pmap.c revision 1.28 1 /* $NetBSD: pmap.c,v 1.28 2001/10/18 18:15:56 rearnsha Exp $ */
2
3 /*
4 * Copyright (c) 2001 Richard Earnshaw
5 * Copyright (c) 2001 Christopher Gilbert
6 * All rights reserved.
7 *
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 3. The name of the company nor the name of the author may be used to
14 * endorse or promote products derived from this software without specific
15 * prior written permission.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
18 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
19 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20 * IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
21 * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
22 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
23 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 */
29
30 /*-
31 * Copyright (c) 1999 The NetBSD Foundation, Inc.
32 * All rights reserved.
33 *
34 * This code is derived from software contributed to The NetBSD Foundation
35 * by Charles M. Hannum.
36 *
37 * Redistribution and use in source and binary forms, with or without
38 * modification, are permitted provided that the following conditions
39 * are met:
40 * 1. Redistributions of source code must retain the above copyright
41 * notice, this list of conditions and the following disclaimer.
42 * 2. Redistributions in binary form must reproduce the above copyright
43 * notice, this list of conditions and the following disclaimer in the
44 * documentation and/or other materials provided with the distribution.
45 * 3. All advertising materials mentioning features or use of this software
46 * must display the following acknowledgement:
47 * This product includes software developed by the NetBSD
48 * Foundation, Inc. and its contributors.
49 * 4. Neither the name of The NetBSD Foundation nor the names of its
50 * contributors may be used to endorse or promote products derived
51 * from this software without specific prior written permission.
52 *
53 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
54 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
55 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
56 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
57 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
58 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
59 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
60 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
61 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
62 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
63 * POSSIBILITY OF SUCH DAMAGE.
64 */
65
66 /*
67 * Copyright (c) 1994-1998 Mark Brinicombe.
68 * Copyright (c) 1994 Brini.
69 * All rights reserved.
70 *
71 * This code is derived from software written for Brini by Mark Brinicombe
72 *
73 * Redistribution and use in source and binary forms, with or without
74 * modification, are permitted provided that the following conditions
75 * are met:
76 * 1. Redistributions of source code must retain the above copyright
77 * notice, this list of conditions and the following disclaimer.
78 * 2. Redistributions in binary form must reproduce the above copyright
79 * notice, this list of conditions and the following disclaimer in the
80 * documentation and/or other materials provided with the distribution.
81 * 3. All advertising materials mentioning features or use of this software
82 * must display the following acknowledgement:
83 * This product includes software developed by Mark Brinicombe.
84 * 4. The name of the author may not be used to endorse or promote products
85 * derived from this software without specific prior written permission.
86 *
87 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
88 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
89 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
90 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
91 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
92 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
93 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
94 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
95 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
96 *
97 * RiscBSD kernel project
98 *
99 * pmap.c
100 *
101 * Machine dependant vm stuff
102 *
103 * Created : 20/09/94
104 */
105
106 /*
107 * Performance improvements, UVM changes, overhauls and part-rewrites
108 * were contributed by Neil A. Carson <neil (at) causality.com>.
109 */
110
111 /*
112 * The dram block info is currently referenced from the bootconfig.
113 * This should be placed in a separate structure.
114 */
115
116 /*
117 * Special compilation symbols
118 * PMAP_DEBUG - Build in pmap_debug_level code
119 */
120
121 /* Include header files */
122
123 #include "opt_pmap_debug.h"
124 #include "opt_ddb.h"
125
126 #include <sys/types.h>
127 #include <sys/param.h>
128 #include <sys/kernel.h>
129 #include <sys/systm.h>
130 #include <sys/proc.h>
131 #include <sys/malloc.h>
132 #include <sys/user.h>
133 #include <sys/pool.h>
134 #include <sys/cdefs.h>
135
136 #include <uvm/uvm.h>
137
138 #include <machine/bootconfig.h>
139 #include <machine/bus.h>
140 #include <machine/pmap.h>
141 #include <machine/pcb.h>
142 #include <machine/param.h>
143 #include <machine/katelib.h>
144
145 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.28 2001/10/18 18:15:56 rearnsha Exp $");
146 #ifdef PMAP_DEBUG
147 #define PDEBUG(_lev_,_stat_) \
148 if (pmap_debug_level >= (_lev_)) \
149 ((_stat_))
150 int pmap_debug_level = -2;
151
152 /*
153 * for switching to potentially finer grained debugging
154 */
155 #define PDB_FOLLOW 0x0001
156 #define PDB_INIT 0x0002
157 #define PDB_ENTER 0x0004
158 #define PDB_REMOVE 0x0008
159 #define PDB_CREATE 0x0010
160 #define PDB_PTPAGE 0x0020
161 #define PDB_ASN 0x0040
162 #define PDB_BITS 0x0080
163 #define PDB_COLLECT 0x0100
164 #define PDB_PROTECT 0x0200
165 #define PDB_BOOTSTRAP 0x1000
166 #define PDB_PARANOIA 0x2000
167 #define PDB_WIRING 0x4000
168 #define PDB_PVDUMP 0x8000
169
170 int debugmap = 0;
171 int pmapdebug = PDB_PARANOIA | PDB_FOLLOW;
172 #define NPDEBUG(_lev_,_stat_) \
173 if (pmapdebug & (_lev_)) \
174 ((_stat_))
175
176 #else /* PMAP_DEBUG */
177 #define PDEBUG(_lev_,_stat_) /* Nothing */
178 #define PDEBUG(_lev_,_stat_) /* Nothing */
179 #endif /* PMAP_DEBUG */
180
181 struct pmap kernel_pmap_store;
182
183 /*
184 * pool that pmap structures are allocated from
185 */
186
187 struct pool pmap_pmap_pool;
188
189 pagehook_t page_hook0;
190 pagehook_t page_hook1;
191 char *memhook;
192 pt_entry_t msgbufpte;
193 extern caddr_t msgbufaddr;
194
195 boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
196 /*
197 * locking data structures
198 */
199
200 static struct lock pmap_main_lock;
201 static struct simplelock pvalloc_lock;
202 #ifdef LOCKDEBUG
203 #define PMAP_MAP_TO_HEAD_LOCK() \
204 (void) spinlockmgr(&pmap_main_lock, LK_SHARED, NULL)
205 #define PMAP_MAP_TO_HEAD_UNLOCK() \
206 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
207
208 #define PMAP_HEAD_TO_MAP_LOCK() \
209 (void) spinlockmgr(&pmap_main_lock, LK_EXCLUSIVE, NULL)
210 #define PMAP_HEAD_TO_MAP_UNLOCK() \
211 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
212 #else
213 #define PMAP_MAP_TO_HEAD_LOCK() /* nothing */
214 #define PMAP_MAP_TO_HEAD_UNLOCK() /* nothing */
215 #define PMAP_HEAD_TO_MAP_LOCK() /* nothing */
216 #define PMAP_HEAD_TO_MAP_UNLOCK() /* nothing */
217 #endif /* LOCKDEBUG */
218
219 /*
220 * pv_page management structures: locked by pvalloc_lock
221 */
222
223 TAILQ_HEAD(pv_pagelist, pv_page);
224 static struct pv_pagelist pv_freepages; /* list of pv_pages with free entrys */
225 static struct pv_pagelist pv_unusedpgs; /* list of unused pv_pages */
226 static int pv_nfpvents; /* # of free pv entries */
227 static struct pv_page *pv_initpage; /* bootstrap page from kernel_map */
228 static vaddr_t pv_cachedva; /* cached VA for later use */
229
230 #define PVE_LOWAT (PVE_PER_PVPAGE / 2) /* free pv_entry low water mark */
231 #define PVE_HIWAT (PVE_LOWAT + (PVE_PER_PVPAGE * 2))
232 /* high water mark */
233
234 /*
235 * local prototypes
236 */
237
238 static struct pv_entry *pmap_add_pvpage __P((struct pv_page *, boolean_t));
239 static struct pv_entry *pmap_alloc_pv __P((struct pmap *, int)); /* see codes below */
240 #define ALLOCPV_NEED 0 /* need PV now */
241 #define ALLOCPV_TRY 1 /* just try to allocate, don't steal */
242 #define ALLOCPV_NONEED 2 /* don't need PV, just growing cache */
243 static struct pv_entry *pmap_alloc_pvpage __P((struct pmap *, int));
244 static void pmap_enter_pv __P((struct pv_head *,
245 struct pv_entry *, struct pmap *,
246 vaddr_t, struct vm_page *, int));
247 static void pmap_free_pv __P((struct pmap *, struct pv_entry *));
248 static void pmap_free_pvs __P((struct pmap *, struct pv_entry *));
249 static void pmap_free_pv_doit __P((struct pv_entry *));
250 static void pmap_free_pvpage __P((void));
251 static boolean_t pmap_is_curpmap __P((struct pmap *));
252 static struct pv_entry *pmap_remove_pv __P((struct pv_head *, struct pmap *,
253 vaddr_t));
254 #define PMAP_REMOVE_ALL 0 /* remove all mappings */
255 #define PMAP_REMOVE_SKIPWIRED 1 /* skip wired mappings */
256
257 vsize_t npages;
258
259 static struct vm_page *pmap_alloc_ptp __P((struct pmap *, vaddr_t, boolean_t));
260 static struct vm_page *pmap_get_ptp __P((struct pmap *, vaddr_t, boolean_t));
261 __inline static void pmap_clearbit __P((paddr_t, unsigned int));
262 __inline static boolean_t pmap_testbit __P((paddr_t, unsigned int));
263
264 extern paddr_t physical_start;
265 extern paddr_t physical_freestart;
266 extern paddr_t physical_end;
267 extern paddr_t physical_freeend;
268 extern unsigned int free_pages;
269 extern int max_processes;
270
271 vaddr_t virtual_start;
272 vaddr_t virtual_end;
273
274 vaddr_t avail_start;
275 vaddr_t avail_end;
276
277 extern pv_addr_t systempage;
278
279 #define ALLOC_PAGE_HOOK(x, s) \
280 x.va = virtual_start; \
281 x.pte = (pt_entry_t *)pmap_pte(pmap_kernel(), virtual_start); \
282 virtual_start += s;
283
284 /* Variables used by the L1 page table queue code */
285 SIMPLEQ_HEAD(l1pt_queue, l1pt);
286 struct l1pt_queue l1pt_static_queue; /* head of our static l1 queue */
287 int l1pt_static_queue_count; /* items in the static l1 queue */
288 int l1pt_static_create_count; /* static l1 items created */
289 struct l1pt_queue l1pt_queue; /* head of our l1 queue */
290 int l1pt_queue_count; /* items in the l1 queue */
291 int l1pt_create_count; /* stat - L1's create count */
292 int l1pt_reuse_count; /* stat - L1's reused count */
293
294 /* Local function prototypes (not used outside this file) */
295 pt_entry_t *pmap_pte __P((struct pmap *pmap, vaddr_t va));
296 void map_pagetable __P((vaddr_t pagetable, vaddr_t va,
297 paddr_t pa, unsigned int flags));
298 void pmap_copy_on_write __P((paddr_t pa));
299 void pmap_pinit __P((struct pmap *));
300 void pmap_freepagedir __P((struct pmap *));
301
302 /* Other function prototypes */
303 extern void bzero_page __P((vaddr_t));
304 extern void bcopy_page __P((vaddr_t, vaddr_t));
305
306 struct l1pt *pmap_alloc_l1pt __P((void));
307 static __inline void pmap_map_in_l1 __P((struct pmap *pmap, vaddr_t va,
308 vaddr_t l2pa, boolean_t));
309
310 static pt_entry_t *pmap_map_ptes __P((struct pmap *));
311 static void pmap_unmap_ptes __P((struct pmap *));
312
313 __inline static void pmap_vac_me_harder __P((struct pmap *, struct pv_head *,
314 pt_entry_t *, boolean_t));
315 static void pmap_vac_me_kpmap __P((struct pmap *, struct pv_head *,
316 pt_entry_t *, boolean_t));
317 static void pmap_vac_me_user __P((struct pmap *, struct pv_head *,
318 pt_entry_t *, boolean_t));
319
320 /*
321 * Cache enable bits in PTE to use on pages that are cacheable.
322 * On most machines this is cacheable/bufferable, but on some, eg arm10, we
323 * can chose between write-through and write-back cacheing.
324 */
325 pt_entry_t pte_cache_mode = (PT_C | PT_B);
326
327 /*
328 * real definition of pv_entry.
329 */
330
331 struct pv_entry {
332 struct pv_entry *pv_next; /* next pv_entry */
333 struct pmap *pv_pmap; /* pmap where mapping lies */
334 vaddr_t pv_va; /* virtual address for mapping */
335 int pv_flags; /* flags */
336 struct vm_page *pv_ptp; /* vm_page for the ptp */
337 };
338
339 /*
340 * pv_entrys are dynamically allocated in chunks from a single page.
341 * we keep track of how many pv_entrys are in use for each page and
342 * we can free pv_entry pages if needed. there is one lock for the
343 * entire allocation system.
344 */
345
346 struct pv_page_info {
347 TAILQ_ENTRY(pv_page) pvpi_list;
348 struct pv_entry *pvpi_pvfree;
349 int pvpi_nfree;
350 };
351
352 /*
353 * number of pv_entry's in a pv_page
354 * (note: won't work on systems where NPBG isn't a constant)
355 */
356
357 #define PVE_PER_PVPAGE ((NBPG - sizeof(struct pv_page_info)) / \
358 sizeof(struct pv_entry))
359
360 /*
361 * a pv_page: where pv_entrys are allocated from
362 */
363
364 struct pv_page {
365 struct pv_page_info pvinfo;
366 struct pv_entry pvents[PVE_PER_PVPAGE];
367 };
368
369 #ifdef MYCROFT_HACK
370 int mycroft_hack = 0;
371 #endif
372
373 /* Function to set the debug level of the pmap code */
374
375 #ifdef PMAP_DEBUG
376 void
377 pmap_debug(level)
378 int level;
379 {
380 pmap_debug_level = level;
381 printf("pmap_debug: level=%d\n", pmap_debug_level);
382 }
383 #endif /* PMAP_DEBUG */
384
385 __inline static boolean_t
386 pmap_is_curpmap(struct pmap *pmap)
387 {
388 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap)
389 || (pmap == pmap_kernel()))
390 return (TRUE);
391 return (FALSE);
392 }
393 #include "isadma.h"
394
395 #if NISADMA > 0
396 /*
397 * Used to protect memory for ISA DMA bounce buffers. If, when loading
398 * pages into the system, memory intersects with any of these ranges,
399 * the intersecting memory will be loaded into a lower-priority free list.
400 */
401 bus_dma_segment_t *pmap_isa_dma_ranges;
402 int pmap_isa_dma_nranges;
403
404 boolean_t pmap_isa_dma_range_intersect __P((paddr_t, psize_t,
405 paddr_t *, psize_t *));
406
407 /*
408 * Check if a memory range intersects with an ISA DMA range, and
409 * return the page-rounded intersection if it does. The intersection
410 * will be placed on a lower-priority free list.
411 */
412 boolean_t
413 pmap_isa_dma_range_intersect(pa, size, pap, sizep)
414 paddr_t pa;
415 psize_t size;
416 paddr_t *pap;
417 psize_t *sizep;
418 {
419 bus_dma_segment_t *ds;
420 int i;
421
422 if (pmap_isa_dma_ranges == NULL)
423 return (FALSE);
424
425 for (i = 0, ds = pmap_isa_dma_ranges;
426 i < pmap_isa_dma_nranges; i++, ds++) {
427 if (ds->ds_addr <= pa && pa < (ds->ds_addr + ds->ds_len)) {
428 /*
429 * Beginning of region intersects with this range.
430 */
431 *pap = trunc_page(pa);
432 *sizep = round_page(min(pa + size,
433 ds->ds_addr + ds->ds_len) - pa);
434 return (TRUE);
435 }
436 if (pa < ds->ds_addr && ds->ds_addr < (pa + size)) {
437 /*
438 * End of region intersects with this range.
439 */
440 *pap = trunc_page(ds->ds_addr);
441 *sizep = round_page(min((pa + size) - ds->ds_addr,
442 ds->ds_len));
443 return (TRUE);
444 }
445 }
446
447 /*
448 * No intersection found.
449 */
450 return (FALSE);
451 }
452 #endif /* NISADMA > 0 */
453
454 /*
455 * p v _ e n t r y f u n c t i o n s
456 */
457
458 /*
459 * pv_entry allocation functions:
460 * the main pv_entry allocation functions are:
461 * pmap_alloc_pv: allocate a pv_entry structure
462 * pmap_free_pv: free one pv_entry
463 * pmap_free_pvs: free a list of pv_entrys
464 *
465 * the rest are helper functions
466 */
467
468 /*
469 * pmap_alloc_pv: inline function to allocate a pv_entry structure
470 * => we lock pvalloc_lock
471 * => if we fail, we call out to pmap_alloc_pvpage
472 * => 3 modes:
473 * ALLOCPV_NEED = we really need a pv_entry, even if we have to steal it
474 * ALLOCPV_TRY = we want a pv_entry, but not enough to steal
475 * ALLOCPV_NONEED = we are trying to grow our free list, don't really need
476 * one now
477 *
478 * "try" is for optional functions like pmap_copy().
479 */
480
481 __inline static struct pv_entry *
482 pmap_alloc_pv(pmap, mode)
483 struct pmap *pmap;
484 int mode;
485 {
486 struct pv_page *pvpage;
487 struct pv_entry *pv;
488
489 simple_lock(&pvalloc_lock);
490
491 if (pv_freepages.tqh_first != NULL) {
492 pvpage = pv_freepages.tqh_first;
493 pvpage->pvinfo.pvpi_nfree--;
494 if (pvpage->pvinfo.pvpi_nfree == 0) {
495 /* nothing left in this one? */
496 TAILQ_REMOVE(&pv_freepages, pvpage, pvinfo.pvpi_list);
497 }
498 pv = pvpage->pvinfo.pvpi_pvfree;
499 #ifdef DIAGNOSTIC
500 if (pv == NULL)
501 panic("pmap_alloc_pv: pvpi_nfree off");
502 #endif
503 pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
504 pv_nfpvents--; /* took one from pool */
505 } else {
506 pv = NULL; /* need more of them */
507 }
508
509 /*
510 * if below low water mark or we didn't get a pv_entry we try and
511 * create more pv_entrys ...
512 */
513
514 if (pv_nfpvents < PVE_LOWAT || pv == NULL) {
515 if (pv == NULL)
516 pv = pmap_alloc_pvpage(pmap, (mode == ALLOCPV_TRY) ?
517 mode : ALLOCPV_NEED);
518 else
519 (void) pmap_alloc_pvpage(pmap, ALLOCPV_NONEED);
520 }
521
522 simple_unlock(&pvalloc_lock);
523 return(pv);
524 }
525
526 /*
527 * pmap_alloc_pvpage: maybe allocate a new pvpage
528 *
529 * if need_entry is false: try and allocate a new pv_page
530 * if need_entry is true: try and allocate a new pv_page and return a
531 * new pv_entry from it. if we are unable to allocate a pv_page
532 * we make a last ditch effort to steal a pv_page from some other
533 * mapping. if that fails, we panic...
534 *
535 * => we assume that the caller holds pvalloc_lock
536 */
537
538 static struct pv_entry *
539 pmap_alloc_pvpage(pmap, mode)
540 struct pmap *pmap;
541 int mode;
542 {
543 struct vm_page *pg;
544 struct pv_page *pvpage;
545 struct pv_entry *pv;
546 int s;
547
548 /*
549 * if we need_entry and we've got unused pv_pages, allocate from there
550 */
551
552 if (mode != ALLOCPV_NONEED && pv_unusedpgs.tqh_first != NULL) {
553
554 /* move it to pv_freepages list */
555 pvpage = pv_unusedpgs.tqh_first;
556 TAILQ_REMOVE(&pv_unusedpgs, pvpage, pvinfo.pvpi_list);
557 TAILQ_INSERT_HEAD(&pv_freepages, pvpage, pvinfo.pvpi_list);
558
559 /* allocate a pv_entry */
560 pvpage->pvinfo.pvpi_nfree--; /* can't go to zero */
561 pv = pvpage->pvinfo.pvpi_pvfree;
562 #ifdef DIAGNOSTIC
563 if (pv == NULL)
564 panic("pmap_alloc_pvpage: pvpi_nfree off");
565 #endif
566 pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
567
568 pv_nfpvents--; /* took one from pool */
569 return(pv);
570 }
571
572 /*
573 * see if we've got a cached unmapped VA that we can map a page in.
574 * if not, try to allocate one.
575 */
576
577
578 if (pv_cachedva == 0) {
579 s = splvm();
580 pv_cachedva = uvm_km_kmemalloc(kmem_map, NULL,
581 PAGE_SIZE, UVM_KMF_TRYLOCK|UVM_KMF_VALLOC);
582 splx(s);
583 if (pv_cachedva == 0) {
584 return (NULL);
585 }
586 }
587
588 pg = uvm_pagealloc(NULL, pv_cachedva - vm_map_min(kernel_map), NULL,
589 UVM_PGA_USERESERVE);
590 if (pg)
591 pg->flags &= ~PG_BUSY; /* never busy */
592
593 if (pg == NULL)
594 return (NULL);
595
596 /*
597 * add a mapping for our new pv_page and free its entrys (save one!)
598 *
599 * NOTE: If we are allocating a PV page for the kernel pmap, the
600 * pmap is already locked! (...but entering the mapping is safe...)
601 */
602
603 pmap_kenter_pa(pv_cachedva, VM_PAGE_TO_PHYS(pg), VM_PROT_ALL);
604 pmap_update(pmap_kernel());
605 pvpage = (struct pv_page *) pv_cachedva;
606 pv_cachedva = 0;
607 return (pmap_add_pvpage(pvpage, mode != ALLOCPV_NONEED));
608 }
609
610 /*
611 * pmap_add_pvpage: add a pv_page's pv_entrys to the free list
612 *
613 * => caller must hold pvalloc_lock
614 * => if need_entry is true, we allocate and return one pv_entry
615 */
616
617 static struct pv_entry *
618 pmap_add_pvpage(pvp, need_entry)
619 struct pv_page *pvp;
620 boolean_t need_entry;
621 {
622 int tofree, lcv;
623
624 /* do we need to return one? */
625 tofree = (need_entry) ? PVE_PER_PVPAGE - 1 : PVE_PER_PVPAGE;
626
627 pvp->pvinfo.pvpi_pvfree = NULL;
628 pvp->pvinfo.pvpi_nfree = tofree;
629 for (lcv = 0 ; lcv < tofree ; lcv++) {
630 pvp->pvents[lcv].pv_next = pvp->pvinfo.pvpi_pvfree;
631 pvp->pvinfo.pvpi_pvfree = &pvp->pvents[lcv];
632 }
633 if (need_entry)
634 TAILQ_INSERT_TAIL(&pv_freepages, pvp, pvinfo.pvpi_list);
635 else
636 TAILQ_INSERT_TAIL(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
637 pv_nfpvents += tofree;
638 return((need_entry) ? &pvp->pvents[lcv] : NULL);
639 }
640
641 /*
642 * pmap_free_pv_doit: actually free a pv_entry
643 *
644 * => do not call this directly! instead use either
645 * 1. pmap_free_pv ==> free a single pv_entry
646 * 2. pmap_free_pvs => free a list of pv_entrys
647 * => we must be holding pvalloc_lock
648 */
649
650 __inline static void
651 pmap_free_pv_doit(pv)
652 struct pv_entry *pv;
653 {
654 struct pv_page *pvp;
655
656 pvp = (struct pv_page *) arm_trunc_page((vaddr_t)pv);
657 pv_nfpvents++;
658 pvp->pvinfo.pvpi_nfree++;
659
660 /* nfree == 1 => fully allocated page just became partly allocated */
661 if (pvp->pvinfo.pvpi_nfree == 1) {
662 TAILQ_INSERT_HEAD(&pv_freepages, pvp, pvinfo.pvpi_list);
663 }
664
665 /* free it */
666 pv->pv_next = pvp->pvinfo.pvpi_pvfree;
667 pvp->pvinfo.pvpi_pvfree = pv;
668
669 /*
670 * are all pv_page's pv_entry's free? move it to unused queue.
671 */
672
673 if (pvp->pvinfo.pvpi_nfree == PVE_PER_PVPAGE) {
674 TAILQ_REMOVE(&pv_freepages, pvp, pvinfo.pvpi_list);
675 TAILQ_INSERT_HEAD(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
676 }
677 }
678
679 /*
680 * pmap_free_pv: free a single pv_entry
681 *
682 * => we gain the pvalloc_lock
683 */
684
685 __inline static void
686 pmap_free_pv(pmap, pv)
687 struct pmap *pmap;
688 struct pv_entry *pv;
689 {
690 simple_lock(&pvalloc_lock);
691 pmap_free_pv_doit(pv);
692
693 /*
694 * Can't free the PV page if the PV entries were associated with
695 * the kernel pmap; the pmap is already locked.
696 */
697 if (pv_nfpvents > PVE_HIWAT && pv_unusedpgs.tqh_first != NULL &&
698 pmap != pmap_kernel())
699 pmap_free_pvpage();
700
701 simple_unlock(&pvalloc_lock);
702 }
703
704 /*
705 * pmap_free_pvs: free a list of pv_entrys
706 *
707 * => we gain the pvalloc_lock
708 */
709
710 __inline static void
711 pmap_free_pvs(pmap, pvs)
712 struct pmap *pmap;
713 struct pv_entry *pvs;
714 {
715 struct pv_entry *nextpv;
716
717 simple_lock(&pvalloc_lock);
718
719 for ( /* null */ ; pvs != NULL ; pvs = nextpv) {
720 nextpv = pvs->pv_next;
721 pmap_free_pv_doit(pvs);
722 }
723
724 /*
725 * Can't free the PV page if the PV entries were associated with
726 * the kernel pmap; the pmap is already locked.
727 */
728 if (pv_nfpvents > PVE_HIWAT && pv_unusedpgs.tqh_first != NULL &&
729 pmap != pmap_kernel())
730 pmap_free_pvpage();
731
732 simple_unlock(&pvalloc_lock);
733 }
734
735
736 /*
737 * pmap_free_pvpage: try and free an unused pv_page structure
738 *
739 * => assume caller is holding the pvalloc_lock and that
740 * there is a page on the pv_unusedpgs list
741 * => if we can't get a lock on the kmem_map we try again later
742 * => note: analysis of MI kmem_map usage [i.e. malloc/free] shows
743 * that if we can lock the kmem_map then we are not already
744 * holding kmem_object's lock.
745 */
746
747 static void
748 pmap_free_pvpage()
749 {
750 int s;
751 struct vm_map *map;
752 struct vm_map_entry *dead_entries;
753 struct pv_page *pvp;
754
755 s = splvm(); /* protect kmem_map */
756
757 pvp = pv_unusedpgs.tqh_first;
758
759 /*
760 * note: watch out for pv_initpage which is allocated out of
761 * kernel_map rather than kmem_map.
762 */
763 if (pvp == pv_initpage)
764 map = kernel_map;
765 else
766 map = kmem_map;
767
768 if (vm_map_lock_try(map)) {
769
770 /* remove pvp from pv_unusedpgs */
771 TAILQ_REMOVE(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
772
773 /* unmap the page */
774 dead_entries = NULL;
775 uvm_unmap_remove(map, (vaddr_t)pvp, ((vaddr_t)pvp) + PAGE_SIZE,
776 &dead_entries);
777 vm_map_unlock(map);
778
779 if (dead_entries != NULL)
780 uvm_unmap_detach(dead_entries, 0);
781
782 pv_nfpvents -= PVE_PER_PVPAGE; /* update free count */
783 }
784
785 if (pvp == pv_initpage)
786 /* no more initpage, we've freed it */
787 pv_initpage = NULL;
788
789 splx(s);
790 }
791
792 /*
793 * main pv_entry manipulation functions:
794 * pmap_enter_pv: enter a mapping onto a pv_head list
795 * pmap_remove_pv: remove a mappiing from a pv_head list
796 *
797 * NOTE: pmap_enter_pv expects to lock the pvh itself
798 * pmap_remove_pv expects te caller to lock the pvh before calling
799 */
800
801 /*
802 * pmap_enter_pv: enter a mapping onto a pv_head lst
803 *
804 * => caller should hold the proper lock on pmap_main_lock
805 * => caller should have pmap locked
806 * => we will gain the lock on the pv_head and allocate the new pv_entry
807 * => caller should adjust ptp's wire_count before calling
808 * => caller should not adjust pmap's wire_count
809 */
810
811 __inline static void
812 pmap_enter_pv(pvh, pve, pmap, va, ptp, flags)
813 struct pv_head *pvh;
814 struct pv_entry *pve; /* preallocated pve for us to use */
815 struct pmap *pmap;
816 vaddr_t va;
817 struct vm_page *ptp; /* PTP in pmap that maps this VA */
818 int flags;
819 {
820 pve->pv_pmap = pmap;
821 pve->pv_va = va;
822 pve->pv_ptp = ptp; /* NULL for kernel pmap */
823 pve->pv_flags = flags;
824 simple_lock(&pvh->pvh_lock); /* lock pv_head */
825 pve->pv_next = pvh->pvh_list; /* add to ... */
826 pvh->pvh_list = pve; /* ... locked list */
827 simple_unlock(&pvh->pvh_lock); /* unlock, done! */
828 if (pve->pv_flags & PT_W)
829 ++pmap->pm_stats.wired_count;
830 }
831
832 /*
833 * pmap_remove_pv: try to remove a mapping from a pv_list
834 *
835 * => caller should hold proper lock on pmap_main_lock
836 * => pmap should be locked
837 * => caller should hold lock on pv_head [so that attrs can be adjusted]
838 * => caller should adjust ptp's wire_count and free PTP if needed
839 * => caller should NOT adjust pmap's wire_count
840 * => we return the removed pve
841 */
842
843 __inline static struct pv_entry *
844 pmap_remove_pv(pvh, pmap, va)
845 struct pv_head *pvh;
846 struct pmap *pmap;
847 vaddr_t va;
848 {
849 struct pv_entry *pve, **prevptr;
850
851 prevptr = &pvh->pvh_list; /* previous pv_entry pointer */
852 pve = *prevptr;
853 while (pve) {
854 if (pve->pv_pmap == pmap && pve->pv_va == va) { /* match? */
855 *prevptr = pve->pv_next; /* remove it! */
856 if (pve->pv_flags & PT_W)
857 --pmap->pm_stats.wired_count;
858 break;
859 }
860 prevptr = &pve->pv_next; /* previous pointer */
861 pve = pve->pv_next; /* advance */
862 }
863 return(pve); /* return removed pve */
864 }
865
866 /*
867 *
868 * pmap_modify_pv: Update pv flags
869 *
870 * => caller should hold lock on pv_head [so that attrs can be adjusted]
871 * => caller should NOT adjust pmap's wire_count
872 * => we return the old flags
873 *
874 * Modify a physical-virtual mapping in the pv table
875 */
876
877 /*__inline */ u_int
878 pmap_modify_pv(pmap, va, pvh, bic_mask, eor_mask)
879 struct pmap *pmap;
880 vaddr_t va;
881 struct pv_head *pvh;
882 u_int bic_mask;
883 u_int eor_mask;
884 {
885 struct pv_entry *npv;
886 u_int flags, oflags;
887
888 /*
889 * There is at least one VA mapping this page.
890 */
891
892 for (npv = pvh->pvh_list; npv; npv = npv->pv_next) {
893 if (pmap == npv->pv_pmap && va == npv->pv_va) {
894 oflags = npv->pv_flags;
895 npv->pv_flags = flags =
896 ((oflags & ~bic_mask) ^ eor_mask);
897 if ((flags ^ oflags) & PT_W) {
898 if (flags & PT_W)
899 ++pmap->pm_stats.wired_count;
900 else
901 --pmap->pm_stats.wired_count;
902 }
903 return (oflags);
904 }
905 }
906 return (0);
907 }
908
909 /*
910 * Map the specified level 2 pagetable into the level 1 page table for
911 * the given pmap to cover a chunk of virtual address space starting from the
912 * address specified.
913 */
914 static /*__inline*/ void
915 pmap_map_in_l1(pmap, va, l2pa, selfref)
916 struct pmap *pmap;
917 vaddr_t va, l2pa;
918 boolean_t selfref;
919 {
920 vaddr_t ptva;
921
922 /* Calculate the index into the L1 page table. */
923 ptva = (va >> PDSHIFT) & ~3;
924
925 PDEBUG(0, printf("wiring %08lx in to pd%p pte0x%lx va0x%lx\n", l2pa,
926 pmap->pm_pdir, L1_PTE(l2pa), ptva));
927
928 /* Map page table into the L1. */
929 pmap->pm_pdir[ptva + 0] = L1_PTE(l2pa + 0x000);
930 pmap->pm_pdir[ptva + 1] = L1_PTE(l2pa + 0x400);
931 pmap->pm_pdir[ptva + 2] = L1_PTE(l2pa + 0x800);
932 pmap->pm_pdir[ptva + 3] = L1_PTE(l2pa + 0xc00);
933
934 PDEBUG(0, printf("pt self reference %lx in %lx\n",
935 L2_PTE_NC_NB(l2pa, AP_KRW), pmap->pm_vptpt));
936
937 /* Map the page table into the page table area. */
938 if (selfref) {
939 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) =
940 L2_PTE_NC_NB(l2pa, AP_KRW);
941 }
942 /* XXX should be a purge */
943 /* cpu_tlb_flushD();*/
944 }
945
946 #if 0
947 static /*__inline*/ void
948 pmap_unmap_in_l1(pmap, va)
949 struct pmap *pmap;
950 vaddr_t va;
951 {
952 vaddr_t ptva;
953
954 /* Calculate the index into the L1 page table. */
955 ptva = (va >> PDSHIFT) & ~3;
956
957 /* Unmap page table from the L1. */
958 pmap->pm_pdir[ptva + 0] = 0;
959 pmap->pm_pdir[ptva + 1] = 0;
960 pmap->pm_pdir[ptva + 2] = 0;
961 pmap->pm_pdir[ptva + 3] = 0;
962
963 /* Unmap the page table from the page table area. */
964 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) = 0;
965
966 /* XXX should be a purge */
967 /* cpu_tlb_flushD();*/
968 }
969 #endif
970
971 /*
972 * Used to map a range of physical addresses into kernel
973 * virtual address space.
974 *
975 * For now, VM is already on, we only need to map the
976 * specified memory.
977 */
978 vaddr_t
979 pmap_map(va, spa, epa, prot)
980 vaddr_t va, spa, epa;
981 int prot;
982 {
983 while (spa < epa) {
984 pmap_kenter_pa(va, spa, prot);
985 va += NBPG;
986 spa += NBPG;
987 }
988 pmap_update(pmap_kernel());
989 return(va);
990 }
991
992
993 /*
994 * void pmap_bootstrap(pd_entry_t *kernel_l1pt, pv_addr_t kernel_ptpt)
995 *
996 * bootstrap the pmap system. This is called from initarm and allows
997 * the pmap system to initailise any structures it requires.
998 *
999 * Currently this sets up the kernel_pmap that is statically allocated
1000 * and also allocated virtual addresses for certain page hooks.
1001 * Currently the only one page hook is allocated that is used
1002 * to zero physical pages of memory.
1003 * It also initialises the start and end address of the kernel data space.
1004 */
1005 extern paddr_t physical_freestart;
1006 extern paddr_t physical_freeend;
1007
1008 char *boot_head;
1009
1010 void
1011 pmap_bootstrap(kernel_l1pt, kernel_ptpt)
1012 pd_entry_t *kernel_l1pt;
1013 pv_addr_t kernel_ptpt;
1014 {
1015 int loop;
1016 paddr_t start, end;
1017 #if NISADMA > 0
1018 paddr_t istart;
1019 psize_t isize;
1020 #endif
1021
1022 pmap_kernel()->pm_pdir = kernel_l1pt;
1023 pmap_kernel()->pm_pptpt = kernel_ptpt.pv_pa;
1024 pmap_kernel()->pm_vptpt = kernel_ptpt.pv_va;
1025 simple_lock_init(&pmap_kernel()->pm_lock);
1026 pmap_kernel()->pm_obj.pgops = NULL;
1027 TAILQ_INIT(&(pmap_kernel()->pm_obj.memq));
1028 pmap_kernel()->pm_obj.uo_npages = 0;
1029 pmap_kernel()->pm_obj.uo_refs = 1;
1030
1031 /*
1032 * Initialize PAGE_SIZE-dependent variables.
1033 */
1034 uvm_setpagesize();
1035
1036 npages = 0;
1037 loop = 0;
1038 while (loop < bootconfig.dramblocks) {
1039 start = (paddr_t)bootconfig.dram[loop].address;
1040 end = start + (bootconfig.dram[loop].pages * NBPG);
1041 if (start < physical_freestart)
1042 start = physical_freestart;
1043 if (end > physical_freeend)
1044 end = physical_freeend;
1045 #if 0
1046 printf("%d: %lx -> %lx\n", loop, start, end - 1);
1047 #endif
1048 #if NISADMA > 0
1049 if (pmap_isa_dma_range_intersect(start, end - start,
1050 &istart, &isize)) {
1051 /*
1052 * Place the pages that intersect with the
1053 * ISA DMA range onto the ISA DMA free list.
1054 */
1055 #if 0
1056 printf(" ISADMA 0x%lx -> 0x%lx\n", istart,
1057 istart + isize - 1);
1058 #endif
1059 uvm_page_physload(atop(istart),
1060 atop(istart + isize), atop(istart),
1061 atop(istart + isize), VM_FREELIST_ISADMA);
1062 npages += atop(istart + isize) - atop(istart);
1063
1064 /*
1065 * Load the pieces that come before
1066 * the intersection into the default
1067 * free list.
1068 */
1069 if (start < istart) {
1070 #if 0
1071 printf(" BEFORE 0x%lx -> 0x%lx\n",
1072 start, istart - 1);
1073 #endif
1074 uvm_page_physload(atop(start),
1075 atop(istart), atop(start),
1076 atop(istart), VM_FREELIST_DEFAULT);
1077 npages += atop(istart) - atop(start);
1078 }
1079
1080 /*
1081 * Load the pieces that come after
1082 * the intersection into the default
1083 * free list.
1084 */
1085 if ((istart + isize) < end) {
1086 #if 0
1087 printf(" AFTER 0x%lx -> 0x%lx\n",
1088 (istart + isize), end - 1);
1089 #endif
1090 uvm_page_physload(atop(istart + isize),
1091 atop(end), atop(istart + isize),
1092 atop(end), VM_FREELIST_DEFAULT);
1093 npages += atop(end) - atop(istart + isize);
1094 }
1095 } else {
1096 uvm_page_physload(atop(start), atop(end),
1097 atop(start), atop(end), VM_FREELIST_DEFAULT);
1098 npages += atop(end) - atop(start);
1099 }
1100 #else /* NISADMA > 0 */
1101 uvm_page_physload(atop(start), atop(end),
1102 atop(start), atop(end), VM_FREELIST_DEFAULT);
1103 npages += atop(end) - atop(start);
1104 #endif /* NISADMA > 0 */
1105 ++loop;
1106 }
1107
1108 #ifdef MYCROFT_HACK
1109 printf("npages = %ld\n", npages);
1110 #endif
1111
1112 virtual_start = KERNEL_VM_BASE;
1113 virtual_end = virtual_start + KERNEL_VM_SIZE - 1;
1114
1115 ALLOC_PAGE_HOOK(page_hook0, NBPG);
1116 ALLOC_PAGE_HOOK(page_hook1, NBPG);
1117
1118 /*
1119 * The mem special device needs a virtual hook but we don't
1120 * need a pte
1121 */
1122 memhook = (char *)virtual_start;
1123 virtual_start += NBPG;
1124
1125 msgbufaddr = (caddr_t)virtual_start;
1126 msgbufpte = (pt_entry_t)pmap_pte(pmap_kernel(), virtual_start);
1127 virtual_start += round_page(MSGBUFSIZE);
1128
1129 /*
1130 * init the static-global locks and global lists.
1131 */
1132 spinlockinit(&pmap_main_lock, "pmaplk", 0);
1133 simple_lock_init(&pvalloc_lock);
1134 TAILQ_INIT(&pv_freepages);
1135 TAILQ_INIT(&pv_unusedpgs);
1136
1137 /*
1138 * compute the number of pages we have and then allocate RAM
1139 * for each pages' pv_head and saved attributes.
1140 */
1141 {
1142 int npages, lcv;
1143 vsize_t s;
1144
1145 npages = 0;
1146 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++)
1147 npages += (vm_physmem[lcv].end - vm_physmem[lcv].start);
1148 s = (vsize_t) (sizeof(struct pv_head) * npages +
1149 sizeof(char) * npages);
1150 s = round_page(s); /* round up */
1151 boot_head = (char *)uvm_pageboot_alloc(s);
1152 bzero((char *)boot_head, s);
1153 if (boot_head == 0)
1154 panic("pmap_init: unable to allocate pv_heads");
1155 }
1156
1157 /*
1158 * initialize the pmap pool.
1159 */
1160
1161 pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl",
1162 0, pool_page_alloc_nointr, pool_page_free_nointr, M_VMPMAP);
1163
1164 cpu_cache_cleanD();
1165 }
1166
1167 /*
1168 * void pmap_init(void)
1169 *
1170 * Initialize the pmap module.
1171 * Called by vm_init() in vm/vm_init.c in order to initialise
1172 * any structures that the pmap system needs to map virtual memory.
1173 */
1174
1175 extern int physmem;
1176
1177 void
1178 pmap_init()
1179 {
1180 int lcv, i;
1181
1182 #ifdef MYCROFT_HACK
1183 printf("physmem = %d\n", physmem);
1184 #endif
1185
1186 /*
1187 * Set the available memory vars - These do not map to real memory
1188 * addresses and cannot as the physical memory is fragmented.
1189 * They are used by ps for %mem calculations.
1190 * One could argue whether this should be the entire memory or just
1191 * the memory that is useable in a user process.
1192 */
1193 avail_start = 0;
1194 avail_end = physmem * NBPG;
1195
1196 /* allocate pv_head stuff first */
1197 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
1198 vm_physmem[lcv].pmseg.pvhead = (struct pv_head *)boot_head;
1199 boot_head = (char *)(vaddr_t)(vm_physmem[lcv].pmseg.pvhead +
1200 (vm_physmem[lcv].end - vm_physmem[lcv].start));
1201 for (i = 0;
1202 i < (vm_physmem[lcv].end - vm_physmem[lcv].start); i++) {
1203 simple_lock_init(
1204 &vm_physmem[lcv].pmseg.pvhead[i].pvh_lock);
1205 }
1206 }
1207
1208 /* now allocate attrs */
1209 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
1210 vm_physmem[lcv].pmseg.attrs = (char *) boot_head;
1211 boot_head = (char *)(vaddr_t)(vm_physmem[lcv].pmseg.attrs +
1212 (vm_physmem[lcv].end - vm_physmem[lcv].start));
1213 }
1214
1215 /*
1216 * now we need to free enough pv_entry structures to allow us to get
1217 * the kmem_map/kmem_object allocated and inited (done after this
1218 * function is finished). to do this we allocate one bootstrap page out
1219 * of kernel_map and use it to provide an initial pool of pv_entry
1220 * structures. we never free this page.
1221 */
1222
1223 pv_initpage = (struct pv_page *) uvm_km_alloc(kernel_map, PAGE_SIZE);
1224 if (pv_initpage == NULL)
1225 panic("pmap_init: pv_initpage");
1226 pv_cachedva = 0; /* a VA we have allocated but not used yet */
1227 pv_nfpvents = 0;
1228 (void) pmap_add_pvpage(pv_initpage, FALSE);
1229
1230 #ifdef MYCROFT_HACK
1231 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) {
1232 printf("physseg[%d] pvent=%p attrs=%p start=%ld end=%ld\n",
1233 lcv,
1234 vm_physmem[lcv].pmseg.pvent, vm_physmem[lcv].pmseg.attrs,
1235 vm_physmem[lcv].start, vm_physmem[lcv].end);
1236 }
1237 #endif
1238 pmap_initialized = TRUE;
1239
1240 /* Initialise our L1 page table queues and counters */
1241 SIMPLEQ_INIT(&l1pt_static_queue);
1242 l1pt_static_queue_count = 0;
1243 l1pt_static_create_count = 0;
1244 SIMPLEQ_INIT(&l1pt_queue);
1245 l1pt_queue_count = 0;
1246 l1pt_create_count = 0;
1247 l1pt_reuse_count = 0;
1248 }
1249
1250 /*
1251 * pmap_postinit()
1252 *
1253 * This routine is called after the vm and kmem subsystems have been
1254 * initialised. This allows the pmap code to perform any initialisation
1255 * that can only be done one the memory allocation is in place.
1256 */
1257
1258 void
1259 pmap_postinit()
1260 {
1261 int loop;
1262 struct l1pt *pt;
1263
1264 #ifdef PMAP_STATIC_L1S
1265 for (loop = 0; loop < PMAP_STATIC_L1S; ++loop) {
1266 #else /* PMAP_STATIC_L1S */
1267 for (loop = 0; loop < max_processes; ++loop) {
1268 #endif /* PMAP_STATIC_L1S */
1269 /* Allocate a L1 page table */
1270 pt = pmap_alloc_l1pt();
1271 if (!pt)
1272 panic("Cannot allocate static L1 page tables\n");
1273
1274 /* Clean it */
1275 bzero((void *)pt->pt_va, PD_SIZE);
1276 pt->pt_flags |= (PTFLAG_STATIC | PTFLAG_CLEAN);
1277 /* Add the page table to the queue */
1278 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pt, pt_queue);
1279 ++l1pt_static_queue_count;
1280 ++l1pt_static_create_count;
1281 }
1282 }
1283
1284
1285 /*
1286 * Create and return a physical map.
1287 *
1288 * If the size specified for the map is zero, the map is an actual physical
1289 * map, and may be referenced by the hardware.
1290 *
1291 * If the size specified is non-zero, the map will be used in software only,
1292 * and is bounded by that size.
1293 */
1294
1295 pmap_t
1296 pmap_create()
1297 {
1298 struct pmap *pmap;
1299
1300 /*
1301 * Fetch pmap entry from the pool
1302 */
1303
1304 pmap = pool_get(&pmap_pmap_pool, PR_WAITOK);
1305 /* XXX is this really needed! */
1306 memset(pmap, 0, sizeof(*pmap));
1307
1308 simple_lock_init(&pmap->pm_obj.vmobjlock);
1309 pmap->pm_obj.pgops = NULL; /* currently not a mappable object */
1310 TAILQ_INIT(&pmap->pm_obj.memq);
1311 pmap->pm_obj.uo_npages = 0;
1312 pmap->pm_obj.uo_refs = 1;
1313 pmap->pm_stats.wired_count = 0;
1314 pmap->pm_stats.resident_count = 1;
1315
1316 /* Now init the machine part of the pmap */
1317 pmap_pinit(pmap);
1318 return(pmap);
1319 }
1320
1321 /*
1322 * pmap_alloc_l1pt()
1323 *
1324 * This routine allocates physical and virtual memory for a L1 page table
1325 * and wires it.
1326 * A l1pt structure is returned to describe the allocated page table.
1327 *
1328 * This routine is allowed to fail if the required memory cannot be allocated.
1329 * In this case NULL is returned.
1330 */
1331
1332 struct l1pt *
1333 pmap_alloc_l1pt(void)
1334 {
1335 paddr_t pa;
1336 vaddr_t va;
1337 struct l1pt *pt;
1338 int error;
1339 struct vm_page *m;
1340 pt_entry_t *ptes;
1341
1342 /* Allocate virtual address space for the L1 page table */
1343 va = uvm_km_valloc(kernel_map, PD_SIZE);
1344 if (va == 0) {
1345 #ifdef DIAGNOSTIC
1346 PDEBUG(0,
1347 printf("pmap: Cannot allocate pageable memory for L1\n"));
1348 #endif /* DIAGNOSTIC */
1349 return(NULL);
1350 }
1351
1352 /* Allocate memory for the l1pt structure */
1353 pt = (struct l1pt *)malloc(sizeof(struct l1pt), M_VMPMAP, M_WAITOK);
1354
1355 /*
1356 * Allocate pages from the VM system.
1357 */
1358 TAILQ_INIT(&pt->pt_plist);
1359 error = uvm_pglistalloc(PD_SIZE, physical_start, physical_end,
1360 PD_SIZE, 0, &pt->pt_plist, 1, M_WAITOK);
1361 if (error) {
1362 #ifdef DIAGNOSTIC
1363 PDEBUG(0,
1364 printf("pmap: Cannot allocate physical mem for L1 (%d)\n",
1365 error));
1366 #endif /* DIAGNOSTIC */
1367 /* Release the resources we already have claimed */
1368 free(pt, M_VMPMAP);
1369 uvm_km_free(kernel_map, va, PD_SIZE);
1370 return(NULL);
1371 }
1372
1373 /* Map our physical pages into our virtual space */
1374 pt->pt_va = va;
1375 m = pt->pt_plist.tqh_first;
1376 ptes = pmap_map_ptes(pmap_kernel());
1377 while (m && va < (pt->pt_va + PD_SIZE)) {
1378 pa = VM_PAGE_TO_PHYS(m);
1379
1380 pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE);
1381
1382 /* Revoke cacheability and bufferability */
1383 /* XXX should be done better than this */
1384 ptes[arm_byte_to_page(va)] &= ~(PT_C | PT_B);
1385
1386 va += NBPG;
1387 m = m->pageq.tqe_next;
1388 }
1389 pmap_unmap_ptes(pmap_kernel());
1390 pmap_update(pmap_kernel());
1391
1392 #ifdef DIAGNOSTIC
1393 if (m)
1394 panic("pmap_alloc_l1pt: pglist not empty\n");
1395 #endif /* DIAGNOSTIC */
1396
1397 pt->pt_flags = 0;
1398 return(pt);
1399 }
1400
1401 /*
1402 * Free a L1 page table previously allocated with pmap_alloc_l1pt().
1403 */
1404 void
1405 pmap_free_l1pt(pt)
1406 struct l1pt *pt;
1407 {
1408 /* Separate the physical memory for the virtual space */
1409 pmap_kremove(pt->pt_va, PD_SIZE);
1410 pmap_update(pmap_kernel());
1411
1412 /* Return the physical memory */
1413 uvm_pglistfree(&pt->pt_plist);
1414
1415 /* Free the virtual space */
1416 uvm_km_free(kernel_map, pt->pt_va, PD_SIZE);
1417
1418 /* Free the l1pt structure */
1419 free(pt, M_VMPMAP);
1420 }
1421
1422 /*
1423 * Allocate a page directory.
1424 * This routine will either allocate a new page directory from the pool
1425 * of L1 page tables currently held by the kernel or it will allocate
1426 * a new one via pmap_alloc_l1pt().
1427 * It will then initialise the l1 page table for use.
1428 */
1429 int
1430 pmap_allocpagedir(pmap)
1431 struct pmap *pmap;
1432 {
1433 paddr_t pa;
1434 struct l1pt *pt;
1435 pt_entry_t *pte;
1436
1437 PDEBUG(0, printf("pmap_allocpagedir(%p)\n", pmap));
1438
1439 /* Do we have any spare L1's lying around ? */
1440 if (l1pt_static_queue_count) {
1441 --l1pt_static_queue_count;
1442 pt = l1pt_static_queue.sqh_first;
1443 SIMPLEQ_REMOVE_HEAD(&l1pt_static_queue, pt, pt_queue);
1444 } else if (l1pt_queue_count) {
1445 --l1pt_queue_count;
1446 pt = l1pt_queue.sqh_first;
1447 SIMPLEQ_REMOVE_HEAD(&l1pt_queue, pt, pt_queue);
1448 ++l1pt_reuse_count;
1449 } else {
1450 pt = pmap_alloc_l1pt();
1451 if (!pt)
1452 return(ENOMEM);
1453 ++l1pt_create_count;
1454 }
1455
1456 /* Store the pointer to the l1 descriptor in the pmap. */
1457 pmap->pm_l1pt = pt;
1458
1459 /* Get the physical address of the start of the l1 */
1460 pa = VM_PAGE_TO_PHYS(pt->pt_plist.tqh_first);
1461
1462 /* Store the virtual address of the l1 in the pmap. */
1463 pmap->pm_pdir = (pd_entry_t *)pt->pt_va;
1464
1465 /* Clean the L1 if it is dirty */
1466 if (!(pt->pt_flags & PTFLAG_CLEAN))
1467 bzero((void *)pmap->pm_pdir, (PD_SIZE - KERNEL_PD_SIZE));
1468
1469 /* Do we already have the kernel mappings ? */
1470 if (!(pt->pt_flags & PTFLAG_KPT)) {
1471 /* Duplicate the kernel mapping i.e. all mappings 0xf0000000+ */
1472
1473 bcopy((char *)pmap_kernel()->pm_pdir + (PD_SIZE - KERNEL_PD_SIZE),
1474 (char *)pmap->pm_pdir + (PD_SIZE - KERNEL_PD_SIZE),
1475 KERNEL_PD_SIZE);
1476 pt->pt_flags |= PTFLAG_KPT;
1477 }
1478
1479 /* Allocate a page table to map all the page tables for this pmap */
1480
1481 #ifdef DIAGNOSTIC
1482 if (pmap->pm_vptpt) {
1483 /* XXX What if we have one already ? */
1484 panic("pmap_allocpagedir: have pt already\n");
1485 }
1486 #endif /* DIAGNOSTIC */
1487 pmap->pm_vptpt = uvm_km_zalloc(kernel_map, NBPG);
1488 if (pmap->pm_vptpt == 0) {
1489 pmap_freepagedir(pmap);
1490 return(ENOMEM);
1491 }
1492
1493 (void) pmap_extract(pmap_kernel(), pmap->pm_vptpt, &pmap->pm_pptpt);
1494 pmap->pm_pptpt &= PG_FRAME;
1495 /* Revoke cacheability and bufferability */
1496 /* XXX should be done better than this */
1497 pte = pmap_pte(pmap_kernel(), pmap->pm_vptpt);
1498 *pte = *pte & ~(PT_C | PT_B);
1499
1500 /* Wire in this page table */
1501 pmap_map_in_l1(pmap, PROCESS_PAGE_TBLS_BASE, pmap->pm_pptpt, TRUE);
1502
1503 pt->pt_flags &= ~PTFLAG_CLEAN; /* L1 is dirty now */
1504
1505 /*
1506 * Map the kernel page tables for 0xf0000000 +
1507 * into the page table used to map the
1508 * pmap's page tables
1509 */
1510 bcopy((char *)(PROCESS_PAGE_TBLS_BASE
1511 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2))
1512 + ((PD_SIZE - KERNEL_PD_SIZE) >> 2)),
1513 (char *)pmap->pm_vptpt + ((PD_SIZE - KERNEL_PD_SIZE) >> 2),
1514 (KERNEL_PD_SIZE >> 2));
1515
1516 return(0);
1517 }
1518
1519
1520 /*
1521 * Initialize a preallocated and zeroed pmap structure,
1522 * such as one in a vmspace structure.
1523 */
1524
1525 void
1526 pmap_pinit(pmap)
1527 struct pmap *pmap;
1528 {
1529 int backoff = 6;
1530 int retry = 10;
1531
1532 PDEBUG(0, printf("pmap_pinit(%p)\n", pmap));
1533
1534 /* Keep looping until we succeed in allocating a page directory */
1535 while (pmap_allocpagedir(pmap) != 0) {
1536 /*
1537 * Ok we failed to allocate a suitable block of memory for an
1538 * L1 page table. This means that either:
1539 * 1. 16KB of virtual address space could not be allocated
1540 * 2. 16KB of physically contiguous memory on a 16KB boundary
1541 * could not be allocated.
1542 *
1543 * Since we cannot fail we will sleep for a while and try
1544 * again.
1545 *
1546 * Searching for a suitable L1 PT is expensive:
1547 * to avoid hogging the system when memory is really
1548 * scarce, use an exponential back-off so that
1549 * eventually we won't retry more than once every 8
1550 * seconds. This should allow other processes to run
1551 * to completion and free up resources.
1552 */
1553 (void) ltsleep(&lbolt, PVM, "l1ptwait", (hz << 3) >> backoff,
1554 NULL);
1555 if (--retry == 0) {
1556 retry = 10;
1557 if (backoff)
1558 --backoff;
1559 }
1560 }
1561
1562 /* Map zero page for the pmap. This will also map the L2 for it */
1563 pmap_enter(pmap, 0x00000000, systempage.pv_pa,
1564 VM_PROT_READ, VM_PROT_READ | PMAP_WIRED);
1565 pmap_update(pmap);
1566 }
1567
1568
1569 void
1570 pmap_freepagedir(pmap)
1571 struct pmap *pmap;
1572 {
1573 /* Free the memory used for the page table mapping */
1574 if (pmap->pm_vptpt != 0)
1575 uvm_km_free(kernel_map, (vaddr_t)pmap->pm_vptpt, NBPG);
1576
1577 /* junk the L1 page table */
1578 if (pmap->pm_l1pt->pt_flags & PTFLAG_STATIC) {
1579 /* Add the page table to the queue */
1580 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pmap->pm_l1pt, pt_queue);
1581 ++l1pt_static_queue_count;
1582 } else if (l1pt_queue_count < 8) {
1583 /* Add the page table to the queue */
1584 SIMPLEQ_INSERT_TAIL(&l1pt_queue, pmap->pm_l1pt, pt_queue);
1585 ++l1pt_queue_count;
1586 } else
1587 pmap_free_l1pt(pmap->pm_l1pt);
1588 }
1589
1590
1591 /*
1592 * Retire the given physical map from service.
1593 * Should only be called if the map contains no valid mappings.
1594 */
1595
1596 void
1597 pmap_destroy(pmap)
1598 struct pmap *pmap;
1599 {
1600 struct vm_page *page;
1601 int count;
1602
1603 if (pmap == NULL)
1604 return;
1605
1606 PDEBUG(0, printf("pmap_destroy(%p)\n", pmap));
1607
1608 /*
1609 * Drop reference count
1610 */
1611 simple_lock(&pmap->pm_obj.vmobjlock);
1612 count = --pmap->pm_obj.uo_refs;
1613 simple_unlock(&pmap->pm_obj.vmobjlock);
1614 if (count > 0) {
1615 return;
1616 }
1617
1618 /*
1619 * reference count is zero, free pmap resources and then free pmap.
1620 */
1621
1622 /* Remove the zero page mapping */
1623 pmap_remove(pmap, 0x00000000, 0x00000000 + NBPG);
1624 pmap_update(pmap);
1625
1626 /*
1627 * Free any page tables still mapped
1628 * This is only temporay until pmap_enter can count the number
1629 * of mappings made in a page table. Then pmap_remove() can
1630 * reduce the count and free the pagetable when the count
1631 * reaches zero. Note that entries in this list should match the
1632 * contents of the ptpt, however this is faster than walking a 1024
1633 * entries looking for pt's
1634 * taken from i386 pmap.c
1635 */
1636 while (pmap->pm_obj.memq.tqh_first != NULL) {
1637 page = pmap->pm_obj.memq.tqh_first;
1638 #ifdef DIAGNOSTIC
1639 if (page->flags & PG_BUSY)
1640 panic("pmap_release: busy page table page");
1641 #endif
1642 /* pmap_page_protect? currently no need for it. */
1643
1644 page->wire_count = 0;
1645 uvm_pagefree(page);
1646 }
1647
1648 /* Free the page dir */
1649 pmap_freepagedir(pmap);
1650
1651 /* return the pmap to the pool */
1652 pool_put(&pmap_pmap_pool, pmap);
1653 }
1654
1655
1656 /*
1657 * void pmap_reference(struct pmap *pmap)
1658 *
1659 * Add a reference to the specified pmap.
1660 */
1661
1662 void
1663 pmap_reference(pmap)
1664 struct pmap *pmap;
1665 {
1666 if (pmap == NULL)
1667 return;
1668
1669 simple_lock(&pmap->pm_lock);
1670 pmap->pm_obj.uo_refs++;
1671 simple_unlock(&pmap->pm_lock);
1672 }
1673
1674 /*
1675 * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
1676 *
1677 * Return the start and end addresses of the kernel's virtual space.
1678 * These values are setup in pmap_bootstrap and are updated as pages
1679 * are allocated.
1680 */
1681
1682 void
1683 pmap_virtual_space(start, end)
1684 vaddr_t *start;
1685 vaddr_t *end;
1686 {
1687 *start = virtual_start;
1688 *end = virtual_end;
1689 }
1690
1691
1692 /*
1693 * Activate the address space for the specified process. If the process
1694 * is the current process, load the new MMU context.
1695 */
1696 void
1697 pmap_activate(p)
1698 struct proc *p;
1699 {
1700 struct pmap *pmap = p->p_vmspace->vm_map.pmap;
1701 struct pcb *pcb = &p->p_addr->u_pcb;
1702
1703 (void) pmap_extract(pmap_kernel(), (vaddr_t)pmap->pm_pdir,
1704 (paddr_t *)&pcb->pcb_pagedir);
1705
1706 PDEBUG(0, printf("pmap_activate: p=%p pmap=%p pcb=%p pdir=%p l1=%p\n",
1707 p, pmap, pcb, pmap->pm_pdir, pcb->pcb_pagedir));
1708
1709 if (p == curproc) {
1710 PDEBUG(0, printf("pmap_activate: setting TTB\n"));
1711 setttb((u_int)pcb->pcb_pagedir);
1712 }
1713 #if 0
1714 pmap->pm_pdchanged = FALSE;
1715 #endif
1716 }
1717
1718
1719 /*
1720 * Deactivate the address space of the specified process.
1721 */
1722 void
1723 pmap_deactivate(p)
1724 struct proc *p;
1725 {
1726 }
1727
1728
1729 /*
1730 * pmap_clean_page()
1731 *
1732 * This is a local function used to work out the best strategy to clean
1733 * a single page referenced by its entry in the PV table. It's used by
1734 * pmap_copy_page, pmap_zero page and maybe some others later on.
1735 *
1736 * Its policy is effectively:
1737 * o If there are no mappings, we don't bother doing anything with the cache.
1738 * o If there is one mapping, we clean just that page.
1739 * o If there are multiple mappings, we clean the entire cache.
1740 *
1741 * So that some functions can be further optimised, it returns 0 if it didn't
1742 * clean the entire cache, or 1 if it did.
1743 *
1744 * XXX One bug in this routine is that if the pv_entry has a single page
1745 * mapped at 0x00000000 a whole cache clean will be performed rather than
1746 * just the 1 page. Since this should not occur in everyday use and if it does
1747 * it will just result in not the most efficient clean for the page.
1748 */
1749 static int
1750 pmap_clean_page(pv, is_src)
1751 struct pv_entry *pv;
1752 boolean_t is_src;
1753 {
1754 struct pmap *pmap;
1755 struct pv_entry *npv;
1756 int cache_needs_cleaning = 0;
1757 vaddr_t page_to_clean = 0;
1758
1759 if (pv == NULL)
1760 /* nothing mapped in so nothing to flush */
1761 return (0);
1762
1763 /* Since we flush the cache each time we change curproc, we
1764 * only need to flush the page if it is in the current pmap.
1765 */
1766 if (curproc)
1767 pmap = curproc->p_vmspace->vm_map.pmap;
1768 else
1769 pmap = pmap_kernel();
1770
1771 for (npv = pv; npv; npv = npv->pv_next) {
1772 if (npv->pv_pmap == pmap) {
1773 /* The page is mapped non-cacheable in
1774 * this map. No need to flush the cache.
1775 */
1776 if (npv->pv_flags & PT_NC) {
1777 #ifdef DIAGNOSTIC
1778 if (cache_needs_cleaning)
1779 panic("pmap_clean_page: "
1780 "cache inconsistency");
1781 #endif
1782 break;
1783 }
1784 #if 0
1785 /* This doesn't work, because pmap_protect
1786 doesn't flush changes on pages that it
1787 has write-protected. */
1788
1789 /* If the page is not writable and this
1790 is the source, then there is no need
1791 to flush it from the cache. */
1792 else if (is_src && ! (npv->pv_flags & PT_Wr))
1793 continue;
1794 #endif
1795 if (cache_needs_cleaning){
1796 page_to_clean = 0;
1797 break;
1798 }
1799 else
1800 page_to_clean = npv->pv_va;
1801 cache_needs_cleaning = 1;
1802 }
1803 }
1804
1805 if (page_to_clean)
1806 cpu_cache_purgeID_rng(page_to_clean, NBPG);
1807 else if (cache_needs_cleaning) {
1808 cpu_cache_purgeID();
1809 return (1);
1810 }
1811 return (0);
1812 }
1813
1814 /*
1815 * pmap_find_pv()
1816 *
1817 * This is a local function that finds a PV head for a given physical page.
1818 * This is a common op, and this function removes loads of ifdefs in the code.
1819 */
1820 static __inline struct pv_head *
1821 pmap_find_pvh(phys)
1822 paddr_t phys;
1823 {
1824 int bank, off;
1825 struct pv_head *pvh;
1826
1827 if ((bank = vm_physseg_find(atop(phys), &off)) == -1)
1828 panic("pmap_find_pv: not a real page, phys=%lx\n", phys);
1829 pvh = &vm_physmem[bank].pmseg.pvhead[off];
1830 return (pvh);
1831 }
1832
1833 /*
1834 * pmap_zero_page()
1835 *
1836 * Zero a given physical page by mapping it at a page hook point.
1837 * In doing the zero page op, the page we zero is mapped cachable, as with
1838 * StrongARM accesses to non-cached pages are non-burst making writing
1839 * _any_ bulk data very slow.
1840 */
1841 void
1842 pmap_zero_page(phys)
1843 paddr_t phys;
1844 {
1845 struct pv_head *pvh;
1846
1847 /* Get an entry for this page, and clean it it. */
1848 pvh = pmap_find_pvh(phys);
1849 simple_lock(&pvh->pvh_lock);
1850 pmap_clean_page(pvh->pvh_list, FALSE);
1851 simple_unlock(&pvh->pvh_lock);
1852
1853 /*
1854 * Hook in the page, zero it, and purge the cache for that
1855 * zeroed page. Invalidate the TLB as needed.
1856 */
1857 *page_hook0.pte = L2_PTE(phys & PG_FRAME, AP_KRW);
1858 cpu_tlb_flushD_SE(page_hook0.va);
1859 bzero_page(page_hook0.va);
1860 cpu_cache_purgeD_rng(page_hook0.va, NBPG);
1861 }
1862
1863 /* pmap_pageidlezero()
1864 *
1865 * The same as above, except that we assume that the page is not
1866 * mapped. This means we never have to flush the cache first. Called
1867 * from the idle loop.
1868 */
1869 boolean_t
1870 pmap_pageidlezero(phys)
1871 paddr_t phys;
1872 {
1873 int i, *ptr;
1874 boolean_t rv = TRUE;
1875
1876 #ifdef DIAGNOSTIC
1877 struct pv_head *pvh;
1878
1879 pvh = pmap_find_pvh(phys);
1880 if (pvh->pvh_list != NULL)
1881 panic("pmap_pageidlezero: zeroing mapped page\n");
1882 #endif
1883
1884 /*
1885 * Hook in the page, zero it, and purge the cache for that
1886 * zeroed page. Invalidate the TLB as needed.
1887 */
1888 *page_hook0.pte = L2_PTE(phys & PG_FRAME, AP_KRW);
1889 cpu_tlb_flushD_SE(page_hook0.va);
1890
1891 for (i = 0, ptr = (int *)page_hook0.va;
1892 i < (NBPG / sizeof(int)); i++) {
1893 if (sched_whichqs != 0) {
1894 /*
1895 * A process has become ready. Abort now,
1896 * so we don't keep it waiting while we
1897 * do slow memory access to finish this
1898 * page.
1899 */
1900 rv = FALSE;
1901 break;
1902 }
1903 *ptr++ = 0;
1904 }
1905
1906 if (rv)
1907 /*
1908 * if we aborted we'll rezero this page again later so don't
1909 * purge it unless we finished it
1910 */
1911 cpu_cache_purgeD_rng(page_hook0.va, NBPG);
1912 return (rv);
1913 }
1914
1915 /*
1916 * pmap_copy_page()
1917 *
1918 * Copy one physical page into another, by mapping the pages into
1919 * hook points. The same comment regarding cachability as in
1920 * pmap_zero_page also applies here.
1921 */
1922 void
1923 pmap_copy_page(src, dest)
1924 paddr_t src;
1925 paddr_t dest;
1926 {
1927 struct pv_head *src_pvh, *dest_pvh;
1928 boolean_t cleanedcache;
1929
1930 /* Get PV entries for the pages, and clean them if needed. */
1931 src_pvh = pmap_find_pvh(src);
1932
1933 simple_lock(&src_pvh->pvh_lock);
1934 cleanedcache = pmap_clean_page(src_pvh->pvh_list, TRUE);
1935 simple_unlock(&src_pvh->pvh_lock);
1936
1937 if (cleanedcache == 0) {
1938 dest_pvh = pmap_find_pvh(dest);
1939 simple_lock(&dest_pvh->pvh_lock);
1940 pmap_clean_page(dest_pvh->pvh_list, FALSE);
1941 simple_unlock(&dest_pvh->pvh_lock);
1942 }
1943 /*
1944 * Map the pages into the page hook points, copy them, and purge
1945 * the cache for the appropriate page. Invalidate the TLB
1946 * as required.
1947 */
1948 *page_hook0.pte = L2_PTE(src & PG_FRAME, AP_KRW);
1949 *page_hook1.pte = L2_PTE(dest & PG_FRAME, AP_KRW);
1950 cpu_tlb_flushD_SE(page_hook0.va);
1951 cpu_tlb_flushD_SE(page_hook1.va);
1952 bcopy_page(page_hook0.va, page_hook1.va);
1953 cpu_cache_purgeD_rng(page_hook0.va, NBPG);
1954 cpu_cache_purgeD_rng(page_hook1.va, NBPG);
1955 }
1956
1957 /*
1958 * int pmap_next_phys_page(paddr_t *addr)
1959 *
1960 * Allocate another physical page returning true or false depending
1961 * on whether a page could be allocated.
1962 */
1963
1964 paddr_t
1965 pmap_next_phys_page(addr)
1966 paddr_t addr;
1967
1968 {
1969 int loop;
1970
1971 if (addr < bootconfig.dram[0].address)
1972 return(bootconfig.dram[0].address);
1973
1974 loop = 0;
1975
1976 while (bootconfig.dram[loop].address != 0
1977 && addr > (bootconfig.dram[loop].address + bootconfig.dram[loop].pages * NBPG))
1978 ++loop;
1979
1980 if (bootconfig.dram[loop].address == 0)
1981 return(0);
1982
1983 addr += NBPG;
1984
1985 if (addr >= (bootconfig.dram[loop].address + bootconfig.dram[loop].pages * NBPG)) {
1986 if (bootconfig.dram[loop + 1].address == 0)
1987 return(0);
1988 addr = bootconfig.dram[loop + 1].address;
1989 }
1990
1991 return(addr);
1992 }
1993
1994 #if 0
1995 void
1996 pmap_pte_addref(pmap, va)
1997 struct pmap *pmap;
1998 vaddr_t va;
1999 {
2000 pd_entry_t *pde;
2001 paddr_t pa;
2002 struct vm_page *m;
2003
2004 if (pmap == pmap_kernel())
2005 return;
2006
2007 pde = pmap_pde(pmap, va & ~(3 << PDSHIFT));
2008 pa = pmap_pte_pa(pde);
2009 m = PHYS_TO_VM_PAGE(pa);
2010 ++m->wire_count;
2011 #ifdef MYCROFT_HACK
2012 printf("addref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
2013 pmap, va, pde, pa, m, m->wire_count);
2014 #endif
2015 }
2016
2017 void
2018 pmap_pte_delref(pmap, va)
2019 struct pmap *pmap;
2020 vaddr_t va;
2021 {
2022 pd_entry_t *pde;
2023 paddr_t pa;
2024 struct vm_page *m;
2025
2026 if (pmap == pmap_kernel())
2027 return;
2028
2029 pde = pmap_pde(pmap, va & ~(3 << PDSHIFT));
2030 pa = pmap_pte_pa(pde);
2031 m = PHYS_TO_VM_PAGE(pa);
2032 --m->wire_count;
2033 #ifdef MYCROFT_HACK
2034 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
2035 pmap, va, pde, pa, m, m->wire_count);
2036 #endif
2037 if (m->wire_count == 0) {
2038 #ifdef MYCROFT_HACK
2039 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p\n",
2040 pmap, va, pde, pa, m);
2041 #endif
2042 pmap_unmap_in_l1(pmap, va);
2043 uvm_pagefree(m);
2044 --pmap->pm_stats.resident_count;
2045 }
2046 }
2047 #else
2048 #define pmap_pte_addref(pmap, va)
2049 #define pmap_pte_delref(pmap, va)
2050 #endif
2051
2052 /*
2053 * Since we have a virtually indexed cache, we may need to inhibit caching if
2054 * there is more than one mapping and at least one of them is writable.
2055 * Since we purge the cache on every context switch, we only need to check for
2056 * other mappings within the same pmap, or kernel_pmap.
2057 * This function is also called when a page is unmapped, to possibly reenable
2058 * caching on any remaining mappings.
2059 *
2060 * The code implements the following logic, where:
2061 *
2062 * KW = # of kernel read/write pages
2063 * KR = # of kernel read only pages
2064 * UW = # of user read/write pages
2065 * UR = # of user read only pages
2066 * OW = # of user read/write pages in another pmap, then
2067 *
2068 * KC = kernel mapping is cacheable
2069 * UC = user mapping is cacheable
2070 *
2071 * KW=0,KR=0 KW=0,KR>0 KW=1,KR=0 KW>1,KR>=0
2072 * +---------------------------------------------
2073 * UW=0,UR=0,OW=0 | --- KC=1 KC=1 KC=0
2074 * UW=0,UR>0,OW=0 | UC=1 KC=1,UC=1 KC=0,UC=0 KC=0,UC=0
2075 * UW=0,UR>0,OW>0 | UC=1 KC=0,UC=1 KC=0,UC=0 KC=0,UC=0
2076 * UW=1,UR=0,OW=0 | UC=1 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0
2077 * UW>1,UR>=0,OW>=0 | UC=0 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0
2078 *
2079 * Note that the pmap must have it's ptes mapped in, and passed with ptes.
2080 */
2081 __inline static void
2082 pmap_vac_me_harder(struct pmap *pmap, struct pv_head *pvh, pt_entry_t *ptes,
2083 boolean_t clear_cache)
2084 {
2085 if (pmap == pmap_kernel())
2086 pmap_vac_me_kpmap(pmap, pvh, ptes, clear_cache);
2087 else
2088 pmap_vac_me_user(pmap, pvh, ptes, clear_cache);
2089 }
2090
2091 static void
2092 pmap_vac_me_kpmap(struct pmap *pmap, struct pv_head *pvh, pt_entry_t *ptes,
2093 boolean_t clear_cache)
2094 {
2095 int user_entries = 0;
2096 int user_writable = 0;
2097 int user_cacheable = 0;
2098 int kernel_entries = 0;
2099 int kernel_writable = 0;
2100 int kernel_cacheable = 0;
2101 struct pv_entry *pv;
2102 struct pmap *last_pmap = pmap;
2103
2104 #ifdef DIAGNOSTIC
2105 if (pmap != pmap_kernel())
2106 panic("pmap_vac_me_kpmap: pmap != pmap_kernel()");
2107 #endif
2108
2109 /*
2110 * Pass one, see if there are both kernel and user pmaps for
2111 * this page. Calculate whether there are user-writable or
2112 * kernel-writable pages.
2113 */
2114 for (pv = pvh->pvh_list; pv != NULL; pv = pv->pv_next) {
2115 if (pv->pv_pmap != pmap) {
2116 user_entries++;
2117 if (pv->pv_flags & PT_Wr)
2118 user_writable++;
2119 if ((pv->pv_flags & PT_NC) == 0)
2120 user_cacheable++;
2121 } else {
2122 kernel_entries++;
2123 if (pv->pv_flags & PT_Wr)
2124 kernel_writable++;
2125 if ((pv->pv_flags & PT_NC) == 0)
2126 kernel_cacheable++;
2127 }
2128 }
2129
2130 /*
2131 * We know we have just been updating a kernel entry, so if
2132 * all user pages are already cacheable, then there is nothing
2133 * further to do.
2134 */
2135 if (kernel_entries == 0 &&
2136 user_cacheable == user_entries)
2137 return;
2138
2139 if (user_entries) {
2140 /*
2141 * Scan over the list again, for each entry, if it
2142 * might not be set correctly, call pmap_vac_me_user
2143 * to recalculate the settings.
2144 */
2145 for (pv = pvh->pvh_list; pv; pv = pv->pv_next) {
2146 /*
2147 * We know kernel mappings will get set
2148 * correctly in other calls. We also know
2149 * that if the pmap is the same as last_pmap
2150 * then we've just handled this entry.
2151 */
2152 if (pv->pv_pmap == pmap || pv->pv_pmap == last_pmap)
2153 continue;
2154 /*
2155 * If there are kernel entries and this page
2156 * is writable but non-cacheable, then we can
2157 * skip this entry also.
2158 */
2159 if (kernel_entries > 0 &&
2160 (pv->pv_flags & (PT_NC | PT_Wr)) ==
2161 (PT_NC | PT_Wr))
2162 continue;
2163 /*
2164 * Similarly if there are no kernel-writable
2165 * entries and the page is already
2166 * read-only/cacheable.
2167 */
2168 if (kernel_writable == 0 &&
2169 (pv->pv_flags & (PT_NC | PT_Wr)) == 0)
2170 continue;
2171 /*
2172 * For some of the remaining cases, we know
2173 * that we must recalculate, but for others we
2174 * can't tell if they are correct or not, so
2175 * we recalculate anyway.
2176 */
2177 pmap_unmap_ptes(last_pmap);
2178 last_pmap = pv->pv_pmap;
2179 ptes = pmap_map_ptes(last_pmap);
2180 pmap_vac_me_user(last_pmap, pvh, ptes,
2181 pmap_is_curpmap(last_pmap));
2182 }
2183 /* Restore the pte mapping that was passed to us. */
2184 if (last_pmap != pmap) {
2185 pmap_unmap_ptes(last_pmap);
2186 ptes = pmap_map_ptes(pmap);
2187 }
2188 if (kernel_entries == 0)
2189 return;
2190 }
2191
2192 pmap_vac_me_user(pmap, pvh, ptes, clear_cache);
2193 return;
2194 }
2195
2196 static void
2197 pmap_vac_me_user(struct pmap *pmap, struct pv_head *pvh, pt_entry_t *ptes,
2198 boolean_t clear_cache)
2199 {
2200 struct pmap *kpmap = pmap_kernel();
2201 struct pv_entry *pv, *npv;
2202 int entries = 0;
2203 int writable = 0;
2204 int cacheable_entries = 0;
2205 int kern_cacheable = 0;
2206 int other_writable = 0;
2207
2208 pv = pvh->pvh_list;
2209 KASSERT(ptes != NULL);
2210
2211 /*
2212 * Count mappings and writable mappings in this pmap.
2213 * Include kernel mappings as part of our own.
2214 * Keep a pointer to the first one.
2215 */
2216 for (npv = pv; npv; npv = npv->pv_next) {
2217 /* Count mappings in the same pmap */
2218 if (pmap == npv->pv_pmap ||
2219 kpmap == npv->pv_pmap) {
2220 if (entries++ == 0)
2221 pv = npv;
2222 /* Cacheable mappings */
2223 if ((npv->pv_flags & PT_NC) == 0) {
2224 cacheable_entries++;
2225 if (kpmap == npv->pv_pmap)
2226 kern_cacheable++;
2227 }
2228 /* Writable mappings */
2229 if (npv->pv_flags & PT_Wr)
2230 ++writable;
2231 } else if (npv->pv_flags & PT_Wr)
2232 other_writable = 1;
2233 }
2234
2235 PDEBUG(3,printf("pmap_vac_me_harder: pmap %p Entries %d, "
2236 "writable %d cacheable %d %s\n", pmap, entries, writable,
2237 cacheable_entries, clear_cache ? "clean" : "no clean"));
2238
2239 /*
2240 * Enable or disable caching as necessary.
2241 * Note: the first entry might be part of the kernel pmap,
2242 * so we can't assume this is indicative of the state of the
2243 * other (maybe non-kpmap) entries.
2244 */
2245 if ((entries > 1 && writable) ||
2246 (entries > 0 && pmap == kpmap && other_writable)) {
2247 if (cacheable_entries == 0)
2248 return;
2249 for (npv = pv; npv; npv = npv->pv_next) {
2250 if ((pmap == npv->pv_pmap
2251 || kpmap == npv->pv_pmap) &&
2252 (npv->pv_flags & PT_NC) == 0) {
2253 ptes[arm_byte_to_page(npv->pv_va)] &=
2254 ~(PT_C | PT_B);
2255 npv->pv_flags |= PT_NC;
2256 /*
2257 * If this page needs flushing from the
2258 * cache, and we aren't going to do it
2259 * below, do it now.
2260 */
2261 if ((cacheable_entries < 4 &&
2262 (clear_cache || npv->pv_pmap == kpmap)) ||
2263 (npv->pv_pmap == kpmap &&
2264 !clear_cache && kern_cacheable < 4)) {
2265 cpu_cache_purgeID_rng(npv->pv_va,
2266 NBPG);
2267 cpu_tlb_flushID_SE(npv->pv_va);
2268 }
2269 }
2270 }
2271 if ((clear_cache && cacheable_entries >= 4) ||
2272 kern_cacheable >= 4) {
2273 cpu_cache_purgeID();
2274 cpu_tlb_flushID();
2275 }
2276 } else if (entries > 0) {
2277 /*
2278 * Turn cacheing back on for some pages. If it is a kernel
2279 * page, only do so if there are no other writable pages.
2280 */
2281 for (npv = pv; npv; npv = npv->pv_next) {
2282 if ((pmap == npv->pv_pmap ||
2283 (kpmap == npv->pv_pmap && other_writable == 0)) &&
2284 (npv->pv_flags & PT_NC)) {
2285 ptes[arm_byte_to_page(npv->pv_va)] |=
2286 pte_cache_mode;
2287 npv->pv_flags &= ~PT_NC;
2288 }
2289 }
2290 }
2291 }
2292
2293 /*
2294 * pmap_remove()
2295 *
2296 * pmap_remove is responsible for nuking a number of mappings for a range
2297 * of virtual address space in the current pmap. To do this efficiently
2298 * is interesting, because in a number of cases a wide virtual address
2299 * range may be supplied that contains few actual mappings. So, the
2300 * optimisations are:
2301 * 1. Try and skip over hunks of address space for which an L1 entry
2302 * does not exist.
2303 * 2. Build up a list of pages we've hit, up to a maximum, so we can
2304 * maybe do just a partial cache clean. This path of execution is
2305 * complicated by the fact that the cache must be flushed _before_
2306 * the PTE is nuked, being a VAC :-)
2307 * 3. Maybe later fast-case a single page, but I don't think this is
2308 * going to make _that_ much difference overall.
2309 */
2310
2311 #define PMAP_REMOVE_CLEAN_LIST_SIZE 3
2312
2313 void
2314 pmap_remove(pmap, sva, eva)
2315 struct pmap *pmap;
2316 vaddr_t sva;
2317 vaddr_t eva;
2318 {
2319 int cleanlist_idx = 0;
2320 struct pagelist {
2321 vaddr_t va;
2322 pt_entry_t *pte;
2323 } cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
2324 pt_entry_t *pte = 0, *ptes;
2325 paddr_t pa;
2326 int pmap_active;
2327 struct pv_head *pvh;
2328
2329 /* Exit quick if there is no pmap */
2330 if (!pmap)
2331 return;
2332
2333 PDEBUG(0, printf("pmap_remove: pmap=%p sva=%08lx eva=%08lx\n", pmap, sva, eva));
2334
2335 sva &= PG_FRAME;
2336 eva &= PG_FRAME;
2337
2338 /*
2339 * we lock in the pmap => pv_head direction
2340 */
2341 PMAP_MAP_TO_HEAD_LOCK();
2342
2343 ptes = pmap_map_ptes(pmap);
2344 /* Get a page table pointer */
2345 while (sva < eva) {
2346 if (pmap_pde_v(pmap_pde(pmap, sva)))
2347 break;
2348 sva = (sva & PD_MASK) + NBPD;
2349 }
2350
2351 pte = &ptes[arm_byte_to_page(sva)];
2352 /* Note if the pmap is active thus require cache and tlb cleans */
2353 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap)
2354 || (pmap == pmap_kernel()))
2355 pmap_active = 1;
2356 else
2357 pmap_active = 0;
2358
2359 /* Now loop along */
2360 while (sva < eva) {
2361 /* Check if we can move to the next PDE (l1 chunk) */
2362 if (!(sva & PT_MASK))
2363 if (!pmap_pde_v(pmap_pde(pmap, sva))) {
2364 sva += NBPD;
2365 pte += arm_byte_to_page(NBPD);
2366 continue;
2367 }
2368
2369 /* We've found a valid PTE, so this page of PTEs has to go. */
2370 if (pmap_pte_v(pte)) {
2371 int bank, off;
2372
2373 /* Update statistics */
2374 --pmap->pm_stats.resident_count;
2375
2376 /*
2377 * Add this page to our cache remove list, if we can.
2378 * If, however the cache remove list is totally full,
2379 * then do a complete cache invalidation taking note
2380 * to backtrack the PTE table beforehand, and ignore
2381 * the lists in future because there's no longer any
2382 * point in bothering with them (we've paid the
2383 * penalty, so will carry on unhindered). Otherwise,
2384 * when we fall out, we just clean the list.
2385 */
2386 PDEBUG(10, printf("remove: inv pte at %p(%x) ", pte, *pte));
2387 pa = pmap_pte_pa(pte);
2388
2389 if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
2390 /* Add to the clean list. */
2391 cleanlist[cleanlist_idx].pte = pte;
2392 cleanlist[cleanlist_idx].va = sva;
2393 cleanlist_idx++;
2394 } else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
2395 int cnt;
2396
2397 /* Nuke everything if needed. */
2398 if (pmap_active) {
2399 cpu_cache_purgeID();
2400 cpu_tlb_flushID();
2401 }
2402
2403 /*
2404 * Roll back the previous PTE list,
2405 * and zero out the current PTE.
2406 */
2407 for (cnt = 0; cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
2408 *cleanlist[cnt].pte = 0;
2409 pmap_pte_delref(pmap, cleanlist[cnt].va);
2410 }
2411 *pte = 0;
2412 pmap_pte_delref(pmap, sva);
2413 cleanlist_idx++;
2414 } else {
2415 /*
2416 * We've already nuked the cache and
2417 * TLB, so just carry on regardless,
2418 * and we won't need to do it again
2419 */
2420 *pte = 0;
2421 pmap_pte_delref(pmap, sva);
2422 }
2423
2424 /*
2425 * Update flags. In a number of circumstances,
2426 * we could cluster a lot of these and do a
2427 * number of sequential pages in one go.
2428 */
2429 if ((bank = vm_physseg_find(atop(pa), &off)) != -1) {
2430 struct pv_entry *pve;
2431 pvh = &vm_physmem[bank].pmseg.pvhead[off];
2432 simple_lock(&pvh->pvh_lock);
2433 pve = pmap_remove_pv(pvh, pmap, sva);
2434 pmap_free_pv(pmap, pve);
2435 pmap_vac_me_harder(pmap, pvh, ptes, FALSE);
2436 simple_unlock(&pvh->pvh_lock);
2437 }
2438 }
2439 sva += NBPG;
2440 pte++;
2441 }
2442
2443 pmap_unmap_ptes(pmap);
2444 /*
2445 * Now, if we've fallen through down to here, chances are that there
2446 * are less than PMAP_REMOVE_CLEAN_LIST_SIZE mappings left.
2447 */
2448 if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
2449 u_int cnt;
2450
2451 for (cnt = 0; cnt < cleanlist_idx; cnt++) {
2452 if (pmap_active) {
2453 cpu_cache_purgeID_rng(cleanlist[cnt].va, NBPG);
2454 *cleanlist[cnt].pte = 0;
2455 cpu_tlb_flushID_SE(cleanlist[cnt].va);
2456 } else
2457 *cleanlist[cnt].pte = 0;
2458 pmap_pte_delref(pmap, cleanlist[cnt].va);
2459 }
2460 }
2461 PMAP_MAP_TO_HEAD_UNLOCK();
2462 }
2463
2464 /*
2465 * Routine: pmap_remove_all
2466 * Function:
2467 * Removes this physical page from
2468 * all physical maps in which it resides.
2469 * Reflects back modify bits to the pager.
2470 */
2471
2472 void
2473 pmap_remove_all(pa)
2474 paddr_t pa;
2475 {
2476 struct pv_entry *pv, *npv;
2477 struct pv_head *pvh;
2478 struct pmap *pmap;
2479 pt_entry_t *pte, *ptes;
2480
2481 PDEBUG(0, printf("pmap_remove_all: pa=%lx ", pa));
2482
2483 /* set pv_head => pmap locking */
2484 PMAP_HEAD_TO_MAP_LOCK();
2485
2486 pvh = pmap_find_pvh(pa);
2487 simple_lock(&pvh->pvh_lock);
2488
2489 pv = pvh->pvh_list;
2490 if (pv == NULL)
2491 {
2492 PDEBUG(0, printf("free page\n"));
2493 simple_unlock(&pvh->pvh_lock);
2494 PMAP_HEAD_TO_MAP_UNLOCK();
2495 return;
2496 }
2497 pmap_clean_page(pv, FALSE);
2498
2499 while (pv) {
2500 pmap = pv->pv_pmap;
2501 ptes = pmap_map_ptes(pmap);
2502 pte = &ptes[arm_byte_to_page(pv->pv_va)];
2503
2504 PDEBUG(0, printf("[%p,%08x,%08lx,%08x] ", pmap, *pte,
2505 pv->pv_va, pv->pv_flags));
2506 #ifdef DEBUG
2507 if (!pmap_pde_v(pmap_pde(pmap, va)) || !pmap_pte_v(pte)
2508 || pmap_pte_pa(pte) != pa)
2509 panic("pmap_remove_all: bad mapping");
2510 #endif /* DEBUG */
2511
2512 /*
2513 * Update statistics
2514 */
2515 --pmap->pm_stats.resident_count;
2516
2517 /* Wired bit */
2518 if (pv->pv_flags & PT_W)
2519 --pmap->pm_stats.wired_count;
2520
2521 /*
2522 * Invalidate the PTEs.
2523 * XXX: should cluster them up and invalidate as many
2524 * as possible at once.
2525 */
2526
2527 #ifdef needednotdone
2528 reduce wiring count on page table pages as references drop
2529 #endif
2530
2531 *pte = 0;
2532 pmap_pte_delref(pmap, pv->pv_va);
2533
2534 npv = pv->pv_next;
2535 pmap_free_pv(pmap, pv);
2536 pv = npv;
2537 pmap_unmap_ptes(pmap);
2538 }
2539 pvh->pvh_list = NULL;
2540 simple_unlock(&pvh->pvh_lock);
2541 PMAP_HEAD_TO_MAP_UNLOCK();
2542
2543 PDEBUG(0, printf("done\n"));
2544 cpu_tlb_flushID();
2545 }
2546
2547
2548 /*
2549 * Set the physical protection on the specified range of this map as requested.
2550 */
2551
2552 void
2553 pmap_protect(pmap, sva, eva, prot)
2554 struct pmap *pmap;
2555 vaddr_t sva;
2556 vaddr_t eva;
2557 vm_prot_t prot;
2558 {
2559 pt_entry_t *pte = NULL, *ptes;
2560 int armprot;
2561 int flush = 0;
2562 paddr_t pa;
2563 int bank, off;
2564 struct pv_head *pvh;
2565
2566 PDEBUG(0, printf("pmap_protect: pmap=%p %08lx->%08lx %x\n",
2567 pmap, sva, eva, prot));
2568
2569 if (~prot & VM_PROT_READ) {
2570 /* Just remove the mappings. */
2571 pmap_remove(pmap, sva, eva);
2572 return;
2573 }
2574 if (prot & VM_PROT_WRITE) {
2575 /*
2576 * If this is a read->write transition, just ignore it and let
2577 * uvm_fault() take care of it later.
2578 */
2579 return;
2580 }
2581
2582 sva &= PG_FRAME;
2583 eva &= PG_FRAME;
2584
2585 /* Need to lock map->head */
2586 PMAP_MAP_TO_HEAD_LOCK();
2587
2588 ptes = pmap_map_ptes(pmap);
2589 /*
2590 * We need to acquire a pointer to a page table page before entering
2591 * the following loop.
2592 */
2593 while (sva < eva) {
2594 if (pmap_pde_v(pmap_pde(pmap, sva)))
2595 break;
2596 sva = (sva & PD_MASK) + NBPD;
2597 }
2598
2599 pte = &ptes[arm_byte_to_page(sva)];
2600
2601 while (sva < eva) {
2602 /* only check once in a while */
2603 if ((sva & PT_MASK) == 0) {
2604 if (!pmap_pde_v(pmap_pde(pmap, sva))) {
2605 /* We can race ahead here, to the next pde. */
2606 sva += NBPD;
2607 pte += arm_byte_to_page(NBPD);
2608 continue;
2609 }
2610 }
2611
2612 if (!pmap_pte_v(pte))
2613 goto next;
2614
2615 flush = 1;
2616
2617 armprot = 0;
2618 if (sva < VM_MAXUSER_ADDRESS)
2619 armprot |= PT_AP(AP_U);
2620 else if (sva < VM_MAX_ADDRESS)
2621 armprot |= PT_AP(AP_W); /* XXX Ekk what is this ? */
2622 *pte = (*pte & 0xfffff00f) | armprot;
2623
2624 pa = pmap_pte_pa(pte);
2625
2626 /* Get the physical page index */
2627
2628 /* Clear write flag */
2629 if ((bank = vm_physseg_find(atop(pa), &off)) != -1) {
2630 pvh = &vm_physmem[bank].pmseg.pvhead[off];
2631 simple_lock(&pvh->pvh_lock);
2632 (void) pmap_modify_pv(pmap, sva, pvh, PT_Wr, 0);
2633 pmap_vac_me_harder(pmap, pvh, ptes, FALSE);
2634 simple_unlock(&pvh->pvh_lock);
2635 }
2636
2637 next:
2638 sva += NBPG;
2639 pte++;
2640 }
2641 pmap_unmap_ptes(pmap);
2642 PMAP_MAP_TO_HEAD_UNLOCK();
2643 if (flush)
2644 cpu_tlb_flushID();
2645 }
2646
2647 /*
2648 * void pmap_enter(struct pmap *pmap, vaddr_t va, paddr_t pa, vm_prot_t prot,
2649 * int flags)
2650 *
2651 * Insert the given physical page (p) at
2652 * the specified virtual address (v) in the
2653 * target physical map with the protection requested.
2654 *
2655 * If specified, the page will be wired down, meaning
2656 * that the related pte can not be reclaimed.
2657 *
2658 * NB: This is the only routine which MAY NOT lazy-evaluate
2659 * or lose information. That is, this routine must actually
2660 * insert this page into the given map NOW.
2661 */
2662
2663 int
2664 pmap_enter(pmap, va, pa, prot, flags)
2665 struct pmap *pmap;
2666 vaddr_t va;
2667 paddr_t pa;
2668 vm_prot_t prot;
2669 int flags;
2670 {
2671 pt_entry_t *pte, *ptes;
2672 u_int npte;
2673 int bank, off;
2674 paddr_t opa;
2675 int nflags;
2676 boolean_t wired = (flags & PMAP_WIRED) != 0;
2677 struct pv_entry *pve;
2678 struct pv_head *pvh;
2679 int error;
2680
2681 PDEBUG(5, printf("pmap_enter: V%08lx P%08lx in pmap %p prot=%08x, wired = %d\n",
2682 va, pa, pmap, prot, wired));
2683
2684 #ifdef DIAGNOSTIC
2685 /* Valid address ? */
2686 if (va >= (KERNEL_VM_BASE + KERNEL_VM_SIZE))
2687 panic("pmap_enter: too big");
2688 if (pmap != pmap_kernel() && va != 0) {
2689 if (va < VM_MIN_ADDRESS || va >= VM_MAXUSER_ADDRESS)
2690 panic("pmap_enter: kernel page in user map");
2691 } else {
2692 if (va >= VM_MIN_ADDRESS && va < VM_MAXUSER_ADDRESS)
2693 panic("pmap_enter: user page in kernel map");
2694 if (va >= VM_MAXUSER_ADDRESS && va < VM_MAX_ADDRESS)
2695 panic("pmap_enter: entering PT page");
2696 }
2697 #endif
2698 /* get lock */
2699 PMAP_MAP_TO_HEAD_LOCK();
2700 /*
2701 * Get a pointer to the pte for this virtual address. If the
2702 * pte pointer is NULL then we are missing the L2 page table
2703 * so we need to create one.
2704 */
2705 /* XXX horrible hack to get us working with lockdebug */
2706 simple_lock(&pmap->pm_obj.vmobjlock);
2707 pte = pmap_pte(pmap, va);
2708 if (!pte) {
2709 struct vm_page *ptp;
2710
2711 /* if failure is allowed then don't try too hard */
2712 ptp = pmap_get_ptp(pmap, va, flags & PMAP_CANFAIL);
2713 if (ptp == NULL) {
2714 if (flags & PMAP_CANFAIL) {
2715 error = ENOMEM;
2716 goto out;
2717 }
2718 panic("pmap_enter: get ptp failed");
2719 }
2720
2721 pte = pmap_pte(pmap, va);
2722 #ifdef DIAGNOSTIC
2723 if (!pte)
2724 panic("pmap_enter: no pte");
2725 #endif
2726 }
2727
2728 nflags = 0;
2729 if (prot & VM_PROT_WRITE)
2730 nflags |= PT_Wr;
2731 if (wired)
2732 nflags |= PT_W;
2733
2734 /* More debugging info */
2735 PDEBUG(5, printf("pmap_enter: pte for V%08lx = V%p (%08x)\n", va, pte,
2736 *pte));
2737
2738 /* Is the pte valid ? If so then this page is already mapped */
2739 if (pmap_pte_v(pte)) {
2740 /* Get the physical address of the current page mapped */
2741 opa = pmap_pte_pa(pte);
2742
2743 #ifdef MYCROFT_HACK
2744 printf("pmap_enter: pmap=%p va=%lx pa=%lx opa=%lx\n", pmap, va, pa, opa);
2745 #endif
2746
2747 /* Are we mapping the same page ? */
2748 if (opa == pa) {
2749 /* All we must be doing is changing the protection */
2750 PDEBUG(0, printf("Case 02 in pmap_enter (V%08lx P%08lx)\n",
2751 va, pa));
2752
2753 /* Has the wiring changed ? */
2754 if ((bank = vm_physseg_find(atop(pa), &off)) != -1) {
2755 pvh = &vm_physmem[bank].pmseg.pvhead[off];
2756 simple_lock(&pvh->pvh_lock);
2757 (void) pmap_modify_pv(pmap, va, pvh,
2758 PT_Wr | PT_W, nflags);
2759 simple_unlock(&pvh->pvh_lock);
2760 } else {
2761 pvh = NULL;
2762 }
2763 } else {
2764 /* We are replacing the page with a new one. */
2765 cpu_cache_purgeID_rng(va, NBPG);
2766
2767 PDEBUG(0, printf("Case 03 in pmap_enter (V%08lx P%08lx P%08lx)\n",
2768 va, pa, opa));
2769
2770 /*
2771 * If it is part of our managed memory then we
2772 * must remove it from the PV list
2773 */
2774 if ((bank = vm_physseg_find(atop(opa), &off)) != -1) {
2775 pvh = &vm_physmem[bank].pmseg.pvhead[off];
2776 simple_lock(&pvh->pvh_lock);
2777 pve = pmap_remove_pv(pvh, pmap, va);
2778 simple_unlock(&pvh->pvh_lock);
2779 } else {
2780 pve = NULL;
2781 }
2782
2783 goto enter;
2784 }
2785 } else {
2786 opa = 0;
2787 pve = NULL;
2788 pmap_pte_addref(pmap, va);
2789
2790 /* pte is not valid so we must be hooking in a new page */
2791 ++pmap->pm_stats.resident_count;
2792
2793 enter:
2794 /*
2795 * Enter on the PV list if part of our managed memory
2796 */
2797 bank = vm_physseg_find(atop(pa), &off);
2798
2799 if (pmap_initialized && (bank != -1)) {
2800 pvh = &vm_physmem[bank].pmseg.pvhead[off];
2801 if (pve == NULL) {
2802 pve = pmap_alloc_pv(pmap, ALLOCPV_NEED);
2803 if (pve == NULL) {
2804 if (flags & PMAP_CANFAIL) {
2805 error = ENOMEM;
2806 goto out;
2807 }
2808 panic("pmap_enter: no pv entries available");
2809 }
2810 }
2811 /* enter_pv locks pvh when adding */
2812 pmap_enter_pv(pvh, pve, pmap, va, NULL, nflags);
2813 } else {
2814 pvh = NULL;
2815 if (pve != NULL)
2816 pmap_free_pv(pmap, pve);
2817 }
2818 }
2819
2820 #ifdef MYCROFT_HACK
2821 if (mycroft_hack)
2822 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);
2823 #endif
2824
2825 /* Construct the pte, giving the correct access. */
2826 npte = (pa & PG_FRAME);
2827
2828 /* VA 0 is magic. */
2829 if (pmap != pmap_kernel() && va != 0)
2830 npte |= PT_AP(AP_U);
2831
2832 if (pmap_initialized && bank != -1) {
2833 #ifdef DIAGNOSTIC
2834 if ((flags & VM_PROT_ALL) & ~prot)
2835 panic("pmap_enter: access_type exceeds prot");
2836 #endif
2837 npte |= pte_cache_mode;
2838 if (flags & VM_PROT_WRITE) {
2839 npte |= L2_SPAGE | PT_AP(AP_W);
2840 vm_physmem[bank].pmseg.attrs[off] |= PT_H | PT_M;
2841 } else if (flags & VM_PROT_ALL) {
2842 npte |= L2_SPAGE;
2843 vm_physmem[bank].pmseg.attrs[off] |= PT_H;
2844 } else
2845 npte |= L2_INVAL;
2846 } else {
2847 if (prot & VM_PROT_WRITE)
2848 npte |= L2_SPAGE | PT_AP(AP_W);
2849 else if (prot & VM_PROT_ALL)
2850 npte |= L2_SPAGE;
2851 else
2852 npte |= L2_INVAL;
2853 }
2854
2855 #ifdef MYCROFT_HACK
2856 if (mycroft_hack)
2857 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);
2858 #endif
2859
2860 *pte = npte;
2861
2862 if (pmap_initialized && bank != -1)
2863 {
2864 boolean_t pmap_active = FALSE;
2865 /* XXX this will change once the whole of pmap_enter uses
2866 * map_ptes
2867 */
2868 ptes = pmap_map_ptes(pmap);
2869 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap)
2870 || (pmap == pmap_kernel()))
2871 pmap_active = TRUE;
2872 simple_lock(&pvh->pvh_lock);
2873 pmap_vac_me_harder(pmap, pvh, ptes, pmap_active);
2874 simple_unlock(&pvh->pvh_lock);
2875 pmap_unmap_ptes(pmap);
2876 }
2877
2878 /* Better flush the TLB ... */
2879 cpu_tlb_flushID_SE(va);
2880 error = 0;
2881 out:
2882 simple_unlock(&pmap->pm_obj.vmobjlock);
2883 PMAP_MAP_TO_HEAD_UNLOCK();
2884 PDEBUG(5, printf("pmap_enter: pte = V%p %08x\n", pte, *pte));
2885
2886 return error;
2887 }
2888
2889 void
2890 pmap_kenter_pa(va, pa, prot)
2891 vaddr_t va;
2892 paddr_t pa;
2893 vm_prot_t prot;
2894 {
2895 struct pmap *pmap = pmap_kernel();
2896 pt_entry_t *pte;
2897 struct vm_page *pg;
2898
2899 if (!pmap_pde_v(pmap_pde(pmap, va))) {
2900
2901 /*
2902 * For the kernel pmaps it would be better to ensure
2903 * that they are always present, and to grow the
2904 * kernel as required.
2905 */
2906
2907 /* must lock the pmap */
2908 simple_lock(&(pmap_kernel()->pm_obj.vmobjlock));
2909 /* Allocate a page table */
2910 pg = uvm_pagealloc(&(pmap_kernel()->pm_obj), 0, NULL,
2911 UVM_PGA_USERESERVE | UVM_PGA_ZERO);
2912 if (pg == NULL) {
2913 panic("pmap_kenter_pa: no free pages");
2914 }
2915 pg->flags &= ~PG_BUSY; /* never busy */
2916
2917 /* Wire this page table into the L1. */
2918 pmap_map_in_l1(pmap, va, VM_PAGE_TO_PHYS(pg), TRUE);
2919 simple_unlock(&(pmap_kernel()->pm_obj.vmobjlock));
2920 }
2921 pte = vtopte(va);
2922 KASSERT(!pmap_pte_v(pte));
2923 *pte = L2_PTE(pa, AP_KRW);
2924 }
2925
2926 void
2927 pmap_kremove(va, len)
2928 vaddr_t va;
2929 vsize_t len;
2930 {
2931 pt_entry_t *pte;
2932
2933 for (len >>= PAGE_SHIFT; len > 0; len--, va += PAGE_SIZE) {
2934
2935 /*
2936 * We assume that we will only be called with small
2937 * regions of memory.
2938 */
2939
2940 KASSERT(pmap_pde_v(pmap_pde(pmap_kernel(), va)));
2941 pte = vtopte(va);
2942 cpu_cache_purgeID_rng(va, PAGE_SIZE);
2943 *pte = 0;
2944 cpu_tlb_flushID_SE(va);
2945 }
2946 }
2947
2948 /*
2949 * pmap_page_protect:
2950 *
2951 * Lower the permission for all mappings to a given page.
2952 */
2953
2954 void
2955 pmap_page_protect(pg, prot)
2956 struct vm_page *pg;
2957 vm_prot_t prot;
2958 {
2959 paddr_t pa = VM_PAGE_TO_PHYS(pg);
2960
2961 PDEBUG(0, printf("pmap_page_protect(pa=%lx, prot=%d)\n", pa, prot));
2962
2963 switch(prot) {
2964 case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
2965 case VM_PROT_READ|VM_PROT_WRITE:
2966 return;
2967
2968 case VM_PROT_READ:
2969 case VM_PROT_READ|VM_PROT_EXECUTE:
2970 pmap_copy_on_write(pa);
2971 break;
2972
2973 default:
2974 pmap_remove_all(pa);
2975 break;
2976 }
2977 }
2978
2979
2980 /*
2981 * Routine: pmap_unwire
2982 * Function: Clear the wired attribute for a map/virtual-address
2983 * pair.
2984 * In/out conditions:
2985 * The mapping must already exist in the pmap.
2986 */
2987
2988 void
2989 pmap_unwire(pmap, va)
2990 struct pmap *pmap;
2991 vaddr_t va;
2992 {
2993 pt_entry_t *pte;
2994 paddr_t pa;
2995 int bank, off;
2996 struct pv_head *pvh;
2997
2998 /*
2999 * Make sure pmap is valid. -dct
3000 */
3001 if (pmap == NULL)
3002 return;
3003
3004 /* Get the pte */
3005 pte = pmap_pte(pmap, va);
3006 if (!pte)
3007 return;
3008
3009 /* Extract the physical address of the page */
3010 pa = pmap_pte_pa(pte);
3011
3012 if ((bank = vm_physseg_find(atop(pa), &off)) == -1)
3013 return;
3014 pvh = &vm_physmem[bank].pmseg.pvhead[off];
3015 simple_lock(&pvh->pvh_lock);
3016 /* Update the wired bit in the pv entry for this page. */
3017 (void) pmap_modify_pv(pmap, va, pvh, PT_W, 0);
3018 simple_unlock(&pvh->pvh_lock);
3019 }
3020
3021 /*
3022 * pt_entry_t *pmap_pte(struct pmap *pmap, vaddr_t va)
3023 *
3024 * Return the pointer to a page table entry corresponding to the supplied
3025 * virtual address.
3026 *
3027 * The page directory is first checked to make sure that a page table
3028 * for the address in question exists and if it does a pointer to the
3029 * entry is returned.
3030 *
3031 * The way this works is that that the kernel page tables are mapped
3032 * into the memory map at ALT_PAGE_TBLS_BASE to ALT_PAGE_TBLS_BASE+4MB.
3033 * This allows page tables to be located quickly.
3034 */
3035 pt_entry_t *
3036 pmap_pte(pmap, va)
3037 struct pmap *pmap;
3038 vaddr_t va;
3039 {
3040 pt_entry_t *ptp;
3041 pt_entry_t *result;
3042
3043 /* The pmap must be valid */
3044 if (!pmap)
3045 return(NULL);
3046
3047 /* Return the address of the pte */
3048 PDEBUG(10, printf("pmap_pte: pmap=%p va=V%08lx pde = V%p (%08X)\n",
3049 pmap, va, pmap_pde(pmap, va), *(pmap_pde(pmap, va))));
3050
3051 /* Do we have a valid pde ? If not we don't have a page table */
3052 if (!pmap_pde_v(pmap_pde(pmap, va))) {
3053 PDEBUG(0, printf("pmap_pte: failed - pde = %p\n",
3054 pmap_pde(pmap, va)));
3055 return(NULL);
3056 }
3057
3058 PDEBUG(10, printf("pmap pagetable = P%08lx current = P%08x\n",
3059 pmap->pm_pptpt, (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE
3060 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) +
3061 (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & PG_FRAME)));
3062
3063 /*
3064 * If the pmap is the kernel pmap or the pmap is the active one
3065 * then we can just return a pointer to entry relative to
3066 * PROCESS_PAGE_TBLS_BASE.
3067 * Otherwise we need to map the page tables to an alternative
3068 * address and reference them there.
3069 */
3070 if (pmap == pmap_kernel() || pmap->pm_pptpt
3071 == (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE
3072 + ((PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) &
3073 ~3) + (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & PG_FRAME)) {
3074 ptp = (pt_entry_t *)PROCESS_PAGE_TBLS_BASE;
3075 } else {
3076 struct proc *p = curproc;
3077
3078 /* If we don't have a valid curproc use proc0 */
3079 /* Perhaps we should just use kernel_pmap instead */
3080 if (p == NULL)
3081 p = &proc0;
3082 #ifdef DIAGNOSTIC
3083 /*
3084 * The pmap should always be valid for the process so
3085 * panic if it is not.
3086 */
3087 if (!p->p_vmspace || !p->p_vmspace->vm_map.pmap) {
3088 printf("pmap_pte: va=%08lx p=%p vm=%p\n",
3089 va, p, p->p_vmspace);
3090 console_debugger();
3091 }
3092 /*
3093 * The pmap for the current process should be mapped. If it
3094 * is not then we have a problem.
3095 */
3096 if (p->p_vmspace->vm_map.pmap->pm_pptpt !=
3097 (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE
3098 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) +
3099 (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & PG_FRAME)) {
3100 printf("pmap pagetable = P%08lx current = P%08x ",
3101 pmap->pm_pptpt, (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE
3102 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) +
3103 (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) &
3104 PG_FRAME));
3105 printf("pptpt=%lx\n", p->p_vmspace->vm_map.pmap->pm_pptpt);
3106 panic("pmap_pte: current and pmap mismatch\n");
3107 }
3108 #endif
3109
3110 ptp = (pt_entry_t *)ALT_PAGE_TBLS_BASE;
3111 pmap_map_in_l1(p->p_vmspace->vm_map.pmap, ALT_PAGE_TBLS_BASE,
3112 pmap->pm_pptpt, FALSE);
3113 cpu_tlb_flushD();
3114 }
3115 PDEBUG(10, printf("page tables base = %p offset=%lx\n", ptp,
3116 ((va >> (PGSHIFT-2)) & ~3)));
3117 result = (pt_entry_t *)((char *)ptp + ((va >> (PGSHIFT-2)) & ~3));
3118 return(result);
3119 }
3120
3121 /*
3122 * Routine: pmap_extract
3123 * Function:
3124 * Extract the physical page address associated
3125 * with the given map/virtual_address pair.
3126 */
3127 boolean_t
3128 pmap_extract(pmap, va, pap)
3129 struct pmap *pmap;
3130 vaddr_t va;
3131 paddr_t *pap;
3132 {
3133 pt_entry_t *pte, *ptes;
3134 paddr_t pa;
3135
3136 PDEBUG(5, printf("pmap_extract: pmap=%p, va=V%08lx\n", pmap, va));
3137
3138 /*
3139 * Get the pte for this virtual address.
3140 */
3141 ptes = pmap_map_ptes(pmap);
3142 pte = &ptes[arm_byte_to_page(va)];
3143
3144 /*
3145 * If there is no pte then there is no page table etc.
3146 * Is the pte valid ? If not then no paged is actually mapped here
3147 */
3148 if (!pmap_pde_v(pmap_pde(pmap, va)) || !pmap_pte_v(pte)){
3149 pmap_unmap_ptes(pmap);
3150 return (FALSE);
3151 }
3152
3153 /* Return the physical address depending on the PTE type */
3154 /* XXX What about L1 section mappings ? */
3155 if ((*(pte) & L2_MASK) == L2_LPAGE) {
3156 /* Extract the physical address from the pte */
3157 pa = (*(pte)) & ~(L2_LPAGE_SIZE - 1);
3158
3159 PDEBUG(5, printf("pmap_extract: LPAGE pa = P%08lx\n",
3160 (pa | (va & (L2_LPAGE_SIZE - 1)))));
3161
3162 if (pap != NULL)
3163 *pap = pa | (va & (L2_LPAGE_SIZE - 1));
3164 } else {
3165 /* Extract the physical address from the pte */
3166 pa = pmap_pte_pa(pte);
3167
3168 PDEBUG(5, printf("pmap_extract: SPAGE pa = P%08lx\n",
3169 (pa | (va & ~PG_FRAME))));
3170
3171 if (pap != NULL)
3172 *pap = pa | (va & ~PG_FRAME);
3173 }
3174 pmap_unmap_ptes(pmap);
3175 return (TRUE);
3176 }
3177
3178
3179 /*
3180 * Copy the range specified by src_addr/len from the source map to the
3181 * range dst_addr/len in the destination map.
3182 *
3183 * This routine is only advisory and need not do anything.
3184 */
3185
3186 void
3187 pmap_copy(dst_pmap, src_pmap, dst_addr, len, src_addr)
3188 struct pmap *dst_pmap;
3189 struct pmap *src_pmap;
3190 vaddr_t dst_addr;
3191 vsize_t len;
3192 vaddr_t src_addr;
3193 {
3194 PDEBUG(0, printf("pmap_copy(%p, %p, %lx, %lx, %lx)\n",
3195 dst_pmap, src_pmap, dst_addr, len, src_addr));
3196 }
3197
3198 #if defined(PMAP_DEBUG)
3199 void
3200 pmap_dump_pvlist(phys, m)
3201 vaddr_t phys;
3202 char *m;
3203 {
3204 struct pv_head *pvh;
3205 struct pv_entry *pv;
3206 int bank, off;
3207
3208 if ((bank = vm_physseg_find(atop(phys), &off)) == -1) {
3209 printf("INVALID PA\n");
3210 return;
3211 }
3212 pvh = &vm_physmem[bank].pmseg.pvhead[off];
3213 simple_lock(&pvh->pvh_lock);
3214 printf("%s %08lx:", m, phys);
3215 if (pvh->pvh_list == NULL) {
3216 printf(" no mappings\n");
3217 return;
3218 }
3219
3220 for (pv = pvh->pvh_list; pv; pv = pv->pv_next)
3221 printf(" pmap %p va %08lx flags %08x", pv->pv_pmap,
3222 pv->pv_va, pv->pv_flags);
3223
3224 printf("\n");
3225 simple_unlock(&pvh->pvh_lock);
3226 }
3227
3228 #endif /* PMAP_DEBUG */
3229
3230 __inline static boolean_t
3231 pmap_testbit(pa, setbits)
3232 paddr_t pa;
3233 unsigned int setbits;
3234 {
3235 int bank, off;
3236
3237 PDEBUG(1, printf("pmap_testbit: pa=%08lx set=%08x\n", pa, setbits));
3238
3239 if ((bank = vm_physseg_find(atop(pa), &off)) == -1)
3240 return(FALSE);
3241
3242 /*
3243 * Check saved info only
3244 */
3245 if (vm_physmem[bank].pmseg.attrs[off] & setbits) {
3246 PDEBUG(0, printf("pmap_attributes = %02x\n",
3247 vm_physmem[bank].pmseg.attrs[off]));
3248 return(TRUE);
3249 }
3250
3251 return(FALSE);
3252 }
3253
3254 static pt_entry_t *
3255 pmap_map_ptes(struct pmap *pmap)
3256 {
3257 struct proc *p;
3258
3259 /* the kernel's pmap is always accessible */
3260 if (pmap == pmap_kernel()) {
3261 return (pt_entry_t *)PROCESS_PAGE_TBLS_BASE ;
3262 }
3263
3264 if (pmap_is_curpmap(pmap)) {
3265 simple_lock(&pmap->pm_obj.vmobjlock);
3266 return (pt_entry_t *)PROCESS_PAGE_TBLS_BASE;
3267 }
3268
3269 p = curproc;
3270
3271 if (p == NULL)
3272 p = &proc0;
3273
3274 /* need to lock both curpmap and pmap: use ordered locking */
3275 if ((unsigned) pmap < (unsigned) curproc->p_vmspace->vm_map.pmap) {
3276 simple_lock(&pmap->pm_obj.vmobjlock);
3277 simple_lock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3278 } else {
3279 simple_lock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3280 simple_lock(&pmap->pm_obj.vmobjlock);
3281 }
3282
3283 pmap_map_in_l1(p->p_vmspace->vm_map.pmap, ALT_PAGE_TBLS_BASE,
3284 pmap->pm_pptpt, FALSE);
3285 cpu_tlb_flushD();
3286 return (pt_entry_t *)ALT_PAGE_TBLS_BASE;
3287 }
3288
3289 /*
3290 * pmap_unmap_ptes: unlock the PTE mapping of "pmap"
3291 */
3292
3293 static void
3294 pmap_unmap_ptes(pmap)
3295 struct pmap *pmap;
3296 {
3297 if (pmap == pmap_kernel()) {
3298 return;
3299 }
3300 if (pmap_is_curpmap(pmap)) {
3301 simple_unlock(&pmap->pm_obj.vmobjlock);
3302 } else {
3303 simple_unlock(&pmap->pm_obj.vmobjlock);
3304 simple_unlock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3305 }
3306 }
3307
3308 /*
3309 * Modify pte bits for all ptes corresponding to the given physical address.
3310 * We use `maskbits' rather than `clearbits' because we're always passing
3311 * constants and the latter would require an extra inversion at run-time.
3312 */
3313
3314 static void
3315 pmap_clearbit(pa, maskbits)
3316 paddr_t pa;
3317 unsigned int maskbits;
3318 {
3319 struct pv_entry *pv;
3320 struct pv_head *pvh;
3321 pt_entry_t *pte;
3322 vaddr_t va;
3323 int bank, off, tlbentry;
3324
3325 PDEBUG(1, printf("pmap_clearbit: pa=%08lx mask=%08x\n",
3326 pa, maskbits));
3327
3328 tlbentry = 0;
3329
3330 if ((bank = vm_physseg_find(atop(pa), &off)) == -1)
3331 return;
3332 PMAP_HEAD_TO_MAP_LOCK();
3333 pvh = &vm_physmem[bank].pmseg.pvhead[off];
3334 simple_lock(&pvh->pvh_lock);
3335
3336 /*
3337 * Clear saved attributes (modify, reference)
3338 */
3339 vm_physmem[bank].pmseg.attrs[off] &= ~maskbits;
3340
3341 if (pvh->pvh_list == NULL) {
3342 simple_unlock(&pvh->pvh_lock);
3343 PMAP_HEAD_TO_MAP_UNLOCK();
3344 return;
3345 }
3346
3347 /*
3348 * Loop over all current mappings setting/clearing as appropos
3349 */
3350 for (pv = pvh->pvh_list; pv; pv = pv->pv_next) {
3351 va = pv->pv_va;
3352 pv->pv_flags &= ~maskbits;
3353 pte = pmap_pte(pv->pv_pmap, va);
3354 KASSERT(pte != NULL);
3355 if (maskbits & (PT_Wr|PT_M))
3356 {
3357 if ((pv->pv_flags & PT_NC))
3358 {
3359 /*
3360 * entry is not cacheable, so reenable the cache,
3361 * nothing to flush
3362 */
3363 *pte |= pte_cache_mode;
3364 pv->pv_flags &= ~PT_NC;
3365 } else {
3366 /*
3367 * entry is cacheable check if pmap is current if it
3368 * is flush it, otherwise it won't be in the cache
3369 */
3370 if (pmap_is_curpmap(pv->pv_pmap))
3371 {
3372 /* entry is in current pmap purge it */
3373 cpu_cache_purgeID_rng(pv->pv_va, NBPG);
3374 }
3375 }
3376
3377 /* make the pte read only */
3378 *pte &= ~PT_AP(AP_W);
3379
3380 if (pmap_is_curpmap(pv->pv_pmap))
3381 /*
3382 * if we had cacheable pte's we'd clean the pte out to
3383 * memory here
3384 */
3385 /*
3386 * flush tlb entry as it's in the current pmap
3387 */
3388 cpu_tlb_flushID_SE(pv->pv_va);
3389
3390 }
3391 if (maskbits & PT_H)
3392 *pte = (*pte & ~L2_MASK) | L2_INVAL;
3393 }
3394 simple_unlock(&pvh->pvh_lock);
3395 PMAP_HEAD_TO_MAP_UNLOCK();
3396 }
3397
3398
3399 boolean_t
3400 pmap_clear_modify(pg)
3401 struct vm_page *pg;
3402 {
3403 paddr_t pa = VM_PAGE_TO_PHYS(pg);
3404 boolean_t rv;
3405
3406 PDEBUG(0, printf("pmap_clear_modify pa=%08lx\n", pa));
3407 rv = pmap_testbit(pa, PT_M);
3408 pmap_clearbit(pa, PT_M);
3409 return rv;
3410 }
3411
3412
3413 boolean_t
3414 pmap_clear_reference(pg)
3415 struct vm_page *pg;
3416 {
3417 paddr_t pa = VM_PAGE_TO_PHYS(pg);
3418 boolean_t rv;
3419
3420 PDEBUG(0, printf("pmap_clear_reference pa=%08lx\n", pa));
3421 rv = pmap_testbit(pa, PT_H);
3422 pmap_clearbit(pa, PT_H);
3423 return rv;
3424 }
3425
3426
3427 void
3428 pmap_copy_on_write(pa)
3429 paddr_t pa;
3430 {
3431 PDEBUG(0, printf("pmap_copy_on_write pa=%08lx\n", pa));
3432 pmap_clearbit(pa, PT_Wr);
3433 }
3434
3435
3436 boolean_t
3437 pmap_is_modified(pg)
3438 struct vm_page *pg;
3439 {
3440 paddr_t pa = VM_PAGE_TO_PHYS(pg);
3441 boolean_t result;
3442
3443 result = pmap_testbit(pa, PT_M);
3444 PDEBUG(1, printf("pmap_is_modified pa=%08lx %x\n", pa, result));
3445 return (result);
3446 }
3447
3448
3449 boolean_t
3450 pmap_is_referenced(pg)
3451 struct vm_page *pg;
3452 {
3453 paddr_t pa = VM_PAGE_TO_PHYS(pg);
3454 boolean_t result;
3455
3456 result = pmap_testbit(pa, PT_H);
3457 PDEBUG(0, printf("pmap_is_referenced pa=%08lx %x\n", pa, result));
3458 return (result);
3459 }
3460
3461
3462 int
3463 pmap_modified_emulation(pmap, va)
3464 struct pmap *pmap;
3465 vaddr_t va;
3466 {
3467 pt_entry_t *pte;
3468 paddr_t pa;
3469 int bank, off;
3470 struct pv_head *pvh;
3471 u_int flags;
3472
3473 PDEBUG(2, printf("pmap_modified_emulation\n"));
3474
3475 /* Get the pte */
3476 pte = pmap_pte(pmap, va);
3477 if (!pte) {
3478 PDEBUG(2, printf("no pte\n"));
3479 return(0);
3480 }
3481
3482 PDEBUG(1, printf("*pte=%08x\n", *pte));
3483
3484 /* Check for a zero pte */
3485 if (*pte == 0)
3486 return(0);
3487
3488 /* This can happen if user code tries to access kernel memory. */
3489 if ((*pte & PT_AP(AP_W)) != 0)
3490 return (0);
3491
3492 /* Extract the physical address of the page */
3493 pa = pmap_pte_pa(pte);
3494 if ((bank = vm_physseg_find(atop(pa), &off)) == -1)
3495 return(0);
3496
3497 PMAP_HEAD_TO_MAP_LOCK();
3498 /* Get the current flags for this page. */
3499 pvh = &vm_physmem[bank].pmseg.pvhead[off];
3500 /* XXX: needed if we hold head->map lock? */
3501 simple_lock(&pvh->pvh_lock);
3502
3503 flags = pmap_modify_pv(pmap, va, pvh, 0, 0);
3504 PDEBUG(2, printf("pmap_modified_emulation: flags = %08x\n", flags));
3505
3506 /*
3507 * Do the flags say this page is writable ? If not then it is a
3508 * genuine write fault. If yes then the write fault is our fault
3509 * as we did not reflect the write access in the PTE. Now we know
3510 * a write has occurred we can correct this and also set the
3511 * modified bit
3512 */
3513 if (~flags & PT_Wr) {
3514 simple_unlock(&pvh->pvh_lock);
3515 PMAP_HEAD_TO_MAP_UNLOCK();
3516 return(0);
3517 }
3518
3519 PDEBUG(0, printf("pmap_modified_emulation: Got a hit va=%08lx, pte = %p (%08x)\n",
3520 va, pte, *pte));
3521 vm_physmem[bank].pmseg.attrs[off] |= PT_H | PT_M;
3522 *pte = (*pte & ~L2_MASK) | L2_SPAGE | PT_AP(AP_W);
3523 PDEBUG(0, printf("->(%08x)\n", *pte));
3524
3525 simple_unlock(&pvh->pvh_lock);
3526 PMAP_HEAD_TO_MAP_UNLOCK();
3527 /* Return, indicating the problem has been dealt with */
3528 cpu_tlb_flushID_SE(va);
3529 return(1);
3530 }
3531
3532
3533 int
3534 pmap_handled_emulation(pmap, va)
3535 struct pmap *pmap;
3536 vaddr_t va;
3537 {
3538 pt_entry_t *pte;
3539 paddr_t pa;
3540 int bank, off;
3541
3542 PDEBUG(2, printf("pmap_handled_emulation\n"));
3543
3544 /* Get the pte */
3545 pte = pmap_pte(pmap, va);
3546 if (!pte) {
3547 PDEBUG(2, printf("no pte\n"));
3548 return(0);
3549 }
3550
3551 PDEBUG(1, printf("*pte=%08x\n", *pte));
3552
3553 /* Check for a zero pte */
3554 if (*pte == 0)
3555 return(0);
3556
3557 /* This can happen if user code tries to access kernel memory. */
3558 if ((*pte & L2_MASK) != L2_INVAL)
3559 return (0);
3560
3561 /* Extract the physical address of the page */
3562 pa = pmap_pte_pa(pte);
3563 if ((bank = vm_physseg_find(atop(pa), &off)) == -1)
3564 return(0);
3565
3566 /*
3567 * Ok we just enable the pte and mark the attibs as handled
3568 */
3569 PDEBUG(0, printf("pmap_handled_emulation: Got a hit va=%08lx pte = %p (%08x)\n",
3570 va, pte, *pte));
3571 vm_physmem[bank].pmseg.attrs[off] |= PT_H;
3572 *pte = (*pte & ~L2_MASK) | L2_SPAGE;
3573 PDEBUG(0, printf("->(%08x)\n", *pte));
3574
3575 /* Return, indicating the problem has been dealt with */
3576 cpu_tlb_flushID_SE(va);
3577 return(1);
3578 }
3579
3580
3581
3582
3583 /*
3584 * pmap_collect: free resources held by a pmap
3585 *
3586 * => optional function.
3587 * => called when a process is swapped out to free memory.
3588 */
3589
3590 void
3591 pmap_collect(pmap)
3592 struct pmap *pmap;
3593 {
3594 }
3595
3596 /*
3597 * Routine: pmap_procwr
3598 *
3599 * Function:
3600 * Synchronize caches corresponding to [addr, addr+len) in p.
3601 *
3602 */
3603 void
3604 pmap_procwr(p, va, len)
3605 struct proc *p;
3606 vaddr_t va;
3607 int len;
3608 {
3609 /* We only need to do anything if it is the current process. */
3610 if (p == curproc)
3611 cpu_cache_syncI_rng(va, len);
3612 }
3613 /*
3614 * PTP functions
3615 */
3616
3617 /*
3618 * pmap_steal_ptp: Steal a PTP from somewhere else.
3619 *
3620 * This is just a placeholder, for now we never steal.
3621 */
3622
3623 static struct vm_page *
3624 pmap_steal_ptp(struct pmap *pmap, vaddr_t va)
3625 {
3626 return (NULL);
3627 }
3628
3629 /*
3630 * pmap_get_ptp: get a PTP (if there isn't one, allocate a new one)
3631 *
3632 * => pmap should NOT be pmap_kernel()
3633 * => pmap should be locked
3634 */
3635
3636 static struct vm_page *
3637 pmap_get_ptp(struct pmap *pmap, vaddr_t va, boolean_t just_try)
3638 {
3639 struct vm_page *ptp;
3640
3641 if (pmap_pde_v(pmap_pde(pmap, va))) {
3642
3643 /* valid... check hint (saves us a PA->PG lookup) */
3644 #if 0
3645 if (pmap->pm_ptphint &&
3646 ((unsigned)pmap_pde(pmap, va) & PG_FRAME) ==
3647 VM_PAGE_TO_PHYS(pmap->pm_ptphint))
3648 return (pmap->pm_ptphint);
3649 #endif
3650 ptp = uvm_pagelookup(&pmap->pm_obj, va);
3651 #ifdef DIAGNOSTIC
3652 if (ptp == NULL)
3653 panic("pmap_get_ptp: unmanaged user PTP");
3654 #endif
3655 // pmap->pm_ptphint = ptp;
3656 return(ptp);
3657 }
3658
3659 /* allocate a new PTP (updates ptphint) */
3660 return(pmap_alloc_ptp(pmap, va, just_try));
3661 }
3662
3663 /*
3664 * pmap_alloc_ptp: allocate a PTP for a PMAP
3665 *
3666 * => pmap should already be locked by caller
3667 * => we use the ptp's wire_count to count the number of active mappings
3668 * in the PTP (we start it at one to prevent any chance this PTP
3669 * will ever leak onto the active/inactive queues)
3670 */
3671
3672 /*__inline */ static struct vm_page *
3673 pmap_alloc_ptp(struct pmap *pmap, vaddr_t va, boolean_t just_try)
3674 {
3675 struct vm_page *ptp;
3676
3677 ptp = uvm_pagealloc(&pmap->pm_obj, va, NULL,
3678 UVM_PGA_USERESERVE|UVM_PGA_ZERO);
3679 if (ptp == NULL) {
3680 if (just_try)
3681 return (NULL);
3682
3683 ptp = pmap_steal_ptp(pmap, va);
3684
3685 if (ptp == NULL)
3686 return (NULL);
3687 /* Stole a page, zero it. */
3688 pmap_zero_page(VM_PAGE_TO_PHYS(ptp));
3689 }
3690
3691 /* got one! */
3692 ptp->flags &= ~PG_BUSY; /* never busy */
3693 ptp->wire_count = 1; /* no mappings yet */
3694 pmap_map_in_l1(pmap, va, VM_PAGE_TO_PHYS(ptp), TRUE);
3695 pmap->pm_stats.resident_count++; /* count PTP as resident */
3696 // pmap->pm_ptphint = ptp;
3697 return (ptp);
3698 }
3699
3700 /* End of pmap.c */
3701