xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision 15c433351f54e7cd5bec8d36c8e89e6a7fa55b26)
1 /*-
2  * Copyright (c) 2001 The NetBSD Foundation, Inc.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to The NetBSD Foundation
6  * by Matt Thomas <matt@3am-software.com> of Allegro Networks, Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
18  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
19  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
21  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
22  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
23  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
24  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
25  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
26  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27  * POSSIBILITY OF SUCH DAMAGE.
28  */
29 /*-
30  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
31  * Copyright (C) 1995, 1996 TooLs GmbH.
32  * All rights reserved.
33  *
34  * Redistribution and use in source and binary forms, with or without
35  * modification, are permitted provided that the following conditions
36  * are met:
37  * 1. Redistributions of source code must retain the above copyright
38  *    notice, this list of conditions and the following disclaimer.
39  * 2. Redistributions in binary form must reproduce the above copyright
40  *    notice, this list of conditions and the following disclaimer in the
41  *    documentation and/or other materials provided with the distribution.
42  * 3. All advertising materials mentioning features or use of this software
43  *    must display the following acknowledgement:
44  *	This product includes software developed by TooLs GmbH.
45  * 4. The name of TooLs GmbH may not be used to endorse or promote products
46  *    derived from this software without specific prior written permission.
47  *
48  * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
49  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
50  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
51  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
52  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
53  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
54  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
55  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
56  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
57  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
58  *
59  * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $
60  */
61 /*-
62  * Copyright (C) 2001 Benno Rice.
63  * All rights reserved.
64  *
65  * Redistribution and use in source and binary forms, with or without
66  * modification, are permitted provided that the following conditions
67  * are met:
68  * 1. Redistributions of source code must retain the above copyright
69  *    notice, this list of conditions and the following disclaimer.
70  * 2. Redistributions in binary form must reproduce the above copyright
71  *    notice, this list of conditions and the following disclaimer in the
72  *    documentation and/or other materials provided with the distribution.
73  *
74  * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR
75  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
76  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
77  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
78  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
79  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
80  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
81  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
82  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
83  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
84  */
85 
86 #include <sys/cdefs.h>
87 __FBSDID("$FreeBSD$");
88 
89 /*
90  * Manages physical address maps.
91  *
92  * Since the information managed by this module is also stored by the
93  * logical address mapping module, this module may throw away valid virtual
94  * to physical mappings at almost any time.  However, invalidations of
95  * mappings must be done as requested.
96  *
97  * In order to cope with hardware architectures which make virtual to
98  * physical map invalidates expensive, this module may delay invalidate
99  * reduced protection operations until such time as they are actually
100  * necessary.  This module is given full information as to which processors
101  * are currently using which maps, and to when physical maps must be made
102  * correct.
103  */
104 
105 #include "opt_kstack_pages.h"
106 
107 #include <sys/param.h>
108 #include <sys/kernel.h>
109 #include <sys/conf.h>
110 #include <sys/queue.h>
111 #include <sys/cpuset.h>
112 #include <sys/kerneldump.h>
113 #include <sys/ktr.h>
114 #include <sys/lock.h>
115 #include <sys/msgbuf.h>
116 #include <sys/mutex.h>
117 #include <sys/proc.h>
118 #include <sys/rwlock.h>
119 #include <sys/sched.h>
120 #include <sys/sysctl.h>
121 #include <sys/systm.h>
122 #include <sys/vmmeter.h>
123 
124 #include <dev/ofw/openfirm.h>
125 
126 #include <vm/vm.h>
127 #include <vm/vm_param.h>
128 #include <vm/vm_kern.h>
129 #include <vm/vm_page.h>
130 #include <vm/vm_map.h>
131 #include <vm/vm_object.h>
132 #include <vm/vm_extern.h>
133 #include <vm/vm_pageout.h>
134 #include <vm/uma.h>
135 
136 #include <machine/cpu.h>
137 #include <machine/platform.h>
138 #include <machine/bat.h>
139 #include <machine/frame.h>
140 #include <machine/md_var.h>
141 #include <machine/psl.h>
142 #include <machine/pte.h>
143 #include <machine/smp.h>
144 #include <machine/sr.h>
145 #include <machine/mmuvar.h>
146 #include <machine/trap.h>
147 
148 #include "mmu_if.h"
149 
150 #define	MOEA_DEBUG
151 
152 #define TODO	panic("%s: not implemented", __func__);
153 
154 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
155 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
156 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
157 
158 struct ofw_map {
159 	vm_offset_t	om_va;
160 	vm_size_t	om_len;
161 	vm_offset_t	om_pa;
162 	u_int		om_mode;
163 };
164 
165 extern unsigned char _etext[];
166 extern unsigned char _end[];
167 
168 /*
169  * Map of physical memory regions.
170  */
171 static struct	mem_region *regions;
172 static struct	mem_region *pregions;
173 static u_int    phys_avail_count;
174 static int	regions_sz, pregions_sz;
175 static struct	ofw_map *translations;
176 
177 /*
178  * Lock for the pteg and pvo tables.
179  */
180 struct mtx	moea_table_mutex;
181 struct mtx	moea_vsid_mutex;
182 
183 /* tlbie instruction synchronization */
184 static struct mtx tlbie_mtx;
185 
186 /*
187  * PTEG data.
188  */
189 static struct	pteg *moea_pteg_table;
190 u_int		moea_pteg_count;
191 u_int		moea_pteg_mask;
192 
193 /*
194  * PVO data.
195  */
196 struct	pvo_head *moea_pvo_table;		/* pvo entries by pteg index */
197 struct	pvo_head moea_pvo_kunmanaged =
198     LIST_HEAD_INITIALIZER(moea_pvo_kunmanaged);	/* list of unmanaged pages */
199 
200 static struct rwlock_padalign pvh_global_lock;
201 
202 uma_zone_t	moea_upvo_zone;	/* zone for pvo entries for unmanaged pages */
203 uma_zone_t	moea_mpvo_zone;	/* zone for pvo entries for managed pages */
204 
205 #define	BPVO_POOL_SIZE	32768
206 static struct	pvo_entry *moea_bpvo_pool;
207 static int	moea_bpvo_pool_index = 0;
208 
209 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
210 static u_int	moea_vsid_bitmap[NPMAPS / VSID_NBPW];
211 
212 static boolean_t moea_initialized = FALSE;
213 
214 /*
215  * Statistics.
216  */
217 u_int	moea_pte_valid = 0;
218 u_int	moea_pte_overflow = 0;
219 u_int	moea_pte_replacements = 0;
220 u_int	moea_pvo_entries = 0;
221 u_int	moea_pvo_enter_calls = 0;
222 u_int	moea_pvo_remove_calls = 0;
223 u_int	moea_pte_spills = 0;
224 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_valid, CTLFLAG_RD, &moea_pte_valid,
225     0, "");
226 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_overflow, CTLFLAG_RD,
227     &moea_pte_overflow, 0, "");
228 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_replacements, CTLFLAG_RD,
229     &moea_pte_replacements, 0, "");
230 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_entries, CTLFLAG_RD, &moea_pvo_entries,
231     0, "");
232 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_enter_calls, CTLFLAG_RD,
233     &moea_pvo_enter_calls, 0, "");
234 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_remove_calls, CTLFLAG_RD,
235     &moea_pvo_remove_calls, 0, "");
236 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_spills, CTLFLAG_RD,
237     &moea_pte_spills, 0, "");
238 
239 /*
240  * Allocate physical memory for use in moea_bootstrap.
241  */
242 static vm_offset_t	moea_bootstrap_alloc(vm_size_t, u_int);
243 
244 /*
245  * PTE calls.
246  */
247 static int		moea_pte_insert(u_int, struct pte *);
248 
249 /*
250  * PVO calls.
251  */
252 static int	moea_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *,
253 		    vm_offset_t, vm_paddr_t, u_int, int);
254 static void	moea_pvo_remove(struct pvo_entry *, int);
255 static struct	pvo_entry *moea_pvo_find_va(pmap_t, vm_offset_t, int *);
256 static struct	pte *moea_pvo_to_pte(const struct pvo_entry *, int);
257 
258 /*
259  * Utility routines.
260  */
261 static int		moea_enter_locked(pmap_t, vm_offset_t, vm_page_t,
262 			    vm_prot_t, u_int, int8_t);
263 static void		moea_syncicache(vm_paddr_t, vm_size_t);
264 static boolean_t	moea_query_bit(vm_page_t, int);
265 static u_int		moea_clear_bit(vm_page_t, int);
266 static void		moea_kremove(mmu_t, vm_offset_t);
267 int		moea_pte_spill(vm_offset_t);
268 
269 /*
270  * Kernel MMU interface
271  */
272 void moea_clear_modify(mmu_t, vm_page_t);
273 void moea_copy_page(mmu_t, vm_page_t, vm_page_t);
274 void moea_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
275     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
276 int moea_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int,
277     int8_t);
278 void moea_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
279     vm_prot_t);
280 void moea_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
281 vm_paddr_t moea_extract(mmu_t, pmap_t, vm_offset_t);
282 vm_page_t moea_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
283 void moea_init(mmu_t);
284 boolean_t moea_is_modified(mmu_t, vm_page_t);
285 boolean_t moea_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
286 boolean_t moea_is_referenced(mmu_t, vm_page_t);
287 int moea_ts_referenced(mmu_t, vm_page_t);
288 vm_offset_t moea_map(mmu_t, vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
289 boolean_t moea_page_exists_quick(mmu_t, pmap_t, vm_page_t);
290 int moea_page_wired_mappings(mmu_t, vm_page_t);
291 void moea_pinit(mmu_t, pmap_t);
292 void moea_pinit0(mmu_t, pmap_t);
293 void moea_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
294 void moea_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
295 void moea_qremove(mmu_t, vm_offset_t, int);
296 void moea_release(mmu_t, pmap_t);
297 void moea_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
298 void moea_remove_all(mmu_t, vm_page_t);
299 void moea_remove_write(mmu_t, vm_page_t);
300 void moea_unwire(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
301 void moea_zero_page(mmu_t, vm_page_t);
302 void moea_zero_page_area(mmu_t, vm_page_t, int, int);
303 void moea_zero_page_idle(mmu_t, vm_page_t);
304 void moea_activate(mmu_t, struct thread *);
305 void moea_deactivate(mmu_t, struct thread *);
306 void moea_cpu_bootstrap(mmu_t, int);
307 void moea_bootstrap(mmu_t, vm_offset_t, vm_offset_t);
308 void *moea_mapdev(mmu_t, vm_paddr_t, vm_size_t);
309 void *moea_mapdev_attr(mmu_t, vm_paddr_t, vm_size_t, vm_memattr_t);
310 void moea_unmapdev(mmu_t, vm_offset_t, vm_size_t);
311 vm_paddr_t moea_kextract(mmu_t, vm_offset_t);
312 void moea_kenter_attr(mmu_t, vm_offset_t, vm_paddr_t, vm_memattr_t);
313 void moea_kenter(mmu_t, vm_offset_t, vm_paddr_t);
314 void moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma);
315 boolean_t moea_dev_direct_mapped(mmu_t, vm_paddr_t, vm_size_t);
316 static void moea_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
317 void moea_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va);
318 void moea_scan_init(mmu_t mmu);
319 vm_offset_t moea_quick_enter_page(mmu_t mmu, vm_page_t m);
320 void moea_quick_remove_page(mmu_t mmu, vm_offset_t addr);
321 
322 static mmu_method_t moea_methods[] = {
323 	MMUMETHOD(mmu_clear_modify,	moea_clear_modify),
324 	MMUMETHOD(mmu_copy_page,	moea_copy_page),
325 	MMUMETHOD(mmu_copy_pages,	moea_copy_pages),
326 	MMUMETHOD(mmu_enter,		moea_enter),
327 	MMUMETHOD(mmu_enter_object,	moea_enter_object),
328 	MMUMETHOD(mmu_enter_quick,	moea_enter_quick),
329 	MMUMETHOD(mmu_extract,		moea_extract),
330 	MMUMETHOD(mmu_extract_and_hold,	moea_extract_and_hold),
331 	MMUMETHOD(mmu_init,		moea_init),
332 	MMUMETHOD(mmu_is_modified,	moea_is_modified),
333 	MMUMETHOD(mmu_is_prefaultable,	moea_is_prefaultable),
334 	MMUMETHOD(mmu_is_referenced,	moea_is_referenced),
335 	MMUMETHOD(mmu_ts_referenced,	moea_ts_referenced),
336 	MMUMETHOD(mmu_map,     		moea_map),
337 	MMUMETHOD(mmu_page_exists_quick,moea_page_exists_quick),
338 	MMUMETHOD(mmu_page_wired_mappings,moea_page_wired_mappings),
339 	MMUMETHOD(mmu_pinit,		moea_pinit),
340 	MMUMETHOD(mmu_pinit0,		moea_pinit0),
341 	MMUMETHOD(mmu_protect,		moea_protect),
342 	MMUMETHOD(mmu_qenter,		moea_qenter),
343 	MMUMETHOD(mmu_qremove,		moea_qremove),
344 	MMUMETHOD(mmu_release,		moea_release),
345 	MMUMETHOD(mmu_remove,		moea_remove),
346 	MMUMETHOD(mmu_remove_all,      	moea_remove_all),
347 	MMUMETHOD(mmu_remove_write,	moea_remove_write),
348 	MMUMETHOD(mmu_sync_icache,	moea_sync_icache),
349 	MMUMETHOD(mmu_unwire,		moea_unwire),
350 	MMUMETHOD(mmu_zero_page,       	moea_zero_page),
351 	MMUMETHOD(mmu_zero_page_area,	moea_zero_page_area),
352 	MMUMETHOD(mmu_zero_page_idle,	moea_zero_page_idle),
353 	MMUMETHOD(mmu_activate,		moea_activate),
354 	MMUMETHOD(mmu_deactivate,      	moea_deactivate),
355 	MMUMETHOD(mmu_page_set_memattr,	moea_page_set_memattr),
356 	MMUMETHOD(mmu_quick_enter_page, moea_quick_enter_page),
357 	MMUMETHOD(mmu_quick_remove_page, moea_quick_remove_page),
358 
359 	/* Internal interfaces */
360 	MMUMETHOD(mmu_bootstrap,       	moea_bootstrap),
361 	MMUMETHOD(mmu_cpu_bootstrap,   	moea_cpu_bootstrap),
362 	MMUMETHOD(mmu_mapdev_attr,	moea_mapdev_attr),
363 	MMUMETHOD(mmu_mapdev,		moea_mapdev),
364 	MMUMETHOD(mmu_unmapdev,		moea_unmapdev),
365 	MMUMETHOD(mmu_kextract,		moea_kextract),
366 	MMUMETHOD(mmu_kenter,		moea_kenter),
367 	MMUMETHOD(mmu_kenter_attr,	moea_kenter_attr),
368 	MMUMETHOD(mmu_dev_direct_mapped,moea_dev_direct_mapped),
369 	MMUMETHOD(mmu_scan_init,	moea_scan_init),
370 	MMUMETHOD(mmu_dumpsys_map,	moea_dumpsys_map),
371 
372 	{ 0, 0 }
373 };
374 
375 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods, 0);
376 
377 static __inline uint32_t
378 moea_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
379 {
380 	uint32_t pte_lo;
381 	int i;
382 
383 	if (ma != VM_MEMATTR_DEFAULT) {
384 		switch (ma) {
385 		case VM_MEMATTR_UNCACHEABLE:
386 			return (PTE_I | PTE_G);
387 		case VM_MEMATTR_CACHEABLE:
388 			return (PTE_M);
389 		case VM_MEMATTR_WRITE_COMBINING:
390 		case VM_MEMATTR_WRITE_BACK:
391 		case VM_MEMATTR_PREFETCHABLE:
392 			return (PTE_I);
393 		case VM_MEMATTR_WRITE_THROUGH:
394 			return (PTE_W | PTE_M);
395 		}
396 	}
397 
398 	/*
399 	 * Assume the page is cache inhibited and access is guarded unless
400 	 * it's in our available memory array.
401 	 */
402 	pte_lo = PTE_I | PTE_G;
403 	for (i = 0; i < pregions_sz; i++) {
404 		if ((pa >= pregions[i].mr_start) &&
405 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
406 			pte_lo = PTE_M;
407 			break;
408 		}
409 	}
410 
411 	return pte_lo;
412 }
413 
414 static void
415 tlbie(vm_offset_t va)
416 {
417 
418 	mtx_lock_spin(&tlbie_mtx);
419 	__asm __volatile("ptesync");
420 	__asm __volatile("tlbie %0" :: "r"(va));
421 	__asm __volatile("eieio; tlbsync; ptesync");
422 	mtx_unlock_spin(&tlbie_mtx);
423 }
424 
425 static void
426 tlbia(void)
427 {
428 	vm_offset_t va;
429 
430 	for (va = 0; va < 0x00040000; va += 0x00001000) {
431 		__asm __volatile("tlbie %0" :: "r"(va));
432 		powerpc_sync();
433 	}
434 	__asm __volatile("tlbsync");
435 	powerpc_sync();
436 }
437 
438 static __inline int
439 va_to_sr(u_int *sr, vm_offset_t va)
440 {
441 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
442 }
443 
444 static __inline u_int
445 va_to_pteg(u_int sr, vm_offset_t addr)
446 {
447 	u_int hash;
448 
449 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
450 	    ADDR_PIDX_SHFT);
451 	return (hash & moea_pteg_mask);
452 }
453 
454 static __inline struct pvo_head *
455 vm_page_to_pvoh(vm_page_t m)
456 {
457 
458 	return (&m->md.mdpg_pvoh);
459 }
460 
461 static __inline void
462 moea_attr_clear(vm_page_t m, int ptebit)
463 {
464 
465 	rw_assert(&pvh_global_lock, RA_WLOCKED);
466 	m->md.mdpg_attrs &= ~ptebit;
467 }
468 
469 static __inline int
470 moea_attr_fetch(vm_page_t m)
471 {
472 
473 	return (m->md.mdpg_attrs);
474 }
475 
476 static __inline void
477 moea_attr_save(vm_page_t m, int ptebit)
478 {
479 
480 	rw_assert(&pvh_global_lock, RA_WLOCKED);
481 	m->md.mdpg_attrs |= ptebit;
482 }
483 
484 static __inline int
485 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
486 {
487 	if (pt->pte_hi == pvo_pt->pte_hi)
488 		return (1);
489 
490 	return (0);
491 }
492 
493 static __inline int
494 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
495 {
496 	return (pt->pte_hi & ~PTE_VALID) ==
497 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
498 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
499 }
500 
501 static __inline void
502 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
503 {
504 
505 	mtx_assert(&moea_table_mutex, MA_OWNED);
506 
507 	/*
508 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
509 	 * set when the real pte is set in memory.
510 	 *
511 	 * Note: Don't set the valid bit for correct operation of tlb update.
512 	 */
513 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
514 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
515 	pt->pte_lo = pte_lo;
516 }
517 
518 static __inline void
519 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
520 {
521 
522 	mtx_assert(&moea_table_mutex, MA_OWNED);
523 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
524 }
525 
526 static __inline void
527 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
528 {
529 
530 	mtx_assert(&moea_table_mutex, MA_OWNED);
531 
532 	/*
533 	 * As shown in Section 7.6.3.2.3
534 	 */
535 	pt->pte_lo &= ~ptebit;
536 	tlbie(va);
537 }
538 
539 static __inline void
540 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
541 {
542 
543 	mtx_assert(&moea_table_mutex, MA_OWNED);
544 	pvo_pt->pte_hi |= PTE_VALID;
545 
546 	/*
547 	 * Update the PTE as defined in section 7.6.3.1.
548 	 * Note that the REF/CHG bits are from pvo_pt and thus should have
549 	 * been saved so this routine can restore them (if desired).
550 	 */
551 	pt->pte_lo = pvo_pt->pte_lo;
552 	powerpc_sync();
553 	pt->pte_hi = pvo_pt->pte_hi;
554 	powerpc_sync();
555 	moea_pte_valid++;
556 }
557 
558 static __inline void
559 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
560 {
561 
562 	mtx_assert(&moea_table_mutex, MA_OWNED);
563 	pvo_pt->pte_hi &= ~PTE_VALID;
564 
565 	/*
566 	 * Force the reg & chg bits back into the PTEs.
567 	 */
568 	powerpc_sync();
569 
570 	/*
571 	 * Invalidate the pte.
572 	 */
573 	pt->pte_hi &= ~PTE_VALID;
574 
575 	tlbie(va);
576 
577 	/*
578 	 * Save the reg & chg bits.
579 	 */
580 	moea_pte_synch(pt, pvo_pt);
581 	moea_pte_valid--;
582 }
583 
584 static __inline void
585 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
586 {
587 
588 	/*
589 	 * Invalidate the PTE
590 	 */
591 	moea_pte_unset(pt, pvo_pt, va);
592 	moea_pte_set(pt, pvo_pt);
593 }
594 
595 /*
596  * Quick sort callout for comparing memory regions.
597  */
598 static int	om_cmp(const void *a, const void *b);
599 
600 static int
601 om_cmp(const void *a, const void *b)
602 {
603 	const struct	ofw_map *mapa;
604 	const struct	ofw_map *mapb;
605 
606 	mapa = a;
607 	mapb = b;
608 	if (mapa->om_pa < mapb->om_pa)
609 		return (-1);
610 	else if (mapa->om_pa > mapb->om_pa)
611 		return (1);
612 	else
613 		return (0);
614 }
615 
616 void
617 moea_cpu_bootstrap(mmu_t mmup, int ap)
618 {
619 	u_int sdr;
620 	int i;
621 
622 	if (ap) {
623 		powerpc_sync();
624 		__asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu));
625 		__asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl));
626 		isync();
627 		__asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu));
628 		__asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl));
629 		isync();
630 	}
631 
632 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
633 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
634 	isync();
635 
636 	__asm __volatile("mtibatu 1,%0" :: "r"(0));
637 	__asm __volatile("mtdbatu 2,%0" :: "r"(0));
638 	__asm __volatile("mtibatu 2,%0" :: "r"(0));
639 	__asm __volatile("mtdbatu 3,%0" :: "r"(0));
640 	__asm __volatile("mtibatu 3,%0" :: "r"(0));
641 	isync();
642 
643 	for (i = 0; i < 16; i++)
644 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
645 	powerpc_sync();
646 
647 	sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10);
648 	__asm __volatile("mtsdr1 %0" :: "r"(sdr));
649 	isync();
650 
651 	tlbia();
652 }
653 
654 void
655 moea_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
656 {
657 	ihandle_t	mmui;
658 	phandle_t	chosen, mmu;
659 	int		sz;
660 	int		i, j;
661 	vm_size_t	size, physsz, hwphyssz;
662 	vm_offset_t	pa, va, off;
663 	void		*dpcpu;
664 	register_t	msr;
665 
666         /*
667          * Set up BAT0 to map the lowest 256 MB area
668          */
669         battable[0x0].batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
670         battable[0x0].batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
671 
672 	/*
673 	 * Map PCI memory space.
674 	 */
675 	battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
676 	battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
677 
678 	battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
679 	battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
680 
681 	battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
682 	battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
683 
684 	battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
685 	battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
686 
687 	/*
688 	 * Map obio devices.
689 	 */
690 	battable[0xf].batl = BATL(0xf0000000, BAT_I|BAT_G, BAT_PP_RW);
691 	battable[0xf].batu = BATU(0xf0000000, BAT_BL_256M, BAT_Vs);
692 
693 	/*
694 	 * Use an IBAT and a DBAT to map the bottom segment of memory
695 	 * where we are. Turn off instruction relocation temporarily
696 	 * to prevent faults while reprogramming the IBAT.
697 	 */
698 	msr = mfmsr();
699 	mtmsr(msr & ~PSL_IR);
700 	__asm (".balign 32; \n"
701 	       "mtibatu 0,%0; mtibatl 0,%1; isync; \n"
702 	       "mtdbatu 0,%0; mtdbatl 0,%1; isync"
703 	    :: "r"(battable[0].batu), "r"(battable[0].batl));
704 	mtmsr(msr);
705 
706 	/* map pci space */
707 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
708 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
709 	isync();
710 
711 	/* set global direct map flag */
712 	hw_direct_map = 1;
713 
714 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
715 	CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
716 
717 	for (i = 0; i < pregions_sz; i++) {
718 		vm_offset_t pa;
719 		vm_offset_t end;
720 
721 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
722 			pregions[i].mr_start,
723 			pregions[i].mr_start + pregions[i].mr_size,
724 			pregions[i].mr_size);
725 		/*
726 		 * Install entries into the BAT table to allow all
727 		 * of physmem to be convered by on-demand BAT entries.
728 		 * The loop will sometimes set the same battable element
729 		 * twice, but that's fine since they won't be used for
730 		 * a while yet.
731 		 */
732 		pa = pregions[i].mr_start & 0xf0000000;
733 		end = pregions[i].mr_start + pregions[i].mr_size;
734 		do {
735                         u_int n = pa >> ADDR_SR_SHFT;
736 
737 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
738 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
739 			pa += SEGMENT_LENGTH;
740 		} while (pa < end);
741 	}
742 
743 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
744 		panic("moea_bootstrap: phys_avail too small");
745 
746 	phys_avail_count = 0;
747 	physsz = 0;
748 	hwphyssz = 0;
749 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
750 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
751 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
752 		    regions[i].mr_start + regions[i].mr_size,
753 		    regions[i].mr_size);
754 		if (hwphyssz != 0 &&
755 		    (physsz + regions[i].mr_size) >= hwphyssz) {
756 			if (physsz < hwphyssz) {
757 				phys_avail[j] = regions[i].mr_start;
758 				phys_avail[j + 1] = regions[i].mr_start +
759 				    hwphyssz - physsz;
760 				physsz = hwphyssz;
761 				phys_avail_count++;
762 			}
763 			break;
764 		}
765 		phys_avail[j] = regions[i].mr_start;
766 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
767 		phys_avail_count++;
768 		physsz += regions[i].mr_size;
769 	}
770 
771 	/* Check for overlap with the kernel and exception vectors */
772 	for (j = 0; j < 2*phys_avail_count; j+=2) {
773 		if (phys_avail[j] < EXC_LAST)
774 			phys_avail[j] += EXC_LAST;
775 
776 		if (kernelstart >= phys_avail[j] &&
777 		    kernelstart < phys_avail[j+1]) {
778 			if (kernelend < phys_avail[j+1]) {
779 				phys_avail[2*phys_avail_count] =
780 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
781 				phys_avail[2*phys_avail_count + 1] =
782 				    phys_avail[j+1];
783 				phys_avail_count++;
784 			}
785 
786 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
787 		}
788 
789 		if (kernelend >= phys_avail[j] &&
790 		    kernelend < phys_avail[j+1]) {
791 			if (kernelstart > phys_avail[j]) {
792 				phys_avail[2*phys_avail_count] = phys_avail[j];
793 				phys_avail[2*phys_avail_count + 1] =
794 				    kernelstart & ~PAGE_MASK;
795 				phys_avail_count++;
796 			}
797 
798 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
799 		}
800 	}
801 
802 	physmem = btoc(physsz);
803 
804 	/*
805 	 * Allocate PTEG table.
806 	 */
807 #ifdef PTEGCOUNT
808 	moea_pteg_count = PTEGCOUNT;
809 #else
810 	moea_pteg_count = 0x1000;
811 
812 	while (moea_pteg_count < physmem)
813 		moea_pteg_count <<= 1;
814 
815 	moea_pteg_count >>= 1;
816 #endif /* PTEGCOUNT */
817 
818 	size = moea_pteg_count * sizeof(struct pteg);
819 	CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
820 	    size);
821 	moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
822 	CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
823 	bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
824 	moea_pteg_mask = moea_pteg_count - 1;
825 
826 	/*
827 	 * Allocate pv/overflow lists.
828 	 */
829 	size = sizeof(struct pvo_head) * moea_pteg_count;
830 	moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
831 	    PAGE_SIZE);
832 	CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
833 	for (i = 0; i < moea_pteg_count; i++)
834 		LIST_INIT(&moea_pvo_table[i]);
835 
836 	/*
837 	 * Initialize the lock that synchronizes access to the pteg and pvo
838 	 * tables.
839 	 */
840 	mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF |
841 	    MTX_RECURSE);
842 	mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF);
843 
844 	mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN);
845 
846 	/*
847 	 * Initialise the unmanaged pvo pool.
848 	 */
849 	moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
850 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
851 	moea_bpvo_pool_index = 0;
852 
853 	/*
854 	 * Make sure kernel vsid is allocated as well as VSID 0.
855 	 */
856 	moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
857 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
858 	moea_vsid_bitmap[0] |= 1;
859 
860 	/*
861 	 * Initialize the kernel pmap (which is statically allocated).
862 	 */
863 	PMAP_LOCK_INIT(kernel_pmap);
864 	for (i = 0; i < 16; i++)
865 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
866 	CPU_FILL(&kernel_pmap->pm_active);
867 	RB_INIT(&kernel_pmap->pmap_pvo);
868 
869  	/*
870 	 * Initialize the global pv list lock.
871 	 */
872 	rw_init(&pvh_global_lock, "pmap pv global");
873 
874 	/*
875 	 * Set up the Open Firmware mappings
876 	 */
877 	chosen = OF_finddevice("/chosen");
878 	if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1 &&
879 	    (mmu = OF_instance_to_package(mmui)) != -1 &&
880 	    (sz = OF_getproplen(mmu, "translations")) != -1) {
881 		translations = NULL;
882 		for (i = 0; phys_avail[i] != 0; i += 2) {
883 			if (phys_avail[i + 1] >= sz) {
884 				translations = (struct ofw_map *)phys_avail[i];
885 				break;
886 			}
887 		}
888 		if (translations == NULL)
889 			panic("moea_bootstrap: no space to copy translations");
890 		bzero(translations, sz);
891 		if (OF_getprop(mmu, "translations", translations, sz) == -1)
892 			panic("moea_bootstrap: can't get ofw translations");
893 		CTR0(KTR_PMAP, "moea_bootstrap: translations");
894 		sz /= sizeof(*translations);
895 		qsort(translations, sz, sizeof (*translations), om_cmp);
896 		for (i = 0; i < sz; i++) {
897 			CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
898 			    translations[i].om_pa, translations[i].om_va,
899 			    translations[i].om_len);
900 
901 			/*
902 			 * If the mapping is 1:1, let the RAM and device
903 			 * on-demand BAT tables take care of the translation.
904 			 */
905 			if (translations[i].om_va == translations[i].om_pa)
906 				continue;
907 
908 			/* Enter the pages */
909 			for (off = 0; off < translations[i].om_len;
910 			    off += PAGE_SIZE)
911 				moea_kenter(mmup, translations[i].om_va + off,
912 					    translations[i].om_pa + off);
913 		}
914 	}
915 
916 	/*
917 	 * Calculate the last available physical address.
918 	 */
919 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
920 		;
921 	Maxmem = powerpc_btop(phys_avail[i + 1]);
922 
923 	moea_cpu_bootstrap(mmup,0);
924 
925 	pmap_bootstrapped++;
926 
927 	/*
928 	 * Set the start and end of kva.
929 	 */
930 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
931 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
932 
933 	/*
934 	 * Allocate a kernel stack with a guard page for thread0 and map it
935 	 * into the kernel page map.
936 	 */
937 	pa = moea_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
938 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
939 	virtual_avail = va + kstack_pages * PAGE_SIZE;
940 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
941 	thread0.td_kstack = va;
942 	thread0.td_kstack_pages = kstack_pages;
943 	for (i = 0; i < kstack_pages; i++) {
944 		moea_kenter(mmup, va, pa);
945 		pa += PAGE_SIZE;
946 		va += PAGE_SIZE;
947 	}
948 
949 	/*
950 	 * Allocate virtual address space for the message buffer.
951 	 */
952 	pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE);
953 	msgbufp = (struct msgbuf *)virtual_avail;
954 	va = virtual_avail;
955 	virtual_avail += round_page(msgbufsize);
956 	while (va < virtual_avail) {
957 		moea_kenter(mmup, va, pa);
958 		pa += PAGE_SIZE;
959 		va += PAGE_SIZE;
960 	}
961 
962 	/*
963 	 * Allocate virtual address space for the dynamic percpu area.
964 	 */
965 	pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
966 	dpcpu = (void *)virtual_avail;
967 	va = virtual_avail;
968 	virtual_avail += DPCPU_SIZE;
969 	while (va < virtual_avail) {
970 		moea_kenter(mmup, va, pa);
971 		pa += PAGE_SIZE;
972 		va += PAGE_SIZE;
973 	}
974 	dpcpu_init(dpcpu, 0);
975 }
976 
977 /*
978  * Activate a user pmap.  The pmap must be activated before it's address
979  * space can be accessed in any way.
980  */
981 void
982 moea_activate(mmu_t mmu, struct thread *td)
983 {
984 	pmap_t	pm, pmr;
985 
986 	/*
987 	 * Load all the data we need up front to encourage the compiler to
988 	 * not issue any loads while we have interrupts disabled below.
989 	 */
990 	pm = &td->td_proc->p_vmspace->vm_pmap;
991 	pmr = pm->pmap_phys;
992 
993 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
994 	PCPU_SET(curpmap, pmr);
995 
996 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
997 }
998 
999 void
1000 moea_deactivate(mmu_t mmu, struct thread *td)
1001 {
1002 	pmap_t	pm;
1003 
1004 	pm = &td->td_proc->p_vmspace->vm_pmap;
1005 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1006 	PCPU_SET(curpmap, NULL);
1007 }
1008 
1009 void
1010 moea_unwire(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1011 {
1012 	struct	pvo_entry key, *pvo;
1013 
1014 	PMAP_LOCK(pm);
1015 	key.pvo_vaddr = sva;
1016 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1017 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1018 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1019 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1020 			panic("moea_unwire: pvo %p is missing PVO_WIRED", pvo);
1021 		pvo->pvo_vaddr &= ~PVO_WIRED;
1022 		pm->pm_stats.wired_count--;
1023 	}
1024 	PMAP_UNLOCK(pm);
1025 }
1026 
1027 void
1028 moea_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1029 {
1030 	vm_offset_t	dst;
1031 	vm_offset_t	src;
1032 
1033 	dst = VM_PAGE_TO_PHYS(mdst);
1034 	src = VM_PAGE_TO_PHYS(msrc);
1035 
1036 	bcopy((void *)src, (void *)dst, PAGE_SIZE);
1037 }
1038 
1039 void
1040 moea_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1041     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1042 {
1043 	void *a_cp, *b_cp;
1044 	vm_offset_t a_pg_offset, b_pg_offset;
1045 	int cnt;
1046 
1047 	while (xfersize > 0) {
1048 		a_pg_offset = a_offset & PAGE_MASK;
1049 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1050 		a_cp = (char *)VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]) +
1051 		    a_pg_offset;
1052 		b_pg_offset = b_offset & PAGE_MASK;
1053 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1054 		b_cp = (char *)VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]) +
1055 		    b_pg_offset;
1056 		bcopy(a_cp, b_cp, cnt);
1057 		a_offset += cnt;
1058 		b_offset += cnt;
1059 		xfersize -= cnt;
1060 	}
1061 }
1062 
1063 /*
1064  * Zero a page of physical memory by temporarily mapping it into the tlb.
1065  */
1066 void
1067 moea_zero_page(mmu_t mmu, vm_page_t m)
1068 {
1069 	vm_offset_t off, pa = VM_PAGE_TO_PHYS(m);
1070 
1071 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1072 		__asm __volatile("dcbz 0,%0" :: "r"(pa + off));
1073 }
1074 
1075 void
1076 moea_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1077 {
1078 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1079 	void *va = (void *)(pa + off);
1080 
1081 	bzero(va, size);
1082 }
1083 
1084 void
1085 moea_zero_page_idle(mmu_t mmu, vm_page_t m)
1086 {
1087 
1088 	moea_zero_page(mmu, m);
1089 }
1090 
1091 vm_offset_t
1092 moea_quick_enter_page(mmu_t mmu, vm_page_t m)
1093 {
1094 
1095 	return (VM_PAGE_TO_PHYS(m));
1096 }
1097 
1098 void
1099 moea_quick_remove_page(mmu_t mmu, vm_offset_t addr)
1100 {
1101 }
1102 
1103 /*
1104  * Map the given physical page at the specified virtual address in the
1105  * target pmap with the protection requested.  If specified the page
1106  * will be wired down.
1107  */
1108 int
1109 moea_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1110     u_int flags, int8_t psind)
1111 {
1112 	int error;
1113 
1114 	for (;;) {
1115 		rw_wlock(&pvh_global_lock);
1116 		PMAP_LOCK(pmap);
1117 		error = moea_enter_locked(pmap, va, m, prot, flags, psind);
1118 		rw_wunlock(&pvh_global_lock);
1119 		PMAP_UNLOCK(pmap);
1120 		if (error != ENOMEM)
1121 			return (KERN_SUCCESS);
1122 		if ((flags & PMAP_ENTER_NOSLEEP) != 0)
1123 			return (KERN_RESOURCE_SHORTAGE);
1124 		VM_OBJECT_ASSERT_UNLOCKED(m->object);
1125 		VM_WAIT;
1126 	}
1127 }
1128 
1129 /*
1130  * Map the given physical page at the specified virtual address in the
1131  * target pmap with the protection requested.  If specified the page
1132  * will be wired down.
1133  *
1134  * The global pvh and pmap must be locked.
1135  */
1136 static int
1137 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1138     u_int flags, int8_t psind __unused)
1139 {
1140 	struct		pvo_head *pvo_head;
1141 	uma_zone_t	zone;
1142 	u_int		pte_lo, pvo_flags;
1143 	int		error;
1144 
1145 	if (pmap_bootstrapped)
1146 		rw_assert(&pvh_global_lock, RA_WLOCKED);
1147 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1148 	if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m))
1149 		VM_OBJECT_ASSERT_LOCKED(m->object);
1150 
1151 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea_initialized) {
1152 		pvo_head = &moea_pvo_kunmanaged;
1153 		zone = moea_upvo_zone;
1154 		pvo_flags = 0;
1155 	} else {
1156 		pvo_head = vm_page_to_pvoh(m);
1157 		zone = moea_mpvo_zone;
1158 		pvo_flags = PVO_MANAGED;
1159 	}
1160 
1161 	pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1162 
1163 	if (prot & VM_PROT_WRITE) {
1164 		pte_lo |= PTE_BW;
1165 		if (pmap_bootstrapped &&
1166 		    (m->oflags & VPO_UNMANAGED) == 0)
1167 			vm_page_aflag_set(m, PGA_WRITEABLE);
1168 	} else
1169 		pte_lo |= PTE_BR;
1170 
1171 	if ((flags & PMAP_ENTER_WIRED) != 0)
1172 		pvo_flags |= PVO_WIRED;
1173 
1174 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1175 	    pte_lo, pvo_flags);
1176 
1177 	/*
1178 	 * Flush the real page from the instruction cache. This has be done
1179 	 * for all user mappings to prevent information leakage via the
1180 	 * instruction cache. moea_pvo_enter() returns ENOENT for the first
1181 	 * mapping for a page.
1182 	 */
1183 	if (pmap != kernel_pmap && error == ENOENT &&
1184 	    (pte_lo & (PTE_I | PTE_G)) == 0)
1185 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1186 
1187 	return (error);
1188 }
1189 
1190 /*
1191  * Maps a sequence of resident pages belonging to the same object.
1192  * The sequence begins with the given page m_start.  This page is
1193  * mapped at the given virtual address start.  Each subsequent page is
1194  * mapped at a virtual address that is offset from start by the same
1195  * amount as the page is offset from m_start within the object.  The
1196  * last page in the sequence is the page with the largest offset from
1197  * m_start that can be mapped at a virtual address less than the given
1198  * virtual address end.  Not every virtual page between start and end
1199  * is mapped; only those for which a resident page exists with the
1200  * corresponding offset from m_start are mapped.
1201  */
1202 void
1203 moea_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1204     vm_page_t m_start, vm_prot_t prot)
1205 {
1206 	vm_page_t m;
1207 	vm_pindex_t diff, psize;
1208 
1209 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1210 
1211 	psize = atop(end - start);
1212 	m = m_start;
1213 	rw_wlock(&pvh_global_lock);
1214 	PMAP_LOCK(pm);
1215 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1216 		moea_enter_locked(pm, start + ptoa(diff), m, prot &
1217 		    (VM_PROT_READ | VM_PROT_EXECUTE), 0, 0);
1218 		m = TAILQ_NEXT(m, listq);
1219 	}
1220 	rw_wunlock(&pvh_global_lock);
1221 	PMAP_UNLOCK(pm);
1222 }
1223 
1224 void
1225 moea_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1226     vm_prot_t prot)
1227 {
1228 
1229 	rw_wlock(&pvh_global_lock);
1230 	PMAP_LOCK(pm);
1231 	moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1232 	    0, 0);
1233 	rw_wunlock(&pvh_global_lock);
1234 	PMAP_UNLOCK(pm);
1235 }
1236 
1237 vm_paddr_t
1238 moea_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1239 {
1240 	struct	pvo_entry *pvo;
1241 	vm_paddr_t pa;
1242 
1243 	PMAP_LOCK(pm);
1244 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1245 	if (pvo == NULL)
1246 		pa = 0;
1247 	else
1248 		pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1249 	PMAP_UNLOCK(pm);
1250 	return (pa);
1251 }
1252 
1253 /*
1254  * Atomically extract and hold the physical page with the given
1255  * pmap and virtual address pair if that mapping permits the given
1256  * protection.
1257  */
1258 vm_page_t
1259 moea_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1260 {
1261 	struct	pvo_entry *pvo;
1262 	vm_page_t m;
1263         vm_paddr_t pa;
1264 
1265 	m = NULL;
1266 	pa = 0;
1267 	PMAP_LOCK(pmap);
1268 retry:
1269 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1270 	if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1271 	    ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1272 	     (prot & VM_PROT_WRITE) == 0)) {
1273 		if (vm_page_pa_tryrelock(pmap, pvo->pvo_pte.pte.pte_lo & PTE_RPGN, &pa))
1274 			goto retry;
1275 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
1276 		vm_page_hold(m);
1277 	}
1278 	PA_UNLOCK_COND(pa);
1279 	PMAP_UNLOCK(pmap);
1280 	return (m);
1281 }
1282 
1283 void
1284 moea_init(mmu_t mmu)
1285 {
1286 
1287 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1288 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1289 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1290 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1291 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1292 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1293 	moea_initialized = TRUE;
1294 }
1295 
1296 boolean_t
1297 moea_is_referenced(mmu_t mmu, vm_page_t m)
1298 {
1299 	boolean_t rv;
1300 
1301 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1302 	    ("moea_is_referenced: page %p is not managed", m));
1303 	rw_wlock(&pvh_global_lock);
1304 	rv = moea_query_bit(m, PTE_REF);
1305 	rw_wunlock(&pvh_global_lock);
1306 	return (rv);
1307 }
1308 
1309 boolean_t
1310 moea_is_modified(mmu_t mmu, vm_page_t m)
1311 {
1312 	boolean_t rv;
1313 
1314 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1315 	    ("moea_is_modified: page %p is not managed", m));
1316 
1317 	/*
1318 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1319 	 * concurrently set while the object is locked.  Thus, if PGA_WRITEABLE
1320 	 * is clear, no PTEs can have PTE_CHG set.
1321 	 */
1322 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1323 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1324 		return (FALSE);
1325 	rw_wlock(&pvh_global_lock);
1326 	rv = moea_query_bit(m, PTE_CHG);
1327 	rw_wunlock(&pvh_global_lock);
1328 	return (rv);
1329 }
1330 
1331 boolean_t
1332 moea_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1333 {
1334 	struct pvo_entry *pvo;
1335 	boolean_t rv;
1336 
1337 	PMAP_LOCK(pmap);
1338 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1339 	rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1340 	PMAP_UNLOCK(pmap);
1341 	return (rv);
1342 }
1343 
1344 void
1345 moea_clear_modify(mmu_t mmu, vm_page_t m)
1346 {
1347 
1348 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1349 	    ("moea_clear_modify: page %p is not managed", m));
1350 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1351 	KASSERT(!vm_page_xbusied(m),
1352 	    ("moea_clear_modify: page %p is exclusive busy", m));
1353 
1354 	/*
1355 	 * If the page is not PGA_WRITEABLE, then no PTEs can have PTE_CHG
1356 	 * set.  If the object containing the page is locked and the page is
1357 	 * not exclusive busied, then PGA_WRITEABLE cannot be concurrently set.
1358 	 */
1359 	if ((m->aflags & PGA_WRITEABLE) == 0)
1360 		return;
1361 	rw_wlock(&pvh_global_lock);
1362 	moea_clear_bit(m, PTE_CHG);
1363 	rw_wunlock(&pvh_global_lock);
1364 }
1365 
1366 /*
1367  * Clear the write and modified bits in each of the given page's mappings.
1368  */
1369 void
1370 moea_remove_write(mmu_t mmu, vm_page_t m)
1371 {
1372 	struct	pvo_entry *pvo;
1373 	struct	pte *pt;
1374 	pmap_t	pmap;
1375 	u_int	lo;
1376 
1377 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1378 	    ("moea_remove_write: page %p is not managed", m));
1379 
1380 	/*
1381 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1382 	 * set by another thread while the object is locked.  Thus,
1383 	 * if PGA_WRITEABLE is clear, no page table entries need updating.
1384 	 */
1385 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1386 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1387 		return;
1388 	rw_wlock(&pvh_global_lock);
1389 	lo = moea_attr_fetch(m);
1390 	powerpc_sync();
1391 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1392 		pmap = pvo->pvo_pmap;
1393 		PMAP_LOCK(pmap);
1394 		if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1395 			pt = moea_pvo_to_pte(pvo, -1);
1396 			pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1397 			pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1398 			if (pt != NULL) {
1399 				moea_pte_synch(pt, &pvo->pvo_pte.pte);
1400 				lo |= pvo->pvo_pte.pte.pte_lo;
1401 				pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1402 				moea_pte_change(pt, &pvo->pvo_pte.pte,
1403 				    pvo->pvo_vaddr);
1404 				mtx_unlock(&moea_table_mutex);
1405 			}
1406 		}
1407 		PMAP_UNLOCK(pmap);
1408 	}
1409 	if ((lo & PTE_CHG) != 0) {
1410 		moea_attr_clear(m, PTE_CHG);
1411 		vm_page_dirty(m);
1412 	}
1413 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1414 	rw_wunlock(&pvh_global_lock);
1415 }
1416 
1417 /*
1418  *	moea_ts_referenced:
1419  *
1420  *	Return a count of reference bits for a page, clearing those bits.
1421  *	It is not necessary for every reference bit to be cleared, but it
1422  *	is necessary that 0 only be returned when there are truly no
1423  *	reference bits set.
1424  *
1425  *	XXX: The exact number of bits to check and clear is a matter that
1426  *	should be tested and standardized at some point in the future for
1427  *	optimal aging of shared pages.
1428  */
1429 int
1430 moea_ts_referenced(mmu_t mmu, vm_page_t m)
1431 {
1432 	int count;
1433 
1434 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1435 	    ("moea_ts_referenced: page %p is not managed", m));
1436 	rw_wlock(&pvh_global_lock);
1437 	count = moea_clear_bit(m, PTE_REF);
1438 	rw_wunlock(&pvh_global_lock);
1439 	return (count);
1440 }
1441 
1442 /*
1443  * Modify the WIMG settings of all mappings for a page.
1444  */
1445 void
1446 moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1447 {
1448 	struct	pvo_entry *pvo;
1449 	struct	pvo_head *pvo_head;
1450 	struct	pte *pt;
1451 	pmap_t	pmap;
1452 	u_int	lo;
1453 
1454 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1455 		m->md.mdpg_cache_attrs = ma;
1456 		return;
1457 	}
1458 
1459 	rw_wlock(&pvh_global_lock);
1460 	pvo_head = vm_page_to_pvoh(m);
1461 	lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1462 
1463 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1464 		pmap = pvo->pvo_pmap;
1465 		PMAP_LOCK(pmap);
1466 		pt = moea_pvo_to_pte(pvo, -1);
1467 		pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1468 		pvo->pvo_pte.pte.pte_lo |= lo;
1469 		if (pt != NULL) {
1470 			moea_pte_change(pt, &pvo->pvo_pte.pte,
1471 			    pvo->pvo_vaddr);
1472 			if (pvo->pvo_pmap == kernel_pmap)
1473 				isync();
1474 		}
1475 		mtx_unlock(&moea_table_mutex);
1476 		PMAP_UNLOCK(pmap);
1477 	}
1478 	m->md.mdpg_cache_attrs = ma;
1479 	rw_wunlock(&pvh_global_lock);
1480 }
1481 
1482 /*
1483  * Map a wired page into kernel virtual address space.
1484  */
1485 void
1486 moea_kenter(mmu_t mmu, vm_offset_t va, vm_paddr_t pa)
1487 {
1488 
1489 	moea_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1490 }
1491 
1492 void
1493 moea_kenter_attr(mmu_t mmu, vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
1494 {
1495 	u_int		pte_lo;
1496 	int		error;
1497 
1498 #if 0
1499 	if (va < VM_MIN_KERNEL_ADDRESS)
1500 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1501 		    va);
1502 #endif
1503 
1504 	pte_lo = moea_calc_wimg(pa, ma);
1505 
1506 	PMAP_LOCK(kernel_pmap);
1507 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1508 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1509 
1510 	if (error != 0 && error != ENOENT)
1511 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1512 		    pa, error);
1513 
1514 	PMAP_UNLOCK(kernel_pmap);
1515 }
1516 
1517 /*
1518  * Extract the physical page address associated with the given kernel virtual
1519  * address.
1520  */
1521 vm_paddr_t
1522 moea_kextract(mmu_t mmu, vm_offset_t va)
1523 {
1524 	struct		pvo_entry *pvo;
1525 	vm_paddr_t pa;
1526 
1527 	/*
1528 	 * Allow direct mappings on 32-bit OEA
1529 	 */
1530 	if (va < VM_MIN_KERNEL_ADDRESS) {
1531 		return (va);
1532 	}
1533 
1534 	PMAP_LOCK(kernel_pmap);
1535 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1536 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1537 	pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1538 	PMAP_UNLOCK(kernel_pmap);
1539 	return (pa);
1540 }
1541 
1542 /*
1543  * Remove a wired page from kernel virtual address space.
1544  */
1545 void
1546 moea_kremove(mmu_t mmu, vm_offset_t va)
1547 {
1548 
1549 	moea_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1550 }
1551 
1552 /*
1553  * Map a range of physical addresses into kernel virtual address space.
1554  *
1555  * The value passed in *virt is a suggested virtual address for the mapping.
1556  * Architectures which can support a direct-mapped physical to virtual region
1557  * can return the appropriate address within that region, leaving '*virt'
1558  * unchanged.  We cannot and therefore do not; *virt is updated with the
1559  * first usable address after the mapped region.
1560  */
1561 vm_offset_t
1562 moea_map(mmu_t mmu, vm_offset_t *virt, vm_paddr_t pa_start,
1563     vm_paddr_t pa_end, int prot)
1564 {
1565 	vm_offset_t	sva, va;
1566 
1567 	sva = *virt;
1568 	va = sva;
1569 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1570 		moea_kenter(mmu, va, pa_start);
1571 	*virt = va;
1572 	return (sva);
1573 }
1574 
1575 /*
1576  * Returns true if the pmap's pv is one of the first
1577  * 16 pvs linked to from this page.  This count may
1578  * be changed upwards or downwards in the future; it
1579  * is only necessary that true be returned for a small
1580  * subset of pmaps for proper page aging.
1581  */
1582 boolean_t
1583 moea_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1584 {
1585         int loops;
1586 	struct pvo_entry *pvo;
1587 	boolean_t rv;
1588 
1589 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1590 	    ("moea_page_exists_quick: page %p is not managed", m));
1591 	loops = 0;
1592 	rv = FALSE;
1593 	rw_wlock(&pvh_global_lock);
1594 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1595 		if (pvo->pvo_pmap == pmap) {
1596 			rv = TRUE;
1597 			break;
1598 		}
1599 		if (++loops >= 16)
1600 			break;
1601 	}
1602 	rw_wunlock(&pvh_global_lock);
1603 	return (rv);
1604 }
1605 
1606 /*
1607  * Return the number of managed mappings to the given physical page
1608  * that are wired.
1609  */
1610 int
1611 moea_page_wired_mappings(mmu_t mmu, vm_page_t m)
1612 {
1613 	struct pvo_entry *pvo;
1614 	int count;
1615 
1616 	count = 0;
1617 	if ((m->oflags & VPO_UNMANAGED) != 0)
1618 		return (count);
1619 	rw_wlock(&pvh_global_lock);
1620 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1621 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1622 			count++;
1623 	rw_wunlock(&pvh_global_lock);
1624 	return (count);
1625 }
1626 
1627 static u_int	moea_vsidcontext;
1628 
1629 void
1630 moea_pinit(mmu_t mmu, pmap_t pmap)
1631 {
1632 	int	i, mask;
1633 	u_int	entropy;
1634 
1635 	KASSERT((int)pmap < VM_MIN_KERNEL_ADDRESS, ("moea_pinit: virt pmap"));
1636 	RB_INIT(&pmap->pmap_pvo);
1637 
1638 	entropy = 0;
1639 	__asm __volatile("mftb %0" : "=r"(entropy));
1640 
1641 	if ((pmap->pmap_phys = (pmap_t)moea_kextract(mmu, (vm_offset_t)pmap))
1642 	    == NULL) {
1643 		pmap->pmap_phys = pmap;
1644 	}
1645 
1646 
1647 	mtx_lock(&moea_vsid_mutex);
1648 	/*
1649 	 * Allocate some segment registers for this pmap.
1650 	 */
1651 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1652 		u_int	hash, n;
1653 
1654 		/*
1655 		 * Create a new value by mutiplying by a prime and adding in
1656 		 * entropy from the timebase register.  This is to make the
1657 		 * VSID more random so that the PT hash function collides
1658 		 * less often.  (Note that the prime casues gcc to do shifts
1659 		 * instead of a multiply.)
1660 		 */
1661 		moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1662 		hash = moea_vsidcontext & (NPMAPS - 1);
1663 		if (hash == 0)		/* 0 is special, avoid it */
1664 			continue;
1665 		n = hash >> 5;
1666 		mask = 1 << (hash & (VSID_NBPW - 1));
1667 		hash = (moea_vsidcontext & 0xfffff);
1668 		if (moea_vsid_bitmap[n] & mask) {	/* collision? */
1669 			/* anything free in this bucket? */
1670 			if (moea_vsid_bitmap[n] == 0xffffffff) {
1671 				entropy = (moea_vsidcontext >> 20);
1672 				continue;
1673 			}
1674 			i = ffs(~moea_vsid_bitmap[n]) - 1;
1675 			mask = 1 << i;
1676 			hash &= 0xfffff & ~(VSID_NBPW - 1);
1677 			hash |= i;
1678 		}
1679 		KASSERT(!(moea_vsid_bitmap[n] & mask),
1680 		    ("Allocating in-use VSID group %#x\n", hash));
1681 		moea_vsid_bitmap[n] |= mask;
1682 		for (i = 0; i < 16; i++)
1683 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1684 		mtx_unlock(&moea_vsid_mutex);
1685 		return;
1686 	}
1687 
1688 	mtx_unlock(&moea_vsid_mutex);
1689 	panic("moea_pinit: out of segments");
1690 }
1691 
1692 /*
1693  * Initialize the pmap associated with process 0.
1694  */
1695 void
1696 moea_pinit0(mmu_t mmu, pmap_t pm)
1697 {
1698 
1699 	PMAP_LOCK_INIT(pm);
1700 	moea_pinit(mmu, pm);
1701 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1702 }
1703 
1704 /*
1705  * Set the physical protection on the specified range of this map as requested.
1706  */
1707 void
1708 moea_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1709     vm_prot_t prot)
1710 {
1711 	struct	pvo_entry *pvo, *tpvo, key;
1712 	struct	pte *pt;
1713 
1714 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1715 	    ("moea_protect: non current pmap"));
1716 
1717 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1718 		moea_remove(mmu, pm, sva, eva);
1719 		return;
1720 	}
1721 
1722 	rw_wlock(&pvh_global_lock);
1723 	PMAP_LOCK(pm);
1724 	key.pvo_vaddr = sva;
1725 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1726 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1727 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1728 
1729 		/*
1730 		 * Grab the PTE pointer before we diddle with the cached PTE
1731 		 * copy.
1732 		 */
1733 		pt = moea_pvo_to_pte(pvo, -1);
1734 		/*
1735 		 * Change the protection of the page.
1736 		 */
1737 		pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1738 		pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1739 
1740 		/*
1741 		 * If the PVO is in the page table, update that pte as well.
1742 		 */
1743 		if (pt != NULL) {
1744 			moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1745 			mtx_unlock(&moea_table_mutex);
1746 		}
1747 	}
1748 	rw_wunlock(&pvh_global_lock);
1749 	PMAP_UNLOCK(pm);
1750 }
1751 
1752 /*
1753  * Map a list of wired pages into kernel virtual address space.  This is
1754  * intended for temporary mappings which do not need page modification or
1755  * references recorded.  Existing mappings in the region are overwritten.
1756  */
1757 void
1758 moea_qenter(mmu_t mmu, vm_offset_t sva, vm_page_t *m, int count)
1759 {
1760 	vm_offset_t va;
1761 
1762 	va = sva;
1763 	while (count-- > 0) {
1764 		moea_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1765 		va += PAGE_SIZE;
1766 		m++;
1767 	}
1768 }
1769 
1770 /*
1771  * Remove page mappings from kernel virtual address space.  Intended for
1772  * temporary mappings entered by moea_qenter.
1773  */
1774 void
1775 moea_qremove(mmu_t mmu, vm_offset_t sva, int count)
1776 {
1777 	vm_offset_t va;
1778 
1779 	va = sva;
1780 	while (count-- > 0) {
1781 		moea_kremove(mmu, va);
1782 		va += PAGE_SIZE;
1783 	}
1784 }
1785 
1786 void
1787 moea_release(mmu_t mmu, pmap_t pmap)
1788 {
1789         int idx, mask;
1790 
1791 	/*
1792 	 * Free segment register's VSID
1793 	 */
1794         if (pmap->pm_sr[0] == 0)
1795                 panic("moea_release");
1796 
1797 	mtx_lock(&moea_vsid_mutex);
1798         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1799         mask = 1 << (idx % VSID_NBPW);
1800         idx /= VSID_NBPW;
1801         moea_vsid_bitmap[idx] &= ~mask;
1802 	mtx_unlock(&moea_vsid_mutex);
1803 }
1804 
1805 /*
1806  * Remove the given range of addresses from the specified map.
1807  */
1808 void
1809 moea_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1810 {
1811 	struct	pvo_entry *pvo, *tpvo, key;
1812 
1813 	rw_wlock(&pvh_global_lock);
1814 	PMAP_LOCK(pm);
1815 	key.pvo_vaddr = sva;
1816 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1817 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1818 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1819 		moea_pvo_remove(pvo, -1);
1820 	}
1821 	PMAP_UNLOCK(pm);
1822 	rw_wunlock(&pvh_global_lock);
1823 }
1824 
1825 /*
1826  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1827  * will reflect changes in pte's back to the vm_page.
1828  */
1829 void
1830 moea_remove_all(mmu_t mmu, vm_page_t m)
1831 {
1832 	struct  pvo_head *pvo_head;
1833 	struct	pvo_entry *pvo, *next_pvo;
1834 	pmap_t	pmap;
1835 
1836 	rw_wlock(&pvh_global_lock);
1837 	pvo_head = vm_page_to_pvoh(m);
1838 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1839 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1840 
1841 		pmap = pvo->pvo_pmap;
1842 		PMAP_LOCK(pmap);
1843 		moea_pvo_remove(pvo, -1);
1844 		PMAP_UNLOCK(pmap);
1845 	}
1846 	if ((m->aflags & PGA_WRITEABLE) && moea_query_bit(m, PTE_CHG)) {
1847 		moea_attr_clear(m, PTE_CHG);
1848 		vm_page_dirty(m);
1849 	}
1850 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1851 	rw_wunlock(&pvh_global_lock);
1852 }
1853 
1854 /*
1855  * Allocate a physical page of memory directly from the phys_avail map.
1856  * Can only be called from moea_bootstrap before avail start and end are
1857  * calculated.
1858  */
1859 static vm_offset_t
1860 moea_bootstrap_alloc(vm_size_t size, u_int align)
1861 {
1862 	vm_offset_t	s, e;
1863 	int		i, j;
1864 
1865 	size = round_page(size);
1866 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1867 		if (align != 0)
1868 			s = (phys_avail[i] + align - 1) & ~(align - 1);
1869 		else
1870 			s = phys_avail[i];
1871 		e = s + size;
1872 
1873 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1874 			continue;
1875 
1876 		if (s == phys_avail[i]) {
1877 			phys_avail[i] += size;
1878 		} else if (e == phys_avail[i + 1]) {
1879 			phys_avail[i + 1] -= size;
1880 		} else {
1881 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1882 				phys_avail[j] = phys_avail[j - 2];
1883 				phys_avail[j + 1] = phys_avail[j - 1];
1884 			}
1885 
1886 			phys_avail[i + 3] = phys_avail[i + 1];
1887 			phys_avail[i + 1] = s;
1888 			phys_avail[i + 2] = e;
1889 			phys_avail_count++;
1890 		}
1891 
1892 		return (s);
1893 	}
1894 	panic("moea_bootstrap_alloc: could not allocate memory");
1895 }
1896 
1897 static void
1898 moea_syncicache(vm_paddr_t pa, vm_size_t len)
1899 {
1900 	__syncicache((void *)pa, len);
1901 }
1902 
1903 static int
1904 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1905     vm_offset_t va, vm_paddr_t pa, u_int pte_lo, int flags)
1906 {
1907 	struct	pvo_entry *pvo;
1908 	u_int	sr;
1909 	int	first;
1910 	u_int	ptegidx;
1911 	int	i;
1912 	int     bootstrap;
1913 
1914 	moea_pvo_enter_calls++;
1915 	first = 0;
1916 	bootstrap = 0;
1917 
1918 	/*
1919 	 * Compute the PTE Group index.
1920 	 */
1921 	va &= ~ADDR_POFF;
1922 	sr = va_to_sr(pm->pm_sr, va);
1923 	ptegidx = va_to_pteg(sr, va);
1924 
1925 	/*
1926 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
1927 	 * there is a mapping.
1928 	 */
1929 	mtx_lock(&moea_table_mutex);
1930 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
1931 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
1932 			if ((pvo->pvo_pte.pte.pte_lo & PTE_RPGN) == pa &&
1933 			    (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
1934 			    (pte_lo & PTE_PP)) {
1935 				/*
1936 				 * The PTE is not changing.  Instead, this may
1937 				 * be a request to change the mapping's wired
1938 				 * attribute.
1939 				 */
1940 				mtx_unlock(&moea_table_mutex);
1941 				if ((flags & PVO_WIRED) != 0 &&
1942 				    (pvo->pvo_vaddr & PVO_WIRED) == 0) {
1943 					pvo->pvo_vaddr |= PVO_WIRED;
1944 					pm->pm_stats.wired_count++;
1945 				} else if ((flags & PVO_WIRED) == 0 &&
1946 				    (pvo->pvo_vaddr & PVO_WIRED) != 0) {
1947 					pvo->pvo_vaddr &= ~PVO_WIRED;
1948 					pm->pm_stats.wired_count--;
1949 				}
1950 				return (0);
1951 			}
1952 			moea_pvo_remove(pvo, -1);
1953 			break;
1954 		}
1955 	}
1956 
1957 	/*
1958 	 * If we aren't overwriting a mapping, try to allocate.
1959 	 */
1960 	if (moea_initialized) {
1961 		pvo = uma_zalloc(zone, M_NOWAIT);
1962 	} else {
1963 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
1964 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
1965 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
1966 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
1967 		}
1968 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
1969 		moea_bpvo_pool_index++;
1970 		bootstrap = 1;
1971 	}
1972 
1973 	if (pvo == NULL) {
1974 		mtx_unlock(&moea_table_mutex);
1975 		return (ENOMEM);
1976 	}
1977 
1978 	moea_pvo_entries++;
1979 	pvo->pvo_vaddr = va;
1980 	pvo->pvo_pmap = pm;
1981 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
1982 	pvo->pvo_vaddr &= ~ADDR_POFF;
1983 	if (flags & PVO_WIRED)
1984 		pvo->pvo_vaddr |= PVO_WIRED;
1985 	if (pvo_head != &moea_pvo_kunmanaged)
1986 		pvo->pvo_vaddr |= PVO_MANAGED;
1987 	if (bootstrap)
1988 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
1989 
1990 	moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo);
1991 
1992 	/*
1993 	 * Add to pmap list
1994 	 */
1995 	RB_INSERT(pvo_tree, &pm->pmap_pvo, pvo);
1996 
1997 	/*
1998 	 * Remember if the list was empty and therefore will be the first
1999 	 * item.
2000 	 */
2001 	if (LIST_FIRST(pvo_head) == NULL)
2002 		first = 1;
2003 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2004 
2005 	if (pvo->pvo_vaddr & PVO_WIRED)
2006 		pm->pm_stats.wired_count++;
2007 	pm->pm_stats.resident_count++;
2008 
2009 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2010 	KASSERT(i < 8, ("Invalid PTE index"));
2011 	if (i >= 0) {
2012 		PVO_PTEGIDX_SET(pvo, i);
2013 	} else {
2014 		panic("moea_pvo_enter: overflow");
2015 		moea_pte_overflow++;
2016 	}
2017 	mtx_unlock(&moea_table_mutex);
2018 
2019 	return (first ? ENOENT : 0);
2020 }
2021 
2022 static void
2023 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
2024 {
2025 	struct	pte *pt;
2026 
2027 	/*
2028 	 * If there is an active pte entry, we need to deactivate it (and
2029 	 * save the ref & cfg bits).
2030 	 */
2031 	pt = moea_pvo_to_pte(pvo, pteidx);
2032 	if (pt != NULL) {
2033 		moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
2034 		mtx_unlock(&moea_table_mutex);
2035 		PVO_PTEGIDX_CLR(pvo);
2036 	} else {
2037 		moea_pte_overflow--;
2038 	}
2039 
2040 	/*
2041 	 * Update our statistics.
2042 	 */
2043 	pvo->pvo_pmap->pm_stats.resident_count--;
2044 	if (pvo->pvo_vaddr & PVO_WIRED)
2045 		pvo->pvo_pmap->pm_stats.wired_count--;
2046 
2047 	/*
2048 	 * Save the REF/CHG bits into their cache if the page is managed.
2049 	 */
2050 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED) {
2051 		struct	vm_page *pg;
2052 
2053 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
2054 		if (pg != NULL) {
2055 			moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2056 			    (PTE_REF | PTE_CHG));
2057 		}
2058 	}
2059 
2060 	/*
2061 	 * Remove this PVO from the PV and pmap lists.
2062 	 */
2063 	LIST_REMOVE(pvo, pvo_vlink);
2064 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2065 
2066 	/*
2067 	 * Remove this from the overflow list and return it to the pool
2068 	 * if we aren't going to reuse it.
2069 	 */
2070 	LIST_REMOVE(pvo, pvo_olink);
2071 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2072 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
2073 		    moea_upvo_zone, pvo);
2074 	moea_pvo_entries--;
2075 	moea_pvo_remove_calls++;
2076 }
2077 
2078 static __inline int
2079 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2080 {
2081 	int	pteidx;
2082 
2083 	/*
2084 	 * We can find the actual pte entry without searching by grabbing
2085 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2086 	 * noticing the HID bit.
2087 	 */
2088 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2089 	if (pvo->pvo_pte.pte.pte_hi & PTE_HID)
2090 		pteidx ^= moea_pteg_mask * 8;
2091 
2092 	return (pteidx);
2093 }
2094 
2095 static struct pvo_entry *
2096 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2097 {
2098 	struct	pvo_entry *pvo;
2099 	int	ptegidx;
2100 	u_int	sr;
2101 
2102 	va &= ~ADDR_POFF;
2103 	sr = va_to_sr(pm->pm_sr, va);
2104 	ptegidx = va_to_pteg(sr, va);
2105 
2106 	mtx_lock(&moea_table_mutex);
2107 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2108 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2109 			if (pteidx_p)
2110 				*pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2111 			break;
2112 		}
2113 	}
2114 	mtx_unlock(&moea_table_mutex);
2115 
2116 	return (pvo);
2117 }
2118 
2119 static struct pte *
2120 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2121 {
2122 	struct	pte *pt;
2123 
2124 	/*
2125 	 * If we haven't been supplied the ptegidx, calculate it.
2126 	 */
2127 	if (pteidx == -1) {
2128 		int	ptegidx;
2129 		u_int	sr;
2130 
2131 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2132 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2133 		pteidx = moea_pvo_pte_index(pvo, ptegidx);
2134 	}
2135 
2136 	pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2137 	mtx_lock(&moea_table_mutex);
2138 
2139 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2140 		panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2141 		    "valid pte index", pvo);
2142 	}
2143 
2144 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2145 		panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2146 		    "pvo but no valid pte", pvo);
2147 	}
2148 
2149 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2150 		if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) {
2151 			panic("moea_pvo_to_pte: pvo %p has valid pte in "
2152 			    "moea_pteg_table %p but invalid in pvo", pvo, pt);
2153 		}
2154 
2155 		if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2156 		    != 0) {
2157 			panic("moea_pvo_to_pte: pvo %p pte does not match "
2158 			    "pte %p in moea_pteg_table", pvo, pt);
2159 		}
2160 
2161 		mtx_assert(&moea_table_mutex, MA_OWNED);
2162 		return (pt);
2163 	}
2164 
2165 	if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) {
2166 		panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2167 		    "moea_pteg_table but valid in pvo: %8x, %8x", pvo, pt, pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2168 	}
2169 
2170 	mtx_unlock(&moea_table_mutex);
2171 	return (NULL);
2172 }
2173 
2174 /*
2175  * XXX: THIS STUFF SHOULD BE IN pte.c?
2176  */
2177 int
2178 moea_pte_spill(vm_offset_t addr)
2179 {
2180 	struct	pvo_entry *source_pvo, *victim_pvo;
2181 	struct	pvo_entry *pvo;
2182 	int	ptegidx, i, j;
2183 	u_int	sr;
2184 	struct	pteg *pteg;
2185 	struct	pte *pt;
2186 
2187 	moea_pte_spills++;
2188 
2189 	sr = mfsrin(addr);
2190 	ptegidx = va_to_pteg(sr, addr);
2191 
2192 	/*
2193 	 * Have to substitute some entry.  Use the primary hash for this.
2194 	 * Use low bits of timebase as random generator.
2195 	 */
2196 	pteg = &moea_pteg_table[ptegidx];
2197 	mtx_lock(&moea_table_mutex);
2198 	__asm __volatile("mftb %0" : "=r"(i));
2199 	i &= 7;
2200 	pt = &pteg->pt[i];
2201 
2202 	source_pvo = NULL;
2203 	victim_pvo = NULL;
2204 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2205 		/*
2206 		 * We need to find a pvo entry for this address.
2207 		 */
2208 		if (source_pvo == NULL &&
2209 		    moea_pte_match(&pvo->pvo_pte.pte, sr, addr,
2210 		    pvo->pvo_pte.pte.pte_hi & PTE_HID)) {
2211 			/*
2212 			 * Now found an entry to be spilled into the pteg.
2213 			 * The PTE is now valid, so we know it's active.
2214 			 */
2215 			j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2216 
2217 			if (j >= 0) {
2218 				PVO_PTEGIDX_SET(pvo, j);
2219 				moea_pte_overflow--;
2220 				mtx_unlock(&moea_table_mutex);
2221 				return (1);
2222 			}
2223 
2224 			source_pvo = pvo;
2225 
2226 			if (victim_pvo != NULL)
2227 				break;
2228 		}
2229 
2230 		/*
2231 		 * We also need the pvo entry of the victim we are replacing
2232 		 * so save the R & C bits of the PTE.
2233 		 */
2234 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2235 		    moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2236 			victim_pvo = pvo;
2237 			if (source_pvo != NULL)
2238 				break;
2239 		}
2240 	}
2241 
2242 	if (source_pvo == NULL) {
2243 		mtx_unlock(&moea_table_mutex);
2244 		return (0);
2245 	}
2246 
2247 	if (victim_pvo == NULL) {
2248 		if ((pt->pte_hi & PTE_HID) == 0)
2249 			panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2250 			    "entry", pt);
2251 
2252 		/*
2253 		 * If this is a secondary PTE, we need to search it's primary
2254 		 * pvo bucket for the matching PVO.
2255 		 */
2256 		LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2257 		    pvo_olink) {
2258 			/*
2259 			 * We also need the pvo entry of the victim we are
2260 			 * replacing so save the R & C bits of the PTE.
2261 			 */
2262 			if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2263 				victim_pvo = pvo;
2264 				break;
2265 			}
2266 		}
2267 
2268 		if (victim_pvo == NULL)
2269 			panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2270 			    "entry", pt);
2271 	}
2272 
2273 	/*
2274 	 * We are invalidating the TLB entry for the EA we are replacing even
2275 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2276 	 * contained in the TLB entry.
2277 	 */
2278 	source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID;
2279 
2280 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2281 	moea_pte_set(pt, &source_pvo->pvo_pte.pte);
2282 
2283 	PVO_PTEGIDX_CLR(victim_pvo);
2284 	PVO_PTEGIDX_SET(source_pvo, i);
2285 	moea_pte_replacements++;
2286 
2287 	mtx_unlock(&moea_table_mutex);
2288 	return (1);
2289 }
2290 
2291 static __inline struct pvo_entry *
2292 moea_pte_spillable_ident(u_int ptegidx)
2293 {
2294 	struct	pte *pt;
2295 	struct	pvo_entry *pvo_walk, *pvo = NULL;
2296 
2297 	LIST_FOREACH(pvo_walk, &moea_pvo_table[ptegidx], pvo_olink) {
2298 		if (pvo_walk->pvo_vaddr & PVO_WIRED)
2299 			continue;
2300 
2301 		if (!(pvo_walk->pvo_pte.pte.pte_hi & PTE_VALID))
2302 			continue;
2303 
2304 		pt = moea_pvo_to_pte(pvo_walk, -1);
2305 
2306 		if (pt == NULL)
2307 			continue;
2308 
2309 		pvo = pvo_walk;
2310 
2311 		mtx_unlock(&moea_table_mutex);
2312 		if (!(pt->pte_lo & PTE_REF))
2313 			return (pvo_walk);
2314 	}
2315 
2316 	return (pvo);
2317 }
2318 
2319 static int
2320 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2321 {
2322 	struct	pte *pt;
2323 	struct	pvo_entry *victim_pvo;
2324 	int	i;
2325 	int	victim_idx;
2326 	u_int	pteg_bkpidx = ptegidx;
2327 
2328 	mtx_assert(&moea_table_mutex, MA_OWNED);
2329 
2330 	/*
2331 	 * First try primary hash.
2332 	 */
2333 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2334 		if ((pt->pte_hi & PTE_VALID) == 0) {
2335 			pvo_pt->pte_hi &= ~PTE_HID;
2336 			moea_pte_set(pt, pvo_pt);
2337 			return (i);
2338 		}
2339 	}
2340 
2341 	/*
2342 	 * Now try secondary hash.
2343 	 */
2344 	ptegidx ^= moea_pteg_mask;
2345 
2346 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2347 		if ((pt->pte_hi & PTE_VALID) == 0) {
2348 			pvo_pt->pte_hi |= PTE_HID;
2349 			moea_pte_set(pt, pvo_pt);
2350 			return (i);
2351 		}
2352 	}
2353 
2354 	/* Try again, but this time try to force a PTE out. */
2355 	ptegidx = pteg_bkpidx;
2356 
2357 	victim_pvo = moea_pte_spillable_ident(ptegidx);
2358 	if (victim_pvo == NULL) {
2359 		ptegidx ^= moea_pteg_mask;
2360 		victim_pvo = moea_pte_spillable_ident(ptegidx);
2361 	}
2362 
2363 	if (victim_pvo == NULL) {
2364 		panic("moea_pte_insert: overflow");
2365 		return (-1);
2366 	}
2367 
2368 	victim_idx = moea_pvo_pte_index(victim_pvo, ptegidx);
2369 
2370 	if (pteg_bkpidx == ptegidx)
2371 		pvo_pt->pte_hi &= ~PTE_HID;
2372 	else
2373 		pvo_pt->pte_hi |= PTE_HID;
2374 
2375 	/*
2376 	 * Synchronize the sacrifice PTE with its PVO, then mark both
2377 	 * invalid. The PVO will be reused when/if the VM system comes
2378 	 * here after a fault.
2379 	 */
2380 	pt = &moea_pteg_table[victim_idx >> 3].pt[victim_idx & 7];
2381 
2382 	if (pt->pte_hi != victim_pvo->pvo_pte.pte.pte_hi)
2383 	    panic("Victim PVO doesn't match PTE! PVO: %8x, PTE: %8x", victim_pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2384 
2385 	/*
2386 	 * Set the new PTE.
2387 	 */
2388 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2389 	PVO_PTEGIDX_CLR(victim_pvo);
2390 	moea_pte_overflow++;
2391 	moea_pte_set(pt, pvo_pt);
2392 
2393 	return (victim_idx & 7);
2394 }
2395 
2396 static boolean_t
2397 moea_query_bit(vm_page_t m, int ptebit)
2398 {
2399 	struct	pvo_entry *pvo;
2400 	struct	pte *pt;
2401 
2402 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2403 	if (moea_attr_fetch(m) & ptebit)
2404 		return (TRUE);
2405 
2406 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2407 
2408 		/*
2409 		 * See if we saved the bit off.  If so, cache it and return
2410 		 * success.
2411 		 */
2412 		if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2413 			moea_attr_save(m, ptebit);
2414 			return (TRUE);
2415 		}
2416 	}
2417 
2418 	/*
2419 	 * No luck, now go through the hard part of looking at the PTEs
2420 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2421 	 * the PTEs.
2422 	 */
2423 	powerpc_sync();
2424 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2425 
2426 		/*
2427 		 * See if this pvo has a valid PTE.  if so, fetch the
2428 		 * REF/CHG bits from the valid PTE.  If the appropriate
2429 		 * ptebit is set, cache it and return success.
2430 		 */
2431 		pt = moea_pvo_to_pte(pvo, -1);
2432 		if (pt != NULL) {
2433 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2434 			mtx_unlock(&moea_table_mutex);
2435 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2436 				moea_attr_save(m, ptebit);
2437 				return (TRUE);
2438 			}
2439 		}
2440 	}
2441 
2442 	return (FALSE);
2443 }
2444 
2445 static u_int
2446 moea_clear_bit(vm_page_t m, int ptebit)
2447 {
2448 	u_int	count;
2449 	struct	pvo_entry *pvo;
2450 	struct	pte *pt;
2451 
2452 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2453 
2454 	/*
2455 	 * Clear the cached value.
2456 	 */
2457 	moea_attr_clear(m, ptebit);
2458 
2459 	/*
2460 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2461 	 * we can reset the right ones).  note that since the pvo entries and
2462 	 * list heads are accessed via BAT0 and are never placed in the page
2463 	 * table, we don't have to worry about further accesses setting the
2464 	 * REF/CHG bits.
2465 	 */
2466 	powerpc_sync();
2467 
2468 	/*
2469 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2470 	 * valid pte clear the ptebit from the valid pte.
2471 	 */
2472 	count = 0;
2473 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2474 		pt = moea_pvo_to_pte(pvo, -1);
2475 		if (pt != NULL) {
2476 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2477 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2478 				count++;
2479 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2480 			}
2481 			mtx_unlock(&moea_table_mutex);
2482 		}
2483 		pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2484 	}
2485 
2486 	return (count);
2487 }
2488 
2489 /*
2490  * Return true if the physical range is encompassed by the battable[idx]
2491  */
2492 static int
2493 moea_bat_mapped(int idx, vm_paddr_t pa, vm_size_t size)
2494 {
2495 	u_int prot;
2496 	u_int32_t start;
2497 	u_int32_t end;
2498 	u_int32_t bat_ble;
2499 
2500 	/*
2501 	 * Return immediately if not a valid mapping
2502 	 */
2503 	if (!(battable[idx].batu & BAT_Vs))
2504 		return (EINVAL);
2505 
2506 	/*
2507 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2508 	 * so it can function as an i/o page
2509 	 */
2510 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2511 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2512 		return (EPERM);
2513 
2514 	/*
2515 	 * The address should be within the BAT range. Assume that the
2516 	 * start address in the BAT has the correct alignment (thus
2517 	 * not requiring masking)
2518 	 */
2519 	start = battable[idx].batl & BAT_PBS;
2520 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2521 	end = start | (bat_ble << 15) | 0x7fff;
2522 
2523 	if ((pa < start) || ((pa + size) > end))
2524 		return (ERANGE);
2525 
2526 	return (0);
2527 }
2528 
2529 boolean_t
2530 moea_dev_direct_mapped(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2531 {
2532 	int i;
2533 
2534 	/*
2535 	 * This currently does not work for entries that
2536 	 * overlap 256M BAT segments.
2537 	 */
2538 
2539 	for(i = 0; i < 16; i++)
2540 		if (moea_bat_mapped(i, pa, size) == 0)
2541 			return (0);
2542 
2543 	return (EFAULT);
2544 }
2545 
2546 /*
2547  * Map a set of physical memory pages into the kernel virtual
2548  * address space. Return a pointer to where it is mapped. This
2549  * routine is intended to be used for mapping device memory,
2550  * NOT real memory.
2551  */
2552 void *
2553 moea_mapdev(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2554 {
2555 
2556 	return (moea_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT));
2557 }
2558 
2559 void *
2560 moea_mapdev_attr(mmu_t mmu, vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
2561 {
2562 	vm_offset_t va, tmpva, ppa, offset;
2563 	int i;
2564 
2565 	ppa = trunc_page(pa);
2566 	offset = pa & PAGE_MASK;
2567 	size = roundup(offset + size, PAGE_SIZE);
2568 
2569 	/*
2570 	 * If the physical address lies within a valid BAT table entry,
2571 	 * return the 1:1 mapping. This currently doesn't work
2572 	 * for regions that overlap 256M BAT segments.
2573 	 */
2574 	for (i = 0; i < 16; i++) {
2575 		if (moea_bat_mapped(i, pa, size) == 0)
2576 			return ((void *) pa);
2577 	}
2578 
2579 	va = kva_alloc(size);
2580 	if (!va)
2581 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2582 
2583 	for (tmpva = va; size > 0;) {
2584 		moea_kenter_attr(mmu, tmpva, ppa, ma);
2585 		tlbie(tmpva);
2586 		size -= PAGE_SIZE;
2587 		tmpva += PAGE_SIZE;
2588 		ppa += PAGE_SIZE;
2589 	}
2590 
2591 	return ((void *)(va + offset));
2592 }
2593 
2594 void
2595 moea_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2596 {
2597 	vm_offset_t base, offset;
2598 
2599 	/*
2600 	 * If this is outside kernel virtual space, then it's a
2601 	 * battable entry and doesn't require unmapping
2602 	 */
2603 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2604 		base = trunc_page(va);
2605 		offset = va & PAGE_MASK;
2606 		size = roundup(offset + size, PAGE_SIZE);
2607 		kva_free(base, size);
2608 	}
2609 }
2610 
2611 static void
2612 moea_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2613 {
2614 	struct pvo_entry *pvo;
2615 	vm_offset_t lim;
2616 	vm_paddr_t pa;
2617 	vm_size_t len;
2618 
2619 	PMAP_LOCK(pm);
2620 	while (sz > 0) {
2621 		lim = round_page(va);
2622 		len = MIN(lim - va, sz);
2623 		pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2624 		if (pvo != NULL) {
2625 			pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) |
2626 			    (va & ADDR_POFF);
2627 			moea_syncicache(pa, len);
2628 		}
2629 		va += len;
2630 		sz -= len;
2631 	}
2632 	PMAP_UNLOCK(pm);
2633 }
2634 
2635 void
2636 moea_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va)
2637 {
2638 
2639 	*va = (void *)pa;
2640 }
2641 
2642 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
2643 
2644 void
2645 moea_scan_init(mmu_t mmu)
2646 {
2647 	struct pvo_entry *pvo;
2648 	vm_offset_t va;
2649 	int i;
2650 
2651 	if (!do_minidump) {
2652 		/* Initialize phys. segments for dumpsys(). */
2653 		memset(&dump_map, 0, sizeof(dump_map));
2654 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
2655 		for (i = 0; i < pregions_sz; i++) {
2656 			dump_map[i].pa_start = pregions[i].mr_start;
2657 			dump_map[i].pa_size = pregions[i].mr_size;
2658 		}
2659 		return;
2660 	}
2661 
2662 	/* Virtual segments for minidumps: */
2663 	memset(&dump_map, 0, sizeof(dump_map));
2664 
2665 	/* 1st: kernel .data and .bss. */
2666 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
2667 	dump_map[0].pa_size =
2668 	    round_page((uintptr_t)_end) - dump_map[0].pa_start;
2669 
2670 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2671 	dump_map[1].pa_start = (vm_paddr_t)msgbufp->msg_ptr;
2672 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
2673 
2674 	/* 3rd: kernel VM. */
2675 	va = dump_map[1].pa_start + dump_map[1].pa_size;
2676 	/* Find start of next chunk (from va). */
2677 	while (va < virtual_end) {
2678 		/* Don't dump the buffer cache. */
2679 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
2680 			va = kmi.buffer_eva;
2681 			continue;
2682 		}
2683 		pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
2684 		if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2685 			break;
2686 		va += PAGE_SIZE;
2687 	}
2688 	if (va < virtual_end) {
2689 		dump_map[2].pa_start = va;
2690 		va += PAGE_SIZE;
2691 		/* Find last page in chunk. */
2692 		while (va < virtual_end) {
2693 			/* Don't run into the buffer cache. */
2694 			if (va == kmi.buffer_sva)
2695 				break;
2696 			pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF,
2697 			    NULL);
2698 			if (pvo == NULL ||
2699 			    !(pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2700 				break;
2701 			va += PAGE_SIZE;
2702 		}
2703 		dump_map[2].pa_size = va - dump_map[2].pa_start;
2704 	}
2705 }
2706