xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision 7ae99f80b6661760c5de3edd330b279f04b092a2)
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/mman.h>
118 #include <sys/msgbuf.h>
119 #include <sys/mutex.h>
120 #include <sys/proc.h>
121 #include <sys/rwlock.h>
122 #include <sys/sched.h>
123 #include <sys/sysctl.h>
124 #include <sys/systm.h>
125 #include <sys/vmmeter.h>
126 
127 #include <dev/ofw/openfirm.h>
128 
129 #include <vm/vm.h>
130 #include <vm/pmap.h>
131 #include <vm/vm_param.h>
132 #include <vm/vm_kern.h>
133 #include <vm/vm_page.h>
134 #include <vm/vm_map.h>
135 #include <vm/vm_object.h>
136 #include <vm/vm_extern.h>
137 #include <vm/vm_page.h>
138 #include <vm/vm_phys.h>
139 #include <vm/vm_pageout.h>
140 #include <vm/uma.h>
141 
142 #include <machine/cpu.h>
143 #include <machine/platform.h>
144 #include <machine/bat.h>
145 #include <machine/frame.h>
146 #include <machine/md_var.h>
147 #include <machine/psl.h>
148 #include <machine/pte.h>
149 #include <machine/smp.h>
150 #include <machine/sr.h>
151 #include <machine/mmuvar.h>
152 #include <machine/trap.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 /* Get physical address from PVO. */
163 #define PVO_PADDR(pvo)		((pvo)->pvo_pte.pte.pte_lo & PTE_RPGN)
164 
165 struct ofw_map {
166 	vm_offset_t	om_va;
167 	vm_size_t	om_len;
168 	vm_offset_t	om_pa;
169 	u_int		om_mode;
170 };
171 
172 extern unsigned char _etext[];
173 extern unsigned char _end[];
174 
175 /*
176  * Map of physical memory regions.
177  */
178 static struct	mem_region *regions;
179 static struct	mem_region *pregions;
180 static u_int    phys_avail_count;
181 static int	regions_sz, pregions_sz;
182 static struct	ofw_map *translations;
183 
184 /*
185  * Lock for the pteg and pvo tables.
186  */
187 struct mtx	moea_table_mutex;
188 struct mtx	moea_vsid_mutex;
189 
190 /* tlbie instruction synchronization */
191 static struct mtx tlbie_mtx;
192 
193 /*
194  * PTEG data.
195  */
196 static struct	pteg *moea_pteg_table;
197 u_int		moea_pteg_count;
198 u_int		moea_pteg_mask;
199 
200 /*
201  * PVO data.
202  */
203 struct	pvo_head *moea_pvo_table;		/* pvo entries by pteg index */
204 struct	pvo_head moea_pvo_kunmanaged =
205     LIST_HEAD_INITIALIZER(moea_pvo_kunmanaged);	/* list of unmanaged pages */
206 
207 static struct rwlock_padalign pvh_global_lock;
208 
209 uma_zone_t	moea_upvo_zone;	/* zone for pvo entries for unmanaged pages */
210 uma_zone_t	moea_mpvo_zone;	/* zone for pvo entries for managed pages */
211 
212 #define	BPVO_POOL_SIZE	32768
213 static struct	pvo_entry *moea_bpvo_pool;
214 static int	moea_bpvo_pool_index = 0;
215 
216 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
217 static u_int	moea_vsid_bitmap[NPMAPS / VSID_NBPW];
218 
219 static boolean_t moea_initialized = FALSE;
220 
221 /*
222  * Statistics.
223  */
224 u_int	moea_pte_valid = 0;
225 u_int	moea_pte_overflow = 0;
226 u_int	moea_pte_replacements = 0;
227 u_int	moea_pvo_entries = 0;
228 u_int	moea_pvo_enter_calls = 0;
229 u_int	moea_pvo_remove_calls = 0;
230 u_int	moea_pte_spills = 0;
231 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_valid, CTLFLAG_RD, &moea_pte_valid,
232     0, "");
233 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_overflow, CTLFLAG_RD,
234     &moea_pte_overflow, 0, "");
235 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_replacements, CTLFLAG_RD,
236     &moea_pte_replacements, 0, "");
237 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_entries, CTLFLAG_RD, &moea_pvo_entries,
238     0, "");
239 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_enter_calls, CTLFLAG_RD,
240     &moea_pvo_enter_calls, 0, "");
241 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_remove_calls, CTLFLAG_RD,
242     &moea_pvo_remove_calls, 0, "");
243 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_spills, CTLFLAG_RD,
244     &moea_pte_spills, 0, "");
245 
246 /*
247  * Allocate physical memory for use in moea_bootstrap.
248  */
249 static vm_offset_t	moea_bootstrap_alloc(vm_size_t, u_int);
250 
251 /*
252  * PTE calls.
253  */
254 static int		moea_pte_insert(u_int, struct pte *);
255 
256 /*
257  * PVO calls.
258  */
259 static int	moea_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *,
260 		    vm_offset_t, vm_paddr_t, u_int, int);
261 static void	moea_pvo_remove(struct pvo_entry *, int);
262 static struct	pvo_entry *moea_pvo_find_va(pmap_t, vm_offset_t, int *);
263 static struct	pte *moea_pvo_to_pte(const struct pvo_entry *, int);
264 
265 /*
266  * Utility routines.
267  */
268 static int		moea_enter_locked(pmap_t, vm_offset_t, vm_page_t,
269 			    vm_prot_t, u_int, int8_t);
270 static void		moea_syncicache(vm_paddr_t, vm_size_t);
271 static boolean_t	moea_query_bit(vm_page_t, int);
272 static u_int		moea_clear_bit(vm_page_t, int);
273 static void		moea_kremove(vm_offset_t);
274 int		moea_pte_spill(vm_offset_t);
275 
276 /*
277  * Kernel MMU interface
278  */
279 void moea_clear_modify(vm_page_t);
280 void moea_copy_page(vm_page_t, vm_page_t);
281 void moea_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
282     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
283 int moea_enter(pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int,
284     int8_t);
285 void moea_enter_object(pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
286     vm_prot_t);
287 void moea_enter_quick(pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
288 vm_paddr_t moea_extract(pmap_t, vm_offset_t);
289 vm_page_t moea_extract_and_hold(pmap_t, vm_offset_t, vm_prot_t);
290 void moea_init(void);
291 boolean_t moea_is_modified(vm_page_t);
292 boolean_t moea_is_prefaultable(pmap_t, vm_offset_t);
293 boolean_t moea_is_referenced(vm_page_t);
294 int moea_ts_referenced(vm_page_t);
295 vm_offset_t moea_map(vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
296 static int moea_mincore(pmap_t, vm_offset_t, vm_paddr_t *);
297 boolean_t moea_page_exists_quick(pmap_t, vm_page_t);
298 void moea_page_init(vm_page_t);
299 int moea_page_wired_mappings(vm_page_t);
300 int moea_pinit(pmap_t);
301 void moea_pinit0(pmap_t);
302 void moea_protect(pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
303 void moea_qenter(vm_offset_t, vm_page_t *, int);
304 void moea_qremove(vm_offset_t, int);
305 void moea_release(pmap_t);
306 void moea_remove(pmap_t, vm_offset_t, vm_offset_t);
307 void moea_remove_all(vm_page_t);
308 void moea_remove_write(vm_page_t);
309 void moea_unwire(pmap_t, vm_offset_t, vm_offset_t);
310 void moea_zero_page(vm_page_t);
311 void moea_zero_page_area(vm_page_t, int, int);
312 void moea_activate(struct thread *);
313 void moea_deactivate(struct thread *);
314 void moea_cpu_bootstrap(int);
315 void moea_bootstrap(vm_offset_t, vm_offset_t);
316 void *moea_mapdev(vm_paddr_t, vm_size_t);
317 void *moea_mapdev_attr(vm_paddr_t, vm_size_t, vm_memattr_t);
318 void moea_unmapdev(void *, vm_size_t);
319 vm_paddr_t moea_kextract(vm_offset_t);
320 void moea_kenter_attr(vm_offset_t, vm_paddr_t, vm_memattr_t);
321 void moea_kenter(vm_offset_t, vm_paddr_t);
322 void moea_page_set_memattr(vm_page_t m, vm_memattr_t ma);
323 boolean_t moea_dev_direct_mapped(vm_paddr_t, vm_size_t);
324 static void moea_sync_icache(pmap_t, vm_offset_t, vm_size_t);
325 void moea_dumpsys_map(vm_paddr_t pa, size_t sz, void **va);
326 void moea_scan_init(void);
327 vm_offset_t moea_quick_enter_page(vm_page_t m);
328 void moea_quick_remove_page(vm_offset_t addr);
329 boolean_t moea_page_is_mapped(vm_page_t m);
330 bool moea_ps_enabled(pmap_t pmap);
331 static int moea_map_user_ptr(pmap_t pm,
332     volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen);
333 static int moea_decode_kernel_ptr(vm_offset_t addr,
334     int *is_user, vm_offset_t *decoded_addr);
335 
336 static struct pmap_funcs moea_methods = {
337 	.clear_modify = moea_clear_modify,
338 	.copy_page = moea_copy_page,
339 	.copy_pages = moea_copy_pages,
340 	.enter = moea_enter,
341 	.enter_object = moea_enter_object,
342 	.enter_quick = moea_enter_quick,
343 	.extract = moea_extract,
344 	.extract_and_hold = moea_extract_and_hold,
345 	.init = moea_init,
346 	.is_modified = moea_is_modified,
347 	.is_prefaultable = moea_is_prefaultable,
348 	.is_referenced = moea_is_referenced,
349 	.ts_referenced = moea_ts_referenced,
350 	.map =      		moea_map,
351 	.page_exists_quick = moea_page_exists_quick,
352 	.page_init = moea_page_init,
353 	.page_wired_mappings = moea_page_wired_mappings,
354 	.pinit = moea_pinit,
355 	.pinit0 = moea_pinit0,
356 	.protect = moea_protect,
357 	.qenter = moea_qenter,
358 	.qremove = moea_qremove,
359 	.release = moea_release,
360 	.remove = moea_remove,
361 	.remove_all = moea_remove_all,
362 	.mincore = moea_mincore,
363 	.remove_write = moea_remove_write,
364 	.sync_icache = moea_sync_icache,
365 	.unwire = moea_unwire,
366 	.zero_page =        	moea_zero_page,
367 	.zero_page_area = moea_zero_page_area,
368 	.activate = moea_activate,
369 	.deactivate =       	moea_deactivate,
370 	.page_set_memattr = moea_page_set_memattr,
371 	.quick_enter_page =  moea_quick_enter_page,
372 	.quick_remove_page =  moea_quick_remove_page,
373 	.page_is_mapped = moea_page_is_mapped,
374 	.ps_enabled = moea_ps_enabled,
375 
376 	/* Internal interfaces */
377 	.bootstrap =        	moea_bootstrap,
378 	.cpu_bootstrap =    	moea_cpu_bootstrap,
379 	.mapdev_attr = moea_mapdev_attr,
380 	.mapdev = moea_mapdev,
381 	.unmapdev = moea_unmapdev,
382 	.kextract = moea_kextract,
383 	.kenter = moea_kenter,
384 	.kenter_attr = moea_kenter_attr,
385 	.dev_direct_mapped = moea_dev_direct_mapped,
386 	.dumpsys_pa_init = moea_scan_init,
387 	.dumpsys_map_chunk = moea_dumpsys_map,
388 	.map_user_ptr = moea_map_user_ptr,
389 	.decode_kernel_ptr =  moea_decode_kernel_ptr,
390 };
391 
392 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods);
393 
394 static __inline uint32_t
395 moea_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
396 {
397 	uint32_t pte_lo;
398 	int i;
399 
400 	if (ma != VM_MEMATTR_DEFAULT) {
401 		switch (ma) {
402 		case VM_MEMATTR_UNCACHEABLE:
403 			return (PTE_I | PTE_G);
404 		case VM_MEMATTR_CACHEABLE:
405 			return (PTE_M);
406 		case VM_MEMATTR_WRITE_COMBINING:
407 		case VM_MEMATTR_WRITE_BACK:
408 		case VM_MEMATTR_PREFETCHABLE:
409 			return (PTE_I);
410 		case VM_MEMATTR_WRITE_THROUGH:
411 			return (PTE_W | PTE_M);
412 		}
413 	}
414 
415 	/*
416 	 * Assume the page is cache inhibited and access is guarded unless
417 	 * it's in our available memory array.
418 	 */
419 	pte_lo = PTE_I | PTE_G;
420 	for (i = 0; i < pregions_sz; i++) {
421 		if ((pa >= pregions[i].mr_start) &&
422 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
423 			pte_lo = PTE_M;
424 			break;
425 		}
426 	}
427 
428 	return pte_lo;
429 }
430 
431 /*
432  * Translate OFW translations into VM attributes.
433  */
434 static __inline vm_memattr_t
435 moea_bootstrap_convert_wimg(uint32_t mode)
436 {
437 
438 	switch (mode) {
439 	case (PTE_I | PTE_G):
440 		/* PCI device memory */
441 		return VM_MEMATTR_UNCACHEABLE;
442 	case (PTE_M):
443 		/* Explicitly coherent */
444 		return VM_MEMATTR_CACHEABLE;
445 	case 0: /* Default claim */
446 	case 2: /* Alternate PP bits set by OF for the original payload */
447 		/* "Normal" memory. */
448 		return VM_MEMATTR_DEFAULT;
449 
450 	default:
451 		/* Err on the side of caution for unknowns */
452 		/* XXX should we panic instead? */
453 		return VM_MEMATTR_UNCACHEABLE;
454 	}
455 }
456 
457 static void
458 tlbie(vm_offset_t va)
459 {
460 
461 	mtx_lock_spin(&tlbie_mtx);
462 	__asm __volatile("ptesync");
463 	__asm __volatile("tlbie %0" :: "r"(va));
464 	__asm __volatile("eieio; tlbsync; ptesync");
465 	mtx_unlock_spin(&tlbie_mtx);
466 }
467 
468 static void
469 tlbia(void)
470 {
471 	vm_offset_t va;
472 
473 	for (va = 0; va < 0x00040000; va += 0x00001000) {
474 		__asm __volatile("tlbie %0" :: "r"(va));
475 		powerpc_sync();
476 	}
477 	__asm __volatile("tlbsync");
478 	powerpc_sync();
479 }
480 
481 static __inline int
482 va_to_sr(u_int *sr, vm_offset_t va)
483 {
484 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
485 }
486 
487 static __inline u_int
488 va_to_pteg(u_int sr, vm_offset_t addr)
489 {
490 	u_int hash;
491 
492 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
493 	    ADDR_PIDX_SHFT);
494 	return (hash & moea_pteg_mask);
495 }
496 
497 static __inline struct pvo_head *
498 vm_page_to_pvoh(vm_page_t m)
499 {
500 
501 	return (&m->md.mdpg_pvoh);
502 }
503 
504 static __inline void
505 moea_attr_clear(vm_page_t m, int ptebit)
506 {
507 
508 	rw_assert(&pvh_global_lock, RA_WLOCKED);
509 	m->md.mdpg_attrs &= ~ptebit;
510 }
511 
512 static __inline int
513 moea_attr_fetch(vm_page_t m)
514 {
515 
516 	return (m->md.mdpg_attrs);
517 }
518 
519 static __inline void
520 moea_attr_save(vm_page_t m, int ptebit)
521 {
522 
523 	rw_assert(&pvh_global_lock, RA_WLOCKED);
524 	m->md.mdpg_attrs |= ptebit;
525 }
526 
527 static __inline int
528 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
529 {
530 	if (pt->pte_hi == pvo_pt->pte_hi)
531 		return (1);
532 
533 	return (0);
534 }
535 
536 static __inline int
537 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
538 {
539 	return (pt->pte_hi & ~PTE_VALID) ==
540 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
541 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
542 }
543 
544 static __inline void
545 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
546 {
547 
548 	mtx_assert(&moea_table_mutex, MA_OWNED);
549 
550 	/*
551 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
552 	 * set when the real pte is set in memory.
553 	 *
554 	 * Note: Don't set the valid bit for correct operation of tlb update.
555 	 */
556 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
557 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
558 	pt->pte_lo = pte_lo;
559 }
560 
561 static __inline void
562 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
563 {
564 
565 	mtx_assert(&moea_table_mutex, MA_OWNED);
566 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
567 }
568 
569 static __inline void
570 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
571 {
572 
573 	mtx_assert(&moea_table_mutex, MA_OWNED);
574 
575 	/*
576 	 * As shown in Section 7.6.3.2.3
577 	 */
578 	pt->pte_lo &= ~ptebit;
579 	tlbie(va);
580 }
581 
582 static __inline void
583 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
584 {
585 
586 	mtx_assert(&moea_table_mutex, MA_OWNED);
587 	pvo_pt->pte_hi |= PTE_VALID;
588 
589 	/*
590 	 * Update the PTE as defined in section 7.6.3.1.
591 	 * Note that the REF/CHG bits are from pvo_pt and thus should have
592 	 * been saved so this routine can restore them (if desired).
593 	 */
594 	pt->pte_lo = pvo_pt->pte_lo;
595 	powerpc_sync();
596 	pt->pte_hi = pvo_pt->pte_hi;
597 	powerpc_sync();
598 	moea_pte_valid++;
599 }
600 
601 static __inline void
602 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
603 {
604 
605 	mtx_assert(&moea_table_mutex, MA_OWNED);
606 	pvo_pt->pte_hi &= ~PTE_VALID;
607 
608 	/*
609 	 * Force the reg & chg bits back into the PTEs.
610 	 */
611 	powerpc_sync();
612 
613 	/*
614 	 * Invalidate the pte.
615 	 */
616 	pt->pte_hi &= ~PTE_VALID;
617 
618 	tlbie(va);
619 
620 	/*
621 	 * Save the reg & chg bits.
622 	 */
623 	moea_pte_synch(pt, pvo_pt);
624 	moea_pte_valid--;
625 }
626 
627 static __inline void
628 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
629 {
630 
631 	/*
632 	 * Invalidate the PTE
633 	 */
634 	moea_pte_unset(pt, pvo_pt, va);
635 	moea_pte_set(pt, pvo_pt);
636 }
637 
638 /*
639  * Quick sort callout for comparing memory regions.
640  */
641 static int	om_cmp(const void *a, const void *b);
642 
643 static int
644 om_cmp(const void *a, const void *b)
645 {
646 	const struct	ofw_map *mapa;
647 	const struct	ofw_map *mapb;
648 
649 	mapa = a;
650 	mapb = b;
651 	if (mapa->om_pa < mapb->om_pa)
652 		return (-1);
653 	else if (mapa->om_pa > mapb->om_pa)
654 		return (1);
655 	else
656 		return (0);
657 }
658 
659 void
660 moea_cpu_bootstrap(int ap)
661 {
662 	u_int sdr;
663 	int i;
664 
665 	if (ap) {
666 		powerpc_sync();
667 		__asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu));
668 		__asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl));
669 		isync();
670 		__asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu));
671 		__asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl));
672 		isync();
673 	}
674 
675 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
676 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
677 	isync();
678 
679 	__asm __volatile("mtibatu 1,%0" :: "r"(0));
680 	__asm __volatile("mtdbatu 2,%0" :: "r"(0));
681 	__asm __volatile("mtibatu 2,%0" :: "r"(0));
682 	__asm __volatile("mtdbatu 3,%0" :: "r"(0));
683 	__asm __volatile("mtibatu 3,%0" :: "r"(0));
684 	isync();
685 
686 	for (i = 0; i < 16; i++)
687 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
688 	powerpc_sync();
689 
690 	sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10);
691 	__asm __volatile("mtsdr1 %0" :: "r"(sdr));
692 	isync();
693 
694 	tlbia();
695 }
696 
697 void
698 moea_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
699 {
700 	ihandle_t	mmui;
701 	phandle_t	chosen, mmu;
702 	int		sz;
703 	int		i, j;
704 	vm_size_t	size, physsz, hwphyssz;
705 	vm_offset_t	pa, va, off;
706 	void		*dpcpu;
707 
708 	/*
709 	 * Map PCI memory space.
710 	 */
711 	battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
712 	battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
713 
714 	battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
715 	battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
716 
717 	battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
718 	battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
719 
720 	battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
721 	battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
722 
723 	powerpc_sync();
724 
725 	/* map pci space */
726 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
727 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
728 	isync();
729 
730 	/* set global direct map flag */
731 	hw_direct_map = 1;
732 
733 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
734 	CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
735 
736 	for (i = 0; i < pregions_sz; i++) {
737 		vm_offset_t pa;
738 		vm_offset_t end;
739 
740 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
741 			pregions[i].mr_start,
742 			pregions[i].mr_start + pregions[i].mr_size,
743 			pregions[i].mr_size);
744 		/*
745 		 * Install entries into the BAT table to allow all
746 		 * of physmem to be convered by on-demand BAT entries.
747 		 * The loop will sometimes set the same battable element
748 		 * twice, but that's fine since they won't be used for
749 		 * a while yet.
750 		 */
751 		pa = pregions[i].mr_start & 0xf0000000;
752 		end = pregions[i].mr_start + pregions[i].mr_size;
753 		do {
754                         u_int n = pa >> ADDR_SR_SHFT;
755 
756 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
757 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
758 			pa += SEGMENT_LENGTH;
759 		} while (pa < end);
760 	}
761 
762 	if (PHYS_AVAIL_ENTRIES < regions_sz)
763 		panic("moea_bootstrap: phys_avail too small");
764 
765 	phys_avail_count = 0;
766 	physsz = 0;
767 	hwphyssz = 0;
768 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
769 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
770 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
771 		    regions[i].mr_start + regions[i].mr_size,
772 		    regions[i].mr_size);
773 		if (hwphyssz != 0 &&
774 		    (physsz + regions[i].mr_size) >= hwphyssz) {
775 			if (physsz < hwphyssz) {
776 				phys_avail[j] = regions[i].mr_start;
777 				phys_avail[j + 1] = regions[i].mr_start +
778 				    hwphyssz - physsz;
779 				physsz = hwphyssz;
780 				phys_avail_count++;
781 			}
782 			break;
783 		}
784 		phys_avail[j] = regions[i].mr_start;
785 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
786 		phys_avail_count++;
787 		physsz += regions[i].mr_size;
788 	}
789 
790 	/* Check for overlap with the kernel and exception vectors */
791 	for (j = 0; j < 2*phys_avail_count; j+=2) {
792 		if (phys_avail[j] < EXC_LAST)
793 			phys_avail[j] += EXC_LAST;
794 
795 		if (kernelstart >= phys_avail[j] &&
796 		    kernelstart < phys_avail[j+1]) {
797 			if (kernelend < phys_avail[j+1]) {
798 				phys_avail[2*phys_avail_count] =
799 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
800 				phys_avail[2*phys_avail_count + 1] =
801 				    phys_avail[j+1];
802 				phys_avail_count++;
803 			}
804 
805 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
806 		}
807 
808 		if (kernelend >= phys_avail[j] &&
809 		    kernelend < phys_avail[j+1]) {
810 			if (kernelstart > phys_avail[j]) {
811 				phys_avail[2*phys_avail_count] = phys_avail[j];
812 				phys_avail[2*phys_avail_count + 1] =
813 				    kernelstart & ~PAGE_MASK;
814 				phys_avail_count++;
815 			}
816 
817 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
818 		}
819 	}
820 
821 	physmem = btoc(physsz);
822 
823 	/*
824 	 * Allocate PTEG table.
825 	 */
826 #ifdef PTEGCOUNT
827 	moea_pteg_count = PTEGCOUNT;
828 #else
829 	moea_pteg_count = 0x1000;
830 
831 	while (moea_pteg_count < physmem)
832 		moea_pteg_count <<= 1;
833 
834 	moea_pteg_count >>= 1;
835 #endif /* PTEGCOUNT */
836 
837 	size = moea_pteg_count * sizeof(struct pteg);
838 	CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
839 	    size);
840 	moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
841 	CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
842 	bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
843 	moea_pteg_mask = moea_pteg_count - 1;
844 
845 	/*
846 	 * Allocate pv/overflow lists.
847 	 */
848 	size = sizeof(struct pvo_head) * moea_pteg_count;
849 	moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
850 	    PAGE_SIZE);
851 	CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
852 	for (i = 0; i < moea_pteg_count; i++)
853 		LIST_INIT(&moea_pvo_table[i]);
854 
855 	/*
856 	 * Initialize the lock that synchronizes access to the pteg and pvo
857 	 * tables.
858 	 */
859 	mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF |
860 	    MTX_RECURSE);
861 	mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF);
862 
863 	mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN);
864 
865 	/*
866 	 * Initialise the unmanaged pvo pool.
867 	 */
868 	moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
869 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
870 	moea_bpvo_pool_index = 0;
871 
872 	/*
873 	 * Make sure kernel vsid is allocated as well as VSID 0.
874 	 */
875 	moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
876 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
877 	moea_vsid_bitmap[0] |= 1;
878 
879 	/*
880 	 * Initialize the kernel pmap (which is statically allocated).
881 	 */
882 	PMAP_LOCK_INIT(kernel_pmap);
883 	for (i = 0; i < 16; i++)
884 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
885 	CPU_FILL(&kernel_pmap->pm_active);
886 	RB_INIT(&kernel_pmap->pmap_pvo);
887 
888  	/*
889 	 * Initialize the global pv list lock.
890 	 */
891 	rw_init(&pvh_global_lock, "pmap pv global");
892 
893 	/*
894 	 * Set up the Open Firmware mappings
895 	 */
896 	chosen = OF_finddevice("/chosen");
897 	if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1 &&
898 	    (mmu = OF_instance_to_package(mmui)) != -1 &&
899 	    (sz = OF_getproplen(mmu, "translations")) != -1) {
900 		translations = NULL;
901 		for (i = 0; phys_avail[i] != 0; i += 2) {
902 			if (phys_avail[i + 1] >= sz) {
903 				translations = (struct ofw_map *)phys_avail[i];
904 				break;
905 			}
906 		}
907 		if (translations == NULL)
908 			panic("moea_bootstrap: no space to copy translations");
909 		bzero(translations, sz);
910 		if (OF_getprop(mmu, "translations", translations, sz) == -1)
911 			panic("moea_bootstrap: can't get ofw translations");
912 		CTR0(KTR_PMAP, "moea_bootstrap: translations");
913 		sz /= sizeof(*translations);
914 		qsort(translations, sz, sizeof (*translations), om_cmp);
915 		for (i = 0; i < sz; i++) {
916 			CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
917 			    translations[i].om_pa, translations[i].om_va,
918 			    translations[i].om_len);
919 
920 			/*
921 			 * If the mapping is 1:1, let the RAM and device
922 			 * on-demand BAT tables take care of the translation.
923 			 *
924 			 * However, always enter mappings for segment 16,
925 			 * which is mixed-protection and therefore not
926 			 * compatible with a BAT entry.
927 			 */
928 			if ((translations[i].om_va >> ADDR_SR_SHFT) != 0xf &&
929 				translations[i].om_va == translations[i].om_pa)
930 					continue;
931 
932 			/* Enter the pages */
933 			for (off = 0; off < translations[i].om_len;
934 			    off += PAGE_SIZE)
935 				moea_kenter_attr(translations[i].om_va + off,
936 				    translations[i].om_pa + off,
937 				    moea_bootstrap_convert_wimg(translations[i].om_mode));
938 		}
939 	}
940 
941 	/*
942 	 * Calculate the last available physical address.
943 	 */
944 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
945 		;
946 	Maxmem = powerpc_btop(phys_avail[i + 1]);
947 
948 	moea_cpu_bootstrap(0);
949 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
950 	pmap_bootstrapped++;
951 
952 	/*
953 	 * Set the start and end of kva.
954 	 */
955 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
956 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
957 
958 	/*
959 	 * Allocate a kernel stack with a guard page for thread0 and map it
960 	 * into the kernel page map.
961 	 */
962 	pa = moea_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
963 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
964 	virtual_avail = va + kstack_pages * PAGE_SIZE;
965 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
966 	thread0.td_kstack = va;
967 	thread0.td_kstack_pages = kstack_pages;
968 	for (i = 0; i < kstack_pages; i++) {
969 		moea_kenter(va, pa);
970 		pa += PAGE_SIZE;
971 		va += PAGE_SIZE;
972 	}
973 
974 	/*
975 	 * Allocate virtual address space for the message buffer.
976 	 */
977 	pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE);
978 	msgbufp = (struct msgbuf *)virtual_avail;
979 	va = virtual_avail;
980 	virtual_avail += round_page(msgbufsize);
981 	while (va < virtual_avail) {
982 		moea_kenter(va, pa);
983 		pa += PAGE_SIZE;
984 		va += PAGE_SIZE;
985 	}
986 
987 	/*
988 	 * Allocate virtual address space for the dynamic percpu area.
989 	 */
990 	pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
991 	dpcpu = (void *)virtual_avail;
992 	va = virtual_avail;
993 	virtual_avail += DPCPU_SIZE;
994 	while (va < virtual_avail) {
995 		moea_kenter(va, pa);
996 		pa += PAGE_SIZE;
997 		va += PAGE_SIZE;
998 	}
999 	dpcpu_init(dpcpu, 0);
1000 }
1001 
1002 /*
1003  * Activate a user pmap.  The pmap must be activated before it's address
1004  * space can be accessed in any way.
1005  */
1006 void
1007 moea_activate(struct thread *td)
1008 {
1009 	pmap_t	pm, pmr;
1010 
1011 	/*
1012 	 * Load all the data we need up front to encourage the compiler to
1013 	 * not issue any loads while we have interrupts disabled below.
1014 	 */
1015 	pm = &td->td_proc->p_vmspace->vm_pmap;
1016 	pmr = pm->pmap_phys;
1017 
1018 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1019 	PCPU_SET(curpmap, pmr);
1020 
1021 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1022 }
1023 
1024 void
1025 moea_deactivate(struct thread *td)
1026 {
1027 	pmap_t	pm;
1028 
1029 	pm = &td->td_proc->p_vmspace->vm_pmap;
1030 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1031 	PCPU_SET(curpmap, NULL);
1032 }
1033 
1034 void
1035 moea_unwire(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1036 {
1037 	struct	pvo_entry key, *pvo;
1038 
1039 	PMAP_LOCK(pm);
1040 	key.pvo_vaddr = sva;
1041 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1042 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1043 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1044 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1045 			panic("moea_unwire: pvo %p is missing PVO_WIRED", pvo);
1046 		pvo->pvo_vaddr &= ~PVO_WIRED;
1047 		pm->pm_stats.wired_count--;
1048 	}
1049 	PMAP_UNLOCK(pm);
1050 }
1051 
1052 void
1053 moea_copy_page(vm_page_t msrc, vm_page_t mdst)
1054 {
1055 	vm_offset_t	dst;
1056 	vm_offset_t	src;
1057 
1058 	dst = VM_PAGE_TO_PHYS(mdst);
1059 	src = VM_PAGE_TO_PHYS(msrc);
1060 
1061 	bcopy((void *)src, (void *)dst, PAGE_SIZE);
1062 }
1063 
1064 void
1065 moea_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
1066     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1067 {
1068 	void *a_cp, *b_cp;
1069 	vm_offset_t a_pg_offset, b_pg_offset;
1070 	int cnt;
1071 
1072 	while (xfersize > 0) {
1073 		a_pg_offset = a_offset & PAGE_MASK;
1074 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1075 		a_cp = (char *)VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]) +
1076 		    a_pg_offset;
1077 		b_pg_offset = b_offset & PAGE_MASK;
1078 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1079 		b_cp = (char *)VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]) +
1080 		    b_pg_offset;
1081 		bcopy(a_cp, b_cp, cnt);
1082 		a_offset += cnt;
1083 		b_offset += cnt;
1084 		xfersize -= cnt;
1085 	}
1086 }
1087 
1088 /*
1089  * Zero a page of physical memory by temporarily mapping it into the tlb.
1090  */
1091 void
1092 moea_zero_page(vm_page_t m)
1093 {
1094 	vm_offset_t off, pa = VM_PAGE_TO_PHYS(m);
1095 
1096 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1097 		__asm __volatile("dcbz 0,%0" :: "r"(pa + off));
1098 }
1099 
1100 void
1101 moea_zero_page_area(vm_page_t m, int off, int size)
1102 {
1103 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1104 	void *va = (void *)(pa + off);
1105 
1106 	bzero(va, size);
1107 }
1108 
1109 vm_offset_t
1110 moea_quick_enter_page(vm_page_t m)
1111 {
1112 
1113 	return (VM_PAGE_TO_PHYS(m));
1114 }
1115 
1116 void
1117 moea_quick_remove_page(vm_offset_t addr)
1118 {
1119 }
1120 
1121 boolean_t
1122 moea_page_is_mapped(vm_page_t m)
1123 {
1124 	return (!LIST_EMPTY(&(m)->md.mdpg_pvoh));
1125 }
1126 
1127 bool
1128 moea_ps_enabled(pmap_t pmap __unused)
1129 {
1130 	return (false);
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 int
1139 moea_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1140     u_int flags, int8_t psind)
1141 {
1142 	int error;
1143 
1144 	for (;;) {
1145 		rw_wlock(&pvh_global_lock);
1146 		PMAP_LOCK(pmap);
1147 		error = moea_enter_locked(pmap, va, m, prot, flags, psind);
1148 		rw_wunlock(&pvh_global_lock);
1149 		PMAP_UNLOCK(pmap);
1150 		if (error != ENOMEM)
1151 			return (KERN_SUCCESS);
1152 		if ((flags & PMAP_ENTER_NOSLEEP) != 0)
1153 			return (KERN_RESOURCE_SHORTAGE);
1154 		VM_OBJECT_ASSERT_UNLOCKED(m->object);
1155 		vm_wait(NULL);
1156 	}
1157 }
1158 
1159 /*
1160  * Map the given physical page at the specified virtual address in the
1161  * target pmap with the protection requested.  If specified the page
1162  * will be wired down.
1163  *
1164  * The global pvh and pmap must be locked.
1165  */
1166 static int
1167 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1168     u_int flags, int8_t psind __unused)
1169 {
1170 	struct		pvo_head *pvo_head;
1171 	uma_zone_t	zone;
1172 	u_int		pte_lo, pvo_flags;
1173 	int		error;
1174 
1175 	if (pmap_bootstrapped)
1176 		rw_assert(&pvh_global_lock, RA_WLOCKED);
1177 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1178 	if ((m->oflags & VPO_UNMANAGED) == 0) {
1179 		if ((flags & PMAP_ENTER_QUICK_LOCKED) == 0)
1180 			VM_PAGE_OBJECT_BUSY_ASSERT(m);
1181 		else
1182 			VM_OBJECT_ASSERT_LOCKED(m->object);
1183 	}
1184 
1185 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea_initialized) {
1186 		pvo_head = &moea_pvo_kunmanaged;
1187 		zone = moea_upvo_zone;
1188 		pvo_flags = 0;
1189 	} else {
1190 		pvo_head = vm_page_to_pvoh(m);
1191 		zone = moea_mpvo_zone;
1192 		pvo_flags = PVO_MANAGED;
1193 	}
1194 
1195 	pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1196 
1197 	if (prot & VM_PROT_WRITE) {
1198 		pte_lo |= PTE_BW;
1199 		if (pmap_bootstrapped &&
1200 		    (m->oflags & VPO_UNMANAGED) == 0)
1201 			vm_page_aflag_set(m, PGA_WRITEABLE);
1202 	} else
1203 		pte_lo |= PTE_BR;
1204 
1205 	if ((flags & PMAP_ENTER_WIRED) != 0)
1206 		pvo_flags |= PVO_WIRED;
1207 
1208 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1209 	    pte_lo, pvo_flags);
1210 
1211 	/*
1212 	 * Flush the real page from the instruction cache. This has be done
1213 	 * for all user mappings to prevent information leakage via the
1214 	 * instruction cache. moea_pvo_enter() returns ENOENT for the first
1215 	 * mapping for a page.
1216 	 */
1217 	if (pmap != kernel_pmap && error == ENOENT &&
1218 	    (pte_lo & (PTE_I | PTE_G)) == 0)
1219 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1220 
1221 	return (error);
1222 }
1223 
1224 /*
1225  * Maps a sequence of resident pages belonging to the same object.
1226  * The sequence begins with the given page m_start.  This page is
1227  * mapped at the given virtual address start.  Each subsequent page is
1228  * mapped at a virtual address that is offset from start by the same
1229  * amount as the page is offset from m_start within the object.  The
1230  * last page in the sequence is the page with the largest offset from
1231  * m_start that can be mapped at a virtual address less than the given
1232  * virtual address end.  Not every virtual page between start and end
1233  * is mapped; only those for which a resident page exists with the
1234  * corresponding offset from m_start are mapped.
1235  */
1236 void
1237 moea_enter_object(pmap_t pm, vm_offset_t start, vm_offset_t end,
1238     vm_page_t m_start, vm_prot_t prot)
1239 {
1240 	vm_page_t m;
1241 	vm_pindex_t diff, psize;
1242 
1243 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1244 
1245 	psize = atop(end - start);
1246 	m = m_start;
1247 	rw_wlock(&pvh_global_lock);
1248 	PMAP_LOCK(pm);
1249 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1250 		moea_enter_locked(pm, start + ptoa(diff), m, prot &
1251 		    (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_QUICK_LOCKED,
1252 		    0);
1253 		m = TAILQ_NEXT(m, listq);
1254 	}
1255 	rw_wunlock(&pvh_global_lock);
1256 	PMAP_UNLOCK(pm);
1257 }
1258 
1259 void
1260 moea_enter_quick(pmap_t pm, vm_offset_t va, vm_page_t m,
1261     vm_prot_t prot)
1262 {
1263 
1264 	rw_wlock(&pvh_global_lock);
1265 	PMAP_LOCK(pm);
1266 	moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1267 	    PMAP_ENTER_QUICK_LOCKED, 0);
1268 	rw_wunlock(&pvh_global_lock);
1269 	PMAP_UNLOCK(pm);
1270 }
1271 
1272 vm_paddr_t
1273 moea_extract(pmap_t pm, vm_offset_t va)
1274 {
1275 	struct	pvo_entry *pvo;
1276 	vm_paddr_t pa;
1277 
1278 	PMAP_LOCK(pm);
1279 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1280 	if (pvo == NULL)
1281 		pa = 0;
1282 	else
1283 		pa = PVO_PADDR(pvo) | (va & ADDR_POFF);
1284 	PMAP_UNLOCK(pm);
1285 	return (pa);
1286 }
1287 
1288 /*
1289  * Atomically extract and hold the physical page with the given
1290  * pmap and virtual address pair if that mapping permits the given
1291  * protection.
1292  */
1293 vm_page_t
1294 moea_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1295 {
1296 	struct	pvo_entry *pvo;
1297 	vm_page_t m;
1298 
1299 	m = NULL;
1300 	PMAP_LOCK(pmap);
1301 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1302 	if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1303 	    ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1304 	     (prot & VM_PROT_WRITE) == 0)) {
1305 		m = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
1306 		if (!vm_page_wire_mapped(m))
1307 			m = NULL;
1308 	}
1309 	PMAP_UNLOCK(pmap);
1310 	return (m);
1311 }
1312 
1313 void
1314 moea_init()
1315 {
1316 
1317 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1318 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1319 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1320 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1321 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1322 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1323 	moea_initialized = TRUE;
1324 }
1325 
1326 boolean_t
1327 moea_is_referenced(vm_page_t m)
1328 {
1329 	boolean_t rv;
1330 
1331 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1332 	    ("moea_is_referenced: page %p is not managed", m));
1333 	rw_wlock(&pvh_global_lock);
1334 	rv = moea_query_bit(m, PTE_REF);
1335 	rw_wunlock(&pvh_global_lock);
1336 	return (rv);
1337 }
1338 
1339 boolean_t
1340 moea_is_modified(vm_page_t m)
1341 {
1342 	boolean_t rv;
1343 
1344 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1345 	    ("moea_is_modified: page %p is not managed", m));
1346 
1347 	/*
1348 	 * If the page is not busied then this check is racy.
1349 	 */
1350 	if (!pmap_page_is_write_mapped(m))
1351 		return (FALSE);
1352 
1353 	rw_wlock(&pvh_global_lock);
1354 	rv = moea_query_bit(m, PTE_CHG);
1355 	rw_wunlock(&pvh_global_lock);
1356 	return (rv);
1357 }
1358 
1359 boolean_t
1360 moea_is_prefaultable(pmap_t pmap, vm_offset_t va)
1361 {
1362 	struct pvo_entry *pvo;
1363 	boolean_t rv;
1364 
1365 	PMAP_LOCK(pmap);
1366 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1367 	rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1368 	PMAP_UNLOCK(pmap);
1369 	return (rv);
1370 }
1371 
1372 void
1373 moea_clear_modify(vm_page_t m)
1374 {
1375 
1376 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1377 	    ("moea_clear_modify: page %p is not managed", m));
1378 	vm_page_assert_busied(m);
1379 
1380 	if (!pmap_page_is_write_mapped(m))
1381 		return;
1382 	rw_wlock(&pvh_global_lock);
1383 	moea_clear_bit(m, PTE_CHG);
1384 	rw_wunlock(&pvh_global_lock);
1385 }
1386 
1387 /*
1388  * Clear the write and modified bits in each of the given page's mappings.
1389  */
1390 void
1391 moea_remove_write(vm_page_t m)
1392 {
1393 	struct	pvo_entry *pvo;
1394 	struct	pte *pt;
1395 	pmap_t	pmap;
1396 	u_int	lo;
1397 
1398 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1399 	    ("moea_remove_write: page %p is not managed", m));
1400 	vm_page_assert_busied(m);
1401 
1402 	if (!pmap_page_is_write_mapped(m))
1403 		return;
1404 	rw_wlock(&pvh_global_lock);
1405 	lo = moea_attr_fetch(m);
1406 	powerpc_sync();
1407 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1408 		pmap = pvo->pvo_pmap;
1409 		PMAP_LOCK(pmap);
1410 		if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1411 			pt = moea_pvo_to_pte(pvo, -1);
1412 			pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1413 			pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1414 			if (pt != NULL) {
1415 				moea_pte_synch(pt, &pvo->pvo_pte.pte);
1416 				lo |= pvo->pvo_pte.pte.pte_lo;
1417 				pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1418 				moea_pte_change(pt, &pvo->pvo_pte.pte,
1419 				    pvo->pvo_vaddr);
1420 				mtx_unlock(&moea_table_mutex);
1421 			}
1422 		}
1423 		PMAP_UNLOCK(pmap);
1424 	}
1425 	if ((lo & PTE_CHG) != 0) {
1426 		moea_attr_clear(m, PTE_CHG);
1427 		vm_page_dirty(m);
1428 	}
1429 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1430 	rw_wunlock(&pvh_global_lock);
1431 }
1432 
1433 /*
1434  *	moea_ts_referenced:
1435  *
1436  *	Return a count of reference bits for a page, clearing those bits.
1437  *	It is not necessary for every reference bit to be cleared, but it
1438  *	is necessary that 0 only be returned when there are truly no
1439  *	reference bits set.
1440  *
1441  *	XXX: The exact number of bits to check and clear is a matter that
1442  *	should be tested and standardized at some point in the future for
1443  *	optimal aging of shared pages.
1444  */
1445 int
1446 moea_ts_referenced(vm_page_t m)
1447 {
1448 	int count;
1449 
1450 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1451 	    ("moea_ts_referenced: page %p is not managed", m));
1452 	rw_wlock(&pvh_global_lock);
1453 	count = moea_clear_bit(m, PTE_REF);
1454 	rw_wunlock(&pvh_global_lock);
1455 	return (count);
1456 }
1457 
1458 /*
1459  * Modify the WIMG settings of all mappings for a page.
1460  */
1461 void
1462 moea_page_set_memattr(vm_page_t m, vm_memattr_t ma)
1463 {
1464 	struct	pvo_entry *pvo;
1465 	struct	pvo_head *pvo_head;
1466 	struct	pte *pt;
1467 	pmap_t	pmap;
1468 	u_int	lo;
1469 
1470 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1471 		m->md.mdpg_cache_attrs = ma;
1472 		return;
1473 	}
1474 
1475 	rw_wlock(&pvh_global_lock);
1476 	pvo_head = vm_page_to_pvoh(m);
1477 	lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1478 
1479 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1480 		pmap = pvo->pvo_pmap;
1481 		PMAP_LOCK(pmap);
1482 		pt = moea_pvo_to_pte(pvo, -1);
1483 		pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1484 		pvo->pvo_pte.pte.pte_lo |= lo;
1485 		if (pt != NULL) {
1486 			moea_pte_change(pt, &pvo->pvo_pte.pte,
1487 			    pvo->pvo_vaddr);
1488 			if (pvo->pvo_pmap == kernel_pmap)
1489 				isync();
1490 		}
1491 		mtx_unlock(&moea_table_mutex);
1492 		PMAP_UNLOCK(pmap);
1493 	}
1494 	m->md.mdpg_cache_attrs = ma;
1495 	rw_wunlock(&pvh_global_lock);
1496 }
1497 
1498 /*
1499  * Map a wired page into kernel virtual address space.
1500  */
1501 void
1502 moea_kenter(vm_offset_t va, vm_paddr_t pa)
1503 {
1504 
1505 	moea_kenter_attr(va, pa, VM_MEMATTR_DEFAULT);
1506 }
1507 
1508 void
1509 moea_kenter_attr(vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
1510 {
1511 	u_int		pte_lo;
1512 	int		error;
1513 
1514 #if 0
1515 	if (va < VM_MIN_KERNEL_ADDRESS)
1516 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1517 		    va);
1518 #endif
1519 
1520 	pte_lo = moea_calc_wimg(pa, ma);
1521 
1522 	PMAP_LOCK(kernel_pmap);
1523 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1524 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1525 
1526 	if (error != 0 && error != ENOENT)
1527 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1528 		    pa, error);
1529 
1530 	PMAP_UNLOCK(kernel_pmap);
1531 }
1532 
1533 /*
1534  * Extract the physical page address associated with the given kernel virtual
1535  * address.
1536  */
1537 vm_paddr_t
1538 moea_kextract(vm_offset_t va)
1539 {
1540 	struct		pvo_entry *pvo;
1541 	vm_paddr_t pa;
1542 
1543 	/*
1544 	 * Allow direct mappings on 32-bit OEA
1545 	 */
1546 	if (va < VM_MIN_KERNEL_ADDRESS) {
1547 		return (va);
1548 	}
1549 
1550 	PMAP_LOCK(kernel_pmap);
1551 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1552 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1553 	pa = PVO_PADDR(pvo) | (va & ADDR_POFF);
1554 	PMAP_UNLOCK(kernel_pmap);
1555 	return (pa);
1556 }
1557 
1558 /*
1559  * Remove a wired page from kernel virtual address space.
1560  */
1561 void
1562 moea_kremove(vm_offset_t va)
1563 {
1564 
1565 	moea_remove(kernel_pmap, va, va + PAGE_SIZE);
1566 }
1567 
1568 /*
1569  * Provide a kernel pointer corresponding to a given userland pointer.
1570  * The returned pointer is valid until the next time this function is
1571  * called in this thread. This is used internally in copyin/copyout.
1572  */
1573 int
1574 moea_map_user_ptr(pmap_t pm, volatile const void *uaddr,
1575     void **kaddr, size_t ulen, size_t *klen)
1576 {
1577 	size_t l;
1578 	register_t vsid;
1579 
1580 	*kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
1581 	l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
1582 	if (l > ulen)
1583 		l = ulen;
1584 	if (klen)
1585 		*klen = l;
1586 	else if (l != ulen)
1587 		return (EFAULT);
1588 
1589 	vsid = va_to_vsid(pm, (vm_offset_t)uaddr);
1590 
1591 	/* Mark segment no-execute */
1592 	vsid |= SR_N;
1593 
1594 	/* If we have already set this VSID, we can just return */
1595 	if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == vsid)
1596 		return (0);
1597 
1598 	__asm __volatile("isync");
1599 	curthread->td_pcb->pcb_cpu.aim.usr_segm =
1600 	    (uintptr_t)uaddr >> ADDR_SR_SHFT;
1601 	curthread->td_pcb->pcb_cpu.aim.usr_vsid = vsid;
1602 	__asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(vsid));
1603 
1604 	return (0);
1605 }
1606 
1607 /*
1608  * Figure out where a given kernel pointer (usually in a fault) points
1609  * to from the VM's perspective, potentially remapping into userland's
1610  * address space.
1611  */
1612 static int
1613 moea_decode_kernel_ptr(vm_offset_t addr, int *is_user,
1614     vm_offset_t *decoded_addr)
1615 {
1616 	vm_offset_t user_sr;
1617 
1618 	if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
1619 		user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
1620 		addr &= ADDR_PIDX | ADDR_POFF;
1621 		addr |= user_sr << ADDR_SR_SHFT;
1622 		*decoded_addr = addr;
1623 		*is_user = 1;
1624 	} else {
1625 		*decoded_addr = addr;
1626 		*is_user = 0;
1627 	}
1628 
1629 	return (0);
1630 }
1631 
1632 /*
1633  * Map a range of physical addresses into kernel virtual address space.
1634  *
1635  * The value passed in *virt is a suggested virtual address for the mapping.
1636  * Architectures which can support a direct-mapped physical to virtual region
1637  * can return the appropriate address within that region, leaving '*virt'
1638  * unchanged.  We cannot and therefore do not; *virt is updated with the
1639  * first usable address after the mapped region.
1640  */
1641 vm_offset_t
1642 moea_map(vm_offset_t *virt, vm_paddr_t pa_start,
1643     vm_paddr_t pa_end, int prot)
1644 {
1645 	vm_offset_t	sva, va;
1646 
1647 	sva = *virt;
1648 	va = sva;
1649 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1650 		moea_kenter(va, pa_start);
1651 	*virt = va;
1652 	return (sva);
1653 }
1654 
1655 /*
1656  * Returns true if the pmap's pv is one of the first
1657  * 16 pvs linked to from this page.  This count may
1658  * be changed upwards or downwards in the future; it
1659  * is only necessary that true be returned for a small
1660  * subset of pmaps for proper page aging.
1661  */
1662 boolean_t
1663 moea_page_exists_quick(pmap_t pmap, vm_page_t m)
1664 {
1665         int loops;
1666 	struct pvo_entry *pvo;
1667 	boolean_t rv;
1668 
1669 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1670 	    ("moea_page_exists_quick: page %p is not managed", m));
1671 	loops = 0;
1672 	rv = FALSE;
1673 	rw_wlock(&pvh_global_lock);
1674 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1675 		if (pvo->pvo_pmap == pmap) {
1676 			rv = TRUE;
1677 			break;
1678 		}
1679 		if (++loops >= 16)
1680 			break;
1681 	}
1682 	rw_wunlock(&pvh_global_lock);
1683 	return (rv);
1684 }
1685 
1686 void
1687 moea_page_init(vm_page_t m)
1688 {
1689 
1690 	m->md.mdpg_attrs = 0;
1691 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
1692 	LIST_INIT(&m->md.mdpg_pvoh);
1693 }
1694 
1695 /*
1696  * Return the number of managed mappings to the given physical page
1697  * that are wired.
1698  */
1699 int
1700 moea_page_wired_mappings(vm_page_t m)
1701 {
1702 	struct pvo_entry *pvo;
1703 	int count;
1704 
1705 	count = 0;
1706 	if ((m->oflags & VPO_UNMANAGED) != 0)
1707 		return (count);
1708 	rw_wlock(&pvh_global_lock);
1709 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1710 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1711 			count++;
1712 	rw_wunlock(&pvh_global_lock);
1713 	return (count);
1714 }
1715 
1716 static u_int	moea_vsidcontext;
1717 
1718 int
1719 moea_pinit(pmap_t pmap)
1720 {
1721 	int	i, mask;
1722 	u_int	entropy;
1723 
1724 	RB_INIT(&pmap->pmap_pvo);
1725 
1726 	entropy = 0;
1727 	__asm __volatile("mftb %0" : "=r"(entropy));
1728 
1729 	if ((pmap->pmap_phys = (pmap_t)moea_kextract((vm_offset_t)pmap))
1730 	    == NULL) {
1731 		pmap->pmap_phys = pmap;
1732 	}
1733 
1734 	mtx_lock(&moea_vsid_mutex);
1735 	/*
1736 	 * Allocate some segment registers for this pmap.
1737 	 */
1738 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1739 		u_int	hash, n;
1740 
1741 		/*
1742 		 * Create a new value by multiplying by a prime and adding in
1743 		 * entropy from the timebase register.  This is to make the
1744 		 * VSID more random so that the PT hash function collides
1745 		 * less often.  (Note that the prime casues gcc to do shifts
1746 		 * instead of a multiply.)
1747 		 */
1748 		moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1749 		hash = moea_vsidcontext & (NPMAPS - 1);
1750 		if (hash == 0)		/* 0 is special, avoid it */
1751 			continue;
1752 		n = hash >> 5;
1753 		mask = 1 << (hash & (VSID_NBPW - 1));
1754 		hash = (moea_vsidcontext & 0xfffff);
1755 		if (moea_vsid_bitmap[n] & mask) {	/* collision? */
1756 			/* anything free in this bucket? */
1757 			if (moea_vsid_bitmap[n] == 0xffffffff) {
1758 				entropy = (moea_vsidcontext >> 20);
1759 				continue;
1760 			}
1761 			i = ffs(~moea_vsid_bitmap[n]) - 1;
1762 			mask = 1 << i;
1763 			hash &= rounddown2(0xfffff, VSID_NBPW);
1764 			hash |= i;
1765 		}
1766 		KASSERT(!(moea_vsid_bitmap[n] & mask),
1767 		    ("Allocating in-use VSID group %#x\n", hash));
1768 		moea_vsid_bitmap[n] |= mask;
1769 		for (i = 0; i < 16; i++)
1770 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1771 		mtx_unlock(&moea_vsid_mutex);
1772 		return (1);
1773 	}
1774 
1775 	mtx_unlock(&moea_vsid_mutex);
1776 	panic("moea_pinit: out of segments");
1777 }
1778 
1779 /*
1780  * Initialize the pmap associated with process 0.
1781  */
1782 void
1783 moea_pinit0(pmap_t pm)
1784 {
1785 
1786 	PMAP_LOCK_INIT(pm);
1787 	moea_pinit(pm);
1788 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1789 }
1790 
1791 /*
1792  * Set the physical protection on the specified range of this map as requested.
1793  */
1794 void
1795 moea_protect(pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1796     vm_prot_t prot)
1797 {
1798 	struct	pvo_entry *pvo, *tpvo, key;
1799 	struct	pte *pt;
1800 
1801 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1802 	    ("moea_protect: non current pmap"));
1803 
1804 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1805 		moea_remove(pm, sva, eva);
1806 		return;
1807 	}
1808 
1809 	rw_wlock(&pvh_global_lock);
1810 	PMAP_LOCK(pm);
1811 	key.pvo_vaddr = sva;
1812 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1813 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1814 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1815 
1816 		/*
1817 		 * Grab the PTE pointer before we diddle with the cached PTE
1818 		 * copy.
1819 		 */
1820 		pt = moea_pvo_to_pte(pvo, -1);
1821 		/*
1822 		 * Change the protection of the page.
1823 		 */
1824 		pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1825 		pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1826 
1827 		/*
1828 		 * If the PVO is in the page table, update that pte as well.
1829 		 */
1830 		if (pt != NULL) {
1831 			moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1832 			mtx_unlock(&moea_table_mutex);
1833 		}
1834 	}
1835 	rw_wunlock(&pvh_global_lock);
1836 	PMAP_UNLOCK(pm);
1837 }
1838 
1839 /*
1840  * Map a list of wired pages into kernel virtual address space.  This is
1841  * intended for temporary mappings which do not need page modification or
1842  * references recorded.  Existing mappings in the region are overwritten.
1843  */
1844 void
1845 moea_qenter(vm_offset_t sva, vm_page_t *m, int count)
1846 {
1847 	vm_offset_t va;
1848 
1849 	va = sva;
1850 	while (count-- > 0) {
1851 		moea_kenter(va, VM_PAGE_TO_PHYS(*m));
1852 		va += PAGE_SIZE;
1853 		m++;
1854 	}
1855 }
1856 
1857 /*
1858  * Remove page mappings from kernel virtual address space.  Intended for
1859  * temporary mappings entered by moea_qenter.
1860  */
1861 void
1862 moea_qremove(vm_offset_t sva, int count)
1863 {
1864 	vm_offset_t va;
1865 
1866 	va = sva;
1867 	while (count-- > 0) {
1868 		moea_kremove(va);
1869 		va += PAGE_SIZE;
1870 	}
1871 }
1872 
1873 void
1874 moea_release(pmap_t pmap)
1875 {
1876         int idx, mask;
1877 
1878 	/*
1879 	 * Free segment register's VSID
1880 	 */
1881         if (pmap->pm_sr[0] == 0)
1882                 panic("moea_release");
1883 
1884 	mtx_lock(&moea_vsid_mutex);
1885         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1886         mask = 1 << (idx % VSID_NBPW);
1887         idx /= VSID_NBPW;
1888         moea_vsid_bitmap[idx] &= ~mask;
1889 	mtx_unlock(&moea_vsid_mutex);
1890 }
1891 
1892 /*
1893  * Remove the given range of addresses from the specified map.
1894  */
1895 void
1896 moea_remove(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1897 {
1898 	struct	pvo_entry *pvo, *tpvo, key;
1899 
1900 	rw_wlock(&pvh_global_lock);
1901 	PMAP_LOCK(pm);
1902 	key.pvo_vaddr = sva;
1903 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1904 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1905 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1906 		moea_pvo_remove(pvo, -1);
1907 	}
1908 	PMAP_UNLOCK(pm);
1909 	rw_wunlock(&pvh_global_lock);
1910 }
1911 
1912 /*
1913  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1914  * will reflect changes in pte's back to the vm_page.
1915  */
1916 void
1917 moea_remove_all(vm_page_t m)
1918 {
1919 	struct  pvo_head *pvo_head;
1920 	struct	pvo_entry *pvo, *next_pvo;
1921 	pmap_t	pmap;
1922 
1923 	rw_wlock(&pvh_global_lock);
1924 	pvo_head = vm_page_to_pvoh(m);
1925 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1926 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1927 
1928 		pmap = pvo->pvo_pmap;
1929 		PMAP_LOCK(pmap);
1930 		moea_pvo_remove(pvo, -1);
1931 		PMAP_UNLOCK(pmap);
1932 	}
1933 	if ((m->a.flags & PGA_WRITEABLE) && moea_query_bit(m, PTE_CHG)) {
1934 		moea_attr_clear(m, PTE_CHG);
1935 		vm_page_dirty(m);
1936 	}
1937 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1938 	rw_wunlock(&pvh_global_lock);
1939 }
1940 
1941 static int
1942 moea_mincore(pmap_t pm, vm_offset_t va, vm_paddr_t *pap)
1943 {
1944 	struct pvo_entry *pvo;
1945 	vm_paddr_t pa;
1946 	vm_page_t m;
1947 	int val;
1948 	bool managed;
1949 
1950 	PMAP_LOCK(pm);
1951 
1952 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1953 	if (pvo != NULL) {
1954 		pa = PVO_PADDR(pvo);
1955 		m = PHYS_TO_VM_PAGE(pa);
1956 		managed = (pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED;
1957 		val = MINCORE_INCORE;
1958 	} else {
1959 		PMAP_UNLOCK(pm);
1960 		return (0);
1961 	}
1962 
1963 	PMAP_UNLOCK(pm);
1964 
1965 	if (m == NULL)
1966 		return (0);
1967 
1968 	if (managed) {
1969 		if (moea_is_modified(m))
1970 			val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
1971 
1972 		if (moea_is_referenced(m))
1973 			val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
1974 	}
1975 
1976 	if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) !=
1977 	    (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) &&
1978 	    managed) {
1979 		*pap = pa;
1980 	}
1981 
1982 	return (val);
1983 }
1984 
1985 /*
1986  * Allocate a physical page of memory directly from the phys_avail map.
1987  * Can only be called from moea_bootstrap before avail start and end are
1988  * calculated.
1989  */
1990 static vm_offset_t
1991 moea_bootstrap_alloc(vm_size_t size, u_int align)
1992 {
1993 	vm_offset_t	s, e;
1994 	int		i, j;
1995 
1996 	size = round_page(size);
1997 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1998 		if (align != 0)
1999 			s = roundup2(phys_avail[i], align);
2000 		else
2001 			s = phys_avail[i];
2002 		e = s + size;
2003 
2004 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2005 			continue;
2006 
2007 		if (s == phys_avail[i]) {
2008 			phys_avail[i] += size;
2009 		} else if (e == phys_avail[i + 1]) {
2010 			phys_avail[i + 1] -= size;
2011 		} else {
2012 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2013 				phys_avail[j] = phys_avail[j - 2];
2014 				phys_avail[j + 1] = phys_avail[j - 1];
2015 			}
2016 
2017 			phys_avail[i + 3] = phys_avail[i + 1];
2018 			phys_avail[i + 1] = s;
2019 			phys_avail[i + 2] = e;
2020 			phys_avail_count++;
2021 		}
2022 
2023 		return (s);
2024 	}
2025 	panic("moea_bootstrap_alloc: could not allocate memory");
2026 }
2027 
2028 static void
2029 moea_syncicache(vm_paddr_t pa, vm_size_t len)
2030 {
2031 	__syncicache((void *)pa, len);
2032 }
2033 
2034 static int
2035 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
2036     vm_offset_t va, vm_paddr_t pa, u_int pte_lo, int flags)
2037 {
2038 	struct	pvo_entry *pvo;
2039 	u_int	sr;
2040 	int	first;
2041 	u_int	ptegidx;
2042 	int	i;
2043 	int     bootstrap;
2044 
2045 	moea_pvo_enter_calls++;
2046 	first = 0;
2047 	bootstrap = 0;
2048 
2049 	/*
2050 	 * Compute the PTE Group index.
2051 	 */
2052 	va &= ~ADDR_POFF;
2053 	sr = va_to_sr(pm->pm_sr, va);
2054 	ptegidx = va_to_pteg(sr, va);
2055 
2056 	/*
2057 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
2058 	 * there is a mapping.
2059 	 */
2060 	mtx_lock(&moea_table_mutex);
2061 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2062 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2063 			if (PVO_PADDR(pvo) == pa &&
2064 			    (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
2065 			    (pte_lo & PTE_PP)) {
2066 				/*
2067 				 * The PTE is not changing.  Instead, this may
2068 				 * be a request to change the mapping's wired
2069 				 * attribute.
2070 				 */
2071 				mtx_unlock(&moea_table_mutex);
2072 				if ((flags & PVO_WIRED) != 0 &&
2073 				    (pvo->pvo_vaddr & PVO_WIRED) == 0) {
2074 					pvo->pvo_vaddr |= PVO_WIRED;
2075 					pm->pm_stats.wired_count++;
2076 				} else if ((flags & PVO_WIRED) == 0 &&
2077 				    (pvo->pvo_vaddr & PVO_WIRED) != 0) {
2078 					pvo->pvo_vaddr &= ~PVO_WIRED;
2079 					pm->pm_stats.wired_count--;
2080 				}
2081 				return (0);
2082 			}
2083 			moea_pvo_remove(pvo, -1);
2084 			break;
2085 		}
2086 	}
2087 
2088 	/*
2089 	 * If we aren't overwriting a mapping, try to allocate.
2090 	 */
2091 	if (moea_initialized) {
2092 		pvo = uma_zalloc(zone, M_NOWAIT);
2093 	} else {
2094 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
2095 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
2096 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
2097 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
2098 		}
2099 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
2100 		moea_bpvo_pool_index++;
2101 		bootstrap = 1;
2102 	}
2103 
2104 	if (pvo == NULL) {
2105 		mtx_unlock(&moea_table_mutex);
2106 		return (ENOMEM);
2107 	}
2108 
2109 	moea_pvo_entries++;
2110 	pvo->pvo_vaddr = va;
2111 	pvo->pvo_pmap = pm;
2112 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
2113 	pvo->pvo_vaddr &= ~ADDR_POFF;
2114 	if (flags & PVO_WIRED)
2115 		pvo->pvo_vaddr |= PVO_WIRED;
2116 	if (pvo_head != &moea_pvo_kunmanaged)
2117 		pvo->pvo_vaddr |= PVO_MANAGED;
2118 	if (bootstrap)
2119 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
2120 
2121 	moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo);
2122 
2123 	/*
2124 	 * Add to pmap list
2125 	 */
2126 	RB_INSERT(pvo_tree, &pm->pmap_pvo, pvo);
2127 
2128 	/*
2129 	 * Remember if the list was empty and therefore will be the first
2130 	 * item.
2131 	 */
2132 	if (LIST_FIRST(pvo_head) == NULL)
2133 		first = 1;
2134 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2135 
2136 	if (pvo->pvo_vaddr & PVO_WIRED)
2137 		pm->pm_stats.wired_count++;
2138 	pm->pm_stats.resident_count++;
2139 
2140 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2141 	KASSERT(i < 8, ("Invalid PTE index"));
2142 	if (i >= 0) {
2143 		PVO_PTEGIDX_SET(pvo, i);
2144 	} else {
2145 		panic("moea_pvo_enter: overflow");
2146 		moea_pte_overflow++;
2147 	}
2148 	mtx_unlock(&moea_table_mutex);
2149 
2150 	return (first ? ENOENT : 0);
2151 }
2152 
2153 static void
2154 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
2155 {
2156 	struct	pte *pt;
2157 
2158 	/*
2159 	 * If there is an active pte entry, we need to deactivate it (and
2160 	 * save the ref & cfg bits).
2161 	 */
2162 	pt = moea_pvo_to_pte(pvo, pteidx);
2163 	if (pt != NULL) {
2164 		moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
2165 		mtx_unlock(&moea_table_mutex);
2166 		PVO_PTEGIDX_CLR(pvo);
2167 	} else {
2168 		moea_pte_overflow--;
2169 	}
2170 
2171 	/*
2172 	 * Update our statistics.
2173 	 */
2174 	pvo->pvo_pmap->pm_stats.resident_count--;
2175 	if (pvo->pvo_vaddr & PVO_WIRED)
2176 		pvo->pvo_pmap->pm_stats.wired_count--;
2177 
2178 	/*
2179 	 * Remove this PVO from the PV and pmap lists.
2180 	 */
2181 	LIST_REMOVE(pvo, pvo_vlink);
2182 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2183 
2184 	/*
2185 	 * Save the REF/CHG bits into their cache if the page is managed.
2186 	 * Clear PGA_WRITEABLE if all mappings of the page have been removed.
2187 	 */
2188 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED) {
2189 		struct vm_page *pg;
2190 
2191 		pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
2192 		if (pg != NULL) {
2193 			moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2194 			    (PTE_REF | PTE_CHG));
2195 			if (LIST_EMPTY(&pg->md.mdpg_pvoh))
2196 				vm_page_aflag_clear(pg, PGA_WRITEABLE);
2197 		}
2198 	}
2199 
2200 	/*
2201 	 * Remove this from the overflow list and return it to the pool
2202 	 * if we aren't going to reuse it.
2203 	 */
2204 	LIST_REMOVE(pvo, pvo_olink);
2205 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2206 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
2207 		    moea_upvo_zone, pvo);
2208 	moea_pvo_entries--;
2209 	moea_pvo_remove_calls++;
2210 }
2211 
2212 static __inline int
2213 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2214 {
2215 	int	pteidx;
2216 
2217 	/*
2218 	 * We can find the actual pte entry without searching by grabbing
2219 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2220 	 * noticing the HID bit.
2221 	 */
2222 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2223 	if (pvo->pvo_pte.pte.pte_hi & PTE_HID)
2224 		pteidx ^= moea_pteg_mask * 8;
2225 
2226 	return (pteidx);
2227 }
2228 
2229 static struct pvo_entry *
2230 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2231 {
2232 	struct	pvo_entry *pvo;
2233 	int	ptegidx;
2234 	u_int	sr;
2235 
2236 	va &= ~ADDR_POFF;
2237 	sr = va_to_sr(pm->pm_sr, va);
2238 	ptegidx = va_to_pteg(sr, va);
2239 
2240 	mtx_lock(&moea_table_mutex);
2241 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2242 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2243 			if (pteidx_p)
2244 				*pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2245 			break;
2246 		}
2247 	}
2248 	mtx_unlock(&moea_table_mutex);
2249 
2250 	return (pvo);
2251 }
2252 
2253 static struct pte *
2254 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2255 {
2256 	struct	pte *pt;
2257 
2258 	/*
2259 	 * If we haven't been supplied the ptegidx, calculate it.
2260 	 */
2261 	if (pteidx == -1) {
2262 		int	ptegidx;
2263 		u_int	sr;
2264 
2265 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2266 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2267 		pteidx = moea_pvo_pte_index(pvo, ptegidx);
2268 	}
2269 
2270 	pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2271 	mtx_lock(&moea_table_mutex);
2272 
2273 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2274 		panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2275 		    "valid pte index", pvo);
2276 	}
2277 
2278 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2279 		panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2280 		    "pvo but no valid pte", pvo);
2281 	}
2282 
2283 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2284 		if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) {
2285 			panic("moea_pvo_to_pte: pvo %p has valid pte in "
2286 			    "moea_pteg_table %p but invalid in pvo", pvo, pt);
2287 		}
2288 
2289 		if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2290 		    != 0) {
2291 			panic("moea_pvo_to_pte: pvo %p pte does not match "
2292 			    "pte %p in moea_pteg_table", pvo, pt);
2293 		}
2294 
2295 		mtx_assert(&moea_table_mutex, MA_OWNED);
2296 		return (pt);
2297 	}
2298 
2299 	if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) {
2300 		panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2301 		    "moea_pteg_table but valid in pvo: %8x, %8x", pvo, pt, pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2302 	}
2303 
2304 	mtx_unlock(&moea_table_mutex);
2305 	return (NULL);
2306 }
2307 
2308 /*
2309  * XXX: THIS STUFF SHOULD BE IN pte.c?
2310  */
2311 int
2312 moea_pte_spill(vm_offset_t addr)
2313 {
2314 	struct	pvo_entry *source_pvo, *victim_pvo;
2315 	struct	pvo_entry *pvo;
2316 	int	ptegidx, i, j;
2317 	u_int	sr;
2318 	struct	pteg *pteg;
2319 	struct	pte *pt;
2320 
2321 	moea_pte_spills++;
2322 
2323 	sr = mfsrin(addr);
2324 	ptegidx = va_to_pteg(sr, addr);
2325 
2326 	/*
2327 	 * Have to substitute some entry.  Use the primary hash for this.
2328 	 * Use low bits of timebase as random generator.
2329 	 */
2330 	pteg = &moea_pteg_table[ptegidx];
2331 	mtx_lock(&moea_table_mutex);
2332 	__asm __volatile("mftb %0" : "=r"(i));
2333 	i &= 7;
2334 	pt = &pteg->pt[i];
2335 
2336 	source_pvo = NULL;
2337 	victim_pvo = NULL;
2338 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2339 		/*
2340 		 * We need to find a pvo entry for this address.
2341 		 */
2342 		if (source_pvo == NULL &&
2343 		    moea_pte_match(&pvo->pvo_pte.pte, sr, addr,
2344 		    pvo->pvo_pte.pte.pte_hi & PTE_HID)) {
2345 			/*
2346 			 * Now found an entry to be spilled into the pteg.
2347 			 * The PTE is now valid, so we know it's active.
2348 			 */
2349 			j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2350 
2351 			if (j >= 0) {
2352 				PVO_PTEGIDX_SET(pvo, j);
2353 				moea_pte_overflow--;
2354 				mtx_unlock(&moea_table_mutex);
2355 				return (1);
2356 			}
2357 
2358 			source_pvo = pvo;
2359 
2360 			if (victim_pvo != NULL)
2361 				break;
2362 		}
2363 
2364 		/*
2365 		 * We also need the pvo entry of the victim we are replacing
2366 		 * so save the R & C bits of the PTE.
2367 		 */
2368 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2369 		    moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2370 			victim_pvo = pvo;
2371 			if (source_pvo != NULL)
2372 				break;
2373 		}
2374 	}
2375 
2376 	if (source_pvo == NULL) {
2377 		mtx_unlock(&moea_table_mutex);
2378 		return (0);
2379 	}
2380 
2381 	if (victim_pvo == NULL) {
2382 		if ((pt->pte_hi & PTE_HID) == 0)
2383 			panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2384 			    "entry", pt);
2385 
2386 		/*
2387 		 * If this is a secondary PTE, we need to search it's primary
2388 		 * pvo bucket for the matching PVO.
2389 		 */
2390 		LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2391 		    pvo_olink) {
2392 			/*
2393 			 * We also need the pvo entry of the victim we are
2394 			 * replacing so save the R & C bits of the PTE.
2395 			 */
2396 			if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2397 				victim_pvo = pvo;
2398 				break;
2399 			}
2400 		}
2401 
2402 		if (victim_pvo == NULL)
2403 			panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2404 			    "entry", pt);
2405 	}
2406 
2407 	/*
2408 	 * We are invalidating the TLB entry for the EA we are replacing even
2409 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2410 	 * contained in the TLB entry.
2411 	 */
2412 	source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID;
2413 
2414 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2415 	moea_pte_set(pt, &source_pvo->pvo_pte.pte);
2416 
2417 	PVO_PTEGIDX_CLR(victim_pvo);
2418 	PVO_PTEGIDX_SET(source_pvo, i);
2419 	moea_pte_replacements++;
2420 
2421 	mtx_unlock(&moea_table_mutex);
2422 	return (1);
2423 }
2424 
2425 static __inline struct pvo_entry *
2426 moea_pte_spillable_ident(u_int ptegidx)
2427 {
2428 	struct	pte *pt;
2429 	struct	pvo_entry *pvo_walk, *pvo = NULL;
2430 
2431 	LIST_FOREACH(pvo_walk, &moea_pvo_table[ptegidx], pvo_olink) {
2432 		if (pvo_walk->pvo_vaddr & PVO_WIRED)
2433 			continue;
2434 
2435 		if (!(pvo_walk->pvo_pte.pte.pte_hi & PTE_VALID))
2436 			continue;
2437 
2438 		pt = moea_pvo_to_pte(pvo_walk, -1);
2439 
2440 		if (pt == NULL)
2441 			continue;
2442 
2443 		pvo = pvo_walk;
2444 
2445 		mtx_unlock(&moea_table_mutex);
2446 		if (!(pt->pte_lo & PTE_REF))
2447 			return (pvo_walk);
2448 	}
2449 
2450 	return (pvo);
2451 }
2452 
2453 static int
2454 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2455 {
2456 	struct	pte *pt;
2457 	struct	pvo_entry *victim_pvo;
2458 	int	i;
2459 	int	victim_idx;
2460 	u_int	pteg_bkpidx = ptegidx;
2461 
2462 	mtx_assert(&moea_table_mutex, MA_OWNED);
2463 
2464 	/*
2465 	 * First try primary hash.
2466 	 */
2467 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2468 		if ((pt->pte_hi & PTE_VALID) == 0) {
2469 			pvo_pt->pte_hi &= ~PTE_HID;
2470 			moea_pte_set(pt, pvo_pt);
2471 			return (i);
2472 		}
2473 	}
2474 
2475 	/*
2476 	 * Now try secondary hash.
2477 	 */
2478 	ptegidx ^= moea_pteg_mask;
2479 
2480 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2481 		if ((pt->pte_hi & PTE_VALID) == 0) {
2482 			pvo_pt->pte_hi |= PTE_HID;
2483 			moea_pte_set(pt, pvo_pt);
2484 			return (i);
2485 		}
2486 	}
2487 
2488 	/* Try again, but this time try to force a PTE out. */
2489 	ptegidx = pteg_bkpidx;
2490 
2491 	victim_pvo = moea_pte_spillable_ident(ptegidx);
2492 	if (victim_pvo == NULL) {
2493 		ptegidx ^= moea_pteg_mask;
2494 		victim_pvo = moea_pte_spillable_ident(ptegidx);
2495 	}
2496 
2497 	if (victim_pvo == NULL) {
2498 		panic("moea_pte_insert: overflow");
2499 		return (-1);
2500 	}
2501 
2502 	victim_idx = moea_pvo_pte_index(victim_pvo, ptegidx);
2503 
2504 	if (pteg_bkpidx == ptegidx)
2505 		pvo_pt->pte_hi &= ~PTE_HID;
2506 	else
2507 		pvo_pt->pte_hi |= PTE_HID;
2508 
2509 	/*
2510 	 * Synchronize the sacrifice PTE with its PVO, then mark both
2511 	 * invalid. The PVO will be reused when/if the VM system comes
2512 	 * here after a fault.
2513 	 */
2514 	pt = &moea_pteg_table[victim_idx >> 3].pt[victim_idx & 7];
2515 
2516 	if (pt->pte_hi != victim_pvo->pvo_pte.pte.pte_hi)
2517 	    panic("Victim PVO doesn't match PTE! PVO: %8x, PTE: %8x", victim_pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2518 
2519 	/*
2520 	 * Set the new PTE.
2521 	 */
2522 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2523 	PVO_PTEGIDX_CLR(victim_pvo);
2524 	moea_pte_overflow++;
2525 	moea_pte_set(pt, pvo_pt);
2526 
2527 	return (victim_idx & 7);
2528 }
2529 
2530 static boolean_t
2531 moea_query_bit(vm_page_t m, int ptebit)
2532 {
2533 	struct	pvo_entry *pvo;
2534 	struct	pte *pt;
2535 
2536 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2537 	if (moea_attr_fetch(m) & ptebit)
2538 		return (TRUE);
2539 
2540 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2541 		/*
2542 		 * See if we saved the bit off.  If so, cache it and return
2543 		 * success.
2544 		 */
2545 		if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2546 			moea_attr_save(m, ptebit);
2547 			return (TRUE);
2548 		}
2549 	}
2550 
2551 	/*
2552 	 * No luck, now go through the hard part of looking at the PTEs
2553 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2554 	 * the PTEs.
2555 	 */
2556 	powerpc_sync();
2557 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2558 		/*
2559 		 * See if this pvo has a valid PTE.  if so, fetch the
2560 		 * REF/CHG bits from the valid PTE.  If the appropriate
2561 		 * ptebit is set, cache it and return success.
2562 		 */
2563 		pt = moea_pvo_to_pte(pvo, -1);
2564 		if (pt != NULL) {
2565 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2566 			mtx_unlock(&moea_table_mutex);
2567 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2568 				moea_attr_save(m, ptebit);
2569 				return (TRUE);
2570 			}
2571 		}
2572 	}
2573 
2574 	return (FALSE);
2575 }
2576 
2577 static u_int
2578 moea_clear_bit(vm_page_t m, int ptebit)
2579 {
2580 	u_int	count;
2581 	struct	pvo_entry *pvo;
2582 	struct	pte *pt;
2583 
2584 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2585 
2586 	/*
2587 	 * Clear the cached value.
2588 	 */
2589 	moea_attr_clear(m, ptebit);
2590 
2591 	/*
2592 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2593 	 * we can reset the right ones).  note that since the pvo entries and
2594 	 * list heads are accessed via BAT0 and are never placed in the page
2595 	 * table, we don't have to worry about further accesses setting the
2596 	 * REF/CHG bits.
2597 	 */
2598 	powerpc_sync();
2599 
2600 	/*
2601 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2602 	 * valid pte clear the ptebit from the valid pte.
2603 	 */
2604 	count = 0;
2605 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2606 		pt = moea_pvo_to_pte(pvo, -1);
2607 		if (pt != NULL) {
2608 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2609 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2610 				count++;
2611 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2612 			}
2613 			mtx_unlock(&moea_table_mutex);
2614 		}
2615 		pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2616 	}
2617 
2618 	return (count);
2619 }
2620 
2621 /*
2622  * Return true if the physical range is encompassed by the battable[idx]
2623  */
2624 static int
2625 moea_bat_mapped(int idx, vm_paddr_t pa, vm_size_t size)
2626 {
2627 	u_int prot;
2628 	u_int32_t start;
2629 	u_int32_t end;
2630 	u_int32_t bat_ble;
2631 
2632 	/*
2633 	 * Return immediately if not a valid mapping
2634 	 */
2635 	if (!(battable[idx].batu & BAT_Vs))
2636 		return (EINVAL);
2637 
2638 	/*
2639 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2640 	 * so it can function as an i/o page
2641 	 */
2642 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2643 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2644 		return (EPERM);
2645 
2646 	/*
2647 	 * The address should be within the BAT range. Assume that the
2648 	 * start address in the BAT has the correct alignment (thus
2649 	 * not requiring masking)
2650 	 */
2651 	start = battable[idx].batl & BAT_PBS;
2652 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2653 	end = start | (bat_ble << 15) | 0x7fff;
2654 
2655 	if ((pa < start) || ((pa + size) > end))
2656 		return (ERANGE);
2657 
2658 	return (0);
2659 }
2660 
2661 boolean_t
2662 moea_dev_direct_mapped(vm_paddr_t pa, vm_size_t size)
2663 {
2664 	int i;
2665 
2666 	/*
2667 	 * This currently does not work for entries that
2668 	 * overlap 256M BAT segments.
2669 	 */
2670 
2671 	for(i = 0; i < 16; i++)
2672 		if (moea_bat_mapped(i, pa, size) == 0)
2673 			return (0);
2674 
2675 	return (EFAULT);
2676 }
2677 
2678 /*
2679  * Map a set of physical memory pages into the kernel virtual
2680  * address space. Return a pointer to where it is mapped. This
2681  * routine is intended to be used for mapping device memory,
2682  * NOT real memory.
2683  */
2684 void *
2685 moea_mapdev(vm_paddr_t pa, vm_size_t size)
2686 {
2687 
2688 	return (moea_mapdev_attr(pa, size, VM_MEMATTR_DEFAULT));
2689 }
2690 
2691 void *
2692 moea_mapdev_attr(vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
2693 {
2694 	vm_offset_t va, tmpva, ppa, offset;
2695 	int i;
2696 
2697 	ppa = trunc_page(pa);
2698 	offset = pa & PAGE_MASK;
2699 	size = roundup(offset + size, PAGE_SIZE);
2700 
2701 	/*
2702 	 * If the physical address lies within a valid BAT table entry,
2703 	 * return the 1:1 mapping. This currently doesn't work
2704 	 * for regions that overlap 256M BAT segments.
2705 	 */
2706 	for (i = 0; i < 16; i++) {
2707 		if (moea_bat_mapped(i, pa, size) == 0)
2708 			return ((void *) pa);
2709 	}
2710 
2711 	va = kva_alloc(size);
2712 	if (!va)
2713 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2714 
2715 	for (tmpva = va; size > 0;) {
2716 		moea_kenter_attr(tmpva, ppa, ma);
2717 		tlbie(tmpva);
2718 		size -= PAGE_SIZE;
2719 		tmpva += PAGE_SIZE;
2720 		ppa += PAGE_SIZE;
2721 	}
2722 
2723 	return ((void *)(va + offset));
2724 }
2725 
2726 void
2727 moea_unmapdev(void *p, vm_size_t size)
2728 {
2729 	vm_offset_t base, offset, va;
2730 
2731 	/*
2732 	 * If this is outside kernel virtual space, then it's a
2733 	 * battable entry and doesn't require unmapping
2734 	 */
2735 	va = (vm_offset_t)p;
2736 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2737 		base = trunc_page(va);
2738 		offset = va & PAGE_MASK;
2739 		size = roundup(offset + size, PAGE_SIZE);
2740 		moea_qremove(base, atop(size));
2741 		kva_free(base, size);
2742 	}
2743 }
2744 
2745 static void
2746 moea_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz)
2747 {
2748 	struct pvo_entry *pvo;
2749 	vm_offset_t lim;
2750 	vm_paddr_t pa;
2751 	vm_size_t len;
2752 
2753 	PMAP_LOCK(pm);
2754 	while (sz > 0) {
2755 		lim = round_page(va + 1);
2756 		len = MIN(lim - va, sz);
2757 		pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2758 		if (pvo != NULL) {
2759 			pa = PVO_PADDR(pvo) | (va & ADDR_POFF);
2760 			moea_syncicache(pa, len);
2761 		}
2762 		va += len;
2763 		sz -= len;
2764 	}
2765 	PMAP_UNLOCK(pm);
2766 }
2767 
2768 void
2769 moea_dumpsys_map(vm_paddr_t pa, size_t sz, void **va)
2770 {
2771 
2772 	*va = (void *)pa;
2773 }
2774 
2775 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
2776 
2777 void
2778 moea_scan_init()
2779 {
2780 	struct pvo_entry *pvo;
2781 	vm_offset_t va;
2782 	int i;
2783 
2784 	if (!do_minidump) {
2785 		/* Initialize phys. segments for dumpsys(). */
2786 		memset(&dump_map, 0, sizeof(dump_map));
2787 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
2788 		for (i = 0; i < pregions_sz; i++) {
2789 			dump_map[i].pa_start = pregions[i].mr_start;
2790 			dump_map[i].pa_size = pregions[i].mr_size;
2791 		}
2792 		return;
2793 	}
2794 
2795 	/* Virtual segments for minidumps: */
2796 	memset(&dump_map, 0, sizeof(dump_map));
2797 
2798 	/* 1st: kernel .data and .bss. */
2799 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
2800 	dump_map[0].pa_size =
2801 	    round_page((uintptr_t)_end) - dump_map[0].pa_start;
2802 
2803 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2804 	dump_map[1].pa_start = (vm_paddr_t)msgbufp->msg_ptr;
2805 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
2806 
2807 	/* 3rd: kernel VM. */
2808 	va = dump_map[1].pa_start + dump_map[1].pa_size;
2809 	/* Find start of next chunk (from va). */
2810 	while (va < virtual_end) {
2811 		/* Don't dump the buffer cache. */
2812 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
2813 			va = kmi.buffer_eva;
2814 			continue;
2815 		}
2816 		pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
2817 		if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2818 			break;
2819 		va += PAGE_SIZE;
2820 	}
2821 	if (va < virtual_end) {
2822 		dump_map[2].pa_start = va;
2823 		va += PAGE_SIZE;
2824 		/* Find last page in chunk. */
2825 		while (va < virtual_end) {
2826 			/* Don't run into the buffer cache. */
2827 			if (va == kmi.buffer_sva)
2828 				break;
2829 			pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF,
2830 			    NULL);
2831 			if (pvo == NULL ||
2832 			    !(pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2833 				break;
2834 			va += PAGE_SIZE;
2835 		}
2836 		dump_map[2].pa_size = va - dump_map[2].pa_start;
2837 	}
2838 }
2839