xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision c47dd3db8cf22114267360dddb4b0334b997db0c)
1 /*-
2  * Copyright (c) 2001 The NetBSD Foundation, Inc.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to The NetBSD Foundation
6  * by Matt Thomas <matt@3am-software.com> of Allegro Networks, Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *        This product includes software developed by the NetBSD
19  *        Foundation, Inc. and its contributors.
20  * 4. Neither the name of The NetBSD Foundation nor the names of its
21  *    contributors may be used to endorse or promote products derived
22  *    from this software without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
25  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
26  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
27  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
28  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
29  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
30  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
31  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
32  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
33  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
34  * POSSIBILITY OF SUCH DAMAGE.
35  */
36 /*-
37  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
38  * Copyright (C) 1995, 1996 TooLs GmbH.
39  * All rights reserved.
40  *
41  * Redistribution and use in source and binary forms, with or without
42  * modification, are permitted provided that the following conditions
43  * are met:
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. All advertising materials mentioning features or use of this software
50  *    must display the following acknowledgement:
51  *	This product includes software developed by TooLs GmbH.
52  * 4. The name of TooLs GmbH may not be used to endorse or promote products
53  *    derived from this software without specific prior written permission.
54  *
55  * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
56  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
57  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
58  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
59  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
60  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
61  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
62  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
63  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
64  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
65  *
66  * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $
67  */
68 /*-
69  * Copyright (C) 2001 Benno Rice.
70  * All rights reserved.
71  *
72  * Redistribution and use in source and binary forms, with or without
73  * modification, are permitted provided that the following conditions
74  * are met:
75  * 1. Redistributions of source code must retain the above copyright
76  *    notice, this list of conditions and the following disclaimer.
77  * 2. Redistributions in binary form must reproduce the above copyright
78  *    notice, this list of conditions and the following disclaimer in the
79  *    documentation and/or other materials provided with the distribution.
80  *
81  * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR
82  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
83  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
84  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
85  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
86  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
87  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
88  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
89  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
90  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
91  */
92 
93 #include <sys/cdefs.h>
94 __FBSDID("$FreeBSD$");
95 
96 /*
97  * Manages physical address maps.
98  *
99  * In addition to hardware address maps, this module is called upon to
100  * provide software-use-only maps which may or may not be stored in the
101  * same form as hardware maps.  These pseudo-maps are used to store
102  * intermediate results from copy operations to and from address spaces.
103  *
104  * Since the information managed by this module is also stored by the
105  * logical address mapping module, this module may throw away valid virtual
106  * to physical mappings at almost any time.  However, invalidations of
107  * mappings must be done as requested.
108  *
109  * In order to cope with hardware architectures which make virtual to
110  * physical map invalidates expensive, this module may delay invalidate
111  * reduced protection operations until such time as they are actually
112  * necessary.  This module is given full information as to which processors
113  * are currently using which maps, and to when physical maps must be made
114  * correct.
115  */
116 
117 #include "opt_kstack_pages.h"
118 
119 #include <sys/param.h>
120 #include <sys/kernel.h>
121 #include <sys/queue.h>
122 #include <sys/cpuset.h>
123 #include <sys/ktr.h>
124 #include <sys/lock.h>
125 #include <sys/msgbuf.h>
126 #include <sys/mutex.h>
127 #include <sys/proc.h>
128 #include <sys/sched.h>
129 #include <sys/sysctl.h>
130 #include <sys/systm.h>
131 #include <sys/vmmeter.h>
132 
133 #include <dev/ofw/openfirm.h>
134 
135 #include <vm/vm.h>
136 #include <vm/vm_param.h>
137 #include <vm/vm_kern.h>
138 #include <vm/vm_page.h>
139 #include <vm/vm_map.h>
140 #include <vm/vm_object.h>
141 #include <vm/vm_extern.h>
142 #include <vm/vm_pageout.h>
143 #include <vm/vm_pager.h>
144 #include <vm/uma.h>
145 
146 #include <machine/cpu.h>
147 #include <machine/platform.h>
148 #include <machine/bat.h>
149 #include <machine/frame.h>
150 #include <machine/md_var.h>
151 #include <machine/psl.h>
152 #include <machine/pte.h>
153 #include <machine/smp.h>
154 #include <machine/sr.h>
155 #include <machine/mmuvar.h>
156 
157 #include "mmu_if.h"
158 
159 #define	MOEA_DEBUG
160 
161 #define TODO	panic("%s: not implemented", __func__);
162 
163 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
164 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
165 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
166 
167 struct ofw_map {
168 	vm_offset_t	om_va;
169 	vm_size_t	om_len;
170 	vm_offset_t	om_pa;
171 	u_int		om_mode;
172 };
173 
174 /*
175  * Map of physical memory regions.
176  */
177 static struct	mem_region *regions;
178 static struct	mem_region *pregions;
179 static u_int    phys_avail_count;
180 static int	regions_sz, pregions_sz;
181 static struct	ofw_map *translations;
182 
183 /*
184  * Lock for the pteg and pvo tables.
185  */
186 struct mtx	moea_table_mutex;
187 struct mtx	moea_vsid_mutex;
188 
189 /* tlbie instruction synchronization */
190 static struct mtx tlbie_mtx;
191 
192 /*
193  * PTEG data.
194  */
195 static struct	pteg *moea_pteg_table;
196 u_int		moea_pteg_count;
197 u_int		moea_pteg_mask;
198 
199 /*
200  * PVO data.
201  */
202 struct	pvo_head *moea_pvo_table;		/* pvo entries by pteg index */
203 struct	pvo_head moea_pvo_kunmanaged =
204     LIST_HEAD_INITIALIZER(moea_pvo_kunmanaged);	/* list of unmanaged pages */
205 
206 uma_zone_t	moea_upvo_zone;	/* zone for pvo entries for unmanaged pages */
207 uma_zone_t	moea_mpvo_zone;	/* zone for pvo entries for managed pages */
208 
209 #define	BPVO_POOL_SIZE	32768
210 static struct	pvo_entry *moea_bpvo_pool;
211 static int	moea_bpvo_pool_index = 0;
212 
213 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
214 static u_int	moea_vsid_bitmap[NPMAPS / VSID_NBPW];
215 
216 static boolean_t moea_initialized = FALSE;
217 
218 /*
219  * Statistics.
220  */
221 u_int	moea_pte_valid = 0;
222 u_int	moea_pte_overflow = 0;
223 u_int	moea_pte_replacements = 0;
224 u_int	moea_pvo_entries = 0;
225 u_int	moea_pvo_enter_calls = 0;
226 u_int	moea_pvo_remove_calls = 0;
227 u_int	moea_pte_spills = 0;
228 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_valid, CTLFLAG_RD, &moea_pte_valid,
229     0, "");
230 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_overflow, CTLFLAG_RD,
231     &moea_pte_overflow, 0, "");
232 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_replacements, CTLFLAG_RD,
233     &moea_pte_replacements, 0, "");
234 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_entries, CTLFLAG_RD, &moea_pvo_entries,
235     0, "");
236 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_enter_calls, CTLFLAG_RD,
237     &moea_pvo_enter_calls, 0, "");
238 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_remove_calls, CTLFLAG_RD,
239     &moea_pvo_remove_calls, 0, "");
240 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_spills, CTLFLAG_RD,
241     &moea_pte_spills, 0, "");
242 
243 /*
244  * Allocate physical memory for use in moea_bootstrap.
245  */
246 static vm_offset_t	moea_bootstrap_alloc(vm_size_t, u_int);
247 
248 /*
249  * PTE calls.
250  */
251 static int		moea_pte_insert(u_int, struct pte *);
252 
253 /*
254  * PVO calls.
255  */
256 static int	moea_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *,
257 		    vm_offset_t, vm_offset_t, u_int, int);
258 static void	moea_pvo_remove(struct pvo_entry *, int);
259 static struct	pvo_entry *moea_pvo_find_va(pmap_t, vm_offset_t, int *);
260 static struct	pte *moea_pvo_to_pte(const struct pvo_entry *, int);
261 
262 /*
263  * Utility routines.
264  */
265 static void		moea_enter_locked(pmap_t, vm_offset_t, vm_page_t,
266 			    vm_prot_t, boolean_t);
267 static void		moea_syncicache(vm_offset_t, vm_size_t);
268 static boolean_t	moea_query_bit(vm_page_t, int);
269 static u_int		moea_clear_bit(vm_page_t, int);
270 static void		moea_kremove(mmu_t, vm_offset_t);
271 int		moea_pte_spill(vm_offset_t);
272 
273 /*
274  * Kernel MMU interface
275  */
276 void moea_change_wiring(mmu_t, pmap_t, vm_offset_t, boolean_t);
277 void moea_clear_modify(mmu_t, vm_page_t);
278 void moea_clear_reference(mmu_t, vm_page_t);
279 void moea_copy_page(mmu_t, vm_page_t, vm_page_t);
280 void moea_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, boolean_t);
281 void moea_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
282     vm_prot_t);
283 void moea_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
284 vm_paddr_t moea_extract(mmu_t, pmap_t, vm_offset_t);
285 vm_page_t moea_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
286 void moea_init(mmu_t);
287 boolean_t moea_is_modified(mmu_t, vm_page_t);
288 boolean_t moea_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
289 boolean_t moea_is_referenced(mmu_t, vm_page_t);
290 boolean_t moea_ts_referenced(mmu_t, vm_page_t);
291 vm_offset_t moea_map(mmu_t, vm_offset_t *, vm_offset_t, vm_offset_t, int);
292 boolean_t moea_page_exists_quick(mmu_t, pmap_t, vm_page_t);
293 int moea_page_wired_mappings(mmu_t, vm_page_t);
294 void moea_pinit(mmu_t, pmap_t);
295 void moea_pinit0(mmu_t, pmap_t);
296 void moea_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
297 void moea_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
298 void moea_qremove(mmu_t, vm_offset_t, int);
299 void moea_release(mmu_t, pmap_t);
300 void moea_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
301 void moea_remove_all(mmu_t, vm_page_t);
302 void moea_remove_write(mmu_t, vm_page_t);
303 void moea_zero_page(mmu_t, vm_page_t);
304 void moea_zero_page_area(mmu_t, vm_page_t, int, int);
305 void moea_zero_page_idle(mmu_t, vm_page_t);
306 void moea_activate(mmu_t, struct thread *);
307 void moea_deactivate(mmu_t, struct thread *);
308 void moea_cpu_bootstrap(mmu_t, int);
309 void moea_bootstrap(mmu_t, vm_offset_t, vm_offset_t);
310 void *moea_mapdev(mmu_t, vm_offset_t, vm_size_t);
311 void *moea_mapdev_attr(mmu_t, vm_offset_t, vm_size_t, vm_memattr_t);
312 void moea_unmapdev(mmu_t, vm_offset_t, vm_size_t);
313 vm_offset_t moea_kextract(mmu_t, vm_offset_t);
314 void moea_kenter_attr(mmu_t, vm_offset_t, vm_offset_t, vm_memattr_t);
315 void moea_kenter(mmu_t, vm_offset_t, vm_offset_t);
316 void moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma);
317 boolean_t moea_dev_direct_mapped(mmu_t, vm_offset_t, vm_size_t);
318 static void moea_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
319 
320 static mmu_method_t moea_methods[] = {
321 	MMUMETHOD(mmu_change_wiring,	moea_change_wiring),
322 	MMUMETHOD(mmu_clear_modify,	moea_clear_modify),
323 	MMUMETHOD(mmu_clear_reference,	moea_clear_reference),
324 	MMUMETHOD(mmu_copy_page,	moea_copy_page),
325 	MMUMETHOD(mmu_enter,		moea_enter),
326 	MMUMETHOD(mmu_enter_object,	moea_enter_object),
327 	MMUMETHOD(mmu_enter_quick,	moea_enter_quick),
328 	MMUMETHOD(mmu_extract,		moea_extract),
329 	MMUMETHOD(mmu_extract_and_hold,	moea_extract_and_hold),
330 	MMUMETHOD(mmu_init,		moea_init),
331 	MMUMETHOD(mmu_is_modified,	moea_is_modified),
332 	MMUMETHOD(mmu_is_prefaultable,	moea_is_prefaultable),
333 	MMUMETHOD(mmu_is_referenced,	moea_is_referenced),
334 	MMUMETHOD(mmu_ts_referenced,	moea_ts_referenced),
335 	MMUMETHOD(mmu_map,     		moea_map),
336 	MMUMETHOD(mmu_page_exists_quick,moea_page_exists_quick),
337 	MMUMETHOD(mmu_page_wired_mappings,moea_page_wired_mappings),
338 	MMUMETHOD(mmu_pinit,		moea_pinit),
339 	MMUMETHOD(mmu_pinit0,		moea_pinit0),
340 	MMUMETHOD(mmu_protect,		moea_protect),
341 	MMUMETHOD(mmu_qenter,		moea_qenter),
342 	MMUMETHOD(mmu_qremove,		moea_qremove),
343 	MMUMETHOD(mmu_release,		moea_release),
344 	MMUMETHOD(mmu_remove,		moea_remove),
345 	MMUMETHOD(mmu_remove_all,      	moea_remove_all),
346 	MMUMETHOD(mmu_remove_write,	moea_remove_write),
347 	MMUMETHOD(mmu_sync_icache,	moea_sync_icache),
348 	MMUMETHOD(mmu_zero_page,       	moea_zero_page),
349 	MMUMETHOD(mmu_zero_page_area,	moea_zero_page_area),
350 	MMUMETHOD(mmu_zero_page_idle,	moea_zero_page_idle),
351 	MMUMETHOD(mmu_activate,		moea_activate),
352 	MMUMETHOD(mmu_deactivate,      	moea_deactivate),
353 	MMUMETHOD(mmu_page_set_memattr,	moea_page_set_memattr),
354 
355 	/* Internal interfaces */
356 	MMUMETHOD(mmu_bootstrap,       	moea_bootstrap),
357 	MMUMETHOD(mmu_cpu_bootstrap,   	moea_cpu_bootstrap),
358 	MMUMETHOD(mmu_mapdev_attr,	moea_mapdev_attr),
359 	MMUMETHOD(mmu_mapdev,		moea_mapdev),
360 	MMUMETHOD(mmu_unmapdev,		moea_unmapdev),
361 	MMUMETHOD(mmu_kextract,		moea_kextract),
362 	MMUMETHOD(mmu_kenter,		moea_kenter),
363 	MMUMETHOD(mmu_kenter_attr,	moea_kenter_attr),
364 	MMUMETHOD(mmu_dev_direct_mapped,moea_dev_direct_mapped),
365 
366 	{ 0, 0 }
367 };
368 
369 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods, 0);
370 
371 static __inline uint32_t
372 moea_calc_wimg(vm_offset_t pa, vm_memattr_t ma)
373 {
374 	uint32_t pte_lo;
375 	int i;
376 
377 	if (ma != VM_MEMATTR_DEFAULT) {
378 		switch (ma) {
379 		case VM_MEMATTR_UNCACHEABLE:
380 			return (PTE_I | PTE_G);
381 		case VM_MEMATTR_WRITE_COMBINING:
382 		case VM_MEMATTR_WRITE_BACK:
383 		case VM_MEMATTR_PREFETCHABLE:
384 			return (PTE_I);
385 		case VM_MEMATTR_WRITE_THROUGH:
386 			return (PTE_W | PTE_M);
387 		}
388 	}
389 
390 	/*
391 	 * Assume the page is cache inhibited and access is guarded unless
392 	 * it's in our available memory array.
393 	 */
394 	pte_lo = PTE_I | PTE_G;
395 	for (i = 0; i < pregions_sz; i++) {
396 		if ((pa >= pregions[i].mr_start) &&
397 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
398 			pte_lo = PTE_M;
399 			break;
400 		}
401 	}
402 
403 	return pte_lo;
404 }
405 
406 static void
407 tlbie(vm_offset_t va)
408 {
409 
410 	mtx_lock_spin(&tlbie_mtx);
411 	__asm __volatile("ptesync");
412 	__asm __volatile("tlbie %0" :: "r"(va));
413 	__asm __volatile("eieio; tlbsync; ptesync");
414 	mtx_unlock_spin(&tlbie_mtx);
415 }
416 
417 static void
418 tlbia(void)
419 {
420 	vm_offset_t va;
421 
422 	for (va = 0; va < 0x00040000; va += 0x00001000) {
423 		__asm __volatile("tlbie %0" :: "r"(va));
424 		powerpc_sync();
425 	}
426 	__asm __volatile("tlbsync");
427 	powerpc_sync();
428 }
429 
430 static __inline int
431 va_to_sr(u_int *sr, vm_offset_t va)
432 {
433 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
434 }
435 
436 static __inline u_int
437 va_to_pteg(u_int sr, vm_offset_t addr)
438 {
439 	u_int hash;
440 
441 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
442 	    ADDR_PIDX_SHFT);
443 	return (hash & moea_pteg_mask);
444 }
445 
446 static __inline struct pvo_head *
447 vm_page_to_pvoh(vm_page_t m)
448 {
449 
450 	return (&m->md.mdpg_pvoh);
451 }
452 
453 static __inline void
454 moea_attr_clear(vm_page_t m, int ptebit)
455 {
456 
457 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
458 	m->md.mdpg_attrs &= ~ptebit;
459 }
460 
461 static __inline int
462 moea_attr_fetch(vm_page_t m)
463 {
464 
465 	return (m->md.mdpg_attrs);
466 }
467 
468 static __inline void
469 moea_attr_save(vm_page_t m, int ptebit)
470 {
471 
472 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
473 	m->md.mdpg_attrs |= ptebit;
474 }
475 
476 static __inline int
477 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
478 {
479 	if (pt->pte_hi == pvo_pt->pte_hi)
480 		return (1);
481 
482 	return (0);
483 }
484 
485 static __inline int
486 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
487 {
488 	return (pt->pte_hi & ~PTE_VALID) ==
489 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
490 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
491 }
492 
493 static __inline void
494 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
495 {
496 
497 	mtx_assert(&moea_table_mutex, MA_OWNED);
498 
499 	/*
500 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
501 	 * set when the real pte is set in memory.
502 	 *
503 	 * Note: Don't set the valid bit for correct operation of tlb update.
504 	 */
505 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
506 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
507 	pt->pte_lo = pte_lo;
508 }
509 
510 static __inline void
511 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
512 {
513 
514 	mtx_assert(&moea_table_mutex, MA_OWNED);
515 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
516 }
517 
518 static __inline void
519 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
520 {
521 
522 	mtx_assert(&moea_table_mutex, MA_OWNED);
523 
524 	/*
525 	 * As shown in Section 7.6.3.2.3
526 	 */
527 	pt->pte_lo &= ~ptebit;
528 	tlbie(va);
529 }
530 
531 static __inline void
532 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
533 {
534 
535 	mtx_assert(&moea_table_mutex, MA_OWNED);
536 	pvo_pt->pte_hi |= PTE_VALID;
537 
538 	/*
539 	 * Update the PTE as defined in section 7.6.3.1.
540 	 * Note that the REF/CHG bits are from pvo_pt and thus should havce
541 	 * been saved so this routine can restore them (if desired).
542 	 */
543 	pt->pte_lo = pvo_pt->pte_lo;
544 	powerpc_sync();
545 	pt->pte_hi = pvo_pt->pte_hi;
546 	powerpc_sync();
547 	moea_pte_valid++;
548 }
549 
550 static __inline void
551 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
552 {
553 
554 	mtx_assert(&moea_table_mutex, MA_OWNED);
555 	pvo_pt->pte_hi &= ~PTE_VALID;
556 
557 	/*
558 	 * Force the reg & chg bits back into the PTEs.
559 	 */
560 	powerpc_sync();
561 
562 	/*
563 	 * Invalidate the pte.
564 	 */
565 	pt->pte_hi &= ~PTE_VALID;
566 
567 	tlbie(va);
568 
569 	/*
570 	 * Save the reg & chg bits.
571 	 */
572 	moea_pte_synch(pt, pvo_pt);
573 	moea_pte_valid--;
574 }
575 
576 static __inline void
577 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
578 {
579 
580 	/*
581 	 * Invalidate the PTE
582 	 */
583 	moea_pte_unset(pt, pvo_pt, va);
584 	moea_pte_set(pt, pvo_pt);
585 }
586 
587 /*
588  * Quick sort callout for comparing memory regions.
589  */
590 static int	mr_cmp(const void *a, const void *b);
591 static int	om_cmp(const void *a, const void *b);
592 
593 static int
594 mr_cmp(const void *a, const void *b)
595 {
596 	const struct	mem_region *regiona;
597 	const struct	mem_region *regionb;
598 
599 	regiona = a;
600 	regionb = b;
601 	if (regiona->mr_start < regionb->mr_start)
602 		return (-1);
603 	else if (regiona->mr_start > regionb->mr_start)
604 		return (1);
605 	else
606 		return (0);
607 }
608 
609 static int
610 om_cmp(const void *a, const void *b)
611 {
612 	const struct	ofw_map *mapa;
613 	const struct	ofw_map *mapb;
614 
615 	mapa = a;
616 	mapb = b;
617 	if (mapa->om_pa < mapb->om_pa)
618 		return (-1);
619 	else if (mapa->om_pa > mapb->om_pa)
620 		return (1);
621 	else
622 		return (0);
623 }
624 
625 void
626 moea_cpu_bootstrap(mmu_t mmup, int ap)
627 {
628 	u_int sdr;
629 	int i;
630 
631 	if (ap) {
632 		powerpc_sync();
633 		__asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu));
634 		__asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl));
635 		isync();
636 		__asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu));
637 		__asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl));
638 		isync();
639 	}
640 
641 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
642 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
643 	isync();
644 
645 	__asm __volatile("mtibatu 1,%0" :: "r"(0));
646 	__asm __volatile("mtdbatu 2,%0" :: "r"(0));
647 	__asm __volatile("mtibatu 2,%0" :: "r"(0));
648 	__asm __volatile("mtdbatu 3,%0" :: "r"(0));
649 	__asm __volatile("mtibatu 3,%0" :: "r"(0));
650 	isync();
651 
652 	for (i = 0; i < 16; i++)
653 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
654 	powerpc_sync();
655 
656 	sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10);
657 	__asm __volatile("mtsdr1 %0" :: "r"(sdr));
658 	isync();
659 
660 	tlbia();
661 }
662 
663 void
664 moea_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
665 {
666 	ihandle_t	mmui;
667 	phandle_t	chosen, mmu;
668 	int		sz;
669 	int		i, j;
670 	vm_size_t	size, physsz, hwphyssz;
671 	vm_offset_t	pa, va, off;
672 	void		*dpcpu;
673 	register_t	msr;
674 
675         /*
676          * Set up BAT0 to map the lowest 256 MB area
677          */
678         battable[0x0].batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
679         battable[0x0].batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
680 
681         /*
682          * Map PCI memory space.
683          */
684         battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
685         battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
686 
687         battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
688         battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
689 
690         battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
691         battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
692 
693         battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
694         battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
695 
696         /*
697          * Map obio devices.
698          */
699         battable[0xf].batl = BATL(0xf0000000, BAT_I|BAT_G, BAT_PP_RW);
700         battable[0xf].batu = BATU(0xf0000000, BAT_BL_256M, BAT_Vs);
701 
702 	/*
703 	 * Use an IBAT and a DBAT to map the bottom segment of memory
704 	 * where we are. Turn off instruction relocation temporarily
705 	 * to prevent faults while reprogramming the IBAT.
706 	 */
707 	msr = mfmsr();
708 	mtmsr(msr & ~PSL_IR);
709 	__asm (".balign 32; \n"
710 	       "mtibatu 0,%0; mtibatl 0,%1; isync; \n"
711 	       "mtdbatu 0,%0; mtdbatl 0,%1; isync"
712 	    :: "r"(battable[0].batu), "r"(battable[0].batl));
713 	mtmsr(msr);
714 
715 	/* map pci space */
716 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
717 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
718 	isync();
719 
720 	/* set global direct map flag */
721 	hw_direct_map = 1;
722 
723 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
724 	CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
725 
726 	qsort(pregions, pregions_sz, sizeof(*pregions), mr_cmp);
727 	for (i = 0; i < pregions_sz; i++) {
728 		vm_offset_t pa;
729 		vm_offset_t end;
730 
731 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
732 			pregions[i].mr_start,
733 			pregions[i].mr_start + pregions[i].mr_size,
734 			pregions[i].mr_size);
735 		/*
736 		 * Install entries into the BAT table to allow all
737 		 * of physmem to be convered by on-demand BAT entries.
738 		 * The loop will sometimes set the same battable element
739 		 * twice, but that's fine since they won't be used for
740 		 * a while yet.
741 		 */
742 		pa = pregions[i].mr_start & 0xf0000000;
743 		end = pregions[i].mr_start + pregions[i].mr_size;
744 		do {
745                         u_int n = pa >> ADDR_SR_SHFT;
746 
747 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
748 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
749 			pa += SEGMENT_LENGTH;
750 		} while (pa < end);
751 	}
752 
753 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
754 		panic("moea_bootstrap: phys_avail too small");
755 	qsort(regions, regions_sz, sizeof(*regions), mr_cmp);
756 	phys_avail_count = 0;
757 	physsz = 0;
758 	hwphyssz = 0;
759 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
760 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
761 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
762 		    regions[i].mr_start + regions[i].mr_size,
763 		    regions[i].mr_size);
764 		if (hwphyssz != 0 &&
765 		    (physsz + regions[i].mr_size) >= hwphyssz) {
766 			if (physsz < hwphyssz) {
767 				phys_avail[j] = regions[i].mr_start;
768 				phys_avail[j + 1] = regions[i].mr_start +
769 				    hwphyssz - physsz;
770 				physsz = hwphyssz;
771 				phys_avail_count++;
772 			}
773 			break;
774 		}
775 		phys_avail[j] = regions[i].mr_start;
776 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
777 		phys_avail_count++;
778 		physsz += regions[i].mr_size;
779 	}
780 	physmem = btoc(physsz);
781 
782 	/*
783 	 * Allocate PTEG table.
784 	 */
785 #ifdef PTEGCOUNT
786 	moea_pteg_count = PTEGCOUNT;
787 #else
788 	moea_pteg_count = 0x1000;
789 
790 	while (moea_pteg_count < physmem)
791 		moea_pteg_count <<= 1;
792 
793 	moea_pteg_count >>= 1;
794 #endif /* PTEGCOUNT */
795 
796 	size = moea_pteg_count * sizeof(struct pteg);
797 	CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
798 	    size);
799 	moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
800 	CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
801 	bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
802 	moea_pteg_mask = moea_pteg_count - 1;
803 
804 	/*
805 	 * Allocate pv/overflow lists.
806 	 */
807 	size = sizeof(struct pvo_head) * moea_pteg_count;
808 	moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
809 	    PAGE_SIZE);
810 	CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
811 	for (i = 0; i < moea_pteg_count; i++)
812 		LIST_INIT(&moea_pvo_table[i]);
813 
814 	/*
815 	 * Initialize the lock that synchronizes access to the pteg and pvo
816 	 * tables.
817 	 */
818 	mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF |
819 	    MTX_RECURSE);
820 	mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF);
821 
822 	mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN);
823 
824 	/*
825 	 * Initialise the unmanaged pvo pool.
826 	 */
827 	moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
828 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
829 	moea_bpvo_pool_index = 0;
830 
831 	/*
832 	 * Make sure kernel vsid is allocated as well as VSID 0.
833 	 */
834 	moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
835 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
836 	moea_vsid_bitmap[0] |= 1;
837 
838 	/*
839 	 * Initialize the kernel pmap (which is statically allocated).
840 	 */
841 	PMAP_LOCK_INIT(kernel_pmap);
842 	for (i = 0; i < 16; i++)
843 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
844 	CPU_FILL(&kernel_pmap->pm_active);
845 
846 	/*
847 	 * Set up the Open Firmware mappings
848 	 */
849 	if ((chosen = OF_finddevice("/chosen")) == -1)
850 		panic("moea_bootstrap: can't find /chosen");
851 	OF_getprop(chosen, "mmu", &mmui, 4);
852 	if ((mmu = OF_instance_to_package(mmui)) == -1)
853 		panic("moea_bootstrap: can't get mmu package");
854 	if ((sz = OF_getproplen(mmu, "translations")) == -1)
855 		panic("moea_bootstrap: can't get ofw translation count");
856 	translations = NULL;
857 	for (i = 0; phys_avail[i] != 0; i += 2) {
858 		if (phys_avail[i + 1] >= sz) {
859 			translations = (struct ofw_map *)phys_avail[i];
860 			break;
861 		}
862 	}
863 	if (translations == NULL)
864 		panic("moea_bootstrap: no space to copy translations");
865 	bzero(translations, sz);
866 	if (OF_getprop(mmu, "translations", translations, sz) == -1)
867 		panic("moea_bootstrap: can't get ofw translations");
868 	CTR0(KTR_PMAP, "moea_bootstrap: translations");
869 	sz /= sizeof(*translations);
870 	qsort(translations, sz, sizeof (*translations), om_cmp);
871 	for (i = 0; i < sz; i++) {
872 		CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
873 		    translations[i].om_pa, translations[i].om_va,
874 		    translations[i].om_len);
875 
876 		/*
877 		 * If the mapping is 1:1, let the RAM and device on-demand
878 		 * BAT tables take care of the translation.
879 		 */
880 		if (translations[i].om_va == translations[i].om_pa)
881 			continue;
882 
883 		/* Enter the pages */
884 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE)
885 			moea_kenter(mmup, translations[i].om_va + off,
886 				    translations[i].om_pa + off);
887 	}
888 
889 	/*
890 	 * Calculate the last available physical address.
891 	 */
892 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
893 		;
894 	Maxmem = powerpc_btop(phys_avail[i + 1]);
895 
896 	moea_cpu_bootstrap(mmup,0);
897 
898 	pmap_bootstrapped++;
899 
900 	/*
901 	 * Set the start and end of kva.
902 	 */
903 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
904 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
905 
906 	/*
907 	 * Allocate a kernel stack with a guard page for thread0 and map it
908 	 * into the kernel page map.
909 	 */
910 	pa = moea_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, PAGE_SIZE);
911 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
912 	virtual_avail = va + KSTACK_PAGES * PAGE_SIZE;
913 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
914 	thread0.td_kstack = va;
915 	thread0.td_kstack_pages = KSTACK_PAGES;
916 	for (i = 0; i < KSTACK_PAGES; i++) {
917 		moea_kenter(mmup, va, pa);
918 		pa += PAGE_SIZE;
919 		va += PAGE_SIZE;
920 	}
921 
922 	/*
923 	 * Allocate virtual address space for the message buffer.
924 	 */
925 	pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE);
926 	msgbufp = (struct msgbuf *)virtual_avail;
927 	va = virtual_avail;
928 	virtual_avail += round_page(msgbufsize);
929 	while (va < virtual_avail) {
930 		moea_kenter(mmup, va, pa);
931 		pa += PAGE_SIZE;
932 		va += PAGE_SIZE;
933 	}
934 
935 	/*
936 	 * Allocate virtual address space for the dynamic percpu area.
937 	 */
938 	pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
939 	dpcpu = (void *)virtual_avail;
940 	va = virtual_avail;
941 	virtual_avail += DPCPU_SIZE;
942 	while (va < virtual_avail) {
943 		moea_kenter(mmup, va, pa);
944 		pa += PAGE_SIZE;
945 		va += PAGE_SIZE;
946 	}
947 	dpcpu_init(dpcpu, 0);
948 }
949 
950 /*
951  * Activate a user pmap.  The pmap must be activated before it's address
952  * space can be accessed in any way.
953  */
954 void
955 moea_activate(mmu_t mmu, struct thread *td)
956 {
957 	pmap_t	pm, pmr;
958 
959 	/*
960 	 * Load all the data we need up front to encourage the compiler to
961 	 * not issue any loads while we have interrupts disabled below.
962 	 */
963 	pm = &td->td_proc->p_vmspace->vm_pmap;
964 	pmr = pm->pmap_phys;
965 
966 	sched_pin();
967 	CPU_OR(&pm->pm_active, PCPU_PTR(cpumask));
968 	sched_unpin();
969 	PCPU_SET(curpmap, pmr);
970 }
971 
972 void
973 moea_deactivate(mmu_t mmu, struct thread *td)
974 {
975 	pmap_t	pm;
976 
977 	pm = &td->td_proc->p_vmspace->vm_pmap;
978 	sched_pin();
979 	CPU_NAND(&pm->pm_active, PCPU_PTR(cpumask));
980 	sched_unpin();
981 	PCPU_SET(curpmap, NULL);
982 }
983 
984 void
985 moea_change_wiring(mmu_t mmu, pmap_t pm, vm_offset_t va, boolean_t wired)
986 {
987 	struct	pvo_entry *pvo;
988 
989 	PMAP_LOCK(pm);
990 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
991 
992 	if (pvo != NULL) {
993 		if (wired) {
994 			if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
995 				pm->pm_stats.wired_count++;
996 			pvo->pvo_vaddr |= PVO_WIRED;
997 		} else {
998 			if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
999 				pm->pm_stats.wired_count--;
1000 			pvo->pvo_vaddr &= ~PVO_WIRED;
1001 		}
1002 	}
1003 	PMAP_UNLOCK(pm);
1004 }
1005 
1006 void
1007 moea_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1008 {
1009 	vm_offset_t	dst;
1010 	vm_offset_t	src;
1011 
1012 	dst = VM_PAGE_TO_PHYS(mdst);
1013 	src = VM_PAGE_TO_PHYS(msrc);
1014 
1015 	kcopy((void *)src, (void *)dst, PAGE_SIZE);
1016 }
1017 
1018 /*
1019  * Zero a page of physical memory by temporarily mapping it into the tlb.
1020  */
1021 void
1022 moea_zero_page(mmu_t mmu, vm_page_t m)
1023 {
1024 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1025 	void *va = (void *)pa;
1026 
1027 	bzero(va, PAGE_SIZE);
1028 }
1029 
1030 void
1031 moea_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1032 {
1033 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1034 	void *va = (void *)(pa + off);
1035 
1036 	bzero(va, size);
1037 }
1038 
1039 void
1040 moea_zero_page_idle(mmu_t mmu, vm_page_t m)
1041 {
1042 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1043 	void *va = (void *)pa;
1044 
1045 	bzero(va, PAGE_SIZE);
1046 }
1047 
1048 /*
1049  * Map the given physical page at the specified virtual address in the
1050  * target pmap with the protection requested.  If specified the page
1051  * will be wired down.
1052  */
1053 void
1054 moea_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1055 	   boolean_t wired)
1056 {
1057 
1058 	vm_page_lock_queues();
1059 	PMAP_LOCK(pmap);
1060 	moea_enter_locked(pmap, va, m, prot, wired);
1061 	vm_page_unlock_queues();
1062 	PMAP_UNLOCK(pmap);
1063 }
1064 
1065 /*
1066  * Map the given physical page at the specified virtual address in the
1067  * target pmap with the protection requested.  If specified the page
1068  * will be wired down.
1069  *
1070  * The page queues and pmap must be locked.
1071  */
1072 static void
1073 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1074     boolean_t wired)
1075 {
1076 	struct		pvo_head *pvo_head;
1077 	uma_zone_t	zone;
1078 	vm_page_t	pg;
1079 	u_int		pte_lo, pvo_flags, was_exec;
1080 	int		error;
1081 
1082 	if (!moea_initialized) {
1083 		pvo_head = &moea_pvo_kunmanaged;
1084 		zone = moea_upvo_zone;
1085 		pvo_flags = 0;
1086 		pg = NULL;
1087 		was_exec = PTE_EXEC;
1088 	} else {
1089 		pvo_head = vm_page_to_pvoh(m);
1090 		pg = m;
1091 		zone = moea_mpvo_zone;
1092 		pvo_flags = PVO_MANAGED;
1093 		was_exec = 0;
1094 	}
1095 	if (pmap_bootstrapped)
1096 		mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1097 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1098 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0 ||
1099 	    (m->oflags & VPO_BUSY) != 0 || VM_OBJECT_LOCKED(m->object),
1100 	    ("moea_enter_locked: page %p is not busy", m));
1101 
1102 	/* XXX change the pvo head for fake pages */
1103 	if ((m->flags & PG_FICTITIOUS) == PG_FICTITIOUS) {
1104 		pvo_flags &= ~PVO_MANAGED;
1105 		pvo_head = &moea_pvo_kunmanaged;
1106 		zone = moea_upvo_zone;
1107 	}
1108 
1109 	/*
1110 	 * If this is a managed page, and it's the first reference to the page,
1111 	 * clear the execness of the page.  Otherwise fetch the execness.
1112 	 */
1113 	if ((pg != NULL) && ((m->flags & PG_FICTITIOUS) == 0)) {
1114 		if (LIST_EMPTY(pvo_head)) {
1115 			moea_attr_clear(pg, PTE_EXEC);
1116 		} else {
1117 			was_exec = moea_attr_fetch(pg) & PTE_EXEC;
1118 		}
1119 	}
1120 
1121 	pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1122 
1123 	if (prot & VM_PROT_WRITE) {
1124 		pte_lo |= PTE_BW;
1125 		if (pmap_bootstrapped &&
1126 		    (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0)
1127 			vm_page_flag_set(m, PG_WRITEABLE);
1128 	} else
1129 		pte_lo |= PTE_BR;
1130 
1131 	if (prot & VM_PROT_EXECUTE)
1132 		pvo_flags |= PVO_EXECUTABLE;
1133 
1134 	if (wired)
1135 		pvo_flags |= PVO_WIRED;
1136 
1137 	if ((m->flags & PG_FICTITIOUS) != 0)
1138 		pvo_flags |= PVO_FAKE;
1139 
1140 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1141 	    pte_lo, pvo_flags);
1142 
1143 	/*
1144 	 * Flush the real page from the instruction cache if this page is
1145 	 * mapped executable and cacheable and was not previously mapped (or
1146 	 * was not mapped executable).
1147 	 */
1148 	if (error == 0 && (pvo_flags & PVO_EXECUTABLE) &&
1149 	    (pte_lo & PTE_I) == 0 && was_exec == 0) {
1150 		/*
1151 		 * Flush the real memory from the cache.
1152 		 */
1153 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1154 		if (pg != NULL)
1155 			moea_attr_save(pg, PTE_EXEC);
1156 	}
1157 
1158 	/* XXX syncicache always until problems are sorted */
1159 	moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1160 }
1161 
1162 /*
1163  * Maps a sequence of resident pages belonging to the same object.
1164  * The sequence begins with the given page m_start.  This page is
1165  * mapped at the given virtual address start.  Each subsequent page is
1166  * mapped at a virtual address that is offset from start by the same
1167  * amount as the page is offset from m_start within the object.  The
1168  * last page in the sequence is the page with the largest offset from
1169  * m_start that can be mapped at a virtual address less than the given
1170  * virtual address end.  Not every virtual page between start and end
1171  * is mapped; only those for which a resident page exists with the
1172  * corresponding offset from m_start are mapped.
1173  */
1174 void
1175 moea_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1176     vm_page_t m_start, vm_prot_t prot)
1177 {
1178 	vm_page_t m;
1179 	vm_pindex_t diff, psize;
1180 
1181 	psize = atop(end - start);
1182 	m = m_start;
1183 	vm_page_lock_queues();
1184 	PMAP_LOCK(pm);
1185 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1186 		moea_enter_locked(pm, start + ptoa(diff), m, prot &
1187 		    (VM_PROT_READ | VM_PROT_EXECUTE), FALSE);
1188 		m = TAILQ_NEXT(m, listq);
1189 	}
1190 	vm_page_unlock_queues();
1191 	PMAP_UNLOCK(pm);
1192 }
1193 
1194 void
1195 moea_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1196     vm_prot_t prot)
1197 {
1198 
1199 	vm_page_lock_queues();
1200 	PMAP_LOCK(pm);
1201 	moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1202 	    FALSE);
1203 	vm_page_unlock_queues();
1204 	PMAP_UNLOCK(pm);
1205 }
1206 
1207 vm_paddr_t
1208 moea_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1209 {
1210 	struct	pvo_entry *pvo;
1211 	vm_paddr_t pa;
1212 
1213 	PMAP_LOCK(pm);
1214 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1215 	if (pvo == NULL)
1216 		pa = 0;
1217 	else
1218 		pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1219 	PMAP_UNLOCK(pm);
1220 	return (pa);
1221 }
1222 
1223 /*
1224  * Atomically extract and hold the physical page with the given
1225  * pmap and virtual address pair if that mapping permits the given
1226  * protection.
1227  */
1228 vm_page_t
1229 moea_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1230 {
1231 	struct	pvo_entry *pvo;
1232 	vm_page_t m;
1233         vm_paddr_t pa;
1234 
1235 	m = NULL;
1236 	pa = 0;
1237 	PMAP_LOCK(pmap);
1238 retry:
1239 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1240 	if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1241 	    ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1242 	     (prot & VM_PROT_WRITE) == 0)) {
1243 		if (vm_page_pa_tryrelock(pmap, pvo->pvo_pte.pte.pte_lo & PTE_RPGN, &pa))
1244 			goto retry;
1245 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
1246 		vm_page_hold(m);
1247 	}
1248 	PA_UNLOCK_COND(pa);
1249 	PMAP_UNLOCK(pmap);
1250 	return (m);
1251 }
1252 
1253 void
1254 moea_init(mmu_t mmu)
1255 {
1256 
1257 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1258 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1259 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1260 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1261 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1262 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1263 	moea_initialized = TRUE;
1264 }
1265 
1266 boolean_t
1267 moea_is_referenced(mmu_t mmu, vm_page_t m)
1268 {
1269 
1270 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1271 	    ("moea_is_referenced: page %p is not managed", m));
1272 	return (moea_query_bit(m, PTE_REF));
1273 }
1274 
1275 boolean_t
1276 moea_is_modified(mmu_t mmu, vm_page_t m)
1277 {
1278 
1279 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1280 	    ("moea_is_modified: page %p is not managed", m));
1281 
1282 	/*
1283 	 * If the page is not VPO_BUSY, then PG_WRITEABLE cannot be
1284 	 * concurrently set while the object is locked.  Thus, if PG_WRITEABLE
1285 	 * is clear, no PTEs can have PTE_CHG set.
1286 	 */
1287 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1288 	if ((m->oflags & VPO_BUSY) == 0 &&
1289 	    (m->flags & PG_WRITEABLE) == 0)
1290 		return (FALSE);
1291 	return (moea_query_bit(m, PTE_CHG));
1292 }
1293 
1294 boolean_t
1295 moea_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1296 {
1297 	struct pvo_entry *pvo;
1298 	boolean_t rv;
1299 
1300 	PMAP_LOCK(pmap);
1301 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1302 	rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1303 	PMAP_UNLOCK(pmap);
1304 	return (rv);
1305 }
1306 
1307 void
1308 moea_clear_reference(mmu_t mmu, vm_page_t m)
1309 {
1310 
1311 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1312 	    ("moea_clear_reference: page %p is not managed", m));
1313 	moea_clear_bit(m, PTE_REF);
1314 }
1315 
1316 void
1317 moea_clear_modify(mmu_t mmu, vm_page_t m)
1318 {
1319 
1320 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1321 	    ("moea_clear_modify: page %p is not managed", m));
1322 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1323 	KASSERT((m->oflags & VPO_BUSY) == 0,
1324 	    ("moea_clear_modify: page %p is busy", m));
1325 
1326 	/*
1327 	 * If the page is not PG_WRITEABLE, then no PTEs can have PTE_CHG
1328 	 * set.  If the object containing the page is locked and the page is
1329 	 * not VPO_BUSY, then PG_WRITEABLE cannot be concurrently set.
1330 	 */
1331 	if ((m->flags & PG_WRITEABLE) == 0)
1332 		return;
1333 	moea_clear_bit(m, PTE_CHG);
1334 }
1335 
1336 /*
1337  * Clear the write and modified bits in each of the given page's mappings.
1338  */
1339 void
1340 moea_remove_write(mmu_t mmu, vm_page_t m)
1341 {
1342 	struct	pvo_entry *pvo;
1343 	struct	pte *pt;
1344 	pmap_t	pmap;
1345 	u_int	lo;
1346 
1347 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1348 	    ("moea_remove_write: page %p is not managed", m));
1349 
1350 	/*
1351 	 * If the page is not VPO_BUSY, then PG_WRITEABLE cannot be set by
1352 	 * another thread while the object is locked.  Thus, if PG_WRITEABLE
1353 	 * is clear, no page table entries need updating.
1354 	 */
1355 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1356 	if ((m->oflags & VPO_BUSY) == 0 &&
1357 	    (m->flags & PG_WRITEABLE) == 0)
1358 		return;
1359 	vm_page_lock_queues();
1360 	lo = moea_attr_fetch(m);
1361 	powerpc_sync();
1362 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1363 		pmap = pvo->pvo_pmap;
1364 		PMAP_LOCK(pmap);
1365 		if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1366 			pt = moea_pvo_to_pte(pvo, -1);
1367 			pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1368 			pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1369 			if (pt != NULL) {
1370 				moea_pte_synch(pt, &pvo->pvo_pte.pte);
1371 				lo |= pvo->pvo_pte.pte.pte_lo;
1372 				pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1373 				moea_pte_change(pt, &pvo->pvo_pte.pte,
1374 				    pvo->pvo_vaddr);
1375 				mtx_unlock(&moea_table_mutex);
1376 			}
1377 		}
1378 		PMAP_UNLOCK(pmap);
1379 	}
1380 	if ((lo & PTE_CHG) != 0) {
1381 		moea_attr_clear(m, PTE_CHG);
1382 		vm_page_dirty(m);
1383 	}
1384 	vm_page_flag_clear(m, PG_WRITEABLE);
1385 	vm_page_unlock_queues();
1386 }
1387 
1388 /*
1389  *	moea_ts_referenced:
1390  *
1391  *	Return a count of reference bits for a page, clearing those bits.
1392  *	It is not necessary for every reference bit to be cleared, but it
1393  *	is necessary that 0 only be returned when there are truly no
1394  *	reference bits set.
1395  *
1396  *	XXX: The exact number of bits to check and clear is a matter that
1397  *	should be tested and standardized at some point in the future for
1398  *	optimal aging of shared pages.
1399  */
1400 boolean_t
1401 moea_ts_referenced(mmu_t mmu, vm_page_t m)
1402 {
1403 
1404 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1405 	    ("moea_ts_referenced: page %p is not managed", m));
1406 	return (moea_clear_bit(m, PTE_REF));
1407 }
1408 
1409 /*
1410  * Modify the WIMG settings of all mappings for a page.
1411  */
1412 void
1413 moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1414 {
1415 	struct	pvo_entry *pvo;
1416 	struct	pvo_head *pvo_head;
1417 	struct	pte *pt;
1418 	pmap_t	pmap;
1419 	u_int	lo;
1420 
1421 	if (m->flags & PG_FICTITIOUS) {
1422 		m->md.mdpg_cache_attrs = ma;
1423 		return;
1424 	}
1425 
1426 	vm_page_lock_queues();
1427 	pvo_head = vm_page_to_pvoh(m);
1428 	lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1429 
1430 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1431 		pmap = pvo->pvo_pmap;
1432 		PMAP_LOCK(pmap);
1433 		pt = moea_pvo_to_pte(pvo, -1);
1434 		pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1435 		pvo->pvo_pte.pte.pte_lo |= lo;
1436 		if (pt != NULL) {
1437 			moea_pte_change(pt, &pvo->pvo_pte.pte,
1438 			    pvo->pvo_vaddr);
1439 			if (pvo->pvo_pmap == kernel_pmap)
1440 				isync();
1441 		}
1442 		mtx_unlock(&moea_table_mutex);
1443 		PMAP_UNLOCK(pmap);
1444 	}
1445 	m->md.mdpg_cache_attrs = ma;
1446 	vm_page_unlock_queues();
1447 }
1448 
1449 /*
1450  * Map a wired page into kernel virtual address space.
1451  */
1452 void
1453 moea_kenter(mmu_t mmu, vm_offset_t va, vm_offset_t pa)
1454 {
1455 
1456 	moea_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1457 }
1458 
1459 void
1460 moea_kenter_attr(mmu_t mmu, vm_offset_t va, vm_offset_t pa, vm_memattr_t ma)
1461 {
1462 	u_int		pte_lo;
1463 	int		error;
1464 
1465 #if 0
1466 	if (va < VM_MIN_KERNEL_ADDRESS)
1467 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1468 		    va);
1469 #endif
1470 
1471 	pte_lo = moea_calc_wimg(pa, ma);
1472 
1473 	PMAP_LOCK(kernel_pmap);
1474 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1475 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1476 
1477 	if (error != 0 && error != ENOENT)
1478 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1479 		    pa, error);
1480 
1481 	/*
1482 	 * Flush the real memory from the instruction cache.
1483 	 */
1484 	if ((pte_lo & (PTE_I | PTE_G)) == 0) {
1485 		moea_syncicache(pa, PAGE_SIZE);
1486 	}
1487 	PMAP_UNLOCK(kernel_pmap);
1488 }
1489 
1490 /*
1491  * Extract the physical page address associated with the given kernel virtual
1492  * address.
1493  */
1494 vm_offset_t
1495 moea_kextract(mmu_t mmu, vm_offset_t va)
1496 {
1497 	struct		pvo_entry *pvo;
1498 	vm_paddr_t pa;
1499 
1500 	/*
1501 	 * Allow direct mappings on 32-bit OEA
1502 	 */
1503 	if (va < VM_MIN_KERNEL_ADDRESS) {
1504 		return (va);
1505 	}
1506 
1507 	PMAP_LOCK(kernel_pmap);
1508 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1509 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1510 	pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1511 	PMAP_UNLOCK(kernel_pmap);
1512 	return (pa);
1513 }
1514 
1515 /*
1516  * Remove a wired page from kernel virtual address space.
1517  */
1518 void
1519 moea_kremove(mmu_t mmu, vm_offset_t va)
1520 {
1521 
1522 	moea_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1523 }
1524 
1525 /*
1526  * Map a range of physical addresses into kernel virtual address space.
1527  *
1528  * The value passed in *virt is a suggested virtual address for the mapping.
1529  * Architectures which can support a direct-mapped physical to virtual region
1530  * can return the appropriate address within that region, leaving '*virt'
1531  * unchanged.  We cannot and therefore do not; *virt is updated with the
1532  * first usable address after the mapped region.
1533  */
1534 vm_offset_t
1535 moea_map(mmu_t mmu, vm_offset_t *virt, vm_offset_t pa_start,
1536     vm_offset_t pa_end, int prot)
1537 {
1538 	vm_offset_t	sva, va;
1539 
1540 	sva = *virt;
1541 	va = sva;
1542 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1543 		moea_kenter(mmu, va, pa_start);
1544 	*virt = va;
1545 	return (sva);
1546 }
1547 
1548 /*
1549  * Returns true if the pmap's pv is one of the first
1550  * 16 pvs linked to from this page.  This count may
1551  * be changed upwards or downwards in the future; it
1552  * is only necessary that true be returned for a small
1553  * subset of pmaps for proper page aging.
1554  */
1555 boolean_t
1556 moea_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1557 {
1558         int loops;
1559 	struct pvo_entry *pvo;
1560 	boolean_t rv;
1561 
1562 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1563 	    ("moea_page_exists_quick: page %p is not managed", m));
1564 	loops = 0;
1565 	rv = FALSE;
1566 	vm_page_lock_queues();
1567 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1568 		if (pvo->pvo_pmap == pmap) {
1569 			rv = TRUE;
1570 			break;
1571 		}
1572 		if (++loops >= 16)
1573 			break;
1574 	}
1575 	vm_page_unlock_queues();
1576 	return (rv);
1577 }
1578 
1579 /*
1580  * Return the number of managed mappings to the given physical page
1581  * that are wired.
1582  */
1583 int
1584 moea_page_wired_mappings(mmu_t mmu, vm_page_t m)
1585 {
1586 	struct pvo_entry *pvo;
1587 	int count;
1588 
1589 	count = 0;
1590 	if ((m->flags & PG_FICTITIOUS) != 0)
1591 		return (count);
1592 	vm_page_lock_queues();
1593 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1594 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1595 			count++;
1596 	vm_page_unlock_queues();
1597 	return (count);
1598 }
1599 
1600 static u_int	moea_vsidcontext;
1601 
1602 void
1603 moea_pinit(mmu_t mmu, pmap_t pmap)
1604 {
1605 	int	i, mask;
1606 	u_int	entropy;
1607 
1608 	KASSERT((int)pmap < VM_MIN_KERNEL_ADDRESS, ("moea_pinit: virt pmap"));
1609 	PMAP_LOCK_INIT(pmap);
1610 
1611 	entropy = 0;
1612 	__asm __volatile("mftb %0" : "=r"(entropy));
1613 
1614 	if ((pmap->pmap_phys = (pmap_t)moea_kextract(mmu, (vm_offset_t)pmap))
1615 	    == NULL) {
1616 		pmap->pmap_phys = pmap;
1617 	}
1618 
1619 
1620 	mtx_lock(&moea_vsid_mutex);
1621 	/*
1622 	 * Allocate some segment registers for this pmap.
1623 	 */
1624 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1625 		u_int	hash, n;
1626 
1627 		/*
1628 		 * Create a new value by mutiplying by a prime and adding in
1629 		 * entropy from the timebase register.  This is to make the
1630 		 * VSID more random so that the PT hash function collides
1631 		 * less often.  (Note that the prime casues gcc to do shifts
1632 		 * instead of a multiply.)
1633 		 */
1634 		moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1635 		hash = moea_vsidcontext & (NPMAPS - 1);
1636 		if (hash == 0)		/* 0 is special, avoid it */
1637 			continue;
1638 		n = hash >> 5;
1639 		mask = 1 << (hash & (VSID_NBPW - 1));
1640 		hash = (moea_vsidcontext & 0xfffff);
1641 		if (moea_vsid_bitmap[n] & mask) {	/* collision? */
1642 			/* anything free in this bucket? */
1643 			if (moea_vsid_bitmap[n] == 0xffffffff) {
1644 				entropy = (moea_vsidcontext >> 20);
1645 				continue;
1646 			}
1647 			i = ffs(~moea_vsid_bitmap[n]) - 1;
1648 			mask = 1 << i;
1649 			hash &= 0xfffff & ~(VSID_NBPW - 1);
1650 			hash |= i;
1651 		}
1652 		moea_vsid_bitmap[n] |= mask;
1653 		for (i = 0; i < 16; i++)
1654 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1655 		mtx_unlock(&moea_vsid_mutex);
1656 		return;
1657 	}
1658 
1659 	mtx_unlock(&moea_vsid_mutex);
1660 	panic("moea_pinit: out of segments");
1661 }
1662 
1663 /*
1664  * Initialize the pmap associated with process 0.
1665  */
1666 void
1667 moea_pinit0(mmu_t mmu, pmap_t pm)
1668 {
1669 
1670 	moea_pinit(mmu, pm);
1671 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1672 }
1673 
1674 /*
1675  * Set the physical protection on the specified range of this map as requested.
1676  */
1677 void
1678 moea_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1679     vm_prot_t prot)
1680 {
1681 	struct	pvo_entry *pvo;
1682 	struct	pte *pt;
1683 	int	pteidx;
1684 
1685 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1686 	    ("moea_protect: non current pmap"));
1687 
1688 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1689 		moea_remove(mmu, pm, sva, eva);
1690 		return;
1691 	}
1692 
1693 	vm_page_lock_queues();
1694 	PMAP_LOCK(pm);
1695 	for (; sva < eva; sva += PAGE_SIZE) {
1696 		pvo = moea_pvo_find_va(pm, sva, &pteidx);
1697 		if (pvo == NULL)
1698 			continue;
1699 
1700 		if ((prot & VM_PROT_EXECUTE) == 0)
1701 			pvo->pvo_vaddr &= ~PVO_EXECUTABLE;
1702 
1703 		/*
1704 		 * Grab the PTE pointer before we diddle with the cached PTE
1705 		 * copy.
1706 		 */
1707 		pt = moea_pvo_to_pte(pvo, pteidx);
1708 		/*
1709 		 * Change the protection of the page.
1710 		 */
1711 		pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1712 		pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1713 
1714 		/*
1715 		 * If the PVO is in the page table, update that pte as well.
1716 		 */
1717 		if (pt != NULL) {
1718 			moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1719 			mtx_unlock(&moea_table_mutex);
1720 		}
1721 	}
1722 	vm_page_unlock_queues();
1723 	PMAP_UNLOCK(pm);
1724 }
1725 
1726 /*
1727  * Map a list of wired pages into kernel virtual address space.  This is
1728  * intended for temporary mappings which do not need page modification or
1729  * references recorded.  Existing mappings in the region are overwritten.
1730  */
1731 void
1732 moea_qenter(mmu_t mmu, vm_offset_t sva, vm_page_t *m, int count)
1733 {
1734 	vm_offset_t va;
1735 
1736 	va = sva;
1737 	while (count-- > 0) {
1738 		moea_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1739 		va += PAGE_SIZE;
1740 		m++;
1741 	}
1742 }
1743 
1744 /*
1745  * Remove page mappings from kernel virtual address space.  Intended for
1746  * temporary mappings entered by moea_qenter.
1747  */
1748 void
1749 moea_qremove(mmu_t mmu, vm_offset_t sva, int count)
1750 {
1751 	vm_offset_t va;
1752 
1753 	va = sva;
1754 	while (count-- > 0) {
1755 		moea_kremove(mmu, va);
1756 		va += PAGE_SIZE;
1757 	}
1758 }
1759 
1760 void
1761 moea_release(mmu_t mmu, pmap_t pmap)
1762 {
1763         int idx, mask;
1764 
1765 	/*
1766 	 * Free segment register's VSID
1767 	 */
1768         if (pmap->pm_sr[0] == 0)
1769                 panic("moea_release");
1770 
1771 	mtx_lock(&moea_vsid_mutex);
1772         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1773         mask = 1 << (idx % VSID_NBPW);
1774         idx /= VSID_NBPW;
1775         moea_vsid_bitmap[idx] &= ~mask;
1776 	mtx_unlock(&moea_vsid_mutex);
1777 	PMAP_LOCK_DESTROY(pmap);
1778 }
1779 
1780 /*
1781  * Remove the given range of addresses from the specified map.
1782  */
1783 void
1784 moea_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1785 {
1786 	struct	pvo_entry *pvo;
1787 	int	pteidx;
1788 
1789 	vm_page_lock_queues();
1790 	PMAP_LOCK(pm);
1791 	for (; sva < eva; sva += PAGE_SIZE) {
1792 		pvo = moea_pvo_find_va(pm, sva, &pteidx);
1793 		if (pvo != NULL) {
1794 			moea_pvo_remove(pvo, pteidx);
1795 		}
1796 	}
1797 	PMAP_UNLOCK(pm);
1798 	vm_page_unlock_queues();
1799 }
1800 
1801 /*
1802  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1803  * will reflect changes in pte's back to the vm_page.
1804  */
1805 void
1806 moea_remove_all(mmu_t mmu, vm_page_t m)
1807 {
1808 	struct  pvo_head *pvo_head;
1809 	struct	pvo_entry *pvo, *next_pvo;
1810 	pmap_t	pmap;
1811 
1812 	vm_page_lock_queues();
1813 	pvo_head = vm_page_to_pvoh(m);
1814 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1815 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1816 
1817 		pmap = pvo->pvo_pmap;
1818 		PMAP_LOCK(pmap);
1819 		moea_pvo_remove(pvo, -1);
1820 		PMAP_UNLOCK(pmap);
1821 	}
1822 	if ((m->flags & PG_WRITEABLE) && moea_is_modified(mmu, m)) {
1823 		moea_attr_clear(m, PTE_CHG);
1824 		vm_page_dirty(m);
1825 	}
1826 	vm_page_flag_clear(m, PG_WRITEABLE);
1827 	vm_page_unlock_queues();
1828 }
1829 
1830 /*
1831  * Allocate a physical page of memory directly from the phys_avail map.
1832  * Can only be called from moea_bootstrap before avail start and end are
1833  * calculated.
1834  */
1835 static vm_offset_t
1836 moea_bootstrap_alloc(vm_size_t size, u_int align)
1837 {
1838 	vm_offset_t	s, e;
1839 	int		i, j;
1840 
1841 	size = round_page(size);
1842 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1843 		if (align != 0)
1844 			s = (phys_avail[i] + align - 1) & ~(align - 1);
1845 		else
1846 			s = phys_avail[i];
1847 		e = s + size;
1848 
1849 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1850 			continue;
1851 
1852 		if (s == phys_avail[i]) {
1853 			phys_avail[i] += size;
1854 		} else if (e == phys_avail[i + 1]) {
1855 			phys_avail[i + 1] -= size;
1856 		} else {
1857 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1858 				phys_avail[j] = phys_avail[j - 2];
1859 				phys_avail[j + 1] = phys_avail[j - 1];
1860 			}
1861 
1862 			phys_avail[i + 3] = phys_avail[i + 1];
1863 			phys_avail[i + 1] = s;
1864 			phys_avail[i + 2] = e;
1865 			phys_avail_count++;
1866 		}
1867 
1868 		return (s);
1869 	}
1870 	panic("moea_bootstrap_alloc: could not allocate memory");
1871 }
1872 
1873 static void
1874 moea_syncicache(vm_offset_t pa, vm_size_t len)
1875 {
1876 	__syncicache((void *)pa, len);
1877 }
1878 
1879 static int
1880 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1881     vm_offset_t va, vm_offset_t pa, u_int pte_lo, int flags)
1882 {
1883 	struct	pvo_entry *pvo;
1884 	u_int	sr;
1885 	int	first;
1886 	u_int	ptegidx;
1887 	int	i;
1888 	int     bootstrap;
1889 
1890 	moea_pvo_enter_calls++;
1891 	first = 0;
1892 	bootstrap = 0;
1893 
1894 	/*
1895 	 * Compute the PTE Group index.
1896 	 */
1897 	va &= ~ADDR_POFF;
1898 	sr = va_to_sr(pm->pm_sr, va);
1899 	ptegidx = va_to_pteg(sr, va);
1900 
1901 	/*
1902 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
1903 	 * there is a mapping.
1904 	 */
1905 	mtx_lock(&moea_table_mutex);
1906 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
1907 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
1908 			if ((pvo->pvo_pte.pte.pte_lo & PTE_RPGN) == pa &&
1909 			    (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
1910 			    (pte_lo & PTE_PP)) {
1911 				mtx_unlock(&moea_table_mutex);
1912 				return (0);
1913 			}
1914 			moea_pvo_remove(pvo, -1);
1915 			break;
1916 		}
1917 	}
1918 
1919 	/*
1920 	 * If we aren't overwriting a mapping, try to allocate.
1921 	 */
1922 	if (moea_initialized) {
1923 		pvo = uma_zalloc(zone, M_NOWAIT);
1924 	} else {
1925 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
1926 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
1927 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
1928 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
1929 		}
1930 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
1931 		moea_bpvo_pool_index++;
1932 		bootstrap = 1;
1933 	}
1934 
1935 	if (pvo == NULL) {
1936 		mtx_unlock(&moea_table_mutex);
1937 		return (ENOMEM);
1938 	}
1939 
1940 	moea_pvo_entries++;
1941 	pvo->pvo_vaddr = va;
1942 	pvo->pvo_pmap = pm;
1943 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
1944 	pvo->pvo_vaddr &= ~ADDR_POFF;
1945 	if (flags & VM_PROT_EXECUTE)
1946 		pvo->pvo_vaddr |= PVO_EXECUTABLE;
1947 	if (flags & PVO_WIRED)
1948 		pvo->pvo_vaddr |= PVO_WIRED;
1949 	if (pvo_head != &moea_pvo_kunmanaged)
1950 		pvo->pvo_vaddr |= PVO_MANAGED;
1951 	if (bootstrap)
1952 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
1953 	if (flags & PVO_FAKE)
1954 		pvo->pvo_vaddr |= PVO_FAKE;
1955 
1956 	moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo);
1957 
1958 	/*
1959 	 * Remember if the list was empty and therefore will be the first
1960 	 * item.
1961 	 */
1962 	if (LIST_FIRST(pvo_head) == NULL)
1963 		first = 1;
1964 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
1965 
1966 	if (pvo->pvo_pte.pte.pte_lo & PVO_WIRED)
1967 		pm->pm_stats.wired_count++;
1968 	pm->pm_stats.resident_count++;
1969 
1970 	/*
1971 	 * We hope this succeeds but it isn't required.
1972 	 */
1973 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
1974 	if (i >= 0) {
1975 		PVO_PTEGIDX_SET(pvo, i);
1976 	} else {
1977 		panic("moea_pvo_enter: overflow");
1978 		moea_pte_overflow++;
1979 	}
1980 	mtx_unlock(&moea_table_mutex);
1981 
1982 	return (first ? ENOENT : 0);
1983 }
1984 
1985 static void
1986 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
1987 {
1988 	struct	pte *pt;
1989 
1990 	/*
1991 	 * If there is an active pte entry, we need to deactivate it (and
1992 	 * save the ref & cfg bits).
1993 	 */
1994 	pt = moea_pvo_to_pte(pvo, pteidx);
1995 	if (pt != NULL) {
1996 		moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1997 		mtx_unlock(&moea_table_mutex);
1998 		PVO_PTEGIDX_CLR(pvo);
1999 	} else {
2000 		moea_pte_overflow--;
2001 	}
2002 
2003 	/*
2004 	 * Update our statistics.
2005 	 */
2006 	pvo->pvo_pmap->pm_stats.resident_count--;
2007 	if (pvo->pvo_pte.pte.pte_lo & PVO_WIRED)
2008 		pvo->pvo_pmap->pm_stats.wired_count--;
2009 
2010 	/*
2011 	 * Save the REF/CHG bits into their cache if the page is managed.
2012 	 */
2013 	if ((pvo->pvo_vaddr & (PVO_MANAGED|PVO_FAKE)) == PVO_MANAGED) {
2014 		struct	vm_page *pg;
2015 
2016 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
2017 		if (pg != NULL) {
2018 			moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2019 			    (PTE_REF | PTE_CHG));
2020 		}
2021 	}
2022 
2023 	/*
2024 	 * Remove this PVO from the PV list.
2025 	 */
2026 	LIST_REMOVE(pvo, pvo_vlink);
2027 
2028 	/*
2029 	 * Remove this from the overflow list and return it to the pool
2030 	 * if we aren't going to reuse it.
2031 	 */
2032 	LIST_REMOVE(pvo, pvo_olink);
2033 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2034 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
2035 		    moea_upvo_zone, pvo);
2036 	moea_pvo_entries--;
2037 	moea_pvo_remove_calls++;
2038 }
2039 
2040 static __inline int
2041 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2042 {
2043 	int	pteidx;
2044 
2045 	/*
2046 	 * We can find the actual pte entry without searching by grabbing
2047 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2048 	 * noticing the HID bit.
2049 	 */
2050 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2051 	if (pvo->pvo_pte.pte.pte_hi & PTE_HID)
2052 		pteidx ^= moea_pteg_mask * 8;
2053 
2054 	return (pteidx);
2055 }
2056 
2057 static struct pvo_entry *
2058 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2059 {
2060 	struct	pvo_entry *pvo;
2061 	int	ptegidx;
2062 	u_int	sr;
2063 
2064 	va &= ~ADDR_POFF;
2065 	sr = va_to_sr(pm->pm_sr, va);
2066 	ptegidx = va_to_pteg(sr, va);
2067 
2068 	mtx_lock(&moea_table_mutex);
2069 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2070 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2071 			if (pteidx_p)
2072 				*pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2073 			break;
2074 		}
2075 	}
2076 	mtx_unlock(&moea_table_mutex);
2077 
2078 	return (pvo);
2079 }
2080 
2081 static struct pte *
2082 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2083 {
2084 	struct	pte *pt;
2085 
2086 	/*
2087 	 * If we haven't been supplied the ptegidx, calculate it.
2088 	 */
2089 	if (pteidx == -1) {
2090 		int	ptegidx;
2091 		u_int	sr;
2092 
2093 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2094 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2095 		pteidx = moea_pvo_pte_index(pvo, ptegidx);
2096 	}
2097 
2098 	pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2099 	mtx_lock(&moea_table_mutex);
2100 
2101 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2102 		panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2103 		    "valid pte index", pvo);
2104 	}
2105 
2106 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2107 		panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2108 		    "pvo but no valid pte", pvo);
2109 	}
2110 
2111 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2112 		if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) {
2113 			panic("moea_pvo_to_pte: pvo %p has valid pte in "
2114 			    "moea_pteg_table %p but invalid in pvo", pvo, pt);
2115 		}
2116 
2117 		if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2118 		    != 0) {
2119 			panic("moea_pvo_to_pte: pvo %p pte does not match "
2120 			    "pte %p in moea_pteg_table", pvo, pt);
2121 		}
2122 
2123 		mtx_assert(&moea_table_mutex, MA_OWNED);
2124 		return (pt);
2125 	}
2126 
2127 	if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) {
2128 		panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2129 		    "moea_pteg_table but valid in pvo", pvo, pt);
2130 	}
2131 
2132 	mtx_unlock(&moea_table_mutex);
2133 	return (NULL);
2134 }
2135 
2136 /*
2137  * XXX: THIS STUFF SHOULD BE IN pte.c?
2138  */
2139 int
2140 moea_pte_spill(vm_offset_t addr)
2141 {
2142 	struct	pvo_entry *source_pvo, *victim_pvo;
2143 	struct	pvo_entry *pvo;
2144 	int	ptegidx, i, j;
2145 	u_int	sr;
2146 	struct	pteg *pteg;
2147 	struct	pte *pt;
2148 
2149 	moea_pte_spills++;
2150 
2151 	sr = mfsrin(addr);
2152 	ptegidx = va_to_pteg(sr, addr);
2153 
2154 	/*
2155 	 * Have to substitute some entry.  Use the primary hash for this.
2156 	 * Use low bits of timebase as random generator.
2157 	 */
2158 	pteg = &moea_pteg_table[ptegidx];
2159 	mtx_lock(&moea_table_mutex);
2160 	__asm __volatile("mftb %0" : "=r"(i));
2161 	i &= 7;
2162 	pt = &pteg->pt[i];
2163 
2164 	source_pvo = NULL;
2165 	victim_pvo = NULL;
2166 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2167 		/*
2168 		 * We need to find a pvo entry for this address.
2169 		 */
2170 		if (source_pvo == NULL &&
2171 		    moea_pte_match(&pvo->pvo_pte.pte, sr, addr,
2172 		    pvo->pvo_pte.pte.pte_hi & PTE_HID)) {
2173 			/*
2174 			 * Now found an entry to be spilled into the pteg.
2175 			 * The PTE is now valid, so we know it's active.
2176 			 */
2177 			j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2178 
2179 			if (j >= 0) {
2180 				PVO_PTEGIDX_SET(pvo, j);
2181 				moea_pte_overflow--;
2182 				mtx_unlock(&moea_table_mutex);
2183 				return (1);
2184 			}
2185 
2186 			source_pvo = pvo;
2187 
2188 			if (victim_pvo != NULL)
2189 				break;
2190 		}
2191 
2192 		/*
2193 		 * We also need the pvo entry of the victim we are replacing
2194 		 * so save the R & C bits of the PTE.
2195 		 */
2196 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2197 		    moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2198 			victim_pvo = pvo;
2199 			if (source_pvo != NULL)
2200 				break;
2201 		}
2202 	}
2203 
2204 	if (source_pvo == NULL) {
2205 		mtx_unlock(&moea_table_mutex);
2206 		return (0);
2207 	}
2208 
2209 	if (victim_pvo == NULL) {
2210 		if ((pt->pte_hi & PTE_HID) == 0)
2211 			panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2212 			    "entry", pt);
2213 
2214 		/*
2215 		 * If this is a secondary PTE, we need to search it's primary
2216 		 * pvo bucket for the matching PVO.
2217 		 */
2218 		LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2219 		    pvo_olink) {
2220 			/*
2221 			 * We also need the pvo entry of the victim we are
2222 			 * replacing so save the R & C bits of the PTE.
2223 			 */
2224 			if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2225 				victim_pvo = pvo;
2226 				break;
2227 			}
2228 		}
2229 
2230 		if (victim_pvo == NULL)
2231 			panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2232 			    "entry", pt);
2233 	}
2234 
2235 	/*
2236 	 * We are invalidating the TLB entry for the EA we are replacing even
2237 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2238 	 * contained in the TLB entry.
2239 	 */
2240 	source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID;
2241 
2242 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2243 	moea_pte_set(pt, &source_pvo->pvo_pte.pte);
2244 
2245 	PVO_PTEGIDX_CLR(victim_pvo);
2246 	PVO_PTEGIDX_SET(source_pvo, i);
2247 	moea_pte_replacements++;
2248 
2249 	mtx_unlock(&moea_table_mutex);
2250 	return (1);
2251 }
2252 
2253 static int
2254 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2255 {
2256 	struct	pte *pt;
2257 	int	i;
2258 
2259 	mtx_assert(&moea_table_mutex, MA_OWNED);
2260 
2261 	/*
2262 	 * First try primary hash.
2263 	 */
2264 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2265 		if ((pt->pte_hi & PTE_VALID) == 0) {
2266 			pvo_pt->pte_hi &= ~PTE_HID;
2267 			moea_pte_set(pt, pvo_pt);
2268 			return (i);
2269 		}
2270 	}
2271 
2272 	/*
2273 	 * Now try secondary hash.
2274 	 */
2275 	ptegidx ^= moea_pteg_mask;
2276 
2277 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2278 		if ((pt->pte_hi & PTE_VALID) == 0) {
2279 			pvo_pt->pte_hi |= PTE_HID;
2280 			moea_pte_set(pt, pvo_pt);
2281 			return (i);
2282 		}
2283 	}
2284 
2285 	panic("moea_pte_insert: overflow");
2286 	return (-1);
2287 }
2288 
2289 static boolean_t
2290 moea_query_bit(vm_page_t m, int ptebit)
2291 {
2292 	struct	pvo_entry *pvo;
2293 	struct	pte *pt;
2294 
2295 	if (moea_attr_fetch(m) & ptebit)
2296 		return (TRUE);
2297 
2298 	vm_page_lock_queues();
2299 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2300 
2301 		/*
2302 		 * See if we saved the bit off.  If so, cache it and return
2303 		 * success.
2304 		 */
2305 		if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2306 			moea_attr_save(m, ptebit);
2307 			vm_page_unlock_queues();
2308 			return (TRUE);
2309 		}
2310 	}
2311 
2312 	/*
2313 	 * No luck, now go through the hard part of looking at the PTEs
2314 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2315 	 * the PTEs.
2316 	 */
2317 	powerpc_sync();
2318 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2319 
2320 		/*
2321 		 * See if this pvo has a valid PTE.  if so, fetch the
2322 		 * REF/CHG bits from the valid PTE.  If the appropriate
2323 		 * ptebit is set, cache it and return success.
2324 		 */
2325 		pt = moea_pvo_to_pte(pvo, -1);
2326 		if (pt != NULL) {
2327 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2328 			mtx_unlock(&moea_table_mutex);
2329 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2330 				moea_attr_save(m, ptebit);
2331 				vm_page_unlock_queues();
2332 				return (TRUE);
2333 			}
2334 		}
2335 	}
2336 
2337 	vm_page_unlock_queues();
2338 	return (FALSE);
2339 }
2340 
2341 static u_int
2342 moea_clear_bit(vm_page_t m, int ptebit)
2343 {
2344 	u_int	count;
2345 	struct	pvo_entry *pvo;
2346 	struct	pte *pt;
2347 
2348 	vm_page_lock_queues();
2349 
2350 	/*
2351 	 * Clear the cached value.
2352 	 */
2353 	moea_attr_clear(m, ptebit);
2354 
2355 	/*
2356 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2357 	 * we can reset the right ones).  note that since the pvo entries and
2358 	 * list heads are accessed via BAT0 and are never placed in the page
2359 	 * table, we don't have to worry about further accesses setting the
2360 	 * REF/CHG bits.
2361 	 */
2362 	powerpc_sync();
2363 
2364 	/*
2365 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2366 	 * valid pte clear the ptebit from the valid pte.
2367 	 */
2368 	count = 0;
2369 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2370 		pt = moea_pvo_to_pte(pvo, -1);
2371 		if (pt != NULL) {
2372 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2373 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2374 				count++;
2375 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2376 			}
2377 			mtx_unlock(&moea_table_mutex);
2378 		}
2379 		pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2380 	}
2381 
2382 	vm_page_unlock_queues();
2383 	return (count);
2384 }
2385 
2386 /*
2387  * Return true if the physical range is encompassed by the battable[idx]
2388  */
2389 static int
2390 moea_bat_mapped(int idx, vm_offset_t pa, vm_size_t size)
2391 {
2392 	u_int prot;
2393 	u_int32_t start;
2394 	u_int32_t end;
2395 	u_int32_t bat_ble;
2396 
2397 	/*
2398 	 * Return immediately if not a valid mapping
2399 	 */
2400 	if (!(battable[idx].batu & BAT_Vs))
2401 		return (EINVAL);
2402 
2403 	/*
2404 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2405 	 * so it can function as an i/o page
2406 	 */
2407 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2408 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2409 		return (EPERM);
2410 
2411 	/*
2412 	 * The address should be within the BAT range. Assume that the
2413 	 * start address in the BAT has the correct alignment (thus
2414 	 * not requiring masking)
2415 	 */
2416 	start = battable[idx].batl & BAT_PBS;
2417 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2418 	end = start | (bat_ble << 15) | 0x7fff;
2419 
2420 	if ((pa < start) || ((pa + size) > end))
2421 		return (ERANGE);
2422 
2423 	return (0);
2424 }
2425 
2426 boolean_t
2427 moea_dev_direct_mapped(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2428 {
2429 	int i;
2430 
2431 	/*
2432 	 * This currently does not work for entries that
2433 	 * overlap 256M BAT segments.
2434 	 */
2435 
2436 	for(i = 0; i < 16; i++)
2437 		if (moea_bat_mapped(i, pa, size) == 0)
2438 			return (0);
2439 
2440 	return (EFAULT);
2441 }
2442 
2443 /*
2444  * Map a set of physical memory pages into the kernel virtual
2445  * address space. Return a pointer to where it is mapped. This
2446  * routine is intended to be used for mapping device memory,
2447  * NOT real memory.
2448  */
2449 void *
2450 moea_mapdev(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2451 {
2452 
2453 	return (moea_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT));
2454 }
2455 
2456 void *
2457 moea_mapdev_attr(mmu_t mmu, vm_offset_t pa, vm_size_t size, vm_memattr_t ma)
2458 {
2459 	vm_offset_t va, tmpva, ppa, offset;
2460 	int i;
2461 
2462 	ppa = trunc_page(pa);
2463 	offset = pa & PAGE_MASK;
2464 	size = roundup(offset + size, PAGE_SIZE);
2465 
2466 	/*
2467 	 * If the physical address lies within a valid BAT table entry,
2468 	 * return the 1:1 mapping. This currently doesn't work
2469 	 * for regions that overlap 256M BAT segments.
2470 	 */
2471 	for (i = 0; i < 16; i++) {
2472 		if (moea_bat_mapped(i, pa, size) == 0)
2473 			return ((void *) pa);
2474 	}
2475 
2476 	va = kmem_alloc_nofault(kernel_map, size);
2477 	if (!va)
2478 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2479 
2480 	for (tmpva = va; size > 0;) {
2481 		moea_kenter_attr(mmu, tmpva, ppa, ma);
2482 		tlbie(tmpva);
2483 		size -= PAGE_SIZE;
2484 		tmpva += PAGE_SIZE;
2485 		ppa += PAGE_SIZE;
2486 	}
2487 
2488 	return ((void *)(va + offset));
2489 }
2490 
2491 void
2492 moea_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2493 {
2494 	vm_offset_t base, offset;
2495 
2496 	/*
2497 	 * If this is outside kernel virtual space, then it's a
2498 	 * battable entry and doesn't require unmapping
2499 	 */
2500 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2501 		base = trunc_page(va);
2502 		offset = va & PAGE_MASK;
2503 		size = roundup(offset + size, PAGE_SIZE);
2504 		kmem_free(kernel_map, base, size);
2505 	}
2506 }
2507 
2508 static void
2509 moea_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2510 {
2511 	struct pvo_entry *pvo;
2512 	vm_offset_t lim;
2513 	vm_paddr_t pa;
2514 	vm_size_t len;
2515 
2516 	PMAP_LOCK(pm);
2517 	while (sz > 0) {
2518 		lim = round_page(va);
2519 		len = MIN(lim - va, sz);
2520 		pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2521 		if (pvo != NULL) {
2522 			pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) |
2523 			    (va & ADDR_POFF);
2524 			moea_syncicache(pa, len);
2525 		}
2526 		va += len;
2527 		sz -= len;
2528 	}
2529 	PMAP_UNLOCK(pm);
2530 }
2531