FreeBSD/Linux Kernel Cross Reference
sys/i386/i386/pmap.c
1 /*
2 * Copyright (c) 1991 Regents of the University of California.
3 * All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
5 * All rights reserved.
6 * Copyright (c) 1994 David Greenman
7 * All rights reserved.
8 *
9 * This code is derived from software contributed to Berkeley by
10 * the Systems Programming Group of the University of Utah Computer
11 * Science Department and William Jolitz of UUNET Technologies Inc.
12 *
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
15 * are met:
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. All advertising materials mentioning features or use of this software
22 * must display the following acknowledgement:
23 * This product includes software developed by the University of
24 * California, Berkeley and its contributors.
25 * 4. Neither the name of the University nor the names of its contributors
26 * may be used to endorse or promote products derived from this software
27 * without specific prior written permission.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 * SUCH DAMAGE.
40 *
41 * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91
42 * $FreeBSD$
43 */
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 /*
78 * Manages physical address maps.
79 *
80 * In addition to hardware address maps, this
81 * module is called upon to provide software-use-only
82 * maps which may or may not be stored in the same
83 * form as hardware maps. These pseudo-maps are
84 * used to store intermediate results from copy
85 * operations to and from address spaces.
86 *
87 * Since the information managed by this module is
88 * also stored by the logical address mapping module,
89 * this module may throw away valid virtual-to-physical
90 * mappings at almost any time. However, invalidations
91 * of virtual-to-physical mappings must be done as
92 * requested.
93 *
94 * In order to cope with hardware architectures which
95 * make virtual-to-physical map invalidates expensive,
96 * this module may delay invalidate or reduced protection
97 * operations until such time as they are actually
98 * necessary. This module is given full information as
99 * to which processors are currently using which maps,
100 * and to when physical maps must be made correct.
101 */
102
103 #include "opt_disable_pse.h"
104 #include "opt_pmap.h"
105 #include "opt_msgbuf.h"
106 #include "opt_user_ldt.h"
107
108 #include <sys/param.h>
109 #include <sys/systm.h>
110 #include <sys/kernel.h>
111 #include <sys/proc.h>
112 #include <sys/msgbuf.h>
113 #include <sys/vmmeter.h>
114 #include <sys/mman.h>
115 #include <sys/malloc.h>
116
117 #include <machine/cpu.h>
118 #include <machine/ipl.h>
119 #include <vm/vm.h>
120 #include <vm/vm_param.h>
121 #include <sys/sysctl.h>
122 #include <sys/lock.h>
123 #include <vm/vm_kern.h>
124 #include <vm/vm_page.h>
125 #include <vm/vm_map.h>
126 #include <vm/vm_object.h>
127 #include <vm/vm_extern.h>
128 #include <vm/vm_pageout.h>
129 #include <vm/vm_pager.h>
130 #include <vm/vm_zone.h>
131
132 #include <sys/user.h>
133
134 #include <machine/cpu.h>
135 #include <machine/cputypes.h>
136 #include <machine/md_var.h>
137 #include <machine/specialreg.h>
138 #if defined(SMP) || defined(APIC_IO)
139 #include <machine/smp.h>
140 #include <machine/apic.h>
141 #include <machine/segments.h>
142 #include <machine/tss.h>
143 #include <machine/globaldata.h>
144 #endif /* SMP || APIC_IO */
145
146 #define PMAP_KEEP_PDIRS
147 #ifndef PMAP_SHPGPERPROC
148 #define PMAP_SHPGPERPROC 200
149 #endif
150
151 #if defined(DIAGNOSTIC)
152 #define PMAP_DIAGNOSTIC
153 #endif
154
155 #define MINPV 2048
156
157 #if !defined(PMAP_DIAGNOSTIC)
158 #define PMAP_INLINE __inline
159 #else
160 #define PMAP_INLINE
161 #endif
162
163 /*
164 * Get PDEs and PTEs for user/kernel address space
165 */
166 #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT]))
167 #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT])
168
169 #define pmap_pde_v(pte) ((*pte & PG_V) != 0)
170 #define pmap_pte_w(pte) ((*pte & PG_W) != 0)
171 #define pmap_pte_m(pte) ((*pte & PG_M) != 0)
172 #define pmap_pte_u(pte) ((*pte & PG_A) != 0)
173 #define pmap_pte_v(pte) ((*pte & PG_V) != 0)
174
175 #define pmap_pte_set_w(pte, v) ((v) ? (*pte |= PG_W) : (*pte &= ~PG_W))
176 #define pmap_pte_set_prot(pte, v) (*pte = (*pte & ~PG_PROT) | (v))
177
178 /*
179 * Given a map and a machine independent protection code,
180 * convert to a vax protection code.
181 */
182 #define pte_prot(m, p) (protection_codes[p])
183 static int protection_codes[8];
184
185 static struct pmap kernel_pmap_store;
186 pmap_t kernel_pmap;
187
188 vm_paddr_t avail_start; /* PA of first available physical page */
189 vm_paddr_t avail_end; /* PA of last available physical page */
190 vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */
191 vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */
192 static boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
193 static int pgeflag; /* PG_G or-in */
194 static int pseflag; /* PG_PS or-in */
195
196 static vm_object_t kptobj;
197
198 static int nkpt;
199 vm_offset_t kernel_vm_end;
200
201 /*
202 * Data for the pv entry allocation mechanism
203 */
204 static vm_zone_t pvzone;
205 static struct vm_zone pvzone_store;
206 static struct vm_object pvzone_obj;
207 static int pv_entry_count = 0, pv_entry_max = 0, pv_entry_high_water = 0;
208 static int pmap_pagedaemon_waken = 0;
209 static struct pv_entry *pvinit;
210
211 /*
212 * All those kernel PT submaps that BSD is so fond of
213 */
214 #ifdef SMP
215 extern pt_entry_t *SMPpt;
216 #define CMAP1 prv_CMAP1
217 #define CMAP2 prv_CMAP2
218 #define CMAP3 prv_CMAP3
219 #define PMAP1 prv_PMAP1
220 #define PMAP2 prv_PMAP2
221 #define CADDR1 prv_CADDR1
222 #define CADDR2 prv_CADDR2
223 #define CADDR3 prv_CADDR3
224 #define PADDR1 prv_PADDR1
225 #define PADDR2 prv_PADDR2
226 #else
227 static pt_entry_t *CMAP1, *CMAP2, *CMAP3;
228 static caddr_t CADDR1, CADDR2, CADDR3;
229 static pd_entry_t *PMAP1;
230 static pt_entry_t *PADDR1;
231 static pd_entry_t *PMAP2;
232 static pt_entry_t *PADDR2;
233 #endif
234
235 static pt_entry_t *ptmmap;
236 caddr_t ptvmmap = 0;
237 static pt_entry_t *msgbufmap;
238 struct msgbuf *msgbufp = 0;
239
240 /*
241 * Crashdump maps.
242 */
243 static pt_entry_t *pt_crashdumpmap;
244 static caddr_t crashdumpmap;
245
246 static pd_entry_t pdir4mb;
247
248 static PMAP_INLINE void free_pv_entry __P((pv_entry_t pv));
249 static pv_entry_t get_pv_entry __P((void));
250 static void i386_protection_init __P((void));
251 static __inline void pmap_changebit __P((vm_page_t m, int bit, boolean_t setem));
252 static void pmap_remove_all __P((vm_page_t m));
253 static vm_page_t pmap_enter_quick __P((pmap_t pmap, vm_offset_t va,
254 vm_page_t m, vm_page_t mpte));
255 static int pmap_remove_pte __P((pmap_t pmap, pt_entry_t* ptq,
256 vm_offset_t sva));
257 static void pmap_remove_page __P((pmap_t pmap, vm_offset_t va));
258 static int pmap_remove_entry __P((pmap_t pmap, vm_page_t m,
259 vm_offset_t va));
260 static boolean_t pmap_testbit __P((vm_page_t m, int bit));
261 static void pmap_insert_entry __P((pmap_t pmap, vm_offset_t va,
262 vm_page_t mpte, vm_page_t m));
263
264 static vm_page_t pmap_allocpte __P((pmap_t pmap, vm_offset_t va));
265 static vm_page_t _pmap_allocpte __P((pmap_t pmap, unsigned ptepindex));
266 static pt_entry_t *pmap_pte_quick __P((pmap_t pmap, vm_offset_t va));
267 static vm_page_t pmap_page_lookup __P((vm_object_t object, vm_pindex_t pindex));
268 static int pmap_unuse_pt __P((pmap_t, vm_offset_t, vm_page_t));
269 static vm_offset_t pmap_kmem_choose(vm_offset_t addr);
270
271 static int pmap_is_current(pmap_t);
272
273 #ifdef PAE
274 static pdpt_entry_t *pmap_alloc_pdpt(void);
275 static void pmap_free_pdpt(pdpt_entry_t *);
276 #endif
277 #if defined(I686_CPU) && !defined(NO_PSE_HACK)
278 static int has_pse_bug = 0; /* Initialized so that it can be patched. */
279 #endif
280
281 /*
282 * Move the kernel virtual free pointer to the next
283 * 4MB. This is used to help improve performance
284 * by using a large (4MB) page for much of the kernel
285 * (.text, .data, .bss)
286 */
287 static vm_offset_t
288 pmap_kmem_choose(vm_offset_t addr)
289 {
290 vm_offset_t newaddr = addr;
291 #if defined(I686_CPU) && !defined(NO_PSE_HACK)
292 /* Deal with un-resolved Pentium4 issues */
293 if (cpu == CPU_686 && (cpu_id & 0xf00) == 0xf00 &&
294 strcmp(cpu_vendor, "GenuineIntel") == 0) {
295 has_pse_bug = 1;
296 return newaddr;
297 }
298 #endif
299 #ifndef DISABLE_PSE
300 if (cpu_feature & CPUID_PSE) {
301 newaddr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
302 }
303 #endif
304 return newaddr;
305 }
306
307 /*
308 * Bootstrap the system enough to run with virtual memory.
309 *
310 * On the i386 this is called after mapping has already been enabled
311 * and just syncs the pmap module with what has already been done.
312 * [We can't call it easily with mapping off since the kernel is not
313 * mapped with PA == VA, hence we would have to relocate every address
314 * from the linked base (virtual) address "KERNBASE" to the actual
315 * (physical) address starting relative to 0]
316 */
317 void
318 pmap_bootstrap(vm_paddr_t firstaddr, vm_paddr_t loadaddr)
319 {
320 vm_offset_t va;
321 pt_entry_t *pte;
322 #ifdef SMP
323 struct globaldata *gd;
324 #endif
325 int i;
326
327 avail_start = firstaddr;
328
329 /*
330 * The calculation of virtual_avail is wrong. It's NKPT*PAGE_SIZE too
331 * large. It should instead be correctly calculated in locore.s and
332 * not based on 'first' (which is a physical address, not a virtual
333 * address, for the start of unused physical memory). The kernel
334 * page tables are NOT double mapped and thus should not be included
335 * in this calculation.
336 */
337 virtual_avail = (vm_offset_t) KERNBASE + firstaddr;
338 virtual_avail = pmap_kmem_choose(virtual_avail);
339
340 virtual_end = VM_MAX_KERNEL_ADDRESS;
341
342 /*
343 * Initialize protection array.
344 */
345 i386_protection_init();
346
347 /*
348 * The kernel's pmap is statically allocated so we don't have to use
349 * pmap_create, which is unlikely to work correctly at this part of
350 * the boot sequence (XXX and which no longer exists).
351 */
352 kernel_pmap = &kernel_pmap_store;
353
354 kernel_pmap->pm_pdir = (pd_entry_t *) (KERNBASE + IdlePTD);
355 kernel_pmap->pm_active = -1; /* don't allow deactivation */
356 #ifdef PAE
357 kernel_pmap->pm_pdpt = IdlePDPT;
358 #endif
359 TAILQ_INIT(&kernel_pmap->pm_pvlist);
360 nkpt = NKPT;
361
362 /*
363 * Reserve some special page table entries/VA space for temporary
364 * mapping of pages.
365 */
366 #define SYSMAP(c, p, v, n) \
367 v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n);
368
369 va = virtual_avail;
370 pte = vtopte(va);
371
372 #ifndef SMP
373 /*
374 * CMAP1/CMAP2/CMAP3 are used for zeroing and copying pages.
375 */
376 SYSMAP(caddr_t, CMAP1, CADDR1, 1)
377 SYSMAP(caddr_t, CMAP2, CADDR2, 1)
378 SYSMAP(caddr_t, CMAP3, CADDR3, 1)
379 *CMAP1 = *CMAP2 = *CMAP3 = 0;
380
381 /*
382 * ptemap is used for pmap_pte
383 */
384 SYSMAP(pd_entry_t *, PMAP1, PADDR1, 1);
385 SYSMAP(pd_entry_t *, PMAP2, PADDR2, 1);
386 #endif
387
388 /*
389 * Crashdump maps.
390 */
391 SYSMAP(caddr_t, pt_crashdumpmap, crashdumpmap, MAXDUMPPGS);
392
393 /*
394 * ptvmmap is used for reading arbitrary physical pages via /dev/mem.
395 * XXX ptmmap is not used.
396 */
397 SYSMAP(caddr_t, ptmmap, ptvmmap, 1)
398
399 /*
400 * msgbufp is used to map the system message buffer.
401 * XXX msgbufmap is not used.
402 */
403 SYSMAP(struct msgbuf *, msgbufmap, msgbufp,
404 atop(round_page(MSGBUF_SIZE)))
405
406 virtual_avail = va;
407
408 for (i = 0; i < NKPT; i++)
409 PTD[i] = 0;
410
411 /*
412 * Initialize the global page flag
413 */
414 pgeflag = 0;
415 #if !defined(SMP) /* XXX - see also mp_machdep.c */
416 if (cpu_feature & CPUID_PGE)
417 pgeflag = PG_G;
418 #endif
419
420 /*
421 * Initialize the 4MB page size flag
422 */
423 pseflag = 0;
424 #ifndef DISABLE_PSE
425 if (cpu_feature & CPUID_PSE)
426 pseflag = PG_PS;
427 #endif
428 #if defined(I686_CPU) && !defined(NO_PSE_HACK)
429 /* Deal with un-resolved Pentium4 issues */
430 if (has_pse_bug)
431 pseflag = 0;
432 #endif
433 /*
434 * The 4MB page version of the initial
435 * kernel page mapping.
436 */
437 if (pseflag) {
438 pd_entry_t ptditmp;
439 /*
440 * Note that we have enabled PSE mode
441 */
442 ptditmp = *(PTmap + i386_btop(KERNBASE));
443 ptditmp &= ~(NBPDR - 1);
444 ptditmp |= PG_V | PG_RW | PG_PS | PG_U | pgeflag;
445 pdir4mb = ptditmp;
446
447 #if !defined(SMP)
448 /*
449 * Enable the PSE mode.
450 */
451 load_cr4(rcr4() | CR4_PSE);
452
453 /*
454 * We can do the mapping here for the single processor
455 * case. We simply ignore the old page table page from
456 * now on.
457 */
458 /*
459 * For SMP, we still need 4K pages to bootstrap APs,
460 * PSE will be enabled as soon as all APs are up.
461 */
462 kernel_pmap->pm_pdir[KPTDI] = PTD[KPTDI] = pdir4mb;
463 invltlb();
464 #endif
465 }
466
467 #ifdef SMP
468 if (cpu_apic_address == 0)
469 panic("pmap_bootstrap: no local apic!");
470
471 /* local apic is mapped on last page */
472 SMPpt[NPTEPG - 1] = PG_V | PG_RW | PG_N | pgeflag |
473 (cpu_apic_address & PG_FRAME);
474
475 /* BSP does this itself, AP's get it pre-set */
476 gd = &SMP_prvspace[0].globaldata;
477 gd->gd_prv_CMAP1 = &SMPpt[1];
478 gd->gd_prv_CMAP2 = &SMPpt[2];
479 gd->gd_prv_CMAP3 = &SMPpt[3];
480 gd->gd_prv_PMAP1 = &SMPpt[4];
481 gd->gd_prv_PMAP2 = &SMPpt[5];
482 gd->gd_prv_CADDR1 = SMP_prvspace[0].CPAGE1;
483 gd->gd_prv_CADDR2 = SMP_prvspace[0].CPAGE2;
484 gd->gd_prv_CADDR3 = SMP_prvspace[0].CPAGE3;
485 gd->gd_prv_PADDR1 = (pt_entry_t *)SMP_prvspace[0].PPAGE1;
486 gd->gd_prv_PADDR2 = (pt_entry_t *)SMP_prvspace[0].PPAGE2;
487 #endif
488
489 invltlb();
490 }
491
492 #ifdef SMP
493 /*
494 * Set 4mb pdir for mp startup
495 */
496 void
497 pmap_set_opt(void)
498 {
499 if (pseflag && (cpu_feature & CPUID_PSE)) {
500 load_cr4(rcr4() | CR4_PSE);
501 if (pdir4mb && cpuid == 0) { /* only on BSP */
502 kernel_pmap->pm_pdir[KPTDI] = PTD[KPTDI] = pdir4mb;
503 cpu_invltlb();
504 }
505 }
506 }
507 #endif
508
509 /*
510 * Initialize the pmap module.
511 * Called by vm_init, to initialize any structures that the pmap
512 * system needs to map virtual memory.
513 * pmap_init has been enhanced to support in a fairly consistant
514 * way, discontiguous physical memory.
515 */
516 void
517 pmap_init(vm_paddr_t phys_start, vm_paddr_t phys_end)
518 {
519 int i;
520 int initial_pvs;
521
522 /*
523 * object for kernel page table pages
524 */
525 kptobj = vm_object_allocate(OBJT_DEFAULT, NKPDE);
526
527 /*
528 * Allocate memory for random pmap data structures. Includes the
529 * pv_head_table.
530 */
531
532 for(i = 0; i < vm_page_array_size; i++) {
533 vm_page_t m;
534
535 m = &vm_page_array[i];
536 TAILQ_INIT(&m->md.pv_list);
537 m->md.pv_list_count = 0;
538 }
539
540 /*
541 * init the pv free list
542 */
543 initial_pvs = vm_page_array_size;
544 if (initial_pvs < MINPV)
545 initial_pvs = MINPV;
546 pvzone = &pvzone_store;
547 pvinit = (struct pv_entry *) kmem_alloc(kernel_map,
548 initial_pvs * sizeof (struct pv_entry));
549 zbootinit(pvzone, "PV ENTRY", sizeof (struct pv_entry), pvinit,
550 vm_page_array_size);
551
552 /*
553 * Now it is safe to enable pv_table recording.
554 */
555 pmap_initialized = TRUE;
556 }
557
558 /*
559 * Initialize the address space (zone) for the pv_entries. Set a
560 * high water mark so that the system can recover from excessive
561 * numbers of pv entries.
562 */
563 void
564 pmap_init2()
565 {
566 int shpgperproc = PMAP_SHPGPERPROC;
567
568 TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc);
569 pv_entry_max = shpgperproc * maxproc + vm_page_array_size;
570 TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max);
571 pv_entry_high_water = 9 * (pv_entry_max / 10);
572 zinitna(pvzone, &pvzone_obj, NULL, 0, pv_entry_max, ZONE_INTERRUPT, 1);
573 }
574
575
576 /***************************************************
577 * Low level helper routines.....
578 ***************************************************/
579
580 #if defined(PMAP_DIAGNOSTIC)
581 /*
582 * This code checks for non-writeable/modified pages.
583 * This should be an invalid condition.
584 */
585 static int
586 pmap_nw_modified(pt_entry_t pte)
587 {
588 return ((pte & (PG_M|PG_RW)) == PG_M);
589 }
590 #endif
591
592 /*
593 * this routine defines the region(s) of memory that should
594 * not be tested for the modified bit.
595 */
596 static PMAP_INLINE int
597 pmap_track_modified(vm_offset_t va)
598 {
599 return (va < clean_sva) || (va >= clean_eva);
600 }
601
602 static PMAP_INLINE void
603 invltlb_1pg(vm_offset_t va)
604 {
605 #if defined(I386_CPU)
606 if (cpu_class == CPUCLASS_386) {
607 invltlb();
608 } else
609 #endif
610 {
611 invlpg(va);
612 }
613 }
614
615 static __inline void
616 pmap_TLB_invalidate(pmap_t pmap, vm_offset_t va)
617 {
618 #if defined(SMP)
619 if (pmap->pm_active & (1 << cpuid))
620 cpu_invlpg((void *)va);
621 if (pmap->pm_active & other_cpus)
622 smp_invltlb();
623 #else
624 if (pmap->pm_active)
625 invltlb_1pg(va);
626 #endif
627 }
628
629 static __inline void
630 pmap_TLB_invalidate_all(pmap_t pmap)
631 {
632 #if defined(SMP)
633 if (pmap->pm_active & (1 << cpuid))
634 cpu_invltlb();
635 if (pmap->pm_active & other_cpus)
636 smp_invltlb();
637 #else
638 if (pmap->pm_active)
639 invltlb();
640 #endif
641 }
642
643 #ifdef PAE
644 static __inline pt_entry_t
645 pte_load(pt_entry_t *pte)
646 {
647 pt_entry_t rv = 0;
648 __asm __volatile(MPLOCKED "cmpxchg8b %1"
649 : "+A" (rv) : "m" (*pte), "b" (0), "c" (0));
650 return rv;
651 }
652
653 static __inline pt_entry_t
654 pte_store(pt_entry_t *pte, pt_entry_t v)
655 {
656 pt_entry_t rv = *pte;
657 __asm __volatile("1:;" MPLOCKED "cmpxchg8b %1; jnz 1b"
658 : "+A" (rv)
659 : "m" (*pte), "b" ((u_int32_t)v), "c" ((u_int32_t)(v >> 32)));
660 return rv;
661 }
662 #else
663 static __inline pt_entry_t
664 pte_load(pt_entry_t *pte)
665 {
666 return *pte;
667 }
668
669 static __inline pt_entry_t
670 pte_store(pt_entry_t *pte, pt_entry_t v)
671 {
672 __asm __volatile("xchgl %1,%0" : "+r" (v) : "m" (*pte));
673 return v;
674 }
675 #endif
676
677 /*
678 * Are we current address space or kernel?
679 */
680 static __inline int
681 pmap_is_current(pmap_t pmap)
682 {
683 return (pmap == kernel_pmap ||
684 (pmap->pm_pdir[PTDPTDI] & PG_FRAME) == (PTDpde[0] & PG_FRAME));
685 }
686
687 /*
688 * Routine: pmap_pte
689 * Function:
690 * Extract the page table entry associated
691 * with the given map/virtual_address pair.
692 * Note: Must be protected by splvm()
693 */
694
695 static pt_entry_t *
696 pmap_pte_quick(pmap_t pmap, vm_offset_t va)
697 {
698 pd_entry_t *pde, newpf;
699
700 #ifdef INVARIANTS
701 if (~cpl & (net_imask | bio_imask | cam_imask))
702 panic("pmap_pte_quick not protected by splvm()");
703 #endif
704 pde = pmap_pde(pmap, va);
705 if (*pde & PG_V) {
706 if (*pde & PG_PS)
707 return (pt_entry_t *)pde;
708 if (pmap_is_current(pmap))
709 return vtopte(va);
710 newpf = *pde & PG_FRAME;
711 if ((*PMAP1 & PG_FRAME) != newpf) {
712 *PMAP1 = newpf | PG_RW | PG_V;
713 #ifdef SMP
714 cpu_invlpg(PADDR1);
715 #else
716 invltlb_1pg((vm_offset_t) PADDR1);
717 #endif
718 }
719 return PADDR1 + (i386_btop(va) & (NPTEPG - 1));
720 }
721 return (0);
722 }
723
724 /*
725 * Routine: pmap_pte
726 * Function:
727 * Extract the page table entry associated
728 * with the given map/virtual_address pair.
729 * Note: Must not be called from interrupts on non-current pmap
730 */
731
732 pt_entry_t *
733 pmap_pte(pmap_t pmap, vm_offset_t va)
734 {
735 pd_entry_t *pde, newpf;
736
737 pde = pmap_pde(pmap, va);
738 if (*pde & PG_V) {
739 if (*pde & PG_PS)
740 return (pt_entry_t *)pde;
741 if (pmap_is_current(pmap))
742 return vtopte(va);
743 #ifdef INVARIANTS
744 if (intr_nesting_level != 0) {
745 panic("pmap_pte called from interrupt");
746 }
747 #endif
748 newpf = *pde & PG_FRAME;
749 if ((*PMAP2 & PG_FRAME) != newpf) {
750 *PMAP2 = newpf | PG_RW | PG_V;
751 #ifdef SMP
752 cpu_invlpg(PADDR2);
753 #else
754 invltlb_1pg((vm_offset_t) PADDR2);
755 #endif
756 }
757 return PADDR2 + (i386_btop(va) & (NPTEPG - 1));
758 }
759 return (0);
760 }
761
762 /*
763 * Routine: pmap_extract
764 * Function:
765 * Extract the physical page address associated
766 * with the given map/virtual_address pair.
767 */
768 vm_paddr_t
769 pmap_extract(pmap_t pmap, vm_offset_t va)
770 {
771 pt_entry_t *pte;
772
773 if (pmap == 0)
774 return 0;
775 pte = pmap_pte(pmap, va);
776 if (pte) {
777 if (*pte & PG_PS)
778 return (*pte & ~PDRMASK) | (va & PDRMASK);
779 return (*pte & PG_FRAME) | (va & PAGE_MASK);
780 }
781 return 0;
782 }
783
784 /***************************************************
785 * Low level mapping routines.....
786 ***************************************************/
787
788 /*
789 * add a wired page to the kva
790 * note that in order for the mapping to take effect -- you
791 * should do a invltlb after doing the pmap_kenter...
792 */
793 PMAP_INLINE void
794 pmap_kenter(vm_offset_t va, vm_paddr_t pa)
795 {
796 pt_entry_t *pte;
797
798 pte = vtopte(va);
799 *pte = pa | PG_RW | PG_V | pgeflag;
800 invltlb_1pg(va);
801 }
802
803 /*
804 * remove a page from the kernel pagetables
805 */
806 PMAP_INLINE void
807 pmap_kremove(vm_offset_t va)
808 {
809 pt_entry_t *pte;
810
811 pte = vtopte(va);
812 *pte = 0;
813 invltlb_1pg(va);
814 }
815
816 /*
817 * Used to map a range of physical addresses into kernel
818 * virtual address space.
819 *
820 * For now, VM is already on, we only need to map the
821 * specified memory.
822 */
823 vm_offset_t
824 pmap_map(vm_offset_t virt, vm_paddr_t start, vm_paddr_t end, int prot)
825 {
826 while (start < end) {
827 pmap_kenter(virt, start);
828 virt += PAGE_SIZE;
829 start += PAGE_SIZE;
830 }
831 return (virt);
832 }
833
834
835 /*
836 * Add a list of wired pages to the kva
837 * this routine is only used for temporary
838 * kernel mappings that do not need to have
839 * page modification or references recorded.
840 * Note that old mappings are simply written
841 * over. The page *must* be wired.
842 */
843 void
844 pmap_qenter(vm_offset_t va, vm_page_t *m, int count)
845 {
846 while (count-- > 0) {
847 pt_entry_t *pte = vtopte(va);
848 *pte = VM_PAGE_TO_PHYS(*m) | PG_RW | PG_V | pgeflag;
849 #ifdef SMP
850 cpu_invlpg((void *)va);
851 #else
852 invltlb_1pg(va);
853 #endif
854 va += PAGE_SIZE;
855 m++;
856 }
857 #ifdef SMP
858 smp_invltlb();
859 #endif
860 }
861
862 /*
863 * this routine jerks page mappings from the
864 * kernel -- it is meant only for temporary mappings.
865 */
866 void
867 pmap_qremove(vm_offset_t va, int count)
868 {
869 while (count-- > 0) {
870 pt_entry_t *pte = vtopte(va);
871 *pte = 0;
872 #ifdef SMP
873 cpu_invlpg((void *)va);
874 #else
875 invltlb_1pg(va);
876 #endif
877 va += PAGE_SIZE;
878 }
879 #ifdef SMP
880 smp_invltlb();
881 #endif
882 }
883
884 static vm_page_t
885 pmap_page_lookup(vm_object_t object, vm_pindex_t pindex)
886 {
887 vm_page_t m;
888 retry:
889 m = vm_page_lookup(object, pindex);
890 if (m && vm_page_sleep_busy(m, FALSE, "pplookp"))
891 goto retry;
892 return m;
893 }
894
895 /*
896 * Create the UPAGES for a new process.
897 * This routine directly affects the fork perf for a process.
898 */
899 void
900 pmap_new_proc(struct proc *p)
901 {
902 int i;
903 vm_object_t upobj;
904 vm_page_t m, ma[UPAGES];
905 vm_offset_t up;
906
907 /*
908 * allocate object for the upages
909 */
910 if ((upobj = p->p_upages_obj) == NULL) {
911 upobj = vm_object_allocate( OBJT_DEFAULT, UPAGES);
912 p->p_upages_obj = upobj;
913 }
914
915 /* get a kernel virtual address for the UPAGES for this proc */
916 if ((up = (vm_offset_t) p->p_addr) == 0) {
917 up = kmem_alloc_nofault(kernel_map, UPAGES * PAGE_SIZE);
918 if (up == 0)
919 panic("pmap_new_proc: u_map allocation failed");
920 p->p_addr = (struct user *) up;
921 }
922
923 for(i = 0; i < UPAGES; i++) {
924 /*
925 * Get a kernel stack page
926 */
927 m = vm_page_grab(upobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
928 ma[i] = m;
929
930 /*
931 * Wire the page
932 */
933 m->wire_count++;
934 cnt.v_wire_count++;
935
936 vm_page_wakeup(m);
937 vm_page_flag_clear(m, PG_ZERO);
938 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
939 m->valid = VM_PAGE_BITS_ALL;
940 }
941 pmap_qenter(up, ma, UPAGES);
942 }
943
944 /*
945 * Dispose the UPAGES for a process that has exited.
946 * This routine directly impacts the exit perf of a process.
947 */
948 void
949 pmap_dispose_proc(struct proc *p)
950 {
951 int i;
952 vm_object_t upobj;
953 vm_page_t m;
954
955 upobj = p->p_upages_obj;
956 pmap_qremove((vm_offset_t) p->p_addr, UPAGES);
957
958 for(i = 0; i < UPAGES; i++) {
959 if ((m = vm_page_lookup(upobj, i)) == NULL)
960 panic("pmap_dispose_proc: upage already missing???");
961
962 vm_page_busy(m);
963 vm_page_unwire(m, 0);
964 vm_page_free(m);
965 }
966
967 /*
968 * If the process got swapped out some of its UPAGES might have gotten
969 * swapped. Just get rid of the object to clean up the swap use
970 * proactively. NOTE! might block waiting for paging I/O to complete.
971 */
972 if (upobj->type == OBJT_SWAP) {
973 p->p_upages_obj = NULL;
974 vm_object_deallocate(upobj);
975 }
976 }
977
978 /*
979 * Allow the UPAGES for a process to be prejudicially paged out.
980 */
981 void
982 pmap_swapout_proc(struct proc *p)
983 {
984 int i;
985 vm_object_t upobj;
986 vm_page_t m;
987
988 upobj = p->p_upages_obj;
989 pmap_qremove((vm_offset_t) p->p_addr, UPAGES);
990
991 /*
992 * let the upages be paged
993 */
994 for(i = 0; i < UPAGES; i++) {
995 if ((m = vm_page_lookup(upobj, i)) == NULL)
996 panic("pmap_swapout_proc: upage already missing???");
997 vm_page_dirty(m);
998 vm_page_unwire(m, 0);
999 }
1000 }
1001
1002 /*
1003 * Bring the UPAGES for a specified process back in.
1004 */
1005 void
1006 pmap_swapin_proc(struct proc *p)
1007 {
1008 int i, rv;
1009 vm_object_t upobj;
1010 vm_page_t m, ma[UPAGES];
1011
1012 upobj = p->p_upages_obj;
1013
1014 for(i = 0; i < UPAGES; i++) {
1015 m = vm_page_grab(upobj, i, VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1016 ma[i] = m;
1017
1018 if (m->valid != VM_PAGE_BITS_ALL) {
1019 rv = vm_pager_get_pages(upobj, &m, 1, 0);
1020 if (rv != VM_PAGER_OK)
1021 panic("pmap_swapin_proc: cannot get upages for proc: %d\n", p->p_pid);
1022 m = vm_page_lookup(upobj, i);
1023 m->valid = VM_PAGE_BITS_ALL;
1024 }
1025
1026 vm_page_wire(m);
1027 vm_page_wakeup(m);
1028 vm_page_flag_set(m, PG_MAPPED | PG_WRITEABLE);
1029 }
1030
1031 pmap_qenter((vm_offset_t) p->p_addr, ma, UPAGES);
1032 }
1033
1034 /***************************************************
1035 * Page table page management routines.....
1036 ***************************************************/
1037
1038 /*
1039 * This routine unholds page table pages, and if the hold count
1040 * drops to zero, then it decrements the wire count.
1041 */
1042 static int
1043 _pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
1044 {
1045 while (vm_page_sleep_busy(m, FALSE, "pmuwpt"))
1046 ;
1047
1048 if (m->hold_count == 0) {
1049 vm_offset_t pteva;
1050 /*
1051 * unmap the page table page
1052 */
1053 pmap->pm_pdir[m->pindex] = 0;
1054 --pmap->pm_stats.resident_count;
1055 if (pmap_is_current(pmap)) {
1056 /*
1057 * Do a invltlb to make the invalidated mapping
1058 * take effect immediately.
1059 */
1060 pteva = UPT_MIN_ADDRESS + i386_ptob(m->pindex);
1061 pmap_TLB_invalidate(pmap, pteva);
1062 }
1063
1064 if (pmap->pm_ptphint == m)
1065 pmap->pm_ptphint = NULL;
1066
1067 /*
1068 * If the page is finally unwired, simply free it.
1069 */
1070 --m->wire_count;
1071 if (m->wire_count == 0) {
1072 vm_page_flash(m);
1073 vm_page_busy(m);
1074 vm_page_free_zero(m);
1075 --cnt.v_wire_count;
1076 }
1077 return 1;
1078 }
1079 return 0;
1080 }
1081
1082 static PMAP_INLINE int
1083 pmap_unwire_pte_hold(pmap_t pmap, vm_page_t m)
1084 {
1085 vm_page_unhold(m);
1086 if (m->hold_count == 0)
1087 return _pmap_unwire_pte_hold(pmap, m);
1088 else
1089 return 0;
1090 }
1091
1092 /*
1093 * After removing a page table entry, this routine is used to
1094 * conditionally free the page, and manage the hold/wire counts.
1095 */
1096 static int
1097 pmap_unuse_pt(pmap_t pmap, vm_offset_t va, vm_page_t mpte)
1098 {
1099 unsigned ptepindex;
1100
1101 if (va >= UPT_MIN_ADDRESS)
1102 return 0;
1103
1104 if (mpte == NULL) {
1105 ptepindex = (va >> PDRSHIFT);
1106 if (pmap->pm_ptphint &&
1107 (pmap->pm_ptphint->pindex == ptepindex)) {
1108 mpte = pmap->pm_ptphint;
1109 } else {
1110 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1111 pmap->pm_ptphint = mpte;
1112 }
1113 }
1114
1115 return pmap_unwire_pte_hold(pmap, mpte);
1116 }
1117
1118 void
1119 pmap_pinit0(pmap_t pmap)
1120 {
1121 pmap->pm_pdir = (pd_entry_t *)(KERNBASE + IdlePTD);
1122 pmap->pm_active = 0;
1123 pmap->pm_ptphint = NULL;
1124 TAILQ_INIT(&pmap->pm_pvlist);
1125 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1126 #ifdef PAE
1127 pmap->pm_pdpt = IdlePDPT;
1128 #endif
1129 }
1130
1131 /*
1132 * Initialize a preallocated and zeroed pmap structure,
1133 * such as one in a vmspace structure.
1134 */
1135 void
1136 pmap_pinit(pmap_t pmap)
1137 {
1138 vm_page_t m, ma[NPGPTD];
1139 vm_paddr_t pa;
1140 int i;
1141
1142 /*
1143 * No need to allocate page table space yet but we do need a valid
1144 * page directory table.
1145 */
1146 if (pmap->pm_pdir == NULL) {
1147 pmap->pm_pdir = (pd_entry_t *)kmem_alloc_pageable(kernel_map,
1148 NPGPTD * PAGE_SIZE);
1149 #ifdef PAE
1150 pmap->pm_pdpt = pmap_alloc_pdpt();
1151 #endif
1152 }
1153
1154 /*
1155 * allocate object for the ptes
1156 */
1157 if (pmap->pm_pteobj == NULL)
1158 pmap->pm_pteobj = vm_object_allocate(OBJT_DEFAULT,
1159 PTDPTDI + NPGPTD);
1160
1161 /*
1162 * allocate the page directory page
1163 */
1164 for (i = 0; i < NPGPTD; i++) {
1165 m = vm_page_grab(pmap->pm_pteobj, PTDPTDI + i,
1166 VM_ALLOC_NORMAL | VM_ALLOC_RETRY);
1167 ma[i] = m;
1168
1169 m->wire_count = 1;
1170 ++cnt.v_wire_count;
1171
1172 vm_page_flag_clear(m, PG_MAPPED | PG_BUSY);
1173 m->valid = VM_PAGE_BITS_ALL;
1174 }
1175
1176 pmap_qenter((vm_offset_t)pmap->pm_pdir, ma, NPGPTD);
1177
1178 for (i = 0; i < NPGPTD; i++) {
1179 if ((ma[i]->flags & PG_ZERO) == 0)
1180 bzero(pmap->pm_pdir + i * NPDEPG, PAGE_SIZE);
1181 }
1182
1183 #ifdef SMP
1184 pmap->pm_pdir[MPPTDI] = PTD[MPPTDI];
1185 #endif
1186
1187 /* install self-referential address mapping entry */
1188 for (i = 0; i < NPGPTD; i++) {
1189 pa = VM_PAGE_TO_PHYS(ma[i]);
1190 pmap->pm_pdir[PTDPTDI + i] = pa | PG_V | PG_RW | PG_A | PG_M;
1191 #ifdef PAE
1192 pmap->pm_pdpt[i] = pa | PG_V;
1193 #endif
1194 }
1195
1196 pmap->pm_active = 0;
1197 pmap->pm_ptphint = NULL;
1198 TAILQ_INIT(&pmap->pm_pvlist);
1199 bzero(&pmap->pm_stats, sizeof pmap->pm_stats);
1200 }
1201
1202 /*
1203 * Wire in kernel global address entries. To avoid a race condition
1204 * between pmap initialization and pmap_growkernel, this procedure
1205 * should be called after the vmspace is attached to the process
1206 * but before this pmap is activated.
1207 */
1208 void
1209 pmap_pinit2(pmap_t pmap)
1210 {
1211 /* XXX copies current process, does not fill in MPPTDI */
1212 bcopy(PTD + KPTDI, pmap->pm_pdir + KPTDI, nkpt * PDESIZE);
1213 }
1214
1215 /*
1216 * this routine is called if the page table page is not
1217 * mapped correctly.
1218 */
1219 static vm_page_t
1220 _pmap_allocpte(pmap_t pmap, unsigned ptepindex)
1221 {
1222 vm_paddr_t ptepa;
1223 vm_page_t m;
1224
1225 /*
1226 * Find or fabricate a new pagetable page
1227 */
1228 m = vm_page_grab(pmap->pm_pteobj, ptepindex,
1229 VM_ALLOC_ZERO);
1230 if (m == NULL) {
1231 VM_WAIT;
1232 /*
1233 * Indicate the need to retry. While waiting, the page table
1234 * page may have been allocated.
1235 */
1236 return (NULL);
1237 }
1238 if ((m->flags & PG_ZERO) == 0)
1239 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1240
1241 KASSERT(m->queue == PQ_NONE,
1242 ("_pmap_allocpte: %p->queue != PQ_NONE", m));
1243
1244 if (m->wire_count == 0)
1245 cnt.v_wire_count++;
1246 m->wire_count++;
1247
1248 /*
1249 * Increment the hold count for the page table page
1250 * (denoting a new mapping.)
1251 */
1252 m->hold_count++;
1253
1254 /*
1255 * Map the pagetable page into the process address space, if
1256 * it isn't already there.
1257 */
1258
1259 pmap->pm_stats.resident_count++;
1260
1261 ptepa = VM_PAGE_TO_PHYS(m);
1262 pmap->pm_pdir[ptepindex] = ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M;
1263
1264 /*
1265 * Set the page table hint
1266 */
1267 pmap->pm_ptphint = m;
1268
1269 m->valid = VM_PAGE_BITS_ALL;
1270 vm_page_flag_clear(m, PG_ZERO);
1271 vm_page_flag_set(m, PG_MAPPED);
1272 vm_page_wakeup(m);
1273
1274 return m;
1275 }
1276
1277 static vm_page_t
1278 pmap_allocpte(pmap_t pmap, vm_offset_t va)
1279 {
1280 unsigned ptepindex;
1281 pd_entry_t pde;
1282 vm_page_t m;
1283
1284 /*
1285 * Calculate pagetable page index
1286 */
1287 ptepindex = va >> PDRSHIFT;
1288
1289 retry:
1290 /*
1291 * Get the page directory entry
1292 */
1293 pde = pmap->pm_pdir[ptepindex];
1294
1295 /*
1296 * This supports switching from a 4MB page to a
1297 * normal 4K page.
1298 */
1299 if (pde & PG_PS) {
1300 pmap->pm_pdir[ptepindex] = 0;
1301 pde = 0;
1302 invltlb();
1303 }
1304
1305 /*
1306 * If the page table page is mapped, we just increment the
1307 * hold count, and activate it.
1308 */
1309 if (pde & PG_V) {
1310 /*
1311 * In order to get the page table page, try the
1312 * hint first.
1313 */
1314 if (pmap->pm_ptphint && pmap->pm_ptphint->pindex == ptepindex) {
1315 m = pmap->pm_ptphint;
1316 } else {
1317 m = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
1318 pmap->pm_ptphint = m;
1319 }
1320 m->hold_count++;
1321 } else {
1322 /*
1323 * Here if the pte page isn't mapped, or if it has
1324 * been deallocated.
1325 */
1326 m = _pmap_allocpte(pmap, ptepindex);
1327 if (m == NULL)
1328 goto retry;
1329 }
1330 return (m);
1331 }
1332
1333
1334 /***************************************************
1335 * Pmap allocation/deallocation routines.
1336 ***************************************************/
1337
1338 /*
1339 * Release any resources held by the given physical map.
1340 * Called when a pmap initialized by pmap_pinit is being released.
1341 * Should only be called if the map contains no valid mappings.
1342 */
1343 void
1344 pmap_release(pmap_t pmap)
1345 {
1346 vm_page_t m;
1347 vm_object_t object = pmap->pm_pteobj;
1348
1349 bzero(pmap->pm_pdir + PTDPTDI, (nkpt + NPGPTD) * PDESIZE);
1350 #ifdef SMP
1351 pmap->pm_pdir[MPPTDI] = 0;
1352 #endif
1353 pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD);
1354
1355 while ((m = TAILQ_FIRST(&object->memq))) {
1356 if (m->pindex < PTDPTDI || m->pindex >= KPTDI)
1357 panic("pmap_release: non ptd page");
1358 m->wire_count--;
1359 cnt.v_wire_count--;
1360 vm_page_busy(m);
1361 vm_page_free_zero(m);
1362 }
1363 }
1364
1365
1366 static int
1367 kvm_size(SYSCTL_HANDLER_ARGS)
1368 {
1369 unsigned long ksize = VM_MAX_KERNEL_ADDRESS - KERNBASE;
1370
1371 return sysctl_handle_long(oidp, &ksize, 0, req);
1372 }
1373 SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG|CTLFLAG_RD,
1374 0, 0, kvm_size, "IU", "Size of KVM");
1375
1376 static int
1377 kvm_free(SYSCTL_HANDLER_ARGS)
1378 {
1379 unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end;
1380
1381 return sysctl_handle_long(oidp, &kfree, 0, req);
1382 }
1383 SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG|CTLFLAG_RD,
1384 0, 0, kvm_free, "IU", "Amount of KVM free");
1385
1386 /*
1387 * grow the number of kernel page table entries, if needed
1388 */
1389 void
1390 pmap_growkernel(vm_offset_t addr)
1391 {
1392 struct proc *p;
1393 struct pmap *pmap;
1394 int s;
1395 vm_paddr_t ptppaddr;
1396 vm_page_t nkpg;
1397 pd_entry_t newpdir;
1398
1399 s = splhigh();
1400 if (kernel_vm_end == 0) {
1401 kernel_vm_end = KERNBASE;
1402 nkpt = 0;
1403 while (pdir_pde(PTD, kernel_vm_end) & PG_V) {
1404 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1405 ~(PAGE_SIZE * NPTEPG - 1);
1406 nkpt++;
1407 }
1408 }
1409 addr = (addr + PAGE_SIZE * NPTEPG) & ~(PAGE_SIZE * NPTEPG - 1);
1410 while (kernel_vm_end < addr) {
1411 if (pdir_pde(PTD, kernel_vm_end) & PG_V) {
1412 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1413 ~(PAGE_SIZE * NPTEPG - 1);
1414 continue;
1415 }
1416
1417 /*
1418 * This index is bogus, but out of the way
1419 */
1420 nkpg = vm_page_alloc(kptobj, nkpt, VM_ALLOC_SYSTEM);
1421 if (!nkpg)
1422 panic("pmap_growkernel: no memory to grow kernel");
1423
1424 nkpt++;
1425
1426 vm_page_wire(nkpg);
1427 ptppaddr = VM_PAGE_TO_PHYS(nkpg);
1428 pmap_zero_page(ptppaddr);
1429 newpdir = ptppaddr | PG_V | PG_RW | PG_A | PG_M;
1430 pdir_pde(PTD, kernel_vm_end) = newpdir;
1431
1432 LIST_FOREACH(p, &allproc, p_list) {
1433 if (p->p_vmspace) {
1434 pmap = vmspace_pmap(p->p_vmspace);
1435 *pmap_pde(pmap, kernel_vm_end) = newpdir;
1436 }
1437 }
1438 *pmap_pde(kernel_pmap, kernel_vm_end) = newpdir;
1439 kernel_vm_end = (kernel_vm_end + PAGE_SIZE * NPTEPG) &
1440 ~(PAGE_SIZE * NPTEPG - 1);
1441 }
1442 splx(s);
1443 }
1444
1445 /***************************************************
1446 * page management routines.
1447 ***************************************************/
1448
1449 /*
1450 * free the pv_entry back to the free list
1451 */
1452 static PMAP_INLINE void
1453 free_pv_entry(pv_entry_t pv)
1454 {
1455 pv_entry_count--;
1456 zfreei(pvzone, pv);
1457 }
1458
1459 /*
1460 * get a new pv_entry, allocating a block from the system
1461 * when needed.
1462 * the memory allocation is performed bypassing the malloc code
1463 * because of the possibility of allocations at interrupt time.
1464 */
1465 static pv_entry_t
1466 get_pv_entry(void)
1467 {
1468 pv_entry_count++;
1469 if (pv_entry_high_water && (pv_entry_count > pv_entry_high_water) &&
1470 (pmap_pagedaemon_waken == 0)) {
1471 pmap_pagedaemon_waken = 1;
1472 wakeup (&vm_pages_needed);
1473 }
1474 return zalloci(pvzone);
1475 }
1476
1477 /*
1478 * This routine is very drastic, but can save the system
1479 * in a pinch.
1480 */
1481 void
1482 pmap_collect()
1483 {
1484 int i;
1485 vm_page_t m;
1486 static int warningdone=0;
1487
1488 if (pmap_pagedaemon_waken == 0)
1489 return;
1490
1491 if (warningdone < 5) {
1492 printf("pmap_collect: collecting pv entries -- suggest increasing PMAP_SHPGPERPROC\n");
1493 warningdone++;
1494 }
1495
1496 for(i = 0; i < vm_page_array_size; i++) {
1497 m = &vm_page_array[i];
1498 if (m->wire_count || m->hold_count || m->busy ||
1499 (m->flags & PG_BUSY))
1500 continue;
1501 pmap_remove_all(m);
1502 }
1503 pmap_pagedaemon_waken = 0;
1504 }
1505
1506
1507 /*
1508 * If it is the first entry on the list, it is actually
1509 * in the header and we must copy the following entry up
1510 * to the header. Otherwise we must search the list for
1511 * the entry. In either case we free the now unused entry.
1512 */
1513
1514 static int
1515 pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va)
1516 {
1517 pv_entry_t pv;
1518 int rtval;
1519 int s;
1520
1521 s = splvm();
1522 if (m->md.pv_list_count < pmap->pm_stats.resident_count) {
1523 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
1524 if (pmap == pv->pv_pmap && va == pv->pv_va)
1525 break;
1526 }
1527 } else {
1528 TAILQ_FOREACH(pv, &pmap->pm_pvlist, pv_plist) {
1529 if (va == pv->pv_va)
1530 break;
1531 }
1532 }
1533
1534 rtval = 0;
1535 if (pv) {
1536 rtval = pmap_unuse_pt(pmap, va, pv->pv_ptem);
1537 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1538 m->md.pv_list_count--;
1539 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
1540 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1541
1542 TAILQ_REMOVE(&pmap->pm_pvlist, pv, pv_plist);
1543 free_pv_entry(pv);
1544 }
1545
1546 splx(s);
1547 return rtval;
1548 }
1549
1550 /*
1551 * Create a pv entry for page at pa for
1552 * (pmap, va).
1553 */
1554 static void
1555 pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t mpte, vm_page_t m)
1556 {
1557 int s;
1558 pv_entry_t pv;
1559
1560 s = splvm();
1561 pv = get_pv_entry();
1562 pv->pv_va = va;
1563 pv->pv_pmap = pmap;
1564 pv->pv_ptem = mpte;
1565
1566 TAILQ_INSERT_TAIL(&pmap->pm_pvlist, pv, pv_plist);
1567 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
1568 m->md.pv_list_count++;
1569
1570 splx(s);
1571 }
1572
1573 /*
1574 * pmap_remove_pte: do the things to unmap a page in a process
1575 */
1576 static int
1577 pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va)
1578 {
1579 pt_entry_t oldpte;
1580 vm_page_t m;
1581
1582 oldpte = pte_store(ptq, 0);
1583 if (oldpte & PG_W)
1584 pmap->pm_stats.wired_count -= 1;
1585 /*
1586 * Machines that don't support invlpg, also don't support
1587 * PG_G.
1588 */
1589 if (oldpte & PG_G)
1590 invlpg(va);
1591 pmap->pm_stats.resident_count -= 1;
1592 if (oldpte & PG_MANAGED) {
1593 m = PHYS_TO_VM_PAGE(oldpte);
1594 if (oldpte & PG_M) {
1595 #if defined(PMAP_DIAGNOSTIC)
1596 if (pmap_nw_modified(oldpte)) {
1597 printf(
1598 "pmap_remove: modified page not writable: va: 0x%x, pte: 0x%x\n",
1599 va, oldpte);
1600 }
1601 #endif
1602 if (pmap_track_modified(va))
1603 vm_page_dirty(m);
1604 }
1605 if (oldpte & PG_A)
1606 vm_page_flag_set(m, PG_REFERENCED);
1607 return pmap_remove_entry(pmap, m, va);
1608 } else {
1609 return pmap_unuse_pt(pmap, va, NULL);
1610 }
1611
1612 return 0;
1613 }
1614
1615 /*
1616 * Remove a single page from a process address space
1617 */
1618 static void
1619 pmap_remove_page(pmap_t pmap, vm_offset_t va)
1620 {
1621 pt_entry_t *pte;
1622
1623 /*
1624 * get a local va for mappings for this pmap.
1625 */
1626 pte = pmap_pte(pmap, va);
1627 if (!pte)
1628 return;
1629 if (*pte & PG_V) {
1630 (void) pmap_remove_pte(pmap, pte, va);
1631 pmap_TLB_invalidate(pmap, va);
1632 }
1633 return;
1634 }
1635
1636 /*
1637 * Remove the given range of addresses from the specified map.
1638 *
1639 * It is assumed that the start and end are properly
1640 * rounded to the page size.
1641 */
1642 void
1643 pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
1644 {
1645 pt_entry_t *pte;
1646 pd_entry_t pde;
1647 vm_offset_t nva;
1648 int anyvalid;
1649
1650 if (pmap == NULL)
1651 return;
1652
1653 if (pmap->pm_stats.resident_count == 0)
1654 return;
1655
1656 /*
1657 * special handling of removing one page. a very
1658 * common operation and easy to short circuit some
1659 * code.
1660 */
1661 if (sva + PAGE_SIZE == eva &&
1662 (pmap->pm_pdir[sva >> PDRSHIFT] & PG_PS) == 0) {
1663 pmap_remove_page(pmap, sva);
1664 return;
1665 }
1666
1667 anyvalid = 0;
1668
1669 /*
1670 * Get a local virtual address for the mappings that are being
1671 * worked with.
1672 */
1673
1674 for (; sva < eva; sva = nva) {
1675 unsigned pdirindex;
1676
1677 /*
1678 * Calculate address for next page table.
1679 */
1680 nva = (sva + NBPDR) & ~PDRMASK;
1681
1682 if (pmap->pm_stats.resident_count == 0)
1683 break;
1684
1685 pdirindex = sva >> PDRSHIFT;
1686 if (((pde = pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1687 pmap->pm_pdir[pdirindex] = 0;
1688 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1689 anyvalid++;
1690 continue;
1691 }
1692
1693 /*
1694 * Weed out invalid mappings. Note: we assume that the page
1695 * directory table is always allocated, and in kernel virtual.
1696 */
1697 if ((pde & PG_V) == 0)
1698 continue;
1699
1700 /*
1701 * Limit our scan to either the end of the va represented
1702 * by the current page table page, or to the end of the
1703 * range being removed.
1704 */
1705 if (nva > eva)
1706 nva = eva;
1707
1708 pte = pmap_pte(pmap, sva);
1709 for (; sva < nva; sva += PAGE_SIZE, pte++) {
1710 if ((*pte & PG_V) == 0)
1711 continue;
1712
1713 anyvalid++;
1714 if (pmap_remove_pte(pmap, pte, sva))
1715 break;
1716 }
1717 }
1718
1719 if (anyvalid)
1720 pmap_TLB_invalidate_all(pmap);
1721 }
1722
1723 /*
1724 * Routine: pmap_remove_all
1725 * Function:
1726 * Removes this physical page from
1727 * all physical maps in which it resides.
1728 * Reflects back modify bits to the pager.
1729 *
1730 * Notes:
1731 * Original versions of this routine were very
1732 * inefficient because they iteratively called
1733 * pmap_remove (slow...)
1734 */
1735
1736 static void
1737 pmap_remove_all(vm_page_t m)
1738 {
1739 pv_entry_t pv;
1740 pt_entry_t *pte, tpte;
1741 int s;
1742
1743 #if defined(PMAP_DIAGNOSTIC)
1744 /*
1745 * XXX this makes pmap_page_protect(NONE) illegal for non-managed
1746 * pages!
1747 */
1748 if (!pmap_initialized || (m->flags & PG_FICTITIOUS)) {
1749 panic("pmap_page_protect: illegal for unmanaged page, va: 0x%x", VM_PAGE_TO_PHYS(m));
1750 }
1751 #endif
1752
1753 s = splvm();
1754 while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
1755 pv->pv_pmap->pm_stats.resident_count--;
1756
1757 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
1758
1759 tpte = pte_store(pte, 0);
1760 if (tpte & PG_W)
1761 pv->pv_pmap->pm_stats.wired_count--;
1762
1763 if (tpte & PG_A)
1764 vm_page_flag_set(m, PG_REFERENCED);
1765
1766 /*
1767 * Update the vm_page_t clean and reference bits.
1768 */
1769 if (tpte & PG_M) {
1770 #if defined(PMAP_DIAGNOSTIC)
1771 if (pmap_nw_modified(tpte)) {
1772 printf(
1773 "pmap_remove_all: modified page not writable: va: 0x%x, pte: 0x%x\n",
1774 pv->pv_va, tpte);
1775 }
1776 #endif
1777 if (pmap_track_modified(pv->pv_va))
1778 vm_page_dirty(m);
1779 }
1780 pmap_TLB_invalidate(pv->pv_pmap, pv->pv_va);
1781
1782 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
1783 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
1784 m->md.pv_list_count--;
1785 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
1786 free_pv_entry(pv);
1787 }
1788
1789 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
1790
1791 splx(s);
1792 }
1793
1794 /*
1795 * Set the physical protection on the
1796 * specified range of this map as requested.
1797 */
1798 void
1799 pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1800 {
1801 pt_entry_t *pte;
1802 pd_entry_t pde;
1803 vm_offset_t nva;
1804 int anychanged;
1805
1806 if (pmap == NULL)
1807 return;
1808
1809 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1810 pmap_remove(pmap, sva, eva);
1811 return;
1812 }
1813
1814 if (prot & VM_PROT_WRITE)
1815 return;
1816
1817 anychanged = 0;
1818
1819 for (; sva < eva; sva = nva) {
1820
1821 unsigned pdirindex;
1822
1823 nva = (sva + NBPDR) & ~PDRMASK;
1824
1825 pdirindex = sva >> PDRSHIFT;
1826 if (((pde = pmap->pm_pdir[pdirindex]) & PG_PS) != 0) {
1827 pmap->pm_pdir[pdirindex] &= ~(PG_M|PG_RW);
1828 pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE;
1829 anychanged++;
1830 continue;
1831 }
1832
1833 /*
1834 * Weed out invalid mappings. Note: we assume that the page
1835 * directory table is always allocated, and in kernel virtual.
1836 */
1837 if ((pde & PG_V) == 0)
1838 continue;
1839
1840 if (nva > eva)
1841 nva = eva;
1842
1843 pte = pmap_pte(pmap, sva);
1844 for (; sva < nva; sva += PAGE_SIZE, pte++) {
1845 pt_entry_t pbits;
1846 vm_page_t m;
1847
1848 pbits = *pte;
1849
1850 if (pbits & PG_MANAGED) {
1851 m = NULL;
1852 if (pbits & PG_A) {
1853 m = PHYS_TO_VM_PAGE(pbits);
1854 vm_page_flag_set(m, PG_REFERENCED);
1855 pbits &= ~PG_A;
1856 }
1857 if (pbits & PG_M) {
1858 if (pmap_track_modified(sva)) {
1859 if (m == NULL)
1860 m = PHYS_TO_VM_PAGE(pbits);
1861 vm_page_dirty(m);
1862 pbits &= ~PG_M;
1863 }
1864 }
1865 }
1866
1867 pbits &= ~PG_RW;
1868
1869 if (pbits != *pte) {
1870 *pte = pbits;
1871 anychanged = 1;
1872 }
1873 }
1874 }
1875 if (anychanged)
1876 pmap_TLB_invalidate_all(pmap);
1877 }
1878
1879 /*
1880 * Insert the given physical page (p) at
1881 * the specified virtual address (v) in the
1882 * target physical map with the protection requested.
1883 *
1884 * If specified, the page will be wired down, meaning
1885 * that the related pte can not be reclaimed.
1886 *
1887 * NB: This is the only routine which MAY NOT lazy-evaluate
1888 * or lose information. That is, this routine must actually
1889 * insert this page into the given map NOW.
1890 */
1891 void
1892 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1893 boolean_t wired)
1894 {
1895 vm_paddr_t pa, opa;
1896 pt_entry_t *pte, origpte, newpte;
1897 vm_page_t mpte;
1898
1899 if (pmap == NULL)
1900 return;
1901
1902 va &= PG_FRAME;
1903 #ifdef PMAP_DIAGNOSTIC
1904 if (va > VM_MAX_KERNEL_ADDRESS)
1905 panic("pmap_enter: toobig");
1906 if ((va >= UPT_MIN_ADDRESS) && (va < UPT_MAX_ADDRESS))
1907 panic("pmap_enter: invalid to pmap_enter page table pages (va: 0x%x)", va);
1908 #endif
1909
1910 mpte = NULL;
1911 /*
1912 * In the case that a page table page is not
1913 * resident, we are creating it here.
1914 */
1915 if (va < UPT_MIN_ADDRESS) {
1916 mpte = pmap_allocpte(pmap, va);
1917 }
1918 #if 0 && defined(PMAP_DIAGNOSTIC)
1919 else {
1920 pd_entry_t *pdeaddr = pmap_pde(pmap, va);
1921 if (((origpte = *pdeaddr) & PG_V) == 0) {
1922 panic("pmap_enter: invalid kernel page table page(0), pdir=%p, pde=%p, va=%p\n",
1923 pmap->pm_pdir[PTDPTDI], origpte, va);
1924 }
1925 if (smp_active) {
1926 pdeaddr = IdlePTDS[cpuid];
1927 if (((newpte = pdeaddr[va >> PDRSHIFT]) & PG_V) == 0) {
1928 if (my_idlePTD != vtophys(pdeaddr))
1929 printf("pde mismatch: %x, %x\n", my_idlePTD, pdeaddr);
1930 printf("cpuid: %d, pdeaddr: 0x%x\n", cpuid, pdeaddr);
1931 panic("pmap_enter: invalid kernel page table page(1), pdir=%p, npde=%p, pde=%p, va=%p\n",
1932 pmap->pm_pdir[PTDPTDI], newpte, origpte, va);
1933 }
1934 }
1935 }
1936 #endif
1937
1938 pte = pmap_pte(pmap, va);
1939
1940 /*
1941 * Page Directory table entry not valid, we need a new PT page
1942 */
1943 if (pte == NULL) {
1944 panic("pmap_enter: invalid page directory pdir=%#llx, va=%#x\n",
1945 (u_int64_t)pmap->pm_pdir[PTDPTDI], va);
1946 }
1947
1948 pa = VM_PAGE_TO_PHYS(m) & PG_FRAME;
1949 origpte = *pte;
1950 opa = origpte & PG_FRAME;
1951
1952 if (origpte & PG_PS)
1953 panic("pmap_enter: attempted pmap_enter on 4MB page");
1954
1955 /*
1956 * Mapping has not changed, must be protection or wiring change.
1957 */
1958 if ((origpte & PG_V) && (opa == pa)) {
1959 /*
1960 * Wiring change, just update stats. We don't worry about
1961 * wiring PT pages as they remain resident as long as there
1962 * are valid mappings in them. Hence, if a user page is wired,
1963 * the PT page will be also.
1964 */
1965 if (wired && ((origpte & PG_W) == 0))
1966 pmap->pm_stats.wired_count++;
1967 else if (!wired && (origpte & PG_W))
1968 pmap->pm_stats.wired_count--;
1969
1970 #if defined(PMAP_DIAGNOSTIC)
1971 if (pmap_nw_modified((pt_entry_t) origpte)) {
1972 printf(
1973 "pmap_enter: modified page not writable: va: 0x%x, pte: 0x%x\n",
1974 va, origpte);
1975 }
1976 #endif
1977
1978 /*
1979 * Remove extra pte reference
1980 */
1981 if (mpte)
1982 mpte->hold_count--;
1983
1984 if ((prot & VM_PROT_WRITE)) {
1985 if ((origpte & PG_RW) == 0) {
1986 *pte |= PG_RW;
1987 pmap_TLB_invalidate(pmap, va);
1988 }
1989 return;
1990 }
1991
1992 /*
1993 * We might be turning off write access to the page,
1994 * so we go ahead and sense modify status.
1995 */
1996 if (origpte & PG_MANAGED) {
1997 if ((origpte & PG_M) && pmap_track_modified(va)) {
1998 vm_page_t om;
1999 om = PHYS_TO_VM_PAGE(opa);
2000 vm_page_dirty(om);
2001 }
2002 pa |= PG_MANAGED;
2003 }
2004 goto validate;
2005 }
2006 /*
2007 * Mapping has changed, invalidate old range and fall through to
2008 * handle validating new mapping.
2009 */
2010 if ((origpte & PG_V)) {
2011 int err;
2012 err = pmap_remove_pte(pmap, pte, va);
2013 if (err)
2014 panic("pmap_enter: pte vanished, va: 0x%x", va);
2015 }
2016
2017 /*
2018 * Enter on the PV list if part of our managed memory. Note that we
2019 * raise IPL while manipulating pv_table since pmap_enter can be
2020 * called at interrupt time.
2021 */
2022 if (pmap_initialized && !(m->flags & (PG_FICTITIOUS|PG_UNMANAGED))) {
2023 pmap_insert_entry(pmap, va, mpte, m);
2024 pa |= PG_MANAGED;
2025 }
2026
2027 /*
2028 * Increment counters
2029 */
2030 pmap->pm_stats.resident_count++;
2031 if (wired)
2032 pmap->pm_stats.wired_count++;
2033
2034 validate:
2035 /*
2036 * Now validate mapping with desired protection/wiring.
2037 */
2038 newpte = pa | pte_prot(pmap, prot) | PG_V;
2039
2040 if (wired)
2041 newpte |= PG_W;
2042 if (va < UPT_MIN_ADDRESS)
2043 newpte |= PG_U;
2044 if (pmap == kernel_pmap)
2045 newpte |= pgeflag;
2046
2047 /*
2048 * if the mapping or permission bits are different, we need
2049 * to update the pte.
2050 */
2051 if ((origpte & ~(PG_M|PG_A)) != newpte) {
2052 *pte = newpte | PG_A;
2053 pmap_TLB_invalidate(pmap, va);
2054 }
2055 }
2056
2057 /*
2058 * this code makes some *MAJOR* assumptions:
2059 * 1. Current pmap & pmap exists.
2060 * 2. Not wired.
2061 * 3. Read access.
2062 * 4. No page table pages.
2063 * 5. Tlbflush is deferred to calling procedure.
2064 * 6. Page IS managed.
2065 * but is *MUCH* faster than pmap_enter...
2066 */
2067
2068 static vm_page_t
2069 pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_page_t mpte)
2070 {
2071 pt_entry_t *pte;
2072 vm_paddr_t pa;
2073
2074 /*
2075 * In the case that a page table page is not
2076 * resident, we are creating it here.
2077 */
2078 if (va < UPT_MIN_ADDRESS) {
2079 unsigned ptepindex;
2080 pd_entry_t pde;
2081
2082 /*
2083 * Calculate pagetable page index
2084 */
2085 ptepindex = va >> PDRSHIFT;
2086 if (mpte && (mpte->pindex == ptepindex)) {
2087 mpte->hold_count++;
2088 } else {
2089 retry:
2090 /*
2091 * Get the page directory entry
2092 */
2093 pde = pmap->pm_pdir[ptepindex];
2094
2095 /*
2096 * If the page table page is mapped, we just increment
2097 * the hold count, and activate it.
2098 */
2099 if ((pde & PG_V)) {
2100 if (pde & PG_PS)
2101 panic("pmap_enter_quick: unexpected mapping into 4MB page");
2102 if (pmap->pm_ptphint &&
2103 (pmap->pm_ptphint->pindex == ptepindex)) {
2104 mpte = pmap->pm_ptphint;
2105 } else {
2106 mpte = pmap_page_lookup(pmap->pm_pteobj, ptepindex);
2107 pmap->pm_ptphint = mpte;
2108 }
2109 if (mpte == NULL)
2110 goto retry;
2111 mpte->hold_count++;
2112 } else {
2113 mpte = _pmap_allocpte(pmap, ptepindex);
2114 if (mpte == NULL)
2115 goto retry;
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.
2126 * But that isn't as quick as vtopte.
2127 */
2128 pte = vtopte(va);
2129 if (*pte) {
2130 if (mpte)
2131 pmap_unwire_pte_hold(pmap, mpte);
2132 return 0;
2133 }
2134
2135 /*
2136 * Enter on the PV list if part of our managed memory. Note that we
2137 * raise IPL while manipulating pv_table since pmap_enter can be
2138 * called at interrupt time.
2139 */
2140 if ((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0)
2141 pmap_insert_entry(pmap, va, mpte, m);
2142
2143 /*
2144 * Increment counters
2145 */
2146 pmap->pm_stats.resident_count++;
2147
2148 pa = VM_PAGE_TO_PHYS(m);
2149
2150 /*
2151 * Now validate mapping with RO protection
2152 */
2153 if (m->flags & (PG_FICTITIOUS|PG_UNMANAGED))
2154 *pte = pa | PG_V | PG_U;
2155 else
2156 *pte = pa | PG_V | PG_U | PG_MANAGED;
2157
2158 return mpte;
2159 }
2160
2161 /*
2162 * Make a temporary mapping for a physical address. This is only intended
2163 * to be used for panic dumps.
2164 */
2165 void *
2166 pmap_kenter_temporary(vm_paddr_t pa, int i)
2167 {
2168 pmap_kenter((vm_offset_t)crashdumpmap + (i * PAGE_SIZE), pa);
2169 return ((void *)crashdumpmap);
2170 }
2171
2172 #define MAX_INIT_PT (96)
2173 /*
2174 * pmap_object_init_pt preloads the ptes for a given object
2175 * into the specified pmap. This eliminates the blast of soft
2176 * faults on process startup and immediately after an mmap.
2177 */
2178 void
2179 pmap_object_init_pt(pmap, addr, prot, object, pindex, size, limit)
2180 pmap_t pmap;
2181 vm_offset_t addr;
2182 vm_prot_t prot;
2183 vm_object_t object;
2184 vm_pindex_t pindex;
2185 vm_size_t size;
2186 int limit;
2187 {
2188 vm_offset_t tmpidx;
2189 int psize;
2190 vm_page_t p, mpte;
2191 int objpgs;
2192
2193 if ((prot & VM_PROT_READ) == 0 || pmap == NULL || object == NULL)
2194 return;
2195
2196 /*
2197 * This code maps large physical mmap regions into the
2198 * processor address space. Note that some shortcuts
2199 * are taken, but the code works.
2200 */
2201 if (pseflag &&
2202 (object->type == OBJT_DEVICE) &&
2203 ((addr & (NBPDR - 1)) == 0) &&
2204 ((size & (NBPDR - 1)) == 0) ) {
2205 int i;
2206 vm_page_t m[1];
2207 unsigned int ptepindex;
2208 int npdes;
2209 pd_entry_t ptepa;
2210
2211 if (pmap->pm_pdir[ptepindex = (addr >> PDRSHIFT)] & PG_V)
2212 return;
2213
2214 retry:
2215 p = vm_page_lookup(object, pindex);
2216 if (p && vm_page_sleep_busy(p, FALSE, "init4p"))
2217 goto retry;
2218
2219 if (p == NULL) {
2220 p = vm_page_alloc(object, pindex, VM_ALLOC_NORMAL);
2221 if (p == NULL)
2222 return;
2223 m[0] = p;
2224
2225 if (vm_pager_get_pages(object, m, 1, 0) != VM_PAGER_OK) {
2226 vm_page_free(p);
2227 return;
2228 }
2229
2230 p = vm_page_lookup(object, pindex);
2231 vm_page_wakeup(p);
2232 }
2233
2234 ptepa = VM_PAGE_TO_PHYS(p);
2235 if (ptepa & (NBPDR - 1)) {
2236 return;
2237 }
2238
2239 p->valid = VM_PAGE_BITS_ALL;
2240
2241 pmap->pm_stats.resident_count += size >> PAGE_SHIFT;
2242 npdes = size >> PDRSHIFT;
2243 for(i = 0; i < npdes; i++) {
2244 pmap->pm_pdir[ptepindex] =
2245 ptepa | PG_U | PG_RW | PG_V | PG_PS;
2246 ptepa += NBPDR;
2247 ptepindex += 1;
2248 }
2249 vm_page_flag_set(p, PG_MAPPED);
2250 invltlb();
2251 return;
2252 }
2253
2254 psize = i386_btop(size);
2255
2256 if ((object->type != OBJT_VNODE) ||
2257 ((limit & MAP_PREFAULT_PARTIAL) && (psize > MAX_INIT_PT) &&
2258 (object->resident_page_count > MAX_INIT_PT))) {
2259 return;
2260 }
2261
2262 if (psize + pindex > object->size) {
2263 if (object->size < pindex)
2264 return;
2265 psize = object->size - pindex;
2266 }
2267
2268 mpte = NULL;
2269 /*
2270 * if we are processing a major portion of the object, then scan the
2271 * entire thing.
2272 */
2273 if (psize > (object->resident_page_count >> 2)) {
2274 objpgs = psize;
2275
2276 for (p = TAILQ_FIRST(&object->memq);
2277 ((objpgs > 0) && (p != NULL));
2278 p = TAILQ_NEXT(p, listq)) {
2279
2280 tmpidx = p->pindex;
2281 if (tmpidx < pindex) {
2282 continue;
2283 }
2284 tmpidx -= pindex;
2285 if (tmpidx >= psize) {
2286 continue;
2287 }
2288 /*
2289 * don't allow an madvise to blow away our really
2290 * free pages allocating pv entries.
2291 */
2292 if ((limit & MAP_PREFAULT_MADVISE) &&
2293 cnt.v_free_count < cnt.v_free_reserved) {
2294 break;
2295 }
2296 if ((p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL &&
2297 (p->busy == 0) &&
2298 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2299 if ((p->queue - p->pc) == PQ_CACHE)
2300 vm_page_deactivate(p);
2301 vm_page_busy(p);
2302 mpte = pmap_enter_quick(pmap,
2303 addr + i386_ptob(tmpidx), p, mpte);
2304 vm_page_flag_set(p, PG_MAPPED);
2305 vm_page_wakeup(p);
2306 }
2307 objpgs -= 1;
2308 }
2309 } else {
2310 /*
2311 * else lookup the pages one-by-one.
2312 */
2313 for (tmpidx = 0; tmpidx < psize; tmpidx += 1) {
2314 /*
2315 * don't allow an madvise to blow away our really
2316 * free pages allocating pv entries.
2317 */
2318 if ((limit & MAP_PREFAULT_MADVISE) &&
2319 cnt.v_free_count < cnt.v_free_reserved) {
2320 break;
2321 }
2322 p = vm_page_lookup(object, tmpidx + pindex);
2323 if (p &&
2324 (p->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL &&
2325 (p->busy == 0) &&
2326 (p->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2327 if ((p->queue - p->pc) == PQ_CACHE)
2328 vm_page_deactivate(p);
2329 vm_page_busy(p);
2330 mpte = pmap_enter_quick(pmap,
2331 addr + i386_ptob(tmpidx), p, mpte);
2332 vm_page_flag_set(p, PG_MAPPED);
2333 vm_page_wakeup(p);
2334 }
2335 }
2336 }
2337 }
2338
2339 /*
2340 * pmap_prefault provides a quick way of clustering
2341 * pagefaults into a processes address space. It is a "cousin"
2342 * of pmap_object_init_pt, except it runs at page fault time instead
2343 * of mmap time.
2344 */
2345 #define PFBAK 4
2346 #define PFFOR 4
2347 #define PAGEORDER_SIZE (PFBAK+PFFOR)
2348
2349 static int pmap_prefault_pageorder[] = {
2350 -PAGE_SIZE, PAGE_SIZE,
2351 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
2352 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
2353 -4 * PAGE_SIZE, 4 * PAGE_SIZE
2354 };
2355
2356 void
2357 pmap_prefault(pmap, addra, entry)
2358 pmap_t pmap;
2359 vm_offset_t addra;
2360 vm_map_entry_t entry;
2361 {
2362 int i;
2363 vm_offset_t starta;
2364 vm_offset_t addr;
2365 vm_pindex_t pindex;
2366 vm_page_t m, mpte;
2367 vm_object_t object;
2368
2369 if (!curproc || (pmap != vmspace_pmap(curproc->p_vmspace)))
2370 return;
2371
2372 object = entry->object.vm_object;
2373
2374 starta = addra - PFBAK * PAGE_SIZE;
2375 if (starta < entry->start) {
2376 starta = entry->start;
2377 } else if (starta > addra) {
2378 starta = 0;
2379 }
2380
2381 mpte = NULL;
2382 for (i = 0; i < PAGEORDER_SIZE; i++) {
2383 vm_object_t lobject;
2384 pt_entry_t *pte;
2385
2386 addr = addra + pmap_prefault_pageorder[i];
2387 if (addr > addra + (PFFOR * PAGE_SIZE))
2388 addr = 0;
2389
2390 if (addr < starta || addr >= entry->end)
2391 continue;
2392
2393 if ((*pmap_pde(pmap, addr) & PG_V) == 0)
2394 continue;
2395
2396 pte = vtopte(addr);
2397 if ((*pte & PG_V))
2398 continue;
2399
2400 pindex = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
2401 lobject = object;
2402 for (m = vm_page_lookup(lobject, pindex);
2403 (!m && (lobject->type == OBJT_DEFAULT) && (lobject->backing_object));
2404 lobject = lobject->backing_object) {
2405 if (lobject->backing_object_offset & PAGE_MASK)
2406 break;
2407 pindex += (lobject->backing_object_offset >> PAGE_SHIFT);
2408 m = vm_page_lookup(lobject->backing_object, pindex);
2409 }
2410
2411 /*
2412 * give-up when a page is not in memory
2413 */
2414 if (m == NULL)
2415 break;
2416
2417 if ((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL &&
2418 (m->busy == 0) &&
2419 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2420
2421 if ((m->queue - m->pc) == PQ_CACHE) {
2422 vm_page_deactivate(m);
2423 }
2424 vm_page_busy(m);
2425 mpte = pmap_enter_quick(pmap, addr, m, mpte);
2426 vm_page_flag_set(m, PG_MAPPED);
2427 vm_page_wakeup(m);
2428 }
2429 }
2430 }
2431
2432 /*
2433 * Routine: pmap_change_wiring
2434 * Function: Change the wiring attribute for a map/virtual-address
2435 * pair.
2436 * In/out conditions:
2437 * The mapping must already exist in the pmap.
2438 */
2439 void
2440 pmap_change_wiring(pmap_t pmap, vm_offset_t va, boolean_t wired)
2441 {
2442 pt_entry_t *pte;
2443
2444 if (pmap == NULL)
2445 return;
2446
2447 pte = pmap_pte(pmap, va);
2448
2449 if (wired && !pmap_pte_w(pte))
2450 pmap->pm_stats.wired_count++;
2451 else if (!wired && pmap_pte_w(pte))
2452 pmap->pm_stats.wired_count--;
2453
2454 /*
2455 * Wiring is not a hardware characteristic so there is no need to
2456 * invalidate TLB.
2457 */
2458 pmap_pte_set_w(pte, wired);
2459 }
2460
2461
2462
2463 /*
2464 * Copy the range specified by src_addr/len
2465 * from the source map to the range dst_addr/len
2466 * in the destination map.
2467 *
2468 * This routine is only advisory and need not do anything.
2469 */
2470
2471 void
2472 pmap_copy(dst_pmap, src_pmap, dst_addr, len, src_addr)
2473 pmap_t dst_pmap, src_pmap;
2474 vm_offset_t dst_addr;
2475 vm_size_t len;
2476 vm_offset_t src_addr;
2477 {
2478 vm_offset_t addr;
2479 vm_offset_t end_addr = src_addr + len;
2480 vm_offset_t pdnxt;
2481 vm_paddr_t src_frame;
2482 vm_page_t m;
2483
2484 if (dst_addr != src_addr)
2485 return;
2486
2487 src_frame = src_pmap->pm_pdir[PTDPTDI] & PG_FRAME;
2488 if (src_frame != (PTDpde[0] & PG_FRAME)) {
2489 return;
2490 }
2491
2492 for(addr = src_addr; addr < end_addr; addr = pdnxt) {
2493 pt_entry_t *src_pte, *dst_pte;
2494 vm_page_t dstmpte, srcmpte;
2495 pd_entry_t srcptepaddr;
2496 unsigned ptepindex;
2497
2498 if (addr >= UPT_MIN_ADDRESS)
2499 panic("pmap_copy: invalid to pmap_copy page tables\n");
2500
2501 /*
2502 * Don't let optional prefaulting of pages make us go
2503 * way below the low water mark of free pages or way
2504 * above high water mark of used pv entries.
2505 */
2506 if (cnt.v_free_count < cnt.v_free_reserved ||
2507 pv_entry_count > pv_entry_high_water)
2508 break;
2509
2510 pdnxt = (addr + NBPDR) & ~(NBPDR - 1);
2511 ptepindex = addr >> PDRSHIFT;
2512
2513 srcptepaddr = src_pmap->pm_pdir[ptepindex];
2514 if (srcptepaddr == 0)
2515 continue;
2516
2517 if (srcptepaddr & PG_PS) {
2518 if (dst_pmap->pm_pdir[ptepindex] == 0) {
2519 dst_pmap->pm_pdir[ptepindex] = srcptepaddr;
2520 dst_pmap->pm_stats.resident_count += NPDEPG;
2521 }
2522 continue;
2523 }
2524
2525 srcmpte = vm_page_lookup(src_pmap->pm_pteobj, ptepindex);
2526 if ((srcmpte == NULL) ||
2527 (srcmpte->hold_count == 0) || (srcmpte->flags & PG_BUSY))
2528 continue;
2529
2530 if (pdnxt > end_addr)
2531 pdnxt = end_addr;
2532
2533 src_pte = vtopte(addr);
2534 while (addr < pdnxt) {
2535 pt_entry_t ptetemp;
2536 ptetemp = *src_pte;
2537 /*
2538 * we only virtual copy managed pages
2539 */
2540 if ((ptetemp & PG_MANAGED) != 0) {
2541 /*
2542 * We have to check after allocpte for the
2543 * pte still being around... allocpte can
2544 * block.
2545 */
2546 dstmpte = pmap_allocpte(dst_pmap, addr);
2547 dst_pte = pmap_pte(dst_pmap, addr);
2548 if ((*dst_pte == 0) && (ptetemp = *src_pte)) {
2549 /*
2550 * Clear the modified and
2551 * accessed (referenced) bits
2552 * during the copy.
2553 */
2554 m = PHYS_TO_VM_PAGE(ptetemp);
2555 *dst_pte = ptetemp & ~(PG_M | PG_A);
2556 dst_pmap->pm_stats.resident_count++;
2557 pmap_insert_entry(dst_pmap, addr,
2558 dstmpte, m);
2559 } else {
2560 pmap_unwire_pte_hold(dst_pmap, dstmpte);
2561 }
2562 if (dstmpte->hold_count >= srcmpte->hold_count)
2563 break;
2564 }
2565 addr += PAGE_SIZE;
2566 src_pte++;
2567 }
2568 }
2569 }
2570
2571 /*
2572 * Routine: pmap_kernel
2573 * Function:
2574 * Returns the physical map handle for the kernel.
2575 */
2576 pmap_t
2577 pmap_kernel()
2578 {
2579 return (kernel_pmap);
2580 }
2581
2582 /*
2583 * pmap_zero_page zeros the specified hardware page by mapping
2584 * the page into KVM and using bzero to clear its contents.
2585 */
2586 void
2587 pmap_zero_page(vm_paddr_t phys)
2588 {
2589 if (*CMAP3)
2590 panic("pmap_zero_page: CMAP3 busy");
2591
2592 *CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2593 #ifdef SMP
2594 cpu_invlpg(CADDR3);
2595 #else
2596 invltlb_1pg((vm_offset_t)CADDR3);
2597 #endif
2598
2599 #if defined(I686_CPU)
2600 if (cpu_class == CPUCLASS_686)
2601 i686_pagezero(CADDR3);
2602 else
2603 #endif
2604 bzero(CADDR3, PAGE_SIZE);
2605 *CMAP3 = 0;
2606 }
2607
2608 /*
2609 * pmap_zero_page_area zeros the specified hardware page by mapping
2610 * the page into KVM and using bzero to clear its contents.
2611 *
2612 * off and size may not cover an area beyond a single hardware page.
2613 */
2614 void
2615 pmap_zero_page_area(vm_paddr_t phys, int off, int size)
2616 {
2617 if (*CMAP3)
2618 panic("pmap_zero_page: CMAP3 busy");
2619
2620 *CMAP3 = PG_V | PG_RW | (phys & PG_FRAME) | PG_A | PG_M;
2621 #ifdef SMP
2622 cpu_invlpg(CADDR3);
2623 #else
2624 invltlb_1pg((vm_offset_t)CADDR3);
2625 #endif
2626
2627 #if defined(I686_CPU)
2628 if (cpu_class == CPUCLASS_686 && off == 0 && size == PAGE_SIZE)
2629 i686_pagezero(CADDR3);
2630 else
2631 #endif
2632 bzero(CADDR3 + off, size);
2633 *CMAP3 = 0;
2634 }
2635
2636 /*
2637 * pmap_copy_page copies the specified (machine independent)
2638 * page by mapping the page into virtual memory and using
2639 * bcopy to copy the page, one machine dependent page at a
2640 * time.
2641 */
2642 void
2643 pmap_copy_page(vm_paddr_t src, vm_paddr_t dst)
2644 {
2645 if (*CMAP1)
2646 panic("pmap_copy_page: CMAP1 busy");
2647 if (*CMAP2)
2648 panic("pmap_copy_page: CMAP2 busy");
2649
2650 *CMAP1 = PG_V | (src & PG_FRAME) | PG_A;
2651 *CMAP2 = PG_V | PG_RW | (dst & PG_FRAME) | PG_A | PG_M;
2652
2653 #ifdef SMP
2654 cpu_invlpg(CADDR1);
2655 cpu_invlpg(CADDR2);
2656 #else
2657 invltlb_1pg((vm_offset_t)CADDR1);
2658 invltlb_1pg((vm_offset_t)CADDR2);
2659 #endif
2660
2661 bcopy(CADDR1, CADDR2, PAGE_SIZE);
2662
2663 *CMAP1 = 0;
2664 *CMAP2 = 0;
2665 }
2666
2667
2668 /*
2669 * Routine: pmap_pageable
2670 * Function:
2671 * Make the specified pages (by pmap, offset)
2672 * pageable (or not) as requested.
2673 *
2674 * A page which is not pageable may not take
2675 * a fault; therefore, its page table entry
2676 * must remain valid for the duration.
2677 *
2678 * This routine is merely advisory; pmap_enter
2679 * will specify that these pages are to be wired
2680 * down (or not) as appropriate.
2681 */
2682 void
2683 pmap_pageable(pmap, sva, eva, pageable)
2684 pmap_t pmap;
2685 vm_offset_t sva, eva;
2686 boolean_t pageable;
2687 {
2688 }
2689
2690 /*
2691 * Returns true if the pmap's pv is one of the first
2692 * 16 pvs linked to from this page. This count may
2693 * be changed upwards or downwards in the future; it
2694 * is only necessary that true be returned for a small
2695 * subset of pmaps for proper page aging.
2696 */
2697 boolean_t
2698 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
2699 {
2700 pv_entry_t pv;
2701 int loops = 0;
2702 int s;
2703
2704 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2705 return FALSE;
2706
2707 s = splvm();
2708
2709 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2710 if (pv->pv_pmap == pmap) {
2711 splx(s);
2712 return TRUE;
2713 }
2714 loops++;
2715 if (loops >= 16)
2716 break;
2717 }
2718 splx(s);
2719 return (FALSE);
2720 }
2721
2722 #define PMAP_REMOVE_PAGES_CURPROC_ONLY
2723 /*
2724 * Remove all pages from specified address space
2725 * this aids process exit speeds. Also, this code
2726 * is special cased for current process only, but
2727 * can have the more generic (and slightly slower)
2728 * mode enabled. This is much faster than pmap_remove
2729 * in the case of running down an entire address space.
2730 */
2731 void
2732 pmap_remove_pages(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
2733 {
2734 pt_entry_t *pte, tpte;
2735 pv_entry_t pv, npv;
2736 int s;
2737 vm_page_t m;
2738
2739 #ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY
2740 if (!curproc || (pmap != vmspace_pmap(curproc->p_vmspace))) {
2741 printf("warning: pmap_remove_pages called with non-current pmap\n");
2742 return;
2743 }
2744 #endif
2745
2746 s = splvm();
2747 for(pv = TAILQ_FIRST(&pmap->pm_pvlist); pv; pv = npv) {
2748
2749 if (pv->pv_va >= eva || pv->pv_va < sva) {
2750 npv = TAILQ_NEXT(pv, pv_plist);
2751 continue;
2752 }
2753
2754 #ifdef PMAP_REMOVE_PAGES_CURPROC_ONLY
2755 pte = vtopte(pv->pv_va);
2756 #else
2757 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2758 #endif
2759 tpte = *pte;
2760
2761 /*
2762 * We cannot remove wired pages from a process' mapping at this time
2763 */
2764 if (tpte & PG_W) {
2765 npv = TAILQ_NEXT(pv, pv_plist);
2766 continue;
2767 }
2768 *pte = 0;
2769
2770 m = PHYS_TO_VM_PAGE(tpte);
2771
2772 KASSERT(m < &vm_page_array[vm_page_array_size],
2773 ("pmap_remove_pages: bad tpte %x", tpte));
2774
2775 pv->pv_pmap->pm_stats.resident_count--;
2776
2777 /*
2778 * Update the vm_page_t clean and reference bits.
2779 */
2780 if (tpte & PG_M) {
2781 vm_page_dirty(m);
2782 }
2783
2784
2785 npv = TAILQ_NEXT(pv, pv_plist);
2786 TAILQ_REMOVE(&pv->pv_pmap->pm_pvlist, pv, pv_plist);
2787
2788 m->md.pv_list_count--;
2789 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2790 if (TAILQ_FIRST(&m->md.pv_list) == NULL) {
2791 vm_page_flag_clear(m, PG_MAPPED | PG_WRITEABLE);
2792 }
2793
2794 pmap_unuse_pt(pv->pv_pmap, pv->pv_va, pv->pv_ptem);
2795 free_pv_entry(pv);
2796 }
2797 splx(s);
2798 pmap_TLB_invalidate_all(pmap);
2799 }
2800
2801 /*
2802 * pmap_testbit tests bits in pte's
2803 * note that the testbit/changebit routines are inline,
2804 * and a lot of things compile-time evaluate.
2805 */
2806 static boolean_t
2807 pmap_testbit(vm_page_t m, int bit)
2808 {
2809 pv_entry_t pv;
2810 pt_entry_t *pte;
2811 int s;
2812
2813 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2814 return FALSE;
2815
2816 if (TAILQ_FIRST(&m->md.pv_list) == NULL)
2817 return FALSE;
2818
2819 s = splvm();
2820
2821 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2822 /*
2823 * if the bit being tested is the modified bit, then
2824 * mark clean_map and ptes as never
2825 * modified.
2826 */
2827 if (bit & (PG_A|PG_M)) {
2828 if (!pmap_track_modified(pv->pv_va))
2829 continue;
2830 }
2831
2832 #if defined(PMAP_DIAGNOSTIC)
2833 if (!pv->pv_pmap) {
2834 printf("Null pmap (tb) at va: 0x%x\n", pv->pv_va);
2835 continue;
2836 }
2837 #endif
2838 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2839 if (*pte & bit) {
2840 splx(s);
2841 return TRUE;
2842 }
2843 }
2844 splx(s);
2845 return (FALSE);
2846 }
2847
2848 /*
2849 * this routine is used to modify bits in ptes
2850 */
2851 static __inline void
2852 pmap_changebit(vm_page_t m, int bit, boolean_t setem)
2853 {
2854 pv_entry_t pv;
2855 pt_entry_t *pte;
2856 int s;
2857
2858 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2859 return;
2860
2861 s = splvm();
2862
2863 /*
2864 * Loop over all current mappings setting/clearing as appropos If
2865 * setting RO do we need to clear the VAC?
2866 */
2867 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
2868 /*
2869 * don't write protect pager mappings
2870 */
2871 if (!setem && (bit == PG_RW)) {
2872 if (!pmap_track_modified(pv->pv_va))
2873 continue;
2874 }
2875
2876 #if defined(PMAP_DIAGNOSTIC)
2877 if (!pv->pv_pmap) {
2878 printf("Null pmap (cb) at va: 0x%x\n", pv->pv_va);
2879 continue;
2880 }
2881 #endif
2882
2883 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2884
2885 if (setem) {
2886 *pte |= bit;
2887 pmap_TLB_invalidate(pv->pv_pmap, pv->pv_va);
2888 } else {
2889 pt_entry_t pbits = *pte;
2890 if (pbits & bit) {
2891 if (bit == PG_RW) {
2892 if (pbits & PG_M) {
2893 vm_page_dirty(m);
2894 }
2895 *pte = pbits & ~(PG_M|PG_RW);
2896 } else {
2897 *pte = pbits & ~bit;
2898 }
2899 pmap_TLB_invalidate(pv->pv_pmap, pv->pv_va);
2900 }
2901 }
2902 }
2903 splx(s);
2904 }
2905
2906 /*
2907 * pmap_page_protect:
2908 *
2909 * Lower the permission for all mappings to a given page.
2910 */
2911 void
2912 pmap_page_protect(vm_page_t m, vm_prot_t prot)
2913 {
2914 if ((prot & VM_PROT_WRITE) == 0) {
2915 if (prot & (VM_PROT_READ | VM_PROT_EXECUTE)) {
2916 pmap_changebit(m, PG_RW, FALSE);
2917 } else {
2918 pmap_remove_all(m);
2919 }
2920 }
2921 }
2922
2923 vm_paddr_t
2924 pmap_phys_address(ppn)
2925 int ppn;
2926 {
2927 return (i386_ptob((vm_paddr_t)ppn));
2928 }
2929
2930 /*
2931 * pmap_ts_referenced:
2932 *
2933 * Return a count of reference bits for a page, clearing those bits.
2934 * It is not necessary for every reference bit to be cleared, but it
2935 * is necessary that 0 only be returned when there are truly no
2936 * reference bits set.
2937 *
2938 * XXX: The exact number of bits to check and clear is a matter that
2939 * should be tested and standardized at some point in the future for
2940 * optimal aging of shared pages.
2941 */
2942 int
2943 pmap_ts_referenced(vm_page_t m)
2944 {
2945 pv_entry_t pv, pvf, pvn;
2946 pt_entry_t *pte;
2947 int s;
2948 int rtval = 0;
2949
2950 if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
2951 return (rtval);
2952
2953 s = splvm();
2954
2955 if ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) {
2956
2957 pvf = pv;
2958
2959 do {
2960 pvn = TAILQ_NEXT(pv, pv_list);
2961
2962 TAILQ_REMOVE(&m->md.pv_list, pv, pv_list);
2963
2964 TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_list);
2965
2966 if (!pmap_track_modified(pv->pv_va))
2967 continue;
2968
2969 pte = pmap_pte_quick(pv->pv_pmap, pv->pv_va);
2970
2971 if (pte && (*pte & PG_A)) {
2972 *pte &= ~PG_A;
2973
2974 pmap_TLB_invalidate(pv->pv_pmap, pv->pv_va);
2975
2976 rtval++;
2977 if (rtval > 4) {
2978 break;
2979 }
2980 }
2981 } while ((pv = pvn) != NULL && pv != pvf);
2982 }
2983 splx(s);
2984
2985 return (rtval);
2986 }
2987
2988 /*
2989 * pmap_is_modified:
2990 *
2991 * Return whether or not the specified physical page was modified
2992 * in any physical maps.
2993 */
2994 boolean_t
2995 pmap_is_modified(vm_page_t m)
2996 {
2997 return pmap_testbit(m, PG_M);
2998 }
2999
3000 /*
3001 * Clear the modify bits on the specified physical page.
3002 */
3003 void
3004 pmap_clear_modify(vm_page_t m)
3005 {
3006 pmap_changebit(m, PG_M, FALSE);
3007 }
3008
3009 /*
3010 * pmap_clear_reference:
3011 *
3012 * Clear the reference bit on the specified physical page.
3013 */
3014 void
3015 pmap_clear_reference(vm_page_t m)
3016 {
3017 pmap_changebit(m, PG_A, FALSE);
3018 }
3019
3020 /*
3021 * Miscellaneous support routines follow
3022 */
3023
3024 static void
3025 i386_protection_init()
3026 {
3027 register int *kp, prot;
3028
3029 kp = protection_codes;
3030 for (prot = 0; prot < 8; prot++) {
3031 switch (prot) {
3032 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_NONE:
3033 /*
3034 * Read access is also 0. There isn't any execute bit,
3035 * so just make it readable.
3036 */
3037 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_NONE:
3038 case VM_PROT_READ | VM_PROT_NONE | VM_PROT_EXECUTE:
3039 case VM_PROT_NONE | VM_PROT_NONE | VM_PROT_EXECUTE:
3040 *kp++ = 0;
3041 break;
3042 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_NONE:
3043 case VM_PROT_NONE | VM_PROT_WRITE | VM_PROT_EXECUTE:
3044 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_NONE:
3045 case VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE:
3046 *kp++ = PG_RW;
3047 break;
3048 }
3049 }
3050 }
3051
3052 /*
3053 * Map a set of physical memory pages into the kernel virtual
3054 * address space. Return a pointer to where it is mapped. This
3055 * routine is intended to be used for mapping device memory,
3056 * NOT real memory.
3057 */
3058 void *
3059 pmap_mapdev(vm_paddr_t pa, vm_size_t size)
3060 {
3061 vm_offset_t va, tmpva, offset;
3062 pt_entry_t *pte;
3063
3064 offset = pa & PAGE_MASK;
3065 size = roundup(offset + size, PAGE_SIZE);
3066
3067 va = kmem_alloc_pageable(kernel_map, size);
3068 if (!va)
3069 panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
3070
3071 pa = pa & PG_FRAME;
3072 for (tmpva = va; size > 0;) {
3073 pte = vtopte(tmpva);
3074 *pte = pa | PG_RW | PG_V | pgeflag;
3075 size -= PAGE_SIZE;
3076 tmpva += PAGE_SIZE;
3077 pa += PAGE_SIZE;
3078 }
3079 invltlb();
3080
3081 return ((void *)(va + offset));
3082 }
3083
3084 void
3085 pmap_unmapdev(vm_offset_t va, vm_size_t size)
3086 {
3087 vm_offset_t base, offset;
3088
3089 base = va & PG_FRAME;
3090 offset = va & PAGE_MASK;
3091 size = roundup(offset + size, PAGE_SIZE);
3092 kmem_free(kernel_map, base, size);
3093 }
3094
3095 /*
3096 * perform the pmap work for mincore
3097 */
3098 int
3099 pmap_mincore(pmap_t pmap, vm_offset_t addr)
3100 {
3101 pt_entry_t *ptep, pte;
3102 vm_page_t m;
3103 int val = 0;
3104
3105 ptep = pmap_pte(pmap, addr);
3106 if (ptep == 0) {
3107 return 0;
3108 }
3109
3110 if ((pte = *ptep) != 0) {
3111 vm_paddr_t pa;
3112
3113 val = MINCORE_INCORE;
3114 if ((pte & PG_MANAGED) == 0)
3115 return val;
3116
3117 pa = pte & PG_FRAME;
3118
3119 m = PHYS_TO_VM_PAGE(pa);
3120
3121 /*
3122 * Modified by us
3123 */
3124 if (pte & PG_M)
3125 val |= MINCORE_MODIFIED|MINCORE_MODIFIED_OTHER;
3126 /*
3127 * Modified by someone
3128 */
3129 else if (m->dirty || pmap_is_modified(m))
3130 val |= MINCORE_MODIFIED_OTHER;
3131 /*
3132 * Referenced by us
3133 */
3134 if (pte & PG_A)
3135 val |= MINCORE_REFERENCED|MINCORE_REFERENCED_OTHER;
3136
3137 /*
3138 * Referenced by someone
3139 */
3140 else if ((m->flags & PG_REFERENCED) || pmap_ts_referenced(m)) {
3141 val |= MINCORE_REFERENCED_OTHER;
3142 vm_page_flag_set(m, PG_REFERENCED);
3143 }
3144 }
3145 return val;
3146 }
3147
3148 void
3149 pmap_activate(struct proc *p)
3150 {
3151 pmap_t pmap;
3152
3153 pmap = vmspace_pmap(p->p_vmspace);
3154 #if defined(SMP)
3155 pmap->pm_active |= 1 << cpuid;
3156 #else
3157 pmap->pm_active |= 1;
3158 #endif
3159 #if defined(SWTCH_OPTIM_STATS)
3160 tlb_flush_count++;
3161 #endif
3162 #ifdef PAE
3163 load_cr3(p->p_addr->u_pcb.pcb_cr3 = vtophys(pmap->pm_pdpt));
3164 #else
3165 load_cr3(p->p_addr->u_pcb.pcb_cr3 = vtophys(pmap->pm_pdir));
3166 #endif
3167 }
3168
3169 vm_offset_t
3170 pmap_addr_hint(vm_object_t obj, vm_offset_t addr, vm_size_t size)
3171 {
3172
3173 if ((obj == NULL) || (size < NBPDR) || (obj->type != OBJT_DEVICE)) {
3174 return addr;
3175 }
3176
3177 addr = (addr + (NBPDR - 1)) & ~(NBPDR - 1);
3178 return addr;
3179 }
3180
3181 #ifdef PAE
3182 /*
3183 * Allocate PDPT in lower 32-bit pages
3184 */
3185 struct pdpt_page {
3186 SLIST_ENTRY(pdpt_page) link;
3187 u_int32_t avail;
3188 vm_paddr_t phys;
3189 u_int32_t bits[4];
3190 };
3191
3192 SLIST_HEAD(,pdpt_page) pdpt_pages = SLIST_HEAD_INITIALIZER(&pdpt_pages);
3193 int pdpt_avail = 0;
3194
3195 static pdpt_entry_t *
3196 pmap_alloc_pdpt()
3197 {
3198 struct pdpt_page *pp;
3199 pdpt_entry_t *pdpt = 0;
3200 int i;
3201
3202 if (pdpt_avail == 0) {
3203 pp = (struct pdpt_page *)contigmalloc(PAGE_SIZE, M_DEVBUF,
3204 M_WAITOK, 0ull, 0xffffffffull, PAGE_SIZE, 0);
3205 if (!pp)
3206 panic("pmap_alloc_pdpt: alloc failed");
3207 pp->phys = vtophys(pp);
3208 pp->avail = PAGE_SIZE / 32 - 1;
3209 pp->bits[0] = 1;
3210 pp->bits[1] = pp->bits[2] = pp->bits[3] = 0;
3211 SLIST_INSERT_HEAD(&pdpt_pages, pp, link);
3212 pdpt_avail += pp->avail;
3213 } else {
3214 SLIST_FOREACH(pp, &pdpt_pages, link) {
3215 if (pp->avail > 0)
3216 break;
3217 }
3218 }
3219
3220 for (i = 0; i < 4; i++) {
3221 int j = ffs(~pp->bits[i]);
3222 if (j == 0)
3223 continue;
3224 pp->bits[i] |= 1 << (j - 1);
3225 pp->avail--;
3226 pdpt_avail--;
3227 pdpt = (pdpt_entry_t *)pp + (32 * i + j - 1) * NPGPTD;
3228 }
3229
3230 return pdpt;
3231 }
3232
3233 #if 0
3234 static void
3235 pmap_free_pdpt(pdpt_entry_t *pdpt)
3236 {
3237 struct pdpt_page *pp;
3238 int i;
3239
3240 pp = (struct pdpt_page *)((vm_offset_t)pdpt & ~PAGE_MASK);
3241 i = (pdpt - (pdpt_entry_t *)pp) / NPGPTD;
3242 pp->bits[i / 32] &= ~(1 << (i % 32));
3243 pp->avail++;
3244 pdpt_avail++;
3245 }
3246 #endif
3247 #endif
3248
3249 #if defined(PMAP_DEBUG)
3250 pmap_pid_dump(int pid)
3251 {
3252 pmap_t pmap;
3253 struct proc *p;
3254 int npte = 0;
3255 int index;
3256 int s;
3257 LIST_FOREACH(p, &allproc, p_list) {
3258 if (p->p_pid != pid)
3259 continue;
3260
3261 if (p->p_vmspace) {
3262 int i,j;
3263 index = 0;
3264 pmap = vmspace_pmap(p->p_vmspace);
3265 for(i=0;i<1024;i++) {
3266 pd_entry_t *pde;
3267 unsigned *pte;
3268 unsigned base = i << PDRSHIFT;
3269
3270 pde = &pmap->pm_pdir[i];
3271 if (pde && pmap_pde_v(pde)) {
3272 s = splvm();
3273 for(j=0;j<1024;j++) {
3274 unsigned va = base + (j << PAGE_SHIFT);
3275 if (va >= (vm_offset_t) VM_MIN_KERNEL_ADDRESS) {
3276 if (index) {
3277 index = 0;
3278 printf("\n");
3279 }
3280 splx(s);
3281 return npte;
3282 }
3283 pte = pmap_pte_quick(pmap, va);
3284 if (pte && pmap_pte_v(pte)) {
3285 vm_offset_t pa;
3286 vm_page_t m;
3287 pa = *(int *)pte;
3288 m = PHYS_TO_VM_PAGE(pa);
3289 printf("va: 0x%x, pt: 0x%x, h: %d, w: %d, f: 0x%x",
3290 va, pa, m->hold_count, m->wire_count, m->flags);
3291 npte++;
3292 index++;
3293 if (index >= 2) {
3294 index = 0;
3295 printf("\n");
3296 } else {
3297 printf(" ");
3298 }
3299 }
3300 }
3301 splx(s);
3302 }
3303 }
3304 }
3305 }
3306 return npte;
3307 }
3308 #endif
3309
3310 #if defined(DEBUG)
3311
3312 static void pads __P((pmap_t pm));
3313 void pmap_pvdump __P((vm_offset_t pa));
3314
3315 /* print address space of pmap*/
3316 static void
3317 pads(pm)
3318 pmap_t pm;
3319 {
3320 unsigned va, i, j;
3321 unsigned *ptep;
3322 int s;
3323
3324 if (pm == kernel_pmap)
3325 return;
3326 s = splvm();
3327 for (i = 0; i < 1024; i++)
3328 if (pm->pm_pdir[i])
3329 for (j = 0; j < 1024; j++) {
3330 va = (i << PDRSHIFT) + (j << PAGE_SHIFT);
3331 if (pm == kernel_pmap && va < KERNBASE)
3332 continue;
3333 if (pm != kernel_pmap && va > UPT_MAX_ADDRESS)
3334 continue;
3335 ptep = pmap_pte_quick(pm, va);
3336 if (pmap_pte_v(ptep))
3337 printf("%x:%x ", va, *(int *) ptep);
3338 };
3339 splx(s);
3340
3341 }
3342
3343 void
3344 pmap_pvdump(pa)
3345 vm_paddr_t pa;
3346 {
3347 register pv_entry_t pv;
3348 vm_page_t m;
3349
3350 printf("pa %x", pa);
3351 m = PHYS_TO_VM_PAGE(pa);
3352 TAILQ_FOREACH(pv, &m->md.pv_list, pv_list) {
3353 #ifdef used_to_be
3354 printf(" -> pmap %p, va %x, flags %x",
3355 (void *)pv->pv_pmap, pv->pv_va, pv->pv_flags);
3356 #endif
3357 printf(" -> pmap %p, va %x", (void *)pv->pv_pmap, pv->pv_va);
3358 pads(pv->pv_pmap);
3359 }
3360 printf(" ");
3361 }
3362 #endif
3363
3364 #if defined(I686_CPU) && !defined(NO_PSE_HACK)
3365 static void note_pse_hack(void *unused);
3366 SYSINIT(note_pse_hack, SI_SUB_INTRINSIC, SI_ORDER_FIRST, note_pse_hack, NULL);
3367
3368 static void
3369 note_pse_hack(void *unused) {
3370 if (!has_pse_bug)
3371 return;
3372 printf("Warning: Pentium 4 CPU: PSE disabled\n");
3373 }
3374 #endif
3375
Cache object: 978ea82db40d00c724a92c4739694f33
|