xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision 38effe887ee979f91ad5abf42a2291558e7ff8d1)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD AND BSD-4-Clause
3  *
4  * Copyright (c) 2001 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Matt Thomas <matt@3am-software.com> of Allegro Networks, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 /*-
32  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
33  * Copyright (C) 1995, 1996 TooLs GmbH.
34  * All rights reserved.
35  *
36  * Redistribution and use in source and binary forms, with or without
37  * modification, are permitted provided that the following conditions
38  * are met:
39  * 1. Redistributions of source code must retain the above copyright
40  *    notice, this list of conditions and the following disclaimer.
41  * 2. Redistributions in binary form must reproduce the above copyright
42  *    notice, this list of conditions and the following disclaimer in the
43  *    documentation and/or other materials provided with the distribution.
44  * 3. All advertising materials mentioning features or use of this software
45  *    must display the following acknowledgement:
46  *	This product includes software developed by TooLs GmbH.
47  * 4. The name of TooLs GmbH may not be used to endorse or promote products
48  *    derived from this software without specific prior written permission.
49  *
50  * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
51  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
52  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
53  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
54  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
55  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
56  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
57  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
58  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
59  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
60  *
61  * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $
62  */
63 /*-
64  * Copyright (C) 2001 Benno Rice.
65  * All rights reserved.
66  *
67  * Redistribution and use in source and binary forms, with or without
68  * modification, are permitted provided that the following conditions
69  * are met:
70  * 1. Redistributions of source code must retain the above copyright
71  *    notice, this list of conditions and the following disclaimer.
72  * 2. Redistributions in binary form must reproduce the above copyright
73  *    notice, this list of conditions and the following disclaimer in the
74  *    documentation and/or other materials provided with the distribution.
75  *
76  * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR
77  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
78  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
79  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
80  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
81  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
82  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
83  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
84  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
85  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
86  */
87 
88 #include <sys/cdefs.h>
89 __FBSDID("$FreeBSD$");
90 
91 /*
92  * Manages physical address maps.
93  *
94  * Since the information managed by this module is also stored by the
95  * logical address mapping module, this module may throw away valid virtual
96  * to physical mappings at almost any time.  However, invalidations of
97  * mappings must be done as requested.
98  *
99  * In order to cope with hardware architectures which make virtual to
100  * physical map invalidates expensive, this module may delay invalidate
101  * reduced protection operations until such time as they are actually
102  * necessary.  This module is given full information as to which processors
103  * are currently using which maps, and to when physical maps must be made
104  * correct.
105  */
106 
107 #include "opt_kstack_pages.h"
108 
109 #include <sys/param.h>
110 #include <sys/kernel.h>
111 #include <sys/conf.h>
112 #include <sys/queue.h>
113 #include <sys/cpuset.h>
114 #include <sys/kerneldump.h>
115 #include <sys/ktr.h>
116 #include <sys/lock.h>
117 #include <sys/msgbuf.h>
118 #include <sys/mutex.h>
119 #include <sys/proc.h>
120 #include <sys/rwlock.h>
121 #include <sys/sched.h>
122 #include <sys/sysctl.h>
123 #include <sys/systm.h>
124 #include <sys/vmmeter.h>
125 
126 #include <dev/ofw/openfirm.h>
127 
128 #include <vm/vm.h>
129 #include <vm/vm_param.h>
130 #include <vm/vm_kern.h>
131 #include <vm/vm_page.h>
132 #include <vm/vm_map.h>
133 #include <vm/vm_object.h>
134 #include <vm/vm_extern.h>
135 #include <vm/vm_page.h>
136 #include <vm/vm_phys.h>
137 #include <vm/vm_pageout.h>
138 #include <vm/uma.h>
139 
140 #include <machine/cpu.h>
141 #include <machine/platform.h>
142 #include <machine/bat.h>
143 #include <machine/frame.h>
144 #include <machine/md_var.h>
145 #include <machine/psl.h>
146 #include <machine/pte.h>
147 #include <machine/smp.h>
148 #include <machine/sr.h>
149 #include <machine/mmuvar.h>
150 #include <machine/trap.h>
151 
152 #include "mmu_if.h"
153 
154 #define	MOEA_DEBUG
155 
156 #define TODO	panic("%s: not implemented", __func__);
157 
158 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
159 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
160 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
161 
162 struct ofw_map {
163 	vm_offset_t	om_va;
164 	vm_size_t	om_len;
165 	vm_offset_t	om_pa;
166 	u_int		om_mode;
167 };
168 
169 extern unsigned char _etext[];
170 extern unsigned char _end[];
171 
172 /*
173  * Map of physical memory regions.
174  */
175 static struct	mem_region *regions;
176 static struct	mem_region *pregions;
177 static u_int    phys_avail_count;
178 static int	regions_sz, pregions_sz;
179 static struct	ofw_map *translations;
180 
181 /*
182  * Lock for the pteg and pvo tables.
183  */
184 struct mtx	moea_table_mutex;
185 struct mtx	moea_vsid_mutex;
186 
187 /* tlbie instruction synchronization */
188 static struct mtx tlbie_mtx;
189 
190 /*
191  * PTEG data.
192  */
193 static struct	pteg *moea_pteg_table;
194 u_int		moea_pteg_count;
195 u_int		moea_pteg_mask;
196 
197 /*
198  * PVO data.
199  */
200 struct	pvo_head *moea_pvo_table;		/* pvo entries by pteg index */
201 struct	pvo_head moea_pvo_kunmanaged =
202     LIST_HEAD_INITIALIZER(moea_pvo_kunmanaged);	/* list of unmanaged pages */
203 
204 static struct rwlock_padalign pvh_global_lock;
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_paddr_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 int		moea_enter_locked(pmap_t, vm_offset_t, vm_page_t,
266 			    vm_prot_t, u_int, int8_t);
267 static void		moea_syncicache(vm_paddr_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_clear_modify(mmu_t, vm_page_t);
277 void moea_copy_page(mmu_t, vm_page_t, vm_page_t);
278 void moea_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
279     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
280 int moea_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int,
281     int8_t);
282 void moea_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
283     vm_prot_t);
284 void moea_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
285 vm_paddr_t moea_extract(mmu_t, pmap_t, vm_offset_t);
286 vm_page_t moea_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
287 void moea_init(mmu_t);
288 boolean_t moea_is_modified(mmu_t, vm_page_t);
289 boolean_t moea_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
290 boolean_t moea_is_referenced(mmu_t, vm_page_t);
291 int moea_ts_referenced(mmu_t, vm_page_t);
292 vm_offset_t moea_map(mmu_t, vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
293 boolean_t moea_page_exists_quick(mmu_t, pmap_t, vm_page_t);
294 void moea_page_init(mmu_t, vm_page_t);
295 int moea_page_wired_mappings(mmu_t, vm_page_t);
296 void moea_pinit(mmu_t, pmap_t);
297 void moea_pinit0(mmu_t, pmap_t);
298 void moea_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
299 void moea_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
300 void moea_qremove(mmu_t, vm_offset_t, int);
301 void moea_release(mmu_t, pmap_t);
302 void moea_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
303 void moea_remove_all(mmu_t, vm_page_t);
304 void moea_remove_write(mmu_t, vm_page_t);
305 void moea_unwire(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
306 void moea_zero_page(mmu_t, vm_page_t);
307 void moea_zero_page_area(mmu_t, vm_page_t, int, int);
308 void moea_activate(mmu_t, struct thread *);
309 void moea_deactivate(mmu_t, struct thread *);
310 void moea_cpu_bootstrap(mmu_t, int);
311 void moea_bootstrap(mmu_t, vm_offset_t, vm_offset_t);
312 void *moea_mapdev(mmu_t, vm_paddr_t, vm_size_t);
313 void *moea_mapdev_attr(mmu_t, vm_paddr_t, vm_size_t, vm_memattr_t);
314 void moea_unmapdev(mmu_t, vm_offset_t, vm_size_t);
315 vm_paddr_t moea_kextract(mmu_t, vm_offset_t);
316 void moea_kenter_attr(mmu_t, vm_offset_t, vm_paddr_t, vm_memattr_t);
317 void moea_kenter(mmu_t, vm_offset_t, vm_paddr_t);
318 void moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma);
319 boolean_t moea_dev_direct_mapped(mmu_t, vm_paddr_t, vm_size_t);
320 static void moea_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
321 void moea_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va);
322 void moea_scan_init(mmu_t mmu);
323 vm_offset_t moea_quick_enter_page(mmu_t mmu, vm_page_t m);
324 void moea_quick_remove_page(mmu_t mmu, vm_offset_t addr);
325 static int moea_map_user_ptr(mmu_t mmu, pmap_t pm,
326     volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen);
327 static int moea_decode_kernel_ptr(mmu_t mmu, vm_offset_t addr,
328     int *is_user, vm_offset_t *decoded_addr);
329 
330 
331 static mmu_method_t moea_methods[] = {
332 	MMUMETHOD(mmu_clear_modify,	moea_clear_modify),
333 	MMUMETHOD(mmu_copy_page,	moea_copy_page),
334 	MMUMETHOD(mmu_copy_pages,	moea_copy_pages),
335 	MMUMETHOD(mmu_enter,		moea_enter),
336 	MMUMETHOD(mmu_enter_object,	moea_enter_object),
337 	MMUMETHOD(mmu_enter_quick,	moea_enter_quick),
338 	MMUMETHOD(mmu_extract,		moea_extract),
339 	MMUMETHOD(mmu_extract_and_hold,	moea_extract_and_hold),
340 	MMUMETHOD(mmu_init,		moea_init),
341 	MMUMETHOD(mmu_is_modified,	moea_is_modified),
342 	MMUMETHOD(mmu_is_prefaultable,	moea_is_prefaultable),
343 	MMUMETHOD(mmu_is_referenced,	moea_is_referenced),
344 	MMUMETHOD(mmu_ts_referenced,	moea_ts_referenced),
345 	MMUMETHOD(mmu_map,     		moea_map),
346 	MMUMETHOD(mmu_page_exists_quick,moea_page_exists_quick),
347 	MMUMETHOD(mmu_page_init,	moea_page_init),
348 	MMUMETHOD(mmu_page_wired_mappings,moea_page_wired_mappings),
349 	MMUMETHOD(mmu_pinit,		moea_pinit),
350 	MMUMETHOD(mmu_pinit0,		moea_pinit0),
351 	MMUMETHOD(mmu_protect,		moea_protect),
352 	MMUMETHOD(mmu_qenter,		moea_qenter),
353 	MMUMETHOD(mmu_qremove,		moea_qremove),
354 	MMUMETHOD(mmu_release,		moea_release),
355 	MMUMETHOD(mmu_remove,		moea_remove),
356 	MMUMETHOD(mmu_remove_all,      	moea_remove_all),
357 	MMUMETHOD(mmu_remove_write,	moea_remove_write),
358 	MMUMETHOD(mmu_sync_icache,	moea_sync_icache),
359 	MMUMETHOD(mmu_unwire,		moea_unwire),
360 	MMUMETHOD(mmu_zero_page,       	moea_zero_page),
361 	MMUMETHOD(mmu_zero_page_area,	moea_zero_page_area),
362 	MMUMETHOD(mmu_activate,		moea_activate),
363 	MMUMETHOD(mmu_deactivate,      	moea_deactivate),
364 	MMUMETHOD(mmu_page_set_memattr,	moea_page_set_memattr),
365 	MMUMETHOD(mmu_quick_enter_page, moea_quick_enter_page),
366 	MMUMETHOD(mmu_quick_remove_page, moea_quick_remove_page),
367 
368 	/* Internal interfaces */
369 	MMUMETHOD(mmu_bootstrap,       	moea_bootstrap),
370 	MMUMETHOD(mmu_cpu_bootstrap,   	moea_cpu_bootstrap),
371 	MMUMETHOD(mmu_mapdev_attr,	moea_mapdev_attr),
372 	MMUMETHOD(mmu_mapdev,		moea_mapdev),
373 	MMUMETHOD(mmu_unmapdev,		moea_unmapdev),
374 	MMUMETHOD(mmu_kextract,		moea_kextract),
375 	MMUMETHOD(mmu_kenter,		moea_kenter),
376 	MMUMETHOD(mmu_kenter_attr,	moea_kenter_attr),
377 	MMUMETHOD(mmu_dev_direct_mapped,moea_dev_direct_mapped),
378 	MMUMETHOD(mmu_scan_init,	moea_scan_init),
379 	MMUMETHOD(mmu_dumpsys_map,	moea_dumpsys_map),
380 	MMUMETHOD(mmu_map_user_ptr,	moea_map_user_ptr),
381 	MMUMETHOD(mmu_decode_kernel_ptr, moea_decode_kernel_ptr),
382 
383 	{ 0, 0 }
384 };
385 
386 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods, 0);
387 
388 static __inline uint32_t
389 moea_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
390 {
391 	uint32_t pte_lo;
392 	int i;
393 
394 	if (ma != VM_MEMATTR_DEFAULT) {
395 		switch (ma) {
396 		case VM_MEMATTR_UNCACHEABLE:
397 			return (PTE_I | PTE_G);
398 		case VM_MEMATTR_CACHEABLE:
399 			return (PTE_M);
400 		case VM_MEMATTR_WRITE_COMBINING:
401 		case VM_MEMATTR_WRITE_BACK:
402 		case VM_MEMATTR_PREFETCHABLE:
403 			return (PTE_I);
404 		case VM_MEMATTR_WRITE_THROUGH:
405 			return (PTE_W | PTE_M);
406 		}
407 	}
408 
409 	/*
410 	 * Assume the page is cache inhibited and access is guarded unless
411 	 * it's in our available memory array.
412 	 */
413 	pte_lo = PTE_I | PTE_G;
414 	for (i = 0; i < pregions_sz; i++) {
415 		if ((pa >= pregions[i].mr_start) &&
416 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
417 			pte_lo = PTE_M;
418 			break;
419 		}
420 	}
421 
422 	return pte_lo;
423 }
424 
425 static void
426 tlbie(vm_offset_t va)
427 {
428 
429 	mtx_lock_spin(&tlbie_mtx);
430 	__asm __volatile("ptesync");
431 	__asm __volatile("tlbie %0" :: "r"(va));
432 	__asm __volatile("eieio; tlbsync; ptesync");
433 	mtx_unlock_spin(&tlbie_mtx);
434 }
435 
436 static void
437 tlbia(void)
438 {
439 	vm_offset_t va;
440 
441 	for (va = 0; va < 0x00040000; va += 0x00001000) {
442 		__asm __volatile("tlbie %0" :: "r"(va));
443 		powerpc_sync();
444 	}
445 	__asm __volatile("tlbsync");
446 	powerpc_sync();
447 }
448 
449 static __inline int
450 va_to_sr(u_int *sr, vm_offset_t va)
451 {
452 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
453 }
454 
455 static __inline u_int
456 va_to_pteg(u_int sr, vm_offset_t addr)
457 {
458 	u_int hash;
459 
460 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
461 	    ADDR_PIDX_SHFT);
462 	return (hash & moea_pteg_mask);
463 }
464 
465 static __inline struct pvo_head *
466 vm_page_to_pvoh(vm_page_t m)
467 {
468 
469 	return (&m->md.mdpg_pvoh);
470 }
471 
472 static __inline void
473 moea_attr_clear(vm_page_t m, int ptebit)
474 {
475 
476 	rw_assert(&pvh_global_lock, RA_WLOCKED);
477 	m->md.mdpg_attrs &= ~ptebit;
478 }
479 
480 static __inline int
481 moea_attr_fetch(vm_page_t m)
482 {
483 
484 	return (m->md.mdpg_attrs);
485 }
486 
487 static __inline void
488 moea_attr_save(vm_page_t m, int ptebit)
489 {
490 
491 	rw_assert(&pvh_global_lock, RA_WLOCKED);
492 	m->md.mdpg_attrs |= ptebit;
493 }
494 
495 static __inline int
496 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
497 {
498 	if (pt->pte_hi == pvo_pt->pte_hi)
499 		return (1);
500 
501 	return (0);
502 }
503 
504 static __inline int
505 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
506 {
507 	return (pt->pte_hi & ~PTE_VALID) ==
508 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
509 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
510 }
511 
512 static __inline void
513 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
514 {
515 
516 	mtx_assert(&moea_table_mutex, MA_OWNED);
517 
518 	/*
519 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
520 	 * set when the real pte is set in memory.
521 	 *
522 	 * Note: Don't set the valid bit for correct operation of tlb update.
523 	 */
524 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
525 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
526 	pt->pte_lo = pte_lo;
527 }
528 
529 static __inline void
530 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
531 {
532 
533 	mtx_assert(&moea_table_mutex, MA_OWNED);
534 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
535 }
536 
537 static __inline void
538 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
539 {
540 
541 	mtx_assert(&moea_table_mutex, MA_OWNED);
542 
543 	/*
544 	 * As shown in Section 7.6.3.2.3
545 	 */
546 	pt->pte_lo &= ~ptebit;
547 	tlbie(va);
548 }
549 
550 static __inline void
551 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
552 {
553 
554 	mtx_assert(&moea_table_mutex, MA_OWNED);
555 	pvo_pt->pte_hi |= PTE_VALID;
556 
557 	/*
558 	 * Update the PTE as defined in section 7.6.3.1.
559 	 * Note that the REF/CHG bits are from pvo_pt and thus should have
560 	 * been saved so this routine can restore them (if desired).
561 	 */
562 	pt->pte_lo = pvo_pt->pte_lo;
563 	powerpc_sync();
564 	pt->pte_hi = pvo_pt->pte_hi;
565 	powerpc_sync();
566 	moea_pte_valid++;
567 }
568 
569 static __inline void
570 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
571 {
572 
573 	mtx_assert(&moea_table_mutex, MA_OWNED);
574 	pvo_pt->pte_hi &= ~PTE_VALID;
575 
576 	/*
577 	 * Force the reg & chg bits back into the PTEs.
578 	 */
579 	powerpc_sync();
580 
581 	/*
582 	 * Invalidate the pte.
583 	 */
584 	pt->pte_hi &= ~PTE_VALID;
585 
586 	tlbie(va);
587 
588 	/*
589 	 * Save the reg & chg bits.
590 	 */
591 	moea_pte_synch(pt, pvo_pt);
592 	moea_pte_valid--;
593 }
594 
595 static __inline void
596 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
597 {
598 
599 	/*
600 	 * Invalidate the PTE
601 	 */
602 	moea_pte_unset(pt, pvo_pt, va);
603 	moea_pte_set(pt, pvo_pt);
604 }
605 
606 /*
607  * Quick sort callout for comparing memory regions.
608  */
609 static int	om_cmp(const void *a, const void *b);
610 
611 static int
612 om_cmp(const void *a, const void *b)
613 {
614 	const struct	ofw_map *mapa;
615 	const struct	ofw_map *mapb;
616 
617 	mapa = a;
618 	mapb = b;
619 	if (mapa->om_pa < mapb->om_pa)
620 		return (-1);
621 	else if (mapa->om_pa > mapb->om_pa)
622 		return (1);
623 	else
624 		return (0);
625 }
626 
627 void
628 moea_cpu_bootstrap(mmu_t mmup, int ap)
629 {
630 	u_int sdr;
631 	int i;
632 
633 	if (ap) {
634 		powerpc_sync();
635 		__asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu));
636 		__asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl));
637 		isync();
638 		__asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu));
639 		__asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl));
640 		isync();
641 	}
642 
643 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
644 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
645 	isync();
646 
647 	__asm __volatile("mtibatu 1,%0" :: "r"(0));
648 	__asm __volatile("mtdbatu 2,%0" :: "r"(0));
649 	__asm __volatile("mtibatu 2,%0" :: "r"(0));
650 	__asm __volatile("mtdbatu 3,%0" :: "r"(0));
651 	__asm __volatile("mtibatu 3,%0" :: "r"(0));
652 	isync();
653 
654 	for (i = 0; i < 16; i++)
655 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
656 	powerpc_sync();
657 
658 	sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10);
659 	__asm __volatile("mtsdr1 %0" :: "r"(sdr));
660 	isync();
661 
662 	tlbia();
663 }
664 
665 void
666 moea_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
667 {
668 	ihandle_t	mmui;
669 	phandle_t	chosen, mmu;
670 	int		sz;
671 	int		i, j;
672 	vm_size_t	size, physsz, hwphyssz;
673 	vm_offset_t	pa, va, off;
674 	void		*dpcpu;
675 	register_t	msr;
676 
677         /*
678          * Set up BAT0 to map the lowest 256 MB area
679          */
680         battable[0x0].batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
681         battable[0x0].batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
682 
683 	/*
684 	 * Map PCI memory space.
685 	 */
686 	battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
687 	battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
688 
689 	battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
690 	battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
691 
692 	battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
693 	battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
694 
695 	battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
696 	battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
697 
698 	/*
699 	 * Map obio devices.
700 	 */
701 	battable[0xf].batl = BATL(0xf0000000, BAT_I|BAT_G, BAT_PP_RW);
702 	battable[0xf].batu = BATU(0xf0000000, BAT_BL_256M, BAT_Vs);
703 
704 	/*
705 	 * Use an IBAT and a DBAT to map the bottom segment of memory
706 	 * where we are. Turn off instruction relocation temporarily
707 	 * to prevent faults while reprogramming the IBAT.
708 	 */
709 	msr = mfmsr();
710 	mtmsr(msr & ~PSL_IR);
711 	__asm (".balign 32; \n"
712 	       "mtibatu 0,%0; mtibatl 0,%1; isync; \n"
713 	       "mtdbatu 0,%0; mtdbatl 0,%1; isync"
714 	    :: "r"(battable[0].batu), "r"(battable[0].batl));
715 	mtmsr(msr);
716 
717 	/* map pci space */
718 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
719 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
720 	isync();
721 
722 	/* set global direct map flag */
723 	hw_direct_map = 1;
724 
725 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
726 	CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
727 
728 	for (i = 0; i < pregions_sz; i++) {
729 		vm_offset_t pa;
730 		vm_offset_t end;
731 
732 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
733 			pregions[i].mr_start,
734 			pregions[i].mr_start + pregions[i].mr_size,
735 			pregions[i].mr_size);
736 		/*
737 		 * Install entries into the BAT table to allow all
738 		 * of physmem to be convered by on-demand BAT entries.
739 		 * The loop will sometimes set the same battable element
740 		 * twice, but that's fine since they won't be used for
741 		 * a while yet.
742 		 */
743 		pa = pregions[i].mr_start & 0xf0000000;
744 		end = pregions[i].mr_start + pregions[i].mr_size;
745 		do {
746                         u_int n = pa >> ADDR_SR_SHFT;
747 
748 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
749 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
750 			pa += SEGMENT_LENGTH;
751 		} while (pa < end);
752 	}
753 
754 	if (PHYS_AVAIL_ENTRIES < regions_sz)
755 		panic("moea_bootstrap: phys_avail too small");
756 
757 	phys_avail_count = 0;
758 	physsz = 0;
759 	hwphyssz = 0;
760 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
761 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
762 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
763 		    regions[i].mr_start + regions[i].mr_size,
764 		    regions[i].mr_size);
765 		if (hwphyssz != 0 &&
766 		    (physsz + regions[i].mr_size) >= hwphyssz) {
767 			if (physsz < hwphyssz) {
768 				phys_avail[j] = regions[i].mr_start;
769 				phys_avail[j + 1] = regions[i].mr_start +
770 				    hwphyssz - physsz;
771 				physsz = hwphyssz;
772 				phys_avail_count++;
773 			}
774 			break;
775 		}
776 		phys_avail[j] = regions[i].mr_start;
777 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
778 		phys_avail_count++;
779 		physsz += regions[i].mr_size;
780 	}
781 
782 	/* Check for overlap with the kernel and exception vectors */
783 	for (j = 0; j < 2*phys_avail_count; j+=2) {
784 		if (phys_avail[j] < EXC_LAST)
785 			phys_avail[j] += EXC_LAST;
786 
787 		if (kernelstart >= phys_avail[j] &&
788 		    kernelstart < phys_avail[j+1]) {
789 			if (kernelend < phys_avail[j+1]) {
790 				phys_avail[2*phys_avail_count] =
791 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
792 				phys_avail[2*phys_avail_count + 1] =
793 				    phys_avail[j+1];
794 				phys_avail_count++;
795 			}
796 
797 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
798 		}
799 
800 		if (kernelend >= phys_avail[j] &&
801 		    kernelend < phys_avail[j+1]) {
802 			if (kernelstart > phys_avail[j]) {
803 				phys_avail[2*phys_avail_count] = phys_avail[j];
804 				phys_avail[2*phys_avail_count + 1] =
805 				    kernelstart & ~PAGE_MASK;
806 				phys_avail_count++;
807 			}
808 
809 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
810 		}
811 	}
812 
813 	physmem = btoc(physsz);
814 
815 	/*
816 	 * Allocate PTEG table.
817 	 */
818 #ifdef PTEGCOUNT
819 	moea_pteg_count = PTEGCOUNT;
820 #else
821 	moea_pteg_count = 0x1000;
822 
823 	while (moea_pteg_count < physmem)
824 		moea_pteg_count <<= 1;
825 
826 	moea_pteg_count >>= 1;
827 #endif /* PTEGCOUNT */
828 
829 	size = moea_pteg_count * sizeof(struct pteg);
830 	CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
831 	    size);
832 	moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
833 	CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
834 	bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
835 	moea_pteg_mask = moea_pteg_count - 1;
836 
837 	/*
838 	 * Allocate pv/overflow lists.
839 	 */
840 	size = sizeof(struct pvo_head) * moea_pteg_count;
841 	moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
842 	    PAGE_SIZE);
843 	CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
844 	for (i = 0; i < moea_pteg_count; i++)
845 		LIST_INIT(&moea_pvo_table[i]);
846 
847 	/*
848 	 * Initialize the lock that synchronizes access to the pteg and pvo
849 	 * tables.
850 	 */
851 	mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF |
852 	    MTX_RECURSE);
853 	mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF);
854 
855 	mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN);
856 
857 	/*
858 	 * Initialise the unmanaged pvo pool.
859 	 */
860 	moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
861 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
862 	moea_bpvo_pool_index = 0;
863 
864 	/*
865 	 * Make sure kernel vsid is allocated as well as VSID 0.
866 	 */
867 	moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
868 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
869 	moea_vsid_bitmap[0] |= 1;
870 
871 	/*
872 	 * Initialize the kernel pmap (which is statically allocated).
873 	 */
874 	PMAP_LOCK_INIT(kernel_pmap);
875 	for (i = 0; i < 16; i++)
876 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
877 	CPU_FILL(&kernel_pmap->pm_active);
878 	RB_INIT(&kernel_pmap->pmap_pvo);
879 
880  	/*
881 	 * Initialize the global pv list lock.
882 	 */
883 	rw_init(&pvh_global_lock, "pmap pv global");
884 
885 	/*
886 	 * Set up the Open Firmware mappings
887 	 */
888 	chosen = OF_finddevice("/chosen");
889 	if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1 &&
890 	    (mmu = OF_instance_to_package(mmui)) != -1 &&
891 	    (sz = OF_getproplen(mmu, "translations")) != -1) {
892 		translations = NULL;
893 		for (i = 0; phys_avail[i] != 0; i += 2) {
894 			if (phys_avail[i + 1] >= sz) {
895 				translations = (struct ofw_map *)phys_avail[i];
896 				break;
897 			}
898 		}
899 		if (translations == NULL)
900 			panic("moea_bootstrap: no space to copy translations");
901 		bzero(translations, sz);
902 		if (OF_getprop(mmu, "translations", translations, sz) == -1)
903 			panic("moea_bootstrap: can't get ofw translations");
904 		CTR0(KTR_PMAP, "moea_bootstrap: translations");
905 		sz /= sizeof(*translations);
906 		qsort(translations, sz, sizeof (*translations), om_cmp);
907 		for (i = 0; i < sz; i++) {
908 			CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
909 			    translations[i].om_pa, translations[i].om_va,
910 			    translations[i].om_len);
911 
912 			/*
913 			 * If the mapping is 1:1, let the RAM and device
914 			 * on-demand BAT tables take care of the translation.
915 			 */
916 			if (translations[i].om_va == translations[i].om_pa)
917 				continue;
918 
919 			/* Enter the pages */
920 			for (off = 0; off < translations[i].om_len;
921 			    off += PAGE_SIZE)
922 				moea_kenter(mmup, translations[i].om_va + off,
923 					    translations[i].om_pa + off);
924 		}
925 	}
926 
927 	/*
928 	 * Calculate the last available physical address.
929 	 */
930 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
931 		;
932 	Maxmem = powerpc_btop(phys_avail[i + 1]);
933 
934 	moea_cpu_bootstrap(mmup,0);
935 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
936 	pmap_bootstrapped++;
937 
938 	/*
939 	 * Set the start and end of kva.
940 	 */
941 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
942 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
943 
944 	/*
945 	 * Allocate a kernel stack with a guard page for thread0 and map it
946 	 * into the kernel page map.
947 	 */
948 	pa = moea_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
949 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
950 	virtual_avail = va + kstack_pages * PAGE_SIZE;
951 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
952 	thread0.td_kstack = va;
953 	thread0.td_kstack_pages = kstack_pages;
954 	for (i = 0; i < kstack_pages; i++) {
955 		moea_kenter(mmup, va, pa);
956 		pa += PAGE_SIZE;
957 		va += PAGE_SIZE;
958 	}
959 
960 	/*
961 	 * Allocate virtual address space for the message buffer.
962 	 */
963 	pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE);
964 	msgbufp = (struct msgbuf *)virtual_avail;
965 	va = virtual_avail;
966 	virtual_avail += round_page(msgbufsize);
967 	while (va < virtual_avail) {
968 		moea_kenter(mmup, va, pa);
969 		pa += PAGE_SIZE;
970 		va += PAGE_SIZE;
971 	}
972 
973 	/*
974 	 * Allocate virtual address space for the dynamic percpu area.
975 	 */
976 	pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
977 	dpcpu = (void *)virtual_avail;
978 	va = virtual_avail;
979 	virtual_avail += DPCPU_SIZE;
980 	while (va < virtual_avail) {
981 		moea_kenter(mmup, va, pa);
982 		pa += PAGE_SIZE;
983 		va += PAGE_SIZE;
984 	}
985 	dpcpu_init(dpcpu, 0);
986 }
987 
988 /*
989  * Activate a user pmap.  The pmap must be activated before it's address
990  * space can be accessed in any way.
991  */
992 void
993 moea_activate(mmu_t mmu, struct thread *td)
994 {
995 	pmap_t	pm, pmr;
996 
997 	/*
998 	 * Load all the data we need up front to encourage the compiler to
999 	 * not issue any loads while we have interrupts disabled below.
1000 	 */
1001 	pm = &td->td_proc->p_vmspace->vm_pmap;
1002 	pmr = pm->pmap_phys;
1003 
1004 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1005 	PCPU_SET(curpmap, pmr);
1006 
1007 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1008 }
1009 
1010 void
1011 moea_deactivate(mmu_t mmu, struct thread *td)
1012 {
1013 	pmap_t	pm;
1014 
1015 	pm = &td->td_proc->p_vmspace->vm_pmap;
1016 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1017 	PCPU_SET(curpmap, NULL);
1018 }
1019 
1020 void
1021 moea_unwire(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1022 {
1023 	struct	pvo_entry key, *pvo;
1024 
1025 	PMAP_LOCK(pm);
1026 	key.pvo_vaddr = sva;
1027 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1028 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1029 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1030 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1031 			panic("moea_unwire: pvo %p is missing PVO_WIRED", pvo);
1032 		pvo->pvo_vaddr &= ~PVO_WIRED;
1033 		pm->pm_stats.wired_count--;
1034 	}
1035 	PMAP_UNLOCK(pm);
1036 }
1037 
1038 void
1039 moea_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1040 {
1041 	vm_offset_t	dst;
1042 	vm_offset_t	src;
1043 
1044 	dst = VM_PAGE_TO_PHYS(mdst);
1045 	src = VM_PAGE_TO_PHYS(msrc);
1046 
1047 	bcopy((void *)src, (void *)dst, PAGE_SIZE);
1048 }
1049 
1050 void
1051 moea_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1052     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1053 {
1054 	void *a_cp, *b_cp;
1055 	vm_offset_t a_pg_offset, b_pg_offset;
1056 	int cnt;
1057 
1058 	while (xfersize > 0) {
1059 		a_pg_offset = a_offset & PAGE_MASK;
1060 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1061 		a_cp = (char *)VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]) +
1062 		    a_pg_offset;
1063 		b_pg_offset = b_offset & PAGE_MASK;
1064 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1065 		b_cp = (char *)VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]) +
1066 		    b_pg_offset;
1067 		bcopy(a_cp, b_cp, cnt);
1068 		a_offset += cnt;
1069 		b_offset += cnt;
1070 		xfersize -= cnt;
1071 	}
1072 }
1073 
1074 /*
1075  * Zero a page of physical memory by temporarily mapping it into the tlb.
1076  */
1077 void
1078 moea_zero_page(mmu_t mmu, vm_page_t m)
1079 {
1080 	vm_offset_t off, pa = VM_PAGE_TO_PHYS(m);
1081 
1082 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1083 		__asm __volatile("dcbz 0,%0" :: "r"(pa + off));
1084 }
1085 
1086 void
1087 moea_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1088 {
1089 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1090 	void *va = (void *)(pa + off);
1091 
1092 	bzero(va, size);
1093 }
1094 
1095 vm_offset_t
1096 moea_quick_enter_page(mmu_t mmu, vm_page_t m)
1097 {
1098 
1099 	return (VM_PAGE_TO_PHYS(m));
1100 }
1101 
1102 void
1103 moea_quick_remove_page(mmu_t mmu, vm_offset_t addr)
1104 {
1105 }
1106 
1107 /*
1108  * Map the given physical page at the specified virtual address in the
1109  * target pmap with the protection requested.  If specified the page
1110  * will be wired down.
1111  */
1112 int
1113 moea_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1114     u_int flags, int8_t psind)
1115 {
1116 	int error;
1117 
1118 	for (;;) {
1119 		rw_wlock(&pvh_global_lock);
1120 		PMAP_LOCK(pmap);
1121 		error = moea_enter_locked(pmap, va, m, prot, flags, psind);
1122 		rw_wunlock(&pvh_global_lock);
1123 		PMAP_UNLOCK(pmap);
1124 		if (error != ENOMEM)
1125 			return (KERN_SUCCESS);
1126 		if ((flags & PMAP_ENTER_NOSLEEP) != 0)
1127 			return (KERN_RESOURCE_SHORTAGE);
1128 		VM_OBJECT_ASSERT_UNLOCKED(m->object);
1129 		vm_wait(NULL);
1130 	}
1131 }
1132 
1133 /*
1134  * Map the given physical page at the specified virtual address in the
1135  * target pmap with the protection requested.  If specified the page
1136  * will be wired down.
1137  *
1138  * The global pvh and pmap must be locked.
1139  */
1140 static int
1141 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1142     u_int flags, int8_t psind __unused)
1143 {
1144 	struct		pvo_head *pvo_head;
1145 	uma_zone_t	zone;
1146 	u_int		pte_lo, pvo_flags;
1147 	int		error;
1148 
1149 	if (pmap_bootstrapped)
1150 		rw_assert(&pvh_global_lock, RA_WLOCKED);
1151 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1152 	if ((m->oflags & VPO_UNMANAGED) == 0) {
1153 		if ((flags & PMAP_ENTER_QUICK_LOCKED) == 0)
1154 			VM_PAGE_OBJECT_BUSY_ASSERT(m);
1155 		else
1156 			VM_OBJECT_ASSERT_LOCKED(m->object);
1157 	}
1158 
1159 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea_initialized) {
1160 		pvo_head = &moea_pvo_kunmanaged;
1161 		zone = moea_upvo_zone;
1162 		pvo_flags = 0;
1163 	} else {
1164 		pvo_head = vm_page_to_pvoh(m);
1165 		zone = moea_mpvo_zone;
1166 		pvo_flags = PVO_MANAGED;
1167 	}
1168 
1169 	pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1170 
1171 	if (prot & VM_PROT_WRITE) {
1172 		pte_lo |= PTE_BW;
1173 		if (pmap_bootstrapped &&
1174 		    (m->oflags & VPO_UNMANAGED) == 0)
1175 			vm_page_aflag_set(m, PGA_WRITEABLE);
1176 	} else
1177 		pte_lo |= PTE_BR;
1178 
1179 	if ((flags & PMAP_ENTER_WIRED) != 0)
1180 		pvo_flags |= PVO_WIRED;
1181 
1182 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1183 	    pte_lo, pvo_flags);
1184 
1185 	/*
1186 	 * Flush the real page from the instruction cache. This has be done
1187 	 * for all user mappings to prevent information leakage via the
1188 	 * instruction cache. moea_pvo_enter() returns ENOENT for the first
1189 	 * mapping for a page.
1190 	 */
1191 	if (pmap != kernel_pmap && error == ENOENT &&
1192 	    (pte_lo & (PTE_I | PTE_G)) == 0)
1193 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1194 
1195 	return (error);
1196 }
1197 
1198 /*
1199  * Maps a sequence of resident pages belonging to the same object.
1200  * The sequence begins with the given page m_start.  This page is
1201  * mapped at the given virtual address start.  Each subsequent page is
1202  * mapped at a virtual address that is offset from start by the same
1203  * amount as the page is offset from m_start within the object.  The
1204  * last page in the sequence is the page with the largest offset from
1205  * m_start that can be mapped at a virtual address less than the given
1206  * virtual address end.  Not every virtual page between start and end
1207  * is mapped; only those for which a resident page exists with the
1208  * corresponding offset from m_start are mapped.
1209  */
1210 void
1211 moea_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1212     vm_page_t m_start, vm_prot_t prot)
1213 {
1214 	vm_page_t m;
1215 	vm_pindex_t diff, psize;
1216 
1217 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1218 
1219 	psize = atop(end - start);
1220 	m = m_start;
1221 	rw_wlock(&pvh_global_lock);
1222 	PMAP_LOCK(pm);
1223 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1224 		moea_enter_locked(pm, start + ptoa(diff), m, prot &
1225 		    (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_QUICK_LOCKED,
1226 		    0);
1227 		m = TAILQ_NEXT(m, listq);
1228 	}
1229 	rw_wunlock(&pvh_global_lock);
1230 	PMAP_UNLOCK(pm);
1231 }
1232 
1233 void
1234 moea_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1235     vm_prot_t prot)
1236 {
1237 
1238 	rw_wlock(&pvh_global_lock);
1239 	PMAP_LOCK(pm);
1240 	moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1241 	    PMAP_ENTER_QUICK_LOCKED, 0);
1242 	rw_wunlock(&pvh_global_lock);
1243 	PMAP_UNLOCK(pm);
1244 }
1245 
1246 vm_paddr_t
1247 moea_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1248 {
1249 	struct	pvo_entry *pvo;
1250 	vm_paddr_t pa;
1251 
1252 	PMAP_LOCK(pm);
1253 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1254 	if (pvo == NULL)
1255 		pa = 0;
1256 	else
1257 		pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1258 	PMAP_UNLOCK(pm);
1259 	return (pa);
1260 }
1261 
1262 /*
1263  * Atomically extract and hold the physical page with the given
1264  * pmap and virtual address pair if that mapping permits the given
1265  * protection.
1266  */
1267 vm_page_t
1268 moea_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1269 {
1270 	struct	pvo_entry *pvo;
1271 	vm_page_t m;
1272 
1273 	m = NULL;
1274 	PMAP_LOCK(pmap);
1275 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1276 	if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1277 	    ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1278 	     (prot & VM_PROT_WRITE) == 0)) {
1279 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
1280 		if (!vm_page_wire_mapped(m))
1281 			m = NULL;
1282 	}
1283 	PMAP_UNLOCK(pmap);
1284 	return (m);
1285 }
1286 
1287 void
1288 moea_init(mmu_t mmu)
1289 {
1290 
1291 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1292 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1293 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1294 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1295 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1296 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1297 	moea_initialized = TRUE;
1298 }
1299 
1300 boolean_t
1301 moea_is_referenced(mmu_t mmu, vm_page_t m)
1302 {
1303 	boolean_t rv;
1304 
1305 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1306 	    ("moea_is_referenced: page %p is not managed", m));
1307 	rw_wlock(&pvh_global_lock);
1308 	rv = moea_query_bit(m, PTE_REF);
1309 	rw_wunlock(&pvh_global_lock);
1310 	return (rv);
1311 }
1312 
1313 boolean_t
1314 moea_is_modified(mmu_t mmu, vm_page_t m)
1315 {
1316 	boolean_t rv;
1317 
1318 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1319 	    ("moea_is_modified: page %p is not managed", m));
1320 
1321 	/*
1322 	 * If the page is not busied then this check is racy.
1323 	 */
1324 	if (!pmap_page_is_write_mapped(m))
1325 		return (FALSE);
1326 
1327 	rw_wlock(&pvh_global_lock);
1328 	rv = moea_query_bit(m, PTE_CHG);
1329 	rw_wunlock(&pvh_global_lock);
1330 	return (rv);
1331 }
1332 
1333 boolean_t
1334 moea_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1335 {
1336 	struct pvo_entry *pvo;
1337 	boolean_t rv;
1338 
1339 	PMAP_LOCK(pmap);
1340 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1341 	rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1342 	PMAP_UNLOCK(pmap);
1343 	return (rv);
1344 }
1345 
1346 void
1347 moea_clear_modify(mmu_t mmu, vm_page_t m)
1348 {
1349 
1350 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1351 	    ("moea_clear_modify: page %p is not managed", m));
1352 	vm_page_assert_busied(m);
1353 
1354 	if (!pmap_page_is_write_mapped(m))
1355 		return;
1356 	rw_wlock(&pvh_global_lock);
1357 	moea_clear_bit(m, PTE_CHG);
1358 	rw_wunlock(&pvh_global_lock);
1359 }
1360 
1361 /*
1362  * Clear the write and modified bits in each of the given page's mappings.
1363  */
1364 void
1365 moea_remove_write(mmu_t mmu, vm_page_t m)
1366 {
1367 	struct	pvo_entry *pvo;
1368 	struct	pte *pt;
1369 	pmap_t	pmap;
1370 	u_int	lo;
1371 
1372 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1373 	    ("moea_remove_write: page %p is not managed", m));
1374 	vm_page_assert_busied(m);
1375 
1376 	if (!pmap_page_is_write_mapped(m))
1377 		return;
1378 	rw_wlock(&pvh_global_lock);
1379 	lo = moea_attr_fetch(m);
1380 	powerpc_sync();
1381 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1382 		pmap = pvo->pvo_pmap;
1383 		PMAP_LOCK(pmap);
1384 		if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1385 			pt = moea_pvo_to_pte(pvo, -1);
1386 			pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1387 			pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1388 			if (pt != NULL) {
1389 				moea_pte_synch(pt, &pvo->pvo_pte.pte);
1390 				lo |= pvo->pvo_pte.pte.pte_lo;
1391 				pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1392 				moea_pte_change(pt, &pvo->pvo_pte.pte,
1393 				    pvo->pvo_vaddr);
1394 				mtx_unlock(&moea_table_mutex);
1395 			}
1396 		}
1397 		PMAP_UNLOCK(pmap);
1398 	}
1399 	if ((lo & PTE_CHG) != 0) {
1400 		moea_attr_clear(m, PTE_CHG);
1401 		vm_page_dirty(m);
1402 	}
1403 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1404 	rw_wunlock(&pvh_global_lock);
1405 }
1406 
1407 /*
1408  *	moea_ts_referenced:
1409  *
1410  *	Return a count of reference bits for a page, clearing those bits.
1411  *	It is not necessary for every reference bit to be cleared, but it
1412  *	is necessary that 0 only be returned when there are truly no
1413  *	reference bits set.
1414  *
1415  *	XXX: The exact number of bits to check and clear is a matter that
1416  *	should be tested and standardized at some point in the future for
1417  *	optimal aging of shared pages.
1418  */
1419 int
1420 moea_ts_referenced(mmu_t mmu, vm_page_t m)
1421 {
1422 	int count;
1423 
1424 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1425 	    ("moea_ts_referenced: page %p is not managed", m));
1426 	rw_wlock(&pvh_global_lock);
1427 	count = moea_clear_bit(m, PTE_REF);
1428 	rw_wunlock(&pvh_global_lock);
1429 	return (count);
1430 }
1431 
1432 /*
1433  * Modify the WIMG settings of all mappings for a page.
1434  */
1435 void
1436 moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1437 {
1438 	struct	pvo_entry *pvo;
1439 	struct	pvo_head *pvo_head;
1440 	struct	pte *pt;
1441 	pmap_t	pmap;
1442 	u_int	lo;
1443 
1444 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1445 		m->md.mdpg_cache_attrs = ma;
1446 		return;
1447 	}
1448 
1449 	rw_wlock(&pvh_global_lock);
1450 	pvo_head = vm_page_to_pvoh(m);
1451 	lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1452 
1453 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1454 		pmap = pvo->pvo_pmap;
1455 		PMAP_LOCK(pmap);
1456 		pt = moea_pvo_to_pte(pvo, -1);
1457 		pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1458 		pvo->pvo_pte.pte.pte_lo |= lo;
1459 		if (pt != NULL) {
1460 			moea_pte_change(pt, &pvo->pvo_pte.pte,
1461 			    pvo->pvo_vaddr);
1462 			if (pvo->pvo_pmap == kernel_pmap)
1463 				isync();
1464 		}
1465 		mtx_unlock(&moea_table_mutex);
1466 		PMAP_UNLOCK(pmap);
1467 	}
1468 	m->md.mdpg_cache_attrs = ma;
1469 	rw_wunlock(&pvh_global_lock);
1470 }
1471 
1472 /*
1473  * Map a wired page into kernel virtual address space.
1474  */
1475 void
1476 moea_kenter(mmu_t mmu, vm_offset_t va, vm_paddr_t pa)
1477 {
1478 
1479 	moea_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1480 }
1481 
1482 void
1483 moea_kenter_attr(mmu_t mmu, vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
1484 {
1485 	u_int		pte_lo;
1486 	int		error;
1487 
1488 #if 0
1489 	if (va < VM_MIN_KERNEL_ADDRESS)
1490 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1491 		    va);
1492 #endif
1493 
1494 	pte_lo = moea_calc_wimg(pa, ma);
1495 
1496 	PMAP_LOCK(kernel_pmap);
1497 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1498 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1499 
1500 	if (error != 0 && error != ENOENT)
1501 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1502 		    pa, error);
1503 
1504 	PMAP_UNLOCK(kernel_pmap);
1505 }
1506 
1507 /*
1508  * Extract the physical page address associated with the given kernel virtual
1509  * address.
1510  */
1511 vm_paddr_t
1512 moea_kextract(mmu_t mmu, vm_offset_t va)
1513 {
1514 	struct		pvo_entry *pvo;
1515 	vm_paddr_t pa;
1516 
1517 	/*
1518 	 * Allow direct mappings on 32-bit OEA
1519 	 */
1520 	if (va < VM_MIN_KERNEL_ADDRESS) {
1521 		return (va);
1522 	}
1523 
1524 	PMAP_LOCK(kernel_pmap);
1525 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1526 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1527 	pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1528 	PMAP_UNLOCK(kernel_pmap);
1529 	return (pa);
1530 }
1531 
1532 /*
1533  * Remove a wired page from kernel virtual address space.
1534  */
1535 void
1536 moea_kremove(mmu_t mmu, vm_offset_t va)
1537 {
1538 
1539 	moea_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1540 }
1541 
1542 /*
1543  * Provide a kernel pointer corresponding to a given userland pointer.
1544  * The returned pointer is valid until the next time this function is
1545  * called in this thread. This is used internally in copyin/copyout.
1546  */
1547 int
1548 moea_map_user_ptr(mmu_t mmu, pmap_t pm, volatile const void *uaddr,
1549     void **kaddr, size_t ulen, size_t *klen)
1550 {
1551 	size_t l;
1552 	register_t vsid;
1553 
1554 	*kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
1555 	l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
1556 	if (l > ulen)
1557 		l = ulen;
1558 	if (klen)
1559 		*klen = l;
1560 	else if (l != ulen)
1561 		return (EFAULT);
1562 
1563 	vsid = va_to_vsid(pm, (vm_offset_t)uaddr);
1564 
1565 	/* Mark segment no-execute */
1566 	vsid |= SR_N;
1567 
1568 	/* If we have already set this VSID, we can just return */
1569 	if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == vsid)
1570 		return (0);
1571 
1572 	__asm __volatile("isync");
1573 	curthread->td_pcb->pcb_cpu.aim.usr_segm =
1574 	    (uintptr_t)uaddr >> ADDR_SR_SHFT;
1575 	curthread->td_pcb->pcb_cpu.aim.usr_vsid = vsid;
1576 	__asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(vsid));
1577 
1578 	return (0);
1579 }
1580 
1581 /*
1582  * Figure out where a given kernel pointer (usually in a fault) points
1583  * to from the VM's perspective, potentially remapping into userland's
1584  * address space.
1585  */
1586 static int
1587 moea_decode_kernel_ptr(mmu_t mmu, vm_offset_t addr, int *is_user,
1588     vm_offset_t *decoded_addr)
1589 {
1590 	vm_offset_t user_sr;
1591 
1592 	if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
1593 		user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
1594 		addr &= ADDR_PIDX | ADDR_POFF;
1595 		addr |= user_sr << ADDR_SR_SHFT;
1596 		*decoded_addr = addr;
1597 		*is_user = 1;
1598 	} else {
1599 		*decoded_addr = addr;
1600 		*is_user = 0;
1601 	}
1602 
1603 	return (0);
1604 }
1605 
1606 /*
1607  * Map a range of physical addresses into kernel virtual address space.
1608  *
1609  * The value passed in *virt is a suggested virtual address for the mapping.
1610  * Architectures which can support a direct-mapped physical to virtual region
1611  * can return the appropriate address within that region, leaving '*virt'
1612  * unchanged.  We cannot and therefore do not; *virt is updated with the
1613  * first usable address after the mapped region.
1614  */
1615 vm_offset_t
1616 moea_map(mmu_t mmu, vm_offset_t *virt, vm_paddr_t pa_start,
1617     vm_paddr_t pa_end, int prot)
1618 {
1619 	vm_offset_t	sva, va;
1620 
1621 	sva = *virt;
1622 	va = sva;
1623 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1624 		moea_kenter(mmu, va, pa_start);
1625 	*virt = va;
1626 	return (sva);
1627 }
1628 
1629 /*
1630  * Returns true if the pmap's pv is one of the first
1631  * 16 pvs linked to from this page.  This count may
1632  * be changed upwards or downwards in the future; it
1633  * is only necessary that true be returned for a small
1634  * subset of pmaps for proper page aging.
1635  */
1636 boolean_t
1637 moea_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1638 {
1639         int loops;
1640 	struct pvo_entry *pvo;
1641 	boolean_t rv;
1642 
1643 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1644 	    ("moea_page_exists_quick: page %p is not managed", m));
1645 	loops = 0;
1646 	rv = FALSE;
1647 	rw_wlock(&pvh_global_lock);
1648 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1649 		if (pvo->pvo_pmap == pmap) {
1650 			rv = TRUE;
1651 			break;
1652 		}
1653 		if (++loops >= 16)
1654 			break;
1655 	}
1656 	rw_wunlock(&pvh_global_lock);
1657 	return (rv);
1658 }
1659 
1660 void
1661 moea_page_init(mmu_t mmu __unused, vm_page_t m)
1662 {
1663 
1664 	m->md.mdpg_attrs = 0;
1665 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
1666 	LIST_INIT(&m->md.mdpg_pvoh);
1667 }
1668 
1669 /*
1670  * Return the number of managed mappings to the given physical page
1671  * that are wired.
1672  */
1673 int
1674 moea_page_wired_mappings(mmu_t mmu, vm_page_t m)
1675 {
1676 	struct pvo_entry *pvo;
1677 	int count;
1678 
1679 	count = 0;
1680 	if ((m->oflags & VPO_UNMANAGED) != 0)
1681 		return (count);
1682 	rw_wlock(&pvh_global_lock);
1683 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1684 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1685 			count++;
1686 	rw_wunlock(&pvh_global_lock);
1687 	return (count);
1688 }
1689 
1690 static u_int	moea_vsidcontext;
1691 
1692 void
1693 moea_pinit(mmu_t mmu, pmap_t pmap)
1694 {
1695 	int	i, mask;
1696 	u_int	entropy;
1697 
1698 	KASSERT((int)pmap < VM_MIN_KERNEL_ADDRESS, ("moea_pinit: virt pmap"));
1699 	RB_INIT(&pmap->pmap_pvo);
1700 
1701 	entropy = 0;
1702 	__asm __volatile("mftb %0" : "=r"(entropy));
1703 
1704 	if ((pmap->pmap_phys = (pmap_t)moea_kextract(mmu, (vm_offset_t)pmap))
1705 	    == NULL) {
1706 		pmap->pmap_phys = pmap;
1707 	}
1708 
1709 
1710 	mtx_lock(&moea_vsid_mutex);
1711 	/*
1712 	 * Allocate some segment registers for this pmap.
1713 	 */
1714 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1715 		u_int	hash, n;
1716 
1717 		/*
1718 		 * Create a new value by mutiplying by a prime and adding in
1719 		 * entropy from the timebase register.  This is to make the
1720 		 * VSID more random so that the PT hash function collides
1721 		 * less often.  (Note that the prime casues gcc to do shifts
1722 		 * instead of a multiply.)
1723 		 */
1724 		moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1725 		hash = moea_vsidcontext & (NPMAPS - 1);
1726 		if (hash == 0)		/* 0 is special, avoid it */
1727 			continue;
1728 		n = hash >> 5;
1729 		mask = 1 << (hash & (VSID_NBPW - 1));
1730 		hash = (moea_vsidcontext & 0xfffff);
1731 		if (moea_vsid_bitmap[n] & mask) {	/* collision? */
1732 			/* anything free in this bucket? */
1733 			if (moea_vsid_bitmap[n] == 0xffffffff) {
1734 				entropy = (moea_vsidcontext >> 20);
1735 				continue;
1736 			}
1737 			i = ffs(~moea_vsid_bitmap[n]) - 1;
1738 			mask = 1 << i;
1739 			hash &= rounddown2(0xfffff, VSID_NBPW);
1740 			hash |= i;
1741 		}
1742 		KASSERT(!(moea_vsid_bitmap[n] & mask),
1743 		    ("Allocating in-use VSID group %#x\n", hash));
1744 		moea_vsid_bitmap[n] |= mask;
1745 		for (i = 0; i < 16; i++)
1746 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1747 		mtx_unlock(&moea_vsid_mutex);
1748 		return;
1749 	}
1750 
1751 	mtx_unlock(&moea_vsid_mutex);
1752 	panic("moea_pinit: out of segments");
1753 }
1754 
1755 /*
1756  * Initialize the pmap associated with process 0.
1757  */
1758 void
1759 moea_pinit0(mmu_t mmu, pmap_t pm)
1760 {
1761 
1762 	PMAP_LOCK_INIT(pm);
1763 	moea_pinit(mmu, pm);
1764 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1765 }
1766 
1767 /*
1768  * Set the physical protection on the specified range of this map as requested.
1769  */
1770 void
1771 moea_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1772     vm_prot_t prot)
1773 {
1774 	struct	pvo_entry *pvo, *tpvo, key;
1775 	struct	pte *pt;
1776 
1777 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1778 	    ("moea_protect: non current pmap"));
1779 
1780 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1781 		moea_remove(mmu, pm, sva, eva);
1782 		return;
1783 	}
1784 
1785 	rw_wlock(&pvh_global_lock);
1786 	PMAP_LOCK(pm);
1787 	key.pvo_vaddr = sva;
1788 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1789 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1790 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1791 
1792 		/*
1793 		 * Grab the PTE pointer before we diddle with the cached PTE
1794 		 * copy.
1795 		 */
1796 		pt = moea_pvo_to_pte(pvo, -1);
1797 		/*
1798 		 * Change the protection of the page.
1799 		 */
1800 		pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1801 		pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1802 
1803 		/*
1804 		 * If the PVO is in the page table, update that pte as well.
1805 		 */
1806 		if (pt != NULL) {
1807 			moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1808 			mtx_unlock(&moea_table_mutex);
1809 		}
1810 	}
1811 	rw_wunlock(&pvh_global_lock);
1812 	PMAP_UNLOCK(pm);
1813 }
1814 
1815 /*
1816  * Map a list of wired pages into kernel virtual address space.  This is
1817  * intended for temporary mappings which do not need page modification or
1818  * references recorded.  Existing mappings in the region are overwritten.
1819  */
1820 void
1821 moea_qenter(mmu_t mmu, vm_offset_t sva, vm_page_t *m, int count)
1822 {
1823 	vm_offset_t va;
1824 
1825 	va = sva;
1826 	while (count-- > 0) {
1827 		moea_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1828 		va += PAGE_SIZE;
1829 		m++;
1830 	}
1831 }
1832 
1833 /*
1834  * Remove page mappings from kernel virtual address space.  Intended for
1835  * temporary mappings entered by moea_qenter.
1836  */
1837 void
1838 moea_qremove(mmu_t mmu, vm_offset_t sva, int count)
1839 {
1840 	vm_offset_t va;
1841 
1842 	va = sva;
1843 	while (count-- > 0) {
1844 		moea_kremove(mmu, va);
1845 		va += PAGE_SIZE;
1846 	}
1847 }
1848 
1849 void
1850 moea_release(mmu_t mmu, pmap_t pmap)
1851 {
1852         int idx, mask;
1853 
1854 	/*
1855 	 * Free segment register's VSID
1856 	 */
1857         if (pmap->pm_sr[0] == 0)
1858                 panic("moea_release");
1859 
1860 	mtx_lock(&moea_vsid_mutex);
1861         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1862         mask = 1 << (idx % VSID_NBPW);
1863         idx /= VSID_NBPW;
1864         moea_vsid_bitmap[idx] &= ~mask;
1865 	mtx_unlock(&moea_vsid_mutex);
1866 }
1867 
1868 /*
1869  * Remove the given range of addresses from the specified map.
1870  */
1871 void
1872 moea_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1873 {
1874 	struct	pvo_entry *pvo, *tpvo, key;
1875 
1876 	rw_wlock(&pvh_global_lock);
1877 	PMAP_LOCK(pm);
1878 	key.pvo_vaddr = sva;
1879 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1880 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1881 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1882 		moea_pvo_remove(pvo, -1);
1883 	}
1884 	PMAP_UNLOCK(pm);
1885 	rw_wunlock(&pvh_global_lock);
1886 }
1887 
1888 /*
1889  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1890  * will reflect changes in pte's back to the vm_page.
1891  */
1892 void
1893 moea_remove_all(mmu_t mmu, vm_page_t m)
1894 {
1895 	struct  pvo_head *pvo_head;
1896 	struct	pvo_entry *pvo, *next_pvo;
1897 	pmap_t	pmap;
1898 
1899 	rw_wlock(&pvh_global_lock);
1900 	pvo_head = vm_page_to_pvoh(m);
1901 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1902 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1903 
1904 		pmap = pvo->pvo_pmap;
1905 		PMAP_LOCK(pmap);
1906 		moea_pvo_remove(pvo, -1);
1907 		PMAP_UNLOCK(pmap);
1908 	}
1909 	if ((m->aflags & PGA_WRITEABLE) && moea_query_bit(m, PTE_CHG)) {
1910 		moea_attr_clear(m, PTE_CHG);
1911 		vm_page_dirty(m);
1912 	}
1913 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1914 	rw_wunlock(&pvh_global_lock);
1915 }
1916 
1917 /*
1918  * Allocate a physical page of memory directly from the phys_avail map.
1919  * Can only be called from moea_bootstrap before avail start and end are
1920  * calculated.
1921  */
1922 static vm_offset_t
1923 moea_bootstrap_alloc(vm_size_t size, u_int align)
1924 {
1925 	vm_offset_t	s, e;
1926 	int		i, j;
1927 
1928 	size = round_page(size);
1929 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1930 		if (align != 0)
1931 			s = roundup2(phys_avail[i], align);
1932 		else
1933 			s = phys_avail[i];
1934 		e = s + size;
1935 
1936 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1937 			continue;
1938 
1939 		if (s == phys_avail[i]) {
1940 			phys_avail[i] += size;
1941 		} else if (e == phys_avail[i + 1]) {
1942 			phys_avail[i + 1] -= size;
1943 		} else {
1944 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1945 				phys_avail[j] = phys_avail[j - 2];
1946 				phys_avail[j + 1] = phys_avail[j - 1];
1947 			}
1948 
1949 			phys_avail[i + 3] = phys_avail[i + 1];
1950 			phys_avail[i + 1] = s;
1951 			phys_avail[i + 2] = e;
1952 			phys_avail_count++;
1953 		}
1954 
1955 		return (s);
1956 	}
1957 	panic("moea_bootstrap_alloc: could not allocate memory");
1958 }
1959 
1960 static void
1961 moea_syncicache(vm_paddr_t pa, vm_size_t len)
1962 {
1963 	__syncicache((void *)pa, len);
1964 }
1965 
1966 static int
1967 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1968     vm_offset_t va, vm_paddr_t pa, u_int pte_lo, int flags)
1969 {
1970 	struct	pvo_entry *pvo;
1971 	u_int	sr;
1972 	int	first;
1973 	u_int	ptegidx;
1974 	int	i;
1975 	int     bootstrap;
1976 
1977 	moea_pvo_enter_calls++;
1978 	first = 0;
1979 	bootstrap = 0;
1980 
1981 	/*
1982 	 * Compute the PTE Group index.
1983 	 */
1984 	va &= ~ADDR_POFF;
1985 	sr = va_to_sr(pm->pm_sr, va);
1986 	ptegidx = va_to_pteg(sr, va);
1987 
1988 	/*
1989 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
1990 	 * there is a mapping.
1991 	 */
1992 	mtx_lock(&moea_table_mutex);
1993 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
1994 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
1995 			if ((pvo->pvo_pte.pte.pte_lo & PTE_RPGN) == pa &&
1996 			    (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
1997 			    (pte_lo & PTE_PP)) {
1998 				/*
1999 				 * The PTE is not changing.  Instead, this may
2000 				 * be a request to change the mapping's wired
2001 				 * attribute.
2002 				 */
2003 				mtx_unlock(&moea_table_mutex);
2004 				if ((flags & PVO_WIRED) != 0 &&
2005 				    (pvo->pvo_vaddr & PVO_WIRED) == 0) {
2006 					pvo->pvo_vaddr |= PVO_WIRED;
2007 					pm->pm_stats.wired_count++;
2008 				} else if ((flags & PVO_WIRED) == 0 &&
2009 				    (pvo->pvo_vaddr & PVO_WIRED) != 0) {
2010 					pvo->pvo_vaddr &= ~PVO_WIRED;
2011 					pm->pm_stats.wired_count--;
2012 				}
2013 				return (0);
2014 			}
2015 			moea_pvo_remove(pvo, -1);
2016 			break;
2017 		}
2018 	}
2019 
2020 	/*
2021 	 * If we aren't overwriting a mapping, try to allocate.
2022 	 */
2023 	if (moea_initialized) {
2024 		pvo = uma_zalloc(zone, M_NOWAIT);
2025 	} else {
2026 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
2027 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
2028 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
2029 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
2030 		}
2031 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
2032 		moea_bpvo_pool_index++;
2033 		bootstrap = 1;
2034 	}
2035 
2036 	if (pvo == NULL) {
2037 		mtx_unlock(&moea_table_mutex);
2038 		return (ENOMEM);
2039 	}
2040 
2041 	moea_pvo_entries++;
2042 	pvo->pvo_vaddr = va;
2043 	pvo->pvo_pmap = pm;
2044 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
2045 	pvo->pvo_vaddr &= ~ADDR_POFF;
2046 	if (flags & PVO_WIRED)
2047 		pvo->pvo_vaddr |= PVO_WIRED;
2048 	if (pvo_head != &moea_pvo_kunmanaged)
2049 		pvo->pvo_vaddr |= PVO_MANAGED;
2050 	if (bootstrap)
2051 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
2052 
2053 	moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo);
2054 
2055 	/*
2056 	 * Add to pmap list
2057 	 */
2058 	RB_INSERT(pvo_tree, &pm->pmap_pvo, pvo);
2059 
2060 	/*
2061 	 * Remember if the list was empty and therefore will be the first
2062 	 * item.
2063 	 */
2064 	if (LIST_FIRST(pvo_head) == NULL)
2065 		first = 1;
2066 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2067 
2068 	if (pvo->pvo_vaddr & PVO_WIRED)
2069 		pm->pm_stats.wired_count++;
2070 	pm->pm_stats.resident_count++;
2071 
2072 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2073 	KASSERT(i < 8, ("Invalid PTE index"));
2074 	if (i >= 0) {
2075 		PVO_PTEGIDX_SET(pvo, i);
2076 	} else {
2077 		panic("moea_pvo_enter: overflow");
2078 		moea_pte_overflow++;
2079 	}
2080 	mtx_unlock(&moea_table_mutex);
2081 
2082 	return (first ? ENOENT : 0);
2083 }
2084 
2085 static void
2086 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
2087 {
2088 	struct	pte *pt;
2089 
2090 	/*
2091 	 * If there is an active pte entry, we need to deactivate it (and
2092 	 * save the ref & cfg bits).
2093 	 */
2094 	pt = moea_pvo_to_pte(pvo, pteidx);
2095 	if (pt != NULL) {
2096 		moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
2097 		mtx_unlock(&moea_table_mutex);
2098 		PVO_PTEGIDX_CLR(pvo);
2099 	} else {
2100 		moea_pte_overflow--;
2101 	}
2102 
2103 	/*
2104 	 * Update our statistics.
2105 	 */
2106 	pvo->pvo_pmap->pm_stats.resident_count--;
2107 	if (pvo->pvo_vaddr & PVO_WIRED)
2108 		pvo->pvo_pmap->pm_stats.wired_count--;
2109 
2110 	/*
2111 	 * Remove this PVO from the PV and pmap lists.
2112 	 */
2113 	LIST_REMOVE(pvo, pvo_vlink);
2114 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2115 
2116 	/*
2117 	 * Save the REF/CHG bits into their cache if the page is managed.
2118 	 * Clear PGA_WRITEABLE if all mappings of the page have been removed.
2119 	 */
2120 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED) {
2121 		struct vm_page *pg;
2122 
2123 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
2124 		if (pg != NULL) {
2125 			moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2126 			    (PTE_REF | PTE_CHG));
2127 			if (LIST_EMPTY(&pg->md.mdpg_pvoh))
2128 				vm_page_aflag_clear(pg, PGA_WRITEABLE);
2129 		}
2130 	}
2131 
2132 	/*
2133 	 * Remove this from the overflow list and return it to the pool
2134 	 * if we aren't going to reuse it.
2135 	 */
2136 	LIST_REMOVE(pvo, pvo_olink);
2137 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2138 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
2139 		    moea_upvo_zone, pvo);
2140 	moea_pvo_entries--;
2141 	moea_pvo_remove_calls++;
2142 }
2143 
2144 static __inline int
2145 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2146 {
2147 	int	pteidx;
2148 
2149 	/*
2150 	 * We can find the actual pte entry without searching by grabbing
2151 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2152 	 * noticing the HID bit.
2153 	 */
2154 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2155 	if (pvo->pvo_pte.pte.pte_hi & PTE_HID)
2156 		pteidx ^= moea_pteg_mask * 8;
2157 
2158 	return (pteidx);
2159 }
2160 
2161 static struct pvo_entry *
2162 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2163 {
2164 	struct	pvo_entry *pvo;
2165 	int	ptegidx;
2166 	u_int	sr;
2167 
2168 	va &= ~ADDR_POFF;
2169 	sr = va_to_sr(pm->pm_sr, va);
2170 	ptegidx = va_to_pteg(sr, va);
2171 
2172 	mtx_lock(&moea_table_mutex);
2173 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2174 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2175 			if (pteidx_p)
2176 				*pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2177 			break;
2178 		}
2179 	}
2180 	mtx_unlock(&moea_table_mutex);
2181 
2182 	return (pvo);
2183 }
2184 
2185 static struct pte *
2186 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2187 {
2188 	struct	pte *pt;
2189 
2190 	/*
2191 	 * If we haven't been supplied the ptegidx, calculate it.
2192 	 */
2193 	if (pteidx == -1) {
2194 		int	ptegidx;
2195 		u_int	sr;
2196 
2197 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2198 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2199 		pteidx = moea_pvo_pte_index(pvo, ptegidx);
2200 	}
2201 
2202 	pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2203 	mtx_lock(&moea_table_mutex);
2204 
2205 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2206 		panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2207 		    "valid pte index", pvo);
2208 	}
2209 
2210 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2211 		panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2212 		    "pvo but no valid pte", pvo);
2213 	}
2214 
2215 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2216 		if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) {
2217 			panic("moea_pvo_to_pte: pvo %p has valid pte in "
2218 			    "moea_pteg_table %p but invalid in pvo", pvo, pt);
2219 		}
2220 
2221 		if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2222 		    != 0) {
2223 			panic("moea_pvo_to_pte: pvo %p pte does not match "
2224 			    "pte %p in moea_pteg_table", pvo, pt);
2225 		}
2226 
2227 		mtx_assert(&moea_table_mutex, MA_OWNED);
2228 		return (pt);
2229 	}
2230 
2231 	if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) {
2232 		panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2233 		    "moea_pteg_table but valid in pvo: %8x, %8x", pvo, pt, pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2234 	}
2235 
2236 	mtx_unlock(&moea_table_mutex);
2237 	return (NULL);
2238 }
2239 
2240 /*
2241  * XXX: THIS STUFF SHOULD BE IN pte.c?
2242  */
2243 int
2244 moea_pte_spill(vm_offset_t addr)
2245 {
2246 	struct	pvo_entry *source_pvo, *victim_pvo;
2247 	struct	pvo_entry *pvo;
2248 	int	ptegidx, i, j;
2249 	u_int	sr;
2250 	struct	pteg *pteg;
2251 	struct	pte *pt;
2252 
2253 	moea_pte_spills++;
2254 
2255 	sr = mfsrin(addr);
2256 	ptegidx = va_to_pteg(sr, addr);
2257 
2258 	/*
2259 	 * Have to substitute some entry.  Use the primary hash for this.
2260 	 * Use low bits of timebase as random generator.
2261 	 */
2262 	pteg = &moea_pteg_table[ptegidx];
2263 	mtx_lock(&moea_table_mutex);
2264 	__asm __volatile("mftb %0" : "=r"(i));
2265 	i &= 7;
2266 	pt = &pteg->pt[i];
2267 
2268 	source_pvo = NULL;
2269 	victim_pvo = NULL;
2270 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2271 		/*
2272 		 * We need to find a pvo entry for this address.
2273 		 */
2274 		if (source_pvo == NULL &&
2275 		    moea_pte_match(&pvo->pvo_pte.pte, sr, addr,
2276 		    pvo->pvo_pte.pte.pte_hi & PTE_HID)) {
2277 			/*
2278 			 * Now found an entry to be spilled into the pteg.
2279 			 * The PTE is now valid, so we know it's active.
2280 			 */
2281 			j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2282 
2283 			if (j >= 0) {
2284 				PVO_PTEGIDX_SET(pvo, j);
2285 				moea_pte_overflow--;
2286 				mtx_unlock(&moea_table_mutex);
2287 				return (1);
2288 			}
2289 
2290 			source_pvo = pvo;
2291 
2292 			if (victim_pvo != NULL)
2293 				break;
2294 		}
2295 
2296 		/*
2297 		 * We also need the pvo entry of the victim we are replacing
2298 		 * so save the R & C bits of the PTE.
2299 		 */
2300 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2301 		    moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2302 			victim_pvo = pvo;
2303 			if (source_pvo != NULL)
2304 				break;
2305 		}
2306 	}
2307 
2308 	if (source_pvo == NULL) {
2309 		mtx_unlock(&moea_table_mutex);
2310 		return (0);
2311 	}
2312 
2313 	if (victim_pvo == NULL) {
2314 		if ((pt->pte_hi & PTE_HID) == 0)
2315 			panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2316 			    "entry", pt);
2317 
2318 		/*
2319 		 * If this is a secondary PTE, we need to search it's primary
2320 		 * pvo bucket for the matching PVO.
2321 		 */
2322 		LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2323 		    pvo_olink) {
2324 			/*
2325 			 * We also need the pvo entry of the victim we are
2326 			 * replacing so save the R & C bits of the PTE.
2327 			 */
2328 			if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2329 				victim_pvo = pvo;
2330 				break;
2331 			}
2332 		}
2333 
2334 		if (victim_pvo == NULL)
2335 			panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2336 			    "entry", pt);
2337 	}
2338 
2339 	/*
2340 	 * We are invalidating the TLB entry for the EA we are replacing even
2341 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2342 	 * contained in the TLB entry.
2343 	 */
2344 	source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID;
2345 
2346 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2347 	moea_pte_set(pt, &source_pvo->pvo_pte.pte);
2348 
2349 	PVO_PTEGIDX_CLR(victim_pvo);
2350 	PVO_PTEGIDX_SET(source_pvo, i);
2351 	moea_pte_replacements++;
2352 
2353 	mtx_unlock(&moea_table_mutex);
2354 	return (1);
2355 }
2356 
2357 static __inline struct pvo_entry *
2358 moea_pte_spillable_ident(u_int ptegidx)
2359 {
2360 	struct	pte *pt;
2361 	struct	pvo_entry *pvo_walk, *pvo = NULL;
2362 
2363 	LIST_FOREACH(pvo_walk, &moea_pvo_table[ptegidx], pvo_olink) {
2364 		if (pvo_walk->pvo_vaddr & PVO_WIRED)
2365 			continue;
2366 
2367 		if (!(pvo_walk->pvo_pte.pte.pte_hi & PTE_VALID))
2368 			continue;
2369 
2370 		pt = moea_pvo_to_pte(pvo_walk, -1);
2371 
2372 		if (pt == NULL)
2373 			continue;
2374 
2375 		pvo = pvo_walk;
2376 
2377 		mtx_unlock(&moea_table_mutex);
2378 		if (!(pt->pte_lo & PTE_REF))
2379 			return (pvo_walk);
2380 	}
2381 
2382 	return (pvo);
2383 }
2384 
2385 static int
2386 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2387 {
2388 	struct	pte *pt;
2389 	struct	pvo_entry *victim_pvo;
2390 	int	i;
2391 	int	victim_idx;
2392 	u_int	pteg_bkpidx = ptegidx;
2393 
2394 	mtx_assert(&moea_table_mutex, MA_OWNED);
2395 
2396 	/*
2397 	 * First try primary hash.
2398 	 */
2399 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2400 		if ((pt->pte_hi & PTE_VALID) == 0) {
2401 			pvo_pt->pte_hi &= ~PTE_HID;
2402 			moea_pte_set(pt, pvo_pt);
2403 			return (i);
2404 		}
2405 	}
2406 
2407 	/*
2408 	 * Now try secondary hash.
2409 	 */
2410 	ptegidx ^= moea_pteg_mask;
2411 
2412 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2413 		if ((pt->pte_hi & PTE_VALID) == 0) {
2414 			pvo_pt->pte_hi |= PTE_HID;
2415 			moea_pte_set(pt, pvo_pt);
2416 			return (i);
2417 		}
2418 	}
2419 
2420 	/* Try again, but this time try to force a PTE out. */
2421 	ptegidx = pteg_bkpidx;
2422 
2423 	victim_pvo = moea_pte_spillable_ident(ptegidx);
2424 	if (victim_pvo == NULL) {
2425 		ptegidx ^= moea_pteg_mask;
2426 		victim_pvo = moea_pte_spillable_ident(ptegidx);
2427 	}
2428 
2429 	if (victim_pvo == NULL) {
2430 		panic("moea_pte_insert: overflow");
2431 		return (-1);
2432 	}
2433 
2434 	victim_idx = moea_pvo_pte_index(victim_pvo, ptegidx);
2435 
2436 	if (pteg_bkpidx == ptegidx)
2437 		pvo_pt->pte_hi &= ~PTE_HID;
2438 	else
2439 		pvo_pt->pte_hi |= PTE_HID;
2440 
2441 	/*
2442 	 * Synchronize the sacrifice PTE with its PVO, then mark both
2443 	 * invalid. The PVO will be reused when/if the VM system comes
2444 	 * here after a fault.
2445 	 */
2446 	pt = &moea_pteg_table[victim_idx >> 3].pt[victim_idx & 7];
2447 
2448 	if (pt->pte_hi != victim_pvo->pvo_pte.pte.pte_hi)
2449 	    panic("Victim PVO doesn't match PTE! PVO: %8x, PTE: %8x", victim_pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2450 
2451 	/*
2452 	 * Set the new PTE.
2453 	 */
2454 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2455 	PVO_PTEGIDX_CLR(victim_pvo);
2456 	moea_pte_overflow++;
2457 	moea_pte_set(pt, pvo_pt);
2458 
2459 	return (victim_idx & 7);
2460 }
2461 
2462 static boolean_t
2463 moea_query_bit(vm_page_t m, int ptebit)
2464 {
2465 	struct	pvo_entry *pvo;
2466 	struct	pte *pt;
2467 
2468 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2469 	if (moea_attr_fetch(m) & ptebit)
2470 		return (TRUE);
2471 
2472 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2473 
2474 		/*
2475 		 * See if we saved the bit off.  If so, cache it and return
2476 		 * success.
2477 		 */
2478 		if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2479 			moea_attr_save(m, ptebit);
2480 			return (TRUE);
2481 		}
2482 	}
2483 
2484 	/*
2485 	 * No luck, now go through the hard part of looking at the PTEs
2486 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2487 	 * the PTEs.
2488 	 */
2489 	powerpc_sync();
2490 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2491 
2492 		/*
2493 		 * See if this pvo has a valid PTE.  if so, fetch the
2494 		 * REF/CHG bits from the valid PTE.  If the appropriate
2495 		 * ptebit is set, cache it and return success.
2496 		 */
2497 		pt = moea_pvo_to_pte(pvo, -1);
2498 		if (pt != NULL) {
2499 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2500 			mtx_unlock(&moea_table_mutex);
2501 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2502 				moea_attr_save(m, ptebit);
2503 				return (TRUE);
2504 			}
2505 		}
2506 	}
2507 
2508 	return (FALSE);
2509 }
2510 
2511 static u_int
2512 moea_clear_bit(vm_page_t m, int ptebit)
2513 {
2514 	u_int	count;
2515 	struct	pvo_entry *pvo;
2516 	struct	pte *pt;
2517 
2518 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2519 
2520 	/*
2521 	 * Clear the cached value.
2522 	 */
2523 	moea_attr_clear(m, ptebit);
2524 
2525 	/*
2526 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2527 	 * we can reset the right ones).  note that since the pvo entries and
2528 	 * list heads are accessed via BAT0 and are never placed in the page
2529 	 * table, we don't have to worry about further accesses setting the
2530 	 * REF/CHG bits.
2531 	 */
2532 	powerpc_sync();
2533 
2534 	/*
2535 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2536 	 * valid pte clear the ptebit from the valid pte.
2537 	 */
2538 	count = 0;
2539 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2540 		pt = moea_pvo_to_pte(pvo, -1);
2541 		if (pt != NULL) {
2542 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2543 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2544 				count++;
2545 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2546 			}
2547 			mtx_unlock(&moea_table_mutex);
2548 		}
2549 		pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2550 	}
2551 
2552 	return (count);
2553 }
2554 
2555 /*
2556  * Return true if the physical range is encompassed by the battable[idx]
2557  */
2558 static int
2559 moea_bat_mapped(int idx, vm_paddr_t pa, vm_size_t size)
2560 {
2561 	u_int prot;
2562 	u_int32_t start;
2563 	u_int32_t end;
2564 	u_int32_t bat_ble;
2565 
2566 	/*
2567 	 * Return immediately if not a valid mapping
2568 	 */
2569 	if (!(battable[idx].batu & BAT_Vs))
2570 		return (EINVAL);
2571 
2572 	/*
2573 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2574 	 * so it can function as an i/o page
2575 	 */
2576 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2577 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2578 		return (EPERM);
2579 
2580 	/*
2581 	 * The address should be within the BAT range. Assume that the
2582 	 * start address in the BAT has the correct alignment (thus
2583 	 * not requiring masking)
2584 	 */
2585 	start = battable[idx].batl & BAT_PBS;
2586 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2587 	end = start | (bat_ble << 15) | 0x7fff;
2588 
2589 	if ((pa < start) || ((pa + size) > end))
2590 		return (ERANGE);
2591 
2592 	return (0);
2593 }
2594 
2595 boolean_t
2596 moea_dev_direct_mapped(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2597 {
2598 	int i;
2599 
2600 	/*
2601 	 * This currently does not work for entries that
2602 	 * overlap 256M BAT segments.
2603 	 */
2604 
2605 	for(i = 0; i < 16; i++)
2606 		if (moea_bat_mapped(i, pa, size) == 0)
2607 			return (0);
2608 
2609 	return (EFAULT);
2610 }
2611 
2612 /*
2613  * Map a set of physical memory pages into the kernel virtual
2614  * address space. Return a pointer to where it is mapped. This
2615  * routine is intended to be used for mapping device memory,
2616  * NOT real memory.
2617  */
2618 void *
2619 moea_mapdev(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2620 {
2621 
2622 	return (moea_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT));
2623 }
2624 
2625 void *
2626 moea_mapdev_attr(mmu_t mmu, vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
2627 {
2628 	vm_offset_t va, tmpva, ppa, offset;
2629 	int i;
2630 
2631 	ppa = trunc_page(pa);
2632 	offset = pa & PAGE_MASK;
2633 	size = roundup(offset + size, PAGE_SIZE);
2634 
2635 	/*
2636 	 * If the physical address lies within a valid BAT table entry,
2637 	 * return the 1:1 mapping. This currently doesn't work
2638 	 * for regions that overlap 256M BAT segments.
2639 	 */
2640 	for (i = 0; i < 16; i++) {
2641 		if (moea_bat_mapped(i, pa, size) == 0)
2642 			return ((void *) pa);
2643 	}
2644 
2645 	va = kva_alloc(size);
2646 	if (!va)
2647 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2648 
2649 	for (tmpva = va; size > 0;) {
2650 		moea_kenter_attr(mmu, tmpva, ppa, ma);
2651 		tlbie(tmpva);
2652 		size -= PAGE_SIZE;
2653 		tmpva += PAGE_SIZE;
2654 		ppa += PAGE_SIZE;
2655 	}
2656 
2657 	return ((void *)(va + offset));
2658 }
2659 
2660 void
2661 moea_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2662 {
2663 	vm_offset_t base, offset;
2664 
2665 	/*
2666 	 * If this is outside kernel virtual space, then it's a
2667 	 * battable entry and doesn't require unmapping
2668 	 */
2669 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2670 		base = trunc_page(va);
2671 		offset = va & PAGE_MASK;
2672 		size = roundup(offset + size, PAGE_SIZE);
2673 		kva_free(base, size);
2674 	}
2675 }
2676 
2677 static void
2678 moea_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2679 {
2680 	struct pvo_entry *pvo;
2681 	vm_offset_t lim;
2682 	vm_paddr_t pa;
2683 	vm_size_t len;
2684 
2685 	PMAP_LOCK(pm);
2686 	while (sz > 0) {
2687 		lim = round_page(va);
2688 		len = MIN(lim - va, sz);
2689 		pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2690 		if (pvo != NULL) {
2691 			pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) |
2692 			    (va & ADDR_POFF);
2693 			moea_syncicache(pa, len);
2694 		}
2695 		va += len;
2696 		sz -= len;
2697 	}
2698 	PMAP_UNLOCK(pm);
2699 }
2700 
2701 void
2702 moea_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va)
2703 {
2704 
2705 	*va = (void *)pa;
2706 }
2707 
2708 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
2709 
2710 void
2711 moea_scan_init(mmu_t mmu)
2712 {
2713 	struct pvo_entry *pvo;
2714 	vm_offset_t va;
2715 	int i;
2716 
2717 	if (!do_minidump) {
2718 		/* Initialize phys. segments for dumpsys(). */
2719 		memset(&dump_map, 0, sizeof(dump_map));
2720 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
2721 		for (i = 0; i < pregions_sz; i++) {
2722 			dump_map[i].pa_start = pregions[i].mr_start;
2723 			dump_map[i].pa_size = pregions[i].mr_size;
2724 		}
2725 		return;
2726 	}
2727 
2728 	/* Virtual segments for minidumps: */
2729 	memset(&dump_map, 0, sizeof(dump_map));
2730 
2731 	/* 1st: kernel .data and .bss. */
2732 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
2733 	dump_map[0].pa_size =
2734 	    round_page((uintptr_t)_end) - dump_map[0].pa_start;
2735 
2736 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2737 	dump_map[1].pa_start = (vm_paddr_t)msgbufp->msg_ptr;
2738 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
2739 
2740 	/* 3rd: kernel VM. */
2741 	va = dump_map[1].pa_start + dump_map[1].pa_size;
2742 	/* Find start of next chunk (from va). */
2743 	while (va < virtual_end) {
2744 		/* Don't dump the buffer cache. */
2745 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
2746 			va = kmi.buffer_eva;
2747 			continue;
2748 		}
2749 		pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
2750 		if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2751 			break;
2752 		va += PAGE_SIZE;
2753 	}
2754 	if (va < virtual_end) {
2755 		dump_map[2].pa_start = va;
2756 		va += PAGE_SIZE;
2757 		/* Find last page in chunk. */
2758 		while (va < virtual_end) {
2759 			/* Don't run into the buffer cache. */
2760 			if (va == kmi.buffer_sva)
2761 				break;
2762 			pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF,
2763 			    NULL);
2764 			if (pvo == NULL ||
2765 			    !(pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2766 				break;
2767 			va += PAGE_SIZE;
2768 		}
2769 		dump_map[2].pa_size = va - dump_map[2].pa_start;
2770 	}
2771 }
2772