xref: /netbsd-src/sys/arch/arm/arm32/pmap.c (revision 82d56013d7b633d116a93943de88e08335357a7c)
1 /*	$NetBSD: pmap.c,v 1.428 2021/03/23 10:21:49 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 1.428 2021/03/23 10:21:49 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 	dsb(sy);
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 appropriate
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 	kpreempt_disable();
2841 	pmap_acquire_page_lock(md);
2842 	struct pv_entry **pvp = &SLIST_FIRST(&md->pvh_list);
2843 	if (*pvp == NULL) {
2844 #ifdef PMAP_CACHE_VIPT
2845 		/*
2846 		 * We *know* the page contents are about to be replaced.
2847 		 * Discard the exec contents
2848 		 */
2849 		if (PV_IS_EXEC_P(md->pvh_attrs))
2850 			PMAPCOUNT(exec_discarded_page_protect);
2851 		md->pvh_attrs &= ~PVF_EXEC;
2852 		PMAP_VALIDATE_MD_PAGE(md);
2853 #endif
2854 		pmap_release_page_lock(md);
2855 		kpreempt_enable();
2856 
2857 		return;
2858 	}
2859 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
2860 	KASSERT(arm_cache_prefer_mask == 0 || pmap_is_page_colored_p(md));
2861 #endif
2862 
2863 	/*
2864 	 * Clear alias counts
2865 	 */
2866 #ifdef PMAP_CACHE_VIVT
2867 	md->k_mappings = 0;
2868 #endif
2869 	md->urw_mappings = md->uro_mappings = 0;
2870 
2871 #ifdef PMAP_CACHE_VIVT
2872 	pmap_clean_page(md, false);
2873 #endif
2874 
2875 	for (pv = *pvp; pv != NULL;) {
2876 		pmap_t pm = pv->pv_pmap;
2877 #ifndef ARM_MMU_EXTENDED
2878 		if (flush == false && pmap_is_current(pm))
2879 			flush = true;
2880 #endif
2881 
2882 #ifdef PMAP_CACHE_VIPT
2883 		if (pm == pmap_kernel() && PV_IS_KENTRY_P(pv->pv_flags)) {
2884 			/* If this was unmanaged mapping, it must be ignored. */
2885 			pvp = &SLIST_NEXT(pv, pv_link);
2886 			pv = *pvp;
2887 			continue;
2888 		}
2889 #endif
2890 
2891 		/*
2892 		 * Try to get a hold on the pmap's lock.  We must do this
2893 		 * while still holding the page locked, to know that the
2894 		 * page is still associated with the pmap and the mapping is
2895 		 * in place.  If a hold can't be had, unlock and wait for
2896 		 * the pmap's lock to become available and retry.  The pmap
2897 		 * must be ref'd over this dance to stop it disappearing
2898 		 * behind us.
2899 		 */
2900 		if (!mutex_tryenter(&pm->pm_lock)) {
2901 			pmap_reference(pm);
2902 			pmap_release_page_lock(md);
2903 			pmap_acquire_pmap_lock(pm);
2904 			/* nothing, just wait for it */
2905 			pmap_release_pmap_lock(pm);
2906 			pmap_destroy(pm);
2907 			/* Restart from the beginning. */
2908 			pmap_acquire_page_lock(md);
2909 			pvp = &SLIST_FIRST(&md->pvh_list);
2910 			pv = *pvp;
2911 			continue;
2912 		}
2913 
2914 		if (pm == pmap_kernel()) {
2915 #ifdef PMAP_CACHE_VIPT
2916 			if (pv->pv_flags & PVF_WRITE)
2917 				md->krw_mappings--;
2918 			else
2919 				md->kro_mappings--;
2920 #endif
2921 			PMAPCOUNT(kernel_unmappings);
2922 		}
2923 		*pvp = SLIST_NEXT(pv, pv_link); /* remove from list */
2924 		PMAPCOUNT(unmappings);
2925 
2926 		pmap_release_page_lock(md);
2927 
2928 		l2b = pmap_get_l2_bucket(pm, pv->pv_va);
2929 		KASSERTMSG(l2b != NULL, "%#lx", pv->pv_va);
2930 
2931 		ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2932 
2933 		/*
2934 		 * Update statistics
2935 		 */
2936 		--pm->pm_stats.resident_count;
2937 
2938 		/* Wired bit */
2939 		if (pv->pv_flags & PVF_WIRED)
2940 			--pm->pm_stats.wired_count;
2941 
2942 		flags |= pv->pv_flags;
2943 
2944 		/*
2945 		 * Invalidate the PTEs.
2946 		 */
2947 		l2pte_reset(ptep);
2948 		PTE_SYNC_CURRENT(pm, ptep);
2949 
2950 #ifdef ARM_MMU_EXTENDED
2951 		pmap_tlb_invalidate_addr(pm, pv->pv_va);
2952 #endif
2953 
2954 		pmap_free_l2_bucket(pm, l2b, PAGE_SIZE / L2_S_SIZE);
2955 
2956 		pmap_release_pmap_lock(pm);
2957 
2958 		pool_put(&pmap_pv_pool, pv);
2959 		pmap_acquire_page_lock(md);
2960 
2961 		/*
2962 		 * Restart at the beginning of the list.
2963 		 */
2964 		pvp = &SLIST_FIRST(&md->pvh_list);
2965 		pv = *pvp;
2966 	}
2967 	/*
2968 	 * if we reach the end of the list and there are still mappings, they
2969 	 * might be able to be cached now.  And they must be kernel mappings.
2970 	 */
2971 	if (!SLIST_EMPTY(&md->pvh_list)) {
2972 		pmap_vac_me_harder(md, pa, pmap_kernel(), 0);
2973 	}
2974 
2975 #ifdef PMAP_CACHE_VIPT
2976 	/*
2977 	 * Its EXEC cache is now gone.
2978 	 */
2979 	if (PV_IS_EXEC_P(md->pvh_attrs))
2980 		PMAPCOUNT(exec_discarded_page_protect);
2981 	md->pvh_attrs &= ~PVF_EXEC;
2982 	KASSERT(md->urw_mappings == 0);
2983 	KASSERT(md->uro_mappings == 0);
2984 #ifndef ARM_MMU_EXTENDED
2985 	if (arm_cache_prefer_mask != 0) {
2986 		if (md->krw_mappings == 0)
2987 			md->pvh_attrs &= ~PVF_WRITE;
2988 		PMAP_VALIDATE_MD_PAGE(md);
2989 	}
2990 #endif /* ARM_MMU_EXTENDED */
2991 #endif /* PMAP_CACHE_VIPT */
2992 	pmap_release_page_lock(md);
2993 
2994 #ifndef ARM_MMU_EXTENDED
2995 	if (flush) {
2996 		/*
2997 		 * Note: We can't use pmap_tlb_flush{I,D}() here since that
2998 		 * would need a subsequent call to pmap_update() to ensure
2999 		 * curpm->pm_cstate.cs_all is reset. Our callers are not
3000 		 * required to do that (see pmap(9)), so we can't modify
3001 		 * the current pmap's state.
3002 		 */
3003 		if (PV_BEEN_EXECD(flags))
3004 			cpu_tlb_flushID();
3005 		else
3006 			cpu_tlb_flushD();
3007 	}
3008 	cpu_cpwait();
3009 #endif /* ARM_MMU_EXTENDED */
3010 
3011 	kpreempt_enable();
3012 }
3013 
3014 /*
3015  * pmap_t pmap_create(void)
3016  *
3017  *      Create a new pmap structure from scratch.
3018  */
3019 pmap_t
3020 pmap_create(void)
3021 {
3022 	pmap_t pm;
3023 
3024 	pm = pool_cache_get(&pmap_cache, PR_WAITOK);
3025 
3026 	mutex_init(&pm->pm_lock, MUTEX_DEFAULT, IPL_NONE);
3027 
3028 	pm->pm_refs = 1;
3029 	pm->pm_stats.wired_count = 0;
3030 	pm->pm_stats.resident_count = 1;
3031 #ifdef ARM_MMU_EXTENDED
3032 #ifdef MULTIPROCESSOR
3033 	kcpuset_create(&pm->pm_active, true);
3034 	kcpuset_create(&pm->pm_onproc, true);
3035 #endif
3036 #else
3037 	pm->pm_cstate.cs_all = 0;
3038 #endif
3039 	pmap_alloc_l1(pm);
3040 
3041 	/*
3042 	 * Note: The pool cache ensures that the pm_l2[] array is already
3043 	 * initialised to zero.
3044 	 */
3045 
3046 	pmap_pinit(pm);
3047 
3048 	return pm;
3049 }
3050 
3051 u_int
3052 arm32_mmap_flags(paddr_t pa)
3053 {
3054 	/*
3055 	 * the upper 8 bits in pmap_enter()'s flags are reserved for MD stuff
3056 	 * and we're using the upper bits in page numbers to pass flags around
3057 	 * so we might as well use the same bits
3058 	 */
3059 	return (u_int)pa & PMAP_MD_MASK;
3060 }
3061 /*
3062  * int pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
3063  *      u_int flags)
3064  *
3065  *      Insert the given physical page (p) at
3066  *      the specified virtual address (v) in the
3067  *      target physical map with the protection requested.
3068  *
3069  *      NB:  This is the only routine which MAY NOT lazy-evaluate
3070  *      or lose information.  That is, this routine must actually
3071  *      insert this page into the given map NOW.
3072  */
3073 int
3074 pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags)
3075 {
3076 	struct l2_bucket *l2b;
3077 	struct vm_page *pg, *opg;
3078 	u_int nflags;
3079 	u_int oflags;
3080 	const bool kpm_p = (pm == pmap_kernel());
3081 #ifdef ARM_HAS_VBAR
3082 	const bool vector_page_p = false;
3083 #else
3084 	const bool vector_page_p = (va == vector_page);
3085 #endif
3086 	struct pmap_page *pp = pmap_pv_tracked(pa);
3087 	struct pv_entry *new_pv = NULL;
3088 	struct pv_entry *old_pv = NULL;
3089 	int error = 0;
3090 
3091 	UVMHIST_FUNC(__func__);
3092 	UVMHIST_CALLARGS(maphist, "pm %#jx va %#jx pa %#jx prot %#jx",
3093 	    (uintptr_t)pm, va, pa, prot);
3094 	UVMHIST_LOG(maphist, "  flag %#jx", flags, 0, 0, 0);
3095 
3096 	KDASSERT((flags & PMAP_WIRED) == 0 || (flags & VM_PROT_ALL) != 0);
3097 	KDASSERT(((va | pa) & PGOFSET) == 0);
3098 
3099 	/*
3100 	 * Get a pointer to the page.  Later on in this function, we
3101 	 * test for a managed page by checking pg != NULL.
3102 	 */
3103 	pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
3104 	/*
3105 	 * if we may need a new pv entry allocate if now, as we can't do it
3106 	 * with the kernel_pmap locked
3107 	 */
3108 	if (pg || pp)
3109 		new_pv = pool_get(&pmap_pv_pool, PR_NOWAIT);
3110 
3111 	nflags = 0;
3112 	if (prot & VM_PROT_WRITE)
3113 		nflags |= PVF_WRITE;
3114 	if (prot & VM_PROT_EXECUTE)
3115 		nflags |= PVF_EXEC;
3116 	if (flags & PMAP_WIRED)
3117 		nflags |= PVF_WIRED;
3118 
3119 	kpreempt_disable();
3120 	pmap_acquire_pmap_lock(pm);
3121 
3122 	/*
3123 	 * Fetch the L2 bucket which maps this page, allocating one if
3124 	 * necessary for user pmaps.
3125 	 */
3126 	if (kpm_p) {
3127 		l2b = pmap_get_l2_bucket(pm, va);
3128 	} else {
3129 		l2b = pmap_alloc_l2_bucket(pm, va);
3130 	}
3131 	if (l2b == NULL) {
3132 		if (flags & PMAP_CANFAIL) {
3133 			pmap_release_pmap_lock(pm);
3134 			kpreempt_enable();
3135 
3136 			error = ENOMEM;
3137 			goto free_pv;
3138 		}
3139 		panic("pmap_enter: failed to allocate L2 bucket");
3140 	}
3141 	pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(va)];
3142 	const pt_entry_t opte = *ptep;
3143 	pt_entry_t npte = pa;
3144 	oflags = 0;
3145 
3146 	if (opte) {
3147 		/*
3148 		 * There is already a mapping at this address.
3149 		 * If the physical address is different, lookup the
3150 		 * vm_page.
3151 		 */
3152 		if (l2pte_pa(opte) != pa) {
3153 			KASSERT(!pmap_pv_tracked(pa));
3154 			opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3155 		} else
3156 			opg = pg;
3157 	} else
3158 		opg = NULL;
3159 
3160 	if (pg || pp) {
3161 		KASSERT((pg != NULL) != (pp != NULL));
3162 		struct vm_page_md *md = (pg != NULL) ? VM_PAGE_TO_MD(pg) :
3163 		    PMAP_PAGE_TO_MD(pp);
3164 
3165 		UVMHIST_LOG(maphist, "  pg %#jx pp %#jx pvh_attrs %#jx "
3166 		    "nflags %#jx", (uintptr_t)pg, (uintptr_t)pp,
3167 		    md->pvh_attrs, nflags);
3168 
3169 		/*
3170 		 * This is to be a managed mapping.
3171 		 */
3172 		pmap_acquire_page_lock(md);
3173 		if ((flags & VM_PROT_ALL) || (md->pvh_attrs & PVF_REF)) {
3174 			/*
3175 			 * - The access type indicates that we don't need
3176 			 *   to do referenced emulation.
3177 			 * OR
3178 			 * - The physical page has already been referenced
3179 			 *   so no need to re-do referenced emulation here.
3180 			 */
3181 			npte |= l2pte_set_readonly(L2_S_PROTO);
3182 
3183 			nflags |= PVF_REF;
3184 
3185 			if ((prot & VM_PROT_WRITE) != 0 &&
3186 			    ((flags & VM_PROT_WRITE) != 0 ||
3187 			     (md->pvh_attrs & PVF_MOD) != 0)) {
3188 				/*
3189 				 * This is a writable mapping, and the
3190 				 * page's mod state indicates it has
3191 				 * already been modified. Make it
3192 				 * writable from the outset.
3193 				 */
3194 				npte = l2pte_set_writable(npte);
3195 				nflags |= PVF_MOD;
3196 			}
3197 
3198 #ifdef ARM_MMU_EXTENDED
3199 			/*
3200 			 * If the page has been cleaned, then the pvh_attrs
3201 			 * will have PVF_EXEC set, so mark it execute so we
3202 			 * don't get an access fault when trying to execute
3203 			 * from it.
3204 			 */
3205 			if (md->pvh_attrs & nflags & PVF_EXEC) {
3206 				npte &= ~L2_XS_XN;
3207 			}
3208 #endif
3209 		} else {
3210 			/*
3211 			 * Need to do page referenced emulation.
3212 			 */
3213 			npte |= L2_TYPE_INV;
3214 		}
3215 
3216 		if (flags & ARM32_MMAP_WRITECOMBINE) {
3217 			npte |= pte_l2_s_wc_mode;
3218 		} else
3219 			npte |= pte_l2_s_cache_mode;
3220 
3221 		if (pg != NULL && pg == opg) {
3222 			/*
3223 			 * We're changing the attrs of an existing mapping.
3224 			 */
3225 			oflags = pmap_modify_pv(md, pa, pm, va,
3226 			    PVF_WRITE | PVF_EXEC | PVF_WIRED |
3227 			    PVF_MOD | PVF_REF, nflags);
3228 
3229 #ifdef PMAP_CACHE_VIVT
3230 			/*
3231 			 * We may need to flush the cache if we're
3232 			 * doing rw-ro...
3233 			 */
3234 			if (pm->pm_cstate.cs_cache_d &&
3235 			    (oflags & PVF_NC) == 0 &&
3236 			    l2pte_writable_p(opte) &&
3237 			    (prot & VM_PROT_WRITE) == 0)
3238 				cpu_dcache_wb_range(va, PAGE_SIZE);
3239 #endif
3240 		} else {
3241 			struct pv_entry *pv;
3242 			/*
3243 			 * New mapping, or changing the backing page
3244 			 * of an existing mapping.
3245 			 */
3246 			if (opg) {
3247 				struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3248 				paddr_t opa = VM_PAGE_TO_PHYS(opg);
3249 
3250 				/*
3251 				 * Replacing an existing mapping with a new one.
3252 				 * It is part of our managed memory so we
3253 				 * must remove it from the PV list
3254 				 */
3255 				pv = pmap_remove_pv(omd, opa, pm, va);
3256 				pmap_vac_me_harder(omd, opa, pm, 0);
3257 				oflags = pv->pv_flags;
3258 
3259 #ifdef PMAP_CACHE_VIVT
3260 				/*
3261 				 * If the old mapping was valid (ref/mod
3262 				 * emulation creates 'invalid' mappings
3263 				 * initially) then make sure to frob
3264 				 * the cache.
3265 				 */
3266 				if (!(oflags & PVF_NC) && l2pte_valid_p(opte)) {
3267 					pmap_cache_wbinv_page(pm, va, true,
3268 					    oflags);
3269 				}
3270 #endif
3271 			} else {
3272 				pv = new_pv;
3273 				new_pv = NULL;
3274 				if (pv == NULL) {
3275 					pmap_release_page_lock(md);
3276 					pmap_release_pmap_lock(pm);
3277 					if ((flags & PMAP_CANFAIL) == 0)
3278 						panic("pmap_enter: "
3279 						    "no pv entries");
3280 
3281 					pmap_free_l2_bucket(pm, l2b, 0);
3282 					UVMHIST_LOG(maphist, "  <-- done (ENOMEM)",
3283 					    0, 0, 0, 0);
3284 					kpreempt_enable();
3285 
3286 					return ENOMEM;
3287 				}
3288 			}
3289 
3290 			pmap_enter_pv(md, pa, pv, pm, va, nflags);
3291 		}
3292 		pmap_release_page_lock(md);
3293 	} else {
3294 		/*
3295 		 * We're mapping an unmanaged page.
3296 		 * These are always readable, and possibly writable, from
3297 		 * the get go as we don't need to track ref/mod status.
3298 		 */
3299 		npte |= l2pte_set_readonly(L2_S_PROTO);
3300 		if (prot & VM_PROT_WRITE)
3301 			npte = l2pte_set_writable(npte);
3302 
3303 		/*
3304 		 * Make sure the vector table is mapped cacheable
3305 		 */
3306 		if ((vector_page_p && !kpm_p)
3307 		    || (flags & ARM32_MMAP_CACHEABLE)) {
3308 			npte |= pte_l2_s_cache_mode;
3309 #ifdef ARM_MMU_EXTENDED
3310 			npte &= ~L2_XS_XN;	/* and executable */
3311 #endif
3312 		} else if (flags & ARM32_MMAP_WRITECOMBINE) {
3313 			npte |= pte_l2_s_wc_mode;
3314 		}
3315 		if (opg) {
3316 			/*
3317 			 * Looks like there's an existing 'managed' mapping
3318 			 * at this address.
3319 			 */
3320 			struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3321 			paddr_t opa = VM_PAGE_TO_PHYS(opg);
3322 
3323 			pmap_acquire_page_lock(omd);
3324 			old_pv = pmap_remove_pv(omd, opa, pm, va);
3325 			pmap_vac_me_harder(omd, opa, pm, 0);
3326 			oflags = old_pv->pv_flags;
3327 			pmap_release_page_lock(omd);
3328 
3329 #ifdef PMAP_CACHE_VIVT
3330 			if (!(oflags & PVF_NC) && l2pte_valid_p(opte)) {
3331 				pmap_cache_wbinv_page(pm, va, true, oflags);
3332 			}
3333 #endif
3334 		}
3335 	}
3336 
3337 	/*
3338 	 * Make sure userland mappings get the right permissions
3339 	 */
3340 	if (!vector_page_p && !kpm_p) {
3341 		npte |= L2_S_PROT_U;
3342 #ifdef ARM_MMU_EXTENDED
3343 		npte |= L2_XS_nG;	/* user pages are not global */
3344 #endif
3345 	}
3346 
3347 	/*
3348 	 * Keep the stats up to date
3349 	 */
3350 	if (opte == 0) {
3351 		l2b->l2b_occupancy += PAGE_SIZE / L2_S_SIZE;
3352 		pm->pm_stats.resident_count++;
3353 	}
3354 
3355 	UVMHIST_LOG(maphist, " opte %#jx npte %#jx", opte, npte, 0, 0);
3356 
3357 #if defined(ARM_MMU_EXTENDED)
3358 	/*
3359 	 * If exec protection was requested but the page hasn't been synced,
3360 	 * sync it now and allow execution from it.
3361 	 */
3362 	if ((nflags & PVF_EXEC) && (npte & L2_XS_XN)) {
3363 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3364 		npte &= ~L2_XS_XN;
3365 		pmap_syncicache_page(md, pa);
3366 		PMAPCOUNT(exec_synced_map);
3367 	}
3368 #endif
3369 	/*
3370 	 * If this is just a wiring change, the two PTEs will be
3371 	 * identical, so there's no need to update the page table.
3372 	 */
3373 	if (npte != opte) {
3374 		l2pte_reset(ptep);
3375 		PTE_SYNC(ptep);
3376 		if (l2pte_valid_p(opte)) {
3377 			pmap_tlb_flush_SE(pm, va, oflags);
3378 		}
3379 		l2pte_set(ptep, npte, 0);
3380 		PTE_SYNC(ptep);
3381 #ifndef ARM_MMU_EXTENDED
3382 		bool is_cached = pmap_is_cached(pm);
3383 		if (is_cached) {
3384 			/*
3385 			 * We only need to frob the cache/tlb if this pmap
3386 			 * is current
3387 			 */
3388 			if (!vector_page_p && l2pte_valid_p(npte)) {
3389 				/*
3390 				 * This mapping is likely to be accessed as
3391 				 * soon as we return to userland. Fix up the
3392 				 * L1 entry to avoid taking another
3393 				 * page/domain fault.
3394 				 */
3395 				pd_entry_t *pdep = pmap_l1_kva(pm)
3396 				     + l1pte_index(va);
3397 				pd_entry_t pde = L1_C_PROTO | l2b->l2b_pa
3398 				    | L1_C_DOM(pmap_domain(pm));
3399 				if (*pdep != pde) {
3400 					l1pte_setone(pdep, pde);
3401 					PDE_SYNC(pdep);
3402 				}
3403 			}
3404 		}
3405 
3406 		UVMHIST_LOG(maphist, "  is_cached %jd cs 0x%08jx",
3407 		    is_cached, pm->pm_cstate.cs_all, 0, 0);
3408 
3409 		if (pg != NULL) {
3410 			struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3411 
3412 			pmap_acquire_page_lock(md);
3413 			pmap_vac_me_harder(md, pa, pm, va);
3414 			pmap_release_page_lock(md);
3415 		}
3416 #endif
3417 	}
3418 #if defined(PMAP_CACHE_VIPT) && defined(DIAGNOSTIC)
3419 	if (pg) {
3420 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3421 
3422 		pmap_acquire_page_lock(md);
3423 #ifndef ARM_MMU_EXTENDED
3424 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
3425 #endif
3426 		PMAP_VALIDATE_MD_PAGE(md);
3427 		pmap_release_page_lock(md);
3428 	}
3429 #endif
3430 
3431 	pmap_release_pmap_lock(pm);
3432 	kpreempt_enable();
3433 
3434 	if (old_pv)
3435 		pool_put(&pmap_pv_pool, old_pv);
3436 free_pv:
3437 	if (new_pv)
3438 		pool_put(&pmap_pv_pool, new_pv);
3439 
3440 	return error;
3441 }
3442 
3443 /*
3444  * pmap_remove()
3445  *
3446  * pmap_remove is responsible for nuking a number of mappings for a range
3447  * of virtual address space in the current pmap. To do this efficiently
3448  * is interesting, because in a number of cases a wide virtual address
3449  * range may be supplied that contains few actual mappings. So, the
3450  * optimisations are:
3451  *  1. Skip over hunks of address space for which no L1 or L2 entry exists.
3452  *  2. Build up a list of pages we've hit, up to a maximum, so we can
3453  *     maybe do just a partial cache clean. This path of execution is
3454  *     complicated by the fact that the cache must be flushed _before_
3455  *     the PTE is nuked, being a VAC :-)
3456  *  3. If we're called after UVM calls pmap_remove_all(), we can defer
3457  *     all invalidations until pmap_update(), since pmap_remove_all() has
3458  *     already flushed the cache.
3459  *  4. Maybe later fast-case a single page, but I don't think this is
3460  *     going to make _that_ much difference overall.
3461  */
3462 
3463 #define	PMAP_REMOVE_CLEAN_LIST_SIZE	3
3464 
3465 void
3466 pmap_remove(pmap_t pm, vaddr_t sva, vaddr_t eva)
3467 {
3468 	SLIST_HEAD(,pv_entry) opv_list;
3469 	struct pv_entry *pv, *npv;
3470 	UVMHIST_FUNC(__func__);
3471 	UVMHIST_CALLARGS(maphist, " (pm=%#jx, sva=%#jx, eva=%#jx)",
3472 	    (uintptr_t)pm, sva, eva, 0);
3473 
3474 #ifdef PMAP_FAULTINFO
3475 	curpcb->pcb_faultinfo.pfi_faultaddr = 0;
3476 	curpcb->pcb_faultinfo.pfi_repeats = 0;
3477 	curpcb->pcb_faultinfo.pfi_faultptep = NULL;
3478 #endif
3479 
3480 	SLIST_INIT(&opv_list);
3481 	/*
3482 	 * we lock in the pmap => pv_head direction
3483 	 */
3484 	kpreempt_disable();
3485 	pmap_acquire_pmap_lock(pm);
3486 
3487 #ifndef ARM_MMU_EXTENDED
3488 	u_int cleanlist_idx, total, cnt;
3489 	struct {
3490 		vaddr_t va;
3491 		pt_entry_t *ptep;
3492 	} cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
3493 
3494 	if (pm->pm_remove_all || !pmap_is_cached(pm)) {
3495 		cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
3496 		if (pm->pm_cstate.cs_tlb == 0)
3497 			pm->pm_remove_all = true;
3498 	} else
3499 		cleanlist_idx = 0;
3500 	total = 0;
3501 #endif
3502 
3503 	while (sva < eva) {
3504 		/*
3505 		 * Do one L2 bucket's worth at a time.
3506 		 */
3507 		vaddr_t next_bucket = L2_NEXT_BUCKET_VA(sva);
3508 		if (next_bucket > eva)
3509 			next_bucket = eva;
3510 
3511 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pm, sva);
3512 		if (l2b == NULL) {
3513 			sva = next_bucket;
3514 			continue;
3515 		}
3516 
3517 		pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(sva)];
3518 		u_int mappings = 0;
3519 
3520 		for (;sva < next_bucket;
3521 		     sva += PAGE_SIZE, ptep += PAGE_SIZE / L2_S_SIZE) {
3522 			pt_entry_t opte = *ptep;
3523 
3524 			if (opte == 0) {
3525 				/* Nothing here, move along */
3526 				continue;
3527 			}
3528 
3529 			u_int flags = PVF_REF;
3530 			paddr_t pa = l2pte_pa(opte);
3531 			struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
3532 
3533 			/*
3534 			 * Update flags. In a number of circumstances,
3535 			 * we could cluster a lot of these and do a
3536 			 * number of sequential pages in one go.
3537 			 */
3538 			if (pg != NULL) {
3539 				struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3540 
3541 				pmap_acquire_page_lock(md);
3542 				pv = pmap_remove_pv(md, pa, pm, sva);
3543 				pmap_vac_me_harder(md, pa, pm, 0);
3544 				pmap_release_page_lock(md);
3545 				if (pv != NULL) {
3546 					if (pm->pm_remove_all == false) {
3547 						flags = pv->pv_flags;
3548 					}
3549 					SLIST_INSERT_HEAD(&opv_list,
3550 					    pv, pv_link);
3551 				}
3552 			}
3553 			mappings += PAGE_SIZE / L2_S_SIZE;
3554 
3555 			if (!l2pte_valid_p(opte)) {
3556 				/*
3557 				 * Ref/Mod emulation is still active for this
3558 				 * mapping, therefore it is has not yet been
3559 				 * accessed. No need to frob the cache/tlb.
3560 				 */
3561 				l2pte_reset(ptep);
3562 				PTE_SYNC_CURRENT(pm, ptep);
3563 				continue;
3564 			}
3565 
3566 #ifdef ARM_MMU_EXTENDED
3567 			l2pte_reset(ptep);
3568 			PTE_SYNC(ptep);
3569 			if (__predict_false(pm->pm_remove_all == false)) {
3570 				pmap_tlb_flush_SE(pm, sva, flags);
3571 			}
3572 #else
3573 			if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
3574 				/* Add to the clean list. */
3575 				cleanlist[cleanlist_idx].ptep = ptep;
3576 				cleanlist[cleanlist_idx].va =
3577 				    sva | (flags & PVF_EXEC);
3578 				cleanlist_idx++;
3579 			} else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
3580 				/* Nuke everything if needed. */
3581 #ifdef PMAP_CACHE_VIVT
3582 				pmap_cache_wbinv_all(pm, PVF_EXEC);
3583 #endif
3584 				/*
3585 				 * Roll back the previous PTE list,
3586 				 * and zero out the current PTE.
3587 				 */
3588 				for (cnt = 0;
3589 				     cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
3590 					l2pte_reset(cleanlist[cnt].ptep);
3591 					PTE_SYNC(cleanlist[cnt].ptep);
3592 				}
3593 				l2pte_reset(ptep);
3594 				PTE_SYNC(ptep);
3595 				cleanlist_idx++;
3596 				pm->pm_remove_all = true;
3597 			} else {
3598 				l2pte_reset(ptep);
3599 				PTE_SYNC(ptep);
3600 				if (pm->pm_remove_all == false) {
3601 					pmap_tlb_flush_SE(pm, sva, flags);
3602 				}
3603 			}
3604 #endif
3605 		}
3606 
3607 #ifndef ARM_MMU_EXTENDED
3608 		/*
3609 		 * Deal with any left overs
3610 		 */
3611 		if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
3612 			total += cleanlist_idx;
3613 			for (cnt = 0; cnt < cleanlist_idx; cnt++) {
3614 				l2pte_reset(cleanlist[cnt].ptep);
3615 				PTE_SYNC_CURRENT(pm, cleanlist[cnt].ptep);
3616 				vaddr_t va = cleanlist[cnt].va;
3617 				if (pm->pm_cstate.cs_all != 0) {
3618 					vaddr_t clva = va & ~PAGE_MASK;
3619 					u_int flags = va & PVF_EXEC;
3620 #ifdef PMAP_CACHE_VIVT
3621 					pmap_cache_wbinv_page(pm, clva, true,
3622 					    PVF_REF | PVF_WRITE | flags);
3623 #endif
3624 					pmap_tlb_flush_SE(pm, clva,
3625 					    PVF_REF | flags);
3626 				}
3627 			}
3628 
3629 			/*
3630 			 * If it looks like we're removing a whole bunch
3631 			 * of mappings, it's faster to just write-back
3632 			 * the whole cache now and defer TLB flushes until
3633 			 * pmap_update() is called.
3634 			 */
3635 			if (total <= PMAP_REMOVE_CLEAN_LIST_SIZE)
3636 				cleanlist_idx = 0;
3637 			else {
3638 				cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
3639 #ifdef PMAP_CACHE_VIVT
3640 				pmap_cache_wbinv_all(pm, PVF_EXEC);
3641 #endif
3642 				pm->pm_remove_all = true;
3643 			}
3644 		}
3645 #endif /* ARM_MMU_EXTENDED */
3646 
3647 		pmap_free_l2_bucket(pm, l2b, mappings);
3648 		pm->pm_stats.resident_count -= mappings / (PAGE_SIZE/L2_S_SIZE);
3649 	}
3650 
3651 	pmap_release_pmap_lock(pm);
3652 	kpreempt_enable();
3653 
3654 	SLIST_FOREACH_SAFE(pv, &opv_list, pv_link, npv) {
3655 		pool_put(&pmap_pv_pool, pv);
3656 	}
3657 }
3658 
3659 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3660 static struct pv_entry *
3661 pmap_kremove_pg(struct vm_page *pg, vaddr_t va)
3662 {
3663 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3664 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
3665 	struct pv_entry *pv;
3666 
3667 	KASSERT(arm_cache_prefer_mask == 0 || md->pvh_attrs & (PVF_COLORED|PVF_NC));
3668 	KASSERT((md->pvh_attrs & PVF_KMPAGE) == 0);
3669 	KASSERT(pmap_page_locked_p(md));
3670 
3671 	pv = pmap_remove_pv(md, pa, pmap_kernel(), va);
3672 	KASSERTMSG(pv, "pg %p (pa #%lx) va %#lx", pg, pa, va);
3673 	KASSERT(PV_IS_KENTRY_P(pv->pv_flags));
3674 
3675 	/*
3676 	 * We are removing a writeable mapping to a cached exec page, if
3677 	 * it's the last mapping then clear its execness otherwise sync
3678 	 * the page to the icache.
3679 	 */
3680 	if ((md->pvh_attrs & (PVF_NC|PVF_EXEC)) == PVF_EXEC
3681 	    && (pv->pv_flags & PVF_WRITE) != 0) {
3682 		if (SLIST_EMPTY(&md->pvh_list)) {
3683 			md->pvh_attrs &= ~PVF_EXEC;
3684 			PMAPCOUNT(exec_discarded_kremove);
3685 		} else {
3686 			pmap_syncicache_page(md, pa);
3687 			PMAPCOUNT(exec_synced_kremove);
3688 		}
3689 	}
3690 	pmap_vac_me_harder(md, pa, pmap_kernel(), 0);
3691 
3692 	return pv;
3693 }
3694 #endif /* PMAP_CACHE_VIPT && !ARM_MMU_EXTENDED */
3695 
3696 /*
3697  * pmap_kenter_pa: enter an unmanaged, wired kernel mapping
3698  *
3699  * We assume there is already sufficient KVM space available
3700  * to do this, as we can't allocate L2 descriptor tables/metadata
3701  * from here.
3702  */
3703 void
3704 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags)
3705 {
3706 #ifdef PMAP_CACHE_VIVT
3707 	struct vm_page *pg = (flags & PMAP_KMPAGE) ? PHYS_TO_VM_PAGE(pa) : NULL;
3708 #endif
3709 #ifdef PMAP_CACHE_VIPT
3710 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
3711 	struct vm_page *opg;
3712 #ifndef ARM_MMU_EXTENDED
3713 	struct pv_entry *pv = NULL;
3714 #endif
3715 #endif
3716 	struct vm_page_md *md = pg != NULL ? VM_PAGE_TO_MD(pg) : NULL;
3717 
3718 	UVMHIST_FUNC(__func__);
3719 
3720 	if (pmap_initialized) {
3721 		UVMHIST_CALLARGS(maphist,
3722 		    "va=%#jx, pa=%#jx, prot=%#jx, flags=%#jx", va, pa, prot,
3723 		     flags);
3724 	}
3725 
3726 	kpreempt_disable();
3727 	pmap_t kpm = pmap_kernel();
3728 	pmap_acquire_pmap_lock(kpm);
3729 	struct l2_bucket * const l2b = pmap_get_l2_bucket(kpm, va);
3730 	const size_t l1slot __diagused = l1pte_index(va);
3731 	KASSERTMSG(l2b != NULL,
3732 	    "va %#lx pa %#lx prot %d maxkvaddr %#lx: l2 %p l2b %p kva %p",
3733 	    va, pa, prot, pmap_curmaxkvaddr, kpm->pm_l2[L2_IDX(l1slot)],
3734 	    kpm->pm_l2[L2_IDX(l1slot)]
3735 		? &kpm->pm_l2[L2_IDX(l1slot)]->l2_bucket[L2_BUCKET(l1slot)]
3736 		: NULL,
3737 	    kpm->pm_l2[L2_IDX(l1slot)]
3738 		? kpm->pm_l2[L2_IDX(l1slot)]->l2_bucket[L2_BUCKET(l1slot)].l2b_kva
3739 		: NULL);
3740 	KASSERT(l2b->l2b_kva != NULL);
3741 
3742 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
3743 	const pt_entry_t opte = *ptep;
3744 
3745 	if (opte == 0) {
3746 		PMAPCOUNT(kenter_mappings);
3747 		l2b->l2b_occupancy += PAGE_SIZE / L2_S_SIZE;
3748 	} else {
3749 		PMAPCOUNT(kenter_remappings);
3750 #ifdef PMAP_CACHE_VIPT
3751 		opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3752 #if !defined(ARM_MMU_EXTENDED) || defined(DIAGNOSTIC)
3753 		struct vm_page_md *omd __diagused = VM_PAGE_TO_MD(opg);
3754 #endif
3755 		if (opg && arm_cache_prefer_mask != 0) {
3756 			KASSERT(opg != pg);
3757 			KASSERT((omd->pvh_attrs & PVF_KMPAGE) == 0);
3758 			KASSERT((flags & PMAP_KMPAGE) == 0);
3759 #ifndef ARM_MMU_EXTENDED
3760 			pmap_acquire_page_lock(omd);
3761 			pv = pmap_kremove_pg(opg, va);
3762 			pmap_release_page_lock(omd);
3763 #endif
3764 		}
3765 #endif
3766 		if (l2pte_valid_p(opte)) {
3767 			l2pte_reset(ptep);
3768 			PTE_SYNC(ptep);
3769 #ifdef PMAP_CACHE_VIVT
3770 			cpu_dcache_wbinv_range(va, PAGE_SIZE);
3771 #endif
3772 			cpu_tlb_flushD_SE(va);
3773 			cpu_cpwait();
3774 		}
3775 	}
3776 	pmap_release_pmap_lock(kpm);
3777 	pt_entry_t npte = L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot);
3778 
3779 	if (flags & PMAP_PTE) {
3780 		KASSERT((flags & PMAP_CACHE_MASK) == 0);
3781 		if (!(flags & PMAP_NOCACHE))
3782 			npte |= pte_l2_s_cache_mode_pt;
3783 	} else {
3784 		switch (flags & (PMAP_CACHE_MASK | PMAP_DEV_MASK)) {
3785 		case PMAP_DEV ... PMAP_DEV | PMAP_CACHE_MASK:
3786 			break;
3787 		case PMAP_NOCACHE:
3788 			npte |= pte_l2_s_nocache_mode;
3789 			break;
3790 		case PMAP_WRITE_COMBINE:
3791 			npte |= pte_l2_s_wc_mode;
3792 			break;
3793 		default:
3794 			npte |= pte_l2_s_cache_mode;
3795 			break;
3796 		}
3797 	}
3798 #ifdef ARM_MMU_EXTENDED
3799 	if (prot & VM_PROT_EXECUTE)
3800 		npte &= ~L2_XS_XN;
3801 #endif
3802 	l2pte_set(ptep, npte, 0);
3803 	PTE_SYNC(ptep);
3804 
3805 	if (pg) {
3806 		if (flags & PMAP_KMPAGE) {
3807 			KASSERT(md->urw_mappings == 0);
3808 			KASSERT(md->uro_mappings == 0);
3809 			KASSERT(md->krw_mappings == 0);
3810 			KASSERT(md->kro_mappings == 0);
3811 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3812 			KASSERT(pv == NULL);
3813 			KASSERT(arm_cache_prefer_mask == 0 || (va & PVF_COLORED) == 0);
3814 			KASSERT((md->pvh_attrs & PVF_NC) == 0);
3815 			/* if there is a color conflict, evict from cache. */
3816 			if (pmap_is_page_colored_p(md)
3817 			    && ((va ^ md->pvh_attrs) & arm_cache_prefer_mask)) {
3818 				PMAPCOUNT(vac_color_change);
3819 				pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
3820 			} else if (md->pvh_attrs & PVF_MULTCLR) {
3821 				/*
3822 				 * If this page has multiple colors, expunge
3823 				 * them.
3824 				 */
3825 				PMAPCOUNT(vac_flush_lots2);
3826 				pmap_flush_page(md, pa, PMAP_FLUSH_SECONDARY);
3827 			}
3828 			/*
3829 			 * Since this is a KMPAGE, there can be no contention
3830 			 * for this page so don't lock it.
3831 			 */
3832 			md->pvh_attrs &= PAGE_SIZE - 1;
3833 			md->pvh_attrs |= PVF_KMPAGE | PVF_COLORED | PVF_DIRTY
3834 			    | (va & arm_cache_prefer_mask);
3835 #else /* !PMAP_CACHE_VIPT || ARM_MMU_EXTENDED */
3836 			md->pvh_attrs |= PVF_KMPAGE;
3837 #endif
3838 			atomic_inc_32(&pmap_kmpages);
3839 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3840 		} else if (arm_cache_prefer_mask != 0) {
3841 			if (pv == NULL) {
3842 				pv = pool_get(&pmap_pv_pool, PR_NOWAIT);
3843 				KASSERT(pv != NULL);
3844 			}
3845 			pmap_acquire_page_lock(md);
3846 			pmap_enter_pv(md, pa, pv, pmap_kernel(), va,
3847 			    PVF_WIRED | PVF_KENTRY
3848 			    | (prot & VM_PROT_WRITE ? PVF_WRITE : 0));
3849 			if ((prot & VM_PROT_WRITE)
3850 			    && !(md->pvh_attrs & PVF_NC))
3851 				md->pvh_attrs |= PVF_DIRTY;
3852 			KASSERT((prot & VM_PROT_WRITE) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
3853 			pmap_vac_me_harder(md, pa, pmap_kernel(), va);
3854 			pmap_release_page_lock(md);
3855 #endif
3856 		}
3857 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3858 	} else {
3859 		if (pv != NULL)
3860 			pool_put(&pmap_pv_pool, pv);
3861 #endif
3862 	}
3863 	kpreempt_enable();
3864 
3865 	if (pmap_initialized) {
3866 		UVMHIST_LOG(maphist, "  <-- done (ptep %#jx: %#jx -> %#jx)",
3867 		    (uintptr_t)ptep, opte, npte, 0);
3868 	}
3869 
3870 }
3871 
3872 void
3873 pmap_kremove(vaddr_t va, vsize_t len)
3874 {
3875 #ifdef UVMHIST
3876 	u_int total_mappings = 0;
3877 #endif
3878 
3879 	PMAPCOUNT(kenter_unmappings);
3880 
3881 	UVMHIST_FUNC(__func__);
3882 	UVMHIST_CALLARGS(maphist, " (va=%#jx, len=%#jx)", va, len, 0, 0);
3883 
3884 	const vaddr_t eva = va + len;
3885 	pmap_t kpm = pmap_kernel();
3886 
3887 	kpreempt_disable();
3888 	pmap_acquire_pmap_lock(kpm);
3889 
3890 	while (va < eva) {
3891 		vaddr_t next_bucket = L2_NEXT_BUCKET_VA(va);
3892 		if (next_bucket > eva)
3893 			next_bucket = eva;
3894 
3895 		struct l2_bucket * const l2b = pmap_get_l2_bucket(kpm, va);
3896 		KDASSERT(l2b != NULL);
3897 
3898 		pt_entry_t * const sptep = &l2b->l2b_kva[l2pte_index(va)];
3899 		pt_entry_t *ptep = sptep;
3900 		u_int mappings = 0;
3901 
3902 		while (va < next_bucket) {
3903 			const pt_entry_t opte = *ptep;
3904 			struct vm_page *opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3905 			if (opg != NULL) {
3906 				struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3907 
3908 				if (omd->pvh_attrs & PVF_KMPAGE) {
3909 					KASSERT(omd->urw_mappings == 0);
3910 					KASSERT(omd->uro_mappings == 0);
3911 					KASSERT(omd->krw_mappings == 0);
3912 					KASSERT(omd->kro_mappings == 0);
3913 					omd->pvh_attrs &= ~PVF_KMPAGE;
3914 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3915 					if (arm_cache_prefer_mask != 0) {
3916 						omd->pvh_attrs &= ~PVF_WRITE;
3917 					}
3918 #endif
3919 					atomic_dec_32(&pmap_kmpages);
3920 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3921 				} else if (arm_cache_prefer_mask != 0) {
3922 					pmap_acquire_page_lock(omd);
3923 					pool_put(&pmap_pv_pool,
3924 					    pmap_kremove_pg(opg, va));
3925 					pmap_release_page_lock(omd);
3926 #endif
3927 				}
3928 			}
3929 			if (l2pte_valid_p(opte)) {
3930 				l2pte_reset(ptep);
3931 				PTE_SYNC(ptep);
3932 #ifdef PMAP_CACHE_VIVT
3933 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
3934 #endif
3935 				cpu_tlb_flushD_SE(va);
3936 
3937 				mappings += PAGE_SIZE / L2_S_SIZE;
3938 			}
3939 			va += PAGE_SIZE;
3940 			ptep += PAGE_SIZE / L2_S_SIZE;
3941 		}
3942 		KDASSERTMSG(mappings <= l2b->l2b_occupancy, "%u %u",
3943 		    mappings, l2b->l2b_occupancy);
3944 		l2b->l2b_occupancy -= mappings;
3945 		//PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
3946 #ifdef UVMHIST
3947 		total_mappings += mappings;
3948 #endif
3949 	}
3950 	pmap_release_pmap_lock(kpm);
3951 	cpu_cpwait();
3952 	kpreempt_enable();
3953 
3954 	UVMHIST_LOG(maphist, "  <--- done (%ju mappings removed)",
3955 	    total_mappings, 0, 0, 0);
3956 }
3957 
3958 bool
3959 pmap_extract(pmap_t pm, vaddr_t va, paddr_t *pap)
3960 {
3961 
3962 	return pmap_extract_coherency(pm, va, pap, NULL);
3963 }
3964 
3965 bool
3966 pmap_extract_coherency(pmap_t pm, vaddr_t va, paddr_t *pap, bool *coherentp)
3967 {
3968 	struct l2_dtable *l2;
3969 	pd_entry_t *pdep, pde;
3970 	pt_entry_t *ptep, pte;
3971 	paddr_t pa;
3972 	u_int l1slot;
3973 	bool coherent;
3974 
3975 	kpreempt_disable();
3976 	pmap_acquire_pmap_lock(pm);
3977 
3978 	l1slot = l1pte_index(va);
3979 	pdep = pmap_l1_kva(pm) + l1slot;
3980 	pde = *pdep;
3981 
3982 	if (l1pte_section_p(pde)) {
3983 		/*
3984 		 * These should only happen for pmap_kernel()
3985 		 */
3986 		KDASSERT(pm == pmap_kernel());
3987 		pmap_release_pmap_lock(pm);
3988 		kpreempt_enable();
3989 
3990 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
3991 		if (l1pte_supersection_p(pde)) {
3992 			pa = (pde & L1_SS_FRAME) | (va & L1_SS_OFFSET);
3993 		} else
3994 #endif
3995 			pa = (pde & L1_S_FRAME) | (va & L1_S_OFFSET);
3996 		coherent = (pde & L1_S_CACHE_MASK) == 0;
3997 	} else {
3998 		/*
3999 		 * Note that we can't rely on the validity of the L1
4000 		 * descriptor as an indication that a mapping exists.
4001 		 * We have to look it up in the L2 dtable.
4002 		 */
4003 		l2 = pm->pm_l2[L2_IDX(l1slot)];
4004 
4005 		if (l2 == NULL ||
4006 		    (ptep = l2->l2_bucket[L2_BUCKET(l1slot)].l2b_kva) == NULL) {
4007 			pmap_release_pmap_lock(pm);
4008 			kpreempt_enable();
4009 
4010 			return false;
4011 		}
4012 
4013 		pte = ptep[l2pte_index(va)];
4014 		pmap_release_pmap_lock(pm);
4015 		kpreempt_enable();
4016 
4017 		if (pte == 0)
4018 			return false;
4019 
4020 		switch (pte & L2_TYPE_MASK) {
4021 		case L2_TYPE_L:
4022 			pa = (pte & L2_L_FRAME) | (va & L2_L_OFFSET);
4023 			coherent = (pte & L2_L_CACHE_MASK) == 0;
4024 			break;
4025 
4026 		default:
4027 			pa = (pte & ~PAGE_MASK) | (va & PAGE_MASK);
4028 			coherent = (pte & L2_S_CACHE_MASK) == 0;
4029 			break;
4030 		}
4031 	}
4032 
4033 	if (pap != NULL)
4034 		*pap = pa;
4035 
4036 	if (coherentp != NULL)
4037 		*coherentp = (pm == pmap_kernel() && coherent);
4038 
4039 	return true;
4040 }
4041 
4042 /*
4043  * pmap_pv_remove: remove an unmanaged pv-tracked page from all pmaps
4044  *	that map it
4045  */
4046 
4047 static void
4048 pmap_pv_remove(paddr_t pa)
4049 {
4050 	struct pmap_page *pp;
4051 
4052 	KASSERT(kpreempt_disabled());
4053 	pp = pmap_pv_tracked(pa);
4054 	if (pp == NULL)
4055 		panic("pmap_pv_protect: page not pv-tracked: 0x%"PRIxPADDR,
4056 		    pa);
4057 
4058 	struct vm_page_md *md = PMAP_PAGE_TO_MD(pp);
4059 	pmap_page_remove(md, pa);
4060 }
4061 
4062 void
4063 pmap_pv_protect(paddr_t pa, vm_prot_t prot)
4064 {
4065 
4066 	/* the only case is remove at the moment */
4067 	KASSERT(prot == VM_PROT_NONE);
4068 	pmap_pv_remove(pa);
4069 }
4070 
4071 void
4072 pmap_protect(pmap_t pm, vaddr_t sva, vaddr_t eva, vm_prot_t prot)
4073 {
4074 	struct l2_bucket *l2b;
4075 	vaddr_t next_bucket;
4076 
4077 	UVMHIST_FUNC(__func__);
4078 	UVMHIST_CALLARGS(maphist, "pm %#jx va %#jx...#%jx prot %#jx",
4079 	    (uintptr_t)pm, sva, eva, prot);
4080 
4081 	if ((prot & VM_PROT_READ) == 0) {
4082 		pmap_remove(pm, sva, eva);
4083 		return;
4084 	}
4085 
4086 	if (prot & VM_PROT_WRITE) {
4087 		/*
4088 		 * If this is a read->write transition, just ignore it and let
4089 		 * uvm_fault() take care of it later.
4090 		 */
4091 		return;
4092 	}
4093 
4094 	kpreempt_disable();
4095 	pmap_acquire_pmap_lock(pm);
4096 
4097 #ifndef ARM_MMU_EXTENDED
4098 	const bool flush = eva - sva >= PAGE_SIZE * 4;
4099 	u_int flags = 0;
4100 #endif
4101 	u_int clr_mask = PVF_WRITE | ((prot & VM_PROT_EXECUTE) ? 0 : PVF_EXEC);
4102 
4103 	while (sva < eva) {
4104 		next_bucket = L2_NEXT_BUCKET_VA(sva);
4105 		if (next_bucket > eva)
4106 			next_bucket = eva;
4107 
4108 		l2b = pmap_get_l2_bucket(pm, sva);
4109 		if (l2b == NULL) {
4110 			sva = next_bucket;
4111 			continue;
4112 		}
4113 
4114 		pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(sva)];
4115 
4116 		while (sva < next_bucket) {
4117 			const pt_entry_t opte = *ptep;
4118 			if (l2pte_valid_p(opte) && l2pte_writable_p(opte)) {
4119 				struct vm_page *pg;
4120 #ifndef ARM_MMU_EXTENDED
4121 				u_int f;
4122 #endif
4123 
4124 #ifdef PMAP_CACHE_VIVT
4125 				/*
4126 				 * OK, at this point, we know we're doing
4127 				 * write-protect operation.  If the pmap is
4128 				 * active, write-back the page.
4129 				 */
4130 				pmap_cache_wbinv_page(pm, sva, false,
4131 				    PVF_REF | PVF_WRITE);
4132 #endif
4133 
4134 				pg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
4135 				pt_entry_t npte = l2pte_set_readonly(opte);
4136 				l2pte_reset(ptep);
4137 				PTE_SYNC(ptep);
4138 #ifdef ARM_MMU_EXTENDED
4139 				pmap_tlb_flush_SE(pm, sva, PVF_REF);
4140 #endif
4141 				l2pte_set(ptep, npte, 0);
4142 				PTE_SYNC(ptep);
4143 
4144 				if (pg != NULL) {
4145 					struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4146 					paddr_t pa = VM_PAGE_TO_PHYS(pg);
4147 
4148 					pmap_acquire_page_lock(md);
4149 #ifndef ARM_MMU_EXTENDED
4150 					f =
4151 #endif
4152 					    pmap_modify_pv(md, pa, pm, sva,
4153 					       clr_mask, 0);
4154 					pmap_vac_me_harder(md, pa, pm, sva);
4155 					pmap_release_page_lock(md);
4156 #ifndef ARM_MMU_EXTENDED
4157 				} else {
4158 					f = PVF_REF | PVF_EXEC;
4159 				}
4160 
4161 				if (flush) {
4162 					flags |= f;
4163 				} else {
4164 					pmap_tlb_flush_SE(pm, sva, f);
4165 #endif
4166 				}
4167 			}
4168 
4169 			sva += PAGE_SIZE;
4170 			ptep += PAGE_SIZE / L2_S_SIZE;
4171 		}
4172 	}
4173 
4174 #ifndef ARM_MMU_EXTENDED
4175 	if (flush) {
4176 		if (PV_BEEN_EXECD(flags)) {
4177 			pmap_tlb_flushID(pm);
4178 		} else if (PV_BEEN_REFD(flags)) {
4179 			pmap_tlb_flushD(pm);
4180 		}
4181 	}
4182 #endif
4183 
4184 	pmap_release_pmap_lock(pm);
4185 	kpreempt_enable();
4186 }
4187 
4188 void
4189 pmap_icache_sync_range(pmap_t pm, vaddr_t sva, vaddr_t eva)
4190 {
4191 	struct l2_bucket *l2b;
4192 	pt_entry_t *ptep;
4193 	vaddr_t next_bucket;
4194 	vsize_t page_size = trunc_page(sva) + PAGE_SIZE - sva;
4195 
4196 	UVMHIST_FUNC(__func__);
4197 	UVMHIST_CALLARGS(maphist, "pm %#jx va %#jx...#%jx",
4198 	    (uintptr_t)pm, sva, eva, 0);
4199 
4200 	pmap_acquire_pmap_lock(pm);
4201 
4202 	while (sva < eva) {
4203 		next_bucket = L2_NEXT_BUCKET_VA(sva);
4204 		if (next_bucket > eva)
4205 			next_bucket = eva;
4206 
4207 		l2b = pmap_get_l2_bucket(pm, sva);
4208 		if (l2b == NULL) {
4209 			sva = next_bucket;
4210 			continue;
4211 		}
4212 
4213 		for (ptep = &l2b->l2b_kva[l2pte_index(sva)];
4214 		     sva < next_bucket;
4215 		     sva += page_size,
4216 		     ptep += PAGE_SIZE / L2_S_SIZE,
4217 		     page_size = PAGE_SIZE) {
4218 			if (l2pte_valid_p(*ptep)) {
4219 				cpu_icache_sync_range(sva,
4220 				    uimin(page_size, eva - sva));
4221 			}
4222 		}
4223 	}
4224 
4225 	pmap_release_pmap_lock(pm);
4226 }
4227 
4228 void
4229 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
4230 {
4231 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4232 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
4233 
4234 	UVMHIST_FUNC(__func__);
4235 	UVMHIST_CALLARGS(maphist, "md %#jx pa %#jx prot %#jx",
4236 	    (uintptr_t)md, pa, prot, 0);
4237 
4238 	switch(prot) {
4239 	case VM_PROT_READ|VM_PROT_WRITE:
4240 #if defined(ARM_MMU_EXTENDED)
4241 		pmap_acquire_page_lock(md);
4242 		pmap_clearbit(md, pa, PVF_EXEC);
4243 		pmap_release_page_lock(md);
4244 		break;
4245 #endif
4246 	case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
4247 		break;
4248 
4249 	case VM_PROT_READ:
4250 #if defined(ARM_MMU_EXTENDED)
4251 		pmap_acquire_page_lock(md);
4252 		pmap_clearbit(md, pa, PVF_WRITE|PVF_EXEC);
4253 		pmap_release_page_lock(md);
4254 		break;
4255 #endif
4256 	case VM_PROT_READ|VM_PROT_EXECUTE:
4257 		pmap_acquire_page_lock(md);
4258 		pmap_clearbit(md, pa, PVF_WRITE);
4259 		pmap_release_page_lock(md);
4260 		break;
4261 
4262 	default:
4263 		pmap_page_remove(md, pa);
4264 		break;
4265 	}
4266 }
4267 
4268 /*
4269  * pmap_clear_modify:
4270  *
4271  *	Clear the "modified" attribute for a page.
4272  */
4273 bool
4274 pmap_clear_modify(struct vm_page *pg)
4275 {
4276 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4277 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
4278 	bool rv;
4279 
4280 	pmap_acquire_page_lock(md);
4281 
4282 	if (md->pvh_attrs & PVF_MOD) {
4283 		rv = true;
4284 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
4285 		/*
4286 		 * If we are going to clear the modified bit and there are
4287 		 * no other modified bits set, flush the page to memory and
4288 		 * mark it clean.
4289 		 */
4290 		if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) == PVF_MOD)
4291 			pmap_flush_page(md, pa, PMAP_CLEAN_PRIMARY);
4292 #endif
4293 		pmap_clearbit(md, pa, PVF_MOD);
4294 	} else {
4295 		rv = false;
4296 	}
4297 	pmap_release_page_lock(md);
4298 
4299 	return rv;
4300 }
4301 
4302 /*
4303  * pmap_clear_reference:
4304  *
4305  *	Clear the "referenced" attribute for a page.
4306  */
4307 bool
4308 pmap_clear_reference(struct vm_page *pg)
4309 {
4310 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4311 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
4312 	bool rv;
4313 
4314 	pmap_acquire_page_lock(md);
4315 
4316 	if (md->pvh_attrs & PVF_REF) {
4317 		rv = true;
4318 		pmap_clearbit(md, pa, PVF_REF);
4319 	} else {
4320 		rv = false;
4321 	}
4322 	pmap_release_page_lock(md);
4323 
4324 	return rv;
4325 }
4326 
4327 /*
4328  * pmap_is_modified:
4329  *
4330  *	Test if a page has the "modified" attribute.
4331  */
4332 /* See <arm/arm32/pmap.h> */
4333 
4334 /*
4335  * pmap_is_referenced:
4336  *
4337  *	Test if a page has the "referenced" attribute.
4338  */
4339 /* See <arm/arm32/pmap.h> */
4340 
4341 #if defined(ARM_MMU_EXTENDED) && 0
4342 int
4343 pmap_prefetchabt_fixup(void *v)
4344 {
4345 	struct trapframe * const tf = v;
4346 	vaddr_t va = trunc_page(tf->tf_pc);
4347 	int rv = ABORT_FIXUP_FAILED;
4348 
4349 	if (!TRAP_USERMODE(tf) && va < VM_MAXUSER_ADDRESS)
4350 		return rv;
4351 
4352 	kpreempt_disable();
4353 	pmap_t pm = curcpu()->ci_pmap_cur;
4354 	const size_t l1slot = l1pte_index(va);
4355 	struct l2_dtable * const l2 = pm->pm_l2[L2_IDX(l1slot)];
4356 	if (l2 == NULL)
4357 		goto out;
4358 
4359 	struct l2_bucket * const l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
4360 	if (l2b->l2b_kva == NULL)
4361 		goto out;
4362 
4363 	/*
4364 	 * Check the PTE itself.
4365 	 */
4366 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
4367 	const pt_entry_t opte = *ptep;
4368 	if ((opte & L2_S_PROT_U) == 0 || (opte & L2_XS_XN) == 0)
4369 		goto out;
4370 
4371 	paddr_t pa = l2pte_pa(opte);
4372 	struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
4373 	KASSERT(pg != NULL);
4374 
4375 	struct vm_page_md * const md = VM_PAGE_TO_MD(pg);
4376 
4377 	pmap_acquire_page_lock(md);
4378 	struct pv_entry * const pv = pmap_find_pv(md, pm, va);
4379 	KASSERT(pv != NULL);
4380 
4381 	if (PV_IS_EXEC_P(pv->pv_flags)) {
4382 		l2pte_reset(ptep);
4383 		PTE_SYNC(ptep);
4384 		pmap_tlb_flush_SE(pm, va, PVF_EXEC | PVF_REF);
4385 		if (!PV_IS_EXEC_P(md->pvh_attrs)) {
4386 			pmap_syncicache_page(md, pa);
4387 		}
4388 		rv = ABORT_FIXUP_RETURN;
4389 		l2pte_set(ptep, opte & ~L2_XS_XN, 0);
4390 		PTE_SYNC(ptep);
4391 	}
4392 	pmap_release_page_lock(md);
4393 
4394   out:
4395 	kpreempt_enable();
4396 
4397 	return rv;
4398 }
4399 #endif
4400 
4401 int
4402 pmap_fault_fixup(pmap_t pm, vaddr_t va, vm_prot_t ftype, int user)
4403 {
4404 	struct l2_dtable *l2;
4405 	struct l2_bucket *l2b;
4406 	paddr_t pa;
4407 	const size_t l1slot = l1pte_index(va);
4408 	int rv = 0;
4409 
4410 	UVMHIST_FUNC(__func__);
4411 	UVMHIST_CALLARGS(maphist, "pm=%#jx, va=%#jx, ftype=%#jx, user=%jd",
4412 	    (uintptr_t)pm, va, ftype, user);
4413 
4414 	va = trunc_page(va);
4415 
4416 	KASSERT(!user || (pm != pmap_kernel()));
4417 
4418 #ifdef ARM_MMU_EXTENDED
4419 	UVMHIST_LOG(maphist, " ti=%#jx pai=%#jx asid=%#jx",
4420 	    (uintptr_t)cpu_tlb_info(curcpu()),
4421 	    (uintptr_t)PMAP_PAI(pm, cpu_tlb_info(curcpu())),
4422 	    (uintptr_t)PMAP_PAI(pm, cpu_tlb_info(curcpu()))->pai_asid, 0);
4423 #endif
4424 
4425 	kpreempt_disable();
4426 	pmap_acquire_pmap_lock(pm);
4427 
4428 	/*
4429 	 * If there is no l2_dtable for this address, then the process
4430 	 * has no business accessing it.
4431 	 *
4432 	 * Note: This will catch userland processes trying to access
4433 	 * kernel addresses.
4434 	 */
4435 	l2 = pm->pm_l2[L2_IDX(l1slot)];
4436 	if (l2 == NULL) {
4437 		UVMHIST_LOG(maphist, " no l2 for l1slot %#jx", l1slot, 0, 0, 0);
4438 		goto out;
4439 	}
4440 
4441 	/*
4442 	 * Likewise if there is no L2 descriptor table
4443 	 */
4444 	l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
4445 	if (l2b->l2b_kva == NULL) {
4446 		UVMHIST_LOG(maphist, " <-- done (no ptep for l1slot %#jx)",
4447 		    l1slot, 0, 0, 0);
4448 		goto out;
4449 	}
4450 
4451 	/*
4452 	 * Check the PTE itself.
4453 	 */
4454 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
4455 	pt_entry_t const opte = *ptep;
4456 	if (opte == 0 || (opte & L2_TYPE_MASK) == L2_TYPE_L) {
4457 		UVMHIST_LOG(maphist, " <-- done (empty pte)",
4458 		    0, 0, 0, 0);
4459 		goto out;
4460 	}
4461 
4462 #ifndef ARM_HAS_VBAR
4463 	/*
4464 	 * Catch a userland access to the vector page mapped at 0x0
4465 	 */
4466 	if (user && (opte & L2_S_PROT_U) == 0) {
4467 		UVMHIST_LOG(maphist, " <-- done (vector_page)", 0, 0, 0, 0);
4468 		goto out;
4469 	}
4470 #endif
4471 
4472 	pa = l2pte_pa(opte);
4473 	UVMHIST_LOG(maphist, " pa %#jx opte %#jx ", pa, opte, 0, 0);
4474 
4475 	if ((ftype & VM_PROT_WRITE) && !l2pte_writable_p(opte)) {
4476 		/*
4477 		 * This looks like a good candidate for "page modified"
4478 		 * emulation...
4479 		 */
4480 		struct pv_entry *pv;
4481 		struct vm_page *pg;
4482 
4483 		/* Extract the physical address of the page */
4484 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) {
4485 			UVMHIST_LOG(maphist, " <-- done (mod/ref unmanaged page)", 0, 0, 0, 0);
4486 			goto out;
4487 		}
4488 
4489 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4490 
4491 		/* Get the current flags for this page. */
4492 		pmap_acquire_page_lock(md);
4493 		pv = pmap_find_pv(md, pm, va);
4494 		if (pv == NULL || PV_IS_KENTRY_P(pv->pv_flags)) {
4495 			pmap_release_page_lock(md);
4496 			UVMHIST_LOG(maphist, " <-- done (mod/ref emul: no PV)", 0, 0, 0, 0);
4497 			goto out;
4498 		}
4499 
4500 		/*
4501 		 * Do the flags say this page is writable? If not then it
4502 		 * is a genuine write fault. If yes then the write fault is
4503 		 * our fault as we did not reflect the write access in the
4504 		 * PTE. Now we know a write has occurred we can correct this
4505 		 * and also set the modified bit
4506 		 */
4507 		if ((pv->pv_flags & PVF_WRITE) == 0) {
4508 			pmap_release_page_lock(md);
4509 			UVMHIST_LOG(maphist, " <-- done (write fault)", 0, 0, 0, 0);
4510 			goto out;
4511 		}
4512 
4513 		md->pvh_attrs |= PVF_REF | PVF_MOD;
4514 		pv->pv_flags |= PVF_REF | PVF_MOD;
4515 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
4516 		/*
4517 		 * If there are cacheable mappings for this page, mark it dirty.
4518 		 */
4519 		if ((md->pvh_attrs & PVF_NC) == 0)
4520 			md->pvh_attrs |= PVF_DIRTY;
4521 #endif
4522 #ifdef ARM_MMU_EXTENDED
4523 		if (md->pvh_attrs & PVF_EXEC) {
4524 			md->pvh_attrs &= ~PVF_EXEC;
4525 			PMAPCOUNT(exec_discarded_modfixup);
4526 		}
4527 #endif
4528 		pmap_release_page_lock(md);
4529 
4530 		/*
4531 		 * Re-enable write permissions for the page.  No need to call
4532 		 * pmap_vac_me_harder(), since this is just a
4533 		 * modified-emulation fault, and the PVF_WRITE bit isn't
4534 		 * changing. We've already set the cacheable bits based on
4535 		 * the assumption that we can write to this page.
4536 		 */
4537 		const pt_entry_t npte =
4538 		    l2pte_set_writable((opte & ~L2_TYPE_MASK) | L2_S_PROTO)
4539 #ifdef ARM_MMU_EXTENDED
4540 		    | (pm != pmap_kernel() ? L2_XS_nG : 0)
4541 #endif
4542 		    | 0;
4543 		l2pte_reset(ptep);
4544 		PTE_SYNC(ptep);
4545 		pmap_tlb_flush_SE(pm, va,
4546 		    (ftype & VM_PROT_EXECUTE) ? PVF_EXEC | PVF_REF : PVF_REF);
4547 		l2pte_set(ptep, npte, 0);
4548 		PTE_SYNC(ptep);
4549 		PMAPCOUNT(fixup_mod);
4550 		rv = 1;
4551 		UVMHIST_LOG(maphist, " <-- done (mod/ref emul: changed pte "
4552 		    "from %#jx to %#jx)", opte, npte, 0, 0);
4553 	} else if ((opte & L2_TYPE_MASK) == L2_TYPE_INV) {
4554 		/*
4555 		 * This looks like a good candidate for "page referenced"
4556 		 * emulation.
4557 		 */
4558 		struct vm_page *pg;
4559 
4560 		/* Extract the physical address of the page */
4561 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) {
4562 			UVMHIST_LOG(maphist, " <-- done (ref emul: unmanaged page)", 0, 0, 0, 0);
4563 			goto out;
4564 		}
4565 
4566 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4567 
4568 		/* Get the current flags for this page. */
4569 		pmap_acquire_page_lock(md);
4570 		struct pv_entry *pv = pmap_find_pv(md, pm, va);
4571 		if (pv == NULL || PV_IS_KENTRY_P(pv->pv_flags)) {
4572 			pmap_release_page_lock(md);
4573 			UVMHIST_LOG(maphist, " <-- done (ref emul no PV)", 0, 0, 0, 0);
4574 			goto out;
4575 		}
4576 
4577 		md->pvh_attrs |= PVF_REF;
4578 		pv->pv_flags |= PVF_REF;
4579 
4580 		pt_entry_t npte =
4581 		    l2pte_set_readonly((opte & ~L2_TYPE_MASK) | L2_S_PROTO);
4582 #ifdef ARM_MMU_EXTENDED
4583 		if (pm != pmap_kernel()) {
4584 			npte |= L2_XS_nG;
4585 		}
4586 		/*
4587 		 * If we got called from prefetch abort, then ftype will have
4588 		 * VM_PROT_EXECUTE set.  Now see if we have no-execute set in
4589 		 * the PTE.
4590 		 */
4591 		if (user && (ftype & VM_PROT_EXECUTE) && (npte & L2_XS_XN)) {
4592 			/*
4593 			 * Is this a mapping of an executable page?
4594 			 */
4595 			if ((pv->pv_flags & PVF_EXEC) == 0) {
4596 				pmap_release_page_lock(md);
4597 				UVMHIST_LOG(maphist, " <-- done (ref emul: no exec)",
4598 				    0, 0, 0, 0);
4599 				goto out;
4600 			}
4601 			/*
4602 			 * If we haven't synced the page, do so now.
4603 			 */
4604 			if ((md->pvh_attrs & PVF_EXEC) == 0) {
4605 				UVMHIST_LOG(maphist, " ref emul: syncicache "
4606 				    "page #%#jx", pa, 0, 0, 0);
4607 				pmap_syncicache_page(md, pa);
4608 				PMAPCOUNT(fixup_exec);
4609 			}
4610 			npte &= ~L2_XS_XN;
4611 		}
4612 #endif /* ARM_MMU_EXTENDED */
4613 		pmap_release_page_lock(md);
4614 		l2pte_reset(ptep);
4615 		PTE_SYNC(ptep);
4616 		pmap_tlb_flush_SE(pm, va,
4617 		    (ftype & VM_PROT_EXECUTE) ? PVF_EXEC | PVF_REF : PVF_REF);
4618 		l2pte_set(ptep, npte, 0);
4619 		PTE_SYNC(ptep);
4620 		PMAPCOUNT(fixup_ref);
4621 		rv = 1;
4622 		UVMHIST_LOG(maphist, " <-- done (ref emul: changed pte from "
4623 		    "%#jx to %#jx)", opte, npte, 0, 0);
4624 #ifdef ARM_MMU_EXTENDED
4625 	} else if (user && (ftype & VM_PROT_EXECUTE) && (opte & L2_XS_XN)) {
4626 		struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
4627 		if (pg == NULL) {
4628 			UVMHIST_LOG(maphist, " <-- done (unmanaged page)", 0, 0, 0, 0);
4629 			goto out;
4630 		}
4631 
4632 		struct vm_page_md * const md = VM_PAGE_TO_MD(pg);
4633 
4634 		/* Get the current flags for this page. */
4635 		pmap_acquire_page_lock(md);
4636 		struct pv_entry * const pv = pmap_find_pv(md, pm, va);
4637 		if (pv == NULL || (pv->pv_flags & PVF_EXEC) == 0) {
4638 			pmap_release_page_lock(md);
4639 			UVMHIST_LOG(maphist, " <-- done (no PV or not EXEC)", 0, 0, 0, 0);
4640 			goto out;
4641 		}
4642 
4643 		/*
4644 		 * If we haven't synced the page, do so now.
4645 		 */
4646 		if ((md->pvh_attrs & PVF_EXEC) == 0) {
4647 			UVMHIST_LOG(maphist, "syncicache page #%#jx",
4648 			    pa, 0, 0, 0);
4649 			pmap_syncicache_page(md, pa);
4650 		}
4651 		pmap_release_page_lock(md);
4652 		/*
4653 		 * Turn off no-execute.
4654 		 */
4655 		KASSERT(opte & L2_XS_nG);
4656 		l2pte_reset(ptep);
4657 		PTE_SYNC(ptep);
4658 		pmap_tlb_flush_SE(pm, va, PVF_EXEC | PVF_REF);
4659 		l2pte_set(ptep, opte & ~L2_XS_XN, 0);
4660 		PTE_SYNC(ptep);
4661 		rv = 1;
4662 		PMAPCOUNT(fixup_exec);
4663 		UVMHIST_LOG(maphist, "exec: changed pte from %#jx to %#jx",
4664 		    opte, opte & ~L2_XS_XN, 0, 0);
4665 #endif
4666 	}
4667 
4668 #ifndef ARM_MMU_EXTENDED
4669 	/*
4670 	 * We know there is a valid mapping here, so simply
4671 	 * fix up the L1 if necessary.
4672 	 */
4673 	pd_entry_t * const pdep = pmap_l1_kva(pm) + l1slot;
4674 	pd_entry_t pde = L1_C_PROTO | l2b->l2b_pa | L1_C_DOM(pmap_domain(pm));
4675 	if (*pdep != pde) {
4676 		l1pte_setone(pdep, pde);
4677 		PDE_SYNC(pdep);
4678 		rv = 1;
4679 		PMAPCOUNT(fixup_pdes);
4680 	}
4681 #endif
4682 
4683 #ifdef CPU_SA110
4684 	/*
4685 	 * There are bugs in the rev K SA110.  This is a check for one
4686 	 * of them.
4687 	 */
4688 	if (rv == 0 && curcpu()->ci_arm_cputype == CPU_ID_SA110 &&
4689 	    curcpu()->ci_arm_cpurev < 3) {
4690 		/* Always current pmap */
4691 		if (l2pte_valid_p(opte)) {
4692 			extern int kernel_debug;
4693 			if (kernel_debug & 1) {
4694 				struct proc *p = curlwp->l_proc;
4695 				printf("prefetch_abort: page is already "
4696 				    "mapped - pte=%p *pte=%08x\n", ptep, opte);
4697 				printf("prefetch_abort: pc=%08lx proc=%p "
4698 				    "process=%s\n", va, p, p->p_comm);
4699 				printf("prefetch_abort: far=%08x fs=%x\n",
4700 				    cpu_faultaddress(), cpu_faultstatus());
4701 			}
4702 #ifdef DDB
4703 			if (kernel_debug & 2)
4704 				Debugger();
4705 #endif
4706 			rv = 1;
4707 		}
4708 	}
4709 #endif /* CPU_SA110 */
4710 
4711 #ifndef ARM_MMU_EXTENDED
4712 	/*
4713 	 * If 'rv == 0' at this point, it generally indicates that there is a
4714 	 * stale TLB entry for the faulting address.  That might be due to a
4715 	 * wrong setting of pmap_needs_pte_sync.  So set it and retry.
4716 	 */
4717 	if (rv == 0
4718 	    && pm->pm_l1->l1_domain_use_count == 1
4719 	    && pmap_needs_pte_sync == 0) {
4720 		pmap_needs_pte_sync = 1;
4721 		PTE_SYNC(ptep);
4722 		PMAPCOUNT(fixup_ptesync);
4723 		rv = 1;
4724 	}
4725 #endif
4726 
4727 #ifndef MULTIPROCESSOR
4728 #if defined(DEBUG) || 1
4729 	/*
4730 	 * If 'rv == 0' at this point, it generally indicates that there is a
4731 	 * stale TLB entry for the faulting address. This happens when two or
4732 	 * more processes are sharing an L1. Since we don't flush the TLB on
4733 	 * a context switch between such processes, we can take domain faults
4734 	 * for mappings which exist at the same VA in both processes. EVEN IF
4735 	 * WE'VE RECENTLY FIXED UP THE CORRESPONDING L1 in pmap_enter(), for
4736 	 * example.
4737 	 *
4738 	 * This is extremely likely to happen if pmap_enter() updated the L1
4739 	 * entry for a recently entered mapping. In this case, the TLB is
4740 	 * flushed for the new mapping, but there may still be TLB entries for
4741 	 * other mappings belonging to other processes in the 1MB range
4742 	 * covered by the L1 entry.
4743 	 *
4744 	 * Since 'rv == 0', we know that the L1 already contains the correct
4745 	 * value, so the fault must be due to a stale TLB entry.
4746 	 *
4747 	 * Since we always need to flush the TLB anyway in the case where we
4748 	 * fixed up the L1, or frobbed the L2 PTE, we effectively deal with
4749 	 * stale TLB entries dynamically.
4750 	 *
4751 	 * However, the above condition can ONLY happen if the current L1 is
4752 	 * being shared. If it happens when the L1 is unshared, it indicates
4753 	 * that other parts of the pmap are not doing their job WRT managing
4754 	 * the TLB.
4755 	 */
4756 	if (rv == 0
4757 #ifndef ARM_MMU_EXTENDED
4758 	    && pm->pm_l1->l1_domain_use_count == 1
4759 #endif
4760 	    && true) {
4761 #ifdef DEBUG
4762 		extern int last_fault_code;
4763 #else
4764 		int last_fault_code = ftype & VM_PROT_EXECUTE
4765 		    ? armreg_ifsr_read()
4766 		    : armreg_dfsr_read();
4767 #endif
4768 		printf("fixup: pm %p, va 0x%lx, ftype %d - nothing to do!\n",
4769 		    pm, va, ftype);
4770 		printf("fixup: l2 %p, l2b %p, ptep %p, pte %#x\n",
4771 		    l2, l2b, ptep, opte);
4772 
4773 #ifndef ARM_MMU_EXTENDED
4774 		printf("fixup: pdep %p, pde %#x, fsr %#x\n",
4775 		    pdep, pde, last_fault_code);
4776 #else
4777 		printf("fixup: pdep %p, pde %#x, ttbcr %#x\n",
4778 		    &pmap_l1_kva(pm)[l1slot], pmap_l1_kva(pm)[l1slot],
4779 		   armreg_ttbcr_read());
4780 		printf("fixup: fsr %#x cpm %p casid %#x contextidr %#x dacr %#x\n",
4781 		    last_fault_code, curcpu()->ci_pmap_cur,
4782 		    curcpu()->ci_pmap_asid_cur,
4783 		    armreg_contextidr_read(), armreg_dacr_read());
4784 #ifdef _ARM_ARCH_7
4785 		if (ftype & VM_PROT_WRITE)
4786 			armreg_ats1cuw_write(va);
4787 		else
4788 			armreg_ats1cur_write(va);
4789 		isb();
4790 		printf("fixup: par %#x\n", armreg_par_read());
4791 #endif
4792 #endif
4793 #ifdef DDB
4794 		extern int kernel_debug;
4795 
4796 		if (kernel_debug & 2) {
4797 			pmap_release_pmap_lock(pm);
4798 #ifdef UVMHIST
4799 			KERNHIST_DUMP(maphist);
4800 #endif
4801 			cpu_Debugger();
4802 			pmap_acquire_pmap_lock(pm);
4803 		}
4804 #endif
4805 	}
4806 #endif
4807 #endif
4808 
4809 #ifndef ARM_MMU_EXTENDED
4810 	/* Flush the TLB in the shared L1 case - see comment above */
4811 	pmap_tlb_flush_SE(pm, va,
4812 	    (ftype & VM_PROT_EXECUTE) ? PVF_EXEC | PVF_REF : PVF_REF);
4813 #endif
4814 
4815 	rv = 1;
4816 
4817 out:
4818 	pmap_release_pmap_lock(pm);
4819 	kpreempt_enable();
4820 
4821 	return rv;
4822 }
4823 
4824 /*
4825  * Routine:	pmap_procwr
4826  *
4827  * Function:
4828  *	Synchronize caches corresponding to [addr, addr+len) in p.
4829  *
4830  */
4831 void
4832 pmap_procwr(struct proc *p, vaddr_t va, int len)
4833 {
4834 #ifndef ARM_MMU_EXTENDED
4835 
4836 	/* We only need to do anything if it is the current process. */
4837 	if (p == curproc)
4838 		cpu_icache_sync_range(va, len);
4839 #endif
4840 }
4841 
4842 /*
4843  * Routine:	pmap_unwire
4844  * Function:	Clear the wired attribute for a map/virtual-address pair.
4845  *
4846  * In/out conditions:
4847  *		The mapping must already exist in the pmap.
4848  */
4849 void
4850 pmap_unwire(pmap_t pm, vaddr_t va)
4851 {
4852 	struct l2_bucket *l2b;
4853 	pt_entry_t *ptep, pte;
4854 	struct vm_page *pg;
4855 	paddr_t pa;
4856 
4857 	UVMHIST_FUNC(__func__);
4858 	UVMHIST_CALLARGS(maphist, "pm %#jx va %#jx", (uintptr_t)pm, va, 0, 0);
4859 
4860 	kpreempt_disable();
4861 	pmap_acquire_pmap_lock(pm);
4862 
4863 	l2b = pmap_get_l2_bucket(pm, va);
4864 	KDASSERT(l2b != NULL);
4865 
4866 	ptep = &l2b->l2b_kva[l2pte_index(va)];
4867 	pte = *ptep;
4868 
4869 	/* Extract the physical address of the page */
4870 	pa = l2pte_pa(pte);
4871 
4872 	if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
4873 		/* Update the wired bit in the pv entry for this page. */
4874 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4875 
4876 		pmap_acquire_page_lock(md);
4877 		(void) pmap_modify_pv(md, pa, pm, va, PVF_WIRED, 0);
4878 		pmap_release_page_lock(md);
4879 	}
4880 
4881 	pmap_release_pmap_lock(pm);
4882 	kpreempt_enable();
4883 
4884 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
4885 }
4886 
4887 #ifdef ARM_MMU_EXTENDED
4888 void
4889 pmap_md_pdetab_activate(pmap_t pm, struct lwp *l)
4890 {
4891 	UVMHIST_FUNC(__func__);
4892 	struct cpu_info * const ci = curcpu();
4893 	struct pmap_asid_info * const pai = PMAP_PAI(pm, cpu_tlb_info(ci));
4894 
4895 	UVMHIST_CALLARGS(maphist, "pm %#jx (pm->pm_l1_pa %08jx asid %ju)",
4896 	    (uintptr_t)pm, pm->pm_l1_pa, pai->pai_asid, 0);
4897 
4898 	/*
4899 	 * Assume that TTBR1 has only global mappings and TTBR0 only
4900 	 * has non-global mappings.  To prevent speculation from doing
4901 	 * evil things we disable translation table walks using TTBR0
4902 	 * before setting the CONTEXTIDR (ASID) or new TTBR0 value.
4903 	 * Once both are set, table walks are reenabled.
4904 	 */
4905 	const uint32_t old_ttbcr = armreg_ttbcr_read();
4906 	armreg_ttbcr_write(old_ttbcr | TTBCR_S_PD0);
4907 	isb();
4908 
4909 	pmap_tlb_asid_acquire(pm, l);
4910 
4911 	cpu_setttb(pm->pm_l1_pa, pai->pai_asid);
4912 	/*
4913 	 * Now we can reenable tablewalks since the CONTEXTIDR and TTRB0
4914 	 * have been updated.
4915 	 */
4916 	isb();
4917 
4918 	if (pm != pmap_kernel()) {
4919 		armreg_ttbcr_write(old_ttbcr & ~TTBCR_S_PD0);
4920 	}
4921 	cpu_cpwait();
4922 
4923 	KASSERTMSG(ci->ci_pmap_asid_cur == pai->pai_asid, "%u vs %u",
4924 	    ci->ci_pmap_asid_cur, pai->pai_asid);
4925 	ci->ci_pmap_cur = pm;
4926 }
4927 
4928 void
4929 pmap_md_pdetab_deactivate(pmap_t pm)
4930 {
4931 
4932 	UVMHIST_FUNC(__func__);
4933 	UVMHIST_CALLARGS(maphist, "pm %#jx", (uintptr_t)pm, 0, 0, 0);
4934 
4935 	kpreempt_disable();
4936 	struct cpu_info * const ci = curcpu();
4937 	/*
4938 	 * Disable translation table walks from TTBR0 while no pmap has been
4939 	 * activated.
4940 	 */
4941 	const uint32_t old_ttbcr = armreg_ttbcr_read();
4942 	armreg_ttbcr_write(old_ttbcr | TTBCR_S_PD0);
4943 	isb();
4944 	pmap_tlb_asid_deactivate(pm);
4945 	cpu_setttb(pmap_kernel()->pm_l1_pa, KERNEL_PID);
4946 	isb();
4947 
4948 	ci->ci_pmap_cur = pmap_kernel();
4949 	KASSERTMSG(ci->ci_pmap_asid_cur == KERNEL_PID, "ci_pmap_asid_cur %u",
4950 	    ci->ci_pmap_asid_cur);
4951 	kpreempt_enable();
4952 }
4953 #endif
4954 
4955 void
4956 pmap_activate(struct lwp *l)
4957 {
4958 	extern int block_userspace_access;
4959 	pmap_t npm = l->l_proc->p_vmspace->vm_map.pmap;
4960 
4961 	UVMHIST_FUNC(__func__);
4962 	UVMHIST_CALLARGS(maphist, "l=%#jx pm=%#jx", (uintptr_t)l,
4963 	    (uintptr_t)npm, 0, 0);
4964 
4965 	struct cpu_info * const ci = curcpu();
4966 
4967 	/*
4968 	 * If activating a non-current lwp or the current lwp is
4969 	 * already active, just return.
4970 	 */
4971 	if (false
4972 	    || l != curlwp
4973 #ifdef ARM_MMU_EXTENDED
4974 	    || (ci->ci_pmap_cur == npm &&
4975 		(npm == pmap_kernel()
4976 		 /* || PMAP_PAI_ASIDVALID_P(pai, cpu_tlb_info(ci)) */))
4977 #else
4978 	    || npm->pm_activated == true
4979 #endif
4980 	    || false) {
4981 		UVMHIST_LOG(maphist, " <-- (same pmap)", (uintptr_t)curlwp,
4982 		    (uintptr_t)l, 0, 0);
4983 		return;
4984 	}
4985 
4986 #ifndef ARM_MMU_EXTENDED
4987 	const uint32_t ndacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2))
4988 	    | (DOMAIN_CLIENT << (pmap_domain(npm) * 2));
4989 
4990 	/*
4991 	 * If TTB and DACR are unchanged, short-circuit all the
4992 	 * TLB/cache management stuff.
4993 	 */
4994 	pmap_t opm = ci->ci_lastlwp
4995 	    ? ci->ci_lastlwp->l_proc->p_vmspace->vm_map.pmap
4996 	    : NULL;
4997 	if (opm != NULL) {
4998 		uint32_t odacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2))
4999 		    | (DOMAIN_CLIENT << (pmap_domain(opm) * 2));
5000 
5001 		if (opm->pm_l1 == npm->pm_l1 && odacr == ndacr)
5002 			goto all_done;
5003 	}
5004 #endif /* !ARM_MMU_EXTENDED */
5005 
5006 	PMAPCOUNT(activations);
5007 	block_userspace_access = 1;
5008 
5009 #ifndef ARM_MMU_EXTENDED
5010 	/*
5011 	 * If switching to a user vmspace which is different to the
5012 	 * most recent one, and the most recent one is potentially
5013 	 * live in the cache, we must write-back and invalidate the
5014 	 * entire cache.
5015 	 */
5016 	pmap_t rpm = ci->ci_pmap_lastuser;
5017 
5018 	/*
5019 	 * XXXSCW: There's a corner case here which can leave turds in the
5020 	 * cache as reported in kern/41058. They're probably left over during
5021 	 * tear-down and switching away from an exiting process. Until the root
5022 	 * cause is identified and fixed, zap the cache when switching pmaps.
5023 	 * This will result in a few unnecessary cache flushes, but that's
5024 	 * better than silently corrupting data.
5025 	 */
5026 #if 0
5027 	if (npm != pmap_kernel() && rpm && npm != rpm &&
5028 	    rpm->pm_cstate.cs_cache) {
5029 		rpm->pm_cstate.cs_cache = 0;
5030 #ifdef PMAP_CACHE_VIVT
5031 		cpu_idcache_wbinv_all();
5032 #endif
5033 	}
5034 #else
5035 	if (rpm) {
5036 		rpm->pm_cstate.cs_cache = 0;
5037 		if (npm == pmap_kernel())
5038 			ci->ci_pmap_lastuser = NULL;
5039 #ifdef PMAP_CACHE_VIVT
5040 		cpu_idcache_wbinv_all();
5041 #endif
5042 	}
5043 #endif
5044 
5045 	/* No interrupts while we frob the TTB/DACR */
5046 	uint32_t oldirqstate = disable_interrupts(IF32_bits);
5047 #endif /* !ARM_MMU_EXTENDED */
5048 
5049 #ifndef ARM_HAS_VBAR
5050 	/*
5051 	 * For ARM_VECTORS_LOW, we MUST, I repeat, MUST fix up the L1
5052 	 * entry corresponding to 'vector_page' in the incoming L1 table
5053 	 * before switching to it otherwise subsequent interrupts/exceptions
5054 	 * (including domain faults!) will jump into hyperspace.
5055 	 */
5056 	if (npm->pm_pl1vec != NULL) {
5057 		cpu_tlb_flushID_SE((u_int)vector_page);
5058 		cpu_cpwait();
5059 		*npm->pm_pl1vec = npm->pm_l1vec;
5060 		PTE_SYNC(npm->pm_pl1vec);
5061 	}
5062 #endif
5063 
5064 #ifdef ARM_MMU_EXTENDED
5065 	pmap_md_pdetab_activate(npm, l);
5066 #else
5067 	cpu_domains(ndacr);
5068 	if (npm == pmap_kernel() || npm == rpm) {
5069 		/*
5070 		 * Switching to a kernel thread, or back to the
5071 		 * same user vmspace as before... Simply update
5072 		 * the TTB (no TLB flush required)
5073 		 */
5074 		cpu_setttb(npm->pm_l1->l1_physaddr, false);
5075 		cpu_cpwait();
5076 	} else {
5077 		/*
5078 		 * Otherwise, update TTB and flush TLB
5079 		 */
5080 		cpu_context_switch(npm->pm_l1->l1_physaddr);
5081 		if (rpm != NULL)
5082 			rpm->pm_cstate.cs_tlb = 0;
5083 	}
5084 
5085 	restore_interrupts(oldirqstate);
5086 #endif /* ARM_MMU_EXTENDED */
5087 
5088 	block_userspace_access = 0;
5089 
5090 #ifndef ARM_MMU_EXTENDED
5091  all_done:
5092 	/*
5093 	 * The new pmap is resident. Make sure it's marked
5094 	 * as resident in the cache/TLB.
5095 	 */
5096 	npm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
5097 	if (npm != pmap_kernel())
5098 		ci->ci_pmap_lastuser = npm;
5099 
5100 	/* The old pmap is not longer active */
5101 	if (opm != npm) {
5102 		if (opm != NULL)
5103 			opm->pm_activated = false;
5104 
5105 		/* But the new one is */
5106 		npm->pm_activated = true;
5107 	}
5108 	ci->ci_pmap_cur = npm;
5109 #endif
5110 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
5111 }
5112 
5113 void
5114 pmap_deactivate(struct lwp *l)
5115 {
5116 	pmap_t pm = l->l_proc->p_vmspace->vm_map.pmap;
5117 
5118 	UVMHIST_FUNC(__func__);
5119 	UVMHIST_CALLARGS(maphist, "l=%#jx (pm=%#jx)", (uintptr_t)l,
5120 		(uintptr_t)pm, 0, 0);
5121 
5122 #ifdef ARM_MMU_EXTENDED
5123 	pmap_md_pdetab_deactivate(pm);
5124 #else
5125 	/*
5126 	 * If the process is exiting, make sure pmap_activate() does
5127 	 * a full MMU context-switch and cache flush, which we might
5128 	 * otherwise skip. See PR port-arm/38950.
5129 	 */
5130 	if (l->l_proc->p_sflag & PS_WEXIT)
5131 		curcpu()->ci_lastlwp = NULL;
5132 
5133 	pm->pm_activated = false;
5134 #endif
5135 	UVMHIST_LOG(maphist, "  <-- done", 0, 0, 0, 0);
5136 }
5137 
5138 void
5139 pmap_update(pmap_t pm)
5140 {
5141 
5142 	UVMHIST_FUNC(__func__);
5143 	UVMHIST_CALLARGS(maphist, "pm=%#jx remove_all %jd", (uintptr_t)pm,
5144 	    pm->pm_remove_all, 0, 0);
5145 
5146 #ifndef ARM_MMU_EXTENDED
5147 	if (pm->pm_remove_all) {
5148 		/*
5149 		 * Finish up the pmap_remove_all() optimisation by flushing
5150 		 * the TLB.
5151 		 */
5152 		pmap_tlb_flushID(pm);
5153 		pm->pm_remove_all = false;
5154 	}
5155 
5156 	if (pmap_is_current(pm)) {
5157 		/*
5158 		 * If we're dealing with a current userland pmap, move its L1
5159 		 * to the end of the LRU.
5160 		 */
5161 		if (pm != pmap_kernel())
5162 			pmap_use_l1(pm);
5163 
5164 		/*
5165 		 * We can assume we're done with frobbing the cache/tlb for
5166 		 * now. Make sure any future pmap ops don't skip cache/tlb
5167 		 * flushes.
5168 		 */
5169 		pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
5170 	}
5171 #else
5172 
5173 	kpreempt_disable();
5174 #if defined(MULTIPROCESSOR) && PMAP_TLB_MAX > 1
5175 	u_int pending = atomic_swap_uint(&pmap->pm_shootdown_pending, 0);
5176 	if (pending && pmap_tlb_shootdown_bystanders(pmap)) {
5177 		PMAP_COUNT(shootdown_ipis);
5178 	}
5179 #endif
5180 
5181 	/*
5182 	 * If pmap_remove_all was called, we deactivated ourselves and released
5183 	 * our ASID.  Now we have to reactivate ourselves.
5184 	 */
5185 	if (__predict_false(pm->pm_remove_all)) {
5186 		pm->pm_remove_all = false;
5187 
5188 		KASSERT(pm != pmap_kernel());
5189 		pmap_md_pdetab_activate(pm, curlwp);
5190 	}
5191 
5192 	if (arm_has_mpext_p)
5193 		armreg_bpiallis_write(0);
5194 	else
5195 		armreg_bpiall_write(0);
5196 
5197 	kpreempt_enable();
5198 
5199 	KASSERTMSG(pm == pmap_kernel()
5200 	    || curcpu()->ci_pmap_cur != pm
5201 	    || pm->pm_pai[0].pai_asid == curcpu()->ci_pmap_asid_cur,
5202 	    "pmap/asid %p/%#x != %s cur pmap/asid %p/%#x", pm,
5203 	    pm->pm_pai[0].pai_asid, curcpu()->ci_data.cpu_name,
5204 	    curcpu()->ci_pmap_cur, curcpu()->ci_pmap_asid_cur);
5205 #endif
5206 
5207 	PMAPCOUNT(updates);
5208 
5209 	/*
5210 	 * make sure TLB/cache operations have completed.
5211 	 */
5212 	cpu_cpwait();
5213 	UVMHIST_LOG(maphist, "  <-- done", 0, 0, 0, 0);
5214 }
5215 
5216 bool
5217 pmap_remove_all(pmap_t pm)
5218 {
5219 
5220 	UVMHIST_FUNC(__func__);
5221 	UVMHIST_CALLARGS(maphist, "(pm=%#jx)", (uintptr_t)pm, 0, 0, 0);
5222 
5223 	KASSERT(pm != pmap_kernel());
5224 
5225 	kpreempt_disable();
5226 	/*
5227 	 * The vmspace described by this pmap is about to be torn down.
5228 	 * Until pmap_update() is called, UVM will only make calls
5229 	 * to pmap_remove(). We can make life much simpler by flushing
5230 	 * the cache now, and deferring TLB invalidation to pmap_update().
5231 	 */
5232 #ifdef PMAP_CACHE_VIVT
5233 	pmap_cache_wbinv_all(pm, PVF_EXEC);
5234 #endif
5235 #ifdef ARM_MMU_EXTENDED
5236 #ifdef MULTIPROCESSOR
5237 	struct cpu_info * const ci = curcpu();
5238 	// This should be the last CPU with this pmap onproc
5239 	KASSERT(!kcpuset_isotherset(pm->pm_onproc, cpu_index(ci)));
5240 	if (kcpuset_isset(pm->pm_onproc, cpu_index(ci)))
5241 #endif
5242 		pmap_tlb_asid_deactivate(pm);
5243 #ifdef MULTIPROCESSOR
5244 	KASSERT(kcpuset_iszero(pm->pm_onproc));
5245 #endif
5246 
5247 	pmap_tlb_asid_release_all(pm);
5248 #endif
5249 	pm->pm_remove_all = true;
5250 	kpreempt_enable();
5251 
5252 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
5253 	return false;
5254 }
5255 
5256 /*
5257  * Retire the given physical map from service.
5258  * Should only be called if the map contains no valid mappings.
5259  */
5260 void
5261 pmap_destroy(pmap_t pm)
5262 {
5263 	UVMHIST_FUNC(__func__);
5264 	UVMHIST_CALLARGS(maphist, "pm=%#jx remove_all %jd", (uintptr_t)pm,
5265 	    pm ? pm->pm_remove_all : 0, 0, 0);
5266 
5267 	if (pm == NULL)
5268 		return;
5269 
5270 	if (pm->pm_remove_all) {
5271 #ifdef ARM_MMU_EXTENDED
5272  		pmap_tlb_asid_release_all(pm);
5273 #else
5274 		pmap_tlb_flushID(pm);
5275 #endif
5276 		pm->pm_remove_all = false;
5277 	}
5278 
5279 	/*
5280 	 * Drop reference count
5281 	 */
5282 	if (atomic_dec_uint_nv(&pm->pm_refs) > 0) {
5283 #ifndef ARM_MMU_EXTENDED
5284 		if (pmap_is_current(pm)) {
5285 			if (pm != pmap_kernel())
5286 				pmap_use_l1(pm);
5287 			pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
5288 		}
5289 #endif
5290 		return;
5291 	}
5292 
5293 	/*
5294 	 * reference count is zero, free pmap resources and then free pmap.
5295 	 */
5296 
5297 #ifndef ARM_HAS_VBAR
5298 	if (vector_page < KERNEL_BASE) {
5299 		KDASSERT(!pmap_is_current(pm));
5300 
5301 		/* Remove the vector page mapping */
5302 		pmap_remove(pm, vector_page, vector_page + PAGE_SIZE);
5303 		pmap_update(pm);
5304 	}
5305 #endif
5306 
5307 	pmap_free_l1(pm);
5308 
5309 #ifdef ARM_MMU_EXTENDED
5310 #ifdef MULTIPROCESSOR
5311 	kcpuset_destroy(pm->pm_active);
5312 	kcpuset_destroy(pm->pm_onproc);
5313 #endif
5314 #else
5315 	struct cpu_info * const ci = curcpu();
5316 	if (ci->ci_pmap_lastuser == pm)
5317 		ci->ci_pmap_lastuser = NULL;
5318 #endif
5319 
5320 	mutex_destroy(&pm->pm_lock);
5321 	pool_cache_put(&pmap_cache, pm);
5322 	UVMHIST_LOG(maphist, "  <-- done", 0, 0, 0, 0);
5323 }
5324 
5325 
5326 /*
5327  * void pmap_reference(pmap_t pm)
5328  *
5329  * Add a reference to the specified pmap.
5330  */
5331 void
5332 pmap_reference(pmap_t pm)
5333 {
5334 
5335 	if (pm == NULL)
5336 		return;
5337 
5338 #ifndef ARM_MMU_EXTENDED
5339 	pmap_use_l1(pm);
5340 #endif
5341 
5342 	atomic_inc_uint(&pm->pm_refs);
5343 }
5344 
5345 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
5346 
5347 static struct evcnt pmap_prefer_nochange_ev =
5348     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap prefer", "nochange");
5349 static struct evcnt pmap_prefer_change_ev =
5350     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap prefer", "change");
5351 
5352 EVCNT_ATTACH_STATIC(pmap_prefer_change_ev);
5353 EVCNT_ATTACH_STATIC(pmap_prefer_nochange_ev);
5354 
5355 void
5356 pmap_prefer(vaddr_t hint, vaddr_t *vap, int td)
5357 {
5358 	vsize_t mask = arm_cache_prefer_mask | (PAGE_SIZE - 1);
5359 	vaddr_t va = *vap;
5360 	vaddr_t diff = (hint - va) & mask;
5361 	if (diff == 0) {
5362 		pmap_prefer_nochange_ev.ev_count++;
5363 	} else {
5364 		pmap_prefer_change_ev.ev_count++;
5365 		if (__predict_false(td))
5366 			va -= mask + 1;
5367 		*vap = va + diff;
5368 	}
5369 }
5370 #endif /* ARM_MMU_V6 | ARM_MMU_V7 */
5371 
5372 /*
5373  * pmap_zero_page()
5374  *
5375  * Zero a given physical page by mapping it at a page hook point.
5376  * In doing the zero page op, the page we zero is mapped cachable, as with
5377  * StrongARM accesses to non-cached pages are non-burst making writing
5378  * _any_ bulk data very slow.
5379  */
5380 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
5381 void
5382 pmap_zero_page_generic(paddr_t pa)
5383 {
5384 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
5385 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
5386 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
5387 #endif
5388 #if defined(PMAP_CACHE_VIPT)
5389 	/* Choose the last page color it had, if any */
5390 	const vsize_t va_offset = md->pvh_attrs & arm_cache_prefer_mask;
5391 #else
5392 	const vsize_t va_offset = 0;
5393 #endif
5394 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
5395 	/*
5396 	 * Is this page mapped at its natural color?
5397 	 * If we have all of memory mapped, then just convert PA to VA.
5398 	 */
5399 	bool okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5400 	   || va_offset == (pa & arm_cache_prefer_mask);
5401 	const vaddr_t vdstp = okcolor
5402 	    ? pmap_direct_mapped_phys(pa, &okcolor, cpu_cdstp(va_offset))
5403 	    : cpu_cdstp(va_offset);
5404 #else
5405 	const bool okcolor = false;
5406 	const vaddr_t vdstp = cpu_cdstp(va_offset);
5407 #endif
5408 	pt_entry_t * const ptep = cpu_cdst_pte(va_offset);
5409 
5410 
5411 #ifdef DEBUG
5412 	if (!SLIST_EMPTY(&md->pvh_list))
5413 		panic("pmap_zero_page: page has mappings");
5414 #endif
5415 
5416 	KDASSERT((pa & PGOFSET) == 0);
5417 
5418 	if (!okcolor) {
5419 		/*
5420 		 * Hook in the page, zero it, and purge the cache for that
5421 		 * zeroed page. Invalidate the TLB as needed.
5422 		 */
5423 		const pt_entry_t npte = L2_S_PROTO | pa | pte_l2_s_cache_mode
5424 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE);
5425 		l2pte_set(ptep, npte, 0);
5426 		PTE_SYNC(ptep);
5427 		cpu_tlb_flushD_SE(vdstp);
5428 		cpu_cpwait();
5429 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS) && defined(PMAP_CACHE_VIPT) \
5430     && !defined(ARM_MMU_EXTENDED)
5431 		/*
5432 		 * If we are direct-mapped and our color isn't ok, then before
5433 		 * we bzero the page invalidate its contents from the cache and
5434 		 * reset the color to its natural color.
5435 		 */
5436 		cpu_dcache_inv_range(vdstp, PAGE_SIZE);
5437 		md->pvh_attrs &= ~arm_cache_prefer_mask;
5438 		md->pvh_attrs |= (pa & arm_cache_prefer_mask);
5439 #endif
5440 	}
5441 	bzero_page(vdstp);
5442 	if (!okcolor) {
5443 		/*
5444 		 * Unmap the page.
5445 		 */
5446 		l2pte_reset(ptep);
5447 		PTE_SYNC(ptep);
5448 		cpu_tlb_flushD_SE(vdstp);
5449 #ifdef PMAP_CACHE_VIVT
5450 		cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
5451 #endif
5452 	}
5453 #ifdef PMAP_CACHE_VIPT
5454 	/*
5455 	 * This page is now cache resident so it now has a page color.
5456 	 * Any contents have been obliterated so clear the EXEC flag.
5457 	 */
5458 #ifndef ARM_MMU_EXTENDED
5459 	if (!pmap_is_page_colored_p(md)) {
5460 		PMAPCOUNT(vac_color_new);
5461 		md->pvh_attrs |= PVF_COLORED;
5462 	}
5463 	md->pvh_attrs |= PVF_DIRTY;
5464 #endif
5465 	if (PV_IS_EXEC_P(md->pvh_attrs)) {
5466 		md->pvh_attrs &= ~PVF_EXEC;
5467 		PMAPCOUNT(exec_discarded_zero);
5468 	}
5469 #endif
5470 }
5471 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
5472 
5473 #if ARM_MMU_XSCALE == 1
5474 void
5475 pmap_zero_page_xscale(paddr_t pa)
5476 {
5477 #ifdef DEBUG
5478 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
5479 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
5480 
5481 	if (!SLIST_EMPTY(&md->pvh_list))
5482 		panic("pmap_zero_page: page has mappings");
5483 #endif
5484 
5485 	KDASSERT((pa & PGOFSET) == 0);
5486 
5487 	/*
5488 	 * Hook in the page, zero it, and purge the cache for that
5489 	 * zeroed page. Invalidate the TLB as needed.
5490 	 */
5491 
5492 	pt_entry_t npte = L2_S_PROTO | pa |
5493 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
5494 	    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
5495 	l2pte_set(cdst_pte, npte, 0);
5496 	PTE_SYNC(cdst_pte);
5497 	cpu_tlb_flushD_SE(cdstp);
5498 	cpu_cpwait();
5499 	bzero_page(cdstp);
5500 	xscale_cache_clean_minidata();
5501 	l2pte_reset(cdst_pte);
5502 	PTE_SYNC(cdst_pte);
5503 }
5504 #endif /* ARM_MMU_XSCALE == 1 */
5505 
5506 /* pmap_pageidlezero()
5507  *
5508  * The same as above, except that we assume that the page is not
5509  * mapped.  This means we never have to flush the cache first.  Called
5510  * from the idle loop.
5511  */
5512 bool
5513 pmap_pageidlezero(paddr_t pa)
5514 {
5515 	bool rv = true;
5516 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
5517 	struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
5518 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
5519 #endif
5520 #ifdef PMAP_CACHE_VIPT
5521 	/* Choose the last page color it had, if any */
5522 	const vsize_t va_offset = md->pvh_attrs & arm_cache_prefer_mask;
5523 #else
5524 	const vsize_t va_offset = 0;
5525 #endif
5526 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
5527 	bool okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5528 	   || va_offset == (pa & arm_cache_prefer_mask);
5529 	const vaddr_t vdstp = okcolor
5530 	    ? pmap_direct_mapped_phys(pa, &okcolor, cpu_cdstp(va_offset))
5531 	    : cpu_cdstp(va_offset);
5532 #else
5533 	const bool okcolor = false;
5534 	const vaddr_t vdstp = cpu_cdstp(va_offset);
5535 #endif
5536 	pt_entry_t * const ptep = cpu_cdst_pte(va_offset);
5537 
5538 
5539 #ifdef DEBUG
5540 	if (!SLIST_EMPTY(&md->pvh_list))
5541 		panic("pmap_pageidlezero: page has mappings");
5542 #endif
5543 
5544 	KDASSERT((pa & PGOFSET) == 0);
5545 
5546 	if (!okcolor) {
5547 		/*
5548 		 * Hook in the page, zero it, and purge the cache for that
5549 		 * zeroed page. Invalidate the TLB as needed.
5550 		 */
5551 		const pt_entry_t npte = L2_S_PROTO | pa |
5552 		    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
5553 		l2pte_set(ptep, npte, 0);
5554 		PTE_SYNC(ptep);
5555 		cpu_tlb_flushD_SE(vdstp);
5556 		cpu_cpwait();
5557 	}
5558 
5559 	uint64_t *ptr = (uint64_t *)vdstp;
5560 	for (size_t i = 0; i < PAGE_SIZE / sizeof(*ptr); i++) {
5561 		if (sched_curcpu_runnable_p() != 0) {
5562 			/*
5563 			 * A process has become ready.  Abort now,
5564 			 * so we don't keep it waiting while we
5565 			 * do slow memory access to finish this
5566 			 * page.
5567 			 */
5568 			rv = false;
5569 			break;
5570 		}
5571 		*ptr++ = 0;
5572 	}
5573 
5574 #ifdef PMAP_CACHE_VIVT
5575 	if (rv)
5576 		/*
5577 		 * if we aborted we'll rezero this page again later so don't
5578 		 * purge it unless we finished it
5579 		 */
5580 		cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
5581 #elif defined(PMAP_CACHE_VIPT)
5582 	/*
5583 	 * This page is now cache resident so it now has a page color.
5584 	 * Any contents have been obliterated so clear the EXEC flag.
5585 	 */
5586 #ifndef ARM_MMU_EXTENDED
5587 	if (!pmap_is_page_colored_p(md)) {
5588 		PMAPCOUNT(vac_color_new);
5589 		md->pvh_attrs |= PVF_COLORED;
5590 	}
5591 #endif
5592 	if (PV_IS_EXEC_P(md->pvh_attrs)) {
5593 		md->pvh_attrs &= ~PVF_EXEC;
5594 		PMAPCOUNT(exec_discarded_zero);
5595 	}
5596 #endif
5597 	/*
5598 	 * Unmap the page.
5599 	 */
5600 	if (!okcolor) {
5601 		l2pte_reset(ptep);
5602 		PTE_SYNC(ptep);
5603 		cpu_tlb_flushD_SE(vdstp);
5604 	}
5605 
5606 	return rv;
5607 }
5608 
5609 /*
5610  * pmap_copy_page()
5611  *
5612  * Copy one physical page into another, by mapping the pages into
5613  * hook points. The same comment regarding cachability as in
5614  * pmap_zero_page also applies here.
5615  */
5616 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
5617 void
5618 pmap_copy_page_generic(paddr_t src, paddr_t dst)
5619 {
5620 	struct vm_page * const src_pg = PHYS_TO_VM_PAGE(src);
5621 	struct vm_page_md *src_md = VM_PAGE_TO_MD(src_pg);
5622 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
5623 	struct vm_page * const dst_pg = PHYS_TO_VM_PAGE(dst);
5624 	struct vm_page_md *dst_md = VM_PAGE_TO_MD(dst_pg);
5625 #endif
5626 #ifdef PMAP_CACHE_VIPT
5627 	const vsize_t src_va_offset = src_md->pvh_attrs & arm_cache_prefer_mask;
5628 	const vsize_t dst_va_offset = dst_md->pvh_attrs & arm_cache_prefer_mask;
5629 #else
5630 	const vsize_t src_va_offset = 0;
5631 	const vsize_t dst_va_offset = 0;
5632 #endif
5633 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
5634 	/*
5635 	 * Is this page mapped at its natural color?
5636 	 * If we have all of memory mapped, then just convert PA to VA.
5637 	 */
5638 	bool src_okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5639 	    || src_va_offset == (src & arm_cache_prefer_mask);
5640 	bool dst_okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5641 	    || dst_va_offset == (dst & arm_cache_prefer_mask);
5642 	const vaddr_t vsrcp = src_okcolor
5643 	    ? pmap_direct_mapped_phys(src, &src_okcolor,
5644 		cpu_csrcp(src_va_offset))
5645 	    : cpu_csrcp(src_va_offset);
5646 	const vaddr_t vdstp = pmap_direct_mapped_phys(dst, &dst_okcolor,
5647 	    cpu_cdstp(dst_va_offset));
5648 #else
5649 	const bool src_okcolor = false;
5650 	const bool dst_okcolor = false;
5651 	const vaddr_t vsrcp = cpu_csrcp(src_va_offset);
5652 	const vaddr_t vdstp = cpu_cdstp(dst_va_offset);
5653 #endif
5654 	pt_entry_t * const src_ptep = cpu_csrc_pte(src_va_offset);
5655 	pt_entry_t * const dst_ptep = cpu_cdst_pte(dst_va_offset);
5656 
5657 #ifdef DEBUG
5658 	if (!SLIST_EMPTY(&dst_md->pvh_list))
5659 		panic("pmap_copy_page: dst page has mappings");
5660 #endif
5661 
5662 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
5663 	KASSERT(arm_cache_prefer_mask == 0 || src_md->pvh_attrs & (PVF_COLORED|PVF_NC));
5664 #endif
5665 	KDASSERT((src & PGOFSET) == 0);
5666 	KDASSERT((dst & PGOFSET) == 0);
5667 
5668 	/*
5669 	 * Clean the source page.  Hold the source page's lock for
5670 	 * the duration of the copy so that no other mappings can
5671 	 * be created while we have a potentially aliased mapping.
5672 	 */
5673 #ifdef PMAP_CACHE_VIVT
5674 	pmap_acquire_page_lock(src_md);
5675 	(void) pmap_clean_page(src_md, true);
5676 	pmap_release_page_lock(src_md);
5677 #endif
5678 
5679 	/*
5680 	 * Map the pages into the page hook points, copy them, and purge
5681 	 * the cache for the appropriate page. Invalidate the TLB
5682 	 * as required.
5683 	 */
5684 	if (!src_okcolor) {
5685 		const pt_entry_t nsrc_pte = L2_S_PROTO
5686 		    | src
5687 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
5688 		    | ((src_md->pvh_attrs & PVF_NC) ? 0 : pte_l2_s_cache_mode)
5689 #else // defined(PMAP_CACHE_VIVT) || defined(ARM_MMU_EXTENDED)
5690 		    | pte_l2_s_cache_mode
5691 #endif
5692 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ);
5693 		l2pte_set(src_ptep, nsrc_pte, 0);
5694 		PTE_SYNC(src_ptep);
5695 		cpu_tlb_flushD_SE(vsrcp);
5696 		cpu_cpwait();
5697 	}
5698 	if (!dst_okcolor) {
5699 		const pt_entry_t ndst_pte = L2_S_PROTO | dst |
5700 		    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
5701 		l2pte_set(dst_ptep, ndst_pte, 0);
5702 		PTE_SYNC(dst_ptep);
5703 		cpu_tlb_flushD_SE(vdstp);
5704 		cpu_cpwait();
5705 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS) && defined(PMAP_CACHE_VIPT)
5706 		/*
5707 		 * If we are direct-mapped and our color isn't ok, then before
5708 		 * we bcopy to the new page invalidate its contents from the
5709 		 * cache and reset its color to its natural color.
5710 		 */
5711 		cpu_dcache_inv_range(vdstp, PAGE_SIZE);
5712 		dst_md->pvh_attrs &= ~arm_cache_prefer_mask;
5713 		dst_md->pvh_attrs |= (dst & arm_cache_prefer_mask);
5714 #endif
5715 	}
5716 	bcopy_page(vsrcp, vdstp);
5717 #ifdef PMAP_CACHE_VIVT
5718 	cpu_dcache_inv_range(vsrcp, PAGE_SIZE);
5719 	cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
5720 #endif
5721 	/*
5722 	 * Unmap the pages.
5723 	 */
5724 	if (!src_okcolor) {
5725 		l2pte_reset(src_ptep);
5726 		PTE_SYNC(src_ptep);
5727 		cpu_tlb_flushD_SE(vsrcp);
5728 		cpu_cpwait();
5729 	}
5730 	if (!dst_okcolor) {
5731 		l2pte_reset(dst_ptep);
5732 		PTE_SYNC(dst_ptep);
5733 		cpu_tlb_flushD_SE(vdstp);
5734 		cpu_cpwait();
5735 	}
5736 #ifdef PMAP_CACHE_VIPT
5737 	/*
5738 	 * Now that the destination page is in the cache, mark it as colored.
5739 	 * If this was an exec page, discard it.
5740 	 */
5741 	pmap_acquire_page_lock(dst_md);
5742 #ifndef ARM_MMU_EXTENDED
5743 	if (arm_pcache.cache_type == CACHE_TYPE_PIPT) {
5744 		dst_md->pvh_attrs &= ~arm_cache_prefer_mask;
5745 		dst_md->pvh_attrs |= (dst & arm_cache_prefer_mask);
5746 	}
5747 	if (!pmap_is_page_colored_p(dst_md)) {
5748 		PMAPCOUNT(vac_color_new);
5749 		dst_md->pvh_attrs |= PVF_COLORED;
5750 	}
5751 	dst_md->pvh_attrs |= PVF_DIRTY;
5752 #endif
5753 	if (PV_IS_EXEC_P(dst_md->pvh_attrs)) {
5754 		dst_md->pvh_attrs &= ~PVF_EXEC;
5755 		PMAPCOUNT(exec_discarded_copy);
5756 	}
5757 	pmap_release_page_lock(dst_md);
5758 #endif
5759 }
5760 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
5761 
5762 #if ARM_MMU_XSCALE == 1
5763 void
5764 pmap_copy_page_xscale(paddr_t src, paddr_t dst)
5765 {
5766 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
5767 	struct vm_page_md *src_md = VM_PAGE_TO_MD(src_pg);
5768 #ifdef DEBUG
5769 	struct vm_page_md *dst_md = VM_PAGE_TO_MD(PHYS_TO_VM_PAGE(dst));
5770 
5771 	if (!SLIST_EMPTY(&dst_md->pvh_list))
5772 		panic("pmap_copy_page: dst page has mappings");
5773 #endif
5774 
5775 	KDASSERT((src & PGOFSET) == 0);
5776 	KDASSERT((dst & PGOFSET) == 0);
5777 
5778 	/*
5779 	 * Clean the source page.  Hold the source page's lock for
5780 	 * the duration of the copy so that no other mappings can
5781 	 * be created while we have a potentially aliased mapping.
5782 	 */
5783 #ifdef PMAP_CACHE_VIVT
5784 	pmap_acquire_page_lock(src_md);
5785 	(void) pmap_clean_page(src_md, true);
5786 	pmap_release_page_lock(src_md);
5787 #endif
5788 
5789 	/*
5790 	 * Map the pages into the page hook points, copy them, and purge
5791 	 * the cache for the appropriate page. Invalidate the TLB
5792 	 * as required.
5793 	 */
5794 	const pt_entry_t nsrc_pte = L2_S_PROTO | src
5795 	    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ)
5796 	    | L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
5797 	l2pte_set(csrc_pte, nsrc_pte, 0);
5798 	PTE_SYNC(csrc_pte);
5799 
5800 	const pt_entry_t ndst_pte = L2_S_PROTO | dst
5801 	    | L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE)
5802 	    | L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
5803 	l2pte_set(cdst_pte, ndst_pte, 0);
5804 	PTE_SYNC(cdst_pte);
5805 
5806 	cpu_tlb_flushD_SE(csrcp);
5807 	cpu_tlb_flushD_SE(cdstp);
5808 	cpu_cpwait();
5809 	bcopy_page(csrcp, cdstp);
5810 	xscale_cache_clean_minidata();
5811 	l2pte_reset(csrc_pte);
5812 	l2pte_reset(cdst_pte);
5813 	PTE_SYNC(csrc_pte);
5814 	PTE_SYNC(cdst_pte);
5815 }
5816 #endif /* ARM_MMU_XSCALE == 1 */
5817 
5818 /*
5819  * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
5820  *
5821  * Return the start and end addresses of the kernel's virtual space.
5822  * These values are setup in pmap_bootstrap and are updated as pages
5823  * are allocated.
5824  */
5825 void
5826 pmap_virtual_space(vaddr_t *start, vaddr_t *end)
5827 {
5828 	*start = virtual_avail;
5829 	*end = virtual_end;
5830 }
5831 
5832 /*
5833  * Helper function for pmap_grow_l2_bucket()
5834  */
5835 static inline int
5836 pmap_grow_map(vaddr_t va, paddr_t *pap)
5837 {
5838 	paddr_t pa;
5839 
5840 	KASSERT((va & PGOFSET) == 0);
5841 
5842 	if (uvm.page_init_done == false) {
5843 #ifdef PMAP_STEAL_MEMORY
5844 		pv_addr_t pv;
5845 		pmap_boot_pagealloc(PAGE_SIZE,
5846 #ifdef PMAP_CACHE_VIPT
5847 		    arm_cache_prefer_mask,
5848 		    va & arm_cache_prefer_mask,
5849 #else
5850 		    0, 0,
5851 #endif
5852 		    &pv);
5853 		pa = pv.pv_pa;
5854 #else
5855 		if (uvm_page_physget(&pa) == false)
5856 			return 1;
5857 #endif	/* PMAP_STEAL_MEMORY */
5858 	} else {
5859 		struct vm_page *pg;
5860 		pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE);
5861 		if (pg == NULL)
5862 			return 1;
5863 		pa = VM_PAGE_TO_PHYS(pg);
5864 		/*
5865 		 * This new page must not have any mappings.
5866 		 */
5867 		struct vm_page_md *md __diagused = VM_PAGE_TO_MD(pg);
5868 		KASSERT(SLIST_EMPTY(&md->pvh_list));
5869 	}
5870 
5871 	/*
5872 	 * Enter it via pmap_kenter_pa and let that routine do the hard work.
5873 	 */
5874 	pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE,
5875 	    PMAP_KMPAGE | PMAP_PTE);
5876 
5877 	if (pap)
5878 		*pap = pa;
5879 
5880 	PMAPCOUNT(pt_mappings);
5881 
5882 	const pmap_t kpm __diagused = pmap_kernel();
5883 	struct l2_bucket * const l2b __diagused = pmap_get_l2_bucket(kpm, va);
5884 	KASSERT(l2b != NULL);
5885 
5886 	pt_entry_t * const ptep __diagused = &l2b->l2b_kva[l2pte_index(va)];
5887 	const pt_entry_t pte __diagused = *ptep;
5888 	KASSERT(l2pte_valid_p(pte));
5889 	KASSERT((pte & L2_S_CACHE_MASK) == pte_l2_s_cache_mode_pt);
5890 
5891 	memset((void *)va, 0, PAGE_SIZE);
5892 
5893 	return 0;
5894 }
5895 
5896 /*
5897  * This is the same as pmap_alloc_l2_bucket(), except that it is only
5898  * used by pmap_growkernel().
5899  */
5900 static inline struct l2_bucket *
5901 pmap_grow_l2_bucket(pmap_t pm, vaddr_t va)
5902 {
5903 	const size_t l1slot = l1pte_index(va);
5904 	struct l2_dtable *l2;
5905 	vaddr_t nva;
5906 
5907 	CTASSERT((PAGE_SIZE % L2_TABLE_SIZE_REAL) == 0);
5908 	if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL) {
5909 		/*
5910 		 * No mapping at this address, as there is
5911 		 * no entry in the L1 table.
5912 		 * Need to allocate a new l2_dtable.
5913 		 */
5914 		nva = pmap_kernel_l2dtable_kva;
5915 		if ((nva & PGOFSET) == 0) {
5916 			/*
5917 			 * Need to allocate a backing page
5918 			 */
5919 			if (pmap_grow_map(nva, NULL))
5920 				return NULL;
5921 		}
5922 
5923 		l2 = (struct l2_dtable *)nva;
5924 		nva += sizeof(struct l2_dtable);
5925 
5926 		if ((nva & PGOFSET) < (pmap_kernel_l2dtable_kva & PGOFSET)) {
5927 			/*
5928 			 * The new l2_dtable straddles a page boundary.
5929 			 * Map in another page to cover it.
5930 			 */
5931 			if (pmap_grow_map(nva & ~PGOFSET, NULL))
5932 				return NULL;
5933 		}
5934 
5935 		pmap_kernel_l2dtable_kva = nva;
5936 
5937 		/*
5938 		 * Link it into the parent pmap
5939 		 */
5940 		pm->pm_l2[L2_IDX(l1slot)] = l2;
5941 	}
5942 
5943 	struct l2_bucket * const l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
5944 
5945 	/*
5946 	 * Fetch pointer to the L2 page table associated with the address.
5947 	 */
5948 	if (l2b->l2b_kva == NULL) {
5949 		pt_entry_t *ptep;
5950 
5951 		/*
5952 		 * No L2 page table has been allocated. Chances are, this
5953 		 * is because we just allocated the l2_dtable, above.
5954 		 */
5955 		nva = pmap_kernel_l2ptp_kva;
5956 		ptep = (pt_entry_t *)nva;
5957 		if ((nva & PGOFSET) == 0) {
5958 			/*
5959 			 * Need to allocate a backing page
5960 			 */
5961 			if (pmap_grow_map(nva, &pmap_kernel_l2ptp_phys))
5962 				return NULL;
5963 			PTE_SYNC_RANGE(ptep, PAGE_SIZE / sizeof(pt_entry_t));
5964 		}
5965 
5966 		l2->l2_occupancy++;
5967 		l2b->l2b_kva = ptep;
5968 		l2b->l2b_l1slot = l1slot;
5969 		l2b->l2b_pa = pmap_kernel_l2ptp_phys;
5970 
5971 		pmap_kernel_l2ptp_kva += L2_TABLE_SIZE_REAL;
5972 		pmap_kernel_l2ptp_phys += L2_TABLE_SIZE_REAL;
5973 	}
5974 
5975 	return l2b;
5976 }
5977 
5978 vaddr_t
5979 pmap_growkernel(vaddr_t maxkvaddr)
5980 {
5981 	UVMHIST_FUNC(__func__);
5982 	UVMHIST_CALLARGS(maphist, "growing kernel from %#jx to %#jx\n",
5983 	    pmap_curmaxkvaddr, maxkvaddr, 0, 0);
5984 
5985 	pmap_t kpm = pmap_kernel();
5986 #ifndef ARM_MMU_EXTENDED
5987 	struct l1_ttable *l1;
5988 #endif
5989 	int s;
5990 
5991 	if (maxkvaddr <= pmap_curmaxkvaddr)
5992 		goto out;		/* we are OK */
5993 
5994 	KDASSERT(maxkvaddr <= virtual_end);
5995 
5996 	/*
5997 	 * whoops!   we need to add kernel PTPs
5998 	 */
5999 
6000 	vaddr_t pmap_maxkvaddr = pmap_curmaxkvaddr;
6001 
6002 	s = splvm();	/* to be safe */
6003 	mutex_enter(&kpm_lock);
6004 
6005 	/* Map 1MB at a time */
6006 	size_t l1slot = l1pte_index(pmap_maxkvaddr);
6007 #ifdef ARM_MMU_EXTENDED
6008 	pd_entry_t * const spdep = &kpm->pm_l1[l1slot];
6009 	pd_entry_t *pdep = spdep;
6010 #endif
6011 	for (;pmap_curmaxkvaddr < maxkvaddr; pmap_curmaxkvaddr += L1_S_SIZE,
6012 #ifdef ARM_MMU_EXTENDED
6013 	     pdep++,
6014 #endif
6015 	     l1slot++) {
6016 		struct l2_bucket *l2b =
6017 		    pmap_grow_l2_bucket(kpm, pmap_curmaxkvaddr);
6018 		KASSERT(l2b != NULL);
6019 
6020 		const pd_entry_t npde = L1_C_PROTO | l2b->l2b_pa
6021 		    | L1_C_DOM(PMAP_DOMAIN_KERNEL);
6022 #ifdef ARM_MMU_EXTENDED
6023 		KASSERT(*pdep == 0);
6024 		l1pte_setone(pdep, npde);
6025 #else
6026 		/* Distribute new L1 entry to all other L1s */
6027 		SLIST_FOREACH(l1, &l1_list, l1_link) {
6028 			pd_entry_t * const pdep = &l1->l1_kva[l1slot];
6029 			l1pte_setone(pdep, npde);
6030 			PDE_SYNC(pdep);
6031 		}
6032 #endif
6033 	}
6034 #ifdef ARM_MMU_EXTENDED
6035 	PDE_SYNC_RANGE(spdep, pdep - spdep);
6036 #endif
6037 
6038 #ifdef PMAP_CACHE_VIVT
6039 	/*
6040 	 * flush out the cache, expensive but growkernel will happen so
6041 	 * rarely
6042 	 */
6043 	cpu_dcache_wbinv_all();
6044 	cpu_tlb_flushD();
6045 	cpu_cpwait();
6046 #endif
6047 
6048 	mutex_exit(&kpm_lock);
6049 	splx(s);
6050 
6051 	kasan_shadow_map((void *)pmap_maxkvaddr,
6052 	    (size_t)(pmap_curmaxkvaddr - pmap_maxkvaddr));
6053 
6054 out:
6055 	return pmap_curmaxkvaddr;
6056 }
6057 
6058 /************************ Utility routines ****************************/
6059 
6060 #ifndef ARM_HAS_VBAR
6061 /*
6062  * vector_page_setprot:
6063  *
6064  *	Manipulate the protection of the vector page.
6065  */
6066 void
6067 vector_page_setprot(int prot)
6068 {
6069 	struct l2_bucket *l2b;
6070 	pt_entry_t *ptep;
6071 
6072 #if defined(CPU_ARMV7) || defined(CPU_ARM11)
6073 	/*
6074 	 * If we are using VBAR to use the vectors in the kernel, then it's
6075 	 * already mapped in the kernel text so no need to anything here.
6076 	 */
6077 	if (vector_page != ARM_VECTORS_LOW && vector_page != ARM_VECTORS_HIGH) {
6078 		KASSERT((armreg_pfr1_read() & ARM_PFR1_SEC_MASK) != 0);
6079 		return;
6080 	}
6081 #endif
6082 
6083 	l2b = pmap_get_l2_bucket(pmap_kernel(), vector_page);
6084 	KASSERT(l2b != NULL);
6085 
6086 	ptep = &l2b->l2b_kva[l2pte_index(vector_page)];
6087 
6088 	const pt_entry_t opte = *ptep;
6089 #ifdef ARM_MMU_EXTENDED
6090 	const pt_entry_t npte = (opte & ~(L2_S_PROT_MASK|L2_XS_XN))
6091 	    | L2_S_PROT(PTE_KERNEL, prot);
6092 #else
6093 	const pt_entry_t npte = (opte & ~L2_S_PROT_MASK)
6094 	    | L2_S_PROT(PTE_KERNEL, prot);
6095 #endif
6096 	l2pte_set(ptep, npte, opte);
6097 	PTE_SYNC(ptep);
6098 	cpu_tlb_flushD_SE(vector_page);
6099 	cpu_cpwait();
6100 }
6101 #endif
6102 
6103 /*
6104  * Fetch pointers to the PDE/PTE for the given pmap/VA pair.
6105  * Returns true if the mapping exists, else false.
6106  *
6107  * NOTE: This function is only used by a couple of arm-specific modules.
6108  * It is not safe to take any pmap locks here, since we could be right
6109  * in the middle of debugging the pmap anyway...
6110  *
6111  * It is possible for this routine to return false even though a valid
6112  * mapping does exist. This is because we don't lock, so the metadata
6113  * state may be inconsistent.
6114  *
6115  * NOTE: We can return a NULL *ptp in the case where the L1 pde is
6116  * a "section" mapping.
6117  */
6118 bool
6119 pmap_get_pde_pte(pmap_t pm, vaddr_t va, pd_entry_t **pdp, pt_entry_t **ptp)
6120 {
6121 	struct l2_dtable *l2;
6122 	pd_entry_t *pdep, pde;
6123 	pt_entry_t *ptep;
6124 	u_short l1slot;
6125 
6126 	if (pm->pm_l1 == NULL)
6127 		return false;
6128 
6129 	l1slot = l1pte_index(va);
6130 	*pdp = pdep = pmap_l1_kva(pm) + l1slot;
6131 	pde = *pdep;
6132 
6133 	if (l1pte_section_p(pde)) {
6134 		*ptp = NULL;
6135 		return true;
6136 	}
6137 
6138 	l2 = pm->pm_l2[L2_IDX(l1slot)];
6139 	if (l2 == NULL ||
6140 	    (ptep = l2->l2_bucket[L2_BUCKET(l1slot)].l2b_kva) == NULL) {
6141 		return false;
6142 	}
6143 
6144 	*ptp = &ptep[l2pte_index(va)];
6145 	return true;
6146 }
6147 
6148 bool
6149 pmap_get_pde(pmap_t pm, vaddr_t va, pd_entry_t **pdp)
6150 {
6151 
6152 	if (pm->pm_l1 == NULL)
6153 		return false;
6154 
6155 	*pdp = pmap_l1_kva(pm) + l1pte_index(va);
6156 
6157 	return true;
6158 }
6159 
6160 /************************ Bootstrapping routines ****************************/
6161 
6162 #ifndef ARM_MMU_EXTENDED
6163 static void
6164 pmap_init_l1(struct l1_ttable *l1, pd_entry_t *l1pt)
6165 {
6166 	int i;
6167 
6168 	l1->l1_kva = l1pt;
6169 	l1->l1_domain_use_count = 0;
6170 	l1->l1_domain_first = 0;
6171 
6172 	for (i = 0; i < PMAP_DOMAINS; i++)
6173 		l1->l1_domain_free[i] = i + 1;
6174 
6175 	/*
6176 	 * Copy the kernel's L1 entries to each new L1.
6177 	 */
6178 	if (pmap_initialized)
6179 		memcpy(l1pt, pmap_l1_kva(pmap_kernel()), L1_TABLE_SIZE);
6180 
6181 	if (pmap_extract(pmap_kernel(), (vaddr_t)l1pt,
6182 	    &l1->l1_physaddr) == false)
6183 		panic("pmap_init_l1: can't get PA of L1 at %p", l1pt);
6184 
6185 	SLIST_INSERT_HEAD(&l1_list, l1, l1_link);
6186 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
6187 }
6188 #endif /* !ARM_MMU_EXTENDED */
6189 
6190 /*
6191  * pmap_bootstrap() is called from the board-specific initarm() routine
6192  * once the kernel L1/L2 descriptors tables have been set up.
6193  *
6194  * This is a somewhat convoluted process since pmap bootstrap is, effectively,
6195  * spread over a number of disparate files/functions.
6196  *
6197  * We are passed the following parameters
6198  *  - vstart
6199  *    1MB-aligned start of managed kernel virtual memory.
6200  *  - vend
6201  *    1MB-aligned end of managed kernel virtual memory.
6202  *
6203  * We use 'kernel_l1pt' to build the metadata (struct l1_ttable and
6204  * struct l2_dtable) necessary to track kernel mappings.
6205  */
6206 #define	PMAP_STATIC_L2_SIZE 16
6207 void
6208 pmap_bootstrap(vaddr_t vstart, vaddr_t vend)
6209 {
6210 	static struct l2_dtable static_l2[PMAP_STATIC_L2_SIZE];
6211 #ifndef ARM_MMU_EXTENDED
6212 	static struct l1_ttable static_l1;
6213 	struct l1_ttable *l1 = &static_l1;
6214 #endif
6215 	struct l2_dtable *l2;
6216 	struct l2_bucket *l2b;
6217 	pd_entry_t *l1pt = (pd_entry_t *) kernel_l1pt.pv_va;
6218 	pmap_t pm = pmap_kernel();
6219 	pt_entry_t *ptep;
6220 	paddr_t pa;
6221 	vsize_t size;
6222 	int nptes, l2idx, l2next = 0;
6223 
6224 #ifdef ARM_MMU_EXTENDED
6225 	KASSERT(pte_l1_s_cache_mode == pte_l1_s_cache_mode_pt);
6226 	KASSERT(pte_l2_s_cache_mode == pte_l2_s_cache_mode_pt);
6227 #endif
6228 
6229 	VPRINTF("kpm ");
6230 	/*
6231 	 * Initialise the kernel pmap object
6232 	 */
6233 	curcpu()->ci_pmap_cur = pm;
6234 #ifdef ARM_MMU_EXTENDED
6235 	pm->pm_l1 = l1pt;
6236 	pm->pm_l1_pa = kernel_l1pt.pv_pa;
6237 	VPRINTF("tlb0 ");
6238 	pmap_tlb_info_init(&pmap_tlb0_info);
6239 #ifdef MULTIPROCESSOR
6240 	VPRINTF("kcpusets ");
6241 	pm->pm_onproc = kcpuset_running;
6242 	pm->pm_active = kcpuset_running;
6243 #endif
6244 #else
6245 	pm->pm_l1 = l1;
6246 #endif
6247 
6248 	VPRINTF("locks ");
6249 	/*
6250 	 * pmap_kenter_pa() and pmap_kremove() may be called from interrupt
6251 	 * context, so its locks have to be at IPL_VM
6252 	 */
6253 	mutex_init(&pmap_lock, MUTEX_DEFAULT, IPL_VM);
6254 	mutex_init(&kpm_lock, MUTEX_DEFAULT, IPL_NONE);
6255 	mutex_init(&pm->pm_lock, MUTEX_DEFAULT, IPL_VM);
6256 	pm->pm_refs = 1;
6257 
6258 	VPRINTF("l1pt ");
6259 	/*
6260 	 * Scan the L1 translation table created by initarm() and create
6261 	 * the required metadata for all valid mappings found in it.
6262 	 */
6263 	for (size_t l1slot = 0;
6264 	     l1slot < L1_TABLE_SIZE / sizeof(pd_entry_t);
6265 	     l1slot++) {
6266 		pd_entry_t pde = l1pt[l1slot];
6267 
6268 		/*
6269 		 * We're only interested in Coarse mappings.
6270 		 * pmap_extract() can deal with section mappings without
6271 		 * recourse to checking L2 metadata.
6272 		 */
6273 		if ((pde & L1_TYPE_MASK) != L1_TYPE_C)
6274 			continue;
6275 
6276 		/*
6277 		 * Lookup the KVA of this L2 descriptor table
6278 		 */
6279 		pa = l1pte_pa(pde);
6280 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
6281 		if (ptep == NULL) {
6282 			panic("pmap_bootstrap: No L2 for va 0x%x, pa 0x%lx",
6283 			    (u_int)l1slot << L1_S_SHIFT, pa);
6284 		}
6285 
6286 		/*
6287 		 * Fetch the associated L2 metadata structure.
6288 		 * Allocate a new one if necessary.
6289 		 */
6290 		if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL) {
6291 			if (l2next == PMAP_STATIC_L2_SIZE)
6292 				panic("pmap_bootstrap: out of static L2s");
6293 			pm->pm_l2[L2_IDX(l1slot)] = l2 = &static_l2[l2next++];
6294 		}
6295 
6296 		/*
6297 		 * One more L1 slot tracked...
6298 		 */
6299 		l2->l2_occupancy++;
6300 
6301 		/*
6302 		 * Fill in the details of the L2 descriptor in the
6303 		 * appropriate bucket.
6304 		 */
6305 		l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
6306 		l2b->l2b_kva = ptep;
6307 		l2b->l2b_pa = pa;
6308 		l2b->l2b_l1slot = l1slot;
6309 
6310 		/*
6311 		 * Establish an initial occupancy count for this descriptor
6312 		 */
6313 		for (l2idx = 0;
6314 		    l2idx < (L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
6315 		    l2idx++) {
6316 			if ((ptep[l2idx] & L2_TYPE_MASK) != L2_TYPE_INV) {
6317 				l2b->l2b_occupancy++;
6318 			}
6319 		}
6320 
6321 		/*
6322 		 * Make sure the descriptor itself has the correct cache mode.
6323 		 * If not, fix it, but whine about the problem. Port-meisters
6324 		 * should consider this a clue to fix up their initarm()
6325 		 * function. :)
6326 		 */
6327 		if (pmap_set_pt_cache_mode(l1pt, (vaddr_t)ptep, 1)) {
6328 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
6329 			    "L2 pte @ %p\n", ptep);
6330 		}
6331 	}
6332 
6333 	VPRINTF("cache(l1pt) ");
6334 	/*
6335 	 * Ensure the primary (kernel) L1 has the correct cache mode for
6336 	 * a page table. Bitch if it is not correctly set.
6337 	 */
6338 	if (pmap_set_pt_cache_mode(l1pt, kernel_l1pt.pv_va,
6339 		    L1_TABLE_SIZE / L2_S_SIZE)) {
6340 		printf("pmap_bootstrap: WARNING! wrong cache mode for "
6341 		    "primary L1 @ 0x%lx\n", kernel_l1pt.pv_va);
6342 	}
6343 
6344 #ifdef PMAP_CACHE_VIVT
6345 	cpu_dcache_wbinv_all();
6346 	cpu_tlb_flushID();
6347 	cpu_cpwait();
6348 #endif
6349 
6350 	/*
6351 	 * now we allocate the "special" VAs which are used for tmp mappings
6352 	 * by the pmap (and other modules).  we allocate the VAs by advancing
6353 	 * virtual_avail (note that there are no pages mapped at these VAs).
6354 	 *
6355 	 * Managed KVM space start from wherever initarm() tells us.
6356 	 */
6357 	virtual_avail = vstart;
6358 	virtual_end = vend;
6359 
6360 	VPRINTF("specials ");
6361 
6362 	pmap_alloc_specials(&virtual_avail, 1, &memhook, NULL);
6363 
6364 #ifdef PMAP_CACHE_VIPT
6365 	/*
6366 	 * If we have a VIPT cache, we need one page/pte per possible alias
6367 	 * page so we won't violate cache aliasing rules.
6368 	 */
6369 	virtual_avail = (virtual_avail + arm_cache_prefer_mask) & ~arm_cache_prefer_mask;
6370 	nptes = (arm_cache_prefer_mask >> L2_S_SHIFT) + 1;
6371 	nptes = roundup(nptes, PAGE_SIZE / L2_S_SIZE);
6372 	if (arm_pcache.icache_type != CACHE_TYPE_PIPT
6373 	    && arm_pcache.icache_way_size > nptes * L2_S_SIZE) {
6374 		nptes = arm_pcache.icache_way_size >> L2_S_SHIFT;
6375 		nptes = roundup(nptes, PAGE_SIZE / L2_S_SIZE);
6376 	}
6377 #else
6378 	nptes = PAGE_SIZE / L2_S_SIZE;
6379 #endif
6380 #ifdef MULTIPROCESSOR
6381 	cnptes = nptes;
6382 	nptes *= arm_cpu_max;
6383 #endif
6384 	pmap_alloc_specials(&virtual_avail, nptes, &csrcp, &csrc_pte);
6385 	pmap_set_pt_cache_mode(l1pt, (vaddr_t)csrc_pte, nptes);
6386 	pmap_alloc_specials(&virtual_avail, nptes, &cdstp, &cdst_pte);
6387 	pmap_set_pt_cache_mode(l1pt, (vaddr_t)cdst_pte, nptes);
6388 	if (msgbufaddr == NULL) {
6389 		pmap_alloc_specials(&virtual_avail,
6390 		    round_page(MSGBUFSIZE) / PAGE_SIZE,
6391 		    (void *)&msgbufaddr, NULL);
6392 	}
6393 
6394 	/*
6395 	 * Allocate a range of kernel virtual address space to be used
6396 	 * for L2 descriptor tables and metadata allocation in
6397 	 * pmap_growkernel().
6398 	 */
6399 	size = howmany(virtual_end - pmap_curmaxkvaddr, L1_S_SIZE);
6400 	pmap_alloc_specials(&virtual_avail,
6401 	    round_page(size * L2_TABLE_SIZE_REAL) / PAGE_SIZE,
6402 	    &pmap_kernel_l2ptp_kva, NULL);
6403 
6404 	size = howmany(size, L2_BUCKET_SIZE);
6405 	pmap_alloc_specials(&virtual_avail,
6406 	    round_page(size * sizeof(struct l2_dtable)) / PAGE_SIZE,
6407 	    &pmap_kernel_l2dtable_kva, NULL);
6408 
6409 #ifndef ARM_MMU_EXTENDED
6410 	/*
6411 	 * init the static-global locks and global pmap list.
6412 	 */
6413 	mutex_init(&l1_lru_lock, MUTEX_DEFAULT, IPL_VM);
6414 
6415 	/*
6416 	 * We can now initialise the first L1's metadata.
6417 	 */
6418 	SLIST_INIT(&l1_list);
6419 	TAILQ_INIT(&l1_lru_list);
6420 	pmap_init_l1(l1, l1pt);
6421 #endif /* ARM_MMU_EXTENDED */
6422 
6423 #ifndef ARM_HAS_VBAR
6424 	/* Set up vector page L1 details, if necessary */
6425 	if (vector_page < KERNEL_BASE) {
6426 		pm->pm_pl1vec = pmap_l1_kva(pm) + l1pte_index(vector_page);
6427 		l2b = pmap_get_l2_bucket(pm, vector_page);
6428 		KDASSERT(l2b != NULL);
6429 		pm->pm_l1vec = l2b->l2b_pa | L1_C_PROTO |
6430 		    L1_C_DOM(pmap_domain(pm));
6431 	} else
6432 		pm->pm_pl1vec = NULL;
6433 #endif
6434 
6435 	VPRINTF("pools ");
6436 	/*
6437 	 * Initialize the pmap cache
6438 	 */
6439 	pool_cache_bootstrap(&pmap_cache, sizeof(struct pmap), 0, 0, 0,
6440 	    "pmappl", NULL, IPL_NONE, pmap_pmap_ctor, NULL, NULL);
6441 
6442 	/*
6443 	 * Initialize the pv pool.
6444 	 */
6445 	pool_init(&pmap_pv_pool, sizeof(struct pv_entry), 0, 0, 0, "pvepl",
6446 	    &pmap_bootstrap_pv_allocator, IPL_NONE);
6447 
6448 	/*
6449 	 * Initialize the L2 dtable pool and cache.
6450 	 */
6451 	pool_cache_bootstrap(&pmap_l2dtable_cache, sizeof(struct l2_dtable), 0,
6452 	    0, 0, "l2dtblpl", NULL, IPL_NONE, pmap_l2dtable_ctor, NULL, NULL);
6453 
6454 	/*
6455 	 * Initialise the L2 descriptor table pool and cache
6456 	 */
6457 	pool_cache_bootstrap(&pmap_l2ptp_cache, L2_TABLE_SIZE_REAL,
6458 	    L2_TABLE_SIZE_REAL, 0, 0, "l2ptppl", NULL, IPL_NONE,
6459 	    pmap_l2ptp_ctor, NULL, NULL);
6460 
6461 	mutex_init(&memlock, MUTEX_DEFAULT, IPL_NONE);
6462 
6463 	cpu_dcache_wbinv_all();
6464 }
6465 
6466 static bool
6467 pmap_set_pt_cache_mode(pd_entry_t *kl1, vaddr_t va, size_t nptes)
6468 {
6469 #ifdef ARM_MMU_EXTENDED
6470 	return false;
6471 #else
6472 	if (pte_l1_s_cache_mode == pte_l1_s_cache_mode_pt
6473 	    && pte_l2_s_cache_mode == pte_l2_s_cache_mode_pt)
6474 		return false;
6475 
6476 	const vaddr_t eva = va + nptes * PAGE_SIZE;
6477 	int rv = 0;
6478 
6479 	while (va < eva) {
6480 		/*
6481 		 * Make sure the descriptor itself has the correct cache mode
6482 		 */
6483 		pd_entry_t * const pdep = &kl1[l1pte_index(va)];
6484 		pd_entry_t pde = *pdep;
6485 
6486 		if (l1pte_section_p(pde)) {
6487 			KASSERT((L1_S_CACHE_MASK & L1_S_V6_SUPER) == 0);
6488 			if ((pde & L1_S_CACHE_MASK) != pte_l1_s_cache_mode_pt) {
6489 				*pdep = (pde & ~L1_S_CACHE_MASK) |
6490 				    pte_l1_s_cache_mode_pt;
6491 				PDE_SYNC(pdep);
6492 				cpu_dcache_wbinv_range((vaddr_t)pdep,
6493 				    sizeof(*pdep));
6494 				rv = 1;
6495 			}
6496 			return rv;
6497 		}
6498 		vaddr_t pa = l1pte_pa(pde);
6499 		pt_entry_t *ptep = (pt_entry_t *)kernel_pt_lookup(pa);
6500 		if (ptep == NULL)
6501 			panic("pmap_bootstrap: No PTP for va %#lx\n", va);
6502 
6503 		ptep += l2pte_index(va);
6504 		const pt_entry_t opte = *ptep;
6505 		if ((opte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
6506 			const pt_entry_t npte = (opte & ~L2_S_CACHE_MASK)
6507 			    | pte_l2_s_cache_mode_pt;
6508 			l2pte_set(ptep, npte, opte);
6509 			PTE_SYNC(ptep);
6510 			cpu_dcache_wbinv_range((vaddr_t)ptep, sizeof(*ptep));
6511 			rv = 1;
6512 		}
6513 		va += PAGE_SIZE;
6514 	}
6515 
6516 	return rv;
6517 #endif
6518 }
6519 
6520 static void
6521 pmap_alloc_specials(vaddr_t *availp, int pages, vaddr_t *vap, pt_entry_t **ptep)
6522 {
6523 	vaddr_t va = *availp;
6524 	struct l2_bucket *l2b;
6525 
6526 	if (ptep) {
6527 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
6528 		if (l2b == NULL)
6529 			panic("pmap_alloc_specials: no l2b for 0x%lx", va);
6530 
6531 		*ptep = &l2b->l2b_kva[l2pte_index(va)];
6532 	}
6533 
6534 	*vap = va;
6535 	*availp = va + (PAGE_SIZE * pages);
6536 }
6537 
6538 void
6539 pmap_init(void)
6540 {
6541 
6542 	/*
6543 	 * Set the available memory vars - These do not map to real memory
6544 	 * addresses and cannot as the physical memory is fragmented.
6545 	 * They are used by ps for %mem calculations.
6546 	 * One could argue whether this should be the entire memory or just
6547 	 * the memory that is useable in a user process.
6548 	 */
6549 	avail_start = ptoa(uvm_physseg_get_avail_start(uvm_physseg_get_first()));
6550 	avail_end = ptoa(uvm_physseg_get_avail_end(uvm_physseg_get_last()));
6551 
6552 	/*
6553 	 * Now we need to free enough pv_entry structures to allow us to get
6554 	 * the kmem_map/kmem_object allocated and inited (done after this
6555 	 * function is finished).  to do this we allocate one bootstrap page out
6556 	 * of kernel_map and use it to provide an initial pool of pv_entry
6557 	 * structures.   we never free this page.
6558 	 */
6559 	pool_setlowat(&pmap_pv_pool, (PAGE_SIZE / sizeof(struct pv_entry)) * 2);
6560 
6561 #ifdef ARM_MMU_EXTENDED
6562 	/*
6563 	 * Initialise the L1 pool and cache.
6564 	 */
6565 
6566 	pool_cache_bootstrap(&pmap_l1tt_cache, L1TT_SIZE, L1TT_SIZE,
6567 	    0, 0, "l1ttpl", &pmap_l1tt_allocator, IPL_NONE, pmap_l1tt_ctor,
6568 	     NULL, NULL);
6569 
6570 	int error __diagused = pmap_maxproc_set(maxproc);
6571 	KASSERT(error == 0);
6572 
6573 	pmap_tlb_info_evcnt_attach(&pmap_tlb0_info);
6574 #endif
6575 
6576 	pmap_initialized = true;
6577 }
6578 
6579 static vaddr_t last_bootstrap_page = 0;
6580 static void *free_bootstrap_pages = NULL;
6581 
6582 static void *
6583 pmap_bootstrap_pv_page_alloc(struct pool *pp, int flags)
6584 {
6585 	extern void *pool_page_alloc(struct pool *, int);
6586 	vaddr_t new_page;
6587 	void *rv;
6588 
6589 	if (pmap_initialized)
6590 		return pool_page_alloc(pp, flags);
6591 
6592 	if (free_bootstrap_pages) {
6593 		rv = free_bootstrap_pages;
6594 		free_bootstrap_pages = *((void **)rv);
6595 		return rv;
6596 	}
6597 
6598 	KASSERT(kernel_map != NULL);
6599 	new_page = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
6600 	    UVM_KMF_WIRED | ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT));
6601 
6602 	KASSERT(new_page > last_bootstrap_page);
6603 	last_bootstrap_page = new_page;
6604 	return (void *)new_page;
6605 }
6606 
6607 static void
6608 pmap_bootstrap_pv_page_free(struct pool *pp, void *v)
6609 {
6610 	extern void pool_page_free(struct pool *, void *);
6611 
6612 	if ((vaddr_t)v <= last_bootstrap_page) {
6613 		*((void **)v) = free_bootstrap_pages;
6614 		free_bootstrap_pages = v;
6615 		return;
6616 	}
6617 
6618 	if (pmap_initialized) {
6619 		pool_page_free(pp, v);
6620 		return;
6621 	}
6622 }
6623 
6624 
6625 #if defined(ARM_MMU_EXTENDED)
6626 static void *
6627 pmap_l1tt_alloc(struct pool *pp, int flags)
6628 {
6629 	struct pglist plist;
6630 	vaddr_t va;
6631 
6632 	const int waitok = flags & PR_WAITOK;
6633 
6634 	int error = uvm_pglistalloc(L1TT_SIZE, 0, -1, L1TT_SIZE, 0, &plist, 1,
6635 	    waitok);
6636 	if (error)
6637 		panic("Cannot allocate L1TT physical pages, %d", error);
6638 
6639 	struct vm_page *pg = TAILQ_FIRST(&plist);
6640 #if !defined( __HAVE_MM_MD_DIRECT_MAPPED_PHYS)
6641 
6642 	/* Allocate a L1 translation table VA */
6643 	va = uvm_km_alloc(kernel_map, L1TT_SIZE, L1TT_SIZE, UVM_KMF_VAONLY);
6644 	if (va == 0)
6645 		panic("Cannot allocate L1TT KVA");
6646 
6647 	const vaddr_t eva = va + L1TT_SIZE;
6648 	vaddr_t mva = va;
6649 	while (pg && mva < eva) {
6650 		paddr_t pa = VM_PAGE_TO_PHYS(pg);
6651 
6652 		pmap_kenter_pa(mva, pa,
6653 		    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE|PMAP_PTE);
6654 
6655 		mva += PAGE_SIZE;
6656 		pg = TAILQ_NEXT(pg, pageq.queue);
6657 	}
6658 	KASSERTMSG(pg == NULL && mva == eva, "pg %p mva %" PRIxVADDR
6659 	    " eva %" PRIxVADDR, pg, mva, eva);
6660 #else
6661 	bool ok;
6662 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
6663 	va = pmap_direct_mapped_phys(pa, &ok, 0);
6664 	KASSERT(ok);
6665 	KASSERT(va >= KERNEL_BASE);
6666 #endif
6667 
6668 	return (void *)va;
6669 }
6670 
6671 static void
6672 pmap_l1tt_free(struct pool *pp, void *v)
6673 {
6674 	vaddr_t va = (vaddr_t)v;
6675 
6676 #if !defined( __HAVE_MM_MD_DIRECT_MAPPED_PHYS)
6677 	uvm_km_free(kernel_map, va, L1TT_SIZE, UVM_KMF_WIRED);
6678 #else
6679 #if defined(KERNEL_BASE_VOFFSET)
6680 	paddr_t pa = va - KERNEL_BASE_VOFFSET;
6681 #else
6682 	paddr_t pa = va - KERNEL_BASE + physical_start;
6683 #endif
6684 	const paddr_t epa = pa + L1TT_SIZE;
6685 
6686 	for (; pa < epa; pa += PAGE_SIZE) {
6687 		struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
6688 		uvm_pagefree(pg);
6689 	}
6690 #endif
6691 }
6692 #endif
6693 
6694 /*
6695  * pmap_postinit()
6696  *
6697  * This routine is called after the vm and kmem subsystems have been
6698  * initialised. This allows the pmap code to perform any initialisation
6699  * that can only be done once the memory allocation is in place.
6700  */
6701 void
6702 pmap_postinit(void)
6703 {
6704 #ifndef ARM_MMU_EXTENDED
6705 	extern paddr_t physical_start, physical_end;
6706 	struct l1_ttable *l1;
6707 	struct pglist plist;
6708 	struct vm_page *m;
6709 	pd_entry_t *pdep;
6710 	vaddr_t va, eva;
6711 	u_int loop, needed;
6712 	int error;
6713 #endif
6714 
6715 	pool_cache_setlowat(&pmap_l2ptp_cache, (PAGE_SIZE / L2_TABLE_SIZE_REAL) * 4);
6716 	pool_cache_setlowat(&pmap_l2dtable_cache,
6717 	    (PAGE_SIZE / sizeof(struct l2_dtable)) * 2);
6718 
6719 #ifndef ARM_MMU_EXTENDED
6720 	needed = (maxproc / PMAP_DOMAINS) + ((maxproc % PMAP_DOMAINS) ? 1 : 0);
6721 	needed -= 1;
6722 
6723 	l1 = kmem_alloc(sizeof(*l1) * needed, KM_SLEEP);
6724 
6725 	for (loop = 0; loop < needed; loop++, l1++) {
6726 		/* Allocate a L1 page table */
6727 		va = uvm_km_alloc(kernel_map, L1_TABLE_SIZE, 0, UVM_KMF_VAONLY);
6728 		if (va == 0)
6729 			panic("Cannot allocate L1 KVM");
6730 
6731 		error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start,
6732 		    physical_end, L1_TABLE_SIZE, 0, &plist, 1, 1);
6733 		if (error)
6734 			panic("Cannot allocate L1 physical pages");
6735 
6736 		m = TAILQ_FIRST(&plist);
6737 		eva = va + L1_TABLE_SIZE;
6738 		pdep = (pd_entry_t *)va;
6739 
6740 		while (m && va < eva) {
6741 			paddr_t pa = VM_PAGE_TO_PHYS(m);
6742 
6743 			pmap_kenter_pa(va, pa,
6744 			    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE|PMAP_PTE);
6745 
6746 			va += PAGE_SIZE;
6747 			m = TAILQ_NEXT(m, pageq.queue);
6748 		}
6749 
6750 #ifdef DIAGNOSTIC
6751 		if (m)
6752 			panic("pmap_alloc_l1pt: pglist not empty");
6753 #endif	/* DIAGNOSTIC */
6754 
6755 		pmap_init_l1(l1, pdep);
6756 	}
6757 
6758 #ifdef DEBUG
6759 	printf("pmap_postinit: Allocated %d static L1 descriptor tables\n",
6760 	    needed);
6761 #endif
6762 #endif /* !ARM_MMU_EXTENDED */
6763 }
6764 
6765 /*
6766  * Note that the following routines are used by board-specific initialisation
6767  * code to configure the initial kernel page tables.
6768  *
6769  */
6770 
6771 /*
6772  * This list exists for the benefit of pmap_map_chunk().  It keeps track
6773  * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
6774  * find them as necessary.
6775  *
6776  * Note that the data on this list MUST remain valid after initarm() returns,
6777  * as pmap_bootstrap() uses it to construct L2 table metadata.
6778  */
6779 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
6780 
6781 static vaddr_t
6782 kernel_pt_lookup(paddr_t pa)
6783 {
6784 	pv_addr_t *pv;
6785 
6786 	SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
6787 		if (pv->pv_pa == (pa & ~PGOFSET))
6788 			return pv->pv_va | (pa & PGOFSET);
6789 	}
6790 	return 0;
6791 }
6792 
6793 /*
6794  * pmap_map_section:
6795  *
6796  *	Create a single section mapping.
6797  */
6798 void
6799 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
6800 {
6801 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
6802 	const size_t l1slot = l1pte_index(va);
6803 	pd_entry_t fl;
6804 
6805 	KASSERT(((va | pa) & L1_S_OFFSET) == 0);
6806 
6807 	switch (cache) {
6808 	case PTE_NOCACHE:
6809 		fl = pte_l1_s_nocache_mode;
6810 		break;
6811 
6812 	case PTE_CACHE:
6813 		fl = pte_l1_s_cache_mode;
6814 		break;
6815 
6816 	case PTE_PAGETABLE:
6817 		fl = pte_l1_s_cache_mode_pt;
6818 		break;
6819 
6820 	case PTE_DEV:
6821 	default:
6822 		fl = 0;
6823 		break;
6824 	}
6825 
6826 	const pd_entry_t npde = L1_S_PROTO | pa |
6827 	    L1_S_PROT(PTE_KERNEL, prot) | fl | L1_S_DOM(PMAP_DOMAIN_KERNEL);
6828 	l1pte_setone(pdep + l1slot, npde);
6829 	PDE_SYNC(pdep + l1slot);
6830 }
6831 
6832 /*
6833  * pmap_map_entry:
6834  *
6835  *	Create a single page mapping.
6836  */
6837 void
6838 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
6839 {
6840 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
6841 	const size_t l1slot = l1pte_index(va);
6842 	pt_entry_t npte;
6843 	pt_entry_t *ptep;
6844 
6845 	KASSERT(((va | pa) & PGOFSET) == 0);
6846 
6847 	switch (cache) {
6848 	case PTE_NOCACHE:
6849 		npte = pte_l2_s_nocache_mode;
6850 		break;
6851 
6852 	case PTE_CACHE:
6853 		npte = pte_l2_s_cache_mode;
6854 		break;
6855 
6856 	case PTE_PAGETABLE:
6857 		npte = pte_l2_s_cache_mode_pt;
6858 		break;
6859 
6860 	default:
6861 		npte = 0;
6862 		break;
6863 	}
6864 
6865 	if ((pdep[l1slot] & L1_TYPE_MASK) != L1_TYPE_C)
6866 		panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
6867 
6868 	ptep = (pt_entry_t *) kernel_pt_lookup(l1pte_pa(pdep[l1slot]));
6869 	if (ptep == NULL)
6870 		panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
6871 
6872 	npte |= L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot);
6873 #ifdef ARM_MMU_EXTENDED
6874 	if (prot & VM_PROT_EXECUTE) {
6875 		npte &= ~L2_XS_XN;
6876 	}
6877 #endif
6878 	ptep += l2pte_index(va);
6879 	l2pte_set(ptep, npte, 0);
6880 	PTE_SYNC(ptep);
6881 }
6882 
6883 /*
6884  * pmap_link_l2pt:
6885  *
6886  *	Link the L2 page table specified by "l2pv" into the L1
6887  *	page table at the slot for "va".
6888  */
6889 void
6890 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
6891 {
6892 	pd_entry_t * const pdep = (pd_entry_t *) l1pt + l1pte_index(va);
6893 
6894 	KASSERT((va & ((L1_S_SIZE * (PAGE_SIZE / L2_T_SIZE)) - 1)) == 0);
6895 	KASSERT((l2pv->pv_pa & PGOFSET) == 0);
6896 
6897 	const pd_entry_t npde = L1_C_DOM(PMAP_DOMAIN_KERNEL) | L1_C_PROTO
6898 	    | l2pv->pv_pa;
6899 
6900 	l1pte_set(pdep, npde);
6901 	PDE_SYNC_RANGE(pdep, PAGE_SIZE / L2_T_SIZE);
6902 
6903 	SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
6904 }
6905 
6906 /*
6907  * pmap_map_chunk:
6908  *
6909  *	Map a chunk of memory using the most efficient mappings
6910  *	possible (section, large page, small page) into the
6911  *	provided L1 and L2 tables at the specified virtual address.
6912  */
6913 vsize_t
6914 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
6915     int prot, int cache)
6916 {
6917 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
6918 	pt_entry_t f1, f2s, f2l;
6919 	vsize_t resid;
6920 
6921 	resid = (size + (PAGE_SIZE - 1)) & ~(PAGE_SIZE - 1);
6922 
6923 	if (l1pt == 0)
6924 		panic("pmap_map_chunk: no L1 table provided");
6925 
6926 // 	VPRINTF("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
6927 // 	    "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
6928 
6929 	switch (cache) {
6930 	case PTE_NOCACHE:
6931 		f1 = pte_l1_s_nocache_mode;
6932 		f2l = pte_l2_l_nocache_mode;
6933 		f2s = pte_l2_s_nocache_mode;
6934 		break;
6935 
6936 	case PTE_CACHE:
6937 		f1 = pte_l1_s_cache_mode;
6938 		f2l = pte_l2_l_cache_mode;
6939 		f2s = pte_l2_s_cache_mode;
6940 		break;
6941 
6942 	case PTE_PAGETABLE:
6943 		f1 = pte_l1_s_cache_mode_pt;
6944 		f2l = pte_l2_l_cache_mode_pt;
6945 		f2s = pte_l2_s_cache_mode_pt;
6946 		break;
6947 
6948 	case PTE_DEV:
6949 	default:
6950 		f1 = 0;
6951 		f2l = 0;
6952 		f2s = 0;
6953 		break;
6954 	}
6955 
6956 	size = resid;
6957 
6958 	while (resid > 0) {
6959 		const size_t l1slot = l1pte_index(va);
6960 #ifdef ARM_MMU_EXTENDED
6961 		/* See if we can use a supersection mapping. */
6962 		if (L1_SS_PROTO && L1_SS_MAPPABLE_P(va, pa, resid)) {
6963 			/* Supersection are always domain 0 */
6964 			const pd_entry_t npde = L1_SS_PROTO | pa
6965 			    | ((prot & VM_PROT_EXECUTE) ? 0 : L1_S_V6_XN)
6966 			    | (va & 0x80000000 ? 0 : L1_S_V6_nG)
6967 			    | L1_S_PROT(PTE_KERNEL, prot) | f1;
6968 			VPRINTF("sS");
6969 			l1pte_set(&pdep[l1slot], npde);
6970 			PDE_SYNC_RANGE(&pdep[l1slot], L1_SS_SIZE / L1_S_SIZE);
6971 //			VPRINTF("\npmap_map_chunk: pa=0x%lx va=0x%lx resid=0x%08lx "
6972 //			    "npdep=%p pde=0x%x\n", pa, va, resid, &pdep[l1slot], npde);
6973 			va += L1_SS_SIZE;
6974 			pa += L1_SS_SIZE;
6975 			resid -= L1_SS_SIZE;
6976 			continue;
6977 		}
6978 #endif
6979 		/* See if we can use a section mapping. */
6980 		if (L1_S_MAPPABLE_P(va, pa, resid)) {
6981 			const pd_entry_t npde = L1_S_PROTO | pa
6982 #ifdef ARM_MMU_EXTENDED
6983 			    | ((prot & VM_PROT_EXECUTE) ? 0 : L1_S_V6_XN)
6984 			    | (va & 0x80000000 ? 0 : L1_S_V6_nG)
6985 #endif
6986 			    | L1_S_PROT(PTE_KERNEL, prot) | f1
6987 			    | L1_S_DOM(PMAP_DOMAIN_KERNEL);
6988 			VPRINTF("S");
6989 			l1pte_set(&pdep[l1slot], npde);
6990 			PDE_SYNC(&pdep[l1slot]);
6991 //			VPRINTF("\npmap_map_chunk: pa=0x%lx va=0x%lx resid=0x%08lx "
6992 //			    "npdep=%p pde=0x%x\n", pa, va, resid, &pdep[l1slot], npde);
6993 			va += L1_S_SIZE;
6994 			pa += L1_S_SIZE;
6995 			resid -= L1_S_SIZE;
6996 			continue;
6997 		}
6998 
6999 		/*
7000 		 * Ok, we're going to use an L2 table.  Make sure
7001 		 * one is actually in the corresponding L1 slot
7002 		 * for the current VA.
7003 		 */
7004 		if ((pdep[l1slot] & L1_TYPE_MASK) != L1_TYPE_C)
7005 			panic("%s: no L2 table for VA %#lx", __func__, va);
7006 
7007 		pt_entry_t *ptep = (pt_entry_t *) kernel_pt_lookup(l1pte_pa(pdep[l1slot]));
7008 		if (ptep == NULL)
7009 			panic("%s: can't find L2 table for VA %#lx", __func__,
7010 			    va);
7011 
7012 		ptep += l2pte_index(va);
7013 
7014 		/* See if we can use a L2 large page mapping. */
7015 		if (L2_L_MAPPABLE_P(va, pa, resid)) {
7016 			const pt_entry_t npte = L2_L_PROTO | pa
7017 #ifdef ARM_MMU_EXTENDED
7018 			    | ((prot & VM_PROT_EXECUTE) ? 0 : L2_XS_L_XN)
7019 			    | (va & 0x80000000 ? 0 : L2_XS_nG)
7020 #endif
7021 			    | L2_L_PROT(PTE_KERNEL, prot) | f2l;
7022 			VPRINTF("L");
7023 			l2pte_set(ptep, npte, 0);
7024 			PTE_SYNC_RANGE(ptep, L2_L_SIZE / L2_S_SIZE);
7025 			va += L2_L_SIZE;
7026 			pa += L2_L_SIZE;
7027 			resid -= L2_L_SIZE;
7028 			continue;
7029 		}
7030 
7031 		VPRINTF("P");
7032 		/* Use a small page mapping. */
7033 		pt_entry_t npte = L2_S_PROTO | pa
7034 #ifdef ARM_MMU_EXTENDED
7035 		    | ((prot & VM_PROT_EXECUTE) ? 0 : L2_XS_XN)
7036 		    | (va & 0x80000000 ? 0 : L2_XS_nG)
7037 #endif
7038 		    | L2_S_PROT(PTE_KERNEL, prot) | f2s;
7039 #ifdef ARM_MMU_EXTENDED
7040 		npte &= ((prot & VM_PROT_EXECUTE) ? ~L2_XS_XN : ~0);
7041 #endif
7042 		l2pte_set(ptep, npte, 0);
7043 		PTE_SYNC(ptep);
7044 		va += PAGE_SIZE;
7045 		pa += PAGE_SIZE;
7046 		resid -= PAGE_SIZE;
7047 	}
7048 	VPRINTF("\n");
7049 	return size;
7050 }
7051 
7052 /*
7053  * pmap_unmap_chunk:
7054  *
7055  *	Unmap a chunk of memory that was previously pmap_map_chunk
7056  */
7057 void
7058 pmap_unmap_chunk(vaddr_t l1pt, vaddr_t va, vsize_t size)
7059 {
7060 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
7061 	const size_t l1slot = l1pte_index(va);
7062 
7063 	KASSERT(size == L1_SS_SIZE || size == L1_S_SIZE);
7064 
7065 	l1pte_set(&pdep[l1slot], 0);
7066 	PDE_SYNC_RANGE(&pdep[l1slot], size / L1_S_SIZE);
7067 
7068 	pmap_tlb_flush_SE(pmap_kernel(), va, PVF_REF);
7069 }
7070 
7071 
7072 
7073 /********************** Static device map routines ***************************/
7074 
7075 static const struct pmap_devmap *pmap_devmap_table;
7076 
7077 /*
7078  * Register the devmap table.  This is provided in case early console
7079  * initialization needs to register mappings created by bootstrap code
7080  * before pmap_devmap_bootstrap() is called.
7081  */
7082 void
7083 pmap_devmap_register(const struct pmap_devmap *table)
7084 {
7085 
7086 	pmap_devmap_table = table;
7087 }
7088 
7089 /*
7090  * Map all of the static regions in the devmap table, and remember
7091  * the devmap table so other parts of the kernel can look up entries
7092  * later.
7093  */
7094 void
7095 pmap_devmap_bootstrap(vaddr_t l1pt, const struct pmap_devmap *table)
7096 {
7097 	int i;
7098 
7099 	pmap_devmap_table = table;
7100 
7101 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
7102 		const struct pmap_devmap *pdp = &pmap_devmap_table[i];
7103 
7104 		KASSERTMSG(VADDR_MAX - pdp->pd_va >= pdp->pd_size - 1, "va %" PRIxVADDR
7105 		    " sz %" PRIxPSIZE, pdp->pd_va, pdp->pd_size);
7106 		KASSERTMSG(PADDR_MAX - pdp->pd_pa >= pdp->pd_size - 1, "pa %" PRIxPADDR
7107 		    " sz %" PRIxPSIZE, pdp->pd_pa, pdp->pd_size);
7108 		VPRINTF("devmap: %08lx -> %08lx @ %08lx\n", pdp->pd_pa,
7109 		    pdp->pd_pa + pdp->pd_size - 1, pdp->pd_va);
7110 
7111 		pmap_map_chunk(l1pt, pdp->pd_va, pdp->pd_pa, pdp->pd_size,
7112 		    pdp->pd_prot, pdp->pd_cache);
7113 	}
7114 }
7115 
7116 const struct pmap_devmap *
7117 pmap_devmap_find_pa(paddr_t pa, psize_t size)
7118 {
7119 	uint64_t endpa;
7120 	int i;
7121 
7122 	if (pmap_devmap_table == NULL)
7123 		return NULL;
7124 
7125 	endpa = (uint64_t)pa + (uint64_t)(size - 1);
7126 
7127 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
7128 		if (pa >= pmap_devmap_table[i].pd_pa &&
7129 		    endpa <= (uint64_t)pmap_devmap_table[i].pd_pa +
7130 			     (uint64_t)(pmap_devmap_table[i].pd_size - 1))
7131 			return &pmap_devmap_table[i];
7132 	}
7133 
7134 	return NULL;
7135 }
7136 
7137 const struct pmap_devmap *
7138 pmap_devmap_find_va(vaddr_t va, vsize_t size)
7139 {
7140 	int i;
7141 
7142 	if (pmap_devmap_table == NULL)
7143 		return NULL;
7144 
7145 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
7146 		if (va >= pmap_devmap_table[i].pd_va &&
7147 		    va + size - 1 <= pmap_devmap_table[i].pd_va +
7148 				     pmap_devmap_table[i].pd_size - 1)
7149 			return &pmap_devmap_table[i];
7150 	}
7151 
7152 	return NULL;
7153 }
7154 
7155 /********************** PTE initialization routines **************************/
7156 
7157 /*
7158  * These routines are called when the CPU type is identified to set up
7159  * the PTE prototypes, cache modes, etc.
7160  *
7161  * The variables are always here, just in case modules need to reference
7162  * them (though, they shouldn't).
7163  */
7164 
7165 pt_entry_t	pte_l1_s_nocache_mode;
7166 pt_entry_t	pte_l1_s_cache_mode;
7167 pt_entry_t	pte_l1_s_wc_mode;
7168 pt_entry_t	pte_l1_s_cache_mode_pt;
7169 pt_entry_t	pte_l1_s_cache_mask;
7170 
7171 pt_entry_t	pte_l2_l_nocache_mode;
7172 pt_entry_t	pte_l2_l_cache_mode;
7173 pt_entry_t	pte_l2_l_wc_mode;
7174 pt_entry_t	pte_l2_l_cache_mode_pt;
7175 pt_entry_t	pte_l2_l_cache_mask;
7176 
7177 pt_entry_t	pte_l2_s_nocache_mode;
7178 pt_entry_t	pte_l2_s_cache_mode;
7179 pt_entry_t	pte_l2_s_wc_mode;
7180 pt_entry_t	pte_l2_s_cache_mode_pt;
7181 pt_entry_t	pte_l2_s_cache_mask;
7182 
7183 pt_entry_t	pte_l1_s_prot_u;
7184 pt_entry_t	pte_l1_s_prot_w;
7185 pt_entry_t	pte_l1_s_prot_ro;
7186 pt_entry_t	pte_l1_s_prot_mask;
7187 
7188 pt_entry_t	pte_l2_s_prot_u;
7189 pt_entry_t	pte_l2_s_prot_w;
7190 pt_entry_t	pte_l2_s_prot_ro;
7191 pt_entry_t	pte_l2_s_prot_mask;
7192 
7193 pt_entry_t	pte_l2_l_prot_u;
7194 pt_entry_t	pte_l2_l_prot_w;
7195 pt_entry_t	pte_l2_l_prot_ro;
7196 pt_entry_t	pte_l2_l_prot_mask;
7197 
7198 pt_entry_t	pte_l1_ss_proto;
7199 pt_entry_t	pte_l1_s_proto;
7200 pt_entry_t	pte_l1_c_proto;
7201 pt_entry_t	pte_l2_s_proto;
7202 
7203 void		(*pmap_copy_page_func)(paddr_t, paddr_t);
7204 void		(*pmap_zero_page_func)(paddr_t);
7205 
7206 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
7207 void
7208 pmap_pte_init_generic(void)
7209 {
7210 
7211 	pte_l1_s_nocache_mode = 0;
7212 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
7213 	pte_l1_s_wc_mode = L1_S_B;
7214 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_generic;
7215 
7216 	pte_l2_l_nocache_mode = 0;
7217 	pte_l2_l_cache_mode = L2_B|L2_C;
7218 	pte_l2_l_wc_mode = L2_B;
7219 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_generic;
7220 
7221 	pte_l2_s_nocache_mode = 0;
7222 	pte_l2_s_cache_mode = L2_B|L2_C;
7223 	pte_l2_s_wc_mode = L2_B;
7224 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_generic;
7225 
7226 	/*
7227 	 * If we have a write-through cache, set B and C.  If
7228 	 * we have a write-back cache, then we assume setting
7229 	 * only C will make those pages write-through (except for those
7230 	 * Cortex CPUs which can read the L1 caches).
7231 	 */
7232 	if (cpufuncs.cf_dcache_wb_range == (void *) cpufunc_nullop
7233 #if ARM_MMU_V7 > 0
7234 	    || CPU_ID_CORTEX_P(curcpu()->ci_arm_cpuid)
7235 #endif
7236 #if ARM_MMU_V6 > 0
7237 	    || CPU_ID_ARM11_P(curcpu()->ci_arm_cpuid) /* arm116 errata 399234 */
7238 #endif
7239 	    || false) {
7240 		pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
7241 		pte_l2_l_cache_mode_pt = L2_B|L2_C;
7242 		pte_l2_s_cache_mode_pt = L2_B|L2_C;
7243 	} else {
7244 		pte_l1_s_cache_mode_pt = L1_S_C;	/* write through */
7245 		pte_l2_l_cache_mode_pt = L2_C;		/* write through */
7246 		pte_l2_s_cache_mode_pt = L2_C;		/* write through */
7247 	}
7248 
7249 	pte_l1_s_prot_u = L1_S_PROT_U_generic;
7250 	pte_l1_s_prot_w = L1_S_PROT_W_generic;
7251 	pte_l1_s_prot_ro = L1_S_PROT_RO_generic;
7252 	pte_l1_s_prot_mask = L1_S_PROT_MASK_generic;
7253 
7254 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
7255 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
7256 	pte_l2_s_prot_ro = L2_S_PROT_RO_generic;
7257 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
7258 
7259 	pte_l2_l_prot_u = L2_L_PROT_U_generic;
7260 	pte_l2_l_prot_w = L2_L_PROT_W_generic;
7261 	pte_l2_l_prot_ro = L2_L_PROT_RO_generic;
7262 	pte_l2_l_prot_mask = L2_L_PROT_MASK_generic;
7263 
7264 	pte_l1_ss_proto = L1_SS_PROTO_generic;
7265 	pte_l1_s_proto = L1_S_PROTO_generic;
7266 	pte_l1_c_proto = L1_C_PROTO_generic;
7267 	pte_l2_s_proto = L2_S_PROTO_generic;
7268 
7269 	pmap_copy_page_func = pmap_copy_page_generic;
7270 	pmap_zero_page_func = pmap_zero_page_generic;
7271 }
7272 
7273 #if defined(CPU_ARM8)
7274 void
7275 pmap_pte_init_arm8(void)
7276 {
7277 
7278 	/*
7279 	 * ARM8 is compatible with generic, but we need to use
7280 	 * the page tables uncached.
7281 	 */
7282 	pmap_pte_init_generic();
7283 
7284 	pte_l1_s_cache_mode_pt = 0;
7285 	pte_l2_l_cache_mode_pt = 0;
7286 	pte_l2_s_cache_mode_pt = 0;
7287 }
7288 #endif /* CPU_ARM8 */
7289 
7290 #if defined(CPU_ARM9) && defined(ARM9_CACHE_WRITE_THROUGH)
7291 void
7292 pmap_pte_init_arm9(void)
7293 {
7294 
7295 	/*
7296 	 * ARM9 is compatible with generic, but we want to use
7297 	 * write-through caching for now.
7298 	 */
7299 	pmap_pte_init_generic();
7300 
7301 	pte_l1_s_cache_mode = L1_S_C;
7302 	pte_l2_l_cache_mode = L2_C;
7303 	pte_l2_s_cache_mode = L2_C;
7304 
7305 	pte_l1_s_wc_mode = L1_S_B;
7306 	pte_l2_l_wc_mode = L2_B;
7307 	pte_l2_s_wc_mode = L2_B;
7308 
7309 	pte_l1_s_cache_mode_pt = L1_S_C;
7310 	pte_l2_l_cache_mode_pt = L2_C;
7311 	pte_l2_s_cache_mode_pt = L2_C;
7312 }
7313 #endif /* CPU_ARM9 && ARM9_CACHE_WRITE_THROUGH */
7314 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
7315 
7316 #if defined(CPU_ARM10)
7317 void
7318 pmap_pte_init_arm10(void)
7319 {
7320 
7321 	/*
7322 	 * ARM10 is compatible with generic, but we want to use
7323 	 * write-through caching for now.
7324 	 */
7325 	pmap_pte_init_generic();
7326 
7327 	pte_l1_s_cache_mode = L1_S_B | L1_S_C;
7328 	pte_l2_l_cache_mode = L2_B | L2_C;
7329 	pte_l2_s_cache_mode = L2_B | L2_C;
7330 
7331 	pte_l1_s_cache_mode = L1_S_B;
7332 	pte_l2_l_cache_mode = L2_B;
7333 	pte_l2_s_cache_mode = L2_B;
7334 
7335 	pte_l1_s_cache_mode_pt = L1_S_C;
7336 	pte_l2_l_cache_mode_pt = L2_C;
7337 	pte_l2_s_cache_mode_pt = L2_C;
7338 
7339 }
7340 #endif /* CPU_ARM10 */
7341 
7342 #if defined(CPU_ARM11) && defined(ARM11_CACHE_WRITE_THROUGH)
7343 void
7344 pmap_pte_init_arm11(void)
7345 {
7346 
7347 	/*
7348 	 * ARM11 is compatible with generic, but we want to use
7349 	 * write-through caching for now.
7350 	 */
7351 	pmap_pte_init_generic();
7352 
7353 	pte_l1_s_cache_mode = L1_S_C;
7354 	pte_l2_l_cache_mode = L2_C;
7355 	pte_l2_s_cache_mode = L2_C;
7356 
7357 	pte_l1_s_wc_mode = L1_S_B;
7358 	pte_l2_l_wc_mode = L2_B;
7359 	pte_l2_s_wc_mode = L2_B;
7360 
7361 	pte_l1_s_cache_mode_pt = L1_S_C;
7362 	pte_l2_l_cache_mode_pt = L2_C;
7363 	pte_l2_s_cache_mode_pt = L2_C;
7364 }
7365 #endif /* CPU_ARM11 && ARM11_CACHE_WRITE_THROUGH */
7366 
7367 #if ARM_MMU_SA1 == 1
7368 void
7369 pmap_pte_init_sa1(void)
7370 {
7371 
7372 	/*
7373 	 * The StrongARM SA-1 cache does not have a write-through
7374 	 * mode.  So, do the generic initialization, then reset
7375 	 * the page table cache mode to B=1,C=1, and note that
7376 	 * the PTEs need to be sync'd.
7377 	 */
7378 	pmap_pte_init_generic();
7379 
7380 	pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
7381 	pte_l2_l_cache_mode_pt = L2_B|L2_C;
7382 	pte_l2_s_cache_mode_pt = L2_B|L2_C;
7383 
7384 	pmap_needs_pte_sync = 1;
7385 }
7386 #endif /* ARM_MMU_SA1 == 1*/
7387 
7388 #if ARM_MMU_XSCALE == 1
7389 #if (ARM_NMMUS > 1)
7390 static u_int xscale_use_minidata;
7391 #endif
7392 
7393 void
7394 pmap_pte_init_xscale(void)
7395 {
7396 	uint32_t auxctl;
7397 	int write_through = 0;
7398 
7399 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
7400 	pte_l1_s_wc_mode = L1_S_B;
7401 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_xscale;
7402 
7403 	pte_l2_l_cache_mode = L2_B|L2_C;
7404 	pte_l2_l_wc_mode = L2_B;
7405 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_xscale;
7406 
7407 	pte_l2_s_cache_mode = L2_B|L2_C;
7408 	pte_l2_s_wc_mode = L2_B;
7409 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_xscale;
7410 
7411 	pte_l1_s_cache_mode_pt = L1_S_C;
7412 	pte_l2_l_cache_mode_pt = L2_C;
7413 	pte_l2_s_cache_mode_pt = L2_C;
7414 
7415 #ifdef XSCALE_CACHE_READ_WRITE_ALLOCATE
7416 	/*
7417 	 * The XScale core has an enhanced mode where writes that
7418 	 * miss the cache cause a cache line to be allocated.  This
7419 	 * is significantly faster than the traditional, write-through
7420 	 * behavior of this case.
7421 	 */
7422 	pte_l1_s_cache_mode |= L1_S_XS_TEX(TEX_XSCALE_X);
7423 	pte_l2_l_cache_mode |= L2_XS_L_TEX(TEX_XSCALE_X);
7424 	pte_l2_s_cache_mode |= L2_XS_T_TEX(TEX_XSCALE_X);
7425 #endif /* XSCALE_CACHE_READ_WRITE_ALLOCATE */
7426 
7427 #ifdef XSCALE_CACHE_WRITE_THROUGH
7428 	/*
7429 	 * Some versions of the XScale core have various bugs in
7430 	 * their cache units, the work-around for which is to run
7431 	 * the cache in write-through mode.  Unfortunately, this
7432 	 * has a major (negative) impact on performance.  So, we
7433 	 * go ahead and run fast-and-loose, in the hopes that we
7434 	 * don't line up the planets in a way that will trip the
7435 	 * bugs.
7436 	 *
7437 	 * However, we give you the option to be slow-but-correct.
7438 	 */
7439 	write_through = 1;
7440 #elif defined(XSCALE_CACHE_WRITE_BACK)
7441 	/* force write back cache mode */
7442 	write_through = 0;
7443 #elif defined(CPU_XSCALE_PXA250) || defined(CPU_XSCALE_PXA270)
7444 	/*
7445 	 * Intel PXA2[15]0 processors are known to have a bug in
7446 	 * write-back cache on revision 4 and earlier (stepping
7447 	 * A[01] and B[012]).  Fixed for C0 and later.
7448 	 */
7449 	{
7450 		uint32_t id, type;
7451 
7452 		id = cpufunc_id();
7453 		type = id & ~(CPU_ID_XSCALE_COREREV_MASK|CPU_ID_REVISION_MASK);
7454 
7455 		if (type == CPU_ID_PXA250 || type == CPU_ID_PXA210) {
7456 			if ((id & CPU_ID_REVISION_MASK) < 5) {
7457 				/* write through for stepping A0-1 and B0-2 */
7458 				write_through = 1;
7459 			}
7460 		}
7461 	}
7462 #endif /* XSCALE_CACHE_WRITE_THROUGH */
7463 
7464 	if (write_through) {
7465 		pte_l1_s_cache_mode = L1_S_C;
7466 		pte_l2_l_cache_mode = L2_C;
7467 		pte_l2_s_cache_mode = L2_C;
7468 	}
7469 
7470 #if (ARM_NMMUS > 1)
7471 	xscale_use_minidata = 1;
7472 #endif
7473 
7474 	pte_l1_s_prot_u = L1_S_PROT_U_xscale;
7475 	pte_l1_s_prot_w = L1_S_PROT_W_xscale;
7476 	pte_l1_s_prot_ro = L1_S_PROT_RO_xscale;
7477 	pte_l1_s_prot_mask = L1_S_PROT_MASK_xscale;
7478 
7479 	pte_l2_s_prot_u = L2_S_PROT_U_xscale;
7480 	pte_l2_s_prot_w = L2_S_PROT_W_xscale;
7481 	pte_l2_s_prot_ro = L2_S_PROT_RO_xscale;
7482 	pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale;
7483 
7484 	pte_l2_l_prot_u = L2_L_PROT_U_xscale;
7485 	pte_l2_l_prot_w = L2_L_PROT_W_xscale;
7486 	pte_l2_l_prot_ro = L2_L_PROT_RO_xscale;
7487 	pte_l2_l_prot_mask = L2_L_PROT_MASK_xscale;
7488 
7489 	pte_l1_ss_proto = L1_SS_PROTO_xscale;
7490 	pte_l1_s_proto = L1_S_PROTO_xscale;
7491 	pte_l1_c_proto = L1_C_PROTO_xscale;
7492 	pte_l2_s_proto = L2_S_PROTO_xscale;
7493 
7494 	pmap_copy_page_func = pmap_copy_page_xscale;
7495 	pmap_zero_page_func = pmap_zero_page_xscale;
7496 
7497 	/*
7498 	 * Disable ECC protection of page table access, for now.
7499 	 */
7500 	auxctl = armreg_auxctl_read();
7501 	auxctl &= ~XSCALE_AUXCTL_P;
7502 	armreg_auxctl_write(auxctl);
7503 }
7504 
7505 /*
7506  * xscale_setup_minidata:
7507  *
7508  *	Set up the mini-data cache clean area.  We require the
7509  *	caller to allocate the right amount of physically and
7510  *	virtually contiguous space.
7511  */
7512 void
7513 xscale_setup_minidata(vaddr_t l1pt, vaddr_t va, paddr_t pa)
7514 {
7515 	extern vaddr_t xscale_minidata_clean_addr;
7516 	extern vsize_t xscale_minidata_clean_size; /* already initialized */
7517 	pd_entry_t *pde = (pd_entry_t *) l1pt;
7518 	vsize_t size;
7519 	uint32_t auxctl;
7520 
7521 	xscale_minidata_clean_addr = va;
7522 
7523 	/* Round it to page size. */
7524 	size = (xscale_minidata_clean_size + L2_S_OFFSET) & L2_S_FRAME;
7525 
7526 	for (; size != 0;
7527 	     va += L2_S_SIZE, pa += L2_S_SIZE, size -= L2_S_SIZE) {
7528 		const size_t l1slot = l1pte_index(va);
7529 		pt_entry_t *ptep = (pt_entry_t *) kernel_pt_lookup(l1pte_pa(pde[l1slot]));
7530 		if (ptep == NULL)
7531 			panic("xscale_setup_minidata: can't find L2 table for "
7532 			    "VA 0x%08lx", va);
7533 
7534 		ptep += l2pte_index(va);
7535 		pt_entry_t opte = *ptep;
7536 		l2pte_set(ptep,
7537 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, VM_PROT_READ)
7538 		    | L2_C | L2_XS_T_TEX(TEX_XSCALE_X), opte);
7539 	}
7540 
7541 	/*
7542 	 * Configure the mini-data cache for write-back with
7543 	 * read/write-allocate.
7544 	 *
7545 	 * NOTE: In order to reconfigure the mini-data cache, we must
7546 	 * make sure it contains no valid data!  In order to do that,
7547 	 * we must issue a global data cache invalidate command!
7548 	 *
7549 	 * WE ASSUME WE ARE RUNNING UN-CACHED WHEN THIS ROUTINE IS CALLED!
7550 	 * THIS IS VERY IMPORTANT!
7551 	 */
7552 
7553 	/* Invalidate data and mini-data. */
7554 	__asm volatile("mcr p15, 0, %0, c7, c6, 0" : : "r" (0));
7555 	auxctl = armreg_auxctl_read();
7556 	auxctl = (auxctl & ~XSCALE_AUXCTL_MD_MASK) | XSCALE_AUXCTL_MD_WB_RWA;
7557 	armreg_auxctl_write(auxctl);
7558 }
7559 
7560 /*
7561  * Change the PTEs for the specified kernel mappings such that they
7562  * will use the mini data cache instead of the main data cache.
7563  */
7564 void
7565 pmap_uarea(vaddr_t va)
7566 {
7567 	vaddr_t next_bucket, eva;
7568 
7569 #if (ARM_NMMUS > 1)
7570 	if (xscale_use_minidata == 0)
7571 		return;
7572 #endif
7573 
7574 	eva = va + USPACE;
7575 
7576 	while (va < eva) {
7577 		next_bucket = L2_NEXT_BUCKET_VA(va);
7578 		if (next_bucket > eva)
7579 			next_bucket = eva;
7580 
7581 		struct l2_bucket *l2b = pmap_get_l2_bucket(pmap_kernel(), va);
7582 		KDASSERT(l2b != NULL);
7583 
7584 		pt_entry_t * const sptep = &l2b->l2b_kva[l2pte_index(va)];
7585 		pt_entry_t *ptep = sptep;
7586 
7587 		while (va < next_bucket) {
7588 			const pt_entry_t opte = *ptep;
7589 			if (!l2pte_minidata_p(opte)) {
7590 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
7591 				cpu_tlb_flushD_SE(va);
7592 				l2pte_set(ptep, opte & ~L2_B, opte);
7593 			}
7594 			ptep += PAGE_SIZE / L2_S_SIZE;
7595 			va += PAGE_SIZE;
7596 		}
7597 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
7598 	}
7599 	cpu_cpwait();
7600 }
7601 #endif /* ARM_MMU_XSCALE == 1 */
7602 
7603 
7604 #if defined(CPU_ARM11MPCORE)
7605 void
7606 pmap_pte_init_arm11mpcore(void)
7607 {
7608 
7609 	/* cache mode is controlled by 5 bits (B, C, TEX[2:0]) */
7610 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_armv6;
7611 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_armv6;
7612 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7613 	/* use extended small page (without APn, with TEX) */
7614 	pte_l2_s_cache_mask = L2_XS_CACHE_MASK_armv6;
7615 #else
7616 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_armv6c;
7617 #endif
7618 
7619 	/* write-back, write-allocate */
7620 	pte_l1_s_cache_mode = L1_S_C | L1_S_B | L1_S_V6_TEX(0x01);
7621 	pte_l2_l_cache_mode = L2_C | L2_B | L2_V6_L_TEX(0x01);
7622 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7623 	pte_l2_s_cache_mode = L2_C | L2_B | L2_V6_XS_TEX(0x01);
7624 #else
7625 	/* no TEX. read-allocate */
7626 	pte_l2_s_cache_mode = L2_C | L2_B;
7627 #endif
7628 	/*
7629 	 * write-back, write-allocate for page tables.
7630 	 */
7631 	pte_l1_s_cache_mode_pt = L1_S_C | L1_S_B | L1_S_V6_TEX(0x01);
7632 	pte_l2_l_cache_mode_pt = L2_C | L2_B | L2_V6_L_TEX(0x01);
7633 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7634 	pte_l2_s_cache_mode_pt = L2_C | L2_B | L2_V6_XS_TEX(0x01);
7635 #else
7636 	pte_l2_s_cache_mode_pt = L2_C | L2_B;
7637 #endif
7638 
7639 	pte_l1_s_prot_u = L1_S_PROT_U_armv6;
7640 	pte_l1_s_prot_w = L1_S_PROT_W_armv6;
7641 	pte_l1_s_prot_ro = L1_S_PROT_RO_armv6;
7642 	pte_l1_s_prot_mask = L1_S_PROT_MASK_armv6;
7643 
7644 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7645 	pte_l2_s_prot_u = L2_S_PROT_U_armv6n;
7646 	pte_l2_s_prot_w = L2_S_PROT_W_armv6n;
7647 	pte_l2_s_prot_ro = L2_S_PROT_RO_armv6n;
7648 	pte_l2_s_prot_mask = L2_S_PROT_MASK_armv6n;
7649 
7650 #else
7651 	/* with AP[0..3] */
7652 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
7653 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
7654 	pte_l2_s_prot_ro = L2_S_PROT_RO_generic;
7655 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
7656 #endif
7657 
7658 #ifdef	ARM11MPCORE_COMPAT_MMU
7659 	/* with AP[0..3] */
7660 	pte_l2_l_prot_u = L2_L_PROT_U_generic;
7661 	pte_l2_l_prot_w = L2_L_PROT_W_generic;
7662 	pte_l2_l_prot_ro = L2_L_PROT_RO_generic;
7663 	pte_l2_l_prot_mask = L2_L_PROT_MASK_generic;
7664 
7665 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
7666 	pte_l1_s_proto = L1_S_PROTO_armv6;
7667 	pte_l1_c_proto = L1_C_PROTO_armv6;
7668 	pte_l2_s_proto = L2_S_PROTO_armv6c;
7669 #else
7670 	pte_l2_l_prot_u = L2_L_PROT_U_armv6n;
7671 	pte_l2_l_prot_w = L2_L_PROT_W_armv6n;
7672 	pte_l2_l_prot_ro = L2_L_PROT_RO_armv6n;
7673 	pte_l2_l_prot_mask = L2_L_PROT_MASK_armv6n;
7674 
7675 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
7676 	pte_l1_s_proto = L1_S_PROTO_armv6;
7677 	pte_l1_c_proto = L1_C_PROTO_armv6;
7678 	pte_l2_s_proto = L2_S_PROTO_armv6n;
7679 #endif
7680 
7681 	pmap_copy_page_func = pmap_copy_page_generic;
7682 	pmap_zero_page_func = pmap_zero_page_generic;
7683 	pmap_needs_pte_sync = 1;
7684 }
7685 #endif	/* CPU_ARM11MPCORE */
7686 
7687 
7688 #if ARM_MMU_V6 == 1
7689 void
7690 pmap_pte_init_armv6(void)
7691 {
7692 	/*
7693 	 * The ARMv6-A MMU is mostly compatible with generic. If the
7694 	 * AP field is zero, that now means "no access" rather than
7695 	 * read-only. The prototypes are a little different because of
7696 	 * the XN bit.
7697 	 */
7698 	pmap_pte_init_generic();
7699 
7700 	pte_l1_s_nocache_mode = L1_S_XS_TEX(1);
7701 	pte_l2_l_nocache_mode = L2_XS_L_TEX(1);
7702 	pte_l2_s_nocache_mode = L2_XS_T_TEX(1);
7703 
7704 #ifdef ARM11_COMPAT_MMU
7705 	/* with AP[0..3] */
7706 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
7707 #else
7708 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_armv6n;
7709 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_armv6n;
7710 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_armv6n;
7711 
7712 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
7713 	pte_l1_s_proto = L1_S_PROTO_armv6;
7714 	pte_l1_c_proto = L1_C_PROTO_armv6;
7715 	pte_l2_s_proto = L2_S_PROTO_armv6n;
7716 
7717 	pte_l1_s_prot_u = L1_S_PROT_U_armv6;
7718 	pte_l1_s_prot_w = L1_S_PROT_W_armv6;
7719 	pte_l1_s_prot_ro = L1_S_PROT_RO_armv6;
7720 	pte_l1_s_prot_mask = L1_S_PROT_MASK_armv6;
7721 
7722 	pte_l2_l_prot_u = L2_L_PROT_U_armv6n;
7723 	pte_l2_l_prot_w = L2_L_PROT_W_armv6n;
7724 	pte_l2_l_prot_ro = L2_L_PROT_RO_armv6n;
7725 	pte_l2_l_prot_mask = L2_L_PROT_MASK_armv6n;
7726 
7727 	pte_l2_s_prot_u = L2_S_PROT_U_armv6n;
7728 	pte_l2_s_prot_w = L2_S_PROT_W_armv6n;
7729 	pte_l2_s_prot_ro = L2_S_PROT_RO_armv6n;
7730 	pte_l2_s_prot_mask = L2_S_PROT_MASK_armv6n;
7731 
7732 #endif
7733 }
7734 #endif /* ARM_MMU_V6 */
7735 
7736 #if ARM_MMU_V7 == 1
7737 void
7738 pmap_pte_init_armv7(void)
7739 {
7740 	/*
7741 	 * The ARMv7-A MMU is mostly compatible with generic. If the
7742 	 * AP field is zero, that now means "no access" rather than
7743 	 * read-only. The prototypes are a little different because of
7744 	 * the XN bit.
7745 	 */
7746 	pmap_pte_init_generic();
7747 
7748 	pmap_needs_pte_sync = 1;
7749 
7750 	pte_l1_s_nocache_mode = L1_S_XS_TEX(1);
7751 	pte_l2_l_nocache_mode = L2_XS_L_TEX(1);
7752 	pte_l2_s_nocache_mode = L2_XS_T_TEX(1);
7753 
7754 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_armv7;
7755 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_armv7;
7756 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_armv7;
7757 
7758 	/*
7759 	 * If the core support coherent walk then updates to translation tables
7760 	 * do not require a clean to the point of unification to ensure
7761 	 * visibility by subsequent translation table walks.  That means we can
7762 	 * map everything shareable and cached and the right thing will happen.
7763 	 */
7764         if (__SHIFTOUT(armreg_mmfr3_read(), __BITS(23,20))) {
7765 		pmap_needs_pte_sync = 0;
7766 
7767 		/*
7768 		 * write-back, no write-allocate, shareable for normal pages.
7769 		 */
7770 		pte_l1_s_cache_mode |= L1_S_V6_S;
7771 		pte_l2_l_cache_mode |= L2_XS_S;
7772 		pte_l2_s_cache_mode |= L2_XS_S;
7773 	}
7774 
7775 	/*
7776 	 * Page tables are just all other memory.  We can use write-back since
7777 	 * pmap_needs_pte_sync is 1 (or the MMU can read out of cache).
7778 	 */
7779 	pte_l1_s_cache_mode_pt = pte_l1_s_cache_mode;
7780 	pte_l2_l_cache_mode_pt = pte_l2_l_cache_mode;
7781 	pte_l2_s_cache_mode_pt = pte_l2_s_cache_mode;
7782 
7783 	/*
7784 	 * Check the Memory Model Features to see if this CPU supports
7785 	 * the TLBIASID coproc op.
7786 	 */
7787 	if (__SHIFTOUT(armreg_mmfr2_read(), __BITS(16,19)) >= 2) {
7788 		arm_has_tlbiasid_p = true;
7789 	} else if (__SHIFTOUT(armreg_mmfr2_read(), __BITS(12,15)) >= 2) {
7790 		arm_has_tlbiasid_p = true;
7791 	}
7792 
7793 	/*
7794 	 * Check the MPIDR to see if this CPU supports MP extensions.
7795 	 */
7796 #ifdef MULTIPROCESSOR
7797 	arm_has_mpext_p = (armreg_mpidr_read() & (MPIDR_MP|MPIDR_U)) == MPIDR_MP;
7798 #else
7799 	arm_has_mpext_p = false;
7800 #endif
7801 
7802 	pte_l1_s_prot_u = L1_S_PROT_U_armv7;
7803 	pte_l1_s_prot_w = L1_S_PROT_W_armv7;
7804 	pte_l1_s_prot_ro = L1_S_PROT_RO_armv7;
7805 	pte_l1_s_prot_mask = L1_S_PROT_MASK_armv7;
7806 
7807 	pte_l2_s_prot_u = L2_S_PROT_U_armv7;
7808 	pte_l2_s_prot_w = L2_S_PROT_W_armv7;
7809 	pte_l2_s_prot_ro = L2_S_PROT_RO_armv7;
7810 	pte_l2_s_prot_mask = L2_S_PROT_MASK_armv7;
7811 
7812 	pte_l2_l_prot_u = L2_L_PROT_U_armv7;
7813 	pte_l2_l_prot_w = L2_L_PROT_W_armv7;
7814 	pte_l2_l_prot_ro = L2_L_PROT_RO_armv7;
7815 	pte_l2_l_prot_mask = L2_L_PROT_MASK_armv7;
7816 
7817 	pte_l1_ss_proto = L1_SS_PROTO_armv7;
7818 	pte_l1_s_proto = L1_S_PROTO_armv7;
7819 	pte_l1_c_proto = L1_C_PROTO_armv7;
7820 	pte_l2_s_proto = L2_S_PROTO_armv7;
7821 
7822 }
7823 #endif /* ARM_MMU_V7 */
7824 
7825 /*
7826  * return the PA of the current L1 table, for use when handling a crash dump
7827  */
7828 uint32_t
7829 pmap_kernel_L1_addr(void)
7830 {
7831 #ifdef ARM_MMU_EXTENDED
7832 	return pmap_kernel()->pm_l1_pa;
7833 #else
7834 	return pmap_kernel()->pm_l1->l1_physaddr;
7835 #endif
7836 }
7837 
7838 #if defined(DDB)
7839 /*
7840  * A couple of ddb-callable functions for dumping pmaps
7841  */
7842 void pmap_dump(pmap_t);
7843 
7844 static pt_entry_t ncptes[64];
7845 static void pmap_dump_ncpg(pmap_t);
7846 
7847 void
7848 pmap_dump(pmap_t pm)
7849 {
7850 	struct l2_dtable *l2;
7851 	struct l2_bucket *l2b;
7852 	pt_entry_t *ptep, pte;
7853 	vaddr_t l2_va, l2b_va, va;
7854 	int i, j, k, occ, rows = 0;
7855 
7856 	if (pm == pmap_kernel())
7857 		printf("pmap_kernel (%p): ", pm);
7858 	else
7859 		printf("user pmap (%p): ", pm);
7860 
7861 #ifdef ARM_MMU_EXTENDED
7862 	printf("l1 at %p\n", pmap_l1_kva(pm));
7863 #else
7864 	printf("domain %d, l1 at %p\n", pmap_domain(pm), pmap_l1_kva(pm));
7865 #endif
7866 
7867 	l2_va = 0;
7868 	for (i = 0; i < L2_SIZE; i++, l2_va += 0x01000000) {
7869 		l2 = pm->pm_l2[i];
7870 
7871 		if (l2 == NULL || l2->l2_occupancy == 0)
7872 			continue;
7873 
7874 		l2b_va = l2_va;
7875 		for (j = 0; j < L2_BUCKET_SIZE; j++, l2b_va += 0x00100000) {
7876 			l2b = &l2->l2_bucket[j];
7877 
7878 			if (l2b->l2b_occupancy == 0 || l2b->l2b_kva == NULL)
7879 				continue;
7880 
7881 			ptep = l2b->l2b_kva;
7882 
7883 			for (k = 0; k < 256 && ptep[k] == 0; k++)
7884 				;
7885 
7886 			k &= ~63;
7887 			occ = l2b->l2b_occupancy;
7888 			va = l2b_va + (k * 4096);
7889 			for (; k < 256; k++, va += 0x1000) {
7890 				char ch = ' ';
7891 				if ((k % 64) == 0) {
7892 					if ((rows % 8) == 0) {
7893 						printf(
7894 "          |0000   |8000   |10000  |18000  |20000  |28000  |30000  |38000\n");
7895 					}
7896 					printf("%08lx: ", va);
7897 				}
7898 
7899 				ncptes[k & 63] = 0;
7900 				pte = ptep[k];
7901 				if (pte == 0) {
7902 					ch = '.';
7903 				} else {
7904 					occ--;
7905 					switch (pte & 0x4c) {
7906 					case 0x00:
7907 						ch = 'N'; /* No cache No buff */
7908 						break;
7909 					case 0x04:
7910 						ch = 'B'; /* No cache buff */
7911 						break;
7912 					case 0x08:
7913 						ch = 'C'; /* Cache No buff */
7914 						break;
7915 					case 0x0c:
7916 						ch = 'F'; /* Cache Buff */
7917 						break;
7918 					case 0x40:
7919 						ch = 'D';
7920 						break;
7921 					case 0x48:
7922 						ch = 'm'; /* Xscale mini-data */
7923 						break;
7924 					default:
7925 						ch = '?';
7926 						break;
7927 					}
7928 
7929 					if ((pte & L2_S_PROT_U) == L2_S_PROT_U)
7930 						ch += 0x20;
7931 
7932 					if ((pte & 0xc) == 0)
7933 						ncptes[k & 63] = pte;
7934 				}
7935 
7936 				if ((k % 64) == 63) {
7937 					rows++;
7938 					printf("%c\n", ch);
7939 					pmap_dump_ncpg(pm);
7940 					if (occ == 0)
7941 						break;
7942 				} else
7943 					printf("%c", ch);
7944 			}
7945 		}
7946 	}
7947 }
7948 
7949 static void
7950 pmap_dump_ncpg(pmap_t pm)
7951 {
7952 	struct vm_page *pg;
7953 	struct vm_page_md *md;
7954 	struct pv_entry *pv;
7955 	int i;
7956 
7957 	for (i = 0; i < 63; i++) {
7958 		if (ncptes[i] == 0)
7959 			continue;
7960 
7961 		pg = PHYS_TO_VM_PAGE(l2pte_pa(ncptes[i]));
7962 		if (pg == NULL)
7963 			continue;
7964 		md = VM_PAGE_TO_MD(pg);
7965 
7966 		printf(" pa 0x%08lx: krw %d kro %d urw %d uro %d\n",
7967 		    VM_PAGE_TO_PHYS(pg),
7968 		    md->krw_mappings, md->kro_mappings,
7969 		    md->urw_mappings, md->uro_mappings);
7970 
7971 		SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
7972 			printf("   %c va 0x%08lx, flags 0x%x\n",
7973 			    (pm == pv->pv_pmap) ? '*' : ' ',
7974 			    pv->pv_va, pv->pv_flags);
7975 		}
7976 	}
7977 }
7978 #endif
7979 
7980 #ifdef PMAP_STEAL_MEMORY
7981 void
7982 pmap_boot_pageadd(pv_addr_t *newpv)
7983 {
7984 	pv_addr_t *pv, *npv;
7985 
7986 	if ((pv = SLIST_FIRST(&pmap_boot_freeq)) != NULL) {
7987 		if (newpv->pv_pa < pv->pv_va) {
7988 			KASSERT(newpv->pv_pa + newpv->pv_size <= pv->pv_pa);
7989 			if (newpv->pv_pa + newpv->pv_size == pv->pv_pa) {
7990 				newpv->pv_size += pv->pv_size;
7991 				SLIST_REMOVE_HEAD(&pmap_boot_freeq, pv_list);
7992 			}
7993 			pv = NULL;
7994 		} else {
7995 			for (; (npv = SLIST_NEXT(pv, pv_list)) != NULL;
7996 			     pv = npv) {
7997 				KASSERT(pv->pv_pa + pv->pv_size < npv->pv_pa);
7998 				KASSERT(pv->pv_pa < newpv->pv_pa);
7999 				if (newpv->pv_pa > npv->pv_pa)
8000 					continue;
8001 				if (pv->pv_pa + pv->pv_size == newpv->pv_pa) {
8002 					pv->pv_size += newpv->pv_size;
8003 					return;
8004 				}
8005 				if (newpv->pv_pa + newpv->pv_size < npv->pv_pa)
8006 					break;
8007 				newpv->pv_size += npv->pv_size;
8008 				SLIST_INSERT_AFTER(pv, newpv, pv_list);
8009 				SLIST_REMOVE_AFTER(newpv, pv_list);
8010 				return;
8011 			}
8012 		}
8013 	}
8014 
8015 	if (pv) {
8016 		SLIST_INSERT_AFTER(pv, newpv, pv_list);
8017 	} else {
8018 		SLIST_INSERT_HEAD(&pmap_boot_freeq, newpv, pv_list);
8019 	}
8020 }
8021 
8022 void
8023 pmap_boot_pagealloc(psize_t amount, psize_t mask, psize_t match,
8024 	pv_addr_t *rpv)
8025 {
8026 	pv_addr_t *pv, **pvp;
8027 
8028 	KASSERT(amount & PGOFSET);
8029 	KASSERT((mask & PGOFSET) == 0);
8030 	KASSERT((match & PGOFSET) == 0);
8031 	KASSERT(amount != 0);
8032 
8033 	for (pvp = &SLIST_FIRST(&pmap_boot_freeq);
8034 	     (pv = *pvp) != NULL;
8035 	     pvp = &SLIST_NEXT(pv, pv_list)) {
8036 		pv_addr_t *newpv;
8037 		psize_t off;
8038 		/*
8039 		 * If this entry is too small to satisfy the request...
8040 		 */
8041 		KASSERT(pv->pv_size > 0);
8042 		if (pv->pv_size < amount)
8043 			continue;
8044 
8045 		for (off = 0; off <= mask; off += PAGE_SIZE) {
8046 			if (((pv->pv_pa + off) & mask) == match
8047 			    && off + amount <= pv->pv_size)
8048 				break;
8049 		}
8050 		if (off > mask)
8051 			continue;
8052 
8053 		rpv->pv_va = pv->pv_va + off;
8054 		rpv->pv_pa = pv->pv_pa + off;
8055 		rpv->pv_size = amount;
8056 		pv->pv_size -= amount;
8057 		if (pv->pv_size == 0) {
8058 			KASSERT(off == 0);
8059 			KASSERT((vaddr_t) pv == rpv->pv_va);
8060 			*pvp = SLIST_NEXT(pv, pv_list);
8061 		} else if (off == 0) {
8062 			KASSERT((vaddr_t) pv == rpv->pv_va);
8063 			newpv = (pv_addr_t *) (rpv->pv_va + amount);
8064 			*newpv = *pv;
8065 			newpv->pv_pa += amount;
8066 			newpv->pv_va += amount;
8067 			*pvp = newpv;
8068 		} else if (off < pv->pv_size) {
8069 			newpv = (pv_addr_t *) (rpv->pv_va + amount);
8070 			*newpv = *pv;
8071 			newpv->pv_size -= off;
8072 			newpv->pv_pa += off + amount;
8073 			newpv->pv_va += off + amount;
8074 
8075 			SLIST_NEXT(pv, pv_list) = newpv;
8076 			pv->pv_size = off;
8077 		} else {
8078 			KASSERT((vaddr_t) pv != rpv->pv_va);
8079 		}
8080 		memset((void *)rpv->pv_va, 0, amount);
8081 		return;
8082 	}
8083 
8084 	if (!uvm_physseg_valid_p(uvm_physseg_get_first()))
8085 		panic("pmap_boot_pagealloc: couldn't allocate memory");
8086 
8087 	for (pvp = &SLIST_FIRST(&pmap_boot_freeq);
8088 	     (pv = *pvp) != NULL;
8089 	     pvp = &SLIST_NEXT(pv, pv_list)) {
8090 		if (SLIST_NEXT(pv, pv_list) == NULL)
8091 			break;
8092 	}
8093 	KASSERT(mask == 0);
8094 
8095 	for (uvm_physseg_t ups = uvm_physseg_get_first();
8096 	    uvm_physseg_valid_p(ups);
8097 	    ups = uvm_physseg_get_next(ups)) {
8098 
8099 		paddr_t spn = uvm_physseg_get_start(ups);
8100 		paddr_t epn = uvm_physseg_get_end(ups);
8101 		if (spn == atop(pv->pv_pa + pv->pv_size)
8102 		    && pv->pv_va + pv->pv_size <= ptoa(epn)) {
8103 			rpv->pv_va = pv->pv_va;
8104 			rpv->pv_pa = pv->pv_pa;
8105 			rpv->pv_size = amount;
8106 			*pvp = NULL;
8107 			pmap_map_chunk(kernel_l1pt.pv_va,
8108 			     ptoa(spn) + (pv->pv_va - pv->pv_pa),
8109 			     ptoa(spn),
8110 			     amount - pv->pv_size,
8111 			     VM_PROT_READ|VM_PROT_WRITE,
8112 			     PTE_CACHE);
8113 
8114 			uvm_physseg_unplug(spn, atop(amount - pv->pv_size));
8115 			memset((void *)rpv->pv_va, 0, rpv->pv_size);
8116 			return;
8117 		}
8118 	}
8119 
8120 	panic("pmap_boot_pagealloc: couldn't allocate memory");
8121 }
8122 
8123 vaddr_t
8124 pmap_steal_memory(vsize_t size, vaddr_t *vstartp, vaddr_t *vendp)
8125 {
8126 	pv_addr_t pv;
8127 
8128 	pmap_boot_pagealloc(size, 0, 0, &pv);
8129 
8130 	return pv.pv_va;
8131 }
8132 #endif /* PMAP_STEAL_MEMORY */
8133 
8134 SYSCTL_SETUP(sysctl_machdep_pmap_setup, "sysctl machdep.kmpages setup")
8135 {
8136 	sysctl_createv(clog, 0, NULL, NULL,
8137 			CTLFLAG_PERMANENT,
8138 			CTLTYPE_NODE, "machdep", NULL,
8139 			NULL, 0, NULL, 0,
8140 			CTL_MACHDEP, CTL_EOL);
8141 
8142 	sysctl_createv(clog, 0, NULL, NULL,
8143 			CTLFLAG_PERMANENT,
8144 			CTLTYPE_INT, "kmpages",
8145 			SYSCTL_DESCR("count of pages allocated to kernel memory allocators"),
8146 			NULL, 0, &pmap_kmpages, 0,
8147 			CTL_MACHDEP, CTL_CREATE, CTL_EOL);
8148 }
8149 
8150 #ifdef PMAP_NEED_ALLOC_POOLPAGE
8151 struct vm_page *
8152 arm_pmap_alloc_poolpage(int flags)
8153 {
8154 	/*
8155 	 * On some systems, only some pages may be "coherent" for dma and we
8156 	 * want to prefer those for pool pages (think mbufs) but fallback to
8157 	 * any page if none is available.
8158 	 */
8159 	if (arm_poolpage_vmfreelist != VM_FREELIST_DEFAULT) {
8160 		return uvm_pagealloc_strat(NULL, 0, NULL, flags,
8161 		    UVM_PGA_STRAT_FALLBACK, arm_poolpage_vmfreelist);
8162 	}
8163 
8164 	return uvm_pagealloc(NULL, 0, NULL, flags);
8165 }
8166 #endif
8167 
8168 #if defined(ARM_MMU_EXTENDED) && defined(MULTIPROCESSOR)
8169 void
8170 pmap_md_tlb_info_attach(struct pmap_tlb_info *ti, struct cpu_info *ci)
8171 {
8172         /* nothing */
8173 }
8174 
8175 int
8176 pic_ipi_shootdown(void *arg)
8177 {
8178 #if PMAP_TLB_NEED_SHOOTDOWN
8179 	pmap_tlb_shootdown_process();
8180 #endif
8181 	return 1;
8182 }
8183 #endif /* ARM_MMU_EXTENDED && MULTIPROCESSOR */
8184 
8185 
8186 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
8187 vaddr_t
8188 pmap_direct_mapped_phys(paddr_t pa, bool *ok_p, vaddr_t va)
8189 {
8190 	bool ok = false;
8191 	if (physical_start <= pa && pa < physical_end) {
8192 #ifdef KERNEL_BASE_VOFFSET
8193 		const vaddr_t newva = pa + KERNEL_BASE_VOFFSET;
8194 #else
8195 		const vaddr_t newva = KERNEL_BASE + pa - physical_start;
8196 #endif
8197 #ifdef ARM_MMU_EXTENDED
8198 		if (newva >= KERNEL_BASE && newva < pmap_directlimit) {
8199 #endif
8200 			va = newva;
8201 			ok = true;
8202 #ifdef ARM_MMU_EXTENDED
8203 		}
8204 #endif
8205 	}
8206 	KASSERT(ok_p);
8207 	*ok_p = ok;
8208 	return va;
8209 }
8210 
8211 vaddr_t
8212 pmap_map_poolpage(paddr_t pa)
8213 {
8214 	bool ok __diagused;
8215 	vaddr_t va = pmap_direct_mapped_phys(pa, &ok, 0);
8216 	KASSERTMSG(ok, "pa %#lx not direct mappable", pa);
8217 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
8218 	if (arm_cache_prefer_mask != 0) {
8219 		struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
8220 		struct vm_page_md * const md = VM_PAGE_TO_MD(pg);
8221 		pmap_acquire_page_lock(md);
8222 		pmap_vac_me_harder(md, pa, pmap_kernel(), va);
8223 		pmap_release_page_lock(md);
8224 	}
8225 #endif
8226 	return va;
8227 }
8228 
8229 paddr_t
8230 pmap_unmap_poolpage(vaddr_t va)
8231 {
8232 	KASSERT(va >= KERNEL_BASE);
8233 #ifdef PMAP_CACHE_VIVT
8234 	cpu_idcache_wbinv_range(va, PAGE_SIZE);
8235 #endif
8236 #if defined(KERNEL_BASE_VOFFSET)
8237         return va - KERNEL_BASE_VOFFSET;
8238 #else
8239         return va - KERNEL_BASE + physical_start;
8240 #endif
8241 }
8242 #endif /* __HAVE_MM_MD_DIRECT_MAPPED_PHYS */
8243