xref: /netbsd-src/sys/arch/arm/arm32/pmap.c (revision b7b7574d3bf8eeb51a1fa3977b59142ec6434a55)
1 /*	$NetBSD: pmap.c,v 1.294 2014/06/15 04:04:01 ozaki-r 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 #ifdef MULTIPROCESSOR
196 #define _INTR_PRIVATE
197 #endif
198 
199 #include <sys/param.h>
200 #include <sys/types.h>
201 #include <sys/kernel.h>
202 #include <sys/systm.h>
203 #include <sys/proc.h>
204 #include <sys/intr.h>
205 #include <sys/pool.h>
206 #include <sys/kmem.h>
207 #include <sys/cdefs.h>
208 #include <sys/cpu.h>
209 #include <sys/sysctl.h>
210 #include <sys/bus.h>
211 #include <sys/atomic.h>
212 #include <sys/kernhist.h>
213 
214 #include <uvm/uvm.h>
215 
216 #include <arm/locore.h>
217 //#include <arm/arm32/katelib.h>
218 
219 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.294 2014/06/15 04:04:01 ozaki-r Exp $");
220 
221 //#define PMAP_DEBUG
222 #ifdef PMAP_DEBUG
223 
224 /* XXX need to get rid of all refs to this */
225 int pmap_debug_level = 0;
226 
227 /*
228  * for switching to potentially finer grained debugging
229  */
230 #define	PDB_FOLLOW	0x0001
231 #define	PDB_INIT	0x0002
232 #define	PDB_ENTER	0x0004
233 #define	PDB_REMOVE	0x0008
234 #define	PDB_CREATE	0x0010
235 #define	PDB_PTPAGE	0x0020
236 #define	PDB_GROWKERN	0x0040
237 #define	PDB_BITS	0x0080
238 #define	PDB_COLLECT	0x0100
239 #define	PDB_PROTECT	0x0200
240 #define	PDB_MAP_L1	0x0400
241 #define	PDB_BOOTSTRAP	0x1000
242 #define	PDB_PARANOIA	0x2000
243 #define	PDB_WIRING	0x4000
244 #define	PDB_PVDUMP	0x8000
245 #define	PDB_VAC		0x10000
246 #define	PDB_KENTER	0x20000
247 #define	PDB_KREMOVE	0x40000
248 #define	PDB_EXEC	0x80000
249 
250 int debugmap = 1;
251 int pmapdebug = 0;
252 #define	NPDEBUG(_lev_,_stat_) \
253 	if (pmapdebug & (_lev_)) \
254         	((_stat_))
255 
256 #else	/* PMAP_DEBUG */
257 #define NPDEBUG(_lev_,_stat_) /* Nothing */
258 #endif	/* PMAP_DEBUG */
259 
260 /*
261  * pmap_kernel() points here
262  */
263 static struct pmap	kernel_pmap_store = {
264 #ifndef ARM_MMU_EXTENDED
265 	.pm_activated = true,
266 	.pm_domain = PMAP_DOMAIN_KERNEL,
267 	.pm_cstate.cs_all = PMAP_CACHE_STATE_ALL,
268 #endif
269 };
270 struct pmap * const	kernel_pmap_ptr = &kernel_pmap_store;
271 #undef pmap_kernel
272 #define pmap_kernel()	(&kernel_pmap_store)
273 #ifdef PMAP_NEED_ALLOC_POOLPAGE
274 int			arm_poolpage_vmfreelist = VM_FREELIST_DEFAULT;
275 #endif
276 
277 /*
278  * Pool and cache that pmap structures are allocated from.
279  * We use a cache to avoid clearing the pm_l2[] array (1KB)
280  * in pmap_create().
281  */
282 static struct pool_cache pmap_cache;
283 static LIST_HEAD(, pmap) pmap_pmaps;
284 
285 /*
286  * Pool of PV structures
287  */
288 static struct pool pmap_pv_pool;
289 static void *pmap_bootstrap_pv_page_alloc(struct pool *, int);
290 static void pmap_bootstrap_pv_page_free(struct pool *, void *);
291 static struct pool_allocator pmap_bootstrap_pv_allocator = {
292 	pmap_bootstrap_pv_page_alloc, pmap_bootstrap_pv_page_free
293 };
294 
295 /*
296  * Pool and cache of l2_dtable structures.
297  * We use a cache to avoid clearing the structures when they're
298  * allocated. (196 bytes)
299  */
300 static struct pool_cache pmap_l2dtable_cache;
301 static vaddr_t pmap_kernel_l2dtable_kva;
302 
303 /*
304  * Pool and cache of L2 page descriptors.
305  * We use a cache to avoid clearing the descriptor table
306  * when they're allocated. (1KB)
307  */
308 static struct pool_cache pmap_l2ptp_cache;
309 static vaddr_t pmap_kernel_l2ptp_kva;
310 static paddr_t pmap_kernel_l2ptp_phys;
311 
312 #ifdef PMAPCOUNTERS
313 #define	PMAP_EVCNT_INITIALIZER(name) \
314 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap", name)
315 
316 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
317 static struct evcnt pmap_ev_vac_clean_one =
318    PMAP_EVCNT_INITIALIZER("clean page (1 color)");
319 static struct evcnt pmap_ev_vac_flush_one =
320    PMAP_EVCNT_INITIALIZER("flush page (1 color)");
321 static struct evcnt pmap_ev_vac_flush_lots =
322    PMAP_EVCNT_INITIALIZER("flush page (2+ colors)");
323 static struct evcnt pmap_ev_vac_flush_lots2 =
324    PMAP_EVCNT_INITIALIZER("flush page (2+ colors, kmpage)");
325 EVCNT_ATTACH_STATIC(pmap_ev_vac_clean_one);
326 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_one);
327 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_lots);
328 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_lots2);
329 
330 static struct evcnt pmap_ev_vac_color_new =
331    PMAP_EVCNT_INITIALIZER("new page color");
332 static struct evcnt pmap_ev_vac_color_reuse =
333    PMAP_EVCNT_INITIALIZER("ok first page color");
334 static struct evcnt pmap_ev_vac_color_ok =
335    PMAP_EVCNT_INITIALIZER("ok page color");
336 static struct evcnt pmap_ev_vac_color_blind =
337    PMAP_EVCNT_INITIALIZER("blind page color");
338 static struct evcnt pmap_ev_vac_color_change =
339    PMAP_EVCNT_INITIALIZER("change page color");
340 static struct evcnt pmap_ev_vac_color_erase =
341    PMAP_EVCNT_INITIALIZER("erase page color");
342 static struct evcnt pmap_ev_vac_color_none =
343    PMAP_EVCNT_INITIALIZER("no page color");
344 static struct evcnt pmap_ev_vac_color_restore =
345    PMAP_EVCNT_INITIALIZER("restore page color");
346 
347 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_new);
348 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_reuse);
349 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_ok);
350 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_blind);
351 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_change);
352 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_erase);
353 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_none);
354 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_restore);
355 #endif
356 
357 static struct evcnt pmap_ev_mappings =
358    PMAP_EVCNT_INITIALIZER("pages mapped");
359 static struct evcnt pmap_ev_unmappings =
360    PMAP_EVCNT_INITIALIZER("pages unmapped");
361 static struct evcnt pmap_ev_remappings =
362    PMAP_EVCNT_INITIALIZER("pages remapped");
363 
364 EVCNT_ATTACH_STATIC(pmap_ev_mappings);
365 EVCNT_ATTACH_STATIC(pmap_ev_unmappings);
366 EVCNT_ATTACH_STATIC(pmap_ev_remappings);
367 
368 static struct evcnt pmap_ev_kernel_mappings =
369    PMAP_EVCNT_INITIALIZER("kernel pages mapped");
370 static struct evcnt pmap_ev_kernel_unmappings =
371    PMAP_EVCNT_INITIALIZER("kernel pages unmapped");
372 static struct evcnt pmap_ev_kernel_remappings =
373    PMAP_EVCNT_INITIALIZER("kernel pages remapped");
374 
375 EVCNT_ATTACH_STATIC(pmap_ev_kernel_mappings);
376 EVCNT_ATTACH_STATIC(pmap_ev_kernel_unmappings);
377 EVCNT_ATTACH_STATIC(pmap_ev_kernel_remappings);
378 
379 static struct evcnt pmap_ev_kenter_mappings =
380    PMAP_EVCNT_INITIALIZER("kenter pages mapped");
381 static struct evcnt pmap_ev_kenter_unmappings =
382    PMAP_EVCNT_INITIALIZER("kenter pages unmapped");
383 static struct evcnt pmap_ev_kenter_remappings =
384    PMAP_EVCNT_INITIALIZER("kenter pages remapped");
385 static struct evcnt pmap_ev_pt_mappings =
386    PMAP_EVCNT_INITIALIZER("page table pages mapped");
387 
388 EVCNT_ATTACH_STATIC(pmap_ev_kenter_mappings);
389 EVCNT_ATTACH_STATIC(pmap_ev_kenter_unmappings);
390 EVCNT_ATTACH_STATIC(pmap_ev_kenter_remappings);
391 EVCNT_ATTACH_STATIC(pmap_ev_pt_mappings);
392 
393 static struct evcnt pmap_ev_fixup_mod =
394    PMAP_EVCNT_INITIALIZER("page modification emulations");
395 static struct evcnt pmap_ev_fixup_ref =
396    PMAP_EVCNT_INITIALIZER("page reference emulations");
397 static struct evcnt pmap_ev_fixup_exec =
398    PMAP_EVCNT_INITIALIZER("exec pages fixed up");
399 static struct evcnt pmap_ev_fixup_pdes =
400    PMAP_EVCNT_INITIALIZER("pdes fixed up");
401 #ifndef ARM_MMU_EXTENDED
402 static struct evcnt pmap_ev_fixup_ptesync =
403    PMAP_EVCNT_INITIALIZER("ptesync fixed");
404 #endif
405 
406 EVCNT_ATTACH_STATIC(pmap_ev_fixup_mod);
407 EVCNT_ATTACH_STATIC(pmap_ev_fixup_ref);
408 EVCNT_ATTACH_STATIC(pmap_ev_fixup_exec);
409 EVCNT_ATTACH_STATIC(pmap_ev_fixup_pdes);
410 #ifndef ARM_MMU_EXTENDED
411 EVCNT_ATTACH_STATIC(pmap_ev_fixup_ptesync);
412 #endif
413 
414 #ifdef PMAP_CACHE_VIPT
415 static struct evcnt pmap_ev_exec_mappings =
416    PMAP_EVCNT_INITIALIZER("exec pages mapped");
417 static struct evcnt pmap_ev_exec_cached =
418    PMAP_EVCNT_INITIALIZER("exec pages cached");
419 
420 EVCNT_ATTACH_STATIC(pmap_ev_exec_mappings);
421 EVCNT_ATTACH_STATIC(pmap_ev_exec_cached);
422 
423 static struct evcnt pmap_ev_exec_synced =
424    PMAP_EVCNT_INITIALIZER("exec pages synced");
425 static struct evcnt pmap_ev_exec_synced_map =
426    PMAP_EVCNT_INITIALIZER("exec pages synced (MP)");
427 #ifndef ARM_MMU_EXTENDED
428 static struct evcnt pmap_ev_exec_synced_unmap =
429    PMAP_EVCNT_INITIALIZER("exec pages synced (UM)");
430 static struct evcnt pmap_ev_exec_synced_remap =
431    PMAP_EVCNT_INITIALIZER("exec pages synced (RM)");
432 static struct evcnt pmap_ev_exec_synced_clearbit =
433    PMAP_EVCNT_INITIALIZER("exec pages synced (DG)");
434 static struct evcnt pmap_ev_exec_synced_kremove =
435    PMAP_EVCNT_INITIALIZER("exec pages synced (KU)");
436 #endif
437 
438 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced);
439 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_map);
440 #ifndef ARM_MMU_EXTENDED
441 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_unmap);
442 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_remap);
443 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_clearbit);
444 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_kremove);
445 #endif
446 
447 static struct evcnt pmap_ev_exec_discarded_unmap =
448    PMAP_EVCNT_INITIALIZER("exec pages discarded (UM)");
449 static struct evcnt pmap_ev_exec_discarded_zero =
450    PMAP_EVCNT_INITIALIZER("exec pages discarded (ZP)");
451 static struct evcnt pmap_ev_exec_discarded_copy =
452    PMAP_EVCNT_INITIALIZER("exec pages discarded (CP)");
453 static struct evcnt pmap_ev_exec_discarded_page_protect =
454    PMAP_EVCNT_INITIALIZER("exec pages discarded (PP)");
455 static struct evcnt pmap_ev_exec_discarded_clearbit =
456    PMAP_EVCNT_INITIALIZER("exec pages discarded (DG)");
457 static struct evcnt pmap_ev_exec_discarded_kremove =
458    PMAP_EVCNT_INITIALIZER("exec pages discarded (KU)");
459 #ifdef ARM_MMU_EXTENDED
460 static struct evcnt pmap_ev_exec_discarded_modfixup =
461    PMAP_EVCNT_INITIALIZER("exec pages discarded (MF)");
462 #endif
463 
464 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_unmap);
465 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_zero);
466 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_copy);
467 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_page_protect);
468 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_clearbit);
469 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_kremove);
470 #ifdef ARM_MMU_EXTENDED
471 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_modfixup);
472 #endif
473 #endif /* PMAP_CACHE_VIPT */
474 
475 static struct evcnt pmap_ev_updates = PMAP_EVCNT_INITIALIZER("updates");
476 static struct evcnt pmap_ev_collects = PMAP_EVCNT_INITIALIZER("collects");
477 static struct evcnt pmap_ev_activations = PMAP_EVCNT_INITIALIZER("activations");
478 
479 EVCNT_ATTACH_STATIC(pmap_ev_updates);
480 EVCNT_ATTACH_STATIC(pmap_ev_collects);
481 EVCNT_ATTACH_STATIC(pmap_ev_activations);
482 
483 #define	PMAPCOUNT(x)	((void)(pmap_ev_##x.ev_count++))
484 #else
485 #define	PMAPCOUNT(x)	((void)0)
486 #endif
487 
488 /*
489  * pmap copy/zero page, and mem(5) hook point
490  */
491 static pt_entry_t *csrc_pte, *cdst_pte;
492 static vaddr_t csrcp, cdstp;
493 #ifdef MULTIPROCESSOR
494 static size_t cnptes;
495 #define	cpu_csrc_pte(o)	(csrc_pte + cnptes * cpu_number() + ((o) >> L2_S_SHIFT))
496 #define	cpu_cdst_pte(o)	(cdst_pte + cnptes * cpu_number() + ((o) >> L2_S_SHIFT))
497 #define	cpu_csrcp(o)	(csrcp + L2_S_SIZE * cnptes * cpu_number() + (o))
498 #define	cpu_cdstp(o)	(cdstp + L2_S_SIZE * cnptes * cpu_number() + (o))
499 #else
500 #define	cpu_csrc_pte(o)	(csrc_pte + ((o) >> L2_S_SHIFT))
501 #define	cpu_cdst_pte(o)	(cdst_pte + ((o) >> L2_S_SHIFT))
502 #define	cpu_csrcp(o)	(csrcp + (o))
503 #define	cpu_cdstp(o)	(cdstp + (o))
504 #endif
505 vaddr_t memhook;			/* used by mem.c & others */
506 kmutex_t memlock __cacheline_aligned;	/* used by mem.c & others */
507 kmutex_t pmap_lock __cacheline_aligned;
508 extern void *msgbufaddr;
509 int pmap_kmpages;
510 /*
511  * Flag to indicate if pmap_init() has done its thing
512  */
513 bool pmap_initialized;
514 
515 #if defined(ARM_MMU_EXTENDED) && defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
516 /*
517  * Start of direct-mapped memory
518  */
519 vaddr_t pmap_directbase = KERNEL_BASE;
520 #endif
521 
522 /*
523  * Misc. locking data structures
524  */
525 
526 static inline void
527 pmap_acquire_pmap_lock(pmap_t pm)
528 {
529 	if (pm == pmap_kernel()) {
530 #ifdef MULTIPROCESSOR
531 		KERNEL_LOCK(1, NULL);
532 #endif
533 	} else {
534 		mutex_enter(pm->pm_lock);
535 	}
536 }
537 
538 static inline void
539 pmap_release_pmap_lock(pmap_t pm)
540 {
541 	if (pm == pmap_kernel()) {
542 #ifdef MULTIPROCESSOR
543 		KERNEL_UNLOCK_ONE(NULL);
544 #endif
545 	} else {
546 		mutex_exit(pm->pm_lock);
547 	}
548 }
549 
550 static inline void
551 pmap_acquire_page_lock(struct vm_page_md *md)
552 {
553 	mutex_enter(&pmap_lock);
554 }
555 
556 static inline void
557 pmap_release_page_lock(struct vm_page_md *md)
558 {
559 	mutex_exit(&pmap_lock);
560 }
561 
562 #ifdef DIAGNOSTIC
563 static inline int
564 pmap_page_locked_p(struct vm_page_md *md)
565 {
566 	return mutex_owned(&pmap_lock);
567 }
568 #endif
569 
570 
571 /*
572  * Metadata for L1 translation tables.
573  */
574 #ifndef ARM_MMU_EXTENDED
575 struct l1_ttable {
576 	/* Entry on the L1 Table list */
577 	SLIST_ENTRY(l1_ttable) l1_link;
578 
579 	/* Entry on the L1 Least Recently Used list */
580 	TAILQ_ENTRY(l1_ttable) l1_lru;
581 
582 	/* Track how many domains are allocated from this L1 */
583 	volatile u_int l1_domain_use_count;
584 
585 	/*
586 	 * A free-list of domain numbers for this L1.
587 	 * We avoid using ffs() and a bitmap to track domains since ffs()
588 	 * is slow on ARM.
589 	 */
590 	uint8_t l1_domain_first;
591 	uint8_t l1_domain_free[PMAP_DOMAINS];
592 
593 	/* Physical address of this L1 page table */
594 	paddr_t l1_physaddr;
595 
596 	/* KVA of this L1 page table */
597 	pd_entry_t *l1_kva;
598 };
599 
600 /*
601  * L1 Page Tables are tracked using a Least Recently Used list.
602  *  - New L1s are allocated from the HEAD.
603  *  - Freed L1s are added to the TAIl.
604  *  - Recently accessed L1s (where an 'access' is some change to one of
605  *    the userland pmaps which owns this L1) are moved to the TAIL.
606  */
607 static TAILQ_HEAD(, l1_ttable) l1_lru_list;
608 static kmutex_t l1_lru_lock __cacheline_aligned;
609 
610 /*
611  * A list of all L1 tables
612  */
613 static SLIST_HEAD(, l1_ttable) l1_list;
614 #endif /* ARM_MMU_EXTENDED */
615 
616 /*
617  * The l2_dtable tracks L2_BUCKET_SIZE worth of L1 slots.
618  *
619  * This is normally 16MB worth L2 page descriptors for any given pmap.
620  * Reference counts are maintained for L2 descriptors so they can be
621  * freed when empty.
622  */
623 struct l2_dtable {
624 	/* The number of L2 page descriptors allocated to this l2_dtable */
625 	u_int l2_occupancy;
626 
627 	/* List of L2 page descriptors */
628 	struct l2_bucket {
629 		pt_entry_t *l2b_kva;	/* KVA of L2 Descriptor Table */
630 		paddr_t l2b_pa;		/* Physical address of same */
631 		u_short l2b_l1slot;	/* This L2 table's L1 index */
632 		u_short l2b_occupancy;	/* How many active descriptors */
633 	} l2_bucket[L2_BUCKET_SIZE];
634 };
635 
636 /*
637  * Given an L1 table index, calculate the corresponding l2_dtable index
638  * and bucket index within the l2_dtable.
639  */
640 #define L2_BUCKET_XSHIFT	(L2_BUCKET_XLOG2 - L1_S_SHIFT)
641 #define L2_BUCKET_XFRAME	(~(vaddr_t)0 << L2_BUCKET_XLOG2)
642 #define L2_BUCKET_IDX(l1slot)	((l1slot) >> L2_BUCKET_XSHIFT)
643 #define L2_IDX(l1slot)		(L2_BUCKET_IDX(l1slot) >> L2_BUCKET_LOG2)
644 #define L2_BUCKET(l1slot)	(L2_BUCKET_IDX(l1slot) & (L2_BUCKET_SIZE - 1))
645 
646 __CTASSERT(0x100000000ULL == ((uint64_t)L2_SIZE * L2_BUCKET_SIZE * L1_S_SIZE));
647 __CTASSERT(L2_BUCKET_XFRAME == ~(L2_BUCKET_XSIZE-1));
648 
649 /*
650  * Given a virtual address, this macro returns the
651  * virtual address required to drop into the next L2 bucket.
652  */
653 #define	L2_NEXT_BUCKET_VA(va)	(((va) & L2_BUCKET_XFRAME) + L2_BUCKET_XSIZE)
654 
655 /*
656  * L2 allocation.
657  */
658 #define	pmap_alloc_l2_dtable()		\
659 	    pool_cache_get(&pmap_l2dtable_cache, PR_NOWAIT)
660 #define	pmap_free_l2_dtable(l2)		\
661 	    pool_cache_put(&pmap_l2dtable_cache, (l2))
662 #define pmap_alloc_l2_ptp(pap)		\
663 	    ((pt_entry_t *)pool_cache_get_paddr(&pmap_l2ptp_cache,\
664 	    PR_NOWAIT, (pap)))
665 
666 /*
667  * We try to map the page tables write-through, if possible.  However, not
668  * all CPUs have a write-through cache mode, so on those we have to sync
669  * the cache when we frob page tables.
670  *
671  * We try to evaluate this at compile time, if possible.  However, it's
672  * not always possible to do that, hence this run-time var.
673  */
674 int	pmap_needs_pte_sync;
675 
676 /*
677  * Real definition of pv_entry.
678  */
679 struct pv_entry {
680 	SLIST_ENTRY(pv_entry) pv_link;	/* next pv_entry */
681 	pmap_t		pv_pmap;        /* pmap where mapping lies */
682 	vaddr_t		pv_va;          /* virtual address for mapping */
683 	u_int		pv_flags;       /* flags */
684 };
685 
686 /*
687  * Macro to determine if a mapping might be resident in the
688  * instruction cache and/or TLB
689  */
690 #if ARM_MMU_V7 > 0 && !defined(ARM_MMU_EXTENDED)
691 /*
692  * Speculative loads by Cortex cores can cause TLB entries to be filled even if
693  * there are no explicit accesses, so there may be always be TLB entries to
694  * flush.  If we used ASIDs then this would not be a problem.
695  */
696 #define	PV_BEEN_EXECD(f)  (((f) & PVF_EXEC) == PVF_EXEC)
697 #else
698 #define	PV_BEEN_EXECD(f)  (((f) & (PVF_REF | PVF_EXEC)) == (PVF_REF | PVF_EXEC))
699 #endif
700 #define	PV_IS_EXEC_P(f)   (((f) & PVF_EXEC) != 0)
701 #define	PV_IS_KENTRY_P(f) (((f) & PVF_KENTRY) != 0)
702 #define	PV_IS_WRITE_P(f)  (((f) & PVF_WRITE) != 0)
703 
704 /*
705  * Macro to determine if a mapping might be resident in the
706  * data cache and/or TLB
707  */
708 #if ARM_MMU_V7 > 0 && !defined(ARM_MMU_EXTENDED)
709 /*
710  * Speculative loads by Cortex cores can cause TLB entries to be filled even if
711  * there are no explicit accesses, so there may be always be TLB entries to
712  * flush.  If we used ASIDs then this would not be a problem.
713  */
714 #define	PV_BEEN_REFD(f)   (1)
715 #else
716 #define	PV_BEEN_REFD(f)   (((f) & PVF_REF) != 0)
717 #endif
718 
719 /*
720  * Local prototypes
721  */
722 static bool		pmap_set_pt_cache_mode(pd_entry_t *, vaddr_t, size_t);
723 static void		pmap_alloc_specials(vaddr_t *, int, vaddr_t *,
724 			    pt_entry_t **);
725 static bool		pmap_is_current(pmap_t) __unused;
726 static bool		pmap_is_cached(pmap_t);
727 static void		pmap_enter_pv(struct vm_page_md *, paddr_t, struct pv_entry *,
728 			    pmap_t, vaddr_t, u_int);
729 static struct pv_entry *pmap_find_pv(struct vm_page_md *, pmap_t, vaddr_t);
730 static struct pv_entry *pmap_remove_pv(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
731 static u_int		pmap_modify_pv(struct vm_page_md *, paddr_t, pmap_t, vaddr_t,
732 			    u_int, u_int);
733 
734 static void		pmap_pinit(pmap_t);
735 static int		pmap_pmap_ctor(void *, void *, int);
736 
737 static void		pmap_alloc_l1(pmap_t);
738 static void		pmap_free_l1(pmap_t);
739 #ifndef ARM_MMU_EXTENDED
740 static void		pmap_use_l1(pmap_t);
741 #endif
742 
743 static struct l2_bucket *pmap_get_l2_bucket(pmap_t, vaddr_t);
744 static struct l2_bucket *pmap_alloc_l2_bucket(pmap_t, vaddr_t);
745 static void		pmap_free_l2_bucket(pmap_t, struct l2_bucket *, u_int);
746 static int		pmap_l2ptp_ctor(void *, void *, int);
747 static int		pmap_l2dtable_ctor(void *, void *, int);
748 
749 static void		pmap_vac_me_harder(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
750 #ifdef PMAP_CACHE_VIVT
751 static void		pmap_vac_me_kpmap(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
752 static void		pmap_vac_me_user(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
753 #endif
754 
755 static void		pmap_clearbit(struct vm_page_md *, paddr_t, u_int);
756 #ifdef PMAP_CACHE_VIVT
757 static bool		pmap_clean_page(struct vm_page_md *, bool);
758 #endif
759 #ifdef PMAP_CACHE_VIPT
760 static void		pmap_syncicache_page(struct vm_page_md *, paddr_t);
761 enum pmap_flush_op {
762 	PMAP_FLUSH_PRIMARY,
763 	PMAP_FLUSH_SECONDARY,
764 	PMAP_CLEAN_PRIMARY
765 };
766 #ifndef ARM_MMU_EXTENDED
767 static void		pmap_flush_page(struct vm_page_md *, paddr_t, enum pmap_flush_op);
768 #endif
769 #endif
770 static void		pmap_page_remove(struct vm_page_md *, paddr_t);
771 
772 #ifndef ARM_MMU_EXTENDED
773 static void		pmap_init_l1(struct l1_ttable *, pd_entry_t *);
774 #endif
775 static vaddr_t		kernel_pt_lookup(paddr_t);
776 
777 
778 /*
779  * Misc variables
780  */
781 vaddr_t virtual_avail;
782 vaddr_t virtual_end;
783 vaddr_t pmap_curmaxkvaddr;
784 
785 paddr_t avail_start;
786 paddr_t avail_end;
787 
788 pv_addrqh_t pmap_boot_freeq = SLIST_HEAD_INITIALIZER(&pmap_boot_freeq);
789 pv_addr_t kernelpages;
790 pv_addr_t kernel_l1pt;
791 pv_addr_t systempage;
792 
793 /* Function to set the debug level of the pmap code */
794 
795 #ifdef PMAP_DEBUG
796 void
797 pmap_debug(int level)
798 {
799 	pmap_debug_level = level;
800 	printf("pmap_debug: level=%d\n", pmap_debug_level);
801 }
802 #endif	/* PMAP_DEBUG */
803 
804 #ifdef PMAP_CACHE_VIPT
805 #define PMAP_VALIDATE_MD_PAGE(md)	\
806 	KASSERTMSG(arm_cache_prefer_mask == 0 || (((md)->pvh_attrs & PVF_WRITE) == 0) == ((md)->urw_mappings + (md)->krw_mappings == 0), \
807 	    "(md) %p: attrs=%#x urw=%u krw=%u", (md), \
808 	    (md)->pvh_attrs, (md)->urw_mappings, (md)->krw_mappings);
809 #endif /* PMAP_CACHE_VIPT */
810 /*
811  * A bunch of routines to conditionally flush the caches/TLB depending
812  * on whether the specified pmap actually needs to be flushed at any
813  * given time.
814  */
815 static inline void
816 pmap_tlb_flush_SE(pmap_t pm, vaddr_t va, u_int flags)
817 {
818 #ifdef ARM_MMU_EXTENDED
819 	pmap_tlb_invalidate_addr(pm, va);
820 #else
821 	if (pm->pm_cstate.cs_tlb_id != 0) {
822 		if (PV_BEEN_EXECD(flags)) {
823 			cpu_tlb_flushID_SE(va);
824 		} else if (PV_BEEN_REFD(flags)) {
825 			cpu_tlb_flushD_SE(va);
826 		}
827 	}
828 #endif /* ARM_MMU_EXTENDED */
829 }
830 
831 static inline void
832 pmap_tlb_flushID(pmap_t pm)
833 {
834 #ifdef ARM_MMU_EXTENDED
835 	pmap_tlb_asid_release_all(pm);
836 #else
837 	if (pm->pm_cstate.cs_tlb_id) {
838 		cpu_tlb_flushID();
839 #if ARM_MMU_V7 == 0
840 		/*
841 		 * Speculative loads by Cortex cores can cause TLB entries to
842 		 * be filled even if there are no explicit accesses, so there
843 		 * may be always be TLB entries to flush.  If we used ASIDs
844 		 * then it would not be a problem.
845 		 * This is not true for other CPUs.
846 		 */
847 		pm->pm_cstate.cs_tlb = 0;
848 #endif /* ARM_MMU_V7 */
849 	}
850 #endif /* ARM_MMU_EXTENDED */
851 }
852 
853 static inline void
854 pmap_tlb_flushD(pmap_t pm)
855 {
856 #ifdef ARM_MMU_EXTENDED
857 	pmap_tlb_asid_release_all(pm);
858 #else
859 	if (pm->pm_cstate.cs_tlb_d) {
860 		cpu_tlb_flushD();
861 #if ARM_MMU_V7 == 0
862 		/*
863 		 * Speculative loads by Cortex cores can cause TLB entries to
864 		 * be filled even if there are no explicit accesses, so there
865 		 * may be always be TLB entries to flush.  If we used ASIDs
866 		 * then it would not be a problem.
867 		 * This is not true for other CPUs.
868 		 */
869 		pm->pm_cstate.cs_tlb_d = 0;
870 #endif /* ARM_MMU_V7 */
871 	}
872 #endif /* ARM_MMU_EXTENDED */
873 }
874 
875 #ifdef PMAP_CACHE_VIVT
876 static inline void
877 pmap_cache_wbinv_page(pmap_t pm, vaddr_t va, bool do_inv, u_int flags)
878 {
879 	if (PV_BEEN_EXECD(flags) && pm->pm_cstate.cs_cache_id) {
880 		cpu_idcache_wbinv_range(va, PAGE_SIZE);
881 	} else if (PV_BEEN_REFD(flags) && pm->pm_cstate.cs_cache_d) {
882 		if (do_inv) {
883 			if (flags & PVF_WRITE)
884 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
885 			else
886 				cpu_dcache_inv_range(va, PAGE_SIZE);
887 		} else if (flags & PVF_WRITE) {
888 			cpu_dcache_wb_range(va, PAGE_SIZE);
889 		}
890 	}
891 }
892 
893 static inline void
894 pmap_cache_wbinv_all(pmap_t pm, u_int flags)
895 {
896 	if (PV_BEEN_EXECD(flags)) {
897 		if (pm->pm_cstate.cs_cache_id) {
898 			cpu_idcache_wbinv_all();
899 			pm->pm_cstate.cs_cache = 0;
900 		}
901 	} else if (pm->pm_cstate.cs_cache_d) {
902 		cpu_dcache_wbinv_all();
903 		pm->pm_cstate.cs_cache_d = 0;
904 	}
905 }
906 #endif /* PMAP_CACHE_VIVT */
907 
908 static inline uint8_t
909 pmap_domain(pmap_t pm)
910 {
911 #ifdef ARM_MMU_EXTENDED
912 	return pm == pmap_kernel() ? PMAP_DOMAIN_KERNEL : PMAP_DOMAIN_USER;
913 #else
914 	return pm->pm_domain;
915 #endif
916 }
917 
918 static inline pd_entry_t *
919 pmap_l1_kva(pmap_t pm)
920 {
921 #ifdef ARM_MMU_EXTENDED
922 	return pm->pm_l1;
923 #else
924 	return pm->pm_l1->l1_kva;
925 #endif
926 }
927 
928 static inline bool
929 pmap_is_current(pmap_t pm)
930 {
931 	if (pm == pmap_kernel() || curproc->p_vmspace->vm_map.pmap == pm)
932 		return true;
933 
934 	return false;
935 }
936 
937 static inline bool
938 pmap_is_cached(pmap_t pm)
939 {
940 #ifdef ARM_MMU_EXTENDED
941 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
942 	if (pm == pmap_kernel() || PMAP_PAI_ASIDVALID_P(PMAP_PAI(pm, ti), ti))
943 		return true;
944 #else
945 	struct cpu_info * const ci = curcpu();
946 	if (pm == pmap_kernel() || ci->ci_pmap_lastuser == NULL
947 	    || ci->ci_pmap_lastuser == pm)
948 		return true;
949 #endif /* ARM_MMU_EXTENDED */
950 
951 	return false;
952 }
953 
954 /*
955  * PTE_SYNC_CURRENT:
956  *
957  *     Make sure the pte is written out to RAM.
958  *     We need to do this for one of two cases:
959  *       - We're dealing with the kernel pmap
960  *       - There is no pmap active in the cache/tlb.
961  *       - The specified pmap is 'active' in the cache/tlb.
962  */
963 #ifdef PMAP_INCLUDE_PTE_SYNC
964 #define	PTE_SYNC_CURRENT(pm, ptep)	\
965 do {					\
966 	if (PMAP_NEEDS_PTE_SYNC && 	\
967 	    pmap_is_cached(pm))		\
968 		PTE_SYNC(ptep);		\
969 } while (/*CONSTCOND*/0)
970 #else
971 #define	PTE_SYNC_CURRENT(pm, ptep)	/* nothing */
972 #endif
973 
974 /*
975  * main pv_entry manipulation functions:
976  *   pmap_enter_pv: enter a mapping onto a vm_page list
977  *   pmap_remove_pv: remove a mapping from a vm_page list
978  *
979  * NOTE: pmap_enter_pv expects to lock the pvh itself
980  *       pmap_remove_pv expects the caller to lock the pvh before calling
981  */
982 
983 /*
984  * pmap_enter_pv: enter a mapping onto a vm_page lst
985  *
986  * => caller should hold the proper lock on pmap_main_lock
987  * => caller should have pmap locked
988  * => we will gain the lock on the vm_page and allocate the new pv_entry
989  * => caller should adjust ptp's wire_count before calling
990  * => caller should not adjust pmap's wire_count
991  */
992 static void
993 pmap_enter_pv(struct vm_page_md *md, paddr_t pa, struct pv_entry *pv, pmap_t pm,
994     vaddr_t va, u_int flags)
995 {
996 	struct pv_entry **pvp;
997 
998 	NPDEBUG(PDB_PVDUMP,
999 	    printf("pmap_enter_pv: pm %p, md %p, flags 0x%x\n", pm, md, flags));
1000 
1001 	pv->pv_pmap = pm;
1002 	pv->pv_va = va;
1003 	pv->pv_flags = flags;
1004 
1005 	pvp = &SLIST_FIRST(&md->pvh_list);
1006 #ifdef PMAP_CACHE_VIPT
1007 	/*
1008 	 * Insert unmanaged entries, writeable first, at the head of
1009 	 * the pv list.
1010 	 */
1011 	if (__predict_true(!PV_IS_KENTRY_P(flags))) {
1012 		while (*pvp != NULL && PV_IS_KENTRY_P((*pvp)->pv_flags))
1013 			pvp = &SLIST_NEXT(*pvp, pv_link);
1014 	}
1015 	if (!PV_IS_WRITE_P(flags)) {
1016 		while (*pvp != NULL && PV_IS_WRITE_P((*pvp)->pv_flags))
1017 			pvp = &SLIST_NEXT(*pvp, pv_link);
1018 	}
1019 #endif
1020 	SLIST_NEXT(pv, pv_link) = *pvp;		/* add to ... */
1021 	*pvp = pv;				/* ... locked list */
1022 	md->pvh_attrs |= flags & (PVF_REF | PVF_MOD);
1023 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
1024 	if ((pv->pv_flags & PVF_KWRITE) == PVF_KWRITE)
1025 		md->pvh_attrs |= PVF_KMOD;
1026 	if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) != PVF_NC)
1027 		md->pvh_attrs |= PVF_DIRTY;
1028 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1029 #endif
1030 	if (pm == pmap_kernel()) {
1031 		PMAPCOUNT(kernel_mappings);
1032 		if (flags & PVF_WRITE)
1033 			md->krw_mappings++;
1034 		else
1035 			md->kro_mappings++;
1036 	} else {
1037 		if (flags & PVF_WRITE)
1038 			md->urw_mappings++;
1039 		else
1040 			md->uro_mappings++;
1041 	}
1042 
1043 #ifdef PMAP_CACHE_VIPT
1044 #ifndef ARM_MMU_EXTENDED
1045 	/*
1046 	 * Even though pmap_vac_me_harder will set PVF_WRITE for us,
1047 	 * do it here as well to keep the mappings & KVF_WRITE consistent.
1048 	 */
1049 	if (arm_cache_prefer_mask != 0 && (flags & PVF_WRITE) != 0) {
1050 		md->pvh_attrs |= PVF_WRITE;
1051 	}
1052 #endif
1053 	/*
1054 	 * If this is an exec mapping and its the first exec mapping
1055 	 * for this page, make sure to sync the I-cache.
1056 	 */
1057 	if (PV_IS_EXEC_P(flags)) {
1058 #ifndef ARM_MMU_EXTENDED
1059 		if (!PV_IS_EXEC_P(md->pvh_attrs)) {
1060 			pmap_syncicache_page(md, pa);
1061 			PMAPCOUNT(exec_synced_map);
1062 		}
1063 #endif
1064 		PMAPCOUNT(exec_mappings);
1065 	}
1066 #endif
1067 
1068 	PMAPCOUNT(mappings);
1069 
1070 	if (pv->pv_flags & PVF_WIRED)
1071 		++pm->pm_stats.wired_count;
1072 }
1073 
1074 /*
1075  *
1076  * pmap_find_pv: Find a pv entry
1077  *
1078  * => caller should hold lock on vm_page
1079  */
1080 static inline struct pv_entry *
1081 pmap_find_pv(struct vm_page_md *md, pmap_t pm, vaddr_t va)
1082 {
1083 	struct pv_entry *pv;
1084 
1085 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1086 		if (pm == pv->pv_pmap && va == pv->pv_va)
1087 			break;
1088 	}
1089 
1090 	return (pv);
1091 }
1092 
1093 /*
1094  * pmap_remove_pv: try to remove a mapping from a pv_list
1095  *
1096  * => caller should hold proper lock on pmap_main_lock
1097  * => pmap should be locked
1098  * => caller should hold lock on vm_page [so that attrs can be adjusted]
1099  * => caller should adjust ptp's wire_count and free PTP if needed
1100  * => caller should NOT adjust pmap's wire_count
1101  * => we return the removed pv
1102  */
1103 static struct pv_entry *
1104 pmap_remove_pv(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1105 {
1106 	struct pv_entry *pv, **prevptr;
1107 
1108 	NPDEBUG(PDB_PVDUMP,
1109 	    printf("pmap_remove_pv: pm %p, md %p, va 0x%08lx\n", pm, md, va));
1110 
1111 	prevptr = &SLIST_FIRST(&md->pvh_list); /* prev pv_entry ptr */
1112 	pv = *prevptr;
1113 
1114 	while (pv) {
1115 		if (pv->pv_pmap == pm && pv->pv_va == va) {	/* match? */
1116 			NPDEBUG(PDB_PVDUMP, printf("pmap_remove_pv: pm %p, md "
1117 			    "%p, flags 0x%x\n", pm, md, pv->pv_flags));
1118 			if (pv->pv_flags & PVF_WIRED) {
1119 				--pm->pm_stats.wired_count;
1120 			}
1121 			*prevptr = SLIST_NEXT(pv, pv_link);	/* remove it! */
1122 			if (pm == pmap_kernel()) {
1123 				PMAPCOUNT(kernel_unmappings);
1124 				if (pv->pv_flags & PVF_WRITE)
1125 					md->krw_mappings--;
1126 				else
1127 					md->kro_mappings--;
1128 			} else {
1129 				if (pv->pv_flags & PVF_WRITE)
1130 					md->urw_mappings--;
1131 				else
1132 					md->uro_mappings--;
1133 			}
1134 
1135 			PMAPCOUNT(unmappings);
1136 #ifdef PMAP_CACHE_VIPT
1137 			if (!(pv->pv_flags & PVF_WRITE))
1138 				break;
1139 			/*
1140 			 * If this page has had an exec mapping, then if
1141 			 * this was the last mapping, discard the contents,
1142 			 * otherwise sync the i-cache for this page.
1143 			 */
1144 			if (PV_IS_EXEC_P(md->pvh_attrs)) {
1145 #ifdef ARM_MMU_EXTENDED
1146 				md->pvh_attrs &= ~PVF_EXEC;
1147 				PMAPCOUNT(exec_discarded_unmap);
1148 #else
1149 				if (SLIST_EMPTY(&md->pvh_list)) {
1150 					md->pvh_attrs &= ~PVF_EXEC;
1151 					PMAPCOUNT(exec_discarded_unmap);
1152 				} else {
1153 					pmap_syncicache_page(md, pa);
1154 					PMAPCOUNT(exec_synced_unmap);
1155 				}
1156 #endif /* ARM_MMU_EXTENDED */
1157 			}
1158 #endif /* PMAP_CACHE_VIPT */
1159 			break;
1160 		}
1161 		prevptr = &SLIST_NEXT(pv, pv_link);	/* previous pointer */
1162 		pv = *prevptr;				/* advance */
1163 	}
1164 
1165 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
1166 	/*
1167 	 * If we no longer have a WRITEABLE KENTRY at the head of list,
1168 	 * clear the KMOD attribute from the page.
1169 	 */
1170 	if (SLIST_FIRST(&md->pvh_list) == NULL
1171 	    || (SLIST_FIRST(&md->pvh_list)->pv_flags & PVF_KWRITE) != PVF_KWRITE)
1172 		md->pvh_attrs &= ~PVF_KMOD;
1173 
1174 	/*
1175 	 * If this was a writeable page and there are no more writeable
1176 	 * mappings (ignoring KMPAGE), clear the WRITE flag and writeback
1177 	 * the contents to memory.
1178 	 */
1179 	if (arm_cache_prefer_mask != 0) {
1180 		if (md->krw_mappings + md->urw_mappings == 0)
1181 			md->pvh_attrs &= ~PVF_WRITE;
1182 		PMAP_VALIDATE_MD_PAGE(md);
1183 	}
1184 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1185 #endif /* PMAP_CACHE_VIPT && !ARM_MMU_EXTENDED */
1186 
1187 	return(pv);				/* return removed pv */
1188 }
1189 
1190 /*
1191  *
1192  * pmap_modify_pv: Update pv flags
1193  *
1194  * => caller should hold lock on vm_page [so that attrs can be adjusted]
1195  * => caller should NOT adjust pmap's wire_count
1196  * => caller must call pmap_vac_me_harder() if writable status of a page
1197  *    may have changed.
1198  * => we return the old flags
1199  *
1200  * Modify a physical-virtual mapping in the pv table
1201  */
1202 static u_int
1203 pmap_modify_pv(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va,
1204     u_int clr_mask, u_int set_mask)
1205 {
1206 	struct pv_entry *npv;
1207 	u_int flags, oflags;
1208 
1209 	KASSERT(!PV_IS_KENTRY_P(clr_mask));
1210 	KASSERT(!PV_IS_KENTRY_P(set_mask));
1211 
1212 	if ((npv = pmap_find_pv(md, pm, va)) == NULL)
1213 		return (0);
1214 
1215 	NPDEBUG(PDB_PVDUMP,
1216 	    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));
1217 
1218 	/*
1219 	 * There is at least one VA mapping this page.
1220 	 */
1221 
1222 	if (clr_mask & (PVF_REF | PVF_MOD)) {
1223 		md->pvh_attrs |= set_mask & (PVF_REF | PVF_MOD);
1224 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
1225 		if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) != PVF_NC)
1226 			md->pvh_attrs |= PVF_DIRTY;
1227 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1228 #endif /* PMAP_CACHE_VIPT && !ARM_MMU_EXTENDED */
1229 	}
1230 
1231 	oflags = npv->pv_flags;
1232 	npv->pv_flags = flags = (oflags & ~clr_mask) | set_mask;
1233 
1234 	if ((flags ^ oflags) & PVF_WIRED) {
1235 		if (flags & PVF_WIRED)
1236 			++pm->pm_stats.wired_count;
1237 		else
1238 			--pm->pm_stats.wired_count;
1239 	}
1240 
1241 	if ((flags ^ oflags) & PVF_WRITE) {
1242 		if (pm == pmap_kernel()) {
1243 			if (flags & PVF_WRITE) {
1244 				md->krw_mappings++;
1245 				md->kro_mappings--;
1246 			} else {
1247 				md->kro_mappings++;
1248 				md->krw_mappings--;
1249 			}
1250 		} else {
1251 			if (flags & PVF_WRITE) {
1252 				md->urw_mappings++;
1253 				md->uro_mappings--;
1254 			} else {
1255 				md->uro_mappings++;
1256 				md->urw_mappings--;
1257 			}
1258 		}
1259 	}
1260 #ifdef PMAP_CACHE_VIPT
1261 	if (arm_cache_prefer_mask != 0) {
1262 		if (md->urw_mappings + md->krw_mappings == 0) {
1263 			md->pvh_attrs &= ~PVF_WRITE;
1264 		} else {
1265 			md->pvh_attrs |= PVF_WRITE;
1266 		}
1267 	}
1268 #ifndef ARM_MMU_EXTENDED
1269 	/*
1270 	 * We have two cases here: the first is from enter_pv (new exec
1271 	 * page), the second is a combined pmap_remove_pv/pmap_enter_pv.
1272 	 * Since in latter, pmap_enter_pv won't do anything, we just have
1273 	 * to do what pmap_remove_pv would do.
1274 	 */
1275 	if ((PV_IS_EXEC_P(flags) && !PV_IS_EXEC_P(md->pvh_attrs))
1276 	    || (PV_IS_EXEC_P(md->pvh_attrs)
1277 		|| (!(flags & PVF_WRITE) && (oflags & PVF_WRITE)))) {
1278 		pmap_syncicache_page(md, pa);
1279 		PMAPCOUNT(exec_synced_remap);
1280 	}
1281 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1282 #endif /* !ARM_MMU_EXTENDED */
1283 #endif /* PMAP_CACHE_VIPT */
1284 
1285 	PMAPCOUNT(remappings);
1286 
1287 	return (oflags);
1288 }
1289 
1290 /*
1291  * Allocate an L1 translation table for the specified pmap.
1292  * This is called at pmap creation time.
1293  */
1294 static void
1295 pmap_alloc_l1(pmap_t pm)
1296 {
1297 #ifdef ARM_MMU_EXTENDED
1298 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
1299 #ifdef PMAP_NEED_ALLOC_POOLPAGE
1300 	struct vm_page *pg = arm_pmap_alloc_poolpage(UVM_PGA_ZERO);
1301 #else
1302 	struct vm_page *pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_ZERO);
1303 #endif
1304 	bool ok __diagused;
1305 	KASSERT(pg != NULL);
1306 	pm->pm_l1_pa = VM_PAGE_TO_PHYS(pg);
1307 	vaddr_t va = pmap_direct_mapped_phys(pm->pm_l1_pa, &ok, 0xdeadbeef);
1308 	KASSERT(ok);
1309 	KASSERT(va >= KERNEL_BASE);
1310 
1311 #else
1312 	KASSERTMSG(kernel_map != NULL, "pm %p", pm);
1313 	vaddr_t va = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
1314 	    UVM_KMF_WIRED|UVM_KMF_ZERO);
1315 	KASSERT(!va);
1316 	pmap_extract(pmap_kernel(), va, &pm->pm_l1_pa);
1317 #endif
1318 	pm->pm_l1 = (pd_entry_t *)va;
1319 #else
1320 	struct l1_ttable *l1;
1321 	uint8_t domain;
1322 
1323 	/*
1324 	 * Remove the L1 at the head of the LRU list
1325 	 */
1326 	mutex_spin_enter(&l1_lru_lock);
1327 	l1 = TAILQ_FIRST(&l1_lru_list);
1328 	KDASSERT(l1 != NULL);
1329 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1330 
1331 	/*
1332 	 * Pick the first available domain number, and update
1333 	 * the link to the next number.
1334 	 */
1335 	domain = l1->l1_domain_first;
1336 	l1->l1_domain_first = l1->l1_domain_free[domain];
1337 
1338 	/*
1339 	 * If there are still free domain numbers in this L1,
1340 	 * put it back on the TAIL of the LRU list.
1341 	 */
1342 	if (++l1->l1_domain_use_count < PMAP_DOMAINS)
1343 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1344 
1345 	mutex_spin_exit(&l1_lru_lock);
1346 
1347 	/*
1348 	 * Fix up the relevant bits in the pmap structure
1349 	 */
1350 	pm->pm_l1 = l1;
1351 	pm->pm_domain = domain + 1;
1352 #endif
1353 }
1354 
1355 /*
1356  * Free an L1 translation table.
1357  * This is called at pmap destruction time.
1358  */
1359 static void
1360 pmap_free_l1(pmap_t pm)
1361 {
1362 #ifdef ARM_MMU_EXTENDED
1363 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
1364 	struct vm_page *pg = PHYS_TO_VM_PAGE(pm->pm_l1_pa);
1365 	uvm_pagefree(pg);
1366 #else
1367 	uvm_km_free(kernel_map, (vaddr_t)pm->pm_l1, PAGE_SIZE, UVM_KMF_WIRED);
1368 #endif
1369 	pm->pm_l1 = NULL;
1370 	pm->pm_l1_pa = 0;
1371 #else
1372 	struct l1_ttable *l1 = pm->pm_l1;
1373 
1374 	mutex_spin_enter(&l1_lru_lock);
1375 
1376 	/*
1377 	 * If this L1 is currently on the LRU list, remove it.
1378 	 */
1379 	if (l1->l1_domain_use_count < PMAP_DOMAINS)
1380 		TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1381 
1382 	/*
1383 	 * Free up the domain number which was allocated to the pmap
1384 	 */
1385 	l1->l1_domain_free[pmap_domain(pm) - 1] = l1->l1_domain_first;
1386 	l1->l1_domain_first = pmap_domain(pm) - 1;
1387 	l1->l1_domain_use_count--;
1388 
1389 	/*
1390 	 * The L1 now must have at least 1 free domain, so add
1391 	 * it back to the LRU list. If the use count is zero,
1392 	 * put it at the head of the list, otherwise it goes
1393 	 * to the tail.
1394 	 */
1395 	if (l1->l1_domain_use_count == 0)
1396 		TAILQ_INSERT_HEAD(&l1_lru_list, l1, l1_lru);
1397 	else
1398 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1399 
1400 	mutex_spin_exit(&l1_lru_lock);
1401 #endif /* ARM_MMU_EXTENDED */
1402 }
1403 
1404 #ifndef ARM_MMU_EXTENDED
1405 static inline void
1406 pmap_use_l1(pmap_t pm)
1407 {
1408 	struct l1_ttable *l1;
1409 
1410 	/*
1411 	 * Do nothing if we're in interrupt context.
1412 	 * Access to an L1 by the kernel pmap must not affect
1413 	 * the LRU list.
1414 	 */
1415 	if (cpu_intr_p() || pm == pmap_kernel())
1416 		return;
1417 
1418 	l1 = pm->pm_l1;
1419 
1420 	/*
1421 	 * If the L1 is not currently on the LRU list, just return
1422 	 */
1423 	if (l1->l1_domain_use_count == PMAP_DOMAINS)
1424 		return;
1425 
1426 	mutex_spin_enter(&l1_lru_lock);
1427 
1428 	/*
1429 	 * Check the use count again, now that we've acquired the lock
1430 	 */
1431 	if (l1->l1_domain_use_count == PMAP_DOMAINS) {
1432 		mutex_spin_exit(&l1_lru_lock);
1433 		return;
1434 	}
1435 
1436 	/*
1437 	 * Move the L1 to the back of the LRU list
1438 	 */
1439 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1440 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1441 
1442 	mutex_spin_exit(&l1_lru_lock);
1443 }
1444 #endif /* !ARM_MMU_EXTENDED */
1445 
1446 /*
1447  * void pmap_free_l2_ptp(pt_entry_t *, paddr_t *)
1448  *
1449  * Free an L2 descriptor table.
1450  */
1451 static inline void
1452 #if defined(PMAP_INCLUDE_PTE_SYNC) && defined(PMAP_CACHE_VIVT)
1453 pmap_free_l2_ptp(bool need_sync, pt_entry_t *l2, paddr_t pa)
1454 #else
1455 pmap_free_l2_ptp(pt_entry_t *l2, paddr_t pa)
1456 #endif
1457 {
1458 #if defined(PMAP_INCLUDE_PTE_SYNC) && defined(PMAP_CACHE_VIVT)
1459 	/*
1460 	 * Note: With a write-back cache, we may need to sync this
1461 	 * L2 table before re-using it.
1462 	 * This is because it may have belonged to a non-current
1463 	 * pmap, in which case the cache syncs would have been
1464 	 * skipped for the pages that were being unmapped. If the
1465 	 * L2 table were then to be immediately re-allocated to
1466 	 * the *current* pmap, it may well contain stale mappings
1467 	 * which have not yet been cleared by a cache write-back
1468 	 * and so would still be visible to the mmu.
1469 	 */
1470 	if (need_sync)
1471 		PTE_SYNC_RANGE(l2, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1472 #endif /* PMAP_INCLUDE_PTE_SYNC && PMAP_CACHE_VIVT */
1473 	pool_cache_put_paddr(&pmap_l2ptp_cache, (void *)l2, pa);
1474 }
1475 
1476 /*
1477  * Returns a pointer to the L2 bucket associated with the specified pmap
1478  * and VA, or NULL if no L2 bucket exists for the address.
1479  */
1480 static inline struct l2_bucket *
1481 pmap_get_l2_bucket(pmap_t pm, vaddr_t va)
1482 {
1483 	const size_t l1slot = l1pte_index(va);
1484 	struct l2_dtable *l2;
1485 	struct l2_bucket *l2b;
1486 
1487 	if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL ||
1488 	    (l2b = &l2->l2_bucket[L2_BUCKET(l1slot)])->l2b_kva == NULL)
1489 		return (NULL);
1490 
1491 	return (l2b);
1492 }
1493 
1494 /*
1495  * Returns a pointer to the L2 bucket associated with the specified pmap
1496  * and VA.
1497  *
1498  * If no L2 bucket exists, perform the necessary allocations to put an L2
1499  * bucket/page table in place.
1500  *
1501  * Note that if a new L2 bucket/page was allocated, the caller *must*
1502  * increment the bucket occupancy counter appropriately *before*
1503  * releasing the pmap's lock to ensure no other thread or cpu deallocates
1504  * the bucket/page in the meantime.
1505  */
1506 static struct l2_bucket *
1507 pmap_alloc_l2_bucket(pmap_t pm, vaddr_t va)
1508 {
1509 	const size_t l1slot = l1pte_index(va);
1510 	struct l2_dtable *l2;
1511 
1512 	if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL) {
1513 		/*
1514 		 * No mapping at this address, as there is
1515 		 * no entry in the L1 table.
1516 		 * Need to allocate a new l2_dtable.
1517 		 */
1518 		if ((l2 = pmap_alloc_l2_dtable()) == NULL)
1519 			return (NULL);
1520 
1521 		/*
1522 		 * Link it into the parent pmap
1523 		 */
1524 		pm->pm_l2[L2_IDX(l1slot)] = l2;
1525 	}
1526 
1527 	struct l2_bucket * const l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
1528 
1529 	/*
1530 	 * Fetch pointer to the L2 page table associated with the address.
1531 	 */
1532 	if (l2b->l2b_kva == NULL) {
1533 		pt_entry_t *ptep;
1534 
1535 		/*
1536 		 * No L2 page table has been allocated. Chances are, this
1537 		 * is because we just allocated the l2_dtable, above.
1538 		 */
1539 		if ((ptep = pmap_alloc_l2_ptp(&l2b->l2b_pa)) == NULL) {
1540 			/*
1541 			 * Oops, no more L2 page tables available at this
1542 			 * time. We may need to deallocate the l2_dtable
1543 			 * if we allocated a new one above.
1544 			 */
1545 			if (l2->l2_occupancy == 0) {
1546 				pm->pm_l2[L2_IDX(l1slot)] = NULL;
1547 				pmap_free_l2_dtable(l2);
1548 			}
1549 			return (NULL);
1550 		}
1551 
1552 		l2->l2_occupancy++;
1553 		l2b->l2b_kva = ptep;
1554 		l2b->l2b_l1slot = l1slot;
1555 
1556 #ifdef ARM_MMU_EXTENDED
1557 		/*
1558 		 * We know there will be a mapping here, so simply
1559 		 * enter this PTP into the L1 now.
1560 		 */
1561 		pd_entry_t * const pdep = pmap_l1_kva(pm) + l1slot;
1562 		pd_entry_t npde = L1_C_PROTO | l2b->l2b_pa
1563 		    | L1_C_DOM(pmap_domain(pm));
1564 		KASSERT(*pdep == 0);
1565 		l1pte_setone(pdep, npde);
1566 		PTE_SYNC(pdep);
1567 #endif
1568 	}
1569 
1570 	return (l2b);
1571 }
1572 
1573 /*
1574  * One or more mappings in the specified L2 descriptor table have just been
1575  * invalidated.
1576  *
1577  * Garbage collect the metadata and descriptor table itself if necessary.
1578  *
1579  * The pmap lock must be acquired when this is called (not necessary
1580  * for the kernel pmap).
1581  */
1582 static void
1583 pmap_free_l2_bucket(pmap_t pm, struct l2_bucket *l2b, u_int count)
1584 {
1585 	KDASSERT(count <= l2b->l2b_occupancy);
1586 
1587 	/*
1588 	 * Update the bucket's reference count according to how many
1589 	 * PTEs the caller has just invalidated.
1590 	 */
1591 	l2b->l2b_occupancy -= count;
1592 
1593 	/*
1594 	 * Note:
1595 	 *
1596 	 * Level 2 page tables allocated to the kernel pmap are never freed
1597 	 * as that would require checking all Level 1 page tables and
1598 	 * removing any references to the Level 2 page table. See also the
1599 	 * comment elsewhere about never freeing bootstrap L2 descriptors.
1600 	 *
1601 	 * We make do with just invalidating the mapping in the L2 table.
1602 	 *
1603 	 * This isn't really a big deal in practice and, in fact, leads
1604 	 * to a performance win over time as we don't need to continually
1605 	 * alloc/free.
1606 	 */
1607 	if (l2b->l2b_occupancy > 0 || pm == pmap_kernel())
1608 		return;
1609 
1610 	/*
1611 	 * There are no more valid mappings in this level 2 page table.
1612 	 * Go ahead and NULL-out the pointer in the bucket, then
1613 	 * free the page table.
1614 	 */
1615 	const size_t l1slot = l2b->l2b_l1slot;
1616 	pt_entry_t * const ptep = l2b->l2b_kva;
1617 	l2b->l2b_kva = NULL;
1618 
1619 	pd_entry_t * const pdep = pmap_l1_kva(pm) + l1slot;
1620 	pd_entry_t pde __diagused = *pdep;
1621 
1622 #ifdef ARM_MMU_EXTENDED
1623 	/*
1624 	 * Invalidate the L1 slot.
1625 	 */
1626 	KASSERT((pde & L1_TYPE_MASK) == L1_TYPE_C);
1627 #else
1628 	/*
1629 	 * If the L1 slot matches the pmap's domain number, then invalidate it.
1630 	 */
1631 	if ((pde & (L1_C_DOM_MASK|L1_TYPE_MASK))
1632 	    == (L1_C_DOM(pmap_domain(pm))|L1_TYPE_C)) {
1633 #endif
1634 		l1pte_setone(pdep, 0);
1635 		PDE_SYNC(pdep);
1636 #ifndef ARM_MMU_EXTENDED
1637 	}
1638 #endif
1639 
1640 	/*
1641 	 * Release the L2 descriptor table back to the pool cache.
1642 	 */
1643 #if defined(PMAP_INCLUDE_PTE_SYNC) && defined(PMAP_CACHE_VIVT)
1644 	pmap_free_l2_ptp(!pmap_is_cached(pm), ptep, l2b->l2b_pa);
1645 #else
1646 	pmap_free_l2_ptp(ptep, l2b->l2b_pa);
1647 #endif
1648 
1649 	/*
1650 	 * Update the reference count in the associated l2_dtable
1651 	 */
1652 	struct l2_dtable * const l2 = pm->pm_l2[L2_IDX(l1slot)];
1653 	if (--l2->l2_occupancy > 0)
1654 		return;
1655 
1656 	/*
1657 	 * There are no more valid mappings in any of the Level 1
1658 	 * slots managed by this l2_dtable. Go ahead and NULL-out
1659 	 * the pointer in the parent pmap and free the l2_dtable.
1660 	 */
1661 	pm->pm_l2[L2_IDX(l1slot)] = NULL;
1662 	pmap_free_l2_dtable(l2);
1663 }
1664 
1665 /*
1666  * Pool cache constructors for L2 descriptor tables, metadata and pmap
1667  * structures.
1668  */
1669 static int
1670 pmap_l2ptp_ctor(void *arg, void *v, int flags)
1671 {
1672 #ifndef PMAP_INCLUDE_PTE_SYNC
1673 	vaddr_t va = (vaddr_t)v & ~PGOFSET;
1674 
1675 	/*
1676 	 * The mappings for these page tables were initially made using
1677 	 * pmap_kenter_pa() by the pool subsystem. Therefore, the cache-
1678 	 * mode will not be right for page table mappings. To avoid
1679 	 * polluting the pmap_kenter_pa() code with a special case for
1680 	 * page tables, we simply fix up the cache-mode here if it's not
1681 	 * correct.
1682 	 */
1683 	if (pte_l2_s_cache_mode != pte_l2_s_cache_mode_pt) {
1684 		const struct l2_bucket * const l2b =
1685 		    pmap_get_l2_bucket(pmap_kernel(), va);
1686 		KASSERTMSG(l2b != NULL, "%#lx", va);
1687 		pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
1688 		const pt_entry_t opte = *ptep;
1689 
1690 		if ((opte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
1691 			/*
1692 			 * Page tables must have the cache-mode set correctly.
1693 			 */
1694 			const pt_entry_t npte = (pte & ~L2_S_CACHE_MASK)
1695 			    | pte_l2_s_cache_mode_pt;
1696 			l2pte_set(ptep, npte, opte);
1697 			PTE_SYNC(ptep);
1698 			cpu_tlb_flushD_SE(va);
1699 			cpu_cpwait();
1700 		}
1701 	}
1702 #endif
1703 
1704 	memset(v, 0, L2_TABLE_SIZE_REAL);
1705 	PTE_SYNC_RANGE(v, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1706 	return (0);
1707 }
1708 
1709 static int
1710 pmap_l2dtable_ctor(void *arg, void *v, int flags)
1711 {
1712 
1713 	memset(v, 0, sizeof(struct l2_dtable));
1714 	return (0);
1715 }
1716 
1717 static int
1718 pmap_pmap_ctor(void *arg, void *v, int flags)
1719 {
1720 
1721 	memset(v, 0, sizeof(struct pmap));
1722 	return (0);
1723 }
1724 
1725 static void
1726 pmap_pinit(pmap_t pm)
1727 {
1728 #ifndef ARM_HAS_VBAR
1729 	struct l2_bucket *l2b;
1730 
1731 	if (vector_page < KERNEL_BASE) {
1732 		/*
1733 		 * Map the vector page.
1734 		 */
1735 		pmap_enter(pm, vector_page, systempage.pv_pa,
1736 		    VM_PROT_READ | VM_PROT_EXECUTE,
1737 		    VM_PROT_READ | VM_PROT_EXECUTE | PMAP_WIRED);
1738 		pmap_update(pm);
1739 
1740 		pm->pm_pl1vec = pmap_l1_kva(pm) + l1pte_index(vector_page);
1741 		l2b = pmap_get_l2_bucket(pm, vector_page);
1742 		KASSERTMSG(l2b != NULL, "%#lx", vector_page);
1743 		pm->pm_l1vec = l2b->l2b_pa | L1_C_PROTO |
1744 		    L1_C_DOM(pmap_domain(pm));
1745 	} else
1746 		pm->pm_pl1vec = NULL;
1747 #endif
1748 }
1749 
1750 #ifdef PMAP_CACHE_VIVT
1751 /*
1752  * Since we have a virtually indexed cache, we may need to inhibit caching if
1753  * there is more than one mapping and at least one of them is writable.
1754  * Since we purge the cache on every context switch, we only need to check for
1755  * other mappings within the same pmap, or kernel_pmap.
1756  * This function is also called when a page is unmapped, to possibly reenable
1757  * caching on any remaining mappings.
1758  *
1759  * The code implements the following logic, where:
1760  *
1761  * KW = # of kernel read/write pages
1762  * KR = # of kernel read only pages
1763  * UW = # of user read/write pages
1764  * UR = # of user read only pages
1765  *
1766  * KC = kernel mapping is cacheable
1767  * UC = user mapping is cacheable
1768  *
1769  *               KW=0,KR=0  KW=0,KR>0  KW=1,KR=0  KW>1,KR>=0
1770  *             +---------------------------------------------
1771  * UW=0,UR=0   | ---        KC=1       KC=1       KC=0
1772  * UW=0,UR>0   | UC=1       KC=1,UC=1  KC=0,UC=0  KC=0,UC=0
1773  * UW=1,UR=0   | UC=1       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1774  * UW>1,UR>=0  | UC=0       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1775  */
1776 
1777 static const int pmap_vac_flags[4][4] = {
1778 	{-1,		0,		0,		PVF_KNC},
1779 	{0,		0,		PVF_NC,		PVF_NC},
1780 	{0,		PVF_NC,		PVF_NC,		PVF_NC},
1781 	{PVF_UNC,	PVF_NC,		PVF_NC,		PVF_NC}
1782 };
1783 
1784 static inline int
1785 pmap_get_vac_flags(const struct vm_page_md *md)
1786 {
1787 	int kidx, uidx;
1788 
1789 	kidx = 0;
1790 	if (md->kro_mappings || md->krw_mappings > 1)
1791 		kidx |= 1;
1792 	if (md->krw_mappings)
1793 		kidx |= 2;
1794 
1795 	uidx = 0;
1796 	if (md->uro_mappings || md->urw_mappings > 1)
1797 		uidx |= 1;
1798 	if (md->urw_mappings)
1799 		uidx |= 2;
1800 
1801 	return (pmap_vac_flags[uidx][kidx]);
1802 }
1803 
1804 static inline void
1805 pmap_vac_me_harder(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1806 {
1807 	int nattr;
1808 
1809 	nattr = pmap_get_vac_flags(md);
1810 
1811 	if (nattr < 0) {
1812 		md->pvh_attrs &= ~PVF_NC;
1813 		return;
1814 	}
1815 
1816 	if (nattr == 0 && (md->pvh_attrs & PVF_NC) == 0)
1817 		return;
1818 
1819 	if (pm == pmap_kernel())
1820 		pmap_vac_me_kpmap(md, pa, pm, va);
1821 	else
1822 		pmap_vac_me_user(md, pa, pm, va);
1823 
1824 	md->pvh_attrs = (md->pvh_attrs & ~PVF_NC) | nattr;
1825 }
1826 
1827 static void
1828 pmap_vac_me_kpmap(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1829 {
1830 	u_int u_cacheable, u_entries;
1831 	struct pv_entry *pv;
1832 	pmap_t last_pmap = pm;
1833 
1834 	/*
1835 	 * Pass one, see if there are both kernel and user pmaps for
1836 	 * this page.  Calculate whether there are user-writable or
1837 	 * kernel-writable pages.
1838 	 */
1839 	u_cacheable = 0;
1840 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1841 		if (pv->pv_pmap != pm && (pv->pv_flags & PVF_NC) == 0)
1842 			u_cacheable++;
1843 	}
1844 
1845 	u_entries = md->urw_mappings + md->uro_mappings;
1846 
1847 	/*
1848 	 * We know we have just been updating a kernel entry, so if
1849 	 * all user pages are already cacheable, then there is nothing
1850 	 * further to do.
1851 	 */
1852 	if (md->k_mappings == 0 && u_cacheable == u_entries)
1853 		return;
1854 
1855 	if (u_entries) {
1856 		/*
1857 		 * Scan over the list again, for each entry, if it
1858 		 * might not be set correctly, call pmap_vac_me_user
1859 		 * to recalculate the settings.
1860 		 */
1861 		SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1862 			/*
1863 			 * We know kernel mappings will get set
1864 			 * correctly in other calls.  We also know
1865 			 * that if the pmap is the same as last_pmap
1866 			 * then we've just handled this entry.
1867 			 */
1868 			if (pv->pv_pmap == pm || pv->pv_pmap == last_pmap)
1869 				continue;
1870 
1871 			/*
1872 			 * If there are kernel entries and this page
1873 			 * is writable but non-cacheable, then we can
1874 			 * skip this entry also.
1875 			 */
1876 			if (md->k_mappings &&
1877 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) ==
1878 			    (PVF_NC | PVF_WRITE))
1879 				continue;
1880 
1881 			/*
1882 			 * Similarly if there are no kernel-writable
1883 			 * entries and the page is already
1884 			 * read-only/cacheable.
1885 			 */
1886 			if (md->krw_mappings == 0 &&
1887 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 0)
1888 				continue;
1889 
1890 			/*
1891 			 * For some of the remaining cases, we know
1892 			 * that we must recalculate, but for others we
1893 			 * can't tell if they are correct or not, so
1894 			 * we recalculate anyway.
1895 			 */
1896 			pmap_vac_me_user(md, pa, (last_pmap = pv->pv_pmap), 0);
1897 		}
1898 
1899 		if (md->k_mappings == 0)
1900 			return;
1901 	}
1902 
1903 	pmap_vac_me_user(md, pa, pm, va);
1904 }
1905 
1906 static void
1907 pmap_vac_me_user(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1908 {
1909 	pmap_t kpmap = pmap_kernel();
1910 	struct pv_entry *pv, *npv = NULL;
1911 	u_int entries = 0;
1912 	u_int writable = 0;
1913 	u_int cacheable_entries = 0;
1914 	u_int kern_cacheable = 0;
1915 	u_int other_writable = 0;
1916 
1917 	/*
1918 	 * Count mappings and writable mappings in this pmap.
1919 	 * Include kernel mappings as part of our own.
1920 	 * Keep a pointer to the first one.
1921 	 */
1922 	npv = NULL;
1923 	KASSERT(pmap_page_locked_p(md));
1924 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1925 		/* Count mappings in the same pmap */
1926 		if (pm == pv->pv_pmap || kpmap == pv->pv_pmap) {
1927 			if (entries++ == 0)
1928 				npv = pv;
1929 
1930 			/* Cacheable mappings */
1931 			if ((pv->pv_flags & PVF_NC) == 0) {
1932 				cacheable_entries++;
1933 				if (kpmap == pv->pv_pmap)
1934 					kern_cacheable++;
1935 			}
1936 
1937 			/* Writable mappings */
1938 			if (pv->pv_flags & PVF_WRITE)
1939 				++writable;
1940 		} else
1941 		if (pv->pv_flags & PVF_WRITE)
1942 			other_writable = 1;
1943 	}
1944 
1945 	/*
1946 	 * Enable or disable caching as necessary.
1947 	 * Note: the first entry might be part of the kernel pmap,
1948 	 * so we can't assume this is indicative of the state of the
1949 	 * other (maybe non-kpmap) entries.
1950 	 */
1951 	if ((entries > 1 && writable) ||
1952 	    (entries > 0 && pm == kpmap && other_writable)) {
1953 		if (cacheable_entries == 0) {
1954 			return;
1955 		}
1956 
1957 		for (pv = npv; pv; pv = SLIST_NEXT(pv, pv_link)) {
1958 			if ((pm != pv->pv_pmap && kpmap != pv->pv_pmap) ||
1959 			    (pv->pv_flags & PVF_NC))
1960 				continue;
1961 
1962 			pv->pv_flags |= PVF_NC;
1963 
1964 			struct l2_bucket * const l2b
1965 			    = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
1966 			KASSERTMSG(l2b != NULL, "%#lx", va);
1967 			pt_entry_t * const ptep
1968 			    = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1969 			const pt_entry_t opte = *ptep;
1970 			pt_entry_t npte = opte & ~L2_S_CACHE_MASK;
1971 
1972 			if ((va != pv->pv_va || pm != pv->pv_pmap)
1973 			    && l2pte_valid_p(npte)) {
1974 #ifdef PMAP_CACHE_VIVT
1975 				pmap_cache_wbinv_page(pv->pv_pmap, pv->pv_va,
1976 				    true, pv->pv_flags);
1977 #endif
1978 				pmap_tlb_flush_SE(pv->pv_pmap, pv->pv_va,
1979 				    pv->pv_flags);
1980 			}
1981 
1982 			l2pte_set(ptep, npte, opte);
1983 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
1984 		}
1985 		cpu_cpwait();
1986 	} else
1987 	if (entries > cacheable_entries) {
1988 		/*
1989 		 * Turn cacheing back on for some pages.  If it is a kernel
1990 		 * page, only do so if there are no other writable pages.
1991 		 */
1992 		for (pv = npv; pv; pv = SLIST_NEXT(pv, pv_link)) {
1993 			if (!(pv->pv_flags & PVF_NC) || (pm != pv->pv_pmap &&
1994 			    (kpmap != pv->pv_pmap || other_writable)))
1995 				continue;
1996 
1997 			pv->pv_flags &= ~PVF_NC;
1998 
1999 			struct l2_bucket * const l2b
2000 			    = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
2001 			KASSERTMSG(l2b != NULL, "%#lx", va);
2002 			pt_entry_t * const ptep
2003 			    = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2004 			const pt_entry_t opte = *ptep;
2005 			pt_entry_t npte = (opte & ~L2_S_CACHE_MASK)
2006 			    | pte_l2_s_cache_mode;
2007 
2008 			if (l2pte_valid_p(opte)) {
2009 				pmap_tlb_flush_SE(pv->pv_pmap, pv->pv_va,
2010 				    pv->pv_flags);
2011 			}
2012 
2013 			l2pte_set(ptep, npte, opte);
2014 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
2015 		}
2016 	}
2017 }
2018 #endif
2019 
2020 #ifdef PMAP_CACHE_VIPT
2021 static void
2022 pmap_vac_me_harder(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
2023 {
2024 #ifndef ARM_MMU_EXTENDED
2025 	struct pv_entry *pv;
2026 	vaddr_t tst_mask;
2027 	bool bad_alias;
2028 	const u_int
2029 	    rw_mappings = md->urw_mappings + md->krw_mappings,
2030 	    ro_mappings = md->uro_mappings + md->kro_mappings;
2031 
2032 	/* do we need to do anything? */
2033 	if (arm_cache_prefer_mask == 0)
2034 		return;
2035 
2036 	NPDEBUG(PDB_VAC, printf("pmap_vac_me_harder: md=%p, pmap=%p va=%08lx\n",
2037 	    md, pm, va));
2038 
2039 	KASSERT(!va || pm);
2040 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2041 
2042 	/* Already a conflict? */
2043 	if (__predict_false(md->pvh_attrs & PVF_NC)) {
2044 		/* just an add, things are already non-cached */
2045 		KASSERT(!(md->pvh_attrs & PVF_DIRTY));
2046 		KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2047 		bad_alias = false;
2048 		if (va) {
2049 			PMAPCOUNT(vac_color_none);
2050 			bad_alias = true;
2051 			KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2052 			goto fixup;
2053 		}
2054 		pv = SLIST_FIRST(&md->pvh_list);
2055 		/* the list can't be empty because it would be cachable */
2056 		if (md->pvh_attrs & PVF_KMPAGE) {
2057 			tst_mask = md->pvh_attrs;
2058 		} else {
2059 			KASSERT(pv);
2060 			tst_mask = pv->pv_va;
2061 			pv = SLIST_NEXT(pv, pv_link);
2062 		}
2063 		/*
2064 		 * Only check for a bad alias if we have writable mappings.
2065 		 */
2066 		tst_mask &= arm_cache_prefer_mask;
2067 		if (rw_mappings > 0) {
2068 			for (; pv && !bad_alias; pv = SLIST_NEXT(pv, pv_link)) {
2069 				/* if there's a bad alias, stop checking. */
2070 				if (tst_mask != (pv->pv_va & arm_cache_prefer_mask))
2071 					bad_alias = true;
2072 			}
2073 			md->pvh_attrs |= PVF_WRITE;
2074 			if (!bad_alias)
2075 				md->pvh_attrs |= PVF_DIRTY;
2076 		} else {
2077 			/*
2078 			 * We have only read-only mappings.  Let's see if there
2079 			 * are multiple colors in use or if we mapped a KMPAGE.
2080 			 * If the latter, we have a bad alias.  If the former,
2081 			 * we need to remember that.
2082 			 */
2083 			for (; pv; pv = SLIST_NEXT(pv, pv_link)) {
2084 				if (tst_mask != (pv->pv_va & arm_cache_prefer_mask)) {
2085 					if (md->pvh_attrs & PVF_KMPAGE)
2086 						bad_alias = true;
2087 					break;
2088 				}
2089 			}
2090 			md->pvh_attrs &= ~PVF_WRITE;
2091 			/*
2092 			 * No KMPAGE and we exited early, so we must have
2093 			 * multiple color mappings.
2094 			 */
2095 			if (!bad_alias && pv != NULL)
2096 				md->pvh_attrs |= PVF_MULTCLR;
2097 		}
2098 
2099 		/* If no conflicting colors, set everything back to cached */
2100 		if (!bad_alias) {
2101 #ifdef DEBUG
2102 			if ((md->pvh_attrs & PVF_WRITE)
2103 			    || ro_mappings < 2) {
2104 				SLIST_FOREACH(pv, &md->pvh_list, pv_link)
2105 					KDASSERT(((tst_mask ^ pv->pv_va) & arm_cache_prefer_mask) == 0);
2106 			}
2107 #endif
2108 			md->pvh_attrs &= (PAGE_SIZE - 1) & ~PVF_NC;
2109 			md->pvh_attrs |= tst_mask | PVF_COLORED;
2110 			/*
2111 			 * Restore DIRTY bit if page is modified
2112 			 */
2113 			if (md->pvh_attrs & PVF_DMOD)
2114 				md->pvh_attrs |= PVF_DIRTY;
2115 			PMAPCOUNT(vac_color_restore);
2116 		} else {
2117 			KASSERT(SLIST_FIRST(&md->pvh_list) != NULL);
2118 			KASSERT(SLIST_NEXT(SLIST_FIRST(&md->pvh_list), pv_link) != NULL);
2119 		}
2120 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2121 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2122 	} else if (!va) {
2123 		KASSERT(pmap_is_page_colored_p(md));
2124 		KASSERT(!(md->pvh_attrs & PVF_WRITE)
2125 		    || (md->pvh_attrs & PVF_DIRTY));
2126 		if (rw_mappings == 0) {
2127 			md->pvh_attrs &= ~PVF_WRITE;
2128 			if (ro_mappings == 1
2129 			    && (md->pvh_attrs & PVF_MULTCLR)) {
2130 				/*
2131 				 * If this is the last readonly mapping
2132 				 * but it doesn't match the current color
2133 				 * for the page, change the current color
2134 				 * to match this last readonly mapping.
2135 				 */
2136 				pv = SLIST_FIRST(&md->pvh_list);
2137 				tst_mask = (md->pvh_attrs ^ pv->pv_va)
2138 				    & arm_cache_prefer_mask;
2139 				if (tst_mask) {
2140 					md->pvh_attrs ^= tst_mask;
2141 					PMAPCOUNT(vac_color_change);
2142 				}
2143 			}
2144 		}
2145 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2146 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2147 		return;
2148 	} else if (!pmap_is_page_colored_p(md)) {
2149 		/* not colored so we just use its color */
2150 		KASSERT(md->pvh_attrs & (PVF_WRITE|PVF_DIRTY));
2151 		KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2152 		PMAPCOUNT(vac_color_new);
2153 		md->pvh_attrs &= PAGE_SIZE - 1;
2154 		md->pvh_attrs |= PVF_COLORED
2155 		    | (va & arm_cache_prefer_mask)
2156 		    | (rw_mappings > 0 ? PVF_WRITE : 0);
2157 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2158 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2159 		return;
2160 	} else if (((md->pvh_attrs ^ va) & arm_cache_prefer_mask) == 0) {
2161 		bad_alias = false;
2162 		if (rw_mappings > 0) {
2163 			/*
2164 			 * We now have writeable mappings and if we have
2165 			 * readonly mappings in more than once color, we have
2166 			 * an aliasing problem.  Regardless mark the page as
2167 			 * writeable.
2168 			 */
2169 			if (md->pvh_attrs & PVF_MULTCLR) {
2170 				if (ro_mappings < 2) {
2171 					/*
2172 					 * If we only have less than two
2173 					 * read-only mappings, just flush the
2174 					 * non-primary colors from the cache.
2175 					 */
2176 					pmap_flush_page(md, pa,
2177 					    PMAP_FLUSH_SECONDARY);
2178 				} else {
2179 					bad_alias = true;
2180 				}
2181 			}
2182 			md->pvh_attrs |= PVF_WRITE;
2183 		}
2184 		/* If no conflicting colors, set everything back to cached */
2185 		if (!bad_alias) {
2186 #ifdef DEBUG
2187 			if (rw_mappings > 0
2188 			    || (md->pvh_attrs & PMAP_KMPAGE)) {
2189 				tst_mask = md->pvh_attrs & arm_cache_prefer_mask;
2190 				SLIST_FOREACH(pv, &md->pvh_list, pv_link)
2191 					KDASSERT(((tst_mask ^ pv->pv_va) & arm_cache_prefer_mask) == 0);
2192 			}
2193 #endif
2194 			if (SLIST_EMPTY(&md->pvh_list))
2195 				PMAPCOUNT(vac_color_reuse);
2196 			else
2197 				PMAPCOUNT(vac_color_ok);
2198 
2199 			/* matching color, just return */
2200 			KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2201 			KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2202 			return;
2203 		}
2204 		KASSERT(SLIST_FIRST(&md->pvh_list) != NULL);
2205 		KASSERT(SLIST_NEXT(SLIST_FIRST(&md->pvh_list), pv_link) != NULL);
2206 
2207 		/* color conflict.  evict from cache. */
2208 
2209 		pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
2210 		md->pvh_attrs &= ~PVF_COLORED;
2211 		md->pvh_attrs |= PVF_NC;
2212 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2213 		KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2214 		PMAPCOUNT(vac_color_erase);
2215 	} else if (rw_mappings == 0
2216 		   && (md->pvh_attrs & PVF_KMPAGE) == 0) {
2217 		KASSERT((md->pvh_attrs & PVF_WRITE) == 0);
2218 
2219 		/*
2220 		 * If the page has dirty cache lines, clean it.
2221 		 */
2222 		if (md->pvh_attrs & PVF_DIRTY)
2223 			pmap_flush_page(md, pa, PMAP_CLEAN_PRIMARY);
2224 
2225 		/*
2226 		 * If this is the first remapping (we know that there are no
2227 		 * writeable mappings), then this is a simple color change.
2228 		 * Otherwise this is a seconary r/o mapping, which means
2229 		 * we don't have to do anything.
2230 		 */
2231 		if (ro_mappings == 1) {
2232 			KASSERT(((md->pvh_attrs ^ va) & arm_cache_prefer_mask) != 0);
2233 			md->pvh_attrs &= PAGE_SIZE - 1;
2234 			md->pvh_attrs |= (va & arm_cache_prefer_mask);
2235 			PMAPCOUNT(vac_color_change);
2236 		} else {
2237 			PMAPCOUNT(vac_color_blind);
2238 		}
2239 		md->pvh_attrs |= PVF_MULTCLR;
2240 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2241 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2242 		return;
2243 	} else {
2244 		if (rw_mappings > 0)
2245 			md->pvh_attrs |= PVF_WRITE;
2246 
2247 		/* color conflict.  evict from cache. */
2248 		pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
2249 
2250 		/* the list can't be empty because this was a enter/modify */
2251 		pv = SLIST_FIRST(&md->pvh_list);
2252 		if ((md->pvh_attrs & PVF_KMPAGE) == 0) {
2253 			KASSERT(pv);
2254 			/*
2255 			 * If there's only one mapped page, change color to the
2256 			 * page's new color and return.  Restore the DIRTY bit
2257 			 * that was erased by pmap_flush_page.
2258 			 */
2259 			if (SLIST_NEXT(pv, pv_link) == NULL) {
2260 				md->pvh_attrs &= PAGE_SIZE - 1;
2261 				md->pvh_attrs |= (va & arm_cache_prefer_mask);
2262 				if (md->pvh_attrs & PVF_DMOD)
2263 					md->pvh_attrs |= PVF_DIRTY;
2264 				PMAPCOUNT(vac_color_change);
2265 				KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2266 				KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2267 				KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2268 				return;
2269 			}
2270 		}
2271 		bad_alias = true;
2272 		md->pvh_attrs &= ~PVF_COLORED;
2273 		md->pvh_attrs |= PVF_NC;
2274 		PMAPCOUNT(vac_color_erase);
2275 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2276 	}
2277 
2278   fixup:
2279 	KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2280 
2281 	/*
2282 	 * Turn cacheing on/off for all pages.
2283 	 */
2284 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
2285 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pv->pv_pmap,
2286 		    pv->pv_va);
2287 		KASSERTMSG(l2b != NULL, "%#lx", va);
2288 		pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2289 		const pt_entry_t opte = *ptep;
2290 		pt_entry_t npte = opte & ~L2_S_CACHE_MASK;
2291 		if (bad_alias) {
2292 			pv->pv_flags |= PVF_NC;
2293 		} else {
2294 			pv->pv_flags &= ~PVF_NC;
2295 			npte |= pte_l2_s_cache_mode;
2296 		}
2297 
2298 		if (opte == npte)	/* only update is there's a change */
2299 			continue;
2300 
2301 		if (l2pte_valid_p(npte)) {
2302 			pmap_tlb_flush_SE(pv->pv_pmap, pv->pv_va, pv->pv_flags);
2303 		}
2304 
2305 		l2pte_set(ptep, npte, opte);
2306 		PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
2307 	}
2308 #endif /* !ARM_MMU_EXTENDED */
2309 }
2310 #endif	/* PMAP_CACHE_VIPT */
2311 
2312 
2313 /*
2314  * Modify pte bits for all ptes corresponding to the given physical address.
2315  * We use `maskbits' rather than `clearbits' because we're always passing
2316  * constants and the latter would require an extra inversion at run-time.
2317  */
2318 static void
2319 pmap_clearbit(struct vm_page_md *md, paddr_t pa, u_int maskbits)
2320 {
2321 	struct pv_entry *pv;
2322 #ifdef PMAP_CACHE_VIPT
2323 	const bool want_syncicache = PV_IS_EXEC_P(md->pvh_attrs);
2324 #ifdef ARM_MMU_EXTENDED
2325 	const u_int execbits = (maskbits & PVF_EXEC) ? L2_XS_XN : 0;
2326 #else
2327 	const u_int execbits = 0;
2328 	bool need_vac_me_harder = false;
2329 	bool need_syncicache = false;
2330 #endif
2331 #else
2332 	const u_int execbits = 0;
2333 #endif
2334 
2335 	NPDEBUG(PDB_BITS,
2336 	    printf("pmap_clearbit: md %p mask 0x%x\n",
2337 	    md, maskbits));
2338 
2339 #ifdef PMAP_CACHE_VIPT
2340 	/*
2341 	 * If we might want to sync the I-cache and we've modified it,
2342 	 * then we know we definitely need to sync or discard it.
2343 	 */
2344 	if (want_syncicache) {
2345 #ifdef ARM_MMU_EXTENDED
2346 		if (md->pvh_attrs & PVF_MOD)
2347 			md->pvh_attrs &= ~PVF_EXEC;
2348 #else
2349 		need_syncicache = md->pvh_attrs & PVF_MOD;
2350 #endif
2351 	}
2352 #endif
2353 	KASSERT(pmap_page_locked_p(md));
2354 
2355 	/*
2356 	 * Clear saved attributes (modify, reference)
2357 	 */
2358 	md->pvh_attrs &= ~(maskbits & (PVF_MOD | PVF_REF));
2359 
2360 	if (SLIST_EMPTY(&md->pvh_list)) {
2361 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
2362 		if (need_syncicache) {
2363 			/*
2364 			 * No one has it mapped, so just discard it.  The next
2365 			 * exec remapping will cause it to be synced.
2366 			 */
2367 			md->pvh_attrs &= ~PVF_EXEC;
2368 			PMAPCOUNT(exec_discarded_clearbit);
2369 		}
2370 #endif
2371 		return;
2372 	}
2373 
2374 	/*
2375 	 * Loop over all current mappings setting/clearing as appropos
2376 	 */
2377 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
2378 		pmap_t pm = pv->pv_pmap;
2379 		const vaddr_t va = pv->pv_va;
2380 		const u_int oflags = pv->pv_flags;
2381 #ifndef ARM_MMU_EXTENDED
2382 		/*
2383 		 * Kernel entries are unmanaged and as such not to be changed.
2384 		 */
2385 		if (PV_IS_KENTRY_P(oflags))
2386 			continue;
2387 #endif
2388 		pv->pv_flags &= ~maskbits;
2389 
2390 		pmap_release_page_lock(md);
2391 		pmap_acquire_pmap_lock(pm);
2392 
2393 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pm, va);
2394 		if (l2b == NULL) {
2395 			pmap_release_pmap_lock(pm);
2396 			pmap_acquire_page_lock(md);
2397 			continue;
2398 		}
2399 		KASSERTMSG(l2b != NULL, "%#lx", va);
2400 
2401 		pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
2402 		const pt_entry_t opte = *ptep;
2403 		pt_entry_t npte = opte | execbits;
2404 
2405 		KASSERT((opte & L2_XS_nG) == (pm == pmap_kernel() ? 0 : L2_XS_nG));
2406 
2407 		NPDEBUG(PDB_BITS,
2408 		    printf( "%s: pv %p, pm %p, va 0x%08lx, flag 0x%x\n",
2409 			__func__, pv, pm, va, oflags));
2410 
2411 		if (maskbits & (PVF_WRITE|PVF_MOD)) {
2412 #ifdef PMAP_CACHE_VIVT
2413 			if ((oflags & PVF_NC)) {
2414 				/*
2415 				 * Entry is not cacheable:
2416 				 *
2417 				 * Don't turn caching on again if this is a
2418 				 * modified emulation. This would be
2419 				 * inconsitent with the settings created by
2420 				 * pmap_vac_me_harder(). Otherwise, it's safe
2421 				 * to re-enable cacheing.
2422 				 *
2423 				 * There's no need to call pmap_vac_me_harder()
2424 				 * here: all pages are losing their write
2425 				 * permission.
2426 				 */
2427 				if (maskbits & PVF_WRITE) {
2428 					npte |= pte_l2_s_cache_mode;
2429 					pv->pv_flags &= ~PVF_NC;
2430 				}
2431 			} else
2432 			if (l2pte_writable_p(opte)) {
2433 				/*
2434 				 * Entry is writable/cacheable: check if pmap
2435 				 * is current if it is flush it, otherwise it
2436 				 * won't be in the cache
2437 				 */
2438 				pmap_cache_wbinv_page(pm, va,
2439 				    (maskbits & PVF_REF) != 0,
2440 				    oflags|PVF_WRITE);
2441 			}
2442 #endif
2443 
2444 			/* make the pte read only */
2445 			npte = l2pte_set_readonly(npte);
2446 
2447 			pmap_acquire_page_lock(md);
2448 #ifdef MULTIPROCESSOR
2449 			pv = pmap_find_pv(md, pm, va);
2450 #endif
2451 			if (pv != NULL && (maskbits & oflags & PVF_WRITE)) {
2452 				/*
2453 				 * Keep alias accounting up to date
2454 				 */
2455 				if (pm == pmap_kernel()) {
2456 					md->krw_mappings--;
2457 					md->kro_mappings++;
2458 				} else {
2459 					md->urw_mappings--;
2460 					md->uro_mappings++;
2461 				}
2462 #ifdef PMAP_CACHE_VIPT
2463 				if (arm_cache_prefer_mask != 0) {
2464 					if (md->urw_mappings + md->krw_mappings == 0) {
2465 						md->pvh_attrs &= ~PVF_WRITE;
2466 					} else {
2467 						PMAP_VALIDATE_MD_PAGE(md);
2468 					}
2469 				}
2470 #ifndef ARM_MMU_EXTENDED
2471 				if (want_syncicache)
2472 					need_syncicache = true;
2473 				need_vac_me_harder = true;
2474 #endif
2475 #endif /* PMAP_CACHE_VIPT */
2476 			}
2477 			pmap_release_page_lock(md);
2478 		}
2479 
2480 		if (maskbits & PVF_REF) {
2481 			if (true
2482 #ifndef ARM_MMU_EXTENDED
2483 			    && (oflags & PVF_NC) == 0
2484 #endif
2485 			    && (maskbits & (PVF_WRITE|PVF_MOD)) == 0
2486 			    && l2pte_valid_p(npte)) {
2487 #ifdef PMAP_CACHE_VIVT
2488 				/*
2489 				 * Check npte here; we may have already
2490 				 * done the wbinv above, and the validity
2491 				 * of the PTE is the same for opte and
2492 				 * npte.
2493 				 */
2494 				pmap_cache_wbinv_page(pm, va, true, oflags);
2495 #endif
2496 			}
2497 
2498 			/*
2499 			 * Make the PTE invalid so that we will take a
2500 			 * page fault the next time the mapping is
2501 			 * referenced.
2502 			 */
2503 			npte &= ~L2_TYPE_MASK;
2504 			npte |= L2_TYPE_INV;
2505 		}
2506 
2507 		if (npte != opte) {
2508 			l2pte_set(ptep, npte, opte);
2509 			PTE_SYNC(ptep);
2510 
2511 			/* Flush the TLB entry if a current pmap. */
2512 			pmap_tlb_flush_SE(pm, va, oflags);
2513 		}
2514 
2515 		pmap_release_pmap_lock(pm);
2516 		pmap_acquire_page_lock(md);
2517 
2518 		NPDEBUG(PDB_BITS,
2519 		    printf("pmap_clearbit: pm %p va 0x%lx opte 0x%08x npte 0x%08x\n",
2520 		    pm, va, opte, npte));
2521 	}
2522 
2523 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
2524 	/*
2525 	 * If we need to sync the I-cache and we haven't done it yet, do it.
2526 	 */
2527 	if (need_syncicache) {
2528 		pmap_release_page_lock(md);
2529 		pmap_syncicache_page(md, pa);
2530 		pmap_acquire_page_lock(md);
2531 		PMAPCOUNT(exec_synced_clearbit);
2532 	}
2533 
2534 	/*
2535 	 * If we are changing this to read-only, we need to call vac_me_harder
2536 	 * so we can change all the read-only pages to cacheable.  We pretend
2537 	 * this as a page deletion.
2538 	 */
2539 	if (need_vac_me_harder) {
2540 		if (md->pvh_attrs & PVF_NC)
2541 			pmap_vac_me_harder(md, pa, NULL, 0);
2542 	}
2543 #endif /* PMAP_CACHE_VIPT && !ARM_MMU_EXTENDED */
2544 }
2545 
2546 /*
2547  * pmap_clean_page()
2548  *
2549  * This is a local function used to work out the best strategy to clean
2550  * a single page referenced by its entry in the PV table. It's used by
2551  * pmap_copy_page, pmap_zero page and maybe some others later on.
2552  *
2553  * Its policy is effectively:
2554  *  o If there are no mappings, we don't bother doing anything with the cache.
2555  *  o If there is one mapping, we clean just that page.
2556  *  o If there are multiple mappings, we clean the entire cache.
2557  *
2558  * So that some functions can be further optimised, it returns 0 if it didn't
2559  * clean the entire cache, or 1 if it did.
2560  *
2561  * XXX One bug in this routine is that if the pv_entry has a single page
2562  * mapped at 0x00000000 a whole cache clean will be performed rather than
2563  * just the 1 page. Since this should not occur in everyday use and if it does
2564  * it will just result in not the most efficient clean for the page.
2565  */
2566 #ifdef PMAP_CACHE_VIVT
2567 static bool
2568 pmap_clean_page(struct vm_page_md *md, bool is_src)
2569 {
2570 	struct pv_entry *pv;
2571 	pmap_t pm_to_clean = NULL;
2572 	bool cache_needs_cleaning = false;
2573 	vaddr_t page_to_clean = 0;
2574 	u_int flags = 0;
2575 
2576 	/*
2577 	 * Since we flush the cache each time we change to a different
2578 	 * user vmspace, we only need to flush the page if it is in the
2579 	 * current pmap.
2580 	 */
2581 	KASSERT(pmap_page_locked_p(md));
2582 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
2583 		if (pmap_is_current(pv->pv_pmap)) {
2584 			flags |= pv->pv_flags;
2585 			/*
2586 			 * The page is mapped non-cacheable in
2587 			 * this map.  No need to flush the cache.
2588 			 */
2589 			if (pv->pv_flags & PVF_NC) {
2590 #ifdef DIAGNOSTIC
2591 				KASSERT(!cache_needs_cleaning);
2592 #endif
2593 				break;
2594 			} else if (is_src && (pv->pv_flags & PVF_WRITE) == 0)
2595 				continue;
2596 			if (cache_needs_cleaning) {
2597 				page_to_clean = 0;
2598 				break;
2599 			} else {
2600 				page_to_clean = pv->pv_va;
2601 				pm_to_clean = pv->pv_pmap;
2602 			}
2603 			cache_needs_cleaning = true;
2604 		}
2605 	}
2606 
2607 	if (page_to_clean) {
2608 		pmap_cache_wbinv_page(pm_to_clean, page_to_clean,
2609 		    !is_src, flags | PVF_REF);
2610 	} else if (cache_needs_cleaning) {
2611 		pmap_t const pm = curproc->p_vmspace->vm_map.pmap;
2612 
2613 		pmap_cache_wbinv_all(pm, flags);
2614 		return true;
2615 	}
2616 	return false;
2617 }
2618 #endif
2619 
2620 #ifdef PMAP_CACHE_VIPT
2621 /*
2622  * Sync a page with the I-cache.  Since this is a VIPT, we must pick the
2623  * right cache alias to make sure we flush the right stuff.
2624  */
2625 void
2626 pmap_syncicache_page(struct vm_page_md *md, paddr_t pa)
2627 {
2628 	pmap_t kpm = pmap_kernel();
2629 	const size_t way_size = arm_pcache.icache_type == CACHE_TYPE_PIPT
2630 	    ? PAGE_SIZE
2631 	    : arm_pcache.icache_way_size;
2632 
2633 	NPDEBUG(PDB_EXEC, printf("pmap_syncicache_page: md=%p (attrs=%#x)\n",
2634 	    md, md->pvh_attrs));
2635 	/*
2636 	 * No need to clean the page if it's non-cached.
2637 	 */
2638 #ifndef ARM_MMU_EXTENDED
2639 	if (md->pvh_attrs & PVF_NC)
2640 		return;
2641 	KASSERT(arm_cache_prefer_mask == 0 || md->pvh_attrs & PVF_COLORED);
2642 #endif
2643 
2644 	pt_entry_t * const ptep = cpu_cdst_pte(0);
2645 	const vaddr_t dstp = cpu_cdstp(0);
2646 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
2647 	if (way_size <= PAGE_SIZE) {
2648 		bool ok = false;
2649 		vaddr_t vdstp = pmap_direct_mapped_phys(pa, &ok, dstp);
2650 		if (ok) {
2651 			cpu_icache_sync_range(vdstp, way_size);
2652 			return;
2653 		}
2654 	}
2655 #endif
2656 
2657 	/*
2658 	 * We don't worry about the color of the exec page, we map the
2659 	 * same page to pages in the way and then do the icache_sync on
2660 	 * the entire way making sure we are cleaned.
2661 	 */
2662 	const pt_entry_t npte = L2_S_PROTO | pa | pte_l2_s_cache_mode
2663 	    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE);
2664 
2665 	for (size_t i = 0, j = 0; i < way_size;
2666 	     i += PAGE_SIZE, j += PAGE_SIZE / L2_S_SIZE) {
2667 		pmap_tlb_flush_SE(kpm, dstp + i, PVF_REF | PVF_EXEC);
2668 		/*
2669 		 * Set up a PTE with to flush these cache lines.
2670 		 */
2671 		l2pte_set(ptep + j, npte, 0);
2672 	}
2673 	PTE_SYNC_RANGE(ptep, way_size / L2_S_SIZE);
2674 
2675 	/*
2676 	 * Flush it.
2677 	 */
2678 	cpu_icache_sync_range(dstp, way_size);
2679 
2680 	for (size_t i = 0, j = 0; i < way_size;
2681 	     i += PAGE_SIZE, j += PAGE_SIZE / L2_S_SIZE) {
2682 		/*
2683 		 * Unmap the page(s).
2684 		 */
2685 		l2pte_reset(ptep + j);
2686 		pmap_tlb_flush_SE(kpm, dstp + i, PVF_REF | PVF_EXEC);
2687 	}
2688 	PTE_SYNC_RANGE(ptep, way_size / L2_S_SIZE);
2689 
2690 	md->pvh_attrs |= PVF_EXEC;
2691 	PMAPCOUNT(exec_synced);
2692 }
2693 
2694 #ifndef ARM_MMU_EXTENDED
2695 void
2696 pmap_flush_page(struct vm_page_md *md, paddr_t pa, enum pmap_flush_op flush)
2697 {
2698 	vsize_t va_offset, end_va;
2699 	bool wbinv_p;
2700 
2701 	if (arm_cache_prefer_mask == 0)
2702 		return;
2703 
2704 	switch (flush) {
2705 	case PMAP_FLUSH_PRIMARY:
2706 		if (md->pvh_attrs & PVF_MULTCLR) {
2707 			va_offset = 0;
2708 			end_va = arm_cache_prefer_mask;
2709 			md->pvh_attrs &= ~PVF_MULTCLR;
2710 			PMAPCOUNT(vac_flush_lots);
2711 		} else {
2712 			va_offset = md->pvh_attrs & arm_cache_prefer_mask;
2713 			end_va = va_offset;
2714 			PMAPCOUNT(vac_flush_one);
2715 		}
2716 		/*
2717 		 * Mark that the page is no longer dirty.
2718 		 */
2719 		md->pvh_attrs &= ~PVF_DIRTY;
2720 		wbinv_p = true;
2721 		break;
2722 	case PMAP_FLUSH_SECONDARY:
2723 		va_offset = 0;
2724 		end_va = arm_cache_prefer_mask;
2725 		wbinv_p = true;
2726 		md->pvh_attrs &= ~PVF_MULTCLR;
2727 		PMAPCOUNT(vac_flush_lots);
2728 		break;
2729 	case PMAP_CLEAN_PRIMARY:
2730 		va_offset = md->pvh_attrs & arm_cache_prefer_mask;
2731 		end_va = va_offset;
2732 		wbinv_p = false;
2733 		/*
2734 		 * Mark that the page is no longer dirty.
2735 		 */
2736 		if ((md->pvh_attrs & PVF_DMOD) == 0)
2737 			md->pvh_attrs &= ~PVF_DIRTY;
2738 		PMAPCOUNT(vac_clean_one);
2739 		break;
2740 	default:
2741 		return;
2742 	}
2743 
2744 	KASSERT(!(md->pvh_attrs & PVF_NC));
2745 
2746 	NPDEBUG(PDB_VAC, printf("pmap_flush_page: md=%p (attrs=%#x)\n",
2747 	    md, md->pvh_attrs));
2748 
2749 	const size_t scache_line_size = arm_scache.dcache_line_size;
2750 
2751 	for (; va_offset <= end_va; va_offset += PAGE_SIZE) {
2752 		pt_entry_t * const ptep = cpu_cdst_pte(va_offset);
2753 		const vaddr_t dstp = cpu_cdstp(va_offset);
2754 		const pt_entry_t opte = *ptep;
2755 
2756 		if (flush == PMAP_FLUSH_SECONDARY
2757 		    && va_offset == (md->pvh_attrs & arm_cache_prefer_mask))
2758 			continue;
2759 
2760 		pmap_tlb_flush_SE(pmap_kernel(), dstp, PVF_REF | PVF_EXEC);
2761 		/*
2762 		 * Set up a PTE with the right coloring to flush
2763 		 * existing cache entries.
2764 		 */
2765 		const pt_entry_t npte = L2_S_PROTO
2766 		    | pa
2767 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE)
2768 		    | pte_l2_s_cache_mode;
2769 		l2pte_set(ptep, npte, opte);
2770 		PTE_SYNC(ptep);
2771 
2772 		/*
2773 		 * Flush it.  Make sure to flush secondary cache too since
2774 		 * bus_dma will ignore uncached pages.
2775 		 */
2776 		if (scache_line_size != 0) {
2777 			cpu_dcache_wb_range(dstp, PAGE_SIZE);
2778 			if (wbinv_p) {
2779 				cpu_sdcache_wbinv_range(dstp, pa, PAGE_SIZE);
2780 				cpu_dcache_inv_range(dstp, PAGE_SIZE);
2781 			} else {
2782 				cpu_sdcache_wb_range(dstp, pa, PAGE_SIZE);
2783 			}
2784 		} else {
2785 			if (wbinv_p) {
2786 				cpu_dcache_wbinv_range(dstp, PAGE_SIZE);
2787 			} else {
2788 				cpu_dcache_wb_range(dstp, PAGE_SIZE);
2789 			}
2790 		}
2791 
2792 		/*
2793 		 * Restore the page table entry since we might have interrupted
2794 		 * pmap_zero_page or pmap_copy_page which was already using
2795 		 * this pte.
2796 		 */
2797 		if (opte) {
2798 			l2pte_set(ptep, opte, npte);
2799 		} else {
2800 			l2pte_reset(ptep);
2801 		}
2802 		PTE_SYNC(ptep);
2803 		pmap_tlb_flush_SE(pmap_kernel(), dstp, PVF_REF | PVF_EXEC);
2804 	}
2805 }
2806 #endif /* ARM_MMU_EXTENDED */
2807 #endif /* PMAP_CACHE_VIPT */
2808 
2809 /*
2810  * Routine:	pmap_page_remove
2811  * Function:
2812  *		Removes this physical page from
2813  *		all physical maps in which it resides.
2814  *		Reflects back modify bits to the pager.
2815  */
2816 static void
2817 pmap_page_remove(struct vm_page_md *md, paddr_t pa)
2818 {
2819 	struct l2_bucket *l2b;
2820 	struct pv_entry *pv;
2821 	pt_entry_t *ptep;
2822 #ifndef ARM_MMU_EXTENDED
2823 	bool flush = false;
2824 #endif
2825 	u_int flags = 0;
2826 
2827 	NPDEBUG(PDB_FOLLOW,
2828 	    printf("pmap_page_remove: md %p (0x%08lx)\n", md,
2829 	    pa));
2830 
2831 	struct pv_entry **pvp = &SLIST_FIRST(&md->pvh_list);
2832 	pmap_acquire_page_lock(md);
2833 	if (*pvp == NULL) {
2834 #ifdef PMAP_CACHE_VIPT
2835 		/*
2836 		 * We *know* the page contents are about to be replaced.
2837 		 * Discard the exec contents
2838 		 */
2839 		if (PV_IS_EXEC_P(md->pvh_attrs))
2840 			PMAPCOUNT(exec_discarded_page_protect);
2841 		md->pvh_attrs &= ~PVF_EXEC;
2842 		PMAP_VALIDATE_MD_PAGE(md);
2843 #endif
2844 		pmap_release_page_lock(md);
2845 		return;
2846 	}
2847 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
2848 	KASSERT(arm_cache_prefer_mask == 0 || pmap_is_page_colored_p(md));
2849 #endif
2850 
2851 	/*
2852 	 * Clear alias counts
2853 	 */
2854 #ifdef PMAP_CACHE_VIVT
2855 	md->k_mappings = 0;
2856 #endif
2857 	md->urw_mappings = md->uro_mappings = 0;
2858 
2859 #ifdef PMAP_CACHE_VIVT
2860 	pmap_clean_page(md, false);
2861 #endif
2862 
2863 	while ((pv = *pvp) != NULL) {
2864 		pmap_t pm = pv->pv_pmap;
2865 #ifndef ARM_MMU_EXTENDED
2866 		if (flush == false && pmap_is_current(pm))
2867 			flush = true;
2868 #endif
2869 
2870 		if (pm == pmap_kernel()) {
2871 #ifdef PMAP_CACHE_VIPT
2872 			/*
2873 			 * If this was unmanaged mapping, it must be preserved.
2874 			 * Move it back on the list and advance the end-of-list
2875 			 * pointer.
2876 			 */
2877 			if (PV_IS_KENTRY_P(pv->pv_flags)) {
2878 				*pvp = pv;
2879 				pvp = &SLIST_NEXT(pv, pv_link);
2880 				continue;
2881 			}
2882 			if (pv->pv_flags & PVF_WRITE)
2883 				md->krw_mappings--;
2884 			else
2885 				md->kro_mappings--;
2886 #endif
2887 			PMAPCOUNT(kernel_unmappings);
2888 		}
2889 		*pvp = SLIST_NEXT(pv, pv_link); /* remove from list */
2890 		PMAPCOUNT(unmappings);
2891 
2892 		pmap_release_page_lock(md);
2893 		pmap_acquire_pmap_lock(pm);
2894 
2895 #ifdef ARM_MMU_EXTENDED
2896 		pmap_tlb_invalidate_addr(pm, pv->pv_va);
2897 #endif
2898 
2899 		l2b = pmap_get_l2_bucket(pm, pv->pv_va);
2900 		KASSERTMSG(l2b != NULL, "%#lx", pv->pv_va);
2901 
2902 		ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2903 
2904 		/*
2905 		 * Update statistics
2906 		 */
2907 		--pm->pm_stats.resident_count;
2908 
2909 		/* Wired bit */
2910 		if (pv->pv_flags & PVF_WIRED)
2911 			--pm->pm_stats.wired_count;
2912 
2913 		flags |= pv->pv_flags;
2914 
2915 		/*
2916 		 * Invalidate the PTEs.
2917 		 */
2918 		l2pte_reset(ptep);
2919 		PTE_SYNC_CURRENT(pm, ptep);
2920 		pmap_free_l2_bucket(pm, l2b, PAGE_SIZE / L2_S_SIZE);
2921 		pmap_release_pmap_lock(pm);
2922 
2923 		pool_put(&pmap_pv_pool, pv);
2924 		pmap_acquire_page_lock(md);
2925 #ifdef MULTIPROCESSOR
2926 		/*
2927 		 * Restart of the beginning of the list.
2928 		 */
2929 		pvp = &SLIST_FIRST(&md->pvh_list);
2930 #endif
2931 	}
2932 	/*
2933 	 * if we reach the end of the list and there are still mappings, they
2934 	 * might be able to be cached now.  And they must be kernel mappings.
2935 	 */
2936 	if (!SLIST_EMPTY(&md->pvh_list)) {
2937 		pmap_vac_me_harder(md, pa, pmap_kernel(), 0);
2938 	}
2939 
2940 #ifdef PMAP_CACHE_VIPT
2941 	/*
2942 	 * Its EXEC cache is now gone.
2943 	 */
2944 	if (PV_IS_EXEC_P(md->pvh_attrs))
2945 		PMAPCOUNT(exec_discarded_page_protect);
2946 	md->pvh_attrs &= ~PVF_EXEC;
2947 	KASSERT(md->urw_mappings == 0);
2948 	KASSERT(md->uro_mappings == 0);
2949 #ifndef ARM_MMU_EXTENDED
2950 	if (arm_cache_prefer_mask != 0) {
2951 		if (md->krw_mappings == 0)
2952 			md->pvh_attrs &= ~PVF_WRITE;
2953 		PMAP_VALIDATE_MD_PAGE(md);
2954 	}
2955 #endif /* ARM_MMU_EXTENDED */
2956 #endif /* PMAP_CACHE_VIPT */
2957 	pmap_release_page_lock(md);
2958 
2959 #ifndef ARM_MMU_EXTENDED
2960 	if (flush) {
2961 		/*
2962 		 * Note: We can't use pmap_tlb_flush{I,D}() here since that
2963 		 * would need a subsequent call to pmap_update() to ensure
2964 		 * curpm->pm_cstate.cs_all is reset. Our callers are not
2965 		 * required to do that (see pmap(9)), so we can't modify
2966 		 * the current pmap's state.
2967 		 */
2968 		if (PV_BEEN_EXECD(flags))
2969 			cpu_tlb_flushID();
2970 		else
2971 			cpu_tlb_flushD();
2972 	}
2973 	cpu_cpwait();
2974 #endif /* ARM_MMU_EXTENDED */
2975 }
2976 
2977 /*
2978  * pmap_t pmap_create(void)
2979  *
2980  *      Create a new pmap structure from scratch.
2981  */
2982 pmap_t
2983 pmap_create(void)
2984 {
2985 	pmap_t pm;
2986 
2987 	pm = pool_cache_get(&pmap_cache, PR_WAITOK);
2988 
2989 	mutex_init(&pm->pm_obj_lock, MUTEX_DEFAULT, IPL_NONE);
2990 	uvm_obj_init(&pm->pm_obj, NULL, false, 1);
2991 	uvm_obj_setlock(&pm->pm_obj, &pm->pm_obj_lock);
2992 
2993 	pm->pm_stats.wired_count = 0;
2994 	pm->pm_stats.resident_count = 1;
2995 #ifdef ARM_MMU_EXTENDED
2996 #ifdef MULTIPROCESSOR
2997 	kcpuset_create(&pm->pm_active, true);
2998 	kcpuset_create(&pm->pm_onproc, true);
2999 #endif
3000 #else
3001 	pm->pm_cstate.cs_all = 0;
3002 #endif
3003 	pmap_alloc_l1(pm);
3004 
3005 	/*
3006 	 * Note: The pool cache ensures that the pm_l2[] array is already
3007 	 * initialised to zero.
3008 	 */
3009 
3010 	pmap_pinit(pm);
3011 
3012 	LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list);
3013 
3014 	return (pm);
3015 }
3016 
3017 u_int
3018 arm32_mmap_flags(paddr_t pa)
3019 {
3020 	/*
3021 	 * the upper 8 bits in pmap_enter()'s flags are reserved for MD stuff
3022 	 * and we're using the upper bits in page numbers to pass flags around
3023 	 * so we might as well use the same bits
3024 	 */
3025 	return (u_int)pa & PMAP_MD_MASK;
3026 }
3027 /*
3028  * int pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
3029  *      u_int flags)
3030  *
3031  *      Insert the given physical page (p) at
3032  *      the specified virtual address (v) in the
3033  *      target physical map with the protection requested.
3034  *
3035  *      NB:  This is the only routine which MAY NOT lazy-evaluate
3036  *      or lose information.  That is, this routine must actually
3037  *      insert this page into the given map NOW.
3038  */
3039 int
3040 pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags)
3041 {
3042 	struct l2_bucket *l2b;
3043 	struct vm_page *pg, *opg;
3044 	u_int nflags;
3045 	u_int oflags;
3046 	const bool kpm_p = (pm == pmap_kernel());
3047 #ifdef ARM_HAS_VBAR
3048 	const bool vector_page_p = false;
3049 #else
3050 	const bool vector_page_p = (va == vector_page);
3051 #endif
3052 
3053 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
3054 
3055 	UVMHIST_LOG(maphist, " (pm %p va %#x pa %#x prot %#x",
3056 	    pm, va, pa, prot);
3057 	UVMHIST_LOG(maphist, "  flag %#x", flags, 0, 0, 0);
3058 
3059 	KDASSERT((flags & PMAP_WIRED) == 0 || (flags & VM_PROT_ALL) != 0);
3060 	KDASSERT(((va | pa) & PGOFSET) == 0);
3061 
3062 	/*
3063 	 * Get a pointer to the page.  Later on in this function, we
3064 	 * test for a managed page by checking pg != NULL.
3065 	 */
3066 	pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
3067 
3068 	nflags = 0;
3069 	if (prot & VM_PROT_WRITE)
3070 		nflags |= PVF_WRITE;
3071 	if (prot & VM_PROT_EXECUTE)
3072 		nflags |= PVF_EXEC;
3073 	if (flags & PMAP_WIRED)
3074 		nflags |= PVF_WIRED;
3075 
3076 	pmap_acquire_pmap_lock(pm);
3077 
3078 	/*
3079 	 * Fetch the L2 bucket which maps this page, allocating one if
3080 	 * necessary for user pmaps.
3081 	 */
3082 	if (kpm_p) {
3083 		l2b = pmap_get_l2_bucket(pm, va);
3084 	} else {
3085 		l2b = pmap_alloc_l2_bucket(pm, va);
3086 	}
3087 	if (l2b == NULL) {
3088 		if (flags & PMAP_CANFAIL) {
3089 			pmap_release_pmap_lock(pm);
3090 			return (ENOMEM);
3091 		}
3092 		panic("pmap_enter: failed to allocate L2 bucket");
3093 	}
3094 	pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(va)];
3095 	const pt_entry_t opte = *ptep;
3096 	pt_entry_t npte = pa;
3097 	oflags = 0;
3098 
3099 	if (opte) {
3100 		/*
3101 		 * There is already a mapping at this address.
3102 		 * If the physical address is different, lookup the
3103 		 * vm_page.
3104 		 */
3105 		if (l2pte_pa(opte) != pa)
3106 			opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3107 		else
3108 			opg = pg;
3109 	} else
3110 		opg = NULL;
3111 
3112 	if (pg) {
3113 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3114 
3115 		/*
3116 		 * This is to be a managed mapping.
3117 		 */
3118 		pmap_acquire_page_lock(md);
3119 		if ((flags & VM_PROT_ALL) || (md->pvh_attrs & PVF_REF)) {
3120 			/*
3121 			 * - The access type indicates that we don't need
3122 			 *   to do referenced emulation.
3123 			 * OR
3124 			 * - The physical page has already been referenced
3125 			 *   so no need to re-do referenced emulation here.
3126 			 */
3127 			npte |= l2pte_set_readonly(L2_S_PROTO);
3128 
3129 			nflags |= PVF_REF;
3130 
3131 			if ((prot & VM_PROT_WRITE) != 0 &&
3132 			    ((flags & VM_PROT_WRITE) != 0 ||
3133 			     (md->pvh_attrs & PVF_MOD) != 0)) {
3134 				/*
3135 				 * This is a writable mapping, and the
3136 				 * page's mod state indicates it has
3137 				 * already been modified. Make it
3138 				 * writable from the outset.
3139 				 */
3140 				npte = l2pte_set_writable(npte);
3141 				nflags |= PVF_MOD;
3142 			}
3143 
3144 #ifdef ARM_MMU_EXTENDED
3145 			/*
3146 			 * If the page has been cleaned, then the pvh_attrs
3147 			 * will have PVF_EXEC set, so mark it execute so we
3148 			 * don't get an access fault when trying to execute
3149 			 * from it.
3150 			 */
3151 			if (md->pvh_attrs & nflags & PVF_EXEC) {
3152 				npte &= ~L2_XS_XN;
3153 			}
3154 #endif
3155 		} else {
3156 			/*
3157 			 * Need to do page referenced emulation.
3158 			 */
3159 			npte |= L2_TYPE_INV;
3160 		}
3161 
3162 		if (flags & ARM32_MMAP_WRITECOMBINE) {
3163 			npte |= pte_l2_s_wc_mode;
3164 		} else
3165 			npte |= pte_l2_s_cache_mode;
3166 
3167 		if (pg == opg) {
3168 			/*
3169 			 * We're changing the attrs of an existing mapping.
3170 			 */
3171 			oflags = pmap_modify_pv(md, pa, pm, va,
3172 			    PVF_WRITE | PVF_EXEC | PVF_WIRED |
3173 			    PVF_MOD | PVF_REF, nflags);
3174 
3175 #ifdef PMAP_CACHE_VIVT
3176 			/*
3177 			 * We may need to flush the cache if we're
3178 			 * doing rw-ro...
3179 			 */
3180 			if (pm->pm_cstate.cs_cache_d &&
3181 			    (oflags & PVF_NC) == 0 &&
3182 			    l2pte_writable_p(opte) &&
3183 			    (prot & VM_PROT_WRITE) == 0)
3184 				cpu_dcache_wb_range(va, PAGE_SIZE);
3185 #endif
3186 		} else {
3187 			struct pv_entry *pv;
3188 			/*
3189 			 * New mapping, or changing the backing page
3190 			 * of an existing mapping.
3191 			 */
3192 			if (opg) {
3193 				struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3194 				paddr_t opa = VM_PAGE_TO_PHYS(opg);
3195 
3196 				/*
3197 				 * Replacing an existing mapping with a new one.
3198 				 * It is part of our managed memory so we
3199 				 * must remove it from the PV list
3200 				 */
3201 				pv = pmap_remove_pv(omd, opa, pm, va);
3202 				pmap_vac_me_harder(omd, opa, pm, 0);
3203 				oflags = pv->pv_flags;
3204 
3205 #ifdef PMAP_CACHE_VIVT
3206 				/*
3207 				 * If the old mapping was valid (ref/mod
3208 				 * emulation creates 'invalid' mappings
3209 				 * initially) then make sure to frob
3210 				 * the cache.
3211 				 */
3212 				if (!(oflags & PVF_NC) && l2pte_valid_p(opte)) {
3213 					pmap_cache_wbinv_page(pm, va, true,
3214 					    oflags);
3215 				}
3216 #endif
3217 			} else {
3218 				pmap_release_page_lock(md);
3219 				pv = pool_get(&pmap_pv_pool, PR_NOWAIT);
3220 				if (pv == NULL) {
3221 					pmap_release_pmap_lock(pm);
3222 					if ((flags & PMAP_CANFAIL) == 0)
3223 						panic("pmap_enter: "
3224 						    "no pv entries");
3225 
3226 					pmap_free_l2_bucket(pm, l2b, 0);
3227 					UVMHIST_LOG(maphist, "  <-- done (ENOMEM)",
3228 					    0, 0, 0, 0);
3229 					return (ENOMEM);
3230 				}
3231 				pmap_acquire_page_lock(md);
3232 			}
3233 
3234 			pmap_enter_pv(md, pa, pv, pm, va, nflags);
3235 		}
3236 		pmap_release_page_lock(md);
3237 	} else {
3238 		/*
3239 		 * We're mapping an unmanaged page.
3240 		 * These are always readable, and possibly writable, from
3241 		 * the get go as we don't need to track ref/mod status.
3242 		 */
3243 		npte |= l2pte_set_readonly(L2_S_PROTO);
3244 		if (prot & VM_PROT_WRITE)
3245 			npte = l2pte_set_writable(npte);
3246 
3247 		/*
3248 		 * Make sure the vector table is mapped cacheable
3249 		 */
3250 		if ((vector_page_p && !kpm_p)
3251 		    || (flags & ARM32_MMAP_CACHEABLE)) {
3252 			npte |= pte_l2_s_cache_mode;
3253 #ifdef ARM_MMU_EXTENDED
3254 			npte &= ~L2_XS_XN;	/* and executable */
3255 #endif
3256 		} else if (flags & ARM32_MMAP_WRITECOMBINE) {
3257 			npte |= pte_l2_s_wc_mode;
3258 		}
3259 		if (opg) {
3260 			/*
3261 			 * Looks like there's an existing 'managed' mapping
3262 			 * at this address.
3263 			 */
3264 			struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3265 			paddr_t opa = VM_PAGE_TO_PHYS(opg);
3266 
3267 			pmap_acquire_page_lock(omd);
3268 			struct pv_entry *pv = pmap_remove_pv(omd, opa, pm, va);
3269 			pmap_vac_me_harder(omd, opa, pm, 0);
3270 			oflags = pv->pv_flags;
3271 			pmap_release_page_lock(omd);
3272 
3273 #ifdef PMAP_CACHE_VIVT
3274 			if (!(oflags & PVF_NC) && l2pte_valid_p(opte)) {
3275 				pmap_cache_wbinv_page(pm, va, true, oflags);
3276 			}
3277 #endif
3278 			pool_put(&pmap_pv_pool, pv);
3279 		}
3280 	}
3281 
3282 	/*
3283 	 * Make sure userland mappings get the right permissions
3284 	 */
3285 	if (!vector_page_p && !kpm_p) {
3286 		npte |= L2_S_PROT_U;
3287 #ifdef ARM_MMU_EXTENDED
3288 		npte |= L2_XS_nG;	/* user pages are not global */
3289 #endif
3290 	}
3291 
3292 	/*
3293 	 * Keep the stats up to date
3294 	 */
3295 	if (opte == 0) {
3296 		l2b->l2b_occupancy += PAGE_SIZE / L2_S_SIZE;
3297 		pm->pm_stats.resident_count++;
3298 	}
3299 
3300 	UVMHIST_LOG(maphist, " opte %#x npte %#x", opte, npte, 0, 0);
3301 
3302 #if defined(ARM_MMU_EXTENDED)
3303 	/*
3304 	 * If exec protection was requested but the page hasn't been synced,
3305 	 * sync it now and allow execution from it.
3306 	 */
3307 	if ((nflags & PVF_EXEC) && (npte & L2_XS_XN)) {
3308 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3309 		npte &= ~L2_XS_XN;
3310 		pmap_syncicache_page(md, pa);
3311 		PMAPCOUNT(exec_synced_map);
3312 	}
3313 #endif
3314 	/*
3315 	 * If this is just a wiring change, the two PTEs will be
3316 	 * identical, so there's no need to update the page table.
3317 	 */
3318 	if (npte != opte) {
3319 
3320 		l2pte_set(ptep, npte, opte);
3321 		PTE_SYNC(ptep);
3322 #ifndef ARM_MMU_EXTENDED
3323 		bool is_cached = pmap_is_cached(pm);
3324 		if (is_cached) {
3325 			/*
3326 			 * We only need to frob the cache/tlb if this pmap
3327 			 * is current
3328 			 */
3329 			if (!vector_page_p && l2pte_valid_p(npte)) {
3330 				/*
3331 				 * This mapping is likely to be accessed as
3332 				 * soon as we return to userland. Fix up the
3333 				 * L1 entry to avoid taking another
3334 				 * page/domain fault.
3335 				 */
3336 				pd_entry_t *pdep = pmap_l1_kva(pm)
3337 				     + l1pte_index(va);
3338 				pd_entry_t pde = L1_C_PROTO | l2b->l2b_pa
3339 				    | L1_C_DOM(pmap_domain(pm));
3340 				if (*pdep != pde) {
3341 					l1pte_setone(pdep, pde);
3342 					PTE_SYNC(pdep);
3343 				}
3344 			}
3345 		}
3346 #endif /* !ARMM_MMU_EXTENDED */
3347 
3348 		pmap_tlb_flush_SE(pm, va, oflags);
3349 
3350 #ifndef ARM_MMU_EXTENDED
3351 		UVMHIST_LOG(maphist, "  is_cached %d cs 0x%08x\n",
3352 		    is_cached, pm->pm_cstate.cs_all, 0, 0);
3353 
3354 		if (pg != NULL) {
3355 			struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3356 
3357 			pmap_acquire_page_lock(md);
3358 			pmap_vac_me_harder(md, pa, pm, va);
3359 			pmap_release_page_lock(md);
3360 		}
3361 #endif
3362 	}
3363 #if defined(PMAP_CACHE_VIPT) && defined(DIAGNOSTIC)
3364 	if (pg) {
3365 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3366 
3367 		pmap_acquire_page_lock(md);
3368 #ifndef ARM_MMU_EXTENDED
3369 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
3370 #endif
3371 		PMAP_VALIDATE_MD_PAGE(md);
3372 		pmap_release_page_lock(md);
3373 	}
3374 #endif
3375 
3376 	pmap_release_pmap_lock(pm);
3377 
3378 	return (0);
3379 }
3380 
3381 /*
3382  * pmap_remove()
3383  *
3384  * pmap_remove is responsible for nuking a number of mappings for a range
3385  * of virtual address space in the current pmap. To do this efficiently
3386  * is interesting, because in a number of cases a wide virtual address
3387  * range may be supplied that contains few actual mappings. So, the
3388  * optimisations are:
3389  *  1. Skip over hunks of address space for which no L1 or L2 entry exists.
3390  *  2. Build up a list of pages we've hit, up to a maximum, so we can
3391  *     maybe do just a partial cache clean. This path of execution is
3392  *     complicated by the fact that the cache must be flushed _before_
3393  *     the PTE is nuked, being a VAC :-)
3394  *  3. If we're called after UVM calls pmap_remove_all(), we can defer
3395  *     all invalidations until pmap_update(), since pmap_remove_all() has
3396  *     already flushed the cache.
3397  *  4. Maybe later fast-case a single page, but I don't think this is
3398  *     going to make _that_ much difference overall.
3399  */
3400 
3401 #define	PMAP_REMOVE_CLEAN_LIST_SIZE	3
3402 
3403 void
3404 pmap_remove(pmap_t pm, vaddr_t sva, vaddr_t eva)
3405 {
3406 	vaddr_t next_bucket;
3407 	u_int cleanlist_idx, total, cnt;
3408 	struct {
3409 		vaddr_t va;
3410 		pt_entry_t *ptep;
3411 	} cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
3412 	u_int mappings;
3413 
3414 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
3415 	UVMHIST_LOG(maphist, " (pm=%p, sva=%#x, eva=%#x)", pm, sva, eva, 0);
3416 
3417 	/*
3418 	 * we lock in the pmap => pv_head direction
3419 	 */
3420 	pmap_acquire_pmap_lock(pm);
3421 
3422 	if (pm->pm_remove_all || !pmap_is_cached(pm)) {
3423 		cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
3424 #ifndef ARM_MMU_EXTENDED
3425 		if (pm->pm_cstate.cs_tlb == 0)
3426 			pm->pm_remove_all = true;
3427 #endif
3428 	} else
3429 		cleanlist_idx = 0;
3430 
3431 	total = 0;
3432 
3433 	while (sva < eva) {
3434 		/*
3435 		 * Do one L2 bucket's worth at a time.
3436 		 */
3437 		next_bucket = L2_NEXT_BUCKET_VA(sva);
3438 		if (next_bucket > eva)
3439 			next_bucket = eva;
3440 
3441 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pm, sva);
3442 		if (l2b == NULL) {
3443 			sva = next_bucket;
3444 			continue;
3445 		}
3446 
3447 		pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(sva)];
3448 
3449 		for (mappings = 0;
3450 		     sva < next_bucket;
3451 		     sva += PAGE_SIZE, ptep += PAGE_SIZE / L2_S_SIZE) {
3452 			pt_entry_t opte = *ptep;
3453 
3454 			if (opte == 0) {
3455 				/* Nothing here, move along */
3456 				continue;
3457 			}
3458 
3459 			u_int flags = PVF_REF;
3460 			paddr_t pa = l2pte_pa(opte);
3461 			struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
3462 
3463 			/*
3464 			 * Update flags. In a number of circumstances,
3465 			 * we could cluster a lot of these and do a
3466 			 * number of sequential pages in one go.
3467 			 */
3468 			if (pg != NULL) {
3469 				struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3470 				struct pv_entry *pv;
3471 
3472 				pmap_acquire_page_lock(md);
3473 				pv = pmap_remove_pv(md, pa, pm, sva);
3474 				pmap_vac_me_harder(md, pa, pm, 0);
3475 				pmap_release_page_lock(md);
3476 				if (pv != NULL) {
3477 					if (pm->pm_remove_all == false) {
3478 						flags = pv->pv_flags;
3479 					}
3480 					pool_put(&pmap_pv_pool, pv);
3481 				}
3482 			}
3483 			mappings += PAGE_SIZE / L2_S_SIZE;
3484 
3485 			if (!l2pte_valid_p(opte)) {
3486 				/*
3487 				 * Ref/Mod emulation is still active for this
3488 				 * mapping, therefore it is has not yet been
3489 				 * accessed. No need to frob the cache/tlb.
3490 				 */
3491 				l2pte_reset(ptep);
3492 				PTE_SYNC_CURRENT(pm, ptep);
3493 				continue;
3494 			}
3495 
3496 #ifdef ARM_MMU_EXTENDED
3497 			if (pm == pmap_kernel()) {
3498 				l2pte_reset(ptep);
3499 				PTE_SYNC(ptep);
3500 				pmap_tlb_flush_SE(pm, sva, flags);
3501 				continue;
3502 			}
3503 #endif
3504 			if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
3505 				/* Add to the clean list. */
3506 				cleanlist[cleanlist_idx].ptep = ptep;
3507 				cleanlist[cleanlist_idx].va =
3508 				    sva | (flags & PVF_EXEC);
3509 				cleanlist_idx++;
3510 			} else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
3511 				/* Nuke everything if needed. */
3512 #ifdef PMAP_CACHE_VIVT
3513 				pmap_cache_wbinv_all(pm, PVF_EXEC);
3514 #endif
3515 				/*
3516 				 * Roll back the previous PTE list,
3517 				 * and zero out the current PTE.
3518 				 */
3519 				for (cnt = 0;
3520 				     cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
3521 					l2pte_reset(cleanlist[cnt].ptep);
3522 					PTE_SYNC(cleanlist[cnt].ptep);
3523 				}
3524 				l2pte_reset(ptep);
3525 				PTE_SYNC(ptep);
3526 				cleanlist_idx++;
3527 				pm->pm_remove_all = true;
3528 			} else {
3529 				l2pte_reset(ptep);
3530 				PTE_SYNC(ptep);
3531 				if (pm->pm_remove_all == false) {
3532 					pmap_tlb_flush_SE(pm, sva, flags);
3533 				}
3534 			}
3535 		}
3536 
3537 		/*
3538 		 * Deal with any left overs
3539 		 */
3540 		if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
3541 			total += cleanlist_idx;
3542 			for (cnt = 0; cnt < cleanlist_idx; cnt++) {
3543 #ifdef ARM_MMU_EXTENDED
3544 				vaddr_t clva = cleanlist[cnt].va;
3545 				pmap_tlb_flush_SE(pm, clva, PVF_REF);
3546 #else
3547 				vaddr_t va = cleanlist[cnt].va;
3548 				if (pm->pm_cstate.cs_all != 0) {
3549 					vaddr_t clva = va & ~PAGE_MASK;
3550 					u_int flags = va & PVF_EXEC;
3551 #ifdef PMAP_CACHE_VIVT
3552 					pmap_cache_wbinv_page(pm, clva, true,
3553 					    PVF_REF | PVF_WRITE | flags);
3554 #endif
3555 					pmap_tlb_flush_SE(pm, clva,
3556 					    PVF_REF | flags);
3557 				}
3558 #endif /* ARM_MMU_EXTENDED */
3559 				l2pte_reset(cleanlist[cnt].ptep);
3560 				PTE_SYNC_CURRENT(pm, cleanlist[cnt].ptep);
3561 			}
3562 
3563 			/*
3564 			 * If it looks like we're removing a whole bunch
3565 			 * of mappings, it's faster to just write-back
3566 			 * the whole cache now and defer TLB flushes until
3567 			 * pmap_update() is called.
3568 			 */
3569 			if (total <= PMAP_REMOVE_CLEAN_LIST_SIZE)
3570 				cleanlist_idx = 0;
3571 			else {
3572 				cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
3573 #ifdef PMAP_CACHE_VIVT
3574 				pmap_cache_wbinv_all(pm, PVF_EXEC);
3575 #endif
3576 				pm->pm_remove_all = true;
3577 			}
3578 		}
3579 
3580 
3581 		pmap_free_l2_bucket(pm, l2b, mappings);
3582 		pm->pm_stats.resident_count -= mappings / (PAGE_SIZE/L2_S_SIZE);
3583 	}
3584 
3585 	pmap_release_pmap_lock(pm);
3586 }
3587 
3588 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3589 static struct pv_entry *
3590 pmap_kremove_pg(struct vm_page *pg, vaddr_t va)
3591 {
3592 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3593 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
3594 	struct pv_entry *pv;
3595 
3596 	KASSERT(arm_cache_prefer_mask == 0 || md->pvh_attrs & (PVF_COLORED|PVF_NC));
3597 	KASSERT((md->pvh_attrs & PVF_KMPAGE) == 0);
3598 	KASSERT(pmap_page_locked_p(md));
3599 
3600 	pv = pmap_remove_pv(md, pa, pmap_kernel(), va);
3601 	KASSERTMSG(pv, "pg %p (pa #%lx) va %#lx", pg, pa, va);
3602 	KASSERT(PV_IS_KENTRY_P(pv->pv_flags));
3603 
3604 	/*
3605 	 * If we are removing a writeable mapping to a cached exec page,
3606 	 * if it's the last mapping then clear it execness other sync
3607 	 * the page to the icache.
3608 	 */
3609 	if ((md->pvh_attrs & (PVF_NC|PVF_EXEC)) == PVF_EXEC
3610 	    && (pv->pv_flags & PVF_WRITE) != 0) {
3611 		if (SLIST_EMPTY(&md->pvh_list)) {
3612 			md->pvh_attrs &= ~PVF_EXEC;
3613 			PMAPCOUNT(exec_discarded_kremove);
3614 		} else {
3615 			pmap_syncicache_page(md, pa);
3616 			PMAPCOUNT(exec_synced_kremove);
3617 		}
3618 	}
3619 	pmap_vac_me_harder(md, pa, pmap_kernel(), 0);
3620 
3621 	return pv;
3622 }
3623 #endif /* PMAP_CACHE_VIPT && !ARM_MMU_EXTENDED */
3624 
3625 /*
3626  * pmap_kenter_pa: enter an unmanaged, wired kernel mapping
3627  *
3628  * We assume there is already sufficient KVM space available
3629  * to do this, as we can't allocate L2 descriptor tables/metadata
3630  * from here.
3631  */
3632 void
3633 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags)
3634 {
3635 #ifdef PMAP_CACHE_VIVT
3636 	struct vm_page *pg = (flags & PMAP_KMPAGE) ? PHYS_TO_VM_PAGE(pa) : NULL;
3637 #endif
3638 #ifdef PMAP_CACHE_VIPT
3639 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
3640 	struct vm_page *opg;
3641 #ifndef ARM_MMU_EXTENDED
3642 	struct pv_entry *pv = NULL;
3643 #endif
3644 #endif
3645 	struct vm_page_md *md = pg != NULL ? VM_PAGE_TO_MD(pg) : NULL;
3646 
3647 	UVMHIST_FUNC(__func__);
3648 
3649 	if (pmap_initialized) {
3650 		UVMHIST_CALLED(maphist);
3651 		UVMHIST_LOG(maphist, " (va=%#x, pa=%#x, prot=%#x, flags=%#x",
3652 		    va, pa, prot, flags);
3653 	}
3654 
3655 	pmap_t kpm = pmap_kernel();
3656 	struct l2_bucket * const l2b = pmap_get_l2_bucket(kpm, va);
3657 	const size_t l1slot __diagused = l1pte_index(va);
3658 	KASSERTMSG(l2b != NULL,
3659 	    "va %#lx pa %#lx prot %d maxkvaddr %#lx: l2 %p l2b %p kva %p",
3660 	    va, pa, prot, pmap_curmaxkvaddr, kpm->pm_l2[L2_IDX(l1slot)],
3661 	    kpm->pm_l2[L2_IDX(l1slot)]
3662 		? &kpm->pm_l2[L2_IDX(l1slot)]->l2_bucket[L2_BUCKET(l1slot)]
3663 		: NULL,
3664 	    kpm->pm_l2[L2_IDX(l1slot)]
3665 		? kpm->pm_l2[L2_IDX(l1slot)]->l2_bucket[L2_BUCKET(l1slot)].l2b_kva
3666 		: NULL);
3667 	KASSERT(l2b->l2b_kva != NULL);
3668 
3669 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
3670 	const pt_entry_t opte = *ptep;
3671 
3672 	if (opte == 0) {
3673 		PMAPCOUNT(kenter_mappings);
3674 		l2b->l2b_occupancy += PAGE_SIZE / L2_S_SIZE;
3675 	} else {
3676 		PMAPCOUNT(kenter_remappings);
3677 #ifdef PMAP_CACHE_VIPT
3678 		opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3679 #if !defined(ARM_MMU_EXTENDED) || defined(DIAGNOSTIC)
3680 		struct vm_page_md *omd __diagused = VM_PAGE_TO_MD(opg);
3681 #endif
3682 		if (opg && arm_cache_prefer_mask != 0) {
3683 			KASSERT(opg != pg);
3684 			KASSERT((omd->pvh_attrs & PVF_KMPAGE) == 0);
3685 			KASSERT((flags & PMAP_KMPAGE) == 0);
3686 #ifndef ARM_MMU_EXTENDED
3687 			pmap_acquire_page_lock(omd);
3688 			pv = pmap_kremove_pg(opg, va);
3689 			pmap_release_page_lock(omd);
3690 #endif
3691 		}
3692 #endif
3693 		if (l2pte_valid_p(opte)) {
3694 #ifdef PMAP_CACHE_VIVT
3695 			cpu_dcache_wbinv_range(va, PAGE_SIZE);
3696 #endif
3697 			cpu_tlb_flushD_SE(va);
3698 			cpu_cpwait();
3699 		}
3700 	}
3701 
3702 	pt_entry_t npte = L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot)
3703 	    | ((flags & PMAP_NOCACHE)
3704 		? 0
3705 		: ((flags & PMAP_PTE)
3706 		    ? pte_l2_s_cache_mode_pt : pte_l2_s_cache_mode));
3707 #ifdef ARM_MMU_EXTENDED
3708 	if (prot & VM_PROT_EXECUTE)
3709 		npte &= ~L2_XS_XN;
3710 #endif
3711 	l2pte_set(ptep, npte, opte);
3712 	PTE_SYNC(ptep);
3713 
3714 	if (pg) {
3715 		if (flags & PMAP_KMPAGE) {
3716 			KASSERT(md->urw_mappings == 0);
3717 			KASSERT(md->uro_mappings == 0);
3718 			KASSERT(md->krw_mappings == 0);
3719 			KASSERT(md->kro_mappings == 0);
3720 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3721 			KASSERT(pv == NULL);
3722 			KASSERT(arm_cache_prefer_mask == 0 || (va & PVF_COLORED) == 0);
3723 			KASSERT((md->pvh_attrs & PVF_NC) == 0);
3724 			/* if there is a color conflict, evict from cache. */
3725 			if (pmap_is_page_colored_p(md)
3726 			    && ((va ^ md->pvh_attrs) & arm_cache_prefer_mask)) {
3727 				PMAPCOUNT(vac_color_change);
3728 				pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
3729 			} else if (md->pvh_attrs & PVF_MULTCLR) {
3730 				/*
3731 				 * If this page has multiple colors, expunge
3732 				 * them.
3733 				 */
3734 				PMAPCOUNT(vac_flush_lots2);
3735 				pmap_flush_page(md, pa, PMAP_FLUSH_SECONDARY);
3736 			}
3737 			/*
3738 			 * Since this is a KMPAGE, there can be no contention
3739 			 * for this page so don't lock it.
3740 			 */
3741 			md->pvh_attrs &= PAGE_SIZE - 1;
3742 			md->pvh_attrs |= PVF_KMPAGE | PVF_COLORED | PVF_DIRTY
3743 			    | (va & arm_cache_prefer_mask);
3744 #else /* !PMAP_CACHE_VIPT || ARM_MMU_EXTENDED */
3745 			md->pvh_attrs |= PVF_KMPAGE;
3746 #endif
3747 			atomic_inc_32(&pmap_kmpages);
3748 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3749 		} else if (arm_cache_prefer_mask != 0) {
3750 			if (pv == NULL) {
3751 				pv = pool_get(&pmap_pv_pool, PR_NOWAIT);
3752 				KASSERT(pv != NULL);
3753 			}
3754 			pmap_acquire_page_lock(md);
3755 			pmap_enter_pv(md, pa, pv, pmap_kernel(), va,
3756 			    PVF_WIRED | PVF_KENTRY
3757 			    | (prot & VM_PROT_WRITE ? PVF_WRITE : 0));
3758 			if ((prot & VM_PROT_WRITE)
3759 			    && !(md->pvh_attrs & PVF_NC))
3760 				md->pvh_attrs |= PVF_DIRTY;
3761 			KASSERT((prot & VM_PROT_WRITE) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
3762 			pmap_vac_me_harder(md, pa, pmap_kernel(), va);
3763 			pmap_release_page_lock(md);
3764 #endif
3765 		}
3766 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3767 	} else {
3768 		if (pv != NULL)
3769 			pool_put(&pmap_pv_pool, pv);
3770 #endif
3771 	}
3772 	KASSERT(md == NULL || !pmap_page_locked_p(md));
3773 	if (pmap_initialized) {
3774 		UVMHIST_LOG(maphist, "  <-- done (ptep %p: %#x -> %#x)",
3775 		    ptep, opte, npte, 0);
3776 	}
3777 
3778 }
3779 
3780 void
3781 pmap_kremove(vaddr_t va, vsize_t len)
3782 {
3783 #ifdef UVMHIST
3784 	u_int total_mappings = 0;
3785 #endif
3786 
3787 	PMAPCOUNT(kenter_unmappings);
3788 
3789 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
3790 
3791 	UVMHIST_LOG(maphist, " (va=%#x, len=%#x)", va, len, 0, 0);
3792 
3793 	const vaddr_t eva = va + len;
3794 
3795 	while (va < eva) {
3796 		vaddr_t next_bucket = L2_NEXT_BUCKET_VA(va);
3797 		if (next_bucket > eva)
3798 			next_bucket = eva;
3799 
3800 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pmap_kernel(), va);
3801 		KDASSERT(l2b != NULL);
3802 
3803 		pt_entry_t * const sptep = &l2b->l2b_kva[l2pte_index(va)];
3804 		pt_entry_t *ptep = sptep;
3805 		u_int mappings = 0;
3806 
3807 		while (va < next_bucket) {
3808 			const pt_entry_t opte = *ptep;
3809 			struct vm_page *opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3810 			if (opg != NULL) {
3811 				struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3812 
3813 				if (omd->pvh_attrs & PVF_KMPAGE) {
3814 					KASSERT(omd->urw_mappings == 0);
3815 					KASSERT(omd->uro_mappings == 0);
3816 					KASSERT(omd->krw_mappings == 0);
3817 					KASSERT(omd->kro_mappings == 0);
3818 					omd->pvh_attrs &= ~PVF_KMPAGE;
3819 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3820 					if (arm_cache_prefer_mask != 0) {
3821 						omd->pvh_attrs &= ~PVF_WRITE;
3822 					}
3823 #endif
3824 					atomic_dec_32(&pmap_kmpages);
3825 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3826 				} else if (arm_cache_prefer_mask != 0) {
3827 					pmap_acquire_page_lock(omd);
3828 					pool_put(&pmap_pv_pool,
3829 					    pmap_kremove_pg(opg, va));
3830 					pmap_release_page_lock(omd);
3831 #endif
3832 				}
3833 			}
3834 			if (l2pte_valid_p(opte)) {
3835 #ifdef PMAP_CACHE_VIVT
3836 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
3837 #endif
3838 				cpu_tlb_flushD_SE(va);
3839 			}
3840 			if (opte) {
3841 				l2pte_reset(ptep);
3842 				mappings += PAGE_SIZE / L2_S_SIZE;
3843 			}
3844 			va += PAGE_SIZE;
3845 			ptep += PAGE_SIZE / L2_S_SIZE;
3846 		}
3847 		KDASSERTMSG(mappings <= l2b->l2b_occupancy, "%u %u",
3848 		    mappings, l2b->l2b_occupancy);
3849 		l2b->l2b_occupancy -= mappings;
3850 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
3851 #ifdef UVMHIST
3852 		total_mappings += mappings;
3853 #endif
3854 	}
3855 	cpu_cpwait();
3856 	UVMHIST_LOG(maphist, "  <--- done (%u mappings removed)",
3857 	    total_mappings, 0, 0, 0);
3858 }
3859 
3860 bool
3861 pmap_extract(pmap_t pm, vaddr_t va, paddr_t *pap)
3862 {
3863 	struct l2_dtable *l2;
3864 	pd_entry_t *pdep, pde;
3865 	pt_entry_t *ptep, pte;
3866 	paddr_t pa;
3867 	u_int l1slot;
3868 
3869 	pmap_acquire_pmap_lock(pm);
3870 
3871 	l1slot = l1pte_index(va);
3872 	pdep = pmap_l1_kva(pm) + l1slot;
3873 	pde = *pdep;
3874 
3875 	if (l1pte_section_p(pde)) {
3876 		/*
3877 		 * These should only happen for pmap_kernel()
3878 		 */
3879 		KDASSERT(pm == pmap_kernel());
3880 		pmap_release_pmap_lock(pm);
3881 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
3882 		if (l1pte_supersection_p(pde)) {
3883 			pa = (pde & L1_SS_FRAME) | (va & L1_SS_OFFSET);
3884 		} else
3885 #endif
3886 			pa = (pde & L1_S_FRAME) | (va & L1_S_OFFSET);
3887 	} else {
3888 		/*
3889 		 * Note that we can't rely on the validity of the L1
3890 		 * descriptor as an indication that a mapping exists.
3891 		 * We have to look it up in the L2 dtable.
3892 		 */
3893 		l2 = pm->pm_l2[L2_IDX(l1slot)];
3894 
3895 		if (l2 == NULL ||
3896 		    (ptep = l2->l2_bucket[L2_BUCKET(l1slot)].l2b_kva) == NULL) {
3897 			pmap_release_pmap_lock(pm);
3898 			return false;
3899 		}
3900 
3901 		pte = ptep[l2pte_index(va)];
3902 		pmap_release_pmap_lock(pm);
3903 
3904 		if (pte == 0)
3905 			return false;
3906 
3907 		switch (pte & L2_TYPE_MASK) {
3908 		case L2_TYPE_L:
3909 			pa = (pte & L2_L_FRAME) | (va & L2_L_OFFSET);
3910 			break;
3911 
3912 		default:
3913 			pa = (pte & ~PAGE_MASK) | (va & PAGE_MASK);
3914 			break;
3915 		}
3916 	}
3917 
3918 	if (pap != NULL)
3919 		*pap = pa;
3920 
3921 	return true;
3922 }
3923 
3924 void
3925 pmap_protect(pmap_t pm, vaddr_t sva, vaddr_t eva, vm_prot_t prot)
3926 {
3927 	struct l2_bucket *l2b;
3928 	vaddr_t next_bucket;
3929 
3930 	NPDEBUG(PDB_PROTECT,
3931 	    printf("pmap_protect: pm %p sva 0x%lx eva 0x%lx prot 0x%x\n",
3932 	    pm, sva, eva, prot));
3933 
3934 	if ((prot & VM_PROT_READ) == 0) {
3935 		pmap_remove(pm, sva, eva);
3936 		return;
3937 	}
3938 
3939 	if (prot & VM_PROT_WRITE) {
3940 		/*
3941 		 * If this is a read->write transition, just ignore it and let
3942 		 * uvm_fault() take care of it later.
3943 		 */
3944 		return;
3945 	}
3946 
3947 	pmap_acquire_pmap_lock(pm);
3948 
3949 	const bool flush = eva - sva >= PAGE_SIZE * 4;
3950 	u_int clr_mask = PVF_WRITE | ((prot & VM_PROT_EXECUTE) ? 0 : PVF_EXEC);
3951 	u_int flags = 0;
3952 
3953 	while (sva < eva) {
3954 		next_bucket = L2_NEXT_BUCKET_VA(sva);
3955 		if (next_bucket > eva)
3956 			next_bucket = eva;
3957 
3958 		l2b = pmap_get_l2_bucket(pm, sva);
3959 		if (l2b == NULL) {
3960 			sva = next_bucket;
3961 			continue;
3962 		}
3963 
3964 		pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(sva)];
3965 
3966 		while (sva < next_bucket) {
3967 			const pt_entry_t opte = *ptep;
3968 			if (l2pte_valid_p(opte) && l2pte_writable_p(opte)) {
3969 				struct vm_page *pg;
3970 				u_int f;
3971 
3972 #ifdef PMAP_CACHE_VIVT
3973 				/*
3974 				 * OK, at this point, we know we're doing
3975 				 * write-protect operation.  If the pmap is
3976 				 * active, write-back the page.
3977 				 */
3978 				pmap_cache_wbinv_page(pm, sva, false,
3979 				    PVF_REF | PVF_WRITE);
3980 #endif
3981 
3982 				pg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3983 				pt_entry_t npte = l2pte_set_readonly(opte);
3984 				l2pte_set(ptep, npte, opte);
3985 				PTE_SYNC(ptep);
3986 
3987 				if (pg != NULL) {
3988 					struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3989 					paddr_t pa = VM_PAGE_TO_PHYS(pg);
3990 
3991 					pmap_acquire_page_lock(md);
3992 					f = pmap_modify_pv(md, pa, pm, sva,
3993 					    clr_mask, 0);
3994 					pmap_vac_me_harder(md, pa, pm, sva);
3995 					pmap_release_page_lock(md);
3996 				} else {
3997 					f = PVF_REF | PVF_EXEC;
3998 				}
3999 
4000 				if (flush) {
4001 					flags |= f;
4002 				} else {
4003 					pmap_tlb_flush_SE(pm, sva, f);
4004 				}
4005 			}
4006 
4007 			sva += PAGE_SIZE;
4008 			ptep += PAGE_SIZE / L2_S_SIZE;
4009 		}
4010 	}
4011 
4012 	if (flush) {
4013 		if (PV_BEEN_EXECD(flags)) {
4014 			pmap_tlb_flushID(pm);
4015 		} else if (PV_BEEN_REFD(flags)) {
4016 			pmap_tlb_flushD(pm);
4017 		}
4018 	}
4019 
4020 	pmap_release_pmap_lock(pm);
4021 }
4022 
4023 void
4024 pmap_icache_sync_range(pmap_t pm, vaddr_t sva, vaddr_t eva)
4025 {
4026 	struct l2_bucket *l2b;
4027 	pt_entry_t *ptep;
4028 	vaddr_t next_bucket;
4029 	vsize_t page_size = trunc_page(sva) + PAGE_SIZE - sva;
4030 
4031 	NPDEBUG(PDB_EXEC,
4032 	    printf("pmap_icache_sync_range: pm %p sva 0x%lx eva 0x%lx\n",
4033 	    pm, sva, eva));
4034 
4035 	pmap_acquire_pmap_lock(pm);
4036 
4037 	while (sva < eva) {
4038 		next_bucket = L2_NEXT_BUCKET_VA(sva);
4039 		if (next_bucket > eva)
4040 			next_bucket = eva;
4041 
4042 		l2b = pmap_get_l2_bucket(pm, sva);
4043 		if (l2b == NULL) {
4044 			sva = next_bucket;
4045 			continue;
4046 		}
4047 
4048 		for (ptep = &l2b->l2b_kva[l2pte_index(sva)];
4049 		     sva < next_bucket;
4050 		     sva += page_size,
4051 		     ptep += PAGE_SIZE / L2_S_SIZE,
4052 		     page_size = PAGE_SIZE) {
4053 			if (l2pte_valid_p(*ptep)) {
4054 				cpu_icache_sync_range(sva,
4055 				    min(page_size, eva - sva));
4056 			}
4057 		}
4058 	}
4059 
4060 	pmap_release_pmap_lock(pm);
4061 }
4062 
4063 void
4064 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
4065 {
4066 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4067 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
4068 
4069 	NPDEBUG(PDB_PROTECT,
4070 	    printf("pmap_page_protect: md %p (0x%08lx), prot 0x%x\n",
4071 	    md, pa, prot));
4072 
4073 	switch(prot) {
4074 	case VM_PROT_READ|VM_PROT_WRITE:
4075 #if defined(ARM_MMU_EXTENDED)
4076 		pmap_acquire_page_lock(md);
4077 		pmap_clearbit(md, pa, PVF_EXEC);
4078 		pmap_release_page_lock(md);
4079 		break;
4080 #endif
4081 	case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
4082 		break;
4083 
4084 	case VM_PROT_READ:
4085 #if defined(ARM_MMU_EXTENDED)
4086 		pmap_acquire_page_lock(md);
4087 		pmap_clearbit(md, pa, PVF_WRITE|PVF_EXEC);
4088 		pmap_release_page_lock(md);
4089 		break;
4090 #endif
4091 	case VM_PROT_READ|VM_PROT_EXECUTE:
4092 		pmap_acquire_page_lock(md);
4093 		pmap_clearbit(md, pa, PVF_WRITE);
4094 		pmap_release_page_lock(md);
4095 		break;
4096 
4097 	default:
4098 		pmap_page_remove(md, pa);
4099 		break;
4100 	}
4101 }
4102 
4103 /*
4104  * pmap_clear_modify:
4105  *
4106  *	Clear the "modified" attribute for a page.
4107  */
4108 bool
4109 pmap_clear_modify(struct vm_page *pg)
4110 {
4111 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4112 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
4113 	bool rv;
4114 
4115 	pmap_acquire_page_lock(md);
4116 
4117 	if (md->pvh_attrs & PVF_MOD) {
4118 		rv = true;
4119 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
4120 		/*
4121 		 * If we are going to clear the modified bit and there are
4122 		 * no other modified bits set, flush the page to memory and
4123 		 * mark it clean.
4124 		 */
4125 		if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) == PVF_MOD)
4126 			pmap_flush_page(md, pa, PMAP_CLEAN_PRIMARY);
4127 #endif
4128 		pmap_clearbit(md, pa, PVF_MOD);
4129 	} else {
4130 		rv = false;
4131 	}
4132 	pmap_release_page_lock(md);
4133 
4134 	return rv;
4135 }
4136 
4137 /*
4138  * pmap_clear_reference:
4139  *
4140  *	Clear the "referenced" attribute for a page.
4141  */
4142 bool
4143 pmap_clear_reference(struct vm_page *pg)
4144 {
4145 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4146 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
4147 	bool rv;
4148 
4149 	pmap_acquire_page_lock(md);
4150 
4151 	if (md->pvh_attrs & PVF_REF) {
4152 		rv = true;
4153 		pmap_clearbit(md, pa, PVF_REF);
4154 	} else {
4155 		rv = false;
4156 	}
4157 	pmap_release_page_lock(md);
4158 
4159 	return rv;
4160 }
4161 
4162 /*
4163  * pmap_is_modified:
4164  *
4165  *	Test if a page has the "modified" attribute.
4166  */
4167 /* See <arm/arm32/pmap.h> */
4168 
4169 /*
4170  * pmap_is_referenced:
4171  *
4172  *	Test if a page has the "referenced" attribute.
4173  */
4174 /* See <arm/arm32/pmap.h> */
4175 
4176 #if defined(ARM_MMU_EXTENDED) && 0
4177 int
4178 pmap_prefetchabt_fixup(void *v)
4179 {
4180 	struct trapframe * const tf = v;
4181 	vaddr_t va = trunc_page(tf->tf_pc);
4182 	int rv = ABORT_FIXUP_FAILED;
4183 
4184 	if (!TRAP_USERMODE(tf) && va < VM_MAXUSER_ADDRESS)
4185 		return rv;
4186 
4187 	kpreempt_disable();
4188 	pmap_t pm = curcpu()->ci_pmap_cur;
4189 	const size_t l1slot = l1pte_index(va);
4190 	struct l2_dtable * const l2 = pm->pm_l2[L2_IDX(l1slot)];
4191 	if (l2 == NULL)
4192 		goto out;
4193 
4194 	struct l2_bucket * const l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
4195 	if (l2b->l2b_kva == NULL)
4196 		goto out;
4197 
4198 	/*
4199 	 * Check the PTE itself.
4200 	 */
4201 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
4202 	const pt_entry_t opte = *ptep;
4203 	if ((opte & L2_S_PROT_U) == 0 || (opte & L2_XS_XN) == 0)
4204 		goto out;
4205 
4206 	paddr_t pa = l2pte_pa(pte);
4207 	struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
4208 	KASSERT(pg != NULL);
4209 
4210 	struct vm_page_md * const md = VM_PAGE_TO_MD(pg);
4211 
4212 	pmap_acquire_page_lock(md);
4213 	struct pv_entry * const pv = pmap_find_pv(md, pm, va);
4214 	KASSERT(pv != NULL);
4215 
4216 	if (PV_IS_EXEC_P(pv->pv_flags)) {
4217 		if (!PV_IS_EXEC_P(md->pvh_attrs)) {
4218 			pmap_syncicache_page(md, pa);
4219 		}
4220 		rv = ABORT_FIXUP_RETURN;
4221 		l2pte_set(ptep, opte & ~L2_XS_XN, opte);
4222 		pmap_tlb_flush_SE(pm, va, PVF_EXEC | PVF_REF);
4223 	}
4224 	pmap_release_page_lock(md);
4225 
4226   out:
4227 	kpreempt_enable();
4228 	return rv;
4229 }
4230 #endif
4231 
4232 int
4233 pmap_fault_fixup(pmap_t pm, vaddr_t va, vm_prot_t ftype, int user)
4234 {
4235 	struct l2_dtable *l2;
4236 	struct l2_bucket *l2b;
4237 	paddr_t pa;
4238 	const size_t l1slot = l1pte_index(va);
4239 	int rv = 0;
4240 
4241 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
4242 
4243 	va = trunc_page(va);
4244 
4245 	KASSERT(!user || (pm != pmap_kernel()));
4246 
4247 	UVMHIST_LOG(maphist, " (pm=%#x, va=%#x, ftype=%#x, user=%d)",
4248 	    pm, va, ftype, user);
4249 #ifdef ARM_MMU_EXTENDED
4250 	UVMHIST_LOG(maphist, " ti=%#x pai=%#x asid=%#x",
4251 	    cpu_tlb_info(curcpu()), PMAP_PAI(pm, cpu_tlb_info(curcpu())),
4252 	    PMAP_PAI(pm, cpu_tlb_info(curcpu()))->pai_asid, 0);
4253 #endif
4254 
4255 	pmap_acquire_pmap_lock(pm);
4256 
4257 	/*
4258 	 * If there is no l2_dtable for this address, then the process
4259 	 * has no business accessing it.
4260 	 *
4261 	 * Note: This will catch userland processes trying to access
4262 	 * kernel addresses.
4263 	 */
4264 	l2 = pm->pm_l2[L2_IDX(l1slot)];
4265 	if (l2 == NULL) {
4266 		UVMHIST_LOG(maphist, " no l2 for l1slot %#x", l1slot, 0, 0, 0);
4267 		goto out;
4268 	}
4269 
4270 	/*
4271 	 * Likewise if there is no L2 descriptor table
4272 	 */
4273 	l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
4274 	if (l2b->l2b_kva == NULL) {
4275 		UVMHIST_LOG(maphist, " <-- done (no ptep for l1slot %#x)", l1slot, 0, 0, 0);
4276 		goto out;
4277 	}
4278 
4279 	/*
4280 	 * Check the PTE itself.
4281 	 */
4282 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
4283 	pt_entry_t const opte = *ptep;
4284 	if (opte == 0 || (opte & L2_TYPE_MASK) == L2_TYPE_L) {
4285 		UVMHIST_LOG(maphist, " <-- done (empty pde for l1slot %#x)", l1slot, 0, 0, 0);
4286 		goto out;
4287 	}
4288 
4289 #ifndef ARM_HAS_VBAR
4290 	/*
4291 	 * Catch a userland access to the vector page mapped at 0x0
4292 	 */
4293 	if (user && (opte & L2_S_PROT_U) == 0) {
4294 		UVMHIST_LOG(maphist, " <-- done (vector_page)", 0, 0, 0, 0);
4295 		goto out;
4296 	}
4297 #endif
4298 
4299 	pa = l2pte_pa(opte);
4300 
4301 	if ((ftype & VM_PROT_WRITE) && !l2pte_writable_p(opte)) {
4302 		/*
4303 		 * This looks like a good candidate for "page modified"
4304 		 * emulation...
4305 		 */
4306 		struct pv_entry *pv;
4307 		struct vm_page *pg;
4308 
4309 		/* Extract the physical address of the page */
4310 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) {
4311 			UVMHIST_LOG(maphist, " <-- done (mod/ref unmanaged page)", 0, 0, 0, 0);
4312 			goto out;
4313 		}
4314 
4315 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4316 
4317 		/* Get the current flags for this page. */
4318 		pmap_acquire_page_lock(md);
4319 		pv = pmap_find_pv(md, pm, va);
4320 		if (pv == NULL || PV_IS_KENTRY_P(pv->pv_flags)) {
4321 			pmap_release_page_lock(md);
4322 			UVMHIST_LOG(maphist, " <-- done (mod/ref emul: no PV)", 0, 0, 0, 0);
4323 			goto out;
4324 		}
4325 
4326 		/*
4327 		 * Do the flags say this page is writable? If not then it
4328 		 * is a genuine write fault. If yes then the write fault is
4329 		 * our fault as we did not reflect the write access in the
4330 		 * PTE. Now we know a write has occurred we can correct this
4331 		 * and also set the modified bit
4332 		 */
4333 		if ((pv->pv_flags & PVF_WRITE) == 0) {
4334 			pmap_release_page_lock(md);
4335 			goto out;
4336 		}
4337 
4338 		md->pvh_attrs |= PVF_REF | PVF_MOD;
4339 		pv->pv_flags |= PVF_REF | PVF_MOD;
4340 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
4341 		/*
4342 		 * If there are cacheable mappings for this page, mark it dirty.
4343 		 */
4344 		if ((md->pvh_attrs & PVF_NC) == 0)
4345 			md->pvh_attrs |= PVF_DIRTY;
4346 #endif
4347 #ifdef ARM_MMU_EXTENDED
4348 		if (md->pvh_attrs & PVF_EXEC) {
4349 			md->pvh_attrs &= ~PVF_EXEC;
4350 			PMAPCOUNT(exec_discarded_modfixup);
4351 		}
4352 #endif
4353 		pmap_release_page_lock(md);
4354 
4355 		/*
4356 		 * Re-enable write permissions for the page.  No need to call
4357 		 * pmap_vac_me_harder(), since this is just a
4358 		 * modified-emulation fault, and the PVF_WRITE bit isn't
4359 		 * changing. We've already set the cacheable bits based on
4360 		 * the assumption that we can write to this page.
4361 		 */
4362 		const pt_entry_t npte =
4363 		    l2pte_set_writable((opte & ~L2_TYPE_MASK) | L2_S_PROTO)
4364 #ifdef ARM_MMU_EXTENDED
4365 		    | (pm != pmap_kernel() ? L2_XS_nG : 0)
4366 #endif
4367 		    | 0;
4368 		l2pte_set(ptep, npte, opte);
4369 		PTE_SYNC(ptep);
4370 		PMAPCOUNT(fixup_mod);
4371 		rv = 1;
4372 		UVMHIST_LOG(maphist, " <-- done (mod/ref emul: changed pte from %#x to %#x)",
4373 		    opte, npte, 0, 0);
4374 	} else if ((opte & L2_TYPE_MASK) == L2_TYPE_INV) {
4375 		/*
4376 		 * This looks like a good candidate for "page referenced"
4377 		 * emulation.
4378 		 */
4379 		struct vm_page *pg;
4380 
4381 		/* Extract the physical address of the page */
4382 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) {
4383 			UVMHIST_LOG(maphist, " <-- done (ref emul: unmanaged page)", 0, 0, 0, 0);
4384 			goto out;
4385 		}
4386 
4387 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4388 
4389 		/* Get the current flags for this page. */
4390 		pmap_acquire_page_lock(md);
4391 		struct pv_entry *pv = pmap_find_pv(md, pm, va);
4392 		if (pv == NULL || PV_IS_KENTRY_P(pv->pv_flags)) {
4393 			pmap_release_page_lock(md);
4394 			UVMHIST_LOG(maphist, " <-- done (ref emul no PV)", 0, 0, 0, 0);
4395 			goto out;
4396 		}
4397 
4398 		md->pvh_attrs |= PVF_REF;
4399 		pv->pv_flags |= PVF_REF;
4400 
4401 		pt_entry_t npte =
4402 		    l2pte_set_readonly((opte & ~L2_TYPE_MASK) | L2_S_PROTO);
4403 #ifdef ARM_MMU_EXTENDED
4404 		if (pm != pmap_kernel()) {
4405 			npte |= L2_XS_nG;
4406 		}
4407 		/*
4408 		 * If we got called from prefetch abort, then ftype will have
4409 		 * VM_PROT_EXECUTE set.  Now see if we have no-execute set in
4410 		 * the PTE.
4411 		 */
4412 		if (user && (ftype & VM_PROT_EXECUTE) && (npte & L2_XS_XN)) {
4413 			/*
4414 			 * Is this a mapping of an executable page?
4415 			 */
4416 			if ((pv->pv_flags & PVF_EXEC) == 0) {
4417 				pmap_release_page_lock(md);
4418 				UVMHIST_LOG(maphist, " <-- done (ref emul: no exec)",
4419 				    0, 0, 0, 0);
4420 				goto out;
4421 			}
4422 			/*
4423 			 * If we haven't synced the page, do so now.
4424 			 */
4425 			if ((md->pvh_attrs & PVF_EXEC) == 0) {
4426 				UVMHIST_LOG(maphist, " ref emul: syncicache page #%#x",
4427 				    pa, 0, 0, 0);
4428 				pmap_syncicache_page(md, pa);
4429 				PMAPCOUNT(fixup_exec);
4430 			}
4431 			npte &= ~L2_XS_XN;
4432 		}
4433 #endif /* ARM_MMU_EXTENDED */
4434 		pmap_release_page_lock(md);
4435 		l2pte_set(ptep, npte, opte);
4436 		PTE_SYNC(ptep);
4437 		PMAPCOUNT(fixup_ref);
4438 		rv = 1;
4439 		UVMHIST_LOG(maphist, " <-- done (ref emul: changed pte from %#x to %#x)",
4440 		    opte, npte, 0, 0);
4441 #ifdef ARM_MMU_EXTENDED
4442 	} else if (user && (ftype & VM_PROT_EXECUTE) && (opte & L2_XS_XN)) {
4443 		struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
4444 		if (pg == NULL) {
4445 			UVMHIST_LOG(maphist, " <-- done (unmanaged page)", 0, 0, 0, 0);
4446 			goto out;
4447 		}
4448 
4449 		struct vm_page_md * const md = VM_PAGE_TO_MD(pg);
4450 
4451 		/* Get the current flags for this page. */
4452 		pmap_acquire_page_lock(md);
4453 		struct pv_entry * const pv = pmap_find_pv(md, pm, va);
4454 		if (pv == NULL || (pv->pv_flags & PVF_EXEC) == 0) {
4455 			pmap_release_page_lock(md);
4456 			UVMHIST_LOG(maphist, " <-- done (no PV or not EXEC)", 0, 0, 0, 0);
4457 			goto out;
4458 		}
4459 
4460 		/*
4461 		 * If we haven't synced the page, do so now.
4462 		 */
4463 		if ((md->pvh_attrs & PVF_EXEC) == 0) {
4464 			UVMHIST_LOG(maphist, "syncicache page #%#x",
4465 			    pa, 0, 0, 0);
4466 			pmap_syncicache_page(md, pa);
4467 		}
4468 		pmap_release_page_lock(md);
4469 		/*
4470 		 * Turn off no-execute.
4471 		 */
4472 		KASSERT(opte & L2_XS_nG);
4473 		l2pte_set(ptep, opte & ~L2_XS_XN, opte);
4474 		PTE_SYNC(ptep);
4475 		rv = 1;
4476 		PMAPCOUNT(fixup_exec);
4477 		UVMHIST_LOG(maphist, "exec: changed pte from %#x to %#x",
4478 		    opte, opte & ~L2_XS_XN, 0, 0);
4479 #endif
4480 	}
4481 
4482 #ifndef ARM_MMU_EXTENDED
4483 	/*
4484 	 * We know there is a valid mapping here, so simply
4485 	 * fix up the L1 if necessary.
4486 	 */
4487 	pd_entry_t * const pdep = pmap_l1_kva(pm) + l1slot;
4488 	pd_entry_t pde = L1_C_PROTO | l2b->l2b_pa | L1_C_DOM(pmap_domain(pm));
4489 	if (*pdep != pde) {
4490 		l1pte_setone(pdep, pde);
4491 		PTE_SYNC(pdep);
4492 		rv = 1;
4493 		PMAPCOUNT(fixup_pdes);
4494 	}
4495 #endif
4496 
4497 #ifdef CPU_SA110
4498 	/*
4499 	 * There are bugs in the rev K SA110.  This is a check for one
4500 	 * of them.
4501 	 */
4502 	if (rv == 0 && curcpu()->ci_arm_cputype == CPU_ID_SA110 &&
4503 	    curcpu()->ci_arm_cpurev < 3) {
4504 		/* Always current pmap */
4505 		if (l2pte_valid_p(opte)) {
4506 			extern int kernel_debug;
4507 			if (kernel_debug & 1) {
4508 				struct proc *p = curlwp->l_proc;
4509 				printf("prefetch_abort: page is already "
4510 				    "mapped - pte=%p *pte=%08x\n", ptep, opte);
4511 				printf("prefetch_abort: pc=%08lx proc=%p "
4512 				    "process=%s\n", va, p, p->p_comm);
4513 				printf("prefetch_abort: far=%08x fs=%x\n",
4514 				    cpu_faultaddress(), cpu_faultstatus());
4515 			}
4516 #ifdef DDB
4517 			if (kernel_debug & 2)
4518 				Debugger();
4519 #endif
4520 			rv = 1;
4521 		}
4522 	}
4523 #endif /* CPU_SA110 */
4524 
4525 #ifndef ARM_MMU_EXTENDED
4526 	/*
4527 	 * If 'rv == 0' at this point, it generally indicates that there is a
4528 	 * stale TLB entry for the faulting address.  That might be due to a
4529 	 * wrong setting of pmap_needs_pte_sync.  So set it and retry.
4530 	 */
4531 	if (rv == 0
4532 	    && pm->pm_l1->l1_domain_use_count == 1
4533 	    && pmap_needs_pte_sync == 0) {
4534 		pmap_needs_pte_sync = 1;
4535 		PTE_SYNC(ptep);
4536 		PMAPCOUNT(fixup_ptesync);
4537 		rv = 1;
4538 	}
4539 #endif
4540 
4541 #if defined(DEBUG) || 1
4542 	/*
4543 	 * If 'rv == 0' at this point, it generally indicates that there is a
4544 	 * stale TLB entry for the faulting address. This happens when two or
4545 	 * more processes are sharing an L1. Since we don't flush the TLB on
4546 	 * a context switch between such processes, we can take domain faults
4547 	 * for mappings which exist at the same VA in both processes. EVEN IF
4548 	 * WE'VE RECENTLY FIXED UP THE CORRESPONDING L1 in pmap_enter(), for
4549 	 * example.
4550 	 *
4551 	 * This is extremely likely to happen if pmap_enter() updated the L1
4552 	 * entry for a recently entered mapping. In this case, the TLB is
4553 	 * flushed for the new mapping, but there may still be TLB entries for
4554 	 * other mappings belonging to other processes in the 1MB range
4555 	 * covered by the L1 entry.
4556 	 *
4557 	 * Since 'rv == 0', we know that the L1 already contains the correct
4558 	 * value, so the fault must be due to a stale TLB entry.
4559 	 *
4560 	 * Since we always need to flush the TLB anyway in the case where we
4561 	 * fixed up the L1, or frobbed the L2 PTE, we effectively deal with
4562 	 * stale TLB entries dynamically.
4563 	 *
4564 	 * However, the above condition can ONLY happen if the current L1 is
4565 	 * being shared. If it happens when the L1 is unshared, it indicates
4566 	 * that other parts of the pmap are not doing their job WRT managing
4567 	 * the TLB.
4568 	 */
4569 	if (rv == 0
4570 #ifndef ARM_MMU_EXTENDED
4571 	    && pm->pm_l1->l1_domain_use_count == 1
4572 #endif
4573 	    && true) {
4574 #ifdef DEBUG
4575 		extern int last_fault_code;
4576 #else
4577 		int last_fault_code = ftype & VM_PROT_EXECUTE
4578 		    ? armreg_ifsr_read()
4579 		    : armreg_dfsr_read();
4580 #endif
4581 		printf("fixup: pm %p, va 0x%lx, ftype %d - nothing to do!\n",
4582 		    pm, va, ftype);
4583 		printf("fixup: l2 %p, l2b %p, ptep %p, pte %#x\n",
4584 		    l2, l2b, ptep, opte);
4585 
4586 #ifndef ARM_MMU_EXTENDED
4587 		printf("fixup: pdep %p, pde %#x, fsr %#x\n",
4588 		    pdep, pde, last_fault_code);
4589 #else
4590 		printf("fixup: pdep %p, pde %#x, ttbcr %#x\n",
4591 		    &pmap_l1_kva(pm)[l1slot], pmap_l1_kva(pm)[l1slot],
4592 		   armreg_ttbcr_read());
4593 		printf("fixup: fsr %#x cpm %p casid %#x contextidr %#x dacr %#x\n",
4594 		    last_fault_code, curcpu()->ci_pmap_cur,
4595 		    curcpu()->ci_pmap_asid_cur,
4596 		    armreg_contextidr_read(), armreg_dacr_read());
4597 #ifdef _ARM_ARCH_7
4598 		if (ftype & VM_PROT_WRITE)
4599 			armreg_ats1cuw_write(va);
4600 		else
4601 			armreg_ats1cur_write(va);
4602 		arm_isb();
4603 		printf("fixup: par %#x\n", armreg_par_read());
4604 #endif
4605 #endif
4606 #ifdef DDB
4607 		extern int kernel_debug;
4608 
4609 		if (kernel_debug & 2) {
4610 			pmap_release_pmap_lock(pm);
4611 #ifdef UVMHIST
4612 			KERNHIST_DUMP(maphist);
4613 #endif
4614 			cpu_Debugger();
4615 			pmap_acquire_pmap_lock(pm);
4616 		}
4617 #endif
4618 	}
4619 #endif
4620 
4621 	pmap_tlb_flush_SE(pm, va,
4622 	    (ftype & VM_PROT_EXECUTE) ? PVF_EXEC | PVF_REF : PVF_REF);
4623 
4624 	rv = 1;
4625 
4626 out:
4627 	pmap_release_pmap_lock(pm);
4628 
4629 	return (rv);
4630 }
4631 
4632 /*
4633  * Routine:	pmap_procwr
4634  *
4635  * Function:
4636  *	Synchronize caches corresponding to [addr, addr+len) in p.
4637  *
4638  */
4639 void
4640 pmap_procwr(struct proc *p, vaddr_t va, int len)
4641 {
4642 	/* We only need to do anything if it is the current process. */
4643 	if (p == curproc)
4644 		cpu_icache_sync_range(va, len);
4645 }
4646 
4647 /*
4648  * Routine:	pmap_unwire
4649  * Function:	Clear the wired attribute for a map/virtual-address pair.
4650  *
4651  * In/out conditions:
4652  *		The mapping must already exist in the pmap.
4653  */
4654 void
4655 pmap_unwire(pmap_t pm, vaddr_t va)
4656 {
4657 	struct l2_bucket *l2b;
4658 	pt_entry_t *ptep, pte;
4659 	struct vm_page *pg;
4660 	paddr_t pa;
4661 
4662 	NPDEBUG(PDB_WIRING, printf("pmap_unwire: pm %p, va 0x%08lx\n", pm, va));
4663 
4664 	pmap_acquire_pmap_lock(pm);
4665 
4666 	l2b = pmap_get_l2_bucket(pm, va);
4667 	KDASSERT(l2b != NULL);
4668 
4669 	ptep = &l2b->l2b_kva[l2pte_index(va)];
4670 	pte = *ptep;
4671 
4672 	/* Extract the physical address of the page */
4673 	pa = l2pte_pa(pte);
4674 
4675 	if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
4676 		/* Update the wired bit in the pv entry for this page. */
4677 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4678 
4679 		pmap_acquire_page_lock(md);
4680 		(void) pmap_modify_pv(md, pa, pm, va, PVF_WIRED, 0);
4681 		pmap_release_page_lock(md);
4682 	}
4683 
4684 	pmap_release_pmap_lock(pm);
4685 }
4686 
4687 void
4688 pmap_activate(struct lwp *l)
4689 {
4690 	struct cpu_info * const ci = curcpu();
4691 	extern int block_userspace_access;
4692 	pmap_t npm = l->l_proc->p_vmspace->vm_map.pmap;
4693 #ifdef ARM_MMU_EXTENDED
4694 	struct pmap_asid_info * const pai = PMAP_PAI(npm, cpu_tlb_info(ci));
4695 #endif
4696 
4697 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
4698 
4699 	UVMHIST_LOG(maphist, "(l=%#x) pm=%#x", l, npm, 0, 0);
4700 
4701 	/*
4702 	 * If activating a non-current lwp or the current lwp is
4703 	 * already active, just return.
4704 	 */
4705 	if (false
4706 	    || l != curlwp
4707 #ifdef ARM_MMU_EXTENDED
4708 	    || (ci->ci_pmap_cur == npm &&
4709 		(npm == pmap_kernel()
4710 		 /* || PMAP_PAI_ASIDVALID_P(pai, cpu_tlb_info(ci)) */))
4711 #else
4712 	    || npm->pm_activated == true
4713 #endif
4714 	    || false) {
4715 		UVMHIST_LOG(maphist, " <-- (same pmap)", curlwp, l, 0, 0);
4716 		return;
4717 	}
4718 
4719 #ifndef ARM_MMU_EXTENDED
4720 	const uint32_t ndacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2))
4721 	    | (DOMAIN_CLIENT << (pmap_domain(npm) * 2));
4722 
4723 	/*
4724 	 * If TTB and DACR are unchanged, short-circuit all the
4725 	 * TLB/cache management stuff.
4726 	 */
4727 	pmap_t opm = ci->ci_lastlwp
4728 	    ? ci->ci_lastlwp->l_proc->p_vmspace->vm_map.pmap
4729 	    : NULL;
4730 	if (opm != NULL) {
4731 		uint32_t odacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2))
4732 		    | (DOMAIN_CLIENT << (pmap_domain(opm) * 2));
4733 
4734 		if (opm->pm_l1 == npm->pm_l1 && odacr == ndacr)
4735 			goto all_done;
4736 	}
4737 #endif /* !ARM_MMU_EXTENDED */
4738 
4739 	PMAPCOUNT(activations);
4740 	block_userspace_access = 1;
4741 
4742 #ifndef ARM_MMU_EXTENDED
4743 	/*
4744 	 * If switching to a user vmspace which is different to the
4745 	 * most recent one, and the most recent one is potentially
4746 	 * live in the cache, we must write-back and invalidate the
4747 	 * entire cache.
4748 	 */
4749 	pmap_t rpm = ci->ci_pmap_lastuser;
4750 #endif
4751 
4752 /*
4753  * XXXSCW: There's a corner case here which can leave turds in the cache as
4754  * reported in kern/41058. They're probably left over during tear-down and
4755  * switching away from an exiting process. Until the root cause is identified
4756  * and fixed, zap the cache when switching pmaps. This will result in a few
4757  * unnecessary cache flushes, but that's better than silently corrupting data.
4758  */
4759 #ifndef ARM_MMU_EXTENDED
4760 #if 0
4761 	if (npm != pmap_kernel() && rpm && npm != rpm &&
4762 	    rpm->pm_cstate.cs_cache) {
4763 		rpm->pm_cstate.cs_cache = 0;
4764 #ifdef PMAP_CACHE_VIVT
4765 		cpu_idcache_wbinv_all();
4766 #endif
4767 	}
4768 #else
4769 	if (rpm) {
4770 		rpm->pm_cstate.cs_cache = 0;
4771 		if (npm == pmap_kernel())
4772 			ci->ci_pmap_lastuser = NULL;
4773 #ifdef PMAP_CACHE_VIVT
4774 		cpu_idcache_wbinv_all();
4775 #endif
4776 	}
4777 #endif
4778 
4779 	/* No interrupts while we frob the TTB/DACR */
4780 	uint32_t oldirqstate = disable_interrupts(IF32_bits);
4781 #endif /* !ARM_MMU_EXTENDED */
4782 
4783 #ifndef ARM_HAS_VBAR
4784 	/*
4785 	 * For ARM_VECTORS_LOW, we MUST, I repeat, MUST fix up the L1
4786 	 * entry corresponding to 'vector_page' in the incoming L1 table
4787 	 * before switching to it otherwise subsequent interrupts/exceptions
4788 	 * (including domain faults!) will jump into hyperspace.
4789 	 */
4790 	if (npm->pm_pl1vec != NULL) {
4791 		cpu_tlb_flushID_SE((u_int)vector_page);
4792 		cpu_cpwait();
4793 		*npm->pm_pl1vec = npm->pm_l1vec;
4794 		PTE_SYNC(npm->pm_pl1vec);
4795 	}
4796 #endif
4797 
4798 #ifdef ARM_MMU_EXTENDED
4799 	/*
4800 	 * Assume that TTBR1 has only global mappings and TTBR0 only has
4801 	 * non-global mappings.  To prevent speculation from doing evil things
4802 	 * we disable translation table walks using TTBR0 before setting the
4803 	 * CONTEXTIDR (ASID) or new TTBR0 value.  Once both are set, table
4804 	 * walks are reenabled.
4805 	 */
4806 	UVMHIST_LOG(maphist, " acquiring asid", 0, 0, 0, 0);
4807 	const uint32_t old_ttbcr = armreg_ttbcr_read();
4808 	armreg_ttbcr_write(old_ttbcr | TTBCR_S_PD0);
4809 	arm_isb();
4810 	pmap_tlb_asid_acquire(npm, l);
4811 	UVMHIST_LOG(maphist, " setting ttbr pa=%#x asid=%#x", npm->pm_l1_pa, pai->pai_asid, 0, 0);
4812 	cpu_setttb(npm->pm_l1_pa, pai->pai_asid);
4813 	/*
4814 	 * Now we can reenable tablewalks since the CONTEXTIDR and TTRB0 have
4815 	 * been updated.
4816 	 */
4817 	arm_isb();
4818 	if (npm != pmap_kernel()) {
4819 		armreg_ttbcr_write(old_ttbcr & ~TTBCR_S_PD0);
4820 	}
4821 	cpu_cpwait();
4822 	ci->ci_pmap_asid_cur = pai->pai_asid;
4823 #else
4824 	cpu_domains(ndacr);
4825 	if (npm == pmap_kernel() || npm == rpm) {
4826 		/*
4827 		 * Switching to a kernel thread, or back to the
4828 		 * same user vmspace as before... Simply update
4829 		 * the TTB (no TLB flush required)
4830 		 */
4831 		cpu_setttb(npm->pm_l1->l1_physaddr, false);
4832 		cpu_cpwait();
4833 	} else {
4834 		/*
4835 		 * Otherwise, update TTB and flush TLB
4836 		 */
4837 		cpu_context_switch(npm->pm_l1->l1_physaddr);
4838 		if (rpm != NULL)
4839 			rpm->pm_cstate.cs_tlb = 0;
4840 	}
4841 
4842 	restore_interrupts(oldirqstate);
4843 #endif /* ARM_MMU_EXTENDED */
4844 
4845 	block_userspace_access = 0;
4846 
4847 #ifndef ARM_MMU_EXTENDED
4848  all_done:
4849 	/*
4850 	 * The new pmap is resident. Make sure it's marked
4851 	 * as resident in the cache/TLB.
4852 	 */
4853 	npm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
4854 	if (npm != pmap_kernel())
4855 		ci->ci_pmap_lastuser = npm;
4856 
4857 	/* The old pmap is not longer active */
4858 	if (opm != npm) {
4859 		if (opm != NULL)
4860 			opm->pm_activated = false;
4861 
4862 		/* But the new one is */
4863 		npm->pm_activated = true;
4864 	}
4865 #endif
4866 	ci->ci_pmap_cur = npm;
4867 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
4868 }
4869 
4870 void
4871 pmap_deactivate(struct lwp *l)
4872 {
4873 	pmap_t pm = l->l_proc->p_vmspace->vm_map.pmap;
4874 
4875 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
4876 
4877 	UVMHIST_LOG(maphist, "(l=%#x) pm=%#x", l, pm, 0, 0);
4878 
4879 #ifdef ARM_MMU_EXTENDED
4880 	kpreempt_disable();
4881 	struct cpu_info * const ci = curcpu();
4882 	struct pmap_asid_info * const pai = PMAP_PAI(pm, cpu_tlb_info(ci));
4883 	/*
4884 	 * Disable translation table walks from TTBR0 while no pmap has been
4885 	 * activated.
4886 	 */
4887 	const uint32_t old_ttbcr = armreg_ttbcr_read();
4888 	armreg_ttbcr_write(old_ttbcr | TTBCR_S_PD0);
4889 	arm_isb();
4890 	pmap_tlb_asid_deactivate(pm);
4891 	cpu_setttb(pmap_kernel()->pm_l1_pa, pai->pai_asid);
4892 	ci->ci_pmap_cur = pmap_kernel();
4893 	kpreempt_enable();
4894 #else
4895 	/*
4896 	 * If the process is exiting, make sure pmap_activate() does
4897 	 * a full MMU context-switch and cache flush, which we might
4898 	 * otherwise skip. See PR port-arm/38950.
4899 	 */
4900 	if (l->l_proc->p_sflag & PS_WEXIT)
4901 		curcpu()->ci_lastlwp = NULL;
4902 
4903 	pm->pm_activated = false;
4904 #endif
4905 	UVMHIST_LOG(maphist, "  <-- done", 0, 0, 0, 0);
4906 }
4907 
4908 void
4909 pmap_update(pmap_t pm)
4910 {
4911 
4912 	if (pm->pm_remove_all) {
4913 #ifdef ARM_MMU_EXTENDED
4914 		KASSERTMSG(curcpu()->ci_pmap_cur != pm || pm->pm_pai[0].pai_asid == curcpu()->ci_pmap_asid_cur, "pmap/asid %p/%#x != %s cur pmap/asid %p/%#x", pm, pm->pm_pai[0].pai_asid, curcpu()->ci_data.cpu_name, curcpu()->ci_pmap_cur, curcpu()->ci_pmap_asid_cur);
4915 		/*
4916 		 * Finish up the pmap_remove_all() optimisation by flushing
4917 		 * all our ASIDs.
4918 		 */
4919 		pmap_tlb_asid_release_all(pm);
4920 #else
4921 		/*
4922 		 * Finish up the pmap_remove_all() optimisation by flushing
4923 		 * the TLB.
4924 		 */
4925 		pmap_tlb_flushID(pm);
4926 #endif
4927 		pm->pm_remove_all = false;
4928 	}
4929 
4930 #ifdef ARM_MMU_EXTENDED
4931 
4932 #if defined(MULTIPROCESSOR) && PMAP_MAX_TLB > 1
4933 	u_int pending = atomic_swap_uint(&pmap->pm_shootdown_pending, 0);
4934 	if (pending && pmap_tlb_shootdown_bystanders(pmap)) {
4935 		PMAP_COUNT(shootdown_ipis);
4936 	}
4937 #endif
4938 	KASSERTMSG(curcpu()->ci_pmap_cur != pm || pm->pm_pai[0].pai_asid == curcpu()->ci_pmap_asid_cur, "pmap/asid %p/%#x != %s cur pmap/asid %p/%#x", pm, pm->pm_pai[0].pai_asid, curcpu()->ci_data.cpu_name, curcpu()->ci_pmap_cur, curcpu()->ci_pmap_asid_cur);
4939 #else
4940 	if (pmap_is_current(pm)) {
4941 		/*
4942 		 * If we're dealing with a current userland pmap, move its L1
4943 		 * to the end of the LRU.
4944 		 */
4945 		if (pm != pmap_kernel())
4946 			pmap_use_l1(pm);
4947 
4948 		/*
4949 		 * We can assume we're done with frobbing the cache/tlb for
4950 		 * now. Make sure any future pmap ops don't skip cache/tlb
4951 		 * flushes.
4952 		 */
4953 		pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
4954 	}
4955 #endif
4956 
4957 	PMAPCOUNT(updates);
4958 
4959 	/*
4960 	 * make sure TLB/cache operations have completed.
4961 	 */
4962 	cpu_cpwait();
4963 }
4964 
4965 void
4966 pmap_remove_all(pmap_t pm)
4967 {
4968 
4969 	/*
4970 	 * The vmspace described by this pmap is about to be torn down.
4971 	 * Until pmap_update() is called, UVM will only make calls
4972 	 * to pmap_remove(). We can make life much simpler by flushing
4973 	 * the cache now, and deferring TLB invalidation to pmap_update().
4974 	 */
4975 #ifdef PMAP_CACHE_VIVT
4976 	pmap_cache_wbinv_all(pm, PVF_EXEC);
4977 #endif
4978 	pm->pm_remove_all = true;
4979 }
4980 
4981 /*
4982  * Retire the given physical map from service.
4983  * Should only be called if the map contains no valid mappings.
4984  */
4985 void
4986 pmap_destroy(pmap_t pm)
4987 {
4988 	u_int count;
4989 
4990 	if (pm == NULL)
4991 		return;
4992 
4993 	if (pm->pm_remove_all) {
4994 		pmap_tlb_flushID(pm);
4995 		pm->pm_remove_all = false;
4996 	}
4997 
4998 	/*
4999 	 * Drop reference count
5000 	 */
5001 	mutex_enter(pm->pm_lock);
5002 	count = --pm->pm_obj.uo_refs;
5003 	mutex_exit(pm->pm_lock);
5004 	if (count > 0) {
5005 #ifndef ARM_MMU_EXTENDED
5006 		if (pmap_is_current(pm)) {
5007 			if (pm != pmap_kernel())
5008 				pmap_use_l1(pm);
5009 			pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
5010 		}
5011 #endif
5012 		return;
5013 	}
5014 
5015 	/*
5016 	 * reference count is zero, free pmap resources and then free pmap.
5017 	 */
5018 
5019 #ifndef ARM_HAS_VBAR
5020 	if (vector_page < KERNEL_BASE) {
5021 		KDASSERT(!pmap_is_current(pm));
5022 
5023 		/* Remove the vector page mapping */
5024 		pmap_remove(pm, vector_page, vector_page + PAGE_SIZE);
5025 		pmap_update(pm);
5026 	}
5027 #endif
5028 
5029 	LIST_REMOVE(pm, pm_list);
5030 
5031 	pmap_free_l1(pm);
5032 
5033 #ifdef ARM_MMU_EXTENDED
5034 #ifdef MULTIPROCESSOR
5035 	kcpuset_destroy(pm->pm_active);
5036 	kcpuset_destroy(pm->pm_onproc);
5037 #endif
5038 #else
5039 	struct cpu_info * const ci = curcpu();
5040 	if (ci->ci_pmap_lastuser == pm)
5041 		ci->ci_pmap_lastuser = NULL;
5042 #endif
5043 
5044 	uvm_obj_destroy(&pm->pm_obj, false);
5045 	mutex_destroy(&pm->pm_obj_lock);
5046 	pool_cache_put(&pmap_cache, pm);
5047 }
5048 
5049 
5050 /*
5051  * void pmap_reference(pmap_t pm)
5052  *
5053  * Add a reference to the specified pmap.
5054  */
5055 void
5056 pmap_reference(pmap_t pm)
5057 {
5058 
5059 	if (pm == NULL)
5060 		return;
5061 
5062 #ifndef ARM_MMU_EXTENDED
5063 	pmap_use_l1(pm);
5064 #endif
5065 
5066 	mutex_enter(pm->pm_lock);
5067 	pm->pm_obj.uo_refs++;
5068 	mutex_exit(pm->pm_lock);
5069 }
5070 
5071 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
5072 
5073 static struct evcnt pmap_prefer_nochange_ev =
5074     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap prefer", "nochange");
5075 static struct evcnt pmap_prefer_change_ev =
5076     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap prefer", "change");
5077 
5078 EVCNT_ATTACH_STATIC(pmap_prefer_change_ev);
5079 EVCNT_ATTACH_STATIC(pmap_prefer_nochange_ev);
5080 
5081 void
5082 pmap_prefer(vaddr_t hint, vaddr_t *vap, int td)
5083 {
5084 	vsize_t mask = arm_cache_prefer_mask | (PAGE_SIZE - 1);
5085 	vaddr_t va = *vap;
5086 	vaddr_t diff = (hint - va) & mask;
5087 	if (diff == 0) {
5088 		pmap_prefer_nochange_ev.ev_count++;
5089 	} else {
5090 		pmap_prefer_change_ev.ev_count++;
5091 		if (__predict_false(td))
5092 			va -= mask + 1;
5093 		*vap = va + diff;
5094 	}
5095 }
5096 #endif /* ARM_MMU_V6 | ARM_MMU_V7 */
5097 
5098 /*
5099  * pmap_zero_page()
5100  *
5101  * Zero a given physical page by mapping it at a page hook point.
5102  * In doing the zero page op, the page we zero is mapped cachable, as with
5103  * StrongARM accesses to non-cached pages are non-burst making writing
5104  * _any_ bulk data very slow.
5105  */
5106 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
5107 void
5108 pmap_zero_page_generic(paddr_t pa)
5109 {
5110 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
5111 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
5112 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
5113 #endif
5114 #if defined(PMAP_CACHE_VIPT)
5115 	/* Choose the last page color it had, if any */
5116 	const vsize_t va_offset = md->pvh_attrs & arm_cache_prefer_mask;
5117 #else
5118 	const vsize_t va_offset = 0;
5119 #endif
5120 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
5121 	/*
5122 	 * Is this page mapped at its natural color?
5123 	 * If we have all of memory mapped, then just convert PA to VA.
5124 	 */
5125 	bool okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5126 	   || va_offset == (pa & arm_cache_prefer_mask);
5127 	const vaddr_t vdstp = okcolor
5128 	    ? pmap_direct_mapped_phys(pa, &okcolor, cpu_cdstp(va_offset))
5129 	    : cpu_cdstp(va_offset);
5130 #else
5131 	const bool okcolor = false;
5132 	const vaddr_t vdstp = cpu_cdstp(va_offset);
5133 #endif
5134 	pt_entry_t * const ptep = cpu_cdst_pte(va_offset);
5135 
5136 
5137 #ifdef DEBUG
5138 	if (!SLIST_EMPTY(&md->pvh_list))
5139 		panic("pmap_zero_page: page has mappings");
5140 #endif
5141 
5142 	KDASSERT((pa & PGOFSET) == 0);
5143 
5144 	if (!okcolor) {
5145 		/*
5146 		 * Hook in the page, zero it, and purge the cache for that
5147 		 * zeroed page. Invalidate the TLB as needed.
5148 		 */
5149 		const pt_entry_t npte = L2_S_PROTO | pa | pte_l2_s_cache_mode
5150 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE);
5151 		l2pte_set(ptep, npte, 0);
5152 		PTE_SYNC(ptep);
5153 		cpu_tlb_flushD_SE(vdstp);
5154 		cpu_cpwait();
5155 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS) && defined(PMAP_CACHE_VIPT) \
5156     && !defined(ARM_MMU_EXTENDED)
5157 		/*
5158 		 * If we are direct-mapped and our color isn't ok, then before
5159 		 * we bzero the page invalidate its contents from the cache and
5160 		 * reset the color to its natural color.
5161 		 */
5162 		cpu_dcache_inv_range(vdstp, PAGE_SIZE);
5163 		md->pvh_attrs &= ~arm_cache_prefer_mask;
5164 		md->pvh_attrs |= (pa & arm_cache_prefer_mask);
5165 #endif
5166 	}
5167 	bzero_page(vdstp);
5168 	if (!okcolor) {
5169 		/*
5170 		 * Unmap the page.
5171 		 */
5172 		l2pte_reset(ptep);
5173 		PTE_SYNC(ptep);
5174 		cpu_tlb_flushD_SE(vdstp);
5175 #ifdef PMAP_CACHE_VIVT
5176 		cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
5177 #endif
5178 	}
5179 #ifdef PMAP_CACHE_VIPT
5180 	/*
5181 	 * This page is now cache resident so it now has a page color.
5182 	 * Any contents have been obliterated so clear the EXEC flag.
5183 	 */
5184 #ifndef ARM_MMU_EXTENDED
5185 	if (!pmap_is_page_colored_p(md)) {
5186 		PMAPCOUNT(vac_color_new);
5187 		md->pvh_attrs |= PVF_COLORED;
5188 	}
5189 	md->pvh_attrs |= PVF_DIRTY;
5190 #endif
5191 	if (PV_IS_EXEC_P(md->pvh_attrs)) {
5192 		md->pvh_attrs &= ~PVF_EXEC;
5193 		PMAPCOUNT(exec_discarded_zero);
5194 	}
5195 #endif
5196 }
5197 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
5198 
5199 #if ARM_MMU_XSCALE == 1
5200 void
5201 pmap_zero_page_xscale(paddr_t pa)
5202 {
5203 #ifdef DEBUG
5204 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
5205 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
5206 
5207 	if (!SLIST_EMPTY(&md->pvh_list))
5208 		panic("pmap_zero_page: page has mappings");
5209 #endif
5210 
5211 	KDASSERT((pa & PGOFSET) == 0);
5212 
5213 	/*
5214 	 * Hook in the page, zero it, and purge the cache for that
5215 	 * zeroed page. Invalidate the TLB as needed.
5216 	 */
5217 
5218 	pt_entry_t npte = L2_S_PROTO | pa |
5219 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
5220 	    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
5221 	l2pte_set(cdst_pte, npte, 0);
5222 	PTE_SYNC(cdst_pte);
5223 	cpu_tlb_flushD_SE(cdstp);
5224 	cpu_cpwait();
5225 	bzero_page(cdstp);
5226 	xscale_cache_clean_minidata();
5227 	l2pte_reset(cdst_pte);
5228 	PTE_SYNC(cdst_pte);
5229 }
5230 #endif /* ARM_MMU_XSCALE == 1 */
5231 
5232 /* pmap_pageidlezero()
5233  *
5234  * The same as above, except that we assume that the page is not
5235  * mapped.  This means we never have to flush the cache first.  Called
5236  * from the idle loop.
5237  */
5238 bool
5239 pmap_pageidlezero(paddr_t pa)
5240 {
5241 	bool rv = true;
5242 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
5243 	struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
5244 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
5245 #endif
5246 #ifdef PMAP_CACHE_VIPT
5247 	/* Choose the last page color it had, if any */
5248 	const vsize_t va_offset = md->pvh_attrs & arm_cache_prefer_mask;
5249 #else
5250 	const vsize_t va_offset = 0;
5251 #endif
5252 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
5253 	bool okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5254 	   || va_offset == (pa & arm_cache_prefer_mask);
5255 	const vaddr_t vdstp = okcolor
5256 	    ? pmap_direct_mapped_phys(pa, &okcolor, cpu_cdstp(va_offset))
5257 	    : cpu_cdstp(va_offset);
5258 #else
5259 	const bool okcolor = false;
5260 	const vaddr_t vdstp = cpu_cdstp(va_offset);
5261 #endif
5262 	pt_entry_t * const ptep = cpu_cdst_pte(va_offset);
5263 
5264 
5265 #ifdef DEBUG
5266 	if (!SLIST_EMPTY(&md->pvh_list))
5267 		panic("pmap_pageidlezero: page has mappings");
5268 #endif
5269 
5270 	KDASSERT((pa & PGOFSET) == 0);
5271 
5272 	if (!okcolor) {
5273 		/*
5274 		 * Hook in the page, zero it, and purge the cache for that
5275 		 * zeroed page. Invalidate the TLB as needed.
5276 		 */
5277 		const pt_entry_t npte = L2_S_PROTO | pa |
5278 		    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
5279 		l2pte_set(ptep, npte, 0);
5280 		PTE_SYNC(ptep);
5281 		cpu_tlb_flushD_SE(vdstp);
5282 		cpu_cpwait();
5283 	}
5284 
5285 	uint64_t *ptr = (uint64_t *)vdstp;
5286 	for (size_t i = 0; i < PAGE_SIZE / sizeof(*ptr); i++) {
5287 		if (sched_curcpu_runnable_p() != 0) {
5288 			/*
5289 			 * A process has become ready.  Abort now,
5290 			 * so we don't keep it waiting while we
5291 			 * do slow memory access to finish this
5292 			 * page.
5293 			 */
5294 			rv = false;
5295 			break;
5296 		}
5297 		*ptr++ = 0;
5298 	}
5299 
5300 #ifdef PMAP_CACHE_VIVT
5301 	if (rv)
5302 		/*
5303 		 * if we aborted we'll rezero this page again later so don't
5304 		 * purge it unless we finished it
5305 		 */
5306 		cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
5307 #elif defined(PMAP_CACHE_VIPT)
5308 	/*
5309 	 * This page is now cache resident so it now has a page color.
5310 	 * Any contents have been obliterated so clear the EXEC flag.
5311 	 */
5312 #ifndef ARM_MMU_EXTENDED
5313 	if (!pmap_is_page_colored_p(md)) {
5314 		PMAPCOUNT(vac_color_new);
5315 		md->pvh_attrs |= PVF_COLORED;
5316 	}
5317 #endif
5318 	if (PV_IS_EXEC_P(md->pvh_attrs)) {
5319 		md->pvh_attrs &= ~PVF_EXEC;
5320 		PMAPCOUNT(exec_discarded_zero);
5321 	}
5322 #endif
5323 	/*
5324 	 * Unmap the page.
5325 	 */
5326 	if (!okcolor) {
5327 		l2pte_reset(ptep);
5328 		PTE_SYNC(ptep);
5329 		cpu_tlb_flushD_SE(vdstp);
5330 	}
5331 
5332 	return rv;
5333 }
5334 
5335 /*
5336  * pmap_copy_page()
5337  *
5338  * Copy one physical page into another, by mapping the pages into
5339  * hook points. The same comment regarding cachability as in
5340  * pmap_zero_page also applies here.
5341  */
5342 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
5343 void
5344 pmap_copy_page_generic(paddr_t src, paddr_t dst)
5345 {
5346 	struct vm_page * const src_pg = PHYS_TO_VM_PAGE(src);
5347 	struct vm_page_md *src_md = VM_PAGE_TO_MD(src_pg);
5348 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
5349 	struct vm_page * const dst_pg = PHYS_TO_VM_PAGE(dst);
5350 	struct vm_page_md *dst_md = VM_PAGE_TO_MD(dst_pg);
5351 #endif
5352 #ifdef PMAP_CACHE_VIPT
5353 	const vsize_t src_va_offset = src_md->pvh_attrs & arm_cache_prefer_mask;
5354 	const vsize_t dst_va_offset = dst_md->pvh_attrs & arm_cache_prefer_mask;
5355 #else
5356 	const vsize_t src_va_offset = 0;
5357 	const vsize_t dst_va_offset = 0;
5358 #endif
5359 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
5360 	/*
5361 	 * Is this page mapped at its natural color?
5362 	 * If we have all of memory mapped, then just convert PA to VA.
5363 	 */
5364 	bool src_okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5365 	    || src_va_offset == (src & arm_cache_prefer_mask);
5366 	bool dst_okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5367 	    || dst_va_offset == (dst & arm_cache_prefer_mask);
5368 	const vaddr_t vsrcp = src_okcolor
5369 	    ? pmap_direct_mapped_phys(src, &src_okcolor,
5370 		cpu_csrcp(src_va_offset))
5371 	    : cpu_csrcp(src_va_offset);
5372 	const vaddr_t vdstp = pmap_direct_mapped_phys(dst, &dst_okcolor,
5373 	    cpu_cdstp(dst_va_offset));
5374 #else
5375 	const bool src_okcolor = false;
5376 	const bool dst_okcolor = false;
5377 	const vaddr_t vsrcp = cpu_csrcp(src_va_offset);
5378 	const vaddr_t vdstp = cpu_cdstp(dst_va_offset);
5379 #endif
5380 	pt_entry_t * const src_ptep = cpu_csrc_pte(src_va_offset);
5381 	pt_entry_t * const dst_ptep = cpu_cdst_pte(dst_va_offset);
5382 
5383 #ifdef DEBUG
5384 	if (!SLIST_EMPTY(&dst_md->pvh_list))
5385 		panic("pmap_copy_page: dst page has mappings");
5386 #endif
5387 
5388 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
5389 	KASSERT(arm_cache_prefer_mask == 0 || src_md->pvh_attrs & (PVF_COLORED|PVF_NC));
5390 #endif
5391 	KDASSERT((src & PGOFSET) == 0);
5392 	KDASSERT((dst & PGOFSET) == 0);
5393 
5394 	/*
5395 	 * Clean the source page.  Hold the source page's lock for
5396 	 * the duration of the copy so that no other mappings can
5397 	 * be created while we have a potentially aliased mapping.
5398 	 */
5399 #ifdef PMAP_CACHE_VIVT
5400 	pmap_acquire_page_lock(src_md);
5401 	(void) pmap_clean_page(src_md, true);
5402 	pmap_release_page_lock(src_md);
5403 #endif
5404 
5405 	/*
5406 	 * Map the pages into the page hook points, copy them, and purge
5407 	 * the cache for the appropriate page. Invalidate the TLB
5408 	 * as required.
5409 	 */
5410 	if (!src_okcolor) {
5411 		const pt_entry_t nsrc_pte = L2_S_PROTO
5412 		    | src
5413 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
5414 		    | ((src_md->pvh_attrs & PVF_NC) ? 0 : pte_l2_s_cache_mode)
5415 #else // defined(PMAP_CACHE_VIVT) || defined(ARM_MMU_EXTENDED)
5416 		    | pte_l2_s_cache_mode
5417 #endif
5418 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ);
5419 		l2pte_set(src_ptep, nsrc_pte, 0);
5420 		PTE_SYNC(src_ptep);
5421 		cpu_tlb_flushD_SE(vsrcp);
5422 		cpu_cpwait();
5423 	}
5424 	if (!dst_okcolor) {
5425 		const pt_entry_t ndst_pte = L2_S_PROTO | dst |
5426 		    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
5427 		l2pte_set(dst_ptep, ndst_pte, 0);
5428 		PTE_SYNC(dst_ptep);
5429 		cpu_tlb_flushD_SE(vdstp);
5430 		cpu_cpwait();
5431 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS) && defined(PMAP_CACHE_VIPT)
5432 		/*
5433 		 * If we are direct-mapped and our color isn't ok, then before
5434 		 * we bcopy to the new page invalidate its contents from the
5435 		 * cache and reset its color to its natural color.
5436 		 */
5437 		cpu_dcache_inv_range(vdstp, PAGE_SIZE);
5438 		dst_md->pvh_attrs &= ~arm_cache_prefer_mask;
5439 		dst_md->pvh_attrs |= (dst & arm_cache_prefer_mask);
5440 #endif
5441 	}
5442 	bcopy_page(vsrcp, vdstp);
5443 #ifdef PMAP_CACHE_VIVT
5444 	cpu_dcache_inv_range(vsrcp, PAGE_SIZE);
5445 	cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
5446 #endif
5447 	/*
5448 	 * Unmap the pages.
5449 	 */
5450 	if (!src_okcolor) {
5451 		l2pte_reset(src_ptep);
5452 		PTE_SYNC(src_ptep);
5453 		cpu_tlb_flushD_SE(vsrcp);
5454 		cpu_cpwait();
5455 	}
5456 	if (!dst_okcolor) {
5457 		l2pte_reset(dst_ptep);
5458 		PTE_SYNC(dst_ptep);
5459 		cpu_tlb_flushD_SE(vdstp);
5460 		cpu_cpwait();
5461 	}
5462 #ifdef PMAP_CACHE_VIPT
5463 	/*
5464 	 * Now that the destination page is in the cache, mark it as colored.
5465 	 * If this was an exec page, discard it.
5466 	 */
5467 	pmap_acquire_page_lock(dst_md);
5468 #ifndef ARM_MMU_EXTENDED
5469 	if (arm_pcache.cache_type == CACHE_TYPE_PIPT) {
5470 		dst_md->pvh_attrs &= ~arm_cache_prefer_mask;
5471 		dst_md->pvh_attrs |= (dst & arm_cache_prefer_mask);
5472 	}
5473 	if (!pmap_is_page_colored_p(dst_md)) {
5474 		PMAPCOUNT(vac_color_new);
5475 		dst_md->pvh_attrs |= PVF_COLORED;
5476 	}
5477 	dst_md->pvh_attrs |= PVF_DIRTY;
5478 #endif
5479 	if (PV_IS_EXEC_P(dst_md->pvh_attrs)) {
5480 		dst_md->pvh_attrs &= ~PVF_EXEC;
5481 		PMAPCOUNT(exec_discarded_copy);
5482 	}
5483 	pmap_release_page_lock(dst_md);
5484 #endif
5485 }
5486 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
5487 
5488 #if ARM_MMU_XSCALE == 1
5489 void
5490 pmap_copy_page_xscale(paddr_t src, paddr_t dst)
5491 {
5492 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
5493 	struct vm_page_md *src_md = VM_PAGE_TO_MD(src_pg);
5494 #ifdef DEBUG
5495 	struct vm_page_md *dst_md = VM_PAGE_TO_MD(PHYS_TO_VM_PAGE(dst));
5496 
5497 	if (!SLIST_EMPTY(&dst_md->pvh_list))
5498 		panic("pmap_copy_page: dst page has mappings");
5499 #endif
5500 
5501 	KDASSERT((src & PGOFSET) == 0);
5502 	KDASSERT((dst & PGOFSET) == 0);
5503 
5504 	/*
5505 	 * Clean the source page.  Hold the source page's lock for
5506 	 * the duration of the copy so that no other mappings can
5507 	 * be created while we have a potentially aliased mapping.
5508 	 */
5509 #ifdef PMAP_CACHE_VIVT
5510 	pmap_acquire_page_lock(src_md);
5511 	(void) pmap_clean_page(src_md, true);
5512 	pmap_release_page_lock(src_md);
5513 #endif
5514 
5515 	/*
5516 	 * Map the pages into the page hook points, copy them, and purge
5517 	 * the cache for the appropriate page. Invalidate the TLB
5518 	 * as required.
5519 	 */
5520 	*csrc_pte = L2_S_PROTO | src |
5521 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
5522 	    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
5523 	PTE_SYNC(csrc_pte);
5524 	*cdst_pte = L2_S_PROTO | dst |
5525 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
5526 	    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
5527 	PTE_SYNC(cdst_pte);
5528 	cpu_tlb_flushD_SE(csrcp);
5529 	cpu_tlb_flushD_SE(cdstp);
5530 	cpu_cpwait();
5531 	bcopy_page(csrcp, cdstp);
5532 	xscale_cache_clean_minidata();
5533 }
5534 #endif /* ARM_MMU_XSCALE == 1 */
5535 
5536 /*
5537  * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
5538  *
5539  * Return the start and end addresses of the kernel's virtual space.
5540  * These values are setup in pmap_bootstrap and are updated as pages
5541  * are allocated.
5542  */
5543 void
5544 pmap_virtual_space(vaddr_t *start, vaddr_t *end)
5545 {
5546 	*start = virtual_avail;
5547 	*end = virtual_end;
5548 }
5549 
5550 /*
5551  * Helper function for pmap_grow_l2_bucket()
5552  */
5553 static inline int
5554 pmap_grow_map(vaddr_t va, paddr_t *pap)
5555 {
5556 	paddr_t pa;
5557 
5558 	if (uvm.page_init_done == false) {
5559 #ifdef PMAP_STEAL_MEMORY
5560 		pv_addr_t pv;
5561 		pmap_boot_pagealloc(PAGE_SIZE,
5562 #ifdef PMAP_CACHE_VIPT
5563 		    arm_cache_prefer_mask,
5564 		    va & arm_cache_prefer_mask,
5565 #else
5566 		    0, 0,
5567 #endif
5568 		    &pv);
5569 		pa = pv.pv_pa;
5570 #else
5571 		if (uvm_page_physget(&pa) == false)
5572 			return (1);
5573 #endif	/* PMAP_STEAL_MEMORY */
5574 	} else {
5575 		struct vm_page *pg;
5576 		pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE);
5577 		if (pg == NULL)
5578 			return (1);
5579 		pa = VM_PAGE_TO_PHYS(pg);
5580 		/*
5581 		 * This new page must not have any mappings.  Enter it via
5582 		 * pmap_kenter_pa and let that routine do the hard work.
5583 		 */
5584 		struct vm_page_md *md __diagused = VM_PAGE_TO_MD(pg);
5585 		KASSERT(SLIST_EMPTY(&md->pvh_list));
5586 		pmap_kenter_pa(va, pa,
5587 		    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE|PMAP_PTE);
5588 	}
5589 
5590 	if (pap)
5591 		*pap = pa;
5592 
5593 	PMAPCOUNT(pt_mappings);
5594 #ifdef DEBUG
5595 	struct l2_bucket * const l2b = pmap_get_l2_bucket(pmap_kernel(), va);
5596 	KDASSERT(l2b != NULL);
5597 
5598 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
5599 	const pt_entry_t opte = *ptep;
5600 	KDASSERT((opte & L2_S_CACHE_MASK) == pte_l2_s_cache_mode_pt);
5601 #endif
5602 	memset((void *)va, 0, PAGE_SIZE);
5603 	return (0);
5604 }
5605 
5606 /*
5607  * This is the same as pmap_alloc_l2_bucket(), except that it is only
5608  * used by pmap_growkernel().
5609  */
5610 static inline struct l2_bucket *
5611 pmap_grow_l2_bucket(pmap_t pm, vaddr_t va)
5612 {
5613 	struct l2_dtable *l2;
5614 	struct l2_bucket *l2b;
5615 	u_short l1slot;
5616 	vaddr_t nva;
5617 
5618 	l1slot = l1pte_index(va);
5619 
5620 	if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL) {
5621 		/*
5622 		 * No mapping at this address, as there is
5623 		 * no entry in the L1 table.
5624 		 * Need to allocate a new l2_dtable.
5625 		 */
5626 		nva = pmap_kernel_l2dtable_kva;
5627 		if ((nva & PGOFSET) == 0) {
5628 			/*
5629 			 * Need to allocate a backing page
5630 			 */
5631 			if (pmap_grow_map(nva, NULL))
5632 				return (NULL);
5633 		}
5634 
5635 		l2 = (struct l2_dtable *)nva;
5636 		nva += sizeof(struct l2_dtable);
5637 
5638 		if ((nva & PGOFSET) < (pmap_kernel_l2dtable_kva & PGOFSET)) {
5639 			/*
5640 			 * The new l2_dtable straddles a page boundary.
5641 			 * Map in another page to cover it.
5642 			 */
5643 			if (pmap_grow_map(nva, NULL))
5644 				return (NULL);
5645 		}
5646 
5647 		pmap_kernel_l2dtable_kva = nva;
5648 
5649 		/*
5650 		 * Link it into the parent pmap
5651 		 */
5652 		pm->pm_l2[L2_IDX(l1slot)] = l2;
5653 	}
5654 
5655 	l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
5656 
5657 	/*
5658 	 * Fetch pointer to the L2 page table associated with the address.
5659 	 */
5660 	if (l2b->l2b_kva == NULL) {
5661 		pt_entry_t *ptep;
5662 
5663 		/*
5664 		 * No L2 page table has been allocated. Chances are, this
5665 		 * is because we just allocated the l2_dtable, above.
5666 		 */
5667 		nva = pmap_kernel_l2ptp_kva;
5668 		ptep = (pt_entry_t *)nva;
5669 		if ((nva & PGOFSET) == 0) {
5670 			/*
5671 			 * Need to allocate a backing page
5672 			 */
5673 			if (pmap_grow_map(nva, &pmap_kernel_l2ptp_phys))
5674 				return (NULL);
5675 			PTE_SYNC_RANGE(ptep, PAGE_SIZE / sizeof(pt_entry_t));
5676 		}
5677 
5678 		l2->l2_occupancy++;
5679 		l2b->l2b_kva = ptep;
5680 		l2b->l2b_l1slot = l1slot;
5681 		l2b->l2b_pa = pmap_kernel_l2ptp_phys;
5682 
5683 		pmap_kernel_l2ptp_kva += L2_TABLE_SIZE_REAL;
5684 		pmap_kernel_l2ptp_phys += L2_TABLE_SIZE_REAL;
5685 	}
5686 
5687 	return (l2b);
5688 }
5689 
5690 vaddr_t
5691 pmap_growkernel(vaddr_t maxkvaddr)
5692 {
5693 	pmap_t kpm = pmap_kernel();
5694 #ifndef ARM_MMU_EXTENDED
5695 	struct l1_ttable *l1;
5696 #endif
5697 	int s;
5698 
5699 	if (maxkvaddr <= pmap_curmaxkvaddr)
5700 		goto out;		/* we are OK */
5701 
5702 	NPDEBUG(PDB_GROWKERN,
5703 	    printf("pmap_growkernel: growing kernel from 0x%lx to 0x%lx\n",
5704 	    pmap_curmaxkvaddr, maxkvaddr));
5705 
5706 	KDASSERT(maxkvaddr <= virtual_end);
5707 
5708 	/*
5709 	 * whoops!   we need to add kernel PTPs
5710 	 */
5711 
5712 	s = splhigh();	/* to be safe */
5713 	mutex_enter(kpm->pm_lock);
5714 
5715 	/* Map 1MB at a time */
5716 	size_t l1slot = l1pte_index(pmap_curmaxkvaddr);
5717 #ifdef ARM_MMU_EXTENDED
5718 	pd_entry_t * const spdep = &kpm->pm_l1[l1slot];
5719 	pd_entry_t *pdep = spdep;
5720 #endif
5721 	for (;pmap_curmaxkvaddr < maxkvaddr; pmap_curmaxkvaddr += L1_S_SIZE,
5722 #ifdef ARM_MMU_EXTENDED
5723 	     pdep++,
5724 #endif
5725 	     l1slot++) {
5726 		struct l2_bucket *l2b =
5727 		    pmap_grow_l2_bucket(kpm, pmap_curmaxkvaddr);
5728 		KASSERT(l2b != NULL);
5729 
5730 		const pd_entry_t npde = L1_C_PROTO | l2b->l2b_pa
5731 		    | L1_C_DOM(PMAP_DOMAIN_KERNEL);
5732 #ifdef ARM_MMU_EXTENDED
5733 		l1pte_setone(pdep, npde);
5734 #else
5735 		/* Distribute new L1 entry to all other L1s */
5736 		SLIST_FOREACH(l1, &l1_list, l1_link) {
5737 			pd_entry_t * const pdep = &l1->l1_kva[l1slot];
5738 			l1pte_setone(pdep, npde);
5739 			PDE_SYNC(pdep);
5740 		}
5741 #endif
5742 	}
5743 #ifdef ARM_MMU_EXTENDED
5744 	PDE_SYNC_RANGE(spdep, pdep - spdep);
5745 #endif
5746 
5747 #ifdef PMAP_CACHE_VIVT
5748 	/*
5749 	 * flush out the cache, expensive but growkernel will happen so
5750 	 * rarely
5751 	 */
5752 	cpu_dcache_wbinv_all();
5753 	cpu_tlb_flushD();
5754 	cpu_cpwait();
5755 #endif
5756 
5757 	mutex_exit(kpm->pm_lock);
5758 	splx(s);
5759 
5760 out:
5761 	return (pmap_curmaxkvaddr);
5762 }
5763 
5764 /************************ Utility routines ****************************/
5765 
5766 #ifndef ARM_HAS_VBAR
5767 /*
5768  * vector_page_setprot:
5769  *
5770  *	Manipulate the protection of the vector page.
5771  */
5772 void
5773 vector_page_setprot(int prot)
5774 {
5775 	struct l2_bucket *l2b;
5776 	pt_entry_t *ptep;
5777 
5778 #if defined(CPU_ARMV7) || defined(CPU_ARM11)
5779 	/*
5780 	 * If we are using VBAR to use the vectors in the kernel, then it's
5781 	 * already mapped in the kernel text so no need to anything here.
5782 	 */
5783 	if (vector_page != ARM_VECTORS_LOW && vector_page != ARM_VECTORS_HIGH) {
5784 		KASSERT((armreg_pfr1_read() & ARM_PFR1_SEC_MASK) != 0);
5785 		return;
5786 	}
5787 #endif
5788 
5789 	l2b = pmap_get_l2_bucket(pmap_kernel(), vector_page);
5790 	KASSERT(l2b != NULL);
5791 
5792 	ptep = &l2b->l2b_kva[l2pte_index(vector_page)];
5793 
5794 	const pt_entry_t opte = *ptep;
5795 #ifdef ARM_MMU_EXTENDED
5796 	const pt_entry_t npte = (opte & ~(L2_S_PROT_MASK|L2_XS_XN))
5797 	    | L2_S_PROT(PTE_KERNEL, prot);
5798 #else
5799 	const pt_entry_t npte = (opte & ~L2_S_PROT_MASK)
5800 	    | L2_S_PROT(PTE_KERNEL, prot);
5801 #endif
5802 	l2pte_set(ptep, npte, opte);
5803 	PTE_SYNC(ptep);
5804 	cpu_tlb_flushD_SE(vector_page);
5805 	cpu_cpwait();
5806 }
5807 #endif
5808 
5809 /*
5810  * Fetch pointers to the PDE/PTE for the given pmap/VA pair.
5811  * Returns true if the mapping exists, else false.
5812  *
5813  * NOTE: This function is only used by a couple of arm-specific modules.
5814  * It is not safe to take any pmap locks here, since we could be right
5815  * in the middle of debugging the pmap anyway...
5816  *
5817  * It is possible for this routine to return false even though a valid
5818  * mapping does exist. This is because we don't lock, so the metadata
5819  * state may be inconsistent.
5820  *
5821  * NOTE: We can return a NULL *ptp in the case where the L1 pde is
5822  * a "section" mapping.
5823  */
5824 bool
5825 pmap_get_pde_pte(pmap_t pm, vaddr_t va, pd_entry_t **pdp, pt_entry_t **ptp)
5826 {
5827 	struct l2_dtable *l2;
5828 	pd_entry_t *pdep, pde;
5829 	pt_entry_t *ptep;
5830 	u_short l1slot;
5831 
5832 	if (pm->pm_l1 == NULL)
5833 		return false;
5834 
5835 	l1slot = l1pte_index(va);
5836 	*pdp = pdep = pmap_l1_kva(pm) + l1slot;
5837 	pde = *pdep;
5838 
5839 	if (l1pte_section_p(pde)) {
5840 		*ptp = NULL;
5841 		return true;
5842 	}
5843 
5844 	l2 = pm->pm_l2[L2_IDX(l1slot)];
5845 	if (l2 == NULL ||
5846 	    (ptep = l2->l2_bucket[L2_BUCKET(l1slot)].l2b_kva) == NULL) {
5847 		return false;
5848 	}
5849 
5850 	*ptp = &ptep[l2pte_index(va)];
5851 	return true;
5852 }
5853 
5854 bool
5855 pmap_get_pde(pmap_t pm, vaddr_t va, pd_entry_t **pdp)
5856 {
5857 
5858 	if (pm->pm_l1 == NULL)
5859 		return false;
5860 
5861 	*pdp = pmap_l1_kva(pm) + l1pte_index(va);
5862 
5863 	return true;
5864 }
5865 
5866 /************************ Bootstrapping routines ****************************/
5867 
5868 #ifndef ARM_MMU_EXTENDED
5869 static void
5870 pmap_init_l1(struct l1_ttable *l1, pd_entry_t *l1pt)
5871 {
5872 	int i;
5873 
5874 	l1->l1_kva = l1pt;
5875 	l1->l1_domain_use_count = 0;
5876 	l1->l1_domain_first = 0;
5877 
5878 	for (i = 0; i < PMAP_DOMAINS; i++)
5879 		l1->l1_domain_free[i] = i + 1;
5880 
5881 	/*
5882 	 * Copy the kernel's L1 entries to each new L1.
5883 	 */
5884 	if (pmap_initialized)
5885 		memcpy(l1pt, pmap_l1_kva(pmap_kernel()), L1_TABLE_SIZE);
5886 
5887 	if (pmap_extract(pmap_kernel(), (vaddr_t)l1pt,
5888 	    &l1->l1_physaddr) == false)
5889 		panic("pmap_init_l1: can't get PA of L1 at %p", l1pt);
5890 
5891 	SLIST_INSERT_HEAD(&l1_list, l1, l1_link);
5892 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
5893 }
5894 #endif /* !ARM_MMU_EXTENDED */
5895 
5896 /*
5897  * pmap_bootstrap() is called from the board-specific initarm() routine
5898  * once the kernel L1/L2 descriptors tables have been set up.
5899  *
5900  * This is a somewhat convoluted process since pmap bootstrap is, effectively,
5901  * spread over a number of disparate files/functions.
5902  *
5903  * We are passed the following parameters
5904  *  - kernel_l1pt
5905  *    This is a pointer to the base of the kernel's L1 translation table.
5906  *  - vstart
5907  *    1MB-aligned start of managed kernel virtual memory.
5908  *  - vend
5909  *    1MB-aligned end of managed kernel virtual memory.
5910  *
5911  * We use the first parameter to build the metadata (struct l1_ttable and
5912  * struct l2_dtable) necessary to track kernel mappings.
5913  */
5914 #define	PMAP_STATIC_L2_SIZE 16
5915 void
5916 pmap_bootstrap(vaddr_t vstart, vaddr_t vend)
5917 {
5918 	static struct l2_dtable static_l2[PMAP_STATIC_L2_SIZE];
5919 #ifndef ARM_MMU_EXTENDED
5920 	static struct l1_ttable static_l1;
5921 	struct l1_ttable *l1 = &static_l1;
5922 #endif
5923 	struct l2_dtable *l2;
5924 	struct l2_bucket *l2b;
5925 	pd_entry_t *l1pt = (pd_entry_t *) kernel_l1pt.pv_va;
5926 	pmap_t pm = pmap_kernel();
5927 	pt_entry_t *ptep;
5928 	paddr_t pa;
5929 	vsize_t size;
5930 	int nptes, l2idx, l2next = 0;
5931 
5932 #ifdef ARM_MMU_EXTENDED
5933 	KASSERT(pte_l1_s_cache_mode == pte_l1_s_cache_mode_pt);
5934 	KASSERT(pte_l2_s_cache_mode == pte_l2_s_cache_mode_pt);
5935 #endif
5936 
5937 #ifdef VERBOSE_INIT_ARM
5938 	printf("kpm ");
5939 #endif
5940 	/*
5941 	 * Initialise the kernel pmap object
5942 	 */
5943 	curcpu()->ci_pmap_cur = pm;
5944 #ifdef ARM_MMU_EXTENDED
5945 	pm->pm_l1 = l1pt;
5946 	pm->pm_l1_pa = kernel_l1pt.pv_pa;
5947 #ifdef VERBOSE_INIT_ARM
5948 	printf("tlb0 ");
5949 #endif
5950 	pmap_tlb_info_init(&pmap_tlb0_info);
5951 #ifdef MULTIPROCESSOR
5952 #ifdef VERBOSE_INIT_ARM
5953 	printf("kcpusets ");
5954 #endif
5955 	pm->pm_onproc = kcpuset_running;
5956 	pm->pm_active = kcpuset_running;
5957 #endif
5958 #else
5959 	pm->pm_l1 = l1;
5960 #endif
5961 
5962 #ifdef VERBOSE_INIT_ARM
5963 	printf("locks ");
5964 #endif
5965 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
5966 	if (arm_cache_prefer_mask != 0) {
5967 		mutex_init(&pmap_lock, MUTEX_DEFAULT, IPL_VM);
5968 	} else {
5969 #endif
5970 		mutex_init(&pmap_lock, MUTEX_DEFAULT, IPL_NONE);
5971 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
5972 	}
5973 #endif
5974 	mutex_init(&pm->pm_obj_lock, MUTEX_DEFAULT, IPL_NONE);
5975 	uvm_obj_init(&pm->pm_obj, NULL, false, 1);
5976 	uvm_obj_setlock(&pm->pm_obj, &pm->pm_obj_lock);
5977 
5978 #ifdef VERBOSE_INIT_ARM
5979 	printf("l1pt ");
5980 #endif
5981 	/*
5982 	 * Scan the L1 translation table created by initarm() and create
5983 	 * the required metadata for all valid mappings found in it.
5984 	 */
5985 	for (size_t l1slot = 0;
5986 	     l1slot < L1_TABLE_SIZE / sizeof(pd_entry_t);
5987 	     l1slot++) {
5988 		pd_entry_t pde = l1pt[l1slot];
5989 
5990 		/*
5991 		 * We're only interested in Coarse mappings.
5992 		 * pmap_extract() can deal with section mappings without
5993 		 * recourse to checking L2 metadata.
5994 		 */
5995 		if ((pde & L1_TYPE_MASK) != L1_TYPE_C)
5996 			continue;
5997 
5998 		/*
5999 		 * Lookup the KVA of this L2 descriptor table
6000 		 */
6001 		pa = l1pte_pa(pde);
6002 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
6003 		if (ptep == NULL) {
6004 			panic("pmap_bootstrap: No L2 for va 0x%x, pa 0x%lx",
6005 			    (u_int)l1slot << L1_S_SHIFT, pa);
6006 		}
6007 
6008 		/*
6009 		 * Fetch the associated L2 metadata structure.
6010 		 * Allocate a new one if necessary.
6011 		 */
6012 		if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL) {
6013 			if (l2next == PMAP_STATIC_L2_SIZE)
6014 				panic("pmap_bootstrap: out of static L2s");
6015 			pm->pm_l2[L2_IDX(l1slot)] = l2 = &static_l2[l2next++];
6016 		}
6017 
6018 		/*
6019 		 * One more L1 slot tracked...
6020 		 */
6021 		l2->l2_occupancy++;
6022 
6023 		/*
6024 		 * Fill in the details of the L2 descriptor in the
6025 		 * appropriate bucket.
6026 		 */
6027 		l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
6028 		l2b->l2b_kva = ptep;
6029 		l2b->l2b_pa = pa;
6030 		l2b->l2b_l1slot = l1slot;
6031 
6032 		/*
6033 		 * Establish an initial occupancy count for this descriptor
6034 		 */
6035 		for (l2idx = 0;
6036 		    l2idx < (L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
6037 		    l2idx++) {
6038 			if ((ptep[l2idx] & L2_TYPE_MASK) != L2_TYPE_INV) {
6039 				l2b->l2b_occupancy++;
6040 			}
6041 		}
6042 
6043 		/*
6044 		 * Make sure the descriptor itself has the correct cache mode.
6045 		 * If not, fix it, but whine about the problem. Port-meisters
6046 		 * should consider this a clue to fix up their initarm()
6047 		 * function. :)
6048 		 */
6049 		if (pmap_set_pt_cache_mode(l1pt, (vaddr_t)ptep, 1)) {
6050 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
6051 			    "L2 pte @ %p\n", ptep);
6052 		}
6053 	}
6054 
6055 #ifdef VERBOSE_INIT_ARM
6056 	printf("cache(l1pt) ");
6057 #endif
6058 	/*
6059 	 * Ensure the primary (kernel) L1 has the correct cache mode for
6060 	 * a page table. Bitch if it is not correctly set.
6061 	 */
6062 	if (pmap_set_pt_cache_mode(l1pt, kernel_l1pt.pv_va,
6063 		    L1_TABLE_SIZE / L2_S_SIZE)) {
6064 		printf("pmap_bootstrap: WARNING! wrong cache mode for "
6065 		    "primary L1 @ 0x%lx\n", kernel_l1pt.pv_va);
6066 	}
6067 
6068 #ifdef PMAP_CACHE_VIVT
6069 	cpu_dcache_wbinv_all();
6070 	cpu_tlb_flushID();
6071 	cpu_cpwait();
6072 #endif
6073 
6074 	/*
6075 	 * now we allocate the "special" VAs which are used for tmp mappings
6076 	 * by the pmap (and other modules).  we allocate the VAs by advancing
6077 	 * virtual_avail (note that there are no pages mapped at these VAs).
6078 	 *
6079 	 * Managed KVM space start from wherever initarm() tells us.
6080 	 */
6081 	virtual_avail = vstart;
6082 	virtual_end = vend;
6083 
6084 #ifdef VERBOSE_INIT_ARM
6085 	printf("specials ");
6086 #endif
6087 #ifdef PMAP_CACHE_VIPT
6088 	/*
6089 	 * If we have a VIPT cache, we need one page/pte per possible alias
6090 	 * page so we won't violate cache aliasing rules.
6091 	 */
6092 	virtual_avail = (virtual_avail + arm_cache_prefer_mask) & ~arm_cache_prefer_mask;
6093 	nptes = (arm_cache_prefer_mask >> L2_S_SHIFT) + 1;
6094 	if (arm_pcache.icache_type != CACHE_TYPE_PIPT
6095 	    && arm_pcache.icache_way_size > nptes * L2_S_SIZE) {
6096 		nptes = arm_pcache.icache_way_size >> L2_S_SHIFT;
6097 	}
6098 #else
6099 	nptes = PAGE_SIZE / L2_S_SIZE;
6100 #endif
6101 #ifdef MULTIPROCESSOR
6102 	cnptes = nptes;
6103 	nptes *= arm_cpu_max;
6104 #endif
6105 	pmap_alloc_specials(&virtual_avail, nptes, &csrcp, &csrc_pte);
6106 	pmap_set_pt_cache_mode(l1pt, (vaddr_t)csrc_pte, nptes);
6107 	pmap_alloc_specials(&virtual_avail, nptes, &cdstp, &cdst_pte);
6108 	pmap_set_pt_cache_mode(l1pt, (vaddr_t)cdst_pte, nptes);
6109 	pmap_alloc_specials(&virtual_avail, nptes, &memhook, NULL);
6110 	if (msgbufaddr == NULL) {
6111 		pmap_alloc_specials(&virtual_avail,
6112 		    round_page(MSGBUFSIZE) / PAGE_SIZE,
6113 		    (void *)&msgbufaddr, NULL);
6114 	}
6115 
6116 	/*
6117 	 * Allocate a range of kernel virtual address space to be used
6118 	 * for L2 descriptor tables and metadata allocation in
6119 	 * pmap_growkernel().
6120 	 */
6121 	size = ((virtual_end - pmap_curmaxkvaddr) + L1_S_OFFSET) / L1_S_SIZE;
6122 	pmap_alloc_specials(&virtual_avail,
6123 	    round_page(size * L2_TABLE_SIZE_REAL) / PAGE_SIZE,
6124 	    &pmap_kernel_l2ptp_kva, NULL);
6125 
6126 	size = (size + (L2_BUCKET_SIZE - 1)) / L2_BUCKET_SIZE;
6127 	pmap_alloc_specials(&virtual_avail,
6128 	    round_page(size * sizeof(struct l2_dtable)) / PAGE_SIZE,
6129 	    &pmap_kernel_l2dtable_kva, NULL);
6130 
6131 #ifndef ARM_MMU_EXTENDED
6132 	/*
6133 	 * init the static-global locks and global pmap list.
6134 	 */
6135 	mutex_init(&l1_lru_lock, MUTEX_DEFAULT, IPL_VM);
6136 
6137 	/*
6138 	 * We can now initialise the first L1's metadata.
6139 	 */
6140 	SLIST_INIT(&l1_list);
6141 	TAILQ_INIT(&l1_lru_list);
6142 	pmap_init_l1(l1, l1pt);
6143 #endif /* ARM_MMU_EXTENDED */
6144 
6145 #ifndef ARM_HAS_VBAR
6146 	/* Set up vector page L1 details, if necessary */
6147 	if (vector_page < KERNEL_BASE) {
6148 		pm->pm_pl1vec = pmap_l1_kva(pm) + l1pte_index(vector_page);
6149 		l2b = pmap_get_l2_bucket(pm, vector_page);
6150 		KDASSERT(l2b != NULL);
6151 		pm->pm_l1vec = l2b->l2b_pa | L1_C_PROTO |
6152 		    L1_C_DOM(pmap_domain(pm));
6153 	} else
6154 		pm->pm_pl1vec = NULL;
6155 #endif
6156 
6157 #ifdef VERBOSE_INIT_ARM
6158 	printf("pools ");
6159 #endif
6160 	/*
6161 	 * Initialize the pmap cache
6162 	 */
6163 	pool_cache_bootstrap(&pmap_cache, sizeof(struct pmap), 0, 0, 0,
6164 	    "pmappl", NULL, IPL_NONE, pmap_pmap_ctor, NULL, NULL);
6165 	LIST_INIT(&pmap_pmaps);
6166 	LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list);
6167 
6168 	/*
6169 	 * Initialize the pv pool.
6170 	 */
6171 	pool_init(&pmap_pv_pool, sizeof(struct pv_entry), 0, 0, 0, "pvepl",
6172 	    &pmap_bootstrap_pv_allocator, IPL_NONE);
6173 
6174 	/*
6175 	 * Initialize the L2 dtable pool and cache.
6176 	 */
6177 	pool_cache_bootstrap(&pmap_l2dtable_cache, sizeof(struct l2_dtable), 0,
6178 	    0, 0, "l2dtblpl", NULL, IPL_NONE, pmap_l2dtable_ctor, NULL, NULL);
6179 
6180 	/*
6181 	 * Initialise the L2 descriptor table pool and cache
6182 	 */
6183 	pool_cache_bootstrap(&pmap_l2ptp_cache, L2_TABLE_SIZE_REAL, 0,
6184 	    L2_TABLE_SIZE_REAL, 0, "l2ptppl", NULL, IPL_NONE,
6185 	    pmap_l2ptp_ctor, NULL, NULL);
6186 
6187 	mutex_init(&memlock, MUTEX_DEFAULT, IPL_NONE);
6188 
6189 	cpu_dcache_wbinv_all();
6190 }
6191 
6192 static bool
6193 pmap_set_pt_cache_mode(pd_entry_t *kl1, vaddr_t va, size_t nptes)
6194 {
6195 #ifdef ARM_MMU_EXTENDED
6196 	return false;
6197 #else
6198 	if (pte_l1_s_cache_mode == pte_l1_s_cache_mode_pt
6199 	    && pte_l2_s_cache_mode == pte_l2_s_cache_mode_pt)
6200 		return false;
6201 
6202 	const vaddr_t eva = va + nptes * PAGE_SIZE;
6203 	int rv = 0;
6204 
6205 	while (va < eva) {
6206 		/*
6207 		 * Make sure the descriptor itself has the correct cache mode
6208 		 */
6209 		pd_entry_t * const pdep = &kl1[l1pte_index(va)];
6210 		pd_entry_t pde = *pdep;
6211 
6212 		if (l1pte_section_p(pde)) {
6213 			__CTASSERT((L1_S_CACHE_MASK & L1_S_V6_SUPER) == 0);
6214 			if ((pde & L1_S_CACHE_MASK) != pte_l1_s_cache_mode_pt) {
6215 				*pdep = (pde & ~L1_S_CACHE_MASK) |
6216 				    pte_l1_s_cache_mode_pt;
6217 				PDE_SYNC(pdep);
6218 				cpu_dcache_wbinv_range((vaddr_t)pdep,
6219 				    sizeof(*pdep));
6220 				rv = 1;
6221 			}
6222 			return rv;
6223 			if (pde & L1_S_V6_SUPER) {
6224 				va = (va & -L1_SS_SIZE) + L1_SS_SIZE;
6225 			} else {
6226 				va = (va & -L1_S_SIZE) + L1_S_SIZE;
6227 			}
6228 			continue;
6229 		}
6230 		vaddr_t pa = l1pte_pa(pde);
6231 		pt_entry_t *ptep = (pt_entry_t *)kernel_pt_lookup(pa);
6232 		if (ptep == NULL)
6233 			panic("pmap_bootstrap: No PTP for va %#lx\n", va);
6234 
6235 		ptep += l2pte_index(va);
6236 		const pt_entry_t opte = *ptep;
6237 		if ((opte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
6238 			const pt_entry_t npte = (opte & ~L2_S_CACHE_MASK)
6239 			    | pte_l2_s_cache_mode_pt;
6240 			l2pte_set(ptep, npte, opte);
6241 			PTE_SYNC(ptep);
6242 			cpu_dcache_wbinv_range((vaddr_t)ptep, sizeof(*ptep));
6243 			rv = 1;
6244 		}
6245 		va += PAGE_SIZE;
6246 	}
6247 
6248 	return (rv);
6249 #endif
6250 }
6251 
6252 static void
6253 pmap_alloc_specials(vaddr_t *availp, int pages, vaddr_t *vap, pt_entry_t **ptep)
6254 {
6255 	vaddr_t va = *availp;
6256 	struct l2_bucket *l2b;
6257 
6258 	if (ptep) {
6259 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
6260 		if (l2b == NULL)
6261 			panic("pmap_alloc_specials: no l2b for 0x%lx", va);
6262 
6263 		if (ptep)
6264 			*ptep = &l2b->l2b_kva[l2pte_index(va)];
6265 	}
6266 
6267 	*vap = va;
6268 	*availp = va + (PAGE_SIZE * pages);
6269 }
6270 
6271 void
6272 pmap_init(void)
6273 {
6274 
6275 	/*
6276 	 * Set the available memory vars - These do not map to real memory
6277 	 * addresses and cannot as the physical memory is fragmented.
6278 	 * They are used by ps for %mem calculations.
6279 	 * One could argue whether this should be the entire memory or just
6280 	 * the memory that is useable in a user process.
6281 	 */
6282 	avail_start = ptoa(VM_PHYSMEM_PTR(0)->start);
6283 	avail_end = ptoa(VM_PHYSMEM_PTR(vm_nphysseg - 1)->end);
6284 
6285 	/*
6286 	 * Now we need to free enough pv_entry structures to allow us to get
6287 	 * the kmem_map/kmem_object allocated and inited (done after this
6288 	 * function is finished).  to do this we allocate one bootstrap page out
6289 	 * of kernel_map and use it to provide an initial pool of pv_entry
6290 	 * structures.   we never free this page.
6291 	 */
6292 	pool_setlowat(&pmap_pv_pool, (PAGE_SIZE / sizeof(struct pv_entry)) * 2);
6293 
6294 #ifdef ARM_MMU_EXTENDED
6295 	pmap_tlb_info_evcnt_attach(&pmap_tlb0_info);
6296 #endif
6297 
6298 	pmap_initialized = true;
6299 }
6300 
6301 static vaddr_t last_bootstrap_page = 0;
6302 static void *free_bootstrap_pages = NULL;
6303 
6304 static void *
6305 pmap_bootstrap_pv_page_alloc(struct pool *pp, int flags)
6306 {
6307 	extern void *pool_page_alloc(struct pool *, int);
6308 	vaddr_t new_page;
6309 	void *rv;
6310 
6311 	if (pmap_initialized)
6312 		return (pool_page_alloc(pp, flags));
6313 
6314 	if (free_bootstrap_pages) {
6315 		rv = free_bootstrap_pages;
6316 		free_bootstrap_pages = *((void **)rv);
6317 		return (rv);
6318 	}
6319 
6320 	KASSERT(kernel_map != NULL);
6321 	new_page = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
6322 	    UVM_KMF_WIRED | ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT));
6323 
6324 	KASSERT(new_page > last_bootstrap_page);
6325 	last_bootstrap_page = new_page;
6326 	return ((void *)new_page);
6327 }
6328 
6329 static void
6330 pmap_bootstrap_pv_page_free(struct pool *pp, void *v)
6331 {
6332 	extern void pool_page_free(struct pool *, void *);
6333 
6334 	if ((vaddr_t)v <= last_bootstrap_page) {
6335 		*((void **)v) = free_bootstrap_pages;
6336 		free_bootstrap_pages = v;
6337 		return;
6338 	}
6339 
6340 	if (pmap_initialized) {
6341 		pool_page_free(pp, v);
6342 		return;
6343 	}
6344 }
6345 
6346 /*
6347  * pmap_postinit()
6348  *
6349  * This routine is called after the vm and kmem subsystems have been
6350  * initialised. This allows the pmap code to perform any initialisation
6351  * that can only be done one the memory allocation is in place.
6352  */
6353 void
6354 pmap_postinit(void)
6355 {
6356 #ifndef ARM_MMU_EXTENDED
6357 	extern paddr_t physical_start, physical_end;
6358 	struct l1_ttable *l1;
6359 	struct pglist plist;
6360 	struct vm_page *m;
6361 	pd_entry_t *pdep;
6362 	vaddr_t va, eva;
6363 	u_int loop, needed;
6364 	int error;
6365 #endif
6366 
6367 	pool_cache_setlowat(&pmap_l2ptp_cache, (PAGE_SIZE / L2_TABLE_SIZE_REAL) * 4);
6368 	pool_cache_setlowat(&pmap_l2dtable_cache,
6369 	    (PAGE_SIZE / sizeof(struct l2_dtable)) * 2);
6370 
6371 #ifndef ARM_MMU_EXTENDED
6372 	needed = (maxproc / PMAP_DOMAINS) + ((maxproc % PMAP_DOMAINS) ? 1 : 0);
6373 	needed -= 1;
6374 
6375 	l1 = kmem_alloc(sizeof(*l1) * needed, KM_SLEEP);
6376 
6377 	for (loop = 0; loop < needed; loop++, l1++) {
6378 		/* Allocate a L1 page table */
6379 		va = uvm_km_alloc(kernel_map, L1_TABLE_SIZE, 0, UVM_KMF_VAONLY);
6380 		if (va == 0)
6381 			panic("Cannot allocate L1 KVM");
6382 
6383 		error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start,
6384 		    physical_end, L1_TABLE_SIZE, 0, &plist, 1, 1);
6385 		if (error)
6386 			panic("Cannot allocate L1 physical pages");
6387 
6388 		m = TAILQ_FIRST(&plist);
6389 		eva = va + L1_TABLE_SIZE;
6390 		pdep = (pd_entry_t *)va;
6391 
6392 		while (m && va < eva) {
6393 			paddr_t pa = VM_PAGE_TO_PHYS(m);
6394 
6395 			pmap_kenter_pa(va, pa,
6396 			    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE|PMAP_PTE);
6397 
6398 			va += PAGE_SIZE;
6399 			m = TAILQ_NEXT(m, pageq.queue);
6400 		}
6401 
6402 #ifdef DIAGNOSTIC
6403 		if (m)
6404 			panic("pmap_alloc_l1pt: pglist not empty");
6405 #endif	/* DIAGNOSTIC */
6406 
6407 		pmap_init_l1(l1, pdep);
6408 	}
6409 
6410 #ifdef DEBUG
6411 	printf("pmap_postinit: Allocated %d static L1 descriptor tables\n",
6412 	    needed);
6413 #endif
6414 #endif /* !ARM_MMU_EXTENDED */
6415 }
6416 
6417 /*
6418  * Note that the following routines are used by board-specific initialisation
6419  * code to configure the initial kernel page tables.
6420  *
6421  * If ARM32_NEW_VM_LAYOUT is *not* defined, they operate on the assumption that
6422  * L2 page-table pages are 4KB in size and use 4 L1 slots. This mimics the
6423  * behaviour of the old pmap, and provides an easy migration path for
6424  * initial bring-up of the new pmap on existing ports. Fortunately,
6425  * pmap_bootstrap() compensates for this hackery. This is only a stop-gap and
6426  * will be deprecated.
6427  *
6428  * If ARM32_NEW_VM_LAYOUT *is* defined, these functions deal with 1KB L2 page
6429  * tables.
6430  */
6431 
6432 /*
6433  * This list exists for the benefit of pmap_map_chunk().  It keeps track
6434  * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
6435  * find them as necessary.
6436  *
6437  * Note that the data on this list MUST remain valid after initarm() returns,
6438  * as pmap_bootstrap() uses it to contruct L2 table metadata.
6439  */
6440 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
6441 
6442 static vaddr_t
6443 kernel_pt_lookup(paddr_t pa)
6444 {
6445 	pv_addr_t *pv;
6446 
6447 	SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
6448 		if (pv->pv_pa == (pa & ~PGOFSET))
6449 			return (pv->pv_va | (pa & PGOFSET));
6450 	}
6451 	return (0);
6452 }
6453 
6454 /*
6455  * pmap_map_section:
6456  *
6457  *	Create a single section mapping.
6458  */
6459 void
6460 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
6461 {
6462 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
6463 	const size_t l1slot = l1pte_index(va);
6464 	pd_entry_t fl;
6465 
6466 	KASSERT(((va | pa) & L1_S_OFFSET) == 0);
6467 
6468 	switch (cache) {
6469 	case PTE_NOCACHE:
6470 	default:
6471 		fl = 0;
6472 		break;
6473 
6474 	case PTE_CACHE:
6475 		fl = pte_l1_s_cache_mode;
6476 		break;
6477 
6478 	case PTE_PAGETABLE:
6479 		fl = pte_l1_s_cache_mode_pt;
6480 		break;
6481 	}
6482 
6483 	const pd_entry_t npde = L1_S_PROTO | pa |
6484 	    L1_S_PROT(PTE_KERNEL, prot) | fl | L1_S_DOM(PMAP_DOMAIN_KERNEL);
6485 	l1pte_setone(pdep + l1slot, npde);
6486 	PDE_SYNC(pdep + l1slot);
6487 }
6488 
6489 /*
6490  * pmap_map_entry:
6491  *
6492  *	Create a single page mapping.
6493  */
6494 void
6495 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
6496 {
6497 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
6498 	const size_t l1slot = l1pte_index(va);
6499 	pt_entry_t npte;
6500 	pt_entry_t *ptep;
6501 
6502 	KASSERT(((va | pa) & PGOFSET) == 0);
6503 
6504 	switch (cache) {
6505 	case PTE_NOCACHE:
6506 	default:
6507 		npte = 0;
6508 		break;
6509 
6510 	case PTE_CACHE:
6511 		npte = pte_l2_s_cache_mode;
6512 		break;
6513 
6514 	case PTE_PAGETABLE:
6515 		npte = pte_l2_s_cache_mode_pt;
6516 		break;
6517 	}
6518 
6519 	if ((pdep[l1slot] & L1_TYPE_MASK) != L1_TYPE_C)
6520 		panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
6521 
6522 	ptep = (pt_entry_t *) kernel_pt_lookup(l1pte_pa(pdep[l1slot]));
6523 	if (ptep == NULL)
6524 		panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
6525 
6526 	npte |= L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot);
6527 #ifdef ARM_MMU_EXTENDED
6528 	if (prot & VM_PROT_EXECUTE) {
6529 		npte &= ~L2_XS_XN;
6530 	}
6531 #endif
6532 	ptep += l2pte_index(va);
6533 	l2pte_set(ptep, npte, 0);
6534 	PTE_SYNC(ptep);
6535 }
6536 
6537 /*
6538  * pmap_link_l2pt:
6539  *
6540  *	Link the L2 page table specified by "l2pv" into the L1
6541  *	page table at the slot for "va".
6542  */
6543 void
6544 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
6545 {
6546 	pd_entry_t * const pdep = (pd_entry_t *) l1pt + l1pte_index(va);
6547 
6548 	KASSERT((va & ((L1_S_SIZE * (PAGE_SIZE / L2_T_SIZE)) - 1)) == 0);
6549 	KASSERT((l2pv->pv_pa & PGOFSET) == 0);
6550 
6551 	const pd_entry_t npde = L1_S_DOM(PMAP_DOMAIN_KERNEL) | L1_C_PROTO
6552 	    | l2pv->pv_pa;
6553 
6554 	l1pte_set(pdep, npde);
6555 	PDE_SYNC_RANGE(pdep, PAGE_SIZE / L2_T_SIZE);
6556 
6557 	SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
6558 }
6559 
6560 /*
6561  * pmap_map_chunk:
6562  *
6563  *	Map a chunk of memory using the most efficient mappings
6564  *	possible (section, large page, small page) into the
6565  *	provided L1 and L2 tables at the specified virtual address.
6566  */
6567 vsize_t
6568 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
6569     int prot, int cache)
6570 {
6571 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
6572 	pt_entry_t f1, f2s, f2l;
6573 	vsize_t resid;
6574 
6575 	resid = (size + (PAGE_SIZE - 1)) & ~(PAGE_SIZE - 1);
6576 
6577 	if (l1pt == 0)
6578 		panic("pmap_map_chunk: no L1 table provided");
6579 
6580 #ifdef VERBOSE_INIT_ARM
6581 	printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
6582 	    "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
6583 #endif
6584 
6585 	switch (cache) {
6586 	case PTE_NOCACHE:
6587 	default:
6588 		f1 = 0;
6589 		f2l = 0;
6590 		f2s = 0;
6591 		break;
6592 
6593 	case PTE_CACHE:
6594 		f1 = pte_l1_s_cache_mode;
6595 		f2l = pte_l2_l_cache_mode;
6596 		f2s = pte_l2_s_cache_mode;
6597 		break;
6598 
6599 	case PTE_PAGETABLE:
6600 		f1 = pte_l1_s_cache_mode_pt;
6601 		f2l = pte_l2_l_cache_mode_pt;
6602 		f2s = pte_l2_s_cache_mode_pt;
6603 		break;
6604 	}
6605 
6606 	size = resid;
6607 
6608 	while (resid > 0) {
6609 		const size_t l1slot = l1pte_index(va);
6610 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
6611 		/* See if we can use a supersection mapping. */
6612 		if (L1_SS_PROTO && L1_SS_MAPPABLE_P(va, pa, resid)) {
6613 			/* Supersection are always domain 0 */
6614 			const pd_entry_t npde = L1_SS_PROTO | pa
6615 #ifdef ARM_MMU_EXTENDED_XXX
6616 			    | ((prot & VM_PROT_EXECUTE) ? 0 : L1_S_V6_XN)
6617 #endif
6618 #ifdef ARM_MMU_EXTENDED
6619 			    | (va & 0x80000000 ? 0 : L1_S_V6_nG)
6620 #endif
6621 			    | L1_S_PROT(PTE_KERNEL, prot) | f1;
6622 #ifdef VERBOSE_INIT_ARM
6623 			printf("sS");
6624 #endif
6625 			l1pte_set(&pdep[l1slot], npde);
6626 			PDE_SYNC_RANGE(&pdep[l1slot], L1_SS_SIZE / L1_S_SIZE);
6627 			va += L1_SS_SIZE;
6628 			pa += L1_SS_SIZE;
6629 			resid -= L1_SS_SIZE;
6630 			continue;
6631 		}
6632 #endif
6633 		/* See if we can use a section mapping. */
6634 		if (L1_S_MAPPABLE_P(va, pa, resid)) {
6635 			const pd_entry_t npde = L1_S_PROTO | pa
6636 #ifdef ARM_MMU_EXTENDED_XXX
6637 			    | ((prot & VM_PROT_EXECUTE) ? 0 : L1_S_V6_XN)
6638 #endif
6639 #ifdef ARM_MMU_EXTENDED
6640 			    | (va & 0x80000000 ? 0 : L1_S_V6_nG)
6641 #endif
6642 			    | L1_S_PROT(PTE_KERNEL, prot) | f1
6643 			    | L1_S_DOM(PMAP_DOMAIN_KERNEL);
6644 #ifdef VERBOSE_INIT_ARM
6645 			printf("S");
6646 #endif
6647 			l1pte_set(&pdep[l1slot], npde);
6648 			PDE_SYNC(&pdep[l1slot]);
6649 			va += L1_S_SIZE;
6650 			pa += L1_S_SIZE;
6651 			resid -= L1_S_SIZE;
6652 			continue;
6653 		}
6654 
6655 		/*
6656 		 * Ok, we're going to use an L2 table.  Make sure
6657 		 * one is actually in the corresponding L1 slot
6658 		 * for the current VA.
6659 		 */
6660 		if ((pdep[l1slot] & L1_TYPE_MASK) != L1_TYPE_C)
6661 			panic("%s: no L2 table for VA %#lx", __func__, va);
6662 
6663 		pt_entry_t *ptep = (pt_entry_t *) kernel_pt_lookup(l1pte_pa(pdep[l1slot]));
6664 		if (ptep == NULL)
6665 			panic("%s: can't find L2 table for VA %#lx", __func__,
6666 			    va);
6667 
6668 		ptep += l2pte_index(va);
6669 
6670 		/* See if we can use a L2 large page mapping. */
6671 		if (L2_L_MAPPABLE_P(va, pa, resid)) {
6672 			const pt_entry_t npte = L2_L_PROTO | pa
6673 #ifdef ARM_MMU_EXTENDED_XXX
6674 			    | ((prot & VM_PROT_EXECUTE) ? 0 : L2_XS_L_XN)
6675 #endif
6676 #ifdef ARM_MMU_EXTENDED
6677 			    | (va & 0x80000000 ? 0 : L2_XS_nG)
6678 #endif
6679 			    | L2_L_PROT(PTE_KERNEL, prot) | f2l;
6680 #ifdef VERBOSE_INIT_ARM
6681 			printf("L");
6682 #endif
6683 			l2pte_set(ptep, npte, 0);
6684 			PTE_SYNC_RANGE(ptep, L2_L_SIZE / L2_S_SIZE);
6685 			va += L2_L_SIZE;
6686 			pa += L2_L_SIZE;
6687 			resid -= L2_L_SIZE;
6688 			continue;
6689 		}
6690 
6691 		/* Use a small page mapping. */
6692 #ifdef VERBOSE_INIT_ARM
6693 		printf("P");
6694 #endif
6695 		const pt_entry_t npte = L2_S_PROTO | pa
6696 #ifdef ARM_MMU_EXTENDED_XXX
6697 		    | ((prot & VM_PROT_EXECUTE) ? 0 : L2_XS_XN)
6698 #endif
6699 #ifdef ARM_MMU_EXTENDED
6700 		    | (va & 0x80000000 ? 0 : L2_XS_nG)
6701 #endif
6702 		    | L2_S_PROT(PTE_KERNEL, prot) | f2s;
6703 		l2pte_set(ptep, npte, 0);
6704 		PTE_SYNC(ptep);
6705 		va += PAGE_SIZE;
6706 		pa += PAGE_SIZE;
6707 		resid -= PAGE_SIZE;
6708 	}
6709 #ifdef VERBOSE_INIT_ARM
6710 	printf("\n");
6711 #endif
6712 	return (size);
6713 }
6714 
6715 /********************** Static device map routines ***************************/
6716 
6717 static const struct pmap_devmap *pmap_devmap_table;
6718 
6719 /*
6720  * Register the devmap table.  This is provided in case early console
6721  * initialization needs to register mappings created by bootstrap code
6722  * before pmap_devmap_bootstrap() is called.
6723  */
6724 void
6725 pmap_devmap_register(const struct pmap_devmap *table)
6726 {
6727 
6728 	pmap_devmap_table = table;
6729 }
6730 
6731 /*
6732  * Map all of the static regions in the devmap table, and remember
6733  * the devmap table so other parts of the kernel can look up entries
6734  * later.
6735  */
6736 void
6737 pmap_devmap_bootstrap(vaddr_t l1pt, const struct pmap_devmap *table)
6738 {
6739 	int i;
6740 
6741 	pmap_devmap_table = table;
6742 
6743 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
6744 #ifdef VERBOSE_INIT_ARM
6745 		printf("devmap: %08lx -> %08lx @ %08lx\n",
6746 		    pmap_devmap_table[i].pd_pa,
6747 		    pmap_devmap_table[i].pd_pa +
6748 			pmap_devmap_table[i].pd_size - 1,
6749 		    pmap_devmap_table[i].pd_va);
6750 #endif
6751 		pmap_map_chunk(l1pt, pmap_devmap_table[i].pd_va,
6752 		    pmap_devmap_table[i].pd_pa,
6753 		    pmap_devmap_table[i].pd_size,
6754 		    pmap_devmap_table[i].pd_prot,
6755 		    pmap_devmap_table[i].pd_cache);
6756 	}
6757 }
6758 
6759 const struct pmap_devmap *
6760 pmap_devmap_find_pa(paddr_t pa, psize_t size)
6761 {
6762 	uint64_t endpa;
6763 	int i;
6764 
6765 	if (pmap_devmap_table == NULL)
6766 		return (NULL);
6767 
6768 	endpa = (uint64_t)pa + (uint64_t)(size - 1);
6769 
6770 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
6771 		if (pa >= pmap_devmap_table[i].pd_pa &&
6772 		    endpa <= (uint64_t)pmap_devmap_table[i].pd_pa +
6773 			     (uint64_t)(pmap_devmap_table[i].pd_size - 1))
6774 			return (&pmap_devmap_table[i]);
6775 	}
6776 
6777 	return (NULL);
6778 }
6779 
6780 const struct pmap_devmap *
6781 pmap_devmap_find_va(vaddr_t va, vsize_t size)
6782 {
6783 	int i;
6784 
6785 	if (pmap_devmap_table == NULL)
6786 		return (NULL);
6787 
6788 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
6789 		if (va >= pmap_devmap_table[i].pd_va &&
6790 		    va + size - 1 <= pmap_devmap_table[i].pd_va +
6791 				     pmap_devmap_table[i].pd_size - 1)
6792 			return (&pmap_devmap_table[i]);
6793 	}
6794 
6795 	return (NULL);
6796 }
6797 
6798 /********************** PTE initialization routines **************************/
6799 
6800 /*
6801  * These routines are called when the CPU type is identified to set up
6802  * the PTE prototypes, cache modes, etc.
6803  *
6804  * The variables are always here, just in case modules need to reference
6805  * them (though, they shouldn't).
6806  */
6807 
6808 pt_entry_t	pte_l1_s_cache_mode;
6809 pt_entry_t	pte_l1_s_wc_mode;
6810 pt_entry_t	pte_l1_s_cache_mode_pt;
6811 pt_entry_t	pte_l1_s_cache_mask;
6812 
6813 pt_entry_t	pte_l2_l_cache_mode;
6814 pt_entry_t	pte_l2_l_wc_mode;
6815 pt_entry_t	pte_l2_l_cache_mode_pt;
6816 pt_entry_t	pte_l2_l_cache_mask;
6817 
6818 pt_entry_t	pte_l2_s_cache_mode;
6819 pt_entry_t	pte_l2_s_wc_mode;
6820 pt_entry_t	pte_l2_s_cache_mode_pt;
6821 pt_entry_t	pte_l2_s_cache_mask;
6822 
6823 pt_entry_t	pte_l1_s_prot_u;
6824 pt_entry_t	pte_l1_s_prot_w;
6825 pt_entry_t	pte_l1_s_prot_ro;
6826 pt_entry_t	pte_l1_s_prot_mask;
6827 
6828 pt_entry_t	pte_l2_s_prot_u;
6829 pt_entry_t	pte_l2_s_prot_w;
6830 pt_entry_t	pte_l2_s_prot_ro;
6831 pt_entry_t	pte_l2_s_prot_mask;
6832 
6833 pt_entry_t	pte_l2_l_prot_u;
6834 pt_entry_t	pte_l2_l_prot_w;
6835 pt_entry_t	pte_l2_l_prot_ro;
6836 pt_entry_t	pte_l2_l_prot_mask;
6837 
6838 pt_entry_t	pte_l1_ss_proto;
6839 pt_entry_t	pte_l1_s_proto;
6840 pt_entry_t	pte_l1_c_proto;
6841 pt_entry_t	pte_l2_s_proto;
6842 
6843 void		(*pmap_copy_page_func)(paddr_t, paddr_t);
6844 void		(*pmap_zero_page_func)(paddr_t);
6845 
6846 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
6847 void
6848 pmap_pte_init_generic(void)
6849 {
6850 
6851 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
6852 	pte_l1_s_wc_mode = L1_S_B;
6853 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_generic;
6854 
6855 	pte_l2_l_cache_mode = L2_B|L2_C;
6856 	pte_l2_l_wc_mode = L2_B;
6857 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_generic;
6858 
6859 	pte_l2_s_cache_mode = L2_B|L2_C;
6860 	pte_l2_s_wc_mode = L2_B;
6861 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_generic;
6862 
6863 	/*
6864 	 * If we have a write-through cache, set B and C.  If
6865 	 * we have a write-back cache, then we assume setting
6866 	 * only C will make those pages write-through (except for those
6867 	 * Cortex CPUs which can read the L1 caches).
6868 	 */
6869 	if (cpufuncs.cf_dcache_wb_range == (void *) cpufunc_nullop
6870 #if ARM_MMU_V7 > 0
6871 	    || CPU_ID_CORTEX_P(curcpu()->ci_arm_cpuid)
6872 #endif
6873 #if ARM_MMU_V6 > 0
6874 	    || CPU_ID_ARM11_P(curcpu()->ci_arm_cpuid) /* arm116 errata 399234 */
6875 #endif
6876 	    || false) {
6877 		pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
6878 		pte_l2_l_cache_mode_pt = L2_B|L2_C;
6879 		pte_l2_s_cache_mode_pt = L2_B|L2_C;
6880 	} else {
6881 		pte_l1_s_cache_mode_pt = L1_S_C;	/* write through */
6882 		pte_l2_l_cache_mode_pt = L2_C;		/* write through */
6883 		pte_l2_s_cache_mode_pt = L2_C;		/* write through */
6884 	}
6885 
6886 	pte_l1_s_prot_u = L1_S_PROT_U_generic;
6887 	pte_l1_s_prot_w = L1_S_PROT_W_generic;
6888 	pte_l1_s_prot_ro = L1_S_PROT_RO_generic;
6889 	pte_l1_s_prot_mask = L1_S_PROT_MASK_generic;
6890 
6891 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
6892 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
6893 	pte_l2_s_prot_ro = L2_S_PROT_RO_generic;
6894 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
6895 
6896 	pte_l2_l_prot_u = L2_L_PROT_U_generic;
6897 	pte_l2_l_prot_w = L2_L_PROT_W_generic;
6898 	pte_l2_l_prot_ro = L2_L_PROT_RO_generic;
6899 	pte_l2_l_prot_mask = L2_L_PROT_MASK_generic;
6900 
6901 	pte_l1_ss_proto = L1_SS_PROTO_generic;
6902 	pte_l1_s_proto = L1_S_PROTO_generic;
6903 	pte_l1_c_proto = L1_C_PROTO_generic;
6904 	pte_l2_s_proto = L2_S_PROTO_generic;
6905 
6906 	pmap_copy_page_func = pmap_copy_page_generic;
6907 	pmap_zero_page_func = pmap_zero_page_generic;
6908 }
6909 
6910 #if defined(CPU_ARM8)
6911 void
6912 pmap_pte_init_arm8(void)
6913 {
6914 
6915 	/*
6916 	 * ARM8 is compatible with generic, but we need to use
6917 	 * the page tables uncached.
6918 	 */
6919 	pmap_pte_init_generic();
6920 
6921 	pte_l1_s_cache_mode_pt = 0;
6922 	pte_l2_l_cache_mode_pt = 0;
6923 	pte_l2_s_cache_mode_pt = 0;
6924 }
6925 #endif /* CPU_ARM8 */
6926 
6927 #if defined(CPU_ARM9) && defined(ARM9_CACHE_WRITE_THROUGH)
6928 void
6929 pmap_pte_init_arm9(void)
6930 {
6931 
6932 	/*
6933 	 * ARM9 is compatible with generic, but we want to use
6934 	 * write-through caching for now.
6935 	 */
6936 	pmap_pte_init_generic();
6937 
6938 	pte_l1_s_cache_mode = L1_S_C;
6939 	pte_l2_l_cache_mode = L2_C;
6940 	pte_l2_s_cache_mode = L2_C;
6941 
6942 	pte_l1_s_wc_mode = L1_S_B;
6943 	pte_l2_l_wc_mode = L2_B;
6944 	pte_l2_s_wc_mode = L2_B;
6945 
6946 	pte_l1_s_cache_mode_pt = L1_S_C;
6947 	pte_l2_l_cache_mode_pt = L2_C;
6948 	pte_l2_s_cache_mode_pt = L2_C;
6949 }
6950 #endif /* CPU_ARM9 && ARM9_CACHE_WRITE_THROUGH */
6951 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
6952 
6953 #if defined(CPU_ARM10)
6954 void
6955 pmap_pte_init_arm10(void)
6956 {
6957 
6958 	/*
6959 	 * ARM10 is compatible with generic, but we want to use
6960 	 * write-through caching for now.
6961 	 */
6962 	pmap_pte_init_generic();
6963 
6964 	pte_l1_s_cache_mode = L1_S_B | L1_S_C;
6965 	pte_l2_l_cache_mode = L2_B | L2_C;
6966 	pte_l2_s_cache_mode = L2_B | L2_C;
6967 
6968 	pte_l1_s_cache_mode = L1_S_B;
6969 	pte_l2_l_cache_mode = L2_B;
6970 	pte_l2_s_cache_mode = L2_B;
6971 
6972 	pte_l1_s_cache_mode_pt = L1_S_C;
6973 	pte_l2_l_cache_mode_pt = L2_C;
6974 	pte_l2_s_cache_mode_pt = L2_C;
6975 
6976 }
6977 #endif /* CPU_ARM10 */
6978 
6979 #if defined(CPU_ARM11) && defined(ARM11_CACHE_WRITE_THROUGH)
6980 void
6981 pmap_pte_init_arm11(void)
6982 {
6983 
6984 	/*
6985 	 * ARM11 is compatible with generic, but we want to use
6986 	 * write-through caching for now.
6987 	 */
6988 	pmap_pte_init_generic();
6989 
6990 	pte_l1_s_cache_mode = L1_S_C;
6991 	pte_l2_l_cache_mode = L2_C;
6992 	pte_l2_s_cache_mode = L2_C;
6993 
6994 	pte_l1_s_wc_mode = L1_S_B;
6995 	pte_l2_l_wc_mode = L2_B;
6996 	pte_l2_s_wc_mode = L2_B;
6997 
6998 	pte_l1_s_cache_mode_pt = L1_S_C;
6999 	pte_l2_l_cache_mode_pt = L2_C;
7000 	pte_l2_s_cache_mode_pt = L2_C;
7001 }
7002 #endif /* CPU_ARM11 && ARM11_CACHE_WRITE_THROUGH */
7003 
7004 #if ARM_MMU_SA1 == 1
7005 void
7006 pmap_pte_init_sa1(void)
7007 {
7008 
7009 	/*
7010 	 * The StrongARM SA-1 cache does not have a write-through
7011 	 * mode.  So, do the generic initialization, then reset
7012 	 * the page table cache mode to B=1,C=1, and note that
7013 	 * the PTEs need to be sync'd.
7014 	 */
7015 	pmap_pte_init_generic();
7016 
7017 	pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
7018 	pte_l2_l_cache_mode_pt = L2_B|L2_C;
7019 	pte_l2_s_cache_mode_pt = L2_B|L2_C;
7020 
7021 	pmap_needs_pte_sync = 1;
7022 }
7023 #endif /* ARM_MMU_SA1 == 1*/
7024 
7025 #if ARM_MMU_XSCALE == 1
7026 #if (ARM_NMMUS > 1)
7027 static u_int xscale_use_minidata;
7028 #endif
7029 
7030 void
7031 pmap_pte_init_xscale(void)
7032 {
7033 	uint32_t auxctl;
7034 	int write_through = 0;
7035 
7036 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
7037 	pte_l1_s_wc_mode = L1_S_B;
7038 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_xscale;
7039 
7040 	pte_l2_l_cache_mode = L2_B|L2_C;
7041 	pte_l2_l_wc_mode = L2_B;
7042 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_xscale;
7043 
7044 	pte_l2_s_cache_mode = L2_B|L2_C;
7045 	pte_l2_s_wc_mode = L2_B;
7046 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_xscale;
7047 
7048 	pte_l1_s_cache_mode_pt = L1_S_C;
7049 	pte_l2_l_cache_mode_pt = L2_C;
7050 	pte_l2_s_cache_mode_pt = L2_C;
7051 
7052 #ifdef XSCALE_CACHE_READ_WRITE_ALLOCATE
7053 	/*
7054 	 * The XScale core has an enhanced mode where writes that
7055 	 * miss the cache cause a cache line to be allocated.  This
7056 	 * is significantly faster than the traditional, write-through
7057 	 * behavior of this case.
7058 	 */
7059 	pte_l1_s_cache_mode |= L1_S_XS_TEX(TEX_XSCALE_X);
7060 	pte_l2_l_cache_mode |= L2_XS_L_TEX(TEX_XSCALE_X);
7061 	pte_l2_s_cache_mode |= L2_XS_T_TEX(TEX_XSCALE_X);
7062 #endif /* XSCALE_CACHE_READ_WRITE_ALLOCATE */
7063 
7064 #ifdef XSCALE_CACHE_WRITE_THROUGH
7065 	/*
7066 	 * Some versions of the XScale core have various bugs in
7067 	 * their cache units, the work-around for which is to run
7068 	 * the cache in write-through mode.  Unfortunately, this
7069 	 * has a major (negative) impact on performance.  So, we
7070 	 * go ahead and run fast-and-loose, in the hopes that we
7071 	 * don't line up the planets in a way that will trip the
7072 	 * bugs.
7073 	 *
7074 	 * However, we give you the option to be slow-but-correct.
7075 	 */
7076 	write_through = 1;
7077 #elif defined(XSCALE_CACHE_WRITE_BACK)
7078 	/* force write back cache mode */
7079 	write_through = 0;
7080 #elif defined(CPU_XSCALE_PXA250) || defined(CPU_XSCALE_PXA270)
7081 	/*
7082 	 * Intel PXA2[15]0 processors are known to have a bug in
7083 	 * write-back cache on revision 4 and earlier (stepping
7084 	 * A[01] and B[012]).  Fixed for C0 and later.
7085 	 */
7086 	{
7087 		uint32_t id, type;
7088 
7089 		id = cpufunc_id();
7090 		type = id & ~(CPU_ID_XSCALE_COREREV_MASK|CPU_ID_REVISION_MASK);
7091 
7092 		if (type == CPU_ID_PXA250 || type == CPU_ID_PXA210) {
7093 			if ((id & CPU_ID_REVISION_MASK) < 5) {
7094 				/* write through for stepping A0-1 and B0-2 */
7095 				write_through = 1;
7096 			}
7097 		}
7098 	}
7099 #endif /* XSCALE_CACHE_WRITE_THROUGH */
7100 
7101 	if (write_through) {
7102 		pte_l1_s_cache_mode = L1_S_C;
7103 		pte_l2_l_cache_mode = L2_C;
7104 		pte_l2_s_cache_mode = L2_C;
7105 	}
7106 
7107 #if (ARM_NMMUS > 1)
7108 	xscale_use_minidata = 1;
7109 #endif
7110 
7111 	pte_l1_s_prot_u = L1_S_PROT_U_xscale;
7112 	pte_l1_s_prot_w = L1_S_PROT_W_xscale;
7113 	pte_l1_s_prot_ro = L1_S_PROT_RO_xscale;
7114 	pte_l1_s_prot_mask = L1_S_PROT_MASK_xscale;
7115 
7116 	pte_l2_s_prot_u = L2_S_PROT_U_xscale;
7117 	pte_l2_s_prot_w = L2_S_PROT_W_xscale;
7118 	pte_l2_s_prot_ro = L2_S_PROT_RO_xscale;
7119 	pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale;
7120 
7121 	pte_l2_l_prot_u = L2_L_PROT_U_xscale;
7122 	pte_l2_l_prot_w = L2_L_PROT_W_xscale;
7123 	pte_l2_l_prot_ro = L2_L_PROT_RO_xscale;
7124 	pte_l2_l_prot_mask = L2_L_PROT_MASK_xscale;
7125 
7126 	pte_l1_ss_proto = L1_SS_PROTO_xscale;
7127 	pte_l1_s_proto = L1_S_PROTO_xscale;
7128 	pte_l1_c_proto = L1_C_PROTO_xscale;
7129 	pte_l2_s_proto = L2_S_PROTO_xscale;
7130 
7131 	pmap_copy_page_func = pmap_copy_page_xscale;
7132 	pmap_zero_page_func = pmap_zero_page_xscale;
7133 
7134 	/*
7135 	 * Disable ECC protection of page table access, for now.
7136 	 */
7137 	__asm volatile("mrc p15, 0, %0, c1, c0, 1" : "=r" (auxctl));
7138 	auxctl &= ~XSCALE_AUXCTL_P;
7139 	__asm volatile("mcr p15, 0, %0, c1, c0, 1" : : "r" (auxctl));
7140 }
7141 
7142 /*
7143  * xscale_setup_minidata:
7144  *
7145  *	Set up the mini-data cache clean area.  We require the
7146  *	caller to allocate the right amount of physically and
7147  *	virtually contiguous space.
7148  */
7149 void
7150 xscale_setup_minidata(vaddr_t l1pt, vaddr_t va, paddr_t pa)
7151 {
7152 	extern vaddr_t xscale_minidata_clean_addr;
7153 	extern vsize_t xscale_minidata_clean_size; /* already initialized */
7154 	pd_entry_t *pde = (pd_entry_t *) l1pt;
7155 	vsize_t size;
7156 	uint32_t auxctl;
7157 
7158 	xscale_minidata_clean_addr = va;
7159 
7160 	/* Round it to page size. */
7161 	size = (xscale_minidata_clean_size + L2_S_OFFSET) & L2_S_FRAME;
7162 
7163 	for (; size != 0;
7164 	     va += L2_S_SIZE, pa += L2_S_SIZE, size -= L2_S_SIZE) {
7165 		const size_t l1slot = l1pte_index(va);
7166 		pt_entry_t *ptep = (pt_entry_t *) kernel_pt_lookup(l1pte_pa(pde[l1slot]));
7167 		if (ptep == NULL)
7168 			panic("xscale_setup_minidata: can't find L2 table for "
7169 			    "VA 0x%08lx", va);
7170 
7171 		ptep += l2pte_index(va);
7172 		pt_entry_t opte = *ptep;
7173 		l2pte_set(ptep,
7174 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, VM_PROT_READ)
7175 		    | L2_C | L2_XS_T_TEX(TEX_XSCALE_X), opte);
7176 	}
7177 
7178 	/*
7179 	 * Configure the mini-data cache for write-back with
7180 	 * read/write-allocate.
7181 	 *
7182 	 * NOTE: In order to reconfigure the mini-data cache, we must
7183 	 * make sure it contains no valid data!  In order to do that,
7184 	 * we must issue a global data cache invalidate command!
7185 	 *
7186 	 * WE ASSUME WE ARE RUNNING UN-CACHED WHEN THIS ROUTINE IS CALLED!
7187 	 * THIS IS VERY IMPORTANT!
7188 	 */
7189 
7190 	/* Invalidate data and mini-data. */
7191 	__asm volatile("mcr p15, 0, %0, c7, c6, 0" : : "r" (0));
7192 	__asm volatile("mrc p15, 0, %0, c1, c0, 1" : "=r" (auxctl));
7193 	auxctl = (auxctl & ~XSCALE_AUXCTL_MD_MASK) | XSCALE_AUXCTL_MD_WB_RWA;
7194 	__asm volatile("mcr p15, 0, %0, c1, c0, 1" : : "r" (auxctl));
7195 }
7196 
7197 /*
7198  * Change the PTEs for the specified kernel mappings such that they
7199  * will use the mini data cache instead of the main data cache.
7200  */
7201 void
7202 pmap_uarea(vaddr_t va)
7203 {
7204 	vaddr_t next_bucket, eva;
7205 
7206 #if (ARM_NMMUS > 1)
7207 	if (xscale_use_minidata == 0)
7208 		return;
7209 #endif
7210 
7211 	eva = va + USPACE;
7212 
7213 	while (va < eva) {
7214 		next_bucket = L2_NEXT_BUCKET_VA(va);
7215 		if (next_bucket > eva)
7216 			next_bucket = eva;
7217 
7218 		struct l2_bucket *l2b = pmap_get_l2_bucket(pmap_kernel(), va);
7219 		KDASSERT(l2b != NULL);
7220 
7221 		pt_entry_t * const sptep = &l2b->l2b_kva[l2pte_index(va)];
7222 		pt_entry_t *ptep = sptep;
7223 
7224 		while (va < next_bucket) {
7225 			const pt_entry_t opte = *ptep;
7226 			if (!l2pte_minidata_p(opte)) {
7227 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
7228 				cpu_tlb_flushD_SE(va);
7229 				l2pte_set(ptep, opte & ~L2_B, opte);
7230 			}
7231 			ptep += PAGE_SIZE / L2_S_SIZE;
7232 			va += PAGE_SIZE;
7233 		}
7234 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
7235 	}
7236 	cpu_cpwait();
7237 }
7238 #endif /* ARM_MMU_XSCALE == 1 */
7239 
7240 
7241 #if defined(CPU_ARM11MPCORE)
7242 
7243 void
7244 pmap_pte_init_arm11mpcore(void)
7245 {
7246 
7247 	/* cache mode is controlled by 5 bits (B, C, TEX) */
7248 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_armv6;
7249 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_armv6;
7250 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7251 	/* use extended small page (without APn, with TEX) */
7252 	pte_l2_s_cache_mask = L2_XS_CACHE_MASK_armv6;
7253 #else
7254 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_armv6c;
7255 #endif
7256 
7257 	/* write-back, write-allocate */
7258 	pte_l1_s_cache_mode = L1_S_C | L1_S_B | L1_S_V6_TEX(0x01);
7259 	pte_l2_l_cache_mode = L2_C | L2_B | L2_V6_L_TEX(0x01);
7260 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7261 	pte_l2_s_cache_mode = L2_C | L2_B | L2_V6_XS_TEX(0x01);
7262 #else
7263 	/* no TEX. read-allocate */
7264 	pte_l2_s_cache_mode = L2_C | L2_B;
7265 #endif
7266 	/*
7267 	 * write-back, write-allocate for page tables.
7268 	 */
7269 	pte_l1_s_cache_mode_pt = L1_S_C | L1_S_B | L1_S_V6_TEX(0x01);
7270 	pte_l2_l_cache_mode_pt = L2_C | L2_B | L2_V6_L_TEX(0x01);
7271 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7272 	pte_l2_s_cache_mode_pt = L2_C | L2_B | L2_V6_XS_TEX(0x01);
7273 #else
7274 	pte_l2_s_cache_mode_pt = L2_C | L2_B;
7275 #endif
7276 
7277 	pte_l1_s_prot_u = L1_S_PROT_U_armv6;
7278 	pte_l1_s_prot_w = L1_S_PROT_W_armv6;
7279 	pte_l1_s_prot_ro = L1_S_PROT_RO_armv6;
7280 	pte_l1_s_prot_mask = L1_S_PROT_MASK_armv6;
7281 
7282 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7283 	pte_l2_s_prot_u = L2_S_PROT_U_armv6n;
7284 	pte_l2_s_prot_w = L2_S_PROT_W_armv6n;
7285 	pte_l2_s_prot_ro = L2_S_PROT_RO_armv6n;
7286 	pte_l2_s_prot_mask = L2_S_PROT_MASK_armv6n;
7287 
7288 #else
7289 	/* with AP[0..3] */
7290 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
7291 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
7292 	pte_l2_s_prot_ro = L2_S_PROT_RO_generic;
7293 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
7294 #endif
7295 
7296 #ifdef	ARM11MPCORE_COMPAT_MMU
7297 	/* with AP[0..3] */
7298 	pte_l2_l_prot_u = L2_L_PROT_U_generic;
7299 	pte_l2_l_prot_w = L2_L_PROT_W_generic;
7300 	pte_l2_l_prot_ro = L2_L_PROT_RO_generic;
7301 	pte_l2_l_prot_mask = L2_L_PROT_MASK_generic;
7302 
7303 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
7304 	pte_l1_s_proto = L1_S_PROTO_armv6;
7305 	pte_l1_c_proto = L1_C_PROTO_armv6;
7306 	pte_l2_s_proto = L2_S_PROTO_armv6c;
7307 #else
7308 	pte_l2_l_prot_u = L2_L_PROT_U_armv6n;
7309 	pte_l2_l_prot_w = L2_L_PROT_W_armv6n;
7310 	pte_l2_l_prot_ro = L2_L_PROT_RO_armv6n;
7311 	pte_l2_l_prot_mask = L2_L_PROT_MASK_armv6n;
7312 
7313 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
7314 	pte_l1_s_proto = L1_S_PROTO_armv6;
7315 	pte_l1_c_proto = L1_C_PROTO_armv6;
7316 	pte_l2_s_proto = L2_S_PROTO_armv6n;
7317 #endif
7318 
7319 	pmap_copy_page_func = pmap_copy_page_generic;
7320 	pmap_zero_page_func = pmap_zero_page_generic;
7321 	pmap_needs_pte_sync = 1;
7322 }
7323 #endif	/* CPU_ARM11MPCORE */
7324 
7325 
7326 #if ARM_MMU_V7 == 1
7327 void
7328 pmap_pte_init_armv7(void)
7329 {
7330 	/*
7331 	 * The ARMv7-A MMU is mostly compatible with generic. If the
7332 	 * AP field is zero, that now means "no access" rather than
7333 	 * read-only. The prototypes are a little different because of
7334 	 * the XN bit.
7335 	 */
7336 	pmap_pte_init_generic();
7337 
7338 	pmap_needs_pte_sync = 1;
7339 
7340 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_armv7;
7341 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_armv7;
7342 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_armv7;
7343 
7344 	/*
7345 	 * If the core support coherent walk then updates to translation tables
7346 	 * do not require a clean to the point of unification to ensure
7347 	 * visibility by subsequent translation table walks.  That means we can
7348 	 * map everything shareable and cached and the right thing will happen.
7349 	 */
7350         if (__SHIFTOUT(armreg_mmfr3_read(), __BITS(23,20))) {
7351 		pmap_needs_pte_sync = 0;
7352 
7353 		/*
7354 		 * write-back, no write-allocate, shareable for normal pages.
7355 		 */
7356 		pte_l1_s_cache_mode |= L1_S_V6_S;
7357 		pte_l2_l_cache_mode |= L2_XS_S;
7358 		pte_l2_s_cache_mode |= L2_XS_S;
7359 	}
7360 
7361 	/*
7362 	 * Page tables are just all other memory.  We can use write-back since
7363 	 * pmap_needs_pte_sync is 1 (or the MMU can read out of cache).
7364 	 */
7365 	pte_l1_s_cache_mode_pt = pte_l1_s_cache_mode;
7366 	pte_l2_l_cache_mode_pt = pte_l2_l_cache_mode;
7367 	pte_l2_s_cache_mode_pt = pte_l2_s_cache_mode;
7368 
7369 	/*
7370 	 * Check the Memory Model Features to see if this CPU supports
7371 	 * the TLBIASID coproc op.
7372 	 */
7373 	if (__SHIFTOUT(armreg_mmfr2_read(), __BITS(16,19)) >= 2) {
7374 		arm_has_tlbiasid_p = true;
7375 	}
7376 
7377 	pte_l1_s_prot_u = L1_S_PROT_U_armv7;
7378 	pte_l1_s_prot_w = L1_S_PROT_W_armv7;
7379 	pte_l1_s_prot_ro = L1_S_PROT_RO_armv7;
7380 	pte_l1_s_prot_mask = L1_S_PROT_MASK_armv7;
7381 
7382 	pte_l2_s_prot_u = L2_S_PROT_U_armv7;
7383 	pte_l2_s_prot_w = L2_S_PROT_W_armv7;
7384 	pte_l2_s_prot_ro = L2_S_PROT_RO_armv7;
7385 	pte_l2_s_prot_mask = L2_S_PROT_MASK_armv7;
7386 
7387 	pte_l2_l_prot_u = L2_L_PROT_U_armv7;
7388 	pte_l2_l_prot_w = L2_L_PROT_W_armv7;
7389 	pte_l2_l_prot_ro = L2_L_PROT_RO_armv7;
7390 	pte_l2_l_prot_mask = L2_L_PROT_MASK_armv7;
7391 
7392 	pte_l1_ss_proto = L1_SS_PROTO_armv7;
7393 	pte_l1_s_proto = L1_S_PROTO_armv7;
7394 	pte_l1_c_proto = L1_C_PROTO_armv7;
7395 	pte_l2_s_proto = L2_S_PROTO_armv7;
7396 
7397 }
7398 #endif /* ARM_MMU_V7 */
7399 
7400 /*
7401  * return the PA of the current L1 table, for use when handling a crash dump
7402  */
7403 uint32_t
7404 pmap_kernel_L1_addr(void)
7405 {
7406 #ifdef ARM_MMU_EXTENDED
7407 	return pmap_kernel()->pm_l1_pa;
7408 #else
7409 	return pmap_kernel()->pm_l1->l1_physaddr;
7410 #endif
7411 }
7412 
7413 #if defined(DDB)
7414 /*
7415  * A couple of ddb-callable functions for dumping pmaps
7416  */
7417 void pmap_dump_all(void);
7418 void pmap_dump(pmap_t);
7419 
7420 void
7421 pmap_dump_all(void)
7422 {
7423 	pmap_t pm;
7424 
7425 	LIST_FOREACH(pm, &pmap_pmaps, pm_list) {
7426 		if (pm == pmap_kernel())
7427 			continue;
7428 		pmap_dump(pm);
7429 		printf("\n");
7430 	}
7431 }
7432 
7433 static pt_entry_t ncptes[64];
7434 static void pmap_dump_ncpg(pmap_t);
7435 
7436 void
7437 pmap_dump(pmap_t pm)
7438 {
7439 	struct l2_dtable *l2;
7440 	struct l2_bucket *l2b;
7441 	pt_entry_t *ptep, pte;
7442 	vaddr_t l2_va, l2b_va, va;
7443 	int i, j, k, occ, rows = 0;
7444 
7445 	if (pm == pmap_kernel())
7446 		printf("pmap_kernel (%p): ", pm);
7447 	else
7448 		printf("user pmap (%p): ", pm);
7449 
7450 #ifdef ARM_MMU_EXTENDED
7451 	printf("l1 at %p\n", pmap_l1_kva(pm));
7452 #else
7453 	printf("domain %d, l1 at %p\n", pmap_domain(pm), pmap_l1_kva(pm));
7454 #endif
7455 
7456 	l2_va = 0;
7457 	for (i = 0; i < L2_SIZE; i++, l2_va += 0x01000000) {
7458 		l2 = pm->pm_l2[i];
7459 
7460 		if (l2 == NULL || l2->l2_occupancy == 0)
7461 			continue;
7462 
7463 		l2b_va = l2_va;
7464 		for (j = 0; j < L2_BUCKET_SIZE; j++, l2b_va += 0x00100000) {
7465 			l2b = &l2->l2_bucket[j];
7466 
7467 			if (l2b->l2b_occupancy == 0 || l2b->l2b_kva == NULL)
7468 				continue;
7469 
7470 			ptep = l2b->l2b_kva;
7471 
7472 			for (k = 0; k < 256 && ptep[k] == 0; k++)
7473 				;
7474 
7475 			k &= ~63;
7476 			occ = l2b->l2b_occupancy;
7477 			va = l2b_va + (k * 4096);
7478 			for (; k < 256; k++, va += 0x1000) {
7479 				char ch = ' ';
7480 				if ((k % 64) == 0) {
7481 					if ((rows % 8) == 0) {
7482 						printf(
7483 "          |0000   |8000   |10000  |18000  |20000  |28000  |30000  |38000\n");
7484 					}
7485 					printf("%08lx: ", va);
7486 				}
7487 
7488 				ncptes[k & 63] = 0;
7489 				pte = ptep[k];
7490 				if (pte == 0) {
7491 					ch = '.';
7492 				} else {
7493 					occ--;
7494 					switch (pte & 0x0c) {
7495 					case 0x00:
7496 						ch = 'D'; /* No cache No buff */
7497 						break;
7498 					case 0x04:
7499 						ch = 'B'; /* No cache buff */
7500 						break;
7501 					case 0x08:
7502 						if (pte & 0x40)
7503 							ch = 'm';
7504 						else
7505 						   ch = 'C'; /* Cache No buff */
7506 						break;
7507 					case 0x0c:
7508 						ch = 'F'; /* Cache Buff */
7509 						break;
7510 					}
7511 
7512 					if ((pte & L2_S_PROT_U) == L2_S_PROT_U)
7513 						ch += 0x20;
7514 
7515 					if ((pte & 0xc) == 0)
7516 						ncptes[k & 63] = pte;
7517 				}
7518 
7519 				if ((k % 64) == 63) {
7520 					rows++;
7521 					printf("%c\n", ch);
7522 					pmap_dump_ncpg(pm);
7523 					if (occ == 0)
7524 						break;
7525 				} else
7526 					printf("%c", ch);
7527 			}
7528 		}
7529 	}
7530 }
7531 
7532 static void
7533 pmap_dump_ncpg(pmap_t pm)
7534 {
7535 	struct vm_page *pg;
7536 	struct vm_page_md *md;
7537 	struct pv_entry *pv;
7538 	int i;
7539 
7540 	for (i = 0; i < 63; i++) {
7541 		if (ncptes[i] == 0)
7542 			continue;
7543 
7544 		pg = PHYS_TO_VM_PAGE(l2pte_pa(ncptes[i]));
7545 		if (pg == NULL)
7546 			continue;
7547 		md = VM_PAGE_TO_MD(pg);
7548 
7549 		printf(" pa 0x%08lx: krw %d kro %d urw %d uro %d\n",
7550 		    VM_PAGE_TO_PHYS(pg),
7551 		    md->krw_mappings, md->kro_mappings,
7552 		    md->urw_mappings, md->uro_mappings);
7553 
7554 		SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
7555 			printf("   %c va 0x%08lx, flags 0x%x\n",
7556 			    (pm == pv->pv_pmap) ? '*' : ' ',
7557 			    pv->pv_va, pv->pv_flags);
7558 		}
7559 	}
7560 }
7561 #endif
7562 
7563 #ifdef PMAP_STEAL_MEMORY
7564 void
7565 pmap_boot_pageadd(pv_addr_t *newpv)
7566 {
7567 	pv_addr_t *pv, *npv;
7568 
7569 	if ((pv = SLIST_FIRST(&pmap_boot_freeq)) != NULL) {
7570 		if (newpv->pv_pa < pv->pv_va) {
7571 			KASSERT(newpv->pv_pa + newpv->pv_size <= pv->pv_pa);
7572 			if (newpv->pv_pa + newpv->pv_size == pv->pv_pa) {
7573 				newpv->pv_size += pv->pv_size;
7574 				SLIST_REMOVE_HEAD(&pmap_boot_freeq, pv_list);
7575 			}
7576 			pv = NULL;
7577 		} else {
7578 			for (; (npv = SLIST_NEXT(pv, pv_list)) != NULL;
7579 			     pv = npv) {
7580 				KASSERT(pv->pv_pa + pv->pv_size < npv->pv_pa);
7581 				KASSERT(pv->pv_pa < newpv->pv_pa);
7582 				if (newpv->pv_pa > npv->pv_pa)
7583 					continue;
7584 				if (pv->pv_pa + pv->pv_size == newpv->pv_pa) {
7585 					pv->pv_size += newpv->pv_size;
7586 					return;
7587 				}
7588 				if (newpv->pv_pa + newpv->pv_size < npv->pv_pa)
7589 					break;
7590 				newpv->pv_size += npv->pv_size;
7591 				SLIST_INSERT_AFTER(pv, newpv, pv_list);
7592 				SLIST_REMOVE_AFTER(newpv, pv_list);
7593 				return;
7594 			}
7595 		}
7596 	}
7597 
7598 	if (pv) {
7599 		SLIST_INSERT_AFTER(pv, newpv, pv_list);
7600 	} else {
7601 		SLIST_INSERT_HEAD(&pmap_boot_freeq, newpv, pv_list);
7602 	}
7603 }
7604 
7605 void
7606 pmap_boot_pagealloc(psize_t amount, psize_t mask, psize_t match,
7607 	pv_addr_t *rpv)
7608 {
7609 	pv_addr_t *pv, **pvp;
7610 	struct vm_physseg *ps;
7611 	size_t i;
7612 
7613 	KASSERT(amount & PGOFSET);
7614 	KASSERT((mask & PGOFSET) == 0);
7615 	KASSERT((match & PGOFSET) == 0);
7616 	KASSERT(amount != 0);
7617 
7618 	for (pvp = &SLIST_FIRST(&pmap_boot_freeq);
7619 	     (pv = *pvp) != NULL;
7620 	     pvp = &SLIST_NEXT(pv, pv_list)) {
7621 		pv_addr_t *newpv;
7622 		psize_t off;
7623 		/*
7624 		 * If this entry is too small to satify the request...
7625 		 */
7626 		KASSERT(pv->pv_size > 0);
7627 		if (pv->pv_size < amount)
7628 			continue;
7629 
7630 		for (off = 0; off <= mask; off += PAGE_SIZE) {
7631 			if (((pv->pv_pa + off) & mask) == match
7632 			    && off + amount <= pv->pv_size)
7633 				break;
7634 		}
7635 		if (off > mask)
7636 			continue;
7637 
7638 		rpv->pv_va = pv->pv_va + off;
7639 		rpv->pv_pa = pv->pv_pa + off;
7640 		rpv->pv_size = amount;
7641 		pv->pv_size -= amount;
7642 		if (pv->pv_size == 0) {
7643 			KASSERT(off == 0);
7644 			KASSERT((vaddr_t) pv == rpv->pv_va);
7645 			*pvp = SLIST_NEXT(pv, pv_list);
7646 		} else if (off == 0) {
7647 			KASSERT((vaddr_t) pv == rpv->pv_va);
7648 			newpv = (pv_addr_t *) (rpv->pv_va + amount);
7649 			*newpv = *pv;
7650 			newpv->pv_pa += amount;
7651 			newpv->pv_va += amount;
7652 			*pvp = newpv;
7653 		} else if (off < pv->pv_size) {
7654 			newpv = (pv_addr_t *) (rpv->pv_va + amount);
7655 			*newpv = *pv;
7656 			newpv->pv_size -= off;
7657 			newpv->pv_pa += off + amount;
7658 			newpv->pv_va += off + amount;
7659 
7660 			SLIST_NEXT(pv, pv_list) = newpv;
7661 			pv->pv_size = off;
7662 		} else {
7663 			KASSERT((vaddr_t) pv != rpv->pv_va);
7664 		}
7665 		memset((void *)rpv->pv_va, 0, amount);
7666 		return;
7667 	}
7668 
7669 	if (vm_nphysseg == 0)
7670 		panic("pmap_boot_pagealloc: couldn't allocate memory");
7671 
7672 	for (pvp = &SLIST_FIRST(&pmap_boot_freeq);
7673 	     (pv = *pvp) != NULL;
7674 	     pvp = &SLIST_NEXT(pv, pv_list)) {
7675 		if (SLIST_NEXT(pv, pv_list) == NULL)
7676 			break;
7677 	}
7678 	KASSERT(mask == 0);
7679 	for (i = 0; i < vm_nphysseg; i++) {
7680 		ps = VM_PHYSMEM_PTR(i);
7681 		if (ps->avail_start == atop(pv->pv_pa + pv->pv_size)
7682 		    && pv->pv_va + pv->pv_size <= ptoa(ps->avail_end)) {
7683 			rpv->pv_va = pv->pv_va;
7684 			rpv->pv_pa = pv->pv_pa;
7685 			rpv->pv_size = amount;
7686 			*pvp = NULL;
7687 			pmap_map_chunk(kernel_l1pt.pv_va,
7688 			     ptoa(ps->avail_start) + (pv->pv_va - pv->pv_pa),
7689 			     ptoa(ps->avail_start),
7690 			     amount - pv->pv_size,
7691 			     VM_PROT_READ|VM_PROT_WRITE,
7692 			     PTE_CACHE);
7693 			ps->avail_start += atop(amount - pv->pv_size);
7694 			/*
7695 			 * If we consumed the entire physseg, remove it.
7696 			 */
7697 			if (ps->avail_start == ps->avail_end) {
7698 				for (--vm_nphysseg; i < vm_nphysseg; i++)
7699 					VM_PHYSMEM_PTR_SWAP(i, i + 1);
7700 			}
7701 			memset((void *)rpv->pv_va, 0, rpv->pv_size);
7702 			return;
7703 		}
7704 	}
7705 
7706 	panic("pmap_boot_pagealloc: couldn't allocate memory");
7707 }
7708 
7709 vaddr_t
7710 pmap_steal_memory(vsize_t size, vaddr_t *vstartp, vaddr_t *vendp)
7711 {
7712 	pv_addr_t pv;
7713 
7714 	pmap_boot_pagealloc(size, 0, 0, &pv);
7715 
7716 	return pv.pv_va;
7717 }
7718 #endif /* PMAP_STEAL_MEMORY */
7719 
7720 SYSCTL_SETUP(sysctl_machdep_pmap_setup, "sysctl machdep.kmpages setup")
7721 {
7722 	sysctl_createv(clog, 0, NULL, NULL,
7723 			CTLFLAG_PERMANENT,
7724 			CTLTYPE_NODE, "machdep", NULL,
7725 			NULL, 0, NULL, 0,
7726 			CTL_MACHDEP, CTL_EOL);
7727 
7728 	sysctl_createv(clog, 0, NULL, NULL,
7729 			CTLFLAG_PERMANENT,
7730 			CTLTYPE_INT, "kmpages",
7731 			SYSCTL_DESCR("count of pages allocated to kernel memory allocators"),
7732 			NULL, 0, &pmap_kmpages, 0,
7733 			CTL_MACHDEP, CTL_CREATE, CTL_EOL);
7734 }
7735 
7736 #ifdef PMAP_NEED_ALLOC_POOLPAGE
7737 struct vm_page *
7738 arm_pmap_alloc_poolpage(int flags)
7739 {
7740 	/*
7741 	 * On some systems, only some pages may be "coherent" for dma and we
7742 	 * want to prefer those for pool pages (think mbufs) but fallback to
7743 	 * any page if none is available.  But we can only fallback if we
7744 	 * aren't direct mapping memory or all of memory can be direct-mapped.
7745 	 * If that isn't true, pool changes can only come from direct-mapped
7746 	 * memory.
7747 	 */
7748 	if (arm_poolpage_vmfreelist != VM_FREELIST_DEFAULT) {
7749 		return uvm_pagealloc_strat(NULL, 0, NULL, flags,
7750 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS) && defined(ARM_MMU_EXTENDED)
7751 		    (pmap_directbase < KERNEL_BASE
7752 			? UVM_PGA_STRAT_ONLY
7753 			: UVM_PGA_STRAT_FALLBACK),
7754 #else
7755 		    UVM_PGA_STRAT_FALLBACK,
7756 #endif
7757 		    arm_poolpage_vmfreelist);
7758 	}
7759 
7760 	return uvm_pagealloc(NULL, 0, NULL, flags);
7761 }
7762 #endif
7763 
7764 #if defined(ARM_MMU_EXTENDED) && defined(MULTIPROCESSOR)
7765 void
7766 pmap_md_tlb_info_attach(struct pmap_tlb_info *ti, struct cpu_info *ci)
7767 {
7768         /* nothing */
7769 }
7770 
7771 int
7772 pic_ipi_shootdown(void *arg)
7773 {
7774 #if PMAP_NEED_TLB_SHOOTDOWN
7775 	pmap_tlb_shootdown_process();
7776 #endif
7777 	return 1;
7778 }
7779 #endif /* ARM_MMU_EXTENDED && MULTIPROCESSOR */
7780 
7781 
7782 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
7783 vaddr_t
7784 pmap_direct_mapped_phys(paddr_t pa, bool *ok_p, vaddr_t va)
7785 {
7786 	bool ok = false;
7787 	if (physical_start <= pa && pa < physical_end) {
7788 #ifdef ARM_MMU_EXTENDED
7789 		const vaddr_t newva = pmap_directbase + pa - physical_start;
7790 		if (newva >= KERNEL_BASE) {
7791 			va = newva;
7792 			ok = true;
7793 		}
7794 #else
7795 		va = KERNEL_BASE + pa - physical_start;
7796 		ok = true;
7797 #endif
7798 	}
7799 	KASSERT(ok_p);
7800 	*ok_p = ok;
7801 	return va;
7802 }
7803 
7804 vaddr_t
7805 pmap_map_poolpage(paddr_t pa)
7806 {
7807 	bool ok __diagused;
7808 	vaddr_t va = pmap_direct_mapped_phys(pa, &ok, 0);
7809 	KASSERT(ok);
7810 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
7811 	if (arm_cache_prefer_mask != 0) {
7812 		struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
7813 		struct vm_page_md * const md = VM_PAGE_TO_MD(pg);
7814 		pmap_acquire_page_lock(md);
7815 		pmap_vac_me_harder(md, pa, pmap_kernel(), va);
7816 		pmap_release_page_lock(md);
7817 	}
7818 #endif
7819 	return va;
7820 }
7821 
7822 paddr_t
7823 pmap_unmap_poolpage(vaddr_t va)
7824 {
7825 	KASSERT(va >= KERNEL_BASE);
7826 #if defined(ARM_MMU_EXTENDED)
7827 	return va - pmap_directbase + physical_start;
7828 #else
7829 #ifdef PMAP_CACHE_VIVT
7830 	cpu_idcache_wbinv_range(va, PAGE_SIZE);
7831 #endif
7832         return va - KERNEL_BASE + physical_start;
7833 #endif
7834 }
7835 #endif /* __HAVE_MM_MD_DIRECT_MAPPED_PHYS */
7836