xref: /netbsd-src/sys/arch/arm/arm32/pmap.c (revision d90047b5d07facf36e6c01dcc0bded8997ce9cc2)
1 /*	$NetBSD: pmap.c,v 1.417 2020/07/10 12:25:09 skrll Exp $	*/
2 
3 /*
4  * Copyright 2003 Wasabi Systems, Inc.
5  * All rights reserved.
6  *
7  * Written by Steve C. Woodford for Wasabi Systems, Inc.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. All advertising materials mentioning features or use of this software
18  *    must display the following acknowledgement:
19  *      This product includes software developed for the NetBSD Project by
20  *      Wasabi Systems, Inc.
21  * 4. The name of Wasabi Systems, Inc. may not be used to endorse
22  *    or promote products derived from this software without specific prior
23  *    written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY WASABI SYSTEMS, INC. ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
27  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
28  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL WASABI SYSTEMS, INC
29  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
30  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
31  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
32  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
33  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
34  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
35  * POSSIBILITY OF SUCH DAMAGE.
36  */
37 
38 /*
39  * Copyright (c) 2002-2003 Wasabi Systems, Inc.
40  * Copyright (c) 2001 Richard Earnshaw
41  * Copyright (c) 2001-2002 Christopher Gilbert
42  * All rights reserved.
43  *
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. The name of the company nor the name of the author may be used to
50  *   endorse or promote products derived from this software without specific
51  *    prior written permission.
52  *
53  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
54  * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
55  * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
56  * IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
57  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
58  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
59  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
60  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
61  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
62  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
63  * SUCH DAMAGE.
64  */
65 
66 /*-
67  * Copyright (c) 1999, 2020 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 /* Include header files */
183 
184 #include "opt_arm_debug.h"
185 #include "opt_cpuoptions.h"
186 #include "opt_ddb.h"
187 #include "opt_lockdebug.h"
188 #include "opt_multiprocessor.h"
189 
190 #ifdef MULTIPROCESSOR
191 #define _INTR_PRIVATE
192 #endif
193 
194 #include <sys/cdefs.h>
195 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.417 2020/07/10 12:25:09 skrll Exp $");
196 
197 #include <sys/param.h>
198 #include <sys/types.h>
199 
200 #include <sys/asan.h>
201 #include <sys/atomic.h>
202 #include <sys/bus.h>
203 #include <sys/cpu.h>
204 #include <sys/intr.h>
205 #include <sys/kernel.h>
206 #include <sys/kernhist.h>
207 #include <sys/kmem.h>
208 #include <sys/pool.h>
209 #include <sys/proc.h>
210 #include <sys/sysctl.h>
211 #include <sys/systm.h>
212 
213 #include <uvm/uvm.h>
214 #include <uvm/pmap/pmap_pvt.h>
215 
216 #include <arm/locore.h>
217 
218 #ifdef DDB
219 #include <arm/db_machdep.h>
220 #endif
221 
222 #ifdef VERBOSE_INIT_ARM
223 #define VPRINTF(...)	printf(__VA_ARGS__)
224 #else
225 #define VPRINTF(...)	__nothing
226 #endif
227 
228 /*
229  * pmap_kernel() points here
230  */
231 static struct pmap	kernel_pmap_store = {
232 #ifndef ARM_MMU_EXTENDED
233 	.pm_activated = true,
234 	.pm_domain = PMAP_DOMAIN_KERNEL,
235 	.pm_cstate.cs_all = PMAP_CACHE_STATE_ALL,
236 #endif
237 };
238 struct pmap * const	kernel_pmap_ptr = &kernel_pmap_store;
239 #undef pmap_kernel
240 #define pmap_kernel()	(&kernel_pmap_store)
241 #ifdef PMAP_NEED_ALLOC_POOLPAGE
242 int			arm_poolpage_vmfreelist = VM_FREELIST_DEFAULT;
243 #endif
244 
245 /*
246  * Pool and cache that pmap structures are allocated from.
247  * We use a cache to avoid clearing the pm_l2[] array (1KB)
248  * in pmap_create().
249  */
250 static struct pool_cache pmap_cache;
251 
252 /*
253  * Pool of PV structures
254  */
255 static struct pool pmap_pv_pool;
256 static void *pmap_bootstrap_pv_page_alloc(struct pool *, int);
257 static void pmap_bootstrap_pv_page_free(struct pool *, void *);
258 static struct pool_allocator pmap_bootstrap_pv_allocator = {
259 	pmap_bootstrap_pv_page_alloc, pmap_bootstrap_pv_page_free
260 };
261 
262 /*
263  * Pool and cache of l2_dtable structures.
264  * We use a cache to avoid clearing the structures when they're
265  * allocated. (196 bytes)
266  */
267 static struct pool_cache pmap_l2dtable_cache;
268 static vaddr_t pmap_kernel_l2dtable_kva;
269 
270 /*
271  * Pool and cache of L2 page descriptors.
272  * We use a cache to avoid clearing the descriptor table
273  * when they're allocated. (1KB)
274  */
275 static struct pool_cache pmap_l2ptp_cache;
276 static vaddr_t pmap_kernel_l2ptp_kva;
277 static paddr_t pmap_kernel_l2ptp_phys;
278 
279 #ifdef PMAPCOUNTERS
280 #define	PMAP_EVCNT_INITIALIZER(name) \
281 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap", name)
282 
283 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
284 static struct evcnt pmap_ev_vac_clean_one =
285    PMAP_EVCNT_INITIALIZER("clean page (1 color)");
286 static struct evcnt pmap_ev_vac_flush_one =
287    PMAP_EVCNT_INITIALIZER("flush page (1 color)");
288 static struct evcnt pmap_ev_vac_flush_lots =
289    PMAP_EVCNT_INITIALIZER("flush page (2+ colors)");
290 static struct evcnt pmap_ev_vac_flush_lots2 =
291    PMAP_EVCNT_INITIALIZER("flush page (2+ colors, kmpage)");
292 EVCNT_ATTACH_STATIC(pmap_ev_vac_clean_one);
293 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_one);
294 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_lots);
295 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_lots2);
296 
297 static struct evcnt pmap_ev_vac_color_new =
298    PMAP_EVCNT_INITIALIZER("new page color");
299 static struct evcnt pmap_ev_vac_color_reuse =
300    PMAP_EVCNT_INITIALIZER("ok first page color");
301 static struct evcnt pmap_ev_vac_color_ok =
302    PMAP_EVCNT_INITIALIZER("ok page color");
303 static struct evcnt pmap_ev_vac_color_blind =
304    PMAP_EVCNT_INITIALIZER("blind page color");
305 static struct evcnt pmap_ev_vac_color_change =
306    PMAP_EVCNT_INITIALIZER("change page color");
307 static struct evcnt pmap_ev_vac_color_erase =
308    PMAP_EVCNT_INITIALIZER("erase page color");
309 static struct evcnt pmap_ev_vac_color_none =
310    PMAP_EVCNT_INITIALIZER("no page color");
311 static struct evcnt pmap_ev_vac_color_restore =
312    PMAP_EVCNT_INITIALIZER("restore page color");
313 
314 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_new);
315 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_reuse);
316 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_ok);
317 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_blind);
318 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_change);
319 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_erase);
320 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_none);
321 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_restore);
322 #endif
323 
324 static struct evcnt pmap_ev_mappings =
325    PMAP_EVCNT_INITIALIZER("pages mapped");
326 static struct evcnt pmap_ev_unmappings =
327    PMAP_EVCNT_INITIALIZER("pages unmapped");
328 static struct evcnt pmap_ev_remappings =
329    PMAP_EVCNT_INITIALIZER("pages remapped");
330 
331 EVCNT_ATTACH_STATIC(pmap_ev_mappings);
332 EVCNT_ATTACH_STATIC(pmap_ev_unmappings);
333 EVCNT_ATTACH_STATIC(pmap_ev_remappings);
334 
335 static struct evcnt pmap_ev_kernel_mappings =
336    PMAP_EVCNT_INITIALIZER("kernel pages mapped");
337 static struct evcnt pmap_ev_kernel_unmappings =
338    PMAP_EVCNT_INITIALIZER("kernel pages unmapped");
339 static struct evcnt pmap_ev_kernel_remappings =
340    PMAP_EVCNT_INITIALIZER("kernel pages remapped");
341 
342 EVCNT_ATTACH_STATIC(pmap_ev_kernel_mappings);
343 EVCNT_ATTACH_STATIC(pmap_ev_kernel_unmappings);
344 EVCNT_ATTACH_STATIC(pmap_ev_kernel_remappings);
345 
346 static struct evcnt pmap_ev_kenter_mappings =
347    PMAP_EVCNT_INITIALIZER("kenter pages mapped");
348 static struct evcnt pmap_ev_kenter_unmappings =
349    PMAP_EVCNT_INITIALIZER("kenter pages unmapped");
350 static struct evcnt pmap_ev_kenter_remappings =
351    PMAP_EVCNT_INITIALIZER("kenter pages remapped");
352 static struct evcnt pmap_ev_pt_mappings =
353    PMAP_EVCNT_INITIALIZER("page table pages mapped");
354 
355 EVCNT_ATTACH_STATIC(pmap_ev_kenter_mappings);
356 EVCNT_ATTACH_STATIC(pmap_ev_kenter_unmappings);
357 EVCNT_ATTACH_STATIC(pmap_ev_kenter_remappings);
358 EVCNT_ATTACH_STATIC(pmap_ev_pt_mappings);
359 
360 static struct evcnt pmap_ev_fixup_mod =
361    PMAP_EVCNT_INITIALIZER("page modification emulations");
362 static struct evcnt pmap_ev_fixup_ref =
363    PMAP_EVCNT_INITIALIZER("page reference emulations");
364 static struct evcnt pmap_ev_fixup_exec =
365    PMAP_EVCNT_INITIALIZER("exec pages fixed up");
366 static struct evcnt pmap_ev_fixup_pdes =
367    PMAP_EVCNT_INITIALIZER("pdes fixed up");
368 #ifndef ARM_MMU_EXTENDED
369 static struct evcnt pmap_ev_fixup_ptesync =
370    PMAP_EVCNT_INITIALIZER("ptesync fixed");
371 #endif
372 
373 EVCNT_ATTACH_STATIC(pmap_ev_fixup_mod);
374 EVCNT_ATTACH_STATIC(pmap_ev_fixup_ref);
375 EVCNT_ATTACH_STATIC(pmap_ev_fixup_exec);
376 EVCNT_ATTACH_STATIC(pmap_ev_fixup_pdes);
377 #ifndef ARM_MMU_EXTENDED
378 EVCNT_ATTACH_STATIC(pmap_ev_fixup_ptesync);
379 #endif
380 
381 #ifdef PMAP_CACHE_VIPT
382 static struct evcnt pmap_ev_exec_mappings =
383    PMAP_EVCNT_INITIALIZER("exec pages mapped");
384 static struct evcnt pmap_ev_exec_cached =
385    PMAP_EVCNT_INITIALIZER("exec pages cached");
386 
387 EVCNT_ATTACH_STATIC(pmap_ev_exec_mappings);
388 EVCNT_ATTACH_STATIC(pmap_ev_exec_cached);
389 
390 static struct evcnt pmap_ev_exec_synced =
391    PMAP_EVCNT_INITIALIZER("exec pages synced");
392 static struct evcnt pmap_ev_exec_synced_map =
393    PMAP_EVCNT_INITIALIZER("exec pages synced (MP)");
394 static struct evcnt pmap_ev_exec_synced_unmap =
395    PMAP_EVCNT_INITIALIZER("exec pages synced (UM)");
396 static struct evcnt pmap_ev_exec_synced_remap =
397    PMAP_EVCNT_INITIALIZER("exec pages synced (RM)");
398 static struct evcnt pmap_ev_exec_synced_clearbit =
399    PMAP_EVCNT_INITIALIZER("exec pages synced (DG)");
400 #ifndef ARM_MMU_EXTENDED
401 static struct evcnt pmap_ev_exec_synced_kremove =
402    PMAP_EVCNT_INITIALIZER("exec pages synced (KU)");
403 #endif
404 
405 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced);
406 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_map);
407 #ifndef ARM_MMU_EXTENDED
408 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_unmap);
409 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_remap);
410 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_clearbit);
411 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_kremove);
412 #endif
413 
414 static struct evcnt pmap_ev_exec_discarded_unmap =
415    PMAP_EVCNT_INITIALIZER("exec pages discarded (UM)");
416 static struct evcnt pmap_ev_exec_discarded_zero =
417    PMAP_EVCNT_INITIALIZER("exec pages discarded (ZP)");
418 static struct evcnt pmap_ev_exec_discarded_copy =
419    PMAP_EVCNT_INITIALIZER("exec pages discarded (CP)");
420 static struct evcnt pmap_ev_exec_discarded_page_protect =
421    PMAP_EVCNT_INITIALIZER("exec pages discarded (PP)");
422 static struct evcnt pmap_ev_exec_discarded_clearbit =
423    PMAP_EVCNT_INITIALIZER("exec pages discarded (DG)");
424 static struct evcnt pmap_ev_exec_discarded_kremove =
425    PMAP_EVCNT_INITIALIZER("exec pages discarded (KU)");
426 #ifdef ARM_MMU_EXTENDED
427 static struct evcnt pmap_ev_exec_discarded_modfixup =
428    PMAP_EVCNT_INITIALIZER("exec pages discarded (MF)");
429 #endif
430 
431 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_unmap);
432 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_zero);
433 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_copy);
434 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_page_protect);
435 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_clearbit);
436 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_kremove);
437 #ifdef ARM_MMU_EXTENDED
438 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_modfixup);
439 #endif
440 #endif /* PMAP_CACHE_VIPT */
441 
442 static struct evcnt pmap_ev_updates = PMAP_EVCNT_INITIALIZER("updates");
443 static struct evcnt pmap_ev_collects = PMAP_EVCNT_INITIALIZER("collects");
444 static struct evcnt pmap_ev_activations = PMAP_EVCNT_INITIALIZER("activations");
445 
446 EVCNT_ATTACH_STATIC(pmap_ev_updates);
447 EVCNT_ATTACH_STATIC(pmap_ev_collects);
448 EVCNT_ATTACH_STATIC(pmap_ev_activations);
449 
450 #define	PMAPCOUNT(x)	((void)(pmap_ev_##x.ev_count++))
451 #else
452 #define	PMAPCOUNT(x)	((void)0)
453 #endif
454 
455 #ifdef ARM_MMU_EXTENDED
456 void pmap_md_pdetab_activate(pmap_t, struct lwp *);
457 void pmap_md_pdetab_deactivate(pmap_t pm);
458 #endif
459 
460 /*
461  * pmap copy/zero page, and mem(5) hook point
462  */
463 static pt_entry_t *csrc_pte, *cdst_pte;
464 static vaddr_t csrcp, cdstp;
465 #ifdef MULTIPROCESSOR
466 static size_t cnptes;
467 #define	cpu_csrc_pte(o)	(csrc_pte + cnptes * cpu_number() + ((o) >> L2_S_SHIFT))
468 #define	cpu_cdst_pte(o)	(cdst_pte + cnptes * cpu_number() + ((o) >> L2_S_SHIFT))
469 #define	cpu_csrcp(o)	(csrcp + L2_S_SIZE * cnptes * cpu_number() + (o))
470 #define	cpu_cdstp(o)	(cdstp + L2_S_SIZE * cnptes * cpu_number() + (o))
471 #else
472 #define	cpu_csrc_pte(o)	(csrc_pte + ((o) >> L2_S_SHIFT))
473 #define	cpu_cdst_pte(o)	(cdst_pte + ((o) >> L2_S_SHIFT))
474 #define	cpu_csrcp(o)	(csrcp + (o))
475 #define	cpu_cdstp(o)	(cdstp + (o))
476 #endif
477 vaddr_t memhook;			/* used by mem.c & others */
478 kmutex_t memlock __cacheline_aligned;	/* used by mem.c & others */
479 kmutex_t pmap_lock __cacheline_aligned;
480 kmutex_t kpm_lock __cacheline_aligned;
481 extern void *msgbufaddr;
482 int pmap_kmpages;
483 /*
484  * Flag to indicate if pmap_init() has done its thing
485  */
486 bool pmap_initialized;
487 
488 #if defined(ARM_MMU_EXTENDED) && defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
489 /*
490  * Virtual end of direct-mapped memory
491  */
492 vaddr_t pmap_directlimit;
493 #endif
494 
495 /*
496  * Misc. locking data structures
497  */
498 
499 static inline void
500 pmap_acquire_pmap_lock(pmap_t pm)
501 {
502 #if defined(MULTIPROCESSOR) && defined(DDB)
503 	if (__predict_false(db_onproc != NULL))
504 		return;
505 #endif
506 
507 	mutex_enter(&pm->pm_lock);
508 }
509 
510 static inline void
511 pmap_release_pmap_lock(pmap_t pm)
512 {
513 #if defined(MULTIPROCESSOR) && defined(DDB)
514 	if (__predict_false(db_onproc != NULL))
515 		return;
516 #endif
517 	mutex_exit(&pm->pm_lock);
518 }
519 
520 static inline void
521 pmap_acquire_page_lock(struct vm_page_md *md)
522 {
523 	mutex_enter(&pmap_lock);
524 }
525 
526 static inline void
527 pmap_release_page_lock(struct vm_page_md *md)
528 {
529 	mutex_exit(&pmap_lock);
530 }
531 
532 #ifdef DIAGNOSTIC
533 static inline int
534 pmap_page_locked_p(struct vm_page_md *md)
535 {
536 	return mutex_owned(&pmap_lock);
537 }
538 #endif
539 
540 
541 /*
542  * Metadata for L1 translation tables.
543  */
544 #ifndef ARM_MMU_EXTENDED
545 struct l1_ttable {
546 	/* Entry on the L1 Table list */
547 	SLIST_ENTRY(l1_ttable) l1_link;
548 
549 	/* Entry on the L1 Least Recently Used list */
550 	TAILQ_ENTRY(l1_ttable) l1_lru;
551 
552 	/* Track how many domains are allocated from this L1 */
553 	volatile u_int l1_domain_use_count;
554 
555 	/*
556 	 * A free-list of domain numbers for this L1.
557 	 * We avoid using ffs() and a bitmap to track domains since ffs()
558 	 * is slow on ARM.
559 	 */
560 	uint8_t l1_domain_first;
561 	uint8_t l1_domain_free[PMAP_DOMAINS];
562 
563 	/* Physical address of this L1 page table */
564 	paddr_t l1_physaddr;
565 
566 	/* KVA of this L1 page table */
567 	pd_entry_t *l1_kva;
568 };
569 
570 /*
571  * L1 Page Tables are tracked using a Least Recently Used list.
572  *  - New L1s are allocated from the HEAD.
573  *  - Freed L1s are added to the TAIL.
574  *  - Recently accessed L1s (where an 'access' is some change to one of
575  *    the userland pmaps which owns this L1) are moved to the TAIL.
576  */
577 static TAILQ_HEAD(, l1_ttable) l1_lru_list;
578 static kmutex_t l1_lru_lock __cacheline_aligned;
579 
580 /*
581  * A list of all L1 tables
582  */
583 static SLIST_HEAD(, l1_ttable) l1_list;
584 #endif /* ARM_MMU_EXTENDED */
585 
586 /*
587  * The l2_dtable tracks L2_BUCKET_SIZE worth of L1 slots.
588  *
589  * This is normally 16MB worth L2 page descriptors for any given pmap.
590  * Reference counts are maintained for L2 descriptors so they can be
591  * freed when empty.
592  */
593 struct l2_bucket {
594 	pt_entry_t *l2b_kva;		/* KVA of L2 Descriptor Table */
595 	paddr_t l2b_pa;			/* Physical address of same */
596 	u_short l2b_l1slot;		/* This L2 table's L1 index */
597 	u_short l2b_occupancy;		/* How many active descriptors */
598 };
599 
600 struct l2_dtable {
601 	/* The number of L2 page descriptors allocated to this l2_dtable */
602 	u_int l2_occupancy;
603 
604 	/* List of L2 page descriptors */
605 	struct l2_bucket l2_bucket[L2_BUCKET_SIZE];
606 };
607 
608 /*
609  * Given an L1 table index, calculate the corresponding l2_dtable index
610  * and bucket index within the l2_dtable.
611  */
612 #define L2_BUCKET_XSHIFT	(L2_BUCKET_XLOG2 - L1_S_SHIFT)
613 #define L2_BUCKET_XFRAME	(~(vaddr_t)0 << L2_BUCKET_XLOG2)
614 #define L2_BUCKET_IDX(l1slot)	((l1slot) >> L2_BUCKET_XSHIFT)
615 #define L2_IDX(l1slot)		(L2_BUCKET_IDX(l1slot) >> L2_BUCKET_LOG2)
616 #define L2_BUCKET(l1slot)	(L2_BUCKET_IDX(l1slot) & (L2_BUCKET_SIZE - 1))
617 
618 __CTASSERT(0x100000000ULL == ((uint64_t)L2_SIZE * L2_BUCKET_SIZE * L1_S_SIZE));
619 __CTASSERT(L2_BUCKET_XFRAME == ~(L2_BUCKET_XSIZE-1));
620 
621 /*
622  * Given a virtual address, this macro returns the
623  * virtual address required to drop into the next L2 bucket.
624  */
625 #define	L2_NEXT_BUCKET_VA(va)	(((va) & L2_BUCKET_XFRAME) + L2_BUCKET_XSIZE)
626 
627 /*
628  * L2 allocation.
629  */
630 #define	pmap_alloc_l2_dtable()		\
631 	    pool_cache_get(&pmap_l2dtable_cache, PR_NOWAIT)
632 #define	pmap_free_l2_dtable(l2)		\
633 	    pool_cache_put(&pmap_l2dtable_cache, (l2))
634 #define pmap_alloc_l2_ptp(pap)		\
635 	    ((pt_entry_t *)pool_cache_get_paddr(&pmap_l2ptp_cache,\
636 	    PR_NOWAIT, (pap)))
637 
638 /*
639  * We try to map the page tables write-through, if possible.  However, not
640  * all CPUs have a write-through cache mode, so on those we have to sync
641  * the cache when we frob page tables.
642  *
643  * We try to evaluate this at compile time, if possible.  However, it's
644  * not always possible to do that, hence this run-time var.
645  */
646 int	pmap_needs_pte_sync;
647 
648 /*
649  * Real definition of pv_entry.
650  */
651 struct pv_entry {
652 	SLIST_ENTRY(pv_entry) pv_link;	/* next pv_entry */
653 	pmap_t		pv_pmap;        /* pmap where mapping lies */
654 	vaddr_t		pv_va;          /* virtual address for mapping */
655 	u_int		pv_flags;       /* flags */
656 };
657 
658 /*
659  * Macros to determine if a mapping might be resident in the
660  * instruction/data cache and/or TLB
661  */
662 #if ARM_MMU_V7 > 0 && !defined(ARM_MMU_EXTENDED)
663 /*
664  * Speculative loads by Cortex cores can cause TLB entries to be filled even if
665  * there are no explicit accesses, so there may be always be TLB entries to
666  * flush.  If we used ASIDs then this would not be a problem.
667  */
668 #define	PV_BEEN_EXECD(f)  (((f) & PVF_EXEC) == PVF_EXEC)
669 #define	PV_BEEN_REFD(f)   (true)
670 #else
671 #define	PV_BEEN_EXECD(f)  (((f) & (PVF_REF | PVF_EXEC)) == (PVF_REF | PVF_EXEC))
672 #define	PV_BEEN_REFD(f)   (((f) & PVF_REF) != 0)
673 #endif
674 #define	PV_IS_EXEC_P(f)   (((f) & PVF_EXEC) != 0)
675 #define	PV_IS_KENTRY_P(f) (((f) & PVF_KENTRY) != 0)
676 #define	PV_IS_WRITE_P(f)  (((f) & PVF_WRITE) != 0)
677 
678 /*
679  * Local prototypes
680  */
681 static bool		pmap_set_pt_cache_mode(pd_entry_t *, vaddr_t, size_t);
682 static void		pmap_alloc_specials(vaddr_t *, int, vaddr_t *,
683 			    pt_entry_t **);
684 static bool		pmap_is_current(pmap_t) __unused;
685 static bool		pmap_is_cached(pmap_t);
686 static void		pmap_enter_pv(struct vm_page_md *, paddr_t, struct pv_entry *,
687 			    pmap_t, vaddr_t, u_int);
688 static struct pv_entry *pmap_find_pv(struct vm_page_md *, pmap_t, vaddr_t);
689 static struct pv_entry *pmap_remove_pv(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
690 static u_int		pmap_modify_pv(struct vm_page_md *, paddr_t, pmap_t, vaddr_t,
691 			    u_int, u_int);
692 
693 static void		pmap_pinit(pmap_t);
694 static int		pmap_pmap_ctor(void *, void *, int);
695 
696 static void		pmap_alloc_l1(pmap_t);
697 static void		pmap_free_l1(pmap_t);
698 #ifndef ARM_MMU_EXTENDED
699 static void		pmap_use_l1(pmap_t);
700 #endif
701 
702 static struct l2_bucket *pmap_get_l2_bucket(pmap_t, vaddr_t);
703 static struct l2_bucket *pmap_alloc_l2_bucket(pmap_t, vaddr_t);
704 static void		pmap_free_l2_bucket(pmap_t, struct l2_bucket *, u_int);
705 static int		pmap_l2ptp_ctor(void *, void *, int);
706 static int		pmap_l2dtable_ctor(void *, void *, int);
707 
708 static void		pmap_vac_me_harder(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
709 #ifdef PMAP_CACHE_VIVT
710 static void		pmap_vac_me_kpmap(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
711 static void		pmap_vac_me_user(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
712 #endif
713 
714 static void		pmap_clearbit(struct vm_page_md *, paddr_t, u_int);
715 #ifdef PMAP_CACHE_VIVT
716 static bool		pmap_clean_page(struct vm_page_md *, bool);
717 #endif
718 #ifdef PMAP_CACHE_VIPT
719 static void		pmap_syncicache_page(struct vm_page_md *, paddr_t);
720 enum pmap_flush_op {
721 	PMAP_FLUSH_PRIMARY,
722 	PMAP_FLUSH_SECONDARY,
723 	PMAP_CLEAN_PRIMARY
724 };
725 #ifndef ARM_MMU_EXTENDED
726 static void		pmap_flush_page(struct vm_page_md *, paddr_t, enum pmap_flush_op);
727 #endif
728 #endif
729 static void		pmap_page_remove(struct vm_page_md *, paddr_t);
730 static void		pmap_pv_remove(paddr_t);
731 
732 #ifndef ARM_MMU_EXTENDED
733 static void		pmap_init_l1(struct l1_ttable *, pd_entry_t *);
734 #endif
735 static vaddr_t		kernel_pt_lookup(paddr_t);
736 
737 #ifdef ARM_MMU_EXTENDED
738 static struct pool_cache pmap_l1tt_cache;
739 
740 static int		pmap_l1tt_ctor(void *, void *, int);
741 static void *		pmap_l1tt_alloc(struct pool *, int);
742 static void		pmap_l1tt_free(struct pool *, void *);
743 
744 static struct pool_allocator pmap_l1tt_allocator = {
745 	.pa_alloc = pmap_l1tt_alloc,
746 	.pa_free = pmap_l1tt_free,
747 	.pa_pagesz = L1TT_SIZE,
748 };
749 #endif
750 
751 /*
752  * Misc variables
753  */
754 vaddr_t virtual_avail;
755 vaddr_t virtual_end;
756 vaddr_t pmap_curmaxkvaddr;
757 
758 paddr_t avail_start;
759 paddr_t avail_end;
760 
761 pv_addrqh_t pmap_boot_freeq = SLIST_HEAD_INITIALIZER(&pmap_boot_freeq);
762 pv_addr_t kernelpages;
763 pv_addr_t kernel_l1pt;
764 pv_addr_t systempage;
765 
766 #ifdef PMAP_CACHE_VIPT
767 #define PMAP_VALIDATE_MD_PAGE(md)	\
768 	KASSERTMSG(arm_cache_prefer_mask == 0 || (((md)->pvh_attrs & PVF_WRITE) == 0) == ((md)->urw_mappings + (md)->krw_mappings == 0), \
769 	    "(md) %p: attrs=%#x urw=%u krw=%u", (md), \
770 	    (md)->pvh_attrs, (md)->urw_mappings, (md)->krw_mappings);
771 #endif /* PMAP_CACHE_VIPT */
772 /*
773  * A bunch of routines to conditionally flush the caches/TLB depending
774  * on whether the specified pmap actually needs to be flushed at any
775  * given time.
776  */
777 static inline void
778 pmap_tlb_flush_SE(pmap_t pm, vaddr_t va, u_int flags)
779 {
780 #ifdef ARM_MMU_EXTENDED
781 	pmap_tlb_invalidate_addr(pm, va);
782 #else
783 	if (pm->pm_cstate.cs_tlb_id != 0) {
784 		if (PV_BEEN_EXECD(flags)) {
785 			cpu_tlb_flushID_SE(va);
786 		} else if (PV_BEEN_REFD(flags)) {
787 			cpu_tlb_flushD_SE(va);
788 		}
789 	}
790 #endif /* ARM_MMU_EXTENDED */
791 }
792 
793 #ifndef ARM_MMU_EXTENDED
794 static inline void
795 pmap_tlb_flushID(pmap_t pm)
796 {
797 	if (pm->pm_cstate.cs_tlb_id) {
798 		cpu_tlb_flushID();
799 #if ARM_MMU_V7 == 0
800 		/*
801 		 * Speculative loads by Cortex cores can cause TLB entries to
802 		 * be filled even if there are no explicit accesses, so there
803 		 * may be always be TLB entries to flush.  If we used ASIDs
804 		 * then it would not be a problem.
805 		 * This is not true for other CPUs.
806 		 */
807 		pm->pm_cstate.cs_tlb = 0;
808 #endif /* ARM_MMU_V7 */
809 	}
810 }
811 
812 static inline void
813 pmap_tlb_flushD(pmap_t pm)
814 {
815 	if (pm->pm_cstate.cs_tlb_d) {
816 		cpu_tlb_flushD();
817 #if ARM_MMU_V7 == 0
818 		/*
819 		 * Speculative loads by Cortex cores can cause TLB entries to
820 		 * be filled even if there are no explicit accesses, so there
821 		 * may be always be TLB entries to flush.  If we used ASIDs
822 		 * then it would not be a problem.
823 		 * This is not true for other CPUs.
824 		 */
825 		pm->pm_cstate.cs_tlb_d = 0;
826 #endif /* ARM_MMU_V7 */
827 	}
828 }
829 #endif /* ARM_MMU_EXTENDED */
830 
831 #ifdef PMAP_CACHE_VIVT
832 static inline void
833 pmap_cache_wbinv_page(pmap_t pm, vaddr_t va, bool do_inv, u_int flags)
834 {
835 	if (PV_BEEN_EXECD(flags) && pm->pm_cstate.cs_cache_id) {
836 		cpu_idcache_wbinv_range(va, PAGE_SIZE);
837 	} else if (PV_BEEN_REFD(flags) && pm->pm_cstate.cs_cache_d) {
838 		if (do_inv) {
839 			if (flags & PVF_WRITE)
840 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
841 			else
842 				cpu_dcache_inv_range(va, PAGE_SIZE);
843 		} else if (flags & PVF_WRITE) {
844 			cpu_dcache_wb_range(va, PAGE_SIZE);
845 		}
846 	}
847 }
848 
849 static inline void
850 pmap_cache_wbinv_all(pmap_t pm, u_int flags)
851 {
852 	if (PV_BEEN_EXECD(flags)) {
853 		if (pm->pm_cstate.cs_cache_id) {
854 			cpu_idcache_wbinv_all();
855 			pm->pm_cstate.cs_cache = 0;
856 		}
857 	} else if (pm->pm_cstate.cs_cache_d) {
858 		cpu_dcache_wbinv_all();
859 		pm->pm_cstate.cs_cache_d = 0;
860 	}
861 }
862 #endif /* PMAP_CACHE_VIVT */
863 
864 static inline uint8_t
865 pmap_domain(pmap_t pm)
866 {
867 #ifdef ARM_MMU_EXTENDED
868 	return pm == pmap_kernel() ? PMAP_DOMAIN_KERNEL : PMAP_DOMAIN_USER;
869 #else
870 	return pm->pm_domain;
871 #endif
872 }
873 
874 static inline pd_entry_t *
875 pmap_l1_kva(pmap_t pm)
876 {
877 #ifdef ARM_MMU_EXTENDED
878 	return pm->pm_l1;
879 #else
880 	return pm->pm_l1->l1_kva;
881 #endif
882 }
883 
884 static inline bool
885 pmap_is_current(pmap_t pm)
886 {
887 	if (pm == pmap_kernel() || curproc->p_vmspace->vm_map.pmap == pm)
888 		return true;
889 
890 	return false;
891 }
892 
893 static inline bool
894 pmap_is_cached(pmap_t pm)
895 {
896 #ifdef ARM_MMU_EXTENDED
897 	if (pm == pmap_kernel())
898 		return true;
899 #ifdef MULTIPROCESSOR
900 	// Is this pmap active on any CPU?
901 	if (!kcpuset_iszero(pm->pm_active))
902 		return true;
903 #else
904 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
905 	// Is this pmap active?
906 	if (PMAP_PAI_ASIDVALID_P(PMAP_PAI(pm, ti), ti))
907 		return true;
908 #endif
909 #else
910 	struct cpu_info * const ci = curcpu();
911 	if (pm == pmap_kernel() || ci->ci_pmap_lastuser == NULL
912 	    || ci->ci_pmap_lastuser == pm)
913 		return true;
914 #endif /* ARM_MMU_EXTENDED */
915 
916 	return false;
917 }
918 
919 /*
920  * PTE_SYNC_CURRENT:
921  *
922  *     Make sure the pte is written out to RAM.
923  *     We need to do this for one of two cases:
924  *       - We're dealing with the kernel pmap
925  *       - There is no pmap active in the cache/tlb.
926  *       - The specified pmap is 'active' in the cache/tlb.
927  */
928 
929 #ifdef PMAP_INCLUDE_PTE_SYNC
930 static inline void
931 pmap_pte_sync_current(pmap_t pm, pt_entry_t *ptep)
932 {
933 	if (PMAP_NEEDS_PTE_SYNC && pmap_is_cached(pm))
934 		PTE_SYNC(ptep);
935 	arm_dsb();
936 }
937 
938 # define PTE_SYNC_CURRENT(pm, ptep)	pmap_pte_sync_current(pm, ptep)
939 #else
940 # define PTE_SYNC_CURRENT(pm, ptep)	__nothing
941 #endif
942 
943 /*
944  * main pv_entry manipulation functions:
945  *   pmap_enter_pv: enter a mapping onto a vm_page list
946  *   pmap_remove_pv: remove a mapping from a vm_page list
947  *
948  * NOTE: pmap_enter_pv expects to lock the pvh itself
949  *       pmap_remove_pv expects the caller to lock the pvh before calling
950  */
951 
952 /*
953  * pmap_enter_pv: enter a mapping onto a vm_page lst
954  *
955  * => caller should hold the proper lock on pmap_main_lock
956  * => caller should have pmap locked
957  * => we will gain the lock on the vm_page and allocate the new pv_entry
958  * => caller should adjust ptp's wire_count before calling
959  * => caller should not adjust pmap's wire_count
960  */
961 static void
962 pmap_enter_pv(struct vm_page_md *md, paddr_t pa, struct pv_entry *pv, pmap_t pm,
963     vaddr_t va, u_int flags)
964 {
965 	UVMHIST_FUNC(__func__);
966 	UVMHIST_CALLARGS(maphist, "md %#jx pa %#jx pm %#jx va %#jx",
967 	    (uintptr_t)md, (uintptr_t)pa, (uintptr_t)pm, va);
968 	UVMHIST_LOG(maphist, "...pv %#jx flags %#jx",
969 	    (uintptr_t)pv, flags, 0, 0);
970 
971 	struct pv_entry **pvp;
972 
973 	pv->pv_pmap = pm;
974 	pv->pv_va = va;
975 	pv->pv_flags = flags;
976 
977 	pvp = &SLIST_FIRST(&md->pvh_list);
978 #ifdef PMAP_CACHE_VIPT
979 	/*
980 	 * Insert unmanaged entries, writeable first, at the head of
981 	 * the pv list.
982 	 */
983 	if (__predict_true(!PV_IS_KENTRY_P(flags))) {
984 		while (*pvp != NULL && PV_IS_KENTRY_P((*pvp)->pv_flags))
985 			pvp = &SLIST_NEXT(*pvp, pv_link);
986 	}
987 	if (!PV_IS_WRITE_P(flags)) {
988 		while (*pvp != NULL && PV_IS_WRITE_P((*pvp)->pv_flags))
989 			pvp = &SLIST_NEXT(*pvp, pv_link);
990 	}
991 #endif
992 	SLIST_NEXT(pv, pv_link) = *pvp;		/* add to ... */
993 	*pvp = pv;				/* ... locked list */
994 	md->pvh_attrs |= flags & (PVF_REF | PVF_MOD);
995 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
996 	if ((pv->pv_flags & PVF_KWRITE) == PVF_KWRITE)
997 		md->pvh_attrs |= PVF_KMOD;
998 	if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) != PVF_NC)
999 		md->pvh_attrs |= PVF_DIRTY;
1000 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1001 #endif
1002 	if (pm == pmap_kernel()) {
1003 		PMAPCOUNT(kernel_mappings);
1004 		if (flags & PVF_WRITE)
1005 			md->krw_mappings++;
1006 		else
1007 			md->kro_mappings++;
1008 	} else {
1009 		if (flags & PVF_WRITE)
1010 			md->urw_mappings++;
1011 		else
1012 			md->uro_mappings++;
1013 	}
1014 
1015 #ifdef PMAP_CACHE_VIPT
1016 #ifndef ARM_MMU_EXTENDED
1017 	/*
1018 	 * Even though pmap_vac_me_harder will set PVF_WRITE for us,
1019 	 * do it here as well to keep the mappings & KVF_WRITE consistent.
1020 	 */
1021 	if (arm_cache_prefer_mask != 0 && (flags & PVF_WRITE) != 0) {
1022 		md->pvh_attrs |= PVF_WRITE;
1023 	}
1024 #endif
1025 	/*
1026 	 * If this is an exec mapping and its the first exec mapping
1027 	 * for this page, make sure to sync the I-cache.
1028 	 */
1029 	if (PV_IS_EXEC_P(flags)) {
1030 		if (!PV_IS_EXEC_P(md->pvh_attrs)) {
1031 			pmap_syncicache_page(md, pa);
1032 			PMAPCOUNT(exec_synced_map);
1033 		}
1034 		PMAPCOUNT(exec_mappings);
1035 	}
1036 #endif
1037 
1038 	PMAPCOUNT(mappings);
1039 
1040 	if (pv->pv_flags & PVF_WIRED)
1041 		++pm->pm_stats.wired_count;
1042 }
1043 
1044 /*
1045  *
1046  * pmap_find_pv: Find a pv entry
1047  *
1048  * => caller should hold lock on vm_page
1049  */
1050 static inline struct pv_entry *
1051 pmap_find_pv(struct vm_page_md *md, pmap_t pm, vaddr_t va)
1052 {
1053 	struct pv_entry *pv;
1054 
1055 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1056 		if (pm == pv->pv_pmap && va == pv->pv_va)
1057 			break;
1058 	}
1059 
1060 	return pv;
1061 }
1062 
1063 /*
1064  * pmap_remove_pv: try to remove a mapping from a pv_list
1065  *
1066  * => caller should hold proper lock on pmap_main_lock
1067  * => pmap should be locked
1068  * => caller should hold lock on vm_page [so that attrs can be adjusted]
1069  * => caller should adjust ptp's wire_count and free PTP if needed
1070  * => caller should NOT adjust pmap's wire_count
1071  * => we return the removed pv
1072  */
1073 static struct pv_entry *
1074 pmap_remove_pv(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1075 {
1076 	UVMHIST_FUNC(__func__);
1077 	UVMHIST_CALLARGS(maphist, "md %#jx pa %#jx pm %#jx va %#jx",
1078 	    (uintptr_t)md, (uintptr_t)pa, (uintptr_t)pm, va);
1079 
1080 	struct pv_entry *pv, **prevptr;
1081 
1082 	prevptr = &SLIST_FIRST(&md->pvh_list); /* prev pv_entry ptr */
1083 	pv = *prevptr;
1084 
1085 	while (pv) {
1086 		if (pv->pv_pmap == pm && pv->pv_va == va) {	/* match? */
1087 			UVMHIST_LOG(maphist, "pm %#jx md %#jx flags %#jx",
1088 			    (uintptr_t)pm, (uintptr_t)md, pv->pv_flags, 0);
1089 			if (pv->pv_flags & PVF_WIRED) {
1090 				--pm->pm_stats.wired_count;
1091 			}
1092 			*prevptr = SLIST_NEXT(pv, pv_link);	/* remove it! */
1093 			if (pm == pmap_kernel()) {
1094 				PMAPCOUNT(kernel_unmappings);
1095 				if (pv->pv_flags & PVF_WRITE)
1096 					md->krw_mappings--;
1097 				else
1098 					md->kro_mappings--;
1099 			} else {
1100 				if (pv->pv_flags & PVF_WRITE)
1101 					md->urw_mappings--;
1102 				else
1103 					md->uro_mappings--;
1104 			}
1105 
1106 			PMAPCOUNT(unmappings);
1107 #ifdef PMAP_CACHE_VIPT
1108 			/*
1109 			 * If this page has had an exec mapping, then if
1110 			 * this was the last mapping, discard the contents,
1111 			 * otherwise sync the i-cache for this page.
1112 			 */
1113 			if (PV_IS_EXEC_P(md->pvh_attrs)) {
1114 				if (SLIST_EMPTY(&md->pvh_list)) {
1115 					md->pvh_attrs &= ~PVF_EXEC;
1116 					PMAPCOUNT(exec_discarded_unmap);
1117 				} else if (pv->pv_flags & PVF_WRITE) {
1118 					pmap_syncicache_page(md, pa);
1119 					PMAPCOUNT(exec_synced_unmap);
1120 				}
1121 			}
1122 #endif /* PMAP_CACHE_VIPT */
1123 			break;
1124 		}
1125 		prevptr = &SLIST_NEXT(pv, pv_link);	/* previous pointer */
1126 		pv = *prevptr;				/* advance */
1127 	}
1128 
1129 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
1130 	/*
1131 	 * If we no longer have a WRITEABLE KENTRY at the head of list,
1132 	 * clear the KMOD attribute from the page.
1133 	 */
1134 	if (SLIST_FIRST(&md->pvh_list) == NULL
1135 	    || (SLIST_FIRST(&md->pvh_list)->pv_flags & PVF_KWRITE) != PVF_KWRITE)
1136 		md->pvh_attrs &= ~PVF_KMOD;
1137 
1138 	/*
1139 	 * If this was a writeable page and there are no more writeable
1140 	 * mappings (ignoring KMPAGE), clear the WRITE flag and writeback
1141 	 * the contents to memory.
1142 	 */
1143 	if (arm_cache_prefer_mask != 0) {
1144 		if (md->krw_mappings + md->urw_mappings == 0)
1145 			md->pvh_attrs &= ~PVF_WRITE;
1146 		PMAP_VALIDATE_MD_PAGE(md);
1147 	}
1148 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1149 #endif /* PMAP_CACHE_VIPT && !ARM_MMU_EXTENDED */
1150 
1151 	/* return removed pv */
1152 	return pv;
1153 }
1154 
1155 /*
1156  *
1157  * pmap_modify_pv: Update pv flags
1158  *
1159  * => caller should hold lock on vm_page [so that attrs can be adjusted]
1160  * => caller should NOT adjust pmap's wire_count
1161  * => caller must call pmap_vac_me_harder() if writable status of a page
1162  *    may have changed.
1163  * => we return the old flags
1164  *
1165  * Modify a physical-virtual mapping in the pv table
1166  */
1167 static u_int
1168 pmap_modify_pv(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va,
1169     u_int clr_mask, u_int set_mask)
1170 {
1171 	struct pv_entry *npv;
1172 	u_int flags, oflags;
1173 	UVMHIST_FUNC(__func__);
1174 	UVMHIST_CALLARGS(maphist, "md %#jx pa %#jx pm %#jx va %#jx",
1175 	    (uintptr_t)md, (uintptr_t)pa, (uintptr_t)pm, va);
1176 	UVMHIST_LOG(maphist, "... clr %#jx set %#jx", clr_mask, set_mask, 0, 0);
1177 
1178 	KASSERT(!PV_IS_KENTRY_P(clr_mask));
1179 	KASSERT(!PV_IS_KENTRY_P(set_mask));
1180 
1181 	if ((npv = pmap_find_pv(md, pm, va)) == NULL) {
1182 		UVMHIST_LOG(maphist, "<--- done (not found)", 0, 0, 0, 0);
1183 		return 0;
1184 	}
1185 
1186 	/*
1187 	 * There is at least one VA mapping this page.
1188 	 */
1189 
1190 	if (clr_mask & (PVF_REF | PVF_MOD)) {
1191 		md->pvh_attrs |= set_mask & (PVF_REF | PVF_MOD);
1192 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
1193 		if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) != PVF_NC)
1194 			md->pvh_attrs |= PVF_DIRTY;
1195 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1196 #endif /* PMAP_CACHE_VIPT && !ARM_MMU_EXTENDED */
1197 	}
1198 
1199 	oflags = npv->pv_flags;
1200 	npv->pv_flags = flags = (oflags & ~clr_mask) | set_mask;
1201 
1202 	if ((flags ^ oflags) & PVF_WIRED) {
1203 		if (flags & PVF_WIRED)
1204 			++pm->pm_stats.wired_count;
1205 		else
1206 			--pm->pm_stats.wired_count;
1207 	}
1208 
1209 	if ((flags ^ oflags) & PVF_WRITE) {
1210 		if (pm == pmap_kernel()) {
1211 			if (flags & PVF_WRITE) {
1212 				md->krw_mappings++;
1213 				md->kro_mappings--;
1214 			} else {
1215 				md->kro_mappings++;
1216 				md->krw_mappings--;
1217 			}
1218 		} else {
1219 			if (flags & PVF_WRITE) {
1220 				md->urw_mappings++;
1221 				md->uro_mappings--;
1222 			} else {
1223 				md->uro_mappings++;
1224 				md->urw_mappings--;
1225 			}
1226 		}
1227 	}
1228 #ifdef PMAP_CACHE_VIPT
1229 	if (arm_cache_prefer_mask != 0) {
1230 		if (md->urw_mappings + md->krw_mappings == 0) {
1231 			md->pvh_attrs &= ~PVF_WRITE;
1232 		} else {
1233 			md->pvh_attrs |= PVF_WRITE;
1234 		}
1235 	}
1236 	/*
1237 	 * We have two cases here: the first is from enter_pv (new exec
1238 	 * page), the second is a combined pmap_remove_pv/pmap_enter_pv.
1239 	 * Since in latter, pmap_enter_pv won't do anything, we just have
1240 	 * to do what pmap_remove_pv would do.
1241 	 */
1242 	if ((PV_IS_EXEC_P(flags) && !PV_IS_EXEC_P(md->pvh_attrs))
1243 	    || (PV_IS_EXEC_P(md->pvh_attrs)
1244 		|| (!(flags & PVF_WRITE) && (oflags & PVF_WRITE)))) {
1245 		pmap_syncicache_page(md, pa);
1246 		PMAPCOUNT(exec_synced_remap);
1247 	}
1248 #ifndef ARM_MMU_EXTENDED
1249 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1250 #endif /* !ARM_MMU_EXTENDED */
1251 #endif /* PMAP_CACHE_VIPT */
1252 
1253 	PMAPCOUNT(remappings);
1254 
1255 	UVMHIST_LOG(maphist, "<--- done", 0, 0, 0, 0);
1256 
1257 	return oflags;
1258 }
1259 
1260 
1261 #if defined(ARM_MMU_EXTENDED)
1262 int
1263 pmap_maxproc_set(int nmaxproc)
1264 {
1265 	static const char pmap_l1ttpool_warnmsg[] =
1266 	    "WARNING: l1ttpool limit reached; increase kern.maxproc";
1267 
1268 	pool_cache_prime(&pmap_l1tt_cache, nmaxproc);
1269 
1270 	/*
1271 	 * Set the hard limit on the pmap_l1tt_cache to the number
1272 	 * of processes the kernel is to support.  Log the limit
1273 	 * reached message max once a minute.
1274 	 */
1275 	pool_cache_sethardlimit(&pmap_l1tt_cache, nmaxproc,
1276 	    pmap_l1ttpool_warnmsg, 60);
1277 
1278 	return 0;
1279 }
1280 
1281 #endif
1282 
1283 /*
1284  * Allocate an L1 translation table for the specified pmap.
1285  * This is called at pmap creation time.
1286  */
1287 static void
1288 pmap_alloc_l1(pmap_t pm)
1289 {
1290 #ifdef ARM_MMU_EXTENDED
1291 	vaddr_t va = (vaddr_t)pool_cache_get_paddr(&pmap_l1tt_cache, PR_WAITOK,
1292 	    &pm->pm_l1_pa);
1293 
1294 	pm->pm_l1 = (pd_entry_t *)va;
1295 	PTE_SYNC_RANGE(pm->pm_l1, L1TT_SIZE / sizeof(pt_entry_t));
1296 #else
1297 	struct l1_ttable *l1;
1298 	uint8_t domain;
1299 
1300 	/*
1301 	 * Remove the L1 at the head of the LRU list
1302 	 */
1303 	mutex_spin_enter(&l1_lru_lock);
1304 	l1 = TAILQ_FIRST(&l1_lru_list);
1305 	KDASSERT(l1 != NULL);
1306 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1307 
1308 	/*
1309 	 * Pick the first available domain number, and update
1310 	 * the link to the next number.
1311 	 */
1312 	domain = l1->l1_domain_first;
1313 	l1->l1_domain_first = l1->l1_domain_free[domain];
1314 
1315 	/*
1316 	 * If there are still free domain numbers in this L1,
1317 	 * put it back on the TAIL of the LRU list.
1318 	 */
1319 	if (++l1->l1_domain_use_count < PMAP_DOMAINS)
1320 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1321 
1322 	mutex_spin_exit(&l1_lru_lock);
1323 
1324 	/*
1325 	 * Fix up the relevant bits in the pmap structure
1326 	 */
1327 	pm->pm_l1 = l1;
1328 	pm->pm_domain = domain + 1;
1329 #endif
1330 }
1331 
1332 /*
1333  * Free an L1 translation table.
1334  * This is called at pmap destruction time.
1335  */
1336 static void
1337 pmap_free_l1(pmap_t pm)
1338 {
1339 #ifdef ARM_MMU_EXTENDED
1340 	pool_cache_put_paddr(&pmap_l1tt_cache, (void *)pm->pm_l1, pm->pm_l1_pa);
1341 
1342 	pm->pm_l1 = NULL;
1343 	pm->pm_l1_pa = 0;
1344 #else
1345 	struct l1_ttable *l1 = pm->pm_l1;
1346 
1347 	mutex_spin_enter(&l1_lru_lock);
1348 
1349 	/*
1350 	 * If this L1 is currently on the LRU list, remove it.
1351 	 */
1352 	if (l1->l1_domain_use_count < PMAP_DOMAINS)
1353 		TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1354 
1355 	/*
1356 	 * Free up the domain number which was allocated to the pmap
1357 	 */
1358 	l1->l1_domain_free[pmap_domain(pm) - 1] = l1->l1_domain_first;
1359 	l1->l1_domain_first = pmap_domain(pm) - 1;
1360 	l1->l1_domain_use_count--;
1361 
1362 	/*
1363 	 * The L1 now must have at least 1 free domain, so add
1364 	 * it back to the LRU list. If the use count is zero,
1365 	 * put it at the head of the list, otherwise it goes
1366 	 * to the tail.
1367 	 */
1368 	if (l1->l1_domain_use_count == 0)
1369 		TAILQ_INSERT_HEAD(&l1_lru_list, l1, l1_lru);
1370 	else
1371 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1372 
1373 	mutex_spin_exit(&l1_lru_lock);
1374 #endif /* ARM_MMU_EXTENDED */
1375 }
1376 
1377 #ifndef ARM_MMU_EXTENDED
1378 static inline void
1379 pmap_use_l1(pmap_t pm)
1380 {
1381 	struct l1_ttable *l1;
1382 
1383 	/*
1384 	 * Do nothing if we're in interrupt context.
1385 	 * Access to an L1 by the kernel pmap must not affect
1386 	 * the LRU list.
1387 	 */
1388 	if (cpu_intr_p() || pm == pmap_kernel())
1389 		return;
1390 
1391 	l1 = pm->pm_l1;
1392 
1393 	/*
1394 	 * If the L1 is not currently on the LRU list, just return
1395 	 */
1396 	if (l1->l1_domain_use_count == PMAP_DOMAINS)
1397 		return;
1398 
1399 	mutex_spin_enter(&l1_lru_lock);
1400 
1401 	/*
1402 	 * Check the use count again, now that we've acquired the lock
1403 	 */
1404 	if (l1->l1_domain_use_count == PMAP_DOMAINS) {
1405 		mutex_spin_exit(&l1_lru_lock);
1406 		return;
1407 	}
1408 
1409 	/*
1410 	 * Move the L1 to the back of the LRU list
1411 	 */
1412 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1413 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1414 
1415 	mutex_spin_exit(&l1_lru_lock);
1416 }
1417 #endif /* !ARM_MMU_EXTENDED */
1418 
1419 /*
1420  * void pmap_free_l2_ptp(pt_entry_t *, paddr_t *)
1421  *
1422  * Free an L2 descriptor table.
1423  */
1424 static inline void
1425 #if defined(PMAP_INCLUDE_PTE_SYNC) && defined(PMAP_CACHE_VIVT)
1426 pmap_free_l2_ptp(bool need_sync, pt_entry_t *l2, paddr_t pa)
1427 #else
1428 pmap_free_l2_ptp(pt_entry_t *l2, paddr_t pa)
1429 #endif
1430 {
1431 #if defined(PMAP_INCLUDE_PTE_SYNC) && defined(PMAP_CACHE_VIVT)
1432 	/*
1433 	 * Note: With a write-back cache, we may need to sync this
1434 	 * L2 table before re-using it.
1435 	 * This is because it may have belonged to a non-current
1436 	 * pmap, in which case the cache syncs would have been
1437 	 * skipped for the pages that were being unmapped. If the
1438 	 * L2 table were then to be immediately re-allocated to
1439 	 * the *current* pmap, it may well contain stale mappings
1440 	 * which have not yet been cleared by a cache write-back
1441 	 * and so would still be visible to the mmu.
1442 	 */
1443 	if (need_sync)
1444 		PTE_SYNC_RANGE(l2, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1445 #endif /* PMAP_INCLUDE_PTE_SYNC && PMAP_CACHE_VIVT */
1446 	pool_cache_put_paddr(&pmap_l2ptp_cache, (void *)l2, pa);
1447 }
1448 
1449 /*
1450  * Returns a pointer to the L2 bucket associated with the specified pmap
1451  * and VA, or NULL if no L2 bucket exists for the address.
1452  */
1453 static inline struct l2_bucket *
1454 pmap_get_l2_bucket(pmap_t pm, vaddr_t va)
1455 {
1456 	const size_t l1slot = l1pte_index(va);
1457 	struct l2_dtable *l2;
1458 	struct l2_bucket *l2b;
1459 
1460 	if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL ||
1461 	    (l2b = &l2->l2_bucket[L2_BUCKET(l1slot)])->l2b_kva == NULL)
1462 		return NULL;
1463 
1464 	return l2b;
1465 }
1466 
1467 /*
1468  * Returns a pointer to the L2 bucket associated with the specified pmap
1469  * and VA.
1470  *
1471  * If no L2 bucket exists, perform the necessary allocations to put an L2
1472  * bucket/page table in place.
1473  *
1474  * Note that if a new L2 bucket/page was allocated, the caller *must*
1475  * increment the bucket occupancy counter appropriately *before*
1476  * releasing the pmap's lock to ensure no other thread or cpu deallocates
1477  * the bucket/page in the meantime.
1478  */
1479 static struct l2_bucket *
1480 pmap_alloc_l2_bucket(pmap_t pm, vaddr_t va)
1481 {
1482 	const size_t l1slot = l1pte_index(va);
1483 	struct l2_dtable *l2;
1484 
1485 	if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL) {
1486 		/*
1487 		 * No mapping at this address, as there is
1488 		 * no entry in the L1 table.
1489 		 * Need to allocate a new l2_dtable.
1490 		 */
1491 		if ((l2 = pmap_alloc_l2_dtable()) == NULL)
1492 			return NULL;
1493 
1494 		/*
1495 		 * Link it into the parent pmap
1496 		 */
1497 		pm->pm_l2[L2_IDX(l1slot)] = l2;
1498 	}
1499 
1500 	struct l2_bucket * const l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
1501 
1502 	/*
1503 	 * Fetch pointer to the L2 page table associated with the address.
1504 	 */
1505 	if (l2b->l2b_kva == NULL) {
1506 		pt_entry_t *ptep;
1507 
1508 		/*
1509 		 * No L2 page table has been allocated. Chances are, this
1510 		 * is because we just allocated the l2_dtable, above.
1511 		 */
1512 		if ((ptep = pmap_alloc_l2_ptp(&l2b->l2b_pa)) == NULL) {
1513 			/*
1514 			 * Oops, no more L2 page tables available at this
1515 			 * time. We may need to deallocate the l2_dtable
1516 			 * if we allocated a new one above.
1517 			 */
1518 			if (l2->l2_occupancy == 0) {
1519 				pm->pm_l2[L2_IDX(l1slot)] = NULL;
1520 				pmap_free_l2_dtable(l2);
1521 			}
1522 			return NULL;
1523 		}
1524 
1525 		l2->l2_occupancy++;
1526 		l2b->l2b_kva = ptep;
1527 		l2b->l2b_l1slot = l1slot;
1528 
1529 #ifdef ARM_MMU_EXTENDED
1530 		/*
1531 		 * We know there will be a mapping here, so simply
1532 		 * enter this PTP into the L1 now.
1533 		 */
1534 		pd_entry_t * const pdep = pmap_l1_kva(pm) + l1slot;
1535 		pd_entry_t npde = L1_C_PROTO | l2b->l2b_pa
1536 		    | L1_C_DOM(pmap_domain(pm));
1537 		KASSERT(*pdep == 0);
1538 		l1pte_setone(pdep, npde);
1539 		PDE_SYNC(pdep);
1540 #endif
1541 	}
1542 
1543 	return l2b;
1544 }
1545 
1546 /*
1547  * One or more mappings in the specified L2 descriptor table have just been
1548  * invalidated.
1549  *
1550  * Garbage collect the metadata and descriptor table itself if necessary.
1551  *
1552  * The pmap lock must be acquired when this is called (not necessary
1553  * for the kernel pmap).
1554  */
1555 static void
1556 pmap_free_l2_bucket(pmap_t pm, struct l2_bucket *l2b, u_int count)
1557 {
1558 	KDASSERT(count <= l2b->l2b_occupancy);
1559 
1560 	/*
1561 	 * Update the bucket's reference count according to how many
1562 	 * PTEs the caller has just invalidated.
1563 	 */
1564 	l2b->l2b_occupancy -= count;
1565 
1566 	/*
1567 	 * Note:
1568 	 *
1569 	 * Level 2 page tables allocated to the kernel pmap are never freed
1570 	 * as that would require checking all Level 1 page tables and
1571 	 * removing any references to the Level 2 page table. See also the
1572 	 * comment elsewhere about never freeing bootstrap L2 descriptors.
1573 	 *
1574 	 * We make do with just invalidating the mapping in the L2 table.
1575 	 *
1576 	 * This isn't really a big deal in practice and, in fact, leads
1577 	 * to a performance win over time as we don't need to continually
1578 	 * alloc/free.
1579 	 */
1580 	if (l2b->l2b_occupancy > 0 || pm == pmap_kernel())
1581 		return;
1582 
1583 	/*
1584 	 * There are no more valid mappings in this level 2 page table.
1585 	 * Go ahead and NULL-out the pointer in the bucket, then
1586 	 * free the page table.
1587 	 */
1588 	const size_t l1slot = l2b->l2b_l1slot;
1589 	pt_entry_t * const ptep = l2b->l2b_kva;
1590 	l2b->l2b_kva = NULL;
1591 
1592 	pd_entry_t * const pdep = pmap_l1_kva(pm) + l1slot;
1593 	pd_entry_t pde __diagused = *pdep;
1594 
1595 #ifdef ARM_MMU_EXTENDED
1596 	/*
1597 	 * Invalidate the L1 slot.
1598 	 */
1599 	KASSERT((pde & L1_TYPE_MASK) == L1_TYPE_C);
1600 #else
1601 	/*
1602 	 * If the L1 slot matches the pmap's domain number, then invalidate it.
1603 	 */
1604 	if ((pde & (L1_C_DOM_MASK|L1_TYPE_MASK))
1605 	    == (L1_C_DOM(pmap_domain(pm))|L1_TYPE_C)) {
1606 #endif
1607 		l1pte_setone(pdep, 0);
1608 		PDE_SYNC(pdep);
1609 #ifndef ARM_MMU_EXTENDED
1610 	}
1611 #endif
1612 
1613 	/*
1614 	 * Release the L2 descriptor table back to the pool cache.
1615 	 */
1616 #if defined(PMAP_INCLUDE_PTE_SYNC) && defined(PMAP_CACHE_VIVT)
1617 	pmap_free_l2_ptp(!pmap_is_cached(pm), ptep, l2b->l2b_pa);
1618 #else
1619 	pmap_free_l2_ptp(ptep, l2b->l2b_pa);
1620 #endif
1621 
1622 	/*
1623 	 * Update the reference count in the associated l2_dtable
1624 	 */
1625 	struct l2_dtable * const l2 = pm->pm_l2[L2_IDX(l1slot)];
1626 	if (--l2->l2_occupancy > 0)
1627 		return;
1628 
1629 	/*
1630 	 * There are no more valid mappings in any of the Level 1
1631 	 * slots managed by this l2_dtable. Go ahead and NULL-out
1632 	 * the pointer in the parent pmap and free the l2_dtable.
1633 	 */
1634 	pm->pm_l2[L2_IDX(l1slot)] = NULL;
1635 	pmap_free_l2_dtable(l2);
1636 }
1637 
1638 #if defined(ARM_MMU_EXTENDED)
1639 /*
1640  * Pool cache constructors for L1 translation tables
1641  */
1642 
1643 static int
1644 pmap_l1tt_ctor(void *arg, void *v, int flags)
1645 {
1646 #ifndef PMAP_INCLUDE_PTE_SYNC
1647 #error not supported
1648 #endif
1649 
1650 	memset(v, 0, L1TT_SIZE);
1651 	PTE_SYNC_RANGE(v, L1TT_SIZE / sizeof(pt_entry_t));
1652 	return 0;
1653 }
1654 #endif
1655 
1656 /*
1657  * Pool cache constructors for L2 descriptor tables, metadata and pmap
1658  * structures.
1659  */
1660 static int
1661 pmap_l2ptp_ctor(void *arg, void *v, int flags)
1662 {
1663 #ifndef PMAP_INCLUDE_PTE_SYNC
1664 	vaddr_t va = (vaddr_t)v & ~PGOFSET;
1665 
1666 	/*
1667 	 * The mappings for these page tables were initially made using
1668 	 * pmap_kenter_pa() by the pool subsystem. Therefore, the cache-
1669 	 * mode will not be right for page table mappings. To avoid
1670 	 * polluting the pmap_kenter_pa() code with a special case for
1671 	 * page tables, we simply fix up the cache-mode here if it's not
1672 	 * correct.
1673 	 */
1674 	if (pte_l2_s_cache_mode != pte_l2_s_cache_mode_pt) {
1675 		const struct l2_bucket * const l2b =
1676 		    pmap_get_l2_bucket(pmap_kernel(), va);
1677 		KASSERTMSG(l2b != NULL, "%#lx", va);
1678 		pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
1679 		const pt_entry_t opte = *ptep;
1680 
1681 		if ((opte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
1682 			/*
1683 			 * Page tables must have the cache-mode set correctly.
1684 			 */
1685 			const pt_entry_t npte = (opte & ~L2_S_CACHE_MASK)
1686 			    | pte_l2_s_cache_mode_pt;
1687 			l2pte_set(ptep, npte, opte);
1688 			PTE_SYNC(ptep);
1689 			cpu_tlb_flushD_SE(va);
1690 			cpu_cpwait();
1691 		}
1692 	}
1693 #endif
1694 
1695 	memset(v, 0, L2_TABLE_SIZE_REAL);
1696 	PTE_SYNC_RANGE(v, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1697 	return 0;
1698 }
1699 
1700 static int
1701 pmap_l2dtable_ctor(void *arg, void *v, int flags)
1702 {
1703 
1704 	memset(v, 0, sizeof(struct l2_dtable));
1705 	return 0;
1706 }
1707 
1708 static int
1709 pmap_pmap_ctor(void *arg, void *v, int flags)
1710 {
1711 
1712 	memset(v, 0, sizeof(struct pmap));
1713 	return 0;
1714 }
1715 
1716 static void
1717 pmap_pinit(pmap_t pm)
1718 {
1719 #ifndef ARM_HAS_VBAR
1720 	struct l2_bucket *l2b;
1721 
1722 	if (vector_page < KERNEL_BASE) {
1723 		/*
1724 		 * Map the vector page.
1725 		 */
1726 		pmap_enter(pm, vector_page, systempage.pv_pa,
1727 		    VM_PROT_READ | VM_PROT_EXECUTE,
1728 		    VM_PROT_READ | VM_PROT_EXECUTE | PMAP_WIRED);
1729 		pmap_update(pm);
1730 
1731 		pm->pm_pl1vec = pmap_l1_kva(pm) + l1pte_index(vector_page);
1732 		l2b = pmap_get_l2_bucket(pm, vector_page);
1733 		KASSERTMSG(l2b != NULL, "%#lx", vector_page);
1734 		pm->pm_l1vec = l2b->l2b_pa | L1_C_PROTO |
1735 		    L1_C_DOM(pmap_domain(pm));
1736 	} else
1737 		pm->pm_pl1vec = NULL;
1738 #endif
1739 }
1740 
1741 #ifdef PMAP_CACHE_VIVT
1742 /*
1743  * Since we have a virtually indexed cache, we may need to inhibit caching if
1744  * there is more than one mapping and at least one of them is writable.
1745  * Since we purge the cache on every context switch, we only need to check for
1746  * other mappings within the same pmap, or kernel_pmap.
1747  * This function is also called when a page is unmapped, to possibly reenable
1748  * caching on any remaining mappings.
1749  *
1750  * The code implements the following logic, where:
1751  *
1752  * KW = # of kernel read/write pages
1753  * KR = # of kernel read only pages
1754  * UW = # of user read/write pages
1755  * UR = # of user read only pages
1756  *
1757  * KC = kernel mapping is cacheable
1758  * UC = user mapping is cacheable
1759  *
1760  *               KW=0,KR=0  KW=0,KR>0  KW=1,KR=0  KW>1,KR>=0
1761  *             +---------------------------------------------
1762  * UW=0,UR=0   | ---        KC=1       KC=1       KC=0
1763  * UW=0,UR>0   | UC=1       KC=1,UC=1  KC=0,UC=0  KC=0,UC=0
1764  * UW=1,UR=0   | UC=1       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1765  * UW>1,UR>=0  | UC=0       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1766  */
1767 
1768 static const int pmap_vac_flags[4][4] = {
1769 	{-1,		0,		0,		PVF_KNC},
1770 	{0,		0,		PVF_NC,		PVF_NC},
1771 	{0,		PVF_NC,		PVF_NC,		PVF_NC},
1772 	{PVF_UNC,	PVF_NC,		PVF_NC,		PVF_NC}
1773 };
1774 
1775 static inline int
1776 pmap_get_vac_flags(const struct vm_page_md *md)
1777 {
1778 	int kidx, uidx;
1779 
1780 	kidx = 0;
1781 	if (md->kro_mappings || md->krw_mappings > 1)
1782 		kidx |= 1;
1783 	if (md->krw_mappings)
1784 		kidx |= 2;
1785 
1786 	uidx = 0;
1787 	if (md->uro_mappings || md->urw_mappings > 1)
1788 		uidx |= 1;
1789 	if (md->urw_mappings)
1790 		uidx |= 2;
1791 
1792 	return pmap_vac_flags[uidx][kidx];
1793 }
1794 
1795 static inline void
1796 pmap_vac_me_harder(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1797 {
1798 	int nattr;
1799 
1800 	nattr = pmap_get_vac_flags(md);
1801 
1802 	if (nattr < 0) {
1803 		md->pvh_attrs &= ~PVF_NC;
1804 		return;
1805 	}
1806 
1807 	if (nattr == 0 && (md->pvh_attrs & PVF_NC) == 0)
1808 		return;
1809 
1810 	if (pm == pmap_kernel())
1811 		pmap_vac_me_kpmap(md, pa, pm, va);
1812 	else
1813 		pmap_vac_me_user(md, pa, pm, va);
1814 
1815 	md->pvh_attrs = (md->pvh_attrs & ~PVF_NC) | nattr;
1816 }
1817 
1818 static void
1819 pmap_vac_me_kpmap(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1820 {
1821 	u_int u_cacheable, u_entries;
1822 	struct pv_entry *pv;
1823 	pmap_t last_pmap = pm;
1824 
1825 	/*
1826 	 * Pass one, see if there are both kernel and user pmaps for
1827 	 * this page.  Calculate whether there are user-writable or
1828 	 * kernel-writable pages.
1829 	 */
1830 	u_cacheable = 0;
1831 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1832 		if (pv->pv_pmap != pm && (pv->pv_flags & PVF_NC) == 0)
1833 			u_cacheable++;
1834 	}
1835 
1836 	u_entries = md->urw_mappings + md->uro_mappings;
1837 
1838 	/*
1839 	 * We know we have just been updating a kernel entry, so if
1840 	 * all user pages are already cacheable, then there is nothing
1841 	 * further to do.
1842 	 */
1843 	if (md->k_mappings == 0 && u_cacheable == u_entries)
1844 		return;
1845 
1846 	if (u_entries) {
1847 		/*
1848 		 * Scan over the list again, for each entry, if it
1849 		 * might not be set correctly, call pmap_vac_me_user
1850 		 * to recalculate the settings.
1851 		 */
1852 		SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1853 			/*
1854 			 * We know kernel mappings will get set
1855 			 * correctly in other calls.  We also know
1856 			 * that if the pmap is the same as last_pmap
1857 			 * then we've just handled this entry.
1858 			 */
1859 			if (pv->pv_pmap == pm || pv->pv_pmap == last_pmap)
1860 				continue;
1861 
1862 			/*
1863 			 * If there are kernel entries and this page
1864 			 * is writable but non-cacheable, then we can
1865 			 * skip this entry also.
1866 			 */
1867 			if (md->k_mappings &&
1868 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) ==
1869 			    (PVF_NC | PVF_WRITE))
1870 				continue;
1871 
1872 			/*
1873 			 * Similarly if there are no kernel-writable
1874 			 * entries and the page is already
1875 			 * read-only/cacheable.
1876 			 */
1877 			if (md->krw_mappings == 0 &&
1878 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 0)
1879 				continue;
1880 
1881 			/*
1882 			 * For some of the remaining cases, we know
1883 			 * that we must recalculate, but for others we
1884 			 * can't tell if they are correct or not, so
1885 			 * we recalculate anyway.
1886 			 */
1887 			pmap_vac_me_user(md, pa, (last_pmap = pv->pv_pmap), 0);
1888 		}
1889 
1890 		if (md->k_mappings == 0)
1891 			return;
1892 	}
1893 
1894 	pmap_vac_me_user(md, pa, pm, va);
1895 }
1896 
1897 static void
1898 pmap_vac_me_user(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1899 {
1900 	pmap_t kpmap = pmap_kernel();
1901 	struct pv_entry *pv, *npv = NULL;
1902 	u_int entries = 0;
1903 	u_int writable = 0;
1904 	u_int cacheable_entries = 0;
1905 	u_int kern_cacheable = 0;
1906 	u_int other_writable = 0;
1907 
1908 	/*
1909 	 * Count mappings and writable mappings in this pmap.
1910 	 * Include kernel mappings as part of our own.
1911 	 * Keep a pointer to the first one.
1912 	 */
1913 	npv = NULL;
1914 	KASSERT(pmap_page_locked_p(md));
1915 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1916 		/* Count mappings in the same pmap */
1917 		if (pm == pv->pv_pmap || kpmap == pv->pv_pmap) {
1918 			if (entries++ == 0)
1919 				npv = pv;
1920 
1921 			/* Cacheable mappings */
1922 			if ((pv->pv_flags & PVF_NC) == 0) {
1923 				cacheable_entries++;
1924 				if (kpmap == pv->pv_pmap)
1925 					kern_cacheable++;
1926 			}
1927 
1928 			/* Writable mappings */
1929 			if (pv->pv_flags & PVF_WRITE)
1930 				++writable;
1931 		} else if (pv->pv_flags & PVF_WRITE)
1932 			other_writable = 1;
1933 	}
1934 
1935 	/*
1936 	 * Enable or disable caching as necessary.
1937 	 * Note: the first entry might be part of the kernel pmap,
1938 	 * so we can't assume this is indicative of the state of the
1939 	 * other (maybe non-kpmap) entries.
1940 	 */
1941 	if ((entries > 1 && writable) ||
1942 	    (entries > 0 && pm == kpmap && other_writable)) {
1943 		if (cacheable_entries == 0) {
1944 			return;
1945 		}
1946 
1947 		for (pv = npv; pv; pv = SLIST_NEXT(pv, pv_link)) {
1948 			if ((pm != pv->pv_pmap && kpmap != pv->pv_pmap) ||
1949 			    (pv->pv_flags & PVF_NC))
1950 				continue;
1951 
1952 			pv->pv_flags |= PVF_NC;
1953 
1954 			struct l2_bucket * const l2b
1955 			    = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
1956 			KASSERTMSG(l2b != NULL, "%#lx", va);
1957 			pt_entry_t * const ptep
1958 			    = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1959 			const pt_entry_t opte = *ptep;
1960 			pt_entry_t npte = opte & ~L2_S_CACHE_MASK;
1961 
1962 			if ((va != pv->pv_va || pm != pv->pv_pmap)
1963 			    && l2pte_valid_p(opte)) {
1964 				pmap_cache_wbinv_page(pv->pv_pmap, pv->pv_va,
1965 				    true, pv->pv_flags);
1966 				pmap_tlb_flush_SE(pv->pv_pmap, pv->pv_va,
1967 				    pv->pv_flags);
1968 			}
1969 
1970 			l2pte_set(ptep, npte, opte);
1971 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
1972 		}
1973 		cpu_cpwait();
1974 	} else if (entries > cacheable_entries) {
1975 		/*
1976 		 * Turn cacheing back on for some pages.  If it is a kernel
1977 		 * page, only do so if there are no other writable pages.
1978 		 */
1979 		for (pv = npv; pv; pv = SLIST_NEXT(pv, pv_link)) {
1980 			if (!(pv->pv_flags & PVF_NC) || (pm != pv->pv_pmap &&
1981 			    (kpmap != pv->pv_pmap || other_writable)))
1982 				continue;
1983 
1984 			pv->pv_flags &= ~PVF_NC;
1985 
1986 			struct l2_bucket * const l2b
1987 			    = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
1988 			KASSERTMSG(l2b != NULL, "%#lx", va);
1989 			pt_entry_t * const ptep
1990 			    = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1991 			const pt_entry_t opte = *ptep;
1992 			pt_entry_t npte = (opte & ~L2_S_CACHE_MASK)
1993 			    | pte_l2_s_cache_mode;
1994 
1995 			if (l2pte_valid_p(opte)) {
1996 				pmap_tlb_flush_SE(pv->pv_pmap, pv->pv_va,
1997 				    pv->pv_flags);
1998 			}
1999 
2000 			l2pte_set(ptep, npte, opte);
2001 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
2002 		}
2003 	}
2004 }
2005 #endif
2006 
2007 #ifdef PMAP_CACHE_VIPT
2008 static void
2009 pmap_vac_me_harder(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
2010 {
2011 
2012 #ifndef ARM_MMU_EXTENDED
2013 	struct pv_entry *pv;
2014 	vaddr_t tst_mask;
2015 	bool bad_alias;
2016 	const u_int
2017 	    rw_mappings = md->urw_mappings + md->krw_mappings,
2018 	    ro_mappings = md->uro_mappings + md->kro_mappings;
2019 
2020 	/* do we need to do anything? */
2021 	if (arm_cache_prefer_mask == 0)
2022 		return;
2023 
2024 	UVMHIST_FUNC(__func__);
2025 	UVMHIST_CALLARGS(maphist, "md %#jx pa %#jx pm %#jx va %#jx",
2026 	    (uintptr_t)md, (uintptr_t)pa, (uintptr_t)pm, va);
2027 
2028 	KASSERT(!va || pm);
2029 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2030 
2031 	/* Already a conflict? */
2032 	if (__predict_false(md->pvh_attrs & PVF_NC)) {
2033 		/* just an add, things are already non-cached */
2034 		KASSERT(!(md->pvh_attrs & PVF_DIRTY));
2035 		KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2036 		bad_alias = false;
2037 		if (va) {
2038 			PMAPCOUNT(vac_color_none);
2039 			bad_alias = true;
2040 			KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2041 			goto fixup;
2042 		}
2043 		pv = SLIST_FIRST(&md->pvh_list);
2044 		/* the list can't be empty because it would be cachable */
2045 		if (md->pvh_attrs & PVF_KMPAGE) {
2046 			tst_mask = md->pvh_attrs;
2047 		} else {
2048 			KASSERT(pv);
2049 			tst_mask = pv->pv_va;
2050 			pv = SLIST_NEXT(pv, pv_link);
2051 		}
2052 		/*
2053 		 * Only check for a bad alias if we have writable mappings.
2054 		 */
2055 		tst_mask &= arm_cache_prefer_mask;
2056 		if (rw_mappings > 0) {
2057 			for (; pv && !bad_alias; pv = SLIST_NEXT(pv, pv_link)) {
2058 				/* if there's a bad alias, stop checking. */
2059 				if (tst_mask != (pv->pv_va & arm_cache_prefer_mask))
2060 					bad_alias = true;
2061 			}
2062 			md->pvh_attrs |= PVF_WRITE;
2063 			if (!bad_alias)
2064 				md->pvh_attrs |= PVF_DIRTY;
2065 		} else {
2066 			/*
2067 			 * We have only read-only mappings.  Let's see if there
2068 			 * are multiple colors in use or if we mapped a KMPAGE.
2069 			 * If the latter, we have a bad alias.  If the former,
2070 			 * we need to remember that.
2071 			 */
2072 			for (; pv; pv = SLIST_NEXT(pv, pv_link)) {
2073 				if (tst_mask != (pv->pv_va & arm_cache_prefer_mask)) {
2074 					if (md->pvh_attrs & PVF_KMPAGE)
2075 						bad_alias = true;
2076 					break;
2077 				}
2078 			}
2079 			md->pvh_attrs &= ~PVF_WRITE;
2080 			/*
2081 			 * No KMPAGE and we exited early, so we must have
2082 			 * multiple color mappings.
2083 			 */
2084 			if (!bad_alias && pv != NULL)
2085 				md->pvh_attrs |= PVF_MULTCLR;
2086 		}
2087 
2088 		/* If no conflicting colors, set everything back to cached */
2089 		if (!bad_alias) {
2090 #ifdef DEBUG
2091 			if ((md->pvh_attrs & PVF_WRITE)
2092 			    || ro_mappings < 2) {
2093 				SLIST_FOREACH(pv, &md->pvh_list, pv_link)
2094 					KDASSERT(((tst_mask ^ pv->pv_va) & arm_cache_prefer_mask) == 0);
2095 			}
2096 #endif
2097 			md->pvh_attrs &= (PAGE_SIZE - 1) & ~PVF_NC;
2098 			md->pvh_attrs |= tst_mask | PVF_COLORED;
2099 			/*
2100 			 * Restore DIRTY bit if page is modified
2101 			 */
2102 			if (md->pvh_attrs & PVF_DMOD)
2103 				md->pvh_attrs |= PVF_DIRTY;
2104 			PMAPCOUNT(vac_color_restore);
2105 		} else {
2106 			KASSERT(SLIST_FIRST(&md->pvh_list) != NULL);
2107 			KASSERT(SLIST_NEXT(SLIST_FIRST(&md->pvh_list), pv_link) != NULL);
2108 		}
2109 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2110 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2111 	} else if (!va) {
2112 		KASSERT(pmap_is_page_colored_p(md));
2113 		KASSERT(!(md->pvh_attrs & PVF_WRITE)
2114 		    || (md->pvh_attrs & PVF_DIRTY));
2115 		if (rw_mappings == 0) {
2116 			md->pvh_attrs &= ~PVF_WRITE;
2117 			if (ro_mappings == 1
2118 			    && (md->pvh_attrs & PVF_MULTCLR)) {
2119 				/*
2120 				 * If this is the last readonly mapping
2121 				 * but it doesn't match the current color
2122 				 * for the page, change the current color
2123 				 * to match this last readonly mapping.
2124 				 */
2125 				pv = SLIST_FIRST(&md->pvh_list);
2126 				tst_mask = (md->pvh_attrs ^ pv->pv_va)
2127 				    & arm_cache_prefer_mask;
2128 				if (tst_mask) {
2129 					md->pvh_attrs ^= tst_mask;
2130 					PMAPCOUNT(vac_color_change);
2131 				}
2132 			}
2133 		}
2134 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2135 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2136 		return;
2137 	} else if (!pmap_is_page_colored_p(md)) {
2138 		/* not colored so we just use its color */
2139 		KASSERT(md->pvh_attrs & (PVF_WRITE|PVF_DIRTY));
2140 		KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2141 		PMAPCOUNT(vac_color_new);
2142 		md->pvh_attrs &= PAGE_SIZE - 1;
2143 		md->pvh_attrs |= PVF_COLORED
2144 		    | (va & arm_cache_prefer_mask)
2145 		    | (rw_mappings > 0 ? PVF_WRITE : 0);
2146 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2147 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2148 		return;
2149 	} else if (((md->pvh_attrs ^ va) & arm_cache_prefer_mask) == 0) {
2150 		bad_alias = false;
2151 		if (rw_mappings > 0) {
2152 			/*
2153 			 * We now have writeable mappings and if we have
2154 			 * readonly mappings in more than once color, we have
2155 			 * an aliasing problem.  Regardless mark the page as
2156 			 * writeable.
2157 			 */
2158 			if (md->pvh_attrs & PVF_MULTCLR) {
2159 				if (ro_mappings < 2) {
2160 					/*
2161 					 * If we only have less than two
2162 					 * read-only mappings, just flush the
2163 					 * non-primary colors from the cache.
2164 					 */
2165 					pmap_flush_page(md, pa,
2166 					    PMAP_FLUSH_SECONDARY);
2167 				} else {
2168 					bad_alias = true;
2169 				}
2170 			}
2171 			md->pvh_attrs |= PVF_WRITE;
2172 		}
2173 		/* If no conflicting colors, set everything back to cached */
2174 		if (!bad_alias) {
2175 #ifdef DEBUG
2176 			if (rw_mappings > 0
2177 			    || (md->pvh_attrs & PMAP_KMPAGE)) {
2178 				tst_mask = md->pvh_attrs & arm_cache_prefer_mask;
2179 				SLIST_FOREACH(pv, &md->pvh_list, pv_link)
2180 					KDASSERT(((tst_mask ^ pv->pv_va) & arm_cache_prefer_mask) == 0);
2181 			}
2182 #endif
2183 			if (SLIST_EMPTY(&md->pvh_list))
2184 				PMAPCOUNT(vac_color_reuse);
2185 			else
2186 				PMAPCOUNT(vac_color_ok);
2187 
2188 			/* matching color, just return */
2189 			KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2190 			KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2191 			return;
2192 		}
2193 		KASSERT(SLIST_FIRST(&md->pvh_list) != NULL);
2194 		KASSERT(SLIST_NEXT(SLIST_FIRST(&md->pvh_list), pv_link) != NULL);
2195 
2196 		/* color conflict.  evict from cache. */
2197 
2198 		pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
2199 		md->pvh_attrs &= ~PVF_COLORED;
2200 		md->pvh_attrs |= PVF_NC;
2201 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2202 		KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2203 		PMAPCOUNT(vac_color_erase);
2204 	} else if (rw_mappings == 0
2205 		   && (md->pvh_attrs & PVF_KMPAGE) == 0) {
2206 		KASSERT((md->pvh_attrs & PVF_WRITE) == 0);
2207 
2208 		/*
2209 		 * If the page has dirty cache lines, clean it.
2210 		 */
2211 		if (md->pvh_attrs & PVF_DIRTY)
2212 			pmap_flush_page(md, pa, PMAP_CLEAN_PRIMARY);
2213 
2214 		/*
2215 		 * If this is the first remapping (we know that there are no
2216 		 * writeable mappings), then this is a simple color change.
2217 		 * Otherwise this is a seconary r/o mapping, which means
2218 		 * we don't have to do anything.
2219 		 */
2220 		if (ro_mappings == 1) {
2221 			KASSERT(((md->pvh_attrs ^ va) & arm_cache_prefer_mask) != 0);
2222 			md->pvh_attrs &= PAGE_SIZE - 1;
2223 			md->pvh_attrs |= (va & arm_cache_prefer_mask);
2224 			PMAPCOUNT(vac_color_change);
2225 		} else {
2226 			PMAPCOUNT(vac_color_blind);
2227 		}
2228 		md->pvh_attrs |= PVF_MULTCLR;
2229 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2230 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2231 		return;
2232 	} else {
2233 		if (rw_mappings > 0)
2234 			md->pvh_attrs |= PVF_WRITE;
2235 
2236 		/* color conflict.  evict from cache. */
2237 		pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
2238 
2239 		/* the list can't be empty because this was a enter/modify */
2240 		pv = SLIST_FIRST(&md->pvh_list);
2241 		if ((md->pvh_attrs & PVF_KMPAGE) == 0) {
2242 			KASSERT(pv);
2243 			/*
2244 			 * If there's only one mapped page, change color to the
2245 			 * page's new color and return.  Restore the DIRTY bit
2246 			 * that was erased by pmap_flush_page.
2247 			 */
2248 			if (SLIST_NEXT(pv, pv_link) == NULL) {
2249 				md->pvh_attrs &= PAGE_SIZE - 1;
2250 				md->pvh_attrs |= (va & arm_cache_prefer_mask);
2251 				if (md->pvh_attrs & PVF_DMOD)
2252 					md->pvh_attrs |= PVF_DIRTY;
2253 				PMAPCOUNT(vac_color_change);
2254 				KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2255 				KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2256 				KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2257 				return;
2258 			}
2259 		}
2260 		bad_alias = true;
2261 		md->pvh_attrs &= ~PVF_COLORED;
2262 		md->pvh_attrs |= PVF_NC;
2263 		PMAPCOUNT(vac_color_erase);
2264 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2265 	}
2266 
2267   fixup:
2268 	KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2269 
2270 	/*
2271 	 * Turn cacheing on/off for all pages.
2272 	 */
2273 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
2274 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pv->pv_pmap,
2275 		    pv->pv_va);
2276 		KASSERTMSG(l2b != NULL, "%#lx", va);
2277 		pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2278 		const pt_entry_t opte = *ptep;
2279 		pt_entry_t npte = opte & ~L2_S_CACHE_MASK;
2280 		if (bad_alias) {
2281 			pv->pv_flags |= PVF_NC;
2282 		} else {
2283 			pv->pv_flags &= ~PVF_NC;
2284 			npte |= pte_l2_s_cache_mode;
2285 		}
2286 
2287 		if (opte == npte)	/* only update is there's a change */
2288 			continue;
2289 
2290 		if (l2pte_valid_p(opte)) {
2291 			pmap_tlb_flush_SE(pv->pv_pmap, pv->pv_va, pv->pv_flags);
2292 		}
2293 
2294 		l2pte_set(ptep, npte, opte);
2295 		PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
2296 	}
2297 #endif /* !ARM_MMU_EXTENDED */
2298 }
2299 #endif	/* PMAP_CACHE_VIPT */
2300 
2301 
2302 /*
2303  * Modify pte bits for all ptes corresponding to the given physical address.
2304  * We use `maskbits' rather than `clearbits' because we're always passing
2305  * constants and the latter would require an extra inversion at run-time.
2306  */
2307 static void
2308 pmap_clearbit(struct vm_page_md *md, paddr_t pa, u_int maskbits)
2309 {
2310 	struct pv_entry *pv;
2311 #ifdef PMAP_CACHE_VIPT
2312 	const bool want_syncicache = PV_IS_EXEC_P(md->pvh_attrs);
2313 	bool need_syncicache = false;
2314 #ifdef ARM_MMU_EXTENDED
2315 	const u_int execbits = (maskbits & PVF_EXEC) ? L2_XS_XN : 0;
2316 #else
2317 	const u_int execbits = 0;
2318 	bool need_vac_me_harder = false;
2319 #endif
2320 #else
2321 	const u_int execbits = 0;
2322 #endif
2323 
2324 	UVMHIST_FUNC(__func__);
2325 	UVMHIST_CALLARGS(maphist, "md %#jx pa %#jx maskbits %#jx",
2326 	    (uintptr_t)md, pa, maskbits, 0);
2327 
2328 #ifdef PMAP_CACHE_VIPT
2329 	/*
2330 	 * If we might want to sync the I-cache and we've modified it,
2331 	 * then we know we definitely need to sync or discard it.
2332 	 */
2333 	if (want_syncicache) {
2334 		if (md->pvh_attrs & PVF_MOD) {
2335 			need_syncicache = true;
2336 		}
2337 	}
2338 #endif
2339 	KASSERT(pmap_page_locked_p(md));
2340 
2341 	/*
2342 	 * Clear saved attributes (modify, reference)
2343 	 */
2344 	md->pvh_attrs &= ~(maskbits & (PVF_MOD | PVF_REF));
2345 
2346 	if (SLIST_EMPTY(&md->pvh_list)) {
2347 #if defined(PMAP_CACHE_VIPT)
2348 		if (need_syncicache) {
2349 			/*
2350 			 * No one has it mapped, so just discard it.  The next
2351 			 * exec remapping will cause it to be synced.
2352 			 */
2353 			md->pvh_attrs &= ~PVF_EXEC;
2354 			PMAPCOUNT(exec_discarded_clearbit);
2355 		}
2356 #endif
2357 		return;
2358 	}
2359 
2360 	/*
2361 	 * Loop over all current mappings setting/clearing as appropos
2362 	 */
2363 	for (pv = SLIST_FIRST(&md->pvh_list); pv != NULL;) {
2364 		pmap_t pm = pv->pv_pmap;
2365 		const vaddr_t va = pv->pv_va;
2366 		const u_int oflags = pv->pv_flags;
2367 #ifndef ARM_MMU_EXTENDED
2368 		/*
2369 		 * Kernel entries are unmanaged and as such not to be changed.
2370 		 */
2371 		if (PV_IS_KENTRY_P(oflags)) {
2372 			pv = SLIST_NEXT(pv, pv_link);
2373 			continue;
2374 		}
2375 #endif
2376 
2377 		/*
2378 		 * Try to get a hold on the pmap's lock.  We must do this
2379 		 * while still holding the page locked, to know that the
2380 		 * page is still associated with the pmap and the mapping is
2381 		 * in place.  If a hold can't be had, unlock and wait for
2382 		 * the pmap's lock to become available and retry.  The pmap
2383 		 * must be ref'd over this dance to stop it disappearing
2384 		 * behind us.
2385 		 */
2386 		if (!mutex_tryenter(&pm->pm_lock)) {
2387 			pmap_reference(pm);
2388 			pmap_release_page_lock(md);
2389 			pmap_acquire_pmap_lock(pm);
2390 			/* nothing, just wait for it */
2391 			pmap_release_pmap_lock(pm);
2392 			pmap_destroy(pm);
2393 			/* Restart from the beginning. */
2394 			pmap_acquire_page_lock(md);
2395 			pv = SLIST_FIRST(&md->pvh_list);
2396 			continue;
2397 		}
2398 		pv->pv_flags &= ~maskbits;
2399 
2400 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pm, va);
2401 		KASSERTMSG(l2b != NULL, "%#lx", va);
2402 
2403 		pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
2404 		const pt_entry_t opte = *ptep;
2405 		pt_entry_t npte = opte | execbits;
2406 
2407 #ifdef ARM_MMU_EXTENDED
2408 		KASSERT((opte & L2_XS_nG) == (pm == pmap_kernel() ? 0 : L2_XS_nG));
2409 #endif
2410 
2411 		UVMHIST_LOG(maphist, "pv %#jx pm %#jx va %#jx flag %#jx",
2412 		    (uintptr_t)pv, (uintptr_t)pm, va, oflags);
2413 
2414 		if (maskbits & (PVF_WRITE|PVF_MOD)) {
2415 #ifdef PMAP_CACHE_VIVT
2416 			if ((oflags & PVF_NC)) {
2417 				/*
2418 				 * Entry is not cacheable:
2419 				 *
2420 				 * Don't turn caching on again if this is a
2421 				 * modified emulation. This would be
2422 				 * inconsitent with the settings created by
2423 				 * pmap_vac_me_harder(). Otherwise, it's safe
2424 				 * to re-enable cacheing.
2425 				 *
2426 				 * There's no need to call pmap_vac_me_harder()
2427 				 * here: all pages are losing their write
2428 				 * permission.
2429 				 */
2430 				if (maskbits & PVF_WRITE) {
2431 					npte |= pte_l2_s_cache_mode;
2432 					pv->pv_flags &= ~PVF_NC;
2433 				}
2434 			} else if (l2pte_writable_p(opte)) {
2435 				/*
2436 				 * Entry is writable/cacheable: check if pmap
2437 				 * is current if it is flush it, otherwise it
2438 				 * won't be in the cache
2439 				 */
2440 				pmap_cache_wbinv_page(pm, va,
2441 				    (maskbits & PVF_REF) != 0,
2442 				    oflags|PVF_WRITE);
2443 			}
2444 #endif
2445 
2446 			/* make the pte read only */
2447 			npte = l2pte_set_readonly(npte);
2448 
2449 			if ((maskbits & oflags & PVF_WRITE)) {
2450 				/*
2451 				 * Keep alias accounting up to date
2452 				 */
2453 				if (pm == pmap_kernel()) {
2454 					md->krw_mappings--;
2455 					md->kro_mappings++;
2456 				} else {
2457 					md->urw_mappings--;
2458 					md->uro_mappings++;
2459 				}
2460 #ifdef PMAP_CACHE_VIPT
2461 				if (arm_cache_prefer_mask != 0) {
2462 					if (md->urw_mappings + md->krw_mappings == 0) {
2463 						md->pvh_attrs &= ~PVF_WRITE;
2464 					} else {
2465 						PMAP_VALIDATE_MD_PAGE(md);
2466 					}
2467 				}
2468 				if (want_syncicache)
2469 					need_syncicache = true;
2470 #ifndef ARM_MMU_EXTENDED
2471 				need_vac_me_harder = true;
2472 #endif
2473 #endif /* PMAP_CACHE_VIPT */
2474 			}
2475 		}
2476 
2477 		if (maskbits & PVF_REF) {
2478 			if (true
2479 #ifndef ARM_MMU_EXTENDED
2480 			    && (oflags & PVF_NC) == 0
2481 #endif
2482 			    && (maskbits & (PVF_WRITE|PVF_MOD)) == 0
2483 			    && l2pte_valid_p(npte)) {
2484 #ifdef PMAP_CACHE_VIVT
2485 				/*
2486 				 * Check npte here; we may have already
2487 				 * done the wbinv above, and the validity
2488 				 * of the PTE is the same for opte and
2489 				 * npte.
2490 				 */
2491 				pmap_cache_wbinv_page(pm, va, true, oflags);
2492 #endif
2493 			}
2494 
2495 			/*
2496 			 * Make the PTE invalid so that we will take a
2497 			 * page fault the next time the mapping is
2498 			 * referenced.
2499 			 */
2500 			npte &= ~L2_TYPE_MASK;
2501 			npte |= L2_TYPE_INV;
2502 		}
2503 
2504 		if (npte != opte) {
2505 			l2pte_reset(ptep);
2506 			PTE_SYNC(ptep);
2507 
2508 			/* Flush the TLB entry if a current pmap. */
2509 			pmap_tlb_flush_SE(pm, va, oflags);
2510 
2511 			l2pte_set(ptep, npte, 0);
2512 			PTE_SYNC(ptep);
2513 		}
2514 
2515 		pmap_release_pmap_lock(pm);
2516 
2517 		UVMHIST_LOG(maphist, "pm %#jx va %#jx opte %#jx npte %#jx",
2518 		    (uintptr_t)pm, va, opte, npte);
2519 
2520 		/* Move to next entry. */
2521 		pv = SLIST_NEXT(pv, pv_link);
2522 	}
2523 
2524 #if defined(PMAP_CACHE_VIPT)
2525 	/*
2526 	 * If we need to sync the I-cache and we haven't done it yet, do it.
2527 	 */
2528 	if (need_syncicache) {
2529 		pmap_syncicache_page(md, pa);
2530 		PMAPCOUNT(exec_synced_clearbit);
2531 	}
2532 #ifndef ARM_MMU_EXTENDED
2533 	/*
2534 	 * If we are changing this to read-only, we need to call vac_me_harder
2535 	 * so we can change all the read-only pages to cacheable.  We pretend
2536 	 * this as a page deletion.
2537 	 */
2538 	if (need_vac_me_harder) {
2539 		if (md->pvh_attrs & PVF_NC)
2540 			pmap_vac_me_harder(md, pa, NULL, 0);
2541 	}
2542 #endif /* !ARM_MMU_EXTENDED */
2543 #endif /* PMAP_CACHE_VIPT */
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 	UVMHIST_FUNC(__func__);
2634 	UVMHIST_CALLARGS(maphist, "md %#jx pa %#jx (attrs=%#jx)",
2635 	    (uintptr_t)md, pa, md->pvh_attrs, 0);
2636 
2637 	/*
2638 	 * No need to clean the page if it's non-cached.
2639 	 */
2640 #ifndef ARM_MMU_EXTENDED
2641 	if (md->pvh_attrs & PVF_NC)
2642 		return;
2643 	KASSERT(arm_cache_prefer_mask == 0 || md->pvh_attrs & PVF_COLORED);
2644 #endif
2645 
2646 	pt_entry_t * const ptep = cpu_cdst_pte(0);
2647 	const vaddr_t dstp = cpu_cdstp(0);
2648 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
2649 	if (way_size <= PAGE_SIZE) {
2650 		bool ok = false;
2651 		vaddr_t vdstp = pmap_direct_mapped_phys(pa, &ok, dstp);
2652 		if (ok) {
2653 			cpu_icache_sync_range(vdstp, way_size);
2654 			return;
2655 		}
2656 	}
2657 #endif
2658 
2659 	/*
2660 	 * We don't worry about the color of the exec page, we map the
2661 	 * same page to pages in the way and then do the icache_sync on
2662 	 * the entire way making sure we are cleaned.
2663 	 */
2664 	const pt_entry_t npte = L2_S_PROTO | pa | pte_l2_s_cache_mode
2665 	    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE);
2666 
2667 	for (size_t i = 0, j = 0; i < way_size;
2668 	     i += PAGE_SIZE, j += PAGE_SIZE / L2_S_SIZE) {
2669 		l2pte_reset(ptep + j);
2670 		PTE_SYNC(ptep + j);
2671 
2672 		pmap_tlb_flush_SE(kpm, dstp + i, PVF_REF | PVF_EXEC);
2673 		/*
2674 		 * Set up a PTE with to flush these cache lines.
2675 		 */
2676 		l2pte_set(ptep + j, npte, 0);
2677 	}
2678 	PTE_SYNC_RANGE(ptep, way_size / L2_S_SIZE);
2679 
2680 	/*
2681 	 * Flush it.
2682 	 */
2683 	cpu_icache_sync_range(dstp, way_size);
2684 
2685 	for (size_t i = 0, j = 0; i < way_size;
2686 	     i += PAGE_SIZE, j += PAGE_SIZE / L2_S_SIZE) {
2687 		/*
2688 		 * Unmap the page(s).
2689 		 */
2690 		l2pte_reset(ptep + j);
2691 		PTE_SYNC(ptep + j);
2692 
2693 		pmap_tlb_flush_SE(kpm, dstp + i, PVF_REF | PVF_EXEC);
2694 	}
2695 
2696 	md->pvh_attrs |= PVF_EXEC;
2697 	PMAPCOUNT(exec_synced);
2698 }
2699 
2700 #ifndef ARM_MMU_EXTENDED
2701 void
2702 pmap_flush_page(struct vm_page_md *md, paddr_t pa, enum pmap_flush_op flush)
2703 {
2704 	vsize_t va_offset, end_va;
2705 	bool wbinv_p;
2706 
2707 	if (arm_cache_prefer_mask == 0)
2708 		return;
2709 
2710 	UVMHIST_FUNC(__func__);
2711 	UVMHIST_CALLARGS(maphist, "md %#jx pa %#jx op %#jx",
2712 	    (uintptr_t)md, pa, op, 0);
2713 
2714 	switch (flush) {
2715 	case PMAP_FLUSH_PRIMARY:
2716 		if (md->pvh_attrs & PVF_MULTCLR) {
2717 			va_offset = 0;
2718 			end_va = arm_cache_prefer_mask;
2719 			md->pvh_attrs &= ~PVF_MULTCLR;
2720 			PMAPCOUNT(vac_flush_lots);
2721 		} else {
2722 			va_offset = md->pvh_attrs & arm_cache_prefer_mask;
2723 			end_va = va_offset;
2724 			PMAPCOUNT(vac_flush_one);
2725 		}
2726 		/*
2727 		 * Mark that the page is no longer dirty.
2728 		 */
2729 		md->pvh_attrs &= ~PVF_DIRTY;
2730 		wbinv_p = true;
2731 		break;
2732 	case PMAP_FLUSH_SECONDARY:
2733 		va_offset = 0;
2734 		end_va = arm_cache_prefer_mask;
2735 		wbinv_p = true;
2736 		md->pvh_attrs &= ~PVF_MULTCLR;
2737 		PMAPCOUNT(vac_flush_lots);
2738 		break;
2739 	case PMAP_CLEAN_PRIMARY:
2740 		va_offset = md->pvh_attrs & arm_cache_prefer_mask;
2741 		end_va = va_offset;
2742 		wbinv_p = false;
2743 		/*
2744 		 * Mark that the page is no longer dirty.
2745 		 */
2746 		if ((md->pvh_attrs & PVF_DMOD) == 0)
2747 			md->pvh_attrs &= ~PVF_DIRTY;
2748 		PMAPCOUNT(vac_clean_one);
2749 		break;
2750 	default:
2751 		return;
2752 	}
2753 
2754 	KASSERT(!(md->pvh_attrs & PVF_NC));
2755 
2756 	UVMHIST_LOG(maphist, "md %#jx (attrs=%#jx)", (uintptr_t)md,
2757 	    md->pvh_attrs, 0, 0);
2758 
2759 	const size_t scache_line_size = arm_scache.dcache_line_size;
2760 
2761 	for (; va_offset <= end_va; va_offset += PAGE_SIZE) {
2762 		pt_entry_t * const ptep = cpu_cdst_pte(va_offset);
2763 		const vaddr_t dstp = cpu_cdstp(va_offset);
2764 		const pt_entry_t opte = *ptep;
2765 
2766 		if (flush == PMAP_FLUSH_SECONDARY
2767 		    && va_offset == (md->pvh_attrs & arm_cache_prefer_mask))
2768 			continue;
2769 
2770 		pmap_tlb_flush_SE(pmap_kernel(), dstp, PVF_REF | PVF_EXEC);
2771 		/*
2772 		 * Set up a PTE with the right coloring to flush
2773 		 * existing cache entries.
2774 		 */
2775 		const pt_entry_t npte = L2_S_PROTO
2776 		    | pa
2777 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE)
2778 		    | pte_l2_s_cache_mode;
2779 		l2pte_set(ptep, npte, opte);
2780 		PTE_SYNC(ptep);
2781 
2782 		/*
2783 		 * Flush it.  Make sure to flush secondary cache too since
2784 		 * bus_dma will ignore uncached pages.
2785 		 */
2786 		if (scache_line_size != 0) {
2787 			cpu_dcache_wb_range(dstp, PAGE_SIZE);
2788 			if (wbinv_p) {
2789 				cpu_sdcache_wbinv_range(dstp, pa, PAGE_SIZE);
2790 				cpu_dcache_inv_range(dstp, PAGE_SIZE);
2791 			} else {
2792 				cpu_sdcache_wb_range(dstp, pa, PAGE_SIZE);
2793 			}
2794 		} else {
2795 			if (wbinv_p) {
2796 				cpu_dcache_wbinv_range(dstp, PAGE_SIZE);
2797 			} else {
2798 				cpu_dcache_wb_range(dstp, PAGE_SIZE);
2799 			}
2800 		}
2801 
2802 		/*
2803 		 * Restore the page table entry since we might have interrupted
2804 		 * pmap_zero_page or pmap_copy_page which was already using
2805 		 * this pte.
2806 		 */
2807 		if (opte) {
2808 			l2pte_set(ptep, opte, npte);
2809 		} else {
2810 			l2pte_reset(ptep);
2811 		}
2812 		PTE_SYNC(ptep);
2813 		pmap_tlb_flush_SE(pmap_kernel(), dstp, PVF_REF | PVF_EXEC);
2814 	}
2815 }
2816 #endif /* ARM_MMU_EXTENDED */
2817 #endif /* PMAP_CACHE_VIPT */
2818 
2819 /*
2820  * Routine:	pmap_page_remove
2821  * Function:
2822  *		Removes this physical page from
2823  *		all physical maps in which it resides.
2824  *		Reflects back modify bits to the pager.
2825  */
2826 static void
2827 pmap_page_remove(struct vm_page_md *md, paddr_t pa)
2828 {
2829 	struct l2_bucket *l2b;
2830 	struct pv_entry *pv;
2831 	pt_entry_t *ptep;
2832 #ifndef ARM_MMU_EXTENDED
2833 	bool flush = false;
2834 #endif
2835 	u_int flags = 0;
2836 
2837 	UVMHIST_FUNC(__func__);
2838 	UVMHIST_CALLARGS(maphist, "md %#jx pa %#jx", (uintptr_t)md, pa, 0, 0);
2839 
2840 	struct pv_entry **pvp = &SLIST_FIRST(&md->pvh_list);
2841 	pmap_acquire_page_lock(md);
2842 	if (*pvp == NULL) {
2843 #ifdef PMAP_CACHE_VIPT
2844 		/*
2845 		 * We *know* the page contents are about to be replaced.
2846 		 * Discard the exec contents
2847 		 */
2848 		if (PV_IS_EXEC_P(md->pvh_attrs))
2849 			PMAPCOUNT(exec_discarded_page_protect);
2850 		md->pvh_attrs &= ~PVF_EXEC;
2851 		PMAP_VALIDATE_MD_PAGE(md);
2852 #endif
2853 		pmap_release_page_lock(md);
2854 		return;
2855 	}
2856 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
2857 	KASSERT(arm_cache_prefer_mask == 0 || pmap_is_page_colored_p(md));
2858 #endif
2859 
2860 	/*
2861 	 * Clear alias counts
2862 	 */
2863 #ifdef PMAP_CACHE_VIVT
2864 	md->k_mappings = 0;
2865 #endif
2866 	md->urw_mappings = md->uro_mappings = 0;
2867 
2868 #ifdef PMAP_CACHE_VIVT
2869 	pmap_clean_page(md, false);
2870 #endif
2871 
2872 	for (pv = *pvp; pv != NULL;) {
2873 		pmap_t pm = pv->pv_pmap;
2874 #ifndef ARM_MMU_EXTENDED
2875 		if (flush == false && pmap_is_current(pm))
2876 			flush = true;
2877 #endif
2878 
2879 #ifdef PMAP_CACHE_VIPT
2880 		if (pm == pmap_kernel() && PV_IS_KENTRY_P(pv->pv_flags)) {
2881 			/* If this was unmanaged mapping, it must be ignored. */
2882 			pvp = &SLIST_NEXT(pv, pv_link);
2883 			pv = *pvp;
2884 			continue;
2885 		}
2886 #endif
2887 
2888 		/*
2889 		 * Try to get a hold on the pmap's lock.  We must do this
2890 		 * while still holding the page locked, to know that the
2891 		 * page is still associated with the pmap and the mapping is
2892 		 * in place.  If a hold can't be had, unlock and wait for
2893 		 * the pmap's lock to become available and retry.  The pmap
2894 		 * must be ref'd over this dance to stop it disappearing
2895 		 * behind us.
2896 		 */
2897 		if (!mutex_tryenter(&pm->pm_lock)) {
2898 			pmap_reference(pm);
2899 			pmap_release_page_lock(md);
2900 			pmap_acquire_pmap_lock(pm);
2901 			/* nothing, just wait for it */
2902 			pmap_release_pmap_lock(pm);
2903 			pmap_destroy(pm);
2904 			/* Restart from the beginning. */
2905 			pmap_acquire_page_lock(md);
2906 			pvp = &SLIST_FIRST(&md->pvh_list);
2907 			pv = *pvp;
2908 			continue;
2909 		}
2910 
2911 		if (pm == pmap_kernel()) {
2912 #ifdef PMAP_CACHE_VIPT
2913 			if (pv->pv_flags & PVF_WRITE)
2914 				md->krw_mappings--;
2915 			else
2916 				md->kro_mappings--;
2917 #endif
2918 			PMAPCOUNT(kernel_unmappings);
2919 		}
2920 		*pvp = SLIST_NEXT(pv, pv_link); /* remove from list */
2921 		PMAPCOUNT(unmappings);
2922 
2923 		pmap_release_page_lock(md);
2924 
2925 		l2b = pmap_get_l2_bucket(pm, pv->pv_va);
2926 		KASSERTMSG(l2b != NULL, "%#lx", pv->pv_va);
2927 
2928 		ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2929 
2930 		/*
2931 		 * Update statistics
2932 		 */
2933 		--pm->pm_stats.resident_count;
2934 
2935 		/* Wired bit */
2936 		if (pv->pv_flags & PVF_WIRED)
2937 			--pm->pm_stats.wired_count;
2938 
2939 		flags |= pv->pv_flags;
2940 
2941 		/*
2942 		 * Invalidate the PTEs.
2943 		 */
2944 		l2pte_reset(ptep);
2945 		PTE_SYNC_CURRENT(pm, ptep);
2946 
2947 #ifdef ARM_MMU_EXTENDED
2948 		pmap_tlb_invalidate_addr(pm, pv->pv_va);
2949 #endif
2950 
2951 		pmap_free_l2_bucket(pm, l2b, PAGE_SIZE / L2_S_SIZE);
2952 
2953 		pmap_release_pmap_lock(pm);
2954 
2955 		pool_put(&pmap_pv_pool, pv);
2956 		pmap_acquire_page_lock(md);
2957 
2958 		/*
2959 		 * Restart at the beginning of the list.
2960 		 */
2961 		pvp = &SLIST_FIRST(&md->pvh_list);
2962 		pv = *pvp;
2963 	}
2964 	/*
2965 	 * if we reach the end of the list and there are still mappings, they
2966 	 * might be able to be cached now.  And they must be kernel mappings.
2967 	 */
2968 	if (!SLIST_EMPTY(&md->pvh_list)) {
2969 		pmap_vac_me_harder(md, pa, pmap_kernel(), 0);
2970 	}
2971 
2972 #ifdef PMAP_CACHE_VIPT
2973 	/*
2974 	 * Its EXEC cache is now gone.
2975 	 */
2976 	if (PV_IS_EXEC_P(md->pvh_attrs))
2977 		PMAPCOUNT(exec_discarded_page_protect);
2978 	md->pvh_attrs &= ~PVF_EXEC;
2979 	KASSERT(md->urw_mappings == 0);
2980 	KASSERT(md->uro_mappings == 0);
2981 #ifndef ARM_MMU_EXTENDED
2982 	if (arm_cache_prefer_mask != 0) {
2983 		if (md->krw_mappings == 0)
2984 			md->pvh_attrs &= ~PVF_WRITE;
2985 		PMAP_VALIDATE_MD_PAGE(md);
2986 	}
2987 #endif /* ARM_MMU_EXTENDED */
2988 #endif /* PMAP_CACHE_VIPT */
2989 	pmap_release_page_lock(md);
2990 
2991 #ifndef ARM_MMU_EXTENDED
2992 	if (flush) {
2993 		/*
2994 		 * Note: We can't use pmap_tlb_flush{I,D}() here since that
2995 		 * would need a subsequent call to pmap_update() to ensure
2996 		 * curpm->pm_cstate.cs_all is reset. Our callers are not
2997 		 * required to do that (see pmap(9)), so we can't modify
2998 		 * the current pmap's state.
2999 		 */
3000 		if (PV_BEEN_EXECD(flags))
3001 			cpu_tlb_flushID();
3002 		else
3003 			cpu_tlb_flushD();
3004 	}
3005 	cpu_cpwait();
3006 #endif /* ARM_MMU_EXTENDED */
3007 }
3008 
3009 /*
3010  * pmap_t pmap_create(void)
3011  *
3012  *      Create a new pmap structure from scratch.
3013  */
3014 pmap_t
3015 pmap_create(void)
3016 {
3017 	pmap_t pm;
3018 
3019 	pm = pool_cache_get(&pmap_cache, PR_WAITOK);
3020 
3021 	mutex_init(&pm->pm_lock, MUTEX_DEFAULT, IPL_NONE);
3022 
3023 	pm->pm_refs = 1;
3024 	pm->pm_stats.wired_count = 0;
3025 	pm->pm_stats.resident_count = 1;
3026 #ifdef ARM_MMU_EXTENDED
3027 #ifdef MULTIPROCESSOR
3028 	kcpuset_create(&pm->pm_active, true);
3029 	kcpuset_create(&pm->pm_onproc, true);
3030 #endif
3031 #else
3032 	pm->pm_cstate.cs_all = 0;
3033 #endif
3034 	pmap_alloc_l1(pm);
3035 
3036 	/*
3037 	 * Note: The pool cache ensures that the pm_l2[] array is already
3038 	 * initialised to zero.
3039 	 */
3040 
3041 	pmap_pinit(pm);
3042 
3043 	return pm;
3044 }
3045 
3046 u_int
3047 arm32_mmap_flags(paddr_t pa)
3048 {
3049 	/*
3050 	 * the upper 8 bits in pmap_enter()'s flags are reserved for MD stuff
3051 	 * and we're using the upper bits in page numbers to pass flags around
3052 	 * so we might as well use the same bits
3053 	 */
3054 	return (u_int)pa & PMAP_MD_MASK;
3055 }
3056 /*
3057  * int pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
3058  *      u_int flags)
3059  *
3060  *      Insert the given physical page (p) at
3061  *      the specified virtual address (v) in the
3062  *      target physical map with the protection requested.
3063  *
3064  *      NB:  This is the only routine which MAY NOT lazy-evaluate
3065  *      or lose information.  That is, this routine must actually
3066  *      insert this page into the given map NOW.
3067  */
3068 int
3069 pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags)
3070 {
3071 	struct l2_bucket *l2b;
3072 	struct vm_page *pg, *opg;
3073 	u_int nflags;
3074 	u_int oflags;
3075 	const bool kpm_p = (pm == pmap_kernel());
3076 #ifdef ARM_HAS_VBAR
3077 	const bool vector_page_p = false;
3078 #else
3079 	const bool vector_page_p = (va == vector_page);
3080 #endif
3081 	struct pmap_page *pp = pmap_pv_tracked(pa);
3082 	struct pv_entry *new_pv = NULL;
3083 	struct pv_entry *old_pv = NULL;
3084 	int error = 0;
3085 
3086 	UVMHIST_FUNC(__func__);
3087 	UVMHIST_CALLARGS(maphist, "pm %#jx va %#jx pa %#jx prot %#jx",
3088 	    (uintptr_t)pm, va, pa, prot);
3089 	UVMHIST_LOG(maphist, "  flag %#jx", flags, 0, 0, 0);
3090 
3091 	KDASSERT((flags & PMAP_WIRED) == 0 || (flags & VM_PROT_ALL) != 0);
3092 	KDASSERT(((va | pa) & PGOFSET) == 0);
3093 
3094 	/*
3095 	 * Get a pointer to the page.  Later on in this function, we
3096 	 * test for a managed page by checking pg != NULL.
3097 	 */
3098 	pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
3099 	/*
3100 	 * if we may need a new pv entry allocate if now, as we can't do it
3101 	 * with the kernel_pmap locked
3102 	 */
3103 	if (pg || pp)
3104 		new_pv = pool_get(&pmap_pv_pool, PR_NOWAIT);
3105 
3106 	nflags = 0;
3107 	if (prot & VM_PROT_WRITE)
3108 		nflags |= PVF_WRITE;
3109 	if (prot & VM_PROT_EXECUTE)
3110 		nflags |= PVF_EXEC;
3111 	if (flags & PMAP_WIRED)
3112 		nflags |= PVF_WIRED;
3113 
3114 	pmap_acquire_pmap_lock(pm);
3115 
3116 	/*
3117 	 * Fetch the L2 bucket which maps this page, allocating one if
3118 	 * necessary for user pmaps.
3119 	 */
3120 	if (kpm_p) {
3121 		l2b = pmap_get_l2_bucket(pm, va);
3122 	} else {
3123 		l2b = pmap_alloc_l2_bucket(pm, va);
3124 	}
3125 	if (l2b == NULL) {
3126 		if (flags & PMAP_CANFAIL) {
3127 			pmap_release_pmap_lock(pm);
3128 			error = ENOMEM;
3129 			goto free_pv;
3130 		}
3131 		panic("pmap_enter: failed to allocate L2 bucket");
3132 	}
3133 	pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(va)];
3134 	const pt_entry_t opte = *ptep;
3135 	pt_entry_t npte = pa;
3136 	oflags = 0;
3137 
3138 	if (opte) {
3139 		/*
3140 		 * There is already a mapping at this address.
3141 		 * If the physical address is different, lookup the
3142 		 * vm_page.
3143 		 */
3144 		if (l2pte_pa(opte) != pa) {
3145 			KASSERT(!pmap_pv_tracked(pa));
3146 			opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3147 		} else
3148 			opg = pg;
3149 	} else
3150 		opg = NULL;
3151 
3152 	if (pg || pp) {
3153 		KASSERT((pg != NULL) != (pp != NULL));
3154 		struct vm_page_md *md = (pg != NULL) ? VM_PAGE_TO_MD(pg) :
3155 		    PMAP_PAGE_TO_MD(pp);
3156 
3157 		/*
3158 		 * This is to be a managed mapping.
3159 		 */
3160 		pmap_acquire_page_lock(md);
3161 		if ((flags & VM_PROT_ALL) || (md->pvh_attrs & PVF_REF)) {
3162 			/*
3163 			 * - The access type indicates that we don't need
3164 			 *   to do referenced emulation.
3165 			 * OR
3166 			 * - The physical page has already been referenced
3167 			 *   so no need to re-do referenced emulation here.
3168 			 */
3169 			npte |= l2pte_set_readonly(L2_S_PROTO);
3170 
3171 			nflags |= PVF_REF;
3172 
3173 			if ((prot & VM_PROT_WRITE) != 0 &&
3174 			    ((flags & VM_PROT_WRITE) != 0 ||
3175 			     (md->pvh_attrs & PVF_MOD) != 0)) {
3176 				/*
3177 				 * This is a writable mapping, and the
3178 				 * page's mod state indicates it has
3179 				 * already been modified. Make it
3180 				 * writable from the outset.
3181 				 */
3182 				npte = l2pte_set_writable(npte);
3183 				nflags |= PVF_MOD;
3184 			}
3185 
3186 #ifdef ARM_MMU_EXTENDED
3187 			/*
3188 			 * If the page has been cleaned, then the pvh_attrs
3189 			 * will have PVF_EXEC set, so mark it execute so we
3190 			 * don't get an access fault when trying to execute
3191 			 * from it.
3192 			 */
3193 			if (md->pvh_attrs & nflags & PVF_EXEC) {
3194 				npte &= ~L2_XS_XN;
3195 			}
3196 #endif
3197 		} else {
3198 			/*
3199 			 * Need to do page referenced emulation.
3200 			 */
3201 			npte |= L2_TYPE_INV;
3202 		}
3203 
3204 		if (flags & ARM32_MMAP_WRITECOMBINE) {
3205 			npte |= pte_l2_s_wc_mode;
3206 		} else
3207 			npte |= pte_l2_s_cache_mode;
3208 
3209 		if (pg != NULL && pg == opg) {
3210 			/*
3211 			 * We're changing the attrs of an existing mapping.
3212 			 */
3213 			oflags = pmap_modify_pv(md, pa, pm, va,
3214 			    PVF_WRITE | PVF_EXEC | PVF_WIRED |
3215 			    PVF_MOD | PVF_REF, nflags);
3216 
3217 #ifdef PMAP_CACHE_VIVT
3218 			/*
3219 			 * We may need to flush the cache if we're
3220 			 * doing rw-ro...
3221 			 */
3222 			if (pm->pm_cstate.cs_cache_d &&
3223 			    (oflags & PVF_NC) == 0 &&
3224 			    l2pte_writable_p(opte) &&
3225 			    (prot & VM_PROT_WRITE) == 0)
3226 				cpu_dcache_wb_range(va, PAGE_SIZE);
3227 #endif
3228 		} else {
3229 			struct pv_entry *pv;
3230 			/*
3231 			 * New mapping, or changing the backing page
3232 			 * of an existing mapping.
3233 			 */
3234 			if (opg) {
3235 				struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3236 				paddr_t opa = VM_PAGE_TO_PHYS(opg);
3237 
3238 				/*
3239 				 * Replacing an existing mapping with a new one.
3240 				 * It is part of our managed memory so we
3241 				 * must remove it from the PV list
3242 				 */
3243 				pv = pmap_remove_pv(omd, opa, pm, va);
3244 				pmap_vac_me_harder(omd, opa, pm, 0);
3245 				oflags = pv->pv_flags;
3246 
3247 #ifdef PMAP_CACHE_VIVT
3248 				/*
3249 				 * If the old mapping was valid (ref/mod
3250 				 * emulation creates 'invalid' mappings
3251 				 * initially) then make sure to frob
3252 				 * the cache.
3253 				 */
3254 				if (!(oflags & PVF_NC) && l2pte_valid_p(opte)) {
3255 					pmap_cache_wbinv_page(pm, va, true,
3256 					    oflags);
3257 				}
3258 #endif
3259 			} else {
3260 				pv = new_pv;
3261 				new_pv = NULL;
3262 				if (pv == NULL) {
3263 					pmap_release_page_lock(md);
3264 					pmap_release_pmap_lock(pm);
3265 					if ((flags & PMAP_CANFAIL) == 0)
3266 						panic("pmap_enter: "
3267 						    "no pv entries");
3268 
3269 					pmap_free_l2_bucket(pm, l2b, 0);
3270 					UVMHIST_LOG(maphist, "  <-- done (ENOMEM)",
3271 					    0, 0, 0, 0);
3272 					return ENOMEM;
3273 				}
3274 			}
3275 
3276 			pmap_enter_pv(md, pa, pv, pm, va, nflags);
3277 		}
3278 		pmap_release_page_lock(md);
3279 	} else {
3280 		/*
3281 		 * We're mapping an unmanaged page.
3282 		 * These are always readable, and possibly writable, from
3283 		 * the get go as we don't need to track ref/mod status.
3284 		 */
3285 		npte |= l2pte_set_readonly(L2_S_PROTO);
3286 		if (prot & VM_PROT_WRITE)
3287 			npte = l2pte_set_writable(npte);
3288 
3289 		/*
3290 		 * Make sure the vector table is mapped cacheable
3291 		 */
3292 		if ((vector_page_p && !kpm_p)
3293 		    || (flags & ARM32_MMAP_CACHEABLE)) {
3294 			npte |= pte_l2_s_cache_mode;
3295 #ifdef ARM_MMU_EXTENDED
3296 			npte &= ~L2_XS_XN;	/* and executable */
3297 #endif
3298 		} else if (flags & ARM32_MMAP_WRITECOMBINE) {
3299 			npte |= pte_l2_s_wc_mode;
3300 		}
3301 		if (opg) {
3302 			/*
3303 			 * Looks like there's an existing 'managed' mapping
3304 			 * at this address.
3305 			 */
3306 			struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3307 			paddr_t opa = VM_PAGE_TO_PHYS(opg);
3308 
3309 			pmap_acquire_page_lock(omd);
3310 			old_pv = pmap_remove_pv(omd, opa, pm, va);
3311 			pmap_vac_me_harder(omd, opa, pm, 0);
3312 			oflags = old_pv->pv_flags;
3313 			pmap_release_page_lock(omd);
3314 
3315 #ifdef PMAP_CACHE_VIVT
3316 			if (!(oflags & PVF_NC) && l2pte_valid_p(opte)) {
3317 				pmap_cache_wbinv_page(pm, va, true, oflags);
3318 			}
3319 #endif
3320 		}
3321 	}
3322 
3323 	/*
3324 	 * Make sure userland mappings get the right permissions
3325 	 */
3326 	if (!vector_page_p && !kpm_p) {
3327 		npte |= L2_S_PROT_U;
3328 #ifdef ARM_MMU_EXTENDED
3329 		npte |= L2_XS_nG;	/* user pages are not global */
3330 #endif
3331 	}
3332 
3333 	/*
3334 	 * Keep the stats up to date
3335 	 */
3336 	if (opte == 0) {
3337 		l2b->l2b_occupancy += PAGE_SIZE / L2_S_SIZE;
3338 		pm->pm_stats.resident_count++;
3339 	}
3340 
3341 	UVMHIST_LOG(maphist, " opte %#jx npte %#jx", opte, npte, 0, 0);
3342 
3343 #if defined(ARM_MMU_EXTENDED)
3344 	/*
3345 	 * If exec protection was requested but the page hasn't been synced,
3346 	 * sync it now and allow execution from it.
3347 	 */
3348 	if ((nflags & PVF_EXEC) && (npte & L2_XS_XN)) {
3349 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3350 		npte &= ~L2_XS_XN;
3351 		pmap_syncicache_page(md, pa);
3352 		PMAPCOUNT(exec_synced_map);
3353 	}
3354 #endif
3355 	/*
3356 	 * If this is just a wiring change, the two PTEs will be
3357 	 * identical, so there's no need to update the page table.
3358 	 */
3359 	if (npte != opte) {
3360 		l2pte_reset(ptep);
3361 		PTE_SYNC(ptep);
3362 		if (l2pte_valid_p(opte)) {
3363 			pmap_tlb_flush_SE(pm, va, oflags);
3364 		}
3365 		l2pte_set(ptep, npte, 0);
3366 		PTE_SYNC(ptep);
3367 #ifndef ARM_MMU_EXTENDED
3368 		bool is_cached = pmap_is_cached(pm);
3369 		if (is_cached) {
3370 			/*
3371 			 * We only need to frob the cache/tlb if this pmap
3372 			 * is current
3373 			 */
3374 			if (!vector_page_p && l2pte_valid_p(npte)) {
3375 				/*
3376 				 * This mapping is likely to be accessed as
3377 				 * soon as we return to userland. Fix up the
3378 				 * L1 entry to avoid taking another
3379 				 * page/domain fault.
3380 				 */
3381 				pd_entry_t *pdep = pmap_l1_kva(pm)
3382 				     + l1pte_index(va);
3383 				pd_entry_t pde = L1_C_PROTO | l2b->l2b_pa
3384 				    | L1_C_DOM(pmap_domain(pm));
3385 				if (*pdep != pde) {
3386 					l1pte_setone(pdep, pde);
3387 					PDE_SYNC(pdep);
3388 				}
3389 			}
3390 		}
3391 
3392 		UVMHIST_LOG(maphist, "  is_cached %jd cs 0x%08jx",
3393 		    is_cached, pm->pm_cstate.cs_all, 0, 0);
3394 
3395 		if (pg != NULL) {
3396 			struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3397 
3398 			pmap_acquire_page_lock(md);
3399 			pmap_vac_me_harder(md, pa, pm, va);
3400 			pmap_release_page_lock(md);
3401 		}
3402 #endif
3403 	}
3404 #if defined(PMAP_CACHE_VIPT) && defined(DIAGNOSTIC)
3405 	if (pg) {
3406 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3407 
3408 		pmap_acquire_page_lock(md);
3409 #ifndef ARM_MMU_EXTENDED
3410 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
3411 #endif
3412 		PMAP_VALIDATE_MD_PAGE(md);
3413 		pmap_release_page_lock(md);
3414 	}
3415 #endif
3416 
3417 	pmap_release_pmap_lock(pm);
3418 
3419 
3420 	if (old_pv)
3421 		pool_put(&pmap_pv_pool, old_pv);
3422 free_pv:
3423 	if (new_pv)
3424 		pool_put(&pmap_pv_pool, new_pv);
3425 	return error;
3426 }
3427 
3428 /*
3429  * pmap_remove()
3430  *
3431  * pmap_remove is responsible for nuking a number of mappings for a range
3432  * of virtual address space in the current pmap. To do this efficiently
3433  * is interesting, because in a number of cases a wide virtual address
3434  * range may be supplied that contains few actual mappings. So, the
3435  * optimisations are:
3436  *  1. Skip over hunks of address space for which no L1 or L2 entry exists.
3437  *  2. Build up a list of pages we've hit, up to a maximum, so we can
3438  *     maybe do just a partial cache clean. This path of execution is
3439  *     complicated by the fact that the cache must be flushed _before_
3440  *     the PTE is nuked, being a VAC :-)
3441  *  3. If we're called after UVM calls pmap_remove_all(), we can defer
3442  *     all invalidations until pmap_update(), since pmap_remove_all() has
3443  *     already flushed the cache.
3444  *  4. Maybe later fast-case a single page, but I don't think this is
3445  *     going to make _that_ much difference overall.
3446  */
3447 
3448 #define	PMAP_REMOVE_CLEAN_LIST_SIZE	3
3449 
3450 void
3451 pmap_remove(pmap_t pm, vaddr_t sva, vaddr_t eva)
3452 {
3453 	SLIST_HEAD(,pv_entry) opv_list;
3454 	struct pv_entry *pv, *npv;
3455 	UVMHIST_FUNC(__func__);
3456 	UVMHIST_CALLARGS(maphist, " (pm=%#jx, sva=%#jx, eva=%#jx)",
3457 	    (uintptr_t)pm, sva, eva, 0);
3458 
3459 #ifdef PMAP_FAULTINFO
3460 	curpcb->pcb_faultinfo.pfi_faultaddr = 0;
3461 	curpcb->pcb_faultinfo.pfi_repeats = 0;
3462 	curpcb->pcb_faultinfo.pfi_faultptep = NULL;
3463 #endif
3464 
3465 	SLIST_INIT(&opv_list);
3466 	/*
3467 	 * we lock in the pmap => pv_head direction
3468 	 */
3469 	pmap_acquire_pmap_lock(pm);
3470 
3471 #ifndef ARM_MMU_EXTENDED
3472 	u_int cleanlist_idx, total, cnt;
3473 	struct {
3474 		vaddr_t va;
3475 		pt_entry_t *ptep;
3476 	} cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
3477 
3478 	if (pm->pm_remove_all || !pmap_is_cached(pm)) {
3479 		cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
3480 		if (pm->pm_cstate.cs_tlb == 0)
3481 			pm->pm_remove_all = true;
3482 	} else
3483 		cleanlist_idx = 0;
3484 	total = 0;
3485 #endif
3486 
3487 	while (sva < eva) {
3488 		/*
3489 		 * Do one L2 bucket's worth at a time.
3490 		 */
3491 		vaddr_t next_bucket = L2_NEXT_BUCKET_VA(sva);
3492 		if (next_bucket > eva)
3493 			next_bucket = eva;
3494 
3495 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pm, sva);
3496 		if (l2b == NULL) {
3497 			sva = next_bucket;
3498 			continue;
3499 		}
3500 
3501 		pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(sva)];
3502 		u_int mappings = 0;
3503 
3504 		for (;sva < next_bucket;
3505 		     sva += PAGE_SIZE, ptep += PAGE_SIZE / L2_S_SIZE) {
3506 			pt_entry_t opte = *ptep;
3507 
3508 			if (opte == 0) {
3509 				/* Nothing here, move along */
3510 				continue;
3511 			}
3512 
3513 			u_int flags = PVF_REF;
3514 			paddr_t pa = l2pte_pa(opte);
3515 			struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
3516 
3517 			/*
3518 			 * Update flags. In a number of circumstances,
3519 			 * we could cluster a lot of these and do a
3520 			 * number of sequential pages in one go.
3521 			 */
3522 			if (pg != NULL) {
3523 				struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3524 
3525 				pmap_acquire_page_lock(md);
3526 				pv = pmap_remove_pv(md, pa, pm, sva);
3527 				pmap_vac_me_harder(md, pa, pm, 0);
3528 				pmap_release_page_lock(md);
3529 				if (pv != NULL) {
3530 					if (pm->pm_remove_all == false) {
3531 						flags = pv->pv_flags;
3532 					}
3533 					SLIST_INSERT_HEAD(&opv_list,
3534 					    pv, pv_link);
3535 				}
3536 			}
3537 			mappings += PAGE_SIZE / L2_S_SIZE;
3538 
3539 			if (!l2pte_valid_p(opte)) {
3540 				/*
3541 				 * Ref/Mod emulation is still active for this
3542 				 * mapping, therefore it is has not yet been
3543 				 * accessed. No need to frob the cache/tlb.
3544 				 */
3545 				l2pte_reset(ptep);
3546 				PTE_SYNC_CURRENT(pm, ptep);
3547 				continue;
3548 			}
3549 
3550 #ifdef ARM_MMU_EXTENDED
3551 			l2pte_reset(ptep);
3552 			PTE_SYNC(ptep);
3553 			if (__predict_false(pm->pm_remove_all == false)) {
3554 				pmap_tlb_flush_SE(pm, sva, flags);
3555 			}
3556 #else
3557 			if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
3558 				/* Add to the clean list. */
3559 				cleanlist[cleanlist_idx].ptep = ptep;
3560 				cleanlist[cleanlist_idx].va =
3561 				    sva | (flags & PVF_EXEC);
3562 				cleanlist_idx++;
3563 			} else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
3564 				/* Nuke everything if needed. */
3565 #ifdef PMAP_CACHE_VIVT
3566 				pmap_cache_wbinv_all(pm, PVF_EXEC);
3567 #endif
3568 				/*
3569 				 * Roll back the previous PTE list,
3570 				 * and zero out the current PTE.
3571 				 */
3572 				for (cnt = 0;
3573 				     cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
3574 					l2pte_reset(cleanlist[cnt].ptep);
3575 					PTE_SYNC(cleanlist[cnt].ptep);
3576 				}
3577 				l2pte_reset(ptep);
3578 				PTE_SYNC(ptep);
3579 				cleanlist_idx++;
3580 				pm->pm_remove_all = true;
3581 			} else {
3582 				l2pte_reset(ptep);
3583 				PTE_SYNC(ptep);
3584 				if (pm->pm_remove_all == false) {
3585 					pmap_tlb_flush_SE(pm, sva, flags);
3586 				}
3587 			}
3588 #endif
3589 		}
3590 
3591 #ifndef ARM_MMU_EXTENDED
3592 		/*
3593 		 * Deal with any left overs
3594 		 */
3595 		if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
3596 			total += cleanlist_idx;
3597 			for (cnt = 0; cnt < cleanlist_idx; cnt++) {
3598 				l2pte_reset(cleanlist[cnt].ptep);
3599 				PTE_SYNC_CURRENT(pm, cleanlist[cnt].ptep);
3600 				vaddr_t va = cleanlist[cnt].va;
3601 				if (pm->pm_cstate.cs_all != 0) {
3602 					vaddr_t clva = va & ~PAGE_MASK;
3603 					u_int flags = va & PVF_EXEC;
3604 #ifdef PMAP_CACHE_VIVT
3605 					pmap_cache_wbinv_page(pm, clva, true,
3606 					    PVF_REF | PVF_WRITE | flags);
3607 #endif
3608 					pmap_tlb_flush_SE(pm, clva,
3609 					    PVF_REF | flags);
3610 				}
3611 			}
3612 
3613 			/*
3614 			 * If it looks like we're removing a whole bunch
3615 			 * of mappings, it's faster to just write-back
3616 			 * the whole cache now and defer TLB flushes until
3617 			 * pmap_update() is called.
3618 			 */
3619 			if (total <= PMAP_REMOVE_CLEAN_LIST_SIZE)
3620 				cleanlist_idx = 0;
3621 			else {
3622 				cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
3623 #ifdef PMAP_CACHE_VIVT
3624 				pmap_cache_wbinv_all(pm, PVF_EXEC);
3625 #endif
3626 				pm->pm_remove_all = true;
3627 			}
3628 		}
3629 #endif /* ARM_MMU_EXTENDED */
3630 
3631 		pmap_free_l2_bucket(pm, l2b, mappings);
3632 		pm->pm_stats.resident_count -= mappings / (PAGE_SIZE/L2_S_SIZE);
3633 	}
3634 
3635 	pmap_release_pmap_lock(pm);
3636 	SLIST_FOREACH_SAFE(pv, &opv_list, pv_link, npv) {
3637 		pool_put(&pmap_pv_pool, pv);
3638 	}
3639 }
3640 
3641 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3642 static struct pv_entry *
3643 pmap_kremove_pg(struct vm_page *pg, vaddr_t va)
3644 {
3645 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3646 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
3647 	struct pv_entry *pv;
3648 
3649 	KASSERT(arm_cache_prefer_mask == 0 || md->pvh_attrs & (PVF_COLORED|PVF_NC));
3650 	KASSERT((md->pvh_attrs & PVF_KMPAGE) == 0);
3651 	KASSERT(pmap_page_locked_p(md));
3652 
3653 	pv = pmap_remove_pv(md, pa, pmap_kernel(), va);
3654 	KASSERTMSG(pv, "pg %p (pa #%lx) va %#lx", pg, pa, va);
3655 	KASSERT(PV_IS_KENTRY_P(pv->pv_flags));
3656 
3657 	/*
3658 	 * We are removing a writeable mapping to a cached exec page, if
3659 	 * it's the last mapping then clear its execness otherwise sync
3660 	 * the page to the icache.
3661 	 */
3662 	if ((md->pvh_attrs & (PVF_NC|PVF_EXEC)) == PVF_EXEC
3663 	    && (pv->pv_flags & PVF_WRITE) != 0) {
3664 		if (SLIST_EMPTY(&md->pvh_list)) {
3665 			md->pvh_attrs &= ~PVF_EXEC;
3666 			PMAPCOUNT(exec_discarded_kremove);
3667 		} else {
3668 			pmap_syncicache_page(md, pa);
3669 			PMAPCOUNT(exec_synced_kremove);
3670 		}
3671 	}
3672 	pmap_vac_me_harder(md, pa, pmap_kernel(), 0);
3673 
3674 	return pv;
3675 }
3676 #endif /* PMAP_CACHE_VIPT && !ARM_MMU_EXTENDED */
3677 
3678 /*
3679  * pmap_kenter_pa: enter an unmanaged, wired kernel mapping
3680  *
3681  * We assume there is already sufficient KVM space available
3682  * to do this, as we can't allocate L2 descriptor tables/metadata
3683  * from here.
3684  */
3685 void
3686 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags)
3687 {
3688 #ifdef PMAP_CACHE_VIVT
3689 	struct vm_page *pg = (flags & PMAP_KMPAGE) ? PHYS_TO_VM_PAGE(pa) : NULL;
3690 #endif
3691 #ifdef PMAP_CACHE_VIPT
3692 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
3693 	struct vm_page *opg;
3694 #ifndef ARM_MMU_EXTENDED
3695 	struct pv_entry *pv = NULL;
3696 #endif
3697 #endif
3698 	struct vm_page_md *md = pg != NULL ? VM_PAGE_TO_MD(pg) : NULL;
3699 
3700 	UVMHIST_FUNC(__func__);
3701 
3702 	if (pmap_initialized) {
3703 		UVMHIST_CALLARGS(maphist,
3704 		    "va=%#jx, pa=%#jx, prot=%#jx, flags=%#jx", va, pa, prot,
3705 		     flags);
3706 	}
3707 
3708 	pmap_t kpm = pmap_kernel();
3709 	pmap_acquire_pmap_lock(kpm);
3710 	struct l2_bucket * const l2b = pmap_get_l2_bucket(kpm, va);
3711 	const size_t l1slot __diagused = l1pte_index(va);
3712 	KASSERTMSG(l2b != NULL,
3713 	    "va %#lx pa %#lx prot %d maxkvaddr %#lx: l2 %p l2b %p kva %p",
3714 	    va, pa, prot, pmap_curmaxkvaddr, kpm->pm_l2[L2_IDX(l1slot)],
3715 	    kpm->pm_l2[L2_IDX(l1slot)]
3716 		? &kpm->pm_l2[L2_IDX(l1slot)]->l2_bucket[L2_BUCKET(l1slot)]
3717 		: NULL,
3718 	    kpm->pm_l2[L2_IDX(l1slot)]
3719 		? kpm->pm_l2[L2_IDX(l1slot)]->l2_bucket[L2_BUCKET(l1slot)].l2b_kva
3720 		: NULL);
3721 	KASSERT(l2b->l2b_kva != NULL);
3722 
3723 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
3724 	const pt_entry_t opte = *ptep;
3725 
3726 	if (opte == 0) {
3727 		PMAPCOUNT(kenter_mappings);
3728 		l2b->l2b_occupancy += PAGE_SIZE / L2_S_SIZE;
3729 	} else {
3730 		PMAPCOUNT(kenter_remappings);
3731 #ifdef PMAP_CACHE_VIPT
3732 		opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3733 #if !defined(ARM_MMU_EXTENDED) || defined(DIAGNOSTIC)
3734 		struct vm_page_md *omd __diagused = VM_PAGE_TO_MD(opg);
3735 #endif
3736 		if (opg && arm_cache_prefer_mask != 0) {
3737 			KASSERT(opg != pg);
3738 			KASSERT((omd->pvh_attrs & PVF_KMPAGE) == 0);
3739 			KASSERT((flags & PMAP_KMPAGE) == 0);
3740 #ifndef ARM_MMU_EXTENDED
3741 			pmap_acquire_page_lock(omd);
3742 			pv = pmap_kremove_pg(opg, va);
3743 			pmap_release_page_lock(omd);
3744 #endif
3745 		}
3746 #endif
3747 		if (l2pte_valid_p(opte)) {
3748 			l2pte_reset(ptep);
3749 			PTE_SYNC(ptep);
3750 #ifdef PMAP_CACHE_VIVT
3751 			cpu_dcache_wbinv_range(va, PAGE_SIZE);
3752 #endif
3753 			cpu_tlb_flushD_SE(va);
3754 			cpu_cpwait();
3755 		}
3756 	}
3757 	pmap_release_pmap_lock(kpm);
3758 	pt_entry_t npte = L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot);
3759 
3760 	if (flags & PMAP_PTE) {
3761 		KASSERT((flags & PMAP_CACHE_MASK) == 0);
3762 		if (!(flags & PMAP_NOCACHE))
3763 			npte |= pte_l2_s_cache_mode_pt;
3764 	} else {
3765 		switch (flags & (PMAP_CACHE_MASK | PMAP_DEV_MASK)) {
3766 		case PMAP_DEV ... PMAP_DEV | PMAP_CACHE_MASK:
3767 			break;
3768 		case PMAP_NOCACHE:
3769 			npte |= pte_l2_s_nocache_mode;
3770 			break;
3771 		case PMAP_WRITE_COMBINE:
3772 			npte |= pte_l2_s_wc_mode;
3773 			break;
3774 		default:
3775 			npte |= pte_l2_s_cache_mode;
3776 			break;
3777 		}
3778 	}
3779 #ifdef ARM_MMU_EXTENDED
3780 	if (prot & VM_PROT_EXECUTE)
3781 		npte &= ~L2_XS_XN;
3782 #endif
3783 	l2pte_set(ptep, npte, 0);
3784 	PTE_SYNC(ptep);
3785 
3786 	if (pg) {
3787 		if (flags & PMAP_KMPAGE) {
3788 			KASSERT(md->urw_mappings == 0);
3789 			KASSERT(md->uro_mappings == 0);
3790 			KASSERT(md->krw_mappings == 0);
3791 			KASSERT(md->kro_mappings == 0);
3792 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3793 			KASSERT(pv == NULL);
3794 			KASSERT(arm_cache_prefer_mask == 0 || (va & PVF_COLORED) == 0);
3795 			KASSERT((md->pvh_attrs & PVF_NC) == 0);
3796 			/* if there is a color conflict, evict from cache. */
3797 			if (pmap_is_page_colored_p(md)
3798 			    && ((va ^ md->pvh_attrs) & arm_cache_prefer_mask)) {
3799 				PMAPCOUNT(vac_color_change);
3800 				pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
3801 			} else if (md->pvh_attrs & PVF_MULTCLR) {
3802 				/*
3803 				 * If this page has multiple colors, expunge
3804 				 * them.
3805 				 */
3806 				PMAPCOUNT(vac_flush_lots2);
3807 				pmap_flush_page(md, pa, PMAP_FLUSH_SECONDARY);
3808 			}
3809 			/*
3810 			 * Since this is a KMPAGE, there can be no contention
3811 			 * for this page so don't lock it.
3812 			 */
3813 			md->pvh_attrs &= PAGE_SIZE - 1;
3814 			md->pvh_attrs |= PVF_KMPAGE | PVF_COLORED | PVF_DIRTY
3815 			    | (va & arm_cache_prefer_mask);
3816 #else /* !PMAP_CACHE_VIPT || ARM_MMU_EXTENDED */
3817 			md->pvh_attrs |= PVF_KMPAGE;
3818 #endif
3819 			atomic_inc_32(&pmap_kmpages);
3820 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3821 		} else if (arm_cache_prefer_mask != 0) {
3822 			if (pv == NULL) {
3823 				pv = pool_get(&pmap_pv_pool, PR_NOWAIT);
3824 				KASSERT(pv != NULL);
3825 			}
3826 			pmap_acquire_page_lock(md);
3827 			pmap_enter_pv(md, pa, pv, pmap_kernel(), va,
3828 			    PVF_WIRED | PVF_KENTRY
3829 			    | (prot & VM_PROT_WRITE ? PVF_WRITE : 0));
3830 			if ((prot & VM_PROT_WRITE)
3831 			    && !(md->pvh_attrs & PVF_NC))
3832 				md->pvh_attrs |= PVF_DIRTY;
3833 			KASSERT((prot & VM_PROT_WRITE) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
3834 			pmap_vac_me_harder(md, pa, pmap_kernel(), va);
3835 			pmap_release_page_lock(md);
3836 #endif
3837 		}
3838 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3839 	} else {
3840 		if (pv != NULL)
3841 			pool_put(&pmap_pv_pool, pv);
3842 #endif
3843 	}
3844 	if (pmap_initialized) {
3845 		UVMHIST_LOG(maphist, "  <-- done (ptep %#jx: %#jx -> %#jx)",
3846 		    (uintptr_t)ptep, opte, npte, 0);
3847 	}
3848 
3849 }
3850 
3851 void
3852 pmap_kremove(vaddr_t va, vsize_t len)
3853 {
3854 #ifdef UVMHIST
3855 	u_int total_mappings = 0;
3856 #endif
3857 
3858 	PMAPCOUNT(kenter_unmappings);
3859 
3860 	UVMHIST_FUNC(__func__);
3861 	UVMHIST_CALLARGS(maphist, " (va=%#jx, len=%#jx)", va, len, 0, 0);
3862 
3863 	const vaddr_t eva = va + len;
3864 	pmap_t kpm = pmap_kernel();
3865 
3866 	pmap_acquire_pmap_lock(kpm);
3867 
3868 	while (va < eva) {
3869 		vaddr_t next_bucket = L2_NEXT_BUCKET_VA(va);
3870 		if (next_bucket > eva)
3871 			next_bucket = eva;
3872 
3873 		struct l2_bucket * const l2b = pmap_get_l2_bucket(kpm, va);
3874 		KDASSERT(l2b != NULL);
3875 
3876 		pt_entry_t * const sptep = &l2b->l2b_kva[l2pte_index(va)];
3877 		pt_entry_t *ptep = sptep;
3878 		u_int mappings = 0;
3879 
3880 		while (va < next_bucket) {
3881 			const pt_entry_t opte = *ptep;
3882 			struct vm_page *opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3883 			if (opg != NULL) {
3884 				struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3885 
3886 				if (omd->pvh_attrs & PVF_KMPAGE) {
3887 					KASSERT(omd->urw_mappings == 0);
3888 					KASSERT(omd->uro_mappings == 0);
3889 					KASSERT(omd->krw_mappings == 0);
3890 					KASSERT(omd->kro_mappings == 0);
3891 					omd->pvh_attrs &= ~PVF_KMPAGE;
3892 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3893 					if (arm_cache_prefer_mask != 0) {
3894 						omd->pvh_attrs &= ~PVF_WRITE;
3895 					}
3896 #endif
3897 					atomic_dec_32(&pmap_kmpages);
3898 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3899 				} else if (arm_cache_prefer_mask != 0) {
3900 					pmap_acquire_page_lock(omd);
3901 					pool_put(&pmap_pv_pool,
3902 					    pmap_kremove_pg(opg, va));
3903 					pmap_release_page_lock(omd);
3904 #endif
3905 				}
3906 			}
3907 			if (l2pte_valid_p(opte)) {
3908 				l2pte_reset(ptep);
3909 				PTE_SYNC(ptep);
3910 #ifdef PMAP_CACHE_VIVT
3911 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
3912 #endif
3913 				cpu_tlb_flushD_SE(va);
3914 
3915 				mappings += PAGE_SIZE / L2_S_SIZE;
3916 			}
3917 			va += PAGE_SIZE;
3918 			ptep += PAGE_SIZE / L2_S_SIZE;
3919 		}
3920 		KDASSERTMSG(mappings <= l2b->l2b_occupancy, "%u %u",
3921 		    mappings, l2b->l2b_occupancy);
3922 		l2b->l2b_occupancy -= mappings;
3923 		//PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
3924 #ifdef UVMHIST
3925 		total_mappings += mappings;
3926 #endif
3927 	}
3928 	pmap_release_pmap_lock(kpm);
3929 	cpu_cpwait();
3930 	UVMHIST_LOG(maphist, "  <--- done (%ju mappings removed)",
3931 	    total_mappings, 0, 0, 0);
3932 }
3933 
3934 bool
3935 pmap_extract(pmap_t pm, vaddr_t va, paddr_t *pap)
3936 {
3937 
3938 	return pmap_extract_coherency(pm, va, pap, NULL);
3939 }
3940 
3941 bool
3942 pmap_extract_coherency(pmap_t pm, vaddr_t va, paddr_t *pap, bool *coherentp)
3943 {
3944 	struct l2_dtable *l2;
3945 	pd_entry_t *pdep, pde;
3946 	pt_entry_t *ptep, pte;
3947 	paddr_t pa;
3948 	u_int l1slot;
3949 	bool coherent;
3950 
3951 	pmap_acquire_pmap_lock(pm);
3952 
3953 	l1slot = l1pte_index(va);
3954 	pdep = pmap_l1_kva(pm) + l1slot;
3955 	pde = *pdep;
3956 
3957 	if (l1pte_section_p(pde)) {
3958 		/*
3959 		 * These should only happen for pmap_kernel()
3960 		 */
3961 		KDASSERT(pm == pmap_kernel());
3962 		pmap_release_pmap_lock(pm);
3963 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
3964 		if (l1pte_supersection_p(pde)) {
3965 			pa = (pde & L1_SS_FRAME) | (va & L1_SS_OFFSET);
3966 		} else
3967 #endif
3968 			pa = (pde & L1_S_FRAME) | (va & L1_S_OFFSET);
3969 		coherent = (pde & L1_S_CACHE_MASK) == 0;
3970 	} else {
3971 		/*
3972 		 * Note that we can't rely on the validity of the L1
3973 		 * descriptor as an indication that a mapping exists.
3974 		 * We have to look it up in the L2 dtable.
3975 		 */
3976 		l2 = pm->pm_l2[L2_IDX(l1slot)];
3977 
3978 		if (l2 == NULL ||
3979 		    (ptep = l2->l2_bucket[L2_BUCKET(l1slot)].l2b_kva) == NULL) {
3980 			pmap_release_pmap_lock(pm);
3981 			return false;
3982 		}
3983 
3984 		pte = ptep[l2pte_index(va)];
3985 		pmap_release_pmap_lock(pm);
3986 
3987 		if (pte == 0)
3988 			return false;
3989 
3990 		switch (pte & L2_TYPE_MASK) {
3991 		case L2_TYPE_L:
3992 			pa = (pte & L2_L_FRAME) | (va & L2_L_OFFSET);
3993 			coherent = (pte & L2_L_CACHE_MASK) == 0;
3994 			break;
3995 
3996 		default:
3997 			pa = (pte & ~PAGE_MASK) | (va & PAGE_MASK);
3998 			coherent = (pte & L2_S_CACHE_MASK) == 0;
3999 			break;
4000 		}
4001 	}
4002 
4003 	if (pap != NULL)
4004 		*pap = pa;
4005 
4006 	if (coherentp != NULL)
4007 		*coherentp = (pm == pmap_kernel() && coherent);
4008 
4009 	return true;
4010 }
4011 
4012 /*
4013  * pmap_pv_remove: remove an unmanaged pv-tracked page from all pmaps
4014  *	that map it
4015  */
4016 
4017 static void
4018 pmap_pv_remove(paddr_t pa)
4019 {
4020 	struct pmap_page *pp;
4021 
4022 	pp = pmap_pv_tracked(pa);
4023 	if (pp == NULL)
4024 		panic("pmap_pv_protect: page not pv-tracked: 0x%"PRIxPADDR,
4025 		    pa);
4026 
4027 	struct vm_page_md *md = PMAP_PAGE_TO_MD(pp);
4028 	pmap_page_remove(md, pa);
4029 }
4030 
4031 void
4032 pmap_pv_protect(paddr_t pa, vm_prot_t prot)
4033 {
4034 
4035 	/* the only case is remove at the moment */
4036 	KASSERT(prot == VM_PROT_NONE);
4037 	pmap_pv_remove(pa);
4038 }
4039 
4040 void
4041 pmap_protect(pmap_t pm, vaddr_t sva, vaddr_t eva, vm_prot_t prot)
4042 {
4043 	struct l2_bucket *l2b;
4044 	vaddr_t next_bucket;
4045 
4046 	UVMHIST_FUNC(__func__);
4047 	UVMHIST_CALLARGS(maphist, "pm %#jx va %#jx...#%jx prot %#jx",
4048 	    (uintptr_t)pm, sva, eva, prot);
4049 
4050 	if ((prot & VM_PROT_READ) == 0) {
4051 		pmap_remove(pm, sva, eva);
4052 		return;
4053 	}
4054 
4055 	if (prot & VM_PROT_WRITE) {
4056 		/*
4057 		 * If this is a read->write transition, just ignore it and let
4058 		 * uvm_fault() take care of it later.
4059 		 */
4060 		return;
4061 	}
4062 
4063 	pmap_acquire_pmap_lock(pm);
4064 
4065 #ifndef ARM_MMU_EXTENDED
4066 	const bool flush = eva - sva >= PAGE_SIZE * 4;
4067 	u_int flags = 0;
4068 #endif
4069 	u_int clr_mask = PVF_WRITE | ((prot & VM_PROT_EXECUTE) ? 0 : PVF_EXEC);
4070 
4071 	while (sva < eva) {
4072 		next_bucket = L2_NEXT_BUCKET_VA(sva);
4073 		if (next_bucket > eva)
4074 			next_bucket = eva;
4075 
4076 		l2b = pmap_get_l2_bucket(pm, sva);
4077 		if (l2b == NULL) {
4078 			sva = next_bucket;
4079 			continue;
4080 		}
4081 
4082 		pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(sva)];
4083 
4084 		while (sva < next_bucket) {
4085 			const pt_entry_t opte = *ptep;
4086 			if (l2pte_valid_p(opte) && l2pte_writable_p(opte)) {
4087 				struct vm_page *pg;
4088 #ifndef ARM_MMU_EXTENDED
4089 				u_int f;
4090 #endif
4091 
4092 #ifdef PMAP_CACHE_VIVT
4093 				/*
4094 				 * OK, at this point, we know we're doing
4095 				 * write-protect operation.  If the pmap is
4096 				 * active, write-back the page.
4097 				 */
4098 				pmap_cache_wbinv_page(pm, sva, false,
4099 				    PVF_REF | PVF_WRITE);
4100 #endif
4101 
4102 				pg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
4103 				pt_entry_t npte = l2pte_set_readonly(opte);
4104 				l2pte_reset(ptep);
4105 				PTE_SYNC(ptep);
4106 #ifdef ARM_MMU_EXTENDED
4107 				pmap_tlb_flush_SE(pm, sva, PVF_REF);
4108 #endif
4109 				l2pte_set(ptep, npte, 0);
4110 				PTE_SYNC(ptep);
4111 
4112 				if (pg != NULL) {
4113 					struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4114 					paddr_t pa = VM_PAGE_TO_PHYS(pg);
4115 
4116 					pmap_acquire_page_lock(md);
4117 #ifndef ARM_MMU_EXTENDED
4118 					f =
4119 #endif
4120 					    pmap_modify_pv(md, pa, pm, sva,
4121 					       clr_mask, 0);
4122 					pmap_vac_me_harder(md, pa, pm, sva);
4123 					pmap_release_page_lock(md);
4124 #ifndef ARM_MMU_EXTENDED
4125 				} else {
4126 					f = PVF_REF | PVF_EXEC;
4127 				}
4128 
4129 				if (flush) {
4130 					flags |= f;
4131 				} else {
4132 					pmap_tlb_flush_SE(pm, sva, f);
4133 #endif
4134 				}
4135 			}
4136 
4137 			sva += PAGE_SIZE;
4138 			ptep += PAGE_SIZE / L2_S_SIZE;
4139 		}
4140 	}
4141 
4142 #ifndef ARM_MMU_EXTENDED
4143 	if (flush) {
4144 		if (PV_BEEN_EXECD(flags)) {
4145 			pmap_tlb_flushID(pm);
4146 		} else if (PV_BEEN_REFD(flags)) {
4147 			pmap_tlb_flushD(pm);
4148 		}
4149 	}
4150 #endif
4151 
4152 	pmap_release_pmap_lock(pm);
4153 }
4154 
4155 void
4156 pmap_icache_sync_range(pmap_t pm, vaddr_t sva, vaddr_t eva)
4157 {
4158 	struct l2_bucket *l2b;
4159 	pt_entry_t *ptep;
4160 	vaddr_t next_bucket;
4161 	vsize_t page_size = trunc_page(sva) + PAGE_SIZE - sva;
4162 
4163 	UVMHIST_FUNC(__func__);
4164 	UVMHIST_CALLARGS(maphist, "pm %#jx va %#jx...#%jx",
4165 	    (uintptr_t)pm, sva, eva, 0);
4166 
4167 	pmap_acquire_pmap_lock(pm);
4168 
4169 	while (sva < eva) {
4170 		next_bucket = L2_NEXT_BUCKET_VA(sva);
4171 		if (next_bucket > eva)
4172 			next_bucket = eva;
4173 
4174 		l2b = pmap_get_l2_bucket(pm, sva);
4175 		if (l2b == NULL) {
4176 			sva = next_bucket;
4177 			continue;
4178 		}
4179 
4180 		for (ptep = &l2b->l2b_kva[l2pte_index(sva)];
4181 		     sva < next_bucket;
4182 		     sva += page_size,
4183 		     ptep += PAGE_SIZE / L2_S_SIZE,
4184 		     page_size = PAGE_SIZE) {
4185 			if (l2pte_valid_p(*ptep)) {
4186 				cpu_icache_sync_range(sva,
4187 				    uimin(page_size, eva - sva));
4188 			}
4189 		}
4190 	}
4191 
4192 	pmap_release_pmap_lock(pm);
4193 }
4194 
4195 void
4196 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
4197 {
4198 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4199 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
4200 
4201 	UVMHIST_FUNC(__func__);
4202 	UVMHIST_CALLARGS(maphist, "md %#jx pa %#jx prot %#jx",
4203 	    (uintptr_t)md, pa, prot, 0);
4204 
4205 	switch(prot) {
4206 	case VM_PROT_READ|VM_PROT_WRITE:
4207 #if defined(ARM_MMU_EXTENDED)
4208 		pmap_acquire_page_lock(md);
4209 		pmap_clearbit(md, pa, PVF_EXEC);
4210 		pmap_release_page_lock(md);
4211 		break;
4212 #endif
4213 	case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
4214 		break;
4215 
4216 	case VM_PROT_READ:
4217 #if defined(ARM_MMU_EXTENDED)
4218 		pmap_acquire_page_lock(md);
4219 		pmap_clearbit(md, pa, PVF_WRITE|PVF_EXEC);
4220 		pmap_release_page_lock(md);
4221 		break;
4222 #endif
4223 	case VM_PROT_READ|VM_PROT_EXECUTE:
4224 		pmap_acquire_page_lock(md);
4225 		pmap_clearbit(md, pa, PVF_WRITE);
4226 		pmap_release_page_lock(md);
4227 		break;
4228 
4229 	default:
4230 		pmap_page_remove(md, pa);
4231 		break;
4232 	}
4233 }
4234 
4235 /*
4236  * pmap_clear_modify:
4237  *
4238  *	Clear the "modified" attribute for a page.
4239  */
4240 bool
4241 pmap_clear_modify(struct vm_page *pg)
4242 {
4243 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4244 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
4245 	bool rv;
4246 
4247 	pmap_acquire_page_lock(md);
4248 
4249 	if (md->pvh_attrs & PVF_MOD) {
4250 		rv = true;
4251 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
4252 		/*
4253 		 * If we are going to clear the modified bit and there are
4254 		 * no other modified bits set, flush the page to memory and
4255 		 * mark it clean.
4256 		 */
4257 		if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) == PVF_MOD)
4258 			pmap_flush_page(md, pa, PMAP_CLEAN_PRIMARY);
4259 #endif
4260 		pmap_clearbit(md, pa, PVF_MOD);
4261 	} else {
4262 		rv = false;
4263 	}
4264 	pmap_release_page_lock(md);
4265 
4266 	return rv;
4267 }
4268 
4269 /*
4270  * pmap_clear_reference:
4271  *
4272  *	Clear the "referenced" attribute for a page.
4273  */
4274 bool
4275 pmap_clear_reference(struct vm_page *pg)
4276 {
4277 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4278 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
4279 	bool rv;
4280 
4281 	pmap_acquire_page_lock(md);
4282 
4283 	if (md->pvh_attrs & PVF_REF) {
4284 		rv = true;
4285 		pmap_clearbit(md, pa, PVF_REF);
4286 	} else {
4287 		rv = false;
4288 	}
4289 	pmap_release_page_lock(md);
4290 
4291 	return rv;
4292 }
4293 
4294 /*
4295  * pmap_is_modified:
4296  *
4297  *	Test if a page has the "modified" attribute.
4298  */
4299 /* See <arm/arm32/pmap.h> */
4300 
4301 /*
4302  * pmap_is_referenced:
4303  *
4304  *	Test if a page has the "referenced" attribute.
4305  */
4306 /* See <arm/arm32/pmap.h> */
4307 
4308 #if defined(ARM_MMU_EXTENDED) && 0
4309 int
4310 pmap_prefetchabt_fixup(void *v)
4311 {
4312 	struct trapframe * const tf = v;
4313 	vaddr_t va = trunc_page(tf->tf_pc);
4314 	int rv = ABORT_FIXUP_FAILED;
4315 
4316 	if (!TRAP_USERMODE(tf) && va < VM_MAXUSER_ADDRESS)
4317 		return rv;
4318 
4319 	kpreempt_disable();
4320 	pmap_t pm = curcpu()->ci_pmap_cur;
4321 	const size_t l1slot = l1pte_index(va);
4322 	struct l2_dtable * const l2 = pm->pm_l2[L2_IDX(l1slot)];
4323 	if (l2 == NULL)
4324 		goto out;
4325 
4326 	struct l2_bucket * const l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
4327 	if (l2b->l2b_kva == NULL)
4328 		goto out;
4329 
4330 	/*
4331 	 * Check the PTE itself.
4332 	 */
4333 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
4334 	const pt_entry_t opte = *ptep;
4335 	if ((opte & L2_S_PROT_U) == 0 || (opte & L2_XS_XN) == 0)
4336 		goto out;
4337 
4338 	paddr_t pa = l2pte_pa(opte);
4339 	struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
4340 	KASSERT(pg != NULL);
4341 
4342 	struct vm_page_md * const md = VM_PAGE_TO_MD(pg);
4343 
4344 	pmap_acquire_page_lock(md);
4345 	struct pv_entry * const pv = pmap_find_pv(md, pm, va);
4346 	KASSERT(pv != NULL);
4347 
4348 	if (PV_IS_EXEC_P(pv->pv_flags)) {
4349 		l2pte_reset(ptep);
4350 		PTE_SYNC(ptep);
4351 		pmap_tlb_flush_SE(pm, va, PVF_EXEC | PVF_REF);
4352 		if (!PV_IS_EXEC_P(md->pvh_attrs)) {
4353 			pmap_syncicache_page(md, pa);
4354 		}
4355 		rv = ABORT_FIXUP_RETURN;
4356 		l2pte_set(ptep, opte & ~L2_XS_XN, 0);
4357 		PTE_SYNC(ptep);
4358 	}
4359 	pmap_release_page_lock(md);
4360 
4361   out:
4362 	kpreempt_enable();
4363 	return rv;
4364 }
4365 #endif
4366 
4367 int
4368 pmap_fault_fixup(pmap_t pm, vaddr_t va, vm_prot_t ftype, int user)
4369 {
4370 	struct l2_dtable *l2;
4371 	struct l2_bucket *l2b;
4372 	paddr_t pa;
4373 	const size_t l1slot = l1pte_index(va);
4374 	int rv = 0;
4375 
4376 	UVMHIST_FUNC(__func__);
4377 	UVMHIST_CALLARGS(maphist, "pm=%#jx, va=%#jx, ftype=%#jx, user=%jd",
4378 	    (uintptr_t)pm, va, ftype, user);
4379 
4380 	va = trunc_page(va);
4381 
4382 	KASSERT(!user || (pm != pmap_kernel()));
4383 
4384 #ifdef ARM_MMU_EXTENDED
4385 	UVMHIST_LOG(maphist, " ti=%#jx pai=%#jx asid=%#jx",
4386 	    (uintptr_t)cpu_tlb_info(curcpu()),
4387 	    (uintptr_t)PMAP_PAI(pm, cpu_tlb_info(curcpu())),
4388 	    (uintptr_t)PMAP_PAI(pm, cpu_tlb_info(curcpu()))->pai_asid, 0);
4389 #endif
4390 
4391 	pmap_acquire_pmap_lock(pm);
4392 
4393 	/*
4394 	 * If there is no l2_dtable for this address, then the process
4395 	 * has no business accessing it.
4396 	 *
4397 	 * Note: This will catch userland processes trying to access
4398 	 * kernel addresses.
4399 	 */
4400 	l2 = pm->pm_l2[L2_IDX(l1slot)];
4401 	if (l2 == NULL) {
4402 		UVMHIST_LOG(maphist, " no l2 for l1slot %#jx", l1slot, 0, 0, 0);
4403 		goto out;
4404 	}
4405 
4406 	/*
4407 	 * Likewise if there is no L2 descriptor table
4408 	 */
4409 	l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
4410 	if (l2b->l2b_kva == NULL) {
4411 		UVMHIST_LOG(maphist, " <-- done (no ptep for l1slot %#jx)",
4412 		    l1slot, 0, 0, 0);
4413 		goto out;
4414 	}
4415 
4416 	/*
4417 	 * Check the PTE itself.
4418 	 */
4419 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
4420 	pt_entry_t const opte = *ptep;
4421 	if (opte == 0 || (opte & L2_TYPE_MASK) == L2_TYPE_L) {
4422 		UVMHIST_LOG(maphist, " <-- done (empty pde for l1slot %#jx)",
4423 		    l1slot, 0, 0, 0);
4424 		goto out;
4425 	}
4426 
4427 #ifndef ARM_HAS_VBAR
4428 	/*
4429 	 * Catch a userland access to the vector page mapped at 0x0
4430 	 */
4431 	if (user && (opte & L2_S_PROT_U) == 0) {
4432 		UVMHIST_LOG(maphist, " <-- done (vector_page)", 0, 0, 0, 0);
4433 		goto out;
4434 	}
4435 #endif
4436 
4437 	pa = l2pte_pa(opte);
4438 
4439 	if ((ftype & VM_PROT_WRITE) && !l2pte_writable_p(opte)) {
4440 		/*
4441 		 * This looks like a good candidate for "page modified"
4442 		 * emulation...
4443 		 */
4444 		struct pv_entry *pv;
4445 		struct vm_page *pg;
4446 
4447 		/* Extract the physical address of the page */
4448 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) {
4449 			UVMHIST_LOG(maphist, " <-- done (mod/ref unmanaged page)", 0, 0, 0, 0);
4450 			goto out;
4451 		}
4452 
4453 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4454 
4455 		/* Get the current flags for this page. */
4456 		pmap_acquire_page_lock(md);
4457 		pv = pmap_find_pv(md, pm, va);
4458 		if (pv == NULL || PV_IS_KENTRY_P(pv->pv_flags)) {
4459 			pmap_release_page_lock(md);
4460 			UVMHIST_LOG(maphist, " <-- done (mod/ref emul: no PV)", 0, 0, 0, 0);
4461 			goto out;
4462 		}
4463 
4464 		/*
4465 		 * Do the flags say this page is writable? If not then it
4466 		 * is a genuine write fault. If yes then the write fault is
4467 		 * our fault as we did not reflect the write access in the
4468 		 * PTE. Now we know a write has occurred we can correct this
4469 		 * and also set the modified bit
4470 		 */
4471 		if ((pv->pv_flags & PVF_WRITE) == 0) {
4472 			pmap_release_page_lock(md);
4473 			goto out;
4474 		}
4475 
4476 		md->pvh_attrs |= PVF_REF | PVF_MOD;
4477 		pv->pv_flags |= PVF_REF | PVF_MOD;
4478 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
4479 		/*
4480 		 * If there are cacheable mappings for this page, mark it dirty.
4481 		 */
4482 		if ((md->pvh_attrs & PVF_NC) == 0)
4483 			md->pvh_attrs |= PVF_DIRTY;
4484 #endif
4485 #ifdef ARM_MMU_EXTENDED
4486 		if (md->pvh_attrs & PVF_EXEC) {
4487 			md->pvh_attrs &= ~PVF_EXEC;
4488 			PMAPCOUNT(exec_discarded_modfixup);
4489 		}
4490 #endif
4491 		pmap_release_page_lock(md);
4492 
4493 		/*
4494 		 * Re-enable write permissions for the page.  No need to call
4495 		 * pmap_vac_me_harder(), since this is just a
4496 		 * modified-emulation fault, and the PVF_WRITE bit isn't
4497 		 * changing. We've already set the cacheable bits based on
4498 		 * the assumption that we can write to this page.
4499 		 */
4500 		const pt_entry_t npte =
4501 		    l2pte_set_writable((opte & ~L2_TYPE_MASK) | L2_S_PROTO)
4502 #ifdef ARM_MMU_EXTENDED
4503 		    | (pm != pmap_kernel() ? L2_XS_nG : 0)
4504 #endif
4505 		    | 0;
4506 		l2pte_reset(ptep);
4507 		PTE_SYNC(ptep);
4508 		pmap_tlb_flush_SE(pm, va,
4509 		    (ftype & VM_PROT_EXECUTE) ? PVF_EXEC | PVF_REF : PVF_REF);
4510 		l2pte_set(ptep, npte, 0);
4511 		PTE_SYNC(ptep);
4512 		PMAPCOUNT(fixup_mod);
4513 		rv = 1;
4514 		UVMHIST_LOG(maphist, " <-- done (mod/ref emul: changed pte "
4515 		    "from %#jx to %#jx)", opte, npte, 0, 0);
4516 	} else if ((opte & L2_TYPE_MASK) == L2_TYPE_INV) {
4517 		/*
4518 		 * This looks like a good candidate for "page referenced"
4519 		 * emulation.
4520 		 */
4521 		struct vm_page *pg;
4522 
4523 		/* Extract the physical address of the page */
4524 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) {
4525 			UVMHIST_LOG(maphist, " <-- done (ref emul: unmanaged page)", 0, 0, 0, 0);
4526 			goto out;
4527 		}
4528 
4529 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4530 
4531 		/* Get the current flags for this page. */
4532 		pmap_acquire_page_lock(md);
4533 		struct pv_entry *pv = pmap_find_pv(md, pm, va);
4534 		if (pv == NULL || PV_IS_KENTRY_P(pv->pv_flags)) {
4535 			pmap_release_page_lock(md);
4536 			UVMHIST_LOG(maphist, " <-- done (ref emul no PV)", 0, 0, 0, 0);
4537 			goto out;
4538 		}
4539 
4540 		md->pvh_attrs |= PVF_REF;
4541 		pv->pv_flags |= PVF_REF;
4542 
4543 		pt_entry_t npte =
4544 		    l2pte_set_readonly((opte & ~L2_TYPE_MASK) | L2_S_PROTO);
4545 #ifdef ARM_MMU_EXTENDED
4546 		if (pm != pmap_kernel()) {
4547 			npte |= L2_XS_nG;
4548 		}
4549 		/*
4550 		 * If we got called from prefetch abort, then ftype will have
4551 		 * VM_PROT_EXECUTE set.  Now see if we have no-execute set in
4552 		 * the PTE.
4553 		 */
4554 		if (user && (ftype & VM_PROT_EXECUTE) && (npte & L2_XS_XN)) {
4555 			/*
4556 			 * Is this a mapping of an executable page?
4557 			 */
4558 			if ((pv->pv_flags & PVF_EXEC) == 0) {
4559 				pmap_release_page_lock(md);
4560 				UVMHIST_LOG(maphist, " <-- done (ref emul: no exec)",
4561 				    0, 0, 0, 0);
4562 				goto out;
4563 			}
4564 			/*
4565 			 * If we haven't synced the page, do so now.
4566 			 */
4567 			if ((md->pvh_attrs & PVF_EXEC) == 0) {
4568 				UVMHIST_LOG(maphist, " ref emul: syncicache "
4569 				    "page #%#jx", pa, 0, 0, 0);
4570 				pmap_syncicache_page(md, pa);
4571 				PMAPCOUNT(fixup_exec);
4572 			}
4573 			npte &= ~L2_XS_XN;
4574 		}
4575 #endif /* ARM_MMU_EXTENDED */
4576 		pmap_release_page_lock(md);
4577 		l2pte_reset(ptep);
4578 		PTE_SYNC(ptep);
4579 		pmap_tlb_flush_SE(pm, va,
4580 		    (ftype & VM_PROT_EXECUTE) ? PVF_EXEC | PVF_REF : PVF_REF);
4581 		l2pte_set(ptep, npte, 0);
4582 		PTE_SYNC(ptep);
4583 		PMAPCOUNT(fixup_ref);
4584 		rv = 1;
4585 		UVMHIST_LOG(maphist, " <-- done (ref emul: changed pte from "
4586 		    "%#jx to %#jx)", opte, npte, 0, 0);
4587 #ifdef ARM_MMU_EXTENDED
4588 	} else if (user && (ftype & VM_PROT_EXECUTE) && (opte & L2_XS_XN)) {
4589 		struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
4590 		if (pg == NULL) {
4591 			UVMHIST_LOG(maphist, " <-- done (unmanaged page)", 0, 0, 0, 0);
4592 			goto out;
4593 		}
4594 
4595 		struct vm_page_md * const md = VM_PAGE_TO_MD(pg);
4596 
4597 		/* Get the current flags for this page. */
4598 		pmap_acquire_page_lock(md);
4599 		struct pv_entry * const pv = pmap_find_pv(md, pm, va);
4600 		if (pv == NULL || (pv->pv_flags & PVF_EXEC) == 0) {
4601 			pmap_release_page_lock(md);
4602 			UVMHIST_LOG(maphist, " <-- done (no PV or not EXEC)", 0, 0, 0, 0);
4603 			goto out;
4604 		}
4605 
4606 		/*
4607 		 * If we haven't synced the page, do so now.
4608 		 */
4609 		if ((md->pvh_attrs & PVF_EXEC) == 0) {
4610 			UVMHIST_LOG(maphist, "syncicache page #%#jx",
4611 			    pa, 0, 0, 0);
4612 			pmap_syncicache_page(md, pa);
4613 		}
4614 		pmap_release_page_lock(md);
4615 		/*
4616 		 * Turn off no-execute.
4617 		 */
4618 		KASSERT(opte & L2_XS_nG);
4619 		l2pte_reset(ptep);
4620 		PTE_SYNC(ptep);
4621 		pmap_tlb_flush_SE(pm, va, PVF_EXEC | PVF_REF);
4622 		l2pte_set(ptep, opte & ~L2_XS_XN, 0);
4623 		PTE_SYNC(ptep);
4624 		rv = 1;
4625 		PMAPCOUNT(fixup_exec);
4626 		UVMHIST_LOG(maphist, "exec: changed pte from %#jx to %#jx",
4627 		    opte, opte & ~L2_XS_XN, 0, 0);
4628 #endif
4629 	}
4630 
4631 #ifndef ARM_MMU_EXTENDED
4632 	/*
4633 	 * We know there is a valid mapping here, so simply
4634 	 * fix up the L1 if necessary.
4635 	 */
4636 	pd_entry_t * const pdep = pmap_l1_kva(pm) + l1slot;
4637 	pd_entry_t pde = L1_C_PROTO | l2b->l2b_pa | L1_C_DOM(pmap_domain(pm));
4638 	if (*pdep != pde) {
4639 		l1pte_setone(pdep, pde);
4640 		PDE_SYNC(pdep);
4641 		rv = 1;
4642 		PMAPCOUNT(fixup_pdes);
4643 	}
4644 #endif
4645 
4646 #ifdef CPU_SA110
4647 	/*
4648 	 * There are bugs in the rev K SA110.  This is a check for one
4649 	 * of them.
4650 	 */
4651 	if (rv == 0 && curcpu()->ci_arm_cputype == CPU_ID_SA110 &&
4652 	    curcpu()->ci_arm_cpurev < 3) {
4653 		/* Always current pmap */
4654 		if (l2pte_valid_p(opte)) {
4655 			extern int kernel_debug;
4656 			if (kernel_debug & 1) {
4657 				struct proc *p = curlwp->l_proc;
4658 				printf("prefetch_abort: page is already "
4659 				    "mapped - pte=%p *pte=%08x\n", ptep, opte);
4660 				printf("prefetch_abort: pc=%08lx proc=%p "
4661 				    "process=%s\n", va, p, p->p_comm);
4662 				printf("prefetch_abort: far=%08x fs=%x\n",
4663 				    cpu_faultaddress(), cpu_faultstatus());
4664 			}
4665 #ifdef DDB
4666 			if (kernel_debug & 2)
4667 				Debugger();
4668 #endif
4669 			rv = 1;
4670 		}
4671 	}
4672 #endif /* CPU_SA110 */
4673 
4674 #ifndef ARM_MMU_EXTENDED
4675 	/*
4676 	 * If 'rv == 0' at this point, it generally indicates that there is a
4677 	 * stale TLB entry for the faulting address.  That might be due to a
4678 	 * wrong setting of pmap_needs_pte_sync.  So set it and retry.
4679 	 */
4680 	if (rv == 0
4681 	    && pm->pm_l1->l1_domain_use_count == 1
4682 	    && pmap_needs_pte_sync == 0) {
4683 		pmap_needs_pte_sync = 1;
4684 		PTE_SYNC(ptep);
4685 		PMAPCOUNT(fixup_ptesync);
4686 		rv = 1;
4687 	}
4688 #endif
4689 
4690 #ifndef MULTIPROCESSOR
4691 #if defined(DEBUG) || 1
4692 	/*
4693 	 * If 'rv == 0' at this point, it generally indicates that there is a
4694 	 * stale TLB entry for the faulting address. This happens when two or
4695 	 * more processes are sharing an L1. Since we don't flush the TLB on
4696 	 * a context switch between such processes, we can take domain faults
4697 	 * for mappings which exist at the same VA in both processes. EVEN IF
4698 	 * WE'VE RECENTLY FIXED UP THE CORRESPONDING L1 in pmap_enter(), for
4699 	 * example.
4700 	 *
4701 	 * This is extremely likely to happen if pmap_enter() updated the L1
4702 	 * entry for a recently entered mapping. In this case, the TLB is
4703 	 * flushed for the new mapping, but there may still be TLB entries for
4704 	 * other mappings belonging to other processes in the 1MB range
4705 	 * covered by the L1 entry.
4706 	 *
4707 	 * Since 'rv == 0', we know that the L1 already contains the correct
4708 	 * value, so the fault must be due to a stale TLB entry.
4709 	 *
4710 	 * Since we always need to flush the TLB anyway in the case where we
4711 	 * fixed up the L1, or frobbed the L2 PTE, we effectively deal with
4712 	 * stale TLB entries dynamically.
4713 	 *
4714 	 * However, the above condition can ONLY happen if the current L1 is
4715 	 * being shared. If it happens when the L1 is unshared, it indicates
4716 	 * that other parts of the pmap are not doing their job WRT managing
4717 	 * the TLB.
4718 	 */
4719 	if (rv == 0
4720 #ifndef ARM_MMU_EXTENDED
4721 	    && pm->pm_l1->l1_domain_use_count == 1
4722 #endif
4723 	    && true) {
4724 #ifdef DEBUG
4725 		extern int last_fault_code;
4726 #else
4727 		int last_fault_code = ftype & VM_PROT_EXECUTE
4728 		    ? armreg_ifsr_read()
4729 		    : armreg_dfsr_read();
4730 #endif
4731 		printf("fixup: pm %p, va 0x%lx, ftype %d - nothing to do!\n",
4732 		    pm, va, ftype);
4733 		printf("fixup: l2 %p, l2b %p, ptep %p, pte %#x\n",
4734 		    l2, l2b, ptep, opte);
4735 
4736 #ifndef ARM_MMU_EXTENDED
4737 		printf("fixup: pdep %p, pde %#x, fsr %#x\n",
4738 		    pdep, pde, last_fault_code);
4739 #else
4740 		printf("fixup: pdep %p, pde %#x, ttbcr %#x\n",
4741 		    &pmap_l1_kva(pm)[l1slot], pmap_l1_kva(pm)[l1slot],
4742 		   armreg_ttbcr_read());
4743 		printf("fixup: fsr %#x cpm %p casid %#x contextidr %#x dacr %#x\n",
4744 		    last_fault_code, curcpu()->ci_pmap_cur,
4745 		    curcpu()->ci_pmap_asid_cur,
4746 		    armreg_contextidr_read(), armreg_dacr_read());
4747 #ifdef _ARM_ARCH_7
4748 		if (ftype & VM_PROT_WRITE)
4749 			armreg_ats1cuw_write(va);
4750 		else
4751 			armreg_ats1cur_write(va);
4752 		arm_isb();
4753 		printf("fixup: par %#x\n", armreg_par_read());
4754 #endif
4755 #endif
4756 #ifdef DDB
4757 		extern int kernel_debug;
4758 
4759 		if (kernel_debug & 2) {
4760 			pmap_release_pmap_lock(pm);
4761 #ifdef UVMHIST
4762 			KERNHIST_DUMP(maphist);
4763 #endif
4764 			cpu_Debugger();
4765 			pmap_acquire_pmap_lock(pm);
4766 		}
4767 #endif
4768 	}
4769 #endif
4770 #endif
4771 
4772 #ifndef ARM_MMU_EXTENDED
4773 	/* Flush the TLB in the shared L1 case - see comment above */
4774 	pmap_tlb_flush_SE(pm, va,
4775 	    (ftype & VM_PROT_EXECUTE) ? PVF_EXEC | PVF_REF : PVF_REF);
4776 #endif
4777 
4778 	rv = 1;
4779 
4780 out:
4781 	pmap_release_pmap_lock(pm);
4782 
4783 	return rv;
4784 }
4785 
4786 /*
4787  * Routine:	pmap_procwr
4788  *
4789  * Function:
4790  *	Synchronize caches corresponding to [addr, addr+len) in p.
4791  *
4792  */
4793 void
4794 pmap_procwr(struct proc *p, vaddr_t va, int len)
4795 {
4796 #ifndef ARM_MMU_EXTENDED
4797 
4798 	/* We only need to do anything if it is the current process. */
4799 	if (p == curproc)
4800 		cpu_icache_sync_range(va, len);
4801 #endif
4802 }
4803 
4804 /*
4805  * Routine:	pmap_unwire
4806  * Function:	Clear the wired attribute for a map/virtual-address pair.
4807  *
4808  * In/out conditions:
4809  *		The mapping must already exist in the pmap.
4810  */
4811 void
4812 pmap_unwire(pmap_t pm, vaddr_t va)
4813 {
4814 	struct l2_bucket *l2b;
4815 	pt_entry_t *ptep, pte;
4816 	struct vm_page *pg;
4817 	paddr_t pa;
4818 
4819 	UVMHIST_FUNC(__func__);
4820 	UVMHIST_CALLARGS(maphist, "pm %#jx va %#jx", (uintptr_t)pm, va, 0, 0);
4821 
4822 	pmap_acquire_pmap_lock(pm);
4823 
4824 	l2b = pmap_get_l2_bucket(pm, va);
4825 	KDASSERT(l2b != NULL);
4826 
4827 	ptep = &l2b->l2b_kva[l2pte_index(va)];
4828 	pte = *ptep;
4829 
4830 	/* Extract the physical address of the page */
4831 	pa = l2pte_pa(pte);
4832 
4833 	if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
4834 		/* Update the wired bit in the pv entry for this page. */
4835 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4836 
4837 		pmap_acquire_page_lock(md);
4838 		(void) pmap_modify_pv(md, pa, pm, va, PVF_WIRED, 0);
4839 		pmap_release_page_lock(md);
4840 	}
4841 
4842 	pmap_release_pmap_lock(pm);
4843 }
4844 
4845 #ifdef ARM_MMU_EXTENDED
4846 void
4847 pmap_md_pdetab_activate(pmap_t pm, struct lwp *l)
4848 {
4849 	UVMHIST_FUNC(__func__);
4850 	struct cpu_info * const ci = curcpu();
4851 	struct pmap_asid_info * const pai = PMAP_PAI(pm, cpu_tlb_info(ci));
4852 
4853 	UVMHIST_CALLARGS(maphist, "pm %#jx (pm->pm_l1_pa %08jx asid %ju)",
4854 	    (uintptr_t)pm, pm->pm_l1_pa, pai->pai_asid, 0);
4855 
4856 	/*
4857 	 * Assume that TTBR1 has only global mappings and TTBR0 only
4858 	 * has non-global mappings.  To prevent speculation from doing
4859 	 * evil things we disable translation table walks using TTBR0
4860 	 * before setting the CONTEXTIDR (ASID) or new TTBR0 value.
4861 	 * Once both are set, table walks are reenabled.
4862 	 */
4863 	const uint32_t old_ttbcr = armreg_ttbcr_read();
4864 	armreg_ttbcr_write(old_ttbcr | TTBCR_S_PD0);
4865 	arm_isb();
4866 
4867 	pmap_tlb_asid_acquire(pm, l);
4868 
4869 	cpu_setttb(pm->pm_l1_pa, pai->pai_asid);
4870 	/*
4871 	 * Now we can reenable tablewalks since the CONTEXTIDR and TTRB0
4872 	 * have been updated.
4873 	 */
4874 	arm_isb();
4875 
4876 	if (pm != pmap_kernel()) {
4877 		armreg_ttbcr_write(old_ttbcr & ~TTBCR_S_PD0);
4878 	}
4879 	cpu_cpwait();
4880 
4881 	KASSERTMSG(ci->ci_pmap_asid_cur == pai->pai_asid, "%u vs %u",
4882 	    ci->ci_pmap_asid_cur, pai->pai_asid);
4883 	ci->ci_pmap_cur = pm;
4884 }
4885 
4886 void
4887 pmap_md_pdetab_deactivate(pmap_t pm)
4888 {
4889 
4890 	UVMHIST_FUNC(__func__);
4891 	UVMHIST_CALLARGS(maphist, "pm %#jx", (uintptr_t)pm, 0, 0, 0);
4892 
4893 	kpreempt_disable();
4894 	struct cpu_info * const ci = curcpu();
4895 	/*
4896 	 * Disable translation table walks from TTBR0 while no pmap has been
4897 	 * activated.
4898 	 */
4899 	const uint32_t old_ttbcr = armreg_ttbcr_read();
4900 	armreg_ttbcr_write(old_ttbcr | TTBCR_S_PD0);
4901 	arm_isb();
4902 	pmap_tlb_asid_deactivate(pm);
4903 	cpu_setttb(pmap_kernel()->pm_l1_pa, KERNEL_PID);
4904 	arm_isb();
4905 
4906 	ci->ci_pmap_cur = pmap_kernel();
4907 	KASSERTMSG(ci->ci_pmap_asid_cur == KERNEL_PID, "ci_pmap_asid_cur %u",
4908 	    ci->ci_pmap_asid_cur);
4909 	kpreempt_enable();
4910 }
4911 #endif
4912 
4913 void
4914 pmap_activate(struct lwp *l)
4915 {
4916 	extern int block_userspace_access;
4917 	pmap_t npm = l->l_proc->p_vmspace->vm_map.pmap;
4918 
4919 	UVMHIST_FUNC(__func__);
4920 	UVMHIST_CALLARGS(maphist, "l=%#jx pm=%#jx", (uintptr_t)l,
4921 	    (uintptr_t)npm, 0, 0);
4922 
4923 	struct cpu_info * const ci = curcpu();
4924 
4925 	/*
4926 	 * If activating a non-current lwp or the current lwp is
4927 	 * already active, just return.
4928 	 */
4929 	if (false
4930 	    || l != curlwp
4931 #ifdef ARM_MMU_EXTENDED
4932 	    || (ci->ci_pmap_cur == npm &&
4933 		(npm == pmap_kernel()
4934 		 /* || PMAP_PAI_ASIDVALID_P(pai, cpu_tlb_info(ci)) */))
4935 #else
4936 	    || npm->pm_activated == true
4937 #endif
4938 	    || false) {
4939 		UVMHIST_LOG(maphist, " <-- (same pmap)", (uintptr_t)curlwp,
4940 		    (uintptr_t)l, 0, 0);
4941 		return;
4942 	}
4943 
4944 #ifndef ARM_MMU_EXTENDED
4945 	const uint32_t ndacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2))
4946 	    | (DOMAIN_CLIENT << (pmap_domain(npm) * 2));
4947 
4948 	/*
4949 	 * If TTB and DACR are unchanged, short-circuit all the
4950 	 * TLB/cache management stuff.
4951 	 */
4952 	pmap_t opm = ci->ci_lastlwp
4953 	    ? ci->ci_lastlwp->l_proc->p_vmspace->vm_map.pmap
4954 	    : NULL;
4955 	if (opm != NULL) {
4956 		uint32_t odacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2))
4957 		    | (DOMAIN_CLIENT << (pmap_domain(opm) * 2));
4958 
4959 		if (opm->pm_l1 == npm->pm_l1 && odacr == ndacr)
4960 			goto all_done;
4961 	}
4962 #endif /* !ARM_MMU_EXTENDED */
4963 
4964 	PMAPCOUNT(activations);
4965 	block_userspace_access = 1;
4966 
4967 #ifndef ARM_MMU_EXTENDED
4968 	/*
4969 	 * If switching to a user vmspace which is different to the
4970 	 * most recent one, and the most recent one is potentially
4971 	 * live in the cache, we must write-back and invalidate the
4972 	 * entire cache.
4973 	 */
4974 	pmap_t rpm = ci->ci_pmap_lastuser;
4975 
4976 	/*
4977 	 * XXXSCW: There's a corner case here which can leave turds in the
4978 	 * cache as reported in kern/41058. They're probably left over during
4979 	 * tear-down and switching away from an exiting process. Until the root
4980 	 * cause is identified and fixed, zap the cache when switching pmaps.
4981 	 * This will result in a few unnecessary cache flushes, but that's
4982 	 * better than silently corrupting data.
4983 	 */
4984 #if 0
4985 	if (npm != pmap_kernel() && rpm && npm != rpm &&
4986 	    rpm->pm_cstate.cs_cache) {
4987 		rpm->pm_cstate.cs_cache = 0;
4988 #ifdef PMAP_CACHE_VIVT
4989 		cpu_idcache_wbinv_all();
4990 #endif
4991 	}
4992 #else
4993 	if (rpm) {
4994 		rpm->pm_cstate.cs_cache = 0;
4995 		if (npm == pmap_kernel())
4996 			ci->ci_pmap_lastuser = NULL;
4997 #ifdef PMAP_CACHE_VIVT
4998 		cpu_idcache_wbinv_all();
4999 #endif
5000 	}
5001 #endif
5002 
5003 	/* No interrupts while we frob the TTB/DACR */
5004 	uint32_t oldirqstate = disable_interrupts(IF32_bits);
5005 #endif /* !ARM_MMU_EXTENDED */
5006 
5007 #ifndef ARM_HAS_VBAR
5008 	/*
5009 	 * For ARM_VECTORS_LOW, we MUST, I repeat, MUST fix up the L1
5010 	 * entry corresponding to 'vector_page' in the incoming L1 table
5011 	 * before switching to it otherwise subsequent interrupts/exceptions
5012 	 * (including domain faults!) will jump into hyperspace.
5013 	 */
5014 	if (npm->pm_pl1vec != NULL) {
5015 		cpu_tlb_flushID_SE((u_int)vector_page);
5016 		cpu_cpwait();
5017 		*npm->pm_pl1vec = npm->pm_l1vec;
5018 		PTE_SYNC(npm->pm_pl1vec);
5019 	}
5020 #endif
5021 
5022 #ifdef ARM_MMU_EXTENDED
5023 	pmap_md_pdetab_activate(npm, l);
5024 #else
5025 	cpu_domains(ndacr);
5026 	if (npm == pmap_kernel() || npm == rpm) {
5027 		/*
5028 		 * Switching to a kernel thread, or back to the
5029 		 * same user vmspace as before... Simply update
5030 		 * the TTB (no TLB flush required)
5031 		 */
5032 		cpu_setttb(npm->pm_l1->l1_physaddr, false);
5033 		cpu_cpwait();
5034 	} else {
5035 		/*
5036 		 * Otherwise, update TTB and flush TLB
5037 		 */
5038 		cpu_context_switch(npm->pm_l1->l1_physaddr);
5039 		if (rpm != NULL)
5040 			rpm->pm_cstate.cs_tlb = 0;
5041 	}
5042 
5043 	restore_interrupts(oldirqstate);
5044 #endif /* ARM_MMU_EXTENDED */
5045 
5046 	block_userspace_access = 0;
5047 
5048 #ifndef ARM_MMU_EXTENDED
5049  all_done:
5050 	/*
5051 	 * The new pmap is resident. Make sure it's marked
5052 	 * as resident in the cache/TLB.
5053 	 */
5054 	npm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
5055 	if (npm != pmap_kernel())
5056 		ci->ci_pmap_lastuser = npm;
5057 
5058 	/* The old pmap is not longer active */
5059 	if (opm != npm) {
5060 		if (opm != NULL)
5061 			opm->pm_activated = false;
5062 
5063 		/* But the new one is */
5064 		npm->pm_activated = true;
5065 	}
5066 	ci->ci_pmap_cur = npm;
5067 #endif
5068 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
5069 }
5070 
5071 void
5072 pmap_deactivate(struct lwp *l)
5073 {
5074 	pmap_t pm = l->l_proc->p_vmspace->vm_map.pmap;
5075 
5076 	UVMHIST_FUNC(__func__);
5077 	UVMHIST_CALLARGS(maphist, "l=%#jx (pm=%#jx)", (uintptr_t)l,
5078 		(uintptr_t)pm, 0, 0);
5079 
5080 #ifdef ARM_MMU_EXTENDED
5081 	pmap_md_pdetab_deactivate(pm);
5082 #else
5083 	/*
5084 	 * If the process is exiting, make sure pmap_activate() does
5085 	 * a full MMU context-switch and cache flush, which we might
5086 	 * otherwise skip. See PR port-arm/38950.
5087 	 */
5088 	if (l->l_proc->p_sflag & PS_WEXIT)
5089 		curcpu()->ci_lastlwp = NULL;
5090 
5091 	pm->pm_activated = false;
5092 #endif
5093 	UVMHIST_LOG(maphist, "  <-- done", 0, 0, 0, 0);
5094 }
5095 
5096 void
5097 pmap_update(pmap_t pm)
5098 {
5099 
5100 	UVMHIST_FUNC(__func__);
5101 	UVMHIST_CALLARGS(maphist, "pm=%#jx remove_all %jd", (uintptr_t)pm,
5102 	    pm->pm_remove_all, 0, 0);
5103 
5104 #ifndef ARM_MMU_EXTENDED
5105 	if (pm->pm_remove_all) {
5106 		/*
5107 		 * Finish up the pmap_remove_all() optimisation by flushing
5108 		 * the TLB.
5109 		 */
5110 		pmap_tlb_flushID(pm);
5111 		pm->pm_remove_all = false;
5112 	}
5113 
5114 	if (pmap_is_current(pm)) {
5115 		/*
5116 		 * If we're dealing with a current userland pmap, move its L1
5117 		 * to the end of the LRU.
5118 		 */
5119 		if (pm != pmap_kernel())
5120 			pmap_use_l1(pm);
5121 
5122 		/*
5123 		 * We can assume we're done with frobbing the cache/tlb for
5124 		 * now. Make sure any future pmap ops don't skip cache/tlb
5125 		 * flushes.
5126 		 */
5127 		pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
5128 	}
5129 #else
5130 
5131 	kpreempt_disable();
5132 #if defined(MULTIPROCESSOR) && PMAP_TLB_MAX > 1
5133 	u_int pending = atomic_swap_uint(&pmap->pm_shootdown_pending, 0);
5134 	if (pending && pmap_tlb_shootdown_bystanders(pmap)) {
5135 		PMAP_COUNT(shootdown_ipis);
5136 	}
5137 #endif
5138 
5139 	/*
5140 	 * If pmap_remove_all was called, we deactivated ourselves and released
5141 	 * our ASID.  Now we have to reactivate ourselves.
5142 	 */
5143 	if (__predict_false(pm->pm_remove_all)) {
5144 		pm->pm_remove_all = false;
5145 
5146 		KASSERT(pm != pmap_kernel());
5147 		pmap_md_pdetab_activate(pm, curlwp);
5148 	}
5149 
5150 	if (arm_has_mpext_p)
5151 		armreg_bpiallis_write(0);
5152 	else
5153 		armreg_bpiall_write(0);
5154 
5155 	kpreempt_enable();
5156 
5157 	KASSERTMSG(pm == pmap_kernel()
5158 	    || curcpu()->ci_pmap_cur != pm
5159 	    || pm->pm_pai[0].pai_asid == curcpu()->ci_pmap_asid_cur,
5160 	    "pmap/asid %p/%#x != %s cur pmap/asid %p/%#x", pm,
5161 	    pm->pm_pai[0].pai_asid, curcpu()->ci_data.cpu_name,
5162 	    curcpu()->ci_pmap_cur, curcpu()->ci_pmap_asid_cur);
5163 #endif
5164 
5165 	PMAPCOUNT(updates);
5166 
5167 	/*
5168 	 * make sure TLB/cache operations have completed.
5169 	 */
5170 	cpu_cpwait();
5171 	UVMHIST_LOG(maphist, "  <-- done", 0, 0, 0, 0);
5172 }
5173 
5174 bool
5175 pmap_remove_all(pmap_t pm)
5176 {
5177 
5178 	/*
5179 	 * The vmspace described by this pmap is about to be torn down.
5180 	 * Until pmap_update() is called, UVM will only make calls
5181 	 * to pmap_remove(). We can make life much simpler by flushing
5182 	 * the cache now, and deferring TLB invalidation to pmap_update().
5183 	 */
5184 #ifdef PMAP_CACHE_VIVT
5185 	pmap_cache_wbinv_all(pm, PVF_EXEC);
5186 #endif
5187 #ifdef ARM_MMU_EXTENDED
5188 #ifdef MULTIPROCESSOR
5189 	struct cpu_info * const ci = curcpu();
5190 	// This should be the last CPU with this pmap onproc
5191 	KASSERT(!kcpuset_isotherset(pm->pm_onproc, cpu_index(ci)));
5192 	if (kcpuset_isset(pm->pm_onproc, cpu_index(ci)))
5193 #endif
5194 		pmap_tlb_asid_deactivate(pm);
5195 #ifdef MULTIPROCESSOR
5196 	KASSERT(kcpuset_iszero(pm->pm_onproc));
5197 #endif
5198 
5199 	pmap_tlb_asid_release_all(pm);
5200 #endif
5201 	pm->pm_remove_all = true;
5202 	return false;
5203 }
5204 
5205 /*
5206  * Retire the given physical map from service.
5207  * Should only be called if the map contains no valid mappings.
5208  */
5209 void
5210 pmap_destroy(pmap_t pm)
5211 {
5212 	UVMHIST_FUNC(__func__);
5213 	UVMHIST_CALLARGS(maphist, "pm=%#jx remove_all %jd", (uintptr_t)pm,
5214 	    pm ? pm->pm_remove_all : 0, 0, 0);
5215 
5216 	if (pm == NULL)
5217 		return;
5218 
5219 	if (pm->pm_remove_all) {
5220 #ifdef ARM_MMU_EXTENDED
5221  		pmap_tlb_asid_release_all(pm);
5222 #else
5223 		pmap_tlb_flushID(pm);
5224 #endif
5225 		pm->pm_remove_all = false;
5226 	}
5227 
5228 	/*
5229 	 * Drop reference count
5230 	 */
5231 	if (atomic_dec_uint_nv(&pm->pm_refs) > 0) {
5232 #ifndef ARM_MMU_EXTENDED
5233 		if (pmap_is_current(pm)) {
5234 			if (pm != pmap_kernel())
5235 				pmap_use_l1(pm);
5236 			pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
5237 		}
5238 #endif
5239 		return;
5240 	}
5241 
5242 	/*
5243 	 * reference count is zero, free pmap resources and then free pmap.
5244 	 */
5245 
5246 #ifndef ARM_HAS_VBAR
5247 	if (vector_page < KERNEL_BASE) {
5248 		KDASSERT(!pmap_is_current(pm));
5249 
5250 		/* Remove the vector page mapping */
5251 		pmap_remove(pm, vector_page, vector_page + PAGE_SIZE);
5252 		pmap_update(pm);
5253 	}
5254 #endif
5255 
5256 	pmap_free_l1(pm);
5257 
5258 #ifdef ARM_MMU_EXTENDED
5259 #ifdef MULTIPROCESSOR
5260 	kcpuset_destroy(pm->pm_active);
5261 	kcpuset_destroy(pm->pm_onproc);
5262 #endif
5263 #else
5264 	struct cpu_info * const ci = curcpu();
5265 	if (ci->ci_pmap_lastuser == pm)
5266 		ci->ci_pmap_lastuser = NULL;
5267 #endif
5268 
5269 	mutex_destroy(&pm->pm_lock);
5270 	pool_cache_put(&pmap_cache, pm);
5271 	UVMHIST_LOG(maphist, "  <-- done", 0, 0, 0, 0);
5272 }
5273 
5274 
5275 /*
5276  * void pmap_reference(pmap_t pm)
5277  *
5278  * Add a reference to the specified pmap.
5279  */
5280 void
5281 pmap_reference(pmap_t pm)
5282 {
5283 
5284 	if (pm == NULL)
5285 		return;
5286 
5287 #ifndef ARM_MMU_EXTENDED
5288 	pmap_use_l1(pm);
5289 #endif
5290 
5291 	atomic_inc_uint(&pm->pm_refs);
5292 }
5293 
5294 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
5295 
5296 static struct evcnt pmap_prefer_nochange_ev =
5297     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap prefer", "nochange");
5298 static struct evcnt pmap_prefer_change_ev =
5299     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap prefer", "change");
5300 
5301 EVCNT_ATTACH_STATIC(pmap_prefer_change_ev);
5302 EVCNT_ATTACH_STATIC(pmap_prefer_nochange_ev);
5303 
5304 void
5305 pmap_prefer(vaddr_t hint, vaddr_t *vap, int td)
5306 {
5307 	vsize_t mask = arm_cache_prefer_mask | (PAGE_SIZE - 1);
5308 	vaddr_t va = *vap;
5309 	vaddr_t diff = (hint - va) & mask;
5310 	if (diff == 0) {
5311 		pmap_prefer_nochange_ev.ev_count++;
5312 	} else {
5313 		pmap_prefer_change_ev.ev_count++;
5314 		if (__predict_false(td))
5315 			va -= mask + 1;
5316 		*vap = va + diff;
5317 	}
5318 }
5319 #endif /* ARM_MMU_V6 | ARM_MMU_V7 */
5320 
5321 /*
5322  * pmap_zero_page()
5323  *
5324  * Zero a given physical page by mapping it at a page hook point.
5325  * In doing the zero page op, the page we zero is mapped cachable, as with
5326  * StrongARM accesses to non-cached pages are non-burst making writing
5327  * _any_ bulk data very slow.
5328  */
5329 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
5330 void
5331 pmap_zero_page_generic(paddr_t pa)
5332 {
5333 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
5334 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
5335 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
5336 #endif
5337 #if defined(PMAP_CACHE_VIPT)
5338 	/* Choose the last page color it had, if any */
5339 	const vsize_t va_offset = md->pvh_attrs & arm_cache_prefer_mask;
5340 #else
5341 	const vsize_t va_offset = 0;
5342 #endif
5343 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
5344 	/*
5345 	 * Is this page mapped at its natural color?
5346 	 * If we have all of memory mapped, then just convert PA to VA.
5347 	 */
5348 	bool okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5349 	   || va_offset == (pa & arm_cache_prefer_mask);
5350 	const vaddr_t vdstp = okcolor
5351 	    ? pmap_direct_mapped_phys(pa, &okcolor, cpu_cdstp(va_offset))
5352 	    : cpu_cdstp(va_offset);
5353 #else
5354 	const bool okcolor = false;
5355 	const vaddr_t vdstp = cpu_cdstp(va_offset);
5356 #endif
5357 	pt_entry_t * const ptep = cpu_cdst_pte(va_offset);
5358 
5359 
5360 #ifdef DEBUG
5361 	if (!SLIST_EMPTY(&md->pvh_list))
5362 		panic("pmap_zero_page: page has mappings");
5363 #endif
5364 
5365 	KDASSERT((pa & PGOFSET) == 0);
5366 
5367 	if (!okcolor) {
5368 		/*
5369 		 * Hook in the page, zero it, and purge the cache for that
5370 		 * zeroed page. Invalidate the TLB as needed.
5371 		 */
5372 		const pt_entry_t npte = L2_S_PROTO | pa | pte_l2_s_cache_mode
5373 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE);
5374 		l2pte_set(ptep, npte, 0);
5375 		PTE_SYNC(ptep);
5376 		cpu_tlb_flushD_SE(vdstp);
5377 		cpu_cpwait();
5378 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS) && defined(PMAP_CACHE_VIPT) \
5379     && !defined(ARM_MMU_EXTENDED)
5380 		/*
5381 		 * If we are direct-mapped and our color isn't ok, then before
5382 		 * we bzero the page invalidate its contents from the cache and
5383 		 * reset the color to its natural color.
5384 		 */
5385 		cpu_dcache_inv_range(vdstp, PAGE_SIZE);
5386 		md->pvh_attrs &= ~arm_cache_prefer_mask;
5387 		md->pvh_attrs |= (pa & arm_cache_prefer_mask);
5388 #endif
5389 	}
5390 	bzero_page(vdstp);
5391 	if (!okcolor) {
5392 		/*
5393 		 * Unmap the page.
5394 		 */
5395 		l2pte_reset(ptep);
5396 		PTE_SYNC(ptep);
5397 		cpu_tlb_flushD_SE(vdstp);
5398 #ifdef PMAP_CACHE_VIVT
5399 		cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
5400 #endif
5401 	}
5402 #ifdef PMAP_CACHE_VIPT
5403 	/*
5404 	 * This page is now cache resident so it now has a page color.
5405 	 * Any contents have been obliterated so clear the EXEC flag.
5406 	 */
5407 #ifndef ARM_MMU_EXTENDED
5408 	if (!pmap_is_page_colored_p(md)) {
5409 		PMAPCOUNT(vac_color_new);
5410 		md->pvh_attrs |= PVF_COLORED;
5411 	}
5412 	md->pvh_attrs |= PVF_DIRTY;
5413 #endif
5414 	if (PV_IS_EXEC_P(md->pvh_attrs)) {
5415 		md->pvh_attrs &= ~PVF_EXEC;
5416 		PMAPCOUNT(exec_discarded_zero);
5417 	}
5418 #endif
5419 }
5420 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
5421 
5422 #if ARM_MMU_XSCALE == 1
5423 void
5424 pmap_zero_page_xscale(paddr_t pa)
5425 {
5426 #ifdef DEBUG
5427 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
5428 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
5429 
5430 	if (!SLIST_EMPTY(&md->pvh_list))
5431 		panic("pmap_zero_page: page has mappings");
5432 #endif
5433 
5434 	KDASSERT((pa & PGOFSET) == 0);
5435 
5436 	/*
5437 	 * Hook in the page, zero it, and purge the cache for that
5438 	 * zeroed page. Invalidate the TLB as needed.
5439 	 */
5440 
5441 	pt_entry_t npte = L2_S_PROTO | pa |
5442 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
5443 	    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
5444 	l2pte_set(cdst_pte, npte, 0);
5445 	PTE_SYNC(cdst_pte);
5446 	cpu_tlb_flushD_SE(cdstp);
5447 	cpu_cpwait();
5448 	bzero_page(cdstp);
5449 	xscale_cache_clean_minidata();
5450 	l2pte_reset(cdst_pte);
5451 	PTE_SYNC(cdst_pte);
5452 }
5453 #endif /* ARM_MMU_XSCALE == 1 */
5454 
5455 /* pmap_pageidlezero()
5456  *
5457  * The same as above, except that we assume that the page is not
5458  * mapped.  This means we never have to flush the cache first.  Called
5459  * from the idle loop.
5460  */
5461 bool
5462 pmap_pageidlezero(paddr_t pa)
5463 {
5464 	bool rv = true;
5465 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
5466 	struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
5467 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
5468 #endif
5469 #ifdef PMAP_CACHE_VIPT
5470 	/* Choose the last page color it had, if any */
5471 	const vsize_t va_offset = md->pvh_attrs & arm_cache_prefer_mask;
5472 #else
5473 	const vsize_t va_offset = 0;
5474 #endif
5475 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
5476 	bool okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5477 	   || va_offset == (pa & arm_cache_prefer_mask);
5478 	const vaddr_t vdstp = okcolor
5479 	    ? pmap_direct_mapped_phys(pa, &okcolor, cpu_cdstp(va_offset))
5480 	    : cpu_cdstp(va_offset);
5481 #else
5482 	const bool okcolor = false;
5483 	const vaddr_t vdstp = cpu_cdstp(va_offset);
5484 #endif
5485 	pt_entry_t * const ptep = cpu_cdst_pte(va_offset);
5486 
5487 
5488 #ifdef DEBUG
5489 	if (!SLIST_EMPTY(&md->pvh_list))
5490 		panic("pmap_pageidlezero: page has mappings");
5491 #endif
5492 
5493 	KDASSERT((pa & PGOFSET) == 0);
5494 
5495 	if (!okcolor) {
5496 		/*
5497 		 * Hook in the page, zero it, and purge the cache for that
5498 		 * zeroed page. Invalidate the TLB as needed.
5499 		 */
5500 		const pt_entry_t npte = L2_S_PROTO | pa |
5501 		    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
5502 		l2pte_set(ptep, npte, 0);
5503 		PTE_SYNC(ptep);
5504 		cpu_tlb_flushD_SE(vdstp);
5505 		cpu_cpwait();
5506 	}
5507 
5508 	uint64_t *ptr = (uint64_t *)vdstp;
5509 	for (size_t i = 0; i < PAGE_SIZE / sizeof(*ptr); i++) {
5510 		if (sched_curcpu_runnable_p() != 0) {
5511 			/*
5512 			 * A process has become ready.  Abort now,
5513 			 * so we don't keep it waiting while we
5514 			 * do slow memory access to finish this
5515 			 * page.
5516 			 */
5517 			rv = false;
5518 			break;
5519 		}
5520 		*ptr++ = 0;
5521 	}
5522 
5523 #ifdef PMAP_CACHE_VIVT
5524 	if (rv)
5525 		/*
5526 		 * if we aborted we'll rezero this page again later so don't
5527 		 * purge it unless we finished it
5528 		 */
5529 		cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
5530 #elif defined(PMAP_CACHE_VIPT)
5531 	/*
5532 	 * This page is now cache resident so it now has a page color.
5533 	 * Any contents have been obliterated so clear the EXEC flag.
5534 	 */
5535 #ifndef ARM_MMU_EXTENDED
5536 	if (!pmap_is_page_colored_p(md)) {
5537 		PMAPCOUNT(vac_color_new);
5538 		md->pvh_attrs |= PVF_COLORED;
5539 	}
5540 #endif
5541 	if (PV_IS_EXEC_P(md->pvh_attrs)) {
5542 		md->pvh_attrs &= ~PVF_EXEC;
5543 		PMAPCOUNT(exec_discarded_zero);
5544 	}
5545 #endif
5546 	/*
5547 	 * Unmap the page.
5548 	 */
5549 	if (!okcolor) {
5550 		l2pte_reset(ptep);
5551 		PTE_SYNC(ptep);
5552 		cpu_tlb_flushD_SE(vdstp);
5553 	}
5554 
5555 	return rv;
5556 }
5557 
5558 /*
5559  * pmap_copy_page()
5560  *
5561  * Copy one physical page into another, by mapping the pages into
5562  * hook points. The same comment regarding cachability as in
5563  * pmap_zero_page also applies here.
5564  */
5565 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
5566 void
5567 pmap_copy_page_generic(paddr_t src, paddr_t dst)
5568 {
5569 	struct vm_page * const src_pg = PHYS_TO_VM_PAGE(src);
5570 	struct vm_page_md *src_md = VM_PAGE_TO_MD(src_pg);
5571 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
5572 	struct vm_page * const dst_pg = PHYS_TO_VM_PAGE(dst);
5573 	struct vm_page_md *dst_md = VM_PAGE_TO_MD(dst_pg);
5574 #endif
5575 #ifdef PMAP_CACHE_VIPT
5576 	const vsize_t src_va_offset = src_md->pvh_attrs & arm_cache_prefer_mask;
5577 	const vsize_t dst_va_offset = dst_md->pvh_attrs & arm_cache_prefer_mask;
5578 #else
5579 	const vsize_t src_va_offset = 0;
5580 	const vsize_t dst_va_offset = 0;
5581 #endif
5582 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
5583 	/*
5584 	 * Is this page mapped at its natural color?
5585 	 * If we have all of memory mapped, then just convert PA to VA.
5586 	 */
5587 	bool src_okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5588 	    || src_va_offset == (src & arm_cache_prefer_mask);
5589 	bool dst_okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5590 	    || dst_va_offset == (dst & arm_cache_prefer_mask);
5591 	const vaddr_t vsrcp = src_okcolor
5592 	    ? pmap_direct_mapped_phys(src, &src_okcolor,
5593 		cpu_csrcp(src_va_offset))
5594 	    : cpu_csrcp(src_va_offset);
5595 	const vaddr_t vdstp = pmap_direct_mapped_phys(dst, &dst_okcolor,
5596 	    cpu_cdstp(dst_va_offset));
5597 #else
5598 	const bool src_okcolor = false;
5599 	const bool dst_okcolor = false;
5600 	const vaddr_t vsrcp = cpu_csrcp(src_va_offset);
5601 	const vaddr_t vdstp = cpu_cdstp(dst_va_offset);
5602 #endif
5603 	pt_entry_t * const src_ptep = cpu_csrc_pte(src_va_offset);
5604 	pt_entry_t * const dst_ptep = cpu_cdst_pte(dst_va_offset);
5605 
5606 #ifdef DEBUG
5607 	if (!SLIST_EMPTY(&dst_md->pvh_list))
5608 		panic("pmap_copy_page: dst page has mappings");
5609 #endif
5610 
5611 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
5612 	KASSERT(arm_cache_prefer_mask == 0 || src_md->pvh_attrs & (PVF_COLORED|PVF_NC));
5613 #endif
5614 	KDASSERT((src & PGOFSET) == 0);
5615 	KDASSERT((dst & PGOFSET) == 0);
5616 
5617 	/*
5618 	 * Clean the source page.  Hold the source page's lock for
5619 	 * the duration of the copy so that no other mappings can
5620 	 * be created while we have a potentially aliased mapping.
5621 	 */
5622 #ifdef PMAP_CACHE_VIVT
5623 	pmap_acquire_page_lock(src_md);
5624 	(void) pmap_clean_page(src_md, true);
5625 	pmap_release_page_lock(src_md);
5626 #endif
5627 
5628 	/*
5629 	 * Map the pages into the page hook points, copy them, and purge
5630 	 * the cache for the appropriate page. Invalidate the TLB
5631 	 * as required.
5632 	 */
5633 	if (!src_okcolor) {
5634 		const pt_entry_t nsrc_pte = L2_S_PROTO
5635 		    | src
5636 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
5637 		    | ((src_md->pvh_attrs & PVF_NC) ? 0 : pte_l2_s_cache_mode)
5638 #else // defined(PMAP_CACHE_VIVT) || defined(ARM_MMU_EXTENDED)
5639 		    | pte_l2_s_cache_mode
5640 #endif
5641 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ);
5642 		l2pte_set(src_ptep, nsrc_pte, 0);
5643 		PTE_SYNC(src_ptep);
5644 		cpu_tlb_flushD_SE(vsrcp);
5645 		cpu_cpwait();
5646 	}
5647 	if (!dst_okcolor) {
5648 		const pt_entry_t ndst_pte = L2_S_PROTO | dst |
5649 		    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
5650 		l2pte_set(dst_ptep, ndst_pte, 0);
5651 		PTE_SYNC(dst_ptep);
5652 		cpu_tlb_flushD_SE(vdstp);
5653 		cpu_cpwait();
5654 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS) && defined(PMAP_CACHE_VIPT)
5655 		/*
5656 		 * If we are direct-mapped and our color isn't ok, then before
5657 		 * we bcopy to the new page invalidate its contents from the
5658 		 * cache and reset its color to its natural color.
5659 		 */
5660 		cpu_dcache_inv_range(vdstp, PAGE_SIZE);
5661 		dst_md->pvh_attrs &= ~arm_cache_prefer_mask;
5662 		dst_md->pvh_attrs |= (dst & arm_cache_prefer_mask);
5663 #endif
5664 	}
5665 	bcopy_page(vsrcp, vdstp);
5666 #ifdef PMAP_CACHE_VIVT
5667 	cpu_dcache_inv_range(vsrcp, PAGE_SIZE);
5668 	cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
5669 #endif
5670 	/*
5671 	 * Unmap the pages.
5672 	 */
5673 	if (!src_okcolor) {
5674 		l2pte_reset(src_ptep);
5675 		PTE_SYNC(src_ptep);
5676 		cpu_tlb_flushD_SE(vsrcp);
5677 		cpu_cpwait();
5678 	}
5679 	if (!dst_okcolor) {
5680 		l2pte_reset(dst_ptep);
5681 		PTE_SYNC(dst_ptep);
5682 		cpu_tlb_flushD_SE(vdstp);
5683 		cpu_cpwait();
5684 	}
5685 #ifdef PMAP_CACHE_VIPT
5686 	/*
5687 	 * Now that the destination page is in the cache, mark it as colored.
5688 	 * If this was an exec page, discard it.
5689 	 */
5690 	pmap_acquire_page_lock(dst_md);
5691 #ifndef ARM_MMU_EXTENDED
5692 	if (arm_pcache.cache_type == CACHE_TYPE_PIPT) {
5693 		dst_md->pvh_attrs &= ~arm_cache_prefer_mask;
5694 		dst_md->pvh_attrs |= (dst & arm_cache_prefer_mask);
5695 	}
5696 	if (!pmap_is_page_colored_p(dst_md)) {
5697 		PMAPCOUNT(vac_color_new);
5698 		dst_md->pvh_attrs |= PVF_COLORED;
5699 	}
5700 	dst_md->pvh_attrs |= PVF_DIRTY;
5701 #endif
5702 	if (PV_IS_EXEC_P(dst_md->pvh_attrs)) {
5703 		dst_md->pvh_attrs &= ~PVF_EXEC;
5704 		PMAPCOUNT(exec_discarded_copy);
5705 	}
5706 	pmap_release_page_lock(dst_md);
5707 #endif
5708 }
5709 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
5710 
5711 #if ARM_MMU_XSCALE == 1
5712 void
5713 pmap_copy_page_xscale(paddr_t src, paddr_t dst)
5714 {
5715 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
5716 	struct vm_page_md *src_md = VM_PAGE_TO_MD(src_pg);
5717 #ifdef DEBUG
5718 	struct vm_page_md *dst_md = VM_PAGE_TO_MD(PHYS_TO_VM_PAGE(dst));
5719 
5720 	if (!SLIST_EMPTY(&dst_md->pvh_list))
5721 		panic("pmap_copy_page: dst page has mappings");
5722 #endif
5723 
5724 	KDASSERT((src & PGOFSET) == 0);
5725 	KDASSERT((dst & PGOFSET) == 0);
5726 
5727 	/*
5728 	 * Clean the source page.  Hold the source page's lock for
5729 	 * the duration of the copy so that no other mappings can
5730 	 * be created while we have a potentially aliased mapping.
5731 	 */
5732 #ifdef PMAP_CACHE_VIVT
5733 	pmap_acquire_page_lock(src_md);
5734 	(void) pmap_clean_page(src_md, true);
5735 	pmap_release_page_lock(src_md);
5736 #endif
5737 
5738 	/*
5739 	 * Map the pages into the page hook points, copy them, and purge
5740 	 * the cache for the appropriate page. Invalidate the TLB
5741 	 * as required.
5742 	 */
5743 	const pt_entry_t nsrc_pte = L2_S_PROTO | src
5744 	    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ)
5745 	    | L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
5746 	l2pte_set(csrc_pte, nsrc_pte, 0);
5747 	PTE_SYNC(csrc_pte);
5748 
5749 	const pt_entry_t ndst_pte = L2_S_PROTO | dst
5750 	    | L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE)
5751 	    | L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
5752 	l2pte_set(cdst_pte, ndst_pte, 0);
5753 	PTE_SYNC(cdst_pte);
5754 
5755 	cpu_tlb_flushD_SE(csrcp);
5756 	cpu_tlb_flushD_SE(cdstp);
5757 	cpu_cpwait();
5758 	bcopy_page(csrcp, cdstp);
5759 	xscale_cache_clean_minidata();
5760 	l2pte_reset(csrc_pte);
5761 	l2pte_reset(cdst_pte);
5762 	PTE_SYNC(csrc_pte);
5763 	PTE_SYNC(cdst_pte);
5764 }
5765 #endif /* ARM_MMU_XSCALE == 1 */
5766 
5767 /*
5768  * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
5769  *
5770  * Return the start and end addresses of the kernel's virtual space.
5771  * These values are setup in pmap_bootstrap and are updated as pages
5772  * are allocated.
5773  */
5774 void
5775 pmap_virtual_space(vaddr_t *start, vaddr_t *end)
5776 {
5777 	*start = virtual_avail;
5778 	*end = virtual_end;
5779 }
5780 
5781 /*
5782  * Helper function for pmap_grow_l2_bucket()
5783  */
5784 static inline int
5785 pmap_grow_map(vaddr_t va, paddr_t *pap)
5786 {
5787 	paddr_t pa;
5788 
5789 	KASSERT((va & PGOFSET) == 0);
5790 
5791 	if (uvm.page_init_done == false) {
5792 #ifdef PMAP_STEAL_MEMORY
5793 		pv_addr_t pv;
5794 		pmap_boot_pagealloc(PAGE_SIZE,
5795 #ifdef PMAP_CACHE_VIPT
5796 		    arm_cache_prefer_mask,
5797 		    va & arm_cache_prefer_mask,
5798 #else
5799 		    0, 0,
5800 #endif
5801 		    &pv);
5802 		pa = pv.pv_pa;
5803 #else
5804 		if (uvm_page_physget(&pa) == false)
5805 			return 1;
5806 #endif	/* PMAP_STEAL_MEMORY */
5807 	} else {
5808 		struct vm_page *pg;
5809 		pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE);
5810 		if (pg == NULL)
5811 			return 1;
5812 		pa = VM_PAGE_TO_PHYS(pg);
5813 		/*
5814 		 * This new page must not have any mappings.
5815 		 */
5816 		struct vm_page_md *md __diagused = VM_PAGE_TO_MD(pg);
5817 		KASSERT(SLIST_EMPTY(&md->pvh_list));
5818 	}
5819 
5820 	/*
5821 	 * Enter it via pmap_kenter_pa and let that routine do the hard work.
5822 	 */
5823 	pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE,
5824 	    PMAP_KMPAGE | PMAP_PTE);
5825 
5826 	if (pap)
5827 		*pap = pa;
5828 
5829 	PMAPCOUNT(pt_mappings);
5830 
5831 	const pmap_t kpm __diagused = pmap_kernel();
5832 	struct l2_bucket * const l2b __diagused = pmap_get_l2_bucket(kpm, va);
5833 	KASSERT(l2b != NULL);
5834 
5835 	pt_entry_t * const ptep __diagused = &l2b->l2b_kva[l2pte_index(va)];
5836 	const pt_entry_t pte __diagused = *ptep;
5837 	KASSERT(l2pte_valid_p(pte));
5838 	KASSERT((pte & L2_S_CACHE_MASK) == pte_l2_s_cache_mode_pt);
5839 
5840 	memset((void *)va, 0, PAGE_SIZE);
5841 
5842 	return 0;
5843 }
5844 
5845 /*
5846  * This is the same as pmap_alloc_l2_bucket(), except that it is only
5847  * used by pmap_growkernel().
5848  */
5849 static inline struct l2_bucket *
5850 pmap_grow_l2_bucket(pmap_t pm, vaddr_t va)
5851 {
5852 	const size_t l1slot = l1pte_index(va);
5853 	struct l2_dtable *l2;
5854 	vaddr_t nva;
5855 
5856 	CTASSERT((PAGE_SIZE % L2_TABLE_SIZE_REAL) == 0);
5857 	if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL) {
5858 		/*
5859 		 * No mapping at this address, as there is
5860 		 * no entry in the L1 table.
5861 		 * Need to allocate a new l2_dtable.
5862 		 */
5863 		nva = pmap_kernel_l2dtable_kva;
5864 		if ((nva & PGOFSET) == 0) {
5865 			/*
5866 			 * Need to allocate a backing page
5867 			 */
5868 			if (pmap_grow_map(nva, NULL))
5869 				return NULL;
5870 		}
5871 
5872 		l2 = (struct l2_dtable *)nva;
5873 		nva += sizeof(struct l2_dtable);
5874 
5875 		if ((nva & PGOFSET) < (pmap_kernel_l2dtable_kva & PGOFSET)) {
5876 			/*
5877 			 * The new l2_dtable straddles a page boundary.
5878 			 * Map in another page to cover it.
5879 			 */
5880 			if (pmap_grow_map(nva & ~PGOFSET, NULL))
5881 				return NULL;
5882 		}
5883 
5884 		pmap_kernel_l2dtable_kva = nva;
5885 
5886 		/*
5887 		 * Link it into the parent pmap
5888 		 */
5889 		pm->pm_l2[L2_IDX(l1slot)] = l2;
5890 	}
5891 
5892 	struct l2_bucket * const l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
5893 
5894 	/*
5895 	 * Fetch pointer to the L2 page table associated with the address.
5896 	 */
5897 	if (l2b->l2b_kva == NULL) {
5898 		pt_entry_t *ptep;
5899 
5900 		/*
5901 		 * No L2 page table has been allocated. Chances are, this
5902 		 * is because we just allocated the l2_dtable, above.
5903 		 */
5904 		nva = pmap_kernel_l2ptp_kva;
5905 		ptep = (pt_entry_t *)nva;
5906 		if ((nva & PGOFSET) == 0) {
5907 			/*
5908 			 * Need to allocate a backing page
5909 			 */
5910 			if (pmap_grow_map(nva, &pmap_kernel_l2ptp_phys))
5911 				return NULL;
5912 			PTE_SYNC_RANGE(ptep, PAGE_SIZE / sizeof(pt_entry_t));
5913 		}
5914 
5915 		l2->l2_occupancy++;
5916 		l2b->l2b_kva = ptep;
5917 		l2b->l2b_l1slot = l1slot;
5918 		l2b->l2b_pa = pmap_kernel_l2ptp_phys;
5919 
5920 		pmap_kernel_l2ptp_kva += L2_TABLE_SIZE_REAL;
5921 		pmap_kernel_l2ptp_phys += L2_TABLE_SIZE_REAL;
5922 	}
5923 
5924 	return l2b;
5925 }
5926 
5927 vaddr_t
5928 pmap_growkernel(vaddr_t maxkvaddr)
5929 {
5930 	UVMHIST_FUNC(__func__);
5931 	UVMHIST_CALLARGS(maphist, "growing kernel from %#jx to %#jx\n",
5932 	    pmap_curmaxkvaddr, maxkvaddr, 0, 0);
5933 
5934 	pmap_t kpm = pmap_kernel();
5935 #ifndef ARM_MMU_EXTENDED
5936 	struct l1_ttable *l1;
5937 #endif
5938 	int s;
5939 
5940 	if (maxkvaddr <= pmap_curmaxkvaddr)
5941 		goto out;		/* we are OK */
5942 
5943 	KDASSERT(maxkvaddr <= virtual_end);
5944 
5945 	/*
5946 	 * whoops!   we need to add kernel PTPs
5947 	 */
5948 
5949 	vaddr_t pmap_maxkvaddr = pmap_curmaxkvaddr;
5950 
5951 	s = splvm();	/* to be safe */
5952 	mutex_enter(&kpm_lock);
5953 
5954 	/* Map 1MB at a time */
5955 	size_t l1slot = l1pte_index(pmap_maxkvaddr);
5956 #ifdef ARM_MMU_EXTENDED
5957 	pd_entry_t * const spdep = &kpm->pm_l1[l1slot];
5958 	pd_entry_t *pdep = spdep;
5959 #endif
5960 	for (;pmap_curmaxkvaddr < maxkvaddr; pmap_curmaxkvaddr += L1_S_SIZE,
5961 #ifdef ARM_MMU_EXTENDED
5962 	     pdep++,
5963 #endif
5964 	     l1slot++) {
5965 		struct l2_bucket *l2b =
5966 		    pmap_grow_l2_bucket(kpm, pmap_curmaxkvaddr);
5967 		KASSERT(l2b != NULL);
5968 
5969 		const pd_entry_t npde = L1_C_PROTO | l2b->l2b_pa
5970 		    | L1_C_DOM(PMAP_DOMAIN_KERNEL);
5971 #ifdef ARM_MMU_EXTENDED
5972 		KASSERT(*pdep == 0);
5973 		l1pte_setone(pdep, npde);
5974 #else
5975 		/* Distribute new L1 entry to all other L1s */
5976 		SLIST_FOREACH(l1, &l1_list, l1_link) {
5977 			pd_entry_t * const pdep = &l1->l1_kva[l1slot];
5978 			l1pte_setone(pdep, npde);
5979 			PDE_SYNC(pdep);
5980 		}
5981 #endif
5982 	}
5983 #ifdef ARM_MMU_EXTENDED
5984 	PDE_SYNC_RANGE(spdep, pdep - spdep);
5985 #endif
5986 
5987 #ifdef PMAP_CACHE_VIVT
5988 	/*
5989 	 * flush out the cache, expensive but growkernel will happen so
5990 	 * rarely
5991 	 */
5992 	cpu_dcache_wbinv_all();
5993 	cpu_tlb_flushD();
5994 	cpu_cpwait();
5995 #endif
5996 
5997 	mutex_exit(&kpm_lock);
5998 	splx(s);
5999 
6000 	kasan_shadow_map((void *)pmap_maxkvaddr,
6001 	    (size_t)(pmap_curmaxkvaddr - pmap_maxkvaddr));
6002 
6003 out:
6004 	return pmap_curmaxkvaddr;
6005 }
6006 
6007 /************************ Utility routines ****************************/
6008 
6009 #ifndef ARM_HAS_VBAR
6010 /*
6011  * vector_page_setprot:
6012  *
6013  *	Manipulate the protection of the vector page.
6014  */
6015 void
6016 vector_page_setprot(int prot)
6017 {
6018 	struct l2_bucket *l2b;
6019 	pt_entry_t *ptep;
6020 
6021 #if defined(CPU_ARMV7) || defined(CPU_ARM11)
6022 	/*
6023 	 * If we are using VBAR to use the vectors in the kernel, then it's
6024 	 * already mapped in the kernel text so no need to anything here.
6025 	 */
6026 	if (vector_page != ARM_VECTORS_LOW && vector_page != ARM_VECTORS_HIGH) {
6027 		KASSERT((armreg_pfr1_read() & ARM_PFR1_SEC_MASK) != 0);
6028 		return;
6029 	}
6030 #endif
6031 
6032 	l2b = pmap_get_l2_bucket(pmap_kernel(), vector_page);
6033 	KASSERT(l2b != NULL);
6034 
6035 	ptep = &l2b->l2b_kva[l2pte_index(vector_page)];
6036 
6037 	const pt_entry_t opte = *ptep;
6038 #ifdef ARM_MMU_EXTENDED
6039 	const pt_entry_t npte = (opte & ~(L2_S_PROT_MASK|L2_XS_XN))
6040 	    | L2_S_PROT(PTE_KERNEL, prot);
6041 #else
6042 	const pt_entry_t npte = (opte & ~L2_S_PROT_MASK)
6043 	    | L2_S_PROT(PTE_KERNEL, prot);
6044 #endif
6045 	l2pte_set(ptep, npte, opte);
6046 	PTE_SYNC(ptep);
6047 	cpu_tlb_flushD_SE(vector_page);
6048 	cpu_cpwait();
6049 }
6050 #endif
6051 
6052 /*
6053  * Fetch pointers to the PDE/PTE for the given pmap/VA pair.
6054  * Returns true if the mapping exists, else false.
6055  *
6056  * NOTE: This function is only used by a couple of arm-specific modules.
6057  * It is not safe to take any pmap locks here, since we could be right
6058  * in the middle of debugging the pmap anyway...
6059  *
6060  * It is possible for this routine to return false even though a valid
6061  * mapping does exist. This is because we don't lock, so the metadata
6062  * state may be inconsistent.
6063  *
6064  * NOTE: We can return a NULL *ptp in the case where the L1 pde is
6065  * a "section" mapping.
6066  */
6067 bool
6068 pmap_get_pde_pte(pmap_t pm, vaddr_t va, pd_entry_t **pdp, pt_entry_t **ptp)
6069 {
6070 	struct l2_dtable *l2;
6071 	pd_entry_t *pdep, pde;
6072 	pt_entry_t *ptep;
6073 	u_short l1slot;
6074 
6075 	if (pm->pm_l1 == NULL)
6076 		return false;
6077 
6078 	l1slot = l1pte_index(va);
6079 	*pdp = pdep = pmap_l1_kva(pm) + l1slot;
6080 	pde = *pdep;
6081 
6082 	if (l1pte_section_p(pde)) {
6083 		*ptp = NULL;
6084 		return true;
6085 	}
6086 
6087 	l2 = pm->pm_l2[L2_IDX(l1slot)];
6088 	if (l2 == NULL ||
6089 	    (ptep = l2->l2_bucket[L2_BUCKET(l1slot)].l2b_kva) == NULL) {
6090 		return false;
6091 	}
6092 
6093 	*ptp = &ptep[l2pte_index(va)];
6094 	return true;
6095 }
6096 
6097 bool
6098 pmap_get_pde(pmap_t pm, vaddr_t va, pd_entry_t **pdp)
6099 {
6100 
6101 	if (pm->pm_l1 == NULL)
6102 		return false;
6103 
6104 	*pdp = pmap_l1_kva(pm) + l1pte_index(va);
6105 
6106 	return true;
6107 }
6108 
6109 /************************ Bootstrapping routines ****************************/
6110 
6111 #ifndef ARM_MMU_EXTENDED
6112 static void
6113 pmap_init_l1(struct l1_ttable *l1, pd_entry_t *l1pt)
6114 {
6115 	int i;
6116 
6117 	l1->l1_kva = l1pt;
6118 	l1->l1_domain_use_count = 0;
6119 	l1->l1_domain_first = 0;
6120 
6121 	for (i = 0; i < PMAP_DOMAINS; i++)
6122 		l1->l1_domain_free[i] = i + 1;
6123 
6124 	/*
6125 	 * Copy the kernel's L1 entries to each new L1.
6126 	 */
6127 	if (pmap_initialized)
6128 		memcpy(l1pt, pmap_l1_kva(pmap_kernel()), L1_TABLE_SIZE);
6129 
6130 	if (pmap_extract(pmap_kernel(), (vaddr_t)l1pt,
6131 	    &l1->l1_physaddr) == false)
6132 		panic("pmap_init_l1: can't get PA of L1 at %p", l1pt);
6133 
6134 	SLIST_INSERT_HEAD(&l1_list, l1, l1_link);
6135 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
6136 }
6137 #endif /* !ARM_MMU_EXTENDED */
6138 
6139 /*
6140  * pmap_bootstrap() is called from the board-specific initarm() routine
6141  * once the kernel L1/L2 descriptors tables have been set up.
6142  *
6143  * This is a somewhat convoluted process since pmap bootstrap is, effectively,
6144  * spread over a number of disparate files/functions.
6145  *
6146  * We are passed the following parameters
6147  *  - vstart
6148  *    1MB-aligned start of managed kernel virtual memory.
6149  *  - vend
6150  *    1MB-aligned end of managed kernel virtual memory.
6151  *
6152  * We use 'kernel_l1pt' to build the metadata (struct l1_ttable and
6153  * struct l2_dtable) necessary to track kernel mappings.
6154  */
6155 #define	PMAP_STATIC_L2_SIZE 16
6156 void
6157 pmap_bootstrap(vaddr_t vstart, vaddr_t vend)
6158 {
6159 	static struct l2_dtable static_l2[PMAP_STATIC_L2_SIZE];
6160 #ifndef ARM_MMU_EXTENDED
6161 	static struct l1_ttable static_l1;
6162 	struct l1_ttable *l1 = &static_l1;
6163 #endif
6164 	struct l2_dtable *l2;
6165 	struct l2_bucket *l2b;
6166 	pd_entry_t *l1pt = (pd_entry_t *) kernel_l1pt.pv_va;
6167 	pmap_t pm = pmap_kernel();
6168 	pt_entry_t *ptep;
6169 	paddr_t pa;
6170 	vsize_t size;
6171 	int nptes, l2idx, l2next = 0;
6172 
6173 #ifdef ARM_MMU_EXTENDED
6174 	KASSERT(pte_l1_s_cache_mode == pte_l1_s_cache_mode_pt);
6175 	KASSERT(pte_l2_s_cache_mode == pte_l2_s_cache_mode_pt);
6176 #endif
6177 
6178 	VPRINTF("kpm ");
6179 	/*
6180 	 * Initialise the kernel pmap object
6181 	 */
6182 	curcpu()->ci_pmap_cur = pm;
6183 #ifdef ARM_MMU_EXTENDED
6184 	pm->pm_l1 = l1pt;
6185 	pm->pm_l1_pa = kernel_l1pt.pv_pa;
6186 	VPRINTF("tlb0 ");
6187 	pmap_tlb_info_init(&pmap_tlb0_info);
6188 #ifdef MULTIPROCESSOR
6189 	VPRINTF("kcpusets ");
6190 	pm->pm_onproc = kcpuset_running;
6191 	pm->pm_active = kcpuset_running;
6192 #endif
6193 #else
6194 	pm->pm_l1 = l1;
6195 #endif
6196 
6197 	VPRINTF("locks ");
6198 	/*
6199 	 * pmap_kenter_pa() and pmap_kremove() may be called from interrupt
6200 	 * context, so its locks have to be at IPL_VM
6201 	 */
6202 	mutex_init(&pmap_lock, MUTEX_DEFAULT, IPL_VM);
6203 	mutex_init(&kpm_lock, MUTEX_DEFAULT, IPL_NONE);
6204 	mutex_init(&pm->pm_lock, MUTEX_DEFAULT, IPL_VM);
6205 	pm->pm_refs = 1;
6206 
6207 	VPRINTF("l1pt ");
6208 	/*
6209 	 * Scan the L1 translation table created by initarm() and create
6210 	 * the required metadata for all valid mappings found in it.
6211 	 */
6212 	for (size_t l1slot = 0;
6213 	     l1slot < L1_TABLE_SIZE / sizeof(pd_entry_t);
6214 	     l1slot++) {
6215 		pd_entry_t pde = l1pt[l1slot];
6216 
6217 		/*
6218 		 * We're only interested in Coarse mappings.
6219 		 * pmap_extract() can deal with section mappings without
6220 		 * recourse to checking L2 metadata.
6221 		 */
6222 		if ((pde & L1_TYPE_MASK) != L1_TYPE_C)
6223 			continue;
6224 
6225 		/*
6226 		 * Lookup the KVA of this L2 descriptor table
6227 		 */
6228 		pa = l1pte_pa(pde);
6229 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
6230 		if (ptep == NULL) {
6231 			panic("pmap_bootstrap: No L2 for va 0x%x, pa 0x%lx",
6232 			    (u_int)l1slot << L1_S_SHIFT, pa);
6233 		}
6234 
6235 		/*
6236 		 * Fetch the associated L2 metadata structure.
6237 		 * Allocate a new one if necessary.
6238 		 */
6239 		if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL) {
6240 			if (l2next == PMAP_STATIC_L2_SIZE)
6241 				panic("pmap_bootstrap: out of static L2s");
6242 			pm->pm_l2[L2_IDX(l1slot)] = l2 = &static_l2[l2next++];
6243 		}
6244 
6245 		/*
6246 		 * One more L1 slot tracked...
6247 		 */
6248 		l2->l2_occupancy++;
6249 
6250 		/*
6251 		 * Fill in the details of the L2 descriptor in the
6252 		 * appropriate bucket.
6253 		 */
6254 		l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
6255 		l2b->l2b_kva = ptep;
6256 		l2b->l2b_pa = pa;
6257 		l2b->l2b_l1slot = l1slot;
6258 
6259 		/*
6260 		 * Establish an initial occupancy count for this descriptor
6261 		 */
6262 		for (l2idx = 0;
6263 		    l2idx < (L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
6264 		    l2idx++) {
6265 			if ((ptep[l2idx] & L2_TYPE_MASK) != L2_TYPE_INV) {
6266 				l2b->l2b_occupancy++;
6267 			}
6268 		}
6269 
6270 		/*
6271 		 * Make sure the descriptor itself has the correct cache mode.
6272 		 * If not, fix it, but whine about the problem. Port-meisters
6273 		 * should consider this a clue to fix up their initarm()
6274 		 * function. :)
6275 		 */
6276 		if (pmap_set_pt_cache_mode(l1pt, (vaddr_t)ptep, 1)) {
6277 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
6278 			    "L2 pte @ %p\n", ptep);
6279 		}
6280 	}
6281 
6282 	VPRINTF("cache(l1pt) ");
6283 	/*
6284 	 * Ensure the primary (kernel) L1 has the correct cache mode for
6285 	 * a page table. Bitch if it is not correctly set.
6286 	 */
6287 	if (pmap_set_pt_cache_mode(l1pt, kernel_l1pt.pv_va,
6288 		    L1_TABLE_SIZE / L2_S_SIZE)) {
6289 		printf("pmap_bootstrap: WARNING! wrong cache mode for "
6290 		    "primary L1 @ 0x%lx\n", kernel_l1pt.pv_va);
6291 	}
6292 
6293 #ifdef PMAP_CACHE_VIVT
6294 	cpu_dcache_wbinv_all();
6295 	cpu_tlb_flushID();
6296 	cpu_cpwait();
6297 #endif
6298 
6299 	/*
6300 	 * now we allocate the "special" VAs which are used for tmp mappings
6301 	 * by the pmap (and other modules).  we allocate the VAs by advancing
6302 	 * virtual_avail (note that there are no pages mapped at these VAs).
6303 	 *
6304 	 * Managed KVM space start from wherever initarm() tells us.
6305 	 */
6306 	virtual_avail = vstart;
6307 	virtual_end = vend;
6308 
6309 	VPRINTF("specials ");
6310 
6311 	pmap_alloc_specials(&virtual_avail, 1, &memhook, NULL);
6312 
6313 #ifdef PMAP_CACHE_VIPT
6314 	/*
6315 	 * If we have a VIPT cache, we need one page/pte per possible alias
6316 	 * page so we won't violate cache aliasing rules.
6317 	 */
6318 	virtual_avail = (virtual_avail + arm_cache_prefer_mask) & ~arm_cache_prefer_mask;
6319 	nptes = (arm_cache_prefer_mask >> L2_S_SHIFT) + 1;
6320 	nptes = roundup(nptes, PAGE_SIZE / L2_S_SIZE);
6321 	if (arm_pcache.icache_type != CACHE_TYPE_PIPT
6322 	    && arm_pcache.icache_way_size > nptes * L2_S_SIZE) {
6323 		nptes = arm_pcache.icache_way_size >> L2_S_SHIFT;
6324 		nptes = roundup(nptes, PAGE_SIZE / L2_S_SIZE);
6325 	}
6326 #else
6327 	nptes = PAGE_SIZE / L2_S_SIZE;
6328 #endif
6329 #ifdef MULTIPROCESSOR
6330 	cnptes = nptes;
6331 	nptes *= arm_cpu_max;
6332 #endif
6333 	pmap_alloc_specials(&virtual_avail, nptes, &csrcp, &csrc_pte);
6334 	pmap_set_pt_cache_mode(l1pt, (vaddr_t)csrc_pte, nptes);
6335 	pmap_alloc_specials(&virtual_avail, nptes, &cdstp, &cdst_pte);
6336 	pmap_set_pt_cache_mode(l1pt, (vaddr_t)cdst_pte, nptes);
6337 	if (msgbufaddr == NULL) {
6338 		pmap_alloc_specials(&virtual_avail,
6339 		    round_page(MSGBUFSIZE) / PAGE_SIZE,
6340 		    (void *)&msgbufaddr, NULL);
6341 	}
6342 
6343 	/*
6344 	 * Allocate a range of kernel virtual address space to be used
6345 	 * for L2 descriptor tables and metadata allocation in
6346 	 * pmap_growkernel().
6347 	 */
6348 	size = howmany(virtual_end - pmap_curmaxkvaddr, L1_S_SIZE);
6349 	pmap_alloc_specials(&virtual_avail,
6350 	    round_page(size * L2_TABLE_SIZE_REAL) / PAGE_SIZE,
6351 	    &pmap_kernel_l2ptp_kva, NULL);
6352 
6353 	size = howmany(size, L2_BUCKET_SIZE);
6354 	pmap_alloc_specials(&virtual_avail,
6355 	    round_page(size * sizeof(struct l2_dtable)) / PAGE_SIZE,
6356 	    &pmap_kernel_l2dtable_kva, NULL);
6357 
6358 #ifndef ARM_MMU_EXTENDED
6359 	/*
6360 	 * init the static-global locks and global pmap list.
6361 	 */
6362 	mutex_init(&l1_lru_lock, MUTEX_DEFAULT, IPL_VM);
6363 
6364 	/*
6365 	 * We can now initialise the first L1's metadata.
6366 	 */
6367 	SLIST_INIT(&l1_list);
6368 	TAILQ_INIT(&l1_lru_list);
6369 	pmap_init_l1(l1, l1pt);
6370 #endif /* ARM_MMU_EXTENDED */
6371 
6372 #ifndef ARM_HAS_VBAR
6373 	/* Set up vector page L1 details, if necessary */
6374 	if (vector_page < KERNEL_BASE) {
6375 		pm->pm_pl1vec = pmap_l1_kva(pm) + l1pte_index(vector_page);
6376 		l2b = pmap_get_l2_bucket(pm, vector_page);
6377 		KDASSERT(l2b != NULL);
6378 		pm->pm_l1vec = l2b->l2b_pa | L1_C_PROTO |
6379 		    L1_C_DOM(pmap_domain(pm));
6380 	} else
6381 		pm->pm_pl1vec = NULL;
6382 #endif
6383 
6384 	VPRINTF("pools ");
6385 	/*
6386 	 * Initialize the pmap cache
6387 	 */
6388 	pool_cache_bootstrap(&pmap_cache, sizeof(struct pmap), 0, 0, 0,
6389 	    "pmappl", NULL, IPL_NONE, pmap_pmap_ctor, NULL, NULL);
6390 
6391 	/*
6392 	 * Initialize the pv pool.
6393 	 */
6394 	pool_init(&pmap_pv_pool, sizeof(struct pv_entry), 0, 0, 0, "pvepl",
6395 	    &pmap_bootstrap_pv_allocator, IPL_NONE);
6396 
6397 	/*
6398 	 * Initialize the L2 dtable pool and cache.
6399 	 */
6400 	pool_cache_bootstrap(&pmap_l2dtable_cache, sizeof(struct l2_dtable), 0,
6401 	    0, 0, "l2dtblpl", NULL, IPL_NONE, pmap_l2dtable_ctor, NULL, NULL);
6402 
6403 	/*
6404 	 * Initialise the L2 descriptor table pool and cache
6405 	 */
6406 	pool_cache_bootstrap(&pmap_l2ptp_cache, L2_TABLE_SIZE_REAL,
6407 	    L2_TABLE_SIZE_REAL, 0, 0, "l2ptppl", NULL, IPL_NONE,
6408 	    pmap_l2ptp_ctor, NULL, NULL);
6409 
6410 	mutex_init(&memlock, MUTEX_DEFAULT, IPL_NONE);
6411 
6412 	cpu_dcache_wbinv_all();
6413 }
6414 
6415 static bool
6416 pmap_set_pt_cache_mode(pd_entry_t *kl1, vaddr_t va, size_t nptes)
6417 {
6418 #ifdef ARM_MMU_EXTENDED
6419 	return false;
6420 #else
6421 	if (pte_l1_s_cache_mode == pte_l1_s_cache_mode_pt
6422 	    && pte_l2_s_cache_mode == pte_l2_s_cache_mode_pt)
6423 		return false;
6424 
6425 	const vaddr_t eva = va + nptes * PAGE_SIZE;
6426 	int rv = 0;
6427 
6428 	while (va < eva) {
6429 		/*
6430 		 * Make sure the descriptor itself has the correct cache mode
6431 		 */
6432 		pd_entry_t * const pdep = &kl1[l1pte_index(va)];
6433 		pd_entry_t pde = *pdep;
6434 
6435 		if (l1pte_section_p(pde)) {
6436 			KASSERT((L1_S_CACHE_MASK & L1_S_V6_SUPER) == 0);
6437 			if ((pde & L1_S_CACHE_MASK) != pte_l1_s_cache_mode_pt) {
6438 				*pdep = (pde & ~L1_S_CACHE_MASK) |
6439 				    pte_l1_s_cache_mode_pt;
6440 				PDE_SYNC(pdep);
6441 				cpu_dcache_wbinv_range((vaddr_t)pdep,
6442 				    sizeof(*pdep));
6443 				rv = 1;
6444 			}
6445 			return rv;
6446 		}
6447 		vaddr_t pa = l1pte_pa(pde);
6448 		pt_entry_t *ptep = (pt_entry_t *)kernel_pt_lookup(pa);
6449 		if (ptep == NULL)
6450 			panic("pmap_bootstrap: No PTP for va %#lx\n", va);
6451 
6452 		ptep += l2pte_index(va);
6453 		const pt_entry_t opte = *ptep;
6454 		if ((opte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
6455 			const pt_entry_t npte = (opte & ~L2_S_CACHE_MASK)
6456 			    | pte_l2_s_cache_mode_pt;
6457 			l2pte_set(ptep, npte, opte);
6458 			PTE_SYNC(ptep);
6459 			cpu_dcache_wbinv_range((vaddr_t)ptep, sizeof(*ptep));
6460 			rv = 1;
6461 		}
6462 		va += PAGE_SIZE;
6463 	}
6464 
6465 	return rv;
6466 #endif
6467 }
6468 
6469 static void
6470 pmap_alloc_specials(vaddr_t *availp, int pages, vaddr_t *vap, pt_entry_t **ptep)
6471 {
6472 	vaddr_t va = *availp;
6473 	struct l2_bucket *l2b;
6474 
6475 	if (ptep) {
6476 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
6477 		if (l2b == NULL)
6478 			panic("pmap_alloc_specials: no l2b for 0x%lx", va);
6479 
6480 		*ptep = &l2b->l2b_kva[l2pte_index(va)];
6481 	}
6482 
6483 	*vap = va;
6484 	*availp = va + (PAGE_SIZE * pages);
6485 }
6486 
6487 void
6488 pmap_init(void)
6489 {
6490 
6491 	/*
6492 	 * Set the available memory vars - These do not map to real memory
6493 	 * addresses and cannot as the physical memory is fragmented.
6494 	 * They are used by ps for %mem calculations.
6495 	 * One could argue whether this should be the entire memory or just
6496 	 * the memory that is useable in a user process.
6497 	 */
6498 	avail_start = ptoa(uvm_physseg_get_avail_start(uvm_physseg_get_first()));
6499 	avail_end = ptoa(uvm_physseg_get_avail_end(uvm_physseg_get_last()));
6500 
6501 	/*
6502 	 * Now we need to free enough pv_entry structures to allow us to get
6503 	 * the kmem_map/kmem_object allocated and inited (done after this
6504 	 * function is finished).  to do this we allocate one bootstrap page out
6505 	 * of kernel_map and use it to provide an initial pool of pv_entry
6506 	 * structures.   we never free this page.
6507 	 */
6508 	pool_setlowat(&pmap_pv_pool, (PAGE_SIZE / sizeof(struct pv_entry)) * 2);
6509 
6510 #ifdef ARM_MMU_EXTENDED
6511 	/*
6512 	 * Initialise the L1 pool and cache.
6513 	 */
6514 
6515 	pool_cache_bootstrap(&pmap_l1tt_cache, L1TT_SIZE, L1TT_SIZE,
6516 	    0, 0, "l1ttpl", &pmap_l1tt_allocator, IPL_NONE, pmap_l1tt_ctor,
6517 	     NULL, NULL);
6518 
6519 	int error __diagused = pmap_maxproc_set(maxproc);
6520 	KASSERT(error == 0);
6521 
6522 	pmap_tlb_info_evcnt_attach(&pmap_tlb0_info);
6523 #endif
6524 
6525 	pmap_initialized = true;
6526 }
6527 
6528 static vaddr_t last_bootstrap_page = 0;
6529 static void *free_bootstrap_pages = NULL;
6530 
6531 static void *
6532 pmap_bootstrap_pv_page_alloc(struct pool *pp, int flags)
6533 {
6534 	extern void *pool_page_alloc(struct pool *, int);
6535 	vaddr_t new_page;
6536 	void *rv;
6537 
6538 	if (pmap_initialized)
6539 		return pool_page_alloc(pp, flags);
6540 
6541 	if (free_bootstrap_pages) {
6542 		rv = free_bootstrap_pages;
6543 		free_bootstrap_pages = *((void **)rv);
6544 		return rv;
6545 	}
6546 
6547 	KASSERT(kernel_map != NULL);
6548 	new_page = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
6549 	    UVM_KMF_WIRED | ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT));
6550 
6551 	KASSERT(new_page > last_bootstrap_page);
6552 	last_bootstrap_page = new_page;
6553 	return (void *)new_page;
6554 }
6555 
6556 static void
6557 pmap_bootstrap_pv_page_free(struct pool *pp, void *v)
6558 {
6559 	extern void pool_page_free(struct pool *, void *);
6560 
6561 	if ((vaddr_t)v <= last_bootstrap_page) {
6562 		*((void **)v) = free_bootstrap_pages;
6563 		free_bootstrap_pages = v;
6564 		return;
6565 	}
6566 
6567 	if (pmap_initialized) {
6568 		pool_page_free(pp, v);
6569 		return;
6570 	}
6571 }
6572 
6573 
6574 #if defined(ARM_MMU_EXTENDED)
6575 static void *
6576 pmap_l1tt_alloc(struct pool *pp, int flags)
6577 {
6578 	struct pglist plist;
6579 	vaddr_t va;
6580 
6581 	const int waitok = flags & PR_WAITOK;
6582 
6583 	int error = uvm_pglistalloc(L1TT_SIZE, 0, -1, L1TT_SIZE, 0, &plist, 1,
6584 	    waitok);
6585 	if (error)
6586 		panic("Cannot allocate L1TT physical pages, %d", error);
6587 
6588 	struct vm_page *pg = TAILQ_FIRST(&plist);
6589 #if !defined( __HAVE_MM_MD_DIRECT_MAPPED_PHYS)
6590 
6591 	/* Allocate a L1 translation table VA */
6592 	va = uvm_km_alloc(kernel_map, L1TT_SIZE, L1TT_SIZE, UVM_KMF_VAONLY);
6593 	if (va == 0)
6594 		panic("Cannot allocate L1TT KVA");
6595 
6596 	const vaddr_t eva = va + L1TT_SIZE;
6597 	vaddr_t mva = va;
6598 	while (pg && mva < eva) {
6599 		paddr_t pa = VM_PAGE_TO_PHYS(pg);
6600 
6601 		pmap_kenter_pa(mva, pa,
6602 		    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE|PMAP_PTE);
6603 
6604 		mva += PAGE_SIZE;
6605 		pg = TAILQ_NEXT(pg, pageq.queue);
6606 	}
6607 	KASSERTMSG(pg == NULL && mva == eva, "pg %p mva %" PRIxVADDR
6608 	    " eva %" PRIxVADDR, pg, mva, eva);
6609 #else
6610 	bool ok;
6611 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
6612 	va = pmap_direct_mapped_phys(pa, &ok, 0);
6613 	KASSERT(ok);
6614 	KASSERT(va >= KERNEL_BASE);
6615 #endif
6616 
6617 	return (void *)va;
6618 }
6619 
6620 static void
6621 pmap_l1tt_free(struct pool *pp, void *v)
6622 {
6623 	vaddr_t va = (vaddr_t)v;
6624 
6625 #if !defined( __HAVE_MM_MD_DIRECT_MAPPED_PHYS)
6626 	uvm_km_free(kernel_map, va, L1TT_SIZE, UVM_KMF_WIRED);
6627 #else
6628 #if defined(KERNEL_BASE_VOFFSET)
6629 	paddr_t pa = va - KERNEL_BASE_VOFFSET;
6630 #else
6631 	paddr_t pa = va - KERNEL_BASE + physical_start;
6632 #endif
6633 	const paddr_t epa = pa + L1TT_SIZE;
6634 
6635 	for (; pa < epa; pa += PAGE_SIZE) {
6636 		struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
6637 		uvm_pagefree(pg);
6638 	}
6639 #endif
6640 }
6641 #endif
6642 
6643 /*
6644  * pmap_postinit()
6645  *
6646  * This routine is called after the vm and kmem subsystems have been
6647  * initialised. This allows the pmap code to perform any initialisation
6648  * that can only be done once the memory allocation is in place.
6649  */
6650 void
6651 pmap_postinit(void)
6652 {
6653 #ifndef ARM_MMU_EXTENDED
6654 	extern paddr_t physical_start, physical_end;
6655 	struct l1_ttable *l1;
6656 	struct pglist plist;
6657 	struct vm_page *m;
6658 	pd_entry_t *pdep;
6659 	vaddr_t va, eva;
6660 	u_int loop, needed;
6661 	int error;
6662 #endif
6663 
6664 	pool_cache_setlowat(&pmap_l2ptp_cache, (PAGE_SIZE / L2_TABLE_SIZE_REAL) * 4);
6665 	pool_cache_setlowat(&pmap_l2dtable_cache,
6666 	    (PAGE_SIZE / sizeof(struct l2_dtable)) * 2);
6667 
6668 #ifndef ARM_MMU_EXTENDED
6669 	needed = (maxproc / PMAP_DOMAINS) + ((maxproc % PMAP_DOMAINS) ? 1 : 0);
6670 	needed -= 1;
6671 
6672 	l1 = kmem_alloc(sizeof(*l1) * needed, KM_SLEEP);
6673 
6674 	for (loop = 0; loop < needed; loop++, l1++) {
6675 		/* Allocate a L1 page table */
6676 		va = uvm_km_alloc(kernel_map, L1_TABLE_SIZE, 0, UVM_KMF_VAONLY);
6677 		if (va == 0)
6678 			panic("Cannot allocate L1 KVM");
6679 
6680 		error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start,
6681 		    physical_end, L1_TABLE_SIZE, 0, &plist, 1, 1);
6682 		if (error)
6683 			panic("Cannot allocate L1 physical pages");
6684 
6685 		m = TAILQ_FIRST(&plist);
6686 		eva = va + L1_TABLE_SIZE;
6687 		pdep = (pd_entry_t *)va;
6688 
6689 		while (m && va < eva) {
6690 			paddr_t pa = VM_PAGE_TO_PHYS(m);
6691 
6692 			pmap_kenter_pa(va, pa,
6693 			    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE|PMAP_PTE);
6694 
6695 			va += PAGE_SIZE;
6696 			m = TAILQ_NEXT(m, pageq.queue);
6697 		}
6698 
6699 #ifdef DIAGNOSTIC
6700 		if (m)
6701 			panic("pmap_alloc_l1pt: pglist not empty");
6702 #endif	/* DIAGNOSTIC */
6703 
6704 		pmap_init_l1(l1, pdep);
6705 	}
6706 
6707 #ifdef DEBUG
6708 	printf("pmap_postinit: Allocated %d static L1 descriptor tables\n",
6709 	    needed);
6710 #endif
6711 #endif /* !ARM_MMU_EXTENDED */
6712 }
6713 
6714 /*
6715  * Note that the following routines are used by board-specific initialisation
6716  * code to configure the initial kernel page tables.
6717  *
6718  */
6719 
6720 /*
6721  * This list exists for the benefit of pmap_map_chunk().  It keeps track
6722  * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
6723  * find them as necessary.
6724  *
6725  * Note that the data on this list MUST remain valid after initarm() returns,
6726  * as pmap_bootstrap() uses it to construct L2 table metadata.
6727  */
6728 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
6729 
6730 static vaddr_t
6731 kernel_pt_lookup(paddr_t pa)
6732 {
6733 	pv_addr_t *pv;
6734 
6735 	SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
6736 		if (pv->pv_pa == (pa & ~PGOFSET))
6737 			return pv->pv_va | (pa & PGOFSET);
6738 	}
6739 	return 0;
6740 }
6741 
6742 /*
6743  * pmap_map_section:
6744  *
6745  *	Create a single section mapping.
6746  */
6747 void
6748 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
6749 {
6750 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
6751 	const size_t l1slot = l1pte_index(va);
6752 	pd_entry_t fl;
6753 
6754 	KASSERT(((va | pa) & L1_S_OFFSET) == 0);
6755 
6756 	switch (cache) {
6757 	case PTE_NOCACHE:
6758 		fl = pte_l1_s_nocache_mode;
6759 		break;
6760 
6761 	case PTE_CACHE:
6762 		fl = pte_l1_s_cache_mode;
6763 		break;
6764 
6765 	case PTE_PAGETABLE:
6766 		fl = pte_l1_s_cache_mode_pt;
6767 		break;
6768 
6769 	case PTE_DEV:
6770 	default:
6771 		fl = 0;
6772 		break;
6773 	}
6774 
6775 	const pd_entry_t npde = L1_S_PROTO | pa |
6776 	    L1_S_PROT(PTE_KERNEL, prot) | fl | L1_S_DOM(PMAP_DOMAIN_KERNEL);
6777 	l1pte_setone(pdep + l1slot, npde);
6778 	PDE_SYNC(pdep + l1slot);
6779 }
6780 
6781 /*
6782  * pmap_map_entry:
6783  *
6784  *	Create a single page mapping.
6785  */
6786 void
6787 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
6788 {
6789 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
6790 	const size_t l1slot = l1pte_index(va);
6791 	pt_entry_t npte;
6792 	pt_entry_t *ptep;
6793 
6794 	KASSERT(((va | pa) & PGOFSET) == 0);
6795 
6796 	switch (cache) {
6797 	case PTE_NOCACHE:
6798 		npte = pte_l2_s_nocache_mode;
6799 		break;
6800 
6801 	case PTE_CACHE:
6802 		npte = pte_l2_s_cache_mode;
6803 		break;
6804 
6805 	case PTE_PAGETABLE:
6806 		npte = pte_l2_s_cache_mode_pt;
6807 		break;
6808 
6809 	default:
6810 		npte = 0;
6811 		break;
6812 	}
6813 
6814 	if ((pdep[l1slot] & L1_TYPE_MASK) != L1_TYPE_C)
6815 		panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
6816 
6817 	ptep = (pt_entry_t *) kernel_pt_lookup(l1pte_pa(pdep[l1slot]));
6818 	if (ptep == NULL)
6819 		panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
6820 
6821 	npte |= L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot);
6822 #ifdef ARM_MMU_EXTENDED
6823 	if (prot & VM_PROT_EXECUTE) {
6824 		npte &= ~L2_XS_XN;
6825 	}
6826 #endif
6827 	ptep += l2pte_index(va);
6828 	l2pte_set(ptep, npte, 0);
6829 	PTE_SYNC(ptep);
6830 }
6831 
6832 /*
6833  * pmap_link_l2pt:
6834  *
6835  *	Link the L2 page table specified by "l2pv" into the L1
6836  *	page table at the slot for "va".
6837  */
6838 void
6839 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
6840 {
6841 	pd_entry_t * const pdep = (pd_entry_t *) l1pt + l1pte_index(va);
6842 
6843 	KASSERT((va & ((L1_S_SIZE * (PAGE_SIZE / L2_T_SIZE)) - 1)) == 0);
6844 	KASSERT((l2pv->pv_pa & PGOFSET) == 0);
6845 
6846 	const pd_entry_t npde = L1_C_DOM(PMAP_DOMAIN_KERNEL) | L1_C_PROTO
6847 	    | l2pv->pv_pa;
6848 
6849 	l1pte_set(pdep, npde);
6850 	PDE_SYNC_RANGE(pdep, PAGE_SIZE / L2_T_SIZE);
6851 
6852 	SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
6853 }
6854 
6855 /*
6856  * pmap_map_chunk:
6857  *
6858  *	Map a chunk of memory using the most efficient mappings
6859  *	possible (section, large page, small page) into the
6860  *	provided L1 and L2 tables at the specified virtual address.
6861  */
6862 vsize_t
6863 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
6864     int prot, int cache)
6865 {
6866 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
6867 	pt_entry_t f1, f2s, f2l;
6868 	vsize_t resid;
6869 
6870 	resid = (size + (PAGE_SIZE - 1)) & ~(PAGE_SIZE - 1);
6871 
6872 	if (l1pt == 0)
6873 		panic("pmap_map_chunk: no L1 table provided");
6874 
6875 // 	VPRINTF("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
6876 // 	    "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
6877 
6878 	switch (cache) {
6879 	case PTE_NOCACHE:
6880 		f1 = pte_l1_s_nocache_mode;
6881 		f2l = pte_l2_l_nocache_mode;
6882 		f2s = pte_l2_s_nocache_mode;
6883 		break;
6884 
6885 	case PTE_CACHE:
6886 		f1 = pte_l1_s_cache_mode;
6887 		f2l = pte_l2_l_cache_mode;
6888 		f2s = pte_l2_s_cache_mode;
6889 		break;
6890 
6891 	case PTE_PAGETABLE:
6892 		f1 = pte_l1_s_cache_mode_pt;
6893 		f2l = pte_l2_l_cache_mode_pt;
6894 		f2s = pte_l2_s_cache_mode_pt;
6895 		break;
6896 
6897 	case PTE_DEV:
6898 	default:
6899 		f1 = 0;
6900 		f2l = 0;
6901 		f2s = 0;
6902 		break;
6903 	}
6904 
6905 	size = resid;
6906 
6907 	while (resid > 0) {
6908 		const size_t l1slot = l1pte_index(va);
6909 #ifdef ARM_MMU_EXTENDED
6910 		/* See if we can use a supersection mapping. */
6911 		if (L1_SS_PROTO && L1_SS_MAPPABLE_P(va, pa, resid)) {
6912 			/* Supersection are always domain 0 */
6913 			const pd_entry_t npde = L1_SS_PROTO | pa
6914 			    | ((prot & VM_PROT_EXECUTE) ? 0 : L1_S_V6_XN)
6915 			    | (va & 0x80000000 ? 0 : L1_S_V6_nG)
6916 			    | L1_S_PROT(PTE_KERNEL, prot) | f1;
6917 			VPRINTF("sS");
6918 			l1pte_set(&pdep[l1slot], npde);
6919 			PDE_SYNC_RANGE(&pdep[l1slot], L1_SS_SIZE / L1_S_SIZE);
6920 //			VPRINTF("\npmap_map_chunk: pa=0x%lx va=0x%lx resid=0x%08lx "
6921 //			    "npdep=%p pde=0x%x\n", pa, va, resid, &pdep[l1slot], npde);
6922 			va += L1_SS_SIZE;
6923 			pa += L1_SS_SIZE;
6924 			resid -= L1_SS_SIZE;
6925 			continue;
6926 		}
6927 #endif
6928 		/* See if we can use a section mapping. */
6929 		if (L1_S_MAPPABLE_P(va, pa, resid)) {
6930 			const pd_entry_t npde = L1_S_PROTO | pa
6931 #ifdef ARM_MMU_EXTENDED
6932 			    | ((prot & VM_PROT_EXECUTE) ? 0 : L1_S_V6_XN)
6933 			    | (va & 0x80000000 ? 0 : L1_S_V6_nG)
6934 #endif
6935 			    | L1_S_PROT(PTE_KERNEL, prot) | f1
6936 			    | L1_S_DOM(PMAP_DOMAIN_KERNEL);
6937 			VPRINTF("S");
6938 			l1pte_set(&pdep[l1slot], npde);
6939 			PDE_SYNC(&pdep[l1slot]);
6940 //			VPRINTF("\npmap_map_chunk: pa=0x%lx va=0x%lx resid=0x%08lx "
6941 //			    "npdep=%p pde=0x%x\n", pa, va, resid, &pdep[l1slot], npde);
6942 			va += L1_S_SIZE;
6943 			pa += L1_S_SIZE;
6944 			resid -= L1_S_SIZE;
6945 			continue;
6946 		}
6947 
6948 		/*
6949 		 * Ok, we're going to use an L2 table.  Make sure
6950 		 * one is actually in the corresponding L1 slot
6951 		 * for the current VA.
6952 		 */
6953 		if ((pdep[l1slot] & L1_TYPE_MASK) != L1_TYPE_C)
6954 			panic("%s: no L2 table for VA %#lx", __func__, va);
6955 
6956 		pt_entry_t *ptep = (pt_entry_t *) kernel_pt_lookup(l1pte_pa(pdep[l1slot]));
6957 		if (ptep == NULL)
6958 			panic("%s: can't find L2 table for VA %#lx", __func__,
6959 			    va);
6960 
6961 		ptep += l2pte_index(va);
6962 
6963 		/* See if we can use a L2 large page mapping. */
6964 		if (L2_L_MAPPABLE_P(va, pa, resid)) {
6965 			const pt_entry_t npte = L2_L_PROTO | pa
6966 #ifdef ARM_MMU_EXTENDED
6967 			    | ((prot & VM_PROT_EXECUTE) ? 0 : L2_XS_L_XN)
6968 			    | (va & 0x80000000 ? 0 : L2_XS_nG)
6969 #endif
6970 			    | L2_L_PROT(PTE_KERNEL, prot) | f2l;
6971 			VPRINTF("L");
6972 			l2pte_set(ptep, npte, 0);
6973 			PTE_SYNC_RANGE(ptep, L2_L_SIZE / L2_S_SIZE);
6974 			va += L2_L_SIZE;
6975 			pa += L2_L_SIZE;
6976 			resid -= L2_L_SIZE;
6977 			continue;
6978 		}
6979 
6980 		VPRINTF("P");
6981 		/* Use a small page mapping. */
6982 		pt_entry_t npte = L2_S_PROTO | pa
6983 #ifdef ARM_MMU_EXTENDED
6984 		    | ((prot & VM_PROT_EXECUTE) ? 0 : L2_XS_XN)
6985 		    | (va & 0x80000000 ? 0 : L2_XS_nG)
6986 #endif
6987 		    | L2_S_PROT(PTE_KERNEL, prot) | f2s;
6988 #ifdef ARM_MMU_EXTENDED
6989 		npte &= ((prot & VM_PROT_EXECUTE) ? ~L2_XS_XN : ~0);
6990 #endif
6991 		l2pte_set(ptep, npte, 0);
6992 		PTE_SYNC(ptep);
6993 		va += PAGE_SIZE;
6994 		pa += PAGE_SIZE;
6995 		resid -= PAGE_SIZE;
6996 	}
6997 	VPRINTF("\n");
6998 	return size;
6999 }
7000 
7001 /*
7002  * pmap_unmap_chunk:
7003  *
7004  *	Unmap a chunk of memory that was previously pmap_map_chunk
7005  */
7006 void
7007 pmap_unmap_chunk(vaddr_t l1pt, vaddr_t va, vsize_t size)
7008 {
7009 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
7010 	const size_t l1slot = l1pte_index(va);
7011 
7012 	KASSERT(size == L1_SS_SIZE || size == L1_S_SIZE);
7013 
7014 	l1pte_set(&pdep[l1slot], 0);
7015 	PDE_SYNC_RANGE(&pdep[l1slot], size / L1_S_SIZE);
7016 
7017 	pmap_tlb_flush_SE(pmap_kernel(), va, PVF_REF);
7018 }
7019 
7020 
7021 
7022 /********************** Static device map routines ***************************/
7023 
7024 static const struct pmap_devmap *pmap_devmap_table;
7025 
7026 /*
7027  * Register the devmap table.  This is provided in case early console
7028  * initialization needs to register mappings created by bootstrap code
7029  * before pmap_devmap_bootstrap() is called.
7030  */
7031 void
7032 pmap_devmap_register(const struct pmap_devmap *table)
7033 {
7034 
7035 	pmap_devmap_table = table;
7036 }
7037 
7038 /*
7039  * Map all of the static regions in the devmap table, and remember
7040  * the devmap table so other parts of the kernel can look up entries
7041  * later.
7042  */
7043 void
7044 pmap_devmap_bootstrap(vaddr_t l1pt, const struct pmap_devmap *table)
7045 {
7046 	int i;
7047 
7048 	pmap_devmap_table = table;
7049 
7050 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
7051 		const struct pmap_devmap *pdp = &pmap_devmap_table[i];
7052 
7053 		KASSERTMSG(VADDR_MAX - pdp->pd_va >= pdp->pd_size - 1, "va %" PRIxVADDR
7054 		    " sz %" PRIxPSIZE, pdp->pd_va, pdp->pd_size);
7055 		KASSERTMSG(PADDR_MAX - pdp->pd_pa >= pdp->pd_size - 1, "pa %" PRIxPADDR
7056 		    " sz %" PRIxPSIZE, pdp->pd_pa, pdp->pd_size);
7057 		VPRINTF("devmap: %08lx -> %08lx @ %08lx\n", pdp->pd_pa,
7058 		    pdp->pd_pa + pdp->pd_size - 1, pdp->pd_va);
7059 
7060 		pmap_map_chunk(l1pt, pdp->pd_va, pdp->pd_pa, pdp->pd_size,
7061 		    pdp->pd_prot, pdp->pd_cache);
7062 	}
7063 }
7064 
7065 const struct pmap_devmap *
7066 pmap_devmap_find_pa(paddr_t pa, psize_t size)
7067 {
7068 	uint64_t endpa;
7069 	int i;
7070 
7071 	if (pmap_devmap_table == NULL)
7072 		return NULL;
7073 
7074 	endpa = (uint64_t)pa + (uint64_t)(size - 1);
7075 
7076 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
7077 		if (pa >= pmap_devmap_table[i].pd_pa &&
7078 		    endpa <= (uint64_t)pmap_devmap_table[i].pd_pa +
7079 			     (uint64_t)(pmap_devmap_table[i].pd_size - 1))
7080 			return &pmap_devmap_table[i];
7081 	}
7082 
7083 	return NULL;
7084 }
7085 
7086 const struct pmap_devmap *
7087 pmap_devmap_find_va(vaddr_t va, vsize_t size)
7088 {
7089 	int i;
7090 
7091 	if (pmap_devmap_table == NULL)
7092 		return NULL;
7093 
7094 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
7095 		if (va >= pmap_devmap_table[i].pd_va &&
7096 		    va + size - 1 <= pmap_devmap_table[i].pd_va +
7097 				     pmap_devmap_table[i].pd_size - 1)
7098 			return &pmap_devmap_table[i];
7099 	}
7100 
7101 	return NULL;
7102 }
7103 
7104 /********************** PTE initialization routines **************************/
7105 
7106 /*
7107  * These routines are called when the CPU type is identified to set up
7108  * the PTE prototypes, cache modes, etc.
7109  *
7110  * The variables are always here, just in case modules need to reference
7111  * them (though, they shouldn't).
7112  */
7113 
7114 pt_entry_t	pte_l1_s_nocache_mode;
7115 pt_entry_t	pte_l1_s_cache_mode;
7116 pt_entry_t	pte_l1_s_wc_mode;
7117 pt_entry_t	pte_l1_s_cache_mode_pt;
7118 pt_entry_t	pte_l1_s_cache_mask;
7119 
7120 pt_entry_t	pte_l2_l_nocache_mode;
7121 pt_entry_t	pte_l2_l_cache_mode;
7122 pt_entry_t	pte_l2_l_wc_mode;
7123 pt_entry_t	pte_l2_l_cache_mode_pt;
7124 pt_entry_t	pte_l2_l_cache_mask;
7125 
7126 pt_entry_t	pte_l2_s_nocache_mode;
7127 pt_entry_t	pte_l2_s_cache_mode;
7128 pt_entry_t	pte_l2_s_wc_mode;
7129 pt_entry_t	pte_l2_s_cache_mode_pt;
7130 pt_entry_t	pte_l2_s_cache_mask;
7131 
7132 pt_entry_t	pte_l1_s_prot_u;
7133 pt_entry_t	pte_l1_s_prot_w;
7134 pt_entry_t	pte_l1_s_prot_ro;
7135 pt_entry_t	pte_l1_s_prot_mask;
7136 
7137 pt_entry_t	pte_l2_s_prot_u;
7138 pt_entry_t	pte_l2_s_prot_w;
7139 pt_entry_t	pte_l2_s_prot_ro;
7140 pt_entry_t	pte_l2_s_prot_mask;
7141 
7142 pt_entry_t	pte_l2_l_prot_u;
7143 pt_entry_t	pte_l2_l_prot_w;
7144 pt_entry_t	pte_l2_l_prot_ro;
7145 pt_entry_t	pte_l2_l_prot_mask;
7146 
7147 pt_entry_t	pte_l1_ss_proto;
7148 pt_entry_t	pte_l1_s_proto;
7149 pt_entry_t	pte_l1_c_proto;
7150 pt_entry_t	pte_l2_s_proto;
7151 
7152 void		(*pmap_copy_page_func)(paddr_t, paddr_t);
7153 void		(*pmap_zero_page_func)(paddr_t);
7154 
7155 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
7156 void
7157 pmap_pte_init_generic(void)
7158 {
7159 
7160 	pte_l1_s_nocache_mode = 0;
7161 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
7162 	pte_l1_s_wc_mode = L1_S_B;
7163 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_generic;
7164 
7165 	pte_l2_l_nocache_mode = 0;
7166 	pte_l2_l_cache_mode = L2_B|L2_C;
7167 	pte_l2_l_wc_mode = L2_B;
7168 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_generic;
7169 
7170 	pte_l2_s_nocache_mode = 0;
7171 	pte_l2_s_cache_mode = L2_B|L2_C;
7172 	pte_l2_s_wc_mode = L2_B;
7173 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_generic;
7174 
7175 	/*
7176 	 * If we have a write-through cache, set B and C.  If
7177 	 * we have a write-back cache, then we assume setting
7178 	 * only C will make those pages write-through (except for those
7179 	 * Cortex CPUs which can read the L1 caches).
7180 	 */
7181 	if (cpufuncs.cf_dcache_wb_range == (void *) cpufunc_nullop
7182 #if ARM_MMU_V7 > 0
7183 	    || CPU_ID_CORTEX_P(curcpu()->ci_arm_cpuid)
7184 #endif
7185 #if ARM_MMU_V6 > 0
7186 	    || CPU_ID_ARM11_P(curcpu()->ci_arm_cpuid) /* arm116 errata 399234 */
7187 #endif
7188 	    || false) {
7189 		pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
7190 		pte_l2_l_cache_mode_pt = L2_B|L2_C;
7191 		pte_l2_s_cache_mode_pt = L2_B|L2_C;
7192 	} else {
7193 		pte_l1_s_cache_mode_pt = L1_S_C;	/* write through */
7194 		pte_l2_l_cache_mode_pt = L2_C;		/* write through */
7195 		pte_l2_s_cache_mode_pt = L2_C;		/* write through */
7196 	}
7197 
7198 	pte_l1_s_prot_u = L1_S_PROT_U_generic;
7199 	pte_l1_s_prot_w = L1_S_PROT_W_generic;
7200 	pte_l1_s_prot_ro = L1_S_PROT_RO_generic;
7201 	pte_l1_s_prot_mask = L1_S_PROT_MASK_generic;
7202 
7203 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
7204 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
7205 	pte_l2_s_prot_ro = L2_S_PROT_RO_generic;
7206 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
7207 
7208 	pte_l2_l_prot_u = L2_L_PROT_U_generic;
7209 	pte_l2_l_prot_w = L2_L_PROT_W_generic;
7210 	pte_l2_l_prot_ro = L2_L_PROT_RO_generic;
7211 	pte_l2_l_prot_mask = L2_L_PROT_MASK_generic;
7212 
7213 	pte_l1_ss_proto = L1_SS_PROTO_generic;
7214 	pte_l1_s_proto = L1_S_PROTO_generic;
7215 	pte_l1_c_proto = L1_C_PROTO_generic;
7216 	pte_l2_s_proto = L2_S_PROTO_generic;
7217 
7218 	pmap_copy_page_func = pmap_copy_page_generic;
7219 	pmap_zero_page_func = pmap_zero_page_generic;
7220 }
7221 
7222 #if defined(CPU_ARM8)
7223 void
7224 pmap_pte_init_arm8(void)
7225 {
7226 
7227 	/*
7228 	 * ARM8 is compatible with generic, but we need to use
7229 	 * the page tables uncached.
7230 	 */
7231 	pmap_pte_init_generic();
7232 
7233 	pte_l1_s_cache_mode_pt = 0;
7234 	pte_l2_l_cache_mode_pt = 0;
7235 	pte_l2_s_cache_mode_pt = 0;
7236 }
7237 #endif /* CPU_ARM8 */
7238 
7239 #if defined(CPU_ARM9) && defined(ARM9_CACHE_WRITE_THROUGH)
7240 void
7241 pmap_pte_init_arm9(void)
7242 {
7243 
7244 	/*
7245 	 * ARM9 is compatible with generic, but we want to use
7246 	 * write-through caching for now.
7247 	 */
7248 	pmap_pte_init_generic();
7249 
7250 	pte_l1_s_cache_mode = L1_S_C;
7251 	pte_l2_l_cache_mode = L2_C;
7252 	pte_l2_s_cache_mode = L2_C;
7253 
7254 	pte_l1_s_wc_mode = L1_S_B;
7255 	pte_l2_l_wc_mode = L2_B;
7256 	pte_l2_s_wc_mode = L2_B;
7257 
7258 	pte_l1_s_cache_mode_pt = L1_S_C;
7259 	pte_l2_l_cache_mode_pt = L2_C;
7260 	pte_l2_s_cache_mode_pt = L2_C;
7261 }
7262 #endif /* CPU_ARM9 && ARM9_CACHE_WRITE_THROUGH */
7263 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
7264 
7265 #if defined(CPU_ARM10)
7266 void
7267 pmap_pte_init_arm10(void)
7268 {
7269 
7270 	/*
7271 	 * ARM10 is compatible with generic, but we want to use
7272 	 * write-through caching for now.
7273 	 */
7274 	pmap_pte_init_generic();
7275 
7276 	pte_l1_s_cache_mode = L1_S_B | L1_S_C;
7277 	pte_l2_l_cache_mode = L2_B | L2_C;
7278 	pte_l2_s_cache_mode = L2_B | L2_C;
7279 
7280 	pte_l1_s_cache_mode = L1_S_B;
7281 	pte_l2_l_cache_mode = L2_B;
7282 	pte_l2_s_cache_mode = L2_B;
7283 
7284 	pte_l1_s_cache_mode_pt = L1_S_C;
7285 	pte_l2_l_cache_mode_pt = L2_C;
7286 	pte_l2_s_cache_mode_pt = L2_C;
7287 
7288 }
7289 #endif /* CPU_ARM10 */
7290 
7291 #if defined(CPU_ARM11) && defined(ARM11_CACHE_WRITE_THROUGH)
7292 void
7293 pmap_pte_init_arm11(void)
7294 {
7295 
7296 	/*
7297 	 * ARM11 is compatible with generic, but we want to use
7298 	 * write-through caching for now.
7299 	 */
7300 	pmap_pte_init_generic();
7301 
7302 	pte_l1_s_cache_mode = L1_S_C;
7303 	pte_l2_l_cache_mode = L2_C;
7304 	pte_l2_s_cache_mode = L2_C;
7305 
7306 	pte_l1_s_wc_mode = L1_S_B;
7307 	pte_l2_l_wc_mode = L2_B;
7308 	pte_l2_s_wc_mode = L2_B;
7309 
7310 	pte_l1_s_cache_mode_pt = L1_S_C;
7311 	pte_l2_l_cache_mode_pt = L2_C;
7312 	pte_l2_s_cache_mode_pt = L2_C;
7313 }
7314 #endif /* CPU_ARM11 && ARM11_CACHE_WRITE_THROUGH */
7315 
7316 #if ARM_MMU_SA1 == 1
7317 void
7318 pmap_pte_init_sa1(void)
7319 {
7320 
7321 	/*
7322 	 * The StrongARM SA-1 cache does not have a write-through
7323 	 * mode.  So, do the generic initialization, then reset
7324 	 * the page table cache mode to B=1,C=1, and note that
7325 	 * the PTEs need to be sync'd.
7326 	 */
7327 	pmap_pte_init_generic();
7328 
7329 	pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
7330 	pte_l2_l_cache_mode_pt = L2_B|L2_C;
7331 	pte_l2_s_cache_mode_pt = L2_B|L2_C;
7332 
7333 	pmap_needs_pte_sync = 1;
7334 }
7335 #endif /* ARM_MMU_SA1 == 1*/
7336 
7337 #if ARM_MMU_XSCALE == 1
7338 #if (ARM_NMMUS > 1)
7339 static u_int xscale_use_minidata;
7340 #endif
7341 
7342 void
7343 pmap_pte_init_xscale(void)
7344 {
7345 	uint32_t auxctl;
7346 	int write_through = 0;
7347 
7348 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
7349 	pte_l1_s_wc_mode = L1_S_B;
7350 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_xscale;
7351 
7352 	pte_l2_l_cache_mode = L2_B|L2_C;
7353 	pte_l2_l_wc_mode = L2_B;
7354 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_xscale;
7355 
7356 	pte_l2_s_cache_mode = L2_B|L2_C;
7357 	pte_l2_s_wc_mode = L2_B;
7358 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_xscale;
7359 
7360 	pte_l1_s_cache_mode_pt = L1_S_C;
7361 	pte_l2_l_cache_mode_pt = L2_C;
7362 	pte_l2_s_cache_mode_pt = L2_C;
7363 
7364 #ifdef XSCALE_CACHE_READ_WRITE_ALLOCATE
7365 	/*
7366 	 * The XScale core has an enhanced mode where writes that
7367 	 * miss the cache cause a cache line to be allocated.  This
7368 	 * is significantly faster than the traditional, write-through
7369 	 * behavior of this case.
7370 	 */
7371 	pte_l1_s_cache_mode |= L1_S_XS_TEX(TEX_XSCALE_X);
7372 	pte_l2_l_cache_mode |= L2_XS_L_TEX(TEX_XSCALE_X);
7373 	pte_l2_s_cache_mode |= L2_XS_T_TEX(TEX_XSCALE_X);
7374 #endif /* XSCALE_CACHE_READ_WRITE_ALLOCATE */
7375 
7376 #ifdef XSCALE_CACHE_WRITE_THROUGH
7377 	/*
7378 	 * Some versions of the XScale core have various bugs in
7379 	 * their cache units, the work-around for which is to run
7380 	 * the cache in write-through mode.  Unfortunately, this
7381 	 * has a major (negative) impact on performance.  So, we
7382 	 * go ahead and run fast-and-loose, in the hopes that we
7383 	 * don't line up the planets in a way that will trip the
7384 	 * bugs.
7385 	 *
7386 	 * However, we give you the option to be slow-but-correct.
7387 	 */
7388 	write_through = 1;
7389 #elif defined(XSCALE_CACHE_WRITE_BACK)
7390 	/* force write back cache mode */
7391 	write_through = 0;
7392 #elif defined(CPU_XSCALE_PXA250) || defined(CPU_XSCALE_PXA270)
7393 	/*
7394 	 * Intel PXA2[15]0 processors are known to have a bug in
7395 	 * write-back cache on revision 4 and earlier (stepping
7396 	 * A[01] and B[012]).  Fixed for C0 and later.
7397 	 */
7398 	{
7399 		uint32_t id, type;
7400 
7401 		id = cpufunc_id();
7402 		type = id & ~(CPU_ID_XSCALE_COREREV_MASK|CPU_ID_REVISION_MASK);
7403 
7404 		if (type == CPU_ID_PXA250 || type == CPU_ID_PXA210) {
7405 			if ((id & CPU_ID_REVISION_MASK) < 5) {
7406 				/* write through for stepping A0-1 and B0-2 */
7407 				write_through = 1;
7408 			}
7409 		}
7410 	}
7411 #endif /* XSCALE_CACHE_WRITE_THROUGH */
7412 
7413 	if (write_through) {
7414 		pte_l1_s_cache_mode = L1_S_C;
7415 		pte_l2_l_cache_mode = L2_C;
7416 		pte_l2_s_cache_mode = L2_C;
7417 	}
7418 
7419 #if (ARM_NMMUS > 1)
7420 	xscale_use_minidata = 1;
7421 #endif
7422 
7423 	pte_l1_s_prot_u = L1_S_PROT_U_xscale;
7424 	pte_l1_s_prot_w = L1_S_PROT_W_xscale;
7425 	pte_l1_s_prot_ro = L1_S_PROT_RO_xscale;
7426 	pte_l1_s_prot_mask = L1_S_PROT_MASK_xscale;
7427 
7428 	pte_l2_s_prot_u = L2_S_PROT_U_xscale;
7429 	pte_l2_s_prot_w = L2_S_PROT_W_xscale;
7430 	pte_l2_s_prot_ro = L2_S_PROT_RO_xscale;
7431 	pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale;
7432 
7433 	pte_l2_l_prot_u = L2_L_PROT_U_xscale;
7434 	pte_l2_l_prot_w = L2_L_PROT_W_xscale;
7435 	pte_l2_l_prot_ro = L2_L_PROT_RO_xscale;
7436 	pte_l2_l_prot_mask = L2_L_PROT_MASK_xscale;
7437 
7438 	pte_l1_ss_proto = L1_SS_PROTO_xscale;
7439 	pte_l1_s_proto = L1_S_PROTO_xscale;
7440 	pte_l1_c_proto = L1_C_PROTO_xscale;
7441 	pte_l2_s_proto = L2_S_PROTO_xscale;
7442 
7443 	pmap_copy_page_func = pmap_copy_page_xscale;
7444 	pmap_zero_page_func = pmap_zero_page_xscale;
7445 
7446 	/*
7447 	 * Disable ECC protection of page table access, for now.
7448 	 */
7449 	auxctl = armreg_auxctl_read();
7450 	auxctl &= ~XSCALE_AUXCTL_P;
7451 	armreg_auxctl_write(auxctl);
7452 }
7453 
7454 /*
7455  * xscale_setup_minidata:
7456  *
7457  *	Set up the mini-data cache clean area.  We require the
7458  *	caller to allocate the right amount of physically and
7459  *	virtually contiguous space.
7460  */
7461 void
7462 xscale_setup_minidata(vaddr_t l1pt, vaddr_t va, paddr_t pa)
7463 {
7464 	extern vaddr_t xscale_minidata_clean_addr;
7465 	extern vsize_t xscale_minidata_clean_size; /* already initialized */
7466 	pd_entry_t *pde = (pd_entry_t *) l1pt;
7467 	vsize_t size;
7468 	uint32_t auxctl;
7469 
7470 	xscale_minidata_clean_addr = va;
7471 
7472 	/* Round it to page size. */
7473 	size = (xscale_minidata_clean_size + L2_S_OFFSET) & L2_S_FRAME;
7474 
7475 	for (; size != 0;
7476 	     va += L2_S_SIZE, pa += L2_S_SIZE, size -= L2_S_SIZE) {
7477 		const size_t l1slot = l1pte_index(va);
7478 		pt_entry_t *ptep = (pt_entry_t *) kernel_pt_lookup(l1pte_pa(pde[l1slot]));
7479 		if (ptep == NULL)
7480 			panic("xscale_setup_minidata: can't find L2 table for "
7481 			    "VA 0x%08lx", va);
7482 
7483 		ptep += l2pte_index(va);
7484 		pt_entry_t opte = *ptep;
7485 		l2pte_set(ptep,
7486 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, VM_PROT_READ)
7487 		    | L2_C | L2_XS_T_TEX(TEX_XSCALE_X), opte);
7488 	}
7489 
7490 	/*
7491 	 * Configure the mini-data cache for write-back with
7492 	 * read/write-allocate.
7493 	 *
7494 	 * NOTE: In order to reconfigure the mini-data cache, we must
7495 	 * make sure it contains no valid data!  In order to do that,
7496 	 * we must issue a global data cache invalidate command!
7497 	 *
7498 	 * WE ASSUME WE ARE RUNNING UN-CACHED WHEN THIS ROUTINE IS CALLED!
7499 	 * THIS IS VERY IMPORTANT!
7500 	 */
7501 
7502 	/* Invalidate data and mini-data. */
7503 	__asm volatile("mcr p15, 0, %0, c7, c6, 0" : : "r" (0));
7504 	auxctl = armreg_auxctl_read();
7505 	auxctl = (auxctl & ~XSCALE_AUXCTL_MD_MASK) | XSCALE_AUXCTL_MD_WB_RWA;
7506 	armreg_auxctl_write(auxctl);
7507 }
7508 
7509 /*
7510  * Change the PTEs for the specified kernel mappings such that they
7511  * will use the mini data cache instead of the main data cache.
7512  */
7513 void
7514 pmap_uarea(vaddr_t va)
7515 {
7516 	vaddr_t next_bucket, eva;
7517 
7518 #if (ARM_NMMUS > 1)
7519 	if (xscale_use_minidata == 0)
7520 		return;
7521 #endif
7522 
7523 	eva = va + USPACE;
7524 
7525 	while (va < eva) {
7526 		next_bucket = L2_NEXT_BUCKET_VA(va);
7527 		if (next_bucket > eva)
7528 			next_bucket = eva;
7529 
7530 		struct l2_bucket *l2b = pmap_get_l2_bucket(pmap_kernel(), va);
7531 		KDASSERT(l2b != NULL);
7532 
7533 		pt_entry_t * const sptep = &l2b->l2b_kva[l2pte_index(va)];
7534 		pt_entry_t *ptep = sptep;
7535 
7536 		while (va < next_bucket) {
7537 			const pt_entry_t opte = *ptep;
7538 			if (!l2pte_minidata_p(opte)) {
7539 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
7540 				cpu_tlb_flushD_SE(va);
7541 				l2pte_set(ptep, opte & ~L2_B, opte);
7542 			}
7543 			ptep += PAGE_SIZE / L2_S_SIZE;
7544 			va += PAGE_SIZE;
7545 		}
7546 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
7547 	}
7548 	cpu_cpwait();
7549 }
7550 #endif /* ARM_MMU_XSCALE == 1 */
7551 
7552 
7553 #if defined(CPU_ARM11MPCORE)
7554 void
7555 pmap_pte_init_arm11mpcore(void)
7556 {
7557 
7558 	/* cache mode is controlled by 5 bits (B, C, TEX[2:0]) */
7559 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_armv6;
7560 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_armv6;
7561 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7562 	/* use extended small page (without APn, with TEX) */
7563 	pte_l2_s_cache_mask = L2_XS_CACHE_MASK_armv6;
7564 #else
7565 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_armv6c;
7566 #endif
7567 
7568 	/* write-back, write-allocate */
7569 	pte_l1_s_cache_mode = L1_S_C | L1_S_B | L1_S_V6_TEX(0x01);
7570 	pte_l2_l_cache_mode = L2_C | L2_B | L2_V6_L_TEX(0x01);
7571 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7572 	pte_l2_s_cache_mode = L2_C | L2_B | L2_V6_XS_TEX(0x01);
7573 #else
7574 	/* no TEX. read-allocate */
7575 	pte_l2_s_cache_mode = L2_C | L2_B;
7576 #endif
7577 	/*
7578 	 * write-back, write-allocate for page tables.
7579 	 */
7580 	pte_l1_s_cache_mode_pt = L1_S_C | L1_S_B | L1_S_V6_TEX(0x01);
7581 	pte_l2_l_cache_mode_pt = L2_C | L2_B | L2_V6_L_TEX(0x01);
7582 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7583 	pte_l2_s_cache_mode_pt = L2_C | L2_B | L2_V6_XS_TEX(0x01);
7584 #else
7585 	pte_l2_s_cache_mode_pt = L2_C | L2_B;
7586 #endif
7587 
7588 	pte_l1_s_prot_u = L1_S_PROT_U_armv6;
7589 	pte_l1_s_prot_w = L1_S_PROT_W_armv6;
7590 	pte_l1_s_prot_ro = L1_S_PROT_RO_armv6;
7591 	pte_l1_s_prot_mask = L1_S_PROT_MASK_armv6;
7592 
7593 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7594 	pte_l2_s_prot_u = L2_S_PROT_U_armv6n;
7595 	pte_l2_s_prot_w = L2_S_PROT_W_armv6n;
7596 	pte_l2_s_prot_ro = L2_S_PROT_RO_armv6n;
7597 	pte_l2_s_prot_mask = L2_S_PROT_MASK_armv6n;
7598 
7599 #else
7600 	/* with AP[0..3] */
7601 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
7602 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
7603 	pte_l2_s_prot_ro = L2_S_PROT_RO_generic;
7604 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
7605 #endif
7606 
7607 #ifdef	ARM11MPCORE_COMPAT_MMU
7608 	/* with AP[0..3] */
7609 	pte_l2_l_prot_u = L2_L_PROT_U_generic;
7610 	pte_l2_l_prot_w = L2_L_PROT_W_generic;
7611 	pte_l2_l_prot_ro = L2_L_PROT_RO_generic;
7612 	pte_l2_l_prot_mask = L2_L_PROT_MASK_generic;
7613 
7614 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
7615 	pte_l1_s_proto = L1_S_PROTO_armv6;
7616 	pte_l1_c_proto = L1_C_PROTO_armv6;
7617 	pte_l2_s_proto = L2_S_PROTO_armv6c;
7618 #else
7619 	pte_l2_l_prot_u = L2_L_PROT_U_armv6n;
7620 	pte_l2_l_prot_w = L2_L_PROT_W_armv6n;
7621 	pte_l2_l_prot_ro = L2_L_PROT_RO_armv6n;
7622 	pte_l2_l_prot_mask = L2_L_PROT_MASK_armv6n;
7623 
7624 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
7625 	pte_l1_s_proto = L1_S_PROTO_armv6;
7626 	pte_l1_c_proto = L1_C_PROTO_armv6;
7627 	pte_l2_s_proto = L2_S_PROTO_armv6n;
7628 #endif
7629 
7630 	pmap_copy_page_func = pmap_copy_page_generic;
7631 	pmap_zero_page_func = pmap_zero_page_generic;
7632 	pmap_needs_pte_sync = 1;
7633 }
7634 #endif	/* CPU_ARM11MPCORE */
7635 
7636 
7637 #if ARM_MMU_V6 == 1
7638 void
7639 pmap_pte_init_armv6(void)
7640 {
7641 	/*
7642 	 * The ARMv6-A MMU is mostly compatible with generic. If the
7643 	 * AP field is zero, that now means "no access" rather than
7644 	 * read-only. The prototypes are a little different because of
7645 	 * the XN bit.
7646 	 */
7647 	pmap_pte_init_generic();
7648 
7649 	pte_l1_s_nocache_mode = L1_S_XS_TEX(1);
7650 	pte_l2_l_nocache_mode = L2_XS_L_TEX(1);
7651 	pte_l2_s_nocache_mode = L2_XS_T_TEX(1);
7652 
7653 #ifdef ARM11_COMPAT_MMU
7654 	/* with AP[0..3] */
7655 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
7656 #else
7657 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_armv6n;
7658 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_armv6n;
7659 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_armv6n;
7660 
7661 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
7662 	pte_l1_s_proto = L1_S_PROTO_armv6;
7663 	pte_l1_c_proto = L1_C_PROTO_armv6;
7664 	pte_l2_s_proto = L2_S_PROTO_armv6n;
7665 
7666 	pte_l1_s_prot_u = L1_S_PROT_U_armv6;
7667 	pte_l1_s_prot_w = L1_S_PROT_W_armv6;
7668 	pte_l1_s_prot_ro = L1_S_PROT_RO_armv6;
7669 	pte_l1_s_prot_mask = L1_S_PROT_MASK_armv6;
7670 
7671 	pte_l2_l_prot_u = L2_L_PROT_U_armv6n;
7672 	pte_l2_l_prot_w = L2_L_PROT_W_armv6n;
7673 	pte_l2_l_prot_ro = L2_L_PROT_RO_armv6n;
7674 	pte_l2_l_prot_mask = L2_L_PROT_MASK_armv6n;
7675 
7676 	pte_l2_s_prot_u = L2_S_PROT_U_armv6n;
7677 	pte_l2_s_prot_w = L2_S_PROT_W_armv6n;
7678 	pte_l2_s_prot_ro = L2_S_PROT_RO_armv6n;
7679 	pte_l2_s_prot_mask = L2_S_PROT_MASK_armv6n;
7680 
7681 #endif
7682 }
7683 #endif /* ARM_MMU_V6 */
7684 
7685 #if ARM_MMU_V7 == 1
7686 void
7687 pmap_pte_init_armv7(void)
7688 {
7689 	/*
7690 	 * The ARMv7-A MMU is mostly compatible with generic. If the
7691 	 * AP field is zero, that now means "no access" rather than
7692 	 * read-only. The prototypes are a little different because of
7693 	 * the XN bit.
7694 	 */
7695 	pmap_pte_init_generic();
7696 
7697 	pmap_needs_pte_sync = 1;
7698 
7699 	pte_l1_s_nocache_mode = L1_S_XS_TEX(1);
7700 	pte_l2_l_nocache_mode = L2_XS_L_TEX(1);
7701 	pte_l2_s_nocache_mode = L2_XS_T_TEX(1);
7702 
7703 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_armv7;
7704 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_armv7;
7705 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_armv7;
7706 
7707 	/*
7708 	 * If the core support coherent walk then updates to translation tables
7709 	 * do not require a clean to the point of unification to ensure
7710 	 * visibility by subsequent translation table walks.  That means we can
7711 	 * map everything shareable and cached and the right thing will happen.
7712 	 */
7713         if (__SHIFTOUT(armreg_mmfr3_read(), __BITS(23,20))) {
7714 		pmap_needs_pte_sync = 0;
7715 
7716 		/*
7717 		 * write-back, no write-allocate, shareable for normal pages.
7718 		 */
7719 		pte_l1_s_cache_mode |= L1_S_V6_S;
7720 		pte_l2_l_cache_mode |= L2_XS_S;
7721 		pte_l2_s_cache_mode |= L2_XS_S;
7722 	}
7723 
7724 	/*
7725 	 * Page tables are just all other memory.  We can use write-back since
7726 	 * pmap_needs_pte_sync is 1 (or the MMU can read out of cache).
7727 	 */
7728 	pte_l1_s_cache_mode_pt = pte_l1_s_cache_mode;
7729 	pte_l2_l_cache_mode_pt = pte_l2_l_cache_mode;
7730 	pte_l2_s_cache_mode_pt = pte_l2_s_cache_mode;
7731 
7732 	/*
7733 	 * Check the Memory Model Features to see if this CPU supports
7734 	 * the TLBIASID coproc op.
7735 	 */
7736 	if (__SHIFTOUT(armreg_mmfr2_read(), __BITS(16,19)) >= 2) {
7737 		arm_has_tlbiasid_p = true;
7738 	} else if (__SHIFTOUT(armreg_mmfr2_read(), __BITS(12,15)) >= 2) {
7739 		arm_has_tlbiasid_p = true;
7740 	}
7741 
7742 	/*
7743 	 * Check the MPIDR to see if this CPU supports MP extensions.
7744 	 */
7745 #ifdef MULTIPROCESSOR
7746 	arm_has_mpext_p = (armreg_mpidr_read() & (MPIDR_MP|MPIDR_U)) == MPIDR_MP;
7747 #else
7748 	arm_has_mpext_p = false;
7749 #endif
7750 
7751 	pte_l1_s_prot_u = L1_S_PROT_U_armv7;
7752 	pte_l1_s_prot_w = L1_S_PROT_W_armv7;
7753 	pte_l1_s_prot_ro = L1_S_PROT_RO_armv7;
7754 	pte_l1_s_prot_mask = L1_S_PROT_MASK_armv7;
7755 
7756 	pte_l2_s_prot_u = L2_S_PROT_U_armv7;
7757 	pte_l2_s_prot_w = L2_S_PROT_W_armv7;
7758 	pte_l2_s_prot_ro = L2_S_PROT_RO_armv7;
7759 	pte_l2_s_prot_mask = L2_S_PROT_MASK_armv7;
7760 
7761 	pte_l2_l_prot_u = L2_L_PROT_U_armv7;
7762 	pte_l2_l_prot_w = L2_L_PROT_W_armv7;
7763 	pte_l2_l_prot_ro = L2_L_PROT_RO_armv7;
7764 	pte_l2_l_prot_mask = L2_L_PROT_MASK_armv7;
7765 
7766 	pte_l1_ss_proto = L1_SS_PROTO_armv7;
7767 	pte_l1_s_proto = L1_S_PROTO_armv7;
7768 	pte_l1_c_proto = L1_C_PROTO_armv7;
7769 	pte_l2_s_proto = L2_S_PROTO_armv7;
7770 
7771 }
7772 #endif /* ARM_MMU_V7 */
7773 
7774 /*
7775  * return the PA of the current L1 table, for use when handling a crash dump
7776  */
7777 uint32_t
7778 pmap_kernel_L1_addr(void)
7779 {
7780 #ifdef ARM_MMU_EXTENDED
7781 	return pmap_kernel()->pm_l1_pa;
7782 #else
7783 	return pmap_kernel()->pm_l1->l1_physaddr;
7784 #endif
7785 }
7786 
7787 #if defined(DDB)
7788 /*
7789  * A couple of ddb-callable functions for dumping pmaps
7790  */
7791 void pmap_dump(pmap_t);
7792 
7793 static pt_entry_t ncptes[64];
7794 static void pmap_dump_ncpg(pmap_t);
7795 
7796 void
7797 pmap_dump(pmap_t pm)
7798 {
7799 	struct l2_dtable *l2;
7800 	struct l2_bucket *l2b;
7801 	pt_entry_t *ptep, pte;
7802 	vaddr_t l2_va, l2b_va, va;
7803 	int i, j, k, occ, rows = 0;
7804 
7805 	if (pm == pmap_kernel())
7806 		printf("pmap_kernel (%p): ", pm);
7807 	else
7808 		printf("user pmap (%p): ", pm);
7809 
7810 #ifdef ARM_MMU_EXTENDED
7811 	printf("l1 at %p\n", pmap_l1_kva(pm));
7812 #else
7813 	printf("domain %d, l1 at %p\n", pmap_domain(pm), pmap_l1_kva(pm));
7814 #endif
7815 
7816 	l2_va = 0;
7817 	for (i = 0; i < L2_SIZE; i++, l2_va += 0x01000000) {
7818 		l2 = pm->pm_l2[i];
7819 
7820 		if (l2 == NULL || l2->l2_occupancy == 0)
7821 			continue;
7822 
7823 		l2b_va = l2_va;
7824 		for (j = 0; j < L2_BUCKET_SIZE; j++, l2b_va += 0x00100000) {
7825 			l2b = &l2->l2_bucket[j];
7826 
7827 			if (l2b->l2b_occupancy == 0 || l2b->l2b_kva == NULL)
7828 				continue;
7829 
7830 			ptep = l2b->l2b_kva;
7831 
7832 			for (k = 0; k < 256 && ptep[k] == 0; k++)
7833 				;
7834 
7835 			k &= ~63;
7836 			occ = l2b->l2b_occupancy;
7837 			va = l2b_va + (k * 4096);
7838 			for (; k < 256; k++, va += 0x1000) {
7839 				char ch = ' ';
7840 				if ((k % 64) == 0) {
7841 					if ((rows % 8) == 0) {
7842 						printf(
7843 "          |0000   |8000   |10000  |18000  |20000  |28000  |30000  |38000\n");
7844 					}
7845 					printf("%08lx: ", va);
7846 				}
7847 
7848 				ncptes[k & 63] = 0;
7849 				pte = ptep[k];
7850 				if (pte == 0) {
7851 					ch = '.';
7852 				} else {
7853 					occ--;
7854 					switch (pte & 0x4c) {
7855 					case 0x00:
7856 						ch = 'N'; /* No cache No buff */
7857 						break;
7858 					case 0x04:
7859 						ch = 'B'; /* No cache buff */
7860 						break;
7861 					case 0x08:
7862 						ch = 'C'; /* Cache No buff */
7863 						break;
7864 					case 0x0c:
7865 						ch = 'F'; /* Cache Buff */
7866 						break;
7867 					case 0x40:
7868 						ch = 'D';
7869 						break;
7870 					case 0x48:
7871 						ch = 'm'; /* Xscale mini-data */
7872 						break;
7873 					default:
7874 						ch = '?';
7875 						break;
7876 					}
7877 
7878 					if ((pte & L2_S_PROT_U) == L2_S_PROT_U)
7879 						ch += 0x20;
7880 
7881 					if ((pte & 0xc) == 0)
7882 						ncptes[k & 63] = pte;
7883 				}
7884 
7885 				if ((k % 64) == 63) {
7886 					rows++;
7887 					printf("%c\n", ch);
7888 					pmap_dump_ncpg(pm);
7889 					if (occ == 0)
7890 						break;
7891 				} else
7892 					printf("%c", ch);
7893 			}
7894 		}
7895 	}
7896 }
7897 
7898 static void
7899 pmap_dump_ncpg(pmap_t pm)
7900 {
7901 	struct vm_page *pg;
7902 	struct vm_page_md *md;
7903 	struct pv_entry *pv;
7904 	int i;
7905 
7906 	for (i = 0; i < 63; i++) {
7907 		if (ncptes[i] == 0)
7908 			continue;
7909 
7910 		pg = PHYS_TO_VM_PAGE(l2pte_pa(ncptes[i]));
7911 		if (pg == NULL)
7912 			continue;
7913 		md = VM_PAGE_TO_MD(pg);
7914 
7915 		printf(" pa 0x%08lx: krw %d kro %d urw %d uro %d\n",
7916 		    VM_PAGE_TO_PHYS(pg),
7917 		    md->krw_mappings, md->kro_mappings,
7918 		    md->urw_mappings, md->uro_mappings);
7919 
7920 		SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
7921 			printf("   %c va 0x%08lx, flags 0x%x\n",
7922 			    (pm == pv->pv_pmap) ? '*' : ' ',
7923 			    pv->pv_va, pv->pv_flags);
7924 		}
7925 	}
7926 }
7927 #endif
7928 
7929 #ifdef PMAP_STEAL_MEMORY
7930 void
7931 pmap_boot_pageadd(pv_addr_t *newpv)
7932 {
7933 	pv_addr_t *pv, *npv;
7934 
7935 	if ((pv = SLIST_FIRST(&pmap_boot_freeq)) != NULL) {
7936 		if (newpv->pv_pa < pv->pv_va) {
7937 			KASSERT(newpv->pv_pa + newpv->pv_size <= pv->pv_pa);
7938 			if (newpv->pv_pa + newpv->pv_size == pv->pv_pa) {
7939 				newpv->pv_size += pv->pv_size;
7940 				SLIST_REMOVE_HEAD(&pmap_boot_freeq, pv_list);
7941 			}
7942 			pv = NULL;
7943 		} else {
7944 			for (; (npv = SLIST_NEXT(pv, pv_list)) != NULL;
7945 			     pv = npv) {
7946 				KASSERT(pv->pv_pa + pv->pv_size < npv->pv_pa);
7947 				KASSERT(pv->pv_pa < newpv->pv_pa);
7948 				if (newpv->pv_pa > npv->pv_pa)
7949 					continue;
7950 				if (pv->pv_pa + pv->pv_size == newpv->pv_pa) {
7951 					pv->pv_size += newpv->pv_size;
7952 					return;
7953 				}
7954 				if (newpv->pv_pa + newpv->pv_size < npv->pv_pa)
7955 					break;
7956 				newpv->pv_size += npv->pv_size;
7957 				SLIST_INSERT_AFTER(pv, newpv, pv_list);
7958 				SLIST_REMOVE_AFTER(newpv, pv_list);
7959 				return;
7960 			}
7961 		}
7962 	}
7963 
7964 	if (pv) {
7965 		SLIST_INSERT_AFTER(pv, newpv, pv_list);
7966 	} else {
7967 		SLIST_INSERT_HEAD(&pmap_boot_freeq, newpv, pv_list);
7968 	}
7969 }
7970 
7971 void
7972 pmap_boot_pagealloc(psize_t amount, psize_t mask, psize_t match,
7973 	pv_addr_t *rpv)
7974 {
7975 	pv_addr_t *pv, **pvp;
7976 
7977 	KASSERT(amount & PGOFSET);
7978 	KASSERT((mask & PGOFSET) == 0);
7979 	KASSERT((match & PGOFSET) == 0);
7980 	KASSERT(amount != 0);
7981 
7982 	for (pvp = &SLIST_FIRST(&pmap_boot_freeq);
7983 	     (pv = *pvp) != NULL;
7984 	     pvp = &SLIST_NEXT(pv, pv_list)) {
7985 		pv_addr_t *newpv;
7986 		psize_t off;
7987 		/*
7988 		 * If this entry is too small to satisfy the request...
7989 		 */
7990 		KASSERT(pv->pv_size > 0);
7991 		if (pv->pv_size < amount)
7992 			continue;
7993 
7994 		for (off = 0; off <= mask; off += PAGE_SIZE) {
7995 			if (((pv->pv_pa + off) & mask) == match
7996 			    && off + amount <= pv->pv_size)
7997 				break;
7998 		}
7999 		if (off > mask)
8000 			continue;
8001 
8002 		rpv->pv_va = pv->pv_va + off;
8003 		rpv->pv_pa = pv->pv_pa + off;
8004 		rpv->pv_size = amount;
8005 		pv->pv_size -= amount;
8006 		if (pv->pv_size == 0) {
8007 			KASSERT(off == 0);
8008 			KASSERT((vaddr_t) pv == rpv->pv_va);
8009 			*pvp = SLIST_NEXT(pv, pv_list);
8010 		} else if (off == 0) {
8011 			KASSERT((vaddr_t) pv == rpv->pv_va);
8012 			newpv = (pv_addr_t *) (rpv->pv_va + amount);
8013 			*newpv = *pv;
8014 			newpv->pv_pa += amount;
8015 			newpv->pv_va += amount;
8016 			*pvp = newpv;
8017 		} else if (off < pv->pv_size) {
8018 			newpv = (pv_addr_t *) (rpv->pv_va + amount);
8019 			*newpv = *pv;
8020 			newpv->pv_size -= off;
8021 			newpv->pv_pa += off + amount;
8022 			newpv->pv_va += off + amount;
8023 
8024 			SLIST_NEXT(pv, pv_list) = newpv;
8025 			pv->pv_size = off;
8026 		} else {
8027 			KASSERT((vaddr_t) pv != rpv->pv_va);
8028 		}
8029 		memset((void *)rpv->pv_va, 0, amount);
8030 		return;
8031 	}
8032 
8033 	if (!uvm_physseg_valid_p(uvm_physseg_get_first()))
8034 		panic("pmap_boot_pagealloc: couldn't allocate memory");
8035 
8036 	for (pvp = &SLIST_FIRST(&pmap_boot_freeq);
8037 	     (pv = *pvp) != NULL;
8038 	     pvp = &SLIST_NEXT(pv, pv_list)) {
8039 		if (SLIST_NEXT(pv, pv_list) == NULL)
8040 			break;
8041 	}
8042 	KASSERT(mask == 0);
8043 
8044 	for (uvm_physseg_t ups = uvm_physseg_get_first();
8045 	    uvm_physseg_valid_p(ups);
8046 	    ups = uvm_physseg_get_next(ups)) {
8047 
8048 		paddr_t spn = uvm_physseg_get_start(ups);
8049 		paddr_t epn = uvm_physseg_get_end(ups);
8050 		if (spn == atop(pv->pv_pa + pv->pv_size)
8051 		    && pv->pv_va + pv->pv_size <= ptoa(epn)) {
8052 			rpv->pv_va = pv->pv_va;
8053 			rpv->pv_pa = pv->pv_pa;
8054 			rpv->pv_size = amount;
8055 			*pvp = NULL;
8056 			pmap_map_chunk(kernel_l1pt.pv_va,
8057 			     ptoa(spn) + (pv->pv_va - pv->pv_pa),
8058 			     ptoa(spn),
8059 			     amount - pv->pv_size,
8060 			     VM_PROT_READ|VM_PROT_WRITE,
8061 			     PTE_CACHE);
8062 
8063 			uvm_physseg_unplug(spn, atop(amount - pv->pv_size));
8064 			memset((void *)rpv->pv_va, 0, rpv->pv_size);
8065 			return;
8066 		}
8067 	}
8068 
8069 	panic("pmap_boot_pagealloc: couldn't allocate memory");
8070 }
8071 
8072 vaddr_t
8073 pmap_steal_memory(vsize_t size, vaddr_t *vstartp, vaddr_t *vendp)
8074 {
8075 	pv_addr_t pv;
8076 
8077 	pmap_boot_pagealloc(size, 0, 0, &pv);
8078 
8079 	return pv.pv_va;
8080 }
8081 #endif /* PMAP_STEAL_MEMORY */
8082 
8083 SYSCTL_SETUP(sysctl_machdep_pmap_setup, "sysctl machdep.kmpages setup")
8084 {
8085 	sysctl_createv(clog, 0, NULL, NULL,
8086 			CTLFLAG_PERMANENT,
8087 			CTLTYPE_NODE, "machdep", NULL,
8088 			NULL, 0, NULL, 0,
8089 			CTL_MACHDEP, CTL_EOL);
8090 
8091 	sysctl_createv(clog, 0, NULL, NULL,
8092 			CTLFLAG_PERMANENT,
8093 			CTLTYPE_INT, "kmpages",
8094 			SYSCTL_DESCR("count of pages allocated to kernel memory allocators"),
8095 			NULL, 0, &pmap_kmpages, 0,
8096 			CTL_MACHDEP, CTL_CREATE, CTL_EOL);
8097 }
8098 
8099 #ifdef PMAP_NEED_ALLOC_POOLPAGE
8100 struct vm_page *
8101 arm_pmap_alloc_poolpage(int flags)
8102 {
8103 	/*
8104 	 * On some systems, only some pages may be "coherent" for dma and we
8105 	 * want to prefer those for pool pages (think mbufs) but fallback to
8106 	 * any page if none is available.
8107 	 */
8108 	if (arm_poolpage_vmfreelist != VM_FREELIST_DEFAULT) {
8109 		return uvm_pagealloc_strat(NULL, 0, NULL, flags,
8110 		    UVM_PGA_STRAT_FALLBACK, arm_poolpage_vmfreelist);
8111 	}
8112 
8113 	return uvm_pagealloc(NULL, 0, NULL, flags);
8114 }
8115 #endif
8116 
8117 #if defined(ARM_MMU_EXTENDED) && defined(MULTIPROCESSOR)
8118 void
8119 pmap_md_tlb_info_attach(struct pmap_tlb_info *ti, struct cpu_info *ci)
8120 {
8121         /* nothing */
8122 }
8123 
8124 int
8125 pic_ipi_shootdown(void *arg)
8126 {
8127 #if PMAP_TLB_NEED_SHOOTDOWN
8128 	pmap_tlb_shootdown_process();
8129 #endif
8130 	return 1;
8131 }
8132 #endif /* ARM_MMU_EXTENDED && MULTIPROCESSOR */
8133 
8134 
8135 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
8136 vaddr_t
8137 pmap_direct_mapped_phys(paddr_t pa, bool *ok_p, vaddr_t va)
8138 {
8139 	bool ok = false;
8140 	if (physical_start <= pa && pa < physical_end) {
8141 #ifdef KERNEL_BASE_VOFFSET
8142 		const vaddr_t newva = pa + KERNEL_BASE_VOFFSET;
8143 #else
8144 		const vaddr_t newva = KERNEL_BASE + pa - physical_start;
8145 #endif
8146 #ifdef ARM_MMU_EXTENDED
8147 		if (newva >= KERNEL_BASE && newva < pmap_directlimit) {
8148 #endif
8149 			va = newva;
8150 			ok = true;
8151 #ifdef ARM_MMU_EXTENDED
8152 		}
8153 #endif
8154 	}
8155 	KASSERT(ok_p);
8156 	*ok_p = ok;
8157 	return va;
8158 }
8159 
8160 vaddr_t
8161 pmap_map_poolpage(paddr_t pa)
8162 {
8163 	bool ok __diagused;
8164 	vaddr_t va = pmap_direct_mapped_phys(pa, &ok, 0);
8165 	KASSERTMSG(ok, "pa %#lx not direct mappable", pa);
8166 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
8167 	if (arm_cache_prefer_mask != 0) {
8168 		struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
8169 		struct vm_page_md * const md = VM_PAGE_TO_MD(pg);
8170 		pmap_acquire_page_lock(md);
8171 		pmap_vac_me_harder(md, pa, pmap_kernel(), va);
8172 		pmap_release_page_lock(md);
8173 	}
8174 #endif
8175 	return va;
8176 }
8177 
8178 paddr_t
8179 pmap_unmap_poolpage(vaddr_t va)
8180 {
8181 	KASSERT(va >= KERNEL_BASE);
8182 #ifdef PMAP_CACHE_VIVT
8183 	cpu_idcache_wbinv_range(va, PAGE_SIZE);
8184 #endif
8185 #if defined(KERNEL_BASE_VOFFSET)
8186         return va - KERNEL_BASE_VOFFSET;
8187 #else
8188         return va - KERNEL_BASE + physical_start;
8189 #endif
8190 }
8191 #endif /* __HAVE_MM_MD_DIRECT_MAPPED_PHYS */
8192