kernel - Fix improper vgone() in procfs and races in truss
[dragonfly.git] / sys / platform / pc64 / x86_64 / pmap.c
CommitLineData
d7f50089 1/*
d7f50089 2 * Copyright (c) 1991 Regents of the University of California.
d7f50089 3 * Copyright (c) 1994 John S. Dyson
d7f50089 4 * Copyright (c) 1994 David Greenman
48ffc236
JG
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.
d7f50089 9 * All rights reserved.
c8fe38ae
MD
10 *
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.
14 *
d7f50089
YY
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
17 * are met:
d7f50089
YY
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
c8fe38ae
MD
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.
30 *
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
d7f50089 41 * SUCH DAMAGE.
c8fe38ae
MD
42 *
43 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
d7f50089 44 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
d7f50089 45 */
c8fe38ae 46
d7f50089 47/*
c8fe38ae
MD
48 * Manages physical address maps.
49 *
50 * In addition to hardware address maps, this
51 * module is called upon to provide software-use-only
52 * maps which may or may not be stored in the same
53 * form as hardware maps. These pseudo-maps are
54 * used to store intermediate results from copy
55 * operations to and from address spaces.
56 *
57 * Since the information managed by this module is
58 * also stored by the logical address mapping module,
59 * this module may throw away valid virtual-to-physical
60 * mappings at almost any time. However, invalidations
61 * of virtual-to-physical mappings must be done as
62 * requested.
63 *
64 * In order to cope with hardware architectures which
65 * make virtual-to-physical map invalidates expensive,
66 * this module may delay invalidate or reduced protection
67 * operations until such time as they are actually
68 * necessary. This module is given full information as
69 * to which processors are currently using which maps,
70 * and to when physical maps must be made correct.
71 */
72
73#if JG
74#include "opt_disable_pse.h"
75#include "opt_pmap.h"
76#endif
77#include "opt_msgbuf.h"
d7f50089 78
c8fe38ae 79#include <sys/param.h>
d7f50089
YY
80#include <sys/systm.h>
81#include <sys/kernel.h>
d7f50089 82#include <sys/proc.h>
c8fe38ae
MD
83#include <sys/msgbuf.h>
84#include <sys/vmmeter.h>
85#include <sys/mman.h>
d7f50089 86
c8fe38ae
MD
87#include <vm/vm.h>
88#include <vm/vm_param.h>
89#include <sys/sysctl.h>
90#include <sys/lock.h>
d7f50089 91#include <vm/vm_kern.h>
c8fe38ae
MD
92#include <vm/vm_page.h>
93#include <vm/vm_map.h>
d7f50089 94#include <vm/vm_object.h>
c8fe38ae 95#include <vm/vm_extern.h>
d7f50089 96#include <vm/vm_pageout.h>
c8fe38ae
MD
97#include <vm/vm_pager.h>
98#include <vm/vm_zone.h>
99
100#include <sys/user.h>
101#include <sys/thread2.h>
102#include <sys/sysref2.h>
d7f50089 103
c8fe38ae 104#include <machine/cputypes.h>
d7f50089 105#include <machine/md_var.h>
c8fe38ae
MD
106#include <machine/specialreg.h>
107#include <machine/smp.h>
108#include <machine_base/apic/apicreg.h>
d7f50089 109#include <machine/globaldata.h>
c8fe38ae
MD
110#include <machine/pmap.h>
111#include <machine/pmap_inval.h>
112
48ffc236
JG
113#include <ddb/ddb.h>
114
c8fe38ae
MD
115#define PMAP_KEEP_PDIRS
116#ifndef PMAP_SHPGPERPROC
117#define PMAP_SHPGPERPROC 200
118#endif
119
120#if defined(DIAGNOSTIC)
121#define PMAP_DIAGNOSTIC
122#endif
123
124#define MINPV 2048
125
c8fe38ae
MD
126/*
127 * Get PDEs and PTEs for user/kernel address space
128 */
48ffc236 129static pd_entry_t *pmap_pde(pmap_t pmap, vm_offset_t va);
c8fe38ae
MD
130#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
131
132#define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
133#define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
134#define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
135#define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
136#define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
137
138
139/*
140 * Given a map and a machine independent protection code,
141 * convert to a vax protection code.
142 */
143#define pte_prot(m, p) \
144 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
145static int protection_codes[8];
d7f50089
YY
146
147struct pmap kernel_pmap;
c8fe38ae 148static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
d7f50089 149
c8fe38ae
MD
150vm_paddr_t avail_start; /* PA of first available physical page */
151vm_paddr_t avail_end; /* PA of last available physical page */
152vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
153vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
154vm_offset_t KvaStart; /* VA start of KVA space */
155vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
156vm_offset_t KvaSize; /* max size of kernel virtual address space */
157static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
158static int pgeflag; /* PG_G or-in */
159static int pseflag; /* PG_PS or-in */
d7f50089 160
c8fe38ae
MD
161static vm_object_t kptobj;
162
48ffc236
JG
163static int ndmpdp;
164static vm_paddr_t dmaplimit;
c8fe38ae
MD
165static int nkpt;
166vm_offset_t kernel_vm_end;
d7f50089 167
48ffc236
JG
168static uint64_t KPDphys; /* phys addr of kernel level 2 */
169uint64_t KPDPphys; /* phys addr of kernel level 3 */
170uint64_t KPML4phys; /* phys addr of kernel level 4 */
171
172static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
173static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
174
d7f50089 175/*
c8fe38ae 176 * Data for the pv entry allocation mechanism
d7f50089 177 */
c8fe38ae
MD
178static vm_zone_t pvzone;
179static struct vm_zone pvzone_store;
180static struct vm_object pvzone_obj;
181static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
182static int pmap_pagedaemon_waken = 0;
183static struct pv_entry *pvinit;
d7f50089
YY
184
185/*
c8fe38ae 186 * All those kernel PT submaps that BSD is so fond of
d7f50089 187 */
c8fe38ae
MD
188pt_entry_t *CMAP1 = 0, *ptmmap;
189caddr_t CADDR1 = 0, ptvmmap = 0;
190static pt_entry_t *msgbufmap;
191struct msgbuf *msgbufp=0;
d7f50089 192
c8fe38ae
MD
193/*
194 * Crashdump maps.
d7f50089 195 */
c8fe38ae
MD
196static pt_entry_t *pt_crashdumpmap;
197static caddr_t crashdumpmap;
198
199extern uint64_t KPTphys;
200extern pt_entry_t *SMPpt;
201extern uint64_t SMPptpa;
202
203#define DISABLE_PSE
204
c8fe38ae 205static pv_entry_t get_pv_entry (void);
bfc09ba0
MD
206static void i386_protection_init (void);
207static void create_pagetables(vm_paddr_t *firstaddr);
208static void pmap_remove_all (vm_page_t m);
209static void pmap_enter_quick (pmap_t pmap, vm_offset_t va, vm_page_t m);
210static int pmap_remove_pte (struct pmap *pmap, pt_entry_t *ptq,
c8fe38ae
MD
211 vm_offset_t sva, pmap_inval_info_t info);
212static void pmap_remove_page (struct pmap *pmap,
213 vm_offset_t va, pmap_inval_info_t info);
bfc09ba0 214static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
c8fe38ae
MD
215 vm_offset_t va, pmap_inval_info_t info);
216static boolean_t pmap_testbit (vm_page_t m, int bit);
217static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
bfc09ba0 218 vm_page_t mpte, vm_page_t m);
c8fe38ae
MD
219
220static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
221
222static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
223static vm_page_t _pmap_allocpte (pmap_t pmap, vm_pindex_t ptepindex);
224static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
225static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
bfc09ba0
MD
226static int _pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
227 pmap_inval_info_t info);
c8fe38ae
MD
228static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
229static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
230
231static unsigned pdir4mb;
d7f50089
YY
232
233/*
c8fe38ae 234 * Move the kernel virtual free pointer to the next
f9cc0f15
JG
235 * 2MB. This is used to help improve performance
236 * by using a large (2MB) page for much of the kernel
c8fe38ae 237 * (.text, .data, .bss)
d7f50089 238 */
bfc09ba0
MD
239static
240vm_offset_t
c8fe38ae 241pmap_kmem_choose(vm_offset_t addr)
d7f50089 242{
c8fe38ae 243 vm_offset_t newaddr = addr;
f9cc0f15
JG
244
245 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
c8fe38ae 246 return newaddr;
d7f50089
YY
247}
248
d7f50089 249/*
c8fe38ae 250 * pmap_pte_quick:
d7f50089 251 *
c8fe38ae
MD
252 * Super fast pmap_pte routine best used when scanning the pv lists.
253 * This eliminates many course-grained invltlb calls. Note that many of
254 * the pv list scans are across different pmaps and it is very wasteful
255 * to do an entire invltlb when checking a single mapping.
256 *
257 * Should only be called while in a critical section.
258 */
48ffc236
JG
259static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
260
bfc09ba0
MD
261static
262pt_entry_t *
c8fe38ae
MD
263pmap_pte_quick(pmap_t pmap, vm_offset_t va)
264{
48ffc236
JG
265 return pmap_pte(pmap, va);
266}
267
268/* Return a non-clipped PD index for a given VA */
bfc09ba0
MD
269static __inline
270vm_pindex_t
48ffc236 271pmap_pde_pindex(vm_offset_t va)
48ffc236
JG
272{
273 return va >> PDRSHIFT;
274}
275
276/* Return various clipped indexes for a given VA */
bfc09ba0
MD
277static __inline
278vm_pindex_t
48ffc236 279pmap_pte_index(vm_offset_t va)
48ffc236
JG
280{
281
282 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
283}
284
bfc09ba0
MD
285static __inline
286vm_pindex_t
48ffc236 287pmap_pde_index(vm_offset_t va)
48ffc236
JG
288{
289
290 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
291}
292
bfc09ba0
MD
293static __inline
294vm_pindex_t
48ffc236 295pmap_pdpe_index(vm_offset_t va)
48ffc236
JG
296{
297
298 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
299}
300
bfc09ba0
MD
301static __inline
302vm_pindex_t
48ffc236 303pmap_pml4e_index(vm_offset_t va)
48ffc236
JG
304{
305
306 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
307}
308
309/* Return a pointer to the PML4 slot that corresponds to a VA */
bfc09ba0
MD
310static __inline
311pml4_entry_t *
48ffc236 312pmap_pml4e(pmap_t pmap, vm_offset_t va)
48ffc236
JG
313{
314
315 return (&pmap->pm_pml4[pmap_pml4e_index(va)]);
316}
317
318/* Return a pointer to the PDP slot that corresponds to a VA */
bfc09ba0
MD
319static __inline
320pdp_entry_t *
48ffc236 321pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va)
48ffc236
JG
322{
323 pdp_entry_t *pdpe;
324
325 pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME);
326 return (&pdpe[pmap_pdpe_index(va)]);
327}
328
329/* Return a pointer to the PDP slot that corresponds to a VA */
bfc09ba0
MD
330static __inline
331pdp_entry_t *
48ffc236 332pmap_pdpe(pmap_t pmap, vm_offset_t va)
48ffc236
JG
333{
334 pml4_entry_t *pml4e;
335
336 pml4e = pmap_pml4e(pmap, va);
337 if ((*pml4e & PG_V) == 0)
338 return NULL;
339 return (pmap_pml4e_to_pdpe(pml4e, va));
340}
341
342/* Return a pointer to the PD slot that corresponds to a VA */
bfc09ba0
MD
343static __inline
344pd_entry_t *
48ffc236 345pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va)
48ffc236
JG
346{
347 pd_entry_t *pde;
348
349 pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME);
350 return (&pde[pmap_pde_index(va)]);
351}
352
353/* Return a pointer to the PD slot that corresponds to a VA */
bfc09ba0
MD
354static __inline
355pd_entry_t *
48ffc236 356pmap_pde(pmap_t pmap, vm_offset_t va)
48ffc236
JG
357{
358 pdp_entry_t *pdpe;
359
360 pdpe = pmap_pdpe(pmap, va);
361 if (pdpe == NULL || (*pdpe & PG_V) == 0)
362 return NULL;
363 return (pmap_pdpe_to_pde(pdpe, va));
364}
365
366/* Return a pointer to the PT slot that corresponds to a VA */
bfc09ba0
MD
367static __inline
368pt_entry_t *
48ffc236 369pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va)
48ffc236
JG
370{
371 pt_entry_t *pte;
372
373 pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME);
374 return (&pte[pmap_pte_index(va)]);
375}
376
377/* Return a pointer to the PT slot that corresponds to a VA */
bfc09ba0
MD
378static __inline
379pt_entry_t *
48ffc236 380pmap_pte(pmap_t pmap, vm_offset_t va)
48ffc236
JG
381{
382 pd_entry_t *pde;
383
384 pde = pmap_pde(pmap, va);
385 if (pde == NULL || (*pde & PG_V) == 0)
386 return NULL;
387 if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */
388 return ((pt_entry_t *)pde);
389 return (pmap_pde_to_pte(pde, va));
390}
391
bfc09ba0
MD
392static __inline
393pt_entry_t *
48ffc236 394vtopte(vm_offset_t va)
48ffc236
JG
395{
396 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
397
398 return (PTmap + ((va >> PAGE_SHIFT) & mask));
c8fe38ae 399}
d7f50089 400
bfc09ba0
MD
401static __inline
402pd_entry_t *
48ffc236 403vtopde(vm_offset_t va)
48ffc236
JG
404{
405 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
406
407 return (PDmap + ((va >> PDRSHIFT) & mask));
408}
c8fe38ae 409
48ffc236 410static uint64_t
c8fe38ae 411allocpages(vm_paddr_t *firstaddr, int n)
d7f50089 412{
48ffc236 413 uint64_t ret;
c8fe38ae
MD
414
415 ret = *firstaddr;
416 bzero((void *)ret, n * PAGE_SIZE);
417 *firstaddr += n * PAGE_SIZE;
418 return (ret);
d7f50089
YY
419}
420
bfc09ba0 421static
c8fe38ae
MD
422void
423create_pagetables(vm_paddr_t *firstaddr)
424{
425 int i;
c8fe38ae
MD
426
427 /* we are running (mostly) V=P at this point */
428
48ffc236
JG
429 /* Allocate pages */
430 KPTphys = allocpages(firstaddr, NKPT);
431 KPML4phys = allocpages(firstaddr, 1);
432 KPDPphys = allocpages(firstaddr, NKPML4E);
433 KPDphys = allocpages(firstaddr, NKPDPE);
434
435 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
436 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
437 ndmpdp = 4;
438 DMPDPphys = allocpages(firstaddr, NDMPML4E);
439 if ((amd_feature & AMDID_PAGE1GB) == 0)
440 DMPDphys = allocpages(firstaddr, ndmpdp);
441 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
442
443 /* Fill in the underlying page table pages */
444 /* Read-only from zero to physfree */
445 /* XXX not fully used, underneath 2M pages */
446 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
447 ((pt_entry_t *)KPTphys)[i] = i << PAGE_SHIFT;
448 ((pt_entry_t *)KPTphys)[i] |= PG_RW | PG_V | PG_G;
449 }
450
451 /* Now map the page tables at their location within PTmap */
452 for (i = 0; i < NKPT; i++) {
453 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
454 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
455 }
456
457 /* Map from zero to end of allocations under 2M pages */
458 /* This replaces some of the KPTphys entries above */
459 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
460 ((pd_entry_t *)KPDphys)[i] = i << PDRSHIFT;
461 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V | PG_PS | PG_G;
462 }
463
464 /* And connect up the PD to the PDP */
465 for (i = 0; i < NKPDPE; i++) {
466 ((pdp_entry_t *)KPDPphys)[i + KPDPI] = KPDphys +
467 (i << PAGE_SHIFT);
468 ((pdp_entry_t *)KPDPphys)[i + KPDPI] |= PG_RW | PG_V | PG_U;
469 }
470
471 /* Now set up the direct map space using either 2MB or 1GB pages */
472 /* Preset PG_M and PG_A because demotion expects it */
473 if ((amd_feature & AMDID_PAGE1GB) == 0) {
474 for (i = 0; i < NPDEPG * ndmpdp; i++) {
475 ((pd_entry_t *)DMPDphys)[i] = (vm_paddr_t)i << PDRSHIFT;
476 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
477 PG_G | PG_M | PG_A;
478 }
479 /* And the direct map space's PDP */
480 for (i = 0; i < ndmpdp; i++) {
481 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
482 (i << PAGE_SHIFT);
483 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
484 }
485 } else {
486 for (i = 0; i < ndmpdp; i++) {
487 ((pdp_entry_t *)DMPDPphys)[i] =
488 (vm_paddr_t)i << PDPSHIFT;
489 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
490 PG_G | PG_M | PG_A;
491 }
492 }
493
494 /* And recursively map PML4 to itself in order to get PTmap */
495 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
496 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
497
498 /* Connect the Direct Map slot up to the PML4 */
499 ((pdp_entry_t *)KPML4phys)[DMPML4I] = DMPDPphys;
500 ((pdp_entry_t *)KPML4phys)[DMPML4I] |= PG_RW | PG_V | PG_U;
501
502 /* Connect the KVA slot up to the PML4 */
503 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
504 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
c8fe38ae
MD
505}
506
507void
f81851b8
MD
508init_paging(vm_paddr_t *firstaddr)
509{
c8fe38ae 510 create_pagetables(firstaddr);
d7f50089
YY
511}
512
513/*
c8fe38ae
MD
514 * Bootstrap the system enough to run with virtual memory.
515 *
516 * On the i386 this is called after mapping has already been enabled
517 * and just syncs the pmap module with what has already been done.
518 * [We can't call it easily with mapping off since the kernel is not
519 * mapped with PA == VA, hence we would have to relocate every address
520 * from the linked base (virtual) address "KERNBASE" to the actual
521 * (physical) address starting relative to 0]
d7f50089
YY
522 */
523void
48ffc236 524pmap_bootstrap(vm_paddr_t *firstaddr)
c8fe38ae
MD
525{
526 vm_offset_t va;
527 pt_entry_t *pte;
528 struct mdglobaldata *gd;
c8fe38ae
MD
529 int pg;
530
48ffc236
JG
531 KvaStart = VM_MIN_KERNEL_ADDRESS;
532 KvaEnd = VM_MAX_KERNEL_ADDRESS;
533 KvaSize = KvaEnd - KvaStart;
534
c8fe38ae
MD
535 avail_start = *firstaddr;
536
537 /*
48ffc236 538 * Create an initial set of page tables to run the kernel in.
c8fe38ae 539 */
48ffc236
JG
540 create_pagetables(firstaddr);
541
c8fe38ae
MD
542 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
543 virtual_start = pmap_kmem_choose(virtual_start);
48ffc236
JG
544
545 virtual_end = VM_MAX_KERNEL_ADDRESS;
546
547 /* XXX do %cr0 as well */
548 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
549 load_cr3(KPML4phys);
c8fe38ae
MD
550
551 /*
552 * Initialize protection array.
553 */
554 i386_protection_init();
555
556 /*
557 * The kernel's pmap is statically allocated so we don't have to use
558 * pmap_create, which is unlikely to work correctly at this part of
559 * the boot sequence (XXX and which no longer exists).
560 */
48ffc236 561 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
c8fe38ae
MD
562 kernel_pmap.pm_count = 1;
563 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
564 TAILQ_INIT(&kernel_pmap.pm_pvlist);
565 nkpt = NKPT;
566
567 /*
568 * Reserve some special page table entries/VA space for temporary
569 * mapping of pages.
570 */
571#define SYSMAP(c, p, v, n) \
572 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
573
574 va = virtual_start;
48ffc236 575#ifdef JG
c8fe38ae 576 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
48ffc236
JG
577#else
578 pte = vtopte(va);
579#endif
c8fe38ae
MD
580
581 /*
582 * CMAP1/CMAP2 are used for zeroing and copying pages.
583 */
584 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
585
586 /*
587 * Crashdump maps.
588 */
589 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
590
591 /*
592 * ptvmmap is used for reading arbitrary physical pages via
593 * /dev/mem.
594 */
595 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
596
597 /*
598 * msgbufp is used to map the system message buffer.
599 * XXX msgbufmap is not used.
600 */
601 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
602 atop(round_page(MSGBUF_SIZE)))
603
604 virtual_start = va;
605
606 *CMAP1 = 0;
c8fe38ae
MD
607
608 /*
609 * PG_G is terribly broken on SMP because we IPI invltlb's in some
610 * cases rather then invl1pg. Actually, I don't even know why it
611 * works under UP because self-referential page table mappings
612 */
613#ifdef SMP
614 pgeflag = 0;
615#else
616 if (cpu_feature & CPUID_PGE)
617 pgeflag = PG_G;
618#endif
619
620/*
621 * Initialize the 4MB page size flag
622 */
623 pseflag = 0;
624/*
625 * The 4MB page version of the initial
626 * kernel page mapping.
627 */
628 pdir4mb = 0;
629
630#if !defined(DISABLE_PSE)
631 if (cpu_feature & CPUID_PSE) {
632 pt_entry_t ptditmp;
633 /*
634 * Note that we have enabled PSE mode
635 */
636 pseflag = PG_PS;
b2b3ffcd 637 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
c8fe38ae
MD
638 ptditmp &= ~(NBPDR - 1);
639 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
640 pdir4mb = ptditmp;
641
642#ifndef SMP
643 /*
644 * Enable the PSE mode. If we are SMP we can't do this
645 * now because the APs will not be able to use it when
646 * they boot up.
647 */
648 load_cr4(rcr4() | CR4_PSE);
649
650 /*
651 * We can do the mapping here for the single processor
652 * case. We simply ignore the old page table page from
653 * now on.
654 */
655 /*
656 * For SMP, we still need 4K pages to bootstrap APs,
657 * PSE will be enabled as soon as all APs are up.
658 */
659 PTD[KPTDI] = (pd_entry_t)ptditmp;
c8fe38ae
MD
660 cpu_invltlb();
661#endif
662 }
663#endif
664#ifdef SMP
665 if (cpu_apic_address == 0)
666 panic("pmap_bootstrap: no local apic!");
c8fe38ae
MD
667#endif
668
669 /*
670 * We need to finish setting up the globaldata page for the BSP.
671 * locore has already populated the page table for the mdglobaldata
672 * portion.
673 */
674 pg = MDGLOBALDATA_BASEALLOC_PAGES;
675 gd = &CPU_prvspace[0].mdglobaldata;
676 gd->gd_CMAP1 = &SMPpt[pg + 0];
677 gd->gd_CMAP2 = &SMPpt[pg + 1];
678 gd->gd_CMAP3 = &SMPpt[pg + 2];
679 gd->gd_PMAP1 = &SMPpt[pg + 3];
680 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
681 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
682 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
683 gd->gd_PADDR1 = (pt_entry_t *)CPU_prvspace[0].PPAGE1;
684
685 cpu_invltlb();
d7f50089
YY
686}
687
c8fe38ae 688#ifdef SMP
d7f50089 689/*
c8fe38ae 690 * Set 4mb pdir for mp startup
d7f50089
YY
691 */
692void
c8fe38ae
MD
693pmap_set_opt(void)
694{
695 if (pseflag && (cpu_feature & CPUID_PSE)) {
696 load_cr4(rcr4() | CR4_PSE);
697 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
c8fe38ae
MD
698 cpu_invltlb();
699 }
700 }
d7f50089 701}
c8fe38ae 702#endif
d7f50089 703
c8fe38ae
MD
704/*
705 * Initialize the pmap module.
706 * Called by vm_init, to initialize any structures that the pmap
707 * system needs to map virtual memory.
708 * pmap_init has been enhanced to support in a fairly consistant
709 * way, discontiguous physical memory.
d7f50089
YY
710 */
711void
c8fe38ae 712pmap_init(void)
d7f50089 713{
c8fe38ae
MD
714 int i;
715 int initial_pvs;
716
717 /*
718 * object for kernel page table pages
719 */
48ffc236
JG
720 /* JG I think the number can be arbitrary */
721 kptobj = vm_object_allocate(OBJT_DEFAULT, 5);
c8fe38ae
MD
722
723 /*
724 * Allocate memory for random pmap data structures. Includes the
725 * pv_head_table.
726 */
727
728 for(i = 0; i < vm_page_array_size; i++) {
729 vm_page_t m;
730
731 m = &vm_page_array[i];
732 TAILQ_INIT(&m->md.pv_list);
733 m->md.pv_list_count = 0;
734 }
735
736 /*
737 * init the pv free list
738 */
739 initial_pvs = vm_page_array_size;
740 if (initial_pvs < MINPV)
741 initial_pvs = MINPV;
742 pvzone = &pvzone_store;
743 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
744 initial_pvs * sizeof (struct pv_entry));
745 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
746 initial_pvs);
747
748 /*
749 * Now it is safe to enable pv_table recording.
750 */
751 pmap_initialized = TRUE;
d887674b 752#ifdef SMP
057877ac 753 lapic = pmap_mapdev_uncacheable(cpu_apic_address, sizeof(struct LAPIC));
d887674b 754#endif
d7f50089
YY
755}
756
c8fe38ae
MD
757/*
758 * Initialize the address space (zone) for the pv_entries. Set a
759 * high water mark so that the system can recover from excessive
760 * numbers of pv entries.
761 */
d7f50089 762void
c8fe38ae 763pmap_init2(void)
d7f50089 764{
c8fe38ae
MD
765 int shpgperproc = PMAP_SHPGPERPROC;
766
767 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
768 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
769 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
770 pv_entry_high_water = 9 * (pv_entry_max / 10);
771 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
d7f50089
YY
772}
773
c8fe38ae
MD
774
775/***************************************************
776 * Low level helper routines.....
777 ***************************************************/
778
779#if defined(PMAP_DIAGNOSTIC)
d7f50089
YY
780
781/*
c8fe38ae
MD
782 * This code checks for non-writeable/modified pages.
783 * This should be an invalid condition.
d7f50089 784 */
bfc09ba0
MD
785static
786int
48ffc236 787pmap_nw_modified(pt_entry_t pte)
d7f50089 788{
c8fe38ae
MD
789 if ((pte & (PG_M|PG_RW)) == PG_M)
790 return 1;
791 else
792 return 0;
d7f50089 793}
c8fe38ae
MD
794#endif
795
d7f50089 796
c8fe38ae
MD
797/*
798 * this routine defines the region(s) of memory that should
799 * not be tested for the modified bit.
800 */
bfc09ba0
MD
801static __inline
802int
c8fe38ae 803pmap_track_modified(vm_offset_t va)
d7f50089 804{
c8fe38ae
MD
805 if ((va < clean_sva) || (va >= clean_eva))
806 return 1;
807 else
808 return 0;
d7f50089
YY
809}
810
d7f50089 811/*
c8fe38ae
MD
812 * pmap_extract:
813 *
814 * Extract the physical page address associated with the map/VA pair.
815 *
816 * This function may not be called from an interrupt if the pmap is
817 * not kernel_pmap.
d7f50089 818 */
c8fe38ae
MD
819vm_paddr_t
820pmap_extract(pmap_t pmap, vm_offset_t va)
d7f50089 821{
48ffc236
JG
822 vm_paddr_t rtval;
823 pt_entry_t *pte;
824 pd_entry_t pde, *pdep;
c8fe38ae 825
48ffc236
JG
826 rtval = 0;
827 pdep = pmap_pde(pmap, va);
828 if (pdep != NULL) {
829 pde = *pdep;
830 if (pde) {
831 if ((pde & PG_PS) != 0) {
832 rtval = (pde & PG_PS_FRAME) | (va & PDRMASK);
833 } else {
834 pte = pmap_pde_to_pte(pdep, va);
835 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
836 }
c8fe38ae 837 }
c8fe38ae 838 }
48ffc236
JG
839 return rtval;
840}
841
842/*
843 * Routine: pmap_kextract
844 * Function:
845 * Extract the physical page address associated
846 * kernel virtual address.
847 */
848vm_paddr_t
849pmap_kextract(vm_offset_t va)
48ffc236
JG
850{
851 pd_entry_t pde;
852 vm_paddr_t pa;
853
854 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
855 pa = DMAP_TO_PHYS(va);
856 } else {
857 pde = *vtopde(va);
858 if (pde & PG_PS) {
859 pa = (pde & PG_PS_FRAME) | (va & PDRMASK);
860 } else {
861 /*
862 * Beware of a concurrent promotion that changes the
863 * PDE at this point! For example, vtopte() must not
864 * be used to access the PTE because it would use the
865 * new PDE. It is, however, safe to use the old PDE
866 * because the page table page is preserved by the
867 * promotion.
868 */
869 pa = *pmap_pde_to_pte(&pde, va);
870 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
871 }
872 }
873 return pa;
d7f50089
YY
874}
875
c8fe38ae
MD
876/***************************************************
877 * Low level mapping routines.....
878 ***************************************************/
879
d7f50089 880/*
c8fe38ae
MD
881 * Routine: pmap_kenter
882 * Function:
883 * Add a wired page to the KVA
884 * NOTE! note that in order for the mapping to take effect -- you
885 * should do an invltlb after doing the pmap_kenter().
d7f50089 886 */
c8fe38ae 887void
d7f50089
YY
888pmap_kenter(vm_offset_t va, vm_paddr_t pa)
889{
c8fe38ae
MD
890 pt_entry_t *pte;
891 pt_entry_t npte;
892 pmap_inval_info info;
893
894 pmap_inval_init(&info);
895 npte = pa | PG_RW | PG_V | pgeflag;
896 pte = vtopte(va);
897 pmap_inval_add(&info, &kernel_pmap, va);
898 *pte = npte;
899 pmap_inval_flush(&info);
d7f50089
YY
900}
901
902/*
c8fe38ae
MD
903 * Routine: pmap_kenter_quick
904 * Function:
905 * Similar to pmap_kenter(), except we only invalidate the
906 * mapping on the current CPU.
d7f50089 907 */
c8fe38ae
MD
908void
909pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
910{
911 pt_entry_t *pte;
912 pt_entry_t npte;
913
914 npte = pa | PG_RW | PG_V | pgeflag;
915 pte = vtopte(va);
916 *pte = npte;
917 cpu_invlpg((void *)va);
918}
919
d7f50089
YY
920void
921pmap_kenter_sync(vm_offset_t va)
922{
c8fe38ae
MD
923 pmap_inval_info info;
924
925 pmap_inval_init(&info);
926 pmap_inval_add(&info, &kernel_pmap, va);
927 pmap_inval_flush(&info);
d7f50089
YY
928}
929
d7f50089
YY
930void
931pmap_kenter_sync_quick(vm_offset_t va)
932{
c8fe38ae 933 cpu_invlpg((void *)va);
d7f50089
YY
934}
935
d7f50089 936/*
c8fe38ae 937 * remove a page from the kernel pagetables
d7f50089
YY
938 */
939void
c8fe38ae 940pmap_kremove(vm_offset_t va)
d7f50089 941{
c8fe38ae
MD
942 pt_entry_t *pte;
943 pmap_inval_info info;
944
945 pmap_inval_init(&info);
946 pte = vtopte(va);
947 pmap_inval_add(&info, &kernel_pmap, va);
948 *pte = 0;
949 pmap_inval_flush(&info);
950}
951
952void
953pmap_kremove_quick(vm_offset_t va)
954{
955 pt_entry_t *pte;
956 pte = vtopte(va);
957 *pte = 0;
958 cpu_invlpg((void *)va);
d7f50089
YY
959}
960
961/*
c8fe38ae 962 * XXX these need to be recoded. They are not used in any critical path.
d7f50089
YY
963 */
964void
c8fe38ae 965pmap_kmodify_rw(vm_offset_t va)
d7f50089 966{
c8fe38ae
MD
967 *vtopte(va) |= PG_RW;
968 cpu_invlpg((void *)va);
d7f50089
YY
969}
970
c8fe38ae
MD
971void
972pmap_kmodify_nc(vm_offset_t va)
973{
974 *vtopte(va) |= PG_N;
975 cpu_invlpg((void *)va);
976}
d7f50089
YY
977
978/*
c8fe38ae
MD
979 * Used to map a range of physical addresses into kernel
980 * virtual address space.
981 *
982 * For now, VM is already on, we only need to map the
983 * specified memory.
d7f50089
YY
984 */
985vm_offset_t
8e5e6f1b 986pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
d7f50089 987{
8fdd3267 988 return PHYS_TO_DMAP(start);
d7f50089
YY
989}
990
c8fe38ae 991
d7f50089 992/*
c8fe38ae
MD
993 * Add a list of wired pages to the kva
994 * this routine is only used for temporary
995 * kernel mappings that do not need to have
996 * page modification or references recorded.
997 * Note that old mappings are simply written
998 * over. The page *must* be wired.
d7f50089
YY
999 */
1000void
c8fe38ae 1001pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
d7f50089 1002{
c8fe38ae
MD
1003 vm_offset_t end_va;
1004
1005 end_va = va + count * PAGE_SIZE;
1006
1007 while (va < end_va) {
1008 pt_entry_t *pte;
1009
1010 pte = vtopte(va);
1011 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1012 cpu_invlpg((void *)va);
1013 va += PAGE_SIZE;
1014 m++;
1015 }
1016#ifdef SMP
1017 smp_invltlb(); /* XXX */
1018#endif
1019}
1020
1021void
1022pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
1023{
1024 vm_offset_t end_va;
1025 cpumask_t cmask = mycpu->gd_cpumask;
1026
1027 end_va = va + count * PAGE_SIZE;
1028
1029 while (va < end_va) {
1030 pt_entry_t *pte;
1031 pt_entry_t pteval;
1032
1033 /*
1034 * Install the new PTE. If the pte changed from the prior
1035 * mapping we must reset the cpu mask and invalidate the page.
1036 * If the pte is the same but we have not seen it on the
1037 * current cpu, invlpg the existing mapping. Otherwise the
1038 * entry is optimal and no invalidation is required.
1039 */
1040 pte = vtopte(va);
1041 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
1042 if (*pte != pteval) {
1043 *mask = 0;
1044 *pte = pteval;
1045 cpu_invlpg((void *)va);
1046 } else if ((*mask & cmask) == 0) {
1047 cpu_invlpg((void *)va);
1048 }
1049 va += PAGE_SIZE;
1050 m++;
1051 }
1052 *mask |= cmask;
d7f50089
YY
1053}
1054
1055/*
7155fc7d 1056 * This routine jerks page mappings from the
c8fe38ae 1057 * kernel -- it is meant only for temporary mappings.
7155fc7d
MD
1058 *
1059 * MPSAFE, INTERRUPT SAFE (cluster callback)
d7f50089 1060 */
c8fe38ae
MD
1061void
1062pmap_qremove(vm_offset_t va, int count)
d7f50089 1063{
c8fe38ae
MD
1064 vm_offset_t end_va;
1065
48ffc236 1066 end_va = va + count * PAGE_SIZE;
c8fe38ae
MD
1067
1068 while (va < end_va) {
1069 pt_entry_t *pte;
1070
1071 pte = vtopte(va);
1072 *pte = 0;
1073 cpu_invlpg((void *)va);
1074 va += PAGE_SIZE;
1075 }
1076#ifdef SMP
1077 smp_invltlb();
1078#endif
d7f50089
YY
1079}
1080
1081/*
c8fe38ae
MD
1082 * This routine works like vm_page_lookup() but also blocks as long as the
1083 * page is busy. This routine does not busy the page it returns.
1084 *
1085 * Unless the caller is managing objects whos pages are in a known state,
1086 * the call should be made with a critical section held so the page's object
1087 * association remains valid on return.
d7f50089 1088 */
bfc09ba0
MD
1089static
1090vm_page_t
c8fe38ae 1091pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
d7f50089 1092{
c8fe38ae
MD
1093 vm_page_t m;
1094
1095 do {
1096 m = vm_page_lookup(object, pindex);
1097 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
1098
1099 return(m);
d7f50089
YY
1100}
1101
1102/*
c8fe38ae
MD
1103 * Create a new thread and optionally associate it with a (new) process.
1104 * NOTE! the new thread's cpu may not equal the current cpu.
d7f50089
YY
1105 */
1106void
c8fe38ae 1107pmap_init_thread(thread_t td)
d7f50089 1108{
c8fe38ae
MD
1109 /* enforce pcb placement */
1110 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
1111 td->td_savefpu = &td->td_pcb->pcb_save;
b2b3ffcd 1112 td->td_sp = (char *)td->td_pcb - 16; /* JG is -16 needed on x86_64? */
d7f50089
YY
1113}
1114
1115/*
c8fe38ae 1116 * This routine directly affects the fork perf for a process.
d7f50089
YY
1117 */
1118void
c8fe38ae 1119pmap_init_proc(struct proc *p)
d7f50089
YY
1120{
1121}
1122
1123/*
c8fe38ae
MD
1124 * Dispose the UPAGES for a process that has exited.
1125 * This routine directly impacts the exit perf of a process.
d7f50089
YY
1126 */
1127void
c8fe38ae 1128pmap_dispose_proc(struct proc *p)
d7f50089 1129{
c8fe38ae 1130 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
d7f50089
YY
1131}
1132
c8fe38ae
MD
1133/***************************************************
1134 * Page table page management routines.....
1135 ***************************************************/
1136
d7f50089 1137/*
c8fe38ae
MD
1138 * This routine unholds page table pages, and if the hold count
1139 * drops to zero, then it decrements the wire count.
d7f50089 1140 */
bfc09ba0
MD
1141static __inline
1142int
1143pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1144 pmap_inval_info_t info)
1145{
1146 KKASSERT(m->hold_count > 0);
1147 if (m->hold_count > 1) {
1148 vm_page_unhold(m);
1149 return 0;
1150 } else {
1151 return _pmap_unwire_pte_hold(pmap, va, m, info);
1152 }
1153}
1154
1155static
1156int
1157_pmap_unwire_pte_hold(pmap_t pmap, vm_offset_t va, vm_page_t m,
1158 pmap_inval_info_t info)
c8fe38ae
MD
1159{
1160 /*
1161 * Wait until we can busy the page ourselves. We cannot have
700e22f7
MD
1162 * any active flushes if we block. We own one hold count on the
1163 * page so it cannot be freed out from under us.
c8fe38ae
MD
1164 */
1165 if (m->flags & PG_BUSY) {
1166 pmap_inval_flush(info);
1167 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1168 ;
1169 }
1170 KASSERT(m->queue == PQ_NONE,
1171 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
1172
700e22f7
MD
1173 /*
1174 * This case can occur if new references were acquired while
1175 * we were blocked.
1176 */
1177 if (m->hold_count > 1) {
1178 KKASSERT(m->hold_count > 1);
1179 vm_page_unhold(m);
1180 return 0;
1181 }
c8fe38ae 1182
700e22f7
MD
1183 /*
1184 * Unmap the page table page
1185 */
1186 KKASSERT(m->hold_count == 1);
1187 vm_page_busy(m);
1188 pmap_inval_add(info, pmap, -1);
c8fe38ae 1189
700e22f7
MD
1190 if (m->pindex >= (NUPDE + NUPDPE)) {
1191 /* PDP page */
1192 pml4_entry_t *pml4;
1193 pml4 = pmap_pml4e(pmap, va);
1194 *pml4 = 0;
1195 } else if (m->pindex >= NUPDE) {
1196 /* PD page */
1197 pdp_entry_t *pdp;
1198 pdp = pmap_pdpe(pmap, va);
1199 *pdp = 0;
1200 } else {
1201 /* PT page */
1202 pd_entry_t *pd;
1203 pd = pmap_pde(pmap, va);
1204 *pd = 0;
1205 }
c8fe38ae 1206
700e22f7
MD
1207 KKASSERT(pmap->pm_stats.resident_count > 0);
1208 --pmap->pm_stats.resident_count;
48ffc236 1209
700e22f7
MD
1210 if (pmap->pm_ptphint == m)
1211 pmap->pm_ptphint = NULL;
1212
1213 if (m->pindex < NUPDE) {
1214 /* We just released a PT, unhold the matching PD */
1215 vm_page_t pdpg;
1216
1217 pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME);
1218 pmap_unwire_pte_hold(pmap, va, pdpg, info);
1219 }
1220 if (m->pindex >= NUPDE && m->pindex < (NUPDE + NUPDPE)) {
1221 /* We just released a PD, unhold the matching PDP */
1222 vm_page_t pdppg;
1223
1224 pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME);
1225 pmap_unwire_pte_hold(pmap, va, pdppg, info);
c8fe38ae 1226 }
700e22f7
MD
1227
1228 /*
1229 * This was our last hold, the page had better be unwired
1230 * after we decrement wire_count.
1231 *
1232 * FUTURE NOTE: shared page directory page could result in
1233 * multiple wire counts.
1234 */
1235 vm_page_unhold(m);
1236 --m->wire_count;
1237 KKASSERT(m->wire_count == 0);
1238 --vmstats.v_wire_count;
1239 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1240 vm_page_flash(m);
1241 vm_page_free_zero(m);
1242
1243 return 1;
c8fe38ae
MD
1244}
1245
c8fe38ae
MD
1246/*
1247 * After removing a page table entry, this routine is used to
1248 * conditionally free the page, and manage the hold/wire counts.
d7f50089 1249 */
bfc09ba0
MD
1250static
1251int
c8fe38ae
MD
1252pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1253 pmap_inval_info_t info)
1254{
1255 vm_pindex_t ptepindex;
700e22f7 1256
48ffc236 1257 if (va >= VM_MAX_USER_ADDRESS)
c8fe38ae
MD
1258 return 0;
1259
1260 if (mpte == NULL) {
48ffc236
JG
1261 ptepindex = pmap_pde_pindex(va);
1262#if JGHINT
c8fe38ae
MD
1263 if (pmap->pm_ptphint &&
1264 (pmap->pm_ptphint->pindex == ptepindex)) {
1265 mpte = pmap->pm_ptphint;
1266 } else {
48ffc236 1267#endif
c8fe38ae 1268 pmap_inval_flush(info);
48ffc236 1269 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
c8fe38ae 1270 pmap->pm_ptphint = mpte;
48ffc236 1271#if JGHINT
c8fe38ae 1272 }
48ffc236 1273#endif
c8fe38ae 1274 }
48ffc236 1275 return pmap_unwire_pte_hold(pmap, va, mpte, info);
c8fe38ae 1276}
d7f50089
YY
1277
1278/*
c8fe38ae
MD
1279 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1280 * it, and IdlePTD, represents the template used to update all other pmaps.
1281 *
1282 * On architectures where the kernel pmap is not integrated into the user
1283 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1284 * kernel_pmap should be used to directly access the kernel_pmap.
d7f50089
YY
1285 */
1286void
c8fe38ae 1287pmap_pinit0(struct pmap *pmap)
d7f50089 1288{
48ffc236 1289 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
c8fe38ae
MD
1290 pmap->pm_count = 1;
1291 pmap->pm_active = 0;
1292 pmap->pm_ptphint = NULL;
1293 TAILQ_INIT(&pmap->pm_pvlist);
1294 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
d7f50089
YY
1295}
1296
1297/*
c8fe38ae
MD
1298 * Initialize a preallocated and zeroed pmap structure,
1299 * such as one in a vmspace structure.
d7f50089
YY
1300 */
1301void
c8fe38ae 1302pmap_pinit(struct pmap *pmap)
d7f50089 1303{
c8fe38ae
MD
1304 vm_page_t ptdpg;
1305
1306 /*
1307 * No need to allocate page table space yet but we do need a valid
1308 * page directory table.
1309 */
48ffc236
JG
1310 if (pmap->pm_pml4 == NULL) {
1311 pmap->pm_pml4 =
1312 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
c8fe38ae
MD
1313 }
1314
1315 /*
1316 * Allocate an object for the ptes
1317 */
1318 if (pmap->pm_pteobj == NULL)
0a5c555b 1319 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, NUPDE + NUPDPE + PML4PML4I + 1);
c8fe38ae
MD
1320
1321 /*
1322 * Allocate the page directory page, unless we already have
1323 * one cached. If we used the cached page the wire_count will
1324 * already be set appropriately.
1325 */
1326 if ((ptdpg = pmap->pm_pdirm) == NULL) {
0a5c555b 1327 ptdpg = vm_page_grab(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I,
c8fe38ae
MD
1328 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1329 pmap->pm_pdirm = ptdpg;
1330 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1331 ptdpg->valid = VM_PAGE_BITS_ALL;
700e22f7
MD
1332 if (ptdpg->wire_count == 0)
1333 ++vmstats.v_wire_count;
c8fe38ae 1334 ptdpg->wire_count = 1;
48ffc236 1335 pmap_kenter((vm_offset_t)pmap->pm_pml4, VM_PAGE_TO_PHYS(ptdpg));
c8fe38ae
MD
1336 }
1337 if ((ptdpg->flags & PG_ZERO) == 0)
48ffc236 1338 bzero(pmap->pm_pml4, PAGE_SIZE);
c8fe38ae 1339
48ffc236
JG
1340 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
1341 pmap->pm_pml4[DMPML4I] = DMPDPphys | PG_RW | PG_V | PG_U;
c8fe38ae
MD
1342
1343 /* install self-referential address mapping entry */
48ffc236 1344 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
c8fe38ae
MD
1345
1346 pmap->pm_count = 1;
1347 pmap->pm_active = 0;
1348 pmap->pm_ptphint = NULL;
1349 TAILQ_INIT(&pmap->pm_pvlist);
1350 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1351 pmap->pm_stats.resident_count = 1;
d7f50089
YY
1352}
1353
1354/*
c8fe38ae
MD
1355 * Clean up a pmap structure so it can be physically freed. This routine
1356 * is called by the vmspace dtor function. A great deal of pmap data is
1357 * left passively mapped to improve vmspace management so we have a bit
1358 * of cleanup work to do here.
d7f50089
YY
1359 */
1360void
c8fe38ae 1361pmap_puninit(pmap_t pmap)
d7f50089 1362{
c8fe38ae
MD
1363 vm_page_t p;
1364
1365 KKASSERT(pmap->pm_active == 0);
1366 if ((p = pmap->pm_pdirm) != NULL) {
48ffc236 1367 KKASSERT(pmap->pm_pml4 != NULL);
bfc09ba0 1368 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
48ffc236 1369 pmap_kremove((vm_offset_t)pmap->pm_pml4);
c8fe38ae
MD
1370 p->wire_count--;
1371 vmstats.v_wire_count--;
1372 KKASSERT((p->flags & PG_BUSY) == 0);
1373 vm_page_busy(p);
1374 vm_page_free_zero(p);
1375 pmap->pm_pdirm = NULL;
1376 }
48ffc236 1377 if (pmap->pm_pml4) {
bfc09ba0 1378 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
48ffc236
JG
1379 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1380 pmap->pm_pml4 = NULL;
c8fe38ae
MD
1381 }
1382 if (pmap->pm_pteobj) {
1383 vm_object_deallocate(pmap->pm_pteobj);
1384 pmap->pm_pteobj = NULL;
1385 }
d7f50089
YY
1386}
1387
1388/*
c8fe38ae
MD
1389 * Wire in kernel global address entries. To avoid a race condition
1390 * between pmap initialization and pmap_growkernel, this procedure
1391 * adds the pmap to the master list (which growkernel scans to update),
1392 * then copies the template.
d7f50089
YY
1393 */
1394void
c8fe38ae 1395pmap_pinit2(struct pmap *pmap)
d7f50089 1396{
c8fe38ae
MD
1397 crit_enter();
1398 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
1399 /* XXX copies current process, does not fill in MPPTDI */
c8fe38ae 1400 crit_exit();
d7f50089
YY
1401}
1402
1403/*
c8fe38ae
MD
1404 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
1405 * 0 on failure (if the procedure had to sleep).
d7f50089 1406 *
c8fe38ae
MD
1407 * When asked to remove the page directory page itself, we actually just
1408 * leave it cached so we do not have to incur the SMP inval overhead of
1409 * removing the kernel mapping. pmap_puninit() will take care of it.
d7f50089 1410 */
bfc09ba0
MD
1411static
1412int
c8fe38ae 1413pmap_release_free_page(struct pmap *pmap, vm_page_t p)
d7f50089 1414{
c8fe38ae
MD
1415 /*
1416 * This code optimizes the case of freeing non-busy
1417 * page-table pages. Those pages are zero now, and
1418 * might as well be placed directly into the zero queue.
1419 */
1420 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
d7f50089 1421 return 0;
d7f50089 1422
c8fe38ae
MD
1423 vm_page_busy(p);
1424
1425 /*
1426 * Remove the page table page from the processes address space.
1427 */
4a4ea614
MD
1428 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1429 /*
1430 * We are the pml4 table itself.
1431 */
1432 /* XXX anything to do here? */
1433 } else if (p->pindex >= (NUPDE + NUPDPE)) {
1b2e0b92 1434 /*
700e22f7
MD
1435 * Remove a PDP page from the PML4. We do not maintain
1436 * hold counts on the PML4 page.
1b2e0b92 1437 */
700e22f7
MD
1438 pml4_entry_t *pml4;
1439 vm_page_t m4;
1440 int idx;
1441
1442 m4 = vm_page_lookup(pmap->pm_pteobj, NUPDE + NUPDPE + PML4PML4I);
1b2e0b92 1443 KKASSERT(m4 != NULL);
700e22f7
MD
1444 pml4 = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m4));
1445 idx = (p->pindex - (NUPDE + NUPDPE)) % NPML4EPG;
1b2e0b92
JG
1446 KKASSERT(pml4[idx] != 0);
1447 pml4[idx] = 0;
1b2e0b92
JG
1448 } else if (p->pindex >= NUPDE) {
1449 /*
700e22f7
MD
1450 * Remove a PD page from the PDP and drop the hold count
1451 * on the PDP. The PDP is left cached in the pmap if
1452 * the hold count drops to 0 so the wire count remains
1453 * intact.
1b2e0b92 1454 */
700e22f7
MD
1455 vm_page_t m3;
1456 pdp_entry_t *pdp;
1457 int idx;
1458
1459 m3 = vm_page_lookup(pmap->pm_pteobj,
1460 NUPDE + NUPDPE + (p->pindex - NUPDE) / NPDPEPG);
1b2e0b92 1461 KKASSERT(m3 != NULL);
700e22f7
MD
1462 pdp = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m3));
1463 idx = (p->pindex - NUPDE) % NPDPEPG;
1b2e0b92
JG
1464 KKASSERT(pdp[idx] != 0);
1465 pdp[idx] = 0;
1466 m3->hold_count--;
1b2e0b92 1467 } else {
700e22f7
MD
1468 /*
1469 * Remove a PT page from the PD and drop the hold count
1470 * on the PD. The PD is left cached in the pmap if
1471 * the hold count drops to 0 so the wire count remains
1472 * intact.
1b2e0b92 1473 */
700e22f7
MD
1474 vm_page_t m2;
1475 pd_entry_t *pd;
1476 int idx;
1477
1478 m2 = vm_page_lookup(pmap->pm_pteobj,
1479 NUPDE + p->pindex / NPDEPG);
1b2e0b92 1480 KKASSERT(m2 != NULL);
700e22f7
MD
1481 pd = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m2));
1482 idx = p->pindex % NPDEPG;
1b2e0b92
JG
1483 pd[idx] = 0;
1484 m2->hold_count--;
1b2e0b92 1485 }
700e22f7
MD
1486
1487 /*
1488 * One fewer mappings in the pmap. p's hold count had better
1489 * be zero.
1490 */
c8fe38ae
MD
1491 KKASSERT(pmap->pm_stats.resident_count > 0);
1492 --pmap->pm_stats.resident_count;
700e22f7 1493 if (p->hold_count)
c8fe38ae 1494 panic("pmap_release: freeing held page table page");
c8fe38ae
MD
1495 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1496 pmap->pm_ptphint = NULL;
1497
1b2e0b92
JG
1498 /*
1499 * We leave the top-level page table page cached, wired, and mapped in
1500 * the pmap until the dtor function (pmap_puninit()) gets called.
1501 * However, still clean it up so we can set PG_ZERO.
1502 */
1503 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
1504 bzero(pmap->pm_pml4, PAGE_SIZE);
1505 vm_page_flag_set(p, PG_ZERO);
1506 vm_page_wakeup(p);
1507 } else {
1508 p->wire_count--;
700e22f7 1509 KKASSERT(p->wire_count == 0);
1b2e0b92
JG
1510 vmstats.v_wire_count--;
1511 /* JG eventually revert to using vm_page_free_zero() */
1512 vm_page_free(p);
1513 }
c8fe38ae
MD
1514 return 1;
1515}
d7f50089
YY
1516
1517/*
e8510e54
MD
1518 * This routine is called when various levels in the page table need to
1519 * be populated. This routine cannot fail.
d7f50089 1520 */
bfc09ba0
MD
1521static
1522vm_page_t
c8fe38ae
MD
1523_pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex)
1524{
e8510e54 1525 vm_page_t m;
c8fe38ae
MD
1526
1527 /*
e8510e54 1528 * Find or fabricate a new pagetable page. This will busy the page.
c8fe38ae
MD
1529 */
1530 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
700e22f7 1531 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
48ffc236
JG
1532 if ((m->flags & PG_ZERO) == 0) {
1533 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1534 }
1535
c8fe38ae
MD
1536 KASSERT(m->queue == PQ_NONE,
1537 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1538
1539 /*
1540 * Increment the hold count for the page we will be returning to
1541 * the caller.
1542 */
1543 m->hold_count++;
700e22f7 1544 if (m->wire_count++ == 0)
c8fe38ae 1545 vmstats.v_wire_count++;
c8fe38ae 1546
c8fe38ae
MD
1547 /*
1548 * Map the pagetable page into the process address space, if
1549 * it isn't already there.
e8510e54
MD
1550 *
1551 * It is possible that someone else got in and mapped the page
1552 * directory page while we were blocked, if so just unbusy and
1553 * return the held page.
c8fe38ae 1554 */
48ffc236 1555 if (ptepindex >= (NUPDE + NUPDPE)) {
e8510e54
MD
1556 /*
1557 * Wire up a new PDP page in the PML4
1558 */
48ffc236 1559 vm_pindex_t pml4index;
700e22f7 1560 pml4_entry_t *pml4;
48ffc236 1561
48ffc236
JG
1562 pml4index = ptepindex - (NUPDE + NUPDPE);
1563 pml4 = &pmap->pm_pml4[pml4index];
e8510e54 1564 if (*pml4 & PG_V) {
700e22f7
MD
1565 if (--m->wire_count == 0)
1566 --vmstats.v_wire_count;
e8510e54
MD
1567 vm_page_wakeup(m);
1568 return(m);
1569 }
48ffc236 1570 *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
48ffc236 1571 } else if (ptepindex >= NUPDE) {
e8510e54
MD
1572 /*
1573 * Wire up a new PD page in the PDP
1574 */
48ffc236
JG
1575 vm_pindex_t pml4index;
1576 vm_pindex_t pdpindex;
e8510e54 1577 vm_page_t pdppg;
48ffc236
JG
1578 pml4_entry_t *pml4;
1579 pdp_entry_t *pdp;
1580
48ffc236
JG
1581 pdpindex = ptepindex - NUPDE;
1582 pml4index = pdpindex >> NPML4EPGSHIFT;
1583
1584 pml4 = &pmap->pm_pml4[pml4index];
1585 if ((*pml4 & PG_V) == 0) {
e8510e54
MD
1586 /*
1587 * Have to allocate a new PDP page, recurse.
1588 * This always succeeds. Returned page will
1589 * be held.
1590 */
1591 pdppg = _pmap_allocpte(pmap,
1592 NUPDE + NUPDPE + pml4index);
48ffc236 1593 } else {
e8510e54
MD
1594 /*
1595 * Add a held reference to the PDP page.
1596 */
48ffc236 1597 pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME);
1b2e0b92 1598 pdppg->hold_count++;
48ffc236 1599 }
c8fe38ae 1600
e8510e54
MD
1601 /*
1602 * Now find the pdp_entry and map the PDP. If the PDP
1603 * has already been mapped unwind and return the
1604 * already-mapped PDP held.
700e22f7
MD
1605 *
1606 * pdppg is left held (hold_count is incremented for
1607 * each PD in the PDP).
e8510e54
MD
1608 */
1609 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
48ffc236 1610 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
e8510e54
MD
1611 if (*pdp & PG_V) {
1612 vm_page_unhold(pdppg);
700e22f7
MD
1613 if (--m->wire_count == 0)
1614 --vmstats.v_wire_count;
e8510e54
MD
1615 vm_page_wakeup(m);
1616 return(m);
1617 }
48ffc236 1618 *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
48ffc236 1619 } else {
e8510e54
MD
1620 /*
1621 * Wire up the new PT page in the PD
1622 */
48ffc236
JG
1623 vm_pindex_t pml4index;
1624 vm_pindex_t pdpindex;
1625 pml4_entry_t *pml4;
1626 pdp_entry_t *pdp;
1627 pd_entry_t *pd;
e8510e54 1628 vm_page_t pdpg;
48ffc236 1629
48ffc236
JG
1630 pdpindex = ptepindex >> NPDPEPGSHIFT;
1631 pml4index = pdpindex >> NPML4EPGSHIFT;
1632
e8510e54
MD
1633 /*
1634 * Locate the PDP page in the PML4, then the PD page in
1635 * the PDP. If either does not exist we simply recurse
1636 * to allocate them.
1637 *
1638 * We can just recurse on the PD page as it will recurse
1639 * on the PDP if necessary.
1640 */
48ffc236
JG
1641 pml4 = &pmap->pm_pml4[pml4index];
1642 if ((*pml4 & PG_V) == 0) {
e8510e54 1643 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
48ffc236
JG
1644 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1645 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
c8fe38ae 1646 } else {
48ffc236
JG
1647 pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME);
1648 pdp = &pdp[pdpindex & ((1ul << NPDPEPGSHIFT) - 1)];
1649 if ((*pdp & PG_V) == 0) {
e8510e54 1650 pdpg = _pmap_allocpte(pmap, NUPDE + pdpindex);
48ffc236 1651 } else {
48ffc236 1652 pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
1b2e0b92 1653 pdpg->hold_count++;
48ffc236 1654 }
c8fe38ae 1655 }
48ffc236 1656
e8510e54
MD
1657 /*
1658 * Now fill in the pte in the PD. If the pte already exists
1659 * (again, if we raced the grab), unhold pdpg and unwire
1660 * m, returning a held m.
700e22f7
MD
1661 *
1662 * pdpg is left held (hold_count is incremented for
1663 * each PT in the PD).
e8510e54
MD
1664 */
1665 pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME);
48ffc236 1666 pd = &pd[ptepindex & ((1ul << NPDEPGSHIFT) - 1)];
700e22f7 1667 if (*pd != 0) {
e8510e54 1668 vm_page_unhold(pdpg);
700e22f7
MD
1669 if (--m->wire_count == 0)
1670 --vmstats.v_wire_count;
e8510e54
MD
1671 vm_page_wakeup(m);
1672 return(m);
1673 }
700e22f7 1674 *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M;
c8fe38ae
MD
1675 }
1676
48ffc236 1677 /*
e8510e54
MD
1678 * We successfully loaded a PDP, PD, or PTE. Set the page table hint,
1679 * valid bits, mapped flag, unbusy, and we're done.
48ffc236
JG
1680 */
1681 pmap->pm_ptphint = m;
700e22f7 1682 ++pmap->pm_stats.resident_count;
48ffc236 1683
c8fe38ae
MD
1684 m->valid = VM_PAGE_BITS_ALL;
1685 vm_page_flag_clear(m, PG_ZERO);
1686 vm_page_flag_set(m, PG_MAPPED);
1687 vm_page_wakeup(m);
1688
e8510e54 1689 return (m);
c8fe38ae
MD
1690}
1691
bfc09ba0
MD
1692static
1693vm_page_t
c8fe38ae 1694pmap_allocpte(pmap_t pmap, vm_offset_t va)
d7f50089 1695{
c8fe38ae 1696 vm_pindex_t ptepindex;
48ffc236 1697 pd_entry_t *pd;
c8fe38ae
MD
1698 vm_page_t m;
1699
1700 /*
1701 * Calculate pagetable page index
1702 */
48ffc236 1703 ptepindex = pmap_pde_pindex(va);
c8fe38ae
MD
1704
1705 /*
1706 * Get the page directory entry
1707 */
48ffc236 1708 pd = pmap_pde(pmap, va);
c8fe38ae
MD
1709
1710 /*
48ffc236 1711 * This supports switching from a 2MB page to a
c8fe38ae
MD
1712 * normal 4K page.
1713 */
48ffc236 1714 if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) {
1b2e0b92 1715 panic("no promotion/demotion yet");
48ffc236
JG
1716 *pd = 0;
1717 pd = NULL;
c8fe38ae
MD
1718 cpu_invltlb();
1719 smp_invltlb();
1720 }
1721
1722 /*
1723 * If the page table page is mapped, we just increment the
1724 * hold count, and activate it.
1725 */
48ffc236
JG
1726 if (pd != NULL && (*pd & PG_V) != 0) {
1727 /* YYY hint is used here on i386 */
1728 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1729 pmap->pm_ptphint = m;
c8fe38ae
MD
1730 m->hold_count++;
1731 return m;
1732 }
1733 /*
1734 * Here if the pte page isn't mapped, or if it has been deallocated.
1735 */
1736 return _pmap_allocpte(pmap, ptepindex);
d7f50089
YY
1737}
1738
c8fe38ae
MD
1739
1740/***************************************************
1741 * Pmap allocation/deallocation routines.
1742 ***************************************************/
1743
d7f50089 1744/*
c8fe38ae
MD
1745 * Release any resources held by the given physical map.
1746 * Called when a pmap initialized by pmap_pinit is being released.
1747 * Should only be called if the map contains no valid mappings.
d7f50089 1748 */
c8fe38ae 1749static int pmap_release_callback(struct vm_page *p, void *data);
d7f50089 1750
c8fe38ae
MD
1751void
1752pmap_release(struct pmap *pmap)
d7f50089 1753{
c8fe38ae
MD
1754 vm_object_t object = pmap->pm_pteobj;
1755 struct rb_vm_page_scan_info info;
1756
1757 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
1758#if defined(DIAGNOSTIC)
1759 if (object->ref_count != 1)
1760 panic("pmap_release: pteobj reference count != 1");
1761#endif
1762
1763 info.pmap = pmap;
1764 info.object = object;
1765 crit_enter();
1766 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1767 crit_exit();
1768
1769 do {
1770 crit_enter();
1771 info.error = 0;
1772 info.mpte = NULL;
1773 info.limit = object->generation;
1774
1775 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1776 pmap_release_callback, &info);
1777 if (info.error == 0 && info.mpte) {
1778 if (!pmap_release_free_page(pmap, info.mpte))
1779 info.error = 1;
1780 }
1781 crit_exit();
1782 } while (info.error);
d7f50089
YY
1783}
1784
bfc09ba0
MD
1785static
1786int
c8fe38ae 1787pmap_release_callback(struct vm_page *p, void *data)
d7f50089 1788{
c8fe38ae
MD
1789 struct rb_vm_page_scan_info *info = data;
1790
0a5c555b 1791 if (p->pindex == NUPDE + NUPDPE + PML4PML4I) {
c8fe38ae
MD
1792 info->mpte = p;
1793 return(0);
1794 }
1795 if (!pmap_release_free_page(info->pmap, p)) {
1796 info->error = 1;
1797 return(-1);
1798 }
1799 if (info->object->generation != info->limit) {
1800 info->error = 1;
1801 return(-1);
1802 }
1803 return(0);
d7f50089
YY
1804}
1805
1806/*
c8fe38ae 1807 * Grow the number of kernel page table entries, if needed.
d7f50089 1808 */
c8fe38ae
MD
1809void
1810pmap_growkernel(vm_offset_t addr)
d7f50089 1811{
48ffc236 1812 vm_paddr_t paddr;
c8fe38ae
MD
1813 vm_offset_t ptppaddr;
1814 vm_page_t nkpg;
48ffc236
JG
1815 pd_entry_t *pde, newpdir;
1816 pdp_entry_t newpdp;
c8fe38ae
MD
1817
1818 crit_enter();
1819 if (kernel_vm_end == 0) {
1820 kernel_vm_end = KERNBASE;
1821 nkpt = 0;
48ffc236 1822 while ((*pmap_pde(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
c8fe38ae
MD
1823 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1824 nkpt++;
48ffc236
JG
1825 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1826 kernel_vm_end = kernel_map.max_offset;
1827 break;
1828 }
c8fe38ae
MD
1829 }
1830 }
48ffc236
JG
1831 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1832 if (addr - 1 >= kernel_map.max_offset)
1833 addr = kernel_map.max_offset;
c8fe38ae 1834 while (kernel_vm_end < addr) {
48ffc236
JG
1835 pde = pmap_pde(&kernel_pmap, kernel_vm_end);
1836 if (pde == NULL) {
1837 /* We need a new PDP entry */
1838 nkpg = vm_page_alloc(kptobj, nkpt,
1839 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM
1840 | VM_ALLOC_INTERRUPT);
1841 if (nkpg == NULL)
1842 panic("pmap_growkernel: no memory to grow kernel");
48ffc236 1843 paddr = VM_PAGE_TO_PHYS(nkpg);
7f2a2740
MD
1844 if ((nkpg->flags & PG_ZERO) == 0)
1845 pmap_zero_page(paddr);
1846 vm_page_flag_clear(nkpg, PG_ZERO);
48ffc236
JG
1847 newpdp = (pdp_entry_t)
1848 (paddr | PG_V | PG_RW | PG_A | PG_M);
1849 *pmap_pdpe(&kernel_pmap, kernel_vm_end) = newpdp;
7f2a2740 1850 nkpt++;
48ffc236
JG
1851 continue; /* try again */
1852 }
1853 if ((*pde & PG_V) != 0) {
c8fe38ae 1854 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
48ffc236
JG
1855 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1856 kernel_vm_end = kernel_map.max_offset;
1857 break;
1858 }
c8fe38ae
MD
1859 continue;
1860 }
1861
1862 /*
1863 * This index is bogus, but out of the way
1864 */
48ffc236 1865 nkpg = vm_page_alloc(kptobj, nkpt,
c8fe38ae
MD
1866 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1867 if (nkpg == NULL)
1868 panic("pmap_growkernel: no memory to grow kernel");
1869
1870 vm_page_wire(nkpg);
1871 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1872 pmap_zero_page(ptppaddr);
7f2a2740 1873 vm_page_flag_clear(nkpg, PG_ZERO);
c8fe38ae 1874 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
c8fe38ae
MD
1875 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
1876 nkpt++;
1877
48ffc236
JG
1878 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1879 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
1880 kernel_vm_end = kernel_map.max_offset;
1881 break;
c8fe38ae 1882 }
c8fe38ae
MD
1883 }
1884 crit_exit();
d7f50089
YY
1885}
1886
1887/*
c8fe38ae
MD
1888 * Retire the given physical map from service.
1889 * Should only be called if the map contains
1890 * no valid mappings.
d7f50089 1891 */
c8fe38ae
MD
1892void
1893pmap_destroy(pmap_t pmap)
d7f50089 1894{
c8fe38ae
MD
1895 int count;
1896
1897 if (pmap == NULL)
1898 return;
1899
1900 count = --pmap->pm_count;
1901 if (count == 0) {
1902 pmap_release(pmap);
1903 panic("destroying a pmap is not yet implemented");
1904 }
d7f50089
YY
1905}
1906
1907/*
c8fe38ae 1908 * Add a reference to the specified pmap.
d7f50089 1909 */
c8fe38ae
MD
1910void
1911pmap_reference(pmap_t pmap)
d7f50089 1912{
c8fe38ae
MD
1913 if (pmap != NULL) {
1914 pmap->pm_count++;
1915 }
d7f50089
YY
1916}
1917
c8fe38ae
MD
1918/***************************************************
1919* page management routines.
1920 ***************************************************/
d7f50089
YY
1921
1922/*
1923 * free the pv_entry back to the free list. This function may be
1924 * called from an interrupt.
1925 */
bfc09ba0
MD
1926static __inline
1927void
d7f50089
YY
1928free_pv_entry(pv_entry_t pv)
1929{
c8fe38ae 1930 pv_entry_count--;
48ffc236 1931 KKASSERT(pv_entry_count >= 0);
c8fe38ae 1932 zfree(pvzone, pv);
d7f50089
YY
1933}
1934
1935/*
1936 * get a new pv_entry, allocating a block from the system
1937 * when needed. This function may be called from an interrupt.
1938 */
bfc09ba0
MD
1939static
1940pv_entry_t
d7f50089
YY
1941get_pv_entry(void)
1942{
c8fe38ae
MD
1943 pv_entry_count++;
1944 if (pv_entry_high_water &&
48ffc236
JG
1945 (pv_entry_count > pv_entry_high_water) &&
1946 (pmap_pagedaemon_waken == 0)) {
c8fe38ae 1947 pmap_pagedaemon_waken = 1;
48ffc236 1948 wakeup(&vm_pages_needed);
c8fe38ae
MD
1949 }
1950 return zalloc(pvzone);
d7f50089
YY
1951}
1952
1953/*
1954 * This routine is very drastic, but can save the system
1955 * in a pinch.
1956 */
1957void
1958pmap_collect(void)
1959{
c8fe38ae
MD
1960 int i;
1961 vm_page_t m;
1962 static int warningdone=0;
1963
1964 if (pmap_pagedaemon_waken == 0)
1965 return;
1966
1967 if (warningdone < 5) {
1968 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1969 warningdone++;
1970 }
1971
1972 for(i = 0; i < vm_page_array_size; i++) {
1973 m = &vm_page_array[i];
1974 if (m->wire_count || m->hold_count || m->busy ||
1975 (m->flags & PG_BUSY))
1976 continue;
1977 pmap_remove_all(m);
1978 }
48ffc236 1979 pmap_pagedaemon_waken = 0;
d7f50089
YY
1980}
1981
c8fe38ae 1982
d7f50089
YY
1983/*
1984 * If it is the first entry on the list, it is actually
1985 * in the header and we must copy the following entry up
1986 * to the header. Otherwise we must search the list for
1987 * the entry. In either case we free the now unused entry.
1988 */
bfc09ba0
MD
1989static
1990int
c8fe38ae
MD
1991pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1992 vm_offset_t va, pmap_inval_info_t info)
1993{
1994 pv_entry_t pv;
1995 int rtval;
1996
1997 crit_enter();
1998 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1999 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2000 if (pmap == pv->pv_pmap && va == pv->pv_va)
2001 break;
2002 }
2003 } else {
2004 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
2005 if (va == pv->pv_va)
2006 break;
2007 }
2008 }
2009
2010 rtval = 0;
48ffc236 2011 /* JGXXX When can 'pv' be NULL? */
c8fe38ae
MD
2012 if (pv) {
2013 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2014 m->md.pv_list_count--;
48ffc236 2015 KKASSERT(m->md.pv_list_count >= 0);
c8fe38ae
MD
2016 if (TAILQ_EMPTY(&m->md.pv_list))
2017 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2018 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2019 ++pmap->pm_generation;
2020 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
2021 free_pv_entry(pv);
2022 }
2023 crit_exit();
2024 return rtval;
d7f50089
YY
2025}
2026
2027/*
c8fe38ae
MD
2028 * Create a pv entry for page at pa for
2029 * (pmap, va).
d7f50089 2030 */
bfc09ba0
MD
2031static
2032void
d7f50089
YY
2033pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
2034{
c8fe38ae
MD
2035 pv_entry_t pv;
2036
2037 crit_enter();
2038 pv = get_pv_entry();
2039 pv->pv_va = va;
2040 pv->pv_pmap = pmap;
2041 pv->pv_ptem = mpte;
2042
2043 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
2044 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2045 m->md.pv_list_count++;
2046
2047 crit_exit();
d7f50089
YY
2048}
2049
2050/*
2051 * pmap_remove_pte: do the things to unmap a page in a process
2052 */
bfc09ba0
MD
2053static
2054int
c8fe38ae
MD
2055pmap_remove_pte(struct pmap *pmap, pt_entry_t *ptq, vm_offset_t va,
2056 pmap_inval_info_t info)
2057{
2058 pt_entry_t oldpte;
2059 vm_page_t m;
2060
2061 pmap_inval_add(info, pmap, va);
2062 oldpte = pte_load_clear(ptq);
2063 if (oldpte & PG_W)
2064 pmap->pm_stats.wired_count -= 1;
2065 /*
2066 * Machines that don't support invlpg, also don't support
2067 * PG_G. XXX PG_G is disabled for SMP so don't worry about
2068 * the SMP case.
2069 */
2070 if (oldpte & PG_G)
2071 cpu_invlpg((void *)va);
2072 KKASSERT(pmap->pm_stats.resident_count > 0);
2073 --pmap->pm_stats.resident_count;
2074 if (oldpte & PG_MANAGED) {
2075 m = PHYS_TO_VM_PAGE(oldpte);
2076 if (oldpte & PG_M) {
2077#if defined(PMAP_DIAGNOSTIC)
2078 if (pmap_nw_modified((pt_entry_t) oldpte)) {
2079 kprintf(
48ffc236 2080 "pmap_remove: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
c8fe38ae
MD
2081 va, oldpte);
2082 }
2083#endif
2084 if (pmap_track_modified(va))
2085 vm_page_dirty(m);
2086 }
2087 if (oldpte & PG_A)
2088 vm_page_flag_set(m, PG_REFERENCED);
2089 return pmap_remove_entry(pmap, m, va, info);
2090 } else {
2091 return pmap_unuse_pt(pmap, va, NULL, info);
2092 }
2093
d7f50089
YY
2094 return 0;
2095}
2096
2097/*
2098 * pmap_remove_page:
2099 *
2100 * Remove a single page from a process address space.
2101 *
2102 * This function may not be called from an interrupt if the pmap is
2103 * not kernel_pmap.
2104 */
bfc09ba0
MD
2105static
2106void
c8fe38ae
MD
2107pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
2108{
48ffc236 2109 pt_entry_t *pte;
c8fe38ae 2110
48ffc236
JG
2111 pte = pmap_pte(pmap, va);
2112 if (pte == NULL)
2113 return;
2114 if ((*pte & PG_V) == 0)
2115 return;
2116 pmap_remove_pte(pmap, pte, va, info);
d7f50089
YY
2117}
2118
2119/*
2120 * pmap_remove:
2121 *
2122 * Remove the given range of addresses from the specified map.
2123 *
2124 * It is assumed that the start and end are properly
2125 * rounded to the page size.
2126 *
2127 * This function may not be called from an interrupt if the pmap is
2128 * not kernel_pmap.
2129 */
2130void
2131pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
2132{
48ffc236
JG
2133 vm_offset_t va_next;
2134 pml4_entry_t *pml4e;
2135 pdp_entry_t *pdpe;
2136 pd_entry_t ptpaddr, *pde;
2137 pt_entry_t *pte;
c8fe38ae
MD
2138 struct pmap_inval_info info;
2139
2140 if (pmap == NULL)
2141 return;
2142
2143 if (pmap->pm_stats.resident_count == 0)
2144 return;
2145
2146 pmap_inval_init(&info);
2147
2148 /*
2149 * special handling of removing one page. a very
2150 * common operation and easy to short circuit some
2151 * code.
2152 */
48ffc236
JG
2153 if (sva + PAGE_SIZE == eva) {
2154 pde = pmap_pde(pmap, sva);
2155 if (pde && (*pde & PG_PS) == 0) {
2156 pmap_remove_page(pmap, sva, &info);
2157 pmap_inval_flush(&info);
2158 return;
2159 }
c8fe38ae
MD
2160 }
2161
48ffc236
JG
2162 for (; sva < eva; sva = va_next) {
2163 pml4e = pmap_pml4e(pmap, sva);
2164 if ((*pml4e & PG_V) == 0) {
2165 va_next = (sva + NBPML4) & ~PML4MASK;
2166 if (va_next < sva)
2167 va_next = eva;
2168 continue;
2169 }
c8fe38ae 2170
48ffc236
JG
2171 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2172 if ((*pdpe & PG_V) == 0) {
2173 va_next = (sva + NBPDP) & ~PDPMASK;
2174 if (va_next < sva)
2175 va_next = eva;
2176 continue;
2177 }
c8fe38ae
MD
2178
2179 /*
2180 * Calculate index for next page table.
2181 */
48ffc236
JG
2182 va_next = (sva + NBPDR) & ~PDRMASK;
2183 if (va_next < sva)
2184 va_next = eva;
c8fe38ae 2185
48ffc236
JG
2186 pde = pmap_pdpe_to_pde(pdpe, sva);
2187 ptpaddr = *pde;
c8fe38ae
MD
2188
2189 /*
48ffc236 2190 * Weed out invalid mappings.
c8fe38ae
MD
2191 */
2192 if (ptpaddr == 0)
2193 continue;
2194
48ffc236
JG
2195 /*
2196 * Check for large page.
2197 */
2198 if ((ptpaddr & PG_PS) != 0) {
2199 /* JG FreeBSD has more complex treatment here */
2200 pmap_inval_add(&info, pmap, -1);
2201 *pde = 0;
2202 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2203 continue;
2204 }
2205
c8fe38ae
MD
2206 /*
2207 * Limit our scan to either the end of the va represented
2208 * by the current page table page, or to the end of the
2209 * range being removed.
2210 */
48ffc236
JG
2211 if (va_next > eva)
2212 va_next = eva;
c8fe38ae
MD
2213
2214 /*
2215 * NOTE: pmap_remove_pte() can block.
2216 */
48ffc236
JG
2217 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2218 sva += PAGE_SIZE) {
2219 if (*pte == 0)
c8fe38ae 2220 continue;
48ffc236 2221 if (pmap_remove_pte(pmap, pte, sva, &info))
c8fe38ae
MD
2222 break;
2223 }
2224 }
2225 pmap_inval_flush(&info);
d7f50089
YY
2226}
2227
2228/*
2229 * pmap_remove_all:
2230 *
c8fe38ae
MD
2231 * Removes this physical page from all physical maps in which it resides.
2232 * Reflects back modify bits to the pager.
d7f50089 2233 *
c8fe38ae 2234 * This routine may not be called from an interrupt.
d7f50089 2235 */
c8fe38ae 2236
bfc09ba0
MD
2237static
2238void
d7f50089
YY
2239pmap_remove_all(vm_page_t m)
2240{
c8fe38ae
MD
2241 struct pmap_inval_info info;
2242 pt_entry_t *pte, tpte;
2243 pv_entry_t pv;
2244
2245 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2246 return;
2247
2248 pmap_inval_init(&info);
2249 crit_enter();
2250 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2251 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
2252 --pv->pv_pmap->pm_stats.resident_count;
2253
2254 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2255 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
2256 tpte = pte_load_clear(pte);
2257
2258 if (tpte & PG_W)
2259 pv->pv_pmap->pm_stats.wired_count--;
2260
2261 if (tpte & PG_A)
2262 vm_page_flag_set(m, PG_REFERENCED);
2263
2264 /*
2265 * Update the vm_page_t clean and reference bits.
2266 */
2267 if (tpte & PG_M) {
2268#if defined(PMAP_DIAGNOSTIC)
48ffc236 2269 if (pmap_nw_modified(tpte)) {
c8fe38ae 2270 kprintf(
48ffc236 2271 "pmap_remove_all: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
c8fe38ae
MD
2272 pv->pv_va, tpte);
2273 }
2274#endif
2275 if (pmap_track_modified(pv->pv_va))
2276 vm_page_dirty(m);
2277 }
2278 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2279 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2280 ++pv->pv_pmap->pm_generation;
2281 m->md.pv_list_count--;
48ffc236 2282 KKASSERT(m->md.pv_list_count >= 0);
c8fe38ae
MD
2283 if (TAILQ_EMPTY(&m->md.pv_list))
2284 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2285 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
2286 free_pv_entry(pv);
2287 }
2288 crit_exit();
2289 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
2290 pmap_inval_flush(&info);
d7f50089
YY
2291}
2292
2293/*
2294 * pmap_protect:
2295 *
2296 * Set the physical protection on the specified range of this map
2297 * as requested.
2298 *
2299 * This function may not be called from an interrupt if the map is
2300 * not the kernel_pmap.
2301 */
2302void
2303pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2304{
48ffc236
JG
2305 vm_offset_t va_next;
2306 pml4_entry_t *pml4e;
2307 pdp_entry_t *pdpe;
2308 pd_entry_t ptpaddr, *pde;
2309 pt_entry_t *pte;
c8fe38ae
MD
2310 pmap_inval_info info;
2311
48ffc236
JG
2312 /* JG review for NX */
2313
c8fe38ae
MD
2314 if (pmap == NULL)
2315 return;
2316
2317 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2318 pmap_remove(pmap, sva, eva);
2319 return;
2320 }
2321
2322 if (prot & VM_PROT_WRITE)
2323 return;
2324
2325 pmap_inval_init(&info);
2326
48ffc236 2327 for (; sva < eva; sva = va_next) {
c8fe38ae 2328
48ffc236
JG
2329 pml4e = pmap_pml4e(pmap, sva);
2330 if ((*pml4e & PG_V) == 0) {
2331 va_next = (sva + NBPML4) & ~PML4MASK;
2332 if (va_next < sva)
2333 va_next = eva;
2334 continue;
2335 }
c8fe38ae 2336
48ffc236
JG
2337 pdpe = pmap_pml4e_to_pdpe(pml4e, sva);
2338 if ((*pdpe & PG_V) == 0) {
2339 va_next = (sva + NBPDP) & ~PDPMASK;
2340 if (va_next < sva)
2341 va_next = eva;
2342 continue;
2343 }
c8fe38ae 2344
48ffc236
JG
2345 va_next = (sva + NBPDR) & ~PDRMASK;
2346 if (va_next < sva)
2347 va_next = eva;
c8fe38ae 2348
48ffc236
JG
2349 pde = pmap_pdpe_to_pde(pdpe, sva);
2350 ptpaddr = *pde;
c8fe38ae 2351
48ffc236
JG
2352 /*
2353 * Check for large page.
2354 */
2355 if ((ptpaddr & PG_PS) != 0) {
c8fe38ae 2356 pmap_inval_add(&info, pmap, -1);
48ffc236 2357 *pde &= ~(PG_M|PG_RW);
c8fe38ae
MD
2358 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
2359 continue;
2360 }
2361
2362 /*
2363 * Weed out invalid mappings. Note: we assume that the page
2364 * directory table is always allocated, and in kernel virtual.
2365 */
2366 if (ptpaddr == 0)
2367 continue;
2368
48ffc236
JG
2369 if (va_next > eva)
2370 va_next = eva;
c8fe38ae 2371
48ffc236
JG
2372 for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++,
2373 sva += PAGE_SIZE) {
2374 pt_entry_t obits, pbits;
c8fe38ae
MD
2375 vm_page_t m;
2376
2377 /*
2378 * XXX non-optimal. Note also that there can be
2379 * no pmap_inval_flush() calls until after we modify
2380 * ptbase[sindex] (or otherwise we have to do another
2381 * pmap_inval_add() call).
2382 */
48ffc236
JG
2383 pmap_inval_add(&info, pmap, sva);
2384 obits = pbits = *pte;
2385 if ((pbits & PG_V) == 0)
2386 continue;
c8fe38ae
MD
2387 if (pbits & PG_MANAGED) {
2388 m = NULL;
2389 if (pbits & PG_A) {
48ffc236 2390 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
c8fe38ae
MD
2391 vm_page_flag_set(m, PG_REFERENCED);
2392 pbits &= ~PG_A;
2393 }
2394 if (pbits & PG_M) {
48ffc236 2395 if (pmap_track_modified(sva)) {
c8fe38ae 2396 if (m == NULL)
3cfe1a9f 2397 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
c8fe38ae
MD
2398 vm_page_dirty(m);
2399 pbits &= ~PG_M;
2400 }
2401 }
2402 }
2403
2404 pbits &= ~PG_RW;
2405
48ffc236
JG
2406 if (pbits != obits) {
2407 *pte = pbits;
c8fe38ae
MD
2408 }
2409 }
2410 }
2411 pmap_inval_flush(&info);
d7f50089
YY
2412}
2413
2414/*
c8fe38ae
MD
2415 * Insert the given physical page (p) at
2416 * the specified virtual address (v) in the
2417 * target physical map with the protection requested.
d7f50089 2418 *
c8fe38ae
MD
2419 * If specified, the page will be wired down, meaning
2420 * that the related pte can not be reclaimed.
d7f50089 2421 *
c8fe38ae
MD
2422 * NB: This is the only routine which MAY NOT lazy-evaluate
2423 * or lose information. That is, this routine must actually
2424 * insert this page into the given map NOW.
d7f50089
YY
2425 */
2426void
2427pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2428 boolean_t wired)
2429{
c8fe38ae 2430 vm_paddr_t pa;
48ffc236 2431 pd_entry_t *pde;
c8fe38ae
MD
2432 pt_entry_t *pte;
2433 vm_paddr_t opa;
48ffc236 2434 pt_entry_t origpte, newpte;
c8fe38ae
MD
2435 vm_page_t mpte;
2436 pmap_inval_info info;
2437
2438 if (pmap == NULL)
2439 return;
2440
48ffc236 2441 va = trunc_page(va);
c8fe38ae
MD
2442#ifdef PMAP_DIAGNOSTIC
2443 if (va >= KvaEnd)
2444 panic("pmap_enter: toobig");
2445 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
48ffc236 2446 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va);
c8fe38ae
MD
2447#endif
2448 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2449 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
48ffc236
JG
2450#ifdef DDB
2451 db_print_backtrace();
2452#endif
c8fe38ae
MD
2453 }
2454 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2455 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
48ffc236
JG
2456#ifdef DDB
2457 db_print_backtrace();
2458#endif
c8fe38ae
MD
2459 }
2460
2461 /*
2462 * In the case that a page table page is not
2463 * resident, we are creating it here.
2464 */
48ffc236 2465 if (va < VM_MAX_USER_ADDRESS)
c8fe38ae
MD
2466 mpte = pmap_allocpte(pmap, va);
2467 else
2468 mpte = NULL;
2469
2470 pmap_inval_init(&info);
48ffc236
JG
2471 pde = pmap_pde(pmap, va);
2472 if (pde != NULL && (*pde & PG_V) != 0) {
2473 if ((*pde & PG_PS) != 0)
2474 panic("pmap_enter: attempted pmap_enter on 2MB page");
2475 pte = pmap_pde_to_pte(pde, va);
2476 } else
2477 panic("pmap_enter: invalid page directory va=%#lx", va);
2478
2479 KKASSERT(pte != NULL);
2480 pa = VM_PAGE_TO_PHYS(m);
48ffc236 2481 origpte = *pte;
c8fe38ae
MD
2482 opa = origpte & PG_FRAME;
2483
c8fe38ae
MD
2484 /*
2485 * Mapping has not changed, must be protection or wiring change.
2486 */
2487 if (origpte && (opa == pa)) {
2488 /*
2489 * Wiring change, just update stats. We don't worry about
2490 * wiring PT pages as they remain resident as long as there
2491 * are valid mappings in them. Hence, if a user page is wired,
2492 * the PT page will be also.
2493 */
2494 if (wired && ((origpte & PG_W) == 0))
2495 pmap->pm_stats.wired_count++;
2496 else if (!wired && (origpte & PG_W))
2497 pmap->pm_stats.wired_count--;
2498
2499#if defined(PMAP_DIAGNOSTIC)
48ffc236 2500 if (pmap_nw_modified(origpte)) {
c8fe38ae 2501 kprintf(
48ffc236 2502 "pmap_enter: modified page not writable: va: 0x%lx, pte: 0x%lx\n",
c8fe38ae
MD
2503 va, origpte);
2504 }
2505#endif
2506
2507 /*
2508 * Remove the extra pte reference. Note that we cannot
2509 * optimize the RO->RW case because we have adjusted the
2510 * wiring count above and may need to adjust the wiring
2511 * bits below.
2512 */
2513 if (mpte)
2514 mpte->hold_count--;
2515
2516 /*
2517 * We might be turning off write access to the page,
2518 * so we go ahead and sense modify status.
2519 */
2520 if (origpte & PG_MANAGED) {
2521 if ((origpte & PG_M) && pmap_track_modified(va)) {
2522 vm_page_t om;
2523 om = PHYS_TO_VM_PAGE(opa);
2524 vm_page_dirty(om);
2525 }
2526 pa |= PG_MANAGED;
2527 KKASSERT(m->flags & PG_MAPPED);
2528 }
2529 goto validate;
2530 }
2531 /*
2532 * Mapping has changed, invalidate old range and fall through to
2533 * handle validating new mapping.
2534 */
2535 if (opa) {
2536 int err;
2537 err = pmap_remove_pte(pmap, pte, va, &info);
2538 if (err)
48ffc236 2539 panic("pmap_enter: pte vanished, va: 0x%lx", va);
c8fe38ae
MD
2540 }
2541
2542 /*
2543 * Enter on the PV list if part of our managed memory. Note that we
2544 * raise IPL while manipulating pv_table since pmap_enter can be
2545 * called at interrupt time.
2546 */
2547 if (pmap_initialized &&
2548 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2549 pmap_insert_entry(pmap, va, mpte, m);
2550 pa |= PG_MANAGED;
2551 vm_page_flag_set(m, PG_MAPPED);
2552 }
2553
2554 /*
2555 * Increment counters
2556 */
2557 ++pmap->pm_stats.resident_count;
2558 if (wired)
2559 pmap->pm_stats.wired_count++;
2560
2561validate:
2562 /*
2563 * Now validate mapping with desired protection/wiring.
2564 */
48ffc236 2565 newpte = (pt_entry_t) (pa | pte_prot(pmap, prot) | PG_V);
c8fe38ae
MD
2566
2567 if (wired)
2568 newpte |= PG_W;
48ffc236 2569 if (va < VM_MAX_USER_ADDRESS)
c8fe38ae
MD
2570 newpte |= PG_U;
2571 if (pmap == &kernel_pmap)
2572 newpte |= pgeflag;
2573
2574 /*
2575 * if the mapping or permission bits are different, we need
2576 * to update the pte.
2577 */
2578 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2579 pmap_inval_add(&info, pmap, va);
2580 *pte = newpte | PG_A;
2581 if (newpte & PG_RW)
2582 vm_page_flag_set(m, PG_WRITEABLE);
2583 }
2584 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
2585 pmap_inval_flush(&info);
d7f50089
YY
2586}
2587
2588/*
c8fe38ae
MD
2589 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2590 * This code also assumes that the pmap has no pre-existing entry for this
2591 * VA.
d7f50089 2592 *
c8fe38ae 2593 * This code currently may only be used on user pmaps, not kernel_pmap.
d7f50089 2594 */
bfc09ba0
MD
2595static
2596void
c8fe38ae 2597pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
d7f50089 2598{
c8fe38ae
MD
2599 pt_entry_t *pte;
2600 vm_paddr_t pa;
2601 vm_page_t mpte;
2602 vm_pindex_t ptepindex;
48ffc236 2603 pd_entry_t *ptepa;
c8fe38ae
MD
2604 pmap_inval_info info;
2605
2606 pmap_inval_init(&info);
2607
2608 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2609 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
48ffc236
JG
2610#ifdef DDB
2611 db_print_backtrace();
2612#endif
c8fe38ae
MD
2613 }
2614 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2615 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
48ffc236
JG
2616#ifdef DDB
2617 db_print_backtrace();
2618#endif
c8fe38ae
MD
2619 }
2620
2621 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2622
2623 /*
2624 * Calculate the page table page (mpte), allocating it if necessary.
2625 *
2626 * A held page table page (mpte), or NULL, is passed onto the
2627 * section following.
2628 */
48ffc236 2629 if (va < VM_MAX_USER_ADDRESS) {
c8fe38ae
MD
2630 /*
2631 * Calculate pagetable page index
2632 */
48ffc236 2633 ptepindex = pmap_pde_pindex(va);
c8fe38ae
MD
2634
2635 do {
2636 /*
2637 * Get the page directory entry
2638 */
48ffc236 2639 ptepa = pmap_pde(pmap, va);
c8fe38ae
MD
2640
2641 /*
2642 * If the page table page is mapped, we just increment
2643 * the hold count, and activate it.
2644 */
48ffc236
JG
2645 if (ptepa && (*ptepa & PG_V) != 0) {
2646 if (*ptepa & PG_PS)
2647 panic("pmap_enter_quick: unexpected mapping into 2MB page");
2648// if (pmap->pm_ptphint &&
2649// (pmap->pm_ptphint->pindex == ptepindex)) {
2650// mpte = pmap->pm_ptphint;
2651// } else {
c8fe38ae
MD
2652 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2653 pmap->pm_ptphint = mpte;
48ffc236 2654// }
c8fe38ae
MD
2655 if (mpte)
2656 mpte->hold_count++;
2657 } else {
2658 mpte = _pmap_allocpte(pmap, ptepindex);
2659 }
2660 } while (mpte == NULL);
2661 } else {
2662 mpte = NULL;
2663 /* this code path is not yet used */
2664 }
2665
2666 /*
2667 * With a valid (and held) page directory page, we can just use
2668 * vtopte() to get to the pte. If the pte is already present
2669 * we do not disturb it.
2670 */
2671 pte = vtopte(va);
2672 if (*pte & PG_V) {
2673 if (mpte)
48ffc236 2674 pmap_unwire_pte_hold(pmap, va, mpte, &info);
c8fe38ae
MD
2675 pa = VM_PAGE_TO_PHYS(m);
2676 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2677 return;
2678 }
2679
2680 /*
2681 * Enter on the PV list if part of our managed memory
2682 */
2683 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2684 pmap_insert_entry(pmap, va, mpte, m);
2685 vm_page_flag_set(m, PG_MAPPED);
2686 }
2687
2688 /*
2689 * Increment counters
2690 */
2691 ++pmap->pm_stats.resident_count;
2692
2693 pa = VM_PAGE_TO_PHYS(m);
2694
2695 /*
2696 * Now validate mapping with RO protection
2697 */
2698 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2699 *pte = pa | PG_V | PG_U;
2700 else
2701 *pte = pa | PG_V | PG_U | PG_MANAGED;
2702/* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2703 pmap_inval_flush(&info);
d7f50089
YY
2704}
2705
2706/*
c8fe38ae
MD
2707 * Make a temporary mapping for a physical address. This is only intended
2708 * to be used for panic dumps.
d7f50089 2709 */
b2b3ffcd 2710/* JG Needed on x86_64? */
c8fe38ae
MD
2711void *
2712pmap_kenter_temporary(vm_paddr_t pa, int i)
d7f50089 2713{
c8fe38ae
MD
2714 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2715 return ((void *)crashdumpmap);
d7f50089
YY
2716}
2717
c8fe38ae
MD
2718#define MAX_INIT_PT (96)
2719
d7f50089
YY
2720/*
2721 * This routine preloads the ptes for a given object into the specified pmap.
2722 * This eliminates the blast of soft faults on process startup and
2723 * immediately after an mmap.
2724 */
2725static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2726
2727void
2728pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2729 vm_object_t object, vm_pindex_t pindex,
2730 vm_size_t size, int limit)
2731{
c8fe38ae
MD
2732 struct rb_vm_page_scan_info info;
2733 struct lwp *lp;
48ffc236 2734 vm_size_t psize;
c8fe38ae
MD
2735
2736 /*
2737 * We can't preinit if read access isn't set or there is no pmap
2738 * or object.
2739 */
2740 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2741 return;
2742
2743 /*
2744 * We can't preinit if the pmap is not the current pmap
2745 */
2746 lp = curthread->td_lwp;
2747 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
2748 return;
2749
b2b3ffcd 2750 psize = x86_64_btop(size);
c8fe38ae
MD
2751
2752 if ((object->type != OBJT_VNODE) ||
2753 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2754 (object->resident_page_count > MAX_INIT_PT))) {
2755 return;
2756 }
2757
2758 if (psize + pindex > object->size) {
2759 if (object->size < pindex)
2760 return;
2761 psize = object->size - pindex;
2762 }
2763
2764 if (psize == 0)
2765 return;
2766
2767 /*
2768 * Use a red-black scan to traverse the requested range and load
2769 * any valid pages found into the pmap.
2770 *
2771 * We cannot safely scan the object's memq unless we are in a
2772 * critical section since interrupts can remove pages from objects.
2773 */
2774 info.start_pindex = pindex;
2775 info.end_pindex = pindex + psize - 1;
2776 info.limit = limit;
2777 info.mpte = NULL;
2778 info.addr = addr;
2779 info.pmap = pmap;
2780
2781 crit_enter();
2782 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2783 pmap_object_init_pt_callback, &info);
2784 crit_exit();
d7f50089
YY
2785}
2786
2787static
2788int
2789pmap_object_init_pt_callback(vm_page_t p, void *data)
2790{
c8fe38ae
MD
2791 struct rb_vm_page_scan_info *info = data;
2792 vm_pindex_t rel_index;
2793 /*
2794 * don't allow an madvise to blow away our really
2795 * free pages allocating pv entries.
2796 */
2797 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2798 vmstats.v_free_count < vmstats.v_free_reserved) {
2799 return(-1);
2800 }
2801 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2802 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2803 if ((p->queue - p->pc) == PQ_CACHE)
2804 vm_page_deactivate(p);
2805 vm_page_busy(p);
2806 rel_index = p->pindex - info->start_pindex;
2807 pmap_enter_quick(info->pmap,
b2b3ffcd 2808 info->addr + x86_64_ptob(rel_index), p);
c8fe38ae
MD
2809 vm_page_wakeup(p);
2810 }
d7f50089
YY
2811 return(0);
2812}
2813
2814/*
2815 * pmap_prefault provides a quick way of clustering pagefaults into a
2816 * processes address space. It is a "cousin" of pmap_object_init_pt,
2817 * except it runs at page fault time instead of mmap time.
2818 */
2819#define PFBAK 4
2820#define PFFOR 4
2821#define PAGEORDER_SIZE (PFBAK+PFFOR)
2822
2823static int pmap_prefault_pageorder[] = {
2824 -PAGE_SIZE, PAGE_SIZE,
2825 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2826 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2827 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2828};
2829
2830void
2831pmap_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry)
2832{
c8fe38ae
MD
2833 int i;
2834 vm_offset_t starta;
2835 vm_offset_t addr;
2836 vm_pindex_t pindex;
2837 vm_page_t m;
2838 vm_object_t object;
2839 struct lwp *lp;
2840
2841 /*
2842 * We do not currently prefault mappings that use virtual page
2843 * tables. We do not prefault foreign pmaps.
2844 */
2845 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
2846 return;
2847 lp = curthread->td_lwp;
2848 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
2849 return;
2850
2851 object = entry->object.vm_object;
2852
2853 starta = addra - PFBAK * PAGE_SIZE;
2854 if (starta < entry->start)
2855 starta = entry->start;
2856 else if (starta > addra)
2857 starta = 0;
2858
2859 /*
2860 * critical section protection is required to maintain the
2861 * page/object association, interrupts can free pages and remove
2862 * them from their objects.
2863 */
2864 crit_enter();
2865 for (i = 0; i < PAGEORDER_SIZE; i++) {
2866 vm_object_t lobject;
2867 pt_entry_t *pte;
665d7d25 2868 pd_entry_t *pde;
c8fe38ae
MD
2869
2870 addr = addra + pmap_prefault_pageorder[i];
2871 if (addr > addra + (PFFOR * PAGE_SIZE))
2872 addr = 0;
2873
2874 if (addr < starta || addr >= entry->end)
2875 continue;
2876
665d7d25
MD
2877 pde = pmap_pde(pmap, addr);
2878 if (pde == NULL || *pde == 0)
c8fe38ae
MD
2879 continue;
2880
2881 pte = vtopte(addr);
2882 if (*pte)
2883 continue;
2884
2885 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2886 lobject = object;
2887
2888 for (m = vm_page_lookup(lobject, pindex);
2889 (!m && (lobject->type == OBJT_DEFAULT) &&
2890 (lobject->backing_object));
2891 lobject = lobject->backing_object
2892 ) {
2893 if (lobject->backing_object_offset & PAGE_MASK)
2894 break;
2895 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2896 m = vm_page_lookup(lobject->backing_object, pindex);
2897 }
2898
2899 /*
2900 * give-up when a page is not in memory
2901 */
2902 if (m == NULL)
2903 break;
2904
2905 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2906 (m->busy == 0) &&
2907 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2908
2909 if ((m->queue - m->pc) == PQ_CACHE) {
2910 vm_page_deactivate(m);
2911 }
2912 vm_page_busy(m);
2913 pmap_enter_quick(pmap, addr, m);
2914 vm_page_wakeup(m);
2915 }
2916 }
2917 crit_exit();
d7f50089
YY
2918}
2919
2920/*
2921 * Routine: pmap_change_wiring
2922 * Function: Change the wiring attribute for a map/virtual-address
2923 * pair.
2924 * In/out conditions:
2925 * The mapping must already exist in the pmap.
2926 */
2927void
2928pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2929{
c8fe38ae
MD
2930 pt_entry_t *pte;
2931
2932 if (pmap == NULL)
2933 return;
2934
2935 pte = pmap_pte(pmap, va);
2936
2937 if (wired && !pmap_pte_w(pte))
2938 pmap->pm_stats.wired_count++;
2939 else if (!wired && pmap_pte_w(pte))
2940 pmap->pm_stats.wired_count--;
2941
2942 /*
2943 * Wiring is not a hardware characteristic so there is no need to
2944 * invalidate TLB. However, in an SMP environment we must use
2945 * a locked bus cycle to update the pte (if we are not using
2946 * the pmap_inval_*() API that is)... it's ok to do this for simple
2947 * wiring changes.
2948 */
2949#ifdef SMP
2950 if (wired)
71577ce5 2951 atomic_set_long(pte, PG_W);
c8fe38ae 2952 else
71577ce5 2953 atomic_clear_long(pte, PG_W);
c8fe38ae
MD
2954#else
2955 if (wired)
71577ce5 2956 atomic_set_long_nonlocked(pte, PG_W);
c8fe38ae 2957 else
71577ce5 2958 atomic_clear_long_nonlocked(pte, PG_W);
c8fe38ae 2959#endif
d7f50089
YY
2960}
2961
c8fe38ae
MD
2962
2963
d7f50089
YY
2964/*
2965 * Copy the range specified by src_addr/len
2966 * from the source map to the range dst_addr/len
2967 * in the destination map.
2968 *
2969 * This routine is only advisory and need not do anything.
2970 */
2971void
2972pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
bfc09ba0 2973 vm_size_t len, vm_offset_t src_addr)
d7f50089 2974{
bfc09ba0
MD
2975 return;
2976#if 0
c8fe38ae
MD
2977 pmap_inval_info info;
2978 vm_offset_t addr;
2979 vm_offset_t end_addr = src_addr + len;
2980 vm_offset_t pdnxt;
2981 pd_entry_t src_frame, dst_frame;
2982 vm_page_t m;
2983
2984 if (dst_addr != src_addr)
2985 return;
48ffc236 2986#if JGPMAP32
c8fe38ae
MD
2987 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
2988 if (src_frame != (PTDpde & PG_FRAME)) {
2989 return;
2990 }
2991
2992 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
2993 if (dst_frame != (APTDpde & PG_FRAME)) {
2994 APTDpde = (pd_entry_t) (dst_frame | PG_RW | PG_V);
2995 /* The page directory is not shared between CPUs */
2996 cpu_invltlb();
2997 }
48ffc236 2998#endif
c8fe38ae
MD
2999 pmap_inval_init(&info);
3000 pmap_inval_add(&info, dst_pmap, -1);
3001 pmap_inval_add(&info, src_pmap, -1);
3002
3003 /*
3004 * critical section protection is required to maintain the page/object
3005 * association, interrupts can free pages and remove them from
3006 * their objects.
3007 */
3008 crit_enter();
3009 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
3010 pt_entry_t *src_pte, *dst_pte;
3011 vm_page_t dstmpte, srcmpte;
3012 vm_offset_t srcptepaddr;
3013 vm_pindex_t ptepindex;
3014
3015 if (addr >= UPT_MIN_ADDRESS)
3016 panic("pmap_copy: invalid to pmap_copy page tables\n");
3017
3018 /*
3019 * Don't let optional prefaulting of pages make us go
3020 * way below the low water mark of free pages or way
3021 * above high water mark of used pv entries.
3022 */
3023 if (vmstats.v_free_count < vmstats.v_free_reserved ||
3024 pv_entry_count > pv_entry_high_water)
3025 break;
3026
3027 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
3028 ptepindex = addr >> PDRSHIFT;
3029
48ffc236 3030#if JGPMAP32
c8fe38ae 3031 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
48ffc236 3032#endif
c8fe38ae
MD
3033 if (srcptepaddr == 0)
3034 continue;
3035
3036 if (srcptepaddr & PG_PS) {
48ffc236 3037#if JGPMAP32
c8fe38ae
MD
3038 if (dst_pmap->pm_pdir[ptepindex] == 0) {
3039 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
3040 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
3041 }
48ffc236 3042#endif
c8fe38ae
MD
3043 continue;
3044 }
3045
3046 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
3047 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
3048 (srcmpte->flags & PG_BUSY)) {
3049 continue;
3050 }
3051
3052 if (pdnxt > end_addr)
3053 pdnxt = end_addr;
3054
3055 src_pte = vtopte(addr);
48ffc236 3056#if JGPMAP32
c8fe38ae 3057 dst_pte = avtopte(addr);
48ffc236 3058#endif
c8fe38ae
MD
3059 while (addr < pdnxt) {
3060 pt_entry_t ptetemp;
3061
3062 ptetemp = *src_pte;
3063 /*
3064 * we only virtual copy managed pages
3065 */
3066 if ((ptetemp & PG_MANAGED) != 0) {
3067 /*
3068 * We have to check after allocpte for the
3069 * pte still being around... allocpte can
3070 * block.
3071 *
3072 * pmap_allocpte() can block. If we lose
3073 * our page directory mappings we stop.
3074 */
3075 dstmpte = pmap_allocpte(dst_pmap, addr);
3076
48ffc236 3077#if JGPMAP32
c8fe38ae
MD
3078 if (src_frame != (PTDpde & PG_FRAME) ||
3079 dst_frame != (APTDpde & PG_FRAME)
3080 ) {
3081 kprintf("WARNING: pmap_copy: detected and corrected race\n");
3082 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3083 goto failed;
3084 } else if ((*dst_pte == 0) &&
3085 (ptetemp = *src_pte) != 0 &&
3086 (ptetemp & PG_MANAGED)) {
3087 /*
3088 * Clear the modified and
3089 * accessed (referenced) bits
3090 * during the copy.
3091 */
3092 m = PHYS_TO_VM_PAGE(ptetemp);
3093 *dst_pte = ptetemp & ~(PG_M | PG_A);
3094 ++dst_pmap->pm_stats.resident_count;
3095 pmap_insert_entry(dst_pmap, addr,
3096 dstmpte, m);
3097 KKASSERT(m->flags & PG_MAPPED);
3098 } else {
3099 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
3100 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
3101 goto failed;
3102 }
48ffc236 3103#endif
c8fe38ae
MD
3104 if (dstmpte->hold_count >= srcmpte->hold_count)
3105 break;
3106 }
3107 addr += PAGE_SIZE;
3108 src_pte++;
3109 dst_pte++;
3110 }
3111 }
3112failed:
3113 crit_exit();
3114 pmap_inval_flush(&info);
f81851b8 3115#endif
d7f50089
YY
3116}
3117
3118/*
3119 * pmap_zero_page:
3120 *
48ffc236 3121 * Zero the specified physical page.
d7f50089
YY
3122 *
3123 * This function may be called from an interrupt and no locking is
3124 * required.
3125 */
3126void
3127pmap_zero_page(vm_paddr_t phys)
3128{
48ffc236 3129 vm_offset_t va = PHYS_TO_DMAP(phys);
c8fe38ae 3130
48ffc236 3131 pagezero((void *)va);
d7f50089
YY
3132}
3133
3134/*
3135 * pmap_page_assertzero:
3136 *
3137 * Assert that a page is empty, panic if it isn't.
3138 */
3139void
3140pmap_page_assertzero(vm_paddr_t phys)
3141{
48ffc236 3142 vm_offset_t virt = PHYS_TO_DMAP(phys);
bfc09ba0 3143 int i;
48ffc236 3144
bfc09ba0
MD
3145 for (i = 0; i < PAGE_SIZE; i += sizeof(long)) {
3146 if (*(long *)((char *)virt + i) != 0) {
3147 panic("pmap_page_assertzero() @ %p not zero!\n", (void *)virt);
c8fe38ae
MD
3148 }
3149 }
d7f50089
YY
3150}
3151
3152/*
3153 * pmap_zero_page:
3154 *
3155 * Zero part of a physical page by mapping it into memory and clearing
3156 * its contents with bzero.
3157 *
3158 * off and size may not cover an area beyond a single hardware page.
3159 */
3160void
3161pmap_zero_page_area(vm_paddr_t phys, int off, int size)
3162{
48ffc236 3163 vm_offset_t virt = PHYS_TO_DMAP(phys);
bfc09ba0 3164
48ffc236 3165 bzero((char *)virt + off, size);
d7f50089
YY
3166}
3167
3168/*
3169 * pmap_copy_page:
3170 *
3171 * Copy the physical page from the source PA to the target PA.
3172 * This function may be called from an interrupt. No locking
3173 * is required.
3174 */
3175void
3176pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
3177{
48ffc236 3178 vm_offset_t src_virt, dst_virt;
c8fe38ae 3179
48ffc236
JG
3180 src_virt = PHYS_TO_DMAP(src);
3181 dst_virt = PHYS_TO_DMAP(dst);
bfc09ba0 3182 bcopy((void *)src_virt, (void *)dst_virt, PAGE_SIZE);
d7f50089
YY
3183}
3184
3185/*
3186 * pmap_copy_page_frag:
3187 *
3188 * Copy the physical page from the source PA to the target PA.
3189 * This function may be called from an interrupt. No locking
3190 * is required.
3191 */
3192void
3193pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
3194{
48ffc236 3195 vm_offset_t src_virt, dst_virt;
c8fe38ae 3196
48ffc236
JG
3197 src_virt = PHYS_TO_DMAP(src);
3198 dst_virt = PHYS_TO_DMAP(dst);
bfc09ba0 3199
48ffc236
JG
3200 bcopy((char *)src_virt + (src & PAGE_MASK),
3201 (char *)dst_virt + (dst & PAGE_MASK),
c8fe38ae 3202 bytes);
d7f50089
YY
3203}
3204
3205/*
3206 * Returns true if the pmap's pv is one of the first
3207 * 16 pvs linked to from this page. This count may
3208 * be changed upwards or downwards in the future; it
3209 * is only necessary that true be returned for a small
3210 * subset of pmaps for proper page aging.
3211 */
3212boolean_t
3213pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
3214{
c8fe38ae
MD
3215 pv_entry_t pv;
3216 int loops = 0;
3217
3218 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3219 return FALSE;
3220
3221 crit_enter();
3222
3223 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3224 if (pv->pv_pmap == pmap) {
3225 crit_exit();
3226 return TRUE;
3227 }
3228 loops++;
3229 if (loops >= 16)
3230 break;
3231 }
3232 crit_exit();
d7f50089
YY
3233 return (FALSE);
3234}
3235
3236/*
3237 * Remove all pages from specified address space
3238 * this aids process exit speeds. Also, this code
3239 * is special cased for current process only, but
3240 * can have the more generic (and slightly slower)
3241 * mode enabled. This is much faster than pmap_remove
3242 * in the case of running down an entire address space.
3243 */
3244void
3245pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
3246{
c8fe38ae
MD
3247 struct lwp *lp;
3248 pt_entry_t *pte, tpte;
3249 pv_entry_t pv, npv;
3250 vm_page_t m;
3251 pmap_inval_info info;
3252 int iscurrentpmap;
48ffc236 3253 int save_generation;
c8fe38ae
MD
3254
3255 lp = curthread->td_lwp;
3256 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
3257 iscurrentpmap = 1;
3258 else
3259 iscurrentpmap = 0;
3260
3261 pmap_inval_init(&info);
3262 crit_enter();
3263 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
3264 if (pv->pv_va >= eva || pv->pv_va < sva) {
3265 npv = TAILQ_NEXT(pv, pv_plist);
3266 continue;
3267 }
3268
3269 KKASSERT(pmap == pv->pv_pmap);
3270
3271 if (iscurrentpmap)
3272 pte = vtopte(pv->pv_va);
3273 else
3274 pte = pmap_pte_quick(pmap, pv->pv_va);
3275 if (pmap->pm_active)
3276 pmap_inval_add(&info, pmap, pv->pv_va);
3277
3278 /*
3279 * We cannot remove wired pages from a process' mapping
3280 * at this time
3281 */
3282 if (*pte & PG_W) {
3283 npv = TAILQ_NEXT(pv, pv_plist);
3284 continue;
3285 }
3286 tpte = pte_load_clear(pte);
3287
48ffc236 3288 m = PHYS_TO_VM_PAGE(tpte & PG_FRAME);
c8fe38ae
MD
3289
3290 KASSERT(m < &vm_page_array[vm_page_array_size],
48ffc236 3291 ("pmap_remove_pages: bad tpte %lx", tpte));
c8fe38ae
MD
3292
3293 KKASSERT(pmap->pm_stats.resident_count > 0);
3294 --pmap->pm_stats.resident_count;
3295
3296 /*
3297 * Update the vm_page_t clean and reference bits.
3298 */
3299 if (tpte & PG_M) {
3300 vm_page_dirty(m);
3301 }
3302
3303 npv = TAILQ_NEXT(pv, pv_plist);
3304 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3305 save_generation = ++pmap->pm_generation;
3306
3307 m->md.pv_list_count--;
3308 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3309 if (TAILQ_EMPTY(&m->md.pv_list))
3310 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3311
3312 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
3313 free_pv_entry(pv);
3314
3315 /*
3316 * Restart the scan if we blocked during the unuse or free
3317 * calls and other removals were made.
3318 */
3319 if (save_generation != pmap->pm_generation) {
3320 kprintf("Warning: pmap_remove_pages race-A avoided\n");
3321 pv = TAILQ_FIRST(&pmap->pm_pvlist);
3322 }
3323 }
3324 pmap_inval_flush(&info);
3325 crit_exit();
d7f50089
YY
3326}
3327
3328/*
c8fe38ae
MD
3329 * pmap_testbit tests bits in pte's
3330 * note that the testbit/clearbit routines are inline,
3331 * and a lot of things compile-time evaluate.
d7f50089 3332 */
bfc09ba0
MD
3333static
3334boolean_t
d7f50089
YY
3335pmap_testbit(vm_page_t m, int bit)
3336{
c8fe38ae
MD
3337 pv_entry_t pv;
3338 pt_entry_t *pte;
3339
3340 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3341 return FALSE;
3342
3343 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3344 return FALSE;
3345
3346 crit_enter();
3347
3348 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3349 /*
3350 * if the bit being tested is the modified bit, then
3351 * mark clean_map and ptes as never
3352 * modified.
3353 */
3354 if (bit & (PG_A|PG_M)) {
3355 if (!pmap_track_modified(pv->pv_va))
3356 continue;
3357 }
3358
3359#if defined(PMAP_DIAGNOSTIC)
48ffc236
JG
3360 if (pv->pv_pmap == NULL) {
3361 kprintf("Null pmap (tb) at va: 0x%lx\n", pv->pv_va);
c8fe38ae
MD
3362 continue;
3363 }
3364#endif
3365 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3366 if (*pte & bit) {
3367 crit_exit();
3368 return TRUE;
3369 }
3370 }
3371 crit_exit();
d7f50089
YY
3372 return (FALSE);
3373}
3374
3375/*
c8fe38ae 3376 * this routine is used to modify bits in ptes
d7f50089 3377 */
bfc09ba0
MD
3378static __inline
3379void
d7f50089
YY
3380pmap_clearbit(vm_page_t m, int bit)
3381{
c8fe38ae
MD
3382 struct pmap_inval_info info;
3383 pv_entry_t pv;
3384 pt_entry_t *pte;
3385 pt_entry_t pbits;
3386
3387 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3388 return;
3389
3390 pmap_inval_init(&info);
3391 crit_enter();
3392
3393 /*
3394 * Loop over all current mappings setting/clearing as appropos If
3395 * setting RO do we need to clear the VAC?
3396 */
3397 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3398 /*
3399 * don't write protect pager mappings
3400 */
3401 if (bit == PG_RW) {
3402 if (!pmap_track_modified(pv->pv_va))
3403 continue;
3404 }
3405
3406#if defined(PMAP_DIAGNOSTIC)
48ffc236
JG
3407 if (pv->pv_pmap == NULL) {
3408 kprintf("Null pmap (cb) at va: 0x%lx\n", pv->pv_va);
c8fe38ae
MD
3409 continue;
3410 }
3411#endif
3412
3413 /*
3414 * Careful here. We can use a locked bus instruction to
3415 * clear PG_A or PG_M safely but we need to synchronize
3416 * with the target cpus when we mess with PG_RW.
3417 *
3418 * We do not have to force synchronization when clearing
3419 * PG_M even for PTEs generated via virtual memory maps,
3420 * because the virtual kernel will invalidate the pmap
3421 * entry when/if it needs to resynchronize the Modify bit.
3422 */
3423 if (bit & PG_RW)
3424 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
3425 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3426again:
3427 pbits = *pte;
3428 if (pbits & bit) {
3429 if (bit == PG_RW) {
3430 if (pbits & PG_M) {
3431 vm_page_dirty(m);
48ffc236 3432 atomic_clear_long(pte, PG_M|PG_RW);
c8fe38ae
MD
3433 } else {
3434 /*
3435 * The cpu may be trying to set PG_M
3436 * simultaniously with our clearing
3437 * of PG_RW.
3438 */
48ffc236 3439 if (!atomic_cmpset_long(pte, pbits,
c8fe38ae
MD
3440 pbits & ~PG_RW))
3441 goto again;
3442 }
3443 } else if (bit == PG_M) {
3444 /*
3445 * We could also clear PG_RW here to force
3446 * a fault on write to redetect PG_M for
3447 * virtual kernels, but it isn't necessary
3448 * since virtual kernels invalidate the pte
3449 * when they clear the VPTE_M bit in their
3450 * virtual page tables.
3451 */
48ffc236 3452 atomic_clear_long(pte, PG_M);
c8fe38ae 3453 } else {
48ffc236 3454 atomic_clear_long(pte, bit);
c8fe38ae
MD
3455 }
3456 }
3457 }
3458 pmap_inval_flush(&info);
3459 crit_exit();
d7f50089
YY
3460}
3461
3462/*
3463 * pmap_page_protect:
3464 *
3465 * Lower the permission for all mappings to a given page.
3466 */
3467void
3468pmap_page_protect(vm_page_t m, vm_prot_t prot)
3469{
48ffc236 3470 /* JG NX support? */
c8fe38ae
MD
3471 if ((prot & VM_PROT_WRITE) == 0) {
3472 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3473 pmap_clearbit(m, PG_RW);
3474 vm_page_flag_clear(m, PG_WRITEABLE);
3475 } else {
3476 pmap_remove_all(m);
3477 }
3478 }
d7f50089
YY
3479}
3480
3481vm_paddr_t
c8fe38ae 3482pmap_phys_address(vm_pindex_t ppn)
d7f50089 3483{
b2b3ffcd 3484 return (x86_64_ptob(ppn));
d7f50089
YY
3485}
3486
3487/*
3488 * pmap_ts_referenced:
3489 *
3490 * Return a count of reference bits for a page, clearing those bits.
3491 * It is not necessary for every reference bit to be cleared, but it
3492 * is necessary that 0 only be returned when there are truly no
3493 * reference bits set.
3494 *
3495 * XXX: The exact number of bits to check and clear is a matter that
3496 * should be tested and standardized at some point in the future for
3497 * optimal aging of shared pages.
3498 */
3499int
3500pmap_ts_referenced(vm_page_t m)
3501{
c8fe38ae
MD
3502 pv_entry_t pv, pvf, pvn;
3503 pt_entry_t *pte;
3504 int rtval = 0;
3505
3506 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3507 return (rtval);
3508
3509 crit_enter();
3510
3511 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3512
3513 pvf = pv;
3514
3515 do {
3516 pvn = TAILQ_NEXT(pv, pv_list);
3517
3518 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3519
3520 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3521
3522 if (!pmap_track_modified(pv->pv_va))
3523 continue;
3524
3525 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3526
3527 if (pte && (*pte & PG_A)) {
3528#ifdef SMP
48ffc236 3529 atomic_clear_long(pte, PG_A);
c8fe38ae 3530#else
48ffc236 3531 atomic_clear_long_nonlocked(pte, PG_A);
c8fe38ae
MD
3532#endif
3533 rtval++;
3534 if (rtval > 4) {
3535 break;
3536 }
3537 }
3538 } while ((pv = pvn) != NULL && pv != pvf);
3539 }
3540 crit_exit();
3541
3542 return (rtval);
d7f50089
YY
3543}
3544
3545/*
3546 * pmap_is_modified:
3547 *
3548 * Return whether or not the specified physical page was modified
3549 * in any physical maps.
3550 */
3551boolean_t
3552pmap_is_modified(vm_page_t m)
3553{
c8fe38ae 3554 return pmap_testbit(m, PG_M);
d7f50089
YY
3555}
3556
3557/*
3558 * Clear the modify bits on the specified physical page.
3559 */
3560void
3561pmap_clear_modify(vm_page_t m)
3562{
c8fe38ae 3563 pmap_clearbit(m, PG_M);
d7f50089
YY
3564}
3565
3566/*
3567 * pmap_clear_reference:
3568 *
3569 * Clear the reference bit on the specified physical page.