pmap.c revision 1.37 1 /* $NetBSD: pmap.c,v 1.37 2002/02/05 21:14:36 thorpej Exp $ */
2
3 /*
4 * Copyright (c) 2002 Wasabi Systems, Inc.
5 * Copyright (c) 2001 Richard Earnshaw
6 * Copyright (c) 2001 Christopher Gilbert
7 * All rights reserved.
8 *
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 * 3. The name of the company nor the name of the author may be used to
15 * endorse or promote products derived from this software without specific
16 * prior written permission.
17 *
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
19 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
20 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21 * IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
22 * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
24 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 * SUCH DAMAGE.
29 */
30
31 /*-
32 * Copyright (c) 1999 The NetBSD Foundation, Inc.
33 * All rights reserved.
34 *
35 * This code is derived from software contributed to The NetBSD Foundation
36 * by Charles M. Hannum.
37 *
38 * Redistribution and use in source and binary forms, with or without
39 * modification, are permitted provided that the following conditions
40 * are met:
41 * 1. Redistributions of source code must retain the above copyright
42 * notice, this list of conditions and the following disclaimer.
43 * 2. Redistributions in binary form must reproduce the above copyright
44 * notice, this list of conditions and the following disclaimer in the
45 * documentation and/or other materials provided with the distribution.
46 * 3. All advertising materials mentioning features or use of this software
47 * must display the following acknowledgement:
48 * This product includes software developed by the NetBSD
49 * Foundation, Inc. and its contributors.
50 * 4. Neither the name of The NetBSD Foundation nor the names of its
51 * contributors may be used to endorse or promote products derived
52 * from this software without specific prior written permission.
53 *
54 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
55 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
56 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
57 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
58 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
59 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
60 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
61 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
62 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
63 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
64 * POSSIBILITY OF SUCH DAMAGE.
65 */
66
67 /*
68 * Copyright (c) 1994-1998 Mark Brinicombe.
69 * Copyright (c) 1994 Brini.
70 * All rights reserved.
71 *
72 * This code is derived from software written for Brini by Mark Brinicombe
73 *
74 * Redistribution and use in source and binary forms, with or without
75 * modification, are permitted provided that the following conditions
76 * are met:
77 * 1. Redistributions of source code must retain the above copyright
78 * notice, this list of conditions and the following disclaimer.
79 * 2. Redistributions in binary form must reproduce the above copyright
80 * notice, this list of conditions and the following disclaimer in the
81 * documentation and/or other materials provided with the distribution.
82 * 3. All advertising materials mentioning features or use of this software
83 * must display the following acknowledgement:
84 * This product includes software developed by Mark Brinicombe.
85 * 4. The name of the author may not be used to endorse or promote products
86 * derived from this software without specific prior written permission.
87 *
88 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
89 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
90 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
91 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
92 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
93 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
94 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
95 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
96 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
97 *
98 * RiscBSD kernel project
99 *
100 * pmap.c
101 *
102 * Machine dependant vm stuff
103 *
104 * Created : 20/09/94
105 */
106
107 /*
108 * Performance improvements, UVM changes, overhauls and part-rewrites
109 * were contributed by Neil A. Carson <neil (at) causality.com>.
110 */
111
112 /*
113 * The dram block info is currently referenced from the bootconfig.
114 * This should be placed in a separate structure.
115 */
116
117 /*
118 * Special compilation symbols
119 * PMAP_DEBUG - Build in pmap_debug_level code
120 */
121
122 /* Include header files */
123
124 #include "opt_pmap_debug.h"
125 #include "opt_ddb.h"
126
127 #include <sys/types.h>
128 #include <sys/param.h>
129 #include <sys/kernel.h>
130 #include <sys/systm.h>
131 #include <sys/proc.h>
132 #include <sys/malloc.h>
133 #include <sys/user.h>
134 #include <sys/pool.h>
135 #include <sys/cdefs.h>
136
137 #include <uvm/uvm.h>
138
139 #include <machine/bootconfig.h>
140 #include <machine/bus.h>
141 #include <machine/pmap.h>
142 #include <machine/pcb.h>
143 #include <machine/param.h>
144 #include <arm/arm32/katelib.h>
145
146 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.37 2002/02/05 21:14:36 thorpej Exp $");
147 #ifdef PMAP_DEBUG
148 #define PDEBUG(_lev_,_stat_) \
149 if (pmap_debug_level >= (_lev_)) \
150 ((_stat_))
151 int pmap_debug_level = -2;
152
153 /*
154 * for switching to potentially finer grained debugging
155 */
156 #define PDB_FOLLOW 0x0001
157 #define PDB_INIT 0x0002
158 #define PDB_ENTER 0x0004
159 #define PDB_REMOVE 0x0008
160 #define PDB_CREATE 0x0010
161 #define PDB_PTPAGE 0x0020
162 #define PDB_ASN 0x0040
163 #define PDB_BITS 0x0080
164 #define PDB_COLLECT 0x0100
165 #define PDB_PROTECT 0x0200
166 #define PDB_BOOTSTRAP 0x1000
167 #define PDB_PARANOIA 0x2000
168 #define PDB_WIRING 0x4000
169 #define PDB_PVDUMP 0x8000
170
171 int debugmap = 0;
172 int pmapdebug = PDB_PARANOIA | PDB_FOLLOW;
173 #define NPDEBUG(_lev_,_stat_) \
174 if (pmapdebug & (_lev_)) \
175 ((_stat_))
176
177 #else /* PMAP_DEBUG */
178 #define PDEBUG(_lev_,_stat_) /* Nothing */
179 #define PDEBUG(_lev_,_stat_) /* Nothing */
180 #endif /* PMAP_DEBUG */
181
182 struct pmap kernel_pmap_store;
183
184 /*
185 * pool that pmap structures are allocated from
186 */
187
188 struct pool pmap_pmap_pool;
189
190 pagehook_t page_hook0;
191 pagehook_t page_hook1;
192 char *memhook;
193 pt_entry_t msgbufpte;
194 extern caddr_t msgbufaddr;
195
196 boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
197 /*
198 * locking data structures
199 */
200
201 static struct lock pmap_main_lock;
202 static struct simplelock pvalloc_lock;
203 #ifdef LOCKDEBUG
204 #define PMAP_MAP_TO_HEAD_LOCK() \
205 (void) spinlockmgr(&pmap_main_lock, LK_SHARED, NULL)
206 #define PMAP_MAP_TO_HEAD_UNLOCK() \
207 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
208
209 #define PMAP_HEAD_TO_MAP_LOCK() \
210 (void) spinlockmgr(&pmap_main_lock, LK_EXCLUSIVE, NULL)
211 #define PMAP_HEAD_TO_MAP_UNLOCK() \
212 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
213 #else
214 #define PMAP_MAP_TO_HEAD_LOCK() /* nothing */
215 #define PMAP_MAP_TO_HEAD_UNLOCK() /* nothing */
216 #define PMAP_HEAD_TO_MAP_LOCK() /* nothing */
217 #define PMAP_HEAD_TO_MAP_UNLOCK() /* nothing */
218 #endif /* LOCKDEBUG */
219
220 /*
221 * pv_page management structures: locked by pvalloc_lock
222 */
223
224 TAILQ_HEAD(pv_pagelist, pv_page);
225 static struct pv_pagelist pv_freepages; /* list of pv_pages with free entrys */
226 static struct pv_pagelist pv_unusedpgs; /* list of unused pv_pages */
227 static int pv_nfpvents; /* # of free pv entries */
228 static struct pv_page *pv_initpage; /* bootstrap page from kernel_map */
229 static vaddr_t pv_cachedva; /* cached VA for later use */
230
231 #define PVE_LOWAT (PVE_PER_PVPAGE / 2) /* free pv_entry low water mark */
232 #define PVE_HIWAT (PVE_LOWAT + (PVE_PER_PVPAGE * 2))
233 /* high water mark */
234
235 /*
236 * local prototypes
237 */
238
239 static struct pv_entry *pmap_add_pvpage __P((struct pv_page *, boolean_t));
240 static struct pv_entry *pmap_alloc_pv __P((struct pmap *, int)); /* see codes below */
241 #define ALLOCPV_NEED 0 /* need PV now */
242 #define ALLOCPV_TRY 1 /* just try to allocate, don't steal */
243 #define ALLOCPV_NONEED 2 /* don't need PV, just growing cache */
244 static struct pv_entry *pmap_alloc_pvpage __P((struct pmap *, int));
245 static void pmap_enter_pv __P((struct vm_page *,
246 struct pv_entry *, struct pmap *,
247 vaddr_t, struct vm_page *, int));
248 static void pmap_free_pv __P((struct pmap *, struct pv_entry *));
249 static void pmap_free_pvs __P((struct pmap *, struct pv_entry *));
250 static void pmap_free_pv_doit __P((struct pv_entry *));
251 static void pmap_free_pvpage __P((void));
252 static boolean_t pmap_is_curpmap __P((struct pmap *));
253 static struct pv_entry *pmap_remove_pv __P((struct vm_page *, struct pmap *,
254 vaddr_t));
255 #define PMAP_REMOVE_ALL 0 /* remove all mappings */
256 #define PMAP_REMOVE_SKIPWIRED 1 /* skip wired mappings */
257
258 static u_int pmap_modify_pv __P((struct pmap *, vaddr_t, struct vm_page *,
259 u_int, u_int));
260
261 static void pmap_free_l1pt __P((struct l1pt *));
262 static int pmap_allocpagedir __P((struct pmap *));
263 static int pmap_clean_page __P((struct pv_entry *, boolean_t));
264 static void pmap_remove_all __P((struct vm_page *));
265
266
267 vsize_t npages;
268
269 static struct vm_page *pmap_alloc_ptp __P((struct pmap *, vaddr_t, boolean_t));
270 static struct vm_page *pmap_get_ptp __P((struct pmap *, vaddr_t, boolean_t));
271 __inline static void pmap_clearbit __P((struct vm_page *, unsigned int));
272 __inline static boolean_t pmap_testbit __P((struct vm_page *, unsigned int));
273
274 extern paddr_t physical_start;
275 extern paddr_t physical_freestart;
276 extern paddr_t physical_end;
277 extern paddr_t physical_freeend;
278 extern unsigned int free_pages;
279 extern int max_processes;
280
281 vaddr_t virtual_start;
282 vaddr_t virtual_end;
283
284 vaddr_t avail_start;
285 vaddr_t avail_end;
286
287 extern pv_addr_t systempage;
288
289 #define ALLOC_PAGE_HOOK(x, s) \
290 x.va = virtual_start; \
291 x.pte = (pt_entry_t *)pmap_pte(pmap_kernel(), virtual_start); \
292 virtual_start += s;
293
294 /* Variables used by the L1 page table queue code */
295 SIMPLEQ_HEAD(l1pt_queue, l1pt);
296 struct l1pt_queue l1pt_static_queue; /* head of our static l1 queue */
297 int l1pt_static_queue_count; /* items in the static l1 queue */
298 int l1pt_static_create_count; /* static l1 items created */
299 struct l1pt_queue l1pt_queue; /* head of our l1 queue */
300 int l1pt_queue_count; /* items in the l1 queue */
301 int l1pt_create_count; /* stat - L1's create count */
302 int l1pt_reuse_count; /* stat - L1's reused count */
303
304 /* Local function prototypes (not used outside this file) */
305 pt_entry_t *pmap_pte __P((struct pmap *pmap, vaddr_t va));
306 void pmap_copy_on_write __P((struct vm_page *));
307 void pmap_pinit __P((struct pmap *));
308 void pmap_freepagedir __P((struct pmap *));
309
310 /* Other function prototypes */
311 extern void bzero_page __P((vaddr_t));
312 extern void bcopy_page __P((vaddr_t, vaddr_t));
313
314 struct l1pt *pmap_alloc_l1pt __P((void));
315 static __inline void pmap_map_in_l1 __P((struct pmap *pmap, vaddr_t va,
316 vaddr_t l2pa, boolean_t));
317
318 static pt_entry_t *pmap_map_ptes __P((struct pmap *));
319 static void pmap_unmap_ptes __P((struct pmap *));
320
321 __inline static void pmap_vac_me_harder __P((struct pmap *, struct vm_page *,
322 pt_entry_t *, boolean_t));
323 static void pmap_vac_me_kpmap __P((struct pmap *, struct vm_page *,
324 pt_entry_t *, boolean_t));
325 static void pmap_vac_me_user __P((struct pmap *, struct vm_page *,
326 pt_entry_t *, boolean_t));
327
328 /*
329 * Cache enable bits in PTE to use on pages that are cacheable.
330 * On most machines this is cacheable/bufferable, but on some, eg arm10, we
331 * can chose between write-through and write-back cacheing.
332 */
333 pt_entry_t pte_cache_mode = (PT_C | PT_B);
334
335 /*
336 * real definition of pv_entry.
337 */
338
339 struct pv_entry {
340 struct pv_entry *pv_next; /* next pv_entry */
341 struct pmap *pv_pmap; /* pmap where mapping lies */
342 vaddr_t pv_va; /* virtual address for mapping */
343 int pv_flags; /* flags */
344 struct vm_page *pv_ptp; /* vm_page for the ptp */
345 };
346
347 /*
348 * pv_entrys are dynamically allocated in chunks from a single page.
349 * we keep track of how many pv_entrys are in use for each page and
350 * we can free pv_entry pages if needed. there is one lock for the
351 * entire allocation system.
352 */
353
354 struct pv_page_info {
355 TAILQ_ENTRY(pv_page) pvpi_list;
356 struct pv_entry *pvpi_pvfree;
357 int pvpi_nfree;
358 };
359
360 /*
361 * number of pv_entry's in a pv_page
362 * (note: won't work on systems where NPBG isn't a constant)
363 */
364
365 #define PVE_PER_PVPAGE ((NBPG - sizeof(struct pv_page_info)) / \
366 sizeof(struct pv_entry))
367
368 /*
369 * a pv_page: where pv_entrys are allocated from
370 */
371
372 struct pv_page {
373 struct pv_page_info pvinfo;
374 struct pv_entry pvents[PVE_PER_PVPAGE];
375 };
376
377 #ifdef MYCROFT_HACK
378 int mycroft_hack = 0;
379 #endif
380
381 /* Function to set the debug level of the pmap code */
382
383 #ifdef PMAP_DEBUG
384 void
385 pmap_debug(level)
386 int level;
387 {
388 pmap_debug_level = level;
389 printf("pmap_debug: level=%d\n", pmap_debug_level);
390 }
391 #endif /* PMAP_DEBUG */
392
393 __inline static boolean_t
394 pmap_is_curpmap(struct pmap *pmap)
395 {
396 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap)
397 || (pmap == pmap_kernel()))
398 return (TRUE);
399 return (FALSE);
400 }
401 #include "isadma.h"
402
403 #if NISADMA > 0
404 /*
405 * Used to protect memory for ISA DMA bounce buffers. If, when loading
406 * pages into the system, memory intersects with any of these ranges,
407 * the intersecting memory will be loaded into a lower-priority free list.
408 */
409 bus_dma_segment_t *pmap_isa_dma_ranges;
410 int pmap_isa_dma_nranges;
411
412 boolean_t pmap_isa_dma_range_intersect __P((paddr_t, psize_t,
413 paddr_t *, psize_t *));
414
415 /*
416 * Check if a memory range intersects with an ISA DMA range, and
417 * return the page-rounded intersection if it does. The intersection
418 * will be placed on a lower-priority free list.
419 */
420 boolean_t
421 pmap_isa_dma_range_intersect(pa, size, pap, sizep)
422 paddr_t pa;
423 psize_t size;
424 paddr_t *pap;
425 psize_t *sizep;
426 {
427 bus_dma_segment_t *ds;
428 int i;
429
430 if (pmap_isa_dma_ranges == NULL)
431 return (FALSE);
432
433 for (i = 0, ds = pmap_isa_dma_ranges;
434 i < pmap_isa_dma_nranges; i++, ds++) {
435 if (ds->ds_addr <= pa && pa < (ds->ds_addr + ds->ds_len)) {
436 /*
437 * Beginning of region intersects with this range.
438 */
439 *pap = trunc_page(pa);
440 *sizep = round_page(min(pa + size,
441 ds->ds_addr + ds->ds_len) - pa);
442 return (TRUE);
443 }
444 if (pa < ds->ds_addr && ds->ds_addr < (pa + size)) {
445 /*
446 * End of region intersects with this range.
447 */
448 *pap = trunc_page(ds->ds_addr);
449 *sizep = round_page(min((pa + size) - ds->ds_addr,
450 ds->ds_len));
451 return (TRUE);
452 }
453 }
454
455 /*
456 * No intersection found.
457 */
458 return (FALSE);
459 }
460 #endif /* NISADMA > 0 */
461
462 /*
463 * p v _ e n t r y f u n c t i o n s
464 */
465
466 /*
467 * pv_entry allocation functions:
468 * the main pv_entry allocation functions are:
469 * pmap_alloc_pv: allocate a pv_entry structure
470 * pmap_free_pv: free one pv_entry
471 * pmap_free_pvs: free a list of pv_entrys
472 *
473 * the rest are helper functions
474 */
475
476 /*
477 * pmap_alloc_pv: inline function to allocate a pv_entry structure
478 * => we lock pvalloc_lock
479 * => if we fail, we call out to pmap_alloc_pvpage
480 * => 3 modes:
481 * ALLOCPV_NEED = we really need a pv_entry, even if we have to steal it
482 * ALLOCPV_TRY = we want a pv_entry, but not enough to steal
483 * ALLOCPV_NONEED = we are trying to grow our free list, don't really need
484 * one now
485 *
486 * "try" is for optional functions like pmap_copy().
487 */
488
489 __inline static struct pv_entry *
490 pmap_alloc_pv(pmap, mode)
491 struct pmap *pmap;
492 int mode;
493 {
494 struct pv_page *pvpage;
495 struct pv_entry *pv;
496
497 simple_lock(&pvalloc_lock);
498
499 if (pv_freepages.tqh_first != NULL) {
500 pvpage = pv_freepages.tqh_first;
501 pvpage->pvinfo.pvpi_nfree--;
502 if (pvpage->pvinfo.pvpi_nfree == 0) {
503 /* nothing left in this one? */
504 TAILQ_REMOVE(&pv_freepages, pvpage, pvinfo.pvpi_list);
505 }
506 pv = pvpage->pvinfo.pvpi_pvfree;
507 #ifdef DIAGNOSTIC
508 if (pv == NULL)
509 panic("pmap_alloc_pv: pvpi_nfree off");
510 #endif
511 pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
512 pv_nfpvents--; /* took one from pool */
513 } else {
514 pv = NULL; /* need more of them */
515 }
516
517 /*
518 * if below low water mark or we didn't get a pv_entry we try and
519 * create more pv_entrys ...
520 */
521
522 if (pv_nfpvents < PVE_LOWAT || pv == NULL) {
523 if (pv == NULL)
524 pv = pmap_alloc_pvpage(pmap, (mode == ALLOCPV_TRY) ?
525 mode : ALLOCPV_NEED);
526 else
527 (void) pmap_alloc_pvpage(pmap, ALLOCPV_NONEED);
528 }
529
530 simple_unlock(&pvalloc_lock);
531 return(pv);
532 }
533
534 /*
535 * pmap_alloc_pvpage: maybe allocate a new pvpage
536 *
537 * if need_entry is false: try and allocate a new pv_page
538 * if need_entry is true: try and allocate a new pv_page and return a
539 * new pv_entry from it. if we are unable to allocate a pv_page
540 * we make a last ditch effort to steal a pv_page from some other
541 * mapping. if that fails, we panic...
542 *
543 * => we assume that the caller holds pvalloc_lock
544 */
545
546 static struct pv_entry *
547 pmap_alloc_pvpage(pmap, mode)
548 struct pmap *pmap;
549 int mode;
550 {
551 struct vm_page *pg;
552 struct pv_page *pvpage;
553 struct pv_entry *pv;
554 int s;
555
556 /*
557 * if we need_entry and we've got unused pv_pages, allocate from there
558 */
559
560 if (mode != ALLOCPV_NONEED && pv_unusedpgs.tqh_first != NULL) {
561
562 /* move it to pv_freepages list */
563 pvpage = pv_unusedpgs.tqh_first;
564 TAILQ_REMOVE(&pv_unusedpgs, pvpage, pvinfo.pvpi_list);
565 TAILQ_INSERT_HEAD(&pv_freepages, pvpage, pvinfo.pvpi_list);
566
567 /* allocate a pv_entry */
568 pvpage->pvinfo.pvpi_nfree--; /* can't go to zero */
569 pv = pvpage->pvinfo.pvpi_pvfree;
570 #ifdef DIAGNOSTIC
571 if (pv == NULL)
572 panic("pmap_alloc_pvpage: pvpi_nfree off");
573 #endif
574 pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
575
576 pv_nfpvents--; /* took one from pool */
577 return(pv);
578 }
579
580 /*
581 * see if we've got a cached unmapped VA that we can map a page in.
582 * if not, try to allocate one.
583 */
584
585
586 if (pv_cachedva == 0) {
587 s = splvm();
588 pv_cachedva = uvm_km_kmemalloc(kmem_map, NULL,
589 PAGE_SIZE, UVM_KMF_TRYLOCK|UVM_KMF_VALLOC);
590 splx(s);
591 if (pv_cachedva == 0) {
592 return (NULL);
593 }
594 }
595
596 pg = uvm_pagealloc(NULL, pv_cachedva - vm_map_min(kernel_map), NULL,
597 UVM_PGA_USERESERVE);
598 if (pg)
599 pg->flags &= ~PG_BUSY; /* never busy */
600
601 if (pg == NULL)
602 return (NULL);
603
604 /*
605 * add a mapping for our new pv_page and free its entrys (save one!)
606 *
607 * NOTE: If we are allocating a PV page for the kernel pmap, the
608 * pmap is already locked! (...but entering the mapping is safe...)
609 */
610
611 pmap_kenter_pa(pv_cachedva, VM_PAGE_TO_PHYS(pg), VM_PROT_ALL);
612 pmap_update(pmap_kernel());
613 pvpage = (struct pv_page *) pv_cachedva;
614 pv_cachedva = 0;
615 return (pmap_add_pvpage(pvpage, mode != ALLOCPV_NONEED));
616 }
617
618 /*
619 * pmap_add_pvpage: add a pv_page's pv_entrys to the free list
620 *
621 * => caller must hold pvalloc_lock
622 * => if need_entry is true, we allocate and return one pv_entry
623 */
624
625 static struct pv_entry *
626 pmap_add_pvpage(pvp, need_entry)
627 struct pv_page *pvp;
628 boolean_t need_entry;
629 {
630 int tofree, lcv;
631
632 /* do we need to return one? */
633 tofree = (need_entry) ? PVE_PER_PVPAGE - 1 : PVE_PER_PVPAGE;
634
635 pvp->pvinfo.pvpi_pvfree = NULL;
636 pvp->pvinfo.pvpi_nfree = tofree;
637 for (lcv = 0 ; lcv < tofree ; lcv++) {
638 pvp->pvents[lcv].pv_next = pvp->pvinfo.pvpi_pvfree;
639 pvp->pvinfo.pvpi_pvfree = &pvp->pvents[lcv];
640 }
641 if (need_entry)
642 TAILQ_INSERT_TAIL(&pv_freepages, pvp, pvinfo.pvpi_list);
643 else
644 TAILQ_INSERT_TAIL(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
645 pv_nfpvents += tofree;
646 return((need_entry) ? &pvp->pvents[lcv] : NULL);
647 }
648
649 /*
650 * pmap_free_pv_doit: actually free a pv_entry
651 *
652 * => do not call this directly! instead use either
653 * 1. pmap_free_pv ==> free a single pv_entry
654 * 2. pmap_free_pvs => free a list of pv_entrys
655 * => we must be holding pvalloc_lock
656 */
657
658 __inline static void
659 pmap_free_pv_doit(pv)
660 struct pv_entry *pv;
661 {
662 struct pv_page *pvp;
663
664 pvp = (struct pv_page *) arm_trunc_page((vaddr_t)pv);
665 pv_nfpvents++;
666 pvp->pvinfo.pvpi_nfree++;
667
668 /* nfree == 1 => fully allocated page just became partly allocated */
669 if (pvp->pvinfo.pvpi_nfree == 1) {
670 TAILQ_INSERT_HEAD(&pv_freepages, pvp, pvinfo.pvpi_list);
671 }
672
673 /* free it */
674 pv->pv_next = pvp->pvinfo.pvpi_pvfree;
675 pvp->pvinfo.pvpi_pvfree = pv;
676
677 /*
678 * are all pv_page's pv_entry's free? move it to unused queue.
679 */
680
681 if (pvp->pvinfo.pvpi_nfree == PVE_PER_PVPAGE) {
682 TAILQ_REMOVE(&pv_freepages, pvp, pvinfo.pvpi_list);
683 TAILQ_INSERT_HEAD(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
684 }
685 }
686
687 /*
688 * pmap_free_pv: free a single pv_entry
689 *
690 * => we gain the pvalloc_lock
691 */
692
693 __inline static void
694 pmap_free_pv(pmap, pv)
695 struct pmap *pmap;
696 struct pv_entry *pv;
697 {
698 simple_lock(&pvalloc_lock);
699 pmap_free_pv_doit(pv);
700
701 /*
702 * Can't free the PV page if the PV entries were associated with
703 * the kernel pmap; the pmap is already locked.
704 */
705 if (pv_nfpvents > PVE_HIWAT && pv_unusedpgs.tqh_first != NULL &&
706 pmap != pmap_kernel())
707 pmap_free_pvpage();
708
709 simple_unlock(&pvalloc_lock);
710 }
711
712 /*
713 * pmap_free_pvs: free a list of pv_entrys
714 *
715 * => we gain the pvalloc_lock
716 */
717
718 __inline static void
719 pmap_free_pvs(pmap, pvs)
720 struct pmap *pmap;
721 struct pv_entry *pvs;
722 {
723 struct pv_entry *nextpv;
724
725 simple_lock(&pvalloc_lock);
726
727 for ( /* null */ ; pvs != NULL ; pvs = nextpv) {
728 nextpv = pvs->pv_next;
729 pmap_free_pv_doit(pvs);
730 }
731
732 /*
733 * Can't free the PV page if the PV entries were associated with
734 * the kernel pmap; the pmap is already locked.
735 */
736 if (pv_nfpvents > PVE_HIWAT && pv_unusedpgs.tqh_first != NULL &&
737 pmap != pmap_kernel())
738 pmap_free_pvpage();
739
740 simple_unlock(&pvalloc_lock);
741 }
742
743
744 /*
745 * pmap_free_pvpage: try and free an unused pv_page structure
746 *
747 * => assume caller is holding the pvalloc_lock and that
748 * there is a page on the pv_unusedpgs list
749 * => if we can't get a lock on the kmem_map we try again later
750 * => note: analysis of MI kmem_map usage [i.e. malloc/free] shows
751 * that if we can lock the kmem_map then we are not already
752 * holding kmem_object's lock.
753 */
754
755 static void
756 pmap_free_pvpage()
757 {
758 int s;
759 struct vm_map *map;
760 struct vm_map_entry *dead_entries;
761 struct pv_page *pvp;
762
763 s = splvm(); /* protect kmem_map */
764
765 pvp = pv_unusedpgs.tqh_first;
766
767 /*
768 * note: watch out for pv_initpage which is allocated out of
769 * kernel_map rather than kmem_map.
770 */
771 if (pvp == pv_initpage)
772 map = kernel_map;
773 else
774 map = kmem_map;
775
776 if (vm_map_lock_try(map)) {
777
778 /* remove pvp from pv_unusedpgs */
779 TAILQ_REMOVE(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
780
781 /* unmap the page */
782 dead_entries = NULL;
783 uvm_unmap_remove(map, (vaddr_t)pvp, ((vaddr_t)pvp) + PAGE_SIZE,
784 &dead_entries);
785 vm_map_unlock(map);
786
787 if (dead_entries != NULL)
788 uvm_unmap_detach(dead_entries, 0);
789
790 pv_nfpvents -= PVE_PER_PVPAGE; /* update free count */
791 }
792
793 if (pvp == pv_initpage)
794 /* no more initpage, we've freed it */
795 pv_initpage = NULL;
796
797 splx(s);
798 }
799
800 /*
801 * main pv_entry manipulation functions:
802 * pmap_enter_pv: enter a mapping onto a vm_page list
803 * pmap_remove_pv: remove a mappiing from a vm_page list
804 *
805 * NOTE: pmap_enter_pv expects to lock the pvh itself
806 * pmap_remove_pv expects te caller to lock the pvh before calling
807 */
808
809 /*
810 * pmap_enter_pv: enter a mapping onto a vm_page lst
811 *
812 * => caller should hold the proper lock on pmap_main_lock
813 * => caller should have pmap locked
814 * => we will gain the lock on the vm_page and allocate the new pv_entry
815 * => caller should adjust ptp's wire_count before calling
816 * => caller should not adjust pmap's wire_count
817 */
818
819 __inline static void
820 pmap_enter_pv(pg, pve, pmap, va, ptp, flags)
821 struct vm_page *pg;
822 struct pv_entry *pve; /* preallocated pve for us to use */
823 struct pmap *pmap;
824 vaddr_t va;
825 struct vm_page *ptp; /* PTP in pmap that maps this VA */
826 int flags;
827 {
828 pve->pv_pmap = pmap;
829 pve->pv_va = va;
830 pve->pv_ptp = ptp; /* NULL for kernel pmap */
831 pve->pv_flags = flags;
832 simple_lock(&pg->mdpage.pvh_slock); /* lock vm_page */
833 pve->pv_next = pg->mdpage.pvh_list; /* add to ... */
834 pg->mdpage.pvh_list = pve; /* ... locked list */
835 simple_unlock(&pg->mdpage.pvh_slock); /* unlock, done! */
836 if (pve->pv_flags & PT_W)
837 ++pmap->pm_stats.wired_count;
838 }
839
840 /*
841 * pmap_remove_pv: try to remove a mapping from a pv_list
842 *
843 * => caller should hold proper lock on pmap_main_lock
844 * => pmap should be locked
845 * => caller should hold lock on vm_page [so that attrs can be adjusted]
846 * => caller should adjust ptp's wire_count and free PTP if needed
847 * => caller should NOT adjust pmap's wire_count
848 * => we return the removed pve
849 */
850
851 __inline static struct pv_entry *
852 pmap_remove_pv(pg, pmap, va)
853 struct vm_page *pg;
854 struct pmap *pmap;
855 vaddr_t va;
856 {
857 struct pv_entry *pve, **prevptr;
858
859 prevptr = &pg->mdpage.pvh_list; /* previous pv_entry pointer */
860 pve = *prevptr;
861 while (pve) {
862 if (pve->pv_pmap == pmap && pve->pv_va == va) { /* match? */
863 *prevptr = pve->pv_next; /* remove it! */
864 if (pve->pv_flags & PT_W)
865 --pmap->pm_stats.wired_count;
866 break;
867 }
868 prevptr = &pve->pv_next; /* previous pointer */
869 pve = pve->pv_next; /* advance */
870 }
871 return(pve); /* return removed pve */
872 }
873
874 /*
875 *
876 * pmap_modify_pv: Update pv flags
877 *
878 * => caller should hold lock on vm_page [so that attrs can be adjusted]
879 * => caller should NOT adjust pmap's wire_count
880 * => caller must call pmap_vac_me_harder() if writable status of a page
881 * may have changed.
882 * => we return the old flags
883 *
884 * Modify a physical-virtual mapping in the pv table
885 */
886
887 /*__inline */
888 static u_int
889 pmap_modify_pv(pmap, va, pg, bic_mask, eor_mask)
890 struct pmap *pmap;
891 vaddr_t va;
892 struct vm_page *pg;
893 u_int bic_mask;
894 u_int eor_mask;
895 {
896 struct pv_entry *npv;
897 u_int flags, oflags;
898
899 /*
900 * There is at least one VA mapping this page.
901 */
902
903 for (npv = pg->mdpage.pvh_list; npv; npv = npv->pv_next) {
904 if (pmap == npv->pv_pmap && va == npv->pv_va) {
905 oflags = npv->pv_flags;
906 npv->pv_flags = flags =
907 ((oflags & ~bic_mask) ^ eor_mask);
908 if ((flags ^ oflags) & PT_W) {
909 if (flags & PT_W)
910 ++pmap->pm_stats.wired_count;
911 else
912 --pmap->pm_stats.wired_count;
913 }
914 return (oflags);
915 }
916 }
917 return (0);
918 }
919
920 /*
921 * Map the specified level 2 pagetable into the level 1 page table for
922 * the given pmap to cover a chunk of virtual address space starting from the
923 * address specified.
924 */
925 static /*__inline*/ void
926 pmap_map_in_l1(pmap, va, l2pa, selfref)
927 struct pmap *pmap;
928 vaddr_t va, l2pa;
929 boolean_t selfref;
930 {
931 vaddr_t ptva;
932
933 /* Calculate the index into the L1 page table. */
934 ptva = (va >> PDSHIFT) & ~3;
935
936 PDEBUG(0, printf("wiring %08lx in to pd%p pte0x%lx va0x%lx\n", l2pa,
937 pmap->pm_pdir, L1_PTE(l2pa), ptva));
938
939 /* Map page table into the L1. */
940 pmap->pm_pdir[ptva + 0] = L1_PTE(l2pa + 0x000);
941 pmap->pm_pdir[ptva + 1] = L1_PTE(l2pa + 0x400);
942 pmap->pm_pdir[ptva + 2] = L1_PTE(l2pa + 0x800);
943 pmap->pm_pdir[ptva + 3] = L1_PTE(l2pa + 0xc00);
944
945 PDEBUG(0, printf("pt self reference %lx in %lx\n",
946 L2_PTE_NC_NB(l2pa, AP_KRW), pmap->pm_vptpt));
947
948 /* Map the page table into the page table area. */
949 if (selfref) {
950 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) =
951 L2_PTE_NC_NB(l2pa, AP_KRW);
952 }
953 /* XXX should be a purge */
954 /* cpu_tlb_flushD();*/
955 }
956
957 #if 0
958 static /*__inline*/ void
959 pmap_unmap_in_l1(pmap, va)
960 struct pmap *pmap;
961 vaddr_t va;
962 {
963 vaddr_t ptva;
964
965 /* Calculate the index into the L1 page table. */
966 ptva = (va >> PDSHIFT) & ~3;
967
968 /* Unmap page table from the L1. */
969 pmap->pm_pdir[ptva + 0] = 0;
970 pmap->pm_pdir[ptva + 1] = 0;
971 pmap->pm_pdir[ptva + 2] = 0;
972 pmap->pm_pdir[ptva + 3] = 0;
973
974 /* Unmap the page table from the page table area. */
975 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) = 0;
976
977 /* XXX should be a purge */
978 /* cpu_tlb_flushD();*/
979 }
980 #endif
981
982 /*
983 * Used to map a range of physical addresses into kernel
984 * virtual address space.
985 *
986 * For now, VM is already on, we only need to map the
987 * specified memory.
988 */
989 vaddr_t
990 pmap_map(va, spa, epa, prot)
991 vaddr_t va, spa, epa;
992 int prot;
993 {
994 while (spa < epa) {
995 pmap_kenter_pa(va, spa, prot);
996 va += NBPG;
997 spa += NBPG;
998 }
999 pmap_update(pmap_kernel());
1000 return(va);
1001 }
1002
1003
1004 /*
1005 * void pmap_bootstrap(pd_entry_t *kernel_l1pt, pv_addr_t kernel_ptpt)
1006 *
1007 * bootstrap the pmap system. This is called from initarm and allows
1008 * the pmap system to initailise any structures it requires.
1009 *
1010 * Currently this sets up the kernel_pmap that is statically allocated
1011 * and also allocated virtual addresses for certain page hooks.
1012 * Currently the only one page hook is allocated that is used
1013 * to zero physical pages of memory.
1014 * It also initialises the start and end address of the kernel data space.
1015 */
1016 extern paddr_t physical_freestart;
1017 extern paddr_t physical_freeend;
1018
1019 char *boot_head;
1020
1021 void
1022 pmap_bootstrap(kernel_l1pt, kernel_ptpt)
1023 pd_entry_t *kernel_l1pt;
1024 pv_addr_t kernel_ptpt;
1025 {
1026 int loop;
1027 paddr_t start, end;
1028 #if NISADMA > 0
1029 paddr_t istart;
1030 psize_t isize;
1031 #endif
1032
1033 pmap_kernel()->pm_pdir = kernel_l1pt;
1034 pmap_kernel()->pm_pptpt = kernel_ptpt.pv_pa;
1035 pmap_kernel()->pm_vptpt = kernel_ptpt.pv_va;
1036 simple_lock_init(&pmap_kernel()->pm_lock);
1037 pmap_kernel()->pm_obj.pgops = NULL;
1038 TAILQ_INIT(&(pmap_kernel()->pm_obj.memq));
1039 pmap_kernel()->pm_obj.uo_npages = 0;
1040 pmap_kernel()->pm_obj.uo_refs = 1;
1041
1042 /*
1043 * Initialize PAGE_SIZE-dependent variables.
1044 */
1045 uvm_setpagesize();
1046
1047 npages = 0;
1048 loop = 0;
1049 while (loop < bootconfig.dramblocks) {
1050 start = (paddr_t)bootconfig.dram[loop].address;
1051 end = start + (bootconfig.dram[loop].pages * NBPG);
1052 if (start < physical_freestart)
1053 start = physical_freestart;
1054 if (end > physical_freeend)
1055 end = physical_freeend;
1056 #if 0
1057 printf("%d: %lx -> %lx\n", loop, start, end - 1);
1058 #endif
1059 #if NISADMA > 0
1060 if (pmap_isa_dma_range_intersect(start, end - start,
1061 &istart, &isize)) {
1062 /*
1063 * Place the pages that intersect with the
1064 * ISA DMA range onto the ISA DMA free list.
1065 */
1066 #if 0
1067 printf(" ISADMA 0x%lx -> 0x%lx\n", istart,
1068 istart + isize - 1);
1069 #endif
1070 uvm_page_physload(atop(istart),
1071 atop(istart + isize), atop(istart),
1072 atop(istart + isize), VM_FREELIST_ISADMA);
1073 npages += atop(istart + isize) - atop(istart);
1074
1075 /*
1076 * Load the pieces that come before
1077 * the intersection into the default
1078 * free list.
1079 */
1080 if (start < istart) {
1081 #if 0
1082 printf(" BEFORE 0x%lx -> 0x%lx\n",
1083 start, istart - 1);
1084 #endif
1085 uvm_page_physload(atop(start),
1086 atop(istart), atop(start),
1087 atop(istart), VM_FREELIST_DEFAULT);
1088 npages += atop(istart) - atop(start);
1089 }
1090
1091 /*
1092 * Load the pieces that come after
1093 * the intersection into the default
1094 * free list.
1095 */
1096 if ((istart + isize) < end) {
1097 #if 0
1098 printf(" AFTER 0x%lx -> 0x%lx\n",
1099 (istart + isize), end - 1);
1100 #endif
1101 uvm_page_physload(atop(istart + isize),
1102 atop(end), atop(istart + isize),
1103 atop(end), VM_FREELIST_DEFAULT);
1104 npages += atop(end) - atop(istart + isize);
1105 }
1106 } else {
1107 uvm_page_physload(atop(start), atop(end),
1108 atop(start), atop(end), VM_FREELIST_DEFAULT);
1109 npages += atop(end) - atop(start);
1110 }
1111 #else /* NISADMA > 0 */
1112 uvm_page_physload(atop(start), atop(end),
1113 atop(start), atop(end), VM_FREELIST_DEFAULT);
1114 npages += atop(end) - atop(start);
1115 #endif /* NISADMA > 0 */
1116 ++loop;
1117 }
1118
1119 #ifdef MYCROFT_HACK
1120 printf("npages = %ld\n", npages);
1121 #endif
1122
1123 virtual_start = KERNEL_VM_BASE;
1124 virtual_end = virtual_start + KERNEL_VM_SIZE - 1;
1125
1126 ALLOC_PAGE_HOOK(page_hook0, NBPG);
1127 ALLOC_PAGE_HOOK(page_hook1, NBPG);
1128
1129 /*
1130 * The mem special device needs a virtual hook but we don't
1131 * need a pte
1132 */
1133 memhook = (char *)virtual_start;
1134 virtual_start += NBPG;
1135
1136 msgbufaddr = (caddr_t)virtual_start;
1137 msgbufpte = (pt_entry_t)pmap_pte(pmap_kernel(), virtual_start);
1138 virtual_start += round_page(MSGBUFSIZE);
1139
1140 /*
1141 * init the static-global locks and global lists.
1142 */
1143 spinlockinit(&pmap_main_lock, "pmaplk", 0);
1144 simple_lock_init(&pvalloc_lock);
1145 TAILQ_INIT(&pv_freepages);
1146 TAILQ_INIT(&pv_unusedpgs);
1147
1148 /*
1149 * initialize the pmap pool.
1150 */
1151
1152 pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl",
1153 0, pool_page_alloc_nointr, pool_page_free_nointr, M_VMPMAP);
1154
1155 cpu_dcache_wbinv_all();
1156 }
1157
1158 /*
1159 * void pmap_init(void)
1160 *
1161 * Initialize the pmap module.
1162 * Called by vm_init() in vm/vm_init.c in order to initialise
1163 * any structures that the pmap system needs to map virtual memory.
1164 */
1165
1166 extern int physmem;
1167
1168 void
1169 pmap_init()
1170 {
1171
1172 /*
1173 * Set the available memory vars - These do not map to real memory
1174 * addresses and cannot as the physical memory is fragmented.
1175 * They are used by ps for %mem calculations.
1176 * One could argue whether this should be the entire memory or just
1177 * the memory that is useable in a user process.
1178 */
1179 avail_start = 0;
1180 avail_end = physmem * NBPG;
1181
1182 /*
1183 * now we need to free enough pv_entry structures to allow us to get
1184 * the kmem_map/kmem_object allocated and inited (done after this
1185 * function is finished). to do this we allocate one bootstrap page out
1186 * of kernel_map and use it to provide an initial pool of pv_entry
1187 * structures. we never free this page.
1188 */
1189
1190 pv_initpage = (struct pv_page *) uvm_km_alloc(kernel_map, PAGE_SIZE);
1191 if (pv_initpage == NULL)
1192 panic("pmap_init: pv_initpage");
1193 pv_cachedva = 0; /* a VA we have allocated but not used yet */
1194 pv_nfpvents = 0;
1195 (void) pmap_add_pvpage(pv_initpage, FALSE);
1196
1197 pmap_initialized = TRUE;
1198
1199 /* Initialise our L1 page table queues and counters */
1200 SIMPLEQ_INIT(&l1pt_static_queue);
1201 l1pt_static_queue_count = 0;
1202 l1pt_static_create_count = 0;
1203 SIMPLEQ_INIT(&l1pt_queue);
1204 l1pt_queue_count = 0;
1205 l1pt_create_count = 0;
1206 l1pt_reuse_count = 0;
1207 }
1208
1209 /*
1210 * pmap_postinit()
1211 *
1212 * This routine is called after the vm and kmem subsystems have been
1213 * initialised. This allows the pmap code to perform any initialisation
1214 * that can only be done one the memory allocation is in place.
1215 */
1216
1217 void
1218 pmap_postinit()
1219 {
1220 int loop;
1221 struct l1pt *pt;
1222
1223 #ifdef PMAP_STATIC_L1S
1224 for (loop = 0; loop < PMAP_STATIC_L1S; ++loop) {
1225 #else /* PMAP_STATIC_L1S */
1226 for (loop = 0; loop < max_processes; ++loop) {
1227 #endif /* PMAP_STATIC_L1S */
1228 /* Allocate a L1 page table */
1229 pt = pmap_alloc_l1pt();
1230 if (!pt)
1231 panic("Cannot allocate static L1 page tables\n");
1232
1233 /* Clean it */
1234 bzero((void *)pt->pt_va, PD_SIZE);
1235 pt->pt_flags |= (PTFLAG_STATIC | PTFLAG_CLEAN);
1236 /* Add the page table to the queue */
1237 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pt, pt_queue);
1238 ++l1pt_static_queue_count;
1239 ++l1pt_static_create_count;
1240 }
1241 }
1242
1243
1244 /*
1245 * Create and return a physical map.
1246 *
1247 * If the size specified for the map is zero, the map is an actual physical
1248 * map, and may be referenced by the hardware.
1249 *
1250 * If the size specified is non-zero, the map will be used in software only,
1251 * and is bounded by that size.
1252 */
1253
1254 pmap_t
1255 pmap_create()
1256 {
1257 struct pmap *pmap;
1258
1259 /*
1260 * Fetch pmap entry from the pool
1261 */
1262
1263 pmap = pool_get(&pmap_pmap_pool, PR_WAITOK);
1264 /* XXX is this really needed! */
1265 memset(pmap, 0, sizeof(*pmap));
1266
1267 simple_lock_init(&pmap->pm_obj.vmobjlock);
1268 pmap->pm_obj.pgops = NULL; /* currently not a mappable object */
1269 TAILQ_INIT(&pmap->pm_obj.memq);
1270 pmap->pm_obj.uo_npages = 0;
1271 pmap->pm_obj.uo_refs = 1;
1272 pmap->pm_stats.wired_count = 0;
1273 pmap->pm_stats.resident_count = 1;
1274
1275 /* Now init the machine part of the pmap */
1276 pmap_pinit(pmap);
1277 return(pmap);
1278 }
1279
1280 /*
1281 * pmap_alloc_l1pt()
1282 *
1283 * This routine allocates physical and virtual memory for a L1 page table
1284 * and wires it.
1285 * A l1pt structure is returned to describe the allocated page table.
1286 *
1287 * This routine is allowed to fail if the required memory cannot be allocated.
1288 * In this case NULL is returned.
1289 */
1290
1291 struct l1pt *
1292 pmap_alloc_l1pt(void)
1293 {
1294 paddr_t pa;
1295 vaddr_t va;
1296 struct l1pt *pt;
1297 int error;
1298 struct vm_page *m;
1299 pt_entry_t *ptes;
1300
1301 /* Allocate virtual address space for the L1 page table */
1302 va = uvm_km_valloc(kernel_map, PD_SIZE);
1303 if (va == 0) {
1304 #ifdef DIAGNOSTIC
1305 PDEBUG(0,
1306 printf("pmap: Cannot allocate pageable memory for L1\n"));
1307 #endif /* DIAGNOSTIC */
1308 return(NULL);
1309 }
1310
1311 /* Allocate memory for the l1pt structure */
1312 pt = (struct l1pt *)malloc(sizeof(struct l1pt), M_VMPMAP, M_WAITOK);
1313
1314 /*
1315 * Allocate pages from the VM system.
1316 */
1317 TAILQ_INIT(&pt->pt_plist);
1318 error = uvm_pglistalloc(PD_SIZE, physical_start, physical_end,
1319 PD_SIZE, 0, &pt->pt_plist, 1, M_WAITOK);
1320 if (error) {
1321 #ifdef DIAGNOSTIC
1322 PDEBUG(0,
1323 printf("pmap: Cannot allocate physical mem for L1 (%d)\n",
1324 error));
1325 #endif /* DIAGNOSTIC */
1326 /* Release the resources we already have claimed */
1327 free(pt, M_VMPMAP);
1328 uvm_km_free(kernel_map, va, PD_SIZE);
1329 return(NULL);
1330 }
1331
1332 /* Map our physical pages into our virtual space */
1333 pt->pt_va = va;
1334 m = pt->pt_plist.tqh_first;
1335 ptes = pmap_map_ptes(pmap_kernel());
1336 while (m && va < (pt->pt_va + PD_SIZE)) {
1337 pa = VM_PAGE_TO_PHYS(m);
1338
1339 pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE);
1340
1341 /* Revoke cacheability and bufferability */
1342 /* XXX should be done better than this */
1343 ptes[arm_byte_to_page(va)] &= ~(PT_C | PT_B);
1344
1345 va += NBPG;
1346 m = m->pageq.tqe_next;
1347 }
1348 pmap_unmap_ptes(pmap_kernel());
1349 pmap_update(pmap_kernel());
1350
1351 #ifdef DIAGNOSTIC
1352 if (m)
1353 panic("pmap_alloc_l1pt: pglist not empty\n");
1354 #endif /* DIAGNOSTIC */
1355
1356 pt->pt_flags = 0;
1357 return(pt);
1358 }
1359
1360 /*
1361 * Free a L1 page table previously allocated with pmap_alloc_l1pt().
1362 */
1363 static void
1364 pmap_free_l1pt(pt)
1365 struct l1pt *pt;
1366 {
1367 /* Separate the physical memory for the virtual space */
1368 pmap_kremove(pt->pt_va, PD_SIZE);
1369 pmap_update(pmap_kernel());
1370
1371 /* Return the physical memory */
1372 uvm_pglistfree(&pt->pt_plist);
1373
1374 /* Free the virtual space */
1375 uvm_km_free(kernel_map, pt->pt_va, PD_SIZE);
1376
1377 /* Free the l1pt structure */
1378 free(pt, M_VMPMAP);
1379 }
1380
1381 /*
1382 * Allocate a page directory.
1383 * This routine will either allocate a new page directory from the pool
1384 * of L1 page tables currently held by the kernel or it will allocate
1385 * a new one via pmap_alloc_l1pt().
1386 * It will then initialise the l1 page table for use.
1387 */
1388 static int
1389 pmap_allocpagedir(pmap)
1390 struct pmap *pmap;
1391 {
1392 paddr_t pa;
1393 struct l1pt *pt;
1394 pt_entry_t *pte;
1395
1396 PDEBUG(0, printf("pmap_allocpagedir(%p)\n", pmap));
1397
1398 /* Do we have any spare L1's lying around ? */
1399 if (l1pt_static_queue_count) {
1400 --l1pt_static_queue_count;
1401 pt = l1pt_static_queue.sqh_first;
1402 SIMPLEQ_REMOVE_HEAD(&l1pt_static_queue, pt, pt_queue);
1403 } else if (l1pt_queue_count) {
1404 --l1pt_queue_count;
1405 pt = l1pt_queue.sqh_first;
1406 SIMPLEQ_REMOVE_HEAD(&l1pt_queue, pt, pt_queue);
1407 ++l1pt_reuse_count;
1408 } else {
1409 pt = pmap_alloc_l1pt();
1410 if (!pt)
1411 return(ENOMEM);
1412 ++l1pt_create_count;
1413 }
1414
1415 /* Store the pointer to the l1 descriptor in the pmap. */
1416 pmap->pm_l1pt = pt;
1417
1418 /* Get the physical address of the start of the l1 */
1419 pa = VM_PAGE_TO_PHYS(pt->pt_plist.tqh_first);
1420
1421 /* Store the virtual address of the l1 in the pmap. */
1422 pmap->pm_pdir = (pd_entry_t *)pt->pt_va;
1423
1424 /* Clean the L1 if it is dirty */
1425 if (!(pt->pt_flags & PTFLAG_CLEAN))
1426 bzero((void *)pmap->pm_pdir, (PD_SIZE - KERNEL_PD_SIZE));
1427
1428 /* Do we already have the kernel mappings ? */
1429 if (!(pt->pt_flags & PTFLAG_KPT)) {
1430 /* Duplicate the kernel mapping i.e. all mappings 0xf0000000+ */
1431
1432 bcopy((char *)pmap_kernel()->pm_pdir + (PD_SIZE - KERNEL_PD_SIZE),
1433 (char *)pmap->pm_pdir + (PD_SIZE - KERNEL_PD_SIZE),
1434 KERNEL_PD_SIZE);
1435 pt->pt_flags |= PTFLAG_KPT;
1436 }
1437
1438 /* Allocate a page table to map all the page tables for this pmap */
1439
1440 #ifdef DIAGNOSTIC
1441 if (pmap->pm_vptpt) {
1442 /* XXX What if we have one already ? */
1443 panic("pmap_allocpagedir: have pt already\n");
1444 }
1445 #endif /* DIAGNOSTIC */
1446 pmap->pm_vptpt = uvm_km_zalloc(kernel_map, NBPG);
1447 if (pmap->pm_vptpt == 0) {
1448 pmap_freepagedir(pmap);
1449 return(ENOMEM);
1450 }
1451
1452 (void) pmap_extract(pmap_kernel(), pmap->pm_vptpt, &pmap->pm_pptpt);
1453 pmap->pm_pptpt &= PG_FRAME;
1454 /* Revoke cacheability and bufferability */
1455 /* XXX should be done better than this */
1456 pte = pmap_pte(pmap_kernel(), pmap->pm_vptpt);
1457 *pte = *pte & ~(PT_C | PT_B);
1458
1459 /* Wire in this page table */
1460 pmap_map_in_l1(pmap, PROCESS_PAGE_TBLS_BASE, pmap->pm_pptpt, TRUE);
1461
1462 pt->pt_flags &= ~PTFLAG_CLEAN; /* L1 is dirty now */
1463
1464 /*
1465 * Map the kernel page tables for 0xf0000000 +
1466 * into the page table used to map the
1467 * pmap's page tables
1468 */
1469 bcopy((char *)(PROCESS_PAGE_TBLS_BASE
1470 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2))
1471 + ((PD_SIZE - KERNEL_PD_SIZE) >> 2)),
1472 (char *)pmap->pm_vptpt + ((PD_SIZE - KERNEL_PD_SIZE) >> 2),
1473 (KERNEL_PD_SIZE >> 2));
1474
1475 return(0);
1476 }
1477
1478
1479 /*
1480 * Initialize a preallocated and zeroed pmap structure,
1481 * such as one in a vmspace structure.
1482 */
1483
1484 void
1485 pmap_pinit(pmap)
1486 struct pmap *pmap;
1487 {
1488 int backoff = 6;
1489 int retry = 10;
1490
1491 PDEBUG(0, printf("pmap_pinit(%p)\n", pmap));
1492
1493 /* Keep looping until we succeed in allocating a page directory */
1494 while (pmap_allocpagedir(pmap) != 0) {
1495 /*
1496 * Ok we failed to allocate a suitable block of memory for an
1497 * L1 page table. This means that either:
1498 * 1. 16KB of virtual address space could not be allocated
1499 * 2. 16KB of physically contiguous memory on a 16KB boundary
1500 * could not be allocated.
1501 *
1502 * Since we cannot fail we will sleep for a while and try
1503 * again.
1504 *
1505 * Searching for a suitable L1 PT is expensive:
1506 * to avoid hogging the system when memory is really
1507 * scarce, use an exponential back-off so that
1508 * eventually we won't retry more than once every 8
1509 * seconds. This should allow other processes to run
1510 * to completion and free up resources.
1511 */
1512 (void) ltsleep(&lbolt, PVM, "l1ptwait", (hz << 3) >> backoff,
1513 NULL);
1514 if (--retry == 0) {
1515 retry = 10;
1516 if (backoff)
1517 --backoff;
1518 }
1519 }
1520
1521 /* Map zero page for the pmap. This will also map the L2 for it */
1522 pmap_enter(pmap, 0x00000000, systempage.pv_pa,
1523 VM_PROT_READ, VM_PROT_READ | PMAP_WIRED);
1524 pmap_update(pmap);
1525 }
1526
1527
1528 void
1529 pmap_freepagedir(pmap)
1530 struct pmap *pmap;
1531 {
1532 /* Free the memory used for the page table mapping */
1533 if (pmap->pm_vptpt != 0)
1534 uvm_km_free(kernel_map, (vaddr_t)pmap->pm_vptpt, NBPG);
1535
1536 /* junk the L1 page table */
1537 if (pmap->pm_l1pt->pt_flags & PTFLAG_STATIC) {
1538 /* Add the page table to the queue */
1539 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pmap->pm_l1pt, pt_queue);
1540 ++l1pt_static_queue_count;
1541 } else if (l1pt_queue_count < 8) {
1542 /* Add the page table to the queue */
1543 SIMPLEQ_INSERT_TAIL(&l1pt_queue, pmap->pm_l1pt, pt_queue);
1544 ++l1pt_queue_count;
1545 } else
1546 pmap_free_l1pt(pmap->pm_l1pt);
1547 }
1548
1549
1550 /*
1551 * Retire the given physical map from service.
1552 * Should only be called if the map contains no valid mappings.
1553 */
1554
1555 void
1556 pmap_destroy(pmap)
1557 struct pmap *pmap;
1558 {
1559 struct vm_page *page;
1560 int count;
1561
1562 if (pmap == NULL)
1563 return;
1564
1565 PDEBUG(0, printf("pmap_destroy(%p)\n", pmap));
1566
1567 /*
1568 * Drop reference count
1569 */
1570 simple_lock(&pmap->pm_obj.vmobjlock);
1571 count = --pmap->pm_obj.uo_refs;
1572 simple_unlock(&pmap->pm_obj.vmobjlock);
1573 if (count > 0) {
1574 return;
1575 }
1576
1577 /*
1578 * reference count is zero, free pmap resources and then free pmap.
1579 */
1580
1581 /* Remove the zero page mapping */
1582 pmap_remove(pmap, 0x00000000, 0x00000000 + NBPG);
1583 pmap_update(pmap);
1584
1585 /*
1586 * Free any page tables still mapped
1587 * This is only temporay until pmap_enter can count the number
1588 * of mappings made in a page table. Then pmap_remove() can
1589 * reduce the count and free the pagetable when the count
1590 * reaches zero. Note that entries in this list should match the
1591 * contents of the ptpt, however this is faster than walking a 1024
1592 * entries looking for pt's
1593 * taken from i386 pmap.c
1594 */
1595 while (pmap->pm_obj.memq.tqh_first != NULL) {
1596 page = pmap->pm_obj.memq.tqh_first;
1597 #ifdef DIAGNOSTIC
1598 if (page->flags & PG_BUSY)
1599 panic("pmap_release: busy page table page");
1600 #endif
1601 /* pmap_page_protect? currently no need for it. */
1602
1603 page->wire_count = 0;
1604 uvm_pagefree(page);
1605 }
1606
1607 /* Free the page dir */
1608 pmap_freepagedir(pmap);
1609
1610 /* return the pmap to the pool */
1611 pool_put(&pmap_pmap_pool, pmap);
1612 }
1613
1614
1615 /*
1616 * void pmap_reference(struct pmap *pmap)
1617 *
1618 * Add a reference to the specified pmap.
1619 */
1620
1621 void
1622 pmap_reference(pmap)
1623 struct pmap *pmap;
1624 {
1625 if (pmap == NULL)
1626 return;
1627
1628 simple_lock(&pmap->pm_lock);
1629 pmap->pm_obj.uo_refs++;
1630 simple_unlock(&pmap->pm_lock);
1631 }
1632
1633 /*
1634 * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
1635 *
1636 * Return the start and end addresses of the kernel's virtual space.
1637 * These values are setup in pmap_bootstrap and are updated as pages
1638 * are allocated.
1639 */
1640
1641 void
1642 pmap_virtual_space(start, end)
1643 vaddr_t *start;
1644 vaddr_t *end;
1645 {
1646 *start = virtual_start;
1647 *end = virtual_end;
1648 }
1649
1650
1651 /*
1652 * Activate the address space for the specified process. If the process
1653 * is the current process, load the new MMU context.
1654 */
1655 void
1656 pmap_activate(p)
1657 struct proc *p;
1658 {
1659 struct pmap *pmap = p->p_vmspace->vm_map.pmap;
1660 struct pcb *pcb = &p->p_addr->u_pcb;
1661
1662 (void) pmap_extract(pmap_kernel(), (vaddr_t)pmap->pm_pdir,
1663 (paddr_t *)&pcb->pcb_pagedir);
1664
1665 PDEBUG(0, printf("pmap_activate: p=%p pmap=%p pcb=%p pdir=%p l1=%p\n",
1666 p, pmap, pcb, pmap->pm_pdir, pcb->pcb_pagedir));
1667
1668 if (p == curproc) {
1669 PDEBUG(0, printf("pmap_activate: setting TTB\n"));
1670 setttb((u_int)pcb->pcb_pagedir);
1671 }
1672 #if 0
1673 pmap->pm_pdchanged = FALSE;
1674 #endif
1675 }
1676
1677
1678 /*
1679 * Deactivate the address space of the specified process.
1680 */
1681 void
1682 pmap_deactivate(p)
1683 struct proc *p;
1684 {
1685 }
1686
1687 /*
1688 * Perform any deferred pmap operations.
1689 */
1690 void
1691 pmap_update(struct pmap *pmap)
1692 {
1693
1694 /*
1695 * We haven't deferred any pmap operations, but we do need to
1696 * make sure TLB/cache operations have completed.
1697 */
1698 cpu_cpwait();
1699 }
1700
1701 /*
1702 * pmap_clean_page()
1703 *
1704 * This is a local function used to work out the best strategy to clean
1705 * a single page referenced by its entry in the PV table. It's used by
1706 * pmap_copy_page, pmap_zero page and maybe some others later on.
1707 *
1708 * Its policy is effectively:
1709 * o If there are no mappings, we don't bother doing anything with the cache.
1710 * o If there is one mapping, we clean just that page.
1711 * o If there are multiple mappings, we clean the entire cache.
1712 *
1713 * So that some functions can be further optimised, it returns 0 if it didn't
1714 * clean the entire cache, or 1 if it did.
1715 *
1716 * XXX One bug in this routine is that if the pv_entry has a single page
1717 * mapped at 0x00000000 a whole cache clean will be performed rather than
1718 * just the 1 page. Since this should not occur in everyday use and if it does
1719 * it will just result in not the most efficient clean for the page.
1720 */
1721 static int
1722 pmap_clean_page(pv, is_src)
1723 struct pv_entry *pv;
1724 boolean_t is_src;
1725 {
1726 struct pmap *pmap;
1727 struct pv_entry *npv;
1728 int cache_needs_cleaning = 0;
1729 vaddr_t page_to_clean = 0;
1730
1731 if (pv == NULL)
1732 /* nothing mapped in so nothing to flush */
1733 return (0);
1734
1735 /* Since we flush the cache each time we change curproc, we
1736 * only need to flush the page if it is in the current pmap.
1737 */
1738 if (curproc)
1739 pmap = curproc->p_vmspace->vm_map.pmap;
1740 else
1741 pmap = pmap_kernel();
1742
1743 for (npv = pv; npv; npv = npv->pv_next) {
1744 if (npv->pv_pmap == pmap) {
1745 /* The page is mapped non-cacheable in
1746 * this map. No need to flush the cache.
1747 */
1748 if (npv->pv_flags & PT_NC) {
1749 #ifdef DIAGNOSTIC
1750 if (cache_needs_cleaning)
1751 panic("pmap_clean_page: "
1752 "cache inconsistency");
1753 #endif
1754 break;
1755 }
1756 #if 0
1757 /* This doesn't work, because pmap_protect
1758 doesn't flush changes on pages that it
1759 has write-protected. */
1760
1761 /* If the page is not writable and this
1762 is the source, then there is no need
1763 to flush it from the cache. */
1764 else if (is_src && ! (npv->pv_flags & PT_Wr))
1765 continue;
1766 #endif
1767 if (cache_needs_cleaning){
1768 page_to_clean = 0;
1769 break;
1770 }
1771 else
1772 page_to_clean = npv->pv_va;
1773 cache_needs_cleaning = 1;
1774 }
1775 }
1776
1777 if (page_to_clean)
1778 cpu_idcache_wbinv_range(page_to_clean, NBPG);
1779 else if (cache_needs_cleaning) {
1780 cpu_idcache_wbinv_all();
1781 return (1);
1782 }
1783 return (0);
1784 }
1785
1786 /*
1787 * pmap_zero_page()
1788 *
1789 * Zero a given physical page by mapping it at a page hook point.
1790 * In doing the zero page op, the page we zero is mapped cachable, as with
1791 * StrongARM accesses to non-cached pages are non-burst making writing
1792 * _any_ bulk data very slow.
1793 */
1794 void
1795 pmap_zero_page(phys)
1796 paddr_t phys;
1797 {
1798 struct vm_page *pg;
1799
1800 /* Get an entry for this page, and clean it it. */
1801 pg = PHYS_TO_VM_PAGE(phys);
1802 simple_lock(&pg->mdpage.pvh_slock);
1803 pmap_clean_page(pg->mdpage.pvh_list, FALSE);
1804 simple_unlock(&pg->mdpage.pvh_slock);
1805
1806 /*
1807 * Hook in the page, zero it, and purge the cache for that
1808 * zeroed page. Invalidate the TLB as needed.
1809 */
1810 *page_hook0.pte = L2_PTE(phys & PG_FRAME, AP_KRW);
1811 cpu_tlb_flushD_SE(page_hook0.va);
1812 cpu_cpwait();
1813 bzero_page(page_hook0.va);
1814 cpu_dcache_wbinv_range(page_hook0.va, NBPG);
1815 }
1816
1817 /* pmap_pageidlezero()
1818 *
1819 * The same as above, except that we assume that the page is not
1820 * mapped. This means we never have to flush the cache first. Called
1821 * from the idle loop.
1822 */
1823 boolean_t
1824 pmap_pageidlezero(phys)
1825 paddr_t phys;
1826 {
1827 int i, *ptr;
1828 boolean_t rv = TRUE;
1829
1830 #ifdef DIAGNOSTIC
1831 struct vm_page *pg;
1832
1833 pg = PHYS_TO_VM_PAGE(phys);
1834 if (pg->mdpage.pvh_list != NULL)
1835 panic("pmap_pageidlezero: zeroing mapped page\n");
1836 #endif
1837
1838 /*
1839 * Hook in the page, zero it, and purge the cache for that
1840 * zeroed page. Invalidate the TLB as needed.
1841 */
1842 *page_hook0.pte = L2_PTE(phys & PG_FRAME, AP_KRW);
1843 cpu_tlb_flushD_SE(page_hook0.va);
1844 cpu_cpwait();
1845
1846 for (i = 0, ptr = (int *)page_hook0.va;
1847 i < (NBPG / sizeof(int)); i++) {
1848 if (sched_whichqs != 0) {
1849 /*
1850 * A process has become ready. Abort now,
1851 * so we don't keep it waiting while we
1852 * do slow memory access to finish this
1853 * page.
1854 */
1855 rv = FALSE;
1856 break;
1857 }
1858 *ptr++ = 0;
1859 }
1860
1861 if (rv)
1862 /*
1863 * if we aborted we'll rezero this page again later so don't
1864 * purge it unless we finished it
1865 */
1866 cpu_dcache_wbinv_range(page_hook0.va, NBPG);
1867 return (rv);
1868 }
1869
1870 /*
1871 * pmap_copy_page()
1872 *
1873 * Copy one physical page into another, by mapping the pages into
1874 * hook points. The same comment regarding cachability as in
1875 * pmap_zero_page also applies here.
1876 */
1877 void
1878 pmap_copy_page(src, dest)
1879 paddr_t src;
1880 paddr_t dest;
1881 {
1882 struct vm_page *src_pg, *dest_pg;
1883 boolean_t cleanedcache;
1884
1885 /* Get PV entries for the pages, and clean them if needed. */
1886 src_pg = PHYS_TO_VM_PAGE(src);
1887
1888 simple_lock(&src_pg->mdpage.pvh_slock);
1889 cleanedcache = pmap_clean_page(src_pg->mdpage.pvh_list, TRUE);
1890 simple_unlock(&src_pg->mdpage.pvh_slock);
1891
1892 if (cleanedcache == 0) {
1893 dest_pg = PHYS_TO_VM_PAGE(dest);
1894 simple_lock(&dest_pg->mdpage.pvh_slock);
1895 pmap_clean_page(dest_pg->mdpage.pvh_list, FALSE);
1896 simple_unlock(&dest_pg->mdpage.pvh_slock);
1897 }
1898 /*
1899 * Map the pages into the page hook points, copy them, and purge
1900 * the cache for the appropriate page. Invalidate the TLB
1901 * as required.
1902 */
1903 *page_hook0.pte = L2_PTE(src & PG_FRAME, AP_KRW);
1904 *page_hook1.pte = L2_PTE(dest & PG_FRAME, AP_KRW);
1905 cpu_tlb_flushD_SE(page_hook0.va);
1906 cpu_tlb_flushD_SE(page_hook1.va);
1907 cpu_cpwait();
1908 bcopy_page(page_hook0.va, page_hook1.va);
1909 cpu_dcache_wbinv_range(page_hook0.va, NBPG);
1910 cpu_dcache_wbinv_range(page_hook1.va, NBPG);
1911 }
1912
1913 #if 0
1914 void
1915 pmap_pte_addref(pmap, va)
1916 struct pmap *pmap;
1917 vaddr_t va;
1918 {
1919 pd_entry_t *pde;
1920 paddr_t pa;
1921 struct vm_page *m;
1922
1923 if (pmap == pmap_kernel())
1924 return;
1925
1926 pde = pmap_pde(pmap, va & ~(3 << PDSHIFT));
1927 pa = pmap_pte_pa(pde);
1928 m = PHYS_TO_VM_PAGE(pa);
1929 ++m->wire_count;
1930 #ifdef MYCROFT_HACK
1931 printf("addref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
1932 pmap, va, pde, pa, m, m->wire_count);
1933 #endif
1934 }
1935
1936 void
1937 pmap_pte_delref(pmap, va)
1938 struct pmap *pmap;
1939 vaddr_t va;
1940 {
1941 pd_entry_t *pde;
1942 paddr_t pa;
1943 struct vm_page *m;
1944
1945 if (pmap == pmap_kernel())
1946 return;
1947
1948 pde = pmap_pde(pmap, va & ~(3 << PDSHIFT));
1949 pa = pmap_pte_pa(pde);
1950 m = PHYS_TO_VM_PAGE(pa);
1951 --m->wire_count;
1952 #ifdef MYCROFT_HACK
1953 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
1954 pmap, va, pde, pa, m, m->wire_count);
1955 #endif
1956 if (m->wire_count == 0) {
1957 #ifdef MYCROFT_HACK
1958 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p\n",
1959 pmap, va, pde, pa, m);
1960 #endif
1961 pmap_unmap_in_l1(pmap, va);
1962 uvm_pagefree(m);
1963 --pmap->pm_stats.resident_count;
1964 }
1965 }
1966 #else
1967 #define pmap_pte_addref(pmap, va)
1968 #define pmap_pte_delref(pmap, va)
1969 #endif
1970
1971 /*
1972 * Since we have a virtually indexed cache, we may need to inhibit caching if
1973 * there is more than one mapping and at least one of them is writable.
1974 * Since we purge the cache on every context switch, we only need to check for
1975 * other mappings within the same pmap, or kernel_pmap.
1976 * This function is also called when a page is unmapped, to possibly reenable
1977 * caching on any remaining mappings.
1978 *
1979 * The code implements the following logic, where:
1980 *
1981 * KW = # of kernel read/write pages
1982 * KR = # of kernel read only pages
1983 * UW = # of user read/write pages
1984 * UR = # of user read only pages
1985 * OW = # of user read/write pages in another pmap, then
1986 *
1987 * KC = kernel mapping is cacheable
1988 * UC = user mapping is cacheable
1989 *
1990 * KW=0,KR=0 KW=0,KR>0 KW=1,KR=0 KW>1,KR>=0
1991 * +---------------------------------------------
1992 * UW=0,UR=0,OW=0 | --- KC=1 KC=1 KC=0
1993 * UW=0,UR>0,OW=0 | UC=1 KC=1,UC=1 KC=0,UC=0 KC=0,UC=0
1994 * UW=0,UR>0,OW>0 | UC=1 KC=0,UC=1 KC=0,UC=0 KC=0,UC=0
1995 * UW=1,UR=0,OW=0 | UC=1 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0
1996 * UW>1,UR>=0,OW>=0 | UC=0 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0
1997 *
1998 * Note that the pmap must have it's ptes mapped in, and passed with ptes.
1999 */
2000 __inline static void
2001 pmap_vac_me_harder(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2002 boolean_t clear_cache)
2003 {
2004 if (pmap == pmap_kernel())
2005 pmap_vac_me_kpmap(pmap, pg, ptes, clear_cache);
2006 else
2007 pmap_vac_me_user(pmap, pg, ptes, clear_cache);
2008 }
2009
2010 static void
2011 pmap_vac_me_kpmap(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2012 boolean_t clear_cache)
2013 {
2014 int user_entries = 0;
2015 int user_writable = 0;
2016 int user_cacheable = 0;
2017 int kernel_entries = 0;
2018 int kernel_writable = 0;
2019 int kernel_cacheable = 0;
2020 struct pv_entry *pv;
2021 struct pmap *last_pmap = pmap;
2022
2023 #ifdef DIAGNOSTIC
2024 if (pmap != pmap_kernel())
2025 panic("pmap_vac_me_kpmap: pmap != pmap_kernel()");
2026 #endif
2027
2028 /*
2029 * Pass one, see if there are both kernel and user pmaps for
2030 * this page. Calculate whether there are user-writable or
2031 * kernel-writable pages.
2032 */
2033 for (pv = pg->mdpage.pvh_list; pv != NULL; pv = pv->pv_next) {
2034 if (pv->pv_pmap != pmap) {
2035 user_entries++;
2036 if (pv->pv_flags & PT_Wr)
2037 user_writable++;
2038 if ((pv->pv_flags & PT_NC) == 0)
2039 user_cacheable++;
2040 } else {
2041 kernel_entries++;
2042 if (pv->pv_flags & PT_Wr)
2043 kernel_writable++;
2044 if ((pv->pv_flags & PT_NC) == 0)
2045 kernel_cacheable++;
2046 }
2047 }
2048
2049 /*
2050 * We know we have just been updating a kernel entry, so if
2051 * all user pages are already cacheable, then there is nothing
2052 * further to do.
2053 */
2054 if (kernel_entries == 0 &&
2055 user_cacheable == user_entries)
2056 return;
2057
2058 if (user_entries) {
2059 /*
2060 * Scan over the list again, for each entry, if it
2061 * might not be set correctly, call pmap_vac_me_user
2062 * to recalculate the settings.
2063 */
2064 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
2065 /*
2066 * We know kernel mappings will get set
2067 * correctly in other calls. We also know
2068 * that if the pmap is the same as last_pmap
2069 * then we've just handled this entry.
2070 */
2071 if (pv->pv_pmap == pmap || pv->pv_pmap == last_pmap)
2072 continue;
2073 /*
2074 * If there are kernel entries and this page
2075 * is writable but non-cacheable, then we can
2076 * skip this entry also.
2077 */
2078 if (kernel_entries > 0 &&
2079 (pv->pv_flags & (PT_NC | PT_Wr)) ==
2080 (PT_NC | PT_Wr))
2081 continue;
2082 /*
2083 * Similarly if there are no kernel-writable
2084 * entries and the page is already
2085 * read-only/cacheable.
2086 */
2087 if (kernel_writable == 0 &&
2088 (pv->pv_flags & (PT_NC | PT_Wr)) == 0)
2089 continue;
2090 /*
2091 * For some of the remaining cases, we know
2092 * that we must recalculate, but for others we
2093 * can't tell if they are correct or not, so
2094 * we recalculate anyway.
2095 */
2096 pmap_unmap_ptes(last_pmap);
2097 last_pmap = pv->pv_pmap;
2098 ptes = pmap_map_ptes(last_pmap);
2099 pmap_vac_me_user(last_pmap, pg, ptes,
2100 pmap_is_curpmap(last_pmap));
2101 }
2102 /* Restore the pte mapping that was passed to us. */
2103 if (last_pmap != pmap) {
2104 pmap_unmap_ptes(last_pmap);
2105 ptes = pmap_map_ptes(pmap);
2106 }
2107 if (kernel_entries == 0)
2108 return;
2109 }
2110
2111 pmap_vac_me_user(pmap, pg, ptes, clear_cache);
2112 return;
2113 }
2114
2115 static void
2116 pmap_vac_me_user(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2117 boolean_t clear_cache)
2118 {
2119 struct pmap *kpmap = pmap_kernel();
2120 struct pv_entry *pv, *npv;
2121 int entries = 0;
2122 int writable = 0;
2123 int cacheable_entries = 0;
2124 int kern_cacheable = 0;
2125 int other_writable = 0;
2126
2127 pv = pg->mdpage.pvh_list;
2128 KASSERT(ptes != NULL);
2129
2130 /*
2131 * Count mappings and writable mappings in this pmap.
2132 * Include kernel mappings as part of our own.
2133 * Keep a pointer to the first one.
2134 */
2135 for (npv = pv; npv; npv = npv->pv_next) {
2136 /* Count mappings in the same pmap */
2137 if (pmap == npv->pv_pmap ||
2138 kpmap == npv->pv_pmap) {
2139 if (entries++ == 0)
2140 pv = npv;
2141 /* Cacheable mappings */
2142 if ((npv->pv_flags & PT_NC) == 0) {
2143 cacheable_entries++;
2144 if (kpmap == npv->pv_pmap)
2145 kern_cacheable++;
2146 }
2147 /* Writable mappings */
2148 if (npv->pv_flags & PT_Wr)
2149 ++writable;
2150 } else if (npv->pv_flags & PT_Wr)
2151 other_writable = 1;
2152 }
2153
2154 PDEBUG(3,printf("pmap_vac_me_harder: pmap %p Entries %d, "
2155 "writable %d cacheable %d %s\n", pmap, entries, writable,
2156 cacheable_entries, clear_cache ? "clean" : "no clean"));
2157
2158 /*
2159 * Enable or disable caching as necessary.
2160 * Note: the first entry might be part of the kernel pmap,
2161 * so we can't assume this is indicative of the state of the
2162 * other (maybe non-kpmap) entries.
2163 */
2164 if ((entries > 1 && writable) ||
2165 (entries > 0 && pmap == kpmap && other_writable)) {
2166 if (cacheable_entries == 0)
2167 return;
2168 for (npv = pv; npv; npv = npv->pv_next) {
2169 if ((pmap == npv->pv_pmap
2170 || kpmap == npv->pv_pmap) &&
2171 (npv->pv_flags & PT_NC) == 0) {
2172 ptes[arm_byte_to_page(npv->pv_va)] &=
2173 ~(PT_C | PT_B);
2174 npv->pv_flags |= PT_NC;
2175 /*
2176 * If this page needs flushing from the
2177 * cache, and we aren't going to do it
2178 * below, do it now.
2179 */
2180 if ((cacheable_entries < 4 &&
2181 (clear_cache || npv->pv_pmap == kpmap)) ||
2182 (npv->pv_pmap == kpmap &&
2183 !clear_cache && kern_cacheable < 4)) {
2184 cpu_idcache_wbinv_range(npv->pv_va,
2185 NBPG);
2186 cpu_tlb_flushID_SE(npv->pv_va);
2187 }
2188 }
2189 }
2190 if ((clear_cache && cacheable_entries >= 4) ||
2191 kern_cacheable >= 4) {
2192 cpu_idcache_wbinv_all();
2193 cpu_tlb_flushID();
2194 }
2195 cpu_cpwait();
2196 } else if (entries > 0) {
2197 /*
2198 * Turn cacheing back on for some pages. If it is a kernel
2199 * page, only do so if there are no other writable pages.
2200 */
2201 for (npv = pv; npv; npv = npv->pv_next) {
2202 if ((pmap == npv->pv_pmap ||
2203 (kpmap == npv->pv_pmap && other_writable == 0)) &&
2204 (npv->pv_flags & PT_NC)) {
2205 ptes[arm_byte_to_page(npv->pv_va)] |=
2206 pte_cache_mode;
2207 npv->pv_flags &= ~PT_NC;
2208 }
2209 }
2210 }
2211 }
2212
2213 /*
2214 * pmap_remove()
2215 *
2216 * pmap_remove is responsible for nuking a number of mappings for a range
2217 * of virtual address space in the current pmap. To do this efficiently
2218 * is interesting, because in a number of cases a wide virtual address
2219 * range may be supplied that contains few actual mappings. So, the
2220 * optimisations are:
2221 * 1. Try and skip over hunks of address space for which an L1 entry
2222 * does not exist.
2223 * 2. Build up a list of pages we've hit, up to a maximum, so we can
2224 * maybe do just a partial cache clean. This path of execution is
2225 * complicated by the fact that the cache must be flushed _before_
2226 * the PTE is nuked, being a VAC :-)
2227 * 3. Maybe later fast-case a single page, but I don't think this is
2228 * going to make _that_ much difference overall.
2229 */
2230
2231 #define PMAP_REMOVE_CLEAN_LIST_SIZE 3
2232
2233 void
2234 pmap_remove(pmap, sva, eva)
2235 struct pmap *pmap;
2236 vaddr_t sva;
2237 vaddr_t eva;
2238 {
2239 int cleanlist_idx = 0;
2240 struct pagelist {
2241 vaddr_t va;
2242 pt_entry_t *pte;
2243 } cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
2244 pt_entry_t *pte = 0, *ptes;
2245 paddr_t pa;
2246 int pmap_active;
2247 struct vm_page *pg;
2248
2249 /* Exit quick if there is no pmap */
2250 if (!pmap)
2251 return;
2252
2253 PDEBUG(0, printf("pmap_remove: pmap=%p sva=%08lx eva=%08lx\n", pmap, sva, eva));
2254
2255 sva &= PG_FRAME;
2256 eva &= PG_FRAME;
2257
2258 /*
2259 * we lock in the pmap => vm_page direction
2260 */
2261 PMAP_MAP_TO_HEAD_LOCK();
2262
2263 ptes = pmap_map_ptes(pmap);
2264 /* Get a page table pointer */
2265 while (sva < eva) {
2266 if (pmap_pde_page(pmap_pde(pmap, sva)))
2267 break;
2268 sva = (sva & PD_MASK) + NBPD;
2269 }
2270
2271 pte = &ptes[arm_byte_to_page(sva)];
2272 /* Note if the pmap is active thus require cache and tlb cleans */
2273 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap)
2274 || (pmap == pmap_kernel()))
2275 pmap_active = 1;
2276 else
2277 pmap_active = 0;
2278
2279 /* Now loop along */
2280 while (sva < eva) {
2281 /* Check if we can move to the next PDE (l1 chunk) */
2282 if (!(sva & PT_MASK))
2283 if (!pmap_pde_page(pmap_pde(pmap, sva))) {
2284 sva += NBPD;
2285 pte += arm_byte_to_page(NBPD);
2286 continue;
2287 }
2288
2289 /* We've found a valid PTE, so this page of PTEs has to go. */
2290 if (pmap_pte_v(pte)) {
2291 /* Update statistics */
2292 --pmap->pm_stats.resident_count;
2293
2294 /*
2295 * Add this page to our cache remove list, if we can.
2296 * If, however the cache remove list is totally full,
2297 * then do a complete cache invalidation taking note
2298 * to backtrack the PTE table beforehand, and ignore
2299 * the lists in future because there's no longer any
2300 * point in bothering with them (we've paid the
2301 * penalty, so will carry on unhindered). Otherwise,
2302 * when we fall out, we just clean the list.
2303 */
2304 PDEBUG(10, printf("remove: inv pte at %p(%x) ", pte, *pte));
2305 pa = pmap_pte_pa(pte);
2306
2307 if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
2308 /* Add to the clean list. */
2309 cleanlist[cleanlist_idx].pte = pte;
2310 cleanlist[cleanlist_idx].va = sva;
2311 cleanlist_idx++;
2312 } else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
2313 int cnt;
2314
2315 /* Nuke everything if needed. */
2316 if (pmap_active) {
2317 cpu_idcache_wbinv_all();
2318 cpu_tlb_flushID();
2319 }
2320
2321 /*
2322 * Roll back the previous PTE list,
2323 * and zero out the current PTE.
2324 */
2325 for (cnt = 0; cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
2326 *cleanlist[cnt].pte = 0;
2327 pmap_pte_delref(pmap, cleanlist[cnt].va);
2328 }
2329 *pte = 0;
2330 pmap_pte_delref(pmap, sva);
2331 cleanlist_idx++;
2332 } else {
2333 /*
2334 * We've already nuked the cache and
2335 * TLB, so just carry on regardless,
2336 * and we won't need to do it again
2337 */
2338 *pte = 0;
2339 pmap_pte_delref(pmap, sva);
2340 }
2341
2342 /*
2343 * Update flags. In a number of circumstances,
2344 * we could cluster a lot of these and do a
2345 * number of sequential pages in one go.
2346 */
2347 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
2348 struct pv_entry *pve;
2349 simple_lock(&pg->mdpage.pvh_slock);
2350 pve = pmap_remove_pv(pg, pmap, sva);
2351 pmap_free_pv(pmap, pve);
2352 pmap_vac_me_harder(pmap, pg, ptes, FALSE);
2353 simple_unlock(&pg->mdpage.pvh_slock);
2354 }
2355 }
2356 sva += NBPG;
2357 pte++;
2358 }
2359
2360 pmap_unmap_ptes(pmap);
2361 /*
2362 * Now, if we've fallen through down to here, chances are that there
2363 * are less than PMAP_REMOVE_CLEAN_LIST_SIZE mappings left.
2364 */
2365 if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
2366 u_int cnt;
2367
2368 for (cnt = 0; cnt < cleanlist_idx; cnt++) {
2369 if (pmap_active) {
2370 cpu_idcache_wbinv_range(cleanlist[cnt].va,
2371 NBPG);
2372 *cleanlist[cnt].pte = 0;
2373 cpu_tlb_flushID_SE(cleanlist[cnt].va);
2374 } else
2375 *cleanlist[cnt].pte = 0;
2376 pmap_pte_delref(pmap, cleanlist[cnt].va);
2377 }
2378 }
2379 PMAP_MAP_TO_HEAD_UNLOCK();
2380 }
2381
2382 /*
2383 * Routine: pmap_remove_all
2384 * Function:
2385 * Removes this physical page from
2386 * all physical maps in which it resides.
2387 * Reflects back modify bits to the pager.
2388 */
2389
2390 static void
2391 pmap_remove_all(pg)
2392 struct vm_page *pg;
2393 {
2394 struct pv_entry *pv, *npv;
2395 struct pmap *pmap;
2396 pt_entry_t *pte, *ptes;
2397
2398 PDEBUG(0, printf("pmap_remove_all: pa=%lx ", VM_PAGE_TO_PHYS(pg)));
2399
2400 /* set vm_page => pmap locking */
2401 PMAP_HEAD_TO_MAP_LOCK();
2402
2403 simple_lock(&pg->mdpage.pvh_slock);
2404
2405 pv = pg->mdpage.pvh_list;
2406 if (pv == NULL) {
2407 PDEBUG(0, printf("free page\n"));
2408 simple_unlock(&pg->mdpage.pvh_slock);
2409 PMAP_HEAD_TO_MAP_UNLOCK();
2410 return;
2411 }
2412 pmap_clean_page(pv, FALSE);
2413
2414 while (pv) {
2415 pmap = pv->pv_pmap;
2416 ptes = pmap_map_ptes(pmap);
2417 pte = &ptes[arm_byte_to_page(pv->pv_va)];
2418
2419 PDEBUG(0, printf("[%p,%08x,%08lx,%08x] ", pmap, *pte,
2420 pv->pv_va, pv->pv_flags));
2421 #ifdef DEBUG
2422 if (!pmap_pde_page(pmap_pde(pmap, pv->pv_va)) ||
2423 !pmap_pte_v(pte) || pmap_pte_pa(pte) != pa)
2424 panic("pmap_remove_all: bad mapping");
2425 #endif /* DEBUG */
2426
2427 /*
2428 * Update statistics
2429 */
2430 --pmap->pm_stats.resident_count;
2431
2432 /* Wired bit */
2433 if (pv->pv_flags & PT_W)
2434 --pmap->pm_stats.wired_count;
2435
2436 /*
2437 * Invalidate the PTEs.
2438 * XXX: should cluster them up and invalidate as many
2439 * as possible at once.
2440 */
2441
2442 #ifdef needednotdone
2443 reduce wiring count on page table pages as references drop
2444 #endif
2445
2446 *pte = 0;
2447 pmap_pte_delref(pmap, pv->pv_va);
2448
2449 npv = pv->pv_next;
2450 pmap_free_pv(pmap, pv);
2451 pv = npv;
2452 pmap_unmap_ptes(pmap);
2453 }
2454 pg->mdpage.pvh_list = NULL;
2455 simple_unlock(&pg->mdpage.pvh_slock);
2456 PMAP_HEAD_TO_MAP_UNLOCK();
2457
2458 PDEBUG(0, printf("done\n"));
2459 cpu_tlb_flushID();
2460 cpu_cpwait();
2461 }
2462
2463
2464 /*
2465 * Set the physical protection on the specified range of this map as requested.
2466 */
2467
2468 void
2469 pmap_protect(pmap, sva, eva, prot)
2470 struct pmap *pmap;
2471 vaddr_t sva;
2472 vaddr_t eva;
2473 vm_prot_t prot;
2474 {
2475 pt_entry_t *pte = NULL, *ptes;
2476 struct vm_page *pg;
2477 int armprot;
2478 int flush = 0;
2479 paddr_t pa;
2480
2481 PDEBUG(0, printf("pmap_protect: pmap=%p %08lx->%08lx %x\n",
2482 pmap, sva, eva, prot));
2483
2484 if (~prot & VM_PROT_READ) {
2485 /* Just remove the mappings. */
2486 pmap_remove(pmap, sva, eva);
2487 /* pmap_update not needed as it should be called by the caller
2488 * of pmap_protect */
2489 return;
2490 }
2491 if (prot & VM_PROT_WRITE) {
2492 /*
2493 * If this is a read->write transition, just ignore it and let
2494 * uvm_fault() take care of it later.
2495 */
2496 return;
2497 }
2498
2499 sva &= PG_FRAME;
2500 eva &= PG_FRAME;
2501
2502 /* Need to lock map->head */
2503 PMAP_MAP_TO_HEAD_LOCK();
2504
2505 ptes = pmap_map_ptes(pmap);
2506 /*
2507 * We need to acquire a pointer to a page table page before entering
2508 * the following loop.
2509 */
2510 while (sva < eva) {
2511 if (pmap_pde_page(pmap_pde(pmap, sva)))
2512 break;
2513 sva = (sva & PD_MASK) + NBPD;
2514 }
2515
2516 pte = &ptes[arm_byte_to_page(sva)];
2517
2518 while (sva < eva) {
2519 /* only check once in a while */
2520 if ((sva & PT_MASK) == 0) {
2521 if (!pmap_pde_page(pmap_pde(pmap, sva))) {
2522 /* We can race ahead here, to the next pde. */
2523 sva += NBPD;
2524 pte += arm_byte_to_page(NBPD);
2525 continue;
2526 }
2527 }
2528
2529 if (!pmap_pte_v(pte))
2530 goto next;
2531
2532 flush = 1;
2533
2534 armprot = 0;
2535 if (sva < VM_MAXUSER_ADDRESS)
2536 armprot |= PT_AP(AP_U);
2537 else if (sva < VM_MAX_ADDRESS)
2538 armprot |= PT_AP(AP_W); /* XXX Ekk what is this ? */
2539 *pte = (*pte & 0xfffff00f) | armprot;
2540
2541 pa = pmap_pte_pa(pte);
2542
2543 /* Get the physical page index */
2544
2545 /* Clear write flag */
2546 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
2547 simple_lock(&pg->mdpage.pvh_slock);
2548 (void) pmap_modify_pv(pmap, sva, pg, PT_Wr, 0);
2549 pmap_vac_me_harder(pmap, pg, ptes, FALSE);
2550 simple_unlock(&pg->mdpage.pvh_slock);
2551 }
2552
2553 next:
2554 sva += NBPG;
2555 pte++;
2556 }
2557 pmap_unmap_ptes(pmap);
2558 PMAP_MAP_TO_HEAD_UNLOCK();
2559 if (flush)
2560 cpu_tlb_flushID();
2561 }
2562
2563 /*
2564 * void pmap_enter(struct pmap *pmap, vaddr_t va, paddr_t pa, vm_prot_t prot,
2565 * int flags)
2566 *
2567 * Insert the given physical page (p) at
2568 * the specified virtual address (v) in the
2569 * target physical map with the protection requested.
2570 *
2571 * If specified, the page will be wired down, meaning
2572 * that the related pte can not be reclaimed.
2573 *
2574 * NB: This is the only routine which MAY NOT lazy-evaluate
2575 * or lose information. That is, this routine must actually
2576 * insert this page into the given map NOW.
2577 */
2578
2579 int
2580 pmap_enter(pmap, va, pa, prot, flags)
2581 struct pmap *pmap;
2582 vaddr_t va;
2583 paddr_t pa;
2584 vm_prot_t prot;
2585 int flags;
2586 {
2587 pt_entry_t *pte, *ptes;
2588 u_int npte;
2589 paddr_t opa;
2590 int nflags;
2591 boolean_t wired = (flags & PMAP_WIRED) != 0;
2592 struct vm_page *pg;
2593 struct pv_entry *pve;
2594 int error;
2595
2596 PDEBUG(5, printf("pmap_enter: V%08lx P%08lx in pmap %p prot=%08x, wired = %d\n",
2597 va, pa, pmap, prot, wired));
2598
2599 #ifdef DIAGNOSTIC
2600 /* Valid address ? */
2601 if (va >= (KERNEL_VM_BASE + KERNEL_VM_SIZE))
2602 panic("pmap_enter: too big");
2603 if (pmap != pmap_kernel() && va != 0) {
2604 if (va < VM_MIN_ADDRESS || va >= VM_MAXUSER_ADDRESS)
2605 panic("pmap_enter: kernel page in user map");
2606 } else {
2607 if (va >= VM_MIN_ADDRESS && va < VM_MAXUSER_ADDRESS)
2608 panic("pmap_enter: user page in kernel map");
2609 if (va >= VM_MAXUSER_ADDRESS && va < VM_MAX_ADDRESS)
2610 panic("pmap_enter: entering PT page");
2611 }
2612 #endif
2613 /*
2614 * Get a pointer to the page. Later on in this function, we
2615 * test for a managed page by checking pg != NULL.
2616 */
2617 pg = PHYS_TO_VM_PAGE(pa);
2618
2619 /* get lock */
2620 PMAP_MAP_TO_HEAD_LOCK();
2621 /*
2622 * Get a pointer to the pte for this virtual address. If the
2623 * pte pointer is NULL then we are missing the L2 page table
2624 * so we need to create one.
2625 */
2626 /* XXX horrible hack to get us working with lockdebug */
2627 simple_lock(&pmap->pm_obj.vmobjlock);
2628 pte = pmap_pte(pmap, va);
2629 if (!pte) {
2630 struct vm_page *ptp;
2631
2632 /* if failure is allowed then don't try too hard */
2633 ptp = pmap_get_ptp(pmap, va, flags & PMAP_CANFAIL);
2634 if (ptp == NULL) {
2635 if (flags & PMAP_CANFAIL) {
2636 error = ENOMEM;
2637 goto out;
2638 }
2639 panic("pmap_enter: get ptp failed");
2640 }
2641
2642 pte = pmap_pte(pmap, va);
2643 #ifdef DIAGNOSTIC
2644 if (!pte)
2645 panic("pmap_enter: no pte");
2646 #endif
2647 }
2648
2649 nflags = 0;
2650 if (prot & VM_PROT_WRITE)
2651 nflags |= PT_Wr;
2652 if (wired)
2653 nflags |= PT_W;
2654
2655 /* More debugging info */
2656 PDEBUG(5, printf("pmap_enter: pte for V%08lx = V%p (%08x)\n", va, pte,
2657 *pte));
2658
2659 /* Is the pte valid ? If so then this page is already mapped */
2660 if (pmap_pte_v(pte)) {
2661 /* Get the physical address of the current page mapped */
2662 opa = pmap_pte_pa(pte);
2663
2664 #ifdef MYCROFT_HACK
2665 printf("pmap_enter: pmap=%p va=%lx pa=%lx opa=%lx\n", pmap, va, pa, opa);
2666 #endif
2667
2668 /* Are we mapping the same page ? */
2669 if (opa == pa) {
2670 /* All we must be doing is changing the protection */
2671 PDEBUG(0, printf("Case 02 in pmap_enter (V%08lx P%08lx)\n",
2672 va, pa));
2673
2674 /* Has the wiring changed ? */
2675 if (pg != NULL) {
2676 simple_lock(&pg->mdpage.pvh_slock);
2677 (void) pmap_modify_pv(pmap, va, pg,
2678 PT_Wr | PT_W, nflags);
2679 simple_unlock(&pg->mdpage.pvh_slock);
2680 }
2681 } else {
2682 /* We are replacing the page with a new one. */
2683 cpu_idcache_wbinv_range(va, NBPG);
2684
2685 PDEBUG(0, printf("Case 03 in pmap_enter (V%08lx P%08lx P%08lx)\n",
2686 va, pa, opa));
2687
2688 /*
2689 * If it is part of our managed memory then we
2690 * must remove it from the PV list
2691 */
2692 if (pg != NULL) {
2693 simple_lock(&pg->mdpage.pvh_slock);
2694 pve = pmap_remove_pv(pg, pmap, va);
2695 simple_unlock(&pg->mdpage.pvh_slock);
2696 } else {
2697 pve = NULL;
2698 }
2699
2700 goto enter;
2701 }
2702 } else {
2703 opa = 0;
2704 pve = NULL;
2705 pmap_pte_addref(pmap, va);
2706
2707 /* pte is not valid so we must be hooking in a new page */
2708 ++pmap->pm_stats.resident_count;
2709
2710 enter:
2711 /*
2712 * Enter on the PV list if part of our managed memory
2713 */
2714 if (pmap_initialized && pg != NULL) {
2715 if (pve == NULL) {
2716 pve = pmap_alloc_pv(pmap, ALLOCPV_NEED);
2717 if (pve == NULL) {
2718 if (flags & PMAP_CANFAIL) {
2719 error = ENOMEM;
2720 goto out;
2721 }
2722 panic("pmap_enter: no pv entries available");
2723 }
2724 }
2725 /* enter_pv locks pvh when adding */
2726 pmap_enter_pv(pg, pve, pmap, va, NULL, nflags);
2727 } else {
2728 pg = NULL;
2729 if (pve != NULL)
2730 pmap_free_pv(pmap, pve);
2731 }
2732 }
2733
2734 #ifdef MYCROFT_HACK
2735 if (mycroft_hack)
2736 printf("pmap_enter: pmap=%p va=%lx pa=%lx opa=%lx bank=%d off=%d pv=%p\n", pmap, va, pa, opa, bank, off, pv);
2737 #endif
2738
2739 /* Construct the pte, giving the correct access. */
2740 npte = (pa & PG_FRAME);
2741
2742 /* VA 0 is magic. */
2743 if (pmap != pmap_kernel() && va != 0)
2744 npte |= PT_AP(AP_U);
2745
2746 if (pmap_initialized && pg != NULL) {
2747 #ifdef DIAGNOSTIC
2748 if ((flags & VM_PROT_ALL) & ~prot)
2749 panic("pmap_enter: access_type exceeds prot");
2750 #endif
2751 npte |= pte_cache_mode;
2752 if (flags & VM_PROT_WRITE) {
2753 npte |= L2_SPAGE | PT_AP(AP_W);
2754 pg->mdpage.pvh_attrs |= PT_H | PT_M;
2755 } else if (flags & VM_PROT_ALL) {
2756 npte |= L2_SPAGE;
2757 pg->mdpage.pvh_attrs |= PT_H;
2758 } else
2759 npte |= L2_INVAL;
2760 } else {
2761 if (prot & VM_PROT_WRITE)
2762 npte |= L2_SPAGE | PT_AP(AP_W);
2763 else if (prot & VM_PROT_ALL)
2764 npte |= L2_SPAGE;
2765 else
2766 npte |= L2_INVAL;
2767 }
2768
2769 #ifdef MYCROFT_HACK
2770 if (mycroft_hack)
2771 printf("pmap_enter: pmap=%p va=%lx pa=%lx prot=%x wired=%d access_type=%x npte=%08x\n", pmap, va, pa, prot, wired, flags & VM_PROT_ALL, npte);
2772 #endif
2773
2774 *pte = npte;
2775
2776 if (pmap_initialized && pg != NULL) {
2777 boolean_t pmap_active = FALSE;
2778 /* XXX this will change once the whole of pmap_enter uses
2779 * map_ptes
2780 */
2781 ptes = pmap_map_ptes(pmap);
2782 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap)
2783 || (pmap == pmap_kernel()))
2784 pmap_active = TRUE;
2785 simple_lock(&pg->mdpage.pvh_slock);
2786 pmap_vac_me_harder(pmap, pg, ptes, pmap_active);
2787 simple_unlock(&pg->mdpage.pvh_slock);
2788 pmap_unmap_ptes(pmap);
2789 }
2790
2791 /* Better flush the TLB ... */
2792 cpu_tlb_flushID_SE(va);
2793 error = 0;
2794 out:
2795 simple_unlock(&pmap->pm_obj.vmobjlock);
2796 PMAP_MAP_TO_HEAD_UNLOCK();
2797 PDEBUG(5, printf("pmap_enter: pte = V%p %08x\n", pte, *pte));
2798
2799 return error;
2800 }
2801
2802 void
2803 pmap_kenter_pa(va, pa, prot)
2804 vaddr_t va;
2805 paddr_t pa;
2806 vm_prot_t prot;
2807 {
2808 struct pmap *pmap = pmap_kernel();
2809 pt_entry_t *pte;
2810 struct vm_page *pg;
2811
2812 if (!pmap_pde_page(pmap_pde(pmap, va))) {
2813
2814 #ifdef DIAGNOSTIC
2815 if (pmap_pde_v(pmap_pde(pmap, va)))
2816 panic("Trying to map kernel page into section mapping"
2817 " VA=%lx PA=%lx", va, pa);
2818 #endif
2819 /*
2820 * For the kernel pmaps it would be better to ensure
2821 * that they are always present, and to grow the
2822 * kernel as required.
2823 */
2824
2825 /* must lock the pmap */
2826 simple_lock(&(pmap_kernel()->pm_obj.vmobjlock));
2827 /* Allocate a page table */
2828 pg = uvm_pagealloc(&(pmap_kernel()->pm_obj), 0, NULL,
2829 UVM_PGA_USERESERVE | UVM_PGA_ZERO);
2830 if (pg == NULL) {
2831 panic("pmap_kenter_pa: no free pages");
2832 }
2833 pg->flags &= ~PG_BUSY; /* never busy */
2834
2835 /* Wire this page table into the L1. */
2836 pmap_map_in_l1(pmap, va, VM_PAGE_TO_PHYS(pg), TRUE);
2837 simple_unlock(&(pmap_kernel()->pm_obj.vmobjlock));
2838 }
2839 pte = vtopte(va);
2840 KASSERT(!pmap_pte_v(pte));
2841 *pte = L2_PTE(pa, AP_KRW);
2842 }
2843
2844 void
2845 pmap_kremove(va, len)
2846 vaddr_t va;
2847 vsize_t len;
2848 {
2849 pt_entry_t *pte;
2850
2851 for (len >>= PAGE_SHIFT; len > 0; len--, va += PAGE_SIZE) {
2852
2853 /*
2854 * We assume that we will only be called with small
2855 * regions of memory.
2856 */
2857
2858 KASSERT(pmap_pde_page(pmap_pde(pmap_kernel(), va)));
2859 pte = vtopte(va);
2860 cpu_idcache_wbinv_range(va, PAGE_SIZE);
2861 *pte = 0;
2862 cpu_tlb_flushID_SE(va);
2863 }
2864 }
2865
2866 /*
2867 * pmap_page_protect:
2868 *
2869 * Lower the permission for all mappings to a given page.
2870 */
2871
2872 void
2873 pmap_page_protect(pg, prot)
2874 struct vm_page *pg;
2875 vm_prot_t prot;
2876 {
2877
2878 PDEBUG(0, printf("pmap_page_protect(pa=%lx, prot=%d)\n",
2879 VM_PAGE_TO_PHYS(pg), prot));
2880
2881 switch(prot) {
2882 case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
2883 case VM_PROT_READ|VM_PROT_WRITE:
2884 return;
2885
2886 case VM_PROT_READ:
2887 case VM_PROT_READ|VM_PROT_EXECUTE:
2888 pmap_copy_on_write(pg);
2889 break;
2890
2891 default:
2892 pmap_remove_all(pg);
2893 break;
2894 }
2895 }
2896
2897
2898 /*
2899 * Routine: pmap_unwire
2900 * Function: Clear the wired attribute for a map/virtual-address
2901 * pair.
2902 * In/out conditions:
2903 * The mapping must already exist in the pmap.
2904 */
2905
2906 void
2907 pmap_unwire(pmap, va)
2908 struct pmap *pmap;
2909 vaddr_t va;
2910 {
2911 pt_entry_t *pte;
2912 paddr_t pa;
2913 struct vm_page *pg;
2914
2915 /*
2916 * Make sure pmap is valid. -dct
2917 */
2918 if (pmap == NULL)
2919 return;
2920
2921 /* Get the pte */
2922 pte = pmap_pte(pmap, va);
2923 if (!pte)
2924 return;
2925
2926 /* Extract the physical address of the page */
2927 pa = pmap_pte_pa(pte);
2928
2929 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
2930 return;
2931
2932 simple_lock(&pg->mdpage.pvh_slock);
2933 /* Update the wired bit in the pv entry for this page. */
2934 (void) pmap_modify_pv(pmap, va, pg, PT_W, 0);
2935 simple_unlock(&pg->mdpage.pvh_slock);
2936 }
2937
2938 /*
2939 * pt_entry_t *pmap_pte(struct pmap *pmap, vaddr_t va)
2940 *
2941 * Return the pointer to a page table entry corresponding to the supplied
2942 * virtual address.
2943 *
2944 * The page directory is first checked to make sure that a page table
2945 * for the address in question exists and if it does a pointer to the
2946 * entry is returned.
2947 *
2948 * The way this works is that that the kernel page tables are mapped
2949 * into the memory map at ALT_PAGE_TBLS_BASE to ALT_PAGE_TBLS_BASE+4MB.
2950 * This allows page tables to be located quickly.
2951 */
2952 pt_entry_t *
2953 pmap_pte(pmap, va)
2954 struct pmap *pmap;
2955 vaddr_t va;
2956 {
2957 pt_entry_t *ptp;
2958 pt_entry_t *result;
2959
2960 /* The pmap must be valid */
2961 if (!pmap)
2962 return(NULL);
2963
2964 /* Return the address of the pte */
2965 PDEBUG(10, printf("pmap_pte: pmap=%p va=V%08lx pde = V%p (%08X)\n",
2966 pmap, va, pmap_pde(pmap, va), *(pmap_pde(pmap, va))));
2967
2968 /* Do we have a valid pde ? If not we don't have a page table */
2969 if (!pmap_pde_page(pmap_pde(pmap, va))) {
2970 PDEBUG(0, printf("pmap_pte: failed - pde = %p\n",
2971 pmap_pde(pmap, va)));
2972 return(NULL);
2973 }
2974
2975 PDEBUG(10, printf("pmap pagetable = P%08lx current = P%08x\n",
2976 pmap->pm_pptpt, (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE
2977 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) +
2978 (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & PG_FRAME)));
2979
2980 /*
2981 * If the pmap is the kernel pmap or the pmap is the active one
2982 * then we can just return a pointer to entry relative to
2983 * PROCESS_PAGE_TBLS_BASE.
2984 * Otherwise we need to map the page tables to an alternative
2985 * address and reference them there.
2986 */
2987 if (pmap == pmap_kernel() || pmap->pm_pptpt
2988 == (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE
2989 + ((PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) &
2990 ~3) + (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & PG_FRAME)) {
2991 ptp = (pt_entry_t *)PROCESS_PAGE_TBLS_BASE;
2992 } else {
2993 struct proc *p = curproc;
2994
2995 /* If we don't have a valid curproc use proc0 */
2996 /* Perhaps we should just use kernel_pmap instead */
2997 if (p == NULL)
2998 p = &proc0;
2999 #ifdef DIAGNOSTIC
3000 /*
3001 * The pmap should always be valid for the process so
3002 * panic if it is not.
3003 */
3004 if (!p->p_vmspace || !p->p_vmspace->vm_map.pmap) {
3005 printf("pmap_pte: va=%08lx p=%p vm=%p\n",
3006 va, p, p->p_vmspace);
3007 console_debugger();
3008 }
3009 /*
3010 * The pmap for the current process should be mapped. If it
3011 * is not then we have a problem.
3012 */
3013 if (p->p_vmspace->vm_map.pmap->pm_pptpt !=
3014 (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE
3015 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) +
3016 (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & PG_FRAME)) {
3017 printf("pmap pagetable = P%08lx current = P%08x ",
3018 pmap->pm_pptpt, (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE
3019 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) +
3020 (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) &
3021 PG_FRAME));
3022 printf("pptpt=%lx\n", p->p_vmspace->vm_map.pmap->pm_pptpt);
3023 panic("pmap_pte: current and pmap mismatch\n");
3024 }
3025 #endif
3026
3027 ptp = (pt_entry_t *)ALT_PAGE_TBLS_BASE;
3028 pmap_map_in_l1(p->p_vmspace->vm_map.pmap, ALT_PAGE_TBLS_BASE,
3029 pmap->pm_pptpt, FALSE);
3030 cpu_tlb_flushD();
3031 cpu_cpwait();
3032 }
3033 PDEBUG(10, printf("page tables base = %p offset=%lx\n", ptp,
3034 ((va >> (PGSHIFT-2)) & ~3)));
3035 result = (pt_entry_t *)((char *)ptp + ((va >> (PGSHIFT-2)) & ~3));
3036 return(result);
3037 }
3038
3039 /*
3040 * Routine: pmap_extract
3041 * Function:
3042 * Extract the physical page address associated
3043 * with the given map/virtual_address pair.
3044 */
3045 boolean_t
3046 pmap_extract(pmap, va, pap)
3047 struct pmap *pmap;
3048 vaddr_t va;
3049 paddr_t *pap;
3050 {
3051 pd_entry_t *pde;
3052 pt_entry_t *pte, *ptes;
3053 paddr_t pa;
3054 boolean_t rv = TRUE;
3055
3056 PDEBUG(5, printf("pmap_extract: pmap=%p, va=V%08lx\n", pmap, va));
3057
3058 /*
3059 * Get the pte for this virtual address.
3060 */
3061 pde = pmap_pde(pmap, va);
3062 ptes = pmap_map_ptes(pmap);
3063 pte = &ptes[arm_byte_to_page(va)];
3064
3065 if (pmap_pde_section(pde)) {
3066 pa = (*pde & PD_MASK) | (va & (L1_SEC_SIZE - 1));
3067 goto out;
3068 } else if (pmap_pde_page(pde) == 0 || pmap_pte_v(pte) == 0) {
3069 rv = FALSE;
3070 goto out;
3071 }
3072
3073 if ((*pte & L2_MASK) == L2_LPAGE) {
3074 /* Extract the physical address from the pte */
3075 pa = *pte & ~(L2_LPAGE_SIZE - 1);
3076
3077 PDEBUG(5, printf("pmap_extract: LPAGE pa = P%08lx\n",
3078 (pa | (va & (L2_LPAGE_SIZE - 1)))));
3079
3080 if (pap != NULL)
3081 *pap = pa | (va & (L2_LPAGE_SIZE - 1));
3082 goto out;
3083 }
3084
3085 /* Extract the physical address from the pte */
3086 pa = pmap_pte_pa(pte);
3087
3088 PDEBUG(5, printf("pmap_extract: SPAGE pa = P%08lx\n",
3089 (pa | (va & ~PG_FRAME))));
3090
3091 if (pap != NULL)
3092 *pap = pa | (va & ~PG_FRAME);
3093 out:
3094 pmap_unmap_ptes(pmap);
3095 return (rv);
3096 }
3097
3098
3099 /*
3100 * Copy the range specified by src_addr/len from the source map to the
3101 * range dst_addr/len in the destination map.
3102 *
3103 * This routine is only advisory and need not do anything.
3104 */
3105
3106 void
3107 pmap_copy(dst_pmap, src_pmap, dst_addr, len, src_addr)
3108 struct pmap *dst_pmap;
3109 struct pmap *src_pmap;
3110 vaddr_t dst_addr;
3111 vsize_t len;
3112 vaddr_t src_addr;
3113 {
3114 PDEBUG(0, printf("pmap_copy(%p, %p, %lx, %lx, %lx)\n",
3115 dst_pmap, src_pmap, dst_addr, len, src_addr));
3116 }
3117
3118 #if defined(PMAP_DEBUG)
3119 void
3120 pmap_dump_pvlist(phys, m)
3121 vaddr_t phys;
3122 char *m;
3123 {
3124 struct vm_page *pg;
3125 struct pv_entry *pv;
3126
3127 if ((pg = PHYS_TO_VM_PAGE(phys)) == NULL) {
3128 printf("INVALID PA\n");
3129 return;
3130 }
3131 simple_lock(&pg->mdpage.pvh_slock);
3132 printf("%s %08lx:", m, phys);
3133 if (pg->mdpage.pvh_list == NULL) {
3134 printf(" no mappings\n");
3135 return;
3136 }
3137
3138 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next)
3139 printf(" pmap %p va %08lx flags %08x", pv->pv_pmap,
3140 pv->pv_va, pv->pv_flags);
3141
3142 printf("\n");
3143 simple_unlock(&pg->mdpage.pvh_slock);
3144 }
3145
3146 #endif /* PMAP_DEBUG */
3147
3148 __inline static boolean_t
3149 pmap_testbit(pg, setbits)
3150 struct vm_page *pg;
3151 unsigned int setbits;
3152 {
3153
3154 PDEBUG(1, printf("pmap_testbit: pa=%08lx set=%08x\n",
3155 VM_PAGE_TO_PHYS(pg), setbits));
3156
3157 /*
3158 * Check saved info only
3159 */
3160 if (pg->mdpage.pvh_attrs & setbits) {
3161 PDEBUG(0, printf("pmap_attributes = %02x\n",
3162 pg->mdpage.pvh_attrs));
3163 return(TRUE);
3164 }
3165
3166 return(FALSE);
3167 }
3168
3169 static pt_entry_t *
3170 pmap_map_ptes(struct pmap *pmap)
3171 {
3172 struct proc *p;
3173
3174 /* the kernel's pmap is always accessible */
3175 if (pmap == pmap_kernel()) {
3176 return (pt_entry_t *)PROCESS_PAGE_TBLS_BASE ;
3177 }
3178
3179 if (pmap_is_curpmap(pmap)) {
3180 simple_lock(&pmap->pm_obj.vmobjlock);
3181 return (pt_entry_t *)PROCESS_PAGE_TBLS_BASE;
3182 }
3183
3184 p = curproc;
3185
3186 if (p == NULL)
3187 p = &proc0;
3188
3189 /* need to lock both curpmap and pmap: use ordered locking */
3190 if ((unsigned) pmap < (unsigned) curproc->p_vmspace->vm_map.pmap) {
3191 simple_lock(&pmap->pm_obj.vmobjlock);
3192 simple_lock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3193 } else {
3194 simple_lock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3195 simple_lock(&pmap->pm_obj.vmobjlock);
3196 }
3197
3198 pmap_map_in_l1(p->p_vmspace->vm_map.pmap, ALT_PAGE_TBLS_BASE,
3199 pmap->pm_pptpt, FALSE);
3200 cpu_tlb_flushD();
3201 cpu_cpwait();
3202 return (pt_entry_t *)ALT_PAGE_TBLS_BASE;
3203 }
3204
3205 /*
3206 * pmap_unmap_ptes: unlock the PTE mapping of "pmap"
3207 */
3208
3209 static void
3210 pmap_unmap_ptes(pmap)
3211 struct pmap *pmap;
3212 {
3213 if (pmap == pmap_kernel()) {
3214 return;
3215 }
3216 if (pmap_is_curpmap(pmap)) {
3217 simple_unlock(&pmap->pm_obj.vmobjlock);
3218 } else {
3219 simple_unlock(&pmap->pm_obj.vmobjlock);
3220 simple_unlock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3221 }
3222 }
3223
3224 /*
3225 * Modify pte bits for all ptes corresponding to the given physical address.
3226 * We use `maskbits' rather than `clearbits' because we're always passing
3227 * constants and the latter would require an extra inversion at run-time.
3228 */
3229
3230 static void
3231 pmap_clearbit(pg, maskbits)
3232 struct vm_page *pg;
3233 unsigned int maskbits;
3234 {
3235 struct pv_entry *pv;
3236 pt_entry_t *pte;
3237 vaddr_t va;
3238 int tlbentry;
3239
3240 PDEBUG(1, printf("pmap_clearbit: pa=%08lx mask=%08x\n",
3241 VM_PAGE_TO_PHYS(pg), maskbits));
3242
3243 tlbentry = 0;
3244
3245 PMAP_HEAD_TO_MAP_LOCK();
3246 simple_lock(&pg->mdpage.pvh_slock);
3247
3248 /*
3249 * Clear saved attributes (modify, reference)
3250 */
3251 pg->mdpage.pvh_attrs &= ~maskbits;
3252
3253 if (pg->mdpage.pvh_list == NULL) {
3254 simple_unlock(&pg->mdpage.pvh_slock);
3255 PMAP_HEAD_TO_MAP_UNLOCK();
3256 return;
3257 }
3258
3259 /*
3260 * Loop over all current mappings setting/clearing as appropos
3261 */
3262 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
3263 va = pv->pv_va;
3264 pv->pv_flags &= ~maskbits;
3265 pte = pmap_pte(pv->pv_pmap, va);
3266 KASSERT(pte != NULL);
3267 if (maskbits & (PT_Wr|PT_M)) {
3268 if ((pv->pv_flags & PT_NC)) {
3269 /*
3270 * Entry is not cacheable: reenable
3271 * the cache, nothing to flush
3272 *
3273 * Don't turn caching on again if this
3274 * is a modified emulation. This
3275 * would be inconsitent with the
3276 * settings created by
3277 * pmap_vac_me_harder().
3278 *
3279 * There's no need to call
3280 * pmap_vac_me_harder() here: all
3281 * pages are loosing their write
3282 * permission.
3283 *
3284 */
3285 if (maskbits & PT_Wr) {
3286 *pte |= pte_cache_mode;
3287 pv->pv_flags &= ~PT_NC;
3288 }
3289 } else if (pmap_is_curpmap(pv->pv_pmap))
3290 /*
3291 * Entry is cacheable: check if pmap is
3292 * current if it is flush it,
3293 * otherwise it won't be in the cache
3294 */
3295 cpu_idcache_wbinv_range(pv->pv_va, NBPG);
3296
3297 /* make the pte read only */
3298 *pte &= ~PT_AP(AP_W);
3299 }
3300
3301 if (maskbits & PT_H)
3302 *pte = (*pte & ~L2_MASK) | L2_INVAL;
3303
3304 if (pmap_is_curpmap(pv->pv_pmap))
3305 /*
3306 * if we had cacheable pte's we'd clean the
3307 * pte out to memory here
3308 *
3309 * flush tlb entry as it's in the current pmap
3310 */
3311 cpu_tlb_flushID_SE(pv->pv_va);
3312 }
3313 cpu_cpwait();
3314
3315 simple_unlock(&pg->mdpage.pvh_slock);
3316 PMAP_HEAD_TO_MAP_UNLOCK();
3317 }
3318
3319
3320 boolean_t
3321 pmap_clear_modify(pg)
3322 struct vm_page *pg;
3323 {
3324 boolean_t rv;
3325
3326 PDEBUG(0, printf("pmap_clear_modify pa=%08lx\n", VM_PAGE_TO_PHYS(pg)));
3327 rv = pmap_testbit(pg, PT_M);
3328 pmap_clearbit(pg, PT_M);
3329 return rv;
3330 }
3331
3332
3333 boolean_t
3334 pmap_clear_reference(pg)
3335 struct vm_page *pg;
3336 {
3337 boolean_t rv;
3338
3339 PDEBUG(0, printf("pmap_clear_reference pa=%08lx\n",
3340 VM_PAGE_TO_PHYS(pg)));
3341 rv = pmap_testbit(pg, PT_H);
3342 pmap_clearbit(pg, PT_H);
3343 return rv;
3344 }
3345
3346
3347 void
3348 pmap_copy_on_write(pg)
3349 struct vm_page *pg;
3350 {
3351 PDEBUG(0, printf("pmap_copy_on_write pa=%08lx\n", VM_PAGE_TO_PHYS(pg)));
3352 pmap_clearbit(pg, PT_Wr);
3353 }
3354
3355
3356 boolean_t
3357 pmap_is_modified(pg)
3358 struct vm_page *pg;
3359 {
3360 boolean_t result;
3361
3362 result = pmap_testbit(pg, PT_M);
3363 PDEBUG(1, printf("pmap_is_modified pa=%08lx %x\n",
3364 VM_PAGE_TO_PHYS(pg), result));
3365 return (result);
3366 }
3367
3368
3369 boolean_t
3370 pmap_is_referenced(pg)
3371 struct vm_page *pg;
3372 {
3373 boolean_t result;
3374
3375 result = pmap_testbit(pg, PT_H);
3376 PDEBUG(0, printf("pmap_is_referenced pa=%08lx %x\n",
3377 VM_PAGE_TO_PHYS(pg), result));
3378 return (result);
3379 }
3380
3381
3382 int
3383 pmap_modified_emulation(pmap, va)
3384 struct pmap *pmap;
3385 vaddr_t va;
3386 {
3387 pt_entry_t *pte;
3388 paddr_t pa;
3389 struct vm_page *pg;
3390 u_int flags;
3391
3392 PDEBUG(2, printf("pmap_modified_emulation\n"));
3393
3394 /* Get the pte */
3395 pte = pmap_pte(pmap, va);
3396 if (!pte) {
3397 PDEBUG(2, printf("no pte\n"));
3398 return(0);
3399 }
3400
3401 PDEBUG(1, printf("*pte=%08x\n", *pte));
3402
3403 /* Check for a zero pte */
3404 if (*pte == 0)
3405 return(0);
3406
3407 /* This can happen if user code tries to access kernel memory. */
3408 if ((*pte & PT_AP(AP_W)) != 0)
3409 return (0);
3410
3411 /* Extract the physical address of the page */
3412 pa = pmap_pte_pa(pte);
3413 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3414 return(0);
3415
3416 /* Get the current flags for this page. */
3417 PMAP_HEAD_TO_MAP_LOCK();
3418 simple_lock(&pg->mdpage.pvh_slock);
3419
3420 flags = pmap_modify_pv(pmap, va, pg, 0, 0);
3421 PDEBUG(2, printf("pmap_modified_emulation: flags = %08x\n", flags));
3422
3423 /*
3424 * Do the flags say this page is writable ? If not then it is a
3425 * genuine write fault. If yes then the write fault is our fault
3426 * as we did not reflect the write access in the PTE. Now we know
3427 * a write has occurred we can correct this and also set the
3428 * modified bit
3429 */
3430 if (~flags & PT_Wr) {
3431 simple_unlock(&pg->mdpage.pvh_slock);
3432 PMAP_HEAD_TO_MAP_UNLOCK();
3433 return(0);
3434 }
3435
3436 PDEBUG(0, printf("pmap_modified_emulation: Got a hit va=%08lx, pte = %p (%08x)\n",
3437 va, pte, *pte));
3438 pg->mdpage.pvh_attrs |= PT_H | PT_M;
3439
3440 /*
3441 * Re-enable write permissions for the page. No need to call
3442 * pmap_vac_me_harder(), since this is just a
3443 * modified-emulation fault, and the PT_Wr bit isn't changing. We've
3444 * already set the cacheable bits based on the assumption that we
3445 * can write to this page.
3446 */
3447 *pte = (*pte & ~L2_MASK) | L2_SPAGE | PT_AP(AP_W);
3448 PDEBUG(0, printf("->(%08x)\n", *pte));
3449
3450 simple_unlock(&pg->mdpage.pvh_slock);
3451 PMAP_HEAD_TO_MAP_UNLOCK();
3452 /* Return, indicating the problem has been dealt with */
3453 cpu_tlb_flushID_SE(va);
3454 cpu_cpwait();
3455 return(1);
3456 }
3457
3458
3459 int
3460 pmap_handled_emulation(pmap, va)
3461 struct pmap *pmap;
3462 vaddr_t va;
3463 {
3464 pt_entry_t *pte;
3465 paddr_t pa;
3466 struct vm_page *pg;
3467
3468 PDEBUG(2, printf("pmap_handled_emulation\n"));
3469
3470 /* Get the pte */
3471 pte = pmap_pte(pmap, va);
3472 if (!pte) {
3473 PDEBUG(2, printf("no pte\n"));
3474 return(0);
3475 }
3476
3477 PDEBUG(1, printf("*pte=%08x\n", *pte));
3478
3479 /* Check for a zero pte */
3480 if (*pte == 0)
3481 return(0);
3482
3483 /* This can happen if user code tries to access kernel memory. */
3484 if ((*pte & L2_MASK) != L2_INVAL)
3485 return (0);
3486
3487 /* Extract the physical address of the page */
3488 pa = pmap_pte_pa(pte);
3489 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3490 return (0);
3491
3492 /*
3493 * Ok we just enable the pte and mark the attibs as handled
3494 */
3495 PDEBUG(0, printf("pmap_handled_emulation: Got a hit va=%08lx pte = %p (%08x)\n",
3496 va, pte, *pte));
3497 pg->mdpage.pvh_attrs |= PT_H;
3498 *pte = (*pte & ~L2_MASK) | L2_SPAGE;
3499 PDEBUG(0, printf("->(%08x)\n", *pte));
3500
3501 /* Return, indicating the problem has been dealt with */
3502 cpu_tlb_flushID_SE(va);
3503 cpu_cpwait();
3504 return(1);
3505 }
3506
3507
3508
3509
3510 /*
3511 * pmap_collect: free resources held by a pmap
3512 *
3513 * => optional function.
3514 * => called when a process is swapped out to free memory.
3515 */
3516
3517 void
3518 pmap_collect(pmap)
3519 struct pmap *pmap;
3520 {
3521 }
3522
3523 /*
3524 * Routine: pmap_procwr
3525 *
3526 * Function:
3527 * Synchronize caches corresponding to [addr, addr+len) in p.
3528 *
3529 */
3530 void
3531 pmap_procwr(p, va, len)
3532 struct proc *p;
3533 vaddr_t va;
3534 int len;
3535 {
3536 /* We only need to do anything if it is the current process. */
3537 if (p == curproc)
3538 cpu_icache_sync_range(va, len);
3539 }
3540 /*
3541 * PTP functions
3542 */
3543
3544 /*
3545 * pmap_steal_ptp: Steal a PTP from somewhere else.
3546 *
3547 * This is just a placeholder, for now we never steal.
3548 */
3549
3550 static struct vm_page *
3551 pmap_steal_ptp(struct pmap *pmap, vaddr_t va)
3552 {
3553 return (NULL);
3554 }
3555
3556 /*
3557 * pmap_get_ptp: get a PTP (if there isn't one, allocate a new one)
3558 *
3559 * => pmap should NOT be pmap_kernel()
3560 * => pmap should be locked
3561 */
3562
3563 static struct vm_page *
3564 pmap_get_ptp(struct pmap *pmap, vaddr_t va, boolean_t just_try)
3565 {
3566 struct vm_page *ptp;
3567
3568 if (pmap_pde_page(pmap_pde(pmap, va))) {
3569
3570 /* valid... check hint (saves us a PA->PG lookup) */
3571 #if 0
3572 if (pmap->pm_ptphint &&
3573 ((unsigned)pmap_pde(pmap, va) & PG_FRAME) ==
3574 VM_PAGE_TO_PHYS(pmap->pm_ptphint))
3575 return (pmap->pm_ptphint);
3576 #endif
3577 ptp = uvm_pagelookup(&pmap->pm_obj, va);
3578 #ifdef DIAGNOSTIC
3579 if (ptp == NULL)
3580 panic("pmap_get_ptp: unmanaged user PTP");
3581 #endif
3582 // pmap->pm_ptphint = ptp;
3583 return(ptp);
3584 }
3585
3586 /* allocate a new PTP (updates ptphint) */
3587 return(pmap_alloc_ptp(pmap, va, just_try));
3588 }
3589
3590 /*
3591 * pmap_alloc_ptp: allocate a PTP for a PMAP
3592 *
3593 * => pmap should already be locked by caller
3594 * => we use the ptp's wire_count to count the number of active mappings
3595 * in the PTP (we start it at one to prevent any chance this PTP
3596 * will ever leak onto the active/inactive queues)
3597 */
3598
3599 /*__inline */ static struct vm_page *
3600 pmap_alloc_ptp(struct pmap *pmap, vaddr_t va, boolean_t just_try)
3601 {
3602 struct vm_page *ptp;
3603
3604 ptp = uvm_pagealloc(&pmap->pm_obj, va, NULL,
3605 UVM_PGA_USERESERVE|UVM_PGA_ZERO);
3606 if (ptp == NULL) {
3607 if (just_try)
3608 return (NULL);
3609
3610 ptp = pmap_steal_ptp(pmap, va);
3611
3612 if (ptp == NULL)
3613 return (NULL);
3614 /* Stole a page, zero it. */
3615 pmap_zero_page(VM_PAGE_TO_PHYS(ptp));
3616 }
3617
3618 /* got one! */
3619 ptp->flags &= ~PG_BUSY; /* never busy */
3620 ptp->wire_count = 1; /* no mappings yet */
3621 pmap_map_in_l1(pmap, va, VM_PAGE_TO_PHYS(ptp), TRUE);
3622 pmap->pm_stats.resident_count++; /* count PTP as resident */
3623 // pmap->pm_ptphint = ptp;
3624 return (ptp);
3625 }
3626
3627 /* End of pmap.c */
3628