pmap.c revision 1.70 1 /* $NetBSD: pmap.c,v 1.70 2002/03/25 04:51:20 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 1.70 2002/03/25 04:51:20 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 vsize_t npages;
289
290 static struct vm_page *pmap_alloc_ptp __P((struct pmap *, vaddr_t));
291 static struct vm_page *pmap_get_ptp __P((struct pmap *, vaddr_t));
292 __inline static void pmap_clearbit __P((struct vm_page *, unsigned int));
293
294 extern paddr_t physical_start;
295 extern paddr_t physical_freestart;
296 extern paddr_t physical_end;
297 extern paddr_t physical_freeend;
298 extern unsigned int free_pages;
299 extern int max_processes;
300
301 vaddr_t virtual_avail;
302 vaddr_t virtual_end;
303 vaddr_t pmap_curmaxkvaddr;
304
305 vaddr_t avail_start;
306 vaddr_t avail_end;
307
308 extern pv_addr_t systempage;
309
310 /* Variables used by the L1 page table queue code */
311 SIMPLEQ_HEAD(l1pt_queue, l1pt);
312 struct l1pt_queue l1pt_static_queue; /* head of our static l1 queue */
313 int l1pt_static_queue_count; /* items in the static l1 queue */
314 int l1pt_static_create_count; /* static l1 items created */
315 struct l1pt_queue l1pt_queue; /* head of our l1 queue */
316 int l1pt_queue_count; /* items in the l1 queue */
317 int l1pt_create_count; /* stat - L1's create count */
318 int l1pt_reuse_count; /* stat - L1's reused count */
319
320 /* Local function prototypes (not used outside this file) */
321 void pmap_copy_on_write __P((struct vm_page *));
322 void pmap_pinit __P((struct pmap *));
323 void pmap_freepagedir __P((struct pmap *));
324
325 /* Other function prototypes */
326 extern void bzero_page __P((vaddr_t));
327 extern void bcopy_page __P((vaddr_t, vaddr_t));
328
329 struct l1pt *pmap_alloc_l1pt __P((void));
330 static __inline void pmap_map_in_l1 __P((struct pmap *pmap, vaddr_t va,
331 vaddr_t l2pa, boolean_t));
332
333 static pt_entry_t *pmap_map_ptes __P((struct pmap *));
334 static void pmap_unmap_ptes __P((struct pmap *));
335
336 __inline static void pmap_vac_me_harder __P((struct pmap *, struct vm_page *,
337 pt_entry_t *, boolean_t));
338 static void pmap_vac_me_kpmap __P((struct pmap *, struct vm_page *,
339 pt_entry_t *, boolean_t));
340 static void pmap_vac_me_user __P((struct pmap *, struct vm_page *,
341 pt_entry_t *, boolean_t));
342
343 /*
344 * Cache enable bits in PTE to use on pages that are cacheable.
345 * On most machines this is cacheable/bufferable, but on some, eg arm10, we
346 * can chose between write-through and write-back cacheing.
347 */
348 pt_entry_t pte_cache_mode = (PT_C | PT_B);
349
350 /*
351 * real definition of pv_entry.
352 */
353
354 struct pv_entry {
355 struct pv_entry *pv_next; /* next pv_entry */
356 struct pmap *pv_pmap; /* pmap where mapping lies */
357 vaddr_t pv_va; /* virtual address for mapping */
358 int pv_flags; /* flags */
359 struct vm_page *pv_ptp; /* vm_page for the ptp */
360 };
361
362 /*
363 * pv_entrys are dynamically allocated in chunks from a single page.
364 * we keep track of how many pv_entrys are in use for each page and
365 * we can free pv_entry pages if needed. there is one lock for the
366 * entire allocation system.
367 */
368
369 struct pv_page_info {
370 TAILQ_ENTRY(pv_page) pvpi_list;
371 struct pv_entry *pvpi_pvfree;
372 int pvpi_nfree;
373 };
374
375 /*
376 * number of pv_entry's in a pv_page
377 * (note: won't work on systems where NPBG isn't a constant)
378 */
379
380 #define PVE_PER_PVPAGE ((NBPG - sizeof(struct pv_page_info)) / \
381 sizeof(struct pv_entry))
382
383 /*
384 * a pv_page: where pv_entrys are allocated from
385 */
386
387 struct pv_page {
388 struct pv_page_info pvinfo;
389 struct pv_entry pvents[PVE_PER_PVPAGE];
390 };
391
392 #ifdef MYCROFT_HACK
393 int mycroft_hack = 0;
394 #endif
395
396 /* Function to set the debug level of the pmap code */
397
398 #ifdef PMAP_DEBUG
399 void
400 pmap_debug(level)
401 int level;
402 {
403 pmap_debug_level = level;
404 printf("pmap_debug: level=%d\n", pmap_debug_level);
405 }
406 #endif /* PMAP_DEBUG */
407
408 __inline static boolean_t
409 pmap_is_curpmap(struct pmap *pmap)
410 {
411
412 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap) ||
413 pmap == pmap_kernel())
414 return (TRUE);
415
416 return (FALSE);
417 }
418
419 #include "isadma.h"
420
421 #if NISADMA > 0
422 /*
423 * Used to protect memory for ISA DMA bounce buffers. If, when loading
424 * pages into the system, memory intersects with any of these ranges,
425 * the intersecting memory will be loaded into a lower-priority free list.
426 */
427 bus_dma_segment_t *pmap_isa_dma_ranges;
428 int pmap_isa_dma_nranges;
429
430 boolean_t pmap_isa_dma_range_intersect __P((paddr_t, psize_t,
431 paddr_t *, psize_t *));
432
433 /*
434 * Check if a memory range intersects with an ISA DMA range, and
435 * return the page-rounded intersection if it does. The intersection
436 * will be placed on a lower-priority free list.
437 */
438 boolean_t
439 pmap_isa_dma_range_intersect(pa, size, pap, sizep)
440 paddr_t pa;
441 psize_t size;
442 paddr_t *pap;
443 psize_t *sizep;
444 {
445 bus_dma_segment_t *ds;
446 int i;
447
448 if (pmap_isa_dma_ranges == NULL)
449 return (FALSE);
450
451 for (i = 0, ds = pmap_isa_dma_ranges;
452 i < pmap_isa_dma_nranges; i++, ds++) {
453 if (ds->ds_addr <= pa && pa < (ds->ds_addr + ds->ds_len)) {
454 /*
455 * Beginning of region intersects with this range.
456 */
457 *pap = trunc_page(pa);
458 *sizep = round_page(min(pa + size,
459 ds->ds_addr + ds->ds_len) - pa);
460 return (TRUE);
461 }
462 if (pa < ds->ds_addr && ds->ds_addr < (pa + size)) {
463 /*
464 * End of region intersects with this range.
465 */
466 *pap = trunc_page(ds->ds_addr);
467 *sizep = round_page(min((pa + size) - ds->ds_addr,
468 ds->ds_len));
469 return (TRUE);
470 }
471 }
472
473 /*
474 * No intersection found.
475 */
476 return (FALSE);
477 }
478 #endif /* NISADMA > 0 */
479
480 /*
481 * p v _ e n t r y f u n c t i o n s
482 */
483
484 /*
485 * pv_entry allocation functions:
486 * the main pv_entry allocation functions are:
487 * pmap_alloc_pv: allocate a pv_entry structure
488 * pmap_free_pv: free one pv_entry
489 * pmap_free_pvs: free a list of pv_entrys
490 *
491 * the rest are helper functions
492 */
493
494 /*
495 * pmap_alloc_pv: inline function to allocate a pv_entry structure
496 * => we lock pvalloc_lock
497 * => if we fail, we call out to pmap_alloc_pvpage
498 * => 3 modes:
499 * ALLOCPV_NEED = we really need a pv_entry, even if we have to steal it
500 * ALLOCPV_TRY = we want a pv_entry, but not enough to steal
501 * ALLOCPV_NONEED = we are trying to grow our free list, don't really need
502 * one now
503 *
504 * "try" is for optional functions like pmap_copy().
505 */
506
507 __inline static struct pv_entry *
508 pmap_alloc_pv(pmap, mode)
509 struct pmap *pmap;
510 int mode;
511 {
512 struct pv_page *pvpage;
513 struct pv_entry *pv;
514
515 simple_lock(&pvalloc_lock);
516
517 pvpage = TAILQ_FIRST(&pv_freepages);
518
519 if (pvpage != NULL) {
520 pvpage->pvinfo.pvpi_nfree--;
521 if (pvpage->pvinfo.pvpi_nfree == 0) {
522 /* nothing left in this one? */
523 TAILQ_REMOVE(&pv_freepages, pvpage, pvinfo.pvpi_list);
524 }
525 pv = pvpage->pvinfo.pvpi_pvfree;
526 KASSERT(pv);
527 pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
528 pv_nfpvents--; /* took one from pool */
529 } else {
530 pv = NULL; /* need more of them */
531 }
532
533 /*
534 * if below low water mark or we didn't get a pv_entry we try and
535 * create more pv_entrys ...
536 */
537
538 if (pv_nfpvents < PVE_LOWAT || pv == NULL) {
539 if (pv == NULL)
540 pv = pmap_alloc_pvpage(pmap, (mode == ALLOCPV_TRY) ?
541 mode : ALLOCPV_NEED);
542 else
543 (void) pmap_alloc_pvpage(pmap, ALLOCPV_NONEED);
544 }
545
546 simple_unlock(&pvalloc_lock);
547 return(pv);
548 }
549
550 /*
551 * pmap_alloc_pvpage: maybe allocate a new pvpage
552 *
553 * if need_entry is false: try and allocate a new pv_page
554 * if need_entry is true: try and allocate a new pv_page and return a
555 * new pv_entry from it. if we are unable to allocate a pv_page
556 * we make a last ditch effort to steal a pv_page from some other
557 * mapping. if that fails, we panic...
558 *
559 * => we assume that the caller holds pvalloc_lock
560 */
561
562 static struct pv_entry *
563 pmap_alloc_pvpage(pmap, mode)
564 struct pmap *pmap;
565 int mode;
566 {
567 struct vm_page *pg;
568 struct pv_page *pvpage;
569 struct pv_entry *pv;
570 int s;
571
572 /*
573 * if we need_entry and we've got unused pv_pages, allocate from there
574 */
575
576 pvpage = TAILQ_FIRST(&pv_unusedpgs);
577 if (mode != ALLOCPV_NONEED && pvpage != NULL) {
578
579 /* move it to pv_freepages list */
580 TAILQ_REMOVE(&pv_unusedpgs, pvpage, pvinfo.pvpi_list);
581 TAILQ_INSERT_HEAD(&pv_freepages, pvpage, pvinfo.pvpi_list);
582
583 /* allocate a pv_entry */
584 pvpage->pvinfo.pvpi_nfree--; /* can't go to zero */
585 pv = pvpage->pvinfo.pvpi_pvfree;
586 KASSERT(pv);
587 pvpage->pvinfo.pvpi_pvfree = pv->pv_next;
588
589 pv_nfpvents--; /* took one from pool */
590 return(pv);
591 }
592
593 /*
594 * see if we've got a cached unmapped VA that we can map a page in.
595 * if not, try to allocate one.
596 */
597
598
599 if (pv_cachedva == 0) {
600 s = splvm();
601 pv_cachedva = uvm_km_kmemalloc(kmem_map, NULL,
602 PAGE_SIZE, UVM_KMF_TRYLOCK|UVM_KMF_VALLOC);
603 splx(s);
604 if (pv_cachedva == 0) {
605 return (NULL);
606 }
607 }
608
609 pg = uvm_pagealloc(NULL, pv_cachedva - vm_map_min(kernel_map), NULL,
610 UVM_PGA_USERESERVE);
611
612 if (pg == NULL)
613 return (NULL);
614 pg->flags &= ~PG_BUSY; /* never busy */
615
616 /*
617 * add a mapping for our new pv_page and free its entrys (save one!)
618 *
619 * NOTE: If we are allocating a PV page for the kernel pmap, the
620 * pmap is already locked! (...but entering the mapping is safe...)
621 */
622
623 pmap_kenter_pa(pv_cachedva, VM_PAGE_TO_PHYS(pg),
624 VM_PROT_READ|VM_PROT_WRITE);
625 pmap_update(pmap_kernel());
626 pvpage = (struct pv_page *) pv_cachedva;
627 pv_cachedva = 0;
628 return (pmap_add_pvpage(pvpage, mode != ALLOCPV_NONEED));
629 }
630
631 /*
632 * pmap_add_pvpage: add a pv_page's pv_entrys to the free list
633 *
634 * => caller must hold pvalloc_lock
635 * => if need_entry is true, we allocate and return one pv_entry
636 */
637
638 static struct pv_entry *
639 pmap_add_pvpage(pvp, need_entry)
640 struct pv_page *pvp;
641 boolean_t need_entry;
642 {
643 int tofree, lcv;
644
645 /* do we need to return one? */
646 tofree = (need_entry) ? PVE_PER_PVPAGE - 1 : PVE_PER_PVPAGE;
647
648 pvp->pvinfo.pvpi_pvfree = NULL;
649 pvp->pvinfo.pvpi_nfree = tofree;
650 for (lcv = 0 ; lcv < tofree ; lcv++) {
651 pvp->pvents[lcv].pv_next = pvp->pvinfo.pvpi_pvfree;
652 pvp->pvinfo.pvpi_pvfree = &pvp->pvents[lcv];
653 }
654 if (need_entry)
655 TAILQ_INSERT_TAIL(&pv_freepages, pvp, pvinfo.pvpi_list);
656 else
657 TAILQ_INSERT_TAIL(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
658 pv_nfpvents += tofree;
659 return((need_entry) ? &pvp->pvents[lcv] : NULL);
660 }
661
662 /*
663 * pmap_free_pv_doit: actually free a pv_entry
664 *
665 * => do not call this directly! instead use either
666 * 1. pmap_free_pv ==> free a single pv_entry
667 * 2. pmap_free_pvs => free a list of pv_entrys
668 * => we must be holding pvalloc_lock
669 */
670
671 __inline static void
672 pmap_free_pv_doit(pv)
673 struct pv_entry *pv;
674 {
675 struct pv_page *pvp;
676
677 pvp = (struct pv_page *) arm_trunc_page((vaddr_t)pv);
678 pv_nfpvents++;
679 pvp->pvinfo.pvpi_nfree++;
680
681 /* nfree == 1 => fully allocated page just became partly allocated */
682 if (pvp->pvinfo.pvpi_nfree == 1) {
683 TAILQ_INSERT_HEAD(&pv_freepages, pvp, pvinfo.pvpi_list);
684 }
685
686 /* free it */
687 pv->pv_next = pvp->pvinfo.pvpi_pvfree;
688 pvp->pvinfo.pvpi_pvfree = pv;
689
690 /*
691 * are all pv_page's pv_entry's free? move it to unused queue.
692 */
693
694 if (pvp->pvinfo.pvpi_nfree == PVE_PER_PVPAGE) {
695 TAILQ_REMOVE(&pv_freepages, pvp, pvinfo.pvpi_list);
696 TAILQ_INSERT_HEAD(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
697 }
698 }
699
700 /*
701 * pmap_free_pv: free a single pv_entry
702 *
703 * => we gain the pvalloc_lock
704 */
705
706 __inline static void
707 pmap_free_pv(pmap, pv)
708 struct pmap *pmap;
709 struct pv_entry *pv;
710 {
711 simple_lock(&pvalloc_lock);
712 pmap_free_pv_doit(pv);
713
714 /*
715 * Can't free the PV page if the PV entries were associated with
716 * the kernel pmap; the pmap is already locked.
717 */
718 if (pv_nfpvents > PVE_HIWAT && TAILQ_FIRST(&pv_unusedpgs) != NULL &&
719 pmap != pmap_kernel())
720 pmap_free_pvpage();
721
722 simple_unlock(&pvalloc_lock);
723 }
724
725 /*
726 * pmap_free_pvs: free a list of pv_entrys
727 *
728 * => we gain the pvalloc_lock
729 */
730
731 __inline static void
732 pmap_free_pvs(pmap, pvs)
733 struct pmap *pmap;
734 struct pv_entry *pvs;
735 {
736 struct pv_entry *nextpv;
737
738 simple_lock(&pvalloc_lock);
739
740 for ( /* null */ ; pvs != NULL ; pvs = nextpv) {
741 nextpv = pvs->pv_next;
742 pmap_free_pv_doit(pvs);
743 }
744
745 /*
746 * Can't free the PV page if the PV entries were associated with
747 * the kernel pmap; the pmap is already locked.
748 */
749 if (pv_nfpvents > PVE_HIWAT && TAILQ_FIRST(&pv_unusedpgs) != NULL &&
750 pmap != pmap_kernel())
751 pmap_free_pvpage();
752
753 simple_unlock(&pvalloc_lock);
754 }
755
756
757 /*
758 * pmap_free_pvpage: try and free an unused pv_page structure
759 *
760 * => assume caller is holding the pvalloc_lock and that
761 * there is a page on the pv_unusedpgs list
762 * => if we can't get a lock on the kmem_map we try again later
763 */
764
765 static void
766 pmap_free_pvpage()
767 {
768 int s;
769 struct vm_map *map;
770 struct vm_map_entry *dead_entries;
771 struct pv_page *pvp;
772
773 s = splvm(); /* protect kmem_map */
774
775 pvp = TAILQ_FIRST(&pv_unusedpgs);
776
777 /*
778 * note: watch out for pv_initpage which is allocated out of
779 * kernel_map rather than kmem_map.
780 */
781 if (pvp == pv_initpage)
782 map = kernel_map;
783 else
784 map = kmem_map;
785 if (vm_map_lock_try(map)) {
786
787 /* remove pvp from pv_unusedpgs */
788 TAILQ_REMOVE(&pv_unusedpgs, pvp, pvinfo.pvpi_list);
789
790 /* unmap the page */
791 dead_entries = NULL;
792 uvm_unmap_remove(map, (vaddr_t)pvp, ((vaddr_t)pvp) + PAGE_SIZE,
793 &dead_entries);
794 vm_map_unlock(map);
795
796 if (dead_entries != NULL)
797 uvm_unmap_detach(dead_entries, 0);
798
799 pv_nfpvents -= PVE_PER_PVPAGE; /* update free count */
800 }
801 if (pvp == pv_initpage)
802 /* no more initpage, we've freed it */
803 pv_initpage = NULL;
804
805 splx(s);
806 }
807
808 /*
809 * main pv_entry manipulation functions:
810 * pmap_enter_pv: enter a mapping onto a vm_page list
811 * pmap_remove_pv: remove a mappiing from a vm_page list
812 *
813 * NOTE: pmap_enter_pv expects to lock the pvh itself
814 * pmap_remove_pv expects te caller to lock the pvh before calling
815 */
816
817 /*
818 * pmap_enter_pv: enter a mapping onto a vm_page lst
819 *
820 * => caller should hold the proper lock on pmap_main_lock
821 * => caller should have pmap locked
822 * => we will gain the lock on the vm_page and allocate the new pv_entry
823 * => caller should adjust ptp's wire_count before calling
824 * => caller should not adjust pmap's wire_count
825 */
826
827 __inline static void
828 pmap_enter_pv(pg, pve, pmap, va, ptp, flags)
829 struct vm_page *pg;
830 struct pv_entry *pve; /* preallocated pve for us to use */
831 struct pmap *pmap;
832 vaddr_t va;
833 struct vm_page *ptp; /* PTP in pmap that maps this VA */
834 int flags;
835 {
836 pve->pv_pmap = pmap;
837 pve->pv_va = va;
838 pve->pv_ptp = ptp; /* NULL for kernel pmap */
839 pve->pv_flags = flags;
840 simple_lock(&pg->mdpage.pvh_slock); /* lock vm_page */
841 pve->pv_next = pg->mdpage.pvh_list; /* add to ... */
842 pg->mdpage.pvh_list = pve; /* ... locked list */
843 simple_unlock(&pg->mdpage.pvh_slock); /* unlock, done! */
844 if (pve->pv_flags & PT_W)
845 ++pmap->pm_stats.wired_count;
846 }
847
848 /*
849 * pmap_remove_pv: try to remove a mapping from a pv_list
850 *
851 * => caller should hold proper lock on pmap_main_lock
852 * => pmap should be locked
853 * => caller should hold lock on vm_page [so that attrs can be adjusted]
854 * => caller should adjust ptp's wire_count and free PTP if needed
855 * => caller should NOT adjust pmap's wire_count
856 * => we return the removed pve
857 */
858
859 __inline static struct pv_entry *
860 pmap_remove_pv(pg, pmap, va)
861 struct vm_page *pg;
862 struct pmap *pmap;
863 vaddr_t va;
864 {
865 struct pv_entry *pve, **prevptr;
866
867 prevptr = &pg->mdpage.pvh_list; /* previous pv_entry pointer */
868 pve = *prevptr;
869 while (pve) {
870 if (pve->pv_pmap == pmap && pve->pv_va == va) { /* match? */
871 *prevptr = pve->pv_next; /* remove it! */
872 if (pve->pv_flags & PT_W)
873 --pmap->pm_stats.wired_count;
874 break;
875 }
876 prevptr = &pve->pv_next; /* previous pointer */
877 pve = pve->pv_next; /* advance */
878 }
879 return(pve); /* return removed pve */
880 }
881
882 /*
883 *
884 * pmap_modify_pv: Update pv flags
885 *
886 * => caller should hold lock on vm_page [so that attrs can be adjusted]
887 * => caller should NOT adjust pmap's wire_count
888 * => caller must call pmap_vac_me_harder() if writable status of a page
889 * may have changed.
890 * => we return the old flags
891 *
892 * Modify a physical-virtual mapping in the pv table
893 */
894
895 /*__inline */
896 static u_int
897 pmap_modify_pv(pmap, va, pg, bic_mask, eor_mask)
898 struct pmap *pmap;
899 vaddr_t va;
900 struct vm_page *pg;
901 u_int bic_mask;
902 u_int eor_mask;
903 {
904 struct pv_entry *npv;
905 u_int flags, oflags;
906
907 /*
908 * There is at least one VA mapping this page.
909 */
910
911 for (npv = pg->mdpage.pvh_list; npv; npv = npv->pv_next) {
912 if (pmap == npv->pv_pmap && va == npv->pv_va) {
913 oflags = npv->pv_flags;
914 npv->pv_flags = flags =
915 ((oflags & ~bic_mask) ^ eor_mask);
916 if ((flags ^ oflags) & PT_W) {
917 if (flags & PT_W)
918 ++pmap->pm_stats.wired_count;
919 else
920 --pmap->pm_stats.wired_count;
921 }
922 return (oflags);
923 }
924 }
925 return (0);
926 }
927
928 /*
929 * Map the specified level 2 pagetable into the level 1 page table for
930 * the given pmap to cover a chunk of virtual address space starting from the
931 * address specified.
932 */
933 static /*__inline*/ void
934 pmap_map_in_l1(pmap, va, l2pa, selfref)
935 struct pmap *pmap;
936 vaddr_t va, l2pa;
937 boolean_t selfref;
938 {
939 vaddr_t ptva;
940
941 /* Calculate the index into the L1 page table. */
942 ptva = (va >> PDSHIFT) & ~3;
943
944 NPDEBUG(PDB_MAP_L1, printf("wiring %08lx in to pd%p pte0x%lx va0x%lx\n", l2pa,
945 pmap->pm_pdir, L1_PTE(l2pa), ptva));
946
947 /* Map page table into the L1. */
948 pmap->pm_pdir[ptva + 0] = L1_PTE(l2pa + 0x000);
949 pmap->pm_pdir[ptva + 1] = L1_PTE(l2pa + 0x400);
950 pmap->pm_pdir[ptva + 2] = L1_PTE(l2pa + 0x800);
951 pmap->pm_pdir[ptva + 3] = L1_PTE(l2pa + 0xc00);
952
953 /* Map the page table into the page table area. */
954 if (selfref) {
955 NPDEBUG(PDB_MAP_L1, printf("pt self reference %lx in %lx\n",
956 L2_PTE_NC_NB(l2pa, AP_KRW), pmap->pm_vptpt));
957 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) =
958 L2_PTE_NC_NB(l2pa, AP_KRW);
959 }
960 /* XXX should be a purge */
961 /* cpu_tlb_flushD();*/
962 }
963
964 #if 0
965 static /*__inline*/ void
966 pmap_unmap_in_l1(pmap, va)
967 struct pmap *pmap;
968 vaddr_t va;
969 {
970 vaddr_t ptva;
971
972 /* Calculate the index into the L1 page table. */
973 ptva = (va >> PDSHIFT) & ~3;
974
975 /* Unmap page table from the L1. */
976 pmap->pm_pdir[ptva + 0] = 0;
977 pmap->pm_pdir[ptva + 1] = 0;
978 pmap->pm_pdir[ptva + 2] = 0;
979 pmap->pm_pdir[ptva + 3] = 0;
980
981 /* Unmap the page table from the page table area. */
982 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) = 0;
983
984 /* XXX should be a purge */
985 /* cpu_tlb_flushD();*/
986 }
987 #endif
988
989 /*
990 * Used to map a range of physical addresses into kernel
991 * virtual address space.
992 *
993 * For now, VM is already on, we only need to map the
994 * specified memory.
995 */
996 vaddr_t
997 pmap_map(va, spa, epa, prot)
998 vaddr_t va, spa, epa;
999 int prot;
1000 {
1001 while (spa < epa) {
1002 pmap_kenter_pa(va, spa, prot);
1003 va += NBPG;
1004 spa += NBPG;
1005 }
1006 pmap_update(pmap_kernel());
1007 return(va);
1008 }
1009
1010
1011 /*
1012 * void pmap_bootstrap(pd_entry_t *kernel_l1pt, pv_addr_t kernel_ptpt)
1013 *
1014 * bootstrap the pmap system. This is called from initarm and allows
1015 * the pmap system to initailise any structures it requires.
1016 *
1017 * Currently this sets up the kernel_pmap that is statically allocated
1018 * and also allocated virtual addresses for certain page hooks.
1019 * Currently the only one page hook is allocated that is used
1020 * to zero physical pages of memory.
1021 * It also initialises the start and end address of the kernel data space.
1022 */
1023 extern paddr_t physical_freestart;
1024 extern paddr_t physical_freeend;
1025
1026 char *boot_head;
1027
1028 void
1029 pmap_bootstrap(kernel_l1pt, kernel_ptpt)
1030 pd_entry_t *kernel_l1pt;
1031 pv_addr_t kernel_ptpt;
1032 {
1033 pt_entry_t *pte;
1034 int loop;
1035 paddr_t start, end;
1036 #if NISADMA > 0
1037 paddr_t istart;
1038 psize_t isize;
1039 #endif
1040
1041 pmap_kernel()->pm_pdir = kernel_l1pt;
1042 pmap_kernel()->pm_pptpt = kernel_ptpt.pv_pa;
1043 pmap_kernel()->pm_vptpt = kernel_ptpt.pv_va;
1044 simple_lock_init(&pmap_kernel()->pm_lock);
1045 pmap_kernel()->pm_obj.pgops = NULL;
1046 TAILQ_INIT(&(pmap_kernel()->pm_obj.memq));
1047 pmap_kernel()->pm_obj.uo_npages = 0;
1048 pmap_kernel()->pm_obj.uo_refs = 1;
1049
1050 /*
1051 * Initialize PAGE_SIZE-dependent variables.
1052 */
1053 uvm_setpagesize();
1054
1055 npages = 0;
1056 loop = 0;
1057 while (loop < bootconfig.dramblocks) {
1058 start = (paddr_t)bootconfig.dram[loop].address;
1059 end = start + (bootconfig.dram[loop].pages * NBPG);
1060 if (start < physical_freestart)
1061 start = physical_freestart;
1062 if (end > physical_freeend)
1063 end = physical_freeend;
1064 #if 0
1065 printf("%d: %lx -> %lx\n", loop, start, end - 1);
1066 #endif
1067 #if NISADMA > 0
1068 if (pmap_isa_dma_range_intersect(start, end - start,
1069 &istart, &isize)) {
1070 /*
1071 * Place the pages that intersect with the
1072 * ISA DMA range onto the ISA DMA free list.
1073 */
1074 #if 0
1075 printf(" ISADMA 0x%lx -> 0x%lx\n", istart,
1076 istart + isize - 1);
1077 #endif
1078 uvm_page_physload(atop(istart),
1079 atop(istart + isize), atop(istart),
1080 atop(istart + isize), VM_FREELIST_ISADMA);
1081 npages += atop(istart + isize) - atop(istart);
1082
1083 /*
1084 * Load the pieces that come before
1085 * the intersection into the default
1086 * free list.
1087 */
1088 if (start < istart) {
1089 #if 0
1090 printf(" BEFORE 0x%lx -> 0x%lx\n",
1091 start, istart - 1);
1092 #endif
1093 uvm_page_physload(atop(start),
1094 atop(istart), atop(start),
1095 atop(istart), VM_FREELIST_DEFAULT);
1096 npages += atop(istart) - atop(start);
1097 }
1098
1099 /*
1100 * Load the pieces that come after
1101 * the intersection into the default
1102 * free list.
1103 */
1104 if ((istart + isize) < end) {
1105 #if 0
1106 printf(" AFTER 0x%lx -> 0x%lx\n",
1107 (istart + isize), end - 1);
1108 #endif
1109 uvm_page_physload(atop(istart + isize),
1110 atop(end), atop(istart + isize),
1111 atop(end), VM_FREELIST_DEFAULT);
1112 npages += atop(end) - atop(istart + isize);
1113 }
1114 } else {
1115 uvm_page_physload(atop(start), atop(end),
1116 atop(start), atop(end), VM_FREELIST_DEFAULT);
1117 npages += atop(end) - atop(start);
1118 }
1119 #else /* NISADMA > 0 */
1120 uvm_page_physload(atop(start), atop(end),
1121 atop(start), atop(end), VM_FREELIST_DEFAULT);
1122 npages += atop(end) - atop(start);
1123 #endif /* NISADMA > 0 */
1124 ++loop;
1125 }
1126
1127 #ifdef MYCROFT_HACK
1128 printf("npages = %ld\n", npages);
1129 #endif
1130
1131 virtual_avail = KERNEL_VM_BASE;
1132 virtual_end = KERNEL_VM_BASE + KERNEL_VM_SIZE - 1;
1133
1134 /*
1135 * now we allocate the "special" VAs which are used for tmp mappings
1136 * by the pmap (and other modules). we allocate the VAs by advancing
1137 * virtual_avail (note that there are no pages mapped at these VAs).
1138 * we find the PTE that maps the allocated VA via the linear PTE
1139 * mapping.
1140 */
1141
1142 pte = ((pt_entry_t *) PTE_BASE) + atop(virtual_avail);
1143
1144 csrcp = virtual_avail; csrc_pte = pte;
1145 virtual_avail += PAGE_SIZE; pte++;
1146
1147 cdstp = virtual_avail; cdst_pte = pte;
1148 virtual_avail += PAGE_SIZE; pte++;
1149
1150 memhook = (char *) virtual_avail; /* don't need pte */
1151 virtual_avail += PAGE_SIZE; pte++;
1152
1153 msgbufaddr = (caddr_t) virtual_avail; /* don't need pte */
1154 virtual_avail += round_page(MSGBUFSIZE);
1155 pte += atop(round_page(MSGBUFSIZE));
1156
1157 /*
1158 * init the static-global locks and global lists.
1159 */
1160 spinlockinit(&pmap_main_lock, "pmaplk", 0);
1161 simple_lock_init(&pvalloc_lock);
1162 simple_lock_init(&pmaps_lock);
1163 LIST_INIT(&pmaps);
1164 TAILQ_INIT(&pv_freepages);
1165 TAILQ_INIT(&pv_unusedpgs);
1166
1167 /*
1168 * initialize the pmap pool.
1169 */
1170
1171 pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl",
1172 &pool_allocator_nointr);
1173
1174 cpu_dcache_wbinv_all();
1175 }
1176
1177 /*
1178 * void pmap_init(void)
1179 *
1180 * Initialize the pmap module.
1181 * Called by vm_init() in vm/vm_init.c in order to initialise
1182 * any structures that the pmap system needs to map virtual memory.
1183 */
1184
1185 extern int physmem;
1186
1187 void
1188 pmap_init()
1189 {
1190
1191 /*
1192 * Set the available memory vars - These do not map to real memory
1193 * addresses and cannot as the physical memory is fragmented.
1194 * They are used by ps for %mem calculations.
1195 * One could argue whether this should be the entire memory or just
1196 * the memory that is useable in a user process.
1197 */
1198 avail_start = 0;
1199 avail_end = physmem * NBPG;
1200
1201 /*
1202 * now we need to free enough pv_entry structures to allow us to get
1203 * the kmem_map/kmem_object allocated and inited (done after this
1204 * function is finished). to do this we allocate one bootstrap page out
1205 * of kernel_map and use it to provide an initial pool of pv_entry
1206 * structures. we never free this page.
1207 */
1208
1209 pv_initpage = (struct pv_page *) uvm_km_alloc(kernel_map, PAGE_SIZE);
1210 if (pv_initpage == NULL)
1211 panic("pmap_init: pv_initpage");
1212 pv_cachedva = 0; /* a VA we have allocated but not used yet */
1213 pv_nfpvents = 0;
1214 (void) pmap_add_pvpage(pv_initpage, FALSE);
1215
1216 pmap_initialized = TRUE;
1217
1218 /* Initialise our L1 page table queues and counters */
1219 SIMPLEQ_INIT(&l1pt_static_queue);
1220 l1pt_static_queue_count = 0;
1221 l1pt_static_create_count = 0;
1222 SIMPLEQ_INIT(&l1pt_queue);
1223 l1pt_queue_count = 0;
1224 l1pt_create_count = 0;
1225 l1pt_reuse_count = 0;
1226 }
1227
1228 /*
1229 * pmap_postinit()
1230 *
1231 * This routine is called after the vm and kmem subsystems have been
1232 * initialised. This allows the pmap code to perform any initialisation
1233 * that can only be done one the memory allocation is in place.
1234 */
1235
1236 void
1237 pmap_postinit()
1238 {
1239 int loop;
1240 struct l1pt *pt;
1241
1242 #ifdef PMAP_STATIC_L1S
1243 for (loop = 0; loop < PMAP_STATIC_L1S; ++loop) {
1244 #else /* PMAP_STATIC_L1S */
1245 for (loop = 0; loop < max_processes; ++loop) {
1246 #endif /* PMAP_STATIC_L1S */
1247 /* Allocate a L1 page table */
1248 pt = pmap_alloc_l1pt();
1249 if (!pt)
1250 panic("Cannot allocate static L1 page tables\n");
1251
1252 /* Clean it */
1253 bzero((void *)pt->pt_va, PD_SIZE);
1254 pt->pt_flags |= (PTFLAG_STATIC | PTFLAG_CLEAN);
1255 /* Add the page table to the queue */
1256 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pt, pt_queue);
1257 ++l1pt_static_queue_count;
1258 ++l1pt_static_create_count;
1259 }
1260 }
1261
1262
1263 /*
1264 * Create and return a physical map.
1265 *
1266 * If the size specified for the map is zero, the map is an actual physical
1267 * map, and may be referenced by the hardware.
1268 *
1269 * If the size specified is non-zero, the map will be used in software only,
1270 * and is bounded by that size.
1271 */
1272
1273 pmap_t
1274 pmap_create()
1275 {
1276 struct pmap *pmap;
1277
1278 /*
1279 * Fetch pmap entry from the pool
1280 */
1281
1282 pmap = pool_get(&pmap_pmap_pool, PR_WAITOK);
1283 /* XXX is this really needed! */
1284 memset(pmap, 0, sizeof(*pmap));
1285
1286 simple_lock_init(&pmap->pm_obj.vmobjlock);
1287 pmap->pm_obj.pgops = NULL; /* currently not a mappable object */
1288 TAILQ_INIT(&pmap->pm_obj.memq);
1289 pmap->pm_obj.uo_npages = 0;
1290 pmap->pm_obj.uo_refs = 1;
1291 pmap->pm_stats.wired_count = 0;
1292 pmap->pm_stats.resident_count = 1;
1293 pmap->pm_ptphint = NULL;
1294
1295 /* Now init the machine part of the pmap */
1296 pmap_pinit(pmap);
1297 return(pmap);
1298 }
1299
1300 /*
1301 * pmap_alloc_l1pt()
1302 *
1303 * This routine allocates physical and virtual memory for a L1 page table
1304 * and wires it.
1305 * A l1pt structure is returned to describe the allocated page table.
1306 *
1307 * This routine is allowed to fail if the required memory cannot be allocated.
1308 * In this case NULL is returned.
1309 */
1310
1311 struct l1pt *
1312 pmap_alloc_l1pt(void)
1313 {
1314 paddr_t pa;
1315 vaddr_t va;
1316 struct l1pt *pt;
1317 int error;
1318 struct vm_page *m;
1319 pt_entry_t *ptes;
1320
1321 /* Allocate virtual address space for the L1 page table */
1322 va = uvm_km_valloc(kernel_map, PD_SIZE);
1323 if (va == 0) {
1324 #ifdef DIAGNOSTIC
1325 PDEBUG(0,
1326 printf("pmap: Cannot allocate pageable memory for L1\n"));
1327 #endif /* DIAGNOSTIC */
1328 return(NULL);
1329 }
1330
1331 /* Allocate memory for the l1pt structure */
1332 pt = (struct l1pt *)malloc(sizeof(struct l1pt), M_VMPMAP, M_WAITOK);
1333
1334 /*
1335 * Allocate pages from the VM system.
1336 */
1337 TAILQ_INIT(&pt->pt_plist);
1338 error = uvm_pglistalloc(PD_SIZE, physical_start, physical_end,
1339 PD_SIZE, 0, &pt->pt_plist, 1, M_WAITOK);
1340 if (error) {
1341 #ifdef DIAGNOSTIC
1342 PDEBUG(0,
1343 printf("pmap: Cannot allocate physical mem for L1 (%d)\n",
1344 error));
1345 #endif /* DIAGNOSTIC */
1346 /* Release the resources we already have claimed */
1347 free(pt, M_VMPMAP);
1348 uvm_km_free(kernel_map, va, PD_SIZE);
1349 return(NULL);
1350 }
1351
1352 /* Map our physical pages into our virtual space */
1353 pt->pt_va = va;
1354 m = TAILQ_FIRST(&pt->pt_plist);
1355 ptes = pmap_map_ptes(pmap_kernel());
1356 while (m && va < (pt->pt_va + PD_SIZE)) {
1357 pa = VM_PAGE_TO_PHYS(m);
1358
1359 pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE);
1360
1361 /* Revoke cacheability and bufferability */
1362 /* XXX should be done better than this */
1363 ptes[arm_btop(va)] &= ~(PT_C | PT_B);
1364
1365 va += NBPG;
1366 m = m->pageq.tqe_next;
1367 }
1368 pmap_unmap_ptes(pmap_kernel());
1369 pmap_update(pmap_kernel());
1370
1371 #ifdef DIAGNOSTIC
1372 if (m)
1373 panic("pmap_alloc_l1pt: pglist not empty\n");
1374 #endif /* DIAGNOSTIC */
1375
1376 pt->pt_flags = 0;
1377 return(pt);
1378 }
1379
1380 /*
1381 * Free a L1 page table previously allocated with pmap_alloc_l1pt().
1382 */
1383 static void
1384 pmap_free_l1pt(pt)
1385 struct l1pt *pt;
1386 {
1387 /* Separate the physical memory for the virtual space */
1388 pmap_kremove(pt->pt_va, PD_SIZE);
1389 pmap_update(pmap_kernel());
1390
1391 /* Return the physical memory */
1392 uvm_pglistfree(&pt->pt_plist);
1393
1394 /* Free the virtual space */
1395 uvm_km_free(kernel_map, pt->pt_va, PD_SIZE);
1396
1397 /* Free the l1pt structure */
1398 free(pt, M_VMPMAP);
1399 }
1400
1401 /*
1402 * Allocate a page directory.
1403 * This routine will either allocate a new page directory from the pool
1404 * of L1 page tables currently held by the kernel or it will allocate
1405 * a new one via pmap_alloc_l1pt().
1406 * It will then initialise the l1 page table for use.
1407 *
1408 * XXX must tidy up and fix this code, not happy about how it does the pmaps_locking
1409 */
1410 static int
1411 pmap_allocpagedir(pmap)
1412 struct pmap *pmap;
1413 {
1414 paddr_t pa;
1415 struct l1pt *pt;
1416 pt_entry_t *pte;
1417
1418 PDEBUG(0, printf("pmap_allocpagedir(%p)\n", pmap));
1419
1420 /* Do we have any spare L1's lying around ? */
1421 if (l1pt_static_queue_count) {
1422 --l1pt_static_queue_count;
1423 pt = l1pt_static_queue.sqh_first;
1424 SIMPLEQ_REMOVE_HEAD(&l1pt_static_queue, pt, pt_queue);
1425 } else if (l1pt_queue_count) {
1426 --l1pt_queue_count;
1427 pt = l1pt_queue.sqh_first;
1428 SIMPLEQ_REMOVE_HEAD(&l1pt_queue, pt, pt_queue);
1429 ++l1pt_reuse_count;
1430 } else {
1431 pt = pmap_alloc_l1pt();
1432 if (!pt)
1433 return(ENOMEM);
1434 ++l1pt_create_count;
1435 }
1436
1437 /* Store the pointer to the l1 descriptor in the pmap. */
1438 pmap->pm_l1pt = pt;
1439
1440 /* Get the physical address of the start of the l1 */
1441 pa = VM_PAGE_TO_PHYS(TAILQ_FIRST(&pt->pt_plist));
1442
1443 /* Store the virtual address of the l1 in the pmap. */
1444 pmap->pm_pdir = (pd_entry_t *)pt->pt_va;
1445
1446 /* Clean the L1 if it is dirty */
1447 if (!(pt->pt_flags & PTFLAG_CLEAN))
1448 bzero((void *)pmap->pm_pdir, (PD_SIZE - KERNEL_PD_SIZE));
1449
1450 /* Allocate a page table to map all the page tables for this pmap */
1451
1452 #ifdef DIAGNOSTIC
1453 if (pmap->pm_vptpt) {
1454 /* XXX What if we have one already ? */
1455 panic("pmap_allocpagedir: have pt already\n");
1456 }
1457 #endif /* DIAGNOSTIC */
1458 pmap->pm_vptpt = uvm_km_zalloc(kernel_map, NBPG);
1459 if (pmap->pm_vptpt == 0) {
1460 pmap_freepagedir(pmap);
1461 return(ENOMEM);
1462 }
1463
1464 /* need to lock this all up for growkernel */
1465 simple_lock(&pmaps_lock);
1466 /* wish we didn't have to keep this locked... */
1467
1468 /* Duplicate the kernel mappings. */
1469 bcopy((char *)pmap_kernel()->pm_pdir + (PD_SIZE - KERNEL_PD_SIZE),
1470 (char *)pmap->pm_pdir + (PD_SIZE - KERNEL_PD_SIZE),
1471 KERNEL_PD_SIZE);
1472
1473 pte = vtopte(pmap->pm_vptpt);
1474 pmap->pm_pptpt = l2pte_pa(*pte);
1475
1476 /* Revoke cacheability and bufferability */
1477 /* XXX should be done better than this */
1478 *pte &= ~(PT_C | PT_B);
1479
1480 /* Wire in this page table */
1481 pmap_map_in_l1(pmap, PTE_BASE, pmap->pm_pptpt, TRUE);
1482
1483 pt->pt_flags &= ~PTFLAG_CLEAN; /* L1 is dirty now */
1484
1485 /*
1486 * Map the kernel page tables into the new PT map.
1487 */
1488 bcopy((char *)(PTE_BASE
1489 + (PTE_BASE >> (PGSHIFT - 2))
1490 + ((PD_SIZE - KERNEL_PD_SIZE) >> 2)),
1491 (char *)pmap->pm_vptpt + ((PD_SIZE - KERNEL_PD_SIZE) >> 2),
1492 (KERNEL_PD_SIZE >> 2));
1493
1494 LIST_INSERT_HEAD(&pmaps, pmap, pm_list);
1495 simple_unlock(&pmaps_lock);
1496
1497 return(0);
1498 }
1499
1500
1501 /*
1502 * Initialize a preallocated and zeroed pmap structure,
1503 * such as one in a vmspace structure.
1504 */
1505
1506 void
1507 pmap_pinit(pmap)
1508 struct pmap *pmap;
1509 {
1510 int backoff = 6;
1511 int retry = 10;
1512
1513 PDEBUG(0, printf("pmap_pinit(%p)\n", pmap));
1514
1515 /* Keep looping until we succeed in allocating a page directory */
1516 while (pmap_allocpagedir(pmap) != 0) {
1517 /*
1518 * Ok we failed to allocate a suitable block of memory for an
1519 * L1 page table. This means that either:
1520 * 1. 16KB of virtual address space could not be allocated
1521 * 2. 16KB of physically contiguous memory on a 16KB boundary
1522 * could not be allocated.
1523 *
1524 * Since we cannot fail we will sleep for a while and try
1525 * again.
1526 *
1527 * Searching for a suitable L1 PT is expensive:
1528 * to avoid hogging the system when memory is really
1529 * scarce, use an exponential back-off so that
1530 * eventually we won't retry more than once every 8
1531 * seconds. This should allow other processes to run
1532 * to completion and free up resources.
1533 */
1534 (void) ltsleep(&lbolt, PVM, "l1ptwait", (hz << 3) >> backoff,
1535 NULL);
1536 if (--retry == 0) {
1537 retry = 10;
1538 if (backoff)
1539 --backoff;
1540 }
1541 }
1542
1543 /* Map zero page for the pmap. This will also map the L2 for it */
1544 pmap_enter(pmap, 0x00000000, systempage.pv_pa,
1545 VM_PROT_READ, VM_PROT_READ | PMAP_WIRED);
1546 pmap_update(pmap);
1547 }
1548
1549
1550 void
1551 pmap_freepagedir(pmap)
1552 struct pmap *pmap;
1553 {
1554 /* Free the memory used for the page table mapping */
1555 if (pmap->pm_vptpt != 0)
1556 uvm_km_free(kernel_map, (vaddr_t)pmap->pm_vptpt, NBPG);
1557
1558 /* junk the L1 page table */
1559 if (pmap->pm_l1pt->pt_flags & PTFLAG_STATIC) {
1560 /* Add the page table to the queue */
1561 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pmap->pm_l1pt, pt_queue);
1562 ++l1pt_static_queue_count;
1563 } else if (l1pt_queue_count < 8) {
1564 /* Add the page table to the queue */
1565 SIMPLEQ_INSERT_TAIL(&l1pt_queue, pmap->pm_l1pt, pt_queue);
1566 ++l1pt_queue_count;
1567 } else
1568 pmap_free_l1pt(pmap->pm_l1pt);
1569 }
1570
1571
1572 /*
1573 * Retire the given physical map from service.
1574 * Should only be called if the map contains no valid mappings.
1575 */
1576
1577 void
1578 pmap_destroy(pmap)
1579 struct pmap *pmap;
1580 {
1581 struct vm_page *page;
1582 int count;
1583
1584 if (pmap == NULL)
1585 return;
1586
1587 PDEBUG(0, printf("pmap_destroy(%p)\n", pmap));
1588
1589 /*
1590 * Drop reference count
1591 */
1592 simple_lock(&pmap->pm_obj.vmobjlock);
1593 count = --pmap->pm_obj.uo_refs;
1594 simple_unlock(&pmap->pm_obj.vmobjlock);
1595 if (count > 0) {
1596 return;
1597 }
1598
1599 /*
1600 * reference count is zero, free pmap resources and then free pmap.
1601 */
1602
1603 /*
1604 * remove it from global list of pmaps
1605 */
1606
1607 simple_lock(&pmaps_lock);
1608 LIST_REMOVE(pmap, pm_list);
1609 simple_unlock(&pmaps_lock);
1610
1611 /* Remove the zero page mapping */
1612 pmap_remove(pmap, 0x00000000, 0x00000000 + NBPG);
1613 pmap_update(pmap);
1614
1615 /*
1616 * Free any page tables still mapped
1617 * This is only temporay until pmap_enter can count the number
1618 * of mappings made in a page table. Then pmap_remove() can
1619 * reduce the count and free the pagetable when the count
1620 * reaches zero. Note that entries in this list should match the
1621 * contents of the ptpt, however this is faster than walking a 1024
1622 * entries looking for pt's
1623 * taken from i386 pmap.c
1624 */
1625 while ((page = TAILQ_FIRST(&pmap->pm_obj.memq)) != NULL) {
1626 KASSERT((page->flags & PG_BUSY) == 0);
1627 page->wire_count = 0;
1628 uvm_pagefree(page);
1629 }
1630
1631 /* Free the page dir */
1632 pmap_freepagedir(pmap);
1633
1634 /* return the pmap to the pool */
1635 pool_put(&pmap_pmap_pool, pmap);
1636 }
1637
1638
1639 /*
1640 * void pmap_reference(struct pmap *pmap)
1641 *
1642 * Add a reference to the specified pmap.
1643 */
1644
1645 void
1646 pmap_reference(pmap)
1647 struct pmap *pmap;
1648 {
1649 if (pmap == NULL)
1650 return;
1651
1652 simple_lock(&pmap->pm_lock);
1653 pmap->pm_obj.uo_refs++;
1654 simple_unlock(&pmap->pm_lock);
1655 }
1656
1657 /*
1658 * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
1659 *
1660 * Return the start and end addresses of the kernel's virtual space.
1661 * These values are setup in pmap_bootstrap and are updated as pages
1662 * are allocated.
1663 */
1664
1665 void
1666 pmap_virtual_space(start, end)
1667 vaddr_t *start;
1668 vaddr_t *end;
1669 {
1670 *start = virtual_avail;
1671 *end = virtual_end;
1672 }
1673
1674
1675 /*
1676 * Activate the address space for the specified process. If the process
1677 * is the current process, load the new MMU context.
1678 */
1679 void
1680 pmap_activate(p)
1681 struct proc *p;
1682 {
1683 struct pmap *pmap = p->p_vmspace->vm_map.pmap;
1684 struct pcb *pcb = &p->p_addr->u_pcb;
1685
1686 (void) pmap_extract(pmap_kernel(), (vaddr_t)pmap->pm_pdir,
1687 (paddr_t *)&pcb->pcb_pagedir);
1688
1689 PDEBUG(0, printf("pmap_activate: p=%p pmap=%p pcb=%p pdir=%p l1=%p\n",
1690 p, pmap, pcb, pmap->pm_pdir, pcb->pcb_pagedir));
1691
1692 if (p == curproc) {
1693 PDEBUG(0, printf("pmap_activate: setting TTB\n"));
1694 setttb((u_int)pcb->pcb_pagedir);
1695 }
1696 #if 0
1697 pmap->pm_pdchanged = FALSE;
1698 #endif
1699 }
1700
1701
1702 /*
1703 * Deactivate the address space of the specified process.
1704 */
1705 void
1706 pmap_deactivate(p)
1707 struct proc *p;
1708 {
1709 }
1710
1711 /*
1712 * Perform any deferred pmap operations.
1713 */
1714 void
1715 pmap_update(struct pmap *pmap)
1716 {
1717
1718 /*
1719 * We haven't deferred any pmap operations, but we do need to
1720 * make sure TLB/cache operations have completed.
1721 */
1722 cpu_cpwait();
1723 }
1724
1725 /*
1726 * pmap_clean_page()
1727 *
1728 * This is a local function used to work out the best strategy to clean
1729 * a single page referenced by its entry in the PV table. It's used by
1730 * pmap_copy_page, pmap_zero page and maybe some others later on.
1731 *
1732 * Its policy is effectively:
1733 * o If there are no mappings, we don't bother doing anything with the cache.
1734 * o If there is one mapping, we clean just that page.
1735 * o If there are multiple mappings, we clean the entire cache.
1736 *
1737 * So that some functions can be further optimised, it returns 0 if it didn't
1738 * clean the entire cache, or 1 if it did.
1739 *
1740 * XXX One bug in this routine is that if the pv_entry has a single page
1741 * mapped at 0x00000000 a whole cache clean will be performed rather than
1742 * just the 1 page. Since this should not occur in everyday use and if it does
1743 * it will just result in not the most efficient clean for the page.
1744 */
1745 static int
1746 pmap_clean_page(pv, is_src)
1747 struct pv_entry *pv;
1748 boolean_t is_src;
1749 {
1750 struct pmap *pmap;
1751 struct pv_entry *npv;
1752 int cache_needs_cleaning = 0;
1753 vaddr_t page_to_clean = 0;
1754
1755 if (pv == NULL)
1756 /* nothing mapped in so nothing to flush */
1757 return (0);
1758
1759 /* Since we flush the cache each time we change curproc, we
1760 * only need to flush the page if it is in the current pmap.
1761 */
1762 if (curproc)
1763 pmap = curproc->p_vmspace->vm_map.pmap;
1764 else
1765 pmap = pmap_kernel();
1766
1767 for (npv = pv; npv; npv = npv->pv_next) {
1768 if (npv->pv_pmap == pmap) {
1769 /* The page is mapped non-cacheable in
1770 * this map. No need to flush the cache.
1771 */
1772 if (npv->pv_flags & PT_NC) {
1773 #ifdef DIAGNOSTIC
1774 if (cache_needs_cleaning)
1775 panic("pmap_clean_page: "
1776 "cache inconsistency");
1777 #endif
1778 break;
1779 }
1780 #if 0
1781 /* This doesn't work, because pmap_protect
1782 doesn't flush changes on pages that it
1783 has write-protected. */
1784
1785 /* If the page is not writable and this
1786 is the source, then there is no need
1787 to flush it from the cache. */
1788 else if (is_src && ! (npv->pv_flags & PT_Wr))
1789 continue;
1790 #endif
1791 if (cache_needs_cleaning){
1792 page_to_clean = 0;
1793 break;
1794 }
1795 else
1796 page_to_clean = npv->pv_va;
1797 cache_needs_cleaning = 1;
1798 }
1799 }
1800
1801 if (page_to_clean)
1802 cpu_idcache_wbinv_range(page_to_clean, NBPG);
1803 else if (cache_needs_cleaning) {
1804 cpu_idcache_wbinv_all();
1805 return (1);
1806 }
1807 return (0);
1808 }
1809
1810 /*
1811 * pmap_zero_page()
1812 *
1813 * Zero a given physical page by mapping it at a page hook point.
1814 * In doing the zero page op, the page we zero is mapped cachable, as with
1815 * StrongARM accesses to non-cached pages are non-burst making writing
1816 * _any_ bulk data very slow.
1817 */
1818 void
1819 pmap_zero_page(phys)
1820 paddr_t phys;
1821 {
1822 struct vm_page *pg;
1823
1824 /* Get an entry for this page, and clean it it. */
1825 pg = PHYS_TO_VM_PAGE(phys);
1826 simple_lock(&pg->mdpage.pvh_slock);
1827 pmap_clean_page(pg->mdpage.pvh_list, FALSE);
1828 simple_unlock(&pg->mdpage.pvh_slock);
1829
1830 /*
1831 * Hook in the page, zero it, and purge the cache for that
1832 * zeroed page. Invalidate the TLB as needed.
1833 */
1834 *cdst_pte = L2_PTE(phys & PG_FRAME, AP_KRW);
1835 cpu_tlb_flushD_SE(cdstp);
1836 cpu_cpwait();
1837 bzero_page(cdstp);
1838 cpu_dcache_wbinv_range(cdstp, NBPG);
1839 }
1840
1841 /* pmap_pageidlezero()
1842 *
1843 * The same as above, except that we assume that the page is not
1844 * mapped. This means we never have to flush the cache first. Called
1845 * from the idle loop.
1846 */
1847 boolean_t
1848 pmap_pageidlezero(phys)
1849 paddr_t phys;
1850 {
1851 int i, *ptr;
1852 boolean_t rv = TRUE;
1853
1854 #ifdef DIAGNOSTIC
1855 struct vm_page *pg;
1856
1857 pg = PHYS_TO_VM_PAGE(phys);
1858 if (pg->mdpage.pvh_list != NULL)
1859 panic("pmap_pageidlezero: zeroing mapped page\n");
1860 #endif
1861
1862 /*
1863 * Hook in the page, zero it, and purge the cache for that
1864 * zeroed page. Invalidate the TLB as needed.
1865 */
1866 *cdst_pte = L2_PTE(phys & PG_FRAME, AP_KRW);
1867 cpu_tlb_flushD_SE(cdstp);
1868 cpu_cpwait();
1869
1870 for (i = 0, ptr = (int *)cdstp;
1871 i < (NBPG / sizeof(int)); i++) {
1872 if (sched_whichqs != 0) {
1873 /*
1874 * A process has become ready. Abort now,
1875 * so we don't keep it waiting while we
1876 * do slow memory access to finish this
1877 * page.
1878 */
1879 rv = FALSE;
1880 break;
1881 }
1882 *ptr++ = 0;
1883 }
1884
1885 if (rv)
1886 /*
1887 * if we aborted we'll rezero this page again later so don't
1888 * purge it unless we finished it
1889 */
1890 cpu_dcache_wbinv_range(cdstp, NBPG);
1891 return (rv);
1892 }
1893
1894 /*
1895 * pmap_copy_page()
1896 *
1897 * Copy one physical page into another, by mapping the pages into
1898 * hook points. The same comment regarding cachability as in
1899 * pmap_zero_page also applies here.
1900 */
1901 void
1902 pmap_copy_page(src, dest)
1903 paddr_t src;
1904 paddr_t dest;
1905 {
1906 struct vm_page *src_pg, *dest_pg;
1907 boolean_t cleanedcache;
1908
1909 /* Get PV entries for the pages, and clean them if needed. */
1910 src_pg = PHYS_TO_VM_PAGE(src);
1911
1912 simple_lock(&src_pg->mdpage.pvh_slock);
1913 cleanedcache = pmap_clean_page(src_pg->mdpage.pvh_list, TRUE);
1914 simple_unlock(&src_pg->mdpage.pvh_slock);
1915
1916 if (cleanedcache == 0) {
1917 dest_pg = PHYS_TO_VM_PAGE(dest);
1918 simple_lock(&dest_pg->mdpage.pvh_slock);
1919 pmap_clean_page(dest_pg->mdpage.pvh_list, FALSE);
1920 simple_unlock(&dest_pg->mdpage.pvh_slock);
1921 }
1922 /*
1923 * Map the pages into the page hook points, copy them, and purge
1924 * the cache for the appropriate page. Invalidate the TLB
1925 * as required.
1926 */
1927 *csrc_pte = L2_PTE(src & PG_FRAME, AP_KR);
1928 *cdst_pte = L2_PTE(dest & PG_FRAME, AP_KRW);
1929 cpu_tlb_flushD_SE(csrcp);
1930 cpu_tlb_flushD_SE(cdstp);
1931 cpu_cpwait();
1932 bcopy_page(csrcp, cdstp);
1933 cpu_dcache_inv_range(csrcp, NBPG);
1934 cpu_dcache_wbinv_range(cdstp, NBPG);
1935 }
1936
1937 #if 0
1938 void
1939 pmap_pte_addref(pmap, va)
1940 struct pmap *pmap;
1941 vaddr_t va;
1942 {
1943 pd_entry_t *pde;
1944 paddr_t pa;
1945 struct vm_page *m;
1946
1947 if (pmap == pmap_kernel())
1948 return;
1949
1950 pde = pmap_pde(pmap, va & ~(3 << PDSHIFT));
1951 pa = pmap_pte_pa(pde);
1952 m = PHYS_TO_VM_PAGE(pa);
1953 ++m->wire_count;
1954 #ifdef MYCROFT_HACK
1955 printf("addref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
1956 pmap, va, pde, pa, m, m->wire_count);
1957 #endif
1958 }
1959
1960 void
1961 pmap_pte_delref(pmap, va)
1962 struct pmap *pmap;
1963 vaddr_t va;
1964 {
1965 pd_entry_t *pde;
1966 paddr_t pa;
1967 struct vm_page *m;
1968
1969 if (pmap == pmap_kernel())
1970 return;
1971
1972 pde = pmap_pde(pmap, va & ~(3 << PDSHIFT));
1973 pa = pmap_pte_pa(pde);
1974 m = PHYS_TO_VM_PAGE(pa);
1975 --m->wire_count;
1976 #ifdef MYCROFT_HACK
1977 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n",
1978 pmap, va, pde, pa, m, m->wire_count);
1979 #endif
1980 if (m->wire_count == 0) {
1981 #ifdef MYCROFT_HACK
1982 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p\n",
1983 pmap, va, pde, pa, m);
1984 #endif
1985 pmap_unmap_in_l1(pmap, va);
1986 uvm_pagefree(m);
1987 --pmap->pm_stats.resident_count;
1988 }
1989 }
1990 #else
1991 #define pmap_pte_addref(pmap, va)
1992 #define pmap_pte_delref(pmap, va)
1993 #endif
1994
1995 /*
1996 * Since we have a virtually indexed cache, we may need to inhibit caching if
1997 * there is more than one mapping and at least one of them is writable.
1998 * Since we purge the cache on every context switch, we only need to check for
1999 * other mappings within the same pmap, or kernel_pmap.
2000 * This function is also called when a page is unmapped, to possibly reenable
2001 * caching on any remaining mappings.
2002 *
2003 * The code implements the following logic, where:
2004 *
2005 * KW = # of kernel read/write pages
2006 * KR = # of kernel read only pages
2007 * UW = # of user read/write pages
2008 * UR = # of user read only pages
2009 * OW = # of user read/write pages in another pmap, then
2010 *
2011 * KC = kernel mapping is cacheable
2012 * UC = user mapping is cacheable
2013 *
2014 * KW=0,KR=0 KW=0,KR>0 KW=1,KR=0 KW>1,KR>=0
2015 * +---------------------------------------------
2016 * UW=0,UR=0,OW=0 | --- KC=1 KC=1 KC=0
2017 * UW=0,UR>0,OW=0 | UC=1 KC=1,UC=1 KC=0,UC=0 KC=0,UC=0
2018 * UW=0,UR>0,OW>0 | UC=1 KC=0,UC=1 KC=0,UC=0 KC=0,UC=0
2019 * UW=1,UR=0,OW=0 | UC=1 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0
2020 * UW>1,UR>=0,OW>=0 | UC=0 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0
2021 *
2022 * Note that the pmap must have it's ptes mapped in, and passed with ptes.
2023 */
2024 __inline static void
2025 pmap_vac_me_harder(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2026 boolean_t clear_cache)
2027 {
2028 if (pmap == pmap_kernel())
2029 pmap_vac_me_kpmap(pmap, pg, ptes, clear_cache);
2030 else
2031 pmap_vac_me_user(pmap, pg, ptes, clear_cache);
2032 }
2033
2034 static void
2035 pmap_vac_me_kpmap(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2036 boolean_t clear_cache)
2037 {
2038 int user_entries = 0;
2039 int user_writable = 0;
2040 int user_cacheable = 0;
2041 int kernel_entries = 0;
2042 int kernel_writable = 0;
2043 int kernel_cacheable = 0;
2044 struct pv_entry *pv;
2045 struct pmap *last_pmap = pmap;
2046
2047 #ifdef DIAGNOSTIC
2048 if (pmap != pmap_kernel())
2049 panic("pmap_vac_me_kpmap: pmap != pmap_kernel()");
2050 #endif
2051
2052 /*
2053 * Pass one, see if there are both kernel and user pmaps for
2054 * this page. Calculate whether there are user-writable or
2055 * kernel-writable pages.
2056 */
2057 for (pv = pg->mdpage.pvh_list; pv != NULL; pv = pv->pv_next) {
2058 if (pv->pv_pmap != pmap) {
2059 user_entries++;
2060 if (pv->pv_flags & PT_Wr)
2061 user_writable++;
2062 if ((pv->pv_flags & PT_NC) == 0)
2063 user_cacheable++;
2064 } else {
2065 kernel_entries++;
2066 if (pv->pv_flags & PT_Wr)
2067 kernel_writable++;
2068 if ((pv->pv_flags & PT_NC) == 0)
2069 kernel_cacheable++;
2070 }
2071 }
2072
2073 /*
2074 * We know we have just been updating a kernel entry, so if
2075 * all user pages are already cacheable, then there is nothing
2076 * further to do.
2077 */
2078 if (kernel_entries == 0 &&
2079 user_cacheable == user_entries)
2080 return;
2081
2082 if (user_entries) {
2083 /*
2084 * Scan over the list again, for each entry, if it
2085 * might not be set correctly, call pmap_vac_me_user
2086 * to recalculate the settings.
2087 */
2088 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
2089 /*
2090 * We know kernel mappings will get set
2091 * correctly in other calls. We also know
2092 * that if the pmap is the same as last_pmap
2093 * then we've just handled this entry.
2094 */
2095 if (pv->pv_pmap == pmap || pv->pv_pmap == last_pmap)
2096 continue;
2097 /*
2098 * If there are kernel entries and this page
2099 * is writable but non-cacheable, then we can
2100 * skip this entry also.
2101 */
2102 if (kernel_entries > 0 &&
2103 (pv->pv_flags & (PT_NC | PT_Wr)) ==
2104 (PT_NC | PT_Wr))
2105 continue;
2106 /*
2107 * Similarly if there are no kernel-writable
2108 * entries and the page is already
2109 * read-only/cacheable.
2110 */
2111 if (kernel_writable == 0 &&
2112 (pv->pv_flags & (PT_NC | PT_Wr)) == 0)
2113 continue;
2114 /*
2115 * For some of the remaining cases, we know
2116 * that we must recalculate, but for others we
2117 * can't tell if they are correct or not, so
2118 * we recalculate anyway.
2119 */
2120 pmap_unmap_ptes(last_pmap);
2121 last_pmap = pv->pv_pmap;
2122 ptes = pmap_map_ptes(last_pmap);
2123 pmap_vac_me_user(last_pmap, pg, ptes,
2124 pmap_is_curpmap(last_pmap));
2125 }
2126 /* Restore the pte mapping that was passed to us. */
2127 if (last_pmap != pmap) {
2128 pmap_unmap_ptes(last_pmap);
2129 ptes = pmap_map_ptes(pmap);
2130 }
2131 if (kernel_entries == 0)
2132 return;
2133 }
2134
2135 pmap_vac_me_user(pmap, pg, ptes, clear_cache);
2136 return;
2137 }
2138
2139 static void
2140 pmap_vac_me_user(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes,
2141 boolean_t clear_cache)
2142 {
2143 struct pmap *kpmap = pmap_kernel();
2144 struct pv_entry *pv, *npv;
2145 int entries = 0;
2146 int writable = 0;
2147 int cacheable_entries = 0;
2148 int kern_cacheable = 0;
2149 int other_writable = 0;
2150
2151 pv = pg->mdpage.pvh_list;
2152 KASSERT(ptes != NULL);
2153
2154 /*
2155 * Count mappings and writable mappings in this pmap.
2156 * Include kernel mappings as part of our own.
2157 * Keep a pointer to the first one.
2158 */
2159 for (npv = pv; npv; npv = npv->pv_next) {
2160 /* Count mappings in the same pmap */
2161 if (pmap == npv->pv_pmap ||
2162 kpmap == npv->pv_pmap) {
2163 if (entries++ == 0)
2164 pv = npv;
2165 /* Cacheable mappings */
2166 if ((npv->pv_flags & PT_NC) == 0) {
2167 cacheable_entries++;
2168 if (kpmap == npv->pv_pmap)
2169 kern_cacheable++;
2170 }
2171 /* Writable mappings */
2172 if (npv->pv_flags & PT_Wr)
2173 ++writable;
2174 } else if (npv->pv_flags & PT_Wr)
2175 other_writable = 1;
2176 }
2177
2178 PDEBUG(3,printf("pmap_vac_me_harder: pmap %p Entries %d, "
2179 "writable %d cacheable %d %s\n", pmap, entries, writable,
2180 cacheable_entries, clear_cache ? "clean" : "no clean"));
2181
2182 /*
2183 * Enable or disable caching as necessary.
2184 * Note: the first entry might be part of the kernel pmap,
2185 * so we can't assume this is indicative of the state of the
2186 * other (maybe non-kpmap) entries.
2187 */
2188 if ((entries > 1 && writable) ||
2189 (entries > 0 && pmap == kpmap && other_writable)) {
2190 if (cacheable_entries == 0)
2191 return;
2192 for (npv = pv; npv; npv = npv->pv_next) {
2193 if ((pmap == npv->pv_pmap
2194 || kpmap == npv->pv_pmap) &&
2195 (npv->pv_flags & PT_NC) == 0) {
2196 ptes[arm_btop(npv->pv_va)] &= ~(PT_C | PT_B);
2197 npv->pv_flags |= PT_NC;
2198 /*
2199 * If this page needs flushing from the
2200 * cache, and we aren't going to do it
2201 * below, do it now.
2202 */
2203 if ((cacheable_entries < 4 &&
2204 (clear_cache || npv->pv_pmap == kpmap)) ||
2205 (npv->pv_pmap == kpmap &&
2206 !clear_cache && kern_cacheable < 4)) {
2207 cpu_idcache_wbinv_range(npv->pv_va,
2208 NBPG);
2209 cpu_tlb_flushID_SE(npv->pv_va);
2210 }
2211 }
2212 }
2213 if ((clear_cache && cacheable_entries >= 4) ||
2214 kern_cacheable >= 4) {
2215 cpu_idcache_wbinv_all();
2216 cpu_tlb_flushID();
2217 }
2218 cpu_cpwait();
2219 } else if (entries > 0) {
2220 /*
2221 * Turn cacheing back on for some pages. If it is a kernel
2222 * page, only do so if there are no other writable pages.
2223 */
2224 for (npv = pv; npv; npv = npv->pv_next) {
2225 if ((pmap == npv->pv_pmap ||
2226 (kpmap == npv->pv_pmap && other_writable == 0)) &&
2227 (npv->pv_flags & PT_NC)) {
2228 ptes[arm_btop(npv->pv_va)] |= pte_cache_mode;
2229 npv->pv_flags &= ~PT_NC;
2230 }
2231 }
2232 }
2233 }
2234
2235 /*
2236 * pmap_remove()
2237 *
2238 * pmap_remove is responsible for nuking a number of mappings for a range
2239 * of virtual address space in the current pmap. To do this efficiently
2240 * is interesting, because in a number of cases a wide virtual address
2241 * range may be supplied that contains few actual mappings. So, the
2242 * optimisations are:
2243 * 1. Try and skip over hunks of address space for which an L1 entry
2244 * does not exist.
2245 * 2. Build up a list of pages we've hit, up to a maximum, so we can
2246 * maybe do just a partial cache clean. This path of execution is
2247 * complicated by the fact that the cache must be flushed _before_
2248 * the PTE is nuked, being a VAC :-)
2249 * 3. Maybe later fast-case a single page, but I don't think this is
2250 * going to make _that_ much difference overall.
2251 */
2252
2253 #define PMAP_REMOVE_CLEAN_LIST_SIZE 3
2254
2255 void
2256 pmap_remove(pmap, sva, eva)
2257 struct pmap *pmap;
2258 vaddr_t sva;
2259 vaddr_t eva;
2260 {
2261 int cleanlist_idx = 0;
2262 struct pagelist {
2263 vaddr_t va;
2264 pt_entry_t *pte;
2265 } cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
2266 pt_entry_t *pte = 0, *ptes;
2267 paddr_t pa;
2268 int pmap_active;
2269 struct vm_page *pg;
2270
2271 /* Exit quick if there is no pmap */
2272 if (!pmap)
2273 return;
2274
2275 PDEBUG(0, printf("pmap_remove: pmap=%p sva=%08lx eva=%08lx\n", pmap, sva, eva));
2276
2277 sva &= PG_FRAME;
2278 eva &= PG_FRAME;
2279
2280 /*
2281 * we lock in the pmap => vm_page direction
2282 */
2283 PMAP_MAP_TO_HEAD_LOCK();
2284
2285 ptes = pmap_map_ptes(pmap);
2286 /* Get a page table pointer */
2287 while (sva < eva) {
2288 if (pmap_pde_page(pmap_pde(pmap, sva)))
2289 break;
2290 sva = (sva & PD_MASK) + NBPD;
2291 }
2292
2293 pte = &ptes[arm_btop(sva)];
2294 /* Note if the pmap is active thus require cache and tlb cleans */
2295 pmap_active = pmap_is_curpmap(pmap);
2296
2297 /* Now loop along */
2298 while (sva < eva) {
2299 /* Check if we can move to the next PDE (l1 chunk) */
2300 if (!(sva & PT_MASK))
2301 if (!pmap_pde_page(pmap_pde(pmap, sva))) {
2302 sva += NBPD;
2303 pte += arm_btop(NBPD);
2304 continue;
2305 }
2306
2307 /* We've found a valid PTE, so this page of PTEs has to go. */
2308 if (pmap_pte_v(pte)) {
2309 /* Update statistics */
2310 --pmap->pm_stats.resident_count;
2311
2312 /*
2313 * Add this page to our cache remove list, if we can.
2314 * If, however the cache remove list is totally full,
2315 * then do a complete cache invalidation taking note
2316 * to backtrack the PTE table beforehand, and ignore
2317 * the lists in future because there's no longer any
2318 * point in bothering with them (we've paid the
2319 * penalty, so will carry on unhindered). Otherwise,
2320 * when we fall out, we just clean the list.
2321 */
2322 PDEBUG(10, printf("remove: inv pte at %p(%x) ", pte, *pte));
2323 pa = pmap_pte_pa(pte);
2324
2325 if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
2326 /* Add to the clean list. */
2327 cleanlist[cleanlist_idx].pte = pte;
2328 cleanlist[cleanlist_idx].va = sva;
2329 cleanlist_idx++;
2330 } else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
2331 int cnt;
2332
2333 /* Nuke everything if needed. */
2334 if (pmap_active) {
2335 cpu_idcache_wbinv_all();
2336 cpu_tlb_flushID();
2337 }
2338
2339 /*
2340 * Roll back the previous PTE list,
2341 * and zero out the current PTE.
2342 */
2343 for (cnt = 0; cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
2344 *cleanlist[cnt].pte = 0;
2345 pmap_pte_delref(pmap, cleanlist[cnt].va);
2346 }
2347 *pte = 0;
2348 pmap_pte_delref(pmap, sva);
2349 cleanlist_idx++;
2350 } else {
2351 /*
2352 * We've already nuked the cache and
2353 * TLB, so just carry on regardless,
2354 * and we won't need to do it again
2355 */
2356 *pte = 0;
2357 pmap_pte_delref(pmap, sva);
2358 }
2359
2360 /*
2361 * Update flags. In a number of circumstances,
2362 * we could cluster a lot of these and do a
2363 * number of sequential pages in one go.
2364 */
2365 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
2366 struct pv_entry *pve;
2367 simple_lock(&pg->mdpage.pvh_slock);
2368 pve = pmap_remove_pv(pg, pmap, sva);
2369 pmap_free_pv(pmap, pve);
2370 pmap_vac_me_harder(pmap, pg, ptes, FALSE);
2371 simple_unlock(&pg->mdpage.pvh_slock);
2372 }
2373 }
2374 sva += NBPG;
2375 pte++;
2376 }
2377
2378 pmap_unmap_ptes(pmap);
2379 /*
2380 * Now, if we've fallen through down to here, chances are that there
2381 * are less than PMAP_REMOVE_CLEAN_LIST_SIZE mappings left.
2382 */
2383 if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
2384 u_int cnt;
2385
2386 for (cnt = 0; cnt < cleanlist_idx; cnt++) {
2387 if (pmap_active) {
2388 cpu_idcache_wbinv_range(cleanlist[cnt].va,
2389 NBPG);
2390 *cleanlist[cnt].pte = 0;
2391 cpu_tlb_flushID_SE(cleanlist[cnt].va);
2392 } else
2393 *cleanlist[cnt].pte = 0;
2394 pmap_pte_delref(pmap, cleanlist[cnt].va);
2395 }
2396 }
2397 PMAP_MAP_TO_HEAD_UNLOCK();
2398 }
2399
2400 /*
2401 * Routine: pmap_remove_all
2402 * Function:
2403 * Removes this physical page from
2404 * all physical maps in which it resides.
2405 * Reflects back modify bits to the pager.
2406 */
2407
2408 static void
2409 pmap_remove_all(pg)
2410 struct vm_page *pg;
2411 {
2412 struct pv_entry *pv, *npv;
2413 struct pmap *pmap;
2414 pt_entry_t *pte, *ptes;
2415
2416 PDEBUG(0, printf("pmap_remove_all: pa=%lx ", VM_PAGE_TO_PHYS(pg)));
2417
2418 /* set vm_page => pmap locking */
2419 PMAP_HEAD_TO_MAP_LOCK();
2420
2421 simple_lock(&pg->mdpage.pvh_slock);
2422
2423 pv = pg->mdpage.pvh_list;
2424 if (pv == NULL) {
2425 PDEBUG(0, printf("free page\n"));
2426 simple_unlock(&pg->mdpage.pvh_slock);
2427 PMAP_HEAD_TO_MAP_UNLOCK();
2428 return;
2429 }
2430 pmap_clean_page(pv, FALSE);
2431
2432 while (pv) {
2433 pmap = pv->pv_pmap;
2434 ptes = pmap_map_ptes(pmap);
2435 pte = &ptes[arm_btop(pv->pv_va)];
2436
2437 PDEBUG(0, printf("[%p,%08x,%08lx,%08x] ", pmap, *pte,
2438 pv->pv_va, pv->pv_flags));
2439 #ifdef DEBUG
2440 if (!pmap_pde_page(pmap_pde(pmap, pv->pv_va)) ||
2441 !pmap_pte_v(pte) || pmap_pte_pa(pte) != pa)
2442 panic("pmap_remove_all: bad mapping");
2443 #endif /* DEBUG */
2444
2445 /*
2446 * Update statistics
2447 */
2448 --pmap->pm_stats.resident_count;
2449
2450 /* Wired bit */
2451 if (pv->pv_flags & PT_W)
2452 --pmap->pm_stats.wired_count;
2453
2454 /*
2455 * Invalidate the PTEs.
2456 * XXX: should cluster them up and invalidate as many
2457 * as possible at once.
2458 */
2459
2460 #ifdef needednotdone
2461 reduce wiring count on page table pages as references drop
2462 #endif
2463
2464 *pte = 0;
2465 pmap_pte_delref(pmap, pv->pv_va);
2466
2467 npv = pv->pv_next;
2468 pmap_free_pv(pmap, pv);
2469 pv = npv;
2470 pmap_unmap_ptes(pmap);
2471 }
2472 pg->mdpage.pvh_list = NULL;
2473 simple_unlock(&pg->mdpage.pvh_slock);
2474 PMAP_HEAD_TO_MAP_UNLOCK();
2475
2476 PDEBUG(0, printf("done\n"));
2477 cpu_tlb_flushID();
2478 cpu_cpwait();
2479 }
2480
2481
2482 /*
2483 * Set the physical protection on the specified range of this map as requested.
2484 */
2485
2486 void
2487 pmap_protect(pmap, sva, eva, prot)
2488 struct pmap *pmap;
2489 vaddr_t sva;
2490 vaddr_t eva;
2491 vm_prot_t prot;
2492 {
2493 pt_entry_t *pte = NULL, *ptes;
2494 struct vm_page *pg;
2495 int armprot;
2496 int flush = 0;
2497 paddr_t pa;
2498
2499 PDEBUG(0, printf("pmap_protect: pmap=%p %08lx->%08lx %x\n",
2500 pmap, sva, eva, prot));
2501
2502 if (~prot & VM_PROT_READ) {
2503 /* Just remove the mappings. */
2504 pmap_remove(pmap, sva, eva);
2505 /* pmap_update not needed as it should be called by the caller
2506 * of pmap_protect */
2507 return;
2508 }
2509 if (prot & VM_PROT_WRITE) {
2510 /*
2511 * If this is a read->write transition, just ignore it and let
2512 * uvm_fault() take care of it later.
2513 */
2514 return;
2515 }
2516
2517 sva &= PG_FRAME;
2518 eva &= PG_FRAME;
2519
2520 /* Need to lock map->head */
2521 PMAP_MAP_TO_HEAD_LOCK();
2522
2523 ptes = pmap_map_ptes(pmap);
2524 /*
2525 * We need to acquire a pointer to a page table page before entering
2526 * the following loop.
2527 */
2528 while (sva < eva) {
2529 if (pmap_pde_page(pmap_pde(pmap, sva)))
2530 break;
2531 sva = (sva & PD_MASK) + NBPD;
2532 }
2533
2534 pte = &ptes[arm_btop(sva)];
2535
2536 while (sva < eva) {
2537 /* only check once in a while */
2538 if ((sva & PT_MASK) == 0) {
2539 if (!pmap_pde_page(pmap_pde(pmap, sva))) {
2540 /* We can race ahead here, to the next pde. */
2541 sva += NBPD;
2542 pte += arm_btop(NBPD);
2543 continue;
2544 }
2545 }
2546
2547 if (!pmap_pte_v(pte))
2548 goto next;
2549
2550 flush = 1;
2551
2552 armprot = 0;
2553 if (sva < VM_MAXUSER_ADDRESS)
2554 armprot |= PT_AP(AP_U);
2555 else if (sva < VM_MAX_ADDRESS)
2556 armprot |= PT_AP(AP_W); /* XXX Ekk what is this ? */
2557 *pte = (*pte & 0xfffff00f) | armprot;
2558
2559 pa = pmap_pte_pa(pte);
2560
2561 /* Get the physical page index */
2562
2563 /* Clear write flag */
2564 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
2565 simple_lock(&pg->mdpage.pvh_slock);
2566 (void) pmap_modify_pv(pmap, sva, pg, PT_Wr, 0);
2567 pmap_vac_me_harder(pmap, pg, ptes, FALSE);
2568 simple_unlock(&pg->mdpage.pvh_slock);
2569 }
2570
2571 next:
2572 sva += NBPG;
2573 pte++;
2574 }
2575 pmap_unmap_ptes(pmap);
2576 PMAP_MAP_TO_HEAD_UNLOCK();
2577 if (flush)
2578 cpu_tlb_flushID();
2579 }
2580
2581 /*
2582 * void pmap_enter(struct pmap *pmap, vaddr_t va, paddr_t pa, vm_prot_t prot,
2583 * int flags)
2584 *
2585 * Insert the given physical page (p) at
2586 * the specified virtual address (v) in the
2587 * target physical map with the protection requested.
2588 *
2589 * If specified, the page will be wired down, meaning
2590 * that the related pte can not be reclaimed.
2591 *
2592 * NB: This is the only routine which MAY NOT lazy-evaluate
2593 * or lose information. That is, this routine must actually
2594 * insert this page into the given map NOW.
2595 */
2596
2597 int
2598 pmap_enter(pmap, va, pa, prot, flags)
2599 struct pmap *pmap;
2600 vaddr_t va;
2601 paddr_t pa;
2602 vm_prot_t prot;
2603 int flags;
2604 {
2605 pt_entry_t *ptes, opte, npte;
2606 paddr_t opa;
2607 boolean_t wired = (flags & PMAP_WIRED) != 0;
2608 struct vm_page *pg;
2609 struct pv_entry *pve;
2610 int error, nflags;
2611
2612 PDEBUG(5, printf("pmap_enter: V%08lx P%08lx in pmap %p prot=%08x, wired = %d\n",
2613 va, pa, pmap, prot, wired));
2614
2615 #ifdef DIAGNOSTIC
2616 /* Valid address ? */
2617 if (va >= (pmap_curmaxkvaddr))
2618 panic("pmap_enter: too big");
2619 if (pmap != pmap_kernel() && va != 0) {
2620 if (va < VM_MIN_ADDRESS || va >= VM_MAXUSER_ADDRESS)
2621 panic("pmap_enter: kernel page in user map");
2622 } else {
2623 if (va >= VM_MIN_ADDRESS && va < VM_MAXUSER_ADDRESS)
2624 panic("pmap_enter: user page in kernel map");
2625 if (va >= VM_MAXUSER_ADDRESS && va < VM_MAX_ADDRESS)
2626 panic("pmap_enter: entering PT page");
2627 }
2628 #endif
2629 /*
2630 * Get a pointer to the page. Later on in this function, we
2631 * test for a managed page by checking pg != NULL.
2632 */
2633 pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
2634
2635 /* get lock */
2636 PMAP_MAP_TO_HEAD_LOCK();
2637
2638 /*
2639 * map the ptes. If there's not already an L2 table for this
2640 * address, allocate one.
2641 */
2642 ptes = pmap_map_ptes(pmap); /* locks pmap */
2643 if (pmap_pde_v(pmap_pde(pmap, va)) == 0) {
2644 struct vm_page *ptp;
2645
2646 /* kernel should be pre-grown */
2647 KASSERT(pmap != pmap_kernel());
2648
2649 /* if failure is allowed then don't try too hard */
2650 ptp = pmap_get_ptp(pmap, va & PD_MASK);
2651 if (ptp == NULL) {
2652 if (flags & PMAP_CANFAIL) {
2653 error = ENOMEM;
2654 goto out;
2655 }
2656 panic("pmap_enter: get ptp failed");
2657 }
2658 }
2659 opte = ptes[arm_btop(va)];
2660
2661 nflags = 0;
2662 if (prot & VM_PROT_WRITE)
2663 nflags |= PT_Wr;
2664 if (wired)
2665 nflags |= PT_W;
2666
2667 /* Is the pte valid ? If so then this page is already mapped */
2668 if (l2pte_valid(opte)) {
2669 /* Get the physical address of the current page mapped */
2670 opa = l2pte_pa(opte);
2671
2672 /* Are we mapping the same page ? */
2673 if (opa == pa) {
2674 /* Has the wiring changed ? */
2675 if (pg != NULL) {
2676 simple_lock(&pg->mdpage.pvh_slock);
2677 (void) pmap_modify_pv(pmap, va, pg,
2678 PT_Wr | PT_W, nflags);
2679 simple_unlock(&pg->mdpage.pvh_slock);
2680 }
2681 } else {
2682 struct vm_page *opg;
2683
2684 /* We are replacing the page with a new one. */
2685 cpu_idcache_wbinv_range(va, NBPG);
2686
2687 /*
2688 * If it is part of our managed memory then we
2689 * must remove it from the PV list
2690 */
2691 if ((opg = PHYS_TO_VM_PAGE(opa)) != NULL) {
2692 simple_lock(&opg->mdpage.pvh_slock);
2693 pve = pmap_remove_pv(opg, pmap, va);
2694 simple_unlock(&opg->mdpage.pvh_slock);
2695 } else {
2696 pve = NULL;
2697 }
2698
2699 goto enter;
2700 }
2701 } else {
2702 opa = 0;
2703 pve = NULL;
2704 pmap_pte_addref(pmap, va);
2705
2706 /* pte is not valid so we must be hooking in a new page */
2707 ++pmap->pm_stats.resident_count;
2708
2709 enter:
2710 /*
2711 * Enter on the PV list if part of our managed memory
2712 */
2713 if (pg != NULL) {
2714 if (pve == NULL) {
2715 pve = pmap_alloc_pv(pmap, ALLOCPV_NEED);
2716 if (pve == NULL) {
2717 if (flags & PMAP_CANFAIL) {
2718 error = ENOMEM;
2719 goto out;
2720 }
2721 panic("pmap_enter: no pv entries "
2722 "available");
2723 }
2724 }
2725 /* enter_pv locks pvh when adding */
2726 pmap_enter_pv(pg, pve, pmap, va, NULL, nflags);
2727 } else {
2728 if (pve != NULL)
2729 pmap_free_pv(pmap, pve);
2730 }
2731 }
2732
2733 /* Construct the pte, giving the correct access. */
2734 npte = (pa & PG_FRAME);
2735
2736 /* VA 0 is magic. */
2737 if (pmap != pmap_kernel() && va != 0)
2738 npte |= PT_AP(AP_U);
2739
2740 if (pg != NULL) {
2741 #ifdef DIAGNOSTIC
2742 if ((flags & VM_PROT_ALL) & ~prot)
2743 panic("pmap_enter: access_type exceeds prot");
2744 #endif
2745 npte |= pte_cache_mode;
2746 if (flags & VM_PROT_WRITE) {
2747 npte |= L2_SPAGE | PT_AP(AP_W);
2748 pg->mdpage.pvh_attrs |= PT_H | PT_M;
2749 } else if (flags & VM_PROT_ALL) {
2750 npte |= L2_SPAGE;
2751 pg->mdpage.pvh_attrs |= PT_H;
2752 } else
2753 npte |= L2_INVAL;
2754 } else {
2755 if (prot & VM_PROT_WRITE)
2756 npte |= L2_SPAGE | PT_AP(AP_W);
2757 else if (prot & VM_PROT_ALL)
2758 npte |= L2_SPAGE;
2759 else
2760 npte |= L2_INVAL;
2761 }
2762
2763 ptes[arm_btop(va)] = npte;
2764
2765 if (pg != NULL) {
2766 simple_lock(&pg->mdpage.pvh_slock);
2767 pmap_vac_me_harder(pmap, pg, ptes, pmap_is_curpmap(pmap));
2768 simple_unlock(&pg->mdpage.pvh_slock);
2769 }
2770
2771 /* Better flush the TLB ... */
2772 cpu_tlb_flushID_SE(va);
2773 error = 0;
2774 out:
2775 pmap_unmap_ptes(pmap); /* unlocks pmap */
2776 PMAP_MAP_TO_HEAD_UNLOCK();
2777
2778 return error;
2779 }
2780
2781 /*
2782 * pmap_kenter_pa: enter a kernel mapping
2783 *
2784 * => no need to lock anything assume va is already allocated
2785 * => should be faster than normal pmap enter function
2786 */
2787 void
2788 pmap_kenter_pa(va, pa, prot)
2789 vaddr_t va;
2790 paddr_t pa;
2791 vm_prot_t prot;
2792 {
2793 pt_entry_t *pte;
2794
2795 pte = vtopte(va);
2796 KASSERT(!pmap_pte_v(pte));
2797 *pte = L2_PTE(pa, AP_KRW);
2798 }
2799
2800 void
2801 pmap_kremove(va, len)
2802 vaddr_t va;
2803 vsize_t len;
2804 {
2805 pt_entry_t *pte;
2806
2807 for (len >>= PAGE_SHIFT; len > 0; len--, va += PAGE_SIZE) {
2808
2809 /*
2810 * We assume that we will only be called with small
2811 * regions of memory.
2812 */
2813
2814 KASSERT(pmap_pde_page(pmap_pde(pmap_kernel(), va)));
2815 pte = vtopte(va);
2816 cpu_idcache_wbinv_range(va, PAGE_SIZE);
2817 *pte = 0;
2818 cpu_tlb_flushID_SE(va);
2819 }
2820 }
2821
2822 /*
2823 * pmap_page_protect:
2824 *
2825 * Lower the permission for all mappings to a given page.
2826 */
2827
2828 void
2829 pmap_page_protect(pg, prot)
2830 struct vm_page *pg;
2831 vm_prot_t prot;
2832 {
2833
2834 PDEBUG(0, printf("pmap_page_protect(pa=%lx, prot=%d)\n",
2835 VM_PAGE_TO_PHYS(pg), prot));
2836
2837 switch(prot) {
2838 case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
2839 case VM_PROT_READ|VM_PROT_WRITE:
2840 return;
2841
2842 case VM_PROT_READ:
2843 case VM_PROT_READ|VM_PROT_EXECUTE:
2844 pmap_copy_on_write(pg);
2845 break;
2846
2847 default:
2848 pmap_remove_all(pg);
2849 break;
2850 }
2851 }
2852
2853
2854 /*
2855 * Routine: pmap_unwire
2856 * Function: Clear the wired attribute for a map/virtual-address
2857 * pair.
2858 * In/out conditions:
2859 * The mapping must already exist in the pmap.
2860 */
2861
2862 void
2863 pmap_unwire(pmap, va)
2864 struct pmap *pmap;
2865 vaddr_t va;
2866 {
2867 pt_entry_t *ptes;
2868 struct vm_page *pg;
2869 paddr_t pa;
2870
2871 PMAP_MAP_TO_HEAD_LOCK();
2872 ptes = pmap_map_ptes(pmap); /* locks pmap */
2873
2874 if (pmap_pde_v(pmap_pde(pmap, va))) {
2875 #ifdef DIAGNOSTIC
2876 if (l2pte_valid(ptes[arm_btop(va)]) == 0)
2877 panic("pmap_unwire: invalid L2 PTE");
2878 #endif
2879 /* Extract the physical address of the page */
2880 pa = l2pte_pa(ptes[arm_btop(va)]);
2881
2882 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
2883 goto out;
2884
2885 /* Update the wired bit in the pv entry for this page. */
2886 simple_lock(&pg->mdpage.pvh_slock);
2887 (void) pmap_modify_pv(pmap, va, pg, PT_W, 0);
2888 simple_unlock(&pg->mdpage.pvh_slock);
2889 }
2890 #ifdef DIAGNOSTIC
2891 else {
2892 panic("pmap_unwire: invalid L1 PTE");
2893 }
2894 #endif
2895 out:
2896 pmap_unmap_ptes(pmap); /* unlocks pmap */
2897 PMAP_MAP_TO_HEAD_UNLOCK();
2898 }
2899
2900 /*
2901 * Routine: pmap_extract
2902 * Function:
2903 * Extract the physical page address associated
2904 * with the given map/virtual_address pair.
2905 */
2906 boolean_t
2907 pmap_extract(pmap, va, pap)
2908 struct pmap *pmap;
2909 vaddr_t va;
2910 paddr_t *pap;
2911 {
2912 pd_entry_t *pde;
2913 pt_entry_t *pte, *ptes;
2914 paddr_t pa;
2915 boolean_t rv = TRUE;
2916
2917 PDEBUG(5, printf("pmap_extract: pmap=%p, va=V%08lx\n", pmap, va));
2918
2919 /*
2920 * Get the pte for this virtual address.
2921 */
2922 pde = pmap_pde(pmap, va);
2923 ptes = pmap_map_ptes(pmap);
2924 pte = &ptes[arm_btop(va)];
2925
2926 if (pmap_pde_section(pde)) {
2927 pa = (*pde & PD_MASK) | (va & (L1_SEC_SIZE - 1));
2928 goto out;
2929 } else if (pmap_pde_page(pde) == 0 || pmap_pte_v(pte) == 0) {
2930 rv = FALSE;
2931 goto out;
2932 }
2933
2934 if ((*pte & L2_MASK) == L2_LPAGE) {
2935 /* Extract the physical address from the pte */
2936 pa = *pte & ~(L2_LPAGE_SIZE - 1);
2937
2938 PDEBUG(5, printf("pmap_extract: LPAGE pa = P%08lx\n",
2939 (pa | (va & (L2_LPAGE_SIZE - 1)))));
2940
2941 if (pap != NULL)
2942 *pap = pa | (va & (L2_LPAGE_SIZE - 1));
2943 goto out;
2944 }
2945
2946 /* Extract the physical address from the pte */
2947 pa = pmap_pte_pa(pte);
2948
2949 PDEBUG(5, printf("pmap_extract: SPAGE pa = P%08lx\n",
2950 (pa | (va & ~PG_FRAME))));
2951
2952 if (pap != NULL)
2953 *pap = pa | (va & ~PG_FRAME);
2954 out:
2955 pmap_unmap_ptes(pmap);
2956 return (rv);
2957 }
2958
2959
2960 /*
2961 * Copy the range specified by src_addr/len from the source map to the
2962 * range dst_addr/len in the destination map.
2963 *
2964 * This routine is only advisory and need not do anything.
2965 */
2966
2967 void
2968 pmap_copy(dst_pmap, src_pmap, dst_addr, len, src_addr)
2969 struct pmap *dst_pmap;
2970 struct pmap *src_pmap;
2971 vaddr_t dst_addr;
2972 vsize_t len;
2973 vaddr_t src_addr;
2974 {
2975 PDEBUG(0, printf("pmap_copy(%p, %p, %lx, %lx, %lx)\n",
2976 dst_pmap, src_pmap, dst_addr, len, src_addr));
2977 }
2978
2979 #if defined(PMAP_DEBUG)
2980 void
2981 pmap_dump_pvlist(phys, m)
2982 vaddr_t phys;
2983 char *m;
2984 {
2985 struct vm_page *pg;
2986 struct pv_entry *pv;
2987
2988 if ((pg = PHYS_TO_VM_PAGE(phys)) == NULL) {
2989 printf("INVALID PA\n");
2990 return;
2991 }
2992 simple_lock(&pg->mdpage.pvh_slock);
2993 printf("%s %08lx:", m, phys);
2994 if (pg->mdpage.pvh_list == NULL) {
2995 printf(" no mappings\n");
2996 return;
2997 }
2998
2999 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next)
3000 printf(" pmap %p va %08lx flags %08x", pv->pv_pmap,
3001 pv->pv_va, pv->pv_flags);
3002
3003 printf("\n");
3004 simple_unlock(&pg->mdpage.pvh_slock);
3005 }
3006
3007 #endif /* PMAP_DEBUG */
3008
3009 static pt_entry_t *
3010 pmap_map_ptes(struct pmap *pmap)
3011 {
3012 struct proc *p;
3013
3014 /* the kernel's pmap is always accessible */
3015 if (pmap == pmap_kernel()) {
3016 return (pt_entry_t *)PTE_BASE ;
3017 }
3018
3019 if (pmap_is_curpmap(pmap)) {
3020 simple_lock(&pmap->pm_obj.vmobjlock);
3021 return (pt_entry_t *)PTE_BASE;
3022 }
3023
3024 p = curproc;
3025
3026 if (p == NULL)
3027 p = &proc0;
3028
3029 /* need to lock both curpmap and pmap: use ordered locking */
3030 if ((unsigned) pmap < (unsigned) curproc->p_vmspace->vm_map.pmap) {
3031 simple_lock(&pmap->pm_obj.vmobjlock);
3032 simple_lock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3033 } else {
3034 simple_lock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3035 simple_lock(&pmap->pm_obj.vmobjlock);
3036 }
3037
3038 pmap_map_in_l1(p->p_vmspace->vm_map.pmap, APTE_BASE,
3039 pmap->pm_pptpt, FALSE);
3040 cpu_tlb_flushD();
3041 cpu_cpwait();
3042 return (pt_entry_t *)APTE_BASE;
3043 }
3044
3045 /*
3046 * pmap_unmap_ptes: unlock the PTE mapping of "pmap"
3047 */
3048
3049 static void
3050 pmap_unmap_ptes(pmap)
3051 struct pmap *pmap;
3052 {
3053 if (pmap == pmap_kernel()) {
3054 return;
3055 }
3056 if (pmap_is_curpmap(pmap)) {
3057 simple_unlock(&pmap->pm_obj.vmobjlock);
3058 } else {
3059 simple_unlock(&pmap->pm_obj.vmobjlock);
3060 simple_unlock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock);
3061 }
3062 }
3063
3064 /*
3065 * Modify pte bits for all ptes corresponding to the given physical address.
3066 * We use `maskbits' rather than `clearbits' because we're always passing
3067 * constants and the latter would require an extra inversion at run-time.
3068 */
3069
3070 static void
3071 pmap_clearbit(pg, maskbits)
3072 struct vm_page *pg;
3073 unsigned int maskbits;
3074 {
3075 struct pv_entry *pv;
3076 pt_entry_t *ptes;
3077 vaddr_t va;
3078 int tlbentry;
3079
3080 PDEBUG(1, printf("pmap_clearbit: pa=%08lx mask=%08x\n",
3081 VM_PAGE_TO_PHYS(pg), maskbits));
3082
3083 tlbentry = 0;
3084
3085 PMAP_HEAD_TO_MAP_LOCK();
3086 simple_lock(&pg->mdpage.pvh_slock);
3087
3088 /*
3089 * Clear saved attributes (modify, reference)
3090 */
3091 pg->mdpage.pvh_attrs &= ~maskbits;
3092
3093 if (pg->mdpage.pvh_list == NULL) {
3094 simple_unlock(&pg->mdpage.pvh_slock);
3095 PMAP_HEAD_TO_MAP_UNLOCK();
3096 return;
3097 }
3098
3099 /*
3100 * Loop over all current mappings setting/clearing as appropos
3101 */
3102 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
3103 va = pv->pv_va;
3104 pv->pv_flags &= ~maskbits;
3105 ptes = pmap_map_ptes(pv->pv_pmap); /* locks pmap */
3106 KASSERT(pmap_pde_v(pmap_pde(pv->pv_pmap, va)));
3107 if (maskbits & (PT_Wr|PT_M)) {
3108 if ((pv->pv_flags & PT_NC)) {
3109 /*
3110 * Entry is not cacheable: reenable
3111 * the cache, nothing to flush
3112 *
3113 * Don't turn caching on again if this
3114 * is a modified emulation. This
3115 * would be inconsitent with the
3116 * settings created by
3117 * pmap_vac_me_harder().
3118 *
3119 * There's no need to call
3120 * pmap_vac_me_harder() here: all
3121 * pages are loosing their write
3122 * permission.
3123 *
3124 */
3125 if (maskbits & PT_Wr) {
3126 ptes[arm_btop(va)] |= pte_cache_mode;
3127 pv->pv_flags &= ~PT_NC;
3128 }
3129 } else if (pmap_is_curpmap(pv->pv_pmap)) {
3130 /*
3131 * Entry is cacheable: check if pmap is
3132 * current if it is flush it,
3133 * otherwise it won't be in the cache
3134 */
3135 cpu_idcache_wbinv_range(pv->pv_va, NBPG);
3136 }
3137
3138 /* make the pte read only */
3139 ptes[arm_btop(va)] &= ~PT_AP(AP_W);
3140 }
3141
3142 if (maskbits & PT_H)
3143 ptes[arm_btop(va)] =
3144 (ptes[arm_btop(va)] & ~L2_MASK) | L2_INVAL;
3145
3146 if (pmap_is_curpmap(pv->pv_pmap)) {
3147 /*
3148 * if we had cacheable pte's we'd clean the
3149 * pte out to memory here
3150 *
3151 * flush tlb entry as it's in the current pmap
3152 */
3153 cpu_tlb_flushID_SE(pv->pv_va);
3154 }
3155 pmap_unmap_ptes(pv->pv_pmap); /* unlocks pmap */
3156 }
3157 cpu_cpwait();
3158
3159 simple_unlock(&pg->mdpage.pvh_slock);
3160 PMAP_HEAD_TO_MAP_UNLOCK();
3161 }
3162
3163 /*
3164 * pmap_clear_modify:
3165 *
3166 * Clear the "modified" attribute for a page.
3167 */
3168 boolean_t
3169 pmap_clear_modify(pg)
3170 struct vm_page *pg;
3171 {
3172 boolean_t rv;
3173
3174 if (pg->mdpage.pvh_attrs & PT_M) {
3175 rv = TRUE;
3176 pmap_clearbit(pg, PT_M);
3177 } else
3178 rv = FALSE;
3179
3180 PDEBUG(0, printf("pmap_clear_modify pa=%08lx -> %d\n",
3181 VM_PAGE_TO_PHYS(pg), rv));
3182
3183 return (rv);
3184 }
3185
3186 /*
3187 * pmap_clear_reference:
3188 *
3189 * Clear the "referenced" attribute for a page.
3190 */
3191 boolean_t
3192 pmap_clear_reference(pg)
3193 struct vm_page *pg;
3194 {
3195 boolean_t rv;
3196
3197 if (pg->mdpage.pvh_attrs & PT_H) {
3198 rv = TRUE;
3199 pmap_clearbit(pg, PT_H);
3200 } else
3201 rv = FALSE;
3202
3203 PDEBUG(0, printf("pmap_clear_reference pa=%08lx -> %d\n",
3204 VM_PAGE_TO_PHYS(pg), rv));
3205
3206 return (rv);
3207 }
3208
3209
3210 void
3211 pmap_copy_on_write(pg)
3212 struct vm_page *pg;
3213 {
3214 PDEBUG(0, printf("pmap_copy_on_write pa=%08lx\n", VM_PAGE_TO_PHYS(pg)));
3215 pmap_clearbit(pg, PT_Wr);
3216 }
3217
3218 /*
3219 * pmap_is_modified:
3220 *
3221 * Test if a page has the "modified" attribute.
3222 */
3223 /* See <arm/arm32/pmap.h> */
3224
3225 /*
3226 * pmap_is_referenced:
3227 *
3228 * Test if a page has the "referenced" attribute.
3229 */
3230 /* See <arm/arm32/pmap.h> */
3231
3232 int
3233 pmap_modified_emulation(pmap, va)
3234 struct pmap *pmap;
3235 vaddr_t va;
3236 {
3237 pt_entry_t *ptes;
3238 struct vm_page *pg;
3239 paddr_t pa;
3240 u_int flags;
3241 int rv = 0;
3242
3243 PDEBUG(2, printf("pmap_modified_emulation\n"));
3244
3245 PMAP_MAP_TO_HEAD_LOCK();
3246 ptes = pmap_map_ptes(pmap); /* locks pmap */
3247
3248 if (pmap_pde_v(pmap_pde(pmap, va)) == 0) {
3249 PDEBUG(2, printf("L1 PTE invalid\n"));
3250 goto out;
3251 }
3252
3253 PDEBUG(1, printf("pte=%08x\n", ptes[arm_btop(va)]));
3254
3255 /* Check for a invalid pte */
3256 if (l2pte_valid(ptes[arm_btop(va)]) == 0)
3257 goto out;
3258
3259 /* This can happen if user code tries to access kernel memory. */
3260 if ((ptes[arm_btop(va)] & PT_AP(AP_W)) != 0)
3261 goto out;
3262
3263 /* Extract the physical address of the page */
3264 pa = l2pte_pa(ptes[arm_btop(va)]);
3265 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3266 goto out;
3267
3268 /* Get the current flags for this page. */
3269 simple_lock(&pg->mdpage.pvh_slock);
3270
3271 flags = pmap_modify_pv(pmap, va, pg, 0, 0);
3272 PDEBUG(2, printf("pmap_modified_emulation: flags = %08x\n", flags));
3273
3274 /*
3275 * Do the flags say this page is writable ? If not then it is a
3276 * genuine write fault. If yes then the write fault is our fault
3277 * as we did not reflect the write access in the PTE. Now we know
3278 * a write has occurred we can correct this and also set the
3279 * modified bit
3280 */
3281 if (~flags & PT_Wr) {
3282 simple_unlock(&pg->mdpage.pvh_slock);
3283 goto out;
3284 }
3285
3286 PDEBUG(0,
3287 printf("pmap_modified_emulation: Got a hit va=%08lx, pte = %08x\n",
3288 va, ptes[arm_btop(va)]));
3289 pg->mdpage.pvh_attrs |= PT_H | PT_M;
3290
3291 /*
3292 * Re-enable write permissions for the page. No need to call
3293 * pmap_vac_me_harder(), since this is just a
3294 * modified-emulation fault, and the PT_Wr bit isn't changing. We've
3295 * already set the cacheable bits based on the assumption that we
3296 * can write to this page.
3297 */
3298 ptes[arm_btop(va)] =
3299 (ptes[arm_btop(va)] & ~L2_MASK) | L2_SPAGE | PT_AP(AP_W);
3300 PDEBUG(0, printf("->(%08x)\n", ptes[arm_btop(va)]));
3301
3302 simple_unlock(&pg->mdpage.pvh_slock);
3303
3304 cpu_tlb_flushID_SE(va);
3305 cpu_cpwait();
3306 rv = 1;
3307 out:
3308 pmap_unmap_ptes(pmap); /* unlocks pmap */
3309 PMAP_MAP_TO_HEAD_UNLOCK();
3310 return (rv);
3311 }
3312
3313 int
3314 pmap_handled_emulation(pmap, va)
3315 struct pmap *pmap;
3316 vaddr_t va;
3317 {
3318 pt_entry_t *ptes;
3319 struct vm_page *pg;
3320 paddr_t pa;
3321 int rv = 0;
3322
3323 PDEBUG(2, printf("pmap_handled_emulation\n"));
3324
3325 PMAP_MAP_TO_HEAD_LOCK();
3326 ptes = pmap_map_ptes(pmap); /* locks pmap */
3327
3328 if (pmap_pde_v(pmap_pde(pmap, va)) == 0) {
3329 PDEBUG(2, printf("L1 PTE invalid\n"));
3330 goto out;
3331 }
3332
3333 PDEBUG(1, printf("pte=%08x\n", ptes[arm_btop(va)]));
3334
3335 /* Check for invalid pte */
3336 if (l2pte_valid(ptes[arm_btop(va)]) == 0)
3337 goto out;
3338
3339 /* This can happen if user code tries to access kernel memory. */
3340 if ((ptes[arm_btop(va)] & L2_MASK) != L2_INVAL)
3341 goto out;
3342
3343 /* Extract the physical address of the page */
3344 pa = l2pte_pa(ptes[arm_btop(va)]);
3345 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3346 goto out;
3347
3348 simple_lock(&pg->mdpage.pvh_slock);
3349
3350 /*
3351 * Ok we just enable the pte and mark the attibs as handled
3352 * XXX Should we traverse the PV list and enable all PTEs?
3353 */
3354 PDEBUG(0,
3355 printf("pmap_handled_emulation: Got a hit va=%08lx pte = %08x\n",
3356 va, ptes[arm_btop(va)]));
3357 pg->mdpage.pvh_attrs |= PT_H;
3358
3359 ptes[arm_btop(va)] = (ptes[arm_btop(va)] & ~L2_MASK) | L2_SPAGE;
3360 PDEBUG(0, printf("->(%08x)\n", ptes[arm_btop(va)]));
3361
3362 simple_unlock(&pg->mdpage.pvh_slock);
3363
3364 cpu_tlb_flushID_SE(va);
3365 cpu_cpwait();
3366 rv = 1;
3367 out:
3368 pmap_unmap_ptes(pmap); /* unlocks pmap */
3369 PMAP_MAP_TO_HEAD_UNLOCK();
3370 return (rv);
3371 }
3372
3373 /*
3374 * pmap_collect: free resources held by a pmap
3375 *
3376 * => optional function.
3377 * => called when a process is swapped out to free memory.
3378 */
3379
3380 void
3381 pmap_collect(pmap)
3382 struct pmap *pmap;
3383 {
3384 }
3385
3386 /*
3387 * Routine: pmap_procwr
3388 *
3389 * Function:
3390 * Synchronize caches corresponding to [addr, addr+len) in p.
3391 *
3392 */
3393 void
3394 pmap_procwr(p, va, len)
3395 struct proc *p;
3396 vaddr_t va;
3397 int len;
3398 {
3399 /* We only need to do anything if it is the current process. */
3400 if (p == curproc)
3401 cpu_icache_sync_range(va, len);
3402 }
3403 /*
3404 * PTP functions
3405 */
3406
3407 /*
3408 * pmap_get_ptp: get a PTP (if there isn't one, allocate a new one)
3409 *
3410 * => pmap should NOT be pmap_kernel()
3411 * => pmap should be locked
3412 */
3413
3414 static struct vm_page *
3415 pmap_get_ptp(struct pmap *pmap, vaddr_t va)
3416 {
3417 struct vm_page *ptp;
3418
3419 if (pmap_pde_page(pmap_pde(pmap, va))) {
3420
3421 /* valid... check hint (saves us a PA->PG lookup) */
3422 if (pmap->pm_ptphint &&
3423 (pmap->pm_pdir[pmap_pdei(va)] & PG_FRAME) ==
3424 VM_PAGE_TO_PHYS(pmap->pm_ptphint))
3425 return (pmap->pm_ptphint);
3426 ptp = uvm_pagelookup(&pmap->pm_obj, va);
3427 #ifdef DIAGNOSTIC
3428 if (ptp == NULL)
3429 panic("pmap_get_ptp: unmanaged user PTP");
3430 #endif
3431 pmap->pm_ptphint = ptp;
3432 return(ptp);
3433 }
3434
3435 /* allocate a new PTP (updates ptphint) */
3436 return(pmap_alloc_ptp(pmap, va));
3437 }
3438
3439 /*
3440 * pmap_alloc_ptp: allocate a PTP for a PMAP
3441 *
3442 * => pmap should already be locked by caller
3443 * => we use the ptp's wire_count to count the number of active mappings
3444 * in the PTP (we start it at one to prevent any chance this PTP
3445 * will ever leak onto the active/inactive queues)
3446 */
3447
3448 /*__inline */ static struct vm_page *
3449 pmap_alloc_ptp(struct pmap *pmap, vaddr_t va)
3450 {
3451 struct vm_page *ptp;
3452
3453 ptp = uvm_pagealloc(&pmap->pm_obj, va, NULL,
3454 UVM_PGA_USERESERVE|UVM_PGA_ZERO);
3455 if (ptp == NULL)
3456 return (NULL);
3457
3458 /* got one! */
3459 ptp->flags &= ~PG_BUSY; /* never busy */
3460 ptp->wire_count = 1; /* no mappings yet */
3461 pmap_map_in_l1(pmap, va, VM_PAGE_TO_PHYS(ptp), TRUE);
3462 pmap->pm_stats.resident_count++; /* count PTP as resident */
3463 pmap->pm_ptphint = ptp;
3464 return (ptp);
3465 }
3466
3467 vaddr_t
3468 pmap_growkernel(maxkvaddr)
3469 vaddr_t maxkvaddr;
3470 {
3471 struct pmap *kpm = pmap_kernel(), *pm;
3472 int s;
3473 paddr_t ptaddr;
3474 struct vm_page *ptp;
3475
3476 if (maxkvaddr <= pmap_curmaxkvaddr)
3477 goto out; /* we are OK */
3478 NPDEBUG(PDB_GROWKERN, printf("pmap_growkernel: growing kernel from %lx to %lx\n",
3479 pmap_curmaxkvaddr, maxkvaddr));
3480
3481 /*
3482 * whoops! we need to add kernel PTPs
3483 */
3484
3485 s = splhigh(); /* to be safe */
3486 simple_lock(&kpm->pm_obj.vmobjlock);
3487 /* due to the way the arm pmap works we map 4MB at a time */
3488 for (/*null*/ ; pmap_curmaxkvaddr < maxkvaddr;
3489 pmap_curmaxkvaddr += 4 * NBPD) {
3490
3491 if (uvm.page_init_done == FALSE) {
3492
3493 /*
3494 * we're growing the kernel pmap early (from
3495 * uvm_pageboot_alloc()). this case must be
3496 * handled a little differently.
3497 */
3498
3499 if (uvm_page_physget(&ptaddr) == FALSE)
3500 panic("pmap_growkernel: out of memory");
3501 pmap_zero_page(ptaddr);
3502
3503 /* map this page in */
3504 pmap_map_in_l1(kpm, pmap_curmaxkvaddr, ptaddr, TRUE);
3505
3506 /* count PTP as resident */
3507 kpm->pm_stats.resident_count++;
3508 continue;
3509 }
3510
3511 /*
3512 * THIS *MUST* BE CODED SO AS TO WORK IN THE
3513 * pmap_initialized == FALSE CASE! WE MAY BE
3514 * INVOKED WHILE pmap_init() IS RUNNING!
3515 */
3516
3517 if ((ptp = pmap_alloc_ptp(kpm, pmap_curmaxkvaddr)) == NULL)
3518 panic("pmap_growkernel: alloc ptp failed");
3519
3520 /* distribute new kernel PTP to all active pmaps */
3521 simple_lock(&pmaps_lock);
3522 LIST_FOREACH(pm, &pmaps, pm_list) {
3523 pmap_map_in_l1(pm, pmap_curmaxkvaddr,
3524 VM_PAGE_TO_PHYS(ptp), TRUE);
3525 }
3526
3527 simple_unlock(&pmaps_lock);
3528 }
3529
3530 /*
3531 * flush out the cache, expensive but growkernel will happen so
3532 * rarely
3533 */
3534 cpu_tlb_flushD();
3535 cpu_cpwait();
3536
3537 simple_unlock(&kpm->pm_obj.vmobjlock);
3538 splx(s);
3539
3540 out:
3541 return (pmap_curmaxkvaddr);
3542 }
3543
3544
3545
3546 /************************ Bootstrapping routines ****************************/
3547
3548 /*
3549 * This list exists for the benefit of pmap_map_chunk(). It keeps track
3550 * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
3551 * find them as necessary.
3552 *
3553 * Note that the data on this list is not valid after initarm() returns.
3554 */
3555 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
3556
3557 static vaddr_t
3558 kernel_pt_lookup(paddr_t pa)
3559 {
3560 pv_addr_t *pv;
3561
3562 SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
3563 if (pv->pv_pa == pa)
3564 return (pv->pv_va);
3565 }
3566 return (0);
3567 }
3568
3569 /*
3570 * pmap_map_section:
3571 *
3572 * Create a single section mapping.
3573 */
3574 void
3575 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
3576 {
3577 pd_entry_t *pde = (pd_entry_t *) l1pt;
3578 pd_entry_t ap = (prot & VM_PROT_WRITE) ? AP_KRW : AP_KR;
3579 pd_entry_t fl = (cache == PTE_CACHE) ? pte_cache_mode : 0;
3580
3581 KASSERT(((va | pa) & (L1_SEC_SIZE - 1)) == 0);
3582
3583 pde[va >> PDSHIFT] = L1_SECPTE(pa & PD_MASK, ap, fl);
3584 }
3585
3586 /*
3587 * pmap_map_entry:
3588 *
3589 * Create a single page mapping.
3590 */
3591 void
3592 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
3593 {
3594 pd_entry_t *pde = (pd_entry_t *) l1pt;
3595 pt_entry_t ap = (prot & VM_PROT_WRITE) ? AP_KRW : AP_KR;
3596 pt_entry_t fl = (cache == PTE_CACHE) ? pte_cache_mode : 0;
3597 pt_entry_t *pte;
3598
3599 KASSERT(((va | pa) & PGOFSET) == 0);
3600
3601 if ((pde[va >> PDSHIFT] & L1_MASK) != L1_PAGE)
3602 panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
3603
3604 pte = (pt_entry_t *)
3605 kernel_pt_lookup(pde[va >> PDSHIFT] & PG_FRAME);
3606 if (pte == NULL)
3607 panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
3608
3609 pte[(va >> PGSHIFT) & 0x3ff] = L2_SPTE(pa & PG_FRAME, ap, fl);
3610 }
3611
3612 /*
3613 * pmap_link_l2pt:
3614 *
3615 * Link the L2 page table specified by "pa" into the L1
3616 * page table at the slot for "va".
3617 */
3618 void
3619 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
3620 {
3621 pd_entry_t *pde = (pd_entry_t *) l1pt;
3622 u_int slot = va >> PDSHIFT;
3623
3624 KASSERT((l2pv->pv_pa & PGOFSET) == 0);
3625
3626 pde[slot + 0] = L1_PTE(l2pv->pv_pa + 0x000);
3627 pde[slot + 1] = L1_PTE(l2pv->pv_pa + 0x400);
3628 pde[slot + 2] = L1_PTE(l2pv->pv_pa + 0x800);
3629 pde[slot + 3] = L1_PTE(l2pv->pv_pa + 0xc00);
3630
3631 SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
3632 }
3633
3634 /*
3635 * pmap_map_chunk:
3636 *
3637 * Map a chunk of memory using the most efficient mappings
3638 * possible (section, large page, small page) into the
3639 * provided L1 and L2 tables at the specified virtual address.
3640 */
3641 vsize_t
3642 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
3643 int prot, int cache)
3644 {
3645 pd_entry_t *pde = (pd_entry_t *) l1pt;
3646 pt_entry_t ap = (prot & VM_PROT_WRITE) ? AP_KRW : AP_KR;
3647 pt_entry_t fl = (cache == PTE_CACHE) ? pte_cache_mode : 0;
3648 pt_entry_t *pte;
3649 vsize_t resid;
3650 int i;
3651
3652 resid = (size + (NBPG - 1)) & ~(NBPG - 1);
3653
3654 if (l1pt == 0)
3655 panic("pmap_map_chunk: no L1 table provided");
3656
3657 #ifdef VERBOSE_INIT_ARM
3658 printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
3659 "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
3660 #endif
3661
3662 size = resid;
3663
3664 while (resid > 0) {
3665 /* See if we can use a section mapping. */
3666 if (((pa | va) & (L1_SEC_SIZE - 1)) == 0 &&
3667 resid >= L1_SEC_SIZE) {
3668 #ifdef VERBOSE_INIT_ARM
3669 printf("S");
3670 #endif
3671 pde[va >> PDSHIFT] = L1_SECPTE(pa, ap, fl);
3672 va += L1_SEC_SIZE;
3673 pa += L1_SEC_SIZE;
3674 resid -= L1_SEC_SIZE;
3675 continue;
3676 }
3677
3678 /*
3679 * Ok, we're going to use an L2 table. Make sure
3680 * one is actually in the corresponding L1 slot
3681 * for the current VA.
3682 */
3683 if ((pde[va >> PDSHIFT] & L1_MASK) != L1_PAGE)
3684 panic("pmap_map_chunk: no L2 table for VA 0x%08lx", va);
3685
3686 pte = (pt_entry_t *)
3687 kernel_pt_lookup(pde[va >> PDSHIFT] & PG_FRAME);
3688 if (pte == NULL)
3689 panic("pmap_map_chunk: can't find L2 table for VA"
3690 "0x%08lx", va);
3691
3692 /* See if we can use a L2 large page mapping. */
3693 if (((pa | va) & (L2_LPAGE_SIZE - 1)) == 0 &&
3694 resid >= L2_LPAGE_SIZE) {
3695 #ifdef VERBOSE_INIT_ARM
3696 printf("L");
3697 #endif
3698 for (i = 0; i < 16; i++) {
3699 pte[((va >> PGSHIFT) & 0x3f0) + i] =
3700 L2_LPTE(pa, ap, fl);
3701 }
3702 va += L2_LPAGE_SIZE;
3703 pa += L2_LPAGE_SIZE;
3704 resid -= L2_LPAGE_SIZE;
3705 continue;
3706 }
3707
3708 /* Use a small page mapping. */
3709 #ifdef VERBOSE_INIT_ARM
3710 printf("P");
3711 #endif
3712 pte[(va >> PGSHIFT) & 0x3ff] = L2_SPTE(pa, ap, fl);
3713 va += NBPG;
3714 pa += NBPG;
3715 resid -= NBPG;
3716 }
3717 #ifdef VERBOSE_INIT_ARM
3718 printf("\n");
3719 #endif
3720 return (size);
3721 }
3722