kernel - simplify vm pager ops, add pre-faulting for zero-fill pages.
[dragonfly.git] / sys / platform / pc32 / i386 / pmap.c
CommitLineData
984263bc
MD
1/*
2 * Copyright (c) 1991 Regents of the University of California.
3 * All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
5 * All rights reserved.
6 * Copyright (c) 1994 David Greenman
7 * All rights reserved.
8 *
9 * This code is derived from software contributed to Berkeley by
10 * the Systems Programming Group of the University of Utah Computer
11 * Science Department and William Jolitz of UUNET Technologies Inc.
12 *
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
15 * are met:
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. All advertising materials mentioning features or use of this software
22 * must display the following acknowledgement:
23 * This product includes software developed by the University of
24 * California, Berkeley and its contributors.
25 * 4. Neither the name of the University nor the names of its contributors
26 * may be used to endorse or promote products derived from this software
27 * without specific prior written permission.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 * SUCH DAMAGE.
40 *
41 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
42 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
cfd17028 43 * $DragonFly: src/sys/platform/pc32/i386/pmap.c,v 1.87 2008/08/25 17:01:38 dillon Exp $
984263bc
MD
44 */
45
46/*
47 * Manages physical address maps.
48 *
49 * In addition to hardware address maps, this
50 * module is called upon to provide software-use-only
51 * maps which may or may not be stored in the same
52 * form as hardware maps. These pseudo-maps are
53 * used to store intermediate results from copy
54 * operations to and from address spaces.
55 *
56 * Since the information managed by this module is
57 * also stored by the logical address mapping module,
58 * this module may throw away valid virtual-to-physical
59 * mappings at almost any time. However, invalidations
60 * of virtual-to-physical mappings must be done as
61 * requested.
62 *
63 * In order to cope with hardware architectures which
64 * make virtual-to-physical map invalidates expensive,
65 * this module may delay invalidate or reduced protection
66 * operations until such time as they are actually
67 * necessary. This module is given full information as
68 * to which processors are currently using which maps,
69 * and to when physical maps must be made correct.
70 */
5926987a
MD
71/*
72 * PMAP_DEBUG - see platform/pc32/include/pmap.h
73 */
984263bc
MD
74
75#include "opt_disable_pse.h"
76#include "opt_pmap.h"
77#include "opt_msgbuf.h"
984263bc
MD
78
79#include <sys/param.h>
80#include <sys/systm.h>
81#include <sys/kernel.h>
82#include <sys/proc.h>
83#include <sys/msgbuf.h>
84#include <sys/vmmeter.h>
85#include <sys/mman.h>
86
87#include <vm/vm.h>
88#include <vm/vm_param.h>
89#include <sys/sysctl.h>
90#include <sys/lock.h>
91#include <vm/vm_kern.h>
92#include <vm/vm_page.h>
93#include <vm/vm_map.h>
94#include <vm/vm_object.h>
95#include <vm/vm_extern.h>
96#include <vm/vm_pageout.h>
97#include <vm/vm_pager.h>
98#include <vm/vm_zone.h>
99
100#include <sys/user.h>
e0e69b7d 101#include <sys/thread2.h>
e3161323 102#include <sys/sysref2.h>
984263bc
MD
103
104#include <machine/cputypes.h>
105#include <machine/md_var.h>
106#include <machine/specialreg.h>
984263bc 107#include <machine/smp.h>
a9295349 108#include <machine_base/apic/apicreg.h>
85100692 109#include <machine/globaldata.h>
0f7a3396
MD
110#include <machine/pmap.h>
111#include <machine/pmap_inval.h>
984263bc
MD
112
113#define PMAP_KEEP_PDIRS
114#ifndef PMAP_SHPGPERPROC
115#define PMAP_SHPGPERPROC 200
116#endif
117
118#if defined(DIAGNOSTIC)
119#define PMAP_DIAGNOSTIC
120#endif
121
122#define MINPV 2048
123
124#if !defined(PMAP_DIAGNOSTIC)
125#define PMAP_INLINE __inline
126#else
127#define PMAP_INLINE
128#endif
129
130/*
131 * Get PDEs and PTEs for user/kernel address space
132 */
133#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
134#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
135
136#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
137#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
138#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
139#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
140#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
141
984263bc
MD
142
143/*
144 * Given a map and a machine independent protection code,
145 * convert to a vax protection code.
146 */
639a9b43
MD
147#define pte_prot(m, p) \
148 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
984263bc
MD
149static int protection_codes[8];
150
fbbaeba3 151struct pmap kernel_pmap;
54a764e8
MD
152static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
153
e880033d 154vm_paddr_t avail_start; /* PA of first available physical page */
6ef943a3 155vm_paddr_t avail_end; /* PA of last available physical page */
e880033d 156vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
984263bc 157vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
c439ad8f
MD
158vm_offset_t KvaStart; /* VA start of KVA space */
159vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
160vm_offset_t KvaSize; /* max size of kernel virtual address space */
984263bc
MD
161static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
162static int pgeflag; /* PG_G or-in */
163static int pseflag; /* PG_PS or-in */
164
165static vm_object_t kptobj;
166
167static int nkpt;
168vm_offset_t kernel_vm_end;
169
170/*
171 * Data for the pv entry allocation mechanism
172 */
173static vm_zone_t pvzone;
174static struct vm_zone pvzone_store;
175static struct vm_object pvzone_obj;
176static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
177static int pmap_pagedaemon_waken = 0;
178static struct pv_entry *pvinit;
179
a93980ab
MD
180/*
181 * Considering all the issues I'm having with pmap caching, if breakage
182 * continues to occur, and for debugging, I've added a sysctl that will
183 * just do an unconditional invltlb.
184 */
185static int dreadful_invltlb;
186
187SYSCTL_INT(_vm, OID_AUTO, dreadful_invltlb,
188 CTLFLAG_RW, &dreadful_invltlb, 0, "");
189
984263bc
MD
190/*
191 * All those kernel PT submaps that BSD is so fond of
192 */
e731d345 193pt_entry_t *CMAP1 = 0, *ptmmap;
984263bc 194caddr_t CADDR1 = 0, ptvmmap = 0;
984263bc
MD
195static pt_entry_t *msgbufmap;
196struct msgbuf *msgbufp=0;
197
198/*
199 * Crashdump maps.
200 */
201static pt_entry_t *pt_crashdumpmap;
202static caddr_t crashdumpmap;
203
984263bc 204extern pt_entry_t *SMPpt;
984263bc 205
3ae0cd58
RG
206static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
207static unsigned * get_ptbase (pmap_t pmap);
208static pv_entry_t get_pv_entry (void);
209static void i386_protection_init (void);
5e8d0349 210static __inline void pmap_clearbit (vm_page_t m, int bit);
3ae0cd58
RG
211
212static void pmap_remove_all (vm_page_t m);
0f7a3396
MD
213static int pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
214 vm_offset_t sva, pmap_inval_info_t info);
215static void pmap_remove_page (struct pmap *pmap,
216 vm_offset_t va, pmap_inval_info_t info);
3ae0cd58 217static int pmap_remove_entry (struct pmap *pmap, vm_page_t m,
0f7a3396 218 vm_offset_t va, pmap_inval_info_t info);
3ae0cd58
RG
219static boolean_t pmap_testbit (vm_page_t m, int bit);
220static void pmap_insert_entry (pmap_t pmap, vm_offset_t va,
221 vm_page_t mpte, vm_page_t m);
222
223static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
224
225static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
226static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
227static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
228static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
0f7a3396 229static int pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
984263bc
MD
230static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
231
232static unsigned pdir4mb;
233
840de426
MD
234/*
235 * Move the kernel virtual free pointer to the next
236 * 4MB. This is used to help improve performance
237 * by using a large (4MB) page for much of the kernel
238 * (.text, .data, .bss)
239 */
240static vm_offset_t
241pmap_kmem_choose(vm_offset_t addr)
242{
243 vm_offset_t newaddr = addr;
244#ifndef DISABLE_PSE
245 if (cpu_feature & CPUID_PSE) {
246 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
247 }
248#endif
249 return newaddr;
250}
251
984263bc 252/*
e0e69b7d
MD
253 * pmap_pte:
254 *
255 * Extract the page table entry associated with the given map/virtual
256 * pair.
257 *
258 * This function may NOT be called from an interrupt.
984263bc 259 */
984263bc 260PMAP_INLINE unsigned *
840de426 261pmap_pte(pmap_t pmap, vm_offset_t va)
984263bc
MD
262{
263 unsigned *pdeaddr;
264
265 if (pmap) {
266 pdeaddr = (unsigned *) pmap_pde(pmap, va);
267 if (*pdeaddr & PG_PS)
268 return pdeaddr;
269 if (*pdeaddr) {
270 return get_ptbase(pmap) + i386_btop(va);
271 }
272 }
273 return (0);
274}
275
276/*
e0e69b7d
MD
277 * pmap_pte_quick:
278 *
c1692ddf
MD
279 * Super fast pmap_pte routine best used when scanning the pv lists.
280 * This eliminates many course-grained invltlb calls. Note that many of
281 * the pv list scans are across different pmaps and it is very wasteful
282 * to do an entire invltlb when checking a single mapping.
e0e69b7d 283 *
c1692ddf
MD
284 * Should only be called while in a critical section.
285 *
286 * Unlike get_ptbase(), this function MAY be called from an interrupt or
287 * interrupt thread.
984263bc 288 */
840de426
MD
289static unsigned *
290pmap_pte_quick(pmap_t pmap, vm_offset_t va)
984263bc 291{
840de426
MD
292 struct mdglobaldata *gd = mdcpu;
293 unsigned pde, newpf;
294
295 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
296 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
297 unsigned index = i386_btop(va);
298 /* are we current address space or kernel? */
fbbaeba3 299 if ((pmap == &kernel_pmap) ||
840de426
MD
300 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
301 return (unsigned *) PTmap + index;
302 }
303 newpf = pde & PG_FRAME;
304 if ( ((* (unsigned *) gd->gd_PMAP1) & PG_FRAME) != newpf) {
305 * (unsigned *) gd->gd_PMAP1 = newpf | PG_RW | PG_V;
306 cpu_invlpg(gd->gd_PADDR1);
307 }
308 return gd->gd_PADDR1 + ((unsigned) index & (NPTEPG - 1));
984263bc 309 }
840de426 310 return (0);
984263bc
MD
311}
312
840de426 313
984263bc
MD
314/*
315 * Bootstrap the system enough to run with virtual memory.
316 *
317 * On the i386 this is called after mapping has already been enabled
318 * and just syncs the pmap module with what has already been done.
319 * [We can't call it easily with mapping off since the kernel is not
320 * mapped with PA == VA, hence we would have to relocate every address
321 * from the linked base (virtual) address "KERNBASE" to the actual
322 * (physical) address starting relative to 0]
323 */
324void
f123d5a1 325pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
984263bc
MD
326{
327 vm_offset_t va;
328 pt_entry_t *pte;
85100692 329 struct mdglobaldata *gd;
984263bc 330 int i;
81c04d07 331 int pg;
984263bc 332
c439ad8f
MD
333 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
334 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
335 KvaEnd = KvaStart + KvaSize;
336
984263bc
MD
337 avail_start = firstaddr;
338
339 /*
e880033d
MD
340 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
341 * too large. It should instead be correctly calculated in locore.s and
984263bc
MD
342 * not based on 'first' (which is a physical address, not a virtual
343 * address, for the start of unused physical memory). The kernel
344 * page tables are NOT double mapped and thus should not be included
345 * in this calculation.
346 */
e880033d
MD
347 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
348 virtual_start = pmap_kmem_choose(virtual_start);
c439ad8f 349 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
984263bc
MD
350
351 /*
352 * Initialize protection array.
353 */
354 i386_protection_init();
355
356 /*
357 * The kernel's pmap is statically allocated so we don't have to use
358 * pmap_create, which is unlikely to work correctly at this part of
359 * the boot sequence (XXX and which no longer exists).
360 */
fbbaeba3
MD
361 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
362 kernel_pmap.pm_count = 1;
363 kernel_pmap.pm_active = (cpumask_t)-1; /* don't allow deactivation */
364 TAILQ_INIT(&kernel_pmap.pm_pvlist);
984263bc
MD
365 nkpt = NKPT;
366
367 /*
368 * Reserve some special page table entries/VA space for temporary
369 * mapping of pages.
370 */
371#define SYSMAP(c, p, v, n) \
372 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
373
e880033d 374 va = virtual_start;
fbbaeba3 375 pte = (pt_entry_t *) pmap_pte(&kernel_pmap, va);
984263bc
MD
376
377 /*
378 * CMAP1/CMAP2 are used for zeroing and copying pages.
379 */
380 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
984263bc
MD
381
382 /*
383 * Crashdump maps.
384 */
385 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
386
e731d345
MD
387 /*
388 * ptvmmap is used for reading arbitrary physical pages via
389 * /dev/mem.
390 */
391 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
392
984263bc
MD
393 /*
394 * msgbufp is used to map the system message buffer.
395 * XXX msgbufmap is not used.
396 */
397 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
398 atop(round_page(MSGBUF_SIZE)))
399
e880033d 400 virtual_start = va;
984263bc 401
17a9f566 402 *(int *) CMAP1 = 0;
984263bc
MD
403 for (i = 0; i < NKPT; i++)
404 PTD[i] = 0;
405
a2a5ad0d
MD
406 /*
407 * PG_G is terribly broken on SMP because we IPI invltlb's in some
408 * cases rather then invl1pg. Actually, I don't even know why it
409 * works under UP because self-referential page table mappings
410 */
411#ifdef SMP
412 pgeflag = 0;
413#else
414 if (cpu_feature & CPUID_PGE)
984263bc 415 pgeflag = PG_G;
a2a5ad0d 416#endif
984263bc
MD
417
418/*
419 * Initialize the 4MB page size flag
420 */
421 pseflag = 0;
422/*
423 * The 4MB page version of the initial
424 * kernel page mapping.
425 */
426 pdir4mb = 0;
427
428#if !defined(DISABLE_PSE)
429 if (cpu_feature & CPUID_PSE) {
430 unsigned ptditmp;
431 /*
432 * Note that we have enabled PSE mode
433 */
434 pseflag = PG_PS;
435 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
436 ptditmp &= ~(NBPDR - 1);
437 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
438 pdir4mb = ptditmp;
439
8a8d5d85
MD
440#ifndef SMP
441 /*
442 * Enable the PSE mode. If we are SMP we can't do this
443 * now because the APs will not be able to use it when
444 * they boot up.
445 */
446 load_cr4(rcr4() | CR4_PSE);
984263bc 447
8a8d5d85
MD
448 /*
449 * We can do the mapping here for the single processor
450 * case. We simply ignore the old page table page from
451 * now on.
452 */
453 /*
454 * For SMP, we still need 4K pages to bootstrap APs,
455 * PSE will be enabled as soon as all APs are up.
456 */
b5b32410 457 PTD[KPTDI] = (pd_entry_t)ptditmp;
fbbaeba3 458 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
0f7a3396 459 cpu_invltlb();
8a8d5d85 460#endif
984263bc
MD
461 }
462#endif
984263bc 463
81c04d07
MD
464 /*
465 * We need to finish setting up the globaldata page for the BSP.
466 * locore has already populated the page table for the mdglobaldata
467 * portion.
468 */
469 pg = MDGLOBALDATA_BASEALLOC_PAGES;
85100692 470 gd = &CPU_prvspace[0].mdglobaldata;
81c04d07
MD
471 gd->gd_CMAP1 = &SMPpt[pg + 0];
472 gd->gd_CMAP2 = &SMPpt[pg + 1];
473 gd->gd_CMAP3 = &SMPpt[pg + 2];
474 gd->gd_PMAP1 = &SMPpt[pg + 3];
c1692ddf 475 gd->gd_GDMAP1 = &PTD[KGDTDI];
85100692
MD
476 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
477 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
478 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
479 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
c1692ddf 480 gd->gd_GDADDR1= (unsigned *)VADDR(KGDTDI, 0);
984263bc 481
0f7a3396 482 cpu_invltlb();
984263bc
MD
483}
484
485#ifdef SMP
486/*
487 * Set 4mb pdir for mp startup
488 */
489void
490pmap_set_opt(void)
491{
492 if (pseflag && (cpu_feature & CPUID_PSE)) {
493 load_cr4(rcr4() | CR4_PSE);
72740893 494 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
fbbaeba3 495 kernel_pmap.pm_pdir[KPTDI] =
984263bc
MD
496 PTD[KPTDI] = (pd_entry_t)pdir4mb;
497 cpu_invltlb();
498 }
499 }
500}
501#endif
502
503/*
504 * Initialize the pmap module.
505 * Called by vm_init, to initialize any structures that the pmap
506 * system needs to map virtual memory.
507 * pmap_init has been enhanced to support in a fairly consistant
508 * way, discontiguous physical memory.
509 */
510void
e7252eda 511pmap_init(void)
984263bc
MD
512{
513 int i;
514 int initial_pvs;
515
516 /*
517 * object for kernel page table pages
518 */
519 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
520
521 /*
522 * Allocate memory for random pmap data structures. Includes the
523 * pv_head_table.
524 */
525
526 for(i = 0; i < vm_page_array_size; i++) {
527 vm_page_t m;
528
529 m = &vm_page_array[i];
530 TAILQ_INIT(&m->md.pv_list);
531 m->md.pv_list_count = 0;
532 }
533
534 /*
535 * init the pv free list
536 */
537 initial_pvs = vm_page_array_size;
538 if (initial_pvs < MINPV)
539 initial_pvs = MINPV;
540 pvzone = &pvzone_store;
e4846942 541 pvinit = (struct pv_entry *) kmem_alloc(&kernel_map,
984263bc
MD
542 initial_pvs * sizeof (struct pv_entry));
543 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
c5a45196 544 initial_pvs);
984263bc
MD
545
546 /*
547 * Now it is safe to enable pv_table recording.
548 */
549 pmap_initialized = TRUE;
550}
551
552/*
553 * Initialize the address space (zone) for the pv_entries. Set a
554 * high water mark so that the system can recover from excessive
555 * numbers of pv entries.
556 */
557void
f123d5a1 558pmap_init2(void)
984263bc
MD
559{
560 int shpgperproc = PMAP_SHPGPERPROC;
561
562 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
563 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
564 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
565 pv_entry_high_water = 9 * (pv_entry_max / 10);
566 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
567}
568
569
570/***************************************************
571 * Low level helper routines.....
572 ***************************************************/
573
5926987a
MD
574#ifdef PMAP_DEBUG
575
576static void
577test_m_maps_pv(vm_page_t m, pv_entry_t pv)
578{
579 pv_entry_t spv;
580
581 crit_enter();
582#ifdef PMAP_DEBUG
583 KKASSERT(pv->pv_m == m);
584#endif
585 TAILQ_FOREACH(spv, &m->md.pv_list, pv_list) {
586 if (pv == spv) {
587 crit_exit();
588 return;
589 }
590 }
591 crit_exit();
592 panic("test_m_maps_pv: failed m %p pv %p\n", m, pv);
593}
594
595static void
596ptbase_assert(struct pmap *pmap)
597{
598 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
599
600 /* are we current address space or kernel? */
601 if (pmap == &kernel_pmap || frame == (((unsigned)PTDpde) & PG_FRAME)) {
602 return;
603 }
c1692ddf 604 KKASSERT(frame == (*mycpu->gd_GDMAP1 & PG_FRAME));
5926987a
MD
605}
606
607#else
608
609#define test_m_maps_pv(m, pv)
610#define ptbase_assert(pmap)
611
612#endif
613
984263bc
MD
614#if defined(PMAP_DIAGNOSTIC)
615
616/*
617 * This code checks for non-writeable/modified pages.
618 * This should be an invalid condition.
619 */
620static int
621pmap_nw_modified(pt_entry_t ptea)
622{
623 int pte;
624
625 pte = (int) ptea;
626
627 if ((pte & (PG_M|PG_RW)) == PG_M)
628 return 1;
629 else
630 return 0;
631}
632#endif
633
634
635/*
636 * this routine defines the region(s) of memory that should
637 * not be tested for the modified bit.
638 */
639static PMAP_INLINE int
640pmap_track_modified(vm_offset_t va)
641{
642 if ((va < clean_sva) || (va >= clean_eva))
643 return 1;
644 else
645 return 0;
646}
647
c1692ddf
MD
648/*
649 * Retrieve the mapped page table base for a particular pmap. Use our self
650 * mapping for the kernel_pmap or our current pmap.
651 *
652 * For foreign pmaps we use the per-cpu page table map. Since this involves
653 * installing a ptd it's actually (per-process x per-cpu). However, we
654 * still cannot depend on our mapping to survive thread switches because
655 * the process might be threaded and switching to another thread for the
656 * same process on the same cpu will allow that other thread to make its
657 * own mapping.
658 *
659 * This could be a bit confusing but the jist is for something like the
660 * vkernel which uses foreign pmaps all the time this represents a pretty
661 * good cache that avoids unnecessary invltlb()s.
662 */
984263bc 663static unsigned *
e0e69b7d 664get_ptbase(pmap_t pmap)
984263bc
MD
665{
666 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
c1692ddf 667 struct mdglobaldata *gd = mdcpu;
984263bc 668
5926987a
MD
669 /*
670 * We can use PTmap if the pmap is our current address space or
671 * the kernel address space.
672 */
fbbaeba3 673 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
984263bc
MD
674 return (unsigned *) PTmap;
675 }
e0e69b7d 676
5926987a 677 /*
c1692ddf
MD
678 * Otherwise we use the per-cpu alternative page table map. Each
679 * cpu gets its own map. Because of this we cannot use this map
680 * from interrupts or threads which can preempt.
be3aecf7
MD
681 *
682 * Even if we already have the map cached we may still have to
683 * invalidate the TLB if another cpu modified a PDE in the map.
5926987a 684 */
c1692ddf
MD
685 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
686 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
e0e69b7d 687
c1692ddf
MD
688 if ((*gd->gd_GDMAP1 & PG_FRAME) != frame) {
689 *gd->gd_GDMAP1 = frame | PG_RW | PG_V;
be3aecf7
MD
690 pmap->pm_cached |= gd->mi.gd_cpumask;
691 cpu_invltlb();
692 } else if ((pmap->pm_cached & gd->mi.gd_cpumask) == 0) {
693 pmap->pm_cached |= gd->mi.gd_cpumask;
984263bc 694 cpu_invltlb();
a93980ab
MD
695 } else if (dreadful_invltlb) {
696 cpu_invltlb();
984263bc 697 }
c1692ddf 698 return ((unsigned *)gd->gd_GDADDR1);
984263bc
MD
699}
700
984263bc 701/*
e0e69b7d
MD
702 * pmap_extract:
703 *
704 * Extract the physical page address associated with the map/VA pair.
705 *
706 * This function may not be called from an interrupt if the pmap is
707 * not kernel_pmap.
984263bc 708 */
6ef943a3 709vm_paddr_t
840de426 710pmap_extract(pmap_t pmap, vm_offset_t va)
984263bc
MD
711{
712 vm_offset_t rtval;
713 vm_offset_t pdirindex;
840de426 714
984263bc
MD
715 pdirindex = va >> PDRSHIFT;
716 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
717 unsigned *pte;
718 if ((rtval & PG_PS) != 0) {
719 rtval &= ~(NBPDR - 1);
720 rtval |= va & (NBPDR - 1);
721 return rtval;
722 }
723 pte = get_ptbase(pmap) + i386_btop(va);
724 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
725 return rtval;
726 }
727 return 0;
f6bf3af1
MD
728}
729
984263bc
MD
730/***************************************************
731 * Low level mapping routines.....
732 ***************************************************/
733
734/*
6d1ec6fa
HP
735 * Routine: pmap_kenter
736 * Function:
737 * Add a wired page to the KVA
738 * NOTE! note that in order for the mapping to take effect -- you
739 * should do an invltlb after doing the pmap_kenter().
984263bc 740 */
24712b90 741void
6ef943a3 742pmap_kenter(vm_offset_t va, vm_paddr_t pa)
984263bc 743{
840de426 744 unsigned *pte;
0f7a3396
MD
745 unsigned npte;
746 pmap_inval_info info;
984263bc 747
0f7a3396 748 pmap_inval_init(&info);
984263bc
MD
749 npte = pa | PG_RW | PG_V | pgeflag;
750 pte = (unsigned *)vtopte(va);
17cde63e 751 pmap_inval_add(&info, &kernel_pmap, va);
984263bc 752 *pte = npte;
0f7a3396 753 pmap_inval_flush(&info);
984263bc
MD
754}
755
6d1ec6fa
HP
756/*
757 * Routine: pmap_kenter_quick
758 * Function:
759 * Similar to pmap_kenter(), except we only invalidate the
760 * mapping on the current CPU.
761 */
24712b90
MD
762void
763pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
764{
765 unsigned *pte;
766 unsigned npte;
767
768 npte = pa | PG_RW | PG_V | pgeflag;
769 pte = (unsigned *)vtopte(va);
770 *pte = npte;
771 cpu_invlpg((void *)va);
772}
773
774void
775pmap_kenter_sync(vm_offset_t va)
776{
777 pmap_inval_info info;
778
779 pmap_inval_init(&info);
fbbaeba3 780 pmap_inval_add(&info, &kernel_pmap, va);
24712b90
MD
781 pmap_inval_flush(&info);
782}
783
784void
785pmap_kenter_sync_quick(vm_offset_t va)
786{
787 cpu_invlpg((void *)va);
788}
789
984263bc
MD
790/*
791 * remove a page from the kernel pagetables
792 */
24712b90 793void
840de426 794pmap_kremove(vm_offset_t va)
984263bc 795{
840de426 796 unsigned *pte;
0f7a3396 797 pmap_inval_info info;
984263bc 798
0f7a3396 799 pmap_inval_init(&info);
984263bc 800 pte = (unsigned *)vtopte(va);
17cde63e 801 pmap_inval_add(&info, &kernel_pmap, va);
984263bc 802 *pte = 0;
0f7a3396 803 pmap_inval_flush(&info);
984263bc
MD
804}
805
24712b90
MD
806void
807pmap_kremove_quick(vm_offset_t va)
808{
809 unsigned *pte;
810 pte = (unsigned *)vtopte(va);
811 *pte = 0;
812 cpu_invlpg((void *)va);
813}
814
9ad680a3
MD
815/*
816 * XXX these need to be recoded. They are not used in any critical path.
817 */
818void
819pmap_kmodify_rw(vm_offset_t va)
820{
821 *vtopte(va) |= PG_RW;
822 cpu_invlpg((void *)va);
823}
824
825void
826pmap_kmodify_nc(vm_offset_t va)
827{
828 *vtopte(va) |= PG_N;
829 cpu_invlpg((void *)va);
830}
831
984263bc
MD
832/*
833 * Used to map a range of physical addresses into kernel
834 * virtual address space.
835 *
836 * For now, VM is already on, we only need to map the
837 * specified memory.
838 */
839vm_offset_t
8e5e6f1b 840pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
984263bc 841{
8e5e6f1b
AH
842 vm_offset_t sva, virt;
843
844 sva = virt = *virtp;
984263bc
MD
845 while (start < end) {
846 pmap_kenter(virt, start);
847 virt += PAGE_SIZE;
848 start += PAGE_SIZE;
849 }
8e5e6f1b
AH
850 *virtp = virt;
851 return (sva);
984263bc
MD
852}
853
854
855/*
856 * Add a list of wired pages to the kva
857 * this routine is only used for temporary
858 * kernel mappings that do not need to have
859 * page modification or references recorded.
860 * Note that old mappings are simply written
861 * over. The page *must* be wired.
862 */
863void
840de426 864pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
984263bc
MD
865{
866 vm_offset_t end_va;
867
868 end_va = va + count * PAGE_SIZE;
869
870 while (va < end_va) {
871 unsigned *pte;
872
873 pte = (unsigned *)vtopte(va);
874 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
984263bc 875 cpu_invlpg((void *)va);
984263bc
MD
876 va += PAGE_SIZE;
877 m++;
878 }
879#ifdef SMP
0f7a3396 880 smp_invltlb(); /* XXX */
984263bc
MD
881#endif
882}
883
8100156a
MD
884void
885pmap_qenter2(vm_offset_t va, vm_page_t *m, int count, cpumask_t *mask)
886{
887 vm_offset_t end_va;
888 cpumask_t cmask = mycpu->gd_cpumask;
889
890 end_va = va + count * PAGE_SIZE;
891
892 while (va < end_va) {
893 unsigned *pte;
894 unsigned pteval;
895
896 /*
897 * Install the new PTE. If the pte changed from the prior
898 * mapping we must reset the cpu mask and invalidate the page.
899 * If the pte is the same but we have not seen it on the
900 * current cpu, invlpg the existing mapping. Otherwise the
901 * entry is optimal and no invalidation is required.
902 */
903 pte = (unsigned *)vtopte(va);
904 pteval = VM_PAGE_TO_PHYS(*m) | PG_A | PG_RW | PG_V | pgeflag;
905 if (*pte != pteval) {
a02705a9 906 *mask = 0;
8100156a
MD
907 *pte = pteval;
908 cpu_invlpg((void *)va);
909 } else if ((*mask & cmask) == 0) {
8100156a
MD
910 cpu_invlpg((void *)va);
911 }
912 va += PAGE_SIZE;
913 m++;
914 }
a02705a9 915 *mask |= cmask;
8100156a
MD
916}
917
984263bc 918/*
7155fc7d 919 * This routine jerks page mappings from the
984263bc 920 * kernel -- it is meant only for temporary mappings.
7155fc7d
MD
921 *
922 * MPSAFE, INTERRUPT SAFE (cluster callback)
984263bc
MD
923 */
924void
840de426 925pmap_qremove(vm_offset_t va, int count)
984263bc
MD
926{
927 vm_offset_t end_va;
928
929 end_va = va + count*PAGE_SIZE;
930
931 while (va < end_va) {
932 unsigned *pte;
933
934 pte = (unsigned *)vtopte(va);
935 *pte = 0;
984263bc 936 cpu_invlpg((void *)va);
984263bc
MD
937 va += PAGE_SIZE;
938 }
939#ifdef SMP
940 smp_invltlb();
941#endif
942}
943
06ecca5a
MD
944/*
945 * This routine works like vm_page_lookup() but also blocks as long as the
946 * page is busy. This routine does not busy the page it returns.
947 *
948 * Unless the caller is managing objects whos pages are in a known state,
654a39f0
MD
949 * the call should be made with a critical section held so the page's object
950 * association remains valid on return.
06ecca5a 951 */
984263bc 952static vm_page_t
840de426 953pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
984263bc
MD
954{
955 vm_page_t m;
06ecca5a 956
17cde63e
MD
957 do {
958 m = vm_page_lookup(object, pindex);
959 } while (m && vm_page_sleep_busy(m, FALSE, "pplookp"));
960
06ecca5a 961 return(m);
984263bc
MD
962}
963
263e4574
MD
964/*
965 * Create a new thread and optionally associate it with a (new) process.
6ef943a3 966 * NOTE! the new thread's cpu may not equal the current cpu.
263e4574 967 */
7d0bac62
MD
968void
969pmap_init_thread(thread_t td)
263e4574 970{
f470d0c8 971 /* enforce pcb placement */
f470d0c8 972 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
65d6ce10 973 td->td_savefpu = &td->td_pcb->pcb_save;
7d0bac62 974 td->td_sp = (char *)td->td_pcb - 16;
263e4574
MD
975}
976
984263bc 977/*
984263bc
MD
978 * This routine directly affects the fork perf for a process.
979 */
980void
13d13d89 981pmap_init_proc(struct proc *p)
984263bc 982{
984263bc
MD
983}
984
985/*
986 * Dispose the UPAGES for a process that has exited.
987 * This routine directly impacts the exit perf of a process.
988 */
c6880072 989void
7e1d4bf4 990pmap_dispose_proc(struct proc *p)
984263bc 991{
f1d1c3fa 992 KASSERT(p->p_lock == 0, ("attempt to dispose referenced proc! %p", p));
984263bc
MD
993}
994
984263bc
MD
995/***************************************************
996 * Page table page management routines.....
997 ***************************************************/
998
999/*
1000 * This routine unholds page table pages, and if the hold count
1001 * drops to zero, then it decrements the wire count.
1002 */
1003static int
0f7a3396 1004_pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
840de426 1005{
17cde63e
MD
1006 /*
1007 * Wait until we can busy the page ourselves. We cannot have
1008 * any active flushes if we block.
1009 */
1010 if (m->flags & PG_BUSY) {
1011 pmap_inval_flush(info);
1012 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1013 ;
1014 }
eec2b734
MD
1015 KASSERT(m->queue == PQ_NONE,
1016 ("_pmap_unwire_pte_hold: %p->queue != PQ_NONE", m));
984263bc 1017
eec2b734 1018 if (m->hold_count == 1) {
984263bc 1019 /*
be3aecf7
MD
1020 * Unmap the page table page.
1021 *
1022 * NOTE: We must clear pm_cached for all cpus, including
1023 * the current one, when clearing a page directory
1024 * entry.
984263bc 1025 */
eec2b734 1026 vm_page_busy(m);
0f7a3396 1027 pmap_inval_add(info, pmap, -1);
2247fe02 1028 KKASSERT(pmap->pm_pdir[m->pindex]);
984263bc 1029 pmap->pm_pdir[m->pindex] = 0;
be3aecf7 1030 pmap->pm_cached = 0;
eec2b734
MD
1031
1032 KKASSERT(pmap->pm_stats.resident_count > 0);
984263bc 1033 --pmap->pm_stats.resident_count;
984263bc
MD
1034
1035 if (pmap->pm_ptphint == m)
1036 pmap->pm_ptphint = NULL;
1037
1038 /*
eec2b734
MD
1039 * This was our last hold, the page had better be unwired
1040 * after we decrement wire_count.
1041 *
1042 * FUTURE NOTE: shared page directory page could result in
1043 * multiple wire counts.
984263bc 1044 */
eec2b734 1045 vm_page_unhold(m);
984263bc 1046 --m->wire_count;
eec2b734
MD
1047 KKASSERT(m->wire_count == 0);
1048 --vmstats.v_wire_count;
17cde63e 1049 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
eec2b734
MD
1050 vm_page_flash(m);
1051 vm_page_free_zero(m);
984263bc 1052 return 1;
17cde63e
MD
1053 } else {
1054 KKASSERT(m->hold_count > 1);
1055 vm_page_unhold(m);
1056 return 0;
984263bc 1057 }
984263bc
MD
1058}
1059
1060static PMAP_INLINE int
0f7a3396 1061pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
984263bc 1062{
eec2b734
MD
1063 KKASSERT(m->hold_count > 0);
1064 if (m->hold_count > 1) {
1065 vm_page_unhold(m);
984263bc 1066 return 0;
eec2b734
MD
1067 } else {
1068 return _pmap_unwire_pte_hold(pmap, m, info);
1069 }
984263bc
MD
1070}
1071
1072/*
1073 * After removing a page table entry, this routine is used to
1074 * conditionally free the page, and manage the hold/wire counts.
5926987a
MD
1075 *
1076 * WARNING: This function can block
984263bc
MD
1077 */
1078static int
0f7a3396
MD
1079pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
1080 pmap_inval_info_t info)
984263bc
MD
1081{
1082 unsigned ptepindex;
1083 if (va >= UPT_MIN_ADDRESS)
1084 return 0;
1085
1086 if (mpte == NULL) {
1087 ptepindex = (va >> PDRSHIFT);
1088 if (pmap->pm_ptphint &&
1089 (pmap->pm_ptphint->pindex == ptepindex)) {
1090 mpte = pmap->pm_ptphint;
1091 } else {
0f7a3396 1092 pmap_inval_flush(info);
984263bc
MD
1093 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1094 pmap->pm_ptphint = mpte;
1095 }
1096 }
1097
0f7a3396 1098 return pmap_unwire_pte_hold(pmap, mpte, info);
984263bc
MD
1099}
1100
54a764e8 1101/*
fbbaeba3
MD
1102 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1103 * it, and IdlePTD, represents the template used to update all other pmaps.
1104 *
1105 * On architectures where the kernel pmap is not integrated into the user
1106 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1107 * kernel_pmap should be used to directly access the kernel_pmap.
54a764e8 1108 */
984263bc 1109void
840de426 1110pmap_pinit0(struct pmap *pmap)
984263bc
MD
1111{
1112 pmap->pm_pdir =
e4846942 1113 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
24712b90 1114 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
984263bc
MD
1115 pmap->pm_count = 1;
1116 pmap->pm_active = 0;
be3aecf7 1117 pmap->pm_cached = 0;
984263bc
MD
1118 pmap->pm_ptphint = NULL;
1119 TAILQ_INIT(&pmap->pm_pvlist);
1120 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1121}
1122
1123/*
1124 * Initialize a preallocated and zeroed pmap structure,
1125 * such as one in a vmspace structure.
1126 */
1127void
840de426 1128pmap_pinit(struct pmap *pmap)
984263bc
MD
1129{
1130 vm_page_t ptdpg;
1131
1132 /*
1133 * No need to allocate page table space yet but we do need a valid
1134 * page directory table.
1135 */
b5b32410 1136 if (pmap->pm_pdir == NULL) {
984263bc 1137 pmap->pm_pdir =
e4846942 1138 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
b5b32410 1139 }
984263bc
MD
1140
1141 /*
c3834cb2 1142 * Allocate an object for the ptes
984263bc
MD
1143 */
1144 if (pmap->pm_pteobj == NULL)
c3834cb2 1145 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
984263bc
MD
1146
1147 /*
c3834cb2
MD
1148 * Allocate the page directory page, unless we already have
1149 * one cached. If we used the cached page the wire_count will
1150 * already be set appropriately.
984263bc 1151 */
c3834cb2
MD
1152 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1153 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1154 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1155 pmap->pm_pdirm = ptdpg;
1156 vm_page_flag_clear(ptdpg, PG_MAPPED | PG_BUSY);
1157 ptdpg->valid = VM_PAGE_BITS_ALL;
1158 ptdpg->wire_count = 1;
1159 ++vmstats.v_wire_count;
1160 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
1161 }
984263bc
MD
1162 if ((ptdpg->flags & PG_ZERO) == 0)
1163 bzero(pmap->pm_pdir, PAGE_SIZE);
5926987a
MD
1164#ifdef PMAP_DEBUG
1165 else
1166 pmap_page_assertzero(VM_PAGE_TO_PHYS(ptdpg));
1167#endif
984263bc 1168
984263bc 1169 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
984263bc
MD
1170
1171 /* install self-referential address mapping entry */
1172 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1173 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1174
1175 pmap->pm_count = 1;
1176 pmap->pm_active = 0;
be3aecf7 1177 pmap->pm_cached = 0;
984263bc
MD
1178 pmap->pm_ptphint = NULL;
1179 TAILQ_INIT(&pmap->pm_pvlist);
1180 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
eec2b734 1181 pmap->pm_stats.resident_count = 1;
984263bc
MD
1182}
1183
e3161323 1184/*
c3834cb2
MD
1185 * Clean up a pmap structure so it can be physically freed. This routine
1186 * is called by the vmspace dtor function. A great deal of pmap data is
1187 * left passively mapped to improve vmspace management so we have a bit
1188 * of cleanup work to do here.
e3161323
MD
1189 */
1190void
1191pmap_puninit(pmap_t pmap)
1192{
c3834cb2
MD
1193 vm_page_t p;
1194
e3161323 1195 KKASSERT(pmap->pm_active == 0);
c3834cb2
MD
1196 if ((p = pmap->pm_pdirm) != NULL) {
1197 KKASSERT(pmap->pm_pdir != NULL);
1198 pmap_kremove((vm_offset_t)pmap->pm_pdir);
1199 p->wire_count--;
1200 vmstats.v_wire_count--;
1201 KKASSERT((p->flags & PG_BUSY) == 0);
1202 vm_page_busy(p);
1203 vm_page_free_zero(p);
1204 pmap->pm_pdirm = NULL;
1205 }
e3161323
MD
1206 if (pmap->pm_pdir) {
1207 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1208 pmap->pm_pdir = NULL;
1209 }
1210 if (pmap->pm_pteobj) {
1211 vm_object_deallocate(pmap->pm_pteobj);
1212 pmap->pm_pteobj = NULL;
1213 }
1214}
1215
984263bc
MD
1216/*
1217 * Wire in kernel global address entries. To avoid a race condition
1218 * between pmap initialization and pmap_growkernel, this procedure
54a764e8
MD
1219 * adds the pmap to the master list (which growkernel scans to update),
1220 * then copies the template.
984263bc
MD
1221 */
1222void
840de426 1223pmap_pinit2(struct pmap *pmap)
984263bc 1224{
54a764e8
MD
1225 crit_enter();
1226 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
984263bc
MD
1227 /* XXX copies current process, does not fill in MPPTDI */
1228 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
54a764e8 1229 crit_exit();
984263bc
MD
1230}
1231
344ad853 1232/*
eec2b734 1233 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
344ad853 1234 * 0 on failure (if the procedure had to sleep).
c3834cb2
MD
1235 *
1236 * When asked to remove the page directory page itself, we actually just
1237 * leave it cached so we do not have to incur the SMP inval overhead of
1238 * removing the kernel mapping. pmap_puninit() will take care of it.
344ad853 1239 */
984263bc 1240static int
840de426 1241pmap_release_free_page(struct pmap *pmap, vm_page_t p)
984263bc
MD
1242{
1243 unsigned *pde = (unsigned *) pmap->pm_pdir;
1244 /*
1245 * This code optimizes the case of freeing non-busy
1246 * page-table pages. Those pages are zero now, and
1247 * might as well be placed directly into the zero queue.
1248 */
1249 if (vm_page_sleep_busy(p, FALSE, "pmaprl"))
1250 return 0;
1251
1252 vm_page_busy(p);
1253
1254 /*
1255 * Remove the page table page from the processes address space.
1256 */
eec2b734 1257 KKASSERT(pmap->pm_stats.resident_count > 0);
2247fe02
MD
1258 KKASSERT(pde[p->pindex]);
1259 pde[p->pindex] = 0;
eec2b734 1260 --pmap->pm_stats.resident_count;
a93980ab 1261 pmap->pm_cached = 0;
984263bc
MD
1262
1263 if (p->hold_count) {
1264 panic("pmap_release: freeing held page table page");
1265 }
c3834cb2
MD
1266 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1267 pmap->pm_ptphint = NULL;
1268
984263bc 1269 /*
c3834cb2
MD
1270 * We leave the page directory page cached, wired, and mapped in
1271 * the pmap until the dtor function (pmap_puninit()) gets called.
1272 * However, still clean it up so we can set PG_ZERO.
c1692ddf
MD
1273 *
1274 * The pmap has already been removed from the pmap_list in the
1275 * PTDPTDI case.
984263bc
MD
1276 */
1277 if (p->pindex == PTDPTDI) {
1278 bzero(pde + KPTDI, nkpt * PTESIZE);
c1692ddf 1279 bzero(pde + KGDTDI, (NPDEPG - KGDTDI) * PTESIZE);
c3834cb2
MD
1280 vm_page_flag_set(p, PG_ZERO);
1281 vm_page_wakeup(p);
1282 } else {
1283 p->wire_count--;
1284 vmstats.v_wire_count--;
1285 vm_page_free_zero(p);
984263bc 1286 }
984263bc
MD
1287 return 1;
1288}
1289
1290/*
1291 * this routine is called if the page table page is not
1292 * mapped correctly.
1293 */
1294static vm_page_t
840de426 1295_pmap_allocpte(pmap_t pmap, unsigned ptepindex)
984263bc
MD
1296{
1297 vm_offset_t pteva, ptepa;
1298 vm_page_t m;
1299
1300 /*
1301 * Find or fabricate a new pagetable page
1302 */
1303 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
dc1fd4b3 1304 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
984263bc
MD
1305
1306 KASSERT(m->queue == PQ_NONE,
1307 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1308
eec2b734
MD
1309 /*
1310 * Increment the hold count for the page we will be returning to
1311 * the caller.
1312 */
1313 m->hold_count++;
1314
1315 /*
1316 * It is possible that someone else got in and mapped by the page
1317 * directory page while we were blocked, if so just unbusy and
1318 * return the held page.
1319 */
1320 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1321 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1322 vm_page_wakeup(m);
1323 return(m);
1324 }
1325
984263bc 1326 if (m->wire_count == 0)
12e4aaff 1327 vmstats.v_wire_count++;
984263bc
MD
1328 m->wire_count++;
1329
984263bc
MD
1330
1331 /*
1332 * Map the pagetable page into the process address space, if
1333 * it isn't already there.
be3aecf7
MD
1334 *
1335 * NOTE: For safety clear pm_cached for all cpus including the
1336 * current one when adding a PDE to the map.
984263bc 1337 */
eec2b734 1338 ++pmap->pm_stats.resident_count;
984263bc
MD
1339
1340 ptepa = VM_PAGE_TO_PHYS(m);
1341 pmap->pm_pdir[ptepindex] =
1342 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
be3aecf7 1343 pmap->pm_cached = 0;
984263bc
MD
1344
1345 /*
1346 * Set the page table hint
1347 */
1348 pmap->pm_ptphint = m;
1349
1350 /*
1351 * Try to use the new mapping, but if we cannot, then
1352 * do it with the routine that maps the page explicitly.
1353 */
1354 if ((m->flags & PG_ZERO) == 0) {
1355 if ((((unsigned)pmap->pm_pdir[PTDPTDI]) & PG_FRAME) ==
1356 (((unsigned) PTDpde) & PG_FRAME)) {
1357 pteva = UPT_MIN_ADDRESS + i386_ptob(ptepindex);
1358 bzero((caddr_t) pteva, PAGE_SIZE);
1359 } else {
1360 pmap_zero_page(ptepa);
1361 }
1362 }
5926987a
MD
1363#ifdef PMAP_DEBUG
1364 else {
1365 pmap_page_assertzero(VM_PAGE_TO_PHYS(m));
1366 }
1367#endif
984263bc
MD
1368
1369 m->valid = VM_PAGE_BITS_ALL;
1370 vm_page_flag_clear(m, PG_ZERO);
1371 vm_page_flag_set(m, PG_MAPPED);
1372 vm_page_wakeup(m);
1373
1374 return m;
1375}
1376
1377static vm_page_t
840de426 1378pmap_allocpte(pmap_t pmap, vm_offset_t va)
984263bc
MD
1379{
1380 unsigned ptepindex;
1381 vm_offset_t ptepa;
1382 vm_page_t m;
1383
1384 /*
1385 * Calculate pagetable page index
1386 */
1387 ptepindex = va >> PDRSHIFT;
1388
1389 /*
1390 * Get the page directory entry
1391 */
1392 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1393
1394 /*
1395 * This supports switching from a 4MB page to a
1396 * normal 4K page.
1397 */
1398 if (ptepa & PG_PS) {
1399 pmap->pm_pdir[ptepindex] = 0;
1400 ptepa = 0;
0f7a3396
MD
1401 cpu_invltlb();
1402 smp_invltlb();
984263bc
MD
1403 }
1404
1405 /*
1406 * If the page table page is mapped, we just increment the
1407 * hold count, and activate it.
1408 */
1409 if (ptepa) {
1410 /*
1411 * In order to get the page table page, try the
1412 * hint first.
1413 */
1414 if (pmap->pm_ptphint &&
1415 (pmap->pm_ptphint->pindex == ptepindex)) {
1416 m = pmap->pm_ptphint;
1417 } else {
1418 m = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
1419 pmap->pm_ptphint = m;
1420 }
1421 m->hold_count++;
1422 return m;
1423 }
1424 /*
1425 * Here if the pte page isn't mapped, or if it has been deallocated.
1426 */
1427 return _pmap_allocpte(pmap, ptepindex);
1428}
1429
1430
1431/***************************************************
1f804340 1432 * Pmap allocation/deallocation routines.
984263bc
MD
1433 ***************************************************/
1434
1435/*
1436 * Release any resources held by the given physical map.
1437 * Called when a pmap initialized by pmap_pinit is being released.
1438 * Should only be called if the map contains no valid mappings.
1439 */
1f804340
MD
1440static int pmap_release_callback(struct vm_page *p, void *data);
1441
984263bc 1442void
840de426 1443pmap_release(struct pmap *pmap)
984263bc 1444{
984263bc 1445 vm_object_t object = pmap->pm_pteobj;
1f804340 1446 struct rb_vm_page_scan_info info;
984263bc 1447
e3161323 1448 KASSERT(pmap->pm_active == 0, ("pmap still active! %08x", pmap->pm_active));
984263bc
MD
1449#if defined(DIAGNOSTIC)
1450 if (object->ref_count != 1)
1451 panic("pmap_release: pteobj reference count != 1");
1452#endif
1453
1f804340
MD
1454 info.pmap = pmap;
1455 info.object = object;
9acd5bbb 1456 crit_enter();
54a764e8 1457 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
1f804340
MD
1458 crit_exit();
1459
1460 do {
1461 crit_enter();
1462 info.error = 0;
1463 info.mpte = NULL;
1464 info.limit = object->generation;
1465
1466 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1467 pmap_release_callback, &info);
1468 if (info.error == 0 && info.mpte) {
1469 if (!pmap_release_free_page(pmap, info.mpte))
1470 info.error = 1;
984263bc 1471 }
344ad853 1472 crit_exit();
1f804340 1473 } while (info.error);
a93980ab 1474 pmap->pm_cached = 0;
1f804340
MD
1475}
1476
1477static int
1478pmap_release_callback(struct vm_page *p, void *data)
1479{
1480 struct rb_vm_page_scan_info *info = data;
1481
1482 if (p->pindex == PTDPTDI) {
1483 info->mpte = p;
1484 return(0);
344ad853 1485 }
1f804340
MD
1486 if (!pmap_release_free_page(info->pmap, p)) {
1487 info->error = 1;
1488 return(-1);
1489 }
1490 if (info->object->generation != info->limit) {
1491 info->error = 1;
1492 return(-1);
1493 }
1494 return(0);
984263bc 1495}
984263bc
MD
1496
1497/*
0e5797fe 1498 * Grow the number of kernel page table entries, if needed.
984263bc 1499 */
0e5797fe 1500
984263bc
MD
1501void
1502pmap_growkernel(vm_offset_t addr)
1503{
54a764e8 1504 struct pmap *pmap;
984263bc
MD
1505 vm_offset_t ptppaddr;
1506 vm_page_t nkpg;
1507 pd_entry_t newpdir;
1508
9acd5bbb 1509 crit_enter();
984263bc
MD
1510 if (kernel_vm_end == 0) {
1511 kernel_vm_end = KERNBASE;
1512 nkpt = 0;
1513 while (pdir_pde(PTD, kernel_vm_end)) {
1514 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1515 nkpt++;
1516 }
1517 }
1518 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1519 while (kernel_vm_end < addr) {
1520 if (pdir_pde(PTD, kernel_vm_end)) {
1521 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1522 continue;
1523 }
1524
1525 /*
1526 * This index is bogus, but out of the way
1527 */
dc1fd4b3
MD
1528 nkpg = vm_page_alloc(kptobj, nkpt,
1529 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM | VM_ALLOC_INTERRUPT);
1530 if (nkpg == NULL)
984263bc
MD
1531 panic("pmap_growkernel: no memory to grow kernel");
1532
984263bc
MD
1533 vm_page_wire(nkpg);
1534 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1535 pmap_zero_page(ptppaddr);
1536 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1537 pdir_pde(PTD, kernel_vm_end) = newpdir;
fbbaeba3 1538 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
0e5797fe
MD
1539 nkpt++;
1540
1541 /*
54a764e8 1542 * This update must be interlocked with pmap_pinit2.
0e5797fe 1543 */
54a764e8
MD
1544 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1545 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1546 }
1547 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1548 ~(PAGE_SIZE * NPTEPG - 1);
984263bc 1549 }
9acd5bbb 1550 crit_exit();
984263bc
MD
1551}
1552
1553/*
1554 * Retire the given physical map from service.
1555 * Should only be called if the map contains
1556 * no valid mappings.
1557 */
1558void
840de426 1559pmap_destroy(pmap_t pmap)
984263bc
MD
1560{
1561 int count;
1562
1563 if (pmap == NULL)
1564 return;
1565
1566 count = --pmap->pm_count;
1567 if (count == 0) {
1568 pmap_release(pmap);
1569 panic("destroying a pmap is not yet implemented");
1570 }
1571}
1572
1573/*
1574 * Add a reference to the specified pmap.
1575 */
1576void
840de426 1577pmap_reference(pmap_t pmap)
984263bc
MD
1578{
1579 if (pmap != NULL) {
1580 pmap->pm_count++;
1581 }
1582}
1583
1584/***************************************************
1585* page management routines.
1586 ***************************************************/
1587
1588/*
8a8d5d85
MD
1589 * free the pv_entry back to the free list. This function may be
1590 * called from an interrupt.
984263bc
MD
1591 */
1592static PMAP_INLINE void
840de426 1593free_pv_entry(pv_entry_t pv)
984263bc 1594{
5926987a
MD
1595#ifdef PMAP_DEBUG
1596 KKASSERT(pv->pv_m != NULL);
1597 pv->pv_m = NULL;
1598#endif
984263bc 1599 pv_entry_count--;
8a8d5d85 1600 zfree(pvzone, pv);
984263bc
MD
1601}
1602
1603/*
1604 * get a new pv_entry, allocating a block from the system
8a8d5d85 1605 * when needed. This function may be called from an interrupt.
984263bc
MD
1606 */
1607static pv_entry_t
1608get_pv_entry(void)
1609{
1610 pv_entry_count++;
1611 if (pv_entry_high_water &&
20479584
MD
1612 (pv_entry_count > pv_entry_high_water) &&
1613 (pmap_pagedaemon_waken == 0)) {
984263bc
MD
1614 pmap_pagedaemon_waken = 1;
1615 wakeup (&vm_pages_needed);
1616 }
8a8d5d85 1617 return zalloc(pvzone);
984263bc
MD
1618}
1619
1620/*
1621 * This routine is very drastic, but can save the system
1622 * in a pinch.
1623 */
1624void
840de426 1625pmap_collect(void)
984263bc
MD
1626{
1627 int i;
1628 vm_page_t m;
1629 static int warningdone=0;
1630
1631 if (pmap_pagedaemon_waken == 0)
1632 return;
20479584 1633 pmap_pagedaemon_waken = 0;
984263bc
MD
1634
1635 if (warningdone < 5) {
26be20a0 1636 kprintf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
984263bc
MD
1637 warningdone++;
1638 }
1639
1640 for(i = 0; i < vm_page_array_size; i++) {
1641 m = &vm_page_array[i];
1642 if (m->wire_count || m->hold_count || m->busy ||
1643 (m->flags & PG_BUSY))
1644 continue;
1645 pmap_remove_all(m);
1646 }
984263bc
MD
1647}
1648
1649
1650/*
1651 * If it is the first entry on the list, it is actually
1652 * in the header and we must copy the following entry up
1653 * to the header. Otherwise we must search the list for
1654 * the entry. In either case we free the now unused entry.
1655 */
984263bc 1656static int
0f7a3396
MD
1657pmap_remove_entry(struct pmap *pmap, vm_page_t m,
1658 vm_offset_t va, pmap_inval_info_t info)
984263bc
MD
1659{
1660 pv_entry_t pv;
1661 int rtval;
984263bc 1662
9acd5bbb 1663 crit_enter();
984263bc
MD
1664 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1665 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1666 if (pmap == pv->pv_pmap && va == pv->pv_va)
1667 break;
1668 }
1669 } else {
1670 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
5926987a
MD
1671#ifdef PMAP_DEBUG
1672 KKASSERT(pv->pv_pmap == pmap);
1673#endif
1674 if (va == pv->pv_va)
984263bc
MD
1675 break;
1676 }
1677 }
5926987a 1678 KKASSERT(pv);
984263bc
MD
1679
1680 rtval = 0;
5926987a
MD
1681 test_m_maps_pv(m, pv);
1682 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1683 m->md.pv_list_count--;
1684 if (TAILQ_EMPTY(&m->md.pv_list))
1685 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1686 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1687 ++pmap->pm_generation;
1688 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
1689 free_pv_entry(pv);
9acd5bbb 1690 crit_exit();
984263bc
MD
1691 return rtval;
1692}
1693
1694/*
1695 * Create a pv entry for page at pa for
1696 * (pmap, va).
1697 */
1698static void
840de426 1699pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
984263bc 1700{
984263bc
MD
1701 pv_entry_t pv;
1702
9acd5bbb 1703 crit_enter();
984263bc 1704 pv = get_pv_entry();
5926987a
MD
1705#ifdef PMAP_DEBUG
1706 KKASSERT(pv->pv_m == NULL);
1707 pv->pv_m = m;
1708#endif
984263bc
MD
1709 pv->pv_va = va;
1710 pv->pv_pmap = pmap;
1711 pv->pv_ptem = mpte;
1712
1713 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1714 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
5926987a 1715 ++pmap->pm_generation;
984263bc
MD
1716 m->md.pv_list_count++;
1717
9acd5bbb 1718 crit_exit();
984263bc
MD
1719}
1720
1721/*
5926987a
MD
1722 * pmap_remove_pte: do the things to unmap a page in a process.
1723 *
1724 * WARNING: This function may block (via pmap_remove_entry/pmap_unuse_pt),
1725 * callers using temporary pmaps must reload them.
984263bc
MD
1726 */
1727static int
0f7a3396 1728pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
5926987a 1729 pmap_inval_info_t info)
984263bc
MD
1730{
1731 unsigned oldpte;
1732 vm_page_t m;
1733
5926987a 1734 ptbase_assert(pmap);
0f7a3396 1735 pmap_inval_add(info, pmap, va);
5926987a 1736 ptbase_assert(pmap);
984263bc 1737 oldpte = loadandclear(ptq);
2247fe02 1738 KKASSERT(oldpte);
984263bc
MD
1739 if (oldpte & PG_W)
1740 pmap->pm_stats.wired_count -= 1;
1741 /*
1742 * Machines that don't support invlpg, also don't support
0f7a3396
MD
1743 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1744 * the SMP case.
984263bc
MD
1745 */
1746 if (oldpte & PG_G)
41a01a4d 1747 cpu_invlpg((void *)va);
eec2b734
MD
1748 KKASSERT(pmap->pm_stats.resident_count > 0);
1749 --pmap->pm_stats.resident_count;
984263bc
MD
1750 if (oldpte & PG_MANAGED) {
1751 m = PHYS_TO_VM_PAGE(oldpte);
1752 if (oldpte & PG_M) {
1753#if defined(PMAP_DIAGNOSTIC)
1754 if (pmap_nw_modified((pt_entry_t) oldpte)) {
d557216f
MD
1755 kprintf("pmap_remove: modified page not "
1756 "writable: va: %p, pte: 0x%lx\n",
1757 (void *)va, (long)oldpte);
984263bc
MD
1758 }
1759#endif
1760 if (pmap_track_modified(va))
1761 vm_page_dirty(m);
1762 }
1763 if (oldpte & PG_A)
1764 vm_page_flag_set(m, PG_REFERENCED);
0f7a3396 1765 return pmap_remove_entry(pmap, m, va, info);
984263bc 1766 } else {
0f7a3396 1767 return pmap_unuse_pt(pmap, va, NULL, info);
984263bc
MD
1768 }
1769
1770 return 0;
1771}
1772
1773/*
e0e69b7d
MD
1774 * pmap_remove_page:
1775 *
5926987a 1776 * Remove a single page from a process address space.
e0e69b7d 1777 *
5926987a
MD
1778 * This function may not be called from an interrupt if the pmap is
1779 * not kernel_pmap.
984263bc
MD
1780 */
1781static void
0f7a3396 1782pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
984263bc 1783{
840de426 1784 unsigned *ptq;
984263bc
MD
1785
1786 /*
e0e69b7d
MD
1787 * if there is no pte for this address, just skip it!!! Otherwise
1788 * get a local va for mappings for this pmap and remove the entry.
984263bc 1789 */
e0e69b7d
MD
1790 if (*pmap_pde(pmap, va) != 0) {
1791 ptq = get_ptbase(pmap) + i386_btop(va);
1792 if (*ptq) {
0f7a3396 1793 pmap_remove_pte(pmap, ptq, va, info);
5926987a 1794 /* ptq invalid */
e0e69b7d 1795 }
984263bc 1796 }
984263bc
MD
1797}
1798
1799/*
0f7a3396 1800 * pmap_remove:
e0e69b7d 1801 *
984263bc
MD
1802 * Remove the given range of addresses from the specified map.
1803 *
1804 * It is assumed that the start and end are properly
1805 * rounded to the page size.
e0e69b7d
MD
1806 *
1807 * This function may not be called from an interrupt if the pmap is
1808 * not kernel_pmap.
984263bc
MD
1809 */
1810void
840de426 1811pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
984263bc 1812{
840de426 1813 unsigned *ptbase;
984263bc
MD
1814 vm_offset_t pdnxt;
1815 vm_offset_t ptpaddr;
1816 vm_offset_t sindex, eindex;
0f7a3396 1817 struct pmap_inval_info info;
984263bc
MD
1818
1819 if (pmap == NULL)
1820 return;
1821
1822 if (pmap->pm_stats.resident_count == 0)
1823 return;
1824
0f7a3396
MD
1825 pmap_inval_init(&info);
1826
984263bc
MD
1827 /*
1828 * special handling of removing one page. a very
1829 * common operation and easy to short circuit some
1830 * code.
1831 */
1832 if (((sva + PAGE_SIZE) == eva) &&
1833 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
0f7a3396
MD
1834 pmap_remove_page(pmap, sva, &info);
1835 pmap_inval_flush(&info);
984263bc
MD
1836 return;
1837 }
1838
984263bc
MD
1839 /*
1840 * Get a local virtual address for the mappings that are being
1841 * worked with.
1842 */
984263bc
MD
1843 sindex = i386_btop(sva);
1844 eindex = i386_btop(eva);
1845
1846 for (; sindex < eindex; sindex = pdnxt) {
1847 unsigned pdirindex;
1848
1849 /*
1850 * Calculate index for next page table.
1851 */
1852 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1853 if (pmap->pm_stats.resident_count == 0)
1854 break;
1855
1856 pdirindex = sindex / NPDEPG;
1857 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
0f7a3396 1858 pmap_inval_add(&info, pmap, -1);
984263bc
MD
1859 pmap->pm_pdir[pdirindex] = 0;
1860 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
be3aecf7 1861 pmap->pm_cached = 0;
984263bc
MD
1862 continue;
1863 }
1864
1865 /*
1866 * Weed out invalid mappings. Note: we assume that the page
1867 * directory table is always allocated, and in kernel virtual.
1868 */
1869 if (ptpaddr == 0)
1870 continue;
1871
1872 /*
1873 * Limit our scan to either the end of the va represented
1874 * by the current page table page, or to the end of the
1875 * range being removed.
1876 */
1877 if (pdnxt > eindex) {
1878 pdnxt = eindex;
1879 }
1880
8790d7d8 1881 /*
5926987a
MD
1882 * NOTE: pmap_remove_pte() can block and wipe the temporary
1883 * ptbase.
8790d7d8 1884 */
0f7a3396 1885 for (; sindex != pdnxt; sindex++) {
984263bc 1886 vm_offset_t va;
8790d7d8
MD
1887
1888 ptbase = get_ptbase(pmap);
0f7a3396 1889 if (ptbase[sindex] == 0)
984263bc 1890 continue;
984263bc 1891 va = i386_ptob(sindex);
0f7a3396 1892 if (pmap_remove_pte(pmap, ptbase + sindex, va, &info))
984263bc
MD
1893 break;
1894 }
1895 }
0f7a3396 1896 pmap_inval_flush(&info);
984263bc
MD
1897}
1898
1899/*
e0e69b7d
MD
1900 * pmap_remove_all:
1901 *
1902 * Removes this physical page from all physical maps in which it resides.
1903 * Reflects back modify bits to the pager.
984263bc 1904 *
e0e69b7d 1905 * This routine may not be called from an interrupt.
984263bc
MD
1906 */
1907
1908static void
840de426 1909pmap_remove_all(vm_page_t m)
984263bc 1910{
0f7a3396 1911 struct pmap_inval_info info;
840de426 1912 unsigned *pte, tpte;
0f7a3396 1913 pv_entry_t pv;
984263bc 1914
bee81bdd
SS
1915 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
1916 return;
984263bc 1917
0f7a3396 1918 pmap_inval_init(&info);
9acd5bbb 1919 crit_enter();
984263bc 1920 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
eec2b734
MD
1921 KKASSERT(pv->pv_pmap->pm_stats.resident_count > 0);
1922 --pv->pv_pmap->pm_stats.resident_count;
984263bc
MD
1923
1924 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
0f7a3396 1925 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
984263bc 1926 tpte = loadandclear(pte);
5926987a
MD
1927#ifdef PMAP_DEBUG
1928 KKASSERT(PHYS_TO_VM_PAGE(tpte) == m);
1929#endif
984263bc
MD
1930 if (tpte & PG_W)
1931 pv->pv_pmap->pm_stats.wired_count--;
1932
1933 if (tpte & PG_A)
1934 vm_page_flag_set(m, PG_REFERENCED);
1935
1936 /*
1937 * Update the vm_page_t clean and reference bits.
1938 */
1939 if (tpte & PG_M) {
1940#if defined(PMAP_DIAGNOSTIC)
1941 if (pmap_nw_modified((pt_entry_t) tpte)) {
d557216f
MD
1942 kprintf("pmap_remove_all: modified page "
1943 "not writable: va: %p, pte: 0x%lx\n",
1944 (void *)pv->pv_va, (long)tpte);
984263bc
MD
1945 }
1946#endif
1947 if (pmap_track_modified(pv->pv_va))
1948 vm_page_dirty(m);
1949 }
5926987a
MD
1950#ifdef PMAP_DEBUG
1951 KKASSERT(pv->pv_m == m);
1952#endif
984263bc 1953 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
8790d7d8
MD
1954 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1955 ++pv->pv_pmap->pm_generation;
984263bc 1956 m->md.pv_list_count--;
17cde63e
MD
1957 if (TAILQ_EMPTY(&m->md.pv_list))
1958 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
0f7a3396 1959 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem, &info);
984263bc
MD
1960 free_pv_entry(pv);
1961 }
9acd5bbb 1962 crit_exit();
17cde63e 1963 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
0f7a3396 1964 pmap_inval_flush(&info);
984263bc
MD
1965}
1966
1967/*
e0e69b7d
MD
1968 * pmap_protect:
1969 *
1970 * Set the physical protection on the specified range of this map
1971 * as requested.
1972 *
1973 * This function may not be called from an interrupt if the map is
1974 * not the kernel_pmap.
984263bc
MD
1975 */
1976void
1977pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1978{
840de426 1979 unsigned *ptbase;
984263bc
MD
1980 vm_offset_t pdnxt, ptpaddr;
1981 vm_pindex_t sindex, eindex;
0f7a3396 1982 pmap_inval_info info;
984263bc
MD
1983
1984 if (pmap == NULL)
1985 return;
1986
1987 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1988 pmap_remove(pmap, sva, eva);
1989 return;
1990 }
1991
1992 if (prot & VM_PROT_WRITE)
1993 return;
1994
0f7a3396 1995 pmap_inval_init(&info);
984263bc
MD
1996
1997 ptbase = get_ptbase(pmap);
1998
1999 sindex = i386_btop(sva);
2000 eindex = i386_btop(eva);
2001
2002 for (; sindex < eindex; sindex = pdnxt) {
2003
2004 unsigned pdirindex;
2005
2006 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
2007
2008 pdirindex = sindex / NPDEPG;
2009 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
0f7a3396 2010 pmap_inval_add(&info, pmap, -1);
55f2596a 2011 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
984263bc 2012 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
984263bc
MD
2013 continue;
2014 }
2015
2016 /*
2017 * Weed out invalid mappings. Note: we assume that the page
2018 * directory table is always allocated, and in kernel virtual.
2019 */
2020 if (ptpaddr == 0)
2021 continue;
2022
2023 if (pdnxt > eindex) {
2024 pdnxt = eindex;
2025 }
2026
2027 for (; sindex != pdnxt; sindex++) {
2028
2029 unsigned pbits;
2030 vm_page_t m;
2031
17cde63e
MD
2032 /*
2033 * XXX non-optimal. Note also that there can be
2034 * no pmap_inval_flush() calls until after we modify
2035 * ptbase[sindex] (or otherwise we have to do another
2036 * pmap_inval_add() call).
2037 */
0f7a3396 2038 pmap_inval_add(&info, pmap, i386_ptob(sindex));
984263bc
MD
2039 pbits = ptbase[sindex];
2040
2041 if (pbits & PG_MANAGED) {
2042 m = NULL;
2043 if (pbits & PG_A) {
2044 m = PHYS_TO_VM_PAGE(pbits);
2045 vm_page_flag_set(m, PG_REFERENCED);
2046 pbits &= ~PG_A;
2047 }
2048 if (pbits & PG_M) {
2049 if (pmap_track_modified(i386_ptob(sindex))) {
2050 if (m == NULL)
2051 m = PHYS_TO_VM_PAGE(pbits);
2052 vm_page_dirty(m);
2053 pbits &= ~PG_M;
2054 }
2055 }
2056 }
2057
2058 pbits &= ~PG_RW;
2059
2060 if (pbits != ptbase[sindex]) {
2061 ptbase[sindex] = pbits;
984263bc
MD
2062 }
2063 }
2064 }
0f7a3396 2065 pmap_inval_flush(&info);
984263bc
MD
2066}
2067
2068/*
2069 * Insert the given physical page (p) at
2070 * the specified virtual address (v) in the
2071 * target physical map with the protection requested.
2072 *
2073 * If specified, the page will be wired down, meaning
2074 * that the related pte can not be reclaimed.
2075 *
2076 * NB: This is the only routine which MAY NOT lazy-evaluate
2077 * or lose information. That is, this routine must actually
2078 * insert this page into the given map NOW.
2079 */
2080void
2081pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2082 boolean_t wired)
2083{
6ef943a3 2084 vm_paddr_t pa;
840de426 2085 unsigned *pte;
6ef943a3 2086 vm_paddr_t opa;
984263bc
MD
2087 vm_offset_t origpte, newpte;
2088 vm_page_t mpte;
0f7a3396 2089 pmap_inval_info info;
984263bc
MD
2090
2091 if (pmap == NULL)
2092 return;
2093
2094 va &= PG_FRAME;
2095#ifdef PMAP_DIAGNOSTIC
c439ad8f 2096 if (va >= KvaEnd)
984263bc 2097 panic("pmap_enter: toobig");
d557216f
MD
2098 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) {
2099 panic("pmap_enter: invalid to pmap_enter page "
2100 "table pages (va: %p)", (void *)va);
2101 }
984263bc 2102#endif
fbbaeba3
MD
2103 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2104 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
1e5fb84b 2105 print_backtrace();
fbbaeba3
MD
2106 }
2107 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2108 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
1e5fb84b 2109 print_backtrace();
fbbaeba3 2110 }
984263bc 2111
984263bc
MD
2112 /*
2113 * In the case that a page table page is not
2114 * resident, we are creating it here.
2115 */
17cde63e 2116 if (va < UPT_MIN_ADDRESS)
984263bc 2117 mpte = pmap_allocpte(pmap, va);
17cde63e
MD
2118 else
2119 mpte = NULL;
984263bc 2120
0f7a3396 2121 pmap_inval_init(&info);
984263bc
MD
2122 pte = pmap_pte(pmap, va);
2123
2124 /*
2125 * Page Directory table entry not valid, we need a new PT page
2126 */
2127 if (pte == NULL) {
d557216f
MD
2128 panic("pmap_enter: invalid page directory pdir=0x%lx, va=%p\n",
2129 (long)pmap->pm_pdir[PTDPTDI], (void *)va);
984263bc
MD
2130 }
2131
2132 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
2133 origpte = *(vm_offset_t *)pte;
2134 opa = origpte & PG_FRAME;
2135
2136 if (origpte & PG_PS)
2137 panic("pmap_enter: attempted pmap_enter on 4MB page");
2138
2139 /*
2140 * Mapping has not changed, must be protection or wiring change.
2141 */
2142 if (origpte && (opa == pa)) {
2143 /*
2144 * Wiring change, just update stats. We don't worry about
2145 * wiring PT pages as they remain resident as long as there
2146 * are valid mappings in them. Hence, if a user page is wired,
2147 * the PT page will be also.
2148 */
2149 if (wired && ((origpte & PG_W) == 0))
2150 pmap->pm_stats.wired_count++;
2151 else if (!wired && (origpte & PG_W))
2152 pmap->pm_stats.wired_count--;
2153
2154#if defined(PMAP_DIAGNOSTIC)
2155 if (pmap_nw_modified((pt_entry_t) origpte)) {
d557216f
MD
2156 kprintf("pmap_enter: modified page not "
2157 "writable: va: %p, pte: 0x%lx\n",
2158 (void *)va, (long )origpte);
984263bc
MD
2159 }
2160#endif
2161
2162 /*
639a9b43
MD
2163 * Remove the extra pte reference. Note that we cannot
2164 * optimize the RO->RW case because we have adjusted the
2165 * wiring count above and may need to adjust the wiring
2166 * bits below.
984263bc
MD
2167 */
2168 if (mpte)
2169 mpte->hold_count--;
2170
984263bc
MD
2171 /*
2172 * We might be turning off write access to the page,
2173 * so we go ahead and sense modify status.
2174 */
2175 if (origpte & PG_MANAGED) {
2176 if ((origpte & PG_M) && pmap_track_modified(va)) {
2177 vm_page_t om;
2178 om = PHYS_TO_VM_PAGE(opa);
2179 vm_page_dirty(om);
2180 }
2181 pa |= PG_MANAGED;
17cde63e 2182 KKASSERT(m->flags & PG_MAPPED);
984263bc
MD
2183 }
2184 goto validate;
2185 }
2186 /*
2187 * Mapping has changed, invalidate old range and fall through to
2188 * handle validating new mapping.
5926987a
MD
2189 *
2190 * Since we have a ref on the page directory page pmap_pte()
2191 * will always return non-NULL.
2192 *
2193 * NOTE: pmap_remove_pte() can block and cause the temporary ptbase
2194 * to get wiped. reload the ptbase. I'm not sure if it is
2195 * also possible to race another pmap_enter() but check for
2196 * that case too.
984263bc 2197 */
5926987a 2198 while (opa) {
984263bc 2199 int err;
5926987a
MD
2200
2201 KKASSERT((origpte & PG_FRAME) ==
2202 (*(vm_offset_t *)pte & PG_FRAME));
0f7a3396 2203 err = pmap_remove_pte(pmap, pte, va, &info);
984263bc 2204 if (err)
d557216f 2205 panic("pmap_enter: pte vanished, va: %p", (void *)va);
5926987a
MD
2206 pte = pmap_pte(pmap, va);
2207 origpte = *(vm_offset_t *)pte;
2208 opa = origpte & PG_FRAME;
2209 if (opa) {
2210 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2211 pmap, (void *)va);
2212 }
984263bc
MD
2213 }
2214
2215 /*
2216 * Enter on the PV list if part of our managed memory. Note that we
2217 * raise IPL while manipulating pv_table since pmap_enter can be
2218 * called at interrupt time.
2219 */
2220 if (pmap_initialized &&
2221 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2222 pmap_insert_entry(pmap, va, mpte, m);
5926987a 2223 ptbase_assert(pmap);
984263bc 2224 pa |= PG_MANAGED;
17cde63e 2225 vm_page_flag_set(m, PG_MAPPED);
984263bc
MD
2226 }
2227
2228 /*
2229 * Increment counters
2230 */
eec2b734 2231 ++pmap->pm_stats.resident_count;
984263bc
MD
2232 if (wired)
2233 pmap->pm_stats.wired_count++;
5926987a 2234 KKASSERT(*pte == 0);
984263bc
MD
2235
2236validate:
2237 /*
2238 * Now validate mapping with desired protection/wiring.
2239 */
5926987a 2240 ptbase_assert(pmap);
984263bc
MD
2241 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2242
2243 if (wired)
2244 newpte |= PG_W;
2245 if (va < UPT_MIN_ADDRESS)
2246 newpte |= PG_U;
fbbaeba3 2247 if (pmap == &kernel_pmap)
984263bc
MD
2248 newpte |= pgeflag;
2249
2250 /*
2251 * if the mapping or permission bits are different, we need
2252 * to update the pte.
2253 */
2254 if ((origpte & ~(PG_M|PG_A)) != newpte) {
17cde63e 2255 pmap_inval_add(&info, pmap, va);
5926987a
MD
2256 ptbase_assert(pmap);
2257 KKASSERT(*pte == 0 ||
2258 (*pte & PG_FRAME) == (newpte & PG_FRAME));
984263bc 2259 *pte = newpte | PG_A;
17cde63e
MD
2260 if (newpte & PG_RW)
2261 vm_page_flag_set(m, PG_WRITEABLE);
984263bc 2262 }
c695044a 2263 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
0f7a3396 2264 pmap_inval_flush(&info);
984263bc
MD
2265}
2266
2267/*
17cde63e
MD
2268 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2269 * This code also assumes that the pmap has no pre-existing entry for this
2270 * VA.
2271 *
2272 * This code currently may only be used on user pmaps, not kernel_pmap.
984263bc 2273 */
1b9d3514 2274void
17cde63e 2275pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
984263bc
MD
2276{
2277 unsigned *pte;
6ef943a3 2278 vm_paddr_t pa;
17cde63e
MD
2279 vm_page_t mpte;
2280 unsigned ptepindex;
2281 vm_offset_t ptepa;
0f7a3396
MD
2282 pmap_inval_info info;
2283
2284 pmap_inval_init(&info);
984263bc 2285
fbbaeba3
MD
2286 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2287 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
1e5fb84b 2288 print_backtrace();
fbbaeba3
MD
2289 }
2290 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2291 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
1e5fb84b 2292 print_backtrace();
fbbaeba3
MD
2293 }
2294
17cde63e
MD
2295 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2296
984263bc 2297 /*
17cde63e
MD
2298 * Calculate the page table page (mpte), allocating it if necessary.
2299 *
2300 * A held page table page (mpte), or NULL, is passed onto the
2301 * section following.
984263bc
MD
2302 */
2303 if (va < UPT_MIN_ADDRESS) {
984263bc
MD
2304 /*
2305 * Calculate pagetable page index
2306 */
2307 ptepindex = va >> PDRSHIFT;
17cde63e
MD
2308
2309 do {
984263bc
MD
2310 /*
2311 * Get the page directory entry
2312 */
2313 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2314
2315 /*
2316 * If the page table page is mapped, we just increment
2317 * the hold count, and activate it.
2318 */
2319 if (ptepa) {
2320 if (ptepa & PG_PS)
2321 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2322 if (pmap->pm_ptphint &&
17cde63e 2323 (pmap->pm_ptphint->pindex == ptepindex)) {
984263bc
MD
2324 mpte = pmap->pm_ptphint;
2325 } else {
2326 mpte = pmap_page_lookup( pmap->pm_pteobj, ptepindex);
2327 pmap->pm_ptphint = mpte;
2328 }
17cde63e
MD
2329 if (mpte)
2330 mpte->hold_count++;
984263bc
MD
2331 } else {
2332 mpte = _pmap_allocpte(pmap, ptepindex);
2333 }
17cde63e 2334 } while (mpte == NULL);
984263bc
MD
2335 } else {
2336 mpte = NULL;
17cde63e 2337 /* this code path is not yet used */
984263bc
MD
2338 }
2339
2340 /*
17cde63e
MD
2341 * With a valid (and held) page directory page, we can just use
2342 * vtopte() to get to the pte. If the pte is already present
2343 * we do not disturb it.
984263bc
MD
2344 */
2345 pte = (unsigned *)vtopte(va);
17cde63e 2346 if (*pte & PG_V) {
984263bc 2347 if (mpte)
0f7a3396 2348 pmap_unwire_pte_hold(pmap, mpte, &info);
17cde63e
MD
2349 pa = VM_PAGE_TO_PHYS(m);
2350 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
2351 return;
984263bc
MD
2352 }
2353
2354 /*
17cde63e 2355 * Enter on the PV list if part of our managed memory
984263bc 2356 */
17cde63e 2357 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
984263bc 2358 pmap_insert_entry(pmap, va, mpte, m);
17cde63e
MD
2359 vm_page_flag_set(m, PG_MAPPED);
2360 }
984263bc
MD
2361
2362 /*
2363 * Increment counters
2364 */
eec2b734 2365 ++pmap->pm_stats.resident_count;
984263bc
MD
2366
2367 pa = VM_PAGE_TO_PHYS(m);
2368
2369 /*
2370 * Now validate mapping with RO protection
2371 */
2372 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2373 *pte = pa | PG_V | PG_U;
2374 else
2375 *pte = pa | PG_V | PG_U | PG_MANAGED;
17cde63e
MD
2376/* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
2377 pmap_inval_flush(&info);
984263bc
MD
2378}
2379
2380/*
2381 * Make a temporary mapping for a physical address. This is only intended
2382 * to be used for panic dumps.
2383 */
2384void *
6ef943a3 2385pmap_kenter_temporary(vm_paddr_t pa, int i)
984263bc
MD
2386{
2387 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2388 return ((void *)crashdumpmap);
2389}
2390
2391#define MAX_INIT_PT (96)
06ecca5a 2392
984263bc 2393/*
06ecca5a
MD
2394 * This routine preloads the ptes for a given object into the specified pmap.
2395 * This eliminates the blast of soft faults on process startup and
2396 * immediately after an mmap.
984263bc 2397 */
1f804340
MD
2398static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2399
984263bc 2400void
083a7402
MD
2401pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2402 vm_object_t object, vm_pindex_t pindex,
2403 vm_size_t size, int limit)
984263bc 2404{
1f804340 2405 struct rb_vm_page_scan_info info;
287ebb09 2406 struct lwp *lp;
984263bc 2407 int psize;
984263bc 2408
54a764e8
MD
2409 /*
2410 * We can't preinit if read access isn't set or there is no pmap
2411 * or object.
2412 */
083a7402 2413 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
984263bc
MD
2414 return;
2415
54a764e8
MD
2416 /*
2417 * We can't preinit if the pmap is not the current pmap
2418 */
287ebb09
MD
2419 lp = curthread->td_lwp;
2420 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
54a764e8
MD
2421 return;
2422
984263bc
MD
2423 psize = i386_btop(size);
2424
2425 if ((object->type != OBJT_VNODE) ||
2426 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2427 (object->resident_page_count > MAX_INIT_PT))) {
2428 return;
2429 }
2430
2431 if (psize + pindex > object->size) {
2432 if (object->size < pindex)
2433 return;
2434 psize = object->size - pindex;
2435 }
2436
1f804340
MD
2437 if (psize == 0)
2438 return;
06ecca5a 2439
984263bc 2440 /*
1f804340
MD
2441 * Use a red-black scan to traverse the requested range and load
2442 * any valid pages found into the pmap.
06ecca5a 2443 *
9acd5bbb
MD
2444 * We cannot safely scan the object's memq unless we are in a
2445 * critical section since interrupts can remove pages from objects.
984263bc 2446 */
1f804340
MD
2447 info.start_pindex = pindex;
2448 info.end_pindex = pindex + psize - 1;
2449 info.limit = limit;
2450 info.mpte = NULL;
2451 info.addr = addr;
2452 info.pmap = pmap;
2453
654a39f0 2454 crit_enter();
1f804340
MD
2455 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2456 pmap_object_init_pt_callback, &info);
2457 crit_exit();
2458}
06ecca5a 2459
1f804340
MD
2460static
2461int
2462pmap_object_init_pt_callback(vm_page_t p, void *data)
2463{
2464 struct rb_vm_page_scan_info *info = data;
2465 vm_pindex_t rel_index;
2466 /*
2467 * don't allow an madvise to blow away our really
2468 * free pages allocating pv entries.
2469 */
2470 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2471 vmstats.v_free_count < vmstats.v_free_reserved) {
2472 return(-1);
984263bc 2473 }
1f804340
MD
2474 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2475 (p->busy == 0) && (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2476 if ((p->queue - p->pc) == PQ_CACHE)
2477 vm_page_deactivate(p);
2478 vm_page_busy(p);
2479 rel_index = p->pindex - info->start_pindex;
17cde63e
MD
2480 pmap_enter_quick(info->pmap,
2481 info->addr + i386_ptob(rel_index), p);
1f804340
MD
2482 vm_page_wakeup(p);
2483 }
2484 return(0);
984263bc
MD
2485}
2486
2487/*
1b9d3514
MD
2488 * Return TRUE if the pmap is in shape to trivially
2489 * pre-fault the specified address.
2490 *
2491 * Returns FALSE if it would be non-trivial or if a
2492 * pte is already loaded into the slot.
984263bc 2493 */
1b9d3514
MD
2494int
2495pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
984263bc 2496{
1b9d3514 2497 unsigned *pte;
984263bc 2498
1b9d3514
MD
2499 if ((*pmap_pde(pmap, addr)) == 0)
2500 return(0);
2501 pte = (unsigned *) vtopte(addr);
2502 if (*pte)
2503 return(0);
2504 return(1);
984263bc
MD
2505}
2506
2507/*
2508 * Routine: pmap_change_wiring
2509 * Function: Change the wiring attribute for a map/virtual-address
2510 * pair.
2511 * In/out conditions:
2512 * The mapping must already exist in the pmap.
2513 */
2514void
840de426 2515pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
984263bc 2516{
840de426 2517 unsigned *pte;
984263bc
MD
2518
2519 if (pmap == NULL)
2520 return;
2521
2522 pte = pmap_pte(pmap, va);
2523
2524 if (wired && !pmap_pte_w(pte))
2525 pmap->pm_stats.wired_count++;
2526 else if (!wired && pmap_pte_w(pte))
2527 pmap->pm_stats.wired_count--;
2528
2529 /*
2530 * Wiring is not a hardware characteristic so there is no need to
0f7a3396
MD
2531 * invalidate TLB. However, in an SMP environment we must use
2532 * a locked bus cycle to update the pte (if we are not using
2533 * the pmap_inval_*() API that is)... it's ok to do this for simple
2534 * wiring changes.
984263bc 2535 */
0f7a3396
MD
2536#ifdef SMP
2537 if (wired)
2538 atomic_set_int(pte, PG_W);
2539 else
2540 atomic_clear_int(pte, PG_W);
2541#else
2542 if (wired)
2543 atomic_set_int_nonlocked(pte, PG_W);
2544 else
2545 atomic_clear_int_nonlocked(pte, PG_W);
2546#endif
984263bc
MD
2547}
2548
2549
2550
2551/*
2552 * Copy the range specified by src_addr/len
2553 * from the source map to the range dst_addr/len
2554 * in the destination map.
2555 *
2556 * This routine is only advisory and need not do anything.
2557 */
984263bc 2558void
840de426
MD
2559pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
2560 vm_size_t len, vm_offset_t src_addr)
984263bc 2561{
5926987a 2562#if 0
0f7a3396 2563 pmap_inval_info info;
984263bc
MD
2564 vm_offset_t addr;
2565 vm_offset_t end_addr = src_addr + len;
2566 vm_offset_t pdnxt;
2567 unsigned src_frame, dst_frame;
2568 vm_page_t m;
5926987a 2569#endif
984263bc
MD
2570
2571 if (dst_addr != src_addr)
2572 return;
17cde63e
MD
2573 /*
2574 * XXX BUGGY. Amoung other things srcmpte is assumed to remain
2575 * valid through blocking calls, and that's just not going to
2576 * be the case.
2577 *
2578 * FIXME!
2579 */
2580 return;
5926987a 2581#if 0
984263bc
MD
2582 src_frame = ((unsigned) src_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
2583 if (src_frame != (((unsigned) PTDpde) & PG_FRAME)) {
2584 return;
2585 }
2586
2587 dst_frame = ((unsigned) dst_pmap->pm_pdir[PTDPTDI]) & PG_FRAME;
c1692ddf
MD
2588 if ((*gd->gd_GDMAP1 & PG_FRAME) != dst_frame) {
2589 *gd->gd_GDMAP1 = dst_frame | PG_RW | PG_V;
984263bc 2590 cpu_invltlb();
be3aecf7 2591 XXX
984263bc 2592 }
0f7a3396
MD
2593 pmap_inval_init(&info);
2594 pmap_inval_add(&info, dst_pmap, -1);
2595 pmap_inval_add(&info, src_pmap, -1);
984263bc 2596
06ecca5a 2597 /*
654a39f0 2598 * critical section protection is required to maintain the page/object
06ecca5a
MD
2599 * association, interrupts can free pages and remove them from
2600 * their objects.
2601 */
654a39f0 2602 crit_enter();
06ecca5a 2603 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
984263bc
MD
2604 unsigned *src_pte, *dst_pte;
2605 vm_page_t dstmpte, srcmpte;
2606 vm_offset_t srcptepaddr;
2607 unsigned ptepindex;
2608
2609 if (addr >= UPT_MIN_ADDRESS)
2610 panic("pmap_copy: invalid to pmap_copy page tables\n");
2611
2612 /*
2613 * Don't let optional prefaulting of pages make us go
2614 * way below the low water mark of free pages or way
2615 * above high water mark of used pv entries.
2616 */
12e4aaff 2617 if (vmstats.v_free_count < vmstats.v_free_reserved ||
984263bc
MD
2618 pv_entry_count > pv_entry_high_water)
2619 break;
2620
2621 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2622 ptepindex = addr >> PDRSHIFT;
2623
2624 srcptepaddr = (vm_offset_t) src_pmap->pm_pdir[ptepindex];
2625 if (srcptepaddr == 0)
2626 continue;
2627
2628 if (srcptepaddr & PG_PS) {
2629 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2630 dst_pmap->pm_pdir[ptepindex] = (pd_entry_t) srcptepaddr;
2631 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
be3aecf7 2632 XXX
984263bc
MD
2633 }
2634 continue;
2635 }
2636
2637 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
17cde63e
MD
2638 if ((srcmpte == NULL) || (srcmpte->hold_count == 0) ||
2639 (srcmpte->flags & PG_BUSY)) {
984263bc 2640 continue;
17cde63e 2641 }
984263bc
MD
2642
2643 if (pdnxt > end_addr)
2644 pdnxt = end_addr;
2645
2646 src_pte = (unsigned *) vtopte(addr);
2647 dst_pte = (unsigned *) avtopte(addr);
2648 while (addr < pdnxt) {
2649 unsigned ptetemp;
5e8d0349 2650
984263bc
MD
2651 ptetemp = *src_pte;
2652 /*
2653 * we only virtual copy managed pages
2654 */
2655 if ((ptetemp & PG_MANAGED) != 0) {
2656 /*
2657 * We have to check after allocpte for the
2658 * pte still being around... allocpte can
2659 * block.
eec2b734
MD
2660 *
2661 * pmap_allocpte() can block. If we lose
2662 * our page directory mappings we stop.
984263bc
MD
2663 */
2664 dstmpte = pmap_allocpte(dst_pmap, addr);
eec2b734
MD
2665
2666 if (src_frame != (((unsigned) PTDpde) & PG_FRAME) ||
c1692ddf 2667 XXX dst_frame != (((unsigned) xxx) & PG_FRAME)
eec2b734
MD
2668 ) {
2669 kprintf("WARNING: pmap_copy: detected and corrected race\n");
2670 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
2671 goto failed;
17cde63e
MD
2672 } else if ((*dst_pte == 0) &&
2673 (ptetemp = *src_pte) != 0 &&
2674 (ptetemp & PG_MANAGED)) {
984263bc
MD
2675 /*
2676 * Clear the modified and
2677 * accessed (referenced) bits
2678 * during the copy.
2679 */
2680 m = PHYS_TO_VM_PAGE(ptetemp);
70fc5283 2681 *dst_pte = ptetemp & ~(PG_M | PG_A);
eec2b734 2682 ++dst_pmap->pm_stats.resident_count;
984263bc
MD
2683 pmap_insert_entry(dst_pmap, addr,
2684 dstmpte, m);
17cde63e 2685 KKASSERT(m->flags & PG_MAPPED);
984263bc 2686 } else {
17cde63e 2687 kprintf("WARNING: pmap_copy: dst_pte race detected and corrected\n");
0f7a3396 2688 pmap_unwire_pte_hold(dst_pmap, dstmpte, &info);
17cde63e 2689 goto failed;
984263bc
MD
2690 }
2691 if (dstmpte->hold_count >= srcmpte->hold_count)
2692 break;
2693 }
2694 addr += PAGE_SIZE;
2695 src_pte++;
2696 dst_pte++;
2697 }
2698 }
eec2b734 2699failed:
654a39f0 2700 crit_exit();
0f7a3396 2701 pmap_inval_flush(&info);
5926987a 2702#endif
984263bc
MD
2703}
2704
984263bc 2705/*
e0e69b7d
MD
2706 * pmap_zero_page:
2707 *
2708 * Zero the specified PA by mapping the page into KVM and clearing its
2709 * contents.
2710 *
2711 * This function may be called from an interrupt and no locking is
2712 * required.
984263bc
MD
2713 */
2714void
6ef943a3 2715pmap_zero_page(vm_paddr_t phys)
984263bc 2716{
85100692 2717 struct mdglobaldata *gd = mdcpu;
17a9f566 2718
e0e69b7d 2719 crit_enter();
85100692
MD
2720 if (*(int *)gd->gd_CMAP3)
2721 panic("pmap_zero_page: CMAP3 busy");
85100692 2722 *(int *)gd->gd_CMAP3 =
17a9f566 2723 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
85100692 2724 cpu_invlpg(gd->gd_CADDR3);
984263bc
MD
2725
2726#if defined(I686_CPU)
2727 if (cpu_class == CPUCLASS_686)
85100692 2728 i686_pagezero(gd->gd_CADDR3);
984263bc
MD
2729 else
2730#endif
85100692 2731 bzero(gd->gd_CADDR3, PAGE_SIZE);
85100692 2732 *(int *) gd->gd_CMAP3 = 0;
e0e69b7d 2733 crit_exit();
8100156a
MD
2734}
2735
2736/*
2737 * pmap_page_assertzero:
2738 *
2739 * Assert that a page is empty, panic if it isn't.
2740 */
2741void
2742pmap_page_assertzero(vm_paddr_t phys)
2743{
2744 struct mdglobaldata *gd = mdcpu;
2745 int i;
2746
2747 crit_enter();
2748 if (*(int *)gd->gd_CMAP3)
2749 panic("pmap_zero_page: CMAP3 busy");
2750 *(int *)gd->gd_CMAP3 =
2751 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2752 cpu_invlpg(gd->gd_CADDR3);
2753 for (i = 0; i < PAGE_SIZE; i += 4) {
2754 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
2755 panic("pmap_page_assertzero() @ %p not zero!\n",
2756 (void *)gd->gd_CADDR3);
2757 }
2758 }
2759 *(int *) gd->gd_CMAP3 = 0;
2760 crit_exit();
984263bc
MD
2761}
2762
2763/*
e0e69b7d
MD
2764 * pmap_zero_page:
2765 *
2766 * Zero part of a physical page by mapping it into memory and clearing
2767 * its contents with bzero.
984263bc
MD
2768 *
2769 * off and size may not cover an area beyond a single hardware page.
2770 */
2771void
6ef943a3 2772pmap_zero_page_area(vm_paddr_t phys, int off, int size)
984263bc 2773{
85100692 2774 struct mdglobaldata *gd = mdcpu;
17a9f566 2775
e0e69b7d 2776 crit_enter();
85100692
MD
2777 if (*(int *) gd->gd_CMAP3)
2778 panic("pmap_zero_page: CMAP3 busy");
85100692
MD
2779 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2780 cpu_invlpg(gd->gd_CADDR3);
984263bc
MD
2781
2782#if defined(I686_CPU)
2783 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
85100692 2784 i686_pagezero(gd->gd_CADDR3);
984263bc
MD
2785 else
2786#endif
85100692 2787 bzero((char *)gd->gd_CADDR3 + off, size);
85100692 2788 *(int *) gd->gd_CMAP3 = 0;
e0e69b7d 2789 crit_exit();
984263bc
MD
2790}
2791
2792/*
e0e69b7d
MD
2793 * pmap_copy_page:
2794 *
2795 * Copy the physical page from the source PA to the target PA.
2796 * This function may be called from an interrupt. No locking
2797 * is required.
984263bc
MD
2798 */
2799void
6ef943a3 2800pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
984263bc 2801{
85100692 2802 struct mdglobaldata *gd = mdcpu;
17a9f566 2803
e0e69b7d 2804 crit_enter();
85100692
MD
2805 if (*(int *) gd->gd_CMAP1)
2806 panic("pmap_copy_page: CMAP1 busy");
2807 if (*(int *) gd->gd_CMAP2)
2808 panic("pmap_copy_page: CMAP2 busy");
984263bc 2809
85100692
MD
2810 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2811 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
984263bc 2812
85100692
MD
2813 cpu_invlpg(gd->gd_CADDR1);
2814 cpu_invlpg(gd->gd_CADDR2);
984263bc 2815
85100692 2816 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
984263bc 2817
85100692
MD
2818 *(int *) gd->gd_CMAP1 = 0;
2819 *(int *) gd->gd_CMAP2 = 0;
e0e69b7d 2820 crit_exit();
984263bc
MD
2821}
2822
f6bf3af1
MD
2823/*
2824 * pmap_copy_page_frag:
2825 *
2826 * Copy the physical page from the source PA to the target PA.
2827 * This function may be called from an interrupt. No locking
2828 * is required.
2829 */
2830void
2831pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2832{
2833 struct mdglobaldata *gd = mdcpu;
2834
2835 crit_enter();
2836 if (*(int *) gd->gd_CMAP1)
2837 panic("pmap_copy_page: CMAP1 busy");
2838 if (*(int *) gd->gd_CMAP2)
2839 panic("pmap_copy_page: CMAP2 busy");
2840
2841 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2842 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2843
2844 cpu_invlpg(gd->gd_CADDR1);
2845 cpu_invlpg(gd->gd_CADDR2);
2846
2847 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2848 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2849 bytes);
2850
2851 *(int *) gd->gd_CMAP1 = 0;
2852 *(int *) gd->gd_CMAP2 = 0;
2853 crit_exit();
2854}
2855
984263bc
MD
2856/*
2857 * Returns true if the pmap's pv is one of the first
2858 * 16 pvs linked to from this page. This count may
2859 * be changed upwards or downwards in the future; it
2860 * is only necessary that true be returned for a small
2861 * subset of pmaps for proper page aging.
2862 */
2863boolean_t
840de426 2864pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
984263bc
MD
2865{
2866 pv_entry_t pv;
2867 int loops = 0;
984263bc
MD
2868
2869 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2870 return FALSE;
2871
9acd5bbb 2872 crit_enter();
984263bc
MD
2873
2874 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2875 if (pv->pv_pmap == pmap) {
9acd5bbb 2876 crit_exit();
984263bc
MD
2877 return TRUE;
2878 }
2879 loops++;
2880 if (loops >= 16)
2881 break;
2882 }
9acd5bbb 2883 crit_exit();
984263bc
MD
2884 return (FALSE);
2885}
2886
984263bc
MD
2887/*
2888 * Remove all pages from specified address space
2889 * this aids process exit speeds. Also, this code
2890 * is special cased for current process only, but
2891 * can have the more generic (and slightly slower)
2892 * mode enabled. This is much faster than pmap_remove
2893 * in the case of running down an entire address space.
2894 */
2895void
840de426 2896pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
984263bc 2897{
287ebb09 2898 struct lwp *lp;
984263bc
MD
2899 unsigned *pte, tpte;
2900 pv_entry_t pv, npv;
984263bc 2901 vm_page_t m;
0f7a3396 2902 pmap_inval_info info;
4a22e893 2903 int iscurrentpmap;
8790d7d8 2904 int32_t save_generation;
984263bc 2905
287ebb09
MD
2906 lp = curthread->td_lwp;
2907 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
4a22e893
MD
2908 iscurrentpmap = 1;
2909 else
2910 iscurrentpmap = 0;
984263bc 2911
0f7a3396 2912 pmap_inval_init(&info);
9acd5bbb 2913 crit_enter();
4a22e893 2914 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
984263bc
MD
2915 if (pv->pv_va >= eva || pv->pv_va < sva) {
2916 npv = TAILQ_NEXT(pv, pv_plist);
2917 continue;
2918 }
2919
8790d7d8
MD
2920 KKASSERT(pmap == pv->pv_pmap);
2921
4a22e893
MD
2922 if (iscurrentpmap)
2923 pte = (unsigned *)vtopte(pv->pv_va);
2924 else
8790d7d8 2925 pte = pmap_pte_quick(pmap, pv->pv_va);
5926987a 2926 KKASSERT(*pte);
4a22e893 2927 if (pmap->pm_active)
8790d7d8 2928 pmap_inval_add(&info, pmap, pv->pv_va);
984263bc 2929
4a22e893
MD
2930 /*
2931 * We cannot remove wired pages from a process' mapping
2932 * at this time
2933 */
17cde63e 2934 if (*pte & PG_W) {
984263bc
MD
2935 npv = TAILQ_NEXT(pv, pv_plist);
2936 continue;
2937 }
2247fe02 2938 KKASSERT(*pte);
17cde63e 2939 tpte = loadandclear(pte);
984263bc
MD
2940
2941 m = PHYS_TO_VM_PAGE(tpte);
5926987a 2942 test_m_maps_pv(m, pv);
984263bc
MD
2943
2944 KASSERT(m < &vm_page_array[vm_page_array_size],
2945 ("pmap_remove_pages: bad tpte %x", tpte));
2946
eec2b734
MD
2947 KKASSERT(pmap->pm_stats.resident_count > 0);
2948 --pmap->pm_stats.resident_count;
984263bc
MD
2949
2950 /*
2951 * Update the vm_page_t clean and reference bits.
2952 */
2953 if (tpte & PG_M) {
2954 vm_page_dirty(m);
2955 }
2956
984263bc 2957 npv = TAILQ_NEXT(pv, pv_plist);
5926987a
MD
2958#ifdef PMAP_DEBUG
2959 KKASSERT(pv->pv_m == m);
2960 KKASSERT(pv->pv_pmap == pmap);
2961#endif
8790d7d8
MD
2962 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2963 save_generation = ++pmap->pm_generation;
984263bc
MD
2964
2965 m->md.pv_list_count--;
2966 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
17cde63e 2967 if (TAILQ_EMPTY(&m->md.pv_list))
984263bc 2968 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
984263bc 2969
8790d7d8 2970 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
984263bc 2971 free_pv_entry(pv);
8790d7d8
MD
2972
2973 /*
2974 * Restart the scan if we blocked during the unuse or free
2975 * calls and other removals were made.
2976 */
2977 if (save_generation != pmap->pm_generation) {
2978 kprintf("Warning: pmap_remove_pages race-A avoided\n");
5926987a 2979 npv = TAILQ_FIRST(&pmap->pm_pvlist);
8790d7d8 2980 }
984263bc 2981 }
0f7a3396 2982 pmap_inval_flush(&info);
9acd5bbb 2983 crit_exit();
984263bc
MD
2984}
2985
2986/*
2987 * pmap_testbit tests bits in pte's
5e8d0349 2988 * note that the testbit/clearbit routines are inline,
984263bc
MD
2989 * and a lot of things compile-time evaluate.
2990 */
2991static boolean_t
840de426 2992pmap_testbit(vm_page_t m, int bit)
984263bc
MD
2993{
2994 pv_entry_t pv;
2995 unsigned *pte;
984263bc
MD
2996
2997 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2998 return FALSE;
2999
3000 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3001 return FALSE;
3002
9acd5bbb 3003 crit_enter();
984263bc
MD
3004
3005 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3006 /*
3007 * if the bit being tested is the modified bit, then
3008 * mark clean_map and ptes as never
3009 * modified.
3010 */
3011 if (bit & (PG_A|PG_M)) {
3012 if (!pmap_track_modified(pv->pv_va))
3013 continue;
3014 }
3015
3016#if defined(PMAP_DIAGNOSTIC)
3017 if (!pv->pv_pmap) {
d557216f
MD
3018 kprintf("Null pmap (tb) at va: %p\n",
3019 (void *)pv->pv_va);
984263bc
MD
3020 continue;
3021 }
3022#endif
3023 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3024 if (*pte & bit) {
9acd5bbb 3025 crit_exit();
984263bc
MD
3026 return TRUE;
3027 }
3028 }
9acd5bbb 3029 crit_exit();
984263bc
MD
3030 return (FALSE);
3031}
3032
3033/*
3034 * this routine is used to modify bits in ptes
3035 */
3036static __inline void
5e8d0349 3037pmap_clearbit(vm_page_t m, int bit)
984263bc 3038{
0f7a3396 3039 struct pmap_inval_info info;
840de426
MD
3040 pv_entry_t pv;
3041 unsigned *pte;
5e8d0349 3042 unsigned pbits;
984263bc
MD
3043
3044 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3045 return;
3046
0f7a3396 3047 pmap_inval_init(&info);
9acd5bbb 3048 crit_enter();
984263bc
MD
3049
3050 /*
3051 * Loop over all current mappings setting/clearing as appropos If
3052 * setting RO do we need to clear the VAC?
3053 */
3054 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3055 /*
3056 * don't write protect pager mappings
3057 */
5e8d0349 3058 if (bit == PG_RW) {
984263bc
MD
3059 if (!pmap_track_modified(pv->pv_va))
3060 continue;
3061 }
3062
3063#if defined(PMAP_DIAGNOSTIC)
3064 if (!pv->pv_pmap) {
d557216f
MD
3065 kprintf("Null pmap (cb) at va: %p\n",
3066 (void *)pv->pv_va);
984263bc
MD
3067 continue;
3068 }
3069#endif
3070
0f7a3396
MD
3071 /*
3072 * Careful here. We can use a locked bus instruction to
3073 * clear PG_A or PG_M safely but we need to synchronize
3074 * with the target cpus when we mess with PG_RW.
70fc5283
MD
3075 *
3076 * We do not have to force synchronization when clearing
3077 * PG_M even for PTEs generated via virtual memory maps,
3078 * because the virtual kernel will invalidate the pmap
3079 * entry when/if it needs to resynchronize the Modify bit.
0f7a3396 3080 */
70fc5283 3081 if (bit & PG_RW)
0f7a3396 3082 pmap_inval_add(&info, pv->pv_pmap, pv->pv_va);
17cde63e
MD
3083 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3084again:
5e8d0349
MD
3085 pbits = *pte;
3086 if (pbits & bit) {
3087 if (bit == PG_RW) {
17cde63e 3088 if (pbits & PG_M) {
5e8d0349 3089 vm_page_dirty(m);
17cde63e
MD
3090 atomic_clear_int(pte, PG_M|PG_RW);
3091 } else {
3092 /*
3093 * The cpu may be trying to set PG_M
3094 * simultaniously with our clearing
3095 * of PG_RW.
3096 */
3097 if (!atomic_cmpset_int(pte, pbits,
3098 pbits & ~PG_RW))
3099 goto again;
3100 }
5e8d0349
MD
3101 } else if (bit == PG_M) {
3102 /*
70fc5283
MD
3103 * We could also clear PG_RW here to force
3104 * a fault on write to redetect PG_M for
3105 * virtual kernels, but it isn't necessary
3106 * since virtual kernels invalidate the pte
3107 * when they clear the VPTE_M bit in their
3108 * virtual page tables.
5e8d0349 3109 */
70fc5283 3110 atomic_clear_int(pte, PG_M);
5e8d0349
MD
3111 } else {
3112 atomic_clear_int(pte, bit);
984263bc
MD
3113 }
3114 }
3115 }
0f7a3396 3116 pmap_inval_flush(&info);
9acd5bbb 3117 crit_exit();
984263bc
MD
3118}
3119
3120/*
3121 * pmap_page_protect:
3122 *
3123 * Lower the permission for all mappings to a given page.
3124 */
3125void
3126pmap_page_protect(vm_page_t m, vm_prot_t prot)
3127{
3128 if ((prot & VM_PROT_WRITE) == 0) {
3129 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
5e8d0349 3130 pmap_clearbit(m, PG_RW);
17cde63e 3131 vm_page_flag_clear(m, PG_WRITEABLE);
984263bc
MD
3132 } else {
3133 pmap_remove_all(m);
3134 }
3135 }
3136}
3137
6ef943a3 3138vm_paddr_t
cfd17028 3139pmap_phys_address(vm_pindex_t ppn)
984263bc
MD
3140{
3141 return (i386_ptob(ppn));
3142}
3143
3144/*
3145 * pmap_ts_referenced:
3146 *
3147 * Return a count of reference bits for a page, clearing those bits.
3148 * It is not necessary for every reference bit to be cleared, but it
3149 * is necessary that 0 only be returned when there are truly no
3150 * reference bits set.
3151 *
3152 * XXX: The exact number of bits to check and clear is a matter that
3153 * should be tested and standardized at some point in the future for
3154 * optimal aging of shared pages.
3155 */
3156int
3157pmap_ts_referenced(vm_page_t m)
3158{
840de426 3159 pv_entry_t pv, pvf, pvn;
984263bc 3160 unsigned *pte;
984263bc
MD
3161 int rtval = 0;
3162
3163 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3164 return (rtval);
3165
9acd5bbb 3166 crit_enter();
984263bc
MD
3167
3168 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3169
3170 pvf = pv;
3171
3172 do {
3173 pvn = TAILQ_NEXT(pv, pv_list);
3174
5926987a 3175 crit_enter();
984263bc 3176 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
984263bc 3177 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
5926987a 3178 crit_exit();
984263bc
MD
3179
3180 if (!pmap_track_modified(pv->pv_va))
3181 continue;
3182
3183 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3184
3185 if (pte && (*pte & PG_A)) {
0f7a3396
MD
3186#ifdef SMP
3187 atomic_clear_int(pte, PG_A);
3188#else
3189 atomic_clear_int_nonlocked(pte, PG_A);
3190#endif
984263bc
MD
3191 rtval++;
3192 if (rtval > 4) {
3193 break;
3194 }
3195 }
3196 } while ((pv = pvn) != NULL && pv != pvf);
3197 }
9acd5bbb 3198 crit_exit();
984263bc
MD
3199
3200 return (rtval);
3201}
3202
3203/*
3204 * pmap_is_modified:
3205 *
3206 * Return whether or not the specified physical page was modified
3207 * in any physical maps.
3208 */
3209boolean_t
3210pmap_is_modified(vm_page_t m)
3211{
3212 return pmap_testbit(m, PG_M);
3213}
3214
3215/*
3216 * Clear the modify bits on the specified physical page.
3217 */
3218void
3219pmap_clear_modify(vm_page_t m)
3220{
5e8d0349 3221 pmap_clearbit(m, PG_M);
984263bc
MD
3222}
3223
3224/*
3225 * pmap_clear_reference:
3226 *
3227 * Clear the reference bit on the specified physical page.
3228 */
3229void
3230pmap_clear_reference(vm_page_t m)
3231{
5e8d0349 3232 pmap_clearbit(m, PG_A);
984263bc
MD
3233}
3234
3235/*
3236 * Miscellaneous support routines follow
3237 */
3238
3239static void
840de426 3240i386_protection_init(void)
984263bc 3241{
840de426 3242 int *kp, prot;
984263bc
MD
3243
3244 kp = protection_codes;
3245 for (prot = 0; prot < 8; prot++) {
3246 switch (prot) {
3247 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3248 /*
3249 * Read access is also 0. There isn't any execute bit,
3250 * so just make it readable.
3251 */
3252 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3253 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3254 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3255 *kp++ = 0;
3256 break;
3257 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3258 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3259 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3260 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3261 *kp++ = PG_RW;
3262 break;
3263 }
3264 }
3265}
3266
3267/*
3268 * Map a set of physical memory pages into the kernel virtual
3269 * address space. Return a pointer to where it is mapped. This
3270 * routine is intended to be used for mapping device memory,
3271 * NOT real memory.
a2a5ad0d
MD
3272 *
3273 * NOTE: we can't use pgeflag unless we invalidate the pages one at
3274 * a time.
984263bc
MD
3275 */
3276void *
6ef943a3 3277pmap_mapdev(vm_paddr_t pa, vm_size_t size)
984263bc
MD
3278{
3279 vm_offset_t va, tmpva, offset;
3280 unsigned *pte;
3281
3282 offset = pa & PAGE_MASK;
3283 size = roundup(offset + size, PAGE_SIZE);
3284
e4846942 3285 va = kmem_alloc_nofault(&kernel_map, size);
984263bc
MD
3286 if (!va)
3287 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3288
3289 pa = pa & PG_FRAME;
3290 for (tmpva = va; size > 0;) {
3291 pte = (unsigned *)vtopte(tmpva);
a2a5ad0d 3292 *pte = pa | PG_RW | PG_V; /* | pgeflag; */
984263bc
MD
3293 size -= PAGE_SIZE;
3294 tmpva += PAGE_SIZE;
3295 pa += PAGE_SIZE;
3296 }
0f7a3396
MD
3297 cpu_invltlb();
3298 smp_invltlb();
984263bc
MD
3299
3300 return ((void *)(va + offset));
3301}
3302
3303void
840de426 3304pmap_unmapdev(vm_offset_t va, vm_size_t size)
984263bc
MD
3305{
3306 vm_offset_t base, offset;
3307
3308 base = va & PG_FRAME;
3309 offset = va & PAGE_MASK;
3310 size = roundup(offset + size, PAGE_SIZE);
0f579831 3311 pmap_qremove(va, size >> PAGE_SHIFT);
e4846942 3312 kmem_free(&kernel_map, base, size);
984263bc
MD
3313}
3314
3315/*
3316 * perform the pmap work for mincore
3317 */
3318int
840de426 3319pmap_mincore(pmap_t pmap, vm_offset_t addr)
984263bc 3320{
984263bc
MD
3321 unsigned *ptep, pte;
3322 vm_page_t m;
3323 int val = 0;
3324
3325 ptep = pmap_pte(pmap, addr);
3326 if (ptep == 0) {
3327 return 0;
3328 }
3329
3330 if ((pte = *ptep) != 0) {
3331 vm_offset_t pa;
3332
3333 val = MINCORE_INCORE;
3334 if ((pte & PG_MANAGED) == 0)
3335 return val;
3336
3337 pa = pte & PG_FRAME;
3338
3339 m = PHYS_TO_VM_PAGE(pa);
3340
3341 /*
3342 * Modified by us
3343 */
3344 if (pte & PG_M)
3345 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3346 /*
3347 * Modified by someone
3348 */
3349 else if (m->dirty || pmap_is_modified(m))
3350 val |= MINCORE_MODIFIED_OTHER;
3351 /*
3352 * Referenced by us
3353 */
3354 if (pte & PG_A)
3355 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3356
3357 /*
3358 * Referenced by someone
3359 */
3360 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3361 val |= MINCORE_REFERENCED_OTHER;
3362 vm_page_flag_set(m, PG_REFERENCED);
3363 }
3364 }
3365 return val;
3366}
3367
e3161323
MD
3368/*
3369 * Replace p->p_vmspace with a new one. If adjrefs is non-zero the new
3370 * vmspace will be