xref: /netbsd-src/sys/arch/arm/arm32/pmap.c (revision d909946ca08dceb44d7d0f22ec9488679695d976)
1 /*	$NetBSD: pmap.c,v 1.340 2016/08/04 08:46:06 skrll Exp $	*/
2 
3 /*
4  * Copyright 2003 Wasabi Systems, Inc.
5  * All rights reserved.
6  *
7  * Written by Steve C. Woodford for Wasabi Systems, Inc.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  * 3. All advertising materials mentioning features or use of this software
18  *    must display the following acknowledgement:
19  *      This product includes software developed for the NetBSD Project by
20  *      Wasabi Systems, Inc.
21  * 4. The name of Wasabi Systems, Inc. may not be used to endorse
22  *    or promote products derived from this software without specific prior
23  *    written permission.
24  *
25  * THIS SOFTWARE IS PROVIDED BY WASABI SYSTEMS, INC. ``AS IS'' AND
26  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
27  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
28  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL WASABI SYSTEMS, INC
29  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
30  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
31  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
32  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
33  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
34  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
35  * POSSIBILITY OF SUCH DAMAGE.
36  */
37 
38 /*
39  * Copyright (c) 2002-2003 Wasabi Systems, Inc.
40  * Copyright (c) 2001 Richard Earnshaw
41  * Copyright (c) 2001-2002 Christopher Gilbert
42  * All rights reserved.
43  *
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. The name of the company nor the name of the author may be used to
50  *    endorse or promote products derived from this software without specific
51  *    prior written permission.
52  *
53  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED
54  * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF
55  * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
56  * IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
57  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
58  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
59  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
60  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
61  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
62  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
63  * SUCH DAMAGE.
64  */
65 
66 /*-
67  * Copyright (c) 1999 The NetBSD Foundation, Inc.
68  * All rights reserved.
69  *
70  * This code is derived from software contributed to The NetBSD Foundation
71  * by Charles M. Hannum.
72  *
73  * Redistribution and use in source and binary forms, with or without
74  * modification, are permitted provided that the following conditions
75  * are met:
76  * 1. Redistributions of source code must retain the above copyright
77  *    notice, this list of conditions and the following disclaimer.
78  * 2. Redistributions in binary form must reproduce the above copyright
79  *    notice, this list of conditions and the following disclaimer in the
80  *    documentation and/or other materials provided with the distribution.
81  *
82  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
83  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
84  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
85  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
86  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
87  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
88  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
89  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
90  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
91  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
92  * POSSIBILITY OF SUCH DAMAGE.
93  */
94 
95 /*
96  * Copyright (c) 1994-1998 Mark Brinicombe.
97  * Copyright (c) 1994 Brini.
98  * All rights reserved.
99  *
100  * This code is derived from software written for Brini by Mark Brinicombe
101  *
102  * Redistribution and use in source and binary forms, with or without
103  * modification, are permitted provided that the following conditions
104  * are met:
105  * 1. Redistributions of source code must retain the above copyright
106  *    notice, this list of conditions and the following disclaimer.
107  * 2. Redistributions in binary form must reproduce the above copyright
108  *    notice, this list of conditions and the following disclaimer in the
109  *    documentation and/or other materials provided with the distribution.
110  * 3. All advertising materials mentioning features or use of this software
111  *    must display the following acknowledgement:
112  *	This product includes software developed by Mark Brinicombe.
113  * 4. The name of the author may not be used to endorse or promote products
114  *    derived from this software without specific prior written permission.
115  *
116  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
117  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
118  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
119  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
120  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
121  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
122  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
123  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
124  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
125  *
126  * RiscBSD kernel project
127  *
128  * pmap.c
129  *
130  * Machine dependent vm stuff
131  *
132  * Created      : 20/09/94
133  */
134 
135 /*
136  * armv6 and VIPT cache support by 3am Software Foundry,
137  * Copyright (c) 2007 Microsoft
138  */
139 
140 /*
141  * Performance improvements, UVM changes, overhauls and part-rewrites
142  * were contributed by Neil A. Carson <neil@causality.com>.
143  */
144 
145 /*
146  * Overhauled again to speedup the pmap, use MMU Domains so that L1 tables
147  * can be shared, and re-work the KVM layout, by Steve Woodford of Wasabi
148  * Systems, Inc.
149  *
150  * There are still a few things outstanding at this time:
151  *
152  *   - There are some unresolved issues for MP systems:
153  *
154  *     o The L1 metadata needs a lock, or more specifically, some places
155  *       need to acquire an exclusive lock when modifying L1 translation
156  *       table entries.
157  *
158  *     o When one cpu modifies an L1 entry, and that L1 table is also
159  *       being used by another cpu, then the latter will need to be told
160  *       that a tlb invalidation may be necessary. (But only if the old
161  *       domain number in the L1 entry being over-written is currently
162  *       the active domain on that cpu). I guess there are lots more tlb
163  *       shootdown issues too...
164  *
165  *     o If the vector_page is at 0x00000000 instead of in kernel VA space,
166  *       then MP systems will lose big-time because of the MMU domain hack.
167  *       The only way this can be solved (apart from moving the vector
168  *       page to 0xffff0000) is to reserve the first 1MB of user address
169  *       space for kernel use only. This would require re-linking all
170  *       applications so that the text section starts above this 1MB
171  *       boundary.
172  *
173  *     o Tracking which VM space is resident in the cache/tlb has not yet
174  *       been implemented for MP systems.
175  *
176  *     o Finally, there is a pathological condition where two cpus running
177  *       two separate processes (not lwps) which happen to share an L1
178  *       can get into a fight over one or more L1 entries. This will result
179  *       in a significant slow-down if both processes are in tight loops.
180  */
181 
182 /*
183  * Special compilation symbols
184  * PMAP_DEBUG		- Build in pmap_debug_level code
185  */
186 
187 /* Include header files */
188 
189 #include "opt_arm_debug.h"
190 #include "opt_cpuoptions.h"
191 #include "opt_pmap_debug.h"
192 #include "opt_ddb.h"
193 #include "opt_lockdebug.h"
194 #include "opt_multiprocessor.h"
195 
196 #ifdef MULTIPROCESSOR
197 #define _INTR_PRIVATE
198 #endif
199 
200 #include <sys/param.h>
201 #include <sys/types.h>
202 #include <sys/kernel.h>
203 #include <sys/systm.h>
204 #include <sys/proc.h>
205 #include <sys/intr.h>
206 #include <sys/pool.h>
207 #include <sys/kmem.h>
208 #include <sys/cdefs.h>
209 #include <sys/cpu.h>
210 #include <sys/sysctl.h>
211 #include <sys/bus.h>
212 #include <sys/atomic.h>
213 #include <sys/kernhist.h>
214 
215 #include <uvm/uvm.h>
216 #include <uvm/pmap/pmap_pvt.h>
217 
218 #include <arm/locore.h>
219 
220 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.340 2016/08/04 08:46:06 skrll Exp $");
221 
222 //#define PMAP_DEBUG
223 #ifdef PMAP_DEBUG
224 
225 /* XXX need to get rid of all refs to this */
226 int pmap_debug_level = 0;
227 
228 /*
229  * for switching to potentially finer grained debugging
230  */
231 #define	PDB_FOLLOW	0x0001
232 #define	PDB_INIT	0x0002
233 #define	PDB_ENTER	0x0004
234 #define	PDB_REMOVE	0x0008
235 #define	PDB_CREATE	0x0010
236 #define	PDB_PTPAGE	0x0020
237 #define	PDB_GROWKERN	0x0040
238 #define	PDB_BITS	0x0080
239 #define	PDB_COLLECT	0x0100
240 #define	PDB_PROTECT	0x0200
241 #define	PDB_MAP_L1	0x0400
242 #define	PDB_BOOTSTRAP	0x1000
243 #define	PDB_PARANOIA	0x2000
244 #define	PDB_WIRING	0x4000
245 #define	PDB_PVDUMP	0x8000
246 #define	PDB_VAC		0x10000
247 #define	PDB_KENTER	0x20000
248 #define	PDB_KREMOVE	0x40000
249 #define	PDB_EXEC	0x80000
250 
251 int debugmap = 1;
252 int pmapdebug = 0;
253 #define	NPDEBUG(_lev_,_stat_) \
254 	if (pmapdebug & (_lev_)) \
255         	((_stat_))
256 
257 #else	/* PMAP_DEBUG */
258 #define NPDEBUG(_lev_,_stat_) /* Nothing */
259 #endif	/* PMAP_DEBUG */
260 
261 /*
262  * pmap_kernel() points here
263  */
264 static struct pmap	kernel_pmap_store = {
265 #ifndef ARM_MMU_EXTENDED
266 	.pm_activated = true,
267 	.pm_domain = PMAP_DOMAIN_KERNEL,
268 	.pm_cstate.cs_all = PMAP_CACHE_STATE_ALL,
269 #endif
270 };
271 struct pmap * const	kernel_pmap_ptr = &kernel_pmap_store;
272 #undef pmap_kernel
273 #define pmap_kernel()	(&kernel_pmap_store)
274 #ifdef PMAP_NEED_ALLOC_POOLPAGE
275 int			arm_poolpage_vmfreelist = VM_FREELIST_DEFAULT;
276 #endif
277 
278 /*
279  * Pool and cache that pmap structures are allocated from.
280  * We use a cache to avoid clearing the pm_l2[] array (1KB)
281  * in pmap_create().
282  */
283 static struct pool_cache pmap_cache;
284 
285 /*
286  * Pool of PV structures
287  */
288 static struct pool pmap_pv_pool;
289 static void *pmap_bootstrap_pv_page_alloc(struct pool *, int);
290 static void pmap_bootstrap_pv_page_free(struct pool *, void *);
291 static struct pool_allocator pmap_bootstrap_pv_allocator = {
292 	pmap_bootstrap_pv_page_alloc, pmap_bootstrap_pv_page_free
293 };
294 
295 /*
296  * Pool and cache of l2_dtable structures.
297  * We use a cache to avoid clearing the structures when they're
298  * allocated. (196 bytes)
299  */
300 static struct pool_cache pmap_l2dtable_cache;
301 static vaddr_t pmap_kernel_l2dtable_kva;
302 
303 /*
304  * Pool and cache of L2 page descriptors.
305  * We use a cache to avoid clearing the descriptor table
306  * when they're allocated. (1KB)
307  */
308 static struct pool_cache pmap_l2ptp_cache;
309 static vaddr_t pmap_kernel_l2ptp_kva;
310 static paddr_t pmap_kernel_l2ptp_phys;
311 
312 #ifdef PMAPCOUNTERS
313 #define	PMAP_EVCNT_INITIALIZER(name) \
314 	EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap", name)
315 
316 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
317 static struct evcnt pmap_ev_vac_clean_one =
318    PMAP_EVCNT_INITIALIZER("clean page (1 color)");
319 static struct evcnt pmap_ev_vac_flush_one =
320    PMAP_EVCNT_INITIALIZER("flush page (1 color)");
321 static struct evcnt pmap_ev_vac_flush_lots =
322    PMAP_EVCNT_INITIALIZER("flush page (2+ colors)");
323 static struct evcnt pmap_ev_vac_flush_lots2 =
324    PMAP_EVCNT_INITIALIZER("flush page (2+ colors, kmpage)");
325 EVCNT_ATTACH_STATIC(pmap_ev_vac_clean_one);
326 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_one);
327 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_lots);
328 EVCNT_ATTACH_STATIC(pmap_ev_vac_flush_lots2);
329 
330 static struct evcnt pmap_ev_vac_color_new =
331    PMAP_EVCNT_INITIALIZER("new page color");
332 static struct evcnt pmap_ev_vac_color_reuse =
333    PMAP_EVCNT_INITIALIZER("ok first page color");
334 static struct evcnt pmap_ev_vac_color_ok =
335    PMAP_EVCNT_INITIALIZER("ok page color");
336 static struct evcnt pmap_ev_vac_color_blind =
337    PMAP_EVCNT_INITIALIZER("blind page color");
338 static struct evcnt pmap_ev_vac_color_change =
339    PMAP_EVCNT_INITIALIZER("change page color");
340 static struct evcnt pmap_ev_vac_color_erase =
341    PMAP_EVCNT_INITIALIZER("erase page color");
342 static struct evcnt pmap_ev_vac_color_none =
343    PMAP_EVCNT_INITIALIZER("no page color");
344 static struct evcnt pmap_ev_vac_color_restore =
345    PMAP_EVCNT_INITIALIZER("restore page color");
346 
347 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_new);
348 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_reuse);
349 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_ok);
350 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_blind);
351 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_change);
352 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_erase);
353 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_none);
354 EVCNT_ATTACH_STATIC(pmap_ev_vac_color_restore);
355 #endif
356 
357 static struct evcnt pmap_ev_mappings =
358    PMAP_EVCNT_INITIALIZER("pages mapped");
359 static struct evcnt pmap_ev_unmappings =
360    PMAP_EVCNT_INITIALIZER("pages unmapped");
361 static struct evcnt pmap_ev_remappings =
362    PMAP_EVCNT_INITIALIZER("pages remapped");
363 
364 EVCNT_ATTACH_STATIC(pmap_ev_mappings);
365 EVCNT_ATTACH_STATIC(pmap_ev_unmappings);
366 EVCNT_ATTACH_STATIC(pmap_ev_remappings);
367 
368 static struct evcnt pmap_ev_kernel_mappings =
369    PMAP_EVCNT_INITIALIZER("kernel pages mapped");
370 static struct evcnt pmap_ev_kernel_unmappings =
371    PMAP_EVCNT_INITIALIZER("kernel pages unmapped");
372 static struct evcnt pmap_ev_kernel_remappings =
373    PMAP_EVCNT_INITIALIZER("kernel pages remapped");
374 
375 EVCNT_ATTACH_STATIC(pmap_ev_kernel_mappings);
376 EVCNT_ATTACH_STATIC(pmap_ev_kernel_unmappings);
377 EVCNT_ATTACH_STATIC(pmap_ev_kernel_remappings);
378 
379 static struct evcnt pmap_ev_kenter_mappings =
380    PMAP_EVCNT_INITIALIZER("kenter pages mapped");
381 static struct evcnt pmap_ev_kenter_unmappings =
382    PMAP_EVCNT_INITIALIZER("kenter pages unmapped");
383 static struct evcnt pmap_ev_kenter_remappings =
384    PMAP_EVCNT_INITIALIZER("kenter pages remapped");
385 static struct evcnt pmap_ev_pt_mappings =
386    PMAP_EVCNT_INITIALIZER("page table pages mapped");
387 
388 EVCNT_ATTACH_STATIC(pmap_ev_kenter_mappings);
389 EVCNT_ATTACH_STATIC(pmap_ev_kenter_unmappings);
390 EVCNT_ATTACH_STATIC(pmap_ev_kenter_remappings);
391 EVCNT_ATTACH_STATIC(pmap_ev_pt_mappings);
392 
393 static struct evcnt pmap_ev_fixup_mod =
394    PMAP_EVCNT_INITIALIZER("page modification emulations");
395 static struct evcnt pmap_ev_fixup_ref =
396    PMAP_EVCNT_INITIALIZER("page reference emulations");
397 static struct evcnt pmap_ev_fixup_exec =
398    PMAP_EVCNT_INITIALIZER("exec pages fixed up");
399 static struct evcnt pmap_ev_fixup_pdes =
400    PMAP_EVCNT_INITIALIZER("pdes fixed up");
401 #ifndef ARM_MMU_EXTENDED
402 static struct evcnt pmap_ev_fixup_ptesync =
403    PMAP_EVCNT_INITIALIZER("ptesync fixed");
404 #endif
405 
406 EVCNT_ATTACH_STATIC(pmap_ev_fixup_mod);
407 EVCNT_ATTACH_STATIC(pmap_ev_fixup_ref);
408 EVCNT_ATTACH_STATIC(pmap_ev_fixup_exec);
409 EVCNT_ATTACH_STATIC(pmap_ev_fixup_pdes);
410 #ifndef ARM_MMU_EXTENDED
411 EVCNT_ATTACH_STATIC(pmap_ev_fixup_ptesync);
412 #endif
413 
414 #ifdef PMAP_CACHE_VIPT
415 static struct evcnt pmap_ev_exec_mappings =
416    PMAP_EVCNT_INITIALIZER("exec pages mapped");
417 static struct evcnt pmap_ev_exec_cached =
418    PMAP_EVCNT_INITIALIZER("exec pages cached");
419 
420 EVCNT_ATTACH_STATIC(pmap_ev_exec_mappings);
421 EVCNT_ATTACH_STATIC(pmap_ev_exec_cached);
422 
423 static struct evcnt pmap_ev_exec_synced =
424    PMAP_EVCNT_INITIALIZER("exec pages synced");
425 static struct evcnt pmap_ev_exec_synced_map =
426    PMAP_EVCNT_INITIALIZER("exec pages synced (MP)");
427 #ifndef ARM_MMU_EXTENDED
428 static struct evcnt pmap_ev_exec_synced_unmap =
429    PMAP_EVCNT_INITIALIZER("exec pages synced (UM)");
430 static struct evcnt pmap_ev_exec_synced_remap =
431    PMAP_EVCNT_INITIALIZER("exec pages synced (RM)");
432 static struct evcnt pmap_ev_exec_synced_clearbit =
433    PMAP_EVCNT_INITIALIZER("exec pages synced (DG)");
434 static struct evcnt pmap_ev_exec_synced_kremove =
435    PMAP_EVCNT_INITIALIZER("exec pages synced (KU)");
436 #endif
437 
438 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced);
439 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_map);
440 #ifndef ARM_MMU_EXTENDED
441 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_unmap);
442 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_remap);
443 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_clearbit);
444 EVCNT_ATTACH_STATIC(pmap_ev_exec_synced_kremove);
445 #endif
446 
447 static struct evcnt pmap_ev_exec_discarded_unmap =
448    PMAP_EVCNT_INITIALIZER("exec pages discarded (UM)");
449 static struct evcnt pmap_ev_exec_discarded_zero =
450    PMAP_EVCNT_INITIALIZER("exec pages discarded (ZP)");
451 static struct evcnt pmap_ev_exec_discarded_copy =
452    PMAP_EVCNT_INITIALIZER("exec pages discarded (CP)");
453 static struct evcnt pmap_ev_exec_discarded_page_protect =
454    PMAP_EVCNT_INITIALIZER("exec pages discarded (PP)");
455 static struct evcnt pmap_ev_exec_discarded_clearbit =
456    PMAP_EVCNT_INITIALIZER("exec pages discarded (DG)");
457 static struct evcnt pmap_ev_exec_discarded_kremove =
458    PMAP_EVCNT_INITIALIZER("exec pages discarded (KU)");
459 #ifdef ARM_MMU_EXTENDED
460 static struct evcnt pmap_ev_exec_discarded_modfixup =
461    PMAP_EVCNT_INITIALIZER("exec pages discarded (MF)");
462 #endif
463 
464 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_unmap);
465 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_zero);
466 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_copy);
467 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_page_protect);
468 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_clearbit);
469 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_kremove);
470 #ifdef ARM_MMU_EXTENDED
471 EVCNT_ATTACH_STATIC(pmap_ev_exec_discarded_modfixup);
472 #endif
473 #endif /* PMAP_CACHE_VIPT */
474 
475 static struct evcnt pmap_ev_updates = PMAP_EVCNT_INITIALIZER("updates");
476 static struct evcnt pmap_ev_collects = PMAP_EVCNT_INITIALIZER("collects");
477 static struct evcnt pmap_ev_activations = PMAP_EVCNT_INITIALIZER("activations");
478 
479 EVCNT_ATTACH_STATIC(pmap_ev_updates);
480 EVCNT_ATTACH_STATIC(pmap_ev_collects);
481 EVCNT_ATTACH_STATIC(pmap_ev_activations);
482 
483 #define	PMAPCOUNT(x)	((void)(pmap_ev_##x.ev_count++))
484 #else
485 #define	PMAPCOUNT(x)	((void)0)
486 #endif
487 
488 /*
489  * pmap copy/zero page, and mem(5) hook point
490  */
491 static pt_entry_t *csrc_pte, *cdst_pte;
492 static vaddr_t csrcp, cdstp;
493 #ifdef MULTIPROCESSOR
494 static size_t cnptes;
495 #define	cpu_csrc_pte(o)	(csrc_pte + cnptes * cpu_number() + ((o) >> L2_S_SHIFT))
496 #define	cpu_cdst_pte(o)	(cdst_pte + cnptes * cpu_number() + ((o) >> L2_S_SHIFT))
497 #define	cpu_csrcp(o)	(csrcp + L2_S_SIZE * cnptes * cpu_number() + (o))
498 #define	cpu_cdstp(o)	(cdstp + L2_S_SIZE * cnptes * cpu_number() + (o))
499 #else
500 #define	cpu_csrc_pte(o)	(csrc_pte + ((o) >> L2_S_SHIFT))
501 #define	cpu_cdst_pte(o)	(cdst_pte + ((o) >> L2_S_SHIFT))
502 #define	cpu_csrcp(o)	(csrcp + (o))
503 #define	cpu_cdstp(o)	(cdstp + (o))
504 #endif
505 vaddr_t memhook;			/* used by mem.c & others */
506 kmutex_t memlock __cacheline_aligned;	/* used by mem.c & others */
507 kmutex_t pmap_lock __cacheline_aligned;
508 extern void *msgbufaddr;
509 int pmap_kmpages;
510 /*
511  * Flag to indicate if pmap_init() has done its thing
512  */
513 bool pmap_initialized;
514 
515 #if defined(ARM_MMU_EXTENDED) && defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
516 /*
517  * Virtual end of direct-mapped memory
518  */
519 vaddr_t pmap_directlimit;
520 #endif
521 
522 /*
523  * Misc. locking data structures
524  */
525 
526 static inline void
527 pmap_acquire_pmap_lock(pmap_t pm)
528 {
529 	if (pm == pmap_kernel()) {
530 #ifdef MULTIPROCESSOR
531 		KERNEL_LOCK(1, NULL);
532 #endif
533 	} else {
534 		mutex_enter(pm->pm_lock);
535 	}
536 }
537 
538 static inline void
539 pmap_release_pmap_lock(pmap_t pm)
540 {
541 	if (pm == pmap_kernel()) {
542 #ifdef MULTIPROCESSOR
543 		KERNEL_UNLOCK_ONE(NULL);
544 #endif
545 	} else {
546 		mutex_exit(pm->pm_lock);
547 	}
548 }
549 
550 static inline void
551 pmap_acquire_page_lock(struct vm_page_md *md)
552 {
553 	mutex_enter(&pmap_lock);
554 }
555 
556 static inline void
557 pmap_release_page_lock(struct vm_page_md *md)
558 {
559 	mutex_exit(&pmap_lock);
560 }
561 
562 #ifdef DIAGNOSTIC
563 static inline int
564 pmap_page_locked_p(struct vm_page_md *md)
565 {
566 	return mutex_owned(&pmap_lock);
567 }
568 #endif
569 
570 
571 /*
572  * Metadata for L1 translation tables.
573  */
574 #ifndef ARM_MMU_EXTENDED
575 struct l1_ttable {
576 	/* Entry on the L1 Table list */
577 	SLIST_ENTRY(l1_ttable) l1_link;
578 
579 	/* Entry on the L1 Least Recently Used list */
580 	TAILQ_ENTRY(l1_ttable) l1_lru;
581 
582 	/* Track how many domains are allocated from this L1 */
583 	volatile u_int l1_domain_use_count;
584 
585 	/*
586 	 * A free-list of domain numbers for this L1.
587 	 * We avoid using ffs() and a bitmap to track domains since ffs()
588 	 * is slow on ARM.
589 	 */
590 	uint8_t l1_domain_first;
591 	uint8_t l1_domain_free[PMAP_DOMAINS];
592 
593 	/* Physical address of this L1 page table */
594 	paddr_t l1_physaddr;
595 
596 	/* KVA of this L1 page table */
597 	pd_entry_t *l1_kva;
598 };
599 
600 /*
601  * L1 Page Tables are tracked using a Least Recently Used list.
602  *  - New L1s are allocated from the HEAD.
603  *  - Freed L1s are added to the TAIl.
604  *  - Recently accessed L1s (where an 'access' is some change to one of
605  *    the userland pmaps which owns this L1) are moved to the TAIL.
606  */
607 static TAILQ_HEAD(, l1_ttable) l1_lru_list;
608 static kmutex_t l1_lru_lock __cacheline_aligned;
609 
610 /*
611  * A list of all L1 tables
612  */
613 static SLIST_HEAD(, l1_ttable) l1_list;
614 #endif /* ARM_MMU_EXTENDED */
615 
616 /*
617  * The l2_dtable tracks L2_BUCKET_SIZE worth of L1 slots.
618  *
619  * This is normally 16MB worth L2 page descriptors for any given pmap.
620  * Reference counts are maintained for L2 descriptors so they can be
621  * freed when empty.
622  */
623 struct l2_bucket {
624 	pt_entry_t *l2b_kva;		/* KVA of L2 Descriptor Table */
625 	paddr_t l2b_pa;			/* Physical address of same */
626 	u_short l2b_l1slot;		/* This L2 table's L1 index */
627 	u_short l2b_occupancy;		/* How many active descriptors */
628 };
629 
630 struct l2_dtable {
631 	/* The number of L2 page descriptors allocated to this l2_dtable */
632 	u_int l2_occupancy;
633 
634 	/* List of L2 page descriptors */
635 	struct l2_bucket l2_bucket[L2_BUCKET_SIZE];
636 };
637 
638 /*
639  * Given an L1 table index, calculate the corresponding l2_dtable index
640  * and bucket index within the l2_dtable.
641  */
642 #define L2_BUCKET_XSHIFT	(L2_BUCKET_XLOG2 - L1_S_SHIFT)
643 #define L2_BUCKET_XFRAME	(~(vaddr_t)0 << L2_BUCKET_XLOG2)
644 #define L2_BUCKET_IDX(l1slot)	((l1slot) >> L2_BUCKET_XSHIFT)
645 #define L2_IDX(l1slot)		(L2_BUCKET_IDX(l1slot) >> L2_BUCKET_LOG2)
646 #define L2_BUCKET(l1slot)	(L2_BUCKET_IDX(l1slot) & (L2_BUCKET_SIZE - 1))
647 
648 __CTASSERT(0x100000000ULL == ((uint64_t)L2_SIZE * L2_BUCKET_SIZE * L1_S_SIZE));
649 __CTASSERT(L2_BUCKET_XFRAME == ~(L2_BUCKET_XSIZE-1));
650 
651 /*
652  * Given a virtual address, this macro returns the
653  * virtual address required to drop into the next L2 bucket.
654  */
655 #define	L2_NEXT_BUCKET_VA(va)	(((va) & L2_BUCKET_XFRAME) + L2_BUCKET_XSIZE)
656 
657 /*
658  * L2 allocation.
659  */
660 #define	pmap_alloc_l2_dtable()		\
661 	    pool_cache_get(&pmap_l2dtable_cache, PR_NOWAIT)
662 #define	pmap_free_l2_dtable(l2)		\
663 	    pool_cache_put(&pmap_l2dtable_cache, (l2))
664 #define pmap_alloc_l2_ptp(pap)		\
665 	    ((pt_entry_t *)pool_cache_get_paddr(&pmap_l2ptp_cache,\
666 	    PR_NOWAIT, (pap)))
667 
668 /*
669  * We try to map the page tables write-through, if possible.  However, not
670  * all CPUs have a write-through cache mode, so on those we have to sync
671  * the cache when we frob page tables.
672  *
673  * We try to evaluate this at compile time, if possible.  However, it's
674  * not always possible to do that, hence this run-time var.
675  */
676 int	pmap_needs_pte_sync;
677 
678 /*
679  * Real definition of pv_entry.
680  */
681 struct pv_entry {
682 	SLIST_ENTRY(pv_entry) pv_link;	/* next pv_entry */
683 	pmap_t		pv_pmap;        /* pmap where mapping lies */
684 	vaddr_t		pv_va;          /* virtual address for mapping */
685 	u_int		pv_flags;       /* flags */
686 };
687 
688 /*
689  * Macros to determine if a mapping might be resident in the
690  * instruction/data cache and/or TLB
691  */
692 #if ARM_MMU_V7 > 0 && !defined(ARM_MMU_EXTENDED)
693 /*
694  * Speculative loads by Cortex cores can cause TLB entries to be filled even if
695  * there are no explicit accesses, so there may be always be TLB entries to
696  * flush.  If we used ASIDs then this would not be a problem.
697  */
698 #define	PV_BEEN_EXECD(f)  (((f) & PVF_EXEC) == PVF_EXEC)
699 #define	PV_BEEN_REFD(f)   (true)
700 #else
701 #define	PV_BEEN_EXECD(f)  (((f) & (PVF_REF | PVF_EXEC)) == (PVF_REF | PVF_EXEC))
702 #define	PV_BEEN_REFD(f)   (((f) & PVF_REF) != 0)
703 #endif
704 #define	PV_IS_EXEC_P(f)   (((f) & PVF_EXEC) != 0)
705 #define	PV_IS_KENTRY_P(f) (((f) & PVF_KENTRY) != 0)
706 #define	PV_IS_WRITE_P(f)  (((f) & PVF_WRITE) != 0)
707 
708 /*
709  * Local prototypes
710  */
711 static bool		pmap_set_pt_cache_mode(pd_entry_t *, vaddr_t, size_t);
712 static void		pmap_alloc_specials(vaddr_t *, int, vaddr_t *,
713 			    pt_entry_t **);
714 static bool		pmap_is_current(pmap_t) __unused;
715 static bool		pmap_is_cached(pmap_t);
716 static void		pmap_enter_pv(struct vm_page_md *, paddr_t, struct pv_entry *,
717 			    pmap_t, vaddr_t, u_int);
718 static struct pv_entry *pmap_find_pv(struct vm_page_md *, pmap_t, vaddr_t);
719 static struct pv_entry *pmap_remove_pv(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
720 static u_int		pmap_modify_pv(struct vm_page_md *, paddr_t, pmap_t, vaddr_t,
721 			    u_int, u_int);
722 
723 static void		pmap_pinit(pmap_t);
724 static int		pmap_pmap_ctor(void *, void *, int);
725 
726 static void		pmap_alloc_l1(pmap_t);
727 static void		pmap_free_l1(pmap_t);
728 #ifndef ARM_MMU_EXTENDED
729 static void		pmap_use_l1(pmap_t);
730 #endif
731 
732 static struct l2_bucket *pmap_get_l2_bucket(pmap_t, vaddr_t);
733 static struct l2_bucket *pmap_alloc_l2_bucket(pmap_t, vaddr_t);
734 static void		pmap_free_l2_bucket(pmap_t, struct l2_bucket *, u_int);
735 static int		pmap_l2ptp_ctor(void *, void *, int);
736 static int		pmap_l2dtable_ctor(void *, void *, int);
737 
738 static void		pmap_vac_me_harder(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
739 #ifdef PMAP_CACHE_VIVT
740 static void		pmap_vac_me_kpmap(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
741 static void		pmap_vac_me_user(struct vm_page_md *, paddr_t, pmap_t, vaddr_t);
742 #endif
743 
744 static void		pmap_clearbit(struct vm_page_md *, paddr_t, u_int);
745 #ifdef PMAP_CACHE_VIVT
746 static bool		pmap_clean_page(struct vm_page_md *, bool);
747 #endif
748 #ifdef PMAP_CACHE_VIPT
749 static void		pmap_syncicache_page(struct vm_page_md *, paddr_t);
750 enum pmap_flush_op {
751 	PMAP_FLUSH_PRIMARY,
752 	PMAP_FLUSH_SECONDARY,
753 	PMAP_CLEAN_PRIMARY
754 };
755 #ifndef ARM_MMU_EXTENDED
756 static void		pmap_flush_page(struct vm_page_md *, paddr_t, enum pmap_flush_op);
757 #endif
758 #endif
759 static void		pmap_page_remove(struct vm_page_md *, paddr_t);
760 static void		pmap_pv_remove(paddr_t);
761 
762 #ifndef ARM_MMU_EXTENDED
763 static void		pmap_init_l1(struct l1_ttable *, pd_entry_t *);
764 #endif
765 static vaddr_t		kernel_pt_lookup(paddr_t);
766 
767 
768 /*
769  * Misc variables
770  */
771 vaddr_t virtual_avail;
772 vaddr_t virtual_end;
773 vaddr_t pmap_curmaxkvaddr;
774 
775 paddr_t avail_start;
776 paddr_t avail_end;
777 
778 pv_addrqh_t pmap_boot_freeq = SLIST_HEAD_INITIALIZER(&pmap_boot_freeq);
779 pv_addr_t kernelpages;
780 pv_addr_t kernel_l1pt;
781 pv_addr_t systempage;
782 
783 /* Function to set the debug level of the pmap code */
784 
785 #ifdef PMAP_DEBUG
786 void
787 pmap_debug(int level)
788 {
789 	pmap_debug_level = level;
790 	printf("pmap_debug: level=%d\n", pmap_debug_level);
791 }
792 #endif	/* PMAP_DEBUG */
793 
794 #ifdef PMAP_CACHE_VIPT
795 #define PMAP_VALIDATE_MD_PAGE(md)	\
796 	KASSERTMSG(arm_cache_prefer_mask == 0 || (((md)->pvh_attrs & PVF_WRITE) == 0) == ((md)->urw_mappings + (md)->krw_mappings == 0), \
797 	    "(md) %p: attrs=%#x urw=%u krw=%u", (md), \
798 	    (md)->pvh_attrs, (md)->urw_mappings, (md)->krw_mappings);
799 #endif /* PMAP_CACHE_VIPT */
800 /*
801  * A bunch of routines to conditionally flush the caches/TLB depending
802  * on whether the specified pmap actually needs to be flushed at any
803  * given time.
804  */
805 static inline void
806 pmap_tlb_flush_SE(pmap_t pm, vaddr_t va, u_int flags)
807 {
808 #ifdef ARM_MMU_EXTENDED
809 	pmap_tlb_invalidate_addr(pm, va);
810 #else
811 	if (pm->pm_cstate.cs_tlb_id != 0) {
812 		if (PV_BEEN_EXECD(flags)) {
813 			cpu_tlb_flushID_SE(va);
814 		} else if (PV_BEEN_REFD(flags)) {
815 			cpu_tlb_flushD_SE(va);
816 		}
817 	}
818 #endif /* ARM_MMU_EXTENDED */
819 }
820 
821 #ifndef ARM_MMU_EXTENDED
822 static inline void
823 pmap_tlb_flushID(pmap_t pm)
824 {
825 	if (pm->pm_cstate.cs_tlb_id) {
826 		cpu_tlb_flushID();
827 #if ARM_MMU_V7 == 0
828 		/*
829 		 * Speculative loads by Cortex cores can cause TLB entries to
830 		 * be filled even if there are no explicit accesses, so there
831 		 * may be always be TLB entries to flush.  If we used ASIDs
832 		 * then it would not be a problem.
833 		 * This is not true for other CPUs.
834 		 */
835 		pm->pm_cstate.cs_tlb = 0;
836 #endif /* ARM_MMU_V7 */
837 	}
838 }
839 
840 static inline void
841 pmap_tlb_flushD(pmap_t pm)
842 {
843 	if (pm->pm_cstate.cs_tlb_d) {
844 		cpu_tlb_flushD();
845 #if ARM_MMU_V7 == 0
846 		/*
847 		 * Speculative loads by Cortex cores can cause TLB entries to
848 		 * be filled even if there are no explicit accesses, so there
849 		 * may be always be TLB entries to flush.  If we used ASIDs
850 		 * then it would not be a problem.
851 		 * This is not true for other CPUs.
852 		 */
853 		pm->pm_cstate.cs_tlb_d = 0;
854 #endif /* ARM_MMU_V7 */
855 	}
856 }
857 #endif /* ARM_MMU_EXTENDED */
858 
859 #ifdef PMAP_CACHE_VIVT
860 static inline void
861 pmap_cache_wbinv_page(pmap_t pm, vaddr_t va, bool do_inv, u_int flags)
862 {
863 	if (PV_BEEN_EXECD(flags) && pm->pm_cstate.cs_cache_id) {
864 		cpu_idcache_wbinv_range(va, PAGE_SIZE);
865 	} else if (PV_BEEN_REFD(flags) && pm->pm_cstate.cs_cache_d) {
866 		if (do_inv) {
867 			if (flags & PVF_WRITE)
868 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
869 			else
870 				cpu_dcache_inv_range(va, PAGE_SIZE);
871 		} else if (flags & PVF_WRITE) {
872 			cpu_dcache_wb_range(va, PAGE_SIZE);
873 		}
874 	}
875 }
876 
877 static inline void
878 pmap_cache_wbinv_all(pmap_t pm, u_int flags)
879 {
880 	if (PV_BEEN_EXECD(flags)) {
881 		if (pm->pm_cstate.cs_cache_id) {
882 			cpu_idcache_wbinv_all();
883 			pm->pm_cstate.cs_cache = 0;
884 		}
885 	} else if (pm->pm_cstate.cs_cache_d) {
886 		cpu_dcache_wbinv_all();
887 		pm->pm_cstate.cs_cache_d = 0;
888 	}
889 }
890 #endif /* PMAP_CACHE_VIVT */
891 
892 static inline uint8_t
893 pmap_domain(pmap_t pm)
894 {
895 #ifdef ARM_MMU_EXTENDED
896 	return pm == pmap_kernel() ? PMAP_DOMAIN_KERNEL : PMAP_DOMAIN_USER;
897 #else
898 	return pm->pm_domain;
899 #endif
900 }
901 
902 static inline pd_entry_t *
903 pmap_l1_kva(pmap_t pm)
904 {
905 #ifdef ARM_MMU_EXTENDED
906 	return pm->pm_l1;
907 #else
908 	return pm->pm_l1->l1_kva;
909 #endif
910 }
911 
912 static inline bool
913 pmap_is_current(pmap_t pm)
914 {
915 	if (pm == pmap_kernel() || curproc->p_vmspace->vm_map.pmap == pm)
916 		return true;
917 
918 	return false;
919 }
920 
921 static inline bool
922 pmap_is_cached(pmap_t pm)
923 {
924 #ifdef ARM_MMU_EXTENDED
925 	if (pm == pmap_kernel())
926 		return true;
927 #ifdef MULTIPROCESSOR
928 	// Is this pmap active on any CPU?
929 	if (!kcpuset_iszero(pm->pm_active))
930 		return true;
931 #else
932 	struct pmap_tlb_info * const ti = cpu_tlb_info(curcpu());
933 	// Is this pmap active?
934 	if (PMAP_PAI_ASIDVALID_P(PMAP_PAI(pm, ti), ti))
935 		return true;
936 #endif
937 #else
938 	struct cpu_info * const ci = curcpu();
939 	if (pm == pmap_kernel() || ci->ci_pmap_lastuser == NULL
940 	    || ci->ci_pmap_lastuser == pm)
941 		return true;
942 #endif /* ARM_MMU_EXTENDED */
943 
944 	return false;
945 }
946 
947 /*
948  * PTE_SYNC_CURRENT:
949  *
950  *     Make sure the pte is written out to RAM.
951  *     We need to do this for one of two cases:
952  *       - We're dealing with the kernel pmap
953  *       - There is no pmap active in the cache/tlb.
954  *       - The specified pmap is 'active' in the cache/tlb.
955  */
956 
957 static inline void
958 pmap_pte_sync_current(pmap_t pm, pt_entry_t *ptep)
959 {
960 	if (PMAP_NEEDS_PTE_SYNC && pmap_is_cached(pm))
961 		PTE_SYNC(ptep);
962 	arm_dsb();
963 }
964 
965 #ifdef PMAP_INCLUDE_PTE_SYNC
966 #define	PTE_SYNC_CURRENT(pm, ptep)	pmap_pte_sync_current(pm, ptep)
967 #else
968 #define	PTE_SYNC_CURRENT(pm, ptep)	/* nothing */
969 #endif
970 
971 /*
972  * main pv_entry manipulation functions:
973  *   pmap_enter_pv: enter a mapping onto a vm_page list
974  *   pmap_remove_pv: remove a mapping from a vm_page list
975  *
976  * NOTE: pmap_enter_pv expects to lock the pvh itself
977  *       pmap_remove_pv expects the caller to lock the pvh before calling
978  */
979 
980 /*
981  * pmap_enter_pv: enter a mapping onto a vm_page lst
982  *
983  * => caller should hold the proper lock on pmap_main_lock
984  * => caller should have pmap locked
985  * => we will gain the lock on the vm_page and allocate the new pv_entry
986  * => caller should adjust ptp's wire_count before calling
987  * => caller should not adjust pmap's wire_count
988  */
989 static void
990 pmap_enter_pv(struct vm_page_md *md, paddr_t pa, struct pv_entry *pv, pmap_t pm,
991     vaddr_t va, u_int flags)
992 {
993 	struct pv_entry **pvp;
994 
995 	NPDEBUG(PDB_PVDUMP,
996 	    printf("pmap_enter_pv: pm %p, md %p, flags 0x%x\n", pm, md, flags));
997 
998 	pv->pv_pmap = pm;
999 	pv->pv_va = va;
1000 	pv->pv_flags = flags;
1001 
1002 	pvp = &SLIST_FIRST(&md->pvh_list);
1003 #ifdef PMAP_CACHE_VIPT
1004 	/*
1005 	 * Insert unmanaged entries, writeable first, at the head of
1006 	 * the pv list.
1007 	 */
1008 	if (__predict_true(!PV_IS_KENTRY_P(flags))) {
1009 		while (*pvp != NULL && PV_IS_KENTRY_P((*pvp)->pv_flags))
1010 			pvp = &SLIST_NEXT(*pvp, pv_link);
1011 	}
1012 	if (!PV_IS_WRITE_P(flags)) {
1013 		while (*pvp != NULL && PV_IS_WRITE_P((*pvp)->pv_flags))
1014 			pvp = &SLIST_NEXT(*pvp, pv_link);
1015 	}
1016 #endif
1017 	SLIST_NEXT(pv, pv_link) = *pvp;		/* add to ... */
1018 	*pvp = pv;				/* ... locked list */
1019 	md->pvh_attrs |= flags & (PVF_REF | PVF_MOD);
1020 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
1021 	if ((pv->pv_flags & PVF_KWRITE) == PVF_KWRITE)
1022 		md->pvh_attrs |= PVF_KMOD;
1023 	if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) != PVF_NC)
1024 		md->pvh_attrs |= PVF_DIRTY;
1025 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1026 #endif
1027 	if (pm == pmap_kernel()) {
1028 		PMAPCOUNT(kernel_mappings);
1029 		if (flags & PVF_WRITE)
1030 			md->krw_mappings++;
1031 		else
1032 			md->kro_mappings++;
1033 	} else {
1034 		if (flags & PVF_WRITE)
1035 			md->urw_mappings++;
1036 		else
1037 			md->uro_mappings++;
1038 	}
1039 
1040 #ifdef PMAP_CACHE_VIPT
1041 #ifndef ARM_MMU_EXTENDED
1042 	/*
1043 	 * Even though pmap_vac_me_harder will set PVF_WRITE for us,
1044 	 * do it here as well to keep the mappings & KVF_WRITE consistent.
1045 	 */
1046 	if (arm_cache_prefer_mask != 0 && (flags & PVF_WRITE) != 0) {
1047 		md->pvh_attrs |= PVF_WRITE;
1048 	}
1049 #endif
1050 	/*
1051 	 * If this is an exec mapping and its the first exec mapping
1052 	 * for this page, make sure to sync the I-cache.
1053 	 */
1054 	if (PV_IS_EXEC_P(flags)) {
1055 #ifndef ARM_MMU_EXTENDED
1056 		if (!PV_IS_EXEC_P(md->pvh_attrs)) {
1057 			pmap_syncicache_page(md, pa);
1058 			PMAPCOUNT(exec_synced_map);
1059 		}
1060 #endif
1061 		PMAPCOUNT(exec_mappings);
1062 	}
1063 #endif
1064 
1065 	PMAPCOUNT(mappings);
1066 
1067 	if (pv->pv_flags & PVF_WIRED)
1068 		++pm->pm_stats.wired_count;
1069 }
1070 
1071 /*
1072  *
1073  * pmap_find_pv: Find a pv entry
1074  *
1075  * => caller should hold lock on vm_page
1076  */
1077 static inline struct pv_entry *
1078 pmap_find_pv(struct vm_page_md *md, pmap_t pm, vaddr_t va)
1079 {
1080 	struct pv_entry *pv;
1081 
1082 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1083 		if (pm == pv->pv_pmap && va == pv->pv_va)
1084 			break;
1085 	}
1086 
1087 	return (pv);
1088 }
1089 
1090 /*
1091  * pmap_remove_pv: try to remove a mapping from a pv_list
1092  *
1093  * => caller should hold proper lock on pmap_main_lock
1094  * => pmap should be locked
1095  * => caller should hold lock on vm_page [so that attrs can be adjusted]
1096  * => caller should adjust ptp's wire_count and free PTP if needed
1097  * => caller should NOT adjust pmap's wire_count
1098  * => we return the removed pv
1099  */
1100 static struct pv_entry *
1101 pmap_remove_pv(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1102 {
1103 	struct pv_entry *pv, **prevptr;
1104 
1105 	NPDEBUG(PDB_PVDUMP,
1106 	    printf("pmap_remove_pv: pm %p, md %p, va 0x%08lx\n", pm, md, va));
1107 
1108 	prevptr = &SLIST_FIRST(&md->pvh_list); /* prev pv_entry ptr */
1109 	pv = *prevptr;
1110 
1111 	while (pv) {
1112 		if (pv->pv_pmap == pm && pv->pv_va == va) {	/* match? */
1113 			NPDEBUG(PDB_PVDUMP, printf("pmap_remove_pv: pm %p, md "
1114 			    "%p, flags 0x%x\n", pm, md, pv->pv_flags));
1115 			if (pv->pv_flags & PVF_WIRED) {
1116 				--pm->pm_stats.wired_count;
1117 			}
1118 			*prevptr = SLIST_NEXT(pv, pv_link);	/* remove it! */
1119 			if (pm == pmap_kernel()) {
1120 				PMAPCOUNT(kernel_unmappings);
1121 				if (pv->pv_flags & PVF_WRITE)
1122 					md->krw_mappings--;
1123 				else
1124 					md->kro_mappings--;
1125 			} else {
1126 				if (pv->pv_flags & PVF_WRITE)
1127 					md->urw_mappings--;
1128 				else
1129 					md->uro_mappings--;
1130 			}
1131 
1132 			PMAPCOUNT(unmappings);
1133 #ifdef PMAP_CACHE_VIPT
1134 			if (!(pv->pv_flags & PVF_WRITE))
1135 				break;
1136 			/*
1137 			 * If this page has had an exec mapping, then if
1138 			 * this was the last mapping, discard the contents,
1139 			 * otherwise sync the i-cache for this page.
1140 			 */
1141 			if (PV_IS_EXEC_P(md->pvh_attrs)) {
1142 #ifdef ARM_MMU_EXTENDED
1143 				md->pvh_attrs &= ~PVF_EXEC;
1144 				PMAPCOUNT(exec_discarded_unmap);
1145 #else
1146 				if (SLIST_EMPTY(&md->pvh_list)) {
1147 					md->pvh_attrs &= ~PVF_EXEC;
1148 					PMAPCOUNT(exec_discarded_unmap);
1149 				} else {
1150 					pmap_syncicache_page(md, pa);
1151 					PMAPCOUNT(exec_synced_unmap);
1152 				}
1153 #endif /* ARM_MMU_EXTENDED */
1154 			}
1155 #endif /* PMAP_CACHE_VIPT */
1156 			break;
1157 		}
1158 		prevptr = &SLIST_NEXT(pv, pv_link);	/* previous pointer */
1159 		pv = *prevptr;				/* advance */
1160 	}
1161 
1162 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
1163 	/*
1164 	 * If we no longer have a WRITEABLE KENTRY at the head of list,
1165 	 * clear the KMOD attribute from the page.
1166 	 */
1167 	if (SLIST_FIRST(&md->pvh_list) == NULL
1168 	    || (SLIST_FIRST(&md->pvh_list)->pv_flags & PVF_KWRITE) != PVF_KWRITE)
1169 		md->pvh_attrs &= ~PVF_KMOD;
1170 
1171 	/*
1172 	 * If this was a writeable page and there are no more writeable
1173 	 * mappings (ignoring KMPAGE), clear the WRITE flag and writeback
1174 	 * the contents to memory.
1175 	 */
1176 	if (arm_cache_prefer_mask != 0) {
1177 		if (md->krw_mappings + md->urw_mappings == 0)
1178 			md->pvh_attrs &= ~PVF_WRITE;
1179 		PMAP_VALIDATE_MD_PAGE(md);
1180 	}
1181 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1182 #endif /* PMAP_CACHE_VIPT && !ARM_MMU_EXTENDED */
1183 
1184 	return(pv);				/* return removed pv */
1185 }
1186 
1187 /*
1188  *
1189  * pmap_modify_pv: Update pv flags
1190  *
1191  * => caller should hold lock on vm_page [so that attrs can be adjusted]
1192  * => caller should NOT adjust pmap's wire_count
1193  * => caller must call pmap_vac_me_harder() if writable status of a page
1194  *    may have changed.
1195  * => we return the old flags
1196  *
1197  * Modify a physical-virtual mapping in the pv table
1198  */
1199 static u_int
1200 pmap_modify_pv(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va,
1201     u_int clr_mask, u_int set_mask)
1202 {
1203 	struct pv_entry *npv;
1204 	u_int flags, oflags;
1205 
1206 	KASSERT(!PV_IS_KENTRY_P(clr_mask));
1207 	KASSERT(!PV_IS_KENTRY_P(set_mask));
1208 
1209 	if ((npv = pmap_find_pv(md, pm, va)) == NULL)
1210 		return (0);
1211 
1212 	NPDEBUG(PDB_PVDUMP,
1213 	    printf("pmap_modify_pv: pm %p, md %p, clr 0x%x, set 0x%x, flags 0x%x\n", pm, md, clr_mask, set_mask, npv->pv_flags));
1214 
1215 	/*
1216 	 * There is at least one VA mapping this page.
1217 	 */
1218 
1219 	if (clr_mask & (PVF_REF | PVF_MOD)) {
1220 		md->pvh_attrs |= set_mask & (PVF_REF | PVF_MOD);
1221 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
1222 		if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) != PVF_NC)
1223 			md->pvh_attrs |= PVF_DIRTY;
1224 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1225 #endif /* PMAP_CACHE_VIPT && !ARM_MMU_EXTENDED */
1226 	}
1227 
1228 	oflags = npv->pv_flags;
1229 	npv->pv_flags = flags = (oflags & ~clr_mask) | set_mask;
1230 
1231 	if ((flags ^ oflags) & PVF_WIRED) {
1232 		if (flags & PVF_WIRED)
1233 			++pm->pm_stats.wired_count;
1234 		else
1235 			--pm->pm_stats.wired_count;
1236 	}
1237 
1238 	if ((flags ^ oflags) & PVF_WRITE) {
1239 		if (pm == pmap_kernel()) {
1240 			if (flags & PVF_WRITE) {
1241 				md->krw_mappings++;
1242 				md->kro_mappings--;
1243 			} else {
1244 				md->kro_mappings++;
1245 				md->krw_mappings--;
1246 			}
1247 		} else {
1248 			if (flags & PVF_WRITE) {
1249 				md->urw_mappings++;
1250 				md->uro_mappings--;
1251 			} else {
1252 				md->uro_mappings++;
1253 				md->urw_mappings--;
1254 			}
1255 		}
1256 	}
1257 #ifdef PMAP_CACHE_VIPT
1258 	if (arm_cache_prefer_mask != 0) {
1259 		if (md->urw_mappings + md->krw_mappings == 0) {
1260 			md->pvh_attrs &= ~PVF_WRITE;
1261 		} else {
1262 			md->pvh_attrs |= PVF_WRITE;
1263 		}
1264 	}
1265 #ifndef ARM_MMU_EXTENDED
1266 	/*
1267 	 * We have two cases here: the first is from enter_pv (new exec
1268 	 * page), the second is a combined pmap_remove_pv/pmap_enter_pv.
1269 	 * Since in latter, pmap_enter_pv won't do anything, we just have
1270 	 * to do what pmap_remove_pv would do.
1271 	 */
1272 	if ((PV_IS_EXEC_P(flags) && !PV_IS_EXEC_P(md->pvh_attrs))
1273 	    || (PV_IS_EXEC_P(md->pvh_attrs)
1274 		|| (!(flags & PVF_WRITE) && (oflags & PVF_WRITE)))) {
1275 		pmap_syncicache_page(md, pa);
1276 		PMAPCOUNT(exec_synced_remap);
1277 	}
1278 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
1279 #endif /* !ARM_MMU_EXTENDED */
1280 #endif /* PMAP_CACHE_VIPT */
1281 
1282 	PMAPCOUNT(remappings);
1283 
1284 	return (oflags);
1285 }
1286 
1287 /*
1288  * Allocate an L1 translation table for the specified pmap.
1289  * This is called at pmap creation time.
1290  */
1291 static void
1292 pmap_alloc_l1(pmap_t pm)
1293 {
1294 #ifdef ARM_MMU_EXTENDED
1295 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
1296 	struct vm_page *pg;
1297 	bool ok __diagused;
1298 	for (;;) {
1299 #ifdef PMAP_NEED_ALLOC_POOLPAGE
1300 		pg = arm_pmap_alloc_poolpage(UVM_PGA_ZERO);
1301 #else
1302 		pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_ZERO);
1303 #endif
1304 		if (pg != NULL)
1305 			break;
1306 		uvm_wait("pmapl1alloc");
1307 	}
1308 	pm->pm_l1_pa = VM_PAGE_TO_PHYS(pg);
1309 	vaddr_t va = pmap_direct_mapped_phys(pm->pm_l1_pa, &ok, 0);
1310 	KASSERT(ok);
1311 	KASSERT(va >= KERNEL_BASE);
1312 
1313 #else
1314 	KASSERTMSG(kernel_map != NULL, "pm %p", pm);
1315 	vaddr_t va = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
1316 	    UVM_KMF_WIRED|UVM_KMF_ZERO);
1317 	KASSERT(va);
1318 	pmap_extract(pmap_kernel(), va, &pm->pm_l1_pa);
1319 #endif
1320 	pm->pm_l1 = (pd_entry_t *)va;
1321 	PTE_SYNC_RANGE(pm->pm_l1, PAGE_SIZE / sizeof(pt_entry_t));
1322 #else
1323 	struct l1_ttable *l1;
1324 	uint8_t domain;
1325 
1326 	/*
1327 	 * Remove the L1 at the head of the LRU list
1328 	 */
1329 	mutex_spin_enter(&l1_lru_lock);
1330 	l1 = TAILQ_FIRST(&l1_lru_list);
1331 	KDASSERT(l1 != NULL);
1332 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1333 
1334 	/*
1335 	 * Pick the first available domain number, and update
1336 	 * the link to the next number.
1337 	 */
1338 	domain = l1->l1_domain_first;
1339 	l1->l1_domain_first = l1->l1_domain_free[domain];
1340 
1341 	/*
1342 	 * If there are still free domain numbers in this L1,
1343 	 * put it back on the TAIL of the LRU list.
1344 	 */
1345 	if (++l1->l1_domain_use_count < PMAP_DOMAINS)
1346 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1347 
1348 	mutex_spin_exit(&l1_lru_lock);
1349 
1350 	/*
1351 	 * Fix up the relevant bits in the pmap structure
1352 	 */
1353 	pm->pm_l1 = l1;
1354 	pm->pm_domain = domain + 1;
1355 #endif
1356 }
1357 
1358 /*
1359  * Free an L1 translation table.
1360  * This is called at pmap destruction time.
1361  */
1362 static void
1363 pmap_free_l1(pmap_t pm)
1364 {
1365 #ifdef ARM_MMU_EXTENDED
1366 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
1367 	struct vm_page *pg = PHYS_TO_VM_PAGE(pm->pm_l1_pa);
1368 	uvm_pagefree(pg);
1369 #else
1370 	uvm_km_free(kernel_map, (vaddr_t)pm->pm_l1, PAGE_SIZE, UVM_KMF_WIRED);
1371 #endif
1372 	pm->pm_l1 = NULL;
1373 	pm->pm_l1_pa = 0;
1374 #else
1375 	struct l1_ttable *l1 = pm->pm_l1;
1376 
1377 	mutex_spin_enter(&l1_lru_lock);
1378 
1379 	/*
1380 	 * If this L1 is currently on the LRU list, remove it.
1381 	 */
1382 	if (l1->l1_domain_use_count < PMAP_DOMAINS)
1383 		TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1384 
1385 	/*
1386 	 * Free up the domain number which was allocated to the pmap
1387 	 */
1388 	l1->l1_domain_free[pmap_domain(pm) - 1] = l1->l1_domain_first;
1389 	l1->l1_domain_first = pmap_domain(pm) - 1;
1390 	l1->l1_domain_use_count--;
1391 
1392 	/*
1393 	 * The L1 now must have at least 1 free domain, so add
1394 	 * it back to the LRU list. If the use count is zero,
1395 	 * put it at the head of the list, otherwise it goes
1396 	 * to the tail.
1397 	 */
1398 	if (l1->l1_domain_use_count == 0)
1399 		TAILQ_INSERT_HEAD(&l1_lru_list, l1, l1_lru);
1400 	else
1401 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1402 
1403 	mutex_spin_exit(&l1_lru_lock);
1404 #endif /* ARM_MMU_EXTENDED */
1405 }
1406 
1407 #ifndef ARM_MMU_EXTENDED
1408 static inline void
1409 pmap_use_l1(pmap_t pm)
1410 {
1411 	struct l1_ttable *l1;
1412 
1413 	/*
1414 	 * Do nothing if we're in interrupt context.
1415 	 * Access to an L1 by the kernel pmap must not affect
1416 	 * the LRU list.
1417 	 */
1418 	if (cpu_intr_p() || pm == pmap_kernel())
1419 		return;
1420 
1421 	l1 = pm->pm_l1;
1422 
1423 	/*
1424 	 * If the L1 is not currently on the LRU list, just return
1425 	 */
1426 	if (l1->l1_domain_use_count == PMAP_DOMAINS)
1427 		return;
1428 
1429 	mutex_spin_enter(&l1_lru_lock);
1430 
1431 	/*
1432 	 * Check the use count again, now that we've acquired the lock
1433 	 */
1434 	if (l1->l1_domain_use_count == PMAP_DOMAINS) {
1435 		mutex_spin_exit(&l1_lru_lock);
1436 		return;
1437 	}
1438 
1439 	/*
1440 	 * Move the L1 to the back of the LRU list
1441 	 */
1442 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
1443 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
1444 
1445 	mutex_spin_exit(&l1_lru_lock);
1446 }
1447 #endif /* !ARM_MMU_EXTENDED */
1448 
1449 /*
1450  * void pmap_free_l2_ptp(pt_entry_t *, paddr_t *)
1451  *
1452  * Free an L2 descriptor table.
1453  */
1454 static inline void
1455 #if defined(PMAP_INCLUDE_PTE_SYNC) && defined(PMAP_CACHE_VIVT)
1456 pmap_free_l2_ptp(bool need_sync, pt_entry_t *l2, paddr_t pa)
1457 #else
1458 pmap_free_l2_ptp(pt_entry_t *l2, paddr_t pa)
1459 #endif
1460 {
1461 #if defined(PMAP_INCLUDE_PTE_SYNC) && defined(PMAP_CACHE_VIVT)
1462 	/*
1463 	 * Note: With a write-back cache, we may need to sync this
1464 	 * L2 table before re-using it.
1465 	 * This is because it may have belonged to a non-current
1466 	 * pmap, in which case the cache syncs would have been
1467 	 * skipped for the pages that were being unmapped. If the
1468 	 * L2 table were then to be immediately re-allocated to
1469 	 * the *current* pmap, it may well contain stale mappings
1470 	 * which have not yet been cleared by a cache write-back
1471 	 * and so would still be visible to the mmu.
1472 	 */
1473 	if (need_sync)
1474 		PTE_SYNC_RANGE(l2, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1475 #endif /* PMAP_INCLUDE_PTE_SYNC && PMAP_CACHE_VIVT */
1476 	pool_cache_put_paddr(&pmap_l2ptp_cache, (void *)l2, pa);
1477 }
1478 
1479 /*
1480  * Returns a pointer to the L2 bucket associated with the specified pmap
1481  * and VA, or NULL if no L2 bucket exists for the address.
1482  */
1483 static inline struct l2_bucket *
1484 pmap_get_l2_bucket(pmap_t pm, vaddr_t va)
1485 {
1486 	const size_t l1slot = l1pte_index(va);
1487 	struct l2_dtable *l2;
1488 	struct l2_bucket *l2b;
1489 
1490 	if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL ||
1491 	    (l2b = &l2->l2_bucket[L2_BUCKET(l1slot)])->l2b_kva == NULL)
1492 		return (NULL);
1493 
1494 	return (l2b);
1495 }
1496 
1497 /*
1498  * Returns a pointer to the L2 bucket associated with the specified pmap
1499  * and VA.
1500  *
1501  * If no L2 bucket exists, perform the necessary allocations to put an L2
1502  * bucket/page table in place.
1503  *
1504  * Note that if a new L2 bucket/page was allocated, the caller *must*
1505  * increment the bucket occupancy counter appropriately *before*
1506  * releasing the pmap's lock to ensure no other thread or cpu deallocates
1507  * the bucket/page in the meantime.
1508  */
1509 static struct l2_bucket *
1510 pmap_alloc_l2_bucket(pmap_t pm, vaddr_t va)
1511 {
1512 	const size_t l1slot = l1pte_index(va);
1513 	struct l2_dtable *l2;
1514 
1515 	if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL) {
1516 		/*
1517 		 * No mapping at this address, as there is
1518 		 * no entry in the L1 table.
1519 		 * Need to allocate a new l2_dtable.
1520 		 */
1521 		if ((l2 = pmap_alloc_l2_dtable()) == NULL)
1522 			return (NULL);
1523 
1524 		/*
1525 		 * Link it into the parent pmap
1526 		 */
1527 		pm->pm_l2[L2_IDX(l1slot)] = l2;
1528 	}
1529 
1530 	struct l2_bucket * const l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
1531 
1532 	/*
1533 	 * Fetch pointer to the L2 page table associated with the address.
1534 	 */
1535 	if (l2b->l2b_kva == NULL) {
1536 		pt_entry_t *ptep;
1537 
1538 		/*
1539 		 * No L2 page table has been allocated. Chances are, this
1540 		 * is because we just allocated the l2_dtable, above.
1541 		 */
1542 		if ((ptep = pmap_alloc_l2_ptp(&l2b->l2b_pa)) == NULL) {
1543 			/*
1544 			 * Oops, no more L2 page tables available at this
1545 			 * time. We may need to deallocate the l2_dtable
1546 			 * if we allocated a new one above.
1547 			 */
1548 			if (l2->l2_occupancy == 0) {
1549 				pm->pm_l2[L2_IDX(l1slot)] = NULL;
1550 				pmap_free_l2_dtable(l2);
1551 			}
1552 			return (NULL);
1553 		}
1554 
1555 		l2->l2_occupancy++;
1556 		l2b->l2b_kva = ptep;
1557 		l2b->l2b_l1slot = l1slot;
1558 
1559 #ifdef ARM_MMU_EXTENDED
1560 		/*
1561 		 * We know there will be a mapping here, so simply
1562 		 * enter this PTP into the L1 now.
1563 		 */
1564 		pd_entry_t * const pdep = pmap_l1_kva(pm) + l1slot;
1565 		pd_entry_t npde = L1_C_PROTO | l2b->l2b_pa
1566 		    | L1_C_DOM(pmap_domain(pm));
1567 		KASSERT(*pdep == 0);
1568 		l1pte_setone(pdep, npde);
1569 		PDE_SYNC(pdep);
1570 #endif
1571 	}
1572 
1573 	return (l2b);
1574 }
1575 
1576 /*
1577  * One or more mappings in the specified L2 descriptor table have just been
1578  * invalidated.
1579  *
1580  * Garbage collect the metadata and descriptor table itself if necessary.
1581  *
1582  * The pmap lock must be acquired when this is called (not necessary
1583  * for the kernel pmap).
1584  */
1585 static void
1586 pmap_free_l2_bucket(pmap_t pm, struct l2_bucket *l2b, u_int count)
1587 {
1588 	KDASSERT(count <= l2b->l2b_occupancy);
1589 
1590 	/*
1591 	 * Update the bucket's reference count according to how many
1592 	 * PTEs the caller has just invalidated.
1593 	 */
1594 	l2b->l2b_occupancy -= count;
1595 
1596 	/*
1597 	 * Note:
1598 	 *
1599 	 * Level 2 page tables allocated to the kernel pmap are never freed
1600 	 * as that would require checking all Level 1 page tables and
1601 	 * removing any references to the Level 2 page table. See also the
1602 	 * comment elsewhere about never freeing bootstrap L2 descriptors.
1603 	 *
1604 	 * We make do with just invalidating the mapping in the L2 table.
1605 	 *
1606 	 * This isn't really a big deal in practice and, in fact, leads
1607 	 * to a performance win over time as we don't need to continually
1608 	 * alloc/free.
1609 	 */
1610 	if (l2b->l2b_occupancy > 0 || pm == pmap_kernel())
1611 		return;
1612 
1613 	/*
1614 	 * There are no more valid mappings in this level 2 page table.
1615 	 * Go ahead and NULL-out the pointer in the bucket, then
1616 	 * free the page table.
1617 	 */
1618 	const size_t l1slot = l2b->l2b_l1slot;
1619 	pt_entry_t * const ptep = l2b->l2b_kva;
1620 	l2b->l2b_kva = NULL;
1621 
1622 	pd_entry_t * const pdep = pmap_l1_kva(pm) + l1slot;
1623 	pd_entry_t pde __diagused = *pdep;
1624 
1625 #ifdef ARM_MMU_EXTENDED
1626 	/*
1627 	 * Invalidate the L1 slot.
1628 	 */
1629 	KASSERT((pde & L1_TYPE_MASK) == L1_TYPE_C);
1630 #else
1631 	/*
1632 	 * If the L1 slot matches the pmap's domain number, then invalidate it.
1633 	 */
1634 	if ((pde & (L1_C_DOM_MASK|L1_TYPE_MASK))
1635 	    == (L1_C_DOM(pmap_domain(pm))|L1_TYPE_C)) {
1636 #endif
1637 		l1pte_setone(pdep, 0);
1638 		PDE_SYNC(pdep);
1639 #ifndef ARM_MMU_EXTENDED
1640 	}
1641 #endif
1642 
1643 	/*
1644 	 * Release the L2 descriptor table back to the pool cache.
1645 	 */
1646 #if defined(PMAP_INCLUDE_PTE_SYNC) && defined(PMAP_CACHE_VIVT)
1647 	pmap_free_l2_ptp(!pmap_is_cached(pm), ptep, l2b->l2b_pa);
1648 #else
1649 	pmap_free_l2_ptp(ptep, l2b->l2b_pa);
1650 #endif
1651 
1652 	/*
1653 	 * Update the reference count in the associated l2_dtable
1654 	 */
1655 	struct l2_dtable * const l2 = pm->pm_l2[L2_IDX(l1slot)];
1656 	if (--l2->l2_occupancy > 0)
1657 		return;
1658 
1659 	/*
1660 	 * There are no more valid mappings in any of the Level 1
1661 	 * slots managed by this l2_dtable. Go ahead and NULL-out
1662 	 * the pointer in the parent pmap and free the l2_dtable.
1663 	 */
1664 	pm->pm_l2[L2_IDX(l1slot)] = NULL;
1665 	pmap_free_l2_dtable(l2);
1666 }
1667 
1668 /*
1669  * Pool cache constructors for L2 descriptor tables, metadata and pmap
1670  * structures.
1671  */
1672 static int
1673 pmap_l2ptp_ctor(void *arg, void *v, int flags)
1674 {
1675 #ifndef PMAP_INCLUDE_PTE_SYNC
1676 	vaddr_t va = (vaddr_t)v & ~PGOFSET;
1677 
1678 	/*
1679 	 * The mappings for these page tables were initially made using
1680 	 * pmap_kenter_pa() by the pool subsystem. Therefore, the cache-
1681 	 * mode will not be right for page table mappings. To avoid
1682 	 * polluting the pmap_kenter_pa() code with a special case for
1683 	 * page tables, we simply fix up the cache-mode here if it's not
1684 	 * correct.
1685 	 */
1686 	if (pte_l2_s_cache_mode != pte_l2_s_cache_mode_pt) {
1687 		const struct l2_bucket * const l2b =
1688 		    pmap_get_l2_bucket(pmap_kernel(), va);
1689 		KASSERTMSG(l2b != NULL, "%#lx", va);
1690 		pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
1691 		const pt_entry_t opte = *ptep;
1692 
1693 		if ((opte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
1694 			/*
1695 			 * Page tables must have the cache-mode set correctly.
1696 			 */
1697 			const pt_entry_t npte = (pte & ~L2_S_CACHE_MASK)
1698 			    | pte_l2_s_cache_mode_pt;
1699 			l2pte_set(ptep, npte, opte);
1700 			PTE_SYNC(ptep);
1701 			cpu_tlb_flushD_SE(va);
1702 			cpu_cpwait();
1703 		}
1704 	}
1705 #endif
1706 
1707 	memset(v, 0, L2_TABLE_SIZE_REAL);
1708 	PTE_SYNC_RANGE(v, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1709 	return (0);
1710 }
1711 
1712 static int
1713 pmap_l2dtable_ctor(void *arg, void *v, int flags)
1714 {
1715 
1716 	memset(v, 0, sizeof(struct l2_dtable));
1717 	return (0);
1718 }
1719 
1720 static int
1721 pmap_pmap_ctor(void *arg, void *v, int flags)
1722 {
1723 
1724 	memset(v, 0, sizeof(struct pmap));
1725 	return (0);
1726 }
1727 
1728 static void
1729 pmap_pinit(pmap_t pm)
1730 {
1731 #ifndef ARM_HAS_VBAR
1732 	struct l2_bucket *l2b;
1733 
1734 	if (vector_page < KERNEL_BASE) {
1735 		/*
1736 		 * Map the vector page.
1737 		 */
1738 		pmap_enter(pm, vector_page, systempage.pv_pa,
1739 		    VM_PROT_READ | VM_PROT_EXECUTE,
1740 		    VM_PROT_READ | VM_PROT_EXECUTE | PMAP_WIRED);
1741 		pmap_update(pm);
1742 
1743 		pm->pm_pl1vec = pmap_l1_kva(pm) + l1pte_index(vector_page);
1744 		l2b = pmap_get_l2_bucket(pm, vector_page);
1745 		KASSERTMSG(l2b != NULL, "%#lx", vector_page);
1746 		pm->pm_l1vec = l2b->l2b_pa | L1_C_PROTO |
1747 		    L1_C_DOM(pmap_domain(pm));
1748 	} else
1749 		pm->pm_pl1vec = NULL;
1750 #endif
1751 }
1752 
1753 #ifdef PMAP_CACHE_VIVT
1754 /*
1755  * Since we have a virtually indexed cache, we may need to inhibit caching if
1756  * there is more than one mapping and at least one of them is writable.
1757  * Since we purge the cache on every context switch, we only need to check for
1758  * other mappings within the same pmap, or kernel_pmap.
1759  * This function is also called when a page is unmapped, to possibly reenable
1760  * caching on any remaining mappings.
1761  *
1762  * The code implements the following logic, where:
1763  *
1764  * KW = # of kernel read/write pages
1765  * KR = # of kernel read only pages
1766  * UW = # of user read/write pages
1767  * UR = # of user read only pages
1768  *
1769  * KC = kernel mapping is cacheable
1770  * UC = user mapping is cacheable
1771  *
1772  *               KW=0,KR=0  KW=0,KR>0  KW=1,KR=0  KW>1,KR>=0
1773  *             +---------------------------------------------
1774  * UW=0,UR=0   | ---        KC=1       KC=1       KC=0
1775  * UW=0,UR>0   | UC=1       KC=1,UC=1  KC=0,UC=0  KC=0,UC=0
1776  * UW=1,UR=0   | UC=1       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1777  * UW>1,UR>=0  | UC=0       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1778  */
1779 
1780 static const int pmap_vac_flags[4][4] = {
1781 	{-1,		0,		0,		PVF_KNC},
1782 	{0,		0,		PVF_NC,		PVF_NC},
1783 	{0,		PVF_NC,		PVF_NC,		PVF_NC},
1784 	{PVF_UNC,	PVF_NC,		PVF_NC,		PVF_NC}
1785 };
1786 
1787 static inline int
1788 pmap_get_vac_flags(const struct vm_page_md *md)
1789 {
1790 	int kidx, uidx;
1791 
1792 	kidx = 0;
1793 	if (md->kro_mappings || md->krw_mappings > 1)
1794 		kidx |= 1;
1795 	if (md->krw_mappings)
1796 		kidx |= 2;
1797 
1798 	uidx = 0;
1799 	if (md->uro_mappings || md->urw_mappings > 1)
1800 		uidx |= 1;
1801 	if (md->urw_mappings)
1802 		uidx |= 2;
1803 
1804 	return (pmap_vac_flags[uidx][kidx]);
1805 }
1806 
1807 static inline void
1808 pmap_vac_me_harder(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1809 {
1810 	int nattr;
1811 
1812 	nattr = pmap_get_vac_flags(md);
1813 
1814 	if (nattr < 0) {
1815 		md->pvh_attrs &= ~PVF_NC;
1816 		return;
1817 	}
1818 
1819 	if (nattr == 0 && (md->pvh_attrs & PVF_NC) == 0)
1820 		return;
1821 
1822 	if (pm == pmap_kernel())
1823 		pmap_vac_me_kpmap(md, pa, pm, va);
1824 	else
1825 		pmap_vac_me_user(md, pa, pm, va);
1826 
1827 	md->pvh_attrs = (md->pvh_attrs & ~PVF_NC) | nattr;
1828 }
1829 
1830 static void
1831 pmap_vac_me_kpmap(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1832 {
1833 	u_int u_cacheable, u_entries;
1834 	struct pv_entry *pv;
1835 	pmap_t last_pmap = pm;
1836 
1837 	/*
1838 	 * Pass one, see if there are both kernel and user pmaps for
1839 	 * this page.  Calculate whether there are user-writable or
1840 	 * kernel-writable pages.
1841 	 */
1842 	u_cacheable = 0;
1843 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1844 		if (pv->pv_pmap != pm && (pv->pv_flags & PVF_NC) == 0)
1845 			u_cacheable++;
1846 	}
1847 
1848 	u_entries = md->urw_mappings + md->uro_mappings;
1849 
1850 	/*
1851 	 * We know we have just been updating a kernel entry, so if
1852 	 * all user pages are already cacheable, then there is nothing
1853 	 * further to do.
1854 	 */
1855 	if (md->k_mappings == 0 && u_cacheable == u_entries)
1856 		return;
1857 
1858 	if (u_entries) {
1859 		/*
1860 		 * Scan over the list again, for each entry, if it
1861 		 * might not be set correctly, call pmap_vac_me_user
1862 		 * to recalculate the settings.
1863 		 */
1864 		SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1865 			/*
1866 			 * We know kernel mappings will get set
1867 			 * correctly in other calls.  We also know
1868 			 * that if the pmap is the same as last_pmap
1869 			 * then we've just handled this entry.
1870 			 */
1871 			if (pv->pv_pmap == pm || pv->pv_pmap == last_pmap)
1872 				continue;
1873 
1874 			/*
1875 			 * If there are kernel entries and this page
1876 			 * is writable but non-cacheable, then we can
1877 			 * skip this entry also.
1878 			 */
1879 			if (md->k_mappings &&
1880 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) ==
1881 			    (PVF_NC | PVF_WRITE))
1882 				continue;
1883 
1884 			/*
1885 			 * Similarly if there are no kernel-writable
1886 			 * entries and the page is already
1887 			 * read-only/cacheable.
1888 			 */
1889 			if (md->krw_mappings == 0 &&
1890 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 0)
1891 				continue;
1892 
1893 			/*
1894 			 * For some of the remaining cases, we know
1895 			 * that we must recalculate, but for others we
1896 			 * can't tell if they are correct or not, so
1897 			 * we recalculate anyway.
1898 			 */
1899 			pmap_vac_me_user(md, pa, (last_pmap = pv->pv_pmap), 0);
1900 		}
1901 
1902 		if (md->k_mappings == 0)
1903 			return;
1904 	}
1905 
1906 	pmap_vac_me_user(md, pa, pm, va);
1907 }
1908 
1909 static void
1910 pmap_vac_me_user(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
1911 {
1912 	pmap_t kpmap = pmap_kernel();
1913 	struct pv_entry *pv, *npv = NULL;
1914 	u_int entries = 0;
1915 	u_int writable = 0;
1916 	u_int cacheable_entries = 0;
1917 	u_int kern_cacheable = 0;
1918 	u_int other_writable = 0;
1919 
1920 	/*
1921 	 * Count mappings and writable mappings in this pmap.
1922 	 * Include kernel mappings as part of our own.
1923 	 * Keep a pointer to the first one.
1924 	 */
1925 	npv = NULL;
1926 	KASSERT(pmap_page_locked_p(md));
1927 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
1928 		/* Count mappings in the same pmap */
1929 		if (pm == pv->pv_pmap || kpmap == pv->pv_pmap) {
1930 			if (entries++ == 0)
1931 				npv = pv;
1932 
1933 			/* Cacheable mappings */
1934 			if ((pv->pv_flags & PVF_NC) == 0) {
1935 				cacheable_entries++;
1936 				if (kpmap == pv->pv_pmap)
1937 					kern_cacheable++;
1938 			}
1939 
1940 			/* Writable mappings */
1941 			if (pv->pv_flags & PVF_WRITE)
1942 				++writable;
1943 		} else
1944 		if (pv->pv_flags & PVF_WRITE)
1945 			other_writable = 1;
1946 	}
1947 
1948 	/*
1949 	 * Enable or disable caching as necessary.
1950 	 * Note: the first entry might be part of the kernel pmap,
1951 	 * so we can't assume this is indicative of the state of the
1952 	 * other (maybe non-kpmap) entries.
1953 	 */
1954 	if ((entries > 1 && writable) ||
1955 	    (entries > 0 && pm == kpmap && other_writable)) {
1956 		if (cacheable_entries == 0) {
1957 			return;
1958 		}
1959 
1960 		for (pv = npv; pv; pv = SLIST_NEXT(pv, pv_link)) {
1961 			if ((pm != pv->pv_pmap && kpmap != pv->pv_pmap) ||
1962 			    (pv->pv_flags & PVF_NC))
1963 				continue;
1964 
1965 			pv->pv_flags |= PVF_NC;
1966 
1967 			struct l2_bucket * const l2b
1968 			    = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
1969 			KASSERTMSG(l2b != NULL, "%#lx", va);
1970 			pt_entry_t * const ptep
1971 			    = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1972 			const pt_entry_t opte = *ptep;
1973 			pt_entry_t npte = opte & ~L2_S_CACHE_MASK;
1974 
1975 			if ((va != pv->pv_va || pm != pv->pv_pmap)
1976 			    && l2pte_valid_p(npte)) {
1977 #ifdef PMAP_CACHE_VIVT
1978 				pmap_cache_wbinv_page(pv->pv_pmap, pv->pv_va,
1979 				    true, pv->pv_flags);
1980 #endif
1981 				pmap_tlb_flush_SE(pv->pv_pmap, pv->pv_va,
1982 				    pv->pv_flags);
1983 			}
1984 
1985 			l2pte_set(ptep, npte, opte);
1986 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
1987 		}
1988 		cpu_cpwait();
1989 	} else
1990 	if (entries > cacheable_entries) {
1991 		/*
1992 		 * Turn cacheing back on for some pages.  If it is a kernel
1993 		 * page, only do so if there are no other writable pages.
1994 		 */
1995 		for (pv = npv; pv; pv = SLIST_NEXT(pv, pv_link)) {
1996 			if (!(pv->pv_flags & PVF_NC) || (pm != pv->pv_pmap &&
1997 			    (kpmap != pv->pv_pmap || other_writable)))
1998 				continue;
1999 
2000 			pv->pv_flags &= ~PVF_NC;
2001 
2002 			struct l2_bucket * const l2b
2003 			    = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
2004 			KASSERTMSG(l2b != NULL, "%#lx", va);
2005 			pt_entry_t * const ptep
2006 			    = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2007 			const pt_entry_t opte = *ptep;
2008 			pt_entry_t npte = (opte & ~L2_S_CACHE_MASK)
2009 			    | pte_l2_s_cache_mode;
2010 
2011 			if (l2pte_valid_p(opte)) {
2012 				pmap_tlb_flush_SE(pv->pv_pmap, pv->pv_va,
2013 				    pv->pv_flags);
2014 			}
2015 
2016 			l2pte_set(ptep, npte, opte);
2017 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
2018 		}
2019 	}
2020 }
2021 #endif
2022 
2023 #ifdef PMAP_CACHE_VIPT
2024 static void
2025 pmap_vac_me_harder(struct vm_page_md *md, paddr_t pa, pmap_t pm, vaddr_t va)
2026 {
2027 #ifndef ARM_MMU_EXTENDED
2028 	struct pv_entry *pv;
2029 	vaddr_t tst_mask;
2030 	bool bad_alias;
2031 	const u_int
2032 	    rw_mappings = md->urw_mappings + md->krw_mappings,
2033 	    ro_mappings = md->uro_mappings + md->kro_mappings;
2034 
2035 	/* do we need to do anything? */
2036 	if (arm_cache_prefer_mask == 0)
2037 		return;
2038 
2039 	NPDEBUG(PDB_VAC, printf("pmap_vac_me_harder: md=%p, pmap=%p va=%08lx\n",
2040 	    md, pm, va));
2041 
2042 	KASSERT(!va || pm);
2043 	KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2044 
2045 	/* Already a conflict? */
2046 	if (__predict_false(md->pvh_attrs & PVF_NC)) {
2047 		/* just an add, things are already non-cached */
2048 		KASSERT(!(md->pvh_attrs & PVF_DIRTY));
2049 		KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2050 		bad_alias = false;
2051 		if (va) {
2052 			PMAPCOUNT(vac_color_none);
2053 			bad_alias = true;
2054 			KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2055 			goto fixup;
2056 		}
2057 		pv = SLIST_FIRST(&md->pvh_list);
2058 		/* the list can't be empty because it would be cachable */
2059 		if (md->pvh_attrs & PVF_KMPAGE) {
2060 			tst_mask = md->pvh_attrs;
2061 		} else {
2062 			KASSERT(pv);
2063 			tst_mask = pv->pv_va;
2064 			pv = SLIST_NEXT(pv, pv_link);
2065 		}
2066 		/*
2067 		 * Only check for a bad alias if we have writable mappings.
2068 		 */
2069 		tst_mask &= arm_cache_prefer_mask;
2070 		if (rw_mappings > 0) {
2071 			for (; pv && !bad_alias; pv = SLIST_NEXT(pv, pv_link)) {
2072 				/* if there's a bad alias, stop checking. */
2073 				if (tst_mask != (pv->pv_va & arm_cache_prefer_mask))
2074 					bad_alias = true;
2075 			}
2076 			md->pvh_attrs |= PVF_WRITE;
2077 			if (!bad_alias)
2078 				md->pvh_attrs |= PVF_DIRTY;
2079 		} else {
2080 			/*
2081 			 * We have only read-only mappings.  Let's see if there
2082 			 * are multiple colors in use or if we mapped a KMPAGE.
2083 			 * If the latter, we have a bad alias.  If the former,
2084 			 * we need to remember that.
2085 			 */
2086 			for (; pv; pv = SLIST_NEXT(pv, pv_link)) {
2087 				if (tst_mask != (pv->pv_va & arm_cache_prefer_mask)) {
2088 					if (md->pvh_attrs & PVF_KMPAGE)
2089 						bad_alias = true;
2090 					break;
2091 				}
2092 			}
2093 			md->pvh_attrs &= ~PVF_WRITE;
2094 			/*
2095 			 * No KMPAGE and we exited early, so we must have
2096 			 * multiple color mappings.
2097 			 */
2098 			if (!bad_alias && pv != NULL)
2099 				md->pvh_attrs |= PVF_MULTCLR;
2100 		}
2101 
2102 		/* If no conflicting colors, set everything back to cached */
2103 		if (!bad_alias) {
2104 #ifdef DEBUG
2105 			if ((md->pvh_attrs & PVF_WRITE)
2106 			    || ro_mappings < 2) {
2107 				SLIST_FOREACH(pv, &md->pvh_list, pv_link)
2108 					KDASSERT(((tst_mask ^ pv->pv_va) & arm_cache_prefer_mask) == 0);
2109 			}
2110 #endif
2111 			md->pvh_attrs &= (PAGE_SIZE - 1) & ~PVF_NC;
2112 			md->pvh_attrs |= tst_mask | PVF_COLORED;
2113 			/*
2114 			 * Restore DIRTY bit if page is modified
2115 			 */
2116 			if (md->pvh_attrs & PVF_DMOD)
2117 				md->pvh_attrs |= PVF_DIRTY;
2118 			PMAPCOUNT(vac_color_restore);
2119 		} else {
2120 			KASSERT(SLIST_FIRST(&md->pvh_list) != NULL);
2121 			KASSERT(SLIST_NEXT(SLIST_FIRST(&md->pvh_list), pv_link) != NULL);
2122 		}
2123 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2124 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2125 	} else if (!va) {
2126 		KASSERT(pmap_is_page_colored_p(md));
2127 		KASSERT(!(md->pvh_attrs & PVF_WRITE)
2128 		    || (md->pvh_attrs & PVF_DIRTY));
2129 		if (rw_mappings == 0) {
2130 			md->pvh_attrs &= ~PVF_WRITE;
2131 			if (ro_mappings == 1
2132 			    && (md->pvh_attrs & PVF_MULTCLR)) {
2133 				/*
2134 				 * If this is the last readonly mapping
2135 				 * but it doesn't match the current color
2136 				 * for the page, change the current color
2137 				 * to match this last readonly mapping.
2138 				 */
2139 				pv = SLIST_FIRST(&md->pvh_list);
2140 				tst_mask = (md->pvh_attrs ^ pv->pv_va)
2141 				    & arm_cache_prefer_mask;
2142 				if (tst_mask) {
2143 					md->pvh_attrs ^= tst_mask;
2144 					PMAPCOUNT(vac_color_change);
2145 				}
2146 			}
2147 		}
2148 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2149 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2150 		return;
2151 	} else if (!pmap_is_page_colored_p(md)) {
2152 		/* not colored so we just use its color */
2153 		KASSERT(md->pvh_attrs & (PVF_WRITE|PVF_DIRTY));
2154 		KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2155 		PMAPCOUNT(vac_color_new);
2156 		md->pvh_attrs &= PAGE_SIZE - 1;
2157 		md->pvh_attrs |= PVF_COLORED
2158 		    | (va & arm_cache_prefer_mask)
2159 		    | (rw_mappings > 0 ? PVF_WRITE : 0);
2160 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2161 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2162 		return;
2163 	} else if (((md->pvh_attrs ^ va) & arm_cache_prefer_mask) == 0) {
2164 		bad_alias = false;
2165 		if (rw_mappings > 0) {
2166 			/*
2167 			 * We now have writeable mappings and if we have
2168 			 * readonly mappings in more than once color, we have
2169 			 * an aliasing problem.  Regardless mark the page as
2170 			 * writeable.
2171 			 */
2172 			if (md->pvh_attrs & PVF_MULTCLR) {
2173 				if (ro_mappings < 2) {
2174 					/*
2175 					 * If we only have less than two
2176 					 * read-only mappings, just flush the
2177 					 * non-primary colors from the cache.
2178 					 */
2179 					pmap_flush_page(md, pa,
2180 					    PMAP_FLUSH_SECONDARY);
2181 				} else {
2182 					bad_alias = true;
2183 				}
2184 			}
2185 			md->pvh_attrs |= PVF_WRITE;
2186 		}
2187 		/* If no conflicting colors, set everything back to cached */
2188 		if (!bad_alias) {
2189 #ifdef DEBUG
2190 			if (rw_mappings > 0
2191 			    || (md->pvh_attrs & PMAP_KMPAGE)) {
2192 				tst_mask = md->pvh_attrs & arm_cache_prefer_mask;
2193 				SLIST_FOREACH(pv, &md->pvh_list, pv_link)
2194 					KDASSERT(((tst_mask ^ pv->pv_va) & arm_cache_prefer_mask) == 0);
2195 			}
2196 #endif
2197 			if (SLIST_EMPTY(&md->pvh_list))
2198 				PMAPCOUNT(vac_color_reuse);
2199 			else
2200 				PMAPCOUNT(vac_color_ok);
2201 
2202 			/* matching color, just return */
2203 			KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2204 			KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2205 			return;
2206 		}
2207 		KASSERT(SLIST_FIRST(&md->pvh_list) != NULL);
2208 		KASSERT(SLIST_NEXT(SLIST_FIRST(&md->pvh_list), pv_link) != NULL);
2209 
2210 		/* color conflict.  evict from cache. */
2211 
2212 		pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
2213 		md->pvh_attrs &= ~PVF_COLORED;
2214 		md->pvh_attrs |= PVF_NC;
2215 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2216 		KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2217 		PMAPCOUNT(vac_color_erase);
2218 	} else if (rw_mappings == 0
2219 		   && (md->pvh_attrs & PVF_KMPAGE) == 0) {
2220 		KASSERT((md->pvh_attrs & PVF_WRITE) == 0);
2221 
2222 		/*
2223 		 * If the page has dirty cache lines, clean it.
2224 		 */
2225 		if (md->pvh_attrs & PVF_DIRTY)
2226 			pmap_flush_page(md, pa, PMAP_CLEAN_PRIMARY);
2227 
2228 		/*
2229 		 * If this is the first remapping (we know that there are no
2230 		 * writeable mappings), then this is a simple color change.
2231 		 * Otherwise this is a seconary r/o mapping, which means
2232 		 * we don't have to do anything.
2233 		 */
2234 		if (ro_mappings == 1) {
2235 			KASSERT(((md->pvh_attrs ^ va) & arm_cache_prefer_mask) != 0);
2236 			md->pvh_attrs &= PAGE_SIZE - 1;
2237 			md->pvh_attrs |= (va & arm_cache_prefer_mask);
2238 			PMAPCOUNT(vac_color_change);
2239 		} else {
2240 			PMAPCOUNT(vac_color_blind);
2241 		}
2242 		md->pvh_attrs |= PVF_MULTCLR;
2243 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2244 		KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2245 		return;
2246 	} else {
2247 		if (rw_mappings > 0)
2248 			md->pvh_attrs |= PVF_WRITE;
2249 
2250 		/* color conflict.  evict from cache. */
2251 		pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
2252 
2253 		/* the list can't be empty because this was a enter/modify */
2254 		pv = SLIST_FIRST(&md->pvh_list);
2255 		if ((md->pvh_attrs & PVF_KMPAGE) == 0) {
2256 			KASSERT(pv);
2257 			/*
2258 			 * If there's only one mapped page, change color to the
2259 			 * page's new color and return.  Restore the DIRTY bit
2260 			 * that was erased by pmap_flush_page.
2261 			 */
2262 			if (SLIST_NEXT(pv, pv_link) == NULL) {
2263 				md->pvh_attrs &= PAGE_SIZE - 1;
2264 				md->pvh_attrs |= (va & arm_cache_prefer_mask);
2265 				if (md->pvh_attrs & PVF_DMOD)
2266 					md->pvh_attrs |= PVF_DIRTY;
2267 				PMAPCOUNT(vac_color_change);
2268 				KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2269 				KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2270 				KASSERT(!(md->pvh_attrs & PVF_MULTCLR));
2271 				return;
2272 			}
2273 		}
2274 		bad_alias = true;
2275 		md->pvh_attrs &= ~PVF_COLORED;
2276 		md->pvh_attrs |= PVF_NC;
2277 		PMAPCOUNT(vac_color_erase);
2278 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
2279 	}
2280 
2281   fixup:
2282 	KASSERT((rw_mappings == 0) == !(md->pvh_attrs & PVF_WRITE));
2283 
2284 	/*
2285 	 * Turn cacheing on/off for all pages.
2286 	 */
2287 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
2288 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pv->pv_pmap,
2289 		    pv->pv_va);
2290 		KASSERTMSG(l2b != NULL, "%#lx", va);
2291 		pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2292 		const pt_entry_t opte = *ptep;
2293 		pt_entry_t npte = opte & ~L2_S_CACHE_MASK;
2294 		if (bad_alias) {
2295 			pv->pv_flags |= PVF_NC;
2296 		} else {
2297 			pv->pv_flags &= ~PVF_NC;
2298 			npte |= pte_l2_s_cache_mode;
2299 		}
2300 
2301 		if (opte == npte)	/* only update is there's a change */
2302 			continue;
2303 
2304 		if (l2pte_valid_p(npte)) {
2305 			pmap_tlb_flush_SE(pv->pv_pmap, pv->pv_va, pv->pv_flags);
2306 		}
2307 
2308 		l2pte_set(ptep, npte, opte);
2309 		PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
2310 	}
2311 #endif /* !ARM_MMU_EXTENDED */
2312 }
2313 #endif	/* PMAP_CACHE_VIPT */
2314 
2315 
2316 /*
2317  * Modify pte bits for all ptes corresponding to the given physical address.
2318  * We use `maskbits' rather than `clearbits' because we're always passing
2319  * constants and the latter would require an extra inversion at run-time.
2320  */
2321 static void
2322 pmap_clearbit(struct vm_page_md *md, paddr_t pa, u_int maskbits)
2323 {
2324 	struct pv_entry *pv;
2325 #ifdef PMAP_CACHE_VIPT
2326 	const bool want_syncicache = PV_IS_EXEC_P(md->pvh_attrs);
2327 #ifdef ARM_MMU_EXTENDED
2328 	const u_int execbits = (maskbits & PVF_EXEC) ? L2_XS_XN : 0;
2329 #else
2330 	const u_int execbits = 0;
2331 	bool need_vac_me_harder = false;
2332 	bool need_syncicache = false;
2333 #endif
2334 #else
2335 	const u_int execbits = 0;
2336 #endif
2337 
2338 	NPDEBUG(PDB_BITS,
2339 	    printf("pmap_clearbit: md %p mask 0x%x\n",
2340 	    md, maskbits));
2341 
2342 #ifdef PMAP_CACHE_VIPT
2343 	/*
2344 	 * If we might want to sync the I-cache and we've modified it,
2345 	 * then we know we definitely need to sync or discard it.
2346 	 */
2347 	if (want_syncicache) {
2348 #ifdef ARM_MMU_EXTENDED
2349 		if (md->pvh_attrs & PVF_MOD)
2350 			md->pvh_attrs &= ~PVF_EXEC;
2351 #else
2352 		need_syncicache = md->pvh_attrs & PVF_MOD;
2353 #endif
2354 	}
2355 #endif
2356 	KASSERT(pmap_page_locked_p(md));
2357 
2358 	/*
2359 	 * Clear saved attributes (modify, reference)
2360 	 */
2361 	md->pvh_attrs &= ~(maskbits & (PVF_MOD | PVF_REF));
2362 
2363 	if (SLIST_EMPTY(&md->pvh_list)) {
2364 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
2365 		if (need_syncicache) {
2366 			/*
2367 			 * No one has it mapped, so just discard it.  The next
2368 			 * exec remapping will cause it to be synced.
2369 			 */
2370 			md->pvh_attrs &= ~PVF_EXEC;
2371 			PMAPCOUNT(exec_discarded_clearbit);
2372 		}
2373 #endif
2374 		return;
2375 	}
2376 
2377 	/*
2378 	 * Loop over all current mappings setting/clearing as appropos
2379 	 */
2380 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
2381 		pmap_t pm = pv->pv_pmap;
2382 		const vaddr_t va = pv->pv_va;
2383 		const u_int oflags = pv->pv_flags;
2384 #ifndef ARM_MMU_EXTENDED
2385 		/*
2386 		 * Kernel entries are unmanaged and as such not to be changed.
2387 		 */
2388 		if (PV_IS_KENTRY_P(oflags))
2389 			continue;
2390 #endif
2391 		pv->pv_flags &= ~maskbits;
2392 
2393 		pmap_release_page_lock(md);
2394 		pmap_acquire_pmap_lock(pm);
2395 
2396 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pm, va);
2397 		if (l2b == NULL) {
2398 			pmap_release_pmap_lock(pm);
2399 			pmap_acquire_page_lock(md);
2400 			continue;
2401 		}
2402 		KASSERTMSG(l2b != NULL, "%#lx", va);
2403 
2404 		pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
2405 		const pt_entry_t opte = *ptep;
2406 		pt_entry_t npte = opte | execbits;
2407 
2408 #ifdef ARM_MMU_EXTENDED
2409 		KASSERT((opte & L2_XS_nG) == (pm == pmap_kernel() ? 0 : L2_XS_nG));
2410 #endif
2411 
2412 		NPDEBUG(PDB_BITS,
2413 		    printf( "%s: pv %p, pm %p, va 0x%08lx, flag 0x%x\n",
2414 			__func__, pv, pm, va, oflags));
2415 
2416 		if (maskbits & (PVF_WRITE|PVF_MOD)) {
2417 #ifdef PMAP_CACHE_VIVT
2418 			if ((oflags & PVF_NC)) {
2419 				/*
2420 				 * Entry is not cacheable:
2421 				 *
2422 				 * Don't turn caching on again if this is a
2423 				 * modified emulation. This would be
2424 				 * inconsitent with the settings created by
2425 				 * pmap_vac_me_harder(). Otherwise, it's safe
2426 				 * to re-enable cacheing.
2427 				 *
2428 				 * There's no need to call pmap_vac_me_harder()
2429 				 * here: all pages are losing their write
2430 				 * permission.
2431 				 */
2432 				if (maskbits & PVF_WRITE) {
2433 					npte |= pte_l2_s_cache_mode;
2434 					pv->pv_flags &= ~PVF_NC;
2435 				}
2436 			} else
2437 			if (l2pte_writable_p(opte)) {
2438 				/*
2439 				 * Entry is writable/cacheable: check if pmap
2440 				 * is current if it is flush it, otherwise it
2441 				 * won't be in the cache
2442 				 */
2443 				pmap_cache_wbinv_page(pm, va,
2444 				    (maskbits & PVF_REF) != 0,
2445 				    oflags|PVF_WRITE);
2446 			}
2447 #endif
2448 
2449 			/* make the pte read only */
2450 			npte = l2pte_set_readonly(npte);
2451 
2452 			pmap_acquire_page_lock(md);
2453 #ifdef MULTIPROCESSOR
2454 			pv = pmap_find_pv(md, pm, va);
2455 #endif
2456 			if (pv != NULL && (maskbits & oflags & PVF_WRITE)) {
2457 				/*
2458 				 * Keep alias accounting up to date
2459 				 */
2460 				if (pm == pmap_kernel()) {
2461 					md->krw_mappings--;
2462 					md->kro_mappings++;
2463 				} else {
2464 					md->urw_mappings--;
2465 					md->uro_mappings++;
2466 				}
2467 #ifdef PMAP_CACHE_VIPT
2468 				if (arm_cache_prefer_mask != 0) {
2469 					if (md->urw_mappings + md->krw_mappings == 0) {
2470 						md->pvh_attrs &= ~PVF_WRITE;
2471 					} else {
2472 						PMAP_VALIDATE_MD_PAGE(md);
2473 					}
2474 				}
2475 #ifndef ARM_MMU_EXTENDED
2476 				if (want_syncicache)
2477 					need_syncicache = true;
2478 				need_vac_me_harder = true;
2479 #endif
2480 #endif /* PMAP_CACHE_VIPT */
2481 			}
2482 			pmap_release_page_lock(md);
2483 		}
2484 
2485 		if (maskbits & PVF_REF) {
2486 			if (true
2487 #ifndef ARM_MMU_EXTENDED
2488 			    && (oflags & PVF_NC) == 0
2489 #endif
2490 			    && (maskbits & (PVF_WRITE|PVF_MOD)) == 0
2491 			    && l2pte_valid_p(npte)) {
2492 #ifdef PMAP_CACHE_VIVT
2493 				/*
2494 				 * Check npte here; we may have already
2495 				 * done the wbinv above, and the validity
2496 				 * of the PTE is the same for opte and
2497 				 * npte.
2498 				 */
2499 				pmap_cache_wbinv_page(pm, va, true, oflags);
2500 #endif
2501 			}
2502 
2503 			/*
2504 			 * Make the PTE invalid so that we will take a
2505 			 * page fault the next time the mapping is
2506 			 * referenced.
2507 			 */
2508 			npte &= ~L2_TYPE_MASK;
2509 			npte |= L2_TYPE_INV;
2510 		}
2511 
2512 		if (npte != opte) {
2513 			l2pte_reset(ptep);
2514 			PTE_SYNC(ptep);
2515 
2516 			/* Flush the TLB entry if a current pmap. */
2517 			pmap_tlb_flush_SE(pm, va, oflags);
2518 
2519 			l2pte_set(ptep, npte, 0);
2520 			PTE_SYNC(ptep);
2521 		}
2522 
2523 		pmap_release_pmap_lock(pm);
2524 		pmap_acquire_page_lock(md);
2525 
2526 		NPDEBUG(PDB_BITS,
2527 		    printf("pmap_clearbit: pm %p va 0x%lx opte 0x%08x npte 0x%08x\n",
2528 		    pm, va, opte, npte));
2529 	}
2530 
2531 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
2532 	/*
2533 	 * If we need to sync the I-cache and we haven't done it yet, do it.
2534 	 */
2535 	if (need_syncicache) {
2536 		pmap_release_page_lock(md);
2537 		pmap_syncicache_page(md, pa);
2538 		pmap_acquire_page_lock(md);
2539 		PMAPCOUNT(exec_synced_clearbit);
2540 	}
2541 
2542 	/*
2543 	 * If we are changing this to read-only, we need to call vac_me_harder
2544 	 * so we can change all the read-only pages to cacheable.  We pretend
2545 	 * this as a page deletion.
2546 	 */
2547 	if (need_vac_me_harder) {
2548 		if (md->pvh_attrs & PVF_NC)
2549 			pmap_vac_me_harder(md, pa, NULL, 0);
2550 	}
2551 #endif /* PMAP_CACHE_VIPT && !ARM_MMU_EXTENDED */
2552 }
2553 
2554 /*
2555  * pmap_clean_page()
2556  *
2557  * This is a local function used to work out the best strategy to clean
2558  * a single page referenced by its entry in the PV table. It's used by
2559  * pmap_copy_page, pmap_zero_page and maybe some others later on.
2560  *
2561  * Its policy is effectively:
2562  *  o If there are no mappings, we don't bother doing anything with the cache.
2563  *  o If there is one mapping, we clean just that page.
2564  *  o If there are multiple mappings, we clean the entire cache.
2565  *
2566  * So that some functions can be further optimised, it returns 0 if it didn't
2567  * clean the entire cache, or 1 if it did.
2568  *
2569  * XXX One bug in this routine is that if the pv_entry has a single page
2570  * mapped at 0x00000000 a whole cache clean will be performed rather than
2571  * just the 1 page. Since this should not occur in everyday use and if it does
2572  * it will just result in not the most efficient clean for the page.
2573  */
2574 #ifdef PMAP_CACHE_VIVT
2575 static bool
2576 pmap_clean_page(struct vm_page_md *md, bool is_src)
2577 {
2578 	struct pv_entry *pv;
2579 	pmap_t pm_to_clean = NULL;
2580 	bool cache_needs_cleaning = false;
2581 	vaddr_t page_to_clean = 0;
2582 	u_int flags = 0;
2583 
2584 	/*
2585 	 * Since we flush the cache each time we change to a different
2586 	 * user vmspace, we only need to flush the page if it is in the
2587 	 * current pmap.
2588 	 */
2589 	KASSERT(pmap_page_locked_p(md));
2590 	SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
2591 		if (pmap_is_current(pv->pv_pmap)) {
2592 			flags |= pv->pv_flags;
2593 			/*
2594 			 * The page is mapped non-cacheable in
2595 			 * this map.  No need to flush the cache.
2596 			 */
2597 			if (pv->pv_flags & PVF_NC) {
2598 #ifdef DIAGNOSTIC
2599 				KASSERT(!cache_needs_cleaning);
2600 #endif
2601 				break;
2602 			} else if (is_src && (pv->pv_flags & PVF_WRITE) == 0)
2603 				continue;
2604 			if (cache_needs_cleaning) {
2605 				page_to_clean = 0;
2606 				break;
2607 			} else {
2608 				page_to_clean = pv->pv_va;
2609 				pm_to_clean = pv->pv_pmap;
2610 			}
2611 			cache_needs_cleaning = true;
2612 		}
2613 	}
2614 
2615 	if (page_to_clean) {
2616 		pmap_cache_wbinv_page(pm_to_clean, page_to_clean,
2617 		    !is_src, flags | PVF_REF);
2618 	} else if (cache_needs_cleaning) {
2619 		pmap_t const pm = curproc->p_vmspace->vm_map.pmap;
2620 
2621 		pmap_cache_wbinv_all(pm, flags);
2622 		return true;
2623 	}
2624 	return false;
2625 }
2626 #endif
2627 
2628 #ifdef PMAP_CACHE_VIPT
2629 /*
2630  * Sync a page with the I-cache.  Since this is a VIPT, we must pick the
2631  * right cache alias to make sure we flush the right stuff.
2632  */
2633 void
2634 pmap_syncicache_page(struct vm_page_md *md, paddr_t pa)
2635 {
2636 	pmap_t kpm = pmap_kernel();
2637 	const size_t way_size = arm_pcache.icache_type == CACHE_TYPE_PIPT
2638 	    ? PAGE_SIZE
2639 	    : arm_pcache.icache_way_size;
2640 
2641 	NPDEBUG(PDB_EXEC, printf("pmap_syncicache_page: md=%p (attrs=%#x)\n",
2642 	    md, md->pvh_attrs));
2643 	/*
2644 	 * No need to clean the page if it's non-cached.
2645 	 */
2646 #ifndef ARM_MMU_EXTENDED
2647 	if (md->pvh_attrs & PVF_NC)
2648 		return;
2649 	KASSERT(arm_cache_prefer_mask == 0 || md->pvh_attrs & PVF_COLORED);
2650 #endif
2651 
2652 	pt_entry_t * const ptep = cpu_cdst_pte(0);
2653 	const vaddr_t dstp = cpu_cdstp(0);
2654 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
2655 	if (way_size <= PAGE_SIZE) {
2656 		bool ok = false;
2657 		vaddr_t vdstp = pmap_direct_mapped_phys(pa, &ok, dstp);
2658 		if (ok) {
2659 			cpu_icache_sync_range(vdstp, way_size);
2660 			return;
2661 		}
2662 	}
2663 #endif
2664 
2665 	/*
2666 	 * We don't worry about the color of the exec page, we map the
2667 	 * same page to pages in the way and then do the icache_sync on
2668 	 * the entire way making sure we are cleaned.
2669 	 */
2670 	const pt_entry_t npte = L2_S_PROTO | pa | pte_l2_s_cache_mode
2671 	    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE);
2672 
2673 	for (size_t i = 0, j = 0; i < way_size;
2674 	     i += PAGE_SIZE, j += PAGE_SIZE / L2_S_SIZE) {
2675 		l2pte_reset(ptep + j);
2676 		PTE_SYNC(ptep + j);
2677 
2678 		pmap_tlb_flush_SE(kpm, dstp + i, PVF_REF | PVF_EXEC);
2679 		/*
2680 		 * Set up a PTE with to flush these cache lines.
2681 		 */
2682 		l2pte_set(ptep + j, npte, 0);
2683 	}
2684 	PTE_SYNC_RANGE(ptep, way_size / L2_S_SIZE);
2685 
2686 	/*
2687 	 * Flush it.
2688 	 */
2689 	cpu_icache_sync_range(dstp, way_size);
2690 
2691 	for (size_t i = 0, j = 0; i < way_size;
2692 	     i += PAGE_SIZE, j += PAGE_SIZE / L2_S_SIZE) {
2693 		/*
2694 		 * Unmap the page(s).
2695 		 */
2696 		l2pte_reset(ptep + j);
2697 		pmap_tlb_flush_SE(kpm, dstp + i, PVF_REF | PVF_EXEC);
2698 	}
2699 	PTE_SYNC_RANGE(ptep, way_size / L2_S_SIZE);
2700 
2701 	md->pvh_attrs |= PVF_EXEC;
2702 	PMAPCOUNT(exec_synced);
2703 }
2704 
2705 #ifndef ARM_MMU_EXTENDED
2706 void
2707 pmap_flush_page(struct vm_page_md *md, paddr_t pa, enum pmap_flush_op flush)
2708 {
2709 	vsize_t va_offset, end_va;
2710 	bool wbinv_p;
2711 
2712 	if (arm_cache_prefer_mask == 0)
2713 		return;
2714 
2715 	switch (flush) {
2716 	case PMAP_FLUSH_PRIMARY:
2717 		if (md->pvh_attrs & PVF_MULTCLR) {
2718 			va_offset = 0;
2719 			end_va = arm_cache_prefer_mask;
2720 			md->pvh_attrs &= ~PVF_MULTCLR;
2721 			PMAPCOUNT(vac_flush_lots);
2722 		} else {
2723 			va_offset = md->pvh_attrs & arm_cache_prefer_mask;
2724 			end_va = va_offset;
2725 			PMAPCOUNT(vac_flush_one);
2726 		}
2727 		/*
2728 		 * Mark that the page is no longer dirty.
2729 		 */
2730 		md->pvh_attrs &= ~PVF_DIRTY;
2731 		wbinv_p = true;
2732 		break;
2733 	case PMAP_FLUSH_SECONDARY:
2734 		va_offset = 0;
2735 		end_va = arm_cache_prefer_mask;
2736 		wbinv_p = true;
2737 		md->pvh_attrs &= ~PVF_MULTCLR;
2738 		PMAPCOUNT(vac_flush_lots);
2739 		break;
2740 	case PMAP_CLEAN_PRIMARY:
2741 		va_offset = md->pvh_attrs & arm_cache_prefer_mask;
2742 		end_va = va_offset;
2743 		wbinv_p = false;
2744 		/*
2745 		 * Mark that the page is no longer dirty.
2746 		 */
2747 		if ((md->pvh_attrs & PVF_DMOD) == 0)
2748 			md->pvh_attrs &= ~PVF_DIRTY;
2749 		PMAPCOUNT(vac_clean_one);
2750 		break;
2751 	default:
2752 		return;
2753 	}
2754 
2755 	KASSERT(!(md->pvh_attrs & PVF_NC));
2756 
2757 	NPDEBUG(PDB_VAC, printf("pmap_flush_page: md=%p (attrs=%#x)\n",
2758 	    md, md->pvh_attrs));
2759 
2760 	const size_t scache_line_size = arm_scache.dcache_line_size;
2761 
2762 	for (; va_offset <= end_va; va_offset += PAGE_SIZE) {
2763 		pt_entry_t * const ptep = cpu_cdst_pte(va_offset);
2764 		const vaddr_t dstp = cpu_cdstp(va_offset);
2765 		const pt_entry_t opte = *ptep;
2766 
2767 		if (flush == PMAP_FLUSH_SECONDARY
2768 		    && va_offset == (md->pvh_attrs & arm_cache_prefer_mask))
2769 			continue;
2770 
2771 		pmap_tlb_flush_SE(pmap_kernel(), dstp, PVF_REF | PVF_EXEC);
2772 		/*
2773 		 * Set up a PTE with the right coloring to flush
2774 		 * existing cache entries.
2775 		 */
2776 		const pt_entry_t npte = L2_S_PROTO
2777 		    | pa
2778 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ|VM_PROT_WRITE)
2779 		    | pte_l2_s_cache_mode;
2780 		l2pte_set(ptep, npte, opte);
2781 		PTE_SYNC(ptep);
2782 
2783 		/*
2784 		 * Flush it.  Make sure to flush secondary cache too since
2785 		 * bus_dma will ignore uncached pages.
2786 		 */
2787 		if (scache_line_size != 0) {
2788 			cpu_dcache_wb_range(dstp, PAGE_SIZE);
2789 			if (wbinv_p) {
2790 				cpu_sdcache_wbinv_range(dstp, pa, PAGE_SIZE);
2791 				cpu_dcache_inv_range(dstp, PAGE_SIZE);
2792 			} else {
2793 				cpu_sdcache_wb_range(dstp, pa, PAGE_SIZE);
2794 			}
2795 		} else {
2796 			if (wbinv_p) {
2797 				cpu_dcache_wbinv_range(dstp, PAGE_SIZE);
2798 			} else {
2799 				cpu_dcache_wb_range(dstp, PAGE_SIZE);
2800 			}
2801 		}
2802 
2803 		/*
2804 		 * Restore the page table entry since we might have interrupted
2805 		 * pmap_zero_page or pmap_copy_page which was already using
2806 		 * this pte.
2807 		 */
2808 		if (opte) {
2809 			l2pte_set(ptep, opte, npte);
2810 		} else {
2811 			l2pte_reset(ptep);
2812 		}
2813 		PTE_SYNC(ptep);
2814 		pmap_tlb_flush_SE(pmap_kernel(), dstp, PVF_REF | PVF_EXEC);
2815 	}
2816 }
2817 #endif /* ARM_MMU_EXTENDED */
2818 #endif /* PMAP_CACHE_VIPT */
2819 
2820 /*
2821  * Routine:	pmap_page_remove
2822  * Function:
2823  *		Removes this physical page from
2824  *		all physical maps in which it resides.
2825  *		Reflects back modify bits to the pager.
2826  */
2827 static void
2828 pmap_page_remove(struct vm_page_md *md, paddr_t pa)
2829 {
2830 	struct l2_bucket *l2b;
2831 	struct pv_entry *pv;
2832 	pt_entry_t *ptep;
2833 #ifndef ARM_MMU_EXTENDED
2834 	bool flush = false;
2835 #endif
2836 	u_int flags = 0;
2837 
2838 	NPDEBUG(PDB_FOLLOW,
2839 	    printf("pmap_page_remove: md %p (0x%08lx)\n", md,
2840 	    pa));
2841 
2842 	struct pv_entry **pvp = &SLIST_FIRST(&md->pvh_list);
2843 	pmap_acquire_page_lock(md);
2844 	if (*pvp == NULL) {
2845 #ifdef PMAP_CACHE_VIPT
2846 		/*
2847 		 * We *know* the page contents are about to be replaced.
2848 		 * Discard the exec contents
2849 		 */
2850 		if (PV_IS_EXEC_P(md->pvh_attrs))
2851 			PMAPCOUNT(exec_discarded_page_protect);
2852 		md->pvh_attrs &= ~PVF_EXEC;
2853 		PMAP_VALIDATE_MD_PAGE(md);
2854 #endif
2855 		pmap_release_page_lock(md);
2856 		return;
2857 	}
2858 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
2859 	KASSERT(arm_cache_prefer_mask == 0 || pmap_is_page_colored_p(md));
2860 #endif
2861 
2862 	/*
2863 	 * Clear alias counts
2864 	 */
2865 #ifdef PMAP_CACHE_VIVT
2866 	md->k_mappings = 0;
2867 #endif
2868 	md->urw_mappings = md->uro_mappings = 0;
2869 
2870 #ifdef PMAP_CACHE_VIVT
2871 	pmap_clean_page(md, false);
2872 #endif
2873 
2874 	while ((pv = *pvp) != NULL) {
2875 		pmap_t pm = pv->pv_pmap;
2876 #ifndef ARM_MMU_EXTENDED
2877 		if (flush == false && pmap_is_current(pm))
2878 			flush = true;
2879 #endif
2880 
2881 		if (pm == pmap_kernel()) {
2882 #ifdef PMAP_CACHE_VIPT
2883 			/*
2884 			 * If this was unmanaged mapping, it must be preserved.
2885 			 * Move it back on the list and advance the end-of-list
2886 			 * pointer.
2887 			 */
2888 			if (PV_IS_KENTRY_P(pv->pv_flags)) {
2889 				*pvp = pv;
2890 				pvp = &SLIST_NEXT(pv, pv_link);
2891 				continue;
2892 			}
2893 			if (pv->pv_flags & PVF_WRITE)
2894 				md->krw_mappings--;
2895 			else
2896 				md->kro_mappings--;
2897 #endif
2898 			PMAPCOUNT(kernel_unmappings);
2899 		}
2900 		*pvp = SLIST_NEXT(pv, pv_link); /* remove from list */
2901 		PMAPCOUNT(unmappings);
2902 
2903 		pmap_release_page_lock(md);
2904 		pmap_acquire_pmap_lock(pm);
2905 
2906 		l2b = pmap_get_l2_bucket(pm, pv->pv_va);
2907 		KASSERTMSG(l2b != NULL, "%#lx", pv->pv_va);
2908 
2909 		ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
2910 
2911 		/*
2912 		 * Update statistics
2913 		 */
2914 		--pm->pm_stats.resident_count;
2915 
2916 		/* Wired bit */
2917 		if (pv->pv_flags & PVF_WIRED)
2918 			--pm->pm_stats.wired_count;
2919 
2920 		flags |= pv->pv_flags;
2921 
2922 		/*
2923 		 * Invalidate the PTEs.
2924 		 */
2925 		l2pte_reset(ptep);
2926 		PTE_SYNC_CURRENT(pm, ptep);
2927 
2928 #ifdef ARM_MMU_EXTENDED
2929 		pmap_tlb_invalidate_addr(pm, pv->pv_va);
2930 #endif
2931 
2932 		pmap_free_l2_bucket(pm, l2b, PAGE_SIZE / L2_S_SIZE);
2933 
2934 		pmap_release_pmap_lock(pm);
2935 
2936 		pool_put(&pmap_pv_pool, pv);
2937 		pmap_acquire_page_lock(md);
2938 #ifdef MULTIPROCESSOR
2939 		/*
2940 		 * Restart of the beginning of the list.
2941 		 */
2942 		pvp = &SLIST_FIRST(&md->pvh_list);
2943 #endif
2944 	}
2945 	/*
2946 	 * if we reach the end of the list and there are still mappings, they
2947 	 * might be able to be cached now.  And they must be kernel mappings.
2948 	 */
2949 	if (!SLIST_EMPTY(&md->pvh_list)) {
2950 		pmap_vac_me_harder(md, pa, pmap_kernel(), 0);
2951 	}
2952 
2953 #ifdef PMAP_CACHE_VIPT
2954 	/*
2955 	 * Its EXEC cache is now gone.
2956 	 */
2957 	if (PV_IS_EXEC_P(md->pvh_attrs))
2958 		PMAPCOUNT(exec_discarded_page_protect);
2959 	md->pvh_attrs &= ~PVF_EXEC;
2960 	KASSERT(md->urw_mappings == 0);
2961 	KASSERT(md->uro_mappings == 0);
2962 #ifndef ARM_MMU_EXTENDED
2963 	if (arm_cache_prefer_mask != 0) {
2964 		if (md->krw_mappings == 0)
2965 			md->pvh_attrs &= ~PVF_WRITE;
2966 		PMAP_VALIDATE_MD_PAGE(md);
2967 	}
2968 #endif /* ARM_MMU_EXTENDED */
2969 #endif /* PMAP_CACHE_VIPT */
2970 	pmap_release_page_lock(md);
2971 
2972 #ifndef ARM_MMU_EXTENDED
2973 	if (flush) {
2974 		/*
2975 		 * Note: We can't use pmap_tlb_flush{I,D}() here since that
2976 		 * would need a subsequent call to pmap_update() to ensure
2977 		 * curpm->pm_cstate.cs_all is reset. Our callers are not
2978 		 * required to do that (see pmap(9)), so we can't modify
2979 		 * the current pmap's state.
2980 		 */
2981 		if (PV_BEEN_EXECD(flags))
2982 			cpu_tlb_flushID();
2983 		else
2984 			cpu_tlb_flushD();
2985 	}
2986 	cpu_cpwait();
2987 #endif /* ARM_MMU_EXTENDED */
2988 }
2989 
2990 /*
2991  * pmap_t pmap_create(void)
2992  *
2993  *      Create a new pmap structure from scratch.
2994  */
2995 pmap_t
2996 pmap_create(void)
2997 {
2998 	pmap_t pm;
2999 
3000 	pm = pool_cache_get(&pmap_cache, PR_WAITOK);
3001 
3002 	mutex_init(&pm->pm_obj_lock, MUTEX_DEFAULT, IPL_NONE);
3003 	uvm_obj_init(&pm->pm_obj, NULL, false, 1);
3004 	uvm_obj_setlock(&pm->pm_obj, &pm->pm_obj_lock);
3005 
3006 	pm->pm_stats.wired_count = 0;
3007 	pm->pm_stats.resident_count = 1;
3008 #ifdef ARM_MMU_EXTENDED
3009 #ifdef MULTIPROCESSOR
3010 	kcpuset_create(&pm->pm_active, true);
3011 	kcpuset_create(&pm->pm_onproc, true);
3012 #endif
3013 #else
3014 	pm->pm_cstate.cs_all = 0;
3015 #endif
3016 	pmap_alloc_l1(pm);
3017 
3018 	/*
3019 	 * Note: The pool cache ensures that the pm_l2[] array is already
3020 	 * initialised to zero.
3021 	 */
3022 
3023 	pmap_pinit(pm);
3024 
3025 	return (pm);
3026 }
3027 
3028 u_int
3029 arm32_mmap_flags(paddr_t pa)
3030 {
3031 	/*
3032 	 * the upper 8 bits in pmap_enter()'s flags are reserved for MD stuff
3033 	 * and we're using the upper bits in page numbers to pass flags around
3034 	 * so we might as well use the same bits
3035 	 */
3036 	return (u_int)pa & PMAP_MD_MASK;
3037 }
3038 /*
3039  * int pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
3040  *      u_int flags)
3041  *
3042  *      Insert the given physical page (p) at
3043  *      the specified virtual address (v) in the
3044  *      target physical map with the protection requested.
3045  *
3046  *      NB:  This is the only routine which MAY NOT lazy-evaluate
3047  *      or lose information.  That is, this routine must actually
3048  *      insert this page into the given map NOW.
3049  */
3050 int
3051 pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags)
3052 {
3053 	struct l2_bucket *l2b;
3054 	struct vm_page *pg, *opg;
3055 	u_int nflags;
3056 	u_int oflags;
3057 	const bool kpm_p = (pm == pmap_kernel());
3058 #ifdef ARM_HAS_VBAR
3059 	const bool vector_page_p = false;
3060 #else
3061 	const bool vector_page_p = (va == vector_page);
3062 #endif
3063 
3064 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
3065 
3066 	UVMHIST_LOG(maphist, " (pm %p va %#x pa %#x prot %#x",
3067 	    pm, va, pa, prot);
3068 	UVMHIST_LOG(maphist, "  flag %#x", flags, 0, 0, 0);
3069 
3070 	KDASSERT((flags & PMAP_WIRED) == 0 || (flags & VM_PROT_ALL) != 0);
3071 	KDASSERT(((va | pa) & PGOFSET) == 0);
3072 
3073 	/*
3074 	 * Get a pointer to the page.  Later on in this function, we
3075 	 * test for a managed page by checking pg != NULL.
3076 	 */
3077 	pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
3078 
3079 	nflags = 0;
3080 	if (prot & VM_PROT_WRITE)
3081 		nflags |= PVF_WRITE;
3082 	if (prot & VM_PROT_EXECUTE)
3083 		nflags |= PVF_EXEC;
3084 	if (flags & PMAP_WIRED)
3085 		nflags |= PVF_WIRED;
3086 
3087 	pmap_acquire_pmap_lock(pm);
3088 
3089 	/*
3090 	 * Fetch the L2 bucket which maps this page, allocating one if
3091 	 * necessary for user pmaps.
3092 	 */
3093 	if (kpm_p) {
3094 		l2b = pmap_get_l2_bucket(pm, va);
3095 	} else {
3096 		l2b = pmap_alloc_l2_bucket(pm, va);
3097 	}
3098 	if (l2b == NULL) {
3099 		if (flags & PMAP_CANFAIL) {
3100 			pmap_release_pmap_lock(pm);
3101 			return (ENOMEM);
3102 		}
3103 		panic("pmap_enter: failed to allocate L2 bucket");
3104 	}
3105 	pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(va)];
3106 	const pt_entry_t opte = *ptep;
3107 	pt_entry_t npte = pa;
3108 	oflags = 0;
3109 
3110 	if (opte) {
3111 		/*
3112 		 * There is already a mapping at this address.
3113 		 * If the physical address is different, lookup the
3114 		 * vm_page.
3115 		 */
3116 		if (l2pte_pa(opte) != pa) {
3117 			KASSERT(!pmap_pv_tracked(pa));
3118 			opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3119 		} else
3120 			opg = pg;
3121 	} else
3122 		opg = NULL;
3123 
3124 	struct pmap_page *pp = pmap_pv_tracked(pa);
3125 
3126 	if (pg || pp) {
3127 		KASSERT((pg != NULL) != (pp != NULL));
3128 		struct vm_page_md *md = (pg != NULL) ? VM_PAGE_TO_MD(pg) :
3129 		    PMAP_PAGE_TO_MD(pp);
3130 
3131 		/*
3132 		 * This is to be a managed mapping.
3133 		 */
3134 		pmap_acquire_page_lock(md);
3135 		if ((flags & VM_PROT_ALL) || (md->pvh_attrs & PVF_REF)) {
3136 			/*
3137 			 * - The access type indicates that we don't need
3138 			 *   to do referenced emulation.
3139 			 * OR
3140 			 * - The physical page has already been referenced
3141 			 *   so no need to re-do referenced emulation here.
3142 			 */
3143 			npte |= l2pte_set_readonly(L2_S_PROTO);
3144 
3145 			nflags |= PVF_REF;
3146 
3147 			if ((prot & VM_PROT_WRITE) != 0 &&
3148 			    ((flags & VM_PROT_WRITE) != 0 ||
3149 			     (md->pvh_attrs & PVF_MOD) != 0)) {
3150 				/*
3151 				 * This is a writable mapping, and the
3152 				 * page's mod state indicates it has
3153 				 * already been modified. Make it
3154 				 * writable from the outset.
3155 				 */
3156 				npte = l2pte_set_writable(npte);
3157 				nflags |= PVF_MOD;
3158 			}
3159 
3160 #ifdef ARM_MMU_EXTENDED
3161 			/*
3162 			 * If the page has been cleaned, then the pvh_attrs
3163 			 * will have PVF_EXEC set, so mark it execute so we
3164 			 * don't get an access fault when trying to execute
3165 			 * from it.
3166 			 */
3167 			if (md->pvh_attrs & nflags & PVF_EXEC) {
3168 				npte &= ~L2_XS_XN;
3169 			}
3170 #endif
3171 		} else {
3172 			/*
3173 			 * Need to do page referenced emulation.
3174 			 */
3175 			npte |= L2_TYPE_INV;
3176 		}
3177 
3178 		if (flags & ARM32_MMAP_WRITECOMBINE) {
3179 			npte |= pte_l2_s_wc_mode;
3180 		} else
3181 			npte |= pte_l2_s_cache_mode;
3182 
3183 		if (pg != NULL && pg == opg) {
3184 			/*
3185 			 * We're changing the attrs of an existing mapping.
3186 			 */
3187 			oflags = pmap_modify_pv(md, pa, pm, va,
3188 			    PVF_WRITE | PVF_EXEC | PVF_WIRED |
3189 			    PVF_MOD | PVF_REF, nflags);
3190 
3191 #ifdef PMAP_CACHE_VIVT
3192 			/*
3193 			 * We may need to flush the cache if we're
3194 			 * doing rw-ro...
3195 			 */
3196 			if (pm->pm_cstate.cs_cache_d &&
3197 			    (oflags & PVF_NC) == 0 &&
3198 			    l2pte_writable_p(opte) &&
3199 			    (prot & VM_PROT_WRITE) == 0)
3200 				cpu_dcache_wb_range(va, PAGE_SIZE);
3201 #endif
3202 		} else {
3203 			struct pv_entry *pv;
3204 			/*
3205 			 * New mapping, or changing the backing page
3206 			 * of an existing mapping.
3207 			 */
3208 			if (opg) {
3209 				struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3210 				paddr_t opa = VM_PAGE_TO_PHYS(opg);
3211 
3212 				/*
3213 				 * Replacing an existing mapping with a new one.
3214 				 * It is part of our managed memory so we
3215 				 * must remove it from the PV list
3216 				 */
3217 				pv = pmap_remove_pv(omd, opa, pm, va);
3218 				pmap_vac_me_harder(omd, opa, pm, 0);
3219 				oflags = pv->pv_flags;
3220 
3221 #ifdef PMAP_CACHE_VIVT
3222 				/*
3223 				 * If the old mapping was valid (ref/mod
3224 				 * emulation creates 'invalid' mappings
3225 				 * initially) then make sure to frob
3226 				 * the cache.
3227 				 */
3228 				if (!(oflags & PVF_NC) && l2pte_valid_p(opte)) {
3229 					pmap_cache_wbinv_page(pm, va, true,
3230 					    oflags);
3231 				}
3232 #endif
3233 			} else {
3234 				pmap_release_page_lock(md);
3235 				pv = pool_get(&pmap_pv_pool, PR_NOWAIT);
3236 				if (pv == NULL) {
3237 					pmap_release_pmap_lock(pm);
3238 					if ((flags & PMAP_CANFAIL) == 0)
3239 						panic("pmap_enter: "
3240 						    "no pv entries");
3241 
3242 					pmap_free_l2_bucket(pm, l2b, 0);
3243 					UVMHIST_LOG(maphist, "  <-- done (ENOMEM)",
3244 					    0, 0, 0, 0);
3245 					return (ENOMEM);
3246 				}
3247 				pmap_acquire_page_lock(md);
3248 			}
3249 
3250 			pmap_enter_pv(md, pa, pv, pm, va, nflags);
3251 		}
3252 		pmap_release_page_lock(md);
3253 	} else {
3254 		/*
3255 		 * We're mapping an unmanaged page.
3256 		 * These are always readable, and possibly writable, from
3257 		 * the get go as we don't need to track ref/mod status.
3258 		 */
3259 		npte |= l2pte_set_readonly(L2_S_PROTO);
3260 		if (prot & VM_PROT_WRITE)
3261 			npte = l2pte_set_writable(npte);
3262 
3263 		/*
3264 		 * Make sure the vector table is mapped cacheable
3265 		 */
3266 		if ((vector_page_p && !kpm_p)
3267 		    || (flags & ARM32_MMAP_CACHEABLE)) {
3268 			npte |= pte_l2_s_cache_mode;
3269 #ifdef ARM_MMU_EXTENDED
3270 			npte &= ~L2_XS_XN;	/* and executable */
3271 #endif
3272 		} else if (flags & ARM32_MMAP_WRITECOMBINE) {
3273 			npte |= pte_l2_s_wc_mode;
3274 		}
3275 		if (opg) {
3276 			/*
3277 			 * Looks like there's an existing 'managed' mapping
3278 			 * at this address.
3279 			 */
3280 			struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3281 			paddr_t opa = VM_PAGE_TO_PHYS(opg);
3282 
3283 			pmap_acquire_page_lock(omd);
3284 			struct pv_entry *pv = pmap_remove_pv(omd, opa, pm, va);
3285 			pmap_vac_me_harder(omd, opa, pm, 0);
3286 			oflags = pv->pv_flags;
3287 			pmap_release_page_lock(omd);
3288 
3289 #ifdef PMAP_CACHE_VIVT
3290 			if (!(oflags & PVF_NC) && l2pte_valid_p(opte)) {
3291 				pmap_cache_wbinv_page(pm, va, true, oflags);
3292 			}
3293 #endif
3294 			pool_put(&pmap_pv_pool, pv);
3295 		}
3296 	}
3297 
3298 	/*
3299 	 * Make sure userland mappings get the right permissions
3300 	 */
3301 	if (!vector_page_p && !kpm_p) {
3302 		npte |= L2_S_PROT_U;
3303 #ifdef ARM_MMU_EXTENDED
3304 		npte |= L2_XS_nG;	/* user pages are not global */
3305 #endif
3306 	}
3307 
3308 	/*
3309 	 * Keep the stats up to date
3310 	 */
3311 	if (opte == 0) {
3312 		l2b->l2b_occupancy += PAGE_SIZE / L2_S_SIZE;
3313 		pm->pm_stats.resident_count++;
3314 	}
3315 
3316 	UVMHIST_LOG(maphist, " opte %#x npte %#x", opte, npte, 0, 0);
3317 
3318 #if defined(ARM_MMU_EXTENDED)
3319 	/*
3320 	 * If exec protection was requested but the page hasn't been synced,
3321 	 * sync it now and allow execution from it.
3322 	 */
3323 	if ((nflags & PVF_EXEC) && (npte & L2_XS_XN)) {
3324 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3325 		npte &= ~L2_XS_XN;
3326 		pmap_syncicache_page(md, pa);
3327 		PMAPCOUNT(exec_synced_map);
3328 	}
3329 #endif
3330 	/*
3331 	 * If this is just a wiring change, the two PTEs will be
3332 	 * identical, so there's no need to update the page table.
3333 	 */
3334 	if (npte != opte) {
3335 		l2pte_reset(ptep);
3336 		PTE_SYNC(ptep);
3337 		if (l2pte_valid_p(opte)) {
3338 			pmap_tlb_flush_SE(pm, va, oflags);
3339 		}
3340 		l2pte_set(ptep, npte, 0);
3341 		PTE_SYNC(ptep);
3342 #ifndef ARM_MMU_EXTENDED
3343 		bool is_cached = pmap_is_cached(pm);
3344 		if (is_cached) {
3345 			/*
3346 			 * We only need to frob the cache/tlb if this pmap
3347 			 * is current
3348 			 */
3349 			if (!vector_page_p && l2pte_valid_p(npte)) {
3350 				/*
3351 				 * This mapping is likely to be accessed as
3352 				 * soon as we return to userland. Fix up the
3353 				 * L1 entry to avoid taking another
3354 				 * page/domain fault.
3355 				 */
3356 				pd_entry_t *pdep = pmap_l1_kva(pm)
3357 				     + l1pte_index(va);
3358 				pd_entry_t pde = L1_C_PROTO | l2b->l2b_pa
3359 				    | L1_C_DOM(pmap_domain(pm));
3360 				if (*pdep != pde) {
3361 					l1pte_setone(pdep, pde);
3362 					PDE_SYNC(pdep);
3363 				}
3364 			}
3365 		}
3366 #endif /* !ARM_MMU_EXTENDED */
3367 
3368 #ifndef ARM_MMU_EXTENDED
3369 		UVMHIST_LOG(maphist, "  is_cached %d cs 0x%08x\n",
3370 		    is_cached, pm->pm_cstate.cs_all, 0, 0);
3371 
3372 		if (pg != NULL) {
3373 			struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3374 
3375 			pmap_acquire_page_lock(md);
3376 			pmap_vac_me_harder(md, pa, pm, va);
3377 			pmap_release_page_lock(md);
3378 		}
3379 #endif
3380 	}
3381 #if defined(PMAP_CACHE_VIPT) && defined(DIAGNOSTIC)
3382 	if (pg) {
3383 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3384 
3385 		pmap_acquire_page_lock(md);
3386 #ifndef ARM_MMU_EXTENDED
3387 		KASSERT((md->pvh_attrs & PVF_DMOD) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
3388 #endif
3389 		PMAP_VALIDATE_MD_PAGE(md);
3390 		pmap_release_page_lock(md);
3391 	}
3392 #endif
3393 
3394 	pmap_release_pmap_lock(pm);
3395 
3396 	return (0);
3397 }
3398 
3399 /*
3400  * pmap_remove()
3401  *
3402  * pmap_remove is responsible for nuking a number of mappings for a range
3403  * of virtual address space in the current pmap. To do this efficiently
3404  * is interesting, because in a number of cases a wide virtual address
3405  * range may be supplied that contains few actual mappings. So, the
3406  * optimisations are:
3407  *  1. Skip over hunks of address space for which no L1 or L2 entry exists.
3408  *  2. Build up a list of pages we've hit, up to a maximum, so we can
3409  *     maybe do just a partial cache clean. This path of execution is
3410  *     complicated by the fact that the cache must be flushed _before_
3411  *     the PTE is nuked, being a VAC :-)
3412  *  3. If we're called after UVM calls pmap_remove_all(), we can defer
3413  *     all invalidations until pmap_update(), since pmap_remove_all() has
3414  *     already flushed the cache.
3415  *  4. Maybe later fast-case a single page, but I don't think this is
3416  *     going to make _that_ much difference overall.
3417  */
3418 
3419 #define	PMAP_REMOVE_CLEAN_LIST_SIZE	3
3420 
3421 void
3422 pmap_remove(pmap_t pm, vaddr_t sva, vaddr_t eva)
3423 {
3424 	vaddr_t next_bucket;
3425 	u_int cleanlist_idx, total, cnt;
3426 	struct {
3427 		vaddr_t va;
3428 		pt_entry_t *ptep;
3429 	} cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
3430 	u_int mappings;
3431 
3432 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
3433 	UVMHIST_LOG(maphist, " (pm=%p, sva=%#x, eva=%#x)", pm, sva, eva, 0);
3434 
3435 	/*
3436 	 * we lock in the pmap => pv_head direction
3437 	 */
3438 	pmap_acquire_pmap_lock(pm);
3439 
3440 	if (pm->pm_remove_all || !pmap_is_cached(pm)) {
3441 		cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
3442 #ifndef ARM_MMU_EXTENDED
3443 		if (pm->pm_cstate.cs_tlb == 0)
3444 			pm->pm_remove_all = true;
3445 #endif
3446 	} else
3447 		cleanlist_idx = 0;
3448 
3449 	total = 0;
3450 
3451 	while (sva < eva) {
3452 		/*
3453 		 * Do one L2 bucket's worth at a time.
3454 		 */
3455 		next_bucket = L2_NEXT_BUCKET_VA(sva);
3456 		if (next_bucket > eva)
3457 			next_bucket = eva;
3458 
3459 		struct l2_bucket * const l2b = pmap_get_l2_bucket(pm, sva);
3460 		if (l2b == NULL) {
3461 			sva = next_bucket;
3462 			continue;
3463 		}
3464 
3465 		pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(sva)];
3466 
3467 		for (mappings = 0;
3468 		     sva < next_bucket;
3469 		     sva += PAGE_SIZE, ptep += PAGE_SIZE / L2_S_SIZE) {
3470 			pt_entry_t opte = *ptep;
3471 
3472 			if (opte == 0) {
3473 				/* Nothing here, move along */
3474 				continue;
3475 			}
3476 
3477 			u_int flags = PVF_REF;
3478 			paddr_t pa = l2pte_pa(opte);
3479 			struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
3480 
3481 			/*
3482 			 * Update flags. In a number of circumstances,
3483 			 * we could cluster a lot of these and do a
3484 			 * number of sequential pages in one go.
3485 			 */
3486 			if (pg != NULL) {
3487 				struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3488 				struct pv_entry *pv;
3489 
3490 				pmap_acquire_page_lock(md);
3491 				pv = pmap_remove_pv(md, pa, pm, sva);
3492 				pmap_vac_me_harder(md, pa, pm, 0);
3493 				pmap_release_page_lock(md);
3494 				if (pv != NULL) {
3495 					if (pm->pm_remove_all == false) {
3496 						flags = pv->pv_flags;
3497 					}
3498 					pool_put(&pmap_pv_pool, pv);
3499 				}
3500 			}
3501 			mappings += PAGE_SIZE / L2_S_SIZE;
3502 
3503 			if (!l2pte_valid_p(opte)) {
3504 				/*
3505 				 * Ref/Mod emulation is still active for this
3506 				 * mapping, therefore it is has not yet been
3507 				 * accessed. No need to frob the cache/tlb.
3508 				 */
3509 				l2pte_reset(ptep);
3510 				PTE_SYNC_CURRENT(pm, ptep);
3511 				continue;
3512 			}
3513 
3514 #ifdef ARM_MMU_EXTENDED
3515 			if (pm == pmap_kernel()) {
3516 				l2pte_reset(ptep);
3517 				PTE_SYNC(ptep);
3518  				pmap_tlb_flush_SE(pm, sva, flags);
3519 				continue;
3520 			}
3521 #endif
3522 			if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
3523 				/* Add to the clean list. */
3524 				cleanlist[cleanlist_idx].ptep = ptep;
3525 				cleanlist[cleanlist_idx].va =
3526 				    sva | (flags & PVF_EXEC);
3527 				cleanlist_idx++;
3528 			} else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
3529 				/* Nuke everything if needed. */
3530 #ifdef PMAP_CACHE_VIVT
3531 				pmap_cache_wbinv_all(pm, PVF_EXEC);
3532 #endif
3533 				/*
3534 				 * Roll back the previous PTE list,
3535 				 * and zero out the current PTE.
3536 				 */
3537 				for (cnt = 0;
3538 				     cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
3539 					l2pte_reset(cleanlist[cnt].ptep);
3540 					PTE_SYNC(cleanlist[cnt].ptep);
3541 				}
3542 				l2pte_reset(ptep);
3543 				PTE_SYNC(ptep);
3544 				cleanlist_idx++;
3545 				pm->pm_remove_all = true;
3546 			} else {
3547 				l2pte_reset(ptep);
3548 				PTE_SYNC(ptep);
3549 				if (pm->pm_remove_all == false) {
3550 					pmap_tlb_flush_SE(pm, sva, flags);
3551 				}
3552 			}
3553 		}
3554 
3555 		/*
3556 		 * Deal with any left overs
3557 		 */
3558 		if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
3559 			total += cleanlist_idx;
3560 			for (cnt = 0; cnt < cleanlist_idx; cnt++) {
3561 				l2pte_reset(cleanlist[cnt].ptep);
3562 				PTE_SYNC_CURRENT(pm, cleanlist[cnt].ptep);
3563 #ifdef ARM_MMU_EXTENDED
3564 				vaddr_t clva = cleanlist[cnt].va;
3565 				pmap_tlb_flush_SE(pm, clva, PVF_REF);
3566 #else
3567 				vaddr_t va = cleanlist[cnt].va;
3568 				if (pm->pm_cstate.cs_all != 0) {
3569 					vaddr_t clva = va & ~PAGE_MASK;
3570 					u_int flags = va & PVF_EXEC;
3571 #ifdef PMAP_CACHE_VIVT
3572 					pmap_cache_wbinv_page(pm, clva, true,
3573 					    PVF_REF | PVF_WRITE | flags);
3574 #endif
3575 					pmap_tlb_flush_SE(pm, clva,
3576 					    PVF_REF | flags);
3577 				}
3578 #endif /* ARM_MMU_EXTENDED */
3579 			}
3580 
3581 			/*
3582 			 * If it looks like we're removing a whole bunch
3583 			 * of mappings, it's faster to just write-back
3584 			 * the whole cache now and defer TLB flushes until
3585 			 * pmap_update() is called.
3586 			 */
3587 			if (total <= PMAP_REMOVE_CLEAN_LIST_SIZE)
3588 				cleanlist_idx = 0;
3589 			else {
3590 				cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
3591 #ifdef PMAP_CACHE_VIVT
3592 				pmap_cache_wbinv_all(pm, PVF_EXEC);
3593 #endif
3594 				pm->pm_remove_all = true;
3595 			}
3596 		}
3597 
3598 
3599 		pmap_free_l2_bucket(pm, l2b, mappings);
3600 		pm->pm_stats.resident_count -= mappings / (PAGE_SIZE/L2_S_SIZE);
3601 	}
3602 
3603 	pmap_release_pmap_lock(pm);
3604 }
3605 
3606 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3607 static struct pv_entry *
3608 pmap_kremove_pg(struct vm_page *pg, vaddr_t va)
3609 {
3610 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
3611 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
3612 	struct pv_entry *pv;
3613 
3614 	KASSERT(arm_cache_prefer_mask == 0 || md->pvh_attrs & (PVF_COLORED|PVF_NC));
3615 	KASSERT((md->pvh_attrs & PVF_KMPAGE) == 0);
3616 	KASSERT(pmap_page_locked_p(md));
3617 
3618 	pv = pmap_remove_pv(md, pa, pmap_kernel(), va);
3619 	KASSERTMSG(pv, "pg %p (pa #%lx) va %#lx", pg, pa, va);
3620 	KASSERT(PV_IS_KENTRY_P(pv->pv_flags));
3621 
3622 	/*
3623 	 * If we are removing a writeable mapping to a cached exec page,
3624 	 * if it's the last mapping then clear it execness other sync
3625 	 * the page to the icache.
3626 	 */
3627 	if ((md->pvh_attrs & (PVF_NC|PVF_EXEC)) == PVF_EXEC
3628 	    && (pv->pv_flags & PVF_WRITE) != 0) {
3629 		if (SLIST_EMPTY(&md->pvh_list)) {
3630 			md->pvh_attrs &= ~PVF_EXEC;
3631 			PMAPCOUNT(exec_discarded_kremove);
3632 		} else {
3633 			pmap_syncicache_page(md, pa);
3634 			PMAPCOUNT(exec_synced_kremove);
3635 		}
3636 	}
3637 	pmap_vac_me_harder(md, pa, pmap_kernel(), 0);
3638 
3639 	return pv;
3640 }
3641 #endif /* PMAP_CACHE_VIPT && !ARM_MMU_EXTENDED */
3642 
3643 /*
3644  * pmap_kenter_pa: enter an unmanaged, wired kernel mapping
3645  *
3646  * We assume there is already sufficient KVM space available
3647  * to do this, as we can't allocate L2 descriptor tables/metadata
3648  * from here.
3649  */
3650 void
3651 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot, u_int flags)
3652 {
3653 #ifdef PMAP_CACHE_VIVT
3654 	struct vm_page *pg = (flags & PMAP_KMPAGE) ? PHYS_TO_VM_PAGE(pa) : NULL;
3655 #endif
3656 #ifdef PMAP_CACHE_VIPT
3657 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
3658 	struct vm_page *opg;
3659 #ifndef ARM_MMU_EXTENDED
3660 	struct pv_entry *pv = NULL;
3661 #endif
3662 #endif
3663 	struct vm_page_md *md = pg != NULL ? VM_PAGE_TO_MD(pg) : NULL;
3664 
3665 	UVMHIST_FUNC(__func__);
3666 
3667 	if (pmap_initialized) {
3668 		UVMHIST_CALLED(maphist);
3669 		UVMHIST_LOG(maphist, " (va=%#x, pa=%#x, prot=%#x, flags=%#x",
3670 		    va, pa, prot, flags);
3671 	}
3672 
3673 	pmap_t kpm = pmap_kernel();
3674 	pmap_acquire_pmap_lock(kpm);
3675 	struct l2_bucket * const l2b = pmap_get_l2_bucket(kpm, va);
3676 	const size_t l1slot __diagused = l1pte_index(va);
3677 	KASSERTMSG(l2b != NULL,
3678 	    "va %#lx pa %#lx prot %d maxkvaddr %#lx: l2 %p l2b %p kva %p",
3679 	    va, pa, prot, pmap_curmaxkvaddr, kpm->pm_l2[L2_IDX(l1slot)],
3680 	    kpm->pm_l2[L2_IDX(l1slot)]
3681 		? &kpm->pm_l2[L2_IDX(l1slot)]->l2_bucket[L2_BUCKET(l1slot)]
3682 		: NULL,
3683 	    kpm->pm_l2[L2_IDX(l1slot)]
3684 		? kpm->pm_l2[L2_IDX(l1slot)]->l2_bucket[L2_BUCKET(l1slot)].l2b_kva
3685 		: NULL);
3686 	KASSERT(l2b->l2b_kva != NULL);
3687 
3688 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
3689 	const pt_entry_t opte = *ptep;
3690 
3691 	if (opte == 0) {
3692 		PMAPCOUNT(kenter_mappings);
3693 		l2b->l2b_occupancy += PAGE_SIZE / L2_S_SIZE;
3694 	} else {
3695 		PMAPCOUNT(kenter_remappings);
3696 #ifdef PMAP_CACHE_VIPT
3697 		opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3698 #if !defined(ARM_MMU_EXTENDED) || defined(DIAGNOSTIC)
3699 		struct vm_page_md *omd __diagused = VM_PAGE_TO_MD(opg);
3700 #endif
3701 		if (opg && arm_cache_prefer_mask != 0) {
3702 			KASSERT(opg != pg);
3703 			KASSERT((omd->pvh_attrs & PVF_KMPAGE) == 0);
3704 			KASSERT((flags & PMAP_KMPAGE) == 0);
3705 #ifndef ARM_MMU_EXTENDED
3706 			pmap_acquire_page_lock(omd);
3707 			pv = pmap_kremove_pg(opg, va);
3708 			pmap_release_page_lock(omd);
3709 #endif
3710 		}
3711 #endif
3712 		if (l2pte_valid_p(opte)) {
3713 			l2pte_reset(ptep);
3714 			PTE_SYNC(ptep);
3715 #ifdef PMAP_CACHE_VIVT
3716 			cpu_dcache_wbinv_range(va, PAGE_SIZE);
3717 #endif
3718 			cpu_tlb_flushD_SE(va);
3719 			cpu_cpwait();
3720 		}
3721 	}
3722 	pmap_release_pmap_lock(kpm);
3723 
3724 	pt_entry_t npte = L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot)
3725 	    | ((flags & PMAP_NOCACHE)
3726 		? 0
3727 		: ((flags & PMAP_PTE)
3728 		    ? pte_l2_s_cache_mode_pt : pte_l2_s_cache_mode));
3729 #ifdef ARM_MMU_EXTENDED
3730 	if (prot & VM_PROT_EXECUTE)
3731 		npte &= ~L2_XS_XN;
3732 #endif
3733 	l2pte_set(ptep, npte, 0);
3734 	PTE_SYNC(ptep);
3735 
3736 	if (pg) {
3737 		if (flags & PMAP_KMPAGE) {
3738 			KASSERT(md->urw_mappings == 0);
3739 			KASSERT(md->uro_mappings == 0);
3740 			KASSERT(md->krw_mappings == 0);
3741 			KASSERT(md->kro_mappings == 0);
3742 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3743 			KASSERT(pv == NULL);
3744 			KASSERT(arm_cache_prefer_mask == 0 || (va & PVF_COLORED) == 0);
3745 			KASSERT((md->pvh_attrs & PVF_NC) == 0);
3746 			/* if there is a color conflict, evict from cache. */
3747 			if (pmap_is_page_colored_p(md)
3748 			    && ((va ^ md->pvh_attrs) & arm_cache_prefer_mask)) {
3749 				PMAPCOUNT(vac_color_change);
3750 				pmap_flush_page(md, pa, PMAP_FLUSH_PRIMARY);
3751 			} else if (md->pvh_attrs & PVF_MULTCLR) {
3752 				/*
3753 				 * If this page has multiple colors, expunge
3754 				 * them.
3755 				 */
3756 				PMAPCOUNT(vac_flush_lots2);
3757 				pmap_flush_page(md, pa, PMAP_FLUSH_SECONDARY);
3758 			}
3759 			/*
3760 			 * Since this is a KMPAGE, there can be no contention
3761 			 * for this page so don't lock it.
3762 			 */
3763 			md->pvh_attrs &= PAGE_SIZE - 1;
3764 			md->pvh_attrs |= PVF_KMPAGE | PVF_COLORED | PVF_DIRTY
3765 			    | (va & arm_cache_prefer_mask);
3766 #else /* !PMAP_CACHE_VIPT || ARM_MMU_EXTENDED */
3767 			md->pvh_attrs |= PVF_KMPAGE;
3768 #endif
3769 			atomic_inc_32(&pmap_kmpages);
3770 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3771 		} else if (arm_cache_prefer_mask != 0) {
3772 			if (pv == NULL) {
3773 				pv = pool_get(&pmap_pv_pool, PR_NOWAIT);
3774 				KASSERT(pv != NULL);
3775 			}
3776 			pmap_acquire_page_lock(md);
3777 			pmap_enter_pv(md, pa, pv, pmap_kernel(), va,
3778 			    PVF_WIRED | PVF_KENTRY
3779 			    | (prot & VM_PROT_WRITE ? PVF_WRITE : 0));
3780 			if ((prot & VM_PROT_WRITE)
3781 			    && !(md->pvh_attrs & PVF_NC))
3782 				md->pvh_attrs |= PVF_DIRTY;
3783 			KASSERT((prot & VM_PROT_WRITE) == 0 || (md->pvh_attrs & (PVF_DIRTY|PVF_NC)));
3784 			pmap_vac_me_harder(md, pa, pmap_kernel(), va);
3785 			pmap_release_page_lock(md);
3786 #endif
3787 		}
3788 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3789 	} else {
3790 		if (pv != NULL)
3791 			pool_put(&pmap_pv_pool, pv);
3792 #endif
3793 	}
3794 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3795 	KASSERT(md == NULL || !pmap_page_locked_p(md));
3796 #endif
3797 	if (pmap_initialized) {
3798 		UVMHIST_LOG(maphist, "  <-- done (ptep %p: %#x -> %#x)",
3799 		    ptep, opte, npte, 0);
3800 	}
3801 
3802 }
3803 
3804 void
3805 pmap_kremove(vaddr_t va, vsize_t len)
3806 {
3807 #ifdef UVMHIST
3808 	u_int total_mappings = 0;
3809 #endif
3810 
3811 	PMAPCOUNT(kenter_unmappings);
3812 
3813 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
3814 
3815 	UVMHIST_LOG(maphist, " (va=%#x, len=%#x)", va, len, 0, 0);
3816 
3817 	const vaddr_t eva = va + len;
3818 
3819 	pmap_acquire_pmap_lock(pmap_kernel());
3820 
3821 	while (va < eva) {
3822 		vaddr_t next_bucket = L2_NEXT_BUCKET_VA(va);
3823 		if (next_bucket > eva)
3824 			next_bucket = eva;
3825 
3826 		pmap_t kpm = pmap_kernel();
3827 		struct l2_bucket * const l2b = pmap_get_l2_bucket(kpm, va);
3828 		KDASSERT(l2b != NULL);
3829 
3830 		pt_entry_t * const sptep = &l2b->l2b_kva[l2pte_index(va)];
3831 		pt_entry_t *ptep = sptep;
3832 		u_int mappings = 0;
3833 
3834 		while (va < next_bucket) {
3835 			const pt_entry_t opte = *ptep;
3836 			struct vm_page *opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
3837 			if (opg != NULL) {
3838 				struct vm_page_md *omd = VM_PAGE_TO_MD(opg);
3839 
3840 				if (omd->pvh_attrs & PVF_KMPAGE) {
3841 					KASSERT(omd->urw_mappings == 0);
3842 					KASSERT(omd->uro_mappings == 0);
3843 					KASSERT(omd->krw_mappings == 0);
3844 					KASSERT(omd->kro_mappings == 0);
3845 					omd->pvh_attrs &= ~PVF_KMPAGE;
3846 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3847 					if (arm_cache_prefer_mask != 0) {
3848 						omd->pvh_attrs &= ~PVF_WRITE;
3849 					}
3850 #endif
3851 					atomic_dec_32(&pmap_kmpages);
3852 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
3853 				} else if (arm_cache_prefer_mask != 0) {
3854 					pmap_acquire_page_lock(omd);
3855 					pool_put(&pmap_pv_pool,
3856 					    pmap_kremove_pg(opg, va));
3857 					pmap_release_page_lock(omd);
3858 #endif
3859 				}
3860 			}
3861 			if (l2pte_valid_p(opte)) {
3862 				l2pte_reset(ptep);
3863 				PTE_SYNC(ptep);
3864 #ifdef PMAP_CACHE_VIVT
3865 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
3866 #endif
3867 				cpu_tlb_flushD_SE(va);
3868 
3869 				mappings += PAGE_SIZE / L2_S_SIZE;
3870 			}
3871 			va += PAGE_SIZE;
3872 			ptep += PAGE_SIZE / L2_S_SIZE;
3873 		}
3874 		KDASSERTMSG(mappings <= l2b->l2b_occupancy, "%u %u",
3875 		    mappings, l2b->l2b_occupancy);
3876 		l2b->l2b_occupancy -= mappings;
3877 		//PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
3878 #ifdef UVMHIST
3879 		total_mappings += mappings;
3880 #endif
3881 	}
3882 	pmap_release_pmap_lock(pmap_kernel());
3883 	cpu_cpwait();
3884 	UVMHIST_LOG(maphist, "  <--- done (%u mappings removed)",
3885 	    total_mappings, 0, 0, 0);
3886 }
3887 
3888 bool
3889 pmap_extract(pmap_t pm, vaddr_t va, paddr_t *pap)
3890 {
3891 	struct l2_dtable *l2;
3892 	pd_entry_t *pdep, pde;
3893 	pt_entry_t *ptep, pte;
3894 	paddr_t pa;
3895 	u_int l1slot;
3896 
3897 	pmap_acquire_pmap_lock(pm);
3898 
3899 	l1slot = l1pte_index(va);
3900 	pdep = pmap_l1_kva(pm) + l1slot;
3901 	pde = *pdep;
3902 
3903 	if (l1pte_section_p(pde)) {
3904 		/*
3905 		 * These should only happen for pmap_kernel()
3906 		 */
3907 		KDASSERT(pm == pmap_kernel());
3908 		pmap_release_pmap_lock(pm);
3909 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
3910 		if (l1pte_supersection_p(pde)) {
3911 			pa = (pde & L1_SS_FRAME) | (va & L1_SS_OFFSET);
3912 		} else
3913 #endif
3914 			pa = (pde & L1_S_FRAME) | (va & L1_S_OFFSET);
3915 	} else {
3916 		/*
3917 		 * Note that we can't rely on the validity of the L1
3918 		 * descriptor as an indication that a mapping exists.
3919 		 * We have to look it up in the L2 dtable.
3920 		 */
3921 		l2 = pm->pm_l2[L2_IDX(l1slot)];
3922 
3923 		if (l2 == NULL ||
3924 		    (ptep = l2->l2_bucket[L2_BUCKET(l1slot)].l2b_kva) == NULL) {
3925 			pmap_release_pmap_lock(pm);
3926 			return false;
3927 		}
3928 
3929 		pte = ptep[l2pte_index(va)];
3930 		pmap_release_pmap_lock(pm);
3931 
3932 		if (pte == 0)
3933 			return false;
3934 
3935 		switch (pte & L2_TYPE_MASK) {
3936 		case L2_TYPE_L:
3937 			pa = (pte & L2_L_FRAME) | (va & L2_L_OFFSET);
3938 			break;
3939 
3940 		default:
3941 			pa = (pte & ~PAGE_MASK) | (va & PAGE_MASK);
3942 			break;
3943 		}
3944 	}
3945 
3946 	if (pap != NULL)
3947 		*pap = pa;
3948 
3949 	return true;
3950 }
3951 
3952 /*
3953  * pmap_pv_remove: remove an unmanaged pv-tracked page from all pmaps
3954  *	that map it
3955  */
3956 
3957 static void
3958 pmap_pv_remove(paddr_t pa)
3959 {
3960 	struct pmap_page *pp;
3961 
3962 	pp = pmap_pv_tracked(pa);
3963 	if (pp == NULL)
3964 		panic("pmap_pv_protect: page not pv-tracked: 0x%"PRIxPADDR,
3965 		    pa);
3966 
3967 	struct vm_page_md *md = PMAP_PAGE_TO_MD(pp);
3968 	pmap_page_remove(md, pa);
3969 }
3970 
3971 void
3972 pmap_pv_protect(paddr_t pa, vm_prot_t prot)
3973 {
3974 
3975 	/* the only case is remove at the moment */
3976 	KASSERT(prot == VM_PROT_NONE);
3977 	pmap_pv_remove(pa);
3978 }
3979 
3980 void
3981 pmap_protect(pmap_t pm, vaddr_t sva, vaddr_t eva, vm_prot_t prot)
3982 {
3983 	struct l2_bucket *l2b;
3984 	vaddr_t next_bucket;
3985 
3986 	NPDEBUG(PDB_PROTECT,
3987 	    printf("pmap_protect: pm %p sva 0x%lx eva 0x%lx prot 0x%x\n",
3988 	    pm, sva, eva, prot));
3989 
3990 	if ((prot & VM_PROT_READ) == 0) {
3991 		pmap_remove(pm, sva, eva);
3992 		return;
3993 	}
3994 
3995 	if (prot & VM_PROT_WRITE) {
3996 		/*
3997 		 * If this is a read->write transition, just ignore it and let
3998 		 * uvm_fault() take care of it later.
3999 		 */
4000 		return;
4001 	}
4002 
4003 	pmap_acquire_pmap_lock(pm);
4004 
4005 #ifndef ARM_MMU_EXTENDED
4006 	const bool flush = eva - sva >= PAGE_SIZE * 4;
4007 	u_int flags = 0;
4008 #endif
4009 	u_int clr_mask = PVF_WRITE | ((prot & VM_PROT_EXECUTE) ? 0 : PVF_EXEC);
4010 
4011 	while (sva < eva) {
4012 		next_bucket = L2_NEXT_BUCKET_VA(sva);
4013 		if (next_bucket > eva)
4014 			next_bucket = eva;
4015 
4016 		l2b = pmap_get_l2_bucket(pm, sva);
4017 		if (l2b == NULL) {
4018 			sva = next_bucket;
4019 			continue;
4020 		}
4021 
4022 		pt_entry_t *ptep = &l2b->l2b_kva[l2pte_index(sva)];
4023 
4024 		while (sva < next_bucket) {
4025 			const pt_entry_t opte = *ptep;
4026 			if (l2pte_valid_p(opte) && l2pte_writable_p(opte)) {
4027 				struct vm_page *pg;
4028 #ifndef ARM_MMU_EXTENDED
4029 				u_int f;
4030 #endif
4031 
4032 #ifdef PMAP_CACHE_VIVT
4033 				/*
4034 				 * OK, at this point, we know we're doing
4035 				 * write-protect operation.  If the pmap is
4036 				 * active, write-back the page.
4037 				 */
4038 				pmap_cache_wbinv_page(pm, sva, false,
4039 				    PVF_REF | PVF_WRITE);
4040 #endif
4041 
4042 				pg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
4043 				pt_entry_t npte = l2pte_set_readonly(opte);
4044 				l2pte_reset(ptep);
4045 				PTE_SYNC(ptep);
4046 #ifdef ARM_MMU_EXTENDED
4047 				pmap_tlb_flush_SE(pm, sva, PVF_REF);
4048 #endif
4049 				l2pte_set(ptep, npte, 0);
4050 				PTE_SYNC(ptep);
4051 
4052 				if (pg != NULL) {
4053 					struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4054 					paddr_t pa = VM_PAGE_TO_PHYS(pg);
4055 
4056 					pmap_acquire_page_lock(md);
4057 #ifndef ARM_MMU_EXTENDED
4058 					f =
4059 #endif
4060 					    pmap_modify_pv(md, pa, pm, sva,
4061 					       clr_mask, 0);
4062 					pmap_vac_me_harder(md, pa, pm, sva);
4063 					pmap_release_page_lock(md);
4064 #ifndef ARM_MMU_EXTENDED
4065 				} else {
4066 					f = PVF_REF | PVF_EXEC;
4067 				}
4068 
4069 				if (flush) {
4070 					flags |= f;
4071 				} else {
4072 					pmap_tlb_flush_SE(pm, sva, f);
4073 #endif
4074 				}
4075 			}
4076 
4077 			sva += PAGE_SIZE;
4078 			ptep += PAGE_SIZE / L2_S_SIZE;
4079 		}
4080 	}
4081 
4082 #ifndef ARM_MMU_EXTENDED
4083 	if (flush) {
4084 		if (PV_BEEN_EXECD(flags)) {
4085 			pmap_tlb_flushID(pm);
4086 		} else if (PV_BEEN_REFD(flags)) {
4087 			pmap_tlb_flushD(pm);
4088 		}
4089 	}
4090 #endif
4091 
4092 	pmap_release_pmap_lock(pm);
4093 }
4094 
4095 void
4096 pmap_icache_sync_range(pmap_t pm, vaddr_t sva, vaddr_t eva)
4097 {
4098 	struct l2_bucket *l2b;
4099 	pt_entry_t *ptep;
4100 	vaddr_t next_bucket;
4101 	vsize_t page_size = trunc_page(sva) + PAGE_SIZE - sva;
4102 
4103 	NPDEBUG(PDB_EXEC,
4104 	    printf("pmap_icache_sync_range: pm %p sva 0x%lx eva 0x%lx\n",
4105 	    pm, sva, eva));
4106 
4107 	pmap_acquire_pmap_lock(pm);
4108 
4109 	while (sva < eva) {
4110 		next_bucket = L2_NEXT_BUCKET_VA(sva);
4111 		if (next_bucket > eva)
4112 			next_bucket = eva;
4113 
4114 		l2b = pmap_get_l2_bucket(pm, sva);
4115 		if (l2b == NULL) {
4116 			sva = next_bucket;
4117 			continue;
4118 		}
4119 
4120 		for (ptep = &l2b->l2b_kva[l2pte_index(sva)];
4121 		     sva < next_bucket;
4122 		     sva += page_size,
4123 		     ptep += PAGE_SIZE / L2_S_SIZE,
4124 		     page_size = PAGE_SIZE) {
4125 			if (l2pte_valid_p(*ptep)) {
4126 				cpu_icache_sync_range(sva,
4127 				    min(page_size, eva - sva));
4128 			}
4129 		}
4130 	}
4131 
4132 	pmap_release_pmap_lock(pm);
4133 }
4134 
4135 void
4136 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
4137 {
4138 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4139 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
4140 
4141 	NPDEBUG(PDB_PROTECT,
4142 	    printf("pmap_page_protect: md %p (0x%08lx), prot 0x%x\n",
4143 	    md, pa, prot));
4144 
4145 	switch(prot) {
4146 	case VM_PROT_READ|VM_PROT_WRITE:
4147 #if defined(ARM_MMU_EXTENDED)
4148 		pmap_acquire_page_lock(md);
4149 		pmap_clearbit(md, pa, PVF_EXEC);
4150 		pmap_release_page_lock(md);
4151 		break;
4152 #endif
4153 	case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
4154 		break;
4155 
4156 	case VM_PROT_READ:
4157 #if defined(ARM_MMU_EXTENDED)
4158 		pmap_acquire_page_lock(md);
4159 		pmap_clearbit(md, pa, PVF_WRITE|PVF_EXEC);
4160 		pmap_release_page_lock(md);
4161 		break;
4162 #endif
4163 	case VM_PROT_READ|VM_PROT_EXECUTE:
4164 		pmap_acquire_page_lock(md);
4165 		pmap_clearbit(md, pa, PVF_WRITE);
4166 		pmap_release_page_lock(md);
4167 		break;
4168 
4169 	default:
4170 		pmap_page_remove(md, pa);
4171 		break;
4172 	}
4173 }
4174 
4175 /*
4176  * pmap_clear_modify:
4177  *
4178  *	Clear the "modified" attribute for a page.
4179  */
4180 bool
4181 pmap_clear_modify(struct vm_page *pg)
4182 {
4183 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4184 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
4185 	bool rv;
4186 
4187 	pmap_acquire_page_lock(md);
4188 
4189 	if (md->pvh_attrs & PVF_MOD) {
4190 		rv = true;
4191 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
4192 		/*
4193 		 * If we are going to clear the modified bit and there are
4194 		 * no other modified bits set, flush the page to memory and
4195 		 * mark it clean.
4196 		 */
4197 		if ((md->pvh_attrs & (PVF_DMOD|PVF_NC)) == PVF_MOD)
4198 			pmap_flush_page(md, pa, PMAP_CLEAN_PRIMARY);
4199 #endif
4200 		pmap_clearbit(md, pa, PVF_MOD);
4201 	} else {
4202 		rv = false;
4203 	}
4204 	pmap_release_page_lock(md);
4205 
4206 	return rv;
4207 }
4208 
4209 /*
4210  * pmap_clear_reference:
4211  *
4212  *	Clear the "referenced" attribute for a page.
4213  */
4214 bool
4215 pmap_clear_reference(struct vm_page *pg)
4216 {
4217 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4218 	paddr_t pa = VM_PAGE_TO_PHYS(pg);
4219 	bool rv;
4220 
4221 	pmap_acquire_page_lock(md);
4222 
4223 	if (md->pvh_attrs & PVF_REF) {
4224 		rv = true;
4225 		pmap_clearbit(md, pa, PVF_REF);
4226 	} else {
4227 		rv = false;
4228 	}
4229 	pmap_release_page_lock(md);
4230 
4231 	return rv;
4232 }
4233 
4234 /*
4235  * pmap_is_modified:
4236  *
4237  *	Test if a page has the "modified" attribute.
4238  */
4239 /* See <arm/arm32/pmap.h> */
4240 
4241 /*
4242  * pmap_is_referenced:
4243  *
4244  *	Test if a page has the "referenced" attribute.
4245  */
4246 /* See <arm/arm32/pmap.h> */
4247 
4248 #if defined(ARM_MMU_EXTENDED) && 0
4249 int
4250 pmap_prefetchabt_fixup(void *v)
4251 {
4252 	struct trapframe * const tf = v;
4253 	vaddr_t va = trunc_page(tf->tf_pc);
4254 	int rv = ABORT_FIXUP_FAILED;
4255 
4256 	if (!TRAP_USERMODE(tf) && va < VM_MAXUSER_ADDRESS)
4257 		return rv;
4258 
4259 	kpreempt_disable();
4260 	pmap_t pm = curcpu()->ci_pmap_cur;
4261 	const size_t l1slot = l1pte_index(va);
4262 	struct l2_dtable * const l2 = pm->pm_l2[L2_IDX(l1slot)];
4263 	if (l2 == NULL)
4264 		goto out;
4265 
4266 	struct l2_bucket * const l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
4267 	if (l2b->l2b_kva == NULL)
4268 		goto out;
4269 
4270 	/*
4271 	 * Check the PTE itself.
4272 	 */
4273 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
4274 	const pt_entry_t opte = *ptep;
4275 	if ((opte & L2_S_PROT_U) == 0 || (opte & L2_XS_XN) == 0)
4276 		goto out;
4277 
4278 	paddr_t pa = l2pte_pa(pte);
4279 	struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
4280 	KASSERT(pg != NULL);
4281 
4282 	struct vm_page_md * const md = VM_PAGE_TO_MD(pg);
4283 
4284 	pmap_acquire_page_lock(md);
4285 	struct pv_entry * const pv = pmap_find_pv(md, pm, va);
4286 	KASSERT(pv != NULL);
4287 
4288 	if (PV_IS_EXEC_P(pv->pv_flags)) {
4289 		l2pte_reset(ptep);
4290 		PTE_SYNC(ptep);
4291 		pmap_tlb_flush_SE(pm, va, PVF_EXEC | PVF_REF);
4292 		if (!PV_IS_EXEC_P(md->pvh_attrs)) {
4293 			pmap_syncicache_page(md, pa);
4294 		}
4295 		rv = ABORT_FIXUP_RETURN;
4296 		l2pte_set(ptep, opte & ~L2_XS_XN, 0);
4297 		PTE_SYNC(ptep);
4298 	}
4299 	pmap_release_page_lock(md);
4300 
4301   out:
4302 	kpreempt_enable();
4303 	return rv;
4304 }
4305 #endif
4306 
4307 int
4308 pmap_fault_fixup(pmap_t pm, vaddr_t va, vm_prot_t ftype, int user)
4309 {
4310 	struct l2_dtable *l2;
4311 	struct l2_bucket *l2b;
4312 	paddr_t pa;
4313 	const size_t l1slot = l1pte_index(va);
4314 	int rv = 0;
4315 
4316 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
4317 
4318 	va = trunc_page(va);
4319 
4320 	KASSERT(!user || (pm != pmap_kernel()));
4321 
4322 	UVMHIST_LOG(maphist, " (pm=%#x, va=%#x, ftype=%#x, user=%d)",
4323 	    pm, va, ftype, user);
4324 #ifdef ARM_MMU_EXTENDED
4325 	UVMHIST_LOG(maphist, " ti=%#x pai=%#x asid=%#x",
4326 	    cpu_tlb_info(curcpu()), PMAP_PAI(pm, cpu_tlb_info(curcpu())),
4327 	    PMAP_PAI(pm, cpu_tlb_info(curcpu()))->pai_asid, 0);
4328 #endif
4329 
4330 	pmap_acquire_pmap_lock(pm);
4331 
4332 	/*
4333 	 * If there is no l2_dtable for this address, then the process
4334 	 * has no business accessing it.
4335 	 *
4336 	 * Note: This will catch userland processes trying to access
4337 	 * kernel addresses.
4338 	 */
4339 	l2 = pm->pm_l2[L2_IDX(l1slot)];
4340 	if (l2 == NULL) {
4341 		UVMHIST_LOG(maphist, " no l2 for l1slot %#x", l1slot, 0, 0, 0);
4342 		goto out;
4343 	}
4344 
4345 	/*
4346 	 * Likewise if there is no L2 descriptor table
4347 	 */
4348 	l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
4349 	if (l2b->l2b_kva == NULL) {
4350 		UVMHIST_LOG(maphist, " <-- done (no ptep for l1slot %#x)", l1slot, 0, 0, 0);
4351 		goto out;
4352 	}
4353 
4354 	/*
4355 	 * Check the PTE itself.
4356 	 */
4357 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
4358 	pt_entry_t const opte = *ptep;
4359 	if (opte == 0 || (opte & L2_TYPE_MASK) == L2_TYPE_L) {
4360 		UVMHIST_LOG(maphist, " <-- done (empty pde for l1slot %#x)", l1slot, 0, 0, 0);
4361 		goto out;
4362 	}
4363 
4364 #ifndef ARM_HAS_VBAR
4365 	/*
4366 	 * Catch a userland access to the vector page mapped at 0x0
4367 	 */
4368 	if (user && (opte & L2_S_PROT_U) == 0) {
4369 		UVMHIST_LOG(maphist, " <-- done (vector_page)", 0, 0, 0, 0);
4370 		goto out;
4371 	}
4372 #endif
4373 
4374 	pa = l2pte_pa(opte);
4375 
4376 	if ((ftype & VM_PROT_WRITE) && !l2pte_writable_p(opte)) {
4377 		/*
4378 		 * This looks like a good candidate for "page modified"
4379 		 * emulation...
4380 		 */
4381 		struct pv_entry *pv;
4382 		struct vm_page *pg;
4383 
4384 		/* Extract the physical address of the page */
4385 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) {
4386 			UVMHIST_LOG(maphist, " <-- done (mod/ref unmanaged page)", 0, 0, 0, 0);
4387 			goto out;
4388 		}
4389 
4390 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4391 
4392 		/* Get the current flags for this page. */
4393 		pmap_acquire_page_lock(md);
4394 		pv = pmap_find_pv(md, pm, va);
4395 		if (pv == NULL || PV_IS_KENTRY_P(pv->pv_flags)) {
4396 			pmap_release_page_lock(md);
4397 			UVMHIST_LOG(maphist, " <-- done (mod/ref emul: no PV)", 0, 0, 0, 0);
4398 			goto out;
4399 		}
4400 
4401 		/*
4402 		 * Do the flags say this page is writable? If not then it
4403 		 * is a genuine write fault. If yes then the write fault is
4404 		 * our fault as we did not reflect the write access in the
4405 		 * PTE. Now we know a write has occurred we can correct this
4406 		 * and also set the modified bit
4407 		 */
4408 		if ((pv->pv_flags & PVF_WRITE) == 0) {
4409 			pmap_release_page_lock(md);
4410 			goto out;
4411 		}
4412 
4413 		md->pvh_attrs |= PVF_REF | PVF_MOD;
4414 		pv->pv_flags |= PVF_REF | PVF_MOD;
4415 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
4416 		/*
4417 		 * If there are cacheable mappings for this page, mark it dirty.
4418 		 */
4419 		if ((md->pvh_attrs & PVF_NC) == 0)
4420 			md->pvh_attrs |= PVF_DIRTY;
4421 #endif
4422 #ifdef ARM_MMU_EXTENDED
4423 		if (md->pvh_attrs & PVF_EXEC) {
4424 			md->pvh_attrs &= ~PVF_EXEC;
4425 			PMAPCOUNT(exec_discarded_modfixup);
4426 		}
4427 #endif
4428 		pmap_release_page_lock(md);
4429 
4430 		/*
4431 		 * Re-enable write permissions for the page.  No need to call
4432 		 * pmap_vac_me_harder(), since this is just a
4433 		 * modified-emulation fault, and the PVF_WRITE bit isn't
4434 		 * changing. We've already set the cacheable bits based on
4435 		 * the assumption that we can write to this page.
4436 		 */
4437 		const pt_entry_t npte =
4438 		    l2pte_set_writable((opte & ~L2_TYPE_MASK) | L2_S_PROTO)
4439 #ifdef ARM_MMU_EXTENDED
4440 		    | (pm != pmap_kernel() ? L2_XS_nG : 0)
4441 #endif
4442 		    | 0;
4443 		l2pte_reset(ptep);
4444 		PTE_SYNC(ptep);
4445 		pmap_tlb_flush_SE(pm, va,
4446 		    (ftype & VM_PROT_EXECUTE) ? PVF_EXEC | PVF_REF : PVF_REF);
4447 		l2pte_set(ptep, npte, 0);
4448 		PTE_SYNC(ptep);
4449 		PMAPCOUNT(fixup_mod);
4450 		rv = 1;
4451 		UVMHIST_LOG(maphist, " <-- done (mod/ref emul: changed pte from %#x to %#x)",
4452 		    opte, npte, 0, 0);
4453 	} else if ((opte & L2_TYPE_MASK) == L2_TYPE_INV) {
4454 		/*
4455 		 * This looks like a good candidate for "page referenced"
4456 		 * emulation.
4457 		 */
4458 		struct vm_page *pg;
4459 
4460 		/* Extract the physical address of the page */
4461 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) {
4462 			UVMHIST_LOG(maphist, " <-- done (ref emul: unmanaged page)", 0, 0, 0, 0);
4463 			goto out;
4464 		}
4465 
4466 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4467 
4468 		/* Get the current flags for this page. */
4469 		pmap_acquire_page_lock(md);
4470 		struct pv_entry *pv = pmap_find_pv(md, pm, va);
4471 		if (pv == NULL || PV_IS_KENTRY_P(pv->pv_flags)) {
4472 			pmap_release_page_lock(md);
4473 			UVMHIST_LOG(maphist, " <-- done (ref emul no PV)", 0, 0, 0, 0);
4474 			goto out;
4475 		}
4476 
4477 		md->pvh_attrs |= PVF_REF;
4478 		pv->pv_flags |= PVF_REF;
4479 
4480 		pt_entry_t npte =
4481 		    l2pte_set_readonly((opte & ~L2_TYPE_MASK) | L2_S_PROTO);
4482 #ifdef ARM_MMU_EXTENDED
4483 		if (pm != pmap_kernel()) {
4484 			npte |= L2_XS_nG;
4485 		}
4486 		/*
4487 		 * If we got called from prefetch abort, then ftype will have
4488 		 * VM_PROT_EXECUTE set.  Now see if we have no-execute set in
4489 		 * the PTE.
4490 		 */
4491 		if (user && (ftype & VM_PROT_EXECUTE) && (npte & L2_XS_XN)) {
4492 			/*
4493 			 * Is this a mapping of an executable page?
4494 			 */
4495 			if ((pv->pv_flags & PVF_EXEC) == 0) {
4496 				pmap_release_page_lock(md);
4497 				UVMHIST_LOG(maphist, " <-- done (ref emul: no exec)",
4498 				    0, 0, 0, 0);
4499 				goto out;
4500 			}
4501 			/*
4502 			 * If we haven't synced the page, do so now.
4503 			 */
4504 			if ((md->pvh_attrs & PVF_EXEC) == 0) {
4505 				UVMHIST_LOG(maphist, " ref emul: syncicache page #%#x",
4506 				    pa, 0, 0, 0);
4507 				pmap_syncicache_page(md, pa);
4508 				PMAPCOUNT(fixup_exec);
4509 			}
4510 			npte &= ~L2_XS_XN;
4511 		}
4512 #endif /* ARM_MMU_EXTENDED */
4513 		pmap_release_page_lock(md);
4514 		l2pte_reset(ptep);
4515 		PTE_SYNC(ptep);
4516 		pmap_tlb_flush_SE(pm, va,
4517 		    (ftype & VM_PROT_EXECUTE) ? PVF_EXEC | PVF_REF : PVF_REF);
4518 		l2pte_set(ptep, npte, 0);
4519 		PTE_SYNC(ptep);
4520 		PMAPCOUNT(fixup_ref);
4521 		rv = 1;
4522 		UVMHIST_LOG(maphist, " <-- done (ref emul: changed pte from %#x to %#x)",
4523 		    opte, npte, 0, 0);
4524 #ifdef ARM_MMU_EXTENDED
4525 	} else if (user && (ftype & VM_PROT_EXECUTE) && (opte & L2_XS_XN)) {
4526 		struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
4527 		if (pg == NULL) {
4528 			UVMHIST_LOG(maphist, " <-- done (unmanaged page)", 0, 0, 0, 0);
4529 			goto out;
4530 		}
4531 
4532 		struct vm_page_md * const md = VM_PAGE_TO_MD(pg);
4533 
4534 		/* Get the current flags for this page. */
4535 		pmap_acquire_page_lock(md);
4536 		struct pv_entry * const pv = pmap_find_pv(md, pm, va);
4537 		if (pv == NULL || (pv->pv_flags & PVF_EXEC) == 0) {
4538 			pmap_release_page_lock(md);
4539 			UVMHIST_LOG(maphist, " <-- done (no PV or not EXEC)", 0, 0, 0, 0);
4540 			goto out;
4541 		}
4542 
4543 		/*
4544 		 * If we haven't synced the page, do so now.
4545 		 */
4546 		if ((md->pvh_attrs & PVF_EXEC) == 0) {
4547 			UVMHIST_LOG(maphist, "syncicache page #%#x",
4548 			    pa, 0, 0, 0);
4549 			pmap_syncicache_page(md, pa);
4550 		}
4551 		pmap_release_page_lock(md);
4552 		/*
4553 		 * Turn off no-execute.
4554 		 */
4555 		KASSERT(opte & L2_XS_nG);
4556 		l2pte_reset(ptep);
4557 		PTE_SYNC(ptep);
4558 		pmap_tlb_flush_SE(pm, va, PVF_EXEC | PVF_REF);
4559 		l2pte_set(ptep, opte & ~L2_XS_XN, 0);
4560 		PTE_SYNC(ptep);
4561 		rv = 1;
4562 		PMAPCOUNT(fixup_exec);
4563 		UVMHIST_LOG(maphist, "exec: changed pte from %#x to %#x",
4564 		    opte, opte & ~L2_XS_XN, 0, 0);
4565 #endif
4566 	}
4567 
4568 #ifndef ARM_MMU_EXTENDED
4569 	/*
4570 	 * We know there is a valid mapping here, so simply
4571 	 * fix up the L1 if necessary.
4572 	 */
4573 	pd_entry_t * const pdep = pmap_l1_kva(pm) + l1slot;
4574 	pd_entry_t pde = L1_C_PROTO | l2b->l2b_pa | L1_C_DOM(pmap_domain(pm));
4575 	if (*pdep != pde) {
4576 		l1pte_setone(pdep, pde);
4577 		PDE_SYNC(pdep);
4578 		rv = 1;
4579 		PMAPCOUNT(fixup_pdes);
4580 	}
4581 #endif
4582 
4583 #ifdef CPU_SA110
4584 	/*
4585 	 * There are bugs in the rev K SA110.  This is a check for one
4586 	 * of them.
4587 	 */
4588 	if (rv == 0 && curcpu()->ci_arm_cputype == CPU_ID_SA110 &&
4589 	    curcpu()->ci_arm_cpurev < 3) {
4590 		/* Always current pmap */
4591 		if (l2pte_valid_p(opte)) {
4592 			extern int kernel_debug;
4593 			if (kernel_debug & 1) {
4594 				struct proc *p = curlwp->l_proc;
4595 				printf("prefetch_abort: page is already "
4596 				    "mapped - pte=%p *pte=%08x\n", ptep, opte);
4597 				printf("prefetch_abort: pc=%08lx proc=%p "
4598 				    "process=%s\n", va, p, p->p_comm);
4599 				printf("prefetch_abort: far=%08x fs=%x\n",
4600 				    cpu_faultaddress(), cpu_faultstatus());
4601 			}
4602 #ifdef DDB
4603 			if (kernel_debug & 2)
4604 				Debugger();
4605 #endif
4606 			rv = 1;
4607 		}
4608 	}
4609 #endif /* CPU_SA110 */
4610 
4611 #ifndef ARM_MMU_EXTENDED
4612 	/*
4613 	 * If 'rv == 0' at this point, it generally indicates that there is a
4614 	 * stale TLB entry for the faulting address.  That might be due to a
4615 	 * wrong setting of pmap_needs_pte_sync.  So set it and retry.
4616 	 */
4617 	if (rv == 0
4618 	    && pm->pm_l1->l1_domain_use_count == 1
4619 	    && pmap_needs_pte_sync == 0) {
4620 		pmap_needs_pte_sync = 1;
4621 		PTE_SYNC(ptep);
4622 		PMAPCOUNT(fixup_ptesync);
4623 		rv = 1;
4624 	}
4625 #endif
4626 
4627 #ifndef MULTIPROCESSOR
4628 #if defined(DEBUG) || 1
4629 	/*
4630 	 * If 'rv == 0' at this point, it generally indicates that there is a
4631 	 * stale TLB entry for the faulting address. This happens when two or
4632 	 * more processes are sharing an L1. Since we don't flush the TLB on
4633 	 * a context switch between such processes, we can take domain faults
4634 	 * for mappings which exist at the same VA in both processes. EVEN IF
4635 	 * WE'VE RECENTLY FIXED UP THE CORRESPONDING L1 in pmap_enter(), for
4636 	 * example.
4637 	 *
4638 	 * This is extremely likely to happen if pmap_enter() updated the L1
4639 	 * entry for a recently entered mapping. In this case, the TLB is
4640 	 * flushed for the new mapping, but there may still be TLB entries for
4641 	 * other mappings belonging to other processes in the 1MB range
4642 	 * covered by the L1 entry.
4643 	 *
4644 	 * Since 'rv == 0', we know that the L1 already contains the correct
4645 	 * value, so the fault must be due to a stale TLB entry.
4646 	 *
4647 	 * Since we always need to flush the TLB anyway in the case where we
4648 	 * fixed up the L1, or frobbed the L2 PTE, we effectively deal with
4649 	 * stale TLB entries dynamically.
4650 	 *
4651 	 * However, the above condition can ONLY happen if the current L1 is
4652 	 * being shared. If it happens when the L1 is unshared, it indicates
4653 	 * that other parts of the pmap are not doing their job WRT managing
4654 	 * the TLB.
4655 	 */
4656 	if (rv == 0
4657 #ifndef ARM_MMU_EXTENDED
4658 	    && pm->pm_l1->l1_domain_use_count == 1
4659 #endif
4660 	    && true) {
4661 #ifdef DEBUG
4662 		extern int last_fault_code;
4663 #else
4664 		int last_fault_code = ftype & VM_PROT_EXECUTE
4665 		    ? armreg_ifsr_read()
4666 		    : armreg_dfsr_read();
4667 #endif
4668 		printf("fixup: pm %p, va 0x%lx, ftype %d - nothing to do!\n",
4669 		    pm, va, ftype);
4670 		printf("fixup: l2 %p, l2b %p, ptep %p, pte %#x\n",
4671 		    l2, l2b, ptep, opte);
4672 
4673 #ifndef ARM_MMU_EXTENDED
4674 		printf("fixup: pdep %p, pde %#x, fsr %#x\n",
4675 		    pdep, pde, last_fault_code);
4676 #else
4677 		printf("fixup: pdep %p, pde %#x, ttbcr %#x\n",
4678 		    &pmap_l1_kva(pm)[l1slot], pmap_l1_kva(pm)[l1slot],
4679 		   armreg_ttbcr_read());
4680 		printf("fixup: fsr %#x cpm %p casid %#x contextidr %#x dacr %#x\n",
4681 		    last_fault_code, curcpu()->ci_pmap_cur,
4682 		    curcpu()->ci_pmap_asid_cur,
4683 		    armreg_contextidr_read(), armreg_dacr_read());
4684 #ifdef _ARM_ARCH_7
4685 		if (ftype & VM_PROT_WRITE)
4686 			armreg_ats1cuw_write(va);
4687 		else
4688 			armreg_ats1cur_write(va);
4689 		arm_isb();
4690 		printf("fixup: par %#x\n", armreg_par_read());
4691 #endif
4692 #endif
4693 #ifdef DDB
4694 		extern int kernel_debug;
4695 
4696 		if (kernel_debug & 2) {
4697 			pmap_release_pmap_lock(pm);
4698 #ifdef UVMHIST
4699 			KERNHIST_DUMP(maphist);
4700 #endif
4701 			cpu_Debugger();
4702 			pmap_acquire_pmap_lock(pm);
4703 		}
4704 #endif
4705 	}
4706 #endif
4707 #endif
4708 
4709 #ifndef ARM_MMU_EXTENDED
4710 	/* Flush the TLB in the shared L1 case - see comment above */
4711 	pmap_tlb_flush_SE(pm, va,
4712 	    (ftype & VM_PROT_EXECUTE) ? PVF_EXEC | PVF_REF : PVF_REF);
4713 #endif
4714 
4715 	rv = 1;
4716 
4717 out:
4718 	pmap_release_pmap_lock(pm);
4719 
4720 	return (rv);
4721 }
4722 
4723 /*
4724  * Routine:	pmap_procwr
4725  *
4726  * Function:
4727  *	Synchronize caches corresponding to [addr, addr+len) in p.
4728  *
4729  */
4730 void
4731 pmap_procwr(struct proc *p, vaddr_t va, int len)
4732 {
4733 	/* We only need to do anything if it is the current process. */
4734 	if (p == curproc)
4735 		cpu_icache_sync_range(va, len);
4736 }
4737 
4738 /*
4739  * Routine:	pmap_unwire
4740  * Function:	Clear the wired attribute for a map/virtual-address pair.
4741  *
4742  * In/out conditions:
4743  *		The mapping must already exist in the pmap.
4744  */
4745 void
4746 pmap_unwire(pmap_t pm, vaddr_t va)
4747 {
4748 	struct l2_bucket *l2b;
4749 	pt_entry_t *ptep, pte;
4750 	struct vm_page *pg;
4751 	paddr_t pa;
4752 
4753 	NPDEBUG(PDB_WIRING, printf("pmap_unwire: pm %p, va 0x%08lx\n", pm, va));
4754 
4755 	pmap_acquire_pmap_lock(pm);
4756 
4757 	l2b = pmap_get_l2_bucket(pm, va);
4758 	KDASSERT(l2b != NULL);
4759 
4760 	ptep = &l2b->l2b_kva[l2pte_index(va)];
4761 	pte = *ptep;
4762 
4763 	/* Extract the physical address of the page */
4764 	pa = l2pte_pa(pte);
4765 
4766 	if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
4767 		/* Update the wired bit in the pv entry for this page. */
4768 		struct vm_page_md *md = VM_PAGE_TO_MD(pg);
4769 
4770 		pmap_acquire_page_lock(md);
4771 		(void) pmap_modify_pv(md, pa, pm, va, PVF_WIRED, 0);
4772 		pmap_release_page_lock(md);
4773 	}
4774 
4775 	pmap_release_pmap_lock(pm);
4776 }
4777 
4778 void
4779 pmap_activate(struct lwp *l)
4780 {
4781 	struct cpu_info * const ci = curcpu();
4782 	extern int block_userspace_access;
4783 	pmap_t npm = l->l_proc->p_vmspace->vm_map.pmap;
4784 #ifdef ARM_MMU_EXTENDED
4785 	struct pmap_asid_info * const pai = PMAP_PAI(npm, cpu_tlb_info(ci));
4786 #endif
4787 
4788 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
4789 
4790 	UVMHIST_LOG(maphist, "(l=%#x) pm=%#x", l, npm, 0, 0);
4791 
4792 	/*
4793 	 * If activating a non-current lwp or the current lwp is
4794 	 * already active, just return.
4795 	 */
4796 	if (false
4797 	    || l != curlwp
4798 #ifdef ARM_MMU_EXTENDED
4799 	    || (ci->ci_pmap_cur == npm &&
4800 		(npm == pmap_kernel()
4801 		 /* || PMAP_PAI_ASIDVALID_P(pai, cpu_tlb_info(ci)) */))
4802 #else
4803 	    || npm->pm_activated == true
4804 #endif
4805 	    || false) {
4806 		UVMHIST_LOG(maphist, " <-- (same pmap)", curlwp, l, 0, 0);
4807 		return;
4808 	}
4809 
4810 #ifndef ARM_MMU_EXTENDED
4811 	const uint32_t ndacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2))
4812 	    | (DOMAIN_CLIENT << (pmap_domain(npm) * 2));
4813 
4814 	/*
4815 	 * If TTB and DACR are unchanged, short-circuit all the
4816 	 * TLB/cache management stuff.
4817 	 */
4818 	pmap_t opm = ci->ci_lastlwp
4819 	    ? ci->ci_lastlwp->l_proc->p_vmspace->vm_map.pmap
4820 	    : NULL;
4821 	if (opm != NULL) {
4822 		uint32_t odacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2))
4823 		    | (DOMAIN_CLIENT << (pmap_domain(opm) * 2));
4824 
4825 		if (opm->pm_l1 == npm->pm_l1 && odacr == ndacr)
4826 			goto all_done;
4827 	}
4828 #endif /* !ARM_MMU_EXTENDED */
4829 
4830 	PMAPCOUNT(activations);
4831 	block_userspace_access = 1;
4832 
4833 #ifndef ARM_MMU_EXTENDED
4834 	/*
4835 	 * If switching to a user vmspace which is different to the
4836 	 * most recent one, and the most recent one is potentially
4837 	 * live in the cache, we must write-back and invalidate the
4838 	 * entire cache.
4839 	 */
4840 	pmap_t rpm = ci->ci_pmap_lastuser;
4841 #endif
4842 
4843 /*
4844  * XXXSCW: There's a corner case here which can leave turds in the cache as
4845  * reported in kern/41058. They're probably left over during tear-down and
4846  * switching away from an exiting process. Until the root cause is identified
4847  * and fixed, zap the cache when switching pmaps. This will result in a few
4848  * unnecessary cache flushes, but that's better than silently corrupting data.
4849  */
4850 #ifndef ARM_MMU_EXTENDED
4851 #if 0
4852 	if (npm != pmap_kernel() && rpm && npm != rpm &&
4853 	    rpm->pm_cstate.cs_cache) {
4854 		rpm->pm_cstate.cs_cache = 0;
4855 #ifdef PMAP_CACHE_VIVT
4856 		cpu_idcache_wbinv_all();
4857 #endif
4858 	}
4859 #else
4860 	if (rpm) {
4861 		rpm->pm_cstate.cs_cache = 0;
4862 		if (npm == pmap_kernel())
4863 			ci->ci_pmap_lastuser = NULL;
4864 #ifdef PMAP_CACHE_VIVT
4865 		cpu_idcache_wbinv_all();
4866 #endif
4867 	}
4868 #endif
4869 
4870 	/* No interrupts while we frob the TTB/DACR */
4871 	uint32_t oldirqstate = disable_interrupts(IF32_bits);
4872 #endif /* !ARM_MMU_EXTENDED */
4873 
4874 #ifndef ARM_HAS_VBAR
4875 	/*
4876 	 * For ARM_VECTORS_LOW, we MUST, I repeat, MUST fix up the L1
4877 	 * entry corresponding to 'vector_page' in the incoming L1 table
4878 	 * before switching to it otherwise subsequent interrupts/exceptions
4879 	 * (including domain faults!) will jump into hyperspace.
4880 	 */
4881 	if (npm->pm_pl1vec != NULL) {
4882 		cpu_tlb_flushID_SE((u_int)vector_page);
4883 		cpu_cpwait();
4884 		*npm->pm_pl1vec = npm->pm_l1vec;
4885 		PTE_SYNC(npm->pm_pl1vec);
4886 	}
4887 #endif
4888 
4889 #ifdef ARM_MMU_EXTENDED
4890 	/*
4891 	 * Assume that TTBR1 has only global mappings and TTBR0 only has
4892 	 * non-global mappings.  To prevent speculation from doing evil things
4893 	 * we disable translation table walks using TTBR0 before setting the
4894 	 * CONTEXTIDR (ASID) or new TTBR0 value.  Once both are set, table
4895 	 * walks are reenabled.
4896 	 */
4897 	UVMHIST_LOG(maphist, " acquiring asid", 0, 0, 0, 0);
4898 	const uint32_t old_ttbcr = armreg_ttbcr_read();
4899 	armreg_ttbcr_write(old_ttbcr | TTBCR_S_PD0);
4900 	arm_isb();
4901 	pmap_tlb_asid_acquire(npm, l);
4902 	UVMHIST_LOG(maphist, " setting ttbr pa=%#x asid=%#x", npm->pm_l1_pa, pai->pai_asid, 0, 0);
4903 	cpu_setttb(npm->pm_l1_pa, pai->pai_asid);
4904 	/*
4905 	 * Now we can reenable tablewalks since the CONTEXTIDR and TTRB0 have
4906 	 * been updated.
4907 	 */
4908 	arm_isb();
4909 	if (npm != pmap_kernel()) {
4910 		armreg_ttbcr_write(old_ttbcr & ~TTBCR_S_PD0);
4911 	}
4912 	cpu_cpwait();
4913 	ci->ci_pmap_asid_cur = pai->pai_asid;
4914 #else
4915 	cpu_domains(ndacr);
4916 	if (npm == pmap_kernel() || npm == rpm) {
4917 		/*
4918 		 * Switching to a kernel thread, or back to the
4919 		 * same user vmspace as before... Simply update
4920 		 * the TTB (no TLB flush required)
4921 		 */
4922 		cpu_setttb(npm->pm_l1->l1_physaddr, false);
4923 		cpu_cpwait();
4924 	} else {
4925 		/*
4926 		 * Otherwise, update TTB and flush TLB
4927 		 */
4928 		cpu_context_switch(npm->pm_l1->l1_physaddr);
4929 		if (rpm != NULL)
4930 			rpm->pm_cstate.cs_tlb = 0;
4931 	}
4932 
4933 	restore_interrupts(oldirqstate);
4934 #endif /* ARM_MMU_EXTENDED */
4935 
4936 	block_userspace_access = 0;
4937 
4938 #ifndef ARM_MMU_EXTENDED
4939  all_done:
4940 	/*
4941 	 * The new pmap is resident. Make sure it's marked
4942 	 * as resident in the cache/TLB.
4943 	 */
4944 	npm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
4945 	if (npm != pmap_kernel())
4946 		ci->ci_pmap_lastuser = npm;
4947 
4948 	/* The old pmap is not longer active */
4949 	if (opm != npm) {
4950 		if (opm != NULL)
4951 			opm->pm_activated = false;
4952 
4953 		/* But the new one is */
4954 		npm->pm_activated = true;
4955 	}
4956 #endif
4957 	ci->ci_pmap_cur = npm;
4958 	UVMHIST_LOG(maphist, " <-- done", 0, 0, 0, 0);
4959 }
4960 
4961 void
4962 pmap_deactivate(struct lwp *l)
4963 {
4964 	pmap_t pm = l->l_proc->p_vmspace->vm_map.pmap;
4965 
4966 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
4967 
4968 	UVMHIST_LOG(maphist, "(l=%#x) pm=%#x", l, pm, 0, 0);
4969 
4970 #ifdef ARM_MMU_EXTENDED
4971 	kpreempt_disable();
4972 	struct cpu_info * const ci = curcpu();
4973 	/*
4974 	 * Disable translation table walks from TTBR0 while no pmap has been
4975 	 * activated.
4976 	 */
4977 	const uint32_t old_ttbcr = armreg_ttbcr_read();
4978 	armreg_ttbcr_write(old_ttbcr | TTBCR_S_PD0);
4979 	arm_isb();
4980 	pmap_tlb_asid_deactivate(pm);
4981 	cpu_setttb(pmap_kernel()->pm_l1_pa, KERNEL_PID);
4982 	ci->ci_pmap_cur = pmap_kernel();
4983 	ci->ci_pmap_asid_cur = KERNEL_PID;
4984 	kpreempt_enable();
4985 #else
4986 	/*
4987 	 * If the process is exiting, make sure pmap_activate() does
4988 	 * a full MMU context-switch and cache flush, which we might
4989 	 * otherwise skip. See PR port-arm/38950.
4990 	 */
4991 	if (l->l_proc->p_sflag & PS_WEXIT)
4992 		curcpu()->ci_lastlwp = NULL;
4993 
4994 	pm->pm_activated = false;
4995 #endif
4996 	UVMHIST_LOG(maphist, "  <-- done", 0, 0, 0, 0);
4997 }
4998 
4999 #ifdef ARM_MMU_EXTENDED
5000 static inline void
5001 pmap_remove_all_complete(pmap_t pm)
5002 {
5003 	KASSERT(pm != pmap_kernel());
5004 
5005 	KASSERTMSG(curcpu()->ci_pmap_cur != pm
5006 	    || pm->pm_pai[0].pai_asid == curcpu()->ci_pmap_asid_cur,
5007 	    "pmap/asid %p/%#x != %s cur pmap/asid %p/%#x", pm,
5008 	    pm->pm_pai[0].pai_asid, curcpu()->ci_data.cpu_name,
5009 	    curcpu()->ci_pmap_cur, curcpu()->ci_pmap_asid_cur);
5010 
5011 	/*
5012 	 * Finish up the pmap_remove_all() optimisation by flushing
5013 	 * all our ASIDs.
5014 	 */
5015 #ifdef MULTIPROCESSOR
5016 	// This should be the last CPU with this pmap onproc
5017 //	KASSERT(!kcpuset_isotherset(pm->pm_onproc, cpu_index(curcpu())));
5018 #if PMAP_TLB_MAX > 1
5019 	for (u_int i = 0; !kcpuset_iszero(pm->pm_active); i++) {
5020 		KASSERT(i < pmap_ntlbs);
5021 		struct pmap_tlb_info * const ti = pmap_tlbs[i];
5022 #else
5023 		struct pmap_tlb_info * const ti = &pmap_tlb0_info;
5024 #endif
5025 		struct cpu_info * const ci = curcpu();
5026 		TLBINFO_LOCK(ti);
5027 		struct pmap_asid_info * const pai = PMAP_PAI(pm, ti);
5028 		if (PMAP_PAI_ASIDVALID_P(pai, ti)) {
5029 			if (kcpuset_isset(pm->pm_onproc, cpu_index(ci))) {
5030 #if PMAP_TLB_MAX == 1
5031 				    KASSERT(cpu_tlb_info(ci) == ti);
5032 
5033 				    tlb_invalidate_asids(pai->pai_asid,
5034 					pai->pai_asid);
5035 #else
5036 				    if (cpu_tlb_info(ci) == ti) {
5037 					    tlb_invalidate_asids(pai->pai_asid,
5038 						pai->pai_asid);
5039 				    } else {
5040 					    pm->pm_shootdown_needed = 1;
5041 				    }
5042 #endif
5043 			}
5044 		}
5045 		TLBINFO_UNLOCK(ti);
5046 
5047 #if PMAP_TLB_MAX > 1
5048 	}
5049 #endif
5050 #else /* MULTIPROCESSOR */
5051 
5052 	struct pmap_asid_info * const pai =
5053 	    PMAP_PAI(pm, cpu_tlb_info(ci));
5054 
5055 	tlb_invalidate_asids(pai->pai_asid, pai->pai_asid);
5056 #endif /* MULTIPROCESSOR */
5057 }
5058 #endif
5059 
5060 void
5061 pmap_update(pmap_t pm)
5062 {
5063 
5064 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
5065 
5066 	UVMHIST_LOG(maphist, "pm=%#x remove_all %d", pm, pm->pm_remove_all, 0,
5067 	    0);
5068 
5069 	if (pm->pm_remove_all) {
5070 #ifdef ARM_MMU_EXTENDED
5071 		pmap_remove_all_complete(pm);
5072 #else
5073 		/*
5074 		 * Finish up the pmap_remove_all() optimisation by flushing
5075 		 * the TLB.
5076 		 */
5077 		pmap_tlb_flushID(pm);
5078 #endif
5079 		pm->pm_remove_all = false;
5080 	}
5081 
5082 #ifdef ARM_MMU_EXTENDED
5083 #if defined(MULTIPROCESSOR)
5084 	armreg_bpiallis_write(0);
5085 #else
5086 	armreg_bpiall_write(0);
5087 #endif
5088 
5089 #if defined(MULTIPROCESSOR) && PMAP_TLB_MAX > 1
5090 	u_int pending = atomic_swap_uint(&pmap->pm_shootdown_pending, 0);
5091 	if (pending && pmap_tlb_shootdown_bystanders(pmap)) {
5092 		PMAP_COUNT(shootdown_ipis);
5093 	}
5094 #endif
5095 	KASSERTMSG(pm == pmap_kernel()
5096 	    || curcpu()->ci_pmap_cur != pm
5097 	    || pm->pm_pai[0].pai_asid == curcpu()->ci_pmap_asid_cur,
5098 	    "pmap/asid %p/%#x != %s cur pmap/asid %p/%#x", pm,
5099 	    pm->pm_pai[0].pai_asid, curcpu()->ci_data.cpu_name,
5100 	    curcpu()->ci_pmap_cur, curcpu()->ci_pmap_asid_cur);
5101 #else
5102 	if (pmap_is_current(pm)) {
5103 		/*
5104 		 * If we're dealing with a current userland pmap, move its L1
5105 		 * to the end of the LRU.
5106 		 */
5107 		if (pm != pmap_kernel())
5108 			pmap_use_l1(pm);
5109 
5110 		/*
5111 		 * We can assume we're done with frobbing the cache/tlb for
5112 		 * now. Make sure any future pmap ops don't skip cache/tlb
5113 		 * flushes.
5114 		 */
5115 		pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
5116 	}
5117 #endif
5118 
5119 	PMAPCOUNT(updates);
5120 
5121 	/*
5122 	 * make sure TLB/cache operations have completed.
5123 	 */
5124 	cpu_cpwait();
5125 	UVMHIST_LOG(maphist, "  <-- done", 0, 0, 0, 0);
5126 }
5127 
5128 void
5129 pmap_remove_all(pmap_t pm)
5130 {
5131 
5132 	/*
5133 	 * The vmspace described by this pmap is about to be torn down.
5134 	 * Until pmap_update() is called, UVM will only make calls
5135 	 * to pmap_remove(). We can make life much simpler by flushing
5136 	 * the cache now, and deferring TLB invalidation to pmap_update().
5137 	 */
5138 #ifdef PMAP_CACHE_VIVT
5139 	pmap_cache_wbinv_all(pm, PVF_EXEC);
5140 #endif
5141 	pm->pm_remove_all = true;
5142 }
5143 
5144 /*
5145  * Retire the given physical map from service.
5146  * Should only be called if the map contains no valid mappings.
5147  */
5148 void
5149 pmap_destroy(pmap_t pm)
5150 {
5151 	UVMHIST_FUNC(__func__); UVMHIST_CALLED(maphist);
5152 
5153 	u_int count;
5154 
5155 	if (pm == NULL)
5156 		return;
5157 
5158 	UVMHIST_LOG(maphist, "pm=%#x remove_all %d", pm, pm->pm_remove_all, 0,
5159 	    0);
5160 
5161 	if (pm->pm_remove_all) {
5162 #ifdef ARM_MMU_EXTENDED
5163 		pmap_remove_all_complete(pm);
5164  		pmap_tlb_asid_release_all(pm);
5165 #else
5166 		pmap_tlb_flushID(pm);
5167 #endif
5168 		pm->pm_remove_all = false;
5169 	}
5170 
5171 	/*
5172 	 * Drop reference count
5173 	 */
5174 	mutex_enter(pm->pm_lock);
5175 	count = --pm->pm_obj.uo_refs;
5176 	mutex_exit(pm->pm_lock);
5177 	if (count > 0) {
5178 #ifndef ARM_MMU_EXTENDED
5179 		if (pmap_is_current(pm)) {
5180 			if (pm != pmap_kernel())
5181 				pmap_use_l1(pm);
5182 			pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
5183 		}
5184 #endif
5185 		return;
5186 	}
5187 
5188 	/*
5189 	 * reference count is zero, free pmap resources and then free pmap.
5190 	 */
5191 
5192 #ifndef ARM_HAS_VBAR
5193 	if (vector_page < KERNEL_BASE) {
5194 		KDASSERT(!pmap_is_current(pm));
5195 
5196 		/* Remove the vector page mapping */
5197 		pmap_remove(pm, vector_page, vector_page + PAGE_SIZE);
5198 		pmap_update(pm);
5199 	}
5200 #endif
5201 
5202 	pmap_free_l1(pm);
5203 
5204 #ifdef ARM_MMU_EXTENDED
5205 #ifdef MULTIPROCESSOR
5206 	kcpuset_destroy(pm->pm_active);
5207 	kcpuset_destroy(pm->pm_onproc);
5208 #endif
5209 #else
5210 	struct cpu_info * const ci = curcpu();
5211 	if (ci->ci_pmap_lastuser == pm)
5212 		ci->ci_pmap_lastuser = NULL;
5213 #endif
5214 
5215 	uvm_obj_destroy(&pm->pm_obj, false);
5216 	mutex_destroy(&pm->pm_obj_lock);
5217 	pool_cache_put(&pmap_cache, pm);
5218 	UVMHIST_LOG(maphist, "  <-- done", 0, 0, 0, 0);
5219 }
5220 
5221 
5222 /*
5223  * void pmap_reference(pmap_t pm)
5224  *
5225  * Add a reference to the specified pmap.
5226  */
5227 void
5228 pmap_reference(pmap_t pm)
5229 {
5230 
5231 	if (pm == NULL)
5232 		return;
5233 
5234 #ifndef ARM_MMU_EXTENDED
5235 	pmap_use_l1(pm);
5236 #endif
5237 
5238 	mutex_enter(pm->pm_lock);
5239 	pm->pm_obj.uo_refs++;
5240 	mutex_exit(pm->pm_lock);
5241 }
5242 
5243 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
5244 
5245 static struct evcnt pmap_prefer_nochange_ev =
5246     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap prefer", "nochange");
5247 static struct evcnt pmap_prefer_change_ev =
5248     EVCNT_INITIALIZER(EVCNT_TYPE_MISC, NULL, "pmap prefer", "change");
5249 
5250 EVCNT_ATTACH_STATIC(pmap_prefer_change_ev);
5251 EVCNT_ATTACH_STATIC(pmap_prefer_nochange_ev);
5252 
5253 void
5254 pmap_prefer(vaddr_t hint, vaddr_t *vap, int td)
5255 {
5256 	vsize_t mask = arm_cache_prefer_mask | (PAGE_SIZE - 1);
5257 	vaddr_t va = *vap;
5258 	vaddr_t diff = (hint - va) & mask;
5259 	if (diff == 0) {
5260 		pmap_prefer_nochange_ev.ev_count++;
5261 	} else {
5262 		pmap_prefer_change_ev.ev_count++;
5263 		if (__predict_false(td))
5264 			va -= mask + 1;
5265 		*vap = va + diff;
5266 	}
5267 }
5268 #endif /* ARM_MMU_V6 | ARM_MMU_V7 */
5269 
5270 /*
5271  * pmap_zero_page()
5272  *
5273  * Zero a given physical page by mapping it at a page hook point.
5274  * In doing the zero page op, the page we zero is mapped cachable, as with
5275  * StrongARM accesses to non-cached pages are non-burst making writing
5276  * _any_ bulk data very slow.
5277  */
5278 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
5279 void
5280 pmap_zero_page_generic(paddr_t pa)
5281 {
5282 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
5283 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
5284 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
5285 #endif
5286 #if defined(PMAP_CACHE_VIPT)
5287 	/* Choose the last page color it had, if any */
5288 	const vsize_t va_offset = md->pvh_attrs & arm_cache_prefer_mask;
5289 #else
5290 	const vsize_t va_offset = 0;
5291 #endif
5292 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
5293 	/*
5294 	 * Is this page mapped at its natural color?
5295 	 * If we have all of memory mapped, then just convert PA to VA.
5296 	 */
5297 	bool okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5298 	   || va_offset == (pa & arm_cache_prefer_mask);
5299 	const vaddr_t vdstp = okcolor
5300 	    ? pmap_direct_mapped_phys(pa, &okcolor, cpu_cdstp(va_offset))
5301 	    : cpu_cdstp(va_offset);
5302 #else
5303 	const bool okcolor = false;
5304 	const vaddr_t vdstp = cpu_cdstp(va_offset);
5305 #endif
5306 	pt_entry_t * const ptep = cpu_cdst_pte(va_offset);
5307 
5308 
5309 #ifdef DEBUG
5310 	if (!SLIST_EMPTY(&md->pvh_list))
5311 		panic("pmap_zero_page: page has mappings");
5312 #endif
5313 
5314 	KDASSERT((pa & PGOFSET) == 0);
5315 
5316 	if (!okcolor) {
5317 		/*
5318 		 * Hook in the page, zero it, and purge the cache for that
5319 		 * zeroed page. Invalidate the TLB as needed.
5320 		 */
5321 		const pt_entry_t npte = L2_S_PROTO | pa | pte_l2_s_cache_mode
5322 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE);
5323 		l2pte_set(ptep, npte, 0);
5324 		PTE_SYNC(ptep);
5325 		cpu_tlb_flushD_SE(vdstp);
5326 		cpu_cpwait();
5327 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS) && defined(PMAP_CACHE_VIPT) \
5328     && !defined(ARM_MMU_EXTENDED)
5329 		/*
5330 		 * If we are direct-mapped and our color isn't ok, then before
5331 		 * we bzero the page invalidate its contents from the cache and
5332 		 * reset the color to its natural color.
5333 		 */
5334 		cpu_dcache_inv_range(vdstp, PAGE_SIZE);
5335 		md->pvh_attrs &= ~arm_cache_prefer_mask;
5336 		md->pvh_attrs |= (pa & arm_cache_prefer_mask);
5337 #endif
5338 	}
5339 	bzero_page(vdstp);
5340 	if (!okcolor) {
5341 		/*
5342 		 * Unmap the page.
5343 		 */
5344 		l2pte_reset(ptep);
5345 		PTE_SYNC(ptep);
5346 		cpu_tlb_flushD_SE(vdstp);
5347 #ifdef PMAP_CACHE_VIVT
5348 		cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
5349 #endif
5350 	}
5351 #ifdef PMAP_CACHE_VIPT
5352 	/*
5353 	 * This page is now cache resident so it now has a page color.
5354 	 * Any contents have been obliterated so clear the EXEC flag.
5355 	 */
5356 #ifndef ARM_MMU_EXTENDED
5357 	if (!pmap_is_page_colored_p(md)) {
5358 		PMAPCOUNT(vac_color_new);
5359 		md->pvh_attrs |= PVF_COLORED;
5360 	}
5361 	md->pvh_attrs |= PVF_DIRTY;
5362 #endif
5363 	if (PV_IS_EXEC_P(md->pvh_attrs)) {
5364 		md->pvh_attrs &= ~PVF_EXEC;
5365 		PMAPCOUNT(exec_discarded_zero);
5366 	}
5367 #endif
5368 }
5369 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
5370 
5371 #if ARM_MMU_XSCALE == 1
5372 void
5373 pmap_zero_page_xscale(paddr_t pa)
5374 {
5375 #ifdef DEBUG
5376 	struct vm_page *pg = PHYS_TO_VM_PAGE(pa);
5377 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
5378 
5379 	if (!SLIST_EMPTY(&md->pvh_list))
5380 		panic("pmap_zero_page: page has mappings");
5381 #endif
5382 
5383 	KDASSERT((pa & PGOFSET) == 0);
5384 
5385 	/*
5386 	 * Hook in the page, zero it, and purge the cache for that
5387 	 * zeroed page. Invalidate the TLB as needed.
5388 	 */
5389 
5390 	pt_entry_t npte = L2_S_PROTO | pa |
5391 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
5392 	    L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
5393 	l2pte_set(cdst_pte, npte, 0);
5394 	PTE_SYNC(cdst_pte);
5395 	cpu_tlb_flushD_SE(cdstp);
5396 	cpu_cpwait();
5397 	bzero_page(cdstp);
5398 	xscale_cache_clean_minidata();
5399 	l2pte_reset(cdst_pte);
5400 	PTE_SYNC(cdst_pte);
5401 }
5402 #endif /* ARM_MMU_XSCALE == 1 */
5403 
5404 /* pmap_pageidlezero()
5405  *
5406  * The same as above, except that we assume that the page is not
5407  * mapped.  This means we never have to flush the cache first.  Called
5408  * from the idle loop.
5409  */
5410 bool
5411 pmap_pageidlezero(paddr_t pa)
5412 {
5413 	bool rv = true;
5414 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
5415 	struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
5416 	struct vm_page_md *md = VM_PAGE_TO_MD(pg);
5417 #endif
5418 #ifdef PMAP_CACHE_VIPT
5419 	/* Choose the last page color it had, if any */
5420 	const vsize_t va_offset = md->pvh_attrs & arm_cache_prefer_mask;
5421 #else
5422 	const vsize_t va_offset = 0;
5423 #endif
5424 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
5425 	bool okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5426 	   || va_offset == (pa & arm_cache_prefer_mask);
5427 	const vaddr_t vdstp = okcolor
5428 	    ? pmap_direct_mapped_phys(pa, &okcolor, cpu_cdstp(va_offset))
5429 	    : cpu_cdstp(va_offset);
5430 #else
5431 	const bool okcolor = false;
5432 	const vaddr_t vdstp = cpu_cdstp(va_offset);
5433 #endif
5434 	pt_entry_t * const ptep = cpu_cdst_pte(va_offset);
5435 
5436 
5437 #ifdef DEBUG
5438 	if (!SLIST_EMPTY(&md->pvh_list))
5439 		panic("pmap_pageidlezero: page has mappings");
5440 #endif
5441 
5442 	KDASSERT((pa & PGOFSET) == 0);
5443 
5444 	if (!okcolor) {
5445 		/*
5446 		 * Hook in the page, zero it, and purge the cache for that
5447 		 * zeroed page. Invalidate the TLB as needed.
5448 		 */
5449 		const pt_entry_t npte = L2_S_PROTO | pa |
5450 		    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
5451 		l2pte_set(ptep, npte, 0);
5452 		PTE_SYNC(ptep);
5453 		cpu_tlb_flushD_SE(vdstp);
5454 		cpu_cpwait();
5455 	}
5456 
5457 	uint64_t *ptr = (uint64_t *)vdstp;
5458 	for (size_t i = 0; i < PAGE_SIZE / sizeof(*ptr); i++) {
5459 		if (sched_curcpu_runnable_p() != 0) {
5460 			/*
5461 			 * A process has become ready.  Abort now,
5462 			 * so we don't keep it waiting while we
5463 			 * do slow memory access to finish this
5464 			 * page.
5465 			 */
5466 			rv = false;
5467 			break;
5468 		}
5469 		*ptr++ = 0;
5470 	}
5471 
5472 #ifdef PMAP_CACHE_VIVT
5473 	if (rv)
5474 		/*
5475 		 * if we aborted we'll rezero this page again later so don't
5476 		 * purge it unless we finished it
5477 		 */
5478 		cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
5479 #elif defined(PMAP_CACHE_VIPT)
5480 	/*
5481 	 * This page is now cache resident so it now has a page color.
5482 	 * Any contents have been obliterated so clear the EXEC flag.
5483 	 */
5484 #ifndef ARM_MMU_EXTENDED
5485 	if (!pmap_is_page_colored_p(md)) {
5486 		PMAPCOUNT(vac_color_new);
5487 		md->pvh_attrs |= PVF_COLORED;
5488 	}
5489 #endif
5490 	if (PV_IS_EXEC_P(md->pvh_attrs)) {
5491 		md->pvh_attrs &= ~PVF_EXEC;
5492 		PMAPCOUNT(exec_discarded_zero);
5493 	}
5494 #endif
5495 	/*
5496 	 * Unmap the page.
5497 	 */
5498 	if (!okcolor) {
5499 		l2pte_reset(ptep);
5500 		PTE_SYNC(ptep);
5501 		cpu_tlb_flushD_SE(vdstp);
5502 	}
5503 
5504 	return rv;
5505 }
5506 
5507 /*
5508  * pmap_copy_page()
5509  *
5510  * Copy one physical page into another, by mapping the pages into
5511  * hook points. The same comment regarding cachability as in
5512  * pmap_zero_page also applies here.
5513  */
5514 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
5515 void
5516 pmap_copy_page_generic(paddr_t src, paddr_t dst)
5517 {
5518 	struct vm_page * const src_pg = PHYS_TO_VM_PAGE(src);
5519 	struct vm_page_md *src_md = VM_PAGE_TO_MD(src_pg);
5520 #if defined(PMAP_CACHE_VIPT) || defined(DEBUG)
5521 	struct vm_page * const dst_pg = PHYS_TO_VM_PAGE(dst);
5522 	struct vm_page_md *dst_md = VM_PAGE_TO_MD(dst_pg);
5523 #endif
5524 #ifdef PMAP_CACHE_VIPT
5525 	const vsize_t src_va_offset = src_md->pvh_attrs & arm_cache_prefer_mask;
5526 	const vsize_t dst_va_offset = dst_md->pvh_attrs & arm_cache_prefer_mask;
5527 #else
5528 	const vsize_t src_va_offset = 0;
5529 	const vsize_t dst_va_offset = 0;
5530 #endif
5531 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS)
5532 	/*
5533 	 * Is this page mapped at its natural color?
5534 	 * If we have all of memory mapped, then just convert PA to VA.
5535 	 */
5536 	bool src_okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5537 	    || src_va_offset == (src & arm_cache_prefer_mask);
5538 	bool dst_okcolor = arm_pcache.dcache_type == CACHE_TYPE_PIPT
5539 	    || dst_va_offset == (dst & arm_cache_prefer_mask);
5540 	const vaddr_t vsrcp = src_okcolor
5541 	    ? pmap_direct_mapped_phys(src, &src_okcolor,
5542 		cpu_csrcp(src_va_offset))
5543 	    : cpu_csrcp(src_va_offset);
5544 	const vaddr_t vdstp = pmap_direct_mapped_phys(dst, &dst_okcolor,
5545 	    cpu_cdstp(dst_va_offset));
5546 #else
5547 	const bool src_okcolor = false;
5548 	const bool dst_okcolor = false;
5549 	const vaddr_t vsrcp = cpu_csrcp(src_va_offset);
5550 	const vaddr_t vdstp = cpu_cdstp(dst_va_offset);
5551 #endif
5552 	pt_entry_t * const src_ptep = cpu_csrc_pte(src_va_offset);
5553 	pt_entry_t * const dst_ptep = cpu_cdst_pte(dst_va_offset);
5554 
5555 #ifdef DEBUG
5556 	if (!SLIST_EMPTY(&dst_md->pvh_list))
5557 		panic("pmap_copy_page: dst page has mappings");
5558 #endif
5559 
5560 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
5561 	KASSERT(arm_cache_prefer_mask == 0 || src_md->pvh_attrs & (PVF_COLORED|PVF_NC));
5562 #endif
5563 	KDASSERT((src & PGOFSET) == 0);
5564 	KDASSERT((dst & PGOFSET) == 0);
5565 
5566 	/*
5567 	 * Clean the source page.  Hold the source page's lock for
5568 	 * the duration of the copy so that no other mappings can
5569 	 * be created while we have a potentially aliased mapping.
5570 	 */
5571 #ifdef PMAP_CACHE_VIVT
5572 	pmap_acquire_page_lock(src_md);
5573 	(void) pmap_clean_page(src_md, true);
5574 	pmap_release_page_lock(src_md);
5575 #endif
5576 
5577 	/*
5578 	 * Map the pages into the page hook points, copy them, and purge
5579 	 * the cache for the appropriate page. Invalidate the TLB
5580 	 * as required.
5581 	 */
5582 	if (!src_okcolor) {
5583 		const pt_entry_t nsrc_pte = L2_S_PROTO
5584 		    | src
5585 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
5586 		    | ((src_md->pvh_attrs & PVF_NC) ? 0 : pte_l2_s_cache_mode)
5587 #else // defined(PMAP_CACHE_VIVT) || defined(ARM_MMU_EXTENDED)
5588 		    | pte_l2_s_cache_mode
5589 #endif
5590 		    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ);
5591 		l2pte_set(src_ptep, nsrc_pte, 0);
5592 		PTE_SYNC(src_ptep);
5593 		cpu_tlb_flushD_SE(vsrcp);
5594 		cpu_cpwait();
5595 	}
5596 	if (!dst_okcolor) {
5597 		const pt_entry_t ndst_pte = L2_S_PROTO | dst |
5598 		    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
5599 		l2pte_set(dst_ptep, ndst_pte, 0);
5600 		PTE_SYNC(dst_ptep);
5601 		cpu_tlb_flushD_SE(vdstp);
5602 		cpu_cpwait();
5603 #if defined(__HAVE_MM_MD_DIRECT_MAPPED_PHYS) && defined(PMAP_CACHE_VIPT)
5604 		/*
5605 		 * If we are direct-mapped and our color isn't ok, then before
5606 		 * we bcopy to the new page invalidate its contents from the
5607 		 * cache and reset its color to its natural color.
5608 		 */
5609 		cpu_dcache_inv_range(vdstp, PAGE_SIZE);
5610 		dst_md->pvh_attrs &= ~arm_cache_prefer_mask;
5611 		dst_md->pvh_attrs |= (dst & arm_cache_prefer_mask);
5612 #endif
5613 	}
5614 	bcopy_page(vsrcp, vdstp);
5615 #ifdef PMAP_CACHE_VIVT
5616 	cpu_dcache_inv_range(vsrcp, PAGE_SIZE);
5617 	cpu_dcache_wbinv_range(vdstp, PAGE_SIZE);
5618 #endif
5619 	/*
5620 	 * Unmap the pages.
5621 	 */
5622 	if (!src_okcolor) {
5623 		l2pte_reset(src_ptep);
5624 		PTE_SYNC(src_ptep);
5625 		cpu_tlb_flushD_SE(vsrcp);
5626 		cpu_cpwait();
5627 	}
5628 	if (!dst_okcolor) {
5629 		l2pte_reset(dst_ptep);
5630 		PTE_SYNC(dst_ptep);
5631 		cpu_tlb_flushD_SE(vdstp);
5632 		cpu_cpwait();
5633 	}
5634 #ifdef PMAP_CACHE_VIPT
5635 	/*
5636 	 * Now that the destination page is in the cache, mark it as colored.
5637 	 * If this was an exec page, discard it.
5638 	 */
5639 	pmap_acquire_page_lock(dst_md);
5640 #ifndef ARM_MMU_EXTENDED
5641 	if (arm_pcache.cache_type == CACHE_TYPE_PIPT) {
5642 		dst_md->pvh_attrs &= ~arm_cache_prefer_mask;
5643 		dst_md->pvh_attrs |= (dst & arm_cache_prefer_mask);
5644 	}
5645 	if (!pmap_is_page_colored_p(dst_md)) {
5646 		PMAPCOUNT(vac_color_new);
5647 		dst_md->pvh_attrs |= PVF_COLORED;
5648 	}
5649 	dst_md->pvh_attrs |= PVF_DIRTY;
5650 #endif
5651 	if (PV_IS_EXEC_P(dst_md->pvh_attrs)) {
5652 		dst_md->pvh_attrs &= ~PVF_EXEC;
5653 		PMAPCOUNT(exec_discarded_copy);
5654 	}
5655 	pmap_release_page_lock(dst_md);
5656 #endif
5657 }
5658 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
5659 
5660 #if ARM_MMU_XSCALE == 1
5661 void
5662 pmap_copy_page_xscale(paddr_t src, paddr_t dst)
5663 {
5664 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
5665 	struct vm_page_md *src_md = VM_PAGE_TO_MD(src_pg);
5666 #ifdef DEBUG
5667 	struct vm_page_md *dst_md = VM_PAGE_TO_MD(PHYS_TO_VM_PAGE(dst));
5668 
5669 	if (!SLIST_EMPTY(&dst_md->pvh_list))
5670 		panic("pmap_copy_page: dst page has mappings");
5671 #endif
5672 
5673 	KDASSERT((src & PGOFSET) == 0);
5674 	KDASSERT((dst & PGOFSET) == 0);
5675 
5676 	/*
5677 	 * Clean the source page.  Hold the source page's lock for
5678 	 * the duration of the copy so that no other mappings can
5679 	 * be created while we have a potentially aliased mapping.
5680 	 */
5681 #ifdef PMAP_CACHE_VIVT
5682 	pmap_acquire_page_lock(src_md);
5683 	(void) pmap_clean_page(src_md, true);
5684 	pmap_release_page_lock(src_md);
5685 #endif
5686 
5687 	/*
5688 	 * Map the pages into the page hook points, copy them, and purge
5689 	 * the cache for the appropriate page. Invalidate the TLB
5690 	 * as required.
5691 	 */
5692 	const pt_entry_t nsrc_pte = L2_S_PROTO | src
5693 	    | L2_S_PROT(PTE_KERNEL, VM_PROT_READ)
5694 	    | L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
5695 	l2pte_set(csrc_pte, nsrc_pte, 0);
5696 	PTE_SYNC(csrc_pte);
5697 
5698 	const pt_entry_t ndst_pte = L2_S_PROTO | dst
5699 	    | L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE)
5700 	    | L2_C | L2_XS_T_TEX(TEX_XSCALE_X);	/* mini-data */
5701 	l2pte_set(cdst_pte, ndst_pte, 0);
5702 	PTE_SYNC(cdst_pte);
5703 
5704 	cpu_tlb_flushD_SE(csrcp);
5705 	cpu_tlb_flushD_SE(cdstp);
5706 	cpu_cpwait();
5707 	bcopy_page(csrcp, cdstp);
5708 	xscale_cache_clean_minidata();
5709 	l2pte_reset(csrc_pte);
5710 	l2pte_reset(cdst_pte);
5711 	PTE_SYNC(csrc_pte);
5712 	PTE_SYNC(cdst_pte);
5713 }
5714 #endif /* ARM_MMU_XSCALE == 1 */
5715 
5716 /*
5717  * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
5718  *
5719  * Return the start and end addresses of the kernel's virtual space.
5720  * These values are setup in pmap_bootstrap and are updated as pages
5721  * are allocated.
5722  */
5723 void
5724 pmap_virtual_space(vaddr_t *start, vaddr_t *end)
5725 {
5726 	*start = virtual_avail;
5727 	*end = virtual_end;
5728 }
5729 
5730 /*
5731  * Helper function for pmap_grow_l2_bucket()
5732  */
5733 static inline int
5734 pmap_grow_map(vaddr_t va, paddr_t *pap)
5735 {
5736 	paddr_t pa;
5737 
5738 	if (uvm.page_init_done == false) {
5739 #ifdef PMAP_STEAL_MEMORY
5740 		pv_addr_t pv;
5741 		pmap_boot_pagealloc(PAGE_SIZE,
5742 #ifdef PMAP_CACHE_VIPT
5743 		    arm_cache_prefer_mask,
5744 		    va & arm_cache_prefer_mask,
5745 #else
5746 		    0, 0,
5747 #endif
5748 		    &pv);
5749 		pa = pv.pv_pa;
5750 #else
5751 		if (uvm_page_physget(&pa) == false)
5752 			return (1);
5753 #endif	/* PMAP_STEAL_MEMORY */
5754 	} else {
5755 		struct vm_page *pg;
5756 		pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE);
5757 		if (pg == NULL)
5758 			return (1);
5759 		pa = VM_PAGE_TO_PHYS(pg);
5760 		/*
5761 		 * This new page must not have any mappings.  Enter it via
5762 		 * pmap_kenter_pa and let that routine do the hard work.
5763 		 */
5764 		struct vm_page_md *md __diagused = VM_PAGE_TO_MD(pg);
5765 		KASSERT(SLIST_EMPTY(&md->pvh_list));
5766 		pmap_kenter_pa(va, pa,
5767 		    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE|PMAP_PTE);
5768 	}
5769 
5770 	if (pap)
5771 		*pap = pa;
5772 
5773 	PMAPCOUNT(pt_mappings);
5774 #ifdef DEBUG
5775 	struct l2_bucket * const l2b = pmap_get_l2_bucket(pmap_kernel(), va);
5776 	KDASSERT(l2b != NULL);
5777 
5778 	pt_entry_t * const ptep = &l2b->l2b_kva[l2pte_index(va)];
5779 	const pt_entry_t opte = *ptep;
5780 	KDASSERT((opte & L2_S_CACHE_MASK) == pte_l2_s_cache_mode_pt);
5781 #endif
5782 	memset((void *)va, 0, PAGE_SIZE);
5783 	return (0);
5784 }
5785 
5786 /*
5787  * This is the same as pmap_alloc_l2_bucket(), except that it is only
5788  * used by pmap_growkernel().
5789  */
5790 static inline struct l2_bucket *
5791 pmap_grow_l2_bucket(pmap_t pm, vaddr_t va)
5792 {
5793 	struct l2_dtable *l2;
5794 	struct l2_bucket *l2b;
5795 	u_short l1slot;
5796 	vaddr_t nva;
5797 
5798 	l1slot = l1pte_index(va);
5799 
5800 	if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL) {
5801 		/*
5802 		 * No mapping at this address, as there is
5803 		 * no entry in the L1 table.
5804 		 * Need to allocate a new l2_dtable.
5805 		 */
5806 		nva = pmap_kernel_l2dtable_kva;
5807 		if ((nva & PGOFSET) == 0) {
5808 			/*
5809 			 * Need to allocate a backing page
5810 			 */
5811 			if (pmap_grow_map(nva, NULL))
5812 				return (NULL);
5813 		}
5814 
5815 		l2 = (struct l2_dtable *)nva;
5816 		nva += sizeof(struct l2_dtable);
5817 
5818 		if ((nva & PGOFSET) < (pmap_kernel_l2dtable_kva & PGOFSET)) {
5819 			/*
5820 			 * The new l2_dtable straddles a page boundary.
5821 			 * Map in another page to cover it.
5822 			 */
5823 			if (pmap_grow_map(nva, NULL))
5824 				return (NULL);
5825 		}
5826 
5827 		pmap_kernel_l2dtable_kva = nva;
5828 
5829 		/*
5830 		 * Link it into the parent pmap
5831 		 */
5832 		pm->pm_l2[L2_IDX(l1slot)] = l2;
5833 	}
5834 
5835 	l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
5836 
5837 	/*
5838 	 * Fetch pointer to the L2 page table associated with the address.
5839 	 */
5840 	if (l2b->l2b_kva == NULL) {
5841 		pt_entry_t *ptep;
5842 
5843 		/*
5844 		 * No L2 page table has been allocated. Chances are, this
5845 		 * is because we just allocated the l2_dtable, above.
5846 		 */
5847 		nva = pmap_kernel_l2ptp_kva;
5848 		ptep = (pt_entry_t *)nva;
5849 		if ((nva & PGOFSET) == 0) {
5850 			/*
5851 			 * Need to allocate a backing page
5852 			 */
5853 			if (pmap_grow_map(nva, &pmap_kernel_l2ptp_phys))
5854 				return (NULL);
5855 			PTE_SYNC_RANGE(ptep, PAGE_SIZE / sizeof(pt_entry_t));
5856 		}
5857 
5858 		l2->l2_occupancy++;
5859 		l2b->l2b_kva = ptep;
5860 		l2b->l2b_l1slot = l1slot;
5861 		l2b->l2b_pa = pmap_kernel_l2ptp_phys;
5862 
5863 		pmap_kernel_l2ptp_kva += L2_TABLE_SIZE_REAL;
5864 		pmap_kernel_l2ptp_phys += L2_TABLE_SIZE_REAL;
5865 	}
5866 
5867 	return (l2b);
5868 }
5869 
5870 vaddr_t
5871 pmap_growkernel(vaddr_t maxkvaddr)
5872 {
5873 	pmap_t kpm = pmap_kernel();
5874 #ifndef ARM_MMU_EXTENDED
5875 	struct l1_ttable *l1;
5876 #endif
5877 	int s;
5878 
5879 	if (maxkvaddr <= pmap_curmaxkvaddr)
5880 		goto out;		/* we are OK */
5881 
5882 	NPDEBUG(PDB_GROWKERN,
5883 	    printf("pmap_growkernel: growing kernel from 0x%lx to 0x%lx\n",
5884 	    pmap_curmaxkvaddr, maxkvaddr));
5885 
5886 	KDASSERT(maxkvaddr <= virtual_end);
5887 
5888 	/*
5889 	 * whoops!   we need to add kernel PTPs
5890 	 */
5891 
5892 	s = splhigh();	/* to be safe */
5893 	mutex_enter(kpm->pm_lock);
5894 
5895 	/* Map 1MB at a time */
5896 	size_t l1slot = l1pte_index(pmap_curmaxkvaddr);
5897 #ifdef ARM_MMU_EXTENDED
5898 	pd_entry_t * const spdep = &kpm->pm_l1[l1slot];
5899 	pd_entry_t *pdep = spdep;
5900 #endif
5901 	for (;pmap_curmaxkvaddr < maxkvaddr; pmap_curmaxkvaddr += L1_S_SIZE,
5902 #ifdef ARM_MMU_EXTENDED
5903 	     pdep++,
5904 #endif
5905 	     l1slot++) {
5906 		struct l2_bucket *l2b =
5907 		    pmap_grow_l2_bucket(kpm, pmap_curmaxkvaddr);
5908 		KASSERT(l2b != NULL);
5909 
5910 		const pd_entry_t npde = L1_C_PROTO | l2b->l2b_pa
5911 		    | L1_C_DOM(PMAP_DOMAIN_KERNEL);
5912 #ifdef ARM_MMU_EXTENDED
5913 		l1pte_setone(pdep, npde);
5914 #else
5915 		/* Distribute new L1 entry to all other L1s */
5916 		SLIST_FOREACH(l1, &l1_list, l1_link) {
5917 			pd_entry_t * const pdep = &l1->l1_kva[l1slot];
5918 			l1pte_setone(pdep, npde);
5919 			PDE_SYNC(pdep);
5920 		}
5921 #endif
5922 	}
5923 #ifdef ARM_MMU_EXTENDED
5924 	PDE_SYNC_RANGE(spdep, pdep - spdep);
5925 #endif
5926 
5927 #ifdef PMAP_CACHE_VIVT
5928 	/*
5929 	 * flush out the cache, expensive but growkernel will happen so
5930 	 * rarely
5931 	 */
5932 	cpu_dcache_wbinv_all();
5933 	cpu_tlb_flushD();
5934 	cpu_cpwait();
5935 #endif
5936 
5937 	mutex_exit(kpm->pm_lock);
5938 	splx(s);
5939 
5940 out:
5941 	return (pmap_curmaxkvaddr);
5942 }
5943 
5944 /************************ Utility routines ****************************/
5945 
5946 #ifndef ARM_HAS_VBAR
5947 /*
5948  * vector_page_setprot:
5949  *
5950  *	Manipulate the protection of the vector page.
5951  */
5952 void
5953 vector_page_setprot(int prot)
5954 {
5955 	struct l2_bucket *l2b;
5956 	pt_entry_t *ptep;
5957 
5958 #if defined(CPU_ARMV7) || defined(CPU_ARM11)
5959 	/*
5960 	 * If we are using VBAR to use the vectors in the kernel, then it's
5961 	 * already mapped in the kernel text so no need to anything here.
5962 	 */
5963 	if (vector_page != ARM_VECTORS_LOW && vector_page != ARM_VECTORS_HIGH) {
5964 		KASSERT((armreg_pfr1_read() & ARM_PFR1_SEC_MASK) != 0);
5965 		return;
5966 	}
5967 #endif
5968 
5969 	l2b = pmap_get_l2_bucket(pmap_kernel(), vector_page);
5970 	KASSERT(l2b != NULL);
5971 
5972 	ptep = &l2b->l2b_kva[l2pte_index(vector_page)];
5973 
5974 	const pt_entry_t opte = *ptep;
5975 #ifdef ARM_MMU_EXTENDED
5976 	const pt_entry_t npte = (opte & ~(L2_S_PROT_MASK|L2_XS_XN))
5977 	    | L2_S_PROT(PTE_KERNEL, prot);
5978 #else
5979 	const pt_entry_t npte = (opte & ~L2_S_PROT_MASK)
5980 	    | L2_S_PROT(PTE_KERNEL, prot);
5981 #endif
5982 	l2pte_set(ptep, npte, opte);
5983 	PTE_SYNC(ptep);
5984 	cpu_tlb_flushD_SE(vector_page);
5985 	cpu_cpwait();
5986 }
5987 #endif
5988 
5989 /*
5990  * Fetch pointers to the PDE/PTE for the given pmap/VA pair.
5991  * Returns true if the mapping exists, else false.
5992  *
5993  * NOTE: This function is only used by a couple of arm-specific modules.
5994  * It is not safe to take any pmap locks here, since we could be right
5995  * in the middle of debugging the pmap anyway...
5996  *
5997  * It is possible for this routine to return false even though a valid
5998  * mapping does exist. This is because we don't lock, so the metadata
5999  * state may be inconsistent.
6000  *
6001  * NOTE: We can return a NULL *ptp in the case where the L1 pde is
6002  * a "section" mapping.
6003  */
6004 bool
6005 pmap_get_pde_pte(pmap_t pm, vaddr_t va, pd_entry_t **pdp, pt_entry_t **ptp)
6006 {
6007 	struct l2_dtable *l2;
6008 	pd_entry_t *pdep, pde;
6009 	pt_entry_t *ptep;
6010 	u_short l1slot;
6011 
6012 	if (pm->pm_l1 == NULL)
6013 		return false;
6014 
6015 	l1slot = l1pte_index(va);
6016 	*pdp = pdep = pmap_l1_kva(pm) + l1slot;
6017 	pde = *pdep;
6018 
6019 	if (l1pte_section_p(pde)) {
6020 		*ptp = NULL;
6021 		return true;
6022 	}
6023 
6024 	l2 = pm->pm_l2[L2_IDX(l1slot)];
6025 	if (l2 == NULL ||
6026 	    (ptep = l2->l2_bucket[L2_BUCKET(l1slot)].l2b_kva) == NULL) {
6027 		return false;
6028 	}
6029 
6030 	*ptp = &ptep[l2pte_index(va)];
6031 	return true;
6032 }
6033 
6034 bool
6035 pmap_get_pde(pmap_t pm, vaddr_t va, pd_entry_t **pdp)
6036 {
6037 
6038 	if (pm->pm_l1 == NULL)
6039 		return false;
6040 
6041 	*pdp = pmap_l1_kva(pm) + l1pte_index(va);
6042 
6043 	return true;
6044 }
6045 
6046 /************************ Bootstrapping routines ****************************/
6047 
6048 #ifndef ARM_MMU_EXTENDED
6049 static void
6050 pmap_init_l1(struct l1_ttable *l1, pd_entry_t *l1pt)
6051 {
6052 	int i;
6053 
6054 	l1->l1_kva = l1pt;
6055 	l1->l1_domain_use_count = 0;
6056 	l1->l1_domain_first = 0;
6057 
6058 	for (i = 0; i < PMAP_DOMAINS; i++)
6059 		l1->l1_domain_free[i] = i + 1;
6060 
6061 	/*
6062 	 * Copy the kernel's L1 entries to each new L1.
6063 	 */
6064 	if (pmap_initialized)
6065 		memcpy(l1pt, pmap_l1_kva(pmap_kernel()), L1_TABLE_SIZE);
6066 
6067 	if (pmap_extract(pmap_kernel(), (vaddr_t)l1pt,
6068 	    &l1->l1_physaddr) == false)
6069 		panic("pmap_init_l1: can't get PA of L1 at %p", l1pt);
6070 
6071 	SLIST_INSERT_HEAD(&l1_list, l1, l1_link);
6072 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
6073 }
6074 #endif /* !ARM_MMU_EXTENDED */
6075 
6076 /*
6077  * pmap_bootstrap() is called from the board-specific initarm() routine
6078  * once the kernel L1/L2 descriptors tables have been set up.
6079  *
6080  * This is a somewhat convoluted process since pmap bootstrap is, effectively,
6081  * spread over a number of disparate files/functions.
6082  *
6083  * We are passed the following parameters
6084  *  - kernel_l1pt
6085  *    This is a pointer to the base of the kernel's L1 translation table.
6086  *  - vstart
6087  *    1MB-aligned start of managed kernel virtual memory.
6088  *  - vend
6089  *    1MB-aligned end of managed kernel virtual memory.
6090  *
6091  * We use the first parameter to build the metadata (struct l1_ttable and
6092  * struct l2_dtable) necessary to track kernel mappings.
6093  */
6094 #define	PMAP_STATIC_L2_SIZE 16
6095 void
6096 pmap_bootstrap(vaddr_t vstart, vaddr_t vend)
6097 {
6098 	static struct l2_dtable static_l2[PMAP_STATIC_L2_SIZE];
6099 #ifndef ARM_MMU_EXTENDED
6100 	static struct l1_ttable static_l1;
6101 	struct l1_ttable *l1 = &static_l1;
6102 #endif
6103 	struct l2_dtable *l2;
6104 	struct l2_bucket *l2b;
6105 	pd_entry_t *l1pt = (pd_entry_t *) kernel_l1pt.pv_va;
6106 	pmap_t pm = pmap_kernel();
6107 	pt_entry_t *ptep;
6108 	paddr_t pa;
6109 	vsize_t size;
6110 	int nptes, l2idx, l2next = 0;
6111 
6112 #ifdef ARM_MMU_EXTENDED
6113 	KASSERT(pte_l1_s_cache_mode == pte_l1_s_cache_mode_pt);
6114 	KASSERT(pte_l2_s_cache_mode == pte_l2_s_cache_mode_pt);
6115 #endif
6116 
6117 #ifdef VERBOSE_INIT_ARM
6118 	printf("kpm ");
6119 #endif
6120 	/*
6121 	 * Initialise the kernel pmap object
6122 	 */
6123 	curcpu()->ci_pmap_cur = pm;
6124 #ifdef ARM_MMU_EXTENDED
6125 	pm->pm_l1 = l1pt;
6126 	pm->pm_l1_pa = kernel_l1pt.pv_pa;
6127 #ifdef VERBOSE_INIT_ARM
6128 	printf("tlb0 ");
6129 #endif
6130 	pmap_tlb_info_init(&pmap_tlb0_info);
6131 #ifdef MULTIPROCESSOR
6132 #ifdef VERBOSE_INIT_ARM
6133 	printf("kcpusets ");
6134 #endif
6135 	pm->pm_onproc = kcpuset_running;
6136 	pm->pm_active = kcpuset_running;
6137 #endif
6138 #else
6139 	pm->pm_l1 = l1;
6140 #endif
6141 
6142 #ifdef VERBOSE_INIT_ARM
6143 	printf("locks ");
6144 #endif
6145 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
6146 	if (arm_cache_prefer_mask != 0) {
6147 		mutex_init(&pmap_lock, MUTEX_DEFAULT, IPL_VM);
6148 	} else {
6149 #endif
6150 		mutex_init(&pmap_lock, MUTEX_DEFAULT, IPL_NONE);
6151 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
6152 	}
6153 #endif
6154 	mutex_init(&pm->pm_obj_lock, MUTEX_DEFAULT, IPL_NONE);
6155 	uvm_obj_init(&pm->pm_obj, NULL, false, 1);
6156 	uvm_obj_setlock(&pm->pm_obj, &pm->pm_obj_lock);
6157 
6158 #ifdef VERBOSE_INIT_ARM
6159 	printf("l1pt ");
6160 #endif
6161 	/*
6162 	 * Scan the L1 translation table created by initarm() and create
6163 	 * the required metadata for all valid mappings found in it.
6164 	 */
6165 	for (size_t l1slot = 0;
6166 	     l1slot < L1_TABLE_SIZE / sizeof(pd_entry_t);
6167 	     l1slot++) {
6168 		pd_entry_t pde = l1pt[l1slot];
6169 
6170 		/*
6171 		 * We're only interested in Coarse mappings.
6172 		 * pmap_extract() can deal with section mappings without
6173 		 * recourse to checking L2 metadata.
6174 		 */
6175 		if ((pde & L1_TYPE_MASK) != L1_TYPE_C)
6176 			continue;
6177 
6178 		/*
6179 		 * Lookup the KVA of this L2 descriptor table
6180 		 */
6181 		pa = l1pte_pa(pde);
6182 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
6183 		if (ptep == NULL) {
6184 			panic("pmap_bootstrap: No L2 for va 0x%x, pa 0x%lx",
6185 			    (u_int)l1slot << L1_S_SHIFT, pa);
6186 		}
6187 
6188 		/*
6189 		 * Fetch the associated L2 metadata structure.
6190 		 * Allocate a new one if necessary.
6191 		 */
6192 		if ((l2 = pm->pm_l2[L2_IDX(l1slot)]) == NULL) {
6193 			if (l2next == PMAP_STATIC_L2_SIZE)
6194 				panic("pmap_bootstrap: out of static L2s");
6195 			pm->pm_l2[L2_IDX(l1slot)] = l2 = &static_l2[l2next++];
6196 		}
6197 
6198 		/*
6199 		 * One more L1 slot tracked...
6200 		 */
6201 		l2->l2_occupancy++;
6202 
6203 		/*
6204 		 * Fill in the details of the L2 descriptor in the
6205 		 * appropriate bucket.
6206 		 */
6207 		l2b = &l2->l2_bucket[L2_BUCKET(l1slot)];
6208 		l2b->l2b_kva = ptep;
6209 		l2b->l2b_pa = pa;
6210 		l2b->l2b_l1slot = l1slot;
6211 
6212 		/*
6213 		 * Establish an initial occupancy count for this descriptor
6214 		 */
6215 		for (l2idx = 0;
6216 		    l2idx < (L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
6217 		    l2idx++) {
6218 			if ((ptep[l2idx] & L2_TYPE_MASK) != L2_TYPE_INV) {
6219 				l2b->l2b_occupancy++;
6220 			}
6221 		}
6222 
6223 		/*
6224 		 * Make sure the descriptor itself has the correct cache mode.
6225 		 * If not, fix it, but whine about the problem. Port-meisters
6226 		 * should consider this a clue to fix up their initarm()
6227 		 * function. :)
6228 		 */
6229 		if (pmap_set_pt_cache_mode(l1pt, (vaddr_t)ptep, 1)) {
6230 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
6231 			    "L2 pte @ %p\n", ptep);
6232 		}
6233 	}
6234 
6235 #ifdef VERBOSE_INIT_ARM
6236 	printf("cache(l1pt) ");
6237 #endif
6238 	/*
6239 	 * Ensure the primary (kernel) L1 has the correct cache mode for
6240 	 * a page table. Bitch if it is not correctly set.
6241 	 */
6242 	if (pmap_set_pt_cache_mode(l1pt, kernel_l1pt.pv_va,
6243 		    L1_TABLE_SIZE / L2_S_SIZE)) {
6244 		printf("pmap_bootstrap: WARNING! wrong cache mode for "
6245 		    "primary L1 @ 0x%lx\n", kernel_l1pt.pv_va);
6246 	}
6247 
6248 #ifdef PMAP_CACHE_VIVT
6249 	cpu_dcache_wbinv_all();
6250 	cpu_tlb_flushID();
6251 	cpu_cpwait();
6252 #endif
6253 
6254 	/*
6255 	 * now we allocate the "special" VAs which are used for tmp mappings
6256 	 * by the pmap (and other modules).  we allocate the VAs by advancing
6257 	 * virtual_avail (note that there are no pages mapped at these VAs).
6258 	 *
6259 	 * Managed KVM space start from wherever initarm() tells us.
6260 	 */
6261 	virtual_avail = vstart;
6262 	virtual_end = vend;
6263 
6264 #ifdef VERBOSE_INIT_ARM
6265 	printf("specials ");
6266 #endif
6267 #ifdef PMAP_CACHE_VIPT
6268 	/*
6269 	 * If we have a VIPT cache, we need one page/pte per possible alias
6270 	 * page so we won't violate cache aliasing rules.
6271 	 */
6272 	virtual_avail = (virtual_avail + arm_cache_prefer_mask) & ~arm_cache_prefer_mask;
6273 	nptes = (arm_cache_prefer_mask >> L2_S_SHIFT) + 1;
6274 	nptes = roundup(nptes, PAGE_SIZE / L2_S_SIZE);
6275 	if (arm_pcache.icache_type != CACHE_TYPE_PIPT
6276 	    && arm_pcache.icache_way_size > nptes * L2_S_SIZE) {
6277 		nptes = arm_pcache.icache_way_size >> L2_S_SHIFT;
6278 		nptes = roundup(nptes, PAGE_SIZE / L2_S_SIZE);
6279 	}
6280 #else
6281 	nptes = PAGE_SIZE / L2_S_SIZE;
6282 #endif
6283 #ifdef MULTIPROCESSOR
6284 	cnptes = nptes;
6285 	nptes *= arm_cpu_max;
6286 #endif
6287 	pmap_alloc_specials(&virtual_avail, nptes, &csrcp, &csrc_pte);
6288 	pmap_set_pt_cache_mode(l1pt, (vaddr_t)csrc_pte, nptes);
6289 	pmap_alloc_specials(&virtual_avail, nptes, &cdstp, &cdst_pte);
6290 	pmap_set_pt_cache_mode(l1pt, (vaddr_t)cdst_pte, nptes);
6291 	pmap_alloc_specials(&virtual_avail, nptes, &memhook, NULL);
6292 	if (msgbufaddr == NULL) {
6293 		pmap_alloc_specials(&virtual_avail,
6294 		    round_page(MSGBUFSIZE) / PAGE_SIZE,
6295 		    (void *)&msgbufaddr, NULL);
6296 	}
6297 
6298 	/*
6299 	 * Allocate a range of kernel virtual address space to be used
6300 	 * for L2 descriptor tables and metadata allocation in
6301 	 * pmap_growkernel().
6302 	 */
6303 	size = ((virtual_end - pmap_curmaxkvaddr) + L1_S_OFFSET) / L1_S_SIZE;
6304 	pmap_alloc_specials(&virtual_avail,
6305 	    round_page(size * L2_TABLE_SIZE_REAL) / PAGE_SIZE,
6306 	    &pmap_kernel_l2ptp_kva, NULL);
6307 
6308 	size = (size + (L2_BUCKET_SIZE - 1)) / L2_BUCKET_SIZE;
6309 	pmap_alloc_specials(&virtual_avail,
6310 	    round_page(size * sizeof(struct l2_dtable)) / PAGE_SIZE,
6311 	    &pmap_kernel_l2dtable_kva, NULL);
6312 
6313 #ifndef ARM_MMU_EXTENDED
6314 	/*
6315 	 * init the static-global locks and global pmap list.
6316 	 */
6317 	mutex_init(&l1_lru_lock, MUTEX_DEFAULT, IPL_VM);
6318 
6319 	/*
6320 	 * We can now initialise the first L1's metadata.
6321 	 */
6322 	SLIST_INIT(&l1_list);
6323 	TAILQ_INIT(&l1_lru_list);
6324 	pmap_init_l1(l1, l1pt);
6325 #endif /* ARM_MMU_EXTENDED */
6326 
6327 #ifndef ARM_HAS_VBAR
6328 	/* Set up vector page L1 details, if necessary */
6329 	if (vector_page < KERNEL_BASE) {
6330 		pm->pm_pl1vec = pmap_l1_kva(pm) + l1pte_index(vector_page);
6331 		l2b = pmap_get_l2_bucket(pm, vector_page);
6332 		KDASSERT(l2b != NULL);
6333 		pm->pm_l1vec = l2b->l2b_pa | L1_C_PROTO |
6334 		    L1_C_DOM(pmap_domain(pm));
6335 	} else
6336 		pm->pm_pl1vec = NULL;
6337 #endif
6338 
6339 #ifdef VERBOSE_INIT_ARM
6340 	printf("pools ");
6341 #endif
6342 	/*
6343 	 * Initialize the pmap cache
6344 	 */
6345 	pool_cache_bootstrap(&pmap_cache, sizeof(struct pmap), 0, 0, 0,
6346 	    "pmappl", NULL, IPL_NONE, pmap_pmap_ctor, NULL, NULL);
6347 
6348 	/*
6349 	 * Initialize the pv pool.
6350 	 */
6351 	pool_init(&pmap_pv_pool, sizeof(struct pv_entry), 0, 0, 0, "pvepl",
6352 	    &pmap_bootstrap_pv_allocator, IPL_NONE);
6353 
6354 	/*
6355 	 * Initialize the L2 dtable pool and cache.
6356 	 */
6357 	pool_cache_bootstrap(&pmap_l2dtable_cache, sizeof(struct l2_dtable), 0,
6358 	    0, 0, "l2dtblpl", NULL, IPL_NONE, pmap_l2dtable_ctor, NULL, NULL);
6359 
6360 	/*
6361 	 * Initialise the L2 descriptor table pool and cache
6362 	 */
6363 	pool_cache_bootstrap(&pmap_l2ptp_cache, L2_TABLE_SIZE_REAL, 0,
6364 	    L2_TABLE_SIZE_REAL, 0, "l2ptppl", NULL, IPL_NONE,
6365 	    pmap_l2ptp_ctor, NULL, NULL);
6366 
6367 	mutex_init(&memlock, MUTEX_DEFAULT, IPL_NONE);
6368 
6369 	cpu_dcache_wbinv_all();
6370 }
6371 
6372 static bool
6373 pmap_set_pt_cache_mode(pd_entry_t *kl1, vaddr_t va, size_t nptes)
6374 {
6375 #ifdef ARM_MMU_EXTENDED
6376 	return false;
6377 #else
6378 	if (pte_l1_s_cache_mode == pte_l1_s_cache_mode_pt
6379 	    && pte_l2_s_cache_mode == pte_l2_s_cache_mode_pt)
6380 		return false;
6381 
6382 	const vaddr_t eva = va + nptes * PAGE_SIZE;
6383 	int rv = 0;
6384 
6385 	while (va < eva) {
6386 		/*
6387 		 * Make sure the descriptor itself has the correct cache mode
6388 		 */
6389 		pd_entry_t * const pdep = &kl1[l1pte_index(va)];
6390 		pd_entry_t pde = *pdep;
6391 
6392 		if (l1pte_section_p(pde)) {
6393 			__CTASSERT((L1_S_CACHE_MASK & L1_S_V6_SUPER) == 0);
6394 			if ((pde & L1_S_CACHE_MASK) != pte_l1_s_cache_mode_pt) {
6395 				*pdep = (pde & ~L1_S_CACHE_MASK) |
6396 				    pte_l1_s_cache_mode_pt;
6397 				PDE_SYNC(pdep);
6398 				cpu_dcache_wbinv_range((vaddr_t)pdep,
6399 				    sizeof(*pdep));
6400 				rv = 1;
6401 			}
6402 			return rv;
6403 		}
6404 		vaddr_t pa = l1pte_pa(pde);
6405 		pt_entry_t *ptep = (pt_entry_t *)kernel_pt_lookup(pa);
6406 		if (ptep == NULL)
6407 			panic("pmap_bootstrap: No PTP for va %#lx\n", va);
6408 
6409 		ptep += l2pte_index(va);
6410 		const pt_entry_t opte = *ptep;
6411 		if ((opte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
6412 			const pt_entry_t npte = (opte & ~L2_S_CACHE_MASK)
6413 			    | pte_l2_s_cache_mode_pt;
6414 			l2pte_set(ptep, npte, opte);
6415 			PTE_SYNC(ptep);
6416 			cpu_dcache_wbinv_range((vaddr_t)ptep, sizeof(*ptep));
6417 			rv = 1;
6418 		}
6419 		va += PAGE_SIZE;
6420 	}
6421 
6422 	return (rv);
6423 #endif
6424 }
6425 
6426 static void
6427 pmap_alloc_specials(vaddr_t *availp, int pages, vaddr_t *vap, pt_entry_t **ptep)
6428 {
6429 	vaddr_t va = *availp;
6430 	struct l2_bucket *l2b;
6431 
6432 	if (ptep) {
6433 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
6434 		if (l2b == NULL)
6435 			panic("pmap_alloc_specials: no l2b for 0x%lx", va);
6436 
6437 		if (ptep)
6438 			*ptep = &l2b->l2b_kva[l2pte_index(va)];
6439 	}
6440 
6441 	*vap = va;
6442 	*availp = va + (PAGE_SIZE * pages);
6443 }
6444 
6445 void
6446 pmap_init(void)
6447 {
6448 
6449 	/*
6450 	 * Set the available memory vars - These do not map to real memory
6451 	 * addresses and cannot as the physical memory is fragmented.
6452 	 * They are used by ps for %mem calculations.
6453 	 * One could argue whether this should be the entire memory or just
6454 	 * the memory that is useable in a user process.
6455 	 */
6456 	avail_start = ptoa(VM_PHYSMEM_PTR(0)->start);
6457 	avail_end = ptoa(VM_PHYSMEM_PTR(vm_nphysseg - 1)->end);
6458 
6459 	/*
6460 	 * Now we need to free enough pv_entry structures to allow us to get
6461 	 * the kmem_map/kmem_object allocated and inited (done after this
6462 	 * function is finished).  to do this we allocate one bootstrap page out
6463 	 * of kernel_map and use it to provide an initial pool of pv_entry
6464 	 * structures.   we never free this page.
6465 	 */
6466 	pool_setlowat(&pmap_pv_pool, (PAGE_SIZE / sizeof(struct pv_entry)) * 2);
6467 
6468 #ifdef ARM_MMU_EXTENDED
6469 	pmap_tlb_info_evcnt_attach(&pmap_tlb0_info);
6470 #endif
6471 
6472 	pmap_initialized = true;
6473 }
6474 
6475 static vaddr_t last_bootstrap_page = 0;
6476 static void *free_bootstrap_pages = NULL;
6477 
6478 static void *
6479 pmap_bootstrap_pv_page_alloc(struct pool *pp, int flags)
6480 {
6481 	extern void *pool_page_alloc(struct pool *, int);
6482 	vaddr_t new_page;
6483 	void *rv;
6484 
6485 	if (pmap_initialized)
6486 		return (pool_page_alloc(pp, flags));
6487 
6488 	if (free_bootstrap_pages) {
6489 		rv = free_bootstrap_pages;
6490 		free_bootstrap_pages = *((void **)rv);
6491 		return (rv);
6492 	}
6493 
6494 	KASSERT(kernel_map != NULL);
6495 	new_page = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
6496 	    UVM_KMF_WIRED | ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT));
6497 
6498 	KASSERT(new_page > last_bootstrap_page);
6499 	last_bootstrap_page = new_page;
6500 	return ((void *)new_page);
6501 }
6502 
6503 static void
6504 pmap_bootstrap_pv_page_free(struct pool *pp, void *v)
6505 {
6506 	extern void pool_page_free(struct pool *, void *);
6507 
6508 	if ((vaddr_t)v <= last_bootstrap_page) {
6509 		*((void **)v) = free_bootstrap_pages;
6510 		free_bootstrap_pages = v;
6511 		return;
6512 	}
6513 
6514 	if (pmap_initialized) {
6515 		pool_page_free(pp, v);
6516 		return;
6517 	}
6518 }
6519 
6520 /*
6521  * pmap_postinit()
6522  *
6523  * This routine is called after the vm and kmem subsystems have been
6524  * initialised. This allows the pmap code to perform any initialisation
6525  * that can only be done one the memory allocation is in place.
6526  */
6527 void
6528 pmap_postinit(void)
6529 {
6530 #ifndef ARM_MMU_EXTENDED
6531 	extern paddr_t physical_start, physical_end;
6532 	struct l1_ttable *l1;
6533 	struct pglist plist;
6534 	struct vm_page *m;
6535 	pd_entry_t *pdep;
6536 	vaddr_t va, eva;
6537 	u_int loop, needed;
6538 	int error;
6539 #endif
6540 
6541 	pool_cache_setlowat(&pmap_l2ptp_cache, (PAGE_SIZE / L2_TABLE_SIZE_REAL) * 4);
6542 	pool_cache_setlowat(&pmap_l2dtable_cache,
6543 	    (PAGE_SIZE / sizeof(struct l2_dtable)) * 2);
6544 
6545 #ifndef ARM_MMU_EXTENDED
6546 	needed = (maxproc / PMAP_DOMAINS) + ((maxproc % PMAP_DOMAINS) ? 1 : 0);
6547 	needed -= 1;
6548 
6549 	l1 = kmem_alloc(sizeof(*l1) * needed, KM_SLEEP);
6550 
6551 	for (loop = 0; loop < needed; loop++, l1++) {
6552 		/* Allocate a L1 page table */
6553 		va = uvm_km_alloc(kernel_map, L1_TABLE_SIZE, 0, UVM_KMF_VAONLY);
6554 		if (va == 0)
6555 			panic("Cannot allocate L1 KVM");
6556 
6557 		error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start,
6558 		    physical_end, L1_TABLE_SIZE, 0, &plist, 1, 1);
6559 		if (error)
6560 			panic("Cannot allocate L1 physical pages");
6561 
6562 		m = TAILQ_FIRST(&plist);
6563 		eva = va + L1_TABLE_SIZE;
6564 		pdep = (pd_entry_t *)va;
6565 
6566 		while (m && va < eva) {
6567 			paddr_t pa = VM_PAGE_TO_PHYS(m);
6568 
6569 			pmap_kenter_pa(va, pa,
6570 			    VM_PROT_READ|VM_PROT_WRITE, PMAP_KMPAGE|PMAP_PTE);
6571 
6572 			va += PAGE_SIZE;
6573 			m = TAILQ_NEXT(m, pageq.queue);
6574 		}
6575 
6576 #ifdef DIAGNOSTIC
6577 		if (m)
6578 			panic("pmap_alloc_l1pt: pglist not empty");
6579 #endif	/* DIAGNOSTIC */
6580 
6581 		pmap_init_l1(l1, pdep);
6582 	}
6583 
6584 #ifdef DEBUG
6585 	printf("pmap_postinit: Allocated %d static L1 descriptor tables\n",
6586 	    needed);
6587 #endif
6588 #endif /* !ARM_MMU_EXTENDED */
6589 }
6590 
6591 /*
6592  * Note that the following routines are used by board-specific initialisation
6593  * code to configure the initial kernel page tables.
6594  *
6595  * If ARM32_NEW_VM_LAYOUT is *not* defined, they operate on the assumption that
6596  * L2 page-table pages are 4KB in size and use 4 L1 slots. This mimics the
6597  * behaviour of the old pmap, and provides an easy migration path for
6598  * initial bring-up of the new pmap on existing ports. Fortunately,
6599  * pmap_bootstrap() compensates for this hackery. This is only a stop-gap and
6600  * will be deprecated.
6601  *
6602  * If ARM32_NEW_VM_LAYOUT *is* defined, these functions deal with 1KB L2 page
6603  * tables.
6604  */
6605 
6606 /*
6607  * This list exists for the benefit of pmap_map_chunk().  It keeps track
6608  * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
6609  * find them as necessary.
6610  *
6611  * Note that the data on this list MUST remain valid after initarm() returns,
6612  * as pmap_bootstrap() uses it to contruct L2 table metadata.
6613  */
6614 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
6615 
6616 static vaddr_t
6617 kernel_pt_lookup(paddr_t pa)
6618 {
6619 	pv_addr_t *pv;
6620 
6621 	SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
6622 		if (pv->pv_pa == (pa & ~PGOFSET))
6623 			return (pv->pv_va | (pa & PGOFSET));
6624 	}
6625 	return (0);
6626 }
6627 
6628 /*
6629  * pmap_map_section:
6630  *
6631  *	Create a single section mapping.
6632  */
6633 void
6634 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
6635 {
6636 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
6637 	const size_t l1slot = l1pte_index(va);
6638 	pd_entry_t fl;
6639 
6640 	KASSERT(((va | pa) & L1_S_OFFSET) == 0);
6641 
6642 	switch (cache) {
6643 	case PTE_NOCACHE:
6644 	default:
6645 		fl = 0;
6646 		break;
6647 
6648 	case PTE_CACHE:
6649 		fl = pte_l1_s_cache_mode;
6650 		break;
6651 
6652 	case PTE_PAGETABLE:
6653 		fl = pte_l1_s_cache_mode_pt;
6654 		break;
6655 	}
6656 
6657 	const pd_entry_t npde = L1_S_PROTO | pa |
6658 	    L1_S_PROT(PTE_KERNEL, prot) | fl | L1_S_DOM(PMAP_DOMAIN_KERNEL);
6659 	l1pte_setone(pdep + l1slot, npde);
6660 	PDE_SYNC(pdep + l1slot);
6661 }
6662 
6663 /*
6664  * pmap_map_entry:
6665  *
6666  *	Create a single page mapping.
6667  */
6668 void
6669 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
6670 {
6671 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
6672 	const size_t l1slot = l1pte_index(va);
6673 	pt_entry_t npte;
6674 	pt_entry_t *ptep;
6675 
6676 	KASSERT(((va | pa) & PGOFSET) == 0);
6677 
6678 	switch (cache) {
6679 	case PTE_NOCACHE:
6680 	default:
6681 		npte = 0;
6682 		break;
6683 
6684 	case PTE_CACHE:
6685 		npte = pte_l2_s_cache_mode;
6686 		break;
6687 
6688 	case PTE_PAGETABLE:
6689 		npte = pte_l2_s_cache_mode_pt;
6690 		break;
6691 	}
6692 
6693 	if ((pdep[l1slot] & L1_TYPE_MASK) != L1_TYPE_C)
6694 		panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
6695 
6696 	ptep = (pt_entry_t *) kernel_pt_lookup(l1pte_pa(pdep[l1slot]));
6697 	if (ptep == NULL)
6698 		panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
6699 
6700 	npte |= L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot);
6701 #ifdef ARM_MMU_EXTENDED
6702 	if (prot & VM_PROT_EXECUTE) {
6703 		npte &= ~L2_XS_XN;
6704 	}
6705 #endif
6706 	ptep += l2pte_index(va);
6707 	l2pte_set(ptep, npte, 0);
6708 	PTE_SYNC(ptep);
6709 }
6710 
6711 /*
6712  * pmap_link_l2pt:
6713  *
6714  *	Link the L2 page table specified by "l2pv" into the L1
6715  *	page table at the slot for "va".
6716  */
6717 void
6718 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
6719 {
6720 	pd_entry_t * const pdep = (pd_entry_t *) l1pt + l1pte_index(va);
6721 
6722 	KASSERT((va & ((L1_S_SIZE * (PAGE_SIZE / L2_T_SIZE)) - 1)) == 0);
6723 	KASSERT((l2pv->pv_pa & PGOFSET) == 0);
6724 
6725 	const pd_entry_t npde = L1_S_DOM(PMAP_DOMAIN_KERNEL) | L1_C_PROTO
6726 	    | l2pv->pv_pa;
6727 
6728 	l1pte_set(pdep, npde);
6729 	PDE_SYNC_RANGE(pdep, PAGE_SIZE / L2_T_SIZE);
6730 
6731 	SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
6732 }
6733 
6734 /*
6735  * pmap_map_chunk:
6736  *
6737  *	Map a chunk of memory using the most efficient mappings
6738  *	possible (section, large page, small page) into the
6739  *	provided L1 and L2 tables at the specified virtual address.
6740  */
6741 vsize_t
6742 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
6743     int prot, int cache)
6744 {
6745 	pd_entry_t * const pdep = (pd_entry_t *) l1pt;
6746 	pt_entry_t f1, f2s, f2l;
6747 	vsize_t resid;
6748 
6749 	resid = (size + (PAGE_SIZE - 1)) & ~(PAGE_SIZE - 1);
6750 
6751 	if (l1pt == 0)
6752 		panic("pmap_map_chunk: no L1 table provided");
6753 
6754 #ifdef VERBOSE_INIT_ARM
6755 	printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
6756 	    "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
6757 #endif
6758 
6759 	switch (cache) {
6760 	case PTE_NOCACHE:
6761 	default:
6762 		f1 = 0;
6763 		f2l = 0;
6764 		f2s = 0;
6765 		break;
6766 
6767 	case PTE_CACHE:
6768 		f1 = pte_l1_s_cache_mode;
6769 		f2l = pte_l2_l_cache_mode;
6770 		f2s = pte_l2_s_cache_mode;
6771 		break;
6772 
6773 	case PTE_PAGETABLE:
6774 		f1 = pte_l1_s_cache_mode_pt;
6775 		f2l = pte_l2_l_cache_mode_pt;
6776 		f2s = pte_l2_s_cache_mode_pt;
6777 		break;
6778 	}
6779 
6780 	size = resid;
6781 
6782 	while (resid > 0) {
6783 		const size_t l1slot = l1pte_index(va);
6784 #if (ARM_MMU_V6 + ARM_MMU_V7) > 0
6785 		/* See if we can use a supersection mapping. */
6786 		if (L1_SS_PROTO && L1_SS_MAPPABLE_P(va, pa, resid)) {
6787 			/* Supersection are always domain 0 */
6788 			const pd_entry_t npde = L1_SS_PROTO | pa
6789 #ifdef ARM_MMU_EXTENDED
6790 			    | ((prot & VM_PROT_EXECUTE) ? 0 : L1_S_V6_XN)
6791 			    | (va & 0x80000000 ? 0 : L1_S_V6_nG)
6792 #endif
6793 			    | L1_S_PROT(PTE_KERNEL, prot) | f1;
6794 #ifdef VERBOSE_INIT_ARM
6795 			printf("sS");
6796 #endif
6797 			l1pte_set(&pdep[l1slot], npde);
6798 			PDE_SYNC_RANGE(&pdep[l1slot], L1_SS_SIZE / L1_S_SIZE);
6799 			va += L1_SS_SIZE;
6800 			pa += L1_SS_SIZE;
6801 			resid -= L1_SS_SIZE;
6802 			continue;
6803 		}
6804 #endif
6805 		/* See if we can use a section mapping. */
6806 		if (L1_S_MAPPABLE_P(va, pa, resid)) {
6807 			const pd_entry_t npde = L1_S_PROTO | pa
6808 #ifdef ARM_MMU_EXTENDED
6809 			    | ((prot & VM_PROT_EXECUTE) ? 0 : L1_S_V6_XN)
6810 			    | (va & 0x80000000 ? 0 : L1_S_V6_nG)
6811 #endif
6812 			    | L1_S_PROT(PTE_KERNEL, prot) | f1
6813 			    | L1_S_DOM(PMAP_DOMAIN_KERNEL);
6814 #ifdef VERBOSE_INIT_ARM
6815 			printf("S");
6816 #endif
6817 			l1pte_set(&pdep[l1slot], npde);
6818 			PDE_SYNC(&pdep[l1slot]);
6819 			va += L1_S_SIZE;
6820 			pa += L1_S_SIZE;
6821 			resid -= L1_S_SIZE;
6822 			continue;
6823 		}
6824 
6825 		/*
6826 		 * Ok, we're going to use an L2 table.  Make sure
6827 		 * one is actually in the corresponding L1 slot
6828 		 * for the current VA.
6829 		 */
6830 		if ((pdep[l1slot] & L1_TYPE_MASK) != L1_TYPE_C)
6831 			panic("%s: no L2 table for VA %#lx", __func__, va);
6832 
6833 		pt_entry_t *ptep = (pt_entry_t *) kernel_pt_lookup(l1pte_pa(pdep[l1slot]));
6834 		if (ptep == NULL)
6835 			panic("%s: can't find L2 table for VA %#lx", __func__,
6836 			    va);
6837 
6838 		ptep += l2pte_index(va);
6839 
6840 		/* See if we can use a L2 large page mapping. */
6841 		if (L2_L_MAPPABLE_P(va, pa, resid)) {
6842 			const pt_entry_t npte = L2_L_PROTO | pa
6843 #ifdef ARM_MMU_EXTENDED
6844 			    | ((prot & VM_PROT_EXECUTE) ? 0 : L2_XS_L_XN)
6845 			    | (va & 0x80000000 ? 0 : L2_XS_nG)
6846 #endif
6847 			    | L2_L_PROT(PTE_KERNEL, prot) | f2l;
6848 #ifdef VERBOSE_INIT_ARM
6849 			printf("L");
6850 #endif
6851 			l2pte_set(ptep, npte, 0);
6852 			PTE_SYNC_RANGE(ptep, L2_L_SIZE / L2_S_SIZE);
6853 			va += L2_L_SIZE;
6854 			pa += L2_L_SIZE;
6855 			resid -= L2_L_SIZE;
6856 			continue;
6857 		}
6858 
6859 #ifdef VERBOSE_INIT_ARM
6860 		printf("P");
6861 #endif
6862 		/* Use a small page mapping. */
6863 		pt_entry_t npte = L2_S_PROTO | pa
6864 #ifdef ARM_MMU_EXTENDED
6865 		    | ((prot & VM_PROT_EXECUTE) ? 0 : L2_XS_XN)
6866 		    | (va & 0x80000000 ? 0 : L2_XS_nG)
6867 #endif
6868 		    | L2_S_PROT(PTE_KERNEL, prot) | f2s;
6869 #ifdef ARM_MMU_EXTENDED
6870 		npte &= ((prot & VM_PROT_EXECUTE) ? ~L2_XS_XN : ~0);
6871 #endif
6872 		l2pte_set(ptep, npte, 0);
6873 		PTE_SYNC(ptep);
6874 		va += PAGE_SIZE;
6875 		pa += PAGE_SIZE;
6876 		resid -= PAGE_SIZE;
6877 	}
6878 #ifdef VERBOSE_INIT_ARM
6879 	printf("\n");
6880 #endif
6881 	return (size);
6882 }
6883 
6884 /********************** Static device map routines ***************************/
6885 
6886 static const struct pmap_devmap *pmap_devmap_table;
6887 
6888 /*
6889  * Register the devmap table.  This is provided in case early console
6890  * initialization needs to register mappings created by bootstrap code
6891  * before pmap_devmap_bootstrap() is called.
6892  */
6893 void
6894 pmap_devmap_register(const struct pmap_devmap *table)
6895 {
6896 
6897 	pmap_devmap_table = table;
6898 }
6899 
6900 /*
6901  * Map all of the static regions in the devmap table, and remember
6902  * the devmap table so other parts of the kernel can look up entries
6903  * later.
6904  */
6905 void
6906 pmap_devmap_bootstrap(vaddr_t l1pt, const struct pmap_devmap *table)
6907 {
6908 	int i;
6909 
6910 	pmap_devmap_table = table;
6911 
6912 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
6913 #ifdef VERBOSE_INIT_ARM
6914 		printf("devmap: %08lx -> %08lx @ %08lx\n",
6915 		    pmap_devmap_table[i].pd_pa,
6916 		    pmap_devmap_table[i].pd_pa +
6917 			pmap_devmap_table[i].pd_size - 1,
6918 		    pmap_devmap_table[i].pd_va);
6919 #endif
6920 		pmap_map_chunk(l1pt, pmap_devmap_table[i].pd_va,
6921 		    pmap_devmap_table[i].pd_pa,
6922 		    pmap_devmap_table[i].pd_size,
6923 		    pmap_devmap_table[i].pd_prot,
6924 		    pmap_devmap_table[i].pd_cache);
6925 	}
6926 }
6927 
6928 const struct pmap_devmap *
6929 pmap_devmap_find_pa(paddr_t pa, psize_t size)
6930 {
6931 	uint64_t endpa;
6932 	int i;
6933 
6934 	if (pmap_devmap_table == NULL)
6935 		return (NULL);
6936 
6937 	endpa = (uint64_t)pa + (uint64_t)(size - 1);
6938 
6939 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
6940 		if (pa >= pmap_devmap_table[i].pd_pa &&
6941 		    endpa <= (uint64_t)pmap_devmap_table[i].pd_pa +
6942 			     (uint64_t)(pmap_devmap_table[i].pd_size - 1))
6943 			return (&pmap_devmap_table[i]);
6944 	}
6945 
6946 	return (NULL);
6947 }
6948 
6949 const struct pmap_devmap *
6950 pmap_devmap_find_va(vaddr_t va, vsize_t size)
6951 {
6952 	int i;
6953 
6954 	if (pmap_devmap_table == NULL)
6955 		return (NULL);
6956 
6957 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
6958 		if (va >= pmap_devmap_table[i].pd_va &&
6959 		    va + size - 1 <= pmap_devmap_table[i].pd_va +
6960 				     pmap_devmap_table[i].pd_size - 1)
6961 			return (&pmap_devmap_table[i]);
6962 	}
6963 
6964 	return (NULL);
6965 }
6966 
6967 /********************** PTE initialization routines **************************/
6968 
6969 /*
6970  * These routines are called when the CPU type is identified to set up
6971  * the PTE prototypes, cache modes, etc.
6972  *
6973  * The variables are always here, just in case modules need to reference
6974  * them (though, they shouldn't).
6975  */
6976 
6977 pt_entry_t	pte_l1_s_cache_mode;
6978 pt_entry_t	pte_l1_s_wc_mode;
6979 pt_entry_t	pte_l1_s_cache_mode_pt;
6980 pt_entry_t	pte_l1_s_cache_mask;
6981 
6982 pt_entry_t	pte_l2_l_cache_mode;
6983 pt_entry_t	pte_l2_l_wc_mode;
6984 pt_entry_t	pte_l2_l_cache_mode_pt;
6985 pt_entry_t	pte_l2_l_cache_mask;
6986 
6987 pt_entry_t	pte_l2_s_cache_mode;
6988 pt_entry_t	pte_l2_s_wc_mode;
6989 pt_entry_t	pte_l2_s_cache_mode_pt;
6990 pt_entry_t	pte_l2_s_cache_mask;
6991 
6992 pt_entry_t	pte_l1_s_prot_u;
6993 pt_entry_t	pte_l1_s_prot_w;
6994 pt_entry_t	pte_l1_s_prot_ro;
6995 pt_entry_t	pte_l1_s_prot_mask;
6996 
6997 pt_entry_t	pte_l2_s_prot_u;
6998 pt_entry_t	pte_l2_s_prot_w;
6999 pt_entry_t	pte_l2_s_prot_ro;
7000 pt_entry_t	pte_l2_s_prot_mask;
7001 
7002 pt_entry_t	pte_l2_l_prot_u;
7003 pt_entry_t	pte_l2_l_prot_w;
7004 pt_entry_t	pte_l2_l_prot_ro;
7005 pt_entry_t	pte_l2_l_prot_mask;
7006 
7007 pt_entry_t	pte_l1_ss_proto;
7008 pt_entry_t	pte_l1_s_proto;
7009 pt_entry_t	pte_l1_c_proto;
7010 pt_entry_t	pte_l2_s_proto;
7011 
7012 void		(*pmap_copy_page_func)(paddr_t, paddr_t);
7013 void		(*pmap_zero_page_func)(paddr_t);
7014 
7015 #if (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6 + ARM_MMU_V7) != 0
7016 void
7017 pmap_pte_init_generic(void)
7018 {
7019 
7020 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
7021 	pte_l1_s_wc_mode = L1_S_B;
7022 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_generic;
7023 
7024 	pte_l2_l_cache_mode = L2_B|L2_C;
7025 	pte_l2_l_wc_mode = L2_B;
7026 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_generic;
7027 
7028 	pte_l2_s_cache_mode = L2_B|L2_C;
7029 	pte_l2_s_wc_mode = L2_B;
7030 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_generic;
7031 
7032 	/*
7033 	 * If we have a write-through cache, set B and C.  If
7034 	 * we have a write-back cache, then we assume setting
7035 	 * only C will make those pages write-through (except for those
7036 	 * Cortex CPUs which can read the L1 caches).
7037 	 */
7038 	if (cpufuncs.cf_dcache_wb_range == (void *) cpufunc_nullop
7039 #if ARM_MMU_V7 > 0
7040 	    || CPU_ID_CORTEX_P(curcpu()->ci_arm_cpuid)
7041 #endif
7042 #if ARM_MMU_V6 > 0
7043 	    || CPU_ID_ARM11_P(curcpu()->ci_arm_cpuid) /* arm116 errata 399234 */
7044 #endif
7045 	    || false) {
7046 		pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
7047 		pte_l2_l_cache_mode_pt = L2_B|L2_C;
7048 		pte_l2_s_cache_mode_pt = L2_B|L2_C;
7049 	} else {
7050 		pte_l1_s_cache_mode_pt = L1_S_C;	/* write through */
7051 		pte_l2_l_cache_mode_pt = L2_C;		/* write through */
7052 		pte_l2_s_cache_mode_pt = L2_C;		/* write through */
7053 	}
7054 
7055 	pte_l1_s_prot_u = L1_S_PROT_U_generic;
7056 	pte_l1_s_prot_w = L1_S_PROT_W_generic;
7057 	pte_l1_s_prot_ro = L1_S_PROT_RO_generic;
7058 	pte_l1_s_prot_mask = L1_S_PROT_MASK_generic;
7059 
7060 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
7061 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
7062 	pte_l2_s_prot_ro = L2_S_PROT_RO_generic;
7063 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
7064 
7065 	pte_l2_l_prot_u = L2_L_PROT_U_generic;
7066 	pte_l2_l_prot_w = L2_L_PROT_W_generic;
7067 	pte_l2_l_prot_ro = L2_L_PROT_RO_generic;
7068 	pte_l2_l_prot_mask = L2_L_PROT_MASK_generic;
7069 
7070 	pte_l1_ss_proto = L1_SS_PROTO_generic;
7071 	pte_l1_s_proto = L1_S_PROTO_generic;
7072 	pte_l1_c_proto = L1_C_PROTO_generic;
7073 	pte_l2_s_proto = L2_S_PROTO_generic;
7074 
7075 	pmap_copy_page_func = pmap_copy_page_generic;
7076 	pmap_zero_page_func = pmap_zero_page_generic;
7077 }
7078 
7079 #if defined(CPU_ARM8)
7080 void
7081 pmap_pte_init_arm8(void)
7082 {
7083 
7084 	/*
7085 	 * ARM8 is compatible with generic, but we need to use
7086 	 * the page tables uncached.
7087 	 */
7088 	pmap_pte_init_generic();
7089 
7090 	pte_l1_s_cache_mode_pt = 0;
7091 	pte_l2_l_cache_mode_pt = 0;
7092 	pte_l2_s_cache_mode_pt = 0;
7093 }
7094 #endif /* CPU_ARM8 */
7095 
7096 #if defined(CPU_ARM9) && defined(ARM9_CACHE_WRITE_THROUGH)
7097 void
7098 pmap_pte_init_arm9(void)
7099 {
7100 
7101 	/*
7102 	 * ARM9 is compatible with generic, but we want to use
7103 	 * write-through caching for now.
7104 	 */
7105 	pmap_pte_init_generic();
7106 
7107 	pte_l1_s_cache_mode = L1_S_C;
7108 	pte_l2_l_cache_mode = L2_C;
7109 	pte_l2_s_cache_mode = L2_C;
7110 
7111 	pte_l1_s_wc_mode = L1_S_B;
7112 	pte_l2_l_wc_mode = L2_B;
7113 	pte_l2_s_wc_mode = L2_B;
7114 
7115 	pte_l1_s_cache_mode_pt = L1_S_C;
7116 	pte_l2_l_cache_mode_pt = L2_C;
7117 	pte_l2_s_cache_mode_pt = L2_C;
7118 }
7119 #endif /* CPU_ARM9 && ARM9_CACHE_WRITE_THROUGH */
7120 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1 + ARM_MMU_V6) != 0 */
7121 
7122 #if defined(CPU_ARM10)
7123 void
7124 pmap_pte_init_arm10(void)
7125 {
7126 
7127 	/*
7128 	 * ARM10 is compatible with generic, but we want to use
7129 	 * write-through caching for now.
7130 	 */
7131 	pmap_pte_init_generic();
7132 
7133 	pte_l1_s_cache_mode = L1_S_B | L1_S_C;
7134 	pte_l2_l_cache_mode = L2_B | L2_C;
7135 	pte_l2_s_cache_mode = L2_B | L2_C;
7136 
7137 	pte_l1_s_cache_mode = L1_S_B;
7138 	pte_l2_l_cache_mode = L2_B;
7139 	pte_l2_s_cache_mode = L2_B;
7140 
7141 	pte_l1_s_cache_mode_pt = L1_S_C;
7142 	pte_l2_l_cache_mode_pt = L2_C;
7143 	pte_l2_s_cache_mode_pt = L2_C;
7144 
7145 }
7146 #endif /* CPU_ARM10 */
7147 
7148 #if defined(CPU_ARM11) && defined(ARM11_CACHE_WRITE_THROUGH)
7149 void
7150 pmap_pte_init_arm11(void)
7151 {
7152 
7153 	/*
7154 	 * ARM11 is compatible with generic, but we want to use
7155 	 * write-through caching for now.
7156 	 */
7157 	pmap_pte_init_generic();
7158 
7159 	pte_l1_s_cache_mode = L1_S_C;
7160 	pte_l2_l_cache_mode = L2_C;
7161 	pte_l2_s_cache_mode = L2_C;
7162 
7163 	pte_l1_s_wc_mode = L1_S_B;
7164 	pte_l2_l_wc_mode = L2_B;
7165 	pte_l2_s_wc_mode = L2_B;
7166 
7167 	pte_l1_s_cache_mode_pt = L1_S_C;
7168 	pte_l2_l_cache_mode_pt = L2_C;
7169 	pte_l2_s_cache_mode_pt = L2_C;
7170 }
7171 #endif /* CPU_ARM11 && ARM11_CACHE_WRITE_THROUGH */
7172 
7173 #if ARM_MMU_SA1 == 1
7174 void
7175 pmap_pte_init_sa1(void)
7176 {
7177 
7178 	/*
7179 	 * The StrongARM SA-1 cache does not have a write-through
7180 	 * mode.  So, do the generic initialization, then reset
7181 	 * the page table cache mode to B=1,C=1, and note that
7182 	 * the PTEs need to be sync'd.
7183 	 */
7184 	pmap_pte_init_generic();
7185 
7186 	pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
7187 	pte_l2_l_cache_mode_pt = L2_B|L2_C;
7188 	pte_l2_s_cache_mode_pt = L2_B|L2_C;
7189 
7190 	pmap_needs_pte_sync = 1;
7191 }
7192 #endif /* ARM_MMU_SA1 == 1*/
7193 
7194 #if ARM_MMU_XSCALE == 1
7195 #if (ARM_NMMUS > 1)
7196 static u_int xscale_use_minidata;
7197 #endif
7198 
7199 void
7200 pmap_pte_init_xscale(void)
7201 {
7202 	uint32_t auxctl;
7203 	int write_through = 0;
7204 
7205 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
7206 	pte_l1_s_wc_mode = L1_S_B;
7207 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_xscale;
7208 
7209 	pte_l2_l_cache_mode = L2_B|L2_C;
7210 	pte_l2_l_wc_mode = L2_B;
7211 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_xscale;
7212 
7213 	pte_l2_s_cache_mode = L2_B|L2_C;
7214 	pte_l2_s_wc_mode = L2_B;
7215 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_xscale;
7216 
7217 	pte_l1_s_cache_mode_pt = L1_S_C;
7218 	pte_l2_l_cache_mode_pt = L2_C;
7219 	pte_l2_s_cache_mode_pt = L2_C;
7220 
7221 #ifdef XSCALE_CACHE_READ_WRITE_ALLOCATE
7222 	/*
7223 	 * The XScale core has an enhanced mode where writes that
7224 	 * miss the cache cause a cache line to be allocated.  This
7225 	 * is significantly faster than the traditional, write-through
7226 	 * behavior of this case.
7227 	 */
7228 	pte_l1_s_cache_mode |= L1_S_XS_TEX(TEX_XSCALE_X);
7229 	pte_l2_l_cache_mode |= L2_XS_L_TEX(TEX_XSCALE_X);
7230 	pte_l2_s_cache_mode |= L2_XS_T_TEX(TEX_XSCALE_X);
7231 #endif /* XSCALE_CACHE_READ_WRITE_ALLOCATE */
7232 
7233 #ifdef XSCALE_CACHE_WRITE_THROUGH
7234 	/*
7235 	 * Some versions of the XScale core have various bugs in
7236 	 * their cache units, the work-around for which is to run
7237 	 * the cache in write-through mode.  Unfortunately, this
7238 	 * has a major (negative) impact on performance.  So, we
7239 	 * go ahead and run fast-and-loose, in the hopes that we
7240 	 * don't line up the planets in a way that will trip the
7241 	 * bugs.
7242 	 *
7243 	 * However, we give you the option to be slow-but-correct.
7244 	 */
7245 	write_through = 1;
7246 #elif defined(XSCALE_CACHE_WRITE_BACK)
7247 	/* force write back cache mode */
7248 	write_through = 0;
7249 #elif defined(CPU_XSCALE_PXA250) || defined(CPU_XSCALE_PXA270)
7250 	/*
7251 	 * Intel PXA2[15]0 processors are known to have a bug in
7252 	 * write-back cache on revision 4 and earlier (stepping
7253 	 * A[01] and B[012]).  Fixed for C0 and later.
7254 	 */
7255 	{
7256 		uint32_t id, type;
7257 
7258 		id = cpufunc_id();
7259 		type = id & ~(CPU_ID_XSCALE_COREREV_MASK|CPU_ID_REVISION_MASK);
7260 
7261 		if (type == CPU_ID_PXA250 || type == CPU_ID_PXA210) {
7262 			if ((id & CPU_ID_REVISION_MASK) < 5) {
7263 				/* write through for stepping A0-1 and B0-2 */
7264 				write_through = 1;
7265 			}
7266 		}
7267 	}
7268 #endif /* XSCALE_CACHE_WRITE_THROUGH */
7269 
7270 	if (write_through) {
7271 		pte_l1_s_cache_mode = L1_S_C;
7272 		pte_l2_l_cache_mode = L2_C;
7273 		pte_l2_s_cache_mode = L2_C;
7274 	}
7275 
7276 #if (ARM_NMMUS > 1)
7277 	xscale_use_minidata = 1;
7278 #endif
7279 
7280 	pte_l1_s_prot_u = L1_S_PROT_U_xscale;
7281 	pte_l1_s_prot_w = L1_S_PROT_W_xscale;
7282 	pte_l1_s_prot_ro = L1_S_PROT_RO_xscale;
7283 	pte_l1_s_prot_mask = L1_S_PROT_MASK_xscale;
7284 
7285 	pte_l2_s_prot_u = L2_S_PROT_U_xscale;
7286 	pte_l2_s_prot_w = L2_S_PROT_W_xscale;
7287 	pte_l2_s_prot_ro = L2_S_PROT_RO_xscale;
7288 	pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale;
7289 
7290 	pte_l2_l_prot_u = L2_L_PROT_U_xscale;
7291 	pte_l2_l_prot_w = L2_L_PROT_W_xscale;
7292 	pte_l2_l_prot_ro = L2_L_PROT_RO_xscale;
7293 	pte_l2_l_prot_mask = L2_L_PROT_MASK_xscale;
7294 
7295 	pte_l1_ss_proto = L1_SS_PROTO_xscale;
7296 	pte_l1_s_proto = L1_S_PROTO_xscale;
7297 	pte_l1_c_proto = L1_C_PROTO_xscale;
7298 	pte_l2_s_proto = L2_S_PROTO_xscale;
7299 
7300 	pmap_copy_page_func = pmap_copy_page_xscale;
7301 	pmap_zero_page_func = pmap_zero_page_xscale;
7302 
7303 	/*
7304 	 * Disable ECC protection of page table access, for now.
7305 	 */
7306 	auxctl = armreg_auxctl_read();
7307 	auxctl &= ~XSCALE_AUXCTL_P;
7308 	armreg_auxctl_write(auxctl);
7309 }
7310 
7311 /*
7312  * xscale_setup_minidata:
7313  *
7314  *	Set up the mini-data cache clean area.  We require the
7315  *	caller to allocate the right amount of physically and
7316  *	virtually contiguous space.
7317  */
7318 void
7319 xscale_setup_minidata(vaddr_t l1pt, vaddr_t va, paddr_t pa)
7320 {
7321 	extern vaddr_t xscale_minidata_clean_addr;
7322 	extern vsize_t xscale_minidata_clean_size; /* already initialized */
7323 	pd_entry_t *pde = (pd_entry_t *) l1pt;
7324 	vsize_t size;
7325 	uint32_t auxctl;
7326 
7327 	xscale_minidata_clean_addr = va;
7328 
7329 	/* Round it to page size. */
7330 	size = (xscale_minidata_clean_size + L2_S_OFFSET) & L2_S_FRAME;
7331 
7332 	for (; size != 0;
7333 	     va += L2_S_SIZE, pa += L2_S_SIZE, size -= L2_S_SIZE) {
7334 		const size_t l1slot = l1pte_index(va);
7335 		pt_entry_t *ptep = (pt_entry_t *) kernel_pt_lookup(l1pte_pa(pde[l1slot]));
7336 		if (ptep == NULL)
7337 			panic("xscale_setup_minidata: can't find L2 table for "
7338 			    "VA 0x%08lx", va);
7339 
7340 		ptep += l2pte_index(va);
7341 		pt_entry_t opte = *ptep;
7342 		l2pte_set(ptep,
7343 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, VM_PROT_READ)
7344 		    | L2_C | L2_XS_T_TEX(TEX_XSCALE_X), opte);
7345 	}
7346 
7347 	/*
7348 	 * Configure the mini-data cache for write-back with
7349 	 * read/write-allocate.
7350 	 *
7351 	 * NOTE: In order to reconfigure the mini-data cache, we must
7352 	 * make sure it contains no valid data!  In order to do that,
7353 	 * we must issue a global data cache invalidate command!
7354 	 *
7355 	 * WE ASSUME WE ARE RUNNING UN-CACHED WHEN THIS ROUTINE IS CALLED!
7356 	 * THIS IS VERY IMPORTANT!
7357 	 */
7358 
7359 	/* Invalidate data and mini-data. */
7360 	__asm volatile("mcr p15, 0, %0, c7, c6, 0" : : "r" (0));
7361 	auxctl = armreg_auxctl_read();
7362 	auxctl = (auxctl & ~XSCALE_AUXCTL_MD_MASK) | XSCALE_AUXCTL_MD_WB_RWA;
7363 	armreg_auxctl_write(auxctl);
7364 }
7365 
7366 /*
7367  * Change the PTEs for the specified kernel mappings such that they
7368  * will use the mini data cache instead of the main data cache.
7369  */
7370 void
7371 pmap_uarea(vaddr_t va)
7372 {
7373 	vaddr_t next_bucket, eva;
7374 
7375 #if (ARM_NMMUS > 1)
7376 	if (xscale_use_minidata == 0)
7377 		return;
7378 #endif
7379 
7380 	eva = va + USPACE;
7381 
7382 	while (va < eva) {
7383 		next_bucket = L2_NEXT_BUCKET_VA(va);
7384 		if (next_bucket > eva)
7385 			next_bucket = eva;
7386 
7387 		struct l2_bucket *l2b = pmap_get_l2_bucket(pmap_kernel(), va);
7388 		KDASSERT(l2b != NULL);
7389 
7390 		pt_entry_t * const sptep = &l2b->l2b_kva[l2pte_index(va)];
7391 		pt_entry_t *ptep = sptep;
7392 
7393 		while (va < next_bucket) {
7394 			const pt_entry_t opte = *ptep;
7395 			if (!l2pte_minidata_p(opte)) {
7396 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
7397 				cpu_tlb_flushD_SE(va);
7398 				l2pte_set(ptep, opte & ~L2_B, opte);
7399 			}
7400 			ptep += PAGE_SIZE / L2_S_SIZE;
7401 			va += PAGE_SIZE;
7402 		}
7403 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
7404 	}
7405 	cpu_cpwait();
7406 }
7407 #endif /* ARM_MMU_XSCALE == 1 */
7408 
7409 
7410 #if defined(CPU_ARM11MPCORE)
7411 
7412 void
7413 pmap_pte_init_arm11mpcore(void)
7414 {
7415 
7416 	/* cache mode is controlled by 5 bits (B, C, TEX) */
7417 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_armv6;
7418 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_armv6;
7419 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7420 	/* use extended small page (without APn, with TEX) */
7421 	pte_l2_s_cache_mask = L2_XS_CACHE_MASK_armv6;
7422 #else
7423 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_armv6c;
7424 #endif
7425 
7426 	/* write-back, write-allocate */
7427 	pte_l1_s_cache_mode = L1_S_C | L1_S_B | L1_S_V6_TEX(0x01);
7428 	pte_l2_l_cache_mode = L2_C | L2_B | L2_V6_L_TEX(0x01);
7429 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7430 	pte_l2_s_cache_mode = L2_C | L2_B | L2_V6_XS_TEX(0x01);
7431 #else
7432 	/* no TEX. read-allocate */
7433 	pte_l2_s_cache_mode = L2_C | L2_B;
7434 #endif
7435 	/*
7436 	 * write-back, write-allocate for page tables.
7437 	 */
7438 	pte_l1_s_cache_mode_pt = L1_S_C | L1_S_B | L1_S_V6_TEX(0x01);
7439 	pte_l2_l_cache_mode_pt = L2_C | L2_B | L2_V6_L_TEX(0x01);
7440 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7441 	pte_l2_s_cache_mode_pt = L2_C | L2_B | L2_V6_XS_TEX(0x01);
7442 #else
7443 	pte_l2_s_cache_mode_pt = L2_C | L2_B;
7444 #endif
7445 
7446 	pte_l1_s_prot_u = L1_S_PROT_U_armv6;
7447 	pte_l1_s_prot_w = L1_S_PROT_W_armv6;
7448 	pte_l1_s_prot_ro = L1_S_PROT_RO_armv6;
7449 	pte_l1_s_prot_mask = L1_S_PROT_MASK_armv6;
7450 
7451 #if defined(ARM11MPCORE_COMPAT_MMU) || defined(ARMV6_EXTENDED_SMALL_PAGE)
7452 	pte_l2_s_prot_u = L2_S_PROT_U_armv6n;
7453 	pte_l2_s_prot_w = L2_S_PROT_W_armv6n;
7454 	pte_l2_s_prot_ro = L2_S_PROT_RO_armv6n;
7455 	pte_l2_s_prot_mask = L2_S_PROT_MASK_armv6n;
7456 
7457 #else
7458 	/* with AP[0..3] */
7459 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
7460 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
7461 	pte_l2_s_prot_ro = L2_S_PROT_RO_generic;
7462 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
7463 #endif
7464 
7465 #ifdef	ARM11MPCORE_COMPAT_MMU
7466 	/* with AP[0..3] */
7467 	pte_l2_l_prot_u = L2_L_PROT_U_generic;
7468 	pte_l2_l_prot_w = L2_L_PROT_W_generic;
7469 	pte_l2_l_prot_ro = L2_L_PROT_RO_generic;
7470 	pte_l2_l_prot_mask = L2_L_PROT_MASK_generic;
7471 
7472 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
7473 	pte_l1_s_proto = L1_S_PROTO_armv6;
7474 	pte_l1_c_proto = L1_C_PROTO_armv6;
7475 	pte_l2_s_proto = L2_S_PROTO_armv6c;
7476 #else
7477 	pte_l2_l_prot_u = L2_L_PROT_U_armv6n;
7478 	pte_l2_l_prot_w = L2_L_PROT_W_armv6n;
7479 	pte_l2_l_prot_ro = L2_L_PROT_RO_armv6n;
7480 	pte_l2_l_prot_mask = L2_L_PROT_MASK_armv6n;
7481 
7482 	pte_l1_ss_proto = L1_SS_PROTO_armv6;
7483 	pte_l1_s_proto = L1_S_PROTO_armv6;
7484 	pte_l1_c_proto = L1_C_PROTO_armv6;
7485 	pte_l2_s_proto = L2_S_PROTO_armv6n;
7486 #endif
7487 
7488 	pmap_copy_page_func = pmap_copy_page_generic;
7489 	pmap_zero_page_func = pmap_zero_page_generic;
7490 	pmap_needs_pte_sync = 1;
7491 }
7492 #endif	/* CPU_ARM11MPCORE */
7493 
7494 
7495 #if ARM_MMU_V7 == 1
7496 void
7497 pmap_pte_init_armv7(void)
7498 {
7499 	/*
7500 	 * The ARMv7-A MMU is mostly compatible with generic. If the
7501 	 * AP field is zero, that now means "no access" rather than
7502 	 * read-only. The prototypes are a little different because of
7503 	 * the XN bit.
7504 	 */
7505 	pmap_pte_init_generic();
7506 
7507 	pmap_needs_pte_sync = 1;
7508 
7509 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_armv7;
7510 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_armv7;
7511 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_armv7;
7512 
7513 	/*
7514 	 * If the core support coherent walk then updates to translation tables
7515 	 * do not require a clean to the point of unification to ensure
7516 	 * visibility by subsequent translation table walks.  That means we can
7517 	 * map everything shareable and cached and the right thing will happen.
7518 	 */
7519         if (__SHIFTOUT(armreg_mmfr3_read(), __BITS(23,20))) {
7520 		pmap_needs_pte_sync = 0;
7521 
7522 		/*
7523 		 * write-back, no write-allocate, shareable for normal pages.
7524 		 */
7525 		pte_l1_s_cache_mode |= L1_S_V6_S;
7526 		pte_l2_l_cache_mode |= L2_XS_S;
7527 		pte_l2_s_cache_mode |= L2_XS_S;
7528 	}
7529 
7530 	/*
7531 	 * Page tables are just all other memory.  We can use write-back since
7532 	 * pmap_needs_pte_sync is 1 (or the MMU can read out of cache).
7533 	 */
7534 	pte_l1_s_cache_mode_pt = pte_l1_s_cache_mode;
7535 	pte_l2_l_cache_mode_pt = pte_l2_l_cache_mode;
7536 	pte_l2_s_cache_mode_pt = pte_l2_s_cache_mode;
7537 
7538 	/*
7539 	 * Check the Memory Model Features to see if this CPU supports
7540 	 * the TLBIASID coproc op.
7541 	 */
7542 	if (__SHIFTOUT(armreg_mmfr2_read(), __BITS(16,19)) >= 2) {
7543 		arm_has_tlbiasid_p = true;
7544 	}
7545 
7546 	pte_l1_s_prot_u = L1_S_PROT_U_armv7;
7547 	pte_l1_s_prot_w = L1_S_PROT_W_armv7;
7548 	pte_l1_s_prot_ro = L1_S_PROT_RO_armv7;
7549 	pte_l1_s_prot_mask = L1_S_PROT_MASK_armv7;
7550 
7551 	pte_l2_s_prot_u = L2_S_PROT_U_armv7;
7552 	pte_l2_s_prot_w = L2_S_PROT_W_armv7;
7553 	pte_l2_s_prot_ro = L2_S_PROT_RO_armv7;
7554 	pte_l2_s_prot_mask = L2_S_PROT_MASK_armv7;
7555 
7556 	pte_l2_l_prot_u = L2_L_PROT_U_armv7;
7557 	pte_l2_l_prot_w = L2_L_PROT_W_armv7;
7558 	pte_l2_l_prot_ro = L2_L_PROT_RO_armv7;
7559 	pte_l2_l_prot_mask = L2_L_PROT_MASK_armv7;
7560 
7561 	pte_l1_ss_proto = L1_SS_PROTO_armv7;
7562 	pte_l1_s_proto = L1_S_PROTO_armv7;
7563 	pte_l1_c_proto = L1_C_PROTO_armv7;
7564 	pte_l2_s_proto = L2_S_PROTO_armv7;
7565 
7566 }
7567 #endif /* ARM_MMU_V7 */
7568 
7569 /*
7570  * return the PA of the current L1 table, for use when handling a crash dump
7571  */
7572 uint32_t
7573 pmap_kernel_L1_addr(void)
7574 {
7575 #ifdef ARM_MMU_EXTENDED
7576 	return pmap_kernel()->pm_l1_pa;
7577 #else
7578 	return pmap_kernel()->pm_l1->l1_physaddr;
7579 #endif
7580 }
7581 
7582 #if defined(DDB)
7583 /*
7584  * A couple of ddb-callable functions for dumping pmaps
7585  */
7586 void pmap_dump(pmap_t);
7587 
7588 static pt_entry_t ncptes[64];
7589 static void pmap_dump_ncpg(pmap_t);
7590 
7591 void
7592 pmap_dump(pmap_t pm)
7593 {
7594 	struct l2_dtable *l2;
7595 	struct l2_bucket *l2b;
7596 	pt_entry_t *ptep, pte;
7597 	vaddr_t l2_va, l2b_va, va;
7598 	int i, j, k, occ, rows = 0;
7599 
7600 	if (pm == pmap_kernel())
7601 		printf("pmap_kernel (%p): ", pm);
7602 	else
7603 		printf("user pmap (%p): ", pm);
7604 
7605 #ifdef ARM_MMU_EXTENDED
7606 	printf("l1 at %p\n", pmap_l1_kva(pm));
7607 #else
7608 	printf("domain %d, l1 at %p\n", pmap_domain(pm), pmap_l1_kva(pm));
7609 #endif
7610 
7611 	l2_va = 0;
7612 	for (i = 0; i < L2_SIZE; i++, l2_va += 0x01000000) {
7613 		l2 = pm->pm_l2[i];
7614 
7615 		if (l2 == NULL || l2->l2_occupancy == 0)
7616 			continue;
7617 
7618 		l2b_va = l2_va;
7619 		for (j = 0; j < L2_BUCKET_SIZE; j++, l2b_va += 0x00100000) {
7620 			l2b = &l2->l2_bucket[j];
7621 
7622 			if (l2b->l2b_occupancy == 0 || l2b->l2b_kva == NULL)
7623 				continue;
7624 
7625 			ptep = l2b->l2b_kva;
7626 
7627 			for (k = 0; k < 256 && ptep[k] == 0; k++)
7628 				;
7629 
7630 			k &= ~63;
7631 			occ = l2b->l2b_occupancy;
7632 			va = l2b_va + (k * 4096);
7633 			for (; k < 256; k++, va += 0x1000) {
7634 				char ch = ' ';
7635 				if ((k % 64) == 0) {
7636 					if ((rows % 8) == 0) {
7637 						printf(
7638 "          |0000   |8000   |10000  |18000  |20000  |28000  |30000  |38000\n");
7639 					}
7640 					printf("%08lx: ", va);
7641 				}
7642 
7643 				ncptes[k & 63] = 0;
7644 				pte = ptep[k];
7645 				if (pte == 0) {
7646 					ch = '.';
7647 				} else {
7648 					occ--;
7649 					switch (pte & 0x0c) {
7650 					case 0x00:
7651 						ch = 'D'; /* No cache No buff */
7652 						break;
7653 					case 0x04:
7654 						ch = 'B'; /* No cache buff */
7655 						break;
7656 					case 0x08:
7657 						if (pte & 0x40)
7658 							ch = 'm';
7659 						else
7660 						   ch = 'C'; /* Cache No buff */
7661 						break;
7662 					case 0x0c:
7663 						ch = 'F'; /* Cache Buff */
7664 						break;
7665 					}
7666 
7667 					if ((pte & L2_S_PROT_U) == L2_S_PROT_U)
7668 						ch += 0x20;
7669 
7670 					if ((pte & 0xc) == 0)
7671 						ncptes[k & 63] = pte;
7672 				}
7673 
7674 				if ((k % 64) == 63) {
7675 					rows++;
7676 					printf("%c\n", ch);
7677 					pmap_dump_ncpg(pm);
7678 					if (occ == 0)
7679 						break;
7680 				} else
7681 					printf("%c", ch);
7682 			}
7683 		}
7684 	}
7685 }
7686 
7687 static void
7688 pmap_dump_ncpg(pmap_t pm)
7689 {
7690 	struct vm_page *pg;
7691 	struct vm_page_md *md;
7692 	struct pv_entry *pv;
7693 	int i;
7694 
7695 	for (i = 0; i < 63; i++) {
7696 		if (ncptes[i] == 0)
7697 			continue;
7698 
7699 		pg = PHYS_TO_VM_PAGE(l2pte_pa(ncptes[i]));
7700 		if (pg == NULL)
7701 			continue;
7702 		md = VM_PAGE_TO_MD(pg);
7703 
7704 		printf(" pa 0x%08lx: krw %d kro %d urw %d uro %d\n",
7705 		    VM_PAGE_TO_PHYS(pg),
7706 		    md->krw_mappings, md->kro_mappings,
7707 		    md->urw_mappings, md->uro_mappings);
7708 
7709 		SLIST_FOREACH(pv, &md->pvh_list, pv_link) {
7710 			printf("   %c va 0x%08lx, flags 0x%x\n",
7711 			    (pm == pv->pv_pmap) ? '*' : ' ',
7712 			    pv->pv_va, pv->pv_flags);
7713 		}
7714 	}
7715 }
7716 #endif
7717 
7718 #ifdef PMAP_STEAL_MEMORY
7719 void
7720 pmap_boot_pageadd(pv_addr_t *newpv)
7721 {
7722 	pv_addr_t *pv, *npv;
7723 
7724 	if ((pv = SLIST_FIRST(&pmap_boot_freeq)) != NULL) {
7725 		if (newpv->pv_pa < pv->pv_va) {
7726 			KASSERT(newpv->pv_pa + newpv->pv_size <= pv->pv_pa);
7727 			if (newpv->pv_pa + newpv->pv_size == pv->pv_pa) {
7728 				newpv->pv_size += pv->pv_size;
7729 				SLIST_REMOVE_HEAD(&pmap_boot_freeq, pv_list);
7730 			}
7731 			pv = NULL;
7732 		} else {
7733 			for (; (npv = SLIST_NEXT(pv, pv_list)) != NULL;
7734 			     pv = npv) {
7735 				KASSERT(pv->pv_pa + pv->pv_size < npv->pv_pa);
7736 				KASSERT(pv->pv_pa < newpv->pv_pa);
7737 				if (newpv->pv_pa > npv->pv_pa)
7738 					continue;
7739 				if (pv->pv_pa + pv->pv_size == newpv->pv_pa) {
7740 					pv->pv_size += newpv->pv_size;
7741 					return;
7742 				}
7743 				if (newpv->pv_pa + newpv->pv_size < npv->pv_pa)
7744 					break;
7745 				newpv->pv_size += npv->pv_size;
7746 				SLIST_INSERT_AFTER(pv, newpv, pv_list);
7747 				SLIST_REMOVE_AFTER(newpv, pv_list);
7748 				return;
7749 			}
7750 		}
7751 	}
7752 
7753 	if (pv) {
7754 		SLIST_INSERT_AFTER(pv, newpv, pv_list);
7755 	} else {
7756 		SLIST_INSERT_HEAD(&pmap_boot_freeq, newpv, pv_list);
7757 	}
7758 }
7759 
7760 void
7761 pmap_boot_pagealloc(psize_t amount, psize_t mask, psize_t match,
7762 	pv_addr_t *rpv)
7763 {
7764 	pv_addr_t *pv, **pvp;
7765 	struct vm_physseg *ps;
7766 	size_t i;
7767 
7768 	KASSERT(amount & PGOFSET);
7769 	KASSERT((mask & PGOFSET) == 0);
7770 	KASSERT((match & PGOFSET) == 0);
7771 	KASSERT(amount != 0);
7772 
7773 	for (pvp = &SLIST_FIRST(&pmap_boot_freeq);
7774 	     (pv = *pvp) != NULL;
7775 	     pvp = &SLIST_NEXT(pv, pv_list)) {
7776 		pv_addr_t *newpv;
7777 		psize_t off;
7778 		/*
7779 		 * If this entry is too small to satify the request...
7780 		 */
7781 		KASSERT(pv->pv_size > 0);
7782 		if (pv->pv_size < amount)
7783 			continue;
7784 
7785 		for (off = 0; off <= mask; off += PAGE_SIZE) {
7786 			if (((pv->pv_pa + off) & mask) == match
7787 			    && off + amount <= pv->pv_size)
7788 				break;
7789 		}
7790 		if (off > mask)
7791 			continue;
7792 
7793 		rpv->pv_va = pv->pv_va + off;
7794 		rpv->pv_pa = pv->pv_pa + off;
7795 		rpv->pv_size = amount;
7796 		pv->pv_size -= amount;
7797 		if (pv->pv_size == 0) {
7798 			KASSERT(off == 0);
7799 			KASSERT((vaddr_t) pv == rpv->pv_va);
7800 			*pvp = SLIST_NEXT(pv, pv_list);
7801 		} else if (off == 0) {
7802 			KASSERT((vaddr_t) pv == rpv->pv_va);
7803 			newpv = (pv_addr_t *) (rpv->pv_va + amount);
7804 			*newpv = *pv;
7805 			newpv->pv_pa += amount;
7806 			newpv->pv_va += amount;
7807 			*pvp = newpv;
7808 		} else if (off < pv->pv_size) {
7809 			newpv = (pv_addr_t *) (rpv->pv_va + amount);
7810 			*newpv = *pv;
7811 			newpv->pv_size -= off;
7812 			newpv->pv_pa += off + amount;
7813 			newpv->pv_va += off + amount;
7814 
7815 			SLIST_NEXT(pv, pv_list) = newpv;
7816 			pv->pv_size = off;
7817 		} else {
7818 			KASSERT((vaddr_t) pv != rpv->pv_va);
7819 		}
7820 		memset((void *)rpv->pv_va, 0, amount);
7821 		return;
7822 	}
7823 
7824 	if (vm_nphysseg == 0)
7825 		panic("pmap_boot_pagealloc: couldn't allocate memory");
7826 
7827 	for (pvp = &SLIST_FIRST(&pmap_boot_freeq);
7828 	     (pv = *pvp) != NULL;
7829 	     pvp = &SLIST_NEXT(pv, pv_list)) {
7830 		if (SLIST_NEXT(pv, pv_list) == NULL)
7831 			break;
7832 	}
7833 	KASSERT(mask == 0);
7834 	for (i = 0; i < vm_nphysseg; i++) {
7835 		ps = VM_PHYSMEM_PTR(i);
7836 		if (ps->avail_start == atop(pv->pv_pa + pv->pv_size)
7837 		    && pv->pv_va + pv->pv_size <= ptoa(ps->avail_end)) {
7838 			rpv->pv_va = pv->pv_va;
7839 			rpv->pv_pa = pv->pv_pa;
7840 			rpv->pv_size = amount;
7841 			*pvp = NULL;
7842 			pmap_map_chunk(kernel_l1pt.pv_va,
7843 			     ptoa(ps->avail_start) + (pv->pv_va - pv->pv_pa),
7844 			     ptoa(ps->avail_start),
7845 			     amount - pv->pv_size,
7846 			     VM_PROT_READ|VM_PROT_WRITE,
7847 			     PTE_CACHE);
7848 			ps->avail_start += atop(amount - pv->pv_size);
7849 			/*
7850 			 * If we consumed the entire physseg, remove it.
7851 			 */
7852 			if (ps->avail_start == ps->avail_end) {
7853 				for (--vm_nphysseg; i < vm_nphysseg; i++)
7854 					VM_PHYSMEM_PTR_SWAP(i, i + 1);
7855 			}
7856 			memset((void *)rpv->pv_va, 0, rpv->pv_size);
7857 			return;
7858 		}
7859 	}
7860 
7861 	panic("pmap_boot_pagealloc: couldn't allocate memory");
7862 }
7863 
7864 vaddr_t
7865 pmap_steal_memory(vsize_t size, vaddr_t *vstartp, vaddr_t *vendp)
7866 {
7867 	pv_addr_t pv;
7868 
7869 	pmap_boot_pagealloc(size, 0, 0, &pv);
7870 
7871 	return pv.pv_va;
7872 }
7873 #endif /* PMAP_STEAL_MEMORY */
7874 
7875 SYSCTL_SETUP(sysctl_machdep_pmap_setup, "sysctl machdep.kmpages setup")
7876 {
7877 	sysctl_createv(clog, 0, NULL, NULL,
7878 			CTLFLAG_PERMANENT,
7879 			CTLTYPE_NODE, "machdep", NULL,
7880 			NULL, 0, NULL, 0,
7881 			CTL_MACHDEP, CTL_EOL);
7882 
7883 	sysctl_createv(clog, 0, NULL, NULL,
7884 			CTLFLAG_PERMANENT,
7885 			CTLTYPE_INT, "kmpages",
7886 			SYSCTL_DESCR("count of pages allocated to kernel memory allocators"),
7887 			NULL, 0, &pmap_kmpages, 0,
7888 			CTL_MACHDEP, CTL_CREATE, CTL_EOL);
7889 }
7890 
7891 #ifdef PMAP_NEED_ALLOC_POOLPAGE
7892 struct vm_page *
7893 arm_pmap_alloc_poolpage(int flags)
7894 {
7895 	/*
7896 	 * On some systems, only some pages may be "coherent" for dma and we
7897 	 * want to prefer those for pool pages (think mbufs) but fallback to
7898 	 * any page if none is available.  But we can only fallback if we
7899 	 * aren't direct mapping memory or all of memory can be direct-mapped.
7900 	 * If that isn't true, pool changes can only come from direct-mapped
7901 	 * memory.
7902 	 */
7903 	if (arm_poolpage_vmfreelist != VM_FREELIST_DEFAULT) {
7904 		return uvm_pagealloc_strat(NULL, 0, NULL, flags,
7905 		    UVM_PGA_STRAT_FALLBACK,
7906 		    arm_poolpage_vmfreelist);
7907 	}
7908 
7909 	return uvm_pagealloc(NULL, 0, NULL, flags);
7910 }
7911 #endif
7912 
7913 #if defined(ARM_MMU_EXTENDED) && defined(MULTIPROCESSOR)
7914 void
7915 pmap_md_tlb_info_attach(struct pmap_tlb_info *ti, struct cpu_info *ci)
7916 {
7917         /* nothing */
7918 }
7919 
7920 int
7921 pic_ipi_shootdown(void *arg)
7922 {
7923 #if PMAP_TLB_NEED_SHOOTDOWN
7924 	pmap_tlb_shootdown_process();
7925 #endif
7926 	return 1;
7927 }
7928 #endif /* ARM_MMU_EXTENDED && MULTIPROCESSOR */
7929 
7930 
7931 #ifdef __HAVE_MM_MD_DIRECT_MAPPED_PHYS
7932 vaddr_t
7933 pmap_direct_mapped_phys(paddr_t pa, bool *ok_p, vaddr_t va)
7934 {
7935 	bool ok = false;
7936 	if (physical_start <= pa && pa < physical_end) {
7937 #ifdef KERNEL_BASE_VOFFSET
7938 		const vaddr_t newva = pa + KERNEL_BASE_VOFFSET;
7939 #else
7940 		const vaddr_t newva = KERNEL_BASE + pa - physical_start;
7941 #endif
7942 #ifdef ARM_MMU_EXTENDED
7943 		if (newva >= KERNEL_BASE && newva < pmap_directlimit) {
7944 #endif
7945 			va = newva;
7946 			ok = true;
7947 #ifdef ARM_MMU_EXTENDED
7948 		}
7949 #endif
7950 	}
7951 	KASSERT(ok_p);
7952 	*ok_p = ok;
7953 	return va;
7954 }
7955 
7956 vaddr_t
7957 pmap_map_poolpage(paddr_t pa)
7958 {
7959 	bool ok __diagused;
7960 	vaddr_t va = pmap_direct_mapped_phys(pa, &ok, 0);
7961 	KASSERTMSG(ok, "pa %#lx not direct mappable", pa);
7962 #if defined(PMAP_CACHE_VIPT) && !defined(ARM_MMU_EXTENDED)
7963 	if (arm_cache_prefer_mask != 0) {
7964 		struct vm_page * const pg = PHYS_TO_VM_PAGE(pa);
7965 		struct vm_page_md * const md = VM_PAGE_TO_MD(pg);
7966 		pmap_acquire_page_lock(md);
7967 		pmap_vac_me_harder(md, pa, pmap_kernel(), va);
7968 		pmap_release_page_lock(md);
7969 	}
7970 #endif
7971 	return va;
7972 }
7973 
7974 paddr_t
7975 pmap_unmap_poolpage(vaddr_t va)
7976 {
7977 	KASSERT(va >= KERNEL_BASE);
7978 #ifdef PMAP_CACHE_VIVT
7979 	cpu_idcache_wbinv_range(va, PAGE_SIZE);
7980 #endif
7981 #if defined(KERNEL_BASE_VOFFSET)
7982         return va - KERNEL_BASE_VOFFSET;
7983 #else
7984         return va - KERNEL_BASE + physical_start;
7985 #endif
7986 }
7987 #endif /* __HAVE_MM_MD_DIRECT_MAPPED_PHYS */
7988