xref: /netbsd-src/sys/arch/arm/arm32/pmap.c (revision bbde328be4e75ea9ad02e9715ea13ca54b797ada)
1 /*	$NetBSD: pmap.c,v 1.212 2010/02/15 07:55:33 skrll Exp $	*/
2 
3 /*
4  * Copyright 2003 Wasabi Systems, Inc.
5  * All rights reserved.
6  *
7  * Written by Steve C. Woodford for Wasabi Systems, Inc.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. All advertising materials mentioning features or use of this software
18  *    must display the following acknowledgement:
19  *      This product includes software developed for the NetBSD Project by
20  *      Wasabi Systems, Inc.
21  * 4. The name of Wasabi Systems, Inc. may not be used to endorse
22  *    or promote products derived from this software without specific prior
23  *    written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY WASABI SYSTEMS, INC. ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
27  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
28  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL WASABI SYSTEMS, INC
29  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
30  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
31  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
32  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
33  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
34  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
35  * POSSIBILITY OF SUCH DAMAGE.
36  */
37 
38 /*
39  * Copyright (c) 2002-2003 Wasabi Systems, Inc.
40  * Copyright (c) 2001 Richard Earnshaw
41  * Copyright (c) 2001-2002 Christopher Gilbert
42  * All rights reserved.
43  *
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. The name of the company nor the name of the author may be used to
50  *    endorse or promote products derived from this software without specific
51  *    prior written permission.
52  *
53  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
54  * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
55  * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
56  * IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
57  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
58  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
59  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
60  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
61  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
62  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
63  * SUCH DAMAGE.
64  */
65 
66 /*-
67  * Copyright (c) 1999 The NetBSD Foundation, Inc.
68  * All rights reserved.
69  *
70  * This code is derived from software contributed to The NetBSD Foundation
71  * by Charles M. Hannum.
72  *
73  * Redistribution and use in source and binary forms, with or without
74  * modification, are permitted provided that the following conditions
75  * are met:
76  * 1. Redistributions of source code must retain the above copyright
77  *    notice, this list of conditions and the following disclaimer.
78  * 2. Redistributions in binary form must reproduce the above copyright
79  *    notice, this list of conditions and the following disclaimer in the
80  *    documentation and/or other materials provided with the distribution.
81  *
82  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
83  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
84  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
85  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
86  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
87  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
88  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
89  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
90  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
91  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
92  * POSSIBILITY OF SUCH DAMAGE.
93  */
94 
95 /*
96  * Copyright (c) 1994-1998 Mark Brinicombe.
97  * Copyright (c) 1994 Brini.
98  * All rights reserved.
99  *
100  * This code is derived from software written for Brini by Mark Brinicombe
101  *
102  * Redistribution and use in source and binary forms, with or without
103  * modification, are permitted provided that the following conditions
104  * are met:
105  * 1. Redistributions of source code must retain the above copyright
106  *    notice, this list of conditions and the following disclaimer.
107  * 2. Redistributions in binary form must reproduce the above copyright
108  *    notice, this list of conditions and the following disclaimer in the
109  *    documentation and/or other materials provided with the distribution.
110  * 3. All advertising materials mentioning features or use of this software
111  *    must display the following acknowledgement:
112  *	This product includes software developed by Mark Brinicombe.
113  * 4. The name of the author may not be used to endorse or promote products
114  *    derived from this software without specific prior written permission.
115  *
116  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
117  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
118  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
119  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
120  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
121  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
122  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
123  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
124  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
125  *
126  * RiscBSD kernel project
127  *
128  * pmap.c
129  *
130  * Machine dependant vm stuff
131  *
132  * Created      : 20/09/94
133  */
134 
135 /*
136  * armv6 and VIPT cache support by 3am Software Foundry,
137  * Copyright (c) 2007 Microsoft
138  */
139 
140 /*
141  * Performance improvements, UVM changes, overhauls and part-rewrites
142  * were contributed by Neil A. Carson <neil@causality.com>.
143  */
144 
145 /*
146  * Overhauled again to speedup the pmap, use MMU Domains so that L1 tables
147  * can be shared, and re-work the KVM layout, by Steve Woodford of Wasabi
148  * Systems, Inc.
149  *
150  * There are still a few things outstanding at this time:
151  *
152  *   - There are some unresolved issues for MP systems:
153  *
154  *     o The L1 metadata needs a lock, or more specifically, some places
155  *       need to acquire an exclusive lock when modifying L1 translation
156  *       table entries.
157  *
158  *     o When one cpu modifies an L1 entry, and that L1 table is also
159  *       being used by another cpu, then the latter will need to be told
160  *       that a tlb invalidation may be necessary. (But only if the old
161  *       domain number in the L1 entry being over-written is currently
162  *       the active domain on that cpu). I guess there are lots more tlb
163  *       shootdown issues too...
164  *
165  *     o If the vector_page is at 0x00000000 instead of 0xffff0000, then
166  *       MP systems will lose big-time because of the MMU domain hack.
167  *       The only way this can be solved (apart from moving the vector
168  *       page to 0xffff0000) is to reserve the first 1MB of user address
169  *       space for kernel use only. This would require re-linking all
170  *       applications so that the text section starts above this 1MB
171  *       boundary.
172  *
173  *     o Tracking which VM space is resident in the cache/tlb has not yet
174  *       been implemented for MP systems.
175  *
176  *     o Finally, there is a pathological condition where two cpus running
177  *       two separate processes (not lwps) which happen to share an L1
178  *       can get into a fight over one or more L1 entries. This will result
179  *       in a significant slow-down if both processes are in tight loops.
180  */
181 
182 /*
183  * Special compilation symbols
184  * PMAP_DEBUG		- Build in pmap_debug_level code
185  */
186 
187 /* Include header files */
188 
189 #include "opt_cpuoptions.h"
190 #include "opt_pmap_debug.h"
191 #include "opt_ddb.h"
192 #include "opt_lockdebug.h"
193 #include "opt_multiprocessor.h"
194 
195 #include <sys/param.h>
196 #include <sys/types.h>
197 #include <sys/kernel.h>
198 #include <sys/systm.h>
199 #include <sys/proc.h>
200 #include <sys/malloc.h>
201 #include <sys/pool.h>
202 #include <sys/cdefs.h>
203 #include <sys/cpu.h>
204 #include <sys/sysctl.h>
205 
206 #include <uvm/uvm.h>
207 
208 #include <machine/bus.h>
209 #include <machine/pmap.h>
210 #include <machine/pcb.h>
211 #include <machine/param.h>
212 #include <arm/arm32/katelib.h>
213 
214 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.212 2010/02/15 07:55:33 skrll Exp $");
215 
216 #ifdef PMAP_DEBUG
217 
218 /* XXX need to get rid of all refs to this */
219 int pmap_debug_level = 0;
220 
221 /*
222  * for switching to potentially finer grained debugging
223  */
224 #define	PDB_FOLLOW	0x0001
225 #define	PDB_INIT	0x0002
226 #define	PDB_ENTER	0x0004
227 #define	PDB_REMOVE	0x0008
228 #define	PDB_CREATE	0x0010
229 #define	PDB_PTPAGE	0x0020
230 #define	PDB_GROWKERN	0x0040
231 #define	PDB_BITS	0x0080
232 #define	PDB_COLLECT	0x0100
233 #define	PDB_PROTECT	0x0200
234 #define	PDB_MAP_L1	0x0400
235 #define	PDB_BOOTSTRAP	0x1000
236 #define	PDB_PARANOIA	0x2000
237 #define	PDB_WIRING	0x4000
238 #define	PDB_PVDUMP	0x8000
239 #define	PDB_VAC		0x10000
240 #define	PDB_KENTER	0x20000
241 #define	PDB_KREMOVE	0x40000
242 #define	PDB_EXEC	0x80000
243 
244 int debugmap = 1;
245 int pmapdebug = 0;
246 #define	NPDEBUG(_lev_,_stat_) \
247 	if (pmapdebug & (_lev_)) \
248         	((_stat_))
249 
250 #else	/* PMAP_DEBUG */
251 #define NPDEBUG(_lev_,_stat_) /* Nothing */
252 #endif	/* PMAP_DEBUG */
253 
254 /*
255  * pmap_kernel() points here
256  */
257 static struct pmap	kernel_pmap_store;
258 struct pmap		*const kernel_pmap_ptr = &kernel_pmap_store;
259 
260 /*
261  * Which pmap is currently 'live' in the cache
262  *
263  * XXXSCW: Fix for SMP ...
264  */
265 static pmap_t pmap_recent_user;
266 
267 /*
268  * Pointer to last active lwp, or NULL if it exited.
269  */
270 struct lwp *pmap_previous_active_lwp;
271 
272 /*
273  * Pool and cache that pmap structures are allocated from.
274  * We use a cache to avoid clearing the pm_l2[] array (1KB)
275  * in pmap_create().
276  */
277 static struct pool_cache pmap_cache;
278 static LIST_HEAD(, pmap) pmap_pmaps;
279 
280 /*
281  * Pool of PV structures
282  */
283 static struct pool pmap_pv_pool;
284 static void *pmap_bootstrap_pv_page_alloc(struct pool *, int);
285 static void pmap_bootstrap_pv_page_free(struct pool *, void *);
286 static struct pool_allocator pmap_bootstrap_pv_allocator = {
287 	pmap_bootstrap_pv_page_alloc, pmap_bootstrap_pv_page_free
288 };
289 
290 /*
291  * Pool and cache of l2_dtable structures.
292  * We use a cache to avoid clearing the structures when they're
293  * allocated. (196 bytes)
294  */
295 static struct pool_cache pmap_l2dtable_cache;
296 static vaddr_t pmap_kernel_l2dtable_kva;
297 
298 /*
299  * Pool and cache of L2 page descriptors.
300  * We use a cache to avoid clearing the descriptor table
301  * when they're allocated. (1KB)
302  */
303 static struct pool_cache pmap_l2ptp_cache;
304 static vaddr_t pmap_kernel_l2ptp_kva;
305 static paddr_t pmap_kernel_l2ptp_phys;
306 
307 #ifdef PMAPCOUNTERS
308 #define	PMAP_EVCNT_INITIALIZER(name) \
309 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap", name)
310 
311 #ifdef PMAP_CACHE_VIPT
312 static struct evcnt pmap_ev_vac_clean_one =
313    PMAP_EVCNT_INITIALIZER("clean page (1 color)");
314 static struct evcnt pmap_ev_vac_flush_one =
315    PMAP_EVCNT_INITIALIZER("flush page (1 color)");
316 static struct evcnt pmap_ev_vac_flush_lots =
317    PMAP_EVCNT_INITIALIZER("flush page (2+ colors)");
318 static struct evcnt pmap_ev_vac_flush_lots2 =
319    PMAP_EVCNT_INITIALIZER("flush page (2+ colors, kmpage)");
320 EVCNT_ATTACH_STATIC(pmap_ev_vac_clean_one);
321 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_one);
322 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_lots);
323 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_lots2);
324 
325 static struct evcnt pmap_ev_vac_color_new =
326    PMAP_EVCNT_INITIALIZER("new page color");
327 static struct evcnt pmap_ev_vac_color_reuse =
328    PMAP_EVCNT_INITIALIZER("ok first page color");
329 static struct evcnt pmap_ev_vac_color_ok =
330    PMAP_EVCNT_INITIALIZER("ok page color");
331 static struct evcnt pmap_ev_vac_color_blind =
332    PMAP_EVCNT_INITIALIZER("blind page color");
333 static struct evcnt pmap_ev_vac_color_change =
334    PMAP_EVCNT_INITIALIZER("change page color");
335 static struct evcnt pmap_ev_vac_color_erase =
336    PMAP_EVCNT_INITIALIZER("erase page color");
337 static struct evcnt pmap_ev_vac_color_none =
338    PMAP_EVCNT_INITIALIZER("no page color");
339 static struct evcnt pmap_ev_vac_color_restore =
340    PMAP_EVCNT_INITIALIZER("restore page color");
341 
342 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_new);
343 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_reuse);
344 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_ok);
345 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_blind);
346 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_change);
347 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_erase);
348 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_none);
349 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_restore);
350 #endif
351 
352 static struct evcnt pmap_ev_mappings =
353    PMAP_EVCNT_INITIALIZER("pages mapped");
354 static struct evcnt pmap_ev_unmappings =
355    PMAP_EVCNT_INITIALIZER("pages unmapped");
356 static struct evcnt pmap_ev_remappings =
357    PMAP_EVCNT_INITIALIZER("pages remapped");
358 
359 EVCNT_ATTACH_STATIC(pmap_ev_mappings);
360 EVCNT_ATTACH_STATIC(pmap_ev_unmappings);
361 EVCNT_ATTACH_STATIC(pmap_ev_remappings);
362 
363 static struct evcnt pmap_ev_kernel_mappings =
364    PMAP_EVCNT_INITIALIZER("kernel pages mapped");
365 static struct evcnt pmap_ev_kernel_unmappings =
366    PMAP_EVCNT_INITIALIZER("kernel pages unmapped");
367 static struct evcnt pmap_ev_kernel_remappings =
368    PMAP_EVCNT_INITIALIZER("kernel pages remapped");
369 
370 EVCNT_ATTACH_STATIC(pmap_ev_kernel_mappings);
371 EVCNT_ATTACH_STATIC(pmap_ev_kernel_unmappings);
372 EVCNT_ATTACH_STATIC(pmap_ev_kernel_remappings);
373 
374 static struct evcnt pmap_ev_kenter_mappings =
375    PMAP_EVCNT_INITIALIZER("kenter pages mapped");
376 static struct evcnt pmap_ev_kenter_unmappings =
377    PMAP_EVCNT_INITIALIZER("kenter pages unmapped");
378 static struct evcnt pmap_ev_kenter_remappings =
379    PMAP_EVCNT_INITIALIZER("kenter pages remapped");
380 static struct evcnt pmap_ev_pt_mappings =
381    PMAP_EVCNT_INITIALIZER("page table pages mapped");
382 
383 EVCNT_ATTACH_STATIC(pmap_ev_kenter_mappings);
384 EVCNT_ATTACH_STATIC(pmap_ev_kenter_unmappings);
385 EVCNT_ATTACH_STATIC(pmap_ev_kenter_remappings);
386 EVCNT_ATTACH_STATIC(pmap_ev_pt_mappings);
387 
388 #ifdef PMAP_CACHE_VIPT
389 static struct evcnt pmap_ev_exec_mappings =
390    PMAP_EVCNT_INITIALIZER("exec pages mapped");
391 static struct evcnt pmap_ev_exec_cached =
392    PMAP_EVCNT_INITIALIZER("exec pages cached");
393 
394 EVCNT_ATTACH_STATIC(pmap_ev_exec_mappings);
395 EVCNT_ATTACH_STATIC(pmap_ev_exec_cached);
396 
397 static struct evcnt pmap_ev_exec_synced =
398    PMAP_EVCNT_INITIALIZER("exec pages synced");
399 static struct evcnt pmap_ev_exec_synced_map =
400    PMAP_EVCNT_INITIALIZER("exec pages synced (MP)");
401 static struct evcnt pmap_ev_exec_synced_unmap =
402    PMAP_EVCNT_INITIALIZER("exec pages synced (UM)");
403 static struct evcnt pmap_ev_exec_synced_remap =
404    PMAP_EVCNT_INITIALIZER("exec pages synced (RM)");
405 static struct evcnt pmap_ev_exec_synced_clearbit =
406    PMAP_EVCNT_INITIALIZER("exec pages synced (DG)");
407 static struct evcnt pmap_ev_exec_synced_kremove =
408    PMAP_EVCNT_INITIALIZER("exec pages synced (KU)");
409 
410 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced);
411 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_map);
412 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_unmap);
413 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_remap);
414 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_clearbit);
415 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_kremove);
416 
417 static struct evcnt pmap_ev_exec_discarded_unmap =
418    PMAP_EVCNT_INITIALIZER("exec pages discarded (UM)");
419 static struct evcnt pmap_ev_exec_discarded_zero =
420    PMAP_EVCNT_INITIALIZER("exec pages discarded (ZP)");
421 static struct evcnt pmap_ev_exec_discarded_copy =
422    PMAP_EVCNT_INITIALIZER("exec pages discarded (CP)");
423 static struct evcnt pmap_ev_exec_discarded_page_protect =
424    PMAP_EVCNT_INITIALIZER("exec pages discarded (PP)");
425 static struct evcnt pmap_ev_exec_discarded_clearbit =
426    PMAP_EVCNT_INITIALIZER("exec pages discarded (DG)");
427 static struct evcnt pmap_ev_exec_discarded_kremove =
428    PMAP_EVCNT_INITIALIZER("exec pages discarded (KU)");
429 
430 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_unmap);
431 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_zero);
432 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_copy);
433 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_page_protect);
434 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_clearbit);
435 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_kremove);
436 #endif /* PMAP_CACHE_VIPT */
437 
438 static struct evcnt pmap_ev_updates = PMAP_EVCNT_INITIALIZER("updates");
439 static struct evcnt pmap_ev_collects = PMAP_EVCNT_INITIALIZER("collects");
440 static struct evcnt pmap_ev_activations = PMAP_EVCNT_INITIALIZER("activations");
441 
442 EVCNT_ATTACH_STATIC(pmap_ev_updates);
443 EVCNT_ATTACH_STATIC(pmap_ev_collects);
444 EVCNT_ATTACH_STATIC(pmap_ev_activations);
445 
446 #define	PMAPCOUNT(x)	((void)(pmap_ev_##x.ev_count++))
447 #else
448 #define	PMAPCOUNT(x)	((void)0)
449 #endif
450 
451 /*
452  * pmap copy/zero page, and mem(5) hook point
453  */
454 static pt_entry_t *csrc_pte, *cdst_pte;
455 static vaddr_t csrcp, cdstp;
456 vaddr_t memhook;			/* used by mem.c */
457 kmutex_t memlock;			/* used by mem.c */
458 void *zeropage;				/* used by mem.c */
459 extern void *msgbufaddr;
460 int pmap_kmpages;
461 /*
462  * Flag to indicate if pmap_init() has done its thing
463  */
464 bool pmap_initialized;
465 
466 /*
467  * Misc. locking data structures
468  */
469 
470 #if 0 /* defined(MULTIPROCESSOR) || defined(LOCKDEBUG) */
471 static struct lock pmap_main_lock;
472 
473 #define PMAP_MAP_TO_HEAD_LOCK() \
474      (void) spinlockmgr(&pmap_main_lock, LK_SHARED, NULL)
475 #define PMAP_MAP_TO_HEAD_UNLOCK() \
476      (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
477 #define PMAP_HEAD_TO_MAP_LOCK() \
478      (void) spinlockmgr(&pmap_main_lock, LK_EXCLUSIVE, NULL)
479 #define PMAP_HEAD_TO_MAP_UNLOCK() \
480      spinlockmgr(&pmap_main_lock, LK_RELEASE, (void *) 0)
481 #else
482 #define PMAP_MAP_TO_HEAD_LOCK()		/* null */
483 #define PMAP_MAP_TO_HEAD_UNLOCK()	/* null */
484 #define PMAP_HEAD_TO_MAP_LOCK()		/* null */
485 #define PMAP_HEAD_TO_MAP_UNLOCK()	/* null */
486 #endif
487 
488 #define	pmap_acquire_pmap_lock(pm)			\
489 	do {						\
490 		if ((pm) != pmap_kernel())		\
491 			mutex_enter(&(pm)->pm_lock);	\
492 	} while (/*CONSTCOND*/0)
493 
494 #define	pmap_release_pmap_lock(pm)			\
495 	do {						\
496 		if ((pm) != pmap_kernel())		\
497 			mutex_exit(&(pm)->pm_lock);	\
498 	} while (/*CONSTCOND*/0)
499 
500 
501 /*
502  * Metadata for L1 translation tables.
503  */
504 struct l1_ttable {
505 	/* Entry on the L1 Table list */
506 	SLIST_ENTRY(l1_ttable) l1_link;
507 
508 	/* Entry on the L1 Least Recently Used list */
509 	TAILQ_ENTRY(l1_ttable) l1_lru;
510 
511 	/* Track how many domains are allocated from this L1 */
512 	volatile u_int l1_domain_use_count;
513 
514 	/*
515 	 * A free-list of domain numbers for this L1.
516 	 * We avoid using ffs() and a bitmap to track domains since ffs()
517 	 * is slow on ARM.
518 	 */
519 	u_int8_t l1_domain_first;
520 	u_int8_t l1_domain_free[PMAP_DOMAINS];
521 
522 	/* Physical address of this L1 page table */
523 	paddr_t l1_physaddr;
524 
525 	/* KVA of this L1 page table */
526 	pd_entry_t *l1_kva;
527 };
528 
529 /*
530  * Convert a virtual address into its L1 table index. That is, the
531  * index used to locate the L2 descriptor table pointer in an L1 table.
532  * This is basically used to index l1->l1_kva[].
533  *
534  * Each L2 descriptor table represents 1MB of VA space.
535  */
536 #define	L1_IDX(va)		(((vaddr_t)(va)) >> L1_S_SHIFT)
537 
538 /*
539  * L1 Page Tables are tracked using a Least Recently Used list.
540  *  - New L1s are allocated from the HEAD.
541  *  - Freed L1s are added to the TAIl.
542  *  - Recently accessed L1s (where an 'access' is some change to one of
543  *    the userland pmaps which owns this L1) are moved to the TAIL.
544  */
545 static TAILQ_HEAD(, l1_ttable) l1_lru_list;
546 static struct simplelock l1_lru_lock;
547 
548 /*
549  * A list of all L1 tables
550  */
551 static SLIST_HEAD(, l1_ttable) l1_list;
552 
553 /*
554  * The l2_dtable tracks L2_BUCKET_SIZE worth of L1 slots.
555  *
556  * This is normally 16MB worth L2 page descriptors for any given pmap.
557  * Reference counts are maintained for L2 descriptors so they can be
558  * freed when empty.
559  */
560 struct l2_dtable {
561 	/* The number of L2 page descriptors allocated to this l2_dtable */
562 	u_int l2_occupancy;
563 
564 	/* List of L2 page descriptors */
565 	struct l2_bucket {
566 		pt_entry_t *l2b_kva;	/* KVA of L2 Descriptor Table */
567 		paddr_t l2b_phys;	/* Physical address of same */
568 		u_short l2b_l1idx;	/* This L2 table's L1 index */
569 		u_short l2b_occupancy;	/* How many active descriptors */
570 	} l2_bucket[L2_BUCKET_SIZE];
571 };
572 
573 /*
574  * Given an L1 table index, calculate the corresponding l2_dtable index
575  * and bucket index within the l2_dtable.
576  */
577 #define	L2_IDX(l1idx)		(((l1idx) >> L2_BUCKET_LOG2) & \
578 				 (L2_SIZE - 1))
579 #define	L2_BUCKET(l1idx)	((l1idx) & (L2_BUCKET_SIZE - 1))
580 
581 /*
582  * Given a virtual address, this macro returns the
583  * virtual address required to drop into the next L2 bucket.
584  */
585 #define	L2_NEXT_BUCKET(va)	(((va) & L1_S_FRAME) + L1_S_SIZE)
586 
587 /*
588  * L2 allocation.
589  */
590 #define	pmap_alloc_l2_dtable()		\
591 	    pool_cache_get(&pmap_l2dtable_cache, PR_NOWAIT)
592 #define	pmap_free_l2_dtable(l2)		\
593 	    pool_cache_put(&pmap_l2dtable_cache, (l2))
594 #define pmap_alloc_l2_ptp(pap)		\
595 	    ((pt_entry_t *)pool_cache_get_paddr(&pmap_l2ptp_cache,\
596 	    PR_NOWAIT, (pap)))
597 
598 /*
599  * We try to map the page tables write-through, if possible.  However, not
600  * all CPUs have a write-through cache mode, so on those we have to sync
601  * the cache when we frob page tables.
602  *
603  * We try to evaluate this at compile time, if possible.  However, it's
604  * not always possible to do that, hence this run-time var.
605  */
606 int	pmap_needs_pte_sync;
607 
608 /*
609  * Real definition of pv_entry.
610  */
611 struct pv_entry {
612 	SLIST_ENTRY(pv_entry) pv_link;	/* next pv_entry */
613 	pmap_t		pv_pmap;        /* pmap where mapping lies */
614 	vaddr_t		pv_va;          /* virtual address for mapping */
615 	u_int		pv_flags;       /* flags */
616 };
617 
618 /*
619  * Macro to determine if a mapping might be resident in the
620  * instruction cache and/or TLB
621  */
622 #define	PV_BEEN_EXECD(f)  (((f) & (PVF_REF | PVF_EXEC)) == (PVF_REF | PVF_EXEC))
623 #define	PV_IS_EXEC_P(f)   (((f) & PVF_EXEC) != 0)
624 
625 /*
626  * Macro to determine if a mapping might be resident in the
627  * data cache and/or TLB
628  */
629 #define	PV_BEEN_REFD(f)   (((f) & PVF_REF) != 0)
630 
631 /*
632  * Local prototypes
633  */
634 static int		pmap_set_pt_cache_mode(pd_entry_t *, vaddr_t);
635 static void		pmap_alloc_specials(vaddr_t *, int, vaddr_t *,
636 			    pt_entry_t **);
637 static bool		pmap_is_current(pmap_t);
638 static bool		pmap_is_cached(pmap_t);
639 static void		pmap_enter_pv(struct vm_page *, struct pv_entry *,
640 			    pmap_t, vaddr_t, u_int);
641 static struct pv_entry *pmap_find_pv(struct vm_page *, pmap_t, vaddr_t);
642 static struct pv_entry *pmap_remove_pv(struct vm_page *, pmap_t, vaddr_t);
643 static u_int		pmap_modify_pv(struct vm_page *, pmap_t, vaddr_t,
644 			    u_int, u_int);
645 
646 static void		pmap_pinit(pmap_t);
647 static int		pmap_pmap_ctor(void *, void *, int);
648 
649 static void		pmap_alloc_l1(pmap_t);
650 static void		pmap_free_l1(pmap_t);
651 static void		pmap_use_l1(pmap_t);
652 
653 static struct l2_bucket *pmap_get_l2_bucket(pmap_t, vaddr_t);
654 static struct l2_bucket *pmap_alloc_l2_bucket(pmap_t, vaddr_t);
655 static void		pmap_free_l2_bucket(pmap_t, struct l2_bucket *, u_int);
656 static int		pmap_l2ptp_ctor(void *, void *, int);
657 static int		pmap_l2dtable_ctor(void *, void *, int);
658 
659 static void		pmap_vac_me_harder(struct vm_page *, pmap_t, vaddr_t);
660 #ifdef PMAP_CACHE_VIVT
661 static void		pmap_vac_me_kpmap(struct vm_page *, pmap_t, vaddr_t);
662 static void		pmap_vac_me_user(struct vm_page *, pmap_t, vaddr_t);
663 #endif
664 
665 static void		pmap_clearbit(struct vm_page *, u_int);
666 #ifdef PMAP_CACHE_VIVT
667 static int		pmap_clean_page(struct pv_entry *, bool);
668 #endif
669 #ifdef PMAP_CACHE_VIPT
670 static void		pmap_syncicache_page(struct vm_page *);
671 enum pmap_flush_op {
672 	PMAP_FLUSH_PRIMARY,
673 	PMAP_FLUSH_SECONDARY,
674 	PMAP_CLEAN_PRIMARY
675 };
676 static void		pmap_flush_page(struct vm_page *, enum pmap_flush_op);
677 #endif
678 static void		pmap_page_remove(struct vm_page *);
679 
680 static void		pmap_init_l1(struct l1_ttable *, pd_entry_t *);
681 static vaddr_t		kernel_pt_lookup(paddr_t);
682 
683 
684 /*
685  * External function prototypes
686  */
687 extern void bzero_page(vaddr_t);
688 extern void bcopy_page(vaddr_t, vaddr_t);
689 
690 /*
691  * Misc variables
692  */
693 vaddr_t virtual_avail;
694 vaddr_t virtual_end;
695 vaddr_t pmap_curmaxkvaddr;
696 
697 paddr_t avail_start;
698 paddr_t avail_end;
699 
700 pv_addrqh_t pmap_boot_freeq = SLIST_HEAD_INITIALIZER(&pmap_boot_freeq);
701 pv_addr_t kernelpages;
702 pv_addr_t kernel_l1pt;
703 pv_addr_t systempage;
704 
705 /* Function to set the debug level of the pmap code */
706 
707 #ifdef PMAP_DEBUG
708 void
709 pmap_debug(int level)
710 {
711 	pmap_debug_level = level;
712 	printf("pmap_debug: level=%d\n", pmap_debug_level);
713 }
714 #endif	/* PMAP_DEBUG */
715 
716 /*
717  * A bunch of routines to conditionally flush the caches/TLB depending
718  * on whether the specified pmap actually needs to be flushed at any
719  * given time.
720  */
721 static inline void
722 pmap_tlb_flushID_SE(pmap_t pm, vaddr_t va)
723 {
724 
725 	if (pm->pm_cstate.cs_tlb_id)
726 		cpu_tlb_flushID_SE(va);
727 }
728 
729 static inline void
730 pmap_tlb_flushD_SE(pmap_t pm, vaddr_t va)
731 {
732 
733 	if (pm->pm_cstate.cs_tlb_d)
734 		cpu_tlb_flushD_SE(va);
735 }
736 
737 static inline void
738 pmap_tlb_flushID(pmap_t pm)
739 {
740 
741 	if (pm->pm_cstate.cs_tlb_id) {
742 		cpu_tlb_flushID();
743 		pm->pm_cstate.cs_tlb = 0;
744 	}
745 }
746 
747 static inline void
748 pmap_tlb_flushD(pmap_t pm)
749 {
750 
751 	if (pm->pm_cstate.cs_tlb_d) {
752 		cpu_tlb_flushD();
753 		pm->pm_cstate.cs_tlb_d = 0;
754 	}
755 }
756 
757 #ifdef PMAP_CACHE_VIVT
758 static inline void
759 pmap_idcache_wbinv_range(pmap_t pm, vaddr_t va, vsize_t len)
760 {
761 	if (pm->pm_cstate.cs_cache_id) {
762 		cpu_idcache_wbinv_range(va, len);
763 	}
764 }
765 
766 static inline void
767 pmap_dcache_wb_range(pmap_t pm, vaddr_t va, vsize_t len,
768     bool do_inv, bool rd_only)
769 {
770 
771 	if (pm->pm_cstate.cs_cache_d) {
772 		if (do_inv) {
773 			if (rd_only)
774 				cpu_dcache_inv_range(va, len);
775 			else
776 				cpu_dcache_wbinv_range(va, len);
777 		} else
778 		if (!rd_only)
779 			cpu_dcache_wb_range(va, len);
780 	}
781 }
782 
783 static inline void
784 pmap_idcache_wbinv_all(pmap_t pm)
785 {
786 	if (pm->pm_cstate.cs_cache_id) {
787 		cpu_idcache_wbinv_all();
788 		pm->pm_cstate.cs_cache = 0;
789 	}
790 }
791 
792 static inline void
793 pmap_dcache_wbinv_all(pmap_t pm)
794 {
795 	if (pm->pm_cstate.cs_cache_d) {
796 		cpu_dcache_wbinv_all();
797 		pm->pm_cstate.cs_cache_d = 0;
798 	}
799 }
800 #endif /* PMAP_CACHE_VIVT */
801 
802 static inline bool
803 pmap_is_current(pmap_t pm)
804 {
805 
806 	if (pm == pmap_kernel() || curproc->p_vmspace->vm_map.pmap == pm)
807 		return true;
808 
809 	return false;
810 }
811 
812 static inline bool
813 pmap_is_cached(pmap_t pm)
814 {
815 
816 	if (pm == pmap_kernel() || pmap_recent_user == NULL ||
817 	    pmap_recent_user == pm)
818 		return (true);
819 
820 	return false;
821 }
822 
823 /*
824  * PTE_SYNC_CURRENT:
825  *
826  *     Make sure the pte is written out to RAM.
827  *     We need to do this for one of two cases:
828  *       - We're dealing with the kernel pmap
829  *       - There is no pmap active in the cache/tlb.
830  *       - The specified pmap is 'active' in the cache/tlb.
831  */
832 #ifdef PMAP_INCLUDE_PTE_SYNC
833 #define	PTE_SYNC_CURRENT(pm, ptep)	\
834 do {					\
835 	if (PMAP_NEEDS_PTE_SYNC && 	\
836 	    pmap_is_cached(pm))		\
837 		PTE_SYNC(ptep);		\
838 } while (/*CONSTCOND*/0)
839 #else
840 #define	PTE_SYNC_CURRENT(pm, ptep)	/* nothing */
841 #endif
842 
843 /*
844  * main pv_entry manipulation functions:
845  *   pmap_enter_pv: enter a mapping onto a vm_page list
846  *   pmap_remove_pv: remove a mappiing from a vm_page list
847  *
848  * NOTE: pmap_enter_pv expects to lock the pvh itself
849  *       pmap_remove_pv expects te caller to lock the pvh before calling
850  */
851 
852 /*
853  * pmap_enter_pv: enter a mapping onto a vm_page lst
854  *
855  * => caller should hold the proper lock on pmap_main_lock
856  * => caller should have pmap locked
857  * => we will gain the lock on the vm_page and allocate the new pv_entry
858  * => caller should adjust ptp's wire_count before calling
859  * => caller should not adjust pmap's wire_count
860  */
861 static void
862 pmap_enter_pv(struct vm_page *pg, struct pv_entry *pv, pmap_t pm,
863     vaddr_t va, u_int flags)
864 {
865 	struct pv_entry **pvp;
866 
867 	NPDEBUG(PDB_PVDUMP,
868 	    printf("pmap_enter_pv: pm %p, pg %p, flags 0x%x\n", pm, pg, flags));
869 
870 	pv->pv_pmap = pm;
871 	pv->pv_va = va;
872 	pv->pv_flags = flags;
873 
874 	simple_lock(&pg->mdpage.pvh_slock);	/* lock vm_page */
875 	pvp = &SLIST_FIRST(&pg->mdpage.pvh_list);
876 #ifdef PMAP_CACHE_VIPT
877 	/*
878 	 * Insert unmanaged entries, writeable first, at the head of
879 	 * the pv list.
880 	 */
881 	if (__predict_true((flags & PVF_KENTRY) == 0)) {
882 		while (*pvp != NULL && (*pvp)->pv_flags & PVF_KENTRY)
883 			pvp = &SLIST_NEXT(*pvp, pv_link);
884 	} else if ((flags & PVF_WRITE) == 0) {
885 		while (*pvp != NULL && (*pvp)->pv_flags & PVF_WRITE)
886 			pvp = &SLIST_NEXT(*pvp, pv_link);
887 	}
888 #endif
889 	SLIST_NEXT(pv, pv_link) = *pvp;		/* add to ... */
890 	*pvp = pv;				/* ... locked list */
891 	pg->mdpage.pvh_attrs |= flags & (PVF_REF | PVF_MOD);
892 #ifdef PMAP_CACHE_VIPT
893 	if ((pv->pv_flags & PVF_KWRITE) == PVF_KWRITE)
894 		pg->mdpage.pvh_attrs |= PVF_KMOD;
895 	if ((pg->mdpage.pvh_attrs & (PVF_DMOD|PVF_NC)) != PVF_NC)
896 		pg->mdpage.pvh_attrs |= PVF_DIRTY;
897 	KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
898 #endif
899 	if (pm == pmap_kernel()) {
900 		PMAPCOUNT(kernel_mappings);
901 		if (flags & PVF_WRITE)
902 			pg->mdpage.krw_mappings++;
903 		else
904 			pg->mdpage.kro_mappings++;
905 	} else {
906 		if (flags & PVF_WRITE)
907 			pg->mdpage.urw_mappings++;
908 		else
909 			pg->mdpage.uro_mappings++;
910 	}
911 
912 #ifdef PMAP_CACHE_VIPT
913 	/*
914 	 * If this is an exec mapping and its the first exec mapping
915 	 * for this page, make sure to sync the I-cache.
916 	 */
917 	if (PV_IS_EXEC_P(flags)) {
918 		if (!PV_IS_EXEC_P(pg->mdpage.pvh_attrs)) {
919 			pmap_syncicache_page(pg);
920 			PMAPCOUNT(exec_synced_map);
921 		}
922 		PMAPCOUNT(exec_mappings);
923 	}
924 #endif
925 
926 	PMAPCOUNT(mappings);
927 	simple_unlock(&pg->mdpage.pvh_slock);	/* unlock, done! */
928 
929 	if (pv->pv_flags & PVF_WIRED)
930 		++pm->pm_stats.wired_count;
931 }
932 
933 /*
934  *
935  * pmap_find_pv: Find a pv entry
936  *
937  * => caller should hold lock on vm_page
938  */
939 static inline struct pv_entry *
940 pmap_find_pv(struct vm_page *pg, pmap_t pm, vaddr_t va)
941 {
942 	struct pv_entry *pv;
943 
944 	SLIST_FOREACH(pv, &pg->mdpage.pvh_list, pv_link) {
945 		if (pm == pv->pv_pmap && va == pv->pv_va)
946 			break;
947 	}
948 
949 	return (pv);
950 }
951 
952 /*
953  * pmap_remove_pv: try to remove a mapping from a pv_list
954  *
955  * => caller should hold proper lock on pmap_main_lock
956  * => pmap should be locked
957  * => caller should hold lock on vm_page [so that attrs can be adjusted]
958  * => caller should adjust ptp's wire_count and free PTP if needed
959  * => caller should NOT adjust pmap's wire_count
960  * => we return the removed pv
961  */
962 static struct pv_entry *
963 pmap_remove_pv(struct vm_page *pg, pmap_t pm, vaddr_t va)
964 {
965 	struct pv_entry *pv, **prevptr;
966 
967 	NPDEBUG(PDB_PVDUMP,
968 	    printf("pmap_remove_pv: pm %p, pg %p, va 0x%08lx\n", pm, pg, va));
969 
970 	prevptr = &SLIST_FIRST(&pg->mdpage.pvh_list); /* prev pv_entry ptr */
971 	pv = *prevptr;
972 
973 	while (pv) {
974 		if (pv->pv_pmap == pm && pv->pv_va == va) {	/* match? */
975 			NPDEBUG(PDB_PVDUMP, printf("pmap_remove_pv: pm %p, pg "
976 			    "%p, flags 0x%x\n", pm, pg, pv->pv_flags));
977 			if (pv->pv_flags & PVF_WIRED) {
978 				--pm->pm_stats.wired_count;
979 			}
980 			*prevptr = SLIST_NEXT(pv, pv_link);	/* remove it! */
981 			if (pm == pmap_kernel()) {
982 				PMAPCOUNT(kernel_unmappings);
983 				if (pv->pv_flags & PVF_WRITE)
984 					pg->mdpage.krw_mappings--;
985 				else
986 					pg->mdpage.kro_mappings--;
987 			} else {
988 				if (pv->pv_flags & PVF_WRITE)
989 					pg->mdpage.urw_mappings--;
990 				else
991 					pg->mdpage.uro_mappings--;
992 			}
993 
994 			PMAPCOUNT(unmappings);
995 #ifdef PMAP_CACHE_VIPT
996 			if (!(pv->pv_flags & PVF_WRITE))
997 				break;
998 			/*
999 			 * If this page has had an exec mapping, then if
1000 			 * this was the last mapping, discard the contents,
1001 			 * otherwise sync the i-cache for this page.
1002 			 */
1003 			if (PV_IS_EXEC_P(pg->mdpage.pvh_attrs)) {
1004 				if (SLIST_EMPTY(&pg->mdpage.pvh_list)) {
1005 					pg->mdpage.pvh_attrs &= ~PVF_EXEC;
1006 					PMAPCOUNT(exec_discarded_unmap);
1007 				} else {
1008 					pmap_syncicache_page(pg);
1009 					PMAPCOUNT(exec_synced_unmap);
1010 				}
1011 			}
1012 #endif /* PMAP_CACHE_VIPT */
1013 			break;
1014 		}
1015 		prevptr = &SLIST_NEXT(pv, pv_link);	/* previous pointer */
1016 		pv = *prevptr;				/* advance */
1017 	}
1018 
1019 #ifdef PMAP_CACHE_VIPT
1020 	/*
1021 	 * If we no longer have a WRITEABLE KENTRY at the head of list,
1022 	 * clear the KMOD attribute from the page.
1023 	 */
1024 	if (SLIST_FIRST(&pg->mdpage.pvh_list) == NULL
1025 	    || (SLIST_FIRST(&pg->mdpage.pvh_list)->pv_flags & PVF_KWRITE) != PVF_KWRITE)
1026 		pg->mdpage.pvh_attrs &= ~PVF_KMOD;
1027 
1028 	/*
1029 	 * If this was a writeable page and there are no more writeable
1030 	 * mappings (ignoring KMPAGE), clear the WRITE flag and writeback
1031 	 * the contents to memory.
1032 	 */
1033 	if (pg->mdpage.krw_mappings + pg->mdpage.urw_mappings == 0)
1034 		pg->mdpage.pvh_attrs &= ~PVF_WRITE;
1035 	KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
1036 #endif /* PMAP_CACHE_VIPT */
1037 
1038 	return(pv);				/* return removed pv */
1039 }
1040 
1041 /*
1042  *
1043  * pmap_modify_pv: Update pv flags
1044  *
1045  * => caller should hold lock on vm_page [so that attrs can be adjusted]
1046  * => caller should NOT adjust pmap's wire_count
1047  * => caller must call pmap_vac_me_harder() if writable status of a page
1048  *    may have changed.
1049  * => we return the old flags
1050  *
1051  * Modify a physical-virtual mapping in the pv table
1052  */
1053 static u_int
1054 pmap_modify_pv(struct vm_page *pg, pmap_t pm, vaddr_t va,
1055     u_int clr_mask, u_int set_mask)
1056 {
1057 	struct pv_entry *npv;
1058 	u_int flags, oflags;
1059 
1060 	KASSERT((clr_mask & PVF_KENTRY) == 0);
1061 	KASSERT((set_mask & PVF_KENTRY) == 0);
1062 
1063 	if ((npv = pmap_find_pv(pg, pm, va)) == NULL)
1064 		return (0);
1065 
1066 	NPDEBUG(PDB_PVDUMP,
1067 	    printf("pmap_modify_pv: pm %p, pg %p, clr 0x%x, set 0x%x, flags 0x%x\n", pm, pg, clr_mask, set_mask, npv->pv_flags));
1068 
1069 	/*
1070 	 * There is at least one VA mapping this page.
1071 	 */
1072 
1073 	if (clr_mask & (PVF_REF | PVF_MOD)) {
1074 		pg->mdpage.pvh_attrs |= set_mask & (PVF_REF | PVF_MOD);
1075 #ifdef PMAP_CACHE_VIPT
1076 		if ((pg->mdpage.pvh_attrs & (PVF_DMOD|PVF_NC)) != PVF_NC)
1077 			pg->mdpage.pvh_attrs |= PVF_DIRTY;
1078 		KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
1079 #endif
1080 	}
1081 
1082 	oflags = npv->pv_flags;
1083 	npv->pv_flags = flags = (oflags & ~clr_mask) | set_mask;
1084 
1085 	if ((flags ^ oflags) & PVF_WIRED) {
1086 		if (flags & PVF_WIRED)
1087 			++pm->pm_stats.wired_count;
1088 		else
1089 			--pm->pm_stats.wired_count;
1090 	}
1091 
1092 	if ((flags ^ oflags) & PVF_WRITE) {
1093 		if (pm == pmap_kernel()) {
1094 			if (flags & PVF_WRITE) {
1095 				pg->mdpage.krw_mappings++;
1096 				pg->mdpage.kro_mappings--;
1097 			} else {
1098 				pg->mdpage.kro_mappings++;
1099 				pg->mdpage.krw_mappings--;
1100 			}
1101 		} else {
1102 			if (flags & PVF_WRITE) {
1103 				pg->mdpage.urw_mappings++;
1104 				pg->mdpage.uro_mappings--;
1105 			} else {
1106 				pg->mdpage.uro_mappings++;
1107 				pg->mdpage.urw_mappings--;
1108 			}
1109 		}
1110 	}
1111 #ifdef PMAP_CACHE_VIPT
1112 	if (pg->mdpage.urw_mappings + pg->mdpage.krw_mappings == 0)
1113 		pg->mdpage.pvh_attrs &= ~PVF_WRITE;
1114 	/*
1115 	 * We have two cases here: the first is from enter_pv (new exec
1116 	 * page), the second is a combined pmap_remove_pv/pmap_enter_pv.
1117 	 * Since in latter, pmap_enter_pv won't do anything, we just have
1118 	 * to do what pmap_remove_pv would do.
1119 	 */
1120 	if ((PV_IS_EXEC_P(flags) && !PV_IS_EXEC_P(pg->mdpage.pvh_attrs))
1121 	    || (PV_IS_EXEC_P(pg->mdpage.pvh_attrs)
1122 		|| (!(flags & PVF_WRITE) && (oflags & PVF_WRITE)))) {
1123 		pmap_syncicache_page(pg);
1124 		PMAPCOUNT(exec_synced_remap);
1125 	}
1126 	KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
1127 #endif
1128 
1129 	PMAPCOUNT(remappings);
1130 
1131 	return (oflags);
1132 }
1133 
1134 /*
1135  * Allocate an L1 translation table for the specified pmap.
1136  * This is called at pmap creation time.
1137  */
1138 static void
1139 pmap_alloc_l1(pmap_t pm)
1140 {
1141 	struct l1_ttable *l1;
1142 	u_int8_t domain;
1143 
1144 	/*
1145 	 * Remove the L1 at the head of the LRU list
1146 	 */
1147 	simple_lock(&l1_lru_lock);
1148 	l1 = TAILQ_FIRST(&l1_lru_list);
1149 	KDASSERT(l1 != NULL);
1150 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1151 
1152 	/*
1153 	 * Pick the first available domain number, and update
1154 	 * the link to the next number.
1155 	 */
1156 	domain = l1->l1_domain_first;
1157 	l1->l1_domain_first = l1->l1_domain_free[domain];
1158 
1159 	/*
1160 	 * If there are still free domain numbers in this L1,
1161 	 * put it back on the TAIL of the LRU list.
1162 	 */
1163 	if (++l1->l1_domain_use_count < PMAP_DOMAINS)
1164 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1165 
1166 	simple_unlock(&l1_lru_lock);
1167 
1168 	/*
1169 	 * Fix up the relevant bits in the pmap structure
1170 	 */
1171 	pm->pm_l1 = l1;
1172 	pm->pm_domain = domain;
1173 }
1174 
1175 /*
1176  * Free an L1 translation table.
1177  * This is called at pmap destruction time.
1178  */
1179 static void
1180 pmap_free_l1(pmap_t pm)
1181 {
1182 	struct l1_ttable *l1 = pm->pm_l1;
1183 
1184 	simple_lock(&l1_lru_lock);
1185 
1186 	/*
1187 	 * If this L1 is currently on the LRU list, remove it.
1188 	 */
1189 	if (l1->l1_domain_use_count < PMAP_DOMAINS)
1190 		TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1191 
1192 	/*
1193 	 * Free up the domain number which was allocated to the pmap
1194 	 */
1195 	l1->l1_domain_free[pm->pm_domain] = l1->l1_domain_first;
1196 	l1->l1_domain_first = pm->pm_domain;
1197 	l1->l1_domain_use_count--;
1198 
1199 	/*
1200 	 * The L1 now must have at least 1 free domain, so add
1201 	 * it back to the LRU list. If the use count is zero,
1202 	 * put it at the head of the list, otherwise it goes
1203 	 * to the tail.
1204 	 */
1205 	if (l1->l1_domain_use_count == 0)
1206 		TAILQ_INSERT_HEAD(&l1_lru_list, l1, l1_lru);
1207 	else
1208 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1209 
1210 	simple_unlock(&l1_lru_lock);
1211 }
1212 
1213 static inline void
1214 pmap_use_l1(pmap_t pm)
1215 {
1216 	struct l1_ttable *l1;
1217 
1218 	/*
1219 	 * Do nothing if we're in interrupt context.
1220 	 * Access to an L1 by the kernel pmap must not affect
1221 	 * the LRU list.
1222 	 */
1223 	if (cpu_intr_p() || pm == pmap_kernel())
1224 		return;
1225 
1226 	l1 = pm->pm_l1;
1227 
1228 	/*
1229 	 * If the L1 is not currently on the LRU list, just return
1230 	 */
1231 	if (l1->l1_domain_use_count == PMAP_DOMAINS)
1232 		return;
1233 
1234 	simple_lock(&l1_lru_lock);
1235 
1236 	/*
1237 	 * Check the use count again, now that we've acquired the lock
1238 	 */
1239 	if (l1->l1_domain_use_count == PMAP_DOMAINS) {
1240 		simple_unlock(&l1_lru_lock);
1241 		return;
1242 	}
1243 
1244 	/*
1245 	 * Move the L1 to the back of the LRU list
1246 	 */
1247 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1248 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1249 
1250 	simple_unlock(&l1_lru_lock);
1251 }
1252 
1253 /*
1254  * void pmap_free_l2_ptp(pt_entry_t *, paddr_t *)
1255  *
1256  * Free an L2 descriptor table.
1257  */
1258 static inline void
1259 #ifndef PMAP_INCLUDE_PTE_SYNC
1260 pmap_free_l2_ptp(pt_entry_t *l2, paddr_t pa)
1261 #else
1262 pmap_free_l2_ptp(bool need_sync, pt_entry_t *l2, paddr_t pa)
1263 #endif
1264 {
1265 #ifdef PMAP_INCLUDE_PTE_SYNC
1266 #ifdef PMAP_CACHE_VIVT
1267 	/*
1268 	 * Note: With a write-back cache, we may need to sync this
1269 	 * L2 table before re-using it.
1270 	 * This is because it may have belonged to a non-current
1271 	 * pmap, in which case the cache syncs would have been
1272 	 * skipped for the pages that were being unmapped. If the
1273 	 * L2 table were then to be immediately re-allocated to
1274 	 * the *current* pmap, it may well contain stale mappings
1275 	 * which have not yet been cleared by a cache write-back
1276 	 * and so would still be visible to the mmu.
1277 	 */
1278 	if (need_sync)
1279 		PTE_SYNC_RANGE(l2, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1280 #endif /* PMAP_CACHE_VIVT */
1281 #endif /* PMAP_INCLUDE_PTE_SYNC */
1282 	pool_cache_put_paddr(&pmap_l2ptp_cache, (void *)l2, pa);
1283 }
1284 
1285 /*
1286  * Returns a pointer to the L2 bucket associated with the specified pmap
1287  * and VA, or NULL if no L2 bucket exists for the address.
1288  */
1289 static inline struct l2_bucket *
1290 pmap_get_l2_bucket(pmap_t pm, vaddr_t va)
1291 {
1292 	struct l2_dtable *l2;
1293 	struct l2_bucket *l2b;
1294 	u_short l1idx;
1295 
1296 	l1idx = L1_IDX(va);
1297 
1298 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL ||
1299 	    (l2b = &l2->l2_bucket[L2_BUCKET(l1idx)])->l2b_kva == NULL)
1300 		return (NULL);
1301 
1302 	return (l2b);
1303 }
1304 
1305 /*
1306  * Returns a pointer to the L2 bucket associated with the specified pmap
1307  * and VA.
1308  *
1309  * If no L2 bucket exists, perform the necessary allocations to put an L2
1310  * bucket/page table in place.
1311  *
1312  * Note that if a new L2 bucket/page was allocated, the caller *must*
1313  * increment the bucket occupancy counter appropriately *before*
1314  * releasing the pmap's lock to ensure no other thread or cpu deallocates
1315  * the bucket/page in the meantime.
1316  */
1317 static struct l2_bucket *
1318 pmap_alloc_l2_bucket(pmap_t pm, vaddr_t va)
1319 {
1320 	struct l2_dtable *l2;
1321 	struct l2_bucket *l2b;
1322 	u_short l1idx;
1323 
1324 	l1idx = L1_IDX(va);
1325 
1326 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
1327 		/*
1328 		 * No mapping at this address, as there is
1329 		 * no entry in the L1 table.
1330 		 * Need to allocate a new l2_dtable.
1331 		 */
1332 		if ((l2 = pmap_alloc_l2_dtable()) == NULL)
1333 			return (NULL);
1334 
1335 		/*
1336 		 * Link it into the parent pmap
1337 		 */
1338 		pm->pm_l2[L2_IDX(l1idx)] = l2;
1339 	}
1340 
1341 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
1342 
1343 	/*
1344 	 * Fetch pointer to the L2 page table associated with the address.
1345 	 */
1346 	if (l2b->l2b_kva == NULL) {
1347 		pt_entry_t *ptep;
1348 
1349 		/*
1350 		 * No L2 page table has been allocated. Chances are, this
1351 		 * is because we just allocated the l2_dtable, above.
1352 		 */
1353 		if ((ptep = pmap_alloc_l2_ptp(&l2b->l2b_phys)) == NULL) {
1354 			/*
1355 			 * Oops, no more L2 page tables available at this
1356 			 * time. We may need to deallocate the l2_dtable
1357 			 * if we allocated a new one above.
1358 			 */
1359 			if (l2->l2_occupancy == 0) {
1360 				pm->pm_l2[L2_IDX(l1idx)] = NULL;
1361 				pmap_free_l2_dtable(l2);
1362 			}
1363 			return (NULL);
1364 		}
1365 
1366 		l2->l2_occupancy++;
1367 		l2b->l2b_kva = ptep;
1368 		l2b->l2b_l1idx = l1idx;
1369 	}
1370 
1371 	return (l2b);
1372 }
1373 
1374 /*
1375  * One or more mappings in the specified L2 descriptor table have just been
1376  * invalidated.
1377  *
1378  * Garbage collect the metadata and descriptor table itself if necessary.
1379  *
1380  * The pmap lock must be acquired when this is called (not necessary
1381  * for the kernel pmap).
1382  */
1383 static void
1384 pmap_free_l2_bucket(pmap_t pm, struct l2_bucket *l2b, u_int count)
1385 {
1386 	struct l2_dtable *l2;
1387 	pd_entry_t *pl1pd, l1pd;
1388 	pt_entry_t *ptep;
1389 	u_short l1idx;
1390 
1391 	KDASSERT(count <= l2b->l2b_occupancy);
1392 
1393 	/*
1394 	 * Update the bucket's reference count according to how many
1395 	 * PTEs the caller has just invalidated.
1396 	 */
1397 	l2b->l2b_occupancy -= count;
1398 
1399 	/*
1400 	 * Note:
1401 	 *
1402 	 * Level 2 page tables allocated to the kernel pmap are never freed
1403 	 * as that would require checking all Level 1 page tables and
1404 	 * removing any references to the Level 2 page table. See also the
1405 	 * comment elsewhere about never freeing bootstrap L2 descriptors.
1406 	 *
1407 	 * We make do with just invalidating the mapping in the L2 table.
1408 	 *
1409 	 * This isn't really a big deal in practice and, in fact, leads
1410 	 * to a performance win over time as we don't need to continually
1411 	 * alloc/free.
1412 	 */
1413 	if (l2b->l2b_occupancy > 0 || pm == pmap_kernel())
1414 		return;
1415 
1416 	/*
1417 	 * There are no more valid mappings in this level 2 page table.
1418 	 * Go ahead and NULL-out the pointer in the bucket, then
1419 	 * free the page table.
1420 	 */
1421 	l1idx = l2b->l2b_l1idx;
1422 	ptep = l2b->l2b_kva;
1423 	l2b->l2b_kva = NULL;
1424 
1425 	pl1pd = &pm->pm_l1->l1_kva[l1idx];
1426 
1427 	/*
1428 	 * If the L1 slot matches the pmap's domain
1429 	 * number, then invalidate it.
1430 	 */
1431 	l1pd = *pl1pd & (L1_TYPE_MASK | L1_C_DOM_MASK);
1432 	if (l1pd == (L1_C_DOM(pm->pm_domain) | L1_TYPE_C)) {
1433 		*pl1pd = 0;
1434 		PTE_SYNC(pl1pd);
1435 	}
1436 
1437 	/*
1438 	 * Release the L2 descriptor table back to the pool cache.
1439 	 */
1440 #ifndef PMAP_INCLUDE_PTE_SYNC
1441 	pmap_free_l2_ptp(ptep, l2b->l2b_phys);
1442 #else
1443 	pmap_free_l2_ptp(!pmap_is_cached(pm), ptep, l2b->l2b_phys);
1444 #endif
1445 
1446 	/*
1447 	 * Update the reference count in the associated l2_dtable
1448 	 */
1449 	l2 = pm->pm_l2[L2_IDX(l1idx)];
1450 	if (--l2->l2_occupancy > 0)
1451 		return;
1452 
1453 	/*
1454 	 * There are no more valid mappings in any of the Level 1
1455 	 * slots managed by this l2_dtable. Go ahead and NULL-out
1456 	 * the pointer in the parent pmap and free the l2_dtable.
1457 	 */
1458 	pm->pm_l2[L2_IDX(l1idx)] = NULL;
1459 	pmap_free_l2_dtable(l2);
1460 }
1461 
1462 /*
1463  * Pool cache constructors for L2 descriptor tables, metadata and pmap
1464  * structures.
1465  */
1466 static int
1467 pmap_l2ptp_ctor(void *arg, void *v, int flags)
1468 {
1469 #ifndef PMAP_INCLUDE_PTE_SYNC
1470 	struct l2_bucket *l2b;
1471 	pt_entry_t *ptep, pte;
1472 	vaddr_t va = (vaddr_t)v & ~PGOFSET;
1473 
1474 	/*
1475 	 * The mappings for these page tables were initially made using
1476 	 * pmap_kenter_pa() by the pool subsystem. Therefore, the cache-
1477 	 * mode will not be right for page table mappings. To avoid
1478 	 * polluting the pmap_kenter_pa() code with a special case for
1479 	 * page tables, we simply fix up the cache-mode here if it's not
1480 	 * correct.
1481 	 */
1482 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
1483 	KDASSERT(l2b != NULL);
1484 	ptep = &l2b->l2b_kva[l2pte_index(va)];
1485 	pte = *ptep;
1486 
1487 	if ((pte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
1488 		/*
1489 		 * Page tables must have the cache-mode set to Write-Thru.
1490 		 */
1491 		*ptep = (pte & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode_pt;
1492 		PTE_SYNC(ptep);
1493 		cpu_tlb_flushD_SE(va);
1494 		cpu_cpwait();
1495 	}
1496 #endif
1497 
1498 	memset(v, 0, L2_TABLE_SIZE_REAL);
1499 	PTE_SYNC_RANGE(v, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1500 	return (0);
1501 }
1502 
1503 static int
1504 pmap_l2dtable_ctor(void *arg, void *v, int flags)
1505 {
1506 
1507 	memset(v, 0, sizeof(struct l2_dtable));
1508 	return (0);
1509 }
1510 
1511 static int
1512 pmap_pmap_ctor(void *arg, void *v, int flags)
1513 {
1514 
1515 	memset(v, 0, sizeof(struct pmap));
1516 	return (0);
1517 }
1518 
1519 static void
1520 pmap_pinit(pmap_t pm)
1521 {
1522 	struct l2_bucket *l2b;
1523 
1524 	if (vector_page < KERNEL_BASE) {
1525 		/*
1526 		 * Map the vector page.
1527 		 */
1528 		pmap_enter(pm, vector_page, systempage.pv_pa,
1529 		    VM_PROT_READ, VM_PROT_READ | PMAP_WIRED);
1530 		pmap_update(pm);
1531 
1532 		pm->pm_pl1vec = &pm->pm_l1->l1_kva[L1_IDX(vector_page)];
1533 		l2b = pmap_get_l2_bucket(pm, vector_page);
1534 		KDASSERT(l2b != NULL);
1535 		pm->pm_l1vec = l2b->l2b_phys | L1_C_PROTO |
1536 		    L1_C_DOM(pm->pm_domain);
1537 	} else
1538 		pm->pm_pl1vec = NULL;
1539 }
1540 
1541 #ifdef PMAP_CACHE_VIVT
1542 /*
1543  * Since we have a virtually indexed cache, we may need to inhibit caching if
1544  * there is more than one mapping and at least one of them is writable.
1545  * Since we purge the cache on every context switch, we only need to check for
1546  * other mappings within the same pmap, or kernel_pmap.
1547  * This function is also called when a page is unmapped, to possibly reenable
1548  * caching on any remaining mappings.
1549  *
1550  * The code implements the following logic, where:
1551  *
1552  * KW = # of kernel read/write pages
1553  * KR = # of kernel read only pages
1554  * UW = # of user read/write pages
1555  * UR = # of user read only pages
1556  *
1557  * KC = kernel mapping is cacheable
1558  * UC = user mapping is cacheable
1559  *
1560  *               KW=0,KR=0  KW=0,KR>0  KW=1,KR=0  KW>1,KR>=0
1561  *             +---------------------------------------------
1562  * UW=0,UR=0   | ---        KC=1       KC=1       KC=0
1563  * UW=0,UR>0   | UC=1       KC=1,UC=1  KC=0,UC=0  KC=0,UC=0
1564  * UW=1,UR=0   | UC=1       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1565  * UW>1,UR>=0  | UC=0       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1566  */
1567 
1568 static const int pmap_vac_flags[4][4] = {
1569 	{-1,		0,		0,		PVF_KNC},
1570 	{0,		0,		PVF_NC,		PVF_NC},
1571 	{0,		PVF_NC,		PVF_NC,		PVF_NC},
1572 	{PVF_UNC,	PVF_NC,		PVF_NC,		PVF_NC}
1573 };
1574 
1575 static inline int
1576 pmap_get_vac_flags(const struct vm_page *pg)
1577 {
1578 	int kidx, uidx;
1579 
1580 	kidx = 0;
1581 	if (pg->mdpage.kro_mappings || pg->mdpage.krw_mappings > 1)
1582 		kidx |= 1;
1583 	if (pg->mdpage.krw_mappings)
1584 		kidx |= 2;
1585 
1586 	uidx = 0;
1587 	if (pg->mdpage.uro_mappings || pg->mdpage.urw_mappings > 1)
1588 		uidx |= 1;
1589 	if (pg->mdpage.urw_mappings)
1590 		uidx |= 2;
1591 
1592 	return (pmap_vac_flags[uidx][kidx]);
1593 }
1594 
1595 static inline void
1596 pmap_vac_me_harder(struct vm_page *pg, pmap_t pm, vaddr_t va)
1597 {
1598 	int nattr;
1599 
1600 	nattr = pmap_get_vac_flags(pg);
1601 
1602 	if (nattr < 0) {
1603 		pg->mdpage.pvh_attrs &= ~PVF_NC;
1604 		return;
1605 	}
1606 
1607 	if (nattr == 0 && (pg->mdpage.pvh_attrs & PVF_NC) == 0)
1608 		return;
1609 
1610 	if (pm == pmap_kernel())
1611 		pmap_vac_me_kpmap(pg, pm, va);
1612 	else
1613 		pmap_vac_me_user(pg, pm, va);
1614 
1615 	pg->mdpage.pvh_attrs = (pg->mdpage.pvh_attrs & ~PVF_NC) | nattr;
1616 }
1617 
1618 static void
1619 pmap_vac_me_kpmap(struct vm_page *pg, pmap_t pm, vaddr_t va)
1620 {
1621 	u_int u_cacheable, u_entries;
1622 	struct pv_entry *pv;
1623 	pmap_t last_pmap = pm;
1624 
1625 	/*
1626 	 * Pass one, see if there are both kernel and user pmaps for
1627 	 * this page.  Calculate whether there are user-writable or
1628 	 * kernel-writable pages.
1629 	 */
1630 	u_cacheable = 0;
1631 	SLIST_FOREACH(pv, &pg->mdpage.pvh_list, pv_link) {
1632 		if (pv->pv_pmap != pm && (pv->pv_flags & PVF_NC) == 0)
1633 			u_cacheable++;
1634 	}
1635 
1636 	u_entries = pg->mdpage.urw_mappings + pg->mdpage.uro_mappings;
1637 
1638 	/*
1639 	 * We know we have just been updating a kernel entry, so if
1640 	 * all user pages are already cacheable, then there is nothing
1641 	 * further to do.
1642 	 */
1643 	if (pg->mdpage.k_mappings == 0 && u_cacheable == u_entries)
1644 		return;
1645 
1646 	if (u_entries) {
1647 		/*
1648 		 * Scan over the list again, for each entry, if it
1649 		 * might not be set correctly, call pmap_vac_me_user
1650 		 * to recalculate the settings.
1651 		 */
1652 		SLIST_FOREACH(pv, &pg->mdpage.pvh_list, pv_link) {
1653 			/*
1654 			 * We know kernel mappings will get set
1655 			 * correctly in other calls.  We also know
1656 			 * that if the pmap is the same as last_pmap
1657 			 * then we've just handled this entry.
1658 			 */
1659 			if (pv->pv_pmap == pm || pv->pv_pmap == last_pmap)
1660 				continue;
1661 
1662 			/*
1663 			 * If there are kernel entries and this page
1664 			 * is writable but non-cacheable, then we can
1665 			 * skip this entry also.
1666 			 */
1667 			if (pg->mdpage.k_mappings &&
1668 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) ==
1669 			    (PVF_NC | PVF_WRITE))
1670 				continue;
1671 
1672 			/*
1673 			 * Similarly if there are no kernel-writable
1674 			 * entries and the page is already
1675 			 * read-only/cacheable.
1676 			 */
1677 			if (pg->mdpage.krw_mappings == 0 &&
1678 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 0)
1679 				continue;
1680 
1681 			/*
1682 			 * For some of the remaining cases, we know
1683 			 * that we must recalculate, but for others we
1684 			 * can't tell if they are correct or not, so
1685 			 * we recalculate anyway.
1686 			 */
1687 			pmap_vac_me_user(pg, (last_pmap = pv->pv_pmap), 0);
1688 		}
1689 
1690 		if (pg->mdpage.k_mappings == 0)
1691 			return;
1692 	}
1693 
1694 	pmap_vac_me_user(pg, pm, va);
1695 }
1696 
1697 static void
1698 pmap_vac_me_user(struct vm_page *pg, pmap_t pm, vaddr_t va)
1699 {
1700 	pmap_t kpmap = pmap_kernel();
1701 	struct pv_entry *pv, *npv = NULL;
1702 	struct l2_bucket *l2b;
1703 	pt_entry_t *ptep, pte;
1704 	u_int entries = 0;
1705 	u_int writable = 0;
1706 	u_int cacheable_entries = 0;
1707 	u_int kern_cacheable = 0;
1708 	u_int other_writable = 0;
1709 
1710 	/*
1711 	 * Count mappings and writable mappings in this pmap.
1712 	 * Include kernel mappings as part of our own.
1713 	 * Keep a pointer to the first one.
1714 	 */
1715 	npv = NULL;
1716 	SLIST_FOREACH(pv, &pg->mdpage.pvh_list, pv_link) {
1717 		/* Count mappings in the same pmap */
1718 		if (pm == pv->pv_pmap || kpmap == pv->pv_pmap) {
1719 			if (entries++ == 0)
1720 				npv = pv;
1721 
1722 			/* Cacheable mappings */
1723 			if ((pv->pv_flags & PVF_NC) == 0) {
1724 				cacheable_entries++;
1725 				if (kpmap == pv->pv_pmap)
1726 					kern_cacheable++;
1727 			}
1728 
1729 			/* Writable mappings */
1730 			if (pv->pv_flags & PVF_WRITE)
1731 				++writable;
1732 		} else
1733 		if (pv->pv_flags & PVF_WRITE)
1734 			other_writable = 1;
1735 	}
1736 
1737 	/*
1738 	 * Enable or disable caching as necessary.
1739 	 * Note: the first entry might be part of the kernel pmap,
1740 	 * so we can't assume this is indicative of the state of the
1741 	 * other (maybe non-kpmap) entries.
1742 	 */
1743 	if ((entries > 1 && writable) ||
1744 	    (entries > 0 && pm == kpmap && other_writable)) {
1745 		if (cacheable_entries == 0)
1746 			return;
1747 
1748 		for (pv = npv; pv; pv = SLIST_NEXT(pv, pv_link)) {
1749 			if ((pm != pv->pv_pmap && kpmap != pv->pv_pmap) ||
1750 			    (pv->pv_flags & PVF_NC))
1751 				continue;
1752 
1753 			pv->pv_flags |= PVF_NC;
1754 
1755 			l2b = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
1756 			KDASSERT(l2b != NULL);
1757 			ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1758 			pte = *ptep & ~L2_S_CACHE_MASK;
1759 
1760 			if ((va != pv->pv_va || pm != pv->pv_pmap) &&
1761 			    l2pte_valid(pte)) {
1762 				if (PV_BEEN_EXECD(pv->pv_flags)) {
1763 #ifdef PMAP_CACHE_VIVT
1764 					pmap_idcache_wbinv_range(pv->pv_pmap,
1765 					    pv->pv_va, PAGE_SIZE);
1766 #endif
1767 					pmap_tlb_flushID_SE(pv->pv_pmap,
1768 					    pv->pv_va);
1769 				} else
1770 				if (PV_BEEN_REFD(pv->pv_flags)) {
1771 #ifdef PMAP_CACHE_VIVT
1772 					pmap_dcache_wb_range(pv->pv_pmap,
1773 					    pv->pv_va, PAGE_SIZE, true,
1774 					    (pv->pv_flags & PVF_WRITE) == 0);
1775 #endif
1776 					pmap_tlb_flushD_SE(pv->pv_pmap,
1777 					    pv->pv_va);
1778 				}
1779 			}
1780 
1781 			*ptep = pte;
1782 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
1783 		}
1784 		cpu_cpwait();
1785 	} else
1786 	if (entries > cacheable_entries) {
1787 		/*
1788 		 * Turn cacheing back on for some pages.  If it is a kernel
1789 		 * page, only do so if there are no other writable pages.
1790 		 */
1791 		for (pv = npv; pv; pv = SLIST_NEXT(pv, pv_link)) {
1792 			if (!(pv->pv_flags & PVF_NC) || (pm != pv->pv_pmap &&
1793 			    (kpmap != pv->pv_pmap || other_writable)))
1794 				continue;
1795 
1796 			pv->pv_flags &= ~PVF_NC;
1797 
1798 			l2b = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
1799 			KDASSERT(l2b != NULL);
1800 			ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1801 			pte = (*ptep & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode;
1802 
1803 			if (l2pte_valid(pte)) {
1804 				if (PV_BEEN_EXECD(pv->pv_flags)) {
1805 					pmap_tlb_flushID_SE(pv->pv_pmap,
1806 					    pv->pv_va);
1807 				} else
1808 				if (PV_BEEN_REFD(pv->pv_flags)) {
1809 					pmap_tlb_flushD_SE(pv->pv_pmap,
1810 					    pv->pv_va);
1811 				}
1812 			}
1813 
1814 			*ptep = pte;
1815 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
1816 		}
1817 	}
1818 }
1819 #endif
1820 
1821 #ifdef PMAP_CACHE_VIPT
1822 static void
1823 pmap_vac_me_harder(struct vm_page *pg, pmap_t pm, vaddr_t va)
1824 {
1825 	struct pv_entry *pv;
1826 	vaddr_t tst_mask;
1827 	bool bad_alias;
1828 	struct l2_bucket *l2b;
1829 	pt_entry_t *ptep, pte, opte;
1830 	const u_int
1831 	    rw_mappings = pg->mdpage.urw_mappings + pg->mdpage.krw_mappings,
1832 	    ro_mappings = pg->mdpage.uro_mappings + pg->mdpage.kro_mappings;
1833 
1834 	/* do we need to do anything? */
1835 	if (arm_cache_prefer_mask == 0)
1836 		return;
1837 
1838 	NPDEBUG(PDB_VAC, printf("pmap_vac_me_harder: pg=%p, pmap=%p va=%08lx\n",
1839 	    pg, pm, va));
1840 
1841 	KASSERT(!va || pm);
1842 	KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
1843 
1844 	/* Already a conflict? */
1845 	if (__predict_false(pg->mdpage.pvh_attrs & PVF_NC)) {
1846 		/* just an add, things are already non-cached */
1847 		KASSERT(!(pg->mdpage.pvh_attrs & PVF_DIRTY));
1848 		KASSERT(!(pg->mdpage.pvh_attrs & PVF_MULTCLR));
1849 		bad_alias = false;
1850 		if (va) {
1851 			PMAPCOUNT(vac_color_none);
1852 			bad_alias = true;
1853 			KASSERT((rw_mappings == 0) == !(pg->mdpage.pvh_attrs & PVF_WRITE));
1854 			goto fixup;
1855 		}
1856 		pv = SLIST_FIRST(&pg->mdpage.pvh_list);
1857 		/* the list can't be empty because it would be cachable */
1858 		if (pg->mdpage.pvh_attrs & PVF_KMPAGE) {
1859 			tst_mask = pg->mdpage.pvh_attrs;
1860 		} else {
1861 			KASSERT(pv);
1862 			tst_mask = pv->pv_va;
1863 			pv = SLIST_NEXT(pv, pv_link);
1864 		}
1865 		/*
1866 		 * Only check for a bad alias if we have writable mappings.
1867 		 */
1868 		tst_mask &= arm_cache_prefer_mask;
1869 		if (rw_mappings > 0 && arm_cache_prefer_mask) {
1870 			for (; pv && !bad_alias; pv = SLIST_NEXT(pv, pv_link)) {
1871 				/* if there's a bad alias, stop checking. */
1872 				if (tst_mask != (pv->pv_va & arm_cache_prefer_mask))
1873 					bad_alias = true;
1874 			}
1875 			pg->mdpage.pvh_attrs |= PVF_WRITE;
1876 			if (!bad_alias)
1877 				pg->mdpage.pvh_attrs |= PVF_DIRTY;
1878 		} else {
1879 			/*
1880 			 * We have only read-only mappings.  Let's see if there
1881 			 * are multiple colors in use or if we mapped a KMPAGE.
1882 			 * If the latter, we have a bad alias.  If the former,
1883 			 * we need to remember that.
1884 			 */
1885 			for (; pv; pv = SLIST_NEXT(pv, pv_link)) {
1886 				if (tst_mask != (pv->pv_va & arm_cache_prefer_mask)) {
1887 					if (pg->mdpage.pvh_attrs & PVF_KMPAGE)
1888 						bad_alias = true;
1889 					break;
1890 				}
1891 			}
1892 			pg->mdpage.pvh_attrs &= ~PVF_WRITE;
1893 			/*
1894 			 * No KMPAGE and we exited early, so we must have
1895 			 * multiple color mappings.
1896 			 */
1897 			if (!bad_alias && pv != NULL)
1898 				pg->mdpage.pvh_attrs |= PVF_MULTCLR;
1899 		}
1900 
1901 		/* If no conflicting colors, set everything back to cached */
1902 		if (!bad_alias) {
1903 #ifdef DEBUG
1904 			if ((pg->mdpage.pvh_attrs & PVF_WRITE)
1905 			    || ro_mappings < 2) {
1906 				SLIST_FOREACH(pv, &pg->mdpage.pvh_list, pv_link)
1907 					KDASSERT(((tst_mask ^ pv->pv_va) & arm_cache_prefer_mask) == 0);
1908 			}
1909 #endif
1910 			pg->mdpage.pvh_attrs &= (PAGE_SIZE - 1) & ~PVF_NC;
1911 			pg->mdpage.pvh_attrs |= tst_mask | PVF_COLORED;
1912 			/*
1913 			 * Restore DIRTY bit if page is modified
1914 			 */
1915 			if (pg->mdpage.pvh_attrs & PVF_DMOD)
1916 				pg->mdpage.pvh_attrs |= PVF_DIRTY;
1917 			PMAPCOUNT(vac_color_restore);
1918 		} else {
1919 			KASSERT(SLIST_FIRST(&pg->mdpage.pvh_list) != NULL);
1920 			KASSERT(SLIST_NEXT(SLIST_FIRST(&pg->mdpage.pvh_list), pv_link) != NULL);
1921 		}
1922 		KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
1923 		KASSERT((rw_mappings == 0) == !(pg->mdpage.pvh_attrs & PVF_WRITE));
1924 	} else if (!va) {
1925 		KASSERT(arm_cache_prefer_mask == 0 || pmap_is_page_colored_p(pg));
1926 		KASSERT(!(pg->mdpage.pvh_attrs & PVF_WRITE)
1927 		    || (pg->mdpage.pvh_attrs & PVF_DIRTY));
1928 		if (rw_mappings == 0) {
1929 			pg->mdpage.pvh_attrs &= ~PVF_WRITE;
1930 			if (ro_mappings == 1
1931 			    && (pg->mdpage.pvh_attrs & PVF_MULTCLR)) {
1932 				/*
1933 				 * If this is the last readonly mapping
1934 				 * but it doesn't match the current color
1935 				 * for the page, change the current color
1936 				 * to match this last readonly mapping.
1937 				 */
1938 				pv = SLIST_FIRST(&pg->mdpage.pvh_list);
1939 				tst_mask = (pg->mdpage.pvh_attrs ^ pv->pv_va)
1940 				    & arm_cache_prefer_mask;
1941 				if (tst_mask) {
1942 					pg->mdpage.pvh_attrs ^= tst_mask;
1943 					PMAPCOUNT(vac_color_change);
1944 				}
1945 			}
1946 		}
1947 		KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
1948 		KASSERT((rw_mappings == 0) == !(pg->mdpage.pvh_attrs & PVF_WRITE));
1949 		return;
1950 	} else if (!pmap_is_page_colored_p(pg)) {
1951 		/* not colored so we just use its color */
1952 		KASSERT(pg->mdpage.pvh_attrs & (PVF_WRITE|PVF_DIRTY));
1953 		KASSERT(!(pg->mdpage.pvh_attrs & PVF_MULTCLR));
1954 		PMAPCOUNT(vac_color_new);
1955 		pg->mdpage.pvh_attrs &= PAGE_SIZE - 1;
1956 		pg->mdpage.pvh_attrs |= PVF_COLORED
1957 		    | (va & arm_cache_prefer_mask)
1958 		    | (rw_mappings > 0 ? PVF_WRITE : 0);
1959 		KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
1960 		KASSERT((rw_mappings == 0) == !(pg->mdpage.pvh_attrs & PVF_WRITE));
1961 		return;
1962 	} else if (((pg->mdpage.pvh_attrs ^ va) & arm_cache_prefer_mask) == 0) {
1963 		bad_alias = false;
1964 		if (rw_mappings > 0) {
1965 			/*
1966 			 * We now have writeable mappings and if we have
1967 			 * readonly mappings in more than once color, we have
1968 			 * an aliasing problem.  Regardless mark the page as
1969 			 * writeable.
1970 			 */
1971 			if (pg->mdpage.pvh_attrs & PVF_MULTCLR) {
1972 				if (ro_mappings < 2) {
1973 					/*
1974 					 * If we only have less than two
1975 					 * read-only mappings, just flush the
1976 					 * non-primary colors from the cache.
1977 					 */
1978 					pmap_flush_page(pg,
1979 					    PMAP_FLUSH_SECONDARY);
1980 				} else {
1981 					bad_alias = true;
1982 				}
1983 			}
1984 			pg->mdpage.pvh_attrs |= PVF_WRITE;
1985 		}
1986 		/* If no conflicting colors, set everything back to cached */
1987 		if (!bad_alias) {
1988 #ifdef DEBUG
1989 			if (rw_mappings > 0
1990 			    || (pg->mdpage.pvh_attrs & PMAP_KMPAGE)) {
1991 				tst_mask = pg->mdpage.pvh_attrs & arm_cache_prefer_mask;
1992 				SLIST_FOREACH(pv, &pg->mdpage.pvh_list, pv_link)
1993 					KDASSERT(((tst_mask ^ pv->pv_va) & arm_cache_prefer_mask) == 0);
1994 			}
1995 #endif
1996 			if (SLIST_EMPTY(&pg->mdpage.pvh_list))
1997 				PMAPCOUNT(vac_color_reuse);
1998 			else
1999 				PMAPCOUNT(vac_color_ok);
2000 
2001 			/* matching color, just return */
2002 			KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
2003 			KASSERT((rw_mappings == 0) == !(pg->mdpage.pvh_attrs & PVF_WRITE));
2004 			return;
2005 		}
2006 		KASSERT(SLIST_FIRST(&pg->mdpage.pvh_list) != NULL);
2007 		KASSERT(SLIST_NEXT(SLIST_FIRST(&pg->mdpage.pvh_list), pv_link) != NULL);
2008 
2009 		/* color conflict.  evict from cache. */
2010 
2011 		pmap_flush_page(pg, PMAP_FLUSH_PRIMARY);
2012 		pg->mdpage.pvh_attrs &= ~PVF_COLORED;
2013 		pg->mdpage.pvh_attrs |= PVF_NC;
2014 		KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
2015 		KASSERT(!(pg->mdpage.pvh_attrs & PVF_MULTCLR));
2016 		PMAPCOUNT(vac_color_erase);
2017 	} else if (rw_mappings == 0
2018 		   && (pg->mdpage.pvh_attrs & PVF_KMPAGE) == 0) {
2019 		KASSERT((pg->mdpage.pvh_attrs & PVF_WRITE) == 0);
2020 
2021 		/*
2022 		 * If the page has dirty cache lines, clean it.
2023 		 */
2024 		if (pg->mdpage.pvh_attrs & PVF_DIRTY)
2025 			pmap_flush_page(pg, PMAP_CLEAN_PRIMARY);
2026 
2027 		/*
2028 		 * If this is the first remapping (we know that there are no
2029 		 * writeable mappings), then this is a simple color change.
2030 		 * Otherwise this is a seconary r/o mapping, which means
2031 		 * we don't have to do anything.
2032 		 */
2033 		if (ro_mappings == 1) {
2034 			KASSERT(((pg->mdpage.pvh_attrs ^ va) & arm_cache_prefer_mask) != 0);
2035 			pg->mdpage.pvh_attrs &= PAGE_SIZE - 1;
2036 			pg->mdpage.pvh_attrs |= (va & arm_cache_prefer_mask);
2037 			PMAPCOUNT(vac_color_change);
2038 		} else {
2039 			PMAPCOUNT(vac_color_blind);
2040 		}
2041 		pg->mdpage.pvh_attrs |= PVF_MULTCLR;
2042 		KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
2043 		KASSERT((rw_mappings == 0) == !(pg->mdpage.pvh_attrs & PVF_WRITE));
2044 		return;
2045 	} else {
2046 		if (rw_mappings > 0)
2047 			pg->mdpage.pvh_attrs |= PVF_WRITE;
2048 
2049 		/* color conflict.  evict from cache. */
2050 		pmap_flush_page(pg, PMAP_FLUSH_PRIMARY);
2051 
2052 		/* the list can't be empty because this was a enter/modify */
2053 		pv = SLIST_FIRST(&pg->mdpage.pvh_list);
2054 		if ((pg->mdpage.pvh_attrs & PVF_KMPAGE) == 0) {
2055 			KASSERT(pv);
2056 			/*
2057 			 * If there's only one mapped page, change color to the
2058 			 * page's new color and return.  Restore the DIRTY bit
2059 			 * that was erased by pmap_flush_page.
2060 			 */
2061 			if (SLIST_NEXT(pv, pv_link) == NULL) {
2062 				pg->mdpage.pvh_attrs &= PAGE_SIZE - 1;
2063 				pg->mdpage.pvh_attrs |= (va & arm_cache_prefer_mask);
2064 				if (pg->mdpage.pvh_attrs & PVF_DMOD)
2065 					pg->mdpage.pvh_attrs |= PVF_DIRTY;
2066 				PMAPCOUNT(vac_color_change);
2067 				KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
2068 				KASSERT((rw_mappings == 0) == !(pg->mdpage.pvh_attrs & PVF_WRITE));
2069 				KASSERT(!(pg->mdpage.pvh_attrs & PVF_MULTCLR));
2070 				return;
2071 			}
2072 		}
2073 		bad_alias = true;
2074 		pg->mdpage.pvh_attrs &= ~PVF_COLORED;
2075 		pg->mdpage.pvh_attrs |= PVF_NC;
2076 		PMAPCOUNT(vac_color_erase);
2077 		KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
2078 	}
2079 
2080   fixup:
2081 	KASSERT((rw_mappings == 0) == !(pg->mdpage.pvh_attrs & PVF_WRITE));
2082 
2083 	/*
2084 	 * Turn cacheing on/off for all pages.
2085 	 */
2086 	SLIST_FOREACH(pv, &pg->mdpage.pvh_list, pv_link) {
2087 		l2b = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
2088 		KDASSERT(l2b != NULL);
2089 		ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2090 		opte = *ptep;
2091 		pte = opte & ~L2_S_CACHE_MASK;
2092 		if (bad_alias) {
2093 			pv->pv_flags |= PVF_NC;
2094 		} else {
2095 			pv->pv_flags &= ~PVF_NC;
2096 			pte |= pte_l2_s_cache_mode;
2097 		}
2098 
2099 		if (opte == pte)	/* only update is there's a change */
2100 			continue;
2101 
2102 		if (l2pte_valid(pte)) {
2103 			if (PV_BEEN_EXECD(pv->pv_flags)) {
2104 				pmap_tlb_flushID_SE(pv->pv_pmap, pv->pv_va);
2105 			} else if (PV_BEEN_REFD(pv->pv_flags)) {
2106 				pmap_tlb_flushD_SE(pv->pv_pmap, pv->pv_va);
2107 			}
2108 		}
2109 
2110 		*ptep = pte;
2111 		PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
2112 	}
2113 }
2114 #endif	/* PMAP_CACHE_VIPT */
2115 
2116 
2117 /*
2118  * Modify pte bits for all ptes corresponding to the given physical address.
2119  * We use `maskbits' rather than `clearbits' because we're always passing
2120  * constants and the latter would require an extra inversion at run-time.
2121  */
2122 static void
2123 pmap_clearbit(struct vm_page *pg, u_int maskbits)
2124 {
2125 	struct l2_bucket *l2b;
2126 	struct pv_entry *pv;
2127 	pt_entry_t *ptep, npte, opte;
2128 	pmap_t pm;
2129 	vaddr_t va;
2130 	u_int oflags;
2131 #ifdef PMAP_CACHE_VIPT
2132 	const bool want_syncicache = PV_IS_EXEC_P(pg->mdpage.pvh_attrs);
2133 	bool need_syncicache = false;
2134 	bool did_syncicache = false;
2135 	bool need_vac_me_harder = false;
2136 #endif
2137 
2138 	NPDEBUG(PDB_BITS,
2139 	    printf("pmap_clearbit: pg %p (0x%08lx) mask 0x%x\n",
2140 	    pg, VM_PAGE_TO_PHYS(pg), maskbits));
2141 
2142 	PMAP_HEAD_TO_MAP_LOCK();
2143 	simple_lock(&pg->mdpage.pvh_slock);
2144 
2145 #ifdef PMAP_CACHE_VIPT
2146 	/*
2147 	 * If we might want to sync the I-cache and we've modified it,
2148 	 * then we know we definitely need to sync or discard it.
2149 	 */
2150 	if (want_syncicache)
2151 		need_syncicache = pg->mdpage.pvh_attrs & PVF_MOD;
2152 #endif
2153 	/*
2154 	 * Clear saved attributes (modify, reference)
2155 	 */
2156 	pg->mdpage.pvh_attrs &= ~(maskbits & (PVF_MOD | PVF_REF));
2157 
2158 	if (SLIST_EMPTY(&pg->mdpage.pvh_list)) {
2159 #ifdef PMAP_CACHE_VIPT
2160 		if (need_syncicache) {
2161 			/*
2162 			 * No one has it mapped, so just discard it.  The next
2163 			 * exec remapping will cause it to be synced.
2164 			 */
2165 			pg->mdpage.pvh_attrs &= ~PVF_EXEC;
2166 			PMAPCOUNT(exec_discarded_clearbit);
2167 		}
2168 #endif
2169 		simple_unlock(&pg->mdpage.pvh_slock);
2170 		PMAP_HEAD_TO_MAP_UNLOCK();
2171 		return;
2172 	}
2173 
2174 	/*
2175 	 * Loop over all current mappings setting/clearing as appropos
2176 	 */
2177 	SLIST_FOREACH(pv, &pg->mdpage.pvh_list, pv_link) {
2178 		va = pv->pv_va;
2179 		pm = pv->pv_pmap;
2180 		oflags = pv->pv_flags;
2181 		/*
2182 		 * Kernel entries are unmanaged and as such not to be changed.
2183 		 */
2184 		if (oflags & PVF_KENTRY)
2185 			continue;
2186 		pv->pv_flags &= ~maskbits;
2187 
2188 		pmap_acquire_pmap_lock(pm);
2189 
2190 		l2b = pmap_get_l2_bucket(pm, va);
2191 		KDASSERT(l2b != NULL);
2192 
2193 		ptep = &l2b->l2b_kva[l2pte_index(va)];
2194 		npte = opte = *ptep;
2195 
2196 		NPDEBUG(PDB_BITS,
2197 		    printf(
2198 		    "pmap_clearbit: pv %p, pm %p, va 0x%08lx, flag 0x%x\n",
2199 		    pv, pv->pv_pmap, pv->pv_va, oflags));
2200 
2201 		if (maskbits & (PVF_WRITE|PVF_MOD)) {
2202 #ifdef PMAP_CACHE_VIVT
2203 			if ((pv->pv_flags & PVF_NC)) {
2204 				/*
2205 				 * Entry is not cacheable:
2206 				 *
2207 				 * Don't turn caching on again if this is a
2208 				 * modified emulation. This would be
2209 				 * inconsitent with the settings created by
2210 				 * pmap_vac_me_harder(). Otherwise, it's safe
2211 				 * to re-enable cacheing.
2212 				 *
2213 				 * There's no need to call pmap_vac_me_harder()
2214 				 * here: all pages are losing their write
2215 				 * permission.
2216 				 */
2217 				if (maskbits & PVF_WRITE) {
2218 					npte |= pte_l2_s_cache_mode;
2219 					pv->pv_flags &= ~PVF_NC;
2220 				}
2221 			} else
2222 			if (opte & L2_S_PROT_W) {
2223 				/*
2224 				 * Entry is writable/cacheable: check if pmap
2225 				 * is current if it is flush it, otherwise it
2226 				 * won't be in the cache
2227 				 */
2228 				if (PV_BEEN_EXECD(oflags))
2229 					pmap_idcache_wbinv_range(pm, pv->pv_va,
2230 					    PAGE_SIZE);
2231 				else
2232 				if (PV_BEEN_REFD(oflags))
2233 					pmap_dcache_wb_range(pm, pv->pv_va,
2234 					    PAGE_SIZE,
2235 					    (maskbits & PVF_REF) != 0, false);
2236 			}
2237 #endif
2238 
2239 			/* make the pte read only */
2240 			npte &= ~L2_S_PROT_W;
2241 
2242 			if (maskbits & oflags & PVF_WRITE) {
2243 				/*
2244 				 * Keep alias accounting up to date
2245 				 */
2246 				if (pv->pv_pmap == pmap_kernel()) {
2247 					pg->mdpage.krw_mappings--;
2248 					pg->mdpage.kro_mappings++;
2249 				} else {
2250 					pg->mdpage.urw_mappings--;
2251 					pg->mdpage.uro_mappings++;
2252 				}
2253 #ifdef PMAP_CACHE_VIPT
2254 				if (pg->mdpage.urw_mappings + pg->mdpage.krw_mappings == 0)
2255 					pg->mdpage.pvh_attrs &= ~PVF_WRITE;
2256 				if (want_syncicache)
2257 					need_syncicache = true;
2258 				need_vac_me_harder = true;
2259 #endif
2260 			}
2261 		}
2262 
2263 		if (maskbits & PVF_REF) {
2264 			if ((pv->pv_flags & PVF_NC) == 0 &&
2265 			    (maskbits & (PVF_WRITE|PVF_MOD)) == 0 &&
2266 			    l2pte_valid(npte)) {
2267 #ifdef PMAP_CACHE_VIVT
2268 				/*
2269 				 * Check npte here; we may have already
2270 				 * done the wbinv above, and the validity
2271 				 * of the PTE is the same for opte and
2272 				 * npte.
2273 				 */
2274 				/* XXXJRT need idcache_inv_range */
2275 				if (PV_BEEN_EXECD(oflags))
2276 					pmap_idcache_wbinv_range(pm,
2277 					    pv->pv_va, PAGE_SIZE);
2278 				else
2279 				if (PV_BEEN_REFD(oflags))
2280 					pmap_dcache_wb_range(pm,
2281 					    pv->pv_va, PAGE_SIZE,
2282 					    true, true);
2283 #endif
2284 			}
2285 
2286 			/*
2287 			 * Make the PTE invalid so that we will take a
2288 			 * page fault the next time the mapping is
2289 			 * referenced.
2290 			 */
2291 			npte &= ~L2_TYPE_MASK;
2292 			npte |= L2_TYPE_INV;
2293 		}
2294 
2295 		if (npte != opte) {
2296 			*ptep = npte;
2297 			PTE_SYNC(ptep);
2298 			/* Flush the TLB entry if a current pmap. */
2299 			if (PV_BEEN_EXECD(oflags))
2300 				pmap_tlb_flushID_SE(pm, pv->pv_va);
2301 			else
2302 			if (PV_BEEN_REFD(oflags))
2303 				pmap_tlb_flushD_SE(pm, pv->pv_va);
2304 		}
2305 
2306 		pmap_release_pmap_lock(pm);
2307 
2308 		NPDEBUG(PDB_BITS,
2309 		    printf("pmap_clearbit: pm %p va 0x%lx opte 0x%08x npte 0x%08x\n",
2310 		    pm, va, opte, npte));
2311 	}
2312 
2313 #ifdef PMAP_CACHE_VIPT
2314 	/*
2315 	 * If we need to sync the I-cache and we haven't done it yet, do it.
2316 	 */
2317 	if (need_syncicache && !did_syncicache) {
2318 		pmap_syncicache_page(pg);
2319 		PMAPCOUNT(exec_synced_clearbit);
2320 	}
2321 	/*
2322 	 * If we are changing this to read-only, we need to call vac_me_harder
2323 	 * so we can change all the read-only pages to cacheable.  We pretend
2324 	 * this as a page deletion.
2325 	 */
2326 	if (need_vac_me_harder) {
2327 		if (pg->mdpage.pvh_attrs & PVF_NC)
2328 			pmap_vac_me_harder(pg, NULL, 0);
2329 	}
2330 #endif
2331 
2332 	simple_unlock(&pg->mdpage.pvh_slock);
2333 	PMAP_HEAD_TO_MAP_UNLOCK();
2334 }
2335 
2336 /*
2337  * pmap_clean_page()
2338  *
2339  * This is a local function used to work out the best strategy to clean
2340  * a single page referenced by its entry in the PV table. It's used by
2341  * pmap_copy_page, pmap_zero page and maybe some others later on.
2342  *
2343  * Its policy is effectively:
2344  *  o If there are no mappings, we don't bother doing anything with the cache.
2345  *  o If there is one mapping, we clean just that page.
2346  *  o If there are multiple mappings, we clean the entire cache.
2347  *
2348  * So that some functions can be further optimised, it returns 0 if it didn't
2349  * clean the entire cache, or 1 if it did.
2350  *
2351  * XXX One bug in this routine is that if the pv_entry has a single page
2352  * mapped at 0x00000000 a whole cache clean will be performed rather than
2353  * just the 1 page. Since this should not occur in everyday use and if it does
2354  * it will just result in not the most efficient clean for the page.
2355  */
2356 #ifdef PMAP_CACHE_VIVT
2357 static int
2358 pmap_clean_page(struct pv_entry *pv, bool is_src)
2359 {
2360 	pmap_t pm_to_clean = NULL;
2361 	struct pv_entry *npv;
2362 	u_int cache_needs_cleaning = 0;
2363 	u_int flags = 0;
2364 	vaddr_t page_to_clean = 0;
2365 
2366 	if (pv == NULL) {
2367 		/* nothing mapped in so nothing to flush */
2368 		return (0);
2369 	}
2370 
2371 	/*
2372 	 * Since we flush the cache each time we change to a different
2373 	 * user vmspace, we only need to flush the page if it is in the
2374 	 * current pmap.
2375 	 */
2376 
2377 	for (npv = pv; npv; npv = SLIST_NEXT(npv, pv_link)) {
2378 		if (pmap_is_current(npv->pv_pmap)) {
2379 			flags |= npv->pv_flags;
2380 			/*
2381 			 * The page is mapped non-cacheable in
2382 			 * this map.  No need to flush the cache.
2383 			 */
2384 			if (npv->pv_flags & PVF_NC) {
2385 #ifdef DIAGNOSTIC
2386 				if (cache_needs_cleaning)
2387 					panic("pmap_clean_page: "
2388 					    "cache inconsistency");
2389 #endif
2390 				break;
2391 			} else if (is_src && (npv->pv_flags & PVF_WRITE) == 0)
2392 				continue;
2393 			if (cache_needs_cleaning) {
2394 				page_to_clean = 0;
2395 				break;
2396 			} else {
2397 				page_to_clean = npv->pv_va;
2398 				pm_to_clean = npv->pv_pmap;
2399 			}
2400 			cache_needs_cleaning = 1;
2401 		}
2402 	}
2403 
2404 	if (page_to_clean) {
2405 		if (PV_BEEN_EXECD(flags))
2406 			pmap_idcache_wbinv_range(pm_to_clean, page_to_clean,
2407 			    PAGE_SIZE);
2408 		else
2409 			pmap_dcache_wb_range(pm_to_clean, page_to_clean,
2410 			    PAGE_SIZE, !is_src, (flags & PVF_WRITE) == 0);
2411 	} else if (cache_needs_cleaning) {
2412 		pmap_t const pm = curproc->p_vmspace->vm_map.pmap;
2413 
2414 		if (PV_BEEN_EXECD(flags))
2415 			pmap_idcache_wbinv_all(pm);
2416 		else
2417 			pmap_dcache_wbinv_all(pm);
2418 		return (1);
2419 	}
2420 	return (0);
2421 }
2422 #endif
2423 
2424 #ifdef PMAP_CACHE_VIPT
2425 /*
2426  * Sync a page with the I-cache.  Since this is a VIPT, we must pick the
2427  * right cache alias to make sure we flush the right stuff.
2428  */
2429 void
2430 pmap_syncicache_page(struct vm_page *pg)
2431 {
2432 	const vsize_t va_offset = pg->mdpage.pvh_attrs & arm_cache_prefer_mask;
2433 	pt_entry_t * const ptep = &cdst_pte[va_offset >> PGSHIFT];
2434 
2435 	NPDEBUG(PDB_EXEC, printf("pmap_syncicache_page: pg=%p (attrs=%#x)\n",
2436 	    pg, pg->mdpage.pvh_attrs));
2437 	/*
2438 	 * No need to clean the page if it's non-cached.
2439 	 */
2440 	if (pg->mdpage.pvh_attrs & PVF_NC)
2441 		return;
2442 	KASSERT(arm_cache_prefer_mask == 0 || pg->mdpage.pvh_attrs & PVF_COLORED);
2443 
2444 	pmap_tlb_flushID_SE(pmap_kernel(), cdstp + va_offset);
2445 	/*
2446 	 * Set up a PTE with the right coloring to flush existing cache lines.
2447 	 */
2448 	*ptep = L2_S_PROTO |
2449 	    VM_PAGE_TO_PHYS(pg)
2450 	    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE)
2451 	    | pte_l2_s_cache_mode;
2452 	PTE_SYNC(ptep);
2453 
2454 	/*
2455 	 * Flush it.
2456 	 */
2457 	cpu_icache_sync_range(cdstp + va_offset, PAGE_SIZE);
2458 	/*
2459 	 * Unmap the page.
2460 	 */
2461 	*ptep = 0;
2462 	PTE_SYNC(ptep);
2463 	pmap_tlb_flushID_SE(pmap_kernel(), cdstp + va_offset);
2464 
2465 	pg->mdpage.pvh_attrs |= PVF_EXEC;
2466 	PMAPCOUNT(exec_synced);
2467 }
2468 
2469 void
2470 pmap_flush_page(struct vm_page *pg, enum pmap_flush_op flush)
2471 {
2472 	vsize_t va_offset, end_va;
2473 	void (*cf)(vaddr_t, vsize_t);
2474 
2475 	if (arm_cache_prefer_mask == 0)
2476 		return;
2477 
2478 	switch (flush) {
2479 	case PMAP_FLUSH_PRIMARY:
2480 		if (pg->mdpage.pvh_attrs & PVF_MULTCLR) {
2481 			va_offset = 0;
2482 			end_va = arm_cache_prefer_mask;
2483 			pg->mdpage.pvh_attrs &= ~PVF_MULTCLR;
2484 			PMAPCOUNT(vac_flush_lots);
2485 		} else {
2486 			va_offset = pg->mdpage.pvh_attrs & arm_cache_prefer_mask;
2487 			end_va = va_offset;
2488 			PMAPCOUNT(vac_flush_one);
2489 		}
2490 		/*
2491 		 * Mark that the page is no longer dirty.
2492 		 */
2493 		pg->mdpage.pvh_attrs &= ~PVF_DIRTY;
2494 		cf = cpufuncs.cf_idcache_wbinv_range;
2495 		break;
2496 	case PMAP_FLUSH_SECONDARY:
2497 		va_offset = 0;
2498 		end_va = arm_cache_prefer_mask;
2499 		cf = cpufuncs.cf_idcache_wbinv_range;
2500 		pg->mdpage.pvh_attrs &= ~PVF_MULTCLR;
2501 		PMAPCOUNT(vac_flush_lots);
2502 		break;
2503 	case PMAP_CLEAN_PRIMARY:
2504 		va_offset = pg->mdpage.pvh_attrs & arm_cache_prefer_mask;
2505 		end_va = va_offset;
2506 		cf = cpufuncs.cf_dcache_wb_range;
2507 		/*
2508 		 * Mark that the page is no longer dirty.
2509 		 */
2510 		if ((pg->mdpage.pvh_attrs & PVF_DMOD) == 0)
2511 			pg->mdpage.pvh_attrs &= ~PVF_DIRTY;
2512 		PMAPCOUNT(vac_clean_one);
2513 		break;
2514 	default:
2515 		return;
2516 	}
2517 
2518 	KASSERT(!(pg->mdpage.pvh_attrs & PVF_NC));
2519 
2520 	NPDEBUG(PDB_VAC, printf("pmap_flush_page: pg=%p (attrs=%#x)\n",
2521 	    pg, pg->mdpage.pvh_attrs));
2522 
2523 	for (; va_offset <= end_va; va_offset += PAGE_SIZE) {
2524 		const size_t pte_offset = va_offset >> PGSHIFT;
2525 		pt_entry_t * const ptep = &cdst_pte[pte_offset];
2526 		const pt_entry_t oldpte = *ptep;
2527 
2528 		if (flush == PMAP_FLUSH_SECONDARY
2529 		    && va_offset == (pg->mdpage.pvh_attrs & arm_cache_prefer_mask))
2530 			continue;
2531 
2532 		pmap_tlb_flushID_SE(pmap_kernel(), cdstp + va_offset);
2533 		/*
2534 		 * Set up a PTE with the right coloring to flush
2535 		 * existing cache entries.
2536 		 */
2537 		*ptep = L2_S_PROTO
2538 		    | VM_PAGE_TO_PHYS(pg)
2539 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE)
2540 		    | pte_l2_s_cache_mode;
2541 		PTE_SYNC(ptep);
2542 
2543 		/*
2544 		 * Flush it.
2545 		 */
2546 		(*cf)(cdstp + va_offset, PAGE_SIZE);
2547 
2548 		/*
2549 		 * Restore the page table entry since we might have interrupted
2550 		 * pmap_zero_page or pmap_copy_page which was already using
2551 		 * this pte.
2552 		 */
2553 		*ptep = oldpte;
2554 		PTE_SYNC(ptep);
2555 		pmap_tlb_flushID_SE(pmap_kernel(), cdstp + va_offset);
2556 	}
2557 }
2558 #endif /* PMAP_CACHE_VIPT */
2559 
2560 /*
2561  * Routine:	pmap_page_remove
2562  * Function:
2563  *		Removes this physical page from
2564  *		all physical maps in which it resides.
2565  *		Reflects back modify bits to the pager.
2566  */
2567 static void
2568 pmap_page_remove(struct vm_page *pg)
2569 {
2570 	struct l2_bucket *l2b;
2571 	struct pv_entry *pv, *npv, **pvp;
2572 	pmap_t pm;
2573 	pt_entry_t *ptep;
2574 	bool flush;
2575 	u_int flags;
2576 
2577 	NPDEBUG(PDB_FOLLOW,
2578 	    printf("pmap_page_remove: pg %p (0x%08lx)\n", pg,
2579 	    VM_PAGE_TO_PHYS(pg)));
2580 
2581 	PMAP_HEAD_TO_MAP_LOCK();
2582 	simple_lock(&pg->mdpage.pvh_slock);
2583 
2584 	pv = SLIST_FIRST(&pg->mdpage.pvh_list);
2585 	if (pv == NULL) {
2586 #ifdef PMAP_CACHE_VIPT
2587 		/*
2588 		 * We *know* the page contents are about to be replaced.
2589 		 * Discard the exec contents
2590 		 */
2591 		if (PV_IS_EXEC_P(pg->mdpage.pvh_attrs))
2592 			PMAPCOUNT(exec_discarded_page_protect);
2593 		pg->mdpage.pvh_attrs &= ~PVF_EXEC;
2594 		KASSERT((pg->mdpage.urw_mappings + pg->mdpage.krw_mappings == 0) == !(pg->mdpage.pvh_attrs & PVF_WRITE));
2595 #endif
2596 		simple_unlock(&pg->mdpage.pvh_slock);
2597 		PMAP_HEAD_TO_MAP_UNLOCK();
2598 		return;
2599 	}
2600 #ifdef PMAP_CACHE_VIPT
2601 	KASSERT(arm_cache_prefer_mask == 0 || pmap_is_page_colored_p(pg));
2602 #endif
2603 
2604 	/*
2605 	 * Clear alias counts
2606 	 */
2607 #ifdef PMAP_CACHE_VIVT
2608 	pg->mdpage.k_mappings = 0;
2609 #endif
2610 	pg->mdpage.urw_mappings = pg->mdpage.uro_mappings = 0;
2611 
2612 	flush = false;
2613 	flags = 0;
2614 
2615 #ifdef PMAP_CACHE_VIVT
2616 	pmap_clean_page(pv, false);
2617 #endif
2618 
2619 	pvp = &SLIST_FIRST(&pg->mdpage.pvh_list);
2620 	while (pv) {
2621 		pm = pv->pv_pmap;
2622 		npv = SLIST_NEXT(pv, pv_link);
2623 		if (flush == false && pmap_is_current(pm))
2624 			flush = true;
2625 
2626 		if (pm == pmap_kernel()) {
2627 #ifdef PMAP_CACHE_VIPT
2628 			/*
2629 			 * If this was unmanaged mapping, it must be preserved.
2630 			 * Move it back on the list and advance the end-of-list
2631 			 * pointer.
2632 			 */
2633 			if (pv->pv_flags & PVF_KENTRY) {
2634 				*pvp = pv;
2635 				pvp = &SLIST_NEXT(pv, pv_link);
2636 				pv = npv;
2637 				continue;
2638 			}
2639 			if (pv->pv_flags & PVF_WRITE)
2640 				pg->mdpage.krw_mappings--;
2641 			else
2642 				pg->mdpage.kro_mappings--;
2643 #endif
2644 			PMAPCOUNT(kernel_unmappings);
2645 		}
2646 		PMAPCOUNT(unmappings);
2647 
2648 		pmap_acquire_pmap_lock(pm);
2649 
2650 		l2b = pmap_get_l2_bucket(pm, pv->pv_va);
2651 		KDASSERT(l2b != NULL);
2652 
2653 		ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2654 
2655 		/*
2656 		 * Update statistics
2657 		 */
2658 		--pm->pm_stats.resident_count;
2659 
2660 		/* Wired bit */
2661 		if (pv->pv_flags & PVF_WIRED)
2662 			--pm->pm_stats.wired_count;
2663 
2664 		flags |= pv->pv_flags;
2665 
2666 		/*
2667 		 * Invalidate the PTEs.
2668 		 */
2669 		*ptep = 0;
2670 		PTE_SYNC_CURRENT(pm, ptep);
2671 		pmap_free_l2_bucket(pm, l2b, 1);
2672 
2673 		pool_put(&pmap_pv_pool, pv);
2674 		pv = npv;
2675 		/*
2676 		 * if we reach the end of the list and there are still
2677 		 * mappings, they might be able to be cached now.
2678 		 */
2679 		if (pv == NULL) {
2680 			*pvp = NULL;
2681 			if (!SLIST_EMPTY(&pg->mdpage.pvh_list))
2682 				pmap_vac_me_harder(pg, pm, 0);
2683 		}
2684 		pmap_release_pmap_lock(pm);
2685 	}
2686 #ifdef PMAP_CACHE_VIPT
2687 	/*
2688 	 * Its EXEC cache is now gone.
2689 	 */
2690 	if (PV_IS_EXEC_P(pg->mdpage.pvh_attrs))
2691 		PMAPCOUNT(exec_discarded_page_protect);
2692 	pg->mdpage.pvh_attrs &= ~PVF_EXEC;
2693 	KASSERT(pg->mdpage.urw_mappings == 0);
2694 	KASSERT(pg->mdpage.uro_mappings == 0);
2695 	if (pg->mdpage.krw_mappings == 0)
2696 		pg->mdpage.pvh_attrs &= ~PVF_WRITE;
2697 	KASSERT((pg->mdpage.urw_mappings + pg->mdpage.krw_mappings == 0) == !(pg->mdpage.pvh_attrs & PVF_WRITE));
2698 #endif
2699 	simple_unlock(&pg->mdpage.pvh_slock);
2700 	PMAP_HEAD_TO_MAP_UNLOCK();
2701 
2702 	if (flush) {
2703 		/*
2704 		 * Note: We can't use pmap_tlb_flush{I,D}() here since that
2705 		 * would need a subsequent call to pmap_update() to ensure
2706 		 * curpm->pm_cstate.cs_all is reset. Our callers are not
2707 		 * required to do that (see pmap(9)), so we can't modify
2708 		 * the current pmap's state.
2709 		 */
2710 		if (PV_BEEN_EXECD(flags))
2711 			cpu_tlb_flushID();
2712 		else
2713 			cpu_tlb_flushD();
2714 	}
2715 	cpu_cpwait();
2716 }
2717 
2718 /*
2719  * pmap_t pmap_create(void)
2720  *
2721  *      Create a new pmap structure from scratch.
2722  */
2723 pmap_t
2724 pmap_create(void)
2725 {
2726 	pmap_t pm;
2727 
2728 	pm = pool_cache_get(&pmap_cache, PR_WAITOK);
2729 
2730 	UVM_OBJ_INIT(&pm->pm_obj, NULL, 1);
2731 	pm->pm_stats.wired_count = 0;
2732 	pm->pm_stats.resident_count = 1;
2733 	pm->pm_cstate.cs_all = 0;
2734 	pmap_alloc_l1(pm);
2735 
2736 	/*
2737 	 * Note: The pool cache ensures that the pm_l2[] array is already
2738 	 * initialised to zero.
2739 	 */
2740 
2741 	pmap_pinit(pm);
2742 
2743 	LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list);
2744 
2745 	return (pm);
2746 }
2747 
2748 /*
2749  * int pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
2750  *      u_int flags)
2751  *
2752  *      Insert the given physical page (p) at
2753  *      the specified virtual address (v) in the
2754  *      target physical map with the protection requested.
2755  *
2756  *      NB:  This is the only routine which MAY NOT lazy-evaluate
2757  *      or lose information.  That is, this routine must actually
2758  *      insert this page into the given map NOW.
2759  */
2760 int
2761 pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags)
2762 {
2763 	struct l2_bucket *l2b;
2764 	struct vm_page *pg, *opg;
2765 	struct pv_entry *pv;
2766 	pt_entry_t *ptep, npte, opte;
2767 	u_int nflags;
2768 	u_int oflags;
2769 
2770 	NPDEBUG(PDB_ENTER, printf("pmap_enter: pm %p va 0x%lx pa 0x%lx prot %x flag %x\n", pm, va, pa, prot, flags));
2771 
2772 	KDASSERT((flags & PMAP_WIRED) == 0 || (flags & VM_PROT_ALL) != 0);
2773 	KDASSERT(((va | pa) & PGOFSET) == 0);
2774 
2775 	/*
2776 	 * Get a pointer to the page.  Later on in this function, we
2777 	 * test for a managed page by checking pg != NULL.
2778 	 */
2779 	pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
2780 
2781 	nflags = 0;
2782 	if (prot & VM_PROT_WRITE)
2783 		nflags |= PVF_WRITE;
2784 	if (prot & VM_PROT_EXECUTE)
2785 		nflags |= PVF_EXEC;
2786 	if (flags & PMAP_WIRED)
2787 		nflags |= PVF_WIRED;
2788 
2789 	PMAP_MAP_TO_HEAD_LOCK();
2790 	pmap_acquire_pmap_lock(pm);
2791 
2792 	/*
2793 	 * Fetch the L2 bucket which maps this page, allocating one if
2794 	 * necessary for user pmaps.
2795 	 */
2796 	if (pm == pmap_kernel())
2797 		l2b = pmap_get_l2_bucket(pm, va);
2798 	else
2799 		l2b = pmap_alloc_l2_bucket(pm, va);
2800 	if (l2b == NULL) {
2801 		if (flags & PMAP_CANFAIL) {
2802 			pmap_release_pmap_lock(pm);
2803 			PMAP_MAP_TO_HEAD_UNLOCK();
2804 			return (ENOMEM);
2805 		}
2806 		panic("pmap_enter: failed to allocate L2 bucket");
2807 	}
2808 	ptep = &l2b->l2b_kva[l2pte_index(va)];
2809 	opte = *ptep;
2810 	npte = pa;
2811 	oflags = 0;
2812 
2813 	if (opte) {
2814 		/*
2815 		 * There is already a mapping at this address.
2816 		 * If the physical address is different, lookup the
2817 		 * vm_page.
2818 		 */
2819 		if (l2pte_pa(opte) != pa)
2820 			opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
2821 		else
2822 			opg = pg;
2823 	} else
2824 		opg = NULL;
2825 
2826 	if (pg) {
2827 		/*
2828 		 * This is to be a managed mapping.
2829 		 */
2830 		if ((flags & VM_PROT_ALL) ||
2831 		    (pg->mdpage.pvh_attrs & PVF_REF)) {
2832 			/*
2833 			 * - The access type indicates that we don't need
2834 			 *   to do referenced emulation.
2835 			 * OR
2836 			 * - The physical page has already been referenced
2837 			 *   so no need to re-do referenced emulation here.
2838 			 */
2839 			npte |= L2_S_PROTO;
2840 
2841 			nflags |= PVF_REF;
2842 
2843 			if ((prot & VM_PROT_WRITE) != 0 &&
2844 			    ((flags & VM_PROT_WRITE) != 0 ||
2845 			     (pg->mdpage.pvh_attrs & PVF_MOD) != 0)) {
2846 				/*
2847 				 * This is a writable mapping, and the
2848 				 * page's mod state indicates it has
2849 				 * already been modified. Make it
2850 				 * writable from the outset.
2851 				 */
2852 				npte |= L2_S_PROT_W;
2853 				nflags |= PVF_MOD;
2854 			}
2855 		} else {
2856 			/*
2857 			 * Need to do page referenced emulation.
2858 			 */
2859 			npte |= L2_TYPE_INV;
2860 		}
2861 
2862 		npte |= pte_l2_s_cache_mode;
2863 
2864 		if (pg == opg) {
2865 			/*
2866 			 * We're changing the attrs of an existing mapping.
2867 			 */
2868 			simple_lock(&pg->mdpage.pvh_slock);
2869 			oflags = pmap_modify_pv(pg, pm, va,
2870 			    PVF_WRITE | PVF_EXEC | PVF_WIRED |
2871 			    PVF_MOD | PVF_REF, nflags);
2872 			simple_unlock(&pg->mdpage.pvh_slock);
2873 
2874 #ifdef PMAP_CACHE_VIVT
2875 			/*
2876 			 * We may need to flush the cache if we're
2877 			 * doing rw-ro...
2878 			 */
2879 			if (pm->pm_cstate.cs_cache_d &&
2880 			    (oflags & PVF_NC) == 0 &&
2881 			    (opte & L2_S_PROT_W) != 0 &&
2882 			    (prot & VM_PROT_WRITE) == 0)
2883 				cpu_dcache_wb_range(va, PAGE_SIZE);
2884 #endif
2885 		} else {
2886 			/*
2887 			 * New mapping, or changing the backing page
2888 			 * of an existing mapping.
2889 			 */
2890 			if (opg) {
2891 				/*
2892 				 * Replacing an existing mapping with a new one.
2893 				 * It is part of our managed memory so we
2894 				 * must remove it from the PV list
2895 				 */
2896 				simple_lock(&opg->mdpage.pvh_slock);
2897 				pv = pmap_remove_pv(opg, pm, va);
2898 				pmap_vac_me_harder(opg, pm, 0);
2899 				simple_unlock(&opg->mdpage.pvh_slock);
2900 				oflags = pv->pv_flags;
2901 
2902 #ifdef PMAP_CACHE_VIVT
2903 				/*
2904 				 * If the old mapping was valid (ref/mod
2905 				 * emulation creates 'invalid' mappings
2906 				 * initially) then make sure to frob
2907 				 * the cache.
2908 				 */
2909 				if ((oflags & PVF_NC) == 0 &&
2910 				    l2pte_valid(opte)) {
2911 					if (PV_BEEN_EXECD(oflags)) {
2912 						pmap_idcache_wbinv_range(pm, va,
2913 						    PAGE_SIZE);
2914 					} else
2915 					if (PV_BEEN_REFD(oflags)) {
2916 						pmap_dcache_wb_range(pm, va,
2917 						    PAGE_SIZE, true,
2918 						    (oflags & PVF_WRITE) == 0);
2919 					}
2920 				}
2921 #endif
2922 			} else
2923 			if ((pv = pool_get(&pmap_pv_pool, PR_NOWAIT)) == NULL){
2924 				if ((flags & PMAP_CANFAIL) == 0)
2925 					panic("pmap_enter: no pv entries");
2926 
2927 				if (pm != pmap_kernel())
2928 					pmap_free_l2_bucket(pm, l2b, 0);
2929 				pmap_release_pmap_lock(pm);
2930 				PMAP_MAP_TO_HEAD_UNLOCK();
2931 				NPDEBUG(PDB_ENTER,
2932 				    printf("pmap_enter: ENOMEM\n"));
2933 				return (ENOMEM);
2934 			}
2935 
2936 			pmap_enter_pv(pg, pv, pm, va, nflags);
2937 		}
2938 	} else {
2939 		/*
2940 		 * We're mapping an unmanaged page.
2941 		 * These are always readable, and possibly writable, from
2942 		 * the get go as we don't need to track ref/mod status.
2943 		 */
2944 		npte |= L2_S_PROTO;
2945 		if (prot & VM_PROT_WRITE)
2946 			npte |= L2_S_PROT_W;
2947 
2948 		/*
2949 		 * Make sure the vector table is mapped cacheable
2950 		 */
2951 		if (pm != pmap_kernel() && va == vector_page)
2952 			npte |= pte_l2_s_cache_mode;
2953 
2954 		if (opg) {
2955 			/*
2956 			 * Looks like there's an existing 'managed' mapping
2957 			 * at this address.
2958 			 */
2959 			simple_lock(&opg->mdpage.pvh_slock);
2960 			pv = pmap_remove_pv(opg, pm, va);
2961 			pmap_vac_me_harder(opg, pm, 0);
2962 			simple_unlock(&opg->mdpage.pvh_slock);
2963 			oflags = pv->pv_flags;
2964 
2965 #ifdef PMAP_CACHE_VIVT
2966 			if ((oflags & PVF_NC) == 0 && l2pte_valid(opte)) {
2967 				if (PV_BEEN_EXECD(oflags))
2968 					pmap_idcache_wbinv_range(pm, va,
2969 					    PAGE_SIZE);
2970 				else
2971 				if (PV_BEEN_REFD(oflags))
2972 					pmap_dcache_wb_range(pm, va, PAGE_SIZE,
2973 					    true, (oflags & PVF_WRITE) == 0);
2974 			}
2975 #endif
2976 			pool_put(&pmap_pv_pool, pv);
2977 		}
2978 	}
2979 
2980 	/*
2981 	 * Make sure userland mappings get the right permissions
2982 	 */
2983 	if (pm != pmap_kernel() && va != vector_page)
2984 		npte |= L2_S_PROT_U;
2985 
2986 	/*
2987 	 * Keep the stats up to date
2988 	 */
2989 	if (opte == 0) {
2990 		l2b->l2b_occupancy++;
2991 		pm->pm_stats.resident_count++;
2992 	}
2993 
2994 	NPDEBUG(PDB_ENTER,
2995 	    printf("pmap_enter: opte 0x%08x npte 0x%08x\n", opte, npte));
2996 
2997 	/*
2998 	 * If this is just a wiring change, the two PTEs will be
2999 	 * identical, so there's no need to update the page table.
3000 	 */
3001 	if (npte != opte) {
3002 		bool is_cached = pmap_is_cached(pm);
3003 
3004 		*ptep = npte;
3005 		if (is_cached) {
3006 			/*
3007 			 * We only need to frob the cache/tlb if this pmap
3008 			 * is current
3009 			 */
3010 			PTE_SYNC(ptep);
3011 			if (va != vector_page && l2pte_valid(npte)) {
3012 				/*
3013 				 * This mapping is likely to be accessed as
3014 				 * soon as we return to userland. Fix up the
3015 				 * L1 entry to avoid taking another
3016 				 * page/domain fault.
3017 				 */
3018 				pd_entry_t *pl1pd, l1pd;
3019 
3020 				pl1pd = &pm->pm_l1->l1_kva[L1_IDX(va)];
3021 				l1pd = l2b->l2b_phys | L1_C_DOM(pm->pm_domain) |
3022 				    L1_C_PROTO;
3023 				if (*pl1pd != l1pd) {
3024 					*pl1pd = l1pd;
3025 					PTE_SYNC(pl1pd);
3026 				}
3027 			}
3028 		}
3029 
3030 		if (PV_BEEN_EXECD(oflags))
3031 			pmap_tlb_flushID_SE(pm, va);
3032 		else
3033 		if (PV_BEEN_REFD(oflags))
3034 			pmap_tlb_flushD_SE(pm, va);
3035 
3036 		NPDEBUG(PDB_ENTER,
3037 		    printf("pmap_enter: is_cached %d cs 0x%08x\n",
3038 		    is_cached, pm->pm_cstate.cs_all));
3039 
3040 		if (pg != NULL) {
3041 			simple_lock(&pg->mdpage.pvh_slock);
3042 			pmap_vac_me_harder(pg, pm, va);
3043 			simple_unlock(&pg->mdpage.pvh_slock);
3044 		}
3045 	}
3046 #if defined(PMAP_CACHE_VIPT) && defined(DIAGNOSTIC)
3047 	if (pg) {
3048 		simple_lock(&pg->mdpage.pvh_slock);
3049 		KASSERT((pg->mdpage.pvh_attrs & PVF_DMOD) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
3050 		KASSERT(((pg->mdpage.pvh_attrs & PVF_WRITE) == 0) == (pg->mdpage.urw_mappings + pg->mdpage.krw_mappings == 0));
3051 		simple_unlock(&pg->mdpage.pvh_slock);
3052 	}
3053 #endif
3054 
3055 	pmap_release_pmap_lock(pm);
3056 	PMAP_MAP_TO_HEAD_UNLOCK();
3057 
3058 	return (0);
3059 }
3060 
3061 /*
3062  * pmap_remove()
3063  *
3064  * pmap_remove is responsible for nuking a number of mappings for a range
3065  * of virtual address space in the current pmap. To do this efficiently
3066  * is interesting, because in a number of cases a wide virtual address
3067  * range may be supplied that contains few actual mappings. So, the
3068  * optimisations are:
3069  *  1. Skip over hunks of address space for which no L1 or L2 entry exists.
3070  *  2. Build up a list of pages we've hit, up to a maximum, so we can
3071  *     maybe do just a partial cache clean. This path of execution is
3072  *     complicated by the fact that the cache must be flushed _before_
3073  *     the PTE is nuked, being a VAC :-)
3074  *  3. If we're called after UVM calls pmap_remove_all(), we can defer
3075  *     all invalidations until pmap_update(), since pmap_remove_all() has
3076  *     already flushed the cache.
3077  *  4. Maybe later fast-case a single page, but I don't think this is
3078  *     going to make _that_ much difference overall.
3079  */
3080 
3081 #define	PMAP_REMOVE_CLEAN_LIST_SIZE	3
3082 
3083 void
3084 pmap_remove(pmap_t pm, vaddr_t sva, vaddr_t eva)
3085 {
3086 	struct l2_bucket *l2b;
3087 	vaddr_t next_bucket;
3088 	pt_entry_t *ptep;
3089 	u_int cleanlist_idx, total, cnt;
3090 	struct {
3091 		vaddr_t va;
3092 		pt_entry_t *ptep;
3093 	} cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
3094 	u_int mappings, is_exec, is_refd;
3095 
3096 	NPDEBUG(PDB_REMOVE, printf("pmap_do_remove: pmap=%p sva=%08lx "
3097 	    "eva=%08lx\n", pm, sva, eva));
3098 
3099 	/*
3100 	 * we lock in the pmap => pv_head direction
3101 	 */
3102 	PMAP_MAP_TO_HEAD_LOCK();
3103 	pmap_acquire_pmap_lock(pm);
3104 
3105 	if (pm->pm_remove_all || !pmap_is_cached(pm)) {
3106 		cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
3107 		if (pm->pm_cstate.cs_tlb == 0)
3108 			pm->pm_remove_all = true;
3109 	} else
3110 		cleanlist_idx = 0;
3111 
3112 	total = 0;
3113 
3114 	while (sva < eva) {
3115 		/*
3116 		 * Do one L2 bucket's worth at a time.
3117 		 */
3118 		next_bucket = L2_NEXT_BUCKET(sva);
3119 		if (next_bucket > eva)
3120 			next_bucket = eva;
3121 
3122 		l2b = pmap_get_l2_bucket(pm, sva);
3123 		if (l2b == NULL) {
3124 			sva = next_bucket;
3125 			continue;
3126 		}
3127 
3128 		ptep = &l2b->l2b_kva[l2pte_index(sva)];
3129 
3130 		for (mappings = 0; sva < next_bucket; sva += PAGE_SIZE, ptep++){
3131 			struct vm_page *pg;
3132 			pt_entry_t pte;
3133 			paddr_t pa;
3134 
3135 			pte = *ptep;
3136 
3137 			if (pte == 0) {
3138 				/* Nothing here, move along */
3139 				continue;
3140 			}
3141 
3142 			pa = l2pte_pa(pte);
3143 			is_exec = 0;
3144 			is_refd = 1;
3145 
3146 			/*
3147 			 * Update flags. In a number of circumstances,
3148 			 * we could cluster a lot of these and do a
3149 			 * number of sequential pages in one go.
3150 			 */
3151 			if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
3152 				struct pv_entry *pv;
3153 				simple_lock(&pg->mdpage.pvh_slock);
3154 				pv = pmap_remove_pv(pg, pm, sva);
3155 				pmap_vac_me_harder(pg, pm, 0);
3156 				simple_unlock(&pg->mdpage.pvh_slock);
3157 				if (pv != NULL) {
3158 					if (pm->pm_remove_all == false) {
3159 						is_exec =
3160 						   PV_BEEN_EXECD(pv->pv_flags);
3161 						is_refd =
3162 						   PV_BEEN_REFD(pv->pv_flags);
3163 					}
3164 					pool_put(&pmap_pv_pool, pv);
3165 				}
3166 			}
3167 			mappings++;
3168 
3169 			if (!l2pte_valid(pte)) {
3170 				/*
3171 				 * Ref/Mod emulation is still active for this
3172 				 * mapping, therefore it is has not yet been
3173 				 * accessed. No need to frob the cache/tlb.
3174 				 */
3175 				*ptep = 0;
3176 				PTE_SYNC_CURRENT(pm, ptep);
3177 				continue;
3178 			}
3179 
3180 			if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
3181 				/* Add to the clean list. */
3182 				cleanlist[cleanlist_idx].ptep = ptep;
3183 				cleanlist[cleanlist_idx].va =
3184 				    sva | (is_exec & 1);
3185 				cleanlist_idx++;
3186 			} else
3187 			if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
3188 				/* Nuke everything if needed. */
3189 #ifdef PMAP_CACHE_VIVT
3190 				pmap_idcache_wbinv_all(pm);
3191 #endif
3192 				pmap_tlb_flushID(pm);
3193 
3194 				/*
3195 				 * Roll back the previous PTE list,
3196 				 * and zero out the current PTE.
3197 				 */
3198 				for (cnt = 0;
3199 				     cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
3200 					*cleanlist[cnt].ptep = 0;
3201 					PTE_SYNC(cleanlist[cnt].ptep);
3202 				}
3203 				*ptep = 0;
3204 				PTE_SYNC(ptep);
3205 				cleanlist_idx++;
3206 				pm->pm_remove_all = true;
3207 			} else {
3208 				*ptep = 0;
3209 				PTE_SYNC(ptep);
3210 				if (pm->pm_remove_all == false) {
3211 					if (is_exec)
3212 						pmap_tlb_flushID_SE(pm, sva);
3213 					else
3214 					if (is_refd)
3215 						pmap_tlb_flushD_SE(pm, sva);
3216 				}
3217 			}
3218 		}
3219 
3220 		/*
3221 		 * Deal with any left overs
3222 		 */
3223 		if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
3224 			total += cleanlist_idx;
3225 			for (cnt = 0; cnt < cleanlist_idx; cnt++) {
3226 				if (pm->pm_cstate.cs_all != 0) {
3227 					vaddr_t clva = cleanlist[cnt].va & ~1;
3228 					if (cleanlist[cnt].va & 1) {
3229 #ifdef PMAP_CACHE_VIVT
3230 						pmap_idcache_wbinv_range(pm,
3231 						    clva, PAGE_SIZE);
3232 #endif
3233 						pmap_tlb_flushID_SE(pm, clva);
3234 					} else {
3235 #ifdef PMAP_CACHE_VIVT
3236 						pmap_dcache_wb_range(pm,
3237 						    clva, PAGE_SIZE, true,
3238 						    false);
3239 #endif
3240 						pmap_tlb_flushD_SE(pm, clva);
3241 					}
3242 				}
3243 				*cleanlist[cnt].ptep = 0;
3244 				PTE_SYNC_CURRENT(pm, cleanlist[cnt].ptep);
3245 			}
3246 
3247 			/*
3248 			 * If it looks like we're removing a whole bunch
3249 			 * of mappings, it's faster to just write-back
3250 			 * the whole cache now and defer TLB flushes until
3251 			 * pmap_update() is called.
3252 			 */
3253 			if (total <= PMAP_REMOVE_CLEAN_LIST_SIZE)
3254 				cleanlist_idx = 0;
3255 			else {
3256 				cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
3257 #ifdef PMAP_CACHE_VIVT
3258 				pmap_idcache_wbinv_all(pm);
3259 #endif
3260 				pm->pm_remove_all = true;
3261 			}
3262 		}
3263 
3264 		pmap_free_l2_bucket(pm, l2b, mappings);
3265 		pm->pm_stats.resident_count -= mappings;
3266 	}
3267 
3268 	pmap_release_pmap_lock(pm);
3269 	PMAP_MAP_TO_HEAD_UNLOCK();
3270 }
3271 
3272 #ifdef PMAP_CACHE_VIPT
3273 static struct pv_entry *
3274 pmap_kremove_pg(struct vm_page *pg, vaddr_t va)
3275 {
3276 	struct pv_entry *pv;
3277 
3278 	simple_lock(&pg->mdpage.pvh_slock);
3279 	KASSERT(arm_cache_prefer_mask == 0 || pg->mdpage.pvh_attrs & (PVF_COLORED|PVF_NC));
3280 	KASSERT((pg->mdpage.pvh_attrs & PVF_KMPAGE) == 0);
3281 
3282 	pv = pmap_remove_pv(pg, pmap_kernel(), va);
3283 	KASSERT(pv);
3284 	KASSERT(pv->pv_flags & PVF_KENTRY);
3285 
3286 	/*
3287 	 * If we are removing a writeable mapping to a cached exec page,
3288 	 * if it's the last mapping then clear it execness other sync
3289 	 * the page to the icache.
3290 	 */
3291 	if ((pg->mdpage.pvh_attrs & (PVF_NC|PVF_EXEC)) == PVF_EXEC
3292 	    && (pv->pv_flags & PVF_WRITE) != 0) {
3293 		if (SLIST_EMPTY(&pg->mdpage.pvh_list)) {
3294 			pg->mdpage.pvh_attrs &= ~PVF_EXEC;
3295 			PMAPCOUNT(exec_discarded_kremove);
3296 		} else {
3297 			pmap_syncicache_page(pg);
3298 			PMAPCOUNT(exec_synced_kremove);
3299 		}
3300 	}
3301 	pmap_vac_me_harder(pg, pmap_kernel(), 0);
3302 	simple_unlock(&pg->mdpage.pvh_slock);
3303 
3304 	return pv;
3305 }
3306 #endif /* PMAP_CACHE_VIPT */
3307 
3308 /*
3309  * pmap_kenter_pa: enter an unmanaged, wired kernel mapping
3310  *
3311  * We assume there is already sufficient KVM space available
3312  * to do this, as we can't allocate L2 descriptor tables/metadata
3313  * from here.
3314  */
3315 void
3316 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags)
3317 {
3318 	struct l2_bucket *l2b;
3319 	pt_entry_t *ptep, opte;
3320 #ifdef PMAP_CACHE_VIVT
3321 	struct vm_page *pg = (prot & PMAP_KMPAGE) ? PHYS_TO_VM_PAGE(pa) : NULL;
3322 #endif
3323 #ifdef PMAP_CACHE_VIPT
3324 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
3325 	struct vm_page *opg;
3326 	struct pv_entry *pv = NULL;
3327 #endif
3328 
3329 	NPDEBUG(PDB_KENTER,
3330 	    printf("pmap_kenter_pa: va 0x%08lx, pa 0x%08lx, prot 0x%x\n",
3331 	    va, pa, prot));
3332 
3333 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
3334 	KDASSERT(l2b != NULL);
3335 
3336 	ptep = &l2b->l2b_kva[l2pte_index(va)];
3337 	opte = *ptep;
3338 
3339 	if (opte == 0) {
3340 		PMAPCOUNT(kenter_mappings);
3341 		l2b->l2b_occupancy++;
3342 	} else {
3343 		PMAPCOUNT(kenter_remappings);
3344 #ifdef PMAP_CACHE_VIPT
3345 		opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3346 		if (opg) {
3347 			KASSERT(opg != pg);
3348 			KASSERT((opg->mdpage.pvh_attrs & PVF_KMPAGE) == 0);
3349 			KASSERT((prot & PMAP_KMPAGE) == 0);
3350 			simple_lock(&opg->mdpage.pvh_slock);
3351 			pv = pmap_kremove_pg(opg, va);
3352 			simple_unlock(&opg->mdpage.pvh_slock);
3353 		}
3354 #endif
3355 		if (l2pte_valid(opte)) {
3356 #ifdef PMAP_CACHE_VIVT
3357 			cpu_dcache_wbinv_range(va, PAGE_SIZE);
3358 #endif
3359 			cpu_tlb_flushD_SE(va);
3360 			cpu_cpwait();
3361 		}
3362 	}
3363 
3364 	*ptep = L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) |
3365 	    pte_l2_s_cache_mode;
3366 	PTE_SYNC(ptep);
3367 
3368 	if (pg) {
3369 		if (prot & PMAP_KMPAGE) {
3370 			simple_lock(&pg->mdpage.pvh_slock);
3371 			KASSERT(pg->mdpage.urw_mappings == 0);
3372 			KASSERT(pg->mdpage.uro_mappings == 0);
3373 			KASSERT(pg->mdpage.krw_mappings == 0);
3374 			KASSERT(pg->mdpage.kro_mappings == 0);
3375 #ifdef PMAP_CACHE_VIPT
3376 			KASSERT(pv == NULL);
3377 			KASSERT(arm_cache_prefer_mask == 0 || (va & PVF_COLORED) == 0);
3378 			KASSERT((pg->mdpage.pvh_attrs & PVF_NC) == 0);
3379 			/* if there is a color conflict, evict from cache. */
3380 			if (pmap_is_page_colored_p(pg)
3381 			    && ((va ^ pg->mdpage.pvh_attrs) & arm_cache_prefer_mask)) {
3382 				PMAPCOUNT(vac_color_change);
3383 				pmap_flush_page(pg, PMAP_FLUSH_PRIMARY);
3384 			} else if (pg->mdpage.pvh_attrs & PVF_MULTCLR) {
3385 				/*
3386 				 * If this page has multiple colors, expunge
3387 				 * them.
3388 				 */
3389 				PMAPCOUNT(vac_flush_lots2);
3390 				pmap_flush_page(pg, PMAP_FLUSH_SECONDARY);
3391 			}
3392 			pg->mdpage.pvh_attrs &= PAGE_SIZE - 1;
3393 			pg->mdpage.pvh_attrs |= PVF_KMPAGE
3394 			    | PVF_COLORED | PVF_DIRTY
3395 			    | (va & arm_cache_prefer_mask);
3396 #endif
3397 #ifdef PMAP_CACHE_VIVT
3398 			pg->mdpage.pvh_attrs |= PVF_KMPAGE;
3399 #endif
3400 			pmap_kmpages++;
3401 			simple_unlock(&pg->mdpage.pvh_slock);
3402 #ifdef PMAP_CACHE_VIPT
3403 		} else {
3404 			if (pv == NULL) {
3405 				pv = pool_get(&pmap_pv_pool, PR_NOWAIT);
3406 				KASSERT(pv != NULL);
3407 			}
3408 			pmap_enter_pv(pg, pv, pmap_kernel(), va,
3409 			    PVF_WIRED | PVF_KENTRY
3410 			    | (prot & VM_PROT_WRITE ? PVF_WRITE : 0));
3411 			if ((prot & VM_PROT_WRITE)
3412 			    && !(pg->mdpage.pvh_attrs & PVF_NC))
3413 				pg->mdpage.pvh_attrs |= PVF_DIRTY;
3414 			KASSERT((prot & VM_PROT_WRITE) == 0 || (pg->mdpage.pvh_attrs & (PVF_DIRTY|PVF_NC)));
3415 			simple_lock(&pg->mdpage.pvh_slock);
3416 			pmap_vac_me_harder(pg, pmap_kernel(), va);
3417 			simple_unlock(&pg->mdpage.pvh_slock);
3418 #endif
3419 		}
3420 #ifdef PMAP_CACHE_VIPT
3421 	} else {
3422 		if (pv != NULL)
3423 			pool_put(&pmap_pv_pool, pv);
3424 #endif
3425 	}
3426 }
3427 
3428 void
3429 pmap_kremove(vaddr_t va, vsize_t len)
3430 {
3431 	struct l2_bucket *l2b;
3432 	pt_entry_t *ptep, *sptep, opte;
3433 	vaddr_t next_bucket, eva;
3434 	u_int mappings;
3435 	struct vm_page *opg;
3436 
3437 	PMAPCOUNT(kenter_unmappings);
3438 
3439 	NPDEBUG(PDB_KREMOVE, printf("pmap_kremove: va 0x%08lx, len 0x%08lx\n",
3440 	    va, len));
3441 
3442 	eva = va + len;
3443 
3444 	while (va < eva) {
3445 		next_bucket = L2_NEXT_BUCKET(va);
3446 		if (next_bucket > eva)
3447 			next_bucket = eva;
3448 
3449 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
3450 		KDASSERT(l2b != NULL);
3451 
3452 		sptep = ptep = &l2b->l2b_kva[l2pte_index(va)];
3453 		mappings = 0;
3454 
3455 		while (va < next_bucket) {
3456 			opte = *ptep;
3457 			opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3458 			if (opg) {
3459 				if (opg->mdpage.pvh_attrs & PVF_KMPAGE) {
3460 					simple_lock(&opg->mdpage.pvh_slock);
3461 					KASSERT(opg->mdpage.urw_mappings == 0);
3462 					KASSERT(opg->mdpage.uro_mappings == 0);
3463 					KASSERT(opg->mdpage.krw_mappings == 0);
3464 					KASSERT(opg->mdpage.kro_mappings == 0);
3465 					opg->mdpage.pvh_attrs &= ~PVF_KMPAGE;
3466 #ifdef PMAP_CACHE_VIPT
3467 					opg->mdpage.pvh_attrs &= ~PVF_WRITE;
3468 #endif
3469 					pmap_kmpages--;
3470 					simple_unlock(&opg->mdpage.pvh_slock);
3471 #ifdef PMAP_CACHE_VIPT
3472 				} else {
3473 					pool_put(&pmap_pv_pool,
3474 					    pmap_kremove_pg(opg, va));
3475 #endif
3476 				}
3477 			}
3478 			if (l2pte_valid(opte)) {
3479 #ifdef PMAP_CACHE_VIVT
3480 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
3481 #endif
3482 				cpu_tlb_flushD_SE(va);
3483 			}
3484 			if (opte) {
3485 				*ptep = 0;
3486 				mappings++;
3487 			}
3488 			va += PAGE_SIZE;
3489 			ptep++;
3490 		}
3491 		KDASSERT(mappings <= l2b->l2b_occupancy);
3492 		l2b->l2b_occupancy -= mappings;
3493 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
3494 	}
3495 	cpu_cpwait();
3496 }
3497 
3498 bool
3499 pmap_extract(pmap_t pm, vaddr_t va, paddr_t *pap)
3500 {
3501 	struct l2_dtable *l2;
3502 	pd_entry_t *pl1pd, l1pd;
3503 	pt_entry_t *ptep, pte;
3504 	paddr_t pa;
3505 	u_int l1idx;
3506 
3507 	pmap_acquire_pmap_lock(pm);
3508 
3509 	l1idx = L1_IDX(va);
3510 	pl1pd = &pm->pm_l1->l1_kva[l1idx];
3511 	l1pd = *pl1pd;
3512 
3513 	if (l1pte_section_p(l1pd)) {
3514 		/*
3515 		 * These should only happen for pmap_kernel()
3516 		 */
3517 		KDASSERT(pm == pmap_kernel());
3518 		pmap_release_pmap_lock(pm);
3519 		pa = (l1pd & L1_S_FRAME) | (va & L1_S_OFFSET);
3520 	} else {
3521 		/*
3522 		 * Note that we can't rely on the validity of the L1
3523 		 * descriptor as an indication that a mapping exists.
3524 		 * We have to look it up in the L2 dtable.
3525 		 */
3526 		l2 = pm->pm_l2[L2_IDX(l1idx)];
3527 
3528 		if (l2 == NULL ||
3529 		    (ptep = l2->l2_bucket[L2_BUCKET(l1idx)].l2b_kva) == NULL) {
3530 			pmap_release_pmap_lock(pm);
3531 			return false;
3532 		}
3533 
3534 		ptep = &ptep[l2pte_index(va)];
3535 		pte = *ptep;
3536 		pmap_release_pmap_lock(pm);
3537 
3538 		if (pte == 0)
3539 			return false;
3540 
3541 		switch (pte & L2_TYPE_MASK) {
3542 		case L2_TYPE_L:
3543 			pa = (pte & L2_L_FRAME) | (va & L2_L_OFFSET);
3544 			break;
3545 
3546 		default:
3547 			pa = (pte & L2_S_FRAME) | (va & L2_S_OFFSET);
3548 			break;
3549 		}
3550 	}
3551 
3552 	if (pap != NULL)
3553 		*pap = pa;
3554 
3555 	return true;
3556 }
3557 
3558 void
3559 pmap_protect(pmap_t pm, vaddr_t sva, vaddr_t eva, vm_prot_t prot)
3560 {
3561 	struct l2_bucket *l2b;
3562 	pt_entry_t *ptep, pte;
3563 	vaddr_t next_bucket;
3564 	u_int flags;
3565 	u_int clr_mask;
3566 	int flush;
3567 
3568 	NPDEBUG(PDB_PROTECT,
3569 	    printf("pmap_protect: pm %p sva 0x%lx eva 0x%lx prot 0x%x\n",
3570 	    pm, sva, eva, prot));
3571 
3572 	if ((prot & VM_PROT_READ) == 0) {
3573 		pmap_remove(pm, sva, eva);
3574 		return;
3575 	}
3576 
3577 	if (prot & VM_PROT_WRITE) {
3578 		/*
3579 		 * If this is a read->write transition, just ignore it and let
3580 		 * uvm_fault() take care of it later.
3581 		 */
3582 		return;
3583 	}
3584 
3585 	PMAP_MAP_TO_HEAD_LOCK();
3586 	pmap_acquire_pmap_lock(pm);
3587 
3588 	flush = ((eva - sva) >= (PAGE_SIZE * 4)) ? 0 : -1;
3589 	flags = 0;
3590 	clr_mask = PVF_WRITE | ((prot & VM_PROT_EXECUTE) ? 0 : PVF_EXEC);
3591 
3592 	while (sva < eva) {
3593 		next_bucket = L2_NEXT_BUCKET(sva);
3594 		if (next_bucket > eva)
3595 			next_bucket = eva;
3596 
3597 		l2b = pmap_get_l2_bucket(pm, sva);
3598 		if (l2b == NULL) {
3599 			sva = next_bucket;
3600 			continue;
3601 		}
3602 
3603 		ptep = &l2b->l2b_kva[l2pte_index(sva)];
3604 
3605 		while (sva < next_bucket) {
3606 			pte = *ptep;
3607 			if (l2pte_valid(pte) != 0 && (pte & L2_S_PROT_W) != 0) {
3608 				struct vm_page *pg;
3609 				u_int f;
3610 
3611 #ifdef PMAP_CACHE_VIVT
3612 				/*
3613 				 * OK, at this point, we know we're doing
3614 				 * write-protect operation.  If the pmap is
3615 				 * active, write-back the page.
3616 				 */
3617 				pmap_dcache_wb_range(pm, sva, PAGE_SIZE,
3618 				    false, false);
3619 #endif
3620 
3621 				pg = PHYS_TO_VM_PAGE(l2pte_pa(pte));
3622 				pte &= ~L2_S_PROT_W;
3623 				*ptep = pte;
3624 				PTE_SYNC(ptep);
3625 
3626 				if (pg != NULL) {
3627 					simple_lock(&pg->mdpage.pvh_slock);
3628 					f = pmap_modify_pv(pg, pm, sva,
3629 					    clr_mask, 0);
3630 					pmap_vac_me_harder(pg, pm, sva);
3631 					simple_unlock(&pg->mdpage.pvh_slock);
3632 				} else
3633 					f = PVF_REF | PVF_EXEC;
3634 
3635 				if (flush >= 0) {
3636 					flush++;
3637 					flags |= f;
3638 				} else
3639 				if (PV_BEEN_EXECD(f))
3640 					pmap_tlb_flushID_SE(pm, sva);
3641 				else
3642 				if (PV_BEEN_REFD(f))
3643 					pmap_tlb_flushD_SE(pm, sva);
3644 			}
3645 
3646 			sva += PAGE_SIZE;
3647 			ptep++;
3648 		}
3649 	}
3650 
3651 	pmap_release_pmap_lock(pm);
3652 	PMAP_MAP_TO_HEAD_UNLOCK();
3653 
3654 	if (flush) {
3655 		if (PV_BEEN_EXECD(flags))
3656 			pmap_tlb_flushID(pm);
3657 		else
3658 		if (PV_BEEN_REFD(flags))
3659 			pmap_tlb_flushD(pm);
3660 	}
3661 }
3662 
3663 void
3664 pmap_icache_sync_range(pmap_t pm, vaddr_t sva, vaddr_t eva)
3665 {
3666 	struct l2_bucket *l2b;
3667 	pt_entry_t *ptep;
3668 	vaddr_t next_bucket;
3669 	vsize_t page_size = trunc_page(sva) + PAGE_SIZE - sva;
3670 
3671 	NPDEBUG(PDB_EXEC,
3672 	    printf("pmap_icache_sync_range: pm %p sva 0x%lx eva 0x%lx\n",
3673 	    pm, sva, eva));
3674 
3675 	PMAP_MAP_TO_HEAD_LOCK();
3676 	pmap_acquire_pmap_lock(pm);
3677 
3678 	while (sva < eva) {
3679 		next_bucket = L2_NEXT_BUCKET(sva);
3680 		if (next_bucket > eva)
3681 			next_bucket = eva;
3682 
3683 		l2b = pmap_get_l2_bucket(pm, sva);
3684 		if (l2b == NULL) {
3685 			sva = next_bucket;
3686 			continue;
3687 		}
3688 
3689 		for (ptep = &l2b->l2b_kva[l2pte_index(sva)];
3690 		     sva < next_bucket;
3691 		     sva += page_size, ptep++, page_size = PAGE_SIZE) {
3692 			if (l2pte_valid(*ptep)) {
3693 				cpu_icache_sync_range(sva,
3694 				    min(page_size, eva - sva));
3695 			}
3696 		}
3697 	}
3698 
3699 	pmap_release_pmap_lock(pm);
3700 	PMAP_MAP_TO_HEAD_UNLOCK();
3701 }
3702 
3703 void
3704 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
3705 {
3706 
3707 	NPDEBUG(PDB_PROTECT,
3708 	    printf("pmap_page_protect: pg %p (0x%08lx), prot 0x%x\n",
3709 	    pg, VM_PAGE_TO_PHYS(pg), prot));
3710 
3711 	switch(prot) {
3712 	case VM_PROT_READ|VM_PROT_WRITE:
3713 #if defined(PMAP_CHECK_VIPT) && defined(PMAP_APX)
3714 		pmap_clearbit(pg, PVF_EXEC);
3715 		break;
3716 #endif
3717 	case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
3718 		break;
3719 
3720 	case VM_PROT_READ:
3721 #if defined(PMAP_CHECK_VIPT) && defined(PMAP_APX)
3722 		pmap_clearbit(pg, PVF_WRITE|PVF_EXEC);
3723 		break;
3724 #endif
3725 	case VM_PROT_READ|VM_PROT_EXECUTE:
3726 		pmap_clearbit(pg, PVF_WRITE);
3727 		break;
3728 
3729 	default:
3730 		pmap_page_remove(pg);
3731 		break;
3732 	}
3733 }
3734 
3735 /*
3736  * pmap_clear_modify:
3737  *
3738  *	Clear the "modified" attribute for a page.
3739  */
3740 bool
3741 pmap_clear_modify(struct vm_page *pg)
3742 {
3743 	bool rv;
3744 
3745 	if (pg->mdpage.pvh_attrs & PVF_MOD) {
3746 		rv = true;
3747 #ifdef PMAP_CACHE_VIPT
3748 		/*
3749 		 * If we are going to clear the modified bit and there are
3750 		 * no other modified bits set, flush the page to memory and
3751 		 * mark it clean.
3752 		 */
3753 		if ((pg->mdpage.pvh_attrs & (PVF_DMOD|PVF_NC)) == PVF_MOD)
3754 			pmap_flush_page(pg, PMAP_CLEAN_PRIMARY);
3755 #endif
3756 		pmap_clearbit(pg, PVF_MOD);
3757 	} else
3758 		rv = false;
3759 
3760 	return (rv);
3761 }
3762 
3763 /*
3764  * pmap_clear_reference:
3765  *
3766  *	Clear the "referenced" attribute for a page.
3767  */
3768 bool
3769 pmap_clear_reference(struct vm_page *pg)
3770 {
3771 	bool rv;
3772 
3773 	if (pg->mdpage.pvh_attrs & PVF_REF) {
3774 		rv = true;
3775 		pmap_clearbit(pg, PVF_REF);
3776 	} else
3777 		rv = false;
3778 
3779 	return (rv);
3780 }
3781 
3782 /*
3783  * pmap_is_modified:
3784  *
3785  *	Test if a page has the "modified" attribute.
3786  */
3787 /* See <arm/arm32/pmap.h> */
3788 
3789 /*
3790  * pmap_is_referenced:
3791  *
3792  *	Test if a page has the "referenced" attribute.
3793  */
3794 /* See <arm/arm32/pmap.h> */
3795 
3796 int
3797 pmap_fault_fixup(pmap_t pm, vaddr_t va, vm_prot_t ftype, int user)
3798 {
3799 	struct l2_dtable *l2;
3800 	struct l2_bucket *l2b;
3801 	pd_entry_t *pl1pd, l1pd;
3802 	pt_entry_t *ptep, pte;
3803 	paddr_t pa;
3804 	u_int l1idx;
3805 	int rv = 0;
3806 
3807 	PMAP_MAP_TO_HEAD_LOCK();
3808 	pmap_acquire_pmap_lock(pm);
3809 
3810 	l1idx = L1_IDX(va);
3811 
3812 	/*
3813 	 * If there is no l2_dtable for this address, then the process
3814 	 * has no business accessing it.
3815 	 *
3816 	 * Note: This will catch userland processes trying to access
3817 	 * kernel addresses.
3818 	 */
3819 	l2 = pm->pm_l2[L2_IDX(l1idx)];
3820 	if (l2 == NULL)
3821 		goto out;
3822 
3823 	/*
3824 	 * Likewise if there is no L2 descriptor table
3825 	 */
3826 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
3827 	if (l2b->l2b_kva == NULL)
3828 		goto out;
3829 
3830 	/*
3831 	 * Check the PTE itself.
3832 	 */
3833 	ptep = &l2b->l2b_kva[l2pte_index(va)];
3834 	pte = *ptep;
3835 	if (pte == 0)
3836 		goto out;
3837 
3838 	/*
3839 	 * Catch a userland access to the vector page mapped at 0x0
3840 	 */
3841 	if (user && (pte & L2_S_PROT_U) == 0)
3842 		goto out;
3843 
3844 	pa = l2pte_pa(pte);
3845 
3846 	if ((ftype & VM_PROT_WRITE) && (pte & L2_S_PROT_W) == 0) {
3847 		/*
3848 		 * This looks like a good candidate for "page modified"
3849 		 * emulation...
3850 		 */
3851 		struct pv_entry *pv;
3852 		struct vm_page *pg;
3853 
3854 		/* Extract the physical address of the page */
3855 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3856 			goto out;
3857 
3858 		/* Get the current flags for this page. */
3859 		simple_lock(&pg->mdpage.pvh_slock);
3860 
3861 		pv = pmap_find_pv(pg, pm, va);
3862 		if (pv == NULL) {
3863 	    		simple_unlock(&pg->mdpage.pvh_slock);
3864 			goto out;
3865 		}
3866 
3867 		/*
3868 		 * Do the flags say this page is writable? If not then it
3869 		 * is a genuine write fault. If yes then the write fault is
3870 		 * our fault as we did not reflect the write access in the
3871 		 * PTE. Now we know a write has occurred we can correct this
3872 		 * and also set the modified bit
3873 		 */
3874 		if ((pv->pv_flags & PVF_WRITE) == 0) {
3875 		    	simple_unlock(&pg->mdpage.pvh_slock);
3876 			goto out;
3877 		}
3878 
3879 		NPDEBUG(PDB_FOLLOW,
3880 		    printf("pmap_fault_fixup: mod emul. pm %p, va 0x%08lx, pa 0x%08lx\n",
3881 		    pm, va, VM_PAGE_TO_PHYS(pg)));
3882 
3883 		pg->mdpage.pvh_attrs |= PVF_REF | PVF_MOD;
3884 		pv->pv_flags |= PVF_REF | PVF_MOD;
3885 #ifdef PMAP_CACHE_VIPT
3886 		/*
3887 		 * If there are cacheable mappings for this page, mark it dirty.
3888 		 */
3889 		if ((pg->mdpage.pvh_attrs & PVF_NC) == 0)
3890 			pg->mdpage.pvh_attrs |= PVF_DIRTY;
3891 #endif
3892 		simple_unlock(&pg->mdpage.pvh_slock);
3893 
3894 		/*
3895 		 * Re-enable write permissions for the page.  No need to call
3896 		 * pmap_vac_me_harder(), since this is just a
3897 		 * modified-emulation fault, and the PVF_WRITE bit isn't
3898 		 * changing. We've already set the cacheable bits based on
3899 		 * the assumption that we can write to this page.
3900 		 */
3901 		*ptep = (pte & ~L2_TYPE_MASK) | L2_S_PROTO | L2_S_PROT_W;
3902 		PTE_SYNC(ptep);
3903 		rv = 1;
3904 	} else
3905 	if ((pte & L2_TYPE_MASK) == L2_TYPE_INV) {
3906 		/*
3907 		 * This looks like a good candidate for "page referenced"
3908 		 * emulation.
3909 		 */
3910 		struct pv_entry *pv;
3911 		struct vm_page *pg;
3912 
3913 		/* Extract the physical address of the page */
3914 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3915 			goto out;
3916 
3917 		/* Get the current flags for this page. */
3918 		simple_lock(&pg->mdpage.pvh_slock);
3919 
3920 		pv = pmap_find_pv(pg, pm, va);
3921 		if (pv == NULL) {
3922 	    		simple_unlock(&pg->mdpage.pvh_slock);
3923 			goto out;
3924 		}
3925 
3926 		pg->mdpage.pvh_attrs |= PVF_REF;
3927 		pv->pv_flags |= PVF_REF;
3928 		simple_unlock(&pg->mdpage.pvh_slock);
3929 
3930 		NPDEBUG(PDB_FOLLOW,
3931 		    printf("pmap_fault_fixup: ref emul. pm %p, va 0x%08lx, pa 0x%08lx\n",
3932 		    pm, va, VM_PAGE_TO_PHYS(pg)));
3933 
3934 		*ptep = (pte & ~L2_TYPE_MASK) | L2_S_PROTO;
3935 		PTE_SYNC(ptep);
3936 		rv = 1;
3937 	}
3938 
3939 	/*
3940 	 * We know there is a valid mapping here, so simply
3941 	 * fix up the L1 if necessary.
3942 	 */
3943 	pl1pd = &pm->pm_l1->l1_kva[l1idx];
3944 	l1pd = l2b->l2b_phys | L1_C_DOM(pm->pm_domain) | L1_C_PROTO;
3945 	if (*pl1pd != l1pd) {
3946 		*pl1pd = l1pd;
3947 		PTE_SYNC(pl1pd);
3948 		rv = 1;
3949 	}
3950 
3951 #ifdef CPU_SA110
3952 	/*
3953 	 * There are bugs in the rev K SA110.  This is a check for one
3954 	 * of them.
3955 	 */
3956 	if (rv == 0 && curcpu()->ci_arm_cputype == CPU_ID_SA110 &&
3957 	    curcpu()->ci_arm_cpurev < 3) {
3958 		/* Always current pmap */
3959 		if (l2pte_valid(pte)) {
3960 			extern int kernel_debug;
3961 			if (kernel_debug & 1) {
3962 				struct proc *p = curlwp->l_proc;
3963 				printf("prefetch_abort: page is already "
3964 				    "mapped - pte=%p *pte=%08x\n", ptep, pte);
3965 				printf("prefetch_abort: pc=%08lx proc=%p "
3966 				    "process=%s\n", va, p, p->p_comm);
3967 				printf("prefetch_abort: far=%08x fs=%x\n",
3968 				    cpu_faultaddress(), cpu_faultstatus());
3969 			}
3970 #ifdef DDB
3971 			if (kernel_debug & 2)
3972 				Debugger();
3973 #endif
3974 			rv = 1;
3975 		}
3976 	}
3977 #endif /* CPU_SA110 */
3978 
3979 #ifdef DEBUG
3980 	/*
3981 	 * If 'rv == 0' at this point, it generally indicates that there is a
3982 	 * stale TLB entry for the faulting address. This happens when two or
3983 	 * more processes are sharing an L1. Since we don't flush the TLB on
3984 	 * a context switch between such processes, we can take domain faults
3985 	 * for mappings which exist at the same VA in both processes. EVEN IF
3986 	 * WE'VE RECENTLY FIXED UP THE CORRESPONDING L1 in pmap_enter(), for
3987 	 * example.
3988 	 *
3989 	 * This is extremely likely to happen if pmap_enter() updated the L1
3990 	 * entry for a recently entered mapping. In this case, the TLB is
3991 	 * flushed for the new mapping, but there may still be TLB entries for
3992 	 * other mappings belonging to other processes in the 1MB range
3993 	 * covered by the L1 entry.
3994 	 *
3995 	 * Since 'rv == 0', we know that the L1 already contains the correct
3996 	 * value, so the fault must be due to a stale TLB entry.
3997 	 *
3998 	 * Since we always need to flush the TLB anyway in the case where we
3999 	 * fixed up the L1, or frobbed the L2 PTE, we effectively deal with
4000 	 * stale TLB entries dynamically.
4001 	 *
4002 	 * However, the above condition can ONLY happen if the current L1 is
4003 	 * being shared. If it happens when the L1 is unshared, it indicates
4004 	 * that other parts of the pmap are not doing their job WRT managing
4005 	 * the TLB.
4006 	 */
4007 	if (rv == 0 && pm->pm_l1->l1_domain_use_count == 1) {
4008 		extern int last_fault_code;
4009 		printf("fixup: pm %p, va 0x%lx, ftype %d - nothing to do!\n",
4010 		    pm, va, ftype);
4011 		printf("fixup: l2 %p, l2b %p, ptep %p, pl1pd %p\n",
4012 		    l2, l2b, ptep, pl1pd);
4013 		printf("fixup: pte 0x%x, l1pd 0x%x, last code 0x%x\n",
4014 		    pte, l1pd, last_fault_code);
4015 #ifdef DDB
4016 		Debugger();
4017 #endif
4018 	}
4019 #endif
4020 
4021 	cpu_tlb_flushID_SE(va);
4022 	cpu_cpwait();
4023 
4024 	rv = 1;
4025 
4026 out:
4027 	pmap_release_pmap_lock(pm);
4028 	PMAP_MAP_TO_HEAD_UNLOCK();
4029 
4030 	return (rv);
4031 }
4032 
4033 /*
4034  * Routine:	pmap_procwr
4035  *
4036  * Function:
4037  *	Synchronize caches corresponding to [addr, addr+len) in p.
4038  *
4039  */
4040 void
4041 pmap_procwr(struct proc *p, vaddr_t va, int len)
4042 {
4043 	/* We only need to do anything if it is the current process. */
4044 	if (p == curproc)
4045 		cpu_icache_sync_range(va, len);
4046 }
4047 
4048 /*
4049  * Routine:	pmap_unwire
4050  * Function:	Clear the wired attribute for a map/virtual-address pair.
4051  *
4052  * In/out conditions:
4053  *		The mapping must already exist in the pmap.
4054  */
4055 void
4056 pmap_unwire(pmap_t pm, vaddr_t va)
4057 {
4058 	struct l2_bucket *l2b;
4059 	pt_entry_t *ptep, pte;
4060 	struct vm_page *pg;
4061 	paddr_t pa;
4062 
4063 	NPDEBUG(PDB_WIRING, printf("pmap_unwire: pm %p, va 0x%08lx\n", pm, va));
4064 
4065 	PMAP_MAP_TO_HEAD_LOCK();
4066 	pmap_acquire_pmap_lock(pm);
4067 
4068 	l2b = pmap_get_l2_bucket(pm, va);
4069 	KDASSERT(l2b != NULL);
4070 
4071 	ptep = &l2b->l2b_kva[l2pte_index(va)];
4072 	pte = *ptep;
4073 
4074 	/* Extract the physical address of the page */
4075 	pa = l2pte_pa(pte);
4076 
4077 	if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
4078 		/* Update the wired bit in the pv entry for this page. */
4079 		simple_lock(&pg->mdpage.pvh_slock);
4080 		(void) pmap_modify_pv(pg, pm, va, PVF_WIRED, 0);
4081 		simple_unlock(&pg->mdpage.pvh_slock);
4082 	}
4083 
4084 	pmap_release_pmap_lock(pm);
4085 	PMAP_MAP_TO_HEAD_UNLOCK();
4086 }
4087 
4088 void
4089 pmap_activate(struct lwp *l)
4090 {
4091 	extern int block_userspace_access;
4092 	pmap_t opm, npm, rpm;
4093 	uint32_t odacr, ndacr;
4094 	int oldirqstate;
4095 
4096 	/*
4097 	 * If activating a non-current lwp or the current lwp is
4098 	 * already active, just return.
4099 	 */
4100 	if (l != curlwp ||
4101 	    l->l_proc->p_vmspace->vm_map.pmap->pm_activated == true)
4102 		return;
4103 
4104 	npm = l->l_proc->p_vmspace->vm_map.pmap;
4105 	ndacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2)) |
4106 	    (DOMAIN_CLIENT << (npm->pm_domain * 2));
4107 
4108 	/*
4109 	 * If TTB and DACR are unchanged, short-circuit all the
4110 	 * TLB/cache management stuff.
4111 	 */
4112 	if (pmap_previous_active_lwp != NULL) {
4113 		opm = pmap_previous_active_lwp->l_proc->p_vmspace->vm_map.pmap;
4114 		odacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2)) |
4115 		    (DOMAIN_CLIENT << (opm->pm_domain * 2));
4116 
4117 		if (opm->pm_l1 == npm->pm_l1 && odacr == ndacr)
4118 			goto all_done;
4119 	} else
4120 		opm = NULL;
4121 
4122 	PMAPCOUNT(activations);
4123 	block_userspace_access = 1;
4124 
4125 	/*
4126 	 * If switching to a user vmspace which is different to the
4127 	 * most recent one, and the most recent one is potentially
4128 	 * live in the cache, we must write-back and invalidate the
4129 	 * entire cache.
4130 	 */
4131 	rpm = pmap_recent_user;
4132 
4133 /*
4134  * XXXSCW: There's a corner case here which can leave turds in the cache as
4135  * reported in kern/41058. They're probably left over during tear-down and
4136  * switching away from an exiting process. Until the root cause is identified
4137  * and fixed, zap the cache when switching pmaps. This will result in a few
4138  * unnecessary cache flushes, but that's better than silently corrupting data.
4139  */
4140 #if 0
4141 	if (npm != pmap_kernel() && rpm && npm != rpm &&
4142 	    rpm->pm_cstate.cs_cache) {
4143 		rpm->pm_cstate.cs_cache = 0;
4144 #ifdef PMAP_CACHE_VIVT
4145 		cpu_idcache_wbinv_all();
4146 #endif
4147 	}
4148 #else
4149 	if (rpm) {
4150 		rpm->pm_cstate.cs_cache = 0;
4151 		if (npm == pmap_kernel())
4152 			pmap_recent_user = NULL;
4153 #ifdef PMAP_CACHE_VIVT
4154 		cpu_idcache_wbinv_all();
4155 #endif
4156 	}
4157 #endif
4158 
4159 	/* No interrupts while we frob the TTB/DACR */
4160 	oldirqstate = disable_interrupts(IF32_bits);
4161 
4162 	/*
4163 	 * For ARM_VECTORS_LOW, we MUST, I repeat, MUST fix up the L1
4164 	 * entry corresponding to 'vector_page' in the incoming L1 table
4165 	 * before switching to it otherwise subsequent interrupts/exceptions
4166 	 * (including domain faults!) will jump into hyperspace.
4167 	 */
4168 	if (npm->pm_pl1vec != NULL) {
4169 		cpu_tlb_flushID_SE((u_int)vector_page);
4170 		cpu_cpwait();
4171 		*npm->pm_pl1vec = npm->pm_l1vec;
4172 		PTE_SYNC(npm->pm_pl1vec);
4173 	}
4174 
4175 	cpu_domains(ndacr);
4176 
4177 	if (npm == pmap_kernel() || npm == rpm) {
4178 		/*
4179 		 * Switching to a kernel thread, or back to the
4180 		 * same user vmspace as before... Simply update
4181 		 * the TTB (no TLB flush required)
4182 		 */
4183 		__asm volatile("mcr p15, 0, %0, c2, c0, 0" ::
4184 		    "r"(npm->pm_l1->l1_physaddr));
4185 		cpu_cpwait();
4186 	} else {
4187 		/*
4188 		 * Otherwise, update TTB and flush TLB
4189 		 */
4190 		cpu_context_switch(npm->pm_l1->l1_physaddr);
4191 		if (rpm != NULL)
4192 			rpm->pm_cstate.cs_tlb = 0;
4193 	}
4194 
4195 	restore_interrupts(oldirqstate);
4196 
4197 	block_userspace_access = 0;
4198 
4199  all_done:
4200 	/*
4201 	 * The new pmap is resident. Make sure it's marked
4202 	 * as resident in the cache/TLB.
4203 	 */
4204 	npm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
4205 	if (npm != pmap_kernel())
4206 		pmap_recent_user = npm;
4207 
4208 	/* The old pmap is not longer active */
4209 	if (opm != NULL)
4210 		opm->pm_activated = false;
4211 
4212 	/* But the new one is */
4213 	npm->pm_activated = true;
4214 }
4215 
4216 void
4217 pmap_deactivate(struct lwp *l)
4218 {
4219 
4220 	/*
4221 	 * If the process is exiting, make sure pmap_activate() does
4222 	 * a full MMU context-switch and cache flush, which we might
4223 	 * otherwise skip. See PR port-arm/38950.
4224 	 */
4225 	if (l->l_proc->p_sflag & PS_WEXIT)
4226 		pmap_previous_active_lwp = NULL;
4227 
4228 	l->l_proc->p_vmspace->vm_map.pmap->pm_activated = false;
4229 }
4230 
4231 void
4232 pmap_update(pmap_t pm)
4233 {
4234 
4235 	if (pm->pm_remove_all) {
4236 		/*
4237 		 * Finish up the pmap_remove_all() optimisation by flushing
4238 		 * the TLB.
4239 		 */
4240 		pmap_tlb_flushID(pm);
4241 		pm->pm_remove_all = false;
4242 	}
4243 
4244 	if (pmap_is_current(pm)) {
4245 		/*
4246 		 * If we're dealing with a current userland pmap, move its L1
4247 		 * to the end of the LRU.
4248 		 */
4249 		if (pm != pmap_kernel())
4250 			pmap_use_l1(pm);
4251 
4252 		/*
4253 		 * We can assume we're done with frobbing the cache/tlb for
4254 		 * now. Make sure any future pmap ops don't skip cache/tlb
4255 		 * flushes.
4256 		 */
4257 		pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
4258 	}
4259 
4260 	PMAPCOUNT(updates);
4261 
4262 	/*
4263 	 * make sure TLB/cache operations have completed.
4264 	 */
4265 	cpu_cpwait();
4266 }
4267 
4268 void
4269 pmap_remove_all(pmap_t pm)
4270 {
4271 
4272 	/*
4273 	 * The vmspace described by this pmap is about to be torn down.
4274 	 * Until pmap_update() is called, UVM will only make calls
4275 	 * to pmap_remove(). We can make life much simpler by flushing
4276 	 * the cache now, and deferring TLB invalidation to pmap_update().
4277 	 */
4278 #ifdef PMAP_CACHE_VIVT
4279 	pmap_idcache_wbinv_all(pm);
4280 #endif
4281 	pm->pm_remove_all = true;
4282 }
4283 
4284 /*
4285  * Retire the given physical map from service.
4286  * Should only be called if the map contains no valid mappings.
4287  */
4288 void
4289 pmap_destroy(pmap_t pm)
4290 {
4291 	u_int count;
4292 
4293 	if (pm == NULL)
4294 		return;
4295 
4296 	if (pm->pm_remove_all) {
4297 		pmap_tlb_flushID(pm);
4298 		pm->pm_remove_all = false;
4299 	}
4300 
4301 	/*
4302 	 * Drop reference count
4303 	 */
4304 	mutex_enter(&pm->pm_lock);
4305 	count = --pm->pm_obj.uo_refs;
4306 	mutex_exit(&pm->pm_lock);
4307 	if (count > 0) {
4308 		if (pmap_is_current(pm)) {
4309 			if (pm != pmap_kernel())
4310 				pmap_use_l1(pm);
4311 			pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
4312 		}
4313 		return;
4314 	}
4315 
4316 	/*
4317 	 * reference count is zero, free pmap resources and then free pmap.
4318 	 */
4319 
4320 	if (vector_page < KERNEL_BASE) {
4321 		KDASSERT(!pmap_is_current(pm));
4322 
4323 		/* Remove the vector page mapping */
4324 		pmap_remove(pm, vector_page, vector_page + PAGE_SIZE);
4325 		pmap_update(pm);
4326 	}
4327 
4328 	LIST_REMOVE(pm, pm_list);
4329 
4330 	pmap_free_l1(pm);
4331 
4332 	if (pmap_recent_user == pm)
4333 		pmap_recent_user = NULL;
4334 
4335 	UVM_OBJ_DESTROY(&pm->pm_obj);
4336 
4337 	/* return the pmap to the pool */
4338 	pool_cache_put(&pmap_cache, pm);
4339 }
4340 
4341 
4342 /*
4343  * void pmap_reference(pmap_t pm)
4344  *
4345  * Add a reference to the specified pmap.
4346  */
4347 void
4348 pmap_reference(pmap_t pm)
4349 {
4350 
4351 	if (pm == NULL)
4352 		return;
4353 
4354 	pmap_use_l1(pm);
4355 
4356 	mutex_enter(&pm->pm_lock);
4357 	pm->pm_obj.uo_refs++;
4358 	mutex_exit(&pm->pm_lock);
4359 }
4360 
4361 #if ARM_MMU_V6 > 0
4362 
4363 static struct evcnt pmap_prefer_nochange_ev =
4364     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap prefer", "nochange");
4365 static struct evcnt pmap_prefer_change_ev =
4366     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap prefer", "change");
4367 
4368 EVCNT_ATTACH_STATIC(pmap_prefer_change_ev);
4369 EVCNT_ATTACH_STATIC(pmap_prefer_nochange_ev);
4370 
4371 void
4372 pmap_prefer(vaddr_t hint, vaddr_t *vap, int td)
4373 {
4374 	vsize_t mask = arm_cache_prefer_mask | (PAGE_SIZE - 1);
4375 	vaddr_t va = *vap;
4376 	vaddr_t diff = (hint - va) & mask;
4377 	if (diff == 0) {
4378 		pmap_prefer_nochange_ev.ev_count++;
4379 	} else {
4380 		pmap_prefer_change_ev.ev_count++;
4381 		if (__predict_false(td))
4382 			va -= mask + 1;
4383 		*vap = va + diff;
4384 	}
4385 }
4386 #endif /* ARM_MMU_V6 */
4387 
4388 /*
4389  * pmap_zero_page()
4390  *
4391  * Zero a given physical page by mapping it at a page hook point.
4392  * In doing the zero page op, the page we zero is mapped cachable, as with
4393  * StrongARM accesses to non-cached pages are non-burst making writing
4394  * _any_ bulk data very slow.
4395  */
4396 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0
4397 void
4398 pmap_zero_page_generic(paddr_t phys)
4399 {
4400 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
4401 	struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
4402 #endif
4403 #ifdef PMAP_CACHE_VIPT
4404 	/* Choose the last page color it had, if any */
4405 	const vsize_t va_offset = pg->mdpage.pvh_attrs & arm_cache_prefer_mask;
4406 #else
4407 	const vsize_t va_offset = 0;
4408 #endif
4409 	pt_entry_t * const ptep = &cdst_pte[va_offset >> PGSHIFT];
4410 
4411 #ifdef DEBUG
4412 	if (!SLIST_EMPTY(&pg->mdpage.pvh_list))
4413 		panic("pmap_zero_page: page has mappings");
4414 #endif
4415 
4416 	KDASSERT((phys & PGOFSET) == 0);
4417 
4418 	/*
4419 	 * Hook in the page, zero it, and purge the cache for that
4420 	 * zeroed page. Invalidate the TLB as needed.
4421 	 */
4422 	*ptep = L2_S_PROTO | phys |
4423 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
4424 	PTE_SYNC(ptep);
4425 	cpu_tlb_flushD_SE(cdstp + va_offset);
4426 	cpu_cpwait();
4427 	bzero_page(cdstp + va_offset);
4428 	/*
4429 	 * Unmap the page.
4430 	 */
4431 	*ptep = 0;
4432 	PTE_SYNC(ptep);
4433 	cpu_tlb_flushD_SE(cdstp + va_offset);
4434 #ifdef PMAP_CACHE_VIVT
4435 	cpu_dcache_wbinv_range(cdstp + va_offset, PAGE_SIZE);
4436 #endif
4437 #ifdef PMAP_CACHE_VIPT
4438 	/*
4439 	 * This page is now cache resident so it now has a page color.
4440 	 * Any contents have been obliterated so clear the EXEC flag.
4441 	 */
4442 	if (!pmap_is_page_colored_p(pg)) {
4443 		PMAPCOUNT(vac_color_new);
4444 		pg->mdpage.pvh_attrs |= PVF_COLORED;
4445 	}
4446 	if (PV_IS_EXEC_P(pg->mdpage.pvh_attrs)) {
4447 		pg->mdpage.pvh_attrs &= ~PVF_EXEC;
4448 		PMAPCOUNT(exec_discarded_zero);
4449 	}
4450 	pg->mdpage.pvh_attrs |= PVF_DIRTY;
4451 #endif
4452 }
4453 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
4454 
4455 #if ARM_MMU_XSCALE == 1
4456 void
4457 pmap_zero_page_xscale(paddr_t phys)
4458 {
4459 #ifdef DEBUG
4460 	struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
4461 
4462 	if (!SLIST_EMPTY(&pg->mdpage.pvh_list))
4463 		panic("pmap_zero_page: page has mappings");
4464 #endif
4465 
4466 	KDASSERT((phys & PGOFSET) == 0);
4467 
4468 	/*
4469 	 * Hook in the page, zero it, and purge the cache for that
4470 	 * zeroed page. Invalidate the TLB as needed.
4471 	 */
4472 	*cdst_pte = L2_S_PROTO | phys |
4473 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
4474 	    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
4475 	PTE_SYNC(cdst_pte);
4476 	cpu_tlb_flushD_SE(cdstp);
4477 	cpu_cpwait();
4478 	bzero_page(cdstp);
4479 	xscale_cache_clean_minidata();
4480 }
4481 #endif /* ARM_MMU_XSCALE == 1 */
4482 
4483 /* pmap_pageidlezero()
4484  *
4485  * The same as above, except that we assume that the page is not
4486  * mapped.  This means we never have to flush the cache first.  Called
4487  * from the idle loop.
4488  */
4489 bool
4490 pmap_pageidlezero(paddr_t phys)
4491 {
4492 	unsigned int i;
4493 	int *ptr;
4494 	bool rv = true;
4495 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
4496 	struct vm_page * const pg = PHYS_TO_VM_PAGE(phys);
4497 #endif
4498 #ifdef PMAP_CACHE_VIPT
4499 	/* Choose the last page color it had, if any */
4500 	const vsize_t va_offset = pg->mdpage.pvh_attrs & arm_cache_prefer_mask;
4501 #else
4502 	const vsize_t va_offset = 0;
4503 #endif
4504 	pt_entry_t * const ptep = &csrc_pte[va_offset >> PGSHIFT];
4505 
4506 
4507 #ifdef DEBUG
4508 	if (!SLIST_EMPTY(&pg->mdpage.pvh_list))
4509 		panic("pmap_pageidlezero: page has mappings");
4510 #endif
4511 
4512 	KDASSERT((phys & PGOFSET) == 0);
4513 
4514 	/*
4515 	 * Hook in the page, zero it, and purge the cache for that
4516 	 * zeroed page. Invalidate the TLB as needed.
4517 	 */
4518 	*ptep = L2_S_PROTO | phys |
4519 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
4520 	PTE_SYNC(ptep);
4521 	cpu_tlb_flushD_SE(cdstp + va_offset);
4522 	cpu_cpwait();
4523 
4524 	for (i = 0, ptr = (int *)(cdstp + va_offset);
4525 			i < (PAGE_SIZE / sizeof(int)); i++) {
4526 		if (sched_curcpu_runnable_p() != 0) {
4527 			/*
4528 			 * A process has become ready.  Abort now,
4529 			 * so we don't keep it waiting while we
4530 			 * do slow memory access to finish this
4531 			 * page.
4532 			 */
4533 			rv = false;
4534 			break;
4535 		}
4536 		*ptr++ = 0;
4537 	}
4538 
4539 #ifdef PMAP_CACHE_VIVT
4540 	if (rv)
4541 		/*
4542 		 * if we aborted we'll rezero this page again later so don't
4543 		 * purge it unless we finished it
4544 		 */
4545 		cpu_dcache_wbinv_range(cdstp, PAGE_SIZE);
4546 #elif defined(PMAP_CACHE_VIPT)
4547 	/*
4548 	 * This page is now cache resident so it now has a page color.
4549 	 * Any contents have been obliterated so clear the EXEC flag.
4550 	 */
4551 	if (!pmap_is_page_colored_p(pg)) {
4552 		PMAPCOUNT(vac_color_new);
4553 		pg->mdpage.pvh_attrs |= PVF_COLORED;
4554 	}
4555 	if (PV_IS_EXEC_P(pg->mdpage.pvh_attrs)) {
4556 		pg->mdpage.pvh_attrs &= ~PVF_EXEC;
4557 		PMAPCOUNT(exec_discarded_zero);
4558 	}
4559 #endif
4560 	/*
4561 	 * Unmap the page.
4562 	 */
4563 	*ptep = 0;
4564 	PTE_SYNC(ptep);
4565 	cpu_tlb_flushD_SE(cdstp + va_offset);
4566 
4567 	return (rv);
4568 }
4569 
4570 /*
4571  * pmap_copy_page()
4572  *
4573  * Copy one physical page into another, by mapping the pages into
4574  * hook points. The same comment regarding cachability as in
4575  * pmap_zero_page also applies here.
4576  */
4577 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0
4578 void
4579 pmap_copy_page_generic(paddr_t src, paddr_t dst)
4580 {
4581 	struct vm_page * const src_pg = PHYS_TO_VM_PAGE(src);
4582 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
4583 	struct vm_page * const dst_pg = PHYS_TO_VM_PAGE(dst);
4584 #endif
4585 #ifdef PMAP_CACHE_VIPT
4586 	const vsize_t src_va_offset = src_pg->mdpage.pvh_attrs & arm_cache_prefer_mask;
4587 	const vsize_t dst_va_offset = dst_pg->mdpage.pvh_attrs & arm_cache_prefer_mask;
4588 #else
4589 	const vsize_t src_va_offset = 0;
4590 	const vsize_t dst_va_offset = 0;
4591 #endif
4592 	pt_entry_t * const src_ptep = &csrc_pte[src_va_offset >> PGSHIFT];
4593 	pt_entry_t * const dst_ptep = &cdst_pte[dst_va_offset >> PGSHIFT];
4594 
4595 #ifdef DEBUG
4596 	if (!SLIST_EMPTY(&dst_pg->mdpage.pvh_list))
4597 		panic("pmap_copy_page: dst page has mappings");
4598 #endif
4599 
4600 #ifdef PMAP_CACHE_VIPT
4601 	KASSERT(arm_cache_prefer_mask == 0 || src_pg->mdpage.pvh_attrs & (PVF_COLORED|PVF_NC));
4602 #endif
4603 	KDASSERT((src & PGOFSET) == 0);
4604 	KDASSERT((dst & PGOFSET) == 0);
4605 
4606 	/*
4607 	 * Clean the source page.  Hold the source page's lock for
4608 	 * the duration of the copy so that no other mappings can
4609 	 * be created while we have a potentially aliased mapping.
4610 	 */
4611 	simple_lock(&src_pg->mdpage.pvh_slock);
4612 #ifdef PMAP_CACHE_VIVT
4613 	(void) pmap_clean_page(SLIST_FIRST(&src_pg->mdpage.pvh_list), true);
4614 #endif
4615 
4616 	/*
4617 	 * Map the pages into the page hook points, copy them, and purge
4618 	 * the cache for the appropriate page. Invalidate the TLB
4619 	 * as required.
4620 	 */
4621 	*src_ptep = L2_S_PROTO
4622 	    | src
4623 #ifdef PMAP_CACHE_VIPT
4624 	    | ((src_pg->mdpage.pvh_attrs & PVF_NC) ? 0 : pte_l2_s_cache_mode)
4625 #endif
4626 #ifdef PMAP_CACHE_VIVT
4627 	    | pte_l2_s_cache_mode
4628 #endif
4629 	    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ);
4630 	*dst_ptep = L2_S_PROTO | dst |
4631 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
4632 	PTE_SYNC(src_ptep);
4633 	PTE_SYNC(dst_ptep);
4634 	cpu_tlb_flushD_SE(csrcp + src_va_offset);
4635 	cpu_tlb_flushD_SE(cdstp + dst_va_offset);
4636 	cpu_cpwait();
4637 	bcopy_page(csrcp + src_va_offset, cdstp + dst_va_offset);
4638 #ifdef PMAP_CACHE_VIVT
4639 	cpu_dcache_inv_range(csrcp + src_va_offset, PAGE_SIZE);
4640 #endif
4641 	simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
4642 #ifdef PMAP_CACHE_VIVT
4643 	cpu_dcache_wbinv_range(cdstp + dst_va_offset, PAGE_SIZE);
4644 #endif
4645 	/*
4646 	 * Unmap the pages.
4647 	 */
4648 	*src_ptep = 0;
4649 	*dst_ptep = 0;
4650 	PTE_SYNC(src_ptep);
4651 	PTE_SYNC(dst_ptep);
4652 	cpu_tlb_flushD_SE(csrcp + src_va_offset);
4653 	cpu_tlb_flushD_SE(cdstp + dst_va_offset);
4654 #ifdef PMAP_CACHE_VIPT
4655 	/*
4656 	 * Now that the destination page is in the cache, mark it as colored.
4657 	 * If this was an exec page, discard it.
4658 	 */
4659 	if (!pmap_is_page_colored_p(dst_pg)) {
4660 		PMAPCOUNT(vac_color_new);
4661 		dst_pg->mdpage.pvh_attrs |= PVF_COLORED;
4662 	}
4663 	if (PV_IS_EXEC_P(dst_pg->mdpage.pvh_attrs)) {
4664 		dst_pg->mdpage.pvh_attrs &= ~PVF_EXEC;
4665 		PMAPCOUNT(exec_discarded_copy);
4666 	}
4667 	dst_pg->mdpage.pvh_attrs |= PVF_DIRTY;
4668 #endif
4669 }
4670 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
4671 
4672 #if ARM_MMU_XSCALE == 1
4673 void
4674 pmap_copy_page_xscale(paddr_t src, paddr_t dst)
4675 {
4676 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
4677 #ifdef DEBUG
4678 	struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst);
4679 
4680 	if (!SLIST_EMPTY(&dst_pg->mdpage.pvh_list))
4681 		panic("pmap_copy_page: dst page has mappings");
4682 #endif
4683 
4684 	KDASSERT((src & PGOFSET) == 0);
4685 	KDASSERT((dst & PGOFSET) == 0);
4686 
4687 	/*
4688 	 * Clean the source page.  Hold the source page's lock for
4689 	 * the duration of the copy so that no other mappings can
4690 	 * be created while we have a potentially aliased mapping.
4691 	 */
4692 	simple_lock(&src_pg->mdpage.pvh_slock);
4693 #ifdef PMAP_CACHE_VIVT
4694 	(void) pmap_clean_page(SLIST_FIRST(&src_pg->mdpage.pvh_list), true);
4695 #endif
4696 
4697 	/*
4698 	 * Map the pages into the page hook points, copy them, and purge
4699 	 * the cache for the appropriate page. Invalidate the TLB
4700 	 * as required.
4701 	 */
4702 	*csrc_pte = L2_S_PROTO | src |
4703 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
4704 	    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
4705 	PTE_SYNC(csrc_pte);
4706 	*cdst_pte = L2_S_PROTO | dst |
4707 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
4708 	    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
4709 	PTE_SYNC(cdst_pte);
4710 	cpu_tlb_flushD_SE(csrcp);
4711 	cpu_tlb_flushD_SE(cdstp);
4712 	cpu_cpwait();
4713 	bcopy_page(csrcp, cdstp);
4714 	simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
4715 	xscale_cache_clean_minidata();
4716 }
4717 #endif /* ARM_MMU_XSCALE == 1 */
4718 
4719 /*
4720  * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
4721  *
4722  * Return the start and end addresses of the kernel's virtual space.
4723  * These values are setup in pmap_bootstrap and are updated as pages
4724  * are allocated.
4725  */
4726 void
4727 pmap_virtual_space(vaddr_t *start, vaddr_t *end)
4728 {
4729 	*start = virtual_avail;
4730 	*end = virtual_end;
4731 }
4732 
4733 /*
4734  * Helper function for pmap_grow_l2_bucket()
4735  */
4736 static inline int
4737 pmap_grow_map(vaddr_t va, pt_entry_t cache_mode, paddr_t *pap)
4738 {
4739 	struct l2_bucket *l2b;
4740 	pt_entry_t *ptep;
4741 	paddr_t pa;
4742 
4743 	if (uvm.page_init_done == false) {
4744 #ifdef PMAP_STEAL_MEMORY
4745 		pv_addr_t pv;
4746 		pmap_boot_pagealloc(PAGE_SIZE,
4747 #ifdef PMAP_CACHE_VIPT
4748 		    arm_cache_prefer_mask,
4749 		    va & arm_cache_prefer_mask,
4750 #else
4751 		    0, 0,
4752 #endif
4753 		    &pv);
4754 		pa = pv.pv_pa;
4755 #else
4756 		if (uvm_page_physget(&pa) == false)
4757 			return (1);
4758 #endif	/* PMAP_STEAL_MEMORY */
4759 	} else {
4760 		struct vm_page *pg;
4761 		pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE);
4762 		if (pg == NULL)
4763 			return (1);
4764 		pa = VM_PAGE_TO_PHYS(pg);
4765 #ifdef PMAP_CACHE_VIPT
4766 		/*
4767 		 * This new page must not have any mappings.  Enter it via
4768 		 * pmap_kenter_pa and let that routine do the hard work.
4769 		 */
4770 		KASSERT(SLIST_EMPTY(&pg->mdpage.pvh_list));
4771 		pmap_kenter_pa(va, pa,
4772 		    VM_PROT_READ|VM_PROT_WRITE|PMAP_KMPAGE, 0);
4773 #endif
4774 	}
4775 
4776 	if (pap)
4777 		*pap = pa;
4778 
4779 	PMAPCOUNT(pt_mappings);
4780 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
4781 	KDASSERT(l2b != NULL);
4782 
4783 	ptep = &l2b->l2b_kva[l2pte_index(va)];
4784 	*ptep = L2_S_PROTO | pa | cache_mode |
4785 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ | VM_PROT_WRITE);
4786 	PTE_SYNC(ptep);
4787 	memset((void *)va, 0, PAGE_SIZE);
4788 	return (0);
4789 }
4790 
4791 /*
4792  * This is the same as pmap_alloc_l2_bucket(), except that it is only
4793  * used by pmap_growkernel().
4794  */
4795 static inline struct l2_bucket *
4796 pmap_grow_l2_bucket(pmap_t pm, vaddr_t va)
4797 {
4798 	struct l2_dtable *l2;
4799 	struct l2_bucket *l2b;
4800 	u_short l1idx;
4801 	vaddr_t nva;
4802 
4803 	l1idx = L1_IDX(va);
4804 
4805 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
4806 		/*
4807 		 * No mapping at this address, as there is
4808 		 * no entry in the L1 table.
4809 		 * Need to allocate a new l2_dtable.
4810 		 */
4811 		nva = pmap_kernel_l2dtable_kva;
4812 		if ((nva & PGOFSET) == 0) {
4813 			/*
4814 			 * Need to allocate a backing page
4815 			 */
4816 			if (pmap_grow_map(nva, pte_l2_s_cache_mode, NULL))
4817 				return (NULL);
4818 		}
4819 
4820 		l2 = (struct l2_dtable *)nva;
4821 		nva += sizeof(struct l2_dtable);
4822 
4823 		if ((nva & PGOFSET) < (pmap_kernel_l2dtable_kva & PGOFSET)) {
4824 			/*
4825 			 * The new l2_dtable straddles a page boundary.
4826 			 * Map in another page to cover it.
4827 			 */
4828 			if (pmap_grow_map(nva, pte_l2_s_cache_mode, NULL))
4829 				return (NULL);
4830 		}
4831 
4832 		pmap_kernel_l2dtable_kva = nva;
4833 
4834 		/*
4835 		 * Link it into the parent pmap
4836 		 */
4837 		pm->pm_l2[L2_IDX(l1idx)] = l2;
4838 	}
4839 
4840 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
4841 
4842 	/*
4843 	 * Fetch pointer to the L2 page table associated with the address.
4844 	 */
4845 	if (l2b->l2b_kva == NULL) {
4846 		pt_entry_t *ptep;
4847 
4848 		/*
4849 		 * No L2 page table has been allocated. Chances are, this
4850 		 * is because we just allocated the l2_dtable, above.
4851 		 */
4852 		nva = pmap_kernel_l2ptp_kva;
4853 		ptep = (pt_entry_t *)nva;
4854 		if ((nva & PGOFSET) == 0) {
4855 			/*
4856 			 * Need to allocate a backing page
4857 			 */
4858 			if (pmap_grow_map(nva, pte_l2_s_cache_mode_pt,
4859 			    &pmap_kernel_l2ptp_phys))
4860 				return (NULL);
4861 			PTE_SYNC_RANGE(ptep, PAGE_SIZE / sizeof(pt_entry_t));
4862 		}
4863 
4864 		l2->l2_occupancy++;
4865 		l2b->l2b_kva = ptep;
4866 		l2b->l2b_l1idx = l1idx;
4867 		l2b->l2b_phys = pmap_kernel_l2ptp_phys;
4868 
4869 		pmap_kernel_l2ptp_kva += L2_TABLE_SIZE_REAL;
4870 		pmap_kernel_l2ptp_phys += L2_TABLE_SIZE_REAL;
4871 	}
4872 
4873 	return (l2b);
4874 }
4875 
4876 vaddr_t
4877 pmap_growkernel(vaddr_t maxkvaddr)
4878 {
4879 	pmap_t kpm = pmap_kernel();
4880 	struct l1_ttable *l1;
4881 	struct l2_bucket *l2b;
4882 	pd_entry_t *pl1pd;
4883 	int s;
4884 
4885 	if (maxkvaddr <= pmap_curmaxkvaddr)
4886 		goto out;		/* we are OK */
4887 
4888 	NPDEBUG(PDB_GROWKERN,
4889 	    printf("pmap_growkernel: growing kernel from 0x%lx to 0x%lx\n",
4890 	    pmap_curmaxkvaddr, maxkvaddr));
4891 
4892 	KDASSERT(maxkvaddr <= virtual_end);
4893 
4894 	/*
4895 	 * whoops!   we need to add kernel PTPs
4896 	 */
4897 
4898 	s = splhigh();	/* to be safe */
4899 	mutex_enter(&kpm->pm_lock);
4900 
4901 	/* Map 1MB at a time */
4902 	for (; pmap_curmaxkvaddr < maxkvaddr; pmap_curmaxkvaddr += L1_S_SIZE) {
4903 
4904 		l2b = pmap_grow_l2_bucket(kpm, pmap_curmaxkvaddr);
4905 		KDASSERT(l2b != NULL);
4906 
4907 		/* Distribute new L1 entry to all other L1s */
4908 		SLIST_FOREACH(l1, &l1_list, l1_link) {
4909 			pl1pd = &l1->l1_kva[L1_IDX(pmap_curmaxkvaddr)];
4910 			*pl1pd = l2b->l2b_phys | L1_C_DOM(PMAP_DOMAIN_KERNEL) |
4911 			    L1_C_PROTO;
4912 			PTE_SYNC(pl1pd);
4913 		}
4914 	}
4915 
4916 	/*
4917 	 * flush out the cache, expensive but growkernel will happen so
4918 	 * rarely
4919 	 */
4920 	cpu_dcache_wbinv_all();
4921 	cpu_tlb_flushD();
4922 	cpu_cpwait();
4923 
4924 	mutex_exit(&kpm->pm_lock);
4925 	splx(s);
4926 
4927 out:
4928 	return (pmap_curmaxkvaddr);
4929 }
4930 
4931 /************************ Utility routines ****************************/
4932 
4933 /*
4934  * vector_page_setprot:
4935  *
4936  *	Manipulate the protection of the vector page.
4937  */
4938 void
4939 vector_page_setprot(int prot)
4940 {
4941 	struct l2_bucket *l2b;
4942 	pt_entry_t *ptep;
4943 
4944 	l2b = pmap_get_l2_bucket(pmap_kernel(), vector_page);
4945 	KDASSERT(l2b != NULL);
4946 
4947 	ptep = &l2b->l2b_kva[l2pte_index(vector_page)];
4948 
4949 	*ptep = (*ptep & ~L1_S_PROT_MASK) | L2_S_PROT(PTE_KERNEL, prot);
4950 	PTE_SYNC(ptep);
4951 	cpu_tlb_flushD_SE(vector_page);
4952 	cpu_cpwait();
4953 }
4954 
4955 /*
4956  * Fetch pointers to the PDE/PTE for the given pmap/VA pair.
4957  * Returns true if the mapping exists, else false.
4958  *
4959  * NOTE: This function is only used by a couple of arm-specific modules.
4960  * It is not safe to take any pmap locks here, since we could be right
4961  * in the middle of debugging the pmap anyway...
4962  *
4963  * It is possible for this routine to return false even though a valid
4964  * mapping does exist. This is because we don't lock, so the metadata
4965  * state may be inconsistent.
4966  *
4967  * NOTE: We can return a NULL *ptp in the case where the L1 pde is
4968  * a "section" mapping.
4969  */
4970 bool
4971 pmap_get_pde_pte(pmap_t pm, vaddr_t va, pd_entry_t **pdp, pt_entry_t **ptp)
4972 {
4973 	struct l2_dtable *l2;
4974 	pd_entry_t *pl1pd, l1pd;
4975 	pt_entry_t *ptep;
4976 	u_short l1idx;
4977 
4978 	if (pm->pm_l1 == NULL)
4979 		return false;
4980 
4981 	l1idx = L1_IDX(va);
4982 	*pdp = pl1pd = &pm->pm_l1->l1_kva[l1idx];
4983 	l1pd = *pl1pd;
4984 
4985 	if (l1pte_section_p(l1pd)) {
4986 		*ptp = NULL;
4987 		return true;
4988 	}
4989 
4990 	if (pm->pm_l2 == NULL)
4991 		return false;
4992 
4993 	l2 = pm->pm_l2[L2_IDX(l1idx)];
4994 
4995 	if (l2 == NULL ||
4996 	    (ptep = l2->l2_bucket[L2_BUCKET(l1idx)].l2b_kva) == NULL) {
4997 		return false;
4998 	}
4999 
5000 	*ptp = &ptep[l2pte_index(va)];
5001 	return true;
5002 }
5003 
5004 bool
5005 pmap_get_pde(pmap_t pm, vaddr_t va, pd_entry_t **pdp)
5006 {
5007 	u_short l1idx;
5008 
5009 	if (pm->pm_l1 == NULL)
5010 		return false;
5011 
5012 	l1idx = L1_IDX(va);
5013 	*pdp = &pm->pm_l1->l1_kva[l1idx];
5014 
5015 	return true;
5016 }
5017 
5018 /************************ Bootstrapping routines ****************************/
5019 
5020 static void
5021 pmap_init_l1(struct l1_ttable *l1, pd_entry_t *l1pt)
5022 {
5023 	int i;
5024 
5025 	l1->l1_kva = l1pt;
5026 	l1->l1_domain_use_count = 0;
5027 	l1->l1_domain_first = 0;
5028 
5029 	for (i = 0; i < PMAP_DOMAINS; i++)
5030 		l1->l1_domain_free[i] = i + 1;
5031 
5032 	/*
5033 	 * Copy the kernel's L1 entries to each new L1.
5034 	 */
5035 	if (pmap_initialized)
5036 		memcpy(l1pt, pmap_kernel()->pm_l1->l1_kva, L1_TABLE_SIZE);
5037 
5038 	if (pmap_extract(pmap_kernel(), (vaddr_t)l1pt,
5039 	    &l1->l1_physaddr) == false)
5040 		panic("pmap_init_l1: can't get PA of L1 at %p", l1pt);
5041 
5042 	SLIST_INSERT_HEAD(&l1_list, l1, l1_link);
5043 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
5044 }
5045 
5046 /*
5047  * pmap_bootstrap() is called from the board-specific initarm() routine
5048  * once the kernel L1/L2 descriptors tables have been set up.
5049  *
5050  * This is a somewhat convoluted process since pmap bootstrap is, effectively,
5051  * spread over a number of disparate files/functions.
5052  *
5053  * We are passed the following parameters
5054  *  - kernel_l1pt
5055  *    This is a pointer to the base of the kernel's L1 translation table.
5056  *  - vstart
5057  *    1MB-aligned start of managed kernel virtual memory.
5058  *  - vend
5059  *    1MB-aligned end of managed kernel virtual memory.
5060  *
5061  * We use the first parameter to build the metadata (struct l1_ttable and
5062  * struct l2_dtable) necessary to track kernel mappings.
5063  */
5064 #define	PMAP_STATIC_L2_SIZE 16
5065 void
5066 pmap_bootstrap(vaddr_t vstart, vaddr_t vend)
5067 {
5068 	static struct l1_ttable static_l1;
5069 	static struct l2_dtable static_l2[PMAP_STATIC_L2_SIZE];
5070 	struct l1_ttable *l1 = &static_l1;
5071 	struct l2_dtable *l2;
5072 	struct l2_bucket *l2b;
5073 	pd_entry_t *l1pt = (pd_entry_t *) kernel_l1pt.pv_va;
5074 	pmap_t pm = pmap_kernel();
5075 	pd_entry_t pde;
5076 	pt_entry_t *ptep;
5077 	paddr_t pa;
5078 	vaddr_t va;
5079 	vsize_t size;
5080 	int nptes, l1idx, l2idx, l2next = 0;
5081 
5082 	/*
5083 	 * Initialise the kernel pmap object
5084 	 */
5085 	pm->pm_l1 = l1;
5086 	pm->pm_domain = PMAP_DOMAIN_KERNEL;
5087 	pm->pm_activated = true;
5088 	pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
5089 	UVM_OBJ_INIT(&pm->pm_obj, NULL, 1);
5090 
5091 	/*
5092 	 * Scan the L1 translation table created by initarm() and create
5093 	 * the required metadata for all valid mappings found in it.
5094 	 */
5095 	for (l1idx = 0; l1idx < (L1_TABLE_SIZE / sizeof(pd_entry_t)); l1idx++) {
5096 		pde = l1pt[l1idx];
5097 
5098 		/*
5099 		 * We're only interested in Coarse mappings.
5100 		 * pmap_extract() can deal with section mappings without
5101 		 * recourse to checking L2 metadata.
5102 		 */
5103 		if ((pde & L1_TYPE_MASK) != L1_TYPE_C)
5104 			continue;
5105 
5106 		/*
5107 		 * Lookup the KVA of this L2 descriptor table
5108 		 */
5109 		pa = (paddr_t)(pde & L1_C_ADDR_MASK);
5110 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
5111 		if (ptep == NULL) {
5112 			panic("pmap_bootstrap: No L2 for va 0x%x, pa 0x%lx",
5113 			    (u_int)l1idx << L1_S_SHIFT, pa);
5114 		}
5115 
5116 		/*
5117 		 * Fetch the associated L2 metadata structure.
5118 		 * Allocate a new one if necessary.
5119 		 */
5120 		if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
5121 			if (l2next == PMAP_STATIC_L2_SIZE)
5122 				panic("pmap_bootstrap: out of static L2s");
5123 			pm->pm_l2[L2_IDX(l1idx)] = l2 = &static_l2[l2next++];
5124 		}
5125 
5126 		/*
5127 		 * One more L1 slot tracked...
5128 		 */
5129 		l2->l2_occupancy++;
5130 
5131 		/*
5132 		 * Fill in the details of the L2 descriptor in the
5133 		 * appropriate bucket.
5134 		 */
5135 		l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
5136 		l2b->l2b_kva = ptep;
5137 		l2b->l2b_phys = pa;
5138 		l2b->l2b_l1idx = l1idx;
5139 
5140 		/*
5141 		 * Establish an initial occupancy count for this descriptor
5142 		 */
5143 		for (l2idx = 0;
5144 		    l2idx < (L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
5145 		    l2idx++) {
5146 			if ((ptep[l2idx] & L2_TYPE_MASK) != L2_TYPE_INV) {
5147 				l2b->l2b_occupancy++;
5148 			}
5149 		}
5150 
5151 		/*
5152 		 * Make sure the descriptor itself has the correct cache mode.
5153 		 * If not, fix it, but whine about the problem. Port-meisters
5154 		 * should consider this a clue to fix up their initarm()
5155 		 * function. :)
5156 		 */
5157 		if (pmap_set_pt_cache_mode(l1pt, (vaddr_t)ptep)) {
5158 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
5159 			    "L2 pte @ %p\n", ptep);
5160 		}
5161 	}
5162 
5163 	/*
5164 	 * Ensure the primary (kernel) L1 has the correct cache mode for
5165 	 * a page table. Bitch if it is not correctly set.
5166 	 */
5167 	for (va = (vaddr_t)l1pt;
5168 	    va < ((vaddr_t)l1pt + L1_TABLE_SIZE); va += PAGE_SIZE) {
5169 		if (pmap_set_pt_cache_mode(l1pt, va))
5170 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
5171 			    "primary L1 @ 0x%lx\n", va);
5172 	}
5173 
5174 	cpu_dcache_wbinv_all();
5175 	cpu_tlb_flushID();
5176 	cpu_cpwait();
5177 
5178 	/*
5179 	 * now we allocate the "special" VAs which are used for tmp mappings
5180 	 * by the pmap (and other modules).  we allocate the VAs by advancing
5181 	 * virtual_avail (note that there are no pages mapped at these VAs).
5182 	 *
5183 	 * Managed KVM space start from wherever initarm() tells us.
5184 	 */
5185 	virtual_avail = vstart;
5186 	virtual_end = vend;
5187 
5188 #ifdef PMAP_CACHE_VIPT
5189 	/*
5190 	 * If we have a VIPT cache, we need one page/pte per possible alias
5191 	 * page so we won't violate cache aliasing rules.
5192 	 */
5193 	virtual_avail = (virtual_avail + arm_cache_prefer_mask) & ~arm_cache_prefer_mask;
5194 	nptes = (arm_cache_prefer_mask >> PGSHIFT) + 1;
5195 #else
5196 	nptes = 1;
5197 #endif
5198 	pmap_alloc_specials(&virtual_avail, nptes, &csrcp, &csrc_pte);
5199 	pmap_set_pt_cache_mode(l1pt, (vaddr_t)csrc_pte);
5200 	pmap_alloc_specials(&virtual_avail, nptes, &cdstp, &cdst_pte);
5201 	pmap_set_pt_cache_mode(l1pt, (vaddr_t)cdst_pte);
5202 	pmap_alloc_specials(&virtual_avail, nptes, &memhook, NULL);
5203 	pmap_alloc_specials(&virtual_avail, round_page(MSGBUFSIZE) / PAGE_SIZE,
5204 	    (void *)&msgbufaddr, NULL);
5205 
5206 	/*
5207 	 * Allocate a range of kernel virtual address space to be used
5208 	 * for L2 descriptor tables and metadata allocation in
5209 	 * pmap_growkernel().
5210 	 */
5211 	size = ((virtual_end - pmap_curmaxkvaddr) + L1_S_OFFSET) / L1_S_SIZE;
5212 	pmap_alloc_specials(&virtual_avail,
5213 	    round_page(size * L2_TABLE_SIZE_REAL) / PAGE_SIZE,
5214 	    &pmap_kernel_l2ptp_kva, NULL);
5215 
5216 	size = (size + (L2_BUCKET_SIZE - 1)) / L2_BUCKET_SIZE;
5217 	pmap_alloc_specials(&virtual_avail,
5218 	    round_page(size * sizeof(struct l2_dtable)) / PAGE_SIZE,
5219 	    &pmap_kernel_l2dtable_kva, NULL);
5220 
5221 	/*
5222 	 * init the static-global locks and global pmap list.
5223 	 */
5224 	/* spinlockinit(&pmap_main_lock, "pmaplk", 0); */
5225 
5226 	/*
5227 	 * We can now initialise the first L1's metadata.
5228 	 */
5229 	SLIST_INIT(&l1_list);
5230 	TAILQ_INIT(&l1_lru_list);
5231 	simple_lock_init(&l1_lru_lock);
5232 	pmap_init_l1(l1, l1pt);
5233 
5234 	/* Set up vector page L1 details, if necessary */
5235 	if (vector_page < KERNEL_BASE) {
5236 		pm->pm_pl1vec = &pm->pm_l1->l1_kva[L1_IDX(vector_page)];
5237 		l2b = pmap_get_l2_bucket(pm, vector_page);
5238 		KDASSERT(l2b != NULL);
5239 		pm->pm_l1vec = l2b->l2b_phys | L1_C_PROTO |
5240 		    L1_C_DOM(pm->pm_domain);
5241 	} else
5242 		pm->pm_pl1vec = NULL;
5243 
5244 	/*
5245 	 * Initialize the pmap cache
5246 	 */
5247 	pool_cache_bootstrap(&pmap_cache, sizeof(struct pmap), 0, 0, 0,
5248 	    "pmappl", NULL, IPL_NONE, pmap_pmap_ctor, NULL, NULL);
5249 	LIST_INIT(&pmap_pmaps);
5250 	LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list);
5251 
5252 	/*
5253 	 * Initialize the pv pool.
5254 	 */
5255 	pool_init(&pmap_pv_pool, sizeof(struct pv_entry), 0, 0, 0, "pvepl",
5256 	    &pmap_bootstrap_pv_allocator, IPL_NONE);
5257 
5258 	/*
5259 	 * Initialize the L2 dtable pool and cache.
5260 	 */
5261 	pool_cache_bootstrap(&pmap_l2dtable_cache, sizeof(struct l2_dtable), 0,
5262 	    0, 0, "l2dtblpl", NULL, IPL_NONE, pmap_l2dtable_ctor, NULL, NULL);
5263 
5264 	/*
5265 	 * Initialise the L2 descriptor table pool and cache
5266 	 */
5267 	pool_cache_bootstrap(&pmap_l2ptp_cache, L2_TABLE_SIZE_REAL, 0,
5268 	    L2_TABLE_SIZE_REAL, 0, "l2ptppl", NULL, IPL_NONE,
5269 	    pmap_l2ptp_ctor, NULL, NULL);
5270 
5271 	cpu_dcache_wbinv_all();
5272 }
5273 
5274 static int
5275 pmap_set_pt_cache_mode(pd_entry_t *kl1, vaddr_t va)
5276 {
5277 	pd_entry_t *pdep, pde;
5278 	pt_entry_t *ptep, pte;
5279 	vaddr_t pa;
5280 	int rv = 0;
5281 
5282 	/*
5283 	 * Make sure the descriptor itself has the correct cache mode
5284 	 */
5285 	pdep = &kl1[L1_IDX(va)];
5286 	pde = *pdep;
5287 
5288 	if (l1pte_section_p(pde)) {
5289 		if ((pde & L1_S_CACHE_MASK) != pte_l1_s_cache_mode_pt) {
5290 			*pdep = (pde & ~L1_S_CACHE_MASK) |
5291 			    pte_l1_s_cache_mode_pt;
5292 			PTE_SYNC(pdep);
5293 			cpu_dcache_wbinv_range((vaddr_t)pdep, sizeof(*pdep));
5294 			rv = 1;
5295 		}
5296 	} else {
5297 		pa = (paddr_t)(pde & L1_C_ADDR_MASK);
5298 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
5299 		if (ptep == NULL)
5300 			panic("pmap_bootstrap: No L2 for L2 @ va %p\n", ptep);
5301 
5302 		ptep = &ptep[l2pte_index(va)];
5303 		pte = *ptep;
5304 		if ((pte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
5305 			*ptep = (pte & ~L2_S_CACHE_MASK) |
5306 			    pte_l2_s_cache_mode_pt;
5307 			PTE_SYNC(ptep);
5308 			cpu_dcache_wbinv_range((vaddr_t)ptep, sizeof(*ptep));
5309 			rv = 1;
5310 		}
5311 	}
5312 
5313 	return (rv);
5314 }
5315 
5316 static void
5317 pmap_alloc_specials(vaddr_t *availp, int pages, vaddr_t *vap, pt_entry_t **ptep)
5318 {
5319 	vaddr_t va = *availp;
5320 	struct l2_bucket *l2b;
5321 
5322 	if (ptep) {
5323 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
5324 		if (l2b == NULL)
5325 			panic("pmap_alloc_specials: no l2b for 0x%lx", va);
5326 
5327 		if (ptep)
5328 			*ptep = &l2b->l2b_kva[l2pte_index(va)];
5329 	}
5330 
5331 	*vap = va;
5332 	*availp = va + (PAGE_SIZE * pages);
5333 }
5334 
5335 void
5336 pmap_init(void)
5337 {
5338 
5339 	/*
5340 	 * Set the available memory vars - These do not map to real memory
5341 	 * addresses and cannot as the physical memory is fragmented.
5342 	 * They are used by ps for %mem calculations.
5343 	 * One could argue whether this should be the entire memory or just
5344 	 * the memory that is useable in a user process.
5345 	 */
5346 	avail_start = ptoa(vm_physmem[0].start);
5347 	avail_end = ptoa(vm_physmem[vm_nphysseg - 1].end);
5348 
5349 	/*
5350 	 * Now we need to free enough pv_entry structures to allow us to get
5351 	 * the kmem_map/kmem_object allocated and inited (done after this
5352 	 * function is finished).  to do this we allocate one bootstrap page out
5353 	 * of kernel_map and use it to provide an initial pool of pv_entry
5354 	 * structures.   we never free this page.
5355 	 */
5356 	pool_setlowat(&pmap_pv_pool,
5357 	    (PAGE_SIZE / sizeof(struct pv_entry)) * 2);
5358 
5359 	mutex_init(&memlock, MUTEX_DEFAULT, IPL_NONE);
5360 	zeropage = (void *)uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
5361 	    UVM_KMF_WIRED|UVM_KMF_ZERO);
5362 
5363 	pmap_initialized = true;
5364 }
5365 
5366 static vaddr_t last_bootstrap_page = 0;
5367 static void *free_bootstrap_pages = NULL;
5368 
5369 static void *
5370 pmap_bootstrap_pv_page_alloc(struct pool *pp, int flags)
5371 {
5372 	extern void *pool_page_alloc(struct pool *, int);
5373 	vaddr_t new_page;
5374 	void *rv;
5375 
5376 	if (pmap_initialized)
5377 		return (pool_page_alloc(pp, flags));
5378 
5379 	if (free_bootstrap_pages) {
5380 		rv = free_bootstrap_pages;
5381 		free_bootstrap_pages = *((void **)rv);
5382 		return (rv);
5383 	}
5384 
5385 	new_page = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
5386 	    UVM_KMF_WIRED | ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT));
5387 
5388 	KASSERT(new_page > last_bootstrap_page);
5389 	last_bootstrap_page = new_page;
5390 	return ((void *)new_page);
5391 }
5392 
5393 static void
5394 pmap_bootstrap_pv_page_free(struct pool *pp, void *v)
5395 {
5396 	extern void pool_page_free(struct pool *, void *);
5397 
5398 	if ((vaddr_t)v <= last_bootstrap_page) {
5399 		*((void **)v) = free_bootstrap_pages;
5400 		free_bootstrap_pages = v;
5401 		return;
5402 	}
5403 
5404 	if (pmap_initialized) {
5405 		pool_page_free(pp, v);
5406 		return;
5407 	}
5408 }
5409 
5410 /*
5411  * pmap_postinit()
5412  *
5413  * This routine is called after the vm and kmem subsystems have been
5414  * initialised. This allows the pmap code to perform any initialisation
5415  * that can only be done one the memory allocation is in place.
5416  */
5417 void
5418 pmap_postinit(void)
5419 {
5420 	extern paddr_t physical_start, physical_end;
5421 	struct l2_bucket *l2b;
5422 	struct l1_ttable *l1;
5423 	struct pglist plist;
5424 	struct vm_page *m;
5425 	pd_entry_t *pl1pt;
5426 	pt_entry_t *ptep, pte;
5427 	vaddr_t va, eva;
5428 	u_int loop, needed;
5429 	int error;
5430 
5431 	pool_cache_setlowat(&pmap_l2ptp_cache,
5432 	    (PAGE_SIZE / L2_TABLE_SIZE_REAL) * 4);
5433 	pool_cache_setlowat(&pmap_l2dtable_cache,
5434 	    (PAGE_SIZE / sizeof(struct l2_dtable)) * 2);
5435 
5436 	needed = (maxproc / PMAP_DOMAINS) + ((maxproc % PMAP_DOMAINS) ? 1 : 0);
5437 	needed -= 1;
5438 
5439 	l1 = malloc(sizeof(*l1) * needed, M_VMPMAP, M_WAITOK);
5440 
5441 	for (loop = 0; loop < needed; loop++, l1++) {
5442 		/* Allocate a L1 page table */
5443 		va = uvm_km_alloc(kernel_map, L1_TABLE_SIZE, 0, UVM_KMF_VAONLY);
5444 		if (va == 0)
5445 			panic("Cannot allocate L1 KVM");
5446 
5447 		error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start,
5448 		    physical_end, L1_TABLE_SIZE, 0, &plist, 1, M_WAITOK);
5449 		if (error)
5450 			panic("Cannot allocate L1 physical pages");
5451 
5452 		m = TAILQ_FIRST(&plist);
5453 		eva = va + L1_TABLE_SIZE;
5454 		pl1pt = (pd_entry_t *)va;
5455 
5456 		while (m && va < eva) {
5457 			paddr_t pa = VM_PAGE_TO_PHYS(m);
5458 
5459 			pmap_kenter_pa(va, pa,
5460 			    VM_PROT_READ|VM_PROT_WRITE|PMAP_KMPAGE, 0);
5461 
5462 			/*
5463 			 * Make sure the L1 descriptor table is mapped
5464 			 * with the cache-mode set to write-through.
5465 			 */
5466 			l2b = pmap_get_l2_bucket(pmap_kernel(), va);
5467 			KDASSERT(l2b != NULL);
5468 			ptep = &l2b->l2b_kva[l2pte_index(va)];
5469 			pte = *ptep;
5470 			pte = (pte & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode_pt;
5471 			*ptep = pte;
5472 			PTE_SYNC(ptep);
5473 			cpu_tlb_flushD_SE(va);
5474 
5475 			va += PAGE_SIZE;
5476 			m = TAILQ_NEXT(m, pageq.queue);
5477 		}
5478 
5479 #ifdef DIAGNOSTIC
5480 		if (m)
5481 			panic("pmap_alloc_l1pt: pglist not empty");
5482 #endif	/* DIAGNOSTIC */
5483 
5484 		pmap_init_l1(l1, pl1pt);
5485 	}
5486 
5487 #ifdef DEBUG
5488 	printf("pmap_postinit: Allocated %d static L1 descriptor tables\n",
5489 	    needed);
5490 #endif
5491 }
5492 
5493 /*
5494  * Note that the following routines are used by board-specific initialisation
5495  * code to configure the initial kernel page tables.
5496  *
5497  * If ARM32_NEW_VM_LAYOUT is *not* defined, they operate on the assumption that
5498  * L2 page-table pages are 4KB in size and use 4 L1 slots. This mimics the
5499  * behaviour of the old pmap, and provides an easy migration path for
5500  * initial bring-up of the new pmap on existing ports. Fortunately,
5501  * pmap_bootstrap() compensates for this hackery. This is only a stop-gap and
5502  * will be deprecated.
5503  *
5504  * If ARM32_NEW_VM_LAYOUT *is* defined, these functions deal with 1KB L2 page
5505  * tables.
5506  */
5507 
5508 /*
5509  * This list exists for the benefit of pmap_map_chunk().  It keeps track
5510  * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
5511  * find them as necessary.
5512  *
5513  * Note that the data on this list MUST remain valid after initarm() returns,
5514  * as pmap_bootstrap() uses it to contruct L2 table metadata.
5515  */
5516 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
5517 
5518 static vaddr_t
5519 kernel_pt_lookup(paddr_t pa)
5520 {
5521 	pv_addr_t *pv;
5522 
5523 	SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
5524 #ifndef ARM32_NEW_VM_LAYOUT
5525 		if (pv->pv_pa == (pa & ~PGOFSET))
5526 			return (pv->pv_va | (pa & PGOFSET));
5527 #else
5528 		if (pv->pv_pa == pa)
5529 			return (pv->pv_va);
5530 #endif
5531 	}
5532 	return (0);
5533 }
5534 
5535 /*
5536  * pmap_map_section:
5537  *
5538  *	Create a single section mapping.
5539  */
5540 void
5541 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
5542 {
5543 	pd_entry_t *pde = (pd_entry_t *) l1pt;
5544 	pd_entry_t fl;
5545 
5546 	KASSERT(((va | pa) & L1_S_OFFSET) == 0);
5547 
5548 	switch (cache) {
5549 	case PTE_NOCACHE:
5550 	default:
5551 		fl = 0;
5552 		break;
5553 
5554 	case PTE_CACHE:
5555 		fl = pte_l1_s_cache_mode;
5556 		break;
5557 
5558 	case PTE_PAGETABLE:
5559 		fl = pte_l1_s_cache_mode_pt;
5560 		break;
5561 	}
5562 
5563 	pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
5564 	    L1_S_PROT(PTE_KERNEL, prot) | fl | L1_S_DOM(PMAP_DOMAIN_KERNEL);
5565 	PTE_SYNC(&pde[va >> L1_S_SHIFT]);
5566 }
5567 
5568 /*
5569  * pmap_map_entry:
5570  *
5571  *	Create a single page mapping.
5572  */
5573 void
5574 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
5575 {
5576 	pd_entry_t *pde = (pd_entry_t *) l1pt;
5577 	pt_entry_t fl;
5578 	pt_entry_t *pte;
5579 
5580 	KASSERT(((va | pa) & PGOFSET) == 0);
5581 
5582 	switch (cache) {
5583 	case PTE_NOCACHE:
5584 	default:
5585 		fl = 0;
5586 		break;
5587 
5588 	case PTE_CACHE:
5589 		fl = pte_l2_s_cache_mode;
5590 		break;
5591 
5592 	case PTE_PAGETABLE:
5593 		fl = pte_l2_s_cache_mode_pt;
5594 		break;
5595 	}
5596 
5597 	if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
5598 		panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
5599 
5600 #ifndef ARM32_NEW_VM_LAYOUT
5601 	pte = (pt_entry_t *)
5602 	    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
5603 #else
5604 	pte = (pt_entry_t *) kernel_pt_lookup(pde[L1_IDX(va)] & L1_C_ADDR_MASK);
5605 #endif
5606 	if (pte == NULL)
5607 		panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
5608 
5609 	fl |= L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot);
5610 #ifndef ARM32_NEW_VM_LAYOUT
5611 	pte += (va >> PGSHIFT) & 0x3ff;
5612 #else
5613 	pte += l2pte_index(va);
5614 	    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | fl;
5615 #endif
5616 	*pte = fl;
5617 	PTE_SYNC(pte);
5618 }
5619 
5620 /*
5621  * pmap_link_l2pt:
5622  *
5623  *	Link the L2 page table specified by "l2pv" into the L1
5624  *	page table at the slot for "va".
5625  */
5626 void
5627 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
5628 {
5629 	pd_entry_t *pde = (pd_entry_t *) l1pt, proto;
5630 	u_int slot = va >> L1_S_SHIFT;
5631 
5632 #ifndef ARM32_NEW_VM_LAYOUT
5633 	KASSERT((va & ((L1_S_SIZE * 4) - 1)) == 0);
5634 	KASSERT((l2pv->pv_pa & PGOFSET) == 0);
5635 #endif
5636 
5637 	proto = L1_S_DOM(PMAP_DOMAIN_KERNEL) | L1_C_PROTO;
5638 
5639 	pde[slot + 0] = proto | (l2pv->pv_pa + 0x000);
5640 #ifdef ARM32_NEW_VM_LAYOUT
5641 	PTE_SYNC(&pde[slot]);
5642 #else
5643 	pde[slot + 1] = proto | (l2pv->pv_pa + 0x400);
5644 	pde[slot + 2] = proto | (l2pv->pv_pa + 0x800);
5645 	pde[slot + 3] = proto | (l2pv->pv_pa + 0xc00);
5646 	PTE_SYNC_RANGE(&pde[slot + 0], 4);
5647 #endif
5648 
5649 	SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
5650 }
5651 
5652 /*
5653  * pmap_map_chunk:
5654  *
5655  *	Map a chunk of memory using the most efficient mappings
5656  *	possible (section, large page, small page) into the
5657  *	provided L1 and L2 tables at the specified virtual address.
5658  */
5659 vsize_t
5660 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
5661     int prot, int cache)
5662 {
5663 	pd_entry_t *pde = (pd_entry_t *) l1pt;
5664 	pt_entry_t *pte, f1, f2s, f2l;
5665 	vsize_t resid;
5666 	int i;
5667 
5668 	resid = (size + (PAGE_SIZE - 1)) & ~(PAGE_SIZE - 1);
5669 
5670 	if (l1pt == 0)
5671 		panic("pmap_map_chunk: no L1 table provided");
5672 
5673 #ifdef VERBOSE_INIT_ARM
5674 	printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
5675 	    "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
5676 #endif
5677 
5678 	switch (cache) {
5679 	case PTE_NOCACHE:
5680 	default:
5681 		f1 = 0;
5682 		f2l = 0;
5683 		f2s = 0;
5684 		break;
5685 
5686 	case PTE_CACHE:
5687 		f1 = pte_l1_s_cache_mode;
5688 		f2l = pte_l2_l_cache_mode;
5689 		f2s = pte_l2_s_cache_mode;
5690 		break;
5691 
5692 	case PTE_PAGETABLE:
5693 		f1 = pte_l1_s_cache_mode_pt;
5694 		f2l = pte_l2_l_cache_mode_pt;
5695 		f2s = pte_l2_s_cache_mode_pt;
5696 		break;
5697 	}
5698 
5699 	size = resid;
5700 
5701 	while (resid > 0) {
5702 		/* See if we can use a section mapping. */
5703 		if (L1_S_MAPPABLE_P(va, pa, resid)) {
5704 #ifdef VERBOSE_INIT_ARM
5705 			printf("S");
5706 #endif
5707 			pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
5708 			    L1_S_PROT(PTE_KERNEL, prot) | f1 |
5709 			    L1_S_DOM(PMAP_DOMAIN_KERNEL);
5710 			PTE_SYNC(&pde[va >> L1_S_SHIFT]);
5711 			va += L1_S_SIZE;
5712 			pa += L1_S_SIZE;
5713 			resid -= L1_S_SIZE;
5714 			continue;
5715 		}
5716 
5717 		/*
5718 		 * Ok, we're going to use an L2 table.  Make sure
5719 		 * one is actually in the corresponding L1 slot
5720 		 * for the current VA.
5721 		 */
5722 		if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
5723 			panic("pmap_map_chunk: no L2 table for VA 0x%08lx", va);
5724 
5725 #ifndef ARM32_NEW_VM_LAYOUT
5726 		pte = (pt_entry_t *)
5727 		    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
5728 #else
5729 		pte = (pt_entry_t *) kernel_pt_lookup(
5730 		    pde[L1_IDX(va)] & L1_C_ADDR_MASK);
5731 #endif
5732 		if (pte == NULL)
5733 			panic("pmap_map_chunk: can't find L2 table for VA"
5734 			    "0x%08lx", va);
5735 
5736 		/* See if we can use a L2 large page mapping. */
5737 		if (L2_L_MAPPABLE_P(va, pa, resid)) {
5738 #ifdef VERBOSE_INIT_ARM
5739 			printf("L");
5740 #endif
5741 			for (i = 0; i < 16; i++) {
5742 #ifndef ARM32_NEW_VM_LAYOUT
5743 				pte[((va >> PGSHIFT) & 0x3f0) + i] =
5744 				    L2_L_PROTO | pa |
5745 				    L2_L_PROT(PTE_KERNEL, prot) | f2l;
5746 				PTE_SYNC(&pte[((va >> PGSHIFT) & 0x3f0) + i]);
5747 #else
5748 				pte[l2pte_index(va) + i] =
5749 				    L2_L_PROTO | pa |
5750 				    L2_L_PROT(PTE_KERNEL, prot) | f2l;
5751 				PTE_SYNC(&pte[l2pte_index(va) + i]);
5752 #endif
5753 			}
5754 			va += L2_L_SIZE;
5755 			pa += L2_L_SIZE;
5756 			resid -= L2_L_SIZE;
5757 			continue;
5758 		}
5759 
5760 		/* Use a small page mapping. */
5761 #ifdef VERBOSE_INIT_ARM
5762 		printf("P");
5763 #endif
5764 #ifndef ARM32_NEW_VM_LAYOUT
5765 		pte[(va >> PGSHIFT) & 0x3ff] =
5766 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | f2s;
5767 		PTE_SYNC(&pte[(va >> PGSHIFT) & 0x3ff]);
5768 #else
5769 		pte[l2pte_index(va)] =
5770 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | f2s;
5771 		PTE_SYNC(&pte[l2pte_index(va)]);
5772 #endif
5773 		va += PAGE_SIZE;
5774 		pa += PAGE_SIZE;
5775 		resid -= PAGE_SIZE;
5776 	}
5777 #ifdef VERBOSE_INIT_ARM
5778 	printf("\n");
5779 #endif
5780 	return (size);
5781 }
5782 
5783 /********************** Static device map routines ***************************/
5784 
5785 static const struct pmap_devmap *pmap_devmap_table;
5786 
5787 /*
5788  * Register the devmap table.  This is provided in case early console
5789  * initialization needs to register mappings created by bootstrap code
5790  * before pmap_devmap_bootstrap() is called.
5791  */
5792 void
5793 pmap_devmap_register(const struct pmap_devmap *table)
5794 {
5795 
5796 	pmap_devmap_table = table;
5797 }
5798 
5799 /*
5800  * Map all of the static regions in the devmap table, and remember
5801  * the devmap table so other parts of the kernel can look up entries
5802  * later.
5803  */
5804 void
5805 pmap_devmap_bootstrap(vaddr_t l1pt, const struct pmap_devmap *table)
5806 {
5807 	int i;
5808 
5809 	pmap_devmap_table = table;
5810 
5811 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
5812 #ifdef VERBOSE_INIT_ARM
5813 		printf("devmap: %08lx -> %08lx @ %08lx\n",
5814 		    pmap_devmap_table[i].pd_pa,
5815 		    pmap_devmap_table[i].pd_pa +
5816 			pmap_devmap_table[i].pd_size - 1,
5817 		    pmap_devmap_table[i].pd_va);
5818 #endif
5819 		pmap_map_chunk(l1pt, pmap_devmap_table[i].pd_va,
5820 		    pmap_devmap_table[i].pd_pa,
5821 		    pmap_devmap_table[i].pd_size,
5822 		    pmap_devmap_table[i].pd_prot,
5823 		    pmap_devmap_table[i].pd_cache);
5824 	}
5825 }
5826 
5827 const struct pmap_devmap *
5828 pmap_devmap_find_pa(paddr_t pa, psize_t size)
5829 {
5830 	uint64_t endpa;
5831 	int i;
5832 
5833 	if (pmap_devmap_table == NULL)
5834 		return (NULL);
5835 
5836 	endpa = (uint64_t)pa + (uint64_t)(size - 1);
5837 
5838 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
5839 		if (pa >= pmap_devmap_table[i].pd_pa &&
5840 		    endpa <= (uint64_t)pmap_devmap_table[i].pd_pa +
5841 			     (uint64_t)(pmap_devmap_table[i].pd_size - 1))
5842 			return (&pmap_devmap_table[i]);
5843 	}
5844 
5845 	return (NULL);
5846 }
5847 
5848 const struct pmap_devmap *
5849 pmap_devmap_find_va(vaddr_t va, vsize_t size)
5850 {
5851 	int i;
5852 
5853 	if (pmap_devmap_table == NULL)
5854 		return (NULL);
5855 
5856 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
5857 		if (va >= pmap_devmap_table[i].pd_va &&
5858 		    va + size - 1 <= pmap_devmap_table[i].pd_va +
5859 				     pmap_devmap_table[i].pd_size - 1)
5860 			return (&pmap_devmap_table[i]);
5861 	}
5862 
5863 	return (NULL);
5864 }
5865 
5866 /********************** PTE initialization routines **************************/
5867 
5868 /*
5869  * These routines are called when the CPU type is identified to set up
5870  * the PTE prototypes, cache modes, etc.
5871  *
5872  * The variables are always here, just in case modules need to reference
5873  * them (though, they shouldn't).
5874  */
5875 
5876 pt_entry_t	pte_l1_s_cache_mode;
5877 pt_entry_t	pte_l1_s_cache_mode_pt;
5878 pt_entry_t	pte_l1_s_cache_mask;
5879 
5880 pt_entry_t	pte_l2_l_cache_mode;
5881 pt_entry_t	pte_l2_l_cache_mode_pt;
5882 pt_entry_t	pte_l2_l_cache_mask;
5883 
5884 pt_entry_t	pte_l2_s_cache_mode;
5885 pt_entry_t	pte_l2_s_cache_mode_pt;
5886 pt_entry_t	pte_l2_s_cache_mask;
5887 
5888 pt_entry_t	pte_l2_s_prot_u;
5889 pt_entry_t	pte_l2_s_prot_w;
5890 pt_entry_t	pte_l2_s_prot_mask;
5891 
5892 pt_entry_t	pte_l1_s_proto;
5893 pt_entry_t	pte_l1_c_proto;
5894 pt_entry_t	pte_l2_s_proto;
5895 
5896 void		(*pmap_copy_page_func)(paddr_t, paddr_t);
5897 void		(*pmap_zero_page_func)(paddr_t);
5898 
5899 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0
5900 void
5901 pmap_pte_init_generic(void)
5902 {
5903 
5904 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
5905 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_generic;
5906 
5907 	pte_l2_l_cache_mode = L2_B|L2_C;
5908 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_generic;
5909 
5910 	pte_l2_s_cache_mode = L2_B|L2_C;
5911 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_generic;
5912 
5913 	/*
5914 	 * If we have a write-through cache, set B and C.  If
5915 	 * we have a write-back cache, then we assume setting
5916 	 * only C will make those pages write-through.
5917 	 */
5918 	if (cpufuncs.cf_dcache_wb_range == (void *) cpufunc_nullop) {
5919 		pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
5920 		pte_l2_l_cache_mode_pt = L2_B|L2_C;
5921 		pte_l2_s_cache_mode_pt = L2_B|L2_C;
5922 	} else {
5923 #if ARM_MMU_V6 > 1
5924 		pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C; /* arm116 errata 399234 */
5925 		pte_l2_l_cache_mode_pt = L2_B|L2_C; /* arm116 errata 399234 */
5926 		pte_l2_s_cache_mode_pt = L2_B|L2_C; /* arm116 errata 399234 */
5927 #else
5928 		pte_l1_s_cache_mode_pt = L1_S_C;
5929 		pte_l2_l_cache_mode_pt = L2_C;
5930 		pte_l2_s_cache_mode_pt = L2_C;
5931 #endif
5932 	}
5933 
5934 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
5935 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
5936 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
5937 
5938 	pte_l1_s_proto = L1_S_PROTO_generic;
5939 	pte_l1_c_proto = L1_C_PROTO_generic;
5940 	pte_l2_s_proto = L2_S_PROTO_generic;
5941 
5942 	pmap_copy_page_func = pmap_copy_page_generic;
5943 	pmap_zero_page_func = pmap_zero_page_generic;
5944 }
5945 
5946 #if defined(CPU_ARM8)
5947 void
5948 pmap_pte_init_arm8(void)
5949 {
5950 
5951 	/*
5952 	 * ARM8 is compatible with generic, but we need to use
5953 	 * the page tables uncached.
5954 	 */
5955 	pmap_pte_init_generic();
5956 
5957 	pte_l1_s_cache_mode_pt = 0;
5958 	pte_l2_l_cache_mode_pt = 0;
5959 	pte_l2_s_cache_mode_pt = 0;
5960 }
5961 #endif /* CPU_ARM8 */
5962 
5963 #if defined(CPU_ARM9) && defined(ARM9_CACHE_WRITE_THROUGH)
5964 void
5965 pmap_pte_init_arm9(void)
5966 {
5967 
5968 	/*
5969 	 * ARM9 is compatible with generic, but we want to use
5970 	 * write-through caching for now.
5971 	 */
5972 	pmap_pte_init_generic();
5973 
5974 	pte_l1_s_cache_mode = L1_S_C;
5975 	pte_l2_l_cache_mode = L2_C;
5976 	pte_l2_s_cache_mode = L2_C;
5977 
5978 	pte_l1_s_cache_mode_pt = L1_S_C;
5979 	pte_l2_l_cache_mode_pt = L2_C;
5980 	pte_l2_s_cache_mode_pt = L2_C;
5981 }
5982 #endif /* CPU_ARM9 && ARM9_CACHE_WRITE_THROUGH */
5983 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
5984 
5985 #if defined(CPU_ARM10)
5986 void
5987 pmap_pte_init_arm10(void)
5988 {
5989 
5990 	/*
5991 	 * ARM10 is compatible with generic, but we want to use
5992 	 * write-through caching for now.
5993 	 */
5994 	pmap_pte_init_generic();
5995 
5996 	pte_l1_s_cache_mode = L1_S_B | L1_S_C;
5997 	pte_l2_l_cache_mode = L2_B | L2_C;
5998 	pte_l2_s_cache_mode = L2_B | L2_C;
5999 
6000 	pte_l1_s_cache_mode_pt = L1_S_C;
6001 	pte_l2_l_cache_mode_pt = L2_C;
6002 	pte_l2_s_cache_mode_pt = L2_C;
6003 
6004 }
6005 #endif /* CPU_ARM10 */
6006 
6007 #if defined(CPU_ARM11) && defined(ARM11_CACHE_WRITE_THROUGH)
6008 void
6009 pmap_pte_init_arm11(void)
6010 {
6011 
6012 	/*
6013 	 * ARM11 is compatible with generic, but we want to use
6014 	 * write-through caching for now.
6015 	 */
6016 	pmap_pte_init_generic();
6017 
6018 	pte_l1_s_cache_mode = L1_S_C;
6019 	pte_l2_l_cache_mode = L2_C;
6020 	pte_l2_s_cache_mode = L2_C;
6021 
6022 	pte_l1_s_cache_mode_pt = L1_S_C;
6023 	pte_l2_l_cache_mode_pt = L2_C;
6024 	pte_l2_s_cache_mode_pt = L2_C;
6025 }
6026 #endif /* CPU_ARM11 && ARM11_CACHE_WRITE_THROUGH */
6027 
6028 #if ARM_MMU_SA1 == 1
6029 void
6030 pmap_pte_init_sa1(void)
6031 {
6032 
6033 	/*
6034 	 * The StrongARM SA-1 cache does not have a write-through
6035 	 * mode.  So, do the generic initialization, then reset
6036 	 * the page table cache mode to B=1,C=1, and note that
6037 	 * the PTEs need to be sync'd.
6038 	 */
6039 	pmap_pte_init_generic();
6040 
6041 	pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
6042 	pte_l2_l_cache_mode_pt = L2_B|L2_C;
6043 	pte_l2_s_cache_mode_pt = L2_B|L2_C;
6044 
6045 	pmap_needs_pte_sync = 1;
6046 }
6047 #endif /* ARM_MMU_SA1 == 1*/
6048 
6049 #if ARM_MMU_XSCALE == 1
6050 #if (ARM_NMMUS > 1)
6051 static u_int xscale_use_minidata;
6052 #endif
6053 
6054 void
6055 pmap_pte_init_xscale(void)
6056 {
6057 	uint32_t auxctl;
6058 	int write_through = 0;
6059 
6060 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
6061 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_xscale;
6062 
6063 	pte_l2_l_cache_mode = L2_B|L2_C;
6064 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_xscale;
6065 
6066 	pte_l2_s_cache_mode = L2_B|L2_C;
6067 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_xscale;
6068 
6069 	pte_l1_s_cache_mode_pt = L1_S_C;
6070 	pte_l2_l_cache_mode_pt = L2_C;
6071 	pte_l2_s_cache_mode_pt = L2_C;
6072 
6073 #ifdef XSCALE_CACHE_READ_WRITE_ALLOCATE
6074 	/*
6075 	 * The XScale core has an enhanced mode where writes that
6076 	 * miss the cache cause a cache line to be allocated.  This
6077 	 * is significantly faster than the traditional, write-through
6078 	 * behavior of this case.
6079 	 */
6080 	pte_l1_s_cache_mode |= L1_S_XS_TEX(TEX_XSCALE_X);
6081 	pte_l2_l_cache_mode |= L2_XS_L_TEX(TEX_XSCALE_X);
6082 	pte_l2_s_cache_mode |= L2_XS_T_TEX(TEX_XSCALE_X);
6083 #endif /* XSCALE_CACHE_READ_WRITE_ALLOCATE */
6084 
6085 #ifdef XSCALE_CACHE_WRITE_THROUGH
6086 	/*
6087 	 * Some versions of the XScale core have various bugs in
6088 	 * their cache units, the work-around for which is to run
6089 	 * the cache in write-through mode.  Unfortunately, this
6090 	 * has a major (negative) impact on performance.  So, we
6091 	 * go ahead and run fast-and-loose, in the hopes that we
6092 	 * don't line up the planets in a way that will trip the
6093 	 * bugs.
6094 	 *
6095 	 * However, we give you the option to be slow-but-correct.
6096 	 */
6097 	write_through = 1;
6098 #elif defined(XSCALE_CACHE_WRITE_BACK)
6099 	/* force write back cache mode */
6100 	write_through = 0;
6101 #elif defined(CPU_XSCALE_PXA250) || defined(CPU_XSCALE_PXA270)
6102 	/*
6103 	 * Intel PXA2[15]0 processors are known to have a bug in
6104 	 * write-back cache on revision 4 and earlier (stepping
6105 	 * A[01] and B[012]).  Fixed for C0 and later.
6106 	 */
6107 	{
6108 		uint32_t id, type;
6109 
6110 		id = cpufunc_id();
6111 		type = id & ~(CPU_ID_XSCALE_COREREV_MASK|CPU_ID_REVISION_MASK);
6112 
6113 		if (type == CPU_ID_PXA250 || type == CPU_ID_PXA210) {
6114 			if ((id & CPU_ID_REVISION_MASK) < 5) {
6115 				/* write through for stepping A0-1 and B0-2 */
6116 				write_through = 1;
6117 			}
6118 		}
6119 	}
6120 #endif /* XSCALE_CACHE_WRITE_THROUGH */
6121 
6122 	if (write_through) {
6123 		pte_l1_s_cache_mode = L1_S_C;
6124 		pte_l2_l_cache_mode = L2_C;
6125 		pte_l2_s_cache_mode = L2_C;
6126 	}
6127 
6128 #if (ARM_NMMUS > 1)
6129 	xscale_use_minidata = 1;
6130 #endif
6131 
6132 	pte_l2_s_prot_u = L2_S_PROT_U_xscale;
6133 	pte_l2_s_prot_w = L2_S_PROT_W_xscale;
6134 	pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale;
6135 
6136 	pte_l1_s_proto = L1_S_PROTO_xscale;
6137 	pte_l1_c_proto = L1_C_PROTO_xscale;
6138 	pte_l2_s_proto = L2_S_PROTO_xscale;
6139 
6140 	pmap_copy_page_func = pmap_copy_page_xscale;
6141 	pmap_zero_page_func = pmap_zero_page_xscale;
6142 
6143 	/*
6144 	 * Disable ECC protection of page table access, for now.
6145 	 */
6146 	__asm volatile("mrc p15, 0, %0, c1, c0, 1" : "=r" (auxctl));
6147 	auxctl &= ~XSCALE_AUXCTL_P;
6148 	__asm volatile("mcr p15, 0, %0, c1, c0, 1" : : "r" (auxctl));
6149 }
6150 
6151 /*
6152  * xscale_setup_minidata:
6153  *
6154  *	Set up the mini-data cache clean area.  We require the
6155  *	caller to allocate the right amount of physically and
6156  *	virtually contiguous space.
6157  */
6158 void
6159 xscale_setup_minidata(vaddr_t l1pt, vaddr_t va, paddr_t pa)
6160 {
6161 	extern vaddr_t xscale_minidata_clean_addr;
6162 	extern vsize_t xscale_minidata_clean_size; /* already initialized */
6163 	pd_entry_t *pde = (pd_entry_t *) l1pt;
6164 	pt_entry_t *pte;
6165 	vsize_t size;
6166 	uint32_t auxctl;
6167 
6168 	xscale_minidata_clean_addr = va;
6169 
6170 	/* Round it to page size. */
6171 	size = (xscale_minidata_clean_size + L2_S_OFFSET) & L2_S_FRAME;
6172 
6173 	for (; size != 0;
6174 	     va += L2_S_SIZE, pa += L2_S_SIZE, size -= L2_S_SIZE) {
6175 #ifndef ARM32_NEW_VM_LAYOUT
6176 		pte = (pt_entry_t *)
6177 		    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
6178 #else
6179 		pte = (pt_entry_t *) kernel_pt_lookup(
6180 		    pde[L1_IDX(va)] & L1_C_ADDR_MASK);
6181 #endif
6182 		if (pte == NULL)
6183 			panic("xscale_setup_minidata: can't find L2 table for "
6184 			    "VA 0x%08lx", va);
6185 #ifndef ARM32_NEW_VM_LAYOUT
6186 		pte[(va >> PGSHIFT) & 0x3ff] =
6187 #else
6188 		pte[l2pte_index(va)] =
6189 #endif
6190 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
6191 		    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);
6192 	}
6193 
6194 	/*
6195 	 * Configure the mini-data cache for write-back with
6196 	 * read/write-allocate.
6197 	 *
6198 	 * NOTE: In order to reconfigure the mini-data cache, we must
6199 	 * make sure it contains no valid data!  In order to do that,
6200 	 * we must issue a global data cache invalidate command!
6201 	 *
6202 	 * WE ASSUME WE ARE RUNNING UN-CACHED WHEN THIS ROUTINE IS CALLED!
6203 	 * THIS IS VERY IMPORTANT!
6204 	 */
6205 
6206 	/* Invalidate data and mini-data. */
6207 	__asm volatile("mcr p15, 0, %0, c7, c6, 0" : : "r" (0));
6208 	__asm volatile("mrc p15, 0, %0, c1, c0, 1" : "=r" (auxctl));
6209 	auxctl = (auxctl & ~XSCALE_AUXCTL_MD_MASK) | XSCALE_AUXCTL_MD_WB_RWA;
6210 	__asm volatile("mcr p15, 0, %0, c1, c0, 1" : : "r" (auxctl));
6211 }
6212 
6213 /*
6214  * Change the PTEs for the specified kernel mappings such that they
6215  * will use the mini data cache instead of the main data cache.
6216  */
6217 void
6218 pmap_uarea(vaddr_t va)
6219 {
6220 	struct l2_bucket *l2b;
6221 	pt_entry_t *ptep, *sptep, pte;
6222 	vaddr_t next_bucket, eva;
6223 
6224 #if (ARM_NMMUS > 1)
6225 	if (xscale_use_minidata == 0)
6226 		return;
6227 #endif
6228 
6229 	eva = va + USPACE;
6230 
6231 	while (va < eva) {
6232 		next_bucket = L2_NEXT_BUCKET(va);
6233 		if (next_bucket > eva)
6234 			next_bucket = eva;
6235 
6236 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
6237 		KDASSERT(l2b != NULL);
6238 
6239 		sptep = ptep = &l2b->l2b_kva[l2pte_index(va)];
6240 
6241 		while (va < next_bucket) {
6242 			pte = *ptep;
6243 			if (!l2pte_minidata(pte)) {
6244 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
6245 				cpu_tlb_flushD_SE(va);
6246 				*ptep = pte & ~L2_B;
6247 			}
6248 			ptep++;
6249 			va += PAGE_SIZE;
6250 		}
6251 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
6252 	}
6253 	cpu_cpwait();
6254 }
6255 #endif /* ARM_MMU_XSCALE == 1 */
6256 
6257 /*
6258  * return the PA of the current L1 table, for use when handling a crash dump
6259  */
6260 uint32_t pmap_kernel_L1_addr(void)
6261 {
6262 	return pmap_kernel()->pm_l1->l1_physaddr;
6263 }
6264 
6265 #if defined(DDB)
6266 /*
6267  * A couple of ddb-callable functions for dumping pmaps
6268  */
6269 void pmap_dump_all(void);
6270 void pmap_dump(pmap_t);
6271 
6272 void
6273 pmap_dump_all(void)
6274 {
6275 	pmap_t pm;
6276 
6277 	LIST_FOREACH(pm, &pmap_pmaps, pm_list) {
6278 		if (pm == pmap_kernel())
6279 			continue;
6280 		pmap_dump(pm);
6281 		printf("\n");
6282 	}
6283 }
6284 
6285 static pt_entry_t ncptes[64];
6286 static void pmap_dump_ncpg(pmap_t);
6287 
6288 void
6289 pmap_dump(pmap_t pm)
6290 {
6291 	struct l2_dtable *l2;
6292 	struct l2_bucket *l2b;
6293 	pt_entry_t *ptep, pte;
6294 	vaddr_t l2_va, l2b_va, va;
6295 	int i, j, k, occ, rows = 0;
6296 
6297 	if (pm == pmap_kernel())
6298 		printf("pmap_kernel (%p): ", pm);
6299 	else
6300 		printf("user pmap (%p): ", pm);
6301 
6302 	printf("domain %d, l1 at %p\n", pm->pm_domain, pm->pm_l1->l1_kva);
6303 
6304 	l2_va = 0;
6305 	for (i = 0; i < L2_SIZE; i++, l2_va += 0x01000000) {
6306 		l2 = pm->pm_l2[i];
6307 
6308 		if (l2 == NULL || l2->l2_occupancy == 0)
6309 			continue;
6310 
6311 		l2b_va = l2_va;
6312 		for (j = 0; j < L2_BUCKET_SIZE; j++, l2b_va += 0x00100000) {
6313 			l2b = &l2->l2_bucket[j];
6314 
6315 			if (l2b->l2b_occupancy == 0 || l2b->l2b_kva == NULL)
6316 				continue;
6317 
6318 			ptep = l2b->l2b_kva;
6319 
6320 			for (k = 0; k < 256 && ptep[k] == 0; k++)
6321 				;
6322 
6323 			k &= ~63;
6324 			occ = l2b->l2b_occupancy;
6325 			va = l2b_va + (k * 4096);
6326 			for (; k < 256; k++, va += 0x1000) {
6327 				char ch = ' ';
6328 				if ((k % 64) == 0) {
6329 					if ((rows % 8) == 0) {
6330 						printf(
6331 "          |0000   |8000   |10000  |18000  |20000  |28000  |30000  |38000\n");
6332 					}
6333 					printf("%08lx: ", va);
6334 				}
6335 
6336 				ncptes[k & 63] = 0;
6337 				pte = ptep[k];
6338 				if (pte == 0) {
6339 					ch = '.';
6340 				} else {
6341 					occ--;
6342 					switch (pte & 0x0c) {
6343 					case 0x00:
6344 						ch = 'D'; /* No cache No buff */
6345 						break;
6346 					case 0x04:
6347 						ch = 'B'; /* No cache buff */
6348 						break;
6349 					case 0x08:
6350 						if (pte & 0x40)
6351 							ch = 'm';
6352 						else
6353 						   ch = 'C'; /* Cache No buff */
6354 						break;
6355 					case 0x0c:
6356 						ch = 'F'; /* Cache Buff */
6357 						break;
6358 					}
6359 
6360 					if ((pte & L2_S_PROT_U) == L2_S_PROT_U)
6361 						ch += 0x20;
6362 
6363 					if ((pte & 0xc) == 0)
6364 						ncptes[k & 63] = pte;
6365 				}
6366 
6367 				if ((k % 64) == 63) {
6368 					rows++;
6369 					printf("%c\n", ch);
6370 					pmap_dump_ncpg(pm);
6371 					if (occ == 0)
6372 						break;
6373 				} else
6374 					printf("%c", ch);
6375 			}
6376 		}
6377 	}
6378 }
6379 
6380 static void
6381 pmap_dump_ncpg(pmap_t pm)
6382 {
6383 	struct vm_page *pg;
6384 	struct pv_entry *pv;
6385 	int i;
6386 
6387 	for (i = 0; i < 63; i++) {
6388 		if (ncptes[i] == 0)
6389 			continue;
6390 
6391 		pg = PHYS_TO_VM_PAGE(l2pte_pa(ncptes[i]));
6392 		if (pg == NULL)
6393 			continue;
6394 
6395 		printf(" pa 0x%08lx: krw %d kro %d urw %d uro %d\n",
6396 		    VM_PAGE_TO_PHYS(pg),
6397 		    pg->mdpage.krw_mappings, pg->mdpage.kro_mappings,
6398 		    pg->mdpage.urw_mappings, pg->mdpage.uro_mappings);
6399 
6400 		SLIST_FOREACH(pv, &pg->mdpage.pvh_list, pv_link) {
6401 			printf("   %c va 0x%08lx, flags 0x%x\n",
6402 			    (pm == pv->pv_pmap) ? '*' : ' ',
6403 			    pv->pv_va, pv->pv_flags);
6404 		}
6405 	}
6406 }
6407 #endif
6408 
6409 #ifdef PMAP_STEAL_MEMORY
6410 void
6411 pmap_boot_pageadd(pv_addr_t *newpv)
6412 {
6413 	pv_addr_t *pv, *npv;
6414 
6415 	if ((pv = SLIST_FIRST(&pmap_boot_freeq)) != NULL) {
6416 		if (newpv->pv_pa < pv->pv_va) {
6417 			KASSERT(newpv->pv_pa + newpv->pv_size <= pv->pv_pa);
6418 			if (newpv->pv_pa + newpv->pv_size == pv->pv_pa) {
6419 				newpv->pv_size += pv->pv_size;
6420 				SLIST_REMOVE_HEAD(&pmap_boot_freeq, pv_list);
6421 			}
6422 			pv = NULL;
6423 		} else {
6424 			for (; (npv = SLIST_NEXT(pv, pv_list)) != NULL;
6425 			     pv = npv) {
6426 				KASSERT(pv->pv_pa + pv->pv_size < npv->pv_pa);
6427 				KASSERT(pv->pv_pa < newpv->pv_pa);
6428 				if (newpv->pv_pa > npv->pv_pa)
6429 					continue;
6430 				if (pv->pv_pa + pv->pv_size == newpv->pv_pa) {
6431 					pv->pv_size += newpv->pv_size;
6432 					return;
6433 				}
6434 				if (newpv->pv_pa + newpv->pv_size < npv->pv_pa)
6435 					break;
6436 				newpv->pv_size += npv->pv_size;
6437 				SLIST_INSERT_AFTER(pv, newpv, pv_list);
6438 				SLIST_REMOVE_AFTER(newpv, pv_list);
6439 				return;
6440 			}
6441 		}
6442 	}
6443 
6444 	if (pv) {
6445 		SLIST_INSERT_AFTER(pv, newpv, pv_list);
6446 	} else {
6447 		SLIST_INSERT_HEAD(&pmap_boot_freeq, newpv, pv_list);
6448 	}
6449 }
6450 
6451 void
6452 pmap_boot_pagealloc(psize_t amount, psize_t mask, psize_t match,
6453 	pv_addr_t *rpv)
6454 {
6455 	pv_addr_t *pv, **pvp;
6456 	struct vm_physseg *ps;
6457 	size_t i;
6458 
6459 	KASSERT(amount & PGOFSET);
6460 	KASSERT((mask & PGOFSET) == 0);
6461 	KASSERT((match & PGOFSET) == 0);
6462 	KASSERT(amount != 0);
6463 
6464 	for (pvp = &SLIST_FIRST(&pmap_boot_freeq);
6465 	     (pv = *pvp) != NULL;
6466 	     pvp = &SLIST_NEXT(pv, pv_list)) {
6467 		pv_addr_t *newpv;
6468 		psize_t off;
6469 		/*
6470 		 * If this entry is too small to satify the request...
6471 		 */
6472 		KASSERT(pv->pv_size > 0);
6473 		if (pv->pv_size < amount)
6474 			continue;
6475 
6476 		for (off = 0; off <= mask; off += PAGE_SIZE) {
6477 			if (((pv->pv_pa + off) & mask) == match
6478 			    && off + amount <= pv->pv_size)
6479 				break;
6480 		}
6481 		if (off > mask)
6482 			continue;
6483 
6484 		rpv->pv_va = pv->pv_va + off;
6485 		rpv->pv_pa = pv->pv_pa + off;
6486 		rpv->pv_size = amount;
6487 		pv->pv_size -= amount;
6488 		if (pv->pv_size == 0) {
6489 			KASSERT(off == 0);
6490 			KASSERT((vaddr_t) pv == rpv->pv_va);
6491 			*pvp = SLIST_NEXT(pv, pv_list);
6492 		} else if (off == 0) {
6493 			KASSERT((vaddr_t) pv == rpv->pv_va);
6494 			newpv = (pv_addr_t *) (rpv->pv_va + amount);
6495 			*newpv = *pv;
6496 			newpv->pv_pa += amount;
6497 			newpv->pv_va += amount;
6498 			*pvp = newpv;
6499 		} else if (off < pv->pv_size) {
6500 			newpv = (pv_addr_t *) (rpv->pv_va + amount);
6501 			*newpv = *pv;
6502 			newpv->pv_size -= off;
6503 			newpv->pv_pa += off + amount;
6504 			newpv->pv_va += off + amount;
6505 
6506 			SLIST_NEXT(pv, pv_list) = newpv;
6507 			pv->pv_size = off;
6508 		} else {
6509 			KASSERT((vaddr_t) pv != rpv->pv_va);
6510 		}
6511 		memset((void *)rpv->pv_va, 0, amount);
6512 		return;
6513 	}
6514 
6515 	if (vm_nphysseg == 0)
6516 		panic("pmap_boot_pagealloc: couldn't allocate memory");
6517 
6518 	for (pvp = &SLIST_FIRST(&pmap_boot_freeq);
6519 	     (pv = *pvp) != NULL;
6520 	     pvp = &SLIST_NEXT(pv, pv_list)) {
6521 		if (SLIST_NEXT(pv, pv_list) == NULL)
6522 			break;
6523 	}
6524 	KASSERT(mask == 0);
6525 	for (ps = vm_physmem, i = 0; i < vm_nphysseg; ps++, i++) {
6526 		if (ps->avail_start == atop(pv->pv_pa + pv->pv_size)
6527 		    && pv->pv_va + pv->pv_size <= ptoa(ps->avail_end)) {
6528 			rpv->pv_va = pv->pv_va;
6529 			rpv->pv_pa = pv->pv_pa;
6530 			rpv->pv_size = amount;
6531 			*pvp = NULL;
6532 			pmap_map_chunk(kernel_l1pt.pv_va,
6533 			     ptoa(ps->avail_start) + (pv->pv_va - pv->pv_pa),
6534 			     ptoa(ps->avail_start),
6535 			     amount - pv->pv_size,
6536 			     VM_PROT_READ|VM_PROT_WRITE,
6537 			     PTE_CACHE);
6538 			ps->avail_start += atop(amount - pv->pv_size);
6539 			/*
6540 			 * If we consumed the entire physseg, remove it.
6541 			 */
6542 			if (ps->avail_start == ps->avail_end) {
6543 				for (--vm_nphysseg; i < vm_nphysseg; i++, ps++)
6544 					ps[0] = ps[1];
6545 			}
6546 			memset((void *)rpv->pv_va, 0, rpv->pv_size);
6547 			return;
6548 		}
6549 	}
6550 
6551 	panic("pmap_boot_pagealloc: couldn't allocate memory");
6552 }
6553 
6554 vaddr_t
6555 pmap_steal_memory(vsize_t size, vaddr_t *vstartp, vaddr_t *vendp)
6556 {
6557 	pv_addr_t pv;
6558 
6559 	pmap_boot_pagealloc(size, 0, 0, &pv);
6560 
6561 	return pv.pv_va;
6562 }
6563 #endif /* PMAP_STEAL_MEMORY */
6564 
6565 SYSCTL_SETUP(sysctl_machdep_pmap_setup, "sysctl machdep.kmpages setup")
6566 {
6567 	sysctl_createv(clog, 0, NULL, NULL,
6568 			CTLFLAG_PERMANENT,
6569 			CTLTYPE_NODE, "machdep", NULL,
6570 			NULL, 0, NULL, 0,
6571 			CTL_MACHDEP, CTL_EOL);
6572 
6573 	sysctl_createv(clog, 0, NULL, NULL,
6574 			CTLFLAG_PERMANENT,
6575 			CTLTYPE_INT, "kmpages",
6576 			SYSCTL_DESCR("count of pages allocated to kernel memory allocators"),
6577 			NULL, 0, &pmap_kmpages, 0,
6578 			CTL_MACHDEP, CTL_CREATE, CTL_EOL);
6579 }
6580