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/5.3/sys/i386/i386/pmap.c 136358 2004-10-10 19:08:00Z 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_cpu.h"
105 #include "opt_pmap.h"
106 #include "opt_msgbuf.h"
107 #include "opt_kstack_pages.h"
108
109 #include <sys/param.h>
110 #include <sys/systm.h>
111 #include <sys/kernel.h>
112 #include <sys/lock.h>
113 #include <sys/malloc.h>
114 #include <sys/mman.h>
115 #include <sys/msgbuf.h>
116 #include <sys/mutex.h>
117 #include <sys/proc.h>
118 #include <sys/sx.h>
119 #include <sys/user.h>
120 #include <sys/vmmeter.h>
121 #include <sys/sched.h>
122 #include <sys/sysctl.h>
123 #ifdef SMP
124 #include <sys/smp.h>
125 #endif
126
127 #include <vm/vm.h>
128 #include <vm/vm_param.h>
129 #include <vm/vm_kern.h>
130 #include <vm/vm_page.h>
131 #include <vm/vm_map.h>
132 #include <vm/vm_object.h>
133 #include <vm/vm_extern.h>
134 #include <vm/vm_pageout.h>
135 #include <vm/vm_pager.h>
136 #include <vm/uma.h>
137
138 #include <machine/cpu.h>
139 #include <machine/cputypes.h>
140 #include <machine/md_var.h>
141 #include <machine/specialreg.h>
142 #ifdef SMP
143 #include <machine/smp.h>
144 #endif
145
146 #if !defined(CPU_ENABLE_SSE) && defined(I686_CPU)
147 #define CPU_ENABLE_SSE
148 #endif
149 #if defined(CPU_DISABLE_SSE)
150 #undef CPU_ENABLE_SSE
151 #endif
152
153 #ifndef PMAP_SHPGPERPROC
154 #define PMAP_SHPGPERPROC 200
155 #endif
156
157 #if defined(DIAGNOSTIC)
158 #define PMAP_DIAGNOSTIC
159 #endif
160
161 #define MINPV 2048
162
163 #if !defined(PMAP_DIAGNOSTIC)
164 #define PMAP_INLINE __inline
165 #else
166 #define PMAP_INLINE
167 #endif
168
169 /*
170 * Get PDEs and PTEs for user/kernel address space
171 */
172 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
173 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
174
175 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
176 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
177 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
178 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
179 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
180
181 #define pmap_pte_set_w(pte, v) ((v) ? atomic_set_int((u_int *)(pte), PG_W) : \
182 atomic_clear_int((u_int *)(pte), PG_W))
183 #define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v)))
184
185 struct pmap kernel_pmap_store;
186 LIST_HEAD(pmaplist, pmap);
187 static struct pmaplist allpmaps;
188 static struct mtx allpmaps_lock;
189
190 vm_paddr_t avail_end; /* PA of last available physical page */
191 vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */
192 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
193 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
194 int pgeflag = 0; /* PG_G or-in */
195 int pseflag = 0; /* PG_PS or-in */
196
197 static int nkpt;
198 vm_offset_t kernel_vm_end;
199 extern u_int32_t KERNend;
200
201 #ifdef PAE
202 static uma_zone_t pdptzone;
203 #endif
204
205 /*
206 * Data for the pv entry allocation mechanism
207 */
208 static uma_zone_t pvzone;
209 static struct vm_object pvzone_obj;
210 static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0;
211 int pmap_pagedaemon_waken;
212
213 /*
214 * All those kernel PT submaps that BSD is so fond of
215 */
216 pt_entry_t *CMAP1 = 0;
217 static pt_entry_t *CMAP2, *CMAP3;
218 caddr_t CADDR1 = 0, ptvmmap = 0;
219 static caddr_t CADDR2, CADDR3;
220 static struct mtx CMAPCADDR12_lock;
221 struct msgbuf *msgbufp = 0;
222
223 /*
224 * Crashdump maps.
225 */
226 static caddr_t crashdumpmap;
227
228 #ifdef SMP
229 extern pt_entry_t *SMPpt;
230 #endif
231 static pt_entry_t *PMAP1 = 0, *PMAP2;
232 static pt_entry_t *PADDR1 = 0, *PADDR2;
233 #ifdef SMP
234 static int PMAP1cpu;
235 static int PMAP1changedcpu;
236 SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD,
237 &PMAP1changedcpu, 0,
238 "Number of times pmap_pte_quick changed CPU with same PMAP1");
239 #endif
240 static int PMAP1changed;
241 SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD,
242 &PMAP1changed, 0,
243 "Number of times pmap_pte_quick changed PMAP1");
244 static int PMAP1unchanged;
245 SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD,
246 &PMAP1unchanged, 0,
247 "Number of times pmap_pte_quick didn't change PMAP1");
248 static struct mtx PMAP2mutex;
249
250 static PMAP_INLINE void free_pv_entry(pv_entry_t pv);
251 static pv_entry_t get_pv_entry(void);
252 static void pmap_clear_ptes(vm_page_t m, int bit);
253
254 static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva);
255 static void pmap_remove_page(struct pmap *pmap, vm_offset_t va);
256 static int pmap_remove_entry(struct pmap *pmap, vm_page_t m,
257 vm_offset_t va);
258 static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m);
259
260 static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags);
261
262 static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags);
263 static int _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m);
264 static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va);
265 static void pmap_pte_release(pt_entry_t *pte);
266 static int pmap_unuse_pt(pmap_t, vm_offset_t);
267 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
268 #ifdef PAE
269 static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait);
270 #endif
271
272 CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t));
273 CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t));
274
275 /*
276 * Move the kernel virtual free pointer to the next
277 * 4MB. This is used to help improve performance
278 * by using a large (4MB) page for much of the kernel
279 * (.text, .data, .bss)
280 */
281 static vm_offset_t
282 pmap_kmem_choose(vm_offset_t addr)
283 {
284 vm_offset_t newaddr = addr;
285
286 #ifndef DISABLE_PSE
287 if (cpu_feature & CPUID_PSE)
288 newaddr = (addr + PDRMASK) & ~PDRMASK;
289 #endif
290 return newaddr;
291 }
292
293 /*
294 * Bootstrap the system enough to run with virtual memory.
295 *
296 * On the i386 this is called after mapping has already been enabled
297 * and just syncs the pmap module with what has already been done.
298 * [We can't call it easily with mapping off since the kernel is not
299 * mapped with PA == VA, hence we would have to relocate every address
300 * from the linked base (virtual) address "KERNBASE" to the actual
301 * (physical) address starting relative to 0]
302 */
303 void
304 pmap_bootstrap(firstaddr, loadaddr)
305 vm_paddr_t firstaddr;
306 vm_paddr_t loadaddr;
307 {
308 vm_offset_t va;
309 pt_entry_t *pte, *unused;
310 int i;
311
312 /*
313 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too
314 * large. It should instead be correctly calculated in locore.s and
315 * not based on 'first' (which is a physical address, not a virtual
316 * address, for the start of unused physical memory). The kernel
317 * page tables are NOT double mapped and thus should not be included
318 * in this calculation.
319 */
320 virtual_avail = (vm_offset_t) KERNBASE + firstaddr;
321 virtual_avail = pmap_kmem_choose(virtual_avail);
322
323 virtual_end = VM_MAX_KERNEL_ADDRESS;
324
325 /*
326 * Initialize the kernel pmap (which is statically allocated).
327 */
328 PMAP_LOCK_INIT(kernel_pmap);
329 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD);
330 #ifdef PAE
331 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT);
332 #endif
333 kernel_pmap->pm_active = -1; /* don't allow deactivation */
334 TAILQ_INIT(&kernel_pmap->pm_pvlist);
335 LIST_INIT(&allpmaps);
336 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN);
337 mtx_lock_spin(&allpmaps_lock);
338 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list);
339 mtx_unlock_spin(&allpmaps_lock);
340 nkpt = NKPT;
341
342 /*
343 * Reserve some special page table entries/VA space for temporary
344 * mapping of pages.
345 */
346 #define SYSMAP(c, p, v, n) \
347 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
348
349 va = virtual_avail;
350 pte = vtopte(va);
351
352 /*
353 * CMAP1/CMAP2 are used for zeroing and copying pages.
354 * CMAP3 is used for the idle process page zeroing.
355 */
356 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
357 SYSMAP(caddr_t, CMAP2, CADDR2, 1)
358 SYSMAP(caddr_t, CMAP3, CADDR3, 1)
359 *CMAP3 = 0;
360
361 mtx_init(&CMAPCADDR12_lock, "CMAPCADDR12", NULL, MTX_DEF);
362
363 /*
364 * Crashdump maps.
365 */
366 SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS)
367
368 /*
369 * ptvmmap is used for reading arbitrary physical pages via /dev/mem.
370 */
371 SYSMAP(caddr_t, unused, ptvmmap, 1)
372
373 /*
374 * msgbufp is used to map the system message buffer.
375 */
376 SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(MSGBUF_SIZE)))
377
378 /*
379 * ptemap is used for pmap_pte_quick
380 */
381 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1);
382 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1);
383
384 mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF);
385
386 virtual_avail = va;
387
388 *CMAP1 = *CMAP2 = 0;
389 for (i = 0; i < NKPT; i++)
390 PTD[i] = 0;
391
392 /* Turn on PG_G on kernel page(s) */
393 pmap_set_pg();
394 }
395
396 /*
397 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on.
398 */
399 void
400 pmap_set_pg(void)
401 {
402 pd_entry_t pdir;
403 pt_entry_t *pte;
404 vm_offset_t va, endva;
405 int i;
406
407 if (pgeflag == 0)
408 return;
409
410 i = KERNLOAD/NBPDR;
411 endva = KERNBASE + KERNend;
412
413 if (pseflag) {
414 va = KERNBASE + KERNLOAD;
415 while (va < endva) {
416 pdir = kernel_pmap->pm_pdir[KPTDI+i];
417 pdir |= pgeflag;
418 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir;
419 invltlb(); /* Play it safe, invltlb() every time */
420 i++;
421 va += NBPDR;
422 }
423 } else {
424 va = (vm_offset_t)btext;
425 while (va < endva) {
426 pte = vtopte(va);
427 if (*pte)
428 *pte |= pgeflag;
429 invltlb(); /* Play it safe, invltlb() every time */
430 va += PAGE_SIZE;
431 }
432 }
433 }
434
435 #ifdef PAE
436
437 static MALLOC_DEFINE(M_PMAPPDPT, "pmap", "pmap pdpt");
438
439 static void *
440 pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
441 {
442 *flags = UMA_SLAB_PRIV;
443 return (contigmalloc(PAGE_SIZE, M_PMAPPDPT, 0, 0x0ULL, 0xffffffffULL,
444 1, 0));
445 }
446 #endif
447
448 /*
449 * Initialize the pmap module.
450 * Called by vm_init, to initialize any structures that the pmap
451 * system needs to map virtual memory.
452 * pmap_init has been enhanced to support in a fairly consistant
453 * way, discontiguous physical memory.
454 */
455 void
456 pmap_init(void)
457 {
458 int i;
459
460 /*
461 * Allocate memory for random pmap data structures. Includes the
462 * pv_head_table.
463 */
464
465 for(i = 0; i < vm_page_array_size; i++) {
466 vm_page_t m;
467
468 m = &vm_page_array[i];
469 TAILQ_INIT(&m->md.pv_list);
470 m->md.pv_list_count = 0;
471 }
472
473 /*
474 * init the pv free list
475 */
476 pvzone = uma_zcreate("PV ENTRY", sizeof (struct pv_entry), NULL, NULL,
477 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE);
478 uma_prealloc(pvzone, MINPV);
479
480 #ifdef PAE
481 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL,
482 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1,
483 UMA_ZONE_VM | UMA_ZONE_NOFREE);
484 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf);
485 #endif
486
487 /*
488 * Now it is safe to enable pv_table recording.
489 */
490 pmap_initialized = TRUE;
491 }
492
493 /*
494 * Initialize the address space (zone) for the pv_entries. Set a
495 * high water mark so that the system can recover from excessive
496 * numbers of pv entries.
497 */
498 void
499 pmap_init2()
500 {
501 int shpgperproc = PMAP_SHPGPERPROC;
502
503 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
504 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
505 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
506 pv_entry_high_water = 9 * (pv_entry_max / 10);
507 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max);
508 }
509
510
511 /***************************************************
512 * Low level helper routines.....
513 ***************************************************/
514
515 #if defined(PMAP_DIAGNOSTIC)
516
517 /*
518 * This code checks for non-writeable/modified pages.
519 * This should be an invalid condition.
520 */
521 static int
522 pmap_nw_modified(pt_entry_t ptea)
523 {
524 int pte;
525
526 pte = (int) ptea;
527
528 if ((pte & (PG_M|PG_RW)) == PG_M)
529 return 1;
530 else
531 return 0;
532 }
533 #endif
534
535
536 /*
537 * this routine defines the region(s) of memory that should
538 * not be tested for the modified bit.
539 */
540 static PMAP_INLINE int
541 pmap_track_modified(vm_offset_t va)
542 {
543 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva))
544 return 1;
545 else
546 return 0;
547 }
548
549 #ifdef I386_CPU
550 /*
551 * i386 only has "invalidate everything" and no SMP to worry about.
552 */
553 PMAP_INLINE void
554 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
555 {
556
557 if (pmap == kernel_pmap || pmap->pm_active)
558 invltlb();
559 }
560
561 PMAP_INLINE void
562 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
563 {
564
565 if (pmap == kernel_pmap || pmap->pm_active)
566 invltlb();
567 }
568
569 PMAP_INLINE void
570 pmap_invalidate_all(pmap_t pmap)
571 {
572
573 if (pmap == kernel_pmap || pmap->pm_active)
574 invltlb();
575 }
576 #else /* !I386_CPU */
577 #ifdef SMP
578 /*
579 * For SMP, these functions have to use the IPI mechanism for coherence.
580 */
581 void
582 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
583 {
584 u_int cpumask;
585 u_int other_cpus;
586
587 if (smp_started) {
588 if (!(read_eflags() & PSL_I))
589 panic("%s: interrupts disabled", __func__);
590 mtx_lock_spin(&smp_rv_mtx);
591 } else
592 critical_enter();
593 /*
594 * We need to disable interrupt preemption but MUST NOT have
595 * interrupts disabled here.
596 * XXX we may need to hold schedlock to get a coherent pm_active
597 * XXX critical sections disable interrupts again
598 */
599 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) {
600 invlpg(va);
601 smp_invlpg(va);
602 } else {
603 cpumask = PCPU_GET(cpumask);
604 other_cpus = PCPU_GET(other_cpus);
605 if (pmap->pm_active & cpumask)
606 invlpg(va);
607 if (pmap->pm_active & other_cpus)
608 smp_masked_invlpg(pmap->pm_active & other_cpus, va);
609 }
610 if (smp_started)
611 mtx_unlock_spin(&smp_rv_mtx);
612 else
613 critical_exit();
614 }
615
616 void
617 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
618 {
619 u_int cpumask;
620 u_int other_cpus;
621 vm_offset_t addr;
622
623 if (smp_started) {
624 if (!(read_eflags() & PSL_I))
625 panic("%s: interrupts disabled", __func__);
626 mtx_lock_spin(&smp_rv_mtx);
627 } else
628 critical_enter();
629 /*
630 * We need to disable interrupt preemption but MUST NOT have
631 * interrupts disabled here.
632 * XXX we may need to hold schedlock to get a coherent pm_active
633 * XXX critical sections disable interrupts again
634 */
635 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) {
636 for (addr = sva; addr < eva; addr += PAGE_SIZE)
637 invlpg(addr);
638 smp_invlpg_range(sva, eva);
639 } else {
640 cpumask = PCPU_GET(cpumask);
641 other_cpus = PCPU_GET(other_cpus);
642 if (pmap->pm_active & cpumask)
643 for (addr = sva; addr < eva; addr += PAGE_SIZE)
644 invlpg(addr);
645 if (pmap->pm_active & other_cpus)
646 smp_masked_invlpg_range(pmap->pm_active & other_cpus,
647 sva, eva);
648 }
649 if (smp_started)
650 mtx_unlock_spin(&smp_rv_mtx);
651 else
652 critical_exit();
653 }
654
655 void
656 pmap_invalidate_all(pmap_t pmap)
657 {
658 u_int cpumask;
659 u_int other_cpus;
660
661 if (smp_started) {
662 if (!(read_eflags() & PSL_I))
663 panic("%s: interrupts disabled", __func__);
664 mtx_lock_spin(&smp_rv_mtx);
665 } else
666 critical_enter();
667 /*
668 * We need to disable interrupt preemption but MUST NOT have
669 * interrupts disabled here.
670 * XXX we may need to hold schedlock to get a coherent pm_active
671 * XXX critical sections disable interrupts again
672 */
673 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) {
674 invltlb();
675 smp_invltlb();
676 } else {
677 cpumask = PCPU_GET(cpumask);
678 other_cpus = PCPU_GET(other_cpus);
679 if (pmap->pm_active & cpumask)
680 invltlb();
681 if (pmap->pm_active & other_cpus)
682 smp_masked_invltlb(pmap->pm_active & other_cpus);
683 }
684 if (smp_started)
685 mtx_unlock_spin(&smp_rv_mtx);
686 else
687 critical_exit();
688 }
689 #else /* !SMP */
690 /*
691 * Normal, non-SMP, 486+ invalidation functions.
692 * We inline these within pmap.c for speed.
693 */
694 PMAP_INLINE void
695 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
696 {
697
698 if (pmap == kernel_pmap || pmap->pm_active)
699 invlpg(va);
700 }
701
702 PMAP_INLINE void
703 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
704 {
705 vm_offset_t addr;
706
707 if (pmap == kernel_pmap || pmap->pm_active)
708 for (addr = sva; addr < eva; addr += PAGE_SIZE)
709 invlpg(addr);
710 }
711
712 PMAP_INLINE void
713 pmap_invalidate_all(pmap_t pmap)
714 {
715
716 if (pmap == kernel_pmap || pmap->pm_active)
717 invltlb();
718 }
719 #endif /* !SMP */
720 #endif /* !I386_CPU */
721
722 /*
723 * Are we current address space or kernel? N.B. We return FALSE when
724 * a pmap's page table is in use because a kernel thread is borrowing
725 * it. The borrowed page table can change spontaneously, making any
726 * dependence on its continued use subject to a race condition.
727 */
728 static __inline int
729 pmap_is_current(pmap_t pmap)
730 {
731
732 return (pmap == kernel_pmap ||
733 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) &&
734 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)));
735 }
736
737 /*
738 * If the given pmap is not the current or kernel pmap, the returned pte must
739 * be released by passing it to pmap_pte_release().
740 */
741 pt_entry_t *
742 pmap_pte(pmap_t pmap, vm_offset_t va)
743 {
744 pd_entry_t newpf;
745 pd_entry_t *pde;
746
747 pde = pmap_pde(pmap, va);
748 if (*pde & PG_PS)
749 return (pde);
750 if (*pde != 0) {
751 /* are we current address space or kernel? */
752 if (pmap_is_current(pmap))
753 return (vtopte(va));
754 mtx_lock(&PMAP2mutex);
755 newpf = *pde & PG_FRAME;
756 if ((*PMAP2 & PG_FRAME) != newpf) {
757 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M;
758 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2);
759 }
760 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1)));
761 }
762 return (0);
763 }
764
765 /*
766 * Releases a pte that was obtained from pmap_pte(). Be prepared for the pte
767 * being NULL.
768 */
769 static __inline void
770 pmap_pte_release(pt_entry_t *pte)
771 {
772
773 if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2)
774 mtx_unlock(&PMAP2mutex);
775 }
776
777 static __inline void
778 invlcaddr(void *caddr)
779 {
780 #ifdef I386_CPU
781 invltlb();
782 #else
783 invlpg((u_int)caddr);
784 #endif
785 }
786
787 /*
788 * Super fast pmap_pte routine best used when scanning
789 * the pv lists. This eliminates many coarse-grained
790 * invltlb calls. Note that many of the pv list
791 * scans are across different pmaps. It is very wasteful
792 * to do an entire invltlb for checking a single mapping.
793 *
794 * If the given pmap is not the current pmap, vm_page_queue_mtx
795 * must be held and curthread pinned to a CPU.
796 */
797 static pt_entry_t *
798 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
799 {
800 pd_entry_t newpf;
801 pd_entry_t *pde;
802
803 pde = pmap_pde(pmap, va);
804 if (*pde & PG_PS)
805 return (pde);
806 if (*pde != 0) {
807 /* are we current address space or kernel? */
808 if (pmap_is_current(pmap))
809 return (vtopte(va));
810 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
811 KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
812 newpf = *pde & PG_FRAME;
813 if ((*PMAP1 & PG_FRAME) != newpf) {
814 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M;
815 #ifdef SMP
816 PMAP1cpu = PCPU_GET(cpuid);
817 #endif
818 invlcaddr(PADDR1);
819 PMAP1changed++;
820 } else
821 #ifdef SMP
822 if (PMAP1cpu != PCPU_GET(cpuid)) {
823 PMAP1cpu = PCPU_GET(cpuid);
824 invlcaddr(PADDR1);
825 PMAP1changedcpu++;
826 } else
827 #endif
828 PMAP1unchanged++;
829 return (PADDR1 + (i386_btop(va) & (NPTEPG - 1)));
830 }
831 return (0);
832 }
833
834 /*
835 * Routine: pmap_extract
836 * Function:
837 * Extract the physical page address associated
838 * with the given map/virtual_address pair.
839 */
840 vm_paddr_t
841 pmap_extract(pmap_t pmap, vm_offset_t va)
842 {
843 vm_paddr_t rtval;
844 pt_entry_t *pte;
845 pd_entry_t pde;
846
847 rtval = 0;
848 PMAP_LOCK(pmap);
849 pde = pmap->pm_pdir[va >> PDRSHIFT];
850 if (pde != 0) {
851 if ((pde & PG_PS) != 0) {
852 rtval = (pde & ~PDRMASK) | (va & PDRMASK);
853 PMAP_UNLOCK(pmap);
854 return rtval;
855 }
856 pte = pmap_pte(pmap, va);
857 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
858 pmap_pte_release(pte);
859 }
860 PMAP_UNLOCK(pmap);
861 return (rtval);
862 }
863
864 /*
865 * Routine: pmap_extract_and_hold
866 * Function:
867 * Atomically extract and hold the physical page
868 * with the given pmap and virtual address pair
869 * if that mapping permits the given protection.
870 */
871 vm_page_t
872 pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
873 {
874 pd_entry_t pde;
875 pt_entry_t pte;
876 vm_page_t m;
877
878 m = NULL;
879 vm_page_lock_queues();
880 PMAP_LOCK(pmap);
881 pde = *pmap_pde(pmap, va);
882 if (pde != 0) {
883 if (pde & PG_PS) {
884 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) {
885 m = PHYS_TO_VM_PAGE((pde & ~PDRMASK) |
886 (va & PDRMASK));
887 vm_page_hold(m);
888 }
889 } else {
890 sched_pin();
891 pte = *pmap_pte_quick(pmap, va);
892 if (pte != 0 &&
893 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) {
894 m = PHYS_TO_VM_PAGE(pte & PG_FRAME);
895 vm_page_hold(m);
896 }
897 sched_unpin();
898 }
899 }
900 vm_page_unlock_queues();
901 PMAP_UNLOCK(pmap);
902 return (m);
903 }
904
905 /***************************************************
906 * Low level mapping routines.....
907 ***************************************************/
908
909 /*
910 * Add a wired page to the kva.
911 * Note: not SMP coherent.
912 */
913 PMAP_INLINE void
914 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
915 {
916 pt_entry_t *pte;
917
918 pte = vtopte(va);
919 pte_store(pte, pa | PG_RW | PG_V | pgeflag);
920 }
921
922 /*
923 * Remove a page from the kernel pagetables.
924 * Note: not SMP coherent.
925 */
926 PMAP_INLINE void
927 pmap_kremove(vm_offset_t va)
928 {
929 pt_entry_t *pte;
930
931 pte = vtopte(va);
932 pte_clear(pte);
933 }
934
935 /*
936 * Used to map a range of physical addresses into kernel
937 * virtual address space.
938 *
939 * The value passed in '*virt' is a suggested virtual address for
940 * the mapping. Architectures which can support a direct-mapped
941 * physical to virtual region can return the appropriate address
942 * within that region, leaving '*virt' unchanged. Other
943 * architectures should map the pages starting at '*virt' and
944 * update '*virt' with the first usable address after the mapped
945 * region.
946 */
947 vm_offset_t
948 pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot)
949 {
950 vm_offset_t va, sva;
951
952 va = sva = *virt;
953 while (start < end) {
954 pmap_kenter(va, start);
955 va += PAGE_SIZE;
956 start += PAGE_SIZE;
957 }
958 pmap_invalidate_range(kernel_pmap, sva, va);
959 *virt = va;
960 return (sva);
961 }
962
963
964 /*
965 * Add a list of wired pages to the kva
966 * this routine is only used for temporary
967 * kernel mappings that do not need to have
968 * page modification or references recorded.
969 * Note that old mappings are simply written
970 * over. The page *must* be wired.
971 * Note: SMP coherent. Uses a ranged shootdown IPI.
972 */
973 void
974 pmap_qenter(vm_offset_t sva, vm_page_t *m, int count)
975 {
976 vm_offset_t va;
977
978 va = sva;
979 while (count-- > 0) {
980 pmap_kenter(va, VM_PAGE_TO_PHYS(*m));
981 va += PAGE_SIZE;
982 m++;
983 }
984 pmap_invalidate_range(kernel_pmap, sva, va);
985 }
986
987 /*
988 * This routine tears out page mappings from the
989 * kernel -- it is meant only for temporary mappings.
990 * Note: SMP coherent. Uses a ranged shootdown IPI.
991 */
992 void
993 pmap_qremove(vm_offset_t sva, int count)
994 {
995 vm_offset_t va;
996
997 va = sva;
998 while (count-- > 0) {
999 pmap_kremove(va);
1000 va += PAGE_SIZE;
1001 }
1002 pmap_invalidate_range(kernel_pmap, sva, va);
1003 }
1004
1005 /***************************************************
1006 * Page table page management routines.....
1007 ***************************************************/
1008
1009 /*
1010 * This routine unholds page table pages, and if the hold count
1011 * drops to zero, then it decrements the wire count.
1012 */
1013 static PMAP_INLINE int
1014 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
1015 {
1016
1017 --m->wire_count;
1018 if (m->wire_count == 0)
1019 return _pmap_unwire_pte_hold(pmap, m);
1020 else
1021 return 0;
1022 }
1023
1024 static int
1025 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
1026 {
1027 vm_offset_t pteva;
1028
1029 /*
1030 * unmap the page table page
1031 */
1032 pmap->pm_pdir[m->pindex] = 0;
1033 --pmap->pm_stats.resident_count;
1034
1035 /*
1036 * Do an invltlb to make the invalidated mapping
1037 * take effect immediately.
1038 */
1039 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex);
1040 pmap_invalidate_page(pmap, pteva);
1041
1042 vm_page_free_zero(m);
1043 atomic_subtract_int(&cnt.v_wire_count, 1);
1044 return 1;
1045 }
1046
1047 /*
1048 * After removing a page table entry, this routine is used to
1049 * conditionally free the page, and manage the hold/wire counts.
1050 */
1051 static int
1052 pmap_unuse_pt(pmap_t pmap, vm_offset_t va)
1053 {
1054 pd_entry_t ptepde;
1055 vm_page_t mpte;
1056
1057 if (va >= VM_MAXUSER_ADDRESS)
1058 return 0;
1059 ptepde = *pmap_pde(pmap, va);
1060 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME);
1061 return pmap_unwire_pte_hold(pmap, mpte);
1062 }
1063
1064 void
1065 pmap_pinit0(pmap)
1066 struct pmap *pmap;
1067 {
1068
1069 PMAP_LOCK_INIT(pmap);
1070 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD);
1071 #ifdef PAE
1072 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT);
1073 #endif
1074 pmap->pm_active = 0;
1075 PCPU_SET(curpmap, pmap);
1076 TAILQ_INIT(&pmap->pm_pvlist);
1077 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1078 mtx_lock_spin(&allpmaps_lock);
1079 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list);
1080 mtx_unlock_spin(&allpmaps_lock);
1081 }
1082
1083 /*
1084 * Initialize a preallocated and zeroed pmap structure,
1085 * such as one in a vmspace structure.
1086 */
1087 void
1088 pmap_pinit(pmap)
1089 register struct pmap *pmap;
1090 {
1091 vm_page_t m, ptdpg[NPGPTD];
1092 vm_paddr_t pa;
1093 static int color;
1094 int i;
1095
1096 PMAP_LOCK_INIT(pmap);
1097
1098 /*
1099 * No need to allocate page table space yet but we do need a valid
1100 * page directory table.
1101 */
1102 if (pmap->pm_pdir == NULL) {
1103 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map,
1104 NBPTD);
1105 #ifdef PAE
1106 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO);
1107 KASSERT(((vm_offset_t)pmap->pm_pdpt &
1108 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0,
1109 ("pmap_pinit: pdpt misaligned"));
1110 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30),
1111 ("pmap_pinit: pdpt above 4g"));
1112 #endif
1113 }
1114
1115 /*
1116 * allocate the page directory page(s)
1117 */
1118 for (i = 0; i < NPGPTD;) {
1119 m = vm_page_alloc(NULL, color++,
1120 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
1121 VM_ALLOC_ZERO);
1122 if (m == NULL)
1123 VM_WAIT;
1124 else {
1125 ptdpg[i++] = m;
1126 }
1127 }
1128
1129 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD);
1130
1131 for (i = 0; i < NPGPTD; i++) {
1132 if ((ptdpg[i]->flags & PG_ZERO) == 0)
1133 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE);
1134 }
1135
1136 mtx_lock_spin(&allpmaps_lock);
1137 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list);
1138 mtx_unlock_spin(&allpmaps_lock);
1139 /* Wire in kernel global address entries. */
1140 /* XXX copies current process, does not fill in MPPTDI */
1141 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t));
1142 #ifdef SMP
1143 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1144 #endif
1145
1146 /* install self-referential address mapping entry(s) */
1147 for (i = 0; i < NPGPTD; i++) {
1148 pa = VM_PAGE_TO_PHYS(ptdpg[i]);
1149 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M;
1150 #ifdef PAE
1151 pmap->pm_pdpt[i] = pa | PG_V;
1152 #endif
1153 }
1154
1155 pmap->pm_active = 0;
1156 TAILQ_INIT(&pmap->pm_pvlist);
1157 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1158 }
1159
1160 /*
1161 * this routine is called if the page table page is not
1162 * mapped correctly.
1163 */
1164 static vm_page_t
1165 _pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags)
1166 {
1167 vm_paddr_t ptepa;
1168 vm_page_t m;
1169
1170 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT ||
1171 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK,
1172 ("_pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK"));
1173
1174 /*
1175 * Allocate a page table page.
1176 */
1177 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ |
1178 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) {
1179 if (flags & M_WAITOK) {
1180 PMAP_UNLOCK(pmap);
1181 vm_page_unlock_queues();
1182 VM_WAIT;
1183 vm_page_lock_queues();
1184 PMAP_LOCK(pmap);
1185 }
1186
1187 /*
1188 * Indicate the need to retry. While waiting, the page table
1189 * page may have been allocated.
1190 */
1191 return (NULL);
1192 }
1193 if ((m->flags & PG_ZERO) == 0)
1194 pmap_zero_page(m);
1195
1196 /*
1197 * Map the pagetable page into the process address space, if
1198 * it isn't already there.
1199 */
1200
1201 pmap->pm_stats.resident_count++;
1202
1203 ptepa = VM_PAGE_TO_PHYS(m);
1204 pmap->pm_pdir[ptepindex] =
1205 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1206
1207 return m;
1208 }
1209
1210 static vm_page_t
1211 pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags)
1212 {
1213 unsigned ptepindex;
1214 pd_entry_t ptepa;
1215 vm_page_t m;
1216
1217 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT ||
1218 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK,
1219 ("pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK"));
1220
1221 /*
1222 * Calculate pagetable page index
1223 */
1224 ptepindex = va >> PDRSHIFT;
1225 retry:
1226 /*
1227 * Get the page directory entry
1228 */
1229 ptepa = pmap->pm_pdir[ptepindex];
1230
1231 /*
1232 * This supports switching from a 4MB page to a
1233 * normal 4K page.
1234 */
1235 if (ptepa & PG_PS) {
1236 pmap->pm_pdir[ptepindex] = 0;
1237 ptepa = 0;
1238 pmap_invalidate_all(kernel_pmap);
1239 }
1240
1241 /*
1242 * If the page table page is mapped, we just increment the
1243 * hold count, and activate it.
1244 */
1245 if (ptepa) {
1246 m = PHYS_TO_VM_PAGE(ptepa);
1247 m->wire_count++;
1248 } else {
1249 /*
1250 * Here if the pte page isn't mapped, or if it has
1251 * been deallocated.
1252 */
1253 m = _pmap_allocpte(pmap, ptepindex, flags);
1254 if (m == NULL && (flags & M_WAITOK))
1255 goto retry;
1256 }
1257 return (m);
1258 }
1259
1260
1261 /***************************************************
1262 * Pmap allocation/deallocation routines.
1263 ***************************************************/
1264
1265 #ifdef SMP
1266 /*
1267 * Deal with a SMP shootdown of other users of the pmap that we are
1268 * trying to dispose of. This can be a bit hairy.
1269 */
1270 static u_int *lazymask;
1271 static u_int lazyptd;
1272 static volatile u_int lazywait;
1273
1274 void pmap_lazyfix_action(void);
1275
1276 void
1277 pmap_lazyfix_action(void)
1278 {
1279 u_int mymask = PCPU_GET(cpumask);
1280
1281 if (rcr3() == lazyptd)
1282 load_cr3(PCPU_GET(curpcb)->pcb_cr3);
1283 atomic_clear_int(lazymask, mymask);
1284 atomic_store_rel_int(&lazywait, 1);
1285 }
1286
1287 static void
1288 pmap_lazyfix_self(u_int mymask)
1289 {
1290
1291 if (rcr3() == lazyptd)
1292 load_cr3(PCPU_GET(curpcb)->pcb_cr3);
1293 atomic_clear_int(lazymask, mymask);
1294 }
1295
1296
1297 static void
1298 pmap_lazyfix(pmap_t pmap)
1299 {
1300 u_int mymask = PCPU_GET(cpumask);
1301 u_int mask;
1302 register u_int spins;
1303
1304 while ((mask = pmap->pm_active) != 0) {
1305 spins = 50000000;
1306 mask = mask & -mask; /* Find least significant set bit */
1307 mtx_lock_spin(&smp_rv_mtx);
1308 #ifdef PAE
1309 lazyptd = vtophys(pmap->pm_pdpt);
1310 #else
1311 lazyptd = vtophys(pmap->pm_pdir);
1312 #endif
1313 if (mask == mymask) {
1314 lazymask = &pmap->pm_active;
1315 pmap_lazyfix_self(mymask);
1316 } else {
1317 atomic_store_rel_int((u_int *)&lazymask,
1318 (u_int)&pmap->pm_active);
1319 atomic_store_rel_int(&lazywait, 0);
1320 ipi_selected(mask, IPI_LAZYPMAP);
1321 while (lazywait == 0) {
1322 ia32_pause();
1323 if (--spins == 0)
1324 break;
1325 }
1326 }
1327 mtx_unlock_spin(&smp_rv_mtx);
1328 if (spins == 0)
1329 printf("pmap_lazyfix: spun for 50000000\n");
1330 }
1331 }
1332
1333 #else /* SMP */
1334
1335 /*
1336 * Cleaning up on uniprocessor is easy. For various reasons, we're
1337 * unlikely to have to even execute this code, including the fact
1338 * that the cleanup is deferred until the parent does a wait(2), which
1339 * means that another userland process has run.
1340 */
1341 static void
1342 pmap_lazyfix(pmap_t pmap)
1343 {
1344 u_int cr3;
1345
1346 cr3 = vtophys(pmap->pm_pdir);
1347 if (cr3 == rcr3()) {
1348 load_cr3(PCPU_GET(curpcb)->pcb_cr3);
1349 pmap->pm_active &= ~(PCPU_GET(cpumask));
1350 }
1351 }
1352 #endif /* SMP */
1353
1354 /*
1355 * Release any resources held by the given physical map.
1356 * Called when a pmap initialized by pmap_pinit is being released.
1357 * Should only be called if the map contains no valid mappings.
1358 */
1359 void
1360 pmap_release(pmap_t pmap)
1361 {
1362 vm_page_t m, ptdpg[NPGPTD];
1363 int i;
1364
1365 KASSERT(pmap->pm_stats.resident_count == 0,
1366 ("pmap_release: pmap resident count %ld != 0",
1367 pmap->pm_stats.resident_count));
1368
1369 pmap_lazyfix(pmap);
1370 mtx_lock_spin(&allpmaps_lock);
1371 LIST_REMOVE(pmap, pm_list);
1372 mtx_unlock_spin(&allpmaps_lock);
1373
1374 for (i = 0; i < NPGPTD; i++)
1375 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]);
1376
1377 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) *
1378 sizeof(*pmap->pm_pdir));
1379 #ifdef SMP
1380 pmap->pm_pdir[MPPTDI] = 0;
1381 #endif
1382
1383 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD);
1384
1385 vm_page_lock_queues();
1386 for (i = 0; i < NPGPTD; i++) {
1387 m = ptdpg[i];
1388 #ifdef PAE
1389 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME),
1390 ("pmap_release: got wrong ptd page"));
1391 #endif
1392 m->wire_count--;
1393 atomic_subtract_int(&cnt.v_wire_count, 1);
1394 vm_page_free_zero(m);
1395 }
1396 vm_page_unlock_queues();
1397 PMAP_LOCK_DESTROY(pmap);
1398 }
1399
1400 static int
1401 kvm_size(SYSCTL_HANDLER_ARGS)
1402 {
1403 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE;
1404
1405 return sysctl_handle_long(oidp, &ksize, 0, req);
1406 }
1407 SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD,
1408 0, 0, kvm_size, "IU", "Size of KVM");
1409
1410 static int
1411 kvm_free(SYSCTL_HANDLER_ARGS)
1412 {
1413 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end;
1414
1415 return sysctl_handle_long(oidp, &kfree, 0, req);
1416 }
1417 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD,
1418 0, 0, kvm_free, "IU", "Amount of KVM free");
1419
1420 /*
1421 * grow the number of kernel page table entries, if needed
1422 */
1423 void
1424 pmap_growkernel(vm_offset_t addr)
1425 {
1426 struct pmap *pmap;
1427 vm_paddr_t ptppaddr;
1428 vm_page_t nkpg;
1429 pd_entry_t newpdir;
1430 pt_entry_t *pde;
1431
1432 mtx_assert(&kernel_map->system_mtx, MA_OWNED);
1433 if (kernel_vm_end == 0) {
1434 kernel_vm_end = KERNBASE;
1435 nkpt = 0;
1436 while (pdir_pde(PTD, kernel_vm_end)) {
1437 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1438 nkpt++;
1439 }
1440 }
1441 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1442 while (kernel_vm_end < addr) {
1443 if (pdir_pde(PTD, kernel_vm_end)) {
1444 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1445 continue;
1446 }
1447
1448 /*
1449 * This index is bogus, but out of the way
1450 */
1451 nkpg = vm_page_alloc(NULL, nkpt,
1452 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
1453 if (!nkpg)
1454 panic("pmap_growkernel: no memory to grow kernel");
1455
1456 nkpt++;
1457
1458 pmap_zero_page(nkpg);
1459 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1460 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1461 pdir_pde(PTD, kernel_vm_end) = newpdir;
1462
1463 mtx_lock_spin(&allpmaps_lock);
1464 LIST_FOREACH(pmap, &allpmaps, pm_list) {
1465 pde = pmap_pde(pmap, kernel_vm_end);
1466 pde_store(pde, newpdir);
1467 }
1468 mtx_unlock_spin(&allpmaps_lock);
1469 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1470 }
1471 }
1472
1473
1474 /***************************************************
1475 * page management routines.
1476 ***************************************************/
1477
1478 /*
1479 * free the pv_entry back to the free list
1480 */
1481 static PMAP_INLINE void
1482 free_pv_entry(pv_entry_t pv)
1483 {
1484 pv_entry_count--;
1485 uma_zfree(pvzone, pv);
1486 }
1487
1488 /*
1489 * get a new pv_entry, allocating a block from the system
1490 * when needed.
1491 * the memory allocation is performed bypassing the malloc code
1492 * because of the possibility of allocations at interrupt time.
1493 */
1494 static pv_entry_t
1495 get_pv_entry(void)
1496 {
1497 pv_entry_count++;
1498 if (pv_entry_high_water &&
1499 (pv_entry_count > pv_entry_high_water) &&
1500 (pmap_pagedaemon_waken == 0)) {
1501 pmap_pagedaemon_waken = 1;
1502 wakeup (&vm_pages_needed);
1503 }
1504 return uma_zalloc(pvzone, M_NOWAIT);
1505 }
1506
1507
1508 static int
1509 pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va)
1510 {
1511 pv_entry_t pv;
1512 int rtval;
1513
1514 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1515 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1516 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1517 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1518 if (pmap == pv->pv_pmap && va == pv->pv_va)
1519 break;
1520 }
1521 } else {
1522 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1523 if (va == pv->pv_va)
1524 break;
1525 }
1526 }
1527
1528 rtval = 0;
1529 if (pv) {
1530 rtval = pmap_unuse_pt(pmap, va);
1531 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1532 m->md.pv_list_count--;
1533 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1534 vm_page_flag_clear(m, PG_WRITEABLE);
1535
1536 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1537 free_pv_entry(pv);
1538 }
1539
1540 return rtval;
1541 }
1542
1543 /*
1544 * Create a pv entry for page at pa for
1545 * (pmap, va).
1546 */
1547 static void
1548 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m)
1549 {
1550 pv_entry_t pv;
1551
1552 pv = get_pv_entry();
1553 pv->pv_va = va;
1554 pv->pv_pmap = pmap;
1555
1556 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1557 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1558 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1559 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1560 m->md.pv_list_count++;
1561 }
1562
1563 /*
1564 * pmap_remove_pte: do the things to unmap a page in a process
1565 */
1566 static int
1567 pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va)
1568 {
1569 pt_entry_t oldpte;
1570 vm_page_t m;
1571
1572 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1573 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
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 return pmap_unuse_pt(pmap, va);
1602 }
1603 }
1604
1605 /*
1606 * Remove a single page from a process address space
1607 */
1608 static void
1609 pmap_remove_page(pmap_t pmap, vm_offset_t va)
1610 {
1611 pt_entry_t *pte;
1612
1613 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1614 KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
1615 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1616 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0)
1617 return;
1618 pmap_remove_pte(pmap, pte, va);
1619 pmap_invalidate_page(pmap, va);
1620 }
1621
1622 /*
1623 * Remove the given range of addresses from the specified map.
1624 *
1625 * It is assumed that the start and end are properly
1626 * rounded to the page size.
1627 */
1628 void
1629 pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1630 {
1631 vm_offset_t pdnxt;
1632 pd_entry_t ptpaddr;
1633 pt_entry_t *pte;
1634 int anyvalid;
1635
1636 /*
1637 * Perform an unsynchronized read. This is, however, safe.
1638 */
1639 if (pmap->pm_stats.resident_count == 0)
1640 return;
1641
1642 anyvalid = 0;
1643
1644 vm_page_lock_queues();
1645 sched_pin();
1646 PMAP_LOCK(pmap);
1647
1648 /*
1649 * special handling of removing one page. a very
1650 * common operation and easy to short circuit some
1651 * code.
1652 */
1653 if ((sva + PAGE_SIZE == eva) &&
1654 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1655 pmap_remove_page(pmap, sva);
1656 goto out;
1657 }
1658
1659 for (; sva < eva; sva = pdnxt) {
1660 unsigned pdirindex;
1661
1662 /*
1663 * Calculate index for next page table.
1664 */
1665 pdnxt = (sva + NBPDR) & ~PDRMASK;
1666 if (pmap->pm_stats.resident_count == 0)
1667 break;
1668
1669 pdirindex = sva >> PDRSHIFT;
1670 ptpaddr = pmap->pm_pdir[pdirindex];
1671
1672 /*
1673 * Weed out invalid mappings. Note: we assume that the page
1674 * directory table is always allocated, and in kernel virtual.
1675 */
1676 if (ptpaddr == 0)
1677 continue;
1678
1679 /*
1680 * Check for large page.
1681 */
1682 if ((ptpaddr & PG_PS) != 0) {
1683 pmap->pm_pdir[pdirindex] = 0;
1684 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1685 anyvalid = 1;
1686 continue;
1687 }
1688
1689 /*
1690 * Limit our scan to either the end of the va represented
1691 * by the current page table page, or to the end of the
1692 * range being removed.
1693 */
1694 if (pdnxt > eva)
1695 pdnxt = eva;
1696
1697 for (; sva != pdnxt; sva += PAGE_SIZE) {
1698 if ((pte = pmap_pte_quick(pmap, sva)) == NULL ||
1699 *pte == 0)
1700 continue;
1701 anyvalid = 1;
1702 if (pmap_remove_pte(pmap, pte, sva))
1703 break;
1704 }
1705 }
1706 out:
1707 sched_unpin();
1708 vm_page_unlock_queues();
1709 if (anyvalid)
1710 pmap_invalidate_all(pmap);
1711 PMAP_UNLOCK(pmap);
1712 }
1713
1714 /*
1715 * Routine: pmap_remove_all
1716 * Function:
1717 * Removes this physical page from
1718 * all physical maps in which it resides.
1719 * Reflects back modify bits to the pager.
1720 *
1721 * Notes:
1722 * Original versions of this routine were very
1723 * inefficient because they iteratively called
1724 * pmap_remove (slow...)
1725 */
1726
1727 void
1728 pmap_remove_all(vm_page_t m)
1729 {
1730 register pv_entry_t pv;
1731 pt_entry_t *pte, tpte;
1732
1733 #if defined(PMAP_DIAGNOSTIC)
1734 /*
1735 * XXX This makes pmap_remove_all() illegal for non-managed pages!
1736 */
1737 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1738 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x",
1739 VM_PAGE_TO_PHYS(m));
1740 }
1741 #endif
1742 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1743 sched_pin();
1744 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1745 PMAP_LOCK(pv->pv_pmap);
1746 pv->pv_pmap->pm_stats.resident_count--;
1747 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1748 tpte = pte_load_clear(pte);
1749 if (tpte & PG_W)
1750 pv->pv_pmap->pm_stats.wired_count--;
1751 if (tpte & PG_A)
1752 vm_page_flag_set(m, PG_REFERENCED);
1753
1754 /*
1755 * Update the vm_page_t clean and reference bits.
1756 */
1757 if (tpte & PG_M) {
1758 #if defined(PMAP_DIAGNOSTIC)
1759 if (pmap_nw_modified((pt_entry_t) tpte)) {
1760 printf(
1761 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1762 pv->pv_va, tpte);
1763 }
1764 #endif
1765 if (pmap_track_modified(pv->pv_va))
1766 vm_page_dirty(m);
1767 }
1768 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
1769 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1770 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1771 m->md.pv_list_count--;
1772 pmap_unuse_pt(pv->pv_pmap, pv->pv_va);
1773 PMAP_UNLOCK(pv->pv_pmap);
1774 free_pv_entry(pv);
1775 }
1776 vm_page_flag_clear(m, PG_WRITEABLE);
1777 sched_unpin();
1778 }
1779
1780 /*
1781 * Set the physical protection on the
1782 * specified range of this map as requested.
1783 */
1784 void
1785 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1786 {
1787 vm_offset_t pdnxt;
1788 pd_entry_t ptpaddr;
1789 int anychanged;
1790
1791 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1792 pmap_remove(pmap, sva, eva);
1793 return;
1794 }
1795
1796 if (prot & VM_PROT_WRITE)
1797 return;
1798
1799 anychanged = 0;
1800
1801 vm_page_lock_queues();
1802 sched_pin();
1803 PMAP_LOCK(pmap);
1804 for (; sva < eva; sva = pdnxt) {
1805 unsigned obits, pbits, pdirindex;
1806
1807 pdnxt = (sva + NBPDR) & ~PDRMASK;
1808
1809 pdirindex = sva >> PDRSHIFT;
1810 ptpaddr = pmap->pm_pdir[pdirindex];
1811
1812 /*
1813 * Weed out invalid mappings. Note: we assume that the page
1814 * directory table is always allocated, and in kernel virtual.
1815 */
1816 if (ptpaddr == 0)
1817 continue;
1818
1819 /*
1820 * Check for large page.
1821 */
1822 if ((ptpaddr & PG_PS) != 0) {
1823 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1824 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1825 anychanged = 1;
1826 continue;
1827 }
1828
1829 if (pdnxt > eva)
1830 pdnxt = eva;
1831
1832 for (; sva != pdnxt; sva += PAGE_SIZE) {
1833 pt_entry_t *pte;
1834 vm_page_t m;
1835
1836 if ((pte = pmap_pte_quick(pmap, sva)) == NULL)
1837 continue;
1838 retry:
1839 /*
1840 * Regardless of whether a pte is 32 or 64 bits in
1841 * size, PG_RW, PG_A, and PG_M are among the least
1842 * significant 32 bits.
1843 */
1844 obits = pbits = *(u_int *)pte;
1845 if (pbits & PG_MANAGED) {
1846 m = NULL;
1847 if (pbits & PG_A) {
1848 m = PHYS_TO_VM_PAGE(pbits);
1849 vm_page_flag_set(m, PG_REFERENCED);
1850 pbits &= ~PG_A;
1851 }
1852 if ((pbits & PG_M) != 0 &&
1853 pmap_track_modified(sva)) {
1854 if (m == NULL)
1855 m = PHYS_TO_VM_PAGE(pbits);
1856 vm_page_dirty(m);
1857 }
1858 }
1859
1860 pbits &= ~(PG_RW | PG_M);
1861
1862 if (pbits != obits) {
1863 if (!atomic_cmpset_int((u_int *)pte, obits,
1864 pbits))
1865 goto retry;
1866 anychanged = 1;
1867 }
1868 }
1869 }
1870 sched_unpin();
1871 vm_page_unlock_queues();
1872 if (anychanged)
1873 pmap_invalidate_all(pmap);
1874 PMAP_UNLOCK(pmap);
1875 }
1876
1877 /*
1878 * Insert the given physical page (p) at
1879 * the specified virtual address (v) in the
1880 * target physical map with the protection requested.
1881 *
1882 * If specified, the page will be wired down, meaning
1883 * that the related pte can not be reclaimed.
1884 *
1885 * NB: This is the only routine which MAY NOT lazy-evaluate
1886 * or lose information. That is, this routine must actually
1887 * insert this page into the given map NOW.
1888 */
1889 void
1890 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1891 boolean_t wired)
1892 {
1893 vm_paddr_t pa;
1894 register pt_entry_t *pte;
1895 vm_paddr_t opa;
1896 pt_entry_t origpte, newpte;
1897 vm_page_t mpte, om;
1898
1899 va &= PG_FRAME;
1900 #ifdef PMAP_DIAGNOSTIC
1901 if (va > VM_MAX_KERNEL_ADDRESS)
1902 panic("pmap_enter: toobig");
1903 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1904 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1905 #endif
1906
1907 mpte = NULL;
1908
1909 vm_page_lock_queues();
1910 PMAP_LOCK(pmap);
1911 sched_pin();
1912
1913 /*
1914 * In the case that a page table page is not
1915 * resident, we are creating it here.
1916 */
1917 if (va < VM_MAXUSER_ADDRESS) {
1918 mpte = pmap_allocpte(pmap, va, M_WAITOK);
1919 }
1920 #if 0 && defined(PMAP_DIAGNOSTIC)
1921 else {
1922 pd_entry_t *pdeaddr = pmap_pde(pmap, va);
1923 origpte = *pdeaddr;
1924 if ((origpte & PG_V) == 0) {
1925 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n",
1926 pmap->pm_pdir[PTDPTDI], origpte, va);
1927 }
1928 }
1929 #endif
1930
1931 pte = pmap_pte_quick(pmap, va);
1932
1933 /*
1934 * Page Directory table entry not valid, we need a new PT page
1935 */
1936 if (pte == NULL) {
1937 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n",
1938 (uintmax_t)pmap->pm_pdir[PTDPTDI], va);
1939 }
1940
1941 pa = VM_PAGE_TO_PHYS(m);
1942 om = NULL;
1943 origpte = *pte;
1944 opa = origpte & PG_FRAME;
1945
1946 if (origpte & PG_PS) {
1947 /*
1948 * Yes, I know this will truncate upper address bits for PAE,
1949 * but I'm actually more interested in the lower bits
1950 */
1951 printf("pmap_enter: va %p, pte %p, origpte %p\n",
1952 (void *)va, (void *)pte, (void *)(uintptr_t)origpte);
1953 panic("pmap_enter: attempted pmap_enter on 4MB page");
1954 }
1955
1956 /*
1957 * Mapping has not changed, must be protection or wiring change.
1958 */
1959 if (origpte && (opa == pa)) {
1960 /*
1961 * Wiring change, just update stats. We don't worry about
1962 * wiring PT pages as they remain resident as long as there
1963 * are valid mappings in them. Hence, if a user page is wired,
1964 * the PT page will be also.
1965 */
1966 if (wired && ((origpte & PG_W) == 0))
1967 pmap->pm_stats.wired_count++;
1968 else if (!wired && (origpte & PG_W))
1969 pmap->pm_stats.wired_count--;
1970
1971 #if defined(PMAP_DIAGNOSTIC)
1972 if (pmap_nw_modified((pt_entry_t) origpte)) {
1973 printf(
1974 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1975 va, origpte);
1976 }
1977 #endif
1978
1979 /*
1980 * Remove extra pte reference
1981 */
1982 if (mpte)
1983 mpte->wire_count--;
1984
1985 /*
1986 * We might be turning off write access to the page,
1987 * so we go ahead and sense modify status.
1988 */
1989 if (origpte & PG_MANAGED) {
1990 om = m;
1991 pa |= PG_MANAGED;
1992 }
1993 goto validate;
1994 }
1995 /*
1996 * Mapping has changed, invalidate old range and fall through to
1997 * handle validating new mapping.
1998 */
1999 if (opa) {
2000 int err;
2001 if (origpte & PG_W)
2002 pmap->pm_stats.wired_count--;
2003 if (origpte & PG_MANAGED) {
2004 om = PHYS_TO_VM_PAGE(opa);
2005 err = pmap_remove_entry(pmap, om, va);
2006 } else
2007 err = pmap_unuse_pt(pmap, va);
2008 if (err)
2009 panic("pmap_enter: pte vanished, va: 0x%x", va);
2010 } else
2011 pmap->pm_stats.resident_count++;
2012
2013 /*
2014 * Enter on the PV list if part of our managed memory. Note that we
2015 * raise IPL while manipulating pv_table since pmap_enter can be
2016 * called at interrupt time.
2017 */
2018 if (pmap_initialized &&
2019 (m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0) {
2020 pmap_insert_entry(pmap, va, m);
2021 pa |= PG_MANAGED;
2022 }
2023
2024 /*
2025 * Increment counters
2026 */
2027 if (wired)
2028 pmap->pm_stats.wired_count++;
2029
2030 validate:
2031 /*
2032 * Now validate mapping with desired protection/wiring.
2033 */
2034 newpte = (pt_entry_t)(pa | PG_V);
2035 if ((prot & VM_PROT_WRITE) != 0)
2036 newpte |= PG_RW;
2037 if (wired)
2038 newpte |= PG_W;
2039 if (va < VM_MAXUSER_ADDRESS)
2040 newpte |= PG_U;
2041 if (pmap == kernel_pmap)
2042 newpte |= pgeflag;
2043
2044 /*
2045 * if the mapping or permission bits are different, we need
2046 * to update the pte.
2047 */
2048 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2049 if (origpte & PG_MANAGED) {
2050 origpte = pte_load_store(pte, newpte | PG_A);
2051 if ((origpte & PG_M) && pmap_track_modified(va))
2052 vm_page_dirty(om);
2053 if (origpte & PG_A)
2054 vm_page_flag_set(om, PG_REFERENCED);
2055 } else
2056 pte_store(pte, newpte | PG_A);
2057 if (origpte) {
2058 pmap_invalidate_page(pmap, va);
2059 }
2060 }
2061 sched_unpin();
2062 vm_page_unlock_queues();
2063 PMAP_UNLOCK(pmap);
2064 }
2065
2066 /*
2067 * this code makes some *MAJOR* assumptions:
2068 * 1. Current pmap & pmap exists.
2069 * 2. Not wired.
2070 * 3. Read access.
2071 * 4. No page table pages.
2072 * 5. Tlbflush is deferred to calling procedure.
2073 * 6. Page IS managed.
2074 * but is *MUCH* faster than pmap_enter...
2075 */
2076
2077 vm_page_t
2078 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
2079 {
2080 pt_entry_t *pte;
2081 vm_paddr_t pa;
2082
2083 vm_page_lock_queues();
2084 PMAP_LOCK(pmap);
2085
2086 /*
2087 * In the case that a page table page is not
2088 * resident, we are creating it here.
2089 */
2090 if (va < VM_MAXUSER_ADDRESS) {
2091 unsigned ptepindex;
2092 pd_entry_t ptepa;
2093
2094 /*
2095 * Calculate pagetable page index
2096 */
2097 ptepindex = va >> PDRSHIFT;
2098 if (mpte && (mpte->pindex == ptepindex)) {
2099 mpte->wire_count++;
2100 } else {
2101 retry:
2102 /*
2103 * Get the page directory entry
2104 */
2105 ptepa = pmap->pm_pdir[ptepindex];
2106
2107 /*
2108 * If the page table page is mapped, we just increment
2109 * the hold count, and activate it.
2110 */
2111 if (ptepa) {
2112 if (ptepa & PG_PS)
2113 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2114 mpte = PHYS_TO_VM_PAGE(ptepa);
2115 mpte->wire_count++;
2116 } else {
2117 mpte = _pmap_allocpte(pmap, ptepindex,
2118 M_WAITOK);
2119 if (mpte == NULL)
2120 goto retry;
2121 }
2122 }
2123 } else {
2124 mpte = NULL;
2125 }
2126
2127 /*
2128 * This call to vtopte makes the assumption that we are
2129 * entering the page into the current pmap. In order to support
2130 * quick entry into any pmap, one would likely use pmap_pte_quick.
2131 * But that isn't as quick as vtopte.
2132 */
2133 pte = vtopte(va);
2134 if (*pte) {
2135 if (mpte != NULL) {
2136 pmap_unwire_pte_hold(pmap, mpte);
2137 mpte = NULL;
2138 }
2139 goto out;
2140 }
2141
2142 /*
2143 * Enter on the PV list if part of our managed memory. Note that we
2144 * raise IPL while manipulating pv_table since pmap_enter can be
2145 * called at interrupt time.
2146 */
2147 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
2148 pmap_insert_entry(pmap, va, m);
2149
2150 /*
2151 * Increment counters
2152 */
2153 pmap->pm_stats.resident_count++;
2154
2155 pa = VM_PAGE_TO_PHYS(m);
2156
2157 /*
2158 * Now validate mapping with RO protection
2159 */
2160 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2161 pte_store(pte, pa | PG_V | PG_U);
2162 else
2163 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED);
2164 out:
2165 vm_page_unlock_queues();
2166 PMAP_UNLOCK(pmap);
2167 return mpte;
2168 }
2169
2170 /*
2171 * Make a temporary mapping for a physical address. This is only intended
2172 * to be used for panic dumps.
2173 */
2174 void *
2175 pmap_kenter_temporary(vm_paddr_t pa, int i)
2176 {
2177 vm_offset_t va;
2178
2179 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE);
2180 pmap_kenter(va, pa);
2181 #ifndef I386_CPU
2182 invlpg(va);
2183 #else
2184 invltlb();
2185 #endif
2186 return ((void *)crashdumpmap);
2187 }
2188
2189 /*
2190 * This code maps large physical mmap regions into the
2191 * processor address space. Note that some shortcuts
2192 * are taken, but the code works.
2193 */
2194 void
2195 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr,
2196 vm_object_t object, vm_pindex_t pindex,
2197 vm_size_t size)
2198 {
2199 vm_page_t p;
2200
2201 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2202 KASSERT(object->type == OBJT_DEVICE,
2203 ("pmap_object_init_pt: non-device object"));
2204 if (pseflag &&
2205 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) {
2206 int i;
2207 vm_page_t m[1];
2208 unsigned int ptepindex;
2209 int npdes;
2210 pd_entry_t ptepa;
2211
2212 PMAP_LOCK(pmap);
2213 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)])
2214 goto out;
2215 PMAP_UNLOCK(pmap);
2216 retry:
2217 p = vm_page_lookup(object, pindex);
2218 if (p != NULL) {
2219 vm_page_lock_queues();
2220 if (vm_page_sleep_if_busy(p, FALSE, "init4p"))
2221 goto retry;
2222 } else {
2223 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
2224 if (p == NULL)
2225 return;
2226 m[0] = p;
2227
2228 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) {
2229 vm_page_lock_queues();
2230 vm_page_free(p);
2231 vm_page_unlock_queues();
2232 return;
2233 }
2234
2235 p = vm_page_lookup(object, pindex);
2236 vm_page_lock_queues();
2237 vm_page_wakeup(p);
2238 }
2239 vm_page_unlock_queues();
2240
2241 ptepa = VM_PAGE_TO_PHYS(p);
2242 if (ptepa & (NBPDR - 1))
2243 return;
2244
2245 p->valid = VM_PAGE_BITS_ALL;
2246
2247 PMAP_LOCK(pmap);
2248 pmap->pm_stats.resident_count += size >> PAGE_SHIFT;
2249 npdes = size >> PDRSHIFT;
2250 for(i = 0; i < npdes; i++) {
2251 pde_store(&pmap->pm_pdir[ptepindex],
2252 ptepa | PG_U | PG_RW | PG_V | PG_PS);
2253 ptepa += NBPDR;
2254 ptepindex += 1;
2255 }
2256 pmap_invalidate_all(pmap);
2257 out:
2258 PMAP_UNLOCK(pmap);
2259 }
2260 }
2261
2262 /*
2263 * Routine: pmap_change_wiring
2264 * Function: Change the wiring attribute for a map/virtual-address
2265 * pair.
2266 * In/out conditions:
2267 * The mapping must already exist in the pmap.
2268 */
2269 void
2270 pmap_change_wiring(pmap, va, wired)
2271 register pmap_t pmap;
2272 vm_offset_t va;
2273 boolean_t wired;
2274 {
2275 register pt_entry_t *pte;
2276
2277 PMAP_LOCK(pmap);
2278 pte = pmap_pte(pmap, va);
2279
2280 if (wired && !pmap_pte_w(pte))
2281 pmap->pm_stats.wired_count++;
2282 else if (!wired && pmap_pte_w(pte))
2283 pmap->pm_stats.wired_count--;
2284
2285 /*
2286 * Wiring is not a hardware characteristic so there is no need to
2287 * invalidate TLB.
2288 */
2289 pmap_pte_set_w(pte, wired);
2290 pmap_pte_release(pte);
2291 PMAP_UNLOCK(pmap);
2292 }
2293
2294
2295
2296 /*
2297 * Copy the range specified by src_addr/len
2298 * from the source map to the range dst_addr/len
2299 * in the destination map.
2300 *
2301 * This routine is only advisory and need not do anything.
2302 */
2303
2304 void
2305 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len,
2306 vm_offset_t src_addr)
2307 {
2308 vm_offset_t addr;
2309 vm_offset_t end_addr = src_addr + len;
2310 vm_offset_t pdnxt;
2311 vm_page_t m;
2312
2313 if (dst_addr != src_addr)
2314 return;
2315
2316 if (!pmap_is_current(src_pmap))
2317 return;
2318
2319 vm_page_lock_queues();
2320 if (dst_pmap < src_pmap) {
2321 PMAP_LOCK(dst_pmap);
2322 PMAP_LOCK(src_pmap);
2323 } else {
2324 PMAP_LOCK(src_pmap);
2325 PMAP_LOCK(dst_pmap);
2326 }
2327 sched_pin();
2328 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2329 pt_entry_t *src_pte, *dst_pte;
2330 vm_page_t dstmpte, srcmpte;
2331 pd_entry_t srcptepaddr;
2332 unsigned ptepindex;
2333
2334 if (addr >= UPT_MIN_ADDRESS)
2335 panic("pmap_copy: invalid to pmap_copy page tables");
2336
2337 /*
2338 * Don't let optional prefaulting of pages make us go
2339 * way below the low water mark of free pages or way
2340 * above high water mark of used pv entries.
2341 */
2342 if (cnt.v_free_count < cnt.v_free_reserved ||
2343 pv_entry_count > pv_entry_high_water)
2344 break;
2345
2346 pdnxt = (addr + NBPDR) & ~PDRMASK;
2347 ptepindex = addr >> PDRSHIFT;
2348
2349 srcptepaddr = src_pmap->pm_pdir[ptepindex];
2350 if (srcptepaddr == 0)
2351 continue;
2352
2353 if (srcptepaddr & PG_PS) {
2354 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2355 dst_pmap->pm_pdir[ptepindex] = srcptepaddr;
2356 dst_pmap->pm_stats.resident_count +=
2357 NBPDR / PAGE_SIZE;
2358 }
2359 continue;
2360 }
2361
2362 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr);
2363 if (srcmpte->wire_count == 0)
2364 panic("pmap_copy: source page table page is unused");
2365
2366 if (pdnxt > end_addr)
2367 pdnxt = end_addr;
2368
2369 src_pte = vtopte(addr);
2370 while (addr < pdnxt) {
2371 pt_entry_t ptetemp;
2372 ptetemp = *src_pte;
2373 /*
2374 * we only virtual copy managed pages
2375 */
2376 if ((ptetemp & PG_MANAGED) != 0) {
2377 /*
2378 * We have to check after allocpte for the
2379 * pte still being around... allocpte can
2380 * block.
2381 */
2382 dstmpte = pmap_allocpte(dst_pmap, addr,
2383 M_NOWAIT);
2384 if (dstmpte == NULL)
2385 break;
2386 dst_pte = pmap_pte_quick(dst_pmap, addr);
2387 if (*dst_pte == 0) {
2388 /*
2389 * Clear the modified and
2390 * accessed (referenced) bits
2391 * during the copy.
2392 */
2393 m = PHYS_TO_VM_PAGE(ptetemp);
2394 *dst_pte = ptetemp & ~(PG_M | PG_A);
2395 dst_pmap->pm_stats.resident_count++;
2396 pmap_insert_entry(dst_pmap, addr, m);
2397 } else
2398 pmap_unwire_pte_hold(dst_pmap, dstmpte);
2399 if (dstmpte->wire_count >= srcmpte->wire_count)
2400 break;
2401 }
2402 addr += PAGE_SIZE;
2403 src_pte++;
2404 }
2405 }
2406 sched_unpin();
2407 vm_page_unlock_queues();
2408 PMAP_UNLOCK(src_pmap);
2409 PMAP_UNLOCK(dst_pmap);
2410 }
2411
2412 static __inline void
2413 pagezero(void *page)
2414 {
2415 #if defined(I686_CPU)
2416 if (cpu_class == CPUCLASS_686) {
2417 #if defined(CPU_ENABLE_SSE)
2418 if (cpu_feature & CPUID_SSE2)
2419 sse2_pagezero(page);
2420 else
2421 #endif
2422 i686_pagezero(page);
2423 } else
2424 #endif
2425 bzero(page, PAGE_SIZE);
2426 }
2427
2428 /*
2429 * pmap_zero_page zeros the specified hardware page by mapping
2430 * the page into KVM and using bzero to clear its contents.
2431 */
2432 void
2433 pmap_zero_page(vm_page_t m)
2434 {
2435
2436 mtx_lock(&CMAPCADDR12_lock);
2437 if (*CMAP2)
2438 panic("pmap_zero_page: CMAP2 busy");
2439 sched_pin();
2440 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M;
2441 invlcaddr(CADDR2);
2442 pagezero(CADDR2);
2443 *CMAP2 = 0;
2444 sched_unpin();
2445 mtx_unlock(&CMAPCADDR12_lock);
2446 }
2447
2448 /*
2449 * pmap_zero_page_area zeros the specified hardware page by mapping
2450 * the page into KVM and using bzero to clear its contents.
2451 *
2452 * off and size may not cover an area beyond a single hardware page.
2453 */
2454 void
2455 pmap_zero_page_area(vm_page_t m, int off, int size)
2456 {
2457
2458 mtx_lock(&CMAPCADDR12_lock);
2459 if (*CMAP2)
2460 panic("pmap_zero_page: CMAP2 busy");
2461 sched_pin();
2462 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M;
2463 invlcaddr(CADDR2);
2464 if (off == 0 && size == PAGE_SIZE)
2465 pagezero(CADDR2);
2466 else
2467 bzero((char *)CADDR2 + off, size);
2468 *CMAP2 = 0;
2469 sched_unpin();
2470 mtx_unlock(&CMAPCADDR12_lock);
2471 }
2472
2473 /*
2474 * pmap_zero_page_idle zeros the specified hardware page by mapping
2475 * the page into KVM and using bzero to clear its contents. This
2476 * is intended to be called from the vm_pagezero process only and
2477 * outside of Giant.
2478 */
2479 void
2480 pmap_zero_page_idle(vm_page_t m)
2481 {
2482
2483 if (*CMAP3)
2484 panic("pmap_zero_page: CMAP3 busy");
2485 sched_pin();
2486 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M;
2487 invlcaddr(CADDR3);
2488 pagezero(CADDR3);
2489 *CMAP3 = 0;
2490 sched_unpin();
2491 }
2492
2493 /*
2494 * pmap_copy_page copies the specified (machine independent)
2495 * page by mapping the page into virtual memory and using
2496 * bcopy to copy the page, one machine dependent page at a
2497 * time.
2498 */
2499 void
2500 pmap_copy_page(vm_page_t src, vm_page_t dst)
2501 {
2502
2503 mtx_lock(&CMAPCADDR12_lock);
2504 if (*CMAP1)
2505 panic("pmap_copy_page: CMAP1 busy");
2506 if (*CMAP2)
2507 panic("pmap_copy_page: CMAP2 busy");
2508 sched_pin();
2509 #ifdef I386_CPU
2510 invltlb();
2511 #else
2512 invlpg((u_int)CADDR1);
2513 invlpg((u_int)CADDR2);
2514 #endif
2515 *CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A;
2516 *CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M;
2517 bcopy(CADDR1, CADDR2, PAGE_SIZE);
2518 *CMAP1 = 0;
2519 *CMAP2 = 0;
2520 sched_unpin();
2521 mtx_unlock(&CMAPCADDR12_lock);
2522 }
2523
2524 /*
2525 * Returns true if the pmap's pv is one of the first
2526 * 16 pvs linked to from this page. This count may
2527 * be changed upwards or downwards in the future; it
2528 * is only necessary that true be returned for a small
2529 * subset of pmaps for proper page aging.
2530 */
2531 boolean_t
2532 pmap_page_exists_quick(pmap, m)
2533 pmap_t pmap;
2534 vm_page_t m;
2535 {
2536 pv_entry_t pv;
2537 int loops = 0;
2538
2539 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2540 return FALSE;
2541
2542 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2543 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2544 if (pv->pv_pmap == pmap) {
2545 return TRUE;
2546 }
2547 loops++;
2548 if (loops >= 16)
2549 break;
2550 }
2551 return (FALSE);
2552 }
2553
2554 #define PMAP_REMOVE_PAGES_CURPROC_ONLY
2555 /*
2556 * Remove all pages from specified address space
2557 * this aids process exit speeds. Also, this code
2558 * is special cased for current process only, but
2559 * can have the more generic (and slightly slower)
2560 * mode enabled. This is much faster than pmap_remove
2561 * in the case of running down an entire address space.
2562 */
2563 void
2564 pmap_remove_pages(pmap, sva, eva)
2565 pmap_t pmap;
2566 vm_offset_t sva, eva;
2567 {
2568 pt_entry_t *pte, tpte;
2569 vm_page_t m;
2570 pv_entry_t pv, npv;
2571
2572 #ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY
2573 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) {
2574 printf("warning: pmap_remove_pages called with non-current pmap\n");
2575 return;
2576 }
2577 #endif
2578 vm_page_lock_queues();
2579 PMAP_LOCK(pmap);
2580 sched_pin();
2581 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2582
2583 if (pv->pv_va >= eva || pv->pv_va < sva) {
2584 npv = TAILQ_NEXT(pv, pv_plist);
2585 continue;
2586 }
2587
2588 #ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY
2589 pte = vtopte(pv->pv_va);
2590 #else
2591 pte = pmap_pte_quick(pmap, pv->pv_va);
2592 #endif
2593 tpte = *pte;
2594
2595 if (tpte == 0) {
2596 printf("TPTE at %p IS ZERO @ VA %08x\n",
2597 pte, pv->pv_va);
2598 panic("bad pte");
2599 }
2600
2601 /*
2602 * We cannot remove wired pages from a process' mapping at this time
2603 */
2604 if (tpte & PG_W) {
2605 npv = TAILQ_NEXT(pv, pv_plist);
2606 continue;
2607 }
2608
2609 m = PHYS_TO_VM_PAGE(tpte);
2610 KASSERT(m->phys_addr == (tpte & PG_FRAME),
2611 ("vm_page_t %p phys_addr mismatch %016jx %016jx",
2612 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte));
2613
2614 KASSERT(m < &vm_page_array[vm_page_array_size],
2615 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte));
2616
2617 pmap->pm_stats.resident_count--;
2618
2619 pte_clear(pte);
2620
2621 /*
2622 * Update the vm_page_t clean and reference bits.
2623 */
2624 if (tpte & PG_M) {
2625 vm_page_dirty(m);
2626 }
2627
2628 npv = TAILQ_NEXT(pv, pv_plist);
2629 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2630
2631 m->md.pv_list_count--;
2632 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2633 if (TAILQ_EMPTY(&m->md.pv_list))
2634 vm_page_flag_clear(m, PG_WRITEABLE);
2635
2636 pmap_unuse_pt(pmap, pv->pv_va);
2637 free_pv_entry(pv);
2638 }
2639 sched_unpin();
2640 pmap_invalidate_all(pmap);
2641 PMAP_UNLOCK(pmap);
2642 vm_page_unlock_queues();
2643 }
2644
2645 /*
2646 * pmap_is_modified:
2647 *
2648 * Return whether or not the specified physical page was modified
2649 * in any physical maps.
2650 */
2651 boolean_t
2652 pmap_is_modified(vm_page_t m)
2653 {
2654 pv_entry_t pv;
2655 pt_entry_t *pte;
2656 boolean_t rv;
2657
2658 rv = FALSE;
2659 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2660 return (rv);
2661
2662 sched_pin();
2663 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2664 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2665 /*
2666 * if the bit being tested is the modified bit, then
2667 * mark clean_map and ptes as never
2668 * modified.
2669 */
2670 if (!pmap_track_modified(pv->pv_va))
2671 continue;
2672 #if defined(PMAP_DIAGNOSTIC)
2673 if (!pv->pv_pmap) {
2674 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2675 continue;
2676 }
2677 #endif
2678 PMAP_LOCK(pv->pv_pmap);
2679 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2680 rv = (*pte & PG_M) != 0;
2681 PMAP_UNLOCK(pv->pv_pmap);
2682 if (rv)
2683 break;
2684 }
2685 sched_unpin();
2686 return (rv);
2687 }
2688
2689 /*
2690 * pmap_is_prefaultable:
2691 *
2692 * Return whether or not the specified virtual address is elgible
2693 * for prefault.
2694 */
2695 boolean_t
2696 pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr)
2697 {
2698 pt_entry_t *pte;
2699 boolean_t rv;
2700
2701 rv = FALSE;
2702 PMAP_LOCK(pmap);
2703 if (*pmap_pde(pmap, addr)) {
2704 pte = vtopte(addr);
2705 rv = *pte == 0;
2706 }
2707 PMAP_UNLOCK(pmap);
2708 return (rv);
2709 }
2710
2711 /*
2712 * Clear the given bit in each of the given page's ptes. The bit is
2713 * expressed as a 32-bit mask. Consequently, if the pte is 64 bits in
2714 * size, only a bit within the least significant 32 can be cleared.
2715 */
2716 static __inline void
2717 pmap_clear_ptes(vm_page_t m, int bit)
2718 {
2719 register pv_entry_t pv;
2720 pt_entry_t pbits, *pte;
2721
2722 if (!pmap_initialized || (m->flags & PG_FICTITIOUS) ||
2723 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0))
2724 return;
2725
2726 sched_pin();
2727 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2728 /*
2729 * Loop over all current mappings setting/clearing as appropos If
2730 * setting RO do we need to clear the VAC?
2731 */
2732 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2733 /*
2734 * don't write protect pager mappings
2735 */
2736 if (bit == PG_RW) {
2737 if (!pmap_track_modified(pv->pv_va))
2738 continue;
2739 }
2740
2741 #if defined(PMAP_DIAGNOSTIC)
2742 if (!pv->pv_pmap) {
2743 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2744 continue;
2745 }
2746 #endif
2747
2748 PMAP_LOCK(pv->pv_pmap);
2749 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2750 retry:
2751 pbits = *pte;
2752 if (pbits & bit) {
2753 if (bit == PG_RW) {
2754 /*
2755 * Regardless of whether a pte is 32 or 64 bits
2756 * in size, PG_RW and PG_M are among the least
2757 * significant 32 bits.
2758 */
2759 if (!atomic_cmpset_int((u_int *)pte, pbits,
2760 pbits & ~(PG_RW | PG_M)))
2761 goto retry;
2762 if (pbits & PG_M) {
2763 vm_page_dirty(m);
2764 }
2765 } else {
2766 atomic_clear_int((u_int *)pte, bit);
2767 }
2768 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
2769 }
2770 PMAP_UNLOCK(pv->pv_pmap);
2771 }
2772 if (bit == PG_RW)
2773 vm_page_flag_clear(m, PG_WRITEABLE);
2774 sched_unpin();
2775 }
2776
2777 /*
2778 * pmap_page_protect:
2779 *
2780 * Lower the permission for all mappings to a given page.
2781 */
2782 void
2783 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2784 {
2785 if ((prot & VM_PROT_WRITE) == 0) {
2786 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2787 pmap_clear_ptes(m, PG_RW);
2788 } else {
2789 pmap_remove_all(m);
2790 }
2791 }
2792 }
2793
2794 /*
2795 * pmap_ts_referenced:
2796 *
2797 * Return a count of reference bits for a page, clearing those bits.
2798 * It is not necessary for every reference bit to be cleared, but it
2799 * is necessary that 0 only be returned when there are truly no
2800 * reference bits set.
2801 *
2802 * XXX: The exact number of bits to check and clear is a matter that
2803 * should be tested and standardized at some point in the future for
2804 * optimal aging of shared pages.
2805 */
2806 int
2807 pmap_ts_referenced(vm_page_t m)
2808 {
2809 register pv_entry_t pv, pvf, pvn;
2810 pt_entry_t *pte;
2811 pt_entry_t v;
2812 int rtval = 0;
2813
2814 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2815 return (rtval);
2816
2817 sched_pin();
2818 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2819 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2820
2821 pvf = pv;
2822
2823 do {
2824 pvn = TAILQ_NEXT(pv, pv_list);
2825
2826 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2827
2828 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2829
2830 if (!pmap_track_modified(pv->pv_va))
2831 continue;
2832
2833 PMAP_LOCK(pv->pv_pmap);
2834 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2835
2836 if (pte && ((v = pte_load(pte)) & PG_A) != 0) {
2837 atomic_clear_int((u_int *)pte, PG_A);
2838 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
2839
2840 rtval++;
2841 if (rtval > 4) {
2842 PMAP_UNLOCK(pv->pv_pmap);
2843 break;
2844 }
2845 }
2846 PMAP_UNLOCK(pv->pv_pmap);
2847 } while ((pv = pvn) != NULL && pv != pvf);
2848 }
2849 sched_unpin();
2850
2851 return (rtval);
2852 }
2853
2854 /*
2855 * Clear the modify bits on the specified physical page.
2856 */
2857 void
2858 pmap_clear_modify(vm_page_t m)
2859 {
2860 pmap_clear_ptes(m, PG_M);
2861 }
2862
2863 /*
2864 * pmap_clear_reference:
2865 *
2866 * Clear the reference bit on the specified physical page.
2867 */
2868 void
2869 pmap_clear_reference(vm_page_t m)
2870 {
2871 pmap_clear_ptes(m, PG_A);
2872 }
2873
2874 /*
2875 * Miscellaneous support routines follow
2876 */
2877
2878 /*
2879 * Map a set of physical memory pages into the kernel virtual
2880 * address space. Return a pointer to where it is mapped. This
2881 * routine is intended to be used for mapping device memory,
2882 * NOT real memory.
2883 */
2884 void *
2885 pmap_mapdev(pa, size)
2886 vm_paddr_t pa;
2887 vm_size_t size;
2888 {
2889 vm_offset_t va, tmpva, offset;
2890
2891 offset = pa & PAGE_MASK;
2892 size = roundup(offset + size, PAGE_SIZE);
2893 pa = pa & PG_FRAME;
2894
2895 if (pa < KERNLOAD && pa + size <= KERNLOAD)
2896 va = KERNBASE + pa;
2897 else
2898 va = kmem_alloc_nofault(kernel_map, size);
2899 if (!va)
2900 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2901
2902 for (tmpva = va; size > 0; ) {
2903 pmap_kenter(tmpva, pa);
2904 size -= PAGE_SIZE;
2905 tmpva += PAGE_SIZE;
2906 pa += PAGE_SIZE;
2907 }
2908 pmap_invalidate_range(kernel_pmap, va, tmpva);
2909 return ((void *)(va + offset));
2910 }
2911
2912 void
2913 pmap_unmapdev(va, size)
2914 vm_offset_t va;
2915 vm_size_t size;
2916 {
2917 vm_offset_t base, offset, tmpva;
2918
2919 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD)
2920 return;
2921 base = va & PG_FRAME;
2922 offset = va & PAGE_MASK;
2923 size = roundup(offset + size, PAGE_SIZE);
2924 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE)
2925 pmap_kremove(tmpva);
2926 pmap_invalidate_range(kernel_pmap, va, tmpva);
2927 kmem_free(kernel_map, base, size);
2928 }
2929
2930 /*
2931 * perform the pmap work for mincore
2932 */
2933 int
2934 pmap_mincore(pmap, addr)
2935 pmap_t pmap;
2936 vm_offset_t addr;
2937 {
2938 pt_entry_t *ptep, pte;
2939 vm_page_t m;
2940 int val = 0;
2941
2942 PMAP_LOCK(pmap);
2943 ptep = pmap_pte(pmap, addr);
2944 pte = (ptep != NULL) ? *ptep : 0;
2945 pmap_pte_release(ptep);
2946 PMAP_UNLOCK(pmap);
2947
2948 if (pte != 0) {
2949 vm_paddr_t pa;
2950
2951 val = MINCORE_INCORE;
2952 if ((pte & PG_MANAGED) == 0)
2953 return val;
2954
2955 pa = pte & PG_FRAME;
2956
2957 m = PHYS_TO_VM_PAGE(pa);
2958
2959 /*
2960 * Modified by us
2961 */
2962 if (pte & PG_M)
2963 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2964 else {
2965 /*
2966 * Modified by someone else
2967 */
2968 vm_page_lock_queues();
2969 if (m->dirty || pmap_is_modified(m))
2970 val |= MINCORE_MODIFIED_OTHER;
2971 vm_page_unlock_queues();
2972 }
2973 /*
2974 * Referenced by us
2975 */
2976 if (pte & PG_A)
2977 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2978 else {
2979 /*
2980 * Referenced by someone else
2981 */
2982 vm_page_lock_queues();
2983 if ((m->flags & PG_REFERENCED) ||
2984 pmap_ts_referenced(m)) {
2985 val |= MINCORE_REFERENCED_OTHER;
2986 vm_page_flag_set(m, PG_REFERENCED);
2987 }
2988 vm_page_unlock_queues();
2989 }
2990 }
2991 return val;
2992 }
2993
2994 void
2995 pmap_activate(struct thread *td)
2996 {
2997 struct proc *p = td->td_proc;
2998 pmap_t pmap, oldpmap;
2999 u_int32_t cr3;
3000
3001 critical_enter();
3002 pmap = vmspace_pmap(td->td_proc->p_vmspace);
3003 oldpmap = PCPU_GET(curpmap);
3004 #if defined(SMP)
3005 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask));
3006 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask));
3007 #else
3008 oldpmap->pm_active &= ~1;
3009 pmap->pm_active |= 1;
3010 #endif
3011 #ifdef PAE
3012 cr3 = vtophys(pmap->pm_pdpt);
3013 #else
3014 cr3 = vtophys(pmap->pm_pdir);
3015 #endif
3016 /* XXXKSE this is wrong.
3017 * pmap_activate is for the current thread on the current cpu
3018 */
3019 if (p->p_flag & P_SA) {
3020 /* Make sure all other cr3 entries are updated. */
3021 /* what if they are running? XXXKSE (maybe abort them) */
3022 FOREACH_THREAD_IN_PROC(p, td) {
3023 td->td_pcb->pcb_cr3 = cr3;
3024 }
3025 } else {
3026 td->td_pcb->pcb_cr3 = cr3;
3027 }
3028 load_cr3(cr3);
3029 PCPU_SET(curpmap, pmap);
3030 critical_exit();
3031 }
3032
3033 vm_offset_t
3034 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3035 {
3036
3037 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3038 return addr;
3039 }
3040
3041 addr = (addr + PDRMASK) & ~PDRMASK;
3042 return addr;
3043 }
3044
3045
3046 #if defined(PMAP_DEBUG)
3047 pmap_pid_dump(int pid)
3048 {
3049 pmap_t pmap;
3050 struct proc *p;
3051 int npte = 0;
3052 int index;
3053
3054 sx_slock(&allproc_lock);
3055 LIST_FOREACH(p, &allproc, p_list) {
3056 if (p->p_pid != pid)
3057 continue;
3058
3059 if (p->p_vmspace) {
3060 int i,j;
3061 index = 0;
3062 pmap = vmspace_pmap(p->p_vmspace);
3063 for (i = 0; i < NPDEPTD; i++) {
3064 pd_entry_t *pde;
3065 pt_entry_t *pte;
3066 vm_offset_t base = i << PDRSHIFT;
3067
3068 pde = &pmap->pm_pdir[i];
3069 if (pde && pmap_pde_v(pde)) {
3070 for (j = 0; j < NPTEPG; j++) {
3071 vm_offset_t va = base + (j << PAGE_SHIFT);
3072 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) {
3073 if (index) {
3074 index = 0;
3075 printf("\n");
3076 }
3077 sx_sunlock(&allproc_lock);
3078 return npte;
3079 }
3080 pte = pmap_pte(pmap, va);
3081 if (pte && pmap_pte_v(pte)) {
3082 pt_entry_t pa;
3083 vm_page_t m;
3084 pa = *pte;
3085 m = PHYS_TO_VM_PAGE(pa);
3086 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x",
3087 va, pa, m->hold_count, m->wire_count, m->flags);
3088 npte++;
3089 index++;
3090 if (index >= 2) {
3091 index = 0;
3092 printf("\n");
3093 } else {
3094 printf(" ");
3095 }
3096 }
3097 }
3098 }
3099 }
3100 }
3101 }
3102 sx_sunlock(&allproc_lock);
3103 return npte;
3104 }
3105 #endif
3106
3107 #if defined(DEBUG)
3108
3109 static void pads(pmap_t pm);
3110 void pmap_pvdump(vm_offset_t pa);
3111
3112 /* print address space of pmap*/
3113 static void
3114 pads(pm)
3115 pmap_t pm;
3116 {
3117 int i, j;
3118 vm_paddr_t va;
3119 pt_entry_t *ptep;
3120
3121 if (pm == kernel_pmap)
3122 return;
3123 for (i = 0; i < NPDEPTD; i++)
3124 if (pm->pm_pdir[i])
3125 for (j = 0; j < NPTEPG; j++) {
3126 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3127 if (pm == kernel_pmap && va < KERNBASE)
3128 continue;
3129 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS)
3130 continue;
3131 ptep = pmap_pte(pm, va);
3132 if (pmap_pte_v(ptep))
3133 printf("%x:%x ", va, *ptep);
3134 };
3135
3136 }
3137
3138 void
3139 pmap_pvdump(pa)
3140 vm_paddr_t pa;
3141 {
3142 pv_entry_t pv;
3143 vm_page_t m;
3144
3145 printf("pa %x", pa);
3146 m = PHYS_TO_VM_PAGE(pa);
3147 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3148 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);
3149 pads(pv->pv_pmap);
3150 }
3151 printf(" ");
3152 }
3153 #endif
Cache object: 2abd7bb2668173ae3f742d10222e22c5
|