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;
139 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
141 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
143 static vm_object_t kptobj;
147 static uint64_t KPDphys; /* phys addr of kernel level 2 */
148 uint64_t KPDPphys; /* phys addr of kernel level 3 */
149 uint64_t KPML4phys; /* phys addr of kernel level 4 */
151 extern int vmm_enabled;
152 extern void *vkernel_stack;
155 * Data for the pv entry allocation mechanism
157 static vm_zone_t pvzone;
158 static struct vm_zone pvzone_store;
159 static struct vm_object pvzone_obj;
160 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
161 static int pmap_pagedaemon_waken = 0;
162 static struct pv_entry *pvinit;
165 * All those kernel PT submaps that BSD is so fond of
167 pt_entry_t *CMAP1 = NULL, *ptmmap;
168 caddr_t CADDR1 = NULL;
169 static pt_entry_t *msgbufmap;
173 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
174 static pv_entry_t get_pv_entry (void);
175 static void i386_protection_init (void);
176 static __inline void pmap_clearbit (vm_page_t m, int bit);
178 static void pmap_remove_all (vm_page_t m);
179 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
181 static void pmap_remove_page (struct pmap *pmap, vm_offset_t va);
182 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
184 static boolean_t pmap_testbit (vm_page_t m, int bit);
185 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
186 vm_page_t mpte, vm_page_t m);
188 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
190 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
191 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
193 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
195 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
196 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t);
201 * Super fast pmap_pte routine best used when scanning the pv lists.
202 * This eliminates many course-grained invltlb calls. Note that many of
203 * the pv list scans are across different pmaps and it is very wasteful
204 * to do an entire invltlb when checking a single mapping.
206 * Should only be called while in a critical section.
209 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
212 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
214 return pmap_pte(pmap, va);
218 /* Return a non-clipped PD index for a given VA */
219 static __inline vm_pindex_t
220 pmap_pde_pindex(vm_offset_t va)
222 return va >> PDRSHIFT;
225 /* Return various clipped indexes for a given VA */
226 static __inline vm_pindex_t
227 pmap_pte_index(vm_offset_t va)
229 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
232 static __inline vm_pindex_t
233 pmap_pde_index(vm_offset_t va)
235 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
238 static __inline vm_pindex_t
239 pmap_pdpe_index(vm_offset_t va)
241 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
244 static __inline vm_pindex_t
245 pmap_pml4e_index(vm_offset_t va)
247 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
250 /* Return a pointer to the PML4 slot that corresponds to a VA */
251 static __inline pml4_entry_t *
252 pmap_pml4e(pmap_t pmap, vm_offset_t va)
254 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
257 /* Return a pointer to the PDP slot that corresponds to a VA */
258 static __inline pdp_entry_t *
259 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
263 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & VPTE_FRAME);
264 return (&pdpe[pmap_pdpe_index(va)]);
267 /* Return a pointer to the PDP slot that corresponds to a VA */
268 static __inline pdp_entry_t *
269 pmap_pdpe(pmap_t pmap, vm_offset_t va)
273 pml4e = pmap_pml4e(pmap, va);
274 if ((*pml4e & VPTE_V) == 0)
276 return (pmap_pml4e_to_pdpe(pml4e, va));
279 /* Return a pointer to the PD slot that corresponds to a VA */
280 static __inline pd_entry_t *
281 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
285 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & VPTE_FRAME);
286 return (&pde[pmap_pde_index(va)]);
289 /* Return a pointer to the PD slot that corresponds to a VA */
290 static __inline pd_entry_t *
291 pmap_pde(pmap_t pmap, vm_offset_t va)
295 pdpe = pmap_pdpe(pmap, va);
296 if (pdpe == NULL || (*pdpe & VPTE_V) == 0)
298 return (pmap_pdpe_to_pde(pdpe, va));
301 /* Return a pointer to the PT slot that corresponds to a VA */
302 static __inline pt_entry_t *
303 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
307 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & VPTE_FRAME);
308 return (&pte[pmap_pte_index(va)]);
311 /* Return a pointer to the PT slot that corresponds to a VA */
312 static __inline pt_entry_t *
313 pmap_pte(pmap_t pmap, vm_offset_t va)
317 pde = pmap_pde(pmap, va);
318 if (pde == NULL || (*pde & VPTE_V) == 0)
320 if ((*pde & VPTE_PS) != 0) /* compat with i386 pmap_pte() */
321 return ((pt_entry_t *)pde);
322 return (pmap_pde_to_pte(pde, va));
327 PMAP_INLINE pt_entry_t *
328 vtopte(vm_offset_t va)
330 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
331 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
333 return (PTmap + ((va >> PAGE_SHIFT) & mask));
336 static __inline pd_entry_t *
337 vtopde(vm_offset_t va)
339 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
340 NPML4EPGSHIFT)) - 1);
342 return (PDmap + ((va >> PDRSHIFT) & mask));
345 static PMAP_INLINE pt_entry_t *
346 vtopte(vm_offset_t va)
349 x = pmap_pte(&kernel_pmap, va);
354 static __inline pd_entry_t *
355 vtopde(vm_offset_t va)
358 x = pmap_pde(&kernel_pmap, va);
365 allocpages(vm_paddr_t *firstaddr, int n)
371 bzero((void *)ret, n * PAGE_SIZE);
373 *firstaddr += n * PAGE_SIZE;
378 create_dmap_vmm(vm_paddr_t *firstaddr)
381 int pml4_stack_index;
388 uint64_t KPDP_DMAP_phys = allocpages(firstaddr, NDMPML4E);
389 uint64_t KPDP_VSTACK_phys = allocpages(firstaddr, 1);
390 uint64_t KPD_VSTACK_phys = allocpages(firstaddr, 1);
392 pml4_entry_t *KPML4virt = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
393 pdp_entry_t *KPDP_DMAP_virt = (pdp_entry_t *)PHYS_TO_DMAP(KPDP_DMAP_phys);
394 pdp_entry_t *KPDP_VSTACK_virt = (pdp_entry_t *)PHYS_TO_DMAP(KPDP_VSTACK_phys);
395 pd_entry_t *KPD_VSTACK_virt = (pd_entry_t *)PHYS_TO_DMAP(KPD_VSTACK_phys);
397 bzero(KPDP_DMAP_virt, NDMPML4E * PAGE_SIZE);
398 bzero(KPDP_VSTACK_virt, 1 * PAGE_SIZE);
399 bzero(KPD_VSTACK_virt, 1 * PAGE_SIZE);
401 do_cpuid(0x80000001, regs);
402 amd_feature = regs[3];
404 /* Build the mappings for the first 512GB */
405 if (amd_feature & AMDID_PAGE1GB) {
406 /* In pages of 1 GB, if supported */
407 for (i = 0; i < NPDPEPG; i++) {
408 KPDP_DMAP_virt[i] = ((uint64_t)i << PDPSHIFT);
409 KPDP_DMAP_virt[i] |= VPTE_RW | VPTE_V | VPTE_PS | VPTE_U;
412 /* In page of 2MB, otherwise */
413 for (i = 0; i < NPDPEPG; i++) {
414 uint64_t KPD_DMAP_phys = allocpages(firstaddr, 1);
415 pd_entry_t *KPD_DMAP_virt = (pd_entry_t *)PHYS_TO_DMAP(KPD_DMAP_phys);
417 bzero(KPD_DMAP_virt, PAGE_SIZE);
419 KPDP_DMAP_virt[i] = KPD_DMAP_phys;
420 KPDP_DMAP_virt[i] |= VPTE_RW | VPTE_V | VPTE_U;
422 /* For each PD, we have to allocate NPTEPG PT */
423 for (j = 0; j < NPTEPG; j++) {
424 KPD_DMAP_virt[j] = (i << PDPSHIFT) | (j << PDRSHIFT);
425 KPD_DMAP_virt[j] |= VPTE_RW | VPTE_V | VPTE_PS | VPTE_U;
430 /* DMAP for the first 512G */
431 KPML4virt[0] = KPDP_DMAP_phys;
432 KPML4virt[0] |= VPTE_RW | VPTE_V | VPTE_U;
434 /* create a 2 MB map of the new stack */
435 pml4_stack_index = (uint64_t)&stack_addr >> PML4SHIFT;
436 KPML4virt[pml4_stack_index] = KPDP_VSTACK_phys;
437 KPML4virt[pml4_stack_index] |= VPTE_RW | VPTE_V | VPTE_U;
439 pdp_stack_index = ((uint64_t)&stack_addr & PML4MASK) >> PDPSHIFT;
440 KPDP_VSTACK_virt[pdp_stack_index] = KPD_VSTACK_phys;
441 KPDP_VSTACK_virt[pdp_stack_index] |= VPTE_RW | VPTE_V | VPTE_U;
443 pd_stack_index = ((uint64_t)&stack_addr & PDPMASK) >> PDRSHIFT;
444 KPD_VSTACK_virt[pd_stack_index] = (uint64_t) vkernel_stack;
445 KPD_VSTACK_virt[pd_stack_index] |= VPTE_RW | VPTE_V | VPTE_U | VPTE_PS;
449 create_pagetables(vm_paddr_t *firstaddr, int64_t ptov_offset)
452 pml4_entry_t *KPML4virt;
453 pdp_entry_t *KPDPvirt;
456 int kpml4i = pmap_pml4e_index(ptov_offset);
457 int kpdpi = pmap_pdpe_index(ptov_offset);
458 int kpdi = pmap_pde_index(ptov_offset);
461 * Calculate NKPT - number of kernel page tables. We have to
462 * accomodoate prealloction of the vm_page_array, dump bitmap,
463 * MSGBUF_SIZE, and other stuff. Be generous.
465 * Maxmem is in pages.
467 nkpt = (Maxmem * (sizeof(struct vm_page) * 2) + MSGBUF_SIZE) / NBPDR;
471 KPML4phys = allocpages(firstaddr, 1);
472 KPDPphys = allocpages(firstaddr, NKPML4E);
473 KPDphys = allocpages(firstaddr, NKPDPE);
474 KPTphys = allocpages(firstaddr, nkpt);
476 KPML4virt = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
477 KPDPvirt = (pdp_entry_t *)PHYS_TO_DMAP(KPDPphys);
478 KPDvirt = (pd_entry_t *)PHYS_TO_DMAP(KPDphys);
479 KPTvirt = (pt_entry_t *)PHYS_TO_DMAP(KPTphys);
481 bzero(KPML4virt, 1 * PAGE_SIZE);
482 bzero(KPDPvirt, NKPML4E * PAGE_SIZE);
483 bzero(KPDvirt, NKPDPE * PAGE_SIZE);
484 bzero(KPTvirt, nkpt * PAGE_SIZE);
486 /* Now map the page tables at their location within PTmap */
487 for (i = 0; i < nkpt; i++) {
488 KPDvirt[i + kpdi] = KPTphys + (i << PAGE_SHIFT);
489 KPDvirt[i + kpdi] |= VPTE_RW | VPTE_V | VPTE_U;
492 /* And connect up the PD to the PDP */
493 for (i = 0; i < NKPDPE; i++) {
494 KPDPvirt[i + kpdpi] = KPDphys + (i << PAGE_SHIFT);
495 KPDPvirt[i + kpdpi] |= VPTE_RW | VPTE_V | VPTE_U;
498 /* And recursively map PML4 to itself in order to get PTmap */
499 KPML4virt[PML4PML4I] = KPML4phys;
500 KPML4virt[PML4PML4I] |= VPTE_RW | VPTE_V | VPTE_U;
502 /* Connect the KVA slot up to the PML4 */
503 KPML4virt[kpml4i] = KPDPphys;
504 KPML4virt[kpml4i] |= VPTE_RW | VPTE_V | VPTE_U;
508 * Typically used to initialize a fictitious page by vm/device_pager.c
511 pmap_page_init(struct vm_page *m)
514 TAILQ_INIT(&m->md.pv_list);
518 * Bootstrap the system enough to run with virtual memory.
520 * On the i386 this is called after mapping has already been enabled
521 * and just syncs the pmap module with what has already been done.
522 * [We can't call it easily with mapping off since the kernel is not
523 * mapped with PA == VA, hence we would have to relocate every address
524 * from the linked base (virtual) address "KERNBASE" to the actual
525 * (physical) address starting relative to 0]
528 pmap_bootstrap(vm_paddr_t *firstaddr, int64_t ptov_offset)
534 * Create an initial set of page tables to run the kernel in.
536 create_pagetables(firstaddr, ptov_offset);
538 /* Create the DMAP for the VMM */
540 create_dmap_vmm(firstaddr);
543 virtual_start = KvaStart;
544 virtual_end = KvaEnd;
547 * Initialize protection array.
549 i386_protection_init();
552 * The kernel's pmap is statically allocated so we don't have to use
553 * pmap_create, which is unlikely to work correctly at this part of
554 * the boot sequence (XXX and which no longer exists).
556 * The kernel_pmap's pm_pteobj is used only for locking and not
559 kernel_pmap.pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys);
560 kernel_pmap.pm_count = 1;
561 /* don't allow deactivation */
562 CPUMASK_ASSALLONES(kernel_pmap.pm_active);
563 kernel_pmap.pm_pteobj = NULL; /* see pmap_init */
564 TAILQ_INIT(&kernel_pmap.pm_pvlist);
565 TAILQ_INIT(&kernel_pmap.pm_pvlist_free);
566 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
567 spin_init(&kernel_pmap.pm_spin, "pmapbootstrap");
570 * Reserve some special page table entries/VA space for temporary
573 #define SYSMAP(c, p, v, n) \
574 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
577 pte = pmap_pte(&kernel_pmap, va);
579 * CMAP1/CMAP2 are used for zeroing and copying pages.
581 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
587 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
591 * ptvmmap is used for reading arbitrary physical pages via
594 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
597 * msgbufp is used to map the system message buffer.
598 * XXX msgbufmap is not used.
600 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
601 atop(round_page(MSGBUF_SIZE)))
606 /* Not ready to do an invltlb yet for VMM*/
613 * Initialize the pmap module.
614 * Called by vm_init, to initialize any structures that the pmap
615 * system needs to map virtual memory.
616 * pmap_init has been enhanced to support in a fairly consistant
617 * way, discontiguous physical memory.
626 * object for kernel page table pages
628 /* JG I think the number can be arbitrary */
629 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
630 kernel_pmap.pm_pteobj = kptobj;
633 * Allocate memory for random pmap data structures. Includes the
636 for(i = 0; i < vm_page_array_size; i++) {
639 m = &vm_page_array[i];
640 TAILQ_INIT(&m->md.pv_list);
641 m->md.pv_list_count = 0;
645 * init the pv free list
647 initial_pvs = vm_page_array_size;
648 if (initial_pvs < MINPV)
650 pvzone = &pvzone_store;
651 pvinit = (struct pv_entry *)
652 kmem_alloc(&kernel_map,
653 initial_pvs * sizeof (struct pv_entry),
655 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
659 * Now it is safe to enable pv_table recording.
661 pmap_initialized = TRUE;
665 * Initialize the address space (zone) for the pv_entries. Set a
666 * high water mark so that the system can recover from excessive
667 * numbers of pv entries.
672 int shpgperproc = PMAP_SHPGPERPROC;
674 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
675 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
676 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
677 pv_entry_high_water = 9 * (pv_entry_max / 10);
678 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT);
682 /***************************************************
683 * Low level helper routines.....
684 ***************************************************/
687 * The modification bit is not tracked for any pages in this range. XXX
688 * such pages in this maps should always use pmap_k*() functions and not
691 * XXX User and kernel address spaces are independant for virtual kernels,
692 * this function only applies to the kernel pmap.
695 pmap_track_modified(pmap_t pmap, vm_offset_t va)
697 if (pmap != &kernel_pmap)
699 if ((va < clean_sva) || (va >= clean_eva))
706 * Extract the physical page address associated with the map/VA pair.
711 pmap_extract(pmap_t pmap, vm_offset_t va)
715 pd_entry_t pde, *pdep;
717 lwkt_gettoken(&vm_token);
719 pdep = pmap_pde(pmap, va);
723 if ((pde & VPTE_PS) != 0) {
725 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
727 pte = pmap_pde_to_pte(pdep, va);
728 rtval = (*pte & VPTE_FRAME) | (va & PAGE_MASK);
732 lwkt_reltoken(&vm_token);
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 --pmap->pm_stats.resident_count;
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 * Wire in kernel global address entries. To avoid a race condition
1290 * between pmap initialization and pmap_growkernel, this procedure
1291 * adds the pmap to the master list (which growkernel scans to update),
1292 * then copies the template.
1294 * In a virtual kernel there are no kernel global address entries.
1299 pmap_pinit2(struct pmap *pmap)
1301 spin_lock(&pmap_spin);
1302 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1303 spin_unlock(&pmap_spin);
1307 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1308 * 0 on failure (if the procedure had to sleep).
1310 * When asked to remove the page directory page itself, we actually just
1311 * leave it cached so we do not have to incur the SMP inval overhead of
1312 * removing the kernel mapping. pmap_puninit() will take care of it.
1315 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1318 * This code optimizes the case of freeing non-busy
1319 * page-table pages. Those pages are zero now, and
1320 * might as well be placed directly into the zero queue.
1322 if (vm_page_busy_try(p, FALSE)) {
1323 vm_page_sleep_busy(p, FALSE, "pmaprl");
1328 * Remove the page table page from the processes address space.
1330 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1332 * We are the pml4 table itself.
1334 /* XXX anything to do here? */
1335 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1337 * We are a PDP page.
1338 * We look for the PML4 entry that points to us.
1340 vm_page_t m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1341 KKASSERT(m4 != NULL);
1342 pml4_entry_t *pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1343 int idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1344 KKASSERT(pml4[idx] != 0);
1347 /* JG What about wire_count? */
1348 } else if (p->pindex >= NUPDE) {
1351 * We look for the PDP entry that points to us.
1353 vm_page_t m3 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1354 KKASSERT(m3 != NULL);
1355 pdp_entry_t *pdp = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1356 int idx = (p->pindex - NUPDE) % NPDPEPG;
1357 KKASSERT(pdp[idx] != 0);
1360 /* JG What about wire_count? */
1362 /* We are a PT page.
1363 * We look for the PD entry that points to us.
1365 vm_page_t m2 = vm_page_lookup(pmap->pm_pteobj, NUPDE + p->pindex / NPDEPG);
1366 KKASSERT(m2 != NULL);
1367 pd_entry_t *pd = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1368 int idx = p->pindex % NPDEPG;
1371 /* JG What about wire_count? */
1373 KKASSERT(pmap->pm_stats.resident_count > 0);
1374 --pmap->pm_stats.resident_count;
1376 if (p->hold_count) {
1377 panic("pmap_release: freeing held pt page "
1378 "pmap=%p pg=%p dmap=%p pi=%ld {%ld,%ld,%ld}",
1379 pmap, p, (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(p)),
1380 p->pindex, NUPDE, NUPDPE, PML4PML4I);
1382 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1383 pmap->pm_ptphint = NULL;
1386 * We leave the top-level page table page cached, wired, and mapped in
1387 * the pmap until the dtor function (pmap_puninit()) gets called.
1388 * However, still clean it up.
1390 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1391 bzero(pmap->pm_pml4, PAGE_SIZE);
1396 atomic_add_int(&vmstats.v_wire_count, -1);
1397 /* JG eventually revert to using vm_page_free_zero() */
1404 * this routine is called if the page table page is not
1408 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1410 vm_page_t m, pdppg, pdpg;
1413 * Find or fabricate a new pagetable page. Handle allocation
1414 * races by checking m->valid.
1416 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1417 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1419 KASSERT(m->queue == PQ_NONE,
1420 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1423 * Increment the hold count for the page we will be returning to
1430 * Map the pagetable page into the process address space, if
1431 * it isn't already there.
1433 ++pmap->pm_stats.resident_count;
1435 if (ptepindex >= (NUPDE + NUPDPE)) {
1437 vm_pindex_t pml4index;
1439 /* Wire up a new PDP page */
1440 pml4index = ptepindex - (NUPDE + NUPDPE);
1441 pml4 = &pmap->pm_pml4[pml4index];
1442 *pml4 = VM_PAGE_TO_PHYS(m) |
1443 VPTE_RW | VPTE_V | VPTE_U |
1445 } else if (ptepindex >= NUPDE) {
1446 vm_pindex_t pml4index;
1447 vm_pindex_t pdpindex;
1451 /* Wire up a new PD page */
1452 pdpindex = ptepindex - NUPDE;
1453 pml4index = pdpindex >> NPML4EPGSHIFT;
1455 pml4 = &pmap->pm_pml4[pml4index];
1456 if ((*pml4 & VPTE_V) == 0) {
1457 /* Have to allocate a new PDP page, recurse */
1458 if (_pmap_allocpte(pmap, NUPDE + NUPDPE + pml4index)
1465 /* Add reference to the PDP page */
1466 pdppg = PHYS_TO_VM_PAGE(*pml4 & VPTE_FRAME);
1467 pdppg->hold_count++;
1469 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & VPTE_FRAME);
1471 /* Now find the pdp page */
1472 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1473 KKASSERT(*pdp == 0); /* JG DEBUG64 */
1474 *pdp = VM_PAGE_TO_PHYS(m) | VPTE_RW | VPTE_V | VPTE_U |
1477 vm_pindex_t pml4index;
1478 vm_pindex_t pdpindex;
1483 /* Wire up a new PT page */
1484 pdpindex = ptepindex >> NPDPEPGSHIFT;
1485 pml4index = pdpindex >> NPML4EPGSHIFT;
1487 /* First, find the pdp and check that its valid. */
1488 pml4 = &pmap->pm_pml4[pml4index];
1489 if ((*pml4 & VPTE_V) == 0) {
1490 /* We miss a PDP page. We ultimately need a PD page.
1491 * Recursively allocating a PD page will allocate
1492 * the missing PDP page and will also allocate
1493 * the PD page we need.
1495 /* Have to allocate a new PD page, recurse */
1496 if (_pmap_allocpte(pmap, NUPDE + pdpindex)
1502 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & VPTE_FRAME);
1503 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1505 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & VPTE_FRAME);
1506 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1507 if ((*pdp & VPTE_V) == 0) {
1508 /* Have to allocate a new PD page, recurse */
1509 if (_pmap_allocpte(pmap, NUPDE + pdpindex)
1516 /* Add reference to the PD page */
1517 pdpg = PHYS_TO_VM_PAGE(*pdp & VPTE_FRAME);
1521 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & VPTE_FRAME);
1523 /* Now we know where the page directory page is */
1524 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1525 KKASSERT(*pd == 0); /* JG DEBUG64 */
1526 *pd = VM_PAGE_TO_PHYS(m) | VPTE_RW | VPTE_V | VPTE_U |
1531 * Set the page table hint
1533 pmap->pm_ptphint = m;
1534 vm_page_flag_set(m, PG_MAPPED);
1541 * Determine the page table page required to access the VA in the pmap
1542 * and allocate it if necessary. Return a held vm_page_t for the page.
1544 * Only used with user pmaps.
1547 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1549 vm_pindex_t ptepindex;
1553 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1556 * Calculate pagetable page index
1558 ptepindex = pmap_pde_pindex(va);
1561 * Get the page directory entry
1563 pd = pmap_pde(pmap, va);
1566 * This supports switching from a 2MB page to a
1569 if (pd != NULL && (*pd & (VPTE_PS | VPTE_V)) == (VPTE_PS | VPTE_V)) {
1570 panic("no promotion/demotion yet");
1578 * If the page table page is mapped, we just increment the
1579 * hold count, and activate it.
1581 if (pd != NULL && (*pd & VPTE_V) != 0) {
1582 /* YYY hint is used here on i386 */
1583 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1584 pmap->pm_ptphint = m;
1590 * Here if the pte page isn't mapped, or if it has been deallocated.
1592 return _pmap_allocpte(pmap, ptepindex);
1596 /***************************************************
1597 * Pmap allocation/deallocation routines.
1598 ***************************************************/
1601 * Release any resources held by the given physical map.
1602 * Called when a pmap initialized by pmap_pinit is being released.
1603 * Should only be called if the map contains no valid mappings.
1605 * Caller must hold pmap->pm_token
1607 static int pmap_release_callback(struct vm_page *p, void *data);
1610 pmap_release(struct pmap *pmap)
1612 vm_object_t object = pmap->pm_pteobj;
1613 struct rb_vm_page_scan_info info;
1615 KKASSERT(pmap != &kernel_pmap);
1617 lwkt_gettoken(&vm_token);
1618 #if defined(DIAGNOSTIC)
1619 if (object->ref_count != 1)
1620 panic("pmap_release: pteobj reference count != 1");
1624 info.object = object;
1626 KASSERT(CPUMASK_TESTZERO(pmap->pm_active),
1627 ("pmap %p still active! %016jx",
1629 (uintmax_t)CPUMASK_LOWMASK(pmap->pm_active)));
1631 spin_lock(&pmap_spin);
1632 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1633 spin_unlock(&pmap_spin);
1635 vm_object_hold(object);
1639 info.limit = object->generation;
1641 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1642 pmap_release_callback, &info);
1643 if (info.error == 0 && info.mpte) {
1644 if (!pmap_release_free_page(pmap, info.mpte))
1647 } while (info.error);
1648 vm_object_drop(object);
1649 lwkt_reltoken(&vm_token);
1653 pmap_release_callback(struct vm_page *p, void *data)
1655 struct rb_vm_page_scan_info *info = data;
1657 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1661 if (!pmap_release_free_page(info->pmap, p)) {
1665 if (info->object->generation != info->limit) {
1673 * Grow the number of kernel page table entries, if needed.
1678 pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
1682 vm_offset_t ptppaddr;
1684 pd_entry_t *pde, newpdir;
1689 vm_object_hold(kptobj);
1690 if (kernel_vm_end == 0) {
1691 kernel_vm_end = KvaStart;
1693 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & VPTE_V) != 0) {
1694 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1696 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1697 kernel_vm_end = kernel_map.max_offset;
1702 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1703 if (addr - 1 >= kernel_map.max_offset)
1704 addr = kernel_map.max_offset;
1705 while (kernel_vm_end < addr) {
1706 pde = pmap_pde(&kernel_pmap, kernel_vm_end);
1708 /* We need a new PDP entry */
1709 nkpg = vm_page_alloc(kptobj, nkpt,
1710 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM
1711 | VM_ALLOC_INTERRUPT);
1713 panic("pmap_growkernel: no memory to "
1716 paddr = VM_PAGE_TO_PHYS(nkpg);
1717 pmap_zero_page(paddr);
1718 newpdp = (pdp_entry_t)(paddr |
1719 VPTE_V | VPTE_RW | VPTE_U |
1721 *pmap_pdpe(&kernel_pmap, kernel_vm_end) = newpdp;
1723 continue; /* try again */
1725 if ((*pde & VPTE_V) != 0) {
1726 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1727 ~(PAGE_SIZE * NPTEPG - 1);
1728 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1729 kernel_vm_end = kernel_map.max_offset;
1736 * This index is bogus, but out of the way
1738 nkpg = vm_page_alloc(kptobj, nkpt,
1741 VM_ALLOC_INTERRUPT);
1743 panic("pmap_growkernel: no memory to grow kernel");
1746 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1747 pmap_zero_page(ptppaddr);
1748 newpdir = (pd_entry_t)(ptppaddr |
1749 VPTE_V | VPTE_RW | VPTE_U |
1751 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1754 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1755 ~(PAGE_SIZE * NPTEPG - 1);
1756 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1757 kernel_vm_end = kernel_map.max_offset;
1761 vm_object_drop(kptobj);
1765 * Add a reference to the specified pmap.
1770 pmap_reference(pmap_t pmap)
1773 lwkt_gettoken(&vm_token);
1775 lwkt_reltoken(&vm_token);
1779 /************************************************************************
1780 * VMSPACE MANAGEMENT *
1781 ************************************************************************
1783 * The VMSPACE management we do in our virtual kernel must be reflected
1784 * in the real kernel. This is accomplished by making vmspace system
1785 * calls to the real kernel.
1788 cpu_vmspace_alloc(struct vmspace *vm)
1795 * If VMM enable, don't do nothing, we
1796 * are able to use real page tables
1801 #define USER_SIZE (VM_MAX_USER_ADDRESS - VM_MIN_USER_ADDRESS)
1803 if (vmspace_create(&vm->vm_pmap, 0, NULL) < 0)
1804 panic("vmspace_create() failed");
1806 rp = vmspace_mmap(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE,
1807 PROT_READ|PROT_WRITE,
1808 MAP_FILE|MAP_SHARED|MAP_VPAGETABLE|MAP_FIXED,
1810 if (rp == MAP_FAILED)
1811 panic("vmspace_mmap: failed");
1812 vmspace_mcontrol(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE,
1814 vpte = VM_PAGE_TO_PHYS(vmspace_pmap(vm)->pm_pdirm) | VPTE_RW | VPTE_V | VPTE_U;
1815 r = vmspace_mcontrol(&vm->vm_pmap, VM_MIN_USER_ADDRESS, USER_SIZE,
1818 panic("vmspace_mcontrol: failed");
1822 cpu_vmspace_free(struct vmspace *vm)
1825 * If VMM enable, don't do nothing, we
1826 * are able to use real page tables
1831 if (vmspace_destroy(&vm->vm_pmap) < 0)
1832 panic("vmspace_destroy() failed");
1835 /***************************************************
1836 * page management routines.
1837 ***************************************************/
1840 * free the pv_entry back to the free list. This function may be
1841 * called from an interrupt.
1843 static __inline void
1844 free_pv_entry(pv_entry_t pv)
1847 KKASSERT(pv_entry_count >= 0);
1852 * get a new pv_entry, allocating a block from the system
1853 * when needed. This function may be called from an interrupt.
1859 if (pv_entry_high_water &&
1860 (pv_entry_count > pv_entry_high_water) &&
1861 (pmap_pagedaemon_waken == 0)) {
1862 pmap_pagedaemon_waken = 1;
1863 wakeup(&vm_pages_needed);
1865 return zalloc(pvzone);
1869 * This routine is very drastic, but can save the system
1879 static int warningdone=0;
1881 if (pmap_pagedaemon_waken == 0)
1883 lwkt_gettoken(&vm_token);
1884 pmap_pagedaemon_waken = 0;
1886 if (warningdone < 5) {
1887 kprintf("pmap_collect: collecting pv entries -- "
1888 "suggest increasing PMAP_SHPGPERPROC\n");
1892 for (i = 0; i < vm_page_array_size; i++) {
1893 m = &vm_page_array[i];
1894 if (m->wire_count || m->hold_count)
1896 if (vm_page_busy_try(m, TRUE) == 0) {
1897 if (m->wire_count == 0 && m->hold_count == 0) {
1903 lwkt_reltoken(&vm_token);
1908 * If it is the first entry on the list, it is actually
1909 * in the header and we must copy the following entry up
1910 * to the header. Otherwise we must search the list for
1911 * the entry. In either case we free the now unused entry.
1913 * caller must hold vm_token.
1916 pmap_remove_entry(struct pmap *pmap, vm_page_t m, vm_offset_t va)
1921 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1922 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1923 if (pmap == pv->pv_pmap && va == pv->pv_va)
1927 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1928 if (va == pv->pv_va)
1934 * Note that pv_ptem is NULL if the page table page itself is not
1935 * managed, even if the page being removed IS managed.
1938 /* JGXXX When can 'pv' be NULL? */
1940 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1941 m->md.pv_list_count--;
1942 atomic_add_int(&m->object->agg_pv_list_count, -1);
1943 KKASSERT(m->md.pv_list_count >= 0);
1944 if (TAILQ_EMPTY(&m->md.pv_list))
1945 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1946 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1947 ++pmap->pm_generation;
1948 KKASSERT(pmap->pm_pteobj != NULL);
1949 vm_object_hold(pmap->pm_pteobj);
1950 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1951 vm_object_drop(pmap->pm_pteobj);
1958 * Create a pv entry for page at pa for (pmap, va). If the page table page
1959 * holding the VA is managed, mpte will be non-NULL.
1962 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1967 pv = get_pv_entry();
1972 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1973 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1974 m->md.pv_list_count++;
1975 atomic_add_int(&m->object->agg_pv_list_count, 1);
1981 * pmap_remove_pte: do the things to unmap a page in a process
1984 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va)
1989 oldpte = pmap_inval_loadandclear(ptq, pmap, va);
1990 if (oldpte & VPTE_WIRED)
1991 --pmap->pm_stats.wired_count;
1992 KKASSERT(pmap->pm_stats.wired_count >= 0);
1996 * Machines that don't support invlpg, also don't support
1997 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2001 cpu_invlpg((void *)va);
2003 KKASSERT(pmap->pm_stats.resident_count > 0);
2004 --pmap->pm_stats.resident_count;
2005 if (oldpte & VPTE_MANAGED) {
2006 m = PHYS_TO_VM_PAGE(oldpte);
2007 if (oldpte & VPTE_M) {
2008 #if defined(PMAP_DIAGNOSTIC)
2009 if (pmap_nw_modified(oldpte)) {
2010 kprintf("pmap_remove: modified page not "
2011 "writable: va: 0x%lx, pte: 0x%lx\n",
2015 if (pmap_track_modified(pmap, va))
2018 if (oldpte & VPTE_A)
2019 vm_page_flag_set(m, PG_REFERENCED);
2020 return pmap_remove_entry(pmap, m, va);
2022 return pmap_unuse_pt(pmap, va, NULL);
2031 * Remove a single page from a process address space.
2033 * This function may not be called from an interrupt if the pmap is
2037 pmap_remove_page(struct pmap *pmap, vm_offset_t va)
2041 pte = pmap_pte(pmap, va);
2044 if ((*pte & VPTE_V) == 0)
2046 pmap_remove_pte(pmap, pte, va);
2050 * Remove the given range of addresses from the specified map.
2052 * It is assumed that the start and end are properly rounded to
2055 * This function may not be called from an interrupt if the pmap is
2061 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2063 vm_offset_t va_next;
2064 pml4_entry_t *pml4e;
2066 pd_entry_t ptpaddr, *pde;
2072 vm_object_hold(pmap->pm_pteobj);
2073 lwkt_gettoken(&vm_token);
2074 KKASSERT(pmap->pm_stats.resident_count >= 0);
2075 if (pmap->pm_stats.resident_count == 0) {
2076 lwkt_reltoken(&vm_token);
2077 vm_object_drop(pmap->pm_pteobj);
2082 * special handling of removing one page. a very
2083 * common operation and easy to short circuit some
2086 if (sva + PAGE_SIZE == eva) {
2087 pde = pmap_pde(pmap, sva);
2088 if (pde && (*pde & VPTE_PS) == 0) {
2089 pmap_remove_page(pmap, sva);
2090 lwkt_reltoken(&vm_token);
2091 vm_object_drop(pmap->pm_pteobj);
2096 for (; sva < eva; sva = va_next) {
2097 pml4e = pmap_pml4e(pmap, sva);
2098 if ((*pml4e & VPTE_V) == 0) {
2099 va_next = (sva + NBPML4) & ~PML4MASK;
2105 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2106 if ((*pdpe & VPTE_V) == 0) {
2107 va_next = (sva + NBPDP) & ~PDPMASK;
2114 * Calculate index for next page table.
2116 va_next = (sva + NBPDR) & ~PDRMASK;
2120 pde = pmap_pdpe_to_pde(pdpe, sva);
2124 * Weed out invalid mappings.
2130 * Check for large page.
2132 if ((ptpaddr & VPTE_PS) != 0) {
2133 /* JG FreeBSD has more complex treatment here */
2134 KKASSERT(*pde != 0);
2135 pmap_inval_pde(pde, pmap, sva);
2136 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2141 * Limit our scan to either the end of the va represented
2142 * by the current page table page, or to the end of the
2143 * range being removed.
2149 * NOTE: pmap_remove_pte() can block.
2151 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2155 if (pmap_remove_pte(pmap, pte, sva))
2159 lwkt_reltoken(&vm_token);
2160 vm_object_drop(pmap->pm_pteobj);
2164 * Removes this physical page from all physical maps in which it resides.
2165 * Reflects back modify bits to the pager.
2167 * This routine may not be called from an interrupt.
2172 pmap_remove_all(vm_page_t m)
2174 pt_entry_t *pte, tpte;
2177 #if defined(PMAP_DIAGNOSTIC)
2179 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
2182 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
2183 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%08llx", (long long)VM_PAGE_TO_PHYS(m));
2187 lwkt_gettoken(&vm_token);
2188 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2189 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2190 --pv->pv_pmap->pm_stats.resident_count;
2192 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2193 KKASSERT(pte != NULL);
2195 tpte = pmap_inval_loadandclear(pte, pv->pv_pmap, pv->pv_va);
2196 if (tpte & VPTE_WIRED)
2197 pv->pv_pmap->pm_stats.wired_count--;
2198 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
2201 vm_page_flag_set(m, PG_REFERENCED);
2204 * Update the vm_page_t clean and reference bits.
2206 if (tpte & VPTE_M) {
2207 #if defined(PMAP_DIAGNOSTIC)
2208 if (pmap_nw_modified(tpte)) {
2210 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2214 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
2217 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2218 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2219 ++pv->pv_pmap->pm_generation;
2220 m->md.pv_list_count--;
2221 atomic_add_int(&m->object->agg_pv_list_count, -1);
2222 KKASSERT(m->md.pv_list_count >= 0);
2223 if (TAILQ_EMPTY(&m->md.pv_list))
2224 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2225 vm_object_hold(pv->pv_pmap->pm_pteobj);
2226 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
2227 vm_object_drop(pv->pv_pmap->pm_pteobj);
2230 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2231 lwkt_reltoken(&vm_token);
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 lwkt_gettoken(&vm_token);
2245 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2246 if (pv->pv_pmap != pmap)
2249 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2250 --pv->pv_pmap->pm_stats.resident_count;
2252 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
2253 KKASSERT(pte != NULL);
2255 tpte = pmap_inval_loadandclear(pte, pv->pv_pmap, pv->pv_va);
2256 if (tpte & VPTE_WIRED)
2257 pv->pv_pmap->pm_stats.wired_count--;
2258 KKASSERT(pv->pv_pmap->pm_stats.wired_count >= 0);
2261 vm_page_flag_set(m, PG_REFERENCED);
2264 * Update the vm_page_t clean and reference bits.
2266 if (tpte & VPTE_M) {
2267 if (pmap_track_modified(pv->pv_pmap, pv->pv_va))
2270 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2271 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2272 ++pv->pv_pmap->pm_generation;
2273 m->md.pv_list_count--;
2274 atomic_add_int(&m->object->agg_pv_list_count, -1);
2275 KKASSERT(m->md.pv_list_count >= 0);
2276 if (TAILQ_EMPTY(&m->md.pv_list))
2277 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2278 vm_object_hold(pv->pv_pmap->pm_pteobj);
2279 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
2280 vm_object_drop(pv->pv_pmap->pm_pteobj);
2284 lwkt_reltoken(&vm_token);
2288 * Set the physical protection on the specified range of this map
2291 * This function may not be called from an interrupt if the map is
2292 * not the kernel_pmap.
2297 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2299 vm_offset_t va_next;
2300 pml4_entry_t *pml4e;
2302 pd_entry_t ptpaddr, *pde;
2305 /* JG review for NX */
2310 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2311 pmap_remove(pmap, sva, eva);
2315 if (prot & VM_PROT_WRITE)
2318 lwkt_gettoken(&vm_token);
2320 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 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2356 * Weed out invalid mappings. Note: we assume that the page
2357 * directory table is always allocated, and in kernel virtual.
2365 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2371 * Clean managed pages and also check the accessed
2372 * bit. Just remove write perms for unmanaged
2373 * pages. Be careful of races, turning off write
2374 * access will force a fault rather then setting
2375 * the modified bit at an unexpected time.
2377 if (*pte & VPTE_MANAGED) {
2378 pbits = pmap_clean_pte(pte, pmap, sva);
2380 if (pbits & VPTE_A) {
2381 m = PHYS_TO_VM_PAGE(pbits & VPTE_FRAME);
2382 vm_page_flag_set(m, PG_REFERENCED);
2383 atomic_clear_long(pte, VPTE_A);
2385 if (pbits & VPTE_M) {
2386 if (pmap_track_modified(pmap, sva)) {
2388 m = PHYS_TO_VM_PAGE(pbits & VPTE_FRAME);
2393 pbits = pmap_setro_pte(pte, pmap, sva);
2397 lwkt_reltoken(&vm_token);
2401 * Enter a managed page into a pmap. If the page is not wired related pmap
2402 * data can be destroyed at any time for later demand-operation.
2404 * Insert the vm_page (m) at virtual address (v) in (pmap), with the
2405 * specified protection, and wire the mapping if requested.
2407 * NOTE: This routine may not lazy-evaluate or lose information. The
2408 * page must actually be inserted into the given map NOW.
2410 * NOTE: When entering a page at a KVA address, the pmap must be the
2416 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2417 boolean_t wired, vm_map_entry_t entry __unused)
2423 pt_entry_t origpte, newpte;
2429 va = trunc_page(va);
2431 vm_object_hold(pmap->pm_pteobj);
2432 lwkt_gettoken(&vm_token);
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 ++pmap->pm_stats.wired_count;
2477 else if (!wired && (origpte & VPTE_WIRED))
2478 --pmap->pm_stats.wired_count;
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 if (pmap_initialized &&
2522 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2523 pmap_insert_entry(pmap, va, mpte, m);
2525 vm_page_flag_set(m, PG_MAPPED);
2529 * Increment counters
2531 ++pmap->pm_stats.resident_count;
2533 pmap->pm_stats.wired_count++;
2537 * Now validate mapping with desired protection/wiring.
2539 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | VPTE_V | VPTE_U);
2542 newpte |= VPTE_WIRED;
2543 // if (pmap != &kernel_pmap)
2547 * If the mapping or permission bits are different from the
2548 * (now cleaned) original pte, an update is needed. We've
2549 * already downgraded or invalidated the page so all we have
2550 * to do now is update the bits.
2552 * XXX should we synchronize RO->RW changes to avoid another
2555 if ((origpte & ~(VPTE_RW|VPTE_M|VPTE_A)) != newpte) {
2556 *pte = newpte | VPTE_A;
2557 if (newpte & VPTE_RW)
2558 vm_page_flag_set(m, PG_WRITEABLE);
2560 KKASSERT((newpte & VPTE_MANAGED) == 0 || (m->flags & PG_MAPPED));
2561 lwkt_reltoken(&vm_token);
2562 vm_object_drop(pmap->pm_pteobj);
2566 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2568 * Currently this routine may only be used on user pmaps, not kernel_pmap.
2573 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2578 vm_pindex_t ptepindex;
2581 KKASSERT(pmap != &kernel_pmap);
2583 KKASSERT(va >= VM_MIN_USER_ADDRESS && va < VM_MAX_USER_ADDRESS);
2586 * Calculate pagetable page index
2588 ptepindex = pmap_pde_pindex(va);
2590 vm_object_hold(pmap->pm_pteobj);
2591 lwkt_gettoken(&vm_token);
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 lwkt_reltoken(&vm_token);
2633 vm_object_drop(pmap->pm_pteobj);
2638 * Enter on the PV list if part of our managed memory
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);
2646 * Increment counters
2648 ++pmap->pm_stats.resident_count;
2650 pa = VM_PAGE_TO_PHYS(m);
2653 * Now validate mapping with RO protection
2655 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2656 *pte = (vpte_t)pa | VPTE_V | VPTE_U;
2658 *pte = (vpte_t)pa | VPTE_V | VPTE_U | VPTE_MANAGED;
2659 /*pmap_inval_add(&info, pmap, va); shouldn't be needed 0->valid */
2660 /*pmap_inval_flush(&info); don't need for vkernel */
2661 lwkt_reltoken(&vm_token);
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 lwkt_gettoken(&vm_token);
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 lwkt_reltoken(&vm_token);
2813 * Change the wiring attribute for a map/virtual-address pair.
2815 * The mapping must already exist in the pmap.
2816 * No other requirements.
2819 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
2820 vm_map_entry_t entry __unused)
2827 lwkt_gettoken(&vm_token);
2828 pte = pmap_pte(pmap, va);
2830 if (wired && !pmap_pte_w(pte))
2831 pmap->pm_stats.wired_count++;
2832 else if (!wired && pmap_pte_w(pte))
2833 pmap->pm_stats.wired_count--;
2836 * Wiring is not a hardware characteristic so there is no need to
2837 * invalidate TLB. However, in an SMP environment we must use
2838 * a locked bus cycle to update the pte (if we are not using
2839 * the pmap_inval_*() API that is)... it's ok to do this for simple
2843 atomic_set_long(pte, VPTE_WIRED);
2845 atomic_clear_long(pte, VPTE_WIRED);
2846 lwkt_reltoken(&vm_token);
2850 * Copy the range specified by src_addr/len
2851 * from the source map to the range dst_addr/len
2852 * in the destination map.
2854 * This routine is only advisory and need not do anything.
2857 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2858 vm_size_t len, vm_offset_t src_addr)
2861 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2862 * valid through blocking calls, and that's just not going to
2873 * Zero the specified physical page.
2875 * This function may be called from an interrupt and no locking is
2879 pmap_zero_page(vm_paddr_t phys)
2881 vm_offset_t va = PHYS_TO_DMAP(phys);
2883 bzero((void *)va, PAGE_SIZE);
2889 * Zero part of a physical page by mapping it into memory and clearing
2890 * its contents with bzero.
2892 * off and size may not cover an area beyond a single hardware page.
2895 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2898 vm_offset_t virt = PHYS_TO_DMAP(phys);
2899 bzero((char *)virt + off, size);
2906 * Copy the physical page from the source PA to the target PA.
2907 * This function may be called from an interrupt. No locking
2911 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2913 vm_offset_t src_virt, dst_virt;
2916 src_virt = PHYS_TO_DMAP(src);
2917 dst_virt = PHYS_TO_DMAP(dst);
2918 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
2923 * pmap_copy_page_frag:
2925 * Copy the physical page from the source PA to the target PA.
2926 * This function may be called from an interrupt. No locking
2930 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2932 vm_offset_t src_virt, dst_virt;
2935 src_virt = PHYS_TO_DMAP(src);
2936 dst_virt = PHYS_TO_DMAP(dst);
2937 bcopy((char *)src_virt + (src & PAGE_MASK),
2938 (char *)dst_virt + (dst & PAGE_MASK),
2944 * Returns true if the pmap's pv is one of the first 16 pvs linked to
2945 * from this page. This count may be changed upwards or downwards
2946 * in the future; it is only necessary that true be returned for a small
2947 * subset of pmaps for proper page aging.
2949 * No other requirements.
2952 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2957 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2961 lwkt_gettoken(&vm_token);
2963 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2964 if (pv->pv_pmap == pmap) {
2965 lwkt_reltoken(&vm_token);
2973 lwkt_reltoken(&vm_token);
2979 * Remove all pages from specified address space this aids process
2980 * exit speeds. Also, this code is special cased for current
2981 * process only, but can have the more generic (and slightly slower)
2982 * mode enabled. This is much faster than pmap_remove in the case
2983 * of running down an entire address space.
2985 * No other requirements.
2988 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2990 pt_entry_t *pte, tpte;
2993 int save_generation;
2995 if (pmap->pm_pteobj)
2996 vm_object_hold(pmap->pm_pteobj);
2997 lwkt_gettoken(&vm_token);
2999 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3000 if (pv->pv_va >= eva || pv->pv_va < sva) {
3001 npv = TAILQ_NEXT(pv, pv_plist);
3005 KKASSERT(pmap == pv->pv_pmap);
3007 pte = pmap_pte(pmap, pv->pv_va);
3010 * We cannot remove wired pages from a process' mapping
3013 if (*pte & VPTE_WIRED) {
3014 npv = TAILQ_NEXT(pv, pv_plist);
3017 tpte = pmap_inval_loadandclear(pte, pmap, pv->pv_va);
3019 m = PHYS_TO_VM_PAGE(tpte & VPTE_FRAME);
3021 KASSERT(m < &vm_page_array[vm_page_array_size],
3022 ("pmap_remove_pages: bad tpte %lx", tpte));
3024 KKASSERT(pmap->pm_stats.resident_count > 0);
3025 --pmap->pm_stats.resident_count;
3028 * Update the vm_page_t clean and reference bits.
3030 if (tpte & VPTE_M) {
3034 npv = TAILQ_NEXT(pv, pv_plist);
3035 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3036 save_generation = ++pmap->pm_generation;
3038 m->md.pv_list_count--;
3039 atomic_add_int(&m->object->agg_pv_list_count, -1);
3040 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3041 if (TAILQ_EMPTY(&m->md.pv_list))
3042 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
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 lwkt_reltoken(&vm_token);
3057 if (pmap->pm_pteobj)
3058 vm_object_drop(pmap->pm_pteobj);
3062 * pmap_testbit tests bits in active mappings of a VM page.
3065 pmap_testbit(vm_page_t m, int bit)
3070 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3073 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3078 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3080 * if the bit being tested is the modified bit, then
3081 * mark clean_map and ptes as never
3084 if (bit & (VPTE_A|VPTE_M)) {
3085 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
3089 #if defined(PMAP_DIAGNOSTIC)
3090 if (pv->pv_pmap == NULL) {
3091 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3095 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
3106 * This routine is used to clear bits in ptes. Certain bits require special
3107 * handling, in particular (on virtual kernels) the VPTE_M (modify) bit.
3109 * This routine is only called with certain VPTE_* bit combinations.
3111 static __inline void
3112 pmap_clearbit(vm_page_t m, int bit)
3118 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3124 * Loop over all current mappings setting/clearing as appropos If
3125 * setting RO do we need to clear the VAC?
3127 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3129 * don't write protect pager mappings
3131 if (bit == VPTE_RW) {
3132 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
3136 #if defined(PMAP_DIAGNOSTIC)
3137 if (pv->pv_pmap == NULL) {
3138 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3144 * Careful here. We can use a locked bus instruction to
3145 * clear VPTE_A or VPTE_M safely but we need to synchronize
3146 * with the target cpus when we mess with VPTE_RW.
3148 * On virtual kernels we must force a new fault-on-write
3149 * in the real kernel if we clear the Modify bit ourselves,
3150 * otherwise the real kernel will not get a new fault and
3151 * will never set our Modify bit again.
3153 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
3155 if (bit == VPTE_RW) {
3157 * We must also clear VPTE_M when clearing
3160 pbits = pmap_clean_pte(pte, pv->pv_pmap,
3164 } else if (bit == VPTE_M) {
3166 * We do not have to make the page read-only
3167 * when clearing the Modify bit. The real
3168 * kernel will make the real PTE read-only
3169 * or otherwise detect the write and set
3170 * our VPTE_M again simply by us invalidating
3171 * the real kernel VA for the pmap (as we did
3172 * above). This allows the real kernel to
3173 * handle the write fault without forwarding
3176 atomic_clear_long(pte, VPTE_M);
3177 } else if ((bit & (VPTE_RW|VPTE_M)) == (VPTE_RW|VPTE_M)) {
3179 * We've been asked to clear W & M, I guess
3180 * the caller doesn't want us to update
3181 * the dirty status of the VM page.
3183 pmap_clean_pte(pte, pv->pv_pmap, pv->pv_va);
3186 * We've been asked to clear bits that do
3187 * not interact with hardware.
3189 atomic_clear_long(pte, bit);
3197 * Lower the permission for all mappings to a given page.
3199 * No other requirements.
3202 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3204 /* JG NX support? */
3205 if ((prot & VM_PROT_WRITE) == 0) {
3206 lwkt_gettoken(&vm_token);
3207 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3208 pmap_clearbit(m, VPTE_RW);
3209 vm_page_flag_clear(m, PG_WRITEABLE);
3213 lwkt_reltoken(&vm_token);
3218 pmap_phys_address(vm_pindex_t ppn)
3220 return (x86_64_ptob(ppn));
3224 * Return a count of reference bits for a page, clearing those bits.
3225 * It is not necessary for every reference bit to be cleared, but it
3226 * is necessary that 0 only be returned when there are truly no
3227 * reference bits set.
3229 * XXX: The exact number of bits to check and clear is a matter that
3230 * should be tested and standardized at some point in the future for
3231 * optimal aging of shared pages.
3233 * No other requirements.
3236 pmap_ts_referenced(vm_page_t m)
3238 pv_entry_t pv, pvf, pvn;
3242 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3246 lwkt_gettoken(&vm_token);
3248 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3253 pvn = TAILQ_NEXT(pv, pv_list);
3255 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3257 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3259 if (!pmap_track_modified(pv->pv_pmap, pv->pv_va))
3262 pte = pmap_pte(pv->pv_pmap, pv->pv_va);
3264 if (pte && (*pte & VPTE_A)) {
3265 atomic_clear_long(pte, VPTE_A);
3271 } while ((pv = pvn) != NULL && pv != pvf);
3273 lwkt_reltoken(&vm_token);
3280 * Return whether or not the specified physical page was modified
3281 * in any physical maps.
3283 * No other requirements.
3286 pmap_is_modified(vm_page_t m)
3290 lwkt_gettoken(&vm_token);
3291 res = pmap_testbit(m, VPTE_M);
3292 lwkt_reltoken(&vm_token);
3297 * Clear the modify bits on the specified physical page.
3299 * No other requirements.
3302 pmap_clear_modify(vm_page_t m)
3304 lwkt_gettoken(&vm_token);
3305 pmap_clearbit(m, VPTE_M);
3306 lwkt_reltoken(&vm_token);
3310 * Clear the reference bit on the specified physical page.
3312 * No other requirements.
3315 pmap_clear_reference(vm_page_t m)
3317 lwkt_gettoken(&vm_token);
3318 pmap_clearbit(m, VPTE_A);
3319 lwkt_reltoken(&vm_token);
3323 * Miscellaneous support routines follow
3327 i386_protection_init(void)
3331 kp = protection_codes;
3332 for (prot = 0; prot < 8; prot++) {
3333 if (prot & VM_PROT_READ)
3334 *kp |= 0; /* if it's VALID is readeable */
3335 if (prot & VM_PROT_WRITE)
3337 if (prot & VM_PROT_EXECUTE)
3338 *kp |= 0; /* if it's VALID is executable */
3344 * Sets the memory attribute for the specified page.
3347 pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma)
3349 /* This is a vkernel, do nothing */
3353 * Change the PAT attribute on an existing kernel memory map. Caller
3354 * must ensure that the virtual memory in question is not accessed
3355 * during the adjustment.
3358 pmap_change_attr(vm_offset_t va, vm_size_t count, int mode)
3360 /* This is a vkernel, do nothing */
3364 * Perform the pmap work for mincore
3366 * No other requirements.
3369 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3371 pt_entry_t *ptep, pte;
3375 lwkt_gettoken(&vm_token);
3376 ptep = pmap_pte(pmap, addr);
3378 if (ptep && (pte = *ptep) != 0) {
3381 val = MINCORE_INCORE;
3382 if ((pte & VPTE_MANAGED) == 0)
3385 pa = pte & VPTE_FRAME;
3387 m = PHYS_TO_VM_PAGE(pa);
3393 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3395 * Modified by someone
3397 else if (m->dirty || pmap_is_modified(m))
3398 val |= MINCORE_MODIFIED_OTHER;
3403 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3406 * Referenced by someone
3408 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3409 val |= MINCORE_REFERENCED_OTHER;
3410 vm_page_flag_set(m, PG_REFERENCED);
3414 lwkt_reltoken(&vm_token);
3419 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3420 * vmspace will be ref'd and the old one will be deref'd.
3422 * Caller must hold vmspace->vm_map.token for oldvm and newvm
3425 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3427 struct vmspace *oldvm;
3431 oldvm = p->p_vmspace;
3432 if (oldvm != newvm) {
3435 p->p_vmspace = newvm;
3436 KKASSERT(p->p_nthreads == 1);
3437 lp = RB_ROOT(&p->p_lwp_tree);
3438 pmap_setlwpvm(lp, newvm);
3446 * Set the vmspace for a LWP. The vmspace is almost universally set the
3447 * same as the process vmspace, but virtual kernels need to swap out contexts
3448 * on a per-lwp basis.
3451 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3453 struct vmspace *oldvm;
3456 oldvm = lp->lwp_vmspace;
3457 if (oldvm != newvm) {
3459 lp->lwp_vmspace = newvm;
3460 if (curthread->td_lwp == lp) {
3461 pmap = vmspace_pmap(newvm);
3462 ATOMIC_CPUMASK_ORBIT(pmap->pm_active, mycpu->gd_cpuid);
3463 if (pmap->pm_active_lock & CPULOCK_EXCL)
3464 pmap_interlock_wait(newvm);
3465 #if defined(SWTCH_OPTIM_STATS)
3468 pmap = vmspace_pmap(oldvm);
3469 ATOMIC_CPUMASK_NANDBIT(pmap->pm_active,
3477 * The swtch code tried to switch in a heavy weight process whos pmap
3478 * is locked by another cpu. We have to wait for the lock to clear before
3479 * the pmap can be used.
3482 pmap_interlock_wait (struct vmspace *vm)
3484 pmap_t pmap = vmspace_pmap(vm);
3486 if (pmap->pm_active_lock & CPULOCK_EXCL) {
3488 while (pmap->pm_active_lock & CPULOCK_EXCL) {
3497 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3500 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3504 addr = roundup2(addr, NBPDR);
3509 * Used by kmalloc/kfree, page already exists at va
3512 pmap_kvtom(vm_offset_t va)
3516 KKASSERT(va >= KvaStart && va < KvaEnd);
3518 return(PHYS_TO_VM_PAGE(*ptep & PG_FRAME));
3522 pmap_object_init(vm_object_t object)
3528 pmap_object_free(vm_object_t object)
3534 pmap_pgscan(struct pmap_pgscan_info *pginfo)
3536 pmap_t pmap = pginfo->pmap;
3537 vm_offset_t sva = pginfo->beg_addr;
3538 vm_offset_t eva = pginfo->end_addr;
3539 vm_offset_t va_next;
3540 pml4_entry_t *pml4e;
3542 pd_entry_t ptpaddr, *pde;
3546 lwkt_gettoken(&vm_token);
3548 for (; sva < eva; sva = va_next) {
3552 pml4e = pmap_pml4e(pmap, sva);
3553 if ((*pml4e & VPTE_V) == 0) {
3554 va_next = (sva + NBPML4) & ~PML4MASK;
3560 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
3561 if ((*pdpe & VPTE_V) == 0) {
3562 va_next = (sva + NBPDP) & ~PDPMASK;
3568 va_next = (sva + NBPDR) & ~PDRMASK;
3572 pde = pmap_pdpe_to_pde(pdpe, sva);
3576 * Check for large page (ignore).
3578 if ((ptpaddr & VPTE_PS) != 0) {
3580 pmap_clean_pde(pde, pmap, sva);
3581 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
3587 * Weed out invalid mappings. Note: we assume that the page
3588 * directory table is always allocated, and in kernel virtual.
3596 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
3602 if ((*pte & VPTE_MANAGED) == 0)
3605 m = PHYS_TO_VM_PAGE(*pte & VPTE_FRAME);
3606 if (vm_page_busy_try(m, TRUE) == 0) {
3607 if (pginfo->callback(pginfo, sva, m) < 0)
3612 lwkt_reltoken(&vm_token);