kernel - Acquire vm_token in pmap_remove_all()
[dragonfly.git] / sys / platform / pc32 / i386 / pmap.c
CommitLineData
984263bc 1/*
4107b0c0
MD
2 * (MPSAFE)
3 *
984263bc
MD
4 * Copyright (c) 1991 Regents of the University of California.
5 * All rights reserved.
6 * Copyright (c) 1994 John S. Dyson
7 * All rights reserved.
8 * Copyright (c) 1994 David Greenman
9 * All rights reserved.
10 *
11 * This code is derived from software contributed to Berkeley by
12 * the Systems Programming Group of the University of Utah Computer
13 * Science Department and William Jolitz of UUNET Technologies Inc.
14 *
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
17 * are met:
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
30 *
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41 * SUCH DAMAGE.
42 *
43 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
44 * $FreeBSD: src/sys/i386/i386/pmap.c,v 1.250.2.18 2002/03/06 22:48:53 silby Exp $
45 */
46
47/*
4107b0c0 48 * Manages physical address maps.
984263bc 49 *
b12defdc 50 * In most cases we hold page table pages busy in order to manipulate them.
984263bc 51 */
5926987a
MD
52/*
53 * PMAP_DEBUG - see platform/pc32/include/pmap.h
54 */
984263bc
MD
55
56#include "opt_disable_pse.h"
57#include "opt_pmap.h"
58#include "opt_msgbuf.h"
984263bc
MD
59
60#include <sys/param.h>
61#include <sys/systm.h>
62#include <sys/kernel.h>
63#include <sys/proc.h>
64#include <sys/msgbuf.h>
65#include <sys/vmmeter.h>
66#include <sys/mman.h>
b12defdc 67#include <sys/thread.h>
984263bc
MD
68
69#include <vm/vm.h>
70#include <vm/vm_param.h>
71#include <sys/sysctl.h>
72#include <sys/lock.h>
73#include <vm/vm_kern.h>
74#include <vm/vm_page.h>
75#include <vm/vm_map.h>
76#include <vm/vm_object.h>
77#include <vm/vm_extern.h>
78#include <vm/vm_pageout.h>
79#include <vm/vm_pager.h>
80#include <vm/vm_zone.h>
81
82#include <sys/user.h>
e0e69b7d 83#include <sys/thread2.h>
e3161323 84#include <sys/sysref2.h>
b12defdc 85#include <sys/spinlock2.h>
90244566 86#include <vm/vm_page2.h>
984263bc
MD
87
88#include <machine/cputypes.h>
89#include <machine/md_var.h>
90#include <machine/specialreg.h>
984263bc 91#include <machine/smp.h>
a9295349 92#include <machine_base/apic/apicreg.h>
85100692 93#include <machine/globaldata.h>
0f7a3396
MD
94#include <machine/pmap.h>
95#include <machine/pmap_inval.h>
984263bc
MD
96
97#define PMAP_KEEP_PDIRS
98#ifndef PMAP_SHPGPERPROC
99#define PMAP_SHPGPERPROC 200
948209ce 100#define PMAP_PVLIMIT 1400000 /* i386 kvm problems */
984263bc
MD
101#endif
102
103#if defined(DIAGNOSTIC)
104#define PMAP_DIAGNOSTIC
105#endif
106
107#define MINPV 2048
108
109#if !defined(PMAP_DIAGNOSTIC)
110#define PMAP_INLINE __inline
111#else
112#define PMAP_INLINE
113#endif
114
115/*
116 * Get PDEs and PTEs for user/kernel address space
117 */
118#define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
119#define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
120
121#define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
122#define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
123#define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
124#define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
125#define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
126
984263bc
MD
127/*
128 * Given a map and a machine independent protection code,
129 * convert to a vax protection code.
130 */
639a9b43
MD
131#define pte_prot(m, p) \
132 (protection_codes[p & (VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE)])
984263bc
MD
133static int protection_codes[8];
134
fbbaeba3 135struct pmap kernel_pmap;
54a764e8
MD
136static TAILQ_HEAD(,pmap) pmap_list = TAILQ_HEAD_INITIALIZER(pmap_list);
137
e880033d 138vm_paddr_t avail_start; /* PA of first available physical page */
6ef943a3 139vm_paddr_t avail_end; /* PA of last available physical page */
e880033d 140vm_offset_t virtual_start; /* VA of first avail page (after kernel bss) */
984263bc 141vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
791c6551
MD
142vm_offset_t virtual2_start;
143vm_offset_t virtual2_end;
c439ad8f
MD
144vm_offset_t KvaStart; /* VA start of KVA space */
145vm_offset_t KvaEnd; /* VA end of KVA space (non-inclusive) */
146vm_offset_t KvaSize; /* max size of kernel virtual address space */
984263bc
MD
147static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
148static int pgeflag; /* PG_G or-in */
149static int pseflag; /* PG_PS or-in */
150
151static vm_object_t kptobj;
152
153static int nkpt;
154vm_offset_t kernel_vm_end;
155
156/*
157 * Data for the pv entry allocation mechanism
158 */
159static vm_zone_t pvzone;
160static struct vm_zone pvzone_store;
161static struct vm_object pvzone_obj;
162static int pv_entry_count=0, pv_entry_max=0, pv_entry_high_water=0;
163static int pmap_pagedaemon_waken = 0;
164static struct pv_entry *pvinit;
165
166/*
a93980ab
MD
167 * Considering all the issues I'm having with pmap caching, if breakage
168 * continues to occur, and for debugging, I've added a sysctl that will
169 * just do an unconditional invltlb.
170 */
171static int dreadful_invltlb;
172
173SYSCTL_INT(_vm, OID_AUTO, dreadful_invltlb,
9733f757 174 CTLFLAG_RW, &dreadful_invltlb, 0, "Debugging sysctl to force invltlb on pmap operations");
a93980ab
MD
175
176/*
984263bc
MD
177 * All those kernel PT submaps that BSD is so fond of
178 */
4090d6ff 179pt_entry_t *CMAP1 = NULL, *ptmmap;
4c0cc8bb 180caddr_t CADDR1 = NULL, ptvmmap = NULL;
984263bc 181static pt_entry_t *msgbufmap;
4090d6ff 182struct msgbuf *msgbufp=NULL;
984263bc
MD
183
184/*
185 * Crashdump maps.
186 */
187static pt_entry_t *pt_crashdumpmap;
188static caddr_t crashdumpmap;
189
984263bc 190extern pt_entry_t *SMPpt;
984263bc 191
3ae0cd58
RG
192static PMAP_INLINE void free_pv_entry (pv_entry_t pv);
193static unsigned * get_ptbase (pmap_t pmap);
194static pv_entry_t get_pv_entry (void);
195static void i386_protection_init (void);
5e8d0349 196static __inline void pmap_clearbit (vm_page_t m, int bit);
3ae0cd58 197
554cf9ac
MD
198static void pmap_remove_all (vm_page_t m);
199static void pmap_remove_pte (struct pmap *pmap, unsigned *ptq,
0f7a3396
MD
200 vm_offset_t sva, pmap_inval_info_t info);
201static void pmap_remove_page (struct pmap *pmap,
202 vm_offset_t va, pmap_inval_info_t info);
554cf9ac 203static void pmap_remove_entry (struct pmap *pmap, vm_page_t m,
0f7a3396 204 vm_offset_t va, pmap_inval_info_t info);
3ae0cd58 205static boolean_t pmap_testbit (vm_page_t m, int bit);
2bb9cc6f
MD
206static void pmap_insert_entry (pmap_t pmap, pv_entry_t pv,
207 vm_offset_t va, vm_page_t mpte, vm_page_t m);
3ae0cd58
RG
208
209static vm_page_t pmap_allocpte (pmap_t pmap, vm_offset_t va);
210
211static int pmap_release_free_page (pmap_t pmap, vm_page_t p);
212static vm_page_t _pmap_allocpte (pmap_t pmap, unsigned ptepindex);
213static unsigned * pmap_pte_quick (pmap_t pmap, vm_offset_t va);
214static vm_page_t pmap_page_lookup (vm_object_t object, vm_pindex_t pindex);
554cf9ac 215static void pmap_unuse_pt (pmap_t, vm_offset_t, vm_page_t, pmap_inval_info_t);
984263bc
MD
216static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
217
3321ee05
MD
218static void pmap_hold(pmap_t pmap);
219static void pmap_drop(pmap_t pmap);
220static void pmap_wait(pmap_t pmap, int count);
221
984263bc
MD
222static unsigned pdir4mb;
223
224/*
840de426
MD
225 * Move the kernel virtual free pointer to the next
226 * 4MB. This is used to help improve performance
227 * by using a large (4MB) page for much of the kernel
228 * (.text, .data, .bss)
229 */
4107b0c0
MD
230static
231vm_offset_t
840de426
MD
232pmap_kmem_choose(vm_offset_t addr)
233{
234 vm_offset_t newaddr = addr;
235#ifndef DISABLE_PSE
236 if (cpu_feature & CPUID_PSE) {
237 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
238 }
239#endif
240 return newaddr;
241}
242
243/*
4107b0c0
MD
244 * This function returns a pointer to the pte entry in the pmap and has
245 * the side effect of potentially retaining a cached mapping of the pmap.
e0e69b7d 246 *
4107b0c0
MD
247 * The caller must hold vm_token and the returned value is only valid
248 * until the caller blocks or releases the token.
984263bc 249 */
4107b0c0
MD
250static
251unsigned *
840de426 252pmap_pte(pmap_t pmap, vm_offset_t va)
984263bc
MD
253{
254 unsigned *pdeaddr;
255
4107b0c0 256 ASSERT_LWKT_TOKEN_HELD(&vm_token);
984263bc
MD
257 if (pmap) {
258 pdeaddr = (unsigned *) pmap_pde(pmap, va);
259 if (*pdeaddr & PG_PS)
260 return pdeaddr;
4107b0c0 261 if (*pdeaddr)
984263bc 262 return get_ptbase(pmap) + i386_btop(va);
984263bc
MD
263 }
264 return (0);
265}
266
267/*
4107b0c0
MD
268 * pmap_pte using the kernel_pmap
269 *
270 * Used for debugging, no requirements.
271 */
272unsigned *
273pmap_kernel_pte(vm_offset_t va)
274{
275 unsigned *pdeaddr;
276
277 pdeaddr = (unsigned *) pmap_pde(&kernel_pmap, va);
278 if (*pdeaddr & PG_PS)
279 return pdeaddr;
280 if (*pdeaddr)
281 return (unsigned *)vtopte(va);
282 return(0);
283}
284
285/*
e0e69b7d
MD
286 * pmap_pte_quick:
287 *
c1692ddf
MD
288 * Super fast pmap_pte routine best used when scanning the pv lists.
289 * This eliminates many course-grained invltlb calls. Note that many of
290 * the pv list scans are across different pmaps and it is very wasteful
291 * to do an entire invltlb when checking a single mapping.
e0e69b7d 292 *
c1692ddf
MD
293 * Should only be called while in a critical section.
294 *
4107b0c0
MD
295 * The caller must hold vm_token and the returned value is only valid
296 * until the caller blocks or releases the token.
984263bc 297 */
4107b0c0
MD
298static
299unsigned *
840de426 300pmap_pte_quick(pmap_t pmap, vm_offset_t va)
984263bc 301{
840de426
MD
302 struct mdglobaldata *gd = mdcpu;
303 unsigned pde, newpf;
304
4107b0c0 305 ASSERT_LWKT_TOKEN_HELD(&vm_token);
840de426
MD
306 if ((pde = (unsigned) pmap->pm_pdir[va >> PDRSHIFT]) != 0) {
307 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
308 unsigned index = i386_btop(va);
309 /* are we current address space or kernel? */
fbbaeba3 310 if ((pmap == &kernel_pmap) ||
840de426
MD
311 (frame == (((unsigned) PTDpde) & PG_FRAME))) {
312 return (unsigned *) PTmap + index;
313 }
314 newpf = pde & PG_FRAME;
4107b0c0
MD
315 if (((*(unsigned *)gd->gd_PMAP1) & PG_FRAME) != newpf) {
316 *(unsigned *)gd->gd_PMAP1 = newpf | PG_RW | PG_V;
840de426
MD
317 cpu_invlpg(gd->gd_PADDR1);
318 }
06bb314f 319 return gd->gd_PADDR1 + (index & (NPTEPG - 1));
984263bc 320 }
840de426 321 return (0);
984263bc
MD
322}
323
840de426 324
984263bc 325/*
4107b0c0 326 * Bootstrap the system enough to run with virtual memory.
984263bc 327 *
4107b0c0
MD
328 * On the i386 this is called after mapping has already been enabled
329 * and just syncs the pmap module with what has already been done.
330 * [We can't call it easily with mapping off since the kernel is not
331 * mapped with PA == VA, hence we would have to relocate every address
332 * from the linked base (virtual) address "KERNBASE" to the actual
333 * (physical) address starting relative to 0]
984263bc
MD
334 */
335void
f123d5a1 336pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
984263bc
MD
337{
338 vm_offset_t va;
339 pt_entry_t *pte;
85100692 340 struct mdglobaldata *gd;
984263bc 341 int i;
81c04d07 342 int pg;
984263bc 343
c439ad8f
MD
344 KvaStart = (vm_offset_t)VADDR(PTDPTDI, 0);
345 KvaSize = (vm_offset_t)VADDR(APTDPTDI, 0) - KvaStart;
346 KvaEnd = KvaStart + KvaSize;
347
984263bc
MD
348 avail_start = firstaddr;
349
350 /*
e880033d
MD
351 * XXX The calculation of virtual_start is wrong. It's NKPT*PAGE_SIZE
352 * too large. It should instead be correctly calculated in locore.s and
984263bc
MD
353 * not based on 'first' (which is a physical address, not a virtual
354 * address, for the start of unused physical memory). The kernel
355 * page tables are NOT double mapped and thus should not be included
356 * in this calculation.
357 */
e880033d
MD
358 virtual_start = (vm_offset_t) KERNBASE + firstaddr;
359 virtual_start = pmap_kmem_choose(virtual_start);
c439ad8f 360 virtual_end = VADDR(KPTDI+NKPDE-1, NPTEPG-1);
984263bc
MD
361
362 /*
363 * Initialize protection array.
364 */
365 i386_protection_init();
366
367 /*
368 * The kernel's pmap is statically allocated so we don't have to use
369 * pmap_create, which is unlikely to work correctly at this part of
370 * the boot sequence (XXX and which no longer exists).
b12defdc
MD
371 *
372 * The kernel_pmap's pm_pteobj is used only for locking and not
373 * for mmu pages.
984263bc 374 */
fbbaeba3
MD
375 kernel_pmap.pm_pdir = (pd_entry_t *)(KERNBASE + (u_int)IdlePTD);
376 kernel_pmap.pm_count = 1;
c2fb025d 377 kernel_pmap.pm_active = (cpumask_t)-1 & ~CPUMASK_LOCK;
b12defdc 378 kernel_pmap.pm_pteobj = &kernel_object;
fbbaeba3 379 TAILQ_INIT(&kernel_pmap.pm_pvlist);
b12defdc
MD
380 TAILQ_INIT(&kernel_pmap.pm_pvlist_free);
381 spin_init(&kernel_pmap.pm_spin);
382 lwkt_token_init(&kernel_pmap.pm_token, "kpmap_tok");
984263bc
MD
383 nkpt = NKPT;
384
385 /*
386 * Reserve some special page table entries/VA space for temporary
387 * mapping of pages.
388 */
389#define SYSMAP(c, p, v, n) \
390 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
391
e880033d 392 va = virtual_start;
4107b0c0 393 pte = (pt_entry_t *) pmap_kernel_pte(va);
984263bc
MD
394
395 /*
396 * CMAP1/CMAP2 are used for zeroing and copying pages.
397 */
398 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
984263bc
MD
399
400 /*
401 * Crashdump maps.
402 */
403 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
404
405 /*
e731d345
MD
406 * ptvmmap is used for reading arbitrary physical pages via
407 * /dev/mem.
408 */
409 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
410
411 /*
984263bc
MD
412 * msgbufp is used to map the system message buffer.
413 * XXX msgbufmap is not used.
414 */
415 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
416 atop(round_page(MSGBUF_SIZE)))
417
e880033d 418 virtual_start = va;
984263bc 419
17a9f566 420 *(int *) CMAP1 = 0;
984263bc
MD
421 for (i = 0; i < NKPT; i++)
422 PTD[i] = 0;
423
a2a5ad0d
MD
424 /*
425 * PG_G is terribly broken on SMP because we IPI invltlb's in some
426 * cases rather then invl1pg. Actually, I don't even know why it
427 * works under UP because self-referential page table mappings
428 */
429#ifdef SMP
430 pgeflag = 0;
431#else
432 if (cpu_feature & CPUID_PGE)
984263bc 433 pgeflag = PG_G;
a2a5ad0d 434#endif
984263bc
MD
435
436/*
437 * Initialize the 4MB page size flag
438 */
439 pseflag = 0;
440/*
441 * The 4MB page version of the initial
442 * kernel page mapping.
443 */
444 pdir4mb = 0;
445
446#if !defined(DISABLE_PSE)
447 if (cpu_feature & CPUID_PSE) {
448 unsigned ptditmp;
449 /*
450 * Note that we have enabled PSE mode
451 */
452 pseflag = PG_PS;
453 ptditmp = *((unsigned *)PTmap + i386_btop(KERNBASE));
454 ptditmp &= ~(NBPDR - 1);
455 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
456 pdir4mb = ptditmp;
457
8a8d5d85
MD
458#ifndef SMP
459 /*
460 * Enable the PSE mode. If we are SMP we can't do this
461 * now because the APs will not be able to use it when
462 * they boot up.
463 */
464 load_cr4(rcr4() | CR4_PSE);
984263bc 465
8a8d5d85
MD
466 /*
467 * We can do the mapping here for the single processor
468 * case. We simply ignore the old page table page from
469 * now on.
470 */
471 /*
472 * For SMP, we still need 4K pages to bootstrap APs,
473 * PSE will be enabled as soon as all APs are up.
474 */
b5b32410 475 PTD[KPTDI] = (pd_entry_t)ptditmp;
fbbaeba3 476 kernel_pmap.pm_pdir[KPTDI] = (pd_entry_t)ptditmp;
0f7a3396 477 cpu_invltlb();
8a8d5d85 478#endif
984263bc
MD
479 }
480#endif
984263bc 481
81c04d07
MD
482 /*
483 * We need to finish setting up the globaldata page for the BSP.
484 * locore has already populated the page table for the mdglobaldata
485 * portion.
486 */
487 pg = MDGLOBALDATA_BASEALLOC_PAGES;
85100692 488 gd = &CPU_prvspace[0].mdglobaldata;
81c04d07
MD
489 gd->gd_CMAP1 = &SMPpt[pg + 0];
490 gd->gd_CMAP2 = &SMPpt[pg + 1];
491 gd->gd_CMAP3 = &SMPpt[pg + 2];
492 gd->gd_PMAP1 = &SMPpt[pg + 3];
9388fcaa 493 gd->gd_GDMAP1 = &PTD[APTDPTDI];
85100692
MD
494 gd->gd_CADDR1 = CPU_prvspace[0].CPAGE1;
495 gd->gd_CADDR2 = CPU_prvspace[0].CPAGE2;
496 gd->gd_CADDR3 = CPU_prvspace[0].CPAGE3;
497 gd->gd_PADDR1 = (unsigned *)CPU_prvspace[0].PPAGE1;
9388fcaa 498 gd->gd_GDADDR1= (unsigned *)VADDR(APTDPTDI, 0);
984263bc 499
0f7a3396 500 cpu_invltlb();
984263bc
MD
501}
502
503#ifdef SMP
504/*
505 * Set 4mb pdir for mp startup
506 */
507void
508pmap_set_opt(void)
509{
510 if (pseflag && (cpu_feature & CPUID_PSE)) {
511 load_cr4(rcr4() | CR4_PSE);
72740893 512 if (pdir4mb && mycpu->gd_cpuid == 0) { /* only on BSP */
fbbaeba3 513 kernel_pmap.pm_pdir[KPTDI] =
984263bc
MD
514 PTD[KPTDI] = (pd_entry_t)pdir4mb;
515 cpu_invltlb();
516 }
517 }
518}
519#endif
520
521/*
4107b0c0
MD
522 * Initialize the pmap module, called by vm_init()
523 *
524 * Called from the low level boot code only.
984263bc
MD
525 */
526void
e7252eda 527pmap_init(void)
984263bc
MD
528{
529 int i;
530 int initial_pvs;
531
532 /*
533 * object for kernel page table pages
534 */
535 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
536
537 /*
538 * Allocate memory for random pmap data structures. Includes the
539 * pv_head_table.
540 */
541
542 for(i = 0; i < vm_page_array_size; i++) {
543 vm_page_t m;
544
545 m = &vm_page_array[i];
546 TAILQ_INIT(&m->md.pv_list);
547 m->md.pv_list_count = 0;
548 }
549
550 /*
551 * init the pv free list
552 */
553 initial_pvs = vm_page_array_size;
554 if (initial_pvs < MINPV)
555 initial_pvs = MINPV;
556 pvzone = &pvzone_store;
948209ce
MD
557 pvinit = (void *)kmem_alloc(&kernel_map,
558 initial_pvs * sizeof (struct pv_entry));
559 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry),
560 pvinit, initial_pvs);
984263bc
MD
561
562 /*
563 * Now it is safe to enable pv_table recording.
564 */
565 pmap_initialized = TRUE;
566}
567
568/*
569 * Initialize the address space (zone) for the pv_entries. Set a
570 * high water mark so that the system can recover from excessive
571 * numbers of pv entries.
4107b0c0
MD
572 *
573 * Called from the low level boot code only.
984263bc
MD
574 */
575void
f123d5a1 576pmap_init2(void)
984263bc
MD
577{
578 int shpgperproc = PMAP_SHPGPERPROC;
948209ce 579 int entry_max;
984263bc
MD
580
581 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
582 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
948209ce
MD
583
584#ifdef PMAP_PVLIMIT
585 /*
586 * Horrible hack for systems with a lot of memory running i386.
587 * the calculated pv_entry_max can wind up eating a ton of KVM
588 * so put a cap on the number of entries if the user did not
589 * change any of the values. This saves about 44MB of KVM on
590 * boxes with 3+GB of ram.
591 *
592 * On the flip side, this makes it more likely that some setups
593 * will run out of pv entries. Those sysads will have to bump
594 * the limit up with vm.pamp.pv_entries or vm.pmap.shpgperproc.
595 */
596 if (shpgperproc == PMAP_SHPGPERPROC) {
597 if (pv_entry_max > PMAP_PVLIMIT)
598 pv_entry_max = PMAP_PVLIMIT;
599 }
600#endif
984263bc
MD
601 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
602 pv_entry_high_water = 9 * (pv_entry_max / 10);
948209ce
MD
603
604 /*
605 * Subtract out pages already installed in the zone (hack)
606 */
607 entry_max = pv_entry_max - vm_page_array_size;
608 if (entry_max <= 0)
609 entry_max = 1;
610
611 zinitna(pvzone, &pvzone_obj, NULL, 0, entry_max, ZONE_INTERRUPT, 1);
984263bc
MD
612}
613
614
615/***************************************************
616 * Low level helper routines.....
617 ***************************************************/
618
5926987a
MD
619#ifdef PMAP_DEBUG
620
621static void
622test_m_maps_pv(vm_page_t m, pv_entry_t pv)
623{
624 pv_entry_t spv;
625
74b9d1ec 626 crit_enter();
5926987a
MD
627#ifdef PMAP_DEBUG
628 KKASSERT(pv->pv_m == m);
629#endif
630 TAILQ_FOREACH(spv, &m->md.pv_list, pv_list) {
74b9d1ec
MD
631 if (pv == spv) {
632 crit_exit();
5926987a 633 return;
74b9d1ec 634 }
5926987a 635 }
74b9d1ec 636 crit_exit();
ed20d0e3 637 panic("test_m_maps_pv: failed m %p pv %p", m, pv);
5926987a
MD
638}
639
640static void
641ptbase_assert(struct pmap *pmap)
642{
643 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
644
645 /* are we current address space or kernel? */
4107b0c0 646 if (pmap == &kernel_pmap || frame == (((unsigned)PTDpde) & PG_FRAME))
5926987a 647 return;
3558dcda 648 KKASSERT(frame == (*mdcpu->gd_GDMAP1 & PG_FRAME));
5926987a
MD
649}
650
651#else
652
653#define test_m_maps_pv(m, pv)
654#define ptbase_assert(pmap)
655
656#endif
657
984263bc
MD
658#if defined(PMAP_DIAGNOSTIC)
659
660/*
661 * This code checks for non-writeable/modified pages.
662 * This should be an invalid condition.
663 */
664static int
665pmap_nw_modified(pt_entry_t ptea)
666{
667 int pte;
668
669 pte = (int) ptea;
670
671 if ((pte & (PG_M|PG_RW)) == PG_M)
672 return 1;
673 else
674 return 0;
675}
676#endif
677
678
679/*
4107b0c0
MD
680 * This routine defines the region(s) of memory that should not be tested
681 * for the modified bit.
682 *
683 * No requirements.
984263bc
MD
684 */
685static PMAP_INLINE int
686pmap_track_modified(vm_offset_t va)
687{
688 if ((va < clean_sva) || (va >= clean_eva))
689 return 1;
690 else
691 return 0;
692}
693
c1692ddf
MD
694/*
695 * Retrieve the mapped page table base for a particular pmap. Use our self
696 * mapping for the kernel_pmap or our current pmap.
697 *
698 * For foreign pmaps we use the per-cpu page table map. Since this involves
699 * installing a ptd it's actually (per-process x per-cpu). However, we
700 * still cannot depend on our mapping to survive thread switches because
701 * the process might be threaded and switching to another thread for the
702 * same process on the same cpu will allow that other thread to make its
703 * own mapping.
704 *
705 * This could be a bit confusing but the jist is for something like the
706 * vkernel which uses foreign pmaps all the time this represents a pretty
707 * good cache that avoids unnecessary invltlb()s.
4107b0c0
MD
708 *
709 * The caller must hold vm_token and the returned value is only valid
710 * until the caller blocks or releases the token.
c1692ddf 711 */
984263bc 712static unsigned *
e0e69b7d 713get_ptbase(pmap_t pmap)
984263bc
MD
714{
715 unsigned frame = (unsigned) pmap->pm_pdir[PTDPTDI] & PG_FRAME;
c1692ddf 716 struct mdglobaldata *gd = mdcpu;
984263bc 717
4107b0c0
MD
718 ASSERT_LWKT_TOKEN_HELD(&vm_token);
719
5926987a
MD
720 /*
721 * We can use PTmap if the pmap is our current address space or
722 * the kernel address space.
723 */
fbbaeba3 724 if (pmap == &kernel_pmap || frame == (((unsigned) PTDpde) & PG_FRAME)) {
984263bc
MD
725 return (unsigned *) PTmap;
726 }
e0e69b7d 727
5926987a 728 /*
c1692ddf
MD
729 * Otherwise we use the per-cpu alternative page table map. Each
730 * cpu gets its own map. Because of this we cannot use this map
731 * from interrupts or threads which can preempt.
be3aecf7
MD
732 *
733 * Even if we already have the map cached we may still have to
734 * invalidate the TLB if another cpu modified a PDE in the map.
5926987a 735 */
c1692ddf
MD
736 KKASSERT(gd->mi.gd_intr_nesting_level == 0 &&
737 (gd->mi.gd_curthread->td_flags & TDF_INTTHREAD) == 0);
e0e69b7d 738
c1692ddf
MD
739 if ((*gd->gd_GDMAP1 & PG_FRAME) != frame) {
740 *gd->gd_GDMAP1 = frame | PG_RW | PG_V;
be3aecf7
MD
741 pmap->pm_cached |= gd->mi.gd_cpumask;
742 cpu_invltlb();
743 } else if ((pmap->pm_cached & gd->mi.gd_cpumask) == 0) {
744 pmap->pm_cached |= gd->mi.gd_cpumask;
984263bc 745 cpu_invltlb();
a93980ab
MD
746 } else if (dreadful_invltlb) {
747 cpu_invltlb();
984263bc 748 }
c1692ddf 749 return ((unsigned *)gd->gd_GDADDR1);
984263bc
MD
750}
751
752/*
e0e69b7d
MD
753 * pmap_extract:
754 *
4107b0c0 755 * Extract the physical page address associated with the map/VA pair.
e0e69b7d 756 *
4107b0c0 757 * The caller may hold vm_token if it desires non-blocking operation.
984263bc 758 */
6ef943a3 759vm_paddr_t
840de426 760pmap_extract(pmap_t pmap, vm_offset_t va)
984263bc
MD
761{
762 vm_offset_t rtval;
763 vm_offset_t pdirindex;
840de426 764
4107b0c0 765 lwkt_gettoken(&vm_token);
984263bc
MD
766 pdirindex = va >> PDRSHIFT;
767 if (pmap && (rtval = (unsigned) pmap->pm_pdir[pdirindex])) {
768 unsigned *pte;
769 if ((rtval & PG_PS) != 0) {
770 rtval &= ~(NBPDR - 1);
771 rtval |= va & (NBPDR - 1);
4107b0c0
MD
772 } else {
773 pte = get_ptbase(pmap) + i386_btop(va);
774 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
984263bc 775 }
4107b0c0
MD
776 } else {
777 rtval = 0;
984263bc 778 }
4107b0c0
MD
779 lwkt_reltoken(&vm_token);
780 return rtval;
f6bf3af1
MD
781}
782
984263bc
MD
783/***************************************************
784 * Low level mapping routines.....
785 ***************************************************/
786
787/*
4107b0c0
MD
788 * Map a wired VM page to a KVA, fully SMP synchronized.
789 *
790 * No requirements, non blocking.
984263bc 791 */
24712b90 792void
6ef943a3 793pmap_kenter(vm_offset_t va, vm_paddr_t pa)
984263bc 794{
840de426 795 unsigned *pte;
0f7a3396
MD
796 unsigned npte;
797 pmap_inval_info info;
984263bc 798
0f7a3396 799 pmap_inval_init(&info);
984263bc
MD
800 npte = pa | PG_RW | PG_V | pgeflag;
801 pte = (unsigned *)vtopte(va);
c2fb025d 802 pmap_inval_interlock(&info, &kernel_pmap, va);
984263bc 803 *pte = npte;
c2fb025d
MD
804 pmap_inval_deinterlock(&info, &kernel_pmap);
805 pmap_inval_done(&info);
984263bc
MD
806}
807
6d1ec6fa 808/*
4107b0c0
MD
809 * Map a wired VM page to a KVA, synchronized on current cpu only.
810 *
811 * No requirements, non blocking.
6d1ec6fa 812 */
24712b90
MD
813void
814pmap_kenter_quick(vm_offset_t va, vm_paddr_t pa)
815{
816 unsigned *pte;
817 unsigned npte;
818
819 npte = pa | PG_RW | PG_V | pgeflag;
820 pte = (unsigned *)vtopte(va);
821 *pte = npte;
822 cpu_invlpg((void *)va);
823}
824
4107b0c0
MD
825/*
826 * Synchronize a previously entered VA on all cpus.
827 *
828 * No requirements, non blocking.
829 */
24712b90
MD
830void
831pmap_kenter_sync(vm_offset_t va)
832{
833 pmap_inval_info info;
834
835 pmap_inval_init(&info);
c2fb025d
MD
836 pmap_inval_interlock(&info, &kernel_pmap, va);
837 pmap_inval_deinterlock(&info, &kernel_pmap);
838 pmap_inval_done(&info);
24712b90
MD
839}
840
4107b0c0
MD
841/*
842 * Synchronize a previously entered VA on the current cpu only.
843 *
844 * No requirements, non blocking.
845 */
24712b90
MD
846void
847pmap_kenter_sync_quick(vm_offset_t va)
848{
849 cpu_invlpg((void *)va);
850}
851
984263bc 852/*
4107b0c0
MD
853 * Remove a page from the kernel pagetables, fully SMP synchronized.
854 *
855 * No requirements, non blocking.
984263bc 856 */
24712b90 857void
840de426 858pmap_kremove(vm_offset_t va)
984263bc 859{
840de426 860 unsigned *pte;
0f7a3396 861 pmap_inval_info info;
984263bc 862
0f7a3396 863 pmap_inval_init(&info);
984263bc 864 pte = (unsigned *)vtopte(va);
c2fb025d 865 pmap_inval_interlock(&info, &kernel_pmap, va);
984263bc 866 *pte = 0;
c2fb025d
MD
867 pmap_inval_deinterlock(&info, &kernel_pmap);
868 pmap_inval_done(&info);
984263bc
MD
869}
870
4107b0c0
MD
871/*
872 * Remove a page from the kernel pagetables, synchronized on current cpu only.
873 *
874 * No requirements, non blocking.
875 */
24712b90
MD
876void
877pmap_kremove_quick(vm_offset_t va)
878{
879 unsigned *pte;
880 pte = (unsigned *)vtopte(va);
881 *pte = 0;
882 cpu_invlpg((void *)va);
883}
884
984263bc 885/*
4107b0c0
MD
886 * Adjust the permissions of a page in the kernel page table,
887 * synchronized on the current cpu only.
888 *
889 * No requirements, non blocking.
9ad680a3
MD
890 */
891void
892pmap_kmodify_rw(vm_offset_t va)
893{
4107b0c0 894 atomic_set_int(vtopte(va), PG_RW);
9ad680a3
MD
895 cpu_invlpg((void *)va);
896}
897
4107b0c0
MD
898/*
899 * Adjust the permissions of a page in the kernel page table,
900 * synchronized on the current cpu only.
901 *
902 * No requirements, non blocking.
903 */
9ad680a3
MD
904void
905pmap_kmodify_nc(vm_offset_t va)
906{
4107b0c0 907 atomic_set_int(vtopte(va), PG_N);
9ad680a3
MD
908 cpu_invlpg((void *)va);
909}
910
911/*
4107b0c0 912 * Map a range of physical addresses into kernel virtual address space.
984263bc 913 *
4107b0c0 914 * No requirements, non blocking.
984263bc
MD
915 */
916vm_offset_t
8e5e6f1b 917pmap_map(vm_offset_t *virtp, vm_paddr_t start, vm_paddr_t end, int prot)
984263bc 918{
8e5e6f1b
AH
919 vm_offset_t sva, virt;
920
921 sva = virt = *virtp;
984263bc
MD
922 while (start < end) {
923 pmap_kenter(virt, start);
924 virt += PAGE_SIZE;
925 start += PAGE_SIZE;
926 }
8e5e6f1b
AH
927 *virtp = virt;
928 return (sva);
984263bc
MD
929}
930
984263bc 931/*
4107b0c0
MD
932 * Add a list of wired pages to the kva, fully SMP synchronized.
933 *
934 * No requirements, non blocking.
984263bc
MD
935 */
936void
840de426 937pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
984263bc
MD
938{
939 vm_offset_t end_va;
940
941 end_va = va + count * PAGE_SIZE;
942
943 while (va < end_va) {
944 unsigned *pte;
945
946 pte = (unsigned *)vtopte(va);
947 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
984263bc 948 cpu_invlpg((void *)va);
984263bc
MD
949 va += PAGE_SIZE;
950 m++;
951 }
952#ifdef SMP
0f7a3396 953 smp_invltlb(); /* XXX */
984263bc
MD
954#endif
955}
956
957/*
4107b0c0 958 * Remove pages from KVA, fully SMP synchronized.
7155fc7d 959 *
4107b0c0 960 * No requirements, non blocking.
984263bc
MD
961 */
962void
840de426 963pmap_qremove(vm_offset_t va, int count)
984263bc
MD
964{
965 vm_offset_t end_va;
966
967 end_va = va + count*PAGE_SIZE;
968
969 while (va < end_va) {
970 unsigned *pte;
971
972 pte = (unsigned *)vtopte(va);
973 *pte = 0;
984263bc 974 cpu_invlpg((void *)va);
984263bc
MD
975 va += PAGE_SIZE;
976 }
977#ifdef SMP
978 smp_invltlb();
979#endif
980}
981
06ecca5a
MD
982/*
983 * This routine works like vm_page_lookup() but also blocks as long as the
984 * page is busy. This routine does not busy the page it returns.
985 *
b12defdc 986 * The caller must hold the object.
06ecca5a 987 */
984263bc 988static vm_page_t
840de426 989pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
984263bc
MD
990{
991 vm_page_t m;
06ecca5a 992
b12defdc
MD
993 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
994 m = vm_page_lookup_busy_wait(object, pindex, FALSE, "pplookp");
17cde63e 995
06ecca5a 996 return(m);
984263bc
MD
997}
998
999/*
263e4574 1000 * Create a new thread and optionally associate it with a (new) process.
6ef943a3 1001 * NOTE! the new thread's cpu may not equal the current cpu.
263e4574 1002 */
7d0bac62
MD
1003void
1004pmap_init_thread(thread_t td)
263e4574 1005{
f470d0c8 1006 /* enforce pcb placement */
f470d0c8 1007 td->td_pcb = (struct pcb *)(td->td_kstack + td->td_kstack_size) - 1;
65d6ce10 1008 td->td_savefpu = &td->td_pcb->pcb_save;
7d0bac62 1009 td->td_sp = (char *)td->td_pcb - 16;
263e4574
MD
1010}
1011
1012/*
984263bc
MD
1013 * This routine directly affects the fork perf for a process.
1014 */
1015void
13d13d89 1016pmap_init_proc(struct proc *p)
984263bc 1017{
984263bc
MD
1018}
1019
984263bc
MD
1020/***************************************************
1021 * Page table page management routines.....
1022 ***************************************************/
1023
1024/*
90244566
MD
1025 * This routine unwires page table pages, removing and freeing the page
1026 * tale page when the wire count drops to 0.
4107b0c0
MD
1027 *
1028 * The caller must hold vm_token.
1029 * This function can block.
984263bc
MD
1030 */
1031static int
90244566 1032_pmap_unwire_pte(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
840de426 1033{
17cde63e
MD
1034 /*
1035 * Wait until we can busy the page ourselves. We cannot have
1036 * any active flushes if we block.
1037 */
b12defdc 1038 vm_page_busy_wait(m, FALSE, "pmuwpt");
eec2b734 1039 KASSERT(m->queue == PQ_NONE,
90244566 1040 ("_pmap_unwire_pte: %p->queue != PQ_NONE", m));
984263bc 1041
90244566 1042 if (m->wire_count == 1) {
984263bc 1043 /*
be3aecf7
MD
1044 * Unmap the page table page.
1045 *
1046 * NOTE: We must clear pm_cached for all cpus, including
1047 * the current one, when clearing a page directory
1048 * entry.
984263bc 1049 */
c2fb025d 1050 pmap_inval_interlock(info, pmap, -1);
2247fe02 1051 KKASSERT(pmap->pm_pdir[m->pindex]);
984263bc 1052 pmap->pm_pdir[m->pindex] = 0;
be3aecf7 1053 pmap->pm_cached = 0;
c2fb025d 1054 pmap_inval_deinterlock(info, pmap);
eec2b734
MD
1055
1056 KKASSERT(pmap->pm_stats.resident_count > 0);
984263bc 1057 --pmap->pm_stats.resident_count;
984263bc
MD
1058
1059 if (pmap->pm_ptphint == m)
1060 pmap->pm_ptphint = NULL;
1061
1062 /*
eec2b734
MD
1063 * This was our last hold, the page had better be unwired
1064 * after we decrement wire_count.
1065 *
1066 * FUTURE NOTE: shared page directory page could result in
1067 * multiple wire counts.
984263bc 1068 */
90244566 1069 vm_page_unwire(m, 0);
17cde63e 1070 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
eec2b734
MD
1071 vm_page_flash(m);
1072 vm_page_free_zero(m);
984263bc 1073 return 1;
17cde63e 1074 } else {
90244566
MD
1075 KKASSERT(m->wire_count > 1);
1076 if (vm_page_unwire_quick(m))
1077 panic("pmap_unwire_pte: Insufficient wire_count");
b12defdc 1078 vm_page_wakeup(m);
17cde63e 1079 return 0;
984263bc 1080 }
984263bc
MD
1081}
1082
4107b0c0
MD
1083/*
1084 * The caller must hold vm_token.
92ba8d28 1085 *
4107b0c0 1086 * This function can block.
92ba8d28
MD
1087 *
1088 * This function can race the wire_count 2->1 case because the page
1089 * is not busied during the unwire_quick operation. An eventual
1090 * pmap_release() will catch the case.
4107b0c0 1091 */
984263bc 1092static PMAP_INLINE int
90244566 1093pmap_unwire_pte(pmap_t pmap, vm_page_t m, pmap_inval_info_t info)
984263bc 1094{
90244566
MD
1095 KKASSERT(m->wire_count > 0);
1096 if (m->wire_count > 1) {
1097 if (vm_page_unwire_quick(m))
1098 panic("pmap_unwire_pte: Insufficient wire_count");
984263bc 1099 return 0;
eec2b734 1100 } else {
90244566 1101 return _pmap_unwire_pte(pmap, m, info);
eec2b734 1102 }
984263bc
MD
1103}
1104
1105/*
4107b0c0 1106 * After removing a (user) page table entry, this routine is used to
984263bc 1107 * conditionally free the page, and manage the hold/wire counts.
5926987a 1108 *
4107b0c0
MD
1109 * The caller must hold vm_token.
1110 * This function can block regardless.
984263bc 1111 */
554cf9ac 1112static void
0f7a3396 1113pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte,
4107b0c0 1114 pmap_inval_info_t info)
984263bc
MD
1115{
1116 unsigned ptepindex;
4107b0c0 1117
b12defdc
MD
1118 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1119
984263bc 1120 if (va >= UPT_MIN_ADDRESS)
554cf9ac 1121 return;
984263bc
MD
1122
1123 if (mpte == NULL) {
1124 ptepindex = (va >> PDRSHIFT);
b1482674
MD
1125 if ((mpte = pmap->pm_ptphint) != NULL &&
1126 mpte->pindex == ptepindex &&
1127 (mpte->flags & PG_BUSY) == 0) {
1128 ; /* use mpte */
984263bc 1129 } else {
b12defdc 1130 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
984263bc 1131 pmap->pm_ptphint = mpte;
b12defdc 1132 vm_page_wakeup(mpte);
984263bc
MD
1133 }
1134 }
554cf9ac 1135 pmap_unwire_pte(pmap, mpte, info);
984263bc
MD
1136}
1137
54a764e8 1138/*
fbbaeba3
MD
1139 * Initialize pmap0/vmspace0. This pmap is not added to pmap_list because
1140 * it, and IdlePTD, represents the template used to update all other pmaps.
1141 *
1142 * On architectures where the kernel pmap is not integrated into the user
1143 * process pmap, this pmap represents the process pmap, not the kernel pmap.
1144 * kernel_pmap should be used to directly access the kernel_pmap.
4107b0c0
MD
1145 *
1146 * No requirements.
54a764e8 1147 */
984263bc 1148void
840de426 1149pmap_pinit0(struct pmap *pmap)
984263bc
MD
1150{
1151 pmap->pm_pdir =
e4846942 1152 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
24712b90 1153 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t) IdlePTD);
984263bc
MD
1154 pmap->pm_count = 1;
1155 pmap->pm_active = 0;
be3aecf7 1156 pmap->pm_cached = 0;
984263bc
MD
1157 pmap->pm_ptphint = NULL;
1158 TAILQ_INIT(&pmap->pm_pvlist);
b12defdc
MD
1159 TAILQ_INIT(&pmap->pm_pvlist_free);
1160 spin_init(&pmap->pm_spin);
1161 lwkt_token_init(&pmap->pm_token, "pmap_tok");
984263bc
MD
1162 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1163}
1164
1165/*
1166 * Initialize a preallocated and zeroed pmap structure,
1167 * such as one in a vmspace structure.
4107b0c0
MD
1168 *
1169 * No requirements.
984263bc
MD
1170 */
1171void
840de426 1172pmap_pinit(struct pmap *pmap)
984263bc
MD
1173{
1174 vm_page_t ptdpg;
1175
1176 /*
1177 * No need to allocate page table space yet but we do need a valid
1178 * page directory table.
1179 */
b5b32410 1180 if (pmap->pm_pdir == NULL) {
984263bc 1181 pmap->pm_pdir =
e4846942 1182 (pd_entry_t *)kmem_alloc_pageable(&kernel_map, PAGE_SIZE);
b5b32410 1183 }
984263bc
MD
1184
1185 /*
c3834cb2 1186 * Allocate an object for the ptes
984263bc
MD
1187 */
1188 if (pmap->pm_pteobj == NULL)
c3834cb2 1189 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
984263bc
MD
1190
1191 /*
c3834cb2
MD
1192 * Allocate the page directory page, unless we already have
1193 * one cached. If we used the cached page the wire_count will
1194 * already be set appropriately.
984263bc 1195 */
c3834cb2
MD
1196 if ((ptdpg = pmap->pm_pdirm) == NULL) {
1197 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
d2d8515b
MD
1198 VM_ALLOC_NORMAL | VM_ALLOC_RETRY |
1199 VM_ALLOC_ZERO);
c3834cb2 1200 pmap->pm_pdirm = ptdpg;
b12defdc
MD
1201 vm_page_flag_clear(ptdpg, PG_MAPPED);
1202 vm_page_wire(ptdpg);
d2d8515b 1203 KKASSERT(ptdpg->valid == VM_PAGE_BITS_ALL);
c3834cb2 1204 pmap_kenter((vm_offset_t)pmap->pm_pdir, VM_PAGE_TO_PHYS(ptdpg));
b12defdc 1205 vm_page_wakeup(ptdpg);
c3834cb2 1206 }
984263bc 1207 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
984263bc
MD
1208
1209 /* install self-referential address mapping entry */
1210 *(unsigned *) (pmap->pm_pdir + PTDPTDI) =
1211 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1212
1213 pmap->pm_count = 1;
1214 pmap->pm_active = 0;
be3aecf7 1215 pmap->pm_cached = 0;
984263bc
MD
1216 pmap->pm_ptphint = NULL;
1217 TAILQ_INIT(&pmap->pm_pvlist);
b12defdc
MD
1218 TAILQ_INIT(&pmap->pm_pvlist_free);
1219 spin_init(&pmap->pm_spin);
1220 lwkt_token_init(&pmap->pm_token, "pmap_tok");
984263bc 1221 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
eec2b734 1222 pmap->pm_stats.resident_count = 1;
984263bc
MD
1223}
1224
1225/*
c3834cb2
MD
1226 * Clean up a pmap structure so it can be physically freed. This routine
1227 * is called by the vmspace dtor function. A great deal of pmap data is
1228 * left passively mapped to improve vmspace management so we have a bit
1229 * of cleanup work to do here.
4107b0c0
MD
1230 *
1231 * No requirements.
e3161323
MD
1232 */
1233void
1234pmap_puninit(pmap_t pmap)
1235{
c3834cb2
MD
1236 vm_page_t p;
1237
3321ee05 1238 pmap_wait(pmap, -1);
e3161323 1239 KKASSERT(pmap->pm_active == 0);
c3834cb2
MD
1240 if ((p = pmap->pm_pdirm) != NULL) {
1241 KKASSERT(pmap->pm_pdir != NULL);
1242 pmap_kremove((vm_offset_t)pmap->pm_pdir);
b12defdc 1243 vm_page_busy_wait(p, FALSE, "pgpun");
90244566 1244 vm_page_unwire(p, 0);
c3834cb2
MD
1245 vm_page_free_zero(p);
1246 pmap->pm_pdirm = NULL;
1247 }
e3161323
MD
1248 if (pmap->pm_pdir) {
1249 kmem_free(&kernel_map, (vm_offset_t)pmap->pm_pdir, PAGE_SIZE);
1250 pmap->pm_pdir = NULL;
1251 }
1252 if (pmap->pm_pteobj) {
1253 vm_object_deallocate(pmap->pm_pteobj);
1254 pmap->pm_pteobj = NULL;
1255 }
1256}
1257
1258/*
984263bc
MD
1259 * Wire in kernel global address entries. To avoid a race condition
1260 * between pmap initialization and pmap_growkernel, this procedure
54a764e8
MD
1261 * adds the pmap to the master list (which growkernel scans to update),
1262 * then copies the template.
4107b0c0
MD
1263 *
1264 * No requirements.
984263bc
MD
1265 */
1266void
840de426 1267pmap_pinit2(struct pmap *pmap)
984263bc 1268{
b12defdc
MD
1269 /*
1270 * XXX copies current process, does not fill in MPPTDI
1271 */
1272 spin_lock(&pmap_spin);
54a764e8 1273 TAILQ_INSERT_TAIL(&pmap_list, pmap, pm_pmnode);
984263bc 1274 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
b12defdc 1275 spin_unlock(&pmap_spin);
984263bc
MD
1276}
1277
344ad853 1278/*
eec2b734 1279 * Attempt to release and free a vm_page in a pmap. Returns 1 on success,
344ad853 1280 * 0 on failure (if the procedure had to sleep).
c3834cb2
MD
1281 *
1282 * When asked to remove the page directory page itself, we actually just
1283 * leave it cached so we do not have to incur the SMP inval overhead of
1284 * removing the kernel mapping. pmap_puninit() will take care of it.
4107b0c0
MD
1285 *
1286 * The caller must hold vm_token.
1287 * This function can block regardless.
344ad853 1288 */
984263bc 1289static int
840de426 1290pmap_release_free_page(struct pmap *pmap, vm_page_t p)
984263bc
MD
1291{
1292 unsigned *pde = (unsigned *) pmap->pm_pdir;
4107b0c0 1293
984263bc
MD
1294 /*
1295 * This code optimizes the case of freeing non-busy
1296 * page-table pages. Those pages are zero now, and
1297 * might as well be placed directly into the zero queue.
1298 */
b12defdc
MD
1299 if (vm_page_busy_try(p, FALSE)) {
1300 vm_page_sleep_busy(p, FALSE, "pmaprl");
984263bc 1301 return 0;
b12defdc 1302 }
984263bc 1303
eec2b734 1304 KKASSERT(pmap->pm_stats.resident_count > 0);
2247fe02 1305 KKASSERT(pde[p->pindex]);
984263bc 1306
b1482674
MD
1307 /*
1308 * page table page's wire_count must be 1. Caller is the pmap
1309 * termination code which holds the pm_pteobj, there is a race
1310 * if someone else is trying to hold the VM object in order to
1311 * clean up a wire_count.
1312 */
90244566 1313 if (p->wire_count != 1) {
b1482674
MD
1314 if (pmap->pm_pteobj->hold_count <= 1)
1315 panic("pmap_release: freeing wired page table page");
1316 kprintf("pmap_release_free_page: unwire race detected\n");
1317 vm_page_wakeup(p);
1318 tsleep(p, 0, "pmapx", 1);
1319 return 0;
984263bc 1320 }
b1482674
MD
1321
1322 /*
1323 * Remove the page table page from the processes address space.
1324 */
1325 pmap->pm_cached = 0;
1326 pde[p->pindex] = 0;
1327 --pmap->pm_stats.resident_count;
c3834cb2
MD
1328 if (pmap->pm_ptphint && (pmap->pm_ptphint->pindex == p->pindex))
1329 pmap->pm_ptphint = NULL;
1330
984263bc 1331 /*
c3834cb2
MD
1332 * We leave the page directory page cached, wired, and mapped in
1333 * the pmap until the dtor function (pmap_puninit()) gets called.
1334 * However, still clean it up so we can set PG_ZERO.
c1692ddf
MD
1335 *
1336 * The pmap has already been removed from the pmap_list in the
1337 * PTDPTDI case.
984263bc
MD
1338 */
1339 if (p->pindex == PTDPTDI) {
1340 bzero(pde + KPTDI, nkpt * PTESIZE);
9388fcaa 1341 bzero(pde + MPPTDI, (NPDEPG - MPPTDI) * PTESIZE);
c3834cb2
MD
1342 vm_page_flag_set(p, PG_ZERO);
1343 vm_page_wakeup(p);
1344 } else {
92ba8d28
MD
1345 /*
1346 * This case can occur if a pmap_unwire_pte() loses a race
1347 * while the page is unbusied.
1348 */
1349 /*panic("pmap_release: page should already be gone %p", p);*/
1350 vm_page_flag_clear(p, PG_MAPPED);
90244566 1351 vm_page_unwire(p, 0);
c3834cb2 1352 vm_page_free_zero(p);
984263bc 1353 }
984263bc
MD
1354 return 1;
1355}
1356
1357/*
4107b0c0
MD
1358 * This routine is called if the page table page is not mapped correctly.
1359 *
1360 * The caller must hold vm_token.
984263bc
MD
1361 */
1362static vm_page_t
840de426 1363_pmap_allocpte(pmap_t pmap, unsigned ptepindex)
984263bc 1364{
480c83b6 1365 vm_offset_t ptepa;
984263bc
MD
1366 vm_page_t m;
1367
1368 /*
d2d8515b
MD
1369 * Find or fabricate a new pagetable page. Setting VM_ALLOC_ZERO
1370 * will zero any new page and mark it valid.
984263bc
MD
1371 */
1372 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
d2d8515b 1373 VM_ALLOC_NORMAL | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
984263bc
MD
1374
1375 KASSERT(m->queue == PQ_NONE,
1376 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1377
eec2b734 1378 /*
90244566 1379 * Increment the wire count for the page we will be returning to
eec2b734
MD
1380 * the caller.
1381 */
90244566 1382 vm_page_wire(m);
eec2b734
MD
1383
1384 /*
1385 * It is possible that someone else got in and mapped by the page
1386 * directory page while we were blocked, if so just unbusy and
90244566 1387 * return the wired page.
eec2b734
MD
1388 */
1389 if ((ptepa = pmap->pm_pdir[ptepindex]) != 0) {
1390 KKASSERT((ptepa & PG_FRAME) == VM_PAGE_TO_PHYS(m));
1391 vm_page_wakeup(m);
1392 return(m);
1393 }
1394
984263bc
MD
1395 /*
1396 * Map the pagetable page into the process address space, if
1397 * it isn't already there.
be3aecf7
MD
1398 *
1399 * NOTE: For safety clear pm_cached for all cpus including the
1400 * current one when adding a PDE to the map.
984263bc 1401 */
eec2b734 1402 ++pmap->pm_stats.resident_count;
984263bc
MD
1403
1404 ptepa = VM_PAGE_TO_PHYS(m);
1405 pmap->pm_pdir[ptepindex] =
1406 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
be3aecf7 1407 pmap->pm_cached = 0;
984263bc
MD
1408
1409 /*
1410 * Set the page table hint
1411 */
1412 pmap->pm_ptphint = m;
984263bc
MD
1413 vm_page_flag_set(m, PG_MAPPED);
1414 vm_page_wakeup(m);
1415
1416 return m;
1417}
1418
4107b0c0
MD
1419/*
1420 * Allocate a page table entry for a va.
1421 *
1422 * The caller must hold vm_token.
1423 */
984263bc 1424static vm_page_t
840de426 1425pmap_allocpte(pmap_t pmap, vm_offset_t va)
984263bc
MD
1426{
1427 unsigned ptepindex;
1428 vm_offset_t ptepa;
b1482674 1429 vm_page_t mpte;
984263bc 1430
b12defdc
MD
1431 ASSERT_LWKT_TOKEN_HELD(vm_object_token(pmap->pm_pteobj));
1432
984263bc
MD
1433 /*
1434 * Calculate pagetable page index
1435 */
1436 ptepindex = va >> PDRSHIFT;
1437
1438 /*
1439 * Get the page directory entry
1440 */
1441 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1442
1443 /*
1444 * This supports switching from a 4MB page to a
1445 * normal 4K page.
1446 */
1447 if (ptepa & PG_PS) {
1448 pmap->pm_pdir[ptepindex] = 0;
1449 ptepa = 0;
0f7a3396 1450 smp_invltlb();
54341a3b 1451 cpu_invltlb();
984263bc
MD
1452 }
1453
1454 /*
1455 * If the page table page is mapped, we just increment the
90244566 1456 * wire count, and activate it.
984263bc
MD
1457 */
1458 if (ptepa) {
1459 /*
1460 * In order to get the page table page, try the
1461 * hint first.
1462 */
b1482674
MD
1463 if ((mpte = pmap->pm_ptphint) != NULL &&
1464 (mpte->pindex == ptepindex) &&
1465 (mpte->flags & PG_BUSY) == 0) {
1466 vm_page_wire_quick(mpte);
984263bc 1467 } else {
b1482674
MD
1468 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1469 pmap->pm_ptphint = mpte;
1470 vm_page_wire_quick(mpte);
1471 vm_page_wakeup(mpte);
984263bc 1472 }
b1482674 1473 return mpte;
984263bc
MD
1474 }
1475 /*
1476 * Here if the pte page isn't mapped, or if it has been deallocated.
1477 */
1478 return _pmap_allocpte(pmap, ptepindex);
1479}
1480
1481
1482/***************************************************
1f804340 1483 * Pmap allocation/deallocation routines.
984263bc
MD
1484 ***************************************************/
1485
1486/*
1487 * Release any resources held by the given physical map.
1488 * Called when a pmap initialized by pmap_pinit is being released.
1489 * Should only be called if the map contains no valid mappings.
4107b0c0 1490 *
b12defdc 1491 * Caller must hold pmap->pm_token
984263bc 1492 */
1f804340
MD
1493static int pmap_release_callback(struct vm_page *p, void *data);
1494
984263bc 1495void
840de426 1496pmap_release(struct pmap *pmap)
984263bc 1497{
984263bc 1498 vm_object_t object = pmap->pm_pteobj;
1f804340 1499 struct rb_vm_page_scan_info info;
984263bc 1500
4107b0c0
MD
1501 KASSERT(pmap->pm_active == 0,
1502 ("pmap still active! %08x", pmap->pm_active));
984263bc
MD
1503#if defined(DIAGNOSTIC)
1504 if (object->ref_count != 1)
1505 panic("pmap_release: pteobj reference count != 1");
1506#endif
1507
1f804340
MD
1508 info.pmap = pmap;
1509 info.object = object;
b12defdc
MD
1510
1511 spin_lock(&pmap_spin);
54a764e8 1512 TAILQ_REMOVE(&pmap_list, pmap, pm_pmnode);
b12defdc 1513 spin_unlock(&pmap_spin);
1f804340 1514
b12defdc 1515 vm_object_hold(object);
1f804340 1516 do {
1f804340
MD
1517 info.error = 0;
1518 info.mpte = NULL;
1519 info.limit = object->generation;
1520
1521 vm_page_rb_tree_RB_SCAN(&object->rb_memq, NULL,
1522 pmap_release_callback, &info);
1523 if (info.error == 0 && info.mpte) {
1524 if (!pmap_release_free_page(pmap, info.mpte))
1525 info.error = 1;
984263bc 1526 }
1f804340 1527 } while (info.error);
2f2d9e58 1528 vm_object_drop(object);
b12defdc
MD
1529
1530 pmap->pm_cached = 0;
1f804340
MD
1531}
1532
4107b0c0
MD
1533/*
1534 * The caller must hold vm_token.
1535 */
1f804340
MD
1536static int
1537pmap_release_callback(struct vm_page *p, void *data)
1538{
1539 struct rb_vm_page_scan_info *info = data;
1540
1541 if (p->pindex == PTDPTDI) {
1542 info->mpte = p;
1543 return(0);
344ad853 1544 }
1f804340
MD
1545 if (!pmap_release_free_page(info->pmap, p)) {
1546 info->error = 1;
1547 return(-1);
1548 }
1549 if (info->object->generation != info->limit) {
1550 info->error = 1;
1551 return(-1);
1552 }
1553 return(0);
984263bc 1554}
984263bc
MD
1555
1556/*
0e5797fe 1557 * Grow the number of kernel page table entries, if needed.
4107b0c0
MD
1558 *
1559 * No requirements.
984263bc
MD
1560 */
1561void
a8cf2878 1562pmap_growkernel(vm_offset_t kstart, vm_offset_t kend)
984263bc 1563{
a8cf2878 1564 vm_offset_t addr = kend;
54a764e8 1565 struct pmap *pmap;
984263bc
MD
1566 vm_offset_t ptppaddr;
1567 vm_page_t nkpg;
1568 pd_entry_t newpdir;
1569
b12defdc 1570 vm_object_hold(kptobj);
984263bc
MD
1571 if (kernel_vm_end == 0) {
1572 kernel_vm_end = KERNBASE;
1573 nkpt = 0;
1574 while (pdir_pde(PTD, kernel_vm_end)) {
4107b0c0
MD
1575 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1576 ~(PAGE_SIZE * NPTEPG - 1);
984263bc
MD
1577 nkpt++;
1578 }
1579 }
1580 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1581 while (kernel_vm_end < addr) {
1582 if (pdir_pde(PTD, kernel_vm_end)) {
4107b0c0
MD
1583 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1584 ~(PAGE_SIZE * NPTEPG - 1);
984263bc
MD
1585 continue;
1586 }
1587
1588 /*
1589 * This index is bogus, but out of the way
1590 */
4107b0c0
MD
1591 nkpg = vm_page_alloc(kptobj, nkpt, VM_ALLOC_NORMAL |
1592 VM_ALLOC_SYSTEM |
1593 VM_ALLOC_INTERRUPT);
dc1fd4b3 1594 if (nkpg == NULL)
984263bc
MD
1595 panic("pmap_growkernel: no memory to grow kernel");
1596
984263bc
MD
1597 vm_page_wire(nkpg);
1598 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1599 pmap_zero_page(ptppaddr);
1600 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1601 pdir_pde(PTD, kernel_vm_end) = newpdir;
fbbaeba3 1602 *pmap_pde(&kernel_pmap, kernel_vm_end) = newpdir;
0e5797fe
MD
1603 nkpt++;
1604
1605 /*
54a764e8 1606 * This update must be interlocked with pmap_pinit2.
0e5797fe 1607 */
b12defdc 1608 spin_lock(&pmap_spin);
54a764e8
MD
1609 TAILQ_FOREACH(pmap, &pmap_list, pm_pmnode) {
1610 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1611 }
b12defdc 1612 spin_unlock(&pmap_spin);
54a764e8
MD
1613 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1614 ~(PAGE_SIZE * NPTEPG - 1);
984263bc 1615 }
b12defdc 1616 vm_object_drop(kptobj);
984263bc
MD
1617}
1618
1619/*
4107b0c0
MD
1620 * Add a reference to the specified pmap.
1621 *
1622 * No requirements.
984263bc
MD
1623 */
1624void
840de426 1625pmap_reference(pmap_t pmap)
984263bc 1626{
4107b0c0
MD
1627 if (pmap) {
1628 lwkt_gettoken(&vm_token);
1629 ++pmap->pm_count;
1630 lwkt_reltoken(&vm_token);
984263bc
MD
1631 }
1632}
1633
3321ee05
MD
1634/*
1635 * vm_token must be held
1636 */
1637static
1638void
1639pmap_hold(pmap_t pmap)
1640{
1641 ++pmap->pm_count;
1642}
1643
1644/*
1645 * vm_token must be held
1646 */
1647static
1648void
1649pmap_drop(pmap_t pmap)
1650{
1651 --pmap->pm_count;
22c4e116 1652 if (pmap->pm_count == (int)0x80000000)
3321ee05
MD
1653 wakeup(pmap);
1654}
1655
1656static
1657void
1658pmap_wait(pmap_t pmap, int count)
1659{
1660 lwkt_gettoken(&vm_token);
1661 pmap->pm_count += count;
1662 if (pmap->pm_count & 0x7FFFFFFF) {
1663 while (pmap->pm_count & 0x7FFFFFFF) {
1664 pmap->pm_count |= 0x80000000;
1665 tsleep(pmap, 0, "pmapd", 0);
1666 pmap->pm_count &= ~0x80000000;
1667 kprintf("pmap_wait: race averted\n");
1668 }
1669 }
1670 lwkt_reltoken(&vm_token);
1671}
1672
984263bc 1673/***************************************************
4107b0c0 1674 * page management routines.
984263bc
MD
1675 ***************************************************/
1676
1677/*
8a8d5d85
MD
1678 * free the pv_entry back to the free list. This function may be
1679 * called from an interrupt.
4107b0c0
MD
1680 *
1681 * The caller must hold vm_token.
984263bc
MD
1682 */
1683static PMAP_INLINE void
840de426 1684free_pv_entry(pv_entry_t pv)
984263bc 1685{
2bb9cc6f
MD
1686 struct mdglobaldata *gd;
1687
5926987a
MD
1688#ifdef PMAP_DEBUG
1689 KKASSERT(pv->pv_m != NULL);
1690 pv->pv_m = NULL;
1691#endif
2bb9cc6f 1692 gd = mdcpu;
984263bc 1693 pv_entry_count--;
2bb9cc6f
MD
1694 if (gd->gd_freepv == NULL)
1695 gd->gd_freepv = pv;
1696 else
1697 zfree(pvzone, pv);
984263bc
MD
1698}
1699
1700/*
1701 * get a new pv_entry, allocating a block from the system
2bb9cc6f
MD
1702 * when needed. This function may be called from an interrupt thread.
1703 *
1704 * THIS FUNCTION CAN BLOCK ON THE ZALLOC TOKEN, serialization of other
1705 * tokens (aka vm_token) to be temporarily lost.
4107b0c0
MD
1706 *
1707 * The caller must hold vm_token.
984263bc
MD
1708 */
1709static pv_entry_t
1710get_pv_entry(void)
1711{
2bb9cc6f
MD
1712 struct mdglobaldata *gd;
1713 pv_entry_t pv;
1714
984263bc
MD
1715 pv_entry_count++;
1716 if (pv_entry_high_water &&
20479584
MD
1717 (pv_entry_count > pv_entry_high_water) &&
1718 (pmap_pagedaemon_waken == 0)) {
984263bc
MD
1719 pmap_pagedaemon_waken = 1;
1720 wakeup (&vm_pages_needed);
1721 }
2bb9cc6f
MD
1722 gd = mdcpu;
1723 if ((pv = gd->gd_freepv) != NULL)
1724 gd->gd_freepv = NULL;
1725 else
1726 pv = zalloc(pvzone);
1727 return pv;
984263bc
MD
1728}
1729
1730/*
1731 * This routine is very drastic, but can save the system
1732 * in a pinch.
4107b0c0
MD
1733 *
1734 * No requirements.
984263bc
MD
1735 */
1736void
840de426 1737pmap_collect(void)
984263bc
MD
1738{
1739 int i;
1740 vm_page_t m;
1741 static int warningdone=0;
1742
1743 if (pmap_pagedaemon_waken == 0)
1744 return;
4107b0c0 1745 lwkt_gettoken(&vm_token);
20479584 1746 pmap_pagedaemon_waken = 0;
984263bc
MD
1747
1748 if (warningdone < 5) {
948209ce
MD
1749 kprintf("pmap_collect: collecting pv entries -- "
1750 "suggest increasing PMAP_SHPGPERPROC\n");
984263bc
MD
1751 warningdone++;
1752 }
1753
b12defdc 1754 for (i = 0; i < vm_page_array_size; i++) {
984263bc 1755 m = &vm_page_array[i];
b12defdc 1756 if (m->wire_count || m->hold_count)
984263bc 1757 continue;
b12defdc
MD
1758 if (vm_page_busy_try(m, TRUE) == 0) {
1759 if (m->wire_count == 0 && m->hold_count == 0) {
1760 pmap_remove_all(m);
1761 }
1762 vm_page_wakeup(m);
4107b0c0 1763 }
984263bc 1764 }
4107b0c0 1765 lwkt_reltoken(&vm_token);
984263bc
MD
1766}
1767
1768
1769/*
b1482674
MD
1770 * Remove the pv entry and unwire the page table page related to the
1771 * pte the caller has cleared from the page table.
4107b0c0
MD
1772 *
1773 * The caller must hold vm_token.
984263bc 1774 */
554cf9ac 1775static void
0f7a3396 1776pmap_remove_entry(struct pmap *pmap, vm_page_t m,
4107b0c0 1777 vm_offset_t va, pmap_inval_info_t info)
984263bc
MD
1778{
1779 pv_entry_t pv;
984263bc 1780
b1482674
MD
1781 /*
1782 * Cannot block
1783 */
4107b0c0 1784 ASSERT_LWKT_TOKEN_HELD(&vm_token);
984263bc
MD
1785 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1786 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1787 if (pmap == pv->pv_pmap && va == pv->pv_va)
1788 break;
1789 }
1790 } else {
1791 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
5926987a
MD
1792#ifdef PMAP_DEBUG
1793 KKASSERT(pv->pv_pmap == pmap);
1794#endif
1795 if (va == pv->pv_va)
984263bc
MD
1796 break;
1797 }
1798 }
5926987a 1799 KKASSERT(pv);
984263bc 1800
b1482674
MD
1801 /*
1802 * Cannot block
1803 */
5926987a
MD
1804 test_m_maps_pv(m, pv);
1805 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1806 m->md.pv_list_count--;
cef01e15
MD
1807 if (m->object)
1808 atomic_add_int(&m->object->agg_pv_list_count, -1);
5926987a
MD
1809 if (TAILQ_EMPTY(&m->md.pv_list))
1810 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1811 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1812 ++pmap->pm_generation;
b1482674
MD
1813
1814 /*
1815 * This can block.
1816 */
b12defdc 1817 vm_object_hold(pmap->pm_pteobj);
554cf9ac 1818 pmap_unuse_pt(pmap, va, pv->pv_ptem, info);
b12defdc 1819 vm_object_drop(pmap->pm_pteobj);
5926987a 1820 free_pv_entry(pv);
984263bc
MD
1821}
1822
1823/*
4107b0c0
MD
1824 * Create a pv entry for page at pa for (pmap, va).
1825 *
1826 * The caller must hold vm_token.
984263bc
MD
1827 */
1828static void
2bb9cc6f
MD
1829pmap_insert_entry(pmap_t pmap, pv_entry_t pv, vm_offset_t va,
1830 vm_page_t mpte, vm_page_t m)
984263bc 1831{
5926987a
MD
1832#ifdef PMAP_DEBUG
1833 KKASSERT(pv->pv_m == NULL);
1834 pv->pv_m = m;
1835#endif
984263bc
MD
1836 pv->pv_va = va;
1837 pv->pv_pmap = pmap;
1838 pv->pv_ptem = mpte;
1839
1840 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1841 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
5926987a 1842 ++pmap->pm_generation;
984263bc 1843 m->md.pv_list_count++;
cef01e15
MD
1844 if (m->object)
1845 atomic_add_int(&m->object->agg_pv_list_count, 1);
984263bc
MD
1846}
1847
1848/*
5926987a
MD
1849 * pmap_remove_pte: do the things to unmap a page in a process.
1850 *
4107b0c0
MD
1851 * The caller must hold vm_token.
1852 *
1853 * WARNING! As with most other pmap functions this one can block, so
1854 * callers using temporary page table mappings must reload
1855 * them.
984263bc 1856 */
554cf9ac 1857static void
0f7a3396 1858pmap_remove_pte(struct pmap *pmap, unsigned *ptq, vm_offset_t va,
5926987a 1859 pmap_inval_info_t info)
984263bc
MD
1860{
1861 unsigned oldpte;
1862 vm_page_t m;
1863
5926987a 1864 ptbase_assert(pmap);
c2fb025d 1865 pmap_inval_interlock(info, pmap, va);
5926987a 1866 ptbase_assert(pmap);
984263bc
MD
1867 oldpte = loadandclear(ptq);
1868 if (oldpte & PG_W)
1869 pmap->pm_stats.wired_count -= 1;
c2fb025d 1870 pmap_inval_deinterlock(info, pmap);
90244566 1871 KKASSERT(oldpte & PG_V);
984263bc
MD
1872 /*
1873 * Machines that don't support invlpg, also don't support
0f7a3396
MD
1874 * PG_G. XXX PG_G is disabled for SMP so don't worry about
1875 * the SMP case.
984263bc
MD
1876 */
1877 if (oldpte & PG_G)
41a01a4d 1878 cpu_invlpg((void *)va);
eec2b734
MD
1879 KKASSERT(pmap->pm_stats.resident_count > 0);
1880 --pmap->pm_stats.resident_count;
984263bc
MD
1881 if (oldpte & PG_MANAGED) {
1882 m = PHYS_TO_VM_PAGE(oldpte);
1883 if (oldpte & PG_M) {
1884#if defined(PMAP_DIAGNOSTIC)
1885 if (pmap_nw_modified((pt_entry_t) oldpte)) {
d557216f
MD
1886 kprintf("pmap_remove: modified page not "
1887 "writable: va: %p, pte: 0x%lx\n",
1888 (void *)va, (long)oldpte);
984263bc
MD
1889 }
1890#endif
1891 if (pmap_track_modified(va))
1892 vm_page_dirty(m);
1893 }
1894 if (oldpte & PG_A)
1895 vm_page_flag_set(m, PG_REFERENCED);
554cf9ac 1896 pmap_remove_entry(pmap, m, va, info);
984263bc 1897 } else {
554cf9ac 1898 pmap_unuse_pt(pmap, va, NULL, info);
984263bc 1899 }
984263bc
MD
1900}
1901
1902/*
5926987a 1903 * Remove a single page from a process address space.
e0e69b7d 1904 *
4107b0c0 1905 * The caller must hold vm_token.
984263bc
MD
1906 */
1907static void
0f7a3396 1908pmap_remove_page(struct pmap *pmap, vm_offset_t va, pmap_inval_info_t info)
984263bc 1909{
840de426 1910 unsigned *ptq;
984263bc
MD
1911
1912 /*
90244566 1913 * If there is no pte for this address, just skip it!!! Otherwise
e0e69b7d 1914 * get a local va for mappings for this pmap and remove the entry.
984263bc 1915 */
e0e69b7d
MD
1916 if (*pmap_pde(pmap, va) != 0) {
1917 ptq = get_ptbase(pmap) + i386_btop(va);
1918 if (*ptq) {
0f7a3396 1919 pmap_remove_pte(pmap, ptq, va, info);
5926987a 1920 /* ptq invalid */
e0e69b7d 1921 }
984263bc 1922 }
984263bc
MD
1923}
1924
1925/*
4107b0c0 1926 * Remove the given range of addresses from the specified map.
984263bc 1927 *
4107b0c0
MD
1928 * It is assumed that the start and end are properly rounded to the page
1929 * size.
e0e69b7d 1930 *
4107b0c0 1931 * No requirements.
984263bc
MD
1932 */
1933void
840de426 1934pmap_remove(struct pmap *pmap, vm_offset_t sva, vm_offset_t eva)
984263bc 1935{
840de426 1936 unsigned *ptbase;
984263bc
MD
1937 vm_offset_t pdnxt;
1938 vm_offset_t ptpaddr;
1939 vm_offset_t sindex, eindex;
0f7a3396 1940 struct pmap_inval_info info;
984263bc
MD
1941
1942 if (pmap == NULL)
1943 return;
1944
b12defdc 1945 vm_object_hold(pmap->pm_pteobj);
4107b0c0
MD
1946 lwkt_gettoken(&vm_token);
1947 if (pmap->pm_stats.resident_count == 0) {
1948 lwkt_reltoken(&vm_token);
b12defdc 1949 vm_object_drop(pmap->pm_pteobj);
984263bc 1950 return;
4107b0c0 1951 }
984263bc 1952
0f7a3396
MD
1953 pmap_inval_init(&info);
1954
984263bc
MD
1955 /*
1956 * special handling of removing one page. a very
1957 * common operation and easy to short circuit some
1958 * code.
1959 */
1960 if (((sva + PAGE_SIZE) == eva) &&
1961 (((unsigned) pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
0f7a3396 1962 pmap_remove_page(pmap, sva, &info);
c2fb025d 1963 pmap_inval_done(&info);
4107b0c0 1964 lwkt_reltoken(&vm_token);
b12defdc 1965 vm_object_drop(pmap->pm_pteobj);
984263bc
MD
1966 return;
1967 }
1968
984263bc
MD
1969 /*
1970 * Get a local virtual address for the mappings that are being
1971 * worked with.
1972 */
984263bc
MD
1973 sindex = i386_btop(sva);
1974 eindex = i386_btop(eva);
1975
554cf9ac 1976 while (sindex < eindex) {
984263bc
MD
1977 unsigned pdirindex;
1978
1979 /*
554cf9ac 1980 * Stop scanning if no pages are left
984263bc 1981 */
984263bc
MD
1982 if (pmap->pm_stats.resident_count == 0)
1983 break;
1984
554cf9ac
MD
1985 /*
1986 * Calculate index for next page table, limited by eindex.
1987 */
1988 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1989 if (pdnxt > eindex)
1990 pdnxt = eindex;
1991
984263bc 1992 pdirindex = sindex / NPDEPG;
554cf9ac
MD
1993 ptpaddr = (unsigned)pmap->pm_pdir[pdirindex];
1994 if (ptpaddr & PG_PS) {
c2fb025d 1995 pmap_inval_interlock(&info, pmap, -1);
984263bc
MD
1996 pmap->pm_pdir[pdirindex] = 0;
1997 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
be3aecf7 1998 pmap->pm_cached = 0;
c2fb025d 1999 pmap_inval_deinterlock(&info, pmap);
554cf9ac 2000 sindex = pdnxt;
984263bc
MD
2001 continue;
2002 }
2003
2004 /*
2005 * Weed out invalid mappings. Note: we assume that the page
2006 * directory table is always allocated, and in kernel virtual.
2007 */
554cf9ac
MD
2008 if (ptpaddr == 0) {
2009 sindex = pdnxt;
984263bc 2010 continue;
984263bc
MD
2011 }
2012
8790d7d8 2013 /*
554cf9ac
MD
2014 * Sub-scan the page table page. pmap_remove_pte() can
2015 * block on us, invalidating ptbase, so we must reload
2016 * ptbase and we must also check whether the page directory
2017 * page is still present.
8790d7d8 2018 */
554cf9ac 2019 while (sindex < pdnxt) {
984263bc 2020 vm_offset_t va;
8790d7d8
MD
2021
2022 ptbase = get_ptbase(pmap);
554cf9ac
MD
2023 if (ptbase[sindex]) {
2024 va = i386_ptob(sindex);
2025 pmap_remove_pte(pmap, ptbase + sindex,
2026 va, &info);
2027 }
2028 if (pmap->pm_pdir[pdirindex] == 0 ||
2029 (pmap->pm_pdir[pdirindex] & PG_PS)) {
984263bc 2030 break;
554cf9ac
MD
2031 }
2032 ++sindex;
984263bc
MD
2033 }
2034 }
c2fb025d 2035 pmap_inval_done(&info);
4107b0c0 2036 lwkt_reltoken(&vm_token);
b12defdc 2037 vm_object_drop(pmap->pm_pteobj);
984263bc
MD
2038}
2039
2040/*
4107b0c0
MD
2041 * Removes this physical page from all physical maps in which it resides.
2042 * Reflects back modify bits to the pager.
984263bc 2043 *
3321ee05 2044 * vm_token must be held by caller.
984263bc 2045 */
984263bc 2046static void
840de426 2047pmap_remove_all(vm_page_t m)
984263bc 2048{
0f7a3396 2049 struct pmap_inval_info info;
840de426 2050 unsigned *pte, tpte;
0f7a3396 2051 pv_entry_t pv;
3321ee05 2052 pmap_t pmap;
984263bc 2053
bee81bdd
SS
2054 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2055 return;
bcc6a8ac
MD
2056 if (TAILQ_EMPTY(&m->md.pv_list))
2057 return;
984263bc 2058
bcc6a8ac 2059 lwkt_getoken(&vm_token);
0f7a3396 2060 pmap_inval_init(&info);
984263bc 2061 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3321ee05
MD
2062 pmap = pv->pv_pmap;
2063 KKASSERT(pmap->pm_stats.resident_count > 0);
2064 --pmap->pm_stats.resident_count;
2065 pmap_hold(pmap);
984263bc 2066
3321ee05
MD
2067 pte = pmap_pte_quick(pmap, pv->pv_va);
2068 pmap_inval_interlock(&info, pmap, pv->pv_va);
984263bc
MD
2069 tpte = loadandclear(pte);
2070 if (tpte & PG_W)
3321ee05
MD
2071 pmap->pm_stats.wired_count--;
2072 pmap_inval_deinterlock(&info, pmap);
984263bc
MD
2073 if (tpte & PG_A)
2074 vm_page_flag_set(m, PG_REFERENCED);
c2fb025d 2075 KKASSERT(PHYS_TO_VM_PAGE(tpte) == m);
984263bc
MD
2076
2077 /*
2078 * Update the vm_page_t clean and reference bits.
2079 */
2080 if (tpte & PG_M) {
2081#if defined(PMAP_DIAGNOSTIC)
2082 if (pmap_nw_modified((pt_entry_t) tpte)) {
d557216f
MD
2083 kprintf("pmap_remove_all: modified page "
2084 "not writable: va: %p, pte: 0x%lx\n",
2085 (void *)pv->pv_va, (long)tpte);
984263bc
MD
2086 }
2087#endif
2088 if (pmap_track_modified(pv->pv_va))
2089 vm_page_dirty(m);
2090 }
5926987a
MD
2091#ifdef PMAP_DEBUG
2092 KKASSERT(pv->pv_m == m);
2093#endif
2bb9cc6f 2094 KKASSERT(pv == TAILQ_FIRST(&m->md.pv_list));
984263bc 2095 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3321ee05
MD
2096 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2097 ++pmap->pm_generation;
984263bc 2098 m->md.pv_list_count--;
cef01e15
MD
2099 if (m->object)
2100 atomic_add_int(&m->object->agg_pv_list_count, -1);
17cde63e
MD
2101 if (TAILQ_EMPTY(&m->md.pv_list))
2102 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
3321ee05
MD
2103 vm_object_hold(pmap->pm_pteobj);
2104 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
2105 vm_object_drop(pmap->pm_pteobj);
984263bc 2106 free_pv_entry(pv);
3321ee05 2107 pmap_drop(pmap);
984263bc 2108 }
17cde63e 2109 KKASSERT((m->flags & (PG_MAPPED|PG_WRITEABLE)) == 0);
c2fb025d 2110 pmap_inval_done(&info);
bcc6a8ac 2111 lwkt_reloken(&vm_token);
984263bc
MD
2112}
2113
2114/*
4107b0c0
MD
2115 * Set the physical protection on the specified range of this map
2116 * as requested.
e0e69b7d 2117 *
4107b0c0 2118 * No requirements.
984263bc
MD
2119 */
2120void
2121pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
2122{
840de426 2123 unsigned *ptbase;
984263bc
MD
2124 vm_offset_t pdnxt, ptpaddr;
2125 vm_pindex_t sindex, eindex;
0f7a3396 2126 pmap_inval_info info;
984263bc
MD
2127
2128 if (pmap == NULL)
2129 return;
2130
2131 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2132 pmap_remove(pmap, sva, eva);
2133 return;
2134 }
2135
2136 if (prot & VM_PROT_WRITE)
2137 return;
2138
4107b0c0 2139 lwkt_gettoken(&vm_token);
0f7a3396 2140 pmap_inval_init(&info);
984263bc
MD
2141
2142 ptbase = get_ptbase(pmap);
2143
2144 sindex = i386_btop(sva);
2145 eindex = i386_btop(eva);
2146
2147 for (; sindex < eindex; sindex = pdnxt) {
984263bc
MD
2148 unsigned pdirindex;
2149
2150 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
2151
2152 pdirindex = sindex / NPDEPG;
2153 if (((ptpaddr = (unsigned) pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
c2fb025d 2154 pmap_inval_interlock(&info, pmap, -1);
55f2596a 2155 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
984263bc 2156 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
c2fb025d 2157 pmap_inval_deinterlock(&info, pmap);
984263bc
MD
2158 continue;
2159 }
2160
2161 /*
2162 * Weed out invalid mappings. Note: we assume that the page
2163 * directory table is always allocated, and in kernel virtual.
2164 */
2165 if (ptpaddr == 0)
2166 continue;
2167
2168 if (pdnxt > eindex) {
2169 pdnxt = eindex;
2170 }
2171
2172 for (; sindex != pdnxt; sindex++) {
984263bc 2173 unsigned pbits;
c2fb025d 2174 unsigned cbits;
984263bc
MD
2175 vm_page_t m;
2176
17cde63e 2177 /*
d5b2d319 2178 * XXX non-optimal.
17cde63e 2179 */
c2fb025d
MD
2180 pmap_inval_interlock(&info, pmap, i386_ptob(sindex));
2181again:
984263bc 2182 pbits = ptbase[sindex];
c2fb025d 2183 cbits = pbits;
984263bc
MD
2184
2185 if (pbits & PG_MANAGED) {
2186 m = NULL;
2187 if (pbits & PG_A) {
2188 m = PHYS_TO_VM_PAGE(pbits);
2189 vm_page_flag_set(m, PG_REFERENCED);
c2fb025d 2190 cbits &= ~PG_A;
984263bc
MD
2191 }
2192 if (pbits & PG_M) {
2193 if (pmap_track_modified(i386_ptob(sindex))) {
2194 if (m == NULL)
2195 m = PHYS_TO_VM_PAGE(pbits);
2196 vm_page_dirty(m);
c2fb025d 2197 cbits &= ~PG_M;
984263bc
MD
2198 }
2199 }
2200 }
c2fb025d
MD
2201 cbits &= ~PG_RW;
2202 if (pbits != cbits &&
2203 !atomic_cmpset_int(ptbase + sindex, pbits, cbits)) {
2204 goto again;
984263bc 2205 }
c2fb025d 2206 pmap_inval_deinterlock(&info, pmap);
984263bc
MD
2207 }
2208 }
c2fb025d 2209 pmap_inval_done(&info);
4107b0c0 2210 lwkt_reltoken(&vm_token);
984263bc
MD
2211}
2212
2213/*
4107b0c0
MD
2214 * Insert the given physical page (p) at the specified virtual address (v)
2215 * in the target physical map with the protection requested.
984263bc 2216 *
4107b0c0
MD
2217 * If specified, the page will be wired down, meaning that the related pte
2218 * cannot be reclaimed.
984263bc 2219 *
4107b0c0 2220 * No requirements.
984263bc
MD
2221 */
2222void
2223pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
921c891e 2224 boolean_t wired, vm_map_entry_t entry __unused)
984263bc 2225{
6ef943a3 2226 vm_paddr_t pa;
840de426 2227 unsigned *pte;
6ef943a3 2228 vm_paddr_t opa;
984263bc
MD
2229 vm_offset_t origpte, newpte;
2230 vm_page_t mpte;
0f7a3396 2231 pmap_inval_info info;
2bb9cc6f 2232 pv_entry_t pv;
984263bc
MD
2233
2234 if (pmap == NULL)
2235 return;
2236
2237 va &= PG_FRAME;
2238#ifdef PMAP_DIAGNOSTIC
c439ad8f 2239 if (va >= KvaEnd)
984263bc 2240 panic("pmap_enter: toobig");
d557216f
MD
2241 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS)) {
2242 panic("pmap_enter: invalid to pmap_enter page "
2243 "table pages (va: %p)", (void *)va);
2244 }
984263bc 2245#endif
fbbaeba3
MD
2246 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2247 kprintf("Warning: pmap_enter called on UVA with kernel_pmap\n");
7ce2998e 2248 print_backtrace(-1);
fbbaeba3
MD
2249 }
2250 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2251 kprintf("Warning: pmap_enter called on KVA without kernel_pmap\n");
7ce2998e 2252 print_backtrace(-1);
fbbaeba3 2253 }
984263bc 2254
b12defdc 2255 vm_object_hold(pmap->pm_pteobj);
4107b0c0
MD
2256 lwkt_gettoken(&vm_token);
2257
984263bc 2258 /*
2bb9cc6f
MD
2259 * This can block, get it before we do anything important.
2260 */
2261 if (pmap_initialized &&
2262 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2263 pv = get_pv_entry();
2264 } else {
2265 pv = NULL;
2266 }
2267
2268 /*
984263bc
MD
2269 * In the case that a page table page is not
2270 * resident, we are creating it here.
2271 */
17cde63e 2272 if (va < UPT_MIN_ADDRESS)
984263bc 2273 mpte = pmap_allocpte(pmap, va);
17cde63e
MD
2274 else
2275 mpte = NULL;
984263bc 2276
b12defdc
MD
2277 if ((prot & VM_PROT_NOSYNC) == 0)
2278 pmap_inval_init(&info);
984263bc
MD
2279 pte = pmap_pte(pmap, va);
2280
2281 /*
2282 * Page Directory table entry not valid, we need a new PT page
2283 */
2284 if (pte == NULL) {
ed20d0e3 2285 panic("pmap_enter: invalid page directory pdir=0x%lx, va=%p",
d557216f 2286 (long)pmap->pm_pdir[PTDPTDI], (void *)va);
984263bc
MD
2287 }
2288
2289 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
2290 origpte = *(vm_offset_t *)pte;
2291 opa = origpte & PG_FRAME;
2292
2293 if (origpte & PG_PS)
2294 panic("pmap_enter: attempted pmap_enter on 4MB page");
2295
2296 /*
2297 * Mapping has not changed, must be protection or wiring change.
2298 */
2299 if (origpte && (opa == pa)) {
2300 /*
2301 * Wiring change, just update stats. We don't worry about
2302 * wiring PT pages as they remain resident as long as there
2303 * are valid mappings in them. Hence, if a user page is wired,
2304 * the PT page will be also.
2305 */
2306 if (wired && ((origpte & PG_W) == 0))
2307 pmap->pm_stats.wired_count++;
2308 else if (!wired && (origpte & PG_W))
2309 pmap->pm_stats.wired_count--;
2310
2311#if defined(PMAP_DIAGNOSTIC)
2312 if (pmap_nw_modified((pt_entry_t) origpte)) {
d557216f
MD
2313 kprintf("pmap_enter: modified page not "
2314 "writable: va: %p, pte: 0x%lx\n",
2315 (void *)va, (long )origpte);
984263bc
MD
2316 }
2317#endif
2318
2319 /*
984263bc
MD
2320 * We might be turning off write access to the page,
2321 * so we go ahead and sense modify status.
2322 */
2323 if (origpte & PG_MANAGED) {
2324 if ((origpte & PG_M) && pmap_track_modified(va)) {
2325 vm_page_t om;
2326 om = PHYS_TO_VM_PAGE(opa);
2327 vm_page_dirty(om);
2328 }
2329 pa |= PG_MANAGED;
17cde63e 2330 KKASSERT(m->flags & PG_MAPPED);
984263bc
MD
2331 }
2332 goto validate;
2333 }
2334 /*
2335 * Mapping has changed, invalidate old range and fall through to
2336 * handle validating new mapping.
5926987a
MD
2337 *
2338 * Since we have a ref on the page directory page pmap_pte()
2339 * will always return non-NULL.
2340 *
2341 * NOTE: pmap_remove_pte() can block and cause the temporary ptbase
2342 * to get wiped. reload the ptbase. I'm not sure if it is
2343 * also possible to race another pmap_enter() but check for
2344 * that case too.
984263bc 2345 */
5926987a 2346 while (opa) {
5926987a
MD
2347 KKASSERT((origpte & PG_FRAME) ==
2348 (*(vm_offset_t *)pte & PG_FRAME));
554cf9ac 2349 pmap_remove_pte(pmap, pte, va, &info);
5926987a
MD
2350 pte = pmap_pte(pmap, va);
2351 origpte = *(vm_offset_t *)pte;
2352 opa = origpte & PG_FRAME;
2353 if (opa) {
2354 kprintf("pmap_enter: Warning, raced pmap %p va %p\n",
2355 pmap, (void *)va);
2356 }
984263bc
MD
2357 }
2358
2359 /*
2360 * Enter on the PV list if part of our managed memory. Note that we
2361 * raise IPL while manipulating pv_table since pmap_enter can be
2362 * called at interrupt time.
2363 */
2364 if (pmap_initialized &&
2365 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2bb9cc6f
MD
2366 pmap_insert_entry(pmap, pv, va, mpte, m);
2367 pv = NULL;
5926987a 2368 ptbase_assert(pmap);
984263bc 2369 pa |= PG_MANAGED;
17cde63e 2370 vm_page_flag_set(m, PG_MAPPED);
984263bc
MD
2371 }
2372
2373 /*
2374 * Increment counters
2375 */
eec2b734 2376 ++pmap->pm_stats.resident_count;
984263bc
MD
2377 if (wired)
2378 pmap->pm_stats.wired_count++;
5926987a 2379 KKASSERT(*pte == 0);
984263bc
MD
2380
2381validate:
2382 /*
2383 * Now validate mapping with desired protection/wiring.
2384 */
5926987a 2385 ptbase_assert(pmap);
984263bc
MD
2386 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2387
2388 if (wired)
2389 newpte |= PG_W;
2390 if (va < UPT_MIN_ADDRESS)
2391 newpte |= PG_U;
fbbaeba3 2392 if (pmap == &kernel_pmap)
984263bc
MD
2393 newpte |= pgeflag;
2394
2395 /*
2bb9cc6f
MD
2396 * If the mapping or permission bits are different, we need
2397 * to update the pte. If the pte is already present we have
2398 * to get rid of the extra wire-count on mpte we had obtained
2399 * above.
b1482674
MD
2400 *
2401 * mpte has a new wire_count, which also serves to prevent the
2402 * page table page from getting ripped out while we work. If we
2403 * are modifying an existing pte instead of installing a new one
2404 * we have to drop it.
984263bc
MD
2405 */
2406 if ((origpte & ~(PG_M|PG_A)) != newpte) {
b12defdc
MD
2407 if (prot & VM_PROT_NOSYNC)
2408 cpu_invlpg((void *)va);
2409 else
2410 pmap_inval_interlock(&info, pmap, va);
5926987a 2411 ptbase_assert(pmap);
2bb9cc6f
MD
2412
2413 if (*pte) {
2414 KKASSERT((*pte & PG_FRAME) == (newpte & PG_FRAME));
2415 if (vm_page_unwire_quick(mpte))
2416 panic("pmap_enter: Insufficient wire_count");
2417 }
2418
984263bc 2419 *pte = newpte | PG_A;
b12defdc
MD
2420 if ((prot & VM_PROT_NOSYNC) == 0)
2421 pmap_inval_deinterlock(&info, pmap);
17cde63e
MD
2422 if (newpte & PG_RW)
2423 vm_page_flag_set(m, PG_WRITEABLE);
b1482674
MD
2424 } else {
2425 if (*pte) {
2426 KKASSERT((*pte & PG_FRAME) == (newpte & PG_FRAME));
2427 if (vm_page_unwire_quick(mpte))
2428 panic("pmap_enter: Insufficient wire_count");
2429 }
984263bc 2430 }
b1482674
MD
2431
2432 /*
2433 * NOTE: mpte invalid after this point if we block.
2434 */
c695044a 2435 KKASSERT((newpte & PG_MANAGED) == 0 || (m->flags & PG_MAPPED));
b12defdc
MD
2436 if ((prot & VM_PROT_NOSYNC) == 0)
2437 pmap_inval_done(&info);
2bb9cc6f
MD
2438 if (pv)
2439 free_pv_entry(pv);
4107b0c0 2440 lwkt_reltoken(&vm_token);
b12defdc 2441 vm_object_drop(pmap->pm_pteobj);
984263bc
MD
2442}
2443
2444/*
17cde63e
MD
2445 * This code works like pmap_enter() but assumes VM_PROT_READ and not-wired.
2446 * This code also assumes that the pmap has no pre-existing entry for this
2447 * VA.
2448 *
2449 * This code currently may only be used on user pmaps, not kernel_pmap.
4107b0c0
MD
2450 *
2451 * No requirements.
984263bc 2452 */
1b9d3514 2453void
17cde63e 2454pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m)
984263bc
MD
2455{
2456 unsigned *pte;
6ef943a3 2457 vm_paddr_t pa;
17cde63e
MD
2458 vm_page_t mpte;
2459 unsigned ptepindex;
2460 vm_offset_t ptepa;
0f7a3396 2461 pmap_inval_info info;
2bb9cc6f 2462 pv_entry_t pv;
0f7a3396 2463
b12defdc 2464 vm_object_hold(pmap->pm_pteobj);
4107b0c0 2465 lwkt_gettoken(&vm_token);
2bb9cc6f
MD
2466
2467 /*
2468 * This can block, get it before we do anything important.
2469 */
2470 if (pmap_initialized &&
2471 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2472 pv = get_pv_entry();
2473 } else {
2474 pv = NULL;
2475 }
2476
0f7a3396 2477 pmap_inval_init(&info);
984263bc 2478
fbbaeba3
MD
2479 if (va < UPT_MAX_ADDRESS && pmap == &kernel_pmap) {
2480 kprintf("Warning: pmap_enter_quick called on UVA with kernel_pmap\n");
7ce2998e 2481 print_backtrace(-1);
fbbaeba3
MD
2482 }
2483 if (va >= UPT_MAX_ADDRESS && pmap != &kernel_pmap) {
2484 kprintf("Warning: pmap_enter_quick called on KVA without kernel_pmap\n");
7ce2998e 2485 print_backtrace(-1);
fbbaeba3
MD
2486 }
2487
17cde63e
MD
2488 KKASSERT(va < UPT_MIN_ADDRESS); /* assert used on user pmaps only */
2489
984263bc 2490 /*
17cde63e
MD
2491 * Calculate the page table page (mpte), allocating it if necessary.
2492 *
2493 * A held page table page (mpte), or NULL, is passed onto the
2494 * section following.
984263bc
MD
2495 */
2496 if (va < UPT_MIN_ADDRESS) {
984263bc
MD
2497 /*
2498 * Calculate pagetable page index
2499 */
2500 ptepindex = va >> PDRSHIFT;
17cde63e
MD
2501
2502 do {
984263bc
MD
2503 /*
2504 * Get the page directory entry
2505 */
2506 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
2507
2508 /*
2509 * If the page table page is mapped, we just increment
90244566 2510 * the wire count, and activate it.
984263bc
MD
2511 */
2512 if (ptepa) {
2513 if (ptepa & PG_PS)
2514 panic("pmap_enter_quick: unexpected mapping into 4MB page");
b1482674
MD
2515 if ((mpte = pmap->pm_ptphint) != NULL &&
2516 (mpte->pindex == ptepindex) &&
2517 (mpte->flags & PG_BUSY) == 0) {
2bb9cc6f 2518 vm_page_wire_quick(mpte);
984263bc 2519 } else {
2bb9cc6f
MD
2520 mpte = pmap_page_lookup(pmap->pm_pteobj,
2521 ptepindex);
984263bc 2522 pmap->pm_ptphint = mpte;
2bb9cc6f 2523 vm_page_wire_quick(mpte);
b12defdc 2524 vm_page_wakeup(mpte);
984263bc 2525 }
984263bc
MD
2526 } else {
2527 mpte = _pmap_allocpte(pmap, ptepindex);
2528 }
17cde63e 2529 } while (mpte == NULL);
984263bc
MD
2530 } else {
2531 mpte = NULL;
17cde63e 2532 /* this code path is not yet used */
984263bc
MD
2533 }
2534
2535 /*
17cde63e
MD
2536 * With a valid (and held) page directory page, we can just use
2537 * vtopte() to get to the pte. If the pte is already present
2538 * we do not disturb it.
984263bc
MD
2539 */
2540 pte = (unsigned *)vtopte(va);
554cf9ac
MD
2541 if (*pte) {
2542 KKASSERT(*pte & PG_V);
17cde63e
MD
2543 pa = VM_PAGE_TO_PHYS(m);
2544 KKASSERT(((*pte ^ pa) & PG_FRAME) == 0);
c2fb025d 2545 pmap_inval_done(&info);
b1482674
MD
2546 if (mpte)
2547 pmap_unwire_pte(pmap, mpte, &info);
b1482674 2548 if (pv) {
2bb9cc6f 2549 free_pv_entry(pv);
b1482674
MD
2550 /* pv = NULL; */
2551 }
554cf9ac
MD
2552 lwkt_reltoken(&vm_token);
2553 vm_object_drop(pmap->pm_pteobj);
17cde63e 2554 return;
984263bc
MD
2555 }
2556
2557 /*
17cde63e 2558 * Enter on the PV list if part of our managed memory
984263bc 2559 */
2bb9cc6f
MD
2560 if (pmap_initialized &&
2561 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2562 pmap_insert_entry(pmap, pv, va, mpte, m);
2563 pv = NULL;
17cde63e
MD
2564 vm_page_flag_set(m, PG_MAPPED);
2565 }
984263bc
MD
2566
2567 /*
2568 * Increment counters
2569 */
eec2b734 2570 ++pmap->pm_stats.resident_count;
984263bc
MD
2571
2572 pa = VM_PAGE_TO_PHYS(m);
2573
2574 /*
2575 * Now validate mapping with RO protection
2576 */
2577 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2578 *pte = pa | PG_V | PG_U;
2579 else
2580 *pte = pa | PG_V | PG_U | PG_MANAGED;
17cde63e 2581/* pmap_inval_add(&info, pmap, va); shouldn't be needed inval->valid */
c2fb025d 2582 pmap_inval_done(&info);
b1482674 2583 if (pv) {
2bb9cc6f 2584 free_pv_entry(pv);
b1482674
MD
2585 /* pv = NULL; */
2586 }
4107b0c0 2587 lwkt_reltoken(&vm_token);
b12defdc 2588 vm_object_drop(pmap->pm_pteobj);
984263bc
MD
2589}
2590
2591/*
2592 * Make a temporary mapping for a physical address. This is only intended
2593 * to be used for panic dumps.
4107b0c0 2594 *
fb8345e6
MD
2595 * The caller is responsible for calling smp_invltlb().
2596 *
4107b0c0 2597 * No requirements.
984263bc
MD
2598 */
2599void *
8e5ea5f7 2600pmap_kenter_temporary(vm_paddr_t pa, long i)
984263bc 2601{
fb8345e6 2602 pmap_kenter_quick((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
984263bc
MD
2603 return ((void *)crashdumpmap);
2604}
2605
2606#define MAX_INIT_PT (96)
06ecca5a 2607
984263bc 2608/*
06ecca5a
MD
2609 * This routine preloads the ptes for a given object into the specified pmap.
2610 * This eliminates the blast of soft faults on process startup and
2611 * immediately after an mmap.
4107b0c0
MD
2612 *
2613 * No requirements.
984263bc 2614 */
1f804340
MD
2615static int pmap_object_init_pt_callback(vm_page_t p, void *data);
2616
984263bc 2617void
083a7402
MD
2618pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_prot_t prot,
2619 vm_object_t object, vm_pindex_t pindex,
2620 vm_size_t size, int limit)
984263bc 2621{
1f804340 2622 struct rb_vm_page_scan_info info;
287ebb09 2623 struct lwp *lp;
984263bc 2624 int psize;
984263bc 2625
54a764e8
MD
2626 /*
2627 * We can't preinit if read access isn't set or there is no pmap
2628 * or object.
2629 */
083a7402 2630 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
984263bc
MD
2631 return;
2632
54a764e8
MD
2633 /*
2634 * We can't preinit if the pmap is not the current pmap
2635 */
287ebb09
MD
2636 lp = curthread->td_lwp;
2637 if (lp == NULL || pmap != vmspace_pmap(lp->lwp_vmspace))
54a764e8
MD
2638 return;
2639
984263bc
MD
2640 psize = i386_btop(size);
2641
2642 if ((object->type != OBJT_VNODE) ||
2643 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2644 (object->resident_page_count > MAX_INIT_PT))) {
2645 return;
2646 }
2647
2648 if (psize + pindex > object->size) {
2649 if (object->size < pindex)
2650 return;
2651 psize = object->size - pindex;
2652 }
2653
1f804340
MD
2654 if (psize == 0)
2655 return;
06ecca5a 2656
984263bc 2657 /*
1f804340
MD
2658 * Use a red-black scan to traverse the requested range and load
2659 * any valid pages found into the pmap.
06ecca5a 2660 *
9acd5bbb
MD
2661 * We cannot safely scan the object's memq unless we are in a
2662 * critical section since interrupts can remove pages from objects.
984263bc 2663 */
1f804340
MD
2664 info.start_pindex = pindex;
2665 info.end_pindex = pindex + psize - 1;
2666 info.limit = limit;
2667 info.mpte = NULL;
2668 info.addr = addr;
2669 info.pmap = pmap;
2670
2f2d9e58 2671 vm_object_hold(object);
1f804340
MD
2672 vm_page_rb_tree_RB_SCAN(&object->rb_memq, rb_vm_page_scancmp,
2673 pmap_object_init_pt_callback, &info);
2f2d9e58 2674 vm_object_drop(object);
1f804340 2675}
06ecca5a 2676
4107b0c0
MD
2677/*
2678 * The caller must hold vm_token.
2679 */
1f804340
MD
2680static
2681int
2682pmap_object_init_pt_callback(vm_page_t p, void *data)
2683{
2684 struct rb_vm_page_scan_info *info = data;
2685 vm_pindex_t rel_index;
2686 /*
2687 * don't allow an madvise to blow away our really
2688 * free pages allocating pv entries.
2689 */
2690 if ((info->limit & MAP_PREFAULT_MADVISE) &&
2691 vmstats.v_free_count < vmstats.v_free_reserved) {
2692 return(-1);
984263bc 2693 }
0d987a03
MD
2694
2695 /*
2696 * Ignore list markers and ignore pages we cannot instantly
2697 * busy (while holding the object token).
2698 */
2699 if (p->flags & PG_MARKER)
2700 return 0;
b12defdc
MD
2701 if (vm_page_busy_try(p, TRUE))
2702 return 0;
1f804340 2703 if (((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
b12defdc 2704 (p->flags & PG_FICTITIOUS) == 0) {
1f804340
MD
2705 if ((p->queue - p->pc) == PQ_CACHE)
2706 vm_page_deactivate(p);
1f804340 2707 rel_index = p->pindex - info->start_pindex;
17cde63e
MD
2708 pmap_enter_quick(info->pmap,
2709 info->addr + i386_ptob(rel_index), p);
1f804340 2710 }
b12defdc 2711 vm_page_wakeup(p);
1f804340 2712 return(0);
984263bc
MD
2713}
2714
2715/*
1b9d3514
MD
2716 * Return TRUE if the pmap is in shape to trivially
2717 * pre-fault the specified address.
2718 *
2719 * Returns FALSE if it would be non-trivial or if a
2720 * pte is already loaded into the slot.
4107b0c0
MD
2721 *
2722 * No requirements.
984263bc 2723 */
1b9d3514
MD
2724int
2725pmap_prefault_ok(pmap_t pmap, vm_offset_t addr)
984263bc 2726{
1b9d3514 2727 unsigned *pte;
4107b0c0 2728 int ret;
984263bc 2729
4107b0c0
MD
2730 lwkt_gettoken(&vm_token);
2731 if ((*pmap_pde(pmap, addr)) == 0) {
2732 ret = 0;
2733 } else {
2734 pte = (unsigned *) vtopte(addr);
2735 ret = (*pte) ? 0 : 1;
2736 }
2737 lwkt_reltoken(&vm_token);
2738 return(ret);
984263bc
MD
2739}
2740
2741/*
4107b0c0
MD
2742 * Change the wiring attribute for a map/virtual-adderss pair. The mapping
2743 * must already exist.
2744 *
2745 * No requirements.
984263bc
MD
2746 */
2747void
921c891e
MD
2748pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired,
2749 vm_map_entry_t entry __unused)
984263bc 2750{
840de426 2751 unsigned *pte;
984263bc
MD
2752
2753 if (pmap == NULL)
2754 return;
2755
4107b0c0 2756 lwkt_gettoken(&vm_token);
984263bc
MD
2757 pte = pmap_pte(pmap, va);
2758
2759 if (wired && !pmap_pte_w(pte))
2760 pmap->pm_stats.wired_count++;
2761 else if (!wired && pmap_pte_w(pte))
2762 pmap->pm_stats.wired_count--;
2763
2764 /*
2765 * Wiring is not a hardware characteristic so there is no need to
0f7a3396
MD
2766 * invalidate TLB. However, in an SMP environment we must use
2767 * a locked bus cycle to update the pte (if we are not using
2768 * the pmap_inval_*() API that is)... it's ok to do this for simple
2769 * wiring changes.
984263bc 2770 */
0f7a3396
MD
2771#ifdef SMP
2772 if (wired)
2773 atomic_set_int(pte, PG_W);
2774 else
2775 atomic_clear_int(pte, PG_W);
2776#else
2777 if (wired)
2778 atomic_set_int_nonlocked(pte, PG_W);
2779 else
2780 atomic_clear_int_nonlocked(pte, PG_W);
2781#endif
4107b0c0 2782 lwkt_reltoken(&vm_token);
984263bc
MD
2783}
2784
984263bc 2785/*
4107b0c0
MD
2786 * Copy the range specified by src_addr/len from the source map to the
2787 * range dst_addr/len in the destination map.
2788 *
2789 * This routine is only advisory and need not do anything.
984263bc 2790 *
4107b0c0 2791 * No requirements.
984263bc 2792 */
984263bc 2793void
840de426 2794pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
4107b0c0 2795 vm_size_t len, vm_offset_t src_addr)
984263bc 2796{
4107b0c0 2797 /* does nothing */
984263bc
MD
2798}
2799
2800/*
4107b0c0
MD
2801 * Zero the specified PA by mapping the page into KVM and clearing its
2802 * contents.
e0e69b7d 2803 *
4107b0c0 2804 * No requirements.
984263bc
MD
2805 */
2806void
6ef943a3 2807pmap_zero_page(vm_paddr_t phys)
984263bc 2808{
85100692 2809 struct mdglobaldata *gd = mdcpu;
17a9f566 2810
e0e69b7d 2811 crit_enter();
85100692
MD
2812 if (*(int *)gd->gd_CMAP3)
2813 panic("pmap_zero_page: CMAP3 busy");
85100692 2814 *(int *)gd->gd_CMAP3 =
17a9f566 2815 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
85100692 2816 cpu_invlpg(gd->gd_CADDR3);
1fa15583 2817 bzero(gd->gd_CADDR3, PAGE_SIZE);
85100692 2818 *(int *) gd->gd_CMAP3 = 0;
e0e69b7d 2819 crit_exit();
8100156a
MD
2820}
2821
2822/*
4107b0c0 2823 * Assert that a page is empty, panic if it isn't.
8100156a 2824 *
4107b0c0 2825 * No requirements.
8100156a
MD
2826 */
2827void
2828pmap_page_assertzero(vm_paddr_t phys)
2829{
2830 struct mdglobaldata *gd = mdcpu;
2831 int i;
2832
2833 crit_enter();
2834 if (*(int *)gd->gd_CMAP3)
2835 panic("pmap_zero_page: CMAP3 busy");
2836 *(int *)gd->gd_CMAP3 =
2837 PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2838 cpu_invlpg(gd->gd_CADDR3);
2839 for (i = 0; i < PAGE_SIZE; i += 4) {
2840 if (*(int *)((char *)gd->gd_CADDR3 + i) != 0) {
ed20d0e3 2841 panic("pmap_page_assertzero() @ %p not zero!",
8100156a
MD
2842 (void *)gd->gd_CADDR3);
2843 }
2844 }
2845 *(int *) gd->gd_CMAP3 = 0;
2846 crit_exit();
984263bc
MD
2847}
2848
2849/*
4107b0c0
MD
2850 * Zero part of a physical page by mapping it into memory and clearing
2851 * its contents with bzero.
e0e69b7d 2852 *
4107b0c0 2853 * off and size may not cover an area beyond a single hardware page.
984263bc 2854 *
4107b0c0 2855 * No requirements.
984263bc
MD
2856 */
2857void
6ef943a3 2858pmap_zero_page_area(vm_paddr_t phys, int off, int size)
984263bc 2859{
85100692 2860 struct mdglobaldata *gd = mdcpu;
17a9f566 2861
e0e69b7d 2862 crit_enter();
85100692
MD
2863 if (*(int *) gd->gd_CMAP3)
2864 panic("pmap_zero_page: CMAP3 busy");
85100692
MD
2865 *(int *) gd->gd_CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2866 cpu_invlpg(gd->gd_CADDR3);
1fa15583 2867 bzero((char *)gd->gd_CADDR3 + off, size);
85100692 2868 *(int *) gd->gd_CMAP3 = 0;
e0e69b7d 2869 crit_exit();
984263bc
MD
2870}
2871
2872/*
4107b0c0
MD
2873 * Copy the physical page from the source PA to the target PA.
2874 * This function may be called from an interrupt. No locking
2875 * is required.
e0e69b7d 2876 *
4107b0c0 2877 * No requirements.
984263bc
MD
2878 */
2879void
6ef943a3 2880pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
984263bc 2881{
85100692 2882 struct mdglobaldata *gd = mdcpu;
17a9f566 2883
e0e69b7d 2884 crit_enter();
85100692
MD
2885 if (*(int *) gd->gd_CMAP1)
2886 panic("pmap_copy_page: CMAP1 busy");
2887 if (*(int *) gd->gd_CMAP2)
2888 panic("pmap_copy_page: CMAP2 busy");
984263bc 2889
85100692
MD
2890 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2891 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
984263bc 2892
85100692
MD
2893 cpu_invlpg(gd->gd_CADDR1);
2894 cpu_invlpg(gd->gd_CADDR2);
984263bc 2895
85100692 2896 bcopy(gd->gd_CADDR1, gd->gd_CADDR2, PAGE_SIZE);
984263bc 2897
85100692
MD
2898 *(int *) gd->gd_CMAP1 = 0;
2899 *(int *) gd->gd_CMAP2 = 0;
e0e69b7d 2900 crit_exit();
984263bc
MD
2901}
2902
f6bf3af1 2903/*
4107b0c0
MD
2904 * Copy the physical page from the source PA to the target PA.
2905 * This function may be called from an interrupt. No locking
2906 * is required.
f6bf3af1 2907 *
4107b0c0 2908 * No requirements.
f6bf3af1
MD
2909 */
2910void
2911pmap_copy_page_frag(vm_paddr_t src, vm_paddr_t dst, size_t bytes)
2912{
2913 struct mdglobaldata *gd = mdcpu;
2914
2915 crit_enter();
2916 if (*(int *) gd->gd_CMAP1)
2917 panic("pmap_copy_page: CMAP1 busy");
2918 if (*(int *) gd->gd_CMAP2)
2919 panic("pmap_copy_page: CMAP2 busy");
2920
2921 *(int *) gd->gd_CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2922 *(int *) gd->gd_CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2923
2924 cpu_invlpg(gd->gd_CADDR1);
2925 cpu_invlpg(gd->gd_CADDR2);
2926
2927 bcopy((char *)gd->gd_CADDR1 + (src & PAGE_MASK),
2928 (char *)gd->gd_CADDR2 + (dst & PAGE_MASK),
2929 bytes);
2930
2931 *(int *) gd->gd_CMAP1 = 0;
2932 *(int *) gd->gd_CMAP2 = 0;
2933 crit_exit();
2934}
2935
984263bc
MD
2936/*
2937 * Returns true if the pmap's pv is one of the first
2938 * 16 pvs linked to from this page. This count may
2939 * be changed upwards or downwards in the future; it
2940 * is only necessary that true be returned for a small
2941 * subset of pmaps for proper page aging.
4107b0c0
MD
2942 *
2943 * No requirements.
984263bc
MD
2944 */
2945boolean_t
840de426 2946pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
984263bc
MD
2947{
2948 pv_entry_t pv;
2949 int loops = 0;
984263bc
MD
2950
2951 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2952 return FALSE;
2953
4107b0c0 2954 lwkt_gettoken(&vm_token);
984263bc
MD
2955 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2956 if (pv->pv_pmap == pmap) {
11502947 2957 lwkt_reltoken(&vm_token);
984263bc
MD
2958 return TRUE;
2959 }
2960 loops++;
2961 if (loops >= 16)
2962 break;
2963 }
4107b0c0 2964 lwkt_reltoken(&vm_token);
984263bc
MD
2965 return (FALSE);
2966}
2967
984263bc
MD
2968/*
2969 * Remove all pages from specified address space
2970 * this aids process exit speeds. Also, this code
2971 * is special cased for current process only, but
2972 * can have the more generic (and slightly slower)
2973 * mode enabled. This is much faster than pmap_remove
2974 * in the case of running down an entire address space.
4107b0c0
MD
2975 *
2976 * No requirements.
984263bc
MD
2977 */
2978void
840de426 2979pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
984263bc 2980{
287ebb09 2981 struct lwp *lp;
984263bc
MD
2982 unsigned *pte, tpte;
2983 pv_entry_t pv, npv;
984263bc 2984 vm_page_t m;
0f7a3396 2985 pmap_inval_info info;
4a22e893 2986 int iscurrentpmap;
8790d7d8 2987 int32_t save_generation;
984263bc 2988
287ebb09
MD
2989 lp = curthread->td_lwp;
2990 if (lp && pmap == vmspace_pmap(lp->lwp_vmspace))
4a22e893
MD
2991 iscurrentpmap = 1;
2992 else
2993 iscurrentpmap = 0;
984263bc 2994
b12defdc
MD
2995 if (pmap->pm_pteobj)
2996 vm_object_hold(pmap->pm_pteobj);
4107b0c0 2997 lwkt_gettoken(&vm_token);
0f7a3396 2998 pmap_inval_init(&info);
b12defdc 2999
4a22e893 3000 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
984263bc
MD
3001 if (pv->pv_va >= eva || pv->pv_va < sva) {
3002 npv = TAILQ_NEXT(pv, pv_plist);
3003 continue;
3004 }
3005
8790d7d8
MD
3006 KKASSERT(pmap == pv->pv_pmap);
3007
4a22e893
MD
3008 if (iscurrentpmap)
3009 pte = (unsigned *)vtopte(pv->pv_va);
3010 else
8790d7d8 3011 pte = pmap_pte_quick(pmap, pv->pv_va);
5926987a 3012 KKASSERT(*pte);
c2fb025d 3013 pmap_inval_interlock(&info, pmap, pv->pv_va);
984263bc 3014
4a22e893
MD
3015 /*
3016 * We cannot remove wired pages from a process' mapping
3017 * at this time
3018 */
17cde63e 3019 if (*pte & PG_W) {
c2fb025d 3020 pmap_inval_deinterlock(&info, pmap);
984263bc
MD
3021 npv = TAILQ_NEXT(pv, pv_plist);
3022 continue;
3023 }
2247fe02 3024 KKASSERT(*pte);
17cde63e 3025 tpte = loadandclear(pte);
c2fb025d 3026 pmap_inval_deinterlock(&info, pmap);
984263bc
MD
3027
3028 m = PHYS_TO_VM_PAGE(tpte);
5926987a 3029 test_m_maps_pv(m, pv);
984263bc
MD
3030
3031 KASSERT(m < &vm_page_array[vm_page_array_size],
3032 ("pmap_remove_pages: bad tpte %x", tpte));
3033
eec2b734
MD
3034 KKASSERT(pmap->pm_stats.resident_count > 0);
3035 --pmap->pm_stats.resident_count;
984263bc
MD
3036
3037 /*
3038 * Update the vm_page_t clean and reference bits.
3039 */
3040 if (tpte & PG_M) {
3041 vm_page_dirty(m);
3042 }
3043
984263bc 3044 npv = TAILQ_NEXT(pv, pv_plist);
5926987a
MD
3045#ifdef PMAP_DEBUG
3046 KKASSERT(pv->pv_m == m);
3047 KKASSERT(pv->pv_pmap == pmap);
3048#endif
8790d7d8
MD
3049 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
3050 save_generation = ++pmap->pm_generation;
984263bc
MD
3051
3052 m->md.pv_list_count--;
cef01e15
MD
3053 if (m->object)
3054 atomic_add_int(&m->object->agg_pv_list_count, -1);
984263bc 3055 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
17cde63e 3056 if (TAILQ_EMPTY(&m->md.pv_list))
984263bc 3057 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
984263bc 3058
8790d7d8 3059 pmap_unuse_pt(pmap, pv->pv_va, pv->pv_ptem, &info);
984263bc 3060 free_pv_entry(pv);
8790d7d8
MD
3061
3062 /*
3063 * Restart the scan if we blocked during the unuse or free
3064 * calls and other removals were made.
3065 */
3066 if (save_generation != pmap->pm_generation) {
3067 kprintf("Warning: pmap_remove_pages race-A avoided\n");
5926987a 3068 npv = TAILQ_FIRST(&pmap->pm_pvlist);
8790d7d8 3069 }
984263bc 3070 }
c2fb025d 3071 pmap_inval_done(&info);
4107b0c0 3072 lwkt_reltoken(&vm_token);
b12defdc
MD
3073 if (pmap->pm_pteobj)
3074 vm_object_drop(pmap->pm_pteobj);
984263bc
MD
3075}
3076
3077/*
3078 * pmap_testbit tests bits in pte's
5e8d0349 3079 * note that the testbit/clearbit routines are inline,
984263bc 3080 * and a lot of things compile-time evaluate.
4107b0c0
MD
3081 *
3082 * The caller must hold vm_token.
984263bc
MD
3083 */
3084static boolean_t
840de426 3085pmap_testbit(vm_page_t m, int bit)
984263bc
MD
3086{
3087 pv_entry_t pv;
3088 unsigned *pte;
984263bc
MD
3089
3090 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3091 return FALSE;
3092
3093 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3094 return FALSE;
3095
984263bc
MD
3096 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3097 /*
3098 * if the bit being tested is the modified bit, then
3099 * mark clean_map and ptes as never
3100 * modified.
3101 */
3102 if (bit & (PG_A|PG_M)) {
3103 if (!pmap_track_modified(pv->pv_va))
3104 continue;
3105 }
3106
3107#if defined(PMAP_DIAGNOSTIC)
3108 if (!pv->pv_pmap) {
d557216f
MD
3109 kprintf("Null pmap (tb) at va: %p\n",
3110 (void *)pv->pv_va);
984263bc
MD
3111 continue;
3112 }
3113#endif
3114 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
74b9d1ec 3115 if (*pte & bit) {
984263bc 3116 return TRUE;
74b9d1ec 3117 }
984263bc 3118 }
984263bc
MD
3119 return (FALSE);
3120}
3121
3122/*
4107b0c0
MD
3123 * This routine is used to modify bits in ptes
3124 *
3125 * The caller must hold vm_token.
984263bc
MD
3126 */
3127static __inline void
5e8d0349 3128pmap_clearbit(vm_page_t m, int bit)
984263bc 3129{
0f7a3396 3130 struct pmap_inval_info info;
840de426
MD
3131 pv_entry_t pv;
3132 unsigned *pte;
5e8d0349 3133 unsigned pbits;
984263bc
MD
3134
3135 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3136 return;
3137
0f7a3396 3138 pmap_inval_init(&info);
984263bc
MD
3139
3140 /*
3141 * Loop over all current mappings setting/clearing as appropos If
3142 * setting RO do we need to clear the VAC?
3143 */
3144 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3145 /*
3146 * don't write protect pager mappings
3147 */
5e8d0349 3148 if (bit == PG_RW) {
984263bc
MD
3149 if (!pmap_track_modified(pv->pv_va))
3150 continue;
3151 }
3152
3153#if defined(PMAP_DIAGNOSTIC)
3154 if (!pv->pv_pmap) {
d557216f
MD
3155 kprintf("Null pmap (cb) at va: %p\n",
3156 (void *)pv->pv_va);
984263bc
MD
3157 continue;
3158 }
3159#endif
3160
0f7a3396
MD
3161 /*
3162 * Careful here. We can use a locked bus instruction to
3163 * clear PG_A or PG_M safely but we need to synchronize
3164 * with the target cpus when we mess with PG_RW.
70fc5283
MD
3165 *
3166 * We do not have to force synchronization when clearing
3167 * PG_M even for PTEs generated via virtual memory maps,
3168 * because the virtual kernel will invalidate the pmap
3169 * entry when/if it needs to resynchronize the Modify bit.
0f7a3396 3170 */
70fc5283 3171 if (bit & PG_RW)
c2fb025d 3172 pmap_inval_interlock(&info, pv->pv_pmap, pv->pv_va);
17cde63e
MD
3173 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3174again:
5e8d0349
MD
3175 pbits = *pte;
3176 if (pbits & bit) {
3177 if (bit == PG_RW) {
17cde63e 3178 if (pbits & PG_M) {
5e8d0349 3179 vm_page_dirty(m);
17cde63e
MD
3180 atomic_clear_int(pte, PG_M|PG_RW);
3181 } else {
3182 /*
3183 * The cpu may be trying to set PG_M
3184 * simultaniously with our clearing
3185 * of PG_RW.
3186 */
3187 if (!atomic_cmpset_int(pte, pbits,
3188 pbits &