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 * Copyright (c) 2005 Alan L. Cox <alc@cs.rice.edu>
9 * All rights reserved.
10 *
11 * This code is derived from software contributed to Berkeley by
12 * the Systems Programming Group of the University of Utah Computer
13 * Science Department and William Jolitz of UUNET Technologies Inc.
14 *
15 * Redistribution and use in source and binary forms, with or without
16 * modification, are permitted provided that the following conditions
17 * are met:
18 * 1. Redistributions of source code must retain the above copyright
19 * notice, this list of conditions and the following disclaimer.
20 * 2. Redistributions in binary form must reproduce the above copyright
21 * notice, this list of conditions and the following disclaimer in the
22 * documentation and/or other materials provided with the distribution.
23 * 3. All advertising materials mentioning features or use of this software
24 * must display the following acknowledgement:
25 * This product includes software developed by the University of
26 * California, Berkeley and its contributors.
27 * 4. Neither the name of the University nor the names of its contributors
28 * may be used to endorse or promote products derived from this software
29 * without specific prior written permission.
30 *
31 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
32 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
33 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
34 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
35 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
36 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
37 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
38 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
39 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
40 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
41 * SUCH DAMAGE.
42 *
43 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
44 */
45 /*-
46 * Copyright (c) 2003 Networks Associates Technology, Inc.
47 * All rights reserved.
48 *
49 * This software was developed for the FreeBSD Project by Jake Burkholder,
50 * Safeport Network Services, and Network Associates Laboratories, the
51 * Security Research Division of Network Associates, Inc. under
52 * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA
53 * CHATS research program.
54 *
55 * Redistribution and use in source and binary forms, with or without
56 * modification, are permitted provided that the following conditions
57 * are met:
58 * 1. Redistributions of source code must retain the above copyright
59 * notice, this list of conditions and the following disclaimer.
60 * 2. Redistributions in binary form must reproduce the above copyright
61 * notice, this list of conditions and the following disclaimer in the
62 * documentation and/or other materials provided with the distribution.
63 *
64 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
65 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
66 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
67 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
68 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
69 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
70 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
71 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
72 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
73 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
74 * SUCH DAMAGE.
75 */
76
77 #include <sys/cdefs.h>
78 __FBSDID("$FreeBSD: releng/6.2/sys/i386/i386/pmap.c 162707 2006-09-27 18:10:16Z alc $");
79
80 /*
81 * Manages physical address maps.
82 *
83 * In addition to hardware address maps, this
84 * module is called upon to provide software-use-only
85 * maps which may or may not be stored in the same
86 * form as hardware maps. These pseudo-maps are
87 * used to store intermediate results from copy
88 * operations to and from address spaces.
89 *
90 * Since the information managed by this module is
91 * also stored by the logical address mapping module,
92 * this module may throw away valid virtual-to-physical
93 * mappings at almost any time. However, invalidations
94 * of virtual-to-physical mappings must be done as
95 * requested.
96 *
97 * In order to cope with hardware architectures which
98 * make virtual-to-physical map invalidates expensive,
99 * this module may delay invalidate or reduced protection
100 * operations until such time as they are actually
101 * necessary. This module is given full information as
102 * to which processors are currently using which maps,
103 * and to when physical maps must be made correct.
104 */
105
106 #include "opt_cpu.h"
107 #include "opt_pmap.h"
108 #include "opt_msgbuf.h"
109 #include "opt_xbox.h"
110
111 #include <sys/param.h>
112 #include <sys/systm.h>
113 #include <sys/kernel.h>
114 #include <sys/lock.h>
115 #include <sys/malloc.h>
116 #include <sys/mman.h>
117 #include <sys/msgbuf.h>
118 #include <sys/mutex.h>
119 #include <sys/proc.h>
120 #include <sys/sx.h>
121 #include <sys/vmmeter.h>
122 #include <sys/sched.h>
123 #include <sys/sysctl.h>
124 #ifdef SMP
125 #include <sys/smp.h>
126 #endif
127
128 #ifdef XBOX
129 #include <machine/xbox.h>
130 #endif
131
132 #include <vm/vm.h>
133 #include <vm/vm_param.h>
134 #include <vm/vm_kern.h>
135 #include <vm/vm_page.h>
136 #include <vm/vm_map.h>
137 #include <vm/vm_object.h>
138 #include <vm/vm_extern.h>
139 #include <vm/vm_pageout.h>
140 #include <vm/vm_pager.h>
141 #include <vm/uma.h>
142
143 #include <machine/cpu.h>
144 #include <machine/cputypes.h>
145 #include <machine/md_var.h>
146 #include <machine/pcb.h>
147 #include <machine/specialreg.h>
148 #ifdef SMP
149 #include <machine/smp.h>
150 #endif
151
152 #if !defined(CPU_DISABLE_SSE) && defined(I686_CPU)
153 #define CPU_ENABLE_SSE
154 #endif
155
156 #ifndef PMAP_SHPGPERPROC
157 #define PMAP_SHPGPERPROC 200
158 #endif
159
160 #if defined(DIAGNOSTIC)
161 #define PMAP_DIAGNOSTIC
162 #endif
163
164 #if !defined(PMAP_DIAGNOSTIC)
165 #define PMAP_INLINE __inline
166 #else
167 #define PMAP_INLINE
168 #endif
169
170 /*
171 * Get PDEs and PTEs for user/kernel address space
172 */
173 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
174 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
175
176 #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0)
177 #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0)
178 #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0)
179 #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0)
180 #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0)
181
182 #define pmap_pte_set_w(pte, v) ((v) ? atomic_set_int((u_int *)(pte), PG_W) : \
183 atomic_clear_int((u_int *)(pte), PG_W))
184 #define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v)))
185
186 struct pmap kernel_pmap_store;
187 LIST_HEAD(pmaplist, pmap);
188 static struct pmaplist allpmaps;
189 static struct mtx allpmaps_lock;
190
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 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 struct sysmaps {
217 struct mtx lock;
218 pt_entry_t *CMAP1;
219 pt_entry_t *CMAP2;
220 caddr_t CADDR1;
221 caddr_t CADDR2;
222 };
223 static struct sysmaps sysmaps_pcpu[MAXCPU];
224 pt_entry_t *CMAP1 = 0;
225 static pt_entry_t *CMAP3;
226 caddr_t CADDR1 = 0, ptvmmap = 0;
227 static caddr_t CADDR3;
228 struct msgbuf *msgbufp = 0;
229
230 /*
231 * Crashdump maps.
232 */
233 static caddr_t crashdumpmap;
234
235 #ifdef SMP
236 extern pt_entry_t *SMPpt;
237 #endif
238 static pt_entry_t *PMAP1 = 0, *PMAP2;
239 static pt_entry_t *PADDR1 = 0, *PADDR2;
240 #ifdef SMP
241 static int PMAP1cpu;
242 static int PMAP1changedcpu;
243 SYSCTL_INT(_debug, OID_AUTO, PMAP1changedcpu, CTLFLAG_RD,
244 &PMAP1changedcpu, 0,
245 "Number of times pmap_pte_quick changed CPU with same PMAP1");
246 #endif
247 static int PMAP1changed;
248 SYSCTL_INT(_debug, OID_AUTO, PMAP1changed, CTLFLAG_RD,
249 &PMAP1changed, 0,
250 "Number of times pmap_pte_quick changed PMAP1");
251 static int PMAP1unchanged;
252 SYSCTL_INT(_debug, OID_AUTO, PMAP1unchanged, CTLFLAG_RD,
253 &PMAP1unchanged, 0,
254 "Number of times pmap_pte_quick didn't change PMAP1");
255 static struct mtx PMAP2mutex;
256
257 static PMAP_INLINE void free_pv_entry(pv_entry_t pv);
258 static pv_entry_t get_pv_entry(void);
259 static void pmap_clear_ptes(vm_page_t m, int bit);
260
261 static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva);
262 static void pmap_remove_page(struct pmap *pmap, vm_offset_t va);
263 static void pmap_remove_entry(struct pmap *pmap, vm_page_t m,
264 vm_offset_t va);
265 static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m);
266
267 static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags);
268
269 static vm_page_t _pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags);
270 static int _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m);
271 static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va);
272 static void pmap_pte_release(pt_entry_t *pte);
273 static int pmap_unuse_pt(pmap_t, vm_offset_t);
274 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
275 #ifdef PAE
276 static void *pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait);
277 #endif
278
279 CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t));
280 CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t));
281
282 /*
283 * Move the kernel virtual free pointer to the next
284 * 4MB. This is used to help improve performance
285 * by using a large (4MB) page for much of the kernel
286 * (.text, .data, .bss)
287 */
288 static vm_offset_t
289 pmap_kmem_choose(vm_offset_t addr)
290 {
291 vm_offset_t newaddr = addr;
292
293 #ifndef DISABLE_PSE
294 if (cpu_feature & CPUID_PSE)
295 newaddr = (addr + PDRMASK) & ~PDRMASK;
296 #endif
297 return newaddr;
298 }
299
300 /*
301 * Bootstrap the system enough to run with virtual memory.
302 *
303 * On the i386 this is called after mapping has already been enabled
304 * and just syncs the pmap module with what has already been done.
305 * [We can't call it easily with mapping off since the kernel is not
306 * mapped with PA == VA, hence we would have to relocate every address
307 * from the linked base (virtual) address "KERNBASE" to the actual
308 * (physical) address starting relative to 0]
309 */
310 void
311 pmap_bootstrap(firstaddr, loadaddr)
312 vm_paddr_t firstaddr;
313 vm_paddr_t loadaddr;
314 {
315 vm_offset_t va;
316 pt_entry_t *pte, *unused;
317 struct sysmaps *sysmaps;
318 int i;
319
320 /*
321 * XXX The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too
322 * large. It should instead be correctly calculated in locore.s and
323 * not based on 'first' (which is a physical address, not a virtual
324 * address, for the start of unused physical memory). The kernel
325 * page tables are NOT double mapped and thus should not be included
326 * in this calculation.
327 */
328 virtual_avail = (vm_offset_t) KERNBASE + firstaddr;
329 virtual_avail = pmap_kmem_choose(virtual_avail);
330
331 virtual_end = VM_MAX_KERNEL_ADDRESS;
332
333 /*
334 * Initialize the kernel pmap (which is statically allocated).
335 */
336 PMAP_LOCK_INIT(kernel_pmap);
337 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + (u_int)IdlePTD);
338 #ifdef PAE
339 kernel_pmap->pm_pdpt = (pdpt_entry_t *) (KERNBASE + (u_int)IdlePDPT);
340 #endif
341 kernel_pmap->pm_active = -1; /* don't allow deactivation */
342 TAILQ_INIT(&kernel_pmap->pm_pvlist);
343 LIST_INIT(&allpmaps);
344 mtx_init(&allpmaps_lock, "allpmaps", NULL, MTX_SPIN);
345 mtx_lock_spin(&allpmaps_lock);
346 LIST_INSERT_HEAD(&allpmaps, kernel_pmap, pm_list);
347 mtx_unlock_spin(&allpmaps_lock);
348 nkpt = NKPT;
349
350 /*
351 * Reserve some special page table entries/VA space for temporary
352 * mapping of pages.
353 */
354 #define SYSMAP(c, p, v, n) \
355 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
356
357 va = virtual_avail;
358 pte = vtopte(va);
359
360 /*
361 * CMAP1/CMAP2 are used for zeroing and copying pages.
362 * CMAP3 is used for the idle process page zeroing.
363 */
364 for (i = 0; i < MAXCPU; i++) {
365 sysmaps = &sysmaps_pcpu[i];
366 mtx_init(&sysmaps->lock, "SYSMAPS", NULL, MTX_DEF);
367 SYSMAP(caddr_t, sysmaps->CMAP1, sysmaps->CADDR1, 1)
368 SYSMAP(caddr_t, sysmaps->CMAP2, sysmaps->CADDR2, 1)
369 }
370 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
371 SYSMAP(caddr_t, CMAP3, CADDR3, 1)
372 *CMAP3 = 0;
373
374 /*
375 * Crashdump maps.
376 */
377 SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS)
378
379 /*
380 * ptvmmap is used for reading arbitrary physical pages via /dev/mem.
381 */
382 SYSMAP(caddr_t, unused, ptvmmap, 1)
383
384 /*
385 * msgbufp is used to map the system message buffer.
386 */
387 SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(MSGBUF_SIZE)))
388
389 /*
390 * ptemap is used for pmap_pte_quick
391 */
392 SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1);
393 SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1);
394
395 mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF);
396
397 virtual_avail = va;
398
399 *CMAP1 = 0;
400
401 #ifdef XBOX
402 /* FIXME: This is gross, but needed for the XBOX. Since we are in such
403 * an early stadium, we cannot yet neatly map video memory ... :-(
404 * Better fixes are very welcome!
405 */
406 if (!arch_i386_is_xbox)
407 #endif
408 for (i = 0; i < NKPT; i++)
409 PTD[i] = 0;
410
411 /* Turn on PG_G on kernel page(s) */
412 pmap_set_pg();
413 }
414
415 /*
416 * Set PG_G on kernel pages. Only the BSP calls this when SMP is turned on.
417 */
418 void
419 pmap_set_pg(void)
420 {
421 pd_entry_t pdir;
422 pt_entry_t *pte;
423 vm_offset_t va, endva;
424 int i;
425
426 if (pgeflag == 0)
427 return;
428
429 i = KERNLOAD/NBPDR;
430 endva = KERNBASE + KERNend;
431
432 if (pseflag) {
433 va = KERNBASE + KERNLOAD;
434 while (va < endva) {
435 pdir = kernel_pmap->pm_pdir[KPTDI+i];
436 pdir |= pgeflag;
437 kernel_pmap->pm_pdir[KPTDI+i] = PTD[KPTDI+i] = pdir;
438 invltlb(); /* Play it safe, invltlb() every time */
439 i++;
440 va += NBPDR;
441 }
442 } else {
443 va = (vm_offset_t)btext;
444 while (va < endva) {
445 pte = vtopte(va);
446 if (*pte)
447 *pte |= pgeflag;
448 invltlb(); /* Play it safe, invltlb() every time */
449 va += PAGE_SIZE;
450 }
451 }
452 }
453
454 /*
455 * Initialize a vm_page's machine-dependent fields.
456 */
457 void
458 pmap_page_init(vm_page_t m)
459 {
460
461 TAILQ_INIT(&m->md.pv_list);
462 m->md.pv_list_count = 0;
463 }
464
465 #ifdef PAE
466
467 static MALLOC_DEFINE(M_PMAPPDPT, "pmap", "pmap pdpt");
468
469 static void *
470 pmap_pdpt_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
471 {
472 *flags = UMA_SLAB_PRIV;
473 return (contigmalloc(PAGE_SIZE, M_PMAPPDPT, 0, 0x0ULL, 0xffffffffULL,
474 1, 0));
475 }
476 #endif
477
478 /*
479 * Initialize the pmap module.
480 * Called by vm_init, to initialize any structures that the pmap
481 * system needs to map virtual memory.
482 */
483 void
484 pmap_init(void)
485 {
486 int shpgperproc = PMAP_SHPGPERPROC;
487
488 /*
489 * Initialize the address space (zone) for the pv entries. Set a
490 * high water mark so that the system can recover from excessive
491 * numbers of pv entries.
492 */
493 pvzone = uma_zcreate("PV ENTRY", sizeof(struct pv_entry), NULL, NULL,
494 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE);
495 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
496 pv_entry_max = shpgperproc * maxproc + cnt.v_page_count;
497 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
498 pv_entry_high_water = 9 * (pv_entry_max / 10);
499 uma_zone_set_obj(pvzone, &pvzone_obj, pv_entry_max);
500
501 #ifdef PAE
502 pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL,
503 NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1,
504 UMA_ZONE_VM | UMA_ZONE_NOFREE);
505 uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf);
506 #endif
507 }
508
509 void
510 pmap_init2()
511 {
512 }
513
514
515 /***************************************************
516 * Low level helper routines.....
517 ***************************************************/
518
519 #if defined(PMAP_DIAGNOSTIC)
520
521 /*
522 * This code checks for non-writeable/modified pages.
523 * This should be an invalid condition.
524 */
525 static int
526 pmap_nw_modified(pt_entry_t ptea)
527 {
528 int pte;
529
530 pte = (int) ptea;
531
532 if ((pte & (PG_M|PG_RW)) == PG_M)
533 return 1;
534 else
535 return 0;
536 }
537 #endif
538
539
540 /*
541 * this routine defines the region(s) of memory that should
542 * not be tested for the modified bit.
543 */
544 static PMAP_INLINE int
545 pmap_track_modified(vm_offset_t va)
546 {
547 if ((va < kmi.clean_sva) || (va >= kmi.clean_eva))
548 return 1;
549 else
550 return 0;
551 }
552
553 #ifdef SMP
554 /*
555 * For SMP, these functions have to use the IPI mechanism for coherence.
556 */
557 void
558 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
559 {
560 u_int cpumask;
561 u_int other_cpus;
562
563 if (smp_started) {
564 if (!(read_eflags() & PSL_I))
565 panic("%s: interrupts disabled", __func__);
566 mtx_lock_spin(&smp_ipi_mtx);
567 } else
568 critical_enter();
569 /*
570 * We need to disable interrupt preemption but MUST NOT have
571 * interrupts disabled here.
572 * XXX we may need to hold schedlock to get a coherent pm_active
573 * XXX critical sections disable interrupts again
574 */
575 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) {
576 invlpg(va);
577 smp_invlpg(va);
578 } else {
579 cpumask = PCPU_GET(cpumask);
580 other_cpus = PCPU_GET(other_cpus);
581 if (pmap->pm_active & cpumask)
582 invlpg(va);
583 if (pmap->pm_active & other_cpus)
584 smp_masked_invlpg(pmap->pm_active & other_cpus, va);
585 }
586 if (smp_started)
587 mtx_unlock_spin(&smp_ipi_mtx);
588 else
589 critical_exit();
590 }
591
592 void
593 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
594 {
595 u_int cpumask;
596 u_int other_cpus;
597 vm_offset_t addr;
598
599 if (smp_started) {
600 if (!(read_eflags() & PSL_I))
601 panic("%s: interrupts disabled", __func__);
602 mtx_lock_spin(&smp_ipi_mtx);
603 } else
604 critical_enter();
605 /*
606 * We need to disable interrupt preemption but MUST NOT have
607 * interrupts disabled here.
608 * XXX we may need to hold schedlock to get a coherent pm_active
609 * XXX critical sections disable interrupts again
610 */
611 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) {
612 for (addr = sva; addr < eva; addr += PAGE_SIZE)
613 invlpg(addr);
614 smp_invlpg_range(sva, eva);
615 } else {
616 cpumask = PCPU_GET(cpumask);
617 other_cpus = PCPU_GET(other_cpus);
618 if (pmap->pm_active & cpumask)
619 for (addr = sva; addr < eva; addr += PAGE_SIZE)
620 invlpg(addr);
621 if (pmap->pm_active & other_cpus)
622 smp_masked_invlpg_range(pmap->pm_active & other_cpus,
623 sva, eva);
624 }
625 if (smp_started)
626 mtx_unlock_spin(&smp_ipi_mtx);
627 else
628 critical_exit();
629 }
630
631 void
632 pmap_invalidate_all(pmap_t pmap)
633 {
634 u_int cpumask;
635 u_int other_cpus;
636
637 if (smp_started) {
638 if (!(read_eflags() & PSL_I))
639 panic("%s: interrupts disabled", __func__);
640 mtx_lock_spin(&smp_ipi_mtx);
641 } else
642 critical_enter();
643 /*
644 * We need to disable interrupt preemption but MUST NOT have
645 * interrupts disabled here.
646 * XXX we may need to hold schedlock to get a coherent pm_active
647 * XXX critical sections disable interrupts again
648 */
649 if (pmap == kernel_pmap || pmap->pm_active == all_cpus) {
650 invltlb();
651 smp_invltlb();
652 } else {
653 cpumask = PCPU_GET(cpumask);
654 other_cpus = PCPU_GET(other_cpus);
655 if (pmap->pm_active & cpumask)
656 invltlb();
657 if (pmap->pm_active & other_cpus)
658 smp_masked_invltlb(pmap->pm_active & other_cpus);
659 }
660 if (smp_started)
661 mtx_unlock_spin(&smp_ipi_mtx);
662 else
663 critical_exit();
664 }
665 #else /* !SMP */
666 /*
667 * Normal, non-SMP, 486+ invalidation functions.
668 * We inline these within pmap.c for speed.
669 */
670 PMAP_INLINE void
671 pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
672 {
673
674 if (pmap == kernel_pmap || pmap->pm_active)
675 invlpg(va);
676 }
677
678 PMAP_INLINE void
679 pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
680 {
681 vm_offset_t addr;
682
683 if (pmap == kernel_pmap || pmap->pm_active)
684 for (addr = sva; addr < eva; addr += PAGE_SIZE)
685 invlpg(addr);
686 }
687
688 PMAP_INLINE void
689 pmap_invalidate_all(pmap_t pmap)
690 {
691
692 if (pmap == kernel_pmap || pmap->pm_active)
693 invltlb();
694 }
695 #endif /* !SMP */
696
697 /*
698 * Are we current address space or kernel? N.B. We return FALSE when
699 * a pmap's page table is in use because a kernel thread is borrowing
700 * it. The borrowed page table can change spontaneously, making any
701 * dependence on its continued use subject to a race condition.
702 */
703 static __inline int
704 pmap_is_current(pmap_t pmap)
705 {
706
707 return (pmap == kernel_pmap ||
708 (pmap == vmspace_pmap(curthread->td_proc->p_vmspace) &&
709 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME)));
710 }
711
712 /*
713 * If the given pmap is not the current or kernel pmap, the returned pte must
714 * be released by passing it to pmap_pte_release().
715 */
716 pt_entry_t *
717 pmap_pte(pmap_t pmap, vm_offset_t va)
718 {
719 pd_entry_t newpf;
720 pd_entry_t *pde;
721
722 pde = pmap_pde(pmap, va);
723 if (*pde & PG_PS)
724 return (pde);
725 if (*pde != 0) {
726 /* are we current address space or kernel? */
727 if (pmap_is_current(pmap))
728 return (vtopte(va));
729 mtx_lock(&PMAP2mutex);
730 newpf = *pde & PG_FRAME;
731 if ((*PMAP2 & PG_FRAME) != newpf) {
732 *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M;
733 pmap_invalidate_page(kernel_pmap, (vm_offset_t)PADDR2);
734 }
735 return (PADDR2 + (i386_btop(va) & (NPTEPG - 1)));
736 }
737 return (0);
738 }
739
740 /*
741 * Releases a pte that was obtained from pmap_pte(). Be prepared for the pte
742 * being NULL.
743 */
744 static __inline void
745 pmap_pte_release(pt_entry_t *pte)
746 {
747
748 if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2)
749 mtx_unlock(&PMAP2mutex);
750 }
751
752 static __inline void
753 invlcaddr(void *caddr)
754 {
755
756 invlpg((u_int)caddr);
757 }
758
759 /*
760 * Super fast pmap_pte routine best used when scanning
761 * the pv lists. This eliminates many coarse-grained
762 * invltlb calls. Note that many of the pv list
763 * scans are across different pmaps. It is very wasteful
764 * to do an entire invltlb for checking a single mapping.
765 *
766 * If the given pmap is not the current pmap, vm_page_queue_mtx
767 * must be held and curthread pinned to a CPU.
768 */
769 static pt_entry_t *
770 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
771 {
772 pd_entry_t newpf;
773 pd_entry_t *pde;
774
775 pde = pmap_pde(pmap, va);
776 if (*pde & PG_PS)
777 return (pde);
778 if (*pde != 0) {
779 /* are we current address space or kernel? */
780 if (pmap_is_current(pmap))
781 return (vtopte(va));
782 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
783 KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
784 newpf = *pde & PG_FRAME;
785 if ((*PMAP1 & PG_FRAME) != newpf) {
786 *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M;
787 #ifdef SMP
788 PMAP1cpu = PCPU_GET(cpuid);
789 #endif
790 invlcaddr(PADDR1);
791 PMAP1changed++;
792 } else
793 #ifdef SMP
794 if (PMAP1cpu != PCPU_GET(cpuid)) {
795 PMAP1cpu = PCPU_GET(cpuid);
796 invlcaddr(PADDR1);
797 PMAP1changedcpu++;
798 } else
799 #endif
800 PMAP1unchanged++;
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_t pmap, vm_offset_t va)
814 {
815 vm_paddr_t rtval;
816 pt_entry_t *pte;
817 pd_entry_t pde;
818
819 rtval = 0;
820 PMAP_LOCK(pmap);
821 pde = pmap->pm_pdir[va >> PDRSHIFT];
822 if (pde != 0) {
823 if ((pde & PG_PS) != 0) {
824 rtval = (pde & ~PDRMASK) | (va & PDRMASK);
825 PMAP_UNLOCK(pmap);
826 return rtval;
827 }
828 pte = pmap_pte(pmap, va);
829 rtval = (*pte & PG_FRAME) | (va & PAGE_MASK);
830 pmap_pte_release(pte);
831 }
832 PMAP_UNLOCK(pmap);
833 return (rtval);
834 }
835
836 /*
837 * Routine: pmap_extract_and_hold
838 * Function:
839 * Atomically extract and hold the physical page
840 * with the given pmap and virtual address pair
841 * if that mapping permits the given protection.
842 */
843 vm_page_t
844 pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
845 {
846 pd_entry_t pde;
847 pt_entry_t pte;
848 vm_page_t m;
849
850 m = NULL;
851 vm_page_lock_queues();
852 PMAP_LOCK(pmap);
853 pde = *pmap_pde(pmap, va);
854 if (pde != 0) {
855 if (pde & PG_PS) {
856 if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) {
857 m = PHYS_TO_VM_PAGE((pde & ~PDRMASK) |
858 (va & PDRMASK));
859 vm_page_hold(m);
860 }
861 } else {
862 sched_pin();
863 pte = *pmap_pte_quick(pmap, va);
864 if (pte != 0 &&
865 ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) {
866 m = PHYS_TO_VM_PAGE(pte & PG_FRAME);
867 vm_page_hold(m);
868 }
869 sched_unpin();
870 }
871 }
872 vm_page_unlock_queues();
873 PMAP_UNLOCK(pmap);
874 return (m);
875 }
876
877 /***************************************************
878 * Low level mapping routines.....
879 ***************************************************/
880
881 /*
882 * Add a wired page to the kva.
883 * Note: not SMP coherent.
884 */
885 PMAP_INLINE void
886 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
887 {
888 pt_entry_t *pte;
889
890 pte = vtopte(va);
891 pte_store(pte, pa | PG_RW | PG_V | pgeflag);
892 }
893
894 /*
895 * Remove a page from the kernel pagetables.
896 * Note: not SMP coherent.
897 */
898 PMAP_INLINE void
899 pmap_kremove(vm_offset_t va)
900 {
901 pt_entry_t *pte;
902
903 pte = vtopte(va);
904 pte_clear(pte);
905 }
906
907 /*
908 * Used to map a range of physical addresses into kernel
909 * virtual address space.
910 *
911 * The value passed in '*virt' is a suggested virtual address for
912 * the mapping. Architectures which can support a direct-mapped
913 * physical to virtual region can return the appropriate address
914 * within that region, leaving '*virt' unchanged. Other
915 * architectures should map the pages starting at '*virt' and
916 * update '*virt' with the first usable address after the mapped
917 * region.
918 */
919 vm_offset_t
920 pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot)
921 {
922 vm_offset_t va, sva;
923
924 va = sva = *virt;
925 while (start < end) {
926 pmap_kenter(va, start);
927 va += PAGE_SIZE;
928 start += PAGE_SIZE;
929 }
930 pmap_invalidate_range(kernel_pmap, sva, va);
931 *virt = va;
932 return (sva);
933 }
934
935
936 /*
937 * Add a list of wired pages to the kva
938 * this routine is only used for temporary
939 * kernel mappings that do not need to have
940 * page modification or references recorded.
941 * Note that old mappings are simply written
942 * over. The page *must* be wired.
943 * Note: SMP coherent. Uses a ranged shootdown IPI.
944 */
945 void
946 pmap_qenter(vm_offset_t sva, vm_page_t *m, int count)
947 {
948 vm_offset_t va;
949
950 va = sva;
951 while (count-- > 0) {
952 pmap_kenter(va, VM_PAGE_TO_PHYS(*m));
953 va += PAGE_SIZE;
954 m++;
955 }
956 pmap_invalidate_range(kernel_pmap, sva, va);
957 }
958
959 /*
960 * This routine tears out page mappings from the
961 * kernel -- it is meant only for temporary mappings.
962 * Note: SMP coherent. Uses a ranged shootdown IPI.
963 */
964 void
965 pmap_qremove(vm_offset_t sva, int count)
966 {
967 vm_offset_t va;
968
969 va = sva;
970 while (count-- > 0) {
971 pmap_kremove(va);
972 va += PAGE_SIZE;
973 }
974 pmap_invalidate_range(kernel_pmap, sva, va);
975 }
976
977 /***************************************************
978 * Page table page management routines.....
979 ***************************************************/
980
981 /*
982 * This routine unholds page table pages, and if the hold count
983 * drops to zero, then it decrements the wire count.
984 */
985 static PMAP_INLINE int
986 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
987 {
988
989 --m->wire_count;
990 if (m->wire_count == 0)
991 return _pmap_unwire_pte_hold(pmap, m);
992 else
993 return 0;
994 }
995
996 static int
997 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
998 {
999 vm_offset_t pteva;
1000
1001 /*
1002 * unmap the page table page
1003 */
1004 pmap->pm_pdir[m->pindex] = 0;
1005 --pmap->pm_stats.resident_count;
1006
1007 /*
1008 * Do an invltlb to make the invalidated mapping
1009 * take effect immediately.
1010 */
1011 pteva = VM_MAXUSER_ADDRESS + i386_ptob(m->pindex);
1012 pmap_invalidate_page(pmap, pteva);
1013
1014 vm_page_free_zero(m);
1015 atomic_subtract_int(&cnt.v_wire_count, 1);
1016 return 1;
1017 }
1018
1019 /*
1020 * After removing a page table entry, this routine is used to
1021 * conditionally free the page, and manage the hold/wire counts.
1022 */
1023 static int
1024 pmap_unuse_pt(pmap_t pmap, vm_offset_t va)
1025 {
1026 pd_entry_t ptepde;
1027 vm_page_t mpte;
1028
1029 if (va >= VM_MAXUSER_ADDRESS)
1030 return 0;
1031 ptepde = *pmap_pde(pmap, va);
1032 mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME);
1033 return pmap_unwire_pte_hold(pmap, mpte);
1034 }
1035
1036 void
1037 pmap_pinit0(pmap)
1038 struct pmap *pmap;
1039 {
1040
1041 PMAP_LOCK_INIT(pmap);
1042 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + (vm_offset_t)IdlePTD);
1043 #ifdef PAE
1044 pmap->pm_pdpt = (pdpt_entry_t *)(KERNBASE + (vm_offset_t)IdlePDPT);
1045 #endif
1046 pmap->pm_active = 0;
1047 PCPU_SET(curpmap, pmap);
1048 TAILQ_INIT(&pmap->pm_pvlist);
1049 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1050 mtx_lock_spin(&allpmaps_lock);
1051 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list);
1052 mtx_unlock_spin(&allpmaps_lock);
1053 }
1054
1055 /*
1056 * Initialize a preallocated and zeroed pmap structure,
1057 * such as one in a vmspace structure.
1058 */
1059 void
1060 pmap_pinit(pmap)
1061 register struct pmap *pmap;
1062 {
1063 vm_page_t m, ptdpg[NPGPTD];
1064 vm_paddr_t pa;
1065 static int color;
1066 int i;
1067
1068 PMAP_LOCK_INIT(pmap);
1069
1070 /*
1071 * No need to allocate page table space yet but we do need a valid
1072 * page directory table.
1073 */
1074 if (pmap->pm_pdir == NULL) {
1075 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_nofault(kernel_map,
1076 NBPTD);
1077 #ifdef PAE
1078 pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO);
1079 KASSERT(((vm_offset_t)pmap->pm_pdpt &
1080 ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0,
1081 ("pmap_pinit: pdpt misaligned"));
1082 KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30),
1083 ("pmap_pinit: pdpt above 4g"));
1084 #endif
1085 }
1086
1087 /*
1088 * allocate the page directory page(s)
1089 */
1090 for (i = 0; i < NPGPTD;) {
1091 m = vm_page_alloc(NULL, color++,
1092 VM_ALLOC_NORMAL | VM_ALLOC_NOOBJ | VM_ALLOC_WIRED |
1093 VM_ALLOC_ZERO);
1094 if (m == NULL)
1095 VM_WAIT;
1096 else {
1097 ptdpg[i++] = m;
1098 }
1099 }
1100
1101 pmap_qenter((vm_offset_t)pmap->pm_pdir, ptdpg, NPGPTD);
1102
1103 for (i = 0; i < NPGPTD; i++) {
1104 if ((ptdpg[i]->flags & PG_ZERO) == 0)
1105 bzero(pmap->pm_pdir + (i * NPDEPG), PAGE_SIZE);
1106 }
1107
1108 mtx_lock_spin(&allpmaps_lock);
1109 LIST_INSERT_HEAD(&allpmaps, pmap, pm_list);
1110 mtx_unlock_spin(&allpmaps_lock);
1111 /* Wire in kernel global address entries. */
1112 /* XXX copies current process, does not fill in MPPTDI */
1113 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * sizeof(pd_entry_t));
1114 #ifdef SMP
1115 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1116 #endif
1117
1118 /* install self-referential address mapping entry(s) */
1119 for (i = 0; i < NPGPTD; i++) {
1120 pa = VM_PAGE_TO_PHYS(ptdpg[i]);
1121 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M;
1122 #ifdef PAE
1123 pmap->pm_pdpt[i] = pa | PG_V;
1124 #endif
1125 }
1126
1127 pmap->pm_active = 0;
1128 TAILQ_INIT(&pmap->pm_pvlist);
1129 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1130 }
1131
1132 /*
1133 * this routine is called if the page table page is not
1134 * mapped correctly.
1135 */
1136 static vm_page_t
1137 _pmap_allocpte(pmap_t pmap, unsigned ptepindex, int flags)
1138 {
1139 vm_paddr_t ptepa;
1140 vm_page_t m;
1141
1142 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT ||
1143 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK,
1144 ("_pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK"));
1145
1146 /*
1147 * Allocate a page table page.
1148 */
1149 if ((m = vm_page_alloc(NULL, ptepindex, VM_ALLOC_NOOBJ |
1150 VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) {
1151 if (flags & M_WAITOK) {
1152 PMAP_UNLOCK(pmap);
1153 vm_page_unlock_queues();
1154 VM_WAIT;
1155 vm_page_lock_queues();
1156 PMAP_LOCK(pmap);
1157 }
1158
1159 /*
1160 * Indicate the need to retry. While waiting, the page table
1161 * page may have been allocated.
1162 */
1163 return (NULL);
1164 }
1165 if ((m->flags & PG_ZERO) == 0)
1166 pmap_zero_page(m);
1167
1168 /*
1169 * Map the pagetable page into the process address space, if
1170 * it isn't already there.
1171 */
1172
1173 pmap->pm_stats.resident_count++;
1174
1175 ptepa = VM_PAGE_TO_PHYS(m);
1176 pmap->pm_pdir[ptepindex] =
1177 (pd_entry_t) (ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M);
1178
1179 return m;
1180 }
1181
1182 static vm_page_t
1183 pmap_allocpte(pmap_t pmap, vm_offset_t va, int flags)
1184 {
1185 unsigned ptepindex;
1186 pd_entry_t ptepa;
1187 vm_page_t m;
1188
1189 KASSERT((flags & (M_NOWAIT | M_WAITOK)) == M_NOWAIT ||
1190 (flags & (M_NOWAIT | M_WAITOK)) == M_WAITOK,
1191 ("pmap_allocpte: flags is neither M_NOWAIT nor M_WAITOK"));
1192
1193 /*
1194 * Calculate pagetable page index
1195 */
1196 ptepindex = va >> PDRSHIFT;
1197 retry:
1198 /*
1199 * Get the page directory entry
1200 */
1201 ptepa = pmap->pm_pdir[ptepindex];
1202
1203 /*
1204 * This supports switching from a 4MB page to a
1205 * normal 4K page.
1206 */
1207 if (ptepa & PG_PS) {
1208 pmap->pm_pdir[ptepindex] = 0;
1209 ptepa = 0;
1210 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1211 pmap_invalidate_all(kernel_pmap);
1212 }
1213
1214 /*
1215 * If the page table page is mapped, we just increment the
1216 * hold count, and activate it.
1217 */
1218 if (ptepa) {
1219 m = PHYS_TO_VM_PAGE(ptepa);
1220 m->wire_count++;
1221 } else {
1222 /*
1223 * Here if the pte page isn't mapped, or if it has
1224 * been deallocated.
1225 */
1226 m = _pmap_allocpte(pmap, ptepindex, flags);
1227 if (m == NULL && (flags & M_WAITOK))
1228 goto retry;
1229 }
1230 return (m);
1231 }
1232
1233
1234 /***************************************************
1235 * Pmap allocation/deallocation routines.
1236 ***************************************************/
1237
1238 #ifdef SMP
1239 /*
1240 * Deal with a SMP shootdown of other users of the pmap that we are
1241 * trying to dispose of. This can be a bit hairy.
1242 */
1243 static u_int *lazymask;
1244 static u_int lazyptd;
1245 static volatile u_int lazywait;
1246
1247 void pmap_lazyfix_action(void);
1248
1249 void
1250 pmap_lazyfix_action(void)
1251 {
1252 u_int mymask = PCPU_GET(cpumask);
1253
1254 if (rcr3() == lazyptd)
1255 load_cr3(PCPU_GET(curpcb)->pcb_cr3);
1256 atomic_clear_int(lazymask, mymask);
1257 atomic_store_rel_int(&lazywait, 1);
1258 }
1259
1260 static void
1261 pmap_lazyfix_self(u_int mymask)
1262 {
1263
1264 if (rcr3() == lazyptd)
1265 load_cr3(PCPU_GET(curpcb)->pcb_cr3);
1266 atomic_clear_int(lazymask, mymask);
1267 }
1268
1269
1270 static void
1271 pmap_lazyfix(pmap_t pmap)
1272 {
1273 u_int mymask;
1274 u_int mask;
1275 register u_int spins;
1276
1277 while ((mask = pmap->pm_active) != 0) {
1278 spins = 50000000;
1279 mask = mask & -mask; /* Find least significant set bit */
1280 mtx_lock_spin(&smp_ipi_mtx);
1281 #ifdef PAE
1282 lazyptd = vtophys(pmap->pm_pdpt);
1283 #else
1284 lazyptd = vtophys(pmap->pm_pdir);
1285 #endif
1286 mymask = PCPU_GET(cpumask);
1287 if (mask == mymask) {
1288 lazymask = &pmap->pm_active;
1289 pmap_lazyfix_self(mymask);
1290 } else {
1291 atomic_store_rel_int((u_int *)&lazymask,
1292 (u_int)&pmap->pm_active);
1293 atomic_store_rel_int(&lazywait, 0);
1294 ipi_selected(mask, IPI_LAZYPMAP);
1295 while (lazywait == 0) {
1296 ia32_pause();
1297 if (--spins == 0)
1298 break;
1299 }
1300 }
1301 mtx_unlock_spin(&smp_ipi_mtx);
1302 if (spins == 0)
1303 printf("pmap_lazyfix: spun for 50000000\n");
1304 }
1305 }
1306
1307 #else /* SMP */
1308
1309 /*
1310 * Cleaning up on uniprocessor is easy. For various reasons, we're
1311 * unlikely to have to even execute this code, including the fact
1312 * that the cleanup is deferred until the parent does a wait(2), which
1313 * means that another userland process has run.
1314 */
1315 static void
1316 pmap_lazyfix(pmap_t pmap)
1317 {
1318 u_int cr3;
1319
1320 cr3 = vtophys(pmap->pm_pdir);
1321 if (cr3 == rcr3()) {
1322 load_cr3(PCPU_GET(curpcb)->pcb_cr3);
1323 pmap->pm_active &= ~(PCPU_GET(cpumask));
1324 }
1325 }
1326 #endif /* SMP */
1327
1328 /*
1329 * Release any resources held by the given physical map.
1330 * Called when a pmap initialized by pmap_pinit is being released.
1331 * Should only be called if the map contains no valid mappings.
1332 */
1333 void
1334 pmap_release(pmap_t pmap)
1335 {
1336 vm_page_t m, ptdpg[NPGPTD];
1337 int i;
1338
1339 KASSERT(pmap->pm_stats.resident_count == 0,
1340 ("pmap_release: pmap resident count %ld != 0",
1341 pmap->pm_stats.resident_count));
1342
1343 pmap_lazyfix(pmap);
1344 mtx_lock_spin(&allpmaps_lock);
1345 LIST_REMOVE(pmap, pm_list);
1346 mtx_unlock_spin(&allpmaps_lock);
1347
1348 for (i = 0; i < NPGPTD; i++)
1349 ptdpg[i] = PHYS_TO_VM_PAGE(pmap->pm_pdir[PTDPTDI + i]);
1350
1351 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) *
1352 sizeof(*pmap->pm_pdir));
1353 #ifdef SMP
1354 pmap->pm_pdir[MPPTDI] = 0;
1355 #endif
1356
1357 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD);
1358
1359 vm_page_lock_queues();
1360 for (i = 0; i < NPGPTD; i++) {
1361 m = ptdpg[i];
1362 #ifdef PAE
1363 KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME),
1364 ("pmap_release: got wrong ptd page"));
1365 #endif
1366 m->wire_count--;
1367 atomic_subtract_int(&cnt.v_wire_count, 1);
1368 vm_page_free_zero(m);
1369 }
1370 vm_page_unlock_queues();
1371 PMAP_LOCK_DESTROY(pmap);
1372 }
1373
1374 static int
1375 kvm_size(SYSCTL_HANDLER_ARGS)
1376 {
1377 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE;
1378
1379 return sysctl_handle_long(oidp, &ksize, 0, req);
1380 }
1381 SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD,
1382 0, 0, kvm_size, "IU", "Size of KVM");
1383
1384 static int
1385 kvm_free(SYSCTL_HANDLER_ARGS)
1386 {
1387 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end;
1388
1389 return sysctl_handle_long(oidp, &kfree, 0, req);
1390 }
1391 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD,
1392 0, 0, kvm_free, "IU", "Amount of KVM free");
1393
1394 /*
1395 * grow the number of kernel page table entries, if needed
1396 */
1397 void
1398 pmap_growkernel(vm_offset_t addr)
1399 {
1400 struct pmap *pmap;
1401 vm_paddr_t ptppaddr;
1402 vm_page_t nkpg;
1403 pd_entry_t newpdir;
1404 pt_entry_t *pde;
1405
1406 mtx_assert(&kernel_map->system_mtx, MA_OWNED);
1407 if (kernel_vm_end == 0) {
1408 kernel_vm_end = KERNBASE;
1409 nkpt = 0;
1410 while (pdir_pde(PTD, kernel_vm_end)) {
1411 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1412 nkpt++;
1413 if (kernel_vm_end - 1 >= kernel_map->max_offset) {
1414 kernel_vm_end = kernel_map->max_offset;
1415 break;
1416 }
1417 }
1418 }
1419 addr = roundup2(addr, PAGE_SIZE * NPTEPG);
1420 if (addr - 1 >= kernel_map->max_offset)
1421 addr = kernel_map->max_offset;
1422 while (kernel_vm_end < addr) {
1423 if (pdir_pde(PTD, kernel_vm_end)) {
1424 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1425 if (kernel_vm_end - 1 >= kernel_map->max_offset) {
1426 kernel_vm_end = kernel_map->max_offset;
1427 break;
1428 }
1429 continue;
1430 }
1431
1432 /*
1433 * This index is bogus, but out of the way
1434 */
1435 nkpg = vm_page_alloc(NULL, nkpt,
1436 VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM | VM_ALLOC_WIRED);
1437 if (!nkpg)
1438 panic("pmap_growkernel: no memory to grow kernel");
1439
1440 nkpt++;
1441
1442 pmap_zero_page(nkpg);
1443 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1444 newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M);
1445 pdir_pde(PTD, kernel_vm_end) = newpdir;
1446
1447 mtx_lock_spin(&allpmaps_lock);
1448 LIST_FOREACH(pmap, &allpmaps, pm_list) {
1449 pde = pmap_pde(pmap, kernel_vm_end);
1450 pde_store(pde, newpdir);
1451 }
1452 mtx_unlock_spin(&allpmaps_lock);
1453 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1454 if (kernel_vm_end - 1 >= kernel_map->max_offset) {
1455 kernel_vm_end = kernel_map->max_offset;
1456 break;
1457 }
1458 }
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_count > pv_entry_high_water) &&
1487 (pmap_pagedaemon_waken == 0)) {
1488 pmap_pagedaemon_waken = 1;
1489 wakeup (&vm_pages_needed);
1490 }
1491 return uma_zalloc(pvzone, M_NOWAIT);
1492 }
1493
1494
1495 static void
1496 pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va)
1497 {
1498 pv_entry_t pv;
1499
1500 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1501 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1502 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1503 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1504 if (pmap == pv->pv_pmap && va == pv->pv_va)
1505 break;
1506 }
1507 } else {
1508 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1509 if (va == pv->pv_va)
1510 break;
1511 }
1512 }
1513 KASSERT(pv != NULL, ("pmap_remove_entry: pv not found"));
1514 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1515 m->md.pv_list_count--;
1516 if (TAILQ_EMPTY(&m->md.pv_list))
1517 vm_page_flag_clear(m, PG_WRITEABLE);
1518 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1519 free_pv_entry(pv);
1520 }
1521
1522 /*
1523 * Create a pv entry for page at pa for
1524 * (pmap, va).
1525 */
1526 static void
1527 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m)
1528 {
1529 pv_entry_t pv;
1530
1531 pv = get_pv_entry();
1532 if (pv == NULL)
1533 panic("no pv entries: increase vm.pmap.shpgperproc");
1534 pv->pv_va = va;
1535 pv->pv_pmap = pmap;
1536
1537 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1538 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1539 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1540 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1541 m->md.pv_list_count++;
1542 }
1543
1544 /*
1545 * pmap_remove_pte: do the things to unmap a page in a process
1546 */
1547 static int
1548 pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va)
1549 {
1550 pt_entry_t oldpte;
1551 vm_page_t m;
1552
1553 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1554 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1555 oldpte = pte_load_clear(ptq);
1556 if (oldpte & PG_W)
1557 pmap->pm_stats.wired_count -= 1;
1558 /*
1559 * Machines that don't support invlpg, also don't support
1560 * PG_G.
1561 */
1562 if (oldpte & PG_G)
1563 pmap_invalidate_page(kernel_pmap, va);
1564 pmap->pm_stats.resident_count -= 1;
1565 if (oldpte & PG_MANAGED) {
1566 m = PHYS_TO_VM_PAGE(oldpte);
1567 if (oldpte & PG_M) {
1568 #if defined(PMAP_DIAGNOSTIC)
1569 if (pmap_nw_modified((pt_entry_t) oldpte)) {
1570 printf(
1571 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1572 va, oldpte);
1573 }
1574 #endif
1575 if (pmap_track_modified(va))
1576 vm_page_dirty(m);
1577 }
1578 if (oldpte & PG_A)
1579 vm_page_flag_set(m, PG_REFERENCED);
1580 pmap_remove_entry(pmap, m, va);
1581 }
1582 return (pmap_unuse_pt(pmap, va));
1583 }
1584
1585 /*
1586 * Remove a single page from a process address space
1587 */
1588 static void
1589 pmap_remove_page(pmap_t pmap, vm_offset_t va)
1590 {
1591 pt_entry_t *pte;
1592
1593 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1594 KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
1595 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1596 if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0)
1597 return;
1598 pmap_remove_pte(pmap, pte, va);
1599 pmap_invalidate_page(pmap, va);
1600 }
1601
1602 /*
1603 * Remove the given range of addresses from the specified map.
1604 *
1605 * It is assumed that the start and end are properly
1606 * rounded to the page size.
1607 */
1608 void
1609 pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1610 {
1611 vm_offset_t pdnxt;
1612 pd_entry_t ptpaddr;
1613 pt_entry_t *pte;
1614 int anyvalid;
1615
1616 /*
1617 * Perform an unsynchronized read. This is, however, safe.
1618 */
1619 if (pmap->pm_stats.resident_count == 0)
1620 return;
1621
1622 anyvalid = 0;
1623
1624 vm_page_lock_queues();
1625 sched_pin();
1626 PMAP_LOCK(pmap);
1627
1628 /*
1629 * special handling of removing one page. a very
1630 * common operation and easy to short circuit some
1631 * code.
1632 */
1633 if ((sva + PAGE_SIZE == eva) &&
1634 ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) {
1635 pmap_remove_page(pmap, sva);
1636 goto out;
1637 }
1638
1639 for (; sva < eva; sva = pdnxt) {
1640 unsigned pdirindex;
1641
1642 /*
1643 * Calculate index for next page table.
1644 */
1645 pdnxt = (sva + NBPDR) & ~PDRMASK;
1646 if (pmap->pm_stats.resident_count == 0)
1647 break;
1648
1649 pdirindex = sva >> PDRSHIFT;
1650 ptpaddr = pmap->pm_pdir[pdirindex];
1651
1652 /*
1653 * Weed out invalid mappings. Note: we assume that the page
1654 * directory table is always allocated, and in kernel virtual.
1655 */
1656 if (ptpaddr == 0)
1657 continue;
1658
1659 /*
1660 * Check for large page.
1661 */
1662 if ((ptpaddr & PG_PS) != 0) {
1663 pmap->pm_pdir[pdirindex] = 0;
1664 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1665 anyvalid = 1;
1666 continue;
1667 }
1668
1669 /*
1670 * Limit our scan to either the end of the va represented
1671 * by the current page table page, or to the end of the
1672 * range being removed.
1673 */
1674 if (pdnxt > eva)
1675 pdnxt = eva;
1676
1677 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++,
1678 sva += PAGE_SIZE) {
1679 if (*pte == 0)
1680 continue;
1681 anyvalid = 1;
1682 if (pmap_remove_pte(pmap, pte, sva))
1683 break;
1684 }
1685 }
1686 out:
1687 sched_unpin();
1688 vm_page_unlock_queues();
1689 if (anyvalid)
1690 pmap_invalidate_all(pmap);
1691 PMAP_UNLOCK(pmap);
1692 }
1693
1694 /*
1695 * Routine: pmap_remove_all
1696 * Function:
1697 * Removes this physical page from
1698 * all physical maps in which it resides.
1699 * Reflects back modify bits to the pager.
1700 *
1701 * Notes:
1702 * Original versions of this routine were very
1703 * inefficient because they iteratively called
1704 * pmap_remove (slow...)
1705 */
1706
1707 void
1708 pmap_remove_all(vm_page_t m)
1709 {
1710 register pv_entry_t pv;
1711 pt_entry_t *pte, tpte;
1712
1713 #if defined(PMAP_DIAGNOSTIC)
1714 /*
1715 * XXX This makes pmap_remove_all() illegal for non-managed pages!
1716 */
1717 if (m->flags & PG_FICTITIOUS) {
1718 panic("pmap_remove_all: illegal for unmanaged page, va: 0x%x",
1719 VM_PAGE_TO_PHYS(m));
1720 }
1721 #endif
1722 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1723 sched_pin();
1724 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1725 PMAP_LOCK(pv->pv_pmap);
1726 pv->pv_pmap->pm_stats.resident_count--;
1727 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1728 tpte = pte_load_clear(pte);
1729 if (tpte & PG_W)
1730 pv->pv_pmap->pm_stats.wired_count--;
1731 if (tpte & PG_A)
1732 vm_page_flag_set(m, PG_REFERENCED);
1733
1734 /*
1735 * Update the vm_page_t clean and reference bits.
1736 */
1737 if (tpte & PG_M) {
1738 #if defined(PMAP_DIAGNOSTIC)
1739 if (pmap_nw_modified((pt_entry_t) tpte)) {
1740 printf(
1741 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1742 pv->pv_va, tpte);
1743 }
1744 #endif
1745 if (pmap_track_modified(pv->pv_va))
1746 vm_page_dirty(m);
1747 }
1748 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
1749 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1750 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1751 m->md.pv_list_count--;
1752 pmap_unuse_pt(pv->pv_pmap, pv->pv_va);
1753 PMAP_UNLOCK(pv->pv_pmap);
1754 free_pv_entry(pv);
1755 }
1756 vm_page_flag_clear(m, PG_WRITEABLE);
1757 sched_unpin();
1758 }
1759
1760 /*
1761 * Set the physical protection on the
1762 * specified range of this map as requested.
1763 */
1764 void
1765 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1766 {
1767 vm_offset_t pdnxt;
1768 pd_entry_t ptpaddr;
1769 pt_entry_t *pte;
1770 int anychanged;
1771
1772 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1773 pmap_remove(pmap, sva, eva);
1774 return;
1775 }
1776
1777 if (prot & VM_PROT_WRITE)
1778 return;
1779
1780 anychanged = 0;
1781
1782 vm_page_lock_queues();
1783 sched_pin();
1784 PMAP_LOCK(pmap);
1785 for (; sva < eva; sva = pdnxt) {
1786 unsigned obits, pbits, pdirindex;
1787
1788 pdnxt = (sva + NBPDR) & ~PDRMASK;
1789
1790 pdirindex = sva >> PDRSHIFT;
1791 ptpaddr = pmap->pm_pdir[pdirindex];
1792
1793 /*
1794 * Weed out invalid mappings. Note: we assume that the page
1795 * directory table is always allocated, and in kernel virtual.
1796 */
1797 if (ptpaddr == 0)
1798 continue;
1799
1800 /*
1801 * Check for large page.
1802 */
1803 if ((ptpaddr & PG_PS) != 0) {
1804 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1805 anychanged = 1;
1806 continue;
1807 }
1808
1809 if (pdnxt > eva)
1810 pdnxt = eva;
1811
1812 for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++,
1813 sva += PAGE_SIZE) {
1814 vm_page_t m;
1815
1816 retry:
1817 /*
1818 * Regardless of whether a pte is 32 or 64 bits in
1819 * size, PG_RW, PG_A, and PG_M are among the least
1820 * significant 32 bits.
1821 */
1822 obits = pbits = *(u_int *)pte;
1823 if (pbits & PG_MANAGED) {
1824 m = NULL;
1825 if (pbits & PG_A) {
1826 m = PHYS_TO_VM_PAGE(*pte);
1827 vm_page_flag_set(m, PG_REFERENCED);
1828 pbits &= ~PG_A;
1829 }
1830 if ((pbits & PG_M) != 0 &&
1831 pmap_track_modified(sva)) {
1832 if (m == NULL)
1833 m = PHYS_TO_VM_PAGE(*pte);
1834 vm_page_dirty(m);
1835 }
1836 }
1837
1838 pbits &= ~(PG_RW | PG_M);
1839
1840 if (pbits != obits) {
1841 if (!atomic_cmpset_int((u_int *)pte, obits,
1842 pbits))
1843 goto retry;
1844 if (obits & PG_G)
1845 pmap_invalidate_page(pmap, sva);
1846 else
1847 anychanged = 1;
1848 }
1849 }
1850 }
1851 sched_unpin();
1852 vm_page_unlock_queues();
1853 if (anychanged)
1854 pmap_invalidate_all(pmap);
1855 PMAP_UNLOCK(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, om;
1879 boolean_t invlva;
1880
1881 va &= PG_FRAME;
1882 #ifdef PMAP_DIAGNOSTIC
1883 if (va > VM_MAX_KERNEL_ADDRESS)
1884 panic("pmap_enter: toobig");
1885 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1886 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1887 #endif
1888
1889 mpte = NULL;
1890
1891 vm_page_lock_queues();
1892 PMAP_LOCK(pmap);
1893 sched_pin();
1894
1895 /*
1896 * In the case that a page table page is not
1897 * resident, we are creating it here.
1898 */
1899 if (va < VM_MAXUSER_ADDRESS) {
1900 mpte = pmap_allocpte(pmap, va, M_WAITOK);
1901 }
1902 #if 0 && defined(PMAP_DIAGNOSTIC)
1903 else {
1904 pd_entry_t *pdeaddr = pmap_pde(pmap, va);
1905 origpte = *pdeaddr;
1906 if ((origpte & PG_V) == 0) {
1907 panic("pmap_enter: invalid kernel page table page, pdir=%p, pde=%p, va=%p\n",
1908 pmap->pm_pdir[PTDPTDI], origpte, va);
1909 }
1910 }
1911 #endif
1912
1913 pte = pmap_pte_quick(pmap, va);
1914
1915 /*
1916 * Page Directory table entry not valid, we need a new PT page
1917 */
1918 if (pte == NULL) {
1919 panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x\n",
1920 (uintmax_t)pmap->pm_pdir[PTDPTDI], va);
1921 }
1922
1923 pa = VM_PAGE_TO_PHYS(m);
1924 om = NULL;
1925 origpte = *pte;
1926 opa = origpte & PG_FRAME;
1927
1928 if (origpte & PG_PS) {
1929 /*
1930 * Yes, I know this will truncate upper address bits for PAE,
1931 * but I'm actually more interested in the lower bits
1932 */
1933 printf("pmap_enter: va %p, pte %p, origpte %p\n",
1934 (void *)va, (void *)pte, (void *)(uintptr_t)origpte);
1935 panic("pmap_enter: attempted pmap_enter on 4MB page");
1936 }
1937
1938 /*
1939 * Mapping has not changed, must be protection or wiring change.
1940 */
1941 if (origpte && (opa == pa)) {
1942 /*
1943 * Wiring change, just update stats. We don't worry about
1944 * wiring PT pages as they remain resident as long as there
1945 * are valid mappings in them. Hence, if a user page is wired,
1946 * the PT page will be also.
1947 */
1948 if (wired && ((origpte & PG_W) == 0))
1949 pmap->pm_stats.wired_count++;
1950 else if (!wired && (origpte & PG_W))
1951 pmap->pm_stats.wired_count--;
1952
1953 /*
1954 * Remove extra pte reference
1955 */
1956 if (mpte)
1957 mpte->wire_count--;
1958
1959 /*
1960 * We might be turning off write access to the page,
1961 * so we go ahead and sense modify status.
1962 */
1963 if (origpte & PG_MANAGED) {
1964 om = m;
1965 pa |= PG_MANAGED;
1966 }
1967 goto validate;
1968 }
1969 /*
1970 * Mapping has changed, invalidate old range and fall through to
1971 * handle validating new mapping.
1972 */
1973 if (opa) {
1974 if (origpte & PG_W)
1975 pmap->pm_stats.wired_count--;
1976 if (origpte & PG_MANAGED) {
1977 om = PHYS_TO_VM_PAGE(opa);
1978 pmap_remove_entry(pmap, om, va);
1979 }
1980 if (mpte != NULL) {
1981 mpte->wire_count--;
1982 KASSERT(mpte->wire_count > 0,
1983 ("pmap_enter: missing reference to page table page,"
1984 " va: 0x%x", va));
1985 }
1986 } else
1987 pmap->pm_stats.resident_count++;
1988
1989 /*
1990 * Enter on the PV list if part of our managed memory.
1991 */
1992 if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0) {
1993 pmap_insert_entry(pmap, va, m);
1994 pa |= PG_MANAGED;
1995 }
1996
1997 /*
1998 * Increment counters
1999 */
2000 if (wired)
2001 pmap->pm_stats.wired_count++;
2002
2003 validate:
2004 /*
2005 * Now validate mapping with desired protection/wiring.
2006 */
2007 newpte = (pt_entry_t)(pa | PG_V);
2008 if ((prot & VM_PROT_WRITE) != 0)
2009 newpte |= PG_RW;
2010 if (wired)
2011 newpte |= PG_W;
2012 if (va < VM_MAXUSER_ADDRESS)
2013 newpte |= PG_U;
2014 if (pmap == kernel_pmap)
2015 newpte |= pgeflag;
2016
2017 /*
2018 * if the mapping or permission bits are different, we need
2019 * to update the pte.
2020 */
2021 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2022 if (origpte & PG_V) {
2023 invlva = FALSE;
2024 origpte = pte_load_store(pte, newpte | PG_A);
2025 if (origpte & PG_A) {
2026 if (origpte & PG_MANAGED)
2027 vm_page_flag_set(om, PG_REFERENCED);
2028 if (opa != VM_PAGE_TO_PHYS(m))
2029 invlva = TRUE;
2030 }
2031 if (origpte & PG_M) {
2032 KASSERT((origpte & PG_RW),
2033 ("pmap_enter: modified page not writable:"
2034 " va: 0x%x, pte: 0x%x", va, origpte));
2035 if ((origpte & PG_MANAGED) &&
2036 pmap_track_modified(va))
2037 vm_page_dirty(om);
2038 if ((prot & VM_PROT_WRITE) == 0)
2039 invlva = TRUE;
2040 }
2041 if (invlva)
2042 pmap_invalidate_page(pmap, va);
2043 } else
2044 pte_store(pte, newpte | PG_A);
2045 }
2046 sched_unpin();
2047 vm_page_unlock_queues();
2048 PMAP_UNLOCK(pmap);
2049 }
2050
2051 /*
2052 * this code makes some *MAJOR* assumptions:
2053 * 1. Current pmap & pmap exists.
2054 * 2. Not wired.
2055 * 3. Read access.
2056 * 4. No page table pages.
2057 * but is *MUCH* faster than pmap_enter...
2058 */
2059
2060 vm_page_t
2061 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
2062 vm_page_t mpte)
2063 {
2064 pt_entry_t *pte;
2065 vm_paddr_t pa;
2066
2067 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2068 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
2069 PMAP_LOCK(pmap);
2070
2071 /*
2072 * In the case that a page table page is not
2073 * resident, we are creating it here.
2074 */
2075 if (va < VM_MAXUSER_ADDRESS) {
2076 unsigned ptepindex;
2077 pd_entry_t ptepa;
2078
2079 /*
2080 * Calculate pagetable page index
2081 */
2082 ptepindex = va >> PDRSHIFT;
2083 if (mpte && (mpte->pindex == ptepindex)) {
2084 mpte->wire_count++;
2085 } else {
2086 retry:
2087 /*
2088 * Get the page directory entry
2089 */
2090 ptepa = pmap->pm_pdir[ptepindex];
2091
2092 /*
2093 * If the page table page is mapped, we just increment
2094 * the hold count, and activate it.
2095 */
2096 if (ptepa) {
2097 if (ptepa & PG_PS)
2098 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2099 mpte = PHYS_TO_VM_PAGE(ptepa);
2100 mpte->wire_count++;
2101 } else {
2102 mpte = _pmap_allocpte(pmap, ptepindex,
2103 M_NOWAIT);
2104 if (mpte == NULL) {
2105 PMAP_UNLOCK(pmap);
2106 vm_page_busy(m);
2107 vm_page_unlock_queues();
2108 VM_OBJECT_UNLOCK(m->object);
2109 VM_WAIT;
2110 VM_OBJECT_LOCK(m->object);
2111 vm_page_lock_queues();
2112 vm_page_wakeup(m);
2113 PMAP_LOCK(pmap);
2114 goto retry;
2115 }
2116 }
2117 }
2118 } else {
2119 mpte = NULL;
2120 }
2121
2122 /*
2123 * This call to vtopte makes the assumption that we are
2124 * entering the page into the current pmap. In order to support
2125 * quick entry into any pmap, one would likely use pmap_pte_quick.
2126 * But that isn't as quick as vtopte.
2127 */
2128 pte = vtopte(va);
2129 if (*pte) {
2130 if (mpte != NULL) {
2131 pmap_unwire_pte_hold(pmap, mpte);
2132 mpte = NULL;
2133 }
2134 goto out;
2135 }
2136
2137 /*
2138 * Enter on the PV list if part of our managed memory. Note that we
2139 * raise IPL while manipulating pv_table since pmap_enter can be
2140 * called at interrupt time.
2141 */
2142 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
2143 pmap_insert_entry(pmap, va, m);
2144
2145 /*
2146 * Increment counters
2147 */
2148 pmap->pm_stats.resident_count++;
2149
2150 pa = VM_PAGE_TO_PHYS(m);
2151
2152 /*
2153 * Now validate mapping with RO protection
2154 */
2155 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2156 pte_store(pte, pa | PG_V | PG_U);
2157 else
2158 pte_store(pte, pa | PG_V | PG_U | PG_MANAGED);
2159 out:
2160 PMAP_UNLOCK(pmap);
2161 return mpte;
2162 }
2163
2164 /*
2165 * Make a temporary mapping for a physical address. This is only intended
2166 * to be used for panic dumps.
2167 */
2168 void *
2169 pmap_kenter_temporary(vm_paddr_t pa, int i)
2170 {
2171 vm_offset_t va;
2172
2173 va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE);
2174 pmap_kenter(va, pa);
2175 invlpg(va);
2176 return ((void *)crashdumpmap);
2177 }
2178
2179 /*
2180 * This code maps large physical mmap regions into the
2181 * processor address space. Note that some shortcuts
2182 * are taken, but the code works.
2183 */
2184 void
2185 pmap_object_init_pt(pmap_t pmap, vm_offset_t addr,
2186 vm_object_t object, vm_pindex_t pindex,
2187 vm_size_t size)
2188 {
2189 vm_page_t p;
2190
2191 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
2192 KASSERT(object->type == OBJT_DEVICE,
2193 ("pmap_object_init_pt: non-device object"));
2194 if (pseflag &&
2195 ((addr & (NBPDR - 1)) == 0) && ((size & (NBPDR - 1)) == 0)) {
2196 int i;
2197 vm_page_t m[1];
2198 unsigned int ptepindex;
2199 int npdes;
2200 pd_entry_t ptepa;
2201
2202 PMAP_LOCK(pmap);
2203 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)])
2204 goto out;
2205 PMAP_UNLOCK(pmap);
2206 retry:
2207 p = vm_page_lookup(object, pindex);
2208 if (p != NULL) {
2209 vm_page_lock_queues();
2210 if (vm_page_sleep_if_busy(p, FALSE, "init4p"))
2211 goto retry;
2212 } else {
2213 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
2214 if (p == NULL)
2215 return;
2216 m[0] = p;
2217
2218 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) {
2219 vm_page_lock_queues();
2220 vm_page_free(p);
2221 vm_page_unlock_queues();
2222 return;
2223 }
2224
2225 p = vm_page_lookup(object, pindex);
2226 vm_page_lock_queues();
2227 vm_page_wakeup(p);
2228 }
2229 vm_page_unlock_queues();
2230
2231 ptepa = VM_PAGE_TO_PHYS(p);
2232 if (ptepa & (NBPDR - 1))
2233 return;
2234
2235 p->valid = VM_PAGE_BITS_ALL;
2236
2237 PMAP_LOCK(pmap);
2238 pmap->pm_stats.resident_count += size >> PAGE_SHIFT;
2239 npdes = size >> PDRSHIFT;
2240 for(i = 0; i < npdes; i++) {
2241 pde_store(&pmap->pm_pdir[ptepindex],
2242 ptepa | PG_U | PG_RW | PG_V | PG_PS);
2243 ptepa += NBPDR;
2244 ptepindex += 1;
2245 }
2246 pmap_invalidate_all(pmap);
2247 out:
2248 PMAP_UNLOCK(pmap);
2249 }
2250 }
2251
2252 /*
2253 * Routine: pmap_change_wiring
2254 * Function: Change the wiring attribute for a map/virtual-address
2255 * pair.
2256 * In/out conditions:
2257 * The mapping must already exist in the pmap.
2258 */
2259 void
2260 pmap_change_wiring(pmap, va, wired)
2261 register pmap_t pmap;
2262 vm_offset_t va;
2263 boolean_t wired;
2264 {
2265 register pt_entry_t *pte;
2266
2267 PMAP_LOCK(pmap);
2268 pte = pmap_pte(pmap, va);
2269
2270 if (wired && !pmap_pte_w(pte))
2271 pmap->pm_stats.wired_count++;
2272 else if (!wired && pmap_pte_w(pte))
2273 pmap->pm_stats.wired_count--;
2274
2275 /*
2276 * Wiring is not a hardware characteristic so there is no need to
2277 * invalidate TLB.
2278 */
2279 pmap_pte_set_w(pte, wired);
2280 pmap_pte_release(pte);
2281 PMAP_UNLOCK(pmap);
2282 }
2283
2284
2285
2286 /*
2287 * Copy the range specified by src_addr/len
2288 * from the source map to the range dst_addr/len
2289 * in the destination map.
2290 *
2291 * This routine is only advisory and need not do anything.
2292 */
2293
2294 void
2295 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len,
2296 vm_offset_t src_addr)
2297 {
2298 vm_offset_t addr;
2299 vm_offset_t end_addr = src_addr + len;
2300 vm_offset_t pdnxt;
2301 vm_page_t m;
2302
2303 if (dst_addr != src_addr)
2304 return;
2305
2306 if (!pmap_is_current(src_pmap))
2307 return;
2308
2309 vm_page_lock_queues();
2310 if (dst_pmap < src_pmap) {
2311 PMAP_LOCK(dst_pmap);
2312 PMAP_LOCK(src_pmap);
2313 } else {
2314 PMAP_LOCK(src_pmap);
2315 PMAP_LOCK(dst_pmap);
2316 }
2317 sched_pin();
2318 for (addr = src_addr; addr < end_addr; addr = pdnxt) {
2319 pt_entry_t *src_pte, *dst_pte;
2320 vm_page_t dstmpte, srcmpte;
2321 pd_entry_t srcptepaddr;
2322 unsigned ptepindex;
2323
2324 if (addr >= UPT_MIN_ADDRESS)
2325 panic("pmap_copy: invalid to pmap_copy page tables");
2326
2327 /*
2328 * Don't let optional prefaulting of pages make us go
2329 * way below the low water mark of free pages or way
2330 * above high water mark of used pv entries.
2331 */
2332 if (cnt.v_free_count < cnt.v_free_reserved ||
2333 pv_entry_count > pv_entry_high_water)
2334 break;
2335
2336 pdnxt = (addr + NBPDR) & ~PDRMASK;
2337 ptepindex = addr >> PDRSHIFT;
2338
2339 srcptepaddr = src_pmap->pm_pdir[ptepindex];
2340 if (srcptepaddr == 0)
2341 continue;
2342
2343 if (srcptepaddr & PG_PS) {
2344 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2345 dst_pmap->pm_pdir[ptepindex] = srcptepaddr &
2346 ~PG_W;
2347 dst_pmap->pm_stats.resident_count +=
2348 NBPDR / PAGE_SIZE;
2349 }
2350 continue;
2351 }
2352
2353 srcmpte = PHYS_TO_VM_PAGE(srcptepaddr);
2354 if (srcmpte->wire_count == 0)
2355 panic("pmap_copy: source page table page is unused");
2356
2357 if (pdnxt > end_addr)
2358 pdnxt = end_addr;
2359
2360 src_pte = vtopte(addr);
2361 while (addr < pdnxt) {
2362 pt_entry_t ptetemp;
2363 ptetemp = *src_pte;
2364 /*
2365 * we only virtual copy managed pages
2366 */
2367 if ((ptetemp & PG_MANAGED) != 0) {
2368 /*
2369 * We have to check after allocpte for the
2370 * pte still being around... allocpte can
2371 * block.
2372 */
2373 dstmpte = pmap_allocpte(dst_pmap, addr,
2374 M_NOWAIT);
2375 if (dstmpte == NULL)
2376 break;
2377 dst_pte = pmap_pte_quick(dst_pmap, addr);
2378 if (*dst_pte == 0) {
2379 /*
2380 * Clear the wired, modified, and
2381 * accessed (referenced) bits
2382 * during the copy.
2383 */
2384 m = PHYS_TO_VM_PAGE(ptetemp);
2385 *dst_pte = ptetemp & ~(PG_W | PG_M |
2386 PG_A);
2387 dst_pmap->pm_stats.resident_count++;
2388 pmap_insert_entry(dst_pmap, addr, m);
2389 } else
2390 pmap_unwire_pte_hold(dst_pmap, dstmpte);
2391 if (dstmpte->wire_count >= srcmpte->wire_count)
2392 break;
2393 }
2394 addr += PAGE_SIZE;
2395 src_pte++;
2396 }
2397 }
2398 sched_unpin();
2399 vm_page_unlock_queues();
2400 PMAP_UNLOCK(src_pmap);
2401 PMAP_UNLOCK(dst_pmap);
2402 }
2403
2404 static __inline void
2405 pagezero(void *page)
2406 {
2407 #if defined(I686_CPU)
2408 if (cpu_class == CPUCLASS_686) {
2409 #if defined(CPU_ENABLE_SSE)
2410 if (cpu_feature & CPUID_SSE2)
2411 sse2_pagezero(page);
2412 else
2413 #endif
2414 i686_pagezero(page);
2415 } else
2416 #endif
2417 bzero(page, PAGE_SIZE);
2418 }
2419
2420 /*
2421 * pmap_zero_page zeros the specified hardware page by mapping
2422 * the page into KVM and using bzero to clear its contents.
2423 */
2424 void
2425 pmap_zero_page(vm_page_t m)
2426 {
2427 struct sysmaps *sysmaps;
2428
2429 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)];
2430 mtx_lock(&sysmaps->lock);
2431 if (*sysmaps->CMAP2)
2432 panic("pmap_zero_page: CMAP2 busy");
2433 sched_pin();
2434 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M;
2435 invlcaddr(sysmaps->CADDR2);
2436 pagezero(sysmaps->CADDR2);
2437 *sysmaps->CMAP2 = 0;
2438 sched_unpin();
2439 mtx_unlock(&sysmaps->lock);
2440 }
2441
2442 /*
2443 * pmap_zero_page_area zeros the specified hardware page by mapping
2444 * the page into KVM and using bzero to clear its contents.
2445 *
2446 * off and size may not cover an area beyond a single hardware page.
2447 */
2448 void
2449 pmap_zero_page_area(vm_page_t m, int off, int size)
2450 {
2451 struct sysmaps *sysmaps;
2452
2453 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)];
2454 mtx_lock(&sysmaps->lock);
2455 if (*sysmaps->CMAP2)
2456 panic("pmap_zero_page: CMAP2 busy");
2457 sched_pin();
2458 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M;
2459 invlcaddr(sysmaps->CADDR2);
2460 if (off == 0 && size == PAGE_SIZE)
2461 pagezero(sysmaps->CADDR2);
2462 else
2463 bzero((char *)sysmaps->CADDR2 + off, size);
2464 *sysmaps->CMAP2 = 0;
2465 sched_unpin();
2466 mtx_unlock(&sysmaps->lock);
2467 }
2468
2469 /*
2470 * pmap_zero_page_idle zeros the specified hardware page by mapping
2471 * the page into KVM and using bzero to clear its contents. This
2472 * is intended to be called from the vm_pagezero process only and
2473 * outside of Giant.
2474 */
2475 void
2476 pmap_zero_page_idle(vm_page_t m)
2477 {
2478
2479 if (*CMAP3)
2480 panic("pmap_zero_page: CMAP3 busy");
2481 sched_pin();
2482 *CMAP3 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M;
2483 invlcaddr(CADDR3);
2484 pagezero(CADDR3);
2485 *CMAP3 = 0;
2486 sched_unpin();
2487 }
2488
2489 /*
2490 * pmap_copy_page copies the specified (machine independent)
2491 * page by mapping the page into virtual memory and using
2492 * bcopy to copy the page, one machine dependent page at a
2493 * time.
2494 */
2495 void
2496 pmap_copy_page(vm_page_t src, vm_page_t dst)
2497 {
2498 struct sysmaps *sysmaps;
2499
2500 sysmaps = &sysmaps_pcpu[PCPU_GET(cpuid)];
2501 mtx_lock(&sysmaps->lock);
2502 if (*sysmaps->CMAP1)
2503 panic("pmap_copy_page: CMAP1 busy");
2504 if (*sysmaps->CMAP2)
2505 panic("pmap_copy_page: CMAP2 busy");
2506 sched_pin();
2507 invlpg((u_int)sysmaps->CADDR1);
2508 invlpg((u_int)sysmaps->CADDR2);
2509 *sysmaps->CMAP1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A;
2510 *sysmaps->CMAP2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M;
2511 bcopy(sysmaps->CADDR1, sysmaps->CADDR2, PAGE_SIZE);
2512 *sysmaps->CMAP1 = 0;
2513 *sysmaps->CMAP2 = 0;
2514 sched_unpin();
2515 mtx_unlock(&sysmaps->lock);
2516 }
2517
2518 /*
2519 * Returns true if the pmap's pv is one of the first
2520 * 16 pvs linked to from this page. This count may
2521 * be changed upwards or downwards in the future; it
2522 * is only necessary that true be returned for a small
2523 * subset of pmaps for proper page aging.
2524 */
2525 boolean_t
2526 pmap_page_exists_quick(pmap, m)
2527 pmap_t pmap;
2528 vm_page_t m;
2529 {
2530 pv_entry_t pv;
2531 int loops = 0;
2532
2533 if (m->flags & PG_FICTITIOUS)
2534 return FALSE;
2535
2536 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2537 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2538 if (pv->pv_pmap == pmap) {
2539 return TRUE;
2540 }
2541 loops++;
2542 if (loops >= 16)
2543 break;
2544 }
2545 return (FALSE);
2546 }
2547
2548 #define PMAP_REMOVE_PAGES_CURPROC_ONLY
2549 /*
2550 * Remove all pages from specified address space
2551 * this aids process exit speeds. Also, this code
2552 * is special cased for current process only, but
2553 * can have the more generic (and slightly slower)
2554 * mode enabled. This is much faster than pmap_remove
2555 * in the case of running down an entire address space.
2556 */
2557 void
2558 pmap_remove_pages(pmap, sva, eva)
2559 pmap_t pmap;
2560 vm_offset_t sva, eva;
2561 {
2562 pt_entry_t *pte, tpte;
2563 vm_page_t m;
2564 pv_entry_t pv, npv;
2565
2566 #ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY
2567 if (pmap != vmspace_pmap(curthread->td_proc->p_vmspace)) {
2568 printf("warning: pmap_remove_pages called with non-current pmap\n");
2569 return;
2570 }
2571 #endif
2572 vm_page_lock_queues();
2573 PMAP_LOCK(pmap);
2574 sched_pin();
2575 for (pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2576
2577 if (pv->pv_va >= eva || pv->pv_va < sva) {
2578 npv = TAILQ_NEXT(pv, pv_plist);
2579 continue;
2580 }
2581
2582 #ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY
2583 pte = vtopte(pv->pv_va);
2584 #else
2585 pte = pmap_pte_quick(pmap, pv->pv_va);
2586 #endif
2587 tpte = *pte;
2588
2589 if (tpte == 0) {
2590 printf("TPTE at %p IS ZERO @ VA %08x\n",
2591 pte, pv->pv_va);
2592 panic("bad pte");
2593 }
2594
2595 /*
2596 * We cannot remove wired pages from a process' mapping at this time
2597 */
2598 if (tpte & PG_W) {
2599 npv = TAILQ_NEXT(pv, pv_plist);
2600 continue;
2601 }
2602
2603 m = PHYS_TO_VM_PAGE(tpte);
2604 KASSERT(m->phys_addr == (tpte & PG_FRAME),
2605 ("vm_page_t %p phys_addr mismatch %016jx %016jx",
2606 m, (uintmax_t)m->phys_addr, (uintmax_t)tpte));
2607
2608 KASSERT(m < &vm_page_array[vm_page_array_size],
2609 ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte));
2610
2611 pmap->pm_stats.resident_count--;
2612
2613 pte_clear(pte);
2614
2615 /*
2616 * Update the vm_page_t clean and reference bits.
2617 */
2618 if (tpte & PG_M) {
2619 vm_page_dirty(m);
2620 }
2621
2622 npv = TAILQ_NEXT(pv, pv_plist);
2623 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
2624
2625 m->md.pv_list_count--;
2626 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2627 if (TAILQ_EMPTY(&m->md.pv_list))
2628 vm_page_flag_clear(m, PG_WRITEABLE);
2629
2630 pmap_unuse_pt(pmap, pv->pv_va);
2631 free_pv_entry(pv);
2632 }
2633 sched_unpin();
2634 pmap_invalidate_all(pmap);
2635 PMAP_UNLOCK(pmap);
2636 vm_page_unlock_queues();
2637 }
2638
2639 /*
2640 * pmap_is_modified:
2641 *
2642 * Return whether or not the specified physical page was modified
2643 * in any physical maps.
2644 */
2645 boolean_t
2646 pmap_is_modified(vm_page_t m)
2647 {
2648 pv_entry_t pv;
2649 pt_entry_t *pte;
2650 boolean_t rv;
2651
2652 rv = FALSE;
2653 if (m->flags & PG_FICTITIOUS)
2654 return (rv);
2655
2656 sched_pin();
2657 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2658 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2659 /*
2660 * if the bit being tested is the modified bit, then
2661 * mark clean_map and ptes as never
2662 * modified.
2663 */
2664 if (!pmap_track_modified(pv->pv_va))
2665 continue;
2666 PMAP_LOCK(pv->pv_pmap);
2667 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2668 rv = (*pte & PG_M) != 0;
2669 PMAP_UNLOCK(pv->pv_pmap);
2670 if (rv)
2671 break;
2672 }
2673 sched_unpin();
2674 return (rv);
2675 }
2676
2677 /*
2678 * pmap_is_prefaultable:
2679 *
2680 * Return whether or not the specified virtual address is elgible
2681 * for prefault.
2682 */
2683 boolean_t
2684 pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr)
2685 {
2686 pt_entry_t *pte;
2687 boolean_t rv;
2688
2689 rv = FALSE;
2690 PMAP_LOCK(pmap);
2691 if (*pmap_pde(pmap, addr)) {
2692 pte = vtopte(addr);
2693 rv = *pte == 0;
2694 }
2695 PMAP_UNLOCK(pmap);
2696 return (rv);
2697 }
2698
2699 /*
2700 * Clear the given bit in each of the given page's ptes. The bit is
2701 * expressed as a 32-bit mask. Consequently, if the pte is 64 bits in
2702 * size, only a bit within the least significant 32 can be cleared.
2703 */
2704 static __inline void
2705 pmap_clear_ptes(vm_page_t m, int bit)
2706 {
2707 register pv_entry_t pv;
2708 pt_entry_t pbits, *pte;
2709
2710 if ((m->flags & PG_FICTITIOUS) ||
2711 (bit == PG_RW && (m->flags & PG_WRITEABLE) == 0))
2712 return;
2713
2714 sched_pin();
2715 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2716 /*
2717 * Loop over all current mappings setting/clearing as appropos If
2718 * setting RO do we need to clear the VAC?
2719 */
2720 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2721 /*
2722 * don't write protect pager mappings
2723 */
2724 if (bit == PG_RW) {
2725 if (!pmap_track_modified(pv->pv_va))
2726 continue;
2727 }
2728
2729 PMAP_LOCK(pv->pv_pmap);
2730 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2731 retry:
2732 pbits = *pte;
2733 if (pbits & bit) {
2734 if (bit == PG_RW) {
2735 /*
2736 * Regardless of whether a pte is 32 or 64 bits
2737 * in size, PG_RW and PG_M are among the least
2738 * significant 32 bits.
2739 */
2740 if (!atomic_cmpset_int((u_int *)pte, pbits,
2741 pbits & ~(PG_RW | PG_M)))
2742 goto retry;
2743 if (pbits & PG_M) {
2744 vm_page_dirty(m);
2745 }
2746 } else {
2747 atomic_clear_int((u_int *)pte, bit);
2748 }
2749 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
2750 }
2751 PMAP_UNLOCK(pv->pv_pmap);
2752 }
2753 if (bit == PG_RW)
2754 vm_page_flag_clear(m, PG_WRITEABLE);
2755 sched_unpin();
2756 }
2757
2758 /*
2759 * pmap_page_protect:
2760 *
2761 * Lower the permission for all mappings to a given page.
2762 */
2763 void
2764 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2765 {
2766 if ((prot & VM_PROT_WRITE) == 0) {
2767 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2768 pmap_clear_ptes(m, PG_RW);
2769 } else {
2770 pmap_remove_all(m);
2771 }
2772 }
2773 }
2774
2775 /*
2776 * pmap_ts_referenced:
2777 *
2778 * Return a count of reference bits for a page, clearing those bits.
2779 * It is not necessary for every reference bit to be cleared, but it
2780 * is necessary that 0 only be returned when there are truly no
2781 * reference bits set.
2782 *
2783 * XXX: The exact number of bits to check and clear is a matter that
2784 * should be tested and standardized at some point in the future for
2785 * optimal aging of shared pages.
2786 */
2787 int
2788 pmap_ts_referenced(vm_page_t m)
2789 {
2790 register pv_entry_t pv, pvf, pvn;
2791 pt_entry_t *pte;
2792 pt_entry_t v;
2793 int rtval = 0;
2794
2795 if (m->flags & PG_FICTITIOUS)
2796 return (rtval);
2797
2798 sched_pin();
2799 mtx_assert(&vm_page_queue_mtx, MA_OWNED);
2800 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2801
2802 pvf = pv;
2803
2804 do {
2805 pvn = TAILQ_NEXT(pv, pv_list);
2806
2807 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2808
2809 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2810
2811 if (!pmap_track_modified(pv->pv_va))
2812 continue;
2813
2814 PMAP_LOCK(pv->pv_pmap);
2815 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2816
2817 if (pte && ((v = pte_load(pte)) & PG_A) != 0) {
2818 atomic_clear_int((u_int *)pte, PG_A);
2819 pmap_invalidate_page(pv->pv_pmap, pv->pv_va);
2820
2821 rtval++;
2822 if (rtval > 4) {
2823 PMAP_UNLOCK(pv->pv_pmap);
2824 break;
2825 }
2826 }
2827 PMAP_UNLOCK(pv->pv_pmap);
2828 } while ((pv = pvn) != NULL && pv != pvf);
2829 }
2830 sched_unpin();
2831
2832 return (rtval);
2833 }
2834
2835 /*
2836 * Clear the modify bits on the specified physical page.
2837 */
2838 void
2839 pmap_clear_modify(vm_page_t m)
2840 {
2841 pmap_clear_ptes(m, PG_M);
2842 }
2843
2844 /*
2845 * pmap_clear_reference:
2846 *
2847 * Clear the reference bit on the specified physical page.
2848 */
2849 void
2850 pmap_clear_reference(vm_page_t m)
2851 {
2852 pmap_clear_ptes(m, PG_A);
2853 }
2854
2855 /*
2856 * Miscellaneous support routines follow
2857 */
2858
2859 /*
2860 * Map a set of physical memory pages into the kernel virtual
2861 * address space. Return a pointer to where it is mapped. This
2862 * routine is intended to be used for mapping device memory,
2863 * NOT real memory.
2864 */
2865 void *
2866 pmap_mapdev(pa, size)
2867 vm_paddr_t pa;
2868 vm_size_t size;
2869 {
2870 vm_offset_t va, tmpva, offset;
2871
2872 offset = pa & PAGE_MASK;
2873 size = roundup(offset + size, PAGE_SIZE);
2874 pa = pa & PG_FRAME;
2875
2876 if (pa < KERNLOAD && pa + size <= KERNLOAD)
2877 va = KERNBASE + pa;
2878 else
2879 va = kmem_alloc_nofault(kernel_map, size);
2880 if (!va)
2881 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2882
2883 for (tmpva = va; size > 0; ) {
2884 pmap_kenter(tmpva, pa);
2885 size -= PAGE_SIZE;
2886 tmpva += PAGE_SIZE;
2887 pa += PAGE_SIZE;
2888 }
2889 pmap_invalidate_range(kernel_pmap, va, tmpva);
2890 return ((void *)(va + offset));
2891 }
2892
2893 void
2894 pmap_unmapdev(va, size)
2895 vm_offset_t va;
2896 vm_size_t size;
2897 {
2898 vm_offset_t base, offset, tmpva;
2899
2900 if (va >= KERNBASE && va + size <= KERNBASE + KERNLOAD)
2901 return;
2902 base = va & PG_FRAME;
2903 offset = va & PAGE_MASK;
2904 size = roundup(offset + size, PAGE_SIZE);
2905 for (tmpva = base; tmpva < (base + size); tmpva += PAGE_SIZE)
2906 pmap_kremove(tmpva);
2907 pmap_invalidate_range(kernel_pmap, va, tmpva);
2908 kmem_free(kernel_map, base, size);
2909 }
2910
2911 /*
2912 * perform the pmap work for mincore
2913 */
2914 int
2915 pmap_mincore(pmap, addr)
2916 pmap_t pmap;
2917 vm_offset_t addr;
2918 {
2919 pt_entry_t *ptep, pte;
2920 vm_page_t m;
2921 int val = 0;
2922
2923 PMAP_LOCK(pmap);
2924 ptep = pmap_pte(pmap, addr);
2925 pte = (ptep != NULL) ? *ptep : 0;
2926 pmap_pte_release(ptep);
2927 PMAP_UNLOCK(pmap);
2928
2929 if (pte != 0) {
2930 vm_paddr_t pa;
2931
2932 val = MINCORE_INCORE;
2933 if ((pte & PG_MANAGED) == 0)
2934 return val;
2935
2936 pa = pte & PG_FRAME;
2937
2938 m = PHYS_TO_VM_PAGE(pa);
2939
2940 /*
2941 * Modified by us
2942 */
2943 if (pte & PG_M)
2944 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
2945 else {
2946 /*
2947 * Modified by someone else
2948 */
2949 vm_page_lock_queues();
2950 if (m->dirty || pmap_is_modified(m))
2951 val |= MINCORE_MODIFIED_OTHER;
2952 vm_page_unlock_queues();
2953 }
2954 /*
2955 * Referenced by us
2956 */
2957 if (pte & PG_A)
2958 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
2959 else {
2960 /*
2961 * Referenced by someone else
2962 */
2963 vm_page_lock_queues();
2964 if ((m->flags & PG_REFERENCED) ||
2965 pmap_ts_referenced(m)) {
2966 val |= MINCORE_REFERENCED_OTHER;
2967 vm_page_flag_set(m, PG_REFERENCED);
2968 }
2969 vm_page_unlock_queues();
2970 }
2971 }
2972 return val;
2973 }
2974
2975 void
2976 pmap_activate(struct thread *td)
2977 {
2978 struct proc *p = td->td_proc;
2979 pmap_t pmap, oldpmap;
2980 u_int32_t cr3;
2981
2982 critical_enter();
2983 pmap = vmspace_pmap(td->td_proc->p_vmspace);
2984 oldpmap = PCPU_GET(curpmap);
2985 #if defined(SMP)
2986 atomic_clear_int(&oldpmap->pm_active, PCPU_GET(cpumask));
2987 atomic_set_int(&pmap->pm_active, PCPU_GET(cpumask));
2988 #else
2989 oldpmap->pm_active &= ~1;
2990 pmap->pm_active |= 1;
2991 #endif
2992 #ifdef PAE
2993 cr3 = vtophys(pmap->pm_pdpt);
2994 #else
2995 cr3 = vtophys(pmap->pm_pdir);
2996 #endif
2997 /* XXXKSE this is wrong.
2998 * pmap_activate is for the current thread on the current cpu
2999 */
3000 if (p->p_flag & P_SA) {
3001 /* Make sure all other cr3 entries are updated. */
3002 /* what if they are running? XXXKSE (maybe abort them) */
3003 FOREACH_THREAD_IN_PROC(p, td) {
3004 td->td_pcb->pcb_cr3 = cr3;
3005 }
3006 } else {
3007 td->td_pcb->pcb_cr3 = cr3;
3008 }
3009 load_cr3(cr3);
3010 PCPU_SET(curpmap, pmap);
3011 critical_exit();
3012 }
3013
3014 vm_offset_t
3015 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3016 {
3017
3018 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3019 return addr;
3020 }
3021
3022 addr = (addr + PDRMASK) & ~PDRMASK;
3023 return addr;
3024 }
3025
3026
3027 #if defined(PMAP_DEBUG)
3028 pmap_pid_dump(int pid)
3029 {
3030 pmap_t pmap;
3031 struct proc *p;
3032 int npte = 0;
3033 int index;
3034
3035 sx_slock(&allproc_lock);
3036 LIST_FOREACH(p, &allproc, p_list) {
3037 if (p->p_pid != pid)
3038 continue;
3039
3040 if (p->p_vmspace) {
3041 int i,j;
3042 index = 0;
3043 pmap = vmspace_pmap(p->p_vmspace);
3044 for (i = 0; i < NPDEPTD; i++) {
3045 pd_entry_t *pde;
3046 pt_entry_t *pte;
3047 vm_offset_t base = i << PDRSHIFT;
3048
3049 pde = &pmap->pm_pdir[i];
3050 if (pde && pmap_pde_v(pde)) {
3051 for (j = 0; j < NPTEPG; j++) {
3052 vm_offset_t va = base + (j << PAGE_SHIFT);
3053 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) {
3054 if (index) {
3055 index = 0;
3056 printf("\n");
3057 }
3058 sx_sunlock(&allproc_lock);
3059 return npte;
3060 }
3061 pte = pmap_pte(pmap, va);
3062 if (pte && pmap_pte_v(pte)) {
3063 pt_entry_t pa;
3064 vm_page_t m;
3065 pa = *pte;
3066 m = PHYS_TO_VM_PAGE(pa);
3067 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x",
3068 va, pa, m->hold_count, m->wire_count, m->flags);
3069 npte++;
3070 index++;
3071 if (index >= 2) {
3072 index = 0;
3073 printf("\n");
3074 } else {
3075 printf(" ");
3076 }
3077 }
3078 }
3079 }
3080 }
3081 }
3082 }
3083 sx_sunlock(&allproc_lock);
3084 return npte;
3085 }
3086 #endif
3087
3088 #if defined(DEBUG)
3089
3090 static void pads(pmap_t pm);
3091 void pmap_pvdump(vm_offset_t pa);
3092
3093 /* print address space of pmap*/
3094 static void
3095 pads(pm)
3096 pmap_t pm;
3097 {
3098 int i, j;
3099 vm_paddr_t va;
3100 pt_entry_t *ptep;
3101
3102 if (pm == kernel_pmap)
3103 return;
3104 for (i = 0; i < NPDEPTD; i++)
3105 if (pm->pm_pdir[i])
3106 for (j = 0; j < NPTEPG; j++) {
3107 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3108 if (pm == kernel_pmap && va < KERNBASE)
3109 continue;
3110 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS)
3111 continue;
3112 ptep = pmap_pte(pm, va);
3113 if (pmap_pte_v(ptep))
3114 printf("%x:%x ", va, *ptep);
3115 };
3116
3117 }
3118
3119 void
3120 pmap_pvdump(pa)
3121 vm_paddr_t pa;
3122 {
3123 pv_entry_t pv;
3124 vm_page_t m;
3125
3126 printf("pa %x", pa);
3127 m = PHYS_TO_VM_PAGE(pa);
3128 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3129 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);
3130 pads(pv->pv_pmap);
3131 }
3132 printf(" ");
3133 }
3134 #endif
Cache object: 0c0760ade3535cbeb804898203ad2053
|