pmap.c revision 1.89 1 /* $NetBSD: pmap.c,v 1.89 2002/04/10 01:30:42 thorpej Exp $ */
2
3 /*
4 * Copyright (c) 2002 Wasabi Systems, Inc.
5 * Copyright (c) 2001 Richard Earnshaw
6 * Copyright (c) 2001 Christopher Gilbert
7 * All rights reserved.
8 *
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
14 * 3. The name of the company nor the name of the author may be used to
15 * endorse or promote products derived from this software without specific
16 * prior written permission.
17 *
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
19 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
20 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
21 * IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
22 * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
23 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
24 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 * SUCH DAMAGE.
29 */
30
31 /*-
32 * Copyright (c) 1999 The NetBSD Foundation, Inc.
33 * All rights reserved.
34 *
35 * This code is derived from software contributed to The NetBSD Foundation
36 * by Charles M. Hannum.
37 *
38 * Redistribution and use in source and binary forms, with or without
39 * modification, are permitted provided that the following conditions
40 * are met:
41 * 1. Redistributions of source code must retain the above copyright
42 * notice, this list of conditions and the following disclaimer.
43 * 2. Redistributions in binary form must reproduce the above copyright
44 * notice, this list of conditions and the following disclaimer in the
45 * documentation and/or other materials provided with the distribution.
46 * 3. All advertising materials mentioning features or use of this software
47 * must display the following acknowledgement:
48 * This product includes software developed by the NetBSD
49 * Foundation, Inc. and its contributors.
50 * 4. Neither the name of The NetBSD Foundation nor the names of its
51 * contributors may be used to endorse or promote products derived
52 * from this software without specific prior written permission.
53 *
54 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
55 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
56 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
57 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
58 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
59 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
60 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
61 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
62 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
63 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
64 * POSSIBILITY OF SUCH DAMAGE.
65 */
66
67 /*
68 * Copyright (c) 1994-1998 Mark Brinicombe.
69 * Copyright (c) 1994 Brini.
70 * All rights reserved.
71 *
72 * This code is derived from software written for Brini by Mark Brinicombe
73 *
74 * Redistribution and use in source and binary forms, with or without
75 * modification, are permitted provided that the following conditions
76 * are met:
77 * 1. Redistributions of source code must retain the above copyright
78 * notice, this list of conditions and the following disclaimer.
79 * 2. Redistributions in binary form must reproduce the above copyright
80 * notice, this list of conditions and the following disclaimer in the
81 * documentation and/or other materials provided with the distribution.
82 * 3. All advertising materials mentioning features or use of this software
83 * must display the following acknowledgement:
84 * This product includes software developed by Mark Brinicombe.
85 * 4. The name of the author may not be used to endorse or promote products
86 * derived from this software without specific prior written permission.
87 *
88 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
89 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
90 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
91 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
92 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
93 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
94 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
95 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
96 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
97 *
98 * RiscBSD kernel project
99 *
100 * pmap.c
101 *
102 * Machine dependant vm stuff
103 *
104 * Created : 20/09/94
105 */
106
107 /*
108 * Performance improvements, UVM changes, overhauls and part-rewrites
109 * were contributed by Neil A. Carson <neil (at) causality.com>.
110 */
111
112 /*
113 * The dram block info is currently referenced from the bootconfig.
114 * This should be placed in a separate structure.
115 */
116
117 /*
118 * Special compilation symbols
119 * PMAP_DEBUG - Build in pmap_debug_level code
120 */
121
122 /* Include header files */
123
124 #include "opt_pmap_debug.h"
125 #include "opt_ddb.h"
126
127 #include <sys/types.h>
128 #include <sys/param.h>
129 #include <sys/kernel.h>
130 #include <sys/systm.h>
131 #include <sys/proc.h>
132 #include <sys/malloc.h>
133 #include <sys/user.h>
134 #include <sys/pool.h>
135 #include <sys/cdefs.h>
136
137 #include <uvm/uvm.h>
138
139 #include <machine/bootconfig.h>
140 #include <machine/bus.h>
141 #include <machine/pmap.h>
142 #include <machine/pcb.h>
143 #include <machine/param.h>
144 #include <arm/arm32/katelib.h>
145
146 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.89 2002/04/10 01:30:42 thorpej Exp $");
147 #ifdef PMAP_DEBUG
148 #define PDEBUG(_lev_,_stat_) \
149 if (pmap_debug_level >= (_lev_)) \
150 ((_stat_))
151 int pmap_debug_level = -2;
152 void pmap_dump_pvlist(vaddr_t phys, char *m);
153
154 /*
155 * for switching to potentially finer grained debugging
156 */
157 #define PDB_FOLLOW 0x0001
158 #define PDB_INIT 0x0002
159 #define PDB_ENTER 0x0004
160 #define PDB_REMOVE 0x0008
161 #define PDB_CREATE 0x0010
162 #define PDB_PTPAGE 0x0020
163 #define PDB_GROWKERN 0x0040
164 #define PDB_BITS 0x0080
165 #define PDB_COLLECT 0x0100
166 #define PDB_PROTECT 0x0200
167 #define PDB_MAP_L1 0x0400
168 #define PDB_BOOTSTRAP 0x1000
169 #define PDB_PARANOIA 0x2000
170 #define PDB_WIRING 0x4000
171 #define PDB_PVDUMP 0x8000
172
173 int debugmap = 0;
174 int pmapdebug = PDB_PARANOIA | PDB_FOLLOW;
175 #define NPDEBUG(_lev_,_stat_) \
176 if (pmapdebug & (_lev_)) \
177 ((_stat_))
178
179 #else /* PMAP_DEBUG */
180 #define PDEBUG(_lev_,_stat_) /* Nothing */
181 #define NPDEBUG(_lev_,_stat_) /* Nothing */
182 #endif /* PMAP_DEBUG */
183
184 struct pmap kernel_pmap_store;
185
186 /*
187 * linked list of all non-kernel pmaps
188 */
189
190 static LIST_HEAD(, pmap) pmaps;
191
192 /*
193 * pool that pmap structures are allocated from
194 */
195
196 struct pool pmap_pmap_pool;
197
198 static pt_entry_t *csrc_pte, *cdst_pte;
199 static vaddr_t csrcp, cdstp;
200
201 char *memhook;
202 extern caddr_t msgbufaddr;
203
204 boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */
205 /*
206 * locking data structures
207 */
208
209 static struct lock pmap_main_lock;
210 static struct simplelock pvalloc_lock;
211 static struct simplelock pmaps_lock;
212 #ifdef LOCKDEBUG
213 #define PMAP_MAP_TO_HEAD_LOCK() \
214 (void) spinlockmgr(&pmap_main_lock, LK_SHARED, NULL)
215 #define PMAP_MAP_TO_HEAD_UNLOCK() \
216 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
217
218 #define PMAP_HEAD_TO_MAP_LOCK() \
219 (void) spinlockmgr(&pmap_main_lock, LK_EXCLUSIVE, NULL)
220 #define PMAP_HEAD_TO_MAP_UNLOCK() \
221 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
222 #else
223 #define PMAP_MAP_TO_HEAD_LOCK() /* nothing */
224 #define PMAP_MAP_TO_HEAD_UNLOCK() /* nothing */
225 #define PMAP_HEAD_TO_MAP_LOCK() /* nothing */
226 #define PMAP_HEAD_TO_MAP_UNLOCK() /* nothing */
227 #endif /* LOCKDEBUG */
228
229 /*
230 * pv_page management structures: locked by pvalloc_lock
231 */
232
233 TAILQ_HEAD(pv_pagelist, pv_page);
234 static struct pv_pagelist pv_freepages; /* list of pv_pages with free entrys */
235 static struct pv_pagelist pv_unusedpgs; /* list of unused pv_pages */
236 static int pv_nfpvents; /* # of free pv entries */
237 static struct pv_page *pv_initpage; /* bootstrap page from kernel_map */
238 static vaddr_t pv_cachedva; /* cached VA for later use */
239
240 #define PVE_LOWAT (PVE_PER_PVPAGE / 2) /* free pv_entry low water mark */
241 #define PVE_HIWAT (PVE_LOWAT + (PVE_PER_PVPAGE * 2))
242 /* high water mark */
243
244 /*
245 * local prototypes
246 */
247
248 static struct pv_entry *pmap_add_pvpage __P((struct pv_page *, boolean_t));
249 static struct pv_entry *pmap_alloc_pv __P((struct pmap *, int)); /* see codes below */
250 #define ALLOCPV_NEED 0 /* need PV now */
251 #define ALLOCPV_TRY 1 /* just try to allocate, don't steal */
252 #define ALLOCPV_NONEED 2 /* don't need PV, just growing cache */
253 static struct pv_entry *pmap_alloc_pvpage __P((struct pmap *, int));
254 static void pmap_enter_pv __P((struct vm_page *,
255 struct pv_entry *, struct pmap *,
256 vaddr_t, struct vm_page *, int));
257 static void pmap_free_pv __P((struct pmap *, struct pv_entry *));
258 static void pmap_free_pvs __P((struct pmap *, struct pv_entry *));
259 static void pmap_free_pv_doit __P((struct pv_entry *));
260 static void pmap_free_pvpage __P((void));
261 static boolean_t pmap_is_curpmap __P((struct pmap *));
262 static struct pv_entry *pmap_remove_pv __P((struct vm_page *, struct pmap *,
263 vaddr_t));
264 #define PMAP_REMOVE_ALL 0 /* remove all mappings */
265 #define PMAP_REMOVE_SKIPWIRED 1 /* skip wired mappings */
266
267 static u_int pmap_modify_pv __P((struct pmap *, vaddr_t, struct vm_page *,
268 u_int, u_int));
269
270 /*
271 * Structure that describes and L1 table.
272 */
273 struct l1pt {
274 SIMPLEQ_ENTRY(l1pt) pt_queue; /* Queue pointers */
275 struct pglist pt_plist; /* Allocated page list */
276 vaddr_t pt_va; /* Allocated virtual address */
277 int pt_flags; /* Flags */
278 };
279 #define PTFLAG_STATIC 0x01 /* Statically allocated */
280 #define PTFLAG_KPT 0x02 /* Kernel pt's are mapped */
281 #define PTFLAG_CLEAN 0x04 /* L1 is clean */
282
283 static void pmap_free_l1pt __P((struct l1pt *));
284 static int pmap_allocpagedir __P((struct pmap *));
285 static int pmap_clean_page __P((struct pv_entry *, boolean_t));
286 static void pmap_remove_all __P((struct vm_page *));
287
288 static struct vm_page *pmap_alloc_ptp __P((struct pmap *, vaddr_t));
289 static struct vm_page *pmap_get_ptp __P((struct pmap *, vaddr_t));
290 __inline static void pmap_clearbit __P((struct vm_page *, unsigned int));
291
292 extern paddr_t physical_start;
293 extern paddr_t physical_freestart;
294 extern paddr_t physical_end;
295 extern paddr_t physical_freeend;
296 extern unsigned int free_pages;
297 extern int max_processes;
298
299 vaddr_t virtual_avail;
300 vaddr_t virtual_end;
301 vaddr_t pmap_curmaxkvaddr;
302
303 vaddr_t avail_start;
304 vaddr_t avail_end;
305
306 extern pv_addr_t systempage;
307
308 /* Variables used by the L1 page table queue code */
309 SIMPLEQ_HEAD(l1pt_queue, l1pt);
310 static struct l1pt_queue l1pt_static_queue; /* head of our static l1 queue */
311 static int l1pt_static_queue_count; /* items in the static l1 queue */
312 static int l1pt_static_create_count; /* static l1 items created */
313 static struct l1pt_queue l1pt_queue; /* head of our l1 queue */
314 static int l1pt_queue_count; /* items in the l1 queue */
315 static int l1pt_create_count; /* stat - L1's create count */
316 static int l1pt_reuse_count; /* stat - L1's reused count */
317
318 /* Local function prototypes (not used outside this file) */
319 void pmap_pinit __P((struct pmap *));
320 void pmap_freepagedir __P((struct pmap *));
321
322 /* Other function prototypes */
323 extern void bzero_page __P((vaddr_t));
324 extern void bcopy_page __P((vaddr_t, vaddr_t));
325
326 struct l1pt *pmap_alloc_l1pt __P((void));
327 static __inline void pmap_map_in_l1 __P((struct pmap *pmap, vaddr_t va,
328 vaddr_t l2pa, boolean_t));
329
330 static pt_entry_t *pmap_map_ptes __P((struct pmap *));
331 static void pmap_unmap_ptes __P((struct pmap *));
332
333 __inline static void pmap_vac_me_harder __P((struct pmap *, struct vm_page *,
334 pt_entry_t *, boolean_t));
335 static void pmap_vac_me_kpmap __P((struct pmap *, struct vm_page *,
336 pt_entry_t *, boolean_t));
337 static void pmap_vac_me_user __P((struct pmap *, struct vm_page *,
338 pt_entry_t *, boolean_t));
339
340 /*
341 * real definition of pv_entry.
342 */
343
344 struct pv_entry {
345 struct pv_entry *pv_next; /* next pv_entry */
346 struct pmap *pv_pmap; /* pmap where mapping lies */
347 vaddr_t pv_va; /* virtual address for mapping */
348 int pv_flags; /* flags */
349 struct vm_page *pv_ptp; /* vm_page for the ptp */
350 };
351
352 /*
353 * pv_entrys are dynamically allocated in chunks from a single page.
354 * we keep track of how many pv_entrys are in use for each page and
355 * we can free pv_entry pages if needed. there is one lock for the
356 * entire allocation system.
357 */
358
359 struct pv_page_info {
360 TAILQ_ENTRY(pv_page) pvpi_list;
361 struct pv_entry *pvpi_pvfree;
362 int pvpi_nfree;
363 };
364
365 /*
366 * number of pv_entry's in a pv_page
367 * (note: won't work on systems where NPBG isn't a constant)
368 */
369
370 #define PVE_PER_PVPAGE ((NBPG - sizeof(struct pv_page_info)) / \
371 sizeof(struct pv_entry))
372
373 /*
374 * a pv_page: where pv_entrys are allocated from
375 */
376
377 struct pv_page {
378 struct pv_page_info pvinfo;
379 struct pv_entry pvents[PVE_PER_PVPAGE];
380 };
381
382 #ifdef MYCROFT_HACK
383 int mycroft_hack = 0;
384 #endif
385
386 /* Function to set the debug level of the pmap code */
387
388 #ifdef PMAP_DEBUG
389 void
390 pmap_debug(int level)
391 {
392 pmap_debug_level = level;
393 printf("pmap_debug: level=%d\n", pmap_debug_level);
394 }
395 #endif /* PMAP_DEBUG */
396
397 __inline static boolean_t
398 pmap_is_curpmap(struct pmap *pmap)
399 {
400
401 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap) ||
402 pmap == pmap_kernel())
403 return (TRUE);
404
405 return (FALSE);
406 }
407
408 #include "isadma.h"
409
410 #if NISADMA > 0
411 /*
412 * Used to protect memory for ISA DMA bounce buffers. If, when loading
413 * pages into the system, memory intersects with any of these ranges,
414 * the intersecting memory will be loaded into a lower-priority free list.
415 */
416 bus_dma_segment_t *pmap_isa_dma_ranges;
417 int pmap_isa_dma_nranges;
418
419 /*
420 * Check if a memory range intersects with an ISA DMA range, and
421 * return the page-rounded intersection if it does. The intersection
422 * will be placed on a lower-priority free list.
423 */
424 static boolean_t
425 pmap_isa_dma_range_intersect(paddr_t pa, psize_t size, paddr_t *pap,
426 psize_t *sizep)
427 {
428 bus_dma_segment_t *ds;
429 int i;
430
431 if (pmap_isa_dma_ranges == NULL)
432 return (FALSE);
433
434 for (i = 0, ds = pmap_isa_dma_ranges;
435 i < pmap_isa_dma_nranges; i++, ds++) {
436 if (ds->ds_addr <= pa && pa < (ds->ds_addr + ds->ds_len)) {
437 /*
438 * Beginning of region intersects with this range.
439 */
440 *pap = trunc_page(pa);
441 *sizep = round_page(min(pa + size,
442 ds->ds_addr + ds->ds_len) - pa);
443 return (TRUE);
444 }
445 if (pa < ds->ds_addr && ds->ds_addr < (pa + size)) {
446 /*
447 * End of region intersects with this range.
448 */
449 *pap = trunc_page(ds->ds_addr);
450 *sizep = round_page(min((pa + size) - ds->ds_addr,
451 ds->ds_len));
452 return (TRUE);
453 }
454 }
455
456 /*
457 * No intersection found.
458 */
459 return (FALSE);
460 }
461 #endif /* NISADMA > 0 */
462
463 /*
464 * p v _ e n t r y f u n c t i o n s
465 */
466
467 /*
468 * pv_entry allocation functions:
469 * the main pv_entry allocation functions are:
470 * pmap_alloc_pv: allocate a pv_entry structure
471 * pmap_free_pv: free one pv_entry
472 * pmap_free_pvs: free a list of pv_entrys
473 *
474 * the rest are helper functions
475 */
476
477 /*
478 * pmap_alloc_pv: inline function to allocate a pv_entry structure
479 * => we lock pvalloc_lock
480 * => if we fail, we call out to pmap_alloc_pvpage
481 * => 3 modes:
482 * ALLOCPV_NEED = we really need a pv_entry, even if we have to steal it
483 * ALLOCPV_TRY = we want a pv_entry, but not enough to steal
484 * ALLOCPV_NONEED = we are trying to grow our free list, don't really need
485 * one now
486 *
487 * "try" is for optional functions like pmap_copy().
488 */
489
490 __inline static struct pv_entry *
491 pmap_alloc_pv(struct pmap *pmap, int mode)
492 {
493 struct pv_page *pvpage;
494 struct pv_entry *pv;
495
496 simple_lock(&pvalloc_lock);
497
498 pvpage = TAILQ_FIRST(&pv_freepages);
499
500 if (pvpage != NULL) {
501 pvpage->pvinfo.pvpi_nfree--;
502 if (pvpage->pvinfo.pvpi_nfree == 0) {
503 /* nothing left in this one? */
504 TAILQ_REMOVE(&pv_freepages, pvpage, pvinfo.pvpi_list);
505 }
506 pv = pvpage->pvinfo.pvpi_pvfree;
507 KASSERT(pv);
508 pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
509 pv_nfpvents--; /* took one from pool */
510 } else {
511 pv = NULL; /* need more of them */
512 }
513
514 /*
515 * if below low water mark or we didn't get a pv_entry we try and
516 * create more pv_entrys ...
517 */
518
519 if (pv_nfpvents < PVE_LOWAT || pv == NULL) {
520 if (pv == NULL)
521 pv = pmap_alloc_pvpage(pmap, (mode == ALLOCPV_TRY) ?
522 mode : ALLOCPV_NEED);
523 else
524 (void) pmap_alloc_pvpage(pmap, ALLOCPV_NONEED);
525 }
526
527 simple_unlock(&pvalloc_lock);
528 return(pv);
529 }
530
531 /*
532 * pmap_alloc_pvpage: maybe allocate a new pvpage
533 *
534 * if need_entry is false: try and allocate a new pv_page
535 * if need_entry is true: try and allocate a new pv_page and return a
536 * new pv_entry from it. if we are unable to allocate a pv_page
537 * we make a last ditch effort to steal a pv_page from some other
538 * mapping. if that fails, we panic...
539 *
540 * => we assume that the caller holds pvalloc_lock
541 */
542
543 static struct pv_entry *
544 pmap_alloc_pvpage(struct pmap *pmap, int mode)
545 {
546 struct vm_page *pg;
547 struct pv_page *pvpage;
548 struct pv_entry *pv;
549 int s;
550
551 /*
552 * if we need_entry and we've got unused pv_pages, allocate from there
553 */
554
555 pvpage = TAILQ_FIRST(&pv_unusedpgs);
556 if (mode != ALLOCPV_NONEED && pvpage != NULL) {
557
558 /* move it to pv_freepages list */
559 TAILQ_REMOVE(&pv_unusedpgs, pvpage, pvinfo.pvpi_list);
560 TAILQ_INSERT_HEAD(&pv_freepages, pvpage, pvinfo.pvpi_list);
561
562 /* allocate a pv_entry */
563 pvpage->pvinfo.pvpi_nfree--; /* can't go to zero */
564 pv = pvpage->pvinfo.pvpi_pvfree;
565 KASSERT(pv);
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
591 if (pg == NULL)
592 return (NULL);
593 pg->flags &= ~PG_BUSY; /* never busy */
594
595 /*
596 * add a mapping for our new pv_page and free its entrys (save one!)
597 *
598 * NOTE: If we are allocating a PV page for the kernel pmap, the
599 * pmap is already locked! (...but entering the mapping is safe...)
600 */
601
602 pmap_kenter_pa(pv_cachedva, VM_PAGE_TO_PHYS(pg),
603 VM_PROT_READ|VM_PROT_WRITE);
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(struct pv_page *pvp, boolean_t need_entry)
619 {
620 int tofree, lcv;
621
622 /* do we need to return one? */
623 tofree = (need_entry) ? PVE_PER_PVPAGE - 1 : PVE_PER_PVPAGE;
624
625 pvp->pvinfo.pvpi_pvfree = NULL;
626 pvp->pvinfo.pvpi_nfree = tofree;
627 for (lcv = 0 ; lcv < tofree ; lcv++) {
628 pvp->pvents[lcv].pv_next = pvp->pvinfo.pvpi_pvfree;
629 pvp->pvinfo.pvpi_pvfree = &pvp->pvents[lcv];
630 }
631 if (need_entry)
632 TAILQ_INSERT_TAIL(&pv_freepages, pvp, pvinfo.pvpi_list);
633 else
634 TAILQ_INSERT_TAIL(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
635 pv_nfpvents += tofree;
636 return((need_entry) ? &pvp->pvents[lcv] : NULL);
637 }
638
639 /*
640 * pmap_free_pv_doit: actually free a pv_entry
641 *
642 * => do not call this directly! instead use either
643 * 1. pmap_free_pv ==> free a single pv_entry
644 * 2. pmap_free_pvs => free a list of pv_entrys
645 * => we must be holding pvalloc_lock
646 */
647
648 __inline static void
649 pmap_free_pv_doit(struct pv_entry *pv)
650 {
651 struct pv_page *pvp;
652
653 pvp = (struct pv_page *) arm_trunc_page((vaddr_t)pv);
654 pv_nfpvents++;
655 pvp->pvinfo.pvpi_nfree++;
656
657 /* nfree == 1 => fully allocated page just became partly allocated */
658 if (pvp->pvinfo.pvpi_nfree == 1) {
659 TAILQ_INSERT_HEAD(&pv_freepages, pvp, pvinfo.pvpi_list);
660 }
661
662 /* free it */
663 pv->pv_next = pvp->pvinfo.pvpi_pvfree;
664 pvp->pvinfo.pvpi_pvfree = pv;
665
666 /*
667 * are all pv_page's pv_entry's free? move it to unused queue.
668 */
669
670 if (pvp->pvinfo.pvpi_nfree == PVE_PER_PVPAGE) {
671 TAILQ_REMOVE(&pv_freepages, pvp, pvinfo.pvpi_list);
672 TAILQ_INSERT_HEAD(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
673 }
674 }
675
676 /*
677 * pmap_free_pv: free a single pv_entry
678 *
679 * => we gain the pvalloc_lock
680 */
681
682 __inline static void
683 pmap_free_pv(struct pmap *pmap, struct pv_entry *pv)
684 {
685 simple_lock(&pvalloc_lock);
686 pmap_free_pv_doit(pv);
687
688 /*
689 * Can't free the PV page if the PV entries were associated with
690 * the kernel pmap; the pmap is already locked.
691 */
692 if (pv_nfpvents > PVE_HIWAT && TAILQ_FIRST(&pv_unusedpgs) != NULL &&
693 pmap != pmap_kernel())
694 pmap_free_pvpage();
695
696 simple_unlock(&pvalloc_lock);
697 }
698
699 /*
700 * pmap_free_pvs: free a list of pv_entrys
701 *
702 * => we gain the pvalloc_lock
703 */
704
705 __inline static void
706 pmap_free_pvs(struct pmap *pmap, struct pv_entry *pvs)
707 {
708 struct pv_entry *nextpv;
709
710 simple_lock(&pvalloc_lock);
711
712 for ( /* null */ ; pvs != NULL ; pvs = nextpv) {
713 nextpv = pvs->pv_next;
714 pmap_free_pv_doit(pvs);
715 }
716
717 /*
718 * Can't free the PV page if the PV entries were associated with
719 * the kernel pmap; the pmap is already locked.
720 */
721 if (pv_nfpvents > PVE_HIWAT && TAILQ_FIRST(&pv_unusedpgs) != NULL &&
722 pmap != pmap_kernel())
723 pmap_free_pvpage();
724
725 simple_unlock(&pvalloc_lock);
726 }
727
728
729 /*
730 * pmap_free_pvpage: try and free an unused pv_page structure
731 *
732 * => assume caller is holding the pvalloc_lock and that
733 * there is a page on the pv_unusedpgs list
734 * => if we can't get a lock on the kmem_map we try again later
735 */
736
737 static void
738 pmap_free_pvpage(void)
739 {
740 int s;
741 struct vm_map *map;
742 struct vm_map_entry *dead_entries;
743 struct pv_page *pvp;
744
745 s = splvm(); /* protect kmem_map */
746
747 pvp = TAILQ_FIRST(&pv_unusedpgs);
748
749 /*
750 * note: watch out for pv_initpage which is allocated out of
751 * kernel_map rather than kmem_map.
752 */
753 if (pvp == pv_initpage)
754 map = kernel_map;
755 else
756 map = kmem_map;
757 if (vm_map_lock_try(map)) {
758
759 /* remove pvp from pv_unusedpgs */
760 TAILQ_REMOVE(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
761
762 /* unmap the page */
763 dead_entries = NULL;
764 uvm_unmap_remove(map, (vaddr_t)pvp, ((vaddr_t)pvp) + PAGE_SIZE,
765 &dead_entries);
766 vm_map_unlock(map);
767
768 if (dead_entries != NULL)
769 uvm_unmap_detach(dead_entries, 0);
770
771 pv_nfpvents -= PVE_PER_PVPAGE; /* update free count */
772 }
773 if (pvp == pv_initpage)
774 /* no more initpage, we've freed it */
775 pv_initpage = NULL;
776
777 splx(s);
778 }
779
780 /*
781 * main pv_entry manipulation functions:
782 * pmap_enter_pv: enter a mapping onto a vm_page list
783 * pmap_remove_pv: remove a mappiing from a vm_page list
784 *
785 * NOTE: pmap_enter_pv expects to lock the pvh itself
786 * pmap_remove_pv expects te caller to lock the pvh before calling
787 */
788
789 /*
790 * pmap_enter_pv: enter a mapping onto a vm_page lst
791 *
792 * => caller should hold the proper lock on pmap_main_lock
793 * => caller should have pmap locked
794 * => we will gain the lock on the vm_page and allocate the new pv_entry
795 * => caller should adjust ptp's wire_count before calling
796 * => caller should not adjust pmap's wire_count
797 */
798
799 __inline static void
800 pmap_enter_pv(struct vm_page *pg, struct pv_entry *pve, struct pmap *pmap,
801 vaddr_t va, struct vm_page *ptp, int flags)
802 {
803 pve->pv_pmap = pmap;
804 pve->pv_va = va;
805 pve->pv_ptp = ptp; /* NULL for kernel pmap */
806 pve->pv_flags = flags;
807 simple_lock(&pg->mdpage.pvh_slock); /* lock vm_page */
808 pve->pv_next = pg->mdpage.pvh_list; /* add to ... */
809 pg->mdpage.pvh_list = pve; /* ... locked list */
810 simple_unlock(&pg->mdpage.pvh_slock); /* unlock, done! */
811 if (pve->pv_flags & PVF_WIRED)
812 ++pmap->pm_stats.wired_count;
813 }
814
815 /*
816 * pmap_remove_pv: try to remove a mapping from a pv_list
817 *
818 * => caller should hold proper lock on pmap_main_lock
819 * => pmap should be locked
820 * => caller should hold lock on vm_page [so that attrs can be adjusted]
821 * => caller should adjust ptp's wire_count and free PTP if needed
822 * => caller should NOT adjust pmap's wire_count
823 * => we return the removed pve
824 */
825
826 __inline static struct pv_entry *
827 pmap_remove_pv(struct vm_page *pg, struct pmap *pmap, vaddr_t va)
828 {
829 struct pv_entry *pve, **prevptr;
830
831 prevptr = &pg->mdpage.pvh_list; /* previous pv_entry pointer */
832 pve = *prevptr;
833 while (pve) {
834 if (pve->pv_pmap == pmap && pve->pv_va == va) { /* match? */
835 *prevptr = pve->pv_next; /* remove it! */
836 if (pve->pv_flags & PVF_WIRED)
837 --pmap->pm_stats.wired_count;
838 break;
839 }
840 prevptr = &pve->pv_next; /* previous pointer */
841 pve = pve->pv_next; /* advance */
842 }
843 return(pve); /* return removed pve */
844 }
845
846 /*
847 *
848 * pmap_modify_pv: Update pv flags
849 *
850 * => caller should hold lock on vm_page [so that attrs can be adjusted]
851 * => caller should NOT adjust pmap's wire_count
852 * => caller must call pmap_vac_me_harder() if writable status of a page
853 * may have changed.
854 * => we return the old flags
855 *
856 * Modify a physical-virtual mapping in the pv table
857 */
858
859 static /* __inline */ u_int
860 pmap_modify_pv(struct pmap *pmap, vaddr_t va, struct vm_page *pg,
861 u_int bic_mask, u_int eor_mask)
862 {
863 struct pv_entry *npv;
864 u_int flags, oflags;
865
866 /*
867 * There is at least one VA mapping this page.
868 */
869
870 for (npv = pg->mdpage.pvh_list; npv; npv = npv->pv_next) {
871 if (pmap == npv->pv_pmap && va == npv->pv_va) {
872 oflags = npv->pv_flags;
873 npv->pv_flags = flags =
874 ((oflags & ~bic_mask) ^ eor_mask);
875 if ((flags ^ oflags) & PVF_WIRED) {
876 if (flags & PVF_WIRED)
877 ++pmap->pm_stats.wired_count;
878 else
879 --pmap->pm_stats.wired_count;
880 }
881 return (oflags);
882 }
883 }
884 return (0);
885 }
886
887 /*
888 * Map the specified level 2 pagetable into the level 1 page table for
889 * the given pmap to cover a chunk of virtual address space starting from the
890 * address specified.
891 */
892 static __inline void
893 pmap_map_in_l1(struct pmap *pmap, vaddr_t va, paddr_t l2pa, boolean_t selfref)
894 {
895 vaddr_t ptva;
896
897 /* Calculate the index into the L1 page table. */
898 ptva = (va >> L1_S_SHIFT) & ~3;
899
900 /* Map page table into the L1. */
901 pmap->pm_pdir[ptva + 0] = L1_C_PROTO | (l2pa + 0x000);
902 pmap->pm_pdir[ptva + 1] = L1_C_PROTO | (l2pa + 0x400);
903 pmap->pm_pdir[ptva + 2] = L1_C_PROTO | (l2pa + 0x800);
904 pmap->pm_pdir[ptva + 3] = L1_C_PROTO | (l2pa + 0xc00);
905
906 /* Map the page table into the page table area. */
907 if (selfref)
908 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) = L2_S_PROTO | l2pa |
909 L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE);
910 }
911
912 #if 0
913 static __inline void
914 pmap_unmap_in_l1(struct pmap *pmap, vaddr_t va)
915 {
916 vaddr_t ptva;
917
918 /* Calculate the index into the L1 page table. */
919 ptva = (va >> L1_S_SHIFT) & ~3;
920
921 /* Unmap page table from the L1. */
922 pmap->pm_pdir[ptva + 0] = 0;
923 pmap->pm_pdir[ptva + 1] = 0;
924 pmap->pm_pdir[ptva + 2] = 0;
925 pmap->pm_pdir[ptva + 3] = 0;
926
927 /* Unmap the page table from the page table area. */
928 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) = 0;
929 }
930 #endif
931
932 /*
933 * Used to map a range of physical addresses into kernel
934 * virtual address space.
935 *
936 * For now, VM is already on, we only need to map the
937 * specified memory.
938 */
939 vaddr_t
940 pmap_map(vaddr_t va, paddr_t spa, paddr_t epa, vm_prot_t prot)
941 {
942 while (spa < epa) {
943 pmap_kenter_pa(va, spa, prot);
944 va += NBPG;
945 spa += NBPG;
946 }
947 pmap_update(pmap_kernel());
948 return(va);
949 }
950
951
952 /*
953 * void pmap_bootstrap(pd_entry_t *kernel_l1pt, pv_addr_t kernel_ptpt)
954 *
955 * bootstrap the pmap system. This is called from initarm and allows
956 * the pmap system to initailise any structures it requires.
957 *
958 * Currently this sets up the kernel_pmap that is statically allocated
959 * and also allocated virtual addresses for certain page hooks.
960 * Currently the only one page hook is allocated that is used
961 * to zero physical pages of memory.
962 * It also initialises the start and end address of the kernel data space.
963 */
964 extern paddr_t physical_freestart;
965 extern paddr_t physical_freeend;
966
967 char *boot_head;
968
969 void
970 pmap_bootstrap(pd_entry_t *kernel_l1pt, pv_addr_t kernel_ptpt)
971 {
972 pt_entry_t *pte;
973 int loop;
974 paddr_t start, end;
975 #if NISADMA > 0
976 paddr_t istart;
977 psize_t isize;
978 #endif
979
980 pmap_kernel()->pm_pdir = kernel_l1pt;
981 pmap_kernel()->pm_pptpt = kernel_ptpt.pv_pa;
982 pmap_kernel()->pm_vptpt = kernel_ptpt.pv_va;
983 simple_lock_init(&pmap_kernel()->pm_lock);
984 pmap_kernel()->pm_obj.pgops = NULL;
985 TAILQ_INIT(&(pmap_kernel()->pm_obj.memq));
986 pmap_kernel()->pm_obj.uo_npages = 0;
987 pmap_kernel()->pm_obj.uo_refs = 1;
988
989 /*
990 * Initialize PAGE_SIZE-dependent variables.
991 */
992 uvm_setpagesize();
993
994 loop = 0;
995 while (loop < bootconfig.dramblocks) {
996 start = (paddr_t)bootconfig.dram[loop].address;
997 end = start + (bootconfig.dram[loop].pages * NBPG);
998 if (start < physical_freestart)
999 start = physical_freestart;
1000 if (end > physical_freeend)
1001 end = physical_freeend;
1002 #if 0
1003 printf("%d: %lx -> %lx\n", loop, start, end - 1);
1004 #endif
1005 #if NISADMA > 0
1006 if (pmap_isa_dma_range_intersect(start, end - start,
1007 &istart, &isize)) {
1008 /*
1009 * Place the pages that intersect with the
1010 * ISA DMA range onto the ISA DMA free list.
1011 */
1012 #if 0
1013 printf(" ISADMA 0x%lx -> 0x%lx\n", istart,
1014 istart + isize - 1);
1015 #endif
1016 uvm_page_physload(atop(istart),
1017 atop(istart + isize), atop(istart),
1018 atop(istart + isize), VM_FREELIST_ISADMA);
1019
1020 /*
1021 * Load the pieces that come before
1022 * the intersection into the default
1023 * free list.
1024 */
1025 if (start < istart) {
1026 #if 0
1027 printf(" BEFORE 0x%lx -> 0x%lx\n",
1028 start, istart - 1);
1029 #endif
1030 uvm_page_physload(atop(start),
1031 atop(istart), atop(start),
1032 atop(istart), VM_FREELIST_DEFAULT);
1033 }
1034
1035 /*
1036 * Load the pieces that come after
1037 * the intersection into the default
1038 * free list.
1039 */
1040 if ((istart + isize) < end) {
1041 #if 0
1042 printf(" AFTER 0x%lx -> 0x%lx\n",
1043 (istart + isize), end - 1);
1044 #endif
1045 uvm_page_physload(atop(istart + isize),
1046 atop(end), atop(istart + isize),
1047 atop(end), VM_FREELIST_DEFAULT);
1048 }
1049 } else {
1050 uvm_page_physload(atop(start), atop(end),
1051 atop(start), atop(end), VM_FREELIST_DEFAULT);
1052 }
1053 #else /* NISADMA > 0 */
1054 uvm_page_physload(atop(start), atop(end),
1055 atop(start), atop(end), VM_FREELIST_DEFAULT);
1056 #endif /* NISADMA > 0 */
1057 ++loop;
1058 }
1059
1060 virtual_avail = KERNEL_VM_BASE;
1061 virtual_end = KERNEL_VM_BASE + KERNEL_VM_SIZE;
1062
1063 /*
1064 * now we allocate the "special" VAs which are used for tmp mappings
1065 * by the pmap (and other modules). we allocate the VAs by advancing
1066 * virtual_avail (note that there are no pages mapped at these VAs).
1067 * we find the PTE that maps the allocated VA via the linear PTE
1068 * mapping.
1069 */
1070
1071 pte = ((pt_entry_t *) PTE_BASE) + atop(virtual_avail);
1072
1073 csrcp = virtual_avail; csrc_pte = pte;
1074 virtual_avail += PAGE_SIZE; pte++;
1075
1076 cdstp = virtual_avail; cdst_pte = pte;
1077 virtual_avail += PAGE_SIZE; pte++;
1078
1079 memhook = (char *) virtual_avail; /* don't need pte */
1080 virtual_avail += PAGE_SIZE; pte++;
1081
1082 msgbufaddr = (caddr_t) virtual_avail; /* don't need pte */
1083 virtual_avail += round_page(MSGBUFSIZE);
1084 pte += atop(round_page(MSGBUFSIZE));
1085
1086 /*
1087 * init the static-global locks and global lists.
1088 */
1089 spinlockinit(&pmap_main_lock, "pmaplk", 0);
1090 simple_lock_init(&pvalloc_lock);
1091 simple_lock_init(&pmaps_lock);
1092 LIST_INIT(&pmaps);
1093 TAILQ_INIT(&pv_freepages);
1094 TAILQ_INIT(&pv_unusedpgs);
1095
1096 /*
1097 * initialize the pmap pool.
1098 */
1099
1100 pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl",
1101 &pool_allocator_nointr);
1102
1103 cpu_dcache_wbinv_all();
1104 }
1105
1106 /*
1107 * void pmap_init(void)
1108 *
1109 * Initialize the pmap module.
1110 * Called by vm_init() in vm/vm_init.c in order to initialise
1111 * any structures that the pmap system needs to map virtual memory.
1112 */
1113
1114 extern int physmem;
1115
1116 void
1117 pmap_init(void)
1118 {
1119
1120 /*
1121 * Set the available memory vars - These do not map to real memory
1122 * addresses and cannot as the physical memory is fragmented.
1123 * They are used by ps for %mem calculations.
1124 * One could argue whether this should be the entire memory or just
1125 * the memory that is useable in a user process.
1126 */
1127 avail_start = 0;
1128 avail_end = physmem * NBPG;
1129
1130 /*
1131 * now we need to free enough pv_entry structures to allow us to get
1132 * the kmem_map/kmem_object allocated and inited (done after this
1133 * function is finished). to do this we allocate one bootstrap page out
1134 * of kernel_map and use it to provide an initial pool of pv_entry
1135 * structures. we never free this page.
1136 */
1137
1138 pv_initpage = (struct pv_page *) uvm_km_alloc(kernel_map, PAGE_SIZE);
1139 if (pv_initpage == NULL)
1140 panic("pmap_init: pv_initpage");
1141 pv_cachedva = 0; /* a VA we have allocated but not used yet */
1142 pv_nfpvents = 0;
1143 (void) pmap_add_pvpage(pv_initpage, FALSE);
1144
1145 pmap_initialized = TRUE;
1146
1147 /* Initialise our L1 page table queues and counters */
1148 SIMPLEQ_INIT(&l1pt_static_queue);
1149 l1pt_static_queue_count = 0;
1150 l1pt_static_create_count = 0;
1151 SIMPLEQ_INIT(&l1pt_queue);
1152 l1pt_queue_count = 0;
1153 l1pt_create_count = 0;
1154 l1pt_reuse_count = 0;
1155 }
1156
1157 /*
1158 * pmap_postinit()
1159 *
1160 * This routine is called after the vm and kmem subsystems have been
1161 * initialised. This allows the pmap code to perform any initialisation
1162 * that can only be done one the memory allocation is in place.
1163 */
1164
1165 void
1166 pmap_postinit(void)
1167 {
1168 int loop;
1169 struct l1pt *pt;
1170
1171 #ifdef PMAP_STATIC_L1S
1172 for (loop = 0; loop < PMAP_STATIC_L1S; ++loop) {
1173 #else /* PMAP_STATIC_L1S */
1174 for (loop = 0; loop < max_processes; ++loop) {
1175 #endif /* PMAP_STATIC_L1S */
1176 /* Allocate a L1 page table */
1177 pt = pmap_alloc_l1pt();
1178 if (!pt)
1179 panic("Cannot allocate static L1 page tables\n");
1180
1181 /* Clean it */
1182 bzero((void *)pt->pt_va, L1_TABLE_SIZE);
1183 pt->pt_flags |= (PTFLAG_STATIC | PTFLAG_CLEAN);
1184 /* Add the page table to the queue */
1185 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pt, pt_queue);
1186 ++l1pt_static_queue_count;
1187 ++l1pt_static_create_count;
1188 }
1189 }
1190
1191
1192 /*
1193 * Create and return a physical map.
1194 *
1195 * If the size specified for the map is zero, the map is an actual physical
1196 * map, and may be referenced by the hardware.
1197 *
1198 * If the size specified is non-zero, the map will be used in software only,
1199 * and is bounded by that size.
1200 */
1201
1202 pmap_t
1203 pmap_create(void)
1204 {
1205 struct pmap *pmap;
1206
1207 /*
1208 * Fetch pmap entry from the pool
1209 */
1210
1211 pmap = pool_get(&pmap_pmap_pool, PR_WAITOK);
1212 /* XXX is this really needed! */
1213 memset(pmap, 0, sizeof(*pmap));
1214
1215 simple_lock_init(&pmap->pm_obj.vmobjlock);
1216 pmap->pm_obj.pgops = NULL; /* currently not a mappable object */
1217 TAILQ_INIT(&pmap->pm_obj.memq);
1218 pmap->pm_obj.uo_npages = 0;
1219 pmap->pm_obj.uo_refs = 1;
1220 pmap->pm_stats.wired_count = 0;
1221 pmap->pm_stats.resident_count = 1;
1222 pmap->pm_ptphint = NULL;
1223
1224 /* Now init the machine part of the pmap */
1225 pmap_pinit(pmap);
1226 return(pmap);
1227 }
1228
1229 /*
1230 * pmap_alloc_l1pt()
1231 *
1232 * This routine allocates physical and virtual memory for a L1 page table
1233 * and wires it.
1234 * A l1pt structure is returned to describe the allocated page table.
1235 *
1236 * This routine is allowed to fail if the required memory cannot be allocated.
1237 * In this case NULL is returned.
1238 */
1239
1240 struct l1pt *
1241 pmap_alloc_l1pt(void)
1242 {
1243 paddr_t pa;
1244 vaddr_t va;
1245 struct l1pt *pt;
1246 int error;
1247 struct vm_page *m;
1248 pt_entry_t *ptes;
1249
1250 /* Allocate virtual address space for the L1 page table */
1251 va = uvm_km_valloc(kernel_map, L1_TABLE_SIZE);
1252 if (va == 0) {
1253 #ifdef DIAGNOSTIC
1254 PDEBUG(0,
1255 printf("pmap: Cannot allocate pageable memory for L1\n"));
1256 #endif /* DIAGNOSTIC */
1257 return(NULL);
1258 }
1259
1260 /* Allocate memory for the l1pt structure */
1261 pt = (struct l1pt *)malloc(sizeof(struct l1pt), M_VMPMAP, M_WAITOK);
1262
1263 /*
1264 * Allocate pages from the VM system.
1265 */
1266 TAILQ_INIT(&pt->pt_plist);
1267 error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start, physical_end,
1268 L1_TABLE_SIZE, 0, &pt->pt_plist, 1, M_WAITOK);
1269 if (error) {
1270 #ifdef DIAGNOSTIC
1271 PDEBUG(0,
1272 printf("pmap: Cannot allocate physical mem for L1 (%d)\n",
1273 error));
1274 #endif /* DIAGNOSTIC */
1275 /* Release the resources we already have claimed */
1276 free(pt, M_VMPMAP);
1277 uvm_km_free(kernel_map, va, L1_TABLE_SIZE);
1278 return(NULL);
1279 }
1280
1281 /* Map our physical pages into our virtual space */
1282 pt->pt_va = va;
1283 m = TAILQ_FIRST(&pt->pt_plist);
1284 ptes = pmap_map_ptes(pmap_kernel());
1285 while (m && va < (pt->pt_va + L1_TABLE_SIZE)) {
1286 pa = VM_PAGE_TO_PHYS(m);
1287
1288 pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE);
1289
1290 /* Revoke cacheability and bufferability */
1291 /* XXX should be done better than this */
1292 ptes[arm_btop(va)] &= ~(L2_C | L2_B);
1293
1294 va += NBPG;
1295 m = m->pageq.tqe_next;
1296 }
1297 pmap_unmap_ptes(pmap_kernel());
1298 pmap_update(pmap_kernel());
1299
1300 #ifdef DIAGNOSTIC
1301 if (m)
1302 panic("pmap_alloc_l1pt: pglist not empty\n");
1303 #endif /* DIAGNOSTIC */
1304
1305 pt->pt_flags = 0;
1306 return(pt);
1307 }
1308
1309 /*
1310 * Free a L1 page table previously allocated with pmap_alloc_l1pt().
1311 */
1312 static void
1313 pmap_free_l1pt(struct l1pt *pt)
1314 {
1315 /* Separate the physical memory for the virtual space */
1316 pmap_kremove(pt->pt_va, L1_TABLE_SIZE);
1317 pmap_update(pmap_kernel());
1318
1319 /* Return the physical memory */
1320 uvm_pglistfree(&pt->pt_plist);
1321
1322 /* Free the virtual space */
1323 uvm_km_free(kernel_map, pt->pt_va, L1_TABLE_SIZE);
1324
1325 /* Free the l1pt structure */
1326 free(pt, M_VMPMAP);
1327 }
1328
1329 /*
1330 * Allocate a page directory.
1331 * This routine will either allocate a new page directory from the pool
1332 * of L1 page tables currently held by the kernel or it will allocate
1333 * a new one via pmap_alloc_l1pt().
1334 * It will then initialise the l1 page table for use.
1335 *
1336 * XXX must tidy up and fix this code, not happy about how it does the pmaps_locking
1337 */
1338 static int
1339 pmap_allocpagedir(struct pmap *pmap)
1340 {
1341 paddr_t pa;
1342 struct l1pt *pt;
1343 pt_entry_t *pte;
1344
1345 PDEBUG(0, printf("pmap_allocpagedir(%p)\n", pmap));
1346
1347 /* Do we have any spare L1's lying around ? */
1348 if (l1pt_static_queue_count) {
1349 --l1pt_static_queue_count;
1350 pt = l1pt_static_queue.sqh_first;
1351 SIMPLEQ_REMOVE_HEAD(&l1pt_static_queue, pt, pt_queue);
1352 } else if (l1pt_queue_count) {
1353 --l1pt_queue_count;
1354 pt = l1pt_queue.sqh_first;
1355 SIMPLEQ_REMOVE_HEAD(&l1pt_queue, pt, pt_queue);
1356 ++l1pt_reuse_count;
1357 } else {
1358 pt = pmap_alloc_l1pt();
1359 if (!pt)
1360 return(ENOMEM);
1361 ++l1pt_create_count;
1362 }
1363
1364 /* Store the pointer to the l1 descriptor in the pmap. */
1365 pmap->pm_l1pt = pt;
1366
1367 /* Get the physical address of the start of the l1 */
1368 pa = VM_PAGE_TO_PHYS(TAILQ_FIRST(&pt->pt_plist));
1369
1370 /* Store the virtual address of the l1 in the pmap. */
1371 pmap->pm_pdir = (pd_entry_t *)pt->pt_va;
1372
1373 /* Clean the L1 if it is dirty */
1374 if (!(pt->pt_flags & PTFLAG_CLEAN))
1375 bzero((void *)pmap->pm_pdir, (L1_TABLE_SIZE - KERNEL_PD_SIZE));
1376
1377 /* Allocate a page table to map all the page tables for this pmap */
1378
1379 #ifdef DIAGNOSTIC
1380 if (pmap->pm_vptpt) {
1381 /* XXX What if we have one already ? */
1382 panic("pmap_allocpagedir: have pt already\n");
1383 }
1384 #endif /* DIAGNOSTIC */
1385 pmap->pm_vptpt = uvm_km_zalloc(kernel_map, NBPG);
1386 if (pmap->pm_vptpt == 0) {
1387 pmap_freepagedir(pmap);
1388 return(ENOMEM);
1389 }
1390
1391 /* need to lock this all up for growkernel */
1392 simple_lock(&pmaps_lock);
1393 /* wish we didn't have to keep this locked... */
1394
1395 /* Duplicate the kernel mappings. */
1396 bcopy((char *)pmap_kernel()->pm_pdir + (L1_TABLE_SIZE - KERNEL_PD_SIZE),
1397 (char *)pmap->pm_pdir + (L1_TABLE_SIZE - KERNEL_PD_SIZE),
1398 KERNEL_PD_SIZE);
1399
1400 pte = vtopte(pmap->pm_vptpt);
1401 pmap->pm_pptpt = l2pte_pa(*pte);
1402
1403 /* Revoke cacheability and bufferability */
1404 /* XXX should be done better than this */
1405 *pte &= ~(L2_C | L2_B);
1406
1407 /* Wire in this page table */
1408 pmap_map_in_l1(pmap, PTE_BASE, pmap->pm_pptpt, TRUE);
1409
1410 pt->pt_flags &= ~PTFLAG_CLEAN; /* L1 is dirty now */
1411
1412 /*
1413 * Map the kernel page tables into the new PT map.
1414 */
1415 bcopy((char *)(PTE_BASE
1416 + (PTE_BASE >> (PGSHIFT - 2))
1417 + ((L1_TABLE_SIZE - KERNEL_PD_SIZE) >> 2)),
1418 (char *)pmap->pm_vptpt + ((L1_TABLE_SIZE - KERNEL_PD_SIZE) >> 2),
1419 (KERNEL_PD_SIZE >> 2));
1420
1421 LIST_INSERT_HEAD(&pmaps, pmap, pm_list);
1422 simple_unlock(&pmaps_lock);
1423
1424 return(0);
1425 }
1426
1427
1428 /*
1429 * Initialize a preallocated and zeroed pmap structure,
1430 * such as one in a vmspace structure.
1431 */
1432
1433 void
1434 pmap_pinit(struct pmap *pmap)
1435 {
1436 int backoff = 6;
1437 int retry = 10;
1438
1439 PDEBUG(0, printf("pmap_pinit(%p)\n", pmap));
1440
1441 /* Keep looping until we succeed in allocating a page directory */
1442 while (pmap_allocpagedir(pmap) != 0) {
1443 /*
1444 * Ok we failed to allocate a suitable block of memory for an
1445 * L1 page table. This means that either:
1446 * 1. 16KB of virtual address space could not be allocated
1447 * 2. 16KB of physically contiguous memory on a 16KB boundary
1448 * could not be allocated.
1449 *
1450 * Since we cannot fail we will sleep for a while and try
1451 * again.
1452 *
1453 * Searching for a suitable L1 PT is expensive:
1454 * to avoid hogging the system when memory is really
1455 * scarce, use an exponential back-off so that
1456 * eventually we won't retry more than once every 8
1457 * seconds. This should allow other processes to run
1458 * to completion and free up resources.
1459 */
1460 (void) ltsleep(&lbolt, PVM, "l1ptwait", (hz << 3) >> backoff,
1461 NULL);
1462 if (--retry == 0) {
1463 retry = 10;
1464 if (backoff)
1465 --backoff;
1466 }
1467 }
1468
1469 if (vector_page < KERNEL_BASE) {
1470 /*
1471 * Map the vector page. This will also allocate and map
1472 * an L2 table for it.
1473 */
1474 pmap_enter(pmap, vector_page, systempage.pv_pa,
1475 VM_PROT_READ, VM_PROT_READ | PMAP_WIRED);
1476 pmap_update(pmap);
1477 }
1478 }
1479
1480
1481 void
1482 pmap_freepagedir(struct pmap *pmap)
1483 {
1484 /* Free the memory used for the page table mapping */
1485 if (pmap->pm_vptpt != 0)
1486 uvm_km_free(kernel_map, (vaddr_t)pmap->pm_vptpt, NBPG);
1487
1488 /* junk the L1 page table */
1489 if (pmap->pm_l1pt->pt_flags & PTFLAG_STATIC) {
1490 /* Add the page table to the queue */
1491 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pmap->pm_l1pt, pt_queue);
1492 ++l1pt_static_queue_count;
1493 } else if (l1pt_queue_count < 8) {
1494 /* Add the page table to the queue */
1495 SIMPLEQ_INSERT_TAIL(&l1pt_queue, pmap->pm_l1pt, pt_queue);
1496 ++l1pt_queue_count;
1497 } else
1498 pmap_free_l1pt(pmap->pm_l1pt);
1499 }
1500
1501
1502 /*
1503 * Retire the given physical map from service.
1504 * Should only be called if the map contains no valid mappings.
1505 */
1506
1507 void
1508 pmap_destroy(struct pmap *pmap)
1509 {
1510 struct vm_page *page;
1511 int count;
1512
1513 if (pmap == NULL)
1514 return;
1515
1516 PDEBUG(0, printf("pmap_destroy(%p)\n", pmap));
1517
1518 /*
1519 * Drop reference count
1520 */
1521 simple_lock(&pmap->pm_obj.vmobjlock);
1522 count = --pmap->pm_obj.uo_refs;
1523 simple_unlock(&pmap->pm_obj.vmobjlock);
1524 if (count > 0) {
1525 return;
1526 }
1527
1528 /*
1529 * reference count is zero, free pmap resources and then free pmap.
1530 */
1531
1532 /*
1533 * remove it from global list of pmaps
1534 */
1535
1536 simple_lock(&pmaps_lock);
1537 LIST_REMOVE(pmap, pm_list);
1538 simple_unlock(&pmaps_lock);
1539
1540 if (vector_page < KERNEL_BASE) {
1541 /* Remove the vector page mapping */
1542 pmap_remove(pmap, vector_page, vector_page + NBPG);
1543 pmap_update(pmap);
1544 }
1545
1546 /*
1547 * Free any page tables still mapped
1548 * This is only temporay until pmap_enter can count the number
1549 * of mappings made in a page table. Then pmap_remove() can
1550 * reduce the count and free the pagetable when the count
1551 * reaches zero. Note that entries in this list should match the
1552 * contents of the ptpt, however this is faster than walking a 1024
1553 * entries looking for pt's
1554 * taken from i386 pmap.c
1555 */
1556 while ((page = TAILQ_FIRST(&pmap->pm_obj.memq)) != NULL) {
1557 KASSERT((page->flags & PG_BUSY) == 0);
1558 page->wire_count = 0;
1559 uvm_pagefree(page);
1560 }
1561
1562 /* Free the page dir */
1563 pmap_freepagedir(pmap);
1564
1565 /* return the pmap to the pool */
1566 pool_put(&pmap_pmap_pool, pmap);
1567 }
1568
1569
1570 /*
1571 * void pmap_reference(struct pmap *pmap)
1572 *
1573 * Add a reference to the specified pmap.
1574 */
1575
1576 void
1577 pmap_reference(struct pmap *pmap)
1578 {
1579 if (pmap == NULL)
1580 return;
1581
1582 simple_lock(&pmap->pm_lock);
1583 pmap->pm_obj.uo_refs++;
1584 simple_unlock(&pmap->pm_lock);
1585 }
1586
1587 /*
1588 * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
1589 *
1590 * Return the start and end addresses of the kernel's virtual space.
1591 * These values are setup in pmap_bootstrap and are updated as pages
1592 * are allocated.
1593 */
1594
1595 void
1596 pmap_virtual_space(vaddr_t *start, vaddr_t *end)
1597 {
1598 *start = virtual_avail;
1599 *end = virtual_end;
1600 }
1601
1602 /*
1603 * Activate the address space for the specified process. If the process
1604 * is the current process, load the new MMU context.
1605 */
1606 void
1607 pmap_activate(struct proc *p)
1608 {
1609 struct pmap *pmap = p->p_vmspace->vm_map.pmap;
1610 struct pcb *pcb = &p->p_addr->u_pcb;
1611
1612 (void) pmap_extract(pmap_kernel(), (vaddr_t)pmap->pm_pdir,
1613 (paddr_t *)&pcb->pcb_pagedir);
1614
1615 PDEBUG(0, printf("pmap_activate: p=%p pmap=%p pcb=%p pdir=%p l1=%p\n",
1616 p, pmap, pcb, pmap->pm_pdir, pcb->pcb_pagedir));
1617
1618 if (p == curproc) {
1619 PDEBUG(0, printf("pmap_activate: setting TTB\n"));
1620 setttb((u_int)pcb->pcb_pagedir);
1621 }
1622 }
1623
1624 /*
1625 * Deactivate the address space of the specified process.
1626 */
1627 void
1628 pmap_deactivate(struct proc *p)
1629 {
1630 }
1631
1632 /*
1633 * Perform any deferred pmap operations.
1634 */
1635 void
1636 pmap_update(struct pmap *pmap)
1637 {
1638
1639 /*
1640 * We haven't deferred any pmap operations, but we do need to
1641 * make sure TLB/cache operations have completed.
1642 */
1643 cpu_cpwait();
1644 }
1645
1646 /*
1647 * pmap_clean_page()
1648 *
1649 * This is a local function used to work out the best strategy to clean
1650 * a single page referenced by its entry in the PV table. It's used by
1651 * pmap_copy_page, pmap_zero page and maybe some others later on.
1652 *
1653 * Its policy is effectively:
1654 * o If there are no mappings, we don't bother doing anything with the cache.
1655 * o If there is one mapping, we clean just that page.
1656 * o If there are multiple mappings, we clean the entire cache.
1657 *
1658 * So that some functions can be further optimised, it returns 0 if it didn't
1659 * clean the entire cache, or 1 if it did.
1660 *
1661 * XXX One bug in this routine is that if the pv_entry has a single page
1662 * mapped at 0x00000000 a whole cache clean will be performed rather than
1663 * just the 1 page. Since this should not occur in everyday use and if it does
1664 * it will just result in not the most efficient clean for the page.
1665 */
1666 static int
1667 pmap_clean_page(struct pv_entry *pv, boolean_t is_src)
1668 {
1669 struct pmap *pmap;
1670 struct pv_entry *npv;
1671 int cache_needs_cleaning = 0;
1672 vaddr_t page_to_clean = 0;
1673
1674 if (pv == NULL)
1675 /* nothing mapped in so nothing to flush */
1676 return (0);
1677
1678 /* Since we flush the cache each time we change curproc, we
1679 * only need to flush the page if it is in the current pmap.
1680 */
1681 if (curproc)
1682 pmap = curproc->p_vmspace->vm_map.pmap;
1683 else
1684 pmap = pmap_kernel();
1685
1686 for (npv = pv; npv; npv = npv->pv_next) {
1687 if (npv->pv_pmap == pmap) {
1688 /* The page is mapped non-cacheable in
1689 * this map. No need to flush the cache.
1690 */
1691 if (npv->pv_flags & PVF_NC) {
1692 #ifdef DIAGNOSTIC
1693 if (cache_needs_cleaning)
1694 panic("pmap_clean_page: "
1695 "cache inconsistency");
1696 #endif
1697 break;
1698 }
1699 #if 0
1700 /* This doesn't work, because pmap_protect
1701 doesn't flush changes on pages that it
1702 has write-protected. */
1703
1704 /* If the page is not writable and this
1705 is the source, then there is no need
1706 to flush it from the cache. */
1707 else if (is_src && ! (npv->pv_flags & PVF_WRITE))
1708 continue;
1709 #endif
1710 if (cache_needs_cleaning){
1711 page_to_clean = 0;
1712 break;
1713 }
1714 else
1715 page_to_clean = npv->pv_va;
1716 cache_needs_cleaning = 1;
1717 }
1718 }
1719
1720 if (page_to_clean)
1721 cpu_idcache_wbinv_range(page_to_clean, NBPG);
1722 else if (cache_needs_cleaning) {
1723 cpu_idcache_wbinv_all();
1724 return (1);
1725 }
1726 return (0);
1727 }
1728
1729 /*
1730 * pmap_zero_page()
1731 *
1732 * Zero a given physical page by mapping it at a page hook point.
1733 * In doing the zero page op, the page we zero is mapped cachable, as with
1734 * StrongARM accesses to non-cached pages are non-burst making writing
1735 * _any_ bulk data very slow.
1736 */
1737 #if ARM_MMU_GENERIC == 1
1738 void
1739 pmap_zero_page_generic(paddr_t phys)
1740 {
1741 #ifdef DEBUG
1742 struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
1743
1744 if (pg->mdpage.pvh_list != NULL)
1745 panic("pmap_zero_page: page has mappings");
1746 #endif
1747
1748 KDASSERT((phys & PGOFSET) == 0);
1749
1750 /*
1751 * Hook in the page, zero it, and purge the cache for that
1752 * zeroed page. Invalidate the TLB as needed.
1753 */
1754 *cdst_pte = L2_S_PROTO | phys |
1755 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
1756 cpu_tlb_flushD_SE(cdstp);
1757 cpu_cpwait();
1758 bzero_page(cdstp);
1759 cpu_dcache_wbinv_range(cdstp, NBPG);
1760 }
1761 #endif /* ARM_MMU_GENERIC == 1 */
1762
1763 #if ARM_MMU_XSCALE == 1
1764 void
1765 pmap_zero_page_xscale(paddr_t phys)
1766 {
1767 #ifdef DEBUG
1768 struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
1769
1770 if (pg->mdpage.pvh_list != NULL)
1771 panic("pmap_zero_page: page has mappings");
1772 #endif
1773
1774 KDASSERT((phys & PGOFSET) == 0);
1775
1776 /*
1777 * Hook in the page, zero it, and purge the cache for that
1778 * zeroed page. Invalidate the TLB as needed.
1779 */
1780 *cdst_pte = L2_S_PROTO | phys |
1781 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
1782 L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X); /* mini-data */
1783 cpu_tlb_flushD_SE(cdstp);
1784 cpu_cpwait();
1785 bzero_page(cdstp);
1786 xscale_cache_clean_minidata();
1787 }
1788 #endif /* ARM_MMU_XSCALE == 1 */
1789
1790 /* pmap_pageidlezero()
1791 *
1792 * The same as above, except that we assume that the page is not
1793 * mapped. This means we never have to flush the cache first. Called
1794 * from the idle loop.
1795 */
1796 boolean_t
1797 pmap_pageidlezero(paddr_t phys)
1798 {
1799 int i, *ptr;
1800 boolean_t rv = TRUE;
1801 #ifdef DEBUG
1802 struct vm_page *pg;
1803
1804 pg = PHYS_TO_VM_PAGE(phys);
1805 if (pg->mdpage.pvh_list != NULL)
1806 panic("pmap_pageidlezero: page has mappings");
1807 #endif
1808
1809 KDASSERT((phys & PGOFSET) == 0);
1810
1811 /*
1812 * Hook in the page, zero it, and purge the cache for that
1813 * zeroed page. Invalidate the TLB as needed.
1814 */
1815 *cdst_pte = L2_S_PROTO | phys |
1816 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
1817 cpu_tlb_flushD_SE(cdstp);
1818 cpu_cpwait();
1819
1820 for (i = 0, ptr = (int *)cdstp;
1821 i < (NBPG / sizeof(int)); i++) {
1822 if (sched_whichqs != 0) {
1823 /*
1824 * A process has become ready. Abort now,
1825 * so we don't keep it waiting while we
1826 * do slow memory access to finish this
1827 * page.
1828 */
1829 rv = FALSE;
1830 break;
1831 }
1832 *ptr++ = 0;
1833 }
1834
1835 if (rv)
1836 /*
1837 * if we aborted we'll rezero this page again later so don't
1838 * purge it unless we finished it
1839 */
1840 cpu_dcache_wbinv_range(cdstp, NBPG);
1841 return (rv);
1842 }
1843
1844 /*
1845 * pmap_copy_page()
1846 *
1847 * Copy one physical page into another, by mapping the pages into
1848 * hook points. The same comment regarding cachability as in
1849 * pmap_zero_page also applies here.
1850 */
1851 #if ARM_MMU_GENERIC == 1
1852 void
1853 pmap_copy_page_generic(paddr_t src, paddr_t dst)
1854 {
1855 struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
1856 #ifdef DEBUG
1857 struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst);
1858
1859 if (dst_pg->mdpage.pvh_list != NULL)
1860 panic("pmap_copy_page: dst page has mappings");
1861 #endif
1862
1863 KDASSERT((src & PGOFSET) == 0);
1864 KDASSERT((dst & PGOFSET) == 0);
1865
1866 /*
1867 * Clean the source page. Hold the source page's lock for
1868 * the duration of the copy so that no other mappings can
1869 * be created while we have a potentially aliased mapping.
1870 */
1871 simple_lock(&src_pg->mdpage.pvh_slock);
1872 (void) pmap_clean_page(src_pg->mdpage.pvh_list, TRUE);
1873
1874 /*
1875 * Map the pages into the page hook points, copy them, and purge
1876 * the cache for the appropriate page. Invalidate the TLB
1877 * as required.
1878 */
1879 *csrc_pte = L2_S_PROTO | src |
1880 L2_S_PROT(PTE_KERNEL, VM_PROT_READ) | pte_l2_s_cache_mode;
1881 *cdst_pte = L2_S_PROTO | dst |
1882 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
1883 cpu_tlb_flushD_SE(csrcp);
1884 cpu_tlb_flushD_SE(cdstp);
1885 cpu_cpwait();
1886 bcopy_page(csrcp, cdstp);
1887 cpu_dcache_inv_range(csrcp, NBPG);
1888 simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
1889 cpu_dcache_wbinv_range(cdstp, NBPG);
1890 }
1891 #endif /* ARM_MMU_GENERIC == 1 */
1892
1893 #if ARM_MMU_XSCALE == 1
1894 void
1895 pmap_copy_page_xscale(paddr_t src, paddr_t dst)
1896 {
1897 struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
1898 #ifdef DEBUG
1899 struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst);
1900
1901 if (dst_pg->mdpage.pvh_list != NULL)
1902 panic("pmap_copy_page: dst page has mappings");
1903 #endif
1904
1905 KDASSERT((src & PGOFSET) == 0);
1906 KDASSERT((dst & PGOFSET) == 0);
1907
1908 /*
1909 * Clean the source page. Hold the source page's lock for
1910 * the duration of the copy so that no other mappings can
1911 * be created while we have a potentially aliased mapping.
1912 */
1913 simple_lock(&src_pg->mdpage.pvh_slock);
1914 (void) pmap_clean_page(src_pg->mdpage.pvh_list, TRUE);
1915
1916 /*
1917 * Map the pages into the page hook points, copy them, and purge
1918 * the cache for the appropriate page. Invalidate the TLB
1919 * as required.
1920 */
1921 *csrc_pte = L2_S_PROTO | src |
1922 L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
1923 L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X); /* mini-data */
1924 *cdst_pte = L2_S_PROTO | dst |
1925 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
1926 L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X); /* mini-data */
1927 cpu_tlb_flushD_SE(csrcp);
1928 cpu_tlb_flushD_SE(cdstp);
1929 cpu_cpwait();
1930 bcopy_page(csrcp, cdstp);
1931 simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
1932 xscale_cache_clean_minidata();
1933 }
1934 #endif /* ARM_MMU_XSCALE == 1 */
1935
1936 #if 0
1937 void
1938 pmap_pte_addref(struct pmap *pmap, vaddr_t va)
1939 {
1940 pd_entry_t *pde;
1941 paddr_t pa;
1942 struct vm_page *m;
1943
1944 if (pmap == pmap_kernel())
1945 return;
1946
1947 pde = pmap_pde(pmap, va & ~(3 << L1_S_SHIFT));
1948 pa = pmap_pte_pa(pde);
1949 m = PHYS_TO_VM_PAGE(pa);
1950 ++m->wire_count;
1951 #ifdef MYCROFT_HACK
1952 printf("addref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
1953 pmap, va, pde, pa, m, m->wire_count);
1954 #endif
1955 }
1956
1957 void
1958 pmap_pte_delref(struct pmap *pmap, vaddr_t va)
1959 {
1960 pd_entry_t *pde;
1961 paddr_t pa;
1962 struct vm_page *m;
1963
1964 if (pmap == pmap_kernel())
1965 return;
1966
1967 pde = pmap_pde(pmap, va & ~(3 << L1_S_SHIFT));
1968 pa = pmap_pte_pa(pde);
1969 m = PHYS_TO_VM_PAGE(pa);
1970 --m->wire_count;
1971 #ifdef MYCROFT_HACK
1972 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
1973 pmap, va, pde, pa, m, m->wire_count);
1974 #endif
1975 if (m->wire_count == 0) {
1976 #ifdef MYCROFT_HACK
1977 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p\n",
1978 pmap, va, pde, pa, m);
1979 #endif
1980 pmap_unmap_in_l1(pmap, va);
1981 uvm_pagefree(m);
1982 --pmap->pm_stats.resident_count;
1983 }
1984 }
1985 #else
1986 #define pmap_pte_addref(pmap, va)
1987 #define pmap_pte_delref(pmap, va)
1988 #endif
1989
1990 /*
1991 * Since we have a virtually indexed cache, we may need to inhibit caching if
1992 * there is more than one mapping and at least one of them is writable.
1993 * Since we purge the cache on every context switch, we only need to check for
1994 * other mappings within the same pmap, or kernel_pmap.
1995 * This function is also called when a page is unmapped, to possibly reenable
1996 * caching on any remaining mappings.
1997 *
1998 * The code implements the following logic, where:
1999 *
2000 * KW = # of kernel read/write pages
2001 * KR = # of kernel read only pages
2002 * UW = # of user read/write pages
2003 * UR = # of user read only pages
2004 * OW = # of user read/write pages in another pmap, then
2005 *
2006 * KC = kernel mapping is cacheable
2007 * UC = user mapping is cacheable
2008 *
2009 * KW=0,KR=0 KW=0,KR>0 KW=1,KR=0 KW>1,KR>=0
2010 * +---------------------------------------------
2011 * UW=0,UR=0,OW=0 | --- KC=1 KC=1 KC=0
2012 * UW=0,UR>0,OW=0 | UC=1 KC=1,UC=1 KC=0,UC=0 KC=0,UC=0
2013 * UW=0,UR>0,OW>0 | UC=1 KC=0,UC=1 KC=0,UC=0 KC=0,UC=0
2014 * UW=1,UR=0,OW=0 | UC=1 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0
2015 * UW>1,UR>=0,OW>=0 | UC=0 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0
2016 *
2017 * Note that the pmap must have it's ptes mapped in, and passed with ptes.
2018 */
2019 __inline static void
2020 pmap_vac_me_harder(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2021 boolean_t clear_cache)
2022 {
2023 if (pmap == pmap_kernel())
2024 pmap_vac_me_kpmap(pmap, pg, ptes, clear_cache);
2025 else
2026 pmap_vac_me_user(pmap, pg, ptes, clear_cache);
2027 }
2028
2029 static void
2030 pmap_vac_me_kpmap(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2031 boolean_t clear_cache)
2032 {
2033 int user_entries = 0;
2034 int user_writable = 0;
2035 int user_cacheable = 0;
2036 int kernel_entries = 0;
2037 int kernel_writable = 0;
2038 int kernel_cacheable = 0;
2039 struct pv_entry *pv;
2040 struct pmap *last_pmap = pmap;
2041
2042 #ifdef DIAGNOSTIC
2043 if (pmap != pmap_kernel())
2044 panic("pmap_vac_me_kpmap: pmap != pmap_kernel()");
2045 #endif
2046
2047 /*
2048 * Pass one, see if there are both kernel and user pmaps for
2049 * this page. Calculate whether there are user-writable or
2050 * kernel-writable pages.
2051 */
2052 for (pv = pg->mdpage.pvh_list; pv != NULL; pv = pv->pv_next) {
2053 if (pv->pv_pmap != pmap) {
2054 user_entries++;
2055 if (pv->pv_flags & PVF_WRITE)
2056 user_writable++;
2057 if ((pv->pv_flags & PVF_NC) == 0)
2058 user_cacheable++;
2059 } else {
2060 kernel_entries++;
2061 if (pv->pv_flags & PVF_WRITE)
2062 kernel_writable++;
2063 if ((pv->pv_flags & PVF_NC) == 0)
2064 kernel_cacheable++;
2065 }
2066 }
2067
2068 /*
2069 * We know we have just been updating a kernel entry, so if
2070 * all user pages are already cacheable, then there is nothing
2071 * further to do.
2072 */
2073 if (kernel_entries == 0 &&
2074 user_cacheable == user_entries)
2075 return;
2076
2077 if (user_entries) {
2078 /*
2079 * Scan over the list again, for each entry, if it
2080 * might not be set correctly, call pmap_vac_me_user
2081 * to recalculate the settings.
2082 */
2083 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
2084 /*
2085 * We know kernel mappings will get set
2086 * correctly in other calls. We also know
2087 * that if the pmap is the same as last_pmap
2088 * then we've just handled this entry.
2089 */
2090 if (pv->pv_pmap == pmap || pv->pv_pmap == last_pmap)
2091 continue;
2092 /*
2093 * If there are kernel entries and this page
2094 * is writable but non-cacheable, then we can
2095 * skip this entry also.
2096 */
2097 if (kernel_entries > 0 &&
2098 (pv->pv_flags & (PVF_NC | PVF_WRITE)) ==
2099 (PVF_NC | PVF_WRITE))
2100 continue;
2101 /*
2102 * Similarly if there are no kernel-writable
2103 * entries and the page is already
2104 * read-only/cacheable.
2105 */
2106 if (kernel_writable == 0 &&
2107 (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 0)
2108 continue;
2109 /*
2110 * For some of the remaining cases, we know
2111 * that we must recalculate, but for others we
2112 * can't tell if they are correct or not, so
2113 * we recalculate anyway.
2114 */
2115 pmap_unmap_ptes(last_pmap);
2116 last_pmap = pv->pv_pmap;
2117 ptes = pmap_map_ptes(last_pmap);
2118 pmap_vac_me_user(last_pmap, pg, ptes,
2119 pmap_is_curpmap(last_pmap));
2120 }
2121 /* Restore the pte mapping that was passed to us. */
2122 if (last_pmap != pmap) {
2123 pmap_unmap_ptes(last_pmap);
2124 ptes = pmap_map_ptes(pmap);
2125 }
2126 if (kernel_entries == 0)
2127 return;
2128 }
2129
2130 pmap_vac_me_user(pmap, pg, ptes, clear_cache);
2131 return;
2132 }
2133
2134 static void
2135 pmap_vac_me_user(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2136 boolean_t clear_cache)
2137 {
2138 struct pmap *kpmap = pmap_kernel();
2139 struct pv_entry *pv, *npv;
2140 int entries = 0;
2141 int writable = 0;
2142 int cacheable_entries = 0;
2143 int kern_cacheable = 0;
2144 int other_writable = 0;
2145
2146 pv = pg->mdpage.pvh_list;
2147 KASSERT(ptes != NULL);
2148
2149 /*
2150 * Count mappings and writable mappings in this pmap.
2151 * Include kernel mappings as part of our own.
2152 * Keep a pointer to the first one.
2153 */
2154 for (npv = pv; npv; npv = npv->pv_next) {
2155 /* Count mappings in the same pmap */
2156 if (pmap == npv->pv_pmap ||
2157 kpmap == npv->pv_pmap) {
2158 if (entries++ == 0)
2159 pv = npv;
2160 /* Cacheable mappings */
2161 if ((npv->pv_flags & PVF_NC) == 0) {
2162 cacheable_entries++;
2163 if (kpmap == npv->pv_pmap)
2164 kern_cacheable++;
2165 }
2166 /* Writable mappings */
2167 if (npv->pv_flags & PVF_WRITE)
2168 ++writable;
2169 } else if (npv->pv_flags & PVF_WRITE)
2170 other_writable = 1;
2171 }
2172
2173 PDEBUG(3,printf("pmap_vac_me_harder: pmap %p Entries %d, "
2174 "writable %d cacheable %d %s\n", pmap, entries, writable,
2175 cacheable_entries, clear_cache ? "clean" : "no clean"));
2176
2177 /*
2178 * Enable or disable caching as necessary.
2179 * Note: the first entry might be part of the kernel pmap,
2180 * so we can't assume this is indicative of the state of the
2181 * other (maybe non-kpmap) entries.
2182 */
2183 if ((entries > 1 && writable) ||
2184 (entries > 0 && pmap == kpmap && other_writable)) {
2185 if (cacheable_entries == 0)
2186 return;
2187 for (npv = pv; npv; npv = npv->pv_next) {
2188 if ((pmap == npv->pv_pmap
2189 || kpmap == npv->pv_pmap) &&
2190 (npv->pv_flags & PVF_NC) == 0) {
2191 ptes[arm_btop(npv->pv_va)] &= ~(L2_C | L2_B);
2192 npv->pv_flags |= PVF_NC;
2193 /*
2194 * If this page needs flushing from the
2195 * cache, and we aren't going to do it
2196 * below, do it now.
2197 */
2198 if ((cacheable_entries < 4 &&
2199 (clear_cache || npv->pv_pmap == kpmap)) ||
2200 (npv->pv_pmap == kpmap &&
2201 !clear_cache && kern_cacheable < 4)) {
2202 cpu_idcache_wbinv_range(npv->pv_va,
2203 NBPG);
2204 cpu_tlb_flushID_SE(npv->pv_va);
2205 }
2206 }
2207 }
2208 if ((clear_cache && cacheable_entries >= 4) ||
2209 kern_cacheable >= 4) {
2210 cpu_idcache_wbinv_all();
2211 cpu_tlb_flushID();
2212 }
2213 cpu_cpwait();
2214 } else if (entries > 0) {
2215 /*
2216 * Turn cacheing back on for some pages. If it is a kernel
2217 * page, only do so if there are no other writable pages.
2218 */
2219 for (npv = pv; npv; npv = npv->pv_next) {
2220 if ((pmap == npv->pv_pmap ||
2221 (kpmap == npv->pv_pmap && other_writable == 0)) &&
2222 (npv->pv_flags & PVF_NC)) {
2223 ptes[arm_btop(npv->pv_va)] |=
2224 pte_l2_s_cache_mode;
2225 npv->pv_flags &= ~PVF_NC;
2226 }
2227 }
2228 }
2229 }
2230
2231 /*
2232 * pmap_remove()
2233 *
2234 * pmap_remove is responsible for nuking a number of mappings for a range
2235 * of virtual address space in the current pmap. To do this efficiently
2236 * is interesting, because in a number of cases a wide virtual address
2237 * range may be supplied that contains few actual mappings. So, the
2238 * optimisations are:
2239 * 1. Try and skip over hunks of address space for which an L1 entry
2240 * does not exist.
2241 * 2. Build up a list of pages we've hit, up to a maximum, so we can
2242 * maybe do just a partial cache clean. This path of execution is
2243 * complicated by the fact that the cache must be flushed _before_
2244 * the PTE is nuked, being a VAC :-)
2245 * 3. Maybe later fast-case a single page, but I don't think this is
2246 * going to make _that_ much difference overall.
2247 */
2248
2249 #define PMAP_REMOVE_CLEAN_LIST_SIZE 3
2250
2251 void
2252 pmap_remove(struct pmap *pmap, vaddr_t sva, vaddr_t eva)
2253 {
2254 int cleanlist_idx = 0;
2255 struct pagelist {
2256 vaddr_t va;
2257 pt_entry_t *pte;
2258 } cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
2259 pt_entry_t *pte = 0, *ptes;
2260 paddr_t pa;
2261 int pmap_active;
2262 struct vm_page *pg;
2263
2264 /* Exit quick if there is no pmap */
2265 if (!pmap)
2266 return;
2267
2268 PDEBUG(0, printf("pmap_remove: pmap=%p sva=%08lx eva=%08lx\n",
2269 pmap, sva, eva));
2270
2271 /*
2272 * we lock in the pmap => vm_page direction
2273 */
2274 PMAP_MAP_TO_HEAD_LOCK();
2275
2276 ptes = pmap_map_ptes(pmap);
2277 /* Get a page table pointer */
2278 while (sva < eva) {
2279 if (pmap_pde_page(pmap_pde(pmap, sva)))
2280 break;
2281 sva = (sva & L1_S_FRAME) + L1_S_SIZE;
2282 }
2283
2284 pte = &ptes[arm_btop(sva)];
2285 /* Note if the pmap is active thus require cache and tlb cleans */
2286 pmap_active = pmap_is_curpmap(pmap);
2287
2288 /* Now loop along */
2289 while (sva < eva) {
2290 /* Check if we can move to the next PDE (l1 chunk) */
2291 if (!(sva & L2_ADDR_BITS))
2292 if (!pmap_pde_page(pmap_pde(pmap, sva))) {
2293 sva += L1_S_SIZE;
2294 pte += arm_btop(L1_S_SIZE);
2295 continue;
2296 }
2297
2298 /* We've found a valid PTE, so this page of PTEs has to go. */
2299 if (pmap_pte_v(pte)) {
2300 /* Update statistics */
2301 --pmap->pm_stats.resident_count;
2302
2303 /*
2304 * Add this page to our cache remove list, if we can.
2305 * If, however the cache remove list is totally full,
2306 * then do a complete cache invalidation taking note
2307 * to backtrack the PTE table beforehand, and ignore
2308 * the lists in future because there's no longer any
2309 * point in bothering with them (we've paid the
2310 * penalty, so will carry on unhindered). Otherwise,
2311 * when we fall out, we just clean the list.
2312 */
2313 PDEBUG(10, printf("remove: inv pte at %p(%x) ", pte, *pte));
2314 pa = pmap_pte_pa(pte);
2315
2316 if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
2317 /* Add to the clean list. */
2318 cleanlist[cleanlist_idx].pte = pte;
2319 cleanlist[cleanlist_idx].va = sva;
2320 cleanlist_idx++;
2321 } else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
2322 int cnt;
2323
2324 /* Nuke everything if needed. */
2325 if (pmap_active) {
2326 cpu_idcache_wbinv_all();
2327 cpu_tlb_flushID();
2328 }
2329
2330 /*
2331 * Roll back the previous PTE list,
2332 * and zero out the current PTE.
2333 */
2334 for (cnt = 0; cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
2335 *cleanlist[cnt].pte = 0;
2336 pmap_pte_delref(pmap, cleanlist[cnt].va);
2337 }
2338 *pte = 0;
2339 pmap_pte_delref(pmap, sva);
2340 cleanlist_idx++;
2341 } else {
2342 /*
2343 * We've already nuked the cache and
2344 * TLB, so just carry on regardless,
2345 * and we won't need to do it again
2346 */
2347 *pte = 0;
2348 pmap_pte_delref(pmap, sva);
2349 }
2350
2351 /*
2352 * Update flags. In a number of circumstances,
2353 * we could cluster a lot of these and do a
2354 * number of sequential pages in one go.
2355 */
2356 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
2357 struct pv_entry *pve;
2358 simple_lock(&pg->mdpage.pvh_slock);
2359 pve = pmap_remove_pv(pg, pmap, sva);
2360 pmap_free_pv(pmap, pve);
2361 pmap_vac_me_harder(pmap, pg, ptes, FALSE);
2362 simple_unlock(&pg->mdpage.pvh_slock);
2363 }
2364 }
2365 sva += NBPG;
2366 pte++;
2367 }
2368
2369 pmap_unmap_ptes(pmap);
2370 /*
2371 * Now, if we've fallen through down to here, chances are that there
2372 * are less than PMAP_REMOVE_CLEAN_LIST_SIZE mappings left.
2373 */
2374 if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
2375 u_int cnt;
2376
2377 for (cnt = 0; cnt < cleanlist_idx; cnt++) {
2378 if (pmap_active) {
2379 cpu_idcache_wbinv_range(cleanlist[cnt].va,
2380 NBPG);
2381 *cleanlist[cnt].pte = 0;
2382 cpu_tlb_flushID_SE(cleanlist[cnt].va);
2383 } else
2384 *cleanlist[cnt].pte = 0;
2385 pmap_pte_delref(pmap, cleanlist[cnt].va);
2386 }
2387 }
2388 PMAP_MAP_TO_HEAD_UNLOCK();
2389 }
2390
2391 /*
2392 * Routine: pmap_remove_all
2393 * Function:
2394 * Removes this physical page from
2395 * all physical maps in which it resides.
2396 * Reflects back modify bits to the pager.
2397 */
2398
2399 static void
2400 pmap_remove_all(struct vm_page *pg)
2401 {
2402 struct pv_entry *pv, *npv;
2403 struct pmap *pmap;
2404 pt_entry_t *pte, *ptes;
2405
2406 PDEBUG(0, printf("pmap_remove_all: pa=%lx ", VM_PAGE_TO_PHYS(pg)));
2407
2408 /* set vm_page => pmap locking */
2409 PMAP_HEAD_TO_MAP_LOCK();
2410
2411 simple_lock(&pg->mdpage.pvh_slock);
2412
2413 pv = pg->mdpage.pvh_list;
2414 if (pv == NULL) {
2415 PDEBUG(0, printf("free page\n"));
2416 simple_unlock(&pg->mdpage.pvh_slock);
2417 PMAP_HEAD_TO_MAP_UNLOCK();
2418 return;
2419 }
2420 pmap_clean_page(pv, FALSE);
2421
2422 while (pv) {
2423 pmap = pv->pv_pmap;
2424 ptes = pmap_map_ptes(pmap);
2425 pte = &ptes[arm_btop(pv->pv_va)];
2426
2427 PDEBUG(0, printf("[%p,%08x,%08lx,%08x] ", pmap, *pte,
2428 pv->pv_va, pv->pv_flags));
2429 #ifdef DEBUG
2430 if (pmap_pde_page(pmap_pde(pmap, pv->pv_va)) == 0 ||
2431 pmap_pte_v(pte) == 0 ||
2432 pmap_pte_pa(pte) != VM_PAGE_TO_PHYS(pg))
2433 panic("pmap_remove_all: bad mapping");
2434 #endif /* DEBUG */
2435
2436 /*
2437 * Update statistics
2438 */
2439 --pmap->pm_stats.resident_count;
2440
2441 /* Wired bit */
2442 if (pv->pv_flags & PVF_WIRED)
2443 --pmap->pm_stats.wired_count;
2444
2445 /*
2446 * Invalidate the PTEs.
2447 * XXX: should cluster them up and invalidate as many
2448 * as possible at once.
2449 */
2450
2451 #ifdef needednotdone
2452 reduce wiring count on page table pages as references drop
2453 #endif
2454
2455 *pte = 0;
2456 pmap_pte_delref(pmap, pv->pv_va);
2457
2458 npv = pv->pv_next;
2459 pmap_free_pv(pmap, pv);
2460 pv = npv;
2461 pmap_unmap_ptes(pmap);
2462 }
2463 pg->mdpage.pvh_list = NULL;
2464 simple_unlock(&pg->mdpage.pvh_slock);
2465 PMAP_HEAD_TO_MAP_UNLOCK();
2466
2467 PDEBUG(0, printf("done\n"));
2468 cpu_tlb_flushID();
2469 cpu_cpwait();
2470 }
2471
2472
2473 /*
2474 * Set the physical protection on the specified range of this map as requested.
2475 */
2476
2477 void
2478 pmap_protect(struct pmap *pmap, vaddr_t sva, vaddr_t eva, vm_prot_t prot)
2479 {
2480 pt_entry_t *pte = NULL, *ptes;
2481 struct vm_page *pg;
2482 int armprot;
2483 int flush = 0;
2484 paddr_t pa;
2485
2486 PDEBUG(0, printf("pmap_protect: pmap=%p %08lx->%08lx %x\n",
2487 pmap, sva, eva, prot));
2488
2489 if (~prot & VM_PROT_READ) {
2490 /* Just remove the mappings. */
2491 pmap_remove(pmap, sva, eva);
2492 /* pmap_update not needed as it should be called by the caller
2493 * of pmap_protect */
2494 return;
2495 }
2496 if (prot & VM_PROT_WRITE) {
2497 /*
2498 * If this is a read->write transition, just ignore it and let
2499 * uvm_fault() take care of it later.
2500 */
2501 return;
2502 }
2503
2504 /* Need to lock map->head */
2505 PMAP_MAP_TO_HEAD_LOCK();
2506
2507 ptes = pmap_map_ptes(pmap);
2508 /*
2509 * We need to acquire a pointer to a page table page before entering
2510 * the following loop.
2511 */
2512 while (sva < eva) {
2513 if (pmap_pde_page(pmap_pde(pmap, sva)))
2514 break;
2515 sva = (sva & L1_S_FRAME) + L1_S_SIZE;
2516 }
2517
2518 pte = &ptes[arm_btop(sva)];
2519
2520 while (sva < eva) {
2521 /* only check once in a while */
2522 if ((sva & L2_ADDR_BITS) == 0) {
2523 if (!pmap_pde_page(pmap_pde(pmap, sva))) {
2524 /* We can race ahead here, to the next pde. */
2525 sva += L1_S_SIZE;
2526 pte += arm_btop(L1_S_SIZE);
2527 continue;
2528 }
2529 }
2530
2531 if (!pmap_pte_v(pte))
2532 goto next;
2533
2534 flush = 1;
2535
2536 armprot = 0;
2537 if (sva < VM_MAXUSER_ADDRESS)
2538 armprot |= L2_S_PROT_U;
2539 else if (sva < VM_MAX_ADDRESS)
2540 armprot |= L2_S_PROT_W; /* XXX Ekk what is this ? */
2541 *pte = (*pte & 0xfffff00f) | armprot;
2542
2543 pa = pmap_pte_pa(pte);
2544
2545 /* Get the physical page index */
2546
2547 /* Clear write flag */
2548 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
2549 simple_lock(&pg->mdpage.pvh_slock);
2550 (void) pmap_modify_pv(pmap, sva, pg, PVF_WRITE, 0);
2551 pmap_vac_me_harder(pmap, pg, ptes, FALSE);
2552 simple_unlock(&pg->mdpage.pvh_slock);
2553 }
2554
2555 next:
2556 sva += NBPG;
2557 pte++;
2558 }
2559 pmap_unmap_ptes(pmap);
2560 PMAP_MAP_TO_HEAD_UNLOCK();
2561 if (flush)
2562 cpu_tlb_flushID();
2563 }
2564
2565 /*
2566 * void pmap_enter(struct pmap *pmap, vaddr_t va, paddr_t pa, vm_prot_t prot,
2567 * int flags)
2568 *
2569 * Insert the given physical page (p) at
2570 * the specified virtual address (v) in the
2571 * target physical map with the protection requested.
2572 *
2573 * If specified, the page will be wired down, meaning
2574 * that the related pte can not be reclaimed.
2575 *
2576 * NB: This is the only routine which MAY NOT lazy-evaluate
2577 * or lose information. That is, this routine must actually
2578 * insert this page into the given map NOW.
2579 */
2580
2581 int
2582 pmap_enter(struct pmap *pmap, vaddr_t va, paddr_t pa, vm_prot_t prot,
2583 int flags)
2584 {
2585 pt_entry_t *ptes, opte, npte;
2586 paddr_t opa;
2587 boolean_t wired = (flags & PMAP_WIRED) != 0;
2588 struct vm_page *pg;
2589 struct pv_entry *pve;
2590 int error, nflags;
2591
2592 PDEBUG(5, printf("pmap_enter: V%08lx P%08lx in pmap %p prot=%08x, wired = %d\n",
2593 va, pa, pmap, prot, wired));
2594
2595 #ifdef DIAGNOSTIC
2596 /* Valid address ? */
2597 if (va >= (pmap_curmaxkvaddr))
2598 panic("pmap_enter: too big");
2599 if (pmap != pmap_kernel() && va != 0) {
2600 if (va < VM_MIN_ADDRESS || va >= VM_MAXUSER_ADDRESS)
2601 panic("pmap_enter: kernel page in user map");
2602 } else {
2603 if (va >= VM_MIN_ADDRESS && va < VM_MAXUSER_ADDRESS)
2604 panic("pmap_enter: user page in kernel map");
2605 if (va >= VM_MAXUSER_ADDRESS && va < VM_MAX_ADDRESS)
2606 panic("pmap_enter: entering PT page");
2607 }
2608 #endif
2609
2610 KDASSERT(((va | pa) & PGOFSET) == 0);
2611
2612 /*
2613 * Get a pointer to the page. Later on in this function, we
2614 * test for a managed page by checking pg != NULL.
2615 */
2616 pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
2617
2618 /* get lock */
2619 PMAP_MAP_TO_HEAD_LOCK();
2620
2621 /*
2622 * map the ptes. If there's not already an L2 table for this
2623 * address, allocate one.
2624 */
2625 ptes = pmap_map_ptes(pmap); /* locks pmap */
2626 if (pmap_pde_v(pmap_pde(pmap, va)) == 0) {
2627 struct vm_page *ptp;
2628
2629 /* kernel should be pre-grown */
2630 KASSERT(pmap != pmap_kernel());
2631
2632 /* if failure is allowed then don't try too hard */
2633 ptp = pmap_get_ptp(pmap, va & L1_S_FRAME);
2634 if (ptp == NULL) {
2635 if (flags & PMAP_CANFAIL) {
2636 error = ENOMEM;
2637 goto out;
2638 }
2639 panic("pmap_enter: get ptp failed");
2640 }
2641 }
2642 opte = ptes[arm_btop(va)];
2643
2644 nflags = 0;
2645 if (prot & VM_PROT_WRITE)
2646 nflags |= PVF_WRITE;
2647 if (wired)
2648 nflags |= PVF_WIRED;
2649
2650 /* Is the pte valid ? If so then this page is already mapped */
2651 if (l2pte_valid(opte)) {
2652 /* Get the physical address of the current page mapped */
2653 opa = l2pte_pa(opte);
2654
2655 /* Are we mapping the same page ? */
2656 if (opa == pa) {
2657 /* Has the wiring changed ? */
2658 if (pg != NULL) {
2659 simple_lock(&pg->mdpage.pvh_slock);
2660 (void) pmap_modify_pv(pmap, va, pg,
2661 PVF_WRITE | PVF_WIRED, nflags);
2662 simple_unlock(&pg->mdpage.pvh_slock);
2663 }
2664 } else {
2665 struct vm_page *opg;
2666
2667 /* We are replacing the page with a new one. */
2668 cpu_idcache_wbinv_range(va, NBPG);
2669
2670 /*
2671 * If it is part of our managed memory then we
2672 * must remove it from the PV list
2673 */
2674 if ((opg = PHYS_TO_VM_PAGE(opa)) != NULL) {
2675 simple_lock(&opg->mdpage.pvh_slock);
2676 pve = pmap_remove_pv(opg, pmap, va);
2677 simple_unlock(&opg->mdpage.pvh_slock);
2678 } else {
2679 pve = NULL;
2680 }
2681
2682 goto enter;
2683 }
2684 } else {
2685 opa = 0;
2686 pve = NULL;
2687 pmap_pte_addref(pmap, va);
2688
2689 /* pte is not valid so we must be hooking in a new page */
2690 ++pmap->pm_stats.resident_count;
2691
2692 enter:
2693 /*
2694 * Enter on the PV list if part of our managed memory
2695 */
2696 if (pg != NULL) {
2697 if (pve == NULL) {
2698 pve = pmap_alloc_pv(pmap, ALLOCPV_NEED);
2699 if (pve == NULL) {
2700 if (flags & PMAP_CANFAIL) {
2701 error = ENOMEM;
2702 goto out;
2703 }
2704 panic("pmap_enter: no pv entries "
2705 "available");
2706 }
2707 }
2708 /* enter_pv locks pvh when adding */
2709 pmap_enter_pv(pg, pve, pmap, va, NULL, nflags);
2710 } else {
2711 if (pve != NULL)
2712 pmap_free_pv(pmap, pve);
2713 }
2714 }
2715
2716 /* Construct the pte, giving the correct access. */
2717 npte = pa;
2718
2719 /* VA 0 is magic. */
2720 if (pmap != pmap_kernel() && va != vector_page)
2721 npte |= L2_S_PROT_U;
2722
2723 if (pg != NULL) {
2724 #ifdef DIAGNOSTIC
2725 if ((flags & VM_PROT_ALL) & ~prot)
2726 panic("pmap_enter: access_type exceeds prot");
2727 #endif
2728 npte |= pte_l2_s_cache_mode;
2729 if (flags & VM_PROT_WRITE) {
2730 npte |= L2_S_PROTO | L2_S_PROT_W;
2731 pg->mdpage.pvh_attrs |= PVF_REF | PVF_MOD;
2732 } else if (flags & VM_PROT_ALL) {
2733 npte |= L2_S_PROTO;
2734 pg->mdpage.pvh_attrs |= PVF_REF;
2735 } else
2736 npte |= L2_TYPE_INV;
2737 } else {
2738 if (prot & VM_PROT_WRITE)
2739 npte |= L2_S_PROTO | L2_S_PROT_W;
2740 else if (prot & VM_PROT_ALL)
2741 npte |= L2_S_PROTO;
2742 else
2743 npte |= L2_TYPE_INV;
2744 }
2745
2746 ptes[arm_btop(va)] = npte;
2747
2748 if (pg != NULL) {
2749 simple_lock(&pg->mdpage.pvh_slock);
2750 pmap_vac_me_harder(pmap, pg, ptes, pmap_is_curpmap(pmap));
2751 simple_unlock(&pg->mdpage.pvh_slock);
2752 }
2753
2754 /* Better flush the TLB ... */
2755 cpu_tlb_flushID_SE(va);
2756 error = 0;
2757 out:
2758 pmap_unmap_ptes(pmap); /* unlocks pmap */
2759 PMAP_MAP_TO_HEAD_UNLOCK();
2760
2761 return error;
2762 }
2763
2764 /*
2765 * pmap_kenter_pa: enter a kernel mapping
2766 *
2767 * => no need to lock anything assume va is already allocated
2768 * => should be faster than normal pmap enter function
2769 */
2770 void
2771 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot)
2772 {
2773 pt_entry_t *pte;
2774
2775 pte = vtopte(va);
2776 KASSERT(!pmap_pte_v(pte));
2777
2778 /* XXX r/w! */
2779 *pte = L2_S_PROTO | pa |
2780 L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE) |
2781 pte_l2_s_cache_mode;
2782 }
2783
2784 void
2785 pmap_kremove(vaddr_t va, vsize_t len)
2786 {
2787 pt_entry_t *pte;
2788
2789 for (len >>= PAGE_SHIFT; len > 0; len--, va += PAGE_SIZE) {
2790
2791 /*
2792 * We assume that we will only be called with small
2793 * regions of memory.
2794 */
2795
2796 KASSERT(pmap_pde_page(pmap_pde(pmap_kernel(), va)));
2797 pte = vtopte(va);
2798 cpu_idcache_wbinv_range(va, PAGE_SIZE);
2799 *pte = 0;
2800 cpu_tlb_flushID_SE(va);
2801 }
2802 }
2803
2804 /*
2805 * pmap_page_protect:
2806 *
2807 * Lower the permission for all mappings to a given page.
2808 */
2809
2810 void
2811 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
2812 {
2813
2814 PDEBUG(0, printf("pmap_page_protect(pa=%lx, prot=%d)\n",
2815 VM_PAGE_TO_PHYS(pg), prot));
2816
2817 switch(prot) {
2818 case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
2819 case VM_PROT_READ|VM_PROT_WRITE:
2820 return;
2821
2822 case VM_PROT_READ:
2823 case VM_PROT_READ|VM_PROT_EXECUTE:
2824 pmap_clearbit(pg, PVF_WRITE);
2825 break;
2826
2827 default:
2828 pmap_remove_all(pg);
2829 break;
2830 }
2831 }
2832
2833
2834 /*
2835 * Routine: pmap_unwire
2836 * Function: Clear the wired attribute for a map/virtual-address
2837 * pair.
2838 * In/out conditions:
2839 * The mapping must already exist in the pmap.
2840 */
2841
2842 void
2843 pmap_unwire(struct pmap *pmap, vaddr_t va)
2844 {
2845 pt_entry_t *ptes;
2846 struct vm_page *pg;
2847 paddr_t pa;
2848
2849 PMAP_MAP_TO_HEAD_LOCK();
2850 ptes = pmap_map_ptes(pmap); /* locks pmap */
2851
2852 if (pmap_pde_v(pmap_pde(pmap, va))) {
2853 #ifdef DIAGNOSTIC
2854 if (l2pte_valid(ptes[arm_btop(va)]) == 0)
2855 panic("pmap_unwire: invalid L2 PTE");
2856 #endif
2857 /* Extract the physical address of the page */
2858 pa = l2pte_pa(ptes[arm_btop(va)]);
2859
2860 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
2861 goto out;
2862
2863 /* Update the wired bit in the pv entry for this page. */
2864 simple_lock(&pg->mdpage.pvh_slock);
2865 (void) pmap_modify_pv(pmap, va, pg, PVF_WIRED, 0);
2866 simple_unlock(&pg->mdpage.pvh_slock);
2867 }
2868 #ifdef DIAGNOSTIC
2869 else {
2870 panic("pmap_unwire: invalid L1 PTE");
2871 }
2872 #endif
2873 out:
2874 pmap_unmap_ptes(pmap); /* unlocks pmap */
2875 PMAP_MAP_TO_HEAD_UNLOCK();
2876 }
2877
2878 /*
2879 * Routine: pmap_extract
2880 * Function:
2881 * Extract the physical page address associated
2882 * with the given map/virtual_address pair.
2883 */
2884 boolean_t
2885 pmap_extract(struct pmap *pmap, vaddr_t va, paddr_t *pap)
2886 {
2887 pd_entry_t *pde;
2888 pt_entry_t *pte, *ptes;
2889 paddr_t pa;
2890
2891 PDEBUG(5, printf("pmap_extract: pmap=%p, va=0x%08lx -> ", pmap, va));
2892
2893 ptes = pmap_map_ptes(pmap); /* locks pmap */
2894
2895 pde = pmap_pde(pmap, va);
2896 pte = &ptes[arm_btop(va)];
2897
2898 if (pmap_pde_section(pde)) {
2899 pa = (*pde & L1_S_FRAME) | (va & L1_S_OFFSET);
2900 PDEBUG(5, printf("section pa=0x%08lx\n", pa));
2901 goto out;
2902 } else if (pmap_pde_page(pde) == 0 || pmap_pte_v(pte) == 0) {
2903 PDEBUG(5, printf("no mapping\n"));
2904 goto failed;
2905 }
2906
2907 if ((*pte & L2_TYPE_MASK) == L2_TYPE_L) {
2908 pa = (*pte & L2_L_FRAME) | (va & L2_L_OFFSET);
2909 PDEBUG(5, printf("large page pa=0x%08lx\n", pa));
2910 goto out;
2911 }
2912
2913 pa = (*pte & L2_S_FRAME) | (va & L2_S_OFFSET);
2914 PDEBUG(5, printf("small page pa=0x%08lx\n", pa));
2915
2916 out:
2917 if (pap != NULL)
2918 *pap = pa;
2919
2920 pmap_unmap_ptes(pmap); /* unlocks pmap */
2921 return (TRUE);
2922
2923 failed:
2924 pmap_unmap_ptes(pmap); /* unlocks pmap */
2925 return (FALSE);
2926 }
2927
2928
2929 /*
2930 * pmap_copy:
2931 *
2932 * Copy the range specified by src_addr/len from the source map to the
2933 * range dst_addr/len in the destination map.
2934 *
2935 * This routine is only advisory and need not do anything.
2936 */
2937 /* Call deleted in <arm/arm32/pmap.h> */
2938
2939 #if defined(PMAP_DEBUG)
2940 void
2941 pmap_dump_pvlist(phys, m)
2942 vaddr_t phys;
2943 char *m;
2944 {
2945 struct vm_page *pg;
2946 struct pv_entry *pv;
2947
2948 if ((pg = PHYS_TO_VM_PAGE(phys)) == NULL) {
2949 printf("INVALID PA\n");
2950 return;
2951 }
2952 simple_lock(&pg->mdpage.pvh_slock);
2953 printf("%s %08lx:", m, phys);
2954 if (pg->mdpage.pvh_list == NULL) {
2955 printf(" no mappings\n");
2956 return;
2957 }
2958
2959 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next)
2960 printf(" pmap %p va %08lx flags %08x", pv->pv_pmap,
2961 pv->pv_va, pv->pv_flags);
2962
2963 printf("\n");
2964 simple_unlock(&pg->mdpage.pvh_slock);
2965 }
2966
2967 #endif /* PMAP_DEBUG */
2968
2969 static pt_entry_t *
2970 pmap_map_ptes(struct pmap *pmap)
2971 {
2972 struct proc *p;
2973
2974 /* the kernel's pmap is always accessible */
2975 if (pmap == pmap_kernel()) {
2976 return (pt_entry_t *)PTE_BASE;
2977 }
2978
2979 if (pmap_is_curpmap(pmap)) {
2980 simple_lock(&pmap->pm_obj.vmobjlock);
2981 return (pt_entry_t *)PTE_BASE;
2982 }
2983
2984 p = curproc;
2985 KDASSERT(p != NULL);
2986
2987 /* need to lock both curpmap and pmap: use ordered locking */
2988 if ((vaddr_t) pmap < (vaddr_t) p->p_vmspace->vm_map.pmap) {
2989 simple_lock(&pmap->pm_obj.vmobjlock);
2990 simple_lock(&p->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
2991 } else {
2992 simple_lock(&p->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
2993 simple_lock(&pmap->pm_obj.vmobjlock);
2994 }
2995
2996 pmap_map_in_l1(p->p_vmspace->vm_map.pmap, APTE_BASE, pmap->pm_pptpt,
2997 FALSE);
2998 cpu_tlb_flushD();
2999 cpu_cpwait();
3000 return (pt_entry_t *)APTE_BASE;
3001 }
3002
3003 /*
3004 * pmap_unmap_ptes: unlock the PTE mapping of "pmap"
3005 */
3006
3007 static void
3008 pmap_unmap_ptes(struct pmap *pmap)
3009 {
3010
3011 if (pmap == pmap_kernel()) {
3012 return;
3013 }
3014 if (pmap_is_curpmap(pmap)) {
3015 simple_unlock(&pmap->pm_obj.vmobjlock);
3016 } else {
3017 KDASSERT(curproc != NULL);
3018 simple_unlock(&pmap->pm_obj.vmobjlock);
3019 simple_unlock(
3020 &curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3021 }
3022 }
3023
3024 /*
3025 * Modify pte bits for all ptes corresponding to the given physical address.
3026 * We use `maskbits' rather than `clearbits' because we're always passing
3027 * constants and the latter would require an extra inversion at run-time.
3028 */
3029
3030 static void
3031 pmap_clearbit(struct vm_page *pg, u_int maskbits)
3032 {
3033 struct pv_entry *pv;
3034 pt_entry_t *ptes;
3035 vaddr_t va;
3036 int tlbentry;
3037
3038 PDEBUG(1, printf("pmap_clearbit: pa=%08lx mask=%08x\n",
3039 VM_PAGE_TO_PHYS(pg), maskbits));
3040
3041 tlbentry = 0;
3042
3043 PMAP_HEAD_TO_MAP_LOCK();
3044 simple_lock(&pg->mdpage.pvh_slock);
3045
3046 /*
3047 * Clear saved attributes (modify, reference)
3048 */
3049 pg->mdpage.pvh_attrs &= ~maskbits;
3050
3051 if (pg->mdpage.pvh_list == NULL) {
3052 simple_unlock(&pg->mdpage.pvh_slock);
3053 PMAP_HEAD_TO_MAP_UNLOCK();
3054 return;
3055 }
3056
3057 /*
3058 * Loop over all current mappings setting/clearing as appropos
3059 */
3060 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
3061 va = pv->pv_va;
3062 pv->pv_flags &= ~maskbits;
3063 ptes = pmap_map_ptes(pv->pv_pmap); /* locks pmap */
3064 KASSERT(pmap_pde_v(pmap_pde(pv->pv_pmap, va)));
3065 if (maskbits & (PVF_WRITE|PVF_MOD)) {
3066 if ((pv->pv_flags & PVF_NC)) {
3067 /*
3068 * Entry is not cacheable: reenable
3069 * the cache, nothing to flush
3070 *
3071 * Don't turn caching on again if this
3072 * is a modified emulation. This
3073 * would be inconsitent with the
3074 * settings created by
3075 * pmap_vac_me_harder().
3076 *
3077 * There's no need to call
3078 * pmap_vac_me_harder() here: all
3079 * pages are loosing their write
3080 * permission.
3081 *
3082 */
3083 if (maskbits & PVF_WRITE) {
3084 ptes[arm_btop(va)] |=
3085 pte_l2_s_cache_mode;
3086 pv->pv_flags &= ~PVF_NC;
3087 }
3088 } else if (pmap_is_curpmap(pv->pv_pmap)) {
3089 /*
3090 * Entry is cacheable: check if pmap is
3091 * current if it is flush it,
3092 * otherwise it won't be in the cache
3093 */
3094 cpu_idcache_wbinv_range(pv->pv_va, NBPG);
3095 }
3096
3097 /* make the pte read only */
3098 ptes[arm_btop(va)] &= ~L2_S_PROT_W;
3099 }
3100
3101 if (maskbits & PVF_REF)
3102 ptes[arm_btop(va)] =
3103 (ptes[arm_btop(va)] & ~L2_TYPE_MASK) | L2_TYPE_INV;
3104
3105 if (pmap_is_curpmap(pv->pv_pmap)) {
3106 /*
3107 * if we had cacheable pte's we'd clean the
3108 * pte out to memory here
3109 *
3110 * flush tlb entry as it's in the current pmap
3111 */
3112 cpu_tlb_flushID_SE(pv->pv_va);
3113 }
3114 pmap_unmap_ptes(pv->pv_pmap); /* unlocks pmap */
3115 }
3116 cpu_cpwait();
3117
3118 simple_unlock(&pg->mdpage.pvh_slock);
3119 PMAP_HEAD_TO_MAP_UNLOCK();
3120 }
3121
3122 /*
3123 * pmap_clear_modify:
3124 *
3125 * Clear the "modified" attribute for a page.
3126 */
3127 boolean_t
3128 pmap_clear_modify(struct vm_page *pg)
3129 {
3130 boolean_t rv;
3131
3132 if (pg->mdpage.pvh_attrs & PVF_MOD) {
3133 rv = TRUE;
3134 pmap_clearbit(pg, PVF_MOD);
3135 } else
3136 rv = FALSE;
3137
3138 PDEBUG(0, printf("pmap_clear_modify pa=%08lx -> %d\n",
3139 VM_PAGE_TO_PHYS(pg), rv));
3140
3141 return (rv);
3142 }
3143
3144 /*
3145 * pmap_clear_reference:
3146 *
3147 * Clear the "referenced" attribute for a page.
3148 */
3149 boolean_t
3150 pmap_clear_reference(struct vm_page *pg)
3151 {
3152 boolean_t rv;
3153
3154 if (pg->mdpage.pvh_attrs & PVF_REF) {
3155 rv = TRUE;
3156 pmap_clearbit(pg, PVF_REF);
3157 } else
3158 rv = FALSE;
3159
3160 PDEBUG(0, printf("pmap_clear_reference pa=%08lx -> %d\n",
3161 VM_PAGE_TO_PHYS(pg), rv));
3162
3163 return (rv);
3164 }
3165
3166 /*
3167 * pmap_is_modified:
3168 *
3169 * Test if a page has the "modified" attribute.
3170 */
3171 /* See <arm/arm32/pmap.h> */
3172
3173 /*
3174 * pmap_is_referenced:
3175 *
3176 * Test if a page has the "referenced" attribute.
3177 */
3178 /* See <arm/arm32/pmap.h> */
3179
3180 int
3181 pmap_modified_emulation(struct pmap *pmap, vaddr_t va)
3182 {
3183 pt_entry_t *ptes;
3184 struct vm_page *pg;
3185 paddr_t pa;
3186 u_int flags;
3187 int rv = 0;
3188
3189 PDEBUG(2, printf("pmap_modified_emulation\n"));
3190
3191 PMAP_MAP_TO_HEAD_LOCK();
3192 ptes = pmap_map_ptes(pmap); /* locks pmap */
3193
3194 if (pmap_pde_v(pmap_pde(pmap, va)) == 0) {
3195 PDEBUG(2, printf("L1 PTE invalid\n"));
3196 goto out;
3197 }
3198
3199 PDEBUG(1, printf("pte=%08x\n", ptes[arm_btop(va)]));
3200
3201 /* Check for a invalid pte */
3202 if (l2pte_valid(ptes[arm_btop(va)]) == 0)
3203 goto out;
3204
3205 /* This can happen if user code tries to access kernel memory. */
3206 if ((ptes[arm_btop(va)] & L2_S_PROT_W) != 0)
3207 goto out;
3208
3209 /* Extract the physical address of the page */
3210 pa = l2pte_pa(ptes[arm_btop(va)]);
3211 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3212 goto out;
3213
3214 /* Get the current flags for this page. */
3215 simple_lock(&pg->mdpage.pvh_slock);
3216
3217 flags = pmap_modify_pv(pmap, va, pg, 0, 0);
3218 PDEBUG(2, printf("pmap_modified_emulation: flags = %08x\n", flags));
3219
3220 /*
3221 * Do the flags say this page is writable ? If not then it is a
3222 * genuine write fault. If yes then the write fault is our fault
3223 * as we did not reflect the write access in the PTE. Now we know
3224 * a write has occurred we can correct this and also set the
3225 * modified bit
3226 */
3227 if (~flags & PVF_WRITE) {
3228 simple_unlock(&pg->mdpage.pvh_slock);
3229 goto out;
3230 }
3231
3232 PDEBUG(0,
3233 printf("pmap_modified_emulation: Got a hit va=%08lx, pte = %08x\n",
3234 va, ptes[arm_btop(va)]));
3235 pg->mdpage.pvh_attrs |= PVF_REF | PVF_MOD;
3236
3237 /*
3238 * Re-enable write permissions for the page. No need to call
3239 * pmap_vac_me_harder(), since this is just a
3240 * modified-emulation fault, and the PVF_WRITE bit isn't changing.
3241 * We've already set the cacheable bits based on the assumption
3242 * that we can write to this page.
3243 */
3244 ptes[arm_btop(va)] =
3245 (ptes[arm_btop(va)] & ~L2_TYPE_MASK) | L2_S_PROTO | L2_S_PROT_W;
3246 PDEBUG(0, printf("->(%08x)\n", ptes[arm_btop(va)]));
3247
3248 simple_unlock(&pg->mdpage.pvh_slock);
3249
3250 cpu_tlb_flushID_SE(va);
3251 cpu_cpwait();
3252 rv = 1;
3253 out:
3254 pmap_unmap_ptes(pmap); /* unlocks pmap */
3255 PMAP_MAP_TO_HEAD_UNLOCK();
3256 return (rv);
3257 }
3258
3259 int
3260 pmap_handled_emulation(struct pmap *pmap, vaddr_t va)
3261 {
3262 pt_entry_t *ptes;
3263 struct vm_page *pg;
3264 paddr_t pa;
3265 int rv = 0;
3266
3267 PDEBUG(2, printf("pmap_handled_emulation\n"));
3268
3269 PMAP_MAP_TO_HEAD_LOCK();
3270 ptes = pmap_map_ptes(pmap); /* locks pmap */
3271
3272 if (pmap_pde_v(pmap_pde(pmap, va)) == 0) {
3273 PDEBUG(2, printf("L1 PTE invalid\n"));
3274 goto out;
3275 }
3276
3277 PDEBUG(1, printf("pte=%08x\n", ptes[arm_btop(va)]));
3278
3279 /* Check for invalid pte */
3280 if (l2pte_valid(ptes[arm_btop(va)]) == 0)
3281 goto out;
3282
3283 /* This can happen if user code tries to access kernel memory. */
3284 if ((ptes[arm_btop(va)] & L2_TYPE_MASK) != L2_TYPE_INV)
3285 goto out;
3286
3287 /* Extract the physical address of the page */
3288 pa = l2pte_pa(ptes[arm_btop(va)]);
3289 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3290 goto out;
3291
3292 simple_lock(&pg->mdpage.pvh_slock);
3293
3294 /*
3295 * Ok we just enable the pte and mark the attibs as handled
3296 * XXX Should we traverse the PV list and enable all PTEs?
3297 */
3298 PDEBUG(0,
3299 printf("pmap_handled_emulation: Got a hit va=%08lx pte = %08x\n",
3300 va, ptes[arm_btop(va)]));
3301 pg->mdpage.pvh_attrs |= PVF_REF;
3302
3303 ptes[arm_btop(va)] = (ptes[arm_btop(va)] & ~L2_TYPE_MASK) | L2_S_PROTO;
3304 PDEBUG(0, printf("->(%08x)\n", ptes[arm_btop(va)]));
3305
3306 simple_unlock(&pg->mdpage.pvh_slock);
3307
3308 cpu_tlb_flushID_SE(va);
3309 cpu_cpwait();
3310 rv = 1;
3311 out:
3312 pmap_unmap_ptes(pmap); /* unlocks pmap */
3313 PMAP_MAP_TO_HEAD_UNLOCK();
3314 return (rv);
3315 }
3316
3317 /*
3318 * pmap_collect: free resources held by a pmap
3319 *
3320 * => optional function.
3321 * => called when a process is swapped out to free memory.
3322 */
3323
3324 void
3325 pmap_collect(struct pmap *pmap)
3326 {
3327 }
3328
3329 /*
3330 * Routine: pmap_procwr
3331 *
3332 * Function:
3333 * Synchronize caches corresponding to [addr, addr+len) in p.
3334 *
3335 */
3336 void
3337 pmap_procwr(struct proc *p, vaddr_t va, int len)
3338 {
3339 /* We only need to do anything if it is the current process. */
3340 if (p == curproc)
3341 cpu_icache_sync_range(va, len);
3342 }
3343 /*
3344 * PTP functions
3345 */
3346
3347 /*
3348 * pmap_get_ptp: get a PTP (if there isn't one, allocate a new one)
3349 *
3350 * => pmap should NOT be pmap_kernel()
3351 * => pmap should be locked
3352 */
3353
3354 static struct vm_page *
3355 pmap_get_ptp(struct pmap *pmap, vaddr_t va)
3356 {
3357 struct vm_page *ptp;
3358
3359 if (pmap_pde_page(pmap_pde(pmap, va))) {
3360
3361 /* valid... check hint (saves us a PA->PG lookup) */
3362 if (pmap->pm_ptphint &&
3363 (pmap->pm_pdir[pmap_pdei(va)] & L2_S_FRAME) ==
3364 VM_PAGE_TO_PHYS(pmap->pm_ptphint))
3365 return (pmap->pm_ptphint);
3366 ptp = uvm_pagelookup(&pmap->pm_obj, va);
3367 #ifdef DIAGNOSTIC
3368 if (ptp == NULL)
3369 panic("pmap_get_ptp: unmanaged user PTP");
3370 #endif
3371 pmap->pm_ptphint = ptp;
3372 return(ptp);
3373 }
3374
3375 /* allocate a new PTP (updates ptphint) */
3376 return(pmap_alloc_ptp(pmap, va));
3377 }
3378
3379 /*
3380 * pmap_alloc_ptp: allocate a PTP for a PMAP
3381 *
3382 * => pmap should already be locked by caller
3383 * => we use the ptp's wire_count to count the number of active mappings
3384 * in the PTP (we start it at one to prevent any chance this PTP
3385 * will ever leak onto the active/inactive queues)
3386 */
3387
3388 /*__inline */ static struct vm_page *
3389 pmap_alloc_ptp(struct pmap *pmap, vaddr_t va)
3390 {
3391 struct vm_page *ptp;
3392
3393 ptp = uvm_pagealloc(&pmap->pm_obj, va, NULL,
3394 UVM_PGA_USERESERVE|UVM_PGA_ZERO);
3395 if (ptp == NULL)
3396 return (NULL);
3397
3398 /* got one! */
3399 ptp->flags &= ~PG_BUSY; /* never busy */
3400 ptp->wire_count = 1; /* no mappings yet */
3401 pmap_map_in_l1(pmap, va, VM_PAGE_TO_PHYS(ptp), TRUE);
3402 pmap->pm_stats.resident_count++; /* count PTP as resident */
3403 pmap->pm_ptphint = ptp;
3404 return (ptp);
3405 }
3406
3407 vaddr_t
3408 pmap_growkernel(vaddr_t maxkvaddr)
3409 {
3410 struct pmap *kpm = pmap_kernel(), *pm;
3411 int s;
3412 paddr_t ptaddr;
3413 struct vm_page *ptp;
3414
3415 if (maxkvaddr <= pmap_curmaxkvaddr)
3416 goto out; /* we are OK */
3417 NPDEBUG(PDB_GROWKERN, printf("pmap_growkernel: growing kernel from %lx to %lx\n",
3418 pmap_curmaxkvaddr, maxkvaddr));
3419
3420 /*
3421 * whoops! we need to add kernel PTPs
3422 */
3423
3424 s = splhigh(); /* to be safe */
3425 simple_lock(&kpm->pm_obj.vmobjlock);
3426 /* due to the way the arm pmap works we map 4MB at a time */
3427 for (/*null*/ ; pmap_curmaxkvaddr < maxkvaddr;
3428 pmap_curmaxkvaddr += 4 * L1_S_SIZE) {
3429
3430 if (uvm.page_init_done == FALSE) {
3431
3432 /*
3433 * we're growing the kernel pmap early (from
3434 * uvm_pageboot_alloc()). this case must be
3435 * handled a little differently.
3436 */
3437
3438 if (uvm_page_physget(&ptaddr) == FALSE)
3439 panic("pmap_growkernel: out of memory");
3440 pmap_zero_page(ptaddr);
3441
3442 /* map this page in */
3443 pmap_map_in_l1(kpm, pmap_curmaxkvaddr, ptaddr, TRUE);
3444
3445 /* count PTP as resident */
3446 kpm->pm_stats.resident_count++;
3447 continue;
3448 }
3449
3450 /*
3451 * THIS *MUST* BE CODED SO AS TO WORK IN THE
3452 * pmap_initialized == FALSE CASE! WE MAY BE
3453 * INVOKED WHILE pmap_init() IS RUNNING!
3454 */
3455
3456 if ((ptp = pmap_alloc_ptp(kpm, pmap_curmaxkvaddr)) == NULL)
3457 panic("pmap_growkernel: alloc ptp failed");
3458
3459 /* distribute new kernel PTP to all active pmaps */
3460 simple_lock(&pmaps_lock);
3461 LIST_FOREACH(pm, &pmaps, pm_list) {
3462 pmap_map_in_l1(pm, pmap_curmaxkvaddr,
3463 VM_PAGE_TO_PHYS(ptp), TRUE);
3464 }
3465
3466 simple_unlock(&pmaps_lock);
3467 }
3468
3469 /*
3470 * flush out the cache, expensive but growkernel will happen so
3471 * rarely
3472 */
3473 cpu_tlb_flushD();
3474 cpu_cpwait();
3475
3476 simple_unlock(&kpm->pm_obj.vmobjlock);
3477 splx(s);
3478
3479 out:
3480 return (pmap_curmaxkvaddr);
3481 }
3482
3483 /************************ Utility routines ****************************/
3484
3485 /*
3486 * vector_page_setprot:
3487 *
3488 * Manipulate the protection of the vector page.
3489 */
3490 void
3491 vector_page_setprot(int prot)
3492 {
3493 pt_entry_t *pte;
3494
3495 pte = vtopte(vector_page);
3496
3497 *pte = (*pte & ~L1_S_PROT_MASK) | L2_S_PROT(PTE_KERNEL, prot);
3498 cpu_tlb_flushD_SE(vector_page);
3499 cpu_cpwait();
3500 }
3501
3502 /************************ Bootstrapping routines ****************************/
3503
3504 /*
3505 * This list exists for the benefit of pmap_map_chunk(). It keeps track
3506 * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
3507 * find them as necessary.
3508 *
3509 * Note that the data on this list is not valid after initarm() returns.
3510 */
3511 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
3512
3513 static vaddr_t
3514 kernel_pt_lookup(paddr_t pa)
3515 {
3516 pv_addr_t *pv;
3517
3518 SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
3519 if (pv->pv_pa == pa)
3520 return (pv->pv_va);
3521 }
3522 return (0);
3523 }
3524
3525 /*
3526 * pmap_map_section:
3527 *
3528 * Create a single section mapping.
3529 */
3530 void
3531 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
3532 {
3533 pd_entry_t *pde = (pd_entry_t *) l1pt;
3534 pd_entry_t fl = (cache == PTE_CACHE) ? pte_l1_s_cache_mode : 0;
3535
3536 KASSERT(((va | pa) & L1_S_OFFSET) == 0);
3537
3538 pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
3539 L1_S_PROT(PTE_KERNEL, prot) | fl;
3540 }
3541
3542 /*
3543 * pmap_map_entry:
3544 *
3545 * Create a single page mapping.
3546 */
3547 void
3548 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
3549 {
3550 pd_entry_t *pde = (pd_entry_t *) l1pt;
3551 pt_entry_t fl = (cache == PTE_CACHE) ? pte_l2_s_cache_mode : 0;
3552 pt_entry_t *pte;
3553
3554 KASSERT(((va | pa) & PGOFSET) == 0);
3555
3556 if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
3557 panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
3558
3559 pte = (pt_entry_t *)
3560 kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
3561 if (pte == NULL)
3562 panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
3563
3564 pte[(va >> PGSHIFT) & 0x3ff] = L2_S_PROTO | pa |
3565 L2_S_PROT(PTE_KERNEL, prot) | fl;
3566 }
3567
3568 /*
3569 * pmap_link_l2pt:
3570 *
3571 * Link the L2 page table specified by "pa" into the L1
3572 * page table at the slot for "va".
3573 */
3574 void
3575 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
3576 {
3577 pd_entry_t *pde = (pd_entry_t *) l1pt;
3578 u_int slot = va >> L1_S_SHIFT;
3579
3580 KASSERT((l2pv->pv_pa & PGOFSET) == 0);
3581
3582 pde[slot + 0] = L1_C_PROTO | (l2pv->pv_pa + 0x000);
3583 pde[slot + 1] = L1_C_PROTO | (l2pv->pv_pa + 0x400);
3584 pde[slot + 2] = L1_C_PROTO | (l2pv->pv_pa + 0x800);
3585 pde[slot + 3] = L1_C_PROTO | (l2pv->pv_pa + 0xc00);
3586
3587 SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
3588 }
3589
3590 /*
3591 * pmap_map_chunk:
3592 *
3593 * Map a chunk of memory using the most efficient mappings
3594 * possible (section, large page, small page) into the
3595 * provided L1 and L2 tables at the specified virtual address.
3596 */
3597 vsize_t
3598 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
3599 int prot, int cache)
3600 {
3601 pd_entry_t *pde = (pd_entry_t *) l1pt;
3602 pt_entry_t *pte, fl;
3603 vsize_t resid;
3604 int i;
3605
3606 resid = (size + (NBPG - 1)) & ~(NBPG - 1);
3607
3608 if (l1pt == 0)
3609 panic("pmap_map_chunk: no L1 table provided");
3610
3611 #ifdef VERBOSE_INIT_ARM
3612 printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
3613 "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
3614 #endif
3615
3616 size = resid;
3617
3618 while (resid > 0) {
3619 /* See if we can use a section mapping. */
3620 if (((pa | va) & L1_S_OFFSET) == 0 &&
3621 resid >= L1_S_SIZE) {
3622 fl = (cache == PTE_CACHE) ? pte_l1_s_cache_mode : 0;
3623 #ifdef VERBOSE_INIT_ARM
3624 printf("S");
3625 #endif
3626 pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
3627 L1_S_PROT(PTE_KERNEL, prot) | fl;
3628 va += L1_S_SIZE;
3629 pa += L1_S_SIZE;
3630 resid -= L1_S_SIZE;
3631 continue;
3632 }
3633
3634 /*
3635 * Ok, we're going to use an L2 table. Make sure
3636 * one is actually in the corresponding L1 slot
3637 * for the current VA.
3638 */
3639 if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
3640 panic("pmap_map_chunk: no L2 table for VA 0x%08lx", va);
3641
3642 pte = (pt_entry_t *)
3643 kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
3644 if (pte == NULL)
3645 panic("pmap_map_chunk: can't find L2 table for VA"
3646 "0x%08lx", va);
3647
3648 /* See if we can use a L2 large page mapping. */
3649 if (((pa | va) & L2_L_OFFSET) == 0 &&
3650 resid >= L2_L_SIZE) {
3651 fl = (cache == PTE_CACHE) ? pte_l2_l_cache_mode : 0;
3652 #ifdef VERBOSE_INIT_ARM
3653 printf("L");
3654 #endif
3655 for (i = 0; i < 16; i++) {
3656 pte[((va >> PGSHIFT) & 0x3f0) + i] =
3657 L2_L_PROTO | pa |
3658 L2_L_PROT(PTE_KERNEL, prot) | fl;
3659 }
3660 va += L2_L_SIZE;
3661 pa += L2_L_SIZE;
3662 resid -= L2_L_SIZE;
3663 continue;
3664 }
3665
3666 /* Use a small page mapping. */
3667 fl = (cache == PTE_CACHE) ? pte_l2_s_cache_mode : 0;
3668 #ifdef VERBOSE_INIT_ARM
3669 printf("P");
3670 #endif
3671 pte[(va >> PGSHIFT) & 0x3ff] = L2_S_PROTO | pa |
3672 L2_S_PROT(PTE_KERNEL, prot) | fl;
3673 va += NBPG;
3674 pa += NBPG;
3675 resid -= NBPG;
3676 }
3677 #ifdef VERBOSE_INIT_ARM
3678 printf("\n");
3679 #endif
3680 return (size);
3681 }
3682
3683 /********************** PTE initialization routines **************************/
3684
3685 /*
3686 * These routines are called when the CPU type is identified to set up
3687 * the PTE prototypes, cache modes, etc.
3688 *
3689 * The variables are always here, just in case LKMs need to reference
3690 * them (though, they shouldn't).
3691 */
3692
3693 pt_entry_t pte_l1_s_cache_mode;
3694 pt_entry_t pte_l1_s_cache_mask;
3695
3696 pt_entry_t pte_l2_l_cache_mode;
3697 pt_entry_t pte_l2_l_cache_mask;
3698
3699 pt_entry_t pte_l2_s_cache_mode;
3700 pt_entry_t pte_l2_s_cache_mask;
3701
3702 pt_entry_t pte_l2_s_prot_u;
3703 pt_entry_t pte_l2_s_prot_w;
3704 pt_entry_t pte_l2_s_prot_mask;
3705
3706 pt_entry_t pte_l1_s_proto;
3707 pt_entry_t pte_l1_c_proto;
3708 pt_entry_t pte_l2_s_proto;
3709
3710 void (*pmap_copy_page_func)(paddr_t, paddr_t);
3711 void (*pmap_zero_page_func)(paddr_t);
3712
3713 #if ARM_MMU_GENERIC == 1
3714 void
3715 pmap_pte_init_generic(void)
3716 {
3717
3718 pte_l1_s_cache_mode = L1_S_B|L1_S_C;
3719 pte_l1_s_cache_mask = L1_S_CACHE_MASK_generic;
3720
3721 pte_l2_l_cache_mode = L2_B|L2_C;
3722 pte_l2_l_cache_mask = L2_L_CACHE_MASK_generic;
3723
3724 pte_l2_s_cache_mode = L2_B|L2_C;
3725 pte_l2_s_cache_mask = L2_S_CACHE_MASK_generic;
3726
3727 pte_l2_s_prot_u = L2_S_PROT_U_generic;
3728 pte_l2_s_prot_w = L2_S_PROT_W_generic;
3729 pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
3730
3731 pte_l1_s_proto = L1_S_PROTO_generic;
3732 pte_l1_c_proto = L1_C_PROTO_generic;
3733 pte_l2_s_proto = L2_S_PROTO_generic;
3734
3735 pmap_copy_page_func = pmap_copy_page_generic;
3736 pmap_zero_page_func = pmap_zero_page_generic;
3737 }
3738
3739 #if defined(CPU_ARM9)
3740 void
3741 pmap_pte_init_arm9(void)
3742 {
3743
3744 /*
3745 * ARM9 is compatible with generic, but we want to use
3746 * write-through caching for now.
3747 */
3748 pmap_pte_init_generic();
3749
3750 pte_l1_s_cache_mode = L1_S_C;
3751 pte_l2_l_cache_mode = L2_C;
3752 pte_l2_s_cache_mode = L2_C;
3753 }
3754 #endif /* CPU_ARM9 */
3755 #endif /* ARM_MMU_GENERIC == 1 */
3756
3757 #if ARM_MMU_XSCALE == 1
3758 void
3759 pmap_pte_init_xscale(void)
3760 {
3761
3762 pte_l1_s_cache_mode = L1_S_B|L1_S_C;
3763 pte_l1_s_cache_mask = L1_S_CACHE_MASK_xscale;
3764
3765 pte_l2_l_cache_mode = L2_B|L2_C;
3766 pte_l2_l_cache_mask = L2_L_CACHE_MASK_xscale;
3767
3768 pte_l2_s_cache_mode = L2_B|L2_C;
3769 pte_l2_s_cache_mask = L2_S_CACHE_MASK_xscale;
3770
3771 pte_l2_s_prot_u = L2_S_PROT_U_xscale;
3772 pte_l2_s_prot_w = L2_S_PROT_W_xscale;
3773 pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale;
3774
3775 pte_l1_s_proto = L1_S_PROTO_xscale;
3776 pte_l1_c_proto = L1_C_PROTO_xscale;
3777 pte_l2_s_proto = L2_S_PROTO_xscale;
3778
3779 pmap_copy_page_func = pmap_copy_page_xscale;
3780 pmap_zero_page_func = pmap_zero_page_xscale;
3781 }
3782
3783 #if defined(CPU_XSCALE_80200)
3784 void
3785 pmap_pte_init_i80200(void)
3786 {
3787
3788 /*
3789 * Use write-through caching on the i80200.
3790 */
3791 pmap_pte_init_xscale();
3792 pte_l1_s_cache_mode = L1_S_C;
3793 pte_l2_l_cache_mode = L2_C;
3794 pte_l2_s_cache_mode = L2_C;
3795 }
3796 #endif /* CPU_XSCALE_80200 */
3797
3798 /*
3799 * xscale_setup_minidata:
3800 *
3801 * Set up the mini-data cache clean area. We require the
3802 * caller to allocate the right amount of physically and
3803 * virtually contiguous space.
3804 */
3805 void
3806 xscale_setup_minidata(vaddr_t l1pt, vaddr_t va, paddr_t pa)
3807 {
3808 extern vaddr_t xscale_minidata_clean_addr;
3809 extern vsize_t xscale_minidata_clean_size; /* already initialized */
3810 pd_entry_t *pde = (pd_entry_t *) l1pt;
3811 pt_entry_t *pte;
3812 vsize_t size;
3813
3814 xscale_minidata_clean_addr = va;
3815
3816 /* Round it to page size. */
3817 size = (xscale_minidata_clean_size + L2_S_OFFSET) & L2_S_FRAME;
3818
3819 for (; size != 0;
3820 va += L2_S_SIZE, pa += L2_S_SIZE, size -= L2_S_SIZE) {
3821 pte = (pt_entry_t *)
3822 kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
3823 if (pte == NULL)
3824 panic("xscale_setup_minidata: can't find L2 table for "
3825 "VA 0x%08lx", va);
3826 pte[(va >> PGSHIFT) & 0x3ff] = L2_S_PROTO | pa |
3827 L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
3828 L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);
3829 }
3830 }
3831 #endif /* ARM_MMU_XSCALE == 1 */
3832