xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision d8b88105c2ccf7686552516877f541efb54fb6c8)
1 /*-
2  * Copyright (c) 2001 The NetBSD Foundation, Inc.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to The NetBSD Foundation
6  * by Matt Thomas <matt@3am-software.com> of Allegro Networks, Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *        This product includes software developed by the NetBSD
19  *        Foundation, Inc. and its contributors.
20  * 4. Neither the name of The NetBSD Foundation nor the names of its
21  *    contributors may be used to endorse or promote products derived
22  *    from this software without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
25  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
26  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
27  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
28  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
29  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
30  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
31  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
32  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
33  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
34  * POSSIBILITY OF SUCH DAMAGE.
35  */
36 /*-
37  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
38  * Copyright (C) 1995, 1996 TooLs GmbH.
39  * All rights reserved.
40  *
41  * Redistribution and use in source and binary forms, with or without
42  * modification, are permitted provided that the following conditions
43  * are met:
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. All advertising materials mentioning features or use of this software
50  *    must display the following acknowledgement:
51  *	This product includes software developed by TooLs GmbH.
52  * 4. The name of TooLs GmbH may not be used to endorse or promote products
53  *    derived from this software without specific prior written permission.
54  *
55  * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
56  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
57  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
58  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
59  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
60  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
61  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
62  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
63  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
64  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
65  *
66  * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $
67  */
68 /*-
69  * Copyright (C) 2001 Benno Rice.
70  * All rights reserved.
71  *
72  * Redistribution and use in source and binary forms, with or without
73  * modification, are permitted provided that the following conditions
74  * are met:
75  * 1. Redistributions of source code must retain the above copyright
76  *    notice, this list of conditions and the following disclaimer.
77  * 2. Redistributions in binary form must reproduce the above copyright
78  *    notice, this list of conditions and the following disclaimer in the
79  *    documentation and/or other materials provided with the distribution.
80  *
81  * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR
82  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
83  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
84  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
85  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
86  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
87  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
88  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
89  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
90  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
91  */
92 
93 #include <sys/cdefs.h>
94 __FBSDID("$FreeBSD$");
95 
96 /*
97  * Manages physical address maps.
98  *
99  * Since the information managed by this module is also stored by the
100  * logical address mapping module, this module may throw away valid virtual
101  * to physical mappings at almost any time.  However, invalidations of
102  * mappings must be done as requested.
103  *
104  * In order to cope with hardware architectures which make virtual to
105  * physical map invalidates expensive, this module may delay invalidate
106  * reduced protection operations until such time as they are actually
107  * necessary.  This module is given full information as to which processors
108  * are currently using which maps, and to when physical maps must be made
109  * correct.
110  */
111 
112 #include "opt_kstack_pages.h"
113 
114 #include <sys/param.h>
115 #include <sys/kernel.h>
116 #include <sys/queue.h>
117 #include <sys/cpuset.h>
118 #include <sys/ktr.h>
119 #include <sys/lock.h>
120 #include <sys/msgbuf.h>
121 #include <sys/mutex.h>
122 #include <sys/proc.h>
123 #include <sys/rwlock.h>
124 #include <sys/sched.h>
125 #include <sys/sysctl.h>
126 #include <sys/systm.h>
127 #include <sys/vmmeter.h>
128 
129 #include <dev/ofw/openfirm.h>
130 
131 #include <vm/vm.h>
132 #include <vm/vm_param.h>
133 #include <vm/vm_kern.h>
134 #include <vm/vm_page.h>
135 #include <vm/vm_map.h>
136 #include <vm/vm_object.h>
137 #include <vm/vm_extern.h>
138 #include <vm/vm_pageout.h>
139 #include <vm/uma.h>
140 
141 #include <machine/cpu.h>
142 #include <machine/platform.h>
143 #include <machine/bat.h>
144 #include <machine/frame.h>
145 #include <machine/md_var.h>
146 #include <machine/psl.h>
147 #include <machine/pte.h>
148 #include <machine/smp.h>
149 #include <machine/sr.h>
150 #include <machine/mmuvar.h>
151 #include <machine/trap.h>
152 
153 #include "mmu_if.h"
154 
155 #define	MOEA_DEBUG
156 
157 #define TODO	panic("%s: not implemented", __func__);
158 
159 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
160 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
161 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
162 
163 struct ofw_map {
164 	vm_offset_t	om_va;
165 	vm_size_t	om_len;
166 	vm_offset_t	om_pa;
167 	u_int		om_mode;
168 };
169 
170 extern unsigned char _etext[];
171 extern unsigned char _end[];
172 
173 extern int dumpsys_minidump;
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_offset_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 void		moea_enter_locked(pmap_t, vm_offset_t, vm_page_t,
269 			    vm_prot_t, boolean_t);
270 static void		moea_syncicache(vm_offset_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(mmu_t, vm_offset_t);
274 int		moea_pte_spill(vm_offset_t);
275 
276 /*
277  * Kernel MMU interface
278  */
279 void moea_change_wiring(mmu_t, pmap_t, vm_offset_t, boolean_t);
280 void moea_clear_modify(mmu_t, vm_page_t);
281 void moea_copy_page(mmu_t, vm_page_t, vm_page_t);
282 void moea_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
283     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
284 void moea_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, boolean_t);
285 void moea_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
286     vm_prot_t);
287 void moea_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
288 vm_paddr_t moea_extract(mmu_t, pmap_t, vm_offset_t);
289 vm_page_t moea_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
290 void moea_init(mmu_t);
291 boolean_t moea_is_modified(mmu_t, vm_page_t);
292 boolean_t moea_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
293 boolean_t moea_is_referenced(mmu_t, vm_page_t);
294 int moea_ts_referenced(mmu_t, vm_page_t);
295 vm_offset_t moea_map(mmu_t, vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
296 boolean_t moea_page_exists_quick(mmu_t, pmap_t, vm_page_t);
297 int moea_page_wired_mappings(mmu_t, vm_page_t);
298 void moea_pinit(mmu_t, pmap_t);
299 void moea_pinit0(mmu_t, pmap_t);
300 void moea_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
301 void moea_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
302 void moea_qremove(mmu_t, vm_offset_t, int);
303 void moea_release(mmu_t, pmap_t);
304 void moea_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
305 void moea_remove_all(mmu_t, vm_page_t);
306 void moea_remove_write(mmu_t, vm_page_t);
307 void moea_zero_page(mmu_t, vm_page_t);
308 void moea_zero_page_area(mmu_t, vm_page_t, int, int);
309 void moea_zero_page_idle(mmu_t, vm_page_t);
310 void moea_activate(mmu_t, struct thread *);
311 void moea_deactivate(mmu_t, struct thread *);
312 void moea_cpu_bootstrap(mmu_t, int);
313 void moea_bootstrap(mmu_t, vm_offset_t, vm_offset_t);
314 void *moea_mapdev(mmu_t, vm_paddr_t, vm_size_t);
315 void *moea_mapdev_attr(mmu_t, vm_offset_t, vm_size_t, vm_memattr_t);
316 void moea_unmapdev(mmu_t, vm_offset_t, vm_size_t);
317 vm_paddr_t moea_kextract(mmu_t, vm_offset_t);
318 void moea_kenter_attr(mmu_t, vm_offset_t, vm_offset_t, vm_memattr_t);
319 void moea_kenter(mmu_t, vm_offset_t, vm_paddr_t);
320 void moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma);
321 boolean_t moea_dev_direct_mapped(mmu_t, vm_paddr_t, vm_size_t);
322 static void moea_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
323 vm_offset_t moea_dumpsys_map(mmu_t mmu, struct pmap_md *md, vm_size_t ofs,
324     vm_size_t *sz);
325 struct pmap_md * moea_scan_md(mmu_t mmu, struct pmap_md *prev);
326 
327 static mmu_method_t moea_methods[] = {
328 	MMUMETHOD(mmu_change_wiring,	moea_change_wiring),
329 	MMUMETHOD(mmu_clear_modify,	moea_clear_modify),
330 	MMUMETHOD(mmu_copy_page,	moea_copy_page),
331 	MMUMETHOD(mmu_copy_pages,	moea_copy_pages),
332 	MMUMETHOD(mmu_enter,		moea_enter),
333 	MMUMETHOD(mmu_enter_object,	moea_enter_object),
334 	MMUMETHOD(mmu_enter_quick,	moea_enter_quick),
335 	MMUMETHOD(mmu_extract,		moea_extract),
336 	MMUMETHOD(mmu_extract_and_hold,	moea_extract_and_hold),
337 	MMUMETHOD(mmu_init,		moea_init),
338 	MMUMETHOD(mmu_is_modified,	moea_is_modified),
339 	MMUMETHOD(mmu_is_prefaultable,	moea_is_prefaultable),
340 	MMUMETHOD(mmu_is_referenced,	moea_is_referenced),
341 	MMUMETHOD(mmu_ts_referenced,	moea_ts_referenced),
342 	MMUMETHOD(mmu_map,     		moea_map),
343 	MMUMETHOD(mmu_page_exists_quick,moea_page_exists_quick),
344 	MMUMETHOD(mmu_page_wired_mappings,moea_page_wired_mappings),
345 	MMUMETHOD(mmu_pinit,		moea_pinit),
346 	MMUMETHOD(mmu_pinit0,		moea_pinit0),
347 	MMUMETHOD(mmu_protect,		moea_protect),
348 	MMUMETHOD(mmu_qenter,		moea_qenter),
349 	MMUMETHOD(mmu_qremove,		moea_qremove),
350 	MMUMETHOD(mmu_release,		moea_release),
351 	MMUMETHOD(mmu_remove,		moea_remove),
352 	MMUMETHOD(mmu_remove_all,      	moea_remove_all),
353 	MMUMETHOD(mmu_remove_write,	moea_remove_write),
354 	MMUMETHOD(mmu_sync_icache,	moea_sync_icache),
355 	MMUMETHOD(mmu_zero_page,       	moea_zero_page),
356 	MMUMETHOD(mmu_zero_page_area,	moea_zero_page_area),
357 	MMUMETHOD(mmu_zero_page_idle,	moea_zero_page_idle),
358 	MMUMETHOD(mmu_activate,		moea_activate),
359 	MMUMETHOD(mmu_deactivate,      	moea_deactivate),
360 	MMUMETHOD(mmu_page_set_memattr,	moea_page_set_memattr),
361 
362 	/* Internal interfaces */
363 	MMUMETHOD(mmu_bootstrap,       	moea_bootstrap),
364 	MMUMETHOD(mmu_cpu_bootstrap,   	moea_cpu_bootstrap),
365 	MMUMETHOD(mmu_mapdev_attr,	moea_mapdev_attr),
366 	MMUMETHOD(mmu_mapdev,		moea_mapdev),
367 	MMUMETHOD(mmu_unmapdev,		moea_unmapdev),
368 	MMUMETHOD(mmu_kextract,		moea_kextract),
369 	MMUMETHOD(mmu_kenter,		moea_kenter),
370 	MMUMETHOD(mmu_kenter_attr,	moea_kenter_attr),
371 	MMUMETHOD(mmu_dev_direct_mapped,moea_dev_direct_mapped),
372 	MMUMETHOD(mmu_scan_md,		moea_scan_md),
373 	MMUMETHOD(mmu_dumpsys_map,	moea_dumpsys_map),
374 
375 	{ 0, 0 }
376 };
377 
378 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods, 0);
379 
380 static __inline uint32_t
381 moea_calc_wimg(vm_offset_t pa, vm_memattr_t ma)
382 {
383 	uint32_t pte_lo;
384 	int i;
385 
386 	if (ma != VM_MEMATTR_DEFAULT) {
387 		switch (ma) {
388 		case VM_MEMATTR_UNCACHEABLE:
389 			return (PTE_I | PTE_G);
390 		case VM_MEMATTR_WRITE_COMBINING:
391 		case VM_MEMATTR_WRITE_BACK:
392 		case VM_MEMATTR_PREFETCHABLE:
393 			return (PTE_I);
394 		case VM_MEMATTR_WRITE_THROUGH:
395 			return (PTE_W | PTE_M);
396 		}
397 	}
398 
399 	/*
400 	 * Assume the page is cache inhibited and access is guarded unless
401 	 * it's in our available memory array.
402 	 */
403 	pte_lo = PTE_I | PTE_G;
404 	for (i = 0; i < pregions_sz; i++) {
405 		if ((pa >= pregions[i].mr_start) &&
406 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
407 			pte_lo = PTE_M;
408 			break;
409 		}
410 	}
411 
412 	return pte_lo;
413 }
414 
415 static void
416 tlbie(vm_offset_t va)
417 {
418 
419 	mtx_lock_spin(&tlbie_mtx);
420 	__asm __volatile("ptesync");
421 	__asm __volatile("tlbie %0" :: "r"(va));
422 	__asm __volatile("eieio; tlbsync; ptesync");
423 	mtx_unlock_spin(&tlbie_mtx);
424 }
425 
426 static void
427 tlbia(void)
428 {
429 	vm_offset_t va;
430 
431 	for (va = 0; va < 0x00040000; va += 0x00001000) {
432 		__asm __volatile("tlbie %0" :: "r"(va));
433 		powerpc_sync();
434 	}
435 	__asm __volatile("tlbsync");
436 	powerpc_sync();
437 }
438 
439 static __inline int
440 va_to_sr(u_int *sr, vm_offset_t va)
441 {
442 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
443 }
444 
445 static __inline u_int
446 va_to_pteg(u_int sr, vm_offset_t addr)
447 {
448 	u_int hash;
449 
450 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
451 	    ADDR_PIDX_SHFT);
452 	return (hash & moea_pteg_mask);
453 }
454 
455 static __inline struct pvo_head *
456 vm_page_to_pvoh(vm_page_t m)
457 {
458 
459 	return (&m->md.mdpg_pvoh);
460 }
461 
462 static __inline void
463 moea_attr_clear(vm_page_t m, int ptebit)
464 {
465 
466 	rw_assert(&pvh_global_lock, RA_WLOCKED);
467 	m->md.mdpg_attrs &= ~ptebit;
468 }
469 
470 static __inline int
471 moea_attr_fetch(vm_page_t m)
472 {
473 
474 	return (m->md.mdpg_attrs);
475 }
476 
477 static __inline void
478 moea_attr_save(vm_page_t m, int ptebit)
479 {
480 
481 	rw_assert(&pvh_global_lock, RA_WLOCKED);
482 	m->md.mdpg_attrs |= ptebit;
483 }
484 
485 static __inline int
486 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
487 {
488 	if (pt->pte_hi == pvo_pt->pte_hi)
489 		return (1);
490 
491 	return (0);
492 }
493 
494 static __inline int
495 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
496 {
497 	return (pt->pte_hi & ~PTE_VALID) ==
498 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
499 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
500 }
501 
502 static __inline void
503 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
504 {
505 
506 	mtx_assert(&moea_table_mutex, MA_OWNED);
507 
508 	/*
509 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
510 	 * set when the real pte is set in memory.
511 	 *
512 	 * Note: Don't set the valid bit for correct operation of tlb update.
513 	 */
514 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
515 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
516 	pt->pte_lo = pte_lo;
517 }
518 
519 static __inline void
520 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
521 {
522 
523 	mtx_assert(&moea_table_mutex, MA_OWNED);
524 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
525 }
526 
527 static __inline void
528 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
529 {
530 
531 	mtx_assert(&moea_table_mutex, MA_OWNED);
532 
533 	/*
534 	 * As shown in Section 7.6.3.2.3
535 	 */
536 	pt->pte_lo &= ~ptebit;
537 	tlbie(va);
538 }
539 
540 static __inline void
541 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
542 {
543 
544 	mtx_assert(&moea_table_mutex, MA_OWNED);
545 	pvo_pt->pte_hi |= PTE_VALID;
546 
547 	/*
548 	 * Update the PTE as defined in section 7.6.3.1.
549 	 * Note that the REF/CHG bits are from pvo_pt and thus should have
550 	 * been saved so this routine can restore them (if desired).
551 	 */
552 	pt->pte_lo = pvo_pt->pte_lo;
553 	powerpc_sync();
554 	pt->pte_hi = pvo_pt->pte_hi;
555 	powerpc_sync();
556 	moea_pte_valid++;
557 }
558 
559 static __inline void
560 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
561 {
562 
563 	mtx_assert(&moea_table_mutex, MA_OWNED);
564 	pvo_pt->pte_hi &= ~PTE_VALID;
565 
566 	/*
567 	 * Force the reg & chg bits back into the PTEs.
568 	 */
569 	powerpc_sync();
570 
571 	/*
572 	 * Invalidate the pte.
573 	 */
574 	pt->pte_hi &= ~PTE_VALID;
575 
576 	tlbie(va);
577 
578 	/*
579 	 * Save the reg & chg bits.
580 	 */
581 	moea_pte_synch(pt, pvo_pt);
582 	moea_pte_valid--;
583 }
584 
585 static __inline void
586 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
587 {
588 
589 	/*
590 	 * Invalidate the PTE
591 	 */
592 	moea_pte_unset(pt, pvo_pt, va);
593 	moea_pte_set(pt, pvo_pt);
594 }
595 
596 /*
597  * Quick sort callout for comparing memory regions.
598  */
599 static int	om_cmp(const void *a, const void *b);
600 
601 static int
602 om_cmp(const void *a, const void *b)
603 {
604 	const struct	ofw_map *mapa;
605 	const struct	ofw_map *mapb;
606 
607 	mapa = a;
608 	mapb = b;
609 	if (mapa->om_pa < mapb->om_pa)
610 		return (-1);
611 	else if (mapa->om_pa > mapb->om_pa)
612 		return (1);
613 	else
614 		return (0);
615 }
616 
617 void
618 moea_cpu_bootstrap(mmu_t mmup, int ap)
619 {
620 	u_int sdr;
621 	int i;
622 
623 	if (ap) {
624 		powerpc_sync();
625 		__asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu));
626 		__asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl));
627 		isync();
628 		__asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu));
629 		__asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl));
630 		isync();
631 	}
632 
633 #ifdef WII
634 	/*
635 	 * Special case for the Wii: don't install the PCI BAT.
636 	 */
637 	if (strcmp(installed_platform(), "wii") != 0) {
638 #endif
639 		__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
640 		__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
641 #ifdef WII
642 	}
643 #endif
644 	isync();
645 
646 	__asm __volatile("mtibatu 1,%0" :: "r"(0));
647 	__asm __volatile("mtdbatu 2,%0" :: "r"(0));
648 	__asm __volatile("mtibatu 2,%0" :: "r"(0));
649 	__asm __volatile("mtdbatu 3,%0" :: "r"(0));
650 	__asm __volatile("mtibatu 3,%0" :: "r"(0));
651 	isync();
652 
653 	for (i = 0; i < 16; i++)
654 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
655 	powerpc_sync();
656 
657 	sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10);
658 	__asm __volatile("mtsdr1 %0" :: "r"(sdr));
659 	isync();
660 
661 	tlbia();
662 }
663 
664 void
665 moea_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
666 {
667 	ihandle_t	mmui;
668 	phandle_t	chosen, mmu;
669 	int		sz;
670 	int		i, j;
671 	vm_size_t	size, physsz, hwphyssz;
672 	vm_offset_t	pa, va, off;
673 	void		*dpcpu;
674 	register_t	msr;
675 
676         /*
677          * Set up BAT0 to map the lowest 256 MB area
678          */
679         battable[0x0].batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
680         battable[0x0].batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
681 
682 	/*
683 	 * Map PCI memory space.
684 	 */
685 	battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
686 	battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
687 
688 	battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
689 	battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
690 
691 	battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
692 	battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
693 
694 	battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
695 	battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
696 
697 	/*
698 	 * Map obio devices.
699 	 */
700 	battable[0xf].batl = BATL(0xf0000000, BAT_I|BAT_G, BAT_PP_RW);
701 	battable[0xf].batu = BATU(0xf0000000, BAT_BL_256M, BAT_Vs);
702 
703 	/*
704 	 * Use an IBAT and a DBAT to map the bottom segment of memory
705 	 * where we are. Turn off instruction relocation temporarily
706 	 * to prevent faults while reprogramming the IBAT.
707 	 */
708 	msr = mfmsr();
709 	mtmsr(msr & ~PSL_IR);
710 	__asm (".balign 32; \n"
711 	       "mtibatu 0,%0; mtibatl 0,%1; isync; \n"
712 	       "mtdbatu 0,%0; mtdbatl 0,%1; isync"
713 	    :: "r"(battable[0].batu), "r"(battable[0].batl));
714 	mtmsr(msr);
715 
716 #ifdef WII
717         if (strcmp(installed_platform(), "wii") != 0) {
718 #endif
719 		/* map pci space */
720 		__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
721 		__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
722 #ifdef WII
723 	}
724 #endif
725 	isync();
726 
727 	/* set global direct map flag */
728 	hw_direct_map = 1;
729 
730 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
731 	CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
732 
733 	for (i = 0; i < pregions_sz; i++) {
734 		vm_offset_t pa;
735 		vm_offset_t end;
736 
737 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
738 			pregions[i].mr_start,
739 			pregions[i].mr_start + pregions[i].mr_size,
740 			pregions[i].mr_size);
741 		/*
742 		 * Install entries into the BAT table to allow all
743 		 * of physmem to be convered by on-demand BAT entries.
744 		 * The loop will sometimes set the same battable element
745 		 * twice, but that's fine since they won't be used for
746 		 * a while yet.
747 		 */
748 		pa = pregions[i].mr_start & 0xf0000000;
749 		end = pregions[i].mr_start + pregions[i].mr_size;
750 		do {
751                         u_int n = pa >> ADDR_SR_SHFT;
752 
753 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
754 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
755 			pa += SEGMENT_LENGTH;
756 		} while (pa < end);
757 	}
758 
759 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
760 		panic("moea_bootstrap: phys_avail too small");
761 
762 	phys_avail_count = 0;
763 	physsz = 0;
764 	hwphyssz = 0;
765 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
766 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
767 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
768 		    regions[i].mr_start + regions[i].mr_size,
769 		    regions[i].mr_size);
770 		if (hwphyssz != 0 &&
771 		    (physsz + regions[i].mr_size) >= hwphyssz) {
772 			if (physsz < hwphyssz) {
773 				phys_avail[j] = regions[i].mr_start;
774 				phys_avail[j + 1] = regions[i].mr_start +
775 				    hwphyssz - physsz;
776 				physsz = hwphyssz;
777 				phys_avail_count++;
778 			}
779 			break;
780 		}
781 		phys_avail[j] = regions[i].mr_start;
782 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
783 		phys_avail_count++;
784 		physsz += regions[i].mr_size;
785 	}
786 
787 	/* Check for overlap with the kernel and exception vectors */
788 	for (j = 0; j < 2*phys_avail_count; j+=2) {
789 		if (phys_avail[j] < EXC_LAST)
790 			phys_avail[j] += EXC_LAST;
791 
792 		if (kernelstart >= phys_avail[j] &&
793 		    kernelstart < phys_avail[j+1]) {
794 			if (kernelend < phys_avail[j+1]) {
795 				phys_avail[2*phys_avail_count] =
796 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
797 				phys_avail[2*phys_avail_count + 1] =
798 				    phys_avail[j+1];
799 				phys_avail_count++;
800 			}
801 
802 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
803 		}
804 
805 		if (kernelend >= phys_avail[j] &&
806 		    kernelend < phys_avail[j+1]) {
807 			if (kernelstart > phys_avail[j]) {
808 				phys_avail[2*phys_avail_count] = phys_avail[j];
809 				phys_avail[2*phys_avail_count + 1] =
810 				    kernelstart & ~PAGE_MASK;
811 				phys_avail_count++;
812 			}
813 
814 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
815 		}
816 	}
817 
818 	physmem = btoc(physsz);
819 
820 	/*
821 	 * Allocate PTEG table.
822 	 */
823 #ifdef PTEGCOUNT
824 	moea_pteg_count = PTEGCOUNT;
825 #else
826 	moea_pteg_count = 0x1000;
827 
828 	while (moea_pteg_count < physmem)
829 		moea_pteg_count <<= 1;
830 
831 	moea_pteg_count >>= 1;
832 #endif /* PTEGCOUNT */
833 
834 	size = moea_pteg_count * sizeof(struct pteg);
835 	CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
836 	    size);
837 	moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
838 	CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
839 	bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
840 	moea_pteg_mask = moea_pteg_count - 1;
841 
842 	/*
843 	 * Allocate pv/overflow lists.
844 	 */
845 	size = sizeof(struct pvo_head) * moea_pteg_count;
846 	moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
847 	    PAGE_SIZE);
848 	CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
849 	for (i = 0; i < moea_pteg_count; i++)
850 		LIST_INIT(&moea_pvo_table[i]);
851 
852 	/*
853 	 * Initialize the lock that synchronizes access to the pteg and pvo
854 	 * tables.
855 	 */
856 	mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF |
857 	    MTX_RECURSE);
858 	mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF);
859 
860 	mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN);
861 
862 	/*
863 	 * Initialise the unmanaged pvo pool.
864 	 */
865 	moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
866 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
867 	moea_bpvo_pool_index = 0;
868 
869 	/*
870 	 * Make sure kernel vsid is allocated as well as VSID 0.
871 	 */
872 	moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
873 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
874 	moea_vsid_bitmap[0] |= 1;
875 
876 	/*
877 	 * Initialize the kernel pmap (which is statically allocated).
878 	 */
879 	PMAP_LOCK_INIT(kernel_pmap);
880 	for (i = 0; i < 16; i++)
881 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
882 	CPU_FILL(&kernel_pmap->pm_active);
883 	RB_INIT(&kernel_pmap->pmap_pvo);
884 
885  	/*
886 	 * Initialize the global pv list lock.
887 	 */
888 	rw_init(&pvh_global_lock, "pmap pv global");
889 
890 	/*
891 	 * Set up the Open Firmware mappings
892 	 */
893 	chosen = OF_finddevice("/chosen");
894 	if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1 &&
895 	    (mmu = OF_instance_to_package(mmui)) != -1 &&
896 	    (sz = OF_getproplen(mmu, "translations")) != -1) {
897 		translations = NULL;
898 		for (i = 0; phys_avail[i] != 0; i += 2) {
899 			if (phys_avail[i + 1] >= sz) {
900 				translations = (struct ofw_map *)phys_avail[i];
901 				break;
902 			}
903 		}
904 		if (translations == NULL)
905 			panic("moea_bootstrap: no space to copy translations");
906 		bzero(translations, sz);
907 		if (OF_getprop(mmu, "translations", translations, sz) == -1)
908 			panic("moea_bootstrap: can't get ofw translations");
909 		CTR0(KTR_PMAP, "moea_bootstrap: translations");
910 		sz /= sizeof(*translations);
911 		qsort(translations, sz, sizeof (*translations), om_cmp);
912 		for (i = 0; i < sz; i++) {
913 			CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
914 			    translations[i].om_pa, translations[i].om_va,
915 			    translations[i].om_len);
916 
917 			/*
918 			 * If the mapping is 1:1, let the RAM and device
919 			 * on-demand BAT tables take care of the translation.
920 			 */
921 			if (translations[i].om_va == translations[i].om_pa)
922 				continue;
923 
924 			/* Enter the pages */
925 			for (off = 0; off < translations[i].om_len;
926 			    off += PAGE_SIZE)
927 				moea_kenter(mmup, translations[i].om_va + off,
928 					    translations[i].om_pa + off);
929 		}
930 	}
931 
932 	/*
933 	 * Calculate the last available physical address.
934 	 */
935 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
936 		;
937 	Maxmem = powerpc_btop(phys_avail[i + 1]);
938 
939 	moea_cpu_bootstrap(mmup,0);
940 
941 	pmap_bootstrapped++;
942 
943 	/*
944 	 * Set the start and end of kva.
945 	 */
946 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
947 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
948 
949 	/*
950 	 * Allocate a kernel stack with a guard page for thread0 and map it
951 	 * into the kernel page map.
952 	 */
953 	pa = moea_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, PAGE_SIZE);
954 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
955 	virtual_avail = va + KSTACK_PAGES * PAGE_SIZE;
956 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
957 	thread0.td_kstack = va;
958 	thread0.td_kstack_pages = KSTACK_PAGES;
959 	for (i = 0; i < KSTACK_PAGES; i++) {
960 		moea_kenter(mmup, va, pa);
961 		pa += PAGE_SIZE;
962 		va += PAGE_SIZE;
963 	}
964 
965 	/*
966 	 * Allocate virtual address space for the message buffer.
967 	 */
968 	pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE);
969 	msgbufp = (struct msgbuf *)virtual_avail;
970 	va = virtual_avail;
971 	virtual_avail += round_page(msgbufsize);
972 	while (va < virtual_avail) {
973 		moea_kenter(mmup, va, pa);
974 		pa += PAGE_SIZE;
975 		va += PAGE_SIZE;
976 	}
977 
978 	/*
979 	 * Allocate virtual address space for the dynamic percpu area.
980 	 */
981 	pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
982 	dpcpu = (void *)virtual_avail;
983 	va = virtual_avail;
984 	virtual_avail += DPCPU_SIZE;
985 	while (va < virtual_avail) {
986 		moea_kenter(mmup, va, pa);
987 		pa += PAGE_SIZE;
988 		va += PAGE_SIZE;
989 	}
990 	dpcpu_init(dpcpu, 0);
991 }
992 
993 /*
994  * Activate a user pmap.  The pmap must be activated before it's address
995  * space can be accessed in any way.
996  */
997 void
998 moea_activate(mmu_t mmu, struct thread *td)
999 {
1000 	pmap_t	pm, pmr;
1001 
1002 	/*
1003 	 * Load all the data we need up front to encourage the compiler to
1004 	 * not issue any loads while we have interrupts disabled below.
1005 	 */
1006 	pm = &td->td_proc->p_vmspace->vm_pmap;
1007 	pmr = pm->pmap_phys;
1008 
1009 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1010 	PCPU_SET(curpmap, pmr);
1011 }
1012 
1013 void
1014 moea_deactivate(mmu_t mmu, struct thread *td)
1015 {
1016 	pmap_t	pm;
1017 
1018 	pm = &td->td_proc->p_vmspace->vm_pmap;
1019 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1020 	PCPU_SET(curpmap, NULL);
1021 }
1022 
1023 void
1024 moea_change_wiring(mmu_t mmu, pmap_t pm, vm_offset_t va, boolean_t wired)
1025 {
1026 	struct	pvo_entry *pvo;
1027 
1028 	PMAP_LOCK(pm);
1029 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1030 
1031 	if (pvo != NULL) {
1032 		if (wired) {
1033 			if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1034 				pm->pm_stats.wired_count++;
1035 			pvo->pvo_vaddr |= PVO_WIRED;
1036 		} else {
1037 			if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1038 				pm->pm_stats.wired_count--;
1039 			pvo->pvo_vaddr &= ~PVO_WIRED;
1040 		}
1041 	}
1042 	PMAP_UNLOCK(pm);
1043 }
1044 
1045 void
1046 moea_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1047 {
1048 	vm_offset_t	dst;
1049 	vm_offset_t	src;
1050 
1051 	dst = VM_PAGE_TO_PHYS(mdst);
1052 	src = VM_PAGE_TO_PHYS(msrc);
1053 
1054 	bcopy((void *)src, (void *)dst, PAGE_SIZE);
1055 }
1056 
1057 void
1058 moea_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1059     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1060 {
1061 	void *a_cp, *b_cp;
1062 	vm_offset_t a_pg_offset, b_pg_offset;
1063 	int cnt;
1064 
1065 	while (xfersize > 0) {
1066 		a_pg_offset = a_offset & PAGE_MASK;
1067 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1068 		a_cp = (char *)VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]) +
1069 		    a_pg_offset;
1070 		b_pg_offset = b_offset & PAGE_MASK;
1071 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1072 		b_cp = (char *)VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]) +
1073 		    b_pg_offset;
1074 		bcopy(a_cp, b_cp, cnt);
1075 		a_offset += cnt;
1076 		b_offset += cnt;
1077 		xfersize -= cnt;
1078 	}
1079 }
1080 
1081 /*
1082  * Zero a page of physical memory by temporarily mapping it into the tlb.
1083  */
1084 void
1085 moea_zero_page(mmu_t mmu, vm_page_t m)
1086 {
1087 	vm_offset_t off, pa = VM_PAGE_TO_PHYS(m);
1088 
1089 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1090 		__asm __volatile("dcbz 0,%0" :: "r"(pa + off));
1091 }
1092 
1093 void
1094 moea_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1095 {
1096 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1097 	void *va = (void *)(pa + off);
1098 
1099 	bzero(va, size);
1100 }
1101 
1102 void
1103 moea_zero_page_idle(mmu_t mmu, vm_page_t m)
1104 {
1105 
1106 	moea_zero_page(mmu, m);
1107 }
1108 
1109 /*
1110  * Map the given physical page at the specified virtual address in the
1111  * target pmap with the protection requested.  If specified the page
1112  * will be wired down.
1113  */
1114 void
1115 moea_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1116 	   boolean_t wired)
1117 {
1118 
1119 	rw_wlock(&pvh_global_lock);
1120 	PMAP_LOCK(pmap);
1121 	moea_enter_locked(pmap, va, m, prot, wired);
1122 	rw_wunlock(&pvh_global_lock);
1123 	PMAP_UNLOCK(pmap);
1124 }
1125 
1126 /*
1127  * Map the given physical page at the specified virtual address in the
1128  * target pmap with the protection requested.  If specified the page
1129  * will be wired down.
1130  *
1131  * The page queues and pmap must be locked.
1132  */
1133 static void
1134 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1135     boolean_t wired)
1136 {
1137 	struct		pvo_head *pvo_head;
1138 	uma_zone_t	zone;
1139 	vm_page_t	pg;
1140 	u_int		pte_lo, pvo_flags;
1141 	int		error;
1142 
1143 	if (!moea_initialized) {
1144 		pvo_head = &moea_pvo_kunmanaged;
1145 		zone = moea_upvo_zone;
1146 		pvo_flags = 0;
1147 		pg = NULL;
1148 	} else {
1149 		pvo_head = vm_page_to_pvoh(m);
1150 		pg = m;
1151 		zone = moea_mpvo_zone;
1152 		pvo_flags = PVO_MANAGED;
1153 	}
1154 	if (pmap_bootstrapped)
1155 		rw_assert(&pvh_global_lock, RA_WLOCKED);
1156 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1157 	if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m))
1158 		VM_OBJECT_ASSERT_LOCKED(m->object);
1159 
1160 	/* XXX change the pvo head for fake pages */
1161 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1162 		pvo_flags &= ~PVO_MANAGED;
1163 		pvo_head = &moea_pvo_kunmanaged;
1164 		zone = moea_upvo_zone;
1165 	}
1166 
1167 	pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1168 
1169 	if (prot & VM_PROT_WRITE) {
1170 		pte_lo |= PTE_BW;
1171 		if (pmap_bootstrapped &&
1172 		    (m->oflags & VPO_UNMANAGED) == 0)
1173 			vm_page_aflag_set(m, PGA_WRITEABLE);
1174 	} else
1175 		pte_lo |= PTE_BR;
1176 
1177 	if (prot & VM_PROT_EXECUTE)
1178 		pvo_flags |= PVO_EXECUTABLE;
1179 
1180 	if (wired)
1181 		pvo_flags |= PVO_WIRED;
1182 
1183 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1184 	    pte_lo, pvo_flags);
1185 
1186 	/*
1187 	 * Flush the real page from the instruction cache. This has be done
1188 	 * for all user mappings to prevent information leakage via the
1189 	 * instruction cache. moea_pvo_enter() returns ENOENT for the first
1190 	 * mapping for a page.
1191 	 */
1192 	if (pmap != kernel_pmap && error == ENOENT &&
1193 	    (pte_lo & (PTE_I | PTE_G)) == 0)
1194 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1195 }
1196 
1197 /*
1198  * Maps a sequence of resident pages belonging to the same object.
1199  * The sequence begins with the given page m_start.  This page is
1200  * mapped at the given virtual address start.  Each subsequent page is
1201  * mapped at a virtual address that is offset from start by the same
1202  * amount as the page is offset from m_start within the object.  The
1203  * last page in the sequence is the page with the largest offset from
1204  * m_start that can be mapped at a virtual address less than the given
1205  * virtual address end.  Not every virtual page between start and end
1206  * is mapped; only those for which a resident page exists with the
1207  * corresponding offset from m_start are mapped.
1208  */
1209 void
1210 moea_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1211     vm_page_t m_start, vm_prot_t prot)
1212 {
1213 	vm_page_t m;
1214 	vm_pindex_t diff, psize;
1215 
1216 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1217 
1218 	psize = atop(end - start);
1219 	m = m_start;
1220 	rw_wlock(&pvh_global_lock);
1221 	PMAP_LOCK(pm);
1222 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1223 		moea_enter_locked(pm, start + ptoa(diff), m, prot &
1224 		    (VM_PROT_READ | VM_PROT_EXECUTE), FALSE);
1225 		m = TAILQ_NEXT(m, listq);
1226 	}
1227 	rw_wunlock(&pvh_global_lock);
1228 	PMAP_UNLOCK(pm);
1229 }
1230 
1231 void
1232 moea_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1233     vm_prot_t prot)
1234 {
1235 
1236 	rw_wlock(&pvh_global_lock);
1237 	PMAP_LOCK(pm);
1238 	moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1239 	    FALSE);
1240 	rw_wunlock(&pvh_global_lock);
1241 	PMAP_UNLOCK(pm);
1242 }
1243 
1244 vm_paddr_t
1245 moea_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1246 {
1247 	struct	pvo_entry *pvo;
1248 	vm_paddr_t pa;
1249 
1250 	PMAP_LOCK(pm);
1251 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1252 	if (pvo == NULL)
1253 		pa = 0;
1254 	else
1255 		pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1256 	PMAP_UNLOCK(pm);
1257 	return (pa);
1258 }
1259 
1260 /*
1261  * Atomically extract and hold the physical page with the given
1262  * pmap and virtual address pair if that mapping permits the given
1263  * protection.
1264  */
1265 vm_page_t
1266 moea_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1267 {
1268 	struct	pvo_entry *pvo;
1269 	vm_page_t m;
1270         vm_paddr_t pa;
1271 
1272 	m = NULL;
1273 	pa = 0;
1274 	PMAP_LOCK(pmap);
1275 retry:
1276 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1277 	if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1278 	    ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1279 	     (prot & VM_PROT_WRITE) == 0)) {
1280 		if (vm_page_pa_tryrelock(pmap, pvo->pvo_pte.pte.pte_lo & PTE_RPGN, &pa))
1281 			goto retry;
1282 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
1283 		vm_page_hold(m);
1284 	}
1285 	PA_UNLOCK_COND(pa);
1286 	PMAP_UNLOCK(pmap);
1287 	return (m);
1288 }
1289 
1290 void
1291 moea_init(mmu_t mmu)
1292 {
1293 
1294 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1295 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1296 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1297 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1298 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1299 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1300 	moea_initialized = TRUE;
1301 }
1302 
1303 boolean_t
1304 moea_is_referenced(mmu_t mmu, vm_page_t m)
1305 {
1306 	boolean_t rv;
1307 
1308 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1309 	    ("moea_is_referenced: page %p is not managed", m));
1310 	rw_wlock(&pvh_global_lock);
1311 	rv = moea_query_bit(m, PTE_REF);
1312 	rw_wunlock(&pvh_global_lock);
1313 	return (rv);
1314 }
1315 
1316 boolean_t
1317 moea_is_modified(mmu_t mmu, vm_page_t m)
1318 {
1319 	boolean_t rv;
1320 
1321 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1322 	    ("moea_is_modified: page %p is not managed", m));
1323 
1324 	/*
1325 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1326 	 * concurrently set while the object is locked.  Thus, if PGA_WRITEABLE
1327 	 * is clear, no PTEs can have PTE_CHG set.
1328 	 */
1329 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1330 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1331 		return (FALSE);
1332 	rw_wlock(&pvh_global_lock);
1333 	rv = moea_query_bit(m, PTE_CHG);
1334 	rw_wunlock(&pvh_global_lock);
1335 	return (rv);
1336 }
1337 
1338 boolean_t
1339 moea_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1340 {
1341 	struct pvo_entry *pvo;
1342 	boolean_t rv;
1343 
1344 	PMAP_LOCK(pmap);
1345 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1346 	rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1347 	PMAP_UNLOCK(pmap);
1348 	return (rv);
1349 }
1350 
1351 void
1352 moea_clear_modify(mmu_t mmu, vm_page_t m)
1353 {
1354 
1355 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1356 	    ("moea_clear_modify: page %p is not managed", m));
1357 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1358 	KASSERT(!vm_page_xbusied(m),
1359 	    ("moea_clear_modify: page %p is exclusive busy", m));
1360 
1361 	/*
1362 	 * If the page is not PGA_WRITEABLE, then no PTEs can have PTE_CHG
1363 	 * set.  If the object containing the page is locked and the page is
1364 	 * not exclusive busied, then PGA_WRITEABLE cannot be concurrently set.
1365 	 */
1366 	if ((m->aflags & PGA_WRITEABLE) == 0)
1367 		return;
1368 	rw_wlock(&pvh_global_lock);
1369 	moea_clear_bit(m, PTE_CHG);
1370 	rw_wunlock(&pvh_global_lock);
1371 }
1372 
1373 /*
1374  * Clear the write and modified bits in each of the given page's mappings.
1375  */
1376 void
1377 moea_remove_write(mmu_t mmu, vm_page_t m)
1378 {
1379 	struct	pvo_entry *pvo;
1380 	struct	pte *pt;
1381 	pmap_t	pmap;
1382 	u_int	lo;
1383 
1384 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1385 	    ("moea_remove_write: page %p is not managed", m));
1386 
1387 	/*
1388 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1389 	 * set by another thread while the object is locked.  Thus,
1390 	 * if PGA_WRITEABLE is clear, no page table entries need updating.
1391 	 */
1392 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1393 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1394 		return;
1395 	rw_wlock(&pvh_global_lock);
1396 	lo = moea_attr_fetch(m);
1397 	powerpc_sync();
1398 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1399 		pmap = pvo->pvo_pmap;
1400 		PMAP_LOCK(pmap);
1401 		if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1402 			pt = moea_pvo_to_pte(pvo, -1);
1403 			pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1404 			pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1405 			if (pt != NULL) {
1406 				moea_pte_synch(pt, &pvo->pvo_pte.pte);
1407 				lo |= pvo->pvo_pte.pte.pte_lo;
1408 				pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1409 				moea_pte_change(pt, &pvo->pvo_pte.pte,
1410 				    pvo->pvo_vaddr);
1411 				mtx_unlock(&moea_table_mutex);
1412 			}
1413 		}
1414 		PMAP_UNLOCK(pmap);
1415 	}
1416 	if ((lo & PTE_CHG) != 0) {
1417 		moea_attr_clear(m, PTE_CHG);
1418 		vm_page_dirty(m);
1419 	}
1420 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1421 	rw_wunlock(&pvh_global_lock);
1422 }
1423 
1424 /*
1425  *	moea_ts_referenced:
1426  *
1427  *	Return a count of reference bits for a page, clearing those bits.
1428  *	It is not necessary for every reference bit to be cleared, but it
1429  *	is necessary that 0 only be returned when there are truly no
1430  *	reference bits set.
1431  *
1432  *	XXX: The exact number of bits to check and clear is a matter that
1433  *	should be tested and standardized at some point in the future for
1434  *	optimal aging of shared pages.
1435  */
1436 int
1437 moea_ts_referenced(mmu_t mmu, vm_page_t m)
1438 {
1439 	int count;
1440 
1441 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1442 	    ("moea_ts_referenced: page %p is not managed", m));
1443 	rw_wlock(&pvh_global_lock);
1444 	count = moea_clear_bit(m, PTE_REF);
1445 	rw_wunlock(&pvh_global_lock);
1446 	return (count);
1447 }
1448 
1449 /*
1450  * Modify the WIMG settings of all mappings for a page.
1451  */
1452 void
1453 moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1454 {
1455 	struct	pvo_entry *pvo;
1456 	struct	pvo_head *pvo_head;
1457 	struct	pte *pt;
1458 	pmap_t	pmap;
1459 	u_int	lo;
1460 
1461 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1462 		m->md.mdpg_cache_attrs = ma;
1463 		return;
1464 	}
1465 
1466 	rw_wlock(&pvh_global_lock);
1467 	pvo_head = vm_page_to_pvoh(m);
1468 	lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1469 
1470 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1471 		pmap = pvo->pvo_pmap;
1472 		PMAP_LOCK(pmap);
1473 		pt = moea_pvo_to_pte(pvo, -1);
1474 		pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1475 		pvo->pvo_pte.pte.pte_lo |= lo;
1476 		if (pt != NULL) {
1477 			moea_pte_change(pt, &pvo->pvo_pte.pte,
1478 			    pvo->pvo_vaddr);
1479 			if (pvo->pvo_pmap == kernel_pmap)
1480 				isync();
1481 		}
1482 		mtx_unlock(&moea_table_mutex);
1483 		PMAP_UNLOCK(pmap);
1484 	}
1485 	m->md.mdpg_cache_attrs = ma;
1486 	rw_wunlock(&pvh_global_lock);
1487 }
1488 
1489 /*
1490  * Map a wired page into kernel virtual address space.
1491  */
1492 void
1493 moea_kenter(mmu_t mmu, vm_offset_t va, vm_paddr_t pa)
1494 {
1495 
1496 	moea_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1497 }
1498 
1499 void
1500 moea_kenter_attr(mmu_t mmu, vm_offset_t va, vm_offset_t pa, vm_memattr_t ma)
1501 {
1502 	u_int		pte_lo;
1503 	int		error;
1504 
1505 #if 0
1506 	if (va < VM_MIN_KERNEL_ADDRESS)
1507 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1508 		    va);
1509 #endif
1510 
1511 	pte_lo = moea_calc_wimg(pa, ma);
1512 
1513 	PMAP_LOCK(kernel_pmap);
1514 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1515 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1516 
1517 	if (error != 0 && error != ENOENT)
1518 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1519 		    pa, error);
1520 
1521 	PMAP_UNLOCK(kernel_pmap);
1522 }
1523 
1524 /*
1525  * Extract the physical page address associated with the given kernel virtual
1526  * address.
1527  */
1528 vm_paddr_t
1529 moea_kextract(mmu_t mmu, vm_offset_t va)
1530 {
1531 	struct		pvo_entry *pvo;
1532 	vm_paddr_t pa;
1533 
1534 	/*
1535 	 * Allow direct mappings on 32-bit OEA
1536 	 */
1537 	if (va < VM_MIN_KERNEL_ADDRESS) {
1538 		return (va);
1539 	}
1540 
1541 	PMAP_LOCK(kernel_pmap);
1542 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1543 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1544 	pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1545 	PMAP_UNLOCK(kernel_pmap);
1546 	return (pa);
1547 }
1548 
1549 /*
1550  * Remove a wired page from kernel virtual address space.
1551  */
1552 void
1553 moea_kremove(mmu_t mmu, vm_offset_t va)
1554 {
1555 
1556 	moea_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1557 }
1558 
1559 /*
1560  * Map a range of physical addresses into kernel virtual address space.
1561  *
1562  * The value passed in *virt is a suggested virtual address for the mapping.
1563  * Architectures which can support a direct-mapped physical to virtual region
1564  * can return the appropriate address within that region, leaving '*virt'
1565  * unchanged.  We cannot and therefore do not; *virt is updated with the
1566  * first usable address after the mapped region.
1567  */
1568 vm_offset_t
1569 moea_map(mmu_t mmu, vm_offset_t *virt, vm_paddr_t pa_start,
1570     vm_paddr_t pa_end, int prot)
1571 {
1572 	vm_offset_t	sva, va;
1573 
1574 	sva = *virt;
1575 	va = sva;
1576 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1577 		moea_kenter(mmu, va, pa_start);
1578 	*virt = va;
1579 	return (sva);
1580 }
1581 
1582 /*
1583  * Returns true if the pmap's pv is one of the first
1584  * 16 pvs linked to from this page.  This count may
1585  * be changed upwards or downwards in the future; it
1586  * is only necessary that true be returned for a small
1587  * subset of pmaps for proper page aging.
1588  */
1589 boolean_t
1590 moea_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1591 {
1592         int loops;
1593 	struct pvo_entry *pvo;
1594 	boolean_t rv;
1595 
1596 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1597 	    ("moea_page_exists_quick: page %p is not managed", m));
1598 	loops = 0;
1599 	rv = FALSE;
1600 	rw_wlock(&pvh_global_lock);
1601 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1602 		if (pvo->pvo_pmap == pmap) {
1603 			rv = TRUE;
1604 			break;
1605 		}
1606 		if (++loops >= 16)
1607 			break;
1608 	}
1609 	rw_wunlock(&pvh_global_lock);
1610 	return (rv);
1611 }
1612 
1613 /*
1614  * Return the number of managed mappings to the given physical page
1615  * that are wired.
1616  */
1617 int
1618 moea_page_wired_mappings(mmu_t mmu, vm_page_t m)
1619 {
1620 	struct pvo_entry *pvo;
1621 	int count;
1622 
1623 	count = 0;
1624 	if ((m->oflags & VPO_UNMANAGED) != 0)
1625 		return (count);
1626 	rw_wlock(&pvh_global_lock);
1627 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1628 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1629 			count++;
1630 	rw_wunlock(&pvh_global_lock);
1631 	return (count);
1632 }
1633 
1634 static u_int	moea_vsidcontext;
1635 
1636 void
1637 moea_pinit(mmu_t mmu, pmap_t pmap)
1638 {
1639 	int	i, mask;
1640 	u_int	entropy;
1641 
1642 	KASSERT((int)pmap < VM_MIN_KERNEL_ADDRESS, ("moea_pinit: virt pmap"));
1643 	RB_INIT(&pmap->pmap_pvo);
1644 
1645 	entropy = 0;
1646 	__asm __volatile("mftb %0" : "=r"(entropy));
1647 
1648 	if ((pmap->pmap_phys = (pmap_t)moea_kextract(mmu, (vm_offset_t)pmap))
1649 	    == NULL) {
1650 		pmap->pmap_phys = pmap;
1651 	}
1652 
1653 
1654 	mtx_lock(&moea_vsid_mutex);
1655 	/*
1656 	 * Allocate some segment registers for this pmap.
1657 	 */
1658 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1659 		u_int	hash, n;
1660 
1661 		/*
1662 		 * Create a new value by mutiplying by a prime and adding in
1663 		 * entropy from the timebase register.  This is to make the
1664 		 * VSID more random so that the PT hash function collides
1665 		 * less often.  (Note that the prime casues gcc to do shifts
1666 		 * instead of a multiply.)
1667 		 */
1668 		moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1669 		hash = moea_vsidcontext & (NPMAPS - 1);
1670 		if (hash == 0)		/* 0 is special, avoid it */
1671 			continue;
1672 		n = hash >> 5;
1673 		mask = 1 << (hash & (VSID_NBPW - 1));
1674 		hash = (moea_vsidcontext & 0xfffff);
1675 		if (moea_vsid_bitmap[n] & mask) {	/* collision? */
1676 			/* anything free in this bucket? */
1677 			if (moea_vsid_bitmap[n] == 0xffffffff) {
1678 				entropy = (moea_vsidcontext >> 20);
1679 				continue;
1680 			}
1681 			i = ffs(~moea_vsid_bitmap[n]) - 1;
1682 			mask = 1 << i;
1683 			hash &= 0xfffff & ~(VSID_NBPW - 1);
1684 			hash |= i;
1685 		}
1686 		KASSERT(!(moea_vsid_bitmap[n] & mask),
1687 		    ("Allocating in-use VSID group %#x\n", hash));
1688 		moea_vsid_bitmap[n] |= mask;
1689 		for (i = 0; i < 16; i++)
1690 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1691 		mtx_unlock(&moea_vsid_mutex);
1692 		return;
1693 	}
1694 
1695 	mtx_unlock(&moea_vsid_mutex);
1696 	panic("moea_pinit: out of segments");
1697 }
1698 
1699 /*
1700  * Initialize the pmap associated with process 0.
1701  */
1702 void
1703 moea_pinit0(mmu_t mmu, pmap_t pm)
1704 {
1705 
1706 	PMAP_LOCK_INIT(pm);
1707 	moea_pinit(mmu, pm);
1708 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1709 }
1710 
1711 /*
1712  * Set the physical protection on the specified range of this map as requested.
1713  */
1714 void
1715 moea_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1716     vm_prot_t prot)
1717 {
1718 	struct	pvo_entry *pvo, *tpvo, key;
1719 	struct	pte *pt;
1720 
1721 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1722 	    ("moea_protect: non current pmap"));
1723 
1724 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1725 		moea_remove(mmu, pm, sva, eva);
1726 		return;
1727 	}
1728 
1729 	rw_wlock(&pvh_global_lock);
1730 	PMAP_LOCK(pm);
1731 	key.pvo_vaddr = sva;
1732 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1733 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1734 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1735 		if ((prot & VM_PROT_EXECUTE) == 0)
1736 			pvo->pvo_vaddr &= ~PVO_EXECUTABLE;
1737 
1738 		/*
1739 		 * Grab the PTE pointer before we diddle with the cached PTE
1740 		 * copy.
1741 		 */
1742 		pt = moea_pvo_to_pte(pvo, -1);
1743 		/*
1744 		 * Change the protection of the page.
1745 		 */
1746 		pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1747 		pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1748 
1749 		/*
1750 		 * If the PVO is in the page table, update that pte as well.
1751 		 */
1752 		if (pt != NULL) {
1753 			moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1754 			mtx_unlock(&moea_table_mutex);
1755 		}
1756 	}
1757 	rw_wunlock(&pvh_global_lock);
1758 	PMAP_UNLOCK(pm);
1759 }
1760 
1761 /*
1762  * Map a list of wired pages into kernel virtual address space.  This is
1763  * intended for temporary mappings which do not need page modification or
1764  * references recorded.  Existing mappings in the region are overwritten.
1765  */
1766 void
1767 moea_qenter(mmu_t mmu, vm_offset_t sva, vm_page_t *m, int count)
1768 {
1769 	vm_offset_t va;
1770 
1771 	va = sva;
1772 	while (count-- > 0) {
1773 		moea_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1774 		va += PAGE_SIZE;
1775 		m++;
1776 	}
1777 }
1778 
1779 /*
1780  * Remove page mappings from kernel virtual address space.  Intended for
1781  * temporary mappings entered by moea_qenter.
1782  */
1783 void
1784 moea_qremove(mmu_t mmu, vm_offset_t sva, int count)
1785 {
1786 	vm_offset_t va;
1787 
1788 	va = sva;
1789 	while (count-- > 0) {
1790 		moea_kremove(mmu, va);
1791 		va += PAGE_SIZE;
1792 	}
1793 }
1794 
1795 void
1796 moea_release(mmu_t mmu, pmap_t pmap)
1797 {
1798         int idx, mask;
1799 
1800 	/*
1801 	 * Free segment register's VSID
1802 	 */
1803         if (pmap->pm_sr[0] == 0)
1804                 panic("moea_release");
1805 
1806 	mtx_lock(&moea_vsid_mutex);
1807         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1808         mask = 1 << (idx % VSID_NBPW);
1809         idx /= VSID_NBPW;
1810         moea_vsid_bitmap[idx] &= ~mask;
1811 	mtx_unlock(&moea_vsid_mutex);
1812 }
1813 
1814 /*
1815  * Remove the given range of addresses from the specified map.
1816  */
1817 void
1818 moea_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1819 {
1820 	struct	pvo_entry *pvo, *tpvo, key;
1821 
1822 	rw_wlock(&pvh_global_lock);
1823 	PMAP_LOCK(pm);
1824 	key.pvo_vaddr = sva;
1825 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1826 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1827 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1828 		moea_pvo_remove(pvo, -1);
1829 	}
1830 	PMAP_UNLOCK(pm);
1831 	rw_wunlock(&pvh_global_lock);
1832 }
1833 
1834 /*
1835  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1836  * will reflect changes in pte's back to the vm_page.
1837  */
1838 void
1839 moea_remove_all(mmu_t mmu, vm_page_t m)
1840 {
1841 	struct  pvo_head *pvo_head;
1842 	struct	pvo_entry *pvo, *next_pvo;
1843 	pmap_t	pmap;
1844 
1845 	rw_wlock(&pvh_global_lock);
1846 	pvo_head = vm_page_to_pvoh(m);
1847 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1848 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1849 
1850 		pmap = pvo->pvo_pmap;
1851 		PMAP_LOCK(pmap);
1852 		moea_pvo_remove(pvo, -1);
1853 		PMAP_UNLOCK(pmap);
1854 	}
1855 	if ((m->aflags & PGA_WRITEABLE) && moea_query_bit(m, PTE_CHG)) {
1856 		moea_attr_clear(m, PTE_CHG);
1857 		vm_page_dirty(m);
1858 	}
1859 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1860 	rw_wunlock(&pvh_global_lock);
1861 }
1862 
1863 /*
1864  * Allocate a physical page of memory directly from the phys_avail map.
1865  * Can only be called from moea_bootstrap before avail start and end are
1866  * calculated.
1867  */
1868 static vm_offset_t
1869 moea_bootstrap_alloc(vm_size_t size, u_int align)
1870 {
1871 	vm_offset_t	s, e;
1872 	int		i, j;
1873 
1874 	size = round_page(size);
1875 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1876 		if (align != 0)
1877 			s = (phys_avail[i] + align - 1) & ~(align - 1);
1878 		else
1879 			s = phys_avail[i];
1880 		e = s + size;
1881 
1882 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1883 			continue;
1884 
1885 		if (s == phys_avail[i]) {
1886 			phys_avail[i] += size;
1887 		} else if (e == phys_avail[i + 1]) {
1888 			phys_avail[i + 1] -= size;
1889 		} else {
1890 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1891 				phys_avail[j] = phys_avail[j - 2];
1892 				phys_avail[j + 1] = phys_avail[j - 1];
1893 			}
1894 
1895 			phys_avail[i + 3] = phys_avail[i + 1];
1896 			phys_avail[i + 1] = s;
1897 			phys_avail[i + 2] = e;
1898 			phys_avail_count++;
1899 		}
1900 
1901 		return (s);
1902 	}
1903 	panic("moea_bootstrap_alloc: could not allocate memory");
1904 }
1905 
1906 static void
1907 moea_syncicache(vm_offset_t pa, vm_size_t len)
1908 {
1909 	__syncicache((void *)pa, len);
1910 }
1911 
1912 static int
1913 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1914     vm_offset_t va, vm_offset_t pa, u_int pte_lo, int flags)
1915 {
1916 	struct	pvo_entry *pvo;
1917 	u_int	sr;
1918 	int	first;
1919 	u_int	ptegidx;
1920 	int	i;
1921 	int     bootstrap;
1922 
1923 	moea_pvo_enter_calls++;
1924 	first = 0;
1925 	bootstrap = 0;
1926 
1927 	/*
1928 	 * Compute the PTE Group index.
1929 	 */
1930 	va &= ~ADDR_POFF;
1931 	sr = va_to_sr(pm->pm_sr, va);
1932 	ptegidx = va_to_pteg(sr, va);
1933 
1934 	/*
1935 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
1936 	 * there is a mapping.
1937 	 */
1938 	mtx_lock(&moea_table_mutex);
1939 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
1940 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
1941 			if ((pvo->pvo_pte.pte.pte_lo & PTE_RPGN) == pa &&
1942 			    (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
1943 			    (pte_lo & PTE_PP)) {
1944 				mtx_unlock(&moea_table_mutex);
1945 				return (0);
1946 			}
1947 			moea_pvo_remove(pvo, -1);
1948 			break;
1949 		}
1950 	}
1951 
1952 	/*
1953 	 * If we aren't overwriting a mapping, try to allocate.
1954 	 */
1955 	if (moea_initialized) {
1956 		pvo = uma_zalloc(zone, M_NOWAIT);
1957 	} else {
1958 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
1959 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
1960 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
1961 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
1962 		}
1963 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
1964 		moea_bpvo_pool_index++;
1965 		bootstrap = 1;
1966 	}
1967 
1968 	if (pvo == NULL) {
1969 		mtx_unlock(&moea_table_mutex);
1970 		return (ENOMEM);
1971 	}
1972 
1973 	moea_pvo_entries++;
1974 	pvo->pvo_vaddr = va;
1975 	pvo->pvo_pmap = pm;
1976 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
1977 	pvo->pvo_vaddr &= ~ADDR_POFF;
1978 	if (flags & VM_PROT_EXECUTE)
1979 		pvo->pvo_vaddr |= PVO_EXECUTABLE;
1980 	if (flags & PVO_WIRED)
1981 		pvo->pvo_vaddr |= PVO_WIRED;
1982 	if (pvo_head != &moea_pvo_kunmanaged)
1983 		pvo->pvo_vaddr |= PVO_MANAGED;
1984 	if (bootstrap)
1985 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
1986 
1987 	moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo);
1988 
1989 	/*
1990 	 * Add to pmap list
1991 	 */
1992 	RB_INSERT(pvo_tree, &pm->pmap_pvo, pvo);
1993 
1994 	/*
1995 	 * Remember if the list was empty and therefore will be the first
1996 	 * item.
1997 	 */
1998 	if (LIST_FIRST(pvo_head) == NULL)
1999 		first = 1;
2000 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2001 
2002 	if (pvo->pvo_pte.pte.pte_lo & PVO_WIRED)
2003 		pm->pm_stats.wired_count++;
2004 	pm->pm_stats.resident_count++;
2005 
2006 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2007 	KASSERT(i < 8, ("Invalid PTE index"));
2008 	if (i >= 0) {
2009 		PVO_PTEGIDX_SET(pvo, i);
2010 	} else {
2011 		panic("moea_pvo_enter: overflow");
2012 		moea_pte_overflow++;
2013 	}
2014 	mtx_unlock(&moea_table_mutex);
2015 
2016 	return (first ? ENOENT : 0);
2017 }
2018 
2019 static void
2020 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
2021 {
2022 	struct	pte *pt;
2023 
2024 	/*
2025 	 * If there is an active pte entry, we need to deactivate it (and
2026 	 * save the ref & cfg bits).
2027 	 */
2028 	pt = moea_pvo_to_pte(pvo, pteidx);
2029 	if (pt != NULL) {
2030 		moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
2031 		mtx_unlock(&moea_table_mutex);
2032 		PVO_PTEGIDX_CLR(pvo);
2033 	} else {
2034 		moea_pte_overflow--;
2035 	}
2036 
2037 	/*
2038 	 * Update our statistics.
2039 	 */
2040 	pvo->pvo_pmap->pm_stats.resident_count--;
2041 	if (pvo->pvo_pte.pte.pte_lo & PVO_WIRED)
2042 		pvo->pvo_pmap->pm_stats.wired_count--;
2043 
2044 	/*
2045 	 * Save the REF/CHG bits into their cache if the page is managed.
2046 	 */
2047 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED) {
2048 		struct	vm_page *pg;
2049 
2050 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
2051 		if (pg != NULL) {
2052 			moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2053 			    (PTE_REF | PTE_CHG));
2054 		}
2055 	}
2056 
2057 	/*
2058 	 * Remove this PVO from the PV and pmap lists.
2059 	 */
2060 	LIST_REMOVE(pvo, pvo_vlink);
2061 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2062 
2063 	/*
2064 	 * Remove this from the overflow list and return it to the pool
2065 	 * if we aren't going to reuse it.
2066 	 */
2067 	LIST_REMOVE(pvo, pvo_olink);
2068 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2069 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
2070 		    moea_upvo_zone, pvo);
2071 	moea_pvo_entries--;
2072 	moea_pvo_remove_calls++;
2073 }
2074 
2075 static __inline int
2076 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2077 {
2078 	int	pteidx;
2079 
2080 	/*
2081 	 * We can find the actual pte entry without searching by grabbing
2082 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2083 	 * noticing the HID bit.
2084 	 */
2085 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2086 	if (pvo->pvo_pte.pte.pte_hi & PTE_HID)
2087 		pteidx ^= moea_pteg_mask * 8;
2088 
2089 	return (pteidx);
2090 }
2091 
2092 static struct pvo_entry *
2093 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2094 {
2095 	struct	pvo_entry *pvo;
2096 	int	ptegidx;
2097 	u_int	sr;
2098 
2099 	va &= ~ADDR_POFF;
2100 	sr = va_to_sr(pm->pm_sr, va);
2101 	ptegidx = va_to_pteg(sr, va);
2102 
2103 	mtx_lock(&moea_table_mutex);
2104 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2105 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2106 			if (pteidx_p)
2107 				*pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2108 			break;
2109 		}
2110 	}
2111 	mtx_unlock(&moea_table_mutex);
2112 
2113 	return (pvo);
2114 }
2115 
2116 static struct pte *
2117 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2118 {
2119 	struct	pte *pt;
2120 
2121 	/*
2122 	 * If we haven't been supplied the ptegidx, calculate it.
2123 	 */
2124 	if (pteidx == -1) {
2125 		int	ptegidx;
2126 		u_int	sr;
2127 
2128 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2129 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2130 		pteidx = moea_pvo_pte_index(pvo, ptegidx);
2131 	}
2132 
2133 	pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2134 	mtx_lock(&moea_table_mutex);
2135 
2136 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2137 		panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2138 		    "valid pte index", pvo);
2139 	}
2140 
2141 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2142 		panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2143 		    "pvo but no valid pte", pvo);
2144 	}
2145 
2146 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2147 		if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) {
2148 			panic("moea_pvo_to_pte: pvo %p has valid pte in "
2149 			    "moea_pteg_table %p but invalid in pvo", pvo, pt);
2150 		}
2151 
2152 		if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2153 		    != 0) {
2154 			panic("moea_pvo_to_pte: pvo %p pte does not match "
2155 			    "pte %p in moea_pteg_table", pvo, pt);
2156 		}
2157 
2158 		mtx_assert(&moea_table_mutex, MA_OWNED);
2159 		return (pt);
2160 	}
2161 
2162 	if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) {
2163 		panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2164 		    "moea_pteg_table but valid in pvo: %8x, %8x", pvo, pt, pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2165 	}
2166 
2167 	mtx_unlock(&moea_table_mutex);
2168 	return (NULL);
2169 }
2170 
2171 /*
2172  * XXX: THIS STUFF SHOULD BE IN pte.c?
2173  */
2174 int
2175 moea_pte_spill(vm_offset_t addr)
2176 {
2177 	struct	pvo_entry *source_pvo, *victim_pvo;
2178 	struct	pvo_entry *pvo;
2179 	int	ptegidx, i, j;
2180 	u_int	sr;
2181 	struct	pteg *pteg;
2182 	struct	pte *pt;
2183 
2184 	moea_pte_spills++;
2185 
2186 	sr = mfsrin(addr);
2187 	ptegidx = va_to_pteg(sr, addr);
2188 
2189 	/*
2190 	 * Have to substitute some entry.  Use the primary hash for this.
2191 	 * Use low bits of timebase as random generator.
2192 	 */
2193 	pteg = &moea_pteg_table[ptegidx];
2194 	mtx_lock(&moea_table_mutex);
2195 	__asm __volatile("mftb %0" : "=r"(i));
2196 	i &= 7;
2197 	pt = &pteg->pt[i];
2198 
2199 	source_pvo = NULL;
2200 	victim_pvo = NULL;
2201 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2202 		/*
2203 		 * We need to find a pvo entry for this address.
2204 		 */
2205 		if (source_pvo == NULL &&
2206 		    moea_pte_match(&pvo->pvo_pte.pte, sr, addr,
2207 		    pvo->pvo_pte.pte.pte_hi & PTE_HID)) {
2208 			/*
2209 			 * Now found an entry to be spilled into the pteg.
2210 			 * The PTE is now valid, so we know it's active.
2211 			 */
2212 			j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2213 
2214 			if (j >= 0) {
2215 				PVO_PTEGIDX_SET(pvo, j);
2216 				moea_pte_overflow--;
2217 				mtx_unlock(&moea_table_mutex);
2218 				return (1);
2219 			}
2220 
2221 			source_pvo = pvo;
2222 
2223 			if (victim_pvo != NULL)
2224 				break;
2225 		}
2226 
2227 		/*
2228 		 * We also need the pvo entry of the victim we are replacing
2229 		 * so save the R & C bits of the PTE.
2230 		 */
2231 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2232 		    moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2233 			victim_pvo = pvo;
2234 			if (source_pvo != NULL)
2235 				break;
2236 		}
2237 	}
2238 
2239 	if (source_pvo == NULL) {
2240 		mtx_unlock(&moea_table_mutex);
2241 		return (0);
2242 	}
2243 
2244 	if (victim_pvo == NULL) {
2245 		if ((pt->pte_hi & PTE_HID) == 0)
2246 			panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2247 			    "entry", pt);
2248 
2249 		/*
2250 		 * If this is a secondary PTE, we need to search it's primary
2251 		 * pvo bucket for the matching PVO.
2252 		 */
2253 		LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2254 		    pvo_olink) {
2255 			/*
2256 			 * We also need the pvo entry of the victim we are
2257 			 * replacing so save the R & C bits of the PTE.
2258 			 */
2259 			if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2260 				victim_pvo = pvo;
2261 				break;
2262 			}
2263 		}
2264 
2265 		if (victim_pvo == NULL)
2266 			panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2267 			    "entry", pt);
2268 	}
2269 
2270 	/*
2271 	 * We are invalidating the TLB entry for the EA we are replacing even
2272 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2273 	 * contained in the TLB entry.
2274 	 */
2275 	source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID;
2276 
2277 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2278 	moea_pte_set(pt, &source_pvo->pvo_pte.pte);
2279 
2280 	PVO_PTEGIDX_CLR(victim_pvo);
2281 	PVO_PTEGIDX_SET(source_pvo, i);
2282 	moea_pte_replacements++;
2283 
2284 	mtx_unlock(&moea_table_mutex);
2285 	return (1);
2286 }
2287 
2288 static __inline struct pvo_entry *
2289 moea_pte_spillable_ident(u_int ptegidx)
2290 {
2291 	struct	pte *pt;
2292 	struct	pvo_entry *pvo_walk, *pvo = NULL;
2293 
2294 	LIST_FOREACH(pvo_walk, &moea_pvo_table[ptegidx], pvo_olink) {
2295 		if (pvo_walk->pvo_vaddr & PVO_WIRED)
2296 			continue;
2297 
2298 		if (!(pvo_walk->pvo_pte.pte.pte_hi & PTE_VALID))
2299 			continue;
2300 
2301 		pt = moea_pvo_to_pte(pvo_walk, -1);
2302 
2303 		if (pt == NULL)
2304 			continue;
2305 
2306 		pvo = pvo_walk;
2307 
2308 		mtx_unlock(&moea_table_mutex);
2309 		if (!(pt->pte_lo & PTE_REF))
2310 			return (pvo_walk);
2311 	}
2312 
2313 	return (pvo);
2314 }
2315 
2316 static int
2317 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2318 {
2319 	struct	pte *pt;
2320 	struct	pvo_entry *victim_pvo;
2321 	int	i;
2322 	int	victim_idx;
2323 	u_int	pteg_bkpidx = ptegidx;
2324 
2325 	mtx_assert(&moea_table_mutex, MA_OWNED);
2326 
2327 	/*
2328 	 * First try primary hash.
2329 	 */
2330 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2331 		if ((pt->pte_hi & PTE_VALID) == 0) {
2332 			pvo_pt->pte_hi &= ~PTE_HID;
2333 			moea_pte_set(pt, pvo_pt);
2334 			return (i);
2335 		}
2336 	}
2337 
2338 	/*
2339 	 * Now try secondary hash.
2340 	 */
2341 	ptegidx ^= moea_pteg_mask;
2342 
2343 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2344 		if ((pt->pte_hi & PTE_VALID) == 0) {
2345 			pvo_pt->pte_hi |= PTE_HID;
2346 			moea_pte_set(pt, pvo_pt);
2347 			return (i);
2348 		}
2349 	}
2350 
2351 	/* Try again, but this time try to force a PTE out. */
2352 	ptegidx = pteg_bkpidx;
2353 
2354 	victim_pvo = moea_pte_spillable_ident(ptegidx);
2355 	if (victim_pvo == NULL) {
2356 		ptegidx ^= moea_pteg_mask;
2357 		victim_pvo = moea_pte_spillable_ident(ptegidx);
2358 	}
2359 
2360 	if (victim_pvo == NULL) {
2361 		panic("moea_pte_insert: overflow");
2362 		return (-1);
2363 	}
2364 
2365 	victim_idx = moea_pvo_pte_index(victim_pvo, ptegidx);
2366 
2367 	if (pteg_bkpidx == ptegidx)
2368 		pvo_pt->pte_hi &= ~PTE_HID;
2369 	else
2370 		pvo_pt->pte_hi |= PTE_HID;
2371 
2372 	/*
2373 	 * Synchronize the sacrifice PTE with its PVO, then mark both
2374 	 * invalid. The PVO will be reused when/if the VM system comes
2375 	 * here after a fault.
2376 	 */
2377 	pt = &moea_pteg_table[victim_idx >> 3].pt[victim_idx & 7];
2378 
2379 	if (pt->pte_hi != victim_pvo->pvo_pte.pte.pte_hi)
2380 	    panic("Victim PVO doesn't match PTE! PVO: %8x, PTE: %8x", victim_pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2381 
2382 	/*
2383 	 * Set the new PTE.
2384 	 */
2385 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2386 	PVO_PTEGIDX_CLR(victim_pvo);
2387 	moea_pte_overflow++;
2388 	moea_pte_set(pt, pvo_pt);
2389 
2390 	return (victim_idx & 7);
2391 }
2392 
2393 static boolean_t
2394 moea_query_bit(vm_page_t m, int ptebit)
2395 {
2396 	struct	pvo_entry *pvo;
2397 	struct	pte *pt;
2398 
2399 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2400 	if (moea_attr_fetch(m) & ptebit)
2401 		return (TRUE);
2402 
2403 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2404 
2405 		/*
2406 		 * See if we saved the bit off.  If so, cache it and return
2407 		 * success.
2408 		 */
2409 		if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2410 			moea_attr_save(m, ptebit);
2411 			return (TRUE);
2412 		}
2413 	}
2414 
2415 	/*
2416 	 * No luck, now go through the hard part of looking at the PTEs
2417 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2418 	 * the PTEs.
2419 	 */
2420 	powerpc_sync();
2421 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2422 
2423 		/*
2424 		 * See if this pvo has a valid PTE.  if so, fetch the
2425 		 * REF/CHG bits from the valid PTE.  If the appropriate
2426 		 * ptebit is set, cache it and return success.
2427 		 */
2428 		pt = moea_pvo_to_pte(pvo, -1);
2429 		if (pt != NULL) {
2430 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2431 			mtx_unlock(&moea_table_mutex);
2432 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2433 				moea_attr_save(m, ptebit);
2434 				return (TRUE);
2435 			}
2436 		}
2437 	}
2438 
2439 	return (FALSE);
2440 }
2441 
2442 static u_int
2443 moea_clear_bit(vm_page_t m, int ptebit)
2444 {
2445 	u_int	count;
2446 	struct	pvo_entry *pvo;
2447 	struct	pte *pt;
2448 
2449 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2450 
2451 	/*
2452 	 * Clear the cached value.
2453 	 */
2454 	moea_attr_clear(m, ptebit);
2455 
2456 	/*
2457 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2458 	 * we can reset the right ones).  note that since the pvo entries and
2459 	 * list heads are accessed via BAT0 and are never placed in the page
2460 	 * table, we don't have to worry about further accesses setting the
2461 	 * REF/CHG bits.
2462 	 */
2463 	powerpc_sync();
2464 
2465 	/*
2466 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2467 	 * valid pte clear the ptebit from the valid pte.
2468 	 */
2469 	count = 0;
2470 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2471 		pt = moea_pvo_to_pte(pvo, -1);
2472 		if (pt != NULL) {
2473 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2474 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2475 				count++;
2476 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2477 			}
2478 			mtx_unlock(&moea_table_mutex);
2479 		}
2480 		pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2481 	}
2482 
2483 	return (count);
2484 }
2485 
2486 /*
2487  * Return true if the physical range is encompassed by the battable[idx]
2488  */
2489 static int
2490 moea_bat_mapped(int idx, vm_offset_t pa, vm_size_t size)
2491 {
2492 	u_int prot;
2493 	u_int32_t start;
2494 	u_int32_t end;
2495 	u_int32_t bat_ble;
2496 
2497 	/*
2498 	 * Return immediately if not a valid mapping
2499 	 */
2500 	if (!(battable[idx].batu & BAT_Vs))
2501 		return (EINVAL);
2502 
2503 	/*
2504 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2505 	 * so it can function as an i/o page
2506 	 */
2507 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2508 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2509 		return (EPERM);
2510 
2511 	/*
2512 	 * The address should be within the BAT range. Assume that the
2513 	 * start address in the BAT has the correct alignment (thus
2514 	 * not requiring masking)
2515 	 */
2516 	start = battable[idx].batl & BAT_PBS;
2517 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2518 	end = start | (bat_ble << 15) | 0x7fff;
2519 
2520 	if ((pa < start) || ((pa + size) > end))
2521 		return (ERANGE);
2522 
2523 	return (0);
2524 }
2525 
2526 boolean_t
2527 moea_dev_direct_mapped(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2528 {
2529 	int i;
2530 
2531 	/*
2532 	 * This currently does not work for entries that
2533 	 * overlap 256M BAT segments.
2534 	 */
2535 
2536 	for(i = 0; i < 16; i++)
2537 		if (moea_bat_mapped(i, pa, size) == 0)
2538 			return (0);
2539 
2540 	return (EFAULT);
2541 }
2542 
2543 /*
2544  * Map a set of physical memory pages into the kernel virtual
2545  * address space. Return a pointer to where it is mapped. This
2546  * routine is intended to be used for mapping device memory,
2547  * NOT real memory.
2548  */
2549 void *
2550 moea_mapdev(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2551 {
2552 
2553 	return (moea_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT));
2554 }
2555 
2556 void *
2557 moea_mapdev_attr(mmu_t mmu, vm_offset_t pa, vm_size_t size, vm_memattr_t ma)
2558 {
2559 	vm_offset_t va, tmpva, ppa, offset;
2560 	int i;
2561 
2562 	ppa = trunc_page(pa);
2563 	offset = pa & PAGE_MASK;
2564 	size = roundup(offset + size, PAGE_SIZE);
2565 
2566 	/*
2567 	 * If the physical address lies within a valid BAT table entry,
2568 	 * return the 1:1 mapping. This currently doesn't work
2569 	 * for regions that overlap 256M BAT segments.
2570 	 */
2571 	for (i = 0; i < 16; i++) {
2572 		if (moea_bat_mapped(i, pa, size) == 0)
2573 			return ((void *) pa);
2574 	}
2575 
2576 	va = kva_alloc(size);
2577 	if (!va)
2578 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2579 
2580 	for (tmpva = va; size > 0;) {
2581 		moea_kenter_attr(mmu, tmpva, ppa, ma);
2582 		tlbie(tmpva);
2583 		size -= PAGE_SIZE;
2584 		tmpva += PAGE_SIZE;
2585 		ppa += PAGE_SIZE;
2586 	}
2587 
2588 	return ((void *)(va + offset));
2589 }
2590 
2591 void
2592 moea_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2593 {
2594 	vm_offset_t base, offset;
2595 
2596 	/*
2597 	 * If this is outside kernel virtual space, then it's a
2598 	 * battable entry and doesn't require unmapping
2599 	 */
2600 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2601 		base = trunc_page(va);
2602 		offset = va & PAGE_MASK;
2603 		size = roundup(offset + size, PAGE_SIZE);
2604 		kva_free(base, size);
2605 	}
2606 }
2607 
2608 static void
2609 moea_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2610 {
2611 	struct pvo_entry *pvo;
2612 	vm_offset_t lim;
2613 	vm_paddr_t pa;
2614 	vm_size_t len;
2615 
2616 	PMAP_LOCK(pm);
2617 	while (sz > 0) {
2618 		lim = round_page(va);
2619 		len = MIN(lim - va, sz);
2620 		pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2621 		if (pvo != NULL) {
2622 			pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) |
2623 			    (va & ADDR_POFF);
2624 			moea_syncicache(pa, len);
2625 		}
2626 		va += len;
2627 		sz -= len;
2628 	}
2629 	PMAP_UNLOCK(pm);
2630 }
2631 
2632 vm_offset_t
2633 moea_dumpsys_map(mmu_t mmu, struct pmap_md *md, vm_size_t ofs,
2634     vm_size_t *sz)
2635 {
2636 	if (md->md_vaddr == ~0UL)
2637 	    return (md->md_paddr + ofs);
2638 	else
2639 	    return (md->md_vaddr + ofs);
2640 }
2641 
2642 struct pmap_md *
2643 moea_scan_md(mmu_t mmu, struct pmap_md *prev)
2644 {
2645 	static struct pmap_md md;
2646 	struct pvo_entry *pvo;
2647 	vm_offset_t va;
2648 
2649 	if (dumpsys_minidump) {
2650 		md.md_paddr = ~0UL;	/* Minidumps use virtual addresses. */
2651 		if (prev == NULL) {
2652 			/* 1st: kernel .data and .bss. */
2653 			md.md_index = 1;
2654 			md.md_vaddr = trunc_page((uintptr_t)_etext);
2655 			md.md_size = round_page((uintptr_t)_end) - md.md_vaddr;
2656 			return (&md);
2657 		}
2658 		switch (prev->md_index) {
2659 		case 1:
2660 			/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2661 			md.md_index = 2;
2662 			md.md_vaddr = (vm_offset_t)msgbufp->msg_ptr;
2663 			md.md_size = round_page(msgbufp->msg_size);
2664 			break;
2665 		case 2:
2666 			/* 3rd: kernel VM. */
2667 			va = prev->md_vaddr + prev->md_size;
2668 			/* Find start of next chunk (from va). */
2669 			while (va < virtual_end) {
2670 				/* Don't dump the buffer cache. */
2671 				if (va >= kmi.buffer_sva &&
2672 				    va < kmi.buffer_eva) {
2673 					va = kmi.buffer_eva;
2674 					continue;
2675 				}
2676 				pvo = moea_pvo_find_va(kernel_pmap,
2677 				    va & ~ADDR_POFF, NULL);
2678 				if (pvo != NULL &&
2679 				    (pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2680 					break;
2681 				va += PAGE_SIZE;
2682 			}
2683 			if (va < virtual_end) {
2684 				md.md_vaddr = va;
2685 				va += PAGE_SIZE;
2686 				/* Find last page in chunk. */
2687 				while (va < virtual_end) {
2688 					/* Don't run into the buffer cache. */
2689 					if (va == kmi.buffer_sva)
2690 						break;
2691 					pvo = moea_pvo_find_va(kernel_pmap,
2692 					    va & ~ADDR_POFF, NULL);
2693 					if (pvo == NULL ||
2694 					    !(pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2695 						break;
2696 					va += PAGE_SIZE;
2697 				}
2698 				md.md_size = va - md.md_vaddr;
2699 				break;
2700 			}
2701 			md.md_index = 3;
2702 			/* FALLTHROUGH */
2703 		default:
2704 			return (NULL);
2705 		}
2706 	} else { /* minidumps */
2707 		mem_regions(&pregions, &pregions_sz,
2708 		    &regions, &regions_sz);
2709 
2710 		if (prev == NULL) {
2711 			/* first physical chunk. */
2712 			md.md_paddr = pregions[0].mr_start;
2713 			md.md_size = pregions[0].mr_size;
2714 			md.md_vaddr = ~0UL;
2715 			md.md_index = 1;
2716 		} else if (md.md_index < pregions_sz) {
2717 			md.md_paddr = pregions[md.md_index].mr_start;
2718 			md.md_size = pregions[md.md_index].mr_size;
2719 			md.md_vaddr = ~0UL;
2720 			md.md_index++;
2721 		} else {
2722 			/* There's no next physical chunk. */
2723 			return (NULL);
2724 		}
2725 	}
2726 
2727 	return (&md);
2728 }
2729