 0);
1833 * get a new pv_entry, allocating a block from the system
1834 * when needed. This function may be called from an interrupt.
1840 if (pv_entry_high_water &&
1841 (pv_entry_count > pv_entry_high_water) &&
1842 (pmap_pagedaemon_waken == 0)) {
1843 pmap_pagedaemon_waken = 1;
1844 wakeup(&vm_pages_needed);
1846 return zalloc(pvzone);
1850 * This routine is very drastic, but can save the system
1860 static int warningdone=0;
1862 if (pmap_pagedaemon_waken == 0)
1864 pmap_pagedaemon_waken = 0;
1866 if (warningdone < 5) {
1867 kprintf("pmap_collect: collecting pv entries -- "
1868 "suggest increasing PMAP_SHPGPERPROC\n");
1872 for (i = 0; i < vm_page_array_size; i++) {
1873 m = &vm_page_array[i];
1874 if (m->wire_count || m->hold_count)
1876 if (vm_page_busy_try(m, TRUE) == 0) {
1877 if (m->wire_count == 0 && m->hold_count == 0) {
1887 * If it is the first entry on the list, it is actually
1888 * in the header and we must copy the following entry up
1889 * to the header. Otherwise we must search the list for
1890 * the entry. In either case we free the now unused entry.
1892 * pmap->pm_pteobj must be held and (m) must be spin-locked by the caller.
1895 pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va)
1900 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1901 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1902 if (pmap == pv->pv_pmap && va == pv->pv_va)
1906 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1907 if (va == pv->pv_va)
1913 * Note that pv_ptem is NULL if the page table page itself is not
1914 * managed, even if the page being removed IS managed.
1917 /* JGXXX When can 'pv' be NULL? */
1919 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1920 m->md.pv_list_count--;
1921 atomic_add_int(&m->object->agg_pv_list_count, -1);
1922 KKASSERT(m->md.pv_list_count >= 0);
1923 if (TAILQ_EMPTY(&m->md.pv_list))
1924 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1925 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1926 atomic_add_int(&pmap->pm_generation, 1);
1927 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1934 * Create a pv entry for page at pa for (pmap, va). If the page table page
1935 * holding the VA is managed, mpte will be non-NULL.
1937 * pmap->pm_pteobj must be held and (m) must be spin-locked by the caller.
1940 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1944 pv = get_pv_entry();
1949 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1950 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1951 m->md.pv_list_count++;
1952 atomic_add_int(&m->object->agg_pv_list_count, 1);
1956 * pmap_remove_pte: do the things to unmap a page in a process
1958 * Caller holds pmap->pm_pteobj
1961 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va)
1967 oldpte = pmap_inval_loadandclear(ptq, pmap, va);
1968 if (oldpte & VPTE_WIRED)
1969 atomic_add_long(&pmap->pm_stats.wired_count, -1);
1970 KKASSERT(pmap->pm_stats.wired_count >= 0);
1974 * Machines that don't support invlpg, also don't support
1975 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1979 cpu_invlpg((void *)va);
1981 KKASSERT(pmap->pm_stats.resident_count > 0);
1982 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1983 if (oldpte & VPTE_MANAGED) {
1984 m = PHYS_TO_VM_PAGE(oldpte);
1985 vm_page_spin_lock(m);
1986 if (oldpte & VPTE_M) {
1987 #if defined(PMAP_DIAGNOSTIC)
1988 if (pmap_nw_modified(oldpte)) {
1989 kprintf("pmap_remove: modified page not "
1990 "writable: va: 0x%lx, pte: 0x%lx\n",
1994 if (pmap_track_modified(pmap, va))
1997 if (oldpte & VPTE_A)
1998 vm_page_flag_set(m, PG_REFERENCED);
1999 error = pmap_remove_entry(pmap, m, va);
2000 vm_page_spin_unlock(m);
2002 error = pmap_unuse_pt(pmap, va, NULL);
2010 * Remove a single page from a process address space.
2012 * This function may not be called from an interrupt if the pmap is
2015 * Caller holds pmap->pm_pteobj
2018 pmap_remove_page(struct pmap *pmap, vm_offset_t va)
2022 pte = pmap_pte(pmap, va);
2025 if ((*pte & VPTE_V) == 0)
2027 pmap_remove_pte(pmap, pte, va);
2031 * Remove the given range of addresses from the specified map.
2033 * It is assumed that the start and end are properly rounded to
2036 * This function may not be called from an interrupt if the pmap is
2042 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2044 vm_offset_t va_next;
2045 pml4_entry_t *pml4e;
2047 pd_entry_t ptpaddr, *pde;
2053 vm_object_hold(pmap->pm_pteobj);
2054 KKASSERT(pmap->pm_stats.resident_count >= 0);
2055 if (pmap->pm_stats.resident_count == 0) {
2056 vm_object_drop(pmap->pm_pteobj);
2061 * special handling of removing one page. a very
2062 * common operation and easy to short circuit some
2065 if (sva + PAGE_SIZE == eva) {
2066 pde = pmap_pde(pmap, sva);
2067 if (pde && (*pde & VPTE_PS) == 0) {
2068 pmap_remove_page(pmap, sva);
2069 vm_object_drop(pmap->pm_pteobj);
2074 for (; sva < eva; sva = va_next) {
2075 pml4e = pmap_pml4e(pmap, sva);
2076 if ((*pml4e & VPTE_V) == 0) {
2077 va_next = (sva + NBPML4) & ~PML4MASK;
2083 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2084 if ((*pdpe & VPTE_V) == 0) {
2085 va_next = (sva + NBPDP) & ~PDPMASK;
2092 * Calculate index for next page table.
2094 va_next = (sva + NBPDR) & ~PDRMASK;
2098 pde = pmap_pdpe_to_pde(pdpe, sva);
2102 * Weed out invalid mappings.
2108 * Check for large page.
2110 if ((ptpaddr & VPTE_PS) != 0) {
2111 /* JG FreeBSD has more complex treatment here */
2112 KKASSERT(*pde != 0);
2113 pmap_inval_pde(pde, pmap, sva);
2114 atomic_add_long(&pmap->pm_stats.resident_count,
2115 -NBPDR / PAGE_SIZE);
2120 * Limit our scan to either the end of the va represented
2121 * by the current page table page, or to the end of the
2122 * range being removed.
2128 * NOTE: pmap_remove_pte() can block.
2130 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2134 if (pmap_remove_pte(pmap, pte, sva))
2138 vm_object_drop(pmap->pm_pteobj);
2142 * Removes this physical page from all physical maps in which it resides.
2143 * Reflects back modify bits to the pager.
2145 * This routine may not be called from an interrupt.
2150 pmap_remove_all(vm_page_t m)
2152 pt_entry_t *pte, tpte;
2157 #if defined(PMAP_DIAGNOSTIC)
2159 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
2162 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
2163 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
2168 vm_page_spin_lock(m);
2169 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2171 pmobj = pmap->pm_pteobj;
2174 * Handle reversed lock ordering
2176 if (vm_object_hold_try(pmobj) == 0) {
2177 refcount_acquire(&pmobj->hold_count);
2178 vm_page_spin_unlock(m);
2179 vm_object_lock(pmobj);
2180 vm_page_spin_lock(m);
2181 if (pv != TAILQ_FIRST(&m->md.pv_list) ||
2182 pmap != pv->pv_pmap ||
2183 pmobj != pmap->pm_pteobj) {
2184 vm_page_spin_unlock(m);
2185 vm_object_drop(pmobj);
2190 KKASSERT(pmap->pm_stats.resident_count > 0);
2191 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2193 pte = pmap_pte(pmap, pv->pv_va);
2194 KKASSERT(pte != NULL);
2196 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2197 if (tpte & VPTE_WIRED)
2198 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2199 KKASSERT(pmap->pm_stats.wired_count >= 0);
2202 vm_page_flag_set(m, PG_REFERENCED);
2205 * Update the vm_page_t clean and reference bits.
2207 if (tpte & VPTE_M) {
2208 #if defined(PMAP_DIAGNOSTIC)
2209 if (pmap_nw_modified(tpte)) {
2211 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2215 if (pmap_track_modified(pmap, pv->pv_va))
2218 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2219 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2220 atomic_add_int(&pmap->pm_generation, 1);
2221 m->md.pv_list_count--;
2222 atomic_add_int(&m->object->agg_pv_list_count, -1);
2223 KKASSERT(m->md.pv_list_count >= 0);
2224 if (TAILQ_EMPTY(&m->md.pv_list))
2225 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2226 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2227 vm_object_drop(pmobj);
2230 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2231 vm_page_spin_unlock(m);
2235 * Removes the page from a particular pmap
2238 pmap_remove_specific(pmap_t pmap, vm_page_t m)
2240 pt_entry_t *pte, tpte;
2243 vm_object_hold(pmap->pm_pteobj);
2245 vm_page_spin_lock(m);
2246 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2247 if (pv->pv_pmap != pmap)
2250 KKASSERT(pmap->pm_stats.resident_count > 0);
2251 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2253 pte = pmap_pte(pmap, pv->pv_va);
2254 KKASSERT(pte != NULL);
2256 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2257 if (tpte & VPTE_WIRED)
2258 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2259 KKASSERT(pmap->pm_stats.wired_count >= 0);
2262 vm_page_flag_set(m, PG_REFERENCED);
2265 * Update the vm_page_t clean and reference bits.
2267 if (tpte & VPTE_M) {
2268 if (pmap_track_modified(pmap, pv->pv_va))
2271 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2272 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2273 atomic_add_int(&pmap->pm_generation, 1);
2274 m->md.pv_list_count--;
2275 atomic_add_int(&m->object->agg_pv_list_count, -1);
2276 KKASSERT(m->md.pv_list_count >= 0);
2277 if (TAILQ_EMPTY(&m->md.pv_list))
2278 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2279 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2280 vm_page_spin_unlock(m);
2284 vm_page_spin_unlock(m);
2285 vm_object_drop(pmap->pm_pteobj);
2289 * Set the physical protection on the specified range of this map
2292 * This function may not be called from an interrupt if the map is
2293 * not the kernel_pmap.
2298 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2300 vm_offset_t va_next;
2301 pml4_entry_t *pml4e;
2303 pd_entry_t ptpaddr, *pde;
2306 /* JG review for NX */
2311 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2312 pmap_remove(pmap, sva, eva);
2316 if (prot & VM_PROT_WRITE)
2319 vm_object_hold(pmap->pm_pteobj);
2321 for (; sva < eva; sva = va_next) {
2322 pml4e = pmap_pml4e(pmap, sva);
2323 if ((*pml4e & VPTE_V) == 0) {
2324 va_next = (sva + NBPML4) & ~PML4MASK;
2330 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2331 if ((*pdpe & VPTE_V) == 0) {
2332 va_next = (sva + NBPDP) & ~PDPMASK;
2338 va_next = (sva + NBPDR) & ~PDRMASK;
2342 pde = pmap_pdpe_to_pde(pdpe, sva);
2346 * Check for large page.
2348 if ((ptpaddr & VPTE_PS) != 0) {
2350 pmap_clean_pde(pde, pmap, sva);
2351 atomic_add_long(&pmap->pm_stats.resident_count,
2352 -NBPDR / PAGE_SIZE);
2357 * Weed out invalid mappings. Note: we assume that the page
2358 * directory table is always allocated, and in kernel virtual.
2366 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2372 * Clean managed pages and also check the accessed
2373 * bit. Just remove write perms for unmanaged
2374 * pages. Be careful of races, turning off write
2375 * access will force a fault rather then setting
2376 * the modified bit at an unexpected time.
2378 if (*pte & VPTE_MANAGED) {
2379 pbits = pmap_clean_pte(pte, pmap, sva);
2381 if (pbits & VPTE_A) {
2382 m = PHYS_TO_VM_PAGE(pbits & VPTE_FRAME);
2383 vm_page_flag_set(m, PG_REFERENCED);
2384 atomic_clear_long(pte, VPTE_A);
2386 if (pbits & VPTE_M) {
2387 if (pmap_track_modified(pmap, sva)) {
2389 m = PHYS_TO_VM_PAGE(pbits & VPTE_FRAME);
2394 pbits = pmap_setro_pte(pte, pmap, sva);
2398 vm_object_drop(pmap->pm_pteobj);
2402 * Enter a managed page into a pmap. If the page is not wired related pmap
2403 * data can be destroyed at any time for later demand-operation.
2405 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
2406 * specified protection, and wire the mapping if requested.
2408 * NOTE: This routine may not lazy-evaluate or lose information. The
2409 * page must actually be inserted into the given map NOW.
2411 * NOTE: When entering a page at a KVA address, the pmap must be the
2417 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2418 boolean_t wired, vm_map_entry_t entry __unused)
2424 pt_entry_t origpte, newpte;
2430 va = trunc_page(va);
2432 vm_object_hold(pmap->pm_pteobj);
2435 * Get the page table page. The kernel_pmap's page table pages
2436 * are preallocated and have no associated vm_page_t.
2438 if (pmap == &kernel_pmap)
2441 mpte = pmap_allocpte(pmap, va);
2443 pde = pmap_pde(pmap, va);
2444 if (pde != NULL && (*pde & VPTE_V) != 0) {
2445 if ((*pde & VPTE_PS) != 0)
2446 panic("pmap_enter: attempted pmap_enter on 2MB page");
2447 pte = pmap_pde_to_pte(pde, va);
2449 panic("pmap_enter: invalid page directory va=%#lx", va);
2452 KKASSERT(pte != NULL);
2454 * Deal with races on the original mapping (though don't worry
2455 * about VPTE_A races) by cleaning it. This will force a fault
2456 * if an attempt is made to write to the page.
2458 pa = VM_PAGE_TO_PHYS(m);
2459 origpte = pmap_clean_pte(pte, pmap, va);
2460 opa = origpte & VPTE_FRAME;
2462 if (origpte & VPTE_PS)
2463 panic("pmap_enter: attempted pmap_enter on 2MB page");
2466 * Mapping has not changed, must be protection or wiring change.
2468 if (origpte && (opa == pa)) {
2470 * Wiring change, just update stats. We don't worry about
2471 * wiring PT pages as they remain resident as long as there
2472 * are valid mappings in them. Hence, if a user page is wired,
2473 * the PT page will be also.
2475 if (wired && ((origpte & VPTE_WIRED) == 0))
2476 atomic_add_long(&pmap->pm_stats.wired_count, 1);
2477 else if (!wired && (origpte & VPTE_WIRED))
2478 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2481 * Remove the extra pte reference. Note that we cannot
2482 * optimize the RO->RW case because we have adjusted the
2483 * wiring count above and may need to adjust the wiring
2490 * We might be turning off write access to the page,
2491 * so we go ahead and sense modify status.
2493 if (origpte & VPTE_MANAGED) {
2494 if ((origpte & VPTE_M) &&
2495 pmap_track_modified(pmap, va)) {
2497 om = PHYS_TO_VM_PAGE(opa);
2501 KKASSERT(m->flags & PG_MAPPED);
2506 * Mapping has changed, invalidate old range and fall through to
2507 * handle validating new mapping.
2511 err = pmap_remove_pte(pmap, pte, va);
2513 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2517 * Enter on the PV list if part of our managed memory. Note that we
2518 * raise IPL while manipulating pv_table since pmap_enter can be
2519 * called at interrupt time.
2521 vm_page_spin_lock(m);
2522 if (pmap_initialized &&
2523 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2524 pmap_insert_entry(pmap, va, mpte, m);
2526 vm_page_flag_set(m, PG_MAPPED);
2528 vm_page_spin_unlock(m);
2531 * Increment counters
2533 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2535 atomic_add_long(&pmap->pm_stats.wired_count, 1);
2539 * Now validate mapping with desired protection/wiring.
2541 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | VPTE_V | VPTE_U);
2544 newpte |= VPTE_WIRED;
2545 // if (pmap != &kernel_pmap)
2549 * If the mapping or permission bits are different from the
2550 * (now cleaned) original pte, an update is needed. We've
2551 * already downgraded or invalidated the page so all we have
2552 * to do now is update the bits.
2554 * XXX should we synchronize RO->RW changes to avoid another
2557 if ((origpte & ~(VPTE_RW|VPTE_M|VPTE_A)) != newpte) {
2558 *pte = newpte | VPTE_A;
2559 if (newpte & VPTE_RW)
2560 vm_page_flag_set(m, PG_WRITEABLE);
2562 KKASSERT((newpte & VPTE_MANAGED) == 0 || (m->flags & PG_MAPPED));
2563 vm_object_drop(pmap->pm_pteobj);
2567 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2569 * Currently this routine may only be used on user pmaps, not kernel_pmap.
2574 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2579 vm_pindex_t ptepindex;
2582 KKASSERT(pmap != &kernel_pmap);
2584 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
2587 * Calculate pagetable page index
2589 ptepindex = pmap_pde_pindex(va);
2591 vm_object_hold(pmap->pm_pteobj);
2595 * Get the page directory entry
2597 ptepa = pmap_pde(pmap, va);
2600 * If the page table page is mapped, we just increment
2601 * the hold count, and activate it.
2603 if (ptepa && (*ptepa & VPTE_V) != 0) {
2604 if (*ptepa & VPTE_PS)
2605 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2606 if (pmap->pm_ptphint &&
2607 (pmap->pm_ptphint->pindex == ptepindex)) {
2608 mpte = pmap->pm_ptphint;
2610 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2611 pmap->pm_ptphint = mpte;
2612 vm_page_wakeup(mpte);
2617 mpte = _pmap_allocpte(pmap, ptepindex);
2619 } while (mpte == NULL);
2622 * Ok, now that the page table page has been validated, get the pte.
2623 * If the pte is already mapped undo mpte's hold_count and
2626 pte = pmap_pte(pmap, va);
2627 if (*pte & VPTE_V) {
2628 KKASSERT(mpte != NULL);
2629 pmap_unwire_pte_hold(pmap, va, mpte);
2630 pa = VM_PAGE_TO_PHYS(m);
2631 KKASSERT(((*pte ^ pa) & VPTE_FRAME) == 0);
2632 vm_object_drop(pmap->pm_pteobj);
2637 * Enter on the PV list if part of our managed memory
2639 vm_page_spin_lock(m);
2640 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2641 pmap_insert_entry(pmap, va, mpte, m);
2642 vm_page_flag_set(m, PG_MAPPED);
2644 vm_page_spin_unlock(m);
2647 * Increment counters
2649 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2651 pa = VM_PAGE_TO_PHYS(m);
2654 * Now validate mapping with RO protection
2656 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2657 *pte = (vpte_t)pa | VPTE_V | VPTE_U;
2659 *pte = (vpte_t)pa | VPTE_V | VPTE_U | VPTE_MANAGED;
2660 /*pmap_inval_add(&info, pmap, va); shouldn't be needed 0->valid */
2661 /*pmap_inval_flush(&info); don't need for vkernel */
2662 vm_object_drop(pmap->pm_pteobj);
2666 * Make a temporary mapping for a physical address. This is only intended
2667 * to be used for panic dumps.
2669 * The caller is responsible for calling smp_invltlb().
2672 pmap_kenter_temporary(vm_paddr_t pa, long i)
2674 pmap_kenter_quick(crashdumpmap + (i * PAGE_SIZE), pa);
2675 return ((void *)crashdumpmap);
2678 #define MAX_INIT_PT (96)
2681 * This routine preloads the ptes for a given object into the specified pmap.
2682 * This eliminates the blast of soft faults on process startup and
2683 * immediately after an mmap.
2687 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2690 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2691 vm_object_t object, vm_pindex_t pindex,
2692 vm_size_t size, int limit)
2694 struct rb_vm_page_scan_info info;
2699 * We can't preinit if read access isn't set or there is no pmap
2702 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2706 * We can't preinit if the pmap is not the current pmap
2708 lp = curthread->td_lwp;
2709 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2712 psize = x86_64_btop(size);
2714 if ((object->type != OBJT_VNODE) ||
2715 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2716 (object->resident_page_count > MAX_INIT_PT))) {
2720 if (psize + pindex > object->size) {
2721 if (object->size < pindex)
2723 psize = object->size - pindex;
2730 * Use a red-black scan to traverse the requested range and load
2731 * any valid pages found into the pmap.
2733 * We cannot safely scan the object's memq unless we are in a
2734 * critical section since interrupts can remove pages from objects.
2736 info.start_pindex = pindex;
2737 info.end_pindex = pindex + psize - 1;
2743 vm_object_hold_shared(object);
2744 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2745 pmap_object_init_pt_callback, &info);
2746 vm_object_drop(object);
2751 pmap_object_init_pt_callback(vm_page_t p, void *data)
2753 struct rb_vm_page_scan_info *info = data;
2754 vm_pindex_t rel_index;
2756 * don't allow an madvise to blow away our really
2757 * free pages allocating pv entries.
2759 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2760 vmstats.v_free_count < vmstats.v_free_reserved) {
2765 * Ignore list markers and ignore pages we cannot instantly
2766 * busy (while holding the object token).
2768 if (p->flags & PG_MARKER)
2770 if (vm_page_busy_try(p, TRUE))
2772 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2773 (p->flags & PG_FICTITIOUS) == 0) {
2774 if ((p->queue - p->pc) == PQ_CACHE)
2775 vm_page_deactivate(p);
2776 rel_index = p->pindex - info->start_pindex;
2777 pmap_enter_quick(info->pmap,
2778 info->addr + x86_64_ptob(rel_index), p);
2785 * Return TRUE if the pmap is in shape to trivially
2786 * pre-fault the specified address.
2788 * Returns FALSE if it would be non-trivial or if a
2789 * pte is already loaded into the slot.
2794 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2800 vm_object_hold(pmap->pm_pteobj);
2801 pde = pmap_pde(pmap, addr);
2802 if (pde == NULL || *pde == 0) {
2805 pte = pmap_pde_to_pte(pde, addr);
2806 ret = (*pte) ? 0 : 1;
2808 vm_object_drop(pmap->pm_pteobj);
2814 * Change the wiring attribute for a map/virtual-address pair.
2816 * The mapping must already exist in the pmap.
2817 * No other requirements.
2820 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
2821 vm_map_entry_t entry __unused)
2828 vm_object_hold(pmap->pm_pteobj);
2829 pte = pmap_pte(pmap, va);
2831 if (wired && !pmap_pte_w(pte))
2832 atomic_add_long(&pmap->pm_stats.wired_count, 1);
2833 else if (!wired && pmap_pte_w(pte))
2834 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2837 * Wiring is not a hardware characteristic so there is no need to
2838 * invalidate TLB. However, in an SMP environment we must use
2839 * a locked bus cycle to update the pte (if we are not using
2840 * the pmap_inval_*() API that is)... it's ok to do this for simple
2844 atomic_set_long(pte, VPTE_WIRED);
2846 atomic_clear_long(pte, VPTE_WIRED);
2847 vm_object_drop(pmap->pm_pteobj);
2851 * Copy the range specified by src_addr/len
2852 * from the source map to the range dst_addr/len
2853 * in the destination map.
2855 * This routine is only advisory and need not do anything.
2858 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2859 vm_size_t len, vm_offset_t src_addr)
2862 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2863 * valid through blocking calls, and that's just not going to
2874 * Zero the specified physical page.
2876 * This function may be called from an interrupt and no locking is
2880 pmap_zero_page(vm_paddr_t phys)
2882 vm_offset_t va = PHYS_TO_DMAP(phys);
2884 bzero((void *)va, PAGE_SIZE);
2890 * Zero part of a physical page by mapping it into memory and clearing
2891 * its contents with bzero.
2893 * off and size may not cover an area beyond a single hardware page.
2896 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2899 vm_offset_t virt = PHYS_TO_DMAP(phys);
2900 bzero((char *)virt + off, size);
2907 * Copy the physical page from the source PA to the target PA.
2908 * This function may be called from an interrupt. No locking
2912 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2914 vm_offset_t src_virt, dst_virt;
2917 src_virt = PHYS_TO_DMAP(src);
2918 dst_virt = PHYS_TO_DMAP(dst);
2919 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
2924 * pmap_copy_page_frag:
2926 * Copy the physical page from the source PA to the target PA.
2927 * This function may be called from an interrupt. No locking
2931 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2933 vm_offset_t src_virt, dst_virt;
2936 src_virt = PHYS_TO_DMAP(src);
2937 dst_virt = PHYS_TO_DMAP(dst);
2938 bcopy((char *)src_virt + (src & PAGE_MASK),
2939 (char *)dst_virt + (dst & PAGE_MASK),
2945 * Returns true if the pmap's pv is one of the first 16 pvs linked to
2946 * from this page. This count may be changed upwards or downwards
2947 * in the future; it is only necessary that true be returned for a small
2948 * subset of pmaps for proper page aging.
2950 * No other requirements.
2953 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2958 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2961 vm_page_spin_lock(m);
2962 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2963 if (pv->pv_pmap == pmap) {
2964 vm_page_spin_unlock(m);
2971 vm_page_spin_unlock(m);
2977 * Remove all pages from specified address space this aids process
2978 * exit speeds. Also, this code is special cased for current
2979 * process only, but can have the more generic (and slightly slower)
2980 * mode enabled. This is much faster than pmap_remove in the case
2981 * of running down an entire address space.
2983 * No other requirements.
2986 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2988 pt_entry_t *pte, tpte;
2991 int save_generation;
2993 if (pmap->pm_pteobj)
2994 vm_object_hold(pmap->pm_pteobj);
2996 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2997 if (pv->pv_va >= eva || pv->pv_va < sva) {
2998 npv = TAILQ_NEXT(pv, pv_plist);
3002 KKASSERT(pmap == pv->pv_pmap);
3004 pte = pmap_pte(pmap, pv->pv_va);
3007 * We cannot remove wired pages from a process' mapping
3010 if (*pte & VPTE_WIRED) {
3011 npv = TAILQ_NEXT(pv, pv_plist);
3014 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
3016 m = PHYS_TO_VM_PAGE(tpte & VPTE_FRAME);
3017 vm_page_spin_lock(m);
3019 KASSERT(m < &vm_page_array[vm_page_array_size],
3020 ("pmap_remove_pages: bad tpte %lx", tpte));
3022 KKASSERT(pmap->pm_stats.resident_count > 0);
3023 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3026 * Update the vm_page_t clean and reference bits.
3028 if (tpte & VPTE_M) {
3032 npv = TAILQ_NEXT(pv, pv_plist);
3033 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3034 atomic_add_int(&pmap->pm_generation, 1);
3035 save_generation = pmap->pm_generation;
3037 m->md.pv_list_count--;
3038 atomic_add_int(&m->object->agg_pv_list_count, -1);
3039 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3040 if (TAILQ_EMPTY(&m->md.pv_list))
3041 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3042 vm_page_spin_unlock(m);
3044 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
3048 * Restart the scan if we blocked during the unuse or free
3049 * calls and other removals were made.
3051 if (save_generation != pmap->pm_generation) {
3052 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3053 npv = TAILQ_FIRST(&pmap->pm_pvlist);
3056 if (pmap->pm_pteobj)
3057 vm_object_drop(pmap->pm_pteobj);
3061 * pmap_testbit tests bits in active mappings of a VM page.
3064 pmap_testbit(vm_page_t m, int bit)
3069 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3072 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3075 vm_page_spin_lock(m);
3076 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3078 * if the bit being tested is the modified bit, then
3079 * mark clean_map and ptes as never
3082 if (bit & (VPTE_A|VPTE_M)) {
3083 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
3087 #if defined(PMAP_DIAGNOSTIC)
3088 if (pv->pv_pmap == NULL) {
3089 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3093 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
3095 vm_page_spin_unlock(m);
3099 vm_page_spin_unlock(m);
3104 * This routine is used to clear bits in ptes. Certain bits require special
3105 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
3107 * This routine is only called with certain VPTE_* bit combinations.
3109 static __inline void
3110 pmap_clearbit(vm_page_t m, int bit)
3116 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3120 * Loop over all current mappings setting/clearing as appropos If
3121 * setting RO do we need to clear the VAC?
3123 vm_page_spin_lock(m);
3124 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3126 * don't write protect pager mappings
3128 if (bit == VPTE_RW) {
3129 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
3133 #if defined(PMAP_DIAGNOSTIC)
3134 if (pv->pv_pmap == NULL) {
3135 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3141 * Careful here. We can use a locked bus instruction to
3142 * clear VPTE_A or VPTE_M safely but we need to synchronize
3143 * with the target cpus when we mess with VPTE_RW.
3145 * On virtual kernels we must force a new fault-on-write
3146 * in the real kernel if we clear the Modify bit ourselves,
3147 * otherwise the real kernel will not get a new fault and
3148 * will never set our Modify bit again.
3150 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
3152 if (bit == VPTE_RW) {
3154 * We must also clear VPTE_M when clearing
3157 pbits = pmap_clean_pte(pte, pv->pv_pmap,
3161 } else if (bit == VPTE_M) {
3163 * We do not have to make the page read-only
3164 * when clearing the Modify bit. The real
3165 * kernel will make the real PTE read-only
3166 * or otherwise detect the write and set
3167 * our VPTE_M again simply by us invalidating
3168 * the real kernel VA for the pmap (as we did
3169 * above). This allows the real kernel to
3170 * handle the write fault without forwarding
3173 atomic_clear_long(pte, VPTE_M);
3174 } else if ((bit & (VPTE_RW|VPTE_M)) == (VPTE_RW|VPTE_M)) {
3176 * We've been asked to clear W & M, I guess
3177 * the caller doesn't want us to update
3178 * the dirty status of the VM page.
3180 pmap_clean_pte(pte, pv->pv_pmap, pv->pv_va);
3183 * We've been asked to clear bits that do
3184 * not interact with hardware.
3186 atomic_clear_long(pte, bit);
3190 vm_page_spin_unlock(m);
3194 * Lower the permission for all mappings to a given page.
3196 * No other requirements.
3199 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3201 /* JG NX support? */
3202 if ((prot & VM_PROT_WRITE) == 0) {
3203 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3204 pmap_clearbit(m, VPTE_RW);
3205 vm_page_flag_clear(m, PG_WRITEABLE);
3213 pmap_phys_address(vm_pindex_t ppn)
3215 return (x86_64_ptob(ppn));
3219 * Return a count of reference bits for a page, clearing those bits.
3220 * It is not necessary for every reference bit to be cleared, but it
3221 * is necessary that 0 only be returned when there are truly no
3222 * reference bits set.
3224 * XXX: The exact number of bits to check and clear is a matter that
3225 * should be tested and standardized at some point in the future for
3226 * optimal aging of shared pages.
3228 * No other requirements.
3231 pmap_ts_referenced(vm_page_t m)
3233 pv_entry_t pv, pvf, pvn;
3237 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3240 vm_page_spin_lock(m);
3242 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3245 pvn = TAILQ_NEXT(pv, pv_list);
3246 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3247 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3249 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
3252 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
3254 if (pte && (*pte & VPTE_A)) {
3255 atomic_clear_long(pte, VPTE_A);
3261 } while ((pv = pvn) != NULL && pv != pvf);
3263 vm_page_spin_unlock(m);
3269 * Return whether or not the specified physical page was modified
3270 * in any physical maps.
3272 * No other requirements.
3275 pmap_is_modified(vm_page_t m)
3279 res = pmap_testbit(m, VPTE_M);
3285 * Clear the modify bits on the specified physical page.
3287 * No other requirements.
3290 pmap_clear_modify(vm_page_t m)
3292 pmap_clearbit(m, VPTE_M);
3296 * Clear the reference bit on the specified physical page.
3298 * No other requirements.
3301 pmap_clear_reference(vm_page_t m)
3303 pmap_clearbit(m, VPTE_A);
3307 * Miscellaneous support routines follow
3311 i386_protection_init(void)
3315 kp = protection_codes;
3316 for (prot = 0; prot < 8; prot++) {
3317 if (prot & VM_PROT_READ)
3318 *kp |= 0; /* if it's VALID is readeable */
3319 if (prot & VM_PROT_WRITE)
3321 if (prot & VM_PROT_EXECUTE)
3322 *kp |= 0; /* if it's VALID is executable */
3328 * Sets the memory attribute for the specified page.
3331 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
3333 /* This is a vkernel, do nothing */
3337 * Change the PAT attribute on an existing kernel memory map. Caller
3338 * must ensure that the virtual memory in question is not accessed
3339 * during the adjustment.
3342 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
3344 /* This is a vkernel, do nothing */
3348 * Perform the pmap work for mincore
3350 * No other requirements.
3353 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3355 pt_entry_t *ptep, pte;
3359 vm_object_hold(pmap->pm_pteobj);
3360 ptep = pmap_pte(pmap, addr);
3362 if (ptep && (pte = *ptep) != 0) {
3365 val = MINCORE_INCORE;
3366 if ((pte & VPTE_MANAGED) == 0)
3369 pa = pte & VPTE_FRAME;
3371 m = PHYS_TO_VM_PAGE(pa);
3377 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3379 * Modified by someone
3381 else if (m->dirty || pmap_is_modified(m))
3382 val |= MINCORE_MODIFIED_OTHER;
3387 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3390 * Referenced by someone
3392 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3393 val |= MINCORE_REFERENCED_OTHER;
3394 vm_page_flag_set(m, PG_REFERENCED);
3398 vm_object_drop(pmap->pm_pteobj);
3404 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3405 * vmspace will be ref'd and the old one will be deref'd.
3407 * Caller must hold vmspace->vm_map.token for oldvm and newvm
3410 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3412 struct vmspace *oldvm;
3416 oldvm = p->p_vmspace;
3417 if (oldvm != newvm) {
3420 p->p_vmspace = newvm;
3421 KKASSERT(p->p_nthreads == 1);
3422 lp = RB_ROOT(&p->p_lwp_tree);
3423 pmap_setlwpvm(lp, newvm);
3431 * Set the vmspace for a LWP. The vmspace is almost universally set the
3432 * same as the process vmspace, but virtual kernels need to swap out contexts
3433 * on a per-lwp basis.
3436 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3438 struct vmspace *oldvm;
3441 oldvm = lp->lwp_vmspace;
3442 if (oldvm != newvm) {
3444 lp->lwp_vmspace = newvm;
3445 if (curthread->td_lwp == lp) {
3446 pmap = vmspace_pmap(newvm);
3447 ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
3448 if (pmap->pm_active_lock & CPULOCK_EXCL)
3449 pmap_interlock_wait(newvm);
3450 #if defined(SWTCH_OPTIM_STATS)
3453 pmap = vmspace_pmap(oldvm);
3454 ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
3462 * The swtch code tried to switch in a heavy weight process whos pmap
3463 * is locked by another cpu. We have to wait for the lock to clear before
3464 * the pmap can be used.
3467 pmap_interlock_wait (struct vmspace *vm)
3469 pmap_t pmap = vmspace_pmap(vm);
3471 if (pmap->pm_active_lock & CPULOCK_EXCL) {
3473 while (pmap->pm_active_lock & CPULOCK_EXCL) {
3482 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3485 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3489 addr = roundup2(addr, NBPDR);
3494 * Used by kmalloc/kfree, page already exists at va
3497 pmap_kvtom(vm_offset_t va)
3501 KKASSERT(va >= KvaStart && va < KvaEnd);
3503 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
3507 pmap_object_init(vm_object_t object)
3513 pmap_object_free(vm_object_t object)
3519 pmap_pgscan(struct pmap_pgscan_info *pginfo)
3521 pmap_t pmap = pginfo->pmap;
3522 vm_offset_t sva = pginfo->beg_addr;
3523 vm_offset_t eva = pginfo->end_addr;
3524 vm_offset_t va_next;
3525 pml4_entry_t *pml4e;
3527 pd_entry_t ptpaddr, *pde;
3531 vm_object_hold(pmap->pm_pteobj);
3533 for (; sva < eva; sva = va_next) {
3537 pml4e = pmap_pml4e(pmap, sva);
3538 if ((*pml4e & VPTE_V) == 0) {
3539 va_next = (sva + NBPML4) & ~PML4MASK;
3545 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
3546 if ((*pdpe & VPTE_V) == 0) {
3547 va_next = (sva + NBPDP) & ~PDPMASK;
3553 va_next = (sva + NBPDR) & ~PDRMASK;
3557 pde = pmap_pdpe_to_pde(pdpe, sva);
3561 * Check for large page (ignore).
3563 if ((ptpaddr & VPTE_PS) != 0) {
3565 pmap_clean_pde(pde, pmap, sva);
3566 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
3572 * Weed out invalid mappings. Note: we assume that the page
3573 * directory table is always allocated, and in kernel virtual.
3581 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
3587 if ((*pte & VPTE_MANAGED) == 0)
3590 m = PHYS_TO_VM_PAGE(*pte & VPTE_FRAME);
3591 if (vm_page_busy_try(m, TRUE) == 0) {
3592 if (pginfo->callback(pginfo, sva, m) < 0)
3597 vm_object_drop(pmap->pm_pteobj);