xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision 8fd7d6577916b57274307607d91b8820c213ca19)
1 /*-
2  * Copyright (c) 2001 The NetBSD Foundation, Inc.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to The NetBSD Foundation
6  * by Matt Thomas <matt@3am-software.com> of Allegro Networks, Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *        This product includes software developed by the NetBSD
19  *        Foundation, Inc. and its contributors.
20  * 4. Neither the name of The NetBSD Foundation nor the names of its
21  *    contributors may be used to endorse or promote products derived
22  *    from this software without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
25  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
26  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
27  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
28  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
29  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
30  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
31  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
32  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
33  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
34  * POSSIBILITY OF SUCH DAMAGE.
35  */
36 /*-
37  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
38  * Copyright (C) 1995, 1996 TooLs GmbH.
39  * All rights reserved.
40  *
41  * Redistribution and use in source and binary forms, with or without
42  * modification, are permitted provided that the following conditions
43  * are met:
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. All advertising materials mentioning features or use of this software
50  *    must display the following acknowledgement:
51  *	This product includes software developed by TooLs GmbH.
52  * 4. The name of TooLs GmbH may not be used to endorse or promote products
53  *    derived from this software without specific prior written permission.
54  *
55  * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
56  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
57  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
58  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
59  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
60  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
61  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
62  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
63  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
64  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
65  *
66  * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $
67  */
68 /*-
69  * Copyright (C) 2001 Benno Rice.
70  * All rights reserved.
71  *
72  * Redistribution and use in source and binary forms, with or without
73  * modification, are permitted provided that the following conditions
74  * are met:
75  * 1. Redistributions of source code must retain the above copyright
76  *    notice, this list of conditions and the following disclaimer.
77  * 2. Redistributions in binary form must reproduce the above copyright
78  *    notice, this list of conditions and the following disclaimer in the
79  *    documentation and/or other materials provided with the distribution.
80  *
81  * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR
82  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
83  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
84  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
85  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
86  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
87  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
88  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
89  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
90  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
91  */
92 
93 #include <sys/cdefs.h>
94 __FBSDID("$FreeBSD$");
95 
96 /*
97  * Manages physical address maps.
98  *
99  * In addition to hardware address maps, this module is called upon to
100  * provide software-use-only maps which may or may not be stored in the
101  * same form as hardware maps.  These pseudo-maps are used to store
102  * intermediate results from copy operations to and from address spaces.
103  *
104  * Since the information managed by this module is also stored by the
105  * logical address mapping module, this module may throw away valid virtual
106  * to physical mappings at almost any time.  However, invalidations of
107  * mappings must be done as requested.
108  *
109  * In order to cope with hardware architectures which make virtual to
110  * physical map invalidates expensive, this module may delay invalidate
111  * reduced protection operations until such time as they are actually
112  * necessary.  This module is given full information as to which processors
113  * are currently using which maps, and to when physical maps must be made
114  * correct.
115  */
116 
117 #include "opt_kstack_pages.h"
118 
119 #include <sys/param.h>
120 #include <sys/kernel.h>
121 #include <sys/ktr.h>
122 #include <sys/lock.h>
123 #include <sys/msgbuf.h>
124 #include <sys/mutex.h>
125 #include <sys/proc.h>
126 #include <sys/sysctl.h>
127 #include <sys/systm.h>
128 #include <sys/vmmeter.h>
129 
130 #include <dev/ofw/openfirm.h>
131 
132 #include <vm/vm.h>
133 #include <vm/vm_param.h>
134 #include <vm/vm_kern.h>
135 #include <vm/vm_page.h>
136 #include <vm/vm_map.h>
137 #include <vm/vm_object.h>
138 #include <vm/vm_extern.h>
139 #include <vm/vm_pageout.h>
140 #include <vm/vm_pager.h>
141 #include <vm/uma.h>
142 
143 #include <machine/cpu.h>
144 #include <machine/platform.h>
145 #include <machine/bat.h>
146 #include <machine/frame.h>
147 #include <machine/md_var.h>
148 #include <machine/psl.h>
149 #include <machine/pte.h>
150 #include <machine/smp.h>
151 #include <machine/sr.h>
152 #include <machine/mmuvar.h>
153 
154 #include "mmu_if.h"
155 
156 #define	MOEA_DEBUG
157 
158 #define TODO	panic("%s: not implemented", __func__);
159 
160 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
161 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
162 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
163 
164 struct ofw_map {
165 	vm_offset_t	om_va;
166 	vm_size_t	om_len;
167 	vm_offset_t	om_pa;
168 	u_int		om_mode;
169 };
170 
171 /*
172  * Map of physical memory regions.
173  */
174 static struct	mem_region *regions;
175 static struct	mem_region *pregions;
176 static u_int    phys_avail_count;
177 static int	regions_sz, pregions_sz;
178 static struct	ofw_map *translations;
179 
180 /*
181  * Lock for the pteg and pvo tables.
182  */
183 struct mtx	moea_table_mutex;
184 struct mtx	moea_vsid_mutex;
185 
186 /* tlbie instruction synchronization */
187 static struct mtx tlbie_mtx;
188 
189 /*
190  * PTEG data.
191  */
192 static struct	pteg *moea_pteg_table;
193 u_int		moea_pteg_count;
194 u_int		moea_pteg_mask;
195 
196 /*
197  * PVO data.
198  */
199 struct	pvo_head *moea_pvo_table;		/* pvo entries by pteg index */
200 struct	pvo_head moea_pvo_kunmanaged =
201     LIST_HEAD_INITIALIZER(moea_pvo_kunmanaged);	/* list of unmanaged pages */
202 
203 uma_zone_t	moea_upvo_zone;	/* zone for pvo entries for unmanaged pages */
204 uma_zone_t	moea_mpvo_zone;	/* zone for pvo entries for managed pages */
205 
206 #define	BPVO_POOL_SIZE	32768
207 static struct	pvo_entry *moea_bpvo_pool;
208 static int	moea_bpvo_pool_index = 0;
209 
210 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
211 static u_int	moea_vsid_bitmap[NPMAPS / VSID_NBPW];
212 
213 static boolean_t moea_initialized = FALSE;
214 
215 /*
216  * Statistics.
217  */
218 u_int	moea_pte_valid = 0;
219 u_int	moea_pte_overflow = 0;
220 u_int	moea_pte_replacements = 0;
221 u_int	moea_pvo_entries = 0;
222 u_int	moea_pvo_enter_calls = 0;
223 u_int	moea_pvo_remove_calls = 0;
224 u_int	moea_pte_spills = 0;
225 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_valid, CTLFLAG_RD, &moea_pte_valid,
226     0, "");
227 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_overflow, CTLFLAG_RD,
228     &moea_pte_overflow, 0, "");
229 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_replacements, CTLFLAG_RD,
230     &moea_pte_replacements, 0, "");
231 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_entries, CTLFLAG_RD, &moea_pvo_entries,
232     0, "");
233 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_enter_calls, CTLFLAG_RD,
234     &moea_pvo_enter_calls, 0, "");
235 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_remove_calls, CTLFLAG_RD,
236     &moea_pvo_remove_calls, 0, "");
237 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_spills, CTLFLAG_RD,
238     &moea_pte_spills, 0, "");
239 
240 /*
241  * Allocate physical memory for use in moea_bootstrap.
242  */
243 static vm_offset_t	moea_bootstrap_alloc(vm_size_t, u_int);
244 
245 /*
246  * PTE calls.
247  */
248 static int		moea_pte_insert(u_int, struct pte *);
249 
250 /*
251  * PVO calls.
252  */
253 static int	moea_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *,
254 		    vm_offset_t, vm_offset_t, u_int, int);
255 static void	moea_pvo_remove(struct pvo_entry *, int);
256 static struct	pvo_entry *moea_pvo_find_va(pmap_t, vm_offset_t, int *);
257 static struct	pte *moea_pvo_to_pte(const struct pvo_entry *, int);
258 
259 /*
260  * Utility routines.
261  */
262 static void		moea_enter_locked(pmap_t, vm_offset_t, vm_page_t,
263 			    vm_prot_t, boolean_t);
264 static void		moea_syncicache(vm_offset_t, vm_size_t);
265 static boolean_t	moea_query_bit(vm_page_t, int);
266 static u_int		moea_clear_bit(vm_page_t, int);
267 static void		moea_kremove(mmu_t, vm_offset_t);
268 int		moea_pte_spill(vm_offset_t);
269 
270 /*
271  * Kernel MMU interface
272  */
273 void moea_change_wiring(mmu_t, pmap_t, vm_offset_t, boolean_t);
274 void moea_clear_modify(mmu_t, vm_page_t);
275 void moea_clear_reference(mmu_t, vm_page_t);
276 void moea_copy_page(mmu_t, vm_page_t, vm_page_t);
277 void moea_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, boolean_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 boolean_t moea_ts_referenced(mmu_t, vm_page_t);
288 vm_offset_t moea_map(mmu_t, vm_offset_t *, vm_offset_t, vm_offset_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_zero_page(mmu_t, vm_page_t);
301 void moea_zero_page_area(mmu_t, vm_page_t, int, int);
302 void moea_zero_page_idle(mmu_t, vm_page_t);
303 void moea_activate(mmu_t, struct thread *);
304 void moea_deactivate(mmu_t, struct thread *);
305 void moea_cpu_bootstrap(mmu_t, int);
306 void moea_bootstrap(mmu_t, vm_offset_t, vm_offset_t);
307 void *moea_mapdev(mmu_t, vm_offset_t, vm_size_t);
308 void *moea_mapdev_attr(mmu_t, vm_offset_t, vm_size_t, vm_memattr_t);
309 void moea_unmapdev(mmu_t, vm_offset_t, vm_size_t);
310 vm_offset_t moea_kextract(mmu_t, vm_offset_t);
311 void moea_kenter_attr(mmu_t, vm_offset_t, vm_offset_t, vm_memattr_t);
312 void moea_kenter(mmu_t, vm_offset_t, vm_offset_t);
313 void moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma);
314 boolean_t moea_dev_direct_mapped(mmu_t, vm_offset_t, vm_size_t);
315 static void moea_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
316 
317 static mmu_method_t moea_methods[] = {
318 	MMUMETHOD(mmu_change_wiring,	moea_change_wiring),
319 	MMUMETHOD(mmu_clear_modify,	moea_clear_modify),
320 	MMUMETHOD(mmu_clear_reference,	moea_clear_reference),
321 	MMUMETHOD(mmu_copy_page,	moea_copy_page),
322 	MMUMETHOD(mmu_enter,		moea_enter),
323 	MMUMETHOD(mmu_enter_object,	moea_enter_object),
324 	MMUMETHOD(mmu_enter_quick,	moea_enter_quick),
325 	MMUMETHOD(mmu_extract,		moea_extract),
326 	MMUMETHOD(mmu_extract_and_hold,	moea_extract_and_hold),
327 	MMUMETHOD(mmu_init,		moea_init),
328 	MMUMETHOD(mmu_is_modified,	moea_is_modified),
329 	MMUMETHOD(mmu_is_prefaultable,	moea_is_prefaultable),
330 	MMUMETHOD(mmu_is_referenced,	moea_is_referenced),
331 	MMUMETHOD(mmu_ts_referenced,	moea_ts_referenced),
332 	MMUMETHOD(mmu_map,     		moea_map),
333 	MMUMETHOD(mmu_page_exists_quick,moea_page_exists_quick),
334 	MMUMETHOD(mmu_page_wired_mappings,moea_page_wired_mappings),
335 	MMUMETHOD(mmu_pinit,		moea_pinit),
336 	MMUMETHOD(mmu_pinit0,		moea_pinit0),
337 	MMUMETHOD(mmu_protect,		moea_protect),
338 	MMUMETHOD(mmu_qenter,		moea_qenter),
339 	MMUMETHOD(mmu_qremove,		moea_qremove),
340 	MMUMETHOD(mmu_release,		moea_release),
341 	MMUMETHOD(mmu_remove,		moea_remove),
342 	MMUMETHOD(mmu_remove_all,      	moea_remove_all),
343 	MMUMETHOD(mmu_remove_write,	moea_remove_write),
344 	MMUMETHOD(mmu_sync_icache,	moea_sync_icache),
345 	MMUMETHOD(mmu_zero_page,       	moea_zero_page),
346 	MMUMETHOD(mmu_zero_page_area,	moea_zero_page_area),
347 	MMUMETHOD(mmu_zero_page_idle,	moea_zero_page_idle),
348 	MMUMETHOD(mmu_activate,		moea_activate),
349 	MMUMETHOD(mmu_deactivate,      	moea_deactivate),
350 	MMUMETHOD(mmu_page_set_memattr,	moea_page_set_memattr),
351 
352 	/* Internal interfaces */
353 	MMUMETHOD(mmu_bootstrap,       	moea_bootstrap),
354 	MMUMETHOD(mmu_cpu_bootstrap,   	moea_cpu_bootstrap),
355 	MMUMETHOD(mmu_mapdev_attr,	moea_mapdev_attr),
356 	MMUMETHOD(mmu_mapdev,		moea_mapdev),
357 	MMUMETHOD(mmu_unmapdev,		moea_unmapdev),
358 	MMUMETHOD(mmu_kextract,		moea_kextract),
359 	MMUMETHOD(mmu_kenter,		moea_kenter),
360 	MMUMETHOD(mmu_kenter_attr,	moea_kenter_attr),
361 	MMUMETHOD(mmu_dev_direct_mapped,moea_dev_direct_mapped),
362 
363 	{ 0, 0 }
364 };
365 
366 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods, 0);
367 
368 static __inline uint32_t
369 moea_calc_wimg(vm_offset_t pa, vm_memattr_t ma)
370 {
371 	uint32_t pte_lo;
372 	int i;
373 
374 	if (ma != VM_MEMATTR_DEFAULT) {
375 		switch (ma) {
376 		case VM_MEMATTR_UNCACHEABLE:
377 			return (PTE_I | PTE_G);
378 		case VM_MEMATTR_WRITE_COMBINING:
379 		case VM_MEMATTR_WRITE_BACK:
380 		case VM_MEMATTR_PREFETCHABLE:
381 			return (PTE_I);
382 		case VM_MEMATTR_WRITE_THROUGH:
383 			return (PTE_W | PTE_M);
384 		}
385 	}
386 
387 	/*
388 	 * Assume the page is cache inhibited and access is guarded unless
389 	 * it's in our available memory array.
390 	 */
391 	pte_lo = PTE_I | PTE_G;
392 	for (i = 0; i < pregions_sz; i++) {
393 		if ((pa >= pregions[i].mr_start) &&
394 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
395 			pte_lo = PTE_M;
396 			break;
397 		}
398 	}
399 
400 	return pte_lo;
401 }
402 
403 static void
404 tlbie(vm_offset_t va)
405 {
406 
407 	mtx_lock_spin(&tlbie_mtx);
408 	__asm __volatile("ptesync");
409 	__asm __volatile("tlbie %0" :: "r"(va));
410 	__asm __volatile("eieio; tlbsync; ptesync");
411 	mtx_unlock_spin(&tlbie_mtx);
412 }
413 
414 static void
415 tlbia(void)
416 {
417 	vm_offset_t va;
418 
419 	for (va = 0; va < 0x00040000; va += 0x00001000) {
420 		__asm __volatile("tlbie %0" :: "r"(va));
421 		powerpc_sync();
422 	}
423 	__asm __volatile("tlbsync");
424 	powerpc_sync();
425 }
426 
427 static __inline int
428 va_to_sr(u_int *sr, vm_offset_t va)
429 {
430 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
431 }
432 
433 static __inline u_int
434 va_to_pteg(u_int sr, vm_offset_t addr)
435 {
436 	u_int hash;
437 
438 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
439 	    ADDR_PIDX_SHFT);
440 	return (hash & moea_pteg_mask);
441 }
442 
443 static __inline struct pvo_head *
444 vm_page_to_pvoh(vm_page_t m)
445 {
446 
447 	return (&m->md.mdpg_pvoh);
448 }
449 
450 static __inline void
451 moea_attr_clear(vm_page_t m, int ptebit)
452 {
453 
454 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
455 	m->md.mdpg_attrs &= ~ptebit;
456 }
457 
458 static __inline int
459 moea_attr_fetch(vm_page_t m)
460 {
461 
462 	return (m->md.mdpg_attrs);
463 }
464 
465 static __inline void
466 moea_attr_save(vm_page_t m, int ptebit)
467 {
468 
469 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
470 	m->md.mdpg_attrs |= ptebit;
471 }
472 
473 static __inline int
474 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
475 {
476 	if (pt->pte_hi == pvo_pt->pte_hi)
477 		return (1);
478 
479 	return (0);
480 }
481 
482 static __inline int
483 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
484 {
485 	return (pt->pte_hi & ~PTE_VALID) ==
486 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
487 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
488 }
489 
490 static __inline void
491 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
492 {
493 
494 	mtx_assert(&moea_table_mutex, MA_OWNED);
495 
496 	/*
497 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
498 	 * set when the real pte is set in memory.
499 	 *
500 	 * Note: Don't set the valid bit for correct operation of tlb update.
501 	 */
502 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
503 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
504 	pt->pte_lo = pte_lo;
505 }
506 
507 static __inline void
508 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
509 {
510 
511 	mtx_assert(&moea_table_mutex, MA_OWNED);
512 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
513 }
514 
515 static __inline void
516 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
517 {
518 
519 	mtx_assert(&moea_table_mutex, MA_OWNED);
520 
521 	/*
522 	 * As shown in Section 7.6.3.2.3
523 	 */
524 	pt->pte_lo &= ~ptebit;
525 	tlbie(va);
526 }
527 
528 static __inline void
529 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
530 {
531 
532 	mtx_assert(&moea_table_mutex, MA_OWNED);
533 	pvo_pt->pte_hi |= PTE_VALID;
534 
535 	/*
536 	 * Update the PTE as defined in section 7.6.3.1.
537 	 * Note that the REF/CHG bits are from pvo_pt and thus should havce
538 	 * been saved so this routine can restore them (if desired).
539 	 */
540 	pt->pte_lo = pvo_pt->pte_lo;
541 	powerpc_sync();
542 	pt->pte_hi = pvo_pt->pte_hi;
543 	powerpc_sync();
544 	moea_pte_valid++;
545 }
546 
547 static __inline void
548 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
549 {
550 
551 	mtx_assert(&moea_table_mutex, MA_OWNED);
552 	pvo_pt->pte_hi &= ~PTE_VALID;
553 
554 	/*
555 	 * Force the reg & chg bits back into the PTEs.
556 	 */
557 	powerpc_sync();
558 
559 	/*
560 	 * Invalidate the pte.
561 	 */
562 	pt->pte_hi &= ~PTE_VALID;
563 
564 	tlbie(va);
565 
566 	/*
567 	 * Save the reg & chg bits.
568 	 */
569 	moea_pte_synch(pt, pvo_pt);
570 	moea_pte_valid--;
571 }
572 
573 static __inline void
574 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
575 {
576 
577 	/*
578 	 * Invalidate the PTE
579 	 */
580 	moea_pte_unset(pt, pvo_pt, va);
581 	moea_pte_set(pt, pvo_pt);
582 }
583 
584 /*
585  * Quick sort callout for comparing memory regions.
586  */
587 static int	mr_cmp(const void *a, const void *b);
588 static int	om_cmp(const void *a, const void *b);
589 
590 static int
591 mr_cmp(const void *a, const void *b)
592 {
593 	const struct	mem_region *regiona;
594 	const struct	mem_region *regionb;
595 
596 	regiona = a;
597 	regionb = b;
598 	if (regiona->mr_start < regionb->mr_start)
599 		return (-1);
600 	else if (regiona->mr_start > regionb->mr_start)
601 		return (1);
602 	else
603 		return (0);
604 }
605 
606 static int
607 om_cmp(const void *a, const void *b)
608 {
609 	const struct	ofw_map *mapa;
610 	const struct	ofw_map *mapb;
611 
612 	mapa = a;
613 	mapb = b;
614 	if (mapa->om_pa < mapb->om_pa)
615 		return (-1);
616 	else if (mapa->om_pa > mapb->om_pa)
617 		return (1);
618 	else
619 		return (0);
620 }
621 
622 void
623 moea_cpu_bootstrap(mmu_t mmup, int ap)
624 {
625 	u_int sdr;
626 	int i;
627 
628 	if (ap) {
629 		powerpc_sync();
630 		__asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu));
631 		__asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl));
632 		isync();
633 		__asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu));
634 		__asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl));
635 		isync();
636 	}
637 
638 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
639 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
640 	isync();
641 
642 	__asm __volatile("mtibatu 1,%0" :: "r"(0));
643 	__asm __volatile("mtdbatu 2,%0" :: "r"(0));
644 	__asm __volatile("mtibatu 2,%0" :: "r"(0));
645 	__asm __volatile("mtdbatu 3,%0" :: "r"(0));
646 	__asm __volatile("mtibatu 3,%0" :: "r"(0));
647 	isync();
648 
649 	for (i = 0; i < 16; i++)
650 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
651 	powerpc_sync();
652 
653 	sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10);
654 	__asm __volatile("mtsdr1 %0" :: "r"(sdr));
655 	isync();
656 
657 	tlbia();
658 }
659 
660 void
661 moea_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
662 {
663 	ihandle_t	mmui;
664 	phandle_t	chosen, mmu;
665 	int		sz;
666 	int		i, j;
667 	vm_size_t	size, physsz, hwphyssz;
668 	vm_offset_t	pa, va, off;
669 	void		*dpcpu;
670 	register_t	msr;
671 
672         /*
673          * Set up BAT0 to map the lowest 256 MB area
674          */
675         battable[0x0].batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
676         battable[0x0].batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
677 
678         /*
679          * Map PCI memory space.
680          */
681         battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
682         battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
683 
684         battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
685         battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
686 
687         battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
688         battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
689 
690         battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
691         battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
692 
693         /*
694          * Map obio devices.
695          */
696         battable[0xf].batl = BATL(0xf0000000, BAT_I|BAT_G, BAT_PP_RW);
697         battable[0xf].batu = BATU(0xf0000000, BAT_BL_256M, BAT_Vs);
698 
699 	/*
700 	 * Use an IBAT and a DBAT to map the bottom segment of memory
701 	 * where we are. Turn off instruction relocation temporarily
702 	 * to prevent faults while reprogramming the IBAT.
703 	 */
704 	msr = mfmsr();
705 	mtmsr(msr & ~PSL_IR);
706 	__asm (".balign 32; \n"
707 	       "mtibatu 0,%0; mtibatl 0,%1; isync; \n"
708 	       "mtdbatu 0,%0; mtdbatl 0,%1; isync"
709 	    :: "r"(battable[0].batu), "r"(battable[0].batl));
710 	mtmsr(msr);
711 
712 	/* map pci space */
713 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
714 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
715 	isync();
716 
717 	/* set global direct map flag */
718 	hw_direct_map = 1;
719 
720 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
721 	CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
722 
723 	qsort(pregions, pregions_sz, sizeof(*pregions), mr_cmp);
724 	for (i = 0; i < pregions_sz; i++) {
725 		vm_offset_t pa;
726 		vm_offset_t end;
727 
728 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
729 			pregions[i].mr_start,
730 			pregions[i].mr_start + pregions[i].mr_size,
731 			pregions[i].mr_size);
732 		/*
733 		 * Install entries into the BAT table to allow all
734 		 * of physmem to be convered by on-demand BAT entries.
735 		 * The loop will sometimes set the same battable element
736 		 * twice, but that's fine since they won't be used for
737 		 * a while yet.
738 		 */
739 		pa = pregions[i].mr_start & 0xf0000000;
740 		end = pregions[i].mr_start + pregions[i].mr_size;
741 		do {
742                         u_int n = pa >> ADDR_SR_SHFT;
743 
744 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
745 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
746 			pa += SEGMENT_LENGTH;
747 		} while (pa < end);
748 	}
749 
750 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
751 		panic("moea_bootstrap: phys_avail too small");
752 	qsort(regions, regions_sz, sizeof(*regions), mr_cmp);
753 	phys_avail_count = 0;
754 	physsz = 0;
755 	hwphyssz = 0;
756 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
757 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
758 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
759 		    regions[i].mr_start + regions[i].mr_size,
760 		    regions[i].mr_size);
761 		if (hwphyssz != 0 &&
762 		    (physsz + regions[i].mr_size) >= hwphyssz) {
763 			if (physsz < hwphyssz) {
764 				phys_avail[j] = regions[i].mr_start;
765 				phys_avail[j + 1] = regions[i].mr_start +
766 				    hwphyssz - physsz;
767 				physsz = hwphyssz;
768 				phys_avail_count++;
769 			}
770 			break;
771 		}
772 		phys_avail[j] = regions[i].mr_start;
773 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
774 		phys_avail_count++;
775 		physsz += regions[i].mr_size;
776 	}
777 	physmem = btoc(physsz);
778 
779 	/*
780 	 * Allocate PTEG table.
781 	 */
782 #ifdef PTEGCOUNT
783 	moea_pteg_count = PTEGCOUNT;
784 #else
785 	moea_pteg_count = 0x1000;
786 
787 	while (moea_pteg_count < physmem)
788 		moea_pteg_count <<= 1;
789 
790 	moea_pteg_count >>= 1;
791 #endif /* PTEGCOUNT */
792 
793 	size = moea_pteg_count * sizeof(struct pteg);
794 	CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
795 	    size);
796 	moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
797 	CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
798 	bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
799 	moea_pteg_mask = moea_pteg_count - 1;
800 
801 	/*
802 	 * Allocate pv/overflow lists.
803 	 */
804 	size = sizeof(struct pvo_head) * moea_pteg_count;
805 	moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
806 	    PAGE_SIZE);
807 	CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
808 	for (i = 0; i < moea_pteg_count; i++)
809 		LIST_INIT(&moea_pvo_table[i]);
810 
811 	/*
812 	 * Initialize the lock that synchronizes access to the pteg and pvo
813 	 * tables.
814 	 */
815 	mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF |
816 	    MTX_RECURSE);
817 	mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF);
818 
819 	mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN);
820 
821 	/*
822 	 * Initialise the unmanaged pvo pool.
823 	 */
824 	moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
825 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
826 	moea_bpvo_pool_index = 0;
827 
828 	/*
829 	 * Make sure kernel vsid is allocated as well as VSID 0.
830 	 */
831 	moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
832 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
833 	moea_vsid_bitmap[0] |= 1;
834 
835 	/*
836 	 * Initialize the kernel pmap (which is statically allocated).
837 	 */
838 	PMAP_LOCK_INIT(kernel_pmap);
839 	for (i = 0; i < 16; i++)
840 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
841 	kernel_pmap->pm_active = ~0;
842 
843 	/*
844 	 * Set up the Open Firmware mappings
845 	 */
846 	if ((chosen = OF_finddevice("/chosen")) == -1)
847 		panic("moea_bootstrap: can't find /chosen");
848 	OF_getprop(chosen, "mmu", &mmui, 4);
849 	if ((mmu = OF_instance_to_package(mmui)) == -1)
850 		panic("moea_bootstrap: can't get mmu package");
851 	if ((sz = OF_getproplen(mmu, "translations")) == -1)
852 		panic("moea_bootstrap: can't get ofw translation count");
853 	translations = NULL;
854 	for (i = 0; phys_avail[i] != 0; i += 2) {
855 		if (phys_avail[i + 1] >= sz) {
856 			translations = (struct ofw_map *)phys_avail[i];
857 			break;
858 		}
859 	}
860 	if (translations == NULL)
861 		panic("moea_bootstrap: no space to copy translations");
862 	bzero(translations, sz);
863 	if (OF_getprop(mmu, "translations", translations, sz) == -1)
864 		panic("moea_bootstrap: can't get ofw translations");
865 	CTR0(KTR_PMAP, "moea_bootstrap: translations");
866 	sz /= sizeof(*translations);
867 	qsort(translations, sz, sizeof (*translations), om_cmp);
868 	for (i = 0; i < sz; i++) {
869 		CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
870 		    translations[i].om_pa, translations[i].om_va,
871 		    translations[i].om_len);
872 
873 		/*
874 		 * If the mapping is 1:1, let the RAM and device on-demand
875 		 * BAT tables take care of the translation.
876 		 */
877 		if (translations[i].om_va == translations[i].om_pa)
878 			continue;
879 
880 		/* Enter the pages */
881 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE)
882 			moea_kenter(mmup, translations[i].om_va + off,
883 				    translations[i].om_pa + off);
884 	}
885 
886 	/*
887 	 * Calculate the last available physical address.
888 	 */
889 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
890 		;
891 	Maxmem = powerpc_btop(phys_avail[i + 1]);
892 
893 	moea_cpu_bootstrap(mmup,0);
894 
895 	pmap_bootstrapped++;
896 
897 	/*
898 	 * Set the start and end of kva.
899 	 */
900 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
901 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
902 
903 	/*
904 	 * Allocate a kernel stack with a guard page for thread0 and map it
905 	 * into the kernel page map.
906 	 */
907 	pa = moea_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, PAGE_SIZE);
908 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
909 	virtual_avail = va + KSTACK_PAGES * PAGE_SIZE;
910 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
911 	thread0.td_kstack = va;
912 	thread0.td_kstack_pages = KSTACK_PAGES;
913 	for (i = 0; i < KSTACK_PAGES; i++) {
914 		moea_kenter(mmup, va, pa);
915 		pa += PAGE_SIZE;
916 		va += PAGE_SIZE;
917 	}
918 
919 	/*
920 	 * Allocate virtual address space for the message buffer.
921 	 */
922 	pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE);
923 	msgbufp = (struct msgbuf *)virtual_avail;
924 	va = virtual_avail;
925 	virtual_avail += round_page(msgbufsize);
926 	while (va < virtual_avail) {
927 		moea_kenter(mmup, va, pa);
928 		pa += PAGE_SIZE;
929 		va += PAGE_SIZE;
930 	}
931 
932 	/*
933 	 * Allocate virtual address space for the dynamic percpu area.
934 	 */
935 	pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
936 	dpcpu = (void *)virtual_avail;
937 	va = virtual_avail;
938 	virtual_avail += DPCPU_SIZE;
939 	while (va < virtual_avail) {
940 		moea_kenter(mmup, va, pa);
941 		pa += PAGE_SIZE;
942 		va += PAGE_SIZE;
943 	}
944 	dpcpu_init(dpcpu, 0);
945 }
946 
947 /*
948  * Activate a user pmap.  The pmap must be activated before it's address
949  * space can be accessed in any way.
950  */
951 void
952 moea_activate(mmu_t mmu, struct thread *td)
953 {
954 	pmap_t	pm, pmr;
955 
956 	/*
957 	 * Load all the data we need up front to encourage the compiler to
958 	 * not issue any loads while we have interrupts disabled below.
959 	 */
960 	pm = &td->td_proc->p_vmspace->vm_pmap;
961 	pmr = pm->pmap_phys;
962 
963 	pm->pm_active |= PCPU_GET(cpumask);
964 	PCPU_SET(curpmap, pmr);
965 }
966 
967 void
968 moea_deactivate(mmu_t mmu, struct thread *td)
969 {
970 	pmap_t	pm;
971 
972 	pm = &td->td_proc->p_vmspace->vm_pmap;
973 	pm->pm_active &= ~PCPU_GET(cpumask);
974 	PCPU_SET(curpmap, NULL);
975 }
976 
977 void
978 moea_change_wiring(mmu_t mmu, pmap_t pm, vm_offset_t va, boolean_t wired)
979 {
980 	struct	pvo_entry *pvo;
981 
982 	PMAP_LOCK(pm);
983 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
984 
985 	if (pvo != NULL) {
986 		if (wired) {
987 			if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
988 				pm->pm_stats.wired_count++;
989 			pvo->pvo_vaddr |= PVO_WIRED;
990 		} else {
991 			if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
992 				pm->pm_stats.wired_count--;
993 			pvo->pvo_vaddr &= ~PVO_WIRED;
994 		}
995 	}
996 	PMAP_UNLOCK(pm);
997 }
998 
999 void
1000 moea_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1001 {
1002 	vm_offset_t	dst;
1003 	vm_offset_t	src;
1004 
1005 	dst = VM_PAGE_TO_PHYS(mdst);
1006 	src = VM_PAGE_TO_PHYS(msrc);
1007 
1008 	kcopy((void *)src, (void *)dst, PAGE_SIZE);
1009 }
1010 
1011 /*
1012  * Zero a page of physical memory by temporarily mapping it into the tlb.
1013  */
1014 void
1015 moea_zero_page(mmu_t mmu, vm_page_t m)
1016 {
1017 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1018 	void *va = (void *)pa;
1019 
1020 	bzero(va, PAGE_SIZE);
1021 }
1022 
1023 void
1024 moea_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1025 {
1026 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1027 	void *va = (void *)(pa + off);
1028 
1029 	bzero(va, size);
1030 }
1031 
1032 void
1033 moea_zero_page_idle(mmu_t mmu, vm_page_t m)
1034 {
1035 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1036 	void *va = (void *)pa;
1037 
1038 	bzero(va, PAGE_SIZE);
1039 }
1040 
1041 /*
1042  * Map the given physical page at the specified virtual address in the
1043  * target pmap with the protection requested.  If specified the page
1044  * will be wired down.
1045  */
1046 void
1047 moea_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1048 	   boolean_t wired)
1049 {
1050 
1051 	vm_page_lock_queues();
1052 	PMAP_LOCK(pmap);
1053 	moea_enter_locked(pmap, va, m, prot, wired);
1054 	vm_page_unlock_queues();
1055 	PMAP_UNLOCK(pmap);
1056 }
1057 
1058 /*
1059  * Map the given physical page at the specified virtual address in the
1060  * target pmap with the protection requested.  If specified the page
1061  * will be wired down.
1062  *
1063  * The page queues and pmap must be locked.
1064  */
1065 static void
1066 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1067     boolean_t wired)
1068 {
1069 	struct		pvo_head *pvo_head;
1070 	uma_zone_t	zone;
1071 	vm_page_t	pg;
1072 	u_int		pte_lo, pvo_flags, was_exec;
1073 	int		error;
1074 
1075 	if (!moea_initialized) {
1076 		pvo_head = &moea_pvo_kunmanaged;
1077 		zone = moea_upvo_zone;
1078 		pvo_flags = 0;
1079 		pg = NULL;
1080 		was_exec = PTE_EXEC;
1081 	} else {
1082 		pvo_head = vm_page_to_pvoh(m);
1083 		pg = m;
1084 		zone = moea_mpvo_zone;
1085 		pvo_flags = PVO_MANAGED;
1086 		was_exec = 0;
1087 	}
1088 	if (pmap_bootstrapped)
1089 		mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1090 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1091 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0 ||
1092 	    (m->oflags & VPO_BUSY) != 0 || VM_OBJECT_LOCKED(m->object),
1093 	    ("moea_enter_locked: page %p is not busy", m));
1094 
1095 	/* XXX change the pvo head for fake pages */
1096 	if ((m->flags & PG_FICTITIOUS) == PG_FICTITIOUS) {
1097 		pvo_flags &= ~PVO_MANAGED;
1098 		pvo_head = &moea_pvo_kunmanaged;
1099 		zone = moea_upvo_zone;
1100 	}
1101 
1102 	/*
1103 	 * If this is a managed page, and it's the first reference to the page,
1104 	 * clear the execness of the page.  Otherwise fetch the execness.
1105 	 */
1106 	if ((pg != NULL) && ((m->flags & PG_FICTITIOUS) == 0)) {
1107 		if (LIST_EMPTY(pvo_head)) {
1108 			moea_attr_clear(pg, PTE_EXEC);
1109 		} else {
1110 			was_exec = moea_attr_fetch(pg) & PTE_EXEC;
1111 		}
1112 	}
1113 
1114 	pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1115 
1116 	if (prot & VM_PROT_WRITE) {
1117 		pte_lo |= PTE_BW;
1118 		if (pmap_bootstrapped &&
1119 		    (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0)
1120 			vm_page_flag_set(m, PG_WRITEABLE);
1121 	} else
1122 		pte_lo |= PTE_BR;
1123 
1124 	if (prot & VM_PROT_EXECUTE)
1125 		pvo_flags |= PVO_EXECUTABLE;
1126 
1127 	if (wired)
1128 		pvo_flags |= PVO_WIRED;
1129 
1130 	if ((m->flags & PG_FICTITIOUS) != 0)
1131 		pvo_flags |= PVO_FAKE;
1132 
1133 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1134 	    pte_lo, pvo_flags);
1135 
1136 	/*
1137 	 * Flush the real page from the instruction cache if this page is
1138 	 * mapped executable and cacheable and was not previously mapped (or
1139 	 * was not mapped executable).
1140 	 */
1141 	if (error == 0 && (pvo_flags & PVO_EXECUTABLE) &&
1142 	    (pte_lo & PTE_I) == 0 && was_exec == 0) {
1143 		/*
1144 		 * Flush the real memory from the cache.
1145 		 */
1146 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1147 		if (pg != NULL)
1148 			moea_attr_save(pg, PTE_EXEC);
1149 	}
1150 
1151 	/* XXX syncicache always until problems are sorted */
1152 	moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1153 }
1154 
1155 /*
1156  * Maps a sequence of resident pages belonging to the same object.
1157  * The sequence begins with the given page m_start.  This page is
1158  * mapped at the given virtual address start.  Each subsequent page is
1159  * mapped at a virtual address that is offset from start by the same
1160  * amount as the page is offset from m_start within the object.  The
1161  * last page in the sequence is the page with the largest offset from
1162  * m_start that can be mapped at a virtual address less than the given
1163  * virtual address end.  Not every virtual page between start and end
1164  * is mapped; only those for which a resident page exists with the
1165  * corresponding offset from m_start are mapped.
1166  */
1167 void
1168 moea_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1169     vm_page_t m_start, vm_prot_t prot)
1170 {
1171 	vm_page_t m;
1172 	vm_pindex_t diff, psize;
1173 
1174 	psize = atop(end - start);
1175 	m = m_start;
1176 	vm_page_lock_queues();
1177 	PMAP_LOCK(pm);
1178 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1179 		moea_enter_locked(pm, start + ptoa(diff), m, prot &
1180 		    (VM_PROT_READ | VM_PROT_EXECUTE), FALSE);
1181 		m = TAILQ_NEXT(m, listq);
1182 	}
1183 	vm_page_unlock_queues();
1184 	PMAP_UNLOCK(pm);
1185 }
1186 
1187 void
1188 moea_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1189     vm_prot_t prot)
1190 {
1191 
1192 	vm_page_lock_queues();
1193 	PMAP_LOCK(pm);
1194 	moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1195 	    FALSE);
1196 	vm_page_unlock_queues();
1197 	PMAP_UNLOCK(pm);
1198 }
1199 
1200 vm_paddr_t
1201 moea_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1202 {
1203 	struct	pvo_entry *pvo;
1204 	vm_paddr_t pa;
1205 
1206 	PMAP_LOCK(pm);
1207 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1208 	if (pvo == NULL)
1209 		pa = 0;
1210 	else
1211 		pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1212 	PMAP_UNLOCK(pm);
1213 	return (pa);
1214 }
1215 
1216 /*
1217  * Atomically extract and hold the physical page with the given
1218  * pmap and virtual address pair if that mapping permits the given
1219  * protection.
1220  */
1221 vm_page_t
1222 moea_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1223 {
1224 	struct	pvo_entry *pvo;
1225 	vm_page_t m;
1226         vm_paddr_t pa;
1227 
1228 	m = NULL;
1229 	pa = 0;
1230 	PMAP_LOCK(pmap);
1231 retry:
1232 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1233 	if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1234 	    ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1235 	     (prot & VM_PROT_WRITE) == 0)) {
1236 		if (vm_page_pa_tryrelock(pmap, pvo->pvo_pte.pte.pte_lo & PTE_RPGN, &pa))
1237 			goto retry;
1238 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
1239 		vm_page_hold(m);
1240 	}
1241 	PA_UNLOCK_COND(pa);
1242 	PMAP_UNLOCK(pmap);
1243 	return (m);
1244 }
1245 
1246 void
1247 moea_init(mmu_t mmu)
1248 {
1249 
1250 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1251 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1252 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1253 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1254 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1255 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1256 	moea_initialized = TRUE;
1257 }
1258 
1259 boolean_t
1260 moea_is_referenced(mmu_t mmu, vm_page_t m)
1261 {
1262 
1263 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1264 	    ("moea_is_referenced: page %p is not managed", m));
1265 	return (moea_query_bit(m, PTE_REF));
1266 }
1267 
1268 boolean_t
1269 moea_is_modified(mmu_t mmu, vm_page_t m)
1270 {
1271 
1272 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1273 	    ("moea_is_modified: page %p is not managed", m));
1274 
1275 	/*
1276 	 * If the page is not VPO_BUSY, then PG_WRITEABLE cannot be
1277 	 * concurrently set while the object is locked.  Thus, if PG_WRITEABLE
1278 	 * is clear, no PTEs can have PTE_CHG set.
1279 	 */
1280 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1281 	if ((m->oflags & VPO_BUSY) == 0 &&
1282 	    (m->flags & PG_WRITEABLE) == 0)
1283 		return (FALSE);
1284 	return (moea_query_bit(m, PTE_CHG));
1285 }
1286 
1287 boolean_t
1288 moea_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1289 {
1290 	struct pvo_entry *pvo;
1291 	boolean_t rv;
1292 
1293 	PMAP_LOCK(pmap);
1294 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1295 	rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1296 	PMAP_UNLOCK(pmap);
1297 	return (rv);
1298 }
1299 
1300 void
1301 moea_clear_reference(mmu_t mmu, vm_page_t m)
1302 {
1303 
1304 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1305 	    ("moea_clear_reference: page %p is not managed", m));
1306 	moea_clear_bit(m, PTE_REF);
1307 }
1308 
1309 void
1310 moea_clear_modify(mmu_t mmu, vm_page_t m)
1311 {
1312 
1313 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1314 	    ("moea_clear_modify: page %p is not managed", m));
1315 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1316 	KASSERT((m->oflags & VPO_BUSY) == 0,
1317 	    ("moea_clear_modify: page %p is busy", m));
1318 
1319 	/*
1320 	 * If the page is not PG_WRITEABLE, then no PTEs can have PTE_CHG
1321 	 * set.  If the object containing the page is locked and the page is
1322 	 * not VPO_BUSY, then PG_WRITEABLE cannot be concurrently set.
1323 	 */
1324 	if ((m->flags & PG_WRITEABLE) == 0)
1325 		return;
1326 	moea_clear_bit(m, PTE_CHG);
1327 }
1328 
1329 /*
1330  * Clear the write and modified bits in each of the given page's mappings.
1331  */
1332 void
1333 moea_remove_write(mmu_t mmu, vm_page_t m)
1334 {
1335 	struct	pvo_entry *pvo;
1336 	struct	pte *pt;
1337 	pmap_t	pmap;
1338 	u_int	lo;
1339 
1340 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1341 	    ("moea_remove_write: page %p is not managed", m));
1342 
1343 	/*
1344 	 * If the page is not VPO_BUSY, then PG_WRITEABLE cannot be set by
1345 	 * another thread while the object is locked.  Thus, if PG_WRITEABLE
1346 	 * is clear, no page table entries need updating.
1347 	 */
1348 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1349 	if ((m->oflags & VPO_BUSY) == 0 &&
1350 	    (m->flags & PG_WRITEABLE) == 0)
1351 		return;
1352 	vm_page_lock_queues();
1353 	lo = moea_attr_fetch(m);
1354 	powerpc_sync();
1355 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1356 		pmap = pvo->pvo_pmap;
1357 		PMAP_LOCK(pmap);
1358 		if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1359 			pt = moea_pvo_to_pte(pvo, -1);
1360 			pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1361 			pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1362 			if (pt != NULL) {
1363 				moea_pte_synch(pt, &pvo->pvo_pte.pte);
1364 				lo |= pvo->pvo_pte.pte.pte_lo;
1365 				pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1366 				moea_pte_change(pt, &pvo->pvo_pte.pte,
1367 				    pvo->pvo_vaddr);
1368 				mtx_unlock(&moea_table_mutex);
1369 			}
1370 		}
1371 		PMAP_UNLOCK(pmap);
1372 	}
1373 	if ((lo & PTE_CHG) != 0) {
1374 		moea_attr_clear(m, PTE_CHG);
1375 		vm_page_dirty(m);
1376 	}
1377 	vm_page_flag_clear(m, PG_WRITEABLE);
1378 	vm_page_unlock_queues();
1379 }
1380 
1381 /*
1382  *	moea_ts_referenced:
1383  *
1384  *	Return a count of reference bits for a page, clearing those bits.
1385  *	It is not necessary for every reference bit to be cleared, but it
1386  *	is necessary that 0 only be returned when there are truly no
1387  *	reference bits set.
1388  *
1389  *	XXX: The exact number of bits to check and clear is a matter that
1390  *	should be tested and standardized at some point in the future for
1391  *	optimal aging of shared pages.
1392  */
1393 boolean_t
1394 moea_ts_referenced(mmu_t mmu, vm_page_t m)
1395 {
1396 
1397 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1398 	    ("moea_ts_referenced: page %p is not managed", m));
1399 	return (moea_clear_bit(m, PTE_REF));
1400 }
1401 
1402 /*
1403  * Modify the WIMG settings of all mappings for a page.
1404  */
1405 void
1406 moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1407 {
1408 	struct	pvo_entry *pvo;
1409 	struct	pvo_head *pvo_head;
1410 	struct	pte *pt;
1411 	pmap_t	pmap;
1412 	u_int	lo;
1413 
1414 	if (m->flags & PG_FICTITIOUS) {
1415 		m->md.mdpg_cache_attrs = ma;
1416 		return;
1417 	}
1418 
1419 	vm_page_lock_queues();
1420 	pvo_head = vm_page_to_pvoh(m);
1421 	lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1422 
1423 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1424 		pmap = pvo->pvo_pmap;
1425 		PMAP_LOCK(pmap);
1426 		pt = moea_pvo_to_pte(pvo, -1);
1427 		pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1428 		pvo->pvo_pte.pte.pte_lo |= lo;
1429 		if (pt != NULL) {
1430 			moea_pte_change(pt, &pvo->pvo_pte.pte,
1431 			    pvo->pvo_vaddr);
1432 			if (pvo->pvo_pmap == kernel_pmap)
1433 				isync();
1434 		}
1435 		mtx_unlock(&moea_table_mutex);
1436 		PMAP_UNLOCK(pmap);
1437 	}
1438 	m->md.mdpg_cache_attrs = ma;
1439 	vm_page_unlock_queues();
1440 }
1441 
1442 /*
1443  * Map a wired page into kernel virtual address space.
1444  */
1445 void
1446 moea_kenter(mmu_t mmu, vm_offset_t va, vm_offset_t pa)
1447 {
1448 
1449 	moea_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1450 }
1451 
1452 void
1453 moea_kenter_attr(mmu_t mmu, vm_offset_t va, vm_offset_t pa, vm_memattr_t ma)
1454 {
1455 	u_int		pte_lo;
1456 	int		error;
1457 
1458 #if 0
1459 	if (va < VM_MIN_KERNEL_ADDRESS)
1460 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1461 		    va);
1462 #endif
1463 
1464 	pte_lo = moea_calc_wimg(pa, ma);
1465 
1466 	PMAP_LOCK(kernel_pmap);
1467 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1468 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1469 
1470 	if (error != 0 && error != ENOENT)
1471 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1472 		    pa, error);
1473 
1474 	/*
1475 	 * Flush the real memory from the instruction cache.
1476 	 */
1477 	if ((pte_lo & (PTE_I | PTE_G)) == 0) {
1478 		moea_syncicache(pa, PAGE_SIZE);
1479 	}
1480 	PMAP_UNLOCK(kernel_pmap);
1481 }
1482 
1483 /*
1484  * Extract the physical page address associated with the given kernel virtual
1485  * address.
1486  */
1487 vm_offset_t
1488 moea_kextract(mmu_t mmu, vm_offset_t va)
1489 {
1490 	struct		pvo_entry *pvo;
1491 	vm_paddr_t pa;
1492 
1493 	/*
1494 	 * Allow direct mappings on 32-bit OEA
1495 	 */
1496 	if (va < VM_MIN_KERNEL_ADDRESS) {
1497 		return (va);
1498 	}
1499 
1500 	PMAP_LOCK(kernel_pmap);
1501 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1502 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1503 	pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1504 	PMAP_UNLOCK(kernel_pmap);
1505 	return (pa);
1506 }
1507 
1508 /*
1509  * Remove a wired page from kernel virtual address space.
1510  */
1511 void
1512 moea_kremove(mmu_t mmu, vm_offset_t va)
1513 {
1514 
1515 	moea_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1516 }
1517 
1518 /*
1519  * Map a range of physical addresses into kernel virtual address space.
1520  *
1521  * The value passed in *virt is a suggested virtual address for the mapping.
1522  * Architectures which can support a direct-mapped physical to virtual region
1523  * can return the appropriate address within that region, leaving '*virt'
1524  * unchanged.  We cannot and therefore do not; *virt is updated with the
1525  * first usable address after the mapped region.
1526  */
1527 vm_offset_t
1528 moea_map(mmu_t mmu, vm_offset_t *virt, vm_offset_t pa_start,
1529     vm_offset_t pa_end, int prot)
1530 {
1531 	vm_offset_t	sva, va;
1532 
1533 	sva = *virt;
1534 	va = sva;
1535 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1536 		moea_kenter(mmu, va, pa_start);
1537 	*virt = va;
1538 	return (sva);
1539 }
1540 
1541 /*
1542  * Returns true if the pmap's pv is one of the first
1543  * 16 pvs linked to from this page.  This count may
1544  * be changed upwards or downwards in the future; it
1545  * is only necessary that true be returned for a small
1546  * subset of pmaps for proper page aging.
1547  */
1548 boolean_t
1549 moea_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1550 {
1551         int loops;
1552 	struct pvo_entry *pvo;
1553 	boolean_t rv;
1554 
1555 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1556 	    ("moea_page_exists_quick: page %p is not managed", m));
1557 	loops = 0;
1558 	rv = FALSE;
1559 	vm_page_lock_queues();
1560 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1561 		if (pvo->pvo_pmap == pmap) {
1562 			rv = TRUE;
1563 			break;
1564 		}
1565 		if (++loops >= 16)
1566 			break;
1567 	}
1568 	vm_page_unlock_queues();
1569 	return (rv);
1570 }
1571 
1572 /*
1573  * Return the number of managed mappings to the given physical page
1574  * that are wired.
1575  */
1576 int
1577 moea_page_wired_mappings(mmu_t mmu, vm_page_t m)
1578 {
1579 	struct pvo_entry *pvo;
1580 	int count;
1581 
1582 	count = 0;
1583 	if ((m->flags & PG_FICTITIOUS) != 0)
1584 		return (count);
1585 	vm_page_lock_queues();
1586 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1587 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1588 			count++;
1589 	vm_page_unlock_queues();
1590 	return (count);
1591 }
1592 
1593 static u_int	moea_vsidcontext;
1594 
1595 void
1596 moea_pinit(mmu_t mmu, pmap_t pmap)
1597 {
1598 	int	i, mask;
1599 	u_int	entropy;
1600 
1601 	KASSERT((int)pmap < VM_MIN_KERNEL_ADDRESS, ("moea_pinit: virt pmap"));
1602 	PMAP_LOCK_INIT(pmap);
1603 
1604 	entropy = 0;
1605 	__asm __volatile("mftb %0" : "=r"(entropy));
1606 
1607 	if ((pmap->pmap_phys = (pmap_t)moea_kextract(mmu, (vm_offset_t)pmap))
1608 	    == NULL) {
1609 		pmap->pmap_phys = pmap;
1610 	}
1611 
1612 
1613 	mtx_lock(&moea_vsid_mutex);
1614 	/*
1615 	 * Allocate some segment registers for this pmap.
1616 	 */
1617 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1618 		u_int	hash, n;
1619 
1620 		/*
1621 		 * Create a new value by mutiplying by a prime and adding in
1622 		 * entropy from the timebase register.  This is to make the
1623 		 * VSID more random so that the PT hash function collides
1624 		 * less often.  (Note that the prime casues gcc to do shifts
1625 		 * instead of a multiply.)
1626 		 */
1627 		moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1628 		hash = moea_vsidcontext & (NPMAPS - 1);
1629 		if (hash == 0)		/* 0 is special, avoid it */
1630 			continue;
1631 		n = hash >> 5;
1632 		mask = 1 << (hash & (VSID_NBPW - 1));
1633 		hash = (moea_vsidcontext & 0xfffff);
1634 		if (moea_vsid_bitmap[n] & mask) {	/* collision? */
1635 			/* anything free in this bucket? */
1636 			if (moea_vsid_bitmap[n] == 0xffffffff) {
1637 				entropy = (moea_vsidcontext >> 20);
1638 				continue;
1639 			}
1640 			i = ffs(~moea_vsid_bitmap[n]) - 1;
1641 			mask = 1 << i;
1642 			hash &= 0xfffff & ~(VSID_NBPW - 1);
1643 			hash |= i;
1644 		}
1645 		moea_vsid_bitmap[n] |= mask;
1646 		for (i = 0; i < 16; i++)
1647 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1648 		mtx_unlock(&moea_vsid_mutex);
1649 		return;
1650 	}
1651 
1652 	mtx_unlock(&moea_vsid_mutex);
1653 	panic("moea_pinit: out of segments");
1654 }
1655 
1656 /*
1657  * Initialize the pmap associated with process 0.
1658  */
1659 void
1660 moea_pinit0(mmu_t mmu, pmap_t pm)
1661 {
1662 
1663 	moea_pinit(mmu, pm);
1664 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1665 }
1666 
1667 /*
1668  * Set the physical protection on the specified range of this map as requested.
1669  */
1670 void
1671 moea_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1672     vm_prot_t prot)
1673 {
1674 	struct	pvo_entry *pvo;
1675 	struct	pte *pt;
1676 	int	pteidx;
1677 
1678 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1679 	    ("moea_protect: non current pmap"));
1680 
1681 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1682 		moea_remove(mmu, pm, sva, eva);
1683 		return;
1684 	}
1685 
1686 	vm_page_lock_queues();
1687 	PMAP_LOCK(pm);
1688 	for (; sva < eva; sva += PAGE_SIZE) {
1689 		pvo = moea_pvo_find_va(pm, sva, &pteidx);
1690 		if (pvo == NULL)
1691 			continue;
1692 
1693 		if ((prot & VM_PROT_EXECUTE) == 0)
1694 			pvo->pvo_vaddr &= ~PVO_EXECUTABLE;
1695 
1696 		/*
1697 		 * Grab the PTE pointer before we diddle with the cached PTE
1698 		 * copy.
1699 		 */
1700 		pt = moea_pvo_to_pte(pvo, pteidx);
1701 		/*
1702 		 * Change the protection of the page.
1703 		 */
1704 		pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1705 		pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1706 
1707 		/*
1708 		 * If the PVO is in the page table, update that pte as well.
1709 		 */
1710 		if (pt != NULL) {
1711 			moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1712 			mtx_unlock(&moea_table_mutex);
1713 		}
1714 	}
1715 	vm_page_unlock_queues();
1716 	PMAP_UNLOCK(pm);
1717 }
1718 
1719 /*
1720  * Map a list of wired pages into kernel virtual address space.  This is
1721  * intended for temporary mappings which do not need page modification or
1722  * references recorded.  Existing mappings in the region are overwritten.
1723  */
1724 void
1725 moea_qenter(mmu_t mmu, vm_offset_t sva, vm_page_t *m, int count)
1726 {
1727 	vm_offset_t va;
1728 
1729 	va = sva;
1730 	while (count-- > 0) {
1731 		moea_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1732 		va += PAGE_SIZE;
1733 		m++;
1734 	}
1735 }
1736 
1737 /*
1738  * Remove page mappings from kernel virtual address space.  Intended for
1739  * temporary mappings entered by moea_qenter.
1740  */
1741 void
1742 moea_qremove(mmu_t mmu, vm_offset_t sva, int count)
1743 {
1744 	vm_offset_t va;
1745 
1746 	va = sva;
1747 	while (count-- > 0) {
1748 		moea_kremove(mmu, va);
1749 		va += PAGE_SIZE;
1750 	}
1751 }
1752 
1753 void
1754 moea_release(mmu_t mmu, pmap_t pmap)
1755 {
1756         int idx, mask;
1757 
1758 	/*
1759 	 * Free segment register's VSID
1760 	 */
1761         if (pmap->pm_sr[0] == 0)
1762                 panic("moea_release");
1763 
1764 	mtx_lock(&moea_vsid_mutex);
1765         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1766         mask = 1 << (idx % VSID_NBPW);
1767         idx /= VSID_NBPW;
1768         moea_vsid_bitmap[idx] &= ~mask;
1769 	mtx_unlock(&moea_vsid_mutex);
1770 	PMAP_LOCK_DESTROY(pmap);
1771 }
1772 
1773 /*
1774  * Remove the given range of addresses from the specified map.
1775  */
1776 void
1777 moea_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1778 {
1779 	struct	pvo_entry *pvo;
1780 	int	pteidx;
1781 
1782 	vm_page_lock_queues();
1783 	PMAP_LOCK(pm);
1784 	for (; sva < eva; sva += PAGE_SIZE) {
1785 		pvo = moea_pvo_find_va(pm, sva, &pteidx);
1786 		if (pvo != NULL) {
1787 			moea_pvo_remove(pvo, pteidx);
1788 		}
1789 	}
1790 	PMAP_UNLOCK(pm);
1791 	vm_page_unlock_queues();
1792 }
1793 
1794 /*
1795  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1796  * will reflect changes in pte's back to the vm_page.
1797  */
1798 void
1799 moea_remove_all(mmu_t mmu, vm_page_t m)
1800 {
1801 	struct  pvo_head *pvo_head;
1802 	struct	pvo_entry *pvo, *next_pvo;
1803 	pmap_t	pmap;
1804 
1805 	vm_page_lock_queues();
1806 	pvo_head = vm_page_to_pvoh(m);
1807 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1808 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1809 
1810 		pmap = pvo->pvo_pmap;
1811 		PMAP_LOCK(pmap);
1812 		moea_pvo_remove(pvo, -1);
1813 		PMAP_UNLOCK(pmap);
1814 	}
1815 	if ((m->flags & PG_WRITEABLE) && moea_is_modified(mmu, m)) {
1816 		moea_attr_clear(m, PTE_CHG);
1817 		vm_page_dirty(m);
1818 	}
1819 	vm_page_flag_clear(m, PG_WRITEABLE);
1820 	vm_page_unlock_queues();
1821 }
1822 
1823 /*
1824  * Allocate a physical page of memory directly from the phys_avail map.
1825  * Can only be called from moea_bootstrap before avail start and end are
1826  * calculated.
1827  */
1828 static vm_offset_t
1829 moea_bootstrap_alloc(vm_size_t size, u_int align)
1830 {
1831 	vm_offset_t	s, e;
1832 	int		i, j;
1833 
1834 	size = round_page(size);
1835 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1836 		if (align != 0)
1837 			s = (phys_avail[i] + align - 1) & ~(align - 1);
1838 		else
1839 			s = phys_avail[i];
1840 		e = s + size;
1841 
1842 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1843 			continue;
1844 
1845 		if (s == phys_avail[i]) {
1846 			phys_avail[i] += size;
1847 		} else if (e == phys_avail[i + 1]) {
1848 			phys_avail[i + 1] -= size;
1849 		} else {
1850 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1851 				phys_avail[j] = phys_avail[j - 2];
1852 				phys_avail[j + 1] = phys_avail[j - 1];
1853 			}
1854 
1855 			phys_avail[i + 3] = phys_avail[i + 1];
1856 			phys_avail[i + 1] = s;
1857 			phys_avail[i + 2] = e;
1858 			phys_avail_count++;
1859 		}
1860 
1861 		return (s);
1862 	}
1863 	panic("moea_bootstrap_alloc: could not allocate memory");
1864 }
1865 
1866 static void
1867 moea_syncicache(vm_offset_t pa, vm_size_t len)
1868 {
1869 	__syncicache((void *)pa, len);
1870 }
1871 
1872 static int
1873 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1874     vm_offset_t va, vm_offset_t pa, u_int pte_lo, int flags)
1875 {
1876 	struct	pvo_entry *pvo;
1877 	u_int	sr;
1878 	int	first;
1879 	u_int	ptegidx;
1880 	int	i;
1881 	int     bootstrap;
1882 
1883 	moea_pvo_enter_calls++;
1884 	first = 0;
1885 	bootstrap = 0;
1886 
1887 	/*
1888 	 * Compute the PTE Group index.
1889 	 */
1890 	va &= ~ADDR_POFF;
1891 	sr = va_to_sr(pm->pm_sr, va);
1892 	ptegidx = va_to_pteg(sr, va);
1893 
1894 	/*
1895 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
1896 	 * there is a mapping.
1897 	 */
1898 	mtx_lock(&moea_table_mutex);
1899 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
1900 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
1901 			if ((pvo->pvo_pte.pte.pte_lo & PTE_RPGN) == pa &&
1902 			    (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
1903 			    (pte_lo & PTE_PP)) {
1904 				mtx_unlock(&moea_table_mutex);
1905 				return (0);
1906 			}
1907 			moea_pvo_remove(pvo, -1);
1908 			break;
1909 		}
1910 	}
1911 
1912 	/*
1913 	 * If we aren't overwriting a mapping, try to allocate.
1914 	 */
1915 	if (moea_initialized) {
1916 		pvo = uma_zalloc(zone, M_NOWAIT);
1917 	} else {
1918 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
1919 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
1920 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
1921 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
1922 		}
1923 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
1924 		moea_bpvo_pool_index++;
1925 		bootstrap = 1;
1926 	}
1927 
1928 	if (pvo == NULL) {
1929 		mtx_unlock(&moea_table_mutex);
1930 		return (ENOMEM);
1931 	}
1932 
1933 	moea_pvo_entries++;
1934 	pvo->pvo_vaddr = va;
1935 	pvo->pvo_pmap = pm;
1936 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
1937 	pvo->pvo_vaddr &= ~ADDR_POFF;
1938 	if (flags & VM_PROT_EXECUTE)
1939 		pvo->pvo_vaddr |= PVO_EXECUTABLE;
1940 	if (flags & PVO_WIRED)
1941 		pvo->pvo_vaddr |= PVO_WIRED;
1942 	if (pvo_head != &moea_pvo_kunmanaged)
1943 		pvo->pvo_vaddr |= PVO_MANAGED;
1944 	if (bootstrap)
1945 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
1946 	if (flags & PVO_FAKE)
1947 		pvo->pvo_vaddr |= PVO_FAKE;
1948 
1949 	moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo);
1950 
1951 	/*
1952 	 * Remember if the list was empty and therefore will be the first
1953 	 * item.
1954 	 */
1955 	if (LIST_FIRST(pvo_head) == NULL)
1956 		first = 1;
1957 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
1958 
1959 	if (pvo->pvo_pte.pte.pte_lo & PVO_WIRED)
1960 		pm->pm_stats.wired_count++;
1961 	pm->pm_stats.resident_count++;
1962 
1963 	/*
1964 	 * We hope this succeeds but it isn't required.
1965 	 */
1966 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
1967 	if (i >= 0) {
1968 		PVO_PTEGIDX_SET(pvo, i);
1969 	} else {
1970 		panic("moea_pvo_enter: overflow");
1971 		moea_pte_overflow++;
1972 	}
1973 	mtx_unlock(&moea_table_mutex);
1974 
1975 	return (first ? ENOENT : 0);
1976 }
1977 
1978 static void
1979 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
1980 {
1981 	struct	pte *pt;
1982 
1983 	/*
1984 	 * If there is an active pte entry, we need to deactivate it (and
1985 	 * save the ref & cfg bits).
1986 	 */
1987 	pt = moea_pvo_to_pte(pvo, pteidx);
1988 	if (pt != NULL) {
1989 		moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1990 		mtx_unlock(&moea_table_mutex);
1991 		PVO_PTEGIDX_CLR(pvo);
1992 	} else {
1993 		moea_pte_overflow--;
1994 	}
1995 
1996 	/*
1997 	 * Update our statistics.
1998 	 */
1999 	pvo->pvo_pmap->pm_stats.resident_count--;
2000 	if (pvo->pvo_pte.pte.pte_lo & PVO_WIRED)
2001 		pvo->pvo_pmap->pm_stats.wired_count--;
2002 
2003 	/*
2004 	 * Save the REF/CHG bits into their cache if the page is managed.
2005 	 */
2006 	if ((pvo->pvo_vaddr & (PVO_MANAGED|PVO_FAKE)) == PVO_MANAGED) {
2007 		struct	vm_page *pg;
2008 
2009 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
2010 		if (pg != NULL) {
2011 			moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2012 			    (PTE_REF | PTE_CHG));
2013 		}
2014 	}
2015 
2016 	/*
2017 	 * Remove this PVO from the PV list.
2018 	 */
2019 	LIST_REMOVE(pvo, pvo_vlink);
2020 
2021 	/*
2022 	 * Remove this from the overflow list and return it to the pool
2023 	 * if we aren't going to reuse it.
2024 	 */
2025 	LIST_REMOVE(pvo, pvo_olink);
2026 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2027 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
2028 		    moea_upvo_zone, pvo);
2029 	moea_pvo_entries--;
2030 	moea_pvo_remove_calls++;
2031 }
2032 
2033 static __inline int
2034 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2035 {
2036 	int	pteidx;
2037 
2038 	/*
2039 	 * We can find the actual pte entry without searching by grabbing
2040 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2041 	 * noticing the HID bit.
2042 	 */
2043 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2044 	if (pvo->pvo_pte.pte.pte_hi & PTE_HID)
2045 		pteidx ^= moea_pteg_mask * 8;
2046 
2047 	return (pteidx);
2048 }
2049 
2050 static struct pvo_entry *
2051 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2052 {
2053 	struct	pvo_entry *pvo;
2054 	int	ptegidx;
2055 	u_int	sr;
2056 
2057 	va &= ~ADDR_POFF;
2058 	sr = va_to_sr(pm->pm_sr, va);
2059 	ptegidx = va_to_pteg(sr, va);
2060 
2061 	mtx_lock(&moea_table_mutex);
2062 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2063 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2064 			if (pteidx_p)
2065 				*pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2066 			break;
2067 		}
2068 	}
2069 	mtx_unlock(&moea_table_mutex);
2070 
2071 	return (pvo);
2072 }
2073 
2074 static struct pte *
2075 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2076 {
2077 	struct	pte *pt;
2078 
2079 	/*
2080 	 * If we haven't been supplied the ptegidx, calculate it.
2081 	 */
2082 	if (pteidx == -1) {
2083 		int	ptegidx;
2084 		u_int	sr;
2085 
2086 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2087 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2088 		pteidx = moea_pvo_pte_index(pvo, ptegidx);
2089 	}
2090 
2091 	pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2092 	mtx_lock(&moea_table_mutex);
2093 
2094 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2095 		panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2096 		    "valid pte index", pvo);
2097 	}
2098 
2099 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2100 		panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2101 		    "pvo but no valid pte", pvo);
2102 	}
2103 
2104 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2105 		if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) {
2106 			panic("moea_pvo_to_pte: pvo %p has valid pte in "
2107 			    "moea_pteg_table %p but invalid in pvo", pvo, pt);
2108 		}
2109 
2110 		if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2111 		    != 0) {
2112 			panic("moea_pvo_to_pte: pvo %p pte does not match "
2113 			    "pte %p in moea_pteg_table", pvo, pt);
2114 		}
2115 
2116 		mtx_assert(&moea_table_mutex, MA_OWNED);
2117 		return (pt);
2118 	}
2119 
2120 	if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) {
2121 		panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2122 		    "moea_pteg_table but valid in pvo", pvo, pt);
2123 	}
2124 
2125 	mtx_unlock(&moea_table_mutex);
2126 	return (NULL);
2127 }
2128 
2129 /*
2130  * XXX: THIS STUFF SHOULD BE IN pte.c?
2131  */
2132 int
2133 moea_pte_spill(vm_offset_t addr)
2134 {
2135 	struct	pvo_entry *source_pvo, *victim_pvo;
2136 	struct	pvo_entry *pvo;
2137 	int	ptegidx, i, j;
2138 	u_int	sr;
2139 	struct	pteg *pteg;
2140 	struct	pte *pt;
2141 
2142 	moea_pte_spills++;
2143 
2144 	sr = mfsrin(addr);
2145 	ptegidx = va_to_pteg(sr, addr);
2146 
2147 	/*
2148 	 * Have to substitute some entry.  Use the primary hash for this.
2149 	 * Use low bits of timebase as random generator.
2150 	 */
2151 	pteg = &moea_pteg_table[ptegidx];
2152 	mtx_lock(&moea_table_mutex);
2153 	__asm __volatile("mftb %0" : "=r"(i));
2154 	i &= 7;
2155 	pt = &pteg->pt[i];
2156 
2157 	source_pvo = NULL;
2158 	victim_pvo = NULL;
2159 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2160 		/*
2161 		 * We need to find a pvo entry for this address.
2162 		 */
2163 		if (source_pvo == NULL &&
2164 		    moea_pte_match(&pvo->pvo_pte.pte, sr, addr,
2165 		    pvo->pvo_pte.pte.pte_hi & PTE_HID)) {
2166 			/*
2167 			 * Now found an entry to be spilled into the pteg.
2168 			 * The PTE is now valid, so we know it's active.
2169 			 */
2170 			j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2171 
2172 			if (j >= 0) {
2173 				PVO_PTEGIDX_SET(pvo, j);
2174 				moea_pte_overflow--;
2175 				mtx_unlock(&moea_table_mutex);
2176 				return (1);
2177 			}
2178 
2179 			source_pvo = pvo;
2180 
2181 			if (victim_pvo != NULL)
2182 				break;
2183 		}
2184 
2185 		/*
2186 		 * We also need the pvo entry of the victim we are replacing
2187 		 * so save the R & C bits of the PTE.
2188 		 */
2189 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2190 		    moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2191 			victim_pvo = pvo;
2192 			if (source_pvo != NULL)
2193 				break;
2194 		}
2195 	}
2196 
2197 	if (source_pvo == NULL) {
2198 		mtx_unlock(&moea_table_mutex);
2199 		return (0);
2200 	}
2201 
2202 	if (victim_pvo == NULL) {
2203 		if ((pt->pte_hi & PTE_HID) == 0)
2204 			panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2205 			    "entry", pt);
2206 
2207 		/*
2208 		 * If this is a secondary PTE, we need to search it's primary
2209 		 * pvo bucket for the matching PVO.
2210 		 */
2211 		LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2212 		    pvo_olink) {
2213 			/*
2214 			 * We also need the pvo entry of the victim we are
2215 			 * replacing so save the R & C bits of the PTE.
2216 			 */
2217 			if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2218 				victim_pvo = pvo;
2219 				break;
2220 			}
2221 		}
2222 
2223 		if (victim_pvo == NULL)
2224 			panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2225 			    "entry", pt);
2226 	}
2227 
2228 	/*
2229 	 * We are invalidating the TLB entry for the EA we are replacing even
2230 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2231 	 * contained in the TLB entry.
2232 	 */
2233 	source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID;
2234 
2235 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2236 	moea_pte_set(pt, &source_pvo->pvo_pte.pte);
2237 
2238 	PVO_PTEGIDX_CLR(victim_pvo);
2239 	PVO_PTEGIDX_SET(source_pvo, i);
2240 	moea_pte_replacements++;
2241 
2242 	mtx_unlock(&moea_table_mutex);
2243 	return (1);
2244 }
2245 
2246 static int
2247 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2248 {
2249 	struct	pte *pt;
2250 	int	i;
2251 
2252 	mtx_assert(&moea_table_mutex, MA_OWNED);
2253 
2254 	/*
2255 	 * First try primary hash.
2256 	 */
2257 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2258 		if ((pt->pte_hi & PTE_VALID) == 0) {
2259 			pvo_pt->pte_hi &= ~PTE_HID;
2260 			moea_pte_set(pt, pvo_pt);
2261 			return (i);
2262 		}
2263 	}
2264 
2265 	/*
2266 	 * Now try secondary hash.
2267 	 */
2268 	ptegidx ^= moea_pteg_mask;
2269 
2270 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2271 		if ((pt->pte_hi & PTE_VALID) == 0) {
2272 			pvo_pt->pte_hi |= PTE_HID;
2273 			moea_pte_set(pt, pvo_pt);
2274 			return (i);
2275 		}
2276 	}
2277 
2278 	panic("moea_pte_insert: overflow");
2279 	return (-1);
2280 }
2281 
2282 static boolean_t
2283 moea_query_bit(vm_page_t m, int ptebit)
2284 {
2285 	struct	pvo_entry *pvo;
2286 	struct	pte *pt;
2287 
2288 	if (moea_attr_fetch(m) & ptebit)
2289 		return (TRUE);
2290 
2291 	vm_page_lock_queues();
2292 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2293 
2294 		/*
2295 		 * See if we saved the bit off.  If so, cache it and return
2296 		 * success.
2297 		 */
2298 		if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2299 			moea_attr_save(m, ptebit);
2300 			vm_page_unlock_queues();
2301 			return (TRUE);
2302 		}
2303 	}
2304 
2305 	/*
2306 	 * No luck, now go through the hard part of looking at the PTEs
2307 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2308 	 * the PTEs.
2309 	 */
2310 	powerpc_sync();
2311 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2312 
2313 		/*
2314 		 * See if this pvo has a valid PTE.  if so, fetch the
2315 		 * REF/CHG bits from the valid PTE.  If the appropriate
2316 		 * ptebit is set, cache it and return success.
2317 		 */
2318 		pt = moea_pvo_to_pte(pvo, -1);
2319 		if (pt != NULL) {
2320 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2321 			mtx_unlock(&moea_table_mutex);
2322 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2323 				moea_attr_save(m, ptebit);
2324 				vm_page_unlock_queues();
2325 				return (TRUE);
2326 			}
2327 		}
2328 	}
2329 
2330 	vm_page_unlock_queues();
2331 	return (FALSE);
2332 }
2333 
2334 static u_int
2335 moea_clear_bit(vm_page_t m, int ptebit)
2336 {
2337 	u_int	count;
2338 	struct	pvo_entry *pvo;
2339 	struct	pte *pt;
2340 
2341 	vm_page_lock_queues();
2342 
2343 	/*
2344 	 * Clear the cached value.
2345 	 */
2346 	moea_attr_clear(m, ptebit);
2347 
2348 	/*
2349 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2350 	 * we can reset the right ones).  note that since the pvo entries and
2351 	 * list heads are accessed via BAT0 and are never placed in the page
2352 	 * table, we don't have to worry about further accesses setting the
2353 	 * REF/CHG bits.
2354 	 */
2355 	powerpc_sync();
2356 
2357 	/*
2358 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2359 	 * valid pte clear the ptebit from the valid pte.
2360 	 */
2361 	count = 0;
2362 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2363 		pt = moea_pvo_to_pte(pvo, -1);
2364 		if (pt != NULL) {
2365 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2366 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2367 				count++;
2368 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2369 			}
2370 			mtx_unlock(&moea_table_mutex);
2371 		}
2372 		pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2373 	}
2374 
2375 	vm_page_unlock_queues();
2376 	return (count);
2377 }
2378 
2379 /*
2380  * Return true if the physical range is encompassed by the battable[idx]
2381  */
2382 static int
2383 moea_bat_mapped(int idx, vm_offset_t pa, vm_size_t size)
2384 {
2385 	u_int prot;
2386 	u_int32_t start;
2387 	u_int32_t end;
2388 	u_int32_t bat_ble;
2389 
2390 	/*
2391 	 * Return immediately if not a valid mapping
2392 	 */
2393 	if (!(battable[idx].batu & BAT_Vs))
2394 		return (EINVAL);
2395 
2396 	/*
2397 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2398 	 * so it can function as an i/o page
2399 	 */
2400 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2401 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2402 		return (EPERM);
2403 
2404 	/*
2405 	 * The address should be within the BAT range. Assume that the
2406 	 * start address in the BAT has the correct alignment (thus
2407 	 * not requiring masking)
2408 	 */
2409 	start = battable[idx].batl & BAT_PBS;
2410 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2411 	end = start | (bat_ble << 15) | 0x7fff;
2412 
2413 	if ((pa < start) || ((pa + size) > end))
2414 		return (ERANGE);
2415 
2416 	return (0);
2417 }
2418 
2419 boolean_t
2420 moea_dev_direct_mapped(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2421 {
2422 	int i;
2423 
2424 	/*
2425 	 * This currently does not work for entries that
2426 	 * overlap 256M BAT segments.
2427 	 */
2428 
2429 	for(i = 0; i < 16; i++)
2430 		if (moea_bat_mapped(i, pa, size) == 0)
2431 			return (0);
2432 
2433 	return (EFAULT);
2434 }
2435 
2436 /*
2437  * Map a set of physical memory pages into the kernel virtual
2438  * address space. Return a pointer to where it is mapped. This
2439  * routine is intended to be used for mapping device memory,
2440  * NOT real memory.
2441  */
2442 void *
2443 moea_mapdev(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2444 {
2445 
2446 	return (moea_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT));
2447 }
2448 
2449 void *
2450 moea_mapdev_attr(mmu_t mmu, vm_offset_t pa, vm_size_t size, vm_memattr_t ma)
2451 {
2452 	vm_offset_t va, tmpva, ppa, offset;
2453 	int i;
2454 
2455 	ppa = trunc_page(pa);
2456 	offset = pa & PAGE_MASK;
2457 	size = roundup(offset + size, PAGE_SIZE);
2458 
2459 	/*
2460 	 * If the physical address lies within a valid BAT table entry,
2461 	 * return the 1:1 mapping. This currently doesn't work
2462 	 * for regions that overlap 256M BAT segments.
2463 	 */
2464 	for (i = 0; i < 16; i++) {
2465 		if (moea_bat_mapped(i, pa, size) == 0)
2466 			return ((void *) pa);
2467 	}
2468 
2469 	va = kmem_alloc_nofault(kernel_map, size);
2470 	if (!va)
2471 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2472 
2473 	for (tmpva = va; size > 0;) {
2474 		moea_kenter_attr(mmu, tmpva, ppa, ma);
2475 		tlbie(tmpva);
2476 		size -= PAGE_SIZE;
2477 		tmpva += PAGE_SIZE;
2478 		ppa += PAGE_SIZE;
2479 	}
2480 
2481 	return ((void *)(va + offset));
2482 }
2483 
2484 void
2485 moea_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2486 {
2487 	vm_offset_t base, offset;
2488 
2489 	/*
2490 	 * If this is outside kernel virtual space, then it's a
2491 	 * battable entry and doesn't require unmapping
2492 	 */
2493 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2494 		base = trunc_page(va);
2495 		offset = va & PAGE_MASK;
2496 		size = roundup(offset + size, PAGE_SIZE);
2497 		kmem_free(kernel_map, base, size);
2498 	}
2499 }
2500 
2501 static void
2502 moea_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2503 {
2504 	struct pvo_entry *pvo;
2505 	vm_offset_t lim;
2506 	vm_paddr_t pa;
2507 	vm_size_t len;
2508 
2509 	PMAP_LOCK(pm);
2510 	while (sz > 0) {
2511 		lim = round_page(va);
2512 		len = MIN(lim - va, sz);
2513 		pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2514 		if (pvo != NULL) {
2515 			pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) |
2516 			    (va & ADDR_POFF);
2517 			moea_syncicache(pa, len);
2518 		}
2519 		va += len;
2520 		sz -= len;
2521 	}
2522 	PMAP_UNLOCK(pm);
2523 }
2524