xref: /netbsd-src/sys/arch/arm/arm32/pmap.c (revision 6a493d6bc668897c91594964a732d38505b70cbb)
1 /*	$NetBSD: pmap.c,v 1.264 2013/09/12 14:45:18 kiyohara 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 dependent 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 in kernel VA space,
166  *       then 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/pool.h>
201 #include <sys/kmem.h>
202 #include <sys/cdefs.h>
203 #include <sys/cpu.h>
204 #include <sys/sysctl.h>
205 #include <sys/bus.h>
206 
207 #include <uvm/uvm.h>
208 
209 #include <arm/locore.h>
210 #include <arm/arm32/katelib.h>
211 
212 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.264 2013/09/12 14:45:18 kiyohara Exp $");
213 
214 #ifdef PMAP_DEBUG
215 
216 /* XXX need to get rid of all refs to this */
217 int pmap_debug_level = 0;
218 
219 /*
220  * for switching to potentially finer grained debugging
221  */
222 #define	PDB_FOLLOW	0x0001
223 #define	PDB_INIT	0x0002
224 #define	PDB_ENTER	0x0004
225 #define	PDB_REMOVE	0x0008
226 #define	PDB_CREATE	0x0010
227 #define	PDB_PTPAGE	0x0020
228 #define	PDB_GROWKERN	0x0040
229 #define	PDB_BITS	0x0080
230 #define	PDB_COLLECT	0x0100
231 #define	PDB_PROTECT	0x0200
232 #define	PDB_MAP_L1	0x0400
233 #define	PDB_BOOTSTRAP	0x1000
234 #define	PDB_PARANOIA	0x2000
235 #define	PDB_WIRING	0x4000
236 #define	PDB_PVDUMP	0x8000
237 #define	PDB_VAC		0x10000
238 #define	PDB_KENTER	0x20000
239 #define	PDB_KREMOVE	0x40000
240 #define	PDB_EXEC	0x80000
241 
242 int debugmap = 1;
243 int pmapdebug = 0;
244 #define	NPDEBUG(_lev_,_stat_) \
245 	if (pmapdebug & (_lev_)) \
246         	((_stat_))
247 
248 #else	/* PMAP_DEBUG */
249 #define NPDEBUG(_lev_,_stat_) /* Nothing */
250 #endif	/* PMAP_DEBUG */
251 
252 /*
253  * pmap_kernel() points here
254  */
255 static struct pmap	kernel_pmap_store;
256 struct pmap		*const kernel_pmap_ptr = &kernel_pmap_store;
257 #ifdef PMAP_NEED_ALLOC_POOLPAGE
258 int			arm_poolpage_vmfreelist = VM_FREELIST_DEFAULT;
259 #endif
260 
261 /*
262  * Which pmap is currently 'live' in the cache
263  *
264  * XXXSCW: Fix for SMP ...
265  */
266 static pmap_t pmap_recent_user;
267 
268 /*
269  * Pointer to last active lwp, or NULL if it exited.
270  */
271 struct lwp *pmap_previous_active_lwp;
272 
273 /*
274  * Pool and cache that pmap structures are allocated from.
275  * We use a cache to avoid clearing the pm_l2[] array (1KB)
276  * in pmap_create().
277  */
278 static struct pool_cache pmap_cache;
279 static LIST_HEAD(, pmap) pmap_pmaps;
280 
281 /*
282  * Pool of PV structures
283  */
284 static struct pool pmap_pv_pool;
285 static void *pmap_bootstrap_pv_page_alloc(struct pool *, int);
286 static void pmap_bootstrap_pv_page_free(struct pool *, void *);
287 static struct pool_allocator pmap_bootstrap_pv_allocator = {
288 	pmap_bootstrap_pv_page_alloc, pmap_bootstrap_pv_page_free
289 };
290 
291 /*
292  * Pool and cache of l2_dtable structures.
293  * We use a cache to avoid clearing the structures when they're
294  * allocated. (196 bytes)
295  */
296 static struct pool_cache pmap_l2dtable_cache;
297 static vaddr_t pmap_kernel_l2dtable_kva;
298 
299 /*
300  * Pool and cache of L2 page descriptors.
301  * We use a cache to avoid clearing the descriptor table
302  * when they're allocated. (1KB)
303  */
304 static struct pool_cache pmap_l2ptp_cache;
305 static vaddr_t pmap_kernel_l2ptp_kva;
306 static paddr_t pmap_kernel_l2ptp_phys;
307 
308 #ifdef PMAPCOUNTERS
309 #define	PMAP_EVCNT_INITIALIZER(name) \
310 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap", name)
311 
312 #ifdef PMAP_CACHE_VIPT
313 static struct evcnt pmap_ev_vac_clean_one =
314    PMAP_EVCNT_INITIALIZER("clean page (1 color)");
315 static struct evcnt pmap_ev_vac_flush_one =
316    PMAP_EVCNT_INITIALIZER("flush page (1 color)");
317 static struct evcnt pmap_ev_vac_flush_lots =
318    PMAP_EVCNT_INITIALIZER("flush page (2+ colors)");
319 static struct evcnt pmap_ev_vac_flush_lots2 =
320    PMAP_EVCNT_INITIALIZER("flush page (2+ colors, kmpage)");
321 EVCNT_ATTACH_STATIC(pmap_ev_vac_clean_one);
322 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_one);
323 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_lots);
324 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_lots2);
325 
326 static struct evcnt pmap_ev_vac_color_new =
327    PMAP_EVCNT_INITIALIZER("new page color");
328 static struct evcnt pmap_ev_vac_color_reuse =
329    PMAP_EVCNT_INITIALIZER("ok first page color");
330 static struct evcnt pmap_ev_vac_color_ok =
331    PMAP_EVCNT_INITIALIZER("ok page color");
332 static struct evcnt pmap_ev_vac_color_blind =
333    PMAP_EVCNT_INITIALIZER("blind page color");
334 static struct evcnt pmap_ev_vac_color_change =
335    PMAP_EVCNT_INITIALIZER("change page color");
336 static struct evcnt pmap_ev_vac_color_erase =
337    PMAP_EVCNT_INITIALIZER("erase page color");
338 static struct evcnt pmap_ev_vac_color_none =
339    PMAP_EVCNT_INITIALIZER("no page color");
340 static struct evcnt pmap_ev_vac_color_restore =
341    PMAP_EVCNT_INITIALIZER("restore page color");
342 
343 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_new);
344 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_reuse);
345 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_ok);
346 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_blind);
347 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_change);
348 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_erase);
349 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_none);
350 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_restore);
351 #endif
352 
353 static struct evcnt pmap_ev_mappings =
354    PMAP_EVCNT_INITIALIZER("pages mapped");
355 static struct evcnt pmap_ev_unmappings =
356    PMAP_EVCNT_INITIALIZER("pages unmapped");
357 static struct evcnt pmap_ev_remappings =
358    PMAP_EVCNT_INITIALIZER("pages remapped");
359 
360 EVCNT_ATTACH_STATIC(pmap_ev_mappings);
361 EVCNT_ATTACH_STATIC(pmap_ev_unmappings);
362 EVCNT_ATTACH_STATIC(pmap_ev_remappings);
363 
364 static struct evcnt pmap_ev_kernel_mappings =
365    PMAP_EVCNT_INITIALIZER("kernel pages mapped");
366 static struct evcnt pmap_ev_kernel_unmappings =
367    PMAP_EVCNT_INITIALIZER("kernel pages unmapped");
368 static struct evcnt pmap_ev_kernel_remappings =
369    PMAP_EVCNT_INITIALIZER("kernel pages remapped");
370 
371 EVCNT_ATTACH_STATIC(pmap_ev_kernel_mappings);
372 EVCNT_ATTACH_STATIC(pmap_ev_kernel_unmappings);
373 EVCNT_ATTACH_STATIC(pmap_ev_kernel_remappings);
374 
375 static struct evcnt pmap_ev_kenter_mappings =
376    PMAP_EVCNT_INITIALIZER("kenter pages mapped");
377 static struct evcnt pmap_ev_kenter_unmappings =
378    PMAP_EVCNT_INITIALIZER("kenter pages unmapped");
379 static struct evcnt pmap_ev_kenter_remappings =
380    PMAP_EVCNT_INITIALIZER("kenter pages remapped");
381 static struct evcnt pmap_ev_pt_mappings =
382    PMAP_EVCNT_INITIALIZER("page table pages mapped");
383 
384 EVCNT_ATTACH_STATIC(pmap_ev_kenter_mappings);
385 EVCNT_ATTACH_STATIC(pmap_ev_kenter_unmappings);
386 EVCNT_ATTACH_STATIC(pmap_ev_kenter_remappings);
387 EVCNT_ATTACH_STATIC(pmap_ev_pt_mappings);
388 
389 #ifdef PMAP_CACHE_VIPT
390 static struct evcnt pmap_ev_exec_mappings =
391    PMAP_EVCNT_INITIALIZER("exec pages mapped");
392 static struct evcnt pmap_ev_exec_cached =
393    PMAP_EVCNT_INITIALIZER("exec pages cached");
394 
395 EVCNT_ATTACH_STATIC(pmap_ev_exec_mappings);
396 EVCNT_ATTACH_STATIC(pmap_ev_exec_cached);
397 
398 static struct evcnt pmap_ev_exec_synced =
399    PMAP_EVCNT_INITIALIZER("exec pages synced");
400 static struct evcnt pmap_ev_exec_synced_map =
401    PMAP_EVCNT_INITIALIZER("exec pages synced (MP)");
402 static struct evcnt pmap_ev_exec_synced_unmap =
403    PMAP_EVCNT_INITIALIZER("exec pages synced (UM)");
404 static struct evcnt pmap_ev_exec_synced_remap =
405    PMAP_EVCNT_INITIALIZER("exec pages synced (RM)");
406 static struct evcnt pmap_ev_exec_synced_clearbit =
407    PMAP_EVCNT_INITIALIZER("exec pages synced (DG)");
408 static struct evcnt pmap_ev_exec_synced_kremove =
409    PMAP_EVCNT_INITIALIZER("exec pages synced (KU)");
410 
411 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced);
412 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_map);
413 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_unmap);
414 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_remap);
415 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_clearbit);
416 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_kremove);
417 
418 static struct evcnt pmap_ev_exec_discarded_unmap =
419    PMAP_EVCNT_INITIALIZER("exec pages discarded (UM)");
420 static struct evcnt pmap_ev_exec_discarded_zero =
421    PMAP_EVCNT_INITIALIZER("exec pages discarded (ZP)");
422 static struct evcnt pmap_ev_exec_discarded_copy =
423    PMAP_EVCNT_INITIALIZER("exec pages discarded (CP)");
424 static struct evcnt pmap_ev_exec_discarded_page_protect =
425    PMAP_EVCNT_INITIALIZER("exec pages discarded (PP)");
426 static struct evcnt pmap_ev_exec_discarded_clearbit =
427    PMAP_EVCNT_INITIALIZER("exec pages discarded (DG)");
428 static struct evcnt pmap_ev_exec_discarded_kremove =
429    PMAP_EVCNT_INITIALIZER("exec pages discarded (KU)");
430 
431 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_unmap);
432 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_zero);
433 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_copy);
434 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_page_protect);
435 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_clearbit);
436 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_kremove);
437 #endif /* PMAP_CACHE_VIPT */
438 
439 static struct evcnt pmap_ev_updates = PMAP_EVCNT_INITIALIZER("updates");
440 static struct evcnt pmap_ev_collects = PMAP_EVCNT_INITIALIZER("collects");
441 static struct evcnt pmap_ev_activations = PMAP_EVCNT_INITIALIZER("activations");
442 
443 EVCNT_ATTACH_STATIC(pmap_ev_updates);
444 EVCNT_ATTACH_STATIC(pmap_ev_collects);
445 EVCNT_ATTACH_STATIC(pmap_ev_activations);
446 
447 #define	PMAPCOUNT(x)	((void)(pmap_ev_##x.ev_count++))
448 #else
449 #define	PMAPCOUNT(x)	((void)0)
450 #endif
451 
452 /*
453  * pmap copy/zero page, and mem(5) hook point
454  */
455 static pt_entry_t *csrc_pte, *cdst_pte;
456 static vaddr_t csrcp, cdstp;
457 vaddr_t memhook;			/* used by mem.c */
458 kmutex_t memlock;			/* used by mem.c */
459 void *zeropage;				/* used by mem.c */
460 extern void *msgbufaddr;
461 int pmap_kmpages;
462 /*
463  * Flag to indicate if pmap_init() has done its thing
464  */
465 bool pmap_initialized;
466 
467 /*
468  * Misc. locking data structures
469  */
470 
471 #define	pmap_acquire_pmap_lock(pm)			\
472 	do {						\
473 		if ((pm) != pmap_kernel())		\
474 			mutex_enter((pm)->pm_lock);	\
475 	} while (/*CONSTCOND*/0)
476 
477 #define	pmap_release_pmap_lock(pm)			\
478 	do {						\
479 		if ((pm) != pmap_kernel())		\
480 			mutex_exit((pm)->pm_lock);	\
481 	} while (/*CONSTCOND*/0)
482 
483 
484 /*
485  * Metadata for L1 translation tables.
486  */
487 struct l1_ttable {
488 	/* Entry on the L1 Table list */
489 	SLIST_ENTRY(l1_ttable) l1_link;
490 
491 	/* Entry on the L1 Least Recently Used list */
492 	TAILQ_ENTRY(l1_ttable) l1_lru;
493 
494 	/* Track how many domains are allocated from this L1 */
495 	volatile u_int l1_domain_use_count;
496 
497 	/*
498 	 * A free-list of domain numbers for this L1.
499 	 * We avoid using ffs() and a bitmap to track domains since ffs()
500 	 * is slow on ARM.
501 	 */
502 	uint8_t l1_domain_first;
503 	uint8_t l1_domain_free[PMAP_DOMAINS];
504 
505 	/* Physical address of this L1 page table */
506 	paddr_t l1_physaddr;
507 
508 	/* KVA of this L1 page table */
509 	pd_entry_t *l1_kva;
510 };
511 
512 /*
513  * Convert a virtual address into its L1 table index. That is, the
514  * index used to locate the L2 descriptor table pointer in an L1 table.
515  * This is basically used to index l1->l1_kva[].
516  *
517  * Each L2 descriptor table represents 1MB of VA space.
518  */
519 #define	L1_IDX(va)		(((vaddr_t)(va)) >> L1_S_SHIFT)
520 
521 /*
522  * L1 Page Tables are tracked using a Least Recently Used list.
523  *  - New L1s are allocated from the HEAD.
524  *  - Freed L1s are added to the TAIl.
525  *  - Recently accessed L1s (where an 'access' is some change to one of
526  *    the userland pmaps which owns this L1) are moved to the TAIL.
527  */
528 static TAILQ_HEAD(, l1_ttable) l1_lru_list;
529 static kmutex_t l1_lru_lock __cacheline_aligned;
530 
531 /*
532  * A list of all L1 tables
533  */
534 static SLIST_HEAD(, l1_ttable) l1_list;
535 
536 /*
537  * The l2_dtable tracks L2_BUCKET_SIZE worth of L1 slots.
538  *
539  * This is normally 16MB worth L2 page descriptors for any given pmap.
540  * Reference counts are maintained for L2 descriptors so they can be
541  * freed when empty.
542  */
543 struct l2_dtable {
544 	/* The number of L2 page descriptors allocated to this l2_dtable */
545 	u_int l2_occupancy;
546 
547 	/* List of L2 page descriptors */
548 	struct l2_bucket {
549 		pt_entry_t *l2b_kva;	/* KVA of L2 Descriptor Table */
550 		paddr_t l2b_phys;	/* Physical address of same */
551 		u_short l2b_l1idx;	/* This L2 table's L1 index */
552 		u_short l2b_occupancy;	/* How many active descriptors */
553 	} l2_bucket[L2_BUCKET_SIZE];
554 };
555 
556 /*
557  * Given an L1 table index, calculate the corresponding l2_dtable index
558  * and bucket index within the l2_dtable.
559  */
560 #define	L2_IDX(l1idx)		(((l1idx) >> L2_BUCKET_LOG2) & \
561 				 (L2_SIZE - 1))
562 #define	L2_BUCKET(l1idx)	((l1idx) & (L2_BUCKET_SIZE - 1))
563 
564 /*
565  * Given a virtual address, this macro returns the
566  * virtual address required to drop into the next L2 bucket.
567  */
568 #define	L2_NEXT_BUCKET(va)	(((va) & L1_S_FRAME) + L1_S_SIZE)
569 
570 /*
571  * L2 allocation.
572  */
573 #define	pmap_alloc_l2_dtable()		\
574 	    pool_cache_get(&pmap_l2dtable_cache, PR_NOWAIT)
575 #define	pmap_free_l2_dtable(l2)		\
576 	    pool_cache_put(&pmap_l2dtable_cache, (l2))
577 #define pmap_alloc_l2_ptp(pap)		\
578 	    ((pt_entry_t *)pool_cache_get_paddr(&pmap_l2ptp_cache,\
579 	    PR_NOWAIT, (pap)))
580 
581 /*
582  * We try to map the page tables write-through, if possible.  However, not
583  * all CPUs have a write-through cache mode, so on those we have to sync
584  * the cache when we frob page tables.
585  *
586  * We try to evaluate this at compile time, if possible.  However, it's
587  * not always possible to do that, hence this run-time var.
588  */
589 int	pmap_needs_pte_sync;
590 
591 /*
592  * Real definition of pv_entry.
593  */
594 struct pv_entry {
595 	SLIST_ENTRY(pv_entry) pv_link;	/* next pv_entry */
596 	pmap_t		pv_pmap;        /* pmap where mapping lies */
597 	vaddr_t		pv_va;          /* virtual address for mapping */
598 	u_int		pv_flags;       /* flags */
599 };
600 
601 /*
602  * Macro to determine if a mapping might be resident in the
603  * instruction cache and/or TLB
604  */
605 #if ARM_MMU_V7 > 0
606 /*
607  * Speculative loads by Cortex cores can cause TLB entries to be filled even if
608  * there are no explicit accesses, so there may be always be TLB entries to
609  * flush.  If we used ASIDs then this would not be a problem.
610  */
611 #define	PV_BEEN_EXECD(f)  (((f) & PVF_EXEC) == PVF_EXEC)
612 #else
613 #define	PV_BEEN_EXECD(f)  (((f) & (PVF_REF | PVF_EXEC)) == (PVF_REF | PVF_EXEC))
614 #endif
615 #define	PV_IS_EXEC_P(f)   (((f) & PVF_EXEC) != 0)
616 
617 /*
618  * Macro to determine if a mapping might be resident in the
619  * data cache and/or TLB
620  */
621 #if ARM_MMU_V7 > 0
622 /*
623  * Speculative loads by Cortex cores can cause TLB entries to be filled even if
624  * there are no explicit accesses, so there may be always be TLB entries to
625  * flush.  If we used ASIDs then this would not be a problem.
626  */
627 #define	PV_BEEN_REFD(f)   (1)
628 #else
629 #define	PV_BEEN_REFD(f)   (((f) & PVF_REF) != 0)
630 #endif
631 
632 /*
633  * Local prototypes
634  */
635 static int		pmap_set_pt_cache_mode(pd_entry_t *, vaddr_t);
636 static void		pmap_alloc_specials(vaddr_t *, int, vaddr_t *,
637 			    pt_entry_t **);
638 static bool		pmap_is_current(pmap_t);
639 static bool		pmap_is_cached(pmap_t);
640 static void		pmap_enter_pv(struct vm_page_md *, paddr_t, struct pv_entry *,
641 			    pmap_t, vaddr_t, u_int);
642 static struct pv_entry *pmap_find_pv(struct vm_page_md *, pmap_t, vaddr_t);
643 static struct pv_entry *pmap_remove_pv(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
644 static u_int		pmap_modify_pv(struct vm_page_md *, paddr_t, pmap_t, vaddr_t,
645 			    u_int, u_int);
646 
647 static void		pmap_pinit(pmap_t);
648 static int		pmap_pmap_ctor(void *, void *, int);
649 
650 static void		pmap_alloc_l1(pmap_t);
651 static void		pmap_free_l1(pmap_t);
652 static void		pmap_use_l1(pmap_t);
653 
654 static struct l2_bucket *pmap_get_l2_bucket(pmap_t, vaddr_t);
655 static struct l2_bucket *pmap_alloc_l2_bucket(pmap_t, vaddr_t);
656 static void		pmap_free_l2_bucket(pmap_t, struct l2_bucket *, u_int);
657 static int		pmap_l2ptp_ctor(void *, void *, int);
658 static int		pmap_l2dtable_ctor(void *, void *, int);
659 
660 static void		pmap_vac_me_harder(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
661 #ifdef PMAP_CACHE_VIVT
662 static void		pmap_vac_me_kpmap(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
663 static void		pmap_vac_me_user(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
664 #endif
665 
666 static void		pmap_clearbit(struct vm_page_md *, paddr_t, u_int);
667 #ifdef PMAP_CACHE_VIVT
668 static int		pmap_clean_page(struct pv_entry *, bool);
669 #endif
670 #ifdef PMAP_CACHE_VIPT
671 static void		pmap_syncicache_page(struct vm_page_md *, paddr_t);
672 enum pmap_flush_op {
673 	PMAP_FLUSH_PRIMARY,
674 	PMAP_FLUSH_SECONDARY,
675 	PMAP_CLEAN_PRIMARY
676 };
677 static void		pmap_flush_page(struct vm_page_md *, paddr_t, enum pmap_flush_op);
678 #endif
679 static void		pmap_page_remove(struct vm_page_md *, paddr_t);
680 
681 static void		pmap_init_l1(struct l1_ttable *, pd_entry_t *);
682 static vaddr_t		kernel_pt_lookup(paddr_t);
683 
684 
685 /*
686  * Misc variables
687  */
688 vaddr_t virtual_avail;
689 vaddr_t virtual_end;
690 vaddr_t pmap_curmaxkvaddr;
691 
692 paddr_t avail_start;
693 paddr_t avail_end;
694 
695 pv_addrqh_t pmap_boot_freeq = SLIST_HEAD_INITIALIZER(&pmap_boot_freeq);
696 pv_addr_t kernelpages;
697 pv_addr_t kernel_l1pt;
698 pv_addr_t systempage;
699 
700 /* Function to set the debug level of the pmap code */
701 
702 #ifdef PMAP_DEBUG
703 void
704 pmap_debug(int level)
705 {
706 	pmap_debug_level = level;
707 	printf("pmap_debug: level=%d\n", pmap_debug_level);
708 }
709 #endif	/* PMAP_DEBUG */
710 
711 #ifdef PMAP_CACHE_VIPT
712 #define PMAP_VALIDATE_MD_PAGE(md)	\
713 	KASSERTMSG(arm_cache_prefer_mask == 0 || (((md)->pvh_attrs & PVF_WRITE) == 0) == ((md)->urw_mappings + (md)->krw_mappings == 0), \
714 	    "(md) %p: attrs=%#x urw=%u krw=%u", (md), \
715 	    (md)->pvh_attrs, (md)->urw_mappings, (md)->krw_mappings);
716 #endif /* PMAP_CACHE_VIPT */
717 /*
718  * A bunch of routines to conditionally flush the caches/TLB depending
719  * on whether the specified pmap actually needs to be flushed at any
720  * given time.
721  */
722 static inline void
723 pmap_tlb_flush_SE(pmap_t pm, vaddr_t va, u_int flags)
724 {
725 	if (pm->pm_cstate.cs_tlb_id != 0) {
726 		if (PV_BEEN_EXECD(flags)) {
727 			cpu_tlb_flushID_SE(va);
728 		} else if (PV_BEEN_REFD(flags)) {
729 			cpu_tlb_flushD_SE(va);
730 		}
731 	}
732 }
733 
734 static inline void
735 pmap_tlb_flushID(pmap_t pm)
736 {
737 	if (pm->pm_cstate.cs_tlb_id) {
738 		cpu_tlb_flushID();
739 #if ARM_MMU_V7 == 0
740 		/*
741 		 * Speculative loads by Cortex cores can cause TLB entries to
742 		 * be filled even if there are no explicit accesses, so there
743 		 * may be always be TLB entries to flush.  If we used ASIDs
744 		 * then it would not be a problem.
745 		 * This is not true for other CPUs.
746 		 */
747 		pm->pm_cstate.cs_tlb = 0;
748 #endif /* ARM_MMU_V7 */
749 	}
750 }
751 
752 static inline void
753 pmap_tlb_flushD(pmap_t pm)
754 {
755 	if (pm->pm_cstate.cs_tlb_d) {
756 		cpu_tlb_flushD();
757 #if ARM_MMU_V7 == 0
758 		/*
759 		 * Speculative loads by Cortex cores can cause TLB entries to
760 		 * be filled even if there are no explicit accesses, so there
761 		 * may be always be TLB entries to flush.  If we used ASIDs
762 		 * then it would not be a problem.
763 		 * This is not true for other CPUs.
764 		 */
765 		pm->pm_cstate.cs_tlb_d = 0;
766 #endif /* ARM_MMU_V7 */
767 	}
768 }
769 
770 #ifdef PMAP_CACHE_VIVT
771 static inline void
772 pmap_cache_wbinv_page(pmap_t pm, vaddr_t va, bool do_inv, u_int flags)
773 {
774 	if (PV_BEEN_EXECD(flags) && pm->pm_cstate.cs_cache_id) {
775 		cpu_idcache_wbinv_range(va, PAGE_SIZE);
776 	} else if (PV_BEEN_REFD(flags) && pm->pm_cstate.cs_cache_d) {
777 		if (do_inv) {
778 			if (flags & PVF_WRITE)
779 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
780 			else
781 				cpu_dcache_inv_range(va, PAGE_SIZE);
782 		} else if (flags & PVF_WRITE) {
783 			cpu_dcache_wb_range(va, PAGE_SIZE);
784 		}
785 	}
786 }
787 
788 static inline void
789 pmap_cache_wbinv_all(pmap_t pm, u_int flags)
790 {
791 	if (PV_BEEN_EXECD(flags)) {
792 		if (pm->pm_cstate.cs_cache_id) {
793 			cpu_idcache_wbinv_all();
794 			pm->pm_cstate.cs_cache = 0;
795 		}
796 	} else if (pm->pm_cstate.cs_cache_d) {
797 		cpu_dcache_wbinv_all();
798 		pm->pm_cstate.cs_cache_d = 0;
799 	}
800 }
801 #endif /* PMAP_CACHE_VIVT */
802 
803 static inline uint8_t
804 pmap_domain(pmap_t pm)
805 {
806 	return pm->pm_domain;
807 }
808 
809 static inline pd_entry_t *
810 pmap_l1_kva(pmap_t pm)
811 {
812 	return pm->pm_l1->l1_kva;
813 }
814 
815 static inline bool
816 pmap_is_current(pmap_t pm)
817 {
818 
819 	if (pm == pmap_kernel() || curproc->p_vmspace->vm_map.pmap == pm)
820 		return true;
821 
822 	return false;
823 }
824 
825 static inline bool
826 pmap_is_cached(pmap_t pm)
827 {
828 
829 	if (pm == pmap_kernel() || pmap_recent_user == NULL ||
830 	    pmap_recent_user == pm)
831 		return (true);
832 
833 	return false;
834 }
835 
836 /*
837  * PTE_SYNC_CURRENT:
838  *
839  *     Make sure the pte is written out to RAM.
840  *     We need to do this for one of two cases:
841  *       - We're dealing with the kernel pmap
842  *       - There is no pmap active in the cache/tlb.
843  *       - The specified pmap is 'active' in the cache/tlb.
844  */
845 #ifdef PMAP_INCLUDE_PTE_SYNC
846 #define	PTE_SYNC_CURRENT(pm, ptep)	\
847 do {					\
848 	if (PMAP_NEEDS_PTE_SYNC && 	\
849 	    pmap_is_cached(pm))		\
850 		PTE_SYNC(ptep);		\
851 } while (/*CONSTCOND*/0)
852 #else
853 #define	PTE_SYNC_CURRENT(pm, ptep)	/* nothing */
854 #endif
855 
856 /*
857  * main pv_entry manipulation functions:
858  *   pmap_enter_pv: enter a mapping onto a vm_page list
859  *   pmap_remove_pv: remove a mapping from a vm_page list
860  *
861  * NOTE: pmap_enter_pv expects to lock the pvh itself
862  *       pmap_remove_pv expects the caller to lock the pvh before calling
863  */
864 
865 /*
866  * pmap_enter_pv: enter a mapping onto a vm_page lst
867  *
868  * => caller should hold the proper lock on pmap_main_lock
869  * => caller should have pmap locked
870  * => we will gain the lock on the vm_page and allocate the new pv_entry
871  * => caller should adjust ptp's wire_count before calling
872  * => caller should not adjust pmap's wire_count
873  */
874 static void
875 pmap_enter_pv(struct vm_page_md *md, paddr_t pa, struct pv_entry *pv, pmap_t pm,
876     vaddr_t va, u_int flags)
877 {
878 	struct pv_entry **pvp;
879 
880 	NPDEBUG(PDB_PVDUMP,
881 	    printf("pmap_enter_pv: pm %p, md %p, flags 0x%x\n", pm, md, flags));
882 
883 	pv->pv_pmap = pm;
884 	pv->pv_va = va;
885 	pv->pv_flags = flags;
886 
887 	pvp = &SLIST_FIRST(&md->pvh_list);
888 #ifdef PMAP_CACHE_VIPT
889 	/*
890 	 * Insert unmanaged entries, writeable first, at the head of
891 	 * the pv list.
892 	 */
893 	if (__predict_true((flags & PVF_KENTRY) == 0)) {
894 		while (*pvp != NULL && (*pvp)->pv_flags & PVF_KENTRY)
895 			pvp = &SLIST_NEXT(*pvp, pv_link);
896 	} else if ((flags & PVF_WRITE) == 0) {
897 		while (*pvp != NULL && (*pvp)->pv_flags & PVF_WRITE)
898 			pvp = &SLIST_NEXT(*pvp, pv_link);
899 	}
900 #endif
901 	SLIST_NEXT(pv, pv_link) = *pvp;		/* add to ... */
902 	*pvp = pv;				/* ... locked list */
903 	md->pvh_attrs |= flags & (PVF_REF | PVF_MOD);
904 #ifdef PMAP_CACHE_VIPT
905 	if ((pv->pv_flags & PVF_KWRITE) == PVF_KWRITE)
906 		md->pvh_attrs |= PVF_KMOD;
907 	if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) != PVF_NC)
908 		md->pvh_attrs |= PVF_DIRTY;
909 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
910 #endif
911 	if (pm == pmap_kernel()) {
912 		PMAPCOUNT(kernel_mappings);
913 		if (flags & PVF_WRITE)
914 			md->krw_mappings++;
915 		else
916 			md->kro_mappings++;
917 	} else {
918 		if (flags & PVF_WRITE)
919 			md->urw_mappings++;
920 		else
921 			md->uro_mappings++;
922 	}
923 
924 #ifdef PMAP_CACHE_VIPT
925 	/*
926 	 * Even though pmap_vac_me_harder will set PVF_WRITE for us,
927 	 * do it here as well to keep the mappings & KVF_WRITE consistent.
928 	 */
929 	if (arm_cache_prefer_mask != 0 && (flags & PVF_WRITE) != 0) {
930 		md->pvh_attrs |= PVF_WRITE;
931 	}
932 	/*
933 	 * If this is an exec mapping and its the first exec mapping
934 	 * for this page, make sure to sync the I-cache.
935 	 */
936 	if (PV_IS_EXEC_P(flags)) {
937 		if (!PV_IS_EXEC_P(md->pvh_attrs)) {
938 			pmap_syncicache_page(md, pa);
939 			PMAPCOUNT(exec_synced_map);
940 		}
941 		PMAPCOUNT(exec_mappings);
942 	}
943 #endif
944 
945 	PMAPCOUNT(mappings);
946 
947 	if (pv->pv_flags & PVF_WIRED)
948 		++pm->pm_stats.wired_count;
949 }
950 
951 /*
952  *
953  * pmap_find_pv: Find a pv entry
954  *
955  * => caller should hold lock on vm_page
956  */
957 static inline struct pv_entry *
958 pmap_find_pv(struct vm_page_md *md, pmap_t pm, vaddr_t va)
959 {
960 	struct pv_entry *pv;
961 
962 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
963 		if (pm == pv->pv_pmap && va == pv->pv_va)
964 			break;
965 	}
966 
967 	return (pv);
968 }
969 
970 /*
971  * pmap_remove_pv: try to remove a mapping from a pv_list
972  *
973  * => caller should hold proper lock on pmap_main_lock
974  * => pmap should be locked
975  * => caller should hold lock on vm_page [so that attrs can be adjusted]
976  * => caller should adjust ptp's wire_count and free PTP if needed
977  * => caller should NOT adjust pmap's wire_count
978  * => we return the removed pv
979  */
980 static struct pv_entry *
981 pmap_remove_pv(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
982 {
983 	struct pv_entry *pv, **prevptr;
984 
985 	NPDEBUG(PDB_PVDUMP,
986 	    printf("pmap_remove_pv: pm %p, md %p, va 0x%08lx\n", pm, md, va));
987 
988 	prevptr = &SLIST_FIRST(&md->pvh_list); /* prev pv_entry ptr */
989 	pv = *prevptr;
990 
991 	while (pv) {
992 		if (pv->pv_pmap == pm && pv->pv_va == va) {	/* match? */
993 			NPDEBUG(PDB_PVDUMP, printf("pmap_remove_pv: pm %p, md "
994 			    "%p, flags 0x%x\n", pm, md, pv->pv_flags));
995 			if (pv->pv_flags & PVF_WIRED) {
996 				--pm->pm_stats.wired_count;
997 			}
998 			*prevptr = SLIST_NEXT(pv, pv_link);	/* remove it! */
999 			if (pm == pmap_kernel()) {
1000 				PMAPCOUNT(kernel_unmappings);
1001 				if (pv->pv_flags & PVF_WRITE)
1002 					md->krw_mappings--;
1003 				else
1004 					md->kro_mappings--;
1005 			} else {
1006 				if (pv->pv_flags & PVF_WRITE)
1007 					md->urw_mappings--;
1008 				else
1009 					md->uro_mappings--;
1010 			}
1011 
1012 			PMAPCOUNT(unmappings);
1013 #ifdef PMAP_CACHE_VIPT
1014 			if (!(pv->pv_flags & PVF_WRITE))
1015 				break;
1016 			/*
1017 			 * If this page has had an exec mapping, then if
1018 			 * this was the last mapping, discard the contents,
1019 			 * otherwise sync the i-cache for this page.
1020 			 */
1021 			if (PV_IS_EXEC_P(md->pvh_attrs)) {
1022 				if (SLIST_EMPTY(&md->pvh_list)) {
1023 					md->pvh_attrs &= ~PVF_EXEC;
1024 					PMAPCOUNT(exec_discarded_unmap);
1025 				} else {
1026 					pmap_syncicache_page(md, pa);
1027 					PMAPCOUNT(exec_synced_unmap);
1028 				}
1029 			}
1030 #endif /* PMAP_CACHE_VIPT */
1031 			break;
1032 		}
1033 		prevptr = &SLIST_NEXT(pv, pv_link);	/* previous pointer */
1034 		pv = *prevptr;				/* advance */
1035 	}
1036 
1037 #ifdef PMAP_CACHE_VIPT
1038 	/*
1039 	 * If we no longer have a WRITEABLE KENTRY at the head of list,
1040 	 * clear the KMOD attribute from the page.
1041 	 */
1042 	if (SLIST_FIRST(&md->pvh_list) == NULL
1043 	    || (SLIST_FIRST(&md->pvh_list)->pv_flags & PVF_KWRITE) != PVF_KWRITE)
1044 		md->pvh_attrs &= ~PVF_KMOD;
1045 
1046 	/*
1047 	 * If this was a writeable page and there are no more writeable
1048 	 * mappings (ignoring KMPAGE), clear the WRITE flag and writeback
1049 	 * the contents to memory.
1050 	 */
1051 	if (arm_cache_prefer_mask != 0) {
1052 		if (md->krw_mappings + md->urw_mappings == 0)
1053 			md->pvh_attrs &= ~PVF_WRITE;
1054 		PMAP_VALIDATE_MD_PAGE(md);
1055 	}
1056 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1057 #endif /* PMAP_CACHE_VIPT */
1058 
1059 	return(pv);				/* return removed pv */
1060 }
1061 
1062 /*
1063  *
1064  * pmap_modify_pv: Update pv flags
1065  *
1066  * => caller should hold lock on vm_page [so that attrs can be adjusted]
1067  * => caller should NOT adjust pmap's wire_count
1068  * => caller must call pmap_vac_me_harder() if writable status of a page
1069  *    may have changed.
1070  * => we return the old flags
1071  *
1072  * Modify a physical-virtual mapping in the pv table
1073  */
1074 static u_int
1075 pmap_modify_pv(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va,
1076     u_int clr_mask, u_int set_mask)
1077 {
1078 	struct pv_entry *npv;
1079 	u_int flags, oflags;
1080 
1081 	KASSERT((clr_mask & PVF_KENTRY) == 0);
1082 	KASSERT((set_mask & PVF_KENTRY) == 0);
1083 
1084 	if ((npv = pmap_find_pv(md, pm, va)) == NULL)
1085 		return (0);
1086 
1087 	NPDEBUG(PDB_PVDUMP,
1088 	    printf("pmap_modify_pv: pm %p, md %p, clr 0x%x, set 0x%x, flags 0x%x\n", pm, md, clr_mask, set_mask, npv->pv_flags));
1089 
1090 	/*
1091 	 * There is at least one VA mapping this page.
1092 	 */
1093 
1094 	if (clr_mask & (PVF_REF | PVF_MOD)) {
1095 		md->pvh_attrs |= set_mask & (PVF_REF | PVF_MOD);
1096 #ifdef PMAP_CACHE_VIPT
1097 		if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) != PVF_NC)
1098 			md->pvh_attrs |= PVF_DIRTY;
1099 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1100 #endif
1101 	}
1102 
1103 	oflags = npv->pv_flags;
1104 	npv->pv_flags = flags = (oflags & ~clr_mask) | set_mask;
1105 
1106 	if ((flags ^ oflags) & PVF_WIRED) {
1107 		if (flags & PVF_WIRED)
1108 			++pm->pm_stats.wired_count;
1109 		else
1110 			--pm->pm_stats.wired_count;
1111 	}
1112 
1113 	if ((flags ^ oflags) & PVF_WRITE) {
1114 		if (pm == pmap_kernel()) {
1115 			if (flags & PVF_WRITE) {
1116 				md->krw_mappings++;
1117 				md->kro_mappings--;
1118 			} else {
1119 				md->kro_mappings++;
1120 				md->krw_mappings--;
1121 			}
1122 		} else {
1123 			if (flags & PVF_WRITE) {
1124 				md->urw_mappings++;
1125 				md->uro_mappings--;
1126 			} else {
1127 				md->uro_mappings++;
1128 				md->urw_mappings--;
1129 			}
1130 		}
1131 	}
1132 #ifdef PMAP_CACHE_VIPT
1133 	if (arm_cache_prefer_mask != 0) {
1134 		if (md->urw_mappings + md->krw_mappings == 0) {
1135 			md->pvh_attrs &= ~PVF_WRITE;
1136 		} else {
1137 			md->pvh_attrs |= PVF_WRITE;
1138 		}
1139 	}
1140 	/*
1141 	 * We have two cases here: the first is from enter_pv (new exec
1142 	 * page), the second is a combined pmap_remove_pv/pmap_enter_pv.
1143 	 * Since in latter, pmap_enter_pv won't do anything, we just have
1144 	 * to do what pmap_remove_pv would do.
1145 	 */
1146 	if ((PV_IS_EXEC_P(flags) && !PV_IS_EXEC_P(md->pvh_attrs))
1147 	    || (PV_IS_EXEC_P(md->pvh_attrs)
1148 		|| (!(flags & PVF_WRITE) && (oflags & PVF_WRITE)))) {
1149 		pmap_syncicache_page(md, pa);
1150 		PMAPCOUNT(exec_synced_remap);
1151 	}
1152 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1153 #endif
1154 
1155 	PMAPCOUNT(remappings);
1156 
1157 	return (oflags);
1158 }
1159 
1160 /*
1161  * Allocate an L1 translation table for the specified pmap.
1162  * This is called at pmap creation time.
1163  */
1164 static void
1165 pmap_alloc_l1(pmap_t pm)
1166 {
1167 	struct l1_ttable *l1;
1168 	uint8_t domain;
1169 
1170 	/*
1171 	 * Remove the L1 at the head of the LRU list
1172 	 */
1173 	mutex_spin_enter(&l1_lru_lock);
1174 	l1 = TAILQ_FIRST(&l1_lru_list);
1175 	KDASSERT(l1 != NULL);
1176 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1177 
1178 	/*
1179 	 * Pick the first available domain number, and update
1180 	 * the link to the next number.
1181 	 */
1182 	domain = l1->l1_domain_first;
1183 	l1->l1_domain_first = l1->l1_domain_free[domain];
1184 
1185 	/*
1186 	 * If there are still free domain numbers in this L1,
1187 	 * put it back on the TAIL of the LRU list.
1188 	 */
1189 	if (++l1->l1_domain_use_count < PMAP_DOMAINS)
1190 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1191 
1192 	mutex_spin_exit(&l1_lru_lock);
1193 
1194 	/*
1195 	 * Fix up the relevant bits in the pmap structure
1196 	 */
1197 	pm->pm_l1 = l1;
1198 	pm->pm_domain = domain + 1;
1199 }
1200 
1201 /*
1202  * Free an L1 translation table.
1203  * This is called at pmap destruction time.
1204  */
1205 static void
1206 pmap_free_l1(pmap_t pm)
1207 {
1208 	struct l1_ttable *l1 = pm->pm_l1;
1209 
1210 	mutex_spin_enter(&l1_lru_lock);
1211 
1212 	/*
1213 	 * If this L1 is currently on the LRU list, remove it.
1214 	 */
1215 	if (l1->l1_domain_use_count < PMAP_DOMAINS)
1216 		TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1217 
1218 	/*
1219 	 * Free up the domain number which was allocated to the pmap
1220 	 */
1221 	l1->l1_domain_free[pmap_domain(pm) - 1] = l1->l1_domain_first;
1222 	l1->l1_domain_first = pmap_domain(pm) - 1;
1223 	l1->l1_domain_use_count--;
1224 
1225 	/*
1226 	 * The L1 now must have at least 1 free domain, so add
1227 	 * it back to the LRU list. If the use count is zero,
1228 	 * put it at the head of the list, otherwise it goes
1229 	 * to the tail.
1230 	 */
1231 	if (l1->l1_domain_use_count == 0)
1232 		TAILQ_INSERT_HEAD(&l1_lru_list, l1, l1_lru);
1233 	else
1234 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1235 
1236 	mutex_spin_exit(&l1_lru_lock);
1237 }
1238 
1239 static inline void
1240 pmap_use_l1(pmap_t pm)
1241 {
1242 	struct l1_ttable *l1;
1243 
1244 	/*
1245 	 * Do nothing if we're in interrupt context.
1246 	 * Access to an L1 by the kernel pmap must not affect
1247 	 * the LRU list.
1248 	 */
1249 	if (cpu_intr_p() || pm == pmap_kernel())
1250 		return;
1251 
1252 	l1 = pm->pm_l1;
1253 
1254 	/*
1255 	 * If the L1 is not currently on the LRU list, just return
1256 	 */
1257 	if (l1->l1_domain_use_count == PMAP_DOMAINS)
1258 		return;
1259 
1260 	mutex_spin_enter(&l1_lru_lock);
1261 
1262 	/*
1263 	 * Check the use count again, now that we've acquired the lock
1264 	 */
1265 	if (l1->l1_domain_use_count == PMAP_DOMAINS) {
1266 		mutex_spin_exit(&l1_lru_lock);
1267 		return;
1268 	}
1269 
1270 	/*
1271 	 * Move the L1 to the back of the LRU list
1272 	 */
1273 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1274 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1275 
1276 	mutex_spin_exit(&l1_lru_lock);
1277 }
1278 
1279 /*
1280  * void pmap_free_l2_ptp(pt_entry_t *, paddr_t *)
1281  *
1282  * Free an L2 descriptor table.
1283  */
1284 static inline void
1285 #ifndef PMAP_INCLUDE_PTE_SYNC
1286 pmap_free_l2_ptp(pt_entry_t *l2, paddr_t pa)
1287 #else
1288 pmap_free_l2_ptp(bool need_sync, pt_entry_t *l2, paddr_t pa)
1289 #endif
1290 {
1291 #ifdef PMAP_INCLUDE_PTE_SYNC
1292 #ifdef PMAP_CACHE_VIVT
1293 	/*
1294 	 * Note: With a write-back cache, we may need to sync this
1295 	 * L2 table before re-using it.
1296 	 * This is because it may have belonged to a non-current
1297 	 * pmap, in which case the cache syncs would have been
1298 	 * skipped for the pages that were being unmapped. If the
1299 	 * L2 table were then to be immediately re-allocated to
1300 	 * the *current* pmap, it may well contain stale mappings
1301 	 * which have not yet been cleared by a cache write-back
1302 	 * and so would still be visible to the mmu.
1303 	 */
1304 	if (need_sync)
1305 		PTE_SYNC_RANGE(l2, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1306 #endif /* PMAP_CACHE_VIVT */
1307 #endif /* PMAP_INCLUDE_PTE_SYNC */
1308 	pool_cache_put_paddr(&pmap_l2ptp_cache, (void *)l2, pa);
1309 }
1310 
1311 /*
1312  * Returns a pointer to the L2 bucket associated with the specified pmap
1313  * and VA, or NULL if no L2 bucket exists for the address.
1314  */
1315 static inline struct l2_bucket *
1316 pmap_get_l2_bucket(pmap_t pm, vaddr_t va)
1317 {
1318 	struct l2_dtable *l2;
1319 	struct l2_bucket *l2b;
1320 	u_short l1idx;
1321 
1322 	l1idx = L1_IDX(va);
1323 
1324 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL ||
1325 	    (l2b = &l2->l2_bucket[L2_BUCKET(l1idx)])->l2b_kva == NULL)
1326 		return (NULL);
1327 
1328 	return (l2b);
1329 }
1330 
1331 /*
1332  * Returns a pointer to the L2 bucket associated with the specified pmap
1333  * and VA.
1334  *
1335  * If no L2 bucket exists, perform the necessary allocations to put an L2
1336  * bucket/page table in place.
1337  *
1338  * Note that if a new L2 bucket/page was allocated, the caller *must*
1339  * increment the bucket occupancy counter appropriately *before*
1340  * releasing the pmap's lock to ensure no other thread or cpu deallocates
1341  * the bucket/page in the meantime.
1342  */
1343 static struct l2_bucket *
1344 pmap_alloc_l2_bucket(pmap_t pm, vaddr_t va)
1345 {
1346 	struct l2_dtable *l2;
1347 	struct l2_bucket *l2b;
1348 	u_short l1idx;
1349 
1350 	l1idx = L1_IDX(va);
1351 
1352 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
1353 		/*
1354 		 * No mapping at this address, as there is
1355 		 * no entry in the L1 table.
1356 		 * Need to allocate a new l2_dtable.
1357 		 */
1358 		if ((l2 = pmap_alloc_l2_dtable()) == NULL)
1359 			return (NULL);
1360 
1361 		/*
1362 		 * Link it into the parent pmap
1363 		 */
1364 		pm->pm_l2[L2_IDX(l1idx)] = l2;
1365 	}
1366 
1367 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
1368 
1369 	/*
1370 	 * Fetch pointer to the L2 page table associated with the address.
1371 	 */
1372 	if (l2b->l2b_kva == NULL) {
1373 		pt_entry_t *ptep;
1374 
1375 		/*
1376 		 * No L2 page table has been allocated. Chances are, this
1377 		 * is because we just allocated the l2_dtable, above.
1378 		 */
1379 		if ((ptep = pmap_alloc_l2_ptp(&l2b->l2b_phys)) == NULL) {
1380 			/*
1381 			 * Oops, no more L2 page tables available at this
1382 			 * time. We may need to deallocate the l2_dtable
1383 			 * if we allocated a new one above.
1384 			 */
1385 			if (l2->l2_occupancy == 0) {
1386 				pm->pm_l2[L2_IDX(l1idx)] = NULL;
1387 				pmap_free_l2_dtable(l2);
1388 			}
1389 			return (NULL);
1390 		}
1391 
1392 		l2->l2_occupancy++;
1393 		l2b->l2b_kva = ptep;
1394 		l2b->l2b_l1idx = l1idx;
1395 	}
1396 
1397 	return (l2b);
1398 }
1399 
1400 /*
1401  * One or more mappings in the specified L2 descriptor table have just been
1402  * invalidated.
1403  *
1404  * Garbage collect the metadata and descriptor table itself if necessary.
1405  *
1406  * The pmap lock must be acquired when this is called (not necessary
1407  * for the kernel pmap).
1408  */
1409 static void
1410 pmap_free_l2_bucket(pmap_t pm, struct l2_bucket *l2b, u_int count)
1411 {
1412 	struct l2_dtable *l2;
1413 	pd_entry_t *pl1pd, l1pd;
1414 	pt_entry_t *ptep;
1415 	u_short l1idx;
1416 
1417 	KDASSERT(count <= l2b->l2b_occupancy);
1418 
1419 	/*
1420 	 * Update the bucket's reference count according to how many
1421 	 * PTEs the caller has just invalidated.
1422 	 */
1423 	l2b->l2b_occupancy -= count;
1424 
1425 	/*
1426 	 * Note:
1427 	 *
1428 	 * Level 2 page tables allocated to the kernel pmap are never freed
1429 	 * as that would require checking all Level 1 page tables and
1430 	 * removing any references to the Level 2 page table. See also the
1431 	 * comment elsewhere about never freeing bootstrap L2 descriptors.
1432 	 *
1433 	 * We make do with just invalidating the mapping in the L2 table.
1434 	 *
1435 	 * This isn't really a big deal in practice and, in fact, leads
1436 	 * to a performance win over time as we don't need to continually
1437 	 * alloc/free.
1438 	 */
1439 	if (l2b->l2b_occupancy > 0 || pm == pmap_kernel())
1440 		return;
1441 
1442 	/*
1443 	 * There are no more valid mappings in this level 2 page table.
1444 	 * Go ahead and NULL-out the pointer in the bucket, then
1445 	 * free the page table.
1446 	 */
1447 	l1idx = l2b->l2b_l1idx;
1448 	ptep = l2b->l2b_kva;
1449 	l2b->l2b_kva = NULL;
1450 
1451 	pl1pd = pmap_l1_kva(pm) + l1idx;
1452 
1453 	/*
1454 	 * If the L1 slot matches the pmap's domain
1455 	 * number, then invalidate it.
1456 	 */
1457 	l1pd = *pl1pd & (L1_TYPE_MASK | L1_C_DOM_MASK);
1458 	if (l1pd == (L1_C_DOM(pmap_domain(pm)) | L1_TYPE_C)) {
1459 		*pl1pd = 0;
1460 		PTE_SYNC(pl1pd);
1461 	}
1462 
1463 	/*
1464 	 * Release the L2 descriptor table back to the pool cache.
1465 	 */
1466 #ifndef PMAP_INCLUDE_PTE_SYNC
1467 	pmap_free_l2_ptp(ptep, l2b->l2b_phys);
1468 #else
1469 	pmap_free_l2_ptp(!pmap_is_cached(pm), ptep, l2b->l2b_phys);
1470 #endif
1471 
1472 	/*
1473 	 * Update the reference count in the associated l2_dtable
1474 	 */
1475 	l2 = pm->pm_l2[L2_IDX(l1idx)];
1476 	if (--l2->l2_occupancy > 0)
1477 		return;
1478 
1479 	/*
1480 	 * There are no more valid mappings in any of the Level 1
1481 	 * slots managed by this l2_dtable. Go ahead and NULL-out
1482 	 * the pointer in the parent pmap and free the l2_dtable.
1483 	 */
1484 	pm->pm_l2[L2_IDX(l1idx)] = NULL;
1485 	pmap_free_l2_dtable(l2);
1486 }
1487 
1488 /*
1489  * Pool cache constructors for L2 descriptor tables, metadata and pmap
1490  * structures.
1491  */
1492 static int
1493 pmap_l2ptp_ctor(void *arg, void *v, int flags)
1494 {
1495 #ifndef PMAP_INCLUDE_PTE_SYNC
1496 	vaddr_t va = (vaddr_t)v & ~PGOFSET;
1497 
1498 	/*
1499 	 * The mappings for these page tables were initially made using
1500 	 * pmap_kenter_pa() by the pool subsystem. Therefore, the cache-
1501 	 * mode will not be right for page table mappings. To avoid
1502 	 * polluting the pmap_kenter_pa() code with a special case for
1503 	 * page tables, we simply fix up the cache-mode here if it's not
1504 	 * correct.
1505 	 */
1506 	struct l2_bucket * const l2b = pmap_get_l2_bucket(pmap_kernel(), va);
1507 	KDASSERT(l2b != NULL);
1508 	pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(va)];
1509 	pt_entry_t opte = *ptep;
1510 
1511 	if ((opte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
1512 		/*
1513 		 * Page tables must have the cache-mode set to Write-Thru.
1514 		 */
1515 		const pt_entry_t npte = (pte & ~L2_S_CACHE_MASK)
1516 		    | pte_l2_s_cache_mode_pt;
1517 		l2pte_set(ptep, npte, opte);
1518 		PTE_SYNC(ptep);
1519 		cpu_tlb_flushD_SE(va);
1520 		cpu_cpwait();
1521 	}
1522 #endif
1523 
1524 	memset(v, 0, L2_TABLE_SIZE_REAL);
1525 	PTE_SYNC_RANGE(v, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1526 	return (0);
1527 }
1528 
1529 static int
1530 pmap_l2dtable_ctor(void *arg, void *v, int flags)
1531 {
1532 
1533 	memset(v, 0, sizeof(struct l2_dtable));
1534 	return (0);
1535 }
1536 
1537 static int
1538 pmap_pmap_ctor(void *arg, void *v, int flags)
1539 {
1540 
1541 	memset(v, 0, sizeof(struct pmap));
1542 	return (0);
1543 }
1544 
1545 static void
1546 pmap_pinit(pmap_t pm)
1547 {
1548 #ifndef ARM_HAS_VBAR
1549 	struct l2_bucket *l2b;
1550 
1551 	if (vector_page < KERNEL_BASE) {
1552 		/*
1553 		 * Map the vector page.
1554 		 */
1555 		pmap_enter(pm, vector_page, systempage.pv_pa,
1556 		    VM_PROT_READ | VM_PROT_EXECUTE,
1557 		    VM_PROT_READ | VM_PROT_EXECUTE | PMAP_WIRED);
1558 		pmap_update(pm);
1559 
1560 		pm->pm_pl1vec = pmap_l1_kva(pm) + L1_IDX(vector_page);
1561 		l2b = pmap_get_l2_bucket(pm, vector_page);
1562 		KDASSERT(l2b != NULL);
1563 		pm->pm_l1vec = l2b->l2b_phys | L1_C_PROTO |
1564 		    L1_C_DOM(pmap_domain(pm));
1565 	} else
1566 		pm->pm_pl1vec = NULL;
1567 #endif
1568 }
1569 
1570 #ifdef PMAP_CACHE_VIVT
1571 /*
1572  * Since we have a virtually indexed cache, we may need to inhibit caching if
1573  * there is more than one mapping and at least one of them is writable.
1574  * Since we purge the cache on every context switch, we only need to check for
1575  * other mappings within the same pmap, or kernel_pmap.
1576  * This function is also called when a page is unmapped, to possibly reenable
1577  * caching on any remaining mappings.
1578  *
1579  * The code implements the following logic, where:
1580  *
1581  * KW = # of kernel read/write pages
1582  * KR = # of kernel read only pages
1583  * UW = # of user read/write pages
1584  * UR = # of user read only pages
1585  *
1586  * KC = kernel mapping is cacheable
1587  * UC = user mapping is cacheable
1588  *
1589  *               KW=0,KR=0  KW=0,KR>0  KW=1,KR=0  KW>1,KR>=0
1590  *             +---------------------------------------------
1591  * UW=0,UR=0   | ---        KC=1       KC=1       KC=0
1592  * UW=0,UR>0   | UC=1       KC=1,UC=1  KC=0,UC=0  KC=0,UC=0
1593  * UW=1,UR=0   | UC=1       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1594  * UW>1,UR>=0  | UC=0       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1595  */
1596 
1597 static const int pmap_vac_flags[4][4] = {
1598 	{-1,		0,		0,		PVF_KNC},
1599 	{0,		0,		PVF_NC,		PVF_NC},
1600 	{0,		PVF_NC,		PVF_NC,		PVF_NC},
1601 	{PVF_UNC,	PVF_NC,		PVF_NC,		PVF_NC}
1602 };
1603 
1604 static inline int
1605 pmap_get_vac_flags(const struct vm_page_md *md)
1606 {
1607 	int kidx, uidx;
1608 
1609 	kidx = 0;
1610 	if (md->kro_mappings || md->krw_mappings > 1)
1611 		kidx |= 1;
1612 	if (md->krw_mappings)
1613 		kidx |= 2;
1614 
1615 	uidx = 0;
1616 	if (md->uro_mappings || md->urw_mappings > 1)
1617 		uidx |= 1;
1618 	if (md->urw_mappings)
1619 		uidx |= 2;
1620 
1621 	return (pmap_vac_flags[uidx][kidx]);
1622 }
1623 
1624 static inline void
1625 pmap_vac_me_harder(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1626 {
1627 	int nattr;
1628 
1629 	nattr = pmap_get_vac_flags(md);
1630 
1631 	if (nattr < 0) {
1632 		md->pvh_attrs &= ~PVF_NC;
1633 		return;
1634 	}
1635 
1636 	if (nattr == 0 && (md->pvh_attrs & PVF_NC) == 0)
1637 		return;
1638 
1639 	if (pm == pmap_kernel())
1640 		pmap_vac_me_kpmap(md, pa, pm, va);
1641 	else
1642 		pmap_vac_me_user(md, pa, pm, va);
1643 
1644 	md->pvh_attrs = (md->pvh_attrs & ~PVF_NC) | nattr;
1645 }
1646 
1647 static void
1648 pmap_vac_me_kpmap(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1649 {
1650 	u_int u_cacheable, u_entries;
1651 	struct pv_entry *pv;
1652 	pmap_t last_pmap = pm;
1653 
1654 	/*
1655 	 * Pass one, see if there are both kernel and user pmaps for
1656 	 * this page.  Calculate whether there are user-writable or
1657 	 * kernel-writable pages.
1658 	 */
1659 	u_cacheable = 0;
1660 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1661 		if (pv->pv_pmap != pm && (pv->pv_flags & PVF_NC) == 0)
1662 			u_cacheable++;
1663 	}
1664 
1665 	u_entries = md->urw_mappings + md->uro_mappings;
1666 
1667 	/*
1668 	 * We know we have just been updating a kernel entry, so if
1669 	 * all user pages are already cacheable, then there is nothing
1670 	 * further to do.
1671 	 */
1672 	if (md->k_mappings == 0 && u_cacheable == u_entries)
1673 		return;
1674 
1675 	if (u_entries) {
1676 		/*
1677 		 * Scan over the list again, for each entry, if it
1678 		 * might not be set correctly, call pmap_vac_me_user
1679 		 * to recalculate the settings.
1680 		 */
1681 		SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1682 			/*
1683 			 * We know kernel mappings will get set
1684 			 * correctly in other calls.  We also know
1685 			 * that if the pmap is the same as last_pmap
1686 			 * then we've just handled this entry.
1687 			 */
1688 			if (pv->pv_pmap == pm || pv->pv_pmap == last_pmap)
1689 				continue;
1690 
1691 			/*
1692 			 * If there are kernel entries and this page
1693 			 * is writable but non-cacheable, then we can
1694 			 * skip this entry also.
1695 			 */
1696 			if (md->k_mappings &&
1697 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) ==
1698 			    (PVF_NC | PVF_WRITE))
1699 				continue;
1700 
1701 			/*
1702 			 * Similarly if there are no kernel-writable
1703 			 * entries and the page is already
1704 			 * read-only/cacheable.
1705 			 */
1706 			if (md->krw_mappings == 0 &&
1707 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 0)
1708 				continue;
1709 
1710 			/*
1711 			 * For some of the remaining cases, we know
1712 			 * that we must recalculate, but for others we
1713 			 * can't tell if they are correct or not, so
1714 			 * we recalculate anyway.
1715 			 */
1716 			pmap_vac_me_user(md, pa, (last_pmap = pv->pv_pmap), 0);
1717 		}
1718 
1719 		if (md->k_mappings == 0)
1720 			return;
1721 	}
1722 
1723 	pmap_vac_me_user(md, pa, pm, va);
1724 }
1725 
1726 static void
1727 pmap_vac_me_user(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1728 {
1729 	pmap_t kpmap = pmap_kernel();
1730 	struct pv_entry *pv, *npv = NULL;
1731 	u_int entries = 0;
1732 	u_int writable = 0;
1733 	u_int cacheable_entries = 0;
1734 	u_int kern_cacheable = 0;
1735 	u_int other_writable = 0;
1736 
1737 	/*
1738 	 * Count mappings and writable mappings in this pmap.
1739 	 * Include kernel mappings as part of our own.
1740 	 * Keep a pointer to the first one.
1741 	 */
1742 	npv = NULL;
1743 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1744 		/* Count mappings in the same pmap */
1745 		if (pm == pv->pv_pmap || kpmap == pv->pv_pmap) {
1746 			if (entries++ == 0)
1747 				npv = pv;
1748 
1749 			/* Cacheable mappings */
1750 			if ((pv->pv_flags & PVF_NC) == 0) {
1751 				cacheable_entries++;
1752 				if (kpmap == pv->pv_pmap)
1753 					kern_cacheable++;
1754 			}
1755 
1756 			/* Writable mappings */
1757 			if (pv->pv_flags & PVF_WRITE)
1758 				++writable;
1759 		} else
1760 		if (pv->pv_flags & PVF_WRITE)
1761 			other_writable = 1;
1762 	}
1763 
1764 	/*
1765 	 * Enable or disable caching as necessary.
1766 	 * Note: the first entry might be part of the kernel pmap,
1767 	 * so we can't assume this is indicative of the state of the
1768 	 * other (maybe non-kpmap) entries.
1769 	 */
1770 	if ((entries > 1 && writable) ||
1771 	    (entries > 0 && pm == kpmap && other_writable)) {
1772 		if (cacheable_entries == 0)
1773 			return;
1774 
1775 		for (pv = npv; pv; pv = SLIST_NEXT(pv, pv_link)) {
1776 			if ((pm != pv->pv_pmap && kpmap != pv->pv_pmap) ||
1777 			    (pv->pv_flags & PVF_NC))
1778 				continue;
1779 
1780 			pv->pv_flags |= PVF_NC;
1781 
1782 			struct l2_bucket * const l2b
1783 			    = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
1784 			KDASSERT(l2b != NULL);
1785 			pt_entry_t * const ptep
1786 			    = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1787 			const pt_entry_t opte = *ptep;
1788 			pt_entry_t npte = opte & ~L2_S_CACHE_MASK;
1789 
1790 			if ((va != pv->pv_va || pm != pv->pv_pmap)
1791 			    && l2pte_valid(npte)) {
1792 #ifdef PMAP_CACHE_VIVT
1793 				pmap_cache_wbinv_page(pv->pv_pmap, pv->pv_va,
1794 				    true, pv->pv_flags);
1795 #endif
1796 				pmap_tlb_flush_SE(pv->pv_pmap, pv->pv_va,
1797 				    pv->pv_flags);
1798 			}
1799 
1800 			l2pte_set(ptep, npte, opte);
1801 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
1802 		}
1803 		cpu_cpwait();
1804 	} else
1805 	if (entries > cacheable_entries) {
1806 		/*
1807 		 * Turn cacheing back on for some pages.  If it is a kernel
1808 		 * page, only do so if there are no other writable pages.
1809 		 */
1810 		for (pv = npv; pv; pv = SLIST_NEXT(pv, pv_link)) {
1811 			if (!(pv->pv_flags & PVF_NC) || (pm != pv->pv_pmap &&
1812 			    (kpmap != pv->pv_pmap || other_writable)))
1813 				continue;
1814 
1815 			pv->pv_flags &= ~PVF_NC;
1816 
1817 			struct l2_bucket * const l2b
1818 			    = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
1819 			KDASSERT(l2b != NULL);
1820 			pt_entry_t * const ptep
1821 			    = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1822 			const pt_entry_t opte = *ptep;
1823 			pt_entry_t npte = (opte & ~L2_S_CACHE_MASK)
1824 			    | pte_l2_s_cache_mode;
1825 
1826 			if (l2pte_valid(opte)) {
1827 				pmap_tlb_flush_SE(pv->pv_pmap, pv->pv_va,
1828 				    pv->pv_flags);
1829 			}
1830 
1831 			l2pte_set(ptep, npte, opte);
1832 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
1833 		}
1834 	}
1835 }
1836 #endif
1837 
1838 #ifdef PMAP_CACHE_VIPT
1839 static void
1840 pmap_vac_me_harder(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1841 {
1842 	struct pv_entry *pv;
1843 	vaddr_t tst_mask;
1844 	bool bad_alias;
1845 	const u_int
1846 	    rw_mappings = md->urw_mappings + md->krw_mappings,
1847 	    ro_mappings = md->uro_mappings + md->kro_mappings;
1848 
1849 	/* do we need to do anything? */
1850 	if (arm_cache_prefer_mask == 0)
1851 		return;
1852 
1853 	NPDEBUG(PDB_VAC, printf("pmap_vac_me_harder: md=%p, pmap=%p va=%08lx\n",
1854 	    md, pm, va));
1855 
1856 	KASSERT(!va || pm);
1857 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1858 
1859 	/* Already a conflict? */
1860 	if (__predict_false(md->pvh_attrs & PVF_NC)) {
1861 		/* just an add, things are already non-cached */
1862 		KASSERT(!(md->pvh_attrs & PVF_DIRTY));
1863 		KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
1864 		bad_alias = false;
1865 		if (va) {
1866 			PMAPCOUNT(vac_color_none);
1867 			bad_alias = true;
1868 			KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
1869 			goto fixup;
1870 		}
1871 		pv = SLIST_FIRST(&md->pvh_list);
1872 		/* the list can't be empty because it would be cachable */
1873 		if (md->pvh_attrs & PVF_KMPAGE) {
1874 			tst_mask = md->pvh_attrs;
1875 		} else {
1876 			KASSERT(pv);
1877 			tst_mask = pv->pv_va;
1878 			pv = SLIST_NEXT(pv, pv_link);
1879 		}
1880 		/*
1881 		 * Only check for a bad alias if we have writable mappings.
1882 		 */
1883 		tst_mask &= arm_cache_prefer_mask;
1884 		if (rw_mappings > 0) {
1885 			for (; pv && !bad_alias; pv = SLIST_NEXT(pv, pv_link)) {
1886 				/* if there's a bad alias, stop checking. */
1887 				if (tst_mask != (pv->pv_va & arm_cache_prefer_mask))
1888 					bad_alias = true;
1889 			}
1890 			md->pvh_attrs |= PVF_WRITE;
1891 			if (!bad_alias)
1892 				md->pvh_attrs |= PVF_DIRTY;
1893 		} else {
1894 			/*
1895 			 * We have only read-only mappings.  Let's see if there
1896 			 * are multiple colors in use or if we mapped a KMPAGE.
1897 			 * If the latter, we have a bad alias.  If the former,
1898 			 * we need to remember that.
1899 			 */
1900 			for (; pv; pv = SLIST_NEXT(pv, pv_link)) {
1901 				if (tst_mask != (pv->pv_va & arm_cache_prefer_mask)) {
1902 					if (md->pvh_attrs & PVF_KMPAGE)
1903 						bad_alias = true;
1904 					break;
1905 				}
1906 			}
1907 			md->pvh_attrs &= ~PVF_WRITE;
1908 			/*
1909 			 * No KMPAGE and we exited early, so we must have
1910 			 * multiple color mappings.
1911 			 */
1912 			if (!bad_alias && pv != NULL)
1913 				md->pvh_attrs |= PVF_MULTCLR;
1914 		}
1915 
1916 		/* If no conflicting colors, set everything back to cached */
1917 		if (!bad_alias) {
1918 #ifdef DEBUG
1919 			if ((md->pvh_attrs & PVF_WRITE)
1920 			    || ro_mappings < 2) {
1921 				SLIST_FOREACH(pv, &md->pvh_list, pv_link)
1922 					KDASSERT(((tst_mask ^ pv->pv_va) & arm_cache_prefer_mask) == 0);
1923 			}
1924 #endif
1925 			md->pvh_attrs &= (PAGE_SIZE - 1) & ~PVF_NC;
1926 			md->pvh_attrs |= tst_mask | PVF_COLORED;
1927 			/*
1928 			 * Restore DIRTY bit if page is modified
1929 			 */
1930 			if (md->pvh_attrs & PVF_DMOD)
1931 				md->pvh_attrs |= PVF_DIRTY;
1932 			PMAPCOUNT(vac_color_restore);
1933 		} else {
1934 			KASSERT(SLIST_FIRST(&md->pvh_list) != NULL);
1935 			KASSERT(SLIST_NEXT(SLIST_FIRST(&md->pvh_list), pv_link) != NULL);
1936 		}
1937 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1938 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
1939 	} else if (!va) {
1940 		KASSERT(pmap_is_page_colored_p(md));
1941 		KASSERT(!(md->pvh_attrs & PVF_WRITE)
1942 		    || (md->pvh_attrs & PVF_DIRTY));
1943 		if (rw_mappings == 0) {
1944 			md->pvh_attrs &= ~PVF_WRITE;
1945 			if (ro_mappings == 1
1946 			    && (md->pvh_attrs & PVF_MULTCLR)) {
1947 				/*
1948 				 * If this is the last readonly mapping
1949 				 * but it doesn't match the current color
1950 				 * for the page, change the current color
1951 				 * to match this last readonly mapping.
1952 				 */
1953 				pv = SLIST_FIRST(&md->pvh_list);
1954 				tst_mask = (md->pvh_attrs ^ pv->pv_va)
1955 				    & arm_cache_prefer_mask;
1956 				if (tst_mask) {
1957 					md->pvh_attrs ^= tst_mask;
1958 					PMAPCOUNT(vac_color_change);
1959 				}
1960 			}
1961 		}
1962 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1963 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
1964 		return;
1965 	} else if (!pmap_is_page_colored_p(md)) {
1966 		/* not colored so we just use its color */
1967 		KASSERT(md->pvh_attrs & (PVF_WRITE|PVF_DIRTY));
1968 		KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
1969 		PMAPCOUNT(vac_color_new);
1970 		md->pvh_attrs &= PAGE_SIZE - 1;
1971 		md->pvh_attrs |= PVF_COLORED
1972 		    | (va & arm_cache_prefer_mask)
1973 		    | (rw_mappings > 0 ? PVF_WRITE : 0);
1974 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1975 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
1976 		return;
1977 	} else if (((md->pvh_attrs ^ va) & arm_cache_prefer_mask) == 0) {
1978 		bad_alias = false;
1979 		if (rw_mappings > 0) {
1980 			/*
1981 			 * We now have writeable mappings and if we have
1982 			 * readonly mappings in more than once color, we have
1983 			 * an aliasing problem.  Regardless mark the page as
1984 			 * writeable.
1985 			 */
1986 			if (md->pvh_attrs & PVF_MULTCLR) {
1987 				if (ro_mappings < 2) {
1988 					/*
1989 					 * If we only have less than two
1990 					 * read-only mappings, just flush the
1991 					 * non-primary colors from the cache.
1992 					 */
1993 					pmap_flush_page(md, pa,
1994 					    PMAP_FLUSH_SECONDARY);
1995 				} else {
1996 					bad_alias = true;
1997 				}
1998 			}
1999 			md->pvh_attrs |= PVF_WRITE;
2000 		}
2001 		/* If no conflicting colors, set everything back to cached */
2002 		if (!bad_alias) {
2003 #ifdef DEBUG
2004 			if (rw_mappings > 0
2005 			    || (md->pvh_attrs & PMAP_KMPAGE)) {
2006 				tst_mask = md->pvh_attrs & arm_cache_prefer_mask;
2007 				SLIST_FOREACH(pv, &md->pvh_list, pv_link)
2008 					KDASSERT(((tst_mask ^ pv->pv_va) & arm_cache_prefer_mask) == 0);
2009 			}
2010 #endif
2011 			if (SLIST_EMPTY(&md->pvh_list))
2012 				PMAPCOUNT(vac_color_reuse);
2013 			else
2014 				PMAPCOUNT(vac_color_ok);
2015 
2016 			/* matching color, just return */
2017 			KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2018 			KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2019 			return;
2020 		}
2021 		KASSERT(SLIST_FIRST(&md->pvh_list) != NULL);
2022 		KASSERT(SLIST_NEXT(SLIST_FIRST(&md->pvh_list), pv_link) != NULL);
2023 
2024 		/* color conflict.  evict from cache. */
2025 
2026 		pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
2027 		md->pvh_attrs &= ~PVF_COLORED;
2028 		md->pvh_attrs |= PVF_NC;
2029 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2030 		KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2031 		PMAPCOUNT(vac_color_erase);
2032 	} else if (rw_mappings == 0
2033 		   && (md->pvh_attrs & PVF_KMPAGE) == 0) {
2034 		KASSERT((md->pvh_attrs & PVF_WRITE) == 0);
2035 
2036 		/*
2037 		 * If the page has dirty cache lines, clean it.
2038 		 */
2039 		if (md->pvh_attrs & PVF_DIRTY)
2040 			pmap_flush_page(md, pa, PMAP_CLEAN_PRIMARY);
2041 
2042 		/*
2043 		 * If this is the first remapping (we know that there are no
2044 		 * writeable mappings), then this is a simple color change.
2045 		 * Otherwise this is a seconary r/o mapping, which means
2046 		 * we don't have to do anything.
2047 		 */
2048 		if (ro_mappings == 1) {
2049 			KASSERT(((md->pvh_attrs ^ va) & arm_cache_prefer_mask) != 0);
2050 			md->pvh_attrs &= PAGE_SIZE - 1;
2051 			md->pvh_attrs |= (va & arm_cache_prefer_mask);
2052 			PMAPCOUNT(vac_color_change);
2053 		} else {
2054 			PMAPCOUNT(vac_color_blind);
2055 		}
2056 		md->pvh_attrs |= PVF_MULTCLR;
2057 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2058 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2059 		return;
2060 	} else {
2061 		if (rw_mappings > 0)
2062 			md->pvh_attrs |= PVF_WRITE;
2063 
2064 		/* color conflict.  evict from cache. */
2065 		pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
2066 
2067 		/* the list can't be empty because this was a enter/modify */
2068 		pv = SLIST_FIRST(&md->pvh_list);
2069 		if ((md->pvh_attrs & PVF_KMPAGE) == 0) {
2070 			KASSERT(pv);
2071 			/*
2072 			 * If there's only one mapped page, change color to the
2073 			 * page's new color and return.  Restore the DIRTY bit
2074 			 * that was erased by pmap_flush_page.
2075 			 */
2076 			if (SLIST_NEXT(pv, pv_link) == NULL) {
2077 				md->pvh_attrs &= PAGE_SIZE - 1;
2078 				md->pvh_attrs |= (va & arm_cache_prefer_mask);
2079 				if (md->pvh_attrs & PVF_DMOD)
2080 					md->pvh_attrs |= PVF_DIRTY;
2081 				PMAPCOUNT(vac_color_change);
2082 				KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2083 				KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2084 				KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2085 				return;
2086 			}
2087 		}
2088 		bad_alias = true;
2089 		md->pvh_attrs &= ~PVF_COLORED;
2090 		md->pvh_attrs |= PVF_NC;
2091 		PMAPCOUNT(vac_color_erase);
2092 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2093 	}
2094 
2095   fixup:
2096 	KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2097 
2098 	/*
2099 	 * Turn cacheing on/off for all pages.
2100 	 */
2101 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
2102 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pv->pv_pmap,
2103 		    pv->pv_va);
2104 		KDASSERT(l2b != NULL);
2105 		pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2106 		const pt_entry_t opte = *ptep;
2107 		pt_entry_t npte = opte & ~L2_S_CACHE_MASK;
2108 		if (bad_alias) {
2109 			pv->pv_flags |= PVF_NC;
2110 		} else {
2111 			pv->pv_flags &= ~PVF_NC;
2112 			npte |= pte_l2_s_cache_mode;
2113 		}
2114 
2115 		if (opte == npte)	/* only update is there's a change */
2116 			continue;
2117 
2118 		if (l2pte_valid(npte)) {
2119 			pmap_tlb_flush_SE(pv->pv_pmap, pv->pv_va, pv->pv_flags);
2120 		}
2121 
2122 		l2pte_set(ptep, npte, opte);
2123 		PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
2124 	}
2125 }
2126 #endif	/* PMAP_CACHE_VIPT */
2127 
2128 
2129 /*
2130  * Modify pte bits for all ptes corresponding to the given physical address.
2131  * We use `maskbits' rather than `clearbits' because we're always passing
2132  * constants and the latter would require an extra inversion at run-time.
2133  */
2134 static void
2135 pmap_clearbit(struct vm_page_md *md, paddr_t pa, u_int maskbits)
2136 {
2137 	struct pv_entry *pv;
2138 	pmap_t pm;
2139 	vaddr_t va;
2140 	u_int oflags;
2141 #ifdef PMAP_CACHE_VIPT
2142 	const bool want_syncicache = PV_IS_EXEC_P(md->pvh_attrs);
2143 	bool need_vac_me_harder = false;
2144 	bool need_syncicache = false;
2145 #endif
2146 
2147 	NPDEBUG(PDB_BITS,
2148 	    printf("pmap_clearbit: md %p mask 0x%x\n",
2149 	    md, maskbits));
2150 
2151 #ifdef PMAP_CACHE_VIPT
2152 	/*
2153 	 * If we might want to sync the I-cache and we've modified it,
2154 	 * then we know we definitely need to sync or discard it.
2155 	 */
2156 	if (want_syncicache) {
2157 		need_syncicache = md->pvh_attrs & PVF_MOD;
2158 	}
2159 #endif
2160 	/*
2161 	 * Clear saved attributes (modify, reference)
2162 	 */
2163 	md->pvh_attrs &= ~(maskbits & (PVF_MOD | PVF_REF));
2164 
2165 	if (SLIST_EMPTY(&md->pvh_list)) {
2166 #if defined(PMAP_CACHE_VIPT)
2167 		if (need_syncicache) {
2168 			/*
2169 			 * No one has it mapped, so just discard it.  The next
2170 			 * exec remapping will cause it to be synced.
2171 			 */
2172 			md->pvh_attrs &= ~PVF_EXEC;
2173 			PMAPCOUNT(exec_discarded_clearbit);
2174 		}
2175 #endif
2176 		return;
2177 	}
2178 
2179 	/*
2180 	 * Loop over all current mappings setting/clearing as appropos
2181 	 */
2182 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
2183 		va = pv->pv_va;
2184 		pm = pv->pv_pmap;
2185 		oflags = pv->pv_flags;
2186 		/*
2187 		 * Kernel entries are unmanaged and as such not to be changed.
2188 		 */
2189 		if (oflags & PVF_KENTRY)
2190 			continue;
2191 		pv->pv_flags &= ~maskbits;
2192 
2193 		pmap_acquire_pmap_lock(pm);
2194 
2195 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pm, va);
2196 		KDASSERT(l2b != NULL);
2197 
2198 		pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
2199 		const pt_entry_t opte = *ptep;
2200 		pt_entry_t npte = opte;
2201 
2202 		NPDEBUG(PDB_BITS,
2203 		    printf(
2204 		    "pmap_clearbit: pv %p, pm %p, va 0x%08lx, flag 0x%x\n",
2205 		    pv, pv->pv_pmap, pv->pv_va, oflags));
2206 
2207 		if (maskbits & (PVF_WRITE|PVF_MOD)) {
2208 #ifdef PMAP_CACHE_VIVT
2209 			if ((pv->pv_flags & PVF_NC)) {
2210 				/*
2211 				 * Entry is not cacheable:
2212 				 *
2213 				 * Don't turn caching on again if this is a
2214 				 * modified emulation. This would be
2215 				 * inconsitent with the settings created by
2216 				 * pmap_vac_me_harder(). Otherwise, it's safe
2217 				 * to re-enable cacheing.
2218 				 *
2219 				 * There's no need to call pmap_vac_me_harder()
2220 				 * here: all pages are losing their write
2221 				 * permission.
2222 				 */
2223 				if (maskbits & PVF_WRITE) {
2224 					npte |= pte_l2_s_cache_mode;
2225 					pv->pv_flags &= ~PVF_NC;
2226 				}
2227 			} else
2228 			if (l2pte_writable_p(opte)) {
2229 				/*
2230 				 * Entry is writable/cacheable: check if pmap
2231 				 * is current if it is flush it, otherwise it
2232 				 * won't be in the cache
2233 				 */
2234 				pmap_cache_wbinv_page(pm, pv->pv_va,
2235 				    (maskbits & PVF_REF) != 0,
2236 				    oflags|PVF_WRITE);
2237 			}
2238 #endif
2239 
2240 			/* make the pte read only */
2241 			npte = l2pte_set_readonly(npte);
2242 
2243 			if (maskbits & oflags & PVF_WRITE) {
2244 				/*
2245 				 * Keep alias accounting up to date
2246 				 */
2247 				if (pv->pv_pmap == pmap_kernel()) {
2248 					md->krw_mappings--;
2249 					md->kro_mappings++;
2250 				} else {
2251 					md->urw_mappings--;
2252 					md->uro_mappings++;
2253 				}
2254 #ifdef PMAP_CACHE_VIPT
2255 				if (arm_cache_prefer_mask != 0) {
2256 					if (md->urw_mappings + md->krw_mappings == 0) {
2257 						md->pvh_attrs &= ~PVF_WRITE;
2258 					} else {
2259 						PMAP_VALIDATE_MD_PAGE(md);
2260 					}
2261 				}
2262 				if (want_syncicache)
2263 					need_syncicache = true;
2264 				need_vac_me_harder = true;
2265 #endif
2266 			}
2267 		}
2268 
2269 		if (maskbits & PVF_REF) {
2270 			if ((pv->pv_flags & PVF_NC) == 0
2271 			    && (maskbits & (PVF_WRITE|PVF_MOD)) == 0
2272 			    && l2pte_valid(npte)) {
2273 #ifdef PMAP_CACHE_VIVT
2274 				/*
2275 				 * Check npte here; we may have already
2276 				 * done the wbinv above, and the validity
2277 				 * of the PTE is the same for opte and
2278 				 * npte.
2279 				 */
2280 				pmap_cache_wbinv_page(pm, pv->pv_va, true,
2281 				    oflags);
2282 #endif
2283 			}
2284 
2285 			/*
2286 			 * Make the PTE invalid so that we will take a
2287 			 * page fault the next time the mapping is
2288 			 * referenced.
2289 			 */
2290 			npte &= ~L2_TYPE_MASK;
2291 			npte |= L2_TYPE_INV;
2292 		}
2293 
2294 		if (npte != opte) {
2295 			l2pte_set(ptep, npte, opte);
2296 			PTE_SYNC(ptep);
2297 
2298 			/* Flush the TLB entry if a current pmap. */
2299 			pmap_tlb_flush_SE(pm, pv->pv_va, oflags);
2300 		}
2301 
2302 		pmap_release_pmap_lock(pm);
2303 
2304 		NPDEBUG(PDB_BITS,
2305 		    printf("pmap_clearbit: pm %p va 0x%lx opte 0x%08x npte 0x%08x\n",
2306 		    pm, va, opte, npte));
2307 	}
2308 
2309 #ifdef PMAP_CACHE_VIPT
2310 	/*
2311 	 * If we need to sync the I-cache and we haven't done it yet, do it.
2312 	 */
2313 	if (need_syncicache) {
2314 		pmap_syncicache_page(md, pa);
2315 		PMAPCOUNT(exec_synced_clearbit);
2316 	}
2317 
2318 	/*
2319 	 * If we are changing this to read-only, we need to call vac_me_harder
2320 	 * so we can change all the read-only pages to cacheable.  We pretend
2321 	 * this as a page deletion.
2322 	 */
2323 	if (need_vac_me_harder) {
2324 		if (md->pvh_attrs & PVF_NC)
2325 			pmap_vac_me_harder(md, pa, NULL, 0);
2326 	}
2327 #endif
2328 }
2329 
2330 /*
2331  * pmap_clean_page()
2332  *
2333  * This is a local function used to work out the best strategy to clean
2334  * a single page referenced by its entry in the PV table. It's used by
2335  * pmap_copy_page, pmap_zero page and maybe some others later on.
2336  *
2337  * Its policy is effectively:
2338  *  o If there are no mappings, we don't bother doing anything with the cache.
2339  *  o If there is one mapping, we clean just that page.
2340  *  o If there are multiple mappings, we clean the entire cache.
2341  *
2342  * So that some functions can be further optimised, it returns 0 if it didn't
2343  * clean the entire cache, or 1 if it did.
2344  *
2345  * XXX One bug in this routine is that if the pv_entry has a single page
2346  * mapped at 0x00000000 a whole cache clean will be performed rather than
2347  * just the 1 page. Since this should not occur in everyday use and if it does
2348  * it will just result in not the most efficient clean for the page.
2349  */
2350 #ifdef PMAP_CACHE_VIVT
2351 static int
2352 pmap_clean_page(struct pv_entry *pv, bool is_src)
2353 {
2354 	pmap_t pm_to_clean = NULL;
2355 	struct pv_entry *npv;
2356 	u_int cache_needs_cleaning = 0;
2357 	u_int flags = 0;
2358 	vaddr_t page_to_clean = 0;
2359 
2360 	if (pv == NULL) {
2361 		/* nothing mapped in so nothing to flush */
2362 		return (0);
2363 	}
2364 
2365 	/*
2366 	 * Since we flush the cache each time we change to a different
2367 	 * user vmspace, we only need to flush the page if it is in the
2368 	 * current pmap.
2369 	 */
2370 
2371 	for (npv = pv; npv; npv = SLIST_NEXT(npv, pv_link)) {
2372 		if (pmap_is_current(npv->pv_pmap)) {
2373 			flags |= npv->pv_flags;
2374 			/*
2375 			 * The page is mapped non-cacheable in
2376 			 * this map.  No need to flush the cache.
2377 			 */
2378 			if (npv->pv_flags & PVF_NC) {
2379 #ifdef DIAGNOSTIC
2380 				if (cache_needs_cleaning)
2381 					panic("pmap_clean_page: "
2382 					    "cache inconsistency");
2383 #endif
2384 				break;
2385 			} else if (is_src && (npv->pv_flags & PVF_WRITE) == 0)
2386 				continue;
2387 			if (cache_needs_cleaning) {
2388 				page_to_clean = 0;
2389 				break;
2390 			} else {
2391 				page_to_clean = npv->pv_va;
2392 				pm_to_clean = npv->pv_pmap;
2393 			}
2394 			cache_needs_cleaning = 1;
2395 		}
2396 	}
2397 
2398 	if (page_to_clean) {
2399 		pmap_cache_wbinv_page(pm_to_clean, page_to_clean,
2400 		    !is_src, flags | PVF_REF);
2401 	} else if (cache_needs_cleaning) {
2402 		pmap_t const pm = curproc->p_vmspace->vm_map.pmap;
2403 
2404 		pmap_cache_wbinv_all(pm, flags);
2405 		return (1);
2406 	}
2407 	return (0);
2408 }
2409 #endif
2410 
2411 #ifdef PMAP_CACHE_VIPT
2412 /*
2413  * Sync a page with the I-cache.  Since this is a VIPT, we must pick the
2414  * right cache alias to make sure we flush the right stuff.
2415  */
2416 void
2417 pmap_syncicache_page(struct vm_page_md *md, paddr_t pa)
2418 {
2419 	const vsize_t va_offset = md->pvh_attrs & arm_cache_prefer_mask;
2420 	pt_entry_t * const ptep = &cdst_pte[va_offset >> PGSHIFT];
2421 
2422 	NPDEBUG(PDB_EXEC, printf("pmap_syncicache_page: md=%p (attrs=%#x)\n",
2423 	    md, md->pvh_attrs));
2424 	/*
2425 	 * No need to clean the page if it's non-cached.
2426 	 */
2427 	if (md->pvh_attrs & PVF_NC)
2428 		return;
2429 	KASSERT(arm_cache_prefer_mask == 0 || md->pvh_attrs & PVF_COLORED);
2430 
2431 	pmap_tlb_flush_SE(pmap_kernel(), cdstp + va_offset, PVF_REF | PVF_EXEC);
2432 	/*
2433 	 * Set up a PTE with the right coloring to flush existing cache lines.
2434 	 */
2435 	const pt_entry_t npte = L2_S_PROTO |
2436 	    pa
2437 	    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE)
2438 	    | pte_l2_s_cache_mode;
2439 	l2pte_set(ptep, npte, 0);
2440 	PTE_SYNC(ptep);
2441 
2442 	/*
2443 	 * Flush it.
2444 	 */
2445 	cpu_icache_sync_range(cdstp + va_offset, PAGE_SIZE);
2446 	/*
2447 	 * Unmap the page.
2448 	 */
2449 	l2pte_reset(ptep);
2450 	PTE_SYNC(ptep);
2451 	pmap_tlb_flush_SE(pmap_kernel(), cdstp + va_offset, PVF_REF | PVF_EXEC);
2452 
2453 	md->pvh_attrs |= PVF_EXEC;
2454 	PMAPCOUNT(exec_synced);
2455 }
2456 
2457 void
2458 pmap_flush_page(struct vm_page_md *md, paddr_t pa, enum pmap_flush_op flush)
2459 {
2460 	vsize_t va_offset, end_va;
2461 	bool wbinv_p;
2462 
2463 	if (arm_cache_prefer_mask == 0)
2464 		return;
2465 
2466 	switch (flush) {
2467 	case PMAP_FLUSH_PRIMARY:
2468 		if (md->pvh_attrs & PVF_MULTCLR) {
2469 			va_offset = 0;
2470 			end_va = arm_cache_prefer_mask;
2471 			md->pvh_attrs &= ~PVF_MULTCLR;
2472 			PMAPCOUNT(vac_flush_lots);
2473 		} else {
2474 			va_offset = md->pvh_attrs & arm_cache_prefer_mask;
2475 			end_va = va_offset;
2476 			PMAPCOUNT(vac_flush_one);
2477 		}
2478 		/*
2479 		 * Mark that the page is no longer dirty.
2480 		 */
2481 		md->pvh_attrs &= ~PVF_DIRTY;
2482 		wbinv_p = true;
2483 		break;
2484 	case PMAP_FLUSH_SECONDARY:
2485 		va_offset = 0;
2486 		end_va = arm_cache_prefer_mask;
2487 		wbinv_p = true;
2488 		md->pvh_attrs &= ~PVF_MULTCLR;
2489 		PMAPCOUNT(vac_flush_lots);
2490 		break;
2491 	case PMAP_CLEAN_PRIMARY:
2492 		va_offset = md->pvh_attrs & arm_cache_prefer_mask;
2493 		end_va = va_offset;
2494 		wbinv_p = false;
2495 		/*
2496 		 * Mark that the page is no longer dirty.
2497 		 */
2498 		if ((md->pvh_attrs & PVF_DMOD) == 0)
2499 			md->pvh_attrs &= ~PVF_DIRTY;
2500 		PMAPCOUNT(vac_clean_one);
2501 		break;
2502 	default:
2503 		return;
2504 	}
2505 
2506 	KASSERT(!(md->pvh_attrs & PVF_NC));
2507 
2508 	NPDEBUG(PDB_VAC, printf("pmap_flush_page: md=%p (attrs=%#x)\n",
2509 	    md, md->pvh_attrs));
2510 
2511 	const size_t scache_line_size = arm_scache.dcache_line_size;
2512 
2513 	for (; va_offset <= end_va; va_offset += PAGE_SIZE) {
2514 		const size_t pte_offset = va_offset >> PGSHIFT;
2515 		pt_entry_t * const ptep = &cdst_pte[pte_offset];
2516 		const pt_entry_t opte = *ptep;
2517 
2518 		if (flush == PMAP_FLUSH_SECONDARY
2519 		    && va_offset == (md->pvh_attrs & arm_cache_prefer_mask))
2520 			continue;
2521 
2522 		pmap_tlb_flush_SE(pmap_kernel(), cdstp + va_offset,
2523 		    PVF_REF | PVF_EXEC);
2524 		/*
2525 		 * Set up a PTE with the right coloring to flush
2526 		 * existing cache entries.
2527 		 */
2528 		const pt_entry_t npte = L2_S_PROTO
2529 		    | pa
2530 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE)
2531 		    | pte_l2_s_cache_mode;
2532 		l2pte_set(ptep, npte, opte);
2533 		PTE_SYNC(ptep);
2534 
2535 		/*
2536 		 * Flush it.  Make sure to flush secondary cache too since
2537 		 * bus_dma will ignore uncached pages.
2538 		 */
2539                 vaddr_t va = cdstp + va_offset;
2540 		if (scache_line_size != 0) {
2541 			cpu_dcache_wb_range(va, PAGE_SIZE);
2542 			if (wbinv_p) {
2543 				cpu_sdcache_wbinv_range(va, pa, PAGE_SIZE);
2544 				cpu_dcache_inv_range(va, PAGE_SIZE);
2545 			} else {
2546 				cpu_sdcache_wb_range(va, pa, PAGE_SIZE);
2547 			}
2548 		} else {
2549 			if (wbinv_p) {
2550 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
2551 			} else {
2552 				cpu_dcache_wb_range(va, PAGE_SIZE);
2553 			}
2554 		}
2555 
2556 		/*
2557 		 * Restore the page table entry since we might have interrupted
2558 		 * pmap_zero_page or pmap_copy_page which was already using
2559 		 * this pte.
2560 		 */
2561 		l2pte_set(ptep, opte, npte);
2562 		PTE_SYNC(ptep);
2563 		pmap_tlb_flush_SE(pmap_kernel(), cdstp + va_offset,
2564 		    PVF_REF | PVF_EXEC);
2565 	}
2566 }
2567 #endif /* PMAP_CACHE_VIPT */
2568 
2569 /*
2570  * Routine:	pmap_page_remove
2571  * Function:
2572  *		Removes this physical page from
2573  *		all physical maps in which it resides.
2574  *		Reflects back modify bits to the pager.
2575  */
2576 static void
2577 pmap_page_remove(struct vm_page_md *md, paddr_t pa)
2578 {
2579 	struct l2_bucket *l2b;
2580 	struct pv_entry *pv, *npv, **pvp;
2581 	pmap_t pm;
2582 	pt_entry_t *ptep;
2583 	bool flush;
2584 	u_int flags;
2585 
2586 	NPDEBUG(PDB_FOLLOW,
2587 	    printf("pmap_page_remove: md %p (0x%08lx)\n", md,
2588 	    pa));
2589 
2590 	pv = SLIST_FIRST(&md->pvh_list);
2591 	if (pv == NULL) {
2592 #ifdef PMAP_CACHE_VIPT
2593 		/*
2594 		 * We *know* the page contents are about to be replaced.
2595 		 * Discard the exec contents
2596 		 */
2597 		if (PV_IS_EXEC_P(md->pvh_attrs))
2598 			PMAPCOUNT(exec_discarded_page_protect);
2599 		md->pvh_attrs &= ~PVF_EXEC;
2600 		PMAP_VALIDATE_MD_PAGE(md);
2601 #endif
2602 		return;
2603 	}
2604 #ifdef PMAP_CACHE_VIPT
2605 	KASSERT(arm_cache_prefer_mask == 0 || pmap_is_page_colored_p(md));
2606 #endif
2607 
2608 	/*
2609 	 * Clear alias counts
2610 	 */
2611 #ifdef PMAP_CACHE_VIVT
2612 	md->k_mappings = 0;
2613 #endif
2614 	md->urw_mappings = md->uro_mappings = 0;
2615 
2616 	flush = false;
2617 	flags = 0;
2618 
2619 #ifdef PMAP_CACHE_VIVT
2620 	pmap_clean_page(pv, false);
2621 #endif
2622 
2623 	pvp = &SLIST_FIRST(&md->pvh_list);
2624 	while (pv) {
2625 		pm = pv->pv_pmap;
2626 		npv = SLIST_NEXT(pv, pv_link);
2627 		if (flush == false && pmap_is_current(pm))
2628 			flush = true;
2629 
2630 		if (pm == pmap_kernel()) {
2631 #ifdef PMAP_CACHE_VIPT
2632 			/*
2633 			 * If this was unmanaged mapping, it must be preserved.
2634 			 * Move it back on the list and advance the end-of-list
2635 			 * pointer.
2636 			 */
2637 			if (pv->pv_flags & PVF_KENTRY) {
2638 				*pvp = pv;
2639 				pvp = &SLIST_NEXT(pv, pv_link);
2640 				pv = npv;
2641 				continue;
2642 			}
2643 			if (pv->pv_flags & PVF_WRITE)
2644 				md->krw_mappings--;
2645 			else
2646 				md->kro_mappings--;
2647 #endif
2648 			PMAPCOUNT(kernel_unmappings);
2649 		}
2650 		PMAPCOUNT(unmappings);
2651 
2652 		pmap_acquire_pmap_lock(pm);
2653 
2654 		l2b = pmap_get_l2_bucket(pm, pv->pv_va);
2655 		KDASSERT(l2b != NULL);
2656 
2657 		ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2658 
2659 		/*
2660 		 * Update statistics
2661 		 */
2662 		--pm->pm_stats.resident_count;
2663 
2664 		/* Wired bit */
2665 		if (pv->pv_flags & PVF_WIRED)
2666 			--pm->pm_stats.wired_count;
2667 
2668 		flags |= pv->pv_flags;
2669 
2670 		/*
2671 		 * Invalidate the PTEs.
2672 		 */
2673 		l2pte_reset(ptep);
2674 		PTE_SYNC_CURRENT(pm, ptep);
2675 		pmap_free_l2_bucket(pm, l2b, 1);
2676 
2677 		pool_put(&pmap_pv_pool, pv);
2678 		pv = npv;
2679 		/*
2680 		 * if we reach the end of the list and there are still
2681 		 * mappings, they might be able to be cached now.
2682 		 */
2683 		if (pv == NULL) {
2684 			*pvp = NULL;
2685 			if (!SLIST_EMPTY(&md->pvh_list))
2686 				pmap_vac_me_harder(md, pa, pm, 0);
2687 		}
2688 		pmap_release_pmap_lock(pm);
2689 	}
2690 #ifdef PMAP_CACHE_VIPT
2691 	/*
2692 	 * Its EXEC cache is now gone.
2693 	 */
2694 	if (PV_IS_EXEC_P(md->pvh_attrs))
2695 		PMAPCOUNT(exec_discarded_page_protect);
2696 	md->pvh_attrs &= ~PVF_EXEC;
2697 	KASSERT(md->urw_mappings == 0);
2698 	KASSERT(md->uro_mappings == 0);
2699 	if (arm_cache_prefer_mask != 0) {
2700 		if (md->krw_mappings == 0)
2701 			md->pvh_attrs &= ~PVF_WRITE;
2702 		PMAP_VALIDATE_MD_PAGE(md);
2703 	}
2704 #endif
2705 
2706 	if (flush) {
2707 		/*
2708 		 * Note: We can't use pmap_tlb_flush{I,D}() here since that
2709 		 * would need a subsequent call to pmap_update() to ensure
2710 		 * curpm->pm_cstate.cs_all is reset. Our callers are not
2711 		 * required to do that (see pmap(9)), so we can't modify
2712 		 * the current pmap's state.
2713 		 */
2714 		if (PV_BEEN_EXECD(flags))
2715 			cpu_tlb_flushID();
2716 		else
2717 			cpu_tlb_flushD();
2718 	}
2719 	cpu_cpwait();
2720 }
2721 
2722 /*
2723  * pmap_t pmap_create(void)
2724  *
2725  *      Create a new pmap structure from scratch.
2726  */
2727 pmap_t
2728 pmap_create(void)
2729 {
2730 	pmap_t pm;
2731 
2732 	pm = pool_cache_get(&pmap_cache, PR_WAITOK);
2733 
2734 	mutex_init(&pm->pm_obj_lock, MUTEX_DEFAULT, IPL_NONE);
2735 	uvm_obj_init(&pm->pm_obj, NULL, false, 1);
2736 	uvm_obj_setlock(&pm->pm_obj, &pm->pm_obj_lock);
2737 
2738 	pm->pm_stats.wired_count = 0;
2739 	pm->pm_stats.resident_count = 1;
2740 	pm->pm_cstate.cs_all = 0;
2741 	pmap_alloc_l1(pm);
2742 
2743 	/*
2744 	 * Note: The pool cache ensures that the pm_l2[] array is already
2745 	 * initialised to zero.
2746 	 */
2747 
2748 	pmap_pinit(pm);
2749 
2750 	LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list);
2751 
2752 	return (pm);
2753 }
2754 
2755 u_int
2756 arm32_mmap_flags(paddr_t pa)
2757 {
2758 	/*
2759 	 * the upper 8 bits in pmap_enter()'s flags are reserved for MD stuff
2760 	 * and we're using the upper bits in page numbers to pass flags around
2761 	 * so we might as well use the same bits
2762 	 */
2763 	return (u_int)pa & PMAP_MD_MASK;
2764 }
2765 /*
2766  * int pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
2767  *      u_int flags)
2768  *
2769  *      Insert the given physical page (p) at
2770  *      the specified virtual address (v) in the
2771  *      target physical map with the protection requested.
2772  *
2773  *      NB:  This is the only routine which MAY NOT lazy-evaluate
2774  *      or lose information.  That is, this routine must actually
2775  *      insert this page into the given map NOW.
2776  */
2777 int
2778 pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags)
2779 {
2780 	struct l2_bucket *l2b;
2781 	struct vm_page *pg, *opg;
2782 	struct pv_entry *pv;
2783 	u_int nflags;
2784 	u_int oflags;
2785 #ifdef ARM_HAS_VBAR
2786 	const bool vector_page_p = false;
2787 #else
2788 	const bool vector_page_p = (va == vector_page);
2789 #endif
2790 
2791 	NPDEBUG(PDB_ENTER, printf("pmap_enter: pm %p va 0x%lx pa 0x%lx prot %x flag %x\n", pm, va, pa, prot, flags));
2792 
2793 	KDASSERT((flags & PMAP_WIRED) == 0 || (flags & VM_PROT_ALL) != 0);
2794 	KDASSERT(((va | pa) & PGOFSET) == 0);
2795 
2796 	/*
2797 	 * Get a pointer to the page.  Later on in this function, we
2798 	 * test for a managed page by checking pg != NULL.
2799 	 */
2800 	pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
2801 
2802 	nflags = 0;
2803 	if (prot & VM_PROT_WRITE)
2804 		nflags |= PVF_WRITE;
2805 	if (prot & VM_PROT_EXECUTE)
2806 		nflags |= PVF_EXEC;
2807 	if (flags & PMAP_WIRED)
2808 		nflags |= PVF_WIRED;
2809 
2810 	pmap_acquire_pmap_lock(pm);
2811 
2812 	/*
2813 	 * Fetch the L2 bucket which maps this page, allocating one if
2814 	 * necessary for user pmaps.
2815 	 */
2816 	if (pm == pmap_kernel())
2817 		l2b = pmap_get_l2_bucket(pm, va);
2818 	else
2819 		l2b = pmap_alloc_l2_bucket(pm, va);
2820 	if (l2b == NULL) {
2821 		if (flags & PMAP_CANFAIL) {
2822 			pmap_release_pmap_lock(pm);
2823 			return (ENOMEM);
2824 		}
2825 		panic("pmap_enter: failed to allocate L2 bucket");
2826 	}
2827 	pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(va)];
2828 	const pt_entry_t opte = *ptep;
2829 	pt_entry_t npte = pa;
2830 	oflags = 0;
2831 
2832 	if (opte) {
2833 		/*
2834 		 * There is already a mapping at this address.
2835 		 * If the physical address is different, lookup the
2836 		 * vm_page.
2837 		 */
2838 		if (l2pte_pa(opte) != pa)
2839 			opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
2840 		else
2841 			opg = pg;
2842 	} else
2843 		opg = NULL;
2844 
2845 	if (pg) {
2846 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
2847 
2848 		/*
2849 		 * This is to be a managed mapping.
2850 		 */
2851 		if ((flags & VM_PROT_ALL) || (md->pvh_attrs & PVF_REF)) {
2852 			/*
2853 			 * - The access type indicates that we don't need
2854 			 *   to do referenced emulation.
2855 			 * OR
2856 			 * - The physical page has already been referenced
2857 			 *   so no need to re-do referenced emulation here.
2858 			 */
2859 			npte |= l2pte_set_readonly(L2_S_PROTO);
2860 
2861 			nflags |= PVF_REF;
2862 
2863 			if ((prot & VM_PROT_WRITE) != 0 &&
2864 			    ((flags & VM_PROT_WRITE) != 0 ||
2865 			     (md->pvh_attrs & PVF_MOD) != 0)) {
2866 				/*
2867 				 * This is a writable mapping, and the
2868 				 * page's mod state indicates it has
2869 				 * already been modified. Make it
2870 				 * writable from the outset.
2871 				 */
2872 				npte = l2pte_set_writable(npte);
2873 				nflags |= PVF_MOD;
2874 			}
2875 		} else {
2876 			/*
2877 			 * Need to do page referenced emulation.
2878 			 */
2879 			npte |= L2_TYPE_INV;
2880 		}
2881 
2882 		if (flags & ARM32_MMAP_WRITECOMBINE) {
2883 			npte |= pte_l2_s_wc_mode;
2884 		} else
2885 			npte |= pte_l2_s_cache_mode;
2886 
2887 		if (pg == opg) {
2888 			/*
2889 			 * We're changing the attrs of an existing mapping.
2890 			 */
2891 #ifdef MULTIPROCESSOR
2892 			KASSERT(uvm_page_locked_p(pg));
2893 #endif
2894 			oflags = pmap_modify_pv(md, pa, pm, va,
2895 			    PVF_WRITE | PVF_EXEC | PVF_WIRED |
2896 			    PVF_MOD | PVF_REF, nflags);
2897 
2898 #ifdef PMAP_CACHE_VIVT
2899 			/*
2900 			 * We may need to flush the cache if we're
2901 			 * doing rw-ro...
2902 			 */
2903 			if (pm->pm_cstate.cs_cache_d &&
2904 			    (oflags & PVF_NC) == 0 &&
2905 			    l2pte_writable_p(opte) &&
2906 			    (prot & VM_PROT_WRITE) == 0)
2907 				cpu_dcache_wb_range(va, PAGE_SIZE);
2908 #endif
2909 		} else {
2910 			/*
2911 			 * New mapping, or changing the backing page
2912 			 * of an existing mapping.
2913 			 */
2914 			if (opg) {
2915 				struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
2916 				paddr_t opa = VM_PAGE_TO_PHYS(opg);
2917 
2918 				/*
2919 				 * Replacing an existing mapping with a new one.
2920 				 * It is part of our managed memory so we
2921 				 * must remove it from the PV list
2922 				 */
2923 #ifdef MULTIPROCESSOR
2924 				KASSERT(uvm_page_locked_p(opg));
2925 #endif
2926 				pv = pmap_remove_pv(omd, opa, pm, va);
2927 				pmap_vac_me_harder(omd, opa, pm, 0);
2928 				oflags = pv->pv_flags;
2929 
2930 #ifdef PMAP_CACHE_VIVT
2931 				/*
2932 				 * If the old mapping was valid (ref/mod
2933 				 * emulation creates 'invalid' mappings
2934 				 * initially) then make sure to frob
2935 				 * the cache.
2936 				 */
2937 				if (!(oflags & PVF_NC) && l2pte_valid(opte)) {
2938 					pmap_cache_wbinv_page(pm, va, true,
2939 					    oflags);
2940 				}
2941 #endif
2942 			} else
2943 			if ((pv = pool_get(&pmap_pv_pool, PR_NOWAIT)) == NULL){
2944 				if ((flags & PMAP_CANFAIL) == 0)
2945 					panic("pmap_enter: no pv entries");
2946 
2947 				if (pm != pmap_kernel())
2948 					pmap_free_l2_bucket(pm, l2b, 0);
2949 				pmap_release_pmap_lock(pm);
2950 				NPDEBUG(PDB_ENTER,
2951 				    printf("pmap_enter: ENOMEM\n"));
2952 				return (ENOMEM);
2953 			}
2954 
2955 #ifdef MULTIPROCESSOR
2956 			KASSERT(uvm_page_locked_p(pg));
2957 #endif
2958 			pmap_enter_pv(md, pa, pv, pm, va, nflags);
2959 		}
2960 	} else {
2961 		/*
2962 		 * We're mapping an unmanaged page.
2963 		 * These are always readable, and possibly writable, from
2964 		 * the get go as we don't need to track ref/mod status.
2965 		 */
2966 		npte |= l2pte_set_readonly(L2_S_PROTO);
2967 		if (prot & VM_PROT_WRITE)
2968 			npte = l2pte_set_writable(npte);
2969 
2970 		/*
2971 		 * Make sure the vector table is mapped cacheable
2972 		 */
2973 		if ((vector_page_p && pm != pmap_kernel())
2974 		    || (flags & ARM32_MMAP_CACHEABLE)) {
2975 			npte |= pte_l2_s_cache_mode;
2976 		} else if (flags & ARM32_MMAP_WRITECOMBINE) {
2977 			npte |= pte_l2_s_wc_mode;
2978 		}
2979 		if (opg) {
2980 			/*
2981 			 * Looks like there's an existing 'managed' mapping
2982 			 * at this address.
2983 			 */
2984 			struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
2985 			paddr_t opa = VM_PAGE_TO_PHYS(opg);
2986 
2987 #ifdef MULTIPROCESSOR
2988 			KASSERT(uvm_page_locked_p(opg));
2989 #endif
2990 			pv = pmap_remove_pv(omd, opa, pm, va);
2991 			pmap_vac_me_harder(omd, opa, pm, 0);
2992 			oflags = pv->pv_flags;
2993 
2994 #ifdef PMAP_CACHE_VIVT
2995 			if (!(oflags & PVF_NC) && l2pte_valid(opte)) {
2996 				pmap_cache_wbinv_page(pm, va, true, oflags);
2997 			}
2998 #endif
2999 			pool_put(&pmap_pv_pool, pv);
3000 		}
3001 	}
3002 
3003 	/*
3004 	 * Make sure userland mappings get the right permissions
3005 	 */
3006 	if (!vector_page_p && pm != pmap_kernel()) {
3007 		npte |= L2_S_PROT_U;
3008 	}
3009 
3010 	/*
3011 	 * Keep the stats up to date
3012 	 */
3013 	if (opte == 0) {
3014 		l2b->l2b_occupancy++;
3015 		pm->pm_stats.resident_count++;
3016 	}
3017 
3018 	NPDEBUG(PDB_ENTER,
3019 	    printf("pmap_enter: opte 0x%08x npte 0x%08x\n", opte, npte));
3020 
3021 	/*
3022 	 * If this is just a wiring change, the two PTEs will be
3023 	 * identical, so there's no need to update the page table.
3024 	 */
3025 	if (npte != opte) {
3026 		bool is_cached = pmap_is_cached(pm);
3027 
3028 		l2pte_set(ptep, npte, opte);
3029 		PTE_SYNC(ptep);
3030 		if (is_cached) {
3031 			/*
3032 			 * We only need to frob the cache/tlb if this pmap
3033 			 * is current
3034 			 */
3035 			if (!vector_page_p && l2pte_valid(npte)) {
3036 				/*
3037 				 * This mapping is likely to be accessed as
3038 				 * soon as we return to userland. Fix up the
3039 				 * L1 entry to avoid taking another
3040 				 * page/domain fault.
3041 				 */
3042 				pd_entry_t *pl1pd, l1pd;
3043 
3044 				pl1pd = pmap_l1_kva(pm) + L1_IDX(va);
3045 				l1pd = l2b->l2b_phys | L1_C_DOM(pmap_domain(pm)) |
3046 				    L1_C_PROTO;
3047 				if (*pl1pd != l1pd) {
3048 					*pl1pd = l1pd;
3049 					PTE_SYNC(pl1pd);
3050 				}
3051 			}
3052 		}
3053 
3054 		pmap_tlb_flush_SE(pm, va, oflags);
3055 
3056 		NPDEBUG(PDB_ENTER,
3057 		    printf("pmap_enter: is_cached %d cs 0x%08x\n",
3058 		    is_cached, pm->pm_cstate.cs_all));
3059 
3060 		if (pg != NULL) {
3061 			struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3062 
3063 #ifdef MULTIPROCESSOR
3064 			KASSERT(uvm_page_locked_p(pg));
3065 #endif
3066 			pmap_vac_me_harder(md, pa, pm, va);
3067 		}
3068 	}
3069 #if defined(PMAP_CACHE_VIPT) && defined(DIAGNOSTIC)
3070 	if (pg) {
3071 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3072 
3073 #ifdef MULTIPROCESSOR
3074 		KASSERT(uvm_page_locked_p(pg));
3075 #endif
3076 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
3077 		PMAP_VALIDATE_MD_PAGE(md);
3078 	}
3079 #endif
3080 
3081 	pmap_release_pmap_lock(pm);
3082 
3083 	return (0);
3084 }
3085 
3086 /*
3087  * pmap_remove()
3088  *
3089  * pmap_remove is responsible for nuking a number of mappings for a range
3090  * of virtual address space in the current pmap. To do this efficiently
3091  * is interesting, because in a number of cases a wide virtual address
3092  * range may be supplied that contains few actual mappings. So, the
3093  * optimisations are:
3094  *  1. Skip over hunks of address space for which no L1 or L2 entry exists.
3095  *  2. Build up a list of pages we've hit, up to a maximum, so we can
3096  *     maybe do just a partial cache clean. This path of execution is
3097  *     complicated by the fact that the cache must be flushed _before_
3098  *     the PTE is nuked, being a VAC :-)
3099  *  3. If we're called after UVM calls pmap_remove_all(), we can defer
3100  *     all invalidations until pmap_update(), since pmap_remove_all() has
3101  *     already flushed the cache.
3102  *  4. Maybe later fast-case a single page, but I don't think this is
3103  *     going to make _that_ much difference overall.
3104  */
3105 
3106 #define	PMAP_REMOVE_CLEAN_LIST_SIZE	3
3107 
3108 void
3109 pmap_remove(pmap_t pm, vaddr_t sva, vaddr_t eva)
3110 {
3111 	vaddr_t next_bucket;
3112 	u_int cleanlist_idx, total, cnt;
3113 	struct {
3114 		vaddr_t va;
3115 		pt_entry_t *ptep;
3116 	} cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
3117 	u_int mappings;
3118 
3119 	NPDEBUG(PDB_REMOVE, printf("pmap_do_remove: pmap=%p sva=%08lx "
3120 	    "eva=%08lx\n", pm, sva, eva));
3121 
3122 	/*
3123 	 * we lock in the pmap => pv_head direction
3124 	 */
3125 	pmap_acquire_pmap_lock(pm);
3126 
3127 	if (pm->pm_remove_all || !pmap_is_cached(pm)) {
3128 		cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
3129 		if (pm->pm_cstate.cs_tlb == 0)
3130 			pm->pm_remove_all = true;
3131 	} else
3132 		cleanlist_idx = 0;
3133 
3134 	total = 0;
3135 
3136 	while (sva < eva) {
3137 		/*
3138 		 * Do one L2 bucket's worth at a time.
3139 		 */
3140 		next_bucket = L2_NEXT_BUCKET(sva);
3141 		if (next_bucket > eva)
3142 			next_bucket = eva;
3143 
3144 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pm, sva);
3145 		if (l2b == NULL) {
3146 			sva = next_bucket;
3147 			continue;
3148 		}
3149 
3150 		pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(sva)];
3151 
3152 		for (mappings = 0;
3153 		     sva < next_bucket;
3154 		     sva += PAGE_SIZE, ptep += PAGE_SIZE / L2_S_SIZE) {
3155 			pt_entry_t opte = *ptep;
3156 
3157 			if (opte == 0) {
3158 				/* Nothing here, move along */
3159 				continue;
3160 			}
3161 
3162 			u_int flags = PVF_REF;
3163 			paddr_t pa = l2pte_pa(opte);
3164 			struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
3165 
3166 			/*
3167 			 * Update flags. In a number of circumstances,
3168 			 * we could cluster a lot of these and do a
3169 			 * number of sequential pages in one go.
3170 			 */
3171 			if (pg != NULL) {
3172 				struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3173 				struct pv_entry *pv;
3174 
3175 #ifdef MULTIPROCESSOR
3176 				KASSERT(uvm_page_locked_p(pg));
3177 #endif
3178 				pv = pmap_remove_pv(md, pa, pm, sva);
3179 				pmap_vac_me_harder(md, pa, pm, 0);
3180 				if (pv != NULL) {
3181 					if (pm->pm_remove_all == false) {
3182 						flags = pv->pv_flags;
3183 					}
3184 					pool_put(&pmap_pv_pool, pv);
3185 				}
3186 			}
3187 			mappings++;
3188 
3189 			if (!l2pte_valid(opte)) {
3190 				/*
3191 				 * Ref/Mod emulation is still active for this
3192 				 * mapping, therefore it is has not yet been
3193 				 * accessed. No need to frob the cache/tlb.
3194 				 */
3195 				l2pte_reset(ptep);
3196 				PTE_SYNC_CURRENT(pm, ptep);
3197 				continue;
3198 			}
3199 
3200 			if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
3201 				/* Add to the clean list. */
3202 				cleanlist[cleanlist_idx].ptep = ptep;
3203 				cleanlist[cleanlist_idx].va =
3204 				    sva | (flags & PVF_EXEC);
3205 				cleanlist_idx++;
3206 			} else
3207 			if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
3208 				/* Nuke everything if needed. */
3209 #ifdef PMAP_CACHE_VIVT
3210 				pmap_cache_wbinv_all(pm, PVF_EXEC);
3211 #endif
3212 				pmap_tlb_flushID(pm);
3213 
3214 				/*
3215 				 * Roll back the previous PTE list,
3216 				 * and zero out the current PTE.
3217 				 */
3218 				for (cnt = 0;
3219 				     cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
3220 					l2pte_reset(cleanlist[cnt].ptep);
3221 					PTE_SYNC(cleanlist[cnt].ptep);
3222 				}
3223 				l2pte_reset(ptep);
3224 				PTE_SYNC(ptep);
3225 				cleanlist_idx++;
3226 				pm->pm_remove_all = true;
3227 			} else {
3228 				l2pte_reset(ptep);
3229 				PTE_SYNC(ptep);
3230 				if (pm->pm_remove_all == false) {
3231 					pmap_tlb_flush_SE(pm, sva, flags);
3232 				}
3233 			}
3234 		}
3235 
3236 		/*
3237 		 * Deal with any left overs
3238 		 */
3239 		if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
3240 			total += cleanlist_idx;
3241 			for (cnt = 0; cnt < cleanlist_idx; cnt++) {
3242 				vaddr_t va = cleanlist[cnt].va;
3243 				if (pm->pm_cstate.cs_all != 0) {
3244 					vaddr_t clva = va & ~PAGE_MASK;
3245 					u_int flags = va & PVF_EXEC;
3246 #ifdef PMAP_CACHE_VIVT
3247 					pmap_cache_wbinv_page(pm, clva, true,
3248 					    PVF_REF | PVF_WRITE | flags);
3249 #endif
3250 					pmap_tlb_flush_SE(pm, clva,
3251 					    PVF_REF | flags);
3252 				}
3253 				l2pte_reset(cleanlist[cnt].ptep);
3254 				PTE_SYNC_CURRENT(pm, cleanlist[cnt].ptep);
3255 			}
3256 
3257 			/*
3258 			 * If it looks like we're removing a whole bunch
3259 			 * of mappings, it's faster to just write-back
3260 			 * the whole cache now and defer TLB flushes until
3261 			 * pmap_update() is called.
3262 			 */
3263 			if (total <= PMAP_REMOVE_CLEAN_LIST_SIZE)
3264 				cleanlist_idx = 0;
3265 			else {
3266 				cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
3267 #ifdef PMAP_CACHE_VIVT
3268 				pmap_cache_wbinv_all(pm, PVF_EXEC);
3269 #endif
3270 				pm->pm_remove_all = true;
3271 			}
3272 		}
3273 
3274 		pmap_free_l2_bucket(pm, l2b, mappings);
3275 		pm->pm_stats.resident_count -= mappings;
3276 	}
3277 
3278 	pmap_release_pmap_lock(pm);
3279 }
3280 
3281 #ifdef PMAP_CACHE_VIPT
3282 static struct pv_entry *
3283 pmap_kremove_pg(struct vm_page *pg, vaddr_t va)
3284 {
3285 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3286 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
3287 	struct pv_entry *pv;
3288 
3289 	KASSERT(arm_cache_prefer_mask == 0 || md->pvh_attrs & (PVF_COLORED|PVF_NC));
3290 	KASSERT((md->pvh_attrs & PVF_KMPAGE) == 0);
3291 
3292 	pv = pmap_remove_pv(md, pa, pmap_kernel(), va);
3293 	KASSERT(pv);
3294 	KASSERT(pv->pv_flags & PVF_KENTRY);
3295 
3296 	/*
3297 	 * If we are removing a writeable mapping to a cached exec page,
3298 	 * if it's the last mapping then clear it execness other sync
3299 	 * the page to the icache.
3300 	 */
3301 	if ((md->pvh_attrs & (PVF_NC|PVF_EXEC)) == PVF_EXEC
3302 	    && (pv->pv_flags & PVF_WRITE) != 0) {
3303 		if (SLIST_EMPTY(&md->pvh_list)) {
3304 			md->pvh_attrs &= ~PVF_EXEC;
3305 			PMAPCOUNT(exec_discarded_kremove);
3306 		} else {
3307 			pmap_syncicache_page(md, pa);
3308 			PMAPCOUNT(exec_synced_kremove);
3309 		}
3310 	}
3311 	pmap_vac_me_harder(md, pa, pmap_kernel(), 0);
3312 
3313 	return pv;
3314 }
3315 #endif /* PMAP_CACHE_VIPT */
3316 
3317 /*
3318  * pmap_kenter_pa: enter an unmanaged, wired kernel mapping
3319  *
3320  * We assume there is already sufficient KVM space available
3321  * to do this, as we can't allocate L2 descriptor tables/metadata
3322  * from here.
3323  */
3324 void
3325 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags)
3326 {
3327 #ifdef PMAP_CACHE_VIVT
3328 	struct vm_page *pg = (flags & PMAP_KMPAGE) ? PHYS_TO_VM_PAGE(pa) : NULL;
3329 #endif
3330 #ifdef PMAP_CACHE_VIPT
3331 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
3332 	struct vm_page *opg;
3333 	struct pv_entry *pv = NULL;
3334 #endif
3335 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3336 
3337 	NPDEBUG(PDB_KENTER,
3338 	    printf("pmap_kenter_pa: va 0x%08lx, pa 0x%08lx, prot 0x%x\n",
3339 	    va, pa, prot));
3340 
3341 	struct l2_bucket * const l2b = pmap_get_l2_bucket(pmap_kernel(), va);
3342 	KDASSERT(l2b != NULL);
3343 
3344 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
3345 	const pt_entry_t opte = *ptep;
3346 
3347 	if (opte == 0) {
3348 		PMAPCOUNT(kenter_mappings);
3349 		l2b->l2b_occupancy++;
3350 	} else {
3351 		PMAPCOUNT(kenter_remappings);
3352 #ifdef PMAP_CACHE_VIPT
3353 		opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3354 #ifdef DIAGNOSTIC
3355 		struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3356 #endif
3357 		if (opg) {
3358 			KASSERT(opg != pg);
3359 			KASSERT((omd->pvh_attrs & PVF_KMPAGE) == 0);
3360 			KASSERT((flags & PMAP_KMPAGE) == 0);
3361 			pv = pmap_kremove_pg(opg, va);
3362 		}
3363 #endif
3364 		if (l2pte_valid(opte)) {
3365 #ifdef PMAP_CACHE_VIVT
3366 			cpu_dcache_wbinv_range(va, PAGE_SIZE);
3367 #endif
3368 			cpu_tlb_flushD_SE(va);
3369 			cpu_cpwait();
3370 		}
3371 	}
3372 
3373 	const pt_entry_t npte = L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot)
3374 	    | ((flags & PMAP_NOCACHE) ? 0 : pte_l2_s_cache_mode);
3375 	l2pte_set(ptep, npte, opte);
3376 	PTE_SYNC(ptep);
3377 
3378 	if (pg) {
3379 #ifdef MULTIPROCESSOR
3380 		KASSERT(uvm_page_locked_p(pg));
3381 #endif
3382 		if (flags & PMAP_KMPAGE) {
3383 			KASSERT(md->urw_mappings == 0);
3384 			KASSERT(md->uro_mappings == 0);
3385 			KASSERT(md->krw_mappings == 0);
3386 			KASSERT(md->kro_mappings == 0);
3387 #ifdef PMAP_CACHE_VIPT
3388 			KASSERT(pv == NULL);
3389 			KASSERT(arm_cache_prefer_mask == 0 || (va & PVF_COLORED) == 0);
3390 			KASSERT((md->pvh_attrs & PVF_NC) == 0);
3391 			/* if there is a color conflict, evict from cache. */
3392 			if (pmap_is_page_colored_p(md)
3393 			    && ((va ^ md->pvh_attrs) & arm_cache_prefer_mask)) {
3394 				PMAPCOUNT(vac_color_change);
3395 				pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
3396 			} else if (md->pvh_attrs & PVF_MULTCLR) {
3397 				/*
3398 				 * If this page has multiple colors, expunge
3399 				 * them.
3400 				 */
3401 				PMAPCOUNT(vac_flush_lots2);
3402 				pmap_flush_page(md, pa, PMAP_FLUSH_SECONDARY);
3403 			}
3404 			md->pvh_attrs &= PAGE_SIZE - 1;
3405 			md->pvh_attrs |= PVF_KMPAGE
3406 			    | PVF_COLORED | PVF_DIRTY
3407 			    | (va & arm_cache_prefer_mask);
3408 #endif
3409 #ifdef PMAP_CACHE_VIVT
3410 			md->pvh_attrs |= PVF_KMPAGE;
3411 #endif
3412 			pmap_kmpages++;
3413 #ifdef PMAP_CACHE_VIPT
3414 		} else {
3415 			if (pv == NULL) {
3416 				pv = pool_get(&pmap_pv_pool, PR_NOWAIT);
3417 				KASSERT(pv != NULL);
3418 			}
3419 			pmap_enter_pv(md, pa, pv, pmap_kernel(), va,
3420 			    PVF_WIRED | PVF_KENTRY
3421 			    | (prot & VM_PROT_WRITE ? PVF_WRITE : 0));
3422 			if ((prot & VM_PROT_WRITE)
3423 			    && !(md->pvh_attrs & PVF_NC))
3424 				md->pvh_attrs |= PVF_DIRTY;
3425 			KASSERT((prot & VM_PROT_WRITE) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
3426 			pmap_vac_me_harder(md, pa, pmap_kernel(), va);
3427 #endif
3428 		}
3429 #ifdef PMAP_CACHE_VIPT
3430 	} else {
3431 		if (pv != NULL)
3432 			pool_put(&pmap_pv_pool, pv);
3433 #endif
3434 	}
3435 }
3436 
3437 void
3438 pmap_kremove(vaddr_t va, vsize_t len)
3439 {
3440 	vaddr_t next_bucket, eva;
3441 	u_int mappings;
3442 
3443 	PMAPCOUNT(kenter_unmappings);
3444 
3445 	NPDEBUG(PDB_KREMOVE, printf("pmap_kremove: va 0x%08lx, len 0x%08lx\n",
3446 	    va, len));
3447 
3448 	eva = va + len;
3449 
3450 	while (va < eva) {
3451 		next_bucket = L2_NEXT_BUCKET(va);
3452 		if (next_bucket > eva)
3453 			next_bucket = eva;
3454 
3455 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pmap_kernel(), va);
3456 		KDASSERT(l2b != NULL);
3457 
3458 		pt_entry_t * const sptep = &l2b->l2b_kva[l2pte_index(va)];
3459 		pt_entry_t *ptep = sptep;
3460 		mappings = 0;
3461 
3462 		while (va < next_bucket) {
3463 			const pt_entry_t opte = *ptep;
3464 			struct vm_page *opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3465 			if (opg != NULL) {
3466 				struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3467 
3468 				if (omd->pvh_attrs & PVF_KMPAGE) {
3469 					KASSERT(omd->urw_mappings == 0);
3470 					KASSERT(omd->uro_mappings == 0);
3471 					KASSERT(omd->krw_mappings == 0);
3472 					KASSERT(omd->kro_mappings == 0);
3473 					omd->pvh_attrs &= ~PVF_KMPAGE;
3474 #ifdef PMAP_CACHE_VIPT
3475 					if (arm_cache_prefer_mask != 0) {
3476 						omd->pvh_attrs &= ~PVF_WRITE;
3477 					}
3478 #endif
3479 					pmap_kmpages--;
3480 #ifdef PMAP_CACHE_VIPT
3481 				} else {
3482 					pool_put(&pmap_pv_pool,
3483 					    pmap_kremove_pg(opg, va));
3484 #endif
3485 				}
3486 			}
3487 			if (l2pte_valid(opte)) {
3488 #ifdef PMAP_CACHE_VIVT
3489 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
3490 #endif
3491 				cpu_tlb_flushD_SE(va);
3492 			}
3493 			if (opte) {
3494 				l2pte_reset(ptep);
3495 				mappings++;
3496 			}
3497 			va += PAGE_SIZE;
3498 			ptep += PAGE_SIZE / L2_S_SIZE;
3499 		}
3500 		KDASSERT(mappings <= l2b->l2b_occupancy);
3501 		l2b->l2b_occupancy -= mappings;
3502 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
3503 	}
3504 	cpu_cpwait();
3505 }
3506 
3507 bool
3508 pmap_extract(pmap_t pm, vaddr_t va, paddr_t *pap)
3509 {
3510 	struct l2_dtable *l2;
3511 	pd_entry_t *pl1pd, l1pd;
3512 	pt_entry_t *ptep, pte;
3513 	paddr_t pa;
3514 	u_int l1idx;
3515 
3516 	pmap_acquire_pmap_lock(pm);
3517 
3518 	l1idx = L1_IDX(va);
3519 	pl1pd = pmap_l1_kva(pm) + l1idx;
3520 	l1pd = *pl1pd;
3521 
3522 	if (l1pte_section_p(l1pd)) {
3523 		/*
3524 		 * These should only happen for pmap_kernel()
3525 		 */
3526 		KDASSERT(pm == pmap_kernel());
3527 		pmap_release_pmap_lock(pm);
3528 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
3529 		if (l1pte_supersection_p(l1pd)) {
3530 			pa = (l1pd & L1_SS_FRAME) | (va & L1_SS_OFFSET);
3531 		} else
3532 #endif
3533 			pa = (l1pd & L1_S_FRAME) | (va & L1_S_OFFSET);
3534 	} else {
3535 		/*
3536 		 * Note that we can't rely on the validity of the L1
3537 		 * descriptor as an indication that a mapping exists.
3538 		 * We have to look it up in the L2 dtable.
3539 		 */
3540 		l2 = pm->pm_l2[L2_IDX(l1idx)];
3541 
3542 		if (l2 == NULL ||
3543 		    (ptep = l2->l2_bucket[L2_BUCKET(l1idx)].l2b_kva) == NULL) {
3544 			pmap_release_pmap_lock(pm);
3545 			return false;
3546 		}
3547 
3548 		ptep = &ptep[l2pte_index(va)];
3549 		pte = *ptep;
3550 		pmap_release_pmap_lock(pm);
3551 
3552 		if (pte == 0)
3553 			return false;
3554 
3555 		switch (pte & L2_TYPE_MASK) {
3556 		case L2_TYPE_L:
3557 			pa = (pte & L2_L_FRAME) | (va & L2_L_OFFSET);
3558 			break;
3559 
3560 		default:
3561 			pa = (pte & L2_S_FRAME) | (va & L2_S_OFFSET);
3562 			break;
3563 		}
3564 	}
3565 
3566 	if (pap != NULL)
3567 		*pap = pa;
3568 
3569 	return true;
3570 }
3571 
3572 void
3573 pmap_protect(pmap_t pm, vaddr_t sva, vaddr_t eva, vm_prot_t prot)
3574 {
3575 	struct l2_bucket *l2b;
3576 	pt_entry_t *ptep, pte;
3577 	vaddr_t next_bucket;
3578 
3579 	NPDEBUG(PDB_PROTECT,
3580 	    printf("pmap_protect: pm %p sva 0x%lx eva 0x%lx prot 0x%x\n",
3581 	    pm, sva, eva, prot));
3582 
3583 	if ((prot & VM_PROT_READ) == 0) {
3584 		pmap_remove(pm, sva, eva);
3585 		return;
3586 	}
3587 
3588 	if (prot & VM_PROT_WRITE) {
3589 		/*
3590 		 * If this is a read->write transition, just ignore it and let
3591 		 * uvm_fault() take care of it later.
3592 		 */
3593 		return;
3594 	}
3595 
3596 	pmap_acquire_pmap_lock(pm);
3597 
3598 	const bool flush = eva - sva >= PAGE_SIZE * 4;
3599 	u_int clr_mask = PVF_WRITE | ((prot & VM_PROT_EXECUTE) ? 0 : PVF_EXEC);
3600 	u_int flags = 0;
3601 
3602 	while (sva < eva) {
3603 		next_bucket = L2_NEXT_BUCKET(sva);
3604 		if (next_bucket > eva)
3605 			next_bucket = eva;
3606 
3607 		l2b = pmap_get_l2_bucket(pm, sva);
3608 		if (l2b == NULL) {
3609 			sva = next_bucket;
3610 			continue;
3611 		}
3612 
3613 		ptep = &l2b->l2b_kva[l2pte_index(sva)];
3614 
3615 		while (sva < next_bucket) {
3616 			pte = *ptep;
3617 			if (l2pte_valid(pte) != 0 && l2pte_writable_p(pte)) {
3618 				struct vm_page *pg;
3619 				u_int f;
3620 
3621 #ifdef PMAP_CACHE_VIVT
3622 				/*
3623 				 * OK, at this point, we know we're doing
3624 				 * write-protect operation.  If the pmap is
3625 				 * active, write-back the page.
3626 				 */
3627 				pmap_cache_wbinv_page(pm, sva, false,
3628 				    PVF_REF | PVF_WRITE);
3629 #endif
3630 
3631 				pg = PHYS_TO_VM_PAGE(l2pte_pa(pte));
3632 				pte = l2pte_set_readonly(pte);
3633 				*ptep = pte;
3634 				PTE_SYNC(ptep);
3635 
3636 				if (pg != NULL) {
3637 					struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3638 					paddr_t pa = VM_PAGE_TO_PHYS(pg);
3639 
3640 #ifdef MULTIPROCESSOR
3641 					KASSERT(uvm_page_locked_p(pg));
3642 #endif
3643 					f = pmap_modify_pv(md, pa, pm, sva,
3644 					    clr_mask, 0);
3645 					pmap_vac_me_harder(md, pa, pm, sva);
3646 				} else {
3647 					f = PVF_REF | PVF_EXEC;
3648 				}
3649 
3650 				if (flush) {
3651 					flags |= f;
3652 				} else {
3653 					pmap_tlb_flush_SE(pm, sva, f);
3654 				}
3655 			}
3656 
3657 			sva += PAGE_SIZE;
3658 			ptep++;
3659 		}
3660 	}
3661 
3662 	if (flush) {
3663 		if (PV_BEEN_EXECD(flags)) {
3664 			pmap_tlb_flushID(pm);
3665 		} else if (PV_BEEN_REFD(flags)) {
3666 			pmap_tlb_flushD(pm);
3667 		}
3668 	}
3669 
3670 	pmap_release_pmap_lock(pm);
3671 }
3672 
3673 void
3674 pmap_icache_sync_range(pmap_t pm, vaddr_t sva, vaddr_t eva)
3675 {
3676 	struct l2_bucket *l2b;
3677 	pt_entry_t *ptep;
3678 	vaddr_t next_bucket;
3679 	vsize_t page_size = trunc_page(sva) + PAGE_SIZE - sva;
3680 
3681 	NPDEBUG(PDB_EXEC,
3682 	    printf("pmap_icache_sync_range: pm %p sva 0x%lx eva 0x%lx\n",
3683 	    pm, sva, eva));
3684 
3685 	pmap_acquire_pmap_lock(pm);
3686 
3687 	while (sva < eva) {
3688 		next_bucket = L2_NEXT_BUCKET(sva);
3689 		if (next_bucket > eva)
3690 			next_bucket = eva;
3691 
3692 		l2b = pmap_get_l2_bucket(pm, sva);
3693 		if (l2b == NULL) {
3694 			sva = next_bucket;
3695 			continue;
3696 		}
3697 
3698 		for (ptep = &l2b->l2b_kva[l2pte_index(sva)];
3699 		     sva < next_bucket;
3700 		     sva += page_size, ptep++, page_size = PAGE_SIZE) {
3701 			if (l2pte_valid(*ptep)) {
3702 				cpu_icache_sync_range(sva,
3703 				    min(page_size, eva - sva));
3704 			}
3705 		}
3706 	}
3707 
3708 	pmap_release_pmap_lock(pm);
3709 }
3710 
3711 void
3712 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
3713 {
3714 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3715 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
3716 
3717 	NPDEBUG(PDB_PROTECT,
3718 	    printf("pmap_page_protect: md %p (0x%08lx), prot 0x%x\n",
3719 	    md, pa, prot));
3720 
3721 #ifdef MULTIPROCESSOR
3722 	KASSERT(uvm_page_locked_p(pg));
3723 #endif
3724 
3725 	switch(prot) {
3726 	case VM_PROT_READ|VM_PROT_WRITE:
3727 #if defined(PMAP_CHECK_VIPT) && defined(PMAP_APX)
3728 		pmap_clearbit(md, pa, PVF_EXEC);
3729 		break;
3730 #endif
3731 	case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
3732 		break;
3733 
3734 	case VM_PROT_READ:
3735 #if defined(PMAP_CHECK_VIPT) && defined(PMAP_APX)
3736 		pmap_clearbit(md, pa, PVF_WRITE|PVF_EXEC);
3737 		break;
3738 #endif
3739 	case VM_PROT_READ|VM_PROT_EXECUTE:
3740 		pmap_clearbit(md, pa, PVF_WRITE);
3741 		break;
3742 
3743 	default:
3744 		pmap_page_remove(md, pa);
3745 		break;
3746 	}
3747 }
3748 
3749 /*
3750  * pmap_clear_modify:
3751  *
3752  *	Clear the "modified" attribute for a page.
3753  */
3754 bool
3755 pmap_clear_modify(struct vm_page *pg)
3756 {
3757 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3758 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
3759 	bool rv;
3760 
3761 #ifdef MULTIPROCESSOR
3762 	KASSERT(uvm_page_locked_p(pg));
3763 #endif
3764 
3765 	if (md->pvh_attrs & PVF_MOD) {
3766 		rv = true;
3767 #ifdef PMAP_CACHE_VIPT
3768 		/*
3769 		 * If we are going to clear the modified bit and there are
3770 		 * no other modified bits set, flush the page to memory and
3771 		 * mark it clean.
3772 		 */
3773 		if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) == PVF_MOD)
3774 			pmap_flush_page(md, pa, PMAP_CLEAN_PRIMARY);
3775 #endif
3776 		pmap_clearbit(md, pa, PVF_MOD);
3777 	} else
3778 		rv = false;
3779 
3780 	return (rv);
3781 }
3782 
3783 /*
3784  * pmap_clear_reference:
3785  *
3786  *	Clear the "referenced" attribute for a page.
3787  */
3788 bool
3789 pmap_clear_reference(struct vm_page *pg)
3790 {
3791 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3792 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
3793 	bool rv;
3794 
3795 #ifdef MULTIPROCESSOR
3796 	KASSERT(uvm_page_locked_p(pg));
3797 #endif
3798 
3799 	if (md->pvh_attrs & PVF_REF) {
3800 		rv = true;
3801 		pmap_clearbit(md, pa, PVF_REF);
3802 	} else
3803 		rv = false;
3804 
3805 	return (rv);
3806 }
3807 
3808 /*
3809  * pmap_is_modified:
3810  *
3811  *	Test if a page has the "modified" attribute.
3812  */
3813 /* See <arm/arm32/pmap.h> */
3814 
3815 /*
3816  * pmap_is_referenced:
3817  *
3818  *	Test if a page has the "referenced" attribute.
3819  */
3820 /* See <arm/arm32/pmap.h> */
3821 
3822 int
3823 pmap_fault_fixup(pmap_t pm, vaddr_t va, vm_prot_t ftype, int user)
3824 {
3825 	struct l2_dtable *l2;
3826 	struct l2_bucket *l2b;
3827 	pd_entry_t *pl1pd, l1pd;
3828 	pt_entry_t *ptep, pte;
3829 	paddr_t pa;
3830 	u_int l1idx;
3831 	int rv = 0;
3832 
3833 	pmap_acquire_pmap_lock(pm);
3834 
3835 	l1idx = L1_IDX(va);
3836 
3837 	/*
3838 	 * If there is no l2_dtable for this address, then the process
3839 	 * has no business accessing it.
3840 	 *
3841 	 * Note: This will catch userland processes trying to access
3842 	 * kernel addresses.
3843 	 */
3844 	l2 = pm->pm_l2[L2_IDX(l1idx)];
3845 	if (l2 == NULL)
3846 		goto out;
3847 
3848 	/*
3849 	 * Likewise if there is no L2 descriptor table
3850 	 */
3851 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
3852 	if (l2b->l2b_kva == NULL)
3853 		goto out;
3854 
3855 	/*
3856 	 * Check the PTE itself.
3857 	 */
3858 	ptep = &l2b->l2b_kva[l2pte_index(va)];
3859 	pte = *ptep;
3860 	if (pte == 0)
3861 		goto out;
3862 
3863 	/*
3864 	 * Catch a userland access to the vector page mapped at 0x0
3865 	 */
3866 	if (user && (pte & L2_S_PROT_U) == 0)
3867 		goto out;
3868 
3869 	pa = l2pte_pa(pte);
3870 
3871 	if ((ftype & VM_PROT_WRITE) && !l2pte_writable_p(pte)) {
3872 		/*
3873 		 * This looks like a good candidate for "page modified"
3874 		 * emulation...
3875 		 */
3876 		struct pv_entry *pv;
3877 		struct vm_page *pg;
3878 
3879 		/* Extract the physical address of the page */
3880 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3881 			goto out;
3882 
3883 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3884 
3885 		/* Get the current flags for this page. */
3886 #ifdef MULTIPROCESSOR
3887 		KASSERT(uvm_page_locked_p(pg));
3888 #endif
3889 
3890 		pv = pmap_find_pv(md, pm, va);
3891 		if (pv == NULL) {
3892 			goto out;
3893 		}
3894 
3895 		/*
3896 		 * Do the flags say this page is writable? If not then it
3897 		 * is a genuine write fault. If yes then the write fault is
3898 		 * our fault as we did not reflect the write access in the
3899 		 * PTE. Now we know a write has occurred we can correct this
3900 		 * and also set the modified bit
3901 		 */
3902 		if ((pv->pv_flags & PVF_WRITE) == 0) {
3903 			goto out;
3904 		}
3905 
3906 		NPDEBUG(PDB_FOLLOW,
3907 		    printf("pmap_fault_fixup: mod emul. pm %p, va 0x%08lx, pa 0x%08lx\n",
3908 		    pm, va, pa));
3909 
3910 		md->pvh_attrs |= PVF_REF | PVF_MOD;
3911 		pv->pv_flags |= PVF_REF | PVF_MOD;
3912 #ifdef PMAP_CACHE_VIPT
3913 		/*
3914 		 * If there are cacheable mappings for this page, mark it dirty.
3915 		 */
3916 		if ((md->pvh_attrs & PVF_NC) == 0)
3917 			md->pvh_attrs |= PVF_DIRTY;
3918 #endif
3919 
3920 		/*
3921 		 * Re-enable write permissions for the page.  No need to call
3922 		 * pmap_vac_me_harder(), since this is just a
3923 		 * modified-emulation fault, and the PVF_WRITE bit isn't
3924 		 * changing. We've already set the cacheable bits based on
3925 		 * the assumption that we can write to this page.
3926 		 */
3927 		*ptep = l2pte_set_writable((pte & ~L2_TYPE_MASK) | L2_S_PROTO);
3928 		PTE_SYNC(ptep);
3929 		rv = 1;
3930 	} else
3931 	if ((pte & L2_TYPE_MASK) == L2_TYPE_INV) {
3932 		/*
3933 		 * This looks like a good candidate for "page referenced"
3934 		 * emulation.
3935 		 */
3936 		struct pv_entry *pv;
3937 		struct vm_page *pg;
3938 
3939 		/* Extract the physical address of the page */
3940 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
3941 			goto out;
3942 
3943 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3944 
3945 		/* Get the current flags for this page. */
3946 #ifdef MULTIPROCESSOR
3947 		KASSERT(uvm_page_locked_p(pg));
3948 #endif
3949 
3950 		pv = pmap_find_pv(md, pm, va);
3951 		if (pv == NULL) {
3952 			goto out;
3953 		}
3954 
3955 		md->pvh_attrs |= PVF_REF;
3956 		pv->pv_flags |= PVF_REF;
3957 
3958 		NPDEBUG(PDB_FOLLOW,
3959 		    printf("pmap_fault_fixup: ref emul. pm %p, va 0x%08lx, pa 0x%08lx\n",
3960 		    pm, va, pa));
3961 
3962 		*ptep = l2pte_set_readonly((pte & ~L2_TYPE_MASK) | L2_S_PROTO);
3963 		PTE_SYNC(ptep);
3964 		rv = 1;
3965 	}
3966 
3967 	/*
3968 	 * We know there is a valid mapping here, so simply
3969 	 * fix up the L1 if necessary.
3970 	 */
3971 	pl1pd = pmap_l1_kva(pm) + l1idx;
3972 	l1pd = l2b->l2b_phys | L1_C_DOM(pmap_domain(pm)) | L1_C_PROTO;
3973 	if (*pl1pd != l1pd) {
3974 		*pl1pd = l1pd;
3975 		PTE_SYNC(pl1pd);
3976 		rv = 1;
3977 	}
3978 
3979 #ifdef CPU_SA110
3980 	/*
3981 	 * There are bugs in the rev K SA110.  This is a check for one
3982 	 * of them.
3983 	 */
3984 	if (rv == 0 && curcpu()->ci_arm_cputype == CPU_ID_SA110 &&
3985 	    curcpu()->ci_arm_cpurev < 3) {
3986 		/* Always current pmap */
3987 		if (l2pte_valid(pte)) {
3988 			extern int kernel_debug;
3989 			if (kernel_debug & 1) {
3990 				struct proc *p = curlwp->l_proc;
3991 				printf("prefetch_abort: page is already "
3992 				    "mapped - pte=%p *pte=%08x\n", ptep, pte);
3993 				printf("prefetch_abort: pc=%08lx proc=%p "
3994 				    "process=%s\n", va, p, p->p_comm);
3995 				printf("prefetch_abort: far=%08x fs=%x\n",
3996 				    cpu_faultaddress(), cpu_faultstatus());
3997 			}
3998 #ifdef DDB
3999 			if (kernel_debug & 2)
4000 				Debugger();
4001 #endif
4002 			rv = 1;
4003 		}
4004 	}
4005 #endif /* CPU_SA110 */
4006 
4007 	/*
4008 	 * If 'rv == 0' at this point, it generally indicates that there is a
4009 	 * stale TLB entry for the faulting address.  That might be due to a
4010 	 * wrong setting of pmap_needs_pte_sync.  So set it and retry.
4011 	 */
4012 	if (rv == 0 && pm->pm_l1->l1_domain_use_count == 1
4013 	    && pmap_needs_pte_sync == 0) {
4014 		pmap_needs_pte_sync = 1;
4015 		PTE_SYNC(ptep);
4016 		rv = 1;
4017 	}
4018 
4019 #ifdef DEBUG
4020 	/*
4021 	 * If 'rv == 0' at this point, it generally indicates that there is a
4022 	 * stale TLB entry for the faulting address. This happens when two or
4023 	 * more processes are sharing an L1. Since we don't flush the TLB on
4024 	 * a context switch between such processes, we can take domain faults
4025 	 * for mappings which exist at the same VA in both processes. EVEN IF
4026 	 * WE'VE RECENTLY FIXED UP THE CORRESPONDING L1 in pmap_enter(), for
4027 	 * example.
4028 	 *
4029 	 * This is extremely likely to happen if pmap_enter() updated the L1
4030 	 * entry for a recently entered mapping. In this case, the TLB is
4031 	 * flushed for the new mapping, but there may still be TLB entries for
4032 	 * other mappings belonging to other processes in the 1MB range
4033 	 * covered by the L1 entry.
4034 	 *
4035 	 * Since 'rv == 0', we know that the L1 already contains the correct
4036 	 * value, so the fault must be due to a stale TLB entry.
4037 	 *
4038 	 * Since we always need to flush the TLB anyway in the case where we
4039 	 * fixed up the L1, or frobbed the L2 PTE, we effectively deal with
4040 	 * stale TLB entries dynamically.
4041 	 *
4042 	 * However, the above condition can ONLY happen if the current L1 is
4043 	 * being shared. If it happens when the L1 is unshared, it indicates
4044 	 * that other parts of the pmap are not doing their job WRT managing
4045 	 * the TLB.
4046 	 */
4047 	if (rv == 0 && pm->pm_l1->l1_domain_use_count == 1) {
4048 		extern int last_fault_code;
4049 		printf("fixup: pm %p, va 0x%lx, ftype %d - nothing to do!\n",
4050 		    pm, va, ftype);
4051 		printf("fixup: l2 %p, l2b %p, ptep %p, pl1pd %p\n",
4052 		    l2, l2b, ptep, pl1pd);
4053 		printf("fixup: pte 0x%x, l1pd 0x%x, last code 0x%x\n",
4054 		    pte, l1pd, last_fault_code);
4055 #ifdef DDB
4056 		extern int kernel_debug;
4057 
4058 		if (kernel_debug & 2)
4059 			Debugger();
4060 #endif
4061 	}
4062 #endif
4063 
4064 	cpu_tlb_flushID_SE(va);
4065 	cpu_cpwait();
4066 
4067 	rv = 1;
4068 
4069 out:
4070 	pmap_release_pmap_lock(pm);
4071 
4072 	return (rv);
4073 }
4074 
4075 /*
4076  * Routine:	pmap_procwr
4077  *
4078  * Function:
4079  *	Synchronize caches corresponding to [addr, addr+len) in p.
4080  *
4081  */
4082 void
4083 pmap_procwr(struct proc *p, vaddr_t va, int len)
4084 {
4085 	/* We only need to do anything if it is the current process. */
4086 	if (p == curproc)
4087 		cpu_icache_sync_range(va, len);
4088 }
4089 
4090 /*
4091  * Routine:	pmap_unwire
4092  * Function:	Clear the wired attribute for a map/virtual-address pair.
4093  *
4094  * In/out conditions:
4095  *		The mapping must already exist in the pmap.
4096  */
4097 void
4098 pmap_unwire(pmap_t pm, vaddr_t va)
4099 {
4100 	struct l2_bucket *l2b;
4101 	pt_entry_t *ptep, pte;
4102 	struct vm_page *pg;
4103 	paddr_t pa;
4104 
4105 	NPDEBUG(PDB_WIRING, printf("pmap_unwire: pm %p, va 0x%08lx\n", pm, va));
4106 
4107 	pmap_acquire_pmap_lock(pm);
4108 
4109 	l2b = pmap_get_l2_bucket(pm, va);
4110 	KDASSERT(l2b != NULL);
4111 
4112 	ptep = &l2b->l2b_kva[l2pte_index(va)];
4113 	pte = *ptep;
4114 
4115 	/* Extract the physical address of the page */
4116 	pa = l2pte_pa(pte);
4117 
4118 	if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
4119 		/* Update the wired bit in the pv entry for this page. */
4120 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4121 
4122 #ifdef MULTIPROCESSOR
4123 		KASSERT(uvm_page_locked_p(pg));
4124 #endif
4125 		(void) pmap_modify_pv(md, pa, pm, va, PVF_WIRED, 0);
4126 	}
4127 
4128 	pmap_release_pmap_lock(pm);
4129 }
4130 
4131 void
4132 pmap_activate(struct lwp *l)
4133 {
4134 	extern int block_userspace_access;
4135 	pmap_t opm, npm, rpm;
4136 	uint32_t odacr, ndacr;
4137 	int oldirqstate;
4138 
4139 	/*
4140 	 * If activating a non-current lwp or the current lwp is
4141 	 * already active, just return.
4142 	 */
4143 	if (l != curlwp ||
4144 	    l->l_proc->p_vmspace->vm_map.pmap->pm_activated == true)
4145 		return;
4146 
4147 	npm = l->l_proc->p_vmspace->vm_map.pmap;
4148 	ndacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2)) |
4149 	    (DOMAIN_CLIENT << (pmap_domain(npm) * 2));
4150 
4151 	/*
4152 	 * If TTB and DACR are unchanged, short-circuit all the
4153 	 * TLB/cache management stuff.
4154 	 */
4155 	if (pmap_previous_active_lwp != NULL) {
4156 		opm = pmap_previous_active_lwp->l_proc->p_vmspace->vm_map.pmap;
4157 		odacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2)) |
4158 		    (DOMAIN_CLIENT << (pmap_domain(opm) * 2));
4159 
4160 		if (opm->pm_l1 == npm->pm_l1 && odacr == ndacr)
4161 			goto all_done;
4162 	} else
4163 		opm = NULL;
4164 
4165 	PMAPCOUNT(activations);
4166 	block_userspace_access = 1;
4167 
4168 	/*
4169 	 * If switching to a user vmspace which is different to the
4170 	 * most recent one, and the most recent one is potentially
4171 	 * live in the cache, we must write-back and invalidate the
4172 	 * entire cache.
4173 	 */
4174 	rpm = pmap_recent_user;
4175 
4176 /*
4177  * XXXSCW: There's a corner case here which can leave turds in the cache as
4178  * reported in kern/41058. They're probably left over during tear-down and
4179  * switching away from an exiting process. Until the root cause is identified
4180  * and fixed, zap the cache when switching pmaps. This will result in a few
4181  * unnecessary cache flushes, but that's better than silently corrupting data.
4182  */
4183 #if 0
4184 	if (npm != pmap_kernel() && rpm && npm != rpm &&
4185 	    rpm->pm_cstate.cs_cache) {
4186 		rpm->pm_cstate.cs_cache = 0;
4187 #ifdef PMAP_CACHE_VIVT
4188 		cpu_idcache_wbinv_all();
4189 #endif
4190 	}
4191 #else
4192 	if (rpm) {
4193 		rpm->pm_cstate.cs_cache = 0;
4194 		if (npm == pmap_kernel())
4195 			pmap_recent_user = NULL;
4196 #ifdef PMAP_CACHE_VIVT
4197 		cpu_idcache_wbinv_all();
4198 #endif
4199 	}
4200 #endif
4201 
4202 	/* No interrupts while we frob the TTB/DACR */
4203 	oldirqstate = disable_interrupts(IF32_bits);
4204 
4205 #ifndef ARM_HAS_VBAR
4206 	/*
4207 	 * For ARM_VECTORS_LOW, we MUST, I repeat, MUST fix up the L1
4208 	 * entry corresponding to 'vector_page' in the incoming L1 table
4209 	 * before switching to it otherwise subsequent interrupts/exceptions
4210 	 * (including domain faults!) will jump into hyperspace.
4211 	 */
4212 	if (npm->pm_pl1vec != NULL) {
4213 		cpu_tlb_flushID_SE((u_int)vector_page);
4214 		cpu_cpwait();
4215 		*npm->pm_pl1vec = npm->pm_l1vec;
4216 		PTE_SYNC(npm->pm_pl1vec);
4217 	}
4218 #endif
4219 
4220 	cpu_domains(ndacr);
4221 
4222 	if (npm == pmap_kernel() || npm == rpm) {
4223 		/*
4224 		 * Switching to a kernel thread, or back to the
4225 		 * same user vmspace as before... Simply update
4226 		 * the TTB (no TLB flush required)
4227 		 */
4228 		cpu_setttb(npm->pm_l1->l1_physaddr, false);
4229 		cpu_cpwait();
4230 	} else {
4231 		/*
4232 		 * Otherwise, update TTB and flush TLB
4233 		 */
4234 		cpu_context_switch(npm->pm_l1->l1_physaddr);
4235 		if (rpm != NULL)
4236 			rpm->pm_cstate.cs_tlb = 0;
4237 	}
4238 
4239 	restore_interrupts(oldirqstate);
4240 
4241 	block_userspace_access = 0;
4242 
4243  all_done:
4244 	/*
4245 	 * The new pmap is resident. Make sure it's marked
4246 	 * as resident in the cache/TLB.
4247 	 */
4248 	npm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
4249 	if (npm != pmap_kernel())
4250 		pmap_recent_user = npm;
4251 
4252 	/* The old pmap is not longer active */
4253 	if (opm != NULL)
4254 		opm->pm_activated = false;
4255 
4256 	/* But the new one is */
4257 	npm->pm_activated = true;
4258 }
4259 
4260 void
4261 pmap_deactivate(struct lwp *l)
4262 {
4263 
4264 	/*
4265 	 * If the process is exiting, make sure pmap_activate() does
4266 	 * a full MMU context-switch and cache flush, which we might
4267 	 * otherwise skip. See PR port-arm/38950.
4268 	 */
4269 	if (l->l_proc->p_sflag & PS_WEXIT)
4270 		pmap_previous_active_lwp = NULL;
4271 
4272 	l->l_proc->p_vmspace->vm_map.pmap->pm_activated = false;
4273 }
4274 
4275 void
4276 pmap_update(pmap_t pm)
4277 {
4278 
4279 	if (pm->pm_remove_all) {
4280 		/*
4281 		 * Finish up the pmap_remove_all() optimisation by flushing
4282 		 * the TLB.
4283 		 */
4284 		pmap_tlb_flushID(pm);
4285 		pm->pm_remove_all = false;
4286 	}
4287 
4288 	if (pmap_is_current(pm)) {
4289 		/*
4290 		 * If we're dealing with a current userland pmap, move its L1
4291 		 * to the end of the LRU.
4292 		 */
4293 		if (pm != pmap_kernel())
4294 			pmap_use_l1(pm);
4295 
4296 		/*
4297 		 * We can assume we're done with frobbing the cache/tlb for
4298 		 * now. Make sure any future pmap ops don't skip cache/tlb
4299 		 * flushes.
4300 		 */
4301 		pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
4302 	}
4303 
4304 	PMAPCOUNT(updates);
4305 
4306 	/*
4307 	 * make sure TLB/cache operations have completed.
4308 	 */
4309 	cpu_cpwait();
4310 }
4311 
4312 void
4313 pmap_remove_all(pmap_t pm)
4314 {
4315 
4316 	/*
4317 	 * The vmspace described by this pmap is about to be torn down.
4318 	 * Until pmap_update() is called, UVM will only make calls
4319 	 * to pmap_remove(). We can make life much simpler by flushing
4320 	 * the cache now, and deferring TLB invalidation to pmap_update().
4321 	 */
4322 #ifdef PMAP_CACHE_VIVT
4323 	pmap_cache_wbinv_all(pm, PVF_EXEC);
4324 #endif
4325 	pm->pm_remove_all = true;
4326 }
4327 
4328 /*
4329  * Retire the given physical map from service.
4330  * Should only be called if the map contains no valid mappings.
4331  */
4332 void
4333 pmap_destroy(pmap_t pm)
4334 {
4335 	u_int count;
4336 
4337 	if (pm == NULL)
4338 		return;
4339 
4340 	if (pm->pm_remove_all) {
4341 		pmap_tlb_flushID(pm);
4342 		pm->pm_remove_all = false;
4343 	}
4344 
4345 	/*
4346 	 * Drop reference count
4347 	 */
4348 	mutex_enter(pm->pm_lock);
4349 	count = --pm->pm_obj.uo_refs;
4350 	mutex_exit(pm->pm_lock);
4351 	if (count > 0) {
4352 		if (pmap_is_current(pm)) {
4353 			if (pm != pmap_kernel())
4354 				pmap_use_l1(pm);
4355 			pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
4356 		}
4357 		return;
4358 	}
4359 
4360 	/*
4361 	 * reference count is zero, free pmap resources and then free pmap.
4362 	 */
4363 
4364 #ifndef ARM_HAS_VBAR
4365 	if (vector_page < KERNEL_BASE) {
4366 		KDASSERT(!pmap_is_current(pm));
4367 
4368 		/* Remove the vector page mapping */
4369 		pmap_remove(pm, vector_page, vector_page + PAGE_SIZE);
4370 		pmap_update(pm);
4371 	}
4372 #endif
4373 
4374 	LIST_REMOVE(pm, pm_list);
4375 
4376 	pmap_free_l1(pm);
4377 
4378 	if (pmap_recent_user == pm)
4379 		pmap_recent_user = NULL;
4380 
4381 	uvm_obj_destroy(&pm->pm_obj, false);
4382 	mutex_destroy(&pm->pm_obj_lock);
4383 	pool_cache_put(&pmap_cache, pm);
4384 }
4385 
4386 
4387 /*
4388  * void pmap_reference(pmap_t pm)
4389  *
4390  * Add a reference to the specified pmap.
4391  */
4392 void
4393 pmap_reference(pmap_t pm)
4394 {
4395 
4396 	if (pm == NULL)
4397 		return;
4398 
4399 	pmap_use_l1(pm);
4400 
4401 	mutex_enter(pm->pm_lock);
4402 	pm->pm_obj.uo_refs++;
4403 	mutex_exit(pm->pm_lock);
4404 }
4405 
4406 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
4407 
4408 static struct evcnt pmap_prefer_nochange_ev =
4409     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap prefer", "nochange");
4410 static struct evcnt pmap_prefer_change_ev =
4411     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap prefer", "change");
4412 
4413 EVCNT_ATTACH_STATIC(pmap_prefer_change_ev);
4414 EVCNT_ATTACH_STATIC(pmap_prefer_nochange_ev);
4415 
4416 void
4417 pmap_prefer(vaddr_t hint, vaddr_t *vap, int td)
4418 {
4419 	vsize_t mask = arm_cache_prefer_mask | (PAGE_SIZE - 1);
4420 	vaddr_t va = *vap;
4421 	vaddr_t diff = (hint - va) & mask;
4422 	if (diff == 0) {
4423 		pmap_prefer_nochange_ev.ev_count++;
4424 	} else {
4425 		pmap_prefer_change_ev.ev_count++;
4426 		if (__predict_false(td))
4427 			va -= mask + 1;
4428 		*vap = va + diff;
4429 	}
4430 }
4431 #endif /* ARM_MMU_V6 | ARM_MMU_V7 */
4432 
4433 /*
4434  * pmap_zero_page()
4435  *
4436  * Zero a given physical page by mapping it at a page hook point.
4437  * In doing the zero page op, the page we zero is mapped cachable, as with
4438  * StrongARM accesses to non-cached pages are non-burst making writing
4439  * _any_ bulk data very slow.
4440  */
4441 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
4442 void
4443 pmap_zero_page_generic(paddr_t phys)
4444 {
4445 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
4446 	struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
4447 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4448 #endif
4449 #if defined(PMAP_CACHE_VIPT)
4450 	/* Choose the last page color it had, if any */
4451 	const vsize_t va_offset = md->pvh_attrs & arm_cache_prefer_mask;
4452 #else
4453 	const vsize_t va_offset = 0;
4454 #endif
4455 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
4456 	/*
4457 	 * Is this page mapped at its natural color?
4458 	 * If we have all of memory mapped, then just convert PA to VA.
4459 	 */
4460 	const bool okcolor = va_offset == (phys & arm_cache_prefer_mask);
4461 	const vaddr_t vdstp = KERNEL_BASE + (phys - physical_start);
4462 #else
4463 	const bool okcolor = false;
4464 	const vaddr_t vdstp = cdstp + va_offset;
4465 #endif
4466 	pt_entry_t * const ptep = &cdst_pte[va_offset >> PGSHIFT];
4467 
4468 
4469 #ifdef DEBUG
4470 	if (!SLIST_EMPTY(&md->pvh_list))
4471 		panic("pmap_zero_page: page has mappings");
4472 #endif
4473 
4474 	KDASSERT((phys & PGOFSET) == 0);
4475 
4476 	if (!okcolor) {
4477 		/*
4478 		 * Hook in the page, zero it, and purge the cache for that
4479 		 * zeroed page. Invalidate the TLB as needed.
4480 		 */
4481 		*ptep = L2_S_PROTO | phys |
4482 		    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
4483 		PTE_SYNC(ptep);
4484 		cpu_tlb_flushD_SE(cdstp + va_offset);
4485 		cpu_cpwait();
4486 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS) && defined(PMAP_CACHE_VIPT)
4487 		/*
4488 		 * If we are direct-mapped and our color isn't ok, then before
4489 		 * we bzero the page invalidate its contents from the cache and
4490 		 * reset the color to its natural color.
4491 		 */
4492 		cpu_dcache_inv_range(cdstp + va_offset, PAGE_SIZE);
4493 		md->pvh_attrs &= ~arm_cache_prefer_mask;
4494 		md->pvh_attrs |= (phys & arm_cache_prefer_mask);
4495 #endif
4496 	}
4497 	bzero_page(vdstp);
4498 	if (!okcolor) {
4499 		/*
4500 		 * Unmap the page.
4501 		 */
4502 		*ptep = 0;
4503 		PTE_SYNC(ptep);
4504 		cpu_tlb_flushD_SE(cdstp + va_offset);
4505 #ifdef PMAP_CACHE_VIVT
4506 		cpu_dcache_wbinv_range(cdstp + va_offset, PAGE_SIZE);
4507 #endif
4508 	}
4509 #ifdef PMAP_CACHE_VIPT
4510 	/*
4511 	 * This page is now cache resident so it now has a page color.
4512 	 * Any contents have been obliterated so clear the EXEC flag.
4513 	 */
4514 	if (!pmap_is_page_colored_p(md)) {
4515 		PMAPCOUNT(vac_color_new);
4516 		md->pvh_attrs |= PVF_COLORED;
4517 	}
4518 	if (PV_IS_EXEC_P(md->pvh_attrs)) {
4519 		md->pvh_attrs &= ~PVF_EXEC;
4520 		PMAPCOUNT(exec_discarded_zero);
4521 	}
4522 	md->pvh_attrs |= PVF_DIRTY;
4523 #endif
4524 }
4525 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
4526 
4527 #if ARM_MMU_XSCALE == 1
4528 void
4529 pmap_zero_page_xscale(paddr_t phys)
4530 {
4531 #ifdef DEBUG
4532 	struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
4533 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4534 
4535 	if (!SLIST_EMPTY(&md->pvh_list))
4536 		panic("pmap_zero_page: page has mappings");
4537 #endif
4538 
4539 	KDASSERT((phys & PGOFSET) == 0);
4540 
4541 	/*
4542 	 * Hook in the page, zero it, and purge the cache for that
4543 	 * zeroed page. Invalidate the TLB as needed.
4544 	 */
4545 	*cdst_pte = L2_S_PROTO | phys |
4546 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
4547 	    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
4548 	PTE_SYNC(cdst_pte);
4549 	cpu_tlb_flushD_SE(cdstp);
4550 	cpu_cpwait();
4551 	bzero_page(cdstp);
4552 	xscale_cache_clean_minidata();
4553 }
4554 #endif /* ARM_MMU_XSCALE == 1 */
4555 
4556 /* pmap_pageidlezero()
4557  *
4558  * The same as above, except that we assume that the page is not
4559  * mapped.  This means we never have to flush the cache first.  Called
4560  * from the idle loop.
4561  */
4562 bool
4563 pmap_pageidlezero(paddr_t phys)
4564 {
4565 	unsigned int i;
4566 	int *ptr;
4567 	bool rv = true;
4568 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
4569 	struct vm_page * const pg = PHYS_TO_VM_PAGE(phys);
4570 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4571 #endif
4572 #ifdef PMAP_CACHE_VIPT
4573 	/* Choose the last page color it had, if any */
4574 	const vsize_t va_offset = md->pvh_attrs & arm_cache_prefer_mask;
4575 #else
4576 	const vsize_t va_offset = 0;
4577 #endif
4578 	pt_entry_t * const ptep = &csrc_pte[va_offset >> PGSHIFT];
4579 
4580 
4581 #ifdef DEBUG
4582 	if (!SLIST_EMPTY(&md->pvh_list))
4583 		panic("pmap_pageidlezero: page has mappings");
4584 #endif
4585 
4586 	KDASSERT((phys & PGOFSET) == 0);
4587 
4588 	/*
4589 	 * Hook in the page, zero it, and purge the cache for that
4590 	 * zeroed page. Invalidate the TLB as needed.
4591 	 */
4592 	*ptep = L2_S_PROTO | phys |
4593 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
4594 	PTE_SYNC(ptep);
4595 	cpu_tlb_flushD_SE(cdstp + va_offset);
4596 	cpu_cpwait();
4597 
4598 	for (i = 0, ptr = (int *)(cdstp + va_offset);
4599 			i < (PAGE_SIZE / sizeof(int)); i++) {
4600 		if (sched_curcpu_runnable_p() != 0) {
4601 			/*
4602 			 * A process has become ready.  Abort now,
4603 			 * so we don't keep it waiting while we
4604 			 * do slow memory access to finish this
4605 			 * page.
4606 			 */
4607 			rv = false;
4608 			break;
4609 		}
4610 		*ptr++ = 0;
4611 	}
4612 
4613 #ifdef PMAP_CACHE_VIVT
4614 	if (rv)
4615 		/*
4616 		 * if we aborted we'll rezero this page again later so don't
4617 		 * purge it unless we finished it
4618 		 */
4619 		cpu_dcache_wbinv_range(cdstp, PAGE_SIZE);
4620 #elif defined(PMAP_CACHE_VIPT)
4621 	/*
4622 	 * This page is now cache resident so it now has a page color.
4623 	 * Any contents have been obliterated so clear the EXEC flag.
4624 	 */
4625 	if (!pmap_is_page_colored_p(md)) {
4626 		PMAPCOUNT(vac_color_new);
4627 		md->pvh_attrs |= PVF_COLORED;
4628 	}
4629 	if (PV_IS_EXEC_P(md->pvh_attrs)) {
4630 		md->pvh_attrs &= ~PVF_EXEC;
4631 		PMAPCOUNT(exec_discarded_zero);
4632 	}
4633 #endif
4634 	/*
4635 	 * Unmap the page.
4636 	 */
4637 	*ptep = 0;
4638 	PTE_SYNC(ptep);
4639 	cpu_tlb_flushD_SE(cdstp + va_offset);
4640 
4641 	return (rv);
4642 }
4643 
4644 /*
4645  * pmap_copy_page()
4646  *
4647  * Copy one physical page into another, by mapping the pages into
4648  * hook points. The same comment regarding cachability as in
4649  * pmap_zero_page also applies here.
4650  */
4651 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
4652 void
4653 pmap_copy_page_generic(paddr_t src, paddr_t dst)
4654 {
4655 	struct vm_page * const src_pg = PHYS_TO_VM_PAGE(src);
4656 	struct vm_page_md *src_md = VM_PAGE_TO_MD(src_pg);
4657 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
4658 	struct vm_page * const dst_pg = PHYS_TO_VM_PAGE(dst);
4659 	struct vm_page_md *dst_md = VM_PAGE_TO_MD(dst_pg);
4660 #endif
4661 #ifdef PMAP_CACHE_VIPT
4662 	const vsize_t src_va_offset = src_md->pvh_attrs & arm_cache_prefer_mask;
4663 	const vsize_t dst_va_offset = dst_md->pvh_attrs & arm_cache_prefer_mask;
4664 #else
4665 	const vsize_t src_va_offset = 0;
4666 	const vsize_t dst_va_offset = 0;
4667 #endif
4668 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
4669 	/*
4670 	 * Is this page mapped at its natural color?
4671 	 * If we have all of memory mapped, then just convert PA to VA.
4672 	 */
4673 	const bool src_okcolor = src_va_offset == (src & arm_cache_prefer_mask);
4674 	const bool dst_okcolor = dst_va_offset == (dst & arm_cache_prefer_mask);
4675 	const vaddr_t vsrcp = src_okcolor
4676 	    ? KERNEL_BASE + (src - physical_start)
4677 	    : csrcp + src_va_offset;
4678 	const vaddr_t vdstp = KERNEL_BASE + (dst - physical_start);
4679 #else
4680 	const bool src_okcolor = false;
4681 	const bool dst_okcolor = false;
4682 	const vaddr_t vsrcp = csrcp + src_va_offset;
4683 	const vaddr_t vdstp = cdstp + dst_va_offset;
4684 #endif
4685 	pt_entry_t * const src_ptep = &csrc_pte[src_va_offset >> PGSHIFT];
4686 	pt_entry_t * const dst_ptep = &cdst_pte[dst_va_offset >> PGSHIFT];
4687 
4688 #ifdef DEBUG
4689 	if (!SLIST_EMPTY(&dst_md->pvh_list))
4690 		panic("pmap_copy_page: dst page has mappings");
4691 #endif
4692 
4693 #ifdef PMAP_CACHE_VIPT
4694 	KASSERT(arm_cache_prefer_mask == 0 || src_md->pvh_attrs & (PVF_COLORED|PVF_NC));
4695 #endif
4696 	KDASSERT((src & PGOFSET) == 0);
4697 	KDASSERT((dst & PGOFSET) == 0);
4698 
4699 	/*
4700 	 * Clean the source page.  Hold the source page's lock for
4701 	 * the duration of the copy so that no other mappings can
4702 	 * be created while we have a potentially aliased mapping.
4703 	 */
4704 #ifdef MULTIPROCESSOR
4705 	KASSERT(uvm_page_locked_p(src_pg));
4706 #endif
4707 #ifdef PMAP_CACHE_VIVT
4708 	(void) pmap_clean_page(SLIST_FIRST(&src_md->pvh_list), true);
4709 #endif
4710 
4711 	/*
4712 	 * Map the pages into the page hook points, copy them, and purge
4713 	 * the cache for the appropriate page. Invalidate the TLB
4714 	 * as required.
4715 	 */
4716 	if (!src_okcolor) {
4717 		*src_ptep = L2_S_PROTO
4718 		    | src
4719 #ifdef PMAP_CACHE_VIPT
4720 		    | ((src_md->pvh_attrs & PVF_NC) ? 0 : pte_l2_s_cache_mode)
4721 #endif
4722 #ifdef PMAP_CACHE_VIVT
4723 		    | pte_l2_s_cache_mode
4724 #endif
4725 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ);
4726 		PTE_SYNC(src_ptep);
4727 		cpu_tlb_flushD_SE(csrcp + src_va_offset);
4728 		cpu_cpwait();
4729 	}
4730 	if (!dst_okcolor) {
4731 		*dst_ptep = L2_S_PROTO | dst |
4732 		    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
4733 		PTE_SYNC(dst_ptep);
4734 		cpu_tlb_flushD_SE(cdstp + dst_va_offset);
4735 		cpu_cpwait();
4736 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS) && defined(PMAP_CACHE_VIPT)
4737 		/*
4738 		 * If we are direct-mapped and our color isn't ok, then before
4739 		 * we bcopy to the new page invalidate its contents from the
4740 		 * cache and reset its color to its natural color.
4741 		 */
4742 		cpu_dcache_inv_range(cdstp + dst_va_offset, PAGE_SIZE);
4743 		dst_md->pvh_attrs &= ~arm_cache_prefer_mask;
4744 		dst_md->pvh_attrs |= (dst & arm_cache_prefer_mask);
4745 #endif
4746 	}
4747 	bcopy_page(vsrcp, vdstp);
4748 #ifdef PMAP_CACHE_VIVT
4749 	cpu_dcache_inv_range(vsrcp, PAGE_SIZE);
4750 	cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
4751 #endif
4752 	/*
4753 	 * Unmap the pages.
4754 	 */
4755 	if (!src_okcolor) {
4756 		*src_ptep = 0;
4757 		PTE_SYNC(src_ptep);
4758 		cpu_tlb_flushD_SE(csrcp + src_va_offset);
4759 		cpu_cpwait();
4760 	}
4761 	if (!dst_okcolor) {
4762 		*dst_ptep = 0;
4763 		PTE_SYNC(dst_ptep);
4764 		cpu_tlb_flushD_SE(cdstp + dst_va_offset);
4765 		cpu_cpwait();
4766 	}
4767 #ifdef PMAP_CACHE_VIPT
4768 	/*
4769 	 * Now that the destination page is in the cache, mark it as colored.
4770 	 * If this was an exec page, discard it.
4771 	 */
4772 	if (!pmap_is_page_colored_p(dst_md)) {
4773 		PMAPCOUNT(vac_color_new);
4774 		dst_md->pvh_attrs |= PVF_COLORED;
4775 	}
4776 	if (PV_IS_EXEC_P(dst_md->pvh_attrs)) {
4777 		dst_md->pvh_attrs &= ~PVF_EXEC;
4778 		PMAPCOUNT(exec_discarded_copy);
4779 	}
4780 	dst_md->pvh_attrs |= PVF_DIRTY;
4781 #endif
4782 }
4783 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
4784 
4785 #if ARM_MMU_XSCALE == 1
4786 void
4787 pmap_copy_page_xscale(paddr_t src, paddr_t dst)
4788 {
4789 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
4790 	struct vm_page_md *src_md = VM_PAGE_TO_MD(src_pg);
4791 #ifdef DEBUG
4792 	struct vm_page_md *dst_md = VM_PAGE_TO_MD(PHYS_TO_VM_PAGE(dst));
4793 
4794 	if (!SLIST_EMPTY(&dst_md->pvh_list))
4795 		panic("pmap_copy_page: dst page has mappings");
4796 #endif
4797 
4798 	KDASSERT((src & PGOFSET) == 0);
4799 	KDASSERT((dst & PGOFSET) == 0);
4800 
4801 	/*
4802 	 * Clean the source page.  Hold the source page's lock for
4803 	 * the duration of the copy so that no other mappings can
4804 	 * be created while we have a potentially aliased mapping.
4805 	 */
4806 #ifdef MULTIPROCESSOR
4807 	KASSERT(uvm_page_locked_p(src_pg));
4808 #endif
4809 #ifdef PMAP_CACHE_VIVT
4810 	(void) pmap_clean_page(SLIST_FIRST(&src_md->pvh_list), true);
4811 #endif
4812 
4813 	/*
4814 	 * Map the pages into the page hook points, copy them, and purge
4815 	 * the cache for the appropriate page. Invalidate the TLB
4816 	 * as required.
4817 	 */
4818 	*csrc_pte = L2_S_PROTO | src |
4819 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
4820 	    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
4821 	PTE_SYNC(csrc_pte);
4822 	*cdst_pte = L2_S_PROTO | dst |
4823 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
4824 	    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
4825 	PTE_SYNC(cdst_pte);
4826 	cpu_tlb_flushD_SE(csrcp);
4827 	cpu_tlb_flushD_SE(cdstp);
4828 	cpu_cpwait();
4829 	bcopy_page(csrcp, cdstp);
4830 	xscale_cache_clean_minidata();
4831 }
4832 #endif /* ARM_MMU_XSCALE == 1 */
4833 
4834 /*
4835  * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
4836  *
4837  * Return the start and end addresses of the kernel's virtual space.
4838  * These values are setup in pmap_bootstrap and are updated as pages
4839  * are allocated.
4840  */
4841 void
4842 pmap_virtual_space(vaddr_t *start, vaddr_t *end)
4843 {
4844 	*start = virtual_avail;
4845 	*end = virtual_end;
4846 }
4847 
4848 /*
4849  * Helper function for pmap_grow_l2_bucket()
4850  */
4851 static inline int
4852 pmap_grow_map(vaddr_t va, pt_entry_t cache_mode, paddr_t *pap)
4853 {
4854 	struct l2_bucket *l2b;
4855 	pt_entry_t *ptep;
4856 	paddr_t pa;
4857 
4858 	if (uvm.page_init_done == false) {
4859 #ifdef PMAP_STEAL_MEMORY
4860 		pv_addr_t pv;
4861 		pmap_boot_pagealloc(PAGE_SIZE,
4862 #ifdef PMAP_CACHE_VIPT
4863 		    arm_cache_prefer_mask,
4864 		    va & arm_cache_prefer_mask,
4865 #else
4866 		    0, 0,
4867 #endif
4868 		    &pv);
4869 		pa = pv.pv_pa;
4870 #else
4871 		if (uvm_page_physget(&pa) == false)
4872 			return (1);
4873 #endif	/* PMAP_STEAL_MEMORY */
4874 	} else {
4875 		struct vm_page *pg;
4876 		pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE);
4877 		if (pg == NULL)
4878 			return (1);
4879 		pa = VM_PAGE_TO_PHYS(pg);
4880 #ifdef PMAP_CACHE_VIPT
4881 #ifdef DIAGNOSTIC
4882 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4883 #endif
4884 		/*
4885 		 * This new page must not have any mappings.  Enter it via
4886 		 * pmap_kenter_pa and let that routine do the hard work.
4887 		 */
4888 		KASSERT(SLIST_EMPTY(&md->pvh_list));
4889 		pmap_kenter_pa(va, pa,
4890 		    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE);
4891 #endif
4892 	}
4893 
4894 	if (pap)
4895 		*pap = pa;
4896 
4897 	PMAPCOUNT(pt_mappings);
4898 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
4899 	KDASSERT(l2b != NULL);
4900 
4901 	ptep = &l2b->l2b_kva[l2pte_index(va)];
4902 	*ptep = L2_S_PROTO | pa | cache_mode |
4903 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ | VM_PROT_WRITE);
4904 	PTE_SYNC(ptep);
4905 	memset((void *)va, 0, PAGE_SIZE);
4906 	return (0);
4907 }
4908 
4909 /*
4910  * This is the same as pmap_alloc_l2_bucket(), except that it is only
4911  * used by pmap_growkernel().
4912  */
4913 static inline struct l2_bucket *
4914 pmap_grow_l2_bucket(pmap_t pm, vaddr_t va)
4915 {
4916 	struct l2_dtable *l2;
4917 	struct l2_bucket *l2b;
4918 	u_short l1idx;
4919 	vaddr_t nva;
4920 
4921 	l1idx = L1_IDX(va);
4922 
4923 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
4924 		/*
4925 		 * No mapping at this address, as there is
4926 		 * no entry in the L1 table.
4927 		 * Need to allocate a new l2_dtable.
4928 		 */
4929 		nva = pmap_kernel_l2dtable_kva;
4930 		if ((nva & PGOFSET) == 0) {
4931 			/*
4932 			 * Need to allocate a backing page
4933 			 */
4934 			if (pmap_grow_map(nva, pte_l2_s_cache_mode, NULL))
4935 				return (NULL);
4936 		}
4937 
4938 		l2 = (struct l2_dtable *)nva;
4939 		nva += sizeof(struct l2_dtable);
4940 
4941 		if ((nva & PGOFSET) < (pmap_kernel_l2dtable_kva & PGOFSET)) {
4942 			/*
4943 			 * The new l2_dtable straddles a page boundary.
4944 			 * Map in another page to cover it.
4945 			 */
4946 			if (pmap_grow_map(nva, pte_l2_s_cache_mode, NULL))
4947 				return (NULL);
4948 		}
4949 
4950 		pmap_kernel_l2dtable_kva = nva;
4951 
4952 		/*
4953 		 * Link it into the parent pmap
4954 		 */
4955 		pm->pm_l2[L2_IDX(l1idx)] = l2;
4956 	}
4957 
4958 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
4959 
4960 	/*
4961 	 * Fetch pointer to the L2 page table associated with the address.
4962 	 */
4963 	if (l2b->l2b_kva == NULL) {
4964 		pt_entry_t *ptep;
4965 
4966 		/*
4967 		 * No L2 page table has been allocated. Chances are, this
4968 		 * is because we just allocated the l2_dtable, above.
4969 		 */
4970 		nva = pmap_kernel_l2ptp_kva;
4971 		ptep = (pt_entry_t *)nva;
4972 		if ((nva & PGOFSET) == 0) {
4973 			/*
4974 			 * Need to allocate a backing page
4975 			 */
4976 			if (pmap_grow_map(nva, pte_l2_s_cache_mode_pt,
4977 			    &pmap_kernel_l2ptp_phys))
4978 				return (NULL);
4979 			PTE_SYNC_RANGE(ptep, PAGE_SIZE / sizeof(pt_entry_t));
4980 		}
4981 
4982 		l2->l2_occupancy++;
4983 		l2b->l2b_kva = ptep;
4984 		l2b->l2b_l1idx = l1idx;
4985 		l2b->l2b_phys = pmap_kernel_l2ptp_phys;
4986 
4987 		pmap_kernel_l2ptp_kva += L2_TABLE_SIZE_REAL;
4988 		pmap_kernel_l2ptp_phys += L2_TABLE_SIZE_REAL;
4989 	}
4990 
4991 	return (l2b);
4992 }
4993 
4994 vaddr_t
4995 pmap_growkernel(vaddr_t maxkvaddr)
4996 {
4997 	pmap_t kpm = pmap_kernel();
4998 	struct l1_ttable *l1;
4999 	struct l2_bucket *l2b;
5000 	pd_entry_t *pl1pd;
5001 	int s;
5002 
5003 	if (maxkvaddr <= pmap_curmaxkvaddr)
5004 		goto out;		/* we are OK */
5005 
5006 	NPDEBUG(PDB_GROWKERN,
5007 	    printf("pmap_growkernel: growing kernel from 0x%lx to 0x%lx\n",
5008 	    pmap_curmaxkvaddr, maxkvaddr));
5009 
5010 	KDASSERT(maxkvaddr <= virtual_end);
5011 
5012 	/*
5013 	 * whoops!   we need to add kernel PTPs
5014 	 */
5015 
5016 	s = splhigh();	/* to be safe */
5017 	mutex_enter(kpm->pm_lock);
5018 
5019 	/* Map 1MB at a time */
5020 	for (; pmap_curmaxkvaddr < maxkvaddr; pmap_curmaxkvaddr += L1_S_SIZE) {
5021 
5022 		l2b = pmap_grow_l2_bucket(kpm, pmap_curmaxkvaddr);
5023 		KDASSERT(l2b != NULL);
5024 
5025 		/* Distribute new L1 entry to all other L1s */
5026 		SLIST_FOREACH(l1, &l1_list, l1_link) {
5027 			pl1pd = &l1->l1_kva[L1_IDX(pmap_curmaxkvaddr)];
5028 			*pl1pd = l2b->l2b_phys | L1_C_DOM(PMAP_DOMAIN_KERNEL) |
5029 			    L1_C_PROTO;
5030 			PTE_SYNC(pl1pd);
5031 		}
5032 	}
5033 
5034 	/*
5035 	 * flush out the cache, expensive but growkernel will happen so
5036 	 * rarely
5037 	 */
5038 	cpu_dcache_wbinv_all();
5039 	cpu_tlb_flushD();
5040 	cpu_cpwait();
5041 
5042 	mutex_exit(kpm->pm_lock);
5043 	splx(s);
5044 
5045 out:
5046 	return (pmap_curmaxkvaddr);
5047 }
5048 
5049 /************************ Utility routines ****************************/
5050 
5051 #ifndef ARM_HAS_VBAR
5052 /*
5053  * vector_page_setprot:
5054  *
5055  *	Manipulate the protection of the vector page.
5056  */
5057 void
5058 vector_page_setprot(int prot)
5059 {
5060 	struct l2_bucket *l2b;
5061 	pt_entry_t *ptep;
5062 
5063 #if defined(CPU_ARMV7) || defined(CPU_ARM11)
5064 	/*
5065 	 * If we are using VBAR to use the vectors in the kernel, then it's
5066 	 * already mapped in the kernel text so no need to anything here.
5067 	 */
5068 	if (vector_page != ARM_VECTORS_LOW && vector_page != ARM_VECTORS_HIGH) {
5069 		KASSERT((armreg_pfr1_read() & ARM_PFR1_SEC_MASK) != 0);
5070 		return;
5071 	}
5072 #endif
5073 
5074 	l2b = pmap_get_l2_bucket(pmap_kernel(), vector_page);
5075 	KDASSERT(l2b != NULL);
5076 
5077 	ptep = &l2b->l2b_kva[l2pte_index(vector_page)];
5078 
5079 	*ptep = (*ptep & ~L2_S_PROT_MASK) | L2_S_PROT(PTE_KERNEL, prot);
5080 	PTE_SYNC(ptep);
5081 	cpu_tlb_flushD_SE(vector_page);
5082 	cpu_cpwait();
5083 }
5084 #endif
5085 
5086 /*
5087  * Fetch pointers to the PDE/PTE for the given pmap/VA pair.
5088  * Returns true if the mapping exists, else false.
5089  *
5090  * NOTE: This function is only used by a couple of arm-specific modules.
5091  * It is not safe to take any pmap locks here, since we could be right
5092  * in the middle of debugging the pmap anyway...
5093  *
5094  * It is possible for this routine to return false even though a valid
5095  * mapping does exist. This is because we don't lock, so the metadata
5096  * state may be inconsistent.
5097  *
5098  * NOTE: We can return a NULL *ptp in the case where the L1 pde is
5099  * a "section" mapping.
5100  */
5101 bool
5102 pmap_get_pde_pte(pmap_t pm, vaddr_t va, pd_entry_t **pdp, pt_entry_t **ptp)
5103 {
5104 	struct l2_dtable *l2;
5105 	pd_entry_t *pl1pd, l1pd;
5106 	pt_entry_t *ptep;
5107 	u_short l1idx;
5108 
5109 	if (pm->pm_l1 == NULL)
5110 		return false;
5111 
5112 	l1idx = L1_IDX(va);
5113 	*pdp = pl1pd = pmap_l1_kva(pm) + l1idx;
5114 	l1pd = *pl1pd;
5115 
5116 	if (l1pte_section_p(l1pd)) {
5117 		*ptp = NULL;
5118 		return true;
5119 	}
5120 
5121 	if (pm->pm_l2 == NULL)
5122 		return false;
5123 
5124 	l2 = pm->pm_l2[L2_IDX(l1idx)];
5125 
5126 	if (l2 == NULL ||
5127 	    (ptep = l2->l2_bucket[L2_BUCKET(l1idx)].l2b_kva) == NULL) {
5128 		return false;
5129 	}
5130 
5131 	*ptp = &ptep[l2pte_index(va)];
5132 	return true;
5133 }
5134 
5135 bool
5136 pmap_get_pde(pmap_t pm, vaddr_t va, pd_entry_t **pdp)
5137 {
5138 
5139 	if (pm->pm_l1 == NULL)
5140 		return false;
5141 
5142 	*pdp = pmap_l1_kva(pm) + L1_IDX(va);
5143 
5144 	return true;
5145 }
5146 
5147 /************************ Bootstrapping routines ****************************/
5148 
5149 static void
5150 pmap_init_l1(struct l1_ttable *l1, pd_entry_t *l1pt)
5151 {
5152 	int i;
5153 
5154 	l1->l1_kva = l1pt;
5155 	l1->l1_domain_use_count = 0;
5156 	l1->l1_domain_first = 0;
5157 
5158 	for (i = 0; i < PMAP_DOMAINS; i++)
5159 		l1->l1_domain_free[i] = i + 1;
5160 
5161 	/*
5162 	 * Copy the kernel's L1 entries to each new L1.
5163 	 */
5164 	if (pmap_initialized)
5165 		memcpy(l1pt, pmap_l1_kva(pmap_kernel()), L1_TABLE_SIZE);
5166 
5167 	if (pmap_extract(pmap_kernel(), (vaddr_t)l1pt,
5168 	    &l1->l1_physaddr) == false)
5169 		panic("pmap_init_l1: can't get PA of L1 at %p", l1pt);
5170 
5171 	SLIST_INSERT_HEAD(&l1_list, l1, l1_link);
5172 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
5173 }
5174 
5175 /*
5176  * pmap_bootstrap() is called from the board-specific initarm() routine
5177  * once the kernel L1/L2 descriptors tables have been set up.
5178  *
5179  * This is a somewhat convoluted process since pmap bootstrap is, effectively,
5180  * spread over a number of disparate files/functions.
5181  *
5182  * We are passed the following parameters
5183  *  - kernel_l1pt
5184  *    This is a pointer to the base of the kernel's L1 translation table.
5185  *  - vstart
5186  *    1MB-aligned start of managed kernel virtual memory.
5187  *  - vend
5188  *    1MB-aligned end of managed kernel virtual memory.
5189  *
5190  * We use the first parameter to build the metadata (struct l1_ttable and
5191  * struct l2_dtable) necessary to track kernel mappings.
5192  */
5193 #define	PMAP_STATIC_L2_SIZE 16
5194 void
5195 pmap_bootstrap(vaddr_t vstart, vaddr_t vend)
5196 {
5197 	static struct l1_ttable static_l1;
5198 	static struct l2_dtable static_l2[PMAP_STATIC_L2_SIZE];
5199 	struct l1_ttable *l1 = &static_l1;
5200 	struct l2_dtable *l2;
5201 	struct l2_bucket *l2b;
5202 	pd_entry_t *l1pt = (pd_entry_t *) kernel_l1pt.pv_va;
5203 	pmap_t pm = pmap_kernel();
5204 	pd_entry_t pde;
5205 	pt_entry_t *ptep;
5206 	paddr_t pa;
5207 	vaddr_t va;
5208 	vsize_t size;
5209 	int nptes, l1idx, l2idx, l2next = 0;
5210 
5211 	/*
5212 	 * Initialise the kernel pmap object
5213 	 */
5214 	pm->pm_l1 = l1;
5215 	pm->pm_domain = PMAP_DOMAIN_KERNEL;
5216 	pm->pm_activated = true;
5217 	pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
5218 
5219 	mutex_init(&pm->pm_obj_lock, MUTEX_DEFAULT, IPL_NONE);
5220 	uvm_obj_init(&pm->pm_obj, NULL, false, 1);
5221 	uvm_obj_setlock(&pm->pm_obj, &pm->pm_obj_lock);
5222 
5223 	/*
5224 	 * Scan the L1 translation table created by initarm() and create
5225 	 * the required metadata for all valid mappings found in it.
5226 	 */
5227 	for (l1idx = 0; l1idx < (L1_TABLE_SIZE / sizeof(pd_entry_t)); l1idx++) {
5228 		pde = l1pt[l1idx];
5229 
5230 		/*
5231 		 * We're only interested in Coarse mappings.
5232 		 * pmap_extract() can deal with section mappings without
5233 		 * recourse to checking L2 metadata.
5234 		 */
5235 		if ((pde & L1_TYPE_MASK) != L1_TYPE_C)
5236 			continue;
5237 
5238 		/*
5239 		 * Lookup the KVA of this L2 descriptor table
5240 		 */
5241 		pa = (paddr_t)(pde & L1_C_ADDR_MASK);
5242 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
5243 		if (ptep == NULL) {
5244 			panic("pmap_bootstrap: No L2 for va 0x%x, pa 0x%lx",
5245 			    (u_int)l1idx << L1_S_SHIFT, pa);
5246 		}
5247 
5248 		/*
5249 		 * Fetch the associated L2 metadata structure.
5250 		 * Allocate a new one if necessary.
5251 		 */
5252 		if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
5253 			if (l2next == PMAP_STATIC_L2_SIZE)
5254 				panic("pmap_bootstrap: out of static L2s");
5255 			pm->pm_l2[L2_IDX(l1idx)] = l2 = &static_l2[l2next++];
5256 		}
5257 
5258 		/*
5259 		 * One more L1 slot tracked...
5260 		 */
5261 		l2->l2_occupancy++;
5262 
5263 		/*
5264 		 * Fill in the details of the L2 descriptor in the
5265 		 * appropriate bucket.
5266 		 */
5267 		l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
5268 		l2b->l2b_kva = ptep;
5269 		l2b->l2b_phys = pa;
5270 		l2b->l2b_l1idx = l1idx;
5271 
5272 		/*
5273 		 * Establish an initial occupancy count for this descriptor
5274 		 */
5275 		for (l2idx = 0;
5276 		    l2idx < (L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
5277 		    l2idx++) {
5278 			if ((ptep[l2idx] & L2_TYPE_MASK) != L2_TYPE_INV) {
5279 				l2b->l2b_occupancy++;
5280 			}
5281 		}
5282 
5283 		/*
5284 		 * Make sure the descriptor itself has the correct cache mode.
5285 		 * If not, fix it, but whine about the problem. Port-meisters
5286 		 * should consider this a clue to fix up their initarm()
5287 		 * function. :)
5288 		 */
5289 		if (pmap_set_pt_cache_mode(l1pt, (vaddr_t)ptep)) {
5290 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
5291 			    "L2 pte @ %p\n", ptep);
5292 		}
5293 	}
5294 
5295 	/*
5296 	 * Ensure the primary (kernel) L1 has the correct cache mode for
5297 	 * a page table. Bitch if it is not correctly set.
5298 	 */
5299 	for (va = (vaddr_t)l1pt;
5300 	    va < ((vaddr_t)l1pt + L1_TABLE_SIZE); va += PAGE_SIZE) {
5301 		if (pmap_set_pt_cache_mode(l1pt, va))
5302 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
5303 			    "primary L1 @ 0x%lx\n", va);
5304 	}
5305 
5306 	cpu_dcache_wbinv_all();
5307 	cpu_tlb_flushID();
5308 	cpu_cpwait();
5309 
5310 	/*
5311 	 * now we allocate the "special" VAs which are used for tmp mappings
5312 	 * by the pmap (and other modules).  we allocate the VAs by advancing
5313 	 * virtual_avail (note that there are no pages mapped at these VAs).
5314 	 *
5315 	 * Managed KVM space start from wherever initarm() tells us.
5316 	 */
5317 	virtual_avail = vstart;
5318 	virtual_end = vend;
5319 
5320 #ifdef PMAP_CACHE_VIPT
5321 	/*
5322 	 * If we have a VIPT cache, we need one page/pte per possible alias
5323 	 * page so we won't violate cache aliasing rules.
5324 	 */
5325 	virtual_avail = (virtual_avail + arm_cache_prefer_mask) & ~arm_cache_prefer_mask;
5326 	nptes = (arm_cache_prefer_mask >> PGSHIFT) + 1;
5327 #else
5328 	nptes = 1;
5329 #endif
5330 	pmap_alloc_specials(&virtual_avail, nptes, &csrcp, &csrc_pte);
5331 	pmap_set_pt_cache_mode(l1pt, (vaddr_t)csrc_pte);
5332 	pmap_alloc_specials(&virtual_avail, nptes, &cdstp, &cdst_pte);
5333 	pmap_set_pt_cache_mode(l1pt, (vaddr_t)cdst_pte);
5334 	pmap_alloc_specials(&virtual_avail, nptes, &memhook, NULL);
5335 	pmap_alloc_specials(&virtual_avail, round_page(MSGBUFSIZE) / PAGE_SIZE,
5336 	    (void *)&msgbufaddr, NULL);
5337 
5338 	/*
5339 	 * Allocate a range of kernel virtual address space to be used
5340 	 * for L2 descriptor tables and metadata allocation in
5341 	 * pmap_growkernel().
5342 	 */
5343 	size = ((virtual_end - pmap_curmaxkvaddr) + L1_S_OFFSET) / L1_S_SIZE;
5344 	pmap_alloc_specials(&virtual_avail,
5345 	    round_page(size * L2_TABLE_SIZE_REAL) / PAGE_SIZE,
5346 	    &pmap_kernel_l2ptp_kva, NULL);
5347 
5348 	size = (size + (L2_BUCKET_SIZE - 1)) / L2_BUCKET_SIZE;
5349 	pmap_alloc_specials(&virtual_avail,
5350 	    round_page(size * sizeof(struct l2_dtable)) / PAGE_SIZE,
5351 	    &pmap_kernel_l2dtable_kva, NULL);
5352 
5353 	/*
5354 	 * init the static-global locks and global pmap list.
5355 	 */
5356 	mutex_init(&l1_lru_lock, MUTEX_DEFAULT, IPL_VM);
5357 
5358 	/*
5359 	 * We can now initialise the first L1's metadata.
5360 	 */
5361 	SLIST_INIT(&l1_list);
5362 	TAILQ_INIT(&l1_lru_list);
5363 	pmap_init_l1(l1, l1pt);
5364 
5365 #ifndef ARM_HAS_VBAR
5366 	/* Set up vector page L1 details, if necessary */
5367 	if (vector_page < KERNEL_BASE) {
5368 		pm->pm_pl1vec = pmap_l1_kva(pm) + L1_IDX(vector_page);
5369 		l2b = pmap_get_l2_bucket(pm, vector_page);
5370 		KDASSERT(l2b != NULL);
5371 		pm->pm_l1vec = l2b->l2b_phys | L1_C_PROTO |
5372 		    L1_C_DOM(pmap_domain(pm));
5373 	} else
5374 		pm->pm_pl1vec = NULL;
5375 #endif
5376 
5377 	/*
5378 	 * Initialize the pmap cache
5379 	 */
5380 	pool_cache_bootstrap(&pmap_cache, sizeof(struct pmap), 0, 0, 0,
5381 	    "pmappl", NULL, IPL_NONE, pmap_pmap_ctor, NULL, NULL);
5382 	LIST_INIT(&pmap_pmaps);
5383 	LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list);
5384 
5385 	/*
5386 	 * Initialize the pv pool.
5387 	 */
5388 	pool_init(&pmap_pv_pool, sizeof(struct pv_entry), 0, 0, 0, "pvepl",
5389 	    &pmap_bootstrap_pv_allocator, IPL_NONE);
5390 
5391 	/*
5392 	 * Initialize the L2 dtable pool and cache.
5393 	 */
5394 	pool_cache_bootstrap(&pmap_l2dtable_cache, sizeof(struct l2_dtable), 0,
5395 	    0, 0, "l2dtblpl", NULL, IPL_NONE, pmap_l2dtable_ctor, NULL, NULL);
5396 
5397 	/*
5398 	 * Initialise the L2 descriptor table pool and cache
5399 	 */
5400 	pool_cache_bootstrap(&pmap_l2ptp_cache, L2_TABLE_SIZE_REAL, 0,
5401 	    L2_TABLE_SIZE_REAL, 0, "l2ptppl", NULL, IPL_NONE,
5402 	    pmap_l2ptp_ctor, NULL, NULL);
5403 
5404 	cpu_dcache_wbinv_all();
5405 }
5406 
5407 static int
5408 pmap_set_pt_cache_mode(pd_entry_t *kl1, vaddr_t va)
5409 {
5410 	pd_entry_t *pdep, pde;
5411 	pt_entry_t *ptep, pte;
5412 	vaddr_t pa;
5413 	int rv = 0;
5414 
5415 	/*
5416 	 * Make sure the descriptor itself has the correct cache mode
5417 	 */
5418 	pdep = &kl1[L1_IDX(va)];
5419 	pde = *pdep;
5420 
5421 	if (l1pte_section_p(pde)) {
5422 		__CTASSERT((L1_S_CACHE_MASK & L1_S_V6_SUPER) == 0);
5423 		if ((pde & L1_S_CACHE_MASK) != pte_l1_s_cache_mode_pt) {
5424 			*pdep = (pde & ~L1_S_CACHE_MASK) |
5425 			    pte_l1_s_cache_mode_pt;
5426 			PTE_SYNC(pdep);
5427 			cpu_dcache_wbinv_range((vaddr_t)pdep, sizeof(*pdep));
5428 			rv = 1;
5429 		}
5430 	} else {
5431 		pa = (paddr_t)(pde & L1_C_ADDR_MASK);
5432 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
5433 		if (ptep == NULL)
5434 			panic("pmap_bootstrap: No L2 for L2 @ va %p\n", ptep);
5435 
5436 		ptep = &ptep[l2pte_index(va)];
5437 		pte = *ptep;
5438 		if ((pte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
5439 			*ptep = (pte & ~L2_S_CACHE_MASK) |
5440 			    pte_l2_s_cache_mode_pt;
5441 			PTE_SYNC(ptep);
5442 			cpu_dcache_wbinv_range((vaddr_t)ptep, sizeof(*ptep));
5443 			rv = 1;
5444 		}
5445 	}
5446 
5447 	return (rv);
5448 }
5449 
5450 static void
5451 pmap_alloc_specials(vaddr_t *availp, int pages, vaddr_t *vap, pt_entry_t **ptep)
5452 {
5453 	vaddr_t va = *availp;
5454 	struct l2_bucket *l2b;
5455 
5456 	if (ptep) {
5457 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
5458 		if (l2b == NULL)
5459 			panic("pmap_alloc_specials: no l2b for 0x%lx", va);
5460 
5461 		if (ptep)
5462 			*ptep = &l2b->l2b_kva[l2pte_index(va)];
5463 	}
5464 
5465 	*vap = va;
5466 	*availp = va + (PAGE_SIZE * pages);
5467 }
5468 
5469 void
5470 pmap_init(void)
5471 {
5472 
5473 	/*
5474 	 * Set the available memory vars - These do not map to real memory
5475 	 * addresses and cannot as the physical memory is fragmented.
5476 	 * They are used by ps for %mem calculations.
5477 	 * One could argue whether this should be the entire memory or just
5478 	 * the memory that is useable in a user process.
5479 	 */
5480 	avail_start = ptoa(VM_PHYSMEM_PTR(0)->start);
5481 	avail_end = ptoa(VM_PHYSMEM_PTR(vm_nphysseg - 1)->end);
5482 
5483 	/*
5484 	 * Now we need to free enough pv_entry structures to allow us to get
5485 	 * the kmem_map/kmem_object allocated and inited (done after this
5486 	 * function is finished).  to do this we allocate one bootstrap page out
5487 	 * of kernel_map and use it to provide an initial pool of pv_entry
5488 	 * structures.   we never free this page.
5489 	 */
5490 	pool_setlowat(&pmap_pv_pool,
5491 	    (PAGE_SIZE / sizeof(struct pv_entry)) * 2);
5492 
5493 	mutex_init(&memlock, MUTEX_DEFAULT, IPL_NONE);
5494 	zeropage = (void *)uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
5495 	    UVM_KMF_WIRED|UVM_KMF_ZERO);
5496 
5497 	pmap_initialized = true;
5498 }
5499 
5500 static vaddr_t last_bootstrap_page = 0;
5501 static void *free_bootstrap_pages = NULL;
5502 
5503 static void *
5504 pmap_bootstrap_pv_page_alloc(struct pool *pp, int flags)
5505 {
5506 	extern void *pool_page_alloc(struct pool *, int);
5507 	vaddr_t new_page;
5508 	void *rv;
5509 
5510 	if (pmap_initialized)
5511 		return (pool_page_alloc(pp, flags));
5512 
5513 	if (free_bootstrap_pages) {
5514 		rv = free_bootstrap_pages;
5515 		free_bootstrap_pages = *((void **)rv);
5516 		return (rv);
5517 	}
5518 
5519 	new_page = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
5520 	    UVM_KMF_WIRED | ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT));
5521 
5522 	KASSERT(new_page > last_bootstrap_page);
5523 	last_bootstrap_page = new_page;
5524 	return ((void *)new_page);
5525 }
5526 
5527 static void
5528 pmap_bootstrap_pv_page_free(struct pool *pp, void *v)
5529 {
5530 	extern void pool_page_free(struct pool *, void *);
5531 
5532 	if ((vaddr_t)v <= last_bootstrap_page) {
5533 		*((void **)v) = free_bootstrap_pages;
5534 		free_bootstrap_pages = v;
5535 		return;
5536 	}
5537 
5538 	if (pmap_initialized) {
5539 		pool_page_free(pp, v);
5540 		return;
5541 	}
5542 }
5543 
5544 /*
5545  * pmap_postinit()
5546  *
5547  * This routine is called after the vm and kmem subsystems have been
5548  * initialised. This allows the pmap code to perform any initialisation
5549  * that can only be done one the memory allocation is in place.
5550  */
5551 void
5552 pmap_postinit(void)
5553 {
5554 	extern paddr_t physical_start, physical_end;
5555 	struct l2_bucket *l2b;
5556 	struct l1_ttable *l1;
5557 	struct pglist plist;
5558 	struct vm_page *m;
5559 	pd_entry_t *pl1pt;
5560 	pt_entry_t *ptep, pte;
5561 	vaddr_t va, eva;
5562 	u_int loop, needed;
5563 	int error;
5564 
5565 	pool_cache_setlowat(&pmap_l2ptp_cache,
5566 	    (PAGE_SIZE / L2_TABLE_SIZE_REAL) * 4);
5567 	pool_cache_setlowat(&pmap_l2dtable_cache,
5568 	    (PAGE_SIZE / sizeof(struct l2_dtable)) * 2);
5569 
5570 	needed = (maxproc / PMAP_DOMAINS) + ((maxproc % PMAP_DOMAINS) ? 1 : 0);
5571 	needed -= 1;
5572 
5573 	l1 = kmem_alloc(sizeof(*l1) * needed, KM_SLEEP);
5574 
5575 	for (loop = 0; loop < needed; loop++, l1++) {
5576 		/* Allocate a L1 page table */
5577 		va = uvm_km_alloc(kernel_map, L1_TABLE_SIZE, 0, UVM_KMF_VAONLY);
5578 		if (va == 0)
5579 			panic("Cannot allocate L1 KVM");
5580 
5581 		error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start,
5582 		    physical_end, L1_TABLE_SIZE, 0, &plist, 1, 1);
5583 		if (error)
5584 			panic("Cannot allocate L1 physical pages");
5585 
5586 		m = TAILQ_FIRST(&plist);
5587 		eva = va + L1_TABLE_SIZE;
5588 		pl1pt = (pd_entry_t *)va;
5589 
5590 		while (m && va < eva) {
5591 			paddr_t pa = VM_PAGE_TO_PHYS(m);
5592 
5593 			pmap_kenter_pa(va, pa,
5594 			    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE);
5595 
5596 			/*
5597 			 * Make sure the L1 descriptor table is mapped
5598 			 * with the cache-mode set to write-through.
5599 			 */
5600 			l2b = pmap_get_l2_bucket(pmap_kernel(), va);
5601 			KDASSERT(l2b != NULL);
5602 			ptep = &l2b->l2b_kva[l2pte_index(va)];
5603 			pte = *ptep;
5604 			pte = (pte & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode_pt;
5605 			*ptep = pte;
5606 			PTE_SYNC(ptep);
5607 			cpu_tlb_flushD_SE(va);
5608 
5609 			va += PAGE_SIZE;
5610 			m = TAILQ_NEXT(m, pageq.queue);
5611 		}
5612 
5613 #ifdef DIAGNOSTIC
5614 		if (m)
5615 			panic("pmap_alloc_l1pt: pglist not empty");
5616 #endif	/* DIAGNOSTIC */
5617 
5618 		pmap_init_l1(l1, pl1pt);
5619 	}
5620 
5621 #ifdef DEBUG
5622 	printf("pmap_postinit: Allocated %d static L1 descriptor tables\n",
5623 	    needed);
5624 #endif
5625 }
5626 
5627 /*
5628  * Note that the following routines are used by board-specific initialisation
5629  * code to configure the initial kernel page tables.
5630  *
5631  * If ARM32_NEW_VM_LAYOUT is *not* defined, they operate on the assumption that
5632  * L2 page-table pages are 4KB in size and use 4 L1 slots. This mimics the
5633  * behaviour of the old pmap, and provides an easy migration path for
5634  * initial bring-up of the new pmap on existing ports. Fortunately,
5635  * pmap_bootstrap() compensates for this hackery. This is only a stop-gap and
5636  * will be deprecated.
5637  *
5638  * If ARM32_NEW_VM_LAYOUT *is* defined, these functions deal with 1KB L2 page
5639  * tables.
5640  */
5641 
5642 /*
5643  * This list exists for the benefit of pmap_map_chunk().  It keeps track
5644  * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
5645  * find them as necessary.
5646  *
5647  * Note that the data on this list MUST remain valid after initarm() returns,
5648  * as pmap_bootstrap() uses it to contruct L2 table metadata.
5649  */
5650 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
5651 
5652 static vaddr_t
5653 kernel_pt_lookup(paddr_t pa)
5654 {
5655 	pv_addr_t *pv;
5656 
5657 	SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
5658 #ifndef ARM32_NEW_VM_LAYOUT
5659 		if (pv->pv_pa == (pa & ~PGOFSET))
5660 			return (pv->pv_va | (pa & PGOFSET));
5661 #else
5662 		if (pv->pv_pa == pa)
5663 			return (pv->pv_va);
5664 #endif
5665 	}
5666 	return (0);
5667 }
5668 
5669 /*
5670  * pmap_map_section:
5671  *
5672  *	Create a single section mapping.
5673  */
5674 void
5675 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
5676 {
5677 	pd_entry_t *pde = (pd_entry_t *) l1pt;
5678 	pd_entry_t fl;
5679 
5680 	KASSERT(((va | pa) & L1_S_OFFSET) == 0);
5681 
5682 	switch (cache) {
5683 	case PTE_NOCACHE:
5684 	default:
5685 		fl = 0;
5686 		break;
5687 
5688 	case PTE_CACHE:
5689 		fl = pte_l1_s_cache_mode;
5690 		break;
5691 
5692 	case PTE_PAGETABLE:
5693 		fl = pte_l1_s_cache_mode_pt;
5694 		break;
5695 	}
5696 
5697 	pde[L1_IDX(va)] = L1_S_PROTO | pa |
5698 	    L1_S_PROT(PTE_KERNEL, prot) | fl | L1_S_DOM(PMAP_DOMAIN_KERNEL);
5699 	PTE_SYNC(&pde[L1_IDX(va)]);
5700 }
5701 
5702 /*
5703  * pmap_map_entry:
5704  *
5705  *	Create a single page mapping.
5706  */
5707 void
5708 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
5709 {
5710 	pd_entry_t *pde = (pd_entry_t *) l1pt;
5711 	pt_entry_t npte;
5712 	pt_entry_t *ptep;
5713 
5714 	KASSERT(((va | pa) & PGOFSET) == 0);
5715 
5716 	switch (cache) {
5717 	case PTE_NOCACHE:
5718 	default:
5719 		npte = 0;
5720 		break;
5721 
5722 	case PTE_CACHE:
5723 		npte = pte_l2_s_cache_mode;
5724 		break;
5725 
5726 	case PTE_PAGETABLE:
5727 		npte = pte_l2_s_cache_mode_pt;
5728 		break;
5729 	}
5730 
5731 	if ((pde[L1_IDX(va)] & L1_TYPE_MASK) != L1_TYPE_C)
5732 		panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
5733 
5734 #ifndef ARM32_NEW_VM_LAYOUT
5735 	ptep = (pt_entry_t *)
5736 	    kernel_pt_lookup(pde[L1_IDX(va)] & L2_S_FRAME);
5737 #else
5738 	ptep = (pt_entry_t *) kernel_pt_lookup(pde[L1_IDX(va)] & L1_C_ADDR_MASK);
5739 #endif
5740 	if (ptep == NULL)
5741 		panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
5742 
5743 	npte |= L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot);
5744 #ifndef ARM32_NEW_VM_LAYOUT
5745 	ptep += (va >> PGSHIFT) & 0x3ff;
5746 #else
5747 	ptep += l2pte_index(va);
5748 #endif
5749 	l2pte_set(ptep, npte, 0);
5750 	PTE_SYNC(ptep);
5751 }
5752 
5753 /*
5754  * pmap_link_l2pt:
5755  *
5756  *	Link the L2 page table specified by "l2pv" into the L1
5757  *	page table at the slot for "va".
5758  */
5759 void
5760 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
5761 {
5762 	pd_entry_t *pde = (pd_entry_t *) l1pt, proto;
5763 	u_int slot = L1_IDX(va);
5764 
5765 #ifndef ARM32_NEW_VM_LAYOUT
5766 	KASSERT((va & ((L1_S_SIZE * 4) - 1)) == 0);
5767 	KASSERT((l2pv->pv_pa & PGOFSET) == 0);
5768 #endif
5769 
5770 	proto = L1_S_DOM(PMAP_DOMAIN_KERNEL) | L1_C_PROTO;
5771 
5772 	pde[slot + 0] = proto | (l2pv->pv_pa + 0x000);
5773 #ifdef ARM32_NEW_VM_LAYOUT
5774 	PTE_SYNC(&pde[slot]);
5775 #else
5776 	for (u_int off = 0, i = 0; off < PAGE_SIZE; off += L2_T_SIZE, i++) {
5777 		pde[slot + i] = proto | (l2pv->pv_pa + off);
5778 	}
5779 	PTE_SYNC_RANGE(&pde[slot + 0], PAGE_SIZE / L2_T_SIZE);
5780 #endif
5781 
5782 	SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
5783 }
5784 
5785 /*
5786  * pmap_map_chunk:
5787  *
5788  *	Map a chunk of memory using the most efficient mappings
5789  *	possible (section, large page, small page) into the
5790  *	provided L1 and L2 tables at the specified virtual address.
5791  */
5792 vsize_t
5793 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
5794     int prot, int cache)
5795 {
5796 	pd_entry_t *pdep = (pd_entry_t *) l1pt;
5797 	pt_entry_t *pte, f1, f2s, f2l;
5798 	vsize_t resid;
5799 	int i;
5800 
5801 	resid = (size + (PAGE_SIZE - 1)) & ~(PAGE_SIZE - 1);
5802 
5803 	if (l1pt == 0)
5804 		panic("pmap_map_chunk: no L1 table provided");
5805 
5806 #ifdef VERBOSE_INIT_ARM
5807 	printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
5808 	    "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
5809 #endif
5810 
5811 	switch (cache) {
5812 	case PTE_NOCACHE:
5813 	default:
5814 		f1 = 0;
5815 		f2l = 0;
5816 		f2s = 0;
5817 		break;
5818 
5819 	case PTE_CACHE:
5820 		f1 = pte_l1_s_cache_mode;
5821 		f2l = pte_l2_l_cache_mode;
5822 		f2s = pte_l2_s_cache_mode;
5823 		break;
5824 
5825 	case PTE_PAGETABLE:
5826 		f1 = pte_l1_s_cache_mode_pt;
5827 		f2l = pte_l2_l_cache_mode_pt;
5828 		f2s = pte_l2_s_cache_mode_pt;
5829 		break;
5830 	}
5831 
5832 	size = resid;
5833 
5834 	while (resid > 0) {
5835 		size_t l1idx = L1_IDX(va);
5836 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
5837 		/* See if we can use a supersection mapping. */
5838 		if (L1_SS_PROTO && L1_SS_MAPPABLE_P(va, pa, resid)) {
5839 			/* Supersection are always domain 0 */
5840 			pd_entry_t pde = L1_SS_PROTO | pa |
5841 			    L1_S_PROT(PTE_KERNEL, prot) | f1;
5842 #ifdef VERBOSE_INIT_ARM
5843 			printf("sS");
5844 #endif
5845 			for (size_t s = l1idx,
5846 			     e = s + L1_SS_SIZE / L1_S_SIZE;
5847 			     s < e;
5848 			     s++) {
5849 				pdep[s] = pde;
5850 				PTE_SYNC(&pdep[s]);
5851 			}
5852 			va += L1_SS_SIZE;
5853 			pa += L1_SS_SIZE;
5854 			resid -= L1_SS_SIZE;
5855 			continue;
5856 		}
5857 #endif
5858 		/* See if we can use a section mapping. */
5859 		if (L1_S_MAPPABLE_P(va, pa, resid)) {
5860 #ifdef VERBOSE_INIT_ARM
5861 			printf("S");
5862 #endif
5863 			pdep[l1idx] = L1_S_PROTO | pa |
5864 			    L1_S_PROT(PTE_KERNEL, prot) | f1 |
5865 			    L1_S_DOM(PMAP_DOMAIN_KERNEL);
5866 			PTE_SYNC(&pdep[l1idx]);
5867 			va += L1_S_SIZE;
5868 			pa += L1_S_SIZE;
5869 			resid -= L1_S_SIZE;
5870 			continue;
5871 		}
5872 
5873 		/*
5874 		 * Ok, we're going to use an L2 table.  Make sure
5875 		 * one is actually in the corresponding L1 slot
5876 		 * for the current VA.
5877 		 */
5878 		if ((pdep[l1idx] & L1_TYPE_MASK) != L1_TYPE_C)
5879 			panic("pmap_map_chunk: no L2 table for VA 0x%08lx", va);
5880 
5881 #ifndef ARM32_NEW_VM_LAYOUT
5882 		pte = (pt_entry_t *)
5883 		    kernel_pt_lookup(pdep[l1idx] & L2_S_FRAME);
5884 #else
5885 		pte = (pt_entry_t *) kernel_pt_lookup(
5886 		    pdep[l1idx] & L1_C_ADDR_MASK);
5887 #endif
5888 		if (pte == NULL)
5889 			panic("pmap_map_chunk: can't find L2 table for VA"
5890 			    "0x%08lx", va);
5891 
5892 		/* See if we can use a L2 large page mapping. */
5893 		if (L2_L_MAPPABLE_P(va, pa, resid)) {
5894 #ifdef VERBOSE_INIT_ARM
5895 			printf("L");
5896 #endif
5897 			for (i = 0; i < 16; i++) {
5898 #ifndef ARM32_NEW_VM_LAYOUT
5899 				pte[((va >> PGSHIFT) & 0x3f0) + i] =
5900 				    L2_L_PROTO | pa |
5901 				    L2_L_PROT(PTE_KERNEL, prot) | f2l;
5902 				PTE_SYNC(&pte[((va >> PGSHIFT) & 0x3f0) + i]);
5903 #else
5904 				pte[l2pte_index(va) + i] =
5905 				    L2_L_PROTO | pa |
5906 				    L2_L_PROT(PTE_KERNEL, prot) | f2l;
5907 				PTE_SYNC(&pte[l2pte_index(va) + i]);
5908 #endif
5909 			}
5910 			va += L2_L_SIZE;
5911 			pa += L2_L_SIZE;
5912 			resid -= L2_L_SIZE;
5913 			continue;
5914 		}
5915 
5916 		/* Use a small page mapping. */
5917 #ifdef VERBOSE_INIT_ARM
5918 		printf("P");
5919 #endif
5920 		pt_entry_t npte =
5921 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | f2s;
5922 #ifndef ARM32_NEW_VM_LAYOUT
5923 		pt_entry_t *ptep = &pte[(va >> PGSHIFT) & 0x3ff];
5924 #else
5925 		pt_entry_t *ptep = &pte[l2pte_index(va)];
5926 #endif
5927 		l2pte_set(ptep, npte, 0);
5928 		PTE_SYNC(ptep);
5929 		va += PAGE_SIZE;
5930 		pa += PAGE_SIZE;
5931 		resid -= PAGE_SIZE;
5932 	}
5933 #ifdef VERBOSE_INIT_ARM
5934 	printf("\n");
5935 #endif
5936 	return (size);
5937 }
5938 
5939 /********************** Static device map routines ***************************/
5940 
5941 static const struct pmap_devmap *pmap_devmap_table;
5942 
5943 /*
5944  * Register the devmap table.  This is provided in case early console
5945  * initialization needs to register mappings created by bootstrap code
5946  * before pmap_devmap_bootstrap() is called.
5947  */
5948 void
5949 pmap_devmap_register(const struct pmap_devmap *table)
5950 {
5951 
5952 	pmap_devmap_table = table;
5953 }
5954 
5955 /*
5956  * Map all of the static regions in the devmap table, and remember
5957  * the devmap table so other parts of the kernel can look up entries
5958  * later.
5959  */
5960 void
5961 pmap_devmap_bootstrap(vaddr_t l1pt, const struct pmap_devmap *table)
5962 {
5963 	int i;
5964 
5965 	pmap_devmap_table = table;
5966 
5967 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
5968 #ifdef VERBOSE_INIT_ARM
5969 		printf("devmap: %08lx -> %08lx @ %08lx\n",
5970 		    pmap_devmap_table[i].pd_pa,
5971 		    pmap_devmap_table[i].pd_pa +
5972 			pmap_devmap_table[i].pd_size - 1,
5973 		    pmap_devmap_table[i].pd_va);
5974 #endif
5975 		pmap_map_chunk(l1pt, pmap_devmap_table[i].pd_va,
5976 		    pmap_devmap_table[i].pd_pa,
5977 		    pmap_devmap_table[i].pd_size,
5978 		    pmap_devmap_table[i].pd_prot,
5979 		    pmap_devmap_table[i].pd_cache);
5980 	}
5981 }
5982 
5983 const struct pmap_devmap *
5984 pmap_devmap_find_pa(paddr_t pa, psize_t size)
5985 {
5986 	uint64_t endpa;
5987 	int i;
5988 
5989 	if (pmap_devmap_table == NULL)
5990 		return (NULL);
5991 
5992 	endpa = (uint64_t)pa + (uint64_t)(size - 1);
5993 
5994 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
5995 		if (pa >= pmap_devmap_table[i].pd_pa &&
5996 		    endpa <= (uint64_t)pmap_devmap_table[i].pd_pa +
5997 			     (uint64_t)(pmap_devmap_table[i].pd_size - 1))
5998 			return (&pmap_devmap_table[i]);
5999 	}
6000 
6001 	return (NULL);
6002 }
6003 
6004 const struct pmap_devmap *
6005 pmap_devmap_find_va(vaddr_t va, vsize_t size)
6006 {
6007 	int i;
6008 
6009 	if (pmap_devmap_table == NULL)
6010 		return (NULL);
6011 
6012 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
6013 		if (va >= pmap_devmap_table[i].pd_va &&
6014 		    va + size - 1 <= pmap_devmap_table[i].pd_va +
6015 				     pmap_devmap_table[i].pd_size - 1)
6016 			return (&pmap_devmap_table[i]);
6017 	}
6018 
6019 	return (NULL);
6020 }
6021 
6022 /********************** PTE initialization routines **************************/
6023 
6024 /*
6025  * These routines are called when the CPU type is identified to set up
6026  * the PTE prototypes, cache modes, etc.
6027  *
6028  * The variables are always here, just in case modules need to reference
6029  * them (though, they shouldn't).
6030  */
6031 
6032 pt_entry_t	pte_l1_s_cache_mode;
6033 pt_entry_t	pte_l1_s_wc_mode;
6034 pt_entry_t	pte_l1_s_cache_mode_pt;
6035 pt_entry_t	pte_l1_s_cache_mask;
6036 
6037 pt_entry_t	pte_l2_l_cache_mode;
6038 pt_entry_t	pte_l2_l_wc_mode;
6039 pt_entry_t	pte_l2_l_cache_mode_pt;
6040 pt_entry_t	pte_l2_l_cache_mask;
6041 
6042 pt_entry_t	pte_l2_s_cache_mode;
6043 pt_entry_t	pte_l2_s_wc_mode;
6044 pt_entry_t	pte_l2_s_cache_mode_pt;
6045 pt_entry_t	pte_l2_s_cache_mask;
6046 
6047 pt_entry_t	pte_l1_s_prot_u;
6048 pt_entry_t	pte_l1_s_prot_w;
6049 pt_entry_t	pte_l1_s_prot_ro;
6050 pt_entry_t	pte_l1_s_prot_mask;
6051 
6052 pt_entry_t	pte_l2_s_prot_u;
6053 pt_entry_t	pte_l2_s_prot_w;
6054 pt_entry_t	pte_l2_s_prot_ro;
6055 pt_entry_t	pte_l2_s_prot_mask;
6056 
6057 pt_entry_t	pte_l2_l_prot_u;
6058 pt_entry_t	pte_l2_l_prot_w;
6059 pt_entry_t	pte_l2_l_prot_ro;
6060 pt_entry_t	pte_l2_l_prot_mask;
6061 
6062 pt_entry_t	pte_l1_ss_proto;
6063 pt_entry_t	pte_l1_s_proto;
6064 pt_entry_t	pte_l1_c_proto;
6065 pt_entry_t	pte_l2_s_proto;
6066 
6067 void		(*pmap_copy_page_func)(paddr_t, paddr_t);
6068 void		(*pmap_zero_page_func)(paddr_t);
6069 
6070 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
6071 void
6072 pmap_pte_init_generic(void)
6073 {
6074 
6075 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
6076 	pte_l1_s_wc_mode = L1_S_B;
6077 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_generic;
6078 
6079 	pte_l2_l_cache_mode = L2_B|L2_C;
6080 	pte_l2_l_wc_mode = L2_B;
6081 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_generic;
6082 
6083 	pte_l2_s_cache_mode = L2_B|L2_C;
6084 	pte_l2_s_wc_mode = L2_B;
6085 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_generic;
6086 
6087 	/*
6088 	 * If we have a write-through cache, set B and C.  If
6089 	 * we have a write-back cache, then we assume setting
6090 	 * only C will make those pages write-through (except for those
6091 	 * Cortex CPUs which can read the L1 caches).
6092 	 */
6093 	if (cpufuncs.cf_dcache_wb_range == (void *) cpufunc_nullop
6094 #if ARM_MMU_V7 > 0
6095 	    || CPU_ID_CORTEX_P(curcpu()->ci_arm_cpuid)
6096 #endif
6097 #if ARM_MMU_V6 > 0
6098 	    || CPU_ID_ARM11_P(curcpu()->ci_arm_cpuid) /* arm116 errata 399234 */
6099 #endif
6100 	    || false) {
6101 		pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
6102 		pte_l2_l_cache_mode_pt = L2_B|L2_C;
6103 		pte_l2_s_cache_mode_pt = L2_B|L2_C;
6104 	} else {
6105 		pte_l1_s_cache_mode_pt = L1_S_C;	/* write through */
6106 		pte_l2_l_cache_mode_pt = L2_C;		/* write through */
6107 		pte_l2_s_cache_mode_pt = L2_C;		/* write through */
6108 	}
6109 
6110 	pte_l1_s_prot_u = L1_S_PROT_U_generic;
6111 	pte_l1_s_prot_w = L1_S_PROT_W_generic;
6112 	pte_l1_s_prot_ro = L1_S_PROT_RO_generic;
6113 	pte_l1_s_prot_mask = L1_S_PROT_MASK_generic;
6114 
6115 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
6116 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
6117 	pte_l2_s_prot_ro = L2_S_PROT_RO_generic;
6118 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
6119 
6120 	pte_l2_l_prot_u = L2_L_PROT_U_generic;
6121 	pte_l2_l_prot_w = L2_L_PROT_W_generic;
6122 	pte_l2_l_prot_ro = L2_L_PROT_RO_generic;
6123 	pte_l2_l_prot_mask = L2_L_PROT_MASK_generic;
6124 
6125 	pte_l1_ss_proto = L1_SS_PROTO_generic;
6126 	pte_l1_s_proto = L1_S_PROTO_generic;
6127 	pte_l1_c_proto = L1_C_PROTO_generic;
6128 	pte_l2_s_proto = L2_S_PROTO_generic;
6129 
6130 	pmap_copy_page_func = pmap_copy_page_generic;
6131 	pmap_zero_page_func = pmap_zero_page_generic;
6132 }
6133 
6134 #if defined(CPU_ARM8)
6135 void
6136 pmap_pte_init_arm8(void)
6137 {
6138 
6139 	/*
6140 	 * ARM8 is compatible with generic, but we need to use
6141 	 * the page tables uncached.
6142 	 */
6143 	pmap_pte_init_generic();
6144 
6145 	pte_l1_s_cache_mode_pt = 0;
6146 	pte_l2_l_cache_mode_pt = 0;
6147 	pte_l2_s_cache_mode_pt = 0;
6148 }
6149 #endif /* CPU_ARM8 */
6150 
6151 #if defined(CPU_ARM9) && defined(ARM9_CACHE_WRITE_THROUGH)
6152 void
6153 pmap_pte_init_arm9(void)
6154 {
6155 
6156 	/*
6157 	 * ARM9 is compatible with generic, but we want to use
6158 	 * write-through caching for now.
6159 	 */
6160 	pmap_pte_init_generic();
6161 
6162 	pte_l1_s_cache_mode = L1_S_C;
6163 	pte_l2_l_cache_mode = L2_C;
6164 	pte_l2_s_cache_mode = L2_C;
6165 
6166 	pte_l1_s_wc_mode = L1_S_B;
6167 	pte_l2_l_wc_mode = L2_B;
6168 	pte_l2_s_wc_mode = L2_B;
6169 
6170 	pte_l1_s_cache_mode_pt = L1_S_C;
6171 	pte_l2_l_cache_mode_pt = L2_C;
6172 	pte_l2_s_cache_mode_pt = L2_C;
6173 }
6174 #endif /* CPU_ARM9 && ARM9_CACHE_WRITE_THROUGH */
6175 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
6176 
6177 #if defined(CPU_ARM10)
6178 void
6179 pmap_pte_init_arm10(void)
6180 {
6181 
6182 	/*
6183 	 * ARM10 is compatible with generic, but we want to use
6184 	 * write-through caching for now.
6185 	 */
6186 	pmap_pte_init_generic();
6187 
6188 	pte_l1_s_cache_mode = L1_S_B | L1_S_C;
6189 	pte_l2_l_cache_mode = L2_B | L2_C;
6190 	pte_l2_s_cache_mode = L2_B | L2_C;
6191 
6192 	pte_l1_s_cache_mode = L1_S_B;
6193 	pte_l2_l_cache_mode = L2_B;
6194 	pte_l2_s_cache_mode = L2_B;
6195 
6196 	pte_l1_s_cache_mode_pt = L1_S_C;
6197 	pte_l2_l_cache_mode_pt = L2_C;
6198 	pte_l2_s_cache_mode_pt = L2_C;
6199 
6200 }
6201 #endif /* CPU_ARM10 */
6202 
6203 #if defined(CPU_ARM11) && defined(ARM11_CACHE_WRITE_THROUGH)
6204 void
6205 pmap_pte_init_arm11(void)
6206 {
6207 
6208 	/*
6209 	 * ARM11 is compatible with generic, but we want to use
6210 	 * write-through caching for now.
6211 	 */
6212 	pmap_pte_init_generic();
6213 
6214 	pte_l1_s_cache_mode = L1_S_C;
6215 	pte_l2_l_cache_mode = L2_C;
6216 	pte_l2_s_cache_mode = L2_C;
6217 
6218 	pte_l1_s_wc_mode = L1_S_B;
6219 	pte_l2_l_wc_mode = L2_B;
6220 	pte_l2_s_wc_mode = L2_B;
6221 
6222 	pte_l1_s_cache_mode_pt = L1_S_C;
6223 	pte_l2_l_cache_mode_pt = L2_C;
6224 	pte_l2_s_cache_mode_pt = L2_C;
6225 }
6226 #endif /* CPU_ARM11 && ARM11_CACHE_WRITE_THROUGH */
6227 
6228 #if ARM_MMU_SA1 == 1
6229 void
6230 pmap_pte_init_sa1(void)
6231 {
6232 
6233 	/*
6234 	 * The StrongARM SA-1 cache does not have a write-through
6235 	 * mode.  So, do the generic initialization, then reset
6236 	 * the page table cache mode to B=1,C=1, and note that
6237 	 * the PTEs need to be sync'd.
6238 	 */
6239 	pmap_pte_init_generic();
6240 
6241 	pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
6242 	pte_l2_l_cache_mode_pt = L2_B|L2_C;
6243 	pte_l2_s_cache_mode_pt = L2_B|L2_C;
6244 
6245 	pmap_needs_pte_sync = 1;
6246 }
6247 #endif /* ARM_MMU_SA1 == 1*/
6248 
6249 #if ARM_MMU_XSCALE == 1
6250 #if (ARM_NMMUS > 1)
6251 static u_int xscale_use_minidata;
6252 #endif
6253 
6254 void
6255 pmap_pte_init_xscale(void)
6256 {
6257 	uint32_t auxctl;
6258 	int write_through = 0;
6259 
6260 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
6261 	pte_l1_s_wc_mode = L1_S_B;
6262 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_xscale;
6263 
6264 	pte_l2_l_cache_mode = L2_B|L2_C;
6265 	pte_l2_l_wc_mode = L2_B;
6266 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_xscale;
6267 
6268 	pte_l2_s_cache_mode = L2_B|L2_C;
6269 	pte_l2_s_wc_mode = L2_B;
6270 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_xscale;
6271 
6272 	pte_l1_s_cache_mode_pt = L1_S_C;
6273 	pte_l2_l_cache_mode_pt = L2_C;
6274 	pte_l2_s_cache_mode_pt = L2_C;
6275 
6276 #ifdef XSCALE_CACHE_READ_WRITE_ALLOCATE
6277 	/*
6278 	 * The XScale core has an enhanced mode where writes that
6279 	 * miss the cache cause a cache line to be allocated.  This
6280 	 * is significantly faster than the traditional, write-through
6281 	 * behavior of this case.
6282 	 */
6283 	pte_l1_s_cache_mode |= L1_S_XS_TEX(TEX_XSCALE_X);
6284 	pte_l2_l_cache_mode |= L2_XS_L_TEX(TEX_XSCALE_X);
6285 	pte_l2_s_cache_mode |= L2_XS_T_TEX(TEX_XSCALE_X);
6286 #endif /* XSCALE_CACHE_READ_WRITE_ALLOCATE */
6287 
6288 #ifdef XSCALE_CACHE_WRITE_THROUGH
6289 	/*
6290 	 * Some versions of the XScale core have various bugs in
6291 	 * their cache units, the work-around for which is to run
6292 	 * the cache in write-through mode.  Unfortunately, this
6293 	 * has a major (negative) impact on performance.  So, we
6294 	 * go ahead and run fast-and-loose, in the hopes that we
6295 	 * don't line up the planets in a way that will trip the
6296 	 * bugs.
6297 	 *
6298 	 * However, we give you the option to be slow-but-correct.
6299 	 */
6300 	write_through = 1;
6301 #elif defined(XSCALE_CACHE_WRITE_BACK)
6302 	/* force write back cache mode */
6303 	write_through = 0;
6304 #elif defined(CPU_XSCALE_PXA250) || defined(CPU_XSCALE_PXA270)
6305 	/*
6306 	 * Intel PXA2[15]0 processors are known to have a bug in
6307 	 * write-back cache on revision 4 and earlier (stepping
6308 	 * A[01] and B[012]).  Fixed for C0 and later.
6309 	 */
6310 	{
6311 		uint32_t id, type;
6312 
6313 		id = cpufunc_id();
6314 		type = id & ~(CPU_ID_XSCALE_COREREV_MASK|CPU_ID_REVISION_MASK);
6315 
6316 		if (type == CPU_ID_PXA250 || type == CPU_ID_PXA210) {
6317 			if ((id & CPU_ID_REVISION_MASK) < 5) {
6318 				/* write through for stepping A0-1 and B0-2 */
6319 				write_through = 1;
6320 			}
6321 		}
6322 	}
6323 #endif /* XSCALE_CACHE_WRITE_THROUGH */
6324 
6325 	if (write_through) {
6326 		pte_l1_s_cache_mode = L1_S_C;
6327 		pte_l2_l_cache_mode = L2_C;
6328 		pte_l2_s_cache_mode = L2_C;
6329 	}
6330 
6331 #if (ARM_NMMUS > 1)
6332 	xscale_use_minidata = 1;
6333 #endif
6334 
6335 	pte_l1_s_prot_u = L1_S_PROT_U_xscale;
6336 	pte_l1_s_prot_w = L1_S_PROT_W_xscale;
6337 	pte_l1_s_prot_ro = L1_S_PROT_RO_xscale;
6338 	pte_l1_s_prot_mask = L1_S_PROT_MASK_xscale;
6339 
6340 	pte_l2_s_prot_u = L2_S_PROT_U_xscale;
6341 	pte_l2_s_prot_w = L2_S_PROT_W_xscale;
6342 	pte_l2_s_prot_ro = L2_S_PROT_RO_xscale;
6343 	pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale;
6344 
6345 	pte_l2_l_prot_u = L2_L_PROT_U_xscale;
6346 	pte_l2_l_prot_w = L2_L_PROT_W_xscale;
6347 	pte_l2_l_prot_ro = L2_L_PROT_RO_xscale;
6348 	pte_l2_l_prot_mask = L2_L_PROT_MASK_xscale;
6349 
6350 	pte_l1_ss_proto = L1_SS_PROTO_xscale;
6351 	pte_l1_s_proto = L1_S_PROTO_xscale;
6352 	pte_l1_c_proto = L1_C_PROTO_xscale;
6353 	pte_l2_s_proto = L2_S_PROTO_xscale;
6354 
6355 	pmap_copy_page_func = pmap_copy_page_xscale;
6356 	pmap_zero_page_func = pmap_zero_page_xscale;
6357 
6358 	/*
6359 	 * Disable ECC protection of page table access, for now.
6360 	 */
6361 	__asm volatile("mrc p15, 0, %0, c1, c0, 1" : "=r" (auxctl));
6362 	auxctl &= ~XSCALE_AUXCTL_P;
6363 	__asm volatile("mcr p15, 0, %0, c1, c0, 1" : : "r" (auxctl));
6364 }
6365 
6366 /*
6367  * xscale_setup_minidata:
6368  *
6369  *	Set up the mini-data cache clean area.  We require the
6370  *	caller to allocate the right amount of physically and
6371  *	virtually contiguous space.
6372  */
6373 void
6374 xscale_setup_minidata(vaddr_t l1pt, vaddr_t va, paddr_t pa)
6375 {
6376 	extern vaddr_t xscale_minidata_clean_addr;
6377 	extern vsize_t xscale_minidata_clean_size; /* already initialized */
6378 	pd_entry_t *pde = (pd_entry_t *) l1pt;
6379 	vsize_t size;
6380 	uint32_t auxctl;
6381 
6382 	xscale_minidata_clean_addr = va;
6383 
6384 	/* Round it to page size. */
6385 	size = (xscale_minidata_clean_size + L2_S_OFFSET) & L2_S_FRAME;
6386 
6387 	for (; size != 0;
6388 	     va += L2_S_SIZE, pa += L2_S_SIZE, size -= L2_S_SIZE) {
6389 		const size_t l1idx = L1_IDX(va);
6390 #ifndef ARM32_NEW_VM_LAYOUT
6391 		pt_entry_t *ptep = (pt_entry_t *)
6392 		    kernel_pt_lookup(pde[l1idx] & L2_S_FRAME);
6393 #else
6394 		pt_entry_t *ptep = (pt_entry_t *) kernel_pt_lookup(
6395 		    pde[l1idx] & L1_C_ADDR_MASK);
6396 #endif
6397 		if (ptep == NULL)
6398 			panic("xscale_setup_minidata: can't find L2 table for "
6399 			    "VA 0x%08lx", va);
6400 
6401 #ifndef ARM32_NEW_VM_LAYOUT
6402 		ptep += (va >> PGSHIFT) & 0x3ff;
6403 #else
6404 		ptep += l2pte_index(va);
6405 #endif
6406 		pt_entry_t opte = *ptep;
6407 		l2pte_set(ptep,
6408 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, VM_PROT_READ)
6409 		    | L2_C | L2_XS_T_TEX(TEX_XSCALE_X), opte);
6410 	}
6411 
6412 	/*
6413 	 * Configure the mini-data cache for write-back with
6414 	 * read/write-allocate.
6415 	 *
6416 	 * NOTE: In order to reconfigure the mini-data cache, we must
6417 	 * make sure it contains no valid data!  In order to do that,
6418 	 * we must issue a global data cache invalidate command!
6419 	 *
6420 	 * WE ASSUME WE ARE RUNNING UN-CACHED WHEN THIS ROUTINE IS CALLED!
6421 	 * THIS IS VERY IMPORTANT!
6422 	 */
6423 
6424 	/* Invalidate data and mini-data. */
6425 	__asm volatile("mcr p15, 0, %0, c7, c6, 0" : : "r" (0));
6426 	__asm volatile("mrc p15, 0, %0, c1, c0, 1" : "=r" (auxctl));
6427 	auxctl = (auxctl & ~XSCALE_AUXCTL_MD_MASK) | XSCALE_AUXCTL_MD_WB_RWA;
6428 	__asm volatile("mcr p15, 0, %0, c1, c0, 1" : : "r" (auxctl));
6429 }
6430 
6431 /*
6432  * Change the PTEs for the specified kernel mappings such that they
6433  * will use the mini data cache instead of the main data cache.
6434  */
6435 void
6436 pmap_uarea(vaddr_t va)
6437 {
6438 	vaddr_t next_bucket, eva;
6439 
6440 #if (ARM_NMMUS > 1)
6441 	if (xscale_use_minidata == 0)
6442 		return;
6443 #endif
6444 
6445 	eva = va + USPACE;
6446 
6447 	while (va < eva) {
6448 		next_bucket = L2_NEXT_BUCKET(va);
6449 		if (next_bucket > eva)
6450 			next_bucket = eva;
6451 
6452 		struct l2_bucket *l2b = pmap_get_l2_bucket(pmap_kernel(), va);
6453 		KDASSERT(l2b != NULL);
6454 
6455 		pt_entry_t * const sptep = &l2b->l2b_kva[l2pte_index(va)];
6456 		pt_entry_t *ptep = sptep;
6457 
6458 		while (va < next_bucket) {
6459 			const pt_entry_t opte = *ptep;
6460 			if (!l2pte_minidata(opte)) {
6461 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
6462 				cpu_tlb_flushD_SE(va);
6463 				l2pte_set(ptep, opte & ~L2_B, opte);
6464 			}
6465 			ptep += PAGE_SIZE / L2_S_SIZE;
6466 			va += PAGE_SIZE;
6467 		}
6468 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
6469 	}
6470 	cpu_cpwait();
6471 }
6472 #endif /* ARM_MMU_XSCALE == 1 */
6473 
6474 
6475 #if defined(CPU_ARM11MPCORE)
6476 
6477 void
6478 pmap_pte_init_arm11mpcore(void)
6479 {
6480 
6481 	/* cache mode is controlled by 5 bits (B, C, TEX) */
6482 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_armv6;
6483 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_armv6;
6484 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
6485 	/* use extended small page (without APn, with TEX) */
6486 	pte_l2_s_cache_mask = L2_XS_CACHE_MASK_armv6;
6487 #else
6488 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_armv6c;
6489 #endif
6490 
6491 	/* write-back, write-allocate */
6492 	pte_l1_s_cache_mode = L1_S_C | L1_S_B | L1_S_V6_TEX(0x01);
6493 	pte_l2_l_cache_mode = L2_C | L2_B | L2_V6_L_TEX(0x01);
6494 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
6495 	pte_l2_s_cache_mode = L2_C | L2_B | L2_V6_XS_TEX(0x01);
6496 #else
6497 	/* no TEX. read-allocate */
6498 	pte_l2_s_cache_mode = L2_C | L2_B;
6499 #endif
6500 	/*
6501 	 * write-back, write-allocate for page tables.
6502 	 */
6503 	pte_l1_s_cache_mode_pt = L1_S_C | L1_S_B | L1_S_V6_TEX(0x01);
6504 	pte_l2_l_cache_mode_pt = L2_C | L2_B | L2_V6_L_TEX(0x01);
6505 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
6506 	pte_l2_s_cache_mode_pt = L2_C | L2_B | L2_V6_XS_TEX(0x01);
6507 #else
6508 	pte_l2_s_cache_mode_pt = L2_C | L2_B;
6509 #endif
6510 
6511 	pte_l1_s_prot_u = L1_S_PROT_U_armv6;
6512 	pte_l1_s_prot_w = L1_S_PROT_W_armv6;
6513 	pte_l1_s_prot_ro = L1_S_PROT_RO_armv6;
6514 	pte_l1_s_prot_mask = L1_S_PROT_MASK_armv6;
6515 
6516 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
6517 	pte_l2_s_prot_u = L2_S_PROT_U_armv6n;
6518 	pte_l2_s_prot_w = L2_S_PROT_W_armv6n;
6519 	pte_l2_s_prot_ro = L2_S_PROT_RO_armv6n;
6520 	pte_l2_s_prot_mask = L2_S_PROT_MASK_armv6n;
6521 
6522 #else
6523 	/* with AP[0..3] */
6524 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
6525 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
6526 	pte_l2_s_prot_ro = L2_S_PROT_RO_generic;
6527 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
6528 #endif
6529 
6530 #ifdef	ARM11MPCORE_COMPAT_MMU
6531 	/* with AP[0..3] */
6532 	pte_l2_l_prot_u = L2_L_PROT_U_generic;
6533 	pte_l2_l_prot_w = L2_L_PROT_W_generic;
6534 	pte_l2_l_prot_ro = L2_L_PROT_RO_generic;
6535 	pte_l2_l_prot_mask = L2_L_PROT_MASK_generic;
6536 
6537 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
6538 	pte_l1_s_proto = L1_S_PROTO_armv6;
6539 	pte_l1_c_proto = L1_C_PROTO_armv6;
6540 	pte_l2_s_proto = L2_S_PROTO_armv6c;
6541 #else
6542 	pte_l2_l_prot_u = L2_L_PROT_U_armv6n;
6543 	pte_l2_l_prot_w = L2_L_PROT_W_armv6n;
6544 	pte_l2_l_prot_ro = L2_L_PROT_RO_armv6n;
6545 	pte_l2_l_prot_mask = L2_L_PROT_MASK_armv6n;
6546 
6547 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
6548 	pte_l1_s_proto = L1_S_PROTO_armv6;
6549 	pte_l1_c_proto = L1_C_PROTO_armv6;
6550 	pte_l2_s_proto = L2_S_PROTO_armv6n;
6551 #endif
6552 
6553 	pmap_copy_page_func = pmap_copy_page_generic;
6554 	pmap_zero_page_func = pmap_zero_page_generic;
6555 	pmap_needs_pte_sync = 1;
6556 }
6557 #endif	/* CPU_ARM11MPCORE */
6558 
6559 
6560 #if ARM_MMU_V7 == 1
6561 void
6562 pmap_pte_init_armv7(void)
6563 {
6564 	/*
6565 	 * The ARMv7-A MMU is mostly compatible with generic. If the
6566 	 * AP field is zero, that now means "no access" rather than
6567 	 * read-only. The prototypes are a little different because of
6568 	 * the XN bit.
6569 	 */
6570 	pmap_pte_init_generic();
6571 
6572 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_armv7;
6573 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_armv7;
6574 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_armv7;
6575 
6576 	if (CPU_ID_CORTEX_A9_P(curcpu()->ci_arm_cpuid)) {
6577 		/*
6578 		 * write-back, no write-allocate, shareable for normal pages.
6579 		 */
6580 		pte_l1_s_cache_mode = L1_S_C | L1_S_B | L1_S_V6_S;
6581 		pte_l2_l_cache_mode = L2_C | L2_B | L2_XS_S;
6582 		pte_l2_s_cache_mode = L2_C | L2_B | L2_XS_S;
6583 
6584 		/*
6585 		 * write-back, no write-allocate, shareable for page tables.
6586 		 */
6587 		pte_l1_s_cache_mode_pt = L1_S_C | L1_S_B | L1_S_V6_S;
6588 		pte_l2_l_cache_mode_pt = L2_C | L2_B | L2_XS_S;
6589 		pte_l2_s_cache_mode_pt = L2_C | L2_B | L2_XS_S;
6590 	}
6591 
6592 	pte_l1_s_prot_u = L1_S_PROT_U_armv7;
6593 	pte_l1_s_prot_w = L1_S_PROT_W_armv7;
6594 	pte_l1_s_prot_ro = L1_S_PROT_RO_armv7;
6595 	pte_l1_s_prot_mask = L1_S_PROT_MASK_armv7;
6596 
6597 	pte_l2_s_prot_u = L2_S_PROT_U_armv7;
6598 	pte_l2_s_prot_w = L2_S_PROT_W_armv7;
6599 	pte_l2_s_prot_ro = L2_S_PROT_RO_armv7;
6600 	pte_l2_s_prot_mask = L2_S_PROT_MASK_armv7;
6601 
6602 	pte_l2_l_prot_u = L2_L_PROT_U_armv7;
6603 	pte_l2_l_prot_w = L2_L_PROT_W_armv7;
6604 	pte_l2_l_prot_ro = L2_L_PROT_RO_armv7;
6605 	pte_l2_l_prot_mask = L2_L_PROT_MASK_armv7;
6606 
6607 	pte_l1_ss_proto = L1_SS_PROTO_armv7;
6608 	pte_l1_s_proto = L1_S_PROTO_armv7;
6609 	pte_l1_c_proto = L1_C_PROTO_armv7;
6610 	pte_l2_s_proto = L2_S_PROTO_armv7;
6611 
6612 	pmap_needs_pte_sync = 1;
6613 }
6614 #endif /* ARM_MMU_V7 */
6615 
6616 /*
6617  * return the PA of the current L1 table, for use when handling a crash dump
6618  */
6619 uint32_t pmap_kernel_L1_addr(void)
6620 {
6621 	return pmap_kernel()->pm_l1->l1_physaddr;
6622 }
6623 
6624 #if defined(DDB)
6625 /*
6626  * A couple of ddb-callable functions for dumping pmaps
6627  */
6628 void pmap_dump_all(void);
6629 void pmap_dump(pmap_t);
6630 
6631 void
6632 pmap_dump_all(void)
6633 {
6634 	pmap_t pm;
6635 
6636 	LIST_FOREACH(pm, &pmap_pmaps, pm_list) {
6637 		if (pm == pmap_kernel())
6638 			continue;
6639 		pmap_dump(pm);
6640 		printf("\n");
6641 	}
6642 }
6643 
6644 static pt_entry_t ncptes[64];
6645 static void pmap_dump_ncpg(pmap_t);
6646 
6647 void
6648 pmap_dump(pmap_t pm)
6649 {
6650 	struct l2_dtable *l2;
6651 	struct l2_bucket *l2b;
6652 	pt_entry_t *ptep, pte;
6653 	vaddr_t l2_va, l2b_va, va;
6654 	int i, j, k, occ, rows = 0;
6655 
6656 	if (pm == pmap_kernel())
6657 		printf("pmap_kernel (%p): ", pm);
6658 	else
6659 		printf("user pmap (%p): ", pm);
6660 
6661 	printf("domain %d, l1 at %p\n", pmap_domain(pm), pmap_l1_kva(pm));
6662 
6663 	l2_va = 0;
6664 	for (i = 0; i < L2_SIZE; i++, l2_va += 0x01000000) {
6665 		l2 = pm->pm_l2[i];
6666 
6667 		if (l2 == NULL || l2->l2_occupancy == 0)
6668 			continue;
6669 
6670 		l2b_va = l2_va;
6671 		for (j = 0; j < L2_BUCKET_SIZE; j++, l2b_va += 0x00100000) {
6672 			l2b = &l2->l2_bucket[j];
6673 
6674 			if (l2b->l2b_occupancy == 0 || l2b->l2b_kva == NULL)
6675 				continue;
6676 
6677 			ptep = l2b->l2b_kva;
6678 
6679 			for (k = 0; k < 256 && ptep[k] == 0; k++)
6680 				;
6681 
6682 			k &= ~63;
6683 			occ = l2b->l2b_occupancy;
6684 			va = l2b_va + (k * 4096);
6685 			for (; k < 256; k++, va += 0x1000) {
6686 				char ch = ' ';
6687 				if ((k % 64) == 0) {
6688 					if ((rows % 8) == 0) {
6689 						printf(
6690 "          |0000   |8000   |10000  |18000  |20000  |28000  |30000  |38000\n");
6691 					}
6692 					printf("%08lx: ", va);
6693 				}
6694 
6695 				ncptes[k & 63] = 0;
6696 				pte = ptep[k];
6697 				if (pte == 0) {
6698 					ch = '.';
6699 				} else {
6700 					occ--;
6701 					switch (pte & 0x0c) {
6702 					case 0x00:
6703 						ch = 'D'; /* No cache No buff */
6704 						break;
6705 					case 0x04:
6706 						ch = 'B'; /* No cache buff */
6707 						break;
6708 					case 0x08:
6709 						if (pte & 0x40)
6710 							ch = 'm';
6711 						else
6712 						   ch = 'C'; /* Cache No buff */
6713 						break;
6714 					case 0x0c:
6715 						ch = 'F'; /* Cache Buff */
6716 						break;
6717 					}
6718 
6719 					if ((pte & L2_S_PROT_U) == L2_S_PROT_U)
6720 						ch += 0x20;
6721 
6722 					if ((pte & 0xc) == 0)
6723 						ncptes[k & 63] = pte;
6724 				}
6725 
6726 				if ((k % 64) == 63) {
6727 					rows++;
6728 					printf("%c\n", ch);
6729 					pmap_dump_ncpg(pm);
6730 					if (occ == 0)
6731 						break;
6732 				} else
6733 					printf("%c", ch);
6734 			}
6735 		}
6736 	}
6737 }
6738 
6739 static void
6740 pmap_dump_ncpg(pmap_t pm)
6741 {
6742 	struct vm_page *pg;
6743 	struct vm_page_md *md;
6744 	struct pv_entry *pv;
6745 	int i;
6746 
6747 	for (i = 0; i < 63; i++) {
6748 		if (ncptes[i] == 0)
6749 			continue;
6750 
6751 		pg = PHYS_TO_VM_PAGE(l2pte_pa(ncptes[i]));
6752 		if (pg == NULL)
6753 			continue;
6754 		md = VM_PAGE_TO_MD(pg);
6755 
6756 		printf(" pa 0x%08lx: krw %d kro %d urw %d uro %d\n",
6757 		    VM_PAGE_TO_PHYS(pg),
6758 		    md->krw_mappings, md->kro_mappings,
6759 		    md->urw_mappings, md->uro_mappings);
6760 
6761 		SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
6762 			printf("   %c va 0x%08lx, flags 0x%x\n",
6763 			    (pm == pv->pv_pmap) ? '*' : ' ',
6764 			    pv->pv_va, pv->pv_flags);
6765 		}
6766 	}
6767 }
6768 #endif
6769 
6770 #ifdef PMAP_STEAL_MEMORY
6771 void
6772 pmap_boot_pageadd(pv_addr_t *newpv)
6773 {
6774 	pv_addr_t *pv, *npv;
6775 
6776 	if ((pv = SLIST_FIRST(&pmap_boot_freeq)) != NULL) {
6777 		if (newpv->pv_pa < pv->pv_va) {
6778 			KASSERT(newpv->pv_pa + newpv->pv_size <= pv->pv_pa);
6779 			if (newpv->pv_pa + newpv->pv_size == pv->pv_pa) {
6780 				newpv->pv_size += pv->pv_size;
6781 				SLIST_REMOVE_HEAD(&pmap_boot_freeq, pv_list);
6782 			}
6783 			pv = NULL;
6784 		} else {
6785 			for (; (npv = SLIST_NEXT(pv, pv_list)) != NULL;
6786 			     pv = npv) {
6787 				KASSERT(pv->pv_pa + pv->pv_size < npv->pv_pa);
6788 				KASSERT(pv->pv_pa < newpv->pv_pa);
6789 				if (newpv->pv_pa > npv->pv_pa)
6790 					continue;
6791 				if (pv->pv_pa + pv->pv_size == newpv->pv_pa) {
6792 					pv->pv_size += newpv->pv_size;
6793 					return;
6794 				}
6795 				if (newpv->pv_pa + newpv->pv_size < npv->pv_pa)
6796 					break;
6797 				newpv->pv_size += npv->pv_size;
6798 				SLIST_INSERT_AFTER(pv, newpv, pv_list);
6799 				SLIST_REMOVE_AFTER(newpv, pv_list);
6800 				return;
6801 			}
6802 		}
6803 	}
6804 
6805 	if (pv) {
6806 		SLIST_INSERT_AFTER(pv, newpv, pv_list);
6807 	} else {
6808 		SLIST_INSERT_HEAD(&pmap_boot_freeq, newpv, pv_list);
6809 	}
6810 }
6811 
6812 void
6813 pmap_boot_pagealloc(psize_t amount, psize_t mask, psize_t match,
6814 	pv_addr_t *rpv)
6815 {
6816 	pv_addr_t *pv, **pvp;
6817 	struct vm_physseg *ps;
6818 	size_t i;
6819 
6820 	KASSERT(amount & PGOFSET);
6821 	KASSERT((mask & PGOFSET) == 0);
6822 	KASSERT((match & PGOFSET) == 0);
6823 	KASSERT(amount != 0);
6824 
6825 	for (pvp = &SLIST_FIRST(&pmap_boot_freeq);
6826 	     (pv = *pvp) != NULL;
6827 	     pvp = &SLIST_NEXT(pv, pv_list)) {
6828 		pv_addr_t *newpv;
6829 		psize_t off;
6830 		/*
6831 		 * If this entry is too small to satify the request...
6832 		 */
6833 		KASSERT(pv->pv_size > 0);
6834 		if (pv->pv_size < amount)
6835 			continue;
6836 
6837 		for (off = 0; off <= mask; off += PAGE_SIZE) {
6838 			if (((pv->pv_pa + off) & mask) == match
6839 			    && off + amount <= pv->pv_size)
6840 				break;
6841 		}
6842 		if (off > mask)
6843 			continue;
6844 
6845 		rpv->pv_va = pv->pv_va + off;
6846 		rpv->pv_pa = pv->pv_pa + off;
6847 		rpv->pv_size = amount;
6848 		pv->pv_size -= amount;
6849 		if (pv->pv_size == 0) {
6850 			KASSERT(off == 0);
6851 			KASSERT((vaddr_t) pv == rpv->pv_va);
6852 			*pvp = SLIST_NEXT(pv, pv_list);
6853 		} else if (off == 0) {
6854 			KASSERT((vaddr_t) pv == rpv->pv_va);
6855 			newpv = (pv_addr_t *) (rpv->pv_va + amount);
6856 			*newpv = *pv;
6857 			newpv->pv_pa += amount;
6858 			newpv->pv_va += amount;
6859 			*pvp = newpv;
6860 		} else if (off < pv->pv_size) {
6861 			newpv = (pv_addr_t *) (rpv->pv_va + amount);
6862 			*newpv = *pv;
6863 			newpv->pv_size -= off;
6864 			newpv->pv_pa += off + amount;
6865 			newpv->pv_va += off + amount;
6866 
6867 			SLIST_NEXT(pv, pv_list) = newpv;
6868 			pv->pv_size = off;
6869 		} else {
6870 			KASSERT((vaddr_t) pv != rpv->pv_va);
6871 		}
6872 		memset((void *)rpv->pv_va, 0, amount);
6873 		return;
6874 	}
6875 
6876 	if (vm_nphysseg == 0)
6877 		panic("pmap_boot_pagealloc: couldn't allocate memory");
6878 
6879 	for (pvp = &SLIST_FIRST(&pmap_boot_freeq);
6880 	     (pv = *pvp) != NULL;
6881 	     pvp = &SLIST_NEXT(pv, pv_list)) {
6882 		if (SLIST_NEXT(pv, pv_list) == NULL)
6883 			break;
6884 	}
6885 	KASSERT(mask == 0);
6886 	for (i = 0; i < vm_nphysseg; i++) {
6887 		ps = VM_PHYSMEM_PTR(i);
6888 		if (ps->avail_start == atop(pv->pv_pa + pv->pv_size)
6889 		    && pv->pv_va + pv->pv_size <= ptoa(ps->avail_end)) {
6890 			rpv->pv_va = pv->pv_va;
6891 			rpv->pv_pa = pv->pv_pa;
6892 			rpv->pv_size = amount;
6893 			*pvp = NULL;
6894 			pmap_map_chunk(kernel_l1pt.pv_va,
6895 			     ptoa(ps->avail_start) + (pv->pv_va - pv->pv_pa),
6896 			     ptoa(ps->avail_start),
6897 			     amount - pv->pv_size,
6898 			     VM_PROT_READ|VM_PROT_WRITE,
6899 			     PTE_CACHE);
6900 			ps->avail_start += atop(amount - pv->pv_size);
6901 			/*
6902 			 * If we consumed the entire physseg, remove it.
6903 			 */
6904 			if (ps->avail_start == ps->avail_end) {
6905 				for (--vm_nphysseg; i < vm_nphysseg; i++)
6906 					VM_PHYSMEM_PTR_SWAP(i, i + 1);
6907 			}
6908 			memset((void *)rpv->pv_va, 0, rpv->pv_size);
6909 			return;
6910 		}
6911 	}
6912 
6913 	panic("pmap_boot_pagealloc: couldn't allocate memory");
6914 }
6915 
6916 vaddr_t
6917 pmap_steal_memory(vsize_t size, vaddr_t *vstartp, vaddr_t *vendp)
6918 {
6919 	pv_addr_t pv;
6920 
6921 	pmap_boot_pagealloc(size, 0, 0, &pv);
6922 
6923 	return pv.pv_va;
6924 }
6925 #endif /* PMAP_STEAL_MEMORY */
6926 
6927 SYSCTL_SETUP(sysctl_machdep_pmap_setup, "sysctl machdep.kmpages setup")
6928 {
6929 	sysctl_createv(clog, 0, NULL, NULL,
6930 			CTLFLAG_PERMANENT,
6931 			CTLTYPE_NODE, "machdep", NULL,
6932 			NULL, 0, NULL, 0,
6933 			CTL_MACHDEP, CTL_EOL);
6934 
6935 	sysctl_createv(clog, 0, NULL, NULL,
6936 			CTLFLAG_PERMANENT,
6937 			CTLTYPE_INT, "kmpages",
6938 			SYSCTL_DESCR("count of pages allocated to kernel memory allocators"),
6939 			NULL, 0, &pmap_kmpages, 0,
6940 			CTL_MACHDEP, CTL_CREATE, CTL_EOL);
6941 }
6942 
6943 #ifdef PMAP_NEED_ALLOC_POOLPAGE
6944 struct vm_page *
6945 arm_pmap_alloc_poolpage(int flags)
6946 {
6947 	/*
6948 	 * On some systems, only some pages may be "coherent" for dma and we
6949 	 * want to prefer those for pool pages (think mbufs) but fallback to
6950 	 * any page if none is available.
6951 	 */
6952 	if (arm_poolpage_vmfreelist != VM_FREELIST_DEFAULT) {
6953 		return uvm_pagealloc_strat(NULL, 0, NULL, flags,
6954 		    UVM_PGA_STRAT_FALLBACK, arm_poolpage_vmfreelist);
6955 	}
6956 
6957 	return uvm_pagealloc(NULL, 0, NULL, flags);
6958 }
6959 #endif
6960