xref: /netbsd-src/sys/arch/arm/arm32/pmap.c (revision 7fa608457b817eca6e0977b37f758ae064f3c99c)
1 /*	$NetBSD: pmap.c,v 1.169 2007/11/08 11:10:28 matt 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  * 3. All advertising materials mentioning features or use of this software
82  *    must display the following acknowledgement:
83  *        This product includes software developed by the NetBSD
84  *        Foundation, Inc. and its contributors.
85  * 4. Neither the name of The NetBSD Foundation nor the names of its
86  *    contributors may be used to endorse or promote products derived
87  *    from this software without specific prior written permission.
88  *
89  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
90  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
91  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
92  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
93  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
94  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
95  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
96  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
97  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
98  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
99  * POSSIBILITY OF SUCH DAMAGE.
100  */
101 
102 /*
103  * Copyright (c) 1994-1998 Mark Brinicombe.
104  * Copyright (c) 1994 Brini.
105  * All rights reserved.
106  *
107  * This code is derived from software written for Brini by Mark Brinicombe
108  *
109  * Redistribution and use in source and binary forms, with or without
110  * modification, are permitted provided that the following conditions
111  * are met:
112  * 1. Redistributions of source code must retain the above copyright
113  *    notice, this list of conditions and the following disclaimer.
114  * 2. Redistributions in binary form must reproduce the above copyright
115  *    notice, this list of conditions and the following disclaimer in the
116  *    documentation and/or other materials provided with the distribution.
117  * 3. All advertising materials mentioning features or use of this software
118  *    must display the following acknowledgement:
119  *	This product includes software developed by Mark Brinicombe.
120  * 4. The name of the author may not be used to endorse or promote products
121  *    derived from this software without specific prior written permission.
122  *
123  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
124  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
125  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
126  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
127  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
128  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
129  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
130  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
131  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
132  *
133  * RiscBSD kernel project
134  *
135  * pmap.c
136  *
137  * Machine dependant vm stuff
138  *
139  * Created      : 20/09/94
140  */
141 
142 /*
143  * Performance improvements, UVM changes, overhauls and part-rewrites
144  * were contributed by Neil A. Carson <neil@causality.com>.
145  */
146 
147 /*
148  * Overhauled again to speedup the pmap, use MMU Domains so that L1 tables
149  * can be shared, and re-work the KVM layout, by Steve Woodford of Wasabi
150  * Systems, Inc.
151  *
152  * There are still a few things outstanding at this time:
153  *
154  *   - There are some unresolved issues for MP systems:
155  *
156  *     o The L1 metadata needs a lock, or more specifically, some places
157  *       need to acquire an exclusive lock when modifying L1 translation
158  *       table entries.
159  *
160  *     o When one cpu modifies an L1 entry, and that L1 table is also
161  *       being used by another cpu, then the latter will need to be told
162  *       that a tlb invalidation may be necessary. (But only if the old
163  *       domain number in the L1 entry being over-written is currently
164  *       the active domain on that cpu). I guess there are lots more tlb
165  *       shootdown issues too...
166  *
167  *     o If the vector_page is at 0x00000000 instead of 0xffff0000, then
168  *       MP systems will lose big-time because of the MMU domain hack.
169  *       The only way this can be solved (apart from moving the vector
170  *       page to 0xffff0000) is to reserve the first 1MB of user address
171  *       space for kernel use only. This would require re-linking all
172  *       applications so that the text section starts above this 1MB
173  *       boundary.
174  *
175  *     o Tracking which VM space is resident in the cache/tlb has not yet
176  *       been implemented for MP systems.
177  *
178  *     o Finally, there is a pathological condition where two cpus running
179  *       two separate processes (not lwps) which happen to share an L1
180  *       can get into a fight over one or more L1 entries. This will result
181  *       in a significant slow-down if both processes are in tight loops.
182  */
183 
184 /*
185  * Special compilation symbols
186  * PMAP_DEBUG		- Build in pmap_debug_level code
187  */
188 
189 /* Include header files */
190 
191 #include "opt_cpuoptions.h"
192 #include "opt_pmap_debug.h"
193 #include "opt_ddb.h"
194 #include "opt_lockdebug.h"
195 #include "opt_multiprocessor.h"
196 
197 #include <sys/types.h>
198 #include <sys/param.h>
199 #include <sys/kernel.h>
200 #include <sys/systm.h>
201 #include <sys/proc.h>
202 #include <sys/malloc.h>
203 #include <sys/user.h>
204 #include <sys/pool.h>
205 #include <sys/cdefs.h>
206 
207 #include <uvm/uvm.h>
208 
209 #include <machine/bus.h>
210 #include <machine/pmap.h>
211 #include <machine/pcb.h>
212 #include <machine/param.h>
213 #include <arm/arm32/katelib.h>
214 
215 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.169 2007/11/08 11:10:28 matt Exp $");
216 
217 #ifdef PMAP_DEBUG
218 
219 /* XXX need to get rid of all refs to this */
220 int pmap_debug_level = 0;
221 
222 /*
223  * for switching to potentially finer grained debugging
224  */
225 #define	PDB_FOLLOW	0x0001
226 #define	PDB_INIT	0x0002
227 #define	PDB_ENTER	0x0004
228 #define	PDB_REMOVE	0x0008
229 #define	PDB_CREATE	0x0010
230 #define	PDB_PTPAGE	0x0020
231 #define	PDB_GROWKERN	0x0040
232 #define	PDB_BITS	0x0080
233 #define	PDB_COLLECT	0x0100
234 #define	PDB_PROTECT	0x0200
235 #define	PDB_MAP_L1	0x0400
236 #define	PDB_BOOTSTRAP	0x1000
237 #define	PDB_PARANOIA	0x2000
238 #define	PDB_WIRING	0x4000
239 #define	PDB_PVDUMP	0x8000
240 #define	PDB_VAC		0x10000
241 #define	PDB_KENTER	0x20000
242 #define	PDB_KREMOVE	0x40000
243 
244 int debugmap = 1;
245 int pmapdebug = 0;
246 #define	NPDEBUG(_lev_,_stat_) \
247 	if (pmapdebug & (_lev_)) \
248         	((_stat_))
249 
250 #else	/* PMAP_DEBUG */
251 #define NPDEBUG(_lev_,_stat_) /* Nothing */
252 #endif	/* PMAP_DEBUG */
253 
254 /*
255  * pmap_kernel() points here
256  */
257 struct pmap     kernel_pmap_store;
258 
259 /*
260  * Which pmap is currently 'live' in the cache
261  *
262  * XXXSCW: Fix for SMP ...
263  */
264 static pmap_t pmap_recent_user;
265 
266 /*
267  * Pool and cache that pmap structures are allocated from.
268  * We use a cache to avoid clearing the pm_l2[] array (1KB)
269  * in pmap_create().
270  */
271 static struct pool_cache pmap_cache;
272 static LIST_HEAD(, pmap) pmap_pmaps;
273 
274 /*
275  * Pool of PV structures
276  */
277 static struct pool pmap_pv_pool;
278 static void *pmap_bootstrap_pv_page_alloc(struct pool *, int);
279 static void pmap_bootstrap_pv_page_free(struct pool *, void *);
280 static struct pool_allocator pmap_bootstrap_pv_allocator = {
281 	pmap_bootstrap_pv_page_alloc, pmap_bootstrap_pv_page_free
282 };
283 
284 /*
285  * Pool and cache of l2_dtable structures.
286  * We use a cache to avoid clearing the structures when they're
287  * allocated. (196 bytes)
288  */
289 static struct pool_cache pmap_l2dtable_cache;
290 static vaddr_t pmap_kernel_l2dtable_kva;
291 
292 /*
293  * Pool and cache of L2 page descriptors.
294  * We use a cache to avoid clearing the descriptor table
295  * when they're allocated. (1KB)
296  */
297 static struct pool_cache pmap_l2ptp_cache;
298 static vaddr_t pmap_kernel_l2ptp_kva;
299 static paddr_t pmap_kernel_l2ptp_phys;
300 
301 /*
302  * pmap copy/zero page, and mem(5) hook point
303  */
304 static pt_entry_t *csrc_pte, *cdst_pte;
305 static vaddr_t csrcp, cdstp;
306 char *memhook;
307 extern void *msgbufaddr;
308 
309 /*
310  * Flag to indicate if pmap_init() has done its thing
311  */
312 bool pmap_initialized;
313 
314 /*
315  * Misc. locking data structures
316  */
317 
318 #if 0 /* defined(MULTIPROCESSOR) || defined(LOCKDEBUG) */
319 static struct lock pmap_main_lock;
320 
321 #define PMAP_MAP_TO_HEAD_LOCK() \
322      (void) spinlockmgr(&pmap_main_lock, LK_SHARED, NULL)
323 #define PMAP_MAP_TO_HEAD_UNLOCK() \
324      (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
325 #define PMAP_HEAD_TO_MAP_LOCK() \
326      (void) spinlockmgr(&pmap_main_lock, LK_EXCLUSIVE, NULL)
327 #define PMAP_HEAD_TO_MAP_UNLOCK() \
328      spinlockmgr(&pmap_main_lock, LK_RELEASE, (void *) 0)
329 #else
330 #define PMAP_MAP_TO_HEAD_LOCK()		/* null */
331 #define PMAP_MAP_TO_HEAD_UNLOCK()	/* null */
332 #define PMAP_HEAD_TO_MAP_LOCK()		/* null */
333 #define PMAP_HEAD_TO_MAP_UNLOCK()	/* null */
334 #endif
335 
336 #define	pmap_acquire_pmap_lock(pm)			\
337 	do {						\
338 		if ((pm) != pmap_kernel())		\
339 			simple_lock(&(pm)->pm_lock);	\
340 	} while (/*CONSTCOND*/0)
341 
342 #define	pmap_release_pmap_lock(pm)			\
343 	do {						\
344 		if ((pm) != pmap_kernel())		\
345 			simple_unlock(&(pm)->pm_lock);	\
346 	} while (/*CONSTCOND*/0)
347 
348 
349 /*
350  * Metadata for L1 translation tables.
351  */
352 struct l1_ttable {
353 	/* Entry on the L1 Table list */
354 	SLIST_ENTRY(l1_ttable) l1_link;
355 
356 	/* Entry on the L1 Least Recently Used list */
357 	TAILQ_ENTRY(l1_ttable) l1_lru;
358 
359 	/* Track how many domains are allocated from this L1 */
360 	volatile u_int l1_domain_use_count;
361 
362 	/*
363 	 * A free-list of domain numbers for this L1.
364 	 * We avoid using ffs() and a bitmap to track domains since ffs()
365 	 * is slow on ARM.
366 	 */
367 	u_int8_t l1_domain_first;
368 	u_int8_t l1_domain_free[PMAP_DOMAINS];
369 
370 	/* Physical address of this L1 page table */
371 	paddr_t l1_physaddr;
372 
373 	/* KVA of this L1 page table */
374 	pd_entry_t *l1_kva;
375 };
376 
377 /*
378  * Convert a virtual address into its L1 table index. That is, the
379  * index used to locate the L2 descriptor table pointer in an L1 table.
380  * This is basically used to index l1->l1_kva[].
381  *
382  * Each L2 descriptor table represents 1MB of VA space.
383  */
384 #define	L1_IDX(va)		(((vaddr_t)(va)) >> L1_S_SHIFT)
385 
386 /*
387  * L1 Page Tables are tracked using a Least Recently Used list.
388  *  - New L1s are allocated from the HEAD.
389  *  - Freed L1s are added to the TAIl.
390  *  - Recently accessed L1s (where an 'access' is some change to one of
391  *    the userland pmaps which owns this L1) are moved to the TAIL.
392  */
393 static TAILQ_HEAD(, l1_ttable) l1_lru_list;
394 static struct simplelock l1_lru_lock;
395 
396 /*
397  * A list of all L1 tables
398  */
399 static SLIST_HEAD(, l1_ttable) l1_list;
400 
401 /*
402  * The l2_dtable tracks L2_BUCKET_SIZE worth of L1 slots.
403  *
404  * This is normally 16MB worth L2 page descriptors for any given pmap.
405  * Reference counts are maintained for L2 descriptors so they can be
406  * freed when empty.
407  */
408 struct l2_dtable {
409 	/* The number of L2 page descriptors allocated to this l2_dtable */
410 	u_int l2_occupancy;
411 
412 	/* List of L2 page descriptors */
413 	struct l2_bucket {
414 		pt_entry_t *l2b_kva;	/* KVA of L2 Descriptor Table */
415 		paddr_t l2b_phys;	/* Physical address of same */
416 		u_short l2b_l1idx;	/* This L2 table's L1 index */
417 		u_short l2b_occupancy;	/* How many active descriptors */
418 	} l2_bucket[L2_BUCKET_SIZE];
419 };
420 
421 /*
422  * Given an L1 table index, calculate the corresponding l2_dtable index
423  * and bucket index within the l2_dtable.
424  */
425 #define	L2_IDX(l1idx)		(((l1idx) >> L2_BUCKET_LOG2) & \
426 				 (L2_SIZE - 1))
427 #define	L2_BUCKET(l1idx)	((l1idx) & (L2_BUCKET_SIZE - 1))
428 
429 /*
430  * Given a virtual address, this macro returns the
431  * virtual address required to drop into the next L2 bucket.
432  */
433 #define	L2_NEXT_BUCKET(va)	(((va) & L1_S_FRAME) + L1_S_SIZE)
434 
435 /*
436  * L2 allocation.
437  */
438 #define	pmap_alloc_l2_dtable()		\
439 	    pool_cache_get(&pmap_l2dtable_cache, PR_NOWAIT)
440 #define	pmap_free_l2_dtable(l2)		\
441 	    pool_cache_put(&pmap_l2dtable_cache, (l2))
442 #define pmap_alloc_l2_ptp(pap)		\
443 	    ((pt_entry_t *)pool_cache_get_paddr(&pmap_l2ptp_cache,\
444 	    PR_NOWAIT, (pap)))
445 
446 /*
447  * We try to map the page tables write-through, if possible.  However, not
448  * all CPUs have a write-through cache mode, so on those we have to sync
449  * the cache when we frob page tables.
450  *
451  * We try to evaluate this at compile time, if possible.  However, it's
452  * not always possible to do that, hence this run-time var.
453  */
454 int	pmap_needs_pte_sync;
455 
456 /*
457  * Real definition of pv_entry.
458  */
459 struct pv_entry {
460 	struct pv_entry *pv_next;       /* next pv_entry */
461 	pmap_t		pv_pmap;        /* pmap where mapping lies */
462 	vaddr_t		pv_va;          /* virtual address for mapping */
463 	u_int		pv_flags;       /* flags */
464 };
465 
466 /*
467  * Macro to determine if a mapping might be resident in the
468  * instruction cache and/or TLB
469  */
470 #define	PV_BEEN_EXECD(f)  (((f) & (PVF_REF | PVF_EXEC)) == (PVF_REF | PVF_EXEC))
471 
472 /*
473  * Macro to determine if a mapping might be resident in the
474  * data cache and/or TLB
475  */
476 #define	PV_BEEN_REFD(f)   (((f) & PVF_REF) != 0)
477 
478 /*
479  * Local prototypes
480  */
481 static int		pmap_set_pt_cache_mode(pd_entry_t *, vaddr_t);
482 static void		pmap_alloc_specials(vaddr_t *, int, vaddr_t *,
483 			    pt_entry_t **);
484 static bool		pmap_is_current(pmap_t);
485 static bool		pmap_is_cached(pmap_t);
486 static void		pmap_enter_pv(struct vm_page *, struct pv_entry *,
487 			    pmap_t, vaddr_t, u_int);
488 static struct pv_entry *pmap_find_pv(struct vm_page *, pmap_t, vaddr_t);
489 static struct pv_entry *pmap_remove_pv(struct vm_page *, pmap_t, vaddr_t, int);
490 static u_int		pmap_modify_pv(struct vm_page *, pmap_t, vaddr_t,
491 			    u_int, u_int);
492 
493 static void		pmap_pinit(pmap_t);
494 static int		pmap_pmap_ctor(void *, void *, int);
495 
496 static void		pmap_alloc_l1(pmap_t);
497 static void		pmap_free_l1(pmap_t);
498 static void		pmap_use_l1(pmap_t);
499 
500 static struct l2_bucket *pmap_get_l2_bucket(pmap_t, vaddr_t);
501 static struct l2_bucket *pmap_alloc_l2_bucket(pmap_t, vaddr_t);
502 static void		pmap_free_l2_bucket(pmap_t, struct l2_bucket *, u_int);
503 static int		pmap_l2ptp_ctor(void *, void *, int);
504 static int		pmap_l2dtable_ctor(void *, void *, int);
505 
506 static void		pmap_vac_me_harder(struct vm_page *, pmap_t, vaddr_t);
507 static void		pmap_vac_me_kpmap(struct vm_page *, pmap_t, vaddr_t);
508 static void		pmap_vac_me_user(struct vm_page *, pmap_t, vaddr_t);
509 
510 static void		pmap_clearbit(struct vm_page *, u_int);
511 static int		pmap_clean_page(struct pv_entry *, bool);
512 static void		pmap_page_remove(struct vm_page *);
513 
514 static void		pmap_init_l1(struct l1_ttable *, pd_entry_t *);
515 static vaddr_t		kernel_pt_lookup(paddr_t);
516 
517 void pmap_switch(struct lwp *, struct lwp *);
518 
519 
520 /*
521  * External function prototypes
522  */
523 extern void bzero_page(vaddr_t);
524 extern void bcopy_page(vaddr_t, vaddr_t);
525 
526 /*
527  * Misc variables
528  */
529 vaddr_t virtual_avail;
530 vaddr_t virtual_end;
531 vaddr_t pmap_curmaxkvaddr;
532 
533 vaddr_t avail_start;
534 vaddr_t avail_end;
535 
536 extern pv_addr_t systempage;
537 
538 /* Function to set the debug level of the pmap code */
539 
540 #ifdef PMAP_DEBUG
541 void
542 pmap_debug(int level)
543 {
544 	pmap_debug_level = level;
545 	printf("pmap_debug: level=%d\n", pmap_debug_level);
546 }
547 #endif	/* PMAP_DEBUG */
548 
549 /*
550  * A bunch of routines to conditionally flush the caches/TLB depending
551  * on whether the specified pmap actually needs to be flushed at any
552  * given time.
553  */
554 static inline void
555 pmap_tlb_flushID_SE(pmap_t pm, vaddr_t va)
556 {
557 
558 	if (pm->pm_cstate.cs_tlb_id)
559 		cpu_tlb_flushID_SE(va);
560 }
561 
562 static inline void
563 pmap_tlb_flushD_SE(pmap_t pm, vaddr_t va)
564 {
565 
566 	if (pm->pm_cstate.cs_tlb_d)
567 		cpu_tlb_flushD_SE(va);
568 }
569 
570 static inline void
571 pmap_tlb_flushID(pmap_t pm)
572 {
573 
574 	if (pm->pm_cstate.cs_tlb_id) {
575 		cpu_tlb_flushID();
576 		pm->pm_cstate.cs_tlb = 0;
577 	}
578 }
579 
580 static inline void
581 pmap_tlb_flushD(pmap_t pm)
582 {
583 
584 	if (pm->pm_cstate.cs_tlb_d) {
585 		cpu_tlb_flushD();
586 		pm->pm_cstate.cs_tlb_d = 0;
587 	}
588 }
589 
590 static inline void
591 pmap_idcache_wbinv_range(pmap_t pm, vaddr_t va, vsize_t len)
592 {
593 
594 	if (pm->pm_cstate.cs_cache_id)
595 		cpu_idcache_wbinv_range(va, len);
596 }
597 
598 static inline void
599 pmap_dcache_wb_range(pmap_t pm, vaddr_t va, vsize_t len,
600     bool do_inv, bool rd_only)
601 {
602 
603 	if (pm->pm_cstate.cs_cache_d) {
604 		if (do_inv) {
605 			if (rd_only)
606 				cpu_dcache_inv_range(va, len);
607 			else
608 				cpu_dcache_wbinv_range(va, len);
609 		} else
610 		if (!rd_only)
611 			cpu_dcache_wb_range(va, len);
612 	}
613 }
614 
615 static inline void
616 pmap_idcache_wbinv_all(pmap_t pm)
617 {
618 
619 	if (pm->pm_cstate.cs_cache_id) {
620 		cpu_idcache_wbinv_all();
621 		pm->pm_cstate.cs_cache = 0;
622 	}
623 }
624 
625 static inline void
626 pmap_dcache_wbinv_all(pmap_t pm)
627 {
628 
629 	if (pm->pm_cstate.cs_cache_d) {
630 		cpu_dcache_wbinv_all();
631 		pm->pm_cstate.cs_cache_d = 0;
632 	}
633 }
634 
635 static inline bool
636 pmap_is_current(pmap_t pm)
637 {
638 
639 	if (pm == pmap_kernel() ||
640 	    (curproc && curproc->p_vmspace->vm_map.pmap == pm))
641 		return (true);
642 
643 	return (false);
644 }
645 
646 static inline bool
647 pmap_is_cached(pmap_t pm)
648 {
649 
650 	if (pm == pmap_kernel() || pmap_recent_user == NULL ||
651 	    pmap_recent_user == pm)
652 		return (true);
653 
654 	return (false);
655 }
656 
657 /*
658  * PTE_SYNC_CURRENT:
659  *
660  *     Make sure the pte is written out to RAM.
661  *     We need to do this for one of two cases:
662  *       - We're dealing with the kernel pmap
663  *       - There is no pmap active in the cache/tlb.
664  *       - The specified pmap is 'active' in the cache/tlb.
665  */
666 #ifdef PMAP_INCLUDE_PTE_SYNC
667 #define	PTE_SYNC_CURRENT(pm, ptep)	\
668 do {					\
669 	if (PMAP_NEEDS_PTE_SYNC && 	\
670 	    pmap_is_cached(pm))		\
671 		PTE_SYNC(ptep);		\
672 } while (/*CONSTCOND*/0)
673 #else
674 #define	PTE_SYNC_CURRENT(pm, ptep)	/* nothing */
675 #endif
676 
677 /*
678  * main pv_entry manipulation functions:
679  *   pmap_enter_pv: enter a mapping onto a vm_page list
680  *   pmap_remove_pv: remove a mappiing from a vm_page list
681  *
682  * NOTE: pmap_enter_pv expects to lock the pvh itself
683  *       pmap_remove_pv expects te caller to lock the pvh before calling
684  */
685 
686 /*
687  * pmap_enter_pv: enter a mapping onto a vm_page lst
688  *
689  * => caller should hold the proper lock on pmap_main_lock
690  * => caller should have pmap locked
691  * => we will gain the lock on the vm_page and allocate the new pv_entry
692  * => caller should adjust ptp's wire_count before calling
693  * => caller should not adjust pmap's wire_count
694  */
695 static void
696 pmap_enter_pv(struct vm_page *pg, struct pv_entry *pve, pmap_t pm,
697     vaddr_t va, u_int flags)
698 {
699 
700 	NPDEBUG(PDB_PVDUMP,
701 	    printf("pmap_enter_pv: pm %p, pg %p, flags 0x%x\n", pm, pg, flags));
702 
703 	pve->pv_pmap = pm;
704 	pve->pv_va = va;
705 	pve->pv_flags = flags;
706 
707 	simple_lock(&pg->mdpage.pvh_slock);	/* lock vm_page */
708 	pve->pv_next = pg->mdpage.pvh_list;	/* add to ... */
709 	pg->mdpage.pvh_list = pve;		/* ... locked list */
710 	pg->mdpage.pvh_attrs |= flags & (PVF_REF | PVF_MOD);
711 	if (pm == pmap_kernel()) {
712 		if (flags & PVF_WRITE)
713 			pg->mdpage.krw_mappings++;
714 		else
715 			pg->mdpage.kro_mappings++;
716 	} else
717 	if (flags & PVF_WRITE)
718 		pg->mdpage.urw_mappings++;
719 	else
720 		pg->mdpage.uro_mappings++;
721 	simple_unlock(&pg->mdpage.pvh_slock);	/* unlock, done! */
722 
723 	if (pve->pv_flags & PVF_WIRED)
724 		++pm->pm_stats.wired_count;
725 }
726 
727 /*
728  *
729  * pmap_find_pv: Find a pv entry
730  *
731  * => caller should hold lock on vm_page
732  */
733 static inline struct pv_entry *
734 pmap_find_pv(struct vm_page *pg, pmap_t pm, vaddr_t va)
735 {
736 	struct pv_entry *pv;
737 
738 	for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
739 		if (pm == pv->pv_pmap && va == pv->pv_va)
740 			break;
741 	}
742 
743 	return (pv);
744 }
745 
746 /*
747  * pmap_remove_pv: try to remove a mapping from a pv_list
748  *
749  * => caller should hold proper lock on pmap_main_lock
750  * => pmap should be locked
751  * => caller should hold lock on vm_page [so that attrs can be adjusted]
752  * => caller should adjust ptp's wire_count and free PTP if needed
753  * => caller should NOT adjust pmap's wire_count
754  * => we return the removed pve
755  */
756 static struct pv_entry *
757 pmap_remove_pv(struct vm_page *pg, pmap_t pm, vaddr_t va, int skip_wired)
758 {
759 	struct pv_entry *pve, **prevptr;
760 
761 	NPDEBUG(PDB_PVDUMP,
762 	    printf("pmap_remove_pv: pm %p, pg %p, va 0x%08lx\n", pm, pg, va));
763 
764 	prevptr = &pg->mdpage.pvh_list;		/* previous pv_entry pointer */
765 	pve = *prevptr;
766 
767 	while (pve) {
768 		if (pve->pv_pmap == pm && pve->pv_va == va) {	/* match? */
769 			NPDEBUG(PDB_PVDUMP, printf("pmap_remove_pv: pm %p, pg "
770 			    "%p, flags 0x%x\n", pm, pg, pve->pv_flags));
771 			if (pve->pv_flags & PVF_WIRED) {
772 				if (skip_wired)
773 					return (NULL);
774 				--pm->pm_stats.wired_count;
775 			}
776 			*prevptr = pve->pv_next;		/* remove it! */
777 			if (pm == pmap_kernel()) {
778 				if (pve->pv_flags & PVF_WRITE)
779 					pg->mdpage.krw_mappings--;
780 				else
781 					pg->mdpage.kro_mappings--;
782 			} else
783 			if (pve->pv_flags & PVF_WRITE)
784 				pg->mdpage.urw_mappings--;
785 			else
786 				pg->mdpage.uro_mappings--;
787 			break;
788 		}
789 		prevptr = &pve->pv_next;		/* previous pointer */
790 		pve = pve->pv_next;			/* advance */
791 	}
792 
793 	return(pve);				/* return removed pve */
794 }
795 
796 /*
797  *
798  * pmap_modify_pv: Update pv flags
799  *
800  * => caller should hold lock on vm_page [so that attrs can be adjusted]
801  * => caller should NOT adjust pmap's wire_count
802  * => caller must call pmap_vac_me_harder() if writable status of a page
803  *    may have changed.
804  * => we return the old flags
805  *
806  * Modify a physical-virtual mapping in the pv table
807  */
808 static u_int
809 pmap_modify_pv(struct vm_page *pg, pmap_t pm, vaddr_t va,
810     u_int clr_mask, u_int set_mask)
811 {
812 	struct pv_entry *npv;
813 	u_int flags, oflags;
814 
815 	if ((npv = pmap_find_pv(pg, pm, va)) == NULL)
816 		return (0);
817 
818 	NPDEBUG(PDB_PVDUMP,
819 	    printf("pmap_modify_pv: pm %p, pg %p, clr 0x%x, set 0x%x, flags 0x%x\n", pm, pg, clr_mask, set_mask, npv->pv_flags));
820 
821 	/*
822 	 * There is at least one VA mapping this page.
823 	 */
824 
825 	if (clr_mask & (PVF_REF | PVF_MOD))
826 		pg->mdpage.pvh_attrs |= set_mask & (PVF_REF | PVF_MOD);
827 
828 	oflags = npv->pv_flags;
829 	npv->pv_flags = flags = (oflags & ~clr_mask) | set_mask;
830 
831 	if ((flags ^ oflags) & PVF_WIRED) {
832 		if (flags & PVF_WIRED)
833 			++pm->pm_stats.wired_count;
834 		else
835 			--pm->pm_stats.wired_count;
836 	}
837 
838 	if ((flags ^ oflags) & PVF_WRITE) {
839 		if (pm == pmap_kernel()) {
840 			if (flags & PVF_WRITE) {
841 				pg->mdpage.krw_mappings++;
842 				pg->mdpage.kro_mappings--;
843 			} else {
844 				pg->mdpage.kro_mappings++;
845 				pg->mdpage.krw_mappings--;
846 			}
847 		} else
848 		if (flags & PVF_WRITE) {
849 			pg->mdpage.urw_mappings++;
850 			pg->mdpage.uro_mappings--;
851 		} else {
852 			pg->mdpage.uro_mappings++;
853 			pg->mdpage.urw_mappings--;
854 		}
855 	}
856 
857 	return (oflags);
858 }
859 
860 /*
861  * Allocate an L1 translation table for the specified pmap.
862  * This is called at pmap creation time.
863  */
864 static void
865 pmap_alloc_l1(pmap_t pm)
866 {
867 	struct l1_ttable *l1;
868 	u_int8_t domain;
869 
870 	/*
871 	 * Remove the L1 at the head of the LRU list
872 	 */
873 	simple_lock(&l1_lru_lock);
874 	l1 = TAILQ_FIRST(&l1_lru_list);
875 	KDASSERT(l1 != NULL);
876 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
877 
878 	/*
879 	 * Pick the first available domain number, and update
880 	 * the link to the next number.
881 	 */
882 	domain = l1->l1_domain_first;
883 	l1->l1_domain_first = l1->l1_domain_free[domain];
884 
885 	/*
886 	 * If there are still free domain numbers in this L1,
887 	 * put it back on the TAIL of the LRU list.
888 	 */
889 	if (++l1->l1_domain_use_count < PMAP_DOMAINS)
890 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
891 
892 	simple_unlock(&l1_lru_lock);
893 
894 	/*
895 	 * Fix up the relevant bits in the pmap structure
896 	 */
897 	pm->pm_l1 = l1;
898 	pm->pm_domain = domain;
899 }
900 
901 /*
902  * Free an L1 translation table.
903  * This is called at pmap destruction time.
904  */
905 static void
906 pmap_free_l1(pmap_t pm)
907 {
908 	struct l1_ttable *l1 = pm->pm_l1;
909 
910 	simple_lock(&l1_lru_lock);
911 
912 	/*
913 	 * If this L1 is currently on the LRU list, remove it.
914 	 */
915 	if (l1->l1_domain_use_count < PMAP_DOMAINS)
916 		TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
917 
918 	/*
919 	 * Free up the domain number which was allocated to the pmap
920 	 */
921 	l1->l1_domain_free[pm->pm_domain] = l1->l1_domain_first;
922 	l1->l1_domain_first = pm->pm_domain;
923 	l1->l1_domain_use_count--;
924 
925 	/*
926 	 * The L1 now must have at least 1 free domain, so add
927 	 * it back to the LRU list. If the use count is zero,
928 	 * put it at the head of the list, otherwise it goes
929 	 * to the tail.
930 	 */
931 	if (l1->l1_domain_use_count == 0)
932 		TAILQ_INSERT_HEAD(&l1_lru_list, l1, l1_lru);
933 	else
934 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
935 
936 	simple_unlock(&l1_lru_lock);
937 }
938 
939 static inline void
940 pmap_use_l1(pmap_t pm)
941 {
942 	struct l1_ttable *l1;
943 
944 	/*
945 	 * Do nothing if we're in interrupt context.
946 	 * Access to an L1 by the kernel pmap must not affect
947 	 * the LRU list.
948 	 */
949 	if (current_intr_depth || pm == pmap_kernel())
950 		return;
951 
952 	l1 = pm->pm_l1;
953 
954 	/*
955 	 * If the L1 is not currently on the LRU list, just return
956 	 */
957 	if (l1->l1_domain_use_count == PMAP_DOMAINS)
958 		return;
959 
960 	simple_lock(&l1_lru_lock);
961 
962 	/*
963 	 * Check the use count again, now that we've acquired the lock
964 	 */
965 	if (l1->l1_domain_use_count == PMAP_DOMAINS) {
966 		simple_unlock(&l1_lru_lock);
967 		return;
968 	}
969 
970 	/*
971 	 * Move the L1 to the back of the LRU list
972 	 */
973 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
974 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
975 
976 	simple_unlock(&l1_lru_lock);
977 }
978 
979 /*
980  * void pmap_free_l2_ptp(pt_entry_t *, paddr_t *)
981  *
982  * Free an L2 descriptor table.
983  */
984 static inline void
985 #ifndef PMAP_INCLUDE_PTE_SYNC
986 pmap_free_l2_ptp(pt_entry_t *l2, paddr_t pa)
987 #else
988 pmap_free_l2_ptp(bool need_sync, pt_entry_t *l2, paddr_t pa)
989 #endif
990 {
991 #ifdef PMAP_INCLUDE_PTE_SYNC
992 	/*
993 	 * Note: With a write-back cache, we may need to sync this
994 	 * L2 table before re-using it.
995 	 * This is because it may have belonged to a non-current
996 	 * pmap, in which case the cache syncs would have been
997 	 * skipped when the pages were being unmapped. If the
998 	 * L2 table were then to be immediately re-allocated to
999 	 * the *current* pmap, it may well contain stale mappings
1000 	 * which have not yet been cleared by a cache write-back
1001 	 * and so would still be visible to the mmu.
1002 	 */
1003 	if (need_sync)
1004 		PTE_SYNC_RANGE(l2, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1005 #endif
1006 	pool_cache_put_paddr(&pmap_l2ptp_cache, (void *)l2, pa);
1007 }
1008 
1009 /*
1010  * Returns a pointer to the L2 bucket associated with the specified pmap
1011  * and VA, or NULL if no L2 bucket exists for the address.
1012  */
1013 static inline struct l2_bucket *
1014 pmap_get_l2_bucket(pmap_t pm, vaddr_t va)
1015 {
1016 	struct l2_dtable *l2;
1017 	struct l2_bucket *l2b;
1018 	u_short l1idx;
1019 
1020 	l1idx = L1_IDX(va);
1021 
1022 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL ||
1023 	    (l2b = &l2->l2_bucket[L2_BUCKET(l1idx)])->l2b_kva == NULL)
1024 		return (NULL);
1025 
1026 	return (l2b);
1027 }
1028 
1029 /*
1030  * Returns a pointer to the L2 bucket associated with the specified pmap
1031  * and VA.
1032  *
1033  * If no L2 bucket exists, perform the necessary allocations to put an L2
1034  * bucket/page table in place.
1035  *
1036  * Note that if a new L2 bucket/page was allocated, the caller *must*
1037  * increment the bucket occupancy counter appropriately *before*
1038  * releasing the pmap's lock to ensure no other thread or cpu deallocates
1039  * the bucket/page in the meantime.
1040  */
1041 static struct l2_bucket *
1042 pmap_alloc_l2_bucket(pmap_t pm, vaddr_t va)
1043 {
1044 	struct l2_dtable *l2;
1045 	struct l2_bucket *l2b;
1046 	u_short l1idx;
1047 
1048 	l1idx = L1_IDX(va);
1049 
1050 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
1051 		/*
1052 		 * No mapping at this address, as there is
1053 		 * no entry in the L1 table.
1054 		 * Need to allocate a new l2_dtable.
1055 		 */
1056 		if ((l2 = pmap_alloc_l2_dtable()) == NULL)
1057 			return (NULL);
1058 
1059 		/*
1060 		 * Link it into the parent pmap
1061 		 */
1062 		pm->pm_l2[L2_IDX(l1idx)] = l2;
1063 	}
1064 
1065 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
1066 
1067 	/*
1068 	 * Fetch pointer to the L2 page table associated with the address.
1069 	 */
1070 	if (l2b->l2b_kva == NULL) {
1071 		pt_entry_t *ptep;
1072 
1073 		/*
1074 		 * No L2 page table has been allocated. Chances are, this
1075 		 * is because we just allocated the l2_dtable, above.
1076 		 */
1077 		if ((ptep = pmap_alloc_l2_ptp(&l2b->l2b_phys)) == NULL) {
1078 			/*
1079 			 * Oops, no more L2 page tables available at this
1080 			 * time. We may need to deallocate the l2_dtable
1081 			 * if we allocated a new one above.
1082 			 */
1083 			if (l2->l2_occupancy == 0) {
1084 				pm->pm_l2[L2_IDX(l1idx)] = NULL;
1085 				pmap_free_l2_dtable(l2);
1086 			}
1087 			return (NULL);
1088 		}
1089 
1090 		l2->l2_occupancy++;
1091 		l2b->l2b_kva = ptep;
1092 		l2b->l2b_l1idx = l1idx;
1093 	}
1094 
1095 	return (l2b);
1096 }
1097 
1098 /*
1099  * One or more mappings in the specified L2 descriptor table have just been
1100  * invalidated.
1101  *
1102  * Garbage collect the metadata and descriptor table itself if necessary.
1103  *
1104  * The pmap lock must be acquired when this is called (not necessary
1105  * for the kernel pmap).
1106  */
1107 static void
1108 pmap_free_l2_bucket(pmap_t pm, struct l2_bucket *l2b, u_int count)
1109 {
1110 	struct l2_dtable *l2;
1111 	pd_entry_t *pl1pd, l1pd;
1112 	pt_entry_t *ptep;
1113 	u_short l1idx;
1114 
1115 	KDASSERT(count <= l2b->l2b_occupancy);
1116 
1117 	/*
1118 	 * Update the bucket's reference count according to how many
1119 	 * PTEs the caller has just invalidated.
1120 	 */
1121 	l2b->l2b_occupancy -= count;
1122 
1123 	/*
1124 	 * Note:
1125 	 *
1126 	 * Level 2 page tables allocated to the kernel pmap are never freed
1127 	 * as that would require checking all Level 1 page tables and
1128 	 * removing any references to the Level 2 page table. See also the
1129 	 * comment elsewhere about never freeing bootstrap L2 descriptors.
1130 	 *
1131 	 * We make do with just invalidating the mapping in the L2 table.
1132 	 *
1133 	 * This isn't really a big deal in practice and, in fact, leads
1134 	 * to a performance win over time as we don't need to continually
1135 	 * alloc/free.
1136 	 */
1137 	if (l2b->l2b_occupancy > 0 || pm == pmap_kernel())
1138 		return;
1139 
1140 	/*
1141 	 * There are no more valid mappings in this level 2 page table.
1142 	 * Go ahead and NULL-out the pointer in the bucket, then
1143 	 * free the page table.
1144 	 */
1145 	l1idx = l2b->l2b_l1idx;
1146 	ptep = l2b->l2b_kva;
1147 	l2b->l2b_kva = NULL;
1148 
1149 	pl1pd = &pm->pm_l1->l1_kva[l1idx];
1150 
1151 	/*
1152 	 * If the L1 slot matches the pmap's domain
1153 	 * number, then invalidate it.
1154 	 */
1155 	l1pd = *pl1pd & (L1_TYPE_MASK | L1_C_DOM_MASK);
1156 	if (l1pd == (L1_C_DOM(pm->pm_domain) | L1_TYPE_C)) {
1157 		*pl1pd = 0;
1158 		PTE_SYNC(pl1pd);
1159 	}
1160 
1161 	/*
1162 	 * Release the L2 descriptor table back to the pool cache.
1163 	 */
1164 #ifndef PMAP_INCLUDE_PTE_SYNC
1165 	pmap_free_l2_ptp(ptep, l2b->l2b_phys);
1166 #else
1167 	pmap_free_l2_ptp(!pmap_is_cached(pm), ptep, l2b->l2b_phys);
1168 #endif
1169 
1170 	/*
1171 	 * Update the reference count in the associated l2_dtable
1172 	 */
1173 	l2 = pm->pm_l2[L2_IDX(l1idx)];
1174 	if (--l2->l2_occupancy > 0)
1175 		return;
1176 
1177 	/*
1178 	 * There are no more valid mappings in any of the Level 1
1179 	 * slots managed by this l2_dtable. Go ahead and NULL-out
1180 	 * the pointer in the parent pmap and free the l2_dtable.
1181 	 */
1182 	pm->pm_l2[L2_IDX(l1idx)] = NULL;
1183 	pmap_free_l2_dtable(l2);
1184 }
1185 
1186 /*
1187  * Pool cache constructors for L2 descriptor tables, metadata and pmap
1188  * structures.
1189  */
1190 static int
1191 pmap_l2ptp_ctor(void *arg, void *v, int flags)
1192 {
1193 #ifndef PMAP_INCLUDE_PTE_SYNC
1194 	struct l2_bucket *l2b;
1195 	pt_entry_t *ptep, pte;
1196 	vaddr_t va = (vaddr_t)v & ~PGOFSET;
1197 
1198 	/*
1199 	 * The mappings for these page tables were initially made using
1200 	 * pmap_kenter_pa() by the pool subsystem. Therefore, the cache-
1201 	 * mode will not be right for page table mappings. To avoid
1202 	 * polluting the pmap_kenter_pa() code with a special case for
1203 	 * page tables, we simply fix up the cache-mode here if it's not
1204 	 * correct.
1205 	 */
1206 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
1207 	KDASSERT(l2b != NULL);
1208 	ptep = &l2b->l2b_kva[l2pte_index(va)];
1209 	pte = *ptep;
1210 
1211 	if ((pte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
1212 		/*
1213 		 * Page tables must have the cache-mode set to Write-Thru.
1214 		 */
1215 		*ptep = (pte & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode_pt;
1216 		PTE_SYNC(ptep);
1217 		cpu_tlb_flushD_SE(va);
1218 		cpu_cpwait();
1219 	}
1220 #endif
1221 
1222 	memset(v, 0, L2_TABLE_SIZE_REAL);
1223 	PTE_SYNC_RANGE(v, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1224 	return (0);
1225 }
1226 
1227 static int
1228 pmap_l2dtable_ctor(void *arg, void *v, int flags)
1229 {
1230 
1231 	memset(v, 0, sizeof(struct l2_dtable));
1232 	return (0);
1233 }
1234 
1235 static int
1236 pmap_pmap_ctor(void *arg, void *v, int flags)
1237 {
1238 
1239 	memset(v, 0, sizeof(struct pmap));
1240 	return (0);
1241 }
1242 
1243 static void
1244 pmap_pinit(pmap_t pm)
1245 {
1246 	struct l2_bucket *l2b;
1247 
1248 	if (vector_page < KERNEL_BASE) {
1249 		/*
1250 		 * Map the vector page.
1251 		 */
1252 		pmap_enter(pm, vector_page, systempage.pv_pa,
1253 		    VM_PROT_READ, VM_PROT_READ | PMAP_WIRED);
1254 		pmap_update(pm);
1255 
1256 		pm->pm_pl1vec = &pm->pm_l1->l1_kva[L1_IDX(vector_page)];
1257 		l2b = pmap_get_l2_bucket(pm, vector_page);
1258 		pm->pm_l1vec = l2b->l2b_phys | L1_C_PROTO |
1259 		    L1_C_DOM(pm->pm_domain);
1260 	} else
1261 		pm->pm_pl1vec = NULL;
1262 }
1263 
1264 /*
1265  * Since we have a virtually indexed cache, we may need to inhibit caching if
1266  * there is more than one mapping and at least one of them is writable.
1267  * Since we purge the cache on every context switch, we only need to check for
1268  * other mappings within the same pmap, or kernel_pmap.
1269  * This function is also called when a page is unmapped, to possibly reenable
1270  * caching on any remaining mappings.
1271  *
1272  * The code implements the following logic, where:
1273  *
1274  * KW = # of kernel read/write pages
1275  * KR = # of kernel read only pages
1276  * UW = # of user read/write pages
1277  * UR = # of user read only pages
1278  *
1279  * KC = kernel mapping is cacheable
1280  * UC = user mapping is cacheable
1281  *
1282  *               KW=0,KR=0  KW=0,KR>0  KW=1,KR=0  KW>1,KR>=0
1283  *             +---------------------------------------------
1284  * UW=0,UR=0   | ---        KC=1       KC=1       KC=0
1285  * UW=0,UR>0   | UC=1       KC=1,UC=1  KC=0,UC=0  KC=0,UC=0
1286  * UW=1,UR=0   | UC=1       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1287  * UW>1,UR>=0  | UC=0       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1288  */
1289 
1290 static const int pmap_vac_flags[4][4] = {
1291 	{-1,		0,		0,		PVF_KNC},
1292 	{0,		0,		PVF_NC,		PVF_NC},
1293 	{0,		PVF_NC,		PVF_NC,		PVF_NC},
1294 	{PVF_UNC,	PVF_NC,		PVF_NC,		PVF_NC}
1295 };
1296 
1297 static inline int
1298 pmap_get_vac_flags(const struct vm_page *pg)
1299 {
1300 	int kidx, uidx;
1301 
1302 	kidx = 0;
1303 	if (pg->mdpage.kro_mappings || pg->mdpage.krw_mappings > 1)
1304 		kidx |= 1;
1305 	if (pg->mdpage.krw_mappings)
1306 		kidx |= 2;
1307 
1308 	uidx = 0;
1309 	if (pg->mdpage.uro_mappings || pg->mdpage.urw_mappings > 1)
1310 		uidx |= 1;
1311 	if (pg->mdpage.urw_mappings)
1312 		uidx |= 2;
1313 
1314 	return (pmap_vac_flags[uidx][kidx]);
1315 }
1316 
1317 static inline void
1318 pmap_vac_me_harder(struct vm_page *pg, pmap_t pm, vaddr_t va)
1319 {
1320 	int nattr;
1321 
1322 	nattr = pmap_get_vac_flags(pg);
1323 
1324 	if (nattr < 0) {
1325 		pg->mdpage.pvh_attrs &= ~PVF_NC;
1326 		return;
1327 	}
1328 
1329 	if (nattr == 0 && (pg->mdpage.pvh_attrs & PVF_NC) == 0)
1330 		return;
1331 
1332 	if (pm == pmap_kernel())
1333 		pmap_vac_me_kpmap(pg, pm, va);
1334 	else
1335 		pmap_vac_me_user(pg, pm, va);
1336 
1337 	pg->mdpage.pvh_attrs = (pg->mdpage.pvh_attrs & ~PVF_NC) | nattr;
1338 }
1339 
1340 static void
1341 pmap_vac_me_kpmap(struct vm_page *pg, pmap_t pm, vaddr_t va)
1342 {
1343 	u_int u_cacheable, u_entries;
1344 	struct pv_entry *pv;
1345 	pmap_t last_pmap = pm;
1346 
1347 	/*
1348 	 * Pass one, see if there are both kernel and user pmaps for
1349 	 * this page.  Calculate whether there are user-writable or
1350 	 * kernel-writable pages.
1351 	 */
1352 	u_cacheable = 0;
1353 	for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
1354 		if (pv->pv_pmap != pm && (pv->pv_flags & PVF_NC) == 0)
1355 			u_cacheable++;
1356 	}
1357 
1358 	u_entries = pg->mdpage.urw_mappings + pg->mdpage.uro_mappings;
1359 
1360 	/*
1361 	 * We know we have just been updating a kernel entry, so if
1362 	 * all user pages are already cacheable, then there is nothing
1363 	 * further to do.
1364 	 */
1365 	if (pg->mdpage.k_mappings == 0 && u_cacheable == u_entries)
1366 		return;
1367 
1368 	if (u_entries) {
1369 		/*
1370 		 * Scan over the list again, for each entry, if it
1371 		 * might not be set correctly, call pmap_vac_me_user
1372 		 * to recalculate the settings.
1373 		 */
1374 		for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
1375 			/*
1376 			 * We know kernel mappings will get set
1377 			 * correctly in other calls.  We also know
1378 			 * that if the pmap is the same as last_pmap
1379 			 * then we've just handled this entry.
1380 			 */
1381 			if (pv->pv_pmap == pm || pv->pv_pmap == last_pmap)
1382 				continue;
1383 
1384 			/*
1385 			 * If there are kernel entries and this page
1386 			 * is writable but non-cacheable, then we can
1387 			 * skip this entry also.
1388 			 */
1389 			if (pg->mdpage.k_mappings &&
1390 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) ==
1391 			    (PVF_NC | PVF_WRITE))
1392 				continue;
1393 
1394 			/*
1395 			 * Similarly if there are no kernel-writable
1396 			 * entries and the page is already
1397 			 * read-only/cacheable.
1398 			 */
1399 			if (pg->mdpage.krw_mappings == 0 &&
1400 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 0)
1401 				continue;
1402 
1403 			/*
1404 			 * For some of the remaining cases, we know
1405 			 * that we must recalculate, but for others we
1406 			 * can't tell if they are correct or not, so
1407 			 * we recalculate anyway.
1408 			 */
1409 			pmap_vac_me_user(pg, (last_pmap = pv->pv_pmap), 0);
1410 		}
1411 
1412 		if (pg->mdpage.k_mappings == 0)
1413 			return;
1414 	}
1415 
1416 	pmap_vac_me_user(pg, pm, va);
1417 }
1418 
1419 static void
1420 pmap_vac_me_user(struct vm_page *pg, pmap_t pm, vaddr_t va)
1421 {
1422 	pmap_t kpmap = pmap_kernel();
1423 	struct pv_entry *pv, *npv;
1424 	struct l2_bucket *l2b;
1425 	pt_entry_t *ptep, pte;
1426 	u_int entries = 0;
1427 	u_int writable = 0;
1428 	u_int cacheable_entries = 0;
1429 	u_int kern_cacheable = 0;
1430 	u_int other_writable = 0;
1431 
1432 	/*
1433 	 * Count mappings and writable mappings in this pmap.
1434 	 * Include kernel mappings as part of our own.
1435 	 * Keep a pointer to the first one.
1436 	 */
1437 	for (pv = npv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
1438 		/* Count mappings in the same pmap */
1439 		if (pm == pv->pv_pmap || kpmap == pv->pv_pmap) {
1440 			if (entries++ == 0)
1441 				npv = pv;
1442 
1443 			/* Cacheable mappings */
1444 			if ((pv->pv_flags & PVF_NC) == 0) {
1445 				cacheable_entries++;
1446 				if (kpmap == pv->pv_pmap)
1447 					kern_cacheable++;
1448 			}
1449 
1450 			/* Writable mappings */
1451 			if (pv->pv_flags & PVF_WRITE)
1452 				++writable;
1453 		} else
1454 		if (pv->pv_flags & PVF_WRITE)
1455 			other_writable = 1;
1456 	}
1457 
1458 	/*
1459 	 * Enable or disable caching as necessary.
1460 	 * Note: the first entry might be part of the kernel pmap,
1461 	 * so we can't assume this is indicative of the state of the
1462 	 * other (maybe non-kpmap) entries.
1463 	 */
1464 	if ((entries > 1 && writable) ||
1465 	    (entries > 0 && pm == kpmap && other_writable)) {
1466 		if (cacheable_entries == 0)
1467 			return;
1468 
1469 		for (pv = npv; pv; pv = pv->pv_next) {
1470 			if ((pm != pv->pv_pmap && kpmap != pv->pv_pmap) ||
1471 			    (pv->pv_flags & PVF_NC))
1472 				continue;
1473 
1474 			pv->pv_flags |= PVF_NC;
1475 
1476 			l2b = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
1477 			ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1478 			pte = *ptep & ~L2_S_CACHE_MASK;
1479 
1480 			if ((va != pv->pv_va || pm != pv->pv_pmap) &&
1481 			    l2pte_valid(pte)) {
1482 				if (PV_BEEN_EXECD(pv->pv_flags)) {
1483 					pmap_idcache_wbinv_range(pv->pv_pmap,
1484 					    pv->pv_va, PAGE_SIZE);
1485 					pmap_tlb_flushID_SE(pv->pv_pmap,
1486 					    pv->pv_va);
1487 				} else
1488 				if (PV_BEEN_REFD(pv->pv_flags)) {
1489 					pmap_dcache_wb_range(pv->pv_pmap,
1490 					    pv->pv_va, PAGE_SIZE, true,
1491 					    (pv->pv_flags & PVF_WRITE) == 0);
1492 					pmap_tlb_flushD_SE(pv->pv_pmap,
1493 					    pv->pv_va);
1494 				}
1495 			}
1496 
1497 			*ptep = pte;
1498 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
1499 		}
1500 		cpu_cpwait();
1501 	} else
1502 	if (entries > cacheable_entries) {
1503 		/*
1504 		 * Turn cacheing back on for some pages.  If it is a kernel
1505 		 * page, only do so if there are no other writable pages.
1506 		 */
1507 		for (pv = npv; pv; pv = pv->pv_next) {
1508 			if (!(pv->pv_flags & PVF_NC) || (pm != pv->pv_pmap &&
1509 			    (kpmap != pv->pv_pmap || other_writable)))
1510 				continue;
1511 
1512 			pv->pv_flags &= ~PVF_NC;
1513 
1514 			l2b = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
1515 			ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1516 			pte = (*ptep & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode;
1517 
1518 			if (l2pte_valid(pte)) {
1519 				if (PV_BEEN_EXECD(pv->pv_flags)) {
1520 					pmap_tlb_flushID_SE(pv->pv_pmap,
1521 					    pv->pv_va);
1522 				} else
1523 				if (PV_BEEN_REFD(pv->pv_flags)) {
1524 					pmap_tlb_flushD_SE(pv->pv_pmap,
1525 					    pv->pv_va);
1526 				}
1527 			}
1528 
1529 			*ptep = pte;
1530 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
1531 		}
1532 	}
1533 }
1534 
1535 /*
1536  * Modify pte bits for all ptes corresponding to the given physical address.
1537  * We use `maskbits' rather than `clearbits' because we're always passing
1538  * constants and the latter would require an extra inversion at run-time.
1539  */
1540 static void
1541 pmap_clearbit(struct vm_page *pg, u_int maskbits)
1542 {
1543 	struct l2_bucket *l2b;
1544 	struct pv_entry *pv;
1545 	pt_entry_t *ptep, npte, opte;
1546 	pmap_t pm;
1547 	vaddr_t va;
1548 	u_int oflags;
1549 
1550 	NPDEBUG(PDB_BITS,
1551 	    printf("pmap_clearbit: pg %p (0x%08lx) mask 0x%x\n",
1552 	    pg, VM_PAGE_TO_PHYS(pg), maskbits));
1553 
1554 	PMAP_HEAD_TO_MAP_LOCK();
1555 	simple_lock(&pg->mdpage.pvh_slock);
1556 
1557 	/*
1558 	 * Clear saved attributes (modify, reference)
1559 	 */
1560 	pg->mdpage.pvh_attrs &= ~(maskbits & (PVF_MOD | PVF_REF));
1561 
1562 	if (pg->mdpage.pvh_list == NULL) {
1563 		simple_unlock(&pg->mdpage.pvh_slock);
1564 		PMAP_HEAD_TO_MAP_UNLOCK();
1565 		return;
1566 	}
1567 
1568 	/*
1569 	 * Loop over all current mappings setting/clearing as appropos
1570 	 */
1571 	for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
1572 		va = pv->pv_va;
1573 		pm = pv->pv_pmap;
1574 		oflags = pv->pv_flags;
1575 		pv->pv_flags &= ~maskbits;
1576 
1577 		pmap_acquire_pmap_lock(pm);
1578 
1579 		l2b = pmap_get_l2_bucket(pm, va);
1580 		KDASSERT(l2b != NULL);
1581 
1582 		ptep = &l2b->l2b_kva[l2pte_index(va)];
1583 		npte = opte = *ptep;
1584 
1585 		NPDEBUG(PDB_BITS,
1586 		    printf(
1587 		    "pmap_clearbit: pv %p, pm %p, va 0x%08lx, flag 0x%x\n",
1588 		    pv, pv->pv_pmap, pv->pv_va, oflags));
1589 
1590 		if (maskbits & (PVF_WRITE|PVF_MOD)) {
1591 			if ((pv->pv_flags & PVF_NC)) {
1592 				/*
1593 				 * Entry is not cacheable:
1594 				 *
1595 				 * Don't turn caching on again if this is a
1596 				 * modified emulation. This would be
1597 				 * inconsitent with the settings created by
1598 				 * pmap_vac_me_harder(). Otherwise, it's safe
1599 				 * to re-enable cacheing.
1600 				 *
1601 				 * There's no need to call pmap_vac_me_harder()
1602 				 * here: all pages are losing their write
1603 				 * permission.
1604 				 */
1605 				if (maskbits & PVF_WRITE) {
1606 					npte |= pte_l2_s_cache_mode;
1607 					pv->pv_flags &= ~PVF_NC;
1608 				}
1609 			} else
1610 			if (opte & L2_S_PROT_W) {
1611 				/*
1612 				 * Entry is writable/cacheable: check if pmap
1613 				 * is current if it is flush it, otherwise it
1614 				 * won't be in the cache
1615 				 */
1616 				if (PV_BEEN_EXECD(oflags))
1617 					pmap_idcache_wbinv_range(pm, pv->pv_va,
1618 					    PAGE_SIZE);
1619 				else
1620 				if (PV_BEEN_REFD(oflags))
1621 					pmap_dcache_wb_range(pm, pv->pv_va,
1622 					    PAGE_SIZE,
1623 					    (maskbits & PVF_REF) ? true : false,
1624 					    false);
1625 			}
1626 
1627 			/* make the pte read only */
1628 			npte &= ~L2_S_PROT_W;
1629 
1630 			if (maskbits & PVF_WRITE) {
1631 				/*
1632 				 * Keep alias accounting up to date
1633 				 */
1634 				if (pv->pv_pmap == pmap_kernel()) {
1635 					if (oflags & PVF_WRITE) {
1636 						pg->mdpage.krw_mappings--;
1637 						pg->mdpage.kro_mappings++;
1638 					}
1639 				} else
1640 				if (oflags & PVF_WRITE) {
1641 					pg->mdpage.urw_mappings--;
1642 					pg->mdpage.uro_mappings++;
1643 				}
1644 			}
1645 		}
1646 
1647 		if (maskbits & PVF_REF) {
1648 			if ((pv->pv_flags & PVF_NC) == 0 &&
1649 			    (maskbits & (PVF_WRITE|PVF_MOD)) == 0) {
1650 				/*
1651 				 * Check npte here; we may have already
1652 				 * done the wbinv above, and the validity
1653 				 * of the PTE is the same for opte and
1654 				 * npte.
1655 				 */
1656 				if (npte & L2_S_PROT_W) {
1657 					if (PV_BEEN_EXECD(oflags))
1658 						pmap_idcache_wbinv_range(pm,
1659 						    pv->pv_va, PAGE_SIZE);
1660 					else
1661 					if (PV_BEEN_REFD(oflags))
1662 						pmap_dcache_wb_range(pm,
1663 						    pv->pv_va, PAGE_SIZE,
1664 						    true, false);
1665 				} else
1666 				if ((npte & L2_TYPE_MASK) != L2_TYPE_INV) {
1667 					/* XXXJRT need idcache_inv_range */
1668 					if (PV_BEEN_EXECD(oflags))
1669 						pmap_idcache_wbinv_range(pm,
1670 						    pv->pv_va, PAGE_SIZE);
1671 					else
1672 					if (PV_BEEN_REFD(oflags))
1673 						pmap_dcache_wb_range(pm,
1674 						    pv->pv_va, PAGE_SIZE,
1675 						    true, true);
1676 				}
1677 			}
1678 
1679 			/*
1680 			 * Make the PTE invalid so that we will take a
1681 			 * page fault the next time the mapping is
1682 			 * referenced.
1683 			 */
1684 			npte &= ~L2_TYPE_MASK;
1685 			npte |= L2_TYPE_INV;
1686 		}
1687 
1688 		if (npte != opte) {
1689 			*ptep = npte;
1690 			PTE_SYNC(ptep);
1691 			/* Flush the TLB entry if a current pmap. */
1692 			if (PV_BEEN_EXECD(oflags))
1693 				pmap_tlb_flushID_SE(pm, pv->pv_va);
1694 			else
1695 			if (PV_BEEN_REFD(oflags))
1696 				pmap_tlb_flushD_SE(pm, pv->pv_va);
1697 		}
1698 
1699 		pmap_release_pmap_lock(pm);
1700 
1701 		NPDEBUG(PDB_BITS,
1702 		    printf("pmap_clearbit: pm %p va 0x%lx opte 0x%08x npte 0x%08x\n",
1703 		    pm, va, opte, npte));
1704 	}
1705 
1706 	simple_unlock(&pg->mdpage.pvh_slock);
1707 	PMAP_HEAD_TO_MAP_UNLOCK();
1708 }
1709 
1710 /*
1711  * pmap_clean_page()
1712  *
1713  * This is a local function used to work out the best strategy to clean
1714  * a single page referenced by its entry in the PV table. It's used by
1715  * pmap_copy_page, pmap_zero page and maybe some others later on.
1716  *
1717  * Its policy is effectively:
1718  *  o If there are no mappings, we don't bother doing anything with the cache.
1719  *  o If there is one mapping, we clean just that page.
1720  *  o If there are multiple mappings, we clean the entire cache.
1721  *
1722  * So that some functions can be further optimised, it returns 0 if it didn't
1723  * clean the entire cache, or 1 if it did.
1724  *
1725  * XXX One bug in this routine is that if the pv_entry has a single page
1726  * mapped at 0x00000000 a whole cache clean will be performed rather than
1727  * just the 1 page. Since this should not occur in everyday use and if it does
1728  * it will just result in not the most efficient clean for the page.
1729  */
1730 static int
1731 pmap_clean_page(struct pv_entry *pv, bool is_src)
1732 {
1733 	pmap_t pm, pm_to_clean = NULL;
1734 	struct pv_entry *npv;
1735 	u_int cache_needs_cleaning = 0;
1736 	u_int flags = 0;
1737 	vaddr_t page_to_clean = 0;
1738 
1739 	if (pv == NULL) {
1740 		/* nothing mapped in so nothing to flush */
1741 		return (0);
1742 	}
1743 
1744 	/*
1745 	 * Since we flush the cache each time we change to a different
1746 	 * user vmspace, we only need to flush the page if it is in the
1747 	 * current pmap.
1748 	 */
1749 	if (curproc)
1750 		pm = curproc->p_vmspace->vm_map.pmap;
1751 	else
1752 		pm = pmap_kernel();
1753 
1754 	for (npv = pv; npv; npv = npv->pv_next) {
1755 		if (npv->pv_pmap == pmap_kernel() || npv->pv_pmap == pm) {
1756 			flags |= npv->pv_flags;
1757 			/*
1758 			 * The page is mapped non-cacheable in
1759 			 * this map.  No need to flush the cache.
1760 			 */
1761 			if (npv->pv_flags & PVF_NC) {
1762 #ifdef DIAGNOSTIC
1763 				if (cache_needs_cleaning)
1764 					panic("pmap_clean_page: "
1765 					    "cache inconsistency");
1766 #endif
1767 				break;
1768 			} else if (is_src && (npv->pv_flags & PVF_WRITE) == 0)
1769 				continue;
1770 			if (cache_needs_cleaning) {
1771 				page_to_clean = 0;
1772 				break;
1773 			} else {
1774 				page_to_clean = npv->pv_va;
1775 				pm_to_clean = npv->pv_pmap;
1776 			}
1777 			cache_needs_cleaning = 1;
1778 		}
1779 	}
1780 
1781 	if (page_to_clean) {
1782 		if (PV_BEEN_EXECD(flags))
1783 			pmap_idcache_wbinv_range(pm_to_clean, page_to_clean,
1784 			    PAGE_SIZE);
1785 		else
1786 			pmap_dcache_wb_range(pm_to_clean, page_to_clean,
1787 			    PAGE_SIZE, !is_src, (flags & PVF_WRITE) == 0);
1788 	} else if (cache_needs_cleaning) {
1789 		if (PV_BEEN_EXECD(flags))
1790 			pmap_idcache_wbinv_all(pm);
1791 		else
1792 			pmap_dcache_wbinv_all(pm);
1793 		return (1);
1794 	}
1795 	return (0);
1796 }
1797 
1798 /*
1799  * Routine:	pmap_page_remove
1800  * Function:
1801  *		Removes this physical page from
1802  *		all physical maps in which it resides.
1803  *		Reflects back modify bits to the pager.
1804  */
1805 static void
1806 pmap_page_remove(struct vm_page *pg)
1807 {
1808 	struct l2_bucket *l2b;
1809 	struct pv_entry *pv, *npv;
1810 	pmap_t pm, curpm;
1811 	pt_entry_t *ptep, pte;
1812 	bool flush;
1813 	u_int flags;
1814 
1815 	NPDEBUG(PDB_FOLLOW,
1816 	    printf("pmap_page_remove: pg %p (0x%08lx)\n", pg,
1817 	    VM_PAGE_TO_PHYS(pg)));
1818 
1819 	PMAP_HEAD_TO_MAP_LOCK();
1820 	simple_lock(&pg->mdpage.pvh_slock);
1821 
1822 	pv = pg->mdpage.pvh_list;
1823 	if (pv == NULL) {
1824 		simple_unlock(&pg->mdpage.pvh_slock);
1825 		PMAP_HEAD_TO_MAP_UNLOCK();
1826 		return;
1827 	}
1828 
1829 	/*
1830 	 * Clear alias counts
1831 	 */
1832 	pg->mdpage.k_mappings = 0;
1833 	pg->mdpage.urw_mappings = pg->mdpage.uro_mappings = 0;
1834 
1835 	flush = false;
1836 	flags = 0;
1837 	if (curproc)
1838 		curpm = curproc->p_vmspace->vm_map.pmap;
1839 	else
1840 		curpm = pmap_kernel();
1841 
1842 	pmap_clean_page(pv, false);
1843 
1844 	while (pv) {
1845 		pm = pv->pv_pmap;
1846 		if (flush == false && (pm == curpm || pm == pmap_kernel()))
1847 			flush = true;
1848 
1849 		pmap_acquire_pmap_lock(pm);
1850 
1851 		l2b = pmap_get_l2_bucket(pm, pv->pv_va);
1852 		KDASSERT(l2b != NULL);
1853 
1854 		ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1855 		pte = *ptep;
1856 
1857 		/*
1858 		 * Update statistics
1859 		 */
1860 		--pm->pm_stats.resident_count;
1861 
1862 		/* Wired bit */
1863 		if (pv->pv_flags & PVF_WIRED)
1864 			--pm->pm_stats.wired_count;
1865 
1866 		flags |= pv->pv_flags;
1867 
1868 		/*
1869 		 * Invalidate the PTEs.
1870 		 */
1871 		*ptep = 0;
1872 		PTE_SYNC_CURRENT(pm, ptep);
1873 		pmap_free_l2_bucket(pm, l2b, 1);
1874 
1875 		npv = pv->pv_next;
1876 		pool_put(&pmap_pv_pool, pv);
1877 		pv = npv;
1878 		pmap_release_pmap_lock(pm);
1879 	}
1880 	pg->mdpage.pvh_list = NULL;
1881 	simple_unlock(&pg->mdpage.pvh_slock);
1882 	PMAP_HEAD_TO_MAP_UNLOCK();
1883 
1884 	if (flush) {
1885 		/*
1886 		 * Note: We can't use pmap_tlb_flush{I,}D() here since that
1887 		 * would need a subsequent call to pmap_update() to ensure
1888 		 * curpm->pm_cstate.cs_all is reset. Our callers are not
1889 		 * required to do that (see pmap(9)), so we can't modify
1890 		 * the current pmap's state.
1891 		 */
1892 		if (PV_BEEN_EXECD(flags))
1893 			cpu_tlb_flushID();
1894 		else
1895 			cpu_tlb_flushD();
1896 	}
1897 	cpu_cpwait();
1898 }
1899 
1900 /*
1901  * pmap_t pmap_create(void)
1902  *
1903  *      Create a new pmap structure from scratch.
1904  */
1905 pmap_t
1906 pmap_create(void)
1907 {
1908 	pmap_t pm;
1909 
1910 	pm = pool_cache_get(&pmap_cache, PR_WAITOK);
1911 
1912 	simple_lock_init(&pm->pm_lock);
1913 	pm->pm_obj.pgops = NULL;	/* currently not a mappable object */
1914 	TAILQ_INIT(&pm->pm_obj.memq);
1915 	pm->pm_obj.uo_npages = 0;
1916 	pm->pm_obj.uo_refs = 1;
1917 	pm->pm_stats.wired_count = 0;
1918 	pm->pm_stats.resident_count = 1;
1919 	pm->pm_cstate.cs_all = 0;
1920 	pmap_alloc_l1(pm);
1921 
1922 	/*
1923 	 * Note: The pool cache ensures that the pm_l2[] array is already
1924 	 * initialised to zero.
1925 	 */
1926 
1927 	pmap_pinit(pm);
1928 
1929 	LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list);
1930 
1931 	return (pm);
1932 }
1933 
1934 /*
1935  * void pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
1936  *     int flags)
1937  *
1938  *      Insert the given physical page (p) at
1939  *      the specified virtual address (v) in the
1940  *      target physical map with the protection requested.
1941  *
1942  *      NB:  This is the only routine which MAY NOT lazy-evaluate
1943  *      or lose information.  That is, this routine must actually
1944  *      insert this page into the given map NOW.
1945  */
1946 int
1947 pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot, int flags)
1948 {
1949 	struct l2_bucket *l2b;
1950 	struct vm_page *pg, *opg;
1951 	struct pv_entry *pve;
1952 	pt_entry_t *ptep, npte, opte;
1953 	u_int nflags;
1954 	u_int oflags;
1955 
1956 	NPDEBUG(PDB_ENTER, printf("pmap_enter: pm %p va 0x%lx pa 0x%lx prot %x flag %x\n", pm, va, pa, prot, flags));
1957 
1958 	KDASSERT((flags & PMAP_WIRED) == 0 || (flags & VM_PROT_ALL) != 0);
1959 	KDASSERT(((va | pa) & PGOFSET) == 0);
1960 
1961 	/*
1962 	 * Get a pointer to the page.  Later on in this function, we
1963 	 * test for a managed page by checking pg != NULL.
1964 	 */
1965 	pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
1966 
1967 	nflags = 0;
1968 	if (prot & VM_PROT_WRITE)
1969 		nflags |= PVF_WRITE;
1970 	if (prot & VM_PROT_EXECUTE)
1971 		nflags |= PVF_EXEC;
1972 	if (flags & PMAP_WIRED)
1973 		nflags |= PVF_WIRED;
1974 
1975 	PMAP_MAP_TO_HEAD_LOCK();
1976 	pmap_acquire_pmap_lock(pm);
1977 
1978 	/*
1979 	 * Fetch the L2 bucket which maps this page, allocating one if
1980 	 * necessary for user pmaps.
1981 	 */
1982 	if (pm == pmap_kernel())
1983 		l2b = pmap_get_l2_bucket(pm, va);
1984 	else
1985 		l2b = pmap_alloc_l2_bucket(pm, va);
1986 	if (l2b == NULL) {
1987 		if (flags & PMAP_CANFAIL) {
1988 			pmap_release_pmap_lock(pm);
1989 			PMAP_MAP_TO_HEAD_UNLOCK();
1990 			return (ENOMEM);
1991 		}
1992 		panic("pmap_enter: failed to allocate L2 bucket");
1993 	}
1994 	ptep = &l2b->l2b_kva[l2pte_index(va)];
1995 	opte = *ptep;
1996 	npte = pa;
1997 	oflags = 0;
1998 
1999 	if (opte) {
2000 		/*
2001 		 * There is already a mapping at this address.
2002 		 * If the physical address is different, lookup the
2003 		 * vm_page.
2004 		 */
2005 		if (l2pte_pa(opte) != pa)
2006 			opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
2007 		else
2008 			opg = pg;
2009 	} else
2010 		opg = NULL;
2011 
2012 	if (pg) {
2013 		/*
2014 		 * This is to be a managed mapping.
2015 		 */
2016 		if ((flags & VM_PROT_ALL) ||
2017 		    (pg->mdpage.pvh_attrs & PVF_REF)) {
2018 			/*
2019 			 * - The access type indicates that we don't need
2020 			 *   to do referenced emulation.
2021 			 * OR
2022 			 * - The physical page has already been referenced
2023 			 *   so no need to re-do referenced emulation here.
2024 			 */
2025 			npte |= L2_S_PROTO;
2026 
2027 			nflags |= PVF_REF;
2028 
2029 			if ((prot & VM_PROT_WRITE) != 0 &&
2030 			    ((flags & VM_PROT_WRITE) != 0 ||
2031 			     (pg->mdpage.pvh_attrs & PVF_MOD) != 0)) {
2032 				/*
2033 				 * This is a writable mapping, and the
2034 				 * page's mod state indicates it has
2035 				 * already been modified. Make it
2036 				 * writable from the outset.
2037 				 */
2038 				npte |= L2_S_PROT_W;
2039 				nflags |= PVF_MOD;
2040 			}
2041 		} else {
2042 			/*
2043 			 * Need to do page referenced emulation.
2044 			 */
2045 			npte |= L2_TYPE_INV;
2046 		}
2047 
2048 		npte |= pte_l2_s_cache_mode;
2049 
2050 		if (pg == opg) {
2051 			/*
2052 			 * We're changing the attrs of an existing mapping.
2053 			 */
2054 			simple_lock(&pg->mdpage.pvh_slock);
2055 			oflags = pmap_modify_pv(pg, pm, va,
2056 			    PVF_WRITE | PVF_EXEC | PVF_WIRED |
2057 			    PVF_MOD | PVF_REF, nflags);
2058 			simple_unlock(&pg->mdpage.pvh_slock);
2059 
2060 			/*
2061 			 * We may need to flush the cache if we're
2062 			 * doing rw-ro...
2063 			 */
2064 			if (pm->pm_cstate.cs_cache_d &&
2065 			    (oflags & PVF_NC) == 0 &&
2066 			    (opte & L2_S_PROT_W) != 0 &&
2067 			    (prot & VM_PROT_WRITE) == 0)
2068 				cpu_dcache_wb_range(va, PAGE_SIZE);
2069 		} else {
2070 			/*
2071 			 * New mapping, or changing the backing page
2072 			 * of an existing mapping.
2073 			 */
2074 			if (opg) {
2075 				/*
2076 				 * Replacing an existing mapping with a new one.
2077 				 * It is part of our managed memory so we
2078 				 * must remove it from the PV list
2079 				 */
2080 				simple_lock(&opg->mdpage.pvh_slock);
2081 				pve = pmap_remove_pv(opg, pm, va, 0);
2082 				pmap_vac_me_harder(opg, pm, 0);
2083 				simple_unlock(&opg->mdpage.pvh_slock);
2084 				oflags = pve->pv_flags;
2085 
2086 				/*
2087 				 * If the old mapping was valid (ref/mod
2088 				 * emulation creates 'invalid' mappings
2089 				 * initially) then make sure to frob
2090 				 * the cache.
2091 				 */
2092 				if ((oflags & PVF_NC) == 0 &&
2093 				    l2pte_valid(opte)) {
2094 					if (PV_BEEN_EXECD(oflags)) {
2095 						pmap_idcache_wbinv_range(pm, va,
2096 						    PAGE_SIZE);
2097 					} else
2098 					if (PV_BEEN_REFD(oflags)) {
2099 						pmap_dcache_wb_range(pm, va,
2100 						    PAGE_SIZE, true,
2101 						    (oflags & PVF_WRITE) == 0);
2102 					}
2103 				}
2104 			} else
2105 			if ((pve = pool_get(&pmap_pv_pool, PR_NOWAIT)) == NULL){
2106 				if ((flags & PMAP_CANFAIL) == 0)
2107 					panic("pmap_enter: no pv entries");
2108 
2109 				if (pm != pmap_kernel())
2110 					pmap_free_l2_bucket(pm, l2b, 0);
2111 				pmap_release_pmap_lock(pm);
2112 				PMAP_MAP_TO_HEAD_UNLOCK();
2113 				NPDEBUG(PDB_ENTER,
2114 				    printf("pmap_enter: ENOMEM\n"));
2115 				return (ENOMEM);
2116 			}
2117 
2118 			pmap_enter_pv(pg, pve, pm, va, nflags);
2119 		}
2120 	} else {
2121 		/*
2122 		 * We're mapping an unmanaged page.
2123 		 * These are always readable, and possibly writable, from
2124 		 * the get go as we don't need to track ref/mod status.
2125 		 */
2126 		npte |= L2_S_PROTO;
2127 		if (prot & VM_PROT_WRITE)
2128 			npte |= L2_S_PROT_W;
2129 
2130 		/*
2131 		 * Make sure the vector table is mapped cacheable
2132 		 */
2133 		if (pm != pmap_kernel() && va == vector_page)
2134 			npte |= pte_l2_s_cache_mode;
2135 
2136 		if (opg) {
2137 			/*
2138 			 * Looks like there's an existing 'managed' mapping
2139 			 * at this address.
2140 			 */
2141 			simple_lock(&opg->mdpage.pvh_slock);
2142 			pve = pmap_remove_pv(opg, pm, va, 0);
2143 			pmap_vac_me_harder(opg, pm, 0);
2144 			simple_unlock(&opg->mdpage.pvh_slock);
2145 			oflags = pve->pv_flags;
2146 
2147 			if ((oflags & PVF_NC) == 0 && l2pte_valid(opte)) {
2148 				if (PV_BEEN_EXECD(oflags))
2149 					pmap_idcache_wbinv_range(pm, va,
2150 					    PAGE_SIZE);
2151 				else
2152 				if (PV_BEEN_REFD(oflags))
2153 					pmap_dcache_wb_range(pm, va, PAGE_SIZE,
2154 					    true, (oflags & PVF_WRITE) == 0);
2155 			}
2156 			pool_put(&pmap_pv_pool, pve);
2157 		}
2158 	}
2159 
2160 	/*
2161 	 * Make sure userland mappings get the right permissions
2162 	 */
2163 	if (pm != pmap_kernel() && va != vector_page)
2164 		npte |= L2_S_PROT_U;
2165 
2166 	/*
2167 	 * Keep the stats up to date
2168 	 */
2169 	if (opte == 0) {
2170 		l2b->l2b_occupancy++;
2171 		pm->pm_stats.resident_count++;
2172 	}
2173 
2174 	NPDEBUG(PDB_ENTER,
2175 	    printf("pmap_enter: opte 0x%08x npte 0x%08x\n", opte, npte));
2176 
2177 	/*
2178 	 * If this is just a wiring change, the two PTEs will be
2179 	 * identical, so there's no need to update the page table.
2180 	 */
2181 	if (npte != opte) {
2182 		bool is_cached = pmap_is_cached(pm);
2183 
2184 		*ptep = npte;
2185 		if (is_cached) {
2186 			/*
2187 			 * We only need to frob the cache/tlb if this pmap
2188 			 * is current
2189 			 */
2190 			PTE_SYNC(ptep);
2191 			if (va != vector_page && l2pte_valid(npte)) {
2192 				/*
2193 				 * This mapping is likely to be accessed as
2194 				 * soon as we return to userland. Fix up the
2195 				 * L1 entry to avoid taking another
2196 				 * page/domain fault.
2197 				 */
2198 				pd_entry_t *pl1pd, l1pd;
2199 
2200 				pl1pd = &pm->pm_l1->l1_kva[L1_IDX(va)];
2201 				l1pd = l2b->l2b_phys | L1_C_DOM(pm->pm_domain) |
2202 				    L1_C_PROTO;
2203 				if (*pl1pd != l1pd) {
2204 					*pl1pd = l1pd;
2205 					PTE_SYNC(pl1pd);
2206 				}
2207 			}
2208 		}
2209 
2210 		if (PV_BEEN_EXECD(oflags))
2211 			pmap_tlb_flushID_SE(pm, va);
2212 		else
2213 		if (PV_BEEN_REFD(oflags))
2214 			pmap_tlb_flushD_SE(pm, va);
2215 
2216 		NPDEBUG(PDB_ENTER,
2217 		    printf("pmap_enter: is_cached %d cs 0x%08x\n",
2218 		    is_cached, pm->pm_cstate.cs_all));
2219 
2220 		if (pg != NULL) {
2221 			simple_lock(&pg->mdpage.pvh_slock);
2222 			pmap_vac_me_harder(pg, pm, va);
2223 			simple_unlock(&pg->mdpage.pvh_slock);
2224 		}
2225 	}
2226 
2227 	pmap_release_pmap_lock(pm);
2228 	PMAP_MAP_TO_HEAD_UNLOCK();
2229 
2230 	return (0);
2231 }
2232 
2233 /*
2234  * pmap_remove()
2235  *
2236  * pmap_remove is responsible for nuking a number of mappings for a range
2237  * of virtual address space in the current pmap. To do this efficiently
2238  * is interesting, because in a number of cases a wide virtual address
2239  * range may be supplied that contains few actual mappings. So, the
2240  * optimisations are:
2241  *  1. Skip over hunks of address space for which no L1 or L2 entry exists.
2242  *  2. Build up a list of pages we've hit, up to a maximum, so we can
2243  *     maybe do just a partial cache clean. This path of execution is
2244  *     complicated by the fact that the cache must be flushed _before_
2245  *     the PTE is nuked, being a VAC :-)
2246  *  3. If we're called after UVM calls pmap_remove_all(), we can defer
2247  *     all invalidations until pmap_update(), since pmap_remove_all() has
2248  *     already flushed the cache.
2249  *  4. Maybe later fast-case a single page, but I don't think this is
2250  *     going to make _that_ much difference overall.
2251  */
2252 
2253 #define	PMAP_REMOVE_CLEAN_LIST_SIZE	3
2254 
2255 void
2256 pmap_do_remove(pmap_t pm, vaddr_t sva, vaddr_t eva, int skip_wired)
2257 {
2258 	struct l2_bucket *l2b;
2259 	vaddr_t next_bucket;
2260 	pt_entry_t *ptep;
2261 	u_int cleanlist_idx, total, cnt;
2262 	struct {
2263 		vaddr_t va;
2264 		pt_entry_t *pte;
2265 	} cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
2266 	u_int mappings, is_exec, is_refd;
2267 
2268 	NPDEBUG(PDB_REMOVE, printf("pmap_do_remove: pmap=%p sva=%08lx "
2269 	    "eva=%08lx\n", pm, sva, eva));
2270 
2271 	/*
2272 	 * we lock in the pmap => pv_head direction
2273 	 */
2274 	PMAP_MAP_TO_HEAD_LOCK();
2275 	pmap_acquire_pmap_lock(pm);
2276 
2277 	if (pm->pm_remove_all || !pmap_is_cached(pm)) {
2278 		cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
2279 		if (pm->pm_cstate.cs_tlb == 0)
2280 			pm->pm_remove_all = true;
2281 	} else
2282 		cleanlist_idx = 0;
2283 
2284 	total = 0;
2285 
2286 	while (sva < eva) {
2287 		/*
2288 		 * Do one L2 bucket's worth at a time.
2289 		 */
2290 		next_bucket = L2_NEXT_BUCKET(sva);
2291 		if (next_bucket > eva)
2292 			next_bucket = eva;
2293 
2294 		l2b = pmap_get_l2_bucket(pm, sva);
2295 		if (l2b == NULL) {
2296 			sva = next_bucket;
2297 			continue;
2298 		}
2299 
2300 		ptep = &l2b->l2b_kva[l2pte_index(sva)];
2301 
2302 		for (mappings = 0; sva < next_bucket; sva += PAGE_SIZE, ptep++){
2303 			struct vm_page *pg;
2304 			pt_entry_t pte;
2305 			paddr_t pa;
2306 
2307 			pte = *ptep;
2308 
2309 			if (pte == 0) {
2310 				/* Nothing here, move along */
2311 				continue;
2312 			}
2313 
2314 			pa = l2pte_pa(pte);
2315 			is_exec = 0;
2316 			is_refd = 1;
2317 
2318 			/*
2319 			 * Update flags. In a number of circumstances,
2320 			 * we could cluster a lot of these and do a
2321 			 * number of sequential pages in one go.
2322 			 */
2323 			if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
2324 				struct pv_entry *pve;
2325 				simple_lock(&pg->mdpage.pvh_slock);
2326 				pve = pmap_remove_pv(pg, pm, sva, skip_wired);
2327 				pmap_vac_me_harder(pg, pm, 0);
2328 				simple_unlock(&pg->mdpage.pvh_slock);
2329 				if (pve != NULL) {
2330 					if (pm->pm_remove_all == false) {
2331 						is_exec =
2332 						   PV_BEEN_EXECD(pve->pv_flags);
2333 						is_refd =
2334 						   PV_BEEN_REFD(pve->pv_flags);
2335 					}
2336 					pool_put(&pmap_pv_pool, pve);
2337 				} else
2338 				if (skip_wired) {
2339 					/* The mapping is wired. Skip it */
2340 					continue;
2341 				}
2342 			} else
2343 			if (skip_wired) {
2344 				/* Unmanaged pages are always wired. */
2345 				continue;
2346 			}
2347 
2348 			mappings++;
2349 
2350 			if (!l2pte_valid(pte)) {
2351 				/*
2352 				 * Ref/Mod emulation is still active for this
2353 				 * mapping, therefore it is has not yet been
2354 				 * accessed. No need to frob the cache/tlb.
2355 				 */
2356 				*ptep = 0;
2357 				PTE_SYNC_CURRENT(pm, ptep);
2358 				continue;
2359 			}
2360 
2361 			if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
2362 				/* Add to the clean list. */
2363 				cleanlist[cleanlist_idx].pte = ptep;
2364 				cleanlist[cleanlist_idx].va =
2365 				    sva | (is_exec & 1);
2366 				cleanlist_idx++;
2367 			} else
2368 			if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
2369 				/* Nuke everything if needed. */
2370 				pmap_idcache_wbinv_all(pm);
2371 				pmap_tlb_flushID(pm);
2372 
2373 				/*
2374 				 * Roll back the previous PTE list,
2375 				 * and zero out the current PTE.
2376 				 */
2377 				for (cnt = 0;
2378 				     cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
2379 					*cleanlist[cnt].pte = 0;
2380 				}
2381 				*ptep = 0;
2382 				PTE_SYNC(ptep);
2383 				cleanlist_idx++;
2384 				pm->pm_remove_all = true;
2385 			} else {
2386 				*ptep = 0;
2387 				PTE_SYNC(ptep);
2388 				if (pm->pm_remove_all == false) {
2389 					if (is_exec)
2390 						pmap_tlb_flushID_SE(pm, sva);
2391 					else
2392 					if (is_refd)
2393 						pmap_tlb_flushD_SE(pm, sva);
2394 				}
2395 			}
2396 		}
2397 
2398 		/*
2399 		 * Deal with any left overs
2400 		 */
2401 		if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
2402 			total += cleanlist_idx;
2403 			for (cnt = 0; cnt < cleanlist_idx; cnt++) {
2404 				if (pm->pm_cstate.cs_all != 0) {
2405 					vaddr_t clva = cleanlist[cnt].va & ~1;
2406 					if (cleanlist[cnt].va & 1) {
2407 						pmap_idcache_wbinv_range(pm,
2408 						    clva, PAGE_SIZE);
2409 						pmap_tlb_flushID_SE(pm, clva);
2410 					} else {
2411 						pmap_dcache_wb_range(pm,
2412 						    clva, PAGE_SIZE, true,
2413 						    false);
2414 						pmap_tlb_flushD_SE(pm, clva);
2415 					}
2416 				}
2417 				*cleanlist[cnt].pte = 0;
2418 				PTE_SYNC_CURRENT(pm, cleanlist[cnt].pte);
2419 			}
2420 
2421 			/*
2422 			 * If it looks like we're removing a whole bunch
2423 			 * of mappings, it's faster to just write-back
2424 			 * the whole cache now and defer TLB flushes until
2425 			 * pmap_update() is called.
2426 			 */
2427 			if (total <= PMAP_REMOVE_CLEAN_LIST_SIZE)
2428 				cleanlist_idx = 0;
2429 			else {
2430 				cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
2431 				pmap_idcache_wbinv_all(pm);
2432 				pm->pm_remove_all = true;
2433 			}
2434 		}
2435 
2436 		pmap_free_l2_bucket(pm, l2b, mappings);
2437 		pm->pm_stats.resident_count -= mappings;
2438 	}
2439 
2440 	pmap_release_pmap_lock(pm);
2441 	PMAP_MAP_TO_HEAD_UNLOCK();
2442 }
2443 
2444 /*
2445  * pmap_kenter_pa: enter an unmanaged, wired kernel mapping
2446  *
2447  * We assume there is already sufficient KVM space available
2448  * to do this, as we can't allocate L2 descriptor tables/metadata
2449  * from here.
2450  */
2451 void
2452 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot)
2453 {
2454 	struct l2_bucket *l2b;
2455 	pt_entry_t *ptep, opte;
2456 
2457 	NPDEBUG(PDB_KENTER,
2458 	    printf("pmap_kenter_pa: va 0x%08lx, pa 0x%08lx, prot 0x%x\n",
2459 	    va, pa, prot));
2460 
2461 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
2462 	KDASSERT(l2b != NULL);
2463 
2464 	ptep = &l2b->l2b_kva[l2pte_index(va)];
2465 	opte = *ptep;
2466 
2467 	if (l2pte_valid(opte)) {
2468 		cpu_dcache_wbinv_range(va, PAGE_SIZE);
2469 		cpu_tlb_flushD_SE(va);
2470 		cpu_cpwait();
2471 	} else
2472 	if (opte == 0)
2473 		l2b->l2b_occupancy++;
2474 
2475 	*ptep = L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) |
2476 	    pte_l2_s_cache_mode;
2477 	PTE_SYNC(ptep);
2478 }
2479 
2480 void
2481 pmap_kremove(vaddr_t va, vsize_t len)
2482 {
2483 	struct l2_bucket *l2b;
2484 	pt_entry_t *ptep, *sptep, opte;
2485 	vaddr_t next_bucket, eva;
2486 	u_int mappings;
2487 
2488 	NPDEBUG(PDB_KREMOVE, printf("pmap_kremove: va 0x%08lx, len 0x%08lx\n",
2489 	    va, len));
2490 
2491 	eva = va + len;
2492 
2493 	while (va < eva) {
2494 		next_bucket = L2_NEXT_BUCKET(va);
2495 		if (next_bucket > eva)
2496 			next_bucket = eva;
2497 
2498 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
2499 		KDASSERT(l2b != NULL);
2500 
2501 		sptep = ptep = &l2b->l2b_kva[l2pte_index(va)];
2502 		mappings = 0;
2503 
2504 		while (va < next_bucket) {
2505 			opte = *ptep;
2506 			if (l2pte_valid(opte)) {
2507 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
2508 				cpu_tlb_flushD_SE(va);
2509 			}
2510 			if (opte) {
2511 				*ptep = 0;
2512 				mappings++;
2513 			}
2514 			va += PAGE_SIZE;
2515 			ptep++;
2516 		}
2517 		KDASSERT(mappings <= l2b->l2b_occupancy);
2518 		l2b->l2b_occupancy -= mappings;
2519 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
2520 	}
2521 	cpu_cpwait();
2522 }
2523 
2524 bool
2525 pmap_extract(pmap_t pm, vaddr_t va, paddr_t *pap)
2526 {
2527 	struct l2_dtable *l2;
2528 	pd_entry_t *pl1pd, l1pd;
2529 	pt_entry_t *ptep, pte;
2530 	paddr_t pa;
2531 	u_int l1idx;
2532 
2533 	pmap_acquire_pmap_lock(pm);
2534 
2535 	l1idx = L1_IDX(va);
2536 	pl1pd = &pm->pm_l1->l1_kva[l1idx];
2537 	l1pd = *pl1pd;
2538 
2539 	if (l1pte_section_p(l1pd)) {
2540 		/*
2541 		 * These should only happen for pmap_kernel()
2542 		 */
2543 		KDASSERT(pm == pmap_kernel());
2544 		pmap_release_pmap_lock(pm);
2545 		pa = (l1pd & L1_S_FRAME) | (va & L1_S_OFFSET);
2546 	} else {
2547 		/*
2548 		 * Note that we can't rely on the validity of the L1
2549 		 * descriptor as an indication that a mapping exists.
2550 		 * We have to look it up in the L2 dtable.
2551 		 */
2552 		l2 = pm->pm_l2[L2_IDX(l1idx)];
2553 
2554 		if (l2 == NULL ||
2555 		    (ptep = l2->l2_bucket[L2_BUCKET(l1idx)].l2b_kva) == NULL) {
2556 			pmap_release_pmap_lock(pm);
2557 			return (false);
2558 		}
2559 
2560 		ptep = &ptep[l2pte_index(va)];
2561 		pte = *ptep;
2562 		pmap_release_pmap_lock(pm);
2563 
2564 		if (pte == 0)
2565 			return (false);
2566 
2567 		switch (pte & L2_TYPE_MASK) {
2568 		case L2_TYPE_L:
2569 			pa = (pte & L2_L_FRAME) | (va & L2_L_OFFSET);
2570 			break;
2571 
2572 		default:
2573 			pa = (pte & L2_S_FRAME) | (va & L2_S_OFFSET);
2574 			break;
2575 		}
2576 	}
2577 
2578 	if (pap != NULL)
2579 		*pap = pa;
2580 
2581 	return (true);
2582 }
2583 
2584 void
2585 pmap_protect(pmap_t pm, vaddr_t sva, vaddr_t eva, vm_prot_t prot)
2586 {
2587 	struct l2_bucket *l2b;
2588 	pt_entry_t *ptep, pte;
2589 	vaddr_t next_bucket;
2590 	u_int flags;
2591 	int flush;
2592 
2593 	NPDEBUG(PDB_PROTECT,
2594 	    printf("pmap_protect: pm %p sva 0x%lx eva 0x%lx prot 0x%x\n",
2595 	    pm, sva, eva, prot));
2596 
2597 	if ((prot & VM_PROT_READ) == 0) {
2598 		pmap_remove(pm, sva, eva);
2599 		return;
2600 	}
2601 
2602 	if (prot & VM_PROT_WRITE) {
2603 		/*
2604 		 * If this is a read->write transition, just ignore it and let
2605 		 * uvm_fault() take care of it later.
2606 		 */
2607 		return;
2608 	}
2609 
2610 	PMAP_MAP_TO_HEAD_LOCK();
2611 	pmap_acquire_pmap_lock(pm);
2612 
2613 	/*
2614 	 * OK, at this point, we know we're doing write-protect operation.
2615 	 * If the pmap is active, write-back the range.
2616 	 */
2617 	pmap_dcache_wb_range(pm, sva, eva - sva, false, false);
2618 
2619 	flush = ((eva - sva) >= (PAGE_SIZE * 4)) ? 0 : -1;
2620 	flags = 0;
2621 
2622 	while (sva < eva) {
2623 		next_bucket = L2_NEXT_BUCKET(sva);
2624 		if (next_bucket > eva)
2625 			next_bucket = eva;
2626 
2627 		l2b = pmap_get_l2_bucket(pm, sva);
2628 		if (l2b == NULL) {
2629 			sva = next_bucket;
2630 			continue;
2631 		}
2632 
2633 		ptep = &l2b->l2b_kva[l2pte_index(sva)];
2634 
2635 		while (sva < next_bucket) {
2636 			if ((pte = *ptep) != 0 && (pte & L2_S_PROT_W) != 0) {
2637 				struct vm_page *pg;
2638 				u_int f;
2639 
2640 				pg = PHYS_TO_VM_PAGE(l2pte_pa(pte));
2641 				pte &= ~L2_S_PROT_W;
2642 				*ptep = pte;
2643 				PTE_SYNC(ptep);
2644 
2645 				if (pg != NULL) {
2646 					simple_lock(&pg->mdpage.pvh_slock);
2647 					f = pmap_modify_pv(pg, pm, sva,
2648 					    PVF_WRITE, 0);
2649 					pmap_vac_me_harder(pg, pm, sva);
2650 					simple_unlock(&pg->mdpage.pvh_slock);
2651 				} else
2652 					f = PVF_REF | PVF_EXEC;
2653 
2654 				if (flush >= 0) {
2655 					flush++;
2656 					flags |= f;
2657 				} else
2658 				if (PV_BEEN_EXECD(f))
2659 					pmap_tlb_flushID_SE(pm, sva);
2660 				else
2661 				if (PV_BEEN_REFD(f))
2662 					pmap_tlb_flushD_SE(pm, sva);
2663 			}
2664 
2665 			sva += PAGE_SIZE;
2666 			ptep++;
2667 		}
2668 	}
2669 
2670 	pmap_release_pmap_lock(pm);
2671 	PMAP_MAP_TO_HEAD_UNLOCK();
2672 
2673 	if (flush) {
2674 		if (PV_BEEN_EXECD(flags))
2675 			pmap_tlb_flushID(pm);
2676 		else
2677 		if (PV_BEEN_REFD(flags))
2678 			pmap_tlb_flushD(pm);
2679 	}
2680 }
2681 
2682 void
2683 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
2684 {
2685 
2686 	NPDEBUG(PDB_PROTECT,
2687 	    printf("pmap_page_protect: pg %p (0x%08lx), prot 0x%x\n",
2688 	    pg, VM_PAGE_TO_PHYS(pg), prot));
2689 
2690 	switch(prot) {
2691 	case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
2692 	case VM_PROT_READ|VM_PROT_WRITE:
2693 		return;
2694 
2695 	case VM_PROT_READ:
2696 	case VM_PROT_READ|VM_PROT_EXECUTE:
2697 		pmap_clearbit(pg, PVF_WRITE);
2698 		break;
2699 
2700 	default:
2701 		pmap_page_remove(pg);
2702 		break;
2703 	}
2704 }
2705 
2706 /*
2707  * pmap_clear_modify:
2708  *
2709  *	Clear the "modified" attribute for a page.
2710  */
2711 bool
2712 pmap_clear_modify(struct vm_page *pg)
2713 {
2714 	bool rv;
2715 
2716 	if (pg->mdpage.pvh_attrs & PVF_MOD) {
2717 		rv = true;
2718 		pmap_clearbit(pg, PVF_MOD);
2719 	} else
2720 		rv = false;
2721 
2722 	return (rv);
2723 }
2724 
2725 /*
2726  * pmap_clear_reference:
2727  *
2728  *	Clear the "referenced" attribute for a page.
2729  */
2730 bool
2731 pmap_clear_reference(struct vm_page *pg)
2732 {
2733 	bool rv;
2734 
2735 	if (pg->mdpage.pvh_attrs & PVF_REF) {
2736 		rv = true;
2737 		pmap_clearbit(pg, PVF_REF);
2738 	} else
2739 		rv = false;
2740 
2741 	return (rv);
2742 }
2743 
2744 /*
2745  * pmap_is_modified:
2746  *
2747  *	Test if a page has the "modified" attribute.
2748  */
2749 /* See <arm/arm32/pmap.h> */
2750 
2751 /*
2752  * pmap_is_referenced:
2753  *
2754  *	Test if a page has the "referenced" attribute.
2755  */
2756 /* See <arm/arm32/pmap.h> */
2757 
2758 int
2759 pmap_fault_fixup(pmap_t pm, vaddr_t va, vm_prot_t ftype, int user)
2760 {
2761 	struct l2_dtable *l2;
2762 	struct l2_bucket *l2b;
2763 	pd_entry_t *pl1pd, l1pd;
2764 	pt_entry_t *ptep, pte;
2765 	paddr_t pa;
2766 	u_int l1idx;
2767 	int rv = 0;
2768 
2769 	PMAP_MAP_TO_HEAD_LOCK();
2770 	pmap_acquire_pmap_lock(pm);
2771 
2772 	l1idx = L1_IDX(va);
2773 
2774 	/*
2775 	 * If there is no l2_dtable for this address, then the process
2776 	 * has no business accessing it.
2777 	 *
2778 	 * Note: This will catch userland processes trying to access
2779 	 * kernel addresses.
2780 	 */
2781 	l2 = pm->pm_l2[L2_IDX(l1idx)];
2782 	if (l2 == NULL)
2783 		goto out;
2784 
2785 	/*
2786 	 * Likewise if there is no L2 descriptor table
2787 	 */
2788 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
2789 	if (l2b->l2b_kva == NULL)
2790 		goto out;
2791 
2792 	/*
2793 	 * Check the PTE itself.
2794 	 */
2795 	ptep = &l2b->l2b_kva[l2pte_index(va)];
2796 	pte = *ptep;
2797 	if (pte == 0)
2798 		goto out;
2799 
2800 	/*
2801 	 * Catch a userland access to the vector page mapped at 0x0
2802 	 */
2803 	if (user && (pte & L2_S_PROT_U) == 0)
2804 		goto out;
2805 
2806 	pa = l2pte_pa(pte);
2807 
2808 	if ((ftype & VM_PROT_WRITE) && (pte & L2_S_PROT_W) == 0) {
2809 		/*
2810 		 * This looks like a good candidate for "page modified"
2811 		 * emulation...
2812 		 */
2813 		struct pv_entry *pv;
2814 		struct vm_page *pg;
2815 
2816 		/* Extract the physical address of the page */
2817 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
2818 			goto out;
2819 
2820 		/* Get the current flags for this page. */
2821 		simple_lock(&pg->mdpage.pvh_slock);
2822 
2823 		pv = pmap_find_pv(pg, pm, va);
2824 		if (pv == NULL) {
2825 	    		simple_unlock(&pg->mdpage.pvh_slock);
2826 			goto out;
2827 		}
2828 
2829 		/*
2830 		 * Do the flags say this page is writable? If not then it
2831 		 * is a genuine write fault. If yes then the write fault is
2832 		 * our fault as we did not reflect the write access in the
2833 		 * PTE. Now we know a write has occurred we can correct this
2834 		 * and also set the modified bit
2835 		 */
2836 		if ((pv->pv_flags & PVF_WRITE) == 0) {
2837 		    	simple_unlock(&pg->mdpage.pvh_slock);
2838 			goto out;
2839 		}
2840 
2841 		NPDEBUG(PDB_FOLLOW,
2842 		    printf("pmap_fault_fixup: mod emul. pm %p, va 0x%08lx, pa 0x%08lx\n",
2843 		    pm, va, VM_PAGE_TO_PHYS(pg)));
2844 
2845 		pg->mdpage.pvh_attrs |= PVF_REF | PVF_MOD;
2846 		pv->pv_flags |= PVF_REF | PVF_MOD;
2847 		simple_unlock(&pg->mdpage.pvh_slock);
2848 
2849 		/*
2850 		 * Re-enable write permissions for the page.  No need to call
2851 		 * pmap_vac_me_harder(), since this is just a
2852 		 * modified-emulation fault, and the PVF_WRITE bit isn't
2853 		 * changing. We've already set the cacheable bits based on
2854 		 * the assumption that we can write to this page.
2855 		 */
2856 		*ptep = (pte & ~L2_TYPE_MASK) | L2_S_PROTO | L2_S_PROT_W;
2857 		PTE_SYNC(ptep);
2858 		rv = 1;
2859 	} else
2860 	if ((pte & L2_TYPE_MASK) == L2_TYPE_INV) {
2861 		/*
2862 		 * This looks like a good candidate for "page referenced"
2863 		 * emulation.
2864 		 */
2865 		struct pv_entry *pv;
2866 		struct vm_page *pg;
2867 
2868 		/* Extract the physical address of the page */
2869 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
2870 			goto out;
2871 
2872 		/* Get the current flags for this page. */
2873 		simple_lock(&pg->mdpage.pvh_slock);
2874 
2875 		pv = pmap_find_pv(pg, pm, va);
2876 		if (pv == NULL) {
2877 	    		simple_unlock(&pg->mdpage.pvh_slock);
2878 			goto out;
2879 		}
2880 
2881 		pg->mdpage.pvh_attrs |= PVF_REF;
2882 		pv->pv_flags |= PVF_REF;
2883 		simple_unlock(&pg->mdpage.pvh_slock);
2884 
2885 		NPDEBUG(PDB_FOLLOW,
2886 		    printf("pmap_fault_fixup: ref emul. pm %p, va 0x%08lx, pa 0x%08lx\n",
2887 		    pm, va, VM_PAGE_TO_PHYS(pg)));
2888 
2889 		*ptep = (pte & ~L2_TYPE_MASK) | L2_S_PROTO;
2890 		PTE_SYNC(ptep);
2891 		rv = 1;
2892 	}
2893 
2894 	/*
2895 	 * We know there is a valid mapping here, so simply
2896 	 * fix up the L1 if necessary.
2897 	 */
2898 	pl1pd = &pm->pm_l1->l1_kva[l1idx];
2899 	l1pd = l2b->l2b_phys | L1_C_DOM(pm->pm_domain) | L1_C_PROTO;
2900 	if (*pl1pd != l1pd) {
2901 		*pl1pd = l1pd;
2902 		PTE_SYNC(pl1pd);
2903 		rv = 1;
2904 	}
2905 
2906 #ifdef CPU_SA110
2907 	/*
2908 	 * There are bugs in the rev K SA110.  This is a check for one
2909 	 * of them.
2910 	 */
2911 	if (rv == 0 && curcpu()->ci_arm_cputype == CPU_ID_SA110 &&
2912 	    curcpu()->ci_arm_cpurev < 3) {
2913 		/* Always current pmap */
2914 		if (l2pte_valid(pte)) {
2915 			extern int kernel_debug;
2916 			if (kernel_debug & 1) {
2917 				struct proc *p = curlwp->l_proc;
2918 				printf("prefetch_abort: page is already "
2919 				    "mapped - pte=%p *pte=%08x\n", ptep, pte);
2920 				printf("prefetch_abort: pc=%08lx proc=%p "
2921 				    "process=%s\n", va, p, p->p_comm);
2922 				printf("prefetch_abort: far=%08x fs=%x\n",
2923 				    cpu_faultaddress(), cpu_faultstatus());
2924 			}
2925 #ifdef DDB
2926 			if (kernel_debug & 2)
2927 				Debugger();
2928 #endif
2929 			rv = 1;
2930 		}
2931 	}
2932 #endif /* CPU_SA110 */
2933 
2934 #ifdef DEBUG
2935 	/*
2936 	 * If 'rv == 0' at this point, it generally indicates that there is a
2937 	 * stale TLB entry for the faulting address. This happens when two or
2938 	 * more processes are sharing an L1. Since we don't flush the TLB on
2939 	 * a context switch between such processes, we can take domain faults
2940 	 * for mappings which exist at the same VA in both processes. EVEN IF
2941 	 * WE'VE RECENTLY FIXED UP THE CORRESPONDING L1 in pmap_enter(), for
2942 	 * example.
2943 	 *
2944 	 * This is extremely likely to happen if pmap_enter() updated the L1
2945 	 * entry for a recently entered mapping. In this case, the TLB is
2946 	 * flushed for the new mapping, but there may still be TLB entries for
2947 	 * other mappings belonging to other processes in the 1MB range
2948 	 * covered by the L1 entry.
2949 	 *
2950 	 * Since 'rv == 0', we know that the L1 already contains the correct
2951 	 * value, so the fault must be due to a stale TLB entry.
2952 	 *
2953 	 * Since we always need to flush the TLB anyway in the case where we
2954 	 * fixed up the L1, or frobbed the L2 PTE, we effectively deal with
2955 	 * stale TLB entries dynamically.
2956 	 *
2957 	 * However, the above condition can ONLY happen if the current L1 is
2958 	 * being shared. If it happens when the L1 is unshared, it indicates
2959 	 * that other parts of the pmap are not doing their job WRT managing
2960 	 * the TLB.
2961 	 */
2962 	if (rv == 0 && pm->pm_l1->l1_domain_use_count == 1) {
2963 		extern int last_fault_code;
2964 		printf("fixup: pm %p, va 0x%lx, ftype %d - nothing to do!\n",
2965 		    pm, va, ftype);
2966 		printf("fixup: l2 %p, l2b %p, ptep %p, pl1pd %p\n",
2967 		    l2, l2b, ptep, pl1pd);
2968 		printf("fixup: pte 0x%x, l1pd 0x%x, last code 0x%x\n",
2969 		    pte, l1pd, last_fault_code);
2970 #ifdef DDB
2971 		Debugger();
2972 #endif
2973 	}
2974 #endif
2975 
2976 	cpu_tlb_flushID_SE(va);
2977 	cpu_cpwait();
2978 
2979 	rv = 1;
2980 
2981 out:
2982 	pmap_release_pmap_lock(pm);
2983 	PMAP_MAP_TO_HEAD_UNLOCK();
2984 
2985 	return (rv);
2986 }
2987 
2988 /*
2989  * pmap_collect: free resources held by a pmap
2990  *
2991  * => optional function.
2992  * => called when a process is swapped out to free memory.
2993  */
2994 void
2995 pmap_collect(pmap_t pm)
2996 {
2997 
2998 	pmap_idcache_wbinv_all(pm);
2999 	pm->pm_remove_all = true;
3000 	pmap_do_remove(pm, VM_MIN_ADDRESS, VM_MAX_ADDRESS, 1);
3001 	pmap_update(pm);
3002 }
3003 
3004 /*
3005  * Routine:	pmap_procwr
3006  *
3007  * Function:
3008  *	Synchronize caches corresponding to [addr, addr+len) in p.
3009  *
3010  */
3011 void
3012 pmap_procwr(struct proc *p, vaddr_t va, int len)
3013 {
3014 	/* We only need to do anything if it is the current process. */
3015 	if (p == curproc)
3016 		cpu_icache_sync_range(va, len);
3017 }
3018 
3019 /*
3020  * Routine:	pmap_unwire
3021  * Function:	Clear the wired attribute for a map/virtual-address pair.
3022  *
3023  * In/out conditions:
3024  *		The mapping must already exist in the pmap.
3025  */
3026 void
3027 pmap_unwire(pmap_t pm, vaddr_t va)
3028 {
3029 	struct l2_bucket *l2b;
3030 	pt_entry_t *ptep, pte;
3031 	struct vm_page *pg;
3032 	paddr_t pa;
3033 
3034 	NPDEBUG(PDB_WIRING, printf("pmap_unwire: pm %p, va 0x%08lx\n", pm, va));
3035 
3036 	PMAP_MAP_TO_HEAD_LOCK();
3037 	pmap_acquire_pmap_lock(pm);
3038 
3039 	l2b = pmap_get_l2_bucket(pm, va);
3040 	KDASSERT(l2b != NULL);
3041 
3042 	ptep = &l2b->l2b_kva[l2pte_index(va)];
3043 	pte = *ptep;
3044 
3045 	/* Extract the physical address of the page */
3046 	pa = l2pte_pa(pte);
3047 
3048 	if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
3049 		/* Update the wired bit in the pv entry for this page. */
3050 		simple_lock(&pg->mdpage.pvh_slock);
3051 		(void) pmap_modify_pv(pg, pm, va, PVF_WIRED, 0);
3052 		simple_unlock(&pg->mdpage.pvh_slock);
3053 	}
3054 
3055 	pmap_release_pmap_lock(pm);
3056 	PMAP_MAP_TO_HEAD_UNLOCK();
3057 }
3058 
3059 void
3060 pmap_switch(struct lwp *olwp, struct lwp *nlwp)
3061 {
3062 	extern int block_userspace_access;
3063 	pmap_t opm, npm, rpm;
3064 	uint32_t odacr, ndacr;
3065 	int oldirqstate;
3066 
3067 	npm = nlwp->l_proc->p_vmspace->vm_map.pmap;
3068 	ndacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2)) |
3069 	    (DOMAIN_CLIENT << (npm->pm_domain * 2));
3070 
3071 	/*
3072 	 * If TTB and DACR are unchanged, short-circuit all the
3073 	 * TLB/cache management stuff.
3074 	 */
3075 	if (olwp != NULL) {
3076 		opm = olwp->l_proc->p_vmspace->vm_map.pmap;
3077 		odacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2)) |
3078 		    (DOMAIN_CLIENT << (opm->pm_domain * 2));
3079 
3080 		if (opm->pm_l1 == npm->pm_l1 && odacr == ndacr)
3081 			goto all_done;
3082 	} else
3083 		opm = NULL;
3084 
3085 	block_userspace_access = 1;
3086 
3087 	/*
3088 	 * If switching to a user vmspace which is different to the
3089 	 * most recent one, and the most recent one is potentially
3090 	 * live in the cache, we must write-back and invalidate the
3091 	 * entire cache.
3092 	 */
3093 	rpm = pmap_recent_user;
3094 	if (npm != pmap_kernel() && rpm && npm != rpm &&
3095 	    rpm->pm_cstate.cs_cache) {
3096 		rpm->pm_cstate.cs_cache = 0;
3097 		cpu_idcache_wbinv_all();
3098 	}
3099 
3100 	/* No interrupts while we frob the TTB/DACR */
3101 	oldirqstate = disable_interrupts(I32_bit | F32_bit);
3102 
3103 	/*
3104 	 * For ARM_VECTORS_LOW, we MUST, I repeat, MUST fix up the L1
3105 	 * entry corresponding to 'vector_page' in the incoming L1 table
3106 	 * before switching to it otherwise subsequent interrupts/exceptions
3107 	 * (including domain faults!) will jump into hyperspace.
3108 	 */
3109 	if (npm->pm_pl1vec != NULL) {
3110 		cpu_tlb_flushID_SE((u_int)vector_page);
3111 		cpu_cpwait();
3112 		*npm->pm_pl1vec = npm->pm_l1vec;
3113 		PTE_SYNC(npm->pm_pl1vec);
3114 	}
3115 
3116 	cpu_domains(ndacr);
3117 
3118 	if (npm == pmap_kernel() || npm == rpm) {
3119 		/*
3120 		 * Switching to a kernel thread, or back to the
3121 		 * same user vmspace as before... Simply update
3122 		 * the TTB (no TLB flush required)
3123 		 */
3124 		__asm volatile("mcr p15, 0, %0, c2, c0, 0" ::
3125 		    "r"(npm->pm_l1->l1_physaddr));
3126 		cpu_cpwait();
3127 	} else {
3128 		/*
3129 		 * Otherwise, update TTB and flush TLB
3130 		 */
3131 		cpu_context_switch(npm->pm_l1->l1_physaddr);
3132 		if (rpm != NULL)
3133 			rpm->pm_cstate.cs_tlb = 0;
3134 	}
3135 
3136 	restore_interrupts(oldirqstate);
3137 
3138 	block_userspace_access = 0;
3139 
3140  all_done:
3141 	/*
3142 	 * The new pmap is resident. Make sure it's marked
3143 	 * as resident in the cache/TLB.
3144 	 */
3145 	npm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
3146 	if (npm != pmap_kernel())
3147 		pmap_recent_user = npm;
3148 
3149 	/* The old pmap is not longer active */
3150 	if (opm != NULL)
3151 		opm->pm_activated = false;
3152 
3153 	/* But the new one is */
3154 	npm->pm_activated = true;
3155 }
3156 
3157 void
3158 pmap_activate(struct lwp *l)
3159 {
3160 
3161 	if (l == curlwp &&
3162 	    l->l_proc->p_vmspace->vm_map.pmap->pm_activated == false)
3163 		pmap_switch(NULL, l);
3164 }
3165 
3166 void
3167 pmap_deactivate(struct lwp *l)
3168 {
3169 
3170 	l->l_proc->p_vmspace->vm_map.pmap->pm_activated = false;
3171 }
3172 
3173 void
3174 pmap_update(pmap_t pm)
3175 {
3176 
3177 	if (pm->pm_remove_all) {
3178 		/*
3179 		 * Finish up the pmap_remove_all() optimisation by flushing
3180 		 * the TLB.
3181 		 */
3182 		pmap_tlb_flushID(pm);
3183 		pm->pm_remove_all = false;
3184 	}
3185 
3186 	if (pmap_is_current(pm)) {
3187 		/*
3188 		 * If we're dealing with a current userland pmap, move its L1
3189 		 * to the end of the LRU.
3190 		 */
3191 		if (pm != pmap_kernel())
3192 			pmap_use_l1(pm);
3193 
3194 		/*
3195 		 * We can assume we're done with frobbing the cache/tlb for
3196 		 * now. Make sure any future pmap ops don't skip cache/tlb
3197 		 * flushes.
3198 		 */
3199 		pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
3200 	}
3201 
3202 	/*
3203 	 * make sure TLB/cache operations have completed.
3204 	 */
3205 	cpu_cpwait();
3206 }
3207 
3208 void
3209 pmap_remove_all(pmap_t pm)
3210 {
3211 
3212 	/*
3213 	 * The vmspace described by this pmap is about to be torn down.
3214 	 * Until pmap_update() is called, UVM will only make calls
3215 	 * to pmap_remove(). We can make life much simpler by flushing
3216 	 * the cache now, and deferring TLB invalidation to pmap_update().
3217 	 */
3218 	pmap_idcache_wbinv_all(pm);
3219 	pm->pm_remove_all = true;
3220 }
3221 
3222 /*
3223  * Retire the given physical map from service.
3224  * Should only be called if the map contains no valid mappings.
3225  */
3226 void
3227 pmap_destroy(pmap_t pm)
3228 {
3229 	u_int count;
3230 
3231 	if (pm == NULL)
3232 		return;
3233 
3234 	if (pm->pm_remove_all) {
3235 		pmap_tlb_flushID(pm);
3236 		pm->pm_remove_all = false;
3237 	}
3238 
3239 	/*
3240 	 * Drop reference count
3241 	 */
3242 	simple_lock(&pm->pm_lock);
3243 	count = --pm->pm_obj.uo_refs;
3244 	simple_unlock(&pm->pm_lock);
3245 	if (count > 0) {
3246 		if (pmap_is_current(pm)) {
3247 			if (pm != pmap_kernel())
3248 				pmap_use_l1(pm);
3249 			pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
3250 		}
3251 		return;
3252 	}
3253 
3254 	/*
3255 	 * reference count is zero, free pmap resources and then free pmap.
3256 	 */
3257 
3258 	if (vector_page < KERNEL_BASE) {
3259 		KDASSERT(!pmap_is_current(pm));
3260 
3261 		/* Remove the vector page mapping */
3262 		pmap_remove(pm, vector_page, vector_page + PAGE_SIZE);
3263 		pmap_update(pm);
3264 	}
3265 
3266 	LIST_REMOVE(pm, pm_list);
3267 
3268 	pmap_free_l1(pm);
3269 
3270 	if (pmap_recent_user == pm)
3271 		pmap_recent_user = NULL;
3272 
3273 	/* return the pmap to the pool */
3274 	pool_cache_put(&pmap_cache, pm);
3275 }
3276 
3277 
3278 /*
3279  * void pmap_reference(pmap_t pm)
3280  *
3281  * Add a reference to the specified pmap.
3282  */
3283 void
3284 pmap_reference(pmap_t pm)
3285 {
3286 
3287 	if (pm == NULL)
3288 		return;
3289 
3290 	pmap_use_l1(pm);
3291 
3292 	simple_lock(&pm->pm_lock);
3293 	pm->pm_obj.uo_refs++;
3294 	simple_unlock(&pm->pm_lock);
3295 }
3296 
3297 /*
3298  * pmap_zero_page()
3299  *
3300  * Zero a given physical page by mapping it at a page hook point.
3301  * In doing the zero page op, the page we zero is mapped cachable, as with
3302  * StrongARM accesses to non-cached pages are non-burst making writing
3303  * _any_ bulk data very slow.
3304  */
3305 #if (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0
3306 void
3307 pmap_zero_page_generic(paddr_t phys)
3308 {
3309 #ifdef DEBUG
3310 	struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
3311 
3312 	if (pg->mdpage.pvh_list != NULL)
3313 		panic("pmap_zero_page: page has mappings");
3314 #endif
3315 
3316 	KDASSERT((phys & PGOFSET) == 0);
3317 
3318 	/*
3319 	 * Hook in the page, zero it, and purge the cache for that
3320 	 * zeroed page. Invalidate the TLB as needed.
3321 	 */
3322 	*cdst_pte = L2_S_PROTO | phys |
3323 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
3324 	PTE_SYNC(cdst_pte);
3325 	cpu_tlb_flushD_SE(cdstp);
3326 	cpu_cpwait();
3327 	bzero_page(cdstp);
3328 	cpu_dcache_wbinv_range(cdstp, PAGE_SIZE);
3329 }
3330 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 */
3331 
3332 #if ARM_MMU_XSCALE == 1
3333 void
3334 pmap_zero_page_xscale(paddr_t phys)
3335 {
3336 #ifdef DEBUG
3337 	struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
3338 
3339 	if (pg->mdpage.pvh_list != NULL)
3340 		panic("pmap_zero_page: page has mappings");
3341 #endif
3342 
3343 	KDASSERT((phys & PGOFSET) == 0);
3344 
3345 	/*
3346 	 * Hook in the page, zero it, and purge the cache for that
3347 	 * zeroed page. Invalidate the TLB as needed.
3348 	 */
3349 	*cdst_pte = L2_S_PROTO | phys |
3350 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
3351 	    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);	/* mini-data */
3352 	PTE_SYNC(cdst_pte);
3353 	cpu_tlb_flushD_SE(cdstp);
3354 	cpu_cpwait();
3355 	bzero_page(cdstp);
3356 	xscale_cache_clean_minidata();
3357 }
3358 #endif /* ARM_MMU_XSCALE == 1 */
3359 
3360 /* pmap_pageidlezero()
3361  *
3362  * The same as above, except that we assume that the page is not
3363  * mapped.  This means we never have to flush the cache first.  Called
3364  * from the idle loop.
3365  */
3366 bool
3367 pmap_pageidlezero(paddr_t phys)
3368 {
3369 	unsigned int i;
3370 	int *ptr;
3371 	bool rv = true;
3372 #ifdef DEBUG
3373 	struct vm_page *pg;
3374 
3375 	pg = PHYS_TO_VM_PAGE(phys);
3376 	if (pg->mdpage.pvh_list != NULL)
3377 		panic("pmap_pageidlezero: page has mappings");
3378 #endif
3379 
3380 	KDASSERT((phys & PGOFSET) == 0);
3381 
3382 	/*
3383 	 * Hook in the page, zero it, and purge the cache for that
3384 	 * zeroed page. Invalidate the TLB as needed.
3385 	 */
3386 	*cdst_pte = L2_S_PROTO | phys |
3387 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
3388 	PTE_SYNC(cdst_pte);
3389 	cpu_tlb_flushD_SE(cdstp);
3390 	cpu_cpwait();
3391 
3392 	for (i = 0, ptr = (int *)cdstp;
3393 			i < (PAGE_SIZE / sizeof(int)); i++) {
3394 		if (sched_curcpu_runnable_p()) {
3395 			/*
3396 			 * A process has become ready.  Abort now,
3397 			 * so we don't keep it waiting while we
3398 			 * do slow memory access to finish this
3399 			 * page.
3400 			 */
3401 			rv = false;
3402 			break;
3403 		}
3404 		*ptr++ = 0;
3405 	}
3406 
3407 	if (rv)
3408 		/*
3409 		 * if we aborted we'll rezero this page again later so don't
3410 		 * purge it unless we finished it
3411 		 */
3412 		cpu_dcache_wbinv_range(cdstp, PAGE_SIZE);
3413 
3414 	return (rv);
3415 }
3416 
3417 /*
3418  * pmap_copy_page()
3419  *
3420  * Copy one physical page into another, by mapping the pages into
3421  * hook points. The same comment regarding cachability as in
3422  * pmap_zero_page also applies here.
3423  */
3424 #if (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0
3425 void
3426 pmap_copy_page_generic(paddr_t src, paddr_t dst)
3427 {
3428 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
3429 #ifdef DEBUG
3430 	struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst);
3431 
3432 	if (dst_pg->mdpage.pvh_list != NULL)
3433 		panic("pmap_copy_page: dst page has mappings");
3434 #endif
3435 
3436 	KDASSERT((src & PGOFSET) == 0);
3437 	KDASSERT((dst & PGOFSET) == 0);
3438 
3439 	/*
3440 	 * Clean the source page.  Hold the source page's lock for
3441 	 * the duration of the copy so that no other mappings can
3442 	 * be created while we have a potentially aliased mapping.
3443 	 */
3444 	simple_lock(&src_pg->mdpage.pvh_slock);
3445 	(void) pmap_clean_page(src_pg->mdpage.pvh_list, true);
3446 
3447 	/*
3448 	 * Map the pages into the page hook points, copy them, and purge
3449 	 * the cache for the appropriate page. Invalidate the TLB
3450 	 * as required.
3451 	 */
3452 	*csrc_pte = L2_S_PROTO | src |
3453 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ) | pte_l2_s_cache_mode;
3454 	PTE_SYNC(csrc_pte);
3455 	*cdst_pte = L2_S_PROTO | dst |
3456 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
3457 	PTE_SYNC(cdst_pte);
3458 	cpu_tlb_flushD_SE(csrcp);
3459 	cpu_tlb_flushD_SE(cdstp);
3460 	cpu_cpwait();
3461 	bcopy_page(csrcp, cdstp);
3462 	cpu_dcache_inv_range(csrcp, PAGE_SIZE);
3463 	simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
3464 	cpu_dcache_wbinv_range(cdstp, PAGE_SIZE);
3465 }
3466 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 */
3467 
3468 #if ARM_MMU_XSCALE == 1
3469 void
3470 pmap_copy_page_xscale(paddr_t src, paddr_t dst)
3471 {
3472 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
3473 #ifdef DEBUG
3474 	struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst);
3475 
3476 	if (dst_pg->mdpage.pvh_list != NULL)
3477 		panic("pmap_copy_page: dst page has mappings");
3478 #endif
3479 
3480 	KDASSERT((src & PGOFSET) == 0);
3481 	KDASSERT((dst & PGOFSET) == 0);
3482 
3483 	/*
3484 	 * Clean the source page.  Hold the source page's lock for
3485 	 * the duration of the copy so that no other mappings can
3486 	 * be created while we have a potentially aliased mapping.
3487 	 */
3488 	simple_lock(&src_pg->mdpage.pvh_slock);
3489 	(void) pmap_clean_page(src_pg->mdpage.pvh_list, true);
3490 
3491 	/*
3492 	 * Map the pages into the page hook points, copy them, and purge
3493 	 * the cache for the appropriate page. Invalidate the TLB
3494 	 * as required.
3495 	 */
3496 	*csrc_pte = L2_S_PROTO | src |
3497 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
3498 	    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);	/* mini-data */
3499 	PTE_SYNC(csrc_pte);
3500 	*cdst_pte = L2_S_PROTO | dst |
3501 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
3502 	    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);	/* mini-data */
3503 	PTE_SYNC(cdst_pte);
3504 	cpu_tlb_flushD_SE(csrcp);
3505 	cpu_tlb_flushD_SE(cdstp);
3506 	cpu_cpwait();
3507 	bcopy_page(csrcp, cdstp);
3508 	simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
3509 	xscale_cache_clean_minidata();
3510 }
3511 #endif /* ARM_MMU_XSCALE == 1 */
3512 
3513 /*
3514  * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
3515  *
3516  * Return the start and end addresses of the kernel's virtual space.
3517  * These values are setup in pmap_bootstrap and are updated as pages
3518  * are allocated.
3519  */
3520 void
3521 pmap_virtual_space(vaddr_t *start, vaddr_t *end)
3522 {
3523 	*start = virtual_avail;
3524 	*end = virtual_end;
3525 }
3526 
3527 /*
3528  * Helper function for pmap_grow_l2_bucket()
3529  */
3530 static inline int
3531 pmap_grow_map(vaddr_t va, pt_entry_t cache_mode, paddr_t *pap)
3532 {
3533 	struct l2_bucket *l2b;
3534 	pt_entry_t *ptep;
3535 	paddr_t pa;
3536 
3537 	if (uvm.page_init_done == false) {
3538 		if (uvm_page_physget(&pa) == false)
3539 			return (1);
3540 	} else {
3541 		struct vm_page *pg;
3542 		pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE);
3543 		if (pg == NULL)
3544 			return (1);
3545 		pa = VM_PAGE_TO_PHYS(pg);
3546 	}
3547 
3548 	if (pap)
3549 		*pap = pa;
3550 
3551 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
3552 	KDASSERT(l2b != NULL);
3553 
3554 	ptep = &l2b->l2b_kva[l2pte_index(va)];
3555 	*ptep = L2_S_PROTO | pa | cache_mode |
3556 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ | VM_PROT_WRITE);
3557 	PTE_SYNC(ptep);
3558 	memset((void *)va, 0, PAGE_SIZE);
3559 	return (0);
3560 }
3561 
3562 /*
3563  * This is the same as pmap_alloc_l2_bucket(), except that it is only
3564  * used by pmap_growkernel().
3565  */
3566 static inline struct l2_bucket *
3567 pmap_grow_l2_bucket(pmap_t pm, vaddr_t va)
3568 {
3569 	struct l2_dtable *l2;
3570 	struct l2_bucket *l2b;
3571 	u_short l1idx;
3572 	vaddr_t nva;
3573 
3574 	l1idx = L1_IDX(va);
3575 
3576 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
3577 		/*
3578 		 * No mapping at this address, as there is
3579 		 * no entry in the L1 table.
3580 		 * Need to allocate a new l2_dtable.
3581 		 */
3582 		nva = pmap_kernel_l2dtable_kva;
3583 		if ((nva & PGOFSET) == 0) {
3584 			/*
3585 			 * Need to allocate a backing page
3586 			 */
3587 			if (pmap_grow_map(nva, pte_l2_s_cache_mode, NULL))
3588 				return (NULL);
3589 		}
3590 
3591 		l2 = (struct l2_dtable *)nva;
3592 		nva += sizeof(struct l2_dtable);
3593 
3594 		if ((nva & PGOFSET) < (pmap_kernel_l2dtable_kva & PGOFSET)) {
3595 			/*
3596 			 * The new l2_dtable straddles a page boundary.
3597 			 * Map in another page to cover it.
3598 			 */
3599 			if (pmap_grow_map(nva, pte_l2_s_cache_mode, NULL))
3600 				return (NULL);
3601 		}
3602 
3603 		pmap_kernel_l2dtable_kva = nva;
3604 
3605 		/*
3606 		 * Link it into the parent pmap
3607 		 */
3608 		pm->pm_l2[L2_IDX(l1idx)] = l2;
3609 	}
3610 
3611 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
3612 
3613 	/*
3614 	 * Fetch pointer to the L2 page table associated with the address.
3615 	 */
3616 	if (l2b->l2b_kva == NULL) {
3617 		pt_entry_t *ptep;
3618 
3619 		/*
3620 		 * No L2 page table has been allocated. Chances are, this
3621 		 * is because we just allocated the l2_dtable, above.
3622 		 */
3623 		nva = pmap_kernel_l2ptp_kva;
3624 		ptep = (pt_entry_t *)nva;
3625 		if ((nva & PGOFSET) == 0) {
3626 			/*
3627 			 * Need to allocate a backing page
3628 			 */
3629 			if (pmap_grow_map(nva, pte_l2_s_cache_mode_pt,
3630 			    &pmap_kernel_l2ptp_phys))
3631 				return (NULL);
3632 			PTE_SYNC_RANGE(ptep, PAGE_SIZE / sizeof(pt_entry_t));
3633 		}
3634 
3635 		l2->l2_occupancy++;
3636 		l2b->l2b_kva = ptep;
3637 		l2b->l2b_l1idx = l1idx;
3638 		l2b->l2b_phys = pmap_kernel_l2ptp_phys;
3639 
3640 		pmap_kernel_l2ptp_kva += L2_TABLE_SIZE_REAL;
3641 		pmap_kernel_l2ptp_phys += L2_TABLE_SIZE_REAL;
3642 	}
3643 
3644 	return (l2b);
3645 }
3646 
3647 vaddr_t
3648 pmap_growkernel(vaddr_t maxkvaddr)
3649 {
3650 	pmap_t kpm = pmap_kernel();
3651 	struct l1_ttable *l1;
3652 	struct l2_bucket *l2b;
3653 	pd_entry_t *pl1pd;
3654 	int s;
3655 
3656 	if (maxkvaddr <= pmap_curmaxkvaddr)
3657 		goto out;		/* we are OK */
3658 
3659 	NPDEBUG(PDB_GROWKERN,
3660 	    printf("pmap_growkernel: growing kernel from 0x%lx to 0x%lx\n",
3661 	    pmap_curmaxkvaddr, maxkvaddr));
3662 
3663 	KDASSERT(maxkvaddr <= virtual_end);
3664 
3665 	/*
3666 	 * whoops!   we need to add kernel PTPs
3667 	 */
3668 
3669 	s = splhigh();	/* to be safe */
3670 	simple_lock(&kpm->pm_lock);
3671 
3672 	/* Map 1MB at a time */
3673 	for (; pmap_curmaxkvaddr < maxkvaddr; pmap_curmaxkvaddr += L1_S_SIZE) {
3674 
3675 		l2b = pmap_grow_l2_bucket(kpm, pmap_curmaxkvaddr);
3676 		KDASSERT(l2b != NULL);
3677 
3678 		/* Distribute new L1 entry to all other L1s */
3679 		SLIST_FOREACH(l1, &l1_list, l1_link) {
3680 			pl1pd = &l1->l1_kva[L1_IDX(pmap_curmaxkvaddr)];
3681 			*pl1pd = l2b->l2b_phys | L1_C_DOM(PMAP_DOMAIN_KERNEL) |
3682 			    L1_C_PROTO;
3683 			PTE_SYNC(pl1pd);
3684 		}
3685 	}
3686 
3687 	/*
3688 	 * flush out the cache, expensive but growkernel will happen so
3689 	 * rarely
3690 	 */
3691 	cpu_dcache_wbinv_all();
3692 	cpu_tlb_flushD();
3693 	cpu_cpwait();
3694 
3695 	simple_unlock(&kpm->pm_lock);
3696 	splx(s);
3697 
3698 out:
3699 	return (pmap_curmaxkvaddr);
3700 }
3701 
3702 /************************ Utility routines ****************************/
3703 
3704 /*
3705  * vector_page_setprot:
3706  *
3707  *	Manipulate the protection of the vector page.
3708  */
3709 void
3710 vector_page_setprot(int prot)
3711 {
3712 	struct l2_bucket *l2b;
3713 	pt_entry_t *ptep;
3714 
3715 	l2b = pmap_get_l2_bucket(pmap_kernel(), vector_page);
3716 	KDASSERT(l2b != NULL);
3717 
3718 	ptep = &l2b->l2b_kva[l2pte_index(vector_page)];
3719 
3720 	*ptep = (*ptep & ~L1_S_PROT_MASK) | L2_S_PROT(PTE_KERNEL, prot);
3721 	PTE_SYNC(ptep);
3722 	cpu_tlb_flushD_SE(vector_page);
3723 	cpu_cpwait();
3724 }
3725 
3726 /*
3727  * Fetch pointers to the PDE/PTE for the given pmap/VA pair.
3728  * Returns true if the mapping exists, else false.
3729  *
3730  * NOTE: This function is only used by a couple of arm-specific modules.
3731  * It is not safe to take any pmap locks here, since we could be right
3732  * in the middle of debugging the pmap anyway...
3733  *
3734  * It is possible for this routine to return false even though a valid
3735  * mapping does exist. This is because we don't lock, so the metadata
3736  * state may be inconsistent.
3737  *
3738  * NOTE: We can return a NULL *ptp in the case where the L1 pde is
3739  * a "section" mapping.
3740  */
3741 bool
3742 pmap_get_pde_pte(pmap_t pm, vaddr_t va, pd_entry_t **pdp, pt_entry_t **ptp)
3743 {
3744 	struct l2_dtable *l2;
3745 	pd_entry_t *pl1pd, l1pd;
3746 	pt_entry_t *ptep;
3747 	u_short l1idx;
3748 
3749 	if (pm->pm_l1 == NULL)
3750 		return (false);
3751 
3752 	l1idx = L1_IDX(va);
3753 	*pdp = pl1pd = &pm->pm_l1->l1_kva[l1idx];
3754 	l1pd = *pl1pd;
3755 
3756 	if (l1pte_section_p(l1pd)) {
3757 		*ptp = NULL;
3758 		return (true);
3759 	}
3760 
3761 	if (pm->pm_l2 == NULL)
3762 		return (false);
3763 
3764 	l2 = pm->pm_l2[L2_IDX(l1idx)];
3765 
3766 	if (l2 == NULL ||
3767 	    (ptep = l2->l2_bucket[L2_BUCKET(l1idx)].l2b_kva) == NULL) {
3768 		return (false);
3769 	}
3770 
3771 	*ptp = &ptep[l2pte_index(va)];
3772 	return (true);
3773 }
3774 
3775 bool
3776 pmap_get_pde(pmap_t pm, vaddr_t va, pd_entry_t **pdp)
3777 {
3778 	u_short l1idx;
3779 
3780 	if (pm->pm_l1 == NULL)
3781 		return (false);
3782 
3783 	l1idx = L1_IDX(va);
3784 	*pdp = &pm->pm_l1->l1_kva[l1idx];
3785 
3786 	return (true);
3787 }
3788 
3789 /************************ Bootstrapping routines ****************************/
3790 
3791 static void
3792 pmap_init_l1(struct l1_ttable *l1, pd_entry_t *l1pt)
3793 {
3794 	int i;
3795 
3796 	l1->l1_kva = l1pt;
3797 	l1->l1_domain_use_count = 0;
3798 	l1->l1_domain_first = 0;
3799 
3800 	for (i = 0; i < PMAP_DOMAINS; i++)
3801 		l1->l1_domain_free[i] = i + 1;
3802 
3803 	/*
3804 	 * Copy the kernel's L1 entries to each new L1.
3805 	 */
3806 	if (pmap_initialized)
3807 		memcpy(l1pt, pmap_kernel()->pm_l1->l1_kva, L1_TABLE_SIZE);
3808 
3809 	if (pmap_extract(pmap_kernel(), (vaddr_t)l1pt,
3810 	    &l1->l1_physaddr) == false)
3811 		panic("pmap_init_l1: can't get PA of L1 at %p", l1pt);
3812 
3813 	SLIST_INSERT_HEAD(&l1_list, l1, l1_link);
3814 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
3815 }
3816 
3817 /*
3818  * pmap_bootstrap() is called from the board-specific initarm() routine
3819  * once the kernel L1/L2 descriptors tables have been set up.
3820  *
3821  * This is a somewhat convoluted process since pmap bootstrap is, effectively,
3822  * spread over a number of disparate files/functions.
3823  *
3824  * We are passed the following parameters
3825  *  - kernel_l1pt
3826  *    This is a pointer to the base of the kernel's L1 translation table.
3827  *  - vstart
3828  *    1MB-aligned start of managed kernel virtual memory.
3829  *  - vend
3830  *    1MB-aligned end of managed kernel virtual memory.
3831  *
3832  * We use the first parameter to build the metadata (struct l1_ttable and
3833  * struct l2_dtable) necessary to track kernel mappings.
3834  */
3835 #define	PMAP_STATIC_L2_SIZE 16
3836 void
3837 pmap_bootstrap(pd_entry_t *kernel_l1pt, vaddr_t vstart, vaddr_t vend)
3838 {
3839 	static struct l1_ttable static_l1;
3840 	static struct l2_dtable static_l2[PMAP_STATIC_L2_SIZE];
3841 	struct l1_ttable *l1 = &static_l1;
3842 	struct l2_dtable *l2;
3843 	struct l2_bucket *l2b;
3844 	pmap_t pm = pmap_kernel();
3845 	pd_entry_t pde;
3846 	pt_entry_t *ptep;
3847 	paddr_t pa;
3848 	vaddr_t va;
3849 	vsize_t size;
3850 	int l1idx, l2idx, l2next = 0;
3851 
3852 	/*
3853 	 * Initialise the kernel pmap object
3854 	 */
3855 	pm->pm_l1 = l1;
3856 	pm->pm_domain = PMAP_DOMAIN_KERNEL;
3857 	pm->pm_activated = true;
3858 	pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
3859 	simple_lock_init(&pm->pm_lock);
3860 	pm->pm_obj.pgops = NULL;
3861 	TAILQ_INIT(&pm->pm_obj.memq);
3862 	pm->pm_obj.uo_npages = 0;
3863 	pm->pm_obj.uo_refs = 1;
3864 
3865 	/*
3866 	 * Scan the L1 translation table created by initarm() and create
3867 	 * the required metadata for all valid mappings found in it.
3868 	 */
3869 	for (l1idx = 0; l1idx < (L1_TABLE_SIZE / sizeof(pd_entry_t)); l1idx++) {
3870 		pde = kernel_l1pt[l1idx];
3871 
3872 		/*
3873 		 * We're only interested in Coarse mappings.
3874 		 * pmap_extract() can deal with section mappings without
3875 		 * recourse to checking L2 metadata.
3876 		 */
3877 		if ((pde & L1_TYPE_MASK) != L1_TYPE_C)
3878 			continue;
3879 
3880 		/*
3881 		 * Lookup the KVA of this L2 descriptor table
3882 		 */
3883 		pa = (paddr_t)(pde & L1_C_ADDR_MASK);
3884 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
3885 		if (ptep == NULL) {
3886 			panic("pmap_bootstrap: No L2 for va 0x%x, pa 0x%lx",
3887 			    (u_int)l1idx << L1_S_SHIFT, pa);
3888 		}
3889 
3890 		/*
3891 		 * Fetch the associated L2 metadata structure.
3892 		 * Allocate a new one if necessary.
3893 		 */
3894 		if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
3895 			if (l2next == PMAP_STATIC_L2_SIZE)
3896 				panic("pmap_bootstrap: out of static L2s");
3897 			pm->pm_l2[L2_IDX(l1idx)] = l2 = &static_l2[l2next++];
3898 		}
3899 
3900 		/*
3901 		 * One more L1 slot tracked...
3902 		 */
3903 		l2->l2_occupancy++;
3904 
3905 		/*
3906 		 * Fill in the details of the L2 descriptor in the
3907 		 * appropriate bucket.
3908 		 */
3909 		l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
3910 		l2b->l2b_kva = ptep;
3911 		l2b->l2b_phys = pa;
3912 		l2b->l2b_l1idx = l1idx;
3913 
3914 		/*
3915 		 * Establish an initial occupancy count for this descriptor
3916 		 */
3917 		for (l2idx = 0;
3918 		    l2idx < (L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
3919 		    l2idx++) {
3920 			if ((ptep[l2idx] & L2_TYPE_MASK) != L2_TYPE_INV) {
3921 				l2b->l2b_occupancy++;
3922 			}
3923 		}
3924 
3925 		/*
3926 		 * Make sure the descriptor itself has the correct cache mode.
3927 		 * If not, fix it, but whine about the problem. Port-meisters
3928 		 * should consider this a clue to fix up their initarm()
3929 		 * function. :)
3930 		 */
3931 		if (pmap_set_pt_cache_mode(kernel_l1pt, (vaddr_t)ptep)) {
3932 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
3933 			    "L2 pte @ %p\n", ptep);
3934 		}
3935 	}
3936 
3937 	/*
3938 	 * Ensure the primary (kernel) L1 has the correct cache mode for
3939 	 * a page table. Bitch if it is not correctly set.
3940 	 */
3941 	for (va = (vaddr_t)kernel_l1pt;
3942 	    va < ((vaddr_t)kernel_l1pt + L1_TABLE_SIZE); va += PAGE_SIZE) {
3943 		if (pmap_set_pt_cache_mode(kernel_l1pt, va))
3944 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
3945 			    "primary L1 @ 0x%lx\n", va);
3946 	}
3947 
3948 	cpu_dcache_wbinv_all();
3949 	cpu_tlb_flushID();
3950 	cpu_cpwait();
3951 
3952 	/*
3953 	 * now we allocate the "special" VAs which are used for tmp mappings
3954 	 * by the pmap (and other modules).  we allocate the VAs by advancing
3955 	 * virtual_avail (note that there are no pages mapped at these VAs).
3956 	 *
3957 	 * Managed KVM space start from wherever initarm() tells us.
3958 	 */
3959 	virtual_avail = vstart;
3960 	virtual_end = vend;
3961 
3962 	pmap_alloc_specials(&virtual_avail, 1, &csrcp, &csrc_pte);
3963 	pmap_set_pt_cache_mode(kernel_l1pt, (vaddr_t)csrc_pte);
3964 	pmap_alloc_specials(&virtual_avail, 1, &cdstp, &cdst_pte);
3965 	pmap_set_pt_cache_mode(kernel_l1pt, (vaddr_t)cdst_pte);
3966 	pmap_alloc_specials(&virtual_avail, 1, (void *)&memhook, NULL);
3967 	pmap_alloc_specials(&virtual_avail, round_page(MSGBUFSIZE) / PAGE_SIZE,
3968 	    (void *)&msgbufaddr, NULL);
3969 
3970 	/*
3971 	 * Allocate a range of kernel virtual address space to be used
3972 	 * for L2 descriptor tables and metadata allocation in
3973 	 * pmap_growkernel().
3974 	 */
3975 	size = ((virtual_end - pmap_curmaxkvaddr) + L1_S_OFFSET) / L1_S_SIZE;
3976 	pmap_alloc_specials(&virtual_avail,
3977 	    round_page(size * L2_TABLE_SIZE_REAL) / PAGE_SIZE,
3978 	    &pmap_kernel_l2ptp_kva, NULL);
3979 
3980 	size = (size + (L2_BUCKET_SIZE - 1)) / L2_BUCKET_SIZE;
3981 	pmap_alloc_specials(&virtual_avail,
3982 	    round_page(size * sizeof(struct l2_dtable)) / PAGE_SIZE,
3983 	    &pmap_kernel_l2dtable_kva, NULL);
3984 
3985 	/*
3986 	 * init the static-global locks and global pmap list.
3987 	 */
3988 	/* spinlockinit(&pmap_main_lock, "pmaplk", 0); */
3989 
3990 	/*
3991 	 * We can now initialise the first L1's metadata.
3992 	 */
3993 	SLIST_INIT(&l1_list);
3994 	TAILQ_INIT(&l1_lru_list);
3995 	simple_lock_init(&l1_lru_lock);
3996 	pmap_init_l1(l1, kernel_l1pt);
3997 
3998 	/* Set up vector page L1 details, if necessary */
3999 	if (vector_page < KERNEL_BASE) {
4000 		pm->pm_pl1vec = &pm->pm_l1->l1_kva[L1_IDX(vector_page)];
4001 		l2b = pmap_get_l2_bucket(pm, vector_page);
4002 		pm->pm_l1vec = l2b->l2b_phys | L1_C_PROTO |
4003 		    L1_C_DOM(pm->pm_domain);
4004 	} else
4005 		pm->pm_pl1vec = NULL;
4006 
4007 	/*
4008 	 * Initialize the pmap cache
4009 	 */
4010 	pool_cache_bootstrap(&pmap_cache, sizeof(struct pmap), 0, 0, 0,
4011 	    "pmappl", NULL, IPL_NONE, pmap_pmap_ctor, NULL, NULL);
4012 	LIST_INIT(&pmap_pmaps);
4013 	LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list);
4014 
4015 	/*
4016 	 * Initialize the pv pool.
4017 	 */
4018 	pool_init(&pmap_pv_pool, sizeof(struct pv_entry), 0, 0, 0, "pvepl",
4019 	    &pmap_bootstrap_pv_allocator, IPL_NONE);
4020 
4021 	/*
4022 	 * Initialize the L2 dtable pool and cache.
4023 	 */
4024 	pool_cache_bootstrap(&pmap_l2dtable_cache, sizeof(struct l2_dtable), 0,
4025 	    0, 0, "l2dtblpl", NULL, IPL_NONE, pmap_l2dtable_ctor, NULL, NULL);
4026 
4027 	/*
4028 	 * Initialise the L2 descriptor table pool and cache
4029 	 */
4030 	pool_cache_bootstrap(&pmap_l2ptp_cache, L2_TABLE_SIZE_REAL, 0,
4031 	    L2_TABLE_SIZE_REAL, 0, "l2ptppl", NULL, IPL_NONE,
4032 	    pmap_l2ptp_ctor, NULL, NULL);
4033 
4034 	cpu_dcache_wbinv_all();
4035 }
4036 
4037 static int
4038 pmap_set_pt_cache_mode(pd_entry_t *kl1, vaddr_t va)
4039 {
4040 	pd_entry_t *pdep, pde;
4041 	pt_entry_t *ptep, pte;
4042 	vaddr_t pa;
4043 	int rv = 0;
4044 
4045 	/*
4046 	 * Make sure the descriptor itself has the correct cache mode
4047 	 */
4048 	pdep = &kl1[L1_IDX(va)];
4049 	pde = *pdep;
4050 
4051 	if (l1pte_section_p(pde)) {
4052 		if ((pde & L1_S_CACHE_MASK) != pte_l1_s_cache_mode_pt) {
4053 			*pdep = (pde & ~L1_S_CACHE_MASK) |
4054 			    pte_l1_s_cache_mode_pt;
4055 			PTE_SYNC(pdep);
4056 			cpu_dcache_wbinv_range((vaddr_t)pdep, sizeof(*pdep));
4057 			rv = 1;
4058 		}
4059 	} else {
4060 		pa = (paddr_t)(pde & L1_C_ADDR_MASK);
4061 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
4062 		if (ptep == NULL)
4063 			panic("pmap_bootstrap: No L2 for L2 @ va %p\n", ptep);
4064 
4065 		ptep = &ptep[l2pte_index(va)];
4066 		pte = *ptep;
4067 		if ((pte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
4068 			*ptep = (pte & ~L2_S_CACHE_MASK) |
4069 			    pte_l2_s_cache_mode_pt;
4070 			PTE_SYNC(ptep);
4071 			cpu_dcache_wbinv_range((vaddr_t)ptep, sizeof(*ptep));
4072 			rv = 1;
4073 		}
4074 	}
4075 
4076 	return (rv);
4077 }
4078 
4079 static void
4080 pmap_alloc_specials(vaddr_t *availp, int pages, vaddr_t *vap, pt_entry_t **ptep)
4081 {
4082 	vaddr_t va = *availp;
4083 	struct l2_bucket *l2b;
4084 
4085 	if (ptep) {
4086 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
4087 		if (l2b == NULL)
4088 			panic("pmap_alloc_specials: no l2b for 0x%lx", va);
4089 
4090 		if (ptep)
4091 			*ptep = &l2b->l2b_kva[l2pte_index(va)];
4092 	}
4093 
4094 	*vap = va;
4095 	*availp = va + (PAGE_SIZE * pages);
4096 }
4097 
4098 void
4099 pmap_init(void)
4100 {
4101 	extern int physmem;
4102 
4103 	/*
4104 	 * Set the available memory vars - These do not map to real memory
4105 	 * addresses and cannot as the physical memory is fragmented.
4106 	 * They are used by ps for %mem calculations.
4107 	 * One could argue whether this should be the entire memory or just
4108 	 * the memory that is useable in a user process.
4109 	 */
4110 	avail_start = 0;
4111 	avail_end = physmem * PAGE_SIZE;
4112 
4113 	/*
4114 	 * Now we need to free enough pv_entry structures to allow us to get
4115 	 * the kmem_map/kmem_object allocated and inited (done after this
4116 	 * function is finished).  to do this we allocate one bootstrap page out
4117 	 * of kernel_map and use it to provide an initial pool of pv_entry
4118 	 * structures.   we never free this page.
4119 	 */
4120 	pool_setlowat(&pmap_pv_pool,
4121 	    (PAGE_SIZE / sizeof(struct pv_entry)) * 2);
4122 
4123 	pmap_initialized = true;
4124 }
4125 
4126 static vaddr_t last_bootstrap_page = 0;
4127 static void *free_bootstrap_pages = NULL;
4128 
4129 static void *
4130 pmap_bootstrap_pv_page_alloc(struct pool *pp, int flags)
4131 {
4132 	extern void *pool_page_alloc(struct pool *, int);
4133 	vaddr_t new_page;
4134 	void *rv;
4135 
4136 	if (pmap_initialized)
4137 		return (pool_page_alloc(pp, flags));
4138 
4139 	if (free_bootstrap_pages) {
4140 		rv = free_bootstrap_pages;
4141 		free_bootstrap_pages = *((void **)rv);
4142 		return (rv);
4143 	}
4144 
4145 	new_page = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
4146 	    UVM_KMF_WIRED | ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT));
4147 
4148 	KASSERT(new_page > last_bootstrap_page);
4149 	last_bootstrap_page = new_page;
4150 	return ((void *)new_page);
4151 }
4152 
4153 static void
4154 pmap_bootstrap_pv_page_free(struct pool *pp, void *v)
4155 {
4156 	extern void pool_page_free(struct pool *, void *);
4157 
4158 	if ((vaddr_t)v <= last_bootstrap_page) {
4159 		*((void **)v) = free_bootstrap_pages;
4160 		free_bootstrap_pages = v;
4161 		return;
4162 	}
4163 
4164 	if (pmap_initialized) {
4165 		pool_page_free(pp, v);
4166 		return;
4167 	}
4168 }
4169 
4170 /*
4171  * pmap_postinit()
4172  *
4173  * This routine is called after the vm and kmem subsystems have been
4174  * initialised. This allows the pmap code to perform any initialisation
4175  * that can only be done one the memory allocation is in place.
4176  */
4177 void
4178 pmap_postinit(void)
4179 {
4180 	extern paddr_t physical_start, physical_end;
4181 	struct l2_bucket *l2b;
4182 	struct l1_ttable *l1;
4183 	struct pglist plist;
4184 	struct vm_page *m;
4185 	pd_entry_t *pl1pt;
4186 	pt_entry_t *ptep, pte;
4187 	vaddr_t va, eva;
4188 	u_int loop, needed;
4189 	int error;
4190 
4191 	pool_cache_setlowat(&pmap_l2ptp_cache,
4192 	    (PAGE_SIZE / L2_TABLE_SIZE_REAL) * 4);
4193 	pool_cache_setlowat(&pmap_l2dtable_cache,
4194 	    (PAGE_SIZE / sizeof(struct l2_dtable)) * 2);
4195 
4196 	needed = (maxproc / PMAP_DOMAINS) + ((maxproc % PMAP_DOMAINS) ? 1 : 0);
4197 	needed -= 1;
4198 
4199 	l1 = malloc(sizeof(*l1) * needed, M_VMPMAP, M_WAITOK);
4200 
4201 	for (loop = 0; loop < needed; loop++, l1++) {
4202 		/* Allocate a L1 page table */
4203 		va = uvm_km_alloc(kernel_map, L1_TABLE_SIZE, 0, UVM_KMF_VAONLY);
4204 		if (va == 0)
4205 			panic("Cannot allocate L1 KVM");
4206 
4207 		error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start,
4208 		    physical_end, L1_TABLE_SIZE, 0, &plist, 1, M_WAITOK);
4209 		if (error)
4210 			panic("Cannot allocate L1 physical pages");
4211 
4212 		m = TAILQ_FIRST(&plist);
4213 		eva = va + L1_TABLE_SIZE;
4214 		pl1pt = (pd_entry_t *)va;
4215 
4216 		while (m && va < eva) {
4217 			paddr_t pa = VM_PAGE_TO_PHYS(m);
4218 
4219 			pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE);
4220 
4221 			/*
4222 			 * Make sure the L1 descriptor table is mapped
4223 			 * with the cache-mode set to write-through.
4224 			 */
4225 			l2b = pmap_get_l2_bucket(pmap_kernel(), va);
4226 			ptep = &l2b->l2b_kva[l2pte_index(va)];
4227 			pte = *ptep;
4228 			pte = (pte & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode_pt;
4229 			*ptep = pte;
4230 			PTE_SYNC(ptep);
4231 			cpu_tlb_flushD_SE(va);
4232 
4233 			va += PAGE_SIZE;
4234 			m = TAILQ_NEXT(m, pageq);
4235 		}
4236 
4237 #ifdef DIAGNOSTIC
4238 		if (m)
4239 			panic("pmap_alloc_l1pt: pglist not empty");
4240 #endif	/* DIAGNOSTIC */
4241 
4242 		pmap_init_l1(l1, pl1pt);
4243 	}
4244 
4245 #ifdef DEBUG
4246 	printf("pmap_postinit: Allocated %d static L1 descriptor tables\n",
4247 	    needed);
4248 #endif
4249 }
4250 
4251 /*
4252  * Note that the following routines are used by board-specific initialisation
4253  * code to configure the initial kernel page tables.
4254  *
4255  * If ARM32_NEW_VM_LAYOUT is *not* defined, they operate on the assumption that
4256  * L2 page-table pages are 4KB in size and use 4 L1 slots. This mimics the
4257  * behaviour of the old pmap, and provides an easy migration path for
4258  * initial bring-up of the new pmap on existing ports. Fortunately,
4259  * pmap_bootstrap() compensates for this hackery. This is only a stop-gap and
4260  * will be deprecated.
4261  *
4262  * If ARM32_NEW_VM_LAYOUT *is* defined, these functions deal with 1KB L2 page
4263  * tables.
4264  */
4265 
4266 /*
4267  * This list exists for the benefit of pmap_map_chunk().  It keeps track
4268  * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
4269  * find them as necessary.
4270  *
4271  * Note that the data on this list MUST remain valid after initarm() returns,
4272  * as pmap_bootstrap() uses it to contruct L2 table metadata.
4273  */
4274 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
4275 
4276 static vaddr_t
4277 kernel_pt_lookup(paddr_t pa)
4278 {
4279 	pv_addr_t *pv;
4280 
4281 	SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
4282 #ifndef ARM32_NEW_VM_LAYOUT
4283 		if (pv->pv_pa == (pa & ~PGOFSET))
4284 			return (pv->pv_va | (pa & PGOFSET));
4285 #else
4286 		if (pv->pv_pa == pa)
4287 			return (pv->pv_va);
4288 #endif
4289 	}
4290 	return (0);
4291 }
4292 
4293 /*
4294  * pmap_map_section:
4295  *
4296  *	Create a single section mapping.
4297  */
4298 void
4299 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
4300 {
4301 	pd_entry_t *pde = (pd_entry_t *) l1pt;
4302 	pd_entry_t fl;
4303 
4304 	KASSERT(((va | pa) & L1_S_OFFSET) == 0);
4305 
4306 	switch (cache) {
4307 	case PTE_NOCACHE:
4308 	default:
4309 		fl = 0;
4310 		break;
4311 
4312 	case PTE_CACHE:
4313 		fl = pte_l1_s_cache_mode;
4314 		break;
4315 
4316 	case PTE_PAGETABLE:
4317 		fl = pte_l1_s_cache_mode_pt;
4318 		break;
4319 	}
4320 
4321 	pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
4322 	    L1_S_PROT(PTE_KERNEL, prot) | fl | L1_S_DOM(PMAP_DOMAIN_KERNEL);
4323 	PTE_SYNC(&pde[va >> L1_S_SHIFT]);
4324 }
4325 
4326 /*
4327  * pmap_map_entry:
4328  *
4329  *	Create a single page mapping.
4330  */
4331 void
4332 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
4333 {
4334 	pd_entry_t *pde = (pd_entry_t *) l1pt;
4335 	pt_entry_t fl;
4336 	pt_entry_t *pte;
4337 
4338 	KASSERT(((va | pa) & PGOFSET) == 0);
4339 
4340 	switch (cache) {
4341 	case PTE_NOCACHE:
4342 	default:
4343 		fl = 0;
4344 		break;
4345 
4346 	case PTE_CACHE:
4347 		fl = pte_l2_s_cache_mode;
4348 		break;
4349 
4350 	case PTE_PAGETABLE:
4351 		fl = pte_l2_s_cache_mode_pt;
4352 		break;
4353 	}
4354 
4355 	if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
4356 		panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
4357 
4358 #ifndef ARM32_NEW_VM_LAYOUT
4359 	pte = (pt_entry_t *)
4360 	    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
4361 #else
4362 	pte = (pt_entry_t *) kernel_pt_lookup(pde[L1_IDX(va)] & L1_C_ADDR_MASK);
4363 #endif
4364 	if (pte == NULL)
4365 		panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
4366 
4367 #ifndef ARM32_NEW_VM_LAYOUT
4368 	pte[(va >> PGSHIFT) & 0x3ff] =
4369 	    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | fl;
4370 	PTE_SYNC(&pte[(va >> PGSHIFT) & 0x3ff]);
4371 #else
4372 	pte[l2pte_index(va)] =
4373 	    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | fl;
4374 	PTE_SYNC(&pte[l2pte_index(va)]);
4375 #endif
4376 }
4377 
4378 /*
4379  * pmap_link_l2pt:
4380  *
4381  *	Link the L2 page table specified by "l2pv" into the L1
4382  *	page table at the slot for "va".
4383  */
4384 void
4385 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
4386 {
4387 	pd_entry_t *pde = (pd_entry_t *) l1pt, proto;
4388 	u_int slot = va >> L1_S_SHIFT;
4389 
4390 #ifndef ARM32_NEW_VM_LAYOUT
4391 	KASSERT((va & ((L1_S_SIZE * 4) - 1)) == 0);
4392 	KASSERT((l2pv->pv_pa & PGOFSET) == 0);
4393 #endif
4394 
4395 	proto = L1_S_DOM(PMAP_DOMAIN_KERNEL) | L1_C_PROTO;
4396 
4397 	pde[slot + 0] = proto | (l2pv->pv_pa + 0x000);
4398 #ifdef ARM32_NEW_VM_LAYOUT
4399 	PTE_SYNC(&pde[slot]);
4400 #else
4401 	pde[slot + 1] = proto | (l2pv->pv_pa + 0x400);
4402 	pde[slot + 2] = proto | (l2pv->pv_pa + 0x800);
4403 	pde[slot + 3] = proto | (l2pv->pv_pa + 0xc00);
4404 	PTE_SYNC_RANGE(&pde[slot + 0], 4);
4405 #endif
4406 
4407 	SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
4408 }
4409 
4410 /*
4411  * pmap_map_chunk:
4412  *
4413  *	Map a chunk of memory using the most efficient mappings
4414  *	possible (section, large page, small page) into the
4415  *	provided L1 and L2 tables at the specified virtual address.
4416  */
4417 vsize_t
4418 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
4419     int prot, int cache)
4420 {
4421 	pd_entry_t *pde = (pd_entry_t *) l1pt;
4422 	pt_entry_t *pte, f1, f2s, f2l;
4423 	vsize_t resid;
4424 	int i;
4425 
4426 	resid = (size + (PAGE_SIZE - 1)) & ~(PAGE_SIZE - 1);
4427 
4428 	if (l1pt == 0)
4429 		panic("pmap_map_chunk: no L1 table provided");
4430 
4431 #ifdef VERBOSE_INIT_ARM
4432 	printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
4433 	    "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
4434 #endif
4435 
4436 	switch (cache) {
4437 	case PTE_NOCACHE:
4438 	default:
4439 		f1 = 0;
4440 		f2l = 0;
4441 		f2s = 0;
4442 		break;
4443 
4444 	case PTE_CACHE:
4445 		f1 = pte_l1_s_cache_mode;
4446 		f2l = pte_l2_l_cache_mode;
4447 		f2s = pte_l2_s_cache_mode;
4448 		break;
4449 
4450 	case PTE_PAGETABLE:
4451 		f1 = pte_l1_s_cache_mode_pt;
4452 		f2l = pte_l2_l_cache_mode_pt;
4453 		f2s = pte_l2_s_cache_mode_pt;
4454 		break;
4455 	}
4456 
4457 	size = resid;
4458 
4459 	while (resid > 0) {
4460 		/* See if we can use a section mapping. */
4461 		if (L1_S_MAPPABLE_P(va, pa, resid)) {
4462 #ifdef VERBOSE_INIT_ARM
4463 			printf("S");
4464 #endif
4465 			pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
4466 			    L1_S_PROT(PTE_KERNEL, prot) | f1 |
4467 			    L1_S_DOM(PMAP_DOMAIN_KERNEL);
4468 			PTE_SYNC(&pde[va >> L1_S_SHIFT]);
4469 			va += L1_S_SIZE;
4470 			pa += L1_S_SIZE;
4471 			resid -= L1_S_SIZE;
4472 			continue;
4473 		}
4474 
4475 		/*
4476 		 * Ok, we're going to use an L2 table.  Make sure
4477 		 * one is actually in the corresponding L1 slot
4478 		 * for the current VA.
4479 		 */
4480 		if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
4481 			panic("pmap_map_chunk: no L2 table for VA 0x%08lx", va);
4482 
4483 #ifndef ARM32_NEW_VM_LAYOUT
4484 		pte = (pt_entry_t *)
4485 		    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
4486 #else
4487 		pte = (pt_entry_t *) kernel_pt_lookup(
4488 		    pde[L1_IDX(va)] & L1_C_ADDR_MASK);
4489 #endif
4490 		if (pte == NULL)
4491 			panic("pmap_map_chunk: can't find L2 table for VA"
4492 			    "0x%08lx", va);
4493 
4494 		/* See if we can use a L2 large page mapping. */
4495 		if (L2_L_MAPPABLE_P(va, pa, resid)) {
4496 #ifdef VERBOSE_INIT_ARM
4497 			printf("L");
4498 #endif
4499 			for (i = 0; i < 16; i++) {
4500 #ifndef ARM32_NEW_VM_LAYOUT
4501 				pte[((va >> PGSHIFT) & 0x3f0) + i] =
4502 				    L2_L_PROTO | pa |
4503 				    L2_L_PROT(PTE_KERNEL, prot) | f2l;
4504 				PTE_SYNC(&pte[((va >> PGSHIFT) & 0x3f0) + i]);
4505 #else
4506 				pte[l2pte_index(va) + i] =
4507 				    L2_L_PROTO | pa |
4508 				    L2_L_PROT(PTE_KERNEL, prot) | f2l;
4509 				PTE_SYNC(&pte[l2pte_index(va) + i]);
4510 #endif
4511 			}
4512 			va += L2_L_SIZE;
4513 			pa += L2_L_SIZE;
4514 			resid -= L2_L_SIZE;
4515 			continue;
4516 		}
4517 
4518 		/* Use a small page mapping. */
4519 #ifdef VERBOSE_INIT_ARM
4520 		printf("P");
4521 #endif
4522 #ifndef ARM32_NEW_VM_LAYOUT
4523 		pte[(va >> PGSHIFT) & 0x3ff] =
4524 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | f2s;
4525 		PTE_SYNC(&pte[(va >> PGSHIFT) & 0x3ff]);
4526 #else
4527 		pte[l2pte_index(va)] =
4528 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | f2s;
4529 		PTE_SYNC(&pte[l2pte_index(va)]);
4530 #endif
4531 		va += PAGE_SIZE;
4532 		pa += PAGE_SIZE;
4533 		resid -= PAGE_SIZE;
4534 	}
4535 #ifdef VERBOSE_INIT_ARM
4536 	printf("\n");
4537 #endif
4538 	return (size);
4539 }
4540 
4541 /********************** Static device map routines ***************************/
4542 
4543 static const struct pmap_devmap *pmap_devmap_table;
4544 
4545 /*
4546  * Register the devmap table.  This is provided in case early console
4547  * initialization needs to register mappings created by bootstrap code
4548  * before pmap_devmap_bootstrap() is called.
4549  */
4550 void
4551 pmap_devmap_register(const struct pmap_devmap *table)
4552 {
4553 
4554 	pmap_devmap_table = table;
4555 }
4556 
4557 /*
4558  * Map all of the static regions in the devmap table, and remember
4559  * the devmap table so other parts of the kernel can look up entries
4560  * later.
4561  */
4562 void
4563 pmap_devmap_bootstrap(vaddr_t l1pt, const struct pmap_devmap *table)
4564 {
4565 	int i;
4566 
4567 	pmap_devmap_table = table;
4568 
4569 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
4570 #ifdef VERBOSE_INIT_ARM
4571 		printf("devmap: %08lx -> %08lx @ %08lx\n",
4572 		    pmap_devmap_table[i].pd_pa,
4573 		    pmap_devmap_table[i].pd_pa +
4574 			pmap_devmap_table[i].pd_size - 1,
4575 		    pmap_devmap_table[i].pd_va);
4576 #endif
4577 		pmap_map_chunk(l1pt, pmap_devmap_table[i].pd_va,
4578 		    pmap_devmap_table[i].pd_pa,
4579 		    pmap_devmap_table[i].pd_size,
4580 		    pmap_devmap_table[i].pd_prot,
4581 		    pmap_devmap_table[i].pd_cache);
4582 	}
4583 }
4584 
4585 const struct pmap_devmap *
4586 pmap_devmap_find_pa(paddr_t pa, psize_t size)
4587 {
4588 	uint64_t endpa;
4589 	int i;
4590 
4591 	if (pmap_devmap_table == NULL)
4592 		return (NULL);
4593 
4594 	endpa = (uint64_t)pa + (uint64_t)(size - 1);
4595 
4596 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
4597 		if (pa >= pmap_devmap_table[i].pd_pa &&
4598 		    endpa <= (uint64_t)pmap_devmap_table[i].pd_pa +
4599 			     (uint64_t)(pmap_devmap_table[i].pd_size - 1))
4600 			return (&pmap_devmap_table[i]);
4601 	}
4602 
4603 	return (NULL);
4604 }
4605 
4606 const struct pmap_devmap *
4607 pmap_devmap_find_va(vaddr_t va, vsize_t size)
4608 {
4609 	int i;
4610 
4611 	if (pmap_devmap_table == NULL)
4612 		return (NULL);
4613 
4614 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
4615 		if (va >= pmap_devmap_table[i].pd_va &&
4616 		    va + size - 1 <= pmap_devmap_table[i].pd_va +
4617 				     pmap_devmap_table[i].pd_size - 1)
4618 			return (&pmap_devmap_table[i]);
4619 	}
4620 
4621 	return (NULL);
4622 }
4623 
4624 /********************** PTE initialization routines **************************/
4625 
4626 /*
4627  * These routines are called when the CPU type is identified to set up
4628  * the PTE prototypes, cache modes, etc.
4629  *
4630  * The variables are always here, just in case LKMs need to reference
4631  * them (though, they shouldn't).
4632  */
4633 
4634 pt_entry_t	pte_l1_s_cache_mode;
4635 pt_entry_t	pte_l1_s_cache_mode_pt;
4636 pt_entry_t	pte_l1_s_cache_mask;
4637 
4638 pt_entry_t	pte_l2_l_cache_mode;
4639 pt_entry_t	pte_l2_l_cache_mode_pt;
4640 pt_entry_t	pte_l2_l_cache_mask;
4641 
4642 pt_entry_t	pte_l2_s_cache_mode;
4643 pt_entry_t	pte_l2_s_cache_mode_pt;
4644 pt_entry_t	pte_l2_s_cache_mask;
4645 
4646 pt_entry_t	pte_l2_s_prot_u;
4647 pt_entry_t	pte_l2_s_prot_w;
4648 pt_entry_t	pte_l2_s_prot_mask;
4649 
4650 pt_entry_t	pte_l1_s_proto;
4651 pt_entry_t	pte_l1_c_proto;
4652 pt_entry_t	pte_l2_s_proto;
4653 
4654 void		(*pmap_copy_page_func)(paddr_t, paddr_t);
4655 void		(*pmap_zero_page_func)(paddr_t);
4656 
4657 #if (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0
4658 void
4659 pmap_pte_init_generic(void)
4660 {
4661 
4662 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
4663 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_generic;
4664 
4665 	pte_l2_l_cache_mode = L2_B|L2_C;
4666 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_generic;
4667 
4668 	pte_l2_s_cache_mode = L2_B|L2_C;
4669 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_generic;
4670 
4671 	/*
4672 	 * If we have a write-through cache, set B and C.  If
4673 	 * we have a write-back cache, then we assume setting
4674 	 * only C will make those pages write-through.
4675 	 */
4676 	if (cpufuncs.cf_dcache_wb_range == (void *) cpufunc_nullop) {
4677 		pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
4678 		pte_l2_l_cache_mode_pt = L2_B|L2_C;
4679 		pte_l2_s_cache_mode_pt = L2_B|L2_C;
4680 	} else {
4681 		pte_l1_s_cache_mode_pt = L1_S_C;
4682 		pte_l2_l_cache_mode_pt = L2_C;
4683 		pte_l2_s_cache_mode_pt = L2_C;
4684 	}
4685 
4686 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
4687 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
4688 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
4689 
4690 	pte_l1_s_proto = L1_S_PROTO_generic;
4691 	pte_l1_c_proto = L1_C_PROTO_generic;
4692 	pte_l2_s_proto = L2_S_PROTO_generic;
4693 
4694 	pmap_copy_page_func = pmap_copy_page_generic;
4695 	pmap_zero_page_func = pmap_zero_page_generic;
4696 }
4697 
4698 #if defined(CPU_ARM8)
4699 void
4700 pmap_pte_init_arm8(void)
4701 {
4702 
4703 	/*
4704 	 * ARM8 is compatible with generic, but we need to use
4705 	 * the page tables uncached.
4706 	 */
4707 	pmap_pte_init_generic();
4708 
4709 	pte_l1_s_cache_mode_pt = 0;
4710 	pte_l2_l_cache_mode_pt = 0;
4711 	pte_l2_s_cache_mode_pt = 0;
4712 }
4713 #endif /* CPU_ARM8 */
4714 
4715 #if defined(CPU_ARM9) && defined(ARM9_CACHE_WRITE_THROUGH)
4716 void
4717 pmap_pte_init_arm9(void)
4718 {
4719 
4720 	/*
4721 	 * ARM9 is compatible with generic, but we want to use
4722 	 * write-through caching for now.
4723 	 */
4724 	pmap_pte_init_generic();
4725 
4726 	pte_l1_s_cache_mode = L1_S_C;
4727 	pte_l2_l_cache_mode = L2_C;
4728 	pte_l2_s_cache_mode = L2_C;
4729 
4730 	pte_l1_s_cache_mode_pt = L1_S_C;
4731 	pte_l2_l_cache_mode_pt = L2_C;
4732 	pte_l2_s_cache_mode_pt = L2_C;
4733 }
4734 #endif /* CPU_ARM9 */
4735 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 */
4736 
4737 #if defined(CPU_ARM10)
4738 void
4739 pmap_pte_init_arm10(void)
4740 {
4741 
4742 	/*
4743 	 * ARM10 is compatible with generic, but we want to use
4744 	 * write-through caching for now.
4745 	 */
4746 	pmap_pte_init_generic();
4747 
4748 	pte_l1_s_cache_mode = L1_S_B | L1_S_C;
4749 	pte_l2_l_cache_mode = L2_B | L2_C;
4750 	pte_l2_s_cache_mode = L2_B | L2_C;
4751 
4752 	pte_l1_s_cache_mode_pt = L1_S_C;
4753 	pte_l2_l_cache_mode_pt = L2_C;
4754 	pte_l2_s_cache_mode_pt = L2_C;
4755 
4756 }
4757 #endif /* CPU_ARM10 */
4758 
4759 #if ARM_MMU_SA1 == 1
4760 void
4761 pmap_pte_init_sa1(void)
4762 {
4763 
4764 	/*
4765 	 * The StrongARM SA-1 cache does not have a write-through
4766 	 * mode.  So, do the generic initialization, then reset
4767 	 * the page table cache mode to B=1,C=1, and note that
4768 	 * the PTEs need to be sync'd.
4769 	 */
4770 	pmap_pte_init_generic();
4771 
4772 	pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
4773 	pte_l2_l_cache_mode_pt = L2_B|L2_C;
4774 	pte_l2_s_cache_mode_pt = L2_B|L2_C;
4775 
4776 	pmap_needs_pte_sync = 1;
4777 }
4778 #endif /* ARM_MMU_SA1 == 1*/
4779 
4780 #if ARM_MMU_XSCALE == 1
4781 #if (ARM_NMMUS > 1)
4782 static u_int xscale_use_minidata;
4783 #endif
4784 
4785 void
4786 pmap_pte_init_xscale(void)
4787 {
4788 	uint32_t auxctl;
4789 	int write_through = 0;
4790 
4791 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
4792 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_xscale;
4793 
4794 	pte_l2_l_cache_mode = L2_B|L2_C;
4795 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_xscale;
4796 
4797 	pte_l2_s_cache_mode = L2_B|L2_C;
4798 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_xscale;
4799 
4800 	pte_l1_s_cache_mode_pt = L1_S_C;
4801 	pte_l2_l_cache_mode_pt = L2_C;
4802 	pte_l2_s_cache_mode_pt = L2_C;
4803 
4804 #ifdef XSCALE_CACHE_READ_WRITE_ALLOCATE
4805 	/*
4806 	 * The XScale core has an enhanced mode where writes that
4807 	 * miss the cache cause a cache line to be allocated.  This
4808 	 * is significantly faster than the traditional, write-through
4809 	 * behavior of this case.
4810 	 */
4811 	pte_l1_s_cache_mode |= L1_S_XSCALE_TEX(TEX_XSCALE_X);
4812 	pte_l2_l_cache_mode |= L2_XSCALE_L_TEX(TEX_XSCALE_X);
4813 	pte_l2_s_cache_mode |= L2_XSCALE_T_TEX(TEX_XSCALE_X);
4814 #endif /* XSCALE_CACHE_READ_WRITE_ALLOCATE */
4815 
4816 #ifdef XSCALE_CACHE_WRITE_THROUGH
4817 	/*
4818 	 * Some versions of the XScale core have various bugs in
4819 	 * their cache units, the work-around for which is to run
4820 	 * the cache in write-through mode.  Unfortunately, this
4821 	 * has a major (negative) impact on performance.  So, we
4822 	 * go ahead and run fast-and-loose, in the hopes that we
4823 	 * don't line up the planets in a way that will trip the
4824 	 * bugs.
4825 	 *
4826 	 * However, we give you the option to be slow-but-correct.
4827 	 */
4828 	write_through = 1;
4829 #elif defined(XSCALE_CACHE_WRITE_BACK)
4830 	/* force write back cache mode */
4831 	write_through = 0;
4832 #elif defined(CPU_XSCALE_PXA250) || defined(CPU_XSCALE_PXA270)
4833 	/*
4834 	 * Intel PXA2[15]0 processors are known to have a bug in
4835 	 * write-back cache on revision 4 and earlier (stepping
4836 	 * A[01] and B[012]).  Fixed for C0 and later.
4837 	 */
4838 	{
4839 		uint32_t id, type;
4840 
4841 		id = cpufunc_id();
4842 		type = id & ~(CPU_ID_XSCALE_COREREV_MASK|CPU_ID_REVISION_MASK);
4843 
4844 		if (type == CPU_ID_PXA250 || type == CPU_ID_PXA210) {
4845 			if ((id & CPU_ID_REVISION_MASK) < 5) {
4846 				/* write through for stepping A0-1 and B0-2 */
4847 				write_through = 1;
4848 			}
4849 		}
4850 	}
4851 #endif /* XSCALE_CACHE_WRITE_THROUGH */
4852 
4853 	if (write_through) {
4854 		pte_l1_s_cache_mode = L1_S_C;
4855 		pte_l2_l_cache_mode = L2_C;
4856 		pte_l2_s_cache_mode = L2_C;
4857 	}
4858 
4859 #if (ARM_NMMUS > 1)
4860 	xscale_use_minidata = 1;
4861 #endif
4862 
4863 	pte_l2_s_prot_u = L2_S_PROT_U_xscale;
4864 	pte_l2_s_prot_w = L2_S_PROT_W_xscale;
4865 	pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale;
4866 
4867 	pte_l1_s_proto = L1_S_PROTO_xscale;
4868 	pte_l1_c_proto = L1_C_PROTO_xscale;
4869 	pte_l2_s_proto = L2_S_PROTO_xscale;
4870 
4871 	pmap_copy_page_func = pmap_copy_page_xscale;
4872 	pmap_zero_page_func = pmap_zero_page_xscale;
4873 
4874 	/*
4875 	 * Disable ECC protection of page table access, for now.
4876 	 */
4877 	__asm volatile("mrc p15, 0, %0, c1, c0, 1" : "=r" (auxctl));
4878 	auxctl &= ~XSCALE_AUXCTL_P;
4879 	__asm volatile("mcr p15, 0, %0, c1, c0, 1" : : "r" (auxctl));
4880 }
4881 
4882 /*
4883  * xscale_setup_minidata:
4884  *
4885  *	Set up the mini-data cache clean area.  We require the
4886  *	caller to allocate the right amount of physically and
4887  *	virtually contiguous space.
4888  */
4889 void
4890 xscale_setup_minidata(vaddr_t l1pt, vaddr_t va, paddr_t pa)
4891 {
4892 	extern vaddr_t xscale_minidata_clean_addr;
4893 	extern vsize_t xscale_minidata_clean_size; /* already initialized */
4894 	pd_entry_t *pde = (pd_entry_t *) l1pt;
4895 	pt_entry_t *pte;
4896 	vsize_t size;
4897 	uint32_t auxctl;
4898 
4899 	xscale_minidata_clean_addr = va;
4900 
4901 	/* Round it to page size. */
4902 	size = (xscale_minidata_clean_size + L2_S_OFFSET) & L2_S_FRAME;
4903 
4904 	for (; size != 0;
4905 	     va += L2_S_SIZE, pa += L2_S_SIZE, size -= L2_S_SIZE) {
4906 #ifndef ARM32_NEW_VM_LAYOUT
4907 		pte = (pt_entry_t *)
4908 		    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
4909 #else
4910 		pte = (pt_entry_t *) kernel_pt_lookup(
4911 		    pde[L1_IDX(va)] & L1_C_ADDR_MASK);
4912 #endif
4913 		if (pte == NULL)
4914 			panic("xscale_setup_minidata: can't find L2 table for "
4915 			    "VA 0x%08lx", va);
4916 #ifndef ARM32_NEW_VM_LAYOUT
4917 		pte[(va >> PGSHIFT) & 0x3ff] =
4918 #else
4919 		pte[l2pte_index(va)] =
4920 #endif
4921 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
4922 		    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);
4923 	}
4924 
4925 	/*
4926 	 * Configure the mini-data cache for write-back with
4927 	 * read/write-allocate.
4928 	 *
4929 	 * NOTE: In order to reconfigure the mini-data cache, we must
4930 	 * make sure it contains no valid data!  In order to do that,
4931 	 * we must issue a global data cache invalidate command!
4932 	 *
4933 	 * WE ASSUME WE ARE RUNNING UN-CACHED WHEN THIS ROUTINE IS CALLED!
4934 	 * THIS IS VERY IMPORTANT!
4935 	 */
4936 
4937 	/* Invalidate data and mini-data. */
4938 	__asm volatile("mcr p15, 0, %0, c7, c6, 0" : : "r" (0));
4939 	__asm volatile("mrc p15, 0, %0, c1, c0, 1" : "=r" (auxctl));
4940 	auxctl = (auxctl & ~XSCALE_AUXCTL_MD_MASK) | XSCALE_AUXCTL_MD_WB_RWA;
4941 	__asm volatile("mcr p15, 0, %0, c1, c0, 1" : : "r" (auxctl));
4942 }
4943 
4944 /*
4945  * Change the PTEs for the specified kernel mappings such that they
4946  * will use the mini data cache instead of the main data cache.
4947  */
4948 void
4949 pmap_uarea(vaddr_t va)
4950 {
4951 	struct l2_bucket *l2b;
4952 	pt_entry_t *ptep, *sptep, pte;
4953 	vaddr_t next_bucket, eva;
4954 
4955 #if (ARM_NMMUS > 1)
4956 	if (xscale_use_minidata == 0)
4957 		return;
4958 #endif
4959 
4960 	eva = va + USPACE;
4961 
4962 	while (va < eva) {
4963 		next_bucket = L2_NEXT_BUCKET(va);
4964 		if (next_bucket > eva)
4965 			next_bucket = eva;
4966 
4967 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
4968 		KDASSERT(l2b != NULL);
4969 
4970 		sptep = ptep = &l2b->l2b_kva[l2pte_index(va)];
4971 
4972 		while (va < next_bucket) {
4973 			pte = *ptep;
4974 			if (!l2pte_minidata(pte)) {
4975 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
4976 				cpu_tlb_flushD_SE(va);
4977 				*ptep = pte & ~L2_B;
4978 			}
4979 			ptep++;
4980 			va += PAGE_SIZE;
4981 		}
4982 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
4983 	}
4984 	cpu_cpwait();
4985 }
4986 #endif /* ARM_MMU_XSCALE == 1 */
4987 
4988 #if defined(DDB)
4989 /*
4990  * A couple of ddb-callable functions for dumping pmaps
4991  */
4992 void pmap_dump_all(void);
4993 void pmap_dump(pmap_t);
4994 
4995 void
4996 pmap_dump_all(void)
4997 {
4998 	pmap_t pm;
4999 
5000 	LIST_FOREACH(pm, &pmap_pmaps, pm_list) {
5001 		if (pm == pmap_kernel())
5002 			continue;
5003 		pmap_dump(pm);
5004 		printf("\n");
5005 	}
5006 }
5007 
5008 static pt_entry_t ncptes[64];
5009 static void pmap_dump_ncpg(pmap_t);
5010 
5011 void
5012 pmap_dump(pmap_t pm)
5013 {
5014 	struct l2_dtable *l2;
5015 	struct l2_bucket *l2b;
5016 	pt_entry_t *ptep, pte;
5017 	vaddr_t l2_va, l2b_va, va;
5018 	int i, j, k, occ, rows = 0;
5019 
5020 	if (pm == pmap_kernel())
5021 		printf("pmap_kernel (%p): ", pm);
5022 	else
5023 		printf("user pmap (%p): ", pm);
5024 
5025 	printf("domain %d, l1 at %p\n", pm->pm_domain, pm->pm_l1->l1_kva);
5026 
5027 	l2_va = 0;
5028 	for (i = 0; i < L2_SIZE; i++, l2_va += 0x01000000) {
5029 		l2 = pm->pm_l2[i];
5030 
5031 		if (l2 == NULL || l2->l2_occupancy == 0)
5032 			continue;
5033 
5034 		l2b_va = l2_va;
5035 		for (j = 0; j < L2_BUCKET_SIZE; j++, l2b_va += 0x00100000) {
5036 			l2b = &l2->l2_bucket[j];
5037 
5038 			if (l2b->l2b_occupancy == 0 || l2b->l2b_kva == NULL)
5039 				continue;
5040 
5041 			ptep = l2b->l2b_kva;
5042 
5043 			for (k = 0; k < 256 && ptep[k] == 0; k++)
5044 				;
5045 
5046 			k &= ~63;
5047 			occ = l2b->l2b_occupancy;
5048 			va = l2b_va + (k * 4096);
5049 			for (; k < 256; k++, va += 0x1000) {
5050 				char ch = ' ';
5051 				if ((k % 64) == 0) {
5052 					if ((rows % 8) == 0) {
5053 						printf(
5054 "          |0000   |8000   |10000  |18000  |20000  |28000  |30000  |38000\n");
5055 					}
5056 					printf("%08lx: ", va);
5057 				}
5058 
5059 				ncptes[k & 63] = 0;
5060 				pte = ptep[k];
5061 				if (pte == 0) {
5062 					ch = '.';
5063 				} else {
5064 					occ--;
5065 					switch (pte & 0x0c) {
5066 					case 0x00:
5067 						ch = 'D'; /* No cache No buff */
5068 						break;
5069 					case 0x04:
5070 						ch = 'B'; /* No cache buff */
5071 						break;
5072 					case 0x08:
5073 						if (pte & 0x40)
5074 							ch = 'm';
5075 						else
5076 						   ch = 'C'; /* Cache No buff */
5077 						break;
5078 					case 0x0c:
5079 						ch = 'F'; /* Cache Buff */
5080 						break;
5081 					}
5082 
5083 					if ((pte & L2_S_PROT_U) == L2_S_PROT_U)
5084 						ch += 0x20;
5085 
5086 					if ((pte & 0xc) == 0)
5087 						ncptes[k & 63] = pte;
5088 				}
5089 
5090 				if ((k % 64) == 63) {
5091 					rows++;
5092 					printf("%c\n", ch);
5093 					pmap_dump_ncpg(pm);
5094 					if (occ == 0)
5095 						break;
5096 				} else
5097 					printf("%c", ch);
5098 			}
5099 		}
5100 	}
5101 }
5102 
5103 static void
5104 pmap_dump_ncpg(pmap_t pm)
5105 {
5106 	struct vm_page *pg;
5107 	struct pv_entry *pv;
5108 	int i;
5109 
5110 	for (i = 0; i < 63; i++) {
5111 		if (ncptes[i] == 0)
5112 			continue;
5113 
5114 		pg = PHYS_TO_VM_PAGE(l2pte_pa(ncptes[i]));
5115 		if (pg == NULL)
5116 			continue;
5117 
5118 		printf(" pa 0x%08lx: krw %d kro %d urw %d uro %d\n",
5119 		    VM_PAGE_TO_PHYS(pg),
5120 		    pg->mdpage.krw_mappings, pg->mdpage.kro_mappings,
5121 		    pg->mdpage.urw_mappings, pg->mdpage.uro_mappings);
5122 
5123 		for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
5124 			printf("   %c va 0x%08lx, flags 0x%x\n",
5125 			    (pm == pv->pv_pmap) ? '*' : ' ',
5126 			    pv->pv_va, pv->pv_flags);
5127 		}
5128 	}
5129 }
5130 #endif
5131