xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision a5570ae0d780680c78ea6f90e101ead47ac11417)
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/queue.h>
110 #include <sys/cpuset.h>
111 #include <sys/ktr.h>
112 #include <sys/lock.h>
113 #include <sys/msgbuf.h>
114 #include <sys/mutex.h>
115 #include <sys/proc.h>
116 #include <sys/rwlock.h>
117 #include <sys/sched.h>
118 #include <sys/sysctl.h>
119 #include <sys/systm.h>
120 #include <sys/vmmeter.h>
121 
122 #include <dev/ofw/openfirm.h>
123 
124 #include <vm/vm.h>
125 #include <vm/vm_param.h>
126 #include <vm/vm_kern.h>
127 #include <vm/vm_page.h>
128 #include <vm/vm_map.h>
129 #include <vm/vm_object.h>
130 #include <vm/vm_extern.h>
131 #include <vm/vm_pageout.h>
132 #include <vm/uma.h>
133 
134 #include <machine/cpu.h>
135 #include <machine/platform.h>
136 #include <machine/bat.h>
137 #include <machine/frame.h>
138 #include <machine/md_var.h>
139 #include <machine/psl.h>
140 #include <machine/pte.h>
141 #include <machine/smp.h>
142 #include <machine/sr.h>
143 #include <machine/mmuvar.h>
144 #include <machine/trap.h>
145 
146 #include "mmu_if.h"
147 
148 #define	MOEA_DEBUG
149 
150 #define TODO	panic("%s: not implemented", __func__);
151 
152 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
153 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
154 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
155 
156 struct ofw_map {
157 	vm_offset_t	om_va;
158 	vm_size_t	om_len;
159 	vm_offset_t	om_pa;
160 	u_int		om_mode;
161 };
162 
163 extern unsigned char _etext[];
164 extern unsigned char _end[];
165 
166 extern int dumpsys_minidump;
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 vm_offset_t moea_dumpsys_map(mmu_t mmu, struct pmap_md *md, vm_size_t ofs,
318     vm_size_t *sz);
319 struct pmap_md * moea_scan_md(mmu_t mmu, struct pmap_md *prev);
320 
321 static mmu_method_t moea_methods[] = {
322 	MMUMETHOD(mmu_clear_modify,	moea_clear_modify),
323 	MMUMETHOD(mmu_copy_page,	moea_copy_page),
324 	MMUMETHOD(mmu_copy_pages,	moea_copy_pages),
325 	MMUMETHOD(mmu_enter,		moea_enter),
326 	MMUMETHOD(mmu_enter_object,	moea_enter_object),
327 	MMUMETHOD(mmu_enter_quick,	moea_enter_quick),
328 	MMUMETHOD(mmu_extract,		moea_extract),
329 	MMUMETHOD(mmu_extract_and_hold,	moea_extract_and_hold),
330 	MMUMETHOD(mmu_init,		moea_init),
331 	MMUMETHOD(mmu_is_modified,	moea_is_modified),
332 	MMUMETHOD(mmu_is_prefaultable,	moea_is_prefaultable),
333 	MMUMETHOD(mmu_is_referenced,	moea_is_referenced),
334 	MMUMETHOD(mmu_ts_referenced,	moea_ts_referenced),
335 	MMUMETHOD(mmu_map,     		moea_map),
336 	MMUMETHOD(mmu_page_exists_quick,moea_page_exists_quick),
337 	MMUMETHOD(mmu_page_wired_mappings,moea_page_wired_mappings),
338 	MMUMETHOD(mmu_pinit,		moea_pinit),
339 	MMUMETHOD(mmu_pinit0,		moea_pinit0),
340 	MMUMETHOD(mmu_protect,		moea_protect),
341 	MMUMETHOD(mmu_qenter,		moea_qenter),
342 	MMUMETHOD(mmu_qremove,		moea_qremove),
343 	MMUMETHOD(mmu_release,		moea_release),
344 	MMUMETHOD(mmu_remove,		moea_remove),
345 	MMUMETHOD(mmu_remove_all,      	moea_remove_all),
346 	MMUMETHOD(mmu_remove_write,	moea_remove_write),
347 	MMUMETHOD(mmu_sync_icache,	moea_sync_icache),
348 	MMUMETHOD(mmu_unwire,		moea_unwire),
349 	MMUMETHOD(mmu_zero_page,       	moea_zero_page),
350 	MMUMETHOD(mmu_zero_page_area,	moea_zero_page_area),
351 	MMUMETHOD(mmu_zero_page_idle,	moea_zero_page_idle),
352 	MMUMETHOD(mmu_activate,		moea_activate),
353 	MMUMETHOD(mmu_deactivate,      	moea_deactivate),
354 	MMUMETHOD(mmu_page_set_memattr,	moea_page_set_memattr),
355 
356 	/* Internal interfaces */
357 	MMUMETHOD(mmu_bootstrap,       	moea_bootstrap),
358 	MMUMETHOD(mmu_cpu_bootstrap,   	moea_cpu_bootstrap),
359 	MMUMETHOD(mmu_mapdev_attr,	moea_mapdev_attr),
360 	MMUMETHOD(mmu_mapdev,		moea_mapdev),
361 	MMUMETHOD(mmu_unmapdev,		moea_unmapdev),
362 	MMUMETHOD(mmu_kextract,		moea_kextract),
363 	MMUMETHOD(mmu_kenter,		moea_kenter),
364 	MMUMETHOD(mmu_kenter_attr,	moea_kenter_attr),
365 	MMUMETHOD(mmu_dev_direct_mapped,moea_dev_direct_mapped),
366 	MMUMETHOD(mmu_scan_md,		moea_scan_md),
367 	MMUMETHOD(mmu_dumpsys_map,	moea_dumpsys_map),
368 
369 	{ 0, 0 }
370 };
371 
372 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods, 0);
373 
374 static __inline uint32_t
375 moea_calc_wimg(vm_offset_t pa, vm_memattr_t ma)
376 {
377 	uint32_t pte_lo;
378 	int i;
379 
380 	if (ma != VM_MEMATTR_DEFAULT) {
381 		switch (ma) {
382 		case VM_MEMATTR_UNCACHEABLE:
383 			return (PTE_I | PTE_G);
384 		case VM_MEMATTR_WRITE_COMBINING:
385 		case VM_MEMATTR_WRITE_BACK:
386 		case VM_MEMATTR_PREFETCHABLE:
387 			return (PTE_I);
388 		case VM_MEMATTR_WRITE_THROUGH:
389 			return (PTE_W | PTE_M);
390 		}
391 	}
392 
393 	/*
394 	 * Assume the page is cache inhibited and access is guarded unless
395 	 * it's in our available memory array.
396 	 */
397 	pte_lo = PTE_I | PTE_G;
398 	for (i = 0; i < pregions_sz; i++) {
399 		if ((pa >= pregions[i].mr_start) &&
400 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
401 			pte_lo = PTE_M;
402 			break;
403 		}
404 	}
405 
406 	return pte_lo;
407 }
408 
409 static void
410 tlbie(vm_offset_t va)
411 {
412 
413 	mtx_lock_spin(&tlbie_mtx);
414 	__asm __volatile("ptesync");
415 	__asm __volatile("tlbie %0" :: "r"(va));
416 	__asm __volatile("eieio; tlbsync; ptesync");
417 	mtx_unlock_spin(&tlbie_mtx);
418 }
419 
420 static void
421 tlbia(void)
422 {
423 	vm_offset_t va;
424 
425 	for (va = 0; va < 0x00040000; va += 0x00001000) {
426 		__asm __volatile("tlbie %0" :: "r"(va));
427 		powerpc_sync();
428 	}
429 	__asm __volatile("tlbsync");
430 	powerpc_sync();
431 }
432 
433 static __inline int
434 va_to_sr(u_int *sr, vm_offset_t va)
435 {
436 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
437 }
438 
439 static __inline u_int
440 va_to_pteg(u_int sr, vm_offset_t addr)
441 {
442 	u_int hash;
443 
444 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
445 	    ADDR_PIDX_SHFT);
446 	return (hash & moea_pteg_mask);
447 }
448 
449 static __inline struct pvo_head *
450 vm_page_to_pvoh(vm_page_t m)
451 {
452 
453 	return (&m->md.mdpg_pvoh);
454 }
455 
456 static __inline void
457 moea_attr_clear(vm_page_t m, int ptebit)
458 {
459 
460 	rw_assert(&pvh_global_lock, RA_WLOCKED);
461 	m->md.mdpg_attrs &= ~ptebit;
462 }
463 
464 static __inline int
465 moea_attr_fetch(vm_page_t m)
466 {
467 
468 	return (m->md.mdpg_attrs);
469 }
470 
471 static __inline void
472 moea_attr_save(vm_page_t m, int ptebit)
473 {
474 
475 	rw_assert(&pvh_global_lock, RA_WLOCKED);
476 	m->md.mdpg_attrs |= ptebit;
477 }
478 
479 static __inline int
480 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
481 {
482 	if (pt->pte_hi == pvo_pt->pte_hi)
483 		return (1);
484 
485 	return (0);
486 }
487 
488 static __inline int
489 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
490 {
491 	return (pt->pte_hi & ~PTE_VALID) ==
492 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
493 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
494 }
495 
496 static __inline void
497 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
498 {
499 
500 	mtx_assert(&moea_table_mutex, MA_OWNED);
501 
502 	/*
503 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
504 	 * set when the real pte is set in memory.
505 	 *
506 	 * Note: Don't set the valid bit for correct operation of tlb update.
507 	 */
508 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
509 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
510 	pt->pte_lo = pte_lo;
511 }
512 
513 static __inline void
514 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
515 {
516 
517 	mtx_assert(&moea_table_mutex, MA_OWNED);
518 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
519 }
520 
521 static __inline void
522 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
523 {
524 
525 	mtx_assert(&moea_table_mutex, MA_OWNED);
526 
527 	/*
528 	 * As shown in Section 7.6.3.2.3
529 	 */
530 	pt->pte_lo &= ~ptebit;
531 	tlbie(va);
532 }
533 
534 static __inline void
535 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
536 {
537 
538 	mtx_assert(&moea_table_mutex, MA_OWNED);
539 	pvo_pt->pte_hi |= PTE_VALID;
540 
541 	/*
542 	 * Update the PTE as defined in section 7.6.3.1.
543 	 * Note that the REF/CHG bits are from pvo_pt and thus should have
544 	 * been saved so this routine can restore them (if desired).
545 	 */
546 	pt->pte_lo = pvo_pt->pte_lo;
547 	powerpc_sync();
548 	pt->pte_hi = pvo_pt->pte_hi;
549 	powerpc_sync();
550 	moea_pte_valid++;
551 }
552 
553 static __inline void
554 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
555 {
556 
557 	mtx_assert(&moea_table_mutex, MA_OWNED);
558 	pvo_pt->pte_hi &= ~PTE_VALID;
559 
560 	/*
561 	 * Force the reg & chg bits back into the PTEs.
562 	 */
563 	powerpc_sync();
564 
565 	/*
566 	 * Invalidate the pte.
567 	 */
568 	pt->pte_hi &= ~PTE_VALID;
569 
570 	tlbie(va);
571 
572 	/*
573 	 * Save the reg & chg bits.
574 	 */
575 	moea_pte_synch(pt, pvo_pt);
576 	moea_pte_valid--;
577 }
578 
579 static __inline void
580 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
581 {
582 
583 	/*
584 	 * Invalidate the PTE
585 	 */
586 	moea_pte_unset(pt, pvo_pt, va);
587 	moea_pte_set(pt, pvo_pt);
588 }
589 
590 /*
591  * Quick sort callout for comparing memory regions.
592  */
593 static int	om_cmp(const void *a, const void *b);
594 
595 static int
596 om_cmp(const void *a, const void *b)
597 {
598 	const struct	ofw_map *mapa;
599 	const struct	ofw_map *mapb;
600 
601 	mapa = a;
602 	mapb = b;
603 	if (mapa->om_pa < mapb->om_pa)
604 		return (-1);
605 	else if (mapa->om_pa > mapb->om_pa)
606 		return (1);
607 	else
608 		return (0);
609 }
610 
611 void
612 moea_cpu_bootstrap(mmu_t mmup, int ap)
613 {
614 	u_int sdr;
615 	int i;
616 
617 	if (ap) {
618 		powerpc_sync();
619 		__asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu));
620 		__asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl));
621 		isync();
622 		__asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu));
623 		__asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl));
624 		isync();
625 	}
626 
627 #ifdef WII
628 	/*
629 	 * Special case for the Wii: don't install the PCI BAT.
630 	 */
631 	if (strcmp(installed_platform(), "wii") != 0) {
632 #endif
633 		__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
634 		__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
635 #ifdef WII
636 	}
637 #endif
638 	isync();
639 
640 	__asm __volatile("mtibatu 1,%0" :: "r"(0));
641 	__asm __volatile("mtdbatu 2,%0" :: "r"(0));
642 	__asm __volatile("mtibatu 2,%0" :: "r"(0));
643 	__asm __volatile("mtdbatu 3,%0" :: "r"(0));
644 	__asm __volatile("mtibatu 3,%0" :: "r"(0));
645 	isync();
646 
647 	for (i = 0; i < 16; i++)
648 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
649 	powerpc_sync();
650 
651 	sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10);
652 	__asm __volatile("mtsdr1 %0" :: "r"(sdr));
653 	isync();
654 
655 	tlbia();
656 }
657 
658 void
659 moea_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
660 {
661 	ihandle_t	mmui;
662 	phandle_t	chosen, mmu;
663 	int		sz;
664 	int		i, j;
665 	vm_size_t	size, physsz, hwphyssz;
666 	vm_offset_t	pa, va, off;
667 	void		*dpcpu;
668 	register_t	msr;
669 
670         /*
671          * Set up BAT0 to map the lowest 256 MB area
672          */
673         battable[0x0].batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
674         battable[0x0].batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
675 
676 	/*
677 	 * Map PCI memory space.
678 	 */
679 	battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
680 	battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
681 
682 	battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
683 	battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
684 
685 	battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
686 	battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
687 
688 	battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
689 	battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
690 
691 	/*
692 	 * Map obio devices.
693 	 */
694 	battable[0xf].batl = BATL(0xf0000000, BAT_I|BAT_G, BAT_PP_RW);
695 	battable[0xf].batu = BATU(0xf0000000, BAT_BL_256M, BAT_Vs);
696 
697 	/*
698 	 * Use an IBAT and a DBAT to map the bottom segment of memory
699 	 * where we are. Turn off instruction relocation temporarily
700 	 * to prevent faults while reprogramming the IBAT.
701 	 */
702 	msr = mfmsr();
703 	mtmsr(msr & ~PSL_IR);
704 	__asm (".balign 32; \n"
705 	       "mtibatu 0,%0; mtibatl 0,%1; isync; \n"
706 	       "mtdbatu 0,%0; mtdbatl 0,%1; isync"
707 	    :: "r"(battable[0].batu), "r"(battable[0].batl));
708 	mtmsr(msr);
709 
710 #ifdef WII
711         if (strcmp(installed_platform(), "wii") != 0) {
712 #endif
713 		/* map pci space */
714 		__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
715 		__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
716 #ifdef WII
717 	}
718 #endif
719 	isync();
720 
721 	/* set global direct map flag */
722 	hw_direct_map = 1;
723 
724 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
725 	CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
726 
727 	for (i = 0; i < pregions_sz; i++) {
728 		vm_offset_t pa;
729 		vm_offset_t end;
730 
731 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
732 			pregions[i].mr_start,
733 			pregions[i].mr_start + pregions[i].mr_size,
734 			pregions[i].mr_size);
735 		/*
736 		 * Install entries into the BAT table to allow all
737 		 * of physmem to be convered by on-demand BAT entries.
738 		 * The loop will sometimes set the same battable element
739 		 * twice, but that's fine since they won't be used for
740 		 * a while yet.
741 		 */
742 		pa = pregions[i].mr_start & 0xf0000000;
743 		end = pregions[i].mr_start + pregions[i].mr_size;
744 		do {
745                         u_int n = pa >> ADDR_SR_SHFT;
746 
747 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
748 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
749 			pa += SEGMENT_LENGTH;
750 		} while (pa < end);
751 	}
752 
753 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
754 		panic("moea_bootstrap: phys_avail too small");
755 
756 	phys_avail_count = 0;
757 	physsz = 0;
758 	hwphyssz = 0;
759 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
760 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
761 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
762 		    regions[i].mr_start + regions[i].mr_size,
763 		    regions[i].mr_size);
764 		if (hwphyssz != 0 &&
765 		    (physsz + regions[i].mr_size) >= hwphyssz) {
766 			if (physsz < hwphyssz) {
767 				phys_avail[j] = regions[i].mr_start;
768 				phys_avail[j + 1] = regions[i].mr_start +
769 				    hwphyssz - physsz;
770 				physsz = hwphyssz;
771 				phys_avail_count++;
772 			}
773 			break;
774 		}
775 		phys_avail[j] = regions[i].mr_start;
776 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
777 		phys_avail_count++;
778 		physsz += regions[i].mr_size;
779 	}
780 
781 	/* Check for overlap with the kernel and exception vectors */
782 	for (j = 0; j < 2*phys_avail_count; j+=2) {
783 		if (phys_avail[j] < EXC_LAST)
784 			phys_avail[j] += EXC_LAST;
785 
786 		if (kernelstart >= phys_avail[j] &&
787 		    kernelstart < phys_avail[j+1]) {
788 			if (kernelend < phys_avail[j+1]) {
789 				phys_avail[2*phys_avail_count] =
790 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
791 				phys_avail[2*phys_avail_count + 1] =
792 				    phys_avail[j+1];
793 				phys_avail_count++;
794 			}
795 
796 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
797 		}
798 
799 		if (kernelend >= phys_avail[j] &&
800 		    kernelend < phys_avail[j+1]) {
801 			if (kernelstart > phys_avail[j]) {
802 				phys_avail[2*phys_avail_count] = phys_avail[j];
803 				phys_avail[2*phys_avail_count + 1] =
804 				    kernelstart & ~PAGE_MASK;
805 				phys_avail_count++;
806 			}
807 
808 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
809 		}
810 	}
811 
812 	physmem = btoc(physsz);
813 
814 	/*
815 	 * Allocate PTEG table.
816 	 */
817 #ifdef PTEGCOUNT
818 	moea_pteg_count = PTEGCOUNT;
819 #else
820 	moea_pteg_count = 0x1000;
821 
822 	while (moea_pteg_count < physmem)
823 		moea_pteg_count <<= 1;
824 
825 	moea_pteg_count >>= 1;
826 #endif /* PTEGCOUNT */
827 
828 	size = moea_pteg_count * sizeof(struct pteg);
829 	CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
830 	    size);
831 	moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
832 	CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
833 	bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
834 	moea_pteg_mask = moea_pteg_count - 1;
835 
836 	/*
837 	 * Allocate pv/overflow lists.
838 	 */
839 	size = sizeof(struct pvo_head) * moea_pteg_count;
840 	moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
841 	    PAGE_SIZE);
842 	CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
843 	for (i = 0; i < moea_pteg_count; i++)
844 		LIST_INIT(&moea_pvo_table[i]);
845 
846 	/*
847 	 * Initialize the lock that synchronizes access to the pteg and pvo
848 	 * tables.
849 	 */
850 	mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF |
851 	    MTX_RECURSE);
852 	mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF);
853 
854 	mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN);
855 
856 	/*
857 	 * Initialise the unmanaged pvo pool.
858 	 */
859 	moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
860 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
861 	moea_bpvo_pool_index = 0;
862 
863 	/*
864 	 * Make sure kernel vsid is allocated as well as VSID 0.
865 	 */
866 	moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
867 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
868 	moea_vsid_bitmap[0] |= 1;
869 
870 	/*
871 	 * Initialize the kernel pmap (which is statically allocated).
872 	 */
873 	PMAP_LOCK_INIT(kernel_pmap);
874 	for (i = 0; i < 16; i++)
875 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
876 	CPU_FILL(&kernel_pmap->pm_active);
877 	RB_INIT(&kernel_pmap->pmap_pvo);
878 
879  	/*
880 	 * Initialize the global pv list lock.
881 	 */
882 	rw_init(&pvh_global_lock, "pmap pv global");
883 
884 	/*
885 	 * Set up the Open Firmware mappings
886 	 */
887 	chosen = OF_finddevice("/chosen");
888 	if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1 &&
889 	    (mmu = OF_instance_to_package(mmui)) != -1 &&
890 	    (sz = OF_getproplen(mmu, "translations")) != -1) {
891 		translations = NULL;
892 		for (i = 0; phys_avail[i] != 0; i += 2) {
893 			if (phys_avail[i + 1] >= sz) {
894 				translations = (struct ofw_map *)phys_avail[i];
895 				break;
896 			}
897 		}
898 		if (translations == NULL)
899 			panic("moea_bootstrap: no space to copy translations");
900 		bzero(translations, sz);
901 		if (OF_getprop(mmu, "translations", translations, sz) == -1)
902 			panic("moea_bootstrap: can't get ofw translations");
903 		CTR0(KTR_PMAP, "moea_bootstrap: translations");
904 		sz /= sizeof(*translations);
905 		qsort(translations, sz, sizeof (*translations), om_cmp);
906 		for (i = 0; i < sz; i++) {
907 			CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
908 			    translations[i].om_pa, translations[i].om_va,
909 			    translations[i].om_len);
910 
911 			/*
912 			 * If the mapping is 1:1, let the RAM and device
913 			 * on-demand BAT tables take care of the translation.
914 			 */
915 			if (translations[i].om_va == translations[i].om_pa)
916 				continue;
917 
918 			/* Enter the pages */
919 			for (off = 0; off < translations[i].om_len;
920 			    off += PAGE_SIZE)
921 				moea_kenter(mmup, translations[i].om_va + off,
922 					    translations[i].om_pa + off);
923 		}
924 	}
925 
926 	/*
927 	 * Calculate the last available physical address.
928 	 */
929 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
930 		;
931 	Maxmem = powerpc_btop(phys_avail[i + 1]);
932 
933 	moea_cpu_bootstrap(mmup,0);
934 
935 	pmap_bootstrapped++;
936 
937 	/*
938 	 * Set the start and end of kva.
939 	 */
940 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
941 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
942 
943 	/*
944 	 * Allocate a kernel stack with a guard page for thread0 and map it
945 	 * into the kernel page map.
946 	 */
947 	pa = moea_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, PAGE_SIZE);
948 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
949 	virtual_avail = va + KSTACK_PAGES * PAGE_SIZE;
950 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
951 	thread0.td_kstack = va;
952 	thread0.td_kstack_pages = KSTACK_PAGES;
953 	for (i = 0; i < KSTACK_PAGES; i++) {
954 		moea_kenter(mmup, va, pa);
955 		pa += PAGE_SIZE;
956 		va += PAGE_SIZE;
957 	}
958 
959 	/*
960 	 * Allocate virtual address space for the message buffer.
961 	 */
962 	pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE);
963 	msgbufp = (struct msgbuf *)virtual_avail;
964 	va = virtual_avail;
965 	virtual_avail += round_page(msgbufsize);
966 	while (va < virtual_avail) {
967 		moea_kenter(mmup, va, pa);
968 		pa += PAGE_SIZE;
969 		va += PAGE_SIZE;
970 	}
971 
972 	/*
973 	 * Allocate virtual address space for the dynamic percpu area.
974 	 */
975 	pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
976 	dpcpu = (void *)virtual_avail;
977 	va = virtual_avail;
978 	virtual_avail += DPCPU_SIZE;
979 	while (va < virtual_avail) {
980 		moea_kenter(mmup, va, pa);
981 		pa += PAGE_SIZE;
982 		va += PAGE_SIZE;
983 	}
984 	dpcpu_init(dpcpu, 0);
985 }
986 
987 /*
988  * Activate a user pmap.  The pmap must be activated before it's address
989  * space can be accessed in any way.
990  */
991 void
992 moea_activate(mmu_t mmu, struct thread *td)
993 {
994 	pmap_t	pm, pmr;
995 
996 	/*
997 	 * Load all the data we need up front to encourage the compiler to
998 	 * not issue any loads while we have interrupts disabled below.
999 	 */
1000 	pm = &td->td_proc->p_vmspace->vm_pmap;
1001 	pmr = pm->pmap_phys;
1002 
1003 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1004 	PCPU_SET(curpmap, pmr);
1005 }
1006 
1007 void
1008 moea_deactivate(mmu_t mmu, struct thread *td)
1009 {
1010 	pmap_t	pm;
1011 
1012 	pm = &td->td_proc->p_vmspace->vm_pmap;
1013 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1014 	PCPU_SET(curpmap, NULL);
1015 }
1016 
1017 void
1018 moea_unwire(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1019 {
1020 	struct	pvo_entry key, *pvo;
1021 
1022 	PMAP_LOCK(pm);
1023 	key.pvo_vaddr = sva;
1024 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1025 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1026 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1027 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1028 			panic("moea_unwire: pvo %p is missing PVO_WIRED", pvo);
1029 		pvo->pvo_vaddr &= ~PVO_WIRED;
1030 		pm->pm_stats.wired_count--;
1031 	}
1032 	PMAP_UNLOCK(pm);
1033 }
1034 
1035 void
1036 moea_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1037 {
1038 	vm_offset_t	dst;
1039 	vm_offset_t	src;
1040 
1041 	dst = VM_PAGE_TO_PHYS(mdst);
1042 	src = VM_PAGE_TO_PHYS(msrc);
1043 
1044 	bcopy((void *)src, (void *)dst, PAGE_SIZE);
1045 }
1046 
1047 void
1048 moea_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1049     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1050 {
1051 	void *a_cp, *b_cp;
1052 	vm_offset_t a_pg_offset, b_pg_offset;
1053 	int cnt;
1054 
1055 	while (xfersize > 0) {
1056 		a_pg_offset = a_offset & PAGE_MASK;
1057 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1058 		a_cp = (char *)VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]) +
1059 		    a_pg_offset;
1060 		b_pg_offset = b_offset & PAGE_MASK;
1061 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1062 		b_cp = (char *)VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]) +
1063 		    b_pg_offset;
1064 		bcopy(a_cp, b_cp, cnt);
1065 		a_offset += cnt;
1066 		b_offset += cnt;
1067 		xfersize -= cnt;
1068 	}
1069 }
1070 
1071 /*
1072  * Zero a page of physical memory by temporarily mapping it into the tlb.
1073  */
1074 void
1075 moea_zero_page(mmu_t mmu, vm_page_t m)
1076 {
1077 	vm_offset_t off, pa = VM_PAGE_TO_PHYS(m);
1078 
1079 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1080 		__asm __volatile("dcbz 0,%0" :: "r"(pa + off));
1081 }
1082 
1083 void
1084 moea_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1085 {
1086 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1087 	void *va = (void *)(pa + off);
1088 
1089 	bzero(va, size);
1090 }
1091 
1092 void
1093 moea_zero_page_idle(mmu_t mmu, vm_page_t m)
1094 {
1095 
1096 	moea_zero_page(mmu, m);
1097 }
1098 
1099 /*
1100  * Map the given physical page at the specified virtual address in the
1101  * target pmap with the protection requested.  If specified the page
1102  * will be wired down.
1103  */
1104 int
1105 moea_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1106     u_int flags, int8_t psind)
1107 {
1108 	int error;
1109 
1110 	for (;;) {
1111 		rw_wlock(&pvh_global_lock);
1112 		PMAP_LOCK(pmap);
1113 		error = moea_enter_locked(pmap, va, m, prot, flags, psind);
1114 		rw_wunlock(&pvh_global_lock);
1115 		PMAP_UNLOCK(pmap);
1116 		if (error != ENOMEM)
1117 			return (KERN_SUCCESS);
1118 		if ((flags & PMAP_ENTER_NOSLEEP) != 0)
1119 			return (KERN_RESOURCE_SHORTAGE);
1120 		VM_OBJECT_ASSERT_UNLOCKED(m->object);
1121 		VM_WAIT;
1122 	}
1123 }
1124 
1125 /*
1126  * Map the given physical page at the specified virtual address in the
1127  * target pmap with the protection requested.  If specified the page
1128  * will be wired down.
1129  *
1130  * The global pvh and pmap must be locked.
1131  */
1132 static int
1133 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1134     u_int flags, int8_t psind __unused)
1135 {
1136 	struct		pvo_head *pvo_head;
1137 	uma_zone_t	zone;
1138 	u_int		pte_lo, pvo_flags;
1139 	int		error;
1140 
1141 	if (pmap_bootstrapped)
1142 		rw_assert(&pvh_global_lock, RA_WLOCKED);
1143 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1144 	if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m))
1145 		VM_OBJECT_ASSERT_LOCKED(m->object);
1146 
1147 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea_initialized) {
1148 		pvo_head = &moea_pvo_kunmanaged;
1149 		zone = moea_upvo_zone;
1150 		pvo_flags = 0;
1151 	} else {
1152 		pvo_head = vm_page_to_pvoh(m);
1153 		zone = moea_mpvo_zone;
1154 		pvo_flags = PVO_MANAGED;
1155 	}
1156 
1157 	pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1158 
1159 	if (prot & VM_PROT_WRITE) {
1160 		pte_lo |= PTE_BW;
1161 		if (pmap_bootstrapped &&
1162 		    (m->oflags & VPO_UNMANAGED) == 0)
1163 			vm_page_aflag_set(m, PGA_WRITEABLE);
1164 	} else
1165 		pte_lo |= PTE_BR;
1166 
1167 	if ((flags & PMAP_ENTER_WIRED) != 0)
1168 		pvo_flags |= PVO_WIRED;
1169 
1170 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1171 	    pte_lo, pvo_flags);
1172 
1173 	/*
1174 	 * Flush the real page from the instruction cache. This has be done
1175 	 * for all user mappings to prevent information leakage via the
1176 	 * instruction cache. moea_pvo_enter() returns ENOENT for the first
1177 	 * mapping for a page.
1178 	 */
1179 	if (pmap != kernel_pmap && error == ENOENT &&
1180 	    (pte_lo & (PTE_I | PTE_G)) == 0)
1181 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1182 
1183 	return (error);
1184 }
1185 
1186 /*
1187  * Maps a sequence of resident pages belonging to the same object.
1188  * The sequence begins with the given page m_start.  This page is
1189  * mapped at the given virtual address start.  Each subsequent page is
1190  * mapped at a virtual address that is offset from start by the same
1191  * amount as the page is offset from m_start within the object.  The
1192  * last page in the sequence is the page with the largest offset from
1193  * m_start that can be mapped at a virtual address less than the given
1194  * virtual address end.  Not every virtual page between start and end
1195  * is mapped; only those for which a resident page exists with the
1196  * corresponding offset from m_start are mapped.
1197  */
1198 void
1199 moea_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1200     vm_page_t m_start, vm_prot_t prot)
1201 {
1202 	vm_page_t m;
1203 	vm_pindex_t diff, psize;
1204 
1205 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1206 
1207 	psize = atop(end - start);
1208 	m = m_start;
1209 	rw_wlock(&pvh_global_lock);
1210 	PMAP_LOCK(pm);
1211 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1212 		moea_enter_locked(pm, start + ptoa(diff), m, prot &
1213 		    (VM_PROT_READ | VM_PROT_EXECUTE), 0, 0);
1214 		m = TAILQ_NEXT(m, listq);
1215 	}
1216 	rw_wunlock(&pvh_global_lock);
1217 	PMAP_UNLOCK(pm);
1218 }
1219 
1220 void
1221 moea_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1222     vm_prot_t prot)
1223 {
1224 
1225 	rw_wlock(&pvh_global_lock);
1226 	PMAP_LOCK(pm);
1227 	moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1228 	    0, 0);
1229 	rw_wunlock(&pvh_global_lock);
1230 	PMAP_UNLOCK(pm);
1231 }
1232 
1233 vm_paddr_t
1234 moea_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1235 {
1236 	struct	pvo_entry *pvo;
1237 	vm_paddr_t pa;
1238 
1239 	PMAP_LOCK(pm);
1240 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1241 	if (pvo == NULL)
1242 		pa = 0;
1243 	else
1244 		pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1245 	PMAP_UNLOCK(pm);
1246 	return (pa);
1247 }
1248 
1249 /*
1250  * Atomically extract and hold the physical page with the given
1251  * pmap and virtual address pair if that mapping permits the given
1252  * protection.
1253  */
1254 vm_page_t
1255 moea_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1256 {
1257 	struct	pvo_entry *pvo;
1258 	vm_page_t m;
1259         vm_paddr_t pa;
1260 
1261 	m = NULL;
1262 	pa = 0;
1263 	PMAP_LOCK(pmap);
1264 retry:
1265 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1266 	if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1267 	    ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1268 	     (prot & VM_PROT_WRITE) == 0)) {
1269 		if (vm_page_pa_tryrelock(pmap, pvo->pvo_pte.pte.pte_lo & PTE_RPGN, &pa))
1270 			goto retry;
1271 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
1272 		vm_page_hold(m);
1273 	}
1274 	PA_UNLOCK_COND(pa);
1275 	PMAP_UNLOCK(pmap);
1276 	return (m);
1277 }
1278 
1279 void
1280 moea_init(mmu_t mmu)
1281 {
1282 
1283 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1284 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1285 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1286 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1287 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1288 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1289 	moea_initialized = TRUE;
1290 }
1291 
1292 boolean_t
1293 moea_is_referenced(mmu_t mmu, vm_page_t m)
1294 {
1295 	boolean_t rv;
1296 
1297 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1298 	    ("moea_is_referenced: page %p is not managed", m));
1299 	rw_wlock(&pvh_global_lock);
1300 	rv = moea_query_bit(m, PTE_REF);
1301 	rw_wunlock(&pvh_global_lock);
1302 	return (rv);
1303 }
1304 
1305 boolean_t
1306 moea_is_modified(mmu_t mmu, vm_page_t m)
1307 {
1308 	boolean_t rv;
1309 
1310 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1311 	    ("moea_is_modified: page %p is not managed", m));
1312 
1313 	/*
1314 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1315 	 * concurrently set while the object is locked.  Thus, if PGA_WRITEABLE
1316 	 * is clear, no PTEs can have PTE_CHG set.
1317 	 */
1318 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1319 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1320 		return (FALSE);
1321 	rw_wlock(&pvh_global_lock);
1322 	rv = moea_query_bit(m, PTE_CHG);
1323 	rw_wunlock(&pvh_global_lock);
1324 	return (rv);
1325 }
1326 
1327 boolean_t
1328 moea_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1329 {
1330 	struct pvo_entry *pvo;
1331 	boolean_t rv;
1332 
1333 	PMAP_LOCK(pmap);
1334 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1335 	rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1336 	PMAP_UNLOCK(pmap);
1337 	return (rv);
1338 }
1339 
1340 void
1341 moea_clear_modify(mmu_t mmu, vm_page_t m)
1342 {
1343 
1344 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1345 	    ("moea_clear_modify: page %p is not managed", m));
1346 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1347 	KASSERT(!vm_page_xbusied(m),
1348 	    ("moea_clear_modify: page %p is exclusive busy", m));
1349 
1350 	/*
1351 	 * If the page is not PGA_WRITEABLE, then no PTEs can have PTE_CHG
1352 	 * set.  If the object containing the page is locked and the page is
1353 	 * not exclusive busied, then PGA_WRITEABLE cannot be concurrently set.
1354 	 */
1355 	if ((m->aflags & PGA_WRITEABLE) == 0)
1356 		return;
1357 	rw_wlock(&pvh_global_lock);
1358 	moea_clear_bit(m, PTE_CHG);
1359 	rw_wunlock(&pvh_global_lock);
1360 }
1361 
1362 /*
1363  * Clear the write and modified bits in each of the given page's mappings.
1364  */
1365 void
1366 moea_remove_write(mmu_t mmu, vm_page_t m)
1367 {
1368 	struct	pvo_entry *pvo;
1369 	struct	pte *pt;
1370 	pmap_t	pmap;
1371 	u_int	lo;
1372 
1373 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1374 	    ("moea_remove_write: page %p is not managed", m));
1375 
1376 	/*
1377 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1378 	 * set by another thread while the object is locked.  Thus,
1379 	 * if PGA_WRITEABLE is clear, no page table entries need updating.
1380 	 */
1381 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1382 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1383 		return;
1384 	rw_wlock(&pvh_global_lock);
1385 	lo = moea_attr_fetch(m);
1386 	powerpc_sync();
1387 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1388 		pmap = pvo->pvo_pmap;
1389 		PMAP_LOCK(pmap);
1390 		if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1391 			pt = moea_pvo_to_pte(pvo, -1);
1392 			pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1393 			pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1394 			if (pt != NULL) {
1395 				moea_pte_synch(pt, &pvo->pvo_pte.pte);
1396 				lo |= pvo->pvo_pte.pte.pte_lo;
1397 				pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1398 				moea_pte_change(pt, &pvo->pvo_pte.pte,
1399 				    pvo->pvo_vaddr);
1400 				mtx_unlock(&moea_table_mutex);
1401 			}
1402 		}
1403 		PMAP_UNLOCK(pmap);
1404 	}
1405 	if ((lo & PTE_CHG) != 0) {
1406 		moea_attr_clear(m, PTE_CHG);
1407 		vm_page_dirty(m);
1408 	}
1409 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1410 	rw_wunlock(&pvh_global_lock);
1411 }
1412 
1413 /*
1414  *	moea_ts_referenced:
1415  *
1416  *	Return a count of reference bits for a page, clearing those bits.
1417  *	It is not necessary for every reference bit to be cleared, but it
1418  *	is necessary that 0 only be returned when there are truly no
1419  *	reference bits set.
1420  *
1421  *	XXX: The exact number of bits to check and clear is a matter that
1422  *	should be tested and standardized at some point in the future for
1423  *	optimal aging of shared pages.
1424  */
1425 int
1426 moea_ts_referenced(mmu_t mmu, vm_page_t m)
1427 {
1428 	int count;
1429 
1430 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1431 	    ("moea_ts_referenced: page %p is not managed", m));
1432 	rw_wlock(&pvh_global_lock);
1433 	count = moea_clear_bit(m, PTE_REF);
1434 	rw_wunlock(&pvh_global_lock);
1435 	return (count);
1436 }
1437 
1438 /*
1439  * Modify the WIMG settings of all mappings for a page.
1440  */
1441 void
1442 moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1443 {
1444 	struct	pvo_entry *pvo;
1445 	struct	pvo_head *pvo_head;
1446 	struct	pte *pt;
1447 	pmap_t	pmap;
1448 	u_int	lo;
1449 
1450 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1451 		m->md.mdpg_cache_attrs = ma;
1452 		return;
1453 	}
1454 
1455 	rw_wlock(&pvh_global_lock);
1456 	pvo_head = vm_page_to_pvoh(m);
1457 	lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1458 
1459 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1460 		pmap = pvo->pvo_pmap;
1461 		PMAP_LOCK(pmap);
1462 		pt = moea_pvo_to_pte(pvo, -1);
1463 		pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1464 		pvo->pvo_pte.pte.pte_lo |= lo;
1465 		if (pt != NULL) {
1466 			moea_pte_change(pt, &pvo->pvo_pte.pte,
1467 			    pvo->pvo_vaddr);
1468 			if (pvo->pvo_pmap == kernel_pmap)
1469 				isync();
1470 		}
1471 		mtx_unlock(&moea_table_mutex);
1472 		PMAP_UNLOCK(pmap);
1473 	}
1474 	m->md.mdpg_cache_attrs = ma;
1475 	rw_wunlock(&pvh_global_lock);
1476 }
1477 
1478 /*
1479  * Map a wired page into kernel virtual address space.
1480  */
1481 void
1482 moea_kenter(mmu_t mmu, vm_offset_t va, vm_paddr_t pa)
1483 {
1484 
1485 	moea_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1486 }
1487 
1488 void
1489 moea_kenter_attr(mmu_t mmu, vm_offset_t va, vm_offset_t pa, vm_memattr_t ma)
1490 {
1491 	u_int		pte_lo;
1492 	int		error;
1493 
1494 #if 0
1495 	if (va < VM_MIN_KERNEL_ADDRESS)
1496 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1497 		    va);
1498 #endif
1499 
1500 	pte_lo = moea_calc_wimg(pa, ma);
1501 
1502 	PMAP_LOCK(kernel_pmap);
1503 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1504 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1505 
1506 	if (error != 0 && error != ENOENT)
1507 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1508 		    pa, error);
1509 
1510 	PMAP_UNLOCK(kernel_pmap);
1511 }
1512 
1513 /*
1514  * Extract the physical page address associated with the given kernel virtual
1515  * address.
1516  */
1517 vm_paddr_t
1518 moea_kextract(mmu_t mmu, vm_offset_t va)
1519 {
1520 	struct		pvo_entry *pvo;
1521 	vm_paddr_t pa;
1522 
1523 	/*
1524 	 * Allow direct mappings on 32-bit OEA
1525 	 */
1526 	if (va < VM_MIN_KERNEL_ADDRESS) {
1527 		return (va);
1528 	}
1529 
1530 	PMAP_LOCK(kernel_pmap);
1531 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1532 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1533 	pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1534 	PMAP_UNLOCK(kernel_pmap);
1535 	return (pa);
1536 }
1537 
1538 /*
1539  * Remove a wired page from kernel virtual address space.
1540  */
1541 void
1542 moea_kremove(mmu_t mmu, vm_offset_t va)
1543 {
1544 
1545 	moea_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1546 }
1547 
1548 /*
1549  * Map a range of physical addresses into kernel virtual address space.
1550  *
1551  * The value passed in *virt is a suggested virtual address for the mapping.
1552  * Architectures which can support a direct-mapped physical to virtual region
1553  * can return the appropriate address within that region, leaving '*virt'
1554  * unchanged.  We cannot and therefore do not; *virt is updated with the
1555  * first usable address after the mapped region.
1556  */
1557 vm_offset_t
1558 moea_map(mmu_t mmu, vm_offset_t *virt, vm_paddr_t pa_start,
1559     vm_paddr_t pa_end, int prot)
1560 {
1561 	vm_offset_t	sva, va;
1562 
1563 	sva = *virt;
1564 	va = sva;
1565 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1566 		moea_kenter(mmu, va, pa_start);
1567 	*virt = va;
1568 	return (sva);
1569 }
1570 
1571 /*
1572  * Returns true if the pmap's pv is one of the first
1573  * 16 pvs linked to from this page.  This count may
1574  * be changed upwards or downwards in the future; it
1575  * is only necessary that true be returned for a small
1576  * subset of pmaps for proper page aging.
1577  */
1578 boolean_t
1579 moea_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1580 {
1581         int loops;
1582 	struct pvo_entry *pvo;
1583 	boolean_t rv;
1584 
1585 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1586 	    ("moea_page_exists_quick: page %p is not managed", m));
1587 	loops = 0;
1588 	rv = FALSE;
1589 	rw_wlock(&pvh_global_lock);
1590 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1591 		if (pvo->pvo_pmap == pmap) {
1592 			rv = TRUE;
1593 			break;
1594 		}
1595 		if (++loops >= 16)
1596 			break;
1597 	}
1598 	rw_wunlock(&pvh_global_lock);
1599 	return (rv);
1600 }
1601 
1602 /*
1603  * Return the number of managed mappings to the given physical page
1604  * that are wired.
1605  */
1606 int
1607 moea_page_wired_mappings(mmu_t mmu, vm_page_t m)
1608 {
1609 	struct pvo_entry *pvo;
1610 	int count;
1611 
1612 	count = 0;
1613 	if ((m->oflags & VPO_UNMANAGED) != 0)
1614 		return (count);
1615 	rw_wlock(&pvh_global_lock);
1616 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1617 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1618 			count++;
1619 	rw_wunlock(&pvh_global_lock);
1620 	return (count);
1621 }
1622 
1623 static u_int	moea_vsidcontext;
1624 
1625 void
1626 moea_pinit(mmu_t mmu, pmap_t pmap)
1627 {
1628 	int	i, mask;
1629 	u_int	entropy;
1630 
1631 	KASSERT((int)pmap < VM_MIN_KERNEL_ADDRESS, ("moea_pinit: virt pmap"));
1632 	RB_INIT(&pmap->pmap_pvo);
1633 
1634 	entropy = 0;
1635 	__asm __volatile("mftb %0" : "=r"(entropy));
1636 
1637 	if ((pmap->pmap_phys = (pmap_t)moea_kextract(mmu, (vm_offset_t)pmap))
1638 	    == NULL) {
1639 		pmap->pmap_phys = pmap;
1640 	}
1641 
1642 
1643 	mtx_lock(&moea_vsid_mutex);
1644 	/*
1645 	 * Allocate some segment registers for this pmap.
1646 	 */
1647 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1648 		u_int	hash, n;
1649 
1650 		/*
1651 		 * Create a new value by mutiplying by a prime and adding in
1652 		 * entropy from the timebase register.  This is to make the
1653 		 * VSID more random so that the PT hash function collides
1654 		 * less often.  (Note that the prime casues gcc to do shifts
1655 		 * instead of a multiply.)
1656 		 */
1657 		moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1658 		hash = moea_vsidcontext & (NPMAPS - 1);
1659 		if (hash == 0)		/* 0 is special, avoid it */
1660 			continue;
1661 		n = hash >> 5;
1662 		mask = 1 << (hash & (VSID_NBPW - 1));
1663 		hash = (moea_vsidcontext & 0xfffff);
1664 		if (moea_vsid_bitmap[n] & mask) {	/* collision? */
1665 			/* anything free in this bucket? */
1666 			if (moea_vsid_bitmap[n] == 0xffffffff) {
1667 				entropy = (moea_vsidcontext >> 20);
1668 				continue;
1669 			}
1670 			i = ffs(~moea_vsid_bitmap[n]) - 1;
1671 			mask = 1 << i;
1672 			hash &= 0xfffff & ~(VSID_NBPW - 1);
1673 			hash |= i;
1674 		}
1675 		KASSERT(!(moea_vsid_bitmap[n] & mask),
1676 		    ("Allocating in-use VSID group %#x\n", hash));
1677 		moea_vsid_bitmap[n] |= mask;
1678 		for (i = 0; i < 16; i++)
1679 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1680 		mtx_unlock(&moea_vsid_mutex);
1681 		return;
1682 	}
1683 
1684 	mtx_unlock(&moea_vsid_mutex);
1685 	panic("moea_pinit: out of segments");
1686 }
1687 
1688 /*
1689  * Initialize the pmap associated with process 0.
1690  */
1691 void
1692 moea_pinit0(mmu_t mmu, pmap_t pm)
1693 {
1694 
1695 	PMAP_LOCK_INIT(pm);
1696 	moea_pinit(mmu, pm);
1697 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1698 }
1699 
1700 /*
1701  * Set the physical protection on the specified range of this map as requested.
1702  */
1703 void
1704 moea_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1705     vm_prot_t prot)
1706 {
1707 	struct	pvo_entry *pvo, *tpvo, key;
1708 	struct	pte *pt;
1709 
1710 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1711 	    ("moea_protect: non current pmap"));
1712 
1713 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1714 		moea_remove(mmu, pm, sva, eva);
1715 		return;
1716 	}
1717 
1718 	rw_wlock(&pvh_global_lock);
1719 	PMAP_LOCK(pm);
1720 	key.pvo_vaddr = sva;
1721 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1722 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1723 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1724 
1725 		/*
1726 		 * Grab the PTE pointer before we diddle with the cached PTE
1727 		 * copy.
1728 		 */
1729 		pt = moea_pvo_to_pte(pvo, -1);
1730 		/*
1731 		 * Change the protection of the page.
1732 		 */
1733 		pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1734 		pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1735 
1736 		/*
1737 		 * If the PVO is in the page table, update that pte as well.
1738 		 */
1739 		if (pt != NULL) {
1740 			moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1741 			mtx_unlock(&moea_table_mutex);
1742 		}
1743 	}
1744 	rw_wunlock(&pvh_global_lock);
1745 	PMAP_UNLOCK(pm);
1746 }
1747 
1748 /*
1749  * Map a list of wired pages into kernel virtual address space.  This is
1750  * intended for temporary mappings which do not need page modification or
1751  * references recorded.  Existing mappings in the region are overwritten.
1752  */
1753 void
1754 moea_qenter(mmu_t mmu, vm_offset_t sva, vm_page_t *m, int count)
1755 {
1756 	vm_offset_t va;
1757 
1758 	va = sva;
1759 	while (count-- > 0) {
1760 		moea_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1761 		va += PAGE_SIZE;
1762 		m++;
1763 	}
1764 }
1765 
1766 /*
1767  * Remove page mappings from kernel virtual address space.  Intended for
1768  * temporary mappings entered by moea_qenter.
1769  */
1770 void
1771 moea_qremove(mmu_t mmu, vm_offset_t sva, int count)
1772 {
1773 	vm_offset_t va;
1774 
1775 	va = sva;
1776 	while (count-- > 0) {
1777 		moea_kremove(mmu, va);
1778 		va += PAGE_SIZE;
1779 	}
1780 }
1781 
1782 void
1783 moea_release(mmu_t mmu, pmap_t pmap)
1784 {
1785         int idx, mask;
1786 
1787 	/*
1788 	 * Free segment register's VSID
1789 	 */
1790         if (pmap->pm_sr[0] == 0)
1791                 panic("moea_release");
1792 
1793 	mtx_lock(&moea_vsid_mutex);
1794         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1795         mask = 1 << (idx % VSID_NBPW);
1796         idx /= VSID_NBPW;
1797         moea_vsid_bitmap[idx] &= ~mask;
1798 	mtx_unlock(&moea_vsid_mutex);
1799 }
1800 
1801 /*
1802  * Remove the given range of addresses from the specified map.
1803  */
1804 void
1805 moea_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1806 {
1807 	struct	pvo_entry *pvo, *tpvo, key;
1808 
1809 	rw_wlock(&pvh_global_lock);
1810 	PMAP_LOCK(pm);
1811 	key.pvo_vaddr = sva;
1812 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1813 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1814 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1815 		moea_pvo_remove(pvo, -1);
1816 	}
1817 	PMAP_UNLOCK(pm);
1818 	rw_wunlock(&pvh_global_lock);
1819 }
1820 
1821 /*
1822  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1823  * will reflect changes in pte's back to the vm_page.
1824  */
1825 void
1826 moea_remove_all(mmu_t mmu, vm_page_t m)
1827 {
1828 	struct  pvo_head *pvo_head;
1829 	struct	pvo_entry *pvo, *next_pvo;
1830 	pmap_t	pmap;
1831 
1832 	rw_wlock(&pvh_global_lock);
1833 	pvo_head = vm_page_to_pvoh(m);
1834 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1835 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1836 
1837 		pmap = pvo->pvo_pmap;
1838 		PMAP_LOCK(pmap);
1839 		moea_pvo_remove(pvo, -1);
1840 		PMAP_UNLOCK(pmap);
1841 	}
1842 	if ((m->aflags & PGA_WRITEABLE) && moea_query_bit(m, PTE_CHG)) {
1843 		moea_attr_clear(m, PTE_CHG);
1844 		vm_page_dirty(m);
1845 	}
1846 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1847 	rw_wunlock(&pvh_global_lock);
1848 }
1849 
1850 /*
1851  * Allocate a physical page of memory directly from the phys_avail map.
1852  * Can only be called from moea_bootstrap before avail start and end are
1853  * calculated.
1854  */
1855 static vm_offset_t
1856 moea_bootstrap_alloc(vm_size_t size, u_int align)
1857 {
1858 	vm_offset_t	s, e;
1859 	int		i, j;
1860 
1861 	size = round_page(size);
1862 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1863 		if (align != 0)
1864 			s = (phys_avail[i] + align - 1) & ~(align - 1);
1865 		else
1866 			s = phys_avail[i];
1867 		e = s + size;
1868 
1869 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1870 			continue;
1871 
1872 		if (s == phys_avail[i]) {
1873 			phys_avail[i] += size;
1874 		} else if (e == phys_avail[i + 1]) {
1875 			phys_avail[i + 1] -= size;
1876 		} else {
1877 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1878 				phys_avail[j] = phys_avail[j - 2];
1879 				phys_avail[j + 1] = phys_avail[j - 1];
1880 			}
1881 
1882 			phys_avail[i + 3] = phys_avail[i + 1];
1883 			phys_avail[i + 1] = s;
1884 			phys_avail[i + 2] = e;
1885 			phys_avail_count++;
1886 		}
1887 
1888 		return (s);
1889 	}
1890 	panic("moea_bootstrap_alloc: could not allocate memory");
1891 }
1892 
1893 static void
1894 moea_syncicache(vm_offset_t pa, vm_size_t len)
1895 {
1896 	__syncicache((void *)pa, len);
1897 }
1898 
1899 static int
1900 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1901     vm_offset_t va, vm_offset_t pa, u_int pte_lo, int flags)
1902 {
1903 	struct	pvo_entry *pvo;
1904 	u_int	sr;
1905 	int	first;
1906 	u_int	ptegidx;
1907 	int	i;
1908 	int     bootstrap;
1909 
1910 	moea_pvo_enter_calls++;
1911 	first = 0;
1912 	bootstrap = 0;
1913 
1914 	/*
1915 	 * Compute the PTE Group index.
1916 	 */
1917 	va &= ~ADDR_POFF;
1918 	sr = va_to_sr(pm->pm_sr, va);
1919 	ptegidx = va_to_pteg(sr, va);
1920 
1921 	/*
1922 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
1923 	 * there is a mapping.
1924 	 */
1925 	mtx_lock(&moea_table_mutex);
1926 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
1927 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
1928 			if ((pvo->pvo_pte.pte.pte_lo & PTE_RPGN) == pa &&
1929 			    (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
1930 			    (pte_lo & PTE_PP)) {
1931 				/*
1932 				 * The PTE is not changing.  Instead, this may
1933 				 * be a request to change the mapping's wired
1934 				 * attribute.
1935 				 */
1936 				mtx_unlock(&moea_table_mutex);
1937 				if ((flags & PVO_WIRED) != 0 &&
1938 				    (pvo->pvo_vaddr & PVO_WIRED) == 0) {
1939 					pvo->pvo_vaddr |= PVO_WIRED;
1940 					pm->pm_stats.wired_count++;
1941 				} else if ((flags & PVO_WIRED) == 0 &&
1942 				    (pvo->pvo_vaddr & PVO_WIRED) != 0) {
1943 					pvo->pvo_vaddr &= ~PVO_WIRED;
1944 					pm->pm_stats.wired_count--;
1945 				}
1946 				return (0);
1947 			}
1948 			moea_pvo_remove(pvo, -1);
1949 			break;
1950 		}
1951 	}
1952 
1953 	/*
1954 	 * If we aren't overwriting a mapping, try to allocate.
1955 	 */
1956 	if (moea_initialized) {
1957 		pvo = uma_zalloc(zone, M_NOWAIT);
1958 	} else {
1959 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
1960 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
1961 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
1962 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
1963 		}
1964 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
1965 		moea_bpvo_pool_index++;
1966 		bootstrap = 1;
1967 	}
1968 
1969 	if (pvo == NULL) {
1970 		mtx_unlock(&moea_table_mutex);
1971 		return (ENOMEM);
1972 	}
1973 
1974 	moea_pvo_entries++;
1975 	pvo->pvo_vaddr = va;
1976 	pvo->pvo_pmap = pm;
1977 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
1978 	pvo->pvo_vaddr &= ~ADDR_POFF;
1979 	if (flags & PVO_WIRED)
1980 		pvo->pvo_vaddr |= PVO_WIRED;
1981 	if (pvo_head != &moea_pvo_kunmanaged)
1982 		pvo->pvo_vaddr |= PVO_MANAGED;
1983 	if (bootstrap)
1984 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
1985 
1986 	moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo);
1987 
1988 	/*
1989 	 * Add to pmap list
1990 	 */
1991 	RB_INSERT(pvo_tree, &pm->pmap_pvo, pvo);
1992 
1993 	/*
1994 	 * Remember if the list was empty and therefore will be the first
1995 	 * item.
1996 	 */
1997 	if (LIST_FIRST(pvo_head) == NULL)
1998 		first = 1;
1999 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2000 
2001 	if (pvo->pvo_vaddr & PVO_WIRED)
2002 		pm->pm_stats.wired_count++;
2003 	pm->pm_stats.resident_count++;
2004 
2005 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2006 	KASSERT(i < 8, ("Invalid PTE index"));
2007 	if (i >= 0) {
2008 		PVO_PTEGIDX_SET(pvo, i);
2009 	} else {
2010 		panic("moea_pvo_enter: overflow");
2011 		moea_pte_overflow++;
2012 	}
2013 	mtx_unlock(&moea_table_mutex);
2014 
2015 	return (first ? ENOENT : 0);
2016 }
2017 
2018 static void
2019 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
2020 {
2021 	struct	pte *pt;
2022 
2023 	/*
2024 	 * If there is an active pte entry, we need to deactivate it (and
2025 	 * save the ref & cfg bits).
2026 	 */
2027 	pt = moea_pvo_to_pte(pvo, pteidx);
2028 	if (pt != NULL) {
2029 		moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
2030 		mtx_unlock(&moea_table_mutex);
2031 		PVO_PTEGIDX_CLR(pvo);
2032 	} else {
2033 		moea_pte_overflow--;
2034 	}
2035 
2036 	/*
2037 	 * Update our statistics.
2038 	 */
2039 	pvo->pvo_pmap->pm_stats.resident_count--;
2040 	if (pvo->pvo_vaddr & PVO_WIRED)
2041 		pvo->pvo_pmap->pm_stats.wired_count--;
2042 
2043 	/*
2044 	 * Save the REF/CHG bits into their cache if the page is managed.
2045 	 */
2046 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED) {
2047 		struct	vm_page *pg;
2048 
2049 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
2050 		if (pg != NULL) {
2051 			moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2052 			    (PTE_REF | PTE_CHG));
2053 		}
2054 	}
2055 
2056 	/*
2057 	 * Remove this PVO from the PV and pmap lists.
2058 	 */
2059 	LIST_REMOVE(pvo, pvo_vlink);
2060 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2061 
2062 	/*
2063 	 * Remove this from the overflow list and return it to the pool
2064 	 * if we aren't going to reuse it.
2065 	 */
2066 	LIST_REMOVE(pvo, pvo_olink);
2067 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2068 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
2069 		    moea_upvo_zone, pvo);
2070 	moea_pvo_entries--;
2071 	moea_pvo_remove_calls++;
2072 }
2073 
2074 static __inline int
2075 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2076 {
2077 	int	pteidx;
2078 
2079 	/*
2080 	 * We can find the actual pte entry without searching by grabbing
2081 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2082 	 * noticing the HID bit.
2083 	 */
2084 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2085 	if (pvo->pvo_pte.pte.pte_hi & PTE_HID)
2086 		pteidx ^= moea_pteg_mask * 8;
2087 
2088 	return (pteidx);
2089 }
2090 
2091 static struct pvo_entry *
2092 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2093 {
2094 	struct	pvo_entry *pvo;
2095 	int	ptegidx;
2096 	u_int	sr;
2097 
2098 	va &= ~ADDR_POFF;
2099 	sr = va_to_sr(pm->pm_sr, va);
2100 	ptegidx = va_to_pteg(sr, va);
2101 
2102 	mtx_lock(&moea_table_mutex);
2103 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2104 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2105 			if (pteidx_p)
2106 				*pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2107 			break;
2108 		}
2109 	}
2110 	mtx_unlock(&moea_table_mutex);
2111 
2112 	return (pvo);
2113 }
2114 
2115 static struct pte *
2116 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2117 {
2118 	struct	pte *pt;
2119 
2120 	/*
2121 	 * If we haven't been supplied the ptegidx, calculate it.
2122 	 */
2123 	if (pteidx == -1) {
2124 		int	ptegidx;
2125 		u_int	sr;
2126 
2127 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2128 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2129 		pteidx = moea_pvo_pte_index(pvo, ptegidx);
2130 	}
2131 
2132 	pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2133 	mtx_lock(&moea_table_mutex);
2134 
2135 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2136 		panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2137 		    "valid pte index", pvo);
2138 	}
2139 
2140 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2141 		panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2142 		    "pvo but no valid pte", pvo);
2143 	}
2144 
2145 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2146 		if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) {
2147 			panic("moea_pvo_to_pte: pvo %p has valid pte in "
2148 			    "moea_pteg_table %p but invalid in pvo", pvo, pt);
2149 		}
2150 
2151 		if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2152 		    != 0) {
2153 			panic("moea_pvo_to_pte: pvo %p pte does not match "
2154 			    "pte %p in moea_pteg_table", pvo, pt);
2155 		}
2156 
2157 		mtx_assert(&moea_table_mutex, MA_OWNED);
2158 		return (pt);
2159 	}
2160 
2161 	if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) {
2162 		panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2163 		    "moea_pteg_table but valid in pvo: %8x, %8x", pvo, pt, pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2164 	}
2165 
2166 	mtx_unlock(&moea_table_mutex);
2167 	return (NULL);
2168 }
2169 
2170 /*
2171  * XXX: THIS STUFF SHOULD BE IN pte.c?
2172  */
2173 int
2174 moea_pte_spill(vm_offset_t addr)
2175 {
2176 	struct	pvo_entry *source_pvo, *victim_pvo;
2177 	struct	pvo_entry *pvo;
2178 	int	ptegidx, i, j;
2179 	u_int	sr;
2180 	struct	pteg *pteg;
2181 	struct	pte *pt;
2182 
2183 	moea_pte_spills++;
2184 
2185 	sr = mfsrin(addr);
2186 	ptegidx = va_to_pteg(sr, addr);
2187 
2188 	/*
2189 	 * Have to substitute some entry.  Use the primary hash for this.
2190 	 * Use low bits of timebase as random generator.
2191 	 */
2192 	pteg = &moea_pteg_table[ptegidx];
2193 	mtx_lock(&moea_table_mutex);
2194 	__asm __volatile("mftb %0" : "=r"(i));
2195 	i &= 7;
2196 	pt = &pteg->pt[i];
2197 
2198 	source_pvo = NULL;
2199 	victim_pvo = NULL;
2200 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2201 		/*
2202 		 * We need to find a pvo entry for this address.
2203 		 */
2204 		if (source_pvo == NULL &&
2205 		    moea_pte_match(&pvo->pvo_pte.pte, sr, addr,
2206 		    pvo->pvo_pte.pte.pte_hi & PTE_HID)) {
2207 			/*
2208 			 * Now found an entry to be spilled into the pteg.
2209 			 * The PTE is now valid, so we know it's active.
2210 			 */
2211 			j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2212 
2213 			if (j >= 0) {
2214 				PVO_PTEGIDX_SET(pvo, j);
2215 				moea_pte_overflow--;
2216 				mtx_unlock(&moea_table_mutex);
2217 				return (1);
2218 			}
2219 
2220 			source_pvo = pvo;
2221 
2222 			if (victim_pvo != NULL)
2223 				break;
2224 		}
2225 
2226 		/*
2227 		 * We also need the pvo entry of the victim we are replacing
2228 		 * so save the R & C bits of the PTE.
2229 		 */
2230 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2231 		    moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2232 			victim_pvo = pvo;
2233 			if (source_pvo != NULL)
2234 				break;
2235 		}
2236 	}
2237 
2238 	if (source_pvo == NULL) {
2239 		mtx_unlock(&moea_table_mutex);
2240 		return (0);
2241 	}
2242 
2243 	if (victim_pvo == NULL) {
2244 		if ((pt->pte_hi & PTE_HID) == 0)
2245 			panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2246 			    "entry", pt);
2247 
2248 		/*
2249 		 * If this is a secondary PTE, we need to search it's primary
2250 		 * pvo bucket for the matching PVO.
2251 		 */
2252 		LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2253 		    pvo_olink) {
2254 			/*
2255 			 * We also need the pvo entry of the victim we are
2256 			 * replacing so save the R & C bits of the PTE.
2257 			 */
2258 			if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2259 				victim_pvo = pvo;
2260 				break;
2261 			}
2262 		}
2263 
2264 		if (victim_pvo == NULL)
2265 			panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2266 			    "entry", pt);
2267 	}
2268 
2269 	/*
2270 	 * We are invalidating the TLB entry for the EA we are replacing even
2271 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2272 	 * contained in the TLB entry.
2273 	 */
2274 	source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID;
2275 
2276 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2277 	moea_pte_set(pt, &source_pvo->pvo_pte.pte);
2278 
2279 	PVO_PTEGIDX_CLR(victim_pvo);
2280 	PVO_PTEGIDX_SET(source_pvo, i);
2281 	moea_pte_replacements++;
2282 
2283 	mtx_unlock(&moea_table_mutex);
2284 	return (1);
2285 }
2286 
2287 static __inline struct pvo_entry *
2288 moea_pte_spillable_ident(u_int ptegidx)
2289 {
2290 	struct	pte *pt;
2291 	struct	pvo_entry *pvo_walk, *pvo = NULL;
2292 
2293 	LIST_FOREACH(pvo_walk, &moea_pvo_table[ptegidx], pvo_olink) {
2294 		if (pvo_walk->pvo_vaddr & PVO_WIRED)
2295 			continue;
2296 
2297 		if (!(pvo_walk->pvo_pte.pte.pte_hi & PTE_VALID))
2298 			continue;
2299 
2300 		pt = moea_pvo_to_pte(pvo_walk, -1);
2301 
2302 		if (pt == NULL)
2303 			continue;
2304 
2305 		pvo = pvo_walk;
2306 
2307 		mtx_unlock(&moea_table_mutex);
2308 		if (!(pt->pte_lo & PTE_REF))
2309 			return (pvo_walk);
2310 	}
2311 
2312 	return (pvo);
2313 }
2314 
2315 static int
2316 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2317 {
2318 	struct	pte *pt;
2319 	struct	pvo_entry *victim_pvo;
2320 	int	i;
2321 	int	victim_idx;
2322 	u_int	pteg_bkpidx = ptegidx;
2323 
2324 	mtx_assert(&moea_table_mutex, MA_OWNED);
2325 
2326 	/*
2327 	 * First try primary hash.
2328 	 */
2329 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2330 		if ((pt->pte_hi & PTE_VALID) == 0) {
2331 			pvo_pt->pte_hi &= ~PTE_HID;
2332 			moea_pte_set(pt, pvo_pt);
2333 			return (i);
2334 		}
2335 	}
2336 
2337 	/*
2338 	 * Now try secondary hash.
2339 	 */
2340 	ptegidx ^= moea_pteg_mask;
2341 
2342 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2343 		if ((pt->pte_hi & PTE_VALID) == 0) {
2344 			pvo_pt->pte_hi |= PTE_HID;
2345 			moea_pte_set(pt, pvo_pt);
2346 			return (i);
2347 		}
2348 	}
2349 
2350 	/* Try again, but this time try to force a PTE out. */
2351 	ptegidx = pteg_bkpidx;
2352 
2353 	victim_pvo = moea_pte_spillable_ident(ptegidx);
2354 	if (victim_pvo == NULL) {
2355 		ptegidx ^= moea_pteg_mask;
2356 		victim_pvo = moea_pte_spillable_ident(ptegidx);
2357 	}
2358 
2359 	if (victim_pvo == NULL) {
2360 		panic("moea_pte_insert: overflow");
2361 		return (-1);
2362 	}
2363 
2364 	victim_idx = moea_pvo_pte_index(victim_pvo, ptegidx);
2365 
2366 	if (pteg_bkpidx == ptegidx)
2367 		pvo_pt->pte_hi &= ~PTE_HID;
2368 	else
2369 		pvo_pt->pte_hi |= PTE_HID;
2370 
2371 	/*
2372 	 * Synchronize the sacrifice PTE with its PVO, then mark both
2373 	 * invalid. The PVO will be reused when/if the VM system comes
2374 	 * here after a fault.
2375 	 */
2376 	pt = &moea_pteg_table[victim_idx >> 3].pt[victim_idx & 7];
2377 
2378 	if (pt->pte_hi != victim_pvo->pvo_pte.pte.pte_hi)
2379 	    panic("Victim PVO doesn't match PTE! PVO: %8x, PTE: %8x", victim_pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2380 
2381 	/*
2382 	 * Set the new PTE.
2383 	 */
2384 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2385 	PVO_PTEGIDX_CLR(victim_pvo);
2386 	moea_pte_overflow++;
2387 	moea_pte_set(pt, pvo_pt);
2388 
2389 	return (victim_idx & 7);
2390 }
2391 
2392 static boolean_t
2393 moea_query_bit(vm_page_t m, int ptebit)
2394 {
2395 	struct	pvo_entry *pvo;
2396 	struct	pte *pt;
2397 
2398 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2399 	if (moea_attr_fetch(m) & ptebit)
2400 		return (TRUE);
2401 
2402 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2403 
2404 		/*
2405 		 * See if we saved the bit off.  If so, cache it and return
2406 		 * success.
2407 		 */
2408 		if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2409 			moea_attr_save(m, ptebit);
2410 			return (TRUE);
2411 		}
2412 	}
2413 
2414 	/*
2415 	 * No luck, now go through the hard part of looking at the PTEs
2416 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2417 	 * the PTEs.
2418 	 */
2419 	powerpc_sync();
2420 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2421 
2422 		/*
2423 		 * See if this pvo has a valid PTE.  if so, fetch the
2424 		 * REF/CHG bits from the valid PTE.  If the appropriate
2425 		 * ptebit is set, cache it and return success.
2426 		 */
2427 		pt = moea_pvo_to_pte(pvo, -1);
2428 		if (pt != NULL) {
2429 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2430 			mtx_unlock(&moea_table_mutex);
2431 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2432 				moea_attr_save(m, ptebit);
2433 				return (TRUE);
2434 			}
2435 		}
2436 	}
2437 
2438 	return (FALSE);
2439 }
2440 
2441 static u_int
2442 moea_clear_bit(vm_page_t m, int ptebit)
2443 {
2444 	u_int	count;
2445 	struct	pvo_entry *pvo;
2446 	struct	pte *pt;
2447 
2448 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2449 
2450 	/*
2451 	 * Clear the cached value.
2452 	 */
2453 	moea_attr_clear(m, ptebit);
2454 
2455 	/*
2456 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2457 	 * we can reset the right ones).  note that since the pvo entries and
2458 	 * list heads are accessed via BAT0 and are never placed in the page
2459 	 * table, we don't have to worry about further accesses setting the
2460 	 * REF/CHG bits.
2461 	 */
2462 	powerpc_sync();
2463 
2464 	/*
2465 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2466 	 * valid pte clear the ptebit from the valid pte.
2467 	 */
2468 	count = 0;
2469 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2470 		pt = moea_pvo_to_pte(pvo, -1);
2471 		if (pt != NULL) {
2472 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2473 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2474 				count++;
2475 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2476 			}
2477 			mtx_unlock(&moea_table_mutex);
2478 		}
2479 		pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2480 	}
2481 
2482 	return (count);
2483 }
2484 
2485 /*
2486  * Return true if the physical range is encompassed by the battable[idx]
2487  */
2488 static int
2489 moea_bat_mapped(int idx, vm_offset_t pa, vm_size_t size)
2490 {
2491 	u_int prot;
2492 	u_int32_t start;
2493 	u_int32_t end;
2494 	u_int32_t bat_ble;
2495 
2496 	/*
2497 	 * Return immediately if not a valid mapping
2498 	 */
2499 	if (!(battable[idx].batu & BAT_Vs))
2500 		return (EINVAL);
2501 
2502 	/*
2503 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2504 	 * so it can function as an i/o page
2505 	 */
2506 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2507 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2508 		return (EPERM);
2509 
2510 	/*
2511 	 * The address should be within the BAT range. Assume that the
2512 	 * start address in the BAT has the correct alignment (thus
2513 	 * not requiring masking)
2514 	 */
2515 	start = battable[idx].batl & BAT_PBS;
2516 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2517 	end = start | (bat_ble << 15) | 0x7fff;
2518 
2519 	if ((pa < start) || ((pa + size) > end))
2520 		return (ERANGE);
2521 
2522 	return (0);
2523 }
2524 
2525 boolean_t
2526 moea_dev_direct_mapped(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2527 {
2528 	int i;
2529 
2530 	/*
2531 	 * This currently does not work for entries that
2532 	 * overlap 256M BAT segments.
2533 	 */
2534 
2535 	for(i = 0; i < 16; i++)
2536 		if (moea_bat_mapped(i, pa, size) == 0)
2537 			return (0);
2538 
2539 	return (EFAULT);
2540 }
2541 
2542 /*
2543  * Map a set of physical memory pages into the kernel virtual
2544  * address space. Return a pointer to where it is mapped. This
2545  * routine is intended to be used for mapping device memory,
2546  * NOT real memory.
2547  */
2548 void *
2549 moea_mapdev(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2550 {
2551 
2552 	return (moea_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT));
2553 }
2554 
2555 void *
2556 moea_mapdev_attr(mmu_t mmu, vm_offset_t pa, vm_size_t size, vm_memattr_t ma)
2557 {
2558 	vm_offset_t va, tmpva, ppa, offset;
2559 	int i;
2560 
2561 	ppa = trunc_page(pa);
2562 	offset = pa & PAGE_MASK;
2563 	size = roundup(offset + size, PAGE_SIZE);
2564 
2565 	/*
2566 	 * If the physical address lies within a valid BAT table entry,
2567 	 * return the 1:1 mapping. This currently doesn't work
2568 	 * for regions that overlap 256M BAT segments.
2569 	 */
2570 	for (i = 0; i < 16; i++) {
2571 		if (moea_bat_mapped(i, pa, size) == 0)
2572 			return ((void *) pa);
2573 	}
2574 
2575 	va = kva_alloc(size);
2576 	if (!va)
2577 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2578 
2579 	for (tmpva = va; size > 0;) {
2580 		moea_kenter_attr(mmu, tmpva, ppa, ma);
2581 		tlbie(tmpva);
2582 		size -= PAGE_SIZE;
2583 		tmpva += PAGE_SIZE;
2584 		ppa += PAGE_SIZE;
2585 	}
2586 
2587 	return ((void *)(va + offset));
2588 }
2589 
2590 void
2591 moea_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2592 {
2593 	vm_offset_t base, offset;
2594 
2595 	/*
2596 	 * If this is outside kernel virtual space, then it's a
2597 	 * battable entry and doesn't require unmapping
2598 	 */
2599 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2600 		base = trunc_page(va);
2601 		offset = va & PAGE_MASK;
2602 		size = roundup(offset + size, PAGE_SIZE);
2603 		kva_free(base, size);
2604 	}
2605 }
2606 
2607 static void
2608 moea_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2609 {
2610 	struct pvo_entry *pvo;
2611 	vm_offset_t lim;
2612 	vm_paddr_t pa;
2613 	vm_size_t len;
2614 
2615 	PMAP_LOCK(pm);
2616 	while (sz > 0) {
2617 		lim = round_page(va);
2618 		len = MIN(lim - va, sz);
2619 		pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2620 		if (pvo != NULL) {
2621 			pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) |
2622 			    (va & ADDR_POFF);
2623 			moea_syncicache(pa, len);
2624 		}
2625 		va += len;
2626 		sz -= len;
2627 	}
2628 	PMAP_UNLOCK(pm);
2629 }
2630 
2631 vm_offset_t
2632 moea_dumpsys_map(mmu_t mmu, struct pmap_md *md, vm_size_t ofs,
2633     vm_size_t *sz)
2634 {
2635 	if (md->md_vaddr == ~0UL)
2636 	    return (md->md_paddr + ofs);
2637 	else
2638 	    return (md->md_vaddr + ofs);
2639 }
2640 
2641 struct pmap_md *
2642 moea_scan_md(mmu_t mmu, struct pmap_md *prev)
2643 {
2644 	static struct pmap_md md;
2645 	struct pvo_entry *pvo;
2646 	vm_offset_t va;
2647 
2648 	if (dumpsys_minidump) {
2649 		md.md_paddr = ~0UL;	/* Minidumps use virtual addresses. */
2650 		if (prev == NULL) {
2651 			/* 1st: kernel .data and .bss. */
2652 			md.md_index = 1;
2653 			md.md_vaddr = trunc_page((uintptr_t)_etext);
2654 			md.md_size = round_page((uintptr_t)_end) - md.md_vaddr;
2655 			return (&md);
2656 		}
2657 		switch (prev->md_index) {
2658 		case 1:
2659 			/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2660 			md.md_index = 2;
2661 			md.md_vaddr = (vm_offset_t)msgbufp->msg_ptr;
2662 			md.md_size = round_page(msgbufp->msg_size);
2663 			break;
2664 		case 2:
2665 			/* 3rd: kernel VM. */
2666 			va = prev->md_vaddr + prev->md_size;
2667 			/* Find start of next chunk (from va). */
2668 			while (va < virtual_end) {
2669 				/* Don't dump the buffer cache. */
2670 				if (va >= kmi.buffer_sva &&
2671 				    va < kmi.buffer_eva) {
2672 					va = kmi.buffer_eva;
2673 					continue;
2674 				}
2675 				pvo = moea_pvo_find_va(kernel_pmap,
2676 				    va & ~ADDR_POFF, NULL);
2677 				if (pvo != NULL &&
2678 				    (pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2679 					break;
2680 				va += PAGE_SIZE;
2681 			}
2682 			if (va < virtual_end) {
2683 				md.md_vaddr = va;
2684 				va += PAGE_SIZE;
2685 				/* Find last page in chunk. */
2686 				while (va < virtual_end) {
2687 					/* Don't run into the buffer cache. */
2688 					if (va == kmi.buffer_sva)
2689 						break;
2690 					pvo = moea_pvo_find_va(kernel_pmap,
2691 					    va & ~ADDR_POFF, NULL);
2692 					if (pvo == NULL ||
2693 					    !(pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2694 						break;
2695 					va += PAGE_SIZE;
2696 				}
2697 				md.md_size = va - md.md_vaddr;
2698 				break;
2699 			}
2700 			md.md_index = 3;
2701 			/* FALLTHROUGH */
2702 		default:
2703 			return (NULL);
2704 		}
2705 	} else { /* minidumps */
2706 		mem_regions(&pregions, &pregions_sz,
2707 		    &regions, &regions_sz);
2708 
2709 		if (prev == NULL) {
2710 			/* first physical chunk. */
2711 			md.md_paddr = pregions[0].mr_start;
2712 			md.md_size = pregions[0].mr_size;
2713 			md.md_vaddr = ~0UL;
2714 			md.md_index = 1;
2715 		} else if (md.md_index < pregions_sz) {
2716 			md.md_paddr = pregions[md.md_index].mr_start;
2717 			md.md_size = pregions[md.md_index].mr_size;
2718 			md.md_vaddr = ~0UL;
2719 			md.md_index++;
2720 		} else {
2721 			/* There's no next physical chunk. */
2722 			return (NULL);
2723 		}
2724 	}
2725 
2726 	return (&md);
2727 }
2728