pmap.c revision 1.49 1 /* $NetBSD: pmap.c,v 1.49 2002/03/05 04:19:59 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 1.49 2002/03/05 04:19:59 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 void pmap_dump_pvlist(vaddr_t phys, char *m);
153
154 /*
155 * for switching to potentially finer grained debugging
156 */
157 #define PDB_FOLLOW 0x0001
158 #define PDB_INIT 0x0002
159 #define PDB_ENTER 0x0004
160 #define PDB_REMOVE 0x0008
161 #define PDB_CREATE 0x0010
162 #define PDB_PTPAGE 0x0020
163 #define PDB_GROWKERN 0x0040
164 #define PDB_BITS 0x0080
165 #define PDB_COLLECT 0x0100
166 #define PDB_PROTECT 0x0200
167 #define PDB_MAP_L1 0x0400
168 #define PDB_BOOTSTRAP 0x1000
169 #define PDB_PARANOIA 0x2000
170 #define PDB_WIRING 0x4000
171 #define PDB_PVDUMP 0x8000
172
173 int debugmap = 0;
174 int pmapdebug = PDB_PARANOIA | PDB_FOLLOW;
175 #define NPDEBUG(_lev_,_stat_) \
176 if (pmapdebug & (_lev_)) \
177 ((_stat_))
178
179 #else /* PMAP_DEBUG */
180 #define PDEBUG(_lev_,_stat_) /* Nothing */
181 #define NPDEBUG(_lev_,_stat_) /* Nothing */
182 #endif /* PMAP_DEBUG */
183
184 struct pmap kernel_pmap_store;
185
186 /*
187 * linked list of all non-kernel pmaps
188 */
189
190 static struct pmap_head pmaps;
191
192 /*
193 * pool that pmap structures are allocated from
194 */
195
196 struct pool pmap_pmap_pool;
197
198 pagehook_t page_hook0;
199 pagehook_t page_hook1;
200 char *memhook;
201 pt_entry_t msgbufpte;
202 extern caddr_t msgbufaddr;
203
204 boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
205 /*
206 * locking data structures
207 */
208
209 static struct lock pmap_main_lock;
210 static struct simplelock pvalloc_lock;
211 static struct simplelock pmaps_lock;
212 #ifdef LOCKDEBUG
213 #define PMAP_MAP_TO_HEAD_LOCK() \
214 (void) spinlockmgr(&pmap_main_lock, LK_SHARED, NULL)
215 #define PMAP_MAP_TO_HEAD_UNLOCK() \
216 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
217
218 #define PMAP_HEAD_TO_MAP_LOCK() \
219 (void) spinlockmgr(&pmap_main_lock, LK_EXCLUSIVE, NULL)
220 #define PMAP_HEAD_TO_MAP_UNLOCK() \
221 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
222 #else
223 #define PMAP_MAP_TO_HEAD_LOCK() /* nothing */
224 #define PMAP_MAP_TO_HEAD_UNLOCK() /* nothing */
225 #define PMAP_HEAD_TO_MAP_LOCK() /* nothing */
226 #define PMAP_HEAD_TO_MAP_UNLOCK() /* nothing */
227 #endif /* LOCKDEBUG */
228
229 /*
230 * pv_page management structures: locked by pvalloc_lock
231 */
232
233 TAILQ_HEAD(pv_pagelist, pv_page);
234 static struct pv_pagelist pv_freepages; /* list of pv_pages with free entrys */
235 static struct pv_pagelist pv_unusedpgs; /* list of unused pv_pages */
236 static int pv_nfpvents; /* # of free pv entries */
237 static struct pv_page *pv_initpage; /* bootstrap page from kernel_map */
238 static vaddr_t pv_cachedva; /* cached VA for later use */
239
240 #define PVE_LOWAT (PVE_PER_PVPAGE / 2) /* free pv_entry low water mark */
241 #define PVE_HIWAT (PVE_LOWAT + (PVE_PER_PVPAGE * 2))
242 /* high water mark */
243
244 /*
245 * local prototypes
246 */
247
248 static struct pv_entry *pmap_add_pvpage __P((struct pv_page *, boolean_t));
249 static struct pv_entry *pmap_alloc_pv __P((struct pmap *, int)); /* see codes below */
250 #define ALLOCPV_NEED 0 /* need PV now */
251 #define ALLOCPV_TRY 1 /* just try to allocate, don't steal */
252 #define ALLOCPV_NONEED 2 /* don't need PV, just growing cache */
253 static struct pv_entry *pmap_alloc_pvpage __P((struct pmap *, int));
254 static void pmap_enter_pv __P((struct vm_page *,
255 struct pv_entry *, struct pmap *,
256 vaddr_t, struct vm_page *, int));
257 static void pmap_free_pv __P((struct pmap *, struct pv_entry *));
258 static void pmap_free_pvs __P((struct pmap *, struct pv_entry *));
259 static void pmap_free_pv_doit __P((struct pv_entry *));
260 static void pmap_free_pvpage __P((void));
261 static boolean_t pmap_is_curpmap __P((struct pmap *));
262 static struct pv_entry *pmap_remove_pv __P((struct vm_page *, struct pmap *,
263 vaddr_t));
264 #define PMAP_REMOVE_ALL 0 /* remove all mappings */
265 #define PMAP_REMOVE_SKIPWIRED 1 /* skip wired mappings */
266
267 static u_int pmap_modify_pv __P((struct pmap *, vaddr_t, struct vm_page *,
268 u_int, u_int));
269
270 static void pmap_free_l1pt __P((struct l1pt *));
271 static int pmap_allocpagedir __P((struct pmap *));
272 static int pmap_clean_page __P((struct pv_entry *, boolean_t));
273 static void pmap_remove_all __P((struct vm_page *));
274
275
276 vsize_t npages;
277
278 static struct vm_page *pmap_alloc_ptp __P((struct pmap *, vaddr_t, boolean_t));
279 static struct vm_page *pmap_get_ptp __P((struct pmap *, vaddr_t, boolean_t));
280 __inline static void pmap_clearbit __P((struct vm_page *, unsigned int));
281 __inline static boolean_t pmap_testbit __P((struct vm_page *, unsigned int));
282
283 extern paddr_t physical_start;
284 extern paddr_t physical_freestart;
285 extern paddr_t physical_end;
286 extern paddr_t physical_freeend;
287 extern unsigned int free_pages;
288 extern int max_processes;
289
290 vaddr_t virtual_start;
291 vaddr_t virtual_end;
292 vaddr_t pmap_curmaxkvaddr;
293
294 vaddr_t avail_start;
295 vaddr_t avail_end;
296
297 extern pv_addr_t systempage;
298
299 #define ALLOC_PAGE_HOOK(x, s) \
300 x.va = virtual_start; \
301 x.pte = (pt_entry_t *)pmap_pte(pmap_kernel(), virtual_start); \
302 virtual_start += s;
303
304 /* Variables used by the L1 page table queue code */
305 SIMPLEQ_HEAD(l1pt_queue, l1pt);
306 struct l1pt_queue l1pt_static_queue; /* head of our static l1 queue */
307 int l1pt_static_queue_count; /* items in the static l1 queue */
308 int l1pt_static_create_count; /* static l1 items created */
309 struct l1pt_queue l1pt_queue; /* head of our l1 queue */
310 int l1pt_queue_count; /* items in the l1 queue */
311 int l1pt_create_count; /* stat - L1's create count */
312 int l1pt_reuse_count; /* stat - L1's reused count */
313
314 /* Local function prototypes (not used outside this file) */
315 pt_entry_t *pmap_pte __P((struct pmap *pmap, vaddr_t va));
316 void pmap_copy_on_write __P((struct vm_page *));
317 void pmap_pinit __P((struct pmap *));
318 void pmap_freepagedir __P((struct pmap *));
319
320 /* Other function prototypes */
321 extern void bzero_page __P((vaddr_t));
322 extern void bcopy_page __P((vaddr_t, vaddr_t));
323
324 struct l1pt *pmap_alloc_l1pt __P((void));
325 static __inline void pmap_map_in_l1 __P((struct pmap *pmap, vaddr_t va,
326 vaddr_t l2pa, boolean_t));
327
328 static pt_entry_t *pmap_map_ptes __P((struct pmap *));
329 static void pmap_unmap_ptes __P((struct pmap *));
330
331 __inline static void pmap_vac_me_harder __P((struct pmap *, struct vm_page *,
332 pt_entry_t *, boolean_t));
333 static void pmap_vac_me_kpmap __P((struct pmap *, struct vm_page *,
334 pt_entry_t *, boolean_t));
335 static void pmap_vac_me_user __P((struct pmap *, struct vm_page *,
336 pt_entry_t *, boolean_t));
337
338 /*
339 * Cache enable bits in PTE to use on pages that are cacheable.
340 * On most machines this is cacheable/bufferable, but on some, eg arm10, we
341 * can chose between write-through and write-back cacheing.
342 */
343 pt_entry_t pte_cache_mode = (PT_C | PT_B);
344
345 /*
346 * real definition of pv_entry.
347 */
348
349 struct pv_entry {
350 struct pv_entry *pv_next; /* next pv_entry */
351 struct pmap *pv_pmap; /* pmap where mapping lies */
352 vaddr_t pv_va; /* virtual address for mapping */
353 int pv_flags; /* flags */
354 struct vm_page *pv_ptp; /* vm_page for the ptp */
355 };
356
357 /*
358 * pv_entrys are dynamically allocated in chunks from a single page.
359 * we keep track of how many pv_entrys are in use for each page and
360 * we can free pv_entry pages if needed. there is one lock for the
361 * entire allocation system.
362 */
363
364 struct pv_page_info {
365 TAILQ_ENTRY(pv_page) pvpi_list;
366 struct pv_entry *pvpi_pvfree;
367 int pvpi_nfree;
368 };
369
370 /*
371 * number of pv_entry's in a pv_page
372 * (note: won't work on systems where NPBG isn't a constant)
373 */
374
375 #define PVE_PER_PVPAGE ((NBPG - sizeof(struct pv_page_info)) / \
376 sizeof(struct pv_entry))
377
378 /*
379 * a pv_page: where pv_entrys are allocated from
380 */
381
382 struct pv_page {
383 struct pv_page_info pvinfo;
384 struct pv_entry pvents[PVE_PER_PVPAGE];
385 };
386
387 #ifdef MYCROFT_HACK
388 int mycroft_hack = 0;
389 #endif
390
391 /* Function to set the debug level of the pmap code */
392
393 #ifdef PMAP_DEBUG
394 void
395 pmap_debug(level)
396 int level;
397 {
398 pmap_debug_level = level;
399 printf("pmap_debug: level=%d\n", pmap_debug_level);
400 }
401 #endif /* PMAP_DEBUG */
402
403 __inline static boolean_t
404 pmap_is_curpmap(struct pmap *pmap)
405 {
406 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap)
407 || (pmap == pmap_kernel()))
408 return (TRUE);
409 return (FALSE);
410 }
411 #include "isadma.h"
412
413 #if NISADMA > 0
414 /*
415 * Used to protect memory for ISA DMA bounce buffers. If, when loading
416 * pages into the system, memory intersects with any of these ranges,
417 * the intersecting memory will be loaded into a lower-priority free list.
418 */
419 bus_dma_segment_t *pmap_isa_dma_ranges;
420 int pmap_isa_dma_nranges;
421
422 boolean_t pmap_isa_dma_range_intersect __P((paddr_t, psize_t,
423 paddr_t *, psize_t *));
424
425 /*
426 * Check if a memory range intersects with an ISA DMA range, and
427 * return the page-rounded intersection if it does. The intersection
428 * will be placed on a lower-priority free list.
429 */
430 boolean_t
431 pmap_isa_dma_range_intersect(pa, size, pap, sizep)
432 paddr_t pa;
433 psize_t size;
434 paddr_t *pap;
435 psize_t *sizep;
436 {
437 bus_dma_segment_t *ds;
438 int i;
439
440 if (pmap_isa_dma_ranges == NULL)
441 return (FALSE);
442
443 for (i = 0, ds = pmap_isa_dma_ranges;
444 i < pmap_isa_dma_nranges; i++, ds++) {
445 if (ds->ds_addr <= pa && pa < (ds->ds_addr + ds->ds_len)) {
446 /*
447 * Beginning of region intersects with this range.
448 */
449 *pap = trunc_page(pa);
450 *sizep = round_page(min(pa + size,
451 ds->ds_addr + ds->ds_len) - pa);
452 return (TRUE);
453 }
454 if (pa < ds->ds_addr && ds->ds_addr < (pa + size)) {
455 /*
456 * End of region intersects with this range.
457 */
458 *pap = trunc_page(ds->ds_addr);
459 *sizep = round_page(min((pa + size) - ds->ds_addr,
460 ds->ds_len));
461 return (TRUE);
462 }
463 }
464
465 /*
466 * No intersection found.
467 */
468 return (FALSE);
469 }
470 #endif /* NISADMA > 0 */
471
472 /*
473 * p v _ e n t r y f u n c t i o n s
474 */
475
476 /*
477 * pv_entry allocation functions:
478 * the main pv_entry allocation functions are:
479 * pmap_alloc_pv: allocate a pv_entry structure
480 * pmap_free_pv: free one pv_entry
481 * pmap_free_pvs: free a list of pv_entrys
482 *
483 * the rest are helper functions
484 */
485
486 /*
487 * pmap_alloc_pv: inline function to allocate a pv_entry structure
488 * => we lock pvalloc_lock
489 * => if we fail, we call out to pmap_alloc_pvpage
490 * => 3 modes:
491 * ALLOCPV_NEED = we really need a pv_entry, even if we have to steal it
492 * ALLOCPV_TRY = we want a pv_entry, but not enough to steal
493 * ALLOCPV_NONEED = we are trying to grow our free list, don't really need
494 * one now
495 *
496 * "try" is for optional functions like pmap_copy().
497 */
498
499 __inline static struct pv_entry *
500 pmap_alloc_pv(pmap, mode)
501 struct pmap *pmap;
502 int mode;
503 {
504 struct pv_page *pvpage;
505 struct pv_entry *pv;
506
507 simple_lock(&pvalloc_lock);
508
509 if (pv_freepages.tqh_first != NULL) {
510 pvpage = pv_freepages.tqh_first;
511 pvpage->pvinfo.pvpi_nfree--;
512 if (pvpage->pvinfo.pvpi_nfree == 0) {
513 /* nothing left in this one? */
514 TAILQ_REMOVE(&pv_freepages, pvpage, pvinfo.pvpi_list);
515 }
516 pv = pvpage->pvinfo.pvpi_pvfree;
517 #ifdef DIAGNOSTIC
518 if (pv == NULL)
519 panic("pmap_alloc_pv: pvpi_nfree off");
520 #endif
521 pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
522 pv_nfpvents--; /* took one from pool */
523 } else {
524 pv = NULL; /* need more of them */
525 }
526
527 /*
528 * if below low water mark or we didn't get a pv_entry we try and
529 * create more pv_entrys ...
530 */
531
532 if (pv_nfpvents < PVE_LOWAT || pv == NULL) {
533 if (pv == NULL)
534 pv = pmap_alloc_pvpage(pmap, (mode == ALLOCPV_TRY) ?
535 mode : ALLOCPV_NEED);
536 else
537 (void) pmap_alloc_pvpage(pmap, ALLOCPV_NONEED);
538 }
539
540 simple_unlock(&pvalloc_lock);
541 return(pv);
542 }
543
544 /*
545 * pmap_alloc_pvpage: maybe allocate a new pvpage
546 *
547 * if need_entry is false: try and allocate a new pv_page
548 * if need_entry is true: try and allocate a new pv_page and return a
549 * new pv_entry from it. if we are unable to allocate a pv_page
550 * we make a last ditch effort to steal a pv_page from some other
551 * mapping. if that fails, we panic...
552 *
553 * => we assume that the caller holds pvalloc_lock
554 */
555
556 static struct pv_entry *
557 pmap_alloc_pvpage(pmap, mode)
558 struct pmap *pmap;
559 int mode;
560 {
561 struct vm_page *pg;
562 struct pv_page *pvpage;
563 struct pv_entry *pv;
564 int s;
565
566 /*
567 * if we need_entry and we've got unused pv_pages, allocate from there
568 */
569
570 if (mode != ALLOCPV_NONEED && pv_unusedpgs.tqh_first != NULL) {
571
572 /* move it to pv_freepages list */
573 pvpage = pv_unusedpgs.tqh_first;
574 TAILQ_REMOVE(&pv_unusedpgs, pvpage, pvinfo.pvpi_list);
575 TAILQ_INSERT_HEAD(&pv_freepages, pvpage, pvinfo.pvpi_list);
576
577 /* allocate a pv_entry */
578 pvpage->pvinfo.pvpi_nfree--; /* can't go to zero */
579 pv = pvpage->pvinfo.pvpi_pvfree;
580 #ifdef DIAGNOSTIC
581 if (pv == NULL)
582 panic("pmap_alloc_pvpage: pvpi_nfree off");
583 #endif
584 pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
585
586 pv_nfpvents--; /* took one from pool */
587 return(pv);
588 }
589
590 /*
591 * see if we've got a cached unmapped VA that we can map a page in.
592 * if not, try to allocate one.
593 */
594
595
596 if (pv_cachedva == 0) {
597 s = splvm();
598 pv_cachedva = uvm_km_kmemalloc(kmem_map, NULL,
599 PAGE_SIZE, UVM_KMF_TRYLOCK|UVM_KMF_VALLOC);
600 splx(s);
601 if (pv_cachedva == 0) {
602 return (NULL);
603 }
604 }
605
606 pg = uvm_pagealloc(NULL, pv_cachedva - vm_map_min(kernel_map), NULL,
607 UVM_PGA_USERESERVE);
608 if (pg)
609 pg->flags &= ~PG_BUSY; /* never busy */
610
611 if (pg == NULL)
612 return (NULL);
613
614 /*
615 * add a mapping for our new pv_page and free its entrys (save one!)
616 *
617 * NOTE: If we are allocating a PV page for the kernel pmap, the
618 * pmap is already locked! (...but entering the mapping is safe...)
619 */
620
621 pmap_kenter_pa(pv_cachedva, VM_PAGE_TO_PHYS(pg), VM_PROT_ALL);
622 pmap_update(pmap_kernel());
623 pvpage = (struct pv_page *) pv_cachedva;
624 pv_cachedva = 0;
625 return (pmap_add_pvpage(pvpage, mode != ALLOCPV_NONEED));
626 }
627
628 /*
629 * pmap_add_pvpage: add a pv_page's pv_entrys to the free list
630 *
631 * => caller must hold pvalloc_lock
632 * => if need_entry is true, we allocate and return one pv_entry
633 */
634
635 static struct pv_entry *
636 pmap_add_pvpage(pvp, need_entry)
637 struct pv_page *pvp;
638 boolean_t need_entry;
639 {
640 int tofree, lcv;
641
642 /* do we need to return one? */
643 tofree = (need_entry) ? PVE_PER_PVPAGE - 1 : PVE_PER_PVPAGE;
644
645 pvp->pvinfo.pvpi_pvfree = NULL;
646 pvp->pvinfo.pvpi_nfree = tofree;
647 for (lcv = 0 ; lcv < tofree ; lcv++) {
648 pvp->pvents[lcv].pv_next = pvp->pvinfo.pvpi_pvfree;
649 pvp->pvinfo.pvpi_pvfree = &pvp->pvents[lcv];
650 }
651 if (need_entry)
652 TAILQ_INSERT_TAIL(&pv_freepages, pvp, pvinfo.pvpi_list);
653 else
654 TAILQ_INSERT_TAIL(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
655 pv_nfpvents += tofree;
656 return((need_entry) ? &pvp->pvents[lcv] : NULL);
657 }
658
659 /*
660 * pmap_free_pv_doit: actually free a pv_entry
661 *
662 * => do not call this directly! instead use either
663 * 1. pmap_free_pv ==> free a single pv_entry
664 * 2. pmap_free_pvs => free a list of pv_entrys
665 * => we must be holding pvalloc_lock
666 */
667
668 __inline static void
669 pmap_free_pv_doit(pv)
670 struct pv_entry *pv;
671 {
672 struct pv_page *pvp;
673
674 pvp = (struct pv_page *) arm_trunc_page((vaddr_t)pv);
675 pv_nfpvents++;
676 pvp->pvinfo.pvpi_nfree++;
677
678 /* nfree == 1 => fully allocated page just became partly allocated */
679 if (pvp->pvinfo.pvpi_nfree == 1) {
680 TAILQ_INSERT_HEAD(&pv_freepages, pvp, pvinfo.pvpi_list);
681 }
682
683 /* free it */
684 pv->pv_next = pvp->pvinfo.pvpi_pvfree;
685 pvp->pvinfo.pvpi_pvfree = pv;
686
687 /*
688 * are all pv_page's pv_entry's free? move it to unused queue.
689 */
690
691 if (pvp->pvinfo.pvpi_nfree == PVE_PER_PVPAGE) {
692 TAILQ_REMOVE(&pv_freepages, pvp, pvinfo.pvpi_list);
693 TAILQ_INSERT_HEAD(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
694 }
695 }
696
697 /*
698 * pmap_free_pv: free a single pv_entry
699 *
700 * => we gain the pvalloc_lock
701 */
702
703 __inline static void
704 pmap_free_pv(pmap, pv)
705 struct pmap *pmap;
706 struct pv_entry *pv;
707 {
708 simple_lock(&pvalloc_lock);
709 pmap_free_pv_doit(pv);
710
711 /*
712 * Can't free the PV page if the PV entries were associated with
713 * the kernel pmap; the pmap is already locked.
714 */
715 if (pv_nfpvents > PVE_HIWAT && pv_unusedpgs.tqh_first != NULL &&
716 pmap != pmap_kernel())
717 pmap_free_pvpage();
718
719 simple_unlock(&pvalloc_lock);
720 }
721
722 /*
723 * pmap_free_pvs: free a list of pv_entrys
724 *
725 * => we gain the pvalloc_lock
726 */
727
728 __inline static void
729 pmap_free_pvs(pmap, pvs)
730 struct pmap *pmap;
731 struct pv_entry *pvs;
732 {
733 struct pv_entry *nextpv;
734
735 simple_lock(&pvalloc_lock);
736
737 for ( /* null */ ; pvs != NULL ; pvs = nextpv) {
738 nextpv = pvs->pv_next;
739 pmap_free_pv_doit(pvs);
740 }
741
742 /*
743 * Can't free the PV page if the PV entries were associated with
744 * the kernel pmap; the pmap is already locked.
745 */
746 if (pv_nfpvents > PVE_HIWAT && pv_unusedpgs.tqh_first != NULL &&
747 pmap != pmap_kernel())
748 pmap_free_pvpage();
749
750 simple_unlock(&pvalloc_lock);
751 }
752
753
754 /*
755 * pmap_free_pvpage: try and free an unused pv_page structure
756 *
757 * => assume caller is holding the pvalloc_lock and that
758 * there is a page on the pv_unusedpgs list
759 * => if we can't get a lock on the kmem_map we try again later
760 * => note: analysis of MI kmem_map usage [i.e. malloc/free] shows
761 * that if we can lock the kmem_map then we are not already
762 * holding kmem_object's lock.
763 */
764
765 static void
766 pmap_free_pvpage()
767 {
768 int s;
769 struct vm_map *map;
770 struct vm_map_entry *dead_entries;
771 struct pv_page *pvp;
772
773 s = splvm(); /* protect kmem_map */
774
775 pvp = pv_unusedpgs.tqh_first;
776
777 /*
778 * note: watch out for pv_initpage which is allocated out of
779 * kernel_map rather than kmem_map.
780 */
781 if (pvp == pv_initpage)
782 map = kernel_map;
783 else
784 map = kmem_map;
785
786 if (vm_map_lock_try(map)) {
787
788 /* remove pvp from pv_unusedpgs */
789 TAILQ_REMOVE(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
790
791 /* unmap the page */
792 dead_entries = NULL;
793 uvm_unmap_remove(map, (vaddr_t)pvp, ((vaddr_t)pvp) + PAGE_SIZE,
794 &dead_entries);
795 vm_map_unlock(map);
796
797 if (dead_entries != NULL)
798 uvm_unmap_detach(dead_entries, 0);
799
800 pv_nfpvents -= PVE_PER_PVPAGE; /* update free count */
801 }
802
803 if (pvp == pv_initpage)
804 /* no more initpage, we've freed it */
805 pv_initpage = NULL;
806
807 splx(s);
808 }
809
810 /*
811 * main pv_entry manipulation functions:
812 * pmap_enter_pv: enter a mapping onto a vm_page list
813 * pmap_remove_pv: remove a mappiing from a vm_page list
814 *
815 * NOTE: pmap_enter_pv expects to lock the pvh itself
816 * pmap_remove_pv expects te caller to lock the pvh before calling
817 */
818
819 /*
820 * pmap_enter_pv: enter a mapping onto a vm_page lst
821 *
822 * => caller should hold the proper lock on pmap_main_lock
823 * => caller should have pmap locked
824 * => we will gain the lock on the vm_page and allocate the new pv_entry
825 * => caller should adjust ptp's wire_count before calling
826 * => caller should not adjust pmap's wire_count
827 */
828
829 __inline static void
830 pmap_enter_pv(pg, pve, pmap, va, ptp, flags)
831 struct vm_page *pg;
832 struct pv_entry *pve; /* preallocated pve for us to use */
833 struct pmap *pmap;
834 vaddr_t va;
835 struct vm_page *ptp; /* PTP in pmap that maps this VA */
836 int flags;
837 {
838 pve->pv_pmap = pmap;
839 pve->pv_va = va;
840 pve->pv_ptp = ptp; /* NULL for kernel pmap */
841 pve->pv_flags = flags;
842 simple_lock(&pg->mdpage.pvh_slock); /* lock vm_page */
843 pve->pv_next = pg->mdpage.pvh_list; /* add to ... */
844 pg->mdpage.pvh_list = pve; /* ... locked list */
845 simple_unlock(&pg->mdpage.pvh_slock); /* unlock, done! */
846 if (pve->pv_flags & PT_W)
847 ++pmap->pm_stats.wired_count;
848 }
849
850 /*
851 * pmap_remove_pv: try to remove a mapping from a pv_list
852 *
853 * => caller should hold proper lock on pmap_main_lock
854 * => pmap should be locked
855 * => caller should hold lock on vm_page [so that attrs can be adjusted]
856 * => caller should adjust ptp's wire_count and free PTP if needed
857 * => caller should NOT adjust pmap's wire_count
858 * => we return the removed pve
859 */
860
861 __inline static struct pv_entry *
862 pmap_remove_pv(pg, pmap, va)
863 struct vm_page *pg;
864 struct pmap *pmap;
865 vaddr_t va;
866 {
867 struct pv_entry *pve, **prevptr;
868
869 prevptr = &pg->mdpage.pvh_list; /* previous pv_entry pointer */
870 pve = *prevptr;
871 while (pve) {
872 if (pve->pv_pmap == pmap && pve->pv_va == va) { /* match? */
873 *prevptr = pve->pv_next; /* remove it! */
874 if (pve->pv_flags & PT_W)
875 --pmap->pm_stats.wired_count;
876 break;
877 }
878 prevptr = &pve->pv_next; /* previous pointer */
879 pve = pve->pv_next; /* advance */
880 }
881 return(pve); /* return removed pve */
882 }
883
884 /*
885 *
886 * pmap_modify_pv: Update pv flags
887 *
888 * => caller should hold lock on vm_page [so that attrs can be adjusted]
889 * => caller should NOT adjust pmap's wire_count
890 * => caller must call pmap_vac_me_harder() if writable status of a page
891 * may have changed.
892 * => we return the old flags
893 *
894 * Modify a physical-virtual mapping in the pv table
895 */
896
897 /*__inline */
898 static u_int
899 pmap_modify_pv(pmap, va, pg, bic_mask, eor_mask)
900 struct pmap *pmap;
901 vaddr_t va;
902 struct vm_page *pg;
903 u_int bic_mask;
904 u_int eor_mask;
905 {
906 struct pv_entry *npv;
907 u_int flags, oflags;
908
909 /*
910 * There is at least one VA mapping this page.
911 */
912
913 for (npv = pg->mdpage.pvh_list; npv; npv = npv->pv_next) {
914 if (pmap == npv->pv_pmap && va == npv->pv_va) {
915 oflags = npv->pv_flags;
916 npv->pv_flags = flags =
917 ((oflags & ~bic_mask) ^ eor_mask);
918 if ((flags ^ oflags) & PT_W) {
919 if (flags & PT_W)
920 ++pmap->pm_stats.wired_count;
921 else
922 --pmap->pm_stats.wired_count;
923 }
924 return (oflags);
925 }
926 }
927 return (0);
928 }
929
930 /*
931 * Map the specified level 2 pagetable into the level 1 page table for
932 * the given pmap to cover a chunk of virtual address space starting from the
933 * address specified.
934 */
935 static /*__inline*/ void
936 pmap_map_in_l1(pmap, va, l2pa, selfref)
937 struct pmap *pmap;
938 vaddr_t va, l2pa;
939 boolean_t selfref;
940 {
941 vaddr_t ptva;
942
943 /* Calculate the index into the L1 page table. */
944 ptva = (va >> PDSHIFT) & ~3;
945
946 NPDEBUG(PDB_MAP_L1, printf("wiring %08lx in to pd%p pte0x%lx va0x%lx\n", l2pa,
947 pmap->pm_pdir, L1_PTE(l2pa), ptva));
948
949 /* Map page table into the L1. */
950 pmap->pm_pdir[ptva + 0] = L1_PTE(l2pa + 0x000);
951 pmap->pm_pdir[ptva + 1] = L1_PTE(l2pa + 0x400);
952 pmap->pm_pdir[ptva + 2] = L1_PTE(l2pa + 0x800);
953 pmap->pm_pdir[ptva + 3] = L1_PTE(l2pa + 0xc00);
954
955 /* Map the page table into the page table area. */
956 if (selfref) {
957 NPDEBUG(PDB_MAP_L1, printf("pt self reference %lx in %lx\n",
958 L2_PTE_NC_NB(l2pa, AP_KRW), pmap->pm_vptpt));
959 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) =
960 L2_PTE_NC_NB(l2pa, AP_KRW);
961 }
962 /* XXX should be a purge */
963 /* cpu_tlb_flushD();*/
964 }
965
966 #if 0
967 static /*__inline*/ void
968 pmap_unmap_in_l1(pmap, va)
969 struct pmap *pmap;
970 vaddr_t va;
971 {
972 vaddr_t ptva;
973
974 /* Calculate the index into the L1 page table. */
975 ptva = (va >> PDSHIFT) & ~3;
976
977 /* Unmap page table from the L1. */
978 pmap->pm_pdir[ptva + 0] = 0;
979 pmap->pm_pdir[ptva + 1] = 0;
980 pmap->pm_pdir[ptva + 2] = 0;
981 pmap->pm_pdir[ptva + 3] = 0;
982
983 /* Unmap the page table from the page table area. */
984 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) = 0;
985
986 /* XXX should be a purge */
987 /* cpu_tlb_flushD();*/
988 }
989 #endif
990
991 /*
992 * Used to map a range of physical addresses into kernel
993 * virtual address space.
994 *
995 * For now, VM is already on, we only need to map the
996 * specified memory.
997 */
998 vaddr_t
999 pmap_map(va, spa, epa, prot)
1000 vaddr_t va, spa, epa;
1001 int prot;
1002 {
1003 while (spa < epa) {
1004 pmap_kenter_pa(va, spa, prot);
1005 va += NBPG;
1006 spa += NBPG;
1007 }
1008 pmap_update(pmap_kernel());
1009 return(va);
1010 }
1011
1012
1013 /*
1014 * void pmap_bootstrap(pd_entry_t *kernel_l1pt, pv_addr_t kernel_ptpt)
1015 *
1016 * bootstrap the pmap system. This is called from initarm and allows
1017 * the pmap system to initailise any structures it requires.
1018 *
1019 * Currently this sets up the kernel_pmap that is statically allocated
1020 * and also allocated virtual addresses for certain page hooks.
1021 * Currently the only one page hook is allocated that is used
1022 * to zero physical pages of memory.
1023 * It also initialises the start and end address of the kernel data space.
1024 */
1025 extern paddr_t physical_freestart;
1026 extern paddr_t physical_freeend;
1027
1028 char *boot_head;
1029
1030 void
1031 pmap_bootstrap(kernel_l1pt, kernel_ptpt)
1032 pd_entry_t *kernel_l1pt;
1033 pv_addr_t kernel_ptpt;
1034 {
1035 int loop;
1036 paddr_t start, end;
1037 #if NISADMA > 0
1038 paddr_t istart;
1039 psize_t isize;
1040 #endif
1041
1042 pmap_kernel()->pm_pdir = kernel_l1pt;
1043 pmap_kernel()->pm_pptpt = kernel_ptpt.pv_pa;
1044 pmap_kernel()->pm_vptpt = kernel_ptpt.pv_va;
1045 simple_lock_init(&pmap_kernel()->pm_lock);
1046 pmap_kernel()->pm_obj.pgops = NULL;
1047 TAILQ_INIT(&(pmap_kernel()->pm_obj.memq));
1048 pmap_kernel()->pm_obj.uo_npages = 0;
1049 pmap_kernel()->pm_obj.uo_refs = 1;
1050
1051 /*
1052 * Initialize PAGE_SIZE-dependent variables.
1053 */
1054 uvm_setpagesize();
1055
1056 npages = 0;
1057 loop = 0;
1058 while (loop < bootconfig.dramblocks) {
1059 start = (paddr_t)bootconfig.dram[loop].address;
1060 end = start + (bootconfig.dram[loop].pages * NBPG);
1061 if (start < physical_freestart)
1062 start = physical_freestart;
1063 if (end > physical_freeend)
1064 end = physical_freeend;
1065 #if 0
1066 printf("%d: %lx -> %lx\n", loop, start, end - 1);
1067 #endif
1068 #if NISADMA > 0
1069 if (pmap_isa_dma_range_intersect(start, end - start,
1070 &istart, &isize)) {
1071 /*
1072 * Place the pages that intersect with the
1073 * ISA DMA range onto the ISA DMA free list.
1074 */
1075 #if 0
1076 printf(" ISADMA 0x%lx -> 0x%lx\n", istart,
1077 istart + isize - 1);
1078 #endif
1079 uvm_page_physload(atop(istart),
1080 atop(istart + isize), atop(istart),
1081 atop(istart + isize), VM_FREELIST_ISADMA);
1082 npages += atop(istart + isize) - atop(istart);
1083
1084 /*
1085 * Load the pieces that come before
1086 * the intersection into the default
1087 * free list.
1088 */
1089 if (start < istart) {
1090 #if 0
1091 printf(" BEFORE 0x%lx -> 0x%lx\n",
1092 start, istart - 1);
1093 #endif
1094 uvm_page_physload(atop(start),
1095 atop(istart), atop(start),
1096 atop(istart), VM_FREELIST_DEFAULT);
1097 npages += atop(istart) - atop(start);
1098 }
1099
1100 /*
1101 * Load the pieces that come after
1102 * the intersection into the default
1103 * free list.
1104 */
1105 if ((istart + isize) < end) {
1106 #if 0
1107 printf(" AFTER 0x%lx -> 0x%lx\n",
1108 (istart + isize), end - 1);
1109 #endif
1110 uvm_page_physload(atop(istart + isize),
1111 atop(end), atop(istart + isize),
1112 atop(end), VM_FREELIST_DEFAULT);
1113 npages += atop(end) - atop(istart + isize);
1114 }
1115 } else {
1116 uvm_page_physload(atop(start), atop(end),
1117 atop(start), atop(end), VM_FREELIST_DEFAULT);
1118 npages += atop(end) - atop(start);
1119 }
1120 #else /* NISADMA > 0 */
1121 uvm_page_physload(atop(start), atop(end),
1122 atop(start), atop(end), VM_FREELIST_DEFAULT);
1123 npages += atop(end) - atop(start);
1124 #endif /* NISADMA > 0 */
1125 ++loop;
1126 }
1127
1128 #ifdef MYCROFT_HACK
1129 printf("npages = %ld\n", npages);
1130 #endif
1131
1132 virtual_start = KERNEL_VM_BASE;
1133 virtual_end = KERNEL_VM_BASE + KERNEL_VM_SIZE - 1;
1134
1135 ALLOC_PAGE_HOOK(page_hook0, NBPG);
1136 ALLOC_PAGE_HOOK(page_hook1, NBPG);
1137
1138 /*
1139 * The mem special device needs a virtual hook but we don't
1140 * need a pte
1141 */
1142 memhook = (char *)virtual_start;
1143 virtual_start += NBPG;
1144
1145 msgbufaddr = (caddr_t)virtual_start;
1146 msgbufpte = (pt_entry_t)pmap_pte(pmap_kernel(), virtual_start);
1147 virtual_start += round_page(MSGBUFSIZE);
1148
1149 /*
1150 * init the static-global locks and global lists.
1151 */
1152 spinlockinit(&pmap_main_lock, "pmaplk", 0);
1153 simple_lock_init(&pvalloc_lock);
1154 simple_lock_init(&pmaps_lock);
1155 LIST_INIT(&pmaps);
1156 TAILQ_INIT(&pv_freepages);
1157 TAILQ_INIT(&pv_unusedpgs);
1158
1159 /*
1160 * initialize the pmap pool.
1161 */
1162
1163 pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl",
1164 0, pool_page_alloc_nointr, pool_page_free_nointr, M_VMPMAP);
1165
1166 cpu_dcache_wbinv_all();
1167 }
1168
1169 /*
1170 * void pmap_init(void)
1171 *
1172 * Initialize the pmap module.
1173 * Called by vm_init() in vm/vm_init.c in order to initialise
1174 * any structures that the pmap system needs to map virtual memory.
1175 */
1176
1177 extern int physmem;
1178
1179 void
1180 pmap_init()
1181 {
1182
1183 /*
1184 * Set the available memory vars - These do not map to real memory
1185 * addresses and cannot as the physical memory is fragmented.
1186 * They are used by ps for %mem calculations.
1187 * One could argue whether this should be the entire memory or just
1188 * the memory that is useable in a user process.
1189 */
1190 avail_start = 0;
1191 avail_end = physmem * NBPG;
1192
1193 /*
1194 * now we need to free enough pv_entry structures to allow us to get
1195 * the kmem_map/kmem_object allocated and inited (done after this
1196 * function is finished). to do this we allocate one bootstrap page out
1197 * of kernel_map and use it to provide an initial pool of pv_entry
1198 * structures. we never free this page.
1199 */
1200
1201 pv_initpage = (struct pv_page *) uvm_km_alloc(kernel_map, PAGE_SIZE);
1202 if (pv_initpage == NULL)
1203 panic("pmap_init: pv_initpage");
1204 pv_cachedva = 0; /* a VA we have allocated but not used yet */
1205 pv_nfpvents = 0;
1206 (void) pmap_add_pvpage(pv_initpage, FALSE);
1207
1208 pmap_initialized = TRUE;
1209
1210 /* Initialise our L1 page table queues and counters */
1211 SIMPLEQ_INIT(&l1pt_static_queue);
1212 l1pt_static_queue_count = 0;
1213 l1pt_static_create_count = 0;
1214 SIMPLEQ_INIT(&l1pt_queue);
1215 l1pt_queue_count = 0;
1216 l1pt_create_count = 0;
1217 l1pt_reuse_count = 0;
1218 }
1219
1220 /*
1221 * pmap_postinit()
1222 *
1223 * This routine is called after the vm and kmem subsystems have been
1224 * initialised. This allows the pmap code to perform any initialisation
1225 * that can only be done one the memory allocation is in place.
1226 */
1227
1228 void
1229 pmap_postinit()
1230 {
1231 int loop;
1232 struct l1pt *pt;
1233
1234 #ifdef PMAP_STATIC_L1S
1235 for (loop = 0; loop < PMAP_STATIC_L1S; ++loop) {
1236 #else /* PMAP_STATIC_L1S */
1237 for (loop = 0; loop < max_processes; ++loop) {
1238 #endif /* PMAP_STATIC_L1S */
1239 /* Allocate a L1 page table */
1240 pt = pmap_alloc_l1pt();
1241 if (!pt)
1242 panic("Cannot allocate static L1 page tables\n");
1243
1244 /* Clean it */
1245 bzero((void *)pt->pt_va, PD_SIZE);
1246 pt->pt_flags |= (PTFLAG_STATIC | PTFLAG_CLEAN);
1247 /* Add the page table to the queue */
1248 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pt, pt_queue);
1249 ++l1pt_static_queue_count;
1250 ++l1pt_static_create_count;
1251 }
1252 }
1253
1254
1255 /*
1256 * Create and return a physical map.
1257 *
1258 * If the size specified for the map is zero, the map is an actual physical
1259 * map, and may be referenced by the hardware.
1260 *
1261 * If the size specified is non-zero, the map will be used in software only,
1262 * and is bounded by that size.
1263 */
1264
1265 pmap_t
1266 pmap_create()
1267 {
1268 struct pmap *pmap;
1269
1270 /*
1271 * Fetch pmap entry from the pool
1272 */
1273
1274 pmap = pool_get(&pmap_pmap_pool, PR_WAITOK);
1275 /* XXX is this really needed! */
1276 memset(pmap, 0, sizeof(*pmap));
1277
1278 simple_lock_init(&pmap->pm_obj.vmobjlock);
1279 pmap->pm_obj.pgops = NULL; /* currently not a mappable object */
1280 TAILQ_INIT(&pmap->pm_obj.memq);
1281 pmap->pm_obj.uo_npages = 0;
1282 pmap->pm_obj.uo_refs = 1;
1283 pmap->pm_stats.wired_count = 0;
1284 pmap->pm_stats.resident_count = 1;
1285
1286 /* Now init the machine part of the pmap */
1287 pmap_pinit(pmap);
1288 return(pmap);
1289 }
1290
1291 /*
1292 * pmap_alloc_l1pt()
1293 *
1294 * This routine allocates physical and virtual memory for a L1 page table
1295 * and wires it.
1296 * A l1pt structure is returned to describe the allocated page table.
1297 *
1298 * This routine is allowed to fail if the required memory cannot be allocated.
1299 * In this case NULL is returned.
1300 */
1301
1302 struct l1pt *
1303 pmap_alloc_l1pt(void)
1304 {
1305 paddr_t pa;
1306 vaddr_t va;
1307 struct l1pt *pt;
1308 int error;
1309 struct vm_page *m;
1310 pt_entry_t *ptes;
1311
1312 /* Allocate virtual address space for the L1 page table */
1313 va = uvm_km_valloc(kernel_map, PD_SIZE);
1314 if (va == 0) {
1315 #ifdef DIAGNOSTIC
1316 PDEBUG(0,
1317 printf("pmap: Cannot allocate pageable memory for L1\n"));
1318 #endif /* DIAGNOSTIC */
1319 return(NULL);
1320 }
1321
1322 /* Allocate memory for the l1pt structure */
1323 pt = (struct l1pt *)malloc(sizeof(struct l1pt), M_VMPMAP, M_WAITOK);
1324
1325 /*
1326 * Allocate pages from the VM system.
1327 */
1328 TAILQ_INIT(&pt->pt_plist);
1329 error = uvm_pglistalloc(PD_SIZE, physical_start, physical_end,
1330 PD_SIZE, 0, &pt->pt_plist, 1, M_WAITOK);
1331 if (error) {
1332 #ifdef DIAGNOSTIC
1333 PDEBUG(0,
1334 printf("pmap: Cannot allocate physical mem for L1 (%d)\n",
1335 error));
1336 #endif /* DIAGNOSTIC */
1337 /* Release the resources we already have claimed */
1338 free(pt, M_VMPMAP);
1339 uvm_km_free(kernel_map, va, PD_SIZE);
1340 return(NULL);
1341 }
1342
1343 /* Map our physical pages into our virtual space */
1344 pt->pt_va = va;
1345 m = pt->pt_plist.tqh_first;
1346 ptes = pmap_map_ptes(pmap_kernel());
1347 while (m && va < (pt->pt_va + PD_SIZE)) {
1348 pa = VM_PAGE_TO_PHYS(m);
1349
1350 pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE);
1351
1352 /* Revoke cacheability and bufferability */
1353 /* XXX should be done better than this */
1354 ptes[arm_byte_to_page(va)] &= ~(PT_C | PT_B);
1355
1356 va += NBPG;
1357 m = m->pageq.tqe_next;
1358 }
1359 pmap_unmap_ptes(pmap_kernel());
1360 pmap_update(pmap_kernel());
1361
1362 #ifdef DIAGNOSTIC
1363 if (m)
1364 panic("pmap_alloc_l1pt: pglist not empty\n");
1365 #endif /* DIAGNOSTIC */
1366
1367 pt->pt_flags = 0;
1368 return(pt);
1369 }
1370
1371 /*
1372 * Free a L1 page table previously allocated with pmap_alloc_l1pt().
1373 */
1374 static void
1375 pmap_free_l1pt(pt)
1376 struct l1pt *pt;
1377 {
1378 /* Separate the physical memory for the virtual space */
1379 pmap_kremove(pt->pt_va, PD_SIZE);
1380 pmap_update(pmap_kernel());
1381
1382 /* Return the physical memory */
1383 uvm_pglistfree(&pt->pt_plist);
1384
1385 /* Free the virtual space */
1386 uvm_km_free(kernel_map, pt->pt_va, PD_SIZE);
1387
1388 /* Free the l1pt structure */
1389 free(pt, M_VMPMAP);
1390 }
1391
1392 /*
1393 * Allocate a page directory.
1394 * This routine will either allocate a new page directory from the pool
1395 * of L1 page tables currently held by the kernel or it will allocate
1396 * a new one via pmap_alloc_l1pt().
1397 * It will then initialise the l1 page table for use.
1398 *
1399 * XXX must tidy up and fix this code, not happy about how it does the pmaps_locking
1400 */
1401 static int
1402 pmap_allocpagedir(pmap)
1403 struct pmap *pmap;
1404 {
1405 paddr_t pa;
1406 struct l1pt *pt;
1407 pt_entry_t *pte;
1408
1409 PDEBUG(0, printf("pmap_allocpagedir(%p)\n", pmap));
1410
1411 /* Do we have any spare L1's lying around ? */
1412 if (l1pt_static_queue_count) {
1413 --l1pt_static_queue_count;
1414 pt = l1pt_static_queue.sqh_first;
1415 SIMPLEQ_REMOVE_HEAD(&l1pt_static_queue, pt, pt_queue);
1416 } else if (l1pt_queue_count) {
1417 --l1pt_queue_count;
1418 pt = l1pt_queue.sqh_first;
1419 SIMPLEQ_REMOVE_HEAD(&l1pt_queue, pt, pt_queue);
1420 ++l1pt_reuse_count;
1421 } else {
1422 pt = pmap_alloc_l1pt();
1423 if (!pt)
1424 return(ENOMEM);
1425 ++l1pt_create_count;
1426 }
1427
1428 /* Store the pointer to the l1 descriptor in the pmap. */
1429 pmap->pm_l1pt = pt;
1430
1431 /* Get the physical address of the start of the l1 */
1432 pa = VM_PAGE_TO_PHYS(pt->pt_plist.tqh_first);
1433
1434 /* Store the virtual address of the l1 in the pmap. */
1435 pmap->pm_pdir = (pd_entry_t *)pt->pt_va;
1436
1437 /* Clean the L1 if it is dirty */
1438 if (!(pt->pt_flags & PTFLAG_CLEAN))
1439 bzero((void *)pmap->pm_pdir, (PD_SIZE - KERNEL_PD_SIZE));
1440
1441 /* Allocate a page table to map all the page tables for this pmap */
1442
1443 #ifdef DIAGNOSTIC
1444 if (pmap->pm_vptpt) {
1445 /* XXX What if we have one already ? */
1446 panic("pmap_allocpagedir: have pt already\n");
1447 }
1448 #endif /* DIAGNOSTIC */
1449 pmap->pm_vptpt = uvm_km_zalloc(kernel_map, NBPG);
1450 if (pmap->pm_vptpt == 0) {
1451 pmap_freepagedir(pmap);
1452 return(ENOMEM);
1453 }
1454
1455 /* need to lock this all up for growkernel */
1456 simple_lock(&pmaps_lock);
1457 /* wish we didn't have to keep this locked... */
1458
1459 /* Duplicate the kernel mapping i.e. all mappings 0xf0000000+ */
1460 bcopy((char *)pmap_kernel()->pm_pdir + (PD_SIZE - KERNEL_PD_SIZE),
1461 (char *)pmap->pm_pdir + (PD_SIZE - KERNEL_PD_SIZE),
1462 KERNEL_PD_SIZE);
1463
1464 (void) pmap_extract(pmap_kernel(), pmap->pm_vptpt, &pmap->pm_pptpt);
1465 pmap->pm_pptpt &= PG_FRAME;
1466 /* Revoke cacheability and bufferability */
1467 /* XXX should be done better than this */
1468 pte = pmap_pte(pmap_kernel(), pmap->pm_vptpt);
1469 *pte = *pte & ~(PT_C | PT_B);
1470
1471 /* Wire in this page table */
1472 pmap_map_in_l1(pmap, PROCESS_PAGE_TBLS_BASE, pmap->pm_pptpt, TRUE);
1473
1474 pt->pt_flags &= ~PTFLAG_CLEAN; /* L1 is dirty now */
1475
1476 /*
1477 * Map the kernel page tables for 0xf0000000 +
1478 * into the page table used to map the
1479 * pmap's page tables
1480 */
1481 bcopy((char *)(PROCESS_PAGE_TBLS_BASE
1482 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2))
1483 + ((PD_SIZE - KERNEL_PD_SIZE) >> 2)),
1484 (char *)pmap->pm_vptpt + ((PD_SIZE - KERNEL_PD_SIZE) >> 2),
1485 (KERNEL_PD_SIZE >> 2));
1486
1487 LIST_INSERT_HEAD(&pmaps, pmap, pm_list);
1488 simple_unlock(&pmaps_lock);
1489
1490 return(0);
1491 }
1492
1493
1494 /*
1495 * Initialize a preallocated and zeroed pmap structure,
1496 * such as one in a vmspace structure.
1497 */
1498
1499 void
1500 pmap_pinit(pmap)
1501 struct pmap *pmap;
1502 {
1503 int backoff = 6;
1504 int retry = 10;
1505
1506 PDEBUG(0, printf("pmap_pinit(%p)\n", pmap));
1507
1508 /* Keep looping until we succeed in allocating a page directory */
1509 while (pmap_allocpagedir(pmap) != 0) {
1510 /*
1511 * Ok we failed to allocate a suitable block of memory for an
1512 * L1 page table. This means that either:
1513 * 1. 16KB of virtual address space could not be allocated
1514 * 2. 16KB of physically contiguous memory on a 16KB boundary
1515 * could not be allocated.
1516 *
1517 * Since we cannot fail we will sleep for a while and try
1518 * again.
1519 *
1520 * Searching for a suitable L1 PT is expensive:
1521 * to avoid hogging the system when memory is really
1522 * scarce, use an exponential back-off so that
1523 * eventually we won't retry more than once every 8
1524 * seconds. This should allow other processes to run
1525 * to completion and free up resources.
1526 */
1527 (void) ltsleep(&lbolt, PVM, "l1ptwait", (hz << 3) >> backoff,
1528 NULL);
1529 if (--retry == 0) {
1530 retry = 10;
1531 if (backoff)
1532 --backoff;
1533 }
1534 }
1535
1536 /* Map zero page for the pmap. This will also map the L2 for it */
1537 pmap_enter(pmap, 0x00000000, systempage.pv_pa,
1538 VM_PROT_READ, VM_PROT_READ | PMAP_WIRED);
1539 pmap_update(pmap);
1540 }
1541
1542
1543 void
1544 pmap_freepagedir(pmap)
1545 struct pmap *pmap;
1546 {
1547 /* Free the memory used for the page table mapping */
1548 if (pmap->pm_vptpt != 0)
1549 uvm_km_free(kernel_map, (vaddr_t)pmap->pm_vptpt, NBPG);
1550
1551 /* junk the L1 page table */
1552 if (pmap->pm_l1pt->pt_flags & PTFLAG_STATIC) {
1553 /* Add the page table to the queue */
1554 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pmap->pm_l1pt, pt_queue);
1555 ++l1pt_static_queue_count;
1556 } else if (l1pt_queue_count < 8) {
1557 /* Add the page table to the queue */
1558 SIMPLEQ_INSERT_TAIL(&l1pt_queue, pmap->pm_l1pt, pt_queue);
1559 ++l1pt_queue_count;
1560 } else
1561 pmap_free_l1pt(pmap->pm_l1pt);
1562 }
1563
1564
1565 /*
1566 * Retire the given physical map from service.
1567 * Should only be called if the map contains no valid mappings.
1568 */
1569
1570 void
1571 pmap_destroy(pmap)
1572 struct pmap *pmap;
1573 {
1574 struct vm_page *page;
1575 int count;
1576
1577 if (pmap == NULL)
1578 return;
1579
1580 PDEBUG(0, printf("pmap_destroy(%p)\n", pmap));
1581
1582 /*
1583 * Drop reference count
1584 */
1585 simple_lock(&pmap->pm_obj.vmobjlock);
1586 count = --pmap->pm_obj.uo_refs;
1587 simple_unlock(&pmap->pm_obj.vmobjlock);
1588 if (count > 0) {
1589 return;
1590 }
1591
1592 /*
1593 * reference count is zero, free pmap resources and then free pmap.
1594 */
1595
1596 /*
1597 * remove it from global list of pmaps
1598 */
1599
1600 simple_lock(&pmaps_lock);
1601 LIST_REMOVE(pmap, pm_list);
1602 simple_unlock(&pmaps_lock);
1603
1604 /* Remove the zero page mapping */
1605 pmap_remove(pmap, 0x00000000, 0x00000000 + NBPG);
1606 pmap_update(pmap);
1607
1608 /*
1609 * Free any page tables still mapped
1610 * This is only temporay until pmap_enter can count the number
1611 * of mappings made in a page table. Then pmap_remove() can
1612 * reduce the count and free the pagetable when the count
1613 * reaches zero. Note that entries in this list should match the
1614 * contents of the ptpt, however this is faster than walking a 1024
1615 * entries looking for pt's
1616 * taken from i386 pmap.c
1617 */
1618 while (pmap->pm_obj.memq.tqh_first != NULL) {
1619 page = pmap->pm_obj.memq.tqh_first;
1620 #ifdef DIAGNOSTIC
1621 if (page->flags & PG_BUSY)
1622 panic("pmap_release: busy page table page");
1623 #endif
1624 /* pmap_page_protect? currently no need for it. */
1625
1626 page->wire_count = 0;
1627 uvm_pagefree(page);
1628 }
1629
1630 /* Free the page dir */
1631 pmap_freepagedir(pmap);
1632
1633 /* return the pmap to the pool */
1634 pool_put(&pmap_pmap_pool, pmap);
1635 }
1636
1637
1638 /*
1639 * void pmap_reference(struct pmap *pmap)
1640 *
1641 * Add a reference to the specified pmap.
1642 */
1643
1644 void
1645 pmap_reference(pmap)
1646 struct pmap *pmap;
1647 {
1648 if (pmap == NULL)
1649 return;
1650
1651 simple_lock(&pmap->pm_lock);
1652 pmap->pm_obj.uo_refs++;
1653 simple_unlock(&pmap->pm_lock);
1654 }
1655
1656 /*
1657 * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
1658 *
1659 * Return the start and end addresses of the kernel's virtual space.
1660 * These values are setup in pmap_bootstrap and are updated as pages
1661 * are allocated.
1662 */
1663
1664 void
1665 pmap_virtual_space(start, end)
1666 vaddr_t *start;
1667 vaddr_t *end;
1668 {
1669 *start = virtual_start;
1670 *end = virtual_end;
1671 }
1672
1673
1674 /*
1675 * Activate the address space for the specified process. If the process
1676 * is the current process, load the new MMU context.
1677 */
1678 void
1679 pmap_activate(p)
1680 struct proc *p;
1681 {
1682 struct pmap *pmap = p->p_vmspace->vm_map.pmap;
1683 struct pcb *pcb = &p->p_addr->u_pcb;
1684
1685 (void) pmap_extract(pmap_kernel(), (vaddr_t)pmap->pm_pdir,
1686 (paddr_t *)&pcb->pcb_pagedir);
1687
1688 PDEBUG(0, printf("pmap_activate: p=%p pmap=%p pcb=%p pdir=%p l1=%p\n",
1689 p, pmap, pcb, pmap->pm_pdir, pcb->pcb_pagedir));
1690
1691 if (p == curproc) {
1692 PDEBUG(0, printf("pmap_activate: setting TTB\n"));
1693 setttb((u_int)pcb->pcb_pagedir);
1694 }
1695 #if 0
1696 pmap->pm_pdchanged = FALSE;
1697 #endif
1698 }
1699
1700
1701 /*
1702 * Deactivate the address space of the specified process.
1703 */
1704 void
1705 pmap_deactivate(p)
1706 struct proc *p;
1707 {
1708 }
1709
1710 /*
1711 * Perform any deferred pmap operations.
1712 */
1713 void
1714 pmap_update(struct pmap *pmap)
1715 {
1716
1717 /*
1718 * We haven't deferred any pmap operations, but we do need to
1719 * make sure TLB/cache operations have completed.
1720 */
1721 cpu_cpwait();
1722 }
1723
1724 /*
1725 * pmap_clean_page()
1726 *
1727 * This is a local function used to work out the best strategy to clean
1728 * a single page referenced by its entry in the PV table. It's used by
1729 * pmap_copy_page, pmap_zero page and maybe some others later on.
1730 *
1731 * Its policy is effectively:
1732 * o If there are no mappings, we don't bother doing anything with the cache.
1733 * o If there is one mapping, we clean just that page.
1734 * o If there are multiple mappings, we clean the entire cache.
1735 *
1736 * So that some functions can be further optimised, it returns 0 if it didn't
1737 * clean the entire cache, or 1 if it did.
1738 *
1739 * XXX One bug in this routine is that if the pv_entry has a single page
1740 * mapped at 0x00000000 a whole cache clean will be performed rather than
1741 * just the 1 page. Since this should not occur in everyday use and if it does
1742 * it will just result in not the most efficient clean for the page.
1743 */
1744 static int
1745 pmap_clean_page(pv, is_src)
1746 struct pv_entry *pv;
1747 boolean_t is_src;
1748 {
1749 struct pmap *pmap;
1750 struct pv_entry *npv;
1751 int cache_needs_cleaning = 0;
1752 vaddr_t page_to_clean = 0;
1753
1754 if (pv == NULL)
1755 /* nothing mapped in so nothing to flush */
1756 return (0);
1757
1758 /* Since we flush the cache each time we change curproc, we
1759 * only need to flush the page if it is in the current pmap.
1760 */
1761 if (curproc)
1762 pmap = curproc->p_vmspace->vm_map.pmap;
1763 else
1764 pmap = pmap_kernel();
1765
1766 for (npv = pv; npv; npv = npv->pv_next) {
1767 if (npv->pv_pmap == pmap) {
1768 /* The page is mapped non-cacheable in
1769 * this map. No need to flush the cache.
1770 */
1771 if (npv->pv_flags & PT_NC) {
1772 #ifdef DIAGNOSTIC
1773 if (cache_needs_cleaning)
1774 panic("pmap_clean_page: "
1775 "cache inconsistency");
1776 #endif
1777 break;
1778 }
1779 #if 0
1780 /* This doesn't work, because pmap_protect
1781 doesn't flush changes on pages that it
1782 has write-protected. */
1783
1784 /* If the page is not writable and this
1785 is the source, then there is no need
1786 to flush it from the cache. */
1787 else if (is_src && ! (npv->pv_flags & PT_Wr))
1788 continue;
1789 #endif
1790 if (cache_needs_cleaning){
1791 page_to_clean = 0;
1792 break;
1793 }
1794 else
1795 page_to_clean = npv->pv_va;
1796 cache_needs_cleaning = 1;
1797 }
1798 }
1799
1800 if (page_to_clean)
1801 cpu_idcache_wbinv_range(page_to_clean, NBPG);
1802 else if (cache_needs_cleaning) {
1803 cpu_idcache_wbinv_all();
1804 return (1);
1805 }
1806 return (0);
1807 }
1808
1809 /*
1810 * pmap_zero_page()
1811 *
1812 * Zero a given physical page by mapping it at a page hook point.
1813 * In doing the zero page op, the page we zero is mapped cachable, as with
1814 * StrongARM accesses to non-cached pages are non-burst making writing
1815 * _any_ bulk data very slow.
1816 */
1817 void
1818 pmap_zero_page(phys)
1819 paddr_t phys;
1820 {
1821 struct vm_page *pg;
1822
1823 /* Get an entry for this page, and clean it it. */
1824 pg = PHYS_TO_VM_PAGE(phys);
1825 simple_lock(&pg->mdpage.pvh_slock);
1826 pmap_clean_page(pg->mdpage.pvh_list, FALSE);
1827 simple_unlock(&pg->mdpage.pvh_slock);
1828
1829 /*
1830 * Hook in the page, zero it, and purge the cache for that
1831 * zeroed page. Invalidate the TLB as needed.
1832 */
1833 *page_hook0.pte = L2_PTE(phys & PG_FRAME, AP_KRW);
1834 cpu_tlb_flushD_SE(page_hook0.va);
1835 cpu_cpwait();
1836 bzero_page(page_hook0.va);
1837 cpu_dcache_wbinv_range(page_hook0.va, NBPG);
1838 }
1839
1840 /* pmap_pageidlezero()
1841 *
1842 * The same as above, except that we assume that the page is not
1843 * mapped. This means we never have to flush the cache first. Called
1844 * from the idle loop.
1845 */
1846 boolean_t
1847 pmap_pageidlezero(phys)
1848 paddr_t phys;
1849 {
1850 int i, *ptr;
1851 boolean_t rv = TRUE;
1852
1853 #ifdef DIAGNOSTIC
1854 struct vm_page *pg;
1855
1856 pg = PHYS_TO_VM_PAGE(phys);
1857 if (pg->mdpage.pvh_list != NULL)
1858 panic("pmap_pageidlezero: zeroing mapped page\n");
1859 #endif
1860
1861 /*
1862 * Hook in the page, zero it, and purge the cache for that
1863 * zeroed page. Invalidate the TLB as needed.
1864 */
1865 *page_hook0.pte = L2_PTE(phys & PG_FRAME, AP_KRW);
1866 cpu_tlb_flushD_SE(page_hook0.va);
1867 cpu_cpwait();
1868
1869 for (i = 0, ptr = (int *)page_hook0.va;
1870 i < (NBPG / sizeof(int)); i++) {
1871 if (sched_whichqs != 0) {
1872 /*
1873 * A process has become ready. Abort now,
1874 * so we don't keep it waiting while we
1875 * do slow memory access to finish this
1876 * page.
1877 */
1878 rv = FALSE;
1879 break;
1880 }
1881 *ptr++ = 0;
1882 }
1883
1884 if (rv)
1885 /*
1886 * if we aborted we'll rezero this page again later so don't
1887 * purge it unless we finished it
1888 */
1889 cpu_dcache_wbinv_range(page_hook0.va, NBPG);
1890 return (rv);
1891 }
1892
1893 /*
1894 * pmap_copy_page()
1895 *
1896 * Copy one physical page into another, by mapping the pages into
1897 * hook points. The same comment regarding cachability as in
1898 * pmap_zero_page also applies here.
1899 */
1900 void
1901 pmap_copy_page(src, dest)
1902 paddr_t src;
1903 paddr_t dest;
1904 {
1905 struct vm_page *src_pg, *dest_pg;
1906 boolean_t cleanedcache;
1907
1908 /* Get PV entries for the pages, and clean them if needed. */
1909 src_pg = PHYS_TO_VM_PAGE(src);
1910
1911 simple_lock(&src_pg->mdpage.pvh_slock);
1912 cleanedcache = pmap_clean_page(src_pg->mdpage.pvh_list, TRUE);
1913 simple_unlock(&src_pg->mdpage.pvh_slock);
1914
1915 if (cleanedcache == 0) {
1916 dest_pg = PHYS_TO_VM_PAGE(dest);
1917 simple_lock(&dest_pg->mdpage.pvh_slock);
1918 pmap_clean_page(dest_pg->mdpage.pvh_list, FALSE);
1919 simple_unlock(&dest_pg->mdpage.pvh_slock);
1920 }
1921 /*
1922 * Map the pages into the page hook points, copy them, and purge
1923 * the cache for the appropriate page. Invalidate the TLB
1924 * as required.
1925 */
1926 *page_hook0.pte = L2_PTE(src & PG_FRAME, AP_KRW);
1927 *page_hook1.pte = L2_PTE(dest & PG_FRAME, AP_KRW);
1928 cpu_tlb_flushD_SE(page_hook0.va);
1929 cpu_tlb_flushD_SE(page_hook1.va);
1930 cpu_cpwait();
1931 bcopy_page(page_hook0.va, page_hook1.va);
1932 cpu_dcache_wbinv_range(page_hook0.va, NBPG);
1933 cpu_dcache_wbinv_range(page_hook1.va, NBPG);
1934 }
1935
1936 #if 0
1937 void
1938 pmap_pte_addref(pmap, va)
1939 struct pmap *pmap;
1940 vaddr_t va;
1941 {
1942 pd_entry_t *pde;
1943 paddr_t pa;
1944 struct vm_page *m;
1945
1946 if (pmap == pmap_kernel())
1947 return;
1948
1949 pde = pmap_pde(pmap, va & ~(3 << PDSHIFT));
1950 pa = pmap_pte_pa(pde);
1951 m = PHYS_TO_VM_PAGE(pa);
1952 ++m->wire_count;
1953 #ifdef MYCROFT_HACK
1954 printf("addref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
1955 pmap, va, pde, pa, m, m->wire_count);
1956 #endif
1957 }
1958
1959 void
1960 pmap_pte_delref(pmap, va)
1961 struct pmap *pmap;
1962 vaddr_t va;
1963 {
1964 pd_entry_t *pde;
1965 paddr_t pa;
1966 struct vm_page *m;
1967
1968 if (pmap == pmap_kernel())
1969 return;
1970
1971 pde = pmap_pde(pmap, va & ~(3 << PDSHIFT));
1972 pa = pmap_pte_pa(pde);
1973 m = PHYS_TO_VM_PAGE(pa);
1974 --m->wire_count;
1975 #ifdef MYCROFT_HACK
1976 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
1977 pmap, va, pde, pa, m, m->wire_count);
1978 #endif
1979 if (m->wire_count == 0) {
1980 #ifdef MYCROFT_HACK
1981 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p\n",
1982 pmap, va, pde, pa, m);
1983 #endif
1984 pmap_unmap_in_l1(pmap, va);
1985 uvm_pagefree(m);
1986 --pmap->pm_stats.resident_count;
1987 }
1988 }
1989 #else
1990 #define pmap_pte_addref(pmap, va)
1991 #define pmap_pte_delref(pmap, va)
1992 #endif
1993
1994 /*
1995 * Since we have a virtually indexed cache, we may need to inhibit caching if
1996 * there is more than one mapping and at least one of them is writable.
1997 * Since we purge the cache on every context switch, we only need to check for
1998 * other mappings within the same pmap, or kernel_pmap.
1999 * This function is also called when a page is unmapped, to possibly reenable
2000 * caching on any remaining mappings.
2001 *
2002 * The code implements the following logic, where:
2003 *
2004 * KW = # of kernel read/write pages
2005 * KR = # of kernel read only pages
2006 * UW = # of user read/write pages
2007 * UR = # of user read only pages
2008 * OW = # of user read/write pages in another pmap, then
2009 *
2010 * KC = kernel mapping is cacheable
2011 * UC = user mapping is cacheable
2012 *
2013 * KW=0,KR=0 KW=0,KR>0 KW=1,KR=0 KW>1,KR>=0
2014 * +---------------------------------------------
2015 * UW=0,UR=0,OW=0 | --- KC=1 KC=1 KC=0
2016 * UW=0,UR>0,OW=0 | UC=1 KC=1,UC=1 KC=0,UC=0 KC=0,UC=0
2017 * UW=0,UR>0,OW>0 | UC=1 KC=0,UC=1 KC=0,UC=0 KC=0,UC=0
2018 * UW=1,UR=0,OW=0 | UC=1 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0
2019 * UW>1,UR>=0,OW>=0 | UC=0 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0
2020 *
2021 * Note that the pmap must have it's ptes mapped in, and passed with ptes.
2022 */
2023 __inline static void
2024 pmap_vac_me_harder(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2025 boolean_t clear_cache)
2026 {
2027 if (pmap == pmap_kernel())
2028 pmap_vac_me_kpmap(pmap, pg, ptes, clear_cache);
2029 else
2030 pmap_vac_me_user(pmap, pg, ptes, clear_cache);
2031 }
2032
2033 static void
2034 pmap_vac_me_kpmap(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2035 boolean_t clear_cache)
2036 {
2037 int user_entries = 0;
2038 int user_writable = 0;
2039 int user_cacheable = 0;
2040 int kernel_entries = 0;
2041 int kernel_writable = 0;
2042 int kernel_cacheable = 0;
2043 struct pv_entry *pv;
2044 struct pmap *last_pmap = pmap;
2045
2046 #ifdef DIAGNOSTIC
2047 if (pmap != pmap_kernel())
2048 panic("pmap_vac_me_kpmap: pmap != pmap_kernel()");
2049 #endif
2050
2051 /*
2052 * Pass one, see if there are both kernel and user pmaps for
2053 * this page. Calculate whether there are user-writable or
2054 * kernel-writable pages.
2055 */
2056 for (pv = pg->mdpage.pvh_list; pv != NULL; pv = pv->pv_next) {
2057 if (pv->pv_pmap != pmap) {
2058 user_entries++;
2059 if (pv->pv_flags & PT_Wr)
2060 user_writable++;
2061 if ((pv->pv_flags & PT_NC) == 0)
2062 user_cacheable++;
2063 } else {
2064 kernel_entries++;
2065 if (pv->pv_flags & PT_Wr)
2066 kernel_writable++;
2067 if ((pv->pv_flags & PT_NC) == 0)
2068 kernel_cacheable++;
2069 }
2070 }
2071
2072 /*
2073 * We know we have just been updating a kernel entry, so if
2074 * all user pages are already cacheable, then there is nothing
2075 * further to do.
2076 */
2077 if (kernel_entries == 0 &&
2078 user_cacheable == user_entries)
2079 return;
2080
2081 if (user_entries) {
2082 /*
2083 * Scan over the list again, for each entry, if it
2084 * might not be set correctly, call pmap_vac_me_user
2085 * to recalculate the settings.
2086 */
2087 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
2088 /*
2089 * We know kernel mappings will get set
2090 * correctly in other calls. We also know
2091 * that if the pmap is the same as last_pmap
2092 * then we've just handled this entry.
2093 */
2094 if (pv->pv_pmap == pmap || pv->pv_pmap == last_pmap)
2095 continue;
2096 /*
2097 * If there are kernel entries and this page
2098 * is writable but non-cacheable, then we can
2099 * skip this entry also.
2100 */
2101 if (kernel_entries > 0 &&
2102 (pv->pv_flags & (PT_NC | PT_Wr)) ==
2103 (PT_NC | PT_Wr))
2104 continue;
2105 /*
2106 * Similarly if there are no kernel-writable
2107 * entries and the page is already
2108 * read-only/cacheable.
2109 */
2110 if (kernel_writable == 0 &&
2111 (pv->pv_flags & (PT_NC | PT_Wr)) == 0)
2112 continue;
2113 /*
2114 * For some of the remaining cases, we know
2115 * that we must recalculate, but for others we
2116 * can't tell if they are correct or not, so
2117 * we recalculate anyway.
2118 */
2119 pmap_unmap_ptes(last_pmap);
2120 last_pmap = pv->pv_pmap;
2121 ptes = pmap_map_ptes(last_pmap);
2122 pmap_vac_me_user(last_pmap, pg, ptes,
2123 pmap_is_curpmap(last_pmap));
2124 }
2125 /* Restore the pte mapping that was passed to us. */
2126 if (last_pmap != pmap) {
2127 pmap_unmap_ptes(last_pmap);
2128 ptes = pmap_map_ptes(pmap);
2129 }
2130 if (kernel_entries == 0)
2131 return;
2132 }
2133
2134 pmap_vac_me_user(pmap, pg, ptes, clear_cache);
2135 return;
2136 }
2137
2138 static void
2139 pmap_vac_me_user(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2140 boolean_t clear_cache)
2141 {
2142 struct pmap *kpmap = pmap_kernel();
2143 struct pv_entry *pv, *npv;
2144 int entries = 0;
2145 int writable = 0;
2146 int cacheable_entries = 0;
2147 int kern_cacheable = 0;
2148 int other_writable = 0;
2149
2150 pv = pg->mdpage.pvh_list;
2151 KASSERT(ptes != NULL);
2152
2153 /*
2154 * Count mappings and writable mappings in this pmap.
2155 * Include kernel mappings as part of our own.
2156 * Keep a pointer to the first one.
2157 */
2158 for (npv = pv; npv; npv = npv->pv_next) {
2159 /* Count mappings in the same pmap */
2160 if (pmap == npv->pv_pmap ||
2161 kpmap == npv->pv_pmap) {
2162 if (entries++ == 0)
2163 pv = npv;
2164 /* Cacheable mappings */
2165 if ((npv->pv_flags & PT_NC) == 0) {
2166 cacheable_entries++;
2167 if (kpmap == npv->pv_pmap)
2168 kern_cacheable++;
2169 }
2170 /* Writable mappings */
2171 if (npv->pv_flags & PT_Wr)
2172 ++writable;
2173 } else if (npv->pv_flags & PT_Wr)
2174 other_writable = 1;
2175 }
2176
2177 PDEBUG(3,printf("pmap_vac_me_harder: pmap %p Entries %d, "
2178 "writable %d cacheable %d %s\n", pmap, entries, writable,
2179 cacheable_entries, clear_cache ? "clean" : "no clean"));
2180
2181 /*
2182 * Enable or disable caching as necessary.
2183 * Note: the first entry might be part of the kernel pmap,
2184 * so we can't assume this is indicative of the state of the
2185 * other (maybe non-kpmap) entries.
2186 */
2187 if ((entries > 1 && writable) ||
2188 (entries > 0 && pmap == kpmap && other_writable)) {
2189 if (cacheable_entries == 0)
2190 return;
2191 for (npv = pv; npv; npv = npv->pv_next) {
2192 if ((pmap == npv->pv_pmap
2193 || kpmap == npv->pv_pmap) &&
2194 (npv->pv_flags & PT_NC) == 0) {
2195 ptes[arm_byte_to_page(npv->pv_va)] &=
2196 ~(PT_C | PT_B);
2197 npv->pv_flags |= PT_NC;
2198 /*
2199 * If this page needs flushing from the
2200 * cache, and we aren't going to do it
2201 * below, do it now.
2202 */
2203 if ((cacheable_entries < 4 &&
2204 (clear_cache || npv->pv_pmap == kpmap)) ||
2205 (npv->pv_pmap == kpmap &&
2206 !clear_cache && kern_cacheable < 4)) {
2207 cpu_idcache_wbinv_range(npv->pv_va,
2208 NBPG);
2209 cpu_tlb_flushID_SE(npv->pv_va);
2210 }
2211 }
2212 }
2213 if ((clear_cache && cacheable_entries >= 4) ||
2214 kern_cacheable >= 4) {
2215 cpu_idcache_wbinv_all();
2216 cpu_tlb_flushID();
2217 }
2218 cpu_cpwait();
2219 } else if (entries > 0) {
2220 /*
2221 * Turn cacheing back on for some pages. If it is a kernel
2222 * page, only do so if there are no other writable pages.
2223 */
2224 for (npv = pv; npv; npv = npv->pv_next) {
2225 if ((pmap == npv->pv_pmap ||
2226 (kpmap == npv->pv_pmap && other_writable == 0)) &&
2227 (npv->pv_flags & PT_NC)) {
2228 ptes[arm_byte_to_page(npv->pv_va)] |=
2229 pte_cache_mode;
2230 npv->pv_flags &= ~PT_NC;
2231 }
2232 }
2233 }
2234 }
2235
2236 /*
2237 * pmap_remove()
2238 *
2239 * pmap_remove is responsible for nuking a number of mappings for a range
2240 * of virtual address space in the current pmap. To do this efficiently
2241 * is interesting, because in a number of cases a wide virtual address
2242 * range may be supplied that contains few actual mappings. So, the
2243 * optimisations are:
2244 * 1. Try and skip over hunks of address space for which an L1 entry
2245 * does not exist.
2246 * 2. Build up a list of pages we've hit, up to a maximum, so we can
2247 * maybe do just a partial cache clean. This path of execution is
2248 * complicated by the fact that the cache must be flushed _before_
2249 * the PTE is nuked, being a VAC :-)
2250 * 3. Maybe later fast-case a single page, but I don't think this is
2251 * going to make _that_ much difference overall.
2252 */
2253
2254 #define PMAP_REMOVE_CLEAN_LIST_SIZE 3
2255
2256 void
2257 pmap_remove(pmap, sva, eva)
2258 struct pmap *pmap;
2259 vaddr_t sva;
2260 vaddr_t eva;
2261 {
2262 int cleanlist_idx = 0;
2263 struct pagelist {
2264 vaddr_t va;
2265 pt_entry_t *pte;
2266 } cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
2267 pt_entry_t *pte = 0, *ptes;
2268 paddr_t pa;
2269 int pmap_active;
2270 struct vm_page *pg;
2271
2272 /* Exit quick if there is no pmap */
2273 if (!pmap)
2274 return;
2275
2276 PDEBUG(0, printf("pmap_remove: pmap=%p sva=%08lx eva=%08lx\n", pmap, sva, eva));
2277
2278 sva &= PG_FRAME;
2279 eva &= PG_FRAME;
2280
2281 /*
2282 * we lock in the pmap => vm_page direction
2283 */
2284 PMAP_MAP_TO_HEAD_LOCK();
2285
2286 ptes = pmap_map_ptes(pmap);
2287 /* Get a page table pointer */
2288 while (sva < eva) {
2289 if (pmap_pde_page(pmap_pde(pmap, sva)))
2290 break;
2291 sva = (sva & PD_MASK) + NBPD;
2292 }
2293
2294 pte = &ptes[arm_byte_to_page(sva)];
2295 /* Note if the pmap is active thus require cache and tlb cleans */
2296 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap)
2297 || (pmap == pmap_kernel()))
2298 pmap_active = 1;
2299 else
2300 pmap_active = 0;
2301
2302 /* Now loop along */
2303 while (sva < eva) {
2304 /* Check if we can move to the next PDE (l1 chunk) */
2305 if (!(sva & PT_MASK))
2306 if (!pmap_pde_page(pmap_pde(pmap, sva))) {
2307 sva += NBPD;
2308 pte += arm_byte_to_page(NBPD);
2309 continue;
2310 }
2311
2312 /* We've found a valid PTE, so this page of PTEs has to go. */
2313 if (pmap_pte_v(pte)) {
2314 /* Update statistics */
2315 --pmap->pm_stats.resident_count;
2316
2317 /*
2318 * Add this page to our cache remove list, if we can.
2319 * If, however the cache remove list is totally full,
2320 * then do a complete cache invalidation taking note
2321 * to backtrack the PTE table beforehand, and ignore
2322 * the lists in future because there's no longer any
2323 * point in bothering with them (we've paid the
2324 * penalty, so will carry on unhindered). Otherwise,
2325 * when we fall out, we just clean the list.
2326 */
2327 PDEBUG(10, printf("remove: inv pte at %p(%x) ", pte, *pte));
2328 pa = pmap_pte_pa(pte);
2329
2330 if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
2331 /* Add to the clean list. */
2332 cleanlist[cleanlist_idx].pte = pte;
2333 cleanlist[cleanlist_idx].va = sva;
2334 cleanlist_idx++;
2335 } else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
2336 int cnt;
2337
2338 /* Nuke everything if needed. */
2339 if (pmap_active) {
2340 cpu_idcache_wbinv_all();
2341 cpu_tlb_flushID();
2342 }
2343
2344 /*
2345 * Roll back the previous PTE list,
2346 * and zero out the current PTE.
2347 */
2348 for (cnt = 0; cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
2349 *cleanlist[cnt].pte = 0;
2350 pmap_pte_delref(pmap, cleanlist[cnt].va);
2351 }
2352 *pte = 0;
2353 pmap_pte_delref(pmap, sva);
2354 cleanlist_idx++;
2355 } else {
2356 /*
2357 * We've already nuked the cache and
2358 * TLB, so just carry on regardless,
2359 * and we won't need to do it again
2360 */
2361 *pte = 0;
2362 pmap_pte_delref(pmap, sva);
2363 }
2364
2365 /*
2366 * Update flags. In a number of circumstances,
2367 * we could cluster a lot of these and do a
2368 * number of sequential pages in one go.
2369 */
2370 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
2371 struct pv_entry *pve;
2372 simple_lock(&pg->mdpage.pvh_slock);
2373 pve = pmap_remove_pv(pg, pmap, sva);
2374 pmap_free_pv(pmap, pve);
2375 pmap_vac_me_harder(pmap, pg, ptes, FALSE);
2376 simple_unlock(&pg->mdpage.pvh_slock);
2377 }
2378 }
2379 sva += NBPG;
2380 pte++;
2381 }
2382
2383 pmap_unmap_ptes(pmap);
2384 /*
2385 * Now, if we've fallen through down to here, chances are that there
2386 * are less than PMAP_REMOVE_CLEAN_LIST_SIZE mappings left.
2387 */
2388 if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
2389 u_int cnt;
2390
2391 for (cnt = 0; cnt < cleanlist_idx; cnt++) {
2392 if (pmap_active) {
2393 cpu_idcache_wbinv_range(cleanlist[cnt].va,
2394 NBPG);
2395 *cleanlist[cnt].pte = 0;
2396 cpu_tlb_flushID_SE(cleanlist[cnt].va);
2397 } else
2398 *cleanlist[cnt].pte = 0;
2399 pmap_pte_delref(pmap, cleanlist[cnt].va);
2400 }
2401 }
2402 PMAP_MAP_TO_HEAD_UNLOCK();
2403 }
2404
2405 /*
2406 * Routine: pmap_remove_all
2407 * Function:
2408 * Removes this physical page from
2409 * all physical maps in which it resides.
2410 * Reflects back modify bits to the pager.
2411 */
2412
2413 static void
2414 pmap_remove_all(pg)
2415 struct vm_page *pg;
2416 {
2417 struct pv_entry *pv, *npv;
2418 struct pmap *pmap;
2419 pt_entry_t *pte, *ptes;
2420
2421 PDEBUG(0, printf("pmap_remove_all: pa=%lx ", VM_PAGE_TO_PHYS(pg)));
2422
2423 /* set vm_page => pmap locking */
2424 PMAP_HEAD_TO_MAP_LOCK();
2425
2426 simple_lock(&pg->mdpage.pvh_slock);
2427
2428 pv = pg->mdpage.pvh_list;
2429 if (pv == NULL) {
2430 PDEBUG(0, printf("free page\n"));
2431 simple_unlock(&pg->mdpage.pvh_slock);
2432 PMAP_HEAD_TO_MAP_UNLOCK();
2433 return;
2434 }
2435 pmap_clean_page(pv, FALSE);
2436
2437 while (pv) {
2438 pmap = pv->pv_pmap;
2439 ptes = pmap_map_ptes(pmap);
2440 pte = &ptes[arm_byte_to_page(pv->pv_va)];
2441
2442 PDEBUG(0, printf("[%p,%08x,%08lx,%08x] ", pmap, *pte,
2443 pv->pv_va, pv->pv_flags));
2444 #ifdef DEBUG
2445 if (!pmap_pde_page(pmap_pde(pmap, pv->pv_va)) ||
2446 !pmap_pte_v(pte) || pmap_pte_pa(pte) != pa)
2447 panic("pmap_remove_all: bad mapping");
2448 #endif /* DEBUG */
2449
2450 /*
2451 * Update statistics
2452 */
2453 --pmap->pm_stats.resident_count;
2454
2455 /* Wired bit */
2456 if (pv->pv_flags & PT_W)
2457 --pmap->pm_stats.wired_count;
2458
2459 /*
2460 * Invalidate the PTEs.
2461 * XXX: should cluster them up and invalidate as many
2462 * as possible at once.
2463 */
2464
2465 #ifdef needednotdone
2466 reduce wiring count on page table pages as references drop
2467 #endif
2468
2469 *pte = 0;
2470 pmap_pte_delref(pmap, pv->pv_va);
2471
2472 npv = pv->pv_next;
2473 pmap_free_pv(pmap, pv);
2474 pv = npv;
2475 pmap_unmap_ptes(pmap);
2476 }
2477 pg->mdpage.pvh_list = NULL;
2478 simple_unlock(&pg->mdpage.pvh_slock);
2479 PMAP_HEAD_TO_MAP_UNLOCK();
2480
2481 PDEBUG(0, printf("done\n"));
2482 cpu_tlb_flushID();
2483 cpu_cpwait();
2484 }
2485
2486
2487 /*
2488 * Set the physical protection on the specified range of this map as requested.
2489 */
2490
2491 void
2492 pmap_protect(pmap, sva, eva, prot)
2493 struct pmap *pmap;
2494 vaddr_t sva;
2495 vaddr_t eva;
2496 vm_prot_t prot;
2497 {
2498 pt_entry_t *pte = NULL, *ptes;
2499 struct vm_page *pg;
2500 int armprot;
2501 int flush = 0;
2502 paddr_t pa;
2503
2504 PDEBUG(0, printf("pmap_protect: pmap=%p %08lx->%08lx %x\n",
2505 pmap, sva, eva, prot));
2506
2507 if (~prot & VM_PROT_READ) {
2508 /* Just remove the mappings. */
2509 pmap_remove(pmap, sva, eva);
2510 /* pmap_update not needed as it should be called by the caller
2511 * of pmap_protect */
2512 return;
2513 }
2514 if (prot & VM_PROT_WRITE) {
2515 /*
2516 * If this is a read->write transition, just ignore it and let
2517 * uvm_fault() take care of it later.
2518 */
2519 return;
2520 }
2521
2522 sva &= PG_FRAME;
2523 eva &= PG_FRAME;
2524
2525 /* Need to lock map->head */
2526 PMAP_MAP_TO_HEAD_LOCK();
2527
2528 ptes = pmap_map_ptes(pmap);
2529 /*
2530 * We need to acquire a pointer to a page table page before entering
2531 * the following loop.
2532 */
2533 while (sva < eva) {
2534 if (pmap_pde_page(pmap_pde(pmap, sva)))
2535 break;
2536 sva = (sva & PD_MASK) + NBPD;
2537 }
2538
2539 pte = &ptes[arm_byte_to_page(sva)];
2540
2541 while (sva < eva) {
2542 /* only check once in a while */
2543 if ((sva & PT_MASK) == 0) {
2544 if (!pmap_pde_page(pmap_pde(pmap, sva))) {
2545 /* We can race ahead here, to the next pde. */
2546 sva += NBPD;
2547 pte += arm_byte_to_page(NBPD);
2548 continue;
2549 }
2550 }
2551
2552 if (!pmap_pte_v(pte))
2553 goto next;
2554
2555 flush = 1;
2556
2557 armprot = 0;
2558 if (sva < VM_MAXUSER_ADDRESS)
2559 armprot |= PT_AP(AP_U);
2560 else if (sva < VM_MAX_ADDRESS)
2561 armprot |= PT_AP(AP_W); /* XXX Ekk what is this ? */
2562 *pte = (*pte & 0xfffff00f) | armprot;
2563
2564 pa = pmap_pte_pa(pte);
2565
2566 /* Get the physical page index */
2567
2568 /* Clear write flag */
2569 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
2570 simple_lock(&pg->mdpage.pvh_slock);
2571 (void) pmap_modify_pv(pmap, sva, pg, PT_Wr, 0);
2572 pmap_vac_me_harder(pmap, pg, ptes, FALSE);
2573 simple_unlock(&pg->mdpage.pvh_slock);
2574 }
2575
2576 next:
2577 sva += NBPG;
2578 pte++;
2579 }
2580 pmap_unmap_ptes(pmap);
2581 PMAP_MAP_TO_HEAD_UNLOCK();
2582 if (flush)
2583 cpu_tlb_flushID();
2584 }
2585
2586 /*
2587 * void pmap_enter(struct pmap *pmap, vaddr_t va, paddr_t pa, vm_prot_t prot,
2588 * int flags)
2589 *
2590 * Insert the given physical page (p) at
2591 * the specified virtual address (v) in the
2592 * target physical map with the protection requested.
2593 *
2594 * If specified, the page will be wired down, meaning
2595 * that the related pte can not be reclaimed.
2596 *
2597 * NB: This is the only routine which MAY NOT lazy-evaluate
2598 * or lose information. That is, this routine must actually
2599 * insert this page into the given map NOW.
2600 */
2601
2602 int
2603 pmap_enter(pmap, va, pa, prot, flags)
2604 struct pmap *pmap;
2605 vaddr_t va;
2606 paddr_t pa;
2607 vm_prot_t prot;
2608 int flags;
2609 {
2610 pt_entry_t *pte, *ptes;
2611 u_int npte;
2612 paddr_t opa;
2613 int nflags;
2614 boolean_t wired = (flags & PMAP_WIRED) != 0;
2615 struct vm_page *pg;
2616 struct pv_entry *pve;
2617 int error;
2618
2619 PDEBUG(5, printf("pmap_enter: V%08lx P%08lx in pmap %p prot=%08x, wired = %d\n",
2620 va, pa, pmap, prot, wired));
2621
2622 #ifdef DIAGNOSTIC
2623 /* Valid address ? */
2624 if (va >= (pmap_curmaxkvaddr))
2625 panic("pmap_enter: too big");
2626 if (pmap != pmap_kernel() && va != 0) {
2627 if (va < VM_MIN_ADDRESS || va >= VM_MAXUSER_ADDRESS)
2628 panic("pmap_enter: kernel page in user map");
2629 } else {
2630 if (va >= VM_MIN_ADDRESS && va < VM_MAXUSER_ADDRESS)
2631 panic("pmap_enter: user page in kernel map");
2632 if (va >= VM_MAXUSER_ADDRESS && va < VM_MAX_ADDRESS)
2633 panic("pmap_enter: entering PT page");
2634 }
2635 #endif
2636 /*
2637 * Get a pointer to the page. Later on in this function, we
2638 * test for a managed page by checking pg != NULL.
2639 */
2640 pg = PHYS_TO_VM_PAGE(pa);
2641
2642 /* get lock */
2643 PMAP_MAP_TO_HEAD_LOCK();
2644 /*
2645 * Get a pointer to the pte for this virtual address. If the
2646 * pte pointer is NULL then we are missing the L2 page table
2647 * so we need to create one.
2648 */
2649 /* XXX horrible hack to get us working with lockdebug */
2650 simple_lock(&pmap->pm_obj.vmobjlock);
2651 pte = pmap_pte(pmap, va);
2652 if (!pte) {
2653 struct vm_page *ptp;
2654 KASSERT(pmap != pmap_kernel()); /* kernel should have pre-grown */
2655
2656 /* if failure is allowed then don't try too hard */
2657 ptp = pmap_get_ptp(pmap, va, flags & PMAP_CANFAIL);
2658 if (ptp == NULL) {
2659 if (flags & PMAP_CANFAIL) {
2660 error = ENOMEM;
2661 goto out;
2662 }
2663 panic("pmap_enter: get ptp failed");
2664 }
2665
2666 pte = pmap_pte(pmap, va);
2667 #ifdef DIAGNOSTIC
2668 if (!pte)
2669 panic("pmap_enter: no pte");
2670 #endif
2671 }
2672
2673 nflags = 0;
2674 if (prot & VM_PROT_WRITE)
2675 nflags |= PT_Wr;
2676 if (wired)
2677 nflags |= PT_W;
2678
2679 /* More debugging info */
2680 PDEBUG(5, printf("pmap_enter: pte for V%08lx = V%p (%08x)\n", va, pte,
2681 *pte));
2682
2683 /* Is the pte valid ? If so then this page is already mapped */
2684 if (pmap_pte_v(pte)) {
2685 /* Get the physical address of the current page mapped */
2686 opa = pmap_pte_pa(pte);
2687
2688 #ifdef MYCROFT_HACK
2689 printf("pmap_enter: pmap=%p va=%lx pa=%lx opa=%lx\n", pmap, va, pa, opa);
2690 #endif
2691
2692 /* Are we mapping the same page ? */
2693 if (opa == pa) {
2694 /* All we must be doing is changing the protection */
2695 PDEBUG(0, printf("Case 02 in pmap_enter (V%08lx P%08lx)\n",
2696 va, pa));
2697
2698 /* Has the wiring changed ? */
2699 if (pg != NULL) {
2700 simple_lock(&pg->mdpage.pvh_slock);
2701 (void) pmap_modify_pv(pmap, va, pg,
2702 PT_Wr | PT_W, nflags);
2703 simple_unlock(&pg->mdpage.pvh_slock);
2704 }
2705 } else {
2706 struct vm_page *opg;
2707
2708 /* We are replacing the page with a new one. */
2709 cpu_idcache_wbinv_range(va, NBPG);
2710
2711 PDEBUG(0, printf("Case 03 in pmap_enter (V%08lx P%08lx P%08lx)\n",
2712 va, pa, opa));
2713
2714 /*
2715 * If it is part of our managed memory then we
2716 * must remove it from the PV list
2717 */
2718 if ((opg = PHYS_TO_VM_PAGE(opa)) != NULL) {
2719 simple_lock(&opg->mdpage.pvh_slock);
2720 pve = pmap_remove_pv(opg, pmap, va);
2721 simple_unlock(&opg->mdpage.pvh_slock);
2722 } else {
2723 pve = NULL;
2724 }
2725
2726 goto enter;
2727 }
2728 } else {
2729 opa = 0;
2730 pve = NULL;
2731 pmap_pte_addref(pmap, va);
2732
2733 /* pte is not valid so we must be hooking in a new page */
2734 ++pmap->pm_stats.resident_count;
2735
2736 enter:
2737 /*
2738 * Enter on the PV list if part of our managed memory
2739 */
2740 if (pmap_initialized && pg != NULL) {
2741 if (pve == NULL) {
2742 pve = pmap_alloc_pv(pmap, ALLOCPV_NEED);
2743 if (pve == NULL) {
2744 if (flags & PMAP_CANFAIL) {
2745 error = ENOMEM;
2746 goto out;
2747 }
2748 panic("pmap_enter: no pv entries available");
2749 }
2750 }
2751 /* enter_pv locks pvh when adding */
2752 pmap_enter_pv(pg, pve, pmap, va, NULL, nflags);
2753 } else {
2754 pg = NULL;
2755 if (pve != NULL)
2756 pmap_free_pv(pmap, pve);
2757 }
2758 }
2759
2760 #ifdef MYCROFT_HACK
2761 if (mycroft_hack)
2762 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);
2763 #endif
2764
2765 /* Construct the pte, giving the correct access. */
2766 npte = (pa & PG_FRAME);
2767
2768 /* VA 0 is magic. */
2769 if (pmap != pmap_kernel() && va != 0)
2770 npte |= PT_AP(AP_U);
2771
2772 if (pmap_initialized && pg != NULL) {
2773 #ifdef DIAGNOSTIC
2774 if ((flags & VM_PROT_ALL) & ~prot)
2775 panic("pmap_enter: access_type exceeds prot");
2776 #endif
2777 npte |= pte_cache_mode;
2778 if (flags & VM_PROT_WRITE) {
2779 npte |= L2_SPAGE | PT_AP(AP_W);
2780 pg->mdpage.pvh_attrs |= PT_H | PT_M;
2781 } else if (flags & VM_PROT_ALL) {
2782 npte |= L2_SPAGE;
2783 pg->mdpage.pvh_attrs |= PT_H;
2784 } else
2785 npte |= L2_INVAL;
2786 } else {
2787 if (prot & VM_PROT_WRITE)
2788 npte |= L2_SPAGE | PT_AP(AP_W);
2789 else if (prot & VM_PROT_ALL)
2790 npte |= L2_SPAGE;
2791 else
2792 npte |= L2_INVAL;
2793 }
2794
2795 #ifdef MYCROFT_HACK
2796 if (mycroft_hack)
2797 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);
2798 #endif
2799
2800 *pte = npte;
2801
2802 if (pmap_initialized && pg != NULL) {
2803 boolean_t pmap_active = FALSE;
2804 /* XXX this will change once the whole of pmap_enter uses
2805 * map_ptes
2806 */
2807 ptes = pmap_map_ptes(pmap);
2808 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap)
2809 || (pmap == pmap_kernel()))
2810 pmap_active = TRUE;
2811 simple_lock(&pg->mdpage.pvh_slock);
2812 pmap_vac_me_harder(pmap, pg, ptes, pmap_active);
2813 simple_unlock(&pg->mdpage.pvh_slock);
2814 pmap_unmap_ptes(pmap);
2815 }
2816
2817 /* Better flush the TLB ... */
2818 cpu_tlb_flushID_SE(va);
2819 error = 0;
2820 out:
2821 simple_unlock(&pmap->pm_obj.vmobjlock);
2822 PMAP_MAP_TO_HEAD_UNLOCK();
2823 PDEBUG(5, printf("pmap_enter: pte = V%p %08x\n", pte, *pte));
2824
2825 return error;
2826 }
2827
2828 /*
2829 * pmap_kenter_pa: enter a kernel mapping
2830 *
2831 * => no need to lock anything assume va is already allocated
2832 * => should be faster than normal pmap enter function
2833 */
2834 void
2835 pmap_kenter_pa(va, pa, prot)
2836 vaddr_t va;
2837 paddr_t pa;
2838 vm_prot_t prot;
2839 {
2840 pt_entry_t *pte;
2841
2842 pte = vtopte(va);
2843 KASSERT(!pmap_pte_v(pte));
2844 *pte = L2_PTE(pa, AP_KRW);
2845 }
2846
2847 void
2848 pmap_kremove(va, len)
2849 vaddr_t va;
2850 vsize_t len;
2851 {
2852 pt_entry_t *pte;
2853
2854 for (len >>= PAGE_SHIFT; len > 0; len--, va += PAGE_SIZE) {
2855
2856 /*
2857 * We assume that we will only be called with small
2858 * regions of memory.
2859 */
2860
2861 KASSERT(pmap_pde_page(pmap_pde(pmap_kernel(), va)));
2862 pte = vtopte(va);
2863 cpu_idcache_wbinv_range(va, PAGE_SIZE);
2864 *pte = 0;
2865 cpu_tlb_flushID_SE(va);
2866 }
2867 }
2868
2869 /*
2870 * pmap_page_protect:
2871 *
2872 * Lower the permission for all mappings to a given page.
2873 */
2874
2875 void
2876 pmap_page_protect(pg, prot)
2877 struct vm_page *pg;
2878 vm_prot_t prot;
2879 {
2880
2881 PDEBUG(0, printf("pmap_page_protect(pa=%lx, prot=%d)\n",
2882 VM_PAGE_TO_PHYS(pg), prot));
2883
2884 switch(prot) {
2885 case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
2886 case VM_PROT_READ|VM_PROT_WRITE:
2887 return;
2888
2889 case VM_PROT_READ:
2890 case VM_PROT_READ|VM_PROT_EXECUTE:
2891 pmap_copy_on_write(pg);
2892 break;
2893
2894 default:
2895 pmap_remove_all(pg);
2896 break;
2897 }
2898 }
2899
2900
2901 /*
2902 * Routine: pmap_unwire
2903 * Function: Clear the wired attribute for a map/virtual-address
2904 * pair.
2905 * In/out conditions:
2906 * The mapping must already exist in the pmap.
2907 */
2908
2909 void
2910 pmap_unwire(pmap, va)
2911 struct pmap *pmap;
2912 vaddr_t va;
2913 {
2914 pt_entry_t *pte;
2915 paddr_t pa;
2916 struct vm_page *pg;
2917
2918 /*
2919 * Make sure pmap is valid. -dct
2920 */
2921 if (pmap == NULL)
2922 return;
2923
2924 /* Get the pte */
2925 pte = pmap_pte(pmap, va);
2926 if (!pte)
2927 return;
2928
2929 /* Extract the physical address of the page */
2930 pa = pmap_pte_pa(pte);
2931
2932 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
2933 return;
2934
2935 simple_lock(&pg->mdpage.pvh_slock);
2936 /* Update the wired bit in the pv entry for this page. */
2937 (void) pmap_modify_pv(pmap, va, pg, PT_W, 0);
2938 simple_unlock(&pg->mdpage.pvh_slock);
2939 }
2940
2941 /*
2942 * pt_entry_t *pmap_pte(struct pmap *pmap, vaddr_t va)
2943 *
2944 * Return the pointer to a page table entry corresponding to the supplied
2945 * virtual address.
2946 *
2947 * The page directory is first checked to make sure that a page table
2948 * for the address in question exists and if it does a pointer to the
2949 * entry is returned.
2950 *
2951 * The way this works is that that the kernel page tables are mapped
2952 * into the memory map at ALT_PAGE_TBLS_BASE to ALT_PAGE_TBLS_BASE+4MB.
2953 * This allows page tables to be located quickly.
2954 */
2955 pt_entry_t *
2956 pmap_pte(pmap, va)
2957 struct pmap *pmap;
2958 vaddr_t va;
2959 {
2960 pt_entry_t *ptp;
2961 pt_entry_t *result;
2962
2963 /* The pmap must be valid */
2964 if (!pmap)
2965 return(NULL);
2966
2967 /* Return the address of the pte */
2968 PDEBUG(10, printf("pmap_pte: pmap=%p va=V%08lx pde = V%p (%08X)\n",
2969 pmap, va, pmap_pde(pmap, va), *(pmap_pde(pmap, va))));
2970
2971 /* Do we have a valid pde ? If not we don't have a page table */
2972 if (!pmap_pde_page(pmap_pde(pmap, va))) {
2973 PDEBUG(0, printf("pmap_pte: failed - pde = %p\n",
2974 pmap_pde(pmap, va)));
2975 return(NULL);
2976 }
2977
2978 PDEBUG(10, printf("pmap pagetable = P%08lx current = P%08x\n",
2979 pmap->pm_pptpt, (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE
2980 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) +
2981 (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & PG_FRAME)));
2982
2983 /*
2984 * If the pmap is the kernel pmap or the pmap is the active one
2985 * then we can just return a pointer to entry relative to
2986 * PROCESS_PAGE_TBLS_BASE.
2987 * Otherwise we need to map the page tables to an alternative
2988 * address and reference them there.
2989 */
2990 if (pmap == pmap_kernel() || pmap->pm_pptpt
2991 == (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE
2992 + ((PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) &
2993 ~3) + (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & PG_FRAME)) {
2994 ptp = (pt_entry_t *)PROCESS_PAGE_TBLS_BASE;
2995 } else {
2996 struct proc *p = curproc;
2997
2998 /* If we don't have a valid curproc use proc0 */
2999 /* Perhaps we should just use kernel_pmap instead */
3000 if (p == NULL)
3001 p = &proc0;
3002 #ifdef DIAGNOSTIC
3003 /*
3004 * The pmap should always be valid for the process so
3005 * panic if it is not.
3006 */
3007 if (!p->p_vmspace || !p->p_vmspace->vm_map.pmap) {
3008 printf("pmap_pte: va=%08lx p=%p vm=%p\n",
3009 va, p, p->p_vmspace);
3010 console_debugger();
3011 }
3012 /*
3013 * The pmap for the current process should be mapped. If it
3014 * is not then we have a problem.
3015 */
3016 if (p->p_vmspace->vm_map.pmap->pm_pptpt !=
3017 (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE
3018 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) +
3019 (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & PG_FRAME)) {
3020 printf("pmap pagetable = P%08lx current = P%08x ",
3021 pmap->pm_pptpt, (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE
3022 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) +
3023 (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) &
3024 PG_FRAME));
3025 printf("pptpt=%lx\n", p->p_vmspace->vm_map.pmap->pm_pptpt);
3026 panic("pmap_pte: current and pmap mismatch\n");
3027 }
3028 #endif
3029
3030 ptp = (pt_entry_t *)ALT_PAGE_TBLS_BASE;
3031 pmap_map_in_l1(p->p_vmspace->vm_map.pmap, ALT_PAGE_TBLS_BASE,
3032 pmap->pm_pptpt, FALSE);
3033 cpu_tlb_flushD();
3034 cpu_cpwait();
3035 }
3036 PDEBUG(10, printf("page tables base = %p offset=%lx\n", ptp,
3037 ((va >> (PGSHIFT-2)) & ~3)));
3038 result = (pt_entry_t *)((char *)ptp + ((va >> (PGSHIFT-2)) & ~3));
3039 return(result);
3040 }
3041
3042 /*
3043 * Routine: pmap_extract
3044 * Function:
3045 * Extract the physical page address associated
3046 * with the given map/virtual_address pair.
3047 */
3048 boolean_t
3049 pmap_extract(pmap, va, pap)
3050 struct pmap *pmap;
3051 vaddr_t va;
3052 paddr_t *pap;
3053 {
3054 pd_entry_t *pde;
3055 pt_entry_t *pte, *ptes;
3056 paddr_t pa;
3057 boolean_t rv = TRUE;
3058
3059 PDEBUG(5, printf("pmap_extract: pmap=%p, va=V%08lx\n", pmap, va));
3060
3061 /*
3062 * Get the pte for this virtual address.
3063 */
3064 pde = pmap_pde(pmap, va);
3065 ptes = pmap_map_ptes(pmap);
3066 pte = &ptes[arm_byte_to_page(va)];
3067
3068 if (pmap_pde_section(pde)) {
3069 pa = (*pde & PD_MASK) | (va & (L1_SEC_SIZE - 1));
3070 goto out;
3071 } else if (pmap_pde_page(pde) == 0 || pmap_pte_v(pte) == 0) {
3072 rv = FALSE;
3073 goto out;
3074 }
3075
3076 if ((*pte & L2_MASK) == L2_LPAGE) {
3077 /* Extract the physical address from the pte */
3078 pa = *pte & ~(L2_LPAGE_SIZE - 1);
3079
3080 PDEBUG(5, printf("pmap_extract: LPAGE pa = P%08lx\n",
3081 (pa | (va & (L2_LPAGE_SIZE - 1)))));
3082
3083 if (pap != NULL)
3084 *pap = pa | (va & (L2_LPAGE_SIZE - 1));
3085 goto out;
3086 }
3087
3088 /* Extract the physical address from the pte */
3089 pa = pmap_pte_pa(pte);
3090
3091 PDEBUG(5, printf("pmap_extract: SPAGE pa = P%08lx\n",
3092 (pa | (va & ~PG_FRAME))));
3093
3094 if (pap != NULL)
3095 *pap = pa | (va & ~PG_FRAME);
3096 out:
3097 pmap_unmap_ptes(pmap);
3098 return (rv);
3099 }
3100
3101
3102 /*
3103 * Copy the range specified by src_addr/len from the source map to the
3104 * range dst_addr/len in the destination map.
3105 *
3106 * This routine is only advisory and need not do anything.
3107 */
3108
3109 void
3110 pmap_copy(dst_pmap, src_pmap, dst_addr, len, src_addr)
3111 struct pmap *dst_pmap;
3112 struct pmap *src_pmap;
3113 vaddr_t dst_addr;
3114 vsize_t len;
3115 vaddr_t src_addr;
3116 {
3117 PDEBUG(0, printf("pmap_copy(%p, %p, %lx, %lx, %lx)\n",
3118 dst_pmap, src_pmap, dst_addr, len, src_addr));
3119 }
3120
3121 #if defined(PMAP_DEBUG)
3122 void
3123 pmap_dump_pvlist(phys, m)
3124 vaddr_t phys;
3125 char *m;
3126 {
3127 struct vm_page *pg;
3128 struct pv_entry *pv;
3129
3130 if ((pg = PHYS_TO_VM_PAGE(phys)) == NULL) {
3131 printf("INVALID PA\n");
3132 return;
3133 }
3134 simple_lock(&pg->mdpage.pvh_slock);
3135 printf("%s %08lx:", m, phys);
3136 if (pg->mdpage.pvh_list == NULL) {
3137 printf(" no mappings\n");
3138 return;
3139 }
3140
3141 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next)
3142 printf(" pmap %p va %08lx flags %08x", pv->pv_pmap,
3143 pv->pv_va, pv->pv_flags);
3144
3145 printf("\n");
3146 simple_unlock(&pg->mdpage.pvh_slock);
3147 }
3148
3149 #endif /* PMAP_DEBUG */
3150
3151 __inline static boolean_t
3152 pmap_testbit(pg, setbits)
3153 struct vm_page *pg;
3154 unsigned int setbits;
3155 {
3156
3157 PDEBUG(1, printf("pmap_testbit: pa=%08lx set=%08x\n",
3158 VM_PAGE_TO_PHYS(pg), setbits));
3159
3160 /*
3161 * Check saved info only
3162 */
3163 if (pg->mdpage.pvh_attrs & setbits) {
3164 PDEBUG(0, printf("pmap_attributes = %02x\n",
3165 pg->mdpage.pvh_attrs));
3166 return(TRUE);
3167 }
3168
3169 return(FALSE);
3170 }
3171
3172 static pt_entry_t *
3173 pmap_map_ptes(struct pmap *pmap)
3174 {
3175 struct proc *p;
3176
3177 /* the kernel's pmap is always accessible */
3178 if (pmap == pmap_kernel()) {
3179 return (pt_entry_t *)PROCESS_PAGE_TBLS_BASE ;
3180 }
3181
3182 if (pmap_is_curpmap(pmap)) {
3183 simple_lock(&pmap->pm_obj.vmobjlock);
3184 return (pt_entry_t *)PROCESS_PAGE_TBLS_BASE;
3185 }
3186
3187 p = curproc;
3188
3189 if (p == NULL)
3190 p = &proc0;
3191
3192 /* need to lock both curpmap and pmap: use ordered locking */
3193 if ((unsigned) pmap < (unsigned) curproc->p_vmspace->vm_map.pmap) {
3194 simple_lock(&pmap->pm_obj.vmobjlock);
3195 simple_lock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3196 } else {
3197 simple_lock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3198 simple_lock(&pmap->pm_obj.vmobjlock);
3199 }
3200
3201 pmap_map_in_l1(p->p_vmspace->vm_map.pmap, ALT_PAGE_TBLS_BASE,
3202 pmap->pm_pptpt, FALSE);
3203 cpu_tlb_flushD();
3204 cpu_cpwait();
3205 return (pt_entry_t *)ALT_PAGE_TBLS_BASE;
3206 }
3207
3208 /*
3209 * pmap_unmap_ptes: unlock the PTE mapping of "pmap"
3210 */
3211
3212 static void
3213 pmap_unmap_ptes(pmap)
3214 struct pmap *pmap;
3215 {
3216 if (pmap == pmap_kernel()) {
3217 return;
3218 }
3219 if (pmap_is_curpmap(pmap)) {
3220 simple_unlock(&pmap->pm_obj.vmobjlock);
3221 } else {
3222 simple_unlock(&pmap->pm_obj.vmobjlock);
3223 simple_unlock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3224 }
3225 }
3226
3227 /*
3228 * Modify pte bits for all ptes corresponding to the given physical address.
3229 * We use `maskbits' rather than `clearbits' because we're always passing
3230 * constants and the latter would require an extra inversion at run-time.
3231 */
3232
3233 static void
3234 pmap_clearbit(pg, maskbits)
3235 struct vm_page *pg;
3236 unsigned int maskbits;
3237 {
3238 struct pv_entry *pv;
3239 pt_entry_t *pte;
3240 vaddr_t va;
3241 int tlbentry;
3242
3243 PDEBUG(1, printf("pmap_clearbit: pa=%08lx mask=%08x\n",
3244 VM_PAGE_TO_PHYS(pg), maskbits));
3245
3246 tlbentry = 0;
3247
3248 PMAP_HEAD_TO_MAP_LOCK();
3249 simple_lock(&pg->mdpage.pvh_slock);
3250
3251 /*
3252 * Clear saved attributes (modify, reference)
3253 */
3254 pg->mdpage.pvh_attrs &= ~maskbits;
3255
3256 if (pg->mdpage.pvh_list == NULL) {
3257 simple_unlock(&pg->mdpage.pvh_slock);
3258 PMAP_HEAD_TO_MAP_UNLOCK();
3259 return;
3260 }
3261
3262 /*
3263 * Loop over all current mappings setting/clearing as appropos
3264 */
3265 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
3266 va = pv->pv_va;
3267 pv->pv_flags &= ~maskbits;
3268 pte = pmap_pte(pv->pv_pmap, va);
3269 KASSERT(pte != NULL);
3270 if (maskbits & (PT_Wr|PT_M)) {
3271 if ((pv->pv_flags & PT_NC)) {
3272 /*
3273 * Entry is not cacheable: reenable
3274 * the cache, nothing to flush
3275 *
3276 * Don't turn caching on again if this
3277 * is a modified emulation. This
3278 * would be inconsitent with the
3279 * settings created by
3280 * pmap_vac_me_harder().
3281 *
3282 * There's no need to call
3283 * pmap_vac_me_harder() here: all
3284 * pages are loosing their write
3285 * permission.
3286 *
3287 */
3288 if (maskbits & PT_Wr) {
3289 *pte |= pte_cache_mode;
3290 pv->pv_flags &= ~PT_NC;
3291 }
3292 } else if (pmap_is_curpmap(pv->pv_pmap))
3293 /*
3294 * Entry is cacheable: check if pmap is
3295 * current if it is flush it,
3296 * otherwise it won't be in the cache
3297 */
3298 cpu_idcache_wbinv_range(pv->pv_va, NBPG);
3299
3300 /* make the pte read only */
3301 *pte &= ~PT_AP(AP_W);
3302 }
3303
3304 if (maskbits & PT_H)
3305 *pte = (*pte & ~L2_MASK) | L2_INVAL;
3306
3307 if (pmap_is_curpmap(pv->pv_pmap))
3308 /*
3309 * if we had cacheable pte's we'd clean the
3310 * pte out to memory here
3311 *
3312 * flush tlb entry as it's in the current pmap
3313 */
3314 cpu_tlb_flushID_SE(pv->pv_va);
3315 }
3316 cpu_cpwait();
3317
3318 simple_unlock(&pg->mdpage.pvh_slock);
3319 PMAP_HEAD_TO_MAP_UNLOCK();
3320 }
3321
3322
3323 boolean_t
3324 pmap_clear_modify(pg)
3325 struct vm_page *pg;
3326 {
3327 boolean_t rv;
3328
3329 PDEBUG(0, printf("pmap_clear_modify pa=%08lx\n", VM_PAGE_TO_PHYS(pg)));
3330 rv = pmap_testbit(pg, PT_M);
3331 pmap_clearbit(pg, PT_M);
3332 return rv;
3333 }
3334
3335
3336 boolean_t
3337 pmap_clear_reference(pg)
3338 struct vm_page *pg;
3339 {
3340 boolean_t rv;
3341
3342 PDEBUG(0, printf("pmap_clear_reference pa=%08lx\n",
3343 VM_PAGE_TO_PHYS(pg)));
3344 rv = pmap_testbit(pg, PT_H);
3345 pmap_clearbit(pg, PT_H);
3346 return rv;
3347 }
3348
3349
3350 void
3351 pmap_copy_on_write(pg)
3352 struct vm_page *pg;
3353 {
3354 PDEBUG(0, printf("pmap_copy_on_write pa=%08lx\n", VM_PAGE_TO_PHYS(pg)));
3355 pmap_clearbit(pg, PT_Wr);
3356 }
3357
3358
3359 boolean_t
3360 pmap_is_modified(pg)
3361 struct vm_page *pg;
3362 {
3363 boolean_t result;
3364
3365 result = pmap_testbit(pg, PT_M);
3366 PDEBUG(1, printf("pmap_is_modified pa=%08lx %x\n",
3367 VM_PAGE_TO_PHYS(pg), result));
3368 return (result);
3369 }
3370
3371
3372 boolean_t
3373 pmap_is_referenced(pg)
3374 struct vm_page *pg;
3375 {
3376 boolean_t result;
3377
3378 result = pmap_testbit(pg, PT_H);
3379 PDEBUG(0, printf("pmap_is_referenced pa=%08lx %x\n",
3380 VM_PAGE_TO_PHYS(pg), result));
3381 return (result);
3382 }
3383
3384
3385 int
3386 pmap_modified_emulation(pmap, va)
3387 struct pmap *pmap;
3388 vaddr_t va;
3389 {
3390 pt_entry_t *pte;
3391 paddr_t pa;
3392 struct vm_page *pg;
3393 u_int flags;
3394
3395 PDEBUG(2, printf("pmap_modified_emulation\n"));
3396
3397 /* Get the pte */
3398 pte = pmap_pte(pmap, va);
3399 if (!pte) {
3400 PDEBUG(2, printf("no pte\n"));
3401 return(0);
3402 }
3403
3404 PDEBUG(1, printf("*pte=%08x\n", *pte));
3405
3406 /* Check for a zero pte */
3407 if (*pte == 0)
3408 return(0);
3409
3410 /* This can happen if user code tries to access kernel memory. */
3411 if ((*pte & PT_AP(AP_W)) != 0)
3412 return (0);
3413
3414 /* Extract the physical address of the page */
3415 pa = pmap_pte_pa(pte);
3416 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3417 return(0);
3418
3419 /* Get the current flags for this page. */
3420 PMAP_HEAD_TO_MAP_LOCK();
3421 simple_lock(&pg->mdpage.pvh_slock);
3422
3423 flags = pmap_modify_pv(pmap, va, pg, 0, 0);
3424 PDEBUG(2, printf("pmap_modified_emulation: flags = %08x\n", flags));
3425
3426 /*
3427 * Do the flags say this page is writable ? If not then it is a
3428 * genuine write fault. If yes then the write fault is our fault
3429 * as we did not reflect the write access in the PTE. Now we know
3430 * a write has occurred we can correct this and also set the
3431 * modified bit
3432 */
3433 if (~flags & PT_Wr) {
3434 simple_unlock(&pg->mdpage.pvh_slock);
3435 PMAP_HEAD_TO_MAP_UNLOCK();
3436 return(0);
3437 }
3438
3439 PDEBUG(0, printf("pmap_modified_emulation: Got a hit va=%08lx, pte = %p (%08x)\n",
3440 va, pte, *pte));
3441 pg->mdpage.pvh_attrs |= PT_H | PT_M;
3442
3443 /*
3444 * Re-enable write permissions for the page. No need to call
3445 * pmap_vac_me_harder(), since this is just a
3446 * modified-emulation fault, and the PT_Wr bit isn't changing. We've
3447 * already set the cacheable bits based on the assumption that we
3448 * can write to this page.
3449 */
3450 *pte = (*pte & ~L2_MASK) | L2_SPAGE | PT_AP(AP_W);
3451 PDEBUG(0, printf("->(%08x)\n", *pte));
3452
3453 simple_unlock(&pg->mdpage.pvh_slock);
3454 PMAP_HEAD_TO_MAP_UNLOCK();
3455 /* Return, indicating the problem has been dealt with */
3456 cpu_tlb_flushID_SE(va);
3457 cpu_cpwait();
3458 return(1);
3459 }
3460
3461
3462 int
3463 pmap_handled_emulation(pmap, va)
3464 struct pmap *pmap;
3465 vaddr_t va;
3466 {
3467 pt_entry_t *pte;
3468 paddr_t pa;
3469 struct vm_page *pg;
3470
3471 PDEBUG(2, printf("pmap_handled_emulation\n"));
3472
3473 /* Get the pte */
3474 pte = pmap_pte(pmap, va);
3475 if (!pte) {
3476 PDEBUG(2, printf("no pte\n"));
3477 return(0);
3478 }
3479
3480 PDEBUG(1, printf("*pte=%08x\n", *pte));
3481
3482 /* Check for a zero pte */
3483 if (*pte == 0)
3484 return(0);
3485
3486 /* This can happen if user code tries to access kernel memory. */
3487 if ((*pte & L2_MASK) != L2_INVAL)
3488 return (0);
3489
3490 /* Extract the physical address of the page */
3491 pa = pmap_pte_pa(pte);
3492 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3493 return (0);
3494
3495 /*
3496 * Ok we just enable the pte and mark the attibs as handled
3497 */
3498 PDEBUG(0, printf("pmap_handled_emulation: Got a hit va=%08lx pte = %p (%08x)\n",
3499 va, pte, *pte));
3500 pg->mdpage.pvh_attrs |= PT_H;
3501 *pte = (*pte & ~L2_MASK) | L2_SPAGE;
3502 PDEBUG(0, printf("->(%08x)\n", *pte));
3503
3504 /* Return, indicating the problem has been dealt with */
3505 cpu_tlb_flushID_SE(va);
3506 cpu_cpwait();
3507 return(1);
3508 }
3509
3510
3511
3512
3513 /*
3514 * pmap_collect: free resources held by a pmap
3515 *
3516 * => optional function.
3517 * => called when a process is swapped out to free memory.
3518 */
3519
3520 void
3521 pmap_collect(pmap)
3522 struct pmap *pmap;
3523 {
3524 }
3525
3526 /*
3527 * Routine: pmap_procwr
3528 *
3529 * Function:
3530 * Synchronize caches corresponding to [addr, addr+len) in p.
3531 *
3532 */
3533 void
3534 pmap_procwr(p, va, len)
3535 struct proc *p;
3536 vaddr_t va;
3537 int len;
3538 {
3539 /* We only need to do anything if it is the current process. */
3540 if (p == curproc)
3541 cpu_icache_sync_range(va, len);
3542 }
3543 /*
3544 * PTP functions
3545 */
3546
3547 /*
3548 * pmap_steal_ptp: Steal a PTP from somewhere else.
3549 *
3550 * This is just a placeholder, for now we never steal.
3551 */
3552
3553 static struct vm_page *
3554 pmap_steal_ptp(struct pmap *pmap, vaddr_t va)
3555 {
3556 return (NULL);
3557 }
3558
3559 /*
3560 * pmap_get_ptp: get a PTP (if there isn't one, allocate a new one)
3561 *
3562 * => pmap should NOT be pmap_kernel()
3563 * => pmap should be locked
3564 */
3565
3566 static struct vm_page *
3567 pmap_get_ptp(struct pmap *pmap, vaddr_t va, boolean_t just_try)
3568 {
3569 struct vm_page *ptp;
3570
3571 if (pmap_pde_page(pmap_pde(pmap, va))) {
3572
3573 /* valid... check hint (saves us a PA->PG lookup) */
3574 #if 0
3575 if (pmap->pm_ptphint &&
3576 ((unsigned)pmap_pde(pmap, va) & PG_FRAME) ==
3577 VM_PAGE_TO_PHYS(pmap->pm_ptphint))
3578 return (pmap->pm_ptphint);
3579 #endif
3580 ptp = uvm_pagelookup(&pmap->pm_obj, va);
3581 #ifdef DIAGNOSTIC
3582 if (ptp == NULL)
3583 panic("pmap_get_ptp: unmanaged user PTP");
3584 #endif
3585 // pmap->pm_ptphint = ptp;
3586 return(ptp);
3587 }
3588
3589 /* allocate a new PTP (updates ptphint) */
3590 return(pmap_alloc_ptp(pmap, va, just_try));
3591 }
3592
3593 /*
3594 * pmap_alloc_ptp: allocate a PTP for a PMAP
3595 *
3596 * => pmap should already be locked by caller
3597 * => we use the ptp's wire_count to count the number of active mappings
3598 * in the PTP (we start it at one to prevent any chance this PTP
3599 * will ever leak onto the active/inactive queues)
3600 */
3601
3602 /*__inline */ static struct vm_page *
3603 pmap_alloc_ptp(struct pmap *pmap, vaddr_t va, boolean_t just_try)
3604 {
3605 struct vm_page *ptp;
3606
3607 ptp = uvm_pagealloc(&pmap->pm_obj, va, NULL,
3608 UVM_PGA_USERESERVE|UVM_PGA_ZERO);
3609 if (ptp == NULL) {
3610 if (just_try)
3611 return (NULL);
3612
3613 ptp = pmap_steal_ptp(pmap, va);
3614
3615 if (ptp == NULL)
3616 return (NULL);
3617 /* Stole a page, zero it. */
3618 pmap_zero_page(VM_PAGE_TO_PHYS(ptp));
3619 }
3620
3621 /* got one! */
3622 ptp->flags &= ~PG_BUSY; /* never busy */
3623 ptp->wire_count = 1; /* no mappings yet */
3624 pmap_map_in_l1(pmap, va, VM_PAGE_TO_PHYS(ptp), TRUE);
3625 pmap->pm_stats.resident_count++; /* count PTP as resident */
3626 // pmap->pm_ptphint = ptp;
3627 return (ptp);
3628 }
3629
3630 vaddr_t
3631 pmap_growkernel(maxkvaddr)
3632 vaddr_t maxkvaddr;
3633 {
3634 struct pmap *kpm = pmap_kernel(), *pm;
3635 int s;
3636 paddr_t ptaddr;
3637 struct vm_page *ptp;
3638
3639 if (maxkvaddr <= pmap_curmaxkvaddr)
3640 goto out; /* we are OK */
3641 NPDEBUG(PDB_GROWKERN, printf("pmap_growkernel: growing kernel from %lx to %lx\n",
3642 pmap_curmaxkvaddr, maxkvaddr));
3643
3644 /*
3645 * whoops! we need to add kernel PTPs
3646 */
3647
3648 s = splhigh(); /* to be safe */
3649 simple_lock(&kpm->pm_obj.vmobjlock);
3650 /* due to the way the arm pmap works we map 4MB at a time */
3651 for (/*null*/ ; pmap_curmaxkvaddr < maxkvaddr ; pmap_curmaxkvaddr += 4 * NBPD) {
3652
3653 if (uvm.page_init_done == FALSE) {
3654
3655 /*
3656 * we're growing the kernel pmap early (from
3657 * uvm_pageboot_alloc()). this case must be
3658 * handled a little differently.
3659 */
3660
3661 if (uvm_page_physget(&ptaddr) == FALSE)
3662 panic("pmap_growkernel: out of memory");
3663 pmap_zero_page(ptaddr);
3664
3665 /* map this page in */
3666 pmap_map_in_l1(kpm, (pmap_curmaxkvaddr + 1), ptaddr, TRUE);
3667
3668 /* count PTP as resident */
3669 kpm->pm_stats.resident_count++;
3670 continue;
3671 }
3672
3673 /*
3674 * THIS *MUST* BE CODED SO AS TO WORK IN THE
3675 * pmap_initialized == FALSE CASE! WE MAY BE
3676 * INVOKED WHILE pmap_init() IS RUNNING!
3677 */
3678
3679 if ((ptp = pmap_alloc_ptp(kpm, (pmap_curmaxkvaddr + 1), FALSE)) == NULL) {
3680 panic("pmap_growkernel: alloc ptp failed");
3681 }
3682
3683 /* distribute new kernel PTP to all active pmaps */
3684 simple_lock(&pmaps_lock);
3685 LIST_FOREACH(pm, &pmaps, pm_list) {
3686 pmap_map_in_l1(pm, (pmap_curmaxkvaddr + 1), VM_PAGE_TO_PHYS(ptp), TRUE);
3687 }
3688
3689 simple_unlock(&pmaps_lock);
3690 }
3691
3692 /*
3693 * flush out the cache, expensive but growkernel will happen so
3694 * rarely
3695 */
3696 cpu_tlb_flushD();
3697 cpu_cpwait();
3698
3699 simple_unlock(&kpm->pm_obj.vmobjlock);
3700 splx(s);
3701
3702 out:
3703 return (pmap_curmaxkvaddr);
3704 }
3705
3706
3707
3708 /************************ Bootstrapping routines ****************************/
3709
3710 /*
3711 * This list exists for the benefit of pmap_map_chunk(). It keeps track
3712 * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
3713 * find them as necessary.
3714 *
3715 * Note that the data on this list is not valid after initarm() returns.
3716 */
3717 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
3718
3719 static vaddr_t
3720 kernel_pt_lookup(paddr_t pa)
3721 {
3722 pv_addr_t *pv;
3723
3724 SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
3725 if (pv->pv_pa == pa)
3726 return (pv->pv_va);
3727 }
3728 return (0);
3729 }
3730
3731 /*
3732 * pmap_map_section:
3733 *
3734 * Create a single section mapping.
3735 */
3736 void
3737 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
3738 {
3739 pd_entry_t *pde = (pd_entry_t *) l1pt;
3740 pd_entry_t ap = (prot & VM_PROT_WRITE) ? AP_KRW : AP_KR;
3741 pd_entry_t fl = (cache == PTE_CACHE) ? pte_cache_mode : 0;
3742
3743 KASSERT(((va | pa) & (L1_SEC_SIZE - 1)) == 0);
3744
3745 pde[va >> PDSHIFT] = L1_SECPTE(pa & PD_MASK, ap, fl);
3746 }
3747
3748 /*
3749 * pmap_map_entry:
3750 *
3751 * Create a single page mapping.
3752 */
3753 void
3754 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
3755 {
3756 pd_entry_t *pde = (pd_entry_t *) l1pt;
3757 pt_entry_t ap = (prot & VM_PROT_WRITE) ? AP_KRW : AP_KR;
3758 pt_entry_t fl = (cache == PTE_CACHE) ? pte_cache_mode : 0;
3759 pt_entry_t *pte;
3760
3761 KASSERT(((va | pa) & PGOFSET) == 0);
3762
3763 if ((pde[va >> PDSHIFT] & L1_MASK) != L1_PAGE)
3764 panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
3765
3766 pte = (pt_entry_t *)
3767 kernel_pt_lookup(pde[va >> PDSHIFT] & PG_FRAME);
3768 if (pte == NULL)
3769 panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
3770
3771 pte[(va >> PGSHIFT) & 0x3ff] = L2_SPTE(pa & PG_FRAME, ap, fl);
3772 }
3773
3774 /*
3775 * pmap_link_l2pt:
3776 *
3777 * Link the L2 page table specified by "pa" into the L1
3778 * page table at the slot for "va".
3779 */
3780 void
3781 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
3782 {
3783 pd_entry_t *pde = (pd_entry_t *) l1pt;
3784 u_int slot = va >> PDSHIFT;
3785
3786 KASSERT((l2pv->pv_pa & PGOFSET) == 0);
3787
3788 pde[slot + 0] = L1_PTE(l2pv->pv_pa + 0x000);
3789 pde[slot + 1] = L1_PTE(l2pv->pv_pa + 0x400);
3790 pde[slot + 2] = L1_PTE(l2pv->pv_pa + 0x800);
3791 pde[slot + 3] = L1_PTE(l2pv->pv_pa + 0xc00);
3792
3793 SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
3794 }
3795
3796 /*
3797 * pmap_map_chunk:
3798 *
3799 * Map a chunk of memory using the most efficient mappings
3800 * possible (section, large page, small page) into the
3801 * provided L1 and L2 tables at the specified virtual address.
3802 */
3803 vsize_t
3804 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
3805 int prot, int cache)
3806 {
3807 pd_entry_t *pde = (pd_entry_t *) l1pt;
3808 pt_entry_t ap = (prot & VM_PROT_WRITE) ? AP_KRW : AP_KR;
3809 pt_entry_t fl = (cache == PTE_CACHE) ? pte_cache_mode : 0;
3810 pt_entry_t *pte;
3811 vsize_t resid;
3812 int i;
3813
3814 resid = (size + (NBPG - 1)) & ~(NBPG - 1);
3815
3816 if (l1pt == 0)
3817 panic("pmap_map_chunk: no L1 table provided");
3818
3819 #ifdef VERBOSE_INIT_ARM
3820 printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
3821 "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
3822 #endif
3823
3824 size = resid;
3825
3826 while (resid > 0) {
3827 /* See if we can use a section mapping. */
3828 if (((pa | va) & (L1_SEC_SIZE - 1)) == 0 &&
3829 resid >= L1_SEC_SIZE) {
3830 #ifdef VERBOSE_INIT_ARM
3831 printf("S");
3832 #endif
3833 pde[va >> PDSHIFT] = L1_SECPTE(pa, ap, fl);
3834 va += L1_SEC_SIZE;
3835 pa += L1_SEC_SIZE;
3836 resid -= L1_SEC_SIZE;
3837 continue;
3838 }
3839
3840 /*
3841 * Ok, we're going to use an L2 table. Make sure
3842 * one is actually in the corresponding L1 slot
3843 * for the current VA.
3844 */
3845 if ((pde[va >> PDSHIFT] & L1_MASK) != L1_PAGE)
3846 panic("pmap_map_chunk: no L2 table for VA 0x%08lx", va);
3847
3848 pte = (pt_entry_t *)
3849 kernel_pt_lookup(pde[va >> PDSHIFT] & PG_FRAME);
3850 if (pte == NULL)
3851 panic("pmap_map_chunk: can't find L2 table for VA"
3852 "0x%08lx", va);
3853
3854 /* See if we can use a L2 large page mapping. */
3855 if (((pa | va) & (L2_LPAGE_SIZE - 1)) == 0 &&
3856 resid >= L2_LPAGE_SIZE) {
3857 #ifdef VERBOSE_INIT_ARM
3858 printf("L");
3859 #endif
3860 for (i = 0; i < 16; i++) {
3861 pte[((va >> PGSHIFT) & 0x3f0) + i] =
3862 L2_LPTE(pa, ap, fl);
3863 }
3864 va += L2_LPAGE_SIZE;
3865 pa += L2_LPAGE_SIZE;
3866 resid -= L2_LPAGE_SIZE;
3867 continue;
3868 }
3869
3870 /* Use a small page mapping. */
3871 #ifdef VERBOSE_INIT_ARM
3872 printf("P");
3873 #endif
3874 pte[(va >> PGSHIFT) & 0x3ff] = L2_SPTE(pa, ap, fl);
3875 va += NBPG;
3876 pa += NBPG;
3877 resid -= NBPG;
3878 }
3879 #ifdef VERBOSE_INIT_ARM
3880 printf("\n");
3881 #endif
3882 return (size);
3883 }
3884