FreeBSD/Linux Kernel Cross Reference
sys/i386/i386/pmap.c
1 /*
2 * Copyright (c) 1991 Regents of the University of California.
3 * All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
5 * All rights reserved.
6 * Copyright (c) 1994 David Greenman
7 * All rights reserved.
8 *
9 * This code is derived from software contributed to Berkeley by
10 * the Systems Programming Group of the University of Utah Computer
11 * Science Department and William Jolitz of UUNET Technologies Inc.
12 *
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
15 * are met:
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. All advertising materials mentioning features or use of this software
22 * must display the following acknowledgement:
23 * This product includes software developed by the University of
24 * California, Berkeley and its contributors.
25 * 4. Neither the name of the University nor the names of its contributors
26 * may be used to endorse or promote products derived from this software
27 * without specific prior written permission.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 * SUCH DAMAGE.
40 *
41 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
42 * $FreeBSD: releng/5.0/sys/i386/i386/pmap.c 107882 2002-12-14 21:08:30Z alc $
43 */
44
45 /*
46 * Manages physical address maps.
47 *
48 * In addition to hardware address maps, this
49 * module is called upon to provide software-use-only
50 * maps which may or may not be stored in the same
51 * form as hardware maps. These pseudo-maps are
52 * used to store intermediate results from copy
53 * operations to and from address spaces.
54 *
55 * Since the information managed by this module is
56 * also stored by the logical address mapping module,
57 * this module may throw away valid virtual-to-physical
58 * mappings at almost any time. However, invalidations
59 * of virtual-to-physical mappings must be done as
60 * requested.
61 *
62 * In order to cope with hardware architectures which
63 * make virtual-to-physical map invalidates expensive,
64 * this module may delay invalidate or reduced protection
65 * operations until such time as they are actually
66 * necessary. This module is given full information as
67 * to which processors are currently using which maps,
68 * and to when physical maps must be made correct.
69 */
70
71 #include "opt_pmap.h"
72 #include "opt_msgbuf.h"
73 #include "opt_kstack_pages.h"
74
75 #include <sys/param.h>
76 #include <sys/systm.h>
77 #include <sys/kernel.h>
78 #include <sys/lock.h>
79 #include <sys/mman.h>
80 #include <sys/msgbuf.h>
81 #include <sys/mutex.h>
82 #include <sys/proc.h>
83 #include <sys/sx.h>
84 #include <sys/user.h>
85 #include <sys/vmmeter.h>
86 #include <sys/sysctl.h>
87 #ifdef SMP
88 #include <sys/smp.h>
89 #endif
90
91 #include <vm/vm.h>
92 #include <vm/vm_param.h>
93 #include <vm/vm_kern.h>
94 #include <vm/vm_page.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_object.h>
97 #include <vm/vm_extern.h>
98 #include <vm/vm_pageout.h>
99 #include <vm/vm_pager.h>
100 #include <vm/uma.h>
101
102 #include <machine/cpu.h>
103 #include <machine/cputypes.h>
104 #include <machine/md_var.h>
105 #include <machine/specialreg.h>
106 #if defined(SMP) || defined(APIC_IO)
107 #include <machine/smp.h>
108 #include <machine/apic.h>
109 #include <machine/segments.h>
110 #include <machine/tss.h>
111 #endif /* SMP || APIC_IO */
112
113 #define PMAP_KEEP_PDIRS
114 #ifndef PMAP_SHPGPERPROC
115 #define PMAP_SHPGPERPROC 200
116 #endif
117
118 #if defined(DIAGNOSTIC)
119 #define PMAP_DIAGNOSTIC
120 #endif
121
122 #define MINPV 2048
123
124 #if !defined(PMAP_DIAGNOSTIC)
125 #define PMAP_INLINE __inline
126 #else
127 #define PMAP_INLINE
128 #endif
129
130 /*
131 * Get PDEs and PTEs for user/kernel address space
132 */
133 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
134 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
135
136 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
137 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
138 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
139 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
140 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
141
142 #define pmap_pte_set_w(pte, v) ((v)?(*(int *)pte |= PG_W):(*(int *)pte &= ~PG_W))
143 #define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v)))
144
145 /*
146 * Given a map and a machine independent protection code,
147 * convert to a vax protection code.
148 */
149 #define pte_prot(m, p) (protection_codes[p])
150 static int protection_codes[8];
151
152 struct pmap kernel_pmap_store;
153 LIST_HEAD(pmaplist, pmap);
154 struct pmaplist allpmaps;
155
156 vm_offset_t avail_start; /* PA of first available physical page */
157 vm_offset_t avail_end; /* PA of last available physical page */
158 vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */
159 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
160 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
161 static int pgeflag; /* PG_G or-in */
162 static int pseflag; /* PG_PS or-in */
163
164 static vm_object_t kptobj;
165
166 static int nkpt;
167 vm_offset_t kernel_vm_end;
168 extern u_int32_t KERNend;
169
170 /*
171 * Data for the pv entry allocation mechanism
172 */
173 static uma_zone_t pvzone;
174 static struct vm_object pvzone_obj;
175 static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0;
176 int pmap_pagedaemon_waken;
177
178 /*
179 * All those kernel PT submaps that BSD is so fond of
180 */
181 pt_entry_t *CMAP1 = 0;
182 static pt_entry_t *CMAP2, *CMAP3, *ptmmap;
183 caddr_t CADDR1 = 0, ptvmmap = 0;
184 static caddr_t CADDR2, CADDR3;
185 static pt_entry_t *msgbufmap;
186 struct msgbuf *msgbufp = 0;
187
188 /*
189 * Crashdump maps.
190 */
191 static pt_entry_t *pt_crashdumpmap;
192 static caddr_t crashdumpmap;
193
194 #ifdef SMP
195 extern pt_entry_t *SMPpt;
196 #endif
197 static pt_entry_t *PMAP1 = 0;
198 static pt_entry_t *PADDR1 = 0;
199
200 static PMAP_INLINE void free_pv_entry(pv_entry_t pv);
201 static pt_entry_t *get_ptbase(pmap_t pmap);
202 static pv_entry_t get_pv_entry(void);
203 static void i386_protection_init(void);
204 static __inline void pmap_changebit(vm_page_t m, int bit, boolean_t setem);
205
206 static vm_page_t pmap_enter_quick(pmap_t pmap, vm_offset_t va,
207 vm_page_t m, vm_page_t mpte);
208 static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva);
209 static void pmap_remove_page(struct pmap *pmap, vm_offset_t va);
210 static int pmap_remove_entry(struct pmap *pmap, vm_page_t m,
211 vm_offset_t va);
212 static boolean_t pmap_testbit(vm_page_t m, int bit);
213 static void pmap_insert_entry(pmap_t pmap, vm_offset_t va,
214 vm_page_t mpte, vm_page_t m);
215
216 static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va);
217
218 static int pmap_release_free_page(pmap_t pmap, vm_page_t p);
219 static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex);
220 static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va);
221 static vm_page_t pmap_page_lookup(vm_object_t object, vm_pindex_t pindex);
222 static int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t);
223 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
224 static void *pmap_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait);
225
226 static pd_entry_t pdir4mb;
227
228 /*
229 * Routine: pmap_pte
230 * Function:
231 * Extract the page table entry associated
232 * with the given map/virtual_address pair.
233 */
234
235 PMAP_INLINE pt_entry_t *
236 pmap_pte(pmap, va)
237 register pmap_t pmap;
238 vm_offset_t va;
239 {
240 pd_entry_t *pdeaddr;
241
242 if (pmap) {
243 pdeaddr = pmap_pde(pmap, va);
244 if (*pdeaddr & PG_PS)
245 return pdeaddr;
246 if (*pdeaddr) {
247 return get_ptbase(pmap) + i386_btop(va);
248 }
249 }
250 return (0);
251 }
252
253 /*
254 * Move the kernel virtual free pointer to the next
255 * 4MB. This is used to help improve performance
256 * by using a large (4MB) page for much of the kernel
257 * (.text, .data, .bss)
258 */
259 static vm_offset_t
260 pmap_kmem_choose(vm_offset_t addr)
261 {
262 vm_offset_t newaddr = addr;
263
264 #ifndef DISABLE_PSE
265 if (cpu_feature & CPUID_PSE)
266 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
267 #endif
268 return newaddr;
269 }
270
271 /*
272 * Bootstrap the system enough to run with virtual memory.
273 *
274 * On the i386 this is called after mapping has already been enabled
275 * and just syncs the pmap module with what has already been done.
276 * [We can't call it easily with mapping off since the kernel is not
277 * mapped with PA == VA, hence we would have to relocate every address
278 * from the linked base (virtual) address "KERNBASE" to the actual
279 * (physical) address starting relative to 0]
280 */
281 void
282 pmap_bootstrap(firstaddr, loadaddr)
283 vm_offset_t firstaddr;
284 vm_offset_t loadaddr;
285 {
286 vm_offset_t va;
287 pt_entry_t *pte;
288 int i;
289
290 avail_start = firstaddr;
291
292 /*
293 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too
294 * large. It should instead be correctly calculated in locore.s and
295 * not based on 'first' (which is a physical address, not a virtual
296 * address, for the start of unused physical memory). The kernel
297 * page tables are NOT double mapped and thus should not be included
298 * in this calculation.
299 */
300 virtual_avail = (vm_offset_t) KERNBASE + firstaddr;
301 virtual_avail = pmap_kmem_choose(virtual_avail);
302
303 virtual_end = VM_MAX_KERNEL_ADDRESS;
304
305 /*
306 * Initialize protection array.
307 */
308 i386_protection_init();
309
310 /*
311 * Initialize the kernel pmap (which is statically allocated).
312 */
313 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD);
314 kernel_pmap->pm_active = -1; /* don't allow deactivation */
315 TAILQ_INIT(&kernel_pmap->pm_pvlist);
316 LIST_INIT(&allpmaps);
317 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list);
318 nkpt = NKPT;
319
320 /*
321 * Reserve some special page table entries/VA space for temporary
322 * mapping of pages.
323 */
324 #define SYSMAP(c, p, v, n) \
325 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
326
327 va = virtual_avail;
328 pte = (pt_entry_t *) pmap_pte(kernel_pmap, va);
329
330 /*
331 * CMAP1/CMAP2 are used for zeroing and copying pages.
332 * CMAP3 is used for the idle process page zeroing.
333 */
334 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
335 SYSMAP(caddr_t, CMAP2, CADDR2, 1)
336 SYSMAP(caddr_t, CMAP3, CADDR3, 1)
337
338 /*
339 * Crashdump maps.
340 */
341 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
342
343 /*
344 * ptvmmap is used for reading arbitrary physical pages via /dev/mem.
345 * XXX ptmmap is not used.
346 */
347 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
348
349 /*
350 * msgbufp is used to map the system message buffer.
351 * XXX msgbufmap is not used.
352 */
353 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
354 atop(round_page(MSGBUF_SIZE)))
355
356 /*
357 * ptemap is used for pmap_pte_quick
358 */
359 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1);
360
361 virtual_avail = va;
362
363 *CMAP1 = *CMAP2 = 0;
364 for (i = 0; i < NKPT; i++)
365 PTD[i] = 0;
366
367 pgeflag = 0;
368 #ifndef DISABLE_PG_G
369 if (cpu_feature & CPUID_PGE)
370 pgeflag = PG_G;
371 #endif
372
373 /*
374 * Initialize the 4MB page size flag
375 */
376 pseflag = 0;
377 /*
378 * The 4MB page version of the initial
379 * kernel page mapping.
380 */
381 pdir4mb = 0;
382
383 #ifndef DISABLE_PSE
384 if (cpu_feature & CPUID_PSE) {
385 pd_entry_t ptditmp;
386 /*
387 * Note that we have enabled PSE mode
388 */
389 pseflag = PG_PS;
390 ptditmp = *(PTmap + i386_btop(KERNBASE));
391 ptditmp &= ~(NBPDR - 1);
392 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
393 pdir4mb = ptditmp;
394 }
395 #endif
396 #ifndef SMP
397 /*
398 * Turn on PGE/PSE. SMP does this later on since the
399 * 4K page tables are required for AP boot (for now).
400 * XXX fixme.
401 */
402 pmap_set_opt();
403 #endif
404 #ifdef SMP
405 if (cpu_apic_address == 0)
406 panic("pmap_bootstrap: no local apic! (non-SMP hardware?)");
407
408 /* local apic is mapped on last page */
409 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N | pgeflag |
410 (cpu_apic_address & PG_FRAME));
411 #endif
412 invltlb();
413 }
414
415 /*
416 * Enable 4MB page mode for MP startup. Turn on PG_G support.
417 * BSP will run this after all the AP's have started up.
418 */
419 void
420 pmap_set_opt(void)
421 {
422 pt_entry_t *pte;
423 vm_offset_t va, endva;
424
425 if (pgeflag && (cpu_feature & CPUID_PGE)) {
426 load_cr4(rcr4() | CR4_PGE);
427 invltlb(); /* Insurance */
428 }
429 #ifndef DISABLE_PSE
430 if (pseflag && (cpu_feature & CPUID_PSE)) {
431 load_cr4(rcr4() | CR4_PSE);
432 invltlb(); /* Insurance */
433 }
434 #endif
435 if (PCPU_GET(cpuid) == 0) {
436 #ifndef DISABLE_PSE
437 if (pdir4mb) {
438 kernel_pmap->pm_pdir[KPTDI] = PTD[KPTDI] = pdir4mb;
439 invltlb(); /* Insurance */
440 }
441 #endif
442 if (pgeflag) {
443 /* Turn on PG_G for text, data, bss pages. */
444 va = (vm_offset_t)btext;
445 #ifndef DISABLE_PSE
446 if (pseflag && (cpu_feature & CPUID_PSE)) {
447 if (va < KERNBASE + (1 << PDRSHIFT))
448 va = KERNBASE + (1 << PDRSHIFT);
449 }
450 #endif
451 endva = KERNBASE + KERNend;
452 while (va < endva) {
453 pte = vtopte(va);
454 if (*pte)
455 *pte |= pgeflag;
456 va += PAGE_SIZE;
457 }
458 invltlb(); /* Insurance */
459 }
460 /*
461 * We do not need to broadcast the invltlb here, because
462 * each AP does it the moment it is released from the boot
463 * lock. See ap_init().
464 */
465 }
466 }
467
468 static void *
469 pmap_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
470 {
471 *flags = UMA_SLAB_PRIV;
472 return (void *)kmem_alloc(kernel_map, bytes);
473 }
474
475 /*
476 * Initialize the pmap module.
477 * Called by vm_init, to initialize any structures that the pmap
478 * system needs to map virtual memory.
479 * pmap_init has been enhanced to support in a fairly consistant
480 * way, discontiguous physical memory.
481 */
482 void
483 pmap_init(phys_start, phys_end)
484 vm_offset_t phys_start, phys_end;
485 {
486 int i;
487 int initial_pvs;
488
489 /*
490 * object for kernel page table pages
491 */
492 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
493
494 /*
495 * Allocate memory for random pmap data structures. Includes the
496 * pv_head_table.
497 */
498
499 for(i = 0; i < vm_page_array_size; i++) {
500 vm_page_t m;
501
502 m = &vm_page_array[i];
503 TAILQ_INIT(&m->md.pv_list);
504 m->md.pv_list_count = 0;
505 }
506
507 /*
508 * init the pv free list
509 */
510 initial_pvs = vm_page_array_size;
511 if (initial_pvs < MINPV)
512 initial_pvs = MINPV;
513 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL,
514 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM);
515 uma_zone_set_allocf(pvzone, pmap_allocf);
516 uma_prealloc(pvzone, initial_pvs);
517
518 /*
519 * Now it is safe to enable pv_table recording.
520 */
521 pmap_initialized = TRUE;
522 }
523
524 /*
525 * Initialize the address space (zone) for the pv_entries. Set a
526 * high water mark so that the system can recover from excessive
527 * numbers of pv entries.
528 */
529 void
530 pmap_init2()
531 {
532 int shpgperproc = PMAP_SHPGPERPROC;
533
534 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
535 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
536 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
537 pv_entry_high_water = 9 * (pv_entry_max / 10);
538 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max);
539 }
540
541
542 /***************************************************
543 * Low level helper routines.....
544 ***************************************************/
545
546 #if defined(PMAP_DIAGNOSTIC)
547
548 /*
549 * This code checks for non-writeable/modified pages.
550 * This should be an invalid condition.
551 */
552 static int
553 pmap_nw_modified(pt_entry_t ptea)
554 {
555 int pte;
556
557 pte = (int) ptea;
558
559 if ((pte & (PG_M|PG_RW)) == PG_M)
560 return 1;
561 else
562 return 0;
563 }
564 #endif
565
566
567 /*
568 * this routine defines the region(s) of memory that should
569 * not be tested for the modified bit.
570 */
571 static PMAP_INLINE int
572 pmap_track_modified(vm_offset_t va)
573 {
574 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva))
575 return 1;
576 else
577 return 0;
578 }
579
580 #ifdef I386_CPU
581 /*
582 * i386 only has "invalidate everything" and no SMP to worry about.
583 */
584 PMAP_INLINE void
585 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
586 {
587
588 if (pmap == kernel_pmap || pmap->pm_active)
589 invltlb();
590 }
591
592 PMAP_INLINE void
593 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
594 {
595
596 if (pmap == kernel_pmap || pmap->pm_active)
597 invltlb();
598 }
599
600 PMAP_INLINE void
601 pmap_invalidate_all(pmap_t pmap)
602 {
603
604 if (pmap == kernel_pmap || pmap->pm_active)
605 invltlb();
606 }
607 #else /* !I386_CPU */
608 #ifdef SMP
609 /*
610 * For SMP, these functions have to use the IPI mechanism for coherence.
611 */
612 void
613 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
614 {
615 u_int cpumask;
616 u_int other_cpus;
617
618 critical_enter();
619 /*
620 * We need to disable interrupt preemption but MUST NOT have
621 * interrupts disabled here.
622 * XXX we may need to hold schedlock to get a coherent pm_active
623 */
624 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) {
625 invlpg(va);
626 smp_invlpg(va);
627 } else {
628 cpumask = PCPU_GET(cpumask);
629 other_cpus = PCPU_GET(other_cpus);
630 if (pmap->pm_active & cpumask)
631 invlpg(va);
632 if (pmap->pm_active & other_cpus)
633 smp_masked_invlpg(pmap->pm_active & other_cpus, va);
634 }
635 critical_exit();
636 }
637
638 void
639 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
640 {
641 u_int cpumask;
642 u_int other_cpus;
643 vm_offset_t addr;
644
645 critical_enter();
646 /*
647 * We need to disable interrupt preemption but MUST NOT have
648 * interrupts disabled here.
649 * XXX we may need to hold schedlock to get a coherent pm_active
650 */
651 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) {
652 for (addr = sva; addr < eva; addr += PAGE_SIZE)
653 invlpg(addr);
654 smp_invlpg_range(sva, eva);
655 } else {
656 cpumask = PCPU_GET(cpumask);
657 other_cpus = PCPU_GET(other_cpus);
658 if (pmap->pm_active & cpumask)
659 for (addr = sva; addr < eva; addr += PAGE_SIZE)
660 invlpg(addr);
661 if (pmap->pm_active & other_cpus)
662 smp_masked_invlpg_range(pmap->pm_active & other_cpus,
663 sva, eva);
664 }
665 critical_exit();
666 }
667
668 void
669 pmap_invalidate_all(pmap_t pmap)
670 {
671 u_int cpumask;
672 u_int other_cpus;
673
674 #ifdef SWTCH_OPTIM_STATS
675 tlb_flush_count++;
676 #endif
677 critical_enter();
678 /*
679 * We need to disable interrupt preemption but MUST NOT have
680 * interrupts disabled here.
681 * XXX we may need to hold schedlock to get a coherent pm_active
682 */
683 if (pmap->pm_active == -1 || pmap->pm_active == all_cpus) {
684 invltlb();
685 smp_invltlb();
686 } else {
687 cpumask = PCPU_GET(cpumask);
688 other_cpus = PCPU_GET(other_cpus);
689 if (pmap->pm_active & cpumask)
690 invltlb();
691 if (pmap->pm_active & other_cpus)
692 smp_masked_invltlb(pmap->pm_active & other_cpus);
693 }
694 critical_exit();
695 }
696 #else /* !SMP */
697 /*
698 * Normal, non-SMP, 486+ invalidation functions.
699 * We inline these within pmap.c for speed.
700 */
701 PMAP_INLINE void
702 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
703 {
704
705 if (pmap == kernel_pmap || pmap->pm_active)
706 invlpg(va);
707 }
708
709 PMAP_INLINE void
710 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
711 {
712 vm_offset_t addr;
713
714 if (pmap == kernel_pmap || pmap->pm_active)
715 for (addr = sva; addr < eva; addr += PAGE_SIZE)
716 invlpg(addr);
717 }
718
719 PMAP_INLINE void
720 pmap_invalidate_all(pmap_t pmap)
721 {
722
723 if (pmap == kernel_pmap || pmap->pm_active)
724 invltlb();
725 }
726 #endif /* !SMP */
727 #endif /* !I386_CPU */
728
729 /*
730 * Return an address which is the base of the Virtual mapping of
731 * all the PTEs for the given pmap. Note this doesn't say that
732 * all the PTEs will be present or that the pages there are valid.
733 * The PTEs are made available by the recursive mapping trick.
734 * It will map in the alternate PTE space if needed.
735 */
736 static pt_entry_t *
737 get_ptbase(pmap)
738 pmap_t pmap;
739 {
740 pd_entry_t frame;
741
742 /* are we current address space or kernel? */
743 if (pmap == kernel_pmap)
744 return PTmap;
745 frame = pmap->pm_pdir[PTDPTDI] & PG_FRAME;
746 if (frame == (PTDpde & PG_FRAME))
747 return PTmap;
748 /* otherwise, we are alternate address space */
749 if (frame != (APTDpde & PG_FRAME)) {
750 APTDpde = (pd_entry_t) (frame | PG_RW | PG_V);
751 pmap_invalidate_all(kernel_pmap); /* XXX Bandaid */
752 }
753 return APTmap;
754 }
755
756 /*
757 * Super fast pmap_pte routine best used when scanning
758 * the pv lists. This eliminates many coarse-grained
759 * invltlb calls. Note that many of the pv list
760 * scans are across different pmaps. It is very wasteful
761 * to do an entire invltlb for checking a single mapping.
762 */
763
764 static pt_entry_t *
765 pmap_pte_quick(pmap, va)
766 register pmap_t pmap;
767 vm_offset_t va;
768 {
769 pd_entry_t pde, newpf;
770 pde = pmap->pm_pdir[va >> PDRSHIFT];
771 if (pde != 0) {
772 pd_entry_t frame = pmap->pm_pdir[PTDPTDI] & PG_FRAME;
773 unsigned index = i386_btop(va);
774 /* are we current address space or kernel? */
775 if (pmap == kernel_pmap || frame == (PTDpde & PG_FRAME))
776 return PTmap + index;
777 newpf = pde & PG_FRAME;
778 if (((*PMAP1) & PG_FRAME) != newpf) {
779 *PMAP1 = newpf | PG_RW | PG_V;
780 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR1);
781 }
782 return PADDR1 + (index & (NPTEPG - 1));
783 }
784 return (0);
785 }
786
787 /*
788 * Routine: pmap_extract
789 * Function:
790 * Extract the physical page address associated
791 * with the given map/virtual_address pair.
792 */
793 vm_offset_t
794 pmap_extract(pmap, va)
795 register pmap_t pmap;
796 vm_offset_t va;
797 {
798 vm_offset_t rtval; /* XXX FIXME */
799 vm_offset_t pdirindex;
800
801 if (pmap == 0)
802 return 0;
803 pdirindex = va >> PDRSHIFT;
804 rtval = pmap->pm_pdir[pdirindex];
805 if (rtval != 0) {
806 pt_entry_t *pte;
807 if ((rtval & PG_PS) != 0) {
808 rtval &= ~(NBPDR - 1);
809 rtval |= va & (NBPDR - 1);
810 return rtval;
811 }
812 pte = get_ptbase(pmap) + i386_btop(va);
813 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
814 return rtval;
815 }
816 return 0;
817
818 }
819
820 /***************************************************
821 * Low level mapping routines.....
822 ***************************************************/
823
824 /*
825 * Add a wired page to the kva.
826 * Note: not SMP coherent.
827 */
828 PMAP_INLINE void
829 pmap_kenter(vm_offset_t va, vm_offset_t pa)
830 {
831 pt_entry_t *pte;
832
833 pte = vtopte(va);
834 *pte = pa | PG_RW | PG_V | pgeflag;
835 }
836
837 /*
838 * Remove a page from the kernel pagetables.
839 * Note: not SMP coherent.
840 */
841 PMAP_INLINE void
842 pmap_kremove(vm_offset_t va)
843 {
844 pt_entry_t *pte;
845
846 pte = vtopte(va);
847 *pte = 0;
848 }
849
850 /*
851 * Used to map a range of physical addresses into kernel
852 * virtual address space.
853 *
854 * The value passed in '*virt' is a suggested virtual address for
855 * the mapping. Architectures which can support a direct-mapped
856 * physical to virtual region can return the appropriate address
857 * within that region, leaving '*virt' unchanged. Other
858 * architectures should map the pages starting at '*virt' and
859 * update '*virt' with the first usable address after the mapped
860 * region.
861 */
862 vm_offset_t
863 pmap_map(vm_offset_t *virt, vm_offset_t start, vm_offset_t end, int prot)
864 {
865 vm_offset_t va, sva;
866
867 va = sva = *virt;
868 while (start < end) {
869 pmap_kenter(va, start);
870 va += PAGE_SIZE;
871 start += PAGE_SIZE;
872 }
873 pmap_invalidate_range(kernel_pmap, sva, va);
874 *virt = va;
875 return (sva);
876 }
877
878
879 /*
880 * Add a list of wired pages to the kva
881 * this routine is only used for temporary
882 * kernel mappings that do not need to have
883 * page modification or references recorded.
884 * Note that old mappings are simply written
885 * over. The page *must* be wired.
886 * Note: SMP coherent. Uses a ranged shootdown IPI.
887 */
888 void
889 pmap_qenter(vm_offset_t sva, vm_page_t *m, int count)
890 {
891 vm_offset_t va;
892
893 va = sva;
894 while (count-- > 0) {
895 pmap_kenter(va, VM_PAGE_TO_PHYS(*m));
896 va += PAGE_SIZE;
897 m++;
898 }
899 pmap_invalidate_range(kernel_pmap, sva, va);
900 }
901
902 /*
903 * This routine tears out page mappings from the
904 * kernel -- it is meant only for temporary mappings.
905 * Note: SMP coherent. Uses a ranged shootdown IPI.
906 */
907 void
908 pmap_qremove(vm_offset_t sva, int count)
909 {
910 vm_offset_t va;
911
912 va = sva;
913 while (count-- > 0) {
914 pmap_kremove(va);
915 va += PAGE_SIZE;
916 }
917 pmap_invalidate_range(kernel_pmap, sva, va);
918 }
919
920 static vm_page_t
921 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
922 {
923 vm_page_t m;
924
925 retry:
926 m = vm_page_lookup(object, pindex);
927 if (m != NULL) {
928 vm_page_lock_queues();
929 if (vm_page_sleep_if_busy(m, FALSE, "pplookp"))
930 goto retry;
931 vm_page_unlock_queues();
932 }
933 return m;
934 }
935
936 #ifndef KSTACK_MAX_PAGES
937 #define KSTACK_MAX_PAGES 32
938 #endif
939
940 /*
941 * Create the kernel stack (including pcb for i386) for a new thread.
942 * This routine directly affects the fork perf for a process and
943 * create performance for a thread.
944 */
945 void
946 pmap_new_thread(struct thread *td, int pages)
947 {
948 int i;
949 vm_page_t ma[KSTACK_MAX_PAGES];
950 vm_object_t ksobj;
951 vm_page_t m;
952 vm_offset_t ks;
953
954 /* Bounds check */
955 if (pages <= 1)
956 pages = KSTACK_PAGES;
957 else if (pages > KSTACK_MAX_PAGES)
958 pages = KSTACK_MAX_PAGES;
959
960 /*
961 * allocate object for the kstack
962 */
963 ksobj = vm_object_allocate(OBJT_DEFAULT, pages);
964 td->td_kstack_obj = ksobj;
965
966 /* get a kernel virtual address for the kstack for this thread */
967 #ifdef KSTACK_GUARD
968 ks = kmem_alloc_nofault(kernel_map, (pages + 1) * PAGE_SIZE);
969 if (ks == 0)
970 panic("pmap_new_thread: kstack allocation failed");
971 if (*vtopte(ks) != 0)
972 pmap_qremove(ks, 1);
973 ks += PAGE_SIZE;
974 td->td_kstack = ks;
975 #else
976 /* get a kernel virtual address for the kstack for this thread */
977 ks = kmem_alloc_nofault(kernel_map, pages * PAGE_SIZE);
978 if (ks == 0)
979 panic("pmap_new_thread: kstack allocation failed");
980 td->td_kstack = ks;
981 #endif
982 /*
983 * Knowing the number of pages allocated is useful when you
984 * want to deallocate them.
985 */
986 td->td_kstack_pages = pages;
987
988 /*
989 * For the length of the stack, link in a real page of ram for each
990 * page of stack.
991 */
992 for (i = 0; i < pages; i++) {
993 /*
994 * Get a kernel stack page
995 */
996 m = vm_page_grab(ksobj, i,
997 VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED);
998 ma[i] = m;
999
1000 vm_page_wakeup(m);
1001 vm_page_flag_clear(m, PG_ZERO);
1002 m->valid = VM_PAGE_BITS_ALL;
1003 }
1004 pmap_qenter(ks, ma, pages);
1005 }
1006
1007 /*
1008 * Dispose the kernel stack for a thread that has exited.
1009 * This routine directly impacts the exit perf of a process and thread.
1010 */
1011 void
1012 pmap_dispose_thread(td)
1013 struct thread *td;
1014 {
1015 int i;
1016 int pages;
1017 vm_object_t ksobj;
1018 vm_offset_t ks;
1019 vm_page_t m;
1020
1021 pages = td->td_kstack_pages;
1022 ksobj = td->td_kstack_obj;
1023 ks = td->td_kstack;
1024 pmap_qremove(ks, pages);
1025 for (i = 0; i < pages; i++) {
1026 m = vm_page_lookup(ksobj, i);
1027 if (m == NULL)
1028 panic("pmap_dispose_thread: kstack already missing?");
1029 vm_page_lock_queues();
1030 vm_page_busy(m);
1031 vm_page_unwire(m, 0);
1032 vm_page_free(m);
1033 vm_page_unlock_queues();
1034 }
1035 /*
1036 * Free the space that this stack was mapped to in the kernel
1037 * address map.
1038 */
1039 #ifdef KSTACK_GUARD
1040 kmem_free(kernel_map, ks - PAGE_SIZE, (pages + 1) * PAGE_SIZE);
1041 #else
1042 kmem_free(kernel_map, ks, pages * PAGE_SIZE);
1043 #endif
1044 vm_object_deallocate(ksobj);
1045 }
1046
1047 /*
1048 * Set up a variable sized alternate kstack. Though it may look MI, it may
1049 * need to be different on certain arches like ia64.
1050 */
1051 void
1052 pmap_new_altkstack(struct thread *td, int pages)
1053 {
1054 /* shuffle the original stack */
1055 td->td_altkstack_obj = td->td_kstack_obj;
1056 td->td_altkstack = td->td_kstack;
1057 td->td_altkstack_pages = td->td_kstack_pages;
1058
1059 pmap_new_thread(td, pages);
1060 }
1061
1062 void
1063 pmap_dispose_altkstack(td)
1064 struct thread *td;
1065 {
1066 pmap_dispose_thread(td);
1067
1068 /* restore the original kstack */
1069 td->td_kstack = td->td_altkstack;
1070 td->td_kstack_obj = td->td_altkstack_obj;
1071 td->td_kstack_pages = td->td_altkstack_pages;
1072 td->td_altkstack = 0;
1073 td->td_altkstack_obj = NULL;
1074 td->td_altkstack_pages = 0;
1075 }
1076
1077 /*
1078 * Allow the Kernel stack for a thread to be prejudicially paged out.
1079 */
1080 void
1081 pmap_swapout_thread(td)
1082 struct thread *td;
1083 {
1084 int i;
1085 int pages;
1086 vm_object_t ksobj;
1087 vm_offset_t ks;
1088 vm_page_t m;
1089
1090 pages = td->td_kstack_pages;
1091 ksobj = td->td_kstack_obj;
1092 ks = td->td_kstack;
1093 pmap_qremove(ks, pages);
1094 for (i = 0; i < pages; i++) {
1095 m = vm_page_lookup(ksobj, i);
1096 if (m == NULL)
1097 panic("pmap_swapout_thread: kstack already missing?");
1098 vm_page_lock_queues();
1099 vm_page_dirty(m);
1100 vm_page_unwire(m, 0);
1101 vm_page_unlock_queues();
1102 }
1103 }
1104
1105 /*
1106 * Bring the kernel stack for a specified thread back in.
1107 */
1108 void
1109 pmap_swapin_thread(td)
1110 struct thread *td;
1111 {
1112 int i, rv;
1113 int pages;
1114 vm_page_t ma[KSTACK_MAX_PAGES];
1115 vm_object_t ksobj;
1116 vm_offset_t ks;
1117 vm_page_t m;
1118
1119 pages = td->td_kstack_pages;
1120 ksobj = td->td_kstack_obj;
1121 ks = td->td_kstack;
1122 for (i = 0; i < pages; i++) {
1123 m = vm_page_grab(ksobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1124 if (m->valid != VM_PAGE_BITS_ALL) {
1125 rv = vm_pager_get_pages(ksobj, &m, 1, 0);
1126 if (rv != VM_PAGER_OK)
1127 panic("pmap_swapin_thread: cannot get kstack for proc: %d\n", td->td_proc->p_pid);
1128 m = vm_page_lookup(ksobj, i);
1129 m->valid = VM_PAGE_BITS_ALL;
1130 }
1131 ma[i] = m;
1132 vm_page_lock_queues();
1133 vm_page_wire(m);
1134 vm_page_wakeup(m);
1135 vm_page_unlock_queues();
1136 }
1137 pmap_qenter(ks, ma, pages);
1138 }
1139
1140 /***************************************************
1141 * Page table page management routines.....
1142 ***************************************************/
1143
1144 /*
1145 * This routine unholds page table pages, and if the hold count
1146 * drops to zero, then it decrements the wire count.
1147 */
1148 static int
1149 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
1150 {
1151
1152 while (vm_page_sleep_if_busy(m, FALSE, "pmuwpt"))
1153 vm_page_lock_queues();
1154
1155 if (m->hold_count == 0) {
1156 vm_offset_t pteva;
1157 /*
1158 * unmap the page table page
1159 */
1160 pmap->pm_pdir[m->pindex] = 0;
1161 --pmap->pm_stats.resident_count;
1162 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) ==
1163 (PTDpde & PG_FRAME)) {
1164 /*
1165 * Do a invltlb to make the invalidated mapping
1166 * take effect immediately.
1167 */
1168 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex);
1169 pmap_invalidate_page(pmap, pteva);
1170 }
1171
1172 if (pmap->pm_ptphint == m)
1173 pmap->pm_ptphint = NULL;
1174
1175 /*
1176 * If the page is finally unwired, simply free it.
1177 */
1178 --m->wire_count;
1179 if (m->wire_count == 0) {
1180 vm_page_busy(m);
1181 vm_page_free_zero(m);
1182 --cnt.v_wire_count;
1183 }
1184 return 1;
1185 }
1186 return 0;
1187 }
1188
1189 static PMAP_INLINE int
1190 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
1191 {
1192 vm_page_unhold(m);
1193 if (m->hold_count == 0)
1194 return _pmap_unwire_pte_hold(pmap, m);
1195 else
1196 return 0;
1197 }
1198
1199 /*
1200 * After removing a page table entry, this routine is used to
1201 * conditionally free the page, and manage the hold/wire counts.
1202 */
1203 static int
1204 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
1205 {
1206 unsigned ptepindex;
1207 if (va >= VM_MAXUSER_ADDRESS)
1208 return 0;
1209
1210 if (mpte == NULL) {
1211 ptepindex = (va >> PDRSHIFT);
1212 if (pmap->pm_ptphint &&
1213 (pmap->pm_ptphint->pindex == ptepindex)) {
1214 mpte = pmap->pm_ptphint;
1215 } else {
1216 while ((mpte = vm_page_lookup(pmap->pm_pteobj, ptepindex)) != NULL &&
1217 vm_page_sleep_if_busy(mpte, FALSE, "pulook"))
1218 vm_page_lock_queues();
1219 pmap->pm_ptphint = mpte;
1220 }
1221 }
1222
1223 return pmap_unwire_pte_hold(pmap, mpte);
1224 }
1225
1226 void
1227 pmap_pinit0(pmap)
1228 struct pmap *pmap;
1229 {
1230 pmap->pm_pdir =
1231 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE);
1232 pmap_kenter((vm_offset_t)pmap->pm_pdir, (vm_offset_t)IdlePTD);
1233 #ifndef I386_CPU
1234 invlpg((vm_offset_t)pmap->pm_pdir);
1235 #else
1236 invltlb();
1237 #endif
1238 pmap->pm_ptphint = NULL;
1239 pmap->pm_active = 0;
1240 TAILQ_INIT(&pmap->pm_pvlist);
1241 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1242 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list);
1243 }
1244
1245 /*
1246 * Initialize a preallocated and zeroed pmap structure,
1247 * such as one in a vmspace structure.
1248 */
1249 void
1250 pmap_pinit(pmap)
1251 register struct pmap *pmap;
1252 {
1253 vm_page_t ptdpg;
1254
1255 /*
1256 * No need to allocate page table space yet but we do need a valid
1257 * page directory table.
1258 */
1259 if (pmap->pm_pdir == NULL)
1260 pmap->pm_pdir =
1261 (pd_entry_t *)kmem_alloc_pageable(kernel_map, PAGE_SIZE);
1262
1263 /*
1264 * allocate object for the ptes
1265 */
1266 if (pmap->pm_pteobj == NULL)
1267 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT, PTDPTDI + 1);
1268
1269 /*
1270 * allocate the page directory page
1271 */
1272 ptdpg = vm_page_grab(pmap->pm_pteobj, PTDPTDI,
1273 VM_ALLOC_NORMAL | VM_ALLOC_RETRY | VM_ALLOC_WIRED | VM_ALLOC_ZERO);
1274 vm_page_flag_clear(ptdpg, PG_BUSY);
1275 ptdpg->valid = VM_PAGE_BITS_ALL;
1276
1277 pmap_qenter((vm_offset_t) pmap->pm_pdir, &ptdpg, 1);
1278 if ((ptdpg->flags & PG_ZERO) == 0)
1279 bzero(pmap->pm_pdir, PAGE_SIZE);
1280
1281 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list);
1282 /* Wire in kernel global address entries. */
1283 /* XXX copies current process, does not fill in MPPTDI */
1284 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PTESIZE);
1285 #ifdef SMP
1286 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1287 #endif
1288
1289 /* install self-referential address mapping entry */
1290 pmap->pm_pdir[PTDPTDI] =
1291 VM_PAGE_TO_PHYS(ptdpg) | PG_V | PG_RW | PG_A | PG_M;
1292
1293 pmap->pm_active = 0;
1294 pmap->pm_ptphint = NULL;
1295 TAILQ_INIT(&pmap->pm_pvlist);
1296 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1297 }
1298
1299 /*
1300 * Wire in kernel global address entries. To avoid a race condition
1301 * between pmap initialization and pmap_growkernel, this procedure
1302 * should be called after the vmspace is attached to the process
1303 * but before this pmap is activated.
1304 */
1305 void
1306 pmap_pinit2(pmap)
1307 struct pmap *pmap;
1308 {
1309 /* XXX: Remove this stub when no longer called */
1310 }
1311
1312 static int
1313 pmap_release_free_page(pmap_t pmap, vm_page_t p)
1314 {
1315 pd_entry_t *pde = pmap->pm_pdir;
1316
1317 /*
1318 * This code optimizes the case of freeing non-busy
1319 * page-table pages. Those pages are zero now, and
1320 * might as well be placed directly into the zero queue.
1321 */
1322 vm_page_lock_queues();
1323 if (vm_page_sleep_if_busy(p, FALSE, "pmaprl"))
1324 return (0);
1325 vm_page_busy(p);
1326
1327 /*
1328 * Remove the page table page from the processes address space.
1329 */
1330 pde[p->pindex] = 0;
1331 pmap->pm_stats.resident_count--;
1332
1333 if (p->hold_count) {
1334 panic("pmap_release: freeing held page table page");
1335 }
1336 /*
1337 * Page directory pages need to have the kernel
1338 * stuff cleared, so they can go into the zero queue also.
1339 */
1340 if (p->pindex == PTDPTDI) {
1341 bzero(pde + KPTDI, nkpt * PTESIZE);
1342 #ifdef SMP
1343 pde[MPPTDI] = 0;
1344 #endif
1345 pde[APTDPTDI] = 0;
1346 pmap_kremove((vm_offset_t) pmap->pm_pdir);
1347 }
1348
1349 if (pmap->pm_ptphint == p)
1350 pmap->pm_ptphint = NULL;
1351
1352 p->wire_count--;
1353 cnt.v_wire_count--;
1354 vm_page_free_zero(p);
1355 vm_page_unlock_queues();
1356 return 1;
1357 }
1358
1359 /*
1360 * this routine is called if the page table page is not
1361 * mapped correctly.
1362 */
1363 static vm_page_t
1364 _pmap_allocpte(pmap, ptepindex)
1365 pmap_t pmap;
1366 unsigned ptepindex;
1367 {
1368 vm_offset_t pteva, ptepa; /* XXXPA */
1369 vm_page_t m;
1370
1371 /*
1372 * Find or fabricate a new pagetable page
1373 */
1374 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1375 VM_ALLOC_WIRED | VM_ALLOC_ZERO | VM_ALLOC_RETRY);
1376
1377 KASSERT(m->queue == PQ_NONE,
1378 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1379
1380 /*
1381 * Increment the hold count for the page table page
1382 * (denoting a new mapping.)
1383 */
1384 m->hold_count++;
1385
1386 /*
1387 * Map the pagetable page into the process address space, if
1388 * it isn't already there.
1389 */
1390
1391 pmap->pm_stats.resident_count++;
1392
1393 ptepa = VM_PAGE_TO_PHYS(m);
1394 pmap->pm_pdir[ptepindex] =
1395 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1396
1397 /*
1398 * Set the page table hint
1399 */
1400 pmap->pm_ptphint = m;
1401
1402 /*
1403 * Try to use the new mapping, but if we cannot, then
1404 * do it with the routine that maps the page explicitly.
1405 */
1406 if ((m->flags & PG_ZERO) == 0) {
1407 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) ==
1408 (PTDpde & PG_FRAME)) {
1409 pteva = VM_MAXUSER_ADDRESS + i386_ptob(ptepindex);
1410 bzero((caddr_t) pteva, PAGE_SIZE);
1411 } else {
1412 pmap_zero_page(m);
1413 }
1414 }
1415
1416 m->valid = VM_PAGE_BITS_ALL;
1417 vm_page_flag_clear(m, PG_ZERO);
1418 vm_page_wakeup(m);
1419
1420 return m;
1421 }
1422
1423 static vm_page_t
1424 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1425 {
1426 unsigned ptepindex;
1427 pd_entry_t ptepa;
1428 vm_page_t m;
1429
1430 /*
1431 * Calculate pagetable page index
1432 */
1433 ptepindex = va >> PDRSHIFT;
1434
1435 /*
1436 * Get the page directory entry
1437 */
1438 ptepa = (vm_offset_t) pmap->pm_pdir[ptepindex];
1439
1440 /*
1441 * This supports switching from a 4MB page to a
1442 * normal 4K page.
1443 */
1444 if (ptepa & PG_PS) {
1445 pmap->pm_pdir[ptepindex] = 0;
1446 ptepa = 0;
1447 pmap_invalidate_all(kernel_pmap);
1448 }
1449
1450 /*
1451 * If the page table page is mapped, we just increment the
1452 * hold count, and activate it.
1453 */
1454 if (ptepa) {
1455 /*
1456 * In order to get the page table page, try the
1457 * hint first.
1458 */
1459 if (pmap->pm_ptphint &&
1460 (pmap->pm_ptphint->pindex == ptepindex)) {
1461 m = pmap->pm_ptphint;
1462 } else {
1463 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1464 pmap->pm_ptphint = m;
1465 }
1466 m->hold_count++;
1467 return m;
1468 }
1469 /*
1470 * Here if the pte page isn't mapped, or if it has been deallocated.
1471 */
1472 return _pmap_allocpte(pmap, ptepindex);
1473 }
1474
1475
1476 /***************************************************
1477 * Pmap allocation/deallocation routines.
1478 ***************************************************/
1479
1480 /*
1481 * Release any resources held by the given physical map.
1482 * Called when a pmap initialized by pmap_pinit is being released.
1483 * Should only be called if the map contains no valid mappings.
1484 */
1485 void
1486 pmap_release(pmap_t pmap)
1487 {
1488 vm_page_t p,n,ptdpg;
1489 vm_object_t object = pmap->pm_pteobj;
1490 int curgeneration;
1491
1492 #if defined(DIAGNOSTIC)
1493 if (object->ref_count != 1)
1494 panic("pmap_release: pteobj reference count != 1");
1495 #endif
1496
1497 ptdpg = NULL;
1498 LIST_REMOVE(pmap, pm_list);
1499 retry:
1500 curgeneration = object->generation;
1501 for (p = TAILQ_FIRST(&object->memq); p != NULL; p = n) {
1502 n = TAILQ_NEXT(p, listq);
1503 if (p->pindex == PTDPTDI) {
1504 ptdpg = p;
1505 continue;
1506 }
1507 while (1) {
1508 if (!pmap_release_free_page(pmap, p) &&
1509 (object->generation != curgeneration))
1510 goto retry;
1511 }
1512 }
1513
1514 if (ptdpg && !pmap_release_free_page(pmap, ptdpg))
1515 goto retry;
1516 }
1517
1518 static int
1519 kvm_size(SYSCTL_HANDLER_ARGS)
1520 {
1521 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE;
1522
1523 return sysctl_handle_long(oidp, &ksize, 0, req);
1524 }
1525 SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD,
1526 0, 0, kvm_size, "IU", "Size of KVM");
1527
1528 static int
1529 kvm_free(SYSCTL_HANDLER_ARGS)
1530 {
1531 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end;
1532
1533 return sysctl_handle_long(oidp, &kfree, 0, req);
1534 }
1535 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD,
1536 0, 0, kvm_free, "IU", "Amount of KVM free");
1537
1538 /*
1539 * grow the number of kernel page table entries, if needed
1540 */
1541 void
1542 pmap_growkernel(vm_offset_t addr)
1543 {
1544 struct pmap *pmap;
1545 int s;
1546 vm_offset_t ptppaddr;
1547 vm_page_t nkpg;
1548 pd_entry_t newpdir;
1549
1550 s = splhigh();
1551 if (kernel_vm_end == 0) {
1552 kernel_vm_end = KERNBASE;
1553 nkpt = 0;
1554 while (pdir_pde(PTD, kernel_vm_end)) {
1555 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1556 nkpt++;
1557 }
1558 }
1559 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1560 while (kernel_vm_end < addr) {
1561 if (pdir_pde(PTD, kernel_vm_end)) {
1562 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1563 continue;
1564 }
1565
1566 /*
1567 * This index is bogus, but out of the way
1568 */
1569 nkpg = vm_page_alloc(kptobj, nkpt,
1570 VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
1571 if (!nkpg)
1572 panic("pmap_growkernel: no memory to grow kernel");
1573
1574 nkpt++;
1575
1576 pmap_zero_page(nkpg);
1577 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1578 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1579 pdir_pde(PTD, kernel_vm_end) = newpdir;
1580
1581 LIST_FOREACH(pmap, &allpmaps, pm_list) {
1582 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1583 }
1584 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1585 }
1586 splx(s);
1587 }
1588
1589
1590 /***************************************************
1591 * page management routines.
1592 ***************************************************/
1593
1594 /*
1595 * free the pv_entry back to the free list
1596 */
1597 static PMAP_INLINE void
1598 free_pv_entry(pv_entry_t pv)
1599 {
1600 pv_entry_count--;
1601 uma_zfree(pvzone, pv);
1602 }
1603
1604 /*
1605 * get a new pv_entry, allocating a block from the system
1606 * when needed.
1607 * the memory allocation is performed bypassing the malloc code
1608 * because of the possibility of allocations at interrupt time.
1609 */
1610 static pv_entry_t
1611 get_pv_entry(void)
1612 {
1613 pv_entry_count++;
1614 if (pv_entry_high_water &&
1615 (pv_entry_count > pv_entry_high_water) &&
1616 (pmap_pagedaemon_waken == 0)) {
1617 pmap_pagedaemon_waken = 1;
1618 wakeup (&vm_pages_needed);
1619 }
1620 return uma_zalloc(pvzone, M_NOWAIT);
1621 }
1622
1623 /*
1624 * If it is the first entry on the list, it is actually
1625 * in the header and we must copy the following entry up
1626 * to the header. Otherwise we must search the list for
1627 * the entry. In either case we free the now unused entry.
1628 */
1629
1630 static int
1631 pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va)
1632 {
1633 pv_entry_t pv;
1634 int rtval;
1635 int s;
1636
1637 s = splvm();
1638 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1639 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1640 if (pmap == pv->pv_pmap && va == pv->pv_va)
1641 break;
1642 }
1643 } else {
1644 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1645 if (va == pv->pv_va)
1646 break;
1647 }
1648 }
1649
1650 rtval = 0;
1651 if (pv) {
1652 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1653 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1654 m->md.pv_list_count--;
1655 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1656 vm_page_flag_clear(m, PG_WRITEABLE);
1657
1658 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1659 free_pv_entry(pv);
1660 }
1661
1662 splx(s);
1663 return rtval;
1664 }
1665
1666 /*
1667 * Create a pv entry for page at pa for
1668 * (pmap, va).
1669 */
1670 static void
1671 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1672 {
1673
1674 int s;
1675 pv_entry_t pv;
1676
1677 s = splvm();
1678 pv = get_pv_entry();
1679 pv->pv_va = va;
1680 pv->pv_pmap = pmap;
1681 pv->pv_ptem = mpte;
1682
1683 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1684 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1685 m->md.pv_list_count++;
1686
1687 splx(s);
1688 }
1689
1690 /*
1691 * pmap_remove_pte: do the things to unmap a page in a process
1692 */
1693 static int
1694 pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va)
1695 {
1696 pt_entry_t oldpte;
1697 vm_page_t m;
1698
1699 oldpte = atomic_readandclear_int(ptq);
1700 if (oldpte & PG_W)
1701 pmap->pm_stats.wired_count -= 1;
1702 /*
1703 * Machines that don't support invlpg, also don't support
1704 * PG_G.
1705 */
1706 if (oldpte & PG_G)
1707 pmap_invalidate_page(kernel_pmap, va);
1708 pmap->pm_stats.resident_count -= 1;
1709 if (oldpte & PG_MANAGED) {
1710 m = PHYS_TO_VM_PAGE(oldpte);
1711 if (oldpte & PG_M) {
1712 #if defined(PMAP_DIAGNOSTIC)
1713 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1714 printf(
1715 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1716 va, oldpte);
1717 }
1718 #endif
1719 if (pmap_track_modified(va))
1720 vm_page_dirty(m);
1721 }
1722 if (oldpte & PG_A)
1723 vm_page_flag_set(m, PG_REFERENCED);
1724 return pmap_remove_entry(pmap, m, va);
1725 } else {
1726 return pmap_unuse_pt(pmap, va, NULL);
1727 }
1728
1729 return 0;
1730 }
1731
1732 /*
1733 * Remove a single page from a process address space
1734 */
1735 static void
1736 pmap_remove_page(pmap_t pmap, vm_offset_t va)
1737 {
1738 register pt_entry_t *ptq;
1739
1740 /*
1741 * if there is no pte for this address, just skip it!!!
1742 */
1743 if (*pmap_pde(pmap, va) == 0) {
1744 return;
1745 }
1746
1747 /*
1748 * get a local va for mappings for this pmap.
1749 */
1750 ptq = get_ptbase(pmap) + i386_btop(va);
1751 if (*ptq) {
1752 (void) pmap_remove_pte(pmap, ptq, va);
1753 pmap_invalidate_page(pmap, va);
1754 }
1755 return;
1756 }
1757
1758 /*
1759 * Remove the given range of addresses from the specified map.
1760 *
1761 * It is assumed that the start and end are properly
1762 * rounded to the page size.
1763 */
1764 void
1765 pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1766 {
1767 register pt_entry_t *ptbase;
1768 vm_offset_t pdnxt;
1769 pd_entry_t ptpaddr;
1770 vm_offset_t sindex, eindex;
1771 int anyvalid;
1772
1773 if (pmap == NULL)
1774 return;
1775
1776 if (pmap->pm_stats.resident_count == 0)
1777 return;
1778
1779 /*
1780 * special handling of removing one page. a very
1781 * common operation and easy to short circuit some
1782 * code.
1783 */
1784 if ((sva + PAGE_SIZE == eva) &&
1785 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1786 pmap_remove_page(pmap, sva);
1787 return;
1788 }
1789
1790 anyvalid = 0;
1791
1792 /*
1793 * Get a local virtual address for the mappings that are being
1794 * worked with.
1795 */
1796 ptbase = get_ptbase(pmap);
1797
1798 sindex = i386_btop(sva);
1799 eindex = i386_btop(eva);
1800
1801 for (; sindex < eindex; sindex = pdnxt) {
1802 unsigned pdirindex;
1803
1804 /*
1805 * Calculate index for next page table.
1806 */
1807 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1808 if (pmap->pm_stats.resident_count == 0)
1809 break;
1810
1811 pdirindex = sindex / NPDEPG;
1812 ptpaddr = pmap->pm_pdir[pdirindex];
1813 if ((ptpaddr & PG_PS) != 0) {
1814 pmap->pm_pdir[pdirindex] = 0;
1815 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1816 anyvalid++;
1817 continue;
1818 }
1819
1820 /*
1821 * Weed out invalid mappings. Note: we assume that the page
1822 * directory table is always allocated, and in kernel virtual.
1823 */
1824 if (ptpaddr == 0)
1825 continue;
1826
1827 /*
1828 * Limit our scan to either the end of the va represented
1829 * by the current page table page, or to the end of the
1830 * range being removed.
1831 */
1832 if (pdnxt > eindex) {
1833 pdnxt = eindex;
1834 }
1835
1836 for (; sindex != pdnxt; sindex++) {
1837 vm_offset_t va;
1838 if (ptbase[sindex] == 0) {
1839 continue;
1840 }
1841 va = i386_ptob(sindex);
1842
1843 anyvalid++;
1844 if (pmap_remove_pte(pmap,
1845 ptbase + sindex, va))
1846 break;
1847 }
1848 }
1849
1850 if (anyvalid)
1851 pmap_invalidate_all(pmap);
1852 }
1853
1854 /*
1855 * Routine: pmap_remove_all
1856 * Function:
1857 * Removes this physical page from
1858 * all physical maps in which it resides.
1859 * Reflects back modify bits to the pager.
1860 *
1861 * Notes:
1862 * Original versions of this routine were very
1863 * inefficient because they iteratively called
1864 * pmap_remove (slow...)
1865 */
1866
1867 void
1868 pmap_remove_all(vm_page_t m)
1869 {
1870 register pv_entry_t pv;
1871 pt_entry_t *pte, tpte;
1872 int s;
1873
1874 #if defined(PMAP_DIAGNOSTIC)
1875 /*
1876 * XXX This makes pmap_remove_all() illegal for non-managed pages!
1877 */
1878 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1879 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x",
1880 VM_PAGE_TO_PHYS(m));
1881 }
1882 #endif
1883 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1884 s = splvm();
1885 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1886 pv->pv_pmap->pm_stats.resident_count--;
1887 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1888 tpte = atomic_readandclear_int(pte);
1889 if (tpte & PG_W)
1890 pv->pv_pmap->pm_stats.wired_count--;
1891 if (tpte & PG_A)
1892 vm_page_flag_set(m, PG_REFERENCED);
1893
1894 /*
1895 * Update the vm_page_t clean and reference bits.
1896 */
1897 if (tpte & PG_M) {
1898 #if defined(PMAP_DIAGNOSTIC)
1899 if (pmap_nw_modified((pt_entry_t) tpte)) {
1900 printf(
1901 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1902 pv->pv_va, tpte);
1903 }
1904 #endif
1905 if (pmap_track_modified(pv->pv_va))
1906 vm_page_dirty(m);
1907 }
1908 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
1909 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1910 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1911 m->md.pv_list_count--;
1912 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
1913 free_pv_entry(pv);
1914 }
1915 vm_page_flag_clear(m, PG_WRITEABLE);
1916 splx(s);
1917 }
1918
1919 /*
1920 * Set the physical protection on the
1921 * specified range of this map as requested.
1922 */
1923 void
1924 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1925 {
1926 register pt_entry_t *ptbase;
1927 vm_offset_t pdnxt;
1928 pd_entry_t ptpaddr;
1929 vm_offset_t sindex, eindex;
1930 int anychanged;
1931
1932 if (pmap == NULL)
1933 return;
1934
1935 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1936 pmap_remove(pmap, sva, eva);
1937 return;
1938 }
1939
1940 if (prot & VM_PROT_WRITE)
1941 return;
1942
1943 anychanged = 0;
1944
1945 ptbase = get_ptbase(pmap);
1946
1947 sindex = i386_btop(sva);
1948 eindex = i386_btop(eva);
1949
1950 for (; sindex < eindex; sindex = pdnxt) {
1951
1952 unsigned pdirindex;
1953
1954 pdnxt = ((sindex + NPTEPG) & ~(NPTEPG - 1));
1955
1956 pdirindex = sindex / NPDEPG;
1957 ptpaddr = pmap->pm_pdir[pdirindex];
1958 if ((ptpaddr & PG_PS) != 0) {
1959 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1960 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1961 anychanged++;
1962 continue;
1963 }
1964
1965 /*
1966 * Weed out invalid mappings. Note: we assume that the page
1967 * directory table is always allocated, and in kernel virtual.
1968 */
1969 if (ptpaddr == 0)
1970 continue;
1971
1972 if (pdnxt > eindex) {
1973 pdnxt = eindex;
1974 }
1975
1976 for (; sindex != pdnxt; sindex++) {
1977
1978 pt_entry_t pbits;
1979 vm_page_t m;
1980
1981 pbits = ptbase[sindex];
1982
1983 if (pbits & PG_MANAGED) {
1984 m = NULL;
1985 if (pbits & PG_A) {
1986 m = PHYS_TO_VM_PAGE(pbits);
1987 vm_page_flag_set(m, PG_REFERENCED);
1988 pbits &= ~PG_A;
1989 }
1990 if (pbits & PG_M) {
1991 if (pmap_track_modified(i386_ptob(sindex))) {
1992 if (m == NULL)
1993 m = PHYS_TO_VM_PAGE(pbits);
1994 vm_page_dirty(m);
1995 pbits &= ~PG_M;
1996 }
1997 }
1998 }
1999
2000 pbits &= ~PG_RW;
2001
2002 if (pbits != ptbase[sindex]) {
2003 ptbase[sindex] = pbits;
2004 anychanged = 1;
2005 }
2006 }
2007 }
2008 if (anychanged)
2009 pmap_invalidate_all(pmap);
2010 }
2011
2012 /*
2013 * Insert the given physical page (p) at
2014 * the specified virtual address (v) in the
2015 * target physical map with the protection requested.
2016 *
2017 * If specified, the page will be wired down, meaning
2018 * that the related pte can not be reclaimed.
2019 *
2020 * NB: This is the only routine which MAY NOT lazy-evaluate
2021 * or lose information. That is, this routine must actually
2022 * insert this page into the given map NOW.
2023 */
2024 void
2025 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2026 boolean_t wired)
2027 {
2028 vm_offset_t pa;
2029 register pt_entry_t *pte;
2030 vm_offset_t opa;
2031 pt_entry_t origpte, newpte;
2032 vm_page_t mpte;
2033
2034 if (pmap == NULL)
2035 return;
2036
2037 va &= PG_FRAME;
2038 #ifdef PMAP_DIAGNOSTIC
2039 if (va > VM_MAX_KERNEL_ADDRESS)
2040 panic("pmap_enter: toobig");
2041 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
2042 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
2043 #endif
2044
2045 mpte = NULL;
2046 /*
2047 * In the case that a page table page is not
2048 * resident, we are creating it here.
2049 */
2050 if (va < VM_MAXUSER_ADDRESS) {
2051 mpte = pmap_allocpte(pmap, va);
2052 }
2053 #if 0 && defined(PMAP_DIAGNOSTIC)
2054 else {
2055 pd_entry_t *pdeaddr = pmap_pde(pmap, va);
2056 origpte = *pdeaddr;
2057 if ((origpte & PG_V) == 0) {
2058 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n",
2059 pmap->pm_pdir[PTDPTDI], origpte, va);
2060 }
2061 }
2062 #endif
2063
2064 pte = pmap_pte(pmap, va);
2065
2066 /*
2067 * Page Directory table entry not valid, we need a new PT page
2068 */
2069 if (pte == NULL) {
2070 panic("pmap_enter: invalid page directory, pdir=%p, va=0x%x\n",
2071 (void *)pmap->pm_pdir[PTDPTDI], va);
2072 }
2073
2074 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
2075 origpte = *(vm_offset_t *)pte;
2076 opa = origpte & PG_FRAME;
2077
2078 if (origpte & PG_PS)
2079 panic("pmap_enter: attempted pmap_enter on 4MB page");
2080
2081 /*
2082 * Mapping has not changed, must be protection or wiring change.
2083 */
2084 if (origpte && (opa == pa)) {
2085 /*
2086 * Wiring change, just update stats. We don't worry about
2087 * wiring PT pages as they remain resident as long as there
2088 * are valid mappings in them. Hence, if a user page is wired,
2089 * the PT page will be also.
2090 */
2091 if (wired && ((origpte & PG_W) == 0))
2092 pmap->pm_stats.wired_count++;
2093 else if (!wired && (origpte & PG_W))
2094 pmap->pm_stats.wired_count--;
2095
2096 #if defined(PMAP_DIAGNOSTIC)
2097 if (pmap_nw_modified((pt_entry_t) origpte)) {
2098 printf(
2099 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
2100 va, origpte);
2101 }
2102 #endif
2103
2104 /*
2105 * Remove extra pte reference
2106 */
2107 if (mpte)
2108 mpte->hold_count--;
2109
2110 if ((prot & VM_PROT_WRITE) && (origpte & PG_V)) {
2111 if ((origpte & PG_RW) == 0) {
2112 *pte |= PG_RW;
2113 pmap_invalidate_page(pmap, va);
2114 }
2115 return;
2116 }
2117
2118 /*
2119 * We might be turning off write access to the page,
2120 * so we go ahead and sense modify status.
2121 */
2122 if (origpte & PG_MANAGED) {
2123 if ((origpte & PG_M) && pmap_track_modified(va)) {
2124 vm_page_t om;
2125 om = PHYS_TO_VM_PAGE(opa);
2126 vm_page_dirty(om);
2127 }
2128 pa |= PG_MANAGED;
2129 }
2130 goto validate;
2131 }
2132 /*
2133 * Mapping has changed, invalidate old range and fall through to
2134 * handle validating new mapping.
2135 */
2136 if (opa) {
2137 int err;
2138 vm_page_lock_queues();
2139 err = pmap_remove_pte(pmap, pte, va);
2140 vm_page_unlock_queues();
2141 if (err)
2142 panic("pmap_enter: pte vanished, va: 0x%x", va);
2143 }
2144
2145 /*
2146 * Enter on the PV list if part of our managed memory. Note that we
2147 * raise IPL while manipulating pv_table since pmap_enter can be
2148 * called at interrupt time.
2149 */
2150 if (pmap_initialized &&
2151 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2152 pmap_insert_entry(pmap, va, mpte, m);
2153 pa |= PG_MANAGED;
2154 }
2155
2156 /*
2157 * Increment counters
2158 */
2159 pmap->pm_stats.resident_count++;
2160 if (wired)
2161 pmap->pm_stats.wired_count++;
2162
2163 validate:
2164 /*
2165 * Now validate mapping with desired protection/wiring.
2166 */
2167 newpte = (vm_offset_t) (pa | pte_prot(pmap, prot) | PG_V);
2168
2169 if (wired)
2170 newpte |= PG_W;
2171 if (va < VM_MAXUSER_ADDRESS)
2172 newpte |= PG_U;
2173 if (pmap == kernel_pmap)
2174 newpte |= pgeflag;
2175
2176 /*
2177 * if the mapping or permission bits are different, we need
2178 * to update the pte.
2179 */
2180 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2181 *pte = newpte | PG_A;
2182 /*if (origpte)*/ {
2183 pmap_invalidate_page(pmap, va);
2184 }
2185 }
2186 }
2187
2188 /*
2189 * this code makes some *MAJOR* assumptions:
2190 * 1. Current pmap & pmap exists.
2191 * 2. Not wired.
2192 * 3. Read access.
2193 * 4. No page table pages.
2194 * 5. Tlbflush is deferred to calling procedure.
2195 * 6. Page IS managed.
2196 * but is *MUCH* faster than pmap_enter...
2197 */
2198
2199 static vm_page_t
2200 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
2201 {
2202 pt_entry_t *pte;
2203 vm_offset_t pa;
2204
2205 /*
2206 * In the case that a page table page is not
2207 * resident, we are creating it here.
2208 */
2209 if (va < VM_MAXUSER_ADDRESS) {
2210 unsigned ptepindex;
2211 pd_entry_t ptepa;
2212
2213 /*
2214 * Calculate pagetable page index
2215 */
2216 ptepindex = va >> PDRSHIFT;
2217 if (mpte && (mpte->pindex == ptepindex)) {
2218 mpte->hold_count++;
2219 } else {
2220 retry:
2221 /*
2222 * Get the page directory entry
2223 */
2224 ptepa = pmap->pm_pdir[ptepindex];
2225
2226 /*
2227 * If the page table page is mapped, we just increment
2228 * the hold count, and activate it.
2229 */
2230 if (ptepa) {
2231 if (ptepa & PG_PS)
2232 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2233 if (pmap->pm_ptphint &&
2234 (pmap->pm_ptphint->pindex == ptepindex)) {
2235 mpte = pmap->pm_ptphint;
2236 } else {
2237 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
2238 pmap->pm_ptphint = mpte;
2239 }
2240 if (mpte == NULL)
2241 goto retry;
2242 mpte->hold_count++;
2243 } else {
2244 mpte = _pmap_allocpte(pmap, ptepindex);
2245 }
2246 }
2247 } else {
2248 mpte = NULL;
2249 }
2250
2251 /*
2252 * This call to vtopte makes the assumption that we are
2253 * entering the page into the current pmap. In order to support
2254 * quick entry into any pmap, one would likely use pmap_pte_quick.
2255 * But that isn't as quick as vtopte.
2256 */
2257 pte = vtopte(va);
2258 if (*pte) {
2259 if (mpte != NULL) {
2260 vm_page_lock_queues();
2261 pmap_unwire_pte_hold(pmap, mpte);
2262 vm_page_unlock_queues();
2263 }
2264 return 0;
2265 }
2266
2267 /*
2268 * Enter on the PV list if part of our managed memory. Note that we
2269 * raise IPL while manipulating pv_table since pmap_enter can be
2270 * called at interrupt time.
2271 */
2272 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
2273 pmap_insert_entry(pmap, va, mpte, m);
2274
2275 /*
2276 * Increment counters
2277 */
2278 pmap->pm_stats.resident_count++;
2279
2280 pa = VM_PAGE_TO_PHYS(m);
2281
2282 /*
2283 * Now validate mapping with RO protection
2284 */
2285 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2286 *pte = pa | PG_V | PG_U;
2287 else
2288 *pte = pa | PG_V | PG_U | PG_MANAGED;
2289
2290 return mpte;
2291 }
2292
2293 /*
2294 * Make a temporary mapping for a physical address. This is only intended
2295 * to be used for panic dumps.
2296 */
2297 void *
2298 pmap_kenter_temporary(vm_offset_t pa, int i)
2299 {
2300 vm_offset_t va;
2301
2302 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE);
2303 pmap_kenter(va, pa);
2304 #ifndef I386_CPU
2305 invlpg(va);
2306 #else
2307 invltlb();
2308 #endif
2309 return ((void *)crashdumpmap);
2310 }
2311
2312 #define MAX_INIT_PT (96)
2313 /*
2314 * pmap_object_init_pt preloads the ptes for a given object
2315 * into the specified pmap. This eliminates the blast of soft
2316 * faults on process startup and immediately after an mmap.
2317 */
2318 void
2319 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr,
2320 vm_object_t object, vm_pindex_t pindex,
2321 vm_size_t size, int limit)
2322 {
2323 vm_offset_t tmpidx;
2324 int psize;
2325 vm_page_t p, mpte;
2326
2327 if (pmap == NULL || object == NULL)
2328 return;
2329
2330 /*
2331 * This code maps large physical mmap regions into the
2332 * processor address space. Note that some shortcuts
2333 * are taken, but the code works.
2334 */
2335 if (pseflag && (object->type == OBJT_DEVICE) &&
2336 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) {
2337 int i;
2338 vm_page_t m[1];
2339 unsigned int ptepindex;
2340 int npdes;
2341 pd_entry_t ptepa;
2342
2343 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)])
2344 return;
2345
2346 retry:
2347 p = vm_page_lookup(object, pindex);
2348 if (p != NULL) {
2349 vm_page_lock_queues();
2350 if (vm_page_sleep_if_busy(p, FALSE, "init4p"))
2351 goto retry;
2352 vm_page_unlock_queues();
2353 } else {
2354 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
2355 if (p == NULL)
2356 return;
2357 m[0] = p;
2358
2359 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) {
2360 vm_page_lock_queues();
2361 vm_page_free(p);
2362 vm_page_unlock_queues();
2363 return;
2364 }
2365
2366 p = vm_page_lookup(object, pindex);
2367 vm_page_wakeup(p);
2368 }
2369
2370 ptepa = VM_PAGE_TO_PHYS(p);
2371 if (ptepa & (NBPDR - 1)) {
2372 return;
2373 }
2374
2375 p->valid = VM_PAGE_BITS_ALL;
2376
2377 pmap->pm_stats.resident_count += size >> PAGE_SHIFT;
2378 npdes = size >> PDRSHIFT;
2379 for(i = 0; i < npdes; i++) {
2380 pmap->pm_pdir[ptepindex] =
2381 ptepa | PG_U | PG_RW | PG_V | PG_PS;
2382 ptepa += NBPDR;
2383 ptepindex += 1;
2384 }
2385 pmap_invalidate_all(kernel_pmap);
2386 return;
2387 }
2388
2389 psize = i386_btop(size);
2390
2391 if ((object->type != OBJT_VNODE) ||
2392 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2393 (object->resident_page_count > MAX_INIT_PT))) {
2394 return;
2395 }
2396
2397 if (psize + pindex > object->size) {
2398 if (object->size < pindex)
2399 return;
2400 psize = object->size - pindex;
2401 }
2402
2403 mpte = NULL;
2404
2405 if ((p = TAILQ_FIRST(&object->memq)) != NULL) {
2406 if (p->pindex < pindex) {
2407 p = vm_page_splay(pindex, object->root);
2408 if ((object->root = p)->pindex < pindex)
2409 p = TAILQ_NEXT(p, listq);
2410 }
2411 }
2412 /*
2413 * Assert: the variable p is either (1) the page with the
2414 * least pindex greater than or equal to the parameter pindex
2415 * or (2) NULL.
2416 */
2417 for (;
2418 p != NULL && (tmpidx = p->pindex - pindex) < psize;
2419 p = TAILQ_NEXT(p, listq)) {
2420 /*
2421 * don't allow an madvise to blow away our really
2422 * free pages allocating pv entries.
2423 */
2424 if ((limit & MAP_PREFAULT_MADVISE) &&
2425 cnt.v_free_count < cnt.v_free_reserved) {
2426 break;
2427 }
2428 vm_page_lock_queues();
2429 if ((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL &&
2430 (p->busy == 0) &&
2431 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2432 if ((p->queue - p->pc) == PQ_CACHE)
2433 vm_page_deactivate(p);
2434 vm_page_busy(p);
2435 vm_page_unlock_queues();
2436 mpte = pmap_enter_quick(pmap,
2437 addr + i386_ptob(tmpidx), p, mpte);
2438 vm_page_lock_queues();
2439 vm_page_wakeup(p);
2440 }
2441 vm_page_unlock_queues();
2442 }
2443 return;
2444 }
2445
2446 /*
2447 * pmap_prefault provides a quick way of clustering
2448 * pagefaults into a processes address space. It is a "cousin"
2449 * of pmap_object_init_pt, except it runs at page fault time instead
2450 * of mmap time.
2451 */
2452 #define PFBAK 4
2453 #define PFFOR 4
2454 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2455
2456 static int pmap_prefault_pageorder[] = {
2457 -1 * PAGE_SIZE, 1 * PAGE_SIZE,
2458 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2459 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2460 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2461 };
2462
2463 void
2464 pmap_prefault(pmap, addra, entry)
2465 pmap_t pmap;
2466 vm_offset_t addra;
2467 vm_map_entry_t entry;
2468 {
2469 int i;
2470 vm_offset_t starta;
2471 vm_offset_t addr;
2472 vm_pindex_t pindex;
2473 vm_page_t m, mpte;
2474 vm_object_t object;
2475
2476 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)))
2477 return;
2478
2479 object = entry->object.vm_object;
2480
2481 starta = addra - PFBAK * PAGE_SIZE;
2482 if (starta < entry->start) {
2483 starta = entry->start;
2484 } else if (starta > addra) {
2485 starta = 0;
2486 }
2487
2488 mpte = NULL;
2489 for (i = 0; i < PAGEORDER_SIZE; i++) {
2490 vm_object_t lobject;
2491 pt_entry_t *pte;
2492
2493 addr = addra + pmap_prefault_pageorder[i];
2494 if (addr > addra + (PFFOR * PAGE_SIZE))
2495 addr = 0;
2496
2497 if (addr < starta || addr >= entry->end)
2498 continue;
2499
2500 if ((*pmap_pde(pmap, addr)) == 0)
2501 continue;
2502
2503 pte = vtopte(addr);
2504 if (*pte)
2505 continue;
2506
2507 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2508 lobject = object;
2509 for (m = vm_page_lookup(lobject, pindex);
2510 (!m && (lobject->type == OBJT_DEFAULT) && (lobject->backing_object));
2511 lobject = lobject->backing_object) {
2512 if (lobject->backing_object_offset & PAGE_MASK)
2513 break;
2514 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2515 m = vm_page_lookup(lobject->backing_object, pindex);
2516 }
2517
2518 /*
2519 * give-up when a page is not in memory
2520 */
2521 if (m == NULL)
2522 break;
2523 vm_page_lock_queues();
2524 if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
2525 (m->busy == 0) &&
2526 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2527
2528 if ((m->queue - m->pc) == PQ_CACHE) {
2529 vm_page_deactivate(m);
2530 }
2531 vm_page_busy(m);
2532 vm_page_unlock_queues();
2533 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2534 vm_page_lock_queues();
2535 vm_page_wakeup(m);
2536 }
2537 vm_page_unlock_queues();
2538 }
2539 }
2540
2541 /*
2542 * Routine: pmap_change_wiring
2543 * Function: Change the wiring attribute for a map/virtual-address
2544 * pair.
2545 * In/out conditions:
2546 * The mapping must already exist in the pmap.
2547 */
2548 void
2549 pmap_change_wiring(pmap, va, wired)
2550 register pmap_t pmap;
2551 vm_offset_t va;
2552 boolean_t wired;
2553 {
2554 register pt_entry_t *pte;
2555
2556 if (pmap == NULL)
2557 return;
2558
2559 pte = pmap_pte(pmap, va);
2560
2561 if (wired && !pmap_pte_w(pte))
2562 pmap->pm_stats.wired_count++;
2563 else if (!wired && pmap_pte_w(pte))
2564 pmap->pm_stats.wired_count--;
2565
2566 /*
2567 * Wiring is not a hardware characteristic so there is no need to
2568 * invalidate TLB.
2569 */
2570 pmap_pte_set_w(pte, wired);
2571 }
2572
2573
2574
2575 /*
2576 * Copy the range specified by src_addr/len
2577 * from the source map to the range dst_addr/len
2578 * in the destination map.
2579 *
2580 * This routine is only advisory and need not do anything.
2581 */
2582
2583 void
2584 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len,
2585 vm_offset_t src_addr)
2586 {
2587 vm_offset_t addr;
2588 vm_offset_t end_addr = src_addr + len;
2589 vm_offset_t pdnxt;
2590 pd_entry_t src_frame, dst_frame;
2591 vm_page_t m;
2592
2593 if (dst_addr != src_addr)
2594 return;
2595
2596 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
2597 if (src_frame != (PTDpde & PG_FRAME))
2598 return;
2599
2600 dst_frame = dst_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
2601 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2602 pt_entry_t *src_pte, *dst_pte;
2603 vm_page_t dstmpte, srcmpte;
2604 pd_entry_t srcptepaddr;
2605 unsigned ptepindex;
2606
2607 if (addr >= UPT_MIN_ADDRESS)
2608 panic("pmap_copy: invalid to pmap_copy page tables\n");
2609
2610 /*
2611 * Don't let optional prefaulting of pages make us go
2612 * way below the low water mark of free pages or way
2613 * above high water mark of used pv entries.
2614 */
2615 if (cnt.v_free_count < cnt.v_free_reserved ||
2616 pv_entry_count > pv_entry_high_water)
2617 break;
2618
2619 pdnxt = ((addr + PAGE_SIZE*NPTEPG) & ~(PAGE_SIZE*NPTEPG - 1));
2620 ptepindex = addr >> PDRSHIFT;
2621
2622 srcptepaddr = src_pmap->pm_pdir[ptepindex];
2623 if (srcptepaddr == 0)
2624 continue;
2625
2626 if (srcptepaddr & PG_PS) {
2627 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2628 dst_pmap->pm_pdir[ptepindex] = srcptepaddr;
2629 dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE;
2630 }
2631 continue;
2632 }
2633
2634 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2635 if ((srcmpte == NULL) ||
2636 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2637 continue;
2638
2639 if (pdnxt > end_addr)
2640 pdnxt = end_addr;
2641
2642 /*
2643 * Have to recheck this before every avtopte() call below
2644 * in case we have blocked and something else used APTDpde.
2645 */
2646 if (dst_frame != (APTDpde & PG_FRAME)) {
2647 APTDpde = dst_frame | PG_RW | PG_V;
2648 pmap_invalidate_all(kernel_pmap); /* XXX Bandaid */
2649 }
2650 src_pte = vtopte(addr);
2651 dst_pte = avtopte(addr);
2652 while (addr < pdnxt) {
2653 pt_entry_t ptetemp;
2654 ptetemp = *src_pte;
2655 /*
2656 * we only virtual copy managed pages
2657 */
2658 if ((ptetemp & PG_MANAGED) != 0) {
2659 /*
2660 * We have to check after allocpte for the
2661 * pte still being around... allocpte can
2662 * block.
2663 */
2664 dstmpte = pmap_allocpte(dst_pmap, addr);
2665 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2666 /*
2667 * Clear the modified and
2668 * accessed (referenced) bits
2669 * during the copy.
2670 */
2671 m = PHYS_TO_VM_PAGE(ptetemp);
2672 *dst_pte = ptetemp & ~(PG_M | PG_A);
2673 dst_pmap->pm_stats.resident_count++;
2674 pmap_insert_entry(dst_pmap, addr,
2675 dstmpte, m);
2676 } else {
2677 vm_page_lock_queues();
2678 pmap_unwire_pte_hold(dst_pmap, dstmpte);
2679 vm_page_unlock_queues();
2680 }
2681 if (dstmpte->hold_count >= srcmpte->hold_count)
2682 break;
2683 }
2684 addr += PAGE_SIZE;
2685 src_pte++;
2686 dst_pte++;
2687 }
2688 }
2689 }
2690
2691 #ifdef SMP
2692
2693 /*
2694 * pmap_zpi_switchin*()
2695 *
2696 * These functions allow us to avoid doing IPIs alltogether in certain
2697 * temporary page-mapping situations (page zeroing). Instead to deal
2698 * with being preempted and moved onto a different cpu we invalidate
2699 * the page when the scheduler switches us in. This does not occur
2700 * very often so we remain relatively optimal with very little effort.
2701 */
2702 static void
2703 pmap_zpi_switchin12(void)
2704 {
2705 invlpg((u_int)CADDR1);
2706 invlpg((u_int)CADDR2);
2707 }
2708
2709 static void
2710 pmap_zpi_switchin2(void)
2711 {
2712 invlpg((u_int)CADDR2);
2713 }
2714
2715 static void
2716 pmap_zpi_switchin3(void)
2717 {
2718 invlpg((u_int)CADDR3);
2719 }
2720
2721 #endif
2722
2723 /*
2724 * pmap_zero_page zeros the specified hardware page by mapping
2725 * the page into KVM and using bzero to clear its contents.
2726 */
2727 void
2728 pmap_zero_page(vm_page_t m)
2729 {
2730 vm_offset_t phys;
2731
2732 phys = VM_PAGE_TO_PHYS(m);
2733 if (*CMAP2)
2734 panic("pmap_zero_page: CMAP2 busy");
2735 *CMAP2 = PG_V | PG_RW | phys | PG_A | PG_M;
2736 #ifdef I386_CPU
2737 invltlb();
2738 #else
2739 #ifdef SMP
2740 curthread->td_switchin = pmap_zpi_switchin2;
2741 #endif
2742 invlpg((u_int)CADDR2);
2743 #endif
2744 #if defined(I686_CPU)
2745 if (cpu_class == CPUCLASS_686)
2746 i686_pagezero(CADDR2);
2747 else
2748 #endif
2749 bzero(CADDR2, PAGE_SIZE);
2750 #ifdef SMP
2751 curthread->td_switchin = NULL;
2752 #endif
2753 *CMAP2 = 0;
2754 }
2755
2756 /*
2757 * pmap_zero_page_area zeros the specified hardware page by mapping
2758 * the page into KVM and using bzero to clear its contents.
2759 *
2760 * off and size may not cover an area beyond a single hardware page.
2761 */
2762 void
2763 pmap_zero_page_area(vm_page_t m, int off, int size)
2764 {
2765 vm_offset_t phys;
2766
2767 phys = VM_PAGE_TO_PHYS(m);
2768 if (*CMAP2)
2769 panic("pmap_zero_page: CMAP2 busy");
2770 *CMAP2 = PG_V | PG_RW | phys | PG_A | PG_M;
2771 #ifdef I386_CPU
2772 invltlb();
2773 #else
2774 #ifdef SMP
2775 curthread->td_switchin = pmap_zpi_switchin2;
2776 #endif
2777 invlpg((u_int)CADDR2);
2778 #endif
2779 #if defined(I686_CPU)
2780 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2781 i686_pagezero(CADDR2);
2782 else
2783 #endif
2784 bzero((char *)CADDR2 + off, size);
2785 #ifdef SMP
2786 curthread->td_switchin = NULL;
2787 #endif
2788 *CMAP2 = 0;
2789 }
2790
2791 /*
2792 * pmap_zero_page_idle zeros the specified hardware page by mapping
2793 * the page into KVM and using bzero to clear its contents. This
2794 * is intended to be called from the vm_pagezero process only and
2795 * outside of Giant.
2796 */
2797 void
2798 pmap_zero_page_idle(vm_page_t m)
2799 {
2800 vm_offset_t phys;
2801
2802 phys = VM_PAGE_TO_PHYS(m);
2803 if (*CMAP3)
2804 panic("pmap_zero_page: CMAP3 busy");
2805 *CMAP3 = PG_V | PG_RW | phys | PG_A | PG_M;
2806 #ifdef I386_CPU
2807 invltlb();
2808 #else
2809 #ifdef SMP
2810 curthread->td_switchin = pmap_zpi_switchin3;
2811 #endif
2812 invlpg((u_int)CADDR3);
2813 #endif
2814 #if defined(I686_CPU)
2815 if (cpu_class == CPUCLASS_686)
2816 i686_pagezero(CADDR3);
2817 else
2818 #endif
2819 bzero(CADDR3, PAGE_SIZE);
2820 #ifdef SMP
2821 curthread->td_switchin = NULL;
2822 #endif
2823 *CMAP3 = 0;
2824 }
2825
2826 /*
2827 * pmap_copy_page copies the specified (machine independent)
2828 * page by mapping the page into virtual memory and using
2829 * bcopy to copy the page, one machine dependent page at a
2830 * time.
2831 */
2832 void
2833 pmap_copy_page(vm_page_t src, vm_page_t dst)
2834 {
2835
2836 if (*CMAP1)
2837 panic("pmap_copy_page: CMAP1 busy");
2838 if (*CMAP2)
2839 panic("pmap_copy_page: CMAP2 busy");
2840 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A;
2841 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M;
2842 #ifdef I386_CPU
2843 invltlb();
2844 #else
2845 #ifdef SMP
2846 curthread->td_switchin = pmap_zpi_switchin12;
2847 #endif
2848 invlpg((u_int)CADDR1);
2849 invlpg((u_int)CADDR2);
2850 #endif
2851 bcopy(CADDR1, CADDR2, PAGE_SIZE);
2852 #ifdef SMP
2853 curthread->td_switchin = NULL;
2854 #endif
2855 *CMAP1 = 0;
2856 *CMAP2 = 0;
2857 }
2858
2859 /*
2860 * Returns true if the pmap's pv is one of the first
2861 * 16 pvs linked to from this page. This count may
2862 * be changed upwards or downwards in the future; it
2863 * is only necessary that true be returned for a small
2864 * subset of pmaps for proper page aging.
2865 */
2866 boolean_t
2867 pmap_page_exists_quick(pmap, m)
2868 pmap_t pmap;
2869 vm_page_t m;
2870 {
2871 pv_entry_t pv;
2872 int loops = 0;
2873 int s;
2874
2875 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2876 return FALSE;
2877
2878 s = splvm();
2879 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2880 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2881 if (pv->pv_pmap == pmap) {
2882 splx(s);
2883 return TRUE;
2884 }
2885 loops++;
2886 if (loops >= 16)
2887 break;
2888 }
2889 splx(s);
2890 return (FALSE);
2891 }
2892
2893 #define PMAP_REMOVE_PAGES_CURPROC_ONLY
2894 /*
2895 * Remove all pages from specified address space
2896 * this aids process exit speeds. Also, this code
2897 * is special cased for current process only, but
2898 * can have the more generic (and slightly slower)
2899 * mode enabled. This is much faster than pmap_remove
2900 * in the case of running down an entire address space.
2901 */
2902 void
2903 pmap_remove_pages(pmap, sva, eva)
2904 pmap_t pmap;
2905 vm_offset_t sva, eva;
2906 {
2907 pt_entry_t *pte, tpte;
2908 vm_page_t m;
2909 pv_entry_t pv, npv;
2910 int s;
2911
2912 #ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY
2913 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) {
2914 printf("warning: pmap_remove_pages called with non-current pmap\n");
2915 return;
2916 }
2917 #endif
2918 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2919 s = splvm();
2920 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2921
2922 if (pv->pv_va >= eva || pv->pv_va < sva) {
2923 npv = TAILQ_NEXT(pv, pv_plist);
2924 continue;
2925 }
2926
2927 #ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY
2928 pte = vtopte(pv->pv_va);
2929 #else
2930 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2931 #endif
2932 tpte = *pte;
2933
2934 if (tpte == 0) {
2935 printf("TPTE at %p IS ZERO @ VA %08x\n",
2936 pte, pv->pv_va);
2937 panic("bad pte");
2938 }
2939
2940 /*
2941 * We cannot remove wired pages from a process' mapping at this time
2942 */
2943 if (tpte & PG_W) {
2944 npv = TAILQ_NEXT(pv, pv_plist);
2945 continue;
2946 }
2947
2948 m = PHYS_TO_VM_PAGE(tpte);
2949 KASSERT(m->phys_addr == (tpte & PG_FRAME),
2950 ("vm_page_t %p phys_addr mismatch %08x %08x",
2951 m, m->phys_addr, tpte));
2952
2953 KASSERT(m < &vm_page_array[vm_page_array_size],
2954 ("pmap_remove_pages: bad tpte %x", tpte));
2955
2956 pv->pv_pmap->pm_stats.resident_count--;
2957
2958 *pte = 0;
2959
2960 /*
2961 * Update the vm_page_t clean and reference bits.
2962 */
2963 if (tpte & PG_M) {
2964 vm_page_dirty(m);
2965 }
2966
2967 npv = TAILQ_NEXT(pv, pv_plist);
2968 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2969
2970 m->md.pv_list_count--;
2971 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2972 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2973 vm_page_flag_clear(m, PG_WRITEABLE);
2974 }
2975
2976 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
2977 free_pv_entry(pv);
2978 }
2979 splx(s);
2980 pmap_invalidate_all(pmap);
2981 }
2982
2983 /*
2984 * pmap_testbit tests bits in pte's
2985 * note that the testbit/changebit routines are inline,
2986 * and a lot of things compile-time evaluate.
2987 */
2988 static boolean_t
2989 pmap_testbit(m, bit)
2990 vm_page_t m;
2991 int bit;
2992 {
2993 pv_entry_t pv;
2994 pt_entry_t *pte;
2995 int s;
2996
2997 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2998 return FALSE;
2999
3000 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
3001 return FALSE;
3002
3003 s = splvm();
3004
3005 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3006 /*
3007 * if the bit being tested is the modified bit, then
3008 * mark clean_map and ptes as never
3009 * modified.
3010 */
3011 if (bit & (PG_A|PG_M)) {
3012 if (!pmap_track_modified(pv->pv_va))
3013 continue;
3014 }
3015
3016 #if defined(PMAP_DIAGNOSTIC)
3017 if (!pv->pv_pmap) {
3018 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
3019 continue;
3020 }
3021 #endif
3022 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3023 if (*pte & bit) {
3024 splx(s);
3025 return TRUE;
3026 }
3027 }
3028 splx(s);
3029 return (FALSE);
3030 }
3031
3032 /*
3033 * this routine is used to modify bits in ptes
3034 */
3035 static __inline void
3036 pmap_changebit(vm_page_t m, int bit, boolean_t setem)
3037 {
3038 register pv_entry_t pv;
3039 register pt_entry_t *pte;
3040 int s;
3041
3042 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) ||
3043 (!setem && bit == PG_RW && (m->flags & PG_WRITEABLE) == 0))
3044 return;
3045
3046 s = splvm();
3047 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3048 /*
3049 * Loop over all current mappings setting/clearing as appropos If
3050 * setting RO do we need to clear the VAC?
3051 */
3052 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3053 /*
3054 * don't write protect pager mappings
3055 */
3056 if (!setem && (bit == PG_RW)) {
3057 if (!pmap_track_modified(pv->pv_va))
3058 continue;
3059 }
3060
3061 #if defined(PMAP_DIAGNOSTIC)
3062 if (!pv->pv_pmap) {
3063 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
3064 continue;
3065 }
3066 #endif
3067
3068 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3069
3070 if (setem) {
3071 *pte |= bit;
3072 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
3073 } else {
3074 pt_entry_t pbits = *pte;
3075 if (pbits & bit) {
3076 if (bit == PG_RW) {
3077 if (pbits & PG_M) {
3078 vm_page_dirty(m);
3079 }
3080 *pte = pbits & ~(PG_M|PG_RW);
3081 } else {
3082 *pte = pbits & ~bit;
3083 }
3084 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
3085 }
3086 }
3087 }
3088 if (!setem && bit == PG_RW)
3089 vm_page_flag_clear(m, PG_WRITEABLE);
3090 splx(s);
3091 }
3092
3093 /*
3094 * pmap_page_protect:
3095 *
3096 * Lower the permission for all mappings to a given page.
3097 */
3098 void
3099 pmap_page_protect(vm_page_t m, vm_prot_t prot)
3100 {
3101 if ((prot & VM_PROT_WRITE) == 0) {
3102 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
3103 pmap_changebit(m, PG_RW, FALSE);
3104 } else {
3105 pmap_remove_all(m);
3106 }
3107 }
3108 }
3109
3110 vm_offset_t
3111 pmap_phys_address(ppn)
3112 int ppn;
3113 {
3114 return (i386_ptob(ppn));
3115 }
3116
3117 /*
3118 * pmap_ts_referenced:
3119 *
3120 * Return a count of reference bits for a page, clearing those bits.
3121 * It is not necessary for every reference bit to be cleared, but it
3122 * is necessary that 0 only be returned when there are truly no
3123 * reference bits set.
3124 *
3125 * XXX: The exact number of bits to check and clear is a matter that
3126 * should be tested and standardized at some point in the future for
3127 * optimal aging of shared pages.
3128 */
3129 int
3130 pmap_ts_referenced(vm_page_t m)
3131 {
3132 register pv_entry_t pv, pvf, pvn;
3133 pt_entry_t *pte;
3134 int s;
3135 int rtval = 0;
3136
3137 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
3138 return (rtval);
3139
3140 s = splvm();
3141 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
3142 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
3143
3144 pvf = pv;
3145
3146 do {
3147 pvn = TAILQ_NEXT(pv, pv_list);
3148
3149 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
3150
3151 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
3152
3153 if (!pmap_track_modified(pv->pv_va))
3154 continue;
3155
3156 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
3157
3158 if (pte && (*pte & PG_A)) {
3159 *pte &= ~PG_A;
3160
3161 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
3162
3163 rtval++;
3164 if (rtval > 4) {
3165 break;
3166 }
3167 }
3168 } while ((pv = pvn) != NULL && pv != pvf);
3169 }
3170 splx(s);
3171
3172 return (rtval);
3173 }
3174
3175 /*
3176 * pmap_is_modified:
3177 *
3178 * Return whether or not the specified physical page was modified
3179 * in any physical maps.
3180 */
3181 boolean_t
3182 pmap_is_modified(vm_page_t m)
3183 {
3184 return pmap_testbit(m, PG_M);
3185 }
3186
3187 /*
3188 * Clear the modify bits on the specified physical page.
3189 */
3190 void
3191 pmap_clear_modify(vm_page_t m)
3192 {
3193 pmap_changebit(m, PG_M, FALSE);
3194 }
3195
3196 /*
3197 * pmap_clear_reference:
3198 *
3199 * Clear the reference bit on the specified physical page.
3200 */
3201 void
3202 pmap_clear_reference(vm_page_t m)
3203 {
3204 pmap_changebit(m, PG_A, FALSE);
3205 }
3206
3207 /*
3208 * Miscellaneous support routines follow
3209 */
3210
3211 static void
3212 i386_protection_init()
3213 {
3214 register int *kp, prot;
3215
3216 kp = protection_codes;
3217 for (prot = 0; prot < 8; prot++) {
3218 switch (prot) {
3219 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3220 /*
3221 * Read access is also 0. There isn't any execute bit,
3222 * so just make it readable.
3223 */
3224 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3225 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3226 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3227 *kp++ = 0;
3228 break;
3229 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3230 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3231 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3232 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3233 *kp++ = PG_RW;
3234 break;
3235 }
3236 }
3237 }
3238
3239 /*
3240 * Map a set of physical memory pages into the kernel virtual
3241 * address space. Return a pointer to where it is mapped. This
3242 * routine is intended to be used for mapping device memory,
3243 * NOT real memory.
3244 */
3245 void *
3246 pmap_mapdev(pa, size)
3247 vm_offset_t pa;
3248 vm_size_t size;
3249 {
3250 vm_offset_t va, tmpva, offset;
3251 pt_entry_t *pte;
3252
3253 offset = pa & PAGE_MASK;
3254 size = roundup(offset + size, PAGE_SIZE);
3255
3256 GIANT_REQUIRED;
3257
3258 va = kmem_alloc_pageable(kernel_map, size);
3259 if (!va)
3260 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3261
3262 pa = pa & PG_FRAME;
3263 for (tmpva = va; size > 0; ) {
3264 pte = vtopte(tmpva);
3265 *pte = pa | PG_RW | PG_V | pgeflag;
3266 size -= PAGE_SIZE;
3267 tmpva += PAGE_SIZE;
3268 pa += PAGE_SIZE;
3269 }
3270 pmap_invalidate_range(kernel_pmap, va, tmpva);
3271 return ((void *)(va + offset));
3272 }
3273
3274 void
3275 pmap_unmapdev(va, size)
3276 vm_offset_t va;
3277 vm_size_t size;
3278 {
3279 vm_offset_t base, offset, tmpva;
3280 pt_entry_t *pte;
3281
3282 base = va & PG_FRAME;
3283 offset = va & PAGE_MASK;
3284 size = roundup(offset + size, PAGE_SIZE);
3285 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE) {
3286 pte = vtopte(tmpva);
3287 *pte = 0;
3288 }
3289 pmap_invalidate_range(kernel_pmap, va, tmpva);
3290 kmem_free(kernel_map, base, size);
3291 }
3292
3293 /*
3294 * perform the pmap work for mincore
3295 */
3296 int
3297 pmap_mincore(pmap, addr)
3298 pmap_t pmap;
3299 vm_offset_t addr;
3300 {
3301 pt_entry_t *ptep, pte;
3302 vm_page_t m;
3303 int val = 0;
3304
3305 ptep = pmap_pte(pmap, addr);
3306 if (ptep == 0) {
3307 return 0;
3308 }
3309
3310 if ((pte = *ptep) != 0) {
3311 vm_offset_t pa;
3312
3313 val = MINCORE_INCORE;
3314 if ((pte & PG_MANAGED) == 0)
3315 return val;
3316
3317 pa = pte & PG_FRAME;
3318
3319 m = PHYS_TO_VM_PAGE(pa);
3320
3321 /*
3322 * Modified by us
3323 */
3324 if (pte & PG_M)
3325 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3326 else {
3327 /*
3328 * Modified by someone else
3329 */
3330 vm_page_lock_queues();
3331 if (m->dirty || pmap_is_modified(m))
3332 val |= MINCORE_MODIFIED_OTHER;
3333 vm_page_unlock_queues();
3334 }
3335 /*
3336 * Referenced by us
3337 */
3338 if (pte & PG_A)
3339 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3340 else {
3341 /*
3342 * Referenced by someone else
3343 */
3344 vm_page_lock_queues();
3345 if ((m->flags & PG_REFERENCED) ||
3346 pmap_ts_referenced(m)) {
3347 val |= MINCORE_REFERENCED_OTHER;
3348 vm_page_flag_set(m, PG_REFERENCED);
3349 }
3350 vm_page_unlock_queues();
3351 }
3352 }
3353 return val;
3354 }
3355
3356 void
3357 pmap_activate(struct thread *td)
3358 {
3359 struct proc *p = td->td_proc;
3360 pmap_t pmap;
3361 u_int32_t cr3;
3362
3363 pmap = vmspace_pmap(td->td_proc->p_vmspace);
3364 #if defined(SMP)
3365 pmap->pm_active |= PCPU_GET(cpumask);
3366 #else
3367 pmap->pm_active |= 1;
3368 #endif
3369 cr3 = vtophys(pmap->pm_pdir);
3370 /* XXXKSE this is wrong.
3371 * pmap_activate is for the current thread on the current cpu
3372 */
3373 if (p->p_flag & P_KSES) {
3374 /* Make sure all other cr3 entries are updated. */
3375 /* what if they are running? XXXKSE (maybe abort them) */
3376 FOREACH_THREAD_IN_PROC(p, td) {
3377 td->td_pcb->pcb_cr3 = cr3;
3378 }
3379 } else {
3380 td->td_pcb->pcb_cr3 = cr3;
3381 }
3382 load_cr3(cr3);
3383 #ifdef SWTCH_OPTIM_STATS
3384 tlb_flush_count++;
3385 #endif
3386 }
3387
3388 vm_offset_t
3389 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3390 {
3391
3392 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3393 return addr;
3394 }
3395
3396 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3397 return addr;
3398 }
3399
3400
3401 #if defined(PMAP_DEBUG)
3402 pmap_pid_dump(int pid)
3403 {
3404 pmap_t pmap;
3405 struct proc *p;
3406 int npte = 0;
3407 int index;
3408
3409 sx_slock(&allproc_lock);
3410 LIST_FOREACH(p, &allproc, p_list) {
3411 if (p->p_pid != pid)
3412 continue;
3413
3414 if (p->p_vmspace) {
3415 int i,j;
3416 index = 0;
3417 pmap = vmspace_pmap(p->p_vmspace);
3418 for (i = 0; i < NPDEPG; i++) {
3419 pd_entry_t *pde;
3420 pt_entry_t *pte;
3421 vm_offset_t base = i << PDRSHIFT;
3422
3423 pde = &pmap->pm_pdir[i];
3424 if (pde && pmap_pde_v(pde)) {
3425 for (j = 0; j < NPTEPG; j++) {
3426 vm_offset_t va = base + (j << PAGE_SHIFT);
3427 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) {
3428 if (index) {
3429 index = 0;
3430 printf("\n");
3431 }
3432 sx_sunlock(&allproc_lock);
3433 return npte;
3434 }
3435 pte = pmap_pte_quick(pmap, va);
3436 if (pte && pmap_pte_v(pte)) {
3437 pt_entry_t pa;
3438 vm_page_t m;
3439 pa = *pte;
3440 m = PHYS_TO_VM_PAGE(pa);
3441 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x",
3442 va, pa, m->hold_count, m->wire_count, m->flags);
3443 npte++;
3444 index++;
3445 if (index >= 2) {
3446 index = 0;
3447 printf("\n");
3448 } else {
3449 printf(" ");
3450 }
3451 }
3452 }
3453 }
3454 }
3455 }
3456 }
3457 sx_sunlock(&allproc_lock);
3458 return npte;
3459 }
3460 #endif
3461
3462 #if defined(DEBUG)
3463
3464 static void pads(pmap_t pm);
3465 void pmap_pvdump(vm_offset_t pa);
3466
3467 /* print address space of pmap*/
3468 static void
3469 pads(pm)
3470 pmap_t pm;
3471 {
3472 int i, j;
3473 vm_offset_t va;
3474 pt_entry_t *ptep;
3475
3476 if (pm == kernel_pmap)
3477 return;
3478 for (i = 0; i < NPDEPG; i++)
3479 if (pm->pm_pdir[i])
3480 for (j = 0; j < NPTEPG; j++) {
3481 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3482 if (pm == kernel_pmap && va < KERNBASE)
3483 continue;
3484 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS)
3485 continue;
3486 ptep = pmap_pte_quick(pm, va);
3487 if (pmap_pte_v(ptep))
3488 printf("%x:%x ", va, *ptep);
3489 };
3490
3491 }
3492
3493 void
3494 pmap_pvdump(pa)
3495 vm_offset_t pa;
3496 {
3497 pv_entry_t pv;
3498 vm_page_t m;
3499
3500 printf("pa %x", pa);
3501 m = PHYS_TO_VM_PAGE(pa);
3502 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3503 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);
3504 pads(pv->pv_pmap);
3505 }
3506 printf(" ");
3507 }
3508 #endif
Cache object: 61c28a6aeffb524bd1febfbd0f8a5a3d
|