2 * Copyright (c) 1991 Regents of the University of California.
3 * Copyright (c) 1994 John S. Dyson
4 * Copyright (c) 1994 David Greenman
5 * Copyright (c) 2003 Peter Wemm
6 * Copyright (c) 2005-2008 Alan L. Cox <alc@cs.rice.edu>
7 * Copyright (c) 2008, 2009 The DragonFly Project.
8 * Copyright (c) 2008, 2009 Jordan Gordeev.
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 $
45 * $DragonFly: src/sys/platform/pc64/amd64/pmap.c,v 1.3 2008/08/29 17:07:10 dillon Exp $
49 * Manages physical address maps.
51 * In addition to hardware address maps, this
52 * module is called upon to provide software-use-only
53 * maps which may or may not be stored in the same
54 * form as hardware maps. These pseudo-maps are
55 * used to store intermediate results from copy
56 * operations to and from address spaces.
58 * Since the information managed by this module is
59 * also stored by the logical address mapping module,
60 * this module may throw away valid virtual-to-physical
61 * mappings at almost any time. However, invalidations
62 * of virtual-to-physical mappings must be done as
65 * In order to cope with hardware architectures which
66 * make virtual-to-physical map invalidates expensive,
67 * this module may delay invalidate or reduced protection
68 * operations until such time as they are actually
69 * necessary. This module is given full information as
70 * to which processors are currently using which maps,
71 * and to when physical maps must be made correct.
75 #include "opt_disable_pse.h"
78 #include "opt_msgbuf.h"
80 #include <sys/param.h>
81 #include <sys/systm.h>
82 #include <sys/kernel.h>
84 #include <sys/msgbuf.h>
85 #include <sys/vmmeter.h>
89 #include <vm/vm_param.h>
90 #include <sys/sysctl.h>
92 #include <vm/vm_kern.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_extern.h>
97 #include <vm/vm_pageout.h>
98 #include <vm/vm_pager.h>
99 #include <vm/vm_zone.h>
101 #include <sys/user.h>
102 #include <sys/thread2.h>
103 #include <sys/sysref2.h>
105 #include <machine/cputypes.h>
106 #include <machine/md_var.h>
107 #include <machine/specialreg.h>
108 #include <machine/smp.h>
109 #include <machine_base/apic/apicreg.h>
110 #include <machine/globaldata.h>
111 #include <machine/pmap.h>
112 #include <machine/pmap_inval.h>
116 #define PMAP_KEEP_PDIRS
117 #ifndef PMAP_SHPGPERPROC
118 #define PMAP_SHPGPERPROC 200
121 #if defined(DIAGNOSTIC)
122 #define PMAP_DIAGNOSTIC
127 #if !defined(PMAP_DIAGNOSTIC)
128 #define PMAP_INLINE __inline
144 * Get PDEs and PTEs for user/kernel address space
147 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
149 static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
150 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
152 #define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
153 #define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
154 #define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
155 #define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
156 #define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
160 * Given a map and a machine independent protection code,
161 * convert to a vax protection code.
163 #define pte_prot(m, p) \
164 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
165 static int protection_codes[8];
167 struct pmap kernel_pmap;
168 static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
170 vm_paddr_t avail_start; /* PA of first available physical page */
171 vm_paddr_t avail_end; /* PA of last available physical page */
172 vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
173 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
174 vm_offset_t KvaStart; /* VA start of KVA space */
175 vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
176 vm_offset_t KvaSize; /* max size of kernel virtual address space */
177 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
178 static int pgeflag; /* PG_G or-in */
179 static int pseflag; /* PG_PS or-in */
181 static vm_object_t kptobj;
184 static vm_paddr_t dmaplimit;
186 vm_offset_t kernel_vm_end;
188 static uint64_t KPDphys; /* phys addr of kernel level 2 */
189 uint64_t KPDPphys; /* phys addr of kernel level 3 */
190 uint64_t KPML4phys; /* phys addr of kernel level 4 */
192 static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
193 static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
196 * Data for the pv entry allocation mechanism
198 static vm_zone_t pvzone;
199 static struct vm_zone pvzone_store;
200 static struct vm_object pvzone_obj;
201 static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
202 static int pmap_pagedaemon_waken = 0;
203 static struct pv_entry *pvinit;
206 * All those kernel PT submaps that BSD is so fond of
208 pt_entry_t *CMAP1 = 0, *ptmmap;
209 caddr_t CADDR1 = 0, ptvmmap = 0;
210 static pt_entry_t *msgbufmap;
211 struct msgbuf *msgbufp=0;
216 static pt_entry_t *pt_crashdumpmap;
217 static caddr_t crashdumpmap;
219 extern uint64_t KPTphys;
220 extern pt_entry_t *SMPpt;
221 extern uint64_t SMPptpa;
225 static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
226 static pv_entry_t get_pv_entry (void);
227 static void i386_protection_init (void);
228 static __inline void pmap_clearbit (vm_page_t m, int bit);
230 static void pmap_remove_all (vm_page_t m);
231 static void pmap_enter_quick (pmap_t pmap, vm_offset_t va, vm_page_t m);
232 static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
233 vm_offset_t sva, pmap_inval_info_t info);
234 static void pmap_remove_page (struct pmap *pmap,
235 vm_offset_t va, pmap_inval_info_t info);
236 static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
237 vm_offset_t va, pmap_inval_info_t info);
238 static boolean_t pmap_testbit (vm_page_t m, int bit);
239 static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
240 vm_page_t mpte, vm_page_t m);
242 static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
244 static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
245 static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
246 static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
247 static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
248 static int pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
249 pmap_inval_info_t info);
250 static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
251 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
253 static unsigned pdir4mb;
256 * Move the kernel virtual free pointer to the next
257 * 4MB. This is used to help improve performance
258 * by using a large (4MB) page for much of the kernel
259 * (.text, .data, .bss)
262 pmap_kmem_choose(vm_offset_t addr)
265 vm_offset_t newaddr = addr;
267 if (cpu_feature & CPUID_PSE) {
268 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
277 * Super fast pmap_pte routine best used when scanning the pv lists.
278 * This eliminates many course-grained invltlb calls. Note that many of
279 * the pv list scans are across different pmaps and it is very wasteful
280 * to do an entire invltlb when checking a single mapping.
282 * Should only be called while in a critical section.
284 static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
287 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
290 return pmap_pte(pmap, va);
293 /* Return a non-clipped PD index for a given VA */
294 static __inline vm_pindex_t
295 pmap_pde_pindex(vm_offset_t va)
298 return va >> PDRSHIFT;
301 /* Return various clipped indexes for a given VA */
302 static __inline vm_pindex_t
303 pmap_pte_index(vm_offset_t va)
307 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
310 static __inline vm_pindex_t
311 pmap_pde_index(vm_offset_t va)
315 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
318 static __inline vm_pindex_t
319 pmap_pdpe_index(vm_offset_t va)
323 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
326 static __inline vm_pindex_t
327 pmap_pml4e_index(vm_offset_t va)
331 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
334 /* Return a pointer to the PML4 slot that corresponds to a VA */
335 static __inline pml4_entry_t *
336 pmap_pml4e(pmap_t pmap, vm_offset_t va)
340 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
343 /* Return a pointer to the PDP slot that corresponds to a VA */
344 static __inline pdp_entry_t *
345 pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
350 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
351 return (&pdpe[pmap_pdpe_index(va)]);
354 /* Return a pointer to the PDP slot that corresponds to a VA */
355 static __inline pdp_entry_t *
356 pmap_pdpe(pmap_t pmap, vm_offset_t va)
361 pml4e = pmap_pml4e(pmap, va);
362 if ((*pml4e & PG_V) == 0)
364 return (pmap_pml4e_to_pdpe(pml4e, va));
367 /* Return a pointer to the PD slot that corresponds to a VA */
368 static __inline pd_entry_t *
369 pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
374 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
375 return (&pde[pmap_pde_index(va)]);
378 /* Return a pointer to the PD slot that corresponds to a VA */
379 static __inline pd_entry_t *
380 pmap_pde(pmap_t pmap, vm_offset_t va)
385 pdpe = pmap_pdpe(pmap, va);
386 if (pdpe == NULL || (*pdpe & PG_V) == 0)
388 return (pmap_pdpe_to_pde(pdpe, va));
391 /* Return a pointer to the PT slot that corresponds to a VA */
392 static __inline pt_entry_t *
393 pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
398 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
399 return (&pte[pmap_pte_index(va)]);
402 /* Return a pointer to the PT slot that corresponds to a VA */
403 static __inline pt_entry_t *
404 pmap_pte(pmap_t pmap, vm_offset_t va)
409 pde = pmap_pde(pmap, va);
410 if (pde == NULL || (*pde & PG_V) == 0)
412 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
413 return ((pt_entry_t *)pde);
414 return (pmap_pde_to_pte(pde, va));
418 PMAP_INLINE pt_entry_t *
419 vtopte(vm_offset_t va)
422 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
424 return (PTmap + ((va >> PAGE_SHIFT) & mask));
427 static __inline pd_entry_t *
428 vtopde(vm_offset_t va)
431 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
433 return (PDmap + ((va >> PDRSHIFT) & mask));
437 allocpages(vm_paddr_t *firstaddr, int n)
443 bzero((void *)ret, n * PAGE_SIZE);
444 *firstaddr += n * PAGE_SIZE;
449 create_pagetables(vm_paddr_t *firstaddr)
454 uint64_t cpu0pp, cpu0idlestk;
455 int idlestk_page_offset = offsetof(struct privatespace, idlestack) / PAGE_SIZE;
457 /* we are running (mostly) V=P at this point */
460 KPTphys = allocpages(firstaddr, NKPT);
461 KPML4phys = allocpages(firstaddr, 1);
462 KPDPphys = allocpages(firstaddr, NKPML4E);
463 KPDphys = allocpages(firstaddr, NKPDPE);
465 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
466 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
468 DMPDPphys = allocpages(firstaddr, NDMPML4E);
469 if ((amd_feature & AMDID_PAGE1GB) == 0)
470 DMPDphys = allocpages(firstaddr, ndmpdp);
471 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
473 /* Fill in the underlying page table pages */
474 /* Read-only from zero to physfree */
475 /* XXX not fully used, underneath 2M pages */
476 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
477 ((pt_entry_t *)KPTphys)[i] = i << PAGE_SHIFT;
478 ((pt_entry_t *)KPTphys)[i] |= PG_RW | PG_V | PG_G;
481 /* Now map the page tables at their location within PTmap */
482 for (i = 0; i < NKPT; i++) {
483 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
484 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
487 /* Map from zero to end of allocations under 2M pages */
488 /* This replaces some of the KPTphys entries above */
489 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
490 ((pd_entry_t *)KPDphys)[i] = i << PDRSHIFT;
491 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V | PG_PS | PG_G;
494 /* And connect up the PD to the PDP */
495 for (i = 0; i < NKPDPE; i++) {
496 ((pdp_entry_t *)KPDPphys)[i + KPDPI] = KPDphys +
498 ((pdp_entry_t *)KPDPphys)[i + KPDPI] |= PG_RW | PG_V | PG_U;
501 /* Now set up the direct map space using either 2MB or 1GB pages */
502 /* Preset PG_M and PG_A because demotion expects it */
503 if ((amd_feature & AMDID_PAGE1GB) == 0) {
504 for (i = 0; i < NPDEPG * ndmpdp; i++) {
505 ((pd_entry_t *)DMPDphys)[i] = (vm_paddr_t)i << PDRSHIFT;
506 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
509 /* And the direct map space's PDP */
510 for (i = 0; i < ndmpdp; i++) {
511 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
513 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
516 for (i = 0; i < ndmpdp; i++) {
517 ((pdp_entry_t *)DMPDPphys)[i] =
518 (vm_paddr_t)i << PDPSHIFT;
519 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
524 /* And recursively map PML4 to itself in order to get PTmap */
525 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
526 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
528 /* Connect the Direct Map slot up to the PML4 */
529 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
530 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
532 /* Connect the KVA slot up to the PML4 */
533 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
534 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
536 common_lvl4_phys = allocpages(firstaddr, 1); /* 512 512G mappings */
537 common_lvl3_phys = allocpages(firstaddr, 1); /* 512 1G mappings */
538 KPTphys = allocpages(firstaddr, NKPT); /* kernel page table */
539 IdlePTD = allocpages(firstaddr, 1); /* kernel page dir */
540 cpu0pp = allocpages(firstaddr, MDGLOBALDATA_BASEALLOC_PAGES);
541 cpu0idlestk = allocpages(firstaddr, UPAGES);
542 SMPptpa = allocpages(firstaddr, 1);
543 SMPpt = (void *)(SMPptpa + KERNBASE);
547 * Load kernel page table with kernel memory mappings
549 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
550 ((pt_entry_t *)KPTphys)[i] = i << PAGE_SHIFT;
551 ((pt_entry_t *)KPTphys)[i] |= PG_RW | PG_V;
555 for (i = 0; i < NKPT; i++) {
556 ((pd_entry_t *)IdlePTD)[i] = KPTphys + (i << PAGE_SHIFT);
557 ((pd_entry_t *)IdlePTD)[i] |= PG_RW | PG_V;
562 * Set up the kernel page table itself.
564 for (i = 0; i < NKPT; i++) {
565 ((pd_entry_t *)IdlePTD)[KPTDI + i] = KPTphys + (i << PAGE_SHIFT);
566 ((pd_entry_t *)IdlePTD)[KPTDI + i] |= PG_RW | PG_V;
570 count = ISA_HOLE_LENGTH >> PAGE_SHIFT;
571 for (i = 0; i < count; i++) {
572 ((pt_entry_t *)KPTphys)[amd64_btop(ISA_HOLE_START) + i] = \
573 (ISA_HOLE_START + i * PAGE_SIZE) | PG_RW | PG_V;
580 ((pd_entry_t *)IdlePTD)[PTDPTDI] = (pd_entry_t)IdlePTD | PG_RW | PG_V;
583 * Map CPU_prvspace[0].mdglobaldata
585 for (i = 0; i < MDGLOBALDATA_BASEALLOC_PAGES; i++) {
586 ((pt_entry_t *)SMPptpa)[i] = \
587 (cpu0pp + i * PAGE_SIZE) | PG_RW | PG_V;
591 * Map CPU_prvspace[0].idlestack
593 for (i = 0; i < UPAGES; i++) {
594 ((pt_entry_t *)SMPptpa)[idlestk_page_offset + i] = \
595 (cpu0idlestk + i * PAGE_SIZE) | PG_RW | PG_V;
601 ((pd_entry_t *)IdlePTD)[MPPTDI] = SMPptpa | PG_RW | PG_V;
606 ((pml4_entry_t *)common_lvl4_phys)[LINKPML4I] = common_lvl3_phys | PG_RW | PG_V | PG_U;
609 * location of "virtual CR3" - a PDP entry that is loaded
610 * with a PD physical address (+ page attributes).
611 * Matt: location of user page directory entry (representing 1G)
613 link_pdpe = &((pdp_entry_t *)common_lvl3_phys)[LINKPDPI];
614 #endif /* JGPMAP32 */
619 init_paging(vm_paddr_t *firstaddr) {
620 create_pagetables(firstaddr);
623 /* switch to the newly created page table */
624 *link_pdpe = IdlePTD | PG_RW | PG_V | PG_U;
625 load_cr3(common_lvl4_phys);
626 link_pdpe = (void *)((char *)link_pdpe + KERNBASE);
628 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
629 KvaEnd = (vm_offset_t)VADDR(APTDPTDI, 0);
630 KvaSize = KvaEnd - KvaStart;
635 * Bootstrap the system enough to run with virtual memory.
637 * On the i386 this is called after mapping has already been enabled
638 * and just syncs the pmap module with what has already been done.
639 * [We can't call it easily with mapping off since the kernel is not
640 * mapped with PA == VA, hence we would have to relocate every address
641 * from the linked base (virtual) address "KERNBASE" to the actual
642 * (physical) address starting relative to 0]
645 pmap_bootstrap(vm_paddr_t *firstaddr)
650 struct mdglobaldata *gd;
654 KvaStart = VM_MIN_KERNEL_ADDRESS;
655 KvaEnd = VM_MAX_KERNEL_ADDRESS;
656 KvaSize = KvaEnd - KvaStart;
658 avail_start = *firstaddr;
661 * Create an initial set of page tables to run the kernel in.
663 create_pagetables(firstaddr);
665 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
666 virtual_start = pmap_kmem_choose(virtual_start);
668 virtual_end = VM_MAX_KERNEL_ADDRESS;
670 /* XXX do %cr0 as well */
671 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
675 * Initialize protection array.
677 i386_protection_init();
680 * The kernel's pmap is statically allocated so we don't have to use
681 * pmap_create, which is unlikely to work correctly at this part of
682 * the boot sequence (XXX and which no longer exists).
685 kernel_pmap.pm_pdir = (pd_entry_t *)(PTOV_OFFSET + (uint64_t)IdlePTD);
687 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
688 kernel_pmap.pm_count = 1;
689 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
690 TAILQ_INIT(&kernel_pmap.pm_pvlist);
694 * Reserve some special page table entries/VA space for temporary
697 #define SYSMAP(c, p, v, n) \
698 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
702 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
708 * CMAP1/CMAP2 are used for zeroing and copying pages.
710 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
715 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
718 * ptvmmap is used for reading arbitrary physical pages via
721 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
724 * msgbufp is used to map the system message buffer.
725 * XXX msgbufmap is not used.
727 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
728 atop(round_page(MSGBUF_SIZE)))
734 for (i = 0; i < NKPT; i++)
739 * PG_G is terribly broken on SMP because we IPI invltlb's in some
740 * cases rather then invl1pg. Actually, I don't even know why it
741 * works under UP because self-referential page table mappings
746 if (cpu_feature & CPUID_PGE)
751 * Initialize the 4MB page size flag
755 * The 4MB page version of the initial
756 * kernel page mapping.
760 #if !defined(DISABLE_PSE)
761 if (cpu_feature & CPUID_PSE) {
764 * Note that we have enabled PSE mode
767 ptditmp = *(PTmap + amd64_btop(KERNBASE));
768 ptditmp &= ~(NBPDR - 1);
769 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
774 * Enable the PSE mode. If we are SMP we can't do this
775 * now because the APs will not be able to use it when
778 load_cr4(rcr4() | CR4_PSE);
781 * We can do the mapping here for the single processor
782 * case. We simply ignore the old page table page from
786 * For SMP, we still need 4K pages to bootstrap APs,
787 * PSE will be enabled as soon as all APs are up.
789 PTD[KPTDI] = (pd_entry_t)ptditmp;
791 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
798 if (cpu_apic_address == 0)
799 panic("pmap_bootstrap: no local apic!");
801 /* local apic is mapped on last page */
802 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag |
803 (cpu_apic_address & PG_FRAME));
807 * We need to finish setting up the globaldata page for the BSP.
808 * locore has already populated the page table for the mdglobaldata
811 pg = MDGLOBALDATA_BASEALLOC_PAGES;
812 gd = &CPU_prvspace[0].mdglobaldata;
813 gd->gd_CMAP1 = &SMPpt[pg + 0];
814 gd->gd_CMAP2 = &SMPpt[pg + 1];
815 gd->gd_CMAP3 = &SMPpt[pg + 2];
816 gd->gd_PMAP1 = &SMPpt[pg + 3];
817 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
818 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
819 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
820 gd->gd_PADDR1 = (pt_entry_t *)CPU_prvspace[0].PPAGE1;
827 * Set 4mb pdir for mp startup
833 if (pseflag && (cpu_feature & CPUID_PSE)) {
834 load_cr4(rcr4() | CR4_PSE);
835 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
837 kernel_pmap.pm_pdir[KPTDI] =
838 PTD[KPTDI] = (pd_entry_t)pdir4mb;
847 * Initialize the pmap module.
848 * Called by vm_init, to initialize any structures that the pmap
849 * system needs to map virtual memory.
850 * pmap_init has been enhanced to support in a fairly consistant
851 * way, discontiguous physical memory.
861 * object for kernel page table pages
863 /* JG I think the number can be arbitrary */
864 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
867 * Allocate memory for random pmap data structures. Includes the
871 for(i = 0; i < vm_page_array_size; i++) {
874 m = &vm_page_array[i];
875 TAILQ_INIT(&m->md.pv_list);
876 m->md.pv_list_count = 0;
880 * init the pv free list
882 initial_pvs = vm_page_array_size;
883 if (initial_pvs < MINPV)
885 pvzone = &pvzone_store;
886 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
887 initial_pvs * sizeof (struct pv_entry));
888 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
892 * Now it is safe to enable pv_table recording.
894 pmap_initialized = TRUE;
898 * Initialize the address space (zone) for the pv_entries. Set a
899 * high water mark so that the system can recover from excessive
900 * numbers of pv entries.
906 int shpgperproc = PMAP_SHPGPERPROC;
908 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
909 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
910 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
911 pv_entry_high_water = 9 * (pv_entry_max / 10);
912 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
916 /***************************************************
917 * Low level helper routines.....
918 ***************************************************/
920 #if defined(PMAP_DIAGNOSTIC)
923 * This code checks for non-writeable/modified pages.
924 * This should be an invalid condition.
927 pmap_nw_modified(pt_entry_t pte)
930 if ((pte & (PG_M|PG_RW)) == PG_M)
939 * this routine defines the region(s) of memory that should
940 * not be tested for the modified bit.
942 static PMAP_INLINE int
943 pmap_track_modified(vm_offset_t va)
946 if ((va < clean_sva) || (va >= clean_eva))
955 * Extract the physical page address associated with the map/VA pair.
957 * This function may not be called from an interrupt if the pmap is
961 pmap_extract(pmap_t pmap, vm_offset_t va)
966 pd_entry_t pde, *pdep;
969 pdep = pmap_pde(pmap, va);
973 if ((pde & PG_PS) != 0) {
974 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
976 pte = pmap_pde_to_pte(pdep, va);
977 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
985 * Routine: pmap_kextract
987 * Extract the physical page address associated
988 * kernel virtual address.
991 pmap_kextract(vm_offset_t va)
997 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
998 pa = DMAP_TO_PHYS(va);
1002 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
1005 * Beware of a concurrent promotion that changes the
1006 * PDE at this point! For example, vtopte() must not
1007 * be used to access the PTE because it would use the
1008 * new PDE. It is, however, safe to use the old PDE
1009 * because the page table page is preserved by the
1012 pa = *pmap_pde_to_pte(&pde, va);
1013 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1019 /***************************************************
1020 * Low level mapping routines.....
1021 ***************************************************/
1024 * Routine: pmap_kenter
1026 * Add a wired page to the KVA
1027 * NOTE! note that in order for the mapping to take effect -- you
1028 * should do an invltlb after doing the pmap_kenter().
1031 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1036 pmap_inval_info info;
1038 pmap_inval_init(&info);
1039 npte = pa | PG_RW | PG_V | pgeflag;
1041 pmap_inval_add(&info, &kernel_pmap, va);
1043 pmap_inval_flush(&info);
1047 * Routine: pmap_kenter_quick
1049 * Similar to pmap_kenter(), except we only invalidate the
1050 * mapping on the current CPU.
1053 pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1059 npte = pa | PG_RW | PG_V | pgeflag;
1062 cpu_invlpg((void *)va);
1066 pmap_kenter_sync(vm_offset_t va)
1069 pmap_inval_info info;
1071 pmap_inval_init(&info);
1072 pmap_inval_add(&info, &kernel_pmap, va);
1073 pmap_inval_flush(&info);
1077 pmap_kenter_sync_quick(vm_offset_t va)
1080 cpu_invlpg((void *)va);
1084 * remove a page from the kernel pagetables
1087 pmap_kremove(vm_offset_t va)
1091 pmap_inval_info info;
1093 pmap_inval_init(&info);
1095 pmap_inval_add(&info, &kernel_pmap, va);
1097 pmap_inval_flush(&info);
1101 pmap_kremove_quick(vm_offset_t va)
1107 cpu_invlpg((void *)va);
1111 * XXX these need to be recoded. They are not used in any critical path.
1114 pmap_kmodify_rw(vm_offset_t va)
1117 *vtopte(va) |= PG_RW;
1118 cpu_invlpg((void *)va);
1122 pmap_kmodify_nc(vm_offset_t va)
1125 *vtopte(va) |= PG_N;
1126 cpu_invlpg((void *)va);
1130 * Used to map a range of physical addresses into kernel
1131 * virtual address space.
1133 * For now, VM is already on, we only need to map the
1137 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
1141 * JG Are callers prepared to get an address in the DMAP,
1142 * instead of the passed-in virt?
1144 while (start < end) {
1145 pmap_kenter(virt, start);
1154 * Add a list of wired pages to the kva
1155 * this routine is only used for temporary
1156 * kernel mappings that do not need to have
1157 * page modification or references recorded.
1158 * Note that old mappings are simply written
1159 * over. The page *must* be wired.
1162 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
1167 end_va = va + count * PAGE_SIZE;
1169 while (va < end_va) {
1173 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1174 cpu_invlpg((void *)va);
1179 smp_invltlb(); /* XXX */
1184 pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
1188 cpumask_t cmask = mycpu->gd_cpumask;
1190 end_va = va + count * PAGE_SIZE;
1192 while (va < end_va) {
1197 * Install the new PTE. If the pte changed from the prior
1198 * mapping we must reset the cpu mask and invalidate the page.
1199 * If the pte is the same but we have not seen it on the
1200 * current cpu, invlpg the existing mapping. Otherwise the
1201 * entry is optimal and no invalidation is required.
1204 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
1205 if (*pte != pteval) {
1208 cpu_invlpg((void *)va);
1209 } else if ((*mask & cmask) == 0) {
1210 cpu_invlpg((void *)va);
1219 * this routine jerks page mappings from the
1220 * kernel -- it is meant only for temporary mappings.
1223 pmap_qremove(vm_offset_t va, int count)
1228 end_va = va + count * PAGE_SIZE;
1230 while (va < end_va) {
1235 cpu_invlpg((void *)va);
1244 * This routine works like vm_page_lookup() but also blocks as long as the
1245 * page is busy. This routine does not busy the page it returns.
1247 * Unless the caller is managing objects whos pages are in a known state,
1248 * the call should be made with a critical section held so the page's object
1249 * association remains valid on return.
1252 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
1258 m = vm_page_lookup(object, pindex);
1259 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
1265 * Create a new thread and optionally associate it with a (new) process.
1266 * NOTE! the new thread's cpu may not equal the current cpu.
1269 pmap_init_thread(thread_t td)
1272 /* enforce pcb placement */
1273 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1274 td->td_savefpu = &td->td_pcb->pcb_save;
1275 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on amd64? */
1279 * This routine directly affects the fork perf for a process.
1282 pmap_init_proc(struct proc *p)
1288 * Dispose the UPAGES for a process that has exited.
1289 * This routine directly impacts the exit perf of a process.
1292 pmap_dispose_proc(struct proc *p)
1295 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
1298 /***************************************************
1299 * Page table page management routines.....
1300 ***************************************************/
1303 * This routine unholds page table pages, and if the hold count
1304 * drops to zero, then it decrements the wire count.
1307 _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m, pmap_inval_info_t info)
1311 * Wait until we can busy the page ourselves. We cannot have
1312 * any active flushes if we block.
1314 if (m->flags & PG_BUSY) {
1315 pmap_inval_flush(info);
1316 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1319 KASSERT(m->queue == PQ_NONE,
1320 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1322 if (m->hold_count == 1) {
1324 * Unmap the page table page
1327 pmap_inval_add(info, pmap, -1);
1329 if (m->pindex >= (NUPDE + NUPDPE)) {
1332 pml4 = pmap_pml4e(pmap, va);
1334 } else if (m->pindex >= NUPDE) {
1337 pdp = pmap_pdpe(pmap, va);
1342 pd = pmap_pde(pmap, va);
1346 KKASSERT(pmap->pm_stats.resident_count > 0);
1347 --pmap->pm_stats.resident_count;
1349 if (pmap->pm_ptphint == m)
1350 pmap->pm_ptphint = NULL;
1353 if (m->pindex < NUPDE) {
1354 /* We just released a PT, unhold the matching PD */
1357 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
1358 pmap_unwire_pte_hold(pmap, va, pdpg, info);
1360 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
1361 /* We just released a PD, unhold the matching PDP */
1364 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
1365 pmap_unwire_pte_hold(pmap, va, pdppg, info);
1370 * This was our last hold, the page had better be unwired
1371 * after we decrement wire_count.
1373 * FUTURE NOTE: shared page directory page could result in
1374 * multiple wire counts.
1378 KKASSERT(m->wire_count == 0);
1379 --vmstats.v_wire_count;
1380 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1382 vm_page_free_zero(m);
1385 KKASSERT(m->hold_count > 1);
1391 static PMAP_INLINE int
1392 pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m, pmap_inval_info_t info)
1395 KKASSERT(m->hold_count > 0);
1396 if (m->hold_count > 1) {
1400 return _pmap_unwire_pte_hold(pmap, va, m, info);
1405 * After removing a page table entry, this routine is used to
1406 * conditionally free the page, and manage the hold/wire counts.
1409 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1410 pmap_inval_info_t info)
1413 /* JG Use FreeBSD/amd64 or FreeBSD/i386 ptepde approaches? */
1414 vm_pindex_t ptepindex;
1415 if (va >= VM_MAX_USER_ADDRESS)
1419 ptepindex = pmap_pde_pindex(va);
1421 if (pmap->pm_ptphint &&
1422 (pmap->pm_ptphint->pindex == ptepindex)) {
1423 mpte = pmap->pm_ptphint;
1426 pmap_inval_flush(info);
1427 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1428 pmap->pm_ptphint = mpte;
1434 return pmap_unwire_pte_hold(pmap, va, mpte, info);
1438 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1439 * it, and IdlePTD, represents the template used to update all other pmaps.
1441 * On architectures where the kernel pmap is not integrated into the user
1442 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1443 * kernel_pmap should be used to directly access the kernel_pmap.
1446 pmap_pinit0(struct pmap *pmap)
1451 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1452 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
1454 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
1456 pmap->pm_active = 0;
1457 pmap->pm_ptphint = NULL;
1458 TAILQ_INIT(&pmap->pm_pvlist);
1459 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1463 * Initialize a preallocated and zeroed pmap structure,
1464 * such as one in a vmspace structure.
1467 pmap_pinit(struct pmap *pmap)
1473 * No need to allocate page table space yet but we do need a valid
1474 * page directory table.
1476 if (pmap->pm_pml4 == NULL) {
1478 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
1482 * Allocate an object for the ptes
1484 if (pmap->pm_pteobj == NULL)
1485 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PML4PML4I + 1);
1488 * Allocate the page directory page, unless we already have
1489 * one cached. If we used the cached page the wire_count will
1490 * already be set appropriately.
1492 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1493 ptdpg = vm_page_grab(pmap->pm_pteobj, PML4PML4I,
1494 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1495 pmap->pm_pdirm = ptdpg;
1496 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1497 ptdpg->valid = VM_PAGE_BITS_ALL;
1498 ptdpg->wire_count = 1;
1499 ++vmstats.v_wire_count;
1500 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
1502 if ((ptdpg->flags & PG_ZERO) == 0)
1503 bzero(pmap->pm_pml4, PAGE_SIZE);
1505 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1506 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
1508 /* install self-referential address mapping entry */
1509 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1512 pmap->pm_active = 0;
1513 pmap->pm_ptphint = NULL;
1514 TAILQ_INIT(&pmap->pm_pvlist);
1515 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1516 pmap->pm_stats.resident_count = 1;
1520 * Clean up a pmap structure so it can be physically freed. This routine
1521 * is called by the vmspace dtor function. A great deal of pmap data is
1522 * left passively mapped to improve vmspace management so we have a bit
1523 * of cleanup work to do here.
1526 pmap_puninit(pmap_t pmap)
1531 KKASSERT(pmap->pm_active == 0);
1532 if ((p = pmap->pm_pdirm) != NULL) {
1533 KKASSERT(pmap->pm_pml4 != NULL);
1534 KKASSERT(pmap->pm_pml4 != (PTOV_OFFSET + KPML4phys));
1535 pmap_kremove((vm_offset_t)pmap->pm_pml4);
1537 vmstats.v_wire_count--;
1538 KKASSERT((p->flags & PG_BUSY) == 0);
1540 vm_page_free_zero(p);
1541 pmap->pm_pdirm = NULL;
1543 if (pmap->pm_pml4) {
1544 KKASSERT(pmap->pm_pml4 != (PTOV_OFFSET + KPML4phys));
1545 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1546 pmap->pm_pml4 = NULL;
1548 if (pmap->pm_pteobj) {
1549 vm_object_deallocate(pmap->pm_pteobj);
1550 pmap->pm_pteobj = NULL;
1555 * Wire in kernel global address entries. To avoid a race condition
1556 * between pmap initialization and pmap_growkernel, this procedure
1557 * adds the pmap to the master list (which growkernel scans to update),
1558 * then copies the template.
1561 pmap_pinit2(struct pmap *pmap)
1565 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1566 /* XXX copies current process, does not fill in MPPTDI */
1568 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1574 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1575 * 0 on failure (if the procedure had to sleep).
1577 * When asked to remove the page directory page itself, we actually just
1578 * leave it cached so we do not have to incur the SMP inval overhead of
1579 * removing the kernel mapping. pmap_puninit() will take care of it.
1582 pmap_release_free_page(struct pmap *pmap, vm_page_t p)
1585 pml4_entry_t *pml4 = pmap->pm_pml4;
1587 * This code optimizes the case of freeing non-busy
1588 * page-table pages. Those pages are zero now, and
1589 * might as well be placed directly into the zero queue.
1591 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1597 * Remove the page table page from the processes address space.
1599 /* JG XXX we need to turn 'pindex' into a page table level
1600 * (PML4, PDP, PD, PT) and index within the page table page
1605 KKASSERT(pmap->pm_stats.resident_count > 0);
1606 --pmap->pm_stats.resident_count;
1608 if (p->hold_count) {
1609 panic("pmap_release: freeing held page table page");
1611 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1612 pmap->pm_ptphint = NULL;
1615 vmstats.v_wire_count--;
1616 vm_page_free_zero(p);
1621 * this routine is called if the page table page is not
1625 _pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1628 vm_page_t m, pdppg, pdpg;
1631 * Find or fabricate a new pagetable page
1633 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1634 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1637 if ((m->flags & PG_ZERO) == 0) {
1638 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1641 KASSERT(m->queue == PQ_NONE,
1642 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1645 * Increment the hold count for the page we will be returning to
1651 * It is possible that someone else got in and mapped by the page
1652 * directory page while we were blocked, if so just unbusy and
1653 * return the held page.
1656 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1657 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1663 if (m->wire_count == 0)
1664 vmstats.v_wire_count++;
1669 * Map the pagetable page into the process address space, if
1670 * it isn't already there.
1673 ++pmap->pm_stats.resident_count;
1676 ptepa = VM_PAGE_TO_PHYS(m);
1677 pmap->pm_pdir[ptepindex] =
1678 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1680 if (ptepindex >= (NUPDE + NUPDPE)) {
1682 vm_pindex_t pml4index;
1684 /* Wire up a new PDPE page */
1685 pml4index = ptepindex - (NUPDE + NUPDPE);
1686 pml4 = &pmap->pm_pml4[pml4index];
1687 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1689 } else if (ptepindex >= NUPDE) {
1690 vm_pindex_t pml4index;
1691 vm_pindex_t pdpindex;
1695 /* Wire up a new PDE page */
1696 pdpindex = ptepindex - NUPDE;
1697 pml4index = pdpindex >> NPML4EPGSHIFT;
1699 pml4 = &pmap->pm_pml4[pml4index];
1700 if ((*pml4 & PG_V) == 0) {
1701 /* Have to allocate a new pdp, recurse */
1702 if (_pmap_allocpte(pmap, NUPDE + NUPDPE + pml4index)
1709 /* Add reference to pdp page */
1710 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1711 pdppg->wire_count++;
1713 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1715 /* Now find the pdp page */
1716 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1717 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1720 vm_pindex_t pml4index;
1721 vm_pindex_t pdpindex;
1726 /* Wire up a new PTE page */
1727 pdpindex = ptepindex >> NPDPEPGSHIFT;
1728 pml4index = pdpindex >> NPML4EPGSHIFT;
1730 /* First, find the pdp and check that its valid. */
1731 pml4 = &pmap->pm_pml4[pml4index];
1732 if ((*pml4 & PG_V) == 0) {
1733 /* Have to allocate a new pd, recurse */
1734 if (_pmap_allocpte(pmap, NUPDE + pdpindex)
1740 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1741 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1743 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1744 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1745 if ((*pdp & PG_V) == 0) {
1746 /* Have to allocate a new pd, recurse */
1747 if (_pmap_allocpte(pmap, NUPDE + pdpindex)
1754 /* Add reference to the pd page */
1755 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1759 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
1761 /* Now we know where the page directory page is */
1762 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
1763 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
1768 * Set the page table hint
1770 pmap->pm_ptphint = m;
1772 m->valid = VM_PAGE_BITS_ALL;
1773 vm_page_flag_clear(m, PG_ZERO);
1774 vm_page_flag_set(m, PG_MAPPED);
1781 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1784 vm_pindex_t ptepindex;
1789 * Calculate pagetable page index
1791 ptepindex = pmap_pde_pindex(va);
1794 * Get the page directory entry
1796 pd = pmap_pde(pmap, va);
1799 * This supports switching from a 2MB page to a
1802 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1810 * If the page table page is mapped, we just increment the
1811 * hold count, and activate it.
1813 if (pd != NULL && (*pd & PG_V) != 0) {
1814 /* YYY hint is used here on i386 */
1815 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1816 pmap->pm_ptphint = m;
1821 * Here if the pte page isn't mapped, or if it has been deallocated.
1823 return _pmap_allocpte(pmap, ptepindex);
1827 /***************************************************
1828 * Pmap allocation/deallocation routines.
1829 ***************************************************/
1832 * Release any resources held by the given physical map.
1833 * Called when a pmap initialized by pmap_pinit is being released.
1834 * Should only be called if the map contains no valid mappings.
1836 static int pmap_release_callback(struct vm_page *p, void *data);
1839 pmap_release(struct pmap *pmap)
1842 vm_object_t object = pmap->pm_pteobj;
1843 struct rb_vm_page_scan_info info;
1845 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1846 #if defined(DIAGNOSTIC)
1847 if (object->ref_count != 1)
1848 panic("pmap_release: pteobj reference count != 1");
1852 info.object = object;
1854 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1861 info.limit = object->generation;
1863 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1864 pmap_release_callback, &info);
1865 if (info.error == 0 && info.mpte) {
1866 if (!pmap_release_free_page(pmap, info.mpte))
1870 } while (info.error);
1874 pmap_release_callback(struct vm_page *p, void *data)
1877 struct rb_vm_page_scan_info *info = data;
1879 if (p->pindex == PML4PML4I) {
1883 if (!pmap_release_free_page(info->pmap, p)) {
1887 if (info->object->generation != info->limit) {
1895 * Grow the number of kernel page table entries, if needed.
1899 pmap_growkernel(vm_offset_t addr)
1904 vm_offset_t ptppaddr;
1906 pd_entry_t *pde, newpdir;
1910 if (kernel_vm_end == 0) {
1911 kernel_vm_end = KERNBASE;
1913 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
1914 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1916 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1917 kernel_vm_end = kernel_map.max_offset;
1922 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1923 if (addr - 1 >= kernel_map.max_offset)
1924 addr = kernel_map.max_offset;
1925 while (kernel_vm_end < addr) {
1926 pde = pmap_pde(&kernel_pmap, kernel_vm_end);
1928 /* We need a new PDP entry */
1929 nkpg = vm_page_alloc(kptobj, nkpt,
1930 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM
1931 | VM_ALLOC_INTERRUPT);
1933 panic("pmap_growkernel: no memory to grow kernel");
1934 if ((nkpg->flags & PG_ZERO) == 0)
1935 pmap_zero_page(nkpg);
1936 paddr = VM_PAGE_TO_PHYS(nkpg);
1937 newpdp = (pdp_entry_t)
1938 (paddr | PG_V | PG_RW | PG_A | PG_M);
1939 *pmap_pdpe(&kernel_pmap, kernel_vm_end) = newpdp;
1940 continue; /* try again */
1942 if ((*pde & PG_V) != 0) {
1943 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1944 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1945 kernel_vm_end = kernel_map.max_offset;
1952 * This index is bogus, but out of the way
1954 nkpg = vm_page_alloc(kptobj, nkpt,
1955 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1957 panic("pmap_growkernel: no memory to grow kernel");
1960 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1961 pmap_zero_page(ptppaddr);
1962 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1963 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1966 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1967 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1968 kernel_vm_end = kernel_map.max_offset;
1976 * Retire the given physical map from service.
1977 * Should only be called if the map contains
1978 * no valid mappings.
1981 pmap_destroy(pmap_t pmap)
1989 count = --pmap->pm_count;
1992 panic("destroying a pmap is not yet implemented");
1997 * Add a reference to the specified pmap.
2000 pmap_reference(pmap_t pmap)
2008 /***************************************************
2009 * page management routines.
2010 ***************************************************/
2013 * free the pv_entry back to the free list. This function may be
2014 * called from an interrupt.
2016 static PMAP_INLINE void
2017 free_pv_entry(pv_entry_t pv)
2021 KKASSERT(pv_entry_count >= 0);
2026 * get a new pv_entry, allocating a block from the system
2027 * when needed. This function may be called from an interrupt.
2034 if (pv_entry_high_water &&
2035 (pv_entry_count > pv_entry_high_water) &&
2036 (pmap_pagedaemon_waken == 0)) {
2037 pmap_pagedaemon_waken = 1;
2038 wakeup(&vm_pages_needed);
2040 return zalloc(pvzone);
2044 * This routine is very drastic, but can save the system
2053 static int warningdone=0;
2055 if (pmap_pagedaemon_waken == 0)
2058 if (warningdone < 5) {
2059 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
2063 for(i = 0; i < vm_page_array_size; i++) {
2064 m = &vm_page_array[i];
2065 if (m->wire_count || m->hold_count || m->busy ||
2066 (m->flags & PG_BUSY))
2070 pmap_pagedaemon_waken = 0;
2075 * If it is the first entry on the list, it is actually
2076 * in the header and we must copy the following entry up
2077 * to the header. Otherwise we must search the list for
2078 * the entry. In either case we free the now unused entry.
2081 pmap_remove_entry(struct pmap *pmap, vm_page_t m,
2082 vm_offset_t va, pmap_inval_info_t info)
2089 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
2090 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2091 if (pmap == pv->pv_pmap && va == pv->pv_va)
2095 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2096 if (va == pv->pv_va)
2102 /* JGXXX When can 'pv' be NULL? */
2104 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2105 m->md.pv_list_count--;
2106 KKASSERT(m->md.pv_list_count >= 0);
2107 if (TAILQ_EMPTY(&m->md.pv_list))
2108 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2109 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2110 ++pmap->pm_generation;
2111 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
2119 * Create a pv entry for page at pa for
2123 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2129 pv = get_pv_entry();
2134 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2135 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2136 m->md.pv_list_count++;
2142 * pmap_remove_pte: do the things to unmap a page in a process
2145 pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2146 pmap_inval_info_t info)
2152 pmap_inval_add(info, pmap, va);
2153 oldpte = pte_load_clear(ptq);
2155 pmap->pm_stats.wired_count -= 1;
2157 * Machines that don't support invlpg, also don't support
2158 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2162 cpu_invlpg((void *)va);
2163 KKASSERT(pmap->pm_stats.resident_count > 0);
2164 --pmap->pm_stats.resident_count;
2165 if (oldpte & PG_MANAGED) {
2166 m = PHYS_TO_VM_PAGE(oldpte);
2167 if (oldpte & PG_M) {
2168 #if defined(PMAP_DIAGNOSTIC)
2169 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2171 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2175 if (pmap_track_modified(va))
2179 vm_page_flag_set(m, PG_REFERENCED);
2180 return pmap_remove_entry(pmap, m, va, info);
2182 return pmap_unuse_pt(pmap, va, NULL, info);
2191 * Remove a single page from a process address space.
2193 * This function may not be called from an interrupt if the pmap is
2197 pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2202 pte = pmap_pte(pmap, va);
2205 if ((*pte & PG_V) == 0)
2207 pmap_remove_pte(pmap, pte, va, info);
2213 * Remove the given range of addresses from the specified map.
2215 * It is assumed that the start and end are properly
2216 * rounded to the page size.
2218 * This function may not be called from an interrupt if the pmap is
2222 pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2225 vm_offset_t va_next;
2226 pml4_entry_t *pml4e;
2228 pd_entry_t ptpaddr, *pde;
2230 struct pmap_inval_info info;
2235 if (pmap->pm_stats.resident_count == 0)
2238 pmap_inval_init(&info);
2241 * special handling of removing one page. a very
2242 * common operation and easy to short circuit some
2245 if (sva + PAGE_SIZE == eva) {
2246 pde = pmap_pde(pmap, sva);
2247 if (pde && (*pde & PG_PS) == 0) {
2248 pmap_remove_page(pmap, sva, &info);
2249 pmap_inval_flush(&info);
2254 for (; sva < eva; sva = va_next) {
2255 pml4e = pmap_pml4e(pmap, sva);
2256 if ((*pml4e & PG_V) == 0) {
2257 va_next = (sva + NBPML4) & ~PML4MASK;
2263 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2264 if ((*pdpe & PG_V) == 0) {
2265 va_next = (sva + NBPDP) & ~PDPMASK;
2272 * Calculate index for next page table.
2274 va_next = (sva + NBPDR) & ~PDRMASK;
2278 pde = pmap_pdpe_to_pde(pdpe, sva);
2282 * Weed out invalid mappings.
2288 * Check for large page.
2290 if ((ptpaddr & PG_PS) != 0) {
2291 /* JG FreeBSD has more complex treatment here */
2292 pmap_inval_add(&info, pmap, -1);
2294 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2299 * Limit our scan to either the end of the va represented
2300 * by the current page table page, or to the end of the
2301 * range being removed.
2307 * NOTE: pmap_remove_pte() can block.
2309 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2313 if (pmap_remove_pte(pmap, pte, sva, &info))
2317 pmap_inval_flush(&info);
2323 * Removes this physical page from all physical maps in which it resides.
2324 * Reflects back modify bits to the pager.
2326 * This routine may not be called from an interrupt.
2330 pmap_remove_all(vm_page_t m)
2333 struct pmap_inval_info info;
2334 pt_entry_t *pte, tpte;
2337 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2340 pmap_inval_init(&info);
2342 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2343 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2344 --pv->pv_pmap->pm_stats.resident_count;
2346 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2347 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2348 tpte = pte_load_clear(pte);
2351 pv->pv_pmap->pm_stats.wired_count--;
2354 vm_page_flag_set(m, PG_REFERENCED);
2357 * Update the vm_page_t clean and reference bits.
2360 #if defined(PMAP_DIAGNOSTIC)
2361 if (pmap_nw_modified(tpte)) {
2363 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2367 if (pmap_track_modified(pv->pv_va))
2370 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2371 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2372 ++pv->pv_pmap->pm_generation;
2373 m->md.pv_list_count--;
2374 KKASSERT(m->md.pv_list_count >= 0);
2375 if (TAILQ_EMPTY(&m->md.pv_list))
2376 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2377 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2381 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2382 pmap_inval_flush(&info);
2388 * Set the physical protection on the specified range of this map
2391 * This function may not be called from an interrupt if the map is
2392 * not the kernel_pmap.
2395 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2398 vm_offset_t va_next;
2399 pml4_entry_t *pml4e;
2401 pd_entry_t ptpaddr, *pde;
2403 pmap_inval_info info;
2405 /* JG review for NX */
2410 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2411 pmap_remove(pmap, sva, eva);
2415 if (prot & VM_PROT_WRITE)
2418 pmap_inval_init(&info);
2420 for (; sva < eva; sva = va_next) {
2422 pml4e = pmap_pml4e(pmap, sva);
2423 if ((*pml4e & PG_V) == 0) {
2424 va_next = (sva + NBPML4) & ~PML4MASK;
2430 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2431 if ((*pdpe & PG_V) == 0) {
2432 va_next = (sva + NBPDP) & ~PDPMASK;
2438 va_next = (sva + NBPDR) & ~PDRMASK;
2442 pde = pmap_pdpe_to_pde(pdpe, sva);
2446 * Check for large page.
2448 if ((ptpaddr & PG_PS) != 0) {
2449 pmap_inval_add(&info, pmap, -1);
2450 *pde &= ~(PG_M|PG_RW);
2451 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2456 * Weed out invalid mappings. Note: we assume that the page
2457 * directory table is always allocated, and in kernel virtual.
2465 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2467 pt_entry_t obits, pbits;
2471 * XXX non-optimal. Note also that there can be
2472 * no pmap_inval_flush() calls until after we modify
2473 * ptbase[sindex] (or otherwise we have to do another
2474 * pmap_inval_add() call).
2476 pmap_inval_add(&info, pmap, sva);
2477 obits = pbits = *pte;
2478 if ((pbits & PG_V) == 0)
2480 if (pbits & PG_MANAGED) {
2483 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
2484 vm_page_flag_set(m, PG_REFERENCED);
2488 if (pmap_track_modified(sva)) {
2490 KKASSERT(pbits == (pbits & PG_FRAME));
2491 m = PHYS_TO_VM_PAGE(pbits);
2500 if (pbits != obits) {
2505 pmap_inval_flush(&info);
2509 * Insert the given physical page (p) at
2510 * the specified virtual address (v) in the
2511 * target physical map with the protection requested.
2513 * If specified, the page will be wired down, meaning
2514 * that the related pte can not be reclaimed.
2516 * NB: This is the only routine which MAY NOT lazy-evaluate
2517 * or lose information. That is, this routine must actually
2518 * insert this page into the given map NOW.
2521 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2529 pt_entry_t origpte, newpte;
2531 pmap_inval_info info;
2536 va = trunc_page(va);
2537 #ifdef PMAP_DIAGNOSTIC
2539 panic("pmap_enter: toobig");
2540 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2541 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
2543 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2544 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
2546 db_print_backtrace();
2549 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2550 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
2552 db_print_backtrace();
2557 * In the case that a page table page is not
2558 * resident, we are creating it here.
2560 if (va < VM_MAX_USER_ADDRESS)
2561 mpte = pmap_allocpte(pmap, va);
2565 pmap_inval_init(&info);
2566 pde = pmap_pde(pmap, va);
2567 if (pde != NULL && (*pde & PG_V) != 0) {
2568 if ((*pde & PG_PS) != 0)
2569 panic("pmap_enter: attempted pmap_enter on 2MB page");
2570 pte = pmap_pde_to_pte(pde, va);
2572 panic("pmap_enter: invalid page directory va=%#lx", va);
2574 KKASSERT(pte != NULL);
2575 pa = VM_PAGE_TO_PHYS(m);
2576 KKASSERT(pa == (pa & PG_FRAME));
2578 opa = origpte & PG_FRAME;
2581 * Mapping has not changed, must be protection or wiring change.
2583 if (origpte && (opa == pa)) {
2585 * Wiring change, just update stats. We don't worry about
2586 * wiring PT pages as they remain resident as long as there
2587 * are valid mappings in them. Hence, if a user page is wired,
2588 * the PT page will be also.
2590 if (wired && ((origpte & PG_W) == 0))
2591 pmap->pm_stats.wired_count++;
2592 else if (!wired && (origpte & PG_W))
2593 pmap->pm_stats.wired_count--;
2595 #if defined(PMAP_DIAGNOSTIC)
2596 if (pmap_nw_modified(origpte)) {
2598 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
2604 * Remove the extra pte reference. Note that we cannot
2605 * optimize the RO->RW case because we have adjusted the
2606 * wiring count above and may need to adjust the wiring
2613 * We might be turning off write access to the page,
2614 * so we go ahead and sense modify status.
2616 if (origpte & PG_MANAGED) {
2617 if ((origpte & PG_M) && pmap_track_modified(va)) {
2619 om = PHYS_TO_VM_PAGE(opa);
2623 KKASSERT(m->flags & PG_MAPPED);
2628 * Mapping has changed, invalidate old range and fall through to
2629 * handle validating new mapping.
2633 err = pmap_remove_pte(pmap, pte, va, &info);
2635 panic("pmap_enter: pte vanished, va: 0x%lx", va);
2639 * Enter on the PV list if part of our managed memory. Note that we
2640 * raise IPL while manipulating pv_table since pmap_enter can be
2641 * called at interrupt time.
2643 if (pmap_initialized &&
2644 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2645 pmap_insert_entry(pmap, va, mpte, m);
2647 vm_page_flag_set(m, PG_MAPPED);
2651 * Increment counters
2653 ++pmap->pm_stats.resident_count;
2655 pmap->pm_stats.wired_count++;
2659 * Now validate mapping with desired protection/wiring.
2661 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
2665 if (va < VM_MAX_USER_ADDRESS)
2667 if (pmap == &kernel_pmap)
2671 * if the mapping or permission bits are different, we need
2672 * to update the pte.
2674 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2675 pmap_inval_add(&info, pmap, va);
2676 *pte = newpte | PG_A;
2678 vm_page_flag_set(m, PG_WRITEABLE);
2680 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2681 pmap_inval_flush(&info);
2685 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2686 * This code also assumes that the pmap has no pre-existing entry for this
2689 * This code currently may only be used on user pmaps, not kernel_pmap.
2692 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
2698 vm_pindex_t ptepindex;
2700 pmap_inval_info info;
2702 pmap_inval_init(&info);
2704 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2705 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
2707 db_print_backtrace();
2710 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2711 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
2713 db_print_backtrace();
2717 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2720 * Calculate the page table page (mpte), allocating it if necessary.
2722 * A held page table page (mpte), or NULL, is passed onto the
2723 * section following.
2725 if (va < VM_MAX_USER_ADDRESS) {
2727 * Calculate pagetable page index
2729 ptepindex = pmap_pde_pindex(va);
2733 * Get the page directory entry
2735 ptepa = pmap_pde(pmap, va);
2738 * If the page table page is mapped, we just increment
2739 * the hold count, and activate it.
2741 if (ptepa && (*ptepa & PG_V) != 0) {
2743 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2744 // if (pmap->pm_ptphint &&
2745 // (pmap->pm_ptphint->pindex == ptepindex)) {
2746 // mpte = pmap->pm_ptphint;
2748 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2749 pmap->pm_ptphint = mpte;
2754 mpte = _pmap_allocpte(pmap, ptepindex);
2756 } while (mpte == NULL);
2759 /* this code path is not yet used */
2763 * With a valid (and held) page directory page, we can just use
2764 * vtopte() to get to the pte. If the pte is already present
2765 * we do not disturb it.
2770 pmap_unwire_pte_hold(pmap, va, mpte, &info);
2771 pa = VM_PAGE_TO_PHYS(m);
2772 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2777 * Enter on the PV list if part of our managed memory
2779 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2780 pmap_insert_entry(pmap, va, mpte, m);
2781 vm_page_flag_set(m, PG_MAPPED);
2785 * Increment counters
2787 ++pmap->pm_stats.resident_count;
2789 pa = VM_PAGE_TO_PHYS(m);
2792 * Now validate mapping with RO protection
2794 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2795 *pte = pa | PG_V | PG_U;
2797 *pte = pa | PG_V | PG_U | PG_MANAGED;
2798 /* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2799 pmap_inval_flush(&info);
2803 * Make a temporary mapping for a physical address. This is only intended
2804 * to be used for panic dumps.
2806 /* JG Needed on amd64? */
2808 pmap_kenter_temporary(vm_paddr_t pa, int i)
2811 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2812 return ((void *)crashdumpmap);
2815 #define MAX_INIT_PT (96)
2818 * This routine preloads the ptes for a given object into the specified pmap.
2819 * This eliminates the blast of soft faults on process startup and
2820 * immediately after an mmap.
2822 static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2825 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2826 vm_object_t object, vm_pindex_t pindex,
2827 vm_size_t size, int limit)
2830 struct rb_vm_page_scan_info info;
2835 * We can't preinit if read access isn't set or there is no pmap
2838 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2842 * We can't preinit if the pmap is not the current pmap
2844 lp = curthread->td_lwp;
2845 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2848 psize = amd64_btop(size);
2850 if ((object->type != OBJT_VNODE) ||
2851 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2852 (object->resident_page_count > MAX_INIT_PT))) {
2856 if (psize + pindex > object->size) {
2857 if (object->size < pindex)
2859 psize = object->size - pindex;
2866 * Use a red-black scan to traverse the requested range and load
2867 * any valid pages found into the pmap.
2869 * We cannot safely scan the object's memq unless we are in a
2870 * critical section since interrupts can remove pages from objects.
2872 info.start_pindex = pindex;
2873 info.end_pindex = pindex + psize - 1;
2880 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2881 pmap_object_init_pt_callback, &info);
2887 pmap_object_init_pt_callback(vm_page_t p, void *data)
2890 struct rb_vm_page_scan_info *info = data;
2891 vm_pindex_t rel_index;
2893 * don't allow an madvise to blow away our really
2894 * free pages allocating pv entries.
2896 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2897 vmstats.v_free_count < vmstats.v_free_reserved) {
2900 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2901 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2902 if ((p->queue - p->pc) == PQ_CACHE)
2903 vm_page_deactivate(p);
2905 rel_index = p->pindex - info->start_pindex;
2906 pmap_enter_quick(info->pmap,
2907 info->addr + amd64_ptob(rel_index), p);
2914 * pmap_prefault provides a quick way of clustering pagefaults into a
2915 * processes address space. It is a "cousin" of pmap_object_init_pt,
2916 * except it runs at page fault time instead of mmap time.
2920 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2922 static int pmap_prefault_pageorder[] = {
2923 -PAGE_SIZE, PAGE_SIZE,
2924 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2925 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2926 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2930 pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2942 * We do not currently prefault mappings that use virtual page
2943 * tables. We do not prefault foreign pmaps.
2945 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2947 lp = curthread->td_lwp;
2948 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2951 object = entry->object.vm_object;
2953 starta = addra - PFBAK * PAGE_SIZE;
2954 if (starta < entry->start)
2955 starta = entry->start;
2956 else if (starta > addra)
2960 * critical section protection is required to maintain the
2961 * page/object association, interrupts can free pages and remove
2962 * them from their objects.
2965 for (i = 0; i < PAGEORDER_SIZE; i++) {
2966 vm_object_t lobject;
2969 addr = addra + pmap_prefault_pageorder[i];
2970 if (addr > addra + (PFFOR * PAGE_SIZE))
2973 if (addr < starta || addr >= entry->end)
2976 if ((*pmap_pde(pmap, addr)) == 0)
2983 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2986 for (m = vm_page_lookup(lobject, pindex);
2987 (!m && (lobject->type == OBJT_DEFAULT) &&
2988 (lobject->backing_object));
2989 lobject = lobject->backing_object
2991 if (lobject->backing_object_offset & PAGE_MASK)
2993 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2994 m = vm_page_lookup(lobject->backing_object, pindex);
2998 * give-up when a page is not in memory
3003 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
3005 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
3007 if ((m->queue - m->pc) == PQ_CACHE) {
3008 vm_page_deactivate(m);
3011 pmap_enter_quick(pmap, addr, m);
3019 * Routine: pmap_change_wiring
3020 * Function: Change the wiring attribute for a map/virtual-address
3022 * In/out conditions:
3023 * The mapping must already exist in the pmap.
3026 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
3034 pte = pmap_pte(pmap, va);
3036 if (wired && !pmap_pte_w(pte))
3037 pmap->pm_stats.wired_count++;
3038 else if (!wired && pmap_pte_w(pte))
3039 pmap->pm_stats.wired_count--;
3042 * Wiring is not a hardware characteristic so there is no need to
3043 * invalidate TLB. However, in an SMP environment we must use
3044 * a locked bus cycle to update the pte (if we are not using
3045 * the pmap_inval_*() API that is)... it's ok to do this for simple
3050 atomic_set_int(pte, PG_W);
3052 atomic_clear_int(pte, PG_W);
3055 atomic_set_int_nonlocked(pte, PG_W);
3057 atomic_clear_int_nonlocked(pte, PG_W);
3064 * Copy the range specified by src_addr/len
3065 * from the source map to the range dst_addr/len
3066 * in the destination map.
3068 * This routine is only advisory and need not do anything.
3071 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
3072 vm_size_t len, vm_offset_t src_addr)
3075 pmap_inval_info info;
3077 vm_offset_t end_addr = src_addr + len;
3079 pd_entry_t src_frame, dst_frame;
3082 if (dst_addr != src_addr)
3085 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
3086 * valid through blocking calls, and that's just not going to
3094 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3095 if (src_frame != (PTDpde & PG_FRAME)) {
3099 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
3100 if (dst_frame != (APTDpde & PG_FRAME)) {
3101 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
3102 /* The page directory is not shared between CPUs */
3106 pmap_inval_init(&info);
3107 pmap_inval_add(&info, dst_pmap, -1);
3108 pmap_inval_add(&info, src_pmap, -1);
3111 * critical section protection is required to maintain the page/object
3112 * association, interrupts can free pages and remove them from
3116 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
3117 pt_entry_t *src_pte, *dst_pte;
3118 vm_page_t dstmpte, srcmpte;
3119 vm_offset_t srcptepaddr;
3120 vm_pindex_t ptepindex;
3122 if (addr >= UPT_MIN_ADDRESS)
3123 panic("pmap_copy: invalid to pmap_copy page tables\n");
3126 * Don't let optional prefaulting of pages make us go
3127 * way below the low water mark of free pages or way
3128 * above high water mark of used pv entries.
3130 if (vmstats.v_free_count < vmstats.v_free_reserved ||
3131 pv_entry_count > pv_entry_high_water)
3134 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
3135 ptepindex = addr >> PDRSHIFT;
3138 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
3140 if (srcptepaddr == 0)
3143 if (srcptepaddr & PG_PS) {
3145 if (dst_pmap->pm_pdir[ptepindex] == 0) {
3146 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
3147 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
3153 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
3154 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
3155 (srcmpte->flags & PG_BUSY)) {
3159 if (pdnxt > end_addr)
3162 src_pte = vtopte(addr);
3164 dst_pte = avtopte(addr);
3166 while (addr < pdnxt) {
3171 * we only virtual copy managed pages
3173 if ((ptetemp & PG_MANAGED) != 0) {
3175 * We have to check after allocpte for the
3176 * pte still being around... allocpte can
3179 * pmap_allocpte() can block. If we lose
3180 * our page directory mappings we stop.
3182 dstmpte = pmap_allocpte(dst_pmap, addr);
3185 if (src_frame != (PTDpde & PG_FRAME) ||
3186 dst_frame != (APTDpde & PG_FRAME)
3188 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3189 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3191 } else if ((*dst_pte == 0) &&
3192 (ptetemp = *src_pte) != 0 &&
3193 (ptetemp & PG_MANAGED)) {
3195 * Clear the modified and
3196 * accessed (referenced) bits
3199 m = PHYS_TO_VM_PAGE(ptetemp);
3200 *dst_pte = ptetemp & ~(PG_M | PG_A);
3201 ++dst_pmap->pm_stats.resident_count;
3202 pmap_insert_entry(dst_pmap, addr,
3204 KKASSERT(m->flags & PG_MAPPED);
3206 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3207 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3211 if (dstmpte->hold_count >= srcmpte->hold_count)
3221 pmap_inval_flush(&info);
3227 * Zero the specified physical page.
3229 * This function may be called from an interrupt and no locking is
3233 pmap_zero_page(vm_paddr_t phys)
3236 vm_offset_t va = PHYS_TO_DMAP(phys);
3238 pagezero((void *)va);
3242 * pmap_page_assertzero:
3244 * Assert that a page is empty, panic if it isn't.
3247 pmap_page_assertzero(vm_paddr_t phys)
3250 struct mdglobaldata *gd = mdcpu;
3254 vm_offset_t virt = PHYS_TO_DMAP(phys);
3256 for (i = 0; i < PAGE_SIZE; i += sizeof(int)) {
3257 if (*(int *)((char *)virt + i) != 0) {
3258 panic("pmap_page_assertzero() @ %p not zero!\n",
3268 * Zero part of a physical page by mapping it into memory and clearing
3269 * its contents with bzero.
3271 * off and size may not cover an area beyond a single hardware page.
3274 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3277 struct mdglobaldata *gd = mdcpu;
3280 vm_offset_t virt = PHYS_TO_DMAP(phys);
3281 bzero((char *)virt + off, size);
3288 * Copy the physical page from the source PA to the target PA.
3289 * This function may be called from an interrupt. No locking
3293 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3296 vm_offset_t src_virt, dst_virt;
3299 src_virt = PHYS_TO_DMAP(src);
3300 dst_virt = PHYS_TO_DMAP(dst);
3301 bcopy(src_virt, dst_virt, PAGE_SIZE);
3306 * pmap_copy_page_frag:
3308 * Copy the physical page from the source PA to the target PA.
3309 * This function may be called from an interrupt. No locking
3313 pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3316 vm_offset_t src_virt, dst_virt;
3319 src_virt = PHYS_TO_DMAP(src);
3320 dst_virt = PHYS_TO_DMAP(dst);
3321 bcopy((char *)src_virt + (src & PAGE_MASK),
3322 (char *)dst_virt + (dst & PAGE_MASK),
3328 * Returns true if the pmap's pv is one of the first
3329 * 16 pvs linked to from this page. This count may
3330 * be changed upwards or downwards in the future; it
3331 * is only necessary that true be returned for a small
3332 * subset of pmaps for proper page aging.
3335 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3341 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3346 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3347 if (pv->pv_pmap == pmap) {
3360 * Remove all pages from specified address space
3361 * this aids process exit speeds. Also, this code
3362 * is special cased for current process only, but
3363 * can have the more generic (and slightly slower)
3364 * mode enabled. This is much faster than pmap_remove
3365 * in the case of running down an entire address space.
3368 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3372 pt_entry_t *pte, tpte;
3375 pmap_inval_info info;
3377 int save_generation;
3379 lp = curthread->td_lwp;
3380 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3385 pmap_inval_init(&info);
3387 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3388 if (pv->pv_va >= eva || pv->pv_va < sva) {
3389 npv = TAILQ_NEXT(pv, pv_plist);
3393 KKASSERT(pmap == pv->pv_pmap);
3396 pte = vtopte(pv->pv_va);
3398 pte = pmap_pte_quick(pmap, pv->pv_va);
3399 if (pmap->pm_active)
3400 pmap_inval_add(&info, pmap, pv->pv_va);
3403 * We cannot remove wired pages from a process' mapping
3407 npv = TAILQ_NEXT(pv, pv_plist);
3410 tpte = pte_load_clear(pte);
3412 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
3414 KASSERT(m < &vm_page_array[vm_page_array_size],
3415 ("pmap_remove_pages: bad tpte %lx", tpte));
3417 KKASSERT(pmap->pm_stats.resident_count > 0);
3418 --pmap->pm_stats.resident_count;
3421 * Update the vm_page_t clean and reference bits.
3427 npv = TAILQ_NEXT(pv, pv_plist);
3428 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3429 save_generation = ++pmap->pm_generation;
3431 m->md.pv_list_count--;
3432 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3433 if (TAILQ_EMPTY(&m->md.pv_list))
3434 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3436 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3440 * Restart the scan if we blocked during the unuse or free
3441 * calls and other removals were made.
3443 if (save_generation != pmap->pm_generation) {
3444 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3445 pv = TAILQ_FIRST(&pmap->pm_pvlist);
3448 pmap_inval_flush(&info);
3453 * pmap_testbit tests bits in pte's
3454 * note that the testbit/clearbit routines are inline,
3455 * and a lot of things compile-time evaluate.
3458 pmap_testbit(vm_page_t m, int bit)
3464 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3467 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3472 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3474 * if the bit being tested is the modified bit, then
3475 * mark clean_map and ptes as never
3478 if (bit & (PG_A|PG_M)) {
3479 if (!pmap_track_modified(pv->pv_va))
3483 #if defined(PMAP_DIAGNOSTIC)
3484 if (pv->pv_pmap == NULL) {
3485 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
3489 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3500 * this routine is used to modify bits in ptes
3502 static __inline void
3503 pmap_clearbit(vm_page_t m, int bit)
3506 struct pmap_inval_info info;
3511 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3514 pmap_inval_init(&info);
3518 * Loop over all current mappings setting/clearing as appropos If
3519 * setting RO do we need to clear the VAC?
3521 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3523 * don't write protect pager mappings
3526 if (!pmap_track_modified(pv->pv_va))
3530 #if defined(PMAP_DIAGNOSTIC)
3531 if (pv->pv_pmap == NULL) {
3532 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
3538 * Careful here. We can use a locked bus instruction to
3539 * clear PG_A or PG_M safely but we need to synchronize
3540 * with the target cpus when we mess with PG_RW.
3542 * We do not have to force synchronization when clearing
3543 * PG_M even for PTEs generated via virtual memory maps,
3544 * because the virtual kernel will invalidate the pmap
3545 * entry when/if it needs to resynchronize the Modify bit.
3548 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
3549 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3556 atomic_clear_long(pte, PG_M|PG_RW);
3559 * The cpu may be trying to set PG_M
3560 * simultaniously with our clearing
3563 if (!atomic_cmpset_long(pte, pbits,
3567 } else if (bit == PG_M) {
3569 * We could also clear PG_RW here to force
3570 * a fault on write to redetect PG_M for
3571 * virtual kernels, but it isn't necessary
3572 * since virtual kernels invalidate the pte
3573 * when they clear the VPTE_M bit in their
3574 * virtual page tables.
3576 atomic_clear_long(pte, PG_M);
3578 atomic_clear_long(pte, bit);
3582 pmap_inval_flush(&info);
3587 * pmap_page_protect:
3589 * Lower the permission for all mappings to a given page.
3592 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3595 /* JG NX support? */
3596 if ((prot & VM_PROT_WRITE) == 0) {
3597 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3598 pmap_clearbit(m, PG_RW);
3599 vm_page_flag_clear(m, PG_WRITEABLE);
3607 pmap_phys_address(vm_pindex_t ppn)
3610 return (amd64_ptob(ppn));
3614 * pmap_ts_referenced:
3616 * Return a count of reference bits for a page, clearing those bits.
3617 * It is not necessary for every reference bit to be cleared, but it
3618 * is necessary that 0 only be returned when there are truly no
3619 * reference bits set.
3621 * XXX: The exact number of bits to check and clear is a matter that
3622 * should be tested and standardized at some point in the future for
3623 * optimal aging of shared pages.
3626 pmap_ts_referenced(vm_page_t m)
3629 pv_entry_t pv, pvf, pvn;
3633 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3638 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3643 pvn = TAILQ_NEXT(pv, pv_list);
3645 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3647 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3649 if (!pmap_track_modified(pv->pv_va))
3652 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3654 if (pte && (*pte & PG_A)) {
3656 atomic_clear_long(pte, PG_A);
3658 atomic_clear_long_nonlocked(pte, PG_A);
3665 } while ((pv = pvn) != NULL && pv != pvf);
3675 * Return whether or not the specified physical page was modified
3676 * in any physical maps.
3679 pmap_is_modified(vm_page_t m)
3682 return pmap_testbit(m, PG_M);
3686 * Clear the modify bits on the specified physical page.
3689 pmap_clear_modify(vm_page_t m)
3692 pmap_clearbit(m, PG_M);
3696 * pmap_clear_reference:
3698 * Clear the reference bit on the specified physical page.
3701 pmap_clear_reference(vm_page_t m)
3704 pmap_clearbit(m, PG_A);
3708 * Miscellaneous support routines follow
3712 i386_protection_init(void)
3717 /* JG NX support may go here; No VM_PROT_EXECUTE ==> set NX bit */
3718 kp = protection_codes;
3719 for (prot = 0; prot < 8; prot++) {
3721 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3723 * Read access is also 0. There isn't any execute bit,
3724 * so just make it readable.
3726 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3727 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3728 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3731 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3732 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3733 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3734 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3742 * Map a set of physical memory pages into the kernel virtual
3743 * address space. Return a pointer to where it is mapped. This
3744 * routine is intended to be used for mapping device memory,
3747 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3751 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3754 vm_offset_t va, tmpva, offset;
3757 offset = pa & PAGE_MASK;
3758 size = roundup(offset + size, PAGE_SIZE);
3760 va = kmem_alloc_nofault(&kernel_map, size);
3762 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3764 pa = pa & ~PAGE_MASK;
3765 for (tmpva = va; size > 0;) {
3766 pte = vtopte(tmpva);
3767 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
3775 return ((void *)(va + offset));
3779 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3782 vm_offset_t base, offset;
3784 base = va & ~PAGE_MASK;
3785 offset = va & PAGE_MASK;
3786 size = roundup(offset + size, PAGE_SIZE);
3787 pmap_qremove(va, size >> PAGE_SHIFT);
3788 kmem_free(&kernel_map, base, size);
3792 * perform the pmap work for mincore
3795 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3798 pt_entry_t *ptep, pte;
3802 ptep = pmap_pte(pmap, addr);
3807 if ((pte = *ptep) != 0) {
3810 val = MINCORE_INCORE;
3811 if ((pte & PG_MANAGED) == 0)
3814 pa = pte & PG_FRAME;
3816 m = PHYS_TO_VM_PAGE(pa);
3822 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3824 * Modified by someone
3826 else if (m->dirty || pmap_is_modified(m))
3827 val |= MINCORE_MODIFIED_OTHER;
3832 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3835 * Referenced by someone
3837 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3838 val |= MINCORE_REFERENCED_OTHER;
3839 vm_page_flag_set(m, PG_REFERENCED);
3846 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3847 * vmspace will be ref'd and the old one will be deref'd.
3849 * The vmspace for all lwps associated with the process will be adjusted
3850 * and cr3 will be reloaded if any lwp is the current lwp.
3853 pmap_replacevm(struct proc *p, struct vmspace *newvm, int adjrefs)
3856 struct vmspace *oldvm;
3860 oldvm = p->p_vmspace;
3861 if (oldvm != newvm) {
3862 p->p_vmspace = newvm;
3863 KKASSERT(p->p_nthreads == 1);
3864 lp = RB_ROOT(&p->p_lwp_tree);
3865 pmap_setlwpvm(lp, newvm);
3867 sysref_get(&newvm->vm_sysref);
3868 sysref_put(&oldvm->vm_sysref);
3875 * Set the vmspace for a LWP. The vmspace is almost universally set the
3876 * same as the process vmspace, but virtual kernels need to swap out contexts
3877 * on a per-lwp basis.
3880 pmap_setlwpvm(struct lwp *lp, struct vmspace *newvm)
3883 struct vmspace *oldvm;
3887 oldvm = lp->lwp_vmspace;
3889 if (oldvm != newvm) {
3890 lp->lwp_vmspace = newvm;
3891 if (curthread->td_lwp == lp) {
3892 pmap = vmspace_pmap(newvm);
3894 atomic_set_int(&pmap->pm_active, 1 << mycpu->gd_cpuid);
3896 pmap->pm_active |= 1;
3898 #if defined(SWTCH_OPTIM_STATS)
3901 curthread->td_pcb->pcb_cr3 = vtophys(pmap->pm_pml4);
3902 load_cr3(curthread->td_pcb->pcb_cr3);
3903 pmap = vmspace_pmap(oldvm);
3905 atomic_clear_int(&pmap->pm_active,
3906 1 << mycpu->gd_cpuid);
3908 pmap->pm_active &= ~1;
3916 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3920 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3924 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3931 static void pads (pmap_t pm);
3932 void pmap_pvdump (vm_paddr_t pa);
3934 /* print address space of pmap*/
3943 if (pm == &kernel_pmap)
3946 for (i = 0; i < NPDEPG; i++) {
3948 if (pm->pm_pdir[i]) {
3949 for (j = 0; j < NPTEPG; j++) {
3950 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3951 if (pm == &kernel_pmap && va < KERNBASE)
3953 if (pm != &kernel_pmap && va > UPT_MAX_ADDRESS)
3955 ptep = pmap_pte_quick(pm, va);
3956 if (pmap_pte_v(ptep))
3957 kprintf("%lx:%lx ", va, *ptep);
3967 pmap_pvdump(vm_paddr_t pa)
3973 kprintf("pa %08llx", (long long)pa);
3974 m = PHYS_TO_VM_PAGE(pa);
3975 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3977 kprintf(" -> pmap %p, va %x, flags %x",
3978 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3980 kprintf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);