kernel - Implement segment pmap optimizations for x86-64 (4)
[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.
921c891e 9 * Copyright (c) 2011-2012 Matthew Dillon
d7f50089 10 * All rights reserved.
c8fe38ae
MD
11 *
12 * This code is derived from software contributed to Berkeley by
13 * the Systems Programming Group of the University of Utah Computer
14 * Science Department and William Jolitz of UUNET Technologies Inc.
15 *
d7f50089
YY
16 * Redistribution and use in source and binary forms, with or without
17 * modification, are permitted provided that the following conditions
18 * are met:
d7f50089
YY
19 * 1. Redistributions of source code must retain the above copyright
20 * notice, this list of conditions and the following disclaimer.
21 * 2. Redistributions in binary form must reproduce the above copyright
c8fe38ae
MD
22 * notice, this list of conditions and the following disclaimer in the
23 * documentation and/or other materials provided with the distribution.
24 * 3. All advertising materials mentioning features or use of this software
25 * must display the following acknowledgement:
26 * This product includes software developed by the University of
27 * California, Berkeley and its contributors.
28 * 4. Neither the name of the University nor the names of its contributors
29 * may be used to endorse or promote products derived from this software
30 * without specific prior written permission.
31 *
32 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
33 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
34 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
35 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
36 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
37 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
38 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
39 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
40 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
41 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
d7f50089 42 * SUCH DAMAGE.
d7f50089
YY
43 */
44/*
90244566 45 * Manage physical address maps for x86-64 systems.
c8fe38ae
MD
46 */
47
48#if JG
49#include "opt_disable_pse.h"
50#include "opt_pmap.h"
51#endif
52#include "opt_msgbuf.h"
d7f50089 53
c8fe38ae 54#include <sys/param.h>
d7f50089
YY
55#include <sys/systm.h>
56#include <sys/kernel.h>
d7f50089 57#include <sys/proc.h>
c8fe38ae
MD
58#include <sys/msgbuf.h>
59#include <sys/vmmeter.h>
60#include <sys/mman.h>
d7f50089 61
c8fe38ae
MD
62#include <vm/vm.h>
63#include <vm/vm_param.h>
64#include <sys/sysctl.h>
65#include <sys/lock.h>
d7f50089 66#include <vm/vm_kern.h>
c8fe38ae
MD
67#include <vm/vm_page.h>
68#include <vm/vm_map.h>
d7f50089 69#include <vm/vm_object.h>
c8fe38ae 70#include <vm/vm_extern.h>
d7f50089 71#include <vm/vm_pageout.h>
c8fe38ae
MD
72#include <vm/vm_pager.h>
73#include <vm/vm_zone.h>
74
75#include <sys/user.h>
76#include <sys/thread2.h>
77#include <sys/sysref2.h>
b12defdc
MD
78#include <sys/spinlock2.h>
79#include <vm/vm_page2.h>
d7f50089 80
c8fe38ae 81#include <machine/cputypes.h>
d7f50089 82#include <machine/md_var.h>
c8fe38ae
MD
83#include <machine/specialreg.h>
84#include <machine/smp.h>
85#include <machine_base/apic/apicreg.h>
d7f50089 86#include <machine/globaldata.h>
c8fe38ae
MD
87#include <machine/pmap.h>
88#include <machine/pmap_inval.h>
7e9313e0 89#include <machine/inttypes.h>
c8fe38ae 90
48ffc236
JG
91#include <ddb/ddb.h>
92
c8fe38ae
MD
93#define PMAP_KEEP_PDIRS
94#ifndef PMAP_SHPGPERPROC
f1d3f422 95#define PMAP_SHPGPERPROC 2000
c8fe38ae
MD
96#endif
97
98#if defined(DIAGNOSTIC)
99#define PMAP_DIAGNOSTIC
100#endif
101
102#define MINPV 2048
103
c8fe38ae 104/*
701c977e
MD
105 * pmap debugging will report who owns a pv lock when blocking.
106 */
107#ifdef PMAP_DEBUG
108
109#define PMAP_DEBUG_DECL ,const char *func, int lineno
110#define PMAP_DEBUG_ARGS , __func__, __LINE__
111#define PMAP_DEBUG_COPY , func, lineno
112
113#define pv_get(pmap, pindex) _pv_get(pmap, pindex \
114 PMAP_DEBUG_ARGS)
115#define pv_lock(pv) _pv_lock(pv \
116 PMAP_DEBUG_ARGS)
117#define pv_hold_try(pv) _pv_hold_try(pv \
118 PMAP_DEBUG_ARGS)
119#define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp \
120 PMAP_DEBUG_ARGS)
121
122#else
123
124#define PMAP_DEBUG_DECL
125#define PMAP_DEBUG_ARGS
126#define PMAP_DEBUG_COPY
127
128#define pv_get(pmap, pindex) _pv_get(pmap, pindex)
129#define pv_lock(pv) _pv_lock(pv)
130#define pv_hold_try(pv) _pv_hold_try(pv)
131#define pv_alloc(pmap, pindex, isnewp) _pv_alloc(pmap, pindex, isnewp)
132
133#endif
134
135/*
c8fe38ae
MD
136 * Get PDEs and PTEs for user/kernel address space
137 */
c8fe38ae
MD
138#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
139
140#define pmap_pde_v(pte) ((*(pd_entry_t *)pte & PG_V) != 0)
141#define pmap_pte_w(pte) ((*(pt_entry_t *)pte & PG_W) != 0)
142#define pmap_pte_m(pte) ((*(pt_entry_t *)pte & PG_M) != 0)
143#define pmap_pte_u(pte) ((*(pt_entry_t *)pte & PG_A) != 0)
144#define pmap_pte_v(pte) ((*(pt_entry_t *)pte & PG_V) != 0)
145
c8fe38ae
MD
146/*
147 * Given a map and a machine independent protection code,
148 * convert to a vax protection code.
149 */
150#define pte_prot(m, p) \
151 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
152static int protection_codes[8];
d7f50089
YY
153
154struct pmap kernel_pmap;
c8fe38ae 155static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
d7f50089 156
921c891e
MD
157MALLOC_DEFINE(M_OBJPMAP, "objpmap", "pmaps associated with VM objects");
158
c8fe38ae
MD
159vm_paddr_t avail_start; /* PA of first available physical page */
160vm_paddr_t avail_end; /* PA of last available physical page */
791c6551
MD
161vm_offset_t virtual2_start; /* cutout free area prior to kernel start */
162vm_offset_t virtual2_end;
c8fe38ae
MD
163vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
164vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
165vm_offset_t KvaStart; /* VA start of KVA space */
166vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
167vm_offset_t KvaSize; /* max size of kernel virtual address space */
168static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
169static int pgeflag; /* PG_G or-in */
170static int pseflag; /* PG_PS or-in */
d7f50089 171
48ffc236
JG
172static int ndmpdp;
173static vm_paddr_t dmaplimit;
c8fe38ae 174static int nkpt;
791c6551 175vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
d7f50089 176
791c6551
MD
177static uint64_t KPTbase;
178static uint64_t KPTphys;
48ffc236 179static uint64_t KPDphys; /* phys addr of kernel level 2 */
791c6551
MD
180static uint64_t KPDbase; /* phys addr of kernel level 2 @ KERNBASE */
181uint64_t KPDPphys; /* phys addr of kernel level 3 */
182uint64_t KPML4phys; /* phys addr of kernel level 4 */
48ffc236
JG
183
184static uint64_t DMPDphys; /* phys addr of direct mapped level 2 */
185static uint64_t DMPDPphys; /* phys addr of direct mapped level 3 */
186
d7f50089 187/*
c8fe38ae 188 * Data for the pv entry allocation mechanism
d7f50089 189 */
c8fe38ae
MD
190static vm_zone_t pvzone;
191static struct vm_zone pvzone_store;
192static struct vm_object pvzone_obj;
701c977e 193static int pv_entry_max=0, pv_entry_high_water=0;
c8fe38ae
MD
194static int pmap_pagedaemon_waken = 0;
195static struct pv_entry *pvinit;
d7f50089
YY
196
197/*
c8fe38ae 198 * All those kernel PT submaps that BSD is so fond of
d7f50089 199 */
4090d6ff 200pt_entry_t *CMAP1 = NULL, *ptmmap;
4c0cc8bb 201caddr_t CADDR1 = NULL, ptvmmap = NULL;
c8fe38ae 202static pt_entry_t *msgbufmap;
4090d6ff 203struct msgbuf *msgbufp=NULL;
d7f50089 204
c8fe38ae
MD
205/*
206 * Crashdump maps.
d7f50089 207 */
c8fe38ae
MD
208static pt_entry_t *pt_crashdumpmap;
209static caddr_t crashdumpmap;
210
b12defdc
MD
211static int pmap_yield_count = 64;
212SYSCTL_INT(_machdep, OID_AUTO, pmap_yield_count, CTLFLAG_RW,
213 &pmap_yield_count, 0, "Yield during init_pt/release");
921c891e
MD
214static int pmap_mmu_optimize = 0;
215SYSCTL_INT(_machdep, OID_AUTO, pmap_mmu_optimize, CTLFLAG_RW,
216 &pmap_mmu_optimize, 0, "Share page table pages when possible");
b12defdc 217
c8fe38ae
MD
218#define DISABLE_PSE
219
701c977e
MD
220static void pv_hold(pv_entry_t pv);
221static int _pv_hold_try(pv_entry_t pv
222 PMAP_DEBUG_DECL);
223static void pv_drop(pv_entry_t pv);
224static void _pv_lock(pv_entry_t pv
225 PMAP_DEBUG_DECL);
226static void pv_unlock(pv_entry_t pv);
227static pv_entry_t _pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew
228 PMAP_DEBUG_DECL);
229static pv_entry_t _pv_get(pmap_t pmap, vm_pindex_t pindex
230 PMAP_DEBUG_DECL);
231static pv_entry_t pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp);
232static pv_entry_t pv_find(pmap_t pmap, vm_pindex_t pindex);
233static void pv_put(pv_entry_t pv);
234static void pv_free(pv_entry_t pv);
235static void *pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex);
236static pv_entry_t pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex,
237 pv_entry_t *pvpp);
921c891e
MD
238static pv_entry_t pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex,
239 pv_entry_t *pvpp, vm_map_entry_t entry, vm_offset_t va);
701c977e
MD
240static void pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp,
241 struct pmap_inval_info *info);
52bb73bc 242static vm_page_t pmap_remove_pv_page(pv_entry_t pv);
921c891e 243static int pmap_release_pv(pv_entry_t pv);
701c977e
MD
244
245static void pmap_remove_callback(pmap_t pmap, struct pmap_inval_info *info,
921c891e
MD
246 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
247 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
701c977e 248static void pmap_protect_callback(pmap_t pmap, struct pmap_inval_info *info,
921c891e
MD
249 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
250 vm_offset_t va, pt_entry_t *ptep, void *arg __unused);
701c977e 251
bfc09ba0
MD
252static void i386_protection_init (void);
253static void create_pagetables(vm_paddr_t *firstaddr);
254static void pmap_remove_all (vm_page_t m);
c8fe38ae 255static boolean_t pmap_testbit (vm_page_t m, int bit);
c8fe38ae 256
c8fe38ae 257static pt_entry_t * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
c8fe38ae
MD
258static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
259
260static unsigned pdir4mb;
d7f50089 261
701c977e
MD
262static int
263pv_entry_compare(pv_entry_t pv1, pv_entry_t pv2)
264{
265 if (pv1->pv_pindex < pv2->pv_pindex)
266 return(-1);
267 if (pv1->pv_pindex > pv2->pv_pindex)
268 return(1);
269 return(0);
270}
271
272RB_GENERATE2(pv_entry_rb_tree, pv_entry, pv_entry,
273 pv_entry_compare, vm_pindex_t, pv_pindex);
274
d7f50089 275/*
c8fe38ae 276 * Move the kernel virtual free pointer to the next
f9cc0f15
JG
277 * 2MB. This is used to help improve performance
278 * by using a large (2MB) page for much of the kernel
c8fe38ae 279 * (.text, .data, .bss)
d7f50089 280 */
bfc09ba0
MD
281static
282vm_offset_t
c8fe38ae 283pmap_kmem_choose(vm_offset_t addr)
d7f50089 284{
c8fe38ae 285 vm_offset_t newaddr = addr;
f9cc0f15
JG
286
287 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
c8fe38ae 288 return newaddr;
d7f50089
YY
289}
290
d7f50089 291/*
c8fe38ae 292 * pmap_pte_quick:
d7f50089 293 *
c8fe38ae
MD
294 * Super fast pmap_pte routine best used when scanning the pv lists.
295 * This eliminates many course-grained invltlb calls. Note that many of
296 * the pv list scans are across different pmaps and it is very wasteful
297 * to do an entire invltlb when checking a single mapping.
c8fe38ae 298 */
48ffc236
JG
299static __inline pt_entry_t *pmap_pte(pmap_t pmap, vm_offset_t va);
300
bfc09ba0
MD
301static
302pt_entry_t *
c8fe38ae
MD
303pmap_pte_quick(pmap_t pmap, vm_offset_t va)
304{
48ffc236
JG
305 return pmap_pte(pmap, va);
306}
307
701c977e
MD
308/*
309 * Returns the pindex of a page table entry (representing a terminal page).
310 * There are NUPTE_TOTAL page table entries possible (a huge number)
311 *
312 * x86-64 has a 48-bit address space, where bit 47 is sign-extended out.
313 * We want to properly translate negative KVAs.
314 */
bfc09ba0
MD
315static __inline
316vm_pindex_t
701c977e 317pmap_pte_pindex(vm_offset_t va)
48ffc236 318{
701c977e 319 return ((va >> PAGE_SHIFT) & (NUPTE_TOTAL - 1));
48ffc236
JG
320}
321
701c977e
MD
322/*
323 * Returns the pindex of a page table.
324 */
bfc09ba0
MD
325static __inline
326vm_pindex_t
701c977e 327pmap_pt_pindex(vm_offset_t va)
48ffc236 328{
701c977e
MD
329 return (NUPTE_TOTAL + ((va >> PDRSHIFT) & (NUPT_TOTAL - 1)));
330}
48ffc236 331
701c977e
MD
332/*
333 * Returns the pindex of a page directory.
334 */
335static __inline
336vm_pindex_t
337pmap_pd_pindex(vm_offset_t va)
338{
339 return (NUPTE_TOTAL + NUPT_TOTAL +
340 ((va >> PDPSHIFT) & (NUPD_TOTAL - 1)));
48ffc236
JG
341}
342
bfc09ba0
MD
343static __inline
344vm_pindex_t
701c977e 345pmap_pdp_pindex(vm_offset_t va)
48ffc236 346{
701c977e
MD
347 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
348 ((va >> PML4SHIFT) & (NUPDP_TOTAL - 1)));
349}
48ffc236 350
701c977e
MD
351static __inline
352vm_pindex_t
353pmap_pml4_pindex(void)
354{
355 return (NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL + NUPDP_TOTAL);
48ffc236
JG
356}
357
701c977e
MD
358/*
359 * Return various clipped indexes for a given VA
360 *
361 * Returns the index of a pte in a page table, representing a terminal
362 * page.
363 */
bfc09ba0
MD
364static __inline
365vm_pindex_t
701c977e 366pmap_pte_index(vm_offset_t va)
48ffc236 367{
701c977e
MD
368 return ((va >> PAGE_SHIFT) & ((1ul << NPTEPGSHIFT) - 1));
369}
48ffc236 370
701c977e
MD
371/*
372 * Returns the index of a pt in a page directory, representing a page
373 * table.
374 */
375static __inline
376vm_pindex_t
377pmap_pt_index(vm_offset_t va)
378{
379 return ((va >> PDRSHIFT) & ((1ul << NPDEPGSHIFT) - 1));
48ffc236
JG
380}
381
701c977e
MD
382/*
383 * Returns the index of a pd in a page directory page, representing a page
384 * directory.
385 */
bfc09ba0
MD
386static __inline
387vm_pindex_t
701c977e 388pmap_pd_index(vm_offset_t va)
48ffc236 389{
701c977e
MD
390 return ((va >> PDPSHIFT) & ((1ul << NPDPEPGSHIFT) - 1));
391}
48ffc236 392
701c977e
MD
393/*
394 * Returns the index of a pdp in the pml4 table, representing a page
395 * directory page.
396 */
397static __inline
398vm_pindex_t
399pmap_pdp_index(vm_offset_t va)
400{
48ffc236
JG
401 return ((va >> PML4SHIFT) & ((1ul << NPML4EPGSHIFT) - 1));
402}
403
701c977e
MD
404/*
405 * Generic procedure to index a pte from a pt, pd, or pdp.
921c891e
MD
406 *
407 * NOTE: Normally passed pindex as pmap_xx_index(). pmap_xx_pindex() is NOT
408 * a page table page index but is instead of PV lookup index.
701c977e
MD
409 */
410static
411void *
412pv_pte_lookup(pv_entry_t pv, vm_pindex_t pindex)
413{
414 pt_entry_t *pte;
415
416 pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pv->pv_m));
417 return(&pte[pindex]);
418}
419
420/*
421 * Return pointer to PDP slot in the PML4
422 */
bfc09ba0
MD
423static __inline
424pml4_entry_t *
701c977e 425pmap_pdp(pmap_t pmap, vm_offset_t va)
48ffc236 426{
701c977e 427 return (&pmap->pm_pml4[pmap_pdp_index(va)]);
48ffc236
JG
428}
429
701c977e
MD
430/*
431 * Return pointer to PD slot in the PDP given a pointer to the PDP
432 */
bfc09ba0
MD
433static __inline
434pdp_entry_t *
eb010d6e 435pmap_pdp_to_pd(pml4_entry_t pdp_pte, vm_offset_t va)
48ffc236 436{
701c977e 437 pdp_entry_t *pd;
48ffc236 438
eb010d6e 439 pd = (pdp_entry_t *)PHYS_TO_DMAP(pdp_pte & PG_FRAME);
701c977e 440 return (&pd[pmap_pd_index(va)]);
48ffc236
JG
441}
442
701c977e 443/*
eb010d6e
MD
444 * Return pointer to PD slot in the PDP.
445 */
bfc09ba0
MD
446static __inline
447pdp_entry_t *
701c977e 448pmap_pd(pmap_t pmap, vm_offset_t va)
48ffc236 449{
701c977e 450 pml4_entry_t *pdp;
48ffc236 451
701c977e
MD
452 pdp = pmap_pdp(pmap, va);
453 if ((*pdp & PG_V) == 0)
48ffc236 454 return NULL;
eb010d6e 455 return (pmap_pdp_to_pd(*pdp, va));
48ffc236
JG
456}
457
701c977e
MD
458/*
459 * Return pointer to PT slot in the PD given a pointer to the PD
460 */
bfc09ba0
MD
461static __inline
462pd_entry_t *
eb010d6e 463pmap_pd_to_pt(pdp_entry_t pd_pte, vm_offset_t va)
48ffc236 464{
701c977e 465 pd_entry_t *pt;
48ffc236 466
eb010d6e 467 pt = (pd_entry_t *)PHYS_TO_DMAP(pd_pte & PG_FRAME);
701c977e 468 return (&pt[pmap_pt_index(va)]);
48ffc236
JG
469}
470
701c977e
MD
471/*
472 * Return pointer to PT slot in the PD
eb010d6e
MD
473 *
474 * SIMPLE PMAP NOTE: Simple pmaps (embedded in objects) do not have PDPs,
475 * so we cannot lookup the PD via the PDP. Instead we
476 * must look it up via the pmap.
701c977e 477 */
bfc09ba0
MD
478static __inline
479pd_entry_t *
701c977e 480pmap_pt(pmap_t pmap, vm_offset_t va)
48ffc236 481{
701c977e 482 pdp_entry_t *pd;
eb010d6e
MD
483 pv_entry_t pv;
484 vm_pindex_t pd_pindex;
485
486 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
487 pd_pindex = pmap_pd_pindex(va);
488 spin_lock(&pmap->pm_spin);
489 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pd_pindex);
490 spin_unlock(&pmap->pm_spin);
491 if (pv == NULL || pv->pv_m == NULL)
492 return NULL;
493 return (pmap_pd_to_pt(VM_PAGE_TO_PHYS(pv->pv_m), va));
494 } else {
495 pd = pmap_pd(pmap, va);
496 if (pd == NULL || (*pd & PG_V) == 0)
497 return NULL;
498 return (pmap_pd_to_pt(*pd, va));
499 }
48ffc236
JG
500}
501
701c977e
MD
502/*
503 * Return pointer to PTE slot in the PT given a pointer to the PT
504 */
bfc09ba0
MD
505static __inline
506pt_entry_t *
eb010d6e 507pmap_pt_to_pte(pd_entry_t pt_pte, vm_offset_t va)
48ffc236
JG
508{
509 pt_entry_t *pte;
510
eb010d6e 511 pte = (pt_entry_t *)PHYS_TO_DMAP(pt_pte & PG_FRAME);
48ffc236
JG
512 return (&pte[pmap_pte_index(va)]);
513}
514
701c977e
MD
515/*
516 * Return pointer to PTE slot in the PT
517 */
bfc09ba0
MD
518static __inline
519pt_entry_t *
48ffc236 520pmap_pte(pmap_t pmap, vm_offset_t va)
48ffc236 521{
701c977e 522 pd_entry_t *pt;
48ffc236 523
701c977e
MD
524 pt = pmap_pt(pmap, va);
525 if (pt == NULL || (*pt & PG_V) == 0)
526 return NULL;
527 if ((*pt & PG_PS) != 0)
528 return ((pt_entry_t *)pt);
eb010d6e 529 return (pmap_pt_to_pte(*pt, va));
48ffc236
JG
530}
531
701c977e
MD
532/*
533 * Of all the layers (PTE, PT, PD, PDP, PML4) the best one to cache is
534 * the PT layer. This will speed up core pmap operations considerably.
535 */
bfc09ba0 536static __inline
701c977e
MD
537void
538pv_cache(pv_entry_t pv, vm_pindex_t pindex)
48ffc236 539{
701c977e
MD
540 if (pindex >= pmap_pt_pindex(0) && pindex <= pmap_pd_pindex(0))
541 pv->pv_pmap->pm_pvhint = pv;
c8fe38ae 542}
d7f50089 543
701c977e
MD
544
545/*
546 * KVM - return address of PT slot in PD
547 */
bfc09ba0
MD
548static __inline
549pd_entry_t *
701c977e 550vtopt(vm_offset_t va)
48ffc236 551{
b12defdc
MD
552 uint64_t mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT +
553 NPML4EPGSHIFT)) - 1);
48ffc236
JG
554
555 return (PDmap + ((va >> PDRSHIFT) & mask));
556}
c8fe38ae 557
701c977e
MD
558/*
559 * KVM - return address of PTE slot in PT
560 */
561static __inline
562pt_entry_t *
563vtopte(vm_offset_t va)
564{
565 uint64_t mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT +
566 NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1);
567
568 return (PTmap + ((va >> PAGE_SHIFT) & mask));
569}
570
48ffc236 571static uint64_t
8e5ea5f7 572allocpages(vm_paddr_t *firstaddr, long n)
d7f50089 573{
48ffc236 574 uint64_t ret;
c8fe38ae
MD
575
576 ret = *firstaddr;
577 bzero((void *)ret, n * PAGE_SIZE);
578 *firstaddr += n * PAGE_SIZE;
579 return (ret);
d7f50089
YY
580}
581
bfc09ba0 582static
c8fe38ae
MD
583void
584create_pagetables(vm_paddr_t *firstaddr)
585{
8e5ea5f7 586 long i; /* must be 64 bits */
da23a592
MD
587 long nkpt_base;
588 long nkpt_phys;
33fb3ba1 589 int j;
c8fe38ae 590
ad54aa11
MD
591 /*
592 * We are running (mostly) V=P at this point
593 *
594 * Calculate NKPT - number of kernel page tables. We have to
595 * accomodoate prealloction of the vm_page_array, dump bitmap,
596 * MSGBUF_SIZE, and other stuff. Be generous.
597 *
598 * Maxmem is in pages.
33fb3ba1
MD
599 *
600 * ndmpdp is the number of 1GB pages we wish to map.
ad54aa11 601 */
86dae8f1
MD
602 ndmpdp = (ptoa(Maxmem) + NBPDP - 1) >> PDPSHIFT;
603 if (ndmpdp < 4) /* Minimum 4GB of dirmap */
604 ndmpdp = 4;
33fb3ba1 605 KKASSERT(ndmpdp <= NKPDPE * NPDEPG);
86dae8f1 606
da23a592
MD
607 /*
608 * Starting at the beginning of kvm (not KERNBASE).
609 */
610 nkpt_phys = (Maxmem * sizeof(struct vm_page) + NBPDR - 1) / NBPDR;
611 nkpt_phys += (Maxmem * sizeof(struct pv_entry) + NBPDR - 1) / NBPDR;
33fb3ba1
MD
612 nkpt_phys += ((nkpt + nkpt + 1 + NKPML4E + NKPDPE + NDMPML4E +
613 ndmpdp) + 511) / 512;
da23a592
MD
614 nkpt_phys += 128;
615
616 /*
617 * Starting at KERNBASE - map 2G worth of page table pages.
618 * KERNBASE is offset -2G from the end of kvm.
619 */
620 nkpt_base = (NPDPEPG - KPDPI) * NPTEPG; /* typically 2 x 512 */
c8fe38ae 621
ad54aa11
MD
622 /*
623 * Allocate pages
624 */
da23a592
MD
625 KPTbase = allocpages(firstaddr, nkpt_base);
626 KPTphys = allocpages(firstaddr, nkpt_phys);
48ffc236
JG
627 KPML4phys = allocpages(firstaddr, 1);
628 KPDPphys = allocpages(firstaddr, NKPML4E);
da23a592 629 KPDphys = allocpages(firstaddr, NKPDPE);
791c6551
MD
630
631 /*
632 * Calculate the page directory base for KERNBASE,
633 * that is where we start populating the page table pages.
634 * Basically this is the end - 2.
635 */
791c6551 636 KPDbase = KPDphys + ((NKPDPE - (NPDPEPG - KPDPI)) << PAGE_SHIFT);
48ffc236 637
48ffc236
JG
638 DMPDPphys = allocpages(firstaddr, NDMPML4E);
639 if ((amd_feature & AMDID_PAGE1GB) == 0)
640 DMPDphys = allocpages(firstaddr, ndmpdp);
641 dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT;
642
791c6551
MD
643 /*
644 * Fill in the underlying page table pages for the area around
645 * KERNBASE. This remaps low physical memory to KERNBASE.
646 *
647 * Read-only from zero to physfree
648 * XXX not fully used, underneath 2M pages
649 */
48ffc236 650 for (i = 0; (i << PAGE_SHIFT) < *firstaddr; i++) {
791c6551
MD
651 ((pt_entry_t *)KPTbase)[i] = i << PAGE_SHIFT;
652 ((pt_entry_t *)KPTbase)[i] |= PG_RW | PG_V | PG_G;
48ffc236
JG
653 }
654
791c6551
MD
655 /*
656 * Now map the initial kernel page tables. One block of page
657 * tables is placed at the beginning of kernel virtual memory,
658 * and another block is placed at KERNBASE to map the kernel binary,
659 * data, bss, and initial pre-allocations.
660 */
da23a592 661 for (i = 0; i < nkpt_base; i++) {
791c6551
MD
662 ((pd_entry_t *)KPDbase)[i] = KPTbase + (i << PAGE_SHIFT);
663 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V;
664 }
da23a592 665 for (i = 0; i < nkpt_phys; i++) {
48ffc236
JG
666 ((pd_entry_t *)KPDphys)[i] = KPTphys + (i << PAGE_SHIFT);
667 ((pd_entry_t *)KPDphys)[i] |= PG_RW | PG_V;
668 }
669
791c6551
MD
670 /*
671 * Map from zero to end of allocations using 2M pages as an
672 * optimization. This will bypass some of the KPTBase pages
673 * above in the KERNBASE area.
674 */
48ffc236 675 for (i = 0; (i << PDRSHIFT) < *firstaddr; i++) {
791c6551
MD
676 ((pd_entry_t *)KPDbase)[i] = i << PDRSHIFT;
677 ((pd_entry_t *)KPDbase)[i] |= PG_RW | PG_V | PG_PS | PG_G;
48ffc236
JG
678 }
679
791c6551
MD
680 /*
681 * And connect up the PD to the PDP. The kernel pmap is expected
682 * to pre-populate all of its PDs. See NKPDPE in vmparam.h.
683 */
48ffc236 684 for (i = 0; i < NKPDPE; i++) {
791c6551
MD
685 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] =
686 KPDphys + (i << PAGE_SHIFT);
687 ((pdp_entry_t *)KPDPphys)[NPDPEPG - NKPDPE + i] |=
688 PG_RW | PG_V | PG_U;
48ffc236
JG
689 }
690
33fb3ba1
MD
691 /*
692 * Now set up the direct map space using either 2MB or 1GB pages
693 * Preset PG_M and PG_A because demotion expects it.
694 *
695 * When filling in entries in the PD pages make sure any excess
696 * entries are set to zero as we allocated enough PD pages
697 */
48ffc236
JG
698 if ((amd_feature & AMDID_PAGE1GB) == 0) {
699 for (i = 0; i < NPDEPG * ndmpdp; i++) {
8e5ea5f7 700 ((pd_entry_t *)DMPDphys)[i] = i << PDRSHIFT;
48ffc236 701 ((pd_entry_t *)DMPDphys)[i] |= PG_RW | PG_V | PG_PS |
33fb3ba1 702 PG_G | PG_M | PG_A;
48ffc236 703 }
33fb3ba1
MD
704
705 /*
706 * And the direct map space's PDP
707 */
48ffc236
JG
708 for (i = 0; i < ndmpdp; i++) {
709 ((pdp_entry_t *)DMPDPphys)[i] = DMPDphys +
33fb3ba1 710 (i << PAGE_SHIFT);
48ffc236
JG
711 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_U;
712 }
713 } else {
714 for (i = 0; i < ndmpdp; i++) {
715 ((pdp_entry_t *)DMPDPphys)[i] =
33fb3ba1 716 (vm_paddr_t)i << PDPSHIFT;
48ffc236 717 ((pdp_entry_t *)DMPDPphys)[i] |= PG_RW | PG_V | PG_PS |
33fb3ba1 718 PG_G | PG_M | PG_A;
48ffc236
JG
719 }
720 }
721
722 /* And recursively map PML4 to itself in order to get PTmap */
723 ((pdp_entry_t *)KPML4phys)[PML4PML4I] = KPML4phys;
724 ((pdp_entry_t *)KPML4phys)[PML4PML4I] |= PG_RW | PG_V | PG_U;
725
33fb3ba1
MD
726 /*
727 * Connect the Direct Map slots up to the PML4
728 */
729 for (j = 0; j < NDMPML4E; ++j) {
730 ((pdp_entry_t *)KPML4phys)[DMPML4I + j] =
731 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
732 PG_RW | PG_V | PG_U;
733 }
48ffc236 734
33fb3ba1
MD
735 /*
736 * Connect the KVA slot up to the PML4
737 */
48ffc236
JG
738 ((pdp_entry_t *)KPML4phys)[KPML4I] = KPDPphys;
739 ((pdp_entry_t *)KPML4phys)[KPML4I] |= PG_RW | PG_V | PG_U;
c8fe38ae
MD
740}
741
d7f50089 742/*
c8fe38ae
MD
743 * Bootstrap the system enough to run with virtual memory.
744 *
745 * On the i386 this is called after mapping has already been enabled
746 * and just syncs the pmap module with what has already been done.
747 * [We can't call it easily with mapping off since the kernel is not
748 * mapped with PA == VA, hence we would have to relocate every address
749 * from the linked base (virtual) address "KERNBASE" to the actual
750 * (physical) address starting relative to 0]
d7f50089
YY
751 */
752void
48ffc236 753pmap_bootstrap(vm_paddr_t *firstaddr)
c8fe38ae
MD
754{
755 vm_offset_t va;
756 pt_entry_t *pte;
757 struct mdglobaldata *gd;
c8fe38ae
MD
758 int pg;
759
48ffc236
JG
760 KvaStart = VM_MIN_KERNEL_ADDRESS;
761 KvaEnd = VM_MAX_KERNEL_ADDRESS;
762 KvaSize = KvaEnd - KvaStart;
763
c8fe38ae
MD
764 avail_start = *firstaddr;
765
766 /*
48ffc236 767 * Create an initial set of page tables to run the kernel in.
c8fe38ae 768 */
48ffc236
JG
769 create_pagetables(firstaddr);
770
791c6551
MD
771 virtual2_start = KvaStart;
772 virtual2_end = PTOV_OFFSET;
773
c8fe38ae
MD
774 virtual_start = (vm_offset_t) PTOV_OFFSET + *firstaddr;
775 virtual_start = pmap_kmem_choose(virtual_start);
48ffc236
JG
776
777 virtual_end = VM_MAX_KERNEL_ADDRESS;
778
779 /* XXX do %cr0 as well */
780 load_cr4(rcr4() | CR4_PGE | CR4_PSE);
781 load_cr3(KPML4phys);
c8fe38ae
MD
782
783 /*
784 * Initialize protection array.
785 */
786 i386_protection_init();
787
788 /*
789 * The kernel's pmap is statically allocated so we don't have to use
790 * pmap_create, which is unlikely to work correctly at this part of
791 * the boot sequence (XXX and which no longer exists).
792 */
48ffc236 793 kernel_pmap.pm_pml4 = (pdp_entry_t *) (PTOV_OFFSET + KPML4phys);
c8fe38ae 794 kernel_pmap.pm_count = 1;
c2fb025d 795 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
701c977e 796 RB_INIT(&kernel_pmap.pm_pvroot);
b12defdc
MD
797 spin_init(&kernel_pmap.pm_spin);
798 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
c8fe38ae
MD
799
800 /*
801 * Reserve some special page table entries/VA space for temporary
802 * mapping of pages.
803 */
804#define SYSMAP(c, p, v, n) \
805 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
806
807 va = virtual_start;
48ffc236 808 pte = vtopte(va);
c8fe38ae
MD
809
810 /*
811 * CMAP1/CMAP2 are used for zeroing and copying pages.
812 */
813 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
814
815 /*
816 * Crashdump maps.
817 */
818 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
819
820 /*
821 * ptvmmap is used for reading arbitrary physical pages via
822 * /dev/mem.
823 */
824 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
825
826 /*
827 * msgbufp is used to map the system message buffer.
828 * XXX msgbufmap is not used.
829 */
830 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
831 atop(round_page(MSGBUF_SIZE)))
832
833 virtual_start = va;
834
835 *CMAP1 = 0;
c8fe38ae
MD
836
837 /*
838 * PG_G is terribly broken on SMP because we IPI invltlb's in some
839 * cases rather then invl1pg. Actually, I don't even know why it
840 * works under UP because self-referential page table mappings
841 */
842#ifdef SMP
843 pgeflag = 0;
844#else
845 if (cpu_feature & CPUID_PGE)
846 pgeflag = PG_G;
847#endif
848
849/*
850 * Initialize the 4MB page size flag
851 */
852 pseflag = 0;
853/*
854 * The 4MB page version of the initial
855 * kernel page mapping.
856 */
857 pdir4mb = 0;
858
859#if !defined(DISABLE_PSE)
860 if (cpu_feature & CPUID_PSE) {
861 pt_entry_t ptditmp;
862 /*
863 * Note that we have enabled PSE mode
864 */
865 pseflag = PG_PS;
b2b3ffcd 866 ptditmp = *(PTmap + x86_64_btop(KERNBASE));
c8fe38ae
MD
867 ptditmp &= ~(NBPDR - 1);
868 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
869 pdir4mb = ptditmp;
870
871#ifndef SMP
872 /*
873 * Enable the PSE mode. If we are SMP we can't do this
874 * now because the APs will not be able to use it when
875 * they boot up.
876 */
877 load_cr4(rcr4() | CR4_PSE);
878
879 /*
880 * We can do the mapping here for the single processor
881 * case. We simply ignore the old page table page from
882 * now on.
883 */
884 /*
885 * For SMP, we still need 4K pages to bootstrap APs,
886 * PSE will be enabled as soon as all APs are up.
887 */
888 PTD[KPTDI] = (pd_entry_t)ptditmp;
c8fe38ae
MD
889 cpu_invltlb();
890#endif
891 }
892#endif
c8fe38ae
MD
893
894 /*
895 * We need to finish setting up the globaldata page for the BSP.
896 * locore has already populated the page table for the mdglobaldata
897 * portion.
898 */
899 pg = MDGLOBALDATA_BASEALLOC_PAGES;
900 gd = &CPU_prvspace[0].mdglobaldata;
c8fe38ae
MD
901
902 cpu_invltlb();
d7f50089
YY
903}
904
c8fe38ae 905#ifdef SMP
d7f50089 906/*
c8fe38ae 907 * Set 4mb pdir for mp startup
d7f50089
YY
908 */
909void
c8fe38ae
MD
910pmap_set_opt(void)
911{
912 if (pseflag && (cpu_feature & CPUID_PSE)) {
913 load_cr4(rcr4() | CR4_PSE);
914 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
c8fe38ae
MD
915 cpu_invltlb();
916 }
917 }
d7f50089 918}
c8fe38ae 919#endif
d7f50089 920
c8fe38ae
MD
921/*
922 * Initialize the pmap module.
923 * Called by vm_init, to initialize any structures that the pmap
924 * system needs to map virtual memory.
925 * pmap_init has been enhanced to support in a fairly consistant
926 * way, discontiguous physical memory.
d7f50089
YY
927 */
928void
c8fe38ae 929pmap_init(void)
d7f50089 930{
c8fe38ae
MD
931 int i;
932 int initial_pvs;
933
934 /*
c8fe38ae
MD
935 * Allocate memory for random pmap data structures. Includes the
936 * pv_head_table.
937 */
938
701c977e 939 for (i = 0; i < vm_page_array_size; i++) {
c8fe38ae
MD
940 vm_page_t m;
941
942 m = &vm_page_array[i];
943 TAILQ_INIT(&m->md.pv_list);
c8fe38ae
MD
944 }
945
946 /*
947 * init the pv free list
948 */
949 initial_pvs = vm_page_array_size;
950 if (initial_pvs < MINPV)
951 initial_pvs = MINPV;
952 pvzone = &pvzone_store;
948209ce
MD
953 pvinit = (void *)kmem_alloc(&kernel_map,
954 initial_pvs * sizeof (struct pv_entry));
955 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
956 pvinit, initial_pvs);
c8fe38ae
MD
957
958 /*
959 * Now it is safe to enable pv_table recording.
960 */
961 pmap_initialized = TRUE;
d7f50089
YY
962}
963
c8fe38ae
MD
964/*
965 * Initialize the address space (zone) for the pv_entries. Set a
966 * high water mark so that the system can recover from excessive
967 * numbers of pv entries.
968 */
d7f50089 969void
c8fe38ae 970pmap_init2(void)
d7f50089 971{
c8fe38ae 972 int shpgperproc = PMAP_SHPGPERPROC;
948209ce 973 int entry_max;
c8fe38ae
MD
974
975 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
976 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
977 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
978 pv_entry_high_water = 9 * (pv_entry_max / 10);
948209ce
MD
979
980 /*
981 * Subtract out pages already installed in the zone (hack)
982 */
983 entry_max = pv_entry_max - vm_page_array_size;
984 if (entry_max <= 0)
985 entry_max = 1;
986
987 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
d7f50089
YY
988}
989
c8fe38ae
MD
990
991/***************************************************
992 * Low level helper routines.....
993 ***************************************************/
994
c8fe38ae
MD
995/*
996 * this routine defines the region(s) of memory that should
997 * not be tested for the modified bit.
998 */
bfc09ba0
MD
999static __inline
1000int
701c977e 1001pmap_track_modified(vm_pindex_t pindex)
d7f50089 1002{
701c977e 1003 vm_offset_t va = (vm_offset_t)pindex << PAGE_SHIFT;
c8fe38ae
MD
1004 if ((va < clean_sva) || (va >= clean_eva))
1005 return 1;
1006 else
1007 return 0;
d7f50089
YY
1008}
1009
d7f50089 1010/*
10d6182e 1011 * Extract the physical page address associated with the map/VA pair.
701c977e 1012 * The page must be wired for this to work reliably.
c8fe38ae 1013 *
701c977e
MD
1014 * XXX for the moment we're using pv_find() instead of pv_get(), as
1015 * callers might be expecting non-blocking operation.
d7f50089 1016 */
c8fe38ae
MD
1017vm_paddr_t
1018pmap_extract(pmap_t pmap, vm_offset_t va)
d7f50089 1019{
48ffc236 1020 vm_paddr_t rtval;
701c977e
MD
1021 pv_entry_t pt_pv;
1022 pt_entry_t *ptep;
c8fe38ae 1023
48ffc236 1024 rtval = 0;
701c977e
MD
1025 if (va >= VM_MAX_USER_ADDRESS) {
1026 /*
1027 * Kernel page directories might be direct-mapped and
1028 * there is typically no PV tracking of pte's
1029 */
1030 pd_entry_t *pt;
1031
1032 pt = pmap_pt(pmap, va);
1033 if (pt && (*pt & PG_V)) {
1034 if (*pt & PG_PS) {
1035 rtval = *pt & PG_PS_FRAME;
1036 rtval |= va & PDRMASK;
48ffc236 1037 } else {
eb010d6e 1038 ptep = pmap_pt_to_pte(*pt, va);
701c977e
MD
1039 if (*pt & PG_V) {
1040 rtval = *ptep & PG_FRAME;
1041 rtval |= va & PAGE_MASK;
1042 }
1043 }
1044 }
1045 } else {
1046 /*
1047 * User pages currently do not direct-map the page directory
1048 * and some pages might not used managed PVs. But all PT's
1049 * will have a PV.
1050 */
1051 pt_pv = pv_find(pmap, pmap_pt_pindex(va));
1052 if (pt_pv) {
1053 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(va));
1054 if (*ptep & PG_V) {
1055 rtval = *ptep & PG_FRAME;
1056 rtval |= va & PAGE_MASK;
48ffc236 1057 }
701c977e 1058 pv_drop(pt_pv);
c8fe38ae 1059 }
c8fe38ae 1060 }
48ffc236
JG
1061 return rtval;
1062}
1063
1064/*
10d6182e 1065 * Extract the physical page address associated kernel virtual address.
48ffc236
JG
1066 */
1067vm_paddr_t
1068pmap_kextract(vm_offset_t va)
48ffc236 1069{
701c977e 1070 pd_entry_t pt; /* pt entry in pd */
48ffc236
JG
1071 vm_paddr_t pa;
1072
1073 if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) {
1074 pa = DMAP_TO_PHYS(va);
1075 } else {
701c977e
MD
1076 pt = *vtopt(va);
1077 if (pt & PG_PS) {
1078 pa = (pt & PG_PS_FRAME) | (va & PDRMASK);
48ffc236
JG
1079 } else {
1080 /*
1081 * Beware of a concurrent promotion that changes the
1082 * PDE at this point! For example, vtopte() must not
1083 * be used to access the PTE because it would use the
1084 * new PDE. It is, however, safe to use the old PDE
1085 * because the page table page is preserved by the
1086 * promotion.
1087 */
eb010d6e 1088 pa = *pmap_pt_to_pte(pt, va);
48ffc236
JG
1089 pa = (pa & PG_FRAME) | (va & PAGE_MASK);
1090 }
1091 }
1092 return pa;
d7f50089
YY
1093}
1094
c8fe38ae
MD
1095/***************************************************
1096 * Low level mapping routines.....
1097 ***************************************************/
1098
d7f50089 1099/*
c8fe38ae
MD
1100 * Routine: pmap_kenter
1101 * Function:
1102 * Add a wired page to the KVA
1103 * NOTE! note that in order for the mapping to take effect -- you
1104 * should do an invltlb after doing the pmap_kenter().
d7f50089 1105 */
c8fe38ae 1106void
d7f50089
YY
1107pmap_kenter(vm_offset_t va, vm_paddr_t pa)
1108{
c8fe38ae
MD
1109 pt_entry_t *pte;
1110 pt_entry_t npte;
1111 pmap_inval_info info;
1112
701c977e 1113 pmap_inval_init(&info); /* XXX remove */
c8fe38ae
MD
1114 npte = pa | PG_RW | PG_V | pgeflag;
1115 pte = vtopte(va);
701c977e 1116 pmap_inval_interlock(&info, &kernel_pmap, va); /* XXX remove */
c8fe38ae 1117 *pte = npte;
701c977e
MD
1118 pmap_inval_deinterlock(&info, &kernel_pmap); /* XXX remove */
1119 pmap_inval_done(&info); /* XXX remove */
d7f50089
YY
1120}
1121
1122/*
c8fe38ae
MD
1123 * Routine: pmap_kenter_quick
1124 * Function:
1125 * Similar to pmap_kenter(), except we only invalidate the
1126 * mapping on the current CPU.
d7f50089
YY
1127 */
1128void
c8fe38ae
MD
1129pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
1130{
1131 pt_entry_t *pte;
1132 pt_entry_t npte;
1133
1134 npte = pa | PG_RW | PG_V | pgeflag;
1135 pte = vtopte(va);
1136 *pte = npte;
1137 cpu_invlpg((void *)va);
1138}
1139
1140void
d7f50089
YY
1141pmap_kenter_sync(vm_offset_t va)
1142{
c8fe38ae
MD
1143 pmap_inval_info info;
1144
1145 pmap_inval_init(&info);
c2fb025d
MD
1146 pmap_inval_interlock(&info, &kernel_pmap, va);
1147 pmap_inval_deinterlock(&info, &kernel_pmap);
1148 pmap_inval_done(&info);
d7f50089
YY
1149}
1150
d7f50089
YY
1151void
1152pmap_kenter_sync_quick(vm_offset_t va)
1153{
c8fe38ae 1154 cpu_invlpg((void *)va);
d7f50089
YY
1155}
1156
d7f50089 1157/*
c8fe38ae 1158 * remove a page from the kernel pagetables
d7f50089
YY
1159 */
1160void
c8fe38ae 1161pmap_kremove(vm_offset_t va)
d7f50089 1162{
c8fe38ae
MD
1163 pt_entry_t *pte;
1164 pmap_inval_info info;
1165
1166 pmap_inval_init(&info);
1167 pte = vtopte(va);
c2fb025d 1168 pmap_inval_interlock(&info, &kernel_pmap, va);
52bb73bc 1169 (void)pte_load_clear(pte);
c2fb025d
MD
1170 pmap_inval_deinterlock(&info, &kernel_pmap);
1171 pmap_inval_done(&info);
c8fe38ae
MD
1172}
1173
1174void
1175pmap_kremove_quick(vm_offset_t va)
1176{
1177 pt_entry_t *pte;
1178 pte = vtopte(va);
52bb73bc 1179 (void)pte_load_clear(pte);
c8fe38ae 1180 cpu_invlpg((void *)va);
d7f50089
YY
1181}
1182
1183/*
c8fe38ae 1184 * XXX these need to be recoded. They are not used in any critical path.
d7f50089
YY
1185 */
1186void
c8fe38ae 1187pmap_kmodify_rw(vm_offset_t va)
d7f50089 1188{
701c977e 1189 atomic_set_long(vtopte(va), PG_RW);
c8fe38ae 1190 cpu_invlpg((void *)va);
d7f50089
YY
1191}
1192
c8fe38ae
MD
1193void
1194pmap_kmodify_nc(vm_offset_t va)
1195{
701c977e 1196 atomic_set_long(vtopte(va), PG_N);
c8fe38ae
MD
1197 cpu_invlpg((void *)va);
1198}
d7f50089
YY
1199
1200/*
ad54aa11
MD
1201 * Used to map a range of physical addresses into kernel virtual
1202 * address space during the low level boot, typically to map the
1203 * dump bitmap, message buffer, and vm_page_array.
c8fe38ae 1204 *
ad54aa11
MD
1205 * These mappings are typically made at some pointer after the end of the
1206 * kernel text+data.
1207 *
1208 * We could return PHYS_TO_DMAP(start) here and not allocate any
1209 * via (*virtp), but then kmem from userland and kernel dumps won't
1210 * have access to the related pointers.
d7f50089
YY
1211 */
1212vm_offset_t
8e5e6f1b 1213pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
d7f50089 1214{
ad54aa11
MD
1215 vm_offset_t va;
1216 vm_offset_t va_start;
1217
1218 /*return PHYS_TO_DMAP(start);*/
1219
1220 va_start = *virtp;
1221 va = va_start;
1222
1223 while (start < end) {
1224 pmap_kenter_quick(va, start);
1225 va += PAGE_SIZE;
1226 start += PAGE_SIZE;
1227 }
1228 *virtp = va;
1229 return va_start;
d7f50089
YY
1230}
1231
c8fe38ae 1232
d7f50089 1233/*
c8fe38ae
MD
1234 * Add a list of wired pages to the kva
1235 * this routine is only used for temporary
1236 * kernel mappings that do not need to have
1237 * page modification or references recorded.
1238 * Note that old mappings are simply written
1239 * over. The page *must* be wired.
d7f50089
YY
1240 */
1241void
c8fe38ae 1242pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
d7f50089 1243{
c8fe38ae
MD
1244 vm_offset_t end_va;
1245
1246 end_va = va + count * PAGE_SIZE;
1247
1248 while (va < end_va) {
1249 pt_entry_t *pte;
1250
1251 pte = vtopte(va);
1252 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
1253 cpu_invlpg((void *)va);
1254 va += PAGE_SIZE;
1255 m++;
1256 }
7d4d6fdb 1257 smp_invltlb();
c8fe38ae
MD
1258}
1259
d7f50089 1260/*
7155fc7d 1261 * This routine jerks page mappings from the
c8fe38ae 1262 * kernel -- it is meant only for temporary mappings.
7155fc7d
MD
1263 *
1264 * MPSAFE, INTERRUPT SAFE (cluster callback)
d7f50089 1265 */
c8fe38ae
MD
1266void
1267pmap_qremove(vm_offset_t va, int count)
d7f50089 1268{
c8fe38ae
MD
1269 vm_offset_t end_va;
1270
48ffc236 1271 end_va = va + count * PAGE_SIZE;
c8fe38ae
MD
1272
1273 while (va < end_va) {
1274 pt_entry_t *pte;
1275
1276 pte = vtopte(va);
52bb73bc 1277 (void)pte_load_clear(pte);
c8fe38ae
MD
1278 cpu_invlpg((void *)va);
1279 va += PAGE_SIZE;
1280 }
c8fe38ae 1281 smp_invltlb();
d7f50089
YY
1282}
1283
1284/*
c8fe38ae
MD
1285 * Create a new thread and optionally associate it with a (new) process.
1286 * NOTE! the new thread's cpu may not equal the current cpu.
d7f50089
YY
1287 */
1288void
c8fe38ae 1289pmap_init_thread(thread_t td)
d7f50089 1290{
d1368d1a 1291 /* enforce pcb placement & alignment */
c8fe38ae 1292 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
d1368d1a 1293 td->td_pcb = (struct pcb *)((intptr_t)td->td_pcb & ~(intptr_t)0xF);
c8fe38ae 1294 td->td_savefpu = &td->td_pcb->pcb_save;
d1368d1a 1295 td->td_sp = (char *)td->td_pcb; /* no -16 */
d7f50089
YY
1296}
1297
1298/*
c8fe38ae 1299 * This routine directly affects the fork perf for a process.
d7f50089
YY
1300 */
1301void
c8fe38ae 1302pmap_init_proc(struct proc *p)
d7f50089
YY
1303{
1304}
1305
1306/*
c8fe38ae
MD
1307 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1308 * it, and IdlePTD, represents the template used to update all other pmaps.
1309 *
1310 * On architectures where the kernel pmap is not integrated into the user
1311 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1312 * kernel_pmap should be used to directly access the kernel_pmap.
d7f50089
YY
1313 */
1314void
c8fe38ae 1315pmap_pinit0(struct pmap *pmap)
d7f50089 1316{
48ffc236 1317 pmap->pm_pml4 = (pml4_entry_t *)(PTOV_OFFSET + KPML4phys);
c8fe38ae
MD
1318 pmap->pm_count = 1;
1319 pmap->pm_active = 0;
701c977e
MD
1320 pmap->pm_pvhint = NULL;
1321 RB_INIT(&pmap->pm_pvroot);
b12defdc
MD
1322 spin_init(&pmap->pm_spin);
1323 lwkt_token_init(&pmap->pm_token, "pmap_tok");
c8fe38ae 1324 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
d7f50089
YY
1325}
1326
1327/*
c8fe38ae
MD
1328 * Initialize a preallocated and zeroed pmap structure,
1329 * such as one in a vmspace structure.
d7f50089 1330 */
921c891e
MD
1331static void
1332pmap_pinit_simple(struct pmap *pmap)
d7f50089 1333{
701c977e
MD
1334 /*
1335 * Misc initialization
1336 */
1337 pmap->pm_count = 1;
1338 pmap->pm_active = 0;
1339 pmap->pm_pvhint = NULL;
921c891e
MD
1340 pmap->pm_flags = PMAP_FLAG_SIMPLE;
1341
1342 /*
1343 * Don't blow up locks/tokens on re-use (XXX fix/use drop code
1344 * for this).
1345 */
701c977e
MD
1346 if (pmap->pm_pmlpv == NULL) {
1347 RB_INIT(&pmap->pm_pvroot);
1348 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1349 spin_init(&pmap->pm_spin);
1350 lwkt_token_init(&pmap->pm_token, "pmap_tok");
1351 }
921c891e
MD
1352}
1353
1354void
1355pmap_pinit(struct pmap *pmap)
1356{
1357 pv_entry_t pv;
1358 int j;
1359
1360 pmap_pinit_simple(pmap);
1361 pmap->pm_flags &= ~PMAP_FLAG_SIMPLE;
c8fe38ae
MD
1362
1363 /*
1364 * No need to allocate page table space yet but we do need a valid
1365 * page directory table.
1366 */
48ffc236
JG
1367 if (pmap->pm_pml4 == NULL) {
1368 pmap->pm_pml4 =
1369 (pml4_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
c8fe38ae
MD
1370 }
1371
1372 /*
701c977e
MD
1373 * Allocate the page directory page, which wires it even though
1374 * it isn't being entered into some higher level page table (it
1375 * being the highest level). If one is already cached we don't
1376 * have to do anything.
c8fe38ae 1377 */
701c977e
MD
1378 if ((pv = pmap->pm_pmlpv) == NULL) {
1379 pv = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1380 pmap->pm_pmlpv = pv;
b12defdc 1381 pmap_kenter((vm_offset_t)pmap->pm_pml4,
701c977e
MD
1382 VM_PAGE_TO_PHYS(pv->pv_m));
1383 pv_put(pv);
33fb3ba1
MD
1384
1385 /*
1386 * Install DMAP and KMAP.
1387 */
1388 for (j = 0; j < NDMPML4E; ++j) {
1389 pmap->pm_pml4[DMPML4I + j] =
1390 (DMPDPphys + ((vm_paddr_t)j << PML4SHIFT)) |
1391 PG_RW | PG_V | PG_U;
1392 }
701c977e 1393 pmap->pm_pml4[KPML4I] = KPDPphys | PG_RW | PG_V | PG_U;
701c977e 1394
33fb3ba1
MD
1395 /*
1396 * install self-referential address mapping entry
1397 */
701c977e
MD
1398 pmap->pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pv->pv_m) |
1399 PG_V | PG_RW | PG_A | PG_M;
1400 } else {
1401 KKASSERT(pv->pv_m->flags & PG_MAPPED);
1402 KKASSERT(pv->pv_m->flags & PG_WRITEABLE);
b12defdc 1403 }
993bac44
MD
1404 KKASSERT(pmap->pm_pml4[255] == 0);
1405 KKASSERT(RB_ROOT(&pmap->pm_pvroot) == pv);
1406 KKASSERT(pv->pv_entry.rbe_left == NULL);
1407 KKASSERT(pv->pv_entry.rbe_right == NULL);
d7f50089
YY
1408}
1409
1410/*
c8fe38ae
MD
1411 * Clean up a pmap structure so it can be physically freed. This routine
1412 * is called by the vmspace dtor function. A great deal of pmap data is
1413 * left passively mapped to improve vmspace management so we have a bit
1414 * of cleanup work to do here.
d7f50089
YY
1415 */
1416void
c8fe38ae 1417pmap_puninit(pmap_t pmap)
d7f50089 1418{
701c977e 1419 pv_entry_t pv;
c8fe38ae
MD
1420 vm_page_t p;
1421
1422 KKASSERT(pmap->pm_active == 0);
701c977e
MD
1423 if ((pv = pmap->pm_pmlpv) != NULL) {
1424 if (pv_hold_try(pv) == 0)
1425 pv_lock(pv);
52bb73bc 1426 p = pmap_remove_pv_page(pv);
701c977e 1427 pv_free(pv);
48ffc236 1428 pmap_kremove((vm_offset_t)pmap->pm_pml4);
b12defdc 1429 vm_page_busy_wait(p, FALSE, "pgpun");
701c977e 1430 KKASSERT(p->flags & (PG_FICTITIOUS|PG_UNMANAGED));
b12defdc 1431 vm_page_unwire(p, 0);
701c977e
MD
1432 vm_page_flag_clear(p, PG_MAPPED | PG_WRITEABLE);
1433
1434 /*
1435 * XXX eventually clean out PML4 static entries and
1436 * use vm_page_free_zero()
1437 */
1438 vm_page_free(p);
1439 pmap->pm_pmlpv = NULL;
c8fe38ae 1440 }
48ffc236 1441 if (pmap->pm_pml4) {
bfc09ba0 1442 KKASSERT(pmap->pm_pml4 != (void *)(PTOV_OFFSET + KPML4phys));
48ffc236
JG
1443 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pml4, PAGE_SIZE);
1444 pmap->pm_pml4 = NULL;
c8fe38ae 1445 }
701c977e
MD
1446 KKASSERT(pmap->pm_stats.resident_count == 0);
1447 KKASSERT(pmap->pm_stats.wired_count == 0);
d7f50089
YY
1448}
1449
1450/*
c8fe38ae
MD
1451 * Wire in kernel global address entries. To avoid a race condition
1452 * between pmap initialization and pmap_growkernel, this procedure
1453 * adds the pmap to the master list (which growkernel scans to update),
1454 * then copies the template.
d7f50089
YY
1455 */
1456void
c8fe38ae 1457pmap_pinit2(struct pmap *pmap)
d7f50089 1458{
b12defdc 1459 spin_lock(&pmap_spin);
c8fe38ae 1460 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
b12defdc 1461 spin_unlock(&pmap_spin);
d7f50089
YY
1462}
1463
1464/*
701c977e
MD
1465 * This routine is called when various levels in the page table need to
1466 * be populated. This routine cannot fail.
d7f50089 1467 *
701c977e
MD
1468 * This function returns two locked pv_entry's, one representing the
1469 * requested pv and one representing the requested pv's parent pv. If
1470 * the pv did not previously exist it will be mapped into its parent
1471 * and wired, otherwise no additional wire count will be added.
d7f50089 1472 */
bfc09ba0 1473static
701c977e
MD
1474pv_entry_t
1475pmap_allocpte(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp)
d7f50089 1476{
701c977e
MD
1477 pt_entry_t *ptep;
1478 pv_entry_t pv;
1479 pv_entry_t pvp;
1480 vm_pindex_t pt_pindex;
1481 vm_page_t m;
1482 int isnew;
921c891e 1483 int ispt;
701c977e 1484
c8fe38ae 1485 /*
701c977e
MD
1486 * If the pv already exists and we aren't being asked for the
1487 * parent page table page we can just return it. A locked+held pv
1488 * is returned.
c8fe38ae 1489 */
921c891e 1490 ispt = 0;
701c977e
MD
1491 pv = pv_alloc(pmap, ptepindex, &isnew);
1492 if (isnew == 0 && pvpp == NULL)
1493 return(pv);
1494
1495 /*
1496 * This is a new PV, we have to resolve its parent page table and
1497 * add an additional wiring to the page if necessary.
1498 */
1499
1500 /*
1501 * Special case terminal PVs. These are not page table pages so
1502 * no vm_page is allocated (the caller supplied the vm_page). If
1503 * pvpp is non-NULL we are being asked to also removed the pt_pv
1504 * for this pv.
1505 *
1506 * Note that pt_pv's are only returned for user VAs. We assert that
1507 * a pt_pv is not being requested for kernel VAs.
1508 */
1509 if (ptepindex < pmap_pt_pindex(0)) {
1510 if (ptepindex >= NUPTE_USER)
1511 KKASSERT(pvpp == NULL);
1512 else
1513 KKASSERT(pvpp != NULL);
1514 if (pvpp) {
1515 pt_pindex = NUPTE_TOTAL + (ptepindex >> NPTEPGSHIFT);
1516 pvp = pmap_allocpte(pmap, pt_pindex, NULL);
1517 if (isnew)
1518 vm_page_wire_quick(pvp->pv_m);
1519 *pvpp = pvp;
1520 } else {
1521 pvp = NULL;
1522 }
1523 return(pv);
b12defdc 1524 }
c8fe38ae
MD
1525
1526 /*
701c977e
MD
1527 * Non-terminal PVs allocate a VM page to represent the page table,
1528 * so we have to resolve pvp and calculate ptepindex for the pvp
1529 * and then for the page table entry index in the pvp for
1530 * fall-through.
c8fe38ae 1531 */
701c977e 1532 if (ptepindex < pmap_pd_pindex(0)) {
4a4ea614 1533 /*
701c977e 1534 * pv is PT, pvp is PD
4a4ea614 1535 */
701c977e
MD
1536 ptepindex = (ptepindex - pmap_pt_pindex(0)) >> NPDEPGSHIFT;
1537 ptepindex += NUPTE_TOTAL + NUPT_TOTAL;
1538 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1539 if (!isnew)
1540 goto notnew;
1541
1b2e0b92 1542 /*
701c977e 1543 * PT index in PD
1b2e0b92 1544 */
701c977e
MD
1545 ptepindex = pv->pv_pindex - pmap_pt_pindex(0);
1546 ptepindex &= ((1ul << NPDEPGSHIFT) - 1);
921c891e 1547 ispt = 1;
701c977e 1548 } else if (ptepindex < pmap_pdp_pindex(0)) {
1b2e0b92 1549 /*
701c977e 1550 * pv is PD, pvp is PDP
921c891e
MD
1551 *
1552 * SIMPLE PMAP NOTE: Simple pmaps do not allocate above
1553 * the PD.
1b2e0b92 1554 */
701c977e
MD
1555 ptepindex = (ptepindex - pmap_pd_pindex(0)) >> NPDPEPGSHIFT;
1556 ptepindex += NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL;
921c891e
MD
1557
1558 if (pmap->pm_flags & PMAP_FLAG_SIMPLE) {
1559 KKASSERT(pvpp == NULL);
1560 pvp = NULL;
1561 } else {
1562 pvp = pmap_allocpte(pmap, ptepindex, NULL);
1563 }
701c977e
MD
1564 if (!isnew)
1565 goto notnew;
1566
1567 /*
1568 * PD index in PDP
1569 */
1570 ptepindex = pv->pv_pindex - pmap_pd_pindex(0);
1571 ptepindex &= ((1ul << NPDPEPGSHIFT) - 1);
1572 } else if (ptepindex < pmap_pml4_pindex()) {
700e22f7 1573 /*
701c977e 1574 * pv is PDP, pvp is the root pml4 table
1b2e0b92 1575 */
701c977e
MD
1576 pvp = pmap_allocpte(pmap, pmap_pml4_pindex(), NULL);
1577 if (!isnew)
1578 goto notnew;
700e22f7 1579
701c977e
MD
1580 /*
1581 * PDP index in PML4
1582 */
1583 ptepindex = pv->pv_pindex - pmap_pdp_pindex(0);
1584 ptepindex &= ((1ul << NPML4EPGSHIFT) - 1);
1585 } else {
1586 /*
1587 * pv represents the top-level PML4, there is no parent.
1588 */
1589 pvp = NULL;
1590 if (!isnew)
1591 goto notnew;
1b2e0b92 1592 }
700e22f7
MD
1593
1594 /*
701c977e
MD
1595 * This code is only reached if isnew is TRUE and this is not a
1596 * terminal PV. We need to allocate a vm_page for the page table
1597 * at this level and enter it into the parent page table.
1598 *
1599 * page table pages are marked PG_WRITEABLE and PG_MAPPED.
1b2e0b92 1600 */
701c977e
MD
1601 for (;;) {
1602 m = vm_page_alloc(NULL, pv->pv_pindex,
1603 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
1604 VM_ALLOC_INTERRUPT);
1605 if (m)
1606 break;
1607 vm_wait(0);
1b2e0b92 1608 }
701c977e
MD
1609 vm_page_spin_lock(m);
1610 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1611 pv->pv_m = m;
1612 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1613 vm_page_spin_unlock(m);
1614 vm_page_unmanage(m); /* m must be spinunlocked */
1615
1616 if ((m->flags & PG_ZERO) == 0) {
1617 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1618 }
1619#ifdef PMAP_DEBUG
1620 else {
1621 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1622 }
1623#endif
1624 m->valid = VM_PAGE_BITS_ALL;
1625 vm_page_flag_clear(m, PG_ZERO);
1626 vm_page_wire(m); /* wire for mapping in parent */
1627
1628 /*
1629 * Wire the page into pvp, bump the wire-count for pvp's page table
1630 * page. Bump the resident_count for the pmap. There is no pvp
1631 * for the top level, address the pm_pml4[] array directly.
1632 *
1633 * If the caller wants the parent we return it, otherwise
1634 * we just put it away.
1635 *
1636 * No interlock is needed for pte 0 -> non-zero.
921c891e
MD
1637 *
1638 * In the situation where *ptep is valid we might have an unmanaged
1639 * page table page shared from another page table which we need to
1640 * unshare before installing our private page table page.
701c977e
MD
1641 */
1642 if (pvp) {
701c977e 1643 ptep = pv_pte_lookup(pvp, ptepindex);
921c891e
MD
1644 if (*ptep & PG_V) {
1645 pt_entry_t pte;
1646 pmap_inval_info info;
1647
1648 kprintf("pmap_allocpte: restate shared pg table pg\n");
1649
1650 if (ispt == 0) {
1651 panic("pmap_allocpte: unexpected pte %p/%d",
1652 pvp, (int)ptepindex);
1653 }
1654 pmap_inval_init(&info);
1655 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
1656 pte = pte_load_clear(ptep);
1657 pmap_inval_deinterlock(&info, pmap);
1658 pmap_inval_done(&info);
1659 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
1660 panic("pmap_allocpte: shared pgtable pg bad wirecount");
1661 } else {
1662 vm_page_wire_quick(pvp->pv_m);
1663 }
701c977e
MD
1664 *ptep = VM_PAGE_TO_PHYS(m) | (PG_U | PG_RW | PG_V |
1665 PG_A | PG_M);
1666 }
1667 vm_page_wakeup(m);
1668notnew:
1669 if (pvpp)
1670 *pvpp = pvp;
1671 else if (pvp)
1672 pv_put(pvp);
1673 return (pv);
1674}
d7f50089
YY
1675
1676/*
921c891e
MD
1677 * This version of pmap_allocpte() checks for possible segment optimizations
1678 * that would allow page-table sharing. It can be called for terminal
1679 * page or page table page ptepindex's.
1680 *
1681 * The function is called with page table page ptepindex's for fictitious
1682 * and unmanaged terminal pages. That is, we don't want to allocate a
1683 * terminal pv, we just want the pt_pv. pvpp is usually passed as NULL
1684 * for this case.
1685 *
1686 * This function can return a pv and *pvpp associated with the passed in pmap
1687 * OR a pv and *pvpp associated with the shared pmap. In the latter case
1688 * an unmanaged page table page will be entered into the pass in pmap.
1689 */
1690static
1691pv_entry_t
1692pmap_allocpte_seg(pmap_t pmap, vm_pindex_t ptepindex, pv_entry_t *pvpp,
1693 vm_map_entry_t entry, vm_offset_t va)
1694{
1695 struct pmap_inval_info info;
1696 vm_object_t object;
1697 pmap_t obpmap;
1698 pmap_t *obpmapp;
1699 vm_offset_t b;
1700 pv_entry_t pte_pv; /* in original or shared pmap */
1701 pv_entry_t pt_pv; /* in original or shared pmap */
1702 pv_entry_t proc_pd_pv; /* in original pmap */
1703 pv_entry_t proc_pt_pv; /* in original pmap */
1704 pv_entry_t xpv; /* PT in shared pmap */
1705 pd_entry_t *pt; /* PT entry in PD of original pmap */
1706 pd_entry_t opte; /* contents of *pt */
1707 pd_entry_t npte; /* contents of *pt */
1708 vm_page_t m;
1709
1710 /*
1711 * Basic tests, require a non-NULL vm_map_entry, require proper
1712 * alignment and type for the vm_map_entry, require that the
1713 * underlying object already be allocated.
1714 *
1715 * We currently allow any type of object to use this optimization.
1716 * The object itself does NOT have to be sized to a multiple of the
1717 * segment size, but the memory mapping does.
1718 */
1719 if (entry == NULL ||
1720 pmap_mmu_optimize == 0 || /* not enabled */
1721 ptepindex >= pmap_pd_pindex(0) || /* not terminal */
1722 entry->inheritance != VM_INHERIT_SHARE || /* not shared */
1723 entry->maptype != VM_MAPTYPE_NORMAL || /* weird map type */
1724 entry->object.vm_object == NULL || /* needs VM object */
1725 (entry->offset & SEG_MASK) || /* must be aligned */
1726 (entry->start & SEG_MASK)) {
1727 return(pmap_allocpte(pmap, ptepindex, pvpp));
1728 }
1729
1730 /*
1731 * Make sure the full segment can be represented.
1732 */
1733 b = va & ~(vm_offset_t)SEG_MASK;
1734 if (b < entry->start && b + SEG_SIZE > entry->end)
1735 return(pmap_allocpte(pmap, ptepindex, pvpp));
1736
1737 /*
1738 * If the full segment can be represented dive the VM object's
1739 * shared pmap, allocating as required.
1740 */
1741 object = entry->object.vm_object;
1742
1743 if (entry->protection & VM_PROT_WRITE)
1744 obpmapp = &object->md.pmap_rw;
1745 else
1746 obpmapp = &object->md.pmap_ro;
1747
1748 /*
1749 * We allocate what appears to be a normal pmap but because portions
1750 * of this pmap are shared with other unrelated pmaps we have to
1751 * set pm_active to point to all cpus.
1752 *
1753 * XXX Currently using pmap_spin to interlock the update, can't use
1754 * vm_object_hold/drop because the token might already be held
1755 * shared OR exclusive and we don't know.
1756 */
1757 while ((obpmap = *obpmapp) == NULL) {
1758 obpmap = kmalloc(sizeof(*obpmap), M_OBJPMAP, M_WAITOK|M_ZERO);
1759 pmap_pinit_simple(obpmap);
1760 pmap_pinit2(obpmap);
1761 spin_lock(&pmap_spin);
1762 if (*obpmapp != NULL) {
1763 /*
1764 * Handle race
1765 */
1766 spin_unlock(&pmap_spin);
1767 pmap_release(obpmap);
1768 pmap_puninit(obpmap);
1769 kfree(obpmap, M_OBJPMAP);
1770 } else {
1771 obpmap->pm_active = smp_active_mask;
1772 *obpmapp = obpmap;
1773 spin_unlock(&pmap_spin);
1774 }
1775 }
1776
1777 /*
1778 * Layering is: PTE, PT, PD, PDP, PML4. We have to return the
1779 * pte/pt using the shared pmap from the object but also adjust
1780 * the process pmap's page table page as a side effect.
1781 */
1782
1783 /*
1784 * Resolve the terminal PTE and PT in the shared pmap. This is what
1785 * we will return. This is true if ptepindex represents a terminal
1786 * page, otherwise pte_pv is actually the PT and pt_pv is actually
1787 * the PD.
1788 */
1789 pt_pv = NULL;
1790 pte_pv = pmap_allocpte(obpmap, ptepindex, &pt_pv);
1791 if (ptepindex >= pmap_pt_pindex(0))
1792 xpv = pte_pv;
1793 else
1794 xpv = pt_pv;
1795
1796 /*
1797 * Resolve the PD in the process pmap so we can properly share the
1798 * page table page. Lock order is bottom-up (leaf first)!
1799 *
1800 * NOTE: proc_pt_pv can be NULL.
1801 */
1802 proc_pt_pv = pv_get(pmap, pmap_pt_pindex(b));
1803 proc_pd_pv = pmap_allocpte(pmap, pmap_pd_pindex(b), NULL);
1804
1805 /*
1806 * xpv is the page table page pv from the shared object
1807 * (for convenience).
1808 *
1809 * Calculate the pte value for the PT to load into the process PD.
1810 * If we have to change it we must properly dispose of the previous
1811 * entry.
1812 */
1813 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
1814 npte = VM_PAGE_TO_PHYS(xpv->pv_m) |
1815 (PG_U | PG_RW | PG_V | PG_A | PG_M);
99c2cc55
MD
1816
1817 /*
1818 * Dispose of previous entry if it was local to the process pmap.
1819 * (This should zero-out *pt)
1820 */
1821 if (proc_pt_pv) {
1822 pmap_release_pv(proc_pt_pv);
1823 proc_pt_pv = NULL;
1824 /* relookup */
1825 pt = pv_pte_lookup(proc_pd_pv, pmap_pt_index(b));
1826 }
1827
1828 /*
1829 * Handle remaining cases.
1830 */
921c891e
MD
1831 if (*pt == 0) {
1832 *pt = npte;
1833 vm_page_wire_quick(xpv->pv_m);
1834 vm_page_wire_quick(proc_pd_pv->pv_m);
1835 atomic_add_long(&pmap->pm_stats.resident_count, 1);
1836 } else if (*pt != npte) {
1837 pmap_inval_init(&info);
1838 pmap_inval_interlock(&info, pmap, (vm_offset_t)-1);
921c891e 1839
99c2cc55
MD
1840 opte = pte_load_clear(pt);
1841 KKASSERT(opte && opte != npte);
1842
1843 *pt = npte;
1844 vm_page_wire_quick(xpv->pv_m); /* pgtable pg that is npte */
1845
1846 /*
1847 * Clean up opte, bump the wire_count for the process
1848 * PD page representing the new entry if it was
1849 * previously empty.
1850 *
1851 * If the entry was not previously empty and we have
1852 * a PT in the proc pmap then opte must match that
1853 * pt. The proc pt must be retired (this is done
1854 * later on in this procedure).
1855 *
1856 * NOTE: replacing valid pte, wire_count on proc_pd_pv
1857 * stays the same.
1858 */
1859 KKASSERT(opte & PG_V);
1860 m = PHYS_TO_VM_PAGE(opte & PG_FRAME);
1861 if (vm_page_unwire_quick(m)) {
1862 panic("pmap_allocpte_seg: "
1863 "bad wire count %p",
1864 m);
921c891e 1865 }
99c2cc55 1866
921c891e
MD
1867 pmap_inval_deinterlock(&info, pmap);
1868 pmap_inval_done(&info);
921c891e
MD
1869 }
1870
1871 /*
1872 * The existing process page table was replaced and must be destroyed
1873 * here.
1874 */
1875 if (proc_pd_pv)
1876 pv_put(proc_pd_pv);
921c891e
MD
1877 if (pvpp)
1878 *pvpp = pt_pv;
1879 else
1880 pv_put(pt_pv);
1881
1882 return (pte_pv);
1883}
1884
1885/*
701c977e
MD
1886 * Release any resources held by the given physical map.
1887 *
1888 * Called when a pmap initialized by pmap_pinit is being released. Should
1889 * only be called if the map contains no valid mappings.
b12defdc 1890 *
701c977e 1891 * Caller must hold pmap->pm_token
d7f50089 1892 */
701c977e
MD
1893struct pmap_release_info {
1894 pmap_t pmap;
1895 int retry;
1896};
1897
1898static int pmap_release_callback(pv_entry_t pv, void *data);
1899
1900void
1901pmap_release(struct pmap *pmap)
c8fe38ae 1902{
701c977e
MD
1903 struct pmap_release_info info;
1904
1905 KASSERT(pmap->pm_active == 0,
1906 ("pmap still active! %016jx", (uintmax_t)pmap->pm_active));
701c977e
MD
1907
1908 spin_lock(&pmap_spin);
1909 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1910 spin_unlock(&pmap_spin);
c8fe38ae
MD
1911
1912 /*
701c977e
MD
1913 * Pull pv's off the RB tree in order from low to high and release
1914 * each page.
c8fe38ae 1915 */
701c977e
MD
1916 info.pmap = pmap;
1917 do {
1918 info.retry = 0;
1919 spin_lock(&pmap->pm_spin);
1920 RB_SCAN(pv_entry_rb_tree, &pmap->pm_pvroot, NULL,
1921 pmap_release_callback, &info);
1922 spin_unlock(&pmap->pm_spin);
1923 } while (info.retry);
1924
a5fc46c9
MD
1925
1926 /*
701c977e
MD
1927 * One resident page (the pml4 page) should remain.
1928 * No wired pages should remain.
a5fc46c9 1929 */
921c891e
MD
1930 KKASSERT(pmap->pm_stats.resident_count ==
1931 ((pmap->pm_flags & PMAP_FLAG_SIMPLE) ? 0 : 1));
1932
701c977e
MD
1933 KKASSERT(pmap->pm_stats.wired_count == 0);
1934}
1935
1936static int
1937pmap_release_callback(pv_entry_t pv, void *data)
1938{
1939 struct pmap_release_info *info = data;
1940 pmap_t pmap = info->pmap;
921c891e 1941 int r;
701c977e
MD
1942
1943 if (pv_hold_try(pv)) {
1944 spin_unlock(&pmap->pm_spin);
1945 } else {
1946 spin_unlock(&pmap->pm_spin);
1947 pv_lock(pv);
1948 if (pv->pv_pmap != pmap) {
1949 pv_put(pv);
1950 spin_lock(&pmap->pm_spin);
1951 info->retry = 1;
1952 return(-1);
a5fc46c9 1953 }
48ffc236 1954 }
921c891e
MD
1955 r = pmap_release_pv(pv);
1956 spin_lock(&pmap->pm_spin);
1957 return(r);
1958}
1959
1960/*
1961 * Called with held (i.e. also locked) pv. This function will dispose of
1962 * the lock along with the pv.
1963 */
1964static int
1965pmap_release_pv(pv_entry_t pv)
1966{
1967 vm_page_t p;
48ffc236 1968
701c977e
MD
1969 /*
1970 * The pmap is currently not spinlocked, pv is held+locked.
1971 * Remove the pv's page from its parent's page table. The
1972 * parent's page table page's wire_count will be decremented.
1973 */
1974 pmap_remove_pv_pte(pv, NULL, NULL);
c8fe38ae
MD
1975
1976 /*
701c977e
MD
1977 * Terminal pvs are unhooked from their vm_pages. Because
1978 * terminal pages aren't page table pages they aren't wired
1979 * by us, so we have to be sure not to unwire them either.
c8fe38ae 1980 */
701c977e 1981 if (pv->pv_pindex < pmap_pt_pindex(0)) {
52bb73bc 1982 pmap_remove_pv_page(pv);
701c977e
MD
1983 goto skip;
1984 }
c8fe38ae 1985
c8fe38ae 1986 /*
701c977e
MD
1987 * We leave the top-level page table page cached, wired, and
1988 * mapped in the pmap until the dtor function (pmap_puninit())
1989 * gets called.
e8510e54 1990 *
701c977e
MD
1991 * Since we are leaving the top-level pv intact we need
1992 * to break out of what would otherwise be an infinite loop.
c8fe38ae 1993 */
701c977e
MD
1994 if (pv->pv_pindex == pmap_pml4_pindex()) {
1995 pv_put(pv);
701c977e
MD
1996 return(-1);
1997 }
1998
1999 /*
2000 * For page table pages (other than the top-level page),
2001 * remove and free the vm_page. The representitive mapping
2002 * removed above by pmap_remove_pv_pte() did not undo the
2003 * last wire_count so we have to do that as well.
2004 */
52bb73bc 2005 p = pmap_remove_pv_page(pv);
701c977e 2006 vm_page_busy_wait(p, FALSE, "pmaprl");
701c977e
MD
2007 if (p->wire_count != 1) {
2008 kprintf("p->wire_count was %016lx %d\n",
2009 pv->pv_pindex, p->wire_count);
2010 }
2011 KKASSERT(p->wire_count == 1);
2012 KKASSERT(p->flags & PG_UNMANAGED);
2013
2014 vm_page_unwire(p, 0);
2015 KKASSERT(p->wire_count == 0);
921c891e
MD
2016
2017 /*
2018 * Theoretically this page, if not the pml4 page, should contain
2019 * all-zeros. But its just too dangerous to mark it PG_ZERO. Free
2020 * normally.
2021 */
701c977e
MD
2022 vm_page_free(p);
2023skip:
2024 pv_free(pv);
921c891e 2025 return 0;
701c977e
MD
2026}
2027
2028/*
2029 * This function will remove the pte associated with a pv from its parent.
2030 * Terminal pv's are supported. The removal will be interlocked if info
2031 * is non-NULL. The caller must dispose of pv instead of just unlocking
2032 * it.
2033 *
2034 * The wire count will be dropped on the parent page table. The wire
2035 * count on the page being removed (pv->pv_m) from the parent page table
2036 * is NOT touched. Note that terminal pages will not have any additional
2037 * wire counts while page table pages will have at least one representing
2038 * the mapping, plus others representing sub-mappings.
2039 *
2040 * NOTE: Cannot be called on kernel page table pages, only KVM terminal
2041 * pages and user page table and terminal pages.
2042 *
2043 * The pv must be locked.
2044 *
2045 * XXX must lock parent pv's if they exist to remove pte XXX
2046 */
2047static
2048void
2049pmap_remove_pv_pte(pv_entry_t pv, pv_entry_t pvp, struct pmap_inval_info *info)
2050{
2051 vm_pindex_t ptepindex = pv->pv_pindex;
2052 pmap_t pmap = pv->pv_pmap;
2053 vm_page_t p;
2054 int gotpvp = 0;
48ffc236 2055
701c977e 2056 KKASSERT(pmap);
48ffc236 2057
701c977e 2058 if (ptepindex == pmap_pml4_pindex()) {
b12defdc 2059 /*
701c977e 2060 * We are the top level pml4 table, there is no parent.
b12defdc 2061 */
701c977e
MD
2062 p = pmap->pm_pmlpv->pv_m;
2063 } else if (ptepindex >= pmap_pdp_pindex(0)) {
e8510e54 2064 /*
701c977e
MD
2065 * Remove a PDP page from the pml4e. This can only occur
2066 * with user page tables. We do not have to lock the
2067 * pml4 PV so just ignore pvp.
e8510e54 2068 */
701c977e
MD
2069 vm_pindex_t pml4_pindex;
2070 vm_pindex_t pdp_index;
2071 pml4_entry_t *pdp;
2072
2073 pdp_index = ptepindex - pmap_pdp_pindex(0);
2074 if (pvp == NULL) {
2075 pml4_pindex = pmap_pml4_pindex();
2076 pvp = pv_get(pv->pv_pmap, pml4_pindex);
921c891e 2077 KKASSERT(pvp);
701c977e 2078 gotpvp = 1;
e8510e54 2079 }
701c977e
MD
2080 pdp = &pmap->pm_pml4[pdp_index & ((1ul << NPML4EPGSHIFT) - 1)];
2081 KKASSERT((*pdp & PG_V) != 0);
2082 p = PHYS_TO_VM_PAGE(*pdp & PG_FRAME);
2083 *pdp = 0;
2084 KKASSERT(info == NULL);
2085 } else if (ptepindex >= pmap_pd_pindex(0)) {
e8510e54 2086 /*
921c891e
MD
2087 * Remove a PD page from the pdp
2088 *
2089 * SIMPLE PMAP NOTE: Non-existant pvp's are ok in the case
2090 * of a simple pmap because it stops at
2091 * the PD page.
e8510e54 2092 */
701c977e
MD
2093 vm_pindex_t pdp_pindex;
2094 vm_pindex_t pd_index;
2095 pdp_entry_t *pd;
48ffc236 2096
701c977e 2097 pd_index = ptepindex - pmap_pd_pindex(0);
48ffc236 2098
701c977e
MD
2099 if (pvp == NULL) {
2100 pdp_pindex = NUPTE_TOTAL + NUPT_TOTAL + NUPD_TOTAL +
2101 (pd_index >> NPML4EPGSHIFT);
2102 pvp = pv_get(pv->pv_pmap, pdp_pindex);
921c891e
MD
2103 if (pvp)
2104 gotpvp = 1;
2105 }
2106 if (pvp) {
2107 pd = pv_pte_lookup(pvp, pd_index &
2108 ((1ul << NPDPEPGSHIFT) - 1));
2109 KKASSERT((*pd & PG_V) != 0);
2110 p = PHYS_TO_VM_PAGE(*pd & PG_FRAME);
2111 *pd = 0;
2112 } else {
2113 KKASSERT(pmap->pm_flags & PMAP_FLAG_SIMPLE);
2114 p = pv->pv_m; /* degenerate test later */
701c977e 2115 }
701c977e
MD
2116 KKASSERT(info == NULL);
2117 } else if (ptepindex >= pmap_pt_pindex(0)) {
e8510e54 2118 /*
701c977e 2119 * Remove a PT page from the pd
e8510e54 2120 */
701c977e
MD
2121 vm_pindex_t pd_pindex;
2122 vm_pindex_t pt_index;
2123 pd_entry_t *pt;
b12defdc 2124
701c977e
MD
2125 pt_index = ptepindex - pmap_pt_pindex(0);
2126
2127 if (pvp == NULL) {
2128 pd_pindex = NUPTE_TOTAL + NUPT_TOTAL +
2129 (pt_index >> NPDPEPGSHIFT);
2130 pvp = pv_get(pv->pv_pmap, pd_pindex);
921c891e 2131 KKASSERT(pvp);
701c977e
MD
2132 gotpvp = 1;
2133 }
2134 pt = pv_pte_lookup(pvp, pt_index & ((1ul << NPDPEPGSHIFT) - 1));
2135 KKASSERT((*pt & PG_V) != 0);
2136 p = PHYS_TO_VM_PAGE(*pt & PG_FRAME);
2137 *pt = 0;
2138 KKASSERT(info == NULL);
2139 } else {
b12defdc 2140 /*
701c977e 2141 * Remove a PTE from the PT page
b12defdc 2142 *
701c977e
MD
2143 * NOTE: pv's must be locked bottom-up to avoid deadlocking.
2144 * pv is a pte_pv so we can safely lock pt_pv.
b12defdc 2145 */
701c977e
MD
2146 vm_pindex_t pt_pindex;
2147 pt_entry_t *ptep;
2148 pt_entry_t pte;
2149 vm_offset_t va;
b12defdc 2150
701c977e
MD
2151 pt_pindex = ptepindex >> NPTEPGSHIFT;
2152 va = (vm_offset_t)ptepindex << PAGE_SHIFT;
2153
2154 if (ptepindex >= NUPTE_USER) {
2155 ptep = vtopte(ptepindex << PAGE_SHIFT);
2156 KKASSERT(pvp == NULL);
c8fe38ae 2157 } else {
701c977e
MD
2158 if (pvp == NULL) {
2159 pt_pindex = NUPTE_TOTAL +
2160 (ptepindex >> NPDPEPGSHIFT);
2161 pvp = pv_get(pv->pv_pmap, pt_pindex);
921c891e 2162 KKASSERT(pvp);
701c977e
MD
2163 gotpvp = 1;
2164 }
2165 ptep = pv_pte_lookup(pvp, ptepindex &
2166 ((1ul << NPDPEPGSHIFT) - 1));
c8fe38ae 2167 }
701c977e
MD
2168
2169 if (info)
2170 pmap_inval_interlock(info, pmap, va);
2171 pte = pte_load_clear(ptep);
2172 if (info)
2173 pmap_inval_deinterlock(info, pmap);
52bb73bc
MD
2174 else
2175 cpu_invlpg((void *)va);
48ffc236 2176
e8510e54 2177 /*
701c977e 2178 * Now update the vm_page_t
e8510e54 2179 */
701c977e
MD
2180 if ((pte & (PG_MANAGED|PG_V)) != (PG_MANAGED|PG_V)) {
2181 kprintf("remove_pte badpte %016lx %016lx %d\n",
2182 pte, pv->pv_pindex,
2183 pv->pv_pindex < pmap_pt_pindex(0));
2184 }
2185 /*KKASSERT((pte & (PG_MANAGED|PG_V)) == (PG_MANAGED|PG_V));*/
2186 p = PHYS_TO_VM_PAGE(pte & PG_FRAME);
2187
2188 if (pte & PG_M) {
2189 if (pmap_track_modified(ptepindex))
2190 vm_page_dirty(p);
2191 }
2192 if (pte & PG_A) {
2193 vm_page_flag_set(p, PG_REFERENCED);
e8510e54 2194 }
701c977e
MD
2195 if (pte & PG_W)
2196 atomic_add_long(&pmap->pm_stats.wired_count, -1);
2197 if (pte & PG_G)
2198 cpu_invlpg((void *)va);
c8fe38ae
MD
2199 }
2200
48ffc236 2201 /*
701c977e
MD
2202 * Unwire the parent page table page. The wire_count cannot go below
2203 * 1 here because the parent page table page is itself still mapped.
2204 *
2205 * XXX remove the assertions later.
48ffc236 2206 */
701c977e
MD
2207 KKASSERT(pv->pv_m == p);
2208 if (pvp && vm_page_unwire_quick(pvp->pv_m))
2209 panic("pmap_remove_pv_pte: Insufficient wire_count");
c8fe38ae 2210
701c977e
MD
2211 if (gotpvp)
2212 pv_put(pvp);
c8fe38ae
MD
2213}
2214
bfc09ba0
MD
2215static
2216vm_page_t
52bb73bc 2217pmap_remove_pv_page(pv_entry_t pv)
d7f50089 2218{
c8fe38ae
MD
2219 vm_page_t m;
2220
701c977e 2221 m = pv->pv_m;
701c977e
MD
2222 KKASSERT(m);
2223 vm_page_spin_lock(m);
2224 pv->pv_m = NULL;
2225 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
c8fe38ae 2226 /*
701c977e
MD
2227 if (m->object)
2228 atomic_add_int(&m->object->agg_pv_list_count, -1);
2229 */
2230 if (TAILQ_EMPTY(&m->md.pv_list))
2231 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2232 vm_page_spin_unlock(m);
52bb73bc 2233 return(m);
d7f50089
YY
2234}
2235
2236/*
c8fe38ae 2237 * Grow the number of kernel page table entries, if needed.
a8cf2878
MD
2238 *
2239 * This routine is always called to validate any address space
2240 * beyond KERNBASE (for kldloads). kernel_vm_end only governs the address
2241 * space below KERNBASE.
d7f50089 2242 */
c8fe38ae 2243void
a8cf2878 2244pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
d7f50089 2245{
48ffc236 2246 vm_paddr_t paddr;
c8fe38ae
MD
2247 vm_offset_t ptppaddr;
2248 vm_page_t nkpg;
701c977e
MD
2249 pd_entry_t *pt, newpt;
2250 pdp_entry_t newpd;
a8cf2878 2251 int update_kernel_vm_end;
c8fe38ae 2252
a8cf2878
MD
2253 /*
2254 * bootstrap kernel_vm_end on first real VM use
2255 */
c8fe38ae 2256 if (kernel_vm_end == 0) {
791c6551 2257 kernel_vm_end = VM_MIN_KERNEL_ADDRESS;
c8fe38ae 2258 nkpt = 0;
701c977e 2259 while ((*pmap_pt(&kernel_pmap, kernel_vm_end) & PG_V) != 0) {
a8cf2878
MD
2260 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
2261 ~(PAGE_SIZE * NPTEPG - 1);
c8fe38ae 2262 nkpt++;
48ffc236
JG
2263 if (kernel_vm_end - 1 >= kernel_map.max_offset) {
2264 kernel_vm_end = kernel_map.max_offset;
2265 break;
2266 }
c8fe38ae
MD
2267 }
2268 }
a8cf2878
MD
2269
2270 /*
2271 * Fill in the gaps. kernel_vm_end is only adjusted for ranges
2272 * below KERNBASE. Ranges above KERNBASE are kldloaded and we
2273 * do not want to force-fill 128G worth of page tables.
2274 */
2275 if (kstart < KERNBASE) {
2276 if (kstart > kernel_vm_end)
2277 kstart = kernel_vm_end;
2278 KKASSERT(kend <= KERNBASE);
2279 update_kernel_vm_end = 1;
2280 } else {
2281 update_kernel_vm_end = 0;
2282 }
2283
2284 kstart = rounddown2(kstart, PAGE_SIZE * NPTEPG);
2285 kend = roundup2(kend, PAGE_SIZE * NPTEPG);
2286
2287 if (kend - 1 >= kernel_map.max_offset)
2288 kend = kernel_map.max_offset;
2289
2290 while (kstart < kend) {
701c977e
MD
2291 pt = pmap_pt(&kernel_pmap, kstart);
2292 if (pt == NULL) {
48ffc236 2293 /* We need a new PDP entry */
701c977e 2294 nkpg = vm_page_alloc(NULL, nkpt,
a8cf2878
MD
2295 VM_ALLOC_NORMAL |
2296 VM_ALLOC_SYSTEM |
2297 VM_ALLOC_INTERRUPT);
2298 if (nkpg == NULL) {
2299 panic("pmap_growkernel: no memory to grow "
2300 "kernel");
2301 }
48ffc236 2302 paddr = VM_PAGE_TO_PHYS(nkpg);
7f2a2740
MD
2303 if ((nkpg->flags & PG_ZERO) == 0)
2304 pmap_zero_page(paddr);
2305 vm_page_flag_clear(nkpg, PG_ZERO);
701c977e 2306 newpd = (pdp_entry_t)
48ffc236 2307 (paddr | PG_V | PG_RW | PG_A | PG_M);
701c977e 2308 *pmap_pd(&kernel_pmap, kstart) = newpd;
7f2a2740 2309 nkpt++;
48ffc236
JG
2310 continue; /* try again */
2311 }
701c977e 2312 if ((*pt & PG_V) != 0) {
a8cf2878
MD
2313 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2314 ~(PAGE_SIZE * NPTEPG - 1);
2315 if (kstart - 1 >= kernel_map.max_offset) {
2316 kstart = kernel_map.max_offset;
48ffc236
JG
2317 break;
2318 }
c8fe38ae
MD
2319 continue;
2320 }
2321
2322 /*
2323 * This index is bogus, but out of the way
2324 */
701c977e 2325 nkpg = vm_page_alloc(NULL, nkpt,
a8cf2878
MD
2326 VM_ALLOC_NORMAL |
2327 VM_ALLOC_SYSTEM |
2328 VM_ALLOC_INTERRUPT);
c8fe38ae
MD
2329 if (nkpg == NULL)
2330 panic("pmap_growkernel: no memory to grow kernel");
2331
2332 vm_page_wire(nkpg);
2333 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
2334 pmap_zero_page(ptppaddr);
7f2a2740 2335 vm_page_flag_clear(nkpg, PG_ZERO);
701c977e
MD
2336 newpt = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
2337 *pmap_pt(&kernel_pmap, kstart) = newpt;
c8fe38ae
MD
2338 nkpt++;
2339
a8cf2878
MD
2340 kstart = (kstart + PAGE_SIZE * NPTEPG) &
2341 ~(PAGE_SIZE * NPTEPG - 1);
2342
2343 if (kstart - 1 >= kernel_map.max_offset) {
2344 kstart = kernel_map.max_offset;
48ffc236 2345 break;
c8fe38ae 2346 }
c8fe38ae 2347 }
a8cf2878
MD
2348
2349 /*
2350 * Only update kernel_vm_end for areas below KERNBASE.
2351 */
2352 if (update_kernel_vm_end && kernel_vm_end < kstart)
2353 kernel_vm_end = kstart;
d7f50089
YY
2354}
2355
2356/*
921c891e 2357 * Add a reference to the specified pmap.
d7f50089 2358 */
c8fe38ae 2359void
921c891e 2360pmap_reference(pmap_t pmap)
d7f50089 2361{
921c891e
MD
2362 if (pmap != NULL) {
2363 lwkt_gettoken(&pmap->pm_token);
2364 ++pmap->pm_count;
2365 lwkt_reltoken(&pmap->pm_token);
c8fe38ae 2366 }
d7f50089
YY
2367}
2368
c8fe38ae 2369void
921c891e 2370pmap_drop(pmap_t pmap)
d7f50089 2371{
c8fe38ae 2372 if (pmap != NULL) {
b12defdc 2373 lwkt_gettoken(&pmap->pm_token);
921c891e 2374 --pmap->pm_count;
b12defdc 2375 lwkt_reltoken(&pmap->pm_token);
c8fe38ae 2376 }
d7f50089
YY
2377}
2378
c8fe38ae 2379/***************************************************
701c977e 2380 * page management routines.
c8fe38ae 2381 ***************************************************/
d7f50089
YY
2382
2383/*
701c977e 2384 * Hold a pv without locking it
d7f50089 2385 */
701c977e
MD
2386static void
2387pv_hold(pv_entry_t pv)
d7f50089 2388{
701c977e
MD
2389 u_int count;
2390
2391 if (atomic_cmpset_int(&pv->pv_hold, 0, 1))
2392 return;
2393
2394 for (;;) {
2395 count = pv->pv_hold;
2396 cpu_ccfence();
2397 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2398 return;
2399 /* retry */
2400 }
d7f50089
YY
2401}
2402
2403/*
701c977e
MD
2404 * Hold a pv_entry, preventing its destruction. TRUE is returned if the pv
2405 * was successfully locked, FALSE if it wasn't. The caller must dispose of
2406 * the pv properly.
2407 *
2408 * Either the pmap->pm_spin or the related vm_page_spin (if traversing a
2409 * pv list via its page) must be held by the caller.
d7f50089 2410 */
701c977e
MD
2411static int
2412_pv_hold_try(pv_entry_t pv PMAP_DEBUG_DECL)
d7f50089 2413{
701c977e
MD
2414 u_int count;
2415
2416 if (atomic_cmpset_int(&pv->pv_hold, 0, PV_HOLD_LOCKED | 1)) {
2417#ifdef PMAP_DEBUG
2418 pv->pv_func = func;
2419 pv->pv_line = lineno;
2420#endif
2421 return TRUE;
2422 }
2423
2424 for (;;) {
2425 count = pv->pv_hold;
2426 cpu_ccfence();
2427 if ((count & PV_HOLD_LOCKED) == 0) {
2428 if (atomic_cmpset_int(&pv->pv_hold, count,
2429 (count + 1) | PV_HOLD_LOCKED)) {
2430#ifdef PMAP_DEBUG
2431 pv->pv_func = func;
2432 pv->pv_line = lineno;
2433#endif
2434 return TRUE;
2435 }
2436 } else {
2437 if (atomic_cmpset_int(&pv->pv_hold, count, count + 1))
2438 return FALSE;
2439 }
2440 /* retry */
c8fe38ae 2441 }
d7f50089
YY
2442}
2443
2444/*
701c977e
MD
2445 * Drop a previously held pv_entry which could not be locked, allowing its
2446 * destruction.
2447 *
2448 * Must not be called with a spinlock held as we might zfree() the pv if it
2449 * is no longer associated with a pmap and this was the last hold count.
d7f50089 2450 */
701c977e
MD
2451static void
2452pv_drop(pv_entry_t pv)
d7f50089 2453{
701c977e 2454 u_int count;
c8fe38ae 2455
701c977e
MD
2456 if (atomic_cmpset_int(&pv->pv_hold, 1, 0)) {
2457 if (pv->pv_pmap == NULL)
2458 zfree(pvzone, pv);
c8fe38ae 2459 return;
c8fe38ae
MD
2460 }
2461
701c977e
MD
2462 for (;;) {
2463 count = pv->pv_hold;
2464 cpu_ccfence();
2465 KKASSERT((count & PV_HOLD_MASK) > 0);
2466 KKASSERT((count & (PV_HOLD_LOCKED | PV_HOLD_MASK)) !=
2467 (PV_HOLD_LOCKED | 1));
2468 if (atomic_cmpset_int(&pv->pv_hold, count, count - 1)) {
2469 if (count == 1 && pv->pv_pmap == NULL)
2470 zfree(pvzone, pv);
2471 return;
b12defdc 2472 }
701c977e 2473 /* retry */
c8fe38ae 2474 }
d7f50089 2475}
c8fe38ae 2476
d7f50089 2477/*
701c977e 2478 * Find or allocate the requested PV entry, returning a locked pv
d7f50089 2479 */
bfc09ba0 2480static
701c977e
MD
2481pv_entry_t
2482_pv_alloc(pmap_t pmap, vm_pindex_t pindex, int *isnew PMAP_DEBUG_DECL)
c8fe38ae
MD
2483{
2484 pv_entry_t pv;
701c977e 2485 pv_entry_t pnew = NULL;
c8fe38ae 2486
701c977e
MD
2487 spin_lock(&pmap->pm_spin);
2488 for (;;) {
2489 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2490 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2491 pindex);
c8fe38ae 2492 }
701c977e
MD
2493 if (pv == NULL) {
2494 if (pnew == NULL) {
2495 spin_unlock(&pmap->pm_spin);
2496 pnew = zalloc(pvzone);
2497 spin_lock(&pmap->pm_spin);
2498 continue;
2499 }
2500 pnew->pv_pmap = pmap;
2501 pnew->pv_pindex = pindex;
2502 pnew->pv_hold = PV_HOLD_LOCKED | 1;
2503#ifdef PMAP_DEBUG
2504 pnew->pv_func = func;
2505 pnew->pv_line = lineno;
2506#endif
2507 pv_entry_rb_tree_RB_INSERT(&pmap->pm_pvroot, pnew);
2508 atomic_add_long(&pmap->pm_stats.resident_count, 1);
2509 spin_unlock(&pmap->pm_spin);
2510 *isnew = 1;
2511 return(pnew);
2512 }
2513 if (pnew) {
2514 spin_unlock(&pmap->pm_spin);
2515 zfree(pvzone, pnew);
2516 pnew = NULL;
2517 spin_lock(&pmap->pm_spin);
2518 continue;
2519 }
2520 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2521 spin_unlock(&pmap->pm_spin);
2522 *isnew = 0;
2523 return(pv);
2524 }
2525 spin_unlock(&pmap->pm_spin);
2526 _pv_lock(pv PMAP_DEBUG_COPY);
2527 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex) {
2528 *isnew = 0;
2529 return(pv);
2530 }
2531 pv_put(pv);
2532 spin_lock(&pmap->pm_spin);
2533 }
c8fe38ae 2534
5926987a 2535
701c977e 2536}
b12defdc 2537
701c977e
MD
2538/*
2539 * Find the requested PV entry, returning a locked+held pv or NULL
2540 */
2541static
2542pv_entry_t
2543_pv_get(pmap_t pmap, vm_pindex_t pindex PMAP_DEBUG_DECL)
2544{
2545 pv_entry_t pv;
5926987a 2546
701c977e
MD
2547 spin_lock(&pmap->pm_spin);
2548 for (;;) {
2549 /*
2550 * Shortcut cache
2551 */
2552 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex) {
2553 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot,
2554 pindex);
2555 }
2556 if (pv == NULL) {
2557 spin_unlock(&pmap->pm_spin);
2558 return NULL;
2559 }
2560 if (_pv_hold_try(pv PMAP_DEBUG_COPY)) {
2561 pv_cache(pv, pindex);
2562 spin_unlock(&pmap->pm_spin);
2563 return(pv);
2564 }
2565 spin_unlock(&pmap->pm_spin);
2566 _pv_lock(pv PMAP_DEBUG_COPY);
2567 if (pv->pv_pmap == pmap && pv->pv_pindex == pindex)
2568 return(pv);
2569 pv_put(pv);
2570 spin_lock(&pmap->pm_spin);
2571 }
d7f50089
YY
2572}
2573
2574/*
701c977e
MD
2575 * Lookup, hold, and attempt to lock (pmap,pindex).
2576 *
2577 * If the entry does not exist NULL is returned and *errorp is set to 0
a5fc46c9 2578 *
701c977e
MD
2579 * If the entry exists and could be successfully locked it is returned and
2580 * errorp is set to 0.
2581 *
2582 * If the entry exists but could NOT be successfully locked it is returned
2583 * held and *errorp is set to 1.
d7f50089 2584 */
bfc09ba0 2585static
701c977e
MD
2586pv_entry_t
2587pv_get_try(pmap_t pmap, vm_pindex_t pindex, int *errorp)
d7f50089 2588{
c8fe38ae
MD
2589 pv_entry_t pv;
2590
701c977e
MD
2591 spin_lock(&pmap->pm_spin);
2592 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2593 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2594 if (pv == NULL) {
2595 spin_unlock(&pmap->pm_spin);
2596 *errorp = 0;
2597 return NULL;
2598 }
2599 if (pv_hold_try(pv)) {
2600 pv_cache(pv, pindex);
2601 spin_unlock(&pmap->pm_spin);
2602 *errorp = 0;
2603 return(pv); /* lock succeeded */
2604 }
2605 spin_unlock(&pmap->pm_spin);
2606 *errorp = 1;
2607 return (pv); /* lock failed */
d7f50089
YY
2608}
2609
2610/*
701c977e 2611 * Find the requested PV entry, returning a held pv or NULL
d7f50089 2612 */
bfc09ba0 2613static
701c977e
MD
2614pv_entry_t
2615pv_find(pmap_t pmap, vm_pindex_t pindex)
c8fe38ae 2616{
701c977e 2617 pv_entry_t pv;
c8fe38ae 2618
701c977e 2619 spin_lock(&pmap->pm_spin);
b12defdc 2620
701c977e
MD
2621 if ((pv = pmap->pm_pvhint) == NULL || pv->pv_pindex != pindex)
2622 pv = pv_entry_rb_tree_RB_LOOKUP(&pmap->pm_pvroot, pindex);
2623 if (pv == NULL) {
2624 spin_unlock(&pmap->pm_spin);
2625 return NULL;
2626 }
2627 pv_hold(pv);
2628 pv_cache(pv, pindex);
2629 spin_unlock(&pmap->pm_spin);
2630 return(pv);
2631}
2632
2633/*
2634 * Lock a held pv, keeping the hold count
2635 */
2636static
2637void
2638_pv_lock(pv_entry_t pv PMAP_DEBUG_DECL)
2639{
2640 u_int count;
2641
2642 for (;;) {
2643 count = pv->pv_hold;
2644 cpu_ccfence();
2645 if ((count & PV_HOLD_LOCKED) == 0) {
2646 if (atomic_cmpset_int(&pv->pv_hold, count,
2647 count | PV_HOLD_LOCKED)) {
2648#ifdef PMAP_DEBUG
2649 pv->pv_func = func;
2650 pv->pv_line = lineno;
2651#endif
2652 return;
c8fe38ae 2653 }
701c977e
MD
2654 continue;
2655 }
2656 tsleep_interlock(pv, 0);
2657 if (atomic_cmpset_int(&pv->pv_hold, count,
2658 count | PV_HOLD_WAITING)) {
2659#ifdef PMAP_DEBUG
2660 kprintf("pv waiting on %s:%d\n",
2661 pv->pv_func, pv->pv_line);
c8fe38ae 2662#endif
701c977e 2663 tsleep(pv, PINTERLOCKED, "pvwait", hz);
c8fe38ae 2664 }
701c977e 2665 /* retry */
b12defdc 2666 }
701c977e 2667}
c8fe38ae 2668
701c977e
MD
2669/*
2670 * Unlock a held and locked pv, keeping the hold count.
2671 */
2672static
2673void
2674pv_unlock(pv_entry_t pv)
2675{
2676 u_int count;
2677
2678 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 1))
2679 return;
2680
2681 for (;;) {
2682 count = pv->pv_hold;
2683 cpu_ccfence();
2684 KKASSERT((count & (PV_HOLD_LOCKED|PV_HOLD_MASK)) >=
2685 (PV_HOLD_LOCKED | 1));
2686 if (atomic_cmpset_int(&pv->pv_hold, count,
2687 count &
2688 ~(PV_HOLD_LOCKED | PV_HOLD_WAITING))) {
2689 if (count & PV_HOLD_WAITING)
2690 wakeup(pv);
2691 break;
2692 }
7ab91d55 2693 }
d7f50089
YY
2694}
2695
2696/*
701c977e
MD
2697 * Unlock and drop a pv. If the pv is no longer associated with a pmap
2698 * and the hold count drops to zero we will free it.
d7f50089 2699 *
701c977e
MD
2700 * Caller should not hold any spin locks. We are protected from hold races
2701 * by virtue of holds only occuring only with a pmap_spin or vm_page_spin
2702 * lock held. A pv cannot be located otherwise.
d7f50089 2703 */
bfc09ba0
MD
2704static
2705void
701c977e 2706pv_put(pv_entry_t pv)
c8fe38ae 2707{
701c977e
MD
2708 if (atomic_cmpset_int(&pv->pv_hold, PV_HOLD_LOCKED | 1, 0)) {
2709 if (pv->pv_pmap == NULL)
2710 zfree(pvzone, pv);
2711 return;
2712 }
2713 pv_unlock(pv);
2714 pv_drop(pv);
2715}
c8fe38ae 2716
701c977e
MD
2717/*
2718 * Unlock, drop, and free a pv, destroying it. The pv is removed from its
2719 * pmap. Any pte operations must have already been completed.
2720 */
2721static
2722void
2723pv_free(pv_entry_t pv)
2724{
2725 pmap_t pmap;
b12defdc 2726
701c977e
MD
2727 KKASSERT(pv->pv_m == NULL);
2728 if ((pmap = pv->pv_pmap) != NULL) {
2729 spin_lock(&pmap->pm_spin);
2730 pv_entry_rb_tree_RB_REMOVE(&pmap->pm_pvroot, pv);
2731 if (pmap->pm_pvhint == pv)
2732 pmap->pm_pvhint = NULL;
2733 atomic_add_long(&pmap->pm_stats.resident_count, -1);
2734 pv->pv_pmap = NULL;
2735 pv->pv_pindex = 0;
2736 spin_unlock(&pmap->pm_spin);
2737 }
2738 pv_put(pv);
2739}
2740
2741/*
2742 * This routine is very drastic, but can save the system
2743 * in a pinch.
2744 */
2745void
2746pmap_collect(void)
2747{
2748 int i;
2749 vm_page_t m;
2750 static int warningdone=0;
2751
2752 if (pmap_pagedaemon_waken == 0)
48ffc236 2753 return;
701c977e
MD
2754 pmap_pagedaemon_waken = 0;
2755 if (warningdone < 5) {
2756 kprintf("pmap_collect: collecting pv entries -- "
2757 "suggest increasing PMAP_SHPGPERPROC\n");
2758 warningdone++;
2759 }
2760
2761 for (i = 0; i < vm_page_array_size; i++) {
2762 m = &vm_page_array[i];
2763 if (m->wire_count || m->hold_count)
2764 continue;
2765 if (vm_page_busy_try(m, TRUE) == 0) {
2766 if (m->wire_count == 0 && m->hold_count == 0) {
2767 pmap_remove_all(m);
2768 }
2769 vm_page_wakeup(m);
2770 }
2771 }
d7f50089
YY
2772}
2773
2774/*
701c977e 2775 * Scan the pmap for active page table entries and issue a callback.
921c891e
MD
2776 * The callback must dispose of pte_pv, whos PTE entry is at *ptep in
2777 * its parent page table.
d7f50089 2778 *
fb4ca018 2779 * pte_pv will be NULL if the page or page table is unmanaged.
921c891e 2780 * pt_pv will point to the page table page containing the pte for the page.
701c977e 2781 *
921c891e
MD
2782 * NOTE! If we come across an unmanaged page TABLE (verses an unmanaged page),
2783 * we pass a NULL pte_pv and we pass a pt_pv pointing to the passed
2784 * process pmap's PD and page to the callback function. This can be
2785 * confusing because the pt_pv is really a pd_pv, and the target page
2786 * table page is simply aliased by the pmap and not owned by it.
d7f50089 2787 *
701c977e 2788 * It is assumed that the start and end are properly rounded to the page size.
fb4ca018
MD
2789 *
2790 * It is assumed that PD pages and above are managed and thus in the RB tree,
2791 * allowing us to use RB_SCAN from the PD pages down for ranged scans.
2792 */
2793struct pmap_scan_info {
2794 struct pmap *pmap;
2795 vm_offset_t sva;
2796 vm_offset_t eva;
2797 vm_pindex_t sva_pd_pindex;
2798 vm_pindex_t eva_pd_pindex;
2799 void (*func)(pmap_t, struct pmap_inval_info *,
2800 pv_entry_t, pv_entry_t, int, vm_offset_t,
2801 pt_entry_t *, void *);
2802 void *arg;
2803 struct pmap_inval_info inval;
2804};
2805
2806static int pmap_scan_cmp(pv_entry_t pv, void *data);
2807static int pmap_scan_callback(pv_entry_t pv, void *data);
2808
701c977e 2809static void
fb4ca018 2810pmap_scan(struct pmap_scan_info *info)
701c977e 2811{
fb4ca018 2812 struct pmap *pmap = info->pmap;
701c977e
MD
2813 pv_entry_t pd_pv; /* A page directory PV */
2814 pv_entry_t pt_pv; /* A page table PV */
2815 pv_entry_t pte_pv; /* A page table entry PV */
2816 pt_entry_t *ptep;
fb4ca018 2817 struct pv_entry dummy_pv;
c8fe38ae
MD
2818
2819 if (pmap == NULL)
2820 return;
2821
701c977e
MD
2822 /*
2823 * Hold the token for stability; if the pmap is empty we have nothing
2824 * to do.
2825 */
b12defdc 2826 lwkt_gettoken(&pmap->pm_token);
701c977e 2827#if 0
10d6182e 2828 if (pmap->pm_stats.resident_count == 0) {
b12defdc 2829 lwkt_reltoken(&pmap->pm_token);
c8fe38ae 2830 return;
10d6182e 2831 }
701c977e 2832#endif
c8fe38ae 2833
fb4ca018 2834 pmap_inval_init(&info->inval);
c8fe38ae
MD
2835
2836 /*
fb4ca018 2837 * Special handling for scanning one page, which is a very common
701c977e 2838 * operation (it is?).
fb4ca018 2839 *
701c977e 2840 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4
c8fe38ae 2841 */
fb4ca018
MD
2842 if (info->sva + PAGE_SIZE == info->eva) {
2843 if (info->sva >= VM_MAX_USER_ADDRESS) {
701c977e
MD
2844 /*
2845 * Kernel mappings do not track wire counts on
fb4ca018
MD
2846 * page table pages and only maintain pd_pv and
2847 * pte_pv levels so pmap_scan() works.
701c977e
MD
2848 */
2849 pt_pv = NULL;
fb4ca018
MD
2850 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
2851 ptep = vtopte(info->sva);
701c977e
MD
2852 } else {
2853 /*
921c891e
MD
2854 * User pages which are unmanaged will not have a
2855 * pte_pv. User page table pages which are unmanaged
2856 * (shared from elsewhere) will also not have a pt_pv.
2857 * The func() callback will pass both pte_pv and pt_pv
2858 * as NULL in that case.
701c977e 2859 */
fb4ca018
MD
2860 pte_pv = pv_get(pmap, pmap_pte_pindex(info->sva));
2861 pt_pv = pv_get(pmap, pmap_pt_pindex(info->sva));
701c977e
MD
2862 if (pt_pv == NULL) {
2863 KKASSERT(pte_pv == NULL);
fb4ca018 2864 pd_pv = pv_get(pmap, pmap_pd_pindex(info->sva));
921c891e
MD
2865 if (pd_pv) {
2866 ptep = pv_pte_lookup(pd_pv,
fb4ca018 2867 pmap_pt_index(info->sva));
921c891e 2868 if (*ptep) {
fb4ca018 2869 info->func(pmap, &info->inval,
921c891e 2870 NULL, pd_pv, 1,
fb4ca018
MD
2871 info->sva, ptep,
2872 info->arg);
921c891e
MD
2873 }
2874 pv_put(pd_pv);
2875 }
701c977e
MD
2876 goto fast_skip;
2877 }
fb4ca018 2878 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(info->sva));
701c977e
MD
2879 }
2880 if (*ptep == 0) {
f2c5d4ab
MD
2881 /*
2882 * Unlike the pv_find() case below we actually
2883 * acquired a locked pv in this case so any
2884 * race should have been resolved. It is expected
2885 * to not exist.
2886 */
701c977e
MD
2887 KKASSERT(pte_pv == NULL);
2888 } else if (pte_pv) {
23b4bd44
MD
2889 KASSERT((*ptep & (PG_MANAGED|PG_V)) == (PG_MANAGED|
2890 PG_V),
2891 ("bad *ptep %016lx sva %016lx pte_pv %p",
fb4ca018
MD
2892 *ptep, info->sva, pte_pv));
2893 info->func(pmap, &info->inval, pte_pv, pt_pv, 0,
2894 info->sva, ptep, info->arg);
701c977e 2895 } else {
23b4bd44
MD
2896 KASSERT((*ptep & (PG_MANAGED|PG_V)) == PG_V,
2897 ("bad *ptep %016lx sva %016lx pte_pv NULL",
fb4ca018
MD
2898 *ptep, info->sva));
2899 info->func(pmap, &info->inval, NULL, pt_pv, 0,
2900 info->sva, ptep, info->arg);
48ffc236 2901 }
701c977e
MD
2902 if (pt_pv)
2903 pv_put(pt_pv);
2904fast_skip:
fb4ca018 2905 pmap_inval_done(&info->inval);
701c977e
MD
2906 lwkt_reltoken(&pmap->pm_token);
2907 return;
c8fe38ae
MD
2908 }
2909
701c977e 2910 /*
fb4ca018
MD
2911 * Nominal scan case, RB_SCAN() for PD pages and iterate from
2912 * there.
2913 */
2914 info->sva_pd_pindex = pmap_pd_pindex(info->sva);
2915 info->eva_pd_pindex = pmap_pd_pindex(info->eva + NBPDP - 1);
2916
2917 if (info->sva >= VM_MAX_USER_ADDRESS) {
2918 /*
2919 * The kernel does not currently maintain any pv_entry's for
2920 * higher-level page tables.
2921 */
2922 bzero(&dummy_pv, sizeof(dummy_pv));
2923 dummy_pv.pv_pindex = info->sva_pd_pindex;
2924 spin_lock(&pmap->pm_spin);
2925 while (dummy_pv.pv_pindex < info->eva_pd_pindex) {
2926 pmap_scan_callback(&dummy_pv, info);
2927 ++dummy_pv.pv_pindex;
2928 }
2929 spin_unlock(&pmap->pm_spin);
2930 } else {
2931 /*
2932 * User page tables maintain local PML4, PDP, and PD
2933 * pv_entry's at the very least. PT pv's might be
2934 * unmanaged and thus not exist. PTE pv's might be
2935 * unmanaged and thus not exist.
2936 */
2937 spin_lock(&pmap->pm_spin);
2938 pv_entry_rb_tree_RB_SCAN(&pmap->pm_pvroot,
2939 pmap_scan_cmp, pmap_scan_callback, info);
2940 spin_unlock(&pmap->pm_spin);
2941 }
2942 pmap_inval_done(&info->inval);
2943 lwkt_reltoken(&pmap->pm_token);
2944}
2945
2946/*
2947 * WARNING! pmap->pm_spin held
2948 */
2949static int
2950pmap_scan_cmp(pv_entry_t pv, void *data)
2951{
2952 struct pmap_scan_info *info = data;
2953 if (pv->pv_pindex < info->sva_pd_pindex)
2954 return(-1);
2955 if (pv->pv_pindex >= info->eva_pd_pindex)
2956 return(1);
2957 return(0);
2958}
2959
2960/*
2961 * WARNING! pmap->pm_spin held
2962 */
2963static int
2964pmap_scan_callback(pv_entry_t pv, void *data)
2965{
2966 struct pmap_scan_info *info = data;
2967 struct pmap *pmap = info->pmap;
2968 pv_entry_t pd_pv; /* A page directory PV */
2969 pv_entry_t pt_pv; /* A page table PV */
2970 pv_entry_t pte_pv; /* A page table entry PV */
2971 pt_entry_t *ptep;
2972 vm_offset_t sva;
2973 vm_offset_t eva;
2974 vm_offset_t va_next;
2975 vm_pindex_t pd_pindex;
2976 int error;
2977
2978 /*
2979 * Pull the PD pindex from the pv before releasing the spinlock.
2980 *
2981 * WARNING: pv is faked for kernel pmap scans.
2982 */
2983 pd_pindex = pv->pv_pindex;
2984 spin_unlock(&pmap->pm_spin);
2985 pv = NULL; /* invalid after spinlock unlocked */
2986
2987 /*
2988 * Calculate the page range within the PD. SIMPLE pmaps are
2989 * direct-mapped for the entire 2^64 address space. Normal pmaps
2990 * reflect the user and kernel address space which requires
2991 * cannonicalization w/regards to converting pd_pindex's back
2992 * into addresses.
2993 */
2994 sva = (pd_pindex - NUPTE_TOTAL - NUPT_TOTAL) << PDPSHIFT;
2995 if ((pmap->pm_flags & PMAP_FLAG_SIMPLE) == 0 &&
2996 (sva & PML4_SIGNMASK)) {
2997 sva |= PML4_SIGNMASK;
2998 }
2999 eva = sva + NBPDP; /* can overflow */
3000 if (sva < info->sva)
3001 sva = info->sva;
3002 if (eva < info->sva || eva > info->eva)
3003 eva = info->eva;
3004
3005 /*
701c977e
MD
3006 * NOTE: kernel mappings do not track page table pages, only
3007 * terminal pages.
3008 *
3009 * NOTE: Locks must be ordered bottom-up. pte,pt,pd,pdp,pml4.
3010 * However, for the scan to be efficient we try to
3011 * cache items top-down.
3012 */
701c977e
MD
3013 pd_pv = NULL;
3014 pt_pv = NULL;
3015
48ffc236 3016 for (; sva < eva; sva = va_next) {
701c977e
MD
3017 lwkt_yield();
3018 if (sva >= VM_MAX_USER_ADDRESS) {
3019 if (pt_pv) {
3020 pv_put(pt_pv);
3021 pt_pv = NULL;
3022 }
3023 goto kernel_skip;
3024 }
3025
3026 /*
fb4ca018
MD
3027 * PD cache (degenerate case if we skip). It is possible
3028 * for the PD to not exist due to races. This is ok.
701c977e
MD
3029 */
3030 if (pd_pv == NULL) {
701c977e
MD
3031 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3032 } else if (pd_pv->pv_pindex != pmap_pd_pindex(sva)) {
3033 pv_put(pd_pv);
701c977e
MD
3034 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3035 }
3036 if (pd_pv == NULL) {
48ffc236
JG
3037 va_next = (sva + NBPDP) & ~PDPMASK;
3038 if (va_next < sva)
3039 va_next = eva;
3040 continue;
3041 }
c8fe38ae
MD
3042
3043 /*
701c977e 3044 * PT cache
c8fe38ae 3045 */
701c977e 3046 if (pt_pv == NULL) {
701c977e
MD
3047 if (pd_pv) {
3048 pv_put(pd_pv);
3049 pd_pv = NULL;
3050 }
3051 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3052 } else if (pt_pv->pv_pindex != pmap_pt_pindex(sva)) {
701c977e
MD
3053 if (pd_pv) {
3054 pv_put(pd_pv);
3055 pd_pv = NULL;
3056 }
3057 pv_put(pt_pv);
3058 pt_pv = pv_get(pmap, pmap_pt_pindex(sva));
3059 }
c8fe38ae
MD
3060
3061 /*
921c891e
MD
3062 * If pt_pv is NULL we either have an shared page table
3063 * page and must issue a callback specific to that case,
3064 * or there is no page table page.
3065 *
3066 * Either way we can skip the page table page.
c8fe38ae 3067 */
701c977e 3068 if (pt_pv == NULL) {
921c891e
MD
3069 /*
3070 * Possible unmanaged (shared from another pmap)
3071 * page table page.
3072 */
3073 if (pd_pv == NULL)
3074 pd_pv = pv_get(pmap, pmap_pd_pindex(sva));
3075 KKASSERT(pd_pv != NULL);
3076 ptep = pv_pte_lookup(pd_pv, pmap_pt_index(sva));
3077 if (*ptep & PG_V) {
fb4ca018
MD
3078 info->func(pmap, &info->inval, NULL, pd_pv, 1,
3079 sva, ptep, info->arg);
921c891e
MD
3080 }
3081
3082 /*
3083 * Done, move to next page table page.
3084 */
701c977e
MD
3085 va_next = (sva + NBPDR) & ~PDRMASK;
3086 if (va_next < sva)
3087 va_next = eva;
c8fe38ae 3088 continue;
701c977e 3089 }
c8fe38ae
MD
3090
3091 /*
701c977e
MD
3092 * From this point in the loop testing pt_pv for non-NULL
3093 * means we are in UVM, else if it is NULL we are in KVM.
fb4ca018
MD
3094 *
3095 * Limit our scan to either the end of the va represented
3096 * by the current page table page, or to the end of the
3097 * range being removed.
48ffc236 3098 */
701c977e
MD
3099kernel_skip:
3100 va_next = (sva + NBPDR) & ~PDRMASK;
3101 if (va_next < sva)
3102 va_next = eva;
fb4ca018
MD
3103 if (va_next > eva)
3104 va_next = eva;
48ffc236
JG
3105
3106 /*
701c977e
MD
3107 * Scan the page table for pages. Some pages may not be
3108 * managed (might not have a pv_entry).
3109 *
3110 * There is no page table management for kernel pages so
3111 * pt_pv will be NULL in that case, but otherwise pt_pv
3112 * is non-NULL, locked, and referenced.
c8fe38ae 3113 */
c8fe38ae 3114
f2c5d4ab
MD
3115 /*
3116 * At this point a non-NULL pt_pv means a UVA, and a NULL
3117 * pt_pv means a KVA.
3118 */
701c977e
MD
3119 if (pt_pv)
3120 ptep = pv_pte_lookup(pt_pv, pmap_pte_index(sva));
3121 else
3122 ptep = vtopte(sva);
3123
3124 while (sva < va_next) {
f2c5d4ab 3125 /*
90244566
MD
3126 * Acquire the related pte_pv, if any. If *ptep == 0
3127 * the related pte_pv should not exist, but if *ptep
3128 * is not zero the pte_pv may or may not exist (e.g.
3129 * will not exist for an unmanaged page).
f2c5d4ab 3130 *
90244566
MD
3131 * However a multitude of races are possible here.
3132 *
3133 * In addition, the (pt_pv, pte_pv) lock order is
3134 * backwards, so we have to be careful in aquiring
3135 * a properly locked pte_pv.
f2c5d4ab 3136 */
fc9ed34d 3137 lwkt_yield();
701c977e
MD
3138 if (pt_pv) {
3139 pte_pv = pv_get_try(pmap, pmap_pte_pindex(sva),
3140 &error);
3141 if (error) {
701c977e
MD
3142 if (pd_pv) {
3143 pv_put(pd_pv);
3144 pd_pv = NULL;
3145 }
3146 pv_put(pt_pv); /* must be non-NULL */
3147 pt_pv = NULL;
3148 pv_lock(pte_pv); /* safe to block now */
3149 pv_put(pte_pv);
3150 pte_pv = NULL;
3151 pt_pv = pv_get(pmap,
3152 pmap_pt_pindex(sva));
921c891e
MD
3153 /*
3154 * pt_pv reloaded, need new ptep
3155 */
3156 KKASSERT(pt_pv != NULL);
3157 ptep = pv_pte_lookup(pt_pv,
3158 pmap_pte_index(sva));
701c977e
MD
3159 continue;
3160 }
3161 } else {
3162 pte_pv = pv_get(pmap, pmap_pte_pindex(sva));
3163 }
3164
3165 /*
90244566 3166 * Ok, if *ptep == 0 we had better NOT have a pte_pv.
a505393f
MD
3167 */
3168 if (*ptep == 0) {
3169 if (pte_pv) {
90244566
MD
3170 kprintf("Unexpected non-NULL pte_pv "
3171 "%p pt_pv %p *ptep = %016lx\n",
3172 pte_pv, pt_pv, *ptep);
3173 panic("Unexpected non-NULL pte_pv");
a505393f 3174 }
90244566
MD
3175 sva += PAGE_SIZE;
3176 ++ptep;
a505393f
MD
3177 continue;
3178 }
3179
3180 /*
90244566
MD
3181 * Ready for the callback. The locked pte_pv (if any)
3182 * is consumed by the callback. pte_pv will exist if
3183 * the page is managed, and will not exist if it
3184 * isn't.
701c977e
MD
3185 */
3186 if (pte_pv) {
23b4bd44
MD
3187 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
3188 (PG_MANAGED|PG_V),
3189 ("bad *ptep %016lx sva %016lx "
3190 "pte_pv %p",
3191 *ptep, sva, pte_pv));
fb4ca018
MD
3192 info->func(pmap, &info->inval, pte_pv, pt_pv, 0,
3193 sva, ptep, info->arg);
701c977e 3194 } else {
23b4bd44
MD
3195 KASSERT((*ptep & (PG_MANAGED|PG_V)) ==
3196 PG_V,
3197 ("bad *ptep %016lx sva %016lx "
3198 "pte_pv NULL",
3199 *ptep, sva));
fb4ca018
MD
3200 info->func(pmap, &info->inval, NULL, pt_pv, 0,
3201 sva, ptep, info->arg);
701c977e 3202 }
f2c5d4ab 3203 pte_pv = NULL;
701c977e
MD
3204 sva += PAGE_SIZE;
3205 ++ptep;
c8fe38ae
MD
3206 }
3207 }
701c977e
MD
3208 if (pd_pv) {
3209 pv_put(pd_pv);
3210 pd_pv = NULL;
3211 }
3212 if (pt_pv) {
3213 pv_put(pt_pv);
3214 pt_pv = NULL;
3215 }
fb4ca018
MD
3216
3217 /*
3218 * Relock before returning.
3219 */
3220 spin_lock(&pmap->pm_spin);
3221 return (0);
701c977e
MD
3222}
3223
3224void
3225pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
3226{
fb4ca018
MD
3227 struct pmap_scan_info info;
3228
3229 info.pmap = pmap;
3230 info.sva = sva;
3231 info.eva = eva;
3232 info.func = pmap_remove_callback;
3233 info.arg = NULL;
3234 pmap_scan(&info);
701c977e
MD
3235}
3236
3237static void
3238pmap_remove_callback(pmap_t pmap, struct pmap_inval_info *info,
921c891e
MD
3239 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3240 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
701c977e
MD
3241{
3242 pt_entry_t pte;
3243
3244 if (pte_pv) {
3245 /*
3246 * This will also drop pt_pv's wire_count. Note that
3247 * terminal pages are not wired based on mmu presence.
3248 */
3249 pmap_remove_pv_pte(pte_pv, pt_pv, info);
52bb73bc 3250 pmap_remove_pv_page(pte_pv);
701c977e 3251 pv_free(pte_pv);
921c891e 3252 } else if (sharept == 0) {
701c977e 3253 /*
921c891e
MD
3254 * Unmanaged page
3255 *
701c977e
MD
3256 * pt_pv's wire_count is still bumped by unmanaged pages
3257 * so we must decrement it manually.
3258 */
3259 pmap_inval_interlock(info, pmap, va);
3260 pte = pte_load_clear(ptep);
3261 pmap_inval_deinterlock(info, pmap);
3262 if (pte & PG_W)
3263 atomic_add_long(&pmap->pm_stats.wired_count, -1);
3264 atomic_add_long(&pmap->pm_stats.resident_count, -1);
921c891e 3265 if (vm_page_unwire_quick(pt_pv->pv_m))
701c977e 3266 panic("pmap_remove: insufficient wirecount");
921c891e
MD
3267 } else {
3268 /*
3269 * Unmanaged page table, pt_pv is actually the pd_pv
3270 * for our pmap (not the share object pmap).
3271 *
3272 * We have to unwire the target page table page and we
3273 * have to unwire our page directory page.
3274 */
3275 pmap_inval_interlock(info, pmap, va);
3276 pte = pte_load_clear(ptep);
3277 pmap_inval_deinterlock(info, pmap);
3278 atomic_add_long(&pmap->pm_stats.resident_count, -1);
3279 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3280 panic("pmap_remove: shared pgtable1 bad wirecount");
3281 if (vm_page_unwire_quick(pt_pv->pv_m))
3282 panic("pmap_remove: shared pgtable2 bad wirecount");
701c977e 3283 }
d7f50089
YY
3284}
3285
3286/*
b12defdc
MD
3287 * Removes this physical page from all physical maps in which it resides.
3288 * Reflects back modify bits to the pager.
d7f50089 3289 *
b12defdc 3290 * This routine may not be called from an interrupt.
d7f50089 3291 */
bfc09ba0
MD
3292static
3293void
d7f50089
YY
3294pmap_remove_all(vm_page_t m)
3295{
c8fe38ae 3296 struct pmap_inval_info info;
c8fe38ae
MD
3297 pv_entry_t pv;
3298
3299 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3300 return;
3301
3302 pmap_inval_init(&info);
701c977e 3303 vm_page_spin_lock(m);
c8fe38ae 3304 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
701c977e
MD
3305 KKASSERT(pv->pv_m == m);
3306 if (pv_hold_try(pv)) {
3307 vm_page_spin_unlock(m);
3308 } else {
3309 vm_page_spin_unlock(m);
3310 pv_lock(pv);
3311 if (pv->pv_m != m) {
3312 pv_put(pv);
3313 vm_page_spin_lock(m);
3314 continue;
3315 }
b12defdc 3316 }
b12defdc 3317 /*
701c977e 3318 * Holding no spinlocks, pv is locked.
b12defdc 3319 */
701c977e 3320 pmap_remove_pv_pte(pv, NULL, &info);
52bb73bc 3321 pmap_remove_pv_page(pv);
701c977e 3322 pv_free(pv);
b12defdc 3323 vm_page_spin_lock(m);
c8fe38ae 3324 }
c8fe38ae 3325 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
52bb73bc 3326 vm_page_spin_unlock(m);
c2fb025d 3327 pmap_inval_done(&info);
d7f50089
YY
3328}
3329
3330/*
921c891e
MD
3331 * Set the physical protection on the specified range of this map
3332 * as requested. This function is typically only used for debug watchpoints
3333 * and COW pages.
d7f50089 3334 *
921c891e
MD
3335 * This function may not be called from an interrupt if the map is
3336 * not the kernel_pmap.
d7f50089 3337 *
921c891e 3338 * NOTE! For shared page table pages we just unmap the page.
d7f50089
YY
3339 */
3340void
3341pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
3342{
fb4ca018 3343 struct pmap_scan_info info;
48ffc236
JG
3344 /* JG review for NX */
3345
c8fe38ae
MD
3346 if (pmap == NULL)
3347 return;
c8fe38ae
MD
3348 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
3349 pmap_remove(pmap, sva, eva);
3350 return;
3351 }
c8fe38ae
MD
3352 if (prot & VM_PROT_WRITE)
3353 return;
fb4ca018
MD
3354 info.pmap = pmap;
3355 info.sva = sva;
3356 info.eva = eva;
3357 info.func = pmap_protect_callback;
3358 info.arg = &prot;
3359 pmap_scan(&info);
701c977e 3360}
c8fe38ae 3361
701c977e
MD
3362static
3363void
3364pmap_protect_callback(pmap_t pmap, struct pmap_inval_info *info,
921c891e
MD
3365 pv_entry_t pte_pv, pv_entry_t pt_pv, int sharept,
3366 vm_offset_t va, pt_entry_t *ptep, void *arg __unused)
701c977e
MD
3367{
3368 pt_entry_t pbits;
3369 pt_entry_t cbits;
921c891e 3370 pt_entry_t pte;
701c977e 3371 vm_page_t m;
c8fe38ae 3372
701c977e
MD
3373 /*
3374 * XXX non-optimal.
3375 */
3376 pmap_inval_interlock(info, pmap, va);
c2fb025d 3377again:
701c977e
MD
3378 pbits = *ptep;
3379 cbits = pbits;
3380 if (pte_pv) {
3381 m = NULL;
3382 if (pbits & PG_A) {
3383 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
3384 KKASSERT(m == pte_pv->pv_m);
3385 vm_page_flag_set(m, PG_REFERENCED);
3386 cbits &= ~PG_A;
3387 }
3388 if (pbits & PG_M) {
3389 if (pmap_track_modified(pte_pv->pv_pindex)) {
3390 if (m == NULL)
48ffc236 3391 m = PHYS_TO_VM_PAGE(pbits & PG_FRAME);
701c977e
MD
3392 vm_page_dirty(m);
3393 cbits &= ~PG_M;
c8fe38ae
MD
3394 }
3395 }
921c891e
MD
3396 } else if (sharept) {
3397 /*
3398 * Unmanaged page table, pt_pv is actually the pd_pv
3399 * for our pmap (not the share object pmap).
3400 *
3401 * When asked to protect something in a shared page table
3402 * page we just unmap the page table page. We have to
3403 * invalidate the tlb in this situation.
3404 */
3405 pte = pte_load_clear(ptep);
3406 pmap_inval_invltlb(info);
3407 if (vm_page_unwire_quick(PHYS_TO_VM_PAGE(pte & PG_FRAME)))
3408 panic("pmap_protect: pgtable1 pg bad wirecount");
3409 if (vm_page_unwire_quick(pt_pv->pv_m))
3410 panic("pmap_protect: pgtable2 pg bad wirecount");
3411 ptep = NULL;
c8fe38ae 3412 }
921c891e
MD
3413 /* else unmanaged page, adjust bits, no wire changes */
3414
3415 if (ptep) {
3416 cbits &= ~PG_RW;