xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision b78ee15e9f04ae15c3e1200df974473167524d17)
1 /*-
2  * Copyright (c) 2001 The NetBSD Foundation, Inc.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to The NetBSD Foundation
6  * by Matt Thomas <matt@3am-software.com> of Allegro Networks, Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
18  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
19  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
21  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
22  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
23  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
24  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
25  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
26  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
27  * POSSIBILITY OF SUCH DAMAGE.
28  */
29 /*-
30  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
31  * Copyright (C) 1995, 1996 TooLs GmbH.
32  * All rights reserved.
33  *
34  * Redistribution and use in source and binary forms, with or without
35  * modification, are permitted provided that the following conditions
36  * are met:
37  * 1. Redistributions of source code must retain the above copyright
38  *    notice, this list of conditions and the following disclaimer.
39  * 2. Redistributions in binary form must reproduce the above copyright
40  *    notice, this list of conditions and the following disclaimer in the
41  *    documentation and/or other materials provided with the distribution.
42  * 3. All advertising materials mentioning features or use of this software
43  *    must display the following acknowledgement:
44  *	This product includes software developed by TooLs GmbH.
45  * 4. The name of TooLs GmbH may not be used to endorse or promote products
46  *    derived from this software without specific prior written permission.
47  *
48  * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
49  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
50  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
51  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
52  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
53  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
54  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
55  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
56  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
57  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
58  *
59  * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $
60  */
61 /*-
62  * Copyright (C) 2001 Benno Rice.
63  * All rights reserved.
64  *
65  * Redistribution and use in source and binary forms, with or without
66  * modification, are permitted provided that the following conditions
67  * are met:
68  * 1. Redistributions of source code must retain the above copyright
69  *    notice, this list of conditions and the following disclaimer.
70  * 2. Redistributions in binary form must reproduce the above copyright
71  *    notice, this list of conditions and the following disclaimer in the
72  *    documentation and/or other materials provided with the distribution.
73  *
74  * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR
75  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
76  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
77  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
78  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
79  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
80  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
81  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
82  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
83  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
84  */
85 
86 #include <sys/cdefs.h>
87 __FBSDID("$FreeBSD$");
88 
89 /*
90  * Manages physical address maps.
91  *
92  * Since the information managed by this module is also stored by the
93  * logical address mapping module, this module may throw away valid virtual
94  * to physical mappings at almost any time.  However, invalidations of
95  * mappings must be done as requested.
96  *
97  * In order to cope with hardware architectures which make virtual to
98  * physical map invalidates expensive, this module may delay invalidate
99  * reduced protection operations until such time as they are actually
100  * necessary.  This module is given full information as to which processors
101  * are currently using which maps, and to when physical maps must be made
102  * correct.
103  */
104 
105 #include "opt_kstack_pages.h"
106 
107 #include <sys/param.h>
108 #include <sys/kernel.h>
109 #include <sys/conf.h>
110 #include <sys/queue.h>
111 #include <sys/cpuset.h>
112 #include <sys/kerneldump.h>
113 #include <sys/ktr.h>
114 #include <sys/lock.h>
115 #include <sys/msgbuf.h>
116 #include <sys/mutex.h>
117 #include <sys/proc.h>
118 #include <sys/rwlock.h>
119 #include <sys/sched.h>
120 #include <sys/sysctl.h>
121 #include <sys/systm.h>
122 #include <sys/vmmeter.h>
123 
124 #include <dev/ofw/openfirm.h>
125 
126 #include <vm/vm.h>
127 #include <vm/vm_param.h>
128 #include <vm/vm_kern.h>
129 #include <vm/vm_page.h>
130 #include <vm/vm_map.h>
131 #include <vm/vm_object.h>
132 #include <vm/vm_extern.h>
133 #include <vm/vm_pageout.h>
134 #include <vm/uma.h>
135 
136 #include <machine/cpu.h>
137 #include <machine/platform.h>
138 #include <machine/bat.h>
139 #include <machine/frame.h>
140 #include <machine/md_var.h>
141 #include <machine/psl.h>
142 #include <machine/pte.h>
143 #include <machine/smp.h>
144 #include <machine/sr.h>
145 #include <machine/mmuvar.h>
146 #include <machine/trap.h>
147 
148 #include "mmu_if.h"
149 
150 #define	MOEA_DEBUG
151 
152 #define TODO	panic("%s: not implemented", __func__);
153 
154 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
155 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
156 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
157 
158 struct ofw_map {
159 	vm_offset_t	om_va;
160 	vm_size_t	om_len;
161 	vm_offset_t	om_pa;
162 	u_int		om_mode;
163 };
164 
165 extern unsigned char _etext[];
166 extern unsigned char _end[];
167 
168 /*
169  * Map of physical memory regions.
170  */
171 static struct	mem_region *regions;
172 static struct	mem_region *pregions;
173 static u_int    phys_avail_count;
174 static int	regions_sz, pregions_sz;
175 static struct	ofw_map *translations;
176 
177 /*
178  * Lock for the pteg and pvo tables.
179  */
180 struct mtx	moea_table_mutex;
181 struct mtx	moea_vsid_mutex;
182 
183 /* tlbie instruction synchronization */
184 static struct mtx tlbie_mtx;
185 
186 /*
187  * PTEG data.
188  */
189 static struct	pteg *moea_pteg_table;
190 u_int		moea_pteg_count;
191 u_int		moea_pteg_mask;
192 
193 /*
194  * PVO data.
195  */
196 struct	pvo_head *moea_pvo_table;		/* pvo entries by pteg index */
197 struct	pvo_head moea_pvo_kunmanaged =
198     LIST_HEAD_INITIALIZER(moea_pvo_kunmanaged);	/* list of unmanaged pages */
199 
200 static struct rwlock_padalign pvh_global_lock;
201 
202 uma_zone_t	moea_upvo_zone;	/* zone for pvo entries for unmanaged pages */
203 uma_zone_t	moea_mpvo_zone;	/* zone for pvo entries for managed pages */
204 
205 #define	BPVO_POOL_SIZE	32768
206 static struct	pvo_entry *moea_bpvo_pool;
207 static int	moea_bpvo_pool_index = 0;
208 
209 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
210 static u_int	moea_vsid_bitmap[NPMAPS / VSID_NBPW];
211 
212 static boolean_t moea_initialized = FALSE;
213 
214 /*
215  * Statistics.
216  */
217 u_int	moea_pte_valid = 0;
218 u_int	moea_pte_overflow = 0;
219 u_int	moea_pte_replacements = 0;
220 u_int	moea_pvo_entries = 0;
221 u_int	moea_pvo_enter_calls = 0;
222 u_int	moea_pvo_remove_calls = 0;
223 u_int	moea_pte_spills = 0;
224 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_valid, CTLFLAG_RD, &moea_pte_valid,
225     0, "");
226 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_overflow, CTLFLAG_RD,
227     &moea_pte_overflow, 0, "");
228 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_replacements, CTLFLAG_RD,
229     &moea_pte_replacements, 0, "");
230 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_entries, CTLFLAG_RD, &moea_pvo_entries,
231     0, "");
232 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_enter_calls, CTLFLAG_RD,
233     &moea_pvo_enter_calls, 0, "");
234 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_remove_calls, CTLFLAG_RD,
235     &moea_pvo_remove_calls, 0, "");
236 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_spills, CTLFLAG_RD,
237     &moea_pte_spills, 0, "");
238 
239 /*
240  * Allocate physical memory for use in moea_bootstrap.
241  */
242 static vm_offset_t	moea_bootstrap_alloc(vm_size_t, u_int);
243 
244 /*
245  * PTE calls.
246  */
247 static int		moea_pte_insert(u_int, struct pte *);
248 
249 /*
250  * PVO calls.
251  */
252 static int	moea_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *,
253 		    vm_offset_t, vm_offset_t, u_int, int);
254 static void	moea_pvo_remove(struct pvo_entry *, int);
255 static struct	pvo_entry *moea_pvo_find_va(pmap_t, vm_offset_t, int *);
256 static struct	pte *moea_pvo_to_pte(const struct pvo_entry *, int);
257 
258 /*
259  * Utility routines.
260  */
261 static int		moea_enter_locked(pmap_t, vm_offset_t, vm_page_t,
262 			    vm_prot_t, u_int, int8_t);
263 static void		moea_syncicache(vm_offset_t, vm_size_t);
264 static boolean_t	moea_query_bit(vm_page_t, int);
265 static u_int		moea_clear_bit(vm_page_t, int);
266 static void		moea_kremove(mmu_t, vm_offset_t);
267 int		moea_pte_spill(vm_offset_t);
268 
269 /*
270  * Kernel MMU interface
271  */
272 void moea_clear_modify(mmu_t, vm_page_t);
273 void moea_copy_page(mmu_t, vm_page_t, vm_page_t);
274 void moea_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
275     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
276 int moea_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int,
277     int8_t);
278 void moea_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
279     vm_prot_t);
280 void moea_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
281 vm_paddr_t moea_extract(mmu_t, pmap_t, vm_offset_t);
282 vm_page_t moea_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
283 void moea_init(mmu_t);
284 boolean_t moea_is_modified(mmu_t, vm_page_t);
285 boolean_t moea_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
286 boolean_t moea_is_referenced(mmu_t, vm_page_t);
287 int moea_ts_referenced(mmu_t, vm_page_t);
288 vm_offset_t moea_map(mmu_t, vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
289 boolean_t moea_page_exists_quick(mmu_t, pmap_t, vm_page_t);
290 int moea_page_wired_mappings(mmu_t, vm_page_t);
291 void moea_pinit(mmu_t, pmap_t);
292 void moea_pinit0(mmu_t, pmap_t);
293 void moea_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
294 void moea_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
295 void moea_qremove(mmu_t, vm_offset_t, int);
296 void moea_release(mmu_t, pmap_t);
297 void moea_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
298 void moea_remove_all(mmu_t, vm_page_t);
299 void moea_remove_write(mmu_t, vm_page_t);
300 void moea_unwire(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
301 void moea_zero_page(mmu_t, vm_page_t);
302 void moea_zero_page_area(mmu_t, vm_page_t, int, int);
303 void moea_zero_page_idle(mmu_t, vm_page_t);
304 void moea_activate(mmu_t, struct thread *);
305 void moea_deactivate(mmu_t, struct thread *);
306 void moea_cpu_bootstrap(mmu_t, int);
307 void moea_bootstrap(mmu_t, vm_offset_t, vm_offset_t);
308 void *moea_mapdev(mmu_t, vm_paddr_t, vm_size_t);
309 void *moea_mapdev_attr(mmu_t, vm_offset_t, vm_size_t, vm_memattr_t);
310 void moea_unmapdev(mmu_t, vm_offset_t, vm_size_t);
311 vm_paddr_t moea_kextract(mmu_t, vm_offset_t);
312 void moea_kenter_attr(mmu_t, vm_offset_t, vm_offset_t, vm_memattr_t);
313 void moea_kenter(mmu_t, vm_offset_t, vm_paddr_t);
314 void moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma);
315 boolean_t moea_dev_direct_mapped(mmu_t, vm_paddr_t, vm_size_t);
316 static void moea_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
317 void moea_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va);
318 void moea_scan_init(mmu_t mmu);
319 
320 static mmu_method_t moea_methods[] = {
321 	MMUMETHOD(mmu_clear_modify,	moea_clear_modify),
322 	MMUMETHOD(mmu_copy_page,	moea_copy_page),
323 	MMUMETHOD(mmu_copy_pages,	moea_copy_pages),
324 	MMUMETHOD(mmu_enter,		moea_enter),
325 	MMUMETHOD(mmu_enter_object,	moea_enter_object),
326 	MMUMETHOD(mmu_enter_quick,	moea_enter_quick),
327 	MMUMETHOD(mmu_extract,		moea_extract),
328 	MMUMETHOD(mmu_extract_and_hold,	moea_extract_and_hold),
329 	MMUMETHOD(mmu_init,		moea_init),
330 	MMUMETHOD(mmu_is_modified,	moea_is_modified),
331 	MMUMETHOD(mmu_is_prefaultable,	moea_is_prefaultable),
332 	MMUMETHOD(mmu_is_referenced,	moea_is_referenced),
333 	MMUMETHOD(mmu_ts_referenced,	moea_ts_referenced),
334 	MMUMETHOD(mmu_map,     		moea_map),
335 	MMUMETHOD(mmu_page_exists_quick,moea_page_exists_quick),
336 	MMUMETHOD(mmu_page_wired_mappings,moea_page_wired_mappings),
337 	MMUMETHOD(mmu_pinit,		moea_pinit),
338 	MMUMETHOD(mmu_pinit0,		moea_pinit0),
339 	MMUMETHOD(mmu_protect,		moea_protect),
340 	MMUMETHOD(mmu_qenter,		moea_qenter),
341 	MMUMETHOD(mmu_qremove,		moea_qremove),
342 	MMUMETHOD(mmu_release,		moea_release),
343 	MMUMETHOD(mmu_remove,		moea_remove),
344 	MMUMETHOD(mmu_remove_all,      	moea_remove_all),
345 	MMUMETHOD(mmu_remove_write,	moea_remove_write),
346 	MMUMETHOD(mmu_sync_icache,	moea_sync_icache),
347 	MMUMETHOD(mmu_unwire,		moea_unwire),
348 	MMUMETHOD(mmu_zero_page,       	moea_zero_page),
349 	MMUMETHOD(mmu_zero_page_area,	moea_zero_page_area),
350 	MMUMETHOD(mmu_zero_page_idle,	moea_zero_page_idle),
351 	MMUMETHOD(mmu_activate,		moea_activate),
352 	MMUMETHOD(mmu_deactivate,      	moea_deactivate),
353 	MMUMETHOD(mmu_page_set_memattr,	moea_page_set_memattr),
354 
355 	/* Internal interfaces */
356 	MMUMETHOD(mmu_bootstrap,       	moea_bootstrap),
357 	MMUMETHOD(mmu_cpu_bootstrap,   	moea_cpu_bootstrap),
358 	MMUMETHOD(mmu_mapdev_attr,	moea_mapdev_attr),
359 	MMUMETHOD(mmu_mapdev,		moea_mapdev),
360 	MMUMETHOD(mmu_unmapdev,		moea_unmapdev),
361 	MMUMETHOD(mmu_kextract,		moea_kextract),
362 	MMUMETHOD(mmu_kenter,		moea_kenter),
363 	MMUMETHOD(mmu_kenter_attr,	moea_kenter_attr),
364 	MMUMETHOD(mmu_dev_direct_mapped,moea_dev_direct_mapped),
365 	MMUMETHOD(mmu_scan_init,	moea_scan_init),
366 	MMUMETHOD(mmu_dumpsys_map,	moea_dumpsys_map),
367 
368 	{ 0, 0 }
369 };
370 
371 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods, 0);
372 
373 static __inline uint32_t
374 moea_calc_wimg(vm_offset_t pa, vm_memattr_t ma)
375 {
376 	uint32_t pte_lo;
377 	int i;
378 
379 	if (ma != VM_MEMATTR_DEFAULT) {
380 		switch (ma) {
381 		case VM_MEMATTR_UNCACHEABLE:
382 			return (PTE_I | PTE_G);
383 		case VM_MEMATTR_WRITE_COMBINING:
384 		case VM_MEMATTR_WRITE_BACK:
385 		case VM_MEMATTR_PREFETCHABLE:
386 			return (PTE_I);
387 		case VM_MEMATTR_WRITE_THROUGH:
388 			return (PTE_W | PTE_M);
389 		}
390 	}
391 
392 	/*
393 	 * Assume the page is cache inhibited and access is guarded unless
394 	 * it's in our available memory array.
395 	 */
396 	pte_lo = PTE_I | PTE_G;
397 	for (i = 0; i < pregions_sz; i++) {
398 		if ((pa >= pregions[i].mr_start) &&
399 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
400 			pte_lo = PTE_M;
401 			break;
402 		}
403 	}
404 
405 	return pte_lo;
406 }
407 
408 static void
409 tlbie(vm_offset_t va)
410 {
411 
412 	mtx_lock_spin(&tlbie_mtx);
413 	__asm __volatile("ptesync");
414 	__asm __volatile("tlbie %0" :: "r"(va));
415 	__asm __volatile("eieio; tlbsync; ptesync");
416 	mtx_unlock_spin(&tlbie_mtx);
417 }
418 
419 static void
420 tlbia(void)
421 {
422 	vm_offset_t va;
423 
424 	for (va = 0; va < 0x00040000; va += 0x00001000) {
425 		__asm __volatile("tlbie %0" :: "r"(va));
426 		powerpc_sync();
427 	}
428 	__asm __volatile("tlbsync");
429 	powerpc_sync();
430 }
431 
432 static __inline int
433 va_to_sr(u_int *sr, vm_offset_t va)
434 {
435 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
436 }
437 
438 static __inline u_int
439 va_to_pteg(u_int sr, vm_offset_t addr)
440 {
441 	u_int hash;
442 
443 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
444 	    ADDR_PIDX_SHFT);
445 	return (hash & moea_pteg_mask);
446 }
447 
448 static __inline struct pvo_head *
449 vm_page_to_pvoh(vm_page_t m)
450 {
451 
452 	return (&m->md.mdpg_pvoh);
453 }
454 
455 static __inline void
456 moea_attr_clear(vm_page_t m, int ptebit)
457 {
458 
459 	rw_assert(&pvh_global_lock, RA_WLOCKED);
460 	m->md.mdpg_attrs &= ~ptebit;
461 }
462 
463 static __inline int
464 moea_attr_fetch(vm_page_t m)
465 {
466 
467 	return (m->md.mdpg_attrs);
468 }
469 
470 static __inline void
471 moea_attr_save(vm_page_t m, int ptebit)
472 {
473 
474 	rw_assert(&pvh_global_lock, RA_WLOCKED);
475 	m->md.mdpg_attrs |= ptebit;
476 }
477 
478 static __inline int
479 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
480 {
481 	if (pt->pte_hi == pvo_pt->pte_hi)
482 		return (1);
483 
484 	return (0);
485 }
486 
487 static __inline int
488 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
489 {
490 	return (pt->pte_hi & ~PTE_VALID) ==
491 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
492 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
493 }
494 
495 static __inline void
496 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
497 {
498 
499 	mtx_assert(&moea_table_mutex, MA_OWNED);
500 
501 	/*
502 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
503 	 * set when the real pte is set in memory.
504 	 *
505 	 * Note: Don't set the valid bit for correct operation of tlb update.
506 	 */
507 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
508 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
509 	pt->pte_lo = pte_lo;
510 }
511 
512 static __inline void
513 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
514 {
515 
516 	mtx_assert(&moea_table_mutex, MA_OWNED);
517 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
518 }
519 
520 static __inline void
521 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
522 {
523 
524 	mtx_assert(&moea_table_mutex, MA_OWNED);
525 
526 	/*
527 	 * As shown in Section 7.6.3.2.3
528 	 */
529 	pt->pte_lo &= ~ptebit;
530 	tlbie(va);
531 }
532 
533 static __inline void
534 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
535 {
536 
537 	mtx_assert(&moea_table_mutex, MA_OWNED);
538 	pvo_pt->pte_hi |= PTE_VALID;
539 
540 	/*
541 	 * Update the PTE as defined in section 7.6.3.1.
542 	 * Note that the REF/CHG bits are from pvo_pt and thus should have
543 	 * been saved so this routine can restore them (if desired).
544 	 */
545 	pt->pte_lo = pvo_pt->pte_lo;
546 	powerpc_sync();
547 	pt->pte_hi = pvo_pt->pte_hi;
548 	powerpc_sync();
549 	moea_pte_valid++;
550 }
551 
552 static __inline void
553 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
554 {
555 
556 	mtx_assert(&moea_table_mutex, MA_OWNED);
557 	pvo_pt->pte_hi &= ~PTE_VALID;
558 
559 	/*
560 	 * Force the reg & chg bits back into the PTEs.
561 	 */
562 	powerpc_sync();
563 
564 	/*
565 	 * Invalidate the pte.
566 	 */
567 	pt->pte_hi &= ~PTE_VALID;
568 
569 	tlbie(va);
570 
571 	/*
572 	 * Save the reg & chg bits.
573 	 */
574 	moea_pte_synch(pt, pvo_pt);
575 	moea_pte_valid--;
576 }
577 
578 static __inline void
579 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
580 {
581 
582 	/*
583 	 * Invalidate the PTE
584 	 */
585 	moea_pte_unset(pt, pvo_pt, va);
586 	moea_pte_set(pt, pvo_pt);
587 }
588 
589 /*
590  * Quick sort callout for comparing memory regions.
591  */
592 static int	om_cmp(const void *a, const void *b);
593 
594 static int
595 om_cmp(const void *a, const void *b)
596 {
597 	const struct	ofw_map *mapa;
598 	const struct	ofw_map *mapb;
599 
600 	mapa = a;
601 	mapb = b;
602 	if (mapa->om_pa < mapb->om_pa)
603 		return (-1);
604 	else if (mapa->om_pa > mapb->om_pa)
605 		return (1);
606 	else
607 		return (0);
608 }
609 
610 void
611 moea_cpu_bootstrap(mmu_t mmup, int ap)
612 {
613 	u_int sdr;
614 	int i;
615 
616 	if (ap) {
617 		powerpc_sync();
618 		__asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu));
619 		__asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl));
620 		isync();
621 		__asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu));
622 		__asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl));
623 		isync();
624 	}
625 
626 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
627 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
628 	isync();
629 
630 	__asm __volatile("mtibatu 1,%0" :: "r"(0));
631 	__asm __volatile("mtdbatu 2,%0" :: "r"(0));
632 	__asm __volatile("mtibatu 2,%0" :: "r"(0));
633 	__asm __volatile("mtdbatu 3,%0" :: "r"(0));
634 	__asm __volatile("mtibatu 3,%0" :: "r"(0));
635 	isync();
636 
637 	for (i = 0; i < 16; i++)
638 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
639 	powerpc_sync();
640 
641 	sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10);
642 	__asm __volatile("mtsdr1 %0" :: "r"(sdr));
643 	isync();
644 
645 	tlbia();
646 }
647 
648 void
649 moea_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
650 {
651 	ihandle_t	mmui;
652 	phandle_t	chosen, mmu;
653 	int		sz;
654 	int		i, j;
655 	vm_size_t	size, physsz, hwphyssz;
656 	vm_offset_t	pa, va, off;
657 	void		*dpcpu;
658 	register_t	msr;
659 
660         /*
661          * Set up BAT0 to map the lowest 256 MB area
662          */
663         battable[0x0].batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
664         battable[0x0].batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
665 
666 	/*
667 	 * Map PCI memory space.
668 	 */
669 	battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
670 	battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
671 
672 	battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
673 	battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
674 
675 	battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
676 	battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
677 
678 	battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
679 	battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
680 
681 	/*
682 	 * Map obio devices.
683 	 */
684 	battable[0xf].batl = BATL(0xf0000000, BAT_I|BAT_G, BAT_PP_RW);
685 	battable[0xf].batu = BATU(0xf0000000, BAT_BL_256M, BAT_Vs);
686 
687 	/*
688 	 * Use an IBAT and a DBAT to map the bottom segment of memory
689 	 * where we are. Turn off instruction relocation temporarily
690 	 * to prevent faults while reprogramming the IBAT.
691 	 */
692 	msr = mfmsr();
693 	mtmsr(msr & ~PSL_IR);
694 	__asm (".balign 32; \n"
695 	       "mtibatu 0,%0; mtibatl 0,%1; isync; \n"
696 	       "mtdbatu 0,%0; mtdbatl 0,%1; isync"
697 	    :: "r"(battable[0].batu), "r"(battable[0].batl));
698 	mtmsr(msr);
699 
700 	/* map pci space */
701 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
702 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
703 	isync();
704 
705 	/* set global direct map flag */
706 	hw_direct_map = 1;
707 
708 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
709 	CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
710 
711 	for (i = 0; i < pregions_sz; i++) {
712 		vm_offset_t pa;
713 		vm_offset_t end;
714 
715 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
716 			pregions[i].mr_start,
717 			pregions[i].mr_start + pregions[i].mr_size,
718 			pregions[i].mr_size);
719 		/*
720 		 * Install entries into the BAT table to allow all
721 		 * of physmem to be convered by on-demand BAT entries.
722 		 * The loop will sometimes set the same battable element
723 		 * twice, but that's fine since they won't be used for
724 		 * a while yet.
725 		 */
726 		pa = pregions[i].mr_start & 0xf0000000;
727 		end = pregions[i].mr_start + pregions[i].mr_size;
728 		do {
729                         u_int n = pa >> ADDR_SR_SHFT;
730 
731 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
732 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
733 			pa += SEGMENT_LENGTH;
734 		} while (pa < end);
735 	}
736 
737 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
738 		panic("moea_bootstrap: phys_avail too small");
739 
740 	phys_avail_count = 0;
741 	physsz = 0;
742 	hwphyssz = 0;
743 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
744 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
745 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
746 		    regions[i].mr_start + regions[i].mr_size,
747 		    regions[i].mr_size);
748 		if (hwphyssz != 0 &&
749 		    (physsz + regions[i].mr_size) >= hwphyssz) {
750 			if (physsz < hwphyssz) {
751 				phys_avail[j] = regions[i].mr_start;
752 				phys_avail[j + 1] = regions[i].mr_start +
753 				    hwphyssz - physsz;
754 				physsz = hwphyssz;
755 				phys_avail_count++;
756 			}
757 			break;
758 		}
759 		phys_avail[j] = regions[i].mr_start;
760 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
761 		phys_avail_count++;
762 		physsz += regions[i].mr_size;
763 	}
764 
765 	/* Check for overlap with the kernel and exception vectors */
766 	for (j = 0; j < 2*phys_avail_count; j+=2) {
767 		if (phys_avail[j] < EXC_LAST)
768 			phys_avail[j] += EXC_LAST;
769 
770 		if (kernelstart >= phys_avail[j] &&
771 		    kernelstart < phys_avail[j+1]) {
772 			if (kernelend < phys_avail[j+1]) {
773 				phys_avail[2*phys_avail_count] =
774 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
775 				phys_avail[2*phys_avail_count + 1] =
776 				    phys_avail[j+1];
777 				phys_avail_count++;
778 			}
779 
780 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
781 		}
782 
783 		if (kernelend >= phys_avail[j] &&
784 		    kernelend < phys_avail[j+1]) {
785 			if (kernelstart > phys_avail[j]) {
786 				phys_avail[2*phys_avail_count] = phys_avail[j];
787 				phys_avail[2*phys_avail_count + 1] =
788 				    kernelstart & ~PAGE_MASK;
789 				phys_avail_count++;
790 			}
791 
792 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
793 		}
794 	}
795 
796 	physmem = btoc(physsz);
797 
798 	/*
799 	 * Allocate PTEG table.
800 	 */
801 #ifdef PTEGCOUNT
802 	moea_pteg_count = PTEGCOUNT;
803 #else
804 	moea_pteg_count = 0x1000;
805 
806 	while (moea_pteg_count < physmem)
807 		moea_pteg_count <<= 1;
808 
809 	moea_pteg_count >>= 1;
810 #endif /* PTEGCOUNT */
811 
812 	size = moea_pteg_count * sizeof(struct pteg);
813 	CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
814 	    size);
815 	moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
816 	CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
817 	bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
818 	moea_pteg_mask = moea_pteg_count - 1;
819 
820 	/*
821 	 * Allocate pv/overflow lists.
822 	 */
823 	size = sizeof(struct pvo_head) * moea_pteg_count;
824 	moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
825 	    PAGE_SIZE);
826 	CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
827 	for (i = 0; i < moea_pteg_count; i++)
828 		LIST_INIT(&moea_pvo_table[i]);
829 
830 	/*
831 	 * Initialize the lock that synchronizes access to the pteg and pvo
832 	 * tables.
833 	 */
834 	mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF |
835 	    MTX_RECURSE);
836 	mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF);
837 
838 	mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN);
839 
840 	/*
841 	 * Initialise the unmanaged pvo pool.
842 	 */
843 	moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
844 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
845 	moea_bpvo_pool_index = 0;
846 
847 	/*
848 	 * Make sure kernel vsid is allocated as well as VSID 0.
849 	 */
850 	moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
851 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
852 	moea_vsid_bitmap[0] |= 1;
853 
854 	/*
855 	 * Initialize the kernel pmap (which is statically allocated).
856 	 */
857 	PMAP_LOCK_INIT(kernel_pmap);
858 	for (i = 0; i < 16; i++)
859 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
860 	CPU_FILL(&kernel_pmap->pm_active);
861 	RB_INIT(&kernel_pmap->pmap_pvo);
862 
863  	/*
864 	 * Initialize the global pv list lock.
865 	 */
866 	rw_init(&pvh_global_lock, "pmap pv global");
867 
868 	/*
869 	 * Set up the Open Firmware mappings
870 	 */
871 	chosen = OF_finddevice("/chosen");
872 	if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1 &&
873 	    (mmu = OF_instance_to_package(mmui)) != -1 &&
874 	    (sz = OF_getproplen(mmu, "translations")) != -1) {
875 		translations = NULL;
876 		for (i = 0; phys_avail[i] != 0; i += 2) {
877 			if (phys_avail[i + 1] >= sz) {
878 				translations = (struct ofw_map *)phys_avail[i];
879 				break;
880 			}
881 		}
882 		if (translations == NULL)
883 			panic("moea_bootstrap: no space to copy translations");
884 		bzero(translations, sz);
885 		if (OF_getprop(mmu, "translations", translations, sz) == -1)
886 			panic("moea_bootstrap: can't get ofw translations");
887 		CTR0(KTR_PMAP, "moea_bootstrap: translations");
888 		sz /= sizeof(*translations);
889 		qsort(translations, sz, sizeof (*translations), om_cmp);
890 		for (i = 0; i < sz; i++) {
891 			CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
892 			    translations[i].om_pa, translations[i].om_va,
893 			    translations[i].om_len);
894 
895 			/*
896 			 * If the mapping is 1:1, let the RAM and device
897 			 * on-demand BAT tables take care of the translation.
898 			 */
899 			if (translations[i].om_va == translations[i].om_pa)
900 				continue;
901 
902 			/* Enter the pages */
903 			for (off = 0; off < translations[i].om_len;
904 			    off += PAGE_SIZE)
905 				moea_kenter(mmup, translations[i].om_va + off,
906 					    translations[i].om_pa + off);
907 		}
908 	}
909 
910 	/*
911 	 * Calculate the last available physical address.
912 	 */
913 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
914 		;
915 	Maxmem = powerpc_btop(phys_avail[i + 1]);
916 
917 	moea_cpu_bootstrap(mmup,0);
918 
919 	pmap_bootstrapped++;
920 
921 	/*
922 	 * Set the start and end of kva.
923 	 */
924 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
925 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
926 
927 	/*
928 	 * Allocate a kernel stack with a guard page for thread0 and map it
929 	 * into the kernel page map.
930 	 */
931 	pa = moea_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, PAGE_SIZE);
932 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
933 	virtual_avail = va + KSTACK_PAGES * PAGE_SIZE;
934 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
935 	thread0.td_kstack = va;
936 	thread0.td_kstack_pages = KSTACK_PAGES;
937 	for (i = 0; i < KSTACK_PAGES; i++) {
938 		moea_kenter(mmup, va, pa);
939 		pa += PAGE_SIZE;
940 		va += PAGE_SIZE;
941 	}
942 
943 	/*
944 	 * Allocate virtual address space for the message buffer.
945 	 */
946 	pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE);
947 	msgbufp = (struct msgbuf *)virtual_avail;
948 	va = virtual_avail;
949 	virtual_avail += round_page(msgbufsize);
950 	while (va < virtual_avail) {
951 		moea_kenter(mmup, va, pa);
952 		pa += PAGE_SIZE;
953 		va += PAGE_SIZE;
954 	}
955 
956 	/*
957 	 * Allocate virtual address space for the dynamic percpu area.
958 	 */
959 	pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
960 	dpcpu = (void *)virtual_avail;
961 	va = virtual_avail;
962 	virtual_avail += DPCPU_SIZE;
963 	while (va < virtual_avail) {
964 		moea_kenter(mmup, va, pa);
965 		pa += PAGE_SIZE;
966 		va += PAGE_SIZE;
967 	}
968 	dpcpu_init(dpcpu, 0);
969 }
970 
971 /*
972  * Activate a user pmap.  The pmap must be activated before it's address
973  * space can be accessed in any way.
974  */
975 void
976 moea_activate(mmu_t mmu, struct thread *td)
977 {
978 	pmap_t	pm, pmr;
979 
980 	/*
981 	 * Load all the data we need up front to encourage the compiler to
982 	 * not issue any loads while we have interrupts disabled below.
983 	 */
984 	pm = &td->td_proc->p_vmspace->vm_pmap;
985 	pmr = pm->pmap_phys;
986 
987 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
988 	PCPU_SET(curpmap, pmr);
989 
990 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
991 }
992 
993 void
994 moea_deactivate(mmu_t mmu, struct thread *td)
995 {
996 	pmap_t	pm;
997 
998 	pm = &td->td_proc->p_vmspace->vm_pmap;
999 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1000 	PCPU_SET(curpmap, NULL);
1001 }
1002 
1003 void
1004 moea_unwire(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1005 {
1006 	struct	pvo_entry key, *pvo;
1007 
1008 	PMAP_LOCK(pm);
1009 	key.pvo_vaddr = sva;
1010 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1011 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1012 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1013 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1014 			panic("moea_unwire: pvo %p is missing PVO_WIRED", pvo);
1015 		pvo->pvo_vaddr &= ~PVO_WIRED;
1016 		pm->pm_stats.wired_count--;
1017 	}
1018 	PMAP_UNLOCK(pm);
1019 }
1020 
1021 void
1022 moea_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1023 {
1024 	vm_offset_t	dst;
1025 	vm_offset_t	src;
1026 
1027 	dst = VM_PAGE_TO_PHYS(mdst);
1028 	src = VM_PAGE_TO_PHYS(msrc);
1029 
1030 	bcopy((void *)src, (void *)dst, PAGE_SIZE);
1031 }
1032 
1033 void
1034 moea_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1035     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1036 {
1037 	void *a_cp, *b_cp;
1038 	vm_offset_t a_pg_offset, b_pg_offset;
1039 	int cnt;
1040 
1041 	while (xfersize > 0) {
1042 		a_pg_offset = a_offset & PAGE_MASK;
1043 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1044 		a_cp = (char *)VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]) +
1045 		    a_pg_offset;
1046 		b_pg_offset = b_offset & PAGE_MASK;
1047 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1048 		b_cp = (char *)VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]) +
1049 		    b_pg_offset;
1050 		bcopy(a_cp, b_cp, cnt);
1051 		a_offset += cnt;
1052 		b_offset += cnt;
1053 		xfersize -= cnt;
1054 	}
1055 }
1056 
1057 /*
1058  * Zero a page of physical memory by temporarily mapping it into the tlb.
1059  */
1060 void
1061 moea_zero_page(mmu_t mmu, vm_page_t m)
1062 {
1063 	vm_offset_t off, pa = VM_PAGE_TO_PHYS(m);
1064 
1065 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1066 		__asm __volatile("dcbz 0,%0" :: "r"(pa + off));
1067 }
1068 
1069 void
1070 moea_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1071 {
1072 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1073 	void *va = (void *)(pa + off);
1074 
1075 	bzero(va, size);
1076 }
1077 
1078 void
1079 moea_zero_page_idle(mmu_t mmu, vm_page_t m)
1080 {
1081 
1082 	moea_zero_page(mmu, m);
1083 }
1084 
1085 /*
1086  * Map the given physical page at the specified virtual address in the
1087  * target pmap with the protection requested.  If specified the page
1088  * will be wired down.
1089  */
1090 int
1091 moea_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1092     u_int flags, int8_t psind)
1093 {
1094 	int error;
1095 
1096 	for (;;) {
1097 		rw_wlock(&pvh_global_lock);
1098 		PMAP_LOCK(pmap);
1099 		error = moea_enter_locked(pmap, va, m, prot, flags, psind);
1100 		rw_wunlock(&pvh_global_lock);
1101 		PMAP_UNLOCK(pmap);
1102 		if (error != ENOMEM)
1103 			return (KERN_SUCCESS);
1104 		if ((flags & PMAP_ENTER_NOSLEEP) != 0)
1105 			return (KERN_RESOURCE_SHORTAGE);
1106 		VM_OBJECT_ASSERT_UNLOCKED(m->object);
1107 		VM_WAIT;
1108 	}
1109 }
1110 
1111 /*
1112  * Map the given physical page at the specified virtual address in the
1113  * target pmap with the protection requested.  If specified the page
1114  * will be wired down.
1115  *
1116  * The global pvh and pmap must be locked.
1117  */
1118 static int
1119 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1120     u_int flags, int8_t psind __unused)
1121 {
1122 	struct		pvo_head *pvo_head;
1123 	uma_zone_t	zone;
1124 	u_int		pte_lo, pvo_flags;
1125 	int		error;
1126 
1127 	if (pmap_bootstrapped)
1128 		rw_assert(&pvh_global_lock, RA_WLOCKED);
1129 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1130 	if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m))
1131 		VM_OBJECT_ASSERT_LOCKED(m->object);
1132 
1133 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea_initialized) {
1134 		pvo_head = &moea_pvo_kunmanaged;
1135 		zone = moea_upvo_zone;
1136 		pvo_flags = 0;
1137 	} else {
1138 		pvo_head = vm_page_to_pvoh(m);
1139 		zone = moea_mpvo_zone;
1140 		pvo_flags = PVO_MANAGED;
1141 	}
1142 
1143 	pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1144 
1145 	if (prot & VM_PROT_WRITE) {
1146 		pte_lo |= PTE_BW;
1147 		if (pmap_bootstrapped &&
1148 		    (m->oflags & VPO_UNMANAGED) == 0)
1149 			vm_page_aflag_set(m, PGA_WRITEABLE);
1150 	} else
1151 		pte_lo |= PTE_BR;
1152 
1153 	if ((flags & PMAP_ENTER_WIRED) != 0)
1154 		pvo_flags |= PVO_WIRED;
1155 
1156 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1157 	    pte_lo, pvo_flags);
1158 
1159 	/*
1160 	 * Flush the real page from the instruction cache. This has be done
1161 	 * for all user mappings to prevent information leakage via the
1162 	 * instruction cache. moea_pvo_enter() returns ENOENT for the first
1163 	 * mapping for a page.
1164 	 */
1165 	if (pmap != kernel_pmap && error == ENOENT &&
1166 	    (pte_lo & (PTE_I | PTE_G)) == 0)
1167 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1168 
1169 	return (error);
1170 }
1171 
1172 /*
1173  * Maps a sequence of resident pages belonging to the same object.
1174  * The sequence begins with the given page m_start.  This page is
1175  * mapped at the given virtual address start.  Each subsequent page is
1176  * mapped at a virtual address that is offset from start by the same
1177  * amount as the page is offset from m_start within the object.  The
1178  * last page in the sequence is the page with the largest offset from
1179  * m_start that can be mapped at a virtual address less than the given
1180  * virtual address end.  Not every virtual page between start and end
1181  * is mapped; only those for which a resident page exists with the
1182  * corresponding offset from m_start are mapped.
1183  */
1184 void
1185 moea_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1186     vm_page_t m_start, vm_prot_t prot)
1187 {
1188 	vm_page_t m;
1189 	vm_pindex_t diff, psize;
1190 
1191 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1192 
1193 	psize = atop(end - start);
1194 	m = m_start;
1195 	rw_wlock(&pvh_global_lock);
1196 	PMAP_LOCK(pm);
1197 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1198 		moea_enter_locked(pm, start + ptoa(diff), m, prot &
1199 		    (VM_PROT_READ | VM_PROT_EXECUTE), 0, 0);
1200 		m = TAILQ_NEXT(m, listq);
1201 	}
1202 	rw_wunlock(&pvh_global_lock);
1203 	PMAP_UNLOCK(pm);
1204 }
1205 
1206 void
1207 moea_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1208     vm_prot_t prot)
1209 {
1210 
1211 	rw_wlock(&pvh_global_lock);
1212 	PMAP_LOCK(pm);
1213 	moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1214 	    0, 0);
1215 	rw_wunlock(&pvh_global_lock);
1216 	PMAP_UNLOCK(pm);
1217 }
1218 
1219 vm_paddr_t
1220 moea_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1221 {
1222 	struct	pvo_entry *pvo;
1223 	vm_paddr_t pa;
1224 
1225 	PMAP_LOCK(pm);
1226 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1227 	if (pvo == NULL)
1228 		pa = 0;
1229 	else
1230 		pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1231 	PMAP_UNLOCK(pm);
1232 	return (pa);
1233 }
1234 
1235 /*
1236  * Atomically extract and hold the physical page with the given
1237  * pmap and virtual address pair if that mapping permits the given
1238  * protection.
1239  */
1240 vm_page_t
1241 moea_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1242 {
1243 	struct	pvo_entry *pvo;
1244 	vm_page_t m;
1245         vm_paddr_t pa;
1246 
1247 	m = NULL;
1248 	pa = 0;
1249 	PMAP_LOCK(pmap);
1250 retry:
1251 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1252 	if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1253 	    ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1254 	     (prot & VM_PROT_WRITE) == 0)) {
1255 		if (vm_page_pa_tryrelock(pmap, pvo->pvo_pte.pte.pte_lo & PTE_RPGN, &pa))
1256 			goto retry;
1257 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
1258 		vm_page_hold(m);
1259 	}
1260 	PA_UNLOCK_COND(pa);
1261 	PMAP_UNLOCK(pmap);
1262 	return (m);
1263 }
1264 
1265 void
1266 moea_init(mmu_t mmu)
1267 {
1268 
1269 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1270 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1271 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1272 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1273 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1274 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1275 	moea_initialized = TRUE;
1276 }
1277 
1278 boolean_t
1279 moea_is_referenced(mmu_t mmu, vm_page_t m)
1280 {
1281 	boolean_t rv;
1282 
1283 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1284 	    ("moea_is_referenced: page %p is not managed", m));
1285 	rw_wlock(&pvh_global_lock);
1286 	rv = moea_query_bit(m, PTE_REF);
1287 	rw_wunlock(&pvh_global_lock);
1288 	return (rv);
1289 }
1290 
1291 boolean_t
1292 moea_is_modified(mmu_t mmu, vm_page_t m)
1293 {
1294 	boolean_t rv;
1295 
1296 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1297 	    ("moea_is_modified: page %p is not managed", m));
1298 
1299 	/*
1300 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1301 	 * concurrently set while the object is locked.  Thus, if PGA_WRITEABLE
1302 	 * is clear, no PTEs can have PTE_CHG set.
1303 	 */
1304 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1305 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1306 		return (FALSE);
1307 	rw_wlock(&pvh_global_lock);
1308 	rv = moea_query_bit(m, PTE_CHG);
1309 	rw_wunlock(&pvh_global_lock);
1310 	return (rv);
1311 }
1312 
1313 boolean_t
1314 moea_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1315 {
1316 	struct pvo_entry *pvo;
1317 	boolean_t rv;
1318 
1319 	PMAP_LOCK(pmap);
1320 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1321 	rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1322 	PMAP_UNLOCK(pmap);
1323 	return (rv);
1324 }
1325 
1326 void
1327 moea_clear_modify(mmu_t mmu, vm_page_t m)
1328 {
1329 
1330 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1331 	    ("moea_clear_modify: page %p is not managed", m));
1332 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1333 	KASSERT(!vm_page_xbusied(m),
1334 	    ("moea_clear_modify: page %p is exclusive busy", m));
1335 
1336 	/*
1337 	 * If the page is not PGA_WRITEABLE, then no PTEs can have PTE_CHG
1338 	 * set.  If the object containing the page is locked and the page is
1339 	 * not exclusive busied, then PGA_WRITEABLE cannot be concurrently set.
1340 	 */
1341 	if ((m->aflags & PGA_WRITEABLE) == 0)
1342 		return;
1343 	rw_wlock(&pvh_global_lock);
1344 	moea_clear_bit(m, PTE_CHG);
1345 	rw_wunlock(&pvh_global_lock);
1346 }
1347 
1348 /*
1349  * Clear the write and modified bits in each of the given page's mappings.
1350  */
1351 void
1352 moea_remove_write(mmu_t mmu, vm_page_t m)
1353 {
1354 	struct	pvo_entry *pvo;
1355 	struct	pte *pt;
1356 	pmap_t	pmap;
1357 	u_int	lo;
1358 
1359 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1360 	    ("moea_remove_write: page %p is not managed", m));
1361 
1362 	/*
1363 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1364 	 * set by another thread while the object is locked.  Thus,
1365 	 * if PGA_WRITEABLE is clear, no page table entries need updating.
1366 	 */
1367 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1368 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1369 		return;
1370 	rw_wlock(&pvh_global_lock);
1371 	lo = moea_attr_fetch(m);
1372 	powerpc_sync();
1373 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1374 		pmap = pvo->pvo_pmap;
1375 		PMAP_LOCK(pmap);
1376 		if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1377 			pt = moea_pvo_to_pte(pvo, -1);
1378 			pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1379 			pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1380 			if (pt != NULL) {
1381 				moea_pte_synch(pt, &pvo->pvo_pte.pte);
1382 				lo |= pvo->pvo_pte.pte.pte_lo;
1383 				pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1384 				moea_pte_change(pt, &pvo->pvo_pte.pte,
1385 				    pvo->pvo_vaddr);
1386 				mtx_unlock(&moea_table_mutex);
1387 			}
1388 		}
1389 		PMAP_UNLOCK(pmap);
1390 	}
1391 	if ((lo & PTE_CHG) != 0) {
1392 		moea_attr_clear(m, PTE_CHG);
1393 		vm_page_dirty(m);
1394 	}
1395 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1396 	rw_wunlock(&pvh_global_lock);
1397 }
1398 
1399 /*
1400  *	moea_ts_referenced:
1401  *
1402  *	Return a count of reference bits for a page, clearing those bits.
1403  *	It is not necessary for every reference bit to be cleared, but it
1404  *	is necessary that 0 only be returned when there are truly no
1405  *	reference bits set.
1406  *
1407  *	XXX: The exact number of bits to check and clear is a matter that
1408  *	should be tested and standardized at some point in the future for
1409  *	optimal aging of shared pages.
1410  */
1411 int
1412 moea_ts_referenced(mmu_t mmu, vm_page_t m)
1413 {
1414 	int count;
1415 
1416 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1417 	    ("moea_ts_referenced: page %p is not managed", m));
1418 	rw_wlock(&pvh_global_lock);
1419 	count = moea_clear_bit(m, PTE_REF);
1420 	rw_wunlock(&pvh_global_lock);
1421 	return (count);
1422 }
1423 
1424 /*
1425  * Modify the WIMG settings of all mappings for a page.
1426  */
1427 void
1428 moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1429 {
1430 	struct	pvo_entry *pvo;
1431 	struct	pvo_head *pvo_head;
1432 	struct	pte *pt;
1433 	pmap_t	pmap;
1434 	u_int	lo;
1435 
1436 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1437 		m->md.mdpg_cache_attrs = ma;
1438 		return;
1439 	}
1440 
1441 	rw_wlock(&pvh_global_lock);
1442 	pvo_head = vm_page_to_pvoh(m);
1443 	lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1444 
1445 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1446 		pmap = pvo->pvo_pmap;
1447 		PMAP_LOCK(pmap);
1448 		pt = moea_pvo_to_pte(pvo, -1);
1449 		pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1450 		pvo->pvo_pte.pte.pte_lo |= lo;
1451 		if (pt != NULL) {
1452 			moea_pte_change(pt, &pvo->pvo_pte.pte,
1453 			    pvo->pvo_vaddr);
1454 			if (pvo->pvo_pmap == kernel_pmap)
1455 				isync();
1456 		}
1457 		mtx_unlock(&moea_table_mutex);
1458 		PMAP_UNLOCK(pmap);
1459 	}
1460 	m->md.mdpg_cache_attrs = ma;
1461 	rw_wunlock(&pvh_global_lock);
1462 }
1463 
1464 /*
1465  * Map a wired page into kernel virtual address space.
1466  */
1467 void
1468 moea_kenter(mmu_t mmu, vm_offset_t va, vm_paddr_t pa)
1469 {
1470 
1471 	moea_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1472 }
1473 
1474 void
1475 moea_kenter_attr(mmu_t mmu, vm_offset_t va, vm_offset_t pa, vm_memattr_t ma)
1476 {
1477 	u_int		pte_lo;
1478 	int		error;
1479 
1480 #if 0
1481 	if (va < VM_MIN_KERNEL_ADDRESS)
1482 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1483 		    va);
1484 #endif
1485 
1486 	pte_lo = moea_calc_wimg(pa, ma);
1487 
1488 	PMAP_LOCK(kernel_pmap);
1489 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1490 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1491 
1492 	if (error != 0 && error != ENOENT)
1493 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1494 		    pa, error);
1495 
1496 	PMAP_UNLOCK(kernel_pmap);
1497 }
1498 
1499 /*
1500  * Extract the physical page address associated with the given kernel virtual
1501  * address.
1502  */
1503 vm_paddr_t
1504 moea_kextract(mmu_t mmu, vm_offset_t va)
1505 {
1506 	struct		pvo_entry *pvo;
1507 	vm_paddr_t pa;
1508 
1509 	/*
1510 	 * Allow direct mappings on 32-bit OEA
1511 	 */
1512 	if (va < VM_MIN_KERNEL_ADDRESS) {
1513 		return (va);
1514 	}
1515 
1516 	PMAP_LOCK(kernel_pmap);
1517 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1518 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1519 	pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1520 	PMAP_UNLOCK(kernel_pmap);
1521 	return (pa);
1522 }
1523 
1524 /*
1525  * Remove a wired page from kernel virtual address space.
1526  */
1527 void
1528 moea_kremove(mmu_t mmu, vm_offset_t va)
1529 {
1530 
1531 	moea_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1532 }
1533 
1534 /*
1535  * Map a range of physical addresses into kernel virtual address space.
1536  *
1537  * The value passed in *virt is a suggested virtual address for the mapping.
1538  * Architectures which can support a direct-mapped physical to virtual region
1539  * can return the appropriate address within that region, leaving '*virt'
1540  * unchanged.  We cannot and therefore do not; *virt is updated with the
1541  * first usable address after the mapped region.
1542  */
1543 vm_offset_t
1544 moea_map(mmu_t mmu, vm_offset_t *virt, vm_paddr_t pa_start,
1545     vm_paddr_t pa_end, int prot)
1546 {
1547 	vm_offset_t	sva, va;
1548 
1549 	sva = *virt;
1550 	va = sva;
1551 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1552 		moea_kenter(mmu, va, pa_start);
1553 	*virt = va;
1554 	return (sva);
1555 }
1556 
1557 /*
1558  * Returns true if the pmap's pv is one of the first
1559  * 16 pvs linked to from this page.  This count may
1560  * be changed upwards or downwards in the future; it
1561  * is only necessary that true be returned for a small
1562  * subset of pmaps for proper page aging.
1563  */
1564 boolean_t
1565 moea_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1566 {
1567         int loops;
1568 	struct pvo_entry *pvo;
1569 	boolean_t rv;
1570 
1571 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1572 	    ("moea_page_exists_quick: page %p is not managed", m));
1573 	loops = 0;
1574 	rv = FALSE;
1575 	rw_wlock(&pvh_global_lock);
1576 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1577 		if (pvo->pvo_pmap == pmap) {
1578 			rv = TRUE;
1579 			break;
1580 		}
1581 		if (++loops >= 16)
1582 			break;
1583 	}
1584 	rw_wunlock(&pvh_global_lock);
1585 	return (rv);
1586 }
1587 
1588 /*
1589  * Return the number of managed mappings to the given physical page
1590  * that are wired.
1591  */
1592 int
1593 moea_page_wired_mappings(mmu_t mmu, vm_page_t m)
1594 {
1595 	struct pvo_entry *pvo;
1596 	int count;
1597 
1598 	count = 0;
1599 	if ((m->oflags & VPO_UNMANAGED) != 0)
1600 		return (count);
1601 	rw_wlock(&pvh_global_lock);
1602 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1603 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1604 			count++;
1605 	rw_wunlock(&pvh_global_lock);
1606 	return (count);
1607 }
1608 
1609 static u_int	moea_vsidcontext;
1610 
1611 void
1612 moea_pinit(mmu_t mmu, pmap_t pmap)
1613 {
1614 	int	i, mask;
1615 	u_int	entropy;
1616 
1617 	KASSERT((int)pmap < VM_MIN_KERNEL_ADDRESS, ("moea_pinit: virt pmap"));
1618 	RB_INIT(&pmap->pmap_pvo);
1619 
1620 	entropy = 0;
1621 	__asm __volatile("mftb %0" : "=r"(entropy));
1622 
1623 	if ((pmap->pmap_phys = (pmap_t)moea_kextract(mmu, (vm_offset_t)pmap))
1624 	    == NULL) {
1625 		pmap->pmap_phys = pmap;
1626 	}
1627 
1628 
1629 	mtx_lock(&moea_vsid_mutex);
1630 	/*
1631 	 * Allocate some segment registers for this pmap.
1632 	 */
1633 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1634 		u_int	hash, n;
1635 
1636 		/*
1637 		 * Create a new value by mutiplying by a prime and adding in
1638 		 * entropy from the timebase register.  This is to make the
1639 		 * VSID more random so that the PT hash function collides
1640 		 * less often.  (Note that the prime casues gcc to do shifts
1641 		 * instead of a multiply.)
1642 		 */
1643 		moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1644 		hash = moea_vsidcontext & (NPMAPS - 1);
1645 		if (hash == 0)		/* 0 is special, avoid it */
1646 			continue;
1647 		n = hash >> 5;
1648 		mask = 1 << (hash & (VSID_NBPW - 1));
1649 		hash = (moea_vsidcontext & 0xfffff);
1650 		if (moea_vsid_bitmap[n] & mask) {	/* collision? */
1651 			/* anything free in this bucket? */
1652 			if (moea_vsid_bitmap[n] == 0xffffffff) {
1653 				entropy = (moea_vsidcontext >> 20);
1654 				continue;
1655 			}
1656 			i = ffs(~moea_vsid_bitmap[n]) - 1;
1657 			mask = 1 << i;
1658 			hash &= 0xfffff & ~(VSID_NBPW - 1);
1659 			hash |= i;
1660 		}
1661 		KASSERT(!(moea_vsid_bitmap[n] & mask),
1662 		    ("Allocating in-use VSID group %#x\n", hash));
1663 		moea_vsid_bitmap[n] |= mask;
1664 		for (i = 0; i < 16; i++)
1665 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1666 		mtx_unlock(&moea_vsid_mutex);
1667 		return;
1668 	}
1669 
1670 	mtx_unlock(&moea_vsid_mutex);
1671 	panic("moea_pinit: out of segments");
1672 }
1673 
1674 /*
1675  * Initialize the pmap associated with process 0.
1676  */
1677 void
1678 moea_pinit0(mmu_t mmu, pmap_t pm)
1679 {
1680 
1681 	PMAP_LOCK_INIT(pm);
1682 	moea_pinit(mmu, pm);
1683 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1684 }
1685 
1686 /*
1687  * Set the physical protection on the specified range of this map as requested.
1688  */
1689 void
1690 moea_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1691     vm_prot_t prot)
1692 {
1693 	struct	pvo_entry *pvo, *tpvo, key;
1694 	struct	pte *pt;
1695 
1696 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1697 	    ("moea_protect: non current pmap"));
1698 
1699 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1700 		moea_remove(mmu, pm, sva, eva);
1701 		return;
1702 	}
1703 
1704 	rw_wlock(&pvh_global_lock);
1705 	PMAP_LOCK(pm);
1706 	key.pvo_vaddr = sva;
1707 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1708 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1709 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1710 
1711 		/*
1712 		 * Grab the PTE pointer before we diddle with the cached PTE
1713 		 * copy.
1714 		 */
1715 		pt = moea_pvo_to_pte(pvo, -1);
1716 		/*
1717 		 * Change the protection of the page.
1718 		 */
1719 		pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1720 		pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1721 
1722 		/*
1723 		 * If the PVO is in the page table, update that pte as well.
1724 		 */
1725 		if (pt != NULL) {
1726 			moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1727 			mtx_unlock(&moea_table_mutex);
1728 		}
1729 	}
1730 	rw_wunlock(&pvh_global_lock);
1731 	PMAP_UNLOCK(pm);
1732 }
1733 
1734 /*
1735  * Map a list of wired pages into kernel virtual address space.  This is
1736  * intended for temporary mappings which do not need page modification or
1737  * references recorded.  Existing mappings in the region are overwritten.
1738  */
1739 void
1740 moea_qenter(mmu_t mmu, vm_offset_t sva, vm_page_t *m, int count)
1741 {
1742 	vm_offset_t va;
1743 
1744 	va = sva;
1745 	while (count-- > 0) {
1746 		moea_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1747 		va += PAGE_SIZE;
1748 		m++;
1749 	}
1750 }
1751 
1752 /*
1753  * Remove page mappings from kernel virtual address space.  Intended for
1754  * temporary mappings entered by moea_qenter.
1755  */
1756 void
1757 moea_qremove(mmu_t mmu, vm_offset_t sva, int count)
1758 {
1759 	vm_offset_t va;
1760 
1761 	va = sva;
1762 	while (count-- > 0) {
1763 		moea_kremove(mmu, va);
1764 		va += PAGE_SIZE;
1765 	}
1766 }
1767 
1768 void
1769 moea_release(mmu_t mmu, pmap_t pmap)
1770 {
1771         int idx, mask;
1772 
1773 	/*
1774 	 * Free segment register's VSID
1775 	 */
1776         if (pmap->pm_sr[0] == 0)
1777                 panic("moea_release");
1778 
1779 	mtx_lock(&moea_vsid_mutex);
1780         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1781         mask = 1 << (idx % VSID_NBPW);
1782         idx /= VSID_NBPW;
1783         moea_vsid_bitmap[idx] &= ~mask;
1784 	mtx_unlock(&moea_vsid_mutex);
1785 }
1786 
1787 /*
1788  * Remove the given range of addresses from the specified map.
1789  */
1790 void
1791 moea_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1792 {
1793 	struct	pvo_entry *pvo, *tpvo, key;
1794 
1795 	rw_wlock(&pvh_global_lock);
1796 	PMAP_LOCK(pm);
1797 	key.pvo_vaddr = sva;
1798 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1799 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1800 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1801 		moea_pvo_remove(pvo, -1);
1802 	}
1803 	PMAP_UNLOCK(pm);
1804 	rw_wunlock(&pvh_global_lock);
1805 }
1806 
1807 /*
1808  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1809  * will reflect changes in pte's back to the vm_page.
1810  */
1811 void
1812 moea_remove_all(mmu_t mmu, vm_page_t m)
1813 {
1814 	struct  pvo_head *pvo_head;
1815 	struct	pvo_entry *pvo, *next_pvo;
1816 	pmap_t	pmap;
1817 
1818 	rw_wlock(&pvh_global_lock);
1819 	pvo_head = vm_page_to_pvoh(m);
1820 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1821 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1822 
1823 		pmap = pvo->pvo_pmap;
1824 		PMAP_LOCK(pmap);
1825 		moea_pvo_remove(pvo, -1);
1826 		PMAP_UNLOCK(pmap);
1827 	}
1828 	if ((m->aflags & PGA_WRITEABLE) && moea_query_bit(m, PTE_CHG)) {
1829 		moea_attr_clear(m, PTE_CHG);
1830 		vm_page_dirty(m);
1831 	}
1832 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1833 	rw_wunlock(&pvh_global_lock);
1834 }
1835 
1836 /*
1837  * Allocate a physical page of memory directly from the phys_avail map.
1838  * Can only be called from moea_bootstrap before avail start and end are
1839  * calculated.
1840  */
1841 static vm_offset_t
1842 moea_bootstrap_alloc(vm_size_t size, u_int align)
1843 {
1844 	vm_offset_t	s, e;
1845 	int		i, j;
1846 
1847 	size = round_page(size);
1848 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1849 		if (align != 0)
1850 			s = (phys_avail[i] + align - 1) & ~(align - 1);
1851 		else
1852 			s = phys_avail[i];
1853 		e = s + size;
1854 
1855 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1856 			continue;
1857 
1858 		if (s == phys_avail[i]) {
1859 			phys_avail[i] += size;
1860 		} else if (e == phys_avail[i + 1]) {
1861 			phys_avail[i + 1] -= size;
1862 		} else {
1863 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1864 				phys_avail[j] = phys_avail[j - 2];
1865 				phys_avail[j + 1] = phys_avail[j - 1];
1866 			}
1867 
1868 			phys_avail[i + 3] = phys_avail[i + 1];
1869 			phys_avail[i + 1] = s;
1870 			phys_avail[i + 2] = e;
1871 			phys_avail_count++;
1872 		}
1873 
1874 		return (s);
1875 	}
1876 	panic("moea_bootstrap_alloc: could not allocate memory");
1877 }
1878 
1879 static void
1880 moea_syncicache(vm_offset_t pa, vm_size_t len)
1881 {
1882 	__syncicache((void *)pa, len);
1883 }
1884 
1885 static int
1886 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1887     vm_offset_t va, vm_offset_t pa, u_int pte_lo, int flags)
1888 {
1889 	struct	pvo_entry *pvo;
1890 	u_int	sr;
1891 	int	first;
1892 	u_int	ptegidx;
1893 	int	i;
1894 	int     bootstrap;
1895 
1896 	moea_pvo_enter_calls++;
1897 	first = 0;
1898 	bootstrap = 0;
1899 
1900 	/*
1901 	 * Compute the PTE Group index.
1902 	 */
1903 	va &= ~ADDR_POFF;
1904 	sr = va_to_sr(pm->pm_sr, va);
1905 	ptegidx = va_to_pteg(sr, va);
1906 
1907 	/*
1908 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
1909 	 * there is a mapping.
1910 	 */
1911 	mtx_lock(&moea_table_mutex);
1912 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
1913 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
1914 			if ((pvo->pvo_pte.pte.pte_lo & PTE_RPGN) == pa &&
1915 			    (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
1916 			    (pte_lo & PTE_PP)) {
1917 				/*
1918 				 * The PTE is not changing.  Instead, this may
1919 				 * be a request to change the mapping's wired
1920 				 * attribute.
1921 				 */
1922 				mtx_unlock(&moea_table_mutex);
1923 				if ((flags & PVO_WIRED) != 0 &&
1924 				    (pvo->pvo_vaddr & PVO_WIRED) == 0) {
1925 					pvo->pvo_vaddr |= PVO_WIRED;
1926 					pm->pm_stats.wired_count++;
1927 				} else if ((flags & PVO_WIRED) == 0 &&
1928 				    (pvo->pvo_vaddr & PVO_WIRED) != 0) {
1929 					pvo->pvo_vaddr &= ~PVO_WIRED;
1930 					pm->pm_stats.wired_count--;
1931 				}
1932 				return (0);
1933 			}
1934 			moea_pvo_remove(pvo, -1);
1935 			break;
1936 		}
1937 	}
1938 
1939 	/*
1940 	 * If we aren't overwriting a mapping, try to allocate.
1941 	 */
1942 	if (moea_initialized) {
1943 		pvo = uma_zalloc(zone, M_NOWAIT);
1944 	} else {
1945 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
1946 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
1947 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
1948 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
1949 		}
1950 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
1951 		moea_bpvo_pool_index++;
1952 		bootstrap = 1;
1953 	}
1954 
1955 	if (pvo == NULL) {
1956 		mtx_unlock(&moea_table_mutex);
1957 		return (ENOMEM);
1958 	}
1959 
1960 	moea_pvo_entries++;
1961 	pvo->pvo_vaddr = va;
1962 	pvo->pvo_pmap = pm;
1963 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
1964 	pvo->pvo_vaddr &= ~ADDR_POFF;
1965 	if (flags & PVO_WIRED)
1966 		pvo->pvo_vaddr |= PVO_WIRED;
1967 	if (pvo_head != &moea_pvo_kunmanaged)
1968 		pvo->pvo_vaddr |= PVO_MANAGED;
1969 	if (bootstrap)
1970 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
1971 
1972 	moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo);
1973 
1974 	/*
1975 	 * Add to pmap list
1976 	 */
1977 	RB_INSERT(pvo_tree, &pm->pmap_pvo, pvo);
1978 
1979 	/*
1980 	 * Remember if the list was empty and therefore will be the first
1981 	 * item.
1982 	 */
1983 	if (LIST_FIRST(pvo_head) == NULL)
1984 		first = 1;
1985 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
1986 
1987 	if (pvo->pvo_vaddr & PVO_WIRED)
1988 		pm->pm_stats.wired_count++;
1989 	pm->pm_stats.resident_count++;
1990 
1991 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
1992 	KASSERT(i < 8, ("Invalid PTE index"));
1993 	if (i >= 0) {
1994 		PVO_PTEGIDX_SET(pvo, i);
1995 	} else {
1996 		panic("moea_pvo_enter: overflow");
1997 		moea_pte_overflow++;
1998 	}
1999 	mtx_unlock(&moea_table_mutex);
2000 
2001 	return (first ? ENOENT : 0);
2002 }
2003 
2004 static void
2005 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
2006 {
2007 	struct	pte *pt;
2008 
2009 	/*
2010 	 * If there is an active pte entry, we need to deactivate it (and
2011 	 * save the ref & cfg bits).
2012 	 */
2013 	pt = moea_pvo_to_pte(pvo, pteidx);
2014 	if (pt != NULL) {
2015 		moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
2016 		mtx_unlock(&moea_table_mutex);
2017 		PVO_PTEGIDX_CLR(pvo);
2018 	} else {
2019 		moea_pte_overflow--;
2020 	}
2021 
2022 	/*
2023 	 * Update our statistics.
2024 	 */
2025 	pvo->pvo_pmap->pm_stats.resident_count--;
2026 	if (pvo->pvo_vaddr & PVO_WIRED)
2027 		pvo->pvo_pmap->pm_stats.wired_count--;
2028 
2029 	/*
2030 	 * Save the REF/CHG bits into their cache if the page is managed.
2031 	 */
2032 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED) {
2033 		struct	vm_page *pg;
2034 
2035 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
2036 		if (pg != NULL) {
2037 			moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2038 			    (PTE_REF | PTE_CHG));
2039 		}
2040 	}
2041 
2042 	/*
2043 	 * Remove this PVO from the PV and pmap lists.
2044 	 */
2045 	LIST_REMOVE(pvo, pvo_vlink);
2046 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2047 
2048 	/*
2049 	 * Remove this from the overflow list and return it to the pool
2050 	 * if we aren't going to reuse it.
2051 	 */
2052 	LIST_REMOVE(pvo, pvo_olink);
2053 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2054 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
2055 		    moea_upvo_zone, pvo);
2056 	moea_pvo_entries--;
2057 	moea_pvo_remove_calls++;
2058 }
2059 
2060 static __inline int
2061 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2062 {
2063 	int	pteidx;
2064 
2065 	/*
2066 	 * We can find the actual pte entry without searching by grabbing
2067 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2068 	 * noticing the HID bit.
2069 	 */
2070 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2071 	if (pvo->pvo_pte.pte.pte_hi & PTE_HID)
2072 		pteidx ^= moea_pteg_mask * 8;
2073 
2074 	return (pteidx);
2075 }
2076 
2077 static struct pvo_entry *
2078 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2079 {
2080 	struct	pvo_entry *pvo;
2081 	int	ptegidx;
2082 	u_int	sr;
2083 
2084 	va &= ~ADDR_POFF;
2085 	sr = va_to_sr(pm->pm_sr, va);
2086 	ptegidx = va_to_pteg(sr, va);
2087 
2088 	mtx_lock(&moea_table_mutex);
2089 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2090 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2091 			if (pteidx_p)
2092 				*pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2093 			break;
2094 		}
2095 	}
2096 	mtx_unlock(&moea_table_mutex);
2097 
2098 	return (pvo);
2099 }
2100 
2101 static struct pte *
2102 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2103 {
2104 	struct	pte *pt;
2105 
2106 	/*
2107 	 * If we haven't been supplied the ptegidx, calculate it.
2108 	 */
2109 	if (pteidx == -1) {
2110 		int	ptegidx;
2111 		u_int	sr;
2112 
2113 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2114 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2115 		pteidx = moea_pvo_pte_index(pvo, ptegidx);
2116 	}
2117 
2118 	pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2119 	mtx_lock(&moea_table_mutex);
2120 
2121 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2122 		panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2123 		    "valid pte index", pvo);
2124 	}
2125 
2126 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2127 		panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2128 		    "pvo but no valid pte", pvo);
2129 	}
2130 
2131 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2132 		if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) {
2133 			panic("moea_pvo_to_pte: pvo %p has valid pte in "
2134 			    "moea_pteg_table %p but invalid in pvo", pvo, pt);
2135 		}
2136 
2137 		if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2138 		    != 0) {
2139 			panic("moea_pvo_to_pte: pvo %p pte does not match "
2140 			    "pte %p in moea_pteg_table", pvo, pt);
2141 		}
2142 
2143 		mtx_assert(&moea_table_mutex, MA_OWNED);
2144 		return (pt);
2145 	}
2146 
2147 	if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) {
2148 		panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2149 		    "moea_pteg_table but valid in pvo: %8x, %8x", pvo, pt, pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2150 	}
2151 
2152 	mtx_unlock(&moea_table_mutex);
2153 	return (NULL);
2154 }
2155 
2156 /*
2157  * XXX: THIS STUFF SHOULD BE IN pte.c?
2158  */
2159 int
2160 moea_pte_spill(vm_offset_t addr)
2161 {
2162 	struct	pvo_entry *source_pvo, *victim_pvo;
2163 	struct	pvo_entry *pvo;
2164 	int	ptegidx, i, j;
2165 	u_int	sr;
2166 	struct	pteg *pteg;
2167 	struct	pte *pt;
2168 
2169 	moea_pte_spills++;
2170 
2171 	sr = mfsrin(addr);
2172 	ptegidx = va_to_pteg(sr, addr);
2173 
2174 	/*
2175 	 * Have to substitute some entry.  Use the primary hash for this.
2176 	 * Use low bits of timebase as random generator.
2177 	 */
2178 	pteg = &moea_pteg_table[ptegidx];
2179 	mtx_lock(&moea_table_mutex);
2180 	__asm __volatile("mftb %0" : "=r"(i));
2181 	i &= 7;
2182 	pt = &pteg->pt[i];
2183 
2184 	source_pvo = NULL;
2185 	victim_pvo = NULL;
2186 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2187 		/*
2188 		 * We need to find a pvo entry for this address.
2189 		 */
2190 		if (source_pvo == NULL &&
2191 		    moea_pte_match(&pvo->pvo_pte.pte, sr, addr,
2192 		    pvo->pvo_pte.pte.pte_hi & PTE_HID)) {
2193 			/*
2194 			 * Now found an entry to be spilled into the pteg.
2195 			 * The PTE is now valid, so we know it's active.
2196 			 */
2197 			j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2198 
2199 			if (j >= 0) {
2200 				PVO_PTEGIDX_SET(pvo, j);
2201 				moea_pte_overflow--;
2202 				mtx_unlock(&moea_table_mutex);
2203 				return (1);
2204 			}
2205 
2206 			source_pvo = pvo;
2207 
2208 			if (victim_pvo != NULL)
2209 				break;
2210 		}
2211 
2212 		/*
2213 		 * We also need the pvo entry of the victim we are replacing
2214 		 * so save the R & C bits of the PTE.
2215 		 */
2216 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2217 		    moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2218 			victim_pvo = pvo;
2219 			if (source_pvo != NULL)
2220 				break;
2221 		}
2222 	}
2223 
2224 	if (source_pvo == NULL) {
2225 		mtx_unlock(&moea_table_mutex);
2226 		return (0);
2227 	}
2228 
2229 	if (victim_pvo == NULL) {
2230 		if ((pt->pte_hi & PTE_HID) == 0)
2231 			panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2232 			    "entry", pt);
2233 
2234 		/*
2235 		 * If this is a secondary PTE, we need to search it's primary
2236 		 * pvo bucket for the matching PVO.
2237 		 */
2238 		LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2239 		    pvo_olink) {
2240 			/*
2241 			 * We also need the pvo entry of the victim we are
2242 			 * replacing so save the R & C bits of the PTE.
2243 			 */
2244 			if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2245 				victim_pvo = pvo;
2246 				break;
2247 			}
2248 		}
2249 
2250 		if (victim_pvo == NULL)
2251 			panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2252 			    "entry", pt);
2253 	}
2254 
2255 	/*
2256 	 * We are invalidating the TLB entry for the EA we are replacing even
2257 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2258 	 * contained in the TLB entry.
2259 	 */
2260 	source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID;
2261 
2262 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2263 	moea_pte_set(pt, &source_pvo->pvo_pte.pte);
2264 
2265 	PVO_PTEGIDX_CLR(victim_pvo);
2266 	PVO_PTEGIDX_SET(source_pvo, i);
2267 	moea_pte_replacements++;
2268 
2269 	mtx_unlock(&moea_table_mutex);
2270 	return (1);
2271 }
2272 
2273 static __inline struct pvo_entry *
2274 moea_pte_spillable_ident(u_int ptegidx)
2275 {
2276 	struct	pte *pt;
2277 	struct	pvo_entry *pvo_walk, *pvo = NULL;
2278 
2279 	LIST_FOREACH(pvo_walk, &moea_pvo_table[ptegidx], pvo_olink) {
2280 		if (pvo_walk->pvo_vaddr & PVO_WIRED)
2281 			continue;
2282 
2283 		if (!(pvo_walk->pvo_pte.pte.pte_hi & PTE_VALID))
2284 			continue;
2285 
2286 		pt = moea_pvo_to_pte(pvo_walk, -1);
2287 
2288 		if (pt == NULL)
2289 			continue;
2290 
2291 		pvo = pvo_walk;
2292 
2293 		mtx_unlock(&moea_table_mutex);
2294 		if (!(pt->pte_lo & PTE_REF))
2295 			return (pvo_walk);
2296 	}
2297 
2298 	return (pvo);
2299 }
2300 
2301 static int
2302 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2303 {
2304 	struct	pte *pt;
2305 	struct	pvo_entry *victim_pvo;
2306 	int	i;
2307 	int	victim_idx;
2308 	u_int	pteg_bkpidx = ptegidx;
2309 
2310 	mtx_assert(&moea_table_mutex, MA_OWNED);
2311 
2312 	/*
2313 	 * First try primary hash.
2314 	 */
2315 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2316 		if ((pt->pte_hi & PTE_VALID) == 0) {
2317 			pvo_pt->pte_hi &= ~PTE_HID;
2318 			moea_pte_set(pt, pvo_pt);
2319 			return (i);
2320 		}
2321 	}
2322 
2323 	/*
2324 	 * Now try secondary hash.
2325 	 */
2326 	ptegidx ^= moea_pteg_mask;
2327 
2328 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2329 		if ((pt->pte_hi & PTE_VALID) == 0) {
2330 			pvo_pt->pte_hi |= PTE_HID;
2331 			moea_pte_set(pt, pvo_pt);
2332 			return (i);
2333 		}
2334 	}
2335 
2336 	/* Try again, but this time try to force a PTE out. */
2337 	ptegidx = pteg_bkpidx;
2338 
2339 	victim_pvo = moea_pte_spillable_ident(ptegidx);
2340 	if (victim_pvo == NULL) {
2341 		ptegidx ^= moea_pteg_mask;
2342 		victim_pvo = moea_pte_spillable_ident(ptegidx);
2343 	}
2344 
2345 	if (victim_pvo == NULL) {
2346 		panic("moea_pte_insert: overflow");
2347 		return (-1);
2348 	}
2349 
2350 	victim_idx = moea_pvo_pte_index(victim_pvo, ptegidx);
2351 
2352 	if (pteg_bkpidx == ptegidx)
2353 		pvo_pt->pte_hi &= ~PTE_HID;
2354 	else
2355 		pvo_pt->pte_hi |= PTE_HID;
2356 
2357 	/*
2358 	 * Synchronize the sacrifice PTE with its PVO, then mark both
2359 	 * invalid. The PVO will be reused when/if the VM system comes
2360 	 * here after a fault.
2361 	 */
2362 	pt = &moea_pteg_table[victim_idx >> 3].pt[victim_idx & 7];
2363 
2364 	if (pt->pte_hi != victim_pvo->pvo_pte.pte.pte_hi)
2365 	    panic("Victim PVO doesn't match PTE! PVO: %8x, PTE: %8x", victim_pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2366 
2367 	/*
2368 	 * Set the new PTE.
2369 	 */
2370 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2371 	PVO_PTEGIDX_CLR(victim_pvo);
2372 	moea_pte_overflow++;
2373 	moea_pte_set(pt, pvo_pt);
2374 
2375 	return (victim_idx & 7);
2376 }
2377 
2378 static boolean_t
2379 moea_query_bit(vm_page_t m, int ptebit)
2380 {
2381 	struct	pvo_entry *pvo;
2382 	struct	pte *pt;
2383 
2384 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2385 	if (moea_attr_fetch(m) & ptebit)
2386 		return (TRUE);
2387 
2388 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2389 
2390 		/*
2391 		 * See if we saved the bit off.  If so, cache it and return
2392 		 * success.
2393 		 */
2394 		if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2395 			moea_attr_save(m, ptebit);
2396 			return (TRUE);
2397 		}
2398 	}
2399 
2400 	/*
2401 	 * No luck, now go through the hard part of looking at the PTEs
2402 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2403 	 * the PTEs.
2404 	 */
2405 	powerpc_sync();
2406 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2407 
2408 		/*
2409 		 * See if this pvo has a valid PTE.  if so, fetch the
2410 		 * REF/CHG bits from the valid PTE.  If the appropriate
2411 		 * ptebit is set, cache it and return success.
2412 		 */
2413 		pt = moea_pvo_to_pte(pvo, -1);
2414 		if (pt != NULL) {
2415 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2416 			mtx_unlock(&moea_table_mutex);
2417 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2418 				moea_attr_save(m, ptebit);
2419 				return (TRUE);
2420 			}
2421 		}
2422 	}
2423 
2424 	return (FALSE);
2425 }
2426 
2427 static u_int
2428 moea_clear_bit(vm_page_t m, int ptebit)
2429 {
2430 	u_int	count;
2431 	struct	pvo_entry *pvo;
2432 	struct	pte *pt;
2433 
2434 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2435 
2436 	/*
2437 	 * Clear the cached value.
2438 	 */
2439 	moea_attr_clear(m, ptebit);
2440 
2441 	/*
2442 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2443 	 * we can reset the right ones).  note that since the pvo entries and
2444 	 * list heads are accessed via BAT0 and are never placed in the page
2445 	 * table, we don't have to worry about further accesses setting the
2446 	 * REF/CHG bits.
2447 	 */
2448 	powerpc_sync();
2449 
2450 	/*
2451 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2452 	 * valid pte clear the ptebit from the valid pte.
2453 	 */
2454 	count = 0;
2455 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2456 		pt = moea_pvo_to_pte(pvo, -1);
2457 		if (pt != NULL) {
2458 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2459 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2460 				count++;
2461 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2462 			}
2463 			mtx_unlock(&moea_table_mutex);
2464 		}
2465 		pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2466 	}
2467 
2468 	return (count);
2469 }
2470 
2471 /*
2472  * Return true if the physical range is encompassed by the battable[idx]
2473  */
2474 static int
2475 moea_bat_mapped(int idx, vm_offset_t pa, vm_size_t size)
2476 {
2477 	u_int prot;
2478 	u_int32_t start;
2479 	u_int32_t end;
2480 	u_int32_t bat_ble;
2481 
2482 	/*
2483 	 * Return immediately if not a valid mapping
2484 	 */
2485 	if (!(battable[idx].batu & BAT_Vs))
2486 		return (EINVAL);
2487 
2488 	/*
2489 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2490 	 * so it can function as an i/o page
2491 	 */
2492 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2493 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2494 		return (EPERM);
2495 
2496 	/*
2497 	 * The address should be within the BAT range. Assume that the
2498 	 * start address in the BAT has the correct alignment (thus
2499 	 * not requiring masking)
2500 	 */
2501 	start = battable[idx].batl & BAT_PBS;
2502 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2503 	end = start | (bat_ble << 15) | 0x7fff;
2504 
2505 	if ((pa < start) || ((pa + size) > end))
2506 		return (ERANGE);
2507 
2508 	return (0);
2509 }
2510 
2511 boolean_t
2512 moea_dev_direct_mapped(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2513 {
2514 	int i;
2515 
2516 	/*
2517 	 * This currently does not work for entries that
2518 	 * overlap 256M BAT segments.
2519 	 */
2520 
2521 	for(i = 0; i < 16; i++)
2522 		if (moea_bat_mapped(i, pa, size) == 0)
2523 			return (0);
2524 
2525 	return (EFAULT);
2526 }
2527 
2528 /*
2529  * Map a set of physical memory pages into the kernel virtual
2530  * address space. Return a pointer to where it is mapped. This
2531  * routine is intended to be used for mapping device memory,
2532  * NOT real memory.
2533  */
2534 void *
2535 moea_mapdev(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2536 {
2537 
2538 	return (moea_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT));
2539 }
2540 
2541 void *
2542 moea_mapdev_attr(mmu_t mmu, vm_offset_t pa, vm_size_t size, vm_memattr_t ma)
2543 {
2544 	vm_offset_t va, tmpva, ppa, offset;
2545 	int i;
2546 
2547 	ppa = trunc_page(pa);
2548 	offset = pa & PAGE_MASK;
2549 	size = roundup(offset + size, PAGE_SIZE);
2550 
2551 	/*
2552 	 * If the physical address lies within a valid BAT table entry,
2553 	 * return the 1:1 mapping. This currently doesn't work
2554 	 * for regions that overlap 256M BAT segments.
2555 	 */
2556 	for (i = 0; i < 16; i++) {
2557 		if (moea_bat_mapped(i, pa, size) == 0)
2558 			return ((void *) pa);
2559 	}
2560 
2561 	va = kva_alloc(size);
2562 	if (!va)
2563 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2564 
2565 	for (tmpva = va; size > 0;) {
2566 		moea_kenter_attr(mmu, tmpva, ppa, ma);
2567 		tlbie(tmpva);
2568 		size -= PAGE_SIZE;
2569 		tmpva += PAGE_SIZE;
2570 		ppa += PAGE_SIZE;
2571 	}
2572 
2573 	return ((void *)(va + offset));
2574 }
2575 
2576 void
2577 moea_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2578 {
2579 	vm_offset_t base, offset;
2580 
2581 	/*
2582 	 * If this is outside kernel virtual space, then it's a
2583 	 * battable entry and doesn't require unmapping
2584 	 */
2585 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2586 		base = trunc_page(va);
2587 		offset = va & PAGE_MASK;
2588 		size = roundup(offset + size, PAGE_SIZE);
2589 		kva_free(base, size);
2590 	}
2591 }
2592 
2593 static void
2594 moea_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2595 {
2596 	struct pvo_entry *pvo;
2597 	vm_offset_t lim;
2598 	vm_paddr_t pa;
2599 	vm_size_t len;
2600 
2601 	PMAP_LOCK(pm);
2602 	while (sz > 0) {
2603 		lim = round_page(va);
2604 		len = MIN(lim - va, sz);
2605 		pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2606 		if (pvo != NULL) {
2607 			pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) |
2608 			    (va & ADDR_POFF);
2609 			moea_syncicache(pa, len);
2610 		}
2611 		va += len;
2612 		sz -= len;
2613 	}
2614 	PMAP_UNLOCK(pm);
2615 }
2616 
2617 void
2618 moea_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va)
2619 {
2620 
2621 	*va = (void *)pa;
2622 }
2623 
2624 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
2625 
2626 void
2627 moea_scan_init(mmu_t mmu)
2628 {
2629 	struct pvo_entry *pvo;
2630 	vm_offset_t va;
2631 	int i;
2632 
2633 	if (!do_minidump) {
2634 		/* Initialize phys. segments for dumpsys(). */
2635 		memset(&dump_map, 0, sizeof(dump_map));
2636 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
2637 		for (i = 0; i < pregions_sz; i++) {
2638 			dump_map[i].pa_start = pregions[i].mr_start;
2639 			dump_map[i].pa_size = pregions[i].mr_size;
2640 		}
2641 		return;
2642 	}
2643 
2644 	/* Virtual segments for minidumps: */
2645 	memset(&dump_map, 0, sizeof(dump_map));
2646 
2647 	/* 1st: kernel .data and .bss. */
2648 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
2649 	dump_map[0].pa_size =
2650 	    round_page((uintptr_t)_end) - dump_map[0].pa_start;
2651 
2652 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2653 	dump_map[1].pa_start = (vm_paddr_t)msgbufp->msg_ptr;
2654 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
2655 
2656 	/* 3rd: kernel VM. */
2657 	va = dump_map[1].pa_start + dump_map[1].pa_size;
2658 	/* Find start of next chunk (from va). */
2659 	while (va < virtual_end) {
2660 		/* Don't dump the buffer cache. */
2661 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
2662 			va = kmi.buffer_eva;
2663 			continue;
2664 		}
2665 		pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
2666 		if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2667 			break;
2668 		va += PAGE_SIZE;
2669 	}
2670 	if (va < virtual_end) {
2671 		dump_map[2].pa_start = va;
2672 		va += PAGE_SIZE;
2673 		/* Find last page in chunk. */
2674 		while (va < virtual_end) {
2675 			/* Don't run into the buffer cache. */
2676 			if (va == kmi.buffer_sva)
2677 				break;
2678 			pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF,
2679 			    NULL);
2680 			if (pvo == NULL ||
2681 			    !(pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2682 				break;
2683 			va += PAGE_SIZE;
2684 		}
2685 		dump_map[2].pa_size = va - dump_map[2].pa_start;
2686 	}
2687 }
2688