pmap.c revision 1.85 1 /* $NetBSD: pmap.c,v 1.85 2002/04/09 21:00:43 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 1.85 2002/04/09 21:00:43 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 * Cache enable bits in PTE to use on pages that are cacheable.
342 * On most machines this is cacheable/bufferable, but on some, eg arm10, we
343 * can chose between write-through and write-back cacheing.
344 */
345 pt_entry_t pte_cache_mode = (L2_C | L2_B);
346
347 /*
348 * real definition of pv_entry.
349 */
350
351 struct pv_entry {
352 struct pv_entry *pv_next; /* next pv_entry */
353 struct pmap *pv_pmap; /* pmap where mapping lies */
354 vaddr_t pv_va; /* virtual address for mapping */
355 int pv_flags; /* flags */
356 struct vm_page *pv_ptp; /* vm_page for the ptp */
357 };
358
359 /*
360 * pv_entrys are dynamically allocated in chunks from a single page.
361 * we keep track of how many pv_entrys are in use for each page and
362 * we can free pv_entry pages if needed. there is one lock for the
363 * entire allocation system.
364 */
365
366 struct pv_page_info {
367 TAILQ_ENTRY(pv_page) pvpi_list;
368 struct pv_entry *pvpi_pvfree;
369 int pvpi_nfree;
370 };
371
372 /*
373 * number of pv_entry's in a pv_page
374 * (note: won't work on systems where NPBG isn't a constant)
375 */
376
377 #define PVE_PER_PVPAGE ((NBPG - sizeof(struct pv_page_info)) / \
378 sizeof(struct pv_entry))
379
380 /*
381 * a pv_page: where pv_entrys are allocated from
382 */
383
384 struct pv_page {
385 struct pv_page_info pvinfo;
386 struct pv_entry pvents[PVE_PER_PVPAGE];
387 };
388
389 #ifdef MYCROFT_HACK
390 int mycroft_hack = 0;
391 #endif
392
393 /* Function to set the debug level of the pmap code */
394
395 #ifdef PMAP_DEBUG
396 void
397 pmap_debug(int level)
398 {
399 pmap_debug_level = level;
400 printf("pmap_debug: level=%d\n", pmap_debug_level);
401 }
402 #endif /* PMAP_DEBUG */
403
404 __inline static boolean_t
405 pmap_is_curpmap(struct pmap *pmap)
406 {
407
408 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap) ||
409 pmap == pmap_kernel())
410 return (TRUE);
411
412 return (FALSE);
413 }
414
415 #include "isadma.h"
416
417 #if NISADMA > 0
418 /*
419 * Used to protect memory for ISA DMA bounce buffers. If, when loading
420 * pages into the system, memory intersects with any of these ranges,
421 * the intersecting memory will be loaded into a lower-priority free list.
422 */
423 bus_dma_segment_t *pmap_isa_dma_ranges;
424 int pmap_isa_dma_nranges;
425
426 /*
427 * Check if a memory range intersects with an ISA DMA range, and
428 * return the page-rounded intersection if it does. The intersection
429 * will be placed on a lower-priority free list.
430 */
431 static boolean_t
432 pmap_isa_dma_range_intersect(paddr_t pa, psize_t size, paddr_t *pap,
433 psize_t *sizep)
434 {
435 bus_dma_segment_t *ds;
436 int i;
437
438 if (pmap_isa_dma_ranges == NULL)
439 return (FALSE);
440
441 for (i = 0, ds = pmap_isa_dma_ranges;
442 i < pmap_isa_dma_nranges; i++, ds++) {
443 if (ds->ds_addr <= pa && pa < (ds->ds_addr + ds->ds_len)) {
444 /*
445 * Beginning of region intersects with this range.
446 */
447 *pap = trunc_page(pa);
448 *sizep = round_page(min(pa + size,
449 ds->ds_addr + ds->ds_len) - pa);
450 return (TRUE);
451 }
452 if (pa < ds->ds_addr && ds->ds_addr < (pa + size)) {
453 /*
454 * End of region intersects with this range.
455 */
456 *pap = trunc_page(ds->ds_addr);
457 *sizep = round_page(min((pa + size) - ds->ds_addr,
458 ds->ds_len));
459 return (TRUE);
460 }
461 }
462
463 /*
464 * No intersection found.
465 */
466 return (FALSE);
467 }
468 #endif /* NISADMA > 0 */
469
470 /*
471 * p v _ e n t r y f u n c t i o n s
472 */
473
474 /*
475 * pv_entry allocation functions:
476 * the main pv_entry allocation functions are:
477 * pmap_alloc_pv: allocate a pv_entry structure
478 * pmap_free_pv: free one pv_entry
479 * pmap_free_pvs: free a list of pv_entrys
480 *
481 * the rest are helper functions
482 */
483
484 /*
485 * pmap_alloc_pv: inline function to allocate a pv_entry structure
486 * => we lock pvalloc_lock
487 * => if we fail, we call out to pmap_alloc_pvpage
488 * => 3 modes:
489 * ALLOCPV_NEED = we really need a pv_entry, even if we have to steal it
490 * ALLOCPV_TRY = we want a pv_entry, but not enough to steal
491 * ALLOCPV_NONEED = we are trying to grow our free list, don't really need
492 * one now
493 *
494 * "try" is for optional functions like pmap_copy().
495 */
496
497 __inline static struct pv_entry *
498 pmap_alloc_pv(struct pmap *pmap, int mode)
499 {
500 struct pv_page *pvpage;
501 struct pv_entry *pv;
502
503 simple_lock(&pvalloc_lock);
504
505 pvpage = TAILQ_FIRST(&pv_freepages);
506
507 if (pvpage != NULL) {
508 pvpage->pvinfo.pvpi_nfree--;
509 if (pvpage->pvinfo.pvpi_nfree == 0) {
510 /* nothing left in this one? */
511 TAILQ_REMOVE(&pv_freepages, pvpage, pvinfo.pvpi_list);
512 }
513 pv = pvpage->pvinfo.pvpi_pvfree;
514 KASSERT(pv);
515 pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
516 pv_nfpvents--; /* took one from pool */
517 } else {
518 pv = NULL; /* need more of them */
519 }
520
521 /*
522 * if below low water mark or we didn't get a pv_entry we try and
523 * create more pv_entrys ...
524 */
525
526 if (pv_nfpvents < PVE_LOWAT || pv == NULL) {
527 if (pv == NULL)
528 pv = pmap_alloc_pvpage(pmap, (mode == ALLOCPV_TRY) ?
529 mode : ALLOCPV_NEED);
530 else
531 (void) pmap_alloc_pvpage(pmap, ALLOCPV_NONEED);
532 }
533
534 simple_unlock(&pvalloc_lock);
535 return(pv);
536 }
537
538 /*
539 * pmap_alloc_pvpage: maybe allocate a new pvpage
540 *
541 * if need_entry is false: try and allocate a new pv_page
542 * if need_entry is true: try and allocate a new pv_page and return a
543 * new pv_entry from it. if we are unable to allocate a pv_page
544 * we make a last ditch effort to steal a pv_page from some other
545 * mapping. if that fails, we panic...
546 *
547 * => we assume that the caller holds pvalloc_lock
548 */
549
550 static struct pv_entry *
551 pmap_alloc_pvpage(struct pmap *pmap, int mode)
552 {
553 struct vm_page *pg;
554 struct pv_page *pvpage;
555 struct pv_entry *pv;
556 int s;
557
558 /*
559 * if we need_entry and we've got unused pv_pages, allocate from there
560 */
561
562 pvpage = TAILQ_FIRST(&pv_unusedpgs);
563 if (mode != ALLOCPV_NONEED && pvpage != NULL) {
564
565 /* move it to pv_freepages list */
566 TAILQ_REMOVE(&pv_unusedpgs, pvpage, pvinfo.pvpi_list);
567 TAILQ_INSERT_HEAD(&pv_freepages, pvpage, pvinfo.pvpi_list);
568
569 /* allocate a pv_entry */
570 pvpage->pvinfo.pvpi_nfree--; /* can't go to zero */
571 pv = pvpage->pvinfo.pvpi_pvfree;
572 KASSERT(pv);
573 pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
574
575 pv_nfpvents--; /* took one from pool */
576 return(pv);
577 }
578
579 /*
580 * see if we've got a cached unmapped VA that we can map a page in.
581 * if not, try to allocate one.
582 */
583
584
585 if (pv_cachedva == 0) {
586 s = splvm();
587 pv_cachedva = uvm_km_kmemalloc(kmem_map, NULL,
588 PAGE_SIZE, UVM_KMF_TRYLOCK|UVM_KMF_VALLOC);
589 splx(s);
590 if (pv_cachedva == 0) {
591 return (NULL);
592 }
593 }
594
595 pg = uvm_pagealloc(NULL, pv_cachedva - vm_map_min(kernel_map), NULL,
596 UVM_PGA_USERESERVE);
597
598 if (pg == NULL)
599 return (NULL);
600 pg->flags &= ~PG_BUSY; /* never busy */
601
602 /*
603 * add a mapping for our new pv_page and free its entrys (save one!)
604 *
605 * NOTE: If we are allocating a PV page for the kernel pmap, the
606 * pmap is already locked! (...but entering the mapping is safe...)
607 */
608
609 pmap_kenter_pa(pv_cachedva, VM_PAGE_TO_PHYS(pg),
610 VM_PROT_READ|VM_PROT_WRITE);
611 pmap_update(pmap_kernel());
612 pvpage = (struct pv_page *) pv_cachedva;
613 pv_cachedva = 0;
614 return (pmap_add_pvpage(pvpage, mode != ALLOCPV_NONEED));
615 }
616
617 /*
618 * pmap_add_pvpage: add a pv_page's pv_entrys to the free list
619 *
620 * => caller must hold pvalloc_lock
621 * => if need_entry is true, we allocate and return one pv_entry
622 */
623
624 static struct pv_entry *
625 pmap_add_pvpage(struct pv_page *pvp, boolean_t need_entry)
626 {
627 int tofree, lcv;
628
629 /* do we need to return one? */
630 tofree = (need_entry) ? PVE_PER_PVPAGE - 1 : PVE_PER_PVPAGE;
631
632 pvp->pvinfo.pvpi_pvfree = NULL;
633 pvp->pvinfo.pvpi_nfree = tofree;
634 for (lcv = 0 ; lcv < tofree ; lcv++) {
635 pvp->pvents[lcv].pv_next = pvp->pvinfo.pvpi_pvfree;
636 pvp->pvinfo.pvpi_pvfree = &pvp->pvents[lcv];
637 }
638 if (need_entry)
639 TAILQ_INSERT_TAIL(&pv_freepages, pvp, pvinfo.pvpi_list);
640 else
641 TAILQ_INSERT_TAIL(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
642 pv_nfpvents += tofree;
643 return((need_entry) ? &pvp->pvents[lcv] : NULL);
644 }
645
646 /*
647 * pmap_free_pv_doit: actually free a pv_entry
648 *
649 * => do not call this directly! instead use either
650 * 1. pmap_free_pv ==> free a single pv_entry
651 * 2. pmap_free_pvs => free a list of pv_entrys
652 * => we must be holding pvalloc_lock
653 */
654
655 __inline static void
656 pmap_free_pv_doit(struct pv_entry *pv)
657 {
658 struct pv_page *pvp;
659
660 pvp = (struct pv_page *) arm_trunc_page((vaddr_t)pv);
661 pv_nfpvents++;
662 pvp->pvinfo.pvpi_nfree++;
663
664 /* nfree == 1 => fully allocated page just became partly allocated */
665 if (pvp->pvinfo.pvpi_nfree == 1) {
666 TAILQ_INSERT_HEAD(&pv_freepages, pvp, pvinfo.pvpi_list);
667 }
668
669 /* free it */
670 pv->pv_next = pvp->pvinfo.pvpi_pvfree;
671 pvp->pvinfo.pvpi_pvfree = pv;
672
673 /*
674 * are all pv_page's pv_entry's free? move it to unused queue.
675 */
676
677 if (pvp->pvinfo.pvpi_nfree == PVE_PER_PVPAGE) {
678 TAILQ_REMOVE(&pv_freepages, pvp, pvinfo.pvpi_list);
679 TAILQ_INSERT_HEAD(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
680 }
681 }
682
683 /*
684 * pmap_free_pv: free a single pv_entry
685 *
686 * => we gain the pvalloc_lock
687 */
688
689 __inline static void
690 pmap_free_pv(struct pmap *pmap, struct pv_entry *pv)
691 {
692 simple_lock(&pvalloc_lock);
693 pmap_free_pv_doit(pv);
694
695 /*
696 * Can't free the PV page if the PV entries were associated with
697 * the kernel pmap; the pmap is already locked.
698 */
699 if (pv_nfpvents > PVE_HIWAT && TAILQ_FIRST(&pv_unusedpgs) != NULL &&
700 pmap != pmap_kernel())
701 pmap_free_pvpage();
702
703 simple_unlock(&pvalloc_lock);
704 }
705
706 /*
707 * pmap_free_pvs: free a list of pv_entrys
708 *
709 * => we gain the pvalloc_lock
710 */
711
712 __inline static void
713 pmap_free_pvs(struct pmap *pmap, struct pv_entry *pvs)
714 {
715 struct pv_entry *nextpv;
716
717 simple_lock(&pvalloc_lock);
718
719 for ( /* null */ ; pvs != NULL ; pvs = nextpv) {
720 nextpv = pvs->pv_next;
721 pmap_free_pv_doit(pvs);
722 }
723
724 /*
725 * Can't free the PV page if the PV entries were associated with
726 * the kernel pmap; the pmap is already locked.
727 */
728 if (pv_nfpvents > PVE_HIWAT && TAILQ_FIRST(&pv_unusedpgs) != NULL &&
729 pmap != pmap_kernel())
730 pmap_free_pvpage();
731
732 simple_unlock(&pvalloc_lock);
733 }
734
735
736 /*
737 * pmap_free_pvpage: try and free an unused pv_page structure
738 *
739 * => assume caller is holding the pvalloc_lock and that
740 * there is a page on the pv_unusedpgs list
741 * => if we can't get a lock on the kmem_map we try again later
742 */
743
744 static void
745 pmap_free_pvpage(void)
746 {
747 int s;
748 struct vm_map *map;
749 struct vm_map_entry *dead_entries;
750 struct pv_page *pvp;
751
752 s = splvm(); /* protect kmem_map */
753
754 pvp = TAILQ_FIRST(&pv_unusedpgs);
755
756 /*
757 * note: watch out for pv_initpage which is allocated out of
758 * kernel_map rather than kmem_map.
759 */
760 if (pvp == pv_initpage)
761 map = kernel_map;
762 else
763 map = kmem_map;
764 if (vm_map_lock_try(map)) {
765
766 /* remove pvp from pv_unusedpgs */
767 TAILQ_REMOVE(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
768
769 /* unmap the page */
770 dead_entries = NULL;
771 uvm_unmap_remove(map, (vaddr_t)pvp, ((vaddr_t)pvp) + PAGE_SIZE,
772 &dead_entries);
773 vm_map_unlock(map);
774
775 if (dead_entries != NULL)
776 uvm_unmap_detach(dead_entries, 0);
777
778 pv_nfpvents -= PVE_PER_PVPAGE; /* update free count */
779 }
780 if (pvp == pv_initpage)
781 /* no more initpage, we've freed it */
782 pv_initpage = NULL;
783
784 splx(s);
785 }
786
787 /*
788 * main pv_entry manipulation functions:
789 * pmap_enter_pv: enter a mapping onto a vm_page list
790 * pmap_remove_pv: remove a mappiing from a vm_page list
791 *
792 * NOTE: pmap_enter_pv expects to lock the pvh itself
793 * pmap_remove_pv expects te caller to lock the pvh before calling
794 */
795
796 /*
797 * pmap_enter_pv: enter a mapping onto a vm_page lst
798 *
799 * => caller should hold the proper lock on pmap_main_lock
800 * => caller should have pmap locked
801 * => we will gain the lock on the vm_page and allocate the new pv_entry
802 * => caller should adjust ptp's wire_count before calling
803 * => caller should not adjust pmap's wire_count
804 */
805
806 __inline static void
807 pmap_enter_pv(struct vm_page *pg, struct pv_entry *pve, struct pmap *pmap,
808 vaddr_t va, struct vm_page *ptp, int flags)
809 {
810 pve->pv_pmap = pmap;
811 pve->pv_va = va;
812 pve->pv_ptp = ptp; /* NULL for kernel pmap */
813 pve->pv_flags = flags;
814 simple_lock(&pg->mdpage.pvh_slock); /* lock vm_page */
815 pve->pv_next = pg->mdpage.pvh_list; /* add to ... */
816 pg->mdpage.pvh_list = pve; /* ... locked list */
817 simple_unlock(&pg->mdpage.pvh_slock); /* unlock, done! */
818 if (pve->pv_flags & PVF_WIRED)
819 ++pmap->pm_stats.wired_count;
820 }
821
822 /*
823 * pmap_remove_pv: try to remove a mapping from a pv_list
824 *
825 * => caller should hold proper lock on pmap_main_lock
826 * => pmap should be locked
827 * => caller should hold lock on vm_page [so that attrs can be adjusted]
828 * => caller should adjust ptp's wire_count and free PTP if needed
829 * => caller should NOT adjust pmap's wire_count
830 * => we return the removed pve
831 */
832
833 __inline static struct pv_entry *
834 pmap_remove_pv(struct vm_page *pg, struct pmap *pmap, vaddr_t va)
835 {
836 struct pv_entry *pve, **prevptr;
837
838 prevptr = &pg->mdpage.pvh_list; /* previous pv_entry pointer */
839 pve = *prevptr;
840 while (pve) {
841 if (pve->pv_pmap == pmap && pve->pv_va == va) { /* match? */
842 *prevptr = pve->pv_next; /* remove it! */
843 if (pve->pv_flags & PVF_WIRED)
844 --pmap->pm_stats.wired_count;
845 break;
846 }
847 prevptr = &pve->pv_next; /* previous pointer */
848 pve = pve->pv_next; /* advance */
849 }
850 return(pve); /* return removed pve */
851 }
852
853 /*
854 *
855 * pmap_modify_pv: Update pv flags
856 *
857 * => caller should hold lock on vm_page [so that attrs can be adjusted]
858 * => caller should NOT adjust pmap's wire_count
859 * => caller must call pmap_vac_me_harder() if writable status of a page
860 * may have changed.
861 * => we return the old flags
862 *
863 * Modify a physical-virtual mapping in the pv table
864 */
865
866 static /* __inline */ u_int
867 pmap_modify_pv(struct pmap *pmap, vaddr_t va, struct vm_page *pg,
868 u_int bic_mask, u_int eor_mask)
869 {
870 struct pv_entry *npv;
871 u_int flags, oflags;
872
873 /*
874 * There is at least one VA mapping this page.
875 */
876
877 for (npv = pg->mdpage.pvh_list; npv; npv = npv->pv_next) {
878 if (pmap == npv->pv_pmap && va == npv->pv_va) {
879 oflags = npv->pv_flags;
880 npv->pv_flags = flags =
881 ((oflags & ~bic_mask) ^ eor_mask);
882 if ((flags ^ oflags) & PVF_WIRED) {
883 if (flags & PVF_WIRED)
884 ++pmap->pm_stats.wired_count;
885 else
886 --pmap->pm_stats.wired_count;
887 }
888 return (oflags);
889 }
890 }
891 return (0);
892 }
893
894 /*
895 * Map the specified level 2 pagetable into the level 1 page table for
896 * the given pmap to cover a chunk of virtual address space starting from the
897 * address specified.
898 */
899 static __inline void
900 pmap_map_in_l1(struct pmap *pmap, vaddr_t va, paddr_t l2pa, boolean_t selfref)
901 {
902 vaddr_t ptva;
903
904 /* Calculate the index into the L1 page table. */
905 ptva = (va >> L1_S_SHIFT) & ~3;
906
907 /* Map page table into the L1. */
908 pmap->pm_pdir[ptva + 0] = L1_C_PROTO | (l2pa + 0x000);
909 pmap->pm_pdir[ptva + 1] = L1_C_PROTO | (l2pa + 0x400);
910 pmap->pm_pdir[ptva + 2] = L1_C_PROTO | (l2pa + 0x800);
911 pmap->pm_pdir[ptva + 3] = L1_C_PROTO | (l2pa + 0xc00);
912
913 /* Map the page table into the page table area. */
914 if (selfref)
915 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) = L2_S_PROTO | l2pa |
916 L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE);
917 }
918
919 #if 0
920 static __inline void
921 pmap_unmap_in_l1(struct pmap *pmap, vaddr_t va)
922 {
923 vaddr_t ptva;
924
925 /* Calculate the index into the L1 page table. */
926 ptva = (va >> L1_S_SHIFT) & ~3;
927
928 /* Unmap page table from the L1. */
929 pmap->pm_pdir[ptva + 0] = 0;
930 pmap->pm_pdir[ptva + 1] = 0;
931 pmap->pm_pdir[ptva + 2] = 0;
932 pmap->pm_pdir[ptva + 3] = 0;
933
934 /* Unmap the page table from the page table area. */
935 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) = 0;
936 }
937 #endif
938
939 /*
940 * Used to map a range of physical addresses into kernel
941 * virtual address space.
942 *
943 * For now, VM is already on, we only need to map the
944 * specified memory.
945 */
946 vaddr_t
947 pmap_map(vaddr_t va, paddr_t spa, paddr_t epa, vm_prot_t prot)
948 {
949 while (spa < epa) {
950 pmap_kenter_pa(va, spa, prot);
951 va += NBPG;
952 spa += NBPG;
953 }
954 pmap_update(pmap_kernel());
955 return(va);
956 }
957
958
959 /*
960 * void pmap_bootstrap(pd_entry_t *kernel_l1pt, pv_addr_t kernel_ptpt)
961 *
962 * bootstrap the pmap system. This is called from initarm and allows
963 * the pmap system to initailise any structures it requires.
964 *
965 * Currently this sets up the kernel_pmap that is statically allocated
966 * and also allocated virtual addresses for certain page hooks.
967 * Currently the only one page hook is allocated that is used
968 * to zero physical pages of memory.
969 * It also initialises the start and end address of the kernel data space.
970 */
971 extern paddr_t physical_freestart;
972 extern paddr_t physical_freeend;
973
974 char *boot_head;
975
976 void
977 pmap_bootstrap(pd_entry_t *kernel_l1pt, pv_addr_t kernel_ptpt)
978 {
979 pt_entry_t *pte;
980 int loop;
981 paddr_t start, end;
982 #if NISADMA > 0
983 paddr_t istart;
984 psize_t isize;
985 #endif
986
987 pmap_kernel()->pm_pdir = kernel_l1pt;
988 pmap_kernel()->pm_pptpt = kernel_ptpt.pv_pa;
989 pmap_kernel()->pm_vptpt = kernel_ptpt.pv_va;
990 simple_lock_init(&pmap_kernel()->pm_lock);
991 pmap_kernel()->pm_obj.pgops = NULL;
992 TAILQ_INIT(&(pmap_kernel()->pm_obj.memq));
993 pmap_kernel()->pm_obj.uo_npages = 0;
994 pmap_kernel()->pm_obj.uo_refs = 1;
995
996 /*
997 * Initialize PAGE_SIZE-dependent variables.
998 */
999 uvm_setpagesize();
1000
1001 loop = 0;
1002 while (loop < bootconfig.dramblocks) {
1003 start = (paddr_t)bootconfig.dram[loop].address;
1004 end = start + (bootconfig.dram[loop].pages * NBPG);
1005 if (start < physical_freestart)
1006 start = physical_freestart;
1007 if (end > physical_freeend)
1008 end = physical_freeend;
1009 #if 0
1010 printf("%d: %lx -> %lx\n", loop, start, end - 1);
1011 #endif
1012 #if NISADMA > 0
1013 if (pmap_isa_dma_range_intersect(start, end - start,
1014 &istart, &isize)) {
1015 /*
1016 * Place the pages that intersect with the
1017 * ISA DMA range onto the ISA DMA free list.
1018 */
1019 #if 0
1020 printf(" ISADMA 0x%lx -> 0x%lx\n", istart,
1021 istart + isize - 1);
1022 #endif
1023 uvm_page_physload(atop(istart),
1024 atop(istart + isize), atop(istart),
1025 atop(istart + isize), VM_FREELIST_ISADMA);
1026
1027 /*
1028 * Load the pieces that come before
1029 * the intersection into the default
1030 * free list.
1031 */
1032 if (start < istart) {
1033 #if 0
1034 printf(" BEFORE 0x%lx -> 0x%lx\n",
1035 start, istart - 1);
1036 #endif
1037 uvm_page_physload(atop(start),
1038 atop(istart), atop(start),
1039 atop(istart), VM_FREELIST_DEFAULT);
1040 }
1041
1042 /*
1043 * Load the pieces that come after
1044 * the intersection into the default
1045 * free list.
1046 */
1047 if ((istart + isize) < end) {
1048 #if 0
1049 printf(" AFTER 0x%lx -> 0x%lx\n",
1050 (istart + isize), end - 1);
1051 #endif
1052 uvm_page_physload(atop(istart + isize),
1053 atop(end), atop(istart + isize),
1054 atop(end), VM_FREELIST_DEFAULT);
1055 }
1056 } else {
1057 uvm_page_physload(atop(start), atop(end),
1058 atop(start), atop(end), VM_FREELIST_DEFAULT);
1059 }
1060 #else /* NISADMA > 0 */
1061 uvm_page_physload(atop(start), atop(end),
1062 atop(start), atop(end), VM_FREELIST_DEFAULT);
1063 #endif /* NISADMA > 0 */
1064 ++loop;
1065 }
1066
1067 virtual_avail = KERNEL_VM_BASE;
1068 virtual_end = KERNEL_VM_BASE + KERNEL_VM_SIZE;
1069
1070 /*
1071 * now we allocate the "special" VAs which are used for tmp mappings
1072 * by the pmap (and other modules). we allocate the VAs by advancing
1073 * virtual_avail (note that there are no pages mapped at these VAs).
1074 * we find the PTE that maps the allocated VA via the linear PTE
1075 * mapping.
1076 */
1077
1078 pte = ((pt_entry_t *) PTE_BASE) + atop(virtual_avail);
1079
1080 csrcp = virtual_avail; csrc_pte = pte;
1081 virtual_avail += PAGE_SIZE; pte++;
1082
1083 cdstp = virtual_avail; cdst_pte = pte;
1084 virtual_avail += PAGE_SIZE; pte++;
1085
1086 memhook = (char *) virtual_avail; /* don't need pte */
1087 virtual_avail += PAGE_SIZE; pte++;
1088
1089 msgbufaddr = (caddr_t) virtual_avail; /* don't need pte */
1090 virtual_avail += round_page(MSGBUFSIZE);
1091 pte += atop(round_page(MSGBUFSIZE));
1092
1093 /*
1094 * init the static-global locks and global lists.
1095 */
1096 spinlockinit(&pmap_main_lock, "pmaplk", 0);
1097 simple_lock_init(&pvalloc_lock);
1098 simple_lock_init(&pmaps_lock);
1099 LIST_INIT(&pmaps);
1100 TAILQ_INIT(&pv_freepages);
1101 TAILQ_INIT(&pv_unusedpgs);
1102
1103 /*
1104 * initialize the pmap pool.
1105 */
1106
1107 pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl",
1108 &pool_allocator_nointr);
1109
1110 cpu_dcache_wbinv_all();
1111 }
1112
1113 /*
1114 * void pmap_init(void)
1115 *
1116 * Initialize the pmap module.
1117 * Called by vm_init() in vm/vm_init.c in order to initialise
1118 * any structures that the pmap system needs to map virtual memory.
1119 */
1120
1121 extern int physmem;
1122
1123 void
1124 pmap_init(void)
1125 {
1126
1127 /*
1128 * Set the available memory vars - These do not map to real memory
1129 * addresses and cannot as the physical memory is fragmented.
1130 * They are used by ps for %mem calculations.
1131 * One could argue whether this should be the entire memory or just
1132 * the memory that is useable in a user process.
1133 */
1134 avail_start = 0;
1135 avail_end = physmem * NBPG;
1136
1137 /*
1138 * now we need to free enough pv_entry structures to allow us to get
1139 * the kmem_map/kmem_object allocated and inited (done after this
1140 * function is finished). to do this we allocate one bootstrap page out
1141 * of kernel_map and use it to provide an initial pool of pv_entry
1142 * structures. we never free this page.
1143 */
1144
1145 pv_initpage = (struct pv_page *) uvm_km_alloc(kernel_map, PAGE_SIZE);
1146 if (pv_initpage == NULL)
1147 panic("pmap_init: pv_initpage");
1148 pv_cachedva = 0; /* a VA we have allocated but not used yet */
1149 pv_nfpvents = 0;
1150 (void) pmap_add_pvpage(pv_initpage, FALSE);
1151
1152 pmap_initialized = TRUE;
1153
1154 /* Initialise our L1 page table queues and counters */
1155 SIMPLEQ_INIT(&l1pt_static_queue);
1156 l1pt_static_queue_count = 0;
1157 l1pt_static_create_count = 0;
1158 SIMPLEQ_INIT(&l1pt_queue);
1159 l1pt_queue_count = 0;
1160 l1pt_create_count = 0;
1161 l1pt_reuse_count = 0;
1162 }
1163
1164 /*
1165 * pmap_postinit()
1166 *
1167 * This routine is called after the vm and kmem subsystems have been
1168 * initialised. This allows the pmap code to perform any initialisation
1169 * that can only be done one the memory allocation is in place.
1170 */
1171
1172 void
1173 pmap_postinit(void)
1174 {
1175 int loop;
1176 struct l1pt *pt;
1177
1178 #ifdef PMAP_STATIC_L1S
1179 for (loop = 0; loop < PMAP_STATIC_L1S; ++loop) {
1180 #else /* PMAP_STATIC_L1S */
1181 for (loop = 0; loop < max_processes; ++loop) {
1182 #endif /* PMAP_STATIC_L1S */
1183 /* Allocate a L1 page table */
1184 pt = pmap_alloc_l1pt();
1185 if (!pt)
1186 panic("Cannot allocate static L1 page tables\n");
1187
1188 /* Clean it */
1189 bzero((void *)pt->pt_va, L1_TABLE_SIZE);
1190 pt->pt_flags |= (PTFLAG_STATIC | PTFLAG_CLEAN);
1191 /* Add the page table to the queue */
1192 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pt, pt_queue);
1193 ++l1pt_static_queue_count;
1194 ++l1pt_static_create_count;
1195 }
1196 }
1197
1198
1199 /*
1200 * Create and return a physical map.
1201 *
1202 * If the size specified for the map is zero, the map is an actual physical
1203 * map, and may be referenced by the hardware.
1204 *
1205 * If the size specified is non-zero, the map will be used in software only,
1206 * and is bounded by that size.
1207 */
1208
1209 pmap_t
1210 pmap_create(void)
1211 {
1212 struct pmap *pmap;
1213
1214 /*
1215 * Fetch pmap entry from the pool
1216 */
1217
1218 pmap = pool_get(&pmap_pmap_pool, PR_WAITOK);
1219 /* XXX is this really needed! */
1220 memset(pmap, 0, sizeof(*pmap));
1221
1222 simple_lock_init(&pmap->pm_obj.vmobjlock);
1223 pmap->pm_obj.pgops = NULL; /* currently not a mappable object */
1224 TAILQ_INIT(&pmap->pm_obj.memq);
1225 pmap->pm_obj.uo_npages = 0;
1226 pmap->pm_obj.uo_refs = 1;
1227 pmap->pm_stats.wired_count = 0;
1228 pmap->pm_stats.resident_count = 1;
1229 pmap->pm_ptphint = NULL;
1230
1231 /* Now init the machine part of the pmap */
1232 pmap_pinit(pmap);
1233 return(pmap);
1234 }
1235
1236 /*
1237 * pmap_alloc_l1pt()
1238 *
1239 * This routine allocates physical and virtual memory for a L1 page table
1240 * and wires it.
1241 * A l1pt structure is returned to describe the allocated page table.
1242 *
1243 * This routine is allowed to fail if the required memory cannot be allocated.
1244 * In this case NULL is returned.
1245 */
1246
1247 struct l1pt *
1248 pmap_alloc_l1pt(void)
1249 {
1250 paddr_t pa;
1251 vaddr_t va;
1252 struct l1pt *pt;
1253 int error;
1254 struct vm_page *m;
1255 pt_entry_t *ptes;
1256
1257 /* Allocate virtual address space for the L1 page table */
1258 va = uvm_km_valloc(kernel_map, L1_TABLE_SIZE);
1259 if (va == 0) {
1260 #ifdef DIAGNOSTIC
1261 PDEBUG(0,
1262 printf("pmap: Cannot allocate pageable memory for L1\n"));
1263 #endif /* DIAGNOSTIC */
1264 return(NULL);
1265 }
1266
1267 /* Allocate memory for the l1pt structure */
1268 pt = (struct l1pt *)malloc(sizeof(struct l1pt), M_VMPMAP, M_WAITOK);
1269
1270 /*
1271 * Allocate pages from the VM system.
1272 */
1273 TAILQ_INIT(&pt->pt_plist);
1274 error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start, physical_end,
1275 L1_TABLE_SIZE, 0, &pt->pt_plist, 1, M_WAITOK);
1276 if (error) {
1277 #ifdef DIAGNOSTIC
1278 PDEBUG(0,
1279 printf("pmap: Cannot allocate physical mem for L1 (%d)\n",
1280 error));
1281 #endif /* DIAGNOSTIC */
1282 /* Release the resources we already have claimed */
1283 free(pt, M_VMPMAP);
1284 uvm_km_free(kernel_map, va, L1_TABLE_SIZE);
1285 return(NULL);
1286 }
1287
1288 /* Map our physical pages into our virtual space */
1289 pt->pt_va = va;
1290 m = TAILQ_FIRST(&pt->pt_plist);
1291 ptes = pmap_map_ptes(pmap_kernel());
1292 while (m && va < (pt->pt_va + L1_TABLE_SIZE)) {
1293 pa = VM_PAGE_TO_PHYS(m);
1294
1295 pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE);
1296
1297 /* Revoke cacheability and bufferability */
1298 /* XXX should be done better than this */
1299 ptes[arm_btop(va)] &= ~(L2_C | L2_B);
1300
1301 va += NBPG;
1302 m = m->pageq.tqe_next;
1303 }
1304 pmap_unmap_ptes(pmap_kernel());
1305 pmap_update(pmap_kernel());
1306
1307 #ifdef DIAGNOSTIC
1308 if (m)
1309 panic("pmap_alloc_l1pt: pglist not empty\n");
1310 #endif /* DIAGNOSTIC */
1311
1312 pt->pt_flags = 0;
1313 return(pt);
1314 }
1315
1316 /*
1317 * Free a L1 page table previously allocated with pmap_alloc_l1pt().
1318 */
1319 static void
1320 pmap_free_l1pt(struct l1pt *pt)
1321 {
1322 /* Separate the physical memory for the virtual space */
1323 pmap_kremove(pt->pt_va, L1_TABLE_SIZE);
1324 pmap_update(pmap_kernel());
1325
1326 /* Return the physical memory */
1327 uvm_pglistfree(&pt->pt_plist);
1328
1329 /* Free the virtual space */
1330 uvm_km_free(kernel_map, pt->pt_va, L1_TABLE_SIZE);
1331
1332 /* Free the l1pt structure */
1333 free(pt, M_VMPMAP);
1334 }
1335
1336 /*
1337 * Allocate a page directory.
1338 * This routine will either allocate a new page directory from the pool
1339 * of L1 page tables currently held by the kernel or it will allocate
1340 * a new one via pmap_alloc_l1pt().
1341 * It will then initialise the l1 page table for use.
1342 *
1343 * XXX must tidy up and fix this code, not happy about how it does the pmaps_locking
1344 */
1345 static int
1346 pmap_allocpagedir(struct pmap *pmap)
1347 {
1348 paddr_t pa;
1349 struct l1pt *pt;
1350 pt_entry_t *pte;
1351
1352 PDEBUG(0, printf("pmap_allocpagedir(%p)\n", pmap));
1353
1354 /* Do we have any spare L1's lying around ? */
1355 if (l1pt_static_queue_count) {
1356 --l1pt_static_queue_count;
1357 pt = l1pt_static_queue.sqh_first;
1358 SIMPLEQ_REMOVE_HEAD(&l1pt_static_queue, pt, pt_queue);
1359 } else if (l1pt_queue_count) {
1360 --l1pt_queue_count;
1361 pt = l1pt_queue.sqh_first;
1362 SIMPLEQ_REMOVE_HEAD(&l1pt_queue, pt, pt_queue);
1363 ++l1pt_reuse_count;
1364 } else {
1365 pt = pmap_alloc_l1pt();
1366 if (!pt)
1367 return(ENOMEM);
1368 ++l1pt_create_count;
1369 }
1370
1371 /* Store the pointer to the l1 descriptor in the pmap. */
1372 pmap->pm_l1pt = pt;
1373
1374 /* Get the physical address of the start of the l1 */
1375 pa = VM_PAGE_TO_PHYS(TAILQ_FIRST(&pt->pt_plist));
1376
1377 /* Store the virtual address of the l1 in the pmap. */
1378 pmap->pm_pdir = (pd_entry_t *)pt->pt_va;
1379
1380 /* Clean the L1 if it is dirty */
1381 if (!(pt->pt_flags & PTFLAG_CLEAN))
1382 bzero((void *)pmap->pm_pdir, (L1_TABLE_SIZE - KERNEL_PD_SIZE));
1383
1384 /* Allocate a page table to map all the page tables for this pmap */
1385
1386 #ifdef DIAGNOSTIC
1387 if (pmap->pm_vptpt) {
1388 /* XXX What if we have one already ? */
1389 panic("pmap_allocpagedir: have pt already\n");
1390 }
1391 #endif /* DIAGNOSTIC */
1392 pmap->pm_vptpt = uvm_km_zalloc(kernel_map, NBPG);
1393 if (pmap->pm_vptpt == 0) {
1394 pmap_freepagedir(pmap);
1395 return(ENOMEM);
1396 }
1397
1398 /* need to lock this all up for growkernel */
1399 simple_lock(&pmaps_lock);
1400 /* wish we didn't have to keep this locked... */
1401
1402 /* Duplicate the kernel mappings. */
1403 bcopy((char *)pmap_kernel()->pm_pdir + (L1_TABLE_SIZE - KERNEL_PD_SIZE),
1404 (char *)pmap->pm_pdir + (L1_TABLE_SIZE - KERNEL_PD_SIZE),
1405 KERNEL_PD_SIZE);
1406
1407 pte = vtopte(pmap->pm_vptpt);
1408 pmap->pm_pptpt = l2pte_pa(*pte);
1409
1410 /* Revoke cacheability and bufferability */
1411 /* XXX should be done better than this */
1412 *pte &= ~(L2_C | L2_B);
1413
1414 /* Wire in this page table */
1415 pmap_map_in_l1(pmap, PTE_BASE, pmap->pm_pptpt, TRUE);
1416
1417 pt->pt_flags &= ~PTFLAG_CLEAN; /* L1 is dirty now */
1418
1419 /*
1420 * Map the kernel page tables into the new PT map.
1421 */
1422 bcopy((char *)(PTE_BASE
1423 + (PTE_BASE >> (PGSHIFT - 2))
1424 + ((L1_TABLE_SIZE - KERNEL_PD_SIZE) >> 2)),
1425 (char *)pmap->pm_vptpt + ((L1_TABLE_SIZE - KERNEL_PD_SIZE) >> 2),
1426 (KERNEL_PD_SIZE >> 2));
1427
1428 LIST_INSERT_HEAD(&pmaps, pmap, pm_list);
1429 simple_unlock(&pmaps_lock);
1430
1431 return(0);
1432 }
1433
1434
1435 /*
1436 * Initialize a preallocated and zeroed pmap structure,
1437 * such as one in a vmspace structure.
1438 */
1439
1440 void
1441 pmap_pinit(struct pmap *pmap)
1442 {
1443 int backoff = 6;
1444 int retry = 10;
1445
1446 PDEBUG(0, printf("pmap_pinit(%p)\n", pmap));
1447
1448 /* Keep looping until we succeed in allocating a page directory */
1449 while (pmap_allocpagedir(pmap) != 0) {
1450 /*
1451 * Ok we failed to allocate a suitable block of memory for an
1452 * L1 page table. This means that either:
1453 * 1. 16KB of virtual address space could not be allocated
1454 * 2. 16KB of physically contiguous memory on a 16KB boundary
1455 * could not be allocated.
1456 *
1457 * Since we cannot fail we will sleep for a while and try
1458 * again.
1459 *
1460 * Searching for a suitable L1 PT is expensive:
1461 * to avoid hogging the system when memory is really
1462 * scarce, use an exponential back-off so that
1463 * eventually we won't retry more than once every 8
1464 * seconds. This should allow other processes to run
1465 * to completion and free up resources.
1466 */
1467 (void) ltsleep(&lbolt, PVM, "l1ptwait", (hz << 3) >> backoff,
1468 NULL);
1469 if (--retry == 0) {
1470 retry = 10;
1471 if (backoff)
1472 --backoff;
1473 }
1474 }
1475
1476 if (vector_page < KERNEL_BASE) {
1477 /*
1478 * Map the vector page. This will also allocate and map
1479 * an L2 table for it.
1480 */
1481 pmap_enter(pmap, vector_page, systempage.pv_pa,
1482 VM_PROT_READ, VM_PROT_READ | PMAP_WIRED);
1483 pmap_update(pmap);
1484 }
1485 }
1486
1487
1488 void
1489 pmap_freepagedir(struct pmap *pmap)
1490 {
1491 /* Free the memory used for the page table mapping */
1492 if (pmap->pm_vptpt != 0)
1493 uvm_km_free(kernel_map, (vaddr_t)pmap->pm_vptpt, NBPG);
1494
1495 /* junk the L1 page table */
1496 if (pmap->pm_l1pt->pt_flags & PTFLAG_STATIC) {
1497 /* Add the page table to the queue */
1498 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pmap->pm_l1pt, pt_queue);
1499 ++l1pt_static_queue_count;
1500 } else if (l1pt_queue_count < 8) {
1501 /* Add the page table to the queue */
1502 SIMPLEQ_INSERT_TAIL(&l1pt_queue, pmap->pm_l1pt, pt_queue);
1503 ++l1pt_queue_count;
1504 } else
1505 pmap_free_l1pt(pmap->pm_l1pt);
1506 }
1507
1508
1509 /*
1510 * Retire the given physical map from service.
1511 * Should only be called if the map contains no valid mappings.
1512 */
1513
1514 void
1515 pmap_destroy(struct pmap *pmap)
1516 {
1517 struct vm_page *page;
1518 int count;
1519
1520 if (pmap == NULL)
1521 return;
1522
1523 PDEBUG(0, printf("pmap_destroy(%p)\n", pmap));
1524
1525 /*
1526 * Drop reference count
1527 */
1528 simple_lock(&pmap->pm_obj.vmobjlock);
1529 count = --pmap->pm_obj.uo_refs;
1530 simple_unlock(&pmap->pm_obj.vmobjlock);
1531 if (count > 0) {
1532 return;
1533 }
1534
1535 /*
1536 * reference count is zero, free pmap resources and then free pmap.
1537 */
1538
1539 /*
1540 * remove it from global list of pmaps
1541 */
1542
1543 simple_lock(&pmaps_lock);
1544 LIST_REMOVE(pmap, pm_list);
1545 simple_unlock(&pmaps_lock);
1546
1547 if (vector_page < KERNEL_BASE) {
1548 /* Remove the vector page mapping */
1549 pmap_remove(pmap, vector_page, vector_page + NBPG);
1550 pmap_update(pmap);
1551 }
1552
1553 /*
1554 * Free any page tables still mapped
1555 * This is only temporay until pmap_enter can count the number
1556 * of mappings made in a page table. Then pmap_remove() can
1557 * reduce the count and free the pagetable when the count
1558 * reaches zero. Note that entries in this list should match the
1559 * contents of the ptpt, however this is faster than walking a 1024
1560 * entries looking for pt's
1561 * taken from i386 pmap.c
1562 */
1563 while ((page = TAILQ_FIRST(&pmap->pm_obj.memq)) != NULL) {
1564 KASSERT((page->flags & PG_BUSY) == 0);
1565 page->wire_count = 0;
1566 uvm_pagefree(page);
1567 }
1568
1569 /* Free the page dir */
1570 pmap_freepagedir(pmap);
1571
1572 /* return the pmap to the pool */
1573 pool_put(&pmap_pmap_pool, pmap);
1574 }
1575
1576
1577 /*
1578 * void pmap_reference(struct pmap *pmap)
1579 *
1580 * Add a reference to the specified pmap.
1581 */
1582
1583 void
1584 pmap_reference(struct pmap *pmap)
1585 {
1586 if (pmap == NULL)
1587 return;
1588
1589 simple_lock(&pmap->pm_lock);
1590 pmap->pm_obj.uo_refs++;
1591 simple_unlock(&pmap->pm_lock);
1592 }
1593
1594 /*
1595 * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
1596 *
1597 * Return the start and end addresses of the kernel's virtual space.
1598 * These values are setup in pmap_bootstrap and are updated as pages
1599 * are allocated.
1600 */
1601
1602 void
1603 pmap_virtual_space(vaddr_t *start, vaddr_t *end)
1604 {
1605 *start = virtual_avail;
1606 *end = virtual_end;
1607 }
1608
1609 /*
1610 * Activate the address space for the specified process. If the process
1611 * is the current process, load the new MMU context.
1612 */
1613 void
1614 pmap_activate(struct proc *p)
1615 {
1616 struct pmap *pmap = p->p_vmspace->vm_map.pmap;
1617 struct pcb *pcb = &p->p_addr->u_pcb;
1618
1619 (void) pmap_extract(pmap_kernel(), (vaddr_t)pmap->pm_pdir,
1620 (paddr_t *)&pcb->pcb_pagedir);
1621
1622 PDEBUG(0, printf("pmap_activate: p=%p pmap=%p pcb=%p pdir=%p l1=%p\n",
1623 p, pmap, pcb, pmap->pm_pdir, pcb->pcb_pagedir));
1624
1625 if (p == curproc) {
1626 PDEBUG(0, printf("pmap_activate: setting TTB\n"));
1627 setttb((u_int)pcb->pcb_pagedir);
1628 }
1629 }
1630
1631 /*
1632 * Deactivate the address space of the specified process.
1633 */
1634 void
1635 pmap_deactivate(struct proc *p)
1636 {
1637 }
1638
1639 /*
1640 * Perform any deferred pmap operations.
1641 */
1642 void
1643 pmap_update(struct pmap *pmap)
1644 {
1645
1646 /*
1647 * We haven't deferred any pmap operations, but we do need to
1648 * make sure TLB/cache operations have completed.
1649 */
1650 cpu_cpwait();
1651 }
1652
1653 /*
1654 * pmap_clean_page()
1655 *
1656 * This is a local function used to work out the best strategy to clean
1657 * a single page referenced by its entry in the PV table. It's used by
1658 * pmap_copy_page, pmap_zero page and maybe some others later on.
1659 *
1660 * Its policy is effectively:
1661 * o If there are no mappings, we don't bother doing anything with the cache.
1662 * o If there is one mapping, we clean just that page.
1663 * o If there are multiple mappings, we clean the entire cache.
1664 *
1665 * So that some functions can be further optimised, it returns 0 if it didn't
1666 * clean the entire cache, or 1 if it did.
1667 *
1668 * XXX One bug in this routine is that if the pv_entry has a single page
1669 * mapped at 0x00000000 a whole cache clean will be performed rather than
1670 * just the 1 page. Since this should not occur in everyday use and if it does
1671 * it will just result in not the most efficient clean for the page.
1672 */
1673 static int
1674 pmap_clean_page(struct pv_entry *pv, boolean_t is_src)
1675 {
1676 struct pmap *pmap;
1677 struct pv_entry *npv;
1678 int cache_needs_cleaning = 0;
1679 vaddr_t page_to_clean = 0;
1680
1681 if (pv == NULL)
1682 /* nothing mapped in so nothing to flush */
1683 return (0);
1684
1685 /* Since we flush the cache each time we change curproc, we
1686 * only need to flush the page if it is in the current pmap.
1687 */
1688 if (curproc)
1689 pmap = curproc->p_vmspace->vm_map.pmap;
1690 else
1691 pmap = pmap_kernel();
1692
1693 for (npv = pv; npv; npv = npv->pv_next) {
1694 if (npv->pv_pmap == pmap) {
1695 /* The page is mapped non-cacheable in
1696 * this map. No need to flush the cache.
1697 */
1698 if (npv->pv_flags & PVF_NC) {
1699 #ifdef DIAGNOSTIC
1700 if (cache_needs_cleaning)
1701 panic("pmap_clean_page: "
1702 "cache inconsistency");
1703 #endif
1704 break;
1705 }
1706 #if 0
1707 /* This doesn't work, because pmap_protect
1708 doesn't flush changes on pages that it
1709 has write-protected. */
1710
1711 /* If the page is not writable and this
1712 is the source, then there is no need
1713 to flush it from the cache. */
1714 else if (is_src && ! (npv->pv_flags & PVF_WRITE))
1715 continue;
1716 #endif
1717 if (cache_needs_cleaning){
1718 page_to_clean = 0;
1719 break;
1720 }
1721 else
1722 page_to_clean = npv->pv_va;
1723 cache_needs_cleaning = 1;
1724 }
1725 }
1726
1727 if (page_to_clean)
1728 cpu_idcache_wbinv_range(page_to_clean, NBPG);
1729 else if (cache_needs_cleaning) {
1730 cpu_idcache_wbinv_all();
1731 return (1);
1732 }
1733 return (0);
1734 }
1735
1736 /*
1737 * pmap_zero_page()
1738 *
1739 * Zero a given physical page by mapping it at a page hook point.
1740 * In doing the zero page op, the page we zero is mapped cachable, as with
1741 * StrongARM accesses to non-cached pages are non-burst making writing
1742 * _any_ bulk data very slow.
1743 */
1744 void
1745 pmap_zero_page(paddr_t phys)
1746 {
1747 #ifdef DEBUG
1748 struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
1749
1750 if (pg->mdpage.pvh_list != NULL)
1751 panic("pmap_zero_page: page has mappings");
1752 #endif
1753
1754 KDASSERT((phys & PGOFSET) == 0);
1755
1756 /*
1757 * Hook in the page, zero it, and purge the cache for that
1758 * zeroed page. Invalidate the TLB as needed.
1759 */
1760 *cdst_pte = L2_S_PROTO | phys |
1761 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_cache_mode;
1762 cpu_tlb_flushD_SE(cdstp);
1763 cpu_cpwait();
1764 bzero_page(cdstp);
1765 cpu_dcache_wbinv_range(cdstp, NBPG);
1766 }
1767
1768 /* pmap_pageidlezero()
1769 *
1770 * The same as above, except that we assume that the page is not
1771 * mapped. This means we never have to flush the cache first. Called
1772 * from the idle loop.
1773 */
1774 boolean_t
1775 pmap_pageidlezero(paddr_t phys)
1776 {
1777 int i, *ptr;
1778 boolean_t rv = TRUE;
1779 #ifdef DEBUG
1780 struct vm_page *pg;
1781
1782 pg = PHYS_TO_VM_PAGE(phys);
1783 if (pg->mdpage.pvh_list != NULL)
1784 panic("pmap_pageidlezero: page has mappings");
1785 #endif
1786
1787 KDASSERT((phys & PGOFSET) == 0);
1788
1789 /*
1790 * Hook in the page, zero it, and purge the cache for that
1791 * zeroed page. Invalidate the TLB as needed.
1792 */
1793 *cdst_pte = L2_S_PROTO | phys |
1794 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_cache_mode;
1795 cpu_tlb_flushD_SE(cdstp);
1796 cpu_cpwait();
1797
1798 for (i = 0, ptr = (int *)cdstp;
1799 i < (NBPG / sizeof(int)); i++) {
1800 if (sched_whichqs != 0) {
1801 /*
1802 * A process has become ready. Abort now,
1803 * so we don't keep it waiting while we
1804 * do slow memory access to finish this
1805 * page.
1806 */
1807 rv = FALSE;
1808 break;
1809 }
1810 *ptr++ = 0;
1811 }
1812
1813 if (rv)
1814 /*
1815 * if we aborted we'll rezero this page again later so don't
1816 * purge it unless we finished it
1817 */
1818 cpu_dcache_wbinv_range(cdstp, NBPG);
1819 return (rv);
1820 }
1821
1822 /*
1823 * pmap_copy_page()
1824 *
1825 * Copy one physical page into another, by mapping the pages into
1826 * hook points. The same comment regarding cachability as in
1827 * pmap_zero_page also applies here.
1828 */
1829 void
1830 pmap_copy_page(paddr_t src, paddr_t dst)
1831 {
1832 struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
1833 #ifdef DEBUG
1834 struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst);
1835
1836 if (dst_pg->mdpage.pvh_list != NULL)
1837 panic("pmap_copy_page: dst page has mappings");
1838 #endif
1839
1840 KDASSERT((src & PGOFSET) == 0);
1841 KDASSERT((dst & PGOFSET) == 0);
1842
1843 /*
1844 * Clean the source page. Hold the source page's lock for
1845 * the duration of the copy so that no other mappings can
1846 * be created while we have a potentially aliased mapping.
1847 */
1848 simple_lock(&src_pg->mdpage.pvh_slock);
1849 (void) pmap_clean_page(src_pg->mdpage.pvh_list, TRUE);
1850
1851 /*
1852 * Map the pages into the page hook points, copy them, and purge
1853 * the cache for the appropriate page. Invalidate the TLB
1854 * as required.
1855 */
1856 *csrc_pte = L2_S_PROTO | src |
1857 L2_S_PROT(PTE_KERNEL, VM_PROT_READ) | pte_cache_mode;
1858 *cdst_pte = L2_S_PROTO | dst |
1859 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_cache_mode;
1860 cpu_tlb_flushD_SE(csrcp);
1861 cpu_tlb_flushD_SE(cdstp);
1862 cpu_cpwait();
1863 bcopy_page(csrcp, cdstp);
1864 cpu_dcache_inv_range(csrcp, NBPG);
1865 simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
1866 cpu_dcache_wbinv_range(cdstp, NBPG);
1867 }
1868
1869 #if 0
1870 void
1871 pmap_pte_addref(struct pmap *pmap, vaddr_t va)
1872 {
1873 pd_entry_t *pde;
1874 paddr_t pa;
1875 struct vm_page *m;
1876
1877 if (pmap == pmap_kernel())
1878 return;
1879
1880 pde = pmap_pde(pmap, va & ~(3 << L1_S_SHIFT));
1881 pa = pmap_pte_pa(pde);
1882 m = PHYS_TO_VM_PAGE(pa);
1883 ++m->wire_count;
1884 #ifdef MYCROFT_HACK
1885 printf("addref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
1886 pmap, va, pde, pa, m, m->wire_count);
1887 #endif
1888 }
1889
1890 void
1891 pmap_pte_delref(struct pmap *pmap, vaddr_t va)
1892 {
1893 pd_entry_t *pde;
1894 paddr_t pa;
1895 struct vm_page *m;
1896
1897 if (pmap == pmap_kernel())
1898 return;
1899
1900 pde = pmap_pde(pmap, va & ~(3 << L1_S_SHIFT));
1901 pa = pmap_pte_pa(pde);
1902 m = PHYS_TO_VM_PAGE(pa);
1903 --m->wire_count;
1904 #ifdef MYCROFT_HACK
1905 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
1906 pmap, va, pde, pa, m, m->wire_count);
1907 #endif
1908 if (m->wire_count == 0) {
1909 #ifdef MYCROFT_HACK
1910 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p\n",
1911 pmap, va, pde, pa, m);
1912 #endif
1913 pmap_unmap_in_l1(pmap, va);
1914 uvm_pagefree(m);
1915 --pmap->pm_stats.resident_count;
1916 }
1917 }
1918 #else
1919 #define pmap_pte_addref(pmap, va)
1920 #define pmap_pte_delref(pmap, va)
1921 #endif
1922
1923 /*
1924 * Since we have a virtually indexed cache, we may need to inhibit caching if
1925 * there is more than one mapping and at least one of them is writable.
1926 * Since we purge the cache on every context switch, we only need to check for
1927 * other mappings within the same pmap, or kernel_pmap.
1928 * This function is also called when a page is unmapped, to possibly reenable
1929 * caching on any remaining mappings.
1930 *
1931 * The code implements the following logic, where:
1932 *
1933 * KW = # of kernel read/write pages
1934 * KR = # of kernel read only pages
1935 * UW = # of user read/write pages
1936 * UR = # of user read only pages
1937 * OW = # of user read/write pages in another pmap, then
1938 *
1939 * KC = kernel mapping is cacheable
1940 * UC = user mapping is cacheable
1941 *
1942 * KW=0,KR=0 KW=0,KR>0 KW=1,KR=0 KW>1,KR>=0
1943 * +---------------------------------------------
1944 * UW=0,UR=0,OW=0 | --- KC=1 KC=1 KC=0
1945 * UW=0,UR>0,OW=0 | UC=1 KC=1,UC=1 KC=0,UC=0 KC=0,UC=0
1946 * UW=0,UR>0,OW>0 | UC=1 KC=0,UC=1 KC=0,UC=0 KC=0,UC=0
1947 * UW=1,UR=0,OW=0 | UC=1 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0
1948 * UW>1,UR>=0,OW>=0 | UC=0 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0
1949 *
1950 * Note that the pmap must have it's ptes mapped in, and passed with ptes.
1951 */
1952 __inline static void
1953 pmap_vac_me_harder(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
1954 boolean_t clear_cache)
1955 {
1956 if (pmap == pmap_kernel())
1957 pmap_vac_me_kpmap(pmap, pg, ptes, clear_cache);
1958 else
1959 pmap_vac_me_user(pmap, pg, ptes, clear_cache);
1960 }
1961
1962 static void
1963 pmap_vac_me_kpmap(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
1964 boolean_t clear_cache)
1965 {
1966 int user_entries = 0;
1967 int user_writable = 0;
1968 int user_cacheable = 0;
1969 int kernel_entries = 0;
1970 int kernel_writable = 0;
1971 int kernel_cacheable = 0;
1972 struct pv_entry *pv;
1973 struct pmap *last_pmap = pmap;
1974
1975 #ifdef DIAGNOSTIC
1976 if (pmap != pmap_kernel())
1977 panic("pmap_vac_me_kpmap: pmap != pmap_kernel()");
1978 #endif
1979
1980 /*
1981 * Pass one, see if there are both kernel and user pmaps for
1982 * this page. Calculate whether there are user-writable or
1983 * kernel-writable pages.
1984 */
1985 for (pv = pg->mdpage.pvh_list; pv != NULL; pv = pv->pv_next) {
1986 if (pv->pv_pmap != pmap) {
1987 user_entries++;
1988 if (pv->pv_flags & PVF_WRITE)
1989 user_writable++;
1990 if ((pv->pv_flags & PVF_NC) == 0)
1991 user_cacheable++;
1992 } else {
1993 kernel_entries++;
1994 if (pv->pv_flags & PVF_WRITE)
1995 kernel_writable++;
1996 if ((pv->pv_flags & PVF_NC) == 0)
1997 kernel_cacheable++;
1998 }
1999 }
2000
2001 /*
2002 * We know we have just been updating a kernel entry, so if
2003 * all user pages are already cacheable, then there is nothing
2004 * further to do.
2005 */
2006 if (kernel_entries == 0 &&
2007 user_cacheable == user_entries)
2008 return;
2009
2010 if (user_entries) {
2011 /*
2012 * Scan over the list again, for each entry, if it
2013 * might not be set correctly, call pmap_vac_me_user
2014 * to recalculate the settings.
2015 */
2016 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
2017 /*
2018 * We know kernel mappings will get set
2019 * correctly in other calls. We also know
2020 * that if the pmap is the same as last_pmap
2021 * then we've just handled this entry.
2022 */
2023 if (pv->pv_pmap == pmap || pv->pv_pmap == last_pmap)
2024 continue;
2025 /*
2026 * If there are kernel entries and this page
2027 * is writable but non-cacheable, then we can
2028 * skip this entry also.
2029 */
2030 if (kernel_entries > 0 &&
2031 (pv->pv_flags & (PVF_NC | PVF_WRITE)) ==
2032 (PVF_NC | PVF_WRITE))
2033 continue;
2034 /*
2035 * Similarly if there are no kernel-writable
2036 * entries and the page is already
2037 * read-only/cacheable.
2038 */
2039 if (kernel_writable == 0 &&
2040 (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 0)
2041 continue;
2042 /*
2043 * For some of the remaining cases, we know
2044 * that we must recalculate, but for others we
2045 * can't tell if they are correct or not, so
2046 * we recalculate anyway.
2047 */
2048 pmap_unmap_ptes(last_pmap);
2049 last_pmap = pv->pv_pmap;
2050 ptes = pmap_map_ptes(last_pmap);
2051 pmap_vac_me_user(last_pmap, pg, ptes,
2052 pmap_is_curpmap(last_pmap));
2053 }
2054 /* Restore the pte mapping that was passed to us. */
2055 if (last_pmap != pmap) {
2056 pmap_unmap_ptes(last_pmap);
2057 ptes = pmap_map_ptes(pmap);
2058 }
2059 if (kernel_entries == 0)
2060 return;
2061 }
2062
2063 pmap_vac_me_user(pmap, pg, ptes, clear_cache);
2064 return;
2065 }
2066
2067 static void
2068 pmap_vac_me_user(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2069 boolean_t clear_cache)
2070 {
2071 struct pmap *kpmap = pmap_kernel();
2072 struct pv_entry *pv, *npv;
2073 int entries = 0;
2074 int writable = 0;
2075 int cacheable_entries = 0;
2076 int kern_cacheable = 0;
2077 int other_writable = 0;
2078
2079 pv = pg->mdpage.pvh_list;
2080 KASSERT(ptes != NULL);
2081
2082 /*
2083 * Count mappings and writable mappings in this pmap.
2084 * Include kernel mappings as part of our own.
2085 * Keep a pointer to the first one.
2086 */
2087 for (npv = pv; npv; npv = npv->pv_next) {
2088 /* Count mappings in the same pmap */
2089 if (pmap == npv->pv_pmap ||
2090 kpmap == npv->pv_pmap) {
2091 if (entries++ == 0)
2092 pv = npv;
2093 /* Cacheable mappings */
2094 if ((npv->pv_flags & PVF_NC) == 0) {
2095 cacheable_entries++;
2096 if (kpmap == npv->pv_pmap)
2097 kern_cacheable++;
2098 }
2099 /* Writable mappings */
2100 if (npv->pv_flags & PVF_WRITE)
2101 ++writable;
2102 } else if (npv->pv_flags & PVF_WRITE)
2103 other_writable = 1;
2104 }
2105
2106 PDEBUG(3,printf("pmap_vac_me_harder: pmap %p Entries %d, "
2107 "writable %d cacheable %d %s\n", pmap, entries, writable,
2108 cacheable_entries, clear_cache ? "clean" : "no clean"));
2109
2110 /*
2111 * Enable or disable caching as necessary.
2112 * Note: the first entry might be part of the kernel pmap,
2113 * so we can't assume this is indicative of the state of the
2114 * other (maybe non-kpmap) entries.
2115 */
2116 if ((entries > 1 && writable) ||
2117 (entries > 0 && pmap == kpmap && other_writable)) {
2118 if (cacheable_entries == 0)
2119 return;
2120 for (npv = pv; npv; npv = npv->pv_next) {
2121 if ((pmap == npv->pv_pmap
2122 || kpmap == npv->pv_pmap) &&
2123 (npv->pv_flags & PVF_NC) == 0) {
2124 ptes[arm_btop(npv->pv_va)] &= ~(L2_C | L2_B);
2125 npv->pv_flags |= PVF_NC;
2126 /*
2127 * If this page needs flushing from the
2128 * cache, and we aren't going to do it
2129 * below, do it now.
2130 */
2131 if ((cacheable_entries < 4 &&
2132 (clear_cache || npv->pv_pmap == kpmap)) ||
2133 (npv->pv_pmap == kpmap &&
2134 !clear_cache && kern_cacheable < 4)) {
2135 cpu_idcache_wbinv_range(npv->pv_va,
2136 NBPG);
2137 cpu_tlb_flushID_SE(npv->pv_va);
2138 }
2139 }
2140 }
2141 if ((clear_cache && cacheable_entries >= 4) ||
2142 kern_cacheable >= 4) {
2143 cpu_idcache_wbinv_all();
2144 cpu_tlb_flushID();
2145 }
2146 cpu_cpwait();
2147 } else if (entries > 0) {
2148 /*
2149 * Turn cacheing back on for some pages. If it is a kernel
2150 * page, only do so if there are no other writable pages.
2151 */
2152 for (npv = pv; npv; npv = npv->pv_next) {
2153 if ((pmap == npv->pv_pmap ||
2154 (kpmap == npv->pv_pmap && other_writable == 0)) &&
2155 (npv->pv_flags & PVF_NC)) {
2156 ptes[arm_btop(npv->pv_va)] |= pte_cache_mode;
2157 npv->pv_flags &= ~PVF_NC;
2158 }
2159 }
2160 }
2161 }
2162
2163 /*
2164 * pmap_remove()
2165 *
2166 * pmap_remove is responsible for nuking a number of mappings for a range
2167 * of virtual address space in the current pmap. To do this efficiently
2168 * is interesting, because in a number of cases a wide virtual address
2169 * range may be supplied that contains few actual mappings. So, the
2170 * optimisations are:
2171 * 1. Try and skip over hunks of address space for which an L1 entry
2172 * does not exist.
2173 * 2. Build up a list of pages we've hit, up to a maximum, so we can
2174 * maybe do just a partial cache clean. This path of execution is
2175 * complicated by the fact that the cache must be flushed _before_
2176 * the PTE is nuked, being a VAC :-)
2177 * 3. Maybe later fast-case a single page, but I don't think this is
2178 * going to make _that_ much difference overall.
2179 */
2180
2181 #define PMAP_REMOVE_CLEAN_LIST_SIZE 3
2182
2183 void
2184 pmap_remove(struct pmap *pmap, vaddr_t sva, vaddr_t eva)
2185 {
2186 int cleanlist_idx = 0;
2187 struct pagelist {
2188 vaddr_t va;
2189 pt_entry_t *pte;
2190 } cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
2191 pt_entry_t *pte = 0, *ptes;
2192 paddr_t pa;
2193 int pmap_active;
2194 struct vm_page *pg;
2195
2196 /* Exit quick if there is no pmap */
2197 if (!pmap)
2198 return;
2199
2200 PDEBUG(0, printf("pmap_remove: pmap=%p sva=%08lx eva=%08lx\n",
2201 pmap, sva, eva));
2202
2203 /*
2204 * we lock in the pmap => vm_page direction
2205 */
2206 PMAP_MAP_TO_HEAD_LOCK();
2207
2208 ptes = pmap_map_ptes(pmap);
2209 /* Get a page table pointer */
2210 while (sva < eva) {
2211 if (pmap_pde_page(pmap_pde(pmap, sva)))
2212 break;
2213 sva = (sva & L1_S_FRAME) + L1_S_SIZE;
2214 }
2215
2216 pte = &ptes[arm_btop(sva)];
2217 /* Note if the pmap is active thus require cache and tlb cleans */
2218 pmap_active = pmap_is_curpmap(pmap);
2219
2220 /* Now loop along */
2221 while (sva < eva) {
2222 /* Check if we can move to the next PDE (l1 chunk) */
2223 if (!(sva & L2_ADDR_BITS))
2224 if (!pmap_pde_page(pmap_pde(pmap, sva))) {
2225 sva += L1_S_SIZE;
2226 pte += arm_btop(L1_S_SIZE);
2227 continue;
2228 }
2229
2230 /* We've found a valid PTE, so this page of PTEs has to go. */
2231 if (pmap_pte_v(pte)) {
2232 /* Update statistics */
2233 --pmap->pm_stats.resident_count;
2234
2235 /*
2236 * Add this page to our cache remove list, if we can.
2237 * If, however the cache remove list is totally full,
2238 * then do a complete cache invalidation taking note
2239 * to backtrack the PTE table beforehand, and ignore
2240 * the lists in future because there's no longer any
2241 * point in bothering with them (we've paid the
2242 * penalty, so will carry on unhindered). Otherwise,
2243 * when we fall out, we just clean the list.
2244 */
2245 PDEBUG(10, printf("remove: inv pte at %p(%x) ", pte, *pte));
2246 pa = pmap_pte_pa(pte);
2247
2248 if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
2249 /* Add to the clean list. */
2250 cleanlist[cleanlist_idx].pte = pte;
2251 cleanlist[cleanlist_idx].va = sva;
2252 cleanlist_idx++;
2253 } else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
2254 int cnt;
2255
2256 /* Nuke everything if needed. */
2257 if (pmap_active) {
2258 cpu_idcache_wbinv_all();
2259 cpu_tlb_flushID();
2260 }
2261
2262 /*
2263 * Roll back the previous PTE list,
2264 * and zero out the current PTE.
2265 */
2266 for (cnt = 0; cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
2267 *cleanlist[cnt].pte = 0;
2268 pmap_pte_delref(pmap, cleanlist[cnt].va);
2269 }
2270 *pte = 0;
2271 pmap_pte_delref(pmap, sva);
2272 cleanlist_idx++;
2273 } else {
2274 /*
2275 * We've already nuked the cache and
2276 * TLB, so just carry on regardless,
2277 * and we won't need to do it again
2278 */
2279 *pte = 0;
2280 pmap_pte_delref(pmap, sva);
2281 }
2282
2283 /*
2284 * Update flags. In a number of circumstances,
2285 * we could cluster a lot of these and do a
2286 * number of sequential pages in one go.
2287 */
2288 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
2289 struct pv_entry *pve;
2290 simple_lock(&pg->mdpage.pvh_slock);
2291 pve = pmap_remove_pv(pg, pmap, sva);
2292 pmap_free_pv(pmap, pve);
2293 pmap_vac_me_harder(pmap, pg, ptes, FALSE);
2294 simple_unlock(&pg->mdpage.pvh_slock);
2295 }
2296 }
2297 sva += NBPG;
2298 pte++;
2299 }
2300
2301 pmap_unmap_ptes(pmap);
2302 /*
2303 * Now, if we've fallen through down to here, chances are that there
2304 * are less than PMAP_REMOVE_CLEAN_LIST_SIZE mappings left.
2305 */
2306 if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
2307 u_int cnt;
2308
2309 for (cnt = 0; cnt < cleanlist_idx; cnt++) {
2310 if (pmap_active) {
2311 cpu_idcache_wbinv_range(cleanlist[cnt].va,
2312 NBPG);
2313 *cleanlist[cnt].pte = 0;
2314 cpu_tlb_flushID_SE(cleanlist[cnt].va);
2315 } else
2316 *cleanlist[cnt].pte = 0;
2317 pmap_pte_delref(pmap, cleanlist[cnt].va);
2318 }
2319 }
2320 PMAP_MAP_TO_HEAD_UNLOCK();
2321 }
2322
2323 /*
2324 * Routine: pmap_remove_all
2325 * Function:
2326 * Removes this physical page from
2327 * all physical maps in which it resides.
2328 * Reflects back modify bits to the pager.
2329 */
2330
2331 static void
2332 pmap_remove_all(struct vm_page *pg)
2333 {
2334 struct pv_entry *pv, *npv;
2335 struct pmap *pmap;
2336 pt_entry_t *pte, *ptes;
2337
2338 PDEBUG(0, printf("pmap_remove_all: pa=%lx ", VM_PAGE_TO_PHYS(pg)));
2339
2340 /* set vm_page => pmap locking */
2341 PMAP_HEAD_TO_MAP_LOCK();
2342
2343 simple_lock(&pg->mdpage.pvh_slock);
2344
2345 pv = pg->mdpage.pvh_list;
2346 if (pv == NULL) {
2347 PDEBUG(0, printf("free page\n"));
2348 simple_unlock(&pg->mdpage.pvh_slock);
2349 PMAP_HEAD_TO_MAP_UNLOCK();
2350 return;
2351 }
2352 pmap_clean_page(pv, FALSE);
2353
2354 while (pv) {
2355 pmap = pv->pv_pmap;
2356 ptes = pmap_map_ptes(pmap);
2357 pte = &ptes[arm_btop(pv->pv_va)];
2358
2359 PDEBUG(0, printf("[%p,%08x,%08lx,%08x] ", pmap, *pte,
2360 pv->pv_va, pv->pv_flags));
2361 #ifdef DEBUG
2362 if (pmap_pde_page(pmap_pde(pmap, pv->pv_va)) == 0 ||
2363 pmap_pte_v(pte) == 0 ||
2364 pmap_pte_pa(pte) != VM_PAGE_TO_PHYS(pg))
2365 panic("pmap_remove_all: bad mapping");
2366 #endif /* DEBUG */
2367
2368 /*
2369 * Update statistics
2370 */
2371 --pmap->pm_stats.resident_count;
2372
2373 /* Wired bit */
2374 if (pv->pv_flags & PVF_WIRED)
2375 --pmap->pm_stats.wired_count;
2376
2377 /*
2378 * Invalidate the PTEs.
2379 * XXX: should cluster them up and invalidate as many
2380 * as possible at once.
2381 */
2382
2383 #ifdef needednotdone
2384 reduce wiring count on page table pages as references drop
2385 #endif
2386
2387 *pte = 0;
2388 pmap_pte_delref(pmap, pv->pv_va);
2389
2390 npv = pv->pv_next;
2391 pmap_free_pv(pmap, pv);
2392 pv = npv;
2393 pmap_unmap_ptes(pmap);
2394 }
2395 pg->mdpage.pvh_list = NULL;
2396 simple_unlock(&pg->mdpage.pvh_slock);
2397 PMAP_HEAD_TO_MAP_UNLOCK();
2398
2399 PDEBUG(0, printf("done\n"));
2400 cpu_tlb_flushID();
2401 cpu_cpwait();
2402 }
2403
2404
2405 /*
2406 * Set the physical protection on the specified range of this map as requested.
2407 */
2408
2409 void
2410 pmap_protect(struct pmap *pmap, vaddr_t sva, vaddr_t eva, vm_prot_t prot)
2411 {
2412 pt_entry_t *pte = NULL, *ptes;
2413 struct vm_page *pg;
2414 int armprot;
2415 int flush = 0;
2416 paddr_t pa;
2417
2418 PDEBUG(0, printf("pmap_protect: pmap=%p %08lx->%08lx %x\n",
2419 pmap, sva, eva, prot));
2420
2421 if (~prot & VM_PROT_READ) {
2422 /* Just remove the mappings. */
2423 pmap_remove(pmap, sva, eva);
2424 /* pmap_update not needed as it should be called by the caller
2425 * of pmap_protect */
2426 return;
2427 }
2428 if (prot & VM_PROT_WRITE) {
2429 /*
2430 * If this is a read->write transition, just ignore it and let
2431 * uvm_fault() take care of it later.
2432 */
2433 return;
2434 }
2435
2436 /* Need to lock map->head */
2437 PMAP_MAP_TO_HEAD_LOCK();
2438
2439 ptes = pmap_map_ptes(pmap);
2440 /*
2441 * We need to acquire a pointer to a page table page before entering
2442 * the following loop.
2443 */
2444 while (sva < eva) {
2445 if (pmap_pde_page(pmap_pde(pmap, sva)))
2446 break;
2447 sva = (sva & L1_S_FRAME) + L1_S_SIZE;
2448 }
2449
2450 pte = &ptes[arm_btop(sva)];
2451
2452 while (sva < eva) {
2453 /* only check once in a while */
2454 if ((sva & L2_ADDR_BITS) == 0) {
2455 if (!pmap_pde_page(pmap_pde(pmap, sva))) {
2456 /* We can race ahead here, to the next pde. */
2457 sva += L1_S_SIZE;
2458 pte += arm_btop(L1_S_SIZE);
2459 continue;
2460 }
2461 }
2462
2463 if (!pmap_pte_v(pte))
2464 goto next;
2465
2466 flush = 1;
2467
2468 armprot = 0;
2469 if (sva < VM_MAXUSER_ADDRESS)
2470 armprot |= L2_S_PROT_U;
2471 else if (sva < VM_MAX_ADDRESS)
2472 armprot |= L2_S_PROT_W; /* XXX Ekk what is this ? */
2473 *pte = (*pte & 0xfffff00f) | armprot;
2474
2475 pa = pmap_pte_pa(pte);
2476
2477 /* Get the physical page index */
2478
2479 /* Clear write flag */
2480 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
2481 simple_lock(&pg->mdpage.pvh_slock);
2482 (void) pmap_modify_pv(pmap, sva, pg, PVF_WRITE, 0);
2483 pmap_vac_me_harder(pmap, pg, ptes, FALSE);
2484 simple_unlock(&pg->mdpage.pvh_slock);
2485 }
2486
2487 next:
2488 sva += NBPG;
2489 pte++;
2490 }
2491 pmap_unmap_ptes(pmap);
2492 PMAP_MAP_TO_HEAD_UNLOCK();
2493 if (flush)
2494 cpu_tlb_flushID();
2495 }
2496
2497 /*
2498 * void pmap_enter(struct pmap *pmap, vaddr_t va, paddr_t pa, vm_prot_t prot,
2499 * int flags)
2500 *
2501 * Insert the given physical page (p) at
2502 * the specified virtual address (v) in the
2503 * target physical map with the protection requested.
2504 *
2505 * If specified, the page will be wired down, meaning
2506 * that the related pte can not be reclaimed.
2507 *
2508 * NB: This is the only routine which MAY NOT lazy-evaluate
2509 * or lose information. That is, this routine must actually
2510 * insert this page into the given map NOW.
2511 */
2512
2513 int
2514 pmap_enter(struct pmap *pmap, vaddr_t va, paddr_t pa, vm_prot_t prot,
2515 int flags)
2516 {
2517 pt_entry_t *ptes, opte, npte;
2518 paddr_t opa;
2519 boolean_t wired = (flags & PMAP_WIRED) != 0;
2520 struct vm_page *pg;
2521 struct pv_entry *pve;
2522 int error, nflags;
2523
2524 PDEBUG(5, printf("pmap_enter: V%08lx P%08lx in pmap %p prot=%08x, wired = %d\n",
2525 va, pa, pmap, prot, wired));
2526
2527 #ifdef DIAGNOSTIC
2528 /* Valid address ? */
2529 if (va >= (pmap_curmaxkvaddr))
2530 panic("pmap_enter: too big");
2531 if (pmap != pmap_kernel() && va != 0) {
2532 if (va < VM_MIN_ADDRESS || va >= VM_MAXUSER_ADDRESS)
2533 panic("pmap_enter: kernel page in user map");
2534 } else {
2535 if (va >= VM_MIN_ADDRESS && va < VM_MAXUSER_ADDRESS)
2536 panic("pmap_enter: user page in kernel map");
2537 if (va >= VM_MAXUSER_ADDRESS && va < VM_MAX_ADDRESS)
2538 panic("pmap_enter: entering PT page");
2539 }
2540 #endif
2541
2542 KDASSERT(((va | pa) & PGOFSET) == 0);
2543
2544 /*
2545 * Get a pointer to the page. Later on in this function, we
2546 * test for a managed page by checking pg != NULL.
2547 */
2548 pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
2549
2550 /* get lock */
2551 PMAP_MAP_TO_HEAD_LOCK();
2552
2553 /*
2554 * map the ptes. If there's not already an L2 table for this
2555 * address, allocate one.
2556 */
2557 ptes = pmap_map_ptes(pmap); /* locks pmap */
2558 if (pmap_pde_v(pmap_pde(pmap, va)) == 0) {
2559 struct vm_page *ptp;
2560
2561 /* kernel should be pre-grown */
2562 KASSERT(pmap != pmap_kernel());
2563
2564 /* if failure is allowed then don't try too hard */
2565 ptp = pmap_get_ptp(pmap, va & L1_S_FRAME);
2566 if (ptp == NULL) {
2567 if (flags & PMAP_CANFAIL) {
2568 error = ENOMEM;
2569 goto out;
2570 }
2571 panic("pmap_enter: get ptp failed");
2572 }
2573 }
2574 opte = ptes[arm_btop(va)];
2575
2576 nflags = 0;
2577 if (prot & VM_PROT_WRITE)
2578 nflags |= PVF_WRITE;
2579 if (wired)
2580 nflags |= PVF_WIRED;
2581
2582 /* Is the pte valid ? If so then this page is already mapped */
2583 if (l2pte_valid(opte)) {
2584 /* Get the physical address of the current page mapped */
2585 opa = l2pte_pa(opte);
2586
2587 /* Are we mapping the same page ? */
2588 if (opa == pa) {
2589 /* Has the wiring changed ? */
2590 if (pg != NULL) {
2591 simple_lock(&pg->mdpage.pvh_slock);
2592 (void) pmap_modify_pv(pmap, va, pg,
2593 PVF_WRITE | PVF_WIRED, nflags);
2594 simple_unlock(&pg->mdpage.pvh_slock);
2595 }
2596 } else {
2597 struct vm_page *opg;
2598
2599 /* We are replacing the page with a new one. */
2600 cpu_idcache_wbinv_range(va, NBPG);
2601
2602 /*
2603 * If it is part of our managed memory then we
2604 * must remove it from the PV list
2605 */
2606 if ((opg = PHYS_TO_VM_PAGE(opa)) != NULL) {
2607 simple_lock(&opg->mdpage.pvh_slock);
2608 pve = pmap_remove_pv(opg, pmap, va);
2609 simple_unlock(&opg->mdpage.pvh_slock);
2610 } else {
2611 pve = NULL;
2612 }
2613
2614 goto enter;
2615 }
2616 } else {
2617 opa = 0;
2618 pve = NULL;
2619 pmap_pte_addref(pmap, va);
2620
2621 /* pte is not valid so we must be hooking in a new page */
2622 ++pmap->pm_stats.resident_count;
2623
2624 enter:
2625 /*
2626 * Enter on the PV list if part of our managed memory
2627 */
2628 if (pg != NULL) {
2629 if (pve == NULL) {
2630 pve = pmap_alloc_pv(pmap, ALLOCPV_NEED);
2631 if (pve == NULL) {
2632 if (flags & PMAP_CANFAIL) {
2633 error = ENOMEM;
2634 goto out;
2635 }
2636 panic("pmap_enter: no pv entries "
2637 "available");
2638 }
2639 }
2640 /* enter_pv locks pvh when adding */
2641 pmap_enter_pv(pg, pve, pmap, va, NULL, nflags);
2642 } else {
2643 if (pve != NULL)
2644 pmap_free_pv(pmap, pve);
2645 }
2646 }
2647
2648 /* Construct the pte, giving the correct access. */
2649 npte = pa;
2650
2651 /* VA 0 is magic. */
2652 if (pmap != pmap_kernel() && va != vector_page)
2653 npte |= L2_S_PROT_U;
2654
2655 if (pg != NULL) {
2656 #ifdef DIAGNOSTIC
2657 if ((flags & VM_PROT_ALL) & ~prot)
2658 panic("pmap_enter: access_type exceeds prot");
2659 #endif
2660 npte |= pte_cache_mode;
2661 if (flags & VM_PROT_WRITE) {
2662 npte |= L2_S_PROTO | L2_S_PROT_W;
2663 pg->mdpage.pvh_attrs |= PVF_REF | PVF_MOD;
2664 } else if (flags & VM_PROT_ALL) {
2665 npte |= L2_S_PROTO;
2666 pg->mdpage.pvh_attrs |= PVF_REF;
2667 } else
2668 npte |= L2_TYPE_INV;
2669 } else {
2670 if (prot & VM_PROT_WRITE)
2671 npte |= L2_S_PROTO | L2_S_PROT_W;
2672 else if (prot & VM_PROT_ALL)
2673 npte |= L2_S_PROTO;
2674 else
2675 npte |= L2_TYPE_INV;
2676 }
2677
2678 ptes[arm_btop(va)] = npte;
2679
2680 if (pg != NULL) {
2681 simple_lock(&pg->mdpage.pvh_slock);
2682 pmap_vac_me_harder(pmap, pg, ptes, pmap_is_curpmap(pmap));
2683 simple_unlock(&pg->mdpage.pvh_slock);
2684 }
2685
2686 /* Better flush the TLB ... */
2687 cpu_tlb_flushID_SE(va);
2688 error = 0;
2689 out:
2690 pmap_unmap_ptes(pmap); /* unlocks pmap */
2691 PMAP_MAP_TO_HEAD_UNLOCK();
2692
2693 return error;
2694 }
2695
2696 /*
2697 * pmap_kenter_pa: enter a kernel mapping
2698 *
2699 * => no need to lock anything assume va is already allocated
2700 * => should be faster than normal pmap enter function
2701 */
2702 void
2703 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot)
2704 {
2705 pt_entry_t *pte;
2706
2707 pte = vtopte(va);
2708 KASSERT(!pmap_pte_v(pte));
2709
2710 /* XXX r/w! */
2711 *pte = L2_S_PROTO | pa |
2712 L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE) | pte_cache_mode;
2713 }
2714
2715 void
2716 pmap_kremove(vaddr_t va, vsize_t len)
2717 {
2718 pt_entry_t *pte;
2719
2720 for (len >>= PAGE_SHIFT; len > 0; len--, va += PAGE_SIZE) {
2721
2722 /*
2723 * We assume that we will only be called with small
2724 * regions of memory.
2725 */
2726
2727 KASSERT(pmap_pde_page(pmap_pde(pmap_kernel(), va)));
2728 pte = vtopte(va);
2729 cpu_idcache_wbinv_range(va, PAGE_SIZE);
2730 *pte = 0;
2731 cpu_tlb_flushID_SE(va);
2732 }
2733 }
2734
2735 /*
2736 * pmap_page_protect:
2737 *
2738 * Lower the permission for all mappings to a given page.
2739 */
2740
2741 void
2742 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
2743 {
2744
2745 PDEBUG(0, printf("pmap_page_protect(pa=%lx, prot=%d)\n",
2746 VM_PAGE_TO_PHYS(pg), prot));
2747
2748 switch(prot) {
2749 case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
2750 case VM_PROT_READ|VM_PROT_WRITE:
2751 return;
2752
2753 case VM_PROT_READ:
2754 case VM_PROT_READ|VM_PROT_EXECUTE:
2755 pmap_clearbit(pg, PVF_WRITE);
2756 break;
2757
2758 default:
2759 pmap_remove_all(pg);
2760 break;
2761 }
2762 }
2763
2764
2765 /*
2766 * Routine: pmap_unwire
2767 * Function: Clear the wired attribute for a map/virtual-address
2768 * pair.
2769 * In/out conditions:
2770 * The mapping must already exist in the pmap.
2771 */
2772
2773 void
2774 pmap_unwire(struct pmap *pmap, vaddr_t va)
2775 {
2776 pt_entry_t *ptes;
2777 struct vm_page *pg;
2778 paddr_t pa;
2779
2780 PMAP_MAP_TO_HEAD_LOCK();
2781 ptes = pmap_map_ptes(pmap); /* locks pmap */
2782
2783 if (pmap_pde_v(pmap_pde(pmap, va))) {
2784 #ifdef DIAGNOSTIC
2785 if (l2pte_valid(ptes[arm_btop(va)]) == 0)
2786 panic("pmap_unwire: invalid L2 PTE");
2787 #endif
2788 /* Extract the physical address of the page */
2789 pa = l2pte_pa(ptes[arm_btop(va)]);
2790
2791 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
2792 goto out;
2793
2794 /* Update the wired bit in the pv entry for this page. */
2795 simple_lock(&pg->mdpage.pvh_slock);
2796 (void) pmap_modify_pv(pmap, va, pg, PVF_WIRED, 0);
2797 simple_unlock(&pg->mdpage.pvh_slock);
2798 }
2799 #ifdef DIAGNOSTIC
2800 else {
2801 panic("pmap_unwire: invalid L1 PTE");
2802 }
2803 #endif
2804 out:
2805 pmap_unmap_ptes(pmap); /* unlocks pmap */
2806 PMAP_MAP_TO_HEAD_UNLOCK();
2807 }
2808
2809 /*
2810 * Routine: pmap_extract
2811 * Function:
2812 * Extract the physical page address associated
2813 * with the given map/virtual_address pair.
2814 */
2815 boolean_t
2816 pmap_extract(struct pmap *pmap, vaddr_t va, paddr_t *pap)
2817 {
2818 pd_entry_t *pde;
2819 pt_entry_t *pte, *ptes;
2820 paddr_t pa;
2821
2822 PDEBUG(5, printf("pmap_extract: pmap=%p, va=0x%08lx -> ", pmap, va));
2823
2824 ptes = pmap_map_ptes(pmap); /* locks pmap */
2825
2826 pde = pmap_pde(pmap, va);
2827 pte = &ptes[arm_btop(va)];
2828
2829 if (pmap_pde_section(pde)) {
2830 pa = (*pde & L1_S_FRAME) | (va & L1_S_OFFSET);
2831 PDEBUG(5, printf("section pa=0x%08lx\n", pa));
2832 goto out;
2833 } else if (pmap_pde_page(pde) == 0 || pmap_pte_v(pte) == 0) {
2834 PDEBUG(5, printf("no mapping\n"));
2835 goto failed;
2836 }
2837
2838 if ((*pte & L2_TYPE_MASK) == L2_TYPE_L) {
2839 pa = (*pte & L2_L_FRAME) | (va & L2_L_OFFSET);
2840 PDEBUG(5, printf("large page pa=0x%08lx\n", pa));
2841 goto out;
2842 }
2843
2844 pa = (*pte & L2_S_FRAME) | (va & L2_S_OFFSET);
2845 PDEBUG(5, printf("small page pa=0x%08lx\n", pa));
2846
2847 out:
2848 if (pap != NULL)
2849 *pap = pa;
2850
2851 pmap_unmap_ptes(pmap); /* unlocks pmap */
2852 return (TRUE);
2853
2854 failed:
2855 pmap_unmap_ptes(pmap); /* unlocks pmap */
2856 return (FALSE);
2857 }
2858
2859
2860 /*
2861 * pmap_copy:
2862 *
2863 * Copy the range specified by src_addr/len from the source map to the
2864 * range dst_addr/len in the destination map.
2865 *
2866 * This routine is only advisory and need not do anything.
2867 */
2868 /* Call deleted in <arm/arm32/pmap.h> */
2869
2870 #if defined(PMAP_DEBUG)
2871 void
2872 pmap_dump_pvlist(phys, m)
2873 vaddr_t phys;
2874 char *m;
2875 {
2876 struct vm_page *pg;
2877 struct pv_entry *pv;
2878
2879 if ((pg = PHYS_TO_VM_PAGE(phys)) == NULL) {
2880 printf("INVALID PA\n");
2881 return;
2882 }
2883 simple_lock(&pg->mdpage.pvh_slock);
2884 printf("%s %08lx:", m, phys);
2885 if (pg->mdpage.pvh_list == NULL) {
2886 printf(" no mappings\n");
2887 return;
2888 }
2889
2890 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next)
2891 printf(" pmap %p va %08lx flags %08x", pv->pv_pmap,
2892 pv->pv_va, pv->pv_flags);
2893
2894 printf("\n");
2895 simple_unlock(&pg->mdpage.pvh_slock);
2896 }
2897
2898 #endif /* PMAP_DEBUG */
2899
2900 static pt_entry_t *
2901 pmap_map_ptes(struct pmap *pmap)
2902 {
2903 struct proc *p;
2904
2905 /* the kernel's pmap is always accessible */
2906 if (pmap == pmap_kernel()) {
2907 return (pt_entry_t *)PTE_BASE;
2908 }
2909
2910 if (pmap_is_curpmap(pmap)) {
2911 simple_lock(&pmap->pm_obj.vmobjlock);
2912 return (pt_entry_t *)PTE_BASE;
2913 }
2914
2915 p = curproc;
2916 KDASSERT(p != NULL);
2917
2918 /* need to lock both curpmap and pmap: use ordered locking */
2919 if ((vaddr_t) pmap < (vaddr_t) p->p_vmspace->vm_map.pmap) {
2920 simple_lock(&pmap->pm_obj.vmobjlock);
2921 simple_lock(&p->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
2922 } else {
2923 simple_lock(&p->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
2924 simple_lock(&pmap->pm_obj.vmobjlock);
2925 }
2926
2927 pmap_map_in_l1(p->p_vmspace->vm_map.pmap, APTE_BASE, pmap->pm_pptpt,
2928 FALSE);
2929 cpu_tlb_flushD();
2930 cpu_cpwait();
2931 return (pt_entry_t *)APTE_BASE;
2932 }
2933
2934 /*
2935 * pmap_unmap_ptes: unlock the PTE mapping of "pmap"
2936 */
2937
2938 static void
2939 pmap_unmap_ptes(struct pmap *pmap)
2940 {
2941
2942 if (pmap == pmap_kernel()) {
2943 return;
2944 }
2945 if (pmap_is_curpmap(pmap)) {
2946 simple_unlock(&pmap->pm_obj.vmobjlock);
2947 } else {
2948 KDASSERT(curproc != NULL);
2949 simple_unlock(&pmap->pm_obj.vmobjlock);
2950 simple_unlock(
2951 &curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
2952 }
2953 }
2954
2955 /*
2956 * Modify pte bits for all ptes corresponding to the given physical address.
2957 * We use `maskbits' rather than `clearbits' because we're always passing
2958 * constants and the latter would require an extra inversion at run-time.
2959 */
2960
2961 static void
2962 pmap_clearbit(struct vm_page *pg, u_int maskbits)
2963 {
2964 struct pv_entry *pv;
2965 pt_entry_t *ptes;
2966 vaddr_t va;
2967 int tlbentry;
2968
2969 PDEBUG(1, printf("pmap_clearbit: pa=%08lx mask=%08x\n",
2970 VM_PAGE_TO_PHYS(pg), maskbits));
2971
2972 tlbentry = 0;
2973
2974 PMAP_HEAD_TO_MAP_LOCK();
2975 simple_lock(&pg->mdpage.pvh_slock);
2976
2977 /*
2978 * Clear saved attributes (modify, reference)
2979 */
2980 pg->mdpage.pvh_attrs &= ~maskbits;
2981
2982 if (pg->mdpage.pvh_list == NULL) {
2983 simple_unlock(&pg->mdpage.pvh_slock);
2984 PMAP_HEAD_TO_MAP_UNLOCK();
2985 return;
2986 }
2987
2988 /*
2989 * Loop over all current mappings setting/clearing as appropos
2990 */
2991 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
2992 va = pv->pv_va;
2993 pv->pv_flags &= ~maskbits;
2994 ptes = pmap_map_ptes(pv->pv_pmap); /* locks pmap */
2995 KASSERT(pmap_pde_v(pmap_pde(pv->pv_pmap, va)));
2996 if (maskbits & (PVF_WRITE|PVF_MOD)) {
2997 if ((pv->pv_flags & PVF_NC)) {
2998 /*
2999 * Entry is not cacheable: reenable
3000 * the cache, nothing to flush
3001 *
3002 * Don't turn caching on again if this
3003 * is a modified emulation. This
3004 * would be inconsitent with the
3005 * settings created by
3006 * pmap_vac_me_harder().
3007 *
3008 * There's no need to call
3009 * pmap_vac_me_harder() here: all
3010 * pages are loosing their write
3011 * permission.
3012 *
3013 */
3014 if (maskbits & PVF_WRITE) {
3015 ptes[arm_btop(va)] |= pte_cache_mode;
3016 pv->pv_flags &= ~PVF_NC;
3017 }
3018 } else if (pmap_is_curpmap(pv->pv_pmap)) {
3019 /*
3020 * Entry is cacheable: check if pmap is
3021 * current if it is flush it,
3022 * otherwise it won't be in the cache
3023 */
3024 cpu_idcache_wbinv_range(pv->pv_va, NBPG);
3025 }
3026
3027 /* make the pte read only */
3028 ptes[arm_btop(va)] &= ~L2_S_PROT_W;
3029 }
3030
3031 if (maskbits & PVF_REF)
3032 ptes[arm_btop(va)] =
3033 (ptes[arm_btop(va)] & ~L2_TYPE_MASK) | L2_TYPE_INV;
3034
3035 if (pmap_is_curpmap(pv->pv_pmap)) {
3036 /*
3037 * if we had cacheable pte's we'd clean the
3038 * pte out to memory here
3039 *
3040 * flush tlb entry as it's in the current pmap
3041 */
3042 cpu_tlb_flushID_SE(pv->pv_va);
3043 }
3044 pmap_unmap_ptes(pv->pv_pmap); /* unlocks pmap */
3045 }
3046 cpu_cpwait();
3047
3048 simple_unlock(&pg->mdpage.pvh_slock);
3049 PMAP_HEAD_TO_MAP_UNLOCK();
3050 }
3051
3052 /*
3053 * pmap_clear_modify:
3054 *
3055 * Clear the "modified" attribute for a page.
3056 */
3057 boolean_t
3058 pmap_clear_modify(struct vm_page *pg)
3059 {
3060 boolean_t rv;
3061
3062 if (pg->mdpage.pvh_attrs & PVF_MOD) {
3063 rv = TRUE;
3064 pmap_clearbit(pg, PVF_MOD);
3065 } else
3066 rv = FALSE;
3067
3068 PDEBUG(0, printf("pmap_clear_modify pa=%08lx -> %d\n",
3069 VM_PAGE_TO_PHYS(pg), rv));
3070
3071 return (rv);
3072 }
3073
3074 /*
3075 * pmap_clear_reference:
3076 *
3077 * Clear the "referenced" attribute for a page.
3078 */
3079 boolean_t
3080 pmap_clear_reference(struct vm_page *pg)
3081 {
3082 boolean_t rv;
3083
3084 if (pg->mdpage.pvh_attrs & PVF_REF) {
3085 rv = TRUE;
3086 pmap_clearbit(pg, PVF_REF);
3087 } else
3088 rv = FALSE;
3089
3090 PDEBUG(0, printf("pmap_clear_reference pa=%08lx -> %d\n",
3091 VM_PAGE_TO_PHYS(pg), rv));
3092
3093 return (rv);
3094 }
3095
3096 /*
3097 * pmap_is_modified:
3098 *
3099 * Test if a page has the "modified" attribute.
3100 */
3101 /* See <arm/arm32/pmap.h> */
3102
3103 /*
3104 * pmap_is_referenced:
3105 *
3106 * Test if a page has the "referenced" attribute.
3107 */
3108 /* See <arm/arm32/pmap.h> */
3109
3110 int
3111 pmap_modified_emulation(struct pmap *pmap, vaddr_t va)
3112 {
3113 pt_entry_t *ptes;
3114 struct vm_page *pg;
3115 paddr_t pa;
3116 u_int flags;
3117 int rv = 0;
3118
3119 PDEBUG(2, printf("pmap_modified_emulation\n"));
3120
3121 PMAP_MAP_TO_HEAD_LOCK();
3122 ptes = pmap_map_ptes(pmap); /* locks pmap */
3123
3124 if (pmap_pde_v(pmap_pde(pmap, va)) == 0) {
3125 PDEBUG(2, printf("L1 PTE invalid\n"));
3126 goto out;
3127 }
3128
3129 PDEBUG(1, printf("pte=%08x\n", ptes[arm_btop(va)]));
3130
3131 /* Check for a invalid pte */
3132 if (l2pte_valid(ptes[arm_btop(va)]) == 0)
3133 goto out;
3134
3135 /* This can happen if user code tries to access kernel memory. */
3136 if ((ptes[arm_btop(va)] & L2_S_PROT_W) != 0)
3137 goto out;
3138
3139 /* Extract the physical address of the page */
3140 pa = l2pte_pa(ptes[arm_btop(va)]);
3141 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3142 goto out;
3143
3144 /* Get the current flags for this page. */
3145 simple_lock(&pg->mdpage.pvh_slock);
3146
3147 flags = pmap_modify_pv(pmap, va, pg, 0, 0);
3148 PDEBUG(2, printf("pmap_modified_emulation: flags = %08x\n", flags));
3149
3150 /*
3151 * Do the flags say this page is writable ? If not then it is a
3152 * genuine write fault. If yes then the write fault is our fault
3153 * as we did not reflect the write access in the PTE. Now we know
3154 * a write has occurred we can correct this and also set the
3155 * modified bit
3156 */
3157 if (~flags & PVF_WRITE) {
3158 simple_unlock(&pg->mdpage.pvh_slock);
3159 goto out;
3160 }
3161
3162 PDEBUG(0,
3163 printf("pmap_modified_emulation: Got a hit va=%08lx, pte = %08x\n",
3164 va, ptes[arm_btop(va)]));
3165 pg->mdpage.pvh_attrs |= PVF_REF | PVF_MOD;
3166
3167 /*
3168 * Re-enable write permissions for the page. No need to call
3169 * pmap_vac_me_harder(), since this is just a
3170 * modified-emulation fault, and the PVF_WRITE bit isn't changing.
3171 * We've already set the cacheable bits based on the assumption
3172 * that we can write to this page.
3173 */
3174 ptes[arm_btop(va)] =
3175 (ptes[arm_btop(va)] & ~L2_TYPE_MASK) | L2_S_PROTO | L2_S_PROT_W;
3176 PDEBUG(0, printf("->(%08x)\n", ptes[arm_btop(va)]));
3177
3178 simple_unlock(&pg->mdpage.pvh_slock);
3179
3180 cpu_tlb_flushID_SE(va);
3181 cpu_cpwait();
3182 rv = 1;
3183 out:
3184 pmap_unmap_ptes(pmap); /* unlocks pmap */
3185 PMAP_MAP_TO_HEAD_UNLOCK();
3186 return (rv);
3187 }
3188
3189 int
3190 pmap_handled_emulation(struct pmap *pmap, vaddr_t va)
3191 {
3192 pt_entry_t *ptes;
3193 struct vm_page *pg;
3194 paddr_t pa;
3195 int rv = 0;
3196
3197 PDEBUG(2, printf("pmap_handled_emulation\n"));
3198
3199 PMAP_MAP_TO_HEAD_LOCK();
3200 ptes = pmap_map_ptes(pmap); /* locks pmap */
3201
3202 if (pmap_pde_v(pmap_pde(pmap, va)) == 0) {
3203 PDEBUG(2, printf("L1 PTE invalid\n"));
3204 goto out;
3205 }
3206
3207 PDEBUG(1, printf("pte=%08x\n", ptes[arm_btop(va)]));
3208
3209 /* Check for invalid pte */
3210 if (l2pte_valid(ptes[arm_btop(va)]) == 0)
3211 goto out;
3212
3213 /* This can happen if user code tries to access kernel memory. */
3214 if ((ptes[arm_btop(va)] & L2_TYPE_MASK) != L2_TYPE_INV)
3215 goto out;
3216
3217 /* Extract the physical address of the page */
3218 pa = l2pte_pa(ptes[arm_btop(va)]);
3219 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3220 goto out;
3221
3222 simple_lock(&pg->mdpage.pvh_slock);
3223
3224 /*
3225 * Ok we just enable the pte and mark the attibs as handled
3226 * XXX Should we traverse the PV list and enable all PTEs?
3227 */
3228 PDEBUG(0,
3229 printf("pmap_handled_emulation: Got a hit va=%08lx pte = %08x\n",
3230 va, ptes[arm_btop(va)]));
3231 pg->mdpage.pvh_attrs |= PVF_REF;
3232
3233 ptes[arm_btop(va)] = (ptes[arm_btop(va)] & ~L2_TYPE_MASK) | L2_S_PROTO;
3234 PDEBUG(0, printf("->(%08x)\n", ptes[arm_btop(va)]));
3235
3236 simple_unlock(&pg->mdpage.pvh_slock);
3237
3238 cpu_tlb_flushID_SE(va);
3239 cpu_cpwait();
3240 rv = 1;
3241 out:
3242 pmap_unmap_ptes(pmap); /* unlocks pmap */
3243 PMAP_MAP_TO_HEAD_UNLOCK();
3244 return (rv);
3245 }
3246
3247 /*
3248 * pmap_collect: free resources held by a pmap
3249 *
3250 * => optional function.
3251 * => called when a process is swapped out to free memory.
3252 */
3253
3254 void
3255 pmap_collect(struct pmap *pmap)
3256 {
3257 }
3258
3259 /*
3260 * Routine: pmap_procwr
3261 *
3262 * Function:
3263 * Synchronize caches corresponding to [addr, addr+len) in p.
3264 *
3265 */
3266 void
3267 pmap_procwr(struct proc *p, vaddr_t va, int len)
3268 {
3269 /* We only need to do anything if it is the current process. */
3270 if (p == curproc)
3271 cpu_icache_sync_range(va, len);
3272 }
3273 /*
3274 * PTP functions
3275 */
3276
3277 /*
3278 * pmap_get_ptp: get a PTP (if there isn't one, allocate a new one)
3279 *
3280 * => pmap should NOT be pmap_kernel()
3281 * => pmap should be locked
3282 */
3283
3284 static struct vm_page *
3285 pmap_get_ptp(struct pmap *pmap, vaddr_t va)
3286 {
3287 struct vm_page *ptp;
3288
3289 if (pmap_pde_page(pmap_pde(pmap, va))) {
3290
3291 /* valid... check hint (saves us a PA->PG lookup) */
3292 if (pmap->pm_ptphint &&
3293 (pmap->pm_pdir[pmap_pdei(va)] & L2_S_FRAME) ==
3294 VM_PAGE_TO_PHYS(pmap->pm_ptphint))
3295 return (pmap->pm_ptphint);
3296 ptp = uvm_pagelookup(&pmap->pm_obj, va);
3297 #ifdef DIAGNOSTIC
3298 if (ptp == NULL)
3299 panic("pmap_get_ptp: unmanaged user PTP");
3300 #endif
3301 pmap->pm_ptphint = ptp;
3302 return(ptp);
3303 }
3304
3305 /* allocate a new PTP (updates ptphint) */
3306 return(pmap_alloc_ptp(pmap, va));
3307 }
3308
3309 /*
3310 * pmap_alloc_ptp: allocate a PTP for a PMAP
3311 *
3312 * => pmap should already be locked by caller
3313 * => we use the ptp's wire_count to count the number of active mappings
3314 * in the PTP (we start it at one to prevent any chance this PTP
3315 * will ever leak onto the active/inactive queues)
3316 */
3317
3318 /*__inline */ static struct vm_page *
3319 pmap_alloc_ptp(struct pmap *pmap, vaddr_t va)
3320 {
3321 struct vm_page *ptp;
3322
3323 ptp = uvm_pagealloc(&pmap->pm_obj, va, NULL,
3324 UVM_PGA_USERESERVE|UVM_PGA_ZERO);
3325 if (ptp == NULL)
3326 return (NULL);
3327
3328 /* got one! */
3329 ptp->flags &= ~PG_BUSY; /* never busy */
3330 ptp->wire_count = 1; /* no mappings yet */
3331 pmap_map_in_l1(pmap, va, VM_PAGE_TO_PHYS(ptp), TRUE);
3332 pmap->pm_stats.resident_count++; /* count PTP as resident */
3333 pmap->pm_ptphint = ptp;
3334 return (ptp);
3335 }
3336
3337 vaddr_t
3338 pmap_growkernel(vaddr_t maxkvaddr)
3339 {
3340 struct pmap *kpm = pmap_kernel(), *pm;
3341 int s;
3342 paddr_t ptaddr;
3343 struct vm_page *ptp;
3344
3345 if (maxkvaddr <= pmap_curmaxkvaddr)
3346 goto out; /* we are OK */
3347 NPDEBUG(PDB_GROWKERN, printf("pmap_growkernel: growing kernel from %lx to %lx\n",
3348 pmap_curmaxkvaddr, maxkvaddr));
3349
3350 /*
3351 * whoops! we need to add kernel PTPs
3352 */
3353
3354 s = splhigh(); /* to be safe */
3355 simple_lock(&kpm->pm_obj.vmobjlock);
3356 /* due to the way the arm pmap works we map 4MB at a time */
3357 for (/*null*/ ; pmap_curmaxkvaddr < maxkvaddr;
3358 pmap_curmaxkvaddr += 4 * L1_S_SIZE) {
3359
3360 if (uvm.page_init_done == FALSE) {
3361
3362 /*
3363 * we're growing the kernel pmap early (from
3364 * uvm_pageboot_alloc()). this case must be
3365 * handled a little differently.
3366 */
3367
3368 if (uvm_page_physget(&ptaddr) == FALSE)
3369 panic("pmap_growkernel: out of memory");
3370 pmap_zero_page(ptaddr);
3371
3372 /* map this page in */
3373 pmap_map_in_l1(kpm, pmap_curmaxkvaddr, ptaddr, TRUE);
3374
3375 /* count PTP as resident */
3376 kpm->pm_stats.resident_count++;
3377 continue;
3378 }
3379
3380 /*
3381 * THIS *MUST* BE CODED SO AS TO WORK IN THE
3382 * pmap_initialized == FALSE CASE! WE MAY BE
3383 * INVOKED WHILE pmap_init() IS RUNNING!
3384 */
3385
3386 if ((ptp = pmap_alloc_ptp(kpm, pmap_curmaxkvaddr)) == NULL)
3387 panic("pmap_growkernel: alloc ptp failed");
3388
3389 /* distribute new kernel PTP to all active pmaps */
3390 simple_lock(&pmaps_lock);
3391 LIST_FOREACH(pm, &pmaps, pm_list) {
3392 pmap_map_in_l1(pm, pmap_curmaxkvaddr,
3393 VM_PAGE_TO_PHYS(ptp), TRUE);
3394 }
3395
3396 simple_unlock(&pmaps_lock);
3397 }
3398
3399 /*
3400 * flush out the cache, expensive but growkernel will happen so
3401 * rarely
3402 */
3403 cpu_tlb_flushD();
3404 cpu_cpwait();
3405
3406 simple_unlock(&kpm->pm_obj.vmobjlock);
3407 splx(s);
3408
3409 out:
3410 return (pmap_curmaxkvaddr);
3411 }
3412
3413 /************************ Utility routines ****************************/
3414
3415 /*
3416 * vector_page_setprot:
3417 *
3418 * Manipulate the protection of the vector page.
3419 */
3420 void
3421 vector_page_setprot(int prot)
3422 {
3423 pt_entry_t *pte;
3424
3425 pte = vtopte(vector_page);
3426
3427 *pte = (*pte & ~L1_S_PROT_MASK) | L2_S_PROT(PTE_KERNEL, prot);
3428 cpu_tlb_flushD_SE(vector_page);
3429 cpu_cpwait();
3430 }
3431
3432 /************************ Bootstrapping routines ****************************/
3433
3434 /*
3435 * This list exists for the benefit of pmap_map_chunk(). It keeps track
3436 * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
3437 * find them as necessary.
3438 *
3439 * Note that the data on this list is not valid after initarm() returns.
3440 */
3441 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
3442
3443 static vaddr_t
3444 kernel_pt_lookup(paddr_t pa)
3445 {
3446 pv_addr_t *pv;
3447
3448 SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
3449 if (pv->pv_pa == pa)
3450 return (pv->pv_va);
3451 }
3452 return (0);
3453 }
3454
3455 /*
3456 * pmap_map_section:
3457 *
3458 * Create a single section mapping.
3459 */
3460 void
3461 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
3462 {
3463 pd_entry_t *pde = (pd_entry_t *) l1pt;
3464 pd_entry_t fl = (cache == PTE_CACHE) ? pte_cache_mode : 0;
3465
3466 KASSERT(((va | pa) & L1_S_OFFSET) == 0);
3467
3468 pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
3469 L1_S_PROT(PTE_KERNEL, prot) | fl;
3470 }
3471
3472 /*
3473 * pmap_map_entry:
3474 *
3475 * Create a single page mapping.
3476 */
3477 void
3478 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
3479 {
3480 pd_entry_t *pde = (pd_entry_t *) l1pt;
3481 pt_entry_t fl = (cache == PTE_CACHE) ? pte_cache_mode : 0;
3482 pt_entry_t *pte;
3483
3484 KASSERT(((va | pa) & PGOFSET) == 0);
3485
3486 if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
3487 panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
3488
3489 pte = (pt_entry_t *)
3490 kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
3491 if (pte == NULL)
3492 panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
3493
3494 pte[(va >> PGSHIFT) & 0x3ff] = L2_S_PROTO | pa |
3495 L2_S_PROT(PTE_KERNEL, prot) | fl;
3496 }
3497
3498 /*
3499 * pmap_link_l2pt:
3500 *
3501 * Link the L2 page table specified by "pa" into the L1
3502 * page table at the slot for "va".
3503 */
3504 void
3505 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
3506 {
3507 pd_entry_t *pde = (pd_entry_t *) l1pt;
3508 u_int slot = va >> L1_S_SHIFT;
3509
3510 KASSERT((l2pv->pv_pa & PGOFSET) == 0);
3511
3512 pde[slot + 0] = L1_C_PROTO | (l2pv->pv_pa + 0x000);
3513 pde[slot + 1] = L1_C_PROTO | (l2pv->pv_pa + 0x400);
3514 pde[slot + 2] = L1_C_PROTO | (l2pv->pv_pa + 0x800);
3515 pde[slot + 3] = L1_C_PROTO | (l2pv->pv_pa + 0xc00);
3516
3517 SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
3518 }
3519
3520 /*
3521 * pmap_map_chunk:
3522 *
3523 * Map a chunk of memory using the most efficient mappings
3524 * possible (section, large page, small page) into the
3525 * provided L1 and L2 tables at the specified virtual address.
3526 */
3527 vsize_t
3528 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
3529 int prot, int cache)
3530 {
3531 pd_entry_t *pde = (pd_entry_t *) l1pt;
3532 pt_entry_t fl = (cache == PTE_CACHE) ? pte_cache_mode : 0;
3533 pt_entry_t *pte;
3534 vsize_t resid;
3535 int i;
3536
3537 resid = (size + (NBPG - 1)) & ~(NBPG - 1);
3538
3539 if (l1pt == 0)
3540 panic("pmap_map_chunk: no L1 table provided");
3541
3542 #ifdef VERBOSE_INIT_ARM
3543 printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
3544 "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
3545 #endif
3546
3547 size = resid;
3548
3549 while (resid > 0) {
3550 /* See if we can use a section mapping. */
3551 if (((pa | va) & L1_S_OFFSET) == 0 &&
3552 resid >= L1_S_SIZE) {
3553 #ifdef VERBOSE_INIT_ARM
3554 printf("S");
3555 #endif
3556 pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
3557 L1_S_PROT(PTE_KERNEL, prot) | fl;
3558 va += L1_S_SIZE;
3559 pa += L1_S_SIZE;
3560 resid -= L1_S_SIZE;
3561 continue;
3562 }
3563
3564 /*
3565 * Ok, we're going to use an L2 table. Make sure
3566 * one is actually in the corresponding L1 slot
3567 * for the current VA.
3568 */
3569 if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
3570 panic("pmap_map_chunk: no L2 table for VA 0x%08lx", va);
3571
3572 pte = (pt_entry_t *)
3573 kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
3574 if (pte == NULL)
3575 panic("pmap_map_chunk: can't find L2 table for VA"
3576 "0x%08lx", va);
3577
3578 /* See if we can use a L2 large page mapping. */
3579 if (((pa | va) & L2_L_OFFSET) == 0 &&
3580 resid >= L2_L_SIZE) {
3581 #ifdef VERBOSE_INIT_ARM
3582 printf("L");
3583 #endif
3584 for (i = 0; i < 16; i++) {
3585 pte[((va >> PGSHIFT) & 0x3f0) + i] =
3586 L2_L_PROTO | pa |
3587 L2_L_PROT(PTE_KERNEL, prot) | fl;
3588 }
3589 va += L2_L_SIZE;
3590 pa += L2_L_SIZE;
3591 resid -= L2_L_SIZE;
3592 continue;
3593 }
3594
3595 /* Use a small page mapping. */
3596 #ifdef VERBOSE_INIT_ARM
3597 printf("P");
3598 #endif
3599 pte[(va >> PGSHIFT) & 0x3ff] = L2_S_PROTO | pa |
3600 L2_S_PROT(PTE_KERNEL, prot) | fl;
3601 va += NBPG;
3602 pa += NBPG;
3603 resid -= NBPG;
3604 }
3605 #ifdef VERBOSE_INIT_ARM
3606 printf("\n");
3607 #endif
3608 return (size);
3609 }
3610
3611 /********************** PTE initialization routines **************************/
3612
3613 /*
3614 * These routines are called when the CPU type is identified to set up
3615 * the PTE prototypes, cache modes, etc.
3616 *
3617 * The variables are always here, just in case LKMs need to reference
3618 * them (though, they shouldn't).
3619 */
3620
3621 pt_entry_t pte_cache_mode;
3622 pt_entry_t pte_cache_mask;
3623
3624 pt_entry_t pte_l2_s_prot_u;
3625 pt_entry_t pte_l2_s_prot_w;
3626 pt_entry_t pte_l2_s_prot_mask;
3627
3628 pt_entry_t pte_l1_s_proto;
3629 pt_entry_t pte_l1_c_proto;
3630 pt_entry_t pte_l2_s_proto;
3631
3632 #if ARM_MMU_GENERIC == 1
3633 void
3634 pmap_pte_init_generic(void)
3635 {
3636
3637 pte_cache_mode = L2_B|L2_C;
3638 pte_cache_mask = L2_CACHE_MASK_generic;
3639
3640 pte_l2_s_prot_u = L2_S_PROT_U_generic;
3641 pte_l2_s_prot_w = L2_S_PROT_W_generic;
3642 pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
3643
3644 pte_l1_s_proto = L1_S_PROTO_generic;
3645 pte_l1_c_proto = L1_C_PROTO_generic;
3646 pte_l2_s_proto = L2_S_PROTO_generic;
3647 }
3648
3649 #if defined(CPU_ARM9)
3650 void
3651 pmap_pte_init_arm9(void)
3652 {
3653
3654 /*
3655 * ARM9 is compatible with generic, but we want to use
3656 * write-through caching for now.
3657 */
3658 pmap_pte_init_generic();
3659 pte_cache_mode = L2_C;
3660 }
3661 #endif /* CPU_ARM9 */
3662 #endif /* ARM_MMU_GENERIC == 1 */
3663
3664 #if ARM_MMU_XSCALE == 1
3665 void
3666 pmap_pte_init_xscale(void)
3667 {
3668
3669 pte_cache_mode = L2_B|L2_C;
3670 pte_cache_mask = L2_CACHE_MASK_xscale;
3671
3672 pte_l2_s_prot_u = L2_S_PROT_U_xscale;
3673 pte_l2_s_prot_w = L2_S_PROT_W_xscale;
3674 pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale;
3675
3676 pte_l1_s_proto = L1_S_PROTO_xscale;
3677 pte_l1_c_proto = L1_C_PROTO_xscale;
3678 pte_l2_s_proto = L2_S_PROTO_xscale;
3679 }
3680
3681 #if defined(CPU_XSCALE_80200)
3682 void
3683 pmap_pte_init_i80200(void)
3684 {
3685
3686 /*
3687 * Use write-through caching on the i80200.
3688 */
3689 pmap_pte_init_xscale();
3690 pte_cache_mode = L2_C;
3691 }
3692 #endif /* CPU_XSCALE_80200 */
3693 #endif /* ARM_MMU_XSCALE == 1 */
3694