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) 2008-2019 The DragonFly Project.
8 * Copyright (c) 2008, 2009 Jordan Gordeev.
11 * This code is derived from software contributed to Berkeley by
12 * the Systems Programming Group of the University of Utah Computer
13 * Science Department and William Jolitz of UUNET Technologies Inc.
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
43 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
44 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
48 * Manages physical address maps.
51 #include "opt_msgbuf.h"
53 #include <sys/param.h>
54 #include <sys/systm.h>
55 #include <sys/kernel.h>
57 #include <sys/msgbuf.h>
58 #include <sys/vmmeter.h>
60 #include <sys/vmspace.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>
75 #include <sys/thread2.h>
76 #include <sys/spinlock2.h>
77 #include <vm/vm_page2.h>
79 #include <machine/cputypes.h>
80 #include <machine/md_var.h>
81 #include <machine/specialreg.h>
82 #include <machine/smp.h>
83 #include <machine/globaldata.h>
84 #include <machine/pcb.h>
85 #include <machine/pmap.h>
86 #include <machine/pmap_inval.h>
94 #define PMAP_KEEP_PDIRS
95 #ifndef PMAP_SHPGPERPROC
96 #define PMAP_SHPGPERPROC 1000
99 #if defined(DIAGNOSTIC)
100 #define PMAP_DIAGNOSTIC
105 #if !defined(PMAP_DIAGNOSTIC)
106 #define PMAP_INLINE __inline
112 * Get PDEs and PTEs for user/kernel address space
114 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
115 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
117 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & VPTE_V) != 0)
118 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & VPTE_WIRED) != 0)
119 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & VPTE_M) != 0)
120 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & VPTE_A) != 0)
121 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & VPTE_V) != 0)
124 * Given a map and a machine independent protection code,
125 * convert to a vax protection code.
127 #define pte_prot(m, p) \
128 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
129 static uint64_t protection_codes[8];
131 static struct pmap kernel_pmap_store;
132 struct pmap *kernel_pmap = &kernel_pmap_store;
134 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
136 static struct vm_object kptobj;
139 static uint64_t KPDphys; /* phys addr of kernel level 2 */
140 uint64_t KPDPphys; /* phys addr of kernel level 3 */
141 uint64_t KPML4phys; /* phys addr of kernel level 4 */
143 extern void *vkernel_stack;
146 * Data for the pv entry allocation mechanism
148 static vm_zone_t pvzone;
149 static struct vm_zone pvzone_store;
150 static vm_pindex_t pv_entry_count = 0;
151 static vm_pindex_t pv_entry_max = 0;
152 static vm_pindex_t pv_entry_high_water = 0;
153 static int pmap_pagedaemon_waken = 0;
154 static struct pv_entry *pvinit;
157 * All those kernel PT submaps that BSD is so fond of
159 pt_entry_t *CMAP1 = NULL, *ptmmap;
160 caddr_t CADDR1 = NULL;
161 static pt_entry_t *msgbufmap;
165 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
166 static pv_entry_t get_pv_entry (void);
167 static void x86_64_protection_init (void);
168 static __inline void pmap_clearbit (vm_page_t m, int bit);
170 static void pmap_remove_all (vm_page_t m);
171 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
172 pt_entry_t oldpte, vm_offset_t sva);
173 static void pmap_remove_page (struct pmap *pmap, vm_offset_t va);
174 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
176 static boolean_t pmap_testbit (vm_page_t m, int bit);
177 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
178 vm_page_t mpte, vm_page_t m, pv_entry_t);
180 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
182 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
183 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
184 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
185 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t);
188 pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
190 if (pv1->pv_va < pv2->pv_va)
192 if (pv1->pv_va > pv2->pv_va)
197 RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
198 pv_entry_compare, vm_offset_t, pv_va);
200 static __inline vm_pindex_t
201 pmap_pt_pindex(vm_offset_t va)
203 return va >> PDRSHIFT;
206 static __inline vm_pindex_t
207 pmap_pte_index(vm_offset_t va)
209 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
212 static __inline vm_pindex_t
213 pmap_pde_index(vm_offset_t va)
215 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
218 static __inline vm_pindex_t
219 pmap_pdpe_index(vm_offset_t va)
221 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
224 static __inline vm_pindex_t
225 pmap_pml4e_index(vm_offset_t va)
227 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
230 /* Return a pointer to the PML4 slot that corresponds to a VA */
231 static __inline pml4_entry_t *
232 pmap_pml4e(pmap_t pmap, vm_offset_t va)
234 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
237 /* Return a pointer to the PDP slot that corresponds to a VA */
238 static __inline pdp_entry_t *
239 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
243 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & VPTE_FRAME);
244 return (&pdpe[pmap_pdpe_index(va)]);
247 /* Return a pointer to the PDP slot that corresponds to a VA */
248 static __inline pdp_entry_t *
249 pmap_pdpe(pmap_t pmap, vm_offset_t va)
253 pml4e = pmap_pml4e(pmap, va);
254 if ((*pml4e & VPTE_V) == 0)
256 return (pmap_pml4e_to_pdpe(pml4e, va));
259 /* Return a pointer to the PD slot that corresponds to a VA */
260 static __inline pd_entry_t *
261 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
265 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & VPTE_FRAME);
266 return (&pde[pmap_pde_index(va)]);
269 /* Return a pointer to the PD slot that corresponds to a VA */
270 static __inline pd_entry_t *
271 pmap_pde(pmap_t pmap, vm_offset_t va)
275 pdpe = pmap_pdpe(pmap, va);
276 if (pdpe == NULL || (*pdpe & VPTE_V) == 0)
278 return (pmap_pdpe_to_pde(pdpe, va));
281 /* Return a pointer to the PT slot that corresponds to a VA */
282 static __inline pt_entry_t *
283 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
287 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & VPTE_FRAME);
288 return (&pte[pmap_pte_index(va)]);
292 * Hold pt_m for page table scans to prevent it from getting reused out
293 * from under us across blocking conditions in the body of the loop.
297 pmap_hold_pt_page(pd_entry_t *pde, vm_offset_t va)
302 pte = (pt_entry_t)*pde;
304 pt_m = PHYS_TO_VM_PAGE(pte & VPTE_FRAME);
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 x86 pmap_pte() */
320 return ((pt_entry_t *)pde);
321 return (pmap_pde_to_pte(pde, va));
324 static PMAP_INLINE pt_entry_t *
325 vtopte(vm_offset_t va)
328 x = pmap_pte(kernel_pmap, va);
333 static __inline pd_entry_t *
334 vtopde(vm_offset_t va)
337 x = pmap_pde(kernel_pmap, va);
343 * Returns the physical address translation from va for a user address.
344 * (vm_paddr_t)-1 is returned on failure.
347 uservtophys(vm_offset_t va)
349 struct vmspace *vm = curproc->p_vmspace;
355 /* XXX No idea how to handle this case in a simple way, just abort */
356 if (PAGE_SIZE - (va & PAGE_MASK) < sizeof(u_int))
357 return ((vm_paddr_t)-1);
359 m = vm_fault_page(&vm->vm_map, trunc_page(va),
360 VM_PROT_READ|VM_PROT_WRITE,
364 return ((vm_paddr_t)-1);
366 pa = VM_PAGE_TO_PHYS(m) | (va & PAGE_MASK);
376 allocpages(vm_paddr_t *firstaddr, int n)
381 /*bzero((void *)ret, n * PAGE_SIZE); not mapped yet */
382 *firstaddr += n * PAGE_SIZE;
387 create_pagetables(vm_paddr_t *firstaddr, int64_t ptov_offset)
390 pml4_entry_t *KPML4virt;
391 pdp_entry_t *KPDPvirt;
394 int kpml4i = pmap_pml4e_index(ptov_offset);
395 int kpdpi = pmap_pdpe_index(ptov_offset);
396 int kpdi = pmap_pde_index(ptov_offset);
399 * Calculate NKPT - number of kernel page tables. We have to
400 * accomodoate prealloction of the vm_page_array, dump bitmap,
401 * MSGBUF_SIZE, and other stuff. Be generous.
403 * Maxmem is in pages.
405 nkpt = (Maxmem * (sizeof(struct vm_page) * 2) + MSGBUF_SIZE) / NBPDR;
409 KPML4phys = allocpages(firstaddr, 1);
410 KPDPphys = allocpages(firstaddr, NKPML4E);
411 KPDphys = allocpages(firstaddr, NKPDPE);
412 KPTphys = allocpages(firstaddr, nkpt);
414 KPML4virt = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
415 KPDPvirt = (pdp_entry_t *)PHYS_TO_DMAP(KPDPphys);
416 KPDvirt = (pd_entry_t *)PHYS_TO_DMAP(KPDphys);
417 KPTvirt = (pt_entry_t *)PHYS_TO_DMAP(KPTphys);
419 bzero(KPML4virt, 1 * PAGE_SIZE);
420 bzero(KPDPvirt, NKPML4E * PAGE_SIZE);
421 bzero(KPDvirt, NKPDPE * PAGE_SIZE);
422 bzero(KPTvirt, nkpt * PAGE_SIZE);
424 /* Now map the page tables at their location within PTmap */
425 for (i = 0; i < nkpt; i++) {
426 KPDvirt[i + kpdi] = KPTphys + (i << PAGE_SHIFT);
427 KPDvirt[i + kpdi] |= VPTE_RW | VPTE_V | VPTE_U;
430 /* And connect up the PD to the PDP */
431 for (i = 0; i < NKPDPE; i++) {
432 KPDPvirt[i + kpdpi] = KPDphys + (i << PAGE_SHIFT);
433 KPDPvirt[i + kpdpi] |= VPTE_RW | VPTE_V | VPTE_U;
436 /* And recursively map PML4 to itself in order to get PTmap */
437 KPML4virt[PML4PML4I] = KPML4phys;
438 KPML4virt[PML4PML4I] |= VPTE_RW | VPTE_V | VPTE_U;
440 /* Connect the KVA slot up to the PML4 */
441 KPML4virt[kpml4i] = KPDPphys;
442 KPML4virt[kpml4i] |= VPTE_RW | VPTE_V | VPTE_U;
446 * Typically used to initialize a fictitious page by vm/device_pager.c
449 pmap_page_init(struct vm_page *m)
452 TAILQ_INIT(&m->md.pv_list);
456 * Bootstrap the system enough to run with virtual memory.
458 * On x86_64 this is called after mapping has already been enabled
459 * and just syncs the pmap module with what has already been done.
460 * [We can't call it easily with mapping off since the kernel is not
461 * mapped with PA == VA, hence we would have to relocate every address
462 * from the linked base (virtual) address "KERNBASE" to the actual
463 * (physical) address starting relative to 0]
466 pmap_bootstrap(vm_paddr_t *firstaddr, int64_t ptov_offset)
472 * Create an initial set of page tables to run the kernel in.
474 create_pagetables(firstaddr, ptov_offset);
476 virtual_start = KvaStart;
477 virtual_end = KvaEnd;
480 * Initialize protection array.
482 x86_64_protection_init();
485 * The kernel's pmap is statically allocated so we don't have to use
486 * pmap_create, which is unlikely to work correctly at this part of
487 * the boot sequence (XXX and which no longer exists).
489 * The kernel_pmap's pm_pteobj is used only for locking and not
492 kernel_pmap->pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
493 kernel_pmap->pm_count = 1;
494 /* don't allow deactivation */
495 CPUMASK_ASSALLONES(kernel_pmap->pm_active);
496 kernel_pmap->pm_pteobj = NULL; /* see pmap_init */
497 RB_INIT(&kernel_pmap->pm_pvroot);
498 spin_init(&kernel_pmap->pm_spin, "pmapbootstrap");
501 * Reserve some special page table entries/VA space for temporary
504 #define SYSMAP(c, p, v, n) \
505 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
508 pte = pmap_pte(kernel_pmap, va);
510 * CMAP1/CMAP2 are used for zeroing and copying pages.
512 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
518 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
522 * ptvmmap is used for reading arbitrary physical pages via
525 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
528 * msgbufp is used to map the system message buffer.
529 * XXX msgbufmap is not used.
531 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
532 atop(round_page(MSGBUF_SIZE)))
541 * Initialize the pmap module.
542 * Called by vm_init, to initialize any structures that the pmap
543 * system needs to map virtual memory.
544 * pmap_init has been enhanced to support in a fairly consistant
545 * way, discontiguous physical memory.
551 vm_pindex_t initial_pvs;
554 * object for kernel page table pages
556 /* JG I think the number can be arbitrary */
557 vm_object_init(&kptobj, 5);
558 kernel_pmap->pm_pteobj = &kptobj;
561 * Allocate memory for random pmap data structures. Includes the
564 for (i = 0; i < vm_page_array_size; i++) {
567 m = &vm_page_array[i];
568 TAILQ_INIT(&m->md.pv_list);
569 m->md.pv_list_count = 0;
573 * init the pv free list
575 initial_pvs = vm_page_array_size;
576 if (initial_pvs < MINPV)
578 pvzone = &pvzone_store;
579 pvinit = (struct pv_entry *)
580 kmem_alloc(kernel_map,
581 initial_pvs * sizeof (struct pv_entry),
583 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
587 * Now it is safe to enable pv_table recording.
589 pmap_initialized = TRUE;
593 * Initialize the address space (zone) for the pv_entries. Set a
594 * high water mark so that the system can recover from excessive
595 * numbers of pv entries.
600 vm_pindex_t shpgperproc = PMAP_SHPGPERPROC;
602 TUNABLE_LONG_FETCH("vm.pmap.shpgperproc", &shpgperproc);
603 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
604 TUNABLE_LONG_FETCH("vm.pmap.pv_entries", &pv_entry_max);
605 pv_entry_high_water = 9 * (pv_entry_max / 10);
606 zinitna(pvzone, NULL, 0, pv_entry_max, ZONE_INTERRUPT);
610 /***************************************************
611 * Low level helper routines.....
612 ***************************************************/
615 * The modification bit is not tracked for any pages in this range. XXX
616 * such pages in this maps should always use pmap_k*() functions and not
619 * XXX User and kernel address spaces are independant for virtual kernels,
620 * this function only applies to the kernel pmap.
623 pmap_track_modified(pmap_t pmap, vm_offset_t va)
625 KKASSERT(pmap != kernel_pmap ||
626 va < clean_sva || va >= clean_eva);
630 * Extract the physical page address associated with the map/VA pair.
635 pmap_extract(pmap_t pmap, vm_offset_t va, void **handlep)
639 pd_entry_t pde, *pdep;
641 vm_object_hold(pmap->pm_pteobj);
643 pdep = pmap_pde(pmap, va);
647 if ((pde & VPTE_PS) != 0) {
649 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
651 pte = pmap_pde_to_pte(pdep, va);
652 rtval = (*pte & VPTE_FRAME) | (va & PAGE_MASK);
657 *handlep = NULL; /* XXX */
658 vm_object_drop(pmap->pm_pteobj);
664 pmap_extract_done(void *handle)
670 vm_object_drop(pmap->pm_pteobj);
675 * Similar to extract but checks protections, SMP-friendly short-cut for
676 * vm_fault_page[_quick]().
678 * WARNING! THE RETURNED PAGE IS ONLY HELD AND NEITHER IT NOR ITS TARGET
679 * DATA IS SUITABLE FOR WRITING. Writing can interfere with
680 * pageouts flushes, msync, etc. The hold_count is not enough
681 * to avoid races against pageouts and other flush code doesn't
682 * care about hold_count.
685 pmap_fault_page_quick(pmap_t pmap __unused, vm_offset_t vaddr __unused,
686 vm_prot_t prot __unused, int *busyp __unused)
692 * Routine: pmap_kextract
694 * Extract the physical page address associated
695 * kernel virtual address.
698 pmap_kextract(vm_offset_t va)
703 KKASSERT(va >= KvaStart && va < KvaEnd);
706 * The DMAP region is not included in [KvaStart, KvaEnd)
709 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
710 pa = DMAP_TO_PHYS(va);
716 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
719 * Beware of a concurrent promotion that changes the
720 * PDE at this point! For example, vtopte() must not
721 * be used to access the PTE because it would use the
722 * new PDE. It is, however, safe to use the old PDE
723 * because the page table page is preserved by the
726 pa = *pmap_pde_to_pte(&pde, va);
727 pa = (pa & VPTE_FRAME) | (va & PAGE_MASK);
735 /***************************************************
736 * Low level mapping routines.....
737 ***************************************************/
740 * Enter a mapping into kernel_pmap. Mappings created in this fashion
741 * are not managed. Mappings must be immediately accessible on all cpus.
743 * Call pmap_inval_pte() to invalidate the virtual pte and clean out the
744 * real pmap and handle related races before storing the new vpte. The
745 * new semantics for kenter require use to do an UNCONDITIONAL invalidation,
746 * because the entry may have previously been cleared without an invalidation.
749 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
754 KKASSERT(va >= KvaStart && va < KvaEnd);
755 npte = pa | VPTE_RW | VPTE_V | VPTE_U;
759 pmap_inval_pte(ptep, kernel_pmap, va);
762 pmap_inval_pte(ptep, kernel_pmap, va);
764 atomic_swap_long(ptep, npte);
768 * Enter an unmanaged KVA mapping for the private use of the current
771 * It is illegal for the mapping to be accessed by other cpus without
772 * proper invalidation.
775 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
781 KKASSERT(va >= KvaStart && va < KvaEnd);
783 npte = (vpte_t)pa | VPTE_RW | VPTE_V | VPTE_U;
787 pmap_inval_pte_quick(ptep, kernel_pmap, va);
793 pmap_inval_pte(pte, kernel_pmap, va);
795 atomic_swap_long(ptep, npte);
801 * Invalidation will occur later, ok to be lazy here.
804 pmap_kenter_noinval(vm_offset_t va, vm_paddr_t pa)
810 KKASSERT(va >= KvaStart && va < KvaEnd);
812 npte = (vpte_t)pa | VPTE_RW | VPTE_V | VPTE_U;
820 atomic_swap_long(ptep, npte);
826 * Remove an unmanaged mapping created with pmap_kenter*().
829 pmap_kremove(vm_offset_t va)
833 KKASSERT(va >= KvaStart && va < KvaEnd);
836 atomic_swap_long(ptep, 0);
837 pmap_inval_pte(ptep, kernel_pmap, va);
841 * Remove an unmanaged mapping created with pmap_kenter*() but synchronize
842 * only with this cpu.
844 * Unfortunately because we optimize new entries by testing VPTE_V later
845 * on, we actually still have to synchronize with all the cpus. XXX maybe
846 * store a junk value and test against 0 in the other places instead?
849 pmap_kremove_quick(vm_offset_t va)
853 KKASSERT(va >= KvaStart && va < KvaEnd);
856 atomic_swap_long(ptep, 0);
857 pmap_inval_pte(ptep, kernel_pmap, va); /* NOT _quick */
861 * Invalidation will occur later, ok to be lazy here.
864 pmap_kremove_noinval(vm_offset_t va)
868 KKASSERT(va >= KvaStart && va < KvaEnd);
871 atomic_swap_long(ptep, 0);
875 * Used to map a range of physical addresses into kernel
876 * virtual address space.
878 * For now, VM is already on, we only need to map the
882 pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
884 return PHYS_TO_DMAP(start);
888 * Map a set of unmanaged VM pages into KVM.
891 _pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count, int doinval)
896 end_va = beg_va + count * PAGE_SIZE;
897 KKASSERT(beg_va >= KvaStart && end_va <= KvaEnd);
899 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
903 atomic_swap_long(ptep, VM_PAGE_TO_PHYS(*m) |
904 VPTE_RW | VPTE_V | VPTE_U);
908 pmap_invalidate_range(kernel_pmap, beg_va, end_va);
909 /* pmap_inval_pte(pte, kernel_pmap, va); */
913 pmap_qenter(vm_offset_t beg_va, vm_page_t *m, int count)
915 _pmap_qenter(beg_va, m, count, 1);
919 pmap_qenter_noinval(vm_offset_t beg_va, vm_page_t *m, int count)
921 _pmap_qenter(beg_va, m, count, 0);
925 * Undo the effects of pmap_qenter*().
928 pmap_qremove(vm_offset_t beg_va, int count)
933 end_va = beg_va + count * PAGE_SIZE;
934 KKASSERT(beg_va >= KvaStart && end_va < KvaEnd);
936 for (va = beg_va; va < end_va; va += PAGE_SIZE) {
940 atomic_swap_long(ptep, 0);
942 pmap_invalidate_range(kernel_pmap, beg_va, end_va);
946 * Unlike the real pmap code, we can't avoid calling the real-kernel.
949 pmap_qremove_quick(vm_offset_t va, int count)
951 pmap_qremove(va, count);
955 pmap_qremove_noinval(vm_offset_t va, int count)
957 pmap_qremove(va, count);
961 * This routine works like vm_page_lookup() but also blocks as long as the
962 * page is busy. This routine does not busy the page it returns.
964 * Unless the caller is managing objects whos pages are in a known state,
965 * the call should be made with a critical section held so the page's object
966 * association remains valid on return.
969 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
973 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
974 m = vm_page_lookup_busy_wait(object, pindex, TRUE, "pplookp");
980 * Create a new thread and optionally associate it with a (new) process.
981 * NOTE! the new thread's cpu may not equal the current cpu.
984 pmap_init_thread(thread_t td)
986 /* enforce pcb placement */
987 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
988 td->td_savefpu = &td->td_pcb->pcb_save;
989 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on x86_64? */
993 * This routine directly affects the fork perf for a process.
996 pmap_init_proc(struct proc *p)
1001 * Unwire a page table which has been removed from the pmap. We own the
1002 * wire_count, so the page cannot go away. The page representing the page
1003 * table is passed in unbusied and must be busied if we cannot trivially
1006 * XXX NOTE! This code is not usually run because we do not currently
1007 * implement dynamic page table page removal. The page in
1008 * its parent assumes at least 1 wire count, so no call to this
1009 * function ever sees a wire count less than 2.
1012 pmap_unwire_pgtable(pmap_t pmap, vm_offset_t va, vm_page_t m)
1015 * Try to unwire optimally. If non-zero is returned the wire_count
1016 * is 1 and we must busy the page to unwire it.
1018 if (vm_page_unwire_quick(m) == 0)
1021 vm_page_busy_wait(m, TRUE, "pmuwpt");
1022 KASSERT(m->queue == PQ_NONE,
1023 ("_pmap_unwire_pgtable: %p->queue != PQ_NONE", m));
1025 if (m->wire_count == 1) {
1027 * Unmap the page table page.
1029 /* pmap_inval_add(info, pmap, -1); */
1031 if (m->pindex >= (NUPT_TOTAL + NUPD_TOTAL)) {
1034 pml4 = pmap_pml4e(pmap, va);
1036 } else if (m->pindex >= NUPT_TOTAL) {
1039 pdp = pmap_pdpe(pmap, va);
1044 pd = pmap_pde(pmap, va);
1048 KKASSERT(pmap->pm_stats.resident_count > 0);
1049 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1051 if (pmap->pm_ptphint == m)
1052 pmap->pm_ptphint = NULL;
1054 if (m->pindex < NUPT_TOTAL) {
1055 /* We just released a PT, unhold the matching PD */
1058 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) &
1060 pmap_unwire_pgtable(pmap, va, pdpg);
1062 if (m->pindex >= NUPT_TOTAL &&
1063 m->pindex < (NUPT_TOTAL + NUPD_TOTAL)) {
1064 /* We just released a PD, unhold the matching PDP */
1067 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) &
1069 pmap_unwire_pgtable(pmap, va, pdppg);
1073 * This was our last wire, the page had better be unwired
1074 * after we decrement wire_count.
1076 * FUTURE NOTE: shared page directory page could result in
1077 * multiple wire counts.
1079 vm_page_unwire(m, 0);
1080 KKASSERT(m->wire_count == 0);
1081 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1086 /* XXX SMP race to 1 if not holding vmobj */
1087 vm_page_unwire(m, 0);
1094 * After removing a page table entry, this routine is used to
1095 * conditionally free the page, and manage the hold/wire counts.
1097 * If not NULL the caller owns a wire_count on mpte, so it can't disappear.
1098 * If NULL the caller owns a wire_count on what would be the mpte, we must
1102 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
1104 vm_pindex_t ptepindex;
1106 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1110 * page table pages in the kernel_pmap are not managed.
1112 if (pmap == kernel_pmap)
1114 ptepindex = pmap_pt_pindex(va);
1115 if (pmap->pm_ptphint &&
1116 (pmap->pm_ptphint->pindex == ptepindex)) {
1117 mpte = pmap->pm_ptphint;
1119 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1120 pmap->pm_ptphint = mpte;
1121 vm_page_wakeup(mpte);
1124 return pmap_unwire_pgtable(pmap, va, mpte);
1128 * Initialize pmap0/vmspace0 . Since process 0 never enters user mode we
1129 * just dummy it up so it works well enough for fork().
1131 * In DragonFly, process pmaps may only be used to manipulate user address
1132 * space, never kernel address space.
1135 pmap_pinit0(struct pmap *pmap)
1141 * Initialize a preallocated and zeroed pmap structure,
1142 * such as one in a vmspace structure.
1145 pmap_pinit(struct pmap *pmap)
1150 * No need to allocate page table space yet but we do need a valid
1151 * page directory table.
1153 if (pmap->pm_pml4 == NULL) {
1154 pmap->pm_pml4 = (pml4_entry_t *)
1155 kmem_alloc_pageable(kernel_map, PAGE_SIZE,
1160 * Allocate an object for the ptes
1162 if (pmap->pm_pteobj == NULL)
1163 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL + 1);
1166 * Allocate the page directory page, unless we already have
1167 * one cached. If we used the cached page the wire_count will
1168 * already be set appropriately.
1170 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1171 ptdpg = vm_page_grab(pmap->pm_pteobj,
1172 NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL,
1173 VM_ALLOC_NORMAL | VM_ALLOC_RETRY |
1175 pmap->pm_pdirm = ptdpg;
1176 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_WRITEABLE);
1177 vm_page_wire(ptdpg);
1178 vm_page_wakeup(ptdpg);
1179 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1182 CPUMASK_ASSZERO(pmap->pm_active);
1183 pmap->pm_ptphint = NULL;
1184 RB_INIT(&pmap->pm_pvroot);
1185 spin_init(&pmap->pm_spin, "pmapinit");
1186 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1187 pmap->pm_stats.resident_count = 1;
1188 pmap->pm_stats.wired_count = 1;
1192 * Clean up a pmap structure so it can be physically freed. This routine
1193 * is called by the vmspace dtor function. A great deal of pmap data is
1194 * left passively mapped to improve vmspace management so we have a bit
1195 * of cleanup work to do here.
1200 pmap_puninit(pmap_t pmap)
1204 KKASSERT(CPUMASK_TESTZERO(pmap->pm_active));
1205 if ((p = pmap->pm_pdirm) != NULL) {
1206 KKASSERT(pmap->pm_pml4 != NULL);
1207 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1208 vm_page_busy_wait(p, TRUE, "pgpun");
1209 vm_page_unwire(p, 0);
1210 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1212 pmap->pm_pdirm = NULL;
1213 atomic_add_long(&pmap->pm_stats.wired_count, -1);
1214 KKASSERT(pmap->pm_stats.wired_count == 0);
1216 if (pmap->pm_pml4) {
1217 kmem_free(kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1218 pmap->pm_pml4 = NULL;
1220 if (pmap->pm_pteobj) {
1221 vm_object_deallocate(pmap->pm_pteobj);
1222 pmap->pm_pteobj = NULL;
1227 * This function is now unused (used to add the pmap to the pmap_list)
1230 pmap_pinit2(struct pmap *pmap)
1235 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1236 * 0 on failure (if the procedure had to sleep).
1238 * When asked to remove the page directory page itself, we actually just
1239 * leave it cached so we do not have to incur the SMP inval overhead of
1240 * removing the kernel mapping. pmap_puninit() will take care of it.
1243 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1246 * This code optimizes the case of freeing non-busy
1247 * page-table pages. Those pages are zero now, and
1248 * might as well be placed directly into the zero queue.
1250 if (vm_page_busy_try(p, TRUE)) {
1251 vm_page_sleep_busy(p, TRUE, "pmaprl");
1256 * Remove the page table page from the processes address space.
1258 if (p->pindex == NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL) {
1260 * We are the pml4 table itself.
1262 /* XXX anything to do here? */
1263 } else if (p->pindex >= (NUPT_TOTAL + NUPD_TOTAL)) {
1265 * We are a PDP page.
1266 * We look for the PML4 entry that points to us.
1272 m4 = vm_page_lookup(pmap->pm_pteobj,
1273 NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
1274 KKASSERT(m4 != NULL);
1275 pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1276 idx = (p->pindex - (NUPT_TOTAL + NUPD_TOTAL)) % NPML4EPG;
1277 KKASSERT(pml4[idx] != 0);
1279 kprintf("pmap_release: Unmapped PML4\n");
1281 vm_page_unwire_quick(m4);
1282 } else if (p->pindex >= NUPT_TOTAL) {
1285 * We look for the PDP entry that points to us.
1291 m3 = vm_page_lookup(pmap->pm_pteobj,
1292 NUPT_TOTAL + NUPD_TOTAL +
1293 (p->pindex - NUPT_TOTAL) / NPDPEPG);
1294 KKASSERT(m3 != NULL);
1295 pdp = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1296 idx = (p->pindex - NUPT_TOTAL) % NPDPEPG;
1297 KKASSERT(pdp[idx] != 0);
1299 kprintf("pmap_release: Unmapped PDP %d\n", idx);
1301 vm_page_unwire_quick(m3);
1303 /* We are a PT page.
1304 * We look for the PD entry that points to us.
1310 m2 = vm_page_lookup(pmap->pm_pteobj,
1311 NUPT_TOTAL + p->pindex / NPDEPG);
1312 KKASSERT(m2 != NULL);
1313 pd = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1314 idx = p->pindex % NPDEPG;
1316 kprintf("pmap_release: Unmapped PD %d\n", idx);
1318 vm_page_unwire_quick(m2);
1320 KKASSERT(pmap->pm_stats.resident_count > 0);
1321 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1323 if (p->wire_count > 1) {
1324 panic("pmap_release: freeing held pt page "
1325 "pmap=%p pg=%p dmap=%p pi=%ld {%ld,%ld,%ld}",
1326 pmap, p, (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(p)),
1327 p->pindex, NUPT_TOTAL, NUPD_TOTAL, NUPDP_TOTAL);
1330 if (pmap->pm_ptphint == p)
1331 pmap->pm_ptphint = NULL;
1334 * We leave the top-level page table page cached, wired, and mapped in
1335 * the pmap until the dtor function (pmap_puninit()) gets called.
1336 * However, still clean it up.
1338 if (p->pindex == NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL) {
1339 bzero(pmap->pm_pml4, PAGE_SIZE);
1342 vm_page_unwire(p, 0);
1343 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1345 atomic_add_long(&pmap->pm_stats.wired_count, -1);
1351 * Locate the requested PT, PD, or PDP page table page.
1353 * Returns a busied page, caller must vm_page_wakeup() when done.
1356 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1365 * Find or fabricate a new pagetable page. A non-zero wire_count
1366 * indicates that the page has already been mapped into its parent.
1368 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1369 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1370 if (m->wire_count != 0)
1374 * Map the page table page into its parent, giving it 1 wire count.
1378 atomic_add_long(&pmap->pm_stats.resident_count, 1);
1379 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1381 data = VM_PAGE_TO_PHYS(m) |
1382 VPTE_RW | VPTE_V | VPTE_U | VPTE_A | VPTE_M | VPTE_WIRED;
1383 atomic_add_long(&pmap->pm_stats.wired_count, 1);
1385 if (ptepindex >= (NUPT_TOTAL + NUPD_TOTAL)) {
1387 * Map PDP into the PML4
1389 pindex = ptepindex - (NUPT_TOTAL + NUPD_TOTAL);
1390 pindex &= (NUPDP_TOTAL - 1);
1391 ptep = (pt_entry_t *)pmap->pm_pml4;
1393 } else if (ptepindex >= NUPT_TOTAL) {
1395 * Map PD into its PDP
1397 pindex = (ptepindex - NUPT_TOTAL) >> NPDPEPGSHIFT;
1398 pindex += NUPT_TOTAL + NUPD_TOTAL;
1399 pm = _pmap_allocpte(pmap, pindex);
1400 pindex = (ptepindex - NUPT_TOTAL) & (NPDPEPG - 1);
1401 ptep = (void *)PHYS_TO_DMAP(pm->phys_addr);
1404 * Map PT into its PD
1406 pindex = ptepindex >> NPDPEPGSHIFT;
1407 pindex += NUPT_TOTAL;
1408 pm = _pmap_allocpte(pmap, pindex);
1409 pindex = ptepindex & (NPTEPG - 1);
1410 ptep = (void *)PHYS_TO_DMAP(pm->phys_addr);
1414 * Install the pte in (pm). (m) prevents races.
1417 data = atomic_swap_long(ptep, data);
1419 vm_page_wire_quick(pm);
1422 pmap->pm_ptphint = pm;
1428 * Determine the page table page required to access the VA in the pmap
1429 * and allocate it if necessary. Return a held vm_page_t for the page.
1431 * Only used with user pmaps.
1434 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1436 vm_pindex_t ptepindex;
1439 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1442 * Calculate pagetable page index, and return the PT page to
1445 ptepindex = pmap_pt_pindex(va);
1446 m = _pmap_allocpte(pmap, ptepindex);
1451 /***************************************************
1452 * Pmap allocation/deallocation routines.
1453 ***************************************************/
1456 * Release any resources held by the given physical map.
1457 * Called when a pmap initialized by pmap_pinit is being released.
1458 * Should only be called if the map contains no valid mappings.
1460 static int pmap_release_callback(struct vm_page *p, void *data);
1463 pmap_release(struct pmap *pmap)
1465 vm_object_t object = pmap->pm_pteobj;
1466 struct rb_vm_page_scan_info info;
1468 KKASSERT(pmap != kernel_pmap);
1470 #if defined(DIAGNOSTIC)
1471 if (object->ref_count != 1)
1472 panic("pmap_release: pteobj reference count != 1");
1476 info.object = object;
1478 KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
1479 ("pmap %p still active! %016jx",
1481 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
1483 vm_object_hold(object);
1487 info.limit = object->generation;
1489 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1490 pmap_release_callback, &info);
1491 if (info.error == 0 && info.mpte) {
1492 if (pmap_release_free_page(pmap, info.mpte))
1495 } while (info.error);
1497 pmap->pm_ptphint = NULL;
1499 KASSERT((pmap->pm_stats.wired_count == (pmap->pm_pdirm != NULL)),
1500 ("pmap_release: dangling count %p %ld",
1501 pmap, pmap->pm_stats.wired_count));
1503 vm_object_drop(object);
1507 pmap_release_callback(struct vm_page *p, void *data)
1509 struct rb_vm_page_scan_info *info = data;
1511 if (p->pindex == NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL) {
1515 if (pmap_release_free_page(info->pmap, p)) {
1519 if (info->object->generation != info->limit) {
1527 * Grow the number of kernel page table entries, if needed.
1529 * kernel_map must be locked exclusively by the caller.
1532 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1536 vm_offset_t ptppaddr;
1538 pd_entry_t *pde, newpdir;
1543 vm_object_hold(&kptobj);
1544 if (kernel_vm_end == 0) {
1545 kernel_vm_end = KvaStart;
1547 while ((*pmap_pde(kernel_pmap, kernel_vm_end) & VPTE_V) != 0) {
1549 rounddown2(kernel_vm_end + PAGE_SIZE * NPTEPG,
1550 PAGE_SIZE * NPTEPG);
1552 if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
1553 kernel_vm_end = vm_map_max(kernel_map);
1558 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1559 if (addr - 1 >= vm_map_max(kernel_map))
1560 addr = vm_map_max(kernel_map);
1561 while (kernel_vm_end < addr) {
1562 pde = pmap_pde(kernel_pmap, kernel_vm_end);
1564 /* We need a new PDP entry */
1565 nkpg = vm_page_alloc(&kptobj, nkpt,
1568 VM_ALLOC_INTERRUPT);
1570 panic("pmap_growkernel: no memory to "
1573 paddr = VM_PAGE_TO_PHYS(nkpg);
1574 pmap_zero_page(paddr);
1575 newpdp = (pdp_entry_t)(paddr |
1576 VPTE_V | VPTE_RW | VPTE_U |
1577 VPTE_A | VPTE_M | VPTE_WIRED);
1578 *pmap_pdpe(kernel_pmap, kernel_vm_end) = newpdp;
1579 atomic_add_long(&kernel_pmap->pm_stats.wired_count, 1);
1581 continue; /* try again */
1583 if ((*pde & VPTE_V) != 0) {
1585 rounddown2(kernel_vm_end + PAGE_SIZE * NPTEPG,
1586 PAGE_SIZE * NPTEPG);
1587 if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
1588 kernel_vm_end = vm_map_max(kernel_map);
1595 * This index is bogus, but out of the way
1597 nkpg = vm_page_alloc(&kptobj, nkpt,
1600 VM_ALLOC_INTERRUPT);
1602 panic("pmap_growkernel: no memory to grow kernel");
1605 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1606 pmap_zero_page(ptppaddr);
1607 newpdir = (pd_entry_t)(ptppaddr |
1608 VPTE_V | VPTE_RW | VPTE_U |
1609 VPTE_A | VPTE_M | VPTE_WIRED);
1610 *pmap_pde(kernel_pmap, kernel_vm_end) = newpdir;
1611 atomic_add_long(&kernel_pmap->pm_stats.wired_count, 1);
1615 rounddown2(kernel_vm_end + PAGE_SIZE * NPTEPG,
1616 PAGE_SIZE * NPTEPG);
1617 if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) {
1618 kernel_vm_end = vm_map_max(kernel_map);
1622 vm_object_drop(&kptobj);
1626 * Add a reference to the specified pmap.
1631 pmap_reference(pmap_t pmap)
1634 atomic_add_int(&pmap->pm_count, 1);
1637 /************************************************************************
1638 * VMSPACE MANAGEMENT *
1639 ************************************************************************
1641 * The VMSPACE management we do in our virtual kernel must be reflected
1642 * in the real kernel. This is accomplished by making vmspace system
1643 * calls to the real kernel.
1646 cpu_vmspace_alloc(struct vmspace *vm)
1652 #define USER_SIZE (VM_MAX_USER_ADDRESS - VM_MIN_USER_ADDRESS)
1654 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
1655 panic("vmspace_create() failed");
1657 rp = vmspace_mmap(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE,
1658 PROT_READ|PROT_WRITE|PROT_EXEC,
1659 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
1661 if (rp == MAP_FAILED)
1662 panic("vmspace_mmap: failed");
1663 vmspace_mcontrol(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE,
1665 vpte = VM_PAGE_TO_PHYS(vmspace_pmap(vm)->pm_pdirm) |
1666 VPTE_RW | VPTE_V | VPTE_U;
1667 r = vmspace_mcontrol(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE,
1670 panic("vmspace_mcontrol: failed");
1674 cpu_vmspace_free(struct vmspace *vm)
1676 if (vmspace_destroy(&vm->vm_pmap) < 0)
1677 panic("vmspace_destroy() failed");
1680 /***************************************************
1681 * page management routines.
1682 ***************************************************/
1685 * free the pv_entry back to the free list. This function may be
1686 * called from an interrupt.
1688 static __inline void
1689 free_pv_entry(pv_entry_t pv)
1691 atomic_add_long(&pv_entry_count, -1);
1696 * get a new pv_entry, allocating a block from the system
1697 * when needed. This function may be called from an interrupt.
1702 atomic_add_long(&pv_entry_count, 1);
1703 if (pv_entry_high_water &&
1704 (pv_entry_count > pv_entry_high_water) &&
1705 atomic_swap_int(&pmap_pagedaemon_waken, 1) == 0) {
1706 wakeup(&vm_pages_needed);
1708 return zalloc(pvzone);
1712 * This routine is very drastic, but can save the system
1722 static int warningdone=0;
1724 if (pmap_pagedaemon_waken == 0)
1726 pmap_pagedaemon_waken = 0;
1728 if (warningdone < 5) {
1729 kprintf("pmap_collect: collecting pv entries -- "
1730 "suggest increasing PMAP_SHPGPERPROC\n");
1734 for (i = 0; i < vm_page_array_size; i++) {
1735 m = &vm_page_array[i];
1736 if (m->wire_count || m->hold_count)
1738 if (vm_page_busy_try(m, TRUE) == 0) {
1739 if (m->wire_count == 0 && m->hold_count == 0) {
1749 * If it is the first entry on the list, it is actually
1750 * in the header and we must copy the following entry up
1751 * to the header. Otherwise we must search the list for
1752 * the entry. In either case we free the now unused entry.
1754 * pmap->pm_pteobj must be held and (m) must be spin-locked by the caller.
1757 pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va)
1762 vm_page_spin_lock(m);
1763 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, va);
1766 * Note that pv_ptem is NULL if the page table page itself is not
1767 * managed, even if the page being removed IS managed.
1771 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1772 if (TAILQ_EMPTY(&m->md.pv_list))
1773 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1774 m->md.pv_list_count--;
1775 KKASSERT(m->md.pv_list_count >= 0);
1776 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
1777 atomic_add_int(&pmap->pm_generation, 1);
1778 vm_page_spin_unlock(m);
1779 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1782 vm_page_spin_unlock(m);
1783 kprintf("pmap_remove_entry: could not find "
1784 "pmap=%p m=%p va=%016jx\n",
1791 * Create a pv entry for page at pa for (pmap, va). If the page table page
1792 * holding the VA is managed, mpte will be non-NULL.
1794 * pmap->pm_pteobj must be held and (m) must be spin-locked by the caller.
1797 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m,
1804 m->md.pv_list_count++;
1805 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1806 pv = pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pv);
1807 vm_page_flag_set(m, PG_MAPPED);
1808 KKASSERT(pv == NULL);
1812 * pmap_remove_pte: do the things to unmap a page in a process
1814 * Caller holds pmap->pm_pteobj and holds the associated page table
1815 * page busy to prevent races.
1818 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, pt_entry_t oldpte,
1825 oldpte = pmap_inval_loadandclear(ptq, pmap, va);
1827 if (oldpte & VPTE_WIRED)
1828 atomic_add_long(&pmap->pm_stats.wired_count, -1);
1829 KKASSERT(pmap->pm_stats.wired_count >= 0);
1833 * Machines that don't support invlpg, also don't support
1834 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1838 cpu_invlpg((void *)va);
1840 KKASSERT(pmap->pm_stats.resident_count > 0);
1841 atomic_add_long(&pmap->pm_stats.resident_count, -1);
1842 if (oldpte & VPTE_MANAGED) {
1843 m = PHYS_TO_VM_PAGE(oldpte);
1846 * NOTE: pmap_remove_entry() will spin-lock the page
1848 if (oldpte & VPTE_M) {
1849 #if defined(PMAP_DIAGNOSTIC)
1850 if (pmap_nw_modified(oldpte)) {
1851 kprintf("pmap_remove: modified page not "
1852 "writable: va: 0x%lx, pte: 0x%lx\n",
1856 pmap_track_modified(pmap, va);
1859 if (oldpte & VPTE_A)
1860 vm_page_flag_set(m, PG_REFERENCED);
1861 error = pmap_remove_entry(pmap, m, va);
1863 error = pmap_unuse_pt(pmap, va, NULL);
1871 * Remove a single page from a process address space.
1873 * This function may not be called from an interrupt if the pmap is
1876 * Caller holds pmap->pm_pteobj
1879 pmap_remove_page(struct pmap *pmap, vm_offset_t va)
1883 pte = pmap_pte(pmap, va);
1886 if ((*pte & VPTE_V) == 0)
1888 pmap_remove_pte(pmap, pte, 0, va);
1892 * Remove the given range of addresses from the specified map.
1894 * It is assumed that the start and end are properly rounded to
1897 * This function may not be called from an interrupt if the pmap is
1903 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
1905 vm_offset_t va_next;
1906 pml4_entry_t *pml4e;
1908 pd_entry_t ptpaddr, *pde;
1915 vm_object_hold(pmap->pm_pteobj);
1916 KKASSERT(pmap->pm_stats.resident_count >= 0);
1917 if (pmap->pm_stats.resident_count == 0) {
1918 vm_object_drop(pmap->pm_pteobj);
1923 * special handling of removing one page. a very
1924 * common operation and easy to short circuit some
1927 if (sva + PAGE_SIZE == eva) {
1928 pde = pmap_pde(pmap, sva);
1929 if (pde && (*pde & VPTE_PS) == 0) {
1930 pmap_remove_page(pmap, sva);
1931 vm_object_drop(pmap->pm_pteobj);
1936 for (; sva < eva; sva = va_next) {
1937 pml4e = pmap_pml4e(pmap, sva);
1938 if ((*pml4e & VPTE_V) == 0) {
1939 va_next = (sva + NBPML4) & ~PML4MASK;
1945 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
1946 if ((*pdpe & VPTE_V) == 0) {
1947 va_next = (sva + NBPDP) & ~PDPMASK;
1954 * Calculate index for next page table.
1956 va_next = (sva + NBPDR) & ~PDRMASK;
1960 pde = pmap_pdpe_to_pde(pdpe, sva);
1964 * Weed out invalid mappings.
1970 * Check for large page.
1972 if ((ptpaddr & VPTE_PS) != 0) {
1973 /* JG FreeBSD has more complex treatment here */
1974 KKASSERT(*pde != 0);
1975 pmap_inval_pde(pde, pmap, sva);
1976 atomic_add_long(&pmap->pm_stats.resident_count,
1977 -NBPDR / PAGE_SIZE);
1982 * Limit our scan to either the end of the va represented
1983 * by the current page table page, or to the end of the
1984 * range being removed.
1990 * NOTE: pmap_remove_pte() can block.
1992 pt_m = pmap_hold_pt_page(pde, sva);
1993 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
1996 if (pmap_remove_pte(pmap, pte, 0, sva))
2000 vm_page_unhold(pt_m);
2002 vm_object_drop(pmap->pm_pteobj);
2006 * Removes this physical page from all physical maps in which it resides.
2007 * Reflects back modify bits to the pager.
2009 * This routine may not be called from an interrupt.
2014 pmap_remove_all(vm_page_t m)
2016 pt_entry_t *pte, tpte;
2021 #if defined(PMAP_DIAGNOSTIC)
2023 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
2026 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
2027 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
2032 vm_page_spin_lock(m);
2033 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2035 pmobj = pmap->pm_pteobj;
2038 * Handle reversed lock ordering
2040 if (vm_object_hold_try(pmobj) == 0) {
2041 refcount_acquire(&pmobj->hold_count);
2042 vm_page_spin_unlock(m);
2043 vm_object_lock(pmobj);
2044 vm_page_spin_lock(m);
2045 if (pv != TAILQ_FIRST(&m->md.pv_list) ||
2046 pmap != pv->pv_pmap ||
2047 pmobj != pmap->pm_pteobj) {
2048 vm_page_spin_unlock(m);
2049 vm_object_drop(pmobj);
2054 KKASSERT(pmap->pm_stats.resident_count > 0);
2055 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2057 pte = pmap_pte(pmap, pv->pv_va);
2058 KKASSERT(pte != NULL);
2060 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2061 if (tpte & VPTE_WIRED)
2062 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2063 KKASSERT(pmap->pm_stats.wired_count >= 0);
2066 vm_page_flag_set(m, PG_REFERENCED);
2069 * Update the vm_page_t clean and reference bits.
2071 if (tpte & VPTE_M) {
2072 #if defined(PMAP_DIAGNOSTIC)
2073 if (pmap_nw_modified(tpte)) {
2075 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2079 pmap_track_modified(pmap, pv->pv_va);
2082 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2083 if (TAILQ_EMPTY(&m->md.pv_list))
2084 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2085 m->md.pv_list_count--;
2086 KKASSERT(m->md.pv_list_count >= 0);
2087 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2088 atomic_add_int(&pmap->pm_generation, 1);
2089 vm_page_spin_unlock(m);
2090 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2093 vm_object_drop(pmobj);
2094 vm_page_spin_lock(m);
2096 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2097 vm_page_spin_unlock(m);
2101 * Removes the page from a particular pmap
2104 pmap_remove_specific(pmap_t pmap, vm_page_t m)
2106 pt_entry_t *pte, tpte;
2109 vm_object_hold(pmap->pm_pteobj);
2111 vm_page_spin_lock(m);
2112 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2113 if (pv->pv_pmap != pmap)
2116 KKASSERT(pmap->pm_stats.resident_count > 0);
2117 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2119 pte = pmap_pte(pmap, pv->pv_va);
2120 KKASSERT(pte != NULL);
2122 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2123 if (tpte & VPTE_WIRED)
2124 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2125 KKASSERT(pmap->pm_stats.wired_count >= 0);
2128 vm_page_flag_set(m, PG_REFERENCED);
2131 * Update the vm_page_t clean and reference bits.
2133 if (tpte & VPTE_M) {
2134 pmap_track_modified(pmap, pv->pv_va);
2137 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2138 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2139 atomic_add_int(&pmap->pm_generation, 1);
2140 m->md.pv_list_count--;
2141 KKASSERT(m->md.pv_list_count >= 0);
2142 if (TAILQ_EMPTY(&m->md.pv_list))
2143 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2144 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2145 vm_page_spin_unlock(m);
2149 vm_page_spin_unlock(m);
2150 vm_object_drop(pmap->pm_pteobj);
2154 * Set the physical protection on the specified range of this map
2157 * This function may not be called from an interrupt if the map is
2158 * not the kernel_pmap.
2163 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2165 vm_offset_t va_next;
2166 pml4_entry_t *pml4e;
2168 pd_entry_t ptpaddr, *pde;
2175 if ((prot & (VM_PROT_READ | VM_PROT_EXECUTE)) == VM_PROT_NONE) {
2176 pmap_remove(pmap, sva, eva);
2180 if (prot & VM_PROT_WRITE)
2183 vm_object_hold(pmap->pm_pteobj);
2185 for (; sva < eva; sva = va_next) {
2186 pml4e = pmap_pml4e(pmap, sva);
2187 if ((*pml4e & VPTE_V) == 0) {
2188 va_next = (sva + NBPML4) & ~PML4MASK;
2194 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2195 if ((*pdpe & VPTE_V) == 0) {
2196 va_next = (sva + NBPDP) & ~PDPMASK;
2202 va_next = (sva + NBPDR) & ~PDRMASK;
2206 pde = pmap_pdpe_to_pde(pdpe, sva);
2211 * Check for large page.
2213 if ((ptpaddr & VPTE_PS) != 0) {
2215 pmap_clean_pde(pde, pmap, sva);
2216 atomic_add_long(&pmap->pm_stats.resident_count,
2217 -NBPDR / PAGE_SIZE);
2223 * Weed out invalid mappings. Note: we assume that the page
2224 * directory table is always allocated, and in kernel virtual.
2232 pt_m = pmap_hold_pt_page(pde, sva);
2233 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2236 * Clean managed pages and also check the accessed
2237 * bit. Just remove write perms for unmanaged
2238 * pages. Be careful of races, turning off write
2239 * access will force a fault rather then setting
2240 * the modified bit at an unexpected time.
2242 pmap_track_modified(pmap, sva);
2243 pmap_clean_pte(pte, pmap, sva, NULL);
2245 vm_page_unhold(pt_m);
2247 vm_object_drop(pmap->pm_pteobj);
2251 * Enter a managed page into a pmap. If the page is not wired related pmap
2252 * data can be destroyed at any time for later demand-operation.
2254 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
2255 * specified protection, and wire the mapping if requested.
2257 * NOTE: This routine may not lazy-evaluate or lose information. The
2258 * page must actually be inserted into the given map NOW.
2260 * NOTE: When entering a page at a KVA address, the pmap must be the
2266 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2267 boolean_t wired, vm_map_entry_t entry __unused)
2272 pt_entry_t origpte, newpte;
2279 va = trunc_page(va);
2281 vm_object_hold(pmap->pm_pteobj);
2284 * Get the page table page. The kernel_pmap's page table pages
2285 * are preallocated and have no associated vm_page_t.
2287 * If not NULL, mpte will be busied and we must vm_page_wakeup()
2288 * to cleanup. There will already be at least one wire count from
2289 * it being mapped into its parent.
2291 if (pmap == kernel_pmap) {
2295 mpte = pmap_allocpte(pmap, va);
2296 pte = (void *)PHYS_TO_DMAP(mpte->phys_addr);
2297 pte += pmap_pte_index(va);
2301 * Deal with races against the kernel's real MMU by cleaning the
2302 * page, even if we are re-entering the same page.
2304 pa = VM_PAGE_TO_PHYS(m);
2305 origpte = pmap_inval_loadandclear(pte, pmap, va);
2306 /*origpte = pmap_clean_pte(pte, pmap, va, NULL);*/
2307 opa = origpte & VPTE_FRAME;
2309 if (origpte & VPTE_PS)
2310 panic("pmap_enter: attempted pmap_enter on 2MB page");
2312 if ((origpte & (VPTE_MANAGED|VPTE_M)) == (VPTE_MANAGED|VPTE_M)) {
2315 pmap_track_modified(pmap, va);
2316 om = PHYS_TO_VM_PAGE(opa);
2321 * Mapping has not changed, must be protection or wiring change.
2323 if (origpte && (opa == pa)) {
2325 * Wiring change, just update stats. We don't worry about
2326 * wiring PT pages as they remain resident as long as there
2327 * are valid mappings in them. Hence, if a user page is wired,
2328 * the PT page will be also.
2330 if (wired && ((origpte & VPTE_WIRED) == 0))
2331 atomic_add_long(&pmap->pm_stats.wired_count, 1);
2332 else if (!wired && (origpte & VPTE_WIRED))
2333 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2335 if (origpte & VPTE_MANAGED) {
2337 KKASSERT(m->flags & PG_MAPPED);
2338 KKASSERT((m->flags & PG_FICTITIOUS) == 0);
2340 KKASSERT((m->flags & PG_FICTITIOUS));
2342 vm_page_spin_lock(m);
2347 * Bump the wire_count for the page table page.
2350 vm_page_wire_quick(mpte);
2353 * Mapping has changed, invalidate old range and fall through to
2354 * handle validating new mapping. Don't inherit anything from
2359 err = pmap_remove_pte(pmap, NULL, origpte, va);
2362 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2366 * Enter on the PV list if part of our managed memory. Note that we
2367 * raise IPL while manipulating pv_table since pmap_enter can be
2368 * called at interrupt time.
2370 if (pmap_initialized) {
2371 if ((m->flags & PG_FICTITIOUS) == 0) {
2373 * WARNING! We are using m's spin-lock as a
2374 * man's pte lock to interlock against
2375 * pmap_page_protect() operations.
2377 * This is a bad hack (obviously).
2379 pv = get_pv_entry();
2380 vm_page_spin_lock(m);
2381 pmap_insert_entry(pmap, va, mpte, m, pv);
2383 /* vm_page_spin_unlock(m); */
2385 vm_page_spin_lock(m);
2388 vm_page_spin_lock(m);
2392 * Increment counters
2394 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2396 atomic_add_long(&pmap->pm_stats.wired_count, 1);
2400 * Now validate mapping with desired protection/wiring.
2402 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | VPTE_V | VPTE_U);
2406 newpte |= VPTE_WIRED;
2407 // if (pmap != kernel_pmap)
2409 if (newpte & VPTE_RW)
2410 vm_page_flag_set(m, PG_WRITEABLE);
2411 KKASSERT((newpte & VPTE_MANAGED) == 0 || (m->flags & PG_MAPPED));
2413 origpte = atomic_swap_long(pte, newpte);
2414 if (origpte & VPTE_M) {
2415 kprintf("pmap [M] race @ %016jx\n", va);
2416 atomic_set_long(pte, VPTE_M);
2418 vm_page_spin_unlock(m);
2421 vm_page_wakeup(mpte);
2422 vm_object_drop(pmap->pm_pteobj);
2426 * Make a temporary mapping for a physical address. This is only intended
2427 * to be used for panic dumps.
2429 * The caller is responsible for calling smp_invltlb().
2432 pmap_kenter_temporary(vm_paddr_t pa, long i)
2434 pmap_kenter_quick(crashdumpmap + (i * PAGE_SIZE), pa);
2435 return ((void *)crashdumpmap);
2438 #define MAX_INIT_PT (96)
2441 * This routine preloads the ptes for a given object into the specified pmap.
2442 * This eliminates the blast of soft faults on process startup and
2443 * immediately after an mmap.
2447 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2450 pmap_object_init_pt(pmap_t pmap, vm_map_entry_t entry,
2451 vm_offset_t addr, vm_size_t size, int limit)
2453 vm_prot_t prot = entry->protection;
2454 vm_object_t object = entry->ba.object;
2455 vm_pindex_t pindex = atop(entry->ba.offset + (addr - entry->ba.start));
2456 struct rb_vm_page_scan_info info;
2461 * We can't preinit if read access isn't set or there is no pmap
2464 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2468 * We can't preinit if the pmap is not the current pmap
2470 lp = curthread->td_lwp;
2471 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2475 * Misc additional checks
2477 psize = x86_64_btop(size);
2479 if ((object->type != OBJT_VNODE) ||
2480 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2481 (object->resident_page_count > MAX_INIT_PT))) {
2485 if (psize + pindex > object->size) {
2486 if (object->size < pindex)
2488 psize = object->size - pindex;
2495 * Use a red-black scan to traverse the requested range and load
2496 * any valid pages found into the pmap.
2498 * We cannot safely scan the object's memq unless we are in a
2499 * critical section since interrupts can remove pages from objects.
2501 info.start_pindex = pindex;
2502 info.end_pindex = pindex + psize - 1;
2509 vm_object_hold_shared(object);
2510 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2511 pmap_object_init_pt_callback, &info);
2512 vm_object_drop(object);
2517 pmap_object_init_pt_callback(vm_page_t p, void *data)
2519 struct rb_vm_page_scan_info *info = data;
2520 vm_pindex_t rel_index;
2522 * don't allow an madvise to blow away our really
2523 * free pages allocating pv entries.
2525 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2526 vmstats.v_free_count < vmstats.v_free_reserved) {
2531 * Ignore list markers and ignore pages we cannot instantly
2532 * busy (while holding the object token).
2534 if (p->flags & PG_MARKER)
2536 if (vm_page_busy_try(p, TRUE))
2538 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2539 (p->flags & PG_FICTITIOUS) == 0) {
2540 if ((p->queue - p->pc) == PQ_CACHE)
2541 vm_page_deactivate(p);
2542 rel_index = p->pindex - info->start_pindex;
2543 pmap_enter(info->pmap, info->addr + x86_64_ptob(rel_index), p,
2544 VM_PROT_READ, FALSE, info->entry);
2551 * Return TRUE if the pmap is in shape to trivially
2552 * pre-fault the specified address.
2554 * Returns FALSE if it would be non-trivial or if a
2555 * pte is already loaded into the slot.
2560 pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
2566 vm_object_hold(pmap->pm_pteobj);
2567 pde = pmap_pde(pmap, addr);
2568 if (pde == NULL || *pde == 0) {
2571 pte = pmap_pde_to_pte(pde, addr);
2572 ret = (*pte) ? 0 : 1;
2574 vm_object_drop(pmap->pm_pteobj);
2580 * Change the wiring attribute for a map/virtual-address pair.
2582 * The mapping must already exist in the pmap.
2583 * No other requirements.
2586 pmap_unwire(pmap_t pmap, vm_offset_t va)
2595 vm_object_hold(pmap->pm_pteobj);
2596 pte = pmap_pte(pmap, va);
2598 if (pte == NULL || (*pte & VPTE_V) == 0) {
2599 vm_object_drop(pmap->pm_pteobj);
2604 * Wiring is not a hardware characteristic so there is no need to
2605 * invalidate TLB. However, in an SMP environment we must use
2606 * a locked bus cycle to update the pte (if we are not using
2607 * the pmap_inval_*() API that is)... it's ok to do this for simple
2610 if (pmap_pte_w(pte))
2611 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2612 /* XXX else return NULL so caller doesn't unwire m ? */
2613 atomic_clear_long(pte, VPTE_WIRED);
2615 pa = *pte & VPTE_FRAME;
2616 m = PHYS_TO_VM_PAGE(pa); /* held by wired count */
2618 vm_object_drop(pmap->pm_pteobj);
2624 * Copy the range specified by src_addr/len
2625 * from the source map to the range dst_addr/len
2626 * in the destination map.
2628 * This routine is only advisory and need not do anything.
2631 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2632 vm_size_t len, vm_offset_t src_addr)
2635 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2636 * valid through blocking calls, and that's just not going to
2647 * Zero the specified physical page.
2649 * This function may be called from an interrupt and no locking is
2653 pmap_zero_page(vm_paddr_t phys)
2655 vm_offset_t va = PHYS_TO_DMAP(phys);
2657 bzero((void *)va, PAGE_SIZE);
2663 * Zero part of a physical page by mapping it into memory and clearing
2664 * its contents with bzero.
2666 * off and size may not cover an area beyond a single hardware page.
2669 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2671 vm_offset_t virt = PHYS_TO_DMAP(phys);
2673 bzero((char *)virt + off, size);
2679 * Copy the physical page from the source PA to the target PA.
2680 * This function may be called from an interrupt. No locking
2684 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2686 vm_offset_t src_virt, dst_virt;
2688 src_virt = PHYS_TO_DMAP(src);
2689 dst_virt = PHYS_TO_DMAP(dst);
2690 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
2694 * pmap_copy_page_frag:
2696 * Copy the physical page from the source PA to the target PA.
2697 * This function may be called from an interrupt. No locking
2701 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2703 vm_offset_t src_virt, dst_virt;
2705 src_virt = PHYS_TO_DMAP(src);
2706 dst_virt = PHYS_TO_DMAP(dst);
2707 bcopy((char *)src_virt + (src & PAGE_MASK),
2708 (char *)dst_virt + (dst & PAGE_MASK),
2713 * Remove all pages from specified address space this aids process
2714 * exit speeds. Also, this code is special cased for current
2715 * process only, but can have the more generic (and slightly slower)
2716 * mode enabled. This is much faster than pmap_remove in the case
2717 * of running down an entire address space.
2719 * No other requirements.
2722 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2724 pmap_remove(pmap, sva, eva);
2726 pt_entry_t *pte, tpte;
2729 int save_generation;
2731 if (pmap->pm_pteobj)
2732 vm_object_hold(pmap->pm_pteobj);
2734 pmap_invalidate_range(pmap, sva, eva);
2736 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2737 if (pv->pv_va >= eva || pv->pv_va < sva) {
2738 npv = TAILQ_NEXT(pv, pv_plist);
2742 KKASSERT(pmap == pv->pv_pmap);
2744 pte = pmap_pte(pmap, pv->pv_va);
2747 * We cannot remove wired pages from a process' mapping
2750 if (*pte & VPTE_WIRED) {
2751 npv = TAILQ_NEXT(pv, pv_plist);
2754 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
2756 m = PHYS_TO_VM_PAGE(tpte & VPTE_FRAME);
2757 vm_page_spin_lock(m);
2759 KASSERT(m < &vm_page_array[vm_page_array_size],
2760 ("pmap_remove_pages: bad tpte %lx", tpte));
2762 KKASSERT(pmap->pm_stats.resident_count > 0);
2763 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2766 * Update the vm_page_t clean and reference bits.
2768 if (tpte & VPTE_M) {
2772 npv = TAILQ_NEXT(pv, pv_plist);
2773 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2774 atomic_add_int(&pmap->pm_generation, 1);
2775 save_generation = pmap->pm_generation;
2776 m->md.pv_list_count--;
2777 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2778 if (TAILQ_EMPTY(&m->md.pv_list))
2779 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2780 vm_page_spin_unlock(m);
2782 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem);
2786 * Restart the scan if we blocked during the unuse or free
2787 * calls and other removals were made.
2789 if (save_generation != pmap->pm_generation) {
2790 kprintf("Warning: pmap_remove_pages race-A avoided\n");
2791 npv = TAILQ_FIRST(&pmap->pm_pvlist);
2794 if (pmap->pm_pteobj)
2795 vm_object_drop(pmap->pm_pteobj);
2796 pmap_remove(pmap, sva, eva);
2801 * pmap_testbit tests bits in active mappings of a VM page.
2804 pmap_testbit(vm_page_t m, int bit)
2809 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2812 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2815 vm_page_spin_lock(m);
2816 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2818 * if the bit being tested is the modified bit, then
2819 * mark clean_map and ptes as never
2822 if (bit & (VPTE_A|VPTE_M))
2823 pmap_track_modified(pv->pv_pmap, pv->pv_va);
2825 #if defined(PMAP_DIAGNOSTIC)
2826 if (pv->pv_pmap == NULL) {
2827 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
2831 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2833 vm_page_spin_unlock(m);
2837 vm_page_spin_unlock(m);
2842 * This routine is used to clear bits in ptes. Certain bits require special
2843 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
2845 * This routine is only called with certain VPTE_* bit combinations.
2847 static __inline void
2848 pmap_clearbit(vm_page_t m, int bit)
2856 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
2858 vm_page_flag_clear(m, PG_WRITEABLE);
2863 * Loop over all current mappings setting/clearing as appropos If
2864 * setting RO do we need to clear the VAC?
2867 vm_page_spin_lock(m);
2868 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2870 * Need the pmap object lock(?)
2873 pmobj = pmap->pm_pteobj;
2875 if (vm_object_hold_try(pmobj) == 0) {
2876 refcount_acquire(&pmobj->hold_count);
2877 vm_page_spin_unlock(m);
2878 vm_object_lock(pmobj);
2879 vm_object_drop(pmobj);
2884 * don't write protect pager mappings
2886 if (bit == VPTE_RW) {
2887 pmap_track_modified(pv->pv_pmap, pv->pv_va);
2890 #if defined(PMAP_DIAGNOSTIC)
2891 if (pv->pv_pmap == NULL) {
2892 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
2893 vm_object_drop(pmobj);
2899 * Careful here. We can use a locked bus instruction to
2900 * clear VPTE_A or VPTE_M safely but we need to synchronize
2901 * with the target cpus when we mess with VPTE_RW.
2903 * On virtual kernels we must force a new fault-on-write
2904 * in the real kernel if we clear the Modify bit ourselves,
2905 * otherwise the real kernel will not get a new fault and
2906 * will never set our Modify bit again.
2908 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2910 if (bit == VPTE_RW) {
2912 * We must also clear VPTE_M when clearing
2913 * VPTE_RW and synchronize its state to
2916 pmap_track_modified(pv->pv_pmap, pv->pv_va);
2917 pbits = pmap_clean_pte(pte, pv->pv_pmap,
2919 } else if (bit == VPTE_M) {
2921 * We must invalidate the real-kernel pte
2922 * when clearing VPTE_M bit to force the
2923 * real-kernel to take a new fault to re-set
2926 atomic_clear_long(pte, VPTE_M);
2927 if (*pte & VPTE_RW) {
2928 pmap_invalidate_range(pv->pv_pmap,
2930 pv->pv_va + PAGE_SIZE);
2932 } else if ((bit & (VPTE_RW|VPTE_M)) ==
2935 * We've been asked to clear W & M, I guess
2936 * the caller doesn't want us to update
2937 * the dirty status of the VM page.
2939 pmap_track_modified(pv->pv_pmap, pv->pv_va);
2940 pmap_clean_pte(pte, pv->pv_pmap, pv->pv_va, m);
2941 panic("shouldn't be called");
2944 * We've been asked to clear bits that do
2945 * not interact with hardware.
2947 atomic_clear_long(pte, bit);
2950 vm_object_drop(pmobj);
2953 vm_page_flag_clear(m, PG_WRITEABLE);
2954 vm_page_spin_unlock(m);
2958 * Lower the permission for all mappings to a given page.
2960 * No other requirements.
2963 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2965 if ((prot & VM_PROT_WRITE) == 0) {
2966 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2967 pmap_clearbit(m, VPTE_RW);
2975 pmap_phys_address(vm_pindex_t ppn)
2977 return (x86_64_ptob(ppn));
2981 * Return a count of reference bits for a page, clearing those bits.
2982 * It is not necessary for every reference bit to be cleared, but it
2983 * is necessary that 0 only be returned when there are truly no
2984 * reference bits set.
2986 * XXX: The exact number of bits to check and clear is a matter that
2987 * should be tested and standardized at some point in the future for
2988 * optimal aging of shared pages.
2990 * No other requirements.
2993 pmap_ts_referenced(vm_page_t m)
2995 pv_entry_t pv, pvf, pvn;
2999 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3002 vm_page_spin_lock(m);
3003 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3006 pvn = TAILQ_NEXT(pv, pv_list);
3007 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3008 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3010 pmap_track_modified(pv->pv_pmap, pv->pv_va);
3011 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
3013 if (pte && (*pte & VPTE_A)) {
3014 atomic_clear_long(pte, VPTE_A);
3020 } while ((pv = pvn) != NULL && pv != pvf);
3022 vm_page_spin_unlock(m);
3028 * Return whether or not the specified physical page was modified
3029 * in any physical maps.
3031 * No other requirements.
3034 pmap_is_modified(vm_page_t m)
3038 res = pmap_testbit(m, VPTE_M);
3044 * Clear the modify bits on the specified physical page. For the vkernel
3045 * we really need to clean the page, which clears VPTE_RW and VPTE_M, in
3046 * order to ensure that we take a fault on the next write to the page.
3047 * Otherwise the page may become dirty without us knowing it.
3049 * No other requirements.
3052 pmap_clear_modify(vm_page_t m)
3054 pmap_clearbit(m, VPTE_RW);
3058 * Clear the reference bit on the specified physical page.
3060 * No other requirements.
3063 pmap_clear_reference(vm_page_t m)
3065 pmap_clearbit(m, VPTE_A);
3069 * Miscellaneous support routines follow
3072 x86_64_protection_init(void)
3077 kp = protection_codes;
3078 for (prot = 0; prot < 8; prot++) {
3079 if (prot & VM_PROT_READ)
3081 if (prot & VM_PROT_WRITE)
3082 *kp |= VPTE_RW; /* R+W */
3083 if (prot && (prot & VM_PROT_EXECUTE) == 0)
3084 *kp |= VPTE_NX; /* NX - !executable */
3090 * Sets the memory attribute for the specified page.
3093 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
3095 /* This is a vkernel, do nothing */
3099 * Change the PAT attribute on an existing kernel memory map. Caller
3100 * must ensure that the virtual memory in question is not accessed
3101 * during the adjustment.
3104 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
3106 /* This is a vkernel, do nothing */
3110 * Perform the pmap work for mincore
3112 * No other requirements.
3115 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3117 pt_entry_t *ptep, pte;
3121 vm_object_hold(pmap->pm_pteobj);
3122 ptep = pmap_pte(pmap, addr);
3124 if (ptep && (pte = *ptep) != 0) {
3127 val = MINCORE_INCORE;
3128 if ((pte & VPTE_MANAGED) == 0)
3131 pa = pte & VPTE_FRAME;
3133 m = PHYS_TO_VM_PAGE(pa);
3139 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3141 * Modified by someone
3143 else if (m->dirty || pmap_is_modified(m))
3144 val |= MINCORE_MODIFIED_OTHER;
3149 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3152 * Referenced by someone
3154 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3155 val |= MINCORE_REFERENCED_OTHER;
3156 vm_page_flag_set(m, PG_REFERENCED);
3160 vm_object_drop(pmap->pm_pteobj);
3166 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3167 * vmspace will be ref'd and the old one will be deref'd.
3169 * Caller must hold vmspace->vm_map.token for oldvm and newvm
3172 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3174 struct vmspace *oldvm;
3177 oldvm = p->p_vmspace;
3178 if (oldvm != newvm) {
3181 KKASSERT((newvm->vm_refcnt & VM_REF_DELETED) == 0);
3182 p->p_vmspace = newvm;
3183 KKASSERT(p->p_nthreads == 1);
3184 lp = RB_ROOT(&p->p_lwp_tree);
3185 pmap_setlwpvm(lp, newvm);
3192 * Set the vmspace for a LWP. The vmspace is almost universally set the
3193 * same as the process vmspace, but virtual kernels need to swap out contexts
3194 * on a per-lwp basis.
3197 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3199 struct vmspace *oldvm;
3202 oldvm = lp->lwp_vmspace;
3203 if (oldvm != newvm) {
3205 KKASSERT((newvm->vm_refcnt & VM_REF_DELETED) == 0);
3206 lp->lwp_vmspace = newvm;
3207 if (curthread->td_lwp == lp) {
3208 pmap = vmspace_pmap(newvm);
3209 ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
3210 if (pmap->pm_active_lock & CPULOCK_EXCL)
3211 pmap_interlock_wait(newvm);
3212 #if defined(SWTCH_OPTIM_STATS)
3215 pmap = vmspace_pmap(oldvm);
3216 ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
3224 * The swtch code tried to switch in a heavy weight process whos pmap
3225 * is locked by another cpu. We have to wait for the lock to clear before
3226 * the pmap can be used.
3229 pmap_interlock_wait (struct vmspace *vm)
3231 pmap_t pmap = vmspace_pmap(vm);
3233 if (pmap->pm_active_lock & CPULOCK_EXCL) {
3235 while (pmap->pm_active_lock & CPULOCK_EXCL) {
3244 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3247 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3251 addr = roundup2(addr, NBPDR);
3256 * Used by kmalloc/kfree, page already exists at va
3259 pmap_kvtom(vm_offset_t va)
3263 KKASSERT(va >= KvaStart && va < KvaEnd);
3265 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
3269 pmap_object_init(vm_object_t object)
3275 pmap_object_free(vm_object_t object)
3281 pmap_pgscan(struct pmap_pgscan_info *pginfo)
3283 pmap_t pmap = pginfo->pmap;
3284 vm_offset_t sva = pginfo->beg_addr;
3285 vm_offset_t eva = pginfo->end_addr;
3286 vm_offset_t va_next;
3287 pml4_entry_t *pml4e;
3289 pd_entry_t ptpaddr, *pde;
3294 vm_object_hold(pmap->pm_pteobj);
3296 for (; sva < eva; sva = va_next) {
3300 pml4e = pmap_pml4e(pmap, sva);
3301 if ((*pml4e & VPTE_V) == 0) {
3302 va_next = (sva + NBPML4) & ~PML4MASK;
3308 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
3309 if ((*pdpe & VPTE_V) == 0) {
3310 va_next = (sva + NBPDP) & ~PDPMASK;
3316 va_next = (sva + NBPDR) & ~PDRMASK;
3320 pde = pmap_pdpe_to_pde(pdpe, sva);
3325 * Check for large page (ignore).
3327 if ((ptpaddr & VPTE_PS) != 0) {
3329 pmap_clean_pde(pde, pmap, sva);
3330 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
3337 * Weed out invalid mappings. Note: we assume that the page
3338 * directory table is always allocated, and in kernel virtual.
3346 pt_m = pmap_hold_pt_page(pde, sva);
3347 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
3353 if ((*pte & VPTE_MANAGED) == 0)
3356 m = PHYS_TO_VM_PAGE(*pte & VPTE_FRAME);
3357 if (vm_page_busy_try(m, TRUE) == 0) {
3358 if (pginfo->callback(pginfo, sva, m) < 0)
3362 vm_page_unhold(pt_m);
3364 vm_object_drop(pmap->pm_pteobj);
3368 pmap_maybethreaded(pmap_t pmap)
3374 * Called while page is hard-busied to clear the PG_MAPPED and PG_WRITEABLE
3377 * vkernel code is using the old pmap style so the flags should already
3381 pmap_mapped_sync(vm_page_t m)