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 */
43 /*-
44 * Copyright (c) 2003 Networks Associates Technology, Inc.
45 * All rights reserved.
46 *
47 * This software was developed for the FreeBSD Project by Jake Burkholder,
48 * Safeport Network Services, and Network Associates Laboratories, the
49 * Security Research Division of Network Associates, Inc. under
50 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA
51 * CHATS research program.
52 *
53 * Redistribution and use in source and binary forms, with or without
54 * modification, are permitted provided that the following conditions
55 * are met:
56 * 1. Redistributions of source code must retain the above copyright
57 * notice, this list of conditions and the following disclaimer.
58 * 2. Redistributions in binary form must reproduce the above copyright
59 * notice, this list of conditions and the following disclaimer in the
60 * documentation and/or other materials provided with the distribution.
61 *
62 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
63 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
64 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
65 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
66 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
67 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
68 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
69 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
70 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
71 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
72 * SUCH DAMAGE.
73 */
74
75 #include <sys/cdefs.h>
76 __FBSDID("$FreeBSD: releng/5.2/sys/i386/i386/pmap.c 122284 2003-11-08 03:01:26Z alc $");
77
78 /*
79 * Manages physical address maps.
80 *
81 * In addition to hardware address maps, this
82 * module is called upon to provide software-use-only
83 * maps which may or may not be stored in the same
84 * form as hardware maps. These pseudo-maps are
85 * used to store intermediate results from copy
86 * operations to and from address spaces.
87 *
88 * Since the information managed by this module is
89 * also stored by the logical address mapping module,
90 * this module may throw away valid virtual-to-physical
91 * mappings at almost any time. However, invalidations
92 * of virtual-to-physical mappings must be done as
93 * requested.
94 *
95 * In order to cope with hardware architectures which
96 * make virtual-to-physical map invalidates expensive,
97 * this module may delay invalidate or reduced protection
98 * operations until such time as they are actually
99 * necessary. This module is given full information as
100 * to which processors are currently using which maps,
101 * and to when physical maps must be made correct.
102 */
103
104 #include "opt_pmap.h"
105 #include "opt_msgbuf.h"
106 #include "opt_kstack_pages.h"
107
108 #include <sys/param.h>
109 #include <sys/systm.h>
110 #include <sys/kernel.h>
111 #include <sys/lock.h>
112 #include <sys/mman.h>
113 #include <sys/msgbuf.h>
114 #include <sys/mutex.h>
115 #include <sys/proc.h>
116 #include <sys/sx.h>
117 #include <sys/user.h>
118 #include <sys/vmmeter.h>
119 #include <sys/sysctl.h>
120 #ifdef SMP
121 #include <sys/smp.h>
122 #endif
123
124 #include <vm/vm.h>
125 #include <vm/vm_param.h>
126 #include <vm/vm_kern.h>
127 #include <vm/vm_page.h>
128 #include <vm/vm_map.h>
129 #include <vm/vm_object.h>
130 #include <vm/vm_extern.h>
131 #include <vm/vm_pageout.h>
132 #include <vm/vm_pager.h>
133 #include <vm/uma.h>
134
135 #include <machine/cpu.h>
136 #include <machine/cputypes.h>
137 #include <machine/md_var.h>
138 #include <machine/specialreg.h>
139 #ifdef SMP
140 #include <machine/smp.h>
141 #endif
142
143 #define PMAP_KEEP_PDIRS
144 #ifndef PMAP_SHPGPERPROC
145 #define PMAP_SHPGPERPROC 200
146 #endif
147
148 #if defined(DIAGNOSTIC)
149 #define PMAP_DIAGNOSTIC
150 #endif
151
152 #define MINPV 2048
153
154 #if !defined(PMAP_DIAGNOSTIC)
155 #define PMAP_INLINE __inline
156 #else
157 #define PMAP_INLINE
158 #endif
159
160 /*
161 * Get PDEs and PTEs for user/kernel address space
162 */
163 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
164 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
165
166 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
167 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
168 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
169 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
170 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
171
172 #define pmap_pte_set_w(pte, v) ((v)?(*(int *)pte |= PG_W):(*(int *)pte &= ~PG_W))
173 #define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v)))
174
175 /*
176 * Given a map and a machine independent protection code,
177 * convert to a vax protection code.
178 */
179 #define pte_prot(m, p) (protection_codes[p])
180 static int protection_codes[8];
181
182 struct pmap kernel_pmap_store;
183 LIST_HEAD(pmaplist, pmap);
184 static struct pmaplist allpmaps;
185 static struct mtx allpmaps_lock;
186 #ifdef SMP
187 static struct mtx lazypmap_lock;
188 #endif
189
190 vm_paddr_t avail_start; /* PA of first available physical page */
191 vm_paddr_t avail_end; /* PA of last available physical page */
192 vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */
193 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
194 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
195 int pgeflag = 0; /* PG_G or-in */
196 int pseflag = 0; /* PG_PS or-in */
197
198 static int nkpt;
199 vm_offset_t kernel_vm_end;
200 extern u_int32_t KERNend;
201
202 #ifdef PAE
203 static uma_zone_t pdptzone;
204 #endif
205
206 /*
207 * Data for the pv entry allocation mechanism
208 */
209 static uma_zone_t pvzone;
210 static struct vm_object pvzone_obj;
211 static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0;
212 int pmap_pagedaemon_waken;
213
214 /*
215 * All those kernel PT submaps that BSD is so fond of
216 */
217 pt_entry_t *CMAP1 = 0;
218 static pt_entry_t *CMAP2, *CMAP3, *ptmmap;
219 caddr_t CADDR1 = 0, ptvmmap = 0;
220 static caddr_t CADDR2, CADDR3;
221 static struct mtx CMAPCADDR12_lock;
222 static pt_entry_t *msgbufmap;
223 struct msgbuf *msgbufp = 0;
224
225 /*
226 * Crashdump maps.
227 */
228 static pt_entry_t *pt_crashdumpmap;
229 static caddr_t crashdumpmap;
230
231 #ifdef SMP
232 extern pt_entry_t *SMPpt;
233 #endif
234 static pt_entry_t *PMAP1 = 0, *PMAP2;
235 static pt_entry_t *PADDR1 = 0, *PADDR2;
236
237 static PMAP_INLINE void free_pv_entry(pv_entry_t pv);
238 static pv_entry_t get_pv_entry(void);
239 static void i386_protection_init(void);
240 static void pmap_clear_ptes(vm_page_t m, int bit)
241 __always_inline;
242
243 static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva);
244 static void pmap_remove_page(struct pmap *pmap, vm_offset_t va);
245 static int pmap_remove_entry(struct pmap *pmap, vm_page_t m,
246 vm_offset_t va);
247 static void pmap_insert_entry(pmap_t pmap, vm_offset_t va,
248 vm_page_t mpte, vm_page_t m);
249
250 static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va);
251
252 static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex);
253 static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va);
254 static int pmap_unuse_pt(pmap_t, vm_offset_t, vm_page_t);
255 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
256 static void *pmap_pv_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait);
257 #ifdef PAE
258 static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait);
259 #endif
260
261 CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t));
262 CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t));
263
264 /*
265 * Move the kernel virtual free pointer to the next
266 * 4MB. This is used to help improve performance
267 * by using a large (4MB) page for much of the kernel
268 * (.text, .data, .bss)
269 */
270 static vm_offset_t
271 pmap_kmem_choose(vm_offset_t addr)
272 {
273 vm_offset_t newaddr = addr;
274
275 #ifndef DISABLE_PSE
276 if (cpu_feature & CPUID_PSE)
277 newaddr = (addr + PDRMASK) & ~PDRMASK;
278 #endif
279 return newaddr;
280 }
281
282 /*
283 * Bootstrap the system enough to run with virtual memory.
284 *
285 * On the i386 this is called after mapping has already been enabled
286 * and just syncs the pmap module with what has already been done.
287 * [We can't call it easily with mapping off since the kernel is not
288 * mapped with PA == VA, hence we would have to relocate every address
289 * from the linked base (virtual) address "KERNBASE" to the actual
290 * (physical) address starting relative to 0]
291 */
292 void
293 pmap_bootstrap(firstaddr, loadaddr)
294 vm_paddr_t firstaddr;
295 vm_paddr_t loadaddr;
296 {
297 vm_offset_t va;
298 pt_entry_t *pte;
299 int i;
300
301 avail_start = firstaddr;
302
303 /*
304 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too
305 * large. It should instead be correctly calculated in locore.s and
306 * not based on 'first' (which is a physical address, not a virtual
307 * address, for the start of unused physical memory). The kernel
308 * page tables are NOT double mapped and thus should not be included
309 * in this calculation.
310 */
311 virtual_avail = (vm_offset_t) KERNBASE + firstaddr;
312 virtual_avail = pmap_kmem_choose(virtual_avail);
313
314 virtual_end = VM_MAX_KERNEL_ADDRESS;
315
316 /*
317 * Initialize protection array.
318 */
319 i386_protection_init();
320
321 /*
322 * Initialize the kernel pmap (which is statically allocated).
323 */
324 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD);
325 #ifdef PAE
326 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT);
327 #endif
328 kernel_pmap->pm_active = -1; /* don't allow deactivation */
329 TAILQ_INIT(&kernel_pmap->pm_pvlist);
330 LIST_INIT(&allpmaps);
331 #ifdef SMP
332 mtx_init(&lazypmap_lock, "lazypmap", NULL, MTX_SPIN);
333 #endif
334 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN);
335 mtx_lock_spin(&allpmaps_lock);
336 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list);
337 mtx_unlock_spin(&allpmaps_lock);
338 nkpt = NKPT;
339
340 /*
341 * Reserve some special page table entries/VA space for temporary
342 * mapping of pages.
343 */
344 #define SYSMAP(c, p, v, n) \
345 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
346
347 va = virtual_avail;
348 pte = vtopte(va);
349
350 /*
351 * CMAP1/CMAP2 are used for zeroing and copying pages.
352 * CMAP3 is used for the idle process page zeroing.
353 */
354 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
355 SYSMAP(caddr_t, CMAP2, CADDR2, 1)
356 SYSMAP(caddr_t, CMAP3, CADDR3, 1)
357 *CMAP3 = 0;
358
359 mtx_init(&CMAPCADDR12_lock, "CMAPCADDR12", NULL, MTX_DEF);
360
361 /*
362 * Crashdump maps.
363 */
364 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
365
366 /*
367 * ptvmmap is used for reading arbitrary physical pages via /dev/mem.
368 * XXX ptmmap is not used.
369 */
370 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
371
372 /*
373 * msgbufp is used to map the system message buffer.
374 * XXX msgbufmap is not used.
375 */
376 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
377 atop(round_page(MSGBUF_SIZE)))
378
379 /*
380 * ptemap is used for pmap_pte_quick
381 */
382 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1);
383 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1);
384
385 virtual_avail = va;
386
387 *CMAP1 = *CMAP2 = 0;
388 for (i = 0; i < NKPT; i++)
389 PTD[i] = 0;
390
391 /* Turn on PG_G on kernel page(s) */
392 pmap_set_pg();
393 }
394
395 /*
396 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on.
397 */
398 void
399 pmap_set_pg(void)
400 {
401 pd_entry_t pdir;
402 pt_entry_t *pte;
403 vm_offset_t va, endva;
404 int i;
405
406 if (pgeflag == 0)
407 return;
408
409 i = KERNLOAD/NBPDR;
410 endva = KERNBASE + KERNend;
411
412 if (pseflag) {
413 va = KERNBASE + KERNLOAD;
414 while (va < endva) {
415 pdir = kernel_pmap->pm_pdir[KPTDI+i];
416 pdir |= pgeflag;
417 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir;
418 invltlb(); /* Play it safe, invltlb() every time */
419 i++;
420 va += NBPDR;
421 }
422 } else {
423 va = (vm_offset_t)btext;
424 while (va < endva) {
425 pte = vtopte(va);
426 if (*pte)
427 *pte |= pgeflag;
428 invltlb(); /* Play it safe, invltlb() every time */
429 va += PAGE_SIZE;
430 }
431 }
432 }
433
434 static void *
435 pmap_pv_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
436 {
437 *flags = UMA_SLAB_PRIV;
438 return (void *)kmem_alloc(kernel_map, bytes);
439 }
440
441 #ifdef PAE
442 static void *
443 pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
444 {
445 *flags = UMA_SLAB_PRIV;
446 return (contigmalloc(PAGE_SIZE, NULL, 0, 0x0ULL, 0xffffffffULL, 1, 0));
447 }
448 #endif
449
450 /*
451 * Initialize the pmap module.
452 * Called by vm_init, to initialize any structures that the pmap
453 * system needs to map virtual memory.
454 * pmap_init has been enhanced to support in a fairly consistant
455 * way, discontiguous physical memory.
456 */
457 void
458 pmap_init(phys_start, phys_end)
459 vm_paddr_t phys_start, phys_end;
460 {
461 int i;
462 int initial_pvs;
463
464 /*
465 * Allocate memory for random pmap data structures. Includes the
466 * pv_head_table.
467 */
468
469 for(i = 0; i < vm_page_array_size; i++) {
470 vm_page_t m;
471
472 m = &vm_page_array[i];
473 TAILQ_INIT(&m->md.pv_list);
474 m->md.pv_list_count = 0;
475 }
476
477 /*
478 * init the pv free list
479 */
480 initial_pvs = vm_page_array_size;
481 if (initial_pvs < MINPV)
482 initial_pvs = MINPV;
483 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL,
484 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE);
485 uma_zone_set_allocf(pvzone, pmap_pv_allocf);
486 uma_prealloc(pvzone, initial_pvs);
487
488 #ifdef PAE
489 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL,
490 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1,
491 UMA_ZONE_VM | UMA_ZONE_NOFREE);
492 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf);
493 #endif
494
495 /*
496 * Now it is safe to enable pv_table recording.
497 */
498 pmap_initialized = TRUE;
499 }
500
501 /*
502 * Initialize the address space (zone) for the pv_entries. Set a
503 * high water mark so that the system can recover from excessive
504 * numbers of pv entries.
505 */
506 void
507 pmap_init2()
508 {
509 int shpgperproc = PMAP_SHPGPERPROC;
510
511 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
512 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
513 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
514 pv_entry_high_water = 9 * (pv_entry_max / 10);
515 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max);
516 }
517
518
519 /***************************************************
520 * Low level helper routines.....
521 ***************************************************/
522
523 #if defined(PMAP_DIAGNOSTIC)
524
525 /*
526 * This code checks for non-writeable/modified pages.
527 * This should be an invalid condition.
528 */
529 static int
530 pmap_nw_modified(pt_entry_t ptea)
531 {
532 int pte;
533
534 pte = (int) ptea;
535
536 if ((pte & (PG_M|PG_RW)) == PG_M)
537 return 1;
538 else
539 return 0;
540 }
541 #endif
542
543
544 /*
545 * this routine defines the region(s) of memory that should
546 * not be tested for the modified bit.
547 */
548 static PMAP_INLINE int
549 pmap_track_modified(vm_offset_t va)
550 {
551 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva))
552 return 1;
553 else
554 return 0;
555 }
556
557 #ifdef I386_CPU
558 /*
559 * i386 only has "invalidate everything" and no SMP to worry about.
560 */
561 PMAP_INLINE void
562 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
563 {
564
565 if (pmap == kernel_pmap || pmap->pm_active)
566 invltlb();
567 }
568
569 PMAP_INLINE void
570 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
571 {
572
573 if (pmap == kernel_pmap || pmap->pm_active)
574 invltlb();
575 }
576
577 PMAP_INLINE void
578 pmap_invalidate_all(pmap_t pmap)
579 {
580
581 if (pmap == kernel_pmap || pmap->pm_active)
582 invltlb();
583 }
584 #else /* !I386_CPU */
585 #ifdef SMP
586 /*
587 * For SMP, these functions have to use the IPI mechanism for coherence.
588 */
589 void
590 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
591 {
592 u_int cpumask;
593 u_int other_cpus;
594
595 if (smp_started) {
596 if (!(read_eflags() & PSL_I))
597 panic("%s: interrupts disabled", __func__);
598 mtx_lock_spin(&smp_tlb_mtx);
599 } else
600 critical_enter();
601 /*
602 * We need to disable interrupt preemption but MUST NOT have
603 * interrupts disabled here.
604 * XXX we may need to hold schedlock to get a coherent pm_active
605 * XXX critical sections disable interrupts again
606 */
607 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) {
608 invlpg(va);
609 smp_invlpg(va);
610 } else {
611 cpumask = PCPU_GET(cpumask);
612 other_cpus = PCPU_GET(other_cpus);
613 if (pmap->pm_active & cpumask)
614 invlpg(va);
615 if (pmap->pm_active & other_cpus)
616 smp_masked_invlpg(pmap->pm_active & other_cpus, va);
617 }
618 if (smp_started)
619 mtx_unlock_spin(&smp_tlb_mtx);
620 else
621 critical_exit();
622 }
623
624 void
625 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
626 {
627 u_int cpumask;
628 u_int other_cpus;
629 vm_offset_t addr;
630
631 if (smp_started) {
632 if (!(read_eflags() & PSL_I))
633 panic("%s: interrupts disabled", __func__);
634 mtx_lock_spin(&smp_tlb_mtx);
635 } else
636 critical_enter();
637 /*
638 * We need to disable interrupt preemption but MUST NOT have
639 * interrupts disabled here.
640 * XXX we may need to hold schedlock to get a coherent pm_active
641 * XXX critical sections disable interrupts again
642 */
643 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) {
644 for (addr = sva; addr < eva; addr += PAGE_SIZE)
645 invlpg(addr);
646 smp_invlpg_range(sva, eva);
647 } else {
648 cpumask = PCPU_GET(cpumask);
649 other_cpus = PCPU_GET(other_cpus);
650 if (pmap->pm_active & cpumask)
651 for (addr = sva; addr < eva; addr += PAGE_SIZE)
652 invlpg(addr);
653 if (pmap->pm_active & other_cpus)
654 smp_masked_invlpg_range(pmap->pm_active & other_cpus,
655 sva, eva);
656 }
657 if (smp_started)
658 mtx_unlock_spin(&smp_tlb_mtx);
659 else
660 critical_exit();
661 }
662
663 void
664 pmap_invalidate_all(pmap_t pmap)
665 {
666 u_int cpumask;
667 u_int other_cpus;
668
669 if (smp_started) {
670 if (!(read_eflags() & PSL_I))
671 panic("%s: interrupts disabled", __func__);
672 mtx_lock_spin(&smp_tlb_mtx);
673 } else
674 critical_enter();
675 /*
676 * We need to disable interrupt preemption but MUST NOT have
677 * interrupts disabled here.
678 * XXX we may need to hold schedlock to get a coherent pm_active
679 * XXX critical sections disable interrupts again
680 */
681 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) {
682 invltlb();
683 smp_invltlb();
684 } else {
685 cpumask = PCPU_GET(cpumask);
686 other_cpus = PCPU_GET(other_cpus);
687 if (pmap->pm_active & cpumask)
688 invltlb();
689 if (pmap->pm_active & other_cpus)
690 smp_masked_invltlb(pmap->pm_active & other_cpus);
691 }
692 if (smp_started)
693 mtx_unlock_spin(&smp_tlb_mtx);
694 else
695 critical_exit();
696 }
697 #else /* !SMP */
698 /*
699 * Normal, non-SMP, 486+ invalidation functions.
700 * We inline these within pmap.c for speed.
701 */
702 PMAP_INLINE void
703 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
704 {
705
706 if (pmap == kernel_pmap || pmap->pm_active)
707 invlpg(va);
708 }
709
710 PMAP_INLINE void
711 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
712 {
713 vm_offset_t addr;
714
715 if (pmap == kernel_pmap || pmap->pm_active)
716 for (addr = sva; addr < eva; addr += PAGE_SIZE)
717 invlpg(addr);
718 }
719
720 PMAP_INLINE void
721 pmap_invalidate_all(pmap_t pmap)
722 {
723
724 if (pmap == kernel_pmap || pmap->pm_active)
725 invltlb();
726 }
727 #endif /* !SMP */
728 #endif /* !I386_CPU */
729
730 /*
731 * Are we current address space or kernel? N.B. We return FALSE when
732 * a pmap's page table is in use because a kernel thread is borrowing
733 * it. The borrowed page table can change spontaneously, making any
734 * dependence on its continued use subject to a race condition.
735 */
736 static __inline int
737 pmap_is_current(pmap_t pmap)
738 {
739
740 return (pmap == kernel_pmap ||
741 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) &&
742 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)));
743 }
744
745 /*
746 * If the given pmap is not the current pmap, Giant must be held.
747 */
748 pt_entry_t *
749 pmap_pte(pmap_t pmap, vm_offset_t va)
750 {
751 pd_entry_t newpf;
752 pd_entry_t *pde;
753
754 pde = pmap_pde(pmap, va);
755 if (*pde & PG_PS)
756 return (pde);
757 if (*pde != 0) {
758 /* are we current address space or kernel? */
759 if (pmap_is_current(pmap))
760 return (vtopte(va));
761 GIANT_REQUIRED;
762 newpf = *pde & PG_FRAME;
763 if ((*PMAP2 & PG_FRAME) != newpf) {
764 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M;
765 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2);
766 }
767 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1)));
768 }
769 return (0);
770 }
771
772 /*
773 * Super fast pmap_pte routine best used when scanning
774 * the pv lists. This eliminates many coarse-grained
775 * invltlb calls. Note that many of the pv list
776 * scans are across different pmaps. It is very wasteful
777 * to do an entire invltlb for checking a single mapping.
778 *
779 * If the given pmap is not the current pmap, vm_page_queue_mtx
780 * must be held.
781 */
782 static pt_entry_t *
783 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
784 {
785 pd_entry_t newpf;
786 pd_entry_t *pde;
787
788 pde = pmap_pde(pmap, va);
789 if (*pde & PG_PS)
790 return (pde);
791 if (*pde != 0) {
792 /* are we current address space or kernel? */
793 if (pmap_is_current(pmap))
794 return (vtopte(va));
795 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
796 newpf = *pde & PG_FRAME;
797 if ((*PMAP1 & PG_FRAME) != newpf) {
798 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M;
799 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR1);
800 }
801 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1)));
802 }
803 return (0);
804 }
805
806 /*
807 * Routine: pmap_extract
808 * Function:
809 * Extract the physical page address associated
810 * with the given map/virtual_address pair.
811 */
812 vm_paddr_t
813 pmap_extract(pmap, va)
814 register pmap_t pmap;
815 vm_offset_t va;
816 {
817 vm_paddr_t rtval;
818 pt_entry_t *pte;
819 pd_entry_t pde;
820
821 if (pmap == 0)
822 return 0;
823 pde = pmap->pm_pdir[va >> PDRSHIFT];
824 if (pde != 0) {
825 if ((pde & PG_PS) != 0) {
826 rtval = (pde & ~PDRMASK) | (va & PDRMASK);
827 return rtval;
828 }
829 pte = pmap_pte(pmap, va);
830 rtval = ((*pte & PG_FRAME) | (va & PAGE_MASK));
831 return rtval;
832 }
833 return 0;
834
835 }
836
837 /*
838 * Routine: pmap_extract_and_hold
839 * Function:
840 * Atomically extract and hold the physical page
841 * with the given pmap and virtual address pair
842 * if that mapping permits the given protection.
843 */
844 vm_page_t
845 pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
846 {
847 vm_paddr_t pa;
848 vm_page_t m;
849
850 m = NULL;
851 mtx_lock(&Giant);
852 if ((pa = pmap_extract(pmap, va)) != 0) {
853 m = PHYS_TO_VM_PAGE(pa);
854 vm_page_lock_queues();
855 vm_page_hold(m);
856 vm_page_unlock_queues();
857 }
858 mtx_unlock(&Giant);
859 return (m);
860 }
861
862 /***************************************************
863 * Low level mapping routines.....
864 ***************************************************/
865
866 /*
867 * Add a wired page to the kva.
868 * Note: not SMP coherent.
869 */
870 PMAP_INLINE void
871 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
872 {
873 pt_entry_t *pte;
874
875 pte = vtopte(va);
876 pte_store(pte, pa | PG_RW | PG_V | pgeflag);
877 }
878
879 /*
880 * Remove a page from the kernel pagetables.
881 * Note: not SMP coherent.
882 */
883 PMAP_INLINE void
884 pmap_kremove(vm_offset_t va)
885 {
886 pt_entry_t *pte;
887
888 pte = vtopte(va);
889 pte_clear(pte);
890 }
891
892 /*
893 * Used to map a range of physical addresses into kernel
894 * virtual address space.
895 *
896 * The value passed in '*virt' is a suggested virtual address for
897 * the mapping. Architectures which can support a direct-mapped
898 * physical to virtual region can return the appropriate address
899 * within that region, leaving '*virt' unchanged. Other
900 * architectures should map the pages starting at '*virt' and
901 * update '*virt' with the first usable address after the mapped
902 * region.
903 */
904 vm_offset_t
905 pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot)
906 {
907 vm_offset_t va, sva;
908
909 va = sva = *virt;
910 while (start < end) {
911 pmap_kenter(va, start);
912 va += PAGE_SIZE;
913 start += PAGE_SIZE;
914 }
915 pmap_invalidate_range(kernel_pmap, sva, va);
916 *virt = va;
917 return (sva);
918 }
919
920
921 /*
922 * Add a list of wired pages to the kva
923 * this routine is only used for temporary
924 * kernel mappings that do not need to have
925 * page modification or references recorded.
926 * Note that old mappings are simply written
927 * over. The page *must* be wired.
928 * Note: SMP coherent. Uses a ranged shootdown IPI.
929 */
930 void
931 pmap_qenter(vm_offset_t sva, vm_page_t *m, int count)
932 {
933 vm_offset_t va;
934
935 va = sva;
936 while (count-- > 0) {
937 pmap_kenter(va, VM_PAGE_TO_PHYS(*m));
938 va += PAGE_SIZE;
939 m++;
940 }
941 pmap_invalidate_range(kernel_pmap, sva, va);
942 }
943
944 /*
945 * This routine tears out page mappings from the
946 * kernel -- it is meant only for temporary mappings.
947 * Note: SMP coherent. Uses a ranged shootdown IPI.
948 */
949 void
950 pmap_qremove(vm_offset_t sva, int count)
951 {
952 vm_offset_t va;
953
954 va = sva;
955 while (count-- > 0) {
956 pmap_kremove(va);
957 va += PAGE_SIZE;
958 }
959 pmap_invalidate_range(kernel_pmap, sva, va);
960 }
961
962 /***************************************************
963 * Page table page management routines.....
964 ***************************************************/
965
966 /*
967 * This routine unholds page table pages, and if the hold count
968 * drops to zero, then it decrements the wire count.
969 */
970 static int
971 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
972 {
973
974 while (vm_page_sleep_if_busy(m, FALSE, "pmuwpt"))
975 vm_page_lock_queues();
976
977 if (m->hold_count == 0) {
978 vm_offset_t pteva;
979 /*
980 * unmap the page table page
981 */
982 pmap->pm_pdir[m->pindex] = 0;
983 --pmap->pm_stats.resident_count;
984 /*
985 * We never unwire a kernel page table page, making a
986 * check for the kernel_pmap unnecessary.
987 */
988 if ((pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)) {
989 /*
990 * Do an invltlb to make the invalidated mapping
991 * take effect immediately.
992 */
993 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex);
994 pmap_invalidate_page(pmap, pteva);
995 }
996
997 /*
998 * If the page is finally unwired, simply free it.
999 */
1000 --m->wire_count;
1001 if (m->wire_count == 0) {
1002 vm_page_busy(m);
1003 vm_page_free_zero(m);
1004 atomic_subtract_int(&cnt.v_wire_count, 1);
1005 }
1006 return 1;
1007 }
1008 return 0;
1009 }
1010
1011 static PMAP_INLINE int
1012 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
1013 {
1014 vm_page_unhold(m);
1015 if (m->hold_count == 0)
1016 return _pmap_unwire_pte_hold(pmap, m);
1017 else
1018 return 0;
1019 }
1020
1021 /*
1022 * After removing a page table entry, this routine is used to
1023 * conditionally free the page, and manage the hold/wire counts.
1024 */
1025 static int
1026 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
1027 {
1028
1029 if (va >= VM_MAXUSER_ADDRESS)
1030 return 0;
1031
1032 return pmap_unwire_pte_hold(pmap, mpte);
1033 }
1034
1035 void
1036 pmap_pinit0(pmap)
1037 struct pmap *pmap;
1038 {
1039
1040 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD);
1041 #ifdef PAE
1042 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT);
1043 #endif
1044 pmap->pm_active = 0;
1045 PCPU_SET(curpmap, pmap);
1046 TAILQ_INIT(&pmap->pm_pvlist);
1047 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1048 mtx_lock_spin(&allpmaps_lock);
1049 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list);
1050 mtx_unlock_spin(&allpmaps_lock);
1051 }
1052
1053 /*
1054 * Initialize a preallocated and zeroed pmap structure,
1055 * such as one in a vmspace structure.
1056 */
1057 void
1058 pmap_pinit(pmap)
1059 register struct pmap *pmap;
1060 {
1061 vm_page_t m, ptdpg[NPGPTD];
1062 vm_paddr_t pa;
1063 static int color;
1064 int i;
1065
1066 /*
1067 * No need to allocate page table space yet but we do need a valid
1068 * page directory table.
1069 */
1070 if (pmap->pm_pdir == NULL) {
1071 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map,
1072 NBPTD);
1073 #ifdef PAE
1074 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO);
1075 KASSERT(((vm_offset_t)pmap->pm_pdpt &
1076 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0,
1077 ("pmap_pinit: pdpt misaligned"));
1078 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30),
1079 ("pmap_pinit: pdpt above 4g"));
1080 #endif
1081 }
1082
1083 /*
1084 * allocate the page directory page(s)
1085 */
1086 for (i = 0; i < NPGPTD;) {
1087 m = vm_page_alloc(NULL, color++,
1088 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
1089 VM_ALLOC_ZERO);
1090 if (m == NULL)
1091 VM_WAIT;
1092 else {
1093 vm_page_lock_queues();
1094 vm_page_flag_clear(m, PG_BUSY);
1095 vm_page_unlock_queues();
1096 ptdpg[i++] = m;
1097 }
1098 }
1099
1100 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD);
1101
1102 for (i = 0; i < NPGPTD; i++) {
1103 if ((ptdpg[i]->flags & PG_ZERO) == 0)
1104 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE);
1105 }
1106
1107 mtx_lock_spin(&allpmaps_lock);
1108 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list);
1109 mtx_unlock_spin(&allpmaps_lock);
1110 /* Wire in kernel global address entries. */
1111 /* XXX copies current process, does not fill in MPPTDI */
1112 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t));
1113 #ifdef SMP
1114 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1115 #endif
1116
1117 /* install self-referential address mapping entry(s) */
1118 for (i = 0; i < NPGPTD; i++) {
1119 pa = VM_PAGE_TO_PHYS(ptdpg[i]);
1120 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M;
1121 #ifdef PAE
1122 pmap->pm_pdpt[i] = pa | PG_V;
1123 #endif
1124 }
1125
1126 pmap->pm_active = 0;
1127 TAILQ_INIT(&pmap->pm_pvlist);
1128 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1129 }
1130
1131 /*
1132 * Wire in kernel global address entries. To avoid a race condition
1133 * between pmap initialization and pmap_growkernel, this procedure
1134 * should be called after the vmspace is attached to the process
1135 * but before this pmap is activated.
1136 */
1137 void
1138 pmap_pinit2(pmap)
1139 struct pmap *pmap;
1140 {
1141 /* XXX: Remove this stub when no longer called */
1142 }
1143
1144 /*
1145 * this routine is called if the page table page is not
1146 * mapped correctly.
1147 */
1148 static vm_page_t
1149 _pmap_allocpte(pmap, ptepindex)
1150 pmap_t pmap;
1151 unsigned ptepindex;
1152 {
1153 vm_paddr_t ptepa;
1154 vm_page_t m;
1155
1156 /*
1157 * Allocate a page table page.
1158 */
1159 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ |
1160 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) {
1161 VM_WAIT;
1162 /*
1163 * Indicate the need to retry. While waiting, the page table
1164 * page may have been allocated.
1165 */
1166 return (NULL);
1167 }
1168 if ((m->flags & PG_ZERO) == 0)
1169 pmap_zero_page(m);
1170
1171 KASSERT(m->queue == PQ_NONE,
1172 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1173
1174 /*
1175 * Increment the hold count for the page table page
1176 * (denoting a new mapping.)
1177 */
1178 m->hold_count++;
1179
1180 /*
1181 * Map the pagetable page into the process address space, if
1182 * it isn't already there.
1183 */
1184
1185 pmap->pm_stats.resident_count++;
1186
1187 ptepa = VM_PAGE_TO_PHYS(m);
1188 pmap->pm_pdir[ptepindex] =
1189 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1190
1191 vm_page_lock_queues();
1192 vm_page_flag_clear(m, PG_ZERO);
1193 vm_page_wakeup(m);
1194 vm_page_unlock_queues();
1195
1196 return m;
1197 }
1198
1199 static vm_page_t
1200 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1201 {
1202 unsigned ptepindex;
1203 pd_entry_t ptepa;
1204 vm_page_t m;
1205
1206 /*
1207 * Calculate pagetable page index
1208 */
1209 ptepindex = va >> PDRSHIFT;
1210 retry:
1211 /*
1212 * Get the page directory entry
1213 */
1214 ptepa = pmap->pm_pdir[ptepindex];
1215
1216 /*
1217 * This supports switching from a 4MB page to a
1218 * normal 4K page.
1219 */
1220 if (ptepa & PG_PS) {
1221 pmap->pm_pdir[ptepindex] = 0;
1222 ptepa = 0;
1223 pmap_invalidate_all(kernel_pmap);
1224 }
1225
1226 /*
1227 * If the page table page is mapped, we just increment the
1228 * hold count, and activate it.
1229 */
1230 if (ptepa) {
1231 m = PHYS_TO_VM_PAGE(ptepa);
1232 m->hold_count++;
1233 } else {
1234 /*
1235 * Here if the pte page isn't mapped, or if it has
1236 * been deallocated.
1237 */
1238 m = _pmap_allocpte(pmap, ptepindex);
1239 if (m == NULL)
1240 goto retry;
1241 }
1242 return (m);
1243 }
1244
1245
1246 /***************************************************
1247 * Pmap allocation/deallocation routines.
1248 ***************************************************/
1249
1250 #ifdef SMP
1251 /*
1252 * Deal with a SMP shootdown of other users of the pmap that we are
1253 * trying to dispose of. This can be a bit hairy.
1254 */
1255 static u_int *lazymask;
1256 static u_int lazyptd;
1257 static volatile u_int lazywait;
1258
1259 void pmap_lazyfix_action(void);
1260
1261 void
1262 pmap_lazyfix_action(void)
1263 {
1264 u_int mymask = PCPU_GET(cpumask);
1265
1266 if (rcr3() == lazyptd)
1267 load_cr3(PCPU_GET(curpcb)->pcb_cr3);
1268 atomic_clear_int(lazymask, mymask);
1269 atomic_store_rel_int(&lazywait, 1);
1270 }
1271
1272 static void
1273 pmap_lazyfix_self(u_int mymask)
1274 {
1275
1276 if (rcr3() == lazyptd)
1277 load_cr3(PCPU_GET(curpcb)->pcb_cr3);
1278 atomic_clear_int(lazymask, mymask);
1279 }
1280
1281
1282 static void
1283 pmap_lazyfix(pmap_t pmap)
1284 {
1285 u_int mymask = PCPU_GET(cpumask);
1286 u_int mask;
1287 register u_int spins;
1288
1289 while ((mask = pmap->pm_active) != 0) {
1290 spins = 50000000;
1291 mask = mask & -mask; /* Find least significant set bit */
1292 mtx_lock_spin(&lazypmap_lock);
1293 #ifdef PAE
1294 lazyptd = vtophys(pmap->pm_pdpt);
1295 #else
1296 lazyptd = vtophys(pmap->pm_pdir);
1297 #endif
1298 if (mask == mymask) {
1299 lazymask = &pmap->pm_active;
1300 pmap_lazyfix_self(mymask);
1301 } else {
1302 atomic_store_rel_int((u_int *)&lazymask,
1303 (u_int)&pmap->pm_active);
1304 atomic_store_rel_int(&lazywait, 0);
1305 ipi_selected(mask, IPI_LAZYPMAP);
1306 while (lazywait == 0) {
1307 ia32_pause();
1308 if (--spins == 0)
1309 break;
1310 }
1311 }
1312 mtx_unlock_spin(&lazypmap_lock);
1313 if (spins == 0)
1314 printf("pmap_lazyfix: spun for 50000000\n");
1315 }
1316 }
1317
1318 #else /* SMP */
1319
1320 /*
1321 * Cleaning up on uniprocessor is easy. For various reasons, we're
1322 * unlikely to have to even execute this code, including the fact
1323 * that the cleanup is deferred until the parent does a wait(2), which
1324 * means that another userland process has run.
1325 */
1326 static void
1327 pmap_lazyfix(pmap_t pmap)
1328 {
1329 u_int cr3;
1330
1331 cr3 = vtophys(pmap->pm_pdir);
1332 if (cr3 == rcr3()) {
1333 load_cr3(PCPU_GET(curpcb)->pcb_cr3);
1334 pmap->pm_active &= ~(PCPU_GET(cpumask));
1335 }
1336 }
1337 #endif /* SMP */
1338
1339 /*
1340 * Release any resources held by the given physical map.
1341 * Called when a pmap initialized by pmap_pinit is being released.
1342 * Should only be called if the map contains no valid mappings.
1343 */
1344 void
1345 pmap_release(pmap_t pmap)
1346 {
1347 vm_page_t m, ptdpg[NPGPTD];
1348 int i;
1349
1350 KASSERT(pmap->pm_stats.resident_count == 0,
1351 ("pmap_release: pmap resident count %ld != 0",
1352 pmap->pm_stats.resident_count));
1353
1354 pmap_lazyfix(pmap);
1355 mtx_lock_spin(&allpmaps_lock);
1356 LIST_REMOVE(pmap, pm_list);
1357 mtx_unlock_spin(&allpmaps_lock);
1358
1359 for (i = 0; i < NPGPTD; i++)
1360 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]);
1361
1362 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) *
1363 sizeof(*pmap->pm_pdir));
1364 #ifdef SMP
1365 pmap->pm_pdir[MPPTDI] = 0;
1366 #endif
1367
1368 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD);
1369
1370 vm_page_lock_queues();
1371 for (i = 0; i < NPGPTD; i++) {
1372 m = ptdpg[i];
1373 #ifdef PAE
1374 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME),
1375 ("pmap_release: got wrong ptd page"));
1376 #endif
1377 m->wire_count--;
1378 atomic_subtract_int(&cnt.v_wire_count, 1);
1379 vm_page_busy(m);
1380 vm_page_free_zero(m);
1381 }
1382 vm_page_unlock_queues();
1383 }
1384
1385 static int
1386 kvm_size(SYSCTL_HANDLER_ARGS)
1387 {
1388 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE;
1389
1390 return sysctl_handle_long(oidp, &ksize, 0, req);
1391 }
1392 SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD,
1393 0, 0, kvm_size, "IU", "Size of KVM");
1394
1395 static int
1396 kvm_free(SYSCTL_HANDLER_ARGS)
1397 {
1398 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end;
1399
1400 return sysctl_handle_long(oidp, &kfree, 0, req);
1401 }
1402 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD,
1403 0, 0, kvm_free, "IU", "Amount of KVM free");
1404
1405 /*
1406 * grow the number of kernel page table entries, if needed
1407 */
1408 void
1409 pmap_growkernel(vm_offset_t addr)
1410 {
1411 struct pmap *pmap;
1412 int s;
1413 vm_paddr_t ptppaddr;
1414 vm_page_t nkpg;
1415 pd_entry_t newpdir;
1416 pt_entry_t *pde;
1417
1418 s = splhigh();
1419 mtx_assert(&kernel_map->system_mtx, MA_OWNED);
1420 if (kernel_vm_end == 0) {
1421 kernel_vm_end = KERNBASE;
1422 nkpt = 0;
1423 while (pdir_pde(PTD, kernel_vm_end)) {
1424 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1425 nkpt++;
1426 }
1427 }
1428 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1429 while (kernel_vm_end < addr) {
1430 if (pdir_pde(PTD, kernel_vm_end)) {
1431 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1432 continue;
1433 }
1434
1435 /*
1436 * This index is bogus, but out of the way
1437 */
1438 nkpg = vm_page_alloc(NULL, nkpt,
1439 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
1440 if (!nkpg)
1441 panic("pmap_growkernel: no memory to grow kernel");
1442
1443 nkpt++;
1444
1445 pmap_zero_page(nkpg);
1446 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1447 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1448 pdir_pde(PTD, kernel_vm_end) = newpdir;
1449
1450 mtx_lock_spin(&allpmaps_lock);
1451 LIST_FOREACH(pmap, &allpmaps, pm_list) {
1452 pde = pmap_pde(pmap, kernel_vm_end);
1453 pde_store(pde, newpdir);
1454 }
1455 mtx_unlock_spin(&allpmaps_lock);
1456 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1457 }
1458 splx(s);
1459 }
1460
1461
1462 /***************************************************
1463 * page management routines.
1464 ***************************************************/
1465
1466 /*
1467 * free the pv_entry back to the free list
1468 */
1469 static PMAP_INLINE void
1470 free_pv_entry(pv_entry_t pv)
1471 {
1472 pv_entry_count--;
1473 uma_zfree(pvzone, pv);
1474 }
1475
1476 /*
1477 * get a new pv_entry, allocating a block from the system
1478 * when needed.
1479 * the memory allocation is performed bypassing the malloc code
1480 * because of the possibility of allocations at interrupt time.
1481 */
1482 static pv_entry_t
1483 get_pv_entry(void)
1484 {
1485 pv_entry_count++;
1486 if (pv_entry_high_water &&
1487 (pv_entry_count > pv_entry_high_water) &&
1488 (pmap_pagedaemon_waken == 0)) {
1489 pmap_pagedaemon_waken = 1;
1490 wakeup (&vm_pages_needed);
1491 }
1492 return uma_zalloc(pvzone, M_NOWAIT);
1493 }
1494
1495 /*
1496 * If it is the first entry on the list, it is actually
1497 * in the header and we must copy the following entry up
1498 * to the header. Otherwise we must search the list for
1499 * the entry. In either case we free the now unused entry.
1500 */
1501
1502 static int
1503 pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va)
1504 {
1505 pv_entry_t pv;
1506 int rtval;
1507 int s;
1508
1509 s = splvm();
1510 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1511 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1512 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1513 if (pmap == pv->pv_pmap && va == pv->pv_va)
1514 break;
1515 }
1516 } else {
1517 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1518 if (va == pv->pv_va)
1519 break;
1520 }
1521 }
1522
1523 rtval = 0;
1524 if (pv) {
1525 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1526 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1527 m->md.pv_list_count--;
1528 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1529 vm_page_flag_clear(m, PG_WRITEABLE);
1530
1531 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1532 free_pv_entry(pv);
1533 }
1534
1535 splx(s);
1536 return rtval;
1537 }
1538
1539 /*
1540 * Create a pv entry for page at pa for
1541 * (pmap, va).
1542 */
1543 static void
1544 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1545 {
1546
1547 int s;
1548 pv_entry_t pv;
1549
1550 s = splvm();
1551 pv = get_pv_entry();
1552 pv->pv_va = va;
1553 pv->pv_pmap = pmap;
1554 pv->pv_ptem = mpte;
1555
1556 vm_page_lock_queues();
1557 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1558 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1559 m->md.pv_list_count++;
1560
1561 vm_page_unlock_queues();
1562 splx(s);
1563 }
1564
1565 /*
1566 * pmap_remove_pte: do the things to unmap a page in a process
1567 */
1568 static int
1569 pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va)
1570 {
1571 pt_entry_t oldpte;
1572 vm_page_t m, mpte;
1573
1574 oldpte = pte_load_clear(ptq);
1575 if (oldpte & PG_W)
1576 pmap->pm_stats.wired_count -= 1;
1577 /*
1578 * Machines that don't support invlpg, also don't support
1579 * PG_G.
1580 */
1581 if (oldpte & PG_G)
1582 pmap_invalidate_page(kernel_pmap, va);
1583 pmap->pm_stats.resident_count -= 1;
1584 if (oldpte & PG_MANAGED) {
1585 m = PHYS_TO_VM_PAGE(oldpte);
1586 if (oldpte & PG_M) {
1587 #if defined(PMAP_DIAGNOSTIC)
1588 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1589 printf(
1590 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1591 va, oldpte);
1592 }
1593 #endif
1594 if (pmap_track_modified(va))
1595 vm_page_dirty(m);
1596 }
1597 if (oldpte & PG_A)
1598 vm_page_flag_set(m, PG_REFERENCED);
1599 return pmap_remove_entry(pmap, m, va);
1600 } else {
1601 mpte = PHYS_TO_VM_PAGE(*pmap_pde(pmap, va));
1602 return pmap_unuse_pt(pmap, va, mpte);
1603 }
1604 }
1605
1606 /*
1607 * Remove a single page from a process address space
1608 */
1609 static void
1610 pmap_remove_page(pmap_t pmap, vm_offset_t va)
1611 {
1612 pt_entry_t *pte;
1613
1614 if ((pte = pmap_pte(pmap, va)) == NULL || *pte == 0)
1615 return;
1616 pmap_remove_pte(pmap, pte, va);
1617 pmap_invalidate_page(pmap, va);
1618 }
1619
1620 /*
1621 * Remove the given range of addresses from the specified map.
1622 *
1623 * It is assumed that the start and end are properly
1624 * rounded to the page size.
1625 */
1626 void
1627 pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1628 {
1629 vm_offset_t pdnxt;
1630 pd_entry_t ptpaddr;
1631 pt_entry_t *pte;
1632 int anyvalid;
1633
1634 if (pmap == NULL)
1635 return;
1636
1637 if (pmap->pm_stats.resident_count == 0)
1638 return;
1639
1640 /*
1641 * special handling of removing one page. a very
1642 * common operation and easy to short circuit some
1643 * code.
1644 */
1645 if ((sva + PAGE_SIZE == eva) &&
1646 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1647 pmap_remove_page(pmap, sva);
1648 return;
1649 }
1650
1651 anyvalid = 0;
1652
1653 for (; sva < eva; sva = pdnxt) {
1654 unsigned pdirindex;
1655
1656 /*
1657 * Calculate index for next page table.
1658 */
1659 pdnxt = (sva + NBPDR) & ~PDRMASK;
1660 if (pmap->pm_stats.resident_count == 0)
1661 break;
1662
1663 pdirindex = sva >> PDRSHIFT;
1664 ptpaddr = pmap->pm_pdir[pdirindex];
1665
1666 /*
1667 * Weed out invalid mappings. Note: we assume that the page
1668 * directory table is always allocated, and in kernel virtual.
1669 */
1670 if (ptpaddr == 0)
1671 continue;
1672
1673 /*
1674 * Check for large page.
1675 */
1676 if ((ptpaddr & PG_PS) != 0) {
1677 pmap->pm_pdir[pdirindex] = 0;
1678 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1679 anyvalid = 1;
1680 continue;
1681 }
1682
1683 /*
1684 * Limit our scan to either the end of the va represented
1685 * by the current page table page, or to the end of the
1686 * range being removed.
1687 */
1688 if (pdnxt > eva)
1689 pdnxt = eva;
1690
1691 for (; sva != pdnxt; sva += PAGE_SIZE) {
1692 if ((pte = pmap_pte(pmap, sva)) == NULL ||
1693 *pte == 0)
1694 continue;
1695 anyvalid = 1;
1696 if (pmap_remove_pte(pmap, pte, sva))
1697 break;
1698 }
1699 }
1700
1701 if (anyvalid)
1702 pmap_invalidate_all(pmap);
1703 }
1704
1705 /*
1706 * Routine: pmap_remove_all
1707 * Function:
1708 * Removes this physical page from
1709 * all physical maps in which it resides.
1710 * Reflects back modify bits to the pager.
1711 *
1712 * Notes:
1713 * Original versions of this routine were very
1714 * inefficient because they iteratively called
1715 * pmap_remove (slow...)
1716 */
1717
1718 void
1719 pmap_remove_all(vm_page_t m)
1720 {
1721 register pv_entry_t pv;
1722 pt_entry_t *pte, tpte;
1723 int s;
1724
1725 #if defined(PMAP_DIAGNOSTIC)
1726 /*
1727 * XXX This makes pmap_remove_all() illegal for non-managed pages!
1728 */
1729 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1730 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x",
1731 VM_PAGE_TO_PHYS(m));
1732 }
1733 #endif
1734 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1735 s = splvm();
1736 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1737 pv->pv_pmap->pm_stats.resident_count--;
1738 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1739 tpte = pte_load_clear(pte);
1740 if (tpte & PG_W)
1741 pv->pv_pmap->pm_stats.wired_count--;
1742 if (tpte & PG_A)
1743 vm_page_flag_set(m, PG_REFERENCED);
1744
1745 /*
1746 * Update the vm_page_t clean and reference bits.
1747 */
1748 if (tpte & PG_M) {
1749 #if defined(PMAP_DIAGNOSTIC)
1750 if (pmap_nw_modified((pt_entry_t) tpte)) {
1751 printf(
1752 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1753 pv->pv_va, tpte);
1754 }
1755 #endif
1756 if (pmap_track_modified(pv->pv_va))
1757 vm_page_dirty(m);
1758 }
1759 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
1760 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1761 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1762 m->md.pv_list_count--;
1763 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
1764 free_pv_entry(pv);
1765 }
1766 vm_page_flag_clear(m, PG_WRITEABLE);
1767 splx(s);
1768 }
1769
1770 /*
1771 * Set the physical protection on the
1772 * specified range of this map as requested.
1773 */
1774 void
1775 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1776 {
1777 vm_offset_t pdnxt;
1778 pd_entry_t ptpaddr;
1779 int anychanged;
1780
1781 if (pmap == NULL)
1782 return;
1783
1784 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1785 pmap_remove(pmap, sva, eva);
1786 return;
1787 }
1788
1789 if (prot & VM_PROT_WRITE)
1790 return;
1791
1792 anychanged = 0;
1793
1794 for (; sva < eva; sva = pdnxt) {
1795 unsigned pdirindex;
1796
1797 pdnxt = (sva + NBPDR) & ~PDRMASK;
1798
1799 pdirindex = sva >> PDRSHIFT;
1800 ptpaddr = pmap->pm_pdir[pdirindex];
1801
1802 /*
1803 * Weed out invalid mappings. Note: we assume that the page
1804 * directory table is always allocated, and in kernel virtual.
1805 */
1806 if (ptpaddr == 0)
1807 continue;
1808
1809 /*
1810 * Check for large page.
1811 */
1812 if ((ptpaddr & PG_PS) != 0) {
1813 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1814 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1815 anychanged = 1;
1816 continue;
1817 }
1818
1819 if (pdnxt > eva)
1820 pdnxt = eva;
1821
1822 for (; sva != pdnxt; sva += PAGE_SIZE) {
1823 pt_entry_t pbits;
1824 pt_entry_t *pte;
1825 vm_page_t m;
1826
1827 if ((pte = pmap_pte(pmap, sva)) == NULL)
1828 continue;
1829 pbits = *pte;
1830 if (pbits & PG_MANAGED) {
1831 m = NULL;
1832 if (pbits & PG_A) {
1833 m = PHYS_TO_VM_PAGE(pbits);
1834 vm_page_flag_set(m, PG_REFERENCED);
1835 pbits &= ~PG_A;
1836 }
1837 if ((pbits & PG_M) != 0 &&
1838 pmap_track_modified(sva)) {
1839 if (m == NULL)
1840 m = PHYS_TO_VM_PAGE(pbits);
1841 vm_page_dirty(m);
1842 pbits &= ~PG_M;
1843 }
1844 }
1845
1846 pbits &= ~PG_RW;
1847
1848 if (pbits != *pte) {
1849 pte_store(pte, pbits);
1850 anychanged = 1;
1851 }
1852 }
1853 }
1854 if (anychanged)
1855 pmap_invalidate_all(pmap);
1856 }
1857
1858 /*
1859 * Insert the given physical page (p) at
1860 * the specified virtual address (v) in the
1861 * target physical map with the protection requested.
1862 *
1863 * If specified, the page will be wired down, meaning
1864 * that the related pte can not be reclaimed.
1865 *
1866 * NB: This is the only routine which MAY NOT lazy-evaluate
1867 * or lose information. That is, this routine must actually
1868 * insert this page into the given map NOW.
1869 */
1870 void
1871 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1872 boolean_t wired)
1873 {
1874 vm_paddr_t pa;
1875 register pt_entry_t *pte;
1876 vm_paddr_t opa;
1877 pt_entry_t origpte, newpte;
1878 vm_page_t mpte;
1879
1880 if (pmap == NULL)
1881 return;
1882
1883 va &= PG_FRAME;
1884 #ifdef PMAP_DIAGNOSTIC
1885 if (va > VM_MAX_KERNEL_ADDRESS)
1886 panic("pmap_enter: toobig");
1887 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1888 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1889 #endif
1890
1891 mpte = NULL;
1892 /*
1893 * In the case that a page table page is not
1894 * resident, we are creating it here.
1895 */
1896 if (va < VM_MAXUSER_ADDRESS) {
1897 mpte = pmap_allocpte(pmap, va);
1898 }
1899 #if 0 && defined(PMAP_DIAGNOSTIC)
1900 else {
1901 pd_entry_t *pdeaddr = pmap_pde(pmap, va);
1902 origpte = *pdeaddr;
1903 if ((origpte & PG_V) == 0) {
1904 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n",
1905 pmap->pm_pdir[PTDPTDI], origpte, va);
1906 }
1907 }
1908 #endif
1909
1910 pte = pmap_pte(pmap, va);
1911
1912 /*
1913 * Page Directory table entry not valid, we need a new PT page
1914 */
1915 if (pte == NULL) {
1916 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n",
1917 (uintmax_t)pmap->pm_pdir[PTDPTDI], va);
1918 }
1919
1920 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
1921 origpte = *pte;
1922 opa = origpte & PG_FRAME;
1923
1924 if (origpte & PG_PS) {
1925 /*
1926 * Yes, I know this will truncate upper address bits for PAE,
1927 * but I'm actually more interested in the lower bits
1928 */
1929 printf("pmap_enter: va %p, pte %p, origpte %p\n",
1930 (void *)va, (void *)pte, (void *)(uintptr_t)origpte);
1931 panic("pmap_enter: attempted pmap_enter on 4MB page");
1932 }
1933
1934 /*
1935 * Mapping has not changed, must be protection or wiring change.
1936 */
1937 if (origpte && (opa == pa)) {
1938 /*
1939 * Wiring change, just update stats. We don't worry about
1940 * wiring PT pages as they remain resident as long as there
1941 * are valid mappings in them. Hence, if a user page is wired,
1942 * the PT page will be also.
1943 */
1944 if (wired && ((origpte & PG_W) == 0))
1945 pmap->pm_stats.wired_count++;
1946 else if (!wired && (origpte & PG_W))
1947 pmap->pm_stats.wired_count--;
1948
1949 #if defined(PMAP_DIAGNOSTIC)
1950 if (pmap_nw_modified((pt_entry_t) origpte)) {
1951 printf(
1952 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1953 va, origpte);
1954 }
1955 #endif
1956
1957 /*
1958 * Remove extra pte reference
1959 */
1960 if (mpte)
1961 mpte->hold_count--;
1962
1963 if ((prot & VM_PROT_WRITE) && (origpte & PG_V)) {
1964 if ((origpte & PG_RW) == 0) {
1965 pte_store(pte, origpte | PG_RW);
1966 pmap_invalidate_page(pmap, va);
1967 }
1968 return;
1969 }
1970
1971 /*
1972 * We might be turning off write access to the page,
1973 * so we go ahead and sense modify status.
1974 */
1975 if (origpte & PG_MANAGED) {
1976 if ((origpte & PG_M) && pmap_track_modified(va)) {
1977 vm_page_t om;
1978 om = PHYS_TO_VM_PAGE(opa);
1979 vm_page_dirty(om);
1980 }
1981 pa |= PG_MANAGED;
1982 }
1983 goto validate;
1984 }
1985 /*
1986 * Mapping has changed, invalidate old range and fall through to
1987 * handle validating new mapping.
1988 */
1989 if (opa) {
1990 int err;
1991 vm_page_lock_queues();
1992 err = pmap_remove_pte(pmap, pte, va);
1993 vm_page_unlock_queues();
1994 if (err)
1995 panic("pmap_enter: pte vanished, va: 0x%x", va);
1996 }
1997
1998 /*
1999 * Enter on the PV list if part of our managed memory. Note that we
2000 * raise IPL while manipulating pv_table since pmap_enter can be
2001 * called at interrupt time.
2002 */
2003 if (pmap_initialized &&
2004 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2005 pmap_insert_entry(pmap, va, mpte, m);
2006 pa |= PG_MANAGED;
2007 }
2008
2009 /*
2010 * Increment counters
2011 */
2012 pmap->pm_stats.resident_count++;
2013 if (wired)
2014 pmap->pm_stats.wired_count++;
2015
2016 validate:
2017 /*
2018 * Now validate mapping with desired protection/wiring.
2019 */
2020 newpte = (pt_entry_t)(pa | pte_prot(pmap, prot) | PG_V);
2021
2022 if (wired)
2023 newpte |= PG_W;
2024 if (va < VM_MAXUSER_ADDRESS)
2025 newpte |= PG_U;
2026 if (pmap == kernel_pmap)
2027 newpte |= pgeflag;
2028
2029 /*
2030 * if the mapping or permission bits are different, we need
2031 * to update the pte.
2032 */
2033 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2034 pte_store(pte, newpte | PG_A);
2035 /*if (origpte)*/ {
2036 pmap_invalidate_page(pmap, va);
2037 }
2038 }
2039 }
2040
2041 /*
2042 * this code makes some *MAJOR* assumptions:
2043 * 1. Current pmap & pmap exists.
2044 * 2. Not wired.
2045 * 3. Read access.
2046 * 4. No page table pages.
2047 * 5. Tlbflush is deferred to calling procedure.
2048 * 6. Page IS managed.
2049 * but is *MUCH* faster than pmap_enter...
2050 */
2051
2052 vm_page_t
2053 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
2054 {
2055 pt_entry_t *pte;
2056 vm_paddr_t pa;
2057
2058 /*
2059 * In the case that a page table page is not
2060 * resident, we are creating it here.
2061 */
2062 if (va < VM_MAXUSER_ADDRESS) {
2063 unsigned ptepindex;
2064 pd_entry_t ptepa;
2065
2066 /*
2067 * Calculate pagetable page index
2068 */
2069 ptepindex = va >> PDRSHIFT;
2070 if (mpte && (mpte->pindex == ptepindex)) {
2071 mpte->hold_count++;
2072 } else {
2073 retry:
2074 /*
2075 * Get the page directory entry
2076 */
2077 ptepa = pmap->pm_pdir[ptepindex];
2078
2079 /*
2080 * If the page table page is mapped, we just increment
2081 * the hold count, and activate it.
2082 */
2083 if (ptepa) {
2084 if (ptepa & PG_PS)
2085 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2086 mpte = PHYS_TO_VM_PAGE(ptepa);
2087 mpte->hold_count++;
2088 } else {
2089 mpte = _pmap_allocpte(pmap, ptepindex);
2090 if (mpte == NULL)
2091 goto retry;
2092 }
2093 }
2094 } else {
2095 mpte = NULL;
2096 }
2097
2098 /*
2099 * This call to vtopte makes the assumption that we are
2100 * entering the page into the current pmap. In order to support
2101 * quick entry into any pmap, one would likely use pmap_pte_quick.
2102 * But that isn't as quick as vtopte.
2103 */
2104 pte = vtopte(va);
2105 if (*pte) {
2106 if (mpte != NULL) {
2107 vm_page_lock_queues();
2108 pmap_unwire_pte_hold(pmap, mpte);
2109 vm_page_unlock_queues();
2110 }
2111 return 0;
2112 }
2113
2114 /*
2115 * Enter on the PV list if part of our managed memory. Note that we
2116 * raise IPL while manipulating pv_table since pmap_enter can be
2117 * called at interrupt time.
2118 */
2119 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
2120 pmap_insert_entry(pmap, va, mpte, m);
2121
2122 /*
2123 * Increment counters
2124 */
2125 pmap->pm_stats.resident_count++;
2126
2127 pa = VM_PAGE_TO_PHYS(m);
2128
2129 /*
2130 * Now validate mapping with RO protection
2131 */
2132 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2133 pte_store(pte, pa | PG_V | PG_U);
2134 else
2135 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED);
2136
2137 return mpte;
2138 }
2139
2140 /*
2141 * Make a temporary mapping for a physical address. This is only intended
2142 * to be used for panic dumps.
2143 */
2144 void *
2145 pmap_kenter_temporary(vm_offset_t pa, int i)
2146 {
2147 vm_offset_t va;
2148
2149 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE);
2150 pmap_kenter(va, pa);
2151 #ifndef I386_CPU
2152 invlpg(va);
2153 #else
2154 invltlb();
2155 #endif
2156 return ((void *)crashdumpmap);
2157 }
2158
2159 /*
2160 * This code maps large physical mmap regions into the
2161 * processor address space. Note that some shortcuts
2162 * are taken, but the code works.
2163 */
2164 void
2165 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr,
2166 vm_object_t object, vm_pindex_t pindex,
2167 vm_size_t size)
2168 {
2169 vm_page_t p;
2170
2171 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2172 KASSERT(object->type == OBJT_DEVICE,
2173 ("pmap_object_init_pt: non-device object"));
2174 if (pseflag &&
2175 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) {
2176 int i;
2177 vm_page_t m[1];
2178 unsigned int ptepindex;
2179 int npdes;
2180 pd_entry_t ptepa;
2181
2182 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)])
2183 return;
2184 retry:
2185 p = vm_page_lookup(object, pindex);
2186 if (p != NULL) {
2187 vm_page_lock_queues();
2188 if (vm_page_sleep_if_busy(p, FALSE, "init4p"))
2189 goto retry;
2190 } else {
2191 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
2192 if (p == NULL)
2193 return;
2194 m[0] = p;
2195
2196 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) {
2197 vm_page_lock_queues();
2198 vm_page_free(p);
2199 vm_page_unlock_queues();
2200 return;
2201 }
2202
2203 p = vm_page_lookup(object, pindex);
2204 vm_page_lock_queues();
2205 vm_page_wakeup(p);
2206 }
2207 vm_page_unlock_queues();
2208
2209 ptepa = VM_PAGE_TO_PHYS(p);
2210 if (ptepa & (NBPDR - 1))
2211 return;
2212
2213 p->valid = VM_PAGE_BITS_ALL;
2214
2215 pmap->pm_stats.resident_count += size >> PAGE_SHIFT;
2216 npdes = size >> PDRSHIFT;
2217 for(i = 0; i < npdes; i++) {
2218 pde_store(&pmap->pm_pdir[ptepindex],
2219 ptepa | PG_U | PG_RW | PG_V | PG_PS);
2220 ptepa += NBPDR;
2221 ptepindex += 1;
2222 }
2223 pmap_invalidate_all(pmap);
2224 }
2225 }
2226
2227 /*
2228 * Routine: pmap_change_wiring
2229 * Function: Change the wiring attribute for a map/virtual-address
2230 * pair.
2231 * In/out conditions:
2232 * The mapping must already exist in the pmap.
2233 */
2234 void
2235 pmap_change_wiring(pmap, va, wired)
2236 register pmap_t pmap;
2237 vm_offset_t va;
2238 boolean_t wired;
2239 {
2240 register pt_entry_t *pte;
2241
2242 if (pmap == NULL)
2243 return;
2244
2245 pte = pmap_pte(pmap, va);
2246
2247 if (wired && !pmap_pte_w(pte))
2248 pmap->pm_stats.wired_count++;
2249 else if (!wired && pmap_pte_w(pte))
2250 pmap->pm_stats.wired_count--;
2251
2252 /*
2253 * Wiring is not a hardware characteristic so there is no need to
2254 * invalidate TLB.
2255 */
2256 pmap_pte_set_w(pte, wired);
2257 }
2258
2259
2260
2261 /*
2262 * Copy the range specified by src_addr/len
2263 * from the source map to the range dst_addr/len
2264 * in the destination map.
2265 *
2266 * This routine is only advisory and need not do anything.
2267 */
2268
2269 void
2270 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len,
2271 vm_offset_t src_addr)
2272 {
2273 vm_offset_t addr;
2274 vm_offset_t end_addr = src_addr + len;
2275 vm_offset_t pdnxt;
2276 vm_page_t m;
2277
2278 if (dst_addr != src_addr)
2279 return;
2280
2281 if (!pmap_is_current(src_pmap))
2282 return;
2283
2284 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2285 pt_entry_t *src_pte, *dst_pte;
2286 vm_page_t dstmpte, srcmpte;
2287 pd_entry_t srcptepaddr;
2288 unsigned ptepindex;
2289
2290 if (addr >= UPT_MIN_ADDRESS)
2291 panic("pmap_copy: invalid to pmap_copy page tables\n");
2292
2293 /*
2294 * Don't let optional prefaulting of pages make us go
2295 * way below the low water mark of free pages or way
2296 * above high water mark of used pv entries.
2297 */
2298 if (cnt.v_free_count < cnt.v_free_reserved ||
2299 pv_entry_count > pv_entry_high_water)
2300 break;
2301
2302 pdnxt = (addr + NBPDR) & ~PDRMASK;
2303 ptepindex = addr >> PDRSHIFT;
2304
2305 srcptepaddr = src_pmap->pm_pdir[ptepindex];
2306 if (srcptepaddr == 0)
2307 continue;
2308
2309 if (srcptepaddr & PG_PS) {
2310 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2311 dst_pmap->pm_pdir[ptepindex] = srcptepaddr;
2312 dst_pmap->pm_stats.resident_count +=
2313 NBPDR / PAGE_SIZE;
2314 }
2315 continue;
2316 }
2317
2318 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr);
2319 if (srcmpte->hold_count == 0 || (srcmpte->flags & PG_BUSY))
2320 continue;
2321
2322 if (pdnxt > end_addr)
2323 pdnxt = end_addr;
2324
2325 src_pte = vtopte(addr);
2326 while (addr < pdnxt) {
2327 pt_entry_t ptetemp;
2328 ptetemp = *src_pte;
2329 /*
2330 * we only virtual copy managed pages
2331 */
2332 if ((ptetemp & PG_MANAGED) != 0) {
2333 /*
2334 * We have to check after allocpte for the
2335 * pte still being around... allocpte can
2336 * block.
2337 */
2338 dstmpte = pmap_allocpte(dst_pmap, addr);
2339 dst_pte = pmap_pte(dst_pmap, addr);
2340 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2341 /*
2342 * Clear the modified and
2343 * accessed (referenced) bits
2344 * during the copy.
2345 */
2346 m = PHYS_TO_VM_PAGE(ptetemp);
2347 *dst_pte = ptetemp & ~(PG_M | PG_A);
2348 dst_pmap->pm_stats.resident_count++;
2349 pmap_insert_entry(dst_pmap, addr,
2350 dstmpte, m);
2351 } else {
2352 vm_page_lock_queues();
2353 pmap_unwire_pte_hold(dst_pmap, dstmpte);
2354 vm_page_unlock_queues();
2355 }
2356 if (dstmpte->hold_count >= srcmpte->hold_count)
2357 break;
2358 }
2359 addr += PAGE_SIZE;
2360 src_pte++;
2361 }
2362 }
2363 }
2364
2365 #ifdef SMP
2366
2367 /*
2368 * pmap_zpi_switchout*()
2369 *
2370 * These functions allow us to avoid doing IPIs alltogether in certain
2371 * temporary page-mapping situations (page zeroing). Instead to deal
2372 * with being preempted and moved onto a different cpu we invalidate
2373 * the page when the scheduler switches us in. This does not occur
2374 * very often so we remain relatively optimal with very little effort.
2375 */
2376 static void
2377 pmap_zpi_switchout12(void)
2378 {
2379 invlpg((u_int)CADDR1);
2380 invlpg((u_int)CADDR2);
2381 }
2382
2383 static void
2384 pmap_zpi_switchout2(void)
2385 {
2386 invlpg((u_int)CADDR2);
2387 }
2388
2389 static void
2390 pmap_zpi_switchout3(void)
2391 {
2392 invlpg((u_int)CADDR3);
2393 }
2394
2395 #endif
2396
2397 static __inline void
2398 pagezero(void *page)
2399 {
2400 #if defined(I686_CPU)
2401 if (cpu_class == CPUCLASS_686) {
2402 #if defined(CPU_ENABLE_SSE)
2403 if (cpu_feature & CPUID_SSE2)
2404 sse2_pagezero(page);
2405 else
2406 #endif
2407 i686_pagezero(page);
2408 } else
2409 #endif
2410 bzero(page, PAGE_SIZE);
2411 }
2412
2413 static __inline void
2414 invlcaddr(void *caddr)
2415 {
2416 #ifdef I386_CPU
2417 invltlb();
2418 #else
2419 invlpg((u_int)caddr);
2420 #endif
2421 }
2422
2423 /*
2424 * pmap_zero_page zeros the specified hardware page by mapping
2425 * the page into KVM and using bzero to clear its contents.
2426 */
2427 void
2428 pmap_zero_page(vm_page_t m)
2429 {
2430
2431 mtx_lock(&CMAPCADDR12_lock);
2432 if (*CMAP2)
2433 panic("pmap_zero_page: CMAP2 busy");
2434 #ifdef SMP
2435 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout2;
2436 #endif
2437 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M;
2438 #ifdef SMP
2439 invlpg((u_int)CADDR2);
2440 #endif
2441 pagezero(CADDR2);
2442 *CMAP2 = 0;
2443 invlcaddr(CADDR2);
2444 #ifdef SMP
2445 curthread->td_pcb->pcb_switchout = NULL;
2446 #endif
2447 mtx_unlock(&CMAPCADDR12_lock);
2448 }
2449
2450 /*
2451 * pmap_zero_page_area zeros the specified hardware page by mapping
2452 * the page into KVM and using bzero to clear its contents.
2453 *
2454 * off and size may not cover an area beyond a single hardware page.
2455 */
2456 void
2457 pmap_zero_page_area(vm_page_t m, int off, int size)
2458 {
2459
2460 mtx_lock(&CMAPCADDR12_lock);
2461 if (*CMAP2)
2462 panic("pmap_zero_page: CMAP2 busy");
2463 #ifdef SMP
2464 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout2;
2465 #endif
2466 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M;
2467 #ifdef SMP
2468 invlpg((u_int)CADDR2);
2469 #endif
2470 if (off == 0 && size == PAGE_SIZE)
2471 pagezero(CADDR2);
2472 else
2473 bzero((char *)CADDR2 + off, size);
2474 *CMAP2 = 0;
2475 invlcaddr(CADDR2);
2476 #ifdef SMP
2477 curthread->td_pcb->pcb_switchout = NULL;
2478 #endif
2479 mtx_unlock(&CMAPCADDR12_lock);
2480 }
2481
2482 /*
2483 * pmap_zero_page_idle zeros the specified hardware page by mapping
2484 * the page into KVM and using bzero to clear its contents. This
2485 * is intended to be called from the vm_pagezero process only and
2486 * outside of Giant.
2487 */
2488 void
2489 pmap_zero_page_idle(vm_page_t m)
2490 {
2491
2492 if (*CMAP3)
2493 panic("pmap_zero_page: CMAP3 busy");
2494 #ifdef SMP
2495 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout3;
2496 #endif
2497 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M;
2498 #ifdef SMP
2499 invlpg((u_int)CADDR3);
2500 #endif
2501 pagezero(CADDR3);
2502 *CMAP3 = 0;
2503 invlcaddr(CADDR3);
2504 #ifdef SMP
2505 curthread->td_pcb->pcb_switchout = NULL;
2506 #endif
2507 }
2508
2509 /*
2510 * pmap_copy_page copies the specified (machine independent)
2511 * page by mapping the page into virtual memory and using
2512 * bcopy to copy the page, one machine dependent page at a
2513 * time.
2514 */
2515 void
2516 pmap_copy_page(vm_page_t src, vm_page_t dst)
2517 {
2518
2519 mtx_lock(&CMAPCADDR12_lock);
2520 if (*CMAP1)
2521 panic("pmap_copy_page: CMAP1 busy");
2522 if (*CMAP2)
2523 panic("pmap_copy_page: CMAP2 busy");
2524 #ifdef SMP
2525 curthread->td_pcb->pcb_switchout = pmap_zpi_switchout12;
2526 #endif
2527 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A;
2528 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M;
2529 #ifdef SMP
2530 invlpg((u_int)CADDR1);
2531 invlpg((u_int)CADDR2);
2532 #endif
2533 bcopy(CADDR1, CADDR2, PAGE_SIZE);
2534 *CMAP1 = 0;
2535 *CMAP2 = 0;
2536 #ifdef I386_CPU
2537 invltlb();
2538 #else
2539 invlpg((u_int)CADDR1);
2540 invlpg((u_int)CADDR2);
2541 #endif
2542 #ifdef SMP
2543 curthread->td_pcb->pcb_switchout = NULL;
2544 #endif
2545 mtx_unlock(&CMAPCADDR12_lock);
2546 }
2547
2548 /*
2549 * Returns true if the pmap's pv is one of the first
2550 * 16 pvs linked to from this page. This count may
2551 * be changed upwards or downwards in the future; it
2552 * is only necessary that true be returned for a small
2553 * subset of pmaps for proper page aging.
2554 */
2555 boolean_t
2556 pmap_page_exists_quick(pmap, m)
2557 pmap_t pmap;
2558 vm_page_t m;
2559 {
2560 pv_entry_t pv;
2561 int loops = 0;
2562 int s;
2563
2564 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2565 return FALSE;
2566
2567 s = splvm();
2568 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2569 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2570 if (pv->pv_pmap == pmap) {
2571 splx(s);
2572 return TRUE;
2573 }
2574 loops++;
2575 if (loops >= 16)
2576 break;
2577 }
2578 splx(s);
2579 return (FALSE);
2580 }
2581
2582 #define PMAP_REMOVE_PAGES_CURPROC_ONLY
2583 /*
2584 * Remove all pages from specified address space
2585 * this aids process exit speeds. Also, this code
2586 * is special cased for current process only, but
2587 * can have the more generic (and slightly slower)
2588 * mode enabled. This is much faster than pmap_remove
2589 * in the case of running down an entire address space.
2590 */
2591 void
2592 pmap_remove_pages(pmap, sva, eva)
2593 pmap_t pmap;
2594 vm_offset_t sva, eva;
2595 {
2596 pt_entry_t *pte, tpte;
2597 vm_page_t m;
2598 pv_entry_t pv, npv;
2599 int s;
2600
2601 #ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY
2602 if (!curthread || (pmap != vmspace_pmap(curthread->td_proc->p_vmspace))) {
2603 printf("warning: pmap_remove_pages called with non-current pmap\n");
2604 return;
2605 }
2606 #endif
2607 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2608 s = splvm();
2609 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2610
2611 if (pv->pv_va >= eva || pv->pv_va < sva) {
2612 npv = TAILQ_NEXT(pv, pv_plist);
2613 continue;
2614 }
2615
2616 #ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY
2617 pte = vtopte(pv->pv_va);
2618 #else
2619 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2620 #endif
2621 tpte = *pte;
2622
2623 if (tpte == 0) {
2624 printf("TPTE at %p IS ZERO @ VA %08x\n",
2625 pte, pv->pv_va);
2626 panic("bad pte");
2627 }
2628
2629 /*
2630 * We cannot remove wired pages from a process' mapping at this time
2631 */
2632 if (tpte & PG_W) {
2633 npv = TAILQ_NEXT(pv, pv_plist);
2634 continue;
2635 }
2636
2637 m = PHYS_TO_VM_PAGE(tpte);
2638 KASSERT(m->phys_addr == (tpte & PG_FRAME),
2639 ("vm_page_t %p phys_addr mismatch %016jx %016jx",
2640 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte));
2641
2642 KASSERT(m < &vm_page_array[vm_page_array_size],
2643 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte));
2644
2645 pv->pv_pmap->pm_stats.resident_count--;
2646
2647 pte_clear(pte);
2648
2649 /*
2650 * Update the vm_page_t clean and reference bits.
2651 */
2652 if (tpte & PG_M) {
2653 vm_page_dirty(m);
2654 }
2655
2656 npv = TAILQ_NEXT(pv, pv_plist);
2657 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2658
2659 m->md.pv_list_count--;
2660 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2661 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2662 vm_page_flag_clear(m, PG_WRITEABLE);
2663 }
2664
2665 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
2666 free_pv_entry(pv);
2667 }
2668 splx(s);
2669 pmap_invalidate_all(pmap);
2670 }
2671
2672 /*
2673 * pmap_is_modified:
2674 *
2675 * Return whether or not the specified physical page was modified
2676 * in any physical maps.
2677 */
2678 boolean_t
2679 pmap_is_modified(vm_page_t m)
2680 {
2681 pv_entry_t pv;
2682 pt_entry_t *pte;
2683 int s;
2684
2685 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2686 return FALSE;
2687
2688 s = splvm();
2689 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2690 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2691 /*
2692 * if the bit being tested is the modified bit, then
2693 * mark clean_map and ptes as never
2694 * modified.
2695 */
2696 if (!pmap_track_modified(pv->pv_va))
2697 continue;
2698 #if defined(PMAP_DIAGNOSTIC)
2699 if (!pv->pv_pmap) {
2700 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2701 continue;
2702 }
2703 #endif
2704 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2705 if (*pte & PG_M) {
2706 splx(s);
2707 return TRUE;
2708 }
2709 }
2710 splx(s);
2711 return (FALSE);
2712 }
2713
2714 /*
2715 * pmap_is_prefaultable:
2716 *
2717 * Return whether or not the specified virtual address is elgible
2718 * for prefault.
2719 */
2720 boolean_t
2721 pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr)
2722 {
2723 pt_entry_t *pte;
2724
2725 if ((*pmap_pde(pmap, addr)) == 0)
2726 return (FALSE);
2727 pte = vtopte(addr);
2728 if (*pte)
2729 return (FALSE);
2730 return (TRUE);
2731 }
2732
2733 /*
2734 * Clear the given bit in each of the given page's ptes.
2735 */
2736 static __inline void
2737 pmap_clear_ptes(vm_page_t m, int bit)
2738 {
2739 register pv_entry_t pv;
2740 pt_entry_t pbits, *pte;
2741 int s;
2742
2743 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) ||
2744 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0))
2745 return;
2746
2747 s = splvm();
2748 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2749 /*
2750 * Loop over all current mappings setting/clearing as appropos If
2751 * setting RO do we need to clear the VAC?
2752 */
2753 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2754 /*
2755 * don't write protect pager mappings
2756 */
2757 if (bit == PG_RW) {
2758 if (!pmap_track_modified(pv->pv_va))
2759 continue;
2760 }
2761
2762 #if defined(PMAP_DIAGNOSTIC)
2763 if (!pv->pv_pmap) {
2764 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2765 continue;
2766 }
2767 #endif
2768
2769 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2770 pbits = *pte;
2771 if (pbits & bit) {
2772 if (bit == PG_RW) {
2773 if (pbits & PG_M) {
2774 vm_page_dirty(m);
2775 }
2776 pte_store(pte, pbits & ~(PG_M|PG_RW));
2777 } else {
2778 pte_store(pte, pbits & ~bit);
2779 }
2780 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
2781 }
2782 }
2783 if (bit == PG_RW)
2784 vm_page_flag_clear(m, PG_WRITEABLE);
2785 splx(s);
2786 }
2787
2788 /*
2789 * pmap_page_protect:
2790 *
2791 * Lower the permission for all mappings to a given page.
2792 */
2793 void
2794 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2795 {
2796 if ((prot & VM_PROT_WRITE) == 0) {
2797 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2798 pmap_clear_ptes(m, PG_RW);
2799 } else {
2800 pmap_remove_all(m);
2801 }
2802 }
2803 }
2804
2805 /*
2806 * pmap_ts_referenced:
2807 *
2808 * Return a count of reference bits for a page, clearing those bits.
2809 * It is not necessary for every reference bit to be cleared, but it
2810 * is necessary that 0 only be returned when there are truly no
2811 * reference bits set.
2812 *
2813 * XXX: The exact number of bits to check and clear is a matter that
2814 * should be tested and standardized at some point in the future for
2815 * optimal aging of shared pages.
2816 */
2817 int
2818 pmap_ts_referenced(vm_page_t m)
2819 {
2820 register pv_entry_t pv, pvf, pvn;
2821 pt_entry_t *pte;
2822 pt_entry_t v;
2823 int s;
2824 int rtval = 0;
2825
2826 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2827 return (rtval);
2828
2829 s = splvm();
2830 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2831 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2832
2833 pvf = pv;
2834
2835 do {
2836 pvn = TAILQ_NEXT(pv, pv_list);
2837
2838 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2839
2840 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2841
2842 if (!pmap_track_modified(pv->pv_va))
2843 continue;
2844
2845 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2846
2847 if (pte && ((v = pte_load(pte)) & PG_A) != 0) {
2848 pte_store(pte, v & ~PG_A);
2849 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
2850
2851 rtval++;
2852 if (rtval > 4) {
2853 break;
2854 }
2855 }
2856 } while ((pv = pvn) != NULL && pv != pvf);
2857 }
2858 splx(s);
2859
2860 return (rtval);
2861 }
2862
2863 /*
2864 * Clear the modify bits on the specified physical page.
2865 */
2866 void
2867 pmap_clear_modify(vm_page_t m)
2868 {
2869 pmap_clear_ptes(m, PG_M);
2870 }
2871
2872 /*
2873 * pmap_clear_reference:
2874 *
2875 * Clear the reference bit on the specified physical page.
2876 */
2877 void
2878 pmap_clear_reference(vm_page_t m)
2879 {
2880 pmap_clear_ptes(m, PG_A);
2881 }
2882
2883 /*
2884 * Miscellaneous support routines follow
2885 */
2886
2887 static void
2888 i386_protection_init()
2889 {
2890 register int *kp, prot;
2891
2892 kp = protection_codes;
2893 for (prot = 0; prot < 8; prot++) {
2894 switch (prot) {
2895 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
2896 /*
2897 * Read access is also 0. There isn't any execute bit,
2898 * so just make it readable.
2899 */
2900 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
2901 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
2902 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
2903 *kp++ = 0;
2904 break;
2905 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
2906 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
2907 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
2908 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
2909 *kp++ = PG_RW;
2910 break;
2911 }
2912 }
2913 }
2914
2915 /*
2916 * Map a set of physical memory pages into the kernel virtual
2917 * address space. Return a pointer to where it is mapped. This
2918 * routine is intended to be used for mapping device memory,
2919 * NOT real memory.
2920 */
2921 void *
2922 pmap_mapdev(pa, size)
2923 vm_paddr_t pa;
2924 vm_size_t size;
2925 {
2926 vm_offset_t va, tmpva, offset;
2927
2928 offset = pa & PAGE_MASK;
2929 size = roundup(offset + size, PAGE_SIZE);
2930 pa = pa & PG_FRAME;
2931
2932 if (pa < KERNLOAD && pa + size <= KERNLOAD)
2933 va = KERNBASE + pa;
2934 else
2935 va = kmem_alloc_nofault(kernel_map, size);
2936 if (!va)
2937 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2938
2939 for (tmpva = va; size > 0; ) {
2940 pmap_kenter(tmpva, pa);
2941 size -= PAGE_SIZE;
2942 tmpva += PAGE_SIZE;
2943 pa += PAGE_SIZE;
2944 }
2945 pmap_invalidate_range(kernel_pmap, va, tmpva);
2946 return ((void *)(va + offset));
2947 }
2948
2949 void
2950 pmap_unmapdev(va, size)
2951 vm_offset_t va;
2952 vm_size_t size;
2953 {
2954 vm_offset_t base, offset, tmpva;
2955
2956 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD)
2957 return;
2958 base = va & PG_FRAME;
2959 offset = va & PAGE_MASK;
2960 size = roundup(offset + size, PAGE_SIZE);
2961 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE)
2962 pmap_kremove(tmpva);
2963 pmap_invalidate_range(kernel_pmap, va, tmpva);
2964 kmem_free(kernel_map, base, size);
2965 }
2966
2967 /*
2968 * perform the pmap work for mincore
2969 */
2970 int
2971 pmap_mincore(pmap, addr)
2972 pmap_t pmap;
2973 vm_offset_t addr;
2974 {
2975 pt_entry_t *ptep, pte;
2976 vm_page_t m;
2977 int val = 0;
2978
2979 ptep = pmap_pte(pmap, addr);
2980 if (ptep == 0) {
2981 return 0;
2982 }
2983
2984 if ((pte = *ptep) != 0) {
2985 vm_paddr_t pa;
2986
2987 val = MINCORE_INCORE;
2988 if ((pte & PG_MANAGED) == 0)
2989 return val;
2990
2991 pa = pte & PG_FRAME;
2992
2993 m = PHYS_TO_VM_PAGE(pa);
2994
2995 /*
2996 * Modified by us
2997 */
2998 if (pte & PG_M)
2999 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3000 else {
3001 /*
3002 * Modified by someone else
3003 */
3004 vm_page_lock_queues();
3005 if (m->dirty || pmap_is_modified(m))
3006 val |= MINCORE_MODIFIED_OTHER;
3007 vm_page_unlock_queues();
3008 }
3009 /*
3010 * Referenced by us
3011 */
3012 if (pte & PG_A)
3013 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3014 else {
3015 /*
3016 * Referenced by someone else
3017 */
3018 vm_page_lock_queues();
3019 if ((m->flags & PG_REFERENCED) ||
3020 pmap_ts_referenced(m)) {
3021 val |= MINCORE_REFERENCED_OTHER;
3022 vm_page_flag_set(m, PG_REFERENCED);
3023 }
3024 vm_page_unlock_queues();
3025 }
3026 }
3027 return val;
3028 }
3029
3030 void
3031 pmap_activate(struct thread *td)
3032 {
3033 struct proc *p = td->td_proc;
3034 pmap_t pmap, oldpmap;
3035 u_int32_t cr3;
3036
3037 critical_enter();
3038 pmap = vmspace_pmap(td->td_proc->p_vmspace);
3039 oldpmap = PCPU_GET(curpmap);
3040 #if defined(SMP)
3041 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask));
3042 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask));
3043 #else
3044 oldpmap->pm_active &= ~1;
3045 pmap->pm_active |= 1;
3046 #endif
3047 #ifdef PAE
3048 cr3 = vtophys(pmap->pm_pdpt);
3049 #else
3050 cr3 = vtophys(pmap->pm_pdir);
3051 #endif
3052 /* XXXKSE this is wrong.
3053 * pmap_activate is for the current thread on the current cpu
3054 */
3055 if (p->p_flag & P_SA) {
3056 /* Make sure all other cr3 entries are updated. */
3057 /* what if they are running? XXXKSE (maybe abort them) */
3058 FOREACH_THREAD_IN_PROC(p, td) {
3059 td->td_pcb->pcb_cr3 = cr3;
3060 }
3061 } else {
3062 td->td_pcb->pcb_cr3 = cr3;
3063 }
3064 load_cr3(cr3);
3065 PCPU_SET(curpmap, pmap);
3066 critical_exit();
3067 }
3068
3069 vm_offset_t
3070 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3071 {
3072
3073 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3074 return addr;
3075 }
3076
3077 addr = (addr + PDRMASK) & ~PDRMASK;
3078 return addr;
3079 }
3080
3081
3082 #if defined(PMAP_DEBUG)
3083 pmap_pid_dump(int pid)
3084 {
3085 pmap_t pmap;
3086 struct proc *p;
3087 int npte = 0;
3088 int index;
3089
3090 sx_slock(&allproc_lock);
3091 LIST_FOREACH(p, &allproc, p_list) {
3092 if (p->p_pid != pid)
3093 continue;
3094
3095 if (p->p_vmspace) {
3096 int i,j;
3097 index = 0;
3098 pmap = vmspace_pmap(p->p_vmspace);
3099 for (i = 0; i < NPDEPTD; i++) {
3100 pd_entry_t *pde;
3101 pt_entry_t *pte;
3102 vm_offset_t base = i << PDRSHIFT;
3103
3104 pde = &pmap->pm_pdir[i];
3105 if (pde && pmap_pde_v(pde)) {
3106 for (j = 0; j < NPTEPG; j++) {
3107 vm_offset_t va = base + (j << PAGE_SHIFT);
3108 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) {
3109 if (index) {
3110 index = 0;
3111 printf("\n");
3112 }
3113 sx_sunlock(&allproc_lock);
3114 return npte;
3115 }
3116 pte = pmap_pte(pmap, va);
3117 if (pte && pmap_pte_v(pte)) {
3118 pt_entry_t pa;
3119 vm_page_t m;
3120 pa = *pte;
3121 m = PHYS_TO_VM_PAGE(pa);
3122 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x",
3123 va, pa, m->hold_count, m->wire_count, m->flags);
3124 npte++;
3125 index++;
3126 if (index >= 2) {
3127 index = 0;
3128 printf("\n");
3129 } else {
3130 printf(" ");
3131 }
3132 }
3133 }
3134 }
3135 }
3136 }
3137 }
3138 sx_sunlock(&allproc_lock);
3139 return npte;
3140 }
3141 #endif
3142
3143 #if defined(DEBUG)
3144
3145 static void pads(pmap_t pm);
3146 void pmap_pvdump(vm_offset_t pa);
3147
3148 /* print address space of pmap*/
3149 static void
3150 pads(pm)
3151 pmap_t pm;
3152 {
3153 int i, j;
3154 vm_paddr_t va;
3155 pt_entry_t *ptep;
3156
3157 if (pm == kernel_pmap)
3158 return;
3159 for (i = 0; i < NPDEPTD; i++)
3160 if (pm->pm_pdir[i])
3161 for (j = 0; j < NPTEPG; j++) {
3162 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3163 if (pm == kernel_pmap && va < KERNBASE)
3164 continue;
3165 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS)
3166 continue;
3167 ptep = pmap_pte(pm, va);
3168 if (pmap_pte_v(ptep))
3169 printf("%x:%x ", va, *ptep);
3170 };
3171
3172 }
3173
3174 void
3175 pmap_pvdump(pa)
3176 vm_paddr_t pa;
3177 {
3178 pv_entry_t pv;
3179 vm_page_t m;
3180
3181 printf("pa %x", pa);
3182 m = PHYS_TO_VM_PAGE(pa);
3183 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3184 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);
3185 pads(pv->pv_pmap);
3186 }
3187 printf(" ");
3188 }
3189 #endif
Cache object: cac0b30ebf2ce76f4c4abc2c4960d89d
|