xref: /netbsd-src/sys/arch/arm/arm32/pmap.c (revision d710132b4b8ce7f7cccaaf660cb16aa16b4077a0)
1 /*	$NetBSD: pmap.c,v 1.137 2003/06/23 11:01:07 martin 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 1.137 2003/06/23 11:01:07 martin Exp $");
216 
217 #ifdef PMAP_DEBUG
218 #define	PDEBUG(_lev_,_stat_) \
219 	if (pmap_debug_level >= (_lev_)) \
220         	((_stat_))
221 int pmap_debug_level = 0;
222 
223 /*
224  * for switching to potentially finer grained debugging
225  */
226 #define	PDB_FOLLOW	0x0001
227 #define	PDB_INIT	0x0002
228 #define	PDB_ENTER	0x0004
229 #define	PDB_REMOVE	0x0008
230 #define	PDB_CREATE	0x0010
231 #define	PDB_PTPAGE	0x0020
232 #define	PDB_GROWKERN	0x0040
233 #define	PDB_BITS	0x0080
234 #define	PDB_COLLECT	0x0100
235 #define	PDB_PROTECT	0x0200
236 #define	PDB_MAP_L1	0x0400
237 #define	PDB_BOOTSTRAP	0x1000
238 #define	PDB_PARANOIA	0x2000
239 #define	PDB_WIRING	0x4000
240 #define	PDB_PVDUMP	0x8000
241 #define	PDB_VAC		0x10000
242 #define	PDB_KENTER	0x20000
243 #define	PDB_KREMOVE	0x40000
244 
245 int debugmap = 1;
246 int pmapdebug = 0;
247 #define	NPDEBUG(_lev_,_stat_) \
248 	if (pmapdebug & (_lev_)) \
249         	((_stat_))
250 
251 #else	/* PMAP_DEBUG */
252 #define	PDEBUG(_lev_,_stat_) /* Nothing */
253 #define NPDEBUG(_lev_,_stat_) /* Nothing */
254 #endif	/* PMAP_DEBUG */
255 
256 /*
257  * pmap_kernel() points here
258  */
259 struct pmap     kernel_pmap_store;
260 
261 /*
262  * Which pmap is currently 'live' in the cache
263  *
264  * XXXSCW: Fix for SMP ...
265  */
266 union pmap_cache_state *pmap_cache_state;
267 
268 /*
269  * Pool and cache that pmap structures are allocated from.
270  * We use a cache to avoid clearing the pm_l2[] array (1KB)
271  * in pmap_create().
272  */
273 static struct pool pmap_pmap_pool;
274 static struct pool_cache pmap_pmap_cache;
275 static LIST_HEAD(, pmap) pmap_pmaps;
276 
277 /*
278  * Pool of PV structures
279  */
280 static struct pool pmap_pv_pool;
281 static void *pmap_bootstrap_pv_page_alloc(struct pool *, int);
282 static void pmap_bootstrap_pv_page_free(struct pool *, void *);
283 static struct pool_allocator pmap_bootstrap_pv_allocator = {
284 	pmap_bootstrap_pv_page_alloc, pmap_bootstrap_pv_page_free
285 };
286 
287 /*
288  * Pool and cache of l2_dtable structures.
289  * We use a cache to avoid clearing the structures when they're
290  * allocated. (196 bytes)
291  */
292 static struct pool pmap_l2dtable_pool;
293 static struct pool_cache pmap_l2dtable_cache;
294 static vaddr_t pmap_kernel_l2dtable_kva;
295 
296 /*
297  * Pool and cache of L2 page descriptors.
298  * We use a cache to avoid clearing the descriptor table
299  * when they're allocated. (1KB)
300  */
301 static struct pool pmap_l2ptp_pool;
302 static struct pool_cache pmap_l2ptp_cache;
303 static vaddr_t pmap_kernel_l2ptp_kva;
304 static paddr_t pmap_kernel_l2ptp_phys;
305 
306 /*
307  * pmap copy/zero page, and mem(5) hook point
308  */
309 static pt_entry_t *csrc_pte, *cdst_pte;
310 static vaddr_t csrcp, cdstp;
311 char *memhook;
312 extern caddr_t msgbufaddr;
313 
314 /*
315  * Flag to indicate if pmap_init() has done its thing
316  */
317 boolean_t pmap_initialized;
318 
319 /*
320  * Misc. locking data structures
321  */
322 
323 #if defined(MULTIPROCESSOR) || defined(LOCKDEBUG)
324 static struct lock pmap_main_lock;
325 
326 #define PMAP_MAP_TO_HEAD_LOCK() \
327      (void) spinlockmgr(&pmap_main_lock, LK_SHARED, NULL)
328 #define PMAP_MAP_TO_HEAD_UNLOCK() \
329      (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL)
330 #define PMAP_HEAD_TO_MAP_LOCK() \
331      (void) spinlockmgr(&pmap_main_lock, LK_EXCLUSIVE, NULL)
332 #define PMAP_HEAD_TO_MAP_UNLOCK() \
333      spinlockmgr(&pmap_main_lock, LK_RELEASE, (void *) 0)
334 #else
335 #define PMAP_MAP_TO_HEAD_LOCK()		/* null */
336 #define PMAP_MAP_TO_HEAD_UNLOCK()	/* null */
337 #define PMAP_HEAD_TO_MAP_LOCK()		/* null */
338 #define PMAP_HEAD_TO_MAP_UNLOCK()	/* null */
339 #endif
340 
341 #define	pmap_acquire_pmap_lock(pm)			\
342 	do {						\
343 		if ((pm) != pmap_kernel())		\
344 			simple_lock(&(pm)->pm_lock);	\
345 	} while (/*CONSTCOND*/0)
346 
347 #define	pmap_release_pmap_lock(pm)			\
348 	do {						\
349 		if ((pm) != pmap_kernel())		\
350 			simple_unlock(&(pm)->pm_lock);	\
351 	} while (/*CONSTCOND*/0)
352 
353 
354 /*
355  * Metadata for L1 translation tables.
356  */
357 struct l1_ttable {
358 	/* Entry on the L1 Table list */
359 	SLIST_ENTRY(l1_ttable) l1_link;
360 
361 	/* Entry on the L1 Least Recently Used list */
362 	TAILQ_ENTRY(l1_ttable) l1_lru;
363 
364 	/* Track how many domains are allocated from this L1 */
365 	volatile u_int l1_domain_use_count;
366 
367 	/*
368 	 * A free-list of domain numbers for this L1.
369 	 * We avoid using ffs() and a bitmap to track domains since ffs()
370 	 * is slow on ARM.
371 	 */
372 	u_int8_t l1_domain_first;
373 	u_int8_t l1_domain_free[PMAP_DOMAINS];
374 
375 	/* Physical address of this L1 page table */
376 	paddr_t l1_physaddr;
377 
378 	/* KVA of this L1 page table */
379 	pd_entry_t *l1_kva;
380 };
381 
382 /*
383  * Convert a virtual address into its L1 table index. That is, the
384  * index used to locate the L2 descriptor table pointer in an L1 table.
385  * This is basically used to index l1->l1_kva[].
386  *
387  * Each L2 descriptor table represents 1MB of VA space.
388  */
389 #define	L1_IDX(va)		(((vaddr_t)(va)) >> L1_S_SHIFT)
390 
391 /*
392  * L1 Page Tables are tracked using a Least Recently Used list.
393  *  - New L1s are allocated from the HEAD.
394  *  - Freed L1s are added to the TAIl.
395  *  - Recently accessed L1s (where an 'access' is some change to one of
396  *    the userland pmaps which owns this L1) are moved to the TAIL.
397  */
398 static TAILQ_HEAD(, l1_ttable) l1_lru_list;
399 static struct simplelock l1_lru_lock;
400 
401 /*
402  * A list of all L1 tables
403  */
404 static SLIST_HEAD(, l1_ttable) l1_list;
405 
406 /*
407  * The l2_dtable tracks L2_BUCKET_SIZE worth of L1 slots.
408  *
409  * This is normally 16MB worth L2 page descriptors for any given pmap.
410  * Reference counts are maintained for L2 descriptors so they can be
411  * freed when empty.
412  */
413 struct l2_dtable {
414 	/* The number of L2 page descriptors allocated to this l2_dtable */
415 	u_int l2_occupancy;
416 
417 	/* List of L2 page descriptors */
418 	struct l2_bucket {
419 		pt_entry_t *l2b_kva;	/* KVA of L2 Descriptor Table */
420 		paddr_t l2b_phys;	/* Physical address of same */
421 		u_short l2b_l1idx;	/* This L2 table's L1 index */
422 		u_short l2b_occupancy;	/* How many active descriptors */
423 	} l2_bucket[L2_BUCKET_SIZE];
424 };
425 
426 /*
427  * Given an L1 table index, calculate the corresponding l2_dtable index
428  * and bucket index within the l2_dtable.
429  */
430 #define	L2_IDX(l1idx)		(((l1idx) >> L2_BUCKET_LOG2) & \
431 				 (L2_SIZE - 1))
432 #define	L2_BUCKET(l1idx)	((l1idx) & (L2_BUCKET_SIZE - 1))
433 
434 /*
435  * Given a virtual address, this macro returns the
436  * virtual address required to drop into the next L2 bucket.
437  */
438 #define	L2_NEXT_BUCKET(va)	(((va) & L1_S_FRAME) + L1_S_SIZE)
439 
440 /*
441  * L2 allocation.
442  */
443 #define	pmap_alloc_l2_dtable()		\
444 	    pool_cache_get(&pmap_l2dtable_cache, PR_NOWAIT)
445 #define	pmap_free_l2_dtable(l2)		\
446 	    pool_cache_put(&pmap_l2dtable_cache, (l2))
447 #define pmap_alloc_l2_ptp(pap)		\
448 	    ((pt_entry_t *)pool_cache_get_paddr(&pmap_l2ptp_cache,\
449 	    PR_NOWAIT, (pap)))
450 
451 /*
452  * We try to map the page tables write-through, if possible.  However, not
453  * all CPUs have a write-through cache mode, so on those we have to sync
454  * the cache when we frob page tables.
455  *
456  * We try to evaluate this at compile time, if possible.  However, it's
457  * not always possible to do that, hence this run-time var.
458  */
459 int	pmap_needs_pte_sync;
460 
461 /*
462  * Real definition of pv_entry.
463  */
464 struct pv_entry {
465 	struct pv_entry *pv_next;       /* next pv_entry */
466 	pmap_t		pv_pmap;        /* pmap where mapping lies */
467 	vaddr_t		pv_va;          /* virtual address for mapping */
468 	u_int		pv_flags;       /* flags */
469 };
470 
471 /*
472  * Macro to determine if a mapping might be resident in the
473  * instruction cache and/or TLB
474  */
475 #define	PV_BEEN_EXECD(f)  (((f) & (PVF_REF | PVF_EXEC)) == (PVF_REF | PVF_EXEC))
476 
477 /*
478  * Macro to determine if a mapping might be resident in the
479  * data cache and/or TLB
480  */
481 #define	PV_BEEN_REFD(f)   (((f) & PVF_REF) != 0)
482 
483 /*
484  * Local prototypes
485  */
486 static int		pmap_set_pt_cache_mode(pd_entry_t *, vaddr_t);
487 static void		pmap_alloc_specials(vaddr_t *, int, vaddr_t *,
488 			    pt_entry_t **);
489 static boolean_t	pmap_is_current(pmap_t);
490 static boolean_t	pmap_is_cached(pmap_t);
491 static void		pmap_enter_pv(struct vm_page *, struct pv_entry *,
492 			    pmap_t, vaddr_t, u_int);
493 static struct pv_entry *pmap_find_pv(struct vm_page *, pmap_t, vaddr_t);
494 static struct pv_entry *pmap_remove_pv(struct vm_page *, pmap_t, vaddr_t);
495 static u_int		pmap_modify_pv(struct vm_page *, pmap_t, vaddr_t,
496 			    u_int, u_int);
497 
498 static void		pmap_pinit(pmap_t);
499 static int		pmap_pmap_ctor(void *, void *, int);
500 
501 static void		pmap_alloc_l1(pmap_t);
502 static void		pmap_free_l1(pmap_t);
503 static void		pmap_use_l1(pmap_t);
504 
505 static struct l2_bucket *pmap_get_l2_bucket(pmap_t, vaddr_t);
506 static struct l2_bucket *pmap_alloc_l2_bucket(pmap_t, vaddr_t);
507 static void		pmap_free_l2_bucket(pmap_t, struct l2_bucket *, u_int);
508 static int		pmap_l2ptp_ctor(void *, void *, int);
509 static int		pmap_l2dtable_ctor(void *, void *, int);
510 
511 static void		pmap_vac_me_harder(struct vm_page *, pmap_t, vaddr_t);
512 static void		pmap_vac_me_kpmap(struct vm_page *, pmap_t, vaddr_t);
513 static void		pmap_vac_me_user(struct vm_page *, pmap_t, vaddr_t);
514 
515 static void		pmap_clearbit(struct vm_page *, u_int);
516 static int		pmap_clean_page(struct pv_entry *, boolean_t);
517 static void		pmap_page_remove(struct vm_page *);
518 
519 static void		pmap_init_l1(struct l1_ttable *, pd_entry_t *);
520 static vaddr_t		kernel_pt_lookup(paddr_t);
521 
522 
523 /*
524  * External function prototypes
525  */
526 extern void bzero_page(vaddr_t);
527 extern void bcopy_page(vaddr_t, vaddr_t);
528 
529 /*
530  * Misc variables
531  */
532 vaddr_t virtual_avail;
533 vaddr_t virtual_end;
534 vaddr_t pmap_curmaxkvaddr;
535 
536 vaddr_t avail_start;
537 vaddr_t avail_end;
538 
539 extern pv_addr_t systempage;
540 
541 /* Function to set the debug level of the pmap code */
542 
543 #ifdef PMAP_DEBUG
544 void
545 pmap_debug(int level)
546 {
547 	pmap_debug_level = level;
548 	printf("pmap_debug: level=%d\n", pmap_debug_level);
549 }
550 #endif	/* PMAP_DEBUG */
551 
552 /*
553  * A bunch of routines to conditionally flush the caches/TLB depending
554  * on whether the specified pmap actually needs to be flushed at any
555  * given time.
556  */
557 static __inline void
558 pmap_tlb_flushID_SE(pmap_t pm, vaddr_t va)
559 {
560 
561 	if (pm->pm_cstate.cs_tlb_id)
562 		cpu_tlb_flushID_SE(va);
563 }
564 
565 static __inline void
566 pmap_tlb_flushD_SE(pmap_t pm, vaddr_t va)
567 {
568 
569 	if (pm->pm_cstate.cs_tlb_d)
570 		cpu_tlb_flushD_SE(va);
571 }
572 
573 static __inline void
574 pmap_tlb_flushID(pmap_t pm)
575 {
576 
577 	if (pm->pm_cstate.cs_tlb_id) {
578 		cpu_tlb_flushID();
579 		pm->pm_cstate.cs_tlb = 0;
580 	}
581 }
582 
583 static __inline void
584 pmap_tlb_flushD(pmap_t pm)
585 {
586 
587 	if (pm->pm_cstate.cs_tlb_d) {
588 		cpu_tlb_flushD();
589 		pm->pm_cstate.cs_tlb_d = 0;
590 	}
591 }
592 
593 static __inline void
594 pmap_idcache_wbinv_range(pmap_t pm, vaddr_t va, vsize_t len)
595 {
596 
597 	if (pm->pm_cstate.cs_cache_id)
598 		cpu_idcache_wbinv_range(va, len);
599 }
600 
601 static __inline void
602 pmap_dcache_wb_range(pmap_t pm, vaddr_t va, vsize_t len,
603     boolean_t do_inv, boolean_t rd_only)
604 {
605 
606 	if (pm->pm_cstate.cs_cache_d) {
607 		if (do_inv) {
608 			if (rd_only)
609 				cpu_dcache_inv_range(va, len);
610 			else
611 				cpu_dcache_wbinv_range(va, len);
612 		} else
613 		if (!rd_only)
614 			cpu_dcache_wb_range(va, len);
615 	}
616 }
617 
618 static __inline void
619 pmap_idcache_wbinv_all(pmap_t pm)
620 {
621 
622 	if (pm->pm_cstate.cs_cache_id) {
623 		cpu_idcache_wbinv_all();
624 		pm->pm_cstate.cs_cache = 0;
625 	}
626 }
627 
628 static __inline void
629 pmap_dcache_wbinv_all(pmap_t pm)
630 {
631 
632 	if (pm->pm_cstate.cs_cache_d) {
633 		cpu_dcache_wbinv_all();
634 		pm->pm_cstate.cs_cache_d = 0;
635 	}
636 }
637 
638 static __inline boolean_t
639 pmap_is_current(pmap_t pm)
640 {
641 
642 	if (pm == pmap_kernel() ||
643 	    (curproc && curproc->p_vmspace->vm_map.pmap == pm))
644 		return (TRUE);
645 
646 	return (FALSE);
647 }
648 
649 static __inline boolean_t
650 pmap_is_cached(pmap_t pm)
651 {
652 
653 	if (pm == pmap_kernel() || pmap_cache_state == NULL ||
654 	   pmap_cache_state == &pm->pm_cstate)
655 		return (TRUE);
656 
657 	return (FALSE);
658 }
659 
660 /*
661  * PTE_SYNC_CURRENT:
662  *
663  *     Make sure the pte is written out to RAM.
664  *     We need to do this for one of two cases:
665  *       - We're dealing with the kernel pmap
666  *       - There is no pmap active in the cache/tlb.
667  *       - The specified pmap is 'active' in the cache/tlb.
668  */
669 #ifdef PMAP_INCLUDE_PTE_SYNC
670 #define	PTE_SYNC_CURRENT(pm, ptep)	\
671 do {					\
672 	if (PMAP_NEEDS_PTE_SYNC && 	\
673 	    pmap_is_cached(pm))		\
674 		PTE_SYNC(ptep);		\
675 } while (/*CONSTCOND*/0)
676 #else
677 #define	PTE_SYNC_CURRENT(pm, ptep)	/* nothing */
678 #endif
679 
680 /*
681  * main pv_entry manipulation functions:
682  *   pmap_enter_pv: enter a mapping onto a vm_page list
683  *   pmap_remove_pv: remove a mappiing from a vm_page list
684  *
685  * NOTE: pmap_enter_pv expects to lock the pvh itself
686  *       pmap_remove_pv expects te caller to lock the pvh before calling
687  */
688 
689 /*
690  * pmap_enter_pv: enter a mapping onto a vm_page lst
691  *
692  * => caller should hold the proper lock on pmap_main_lock
693  * => caller should have pmap locked
694  * => we will gain the lock on the vm_page and allocate the new pv_entry
695  * => caller should adjust ptp's wire_count before calling
696  * => caller should not adjust pmap's wire_count
697  */
698 static void
699 pmap_enter_pv(struct vm_page *pg, struct pv_entry *pve, pmap_t pm,
700     vaddr_t va, u_int flags)
701 {
702 
703 	NPDEBUG(PDB_PVDUMP,
704 	    printf("pmap_enter_pv: pm %p, pg %p, flags 0x%x\n", pm, pg, flags));
705 
706 	pve->pv_pmap = pm;
707 	pve->pv_va = va;
708 	pve->pv_flags = flags;
709 
710 	simple_lock(&pg->mdpage.pvh_slock);	/* lock vm_page */
711 	pve->pv_next = pg->mdpage.pvh_list;	/* add to ... */
712 	pg->mdpage.pvh_list = pve;		/* ... locked list */
713 	pg->mdpage.pvh_attrs |= flags & (PVF_REF | PVF_MOD);
714 	if (pm == pmap_kernel()) {
715 		if (flags & PVF_WRITE)
716 			pg->mdpage.krw_mappings++;
717 		else
718 			pg->mdpage.kro_mappings++;
719 	} else
720 	if (flags & PVF_WRITE)
721 		pg->mdpage.urw_mappings++;
722 	else
723 		pg->mdpage.uro_mappings++;
724 	simple_unlock(&pg->mdpage.pvh_slock);	/* unlock, done! */
725 
726 	if (pve->pv_flags & PVF_WIRED)
727 		++pm->pm_stats.wired_count;
728 }
729 
730 /*
731  *
732  * pmap_find_pv: Find a pv entry
733  *
734  * => caller should hold lock on vm_page
735  */
736 static __inline struct pv_entry *
737 pmap_find_pv(struct vm_page *pg, pmap_t pm, vaddr_t va)
738 {
739 	struct pv_entry *pv;
740 
741 	for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
742 		if (pm == pv->pv_pmap && va == pv->pv_va)
743 			break;
744 	}
745 
746 	return (pv);
747 }
748 
749 /*
750  * pmap_remove_pv: try to remove a mapping from a pv_list
751  *
752  * => caller should hold proper lock on pmap_main_lock
753  * => pmap should be locked
754  * => caller should hold lock on vm_page [so that attrs can be adjusted]
755  * => caller should adjust ptp's wire_count and free PTP if needed
756  * => caller should NOT adjust pmap's wire_count
757  * => we return the removed pve
758  */
759 static struct pv_entry *
760 pmap_remove_pv(struct vm_page *pg, pmap_t pm, vaddr_t va)
761 {
762 	struct pv_entry *pve, **prevptr;
763 
764 	NPDEBUG(PDB_PVDUMP,
765 	    printf("pmap_remove_pv: pm %p, pg %p, va 0x%08lx\n", pm, pg, va));
766 
767 	prevptr = &pg->mdpage.pvh_list;		/* previous pv_entry pointer */
768 	pve = *prevptr;
769 
770 	while (pve) {
771 		if (pve->pv_pmap == pm && pve->pv_va == va) {	/* match? */
772 			NPDEBUG(PDB_PVDUMP,
773 			    printf("pmap_remove_pv: pm %p, pg %p, flags 0x%x\n", pm, pg, pve->pv_flags));
774 			*prevptr = pve->pv_next;		/* remove it! */
775 			if (pve->pv_flags & PVF_WIRED)
776 			    --pm->pm_stats.wired_count;
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 static void
861 pmap_pinit(pmap_t pm)
862 {
863 
864 	if (vector_page < KERNEL_BASE) {
865 		/*
866 		 * Map the vector page.
867 		 */
868 		pmap_enter(pm, vector_page, systempage.pv_pa,
869 		    VM_PROT_READ, VM_PROT_READ | PMAP_WIRED);
870 		pmap_update(pm);
871 	}
872 }
873 
874 /*
875  * Allocate an L1 translation table for the specified pmap.
876  * This is called at pmap creation time.
877  */
878 static void
879 pmap_alloc_l1(pmap_t pm)
880 {
881 	struct l1_ttable *l1;
882 	u_int8_t domain;
883 
884 	/*
885 	 * Remove the L1 at the head of the LRU list
886 	 */
887 	simple_lock(&l1_lru_lock);
888 	l1 = TAILQ_FIRST(&l1_lru_list);
889 	KDASSERT(l1 != NULL);
890 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
891 
892 	/*
893 	 * Pick the first available domain number, and update
894 	 * the link to the next number.
895 	 */
896 	domain = l1->l1_domain_first;
897 	l1->l1_domain_first = l1->l1_domain_free[domain];
898 
899 	/*
900 	 * If there are still free domain numbers in this L1,
901 	 * put it back on the TAIL of the LRU list.
902 	 */
903 	if (++l1->l1_domain_use_count < PMAP_DOMAINS)
904 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
905 
906 	simple_unlock(&l1_lru_lock);
907 
908 	/*
909 	 * Fix up the relevant bits in the pmap structure
910 	 */
911 	pm->pm_l1 = l1;
912 	pm->pm_domain = domain;
913 }
914 
915 /*
916  * Free an L1 translation table.
917  * This is called at pmap destruction time.
918  */
919 static void
920 pmap_free_l1(pmap_t pm)
921 {
922 	struct l1_ttable *l1 = pm->pm_l1;
923 
924 	simple_lock(&l1_lru_lock);
925 
926 	/*
927 	 * If this L1 is currently on the LRU list, remove it.
928 	 */
929 	if (l1->l1_domain_use_count < PMAP_DOMAINS)
930 		TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
931 
932 	/*
933 	 * Free up the domain number which was allocated to the pmap
934 	 */
935 	l1->l1_domain_free[pm->pm_domain] = l1->l1_domain_first;
936 	l1->l1_domain_first = pm->pm_domain;
937 	l1->l1_domain_use_count--;
938 
939 	/*
940 	 * The L1 now must have at least 1 free domain, so add
941 	 * it back to the LRU list. If the use count is zero,
942 	 * put it at the head of the list, otherwise it goes
943 	 * to the tail.
944 	 */
945 	if (l1->l1_domain_use_count == 0)
946 		TAILQ_INSERT_HEAD(&l1_lru_list, l1, l1_lru);
947 	else
948 		TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
949 
950 	simple_unlock(&l1_lru_lock);
951 }
952 
953 static __inline void
954 pmap_use_l1(pmap_t pm)
955 {
956 	struct l1_ttable *l1;
957 
958 	/*
959 	 * Do nothing if we're in interrupt context.
960 	 * Access to an L1 by the kernel pmap must not affect
961 	 * the LRU list.
962 	 */
963 	if (current_intr_depth || pm == pmap_kernel())
964 		return;
965 
966 	l1 = pm->pm_l1;
967 
968 	/*
969 	 * If the L1 is not currently on the LRU list, just return
970 	 */
971 	if (l1->l1_domain_use_count == PMAP_DOMAINS)
972 		return;
973 
974 	simple_lock(&l1_lru_lock);
975 
976 	/*
977 	 * Check the use count again, now that we've acquired the lock
978 	 */
979 	if (l1->l1_domain_use_count == PMAP_DOMAINS) {
980 		simple_unlock(&l1_lru_lock);
981 		return;
982 	}
983 
984 	/*
985 	 * Move the L1 to the back of the LRU list
986 	 */
987 	TAILQ_REMOVE(&l1_lru_list, l1, l1_lru);
988 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
989 
990 	simple_unlock(&l1_lru_lock);
991 }
992 
993 /*
994  * void pmap_free_l2_ptp(pt_entry_t *, paddr_t *)
995  *
996  * Free an L2 descriptor table.
997  */
998 static __inline void
999 #ifndef PMAP_INCLUDE_PTE_SYNC
1000 pmap_free_l2_ptp(pt_entry_t *l2, paddr_t pa)
1001 #else
1002 pmap_free_l2_ptp(boolean_t need_sync, pt_entry_t *l2, paddr_t pa)
1003 #endif
1004 {
1005 #ifdef PMAP_INCLUDE_PTE_SYNC
1006 	/*
1007 	 * Note: With a write-back cache, we may need to sync this
1008 	 * L2 table before re-using it.
1009 	 * This is because it may have belonged to a non-current
1010 	 * pmap, in which case the cache syncs would have been
1011 	 * skipped when the pages were being unmapped. If the
1012 	 * L2 table were then to be immediately re-allocated to
1013 	 * the *current* pmap, it may well contain stale mappings
1014 	 * which have not yet been cleared by a cache write-back
1015 	 * and so would still be visible to the mmu.
1016 	 */
1017 	if (need_sync)
1018 		PTE_SYNC_RANGE(l2, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1019 #endif
1020 	pool_cache_put_paddr(&pmap_l2ptp_cache, (void *)l2, pa);
1021 }
1022 
1023 /*
1024  * Returns a pointer to the L2 bucket associated with the specified pmap
1025  * and VA, or NULL if no L2 bucket exists for the address.
1026  */
1027 static __inline struct l2_bucket *
1028 pmap_get_l2_bucket(pmap_t pm, vaddr_t va)
1029 {
1030 	struct l2_dtable *l2;
1031 	struct l2_bucket *l2b;
1032 	u_short l1idx;
1033 
1034 	l1idx = L1_IDX(va);
1035 
1036 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL ||
1037 	    (l2b = &l2->l2_bucket[L2_BUCKET(l1idx)])->l2b_kva == NULL)
1038 		return (NULL);
1039 
1040 	return (l2b);
1041 }
1042 
1043 /*
1044  * Returns a pointer to the L2 bucket associated with the specified pmap
1045  * and VA.
1046  *
1047  * If no L2 bucket exists, perform the necessary allocations to put an L2
1048  * bucket/page table in place.
1049  *
1050  * Note that if a new L2 bucket/page was allocated, the caller *must*
1051  * increment the bucket occupancy counter appropriately *before*
1052  * releasing the pmap's lock to ensure no other thread or cpu deallocates
1053  * the bucket/page in the meantime.
1054  */
1055 static struct l2_bucket *
1056 pmap_alloc_l2_bucket(pmap_t pm, vaddr_t va)
1057 {
1058 	struct l2_dtable *l2;
1059 	struct l2_bucket *l2b;
1060 	u_short l1idx;
1061 
1062 	l1idx = L1_IDX(va);
1063 
1064 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
1065 		/*
1066 		 * No mapping at this address, as there is
1067 		 * no entry in the L1 table.
1068 		 * Need to allocate a new l2_dtable.
1069 		 */
1070 		if ((l2 = pmap_alloc_l2_dtable()) == NULL)
1071 			return (NULL);
1072 
1073 		/*
1074 		 * Link it into the parent pmap
1075 		 */
1076 		pm->pm_l2[L2_IDX(l1idx)] = l2;
1077 	}
1078 
1079 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
1080 
1081 	/*
1082 	 * Fetch pointer to the L2 page table associated with the address.
1083 	 */
1084 	if (l2b->l2b_kva == NULL) {
1085 		pt_entry_t *ptep;
1086 
1087 		/*
1088 		 * No L2 page table has been allocated. Chances are, this
1089 		 * is because we just allocated the l2_dtable, above.
1090 		 */
1091 		if ((ptep = pmap_alloc_l2_ptp(&l2b->l2b_phys)) == NULL) {
1092 			/*
1093 			 * Oops, no more L2 page tables available at this
1094 			 * time. We may need to deallocate the l2_dtable
1095 			 * if we allocated a new one above.
1096 			 */
1097 			if (l2->l2_occupancy == 0) {
1098 				pm->pm_l2[L2_IDX(l1idx)] = NULL;
1099 				pmap_free_l2_dtable(l2);
1100 			}
1101 			return (NULL);
1102 		}
1103 
1104 		l2->l2_occupancy++;
1105 		l2b->l2b_kva = ptep;
1106 		l2b->l2b_l1idx = l1idx;
1107 	}
1108 
1109 	return (l2b);
1110 }
1111 
1112 /*
1113  * One or more mappings in the specified L2 descriptor table have just been
1114  * invalidated.
1115  *
1116  * Garbage collect the metadata and descriptor table itself if necessary.
1117  *
1118  * The pmap lock must be acquired when this is called (not necessary
1119  * for the kernel pmap).
1120  */
1121 static void
1122 pmap_free_l2_bucket(pmap_t pm, struct l2_bucket *l2b, u_int count)
1123 {
1124 	struct l2_dtable *l2;
1125 	pd_entry_t *pl1pd, l1pd;
1126 	pt_entry_t *ptep;
1127 	u_short l1idx;
1128 
1129 	KDASSERT(count <= l2b->l2b_occupancy);
1130 
1131 	/*
1132 	 * Update the bucket's reference count according to how many
1133 	 * PTEs the caller has just invalidated.
1134 	 */
1135 	l2b->l2b_occupancy -= count;
1136 
1137 	/*
1138 	 * Note:
1139 	 *
1140 	 * Level 2 page tables allocated to the kernel pmap are never freed
1141 	 * as that would require checking all Level 1 page tables and
1142 	 * removing any references to the Level 2 page table. See also the
1143 	 * comment elsewhere about never freeing bootstrap L2 descriptors.
1144 	 *
1145 	 * We make do with just invalidating the mapping in the L2 table.
1146 	 *
1147 	 * This isn't really a big deal in practice and, in fact, leads
1148 	 * to a performance win over time as we don't need to continually
1149 	 * alloc/free.
1150 	 */
1151 	if (l2b->l2b_occupancy > 0 || pm == pmap_kernel())
1152 		return;
1153 
1154 	/*
1155 	 * There are no more valid mappings in this level 2 page table.
1156 	 * Go ahead and NULL-out the pointer in the bucket, then
1157 	 * free the page table.
1158 	 */
1159 	l1idx = l2b->l2b_l1idx;
1160 	ptep = l2b->l2b_kva;
1161 	l2b->l2b_kva = NULL;
1162 
1163 	pl1pd = &pm->pm_l1->l1_kva[l1idx];
1164 
1165 	/*
1166 	 * If the L1 slot matches the pmap's domain
1167 	 * number, then invalidate it.
1168 	 */
1169 	l1pd = *pl1pd & (L1_TYPE_MASK | L1_C_DOM_MASK);
1170 	if (l1pd == (L1_C_DOM(pm->pm_domain) | L1_TYPE_C)) {
1171 		*pl1pd = 0;
1172 		PTE_SYNC(pl1pd);
1173 	}
1174 
1175 	/*
1176 	 * Release the L2 descriptor table back to the pool cache.
1177 	 */
1178 #ifndef PMAP_INCLUDE_PTE_SYNC
1179 	pmap_free_l2_ptp(ptep, l2b->l2b_phys);
1180 #else
1181 	pmap_free_l2_ptp(!pmap_is_cached(pm), ptep, l2b->l2b_phys);
1182 #endif
1183 
1184 	/*
1185 	 * Update the reference count in the associated l2_dtable
1186 	 */
1187 	l2 = pm->pm_l2[L2_IDX(l1idx)];
1188 	if (--l2->l2_occupancy > 0)
1189 		return;
1190 
1191 	/*
1192 	 * There are no more valid mappings in any of the Level 1
1193 	 * slots managed by this l2_dtable. Go ahead and NULL-out
1194 	 * the pointer in the parent pmap and free the l2_dtable.
1195 	 */
1196 	pm->pm_l2[L2_IDX(l1idx)] = NULL;
1197 	pmap_free_l2_dtable(l2);
1198 }
1199 
1200 /*
1201  * Pool cache constructors for L2 descriptor tables, metadata and pmap
1202  * structures.
1203  */
1204 static int
1205 pmap_l2ptp_ctor(void *arg, void *v, int flags)
1206 {
1207 #ifndef PMAP_INCLUDE_PTE_SYNC
1208 	struct l2_bucket *l2b;
1209 	pt_entry_t *ptep, pte;
1210 	vaddr_t va = (vaddr_t)v & ~PGOFSET;
1211 
1212 	/*
1213 	 * The mappings for these page tables were initially made using
1214 	 * pmap_kenter_pa() by the pool subsystem. Therefore, the cache-
1215 	 * mode will not be right for page table mappings. To avoid
1216 	 * polluting the pmap_kenter_pa() code with a special case for
1217 	 * page tables, we simply fix up the cache-mode here if it's not
1218 	 * correct.
1219 	 */
1220 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
1221 	KDASSERT(l2b != NULL);
1222 	ptep = &l2b->l2b_kva[l2pte_index(va)];
1223 	pte = *ptep;
1224 
1225 	if ((pte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
1226 		/*
1227 		 * Page tables must have the cache-mode set to Write-Thru.
1228 		 */
1229 		*ptep = (pte & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode_pt;
1230 		PTE_SYNC(ptep);
1231 		cpu_tlb_flushD_SE(va);
1232 		cpu_cpwait();
1233 	}
1234 #endif
1235 
1236 	memset(v, 0, L2_TABLE_SIZE_REAL);
1237 	PTE_SYNC_RANGE(v, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
1238 	return (0);
1239 }
1240 
1241 static int
1242 pmap_l2dtable_ctor(void *arg, void *v, int flags)
1243 {
1244 
1245 	memset(v, 0, sizeof(struct l2_dtable));
1246 	return (0);
1247 }
1248 
1249 static int
1250 pmap_pmap_ctor(void *arg, void *v, int flags)
1251 {
1252 
1253 	memset(v, 0, sizeof(struct pmap));
1254 	return (0);
1255 }
1256 
1257 /*
1258  * Since we have a virtually indexed cache, we may need to inhibit caching if
1259  * there is more than one mapping and at least one of them is writable.
1260  * Since we purge the cache on every context switch, we only need to check for
1261  * other mappings within the same pmap, or kernel_pmap.
1262  * This function is also called when a page is unmapped, to possibly reenable
1263  * caching on any remaining mappings.
1264  *
1265  * The code implements the following logic, where:
1266  *
1267  * KW = # of kernel read/write pages
1268  * KR = # of kernel read only pages
1269  * UW = # of user read/write pages
1270  * UR = # of user read only pages
1271  *
1272  * KC = kernel mapping is cacheable
1273  * UC = user mapping is cacheable
1274  *
1275  *               KW=0,KR=0  KW=0,KR>0  KW=1,KR=0  KW>1,KR>=0
1276  *             +---------------------------------------------
1277  * UW=0,UR=0   | ---        KC=1       KC=1       KC=0
1278  * UW=0,UR>0   | UC=1       KC=1,UC=1  KC=0,UC=0  KC=0,UC=0
1279  * UW=1,UR=0   | UC=1       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1280  * UW>1,UR>=0  | UC=0       KC=0,UC=0  KC=0,UC=0  KC=0,UC=0
1281  */
1282 
1283 static const int pmap_vac_flags[4][4] = {
1284 	{-1,		0,		0,		PVF_KNC},
1285 	{0,		0,		PVF_NC,		PVF_NC},
1286 	{0,		PVF_NC,		PVF_NC,		PVF_NC},
1287 	{PVF_UNC,	PVF_NC,		PVF_NC,		PVF_NC}
1288 };
1289 
1290 static __inline int
1291 pmap_get_vac_flags(const struct vm_page *pg)
1292 {
1293 	int kidx, uidx;
1294 
1295 	kidx = 0;
1296 	if (pg->mdpage.kro_mappings || pg->mdpage.krw_mappings > 1)
1297 		kidx |= 1;
1298 	if (pg->mdpage.krw_mappings)
1299 		kidx |= 2;
1300 
1301 	uidx = 0;
1302 	if (pg->mdpage.uro_mappings || pg->mdpage.urw_mappings > 1)
1303 		uidx |= 1;
1304 	if (pg->mdpage.urw_mappings)
1305 		uidx |= 2;
1306 
1307 	return (pmap_vac_flags[uidx][kidx]);
1308 }
1309 
1310 static __inline void
1311 pmap_vac_me_harder(struct vm_page *pg, pmap_t pm, vaddr_t va)
1312 {
1313 	int nattr;
1314 
1315 	nattr = pmap_get_vac_flags(pg);
1316 
1317 	if (nattr < 0) {
1318 		pg->mdpage.pvh_attrs &= ~PVF_NC;
1319 		return;
1320 	}
1321 
1322 	if (nattr == 0 && (pg->mdpage.pvh_attrs & PVF_NC) == 0)
1323 		return;
1324 
1325 	if (pm == pmap_kernel())
1326 		pmap_vac_me_kpmap(pg, pm, va);
1327 	else
1328 		pmap_vac_me_user(pg, pm, va);
1329 
1330 	pg->mdpage.pvh_attrs = (pg->mdpage.pvh_attrs & ~PVF_NC) | nattr;
1331 }
1332 
1333 static void
1334 pmap_vac_me_kpmap(struct vm_page *pg, pmap_t pm, vaddr_t va)
1335 {
1336 	u_int u_cacheable, u_entries;
1337 	struct pv_entry *pv;
1338 	pmap_t last_pmap = pm;
1339 
1340 	/*
1341 	 * Pass one, see if there are both kernel and user pmaps for
1342 	 * this page.  Calculate whether there are user-writable or
1343 	 * kernel-writable pages.
1344 	 */
1345 	u_cacheable = 0;
1346 	for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
1347 		if (pv->pv_pmap != pm && (pv->pv_flags & PVF_NC) == 0)
1348 			u_cacheable++;
1349 	}
1350 
1351 	u_entries = pg->mdpage.urw_mappings + pg->mdpage.uro_mappings;
1352 
1353 	/*
1354 	 * We know we have just been updating a kernel entry, so if
1355 	 * all user pages are already cacheable, then there is nothing
1356 	 * further to do.
1357 	 */
1358 	if (pg->mdpage.k_mappings == 0 && u_cacheable == u_entries)
1359 		return;
1360 
1361 	if (u_entries) {
1362 		/*
1363 		 * Scan over the list again, for each entry, if it
1364 		 * might not be set correctly, call pmap_vac_me_user
1365 		 * to recalculate the settings.
1366 		 */
1367 		for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
1368 			/*
1369 			 * We know kernel mappings will get set
1370 			 * correctly in other calls.  We also know
1371 			 * that if the pmap is the same as last_pmap
1372 			 * then we've just handled this entry.
1373 			 */
1374 			if (pv->pv_pmap == pm || pv->pv_pmap == last_pmap)
1375 				continue;
1376 
1377 			/*
1378 			 * If there are kernel entries and this page
1379 			 * is writable but non-cacheable, then we can
1380 			 * skip this entry also.
1381 			 */
1382 			if (pg->mdpage.k_mappings &&
1383 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) ==
1384 			    (PVF_NC | PVF_WRITE))
1385 				continue;
1386 
1387 			/*
1388 			 * Similarly if there are no kernel-writable
1389 			 * entries and the page is already
1390 			 * read-only/cacheable.
1391 			 */
1392 			if (pg->mdpage.krw_mappings == 0 &&
1393 			    (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 0)
1394 				continue;
1395 
1396 			/*
1397 			 * For some of the remaining cases, we know
1398 			 * that we must recalculate, but for others we
1399 			 * can't tell if they are correct or not, so
1400 			 * we recalculate anyway.
1401 			 */
1402 			pmap_vac_me_user(pg, (last_pmap = pv->pv_pmap), 0);
1403 		}
1404 
1405 		if (pg->mdpage.k_mappings == 0)
1406 			return;
1407 	}
1408 
1409 	pmap_vac_me_user(pg, pm, va);
1410 }
1411 
1412 static void
1413 pmap_vac_me_user(struct vm_page *pg, pmap_t pm, vaddr_t va)
1414 {
1415 	pmap_t kpmap = pmap_kernel();
1416 	struct pv_entry *pv, *npv;
1417 	struct l2_bucket *l2b;
1418 	pt_entry_t *ptep, pte;
1419 	u_int entries = 0;
1420 	u_int writable = 0;
1421 	u_int cacheable_entries = 0;
1422 	u_int kern_cacheable = 0;
1423 	u_int other_writable = 0;
1424 
1425 	/*
1426 	 * Count mappings and writable mappings in this pmap.
1427 	 * Include kernel mappings as part of our own.
1428 	 * Keep a pointer to the first one.
1429 	 */
1430 	for (pv = npv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
1431 		/* Count mappings in the same pmap */
1432 		if (pm == pv->pv_pmap || kpmap == pv->pv_pmap) {
1433 			if (entries++ == 0)
1434 				npv = pv;
1435 
1436 			/* Cacheable mappings */
1437 			if ((pv->pv_flags & PVF_NC) == 0) {
1438 				cacheable_entries++;
1439 				if (kpmap == pv->pv_pmap)
1440 					kern_cacheable++;
1441 			}
1442 
1443 			/* Writable mappings */
1444 			if (pv->pv_flags & PVF_WRITE)
1445 				++writable;
1446 		} else
1447 		if (pv->pv_flags & PVF_WRITE)
1448 			other_writable = 1;
1449 	}
1450 
1451 	/*
1452 	 * Enable or disable caching as necessary.
1453 	 * Note: the first entry might be part of the kernel pmap,
1454 	 * so we can't assume this is indicative of the state of the
1455 	 * other (maybe non-kpmap) entries.
1456 	 */
1457 	if ((entries > 1 && writable) ||
1458 	    (entries > 0 && pm == kpmap && other_writable)) {
1459 		if (cacheable_entries == 0)
1460 			return;
1461 
1462 		for (pv = npv; pv; pv = pv->pv_next) {
1463 			if ((pm != pv->pv_pmap && kpmap != pv->pv_pmap) ||
1464 			    (pv->pv_flags & PVF_NC))
1465 				continue;
1466 
1467 			pv->pv_flags |= PVF_NC;
1468 
1469 			l2b = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
1470 			ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1471 			pte = *ptep & ~L2_S_CACHE_MASK;
1472 
1473 			if ((va != pv->pv_va || pm != pv->pv_pmap) &&
1474 			    l2pte_valid(pte)) {
1475 				if (PV_BEEN_EXECD(pv->pv_flags)) {
1476 					pmap_idcache_wbinv_range(pv->pv_pmap,
1477 					    pv->pv_va, PAGE_SIZE);
1478 					pmap_tlb_flushID_SE(pv->pv_pmap,
1479 					    pv->pv_va);
1480 				} else
1481 				if (PV_BEEN_REFD(pv->pv_flags)) {
1482 					pmap_dcache_wb_range(pv->pv_pmap,
1483 					    pv->pv_va, PAGE_SIZE, TRUE,
1484 					    (pv->pv_flags & PVF_WRITE) == 0);
1485 					pmap_tlb_flushD_SE(pv->pv_pmap,
1486 					    pv->pv_va);
1487 				}
1488 			}
1489 
1490 			*ptep = pte;
1491 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
1492 		}
1493 		cpu_cpwait();
1494 	} else
1495 	if (entries > cacheable_entries) {
1496 		/*
1497 		 * Turn cacheing back on for some pages.  If it is a kernel
1498 		 * page, only do so if there are no other writable pages.
1499 		 */
1500 		for (pv = npv; pv; pv = pv->pv_next) {
1501 			if (!(pv->pv_flags & PVF_NC) || (pm != pv->pv_pmap &&
1502 			    (kpmap != pv->pv_pmap || other_writable)))
1503 				continue;
1504 
1505 			pv->pv_flags &= ~PVF_NC;
1506 
1507 			l2b = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va);
1508 			ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1509 			pte = (*ptep & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode;
1510 
1511 			if (l2pte_valid(pte)) {
1512 				if (PV_BEEN_EXECD(pv->pv_flags)) {
1513 					pmap_tlb_flushID_SE(pv->pv_pmap,
1514 					    pv->pv_va);
1515 				} else
1516 				if (PV_BEEN_REFD(pv->pv_flags)) {
1517 					pmap_tlb_flushD_SE(pv->pv_pmap,
1518 					    pv->pv_va);
1519 				}
1520 			}
1521 
1522 			*ptep = pte;
1523 			PTE_SYNC_CURRENT(pv->pv_pmap, ptep);
1524 		}
1525 	}
1526 }
1527 
1528 /*
1529  * Modify pte bits for all ptes corresponding to the given physical address.
1530  * We use `maskbits' rather than `clearbits' because we're always passing
1531  * constants and the latter would require an extra inversion at run-time.
1532  */
1533 static void
1534 pmap_clearbit(struct vm_page *pg, u_int maskbits)
1535 {
1536 	struct l2_bucket *l2b;
1537 	struct pv_entry *pv;
1538 	pt_entry_t *ptep, npte, opte;
1539 	pmap_t pm;
1540 	vaddr_t va;
1541 	u_int oflags;
1542 
1543 	NPDEBUG(PDB_BITS,
1544 	    printf("pmap_clearbit: pg %p (0x%08lx) mask 0x%x\n",
1545 	    pg, pg->phys_addr, maskbits));
1546 
1547 	PMAP_HEAD_TO_MAP_LOCK();
1548 	simple_lock(&pg->mdpage.pvh_slock);
1549 
1550 	/*
1551 	 * Clear saved attributes (modify, reference)
1552 	 */
1553 	pg->mdpage.pvh_attrs &= ~(maskbits & (PVF_MOD | PVF_REF));
1554 
1555 	if (pg->mdpage.pvh_list == NULL) {
1556 		simple_unlock(&pg->mdpage.pvh_slock);
1557 		PMAP_HEAD_TO_MAP_UNLOCK();
1558 		return;
1559 	}
1560 
1561 	/*
1562 	 * Loop over all current mappings setting/clearing as appropos
1563 	 */
1564 	for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
1565 		va = pv->pv_va;
1566 		pm = pv->pv_pmap;
1567 		oflags = pv->pv_flags;
1568 		pv->pv_flags &= ~maskbits;
1569 
1570 		pmap_acquire_pmap_lock(pm);
1571 
1572 		l2b = pmap_get_l2_bucket(pm, va);
1573 		KDASSERT(l2b != NULL);
1574 
1575 		ptep = &l2b->l2b_kva[l2pte_index(va)];
1576 		npte = opte = *ptep;
1577 
1578 		NPDEBUG(PDB_BITS,
1579 		    printf(
1580 		    "pmap_clearbit: pv %p, pm %p, va 0x%08lx, flag 0x%x\n",
1581 		    pv, pv->pv_pmap, pv->pv_va, oflags));
1582 
1583 		if (maskbits & (PVF_WRITE|PVF_MOD)) {
1584 			if ((pv->pv_flags & PVF_NC)) {
1585 				/*
1586 				 * Entry is not cacheable:
1587 				 *
1588 				 * Don't turn caching on again if this is a
1589 				 * modified emulation. This would be
1590 				 * inconsitent with the settings created by
1591 				 * pmap_vac_me_harder(). Otherwise, it's safe
1592 				 * to re-enable cacheing.
1593 				 *
1594 				 * There's no need to call pmap_vac_me_harder()
1595 				 * here: all pages are losing their write
1596 				 * permission.
1597 				 */
1598 				if (maskbits & PVF_WRITE) {
1599 					npte |= pte_l2_s_cache_mode;
1600 					pv->pv_flags &= ~PVF_NC;
1601 				}
1602 			} else
1603 			if (opte & L2_S_PROT_W) {
1604 				/*
1605 				 * Entry is writable/cacheable: check if pmap
1606 				 * is current if it is flush it, otherwise it
1607 				 * won't be in the cache
1608 				 */
1609 				if (PV_BEEN_EXECD(oflags))
1610 					pmap_idcache_wbinv_range(pm, pv->pv_va,
1611 					    PAGE_SIZE);
1612 				else
1613 				if (PV_BEEN_REFD(oflags))
1614 					pmap_dcache_wb_range(pm, pv->pv_va,
1615 					    PAGE_SIZE,
1616 					    (maskbits & PVF_REF) ? TRUE : FALSE,
1617 					    FALSE);
1618 			}
1619 
1620 			/* make the pte read only */
1621 			npte &= ~L2_S_PROT_W;
1622 
1623 			if (maskbits & PVF_WRITE) {
1624 				/*
1625 				 * Keep alias accounting up to date
1626 				 */
1627 				if (pv->pv_pmap == pmap_kernel()) {
1628 					if (oflags & PVF_WRITE) {
1629 						pg->mdpage.krw_mappings--;
1630 						pg->mdpage.kro_mappings++;
1631 					}
1632 				} else
1633 				if (oflags & PVF_WRITE) {
1634 					pg->mdpage.urw_mappings--;
1635 					pg->mdpage.uro_mappings++;
1636 				}
1637 			}
1638 		}
1639 
1640 		if (maskbits & PVF_REF) {
1641 			if ((pv->pv_flags & PVF_NC) == 0 &&
1642 			    (maskbits & (PVF_WRITE|PVF_MOD)) == 0) {
1643 				/*
1644 				 * Check npte here; we may have already
1645 				 * done the wbinv above, and the validity
1646 				 * of the PTE is the same for opte and
1647 				 * npte.
1648 				 */
1649 				if (npte & L2_S_PROT_W) {
1650 					if (PV_BEEN_EXECD(oflags))
1651 						pmap_idcache_wbinv_range(pm,
1652 						    pv->pv_va, PAGE_SIZE);
1653 					else
1654 					if (PV_BEEN_REFD(oflags))
1655 						pmap_dcache_wb_range(pm,
1656 						    pv->pv_va, PAGE_SIZE,
1657 						    TRUE, FALSE);
1658 				} else
1659 				if ((npte & L2_TYPE_MASK) != L2_TYPE_INV) {
1660 					/* XXXJRT need idcache_inv_range */
1661 					if (PV_BEEN_EXECD(oflags))
1662 						pmap_idcache_wbinv_range(pm,
1663 						    pv->pv_va, PAGE_SIZE);
1664 					else
1665 					if (PV_BEEN_REFD(oflags))
1666 						pmap_dcache_wb_range(pm,
1667 						    pv->pv_va, PAGE_SIZE,
1668 						    TRUE, TRUE);
1669 				}
1670 			}
1671 
1672 			/*
1673 			 * Make the PTE invalid so that we will take a
1674 			 * page fault the next time the mapping is
1675 			 * referenced.
1676 			 */
1677 			npte &= ~L2_TYPE_MASK;
1678 			npte |= L2_TYPE_INV;
1679 		}
1680 
1681 		if (npte != opte) {
1682 			*ptep = npte;
1683 			PTE_SYNC(ptep);
1684 			/* Flush the TLB entry if a current pmap. */
1685 			if (PV_BEEN_EXECD(oflags))
1686 				pmap_tlb_flushID_SE(pm, pv->pv_va);
1687 			else
1688 			if (PV_BEEN_REFD(oflags))
1689 				pmap_tlb_flushD_SE(pm, pv->pv_va);
1690 		}
1691 
1692 		pmap_release_pmap_lock(pm);
1693 
1694 		NPDEBUG(PDB_BITS,
1695 		    printf("pmap_clearbit: pm %p va 0x%lx opte 0x%08x npte 0x%08x\n",
1696 		    pm, va, opte, npte));
1697 	}
1698 
1699 	simple_unlock(&pg->mdpage.pvh_slock);
1700 	PMAP_HEAD_TO_MAP_UNLOCK();
1701 }
1702 
1703 /*
1704  * pmap_clean_page()
1705  *
1706  * This is a local function used to work out the best strategy to clean
1707  * a single page referenced by its entry in the PV table. It's used by
1708  * pmap_copy_page, pmap_zero page and maybe some others later on.
1709  *
1710  * Its policy is effectively:
1711  *  o If there are no mappings, we don't bother doing anything with the cache.
1712  *  o If there is one mapping, we clean just that page.
1713  *  o If there are multiple mappings, we clean the entire cache.
1714  *
1715  * So that some functions can be further optimised, it returns 0 if it didn't
1716  * clean the entire cache, or 1 if it did.
1717  *
1718  * XXX One bug in this routine is that if the pv_entry has a single page
1719  * mapped at 0x00000000 a whole cache clean will be performed rather than
1720  * just the 1 page. Since this should not occur in everyday use and if it does
1721  * it will just result in not the most efficient clean for the page.
1722  */
1723 static int
1724 pmap_clean_page(struct pv_entry *pv, boolean_t is_src)
1725 {
1726 	pmap_t pm, pm_to_clean = NULL;
1727 	struct pv_entry *npv;
1728 	u_int cache_needs_cleaning = 0;
1729 	u_int flags = 0;
1730 	vaddr_t page_to_clean = 0;
1731 
1732 	if (pv == NULL) {
1733 		/* nothing mapped in so nothing to flush */
1734 		return (0);
1735 	}
1736 
1737 	/*
1738 	 * Since we flush the cache each time we change to a different
1739 	 * user vmspace, we only need to flush the page if it is in the
1740 	 * current pmap.
1741 	 */
1742 	if (curproc)
1743 		pm = curproc->p_vmspace->vm_map.pmap;
1744 	else
1745 		pm = pmap_kernel();
1746 
1747 	for (npv = pv; npv; npv = npv->pv_next) {
1748 		if (npv->pv_pmap == pmap_kernel() || npv->pv_pmap == pm) {
1749 			flags |= npv->pv_flags;
1750 			/*
1751 			 * The page is mapped non-cacheable in
1752 			 * this map.  No need to flush the cache.
1753 			 */
1754 			if (npv->pv_flags & PVF_NC) {
1755 #ifdef DIAGNOSTIC
1756 				if (cache_needs_cleaning)
1757 					panic("pmap_clean_page: "
1758 					    "cache inconsistency");
1759 #endif
1760 				break;
1761 			} else if (is_src && (npv->pv_flags & PVF_WRITE) == 0)
1762 				continue;
1763 			if (cache_needs_cleaning) {
1764 				page_to_clean = 0;
1765 				break;
1766 			} else {
1767 				page_to_clean = npv->pv_va;
1768 				pm_to_clean = npv->pv_pmap;
1769 			}
1770 			cache_needs_cleaning = 1;
1771 		}
1772 	}
1773 
1774 	if (page_to_clean) {
1775 		if (PV_BEEN_EXECD(flags))
1776 			pmap_idcache_wbinv_range(pm_to_clean, page_to_clean,
1777 			    PAGE_SIZE);
1778 		else
1779 			pmap_dcache_wb_range(pm_to_clean, page_to_clean,
1780 			    PAGE_SIZE, !is_src, (flags & PVF_WRITE) == 0);
1781 	} else if (cache_needs_cleaning) {
1782 		if (PV_BEEN_EXECD(flags))
1783 			pmap_idcache_wbinv_all(pm);
1784 		else
1785 			pmap_dcache_wbinv_all(pm);
1786 		return (1);
1787 	}
1788 	return (0);
1789 }
1790 
1791 /*
1792  * Routine:	pmap_page_remove
1793  * Function:
1794  *		Removes this physical page from
1795  *		all physical maps in which it resides.
1796  *		Reflects back modify bits to the pager.
1797  */
1798 static void
1799 pmap_page_remove(struct vm_page *pg)
1800 {
1801 	struct l2_bucket *l2b;
1802 	struct pv_entry *pv, *npv;
1803 	pmap_t pm, curpm;
1804 	pt_entry_t *ptep, pte;
1805 	boolean_t flush;
1806 	u_int flags;
1807 
1808 	NPDEBUG(PDB_FOLLOW,
1809 	    printf("pmap_page_remove: pg %p (0x%08lx)\n", pg, pg->phys_addr));
1810 
1811 	PMAP_HEAD_TO_MAP_LOCK();
1812 	simple_lock(&pg->mdpage.pvh_slock);
1813 
1814 	pv = pg->mdpage.pvh_list;
1815 	if (pv == NULL) {
1816 		simple_unlock(&pg->mdpage.pvh_slock);
1817 		PMAP_HEAD_TO_MAP_UNLOCK();
1818 		return;
1819 	}
1820 
1821 	/*
1822 	 * Clear alias counts
1823 	 */
1824 	pg->mdpage.k_mappings = 0;
1825 	pg->mdpage.urw_mappings = pg->mdpage.uro_mappings = 0;
1826 
1827 	flush = FALSE;
1828 	flags = 0;
1829 	if (curproc)
1830 		curpm = curproc->p_vmspace->vm_map.pmap;
1831 	else
1832 		curpm = pmap_kernel();
1833 
1834 	pmap_clean_page(pv, FALSE);
1835 
1836 	while (pv) {
1837 		pm = pv->pv_pmap;
1838 		if (flush == FALSE && (pm == curpm || pm == pmap_kernel()))
1839 			flush = TRUE;
1840 
1841 		pmap_acquire_pmap_lock(pm);
1842 
1843 		l2b = pmap_get_l2_bucket(pm, pv->pv_va);
1844 		KDASSERT(l2b != NULL);
1845 
1846 		ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)];
1847 		pte = *ptep;
1848 
1849 		/*
1850 		 * Update statistics
1851 		 */
1852 		--pm->pm_stats.resident_count;
1853 
1854 		/* Wired bit */
1855 		if (pv->pv_flags & PVF_WIRED)
1856 			--pm->pm_stats.wired_count;
1857 
1858 		flags |= pv->pv_flags;
1859 
1860 		/*
1861 		 * Invalidate the PTEs.
1862 		 */
1863 		*ptep = 0;
1864 		PTE_SYNC_CURRENT(pm, ptep);
1865 		pmap_free_l2_bucket(pm, l2b, 1);
1866 
1867 		npv = pv->pv_next;
1868 		pool_put(&pmap_pv_pool, pv);
1869 		pv = npv;
1870 		pmap_release_pmap_lock(pm);
1871 	}
1872 	pg->mdpage.pvh_list = NULL;
1873 	simple_unlock(&pg->mdpage.pvh_slock);
1874 	PMAP_HEAD_TO_MAP_UNLOCK();
1875 
1876 	if (flush) {
1877 		if (PV_BEEN_EXECD(flags))
1878 			pmap_tlb_flushID(curpm);
1879 		else
1880 			pmap_tlb_flushD(curpm);
1881 	}
1882 	cpu_cpwait();
1883 }
1884 
1885 /*
1886  * pmap_t pmap_create(void)
1887  *
1888  *      Create a new pmap structure from scratch.
1889  */
1890 pmap_t
1891 pmap_create(void)
1892 {
1893 	pmap_t pm;
1894 
1895 	pm = pool_cache_get(&pmap_pmap_cache, PR_WAITOK);
1896 
1897 	simple_lock_init(&pm->pm_lock);
1898 	pm->pm_obj.pgops = NULL;	/* currently not a mappable object */
1899 	TAILQ_INIT(&pm->pm_obj.memq);
1900 	pm->pm_obj.uo_npages = 0;
1901 	pm->pm_obj.uo_refs = 1;
1902 	pm->pm_stats.wired_count = 0;
1903 	pm->pm_stats.resident_count = 1;
1904 	pm->pm_cstate.cs_all = 0;
1905 	pmap_alloc_l1(pm);
1906 
1907 	/*
1908 	 * Note: The pool cache ensures that the pm_l2[] array is already
1909 	 * initialised to zero.
1910 	 */
1911 
1912 	pmap_pinit(pm);
1913 
1914 	LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list);
1915 
1916 	return (pm);
1917 }
1918 
1919 /*
1920  * void pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot,
1921  *     int flags)
1922  *
1923  *      Insert the given physical page (p) at
1924  *      the specified virtual address (v) in the
1925  *      target physical map with the protection requested.
1926  *
1927  *      NB:  This is the only routine which MAY NOT lazy-evaluate
1928  *      or lose information.  That is, this routine must actually
1929  *      insert this page into the given map NOW.
1930  */
1931 int
1932 pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot, int flags)
1933 {
1934 	struct l2_bucket *l2b;
1935 	struct vm_page *pg, *opg;
1936 	struct pv_entry *pve;
1937 	pt_entry_t *ptep, npte, opte;
1938 	u_int nflags;
1939 	u_int oflags;
1940 
1941 	NPDEBUG(PDB_ENTER, printf("pmap_enter: pm %p va 0x%lx pa 0x%lx prot %x flag %x\n", pm, va, pa, prot, flags));
1942 
1943 	KDASSERT((flags & PMAP_WIRED) == 0 || (flags & VM_PROT_ALL) != 0);
1944 	KDASSERT(((va | pa) & PGOFSET) == 0);
1945 
1946 	/*
1947 	 * Get a pointer to the page.  Later on in this function, we
1948 	 * test for a managed page by checking pg != NULL.
1949 	 */
1950 	pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL;
1951 
1952 	nflags = 0;
1953 	if (prot & VM_PROT_WRITE)
1954 		nflags |= PVF_WRITE;
1955 	if (prot & VM_PROT_EXECUTE)
1956 		nflags |= PVF_EXEC;
1957 	if (flags & PMAP_WIRED)
1958 		nflags |= PVF_WIRED;
1959 
1960 	PMAP_MAP_TO_HEAD_LOCK();
1961 	pmap_acquire_pmap_lock(pm);
1962 
1963 	/*
1964 	 * Fetch the L2 bucket which maps this page, allocating one if
1965 	 * necessary for user pmaps.
1966 	 */
1967 	if (pm == pmap_kernel())
1968 		l2b = pmap_get_l2_bucket(pm, va);
1969 	else
1970 		l2b = pmap_alloc_l2_bucket(pm, va);
1971 	if (l2b == NULL) {
1972 		if (flags & PMAP_CANFAIL) {
1973 			pmap_release_pmap_lock(pm);
1974 			PMAP_MAP_TO_HEAD_UNLOCK();
1975 			return (ENOMEM);
1976 		}
1977 		panic("pmap_enter: failed to allocate L2 bucket");
1978 	}
1979 	ptep = &l2b->l2b_kva[l2pte_index(va)];
1980 	opte = *ptep;
1981 	npte = pa;
1982 	oflags = 0;
1983 
1984 	if (opte) {
1985 		/*
1986 		 * There is already a mapping at this address.
1987 		 * If the physical address is different, lookup the
1988 		 * vm_page.
1989 		 */
1990 		if (l2pte_pa(opte) != pa)
1991 			opg = PHYS_TO_VM_PAGE(l2pte_pa(opte));
1992 		else
1993 			opg = pg;
1994 	} else
1995 		opg = NULL;
1996 
1997 	if (pg) {
1998 		/*
1999 		 * This is to be a managed mapping.
2000 		 */
2001 		if ((flags & VM_PROT_ALL) ||
2002 		    (pg->mdpage.pvh_attrs & PVF_REF)) {
2003 			/*
2004 			 * - The access type indicates that we don't need
2005 			 *   to do referenced emulation.
2006 			 * OR
2007 			 * - The physical page has already been referenced
2008 			 *   so no need to re-do referenced emulation here.
2009 			 */
2010 			npte |= L2_S_PROTO;
2011 
2012 			nflags |= PVF_REF;
2013 
2014 			if ((prot & VM_PROT_WRITE) != 0 &&
2015 			    ((flags & VM_PROT_WRITE) != 0 ||
2016 			     (pg->mdpage.pvh_attrs & PVF_MOD) != 0)) {
2017 				/*
2018 				 * This is a writable mapping, and the
2019 				 * page's mod state indicates it has
2020 				 * already been modified. Make it
2021 				 * writable from the outset.
2022 				 */
2023 				npte |= L2_S_PROT_W;
2024 				nflags |= PVF_MOD;
2025 			}
2026 		} else {
2027 			/*
2028 			 * Need to do page referenced emulation.
2029 			 */
2030 			npte |= L2_TYPE_INV;
2031 		}
2032 
2033 		npte |= pte_l2_s_cache_mode;
2034 
2035 		if (pg == opg) {
2036 			/*
2037 			 * We're changing the attrs of an existing mapping.
2038 			 */
2039 			simple_lock(&pg->mdpage.pvh_slock);
2040 			oflags = pmap_modify_pv(pg, pm, va,
2041 			    PVF_WRITE | PVF_EXEC | PVF_WIRED |
2042 			    PVF_MOD | PVF_REF, nflags);
2043 			simple_unlock(&pg->mdpage.pvh_slock);
2044 
2045 			/*
2046 			 * We may need to flush the cache if we're
2047 			 * doing rw-ro...
2048 			 */
2049 			if (pm->pm_cstate.cs_cache_d &&
2050 			    (oflags & PVF_NC) == 0 &&
2051 			    (opte & L2_S_PROT_W) != 0 &&
2052 			    (prot & VM_PROT_WRITE) == 0)
2053 				cpu_dcache_wb_range(va, PAGE_SIZE);
2054 		} else {
2055 			/*
2056 			 * New mapping, or changing the backing page
2057 			 * of an existing mapping.
2058 			 */
2059 			if (opg) {
2060 				/*
2061 				 * Replacing an existing mapping with a new one.
2062 				 * It is part of our managed memory so we
2063 				 * must remove it from the PV list
2064 				 */
2065 				simple_lock(&opg->mdpage.pvh_slock);
2066 				pve = pmap_remove_pv(opg, pm, va);
2067 				pmap_vac_me_harder(opg, pm, 0);
2068 				simple_unlock(&opg->mdpage.pvh_slock);
2069 				oflags = pve->pv_flags;
2070 
2071 				/*
2072 				 * If the old mapping was valid (ref/mod
2073 				 * emulation creates 'invalid' mappings
2074 				 * initially) then make sure to frob
2075 				 * the cache.
2076 				 */
2077 				if ((oflags & PVF_NC) == 0 &&
2078 				    l2pte_valid(opte)) {
2079 					if (PV_BEEN_EXECD(oflags)) {
2080 						pmap_idcache_wbinv_range(pm, va,
2081 						    PAGE_SIZE);
2082 					} else
2083 					if (PV_BEEN_REFD(oflags)) {
2084 						pmap_dcache_wb_range(pm, va,
2085 						    PAGE_SIZE, TRUE,
2086 						    (oflags & PVF_WRITE) == 0);
2087 					}
2088 				}
2089 			} else
2090 			if ((pve = pool_get(&pmap_pv_pool, PR_NOWAIT)) == NULL){
2091 				if ((flags & PMAP_CANFAIL) == 0)
2092 					panic("pmap_enter: no pv entries");
2093 
2094 				if (pm != pmap_kernel())
2095 					pmap_free_l2_bucket(pm, l2b, 0);
2096 				pmap_release_pmap_lock(pm);
2097 				PMAP_MAP_TO_HEAD_UNLOCK();
2098 				NPDEBUG(PDB_ENTER,
2099 				    printf("pmap_enter: ENOMEM\n"));
2100 				return (ENOMEM);
2101 			}
2102 
2103 			pmap_enter_pv(pg, pve, pm, va, nflags);
2104 		}
2105 	} else {
2106 		/*
2107 		 * We're mapping an unmanaged page.
2108 		 * These are always readable, and possibly writable, from
2109 		 * the get go as we don't need to track ref/mod status.
2110 		 */
2111 		npte |= L2_S_PROTO;
2112 		if (prot & VM_PROT_WRITE)
2113 			npte |= L2_S_PROT_W;
2114 
2115 		/*
2116 		 * Make sure the vector table is mapped cacheable
2117 		 */
2118 		if (pm != pmap_kernel() && va == vector_page)
2119 			npte |= pte_l2_s_cache_mode;
2120 
2121 		if (opg) {
2122 			/*
2123 			 * Looks like there's an existing 'managed' mapping
2124 			 * at this address.
2125 			 */
2126 			simple_lock(&opg->mdpage.pvh_slock);
2127 			pve = pmap_remove_pv(opg, pm, va);
2128 			pmap_vac_me_harder(opg, pm, 0);
2129 			simple_unlock(&opg->mdpage.pvh_slock);
2130 			oflags = pve->pv_flags;
2131 
2132 			if ((oflags & PVF_NC) == 0 && l2pte_valid(opte)) {
2133 				if (PV_BEEN_EXECD(oflags))
2134 					pmap_idcache_wbinv_range(pm, va,
2135 					    PAGE_SIZE);
2136 				else
2137 				if (PV_BEEN_REFD(oflags))
2138 					pmap_dcache_wb_range(pm, va, PAGE_SIZE,
2139 					    TRUE, (oflags & PVF_WRITE) == 0);
2140 			}
2141 			pool_put(&pmap_pv_pool, pve);
2142 		}
2143 	}
2144 
2145 	/*
2146 	 * Make sure userland mappings get the right permissions
2147 	 */
2148 	if (pm != pmap_kernel() && va != vector_page)
2149 		npte |= L2_S_PROT_U;
2150 
2151 	/*
2152 	 * Keep the stats up to date
2153 	 */
2154 	if (opte == 0) {
2155 		l2b->l2b_occupancy++;
2156 		pm->pm_stats.resident_count++;
2157 	}
2158 
2159 	NPDEBUG(PDB_ENTER,
2160 	    printf("pmap_enter: opte 0x%08x npte 0x%08x\n", opte, npte));
2161 
2162 	/*
2163 	 * If this is just a wiring change, the two PTEs will be
2164 	 * identical, so there's no need to update the page table.
2165 	 */
2166 	if (npte != opte) {
2167 		boolean_t is_cached = pmap_is_cached(pm);
2168 
2169 		*ptep = npte;
2170 		if (is_cached) {
2171 			/*
2172 			 * We only need to frob the cache/tlb if this pmap
2173 			 * is current
2174 			 */
2175 			PTE_SYNC(ptep);
2176 			if (va != vector_page && l2pte_valid(npte)) {
2177 				/*
2178 				 * This mapping is likely to be accessed as
2179 				 * soon as we return to userland. Fix up the
2180 				 * L1 entry to avoid taking another
2181 				 * page/domain fault.
2182 				 */
2183 				pd_entry_t *pl1pd, l1pd;
2184 
2185 				pl1pd = &pm->pm_l1->l1_kva[L1_IDX(va)];
2186 				l1pd = l2b->l2b_phys | L1_C_DOM(pm->pm_domain) |
2187 				    L1_C_PROTO;
2188 				if (*pl1pd != l1pd) {
2189 					*pl1pd = l1pd;
2190 					PTE_SYNC(pl1pd);
2191 				}
2192 			}
2193 		}
2194 
2195 		if (PV_BEEN_EXECD(oflags))
2196 			pmap_tlb_flushID_SE(pm, va);
2197 		else
2198 		if (PV_BEEN_REFD(oflags))
2199 			pmap_tlb_flushD_SE(pm, va);
2200 
2201 		NPDEBUG(PDB_ENTER,
2202 		    printf("pmap_enter: is_cached %d cs 0x%08x\n",
2203 		    is_cached, pm->pm_cstate.cs_all));
2204 
2205 		if (pg != NULL) {
2206 			simple_lock(&pg->mdpage.pvh_slock);
2207 			pmap_vac_me_harder(pg, pm, va);
2208 			simple_unlock(&pg->mdpage.pvh_slock);
2209 		}
2210 	}
2211 
2212 	pmap_release_pmap_lock(pm);
2213 	PMAP_MAP_TO_HEAD_UNLOCK();
2214 
2215 	return (0);
2216 }
2217 
2218 /*
2219  * pmap_remove()
2220  *
2221  * pmap_remove is responsible for nuking a number of mappings for a range
2222  * of virtual address space in the current pmap. To do this efficiently
2223  * is interesting, because in a number of cases a wide virtual address
2224  * range may be supplied that contains few actual mappings. So, the
2225  * optimisations are:
2226  *  1. Skip over hunks of address space for which no L1 or L2 entry exists.
2227  *  2. Build up a list of pages we've hit, up to a maximum, so we can
2228  *     maybe do just a partial cache clean. This path of execution is
2229  *     complicated by the fact that the cache must be flushed _before_
2230  *     the PTE is nuked, being a VAC :-)
2231  *  3. If we're called after UVM calls pmap_remove_all(), we can defer
2232  *     all invalidations until pmap_update(), since pmap_remove_all() has
2233  *     already flushed the cache.
2234  *  4. Maybe later fast-case a single page, but I don't think this is
2235  *     going to make _that_ much difference overall.
2236  */
2237 
2238 #define	PMAP_REMOVE_CLEAN_LIST_SIZE	3
2239 
2240 void
2241 pmap_remove(pmap_t pm, vaddr_t sva, vaddr_t eva)
2242 {
2243 	struct l2_bucket *l2b;
2244 	vaddr_t next_bucket;
2245 	pt_entry_t *ptep;
2246 	u_int cleanlist_idx, total, cnt;
2247 	struct {
2248 		vaddr_t va;
2249 		pt_entry_t *pte;
2250 	} cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE];
2251 	u_int mappings, is_exec, is_refd;
2252 
2253 	NPDEBUG(PDB_REMOVE, printf("pmap_remove: pmap=%p sva=%08lx eva=%08lx\n",
2254 	    pm, sva, eva));
2255 
2256 	/*
2257 	 * we lock in the pmap => pv_head direction
2258 	 */
2259 	PMAP_MAP_TO_HEAD_LOCK();
2260 	pmap_acquire_pmap_lock(pm);
2261 
2262 	if (pm->pm_remove_all || !pmap_is_cached(pm)) {
2263 		cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
2264 		if (pm->pm_cstate.cs_tlb == 0)
2265 			pm->pm_remove_all = TRUE;
2266 	} else
2267 		cleanlist_idx = 0;
2268 
2269 	total = 0;
2270 
2271 	while (sva < eva) {
2272 		/*
2273 		 * Do one L2 bucket's worth at a time.
2274 		 */
2275 		next_bucket = L2_NEXT_BUCKET(sva);
2276 		if (next_bucket > eva)
2277 			next_bucket = eva;
2278 
2279 		l2b = pmap_get_l2_bucket(pm, sva);
2280 		if (l2b == NULL) {
2281 			sva = next_bucket;
2282 			continue;
2283 		}
2284 
2285 		ptep = &l2b->l2b_kva[l2pte_index(sva)];
2286 		mappings = 0;
2287 
2288 		while (sva < next_bucket) {
2289 			struct vm_page *pg;
2290 			pt_entry_t pte;
2291 			paddr_t pa;
2292 
2293 			pte = *ptep;
2294 
2295 			if (pte == 0) {
2296 				/*
2297 				 * Nothing here, move along
2298 				 */
2299 				sva += PAGE_SIZE;
2300 				ptep++;
2301 				continue;
2302 			}
2303 
2304 			pm->pm_stats.resident_count--;
2305 			pa = l2pte_pa(pte);
2306 			is_exec = 0;
2307 			is_refd = 1;
2308 
2309 			/*
2310 			 * Update flags. In a number of circumstances,
2311 			 * we could cluster a lot of these and do a
2312 			 * number of sequential pages in one go.
2313 			 */
2314 			if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
2315 				struct pv_entry *pve;
2316 				simple_lock(&pg->mdpage.pvh_slock);
2317 				pve = pmap_remove_pv(pg, pm, sva);
2318 				pmap_vac_me_harder(pg, pm, 0);
2319 				simple_unlock(&pg->mdpage.pvh_slock);
2320 				if (pve != NULL) {
2321 					if (pm->pm_remove_all == FALSE) {
2322 						is_exec =
2323 						   PV_BEEN_EXECD(pve->pv_flags);
2324 						is_refd =
2325 						   PV_BEEN_REFD(pve->pv_flags);
2326 					}
2327 					pool_put(&pmap_pv_pool, pve);
2328 				}
2329 			}
2330 
2331 			if (!l2pte_valid(pte)) {
2332 				*ptep = 0;
2333 				PTE_SYNC_CURRENT(pm, ptep);
2334 				sva += PAGE_SIZE;
2335 				ptep++;
2336 				mappings++;
2337 				continue;
2338 			}
2339 
2340 			if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) {
2341 				/* Add to the clean list. */
2342 				cleanlist[cleanlist_idx].pte = ptep;
2343 				cleanlist[cleanlist_idx].va =
2344 				    sva | (is_exec & 1);
2345 				cleanlist_idx++;
2346 			} else
2347 			if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) {
2348 				/* Nuke everything if needed. */
2349 				pmap_idcache_wbinv_all(pm);
2350 				pmap_tlb_flushID(pm);
2351 
2352 				/*
2353 				 * Roll back the previous PTE list,
2354 				 * and zero out the current PTE.
2355 				 */
2356 				for (cnt = 0;
2357 				     cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) {
2358 					*cleanlist[cnt].pte = 0;
2359 				}
2360 				*ptep = 0;
2361 				PTE_SYNC(ptep);
2362 				cleanlist_idx++;
2363 				pm->pm_remove_all = TRUE;
2364 			} else {
2365 				*ptep = 0;
2366 				PTE_SYNC(ptep);
2367 				if (pm->pm_remove_all == FALSE) {
2368 					if (is_exec)
2369 						pmap_tlb_flushID_SE(pm, sva);
2370 					else
2371 					if (is_refd)
2372 						pmap_tlb_flushD_SE(pm, sva);
2373 				}
2374 			}
2375 
2376 			sva += PAGE_SIZE;
2377 			ptep++;
2378 			mappings++;
2379 		}
2380 
2381 		/*
2382 		 * Deal with any left overs
2383 		 */
2384 		if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) {
2385 			total += cleanlist_idx;
2386 			for (cnt = 0; cnt < cleanlist_idx; cnt++) {
2387 				if (pm->pm_cstate.cs_all != 0) {
2388 					vaddr_t clva = cleanlist[cnt].va & ~1;
2389 					if (cleanlist[cnt].va & 1) {
2390 						pmap_idcache_wbinv_range(pm,
2391 						    clva, PAGE_SIZE);
2392 						pmap_tlb_flushID_SE(pm, clva);
2393 					} else {
2394 						pmap_dcache_wb_range(pm,
2395 						    clva, PAGE_SIZE, TRUE,
2396 						    FALSE);
2397 						pmap_tlb_flushD_SE(pm, clva);
2398 					}
2399 				}
2400 				*cleanlist[cnt].pte = 0;
2401 				PTE_SYNC_CURRENT(pm, cleanlist[cnt].pte);
2402 			}
2403 
2404 			/*
2405 			 * If it looks like we're removing a whole bunch
2406 			 * of mappings, it's faster to just write-back
2407 			 * the whole cache now and defer TLB flushes until
2408 			 * pmap_update() is called.
2409 			 */
2410 			if (total <= PMAP_REMOVE_CLEAN_LIST_SIZE)
2411 				cleanlist_idx = 0;
2412 			else {
2413 				cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1;
2414 				pmap_idcache_wbinv_all(pm);
2415 				pm->pm_remove_all = TRUE;
2416 			}
2417 		}
2418 
2419 		pmap_free_l2_bucket(pm, l2b, mappings);
2420 	}
2421 
2422 	pmap_release_pmap_lock(pm);
2423 	PMAP_MAP_TO_HEAD_UNLOCK();
2424 }
2425 
2426 /*
2427  * pmap_kenter_pa: enter an unmanaged, wired kernel mapping
2428  *
2429  * We assume there is already sufficient KVM space available
2430  * to do this, as we can't allocate L2 descriptor tables/metadata
2431  * from here.
2432  */
2433 void
2434 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot)
2435 {
2436 	struct l2_bucket *l2b;
2437 	pt_entry_t *ptep, opte;
2438 
2439 	NPDEBUG(PDB_KENTER,
2440 	    printf("pmap_kenter_pa: va 0x%08lx, pa 0x%08lx, prot 0x%x\n",
2441 	    va, pa, prot));
2442 
2443 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
2444 	KDASSERT(l2b != NULL);
2445 
2446 	ptep = &l2b->l2b_kva[l2pte_index(va)];
2447 	opte = *ptep;
2448 
2449 	if (l2pte_valid(opte)) {
2450 		cpu_dcache_wbinv_range(va, PAGE_SIZE);
2451 		cpu_tlb_flushD_SE(va);
2452 		cpu_cpwait();
2453 	} else
2454 	if (opte == 0)
2455 		l2b->l2b_occupancy++;
2456 
2457 	*ptep = L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) |
2458 	    pte_l2_s_cache_mode;
2459 	PTE_SYNC(ptep);
2460 }
2461 
2462 void
2463 pmap_kremove(vaddr_t va, vsize_t len)
2464 {
2465 	struct l2_bucket *l2b;
2466 	pt_entry_t *ptep, *sptep, opte;
2467 	vaddr_t next_bucket, eva;
2468 	u_int mappings;
2469 
2470 	NPDEBUG(PDB_KREMOVE, printf("pmap_kremove: va 0x%08lx, len 0x%08lx\n",
2471 	    va, len));
2472 
2473 	eva = va + len;
2474 
2475 	while (va < eva) {
2476 		next_bucket = L2_NEXT_BUCKET(va);
2477 		if (next_bucket > eva)
2478 			next_bucket = eva;
2479 
2480 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
2481 		KDASSERT(l2b != NULL);
2482 
2483 		sptep = ptep = &l2b->l2b_kva[l2pte_index(va)];
2484 		mappings = 0;
2485 
2486 		while (va < next_bucket) {
2487 			opte = *ptep;
2488 			if (l2pte_valid(opte)) {
2489 				cpu_dcache_wbinv_range(va, PAGE_SIZE);
2490 				cpu_tlb_flushD_SE(va);
2491 			}
2492 			if (opte) {
2493 				*ptep = 0;
2494 				mappings++;
2495 			}
2496 			va += PAGE_SIZE;
2497 			ptep++;
2498 		}
2499 		KDASSERT(mappings <= l2b->l2b_occupancy);
2500 		l2b->l2b_occupancy -= mappings;
2501 		PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep));
2502 	}
2503 	cpu_cpwait();
2504 }
2505 
2506 boolean_t
2507 pmap_extract(pmap_t pm, vaddr_t va, paddr_t *pap)
2508 {
2509 	struct l2_dtable *l2;
2510 	pd_entry_t *pl1pd, l1pd;
2511 	pt_entry_t *ptep, pte;
2512 	paddr_t pa;
2513 	u_int l1idx;
2514 
2515 	pmap_acquire_pmap_lock(pm);
2516 
2517 	l1idx = L1_IDX(va);
2518 	pl1pd = &pm->pm_l1->l1_kva[l1idx];
2519 	l1pd = *pl1pd;
2520 
2521 	if (l1pte_section_p(l1pd)) {
2522 		/*
2523 		 * These should only happen for pmap_kernel()
2524 		 */
2525 		KDASSERT(pm == pmap_kernel());
2526 		pmap_release_pmap_lock(pm);
2527 		pa = (l1pd & L1_S_FRAME) | (va & L1_S_OFFSET);
2528 	} else {
2529 		/*
2530 		 * Note that we can't rely on the validity of the L1
2531 		 * descriptor as an indication that a mapping exists.
2532 		 * We have to look it up in the L2 dtable.
2533 		 */
2534 		l2 = pm->pm_l2[L2_IDX(l1idx)];
2535 
2536 		if (l2 == NULL ||
2537 		    (ptep = l2->l2_bucket[L2_BUCKET(l1idx)].l2b_kva) == NULL) {
2538 			pmap_release_pmap_lock(pm);
2539 			return (FALSE);
2540 		}
2541 
2542 		ptep = &ptep[l2pte_index(va)];
2543 		pte = *ptep;
2544 		pmap_release_pmap_lock(pm);
2545 
2546 		if (pte == 0)
2547 			return (FALSE);
2548 
2549 		switch (pte & L2_TYPE_MASK) {
2550 		case L2_TYPE_L:
2551 			pa = (pte & L2_L_FRAME) | (va & L2_L_OFFSET);
2552 			break;
2553 
2554 		default:
2555 			pa = (pte & L2_S_FRAME) | (va & L2_S_OFFSET);
2556 			break;
2557 		}
2558 	}
2559 
2560 	if (pap != NULL)
2561 		*pap = pa;
2562 
2563 	return (TRUE);
2564 }
2565 
2566 void
2567 pmap_protect(pmap_t pm, vaddr_t sva, vaddr_t eva, vm_prot_t prot)
2568 {
2569 	struct l2_bucket *l2b;
2570 	pt_entry_t *ptep, pte;
2571 	vaddr_t next_bucket;
2572 	u_int flags;
2573 	int flush;
2574 
2575 	NPDEBUG(PDB_PROTECT,
2576 	    printf("pmap_protect: pm %p sva 0x%lx eva 0x%lx prot 0x%x\n",
2577 	    pm, sva, eva, prot));
2578 
2579 	if ((prot & VM_PROT_READ) == 0) {
2580 		pmap_remove(pm, sva, eva);
2581 		return;
2582 	}
2583 
2584 	if (prot & VM_PROT_WRITE) {
2585 		/*
2586 		 * If this is a read->write transition, just ignore it and let
2587 		 * uvm_fault() take care of it later.
2588 		 */
2589 		return;
2590 	}
2591 
2592 	PMAP_MAP_TO_HEAD_LOCK();
2593 	pmap_acquire_pmap_lock(pm);
2594 
2595 	/*
2596 	 * OK, at this point, we know we're doing write-protect operation.
2597 	 * If the pmap is active, write-back the range.
2598 	 */
2599 	pmap_dcache_wb_range(pm, sva, eva - sva, FALSE, FALSE);
2600 
2601 	flush = ((eva - sva) >= (PAGE_SIZE * 4)) ? 0 : -1;
2602 	flags = 0;
2603 
2604 	while (sva < eva) {
2605 		next_bucket = L2_NEXT_BUCKET(sva);
2606 		if (next_bucket > eva)
2607 			next_bucket = eva;
2608 
2609 		l2b = pmap_get_l2_bucket(pm, sva);
2610 		if (l2b == NULL) {
2611 			sva = next_bucket;
2612 			continue;
2613 		}
2614 
2615 		ptep = &l2b->l2b_kva[l2pte_index(sva)];
2616 
2617 		while (sva < next_bucket) {
2618 			if ((pte = *ptep) != 0 && (pte & L2_S_PROT_W) != 0) {
2619 				struct vm_page *pg;
2620 				u_int f;
2621 
2622 				pg = PHYS_TO_VM_PAGE(l2pte_pa(pte));
2623 				pte &= ~L2_S_PROT_W;
2624 				*ptep = pte;
2625 				PTE_SYNC(ptep);
2626 
2627 				if (pg != NULL) {
2628 					simple_lock(&pg->mdpage.pvh_slock);
2629 					f = pmap_modify_pv(pg, pm, sva,
2630 					    PVF_WRITE, 0);
2631 					pmap_vac_me_harder(pg, pm, sva);
2632 					simple_unlock(&pg->mdpage.pvh_slock);
2633 				} else
2634 					f = PVF_REF | PVF_EXEC;
2635 
2636 				if (flush >= 0) {
2637 					flush++;
2638 					flags |= f;
2639 				} else
2640 				if (PV_BEEN_EXECD(f))
2641 					pmap_tlb_flushID_SE(pm, sva);
2642 				else
2643 				if (PV_BEEN_REFD(f))
2644 					pmap_tlb_flushD_SE(pm, sva);
2645 			}
2646 
2647 			sva += PAGE_SIZE;
2648 			ptep++;
2649 		}
2650 	}
2651 
2652 	pmap_release_pmap_lock(pm);
2653 	PMAP_MAP_TO_HEAD_UNLOCK();
2654 
2655 	if (flush) {
2656 		if (PV_BEEN_EXECD(flags))
2657 			pmap_tlb_flushID(pm);
2658 		else
2659 		if (PV_BEEN_REFD(flags))
2660 			pmap_tlb_flushD(pm);
2661 	}
2662 }
2663 
2664 void
2665 pmap_page_protect(struct vm_page *pg, vm_prot_t prot)
2666 {
2667 
2668 	NPDEBUG(PDB_PROTECT,
2669 	    printf("pmap_page_protect: pg %p (0x%08lx), prot 0x%x\n",
2670 	    pg, pg->phys_addr, prot));
2671 
2672 	switch(prot) {
2673 	case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE:
2674 	case VM_PROT_READ|VM_PROT_WRITE:
2675 		return;
2676 
2677 	case VM_PROT_READ:
2678 	case VM_PROT_READ|VM_PROT_EXECUTE:
2679 		pmap_clearbit(pg, PVF_WRITE);
2680 		break;
2681 
2682 	default:
2683 		pmap_page_remove(pg);
2684 		break;
2685 	}
2686 }
2687 
2688 /*
2689  * pmap_clear_modify:
2690  *
2691  *	Clear the "modified" attribute for a page.
2692  */
2693 boolean_t
2694 pmap_clear_modify(struct vm_page *pg)
2695 {
2696 	boolean_t rv;
2697 
2698 	if (pg->mdpage.pvh_attrs & PVF_MOD) {
2699 		rv = TRUE;
2700 		pmap_clearbit(pg, PVF_MOD);
2701 	} else
2702 		rv = FALSE;
2703 
2704 	return (rv);
2705 }
2706 
2707 /*
2708  * pmap_clear_reference:
2709  *
2710  *	Clear the "referenced" attribute for a page.
2711  */
2712 boolean_t
2713 pmap_clear_reference(struct vm_page *pg)
2714 {
2715 	boolean_t rv;
2716 
2717 	if (pg->mdpage.pvh_attrs & PVF_REF) {
2718 		rv = TRUE;
2719 		pmap_clearbit(pg, PVF_REF);
2720 	} else
2721 		rv = FALSE;
2722 
2723 	return (rv);
2724 }
2725 
2726 /*
2727  * pmap_is_modified:
2728  *
2729  *	Test if a page has the "modified" attribute.
2730  */
2731 /* See <arm/arm32/pmap.h> */
2732 
2733 /*
2734  * pmap_is_referenced:
2735  *
2736  *	Test if a page has the "referenced" attribute.
2737  */
2738 /* See <arm/arm32/pmap.h> */
2739 
2740 int
2741 pmap_fault_fixup(pmap_t pm, vaddr_t va, vm_prot_t ftype, int user)
2742 {
2743 	struct l2_dtable *l2;
2744 	struct l2_bucket *l2b;
2745 	pd_entry_t *pl1pd, l1pd;
2746 	pt_entry_t *ptep, pte;
2747 	paddr_t pa;
2748 	u_int l1idx;
2749 	int rv = 0;
2750 
2751 	PMAP_MAP_TO_HEAD_LOCK();
2752 	pmap_acquire_pmap_lock(pm);
2753 
2754 	l1idx = L1_IDX(va);
2755 
2756 	/*
2757 	 * If there is no l2_dtable for this address, then the process
2758 	 * has no business accessing it.
2759 	 *
2760 	 * Note: This will catch userland processes trying to access
2761 	 * kernel addresses.
2762 	 */
2763 	l2 = pm->pm_l2[L2_IDX(l1idx)];
2764 	if (l2 == NULL)
2765 		goto out;
2766 
2767 	/*
2768 	 * Likewise if there is no L2 descriptor table
2769 	 */
2770 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
2771 	if (l2b->l2b_kva == NULL)
2772 		goto out;
2773 
2774 	/*
2775 	 * Check the PTE itself.
2776 	 */
2777 	ptep = &l2b->l2b_kva[l2pte_index(va)];
2778 	pte = *ptep;
2779 	if (pte == 0)
2780 		goto out;
2781 
2782 	/*
2783 	 * Catch a userland access to the vector page mapped at 0x0
2784 	 */
2785 	if (user && (pte & L2_S_PROT_U) == 0)
2786 		goto out;
2787 
2788 	pa = l2pte_pa(pte);
2789 
2790 	if ((ftype & VM_PROT_WRITE) && (pte & L2_S_PROT_W) == 0) {
2791 		/*
2792 		 * This looks like a good candidate for "page modified"
2793 		 * emulation...
2794 		 */
2795 		struct pv_entry *pv;
2796 		struct vm_page *pg;
2797 
2798 		/* Extract the physical address of the page */
2799 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
2800 			goto out;
2801 
2802 		/* Get the current flags for this page. */
2803 		simple_lock(&pg->mdpage.pvh_slock);
2804 
2805 		pv = pmap_find_pv(pg, pm, va);
2806 		if (pv == NULL) {
2807 	    		simple_unlock(&pg->mdpage.pvh_slock);
2808 			goto out;
2809 		}
2810 
2811 		/*
2812 		 * Do the flags say this page is writable? If not then it
2813 		 * is a genuine write fault. If yes then the write fault is
2814 		 * our fault as we did not reflect the write access in the
2815 		 * PTE. Now we know a write has occurred we can correct this
2816 		 * and also set the modified bit
2817 		 */
2818 		if ((pv->pv_flags & PVF_WRITE) == 0) {
2819 		    	simple_unlock(&pg->mdpage.pvh_slock);
2820 			goto out;
2821 		}
2822 
2823 		NPDEBUG(PDB_FOLLOW,
2824 		    printf("pmap_fault_fixup: mod emul. pm %p, va 0x%08lx, pa 0x%08lx\n",
2825 		    pm, va, pg->phys_addr));
2826 
2827 		pg->mdpage.pvh_attrs |= PVF_REF | PVF_MOD;
2828 		pv->pv_flags |= PVF_REF | PVF_MOD;
2829 		simple_unlock(&pg->mdpage.pvh_slock);
2830 
2831 		/*
2832 		 * Re-enable write permissions for the page.  No need to call
2833 		 * pmap_vac_me_harder(), since this is just a
2834 		 * modified-emulation fault, and the PVF_WRITE bit isn't
2835 		 * changing. We've already set the cacheable bits based on
2836 		 * the assumption that we can write to this page.
2837 		 */
2838 		*ptep = (pte & ~L2_TYPE_MASK) | L2_S_PROTO | L2_S_PROT_W;
2839 		PTE_SYNC(ptep);
2840 		rv = 1;
2841 	} else
2842 	if ((pte & L2_TYPE_MASK) == L2_TYPE_INV) {
2843 		/*
2844 		 * This looks like a good candidate for "page referenced"
2845 		 * emulation.
2846 		 */
2847 		struct pv_entry *pv;
2848 		struct vm_page *pg;
2849 
2850 		/* Extract the physical address of the page */
2851 		if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL)
2852 			goto out;
2853 
2854 		/* Get the current flags for this page. */
2855 		simple_lock(&pg->mdpage.pvh_slock);
2856 
2857 		pv = pmap_find_pv(pg, pm, va);
2858 		if (pv == NULL) {
2859 	    		simple_unlock(&pg->mdpage.pvh_slock);
2860 			goto out;
2861 		}
2862 
2863 		pg->mdpage.pvh_attrs |= PVF_REF;
2864 		pv->pv_flags |= PVF_REF;
2865 		simple_unlock(&pg->mdpage.pvh_slock);
2866 
2867 		NPDEBUG(PDB_FOLLOW,
2868 		    printf("pmap_fault_fixup: ref emul. pm %p, va 0x%08lx, pa 0x%08lx\n",
2869 		    pm, va, pg->phys_addr));
2870 
2871 		*ptep = (pte & ~L2_TYPE_MASK) | L2_S_PROTO;
2872 		PTE_SYNC(ptep);
2873 		rv = 1;
2874 	}
2875 
2876 	/*
2877 	 * We know there is a valid mapping here, so simply
2878 	 * fix up the L1 if necessary.
2879 	 */
2880 	pl1pd = &pm->pm_l1->l1_kva[l1idx];
2881 	l1pd = l2b->l2b_phys | L1_C_DOM(pm->pm_domain) | L1_C_PROTO;
2882 	if (*pl1pd != l1pd) {
2883 		*pl1pd = l1pd;
2884 		PTE_SYNC(pl1pd);
2885 		rv = 1;
2886 	}
2887 
2888 #ifdef CPU_SA110
2889 	/*
2890 	 * There are bugs in the rev K SA110.  This is a check for one
2891 	 * of them.
2892 	 */
2893 	if (rv == 0 && curcpu()->ci_arm_cputype == CPU_ID_SA110 &&
2894 	    curcpu()->ci_arm_cpurev < 3) {
2895 		/* Always current pmap */
2896 		if (l2pte_valid(pte)) {
2897 			extern int kernel_debug;
2898 			if (kernel_debug & 1) {
2899 				struct proc *p = curlwp->l_proc;
2900 				printf("prefetch_abort: page is already "
2901 				    "mapped - pte=%p *pte=%08x\n", ptep, pte);
2902 				printf("prefetch_abort: pc=%08lx proc=%p "
2903 				    "process=%s\n", va, p, p->p_comm);
2904 				printf("prefetch_abort: far=%08x fs=%x\n",
2905 				    cpu_faultaddress(), cpu_faultstatus());
2906 			}
2907 #ifdef DDB
2908 			if (kernel_debug & 2)
2909 				Debugger();
2910 #endif
2911 			rv = 1;
2912 		}
2913 	}
2914 #endif /* CPU_SA110 */
2915 
2916 #ifdef DEBUG
2917 	/*
2918 	 * If 'rv == 0' at this point, it generally indicates that there is a
2919 	 * stale TLB entry for the faulting address. This happens when two or
2920 	 * more processes are sharing an L1. Since we don't flush the TLB on
2921 	 * a context switch between such processes, we can take domain faults
2922 	 * for mappings which exist at the same VA in both processes. EVEN IF
2923 	 * WE'VE RECENTLY FIXED UP THE CORRESPONDING L1 in pmap_enter(), for
2924 	 * example.
2925 	 *
2926 	 * This is extremely likely to happen if pmap_enter() updated the L1
2927 	 * entry for a recently entered mapping. In this case, the TLB is
2928 	 * flushed for the new mapping, but there may still be TLB entries for
2929 	 * other mappings belonging to other processes in the 1MB range
2930 	 * covered by the L1 entry.
2931 	 *
2932 	 * Since 'rv == 0', we know that the L1 already contains the correct
2933 	 * value, so the fault must be due to a stale TLB entry.
2934 	 *
2935 	 * Since we always need to flush the TLB anyway in the case where we
2936 	 * fixed up the L1, or frobbed the L2 PTE, we effectively deal with
2937 	 * stale TLB entries dynamically.
2938 	 *
2939 	 * However, the above condition can ONLY happen if the current L1 is
2940 	 * being shared. If it happens when the L1 is unshared, it indicates
2941 	 * that other parts of the pmap are not doing their job WRT managing
2942 	 * the TLB.
2943 	 */
2944 	if (rv == 0 && pm->pm_l1->l1_domain_use_count == 1) {
2945 		extern int last_fault_code;
2946 		printf("fixup: pm %p, va 0x%lx, ftype %d - nothing to do!\n",
2947 		    pm, va, ftype);
2948 		printf("fixup: l2 %p, l2b %p, ptep %p, pl1pd %p\n",
2949 		    l2, l2b, ptep, pl1pd);
2950 		printf("fixup: pte 0x%x, l1pd 0x%x, last code 0x%x\n",
2951 		    pte, l1pd, last_fault_code);
2952 #ifdef DDB
2953 		Debugger();
2954 #endif
2955 	}
2956 #endif
2957 
2958 	cpu_tlb_flushID_SE(va);
2959 	cpu_cpwait();
2960 
2961 	rv = 1;
2962 
2963 out:
2964 	pmap_release_pmap_lock(pm);
2965 	PMAP_MAP_TO_HEAD_UNLOCK();
2966 
2967 	return (rv);
2968 }
2969 
2970 /*
2971  * pmap_collect: free resources held by a pmap
2972  *
2973  * => optional function.
2974  * => called when a process is swapped out to free memory.
2975  */
2976 void
2977 pmap_collect(pmap_t pm)
2978 {
2979 	/*
2980 	 * Nothing to do.
2981 	 * We don't even need to free-up the process' L1.
2982 	 */
2983 }
2984 
2985 /*
2986  * Routine:	pmap_procwr
2987  *
2988  * Function:
2989  *	Synchronize caches corresponding to [addr, addr+len) in p.
2990  *
2991  */
2992 void
2993 pmap_procwr(struct proc *p, vaddr_t va, int len)
2994 {
2995 	/* We only need to do anything if it is the current process. */
2996 	if (p == curproc)
2997 		cpu_icache_sync_range(va, len);
2998 }
2999 
3000 /*
3001  * Routine:	pmap_unwire
3002  * Function:	Clear the wired attribute for a map/virtual-address pair.
3003  *
3004  * In/out conditions:
3005  *		The mapping must already exist in the pmap.
3006  */
3007 void
3008 pmap_unwire(pmap_t pm, vaddr_t va)
3009 {
3010 	struct l2_bucket *l2b;
3011 	pt_entry_t *ptep, pte;
3012 	struct vm_page *pg;
3013 	paddr_t pa;
3014 
3015 	NPDEBUG(PDB_WIRING, printf("pmap_unwire: pm %p, va 0x%08lx\n", pm, va));
3016 
3017 	PMAP_MAP_TO_HEAD_LOCK();
3018 	pmap_acquire_pmap_lock(pm);
3019 
3020 	l2b = pmap_get_l2_bucket(pm, va);
3021 	KDASSERT(l2b != NULL);
3022 
3023 	ptep = &l2b->l2b_kva[l2pte_index(va)];
3024 	pte = *ptep;
3025 
3026 	/* Extract the physical address of the page */
3027 	pa = l2pte_pa(pte);
3028 
3029 	if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) {
3030 		/* Update the wired bit in the pv entry for this page. */
3031 		simple_lock(&pg->mdpage.pvh_slock);
3032 		(void) pmap_modify_pv(pg, pm, va, PVF_WIRED, 0);
3033 		simple_unlock(&pg->mdpage.pvh_slock);
3034 	}
3035 
3036 	pmap_release_pmap_lock(pm);
3037 	PMAP_MAP_TO_HEAD_UNLOCK();
3038 }
3039 
3040 void
3041 pmap_activate(struct lwp *l)
3042 {
3043 	pmap_t pm;
3044 	struct pcb *pcb;
3045 	int s;
3046 
3047 	pm = l->l_proc->p_vmspace->vm_map.pmap;
3048 	pcb = &l->l_addr->u_pcb;
3049 
3050 	pmap_set_pcb_pagedir(pm, pcb);
3051 
3052 	if (l == curlwp) {
3053 		u_int cur_dacr, cur_ttb;
3054 
3055 		__asm __volatile("mrc p15, 0, %0, c2, c0, 0" : "=r"(cur_ttb));
3056 		__asm __volatile("mrc p15, 0, %0, c3, c0, 0" : "=r"(cur_dacr));
3057 
3058 		cur_ttb &= ~(L1_TABLE_SIZE - 1);
3059 
3060 		if (cur_ttb == (u_int)pcb->pcb_pagedir &&
3061 		    cur_dacr == pcb->pcb_dacr) {
3062 			/*
3063 			 * No need to switch address spaces.
3064 			 */
3065 			return;
3066 		}
3067 
3068 		s = splhigh();
3069 		pmap_acquire_pmap_lock(pm);
3070 		disable_interrupts(I32_bit | F32_bit);
3071 
3072 		/*
3073 		 * We MUST, I repeat, MUST fix up the L1 entry corresponding
3074 		 * to 'vector_page' in the incoming L1 table before switching
3075 		 * to it otherwise subsequent interrupts/exceptions (including
3076 		 * domain faults!) will jump into hyperspace.
3077 		 */
3078 		if (pcb->pcb_pl1vec) {
3079 			*pcb->pcb_pl1vec = pcb->pcb_l1vec;
3080 			/*
3081 			 * Don't need to PTE_SYNC() at this point since
3082 			 * cpu_setttb() is about to flush both the cache
3083 			 * and the TLB.
3084 			 */
3085 		}
3086 
3087 		cpu_domains(pcb->pcb_dacr);
3088 		cpu_setttb(pcb->pcb_pagedir);
3089 
3090 		enable_interrupts(I32_bit | F32_bit);
3091 
3092 		/*
3093 		 * Flag any previous userland pmap as being NOT
3094 		 * resident in the cache/tlb.
3095 		 */
3096 		if (pmap_cache_state && pmap_cache_state != &pm->pm_cstate)
3097 			pmap_cache_state->cs_all = 0;
3098 
3099 		/*
3100 		 * The new pmap, however, IS resident.
3101 		 */
3102 		pmap_cache_state = &pm->pm_cstate;
3103 		pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
3104 		pmap_release_pmap_lock(pm);
3105 		splx(s);
3106 	}
3107 }
3108 
3109 void
3110 pmap_deactivate(struct lwp *l)
3111 {
3112 }
3113 
3114 void
3115 pmap_update(pmap_t pm)
3116 {
3117 
3118 	if (pm->pm_remove_all) {
3119 		/*
3120 		 * Finish up the pmap_remove_all() optimisation by flushing
3121 		 * the TLB.
3122 		 */
3123 		pmap_tlb_flushID(pm);
3124 		pm->pm_remove_all = FALSE;
3125 	}
3126 
3127 	if (pmap_is_current(pm)) {
3128 		/*
3129 		 * If we're dealing with a current userland pmap, move its L1
3130 		 * to the end of the LRU.
3131 		 */
3132 		if (pm != pmap_kernel())
3133 			pmap_use_l1(pm);
3134 
3135 		/*
3136 		 * We can assume we're done with frobbing the cache/tlb for
3137 		 * now. Make sure any future pmap ops don't skip cache/tlb
3138 		 * flushes.
3139 		 */
3140 		pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
3141 	}
3142 
3143 	/*
3144 	 * make sure TLB/cache operations have completed.
3145 	 */
3146 	cpu_cpwait();
3147 }
3148 
3149 void
3150 pmap_remove_all(pmap_t pm)
3151 {
3152 
3153 	/*
3154 	 * The vmspace described by this pmap is about to be torn down.
3155 	 * Until pmap_update() is called, UVM will only make calls
3156 	 * to pmap_remove(). We can make life much simpler by flushing
3157 	 * the cache now, and deferring TLB invalidation to pmap_update().
3158 	 */
3159 	pmap_idcache_wbinv_all(pm);
3160 	pm->pm_remove_all = TRUE;
3161 }
3162 
3163 /*
3164  * Retire the given physical map from service.
3165  * Should only be called if the map contains no valid mappings.
3166  */
3167 void
3168 pmap_destroy(pmap_t pm)
3169 {
3170 	u_int count;
3171 
3172 	if (pm == NULL)
3173 		return;
3174 
3175 	if (pm->pm_remove_all) {
3176 		pmap_tlb_flushID(pm);
3177 		pm->pm_remove_all = FALSE;
3178 	}
3179 
3180 	/*
3181 	 * Drop reference count
3182 	 */
3183 	simple_lock(&pm->pm_lock);
3184 	count = --pm->pm_obj.uo_refs;
3185 	simple_unlock(&pm->pm_lock);
3186 	if (count > 0) {
3187 		if (pmap_is_current(pm)) {
3188 			if (pm != pmap_kernel())
3189 				pmap_use_l1(pm);
3190 			pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
3191 		}
3192 		return;
3193 	}
3194 
3195 	/*
3196 	 * reference count is zero, free pmap resources and then free pmap.
3197 	 */
3198 
3199 	if (vector_page < KERNEL_BASE) {
3200 		/* Remove the vector page mapping */
3201 		pmap_remove(pm, vector_page, vector_page + PAGE_SIZE);
3202 		pmap_update(pm);
3203 	}
3204 
3205 	LIST_REMOVE(pm, pm_list);
3206 
3207 	pmap_free_l1(pm);
3208 
3209 	/* return the pmap to the pool */
3210 	pool_cache_put(&pmap_pmap_cache, pm);
3211 }
3212 
3213 
3214 /*
3215  * void pmap_reference(pmap_t pm)
3216  *
3217  * Add a reference to the specified pmap.
3218  */
3219 void
3220 pmap_reference(pmap_t pm)
3221 {
3222 
3223 	if (pm == NULL)
3224 		return;
3225 
3226 	pmap_use_l1(pm);
3227 
3228 	simple_lock(&pm->pm_lock);
3229 	pm->pm_obj.uo_refs++;
3230 	simple_unlock(&pm->pm_lock);
3231 }
3232 
3233 /*
3234  * pmap_zero_page()
3235  *
3236  * Zero a given physical page by mapping it at a page hook point.
3237  * In doing the zero page op, the page we zero is mapped cachable, as with
3238  * StrongARM accesses to non-cached pages are non-burst making writing
3239  * _any_ bulk data very slow.
3240  */
3241 #if (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0
3242 void
3243 pmap_zero_page_generic(paddr_t phys)
3244 {
3245 #ifdef DEBUG
3246 	struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
3247 
3248 	if (pg->mdpage.pvh_list != NULL)
3249 		panic("pmap_zero_page: page has mappings");
3250 #endif
3251 
3252 	KDASSERT((phys & PGOFSET) == 0);
3253 
3254 	/*
3255 	 * Hook in the page, zero it, and purge the cache for that
3256 	 * zeroed page. Invalidate the TLB as needed.
3257 	 */
3258 	*cdst_pte = L2_S_PROTO | phys |
3259 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
3260 	PTE_SYNC(cdst_pte);
3261 	cpu_tlb_flushD_SE(cdstp);
3262 	cpu_cpwait();
3263 	bzero_page(cdstp);
3264 	cpu_dcache_wbinv_range(cdstp, PAGE_SIZE);
3265 }
3266 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 */
3267 
3268 #if ARM_MMU_XSCALE == 1
3269 void
3270 pmap_zero_page_xscale(paddr_t phys)
3271 {
3272 #ifdef DEBUG
3273 	struct vm_page *pg = PHYS_TO_VM_PAGE(phys);
3274 
3275 	if (pg->mdpage.pvh_list != NULL)
3276 		panic("pmap_zero_page: page has mappings");
3277 #endif
3278 
3279 	KDASSERT((phys & PGOFSET) == 0);
3280 
3281 	/*
3282 	 * Hook in the page, zero it, and purge the cache for that
3283 	 * zeroed page. Invalidate the TLB as needed.
3284 	 */
3285 	*cdst_pte = L2_S_PROTO | phys |
3286 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
3287 	    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);	/* mini-data */
3288 	PTE_SYNC(cdst_pte);
3289 	cpu_tlb_flushD_SE(cdstp);
3290 	cpu_cpwait();
3291 	bzero_page(cdstp);
3292 	xscale_cache_clean_minidata();
3293 }
3294 #endif /* ARM_MMU_XSCALE == 1 */
3295 
3296 /* pmap_pageidlezero()
3297  *
3298  * The same as above, except that we assume that the page is not
3299  * mapped.  This means we never have to flush the cache first.  Called
3300  * from the idle loop.
3301  */
3302 boolean_t
3303 pmap_pageidlezero(paddr_t phys)
3304 {
3305 	unsigned int i;
3306 	int *ptr;
3307 	boolean_t rv = TRUE;
3308 #ifdef DEBUG
3309 	struct vm_page *pg;
3310 
3311 	pg = PHYS_TO_VM_PAGE(phys);
3312 	if (pg->mdpage.pvh_list != NULL)
3313 		panic("pmap_pageidlezero: 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 
3328 	for (i = 0, ptr = (int *)cdstp;
3329 			i < (PAGE_SIZE / sizeof(int)); i++) {
3330 		if (sched_whichqs != 0) {
3331 			/*
3332 			 * A process has become ready.  Abort now,
3333 			 * so we don't keep it waiting while we
3334 			 * do slow memory access to finish this
3335 			 * page.
3336 			 */
3337 			rv = FALSE;
3338 			break;
3339 		}
3340 		*ptr++ = 0;
3341 	}
3342 
3343 	if (rv)
3344 		/*
3345 		 * if we aborted we'll rezero this page again later so don't
3346 		 * purge it unless we finished it
3347 		 */
3348 		cpu_dcache_wbinv_range(cdstp, PAGE_SIZE);
3349 
3350 	return (rv);
3351 }
3352 
3353 /*
3354  * pmap_copy_page()
3355  *
3356  * Copy one physical page into another, by mapping the pages into
3357  * hook points. The same comment regarding cachability as in
3358  * pmap_zero_page also applies here.
3359  */
3360 #if (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0
3361 void
3362 pmap_copy_page_generic(paddr_t src, paddr_t dst)
3363 {
3364 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
3365 #ifdef DEBUG
3366 	struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst);
3367 
3368 	if (dst_pg->mdpage.pvh_list != NULL)
3369 		panic("pmap_copy_page: dst page has mappings");
3370 #endif
3371 
3372 	KDASSERT((src & PGOFSET) == 0);
3373 	KDASSERT((dst & PGOFSET) == 0);
3374 
3375 	/*
3376 	 * Clean the source page.  Hold the source page's lock for
3377 	 * the duration of the copy so that no other mappings can
3378 	 * be created while we have a potentially aliased mapping.
3379 	 */
3380 	simple_lock(&src_pg->mdpage.pvh_slock);
3381 	(void) pmap_clean_page(src_pg->mdpage.pvh_list, TRUE);
3382 
3383 	/*
3384 	 * Map the pages into the page hook points, copy them, and purge
3385 	 * the cache for the appropriate page. Invalidate the TLB
3386 	 * as required.
3387 	 */
3388 	*csrc_pte = L2_S_PROTO | src |
3389 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ) | pte_l2_s_cache_mode;
3390 	PTE_SYNC(csrc_pte);
3391 	*cdst_pte = L2_S_PROTO | dst |
3392 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode;
3393 	PTE_SYNC(cdst_pte);
3394 	cpu_tlb_flushD_SE(csrcp);
3395 	cpu_tlb_flushD_SE(cdstp);
3396 	cpu_cpwait();
3397 	bcopy_page(csrcp, cdstp);
3398 	cpu_dcache_inv_range(csrcp, PAGE_SIZE);
3399 	simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
3400 	cpu_dcache_wbinv_range(cdstp, PAGE_SIZE);
3401 }
3402 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 */
3403 
3404 #if ARM_MMU_XSCALE == 1
3405 void
3406 pmap_copy_page_xscale(paddr_t src, paddr_t dst)
3407 {
3408 	struct vm_page *src_pg = PHYS_TO_VM_PAGE(src);
3409 #ifdef DEBUG
3410 	struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst);
3411 
3412 	if (dst_pg->mdpage.pvh_list != NULL)
3413 		panic("pmap_copy_page: dst page has mappings");
3414 #endif
3415 
3416 	KDASSERT((src & PGOFSET) == 0);
3417 	KDASSERT((dst & PGOFSET) == 0);
3418 
3419 	/*
3420 	 * Clean the source page.  Hold the source page's lock for
3421 	 * the duration of the copy so that no other mappings can
3422 	 * be created while we have a potentially aliased mapping.
3423 	 */
3424 	simple_lock(&src_pg->mdpage.pvh_slock);
3425 	(void) pmap_clean_page(src_pg->mdpage.pvh_list, TRUE);
3426 
3427 	/*
3428 	 * Map the pages into the page hook points, copy them, and purge
3429 	 * the cache for the appropriate page. Invalidate the TLB
3430 	 * as required.
3431 	 */
3432 	*csrc_pte = L2_S_PROTO | src |
3433 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
3434 	    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);	/* mini-data */
3435 	PTE_SYNC(csrc_pte);
3436 	*cdst_pte = L2_S_PROTO | dst |
3437 	    L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) |
3438 	    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);	/* mini-data */
3439 	PTE_SYNC(cdst_pte);
3440 	cpu_tlb_flushD_SE(csrcp);
3441 	cpu_tlb_flushD_SE(cdstp);
3442 	cpu_cpwait();
3443 	bcopy_page(csrcp, cdstp);
3444 	simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */
3445 	xscale_cache_clean_minidata();
3446 }
3447 #endif /* ARM_MMU_XSCALE == 1 */
3448 
3449 /*
3450  * void pmap_virtual_space(vaddr_t *start, vaddr_t *end)
3451  *
3452  * Return the start and end addresses of the kernel's virtual space.
3453  * These values are setup in pmap_bootstrap and are updated as pages
3454  * are allocated.
3455  */
3456 void
3457 pmap_virtual_space(vaddr_t *start, vaddr_t *end)
3458 {
3459 	*start = virtual_avail;
3460 	*end = virtual_end;
3461 }
3462 
3463 /*
3464  * Helper function for pmap_grow_l2_bucket()
3465  */
3466 static __inline int
3467 pmap_grow_map(vaddr_t va, pt_entry_t cache_mode, paddr_t *pap)
3468 {
3469 	struct l2_bucket *l2b;
3470 	pt_entry_t *ptep;
3471 	paddr_t pa;
3472 
3473 	if (uvm.page_init_done == FALSE) {
3474 		if (uvm_page_physget(&pa) == FALSE)
3475 			return (1);
3476 	} else {
3477 		struct vm_page *pg;
3478 		pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE);
3479 		if (pg == NULL)
3480 			return (1);
3481 		pa = VM_PAGE_TO_PHYS(pg);
3482 	}
3483 
3484 	if (pap)
3485 		*pap = pa;
3486 
3487 	l2b = pmap_get_l2_bucket(pmap_kernel(), va);
3488 	KDASSERT(l2b != NULL);
3489 
3490 	ptep = &l2b->l2b_kva[l2pte_index(va)];
3491 	*ptep = L2_S_PROTO | pa | cache_mode |
3492 	    L2_S_PROT(PTE_KERNEL, VM_PROT_READ | VM_PROT_WRITE);
3493 	PTE_SYNC(ptep);
3494 	memset((void *)va, 0, PAGE_SIZE);
3495 	return (0);
3496 }
3497 
3498 /*
3499  * This is the same as pmap_alloc_l2_bucket(), except that it is only
3500  * used by pmap_growkernel().
3501  */
3502 static __inline struct l2_bucket *
3503 pmap_grow_l2_bucket(pmap_t pm, vaddr_t va)
3504 {
3505 	struct l2_dtable *l2;
3506 	struct l2_bucket *l2b;
3507 	u_short l1idx;
3508 	vaddr_t nva;
3509 
3510 	l1idx = L1_IDX(va);
3511 
3512 	if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
3513 		/*
3514 		 * No mapping at this address, as there is
3515 		 * no entry in the L1 table.
3516 		 * Need to allocate a new l2_dtable.
3517 		 */
3518 		nva = pmap_kernel_l2dtable_kva;
3519 		if ((nva & PGOFSET) == 0) {
3520 			/*
3521 			 * Need to allocate a backing page
3522 			 */
3523 			if (pmap_grow_map(nva, pte_l2_s_cache_mode, NULL))
3524 				return (NULL);
3525 		}
3526 
3527 		l2 = (struct l2_dtable *)nva;
3528 		nva += sizeof(struct l2_dtable);
3529 
3530 		if ((nva & PGOFSET) < (pmap_kernel_l2dtable_kva & PGOFSET)) {
3531 			/*
3532 			 * The new l2_dtable straddles a page boundary.
3533 			 * Map in another page to cover it.
3534 			 */
3535 			if (pmap_grow_map(nva, pte_l2_s_cache_mode, NULL))
3536 				return (NULL);
3537 		}
3538 
3539 		pmap_kernel_l2dtable_kva = nva;
3540 
3541 		/*
3542 		 * Link it into the parent pmap
3543 		 */
3544 		pm->pm_l2[L2_IDX(l1idx)] = l2;
3545 	}
3546 
3547 	l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
3548 
3549 	/*
3550 	 * Fetch pointer to the L2 page table associated with the address.
3551 	 */
3552 	if (l2b->l2b_kva == NULL) {
3553 		pt_entry_t *ptep;
3554 
3555 		/*
3556 		 * No L2 page table has been allocated. Chances are, this
3557 		 * is because we just allocated the l2_dtable, above.
3558 		 */
3559 		nva = pmap_kernel_l2ptp_kva;
3560 		ptep = (pt_entry_t *)nva;
3561 		if ((nva & PGOFSET) == 0) {
3562 			/*
3563 			 * Need to allocate a backing page
3564 			 */
3565 			if (pmap_grow_map(nva, pte_l2_s_cache_mode_pt,
3566 			    &pmap_kernel_l2ptp_phys))
3567 				return (NULL);
3568 			PTE_SYNC_RANGE(ptep, PAGE_SIZE / sizeof(pt_entry_t));
3569 		}
3570 
3571 		l2->l2_occupancy++;
3572 		l2b->l2b_kva = ptep;
3573 		l2b->l2b_l1idx = l1idx;
3574 		l2b->l2b_phys = pmap_kernel_l2ptp_phys;
3575 
3576 		pmap_kernel_l2ptp_kva += L2_TABLE_SIZE_REAL;
3577 		pmap_kernel_l2ptp_phys += L2_TABLE_SIZE_REAL;
3578 	}
3579 
3580 	return (l2b);
3581 }
3582 
3583 vaddr_t
3584 pmap_growkernel(vaddr_t maxkvaddr)
3585 {
3586 	pmap_t kpm = pmap_kernel();
3587 	struct l1_ttable *l1;
3588 	struct l2_bucket *l2b;
3589 	pd_entry_t *pl1pd;
3590 	int s;
3591 
3592 	if (maxkvaddr <= pmap_curmaxkvaddr)
3593 		goto out;		/* we are OK */
3594 
3595 	NPDEBUG(PDB_GROWKERN,
3596 	    printf("pmap_growkernel: growing kernel from 0x%lx to 0x%lx\n",
3597 	    pmap_curmaxkvaddr, maxkvaddr));
3598 
3599 	KDASSERT(maxkvaddr <= virtual_end);
3600 
3601 	/*
3602 	 * whoops!   we need to add kernel PTPs
3603 	 */
3604 
3605 	s = splhigh();	/* to be safe */
3606 	simple_lock(&kpm->pm_lock);
3607 
3608 	/* Map 1MB at a time */
3609 	for (; pmap_curmaxkvaddr < maxkvaddr; pmap_curmaxkvaddr += L1_S_SIZE) {
3610 
3611 		l2b = pmap_grow_l2_bucket(kpm, pmap_curmaxkvaddr);
3612 		KDASSERT(l2b != NULL);
3613 
3614 		/* Distribute new L1 entry to all other L1s */
3615 		SLIST_FOREACH(l1, &l1_list, l1_link) {
3616 			pl1pd = &l1->l1_kva[L1_IDX(pmap_curmaxkvaddr)];
3617 			*pl1pd = l2b->l2b_phys | L1_C_DOM(PMAP_DOMAIN_KERNEL) |
3618 			    L1_C_PROTO;
3619 			PTE_SYNC(pl1pd);
3620 		}
3621 	}
3622 
3623 	/*
3624 	 * flush out the cache, expensive but growkernel will happen so
3625 	 * rarely
3626 	 */
3627 	cpu_dcache_wbinv_all();
3628 	cpu_tlb_flushD();
3629 	cpu_cpwait();
3630 
3631 	simple_unlock(&kpm->pm_lock);
3632 	splx(s);
3633 
3634 out:
3635 	return (pmap_curmaxkvaddr);
3636 }
3637 
3638 /************************ Utility routines ****************************/
3639 
3640 /*
3641  * vector_page_setprot:
3642  *
3643  *	Manipulate the protection of the vector page.
3644  */
3645 void
3646 vector_page_setprot(int prot)
3647 {
3648 	struct l2_bucket *l2b;
3649 	pt_entry_t *ptep;
3650 
3651 	l2b = pmap_get_l2_bucket(pmap_kernel(), vector_page);
3652 	KDASSERT(l2b != NULL);
3653 
3654 	ptep = &l2b->l2b_kva[l2pte_index(vector_page)];
3655 
3656 	*ptep = (*ptep & ~L1_S_PROT_MASK) | L2_S_PROT(PTE_KERNEL, prot);
3657 	PTE_SYNC(ptep);
3658 	cpu_tlb_flushD_SE(vector_page);
3659 	cpu_cpwait();
3660 }
3661 
3662 /*
3663  * This is used to stuff certain critical values into the PCB where they
3664  * can be accessed quickly from cpu_switch() et al.
3665  */
3666 void
3667 pmap_set_pcb_pagedir(pmap_t pm, struct pcb *pcb)
3668 {
3669 	struct l2_bucket *l2b;
3670 
3671 	KDASSERT(pm->pm_l1);
3672 
3673 	pcb->pcb_pagedir = pm->pm_l1->l1_physaddr;
3674 	pcb->pcb_dacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2)) |
3675 	    (DOMAIN_CLIENT << (pm->pm_domain * 2));
3676 	pcb->pcb_cstate = (void *)&pm->pm_cstate;
3677 
3678 	if (vector_page < KERNEL_BASE) {
3679 		pcb->pcb_pl1vec = &pm->pm_l1->l1_kva[L1_IDX(vector_page)];
3680 		l2b = pmap_get_l2_bucket(pm, vector_page);
3681 		pcb->pcb_l1vec = l2b->l2b_phys | L1_C_PROTO |
3682 		    L1_C_DOM(pm->pm_domain);
3683 	} else
3684 		pcb->pcb_pl1vec = NULL;
3685 }
3686 
3687 /*
3688  * Fetch pointers to the PDE/PTE for the given pmap/VA pair.
3689  * Returns TRUE if the mapping exists, else FALSE.
3690  *
3691  * NOTE: This function is only used by a couple of arm-specific modules.
3692  * It is not safe to take any pmap locks here, since we could be right
3693  * in the middle of debugging the pmap anyway...
3694  *
3695  * It is possible for this routine to return FALSE even though a valid
3696  * mapping does exist. This is because we don't lock, so the metadata
3697  * state may be inconsistent.
3698  *
3699  * NOTE: We can return a NULL *ptp in the case where the L1 pde is
3700  * a "section" mapping.
3701  */
3702 boolean_t
3703 pmap_get_pde_pte(pmap_t pm, vaddr_t va, pd_entry_t **pdp, pt_entry_t **ptp)
3704 {
3705 	struct l2_dtable *l2;
3706 	pd_entry_t *pl1pd, l1pd;
3707 	pt_entry_t *ptep;
3708 	u_short l1idx;
3709 
3710 	if (pm->pm_l1 == NULL)
3711 		return (FALSE);
3712 
3713 	l1idx = L1_IDX(va);
3714 	*pdp = pl1pd = &pm->pm_l1->l1_kva[l1idx];
3715 	l1pd = *pl1pd;
3716 
3717 	if (l1pte_section_p(l1pd)) {
3718 		*ptp = NULL;
3719 		return (TRUE);
3720 	}
3721 
3722 	if (pm->pm_l2 == NULL)
3723 		return (FALSE);
3724 
3725 	l2 = pm->pm_l2[L2_IDX(l1idx)];
3726 
3727 	if (l2 == NULL ||
3728 	    (ptep = l2->l2_bucket[L2_BUCKET(l1idx)].l2b_kva) == NULL) {
3729 		return (FALSE);
3730 	}
3731 
3732 	*ptp = &ptep[l2pte_index(va)];
3733 	return (TRUE);
3734 }
3735 
3736 boolean_t
3737 pmap_get_pde(pmap_t pm, vaddr_t va, pd_entry_t **pdp)
3738 {
3739 	u_short l1idx;
3740 
3741 	if (pm->pm_l1 == NULL)
3742 		return (FALSE);
3743 
3744 	l1idx = L1_IDX(va);
3745 	*pdp = &pm->pm_l1->l1_kva[l1idx];
3746 
3747 	return (TRUE);
3748 }
3749 
3750 /************************ Bootstrapping routines ****************************/
3751 
3752 static void
3753 pmap_init_l1(struct l1_ttable *l1, pd_entry_t *l1pt)
3754 {
3755 	int i;
3756 
3757 	l1->l1_kva = l1pt;
3758 	l1->l1_domain_use_count = 0;
3759 	l1->l1_domain_first = 0;
3760 
3761 	for (i = 0; i < PMAP_DOMAINS; i++)
3762 		l1->l1_domain_free[i] = i + 1;
3763 
3764 	/*
3765 	 * Copy the kernel's L1 entries to each new L1.
3766 	 */
3767 	if (pmap_initialized)
3768 		memcpy(l1pt, pmap_kernel()->pm_l1->l1_kva, L1_TABLE_SIZE);
3769 
3770 	if (pmap_extract(pmap_kernel(), (vaddr_t)l1pt,
3771 	    &l1->l1_physaddr) == FALSE)
3772 		panic("pmap_init_l1: can't get PA of L1 at %p", l1pt);
3773 
3774 	SLIST_INSERT_HEAD(&l1_list, l1, l1_link);
3775 	TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru);
3776 }
3777 
3778 /*
3779  * pmap_bootstrap() is called from the board-specific initarm() routine
3780  * once the kernel L1/L2 descriptors tables have been set up.
3781  *
3782  * This is a somewhat convoluted process since pmap bootstrap is, effectively,
3783  * spread over a number of disparate files/functions.
3784  *
3785  * We are passed the following parameters
3786  *  - kernel_l1pt
3787  *    This is a pointer to the base of the kernel's L1 translation table.
3788  *  - vstart
3789  *    1MB-aligned start of managed kernel virtual memory.
3790  *  - vend
3791  *    1MB-aligned end of managed kernel virtual memory.
3792  *
3793  * We use the first parameter to build the metadata (struct l1_ttable and
3794  * struct l2_dtable) necessary to track kernel mappings.
3795  */
3796 #define	PMAP_STATIC_L2_SIZE 16
3797 void
3798 pmap_bootstrap(pd_entry_t *kernel_l1pt, vaddr_t vstart, vaddr_t vend)
3799 {
3800 	static struct l1_ttable static_l1;
3801 	static struct l2_dtable static_l2[PMAP_STATIC_L2_SIZE];
3802 	struct l1_ttable *l1 = &static_l1;
3803 	struct l2_dtable *l2;
3804 	struct l2_bucket *l2b;
3805 	pmap_t pm = pmap_kernel();
3806 	pd_entry_t pde;
3807 	pt_entry_t *ptep;
3808 	paddr_t pa;
3809 	vaddr_t va;
3810 	vsize_t size;
3811 	int l1idx, l2idx, l2next = 0;
3812 
3813 	/*
3814 	 * Initialise the kernel pmap object
3815 	 */
3816 	pm->pm_l1 = l1;
3817 	pm->pm_domain = PMAP_DOMAIN_KERNEL;
3818 	pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL;
3819 	simple_lock_init(&pm->pm_lock);
3820 	pm->pm_obj.pgops = NULL;
3821 	TAILQ_INIT(&pm->pm_obj.memq);
3822 	pm->pm_obj.uo_npages = 0;
3823 	pm->pm_obj.uo_refs = 1;
3824 
3825 	/*
3826 	 * Scan the L1 translation table created by initarm() and create
3827 	 * the required metadata for all valid mappings found in it.
3828 	 */
3829 	for (l1idx = 0; l1idx < (L1_TABLE_SIZE / sizeof(pd_entry_t)); l1idx++) {
3830 		pde = kernel_l1pt[l1idx];
3831 
3832 		/*
3833 		 * We're only interested in Coarse mappings.
3834 		 * pmap_extract() can deal with section mappings without
3835 		 * recourse to checking L2 metadata.
3836 		 */
3837 		if ((pde & L1_TYPE_MASK) != L1_TYPE_C)
3838 			continue;
3839 
3840 		/*
3841 		 * Lookup the KVA of this L2 descriptor table
3842 		 */
3843 		pa = (paddr_t)(pde & L1_C_ADDR_MASK);
3844 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
3845 		if (ptep == NULL) {
3846 			panic("pmap_bootstrap: No L2 for va 0x%x, pa 0x%lx",
3847 			    (u_int)l1idx << L1_S_SHIFT, pa);
3848 		}
3849 
3850 		/*
3851 		 * Fetch the associated L2 metadata structure.
3852 		 * Allocate a new one if necessary.
3853 		 */
3854 		if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) {
3855 			if (l2next == PMAP_STATIC_L2_SIZE)
3856 				panic("pmap_bootstrap: out of static L2s");
3857 			pm->pm_l2[L2_IDX(l1idx)] = l2 = &static_l2[l2next++];
3858 		}
3859 
3860 		/*
3861 		 * One more L1 slot tracked...
3862 		 */
3863 		l2->l2_occupancy++;
3864 
3865 		/*
3866 		 * Fill in the details of the L2 descriptor in the
3867 		 * appropriate bucket.
3868 		 */
3869 		l2b = &l2->l2_bucket[L2_BUCKET(l1idx)];
3870 		l2b->l2b_kva = ptep;
3871 		l2b->l2b_phys = pa;
3872 		l2b->l2b_l1idx = l1idx;
3873 
3874 		/*
3875 		 * Establish an initial occupancy count for this descriptor
3876 		 */
3877 		for (l2idx = 0;
3878 		    l2idx < (L2_TABLE_SIZE_REAL / sizeof(pt_entry_t));
3879 		    l2idx++) {
3880 			if ((ptep[l2idx] & L2_TYPE_MASK) != L2_TYPE_INV) {
3881 				l2b->l2b_occupancy++;
3882 			}
3883 		}
3884 
3885 		/*
3886 		 * Make sure the descriptor itself has the correct cache mode.
3887 		 * If not, fix it, but bitch about the problem. Port-meisters
3888 		 * should consider this a clue to fix up their initarm()
3889 		 * function. :)
3890 		 */
3891 		if (pmap_set_pt_cache_mode(kernel_l1pt, (vaddr_t)ptep)) {
3892 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
3893 			    "L2 pte @ %p\n", ptep);
3894 		}
3895 	}
3896 
3897 	/*
3898 	 * Ensure the primary (kernel) L1 has the correct cache mode for
3899 	 * a page table. Bitch if it is not correctly set.
3900 	 */
3901 	for (va = (vaddr_t)kernel_l1pt;
3902 	    va < ((vaddr_t)kernel_l1pt + L1_TABLE_SIZE); va += PAGE_SIZE) {
3903 		if (pmap_set_pt_cache_mode(kernel_l1pt, va))
3904 			printf("pmap_bootstrap: WARNING! wrong cache mode for "
3905 			    "primary L1 @ 0x%lx\n", va);
3906 	}
3907 
3908 	cpu_dcache_wbinv_all();
3909 	cpu_tlb_flushID();
3910 	cpu_cpwait();
3911 
3912 	/*
3913 	 * now we allocate the "special" VAs which are used for tmp mappings
3914 	 * by the pmap (and other modules).  we allocate the VAs by advancing
3915 	 * virtual_avail (note that there are no pages mapped at these VAs).
3916 	 *
3917 	 * Managed KVM space start from wherever initarm() tells us.
3918 	 */
3919 	virtual_avail = vstart;
3920 	virtual_end = vend;
3921 
3922 	pmap_alloc_specials(&virtual_avail, 1, &csrcp, &csrc_pte);
3923 	pmap_set_pt_cache_mode(kernel_l1pt, (vaddr_t)csrc_pte);
3924 	pmap_alloc_specials(&virtual_avail, 1, &cdstp, &cdst_pte);
3925 	pmap_set_pt_cache_mode(kernel_l1pt, (vaddr_t)cdst_pte);
3926 	pmap_alloc_specials(&virtual_avail, 1, (vaddr_t *)&memhook, NULL);
3927 	pmap_alloc_specials(&virtual_avail, round_page(MSGBUFSIZE) / PAGE_SIZE,
3928 	    (vaddr_t *)&msgbufaddr, NULL);
3929 
3930 	/*
3931 	 * Allocate a range of kernel virtual address space to be used
3932 	 * for L2 descriptor tables and metadata allocation in
3933 	 * pmap_growkernel().
3934 	 */
3935 	size = ((virtual_end - pmap_curmaxkvaddr) + L1_S_OFFSET) / L1_S_SIZE;
3936 	pmap_alloc_specials(&virtual_avail,
3937 	    round_page(size * L2_TABLE_SIZE_REAL) / PAGE_SIZE,
3938 	    &pmap_kernel_l2ptp_kva, NULL);
3939 
3940 	size = (size + (L2_BUCKET_SIZE - 1)) / L2_BUCKET_SIZE;
3941 	pmap_alloc_specials(&virtual_avail,
3942 	    round_page(size * sizeof(struct l2_dtable)) / PAGE_SIZE,
3943 	    &pmap_kernel_l2dtable_kva, NULL);
3944 
3945 	/*
3946 	 * init the static-global locks and global pmap list.
3947 	 */
3948 #if defined(MULTIPROCESSOR) || defined(LOCKDEBUG)
3949 	spinlockinit(&pmap_main_lock, "pmaplk", 0);
3950 #endif
3951 
3952 	/*
3953 	 * We can now initialise the first L1's metadata.
3954 	 */
3955 	SLIST_INIT(&l1_list);
3956 	TAILQ_INIT(&l1_lru_list);
3957 	simple_lock_init(&l1_lru_lock);
3958 	pmap_init_l1(l1, kernel_l1pt);
3959 
3960 	/*
3961 	 * Initialize the pmap pool and cache
3962 	 */
3963 	pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl",
3964 	    &pool_allocator_nointr);
3965 	pool_cache_init(&pmap_pmap_cache, &pmap_pmap_pool,
3966 	    pmap_pmap_ctor, NULL, NULL);
3967 	LIST_INIT(&pmap_pmaps);
3968 	LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list);
3969 
3970 	/*
3971 	 * Initialize the pv pool.
3972 	 */
3973 	pool_init(&pmap_pv_pool, sizeof(struct pv_entry), 0, 0, 0, "pvepl",
3974 	    &pmap_bootstrap_pv_allocator);
3975 
3976 	/*
3977 	 * Initialize the L2 dtable pool and cache.
3978 	 */
3979 	pool_init(&pmap_l2dtable_pool, sizeof(struct l2_dtable), 0, 0, 0,
3980 	    "l2dtblpl", NULL);
3981 	pool_cache_init(&pmap_l2dtable_cache, &pmap_l2dtable_pool,
3982 	    pmap_l2dtable_ctor, NULL, NULL);
3983 
3984 	/*
3985 	 * Initialise the L2 descriptor table pool and cache
3986 	 */
3987 	pool_init(&pmap_l2ptp_pool, L2_TABLE_SIZE_REAL, 0, L2_TABLE_SIZE_REAL,
3988 	    0, "l2ptppl", NULL);
3989 	pool_cache_init(&pmap_l2ptp_cache, &pmap_l2ptp_pool,
3990 	    pmap_l2ptp_ctor, NULL, NULL);
3991 
3992 	cpu_dcache_wbinv_all();
3993 }
3994 
3995 static int
3996 pmap_set_pt_cache_mode(pd_entry_t *kl1, vaddr_t va)
3997 {
3998 	pd_entry_t *pdep, pde;
3999 	pt_entry_t *ptep, pte;
4000 	vaddr_t pa;
4001 	int rv = 0;
4002 
4003 	/*
4004 	 * Make sure the descriptor itself has the correct cache mode
4005 	 */
4006 	pdep = &kl1[L1_IDX(va)];
4007 	pde = *pdep;
4008 
4009 	if (l1pte_section_p(pde)) {
4010 		if ((pde & L1_S_CACHE_MASK) != pte_l1_s_cache_mode_pt) {
4011 			*pdep = (pde & ~L1_S_CACHE_MASK) |
4012 			    pte_l1_s_cache_mode_pt;
4013 			PTE_SYNC(pdep);
4014 			cpu_dcache_wbinv_range((vaddr_t)pdep, sizeof(*pdep));
4015 			rv = 1;
4016 		}
4017 	} else {
4018 		pa = (paddr_t)(pde & L1_C_ADDR_MASK);
4019 		ptep = (pt_entry_t *)kernel_pt_lookup(pa);
4020 		if (ptep == NULL)
4021 			panic("pmap_bootstrap: No L2 for L2 @ va %p\n", ptep);
4022 
4023 		ptep = &ptep[l2pte_index(va)];
4024 		pte = *ptep;
4025 		if ((pte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) {
4026 			*ptep = (pte & ~L2_S_CACHE_MASK) |
4027 			    pte_l2_s_cache_mode_pt;
4028 			PTE_SYNC(ptep);
4029 			cpu_dcache_wbinv_range((vaddr_t)ptep, sizeof(*ptep));
4030 			rv = 1;
4031 		}
4032 	}
4033 
4034 	return (rv);
4035 }
4036 
4037 static void
4038 pmap_alloc_specials(vaddr_t *availp, int pages, vaddr_t *vap, pt_entry_t **ptep)
4039 {
4040 	vaddr_t va = *availp;
4041 	struct l2_bucket *l2b;
4042 
4043 	if (ptep) {
4044 		l2b = pmap_get_l2_bucket(pmap_kernel(), va);
4045 		if (l2b == NULL)
4046 			panic("pmap_alloc_specials: no l2b for 0x%lx", va);
4047 
4048 		if (ptep)
4049 			*ptep = &l2b->l2b_kva[l2pte_index(va)];
4050 	}
4051 
4052 	*vap = va;
4053 	*availp = va + (PAGE_SIZE * pages);
4054 }
4055 
4056 void
4057 pmap_init(void)
4058 {
4059 	extern int physmem;
4060 
4061 	/*
4062 	 * Set the available memory vars - These do not map to real memory
4063 	 * addresses and cannot as the physical memory is fragmented.
4064 	 * They are used by ps for %mem calculations.
4065 	 * One could argue whether this should be the entire memory or just
4066 	 * the memory that is useable in a user process.
4067 	 */
4068 	avail_start = 0;
4069 	avail_end = physmem * PAGE_SIZE;
4070 
4071 	/*
4072 	 * Now we need to free enough pv_entry structures to allow us to get
4073 	 * the kmem_map/kmem_object allocated and inited (done after this
4074 	 * function is finished).  to do this we allocate one bootstrap page out
4075 	 * of kernel_map and use it to provide an initial pool of pv_entry
4076 	 * structures.   we never free this page.
4077 	 */
4078 	pool_setlowat(&pmap_pv_pool,
4079 	    (PAGE_SIZE / sizeof(struct pv_entry)) * 2);
4080 
4081 	pmap_initialized = TRUE;
4082 }
4083 
4084 static vaddr_t last_bootstrap_page = 0;
4085 static void *free_bootstrap_pages = NULL;
4086 
4087 static void *
4088 pmap_bootstrap_pv_page_alloc(struct pool *pp, int flags)
4089 {
4090 	extern void *pool_page_alloc(struct pool *, int);
4091 	vaddr_t new_page;
4092 	void *rv;
4093 
4094 	if (pmap_initialized)
4095 		return (pool_page_alloc(pp, flags));
4096 
4097 	if (free_bootstrap_pages) {
4098 		rv = free_bootstrap_pages;
4099 		free_bootstrap_pages = *((void **)rv);
4100 		return (rv);
4101 	}
4102 
4103 	new_page = uvm_km_kmemalloc(kernel_map, NULL, PAGE_SIZE,
4104 	    (flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT);
4105 
4106 	KASSERT(new_page > last_bootstrap_page);
4107 	last_bootstrap_page = new_page;
4108 	return ((void *)new_page);
4109 }
4110 
4111 static void
4112 pmap_bootstrap_pv_page_free(struct pool *pp, void *v)
4113 {
4114 	extern void pool_page_free(struct pool *, void *);
4115 
4116 	if (pmap_initialized) {
4117 		pool_page_free(pp, v);
4118 		return;
4119 	}
4120 
4121 	if ((vaddr_t)v < last_bootstrap_page) {
4122 		*((void **)v) = free_bootstrap_pages;
4123 		free_bootstrap_pages = v;
4124 		return;
4125 	}
4126 }
4127 
4128 /*
4129  * pmap_postinit()
4130  *
4131  * This routine is called after the vm and kmem subsystems have been
4132  * initialised. This allows the pmap code to perform any initialisation
4133  * that can only be done one the memory allocation is in place.
4134  */
4135 void
4136 pmap_postinit(void)
4137 {
4138 	extern paddr_t physical_start, physical_end;
4139 	struct l2_bucket *l2b;
4140 	struct l1_ttable *l1;
4141 	struct pglist plist;
4142 	struct vm_page *m;
4143 	pd_entry_t *pl1pt;
4144 	pt_entry_t *ptep, pte;
4145 	vaddr_t va, eva;
4146 	u_int loop, needed;
4147 	int error;
4148 
4149 	pool_setlowat(&pmap_l2ptp_pool,
4150 	    (PAGE_SIZE / L2_TABLE_SIZE_REAL) * 4);
4151 	pool_setlowat(&pmap_l2dtable_pool,
4152 	    (PAGE_SIZE / sizeof(struct l2_dtable)) * 2);
4153 
4154 	needed = (maxproc / PMAP_DOMAINS) + ((maxproc % PMAP_DOMAINS) ? 1 : 0);
4155 	needed -= 1;
4156 
4157 	l1 = malloc(sizeof(*l1) * needed, M_VMPMAP, M_WAITOK);
4158 
4159 	for (loop = 0; loop < needed; loop++, l1++) {
4160 		/* Allocate a L1 page table */
4161 		va = uvm_km_valloc(kernel_map, L1_TABLE_SIZE);
4162 		if (va == 0)
4163 			panic("Cannot allocate L1 KVM");
4164 
4165 		error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start,
4166 		    physical_end, L1_TABLE_SIZE, 0, &plist, 1, M_WAITOK);
4167 		if (error)
4168 			panic("Cannot allocate L1 physical pages");
4169 
4170 		m = TAILQ_FIRST(&plist);
4171 		eva = va + L1_TABLE_SIZE;
4172 		pl1pt = (pd_entry_t *)va;
4173 
4174 		while (m && va < eva) {
4175 			paddr_t pa = VM_PAGE_TO_PHYS(m);
4176 
4177 			pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE);
4178 
4179 			/*
4180 			 * Make sure the L1 descriptor table is mapped
4181 			 * with the cache-mode set to write-through.
4182 			 */
4183 			l2b = pmap_get_l2_bucket(pmap_kernel(), va);
4184 			ptep = &l2b->l2b_kva[l2pte_index(va)];
4185 			pte = *ptep;
4186 			pte = (pte & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode_pt;
4187 			*ptep = pte;
4188 			PTE_SYNC(ptep);
4189 			cpu_tlb_flushD_SE(va);
4190 
4191 			va += PAGE_SIZE;
4192 			m = m->pageq.tqe_next;
4193 		}
4194 
4195 #ifdef DIAGNOSTIC
4196 		if (m)
4197 			panic("pmap_alloc_l1pt: pglist not empty");
4198 #endif	/* DIAGNOSTIC */
4199 
4200 		pmap_init_l1(l1, pl1pt);
4201 	}
4202 
4203 #ifdef DEBUG
4204 	printf("pmap_postinit: Allocated %d static L1 descriptor tables\n",
4205 	    needed);
4206 #endif
4207 }
4208 
4209 /*
4210  * Note that the following routines are used by board-specific initialisation
4211  * code to configure the initial kernel page tables.
4212  *
4213  * If ARM32_NEW_VM_LAYOUT is *not* defined, they operate on the assumption that
4214  * L2 page-table pages are 4KB in size and use 4 L1 slots. This mimics the
4215  * behaviour of the old pmap, and provides an easy migration path for
4216  * initial bring-up of the new pmap on existing ports. Fortunately,
4217  * pmap_bootstrap() compensates for this hackery. This is only a stop-gap and
4218  * will be deprecated.
4219  *
4220  * If ARM32_NEW_VM_LAYOUT *is* defined, these functions deal with 1KB L2 page
4221  * tables.
4222  */
4223 
4224 /*
4225  * This list exists for the benefit of pmap_map_chunk().  It keeps track
4226  * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can
4227  * find them as necessary.
4228  *
4229  * Note that the data on this list MUST remain valid after initarm() returns,
4230  * as pmap_bootstrap() uses it to contruct L2 table metadata.
4231  */
4232 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list);
4233 
4234 static vaddr_t
4235 kernel_pt_lookup(paddr_t pa)
4236 {
4237 	pv_addr_t *pv;
4238 
4239 	SLIST_FOREACH(pv, &kernel_pt_list, pv_list) {
4240 #ifndef ARM32_NEW_VM_LAYOUT
4241 		if (pv->pv_pa == (pa & ~PGOFSET))
4242 			return (pv->pv_va | (pa & PGOFSET));
4243 #else
4244 		if (pv->pv_pa == pa)
4245 			return (pv->pv_va);
4246 #endif
4247 	}
4248 	return (0);
4249 }
4250 
4251 /*
4252  * pmap_map_section:
4253  *
4254  *	Create a single section mapping.
4255  */
4256 void
4257 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
4258 {
4259 	pd_entry_t *pde = (pd_entry_t *) l1pt;
4260 	pd_entry_t fl;
4261 
4262 	KASSERT(((va | pa) & L1_S_OFFSET) == 0);
4263 
4264 	switch (cache) {
4265 	case PTE_NOCACHE:
4266 	default:
4267 		fl = 0;
4268 		break;
4269 
4270 	case PTE_CACHE:
4271 		fl = pte_l1_s_cache_mode;
4272 		break;
4273 
4274 	case PTE_PAGETABLE:
4275 		fl = pte_l1_s_cache_mode_pt;
4276 		break;
4277 	}
4278 
4279 	pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
4280 	    L1_S_PROT(PTE_KERNEL, prot) | fl | L1_S_DOM(PMAP_DOMAIN_KERNEL);
4281 	PTE_SYNC(&pde[va >> L1_S_SHIFT]);
4282 }
4283 
4284 /*
4285  * pmap_map_entry:
4286  *
4287  *	Create a single page mapping.
4288  */
4289 void
4290 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache)
4291 {
4292 	pd_entry_t *pde = (pd_entry_t *) l1pt;
4293 	pt_entry_t fl;
4294 	pt_entry_t *pte;
4295 
4296 	KASSERT(((va | pa) & PGOFSET) == 0);
4297 
4298 	switch (cache) {
4299 	case PTE_NOCACHE:
4300 	default:
4301 		fl = 0;
4302 		break;
4303 
4304 	case PTE_CACHE:
4305 		fl = pte_l2_s_cache_mode;
4306 		break;
4307 
4308 	case PTE_PAGETABLE:
4309 		fl = pte_l2_s_cache_mode_pt;
4310 		break;
4311 	}
4312 
4313 	if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
4314 		panic("pmap_map_entry: no L2 table for VA 0x%08lx", va);
4315 
4316 #ifndef ARM32_NEW_VM_LAYOUT
4317 	pte = (pt_entry_t *)
4318 	    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
4319 #else
4320 	pte = (pt_entry_t *) kernel_pt_lookup(pde[L1_IDX(va)] & L1_C_ADDR_MASK);
4321 #endif
4322 	if (pte == NULL)
4323 		panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va);
4324 
4325 #ifndef ARM32_NEW_VM_LAYOUT
4326 	pte[(va >> PGSHIFT) & 0x3ff] =
4327 	    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | fl;
4328 	PTE_SYNC(&pte[(va >> PGSHIFT) & 0x3ff]);
4329 #else
4330 	pte[l2pte_index(va)] =
4331 	    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | fl;
4332 	PTE_SYNC(&pte[l2pte_index(va)]);
4333 #endif
4334 }
4335 
4336 /*
4337  * pmap_link_l2pt:
4338  *
4339  *	Link the L2 page table specified by "l2pv" into the L1
4340  *	page table at the slot for "va".
4341  */
4342 void
4343 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv)
4344 {
4345 	pd_entry_t *pde = (pd_entry_t *) l1pt, proto;
4346 	u_int slot = va >> L1_S_SHIFT;
4347 
4348 #ifndef ARM32_NEW_VM_LAYOUT
4349 	KASSERT((va & ((L1_S_SIZE * 4) - 1)) == 0);
4350 	KASSERT((l2pv->pv_pa & PGOFSET) == 0);
4351 #endif
4352 
4353 	proto = L1_S_DOM(PMAP_DOMAIN_KERNEL) | L1_C_PROTO;
4354 
4355 	pde[slot + 0] = proto | (l2pv->pv_pa + 0x000);
4356 #ifdef ARM32_NEW_VM_LAYOUT
4357 	PTE_SYNC(&pde[slot]);
4358 #else
4359 	pde[slot + 1] = proto | (l2pv->pv_pa + 0x400);
4360 	pde[slot + 2] = proto | (l2pv->pv_pa + 0x800);
4361 	pde[slot + 3] = proto | (l2pv->pv_pa + 0xc00);
4362 	PTE_SYNC_RANGE(&pde[slot + 0], 4);
4363 #endif
4364 
4365 	SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list);
4366 }
4367 
4368 /*
4369  * pmap_map_chunk:
4370  *
4371  *	Map a chunk of memory using the most efficient mappings
4372  *	possible (section, large page, small page) into the
4373  *	provided L1 and L2 tables at the specified virtual address.
4374  */
4375 vsize_t
4376 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size,
4377     int prot, int cache)
4378 {
4379 	pd_entry_t *pde = (pd_entry_t *) l1pt;
4380 	pt_entry_t *pte, f1, f2s, f2l;
4381 	vsize_t resid;
4382 	int i;
4383 
4384 	resid = (size + (PAGE_SIZE - 1)) & ~(PAGE_SIZE - 1);
4385 
4386 	if (l1pt == 0)
4387 		panic("pmap_map_chunk: no L1 table provided");
4388 
4389 #ifdef VERBOSE_INIT_ARM
4390 	printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx "
4391 	    "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache);
4392 #endif
4393 
4394 	switch (cache) {
4395 	case PTE_NOCACHE:
4396 	default:
4397 		f1 = 0;
4398 		f2l = 0;
4399 		f2s = 0;
4400 		break;
4401 
4402 	case PTE_CACHE:
4403 		f1 = pte_l1_s_cache_mode;
4404 		f2l = pte_l2_l_cache_mode;
4405 		f2s = pte_l2_s_cache_mode;
4406 		break;
4407 
4408 	case PTE_PAGETABLE:
4409 		f1 = pte_l1_s_cache_mode_pt;
4410 		f2l = pte_l2_l_cache_mode_pt;
4411 		f2s = pte_l2_s_cache_mode_pt;
4412 		break;
4413 	}
4414 
4415 	size = resid;
4416 
4417 	while (resid > 0) {
4418 		/* See if we can use a section mapping. */
4419 		if (L1_S_MAPPABLE_P(va, pa, resid)) {
4420 #ifdef VERBOSE_INIT_ARM
4421 			printf("S");
4422 #endif
4423 			pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa |
4424 			    L1_S_PROT(PTE_KERNEL, prot) | f1 |
4425 			    L1_S_DOM(PMAP_DOMAIN_KERNEL);
4426 			PTE_SYNC(&pde[va >> L1_S_SHIFT]);
4427 			va += L1_S_SIZE;
4428 			pa += L1_S_SIZE;
4429 			resid -= L1_S_SIZE;
4430 			continue;
4431 		}
4432 
4433 		/*
4434 		 * Ok, we're going to use an L2 table.  Make sure
4435 		 * one is actually in the corresponding L1 slot
4436 		 * for the current VA.
4437 		 */
4438 		if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C)
4439 			panic("pmap_map_chunk: no L2 table for VA 0x%08lx", va);
4440 
4441 #ifndef ARM32_NEW_VM_LAYOUT
4442 		pte = (pt_entry_t *)
4443 		    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
4444 #else
4445 		pte = (pt_entry_t *) kernel_pt_lookup(
4446 		    pde[L1_IDX(va)] & L1_C_ADDR_MASK);
4447 #endif
4448 		if (pte == NULL)
4449 			panic("pmap_map_chunk: can't find L2 table for VA"
4450 			    "0x%08lx", va);
4451 
4452 		/* See if we can use a L2 large page mapping. */
4453 		if (L2_L_MAPPABLE_P(va, pa, resid)) {
4454 #ifdef VERBOSE_INIT_ARM
4455 			printf("L");
4456 #endif
4457 			for (i = 0; i < 16; i++) {
4458 #ifndef ARM32_NEW_VM_LAYOUT
4459 				pte[((va >> PGSHIFT) & 0x3f0) + i] =
4460 				    L2_L_PROTO | pa |
4461 				    L2_L_PROT(PTE_KERNEL, prot) | f2l;
4462 				PTE_SYNC(&pte[((va >> PGSHIFT) & 0x3f0) + i]);
4463 #else
4464 				pte[l2pte_index(va) + i] =
4465 				    L2_L_PROTO | pa |
4466 				    L2_L_PROT(PTE_KERNEL, prot) | f2l;
4467 				PTE_SYNC(&pte[l2pte_index(va) + i]);
4468 #endif
4469 			}
4470 			va += L2_L_SIZE;
4471 			pa += L2_L_SIZE;
4472 			resid -= L2_L_SIZE;
4473 			continue;
4474 		}
4475 
4476 		/* Use a small page mapping. */
4477 #ifdef VERBOSE_INIT_ARM
4478 		printf("P");
4479 #endif
4480 #ifndef ARM32_NEW_VM_LAYOUT
4481 		pte[(va >> PGSHIFT) & 0x3ff] =
4482 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | f2s;
4483 		PTE_SYNC(&pte[(va >> PGSHIFT) & 0x3ff]);
4484 #else
4485 		pte[l2pte_index(va)] =
4486 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | f2s;
4487 		PTE_SYNC(&pte[l2pte_index(va)]);
4488 #endif
4489 		va += PAGE_SIZE;
4490 		pa += PAGE_SIZE;
4491 		resid -= PAGE_SIZE;
4492 	}
4493 #ifdef VERBOSE_INIT_ARM
4494 	printf("\n");
4495 #endif
4496 	return (size);
4497 }
4498 
4499 /********************** Static device map routines ***************************/
4500 
4501 static const struct pmap_devmap *pmap_devmap_table;
4502 
4503 /*
4504  * Register the devmap table.  This is provided in case early console
4505  * initialization needs to register mappings created by bootstrap code
4506  * before pmap_devmap_bootstrap() is called.
4507  */
4508 void
4509 pmap_devmap_register(const struct pmap_devmap *table)
4510 {
4511 
4512 	pmap_devmap_table = table;
4513 }
4514 
4515 /*
4516  * Map all of the static regions in the devmap table, and remember
4517  * the devmap table so other parts of the kernel can look up entries
4518  * later.
4519  */
4520 void
4521 pmap_devmap_bootstrap(vaddr_t l1pt, const struct pmap_devmap *table)
4522 {
4523 	int i;
4524 
4525 	pmap_devmap_table = table;
4526 
4527 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
4528 #ifdef VERBOSE_INIT_ARM
4529 		printf("devmap: %08lx -> %08lx @ %08lx\n",
4530 		    pmap_devmap_table[i].pd_pa,
4531 		    pmap_devmap_table[i].pd_pa +
4532 			pmap_devmap_table[i].pd_size - 1,
4533 		    pmap_devmap_table[i].pd_va);
4534 #endif
4535 		pmap_map_chunk(l1pt, pmap_devmap_table[i].pd_va,
4536 		    pmap_devmap_table[i].pd_pa,
4537 		    pmap_devmap_table[i].pd_size,
4538 		    pmap_devmap_table[i].pd_prot,
4539 		    pmap_devmap_table[i].pd_cache);
4540 	}
4541 }
4542 
4543 const struct pmap_devmap *
4544 pmap_devmap_find_pa(paddr_t pa, psize_t size)
4545 {
4546 	int i;
4547 
4548 	if (pmap_devmap_table == NULL)
4549 		return (NULL);
4550 
4551 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
4552 		if (pa >= pmap_devmap_table[i].pd_pa &&
4553 		    pa + size <= pmap_devmap_table[i].pd_pa +
4554 				 pmap_devmap_table[i].pd_size)
4555 			return (&pmap_devmap_table[i]);
4556 	}
4557 
4558 	return (NULL);
4559 }
4560 
4561 const struct pmap_devmap *
4562 pmap_devmap_find_va(vaddr_t va, vsize_t size)
4563 {
4564 	int i;
4565 
4566 	if (pmap_devmap_table == NULL)
4567 		return (NULL);
4568 
4569 	for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) {
4570 		if (va >= pmap_devmap_table[i].pd_va &&
4571 		    va + size <= pmap_devmap_table[i].pd_va +
4572 				 pmap_devmap_table[i].pd_size)
4573 			return (&pmap_devmap_table[i]);
4574 	}
4575 
4576 	return (NULL);
4577 }
4578 
4579 /********************** PTE initialization routines **************************/
4580 
4581 /*
4582  * These routines are called when the CPU type is identified to set up
4583  * the PTE prototypes, cache modes, etc.
4584  *
4585  * The variables are always here, just in case LKMs need to reference
4586  * them (though, they shouldn't).
4587  */
4588 
4589 pt_entry_t	pte_l1_s_cache_mode;
4590 pt_entry_t	pte_l1_s_cache_mode_pt;
4591 pt_entry_t	pte_l1_s_cache_mask;
4592 
4593 pt_entry_t	pte_l2_l_cache_mode;
4594 pt_entry_t	pte_l2_l_cache_mode_pt;
4595 pt_entry_t	pte_l2_l_cache_mask;
4596 
4597 pt_entry_t	pte_l2_s_cache_mode;
4598 pt_entry_t	pte_l2_s_cache_mode_pt;
4599 pt_entry_t	pte_l2_s_cache_mask;
4600 
4601 pt_entry_t	pte_l2_s_prot_u;
4602 pt_entry_t	pte_l2_s_prot_w;
4603 pt_entry_t	pte_l2_s_prot_mask;
4604 
4605 pt_entry_t	pte_l1_s_proto;
4606 pt_entry_t	pte_l1_c_proto;
4607 pt_entry_t	pte_l2_s_proto;
4608 
4609 void		(*pmap_copy_page_func)(paddr_t, paddr_t);
4610 void		(*pmap_zero_page_func)(paddr_t);
4611 
4612 #if (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0
4613 void
4614 pmap_pte_init_generic(void)
4615 {
4616 
4617 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
4618 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_generic;
4619 
4620 	pte_l2_l_cache_mode = L2_B|L2_C;
4621 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_generic;
4622 
4623 	pte_l2_s_cache_mode = L2_B|L2_C;
4624 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_generic;
4625 
4626 	/*
4627 	 * If we have a write-through cache, set B and C.  If
4628 	 * we have a write-back cache, then we assume setting
4629 	 * only C will make those pages write-through.
4630 	 */
4631 	if (cpufuncs.cf_dcache_wb_range == (void *) cpufunc_nullop) {
4632 		pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
4633 		pte_l2_l_cache_mode_pt = L2_B|L2_C;
4634 		pte_l2_s_cache_mode_pt = L2_B|L2_C;
4635 	} else {
4636 		pte_l1_s_cache_mode_pt = L1_S_C;
4637 		pte_l2_l_cache_mode_pt = L2_C;
4638 		pte_l2_s_cache_mode_pt = L2_C;
4639 	}
4640 
4641 	pte_l2_s_prot_u = L2_S_PROT_U_generic;
4642 	pte_l2_s_prot_w = L2_S_PROT_W_generic;
4643 	pte_l2_s_prot_mask = L2_S_PROT_MASK_generic;
4644 
4645 	pte_l1_s_proto = L1_S_PROTO_generic;
4646 	pte_l1_c_proto = L1_C_PROTO_generic;
4647 	pte_l2_s_proto = L2_S_PROTO_generic;
4648 
4649 	pmap_copy_page_func = pmap_copy_page_generic;
4650 	pmap_zero_page_func = pmap_zero_page_generic;
4651 }
4652 
4653 #if defined(CPU_ARM8)
4654 void
4655 pmap_pte_init_arm8(void)
4656 {
4657 
4658 	/*
4659 	 * ARM8 is compatible with generic, but we need to use
4660 	 * the page tables uncached.
4661 	 */
4662 	pmap_pte_init_generic();
4663 
4664 	pte_l1_s_cache_mode_pt = 0;
4665 	pte_l2_l_cache_mode_pt = 0;
4666 	pte_l2_s_cache_mode_pt = 0;
4667 }
4668 #endif /* CPU_ARM8 */
4669 
4670 #if defined(CPU_ARM9)
4671 void
4672 pmap_pte_init_arm9(void)
4673 {
4674 
4675 	/*
4676 	 * ARM9 is compatible with generic, but we want to use
4677 	 * write-through caching for now.
4678 	 */
4679 	pmap_pte_init_generic();
4680 
4681 	pte_l1_s_cache_mode = L1_S_C;
4682 	pte_l2_l_cache_mode = L2_C;
4683 	pte_l2_s_cache_mode = L2_C;
4684 
4685 	pte_l1_s_cache_mode_pt = L1_S_C;
4686 	pte_l2_l_cache_mode_pt = L2_C;
4687 	pte_l2_s_cache_mode_pt = L2_C;
4688 }
4689 #endif /* CPU_ARM9 */
4690 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 */
4691 
4692 #if ARM_MMU_SA1 == 1
4693 void
4694 pmap_pte_init_sa1(void)
4695 {
4696 
4697 	/*
4698 	 * The StrongARM SA-1 cache does not have a write-through
4699 	 * mode.  So, do the generic initialization, then reset
4700 	 * the page table cache mode to B=1,C=1, and note that
4701 	 * the PTEs need to be sync'd.
4702 	 */
4703 	pmap_pte_init_generic();
4704 
4705 	pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C;
4706 	pte_l2_l_cache_mode_pt = L2_B|L2_C;
4707 	pte_l2_s_cache_mode_pt = L2_B|L2_C;
4708 
4709 	pmap_needs_pte_sync = 1;
4710 }
4711 #endif /* ARM_MMU_SA1 == 1*/
4712 
4713 #if ARM_MMU_XSCALE == 1
4714 void
4715 pmap_pte_init_xscale(void)
4716 {
4717 	uint32_t auxctl;
4718 	int write_through = 0;
4719 
4720 	pte_l1_s_cache_mode = L1_S_B|L1_S_C;
4721 	pte_l1_s_cache_mask = L1_S_CACHE_MASK_xscale;
4722 
4723 	pte_l2_l_cache_mode = L2_B|L2_C;
4724 	pte_l2_l_cache_mask = L2_L_CACHE_MASK_xscale;
4725 
4726 	pte_l2_s_cache_mode = L2_B|L2_C;
4727 	pte_l2_s_cache_mask = L2_S_CACHE_MASK_xscale;
4728 
4729 	pte_l1_s_cache_mode_pt = L1_S_C;
4730 	pte_l2_l_cache_mode_pt = L2_C;
4731 	pte_l2_s_cache_mode_pt = L2_C;
4732 
4733 #ifdef XSCALE_CACHE_READ_WRITE_ALLOCATE
4734 	/*
4735 	 * The XScale core has an enhanced mode where writes that
4736 	 * miss the cache cause a cache line to be allocated.  This
4737 	 * is significantly faster than the traditional, write-through
4738 	 * behavior of this case.
4739 	 */
4740 	pte_l1_s_cache_mode |= L1_S_XSCALE_TEX(TEX_XSCALE_X);
4741 	pte_l2_l_cache_mode |= L2_XSCALE_L_TEX(TEX_XSCALE_X);
4742 	pte_l2_s_cache_mode |= L2_XSCALE_T_TEX(TEX_XSCALE_X);
4743 #endif /* XSCALE_CACHE_READ_WRITE_ALLOCATE */
4744 
4745 #ifdef XSCALE_CACHE_WRITE_THROUGH
4746 	/*
4747 	 * Some versions of the XScale core have various bugs in
4748 	 * their cache units, the work-around for which is to run
4749 	 * the cache in write-through mode.  Unfortunately, this
4750 	 * has a major (negative) impact on performance.  So, we
4751 	 * go ahead and run fast-and-loose, in the hopes that we
4752 	 * don't line up the planets in a way that will trip the
4753 	 * bugs.
4754 	 *
4755 	 * However, we give you the option to be slow-but-correct.
4756 	 */
4757 	write_through = 1;
4758 #elif defined(XSCALE_CACHE_WRITE_BACK)
4759 	/* force write back cache mode */
4760 	write_through = 0;
4761 #elif defined(CPU_XSCALE_PXA2X0)
4762 	/*
4763 	 * Intel PXA2[15]0 processors are known to have a bug in
4764 	 * write-back cache on revision 4 and earlier (stepping
4765 	 * A[01] and B[012]).  Fixed for C0 and later.
4766 	 */
4767 	{
4768 		uint32_t id, type;
4769 
4770 		id = cpufunc_id();
4771 		type = id & ~(CPU_ID_XSCALE_COREREV_MASK|CPU_ID_REVISION_MASK);
4772 
4773 		if (type == CPU_ID_PXA250 || type == CPU_ID_PXA210) {
4774 			if ((id & CPU_ID_REVISION_MASK) < 5) {
4775 				/* write through for stepping A0-1 and B0-2 */
4776 				write_through = 1;
4777 			}
4778 		}
4779 	}
4780 #endif /* XSCALE_CACHE_WRITE_THROUGH */
4781 
4782 	if (write_through) {
4783 		pte_l1_s_cache_mode = L1_S_C;
4784 		pte_l2_l_cache_mode = L2_C;
4785 		pte_l2_s_cache_mode = L2_C;
4786 	}
4787 
4788 	pte_l2_s_prot_u = L2_S_PROT_U_xscale;
4789 	pte_l2_s_prot_w = L2_S_PROT_W_xscale;
4790 	pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale;
4791 
4792 	pte_l1_s_proto = L1_S_PROTO_xscale;
4793 	pte_l1_c_proto = L1_C_PROTO_xscale;
4794 	pte_l2_s_proto = L2_S_PROTO_xscale;
4795 
4796 	pmap_copy_page_func = pmap_copy_page_xscale;
4797 	pmap_zero_page_func = pmap_zero_page_xscale;
4798 
4799 	/*
4800 	 * Disable ECC protection of page table access, for now.
4801 	 */
4802 	__asm __volatile("mrc p15, 0, %0, c1, c0, 1"
4803 		: "=r" (auxctl));
4804 	auxctl &= ~XSCALE_AUXCTL_P;
4805 	__asm __volatile("mcr p15, 0, %0, c1, c0, 1"
4806 		:
4807 		: "r" (auxctl));
4808 }
4809 
4810 /*
4811  * xscale_setup_minidata:
4812  *
4813  *	Set up the mini-data cache clean area.  We require the
4814  *	caller to allocate the right amount of physically and
4815  *	virtually contiguous space.
4816  */
4817 void
4818 xscale_setup_minidata(vaddr_t l1pt, vaddr_t va, paddr_t pa)
4819 {
4820 	extern vaddr_t xscale_minidata_clean_addr;
4821 	extern vsize_t xscale_minidata_clean_size; /* already initialized */
4822 	pd_entry_t *pde = (pd_entry_t *) l1pt;
4823 	pt_entry_t *pte;
4824 	vsize_t size;
4825 	uint32_t auxctl;
4826 
4827 	xscale_minidata_clean_addr = va;
4828 
4829 	/* Round it to page size. */
4830 	size = (xscale_minidata_clean_size + L2_S_OFFSET) & L2_S_FRAME;
4831 
4832 	for (; size != 0;
4833 	     va += L2_S_SIZE, pa += L2_S_SIZE, size -= L2_S_SIZE) {
4834 #ifndef ARM32_NEW_VM_LAYOUT
4835 		pte = (pt_entry_t *)
4836 		    kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME);
4837 #else
4838 		pte = (pt_entry_t *) kernel_pt_lookup(
4839 		    pde[L1_IDX(va)] & L1_C_ADDR_MASK);
4840 #endif
4841 		if (pte == NULL)
4842 			panic("xscale_setup_minidata: can't find L2 table for "
4843 			    "VA 0x%08lx", va);
4844 #ifndef ARM32_NEW_VM_LAYOUT
4845 		pte[(va >> PGSHIFT) & 0x3ff] =
4846 #else
4847 		pte[l2pte_index(va)] =
4848 #endif
4849 		    L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, VM_PROT_READ) |
4850 		    L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X);
4851 	}
4852 
4853 	/*
4854 	 * Configure the mini-data cache for write-back with
4855 	 * read/write-allocate.
4856 	 *
4857 	 * NOTE: In order to reconfigure the mini-data cache, we must
4858 	 * make sure it contains no valid data!  In order to do that,
4859 	 * we must issue a global data cache invalidate command!
4860 	 *
4861 	 * WE ASSUME WE ARE RUNNING UN-CACHED WHEN THIS ROUTINE IS CALLED!
4862 	 * THIS IS VERY IMPORTANT!
4863 	 */
4864 
4865 	/* Invalidate data and mini-data. */
4866 	__asm __volatile("mcr p15, 0, %0, c7, c6, 0"
4867 		:
4868 		: "r" (auxctl));
4869 
4870 
4871 	__asm __volatile("mrc p15, 0, %0, c1, c0, 1"
4872 		: "=r" (auxctl));
4873 	auxctl = (auxctl & ~XSCALE_AUXCTL_MD_MASK) | XSCALE_AUXCTL_MD_WB_RWA;
4874 	__asm __volatile("mcr p15, 0, %0, c1, c0, 1"
4875 		:
4876 		: "r" (auxctl));
4877 }
4878 #endif /* ARM_MMU_XSCALE == 1 */
4879 
4880 #if defined(DDB)
4881 /*
4882  * A couple of ddb-callable functions for dumping pmaps
4883  */
4884 void pmap_dump_all(void);
4885 void pmap_dump(pmap_t);
4886 
4887 void
4888 pmap_dump_all(void)
4889 {
4890 	pmap_t pm;
4891 
4892 	LIST_FOREACH(pm, &pmap_pmaps, pm_list) {
4893 		if (pm == pmap_kernel())
4894 			continue;
4895 		pmap_dump(pm);
4896 		printf("\n");
4897 	}
4898 }
4899 
4900 static pt_entry_t ncptes[64];
4901 static void pmap_dump_ncpg(pmap_t);
4902 
4903 void
4904 pmap_dump(pmap_t pm)
4905 {
4906 	struct l2_dtable *l2;
4907 	struct l2_bucket *l2b;
4908 	pt_entry_t *ptep, pte;
4909 	vaddr_t l2_va, l2b_va, va;
4910 	int i, j, k, occ, rows = 0;
4911 	char ch;
4912 
4913 	if (pm == pmap_kernel())
4914 		printf("pmap_kernel (%p): ", pm);
4915 	else
4916 		printf("user pmap (%p): ", pm);
4917 
4918 	printf("domain %d, l1 at %p\n", pm->pm_domain, pm->pm_l1->l1_kva);
4919 
4920 	l2_va = 0;
4921 	for (i = 0; i < L2_SIZE; i++, l2_va += 0x01000000) {
4922 		l2 = pm->pm_l2[i];
4923 
4924 		if (l2 == NULL || l2->l2_occupancy == 0)
4925 			continue;
4926 
4927 		l2b_va = l2_va;
4928 		for (j = 0; j < L2_BUCKET_SIZE; j++, l2b_va += 0x00100000) {
4929 			l2b = &l2->l2_bucket[j];
4930 
4931 			if (l2b->l2b_occupancy == 0 || l2b->l2b_kva == NULL)
4932 				continue;
4933 
4934 			ptep = l2b->l2b_kva;
4935 
4936 			for (k = 0; k < 256 && ptep[k] == 0; k++)
4937 				;
4938 
4939 			k &= ~63;
4940 			occ = l2b->l2b_occupancy;
4941 			va = l2b_va + (k * 4096);
4942 			for (; k < 256; k++, va += 0x1000) {
4943 				if ((k % 64) == 0) {
4944 					if ((rows % 8) == 0) {
4945 						printf(
4946 "          |0000   |8000   |10000  |18000  |20000  |28000  |30000  |38000\n");
4947 					}
4948 					printf("%08lx: ", va);
4949 				}
4950 
4951 				ncptes[k & 63] = 0;
4952 				pte = ptep[k];
4953 				if (pte == 0) {
4954 					ch = '.';
4955 				} else {
4956 					occ--;
4957 					switch (pte & 0x0c) {
4958 					case 0x00:
4959 						ch = 'D'; /* No cache No buff */
4960 						break;
4961 					case 0x04:
4962 						ch = 'B'; /* No cache buff */
4963 						break;
4964 					case 0x08:
4965 						ch = 'C'; /* Cache No buff */
4966 						break;
4967 					case 0x0c:
4968 						ch = 'F'; /* Cache Buff */
4969 						break;
4970 					}
4971 
4972 					if ((pte & L2_S_PROT_U) == L2_S_PROT_U)
4973 						ch += 0x20;
4974 
4975 					if ((pte & 0xc) == 0)
4976 						ncptes[k & 63] = pte;
4977 				}
4978 
4979 				if ((k % 64) == 63) {
4980 					rows++;
4981 					printf("%c\n", ch);
4982 					pmap_dump_ncpg(pm);
4983 					if (occ == 0)
4984 						break;
4985 				} else
4986 					printf("%c", ch);
4987 			}
4988 		}
4989 	}
4990 }
4991 
4992 static void
4993 pmap_dump_ncpg(pmap_t pm)
4994 {
4995 	struct vm_page *pg;
4996 	struct pv_entry *pv;
4997 	int i;
4998 
4999 	for (i = 0; i < 63; i++) {
5000 		if (ncptes[i] == 0)
5001 			continue;
5002 
5003 		pg = PHYS_TO_VM_PAGE(l2pte_pa(ncptes[i]));
5004 		if (pg == NULL)
5005 			continue;
5006 
5007 		printf(" pa 0x%08lx: krw %d kro %d urw %d uro %d\n",
5008 		    pg->phys_addr,
5009 		    pg->mdpage.krw_mappings, pg->mdpage.kro_mappings,
5010 		    pg->mdpage.urw_mappings, pg->mdpage.uro_mappings);
5011 
5012 		for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) {
5013 			printf("   %c va 0x%08lx, flags 0x%x\n",
5014 			    (pm == pv->pv_pmap) ? '*' : ' ',
5015 			    pv->pv_va, pv->pv_flags);
5016 		}
5017 	}
5018 }
5019 #endif
5020