xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision 3e65b9c6e6b7b2081d54e1dc40983c3c00eaf738)
1 /*-
2  * Copyright (c) 2001 The NetBSD Foundation, Inc.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to The NetBSD Foundation
6  * by Matt Thomas <matt@3am-software.com> of Allegro Networks, Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *        This product includes software developed by the NetBSD
19  *        Foundation, Inc. and its contributors.
20  * 4. Neither the name of The NetBSD Foundation nor the names of its
21  *    contributors may be used to endorse or promote products derived
22  *    from this software without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
25  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
26  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
27  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
28  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
29  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
30  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
31  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
32  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
33  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
34  * POSSIBILITY OF SUCH DAMAGE.
35  */
36 /*-
37  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
38  * Copyright (C) 1995, 1996 TooLs GmbH.
39  * All rights reserved.
40  *
41  * Redistribution and use in source and binary forms, with or without
42  * modification, are permitted provided that the following conditions
43  * are met:
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. All advertising materials mentioning features or use of this software
50  *    must display the following acknowledgement:
51  *	This product includes software developed by TooLs GmbH.
52  * 4. The name of TooLs GmbH may not be used to endorse or promote products
53  *    derived from this software without specific prior written permission.
54  *
55  * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
56  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
57  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
58  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
59  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
60  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
61  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
62  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
63  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
64  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
65  *
66  * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $
67  */
68 /*-
69  * Copyright (C) 2001 Benno Rice.
70  * All rights reserved.
71  *
72  * Redistribution and use in source and binary forms, with or without
73  * modification, are permitted provided that the following conditions
74  * are met:
75  * 1. Redistributions of source code must retain the above copyright
76  *    notice, this list of conditions and the following disclaimer.
77  * 2. Redistributions in binary form must reproduce the above copyright
78  *    notice, this list of conditions and the following disclaimer in the
79  *    documentation and/or other materials provided with the distribution.
80  *
81  * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR
82  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
83  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
84  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
85  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
86  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
87  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
88  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
89  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
90  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
91  */
92 
93 #include <sys/cdefs.h>
94 __FBSDID("$FreeBSD$");
95 
96 /*
97  * Manages physical address maps.
98  *
99  * In addition to hardware address maps, this module is called upon to
100  * provide software-use-only maps which may or may not be stored in the
101  * same form as hardware maps.  These pseudo-maps are used to store
102  * intermediate results from copy operations to and from address spaces.
103  *
104  * Since the information managed by this module is also stored by the
105  * logical address mapping module, this module may throw away valid virtual
106  * to physical mappings at almost any time.  However, invalidations of
107  * mappings must be done as requested.
108  *
109  * In order to cope with hardware architectures which make virtual to
110  * physical map invalidates expensive, this module may delay invalidate
111  * reduced protection operations until such time as they are actually
112  * necessary.  This module is given full information as to which processors
113  * are currently using which maps, and to when physical maps must be made
114  * correct.
115  */
116 
117 #include "opt_kstack_pages.h"
118 
119 #include <sys/param.h>
120 #include <sys/kernel.h>
121 #include <sys/queue.h>
122 #include <sys/cpuset.h>
123 #include <sys/ktr.h>
124 #include <sys/lock.h>
125 #include <sys/msgbuf.h>
126 #include <sys/mutex.h>
127 #include <sys/proc.h>
128 #include <sys/sched.h>
129 #include <sys/sysctl.h>
130 #include <sys/systm.h>
131 #include <sys/vmmeter.h>
132 
133 #include <dev/ofw/openfirm.h>
134 
135 #include <vm/vm.h>
136 #include <vm/vm_param.h>
137 #include <vm/vm_kern.h>
138 #include <vm/vm_page.h>
139 #include <vm/vm_map.h>
140 #include <vm/vm_object.h>
141 #include <vm/vm_extern.h>
142 #include <vm/vm_pageout.h>
143 #include <vm/vm_pager.h>
144 #include <vm/uma.h>
145 
146 #include <machine/cpu.h>
147 #include <machine/platform.h>
148 #include <machine/bat.h>
149 #include <machine/frame.h>
150 #include <machine/md_var.h>
151 #include <machine/psl.h>
152 #include <machine/pte.h>
153 #include <machine/smp.h>
154 #include <machine/sr.h>
155 #include <machine/mmuvar.h>
156 #include <machine/trap_aim.h>
157 
158 #include "mmu_if.h"
159 
160 #define	MOEA_DEBUG
161 
162 #define TODO	panic("%s: not implemented", __func__);
163 
164 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
165 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
166 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
167 
168 struct ofw_map {
169 	vm_offset_t	om_va;
170 	vm_size_t	om_len;
171 	vm_offset_t	om_pa;
172 	u_int		om_mode;
173 };
174 
175 /*
176  * Map of physical memory regions.
177  */
178 static struct	mem_region *regions;
179 static struct	mem_region *pregions;
180 static u_int    phys_avail_count;
181 static int	regions_sz, pregions_sz;
182 static struct	ofw_map *translations;
183 
184 /*
185  * Lock for the pteg and pvo tables.
186  */
187 struct mtx	moea_table_mutex;
188 struct mtx	moea_vsid_mutex;
189 
190 /* tlbie instruction synchronization */
191 static struct mtx tlbie_mtx;
192 
193 /*
194  * PTEG data.
195  */
196 static struct	pteg *moea_pteg_table;
197 u_int		moea_pteg_count;
198 u_int		moea_pteg_mask;
199 
200 /*
201  * PVO data.
202  */
203 struct	pvo_head *moea_pvo_table;		/* pvo entries by pteg index */
204 struct	pvo_head moea_pvo_kunmanaged =
205     LIST_HEAD_INITIALIZER(moea_pvo_kunmanaged);	/* list of unmanaged pages */
206 
207 uma_zone_t	moea_upvo_zone;	/* zone for pvo entries for unmanaged pages */
208 uma_zone_t	moea_mpvo_zone;	/* zone for pvo entries for managed pages */
209 
210 #define	BPVO_POOL_SIZE	32768
211 static struct	pvo_entry *moea_bpvo_pool;
212 static int	moea_bpvo_pool_index = 0;
213 
214 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
215 static u_int	moea_vsid_bitmap[NPMAPS / VSID_NBPW];
216 
217 static boolean_t moea_initialized = FALSE;
218 
219 /*
220  * Statistics.
221  */
222 u_int	moea_pte_valid = 0;
223 u_int	moea_pte_overflow = 0;
224 u_int	moea_pte_replacements = 0;
225 u_int	moea_pvo_entries = 0;
226 u_int	moea_pvo_enter_calls = 0;
227 u_int	moea_pvo_remove_calls = 0;
228 u_int	moea_pte_spills = 0;
229 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_valid, CTLFLAG_RD, &moea_pte_valid,
230     0, "");
231 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_overflow, CTLFLAG_RD,
232     &moea_pte_overflow, 0, "");
233 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_replacements, CTLFLAG_RD,
234     &moea_pte_replacements, 0, "");
235 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_entries, CTLFLAG_RD, &moea_pvo_entries,
236     0, "");
237 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_enter_calls, CTLFLAG_RD,
238     &moea_pvo_enter_calls, 0, "");
239 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_remove_calls, CTLFLAG_RD,
240     &moea_pvo_remove_calls, 0, "");
241 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_spills, CTLFLAG_RD,
242     &moea_pte_spills, 0, "");
243 
244 /*
245  * Allocate physical memory for use in moea_bootstrap.
246  */
247 static vm_offset_t	moea_bootstrap_alloc(vm_size_t, u_int);
248 
249 /*
250  * PTE calls.
251  */
252 static int		moea_pte_insert(u_int, struct pte *);
253 
254 /*
255  * PVO calls.
256  */
257 static int	moea_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *,
258 		    vm_offset_t, vm_offset_t, u_int, int);
259 static void	moea_pvo_remove(struct pvo_entry *, int);
260 static struct	pvo_entry *moea_pvo_find_va(pmap_t, vm_offset_t, int *);
261 static struct	pte *moea_pvo_to_pte(const struct pvo_entry *, int);
262 
263 /*
264  * Utility routines.
265  */
266 static void		moea_enter_locked(pmap_t, vm_offset_t, vm_page_t,
267 			    vm_prot_t, boolean_t);
268 static void		moea_syncicache(vm_offset_t, vm_size_t);
269 static boolean_t	moea_query_bit(vm_page_t, int);
270 static u_int		moea_clear_bit(vm_page_t, int);
271 static void		moea_kremove(mmu_t, vm_offset_t);
272 int		moea_pte_spill(vm_offset_t);
273 
274 /*
275  * Kernel MMU interface
276  */
277 void moea_change_wiring(mmu_t, pmap_t, vm_offset_t, boolean_t);
278 void moea_clear_modify(mmu_t, vm_page_t);
279 void moea_clear_reference(mmu_t, vm_page_t);
280 void moea_copy_page(mmu_t, vm_page_t, vm_page_t);
281 void moea_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, boolean_t);
282 void moea_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
283     vm_prot_t);
284 void moea_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
285 vm_paddr_t moea_extract(mmu_t, pmap_t, vm_offset_t);
286 vm_page_t moea_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
287 void moea_init(mmu_t);
288 boolean_t moea_is_modified(mmu_t, vm_page_t);
289 boolean_t moea_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
290 boolean_t moea_is_referenced(mmu_t, vm_page_t);
291 boolean_t moea_ts_referenced(mmu_t, vm_page_t);
292 vm_offset_t moea_map(mmu_t, vm_offset_t *, vm_offset_t, vm_offset_t, int);
293 boolean_t moea_page_exists_quick(mmu_t, pmap_t, vm_page_t);
294 int moea_page_wired_mappings(mmu_t, vm_page_t);
295 void moea_pinit(mmu_t, pmap_t);
296 void moea_pinit0(mmu_t, pmap_t);
297 void moea_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
298 void moea_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
299 void moea_qremove(mmu_t, vm_offset_t, int);
300 void moea_release(mmu_t, pmap_t);
301 void moea_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
302 void moea_remove_all(mmu_t, vm_page_t);
303 void moea_remove_write(mmu_t, vm_page_t);
304 void moea_zero_page(mmu_t, vm_page_t);
305 void moea_zero_page_area(mmu_t, vm_page_t, int, int);
306 void moea_zero_page_idle(mmu_t, vm_page_t);
307 void moea_activate(mmu_t, struct thread *);
308 void moea_deactivate(mmu_t, struct thread *);
309 void moea_cpu_bootstrap(mmu_t, int);
310 void moea_bootstrap(mmu_t, vm_offset_t, vm_offset_t);
311 void *moea_mapdev(mmu_t, vm_offset_t, vm_size_t);
312 void *moea_mapdev_attr(mmu_t, vm_offset_t, vm_size_t, vm_memattr_t);
313 void moea_unmapdev(mmu_t, vm_offset_t, vm_size_t);
314 vm_offset_t moea_kextract(mmu_t, vm_offset_t);
315 void moea_kenter_attr(mmu_t, vm_offset_t, vm_offset_t, vm_memattr_t);
316 void moea_kenter(mmu_t, vm_offset_t, vm_offset_t);
317 void moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma);
318 boolean_t moea_dev_direct_mapped(mmu_t, vm_offset_t, vm_size_t);
319 static void moea_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
320 
321 static mmu_method_t moea_methods[] = {
322 	MMUMETHOD(mmu_change_wiring,	moea_change_wiring),
323 	MMUMETHOD(mmu_clear_modify,	moea_clear_modify),
324 	MMUMETHOD(mmu_clear_reference,	moea_clear_reference),
325 	MMUMETHOD(mmu_copy_page,	moea_copy_page),
326 	MMUMETHOD(mmu_enter,		moea_enter),
327 	MMUMETHOD(mmu_enter_object,	moea_enter_object),
328 	MMUMETHOD(mmu_enter_quick,	moea_enter_quick),
329 	MMUMETHOD(mmu_extract,		moea_extract),
330 	MMUMETHOD(mmu_extract_and_hold,	moea_extract_and_hold),
331 	MMUMETHOD(mmu_init,		moea_init),
332 	MMUMETHOD(mmu_is_modified,	moea_is_modified),
333 	MMUMETHOD(mmu_is_prefaultable,	moea_is_prefaultable),
334 	MMUMETHOD(mmu_is_referenced,	moea_is_referenced),
335 	MMUMETHOD(mmu_ts_referenced,	moea_ts_referenced),
336 	MMUMETHOD(mmu_map,     		moea_map),
337 	MMUMETHOD(mmu_page_exists_quick,moea_page_exists_quick),
338 	MMUMETHOD(mmu_page_wired_mappings,moea_page_wired_mappings),
339 	MMUMETHOD(mmu_pinit,		moea_pinit),
340 	MMUMETHOD(mmu_pinit0,		moea_pinit0),
341 	MMUMETHOD(mmu_protect,		moea_protect),
342 	MMUMETHOD(mmu_qenter,		moea_qenter),
343 	MMUMETHOD(mmu_qremove,		moea_qremove),
344 	MMUMETHOD(mmu_release,		moea_release),
345 	MMUMETHOD(mmu_remove,		moea_remove),
346 	MMUMETHOD(mmu_remove_all,      	moea_remove_all),
347 	MMUMETHOD(mmu_remove_write,	moea_remove_write),
348 	MMUMETHOD(mmu_sync_icache,	moea_sync_icache),
349 	MMUMETHOD(mmu_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 
367 	{ 0, 0 }
368 };
369 
370 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods, 0);
371 
372 static __inline uint32_t
373 moea_calc_wimg(vm_offset_t pa, vm_memattr_t ma)
374 {
375 	uint32_t pte_lo;
376 	int i;
377 
378 	if (ma != VM_MEMATTR_DEFAULT) {
379 		switch (ma) {
380 		case VM_MEMATTR_UNCACHEABLE:
381 			return (PTE_I | PTE_G);
382 		case VM_MEMATTR_WRITE_COMBINING:
383 		case VM_MEMATTR_WRITE_BACK:
384 		case VM_MEMATTR_PREFETCHABLE:
385 			return (PTE_I);
386 		case VM_MEMATTR_WRITE_THROUGH:
387 			return (PTE_W | PTE_M);
388 		}
389 	}
390 
391 	/*
392 	 * Assume the page is cache inhibited and access is guarded unless
393 	 * it's in our available memory array.
394 	 */
395 	pte_lo = PTE_I | PTE_G;
396 	for (i = 0; i < pregions_sz; i++) {
397 		if ((pa >= pregions[i].mr_start) &&
398 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
399 			pte_lo = PTE_M;
400 			break;
401 		}
402 	}
403 
404 	return pte_lo;
405 }
406 
407 static void
408 tlbie(vm_offset_t va)
409 {
410 
411 	mtx_lock_spin(&tlbie_mtx);
412 	__asm __volatile("ptesync");
413 	__asm __volatile("tlbie %0" :: "r"(va));
414 	__asm __volatile("eieio; tlbsync; ptesync");
415 	mtx_unlock_spin(&tlbie_mtx);
416 }
417 
418 static void
419 tlbia(void)
420 {
421 	vm_offset_t va;
422 
423 	for (va = 0; va < 0x00040000; va += 0x00001000) {
424 		__asm __volatile("tlbie %0" :: "r"(va));
425 		powerpc_sync();
426 	}
427 	__asm __volatile("tlbsync");
428 	powerpc_sync();
429 }
430 
431 static __inline int
432 va_to_sr(u_int *sr, vm_offset_t va)
433 {
434 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
435 }
436 
437 static __inline u_int
438 va_to_pteg(u_int sr, vm_offset_t addr)
439 {
440 	u_int hash;
441 
442 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
443 	    ADDR_PIDX_SHFT);
444 	return (hash & moea_pteg_mask);
445 }
446 
447 static __inline struct pvo_head *
448 vm_page_to_pvoh(vm_page_t m)
449 {
450 
451 	return (&m->md.mdpg_pvoh);
452 }
453 
454 static __inline void
455 moea_attr_clear(vm_page_t m, int ptebit)
456 {
457 
458 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
459 	m->md.mdpg_attrs &= ~ptebit;
460 }
461 
462 static __inline int
463 moea_attr_fetch(vm_page_t m)
464 {
465 
466 	return (m->md.mdpg_attrs);
467 }
468 
469 static __inline void
470 moea_attr_save(vm_page_t m, int ptebit)
471 {
472 
473 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
474 	m->md.mdpg_attrs |= ptebit;
475 }
476 
477 static __inline int
478 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
479 {
480 	if (pt->pte_hi == pvo_pt->pte_hi)
481 		return (1);
482 
483 	return (0);
484 }
485 
486 static __inline int
487 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
488 {
489 	return (pt->pte_hi & ~PTE_VALID) ==
490 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
491 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
492 }
493 
494 static __inline void
495 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
496 {
497 
498 	mtx_assert(&moea_table_mutex, MA_OWNED);
499 
500 	/*
501 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
502 	 * set when the real pte is set in memory.
503 	 *
504 	 * Note: Don't set the valid bit for correct operation of tlb update.
505 	 */
506 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
507 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
508 	pt->pte_lo = pte_lo;
509 }
510 
511 static __inline void
512 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
513 {
514 
515 	mtx_assert(&moea_table_mutex, MA_OWNED);
516 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
517 }
518 
519 static __inline void
520 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
521 {
522 
523 	mtx_assert(&moea_table_mutex, MA_OWNED);
524 
525 	/*
526 	 * As shown in Section 7.6.3.2.3
527 	 */
528 	pt->pte_lo &= ~ptebit;
529 	tlbie(va);
530 }
531 
532 static __inline void
533 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
534 {
535 
536 	mtx_assert(&moea_table_mutex, MA_OWNED);
537 	pvo_pt->pte_hi |= PTE_VALID;
538 
539 	/*
540 	 * Update the PTE as defined in section 7.6.3.1.
541 	 * Note that the REF/CHG bits are from pvo_pt and thus should havce
542 	 * been saved so this routine can restore them (if desired).
543 	 */
544 	pt->pte_lo = pvo_pt->pte_lo;
545 	powerpc_sync();
546 	pt->pte_hi = pvo_pt->pte_hi;
547 	powerpc_sync();
548 	moea_pte_valid++;
549 }
550 
551 static __inline void
552 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
553 {
554 
555 	mtx_assert(&moea_table_mutex, MA_OWNED);
556 	pvo_pt->pte_hi &= ~PTE_VALID;
557 
558 	/*
559 	 * Force the reg & chg bits back into the PTEs.
560 	 */
561 	powerpc_sync();
562 
563 	/*
564 	 * Invalidate the pte.
565 	 */
566 	pt->pte_hi &= ~PTE_VALID;
567 
568 	tlbie(va);
569 
570 	/*
571 	 * Save the reg & chg bits.
572 	 */
573 	moea_pte_synch(pt, pvo_pt);
574 	moea_pte_valid--;
575 }
576 
577 static __inline void
578 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
579 {
580 
581 	/*
582 	 * Invalidate the PTE
583 	 */
584 	moea_pte_unset(pt, pvo_pt, va);
585 	moea_pte_set(pt, pvo_pt);
586 }
587 
588 /*
589  * Quick sort callout for comparing memory regions.
590  */
591 static int	om_cmp(const void *a, const void *b);
592 
593 static int
594 om_cmp(const void *a, const void *b)
595 {
596 	const struct	ofw_map *mapa;
597 	const struct	ofw_map *mapb;
598 
599 	mapa = a;
600 	mapb = b;
601 	if (mapa->om_pa < mapb->om_pa)
602 		return (-1);
603 	else if (mapa->om_pa > mapb->om_pa)
604 		return (1);
605 	else
606 		return (0);
607 }
608 
609 void
610 moea_cpu_bootstrap(mmu_t mmup, int ap)
611 {
612 	u_int sdr;
613 	int i;
614 
615 	if (ap) {
616 		powerpc_sync();
617 		__asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu));
618 		__asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl));
619 		isync();
620 		__asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu));
621 		__asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl));
622 		isync();
623 	}
624 
625 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
626 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
627 	isync();
628 
629 	__asm __volatile("mtibatu 1,%0" :: "r"(0));
630 	__asm __volatile("mtdbatu 2,%0" :: "r"(0));
631 	__asm __volatile("mtibatu 2,%0" :: "r"(0));
632 	__asm __volatile("mtdbatu 3,%0" :: "r"(0));
633 	__asm __volatile("mtibatu 3,%0" :: "r"(0));
634 	isync();
635 
636 	for (i = 0; i < 16; i++)
637 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
638 	powerpc_sync();
639 
640 	sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10);
641 	__asm __volatile("mtsdr1 %0" :: "r"(sdr));
642 	isync();
643 
644 	tlbia();
645 }
646 
647 void
648 moea_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
649 {
650 	ihandle_t	mmui;
651 	phandle_t	chosen, mmu;
652 	int		sz;
653 	int		i, j;
654 	vm_size_t	size, physsz, hwphyssz;
655 	vm_offset_t	pa, va, off;
656 	void		*dpcpu;
657 	register_t	msr;
658 
659         /*
660          * Set up BAT0 to map the lowest 256 MB area
661          */
662         battable[0x0].batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
663         battable[0x0].batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
664 
665         /*
666          * Map PCI memory space.
667          */
668         battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
669         battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
670 
671         battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
672         battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
673 
674         battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
675         battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
676 
677         battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
678         battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
679 
680         /*
681          * Map obio devices.
682          */
683         battable[0xf].batl = BATL(0xf0000000, BAT_I|BAT_G, BAT_PP_RW);
684         battable[0xf].batu = BATU(0xf0000000, BAT_BL_256M, BAT_Vs);
685 
686 	/*
687 	 * Use an IBAT and a DBAT to map the bottom segment of memory
688 	 * where we are. Turn off instruction relocation temporarily
689 	 * to prevent faults while reprogramming the IBAT.
690 	 */
691 	msr = mfmsr();
692 	mtmsr(msr & ~PSL_IR);
693 	__asm (".balign 32; \n"
694 	       "mtibatu 0,%0; mtibatl 0,%1; isync; \n"
695 	       "mtdbatu 0,%0; mtdbatl 0,%1; isync"
696 	    :: "r"(battable[0].batu), "r"(battable[0].batl));
697 	mtmsr(msr);
698 
699 	/* map pci space */
700 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
701 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
702 	isync();
703 
704 	/* set global direct map flag */
705 	hw_direct_map = 1;
706 
707 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
708 	CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
709 
710 	for (i = 0; i < pregions_sz; i++) {
711 		vm_offset_t pa;
712 		vm_offset_t end;
713 
714 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
715 			pregions[i].mr_start,
716 			pregions[i].mr_start + pregions[i].mr_size,
717 			pregions[i].mr_size);
718 		/*
719 		 * Install entries into the BAT table to allow all
720 		 * of physmem to be convered by on-demand BAT entries.
721 		 * The loop will sometimes set the same battable element
722 		 * twice, but that's fine since they won't be used for
723 		 * a while yet.
724 		 */
725 		pa = pregions[i].mr_start & 0xf0000000;
726 		end = pregions[i].mr_start + pregions[i].mr_size;
727 		do {
728                         u_int n = pa >> ADDR_SR_SHFT;
729 
730 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
731 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
732 			pa += SEGMENT_LENGTH;
733 		} while (pa < end);
734 	}
735 
736 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
737 		panic("moea_bootstrap: phys_avail too small");
738 
739 	phys_avail_count = 0;
740 	physsz = 0;
741 	hwphyssz = 0;
742 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
743 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
744 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
745 		    regions[i].mr_start + regions[i].mr_size,
746 		    regions[i].mr_size);
747 		if (hwphyssz != 0 &&
748 		    (physsz + regions[i].mr_size) >= hwphyssz) {
749 			if (physsz < hwphyssz) {
750 				phys_avail[j] = regions[i].mr_start;
751 				phys_avail[j + 1] = regions[i].mr_start +
752 				    hwphyssz - physsz;
753 				physsz = hwphyssz;
754 				phys_avail_count++;
755 			}
756 			break;
757 		}
758 		phys_avail[j] = regions[i].mr_start;
759 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
760 		phys_avail_count++;
761 		physsz += regions[i].mr_size;
762 	}
763 
764 	/* Check for overlap with the kernel and exception vectors */
765 	for (j = 0; j < 2*phys_avail_count; j+=2) {
766 		if (phys_avail[j] < EXC_LAST)
767 			phys_avail[j] += EXC_LAST;
768 
769 		if (kernelstart >= phys_avail[j] &&
770 		    kernelstart < phys_avail[j+1]) {
771 			if (kernelend < phys_avail[j+1]) {
772 				phys_avail[2*phys_avail_count] =
773 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
774 				phys_avail[2*phys_avail_count + 1] =
775 				    phys_avail[j+1];
776 				phys_avail_count++;
777 			}
778 
779 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
780 		}
781 
782 		if (kernelend >= phys_avail[j] &&
783 		    kernelend < phys_avail[j+1]) {
784 			if (kernelstart > phys_avail[j]) {
785 				phys_avail[2*phys_avail_count] = phys_avail[j];
786 				phys_avail[2*phys_avail_count + 1] =
787 				    kernelstart & ~PAGE_MASK;
788 				phys_avail_count++;
789 			}
790 
791 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
792 		}
793 	}
794 
795 	physmem = btoc(physsz);
796 
797 	/*
798 	 * Allocate PTEG table.
799 	 */
800 #ifdef PTEGCOUNT
801 	moea_pteg_count = PTEGCOUNT;
802 #else
803 	moea_pteg_count = 0x1000;
804 
805 	while (moea_pteg_count < physmem)
806 		moea_pteg_count <<= 1;
807 
808 	moea_pteg_count >>= 1;
809 #endif /* PTEGCOUNT */
810 
811 	size = moea_pteg_count * sizeof(struct pteg);
812 	CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
813 	    size);
814 	moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
815 	CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
816 	bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
817 	moea_pteg_mask = moea_pteg_count - 1;
818 
819 	/*
820 	 * Allocate pv/overflow lists.
821 	 */
822 	size = sizeof(struct pvo_head) * moea_pteg_count;
823 	moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
824 	    PAGE_SIZE);
825 	CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
826 	for (i = 0; i < moea_pteg_count; i++)
827 		LIST_INIT(&moea_pvo_table[i]);
828 
829 	/*
830 	 * Initialize the lock that synchronizes access to the pteg and pvo
831 	 * tables.
832 	 */
833 	mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF |
834 	    MTX_RECURSE);
835 	mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF);
836 
837 	mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN);
838 
839 	/*
840 	 * Initialise the unmanaged pvo pool.
841 	 */
842 	moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
843 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
844 	moea_bpvo_pool_index = 0;
845 
846 	/*
847 	 * Make sure kernel vsid is allocated as well as VSID 0.
848 	 */
849 	moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
850 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
851 	moea_vsid_bitmap[0] |= 1;
852 
853 	/*
854 	 * Initialize the kernel pmap (which is statically allocated).
855 	 */
856 	PMAP_LOCK_INIT(kernel_pmap);
857 	for (i = 0; i < 16; i++)
858 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
859 	CPU_FILL(&kernel_pmap->pm_active);
860 	LIST_INIT(&kernel_pmap->pmap_pvo);
861 
862 	/*
863 	 * Set up the Open Firmware mappings
864 	 */
865 	chosen = OF_finddevice("/chosen");
866 	if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1 &&
867 	    (mmu = OF_instance_to_package(mmui)) != -1 &&
868 	    (sz = OF_getproplen(mmu, "translations")) != -1) {
869 		translations = NULL;
870 		for (i = 0; phys_avail[i] != 0; i += 2) {
871 			if (phys_avail[i + 1] >= sz) {
872 				translations = (struct ofw_map *)phys_avail[i];
873 				break;
874 			}
875 		}
876 		if (translations == NULL)
877 			panic("moea_bootstrap: no space to copy translations");
878 		bzero(translations, sz);
879 		if (OF_getprop(mmu, "translations", translations, sz) == -1)
880 			panic("moea_bootstrap: can't get ofw translations");
881 		CTR0(KTR_PMAP, "moea_bootstrap: translations");
882 		sz /= sizeof(*translations);
883 		qsort(translations, sz, sizeof (*translations), om_cmp);
884 		for (i = 0; i < sz; i++) {
885 			CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
886 			    translations[i].om_pa, translations[i].om_va,
887 			    translations[i].om_len);
888 
889 			/*
890 			 * If the mapping is 1:1, let the RAM and device
891 			 * on-demand BAT tables take care of the translation.
892 			 */
893 			if (translations[i].om_va == translations[i].om_pa)
894 				continue;
895 
896 			/* Enter the pages */
897 			for (off = 0; off < translations[i].om_len;
898 			    off += PAGE_SIZE)
899 				moea_kenter(mmup, translations[i].om_va + off,
900 					    translations[i].om_pa + off);
901 		}
902 	}
903 
904 	/*
905 	 * Calculate the last available physical address.
906 	 */
907 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
908 		;
909 	Maxmem = powerpc_btop(phys_avail[i + 1]);
910 
911 	moea_cpu_bootstrap(mmup,0);
912 
913 	pmap_bootstrapped++;
914 
915 	/*
916 	 * Set the start and end of kva.
917 	 */
918 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
919 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
920 
921 	/*
922 	 * Allocate a kernel stack with a guard page for thread0 and map it
923 	 * into the kernel page map.
924 	 */
925 	pa = moea_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, PAGE_SIZE);
926 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
927 	virtual_avail = va + KSTACK_PAGES * PAGE_SIZE;
928 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
929 	thread0.td_kstack = va;
930 	thread0.td_kstack_pages = KSTACK_PAGES;
931 	for (i = 0; i < KSTACK_PAGES; i++) {
932 		moea_kenter(mmup, va, pa);
933 		pa += PAGE_SIZE;
934 		va += PAGE_SIZE;
935 	}
936 
937 	/*
938 	 * Allocate virtual address space for the message buffer.
939 	 */
940 	pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE);
941 	msgbufp = (struct msgbuf *)virtual_avail;
942 	va = virtual_avail;
943 	virtual_avail += round_page(msgbufsize);
944 	while (va < virtual_avail) {
945 		moea_kenter(mmup, va, pa);
946 		pa += PAGE_SIZE;
947 		va += PAGE_SIZE;
948 	}
949 
950 	/*
951 	 * Allocate virtual address space for the dynamic percpu area.
952 	 */
953 	pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
954 	dpcpu = (void *)virtual_avail;
955 	va = virtual_avail;
956 	virtual_avail += DPCPU_SIZE;
957 	while (va < virtual_avail) {
958 		moea_kenter(mmup, va, pa);
959 		pa += PAGE_SIZE;
960 		va += PAGE_SIZE;
961 	}
962 	dpcpu_init(dpcpu, 0);
963 }
964 
965 /*
966  * Activate a user pmap.  The pmap must be activated before it's address
967  * space can be accessed in any way.
968  */
969 void
970 moea_activate(mmu_t mmu, struct thread *td)
971 {
972 	pmap_t	pm, pmr;
973 
974 	/*
975 	 * Load all the data we need up front to encourage the compiler to
976 	 * not issue any loads while we have interrupts disabled below.
977 	 */
978 	pm = &td->td_proc->p_vmspace->vm_pmap;
979 	pmr = pm->pmap_phys;
980 
981 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
982 	PCPU_SET(curpmap, pmr);
983 }
984 
985 void
986 moea_deactivate(mmu_t mmu, struct thread *td)
987 {
988 	pmap_t	pm;
989 
990 	pm = &td->td_proc->p_vmspace->vm_pmap;
991 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
992 	PCPU_SET(curpmap, NULL);
993 }
994 
995 void
996 moea_change_wiring(mmu_t mmu, pmap_t pm, vm_offset_t va, boolean_t wired)
997 {
998 	struct	pvo_entry *pvo;
999 
1000 	PMAP_LOCK(pm);
1001 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1002 
1003 	if (pvo != NULL) {
1004 		if (wired) {
1005 			if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1006 				pm->pm_stats.wired_count++;
1007 			pvo->pvo_vaddr |= PVO_WIRED;
1008 		} else {
1009 			if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1010 				pm->pm_stats.wired_count--;
1011 			pvo->pvo_vaddr &= ~PVO_WIRED;
1012 		}
1013 	}
1014 	PMAP_UNLOCK(pm);
1015 }
1016 
1017 void
1018 moea_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1019 {
1020 	vm_offset_t	dst;
1021 	vm_offset_t	src;
1022 
1023 	dst = VM_PAGE_TO_PHYS(mdst);
1024 	src = VM_PAGE_TO_PHYS(msrc);
1025 
1026 	kcopy((void *)src, (void *)dst, PAGE_SIZE);
1027 }
1028 
1029 /*
1030  * Zero a page of physical memory by temporarily mapping it into the tlb.
1031  */
1032 void
1033 moea_zero_page(mmu_t mmu, vm_page_t m)
1034 {
1035 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1036 	void *va = (void *)pa;
1037 
1038 	bzero(va, PAGE_SIZE);
1039 }
1040 
1041 void
1042 moea_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1043 {
1044 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1045 	void *va = (void *)(pa + off);
1046 
1047 	bzero(va, size);
1048 }
1049 
1050 void
1051 moea_zero_page_idle(mmu_t mmu, vm_page_t m)
1052 {
1053 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1054 	void *va = (void *)pa;
1055 
1056 	bzero(va, PAGE_SIZE);
1057 }
1058 
1059 /*
1060  * Map the given physical page at the specified virtual address in the
1061  * target pmap with the protection requested.  If specified the page
1062  * will be wired down.
1063  */
1064 void
1065 moea_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1066 	   boolean_t wired)
1067 {
1068 
1069 	vm_page_lock_queues();
1070 	PMAP_LOCK(pmap);
1071 	moea_enter_locked(pmap, va, m, prot, wired);
1072 	vm_page_unlock_queues();
1073 	PMAP_UNLOCK(pmap);
1074 }
1075 
1076 /*
1077  * Map the given physical page at the specified virtual address in the
1078  * target pmap with the protection requested.  If specified the page
1079  * will be wired down.
1080  *
1081  * The page queues and pmap must be locked.
1082  */
1083 static void
1084 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1085     boolean_t wired)
1086 {
1087 	struct		pvo_head *pvo_head;
1088 	uma_zone_t	zone;
1089 	vm_page_t	pg;
1090 	u_int		pte_lo, pvo_flags, was_exec;
1091 	int		error;
1092 
1093 	if (!moea_initialized) {
1094 		pvo_head = &moea_pvo_kunmanaged;
1095 		zone = moea_upvo_zone;
1096 		pvo_flags = 0;
1097 		pg = NULL;
1098 		was_exec = PTE_EXEC;
1099 	} else {
1100 		pvo_head = vm_page_to_pvoh(m);
1101 		pg = m;
1102 		zone = moea_mpvo_zone;
1103 		pvo_flags = PVO_MANAGED;
1104 		was_exec = 0;
1105 	}
1106 	if (pmap_bootstrapped)
1107 		mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1108 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1109 	KASSERT((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) != 0 ||
1110 	    VM_OBJECT_LOCKED(m->object),
1111 	    ("moea_enter_locked: page %p is not busy", m));
1112 
1113 	/* XXX change the pvo head for fake pages */
1114 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1115 		pvo_flags &= ~PVO_MANAGED;
1116 		pvo_head = &moea_pvo_kunmanaged;
1117 		zone = moea_upvo_zone;
1118 	}
1119 
1120 	/*
1121 	 * If this is a managed page, and it's the first reference to the page,
1122 	 * clear the execness of the page.  Otherwise fetch the execness.
1123 	 */
1124 	if ((pg != NULL) && ((m->oflags & VPO_UNMANAGED) == 0)) {
1125 		if (LIST_EMPTY(pvo_head)) {
1126 			moea_attr_clear(pg, PTE_EXEC);
1127 		} else {
1128 			was_exec = moea_attr_fetch(pg) & PTE_EXEC;
1129 		}
1130 	}
1131 
1132 	pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1133 
1134 	if (prot & VM_PROT_WRITE) {
1135 		pte_lo |= PTE_BW;
1136 		if (pmap_bootstrapped &&
1137 		    (m->oflags & VPO_UNMANAGED) == 0)
1138 			vm_page_aflag_set(m, PGA_WRITEABLE);
1139 	} else
1140 		pte_lo |= PTE_BR;
1141 
1142 	if (prot & VM_PROT_EXECUTE)
1143 		pvo_flags |= PVO_EXECUTABLE;
1144 
1145 	if (wired)
1146 		pvo_flags |= PVO_WIRED;
1147 
1148 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1149 	    pte_lo, pvo_flags);
1150 
1151 	/*
1152 	 * Flush the real page from the instruction cache if this page is
1153 	 * mapped executable and cacheable and was not previously mapped (or
1154 	 * was not mapped executable).
1155 	 */
1156 	if (error == 0 && (pvo_flags & PVO_EXECUTABLE) &&
1157 	    (pte_lo & PTE_I) == 0 && was_exec == 0) {
1158 		/*
1159 		 * Flush the real memory from the cache.
1160 		 */
1161 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1162 		if (pg != NULL)
1163 			moea_attr_save(pg, PTE_EXEC);
1164 	}
1165 
1166 	/* XXX syncicache always until problems are sorted */
1167 	moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1168 }
1169 
1170 /*
1171  * Maps a sequence of resident pages belonging to the same object.
1172  * The sequence begins with the given page m_start.  This page is
1173  * mapped at the given virtual address start.  Each subsequent page is
1174  * mapped at a virtual address that is offset from start by the same
1175  * amount as the page is offset from m_start within the object.  The
1176  * last page in the sequence is the page with the largest offset from
1177  * m_start that can be mapped at a virtual address less than the given
1178  * virtual address end.  Not every virtual page between start and end
1179  * is mapped; only those for which a resident page exists with the
1180  * corresponding offset from m_start are mapped.
1181  */
1182 void
1183 moea_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1184     vm_page_t m_start, vm_prot_t prot)
1185 {
1186 	vm_page_t m;
1187 	vm_pindex_t diff, psize;
1188 
1189 	psize = atop(end - start);
1190 	m = m_start;
1191 	vm_page_lock_queues();
1192 	PMAP_LOCK(pm);
1193 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1194 		moea_enter_locked(pm, start + ptoa(diff), m, prot &
1195 		    (VM_PROT_READ | VM_PROT_EXECUTE), FALSE);
1196 		m = TAILQ_NEXT(m, listq);
1197 	}
1198 	vm_page_unlock_queues();
1199 	PMAP_UNLOCK(pm);
1200 }
1201 
1202 void
1203 moea_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1204     vm_prot_t prot)
1205 {
1206 
1207 	vm_page_lock_queues();
1208 	PMAP_LOCK(pm);
1209 	moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1210 	    FALSE);
1211 	vm_page_unlock_queues();
1212 	PMAP_UNLOCK(pm);
1213 }
1214 
1215 vm_paddr_t
1216 moea_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1217 {
1218 	struct	pvo_entry *pvo;
1219 	vm_paddr_t pa;
1220 
1221 	PMAP_LOCK(pm);
1222 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1223 	if (pvo == NULL)
1224 		pa = 0;
1225 	else
1226 		pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1227 	PMAP_UNLOCK(pm);
1228 	return (pa);
1229 }
1230 
1231 /*
1232  * Atomically extract and hold the physical page with the given
1233  * pmap and virtual address pair if that mapping permits the given
1234  * protection.
1235  */
1236 vm_page_t
1237 moea_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1238 {
1239 	struct	pvo_entry *pvo;
1240 	vm_page_t m;
1241         vm_paddr_t pa;
1242 
1243 	m = NULL;
1244 	pa = 0;
1245 	PMAP_LOCK(pmap);
1246 retry:
1247 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1248 	if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1249 	    ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1250 	     (prot & VM_PROT_WRITE) == 0)) {
1251 		if (vm_page_pa_tryrelock(pmap, pvo->pvo_pte.pte.pte_lo & PTE_RPGN, &pa))
1252 			goto retry;
1253 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
1254 		vm_page_hold(m);
1255 	}
1256 	PA_UNLOCK_COND(pa);
1257 	PMAP_UNLOCK(pmap);
1258 	return (m);
1259 }
1260 
1261 void
1262 moea_init(mmu_t mmu)
1263 {
1264 
1265 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1266 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1267 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1268 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1269 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1270 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1271 	moea_initialized = TRUE;
1272 }
1273 
1274 boolean_t
1275 moea_is_referenced(mmu_t mmu, vm_page_t m)
1276 {
1277 
1278 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1279 	    ("moea_is_referenced: page %p is not managed", m));
1280 	return (moea_query_bit(m, PTE_REF));
1281 }
1282 
1283 boolean_t
1284 moea_is_modified(mmu_t mmu, vm_page_t m)
1285 {
1286 
1287 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1288 	    ("moea_is_modified: page %p is not managed", m));
1289 
1290 	/*
1291 	 * If the page is not VPO_BUSY, then PGA_WRITEABLE cannot be
1292 	 * concurrently set while the object is locked.  Thus, if PGA_WRITEABLE
1293 	 * is clear, no PTEs can have PTE_CHG set.
1294 	 */
1295 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1296 	if ((m->oflags & VPO_BUSY) == 0 &&
1297 	    (m->aflags & PGA_WRITEABLE) == 0)
1298 		return (FALSE);
1299 	return (moea_query_bit(m, PTE_CHG));
1300 }
1301 
1302 boolean_t
1303 moea_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1304 {
1305 	struct pvo_entry *pvo;
1306 	boolean_t rv;
1307 
1308 	PMAP_LOCK(pmap);
1309 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1310 	rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1311 	PMAP_UNLOCK(pmap);
1312 	return (rv);
1313 }
1314 
1315 void
1316 moea_clear_reference(mmu_t mmu, vm_page_t m)
1317 {
1318 
1319 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1320 	    ("moea_clear_reference: page %p is not managed", m));
1321 	moea_clear_bit(m, PTE_REF);
1322 }
1323 
1324 void
1325 moea_clear_modify(mmu_t mmu, vm_page_t m)
1326 {
1327 
1328 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1329 	    ("moea_clear_modify: page %p is not managed", m));
1330 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1331 	KASSERT((m->oflags & VPO_BUSY) == 0,
1332 	    ("moea_clear_modify: page %p is busy", m));
1333 
1334 	/*
1335 	 * If the page is not PGA_WRITEABLE, then no PTEs can have PTE_CHG
1336 	 * set.  If the object containing the page is locked and the page is
1337 	 * not VPO_BUSY, then PGA_WRITEABLE cannot be concurrently set.
1338 	 */
1339 	if ((m->aflags & PGA_WRITEABLE) == 0)
1340 		return;
1341 	moea_clear_bit(m, PTE_CHG);
1342 }
1343 
1344 /*
1345  * Clear the write and modified bits in each of the given page's mappings.
1346  */
1347 void
1348 moea_remove_write(mmu_t mmu, vm_page_t m)
1349 {
1350 	struct	pvo_entry *pvo;
1351 	struct	pte *pt;
1352 	pmap_t	pmap;
1353 	u_int	lo;
1354 
1355 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1356 	    ("moea_remove_write: page %p is not managed", m));
1357 
1358 	/*
1359 	 * If the page is not VPO_BUSY, then PGA_WRITEABLE cannot be set by
1360 	 * another thread while the object is locked.  Thus, if PGA_WRITEABLE
1361 	 * is clear, no page table entries need updating.
1362 	 */
1363 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1364 	if ((m->oflags & VPO_BUSY) == 0 &&
1365 	    (m->aflags & PGA_WRITEABLE) == 0)
1366 		return;
1367 	vm_page_lock_queues();
1368 	lo = moea_attr_fetch(m);
1369 	powerpc_sync();
1370 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1371 		pmap = pvo->pvo_pmap;
1372 		PMAP_LOCK(pmap);
1373 		if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1374 			pt = moea_pvo_to_pte(pvo, -1);
1375 			pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1376 			pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1377 			if (pt != NULL) {
1378 				moea_pte_synch(pt, &pvo->pvo_pte.pte);
1379 				lo |= pvo->pvo_pte.pte.pte_lo;
1380 				pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1381 				moea_pte_change(pt, &pvo->pvo_pte.pte,
1382 				    pvo->pvo_vaddr);
1383 				mtx_unlock(&moea_table_mutex);
1384 			}
1385 		}
1386 		PMAP_UNLOCK(pmap);
1387 	}
1388 	if ((lo & PTE_CHG) != 0) {
1389 		moea_attr_clear(m, PTE_CHG);
1390 		vm_page_dirty(m);
1391 	}
1392 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1393 	vm_page_unlock_queues();
1394 }
1395 
1396 /*
1397  *	moea_ts_referenced:
1398  *
1399  *	Return a count of reference bits for a page, clearing those bits.
1400  *	It is not necessary for every reference bit to be cleared, but it
1401  *	is necessary that 0 only be returned when there are truly no
1402  *	reference bits set.
1403  *
1404  *	XXX: The exact number of bits to check and clear is a matter that
1405  *	should be tested and standardized at some point in the future for
1406  *	optimal aging of shared pages.
1407  */
1408 boolean_t
1409 moea_ts_referenced(mmu_t mmu, vm_page_t m)
1410 {
1411 
1412 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1413 	    ("moea_ts_referenced: page %p is not managed", m));
1414 	return (moea_clear_bit(m, PTE_REF));
1415 }
1416 
1417 /*
1418  * Modify the WIMG settings of all mappings for a page.
1419  */
1420 void
1421 moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1422 {
1423 	struct	pvo_entry *pvo;
1424 	struct	pvo_head *pvo_head;
1425 	struct	pte *pt;
1426 	pmap_t	pmap;
1427 	u_int	lo;
1428 
1429 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1430 		m->md.mdpg_cache_attrs = ma;
1431 		return;
1432 	}
1433 
1434 	vm_page_lock_queues();
1435 	pvo_head = vm_page_to_pvoh(m);
1436 	lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1437 
1438 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1439 		pmap = pvo->pvo_pmap;
1440 		PMAP_LOCK(pmap);
1441 		pt = moea_pvo_to_pte(pvo, -1);
1442 		pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1443 		pvo->pvo_pte.pte.pte_lo |= lo;
1444 		if (pt != NULL) {
1445 			moea_pte_change(pt, &pvo->pvo_pte.pte,
1446 			    pvo->pvo_vaddr);
1447 			if (pvo->pvo_pmap == kernel_pmap)
1448 				isync();
1449 		}
1450 		mtx_unlock(&moea_table_mutex);
1451 		PMAP_UNLOCK(pmap);
1452 	}
1453 	m->md.mdpg_cache_attrs = ma;
1454 	vm_page_unlock_queues();
1455 }
1456 
1457 /*
1458  * Map a wired page into kernel virtual address space.
1459  */
1460 void
1461 moea_kenter(mmu_t mmu, vm_offset_t va, vm_offset_t pa)
1462 {
1463 
1464 	moea_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1465 }
1466 
1467 void
1468 moea_kenter_attr(mmu_t mmu, vm_offset_t va, vm_offset_t pa, vm_memattr_t ma)
1469 {
1470 	u_int		pte_lo;
1471 	int		error;
1472 
1473 #if 0
1474 	if (va < VM_MIN_KERNEL_ADDRESS)
1475 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1476 		    va);
1477 #endif
1478 
1479 	pte_lo = moea_calc_wimg(pa, ma);
1480 
1481 	PMAP_LOCK(kernel_pmap);
1482 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1483 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1484 
1485 	if (error != 0 && error != ENOENT)
1486 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1487 		    pa, error);
1488 
1489 	/*
1490 	 * Flush the real memory from the instruction cache.
1491 	 */
1492 	if ((pte_lo & (PTE_I | PTE_G)) == 0) {
1493 		moea_syncicache(pa, PAGE_SIZE);
1494 	}
1495 	PMAP_UNLOCK(kernel_pmap);
1496 }
1497 
1498 /*
1499  * Extract the physical page address associated with the given kernel virtual
1500  * address.
1501  */
1502 vm_offset_t
1503 moea_kextract(mmu_t mmu, vm_offset_t va)
1504 {
1505 	struct		pvo_entry *pvo;
1506 	vm_paddr_t pa;
1507 
1508 	/*
1509 	 * Allow direct mappings on 32-bit OEA
1510 	 */
1511 	if (va < VM_MIN_KERNEL_ADDRESS) {
1512 		return (va);
1513 	}
1514 
1515 	PMAP_LOCK(kernel_pmap);
1516 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1517 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1518 	pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1519 	PMAP_UNLOCK(kernel_pmap);
1520 	return (pa);
1521 }
1522 
1523 /*
1524  * Remove a wired page from kernel virtual address space.
1525  */
1526 void
1527 moea_kremove(mmu_t mmu, vm_offset_t va)
1528 {
1529 
1530 	moea_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1531 }
1532 
1533 /*
1534  * Map a range of physical addresses into kernel virtual address space.
1535  *
1536  * The value passed in *virt is a suggested virtual address for the mapping.
1537  * Architectures which can support a direct-mapped physical to virtual region
1538  * can return the appropriate address within that region, leaving '*virt'
1539  * unchanged.  We cannot and therefore do not; *virt is updated with the
1540  * first usable address after the mapped region.
1541  */
1542 vm_offset_t
1543 moea_map(mmu_t mmu, vm_offset_t *virt, vm_offset_t pa_start,
1544     vm_offset_t pa_end, int prot)
1545 {
1546 	vm_offset_t	sva, va;
1547 
1548 	sva = *virt;
1549 	va = sva;
1550 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1551 		moea_kenter(mmu, va, pa_start);
1552 	*virt = va;
1553 	return (sva);
1554 }
1555 
1556 /*
1557  * Returns true if the pmap's pv is one of the first
1558  * 16 pvs linked to from this page.  This count may
1559  * be changed upwards or downwards in the future; it
1560  * is only necessary that true be returned for a small
1561  * subset of pmaps for proper page aging.
1562  */
1563 boolean_t
1564 moea_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1565 {
1566         int loops;
1567 	struct pvo_entry *pvo;
1568 	boolean_t rv;
1569 
1570 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1571 	    ("moea_page_exists_quick: page %p is not managed", m));
1572 	loops = 0;
1573 	rv = FALSE;
1574 	vm_page_lock_queues();
1575 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1576 		if (pvo->pvo_pmap == pmap) {
1577 			rv = TRUE;
1578 			break;
1579 		}
1580 		if (++loops >= 16)
1581 			break;
1582 	}
1583 	vm_page_unlock_queues();
1584 	return (rv);
1585 }
1586 
1587 /*
1588  * Return the number of managed mappings to the given physical page
1589  * that are wired.
1590  */
1591 int
1592 moea_page_wired_mappings(mmu_t mmu, vm_page_t m)
1593 {
1594 	struct pvo_entry *pvo;
1595 	int count;
1596 
1597 	count = 0;
1598 	if ((m->oflags & VPO_UNMANAGED) != 0)
1599 		return (count);
1600 	vm_page_lock_queues();
1601 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1602 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1603 			count++;
1604 	vm_page_unlock_queues();
1605 	return (count);
1606 }
1607 
1608 static u_int	moea_vsidcontext;
1609 
1610 void
1611 moea_pinit(mmu_t mmu, pmap_t pmap)
1612 {
1613 	int	i, mask;
1614 	u_int	entropy;
1615 
1616 	KASSERT((int)pmap < VM_MIN_KERNEL_ADDRESS, ("moea_pinit: virt pmap"));
1617 	PMAP_LOCK_INIT(pmap);
1618 	LIST_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 	moea_pinit(mmu, pm);
1682 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1683 }
1684 
1685 /*
1686  * Set the physical protection on the specified range of this map as requested.
1687  */
1688 void
1689 moea_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1690     vm_prot_t prot)
1691 {
1692 	struct	pvo_entry *pvo;
1693 	struct	pte *pt;
1694 	int	pteidx;
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 	vm_page_lock_queues();
1705 	PMAP_LOCK(pm);
1706 	for (; sva < eva; sva += PAGE_SIZE) {
1707 		pvo = moea_pvo_find_va(pm, sva, &pteidx);
1708 		if (pvo == NULL)
1709 			continue;
1710 
1711 		if ((prot & VM_PROT_EXECUTE) == 0)
1712 			pvo->pvo_vaddr &= ~PVO_EXECUTABLE;
1713 
1714 		/*
1715 		 * Grab the PTE pointer before we diddle with the cached PTE
1716 		 * copy.
1717 		 */
1718 		pt = moea_pvo_to_pte(pvo, pteidx);
1719 		/*
1720 		 * Change the protection of the page.
1721 		 */
1722 		pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1723 		pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1724 
1725 		/*
1726 		 * If the PVO is in the page table, update that pte as well.
1727 		 */
1728 		if (pt != NULL) {
1729 			moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1730 			mtx_unlock(&moea_table_mutex);
1731 		}
1732 	}
1733 	vm_page_unlock_queues();
1734 	PMAP_UNLOCK(pm);
1735 }
1736 
1737 /*
1738  * Map a list of wired pages into kernel virtual address space.  This is
1739  * intended for temporary mappings which do not need page modification or
1740  * references recorded.  Existing mappings in the region are overwritten.
1741  */
1742 void
1743 moea_qenter(mmu_t mmu, vm_offset_t sva, vm_page_t *m, int count)
1744 {
1745 	vm_offset_t va;
1746 
1747 	va = sva;
1748 	while (count-- > 0) {
1749 		moea_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1750 		va += PAGE_SIZE;
1751 		m++;
1752 	}
1753 }
1754 
1755 /*
1756  * Remove page mappings from kernel virtual address space.  Intended for
1757  * temporary mappings entered by moea_qenter.
1758  */
1759 void
1760 moea_qremove(mmu_t mmu, vm_offset_t sva, int count)
1761 {
1762 	vm_offset_t va;
1763 
1764 	va = sva;
1765 	while (count-- > 0) {
1766 		moea_kremove(mmu, va);
1767 		va += PAGE_SIZE;
1768 	}
1769 }
1770 
1771 void
1772 moea_release(mmu_t mmu, pmap_t pmap)
1773 {
1774         int idx, mask;
1775 
1776 	/*
1777 	 * Free segment register's VSID
1778 	 */
1779         if (pmap->pm_sr[0] == 0)
1780                 panic("moea_release");
1781 
1782 	mtx_lock(&moea_vsid_mutex);
1783         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1784         mask = 1 << (idx % VSID_NBPW);
1785         idx /= VSID_NBPW;
1786         moea_vsid_bitmap[idx] &= ~mask;
1787 	mtx_unlock(&moea_vsid_mutex);
1788 	PMAP_LOCK_DESTROY(pmap);
1789 }
1790 
1791 /*
1792  * Remove the given range of addresses from the specified map.
1793  */
1794 void
1795 moea_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1796 {
1797 	struct	pvo_entry *pvo;
1798 	int	pteidx;
1799 
1800 	vm_page_lock_queues();
1801 	PMAP_LOCK(pm);
1802 	if ((eva - sva)/PAGE_SIZE < 10) {
1803 		for (; sva < eva; sva += PAGE_SIZE) {
1804 			pvo = moea_pvo_find_va(pm, sva, &pteidx);
1805 			if (pvo != NULL)
1806 				moea_pvo_remove(pvo, pteidx);
1807 		}
1808 	} else {
1809 		LIST_FOREACH(pvo, &pm->pmap_pvo, pvo_plink) {
1810 			if (PVO_VADDR(pvo) < sva || PVO_VADDR(pvo) >= eva)
1811 				continue;
1812 			moea_pvo_remove(pvo, -1);
1813 		}
1814 	}
1815 	PMAP_UNLOCK(pm);
1816 	vm_page_unlock_queues();
1817 }
1818 
1819 /*
1820  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1821  * will reflect changes in pte's back to the vm_page.
1822  */
1823 void
1824 moea_remove_all(mmu_t mmu, vm_page_t m)
1825 {
1826 	struct  pvo_head *pvo_head;
1827 	struct	pvo_entry *pvo, *next_pvo;
1828 	pmap_t	pmap;
1829 
1830 	vm_page_lock_queues();
1831 	pvo_head = vm_page_to_pvoh(m);
1832 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1833 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1834 
1835 		pmap = pvo->pvo_pmap;
1836 		PMAP_LOCK(pmap);
1837 		moea_pvo_remove(pvo, -1);
1838 		PMAP_UNLOCK(pmap);
1839 	}
1840 	if ((m->aflags & PGA_WRITEABLE) && moea_is_modified(mmu, m)) {
1841 		moea_attr_clear(m, PTE_CHG);
1842 		vm_page_dirty(m);
1843 	}
1844 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1845 	vm_page_unlock_queues();
1846 }
1847 
1848 /*
1849  * Allocate a physical page of memory directly from the phys_avail map.
1850  * Can only be called from moea_bootstrap before avail start and end are
1851  * calculated.
1852  */
1853 static vm_offset_t
1854 moea_bootstrap_alloc(vm_size_t size, u_int align)
1855 {
1856 	vm_offset_t	s, e;
1857 	int		i, j;
1858 
1859 	size = round_page(size);
1860 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1861 		if (align != 0)
1862 			s = (phys_avail[i] + align - 1) & ~(align - 1);
1863 		else
1864 			s = phys_avail[i];
1865 		e = s + size;
1866 
1867 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1868 			continue;
1869 
1870 		if (s == phys_avail[i]) {
1871 			phys_avail[i] += size;
1872 		} else if (e == phys_avail[i + 1]) {
1873 			phys_avail[i + 1] -= size;
1874 		} else {
1875 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1876 				phys_avail[j] = phys_avail[j - 2];
1877 				phys_avail[j + 1] = phys_avail[j - 1];
1878 			}
1879 
1880 			phys_avail[i + 3] = phys_avail[i + 1];
1881 			phys_avail[i + 1] = s;
1882 			phys_avail[i + 2] = e;
1883 			phys_avail_count++;
1884 		}
1885 
1886 		return (s);
1887 	}
1888 	panic("moea_bootstrap_alloc: could not allocate memory");
1889 }
1890 
1891 static void
1892 moea_syncicache(vm_offset_t pa, vm_size_t len)
1893 {
1894 	__syncicache((void *)pa, len);
1895 }
1896 
1897 static int
1898 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1899     vm_offset_t va, vm_offset_t pa, u_int pte_lo, int flags)
1900 {
1901 	struct	pvo_entry *pvo;
1902 	u_int	sr;
1903 	int	first;
1904 	u_int	ptegidx;
1905 	int	i;
1906 	int     bootstrap;
1907 
1908 	moea_pvo_enter_calls++;
1909 	first = 0;
1910 	bootstrap = 0;
1911 
1912 	/*
1913 	 * Compute the PTE Group index.
1914 	 */
1915 	va &= ~ADDR_POFF;
1916 	sr = va_to_sr(pm->pm_sr, va);
1917 	ptegidx = va_to_pteg(sr, va);
1918 
1919 	/*
1920 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
1921 	 * there is a mapping.
1922 	 */
1923 	mtx_lock(&moea_table_mutex);
1924 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
1925 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
1926 			if ((pvo->pvo_pte.pte.pte_lo & PTE_RPGN) == pa &&
1927 			    (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
1928 			    (pte_lo & PTE_PP)) {
1929 				mtx_unlock(&moea_table_mutex);
1930 				return (0);
1931 			}
1932 			moea_pvo_remove(pvo, -1);
1933 			break;
1934 		}
1935 	}
1936 
1937 	/*
1938 	 * If we aren't overwriting a mapping, try to allocate.
1939 	 */
1940 	if (moea_initialized) {
1941 		pvo = uma_zalloc(zone, M_NOWAIT);
1942 	} else {
1943 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
1944 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
1945 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
1946 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
1947 		}
1948 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
1949 		moea_bpvo_pool_index++;
1950 		bootstrap = 1;
1951 	}
1952 
1953 	if (pvo == NULL) {
1954 		mtx_unlock(&moea_table_mutex);
1955 		return (ENOMEM);
1956 	}
1957 
1958 	moea_pvo_entries++;
1959 	pvo->pvo_vaddr = va;
1960 	pvo->pvo_pmap = pm;
1961 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
1962 	pvo->pvo_vaddr &= ~ADDR_POFF;
1963 	if (flags & VM_PROT_EXECUTE)
1964 		pvo->pvo_vaddr |= PVO_EXECUTABLE;
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 	LIST_INSERT_HEAD(&pm->pmap_pvo, pvo, pvo_plink);
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_pte.pte.pte_lo & PVO_WIRED)
1988 		pm->pm_stats.wired_count++;
1989 	pm->pm_stats.resident_count++;
1990 
1991 	/*
1992 	 * We hope this succeeds but it isn't required.
1993 	 */
1994 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
1995 	if (i >= 0) {
1996 		PVO_PTEGIDX_SET(pvo, i);
1997 	} else {
1998 		panic("moea_pvo_enter: overflow");
1999 		moea_pte_overflow++;
2000 	}
2001 	mtx_unlock(&moea_table_mutex);
2002 
2003 	return (first ? ENOENT : 0);
2004 }
2005 
2006 static void
2007 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
2008 {
2009 	struct	pte *pt;
2010 
2011 	/*
2012 	 * If there is an active pte entry, we need to deactivate it (and
2013 	 * save the ref & cfg bits).
2014 	 */
2015 	pt = moea_pvo_to_pte(pvo, pteidx);
2016 	if (pt != NULL) {
2017 		moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
2018 		mtx_unlock(&moea_table_mutex);
2019 		PVO_PTEGIDX_CLR(pvo);
2020 	} else {
2021 		moea_pte_overflow--;
2022 	}
2023 
2024 	/*
2025 	 * Update our statistics.
2026 	 */
2027 	pvo->pvo_pmap->pm_stats.resident_count--;
2028 	if (pvo->pvo_pte.pte.pte_lo & PVO_WIRED)
2029 		pvo->pvo_pmap->pm_stats.wired_count--;
2030 
2031 	/*
2032 	 * Save the REF/CHG bits into their cache if the page is managed.
2033 	 */
2034 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED) {
2035 		struct	vm_page *pg;
2036 
2037 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
2038 		if (pg != NULL) {
2039 			moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2040 			    (PTE_REF | PTE_CHG));
2041 		}
2042 	}
2043 
2044 	/*
2045 	 * Remove this PVO from the PV and pmap lists.
2046 	 */
2047 	LIST_REMOVE(pvo, pvo_vlink);
2048 	LIST_REMOVE(pvo, pvo_plink);
2049 
2050 	/*
2051 	 * Remove this from the overflow list and return it to the pool
2052 	 * if we aren't going to reuse it.
2053 	 */
2054 	LIST_REMOVE(pvo, pvo_olink);
2055 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2056 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
2057 		    moea_upvo_zone, pvo);
2058 	moea_pvo_entries--;
2059 	moea_pvo_remove_calls++;
2060 }
2061 
2062 static __inline int
2063 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2064 {
2065 	int	pteidx;
2066 
2067 	/*
2068 	 * We can find the actual pte entry without searching by grabbing
2069 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2070 	 * noticing the HID bit.
2071 	 */
2072 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2073 	if (pvo->pvo_pte.pte.pte_hi & PTE_HID)
2074 		pteidx ^= moea_pteg_mask * 8;
2075 
2076 	return (pteidx);
2077 }
2078 
2079 static struct pvo_entry *
2080 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2081 {
2082 	struct	pvo_entry *pvo;
2083 	int	ptegidx;
2084 	u_int	sr;
2085 
2086 	va &= ~ADDR_POFF;
2087 	sr = va_to_sr(pm->pm_sr, va);
2088 	ptegidx = va_to_pteg(sr, va);
2089 
2090 	mtx_lock(&moea_table_mutex);
2091 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2092 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2093 			if (pteidx_p)
2094 				*pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2095 			break;
2096 		}
2097 	}
2098 	mtx_unlock(&moea_table_mutex);
2099 
2100 	return (pvo);
2101 }
2102 
2103 static struct pte *
2104 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2105 {
2106 	struct	pte *pt;
2107 
2108 	/*
2109 	 * If we haven't been supplied the ptegidx, calculate it.
2110 	 */
2111 	if (pteidx == -1) {
2112 		int	ptegidx;
2113 		u_int	sr;
2114 
2115 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2116 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2117 		pteidx = moea_pvo_pte_index(pvo, ptegidx);
2118 	}
2119 
2120 	pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2121 	mtx_lock(&moea_table_mutex);
2122 
2123 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2124 		panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2125 		    "valid pte index", pvo);
2126 	}
2127 
2128 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2129 		panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2130 		    "pvo but no valid pte", pvo);
2131 	}
2132 
2133 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2134 		if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) {
2135 			panic("moea_pvo_to_pte: pvo %p has valid pte in "
2136 			    "moea_pteg_table %p but invalid in pvo", pvo, pt);
2137 		}
2138 
2139 		if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2140 		    != 0) {
2141 			panic("moea_pvo_to_pte: pvo %p pte does not match "
2142 			    "pte %p in moea_pteg_table", pvo, pt);
2143 		}
2144 
2145 		mtx_assert(&moea_table_mutex, MA_OWNED);
2146 		return (pt);
2147 	}
2148 
2149 	if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) {
2150 		panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2151 		    "moea_pteg_table but valid in pvo", pvo, pt);
2152 	}
2153 
2154 	mtx_unlock(&moea_table_mutex);
2155 	return (NULL);
2156 }
2157 
2158 /*
2159  * XXX: THIS STUFF SHOULD BE IN pte.c?
2160  */
2161 int
2162 moea_pte_spill(vm_offset_t addr)
2163 {
2164 	struct	pvo_entry *source_pvo, *victim_pvo;
2165 	struct	pvo_entry *pvo;
2166 	int	ptegidx, i, j;
2167 	u_int	sr;
2168 	struct	pteg *pteg;
2169 	struct	pte *pt;
2170 
2171 	moea_pte_spills++;
2172 
2173 	sr = mfsrin(addr);
2174 	ptegidx = va_to_pteg(sr, addr);
2175 
2176 	/*
2177 	 * Have to substitute some entry.  Use the primary hash for this.
2178 	 * Use low bits of timebase as random generator.
2179 	 */
2180 	pteg = &moea_pteg_table[ptegidx];
2181 	mtx_lock(&moea_table_mutex);
2182 	__asm __volatile("mftb %0" : "=r"(i));
2183 	i &= 7;
2184 	pt = &pteg->pt[i];
2185 
2186 	source_pvo = NULL;
2187 	victim_pvo = NULL;
2188 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2189 		/*
2190 		 * We need to find a pvo entry for this address.
2191 		 */
2192 		if (source_pvo == NULL &&
2193 		    moea_pte_match(&pvo->pvo_pte.pte, sr, addr,
2194 		    pvo->pvo_pte.pte.pte_hi & PTE_HID)) {
2195 			/*
2196 			 * Now found an entry to be spilled into the pteg.
2197 			 * The PTE is now valid, so we know it's active.
2198 			 */
2199 			j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2200 
2201 			if (j >= 0) {
2202 				PVO_PTEGIDX_SET(pvo, j);
2203 				moea_pte_overflow--;
2204 				mtx_unlock(&moea_table_mutex);
2205 				return (1);
2206 			}
2207 
2208 			source_pvo = pvo;
2209 
2210 			if (victim_pvo != NULL)
2211 				break;
2212 		}
2213 
2214 		/*
2215 		 * We also need the pvo entry of the victim we are replacing
2216 		 * so save the R & C bits of the PTE.
2217 		 */
2218 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2219 		    moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2220 			victim_pvo = pvo;
2221 			if (source_pvo != NULL)
2222 				break;
2223 		}
2224 	}
2225 
2226 	if (source_pvo == NULL) {
2227 		mtx_unlock(&moea_table_mutex);
2228 		return (0);
2229 	}
2230 
2231 	if (victim_pvo == NULL) {
2232 		if ((pt->pte_hi & PTE_HID) == 0)
2233 			panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2234 			    "entry", pt);
2235 
2236 		/*
2237 		 * If this is a secondary PTE, we need to search it's primary
2238 		 * pvo bucket for the matching PVO.
2239 		 */
2240 		LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2241 		    pvo_olink) {
2242 			/*
2243 			 * We also need the pvo entry of the victim we are
2244 			 * replacing so save the R & C bits of the PTE.
2245 			 */
2246 			if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2247 				victim_pvo = pvo;
2248 				break;
2249 			}
2250 		}
2251 
2252 		if (victim_pvo == NULL)
2253 			panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2254 			    "entry", pt);
2255 	}
2256 
2257 	/*
2258 	 * We are invalidating the TLB entry for the EA we are replacing even
2259 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2260 	 * contained in the TLB entry.
2261 	 */
2262 	source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID;
2263 
2264 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2265 	moea_pte_set(pt, &source_pvo->pvo_pte.pte);
2266 
2267 	PVO_PTEGIDX_CLR(victim_pvo);
2268 	PVO_PTEGIDX_SET(source_pvo, i);
2269 	moea_pte_replacements++;
2270 
2271 	mtx_unlock(&moea_table_mutex);
2272 	return (1);
2273 }
2274 
2275 static int
2276 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2277 {
2278 	struct	pte *pt;
2279 	int	i;
2280 
2281 	mtx_assert(&moea_table_mutex, MA_OWNED);
2282 
2283 	/*
2284 	 * First try primary hash.
2285 	 */
2286 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2287 		if ((pt->pte_hi & PTE_VALID) == 0) {
2288 			pvo_pt->pte_hi &= ~PTE_HID;
2289 			moea_pte_set(pt, pvo_pt);
2290 			return (i);
2291 		}
2292 	}
2293 
2294 	/*
2295 	 * Now try secondary hash.
2296 	 */
2297 	ptegidx ^= moea_pteg_mask;
2298 
2299 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2300 		if ((pt->pte_hi & PTE_VALID) == 0) {
2301 			pvo_pt->pte_hi |= PTE_HID;
2302 			moea_pte_set(pt, pvo_pt);
2303 			return (i);
2304 		}
2305 	}
2306 
2307 	panic("moea_pte_insert: overflow");
2308 	return (-1);
2309 }
2310 
2311 static boolean_t
2312 moea_query_bit(vm_page_t m, int ptebit)
2313 {
2314 	struct	pvo_entry *pvo;
2315 	struct	pte *pt;
2316 
2317 	if (moea_attr_fetch(m) & ptebit)
2318 		return (TRUE);
2319 
2320 	vm_page_lock_queues();
2321 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2322 
2323 		/*
2324 		 * See if we saved the bit off.  If so, cache it and return
2325 		 * success.
2326 		 */
2327 		if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2328 			moea_attr_save(m, ptebit);
2329 			vm_page_unlock_queues();
2330 			return (TRUE);
2331 		}
2332 	}
2333 
2334 	/*
2335 	 * No luck, now go through the hard part of looking at the PTEs
2336 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2337 	 * the PTEs.
2338 	 */
2339 	powerpc_sync();
2340 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2341 
2342 		/*
2343 		 * See if this pvo has a valid PTE.  if so, fetch the
2344 		 * REF/CHG bits from the valid PTE.  If the appropriate
2345 		 * ptebit is set, cache it and return success.
2346 		 */
2347 		pt = moea_pvo_to_pte(pvo, -1);
2348 		if (pt != NULL) {
2349 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2350 			mtx_unlock(&moea_table_mutex);
2351 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2352 				moea_attr_save(m, ptebit);
2353 				vm_page_unlock_queues();
2354 				return (TRUE);
2355 			}
2356 		}
2357 	}
2358 
2359 	vm_page_unlock_queues();
2360 	return (FALSE);
2361 }
2362 
2363 static u_int
2364 moea_clear_bit(vm_page_t m, int ptebit)
2365 {
2366 	u_int	count;
2367 	struct	pvo_entry *pvo;
2368 	struct	pte *pt;
2369 
2370 	vm_page_lock_queues();
2371 
2372 	/*
2373 	 * Clear the cached value.
2374 	 */
2375 	moea_attr_clear(m, ptebit);
2376 
2377 	/*
2378 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2379 	 * we can reset the right ones).  note that since the pvo entries and
2380 	 * list heads are accessed via BAT0 and are never placed in the page
2381 	 * table, we don't have to worry about further accesses setting the
2382 	 * REF/CHG bits.
2383 	 */
2384 	powerpc_sync();
2385 
2386 	/*
2387 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2388 	 * valid pte clear the ptebit from the valid pte.
2389 	 */
2390 	count = 0;
2391 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2392 		pt = moea_pvo_to_pte(pvo, -1);
2393 		if (pt != NULL) {
2394 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2395 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2396 				count++;
2397 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2398 			}
2399 			mtx_unlock(&moea_table_mutex);
2400 		}
2401 		pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2402 	}
2403 
2404 	vm_page_unlock_queues();
2405 	return (count);
2406 }
2407 
2408 /*
2409  * Return true if the physical range is encompassed by the battable[idx]
2410  */
2411 static int
2412 moea_bat_mapped(int idx, vm_offset_t pa, vm_size_t size)
2413 {
2414 	u_int prot;
2415 	u_int32_t start;
2416 	u_int32_t end;
2417 	u_int32_t bat_ble;
2418 
2419 	/*
2420 	 * Return immediately if not a valid mapping
2421 	 */
2422 	if (!(battable[idx].batu & BAT_Vs))
2423 		return (EINVAL);
2424 
2425 	/*
2426 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2427 	 * so it can function as an i/o page
2428 	 */
2429 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2430 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2431 		return (EPERM);
2432 
2433 	/*
2434 	 * The address should be within the BAT range. Assume that the
2435 	 * start address in the BAT has the correct alignment (thus
2436 	 * not requiring masking)
2437 	 */
2438 	start = battable[idx].batl & BAT_PBS;
2439 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2440 	end = start | (bat_ble << 15) | 0x7fff;
2441 
2442 	if ((pa < start) || ((pa + size) > end))
2443 		return (ERANGE);
2444 
2445 	return (0);
2446 }
2447 
2448 boolean_t
2449 moea_dev_direct_mapped(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2450 {
2451 	int i;
2452 
2453 	/*
2454 	 * This currently does not work for entries that
2455 	 * overlap 256M BAT segments.
2456 	 */
2457 
2458 	for(i = 0; i < 16; i++)
2459 		if (moea_bat_mapped(i, pa, size) == 0)
2460 			return (0);
2461 
2462 	return (EFAULT);
2463 }
2464 
2465 /*
2466  * Map a set of physical memory pages into the kernel virtual
2467  * address space. Return a pointer to where it is mapped. This
2468  * routine is intended to be used for mapping device memory,
2469  * NOT real memory.
2470  */
2471 void *
2472 moea_mapdev(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2473 {
2474 
2475 	return (moea_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT));
2476 }
2477 
2478 void *
2479 moea_mapdev_attr(mmu_t mmu, vm_offset_t pa, vm_size_t size, vm_memattr_t ma)
2480 {
2481 	vm_offset_t va, tmpva, ppa, offset;
2482 	int i;
2483 
2484 	ppa = trunc_page(pa);
2485 	offset = pa & PAGE_MASK;
2486 	size = roundup(offset + size, PAGE_SIZE);
2487 
2488 	/*
2489 	 * If the physical address lies within a valid BAT table entry,
2490 	 * return the 1:1 mapping. This currently doesn't work
2491 	 * for regions that overlap 256M BAT segments.
2492 	 */
2493 	for (i = 0; i < 16; i++) {
2494 		if (moea_bat_mapped(i, pa, size) == 0)
2495 			return ((void *) pa);
2496 	}
2497 
2498 	va = kmem_alloc_nofault(kernel_map, size);
2499 	if (!va)
2500 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2501 
2502 	for (tmpva = va; size > 0;) {
2503 		moea_kenter_attr(mmu, tmpva, ppa, ma);
2504 		tlbie(tmpva);
2505 		size -= PAGE_SIZE;
2506 		tmpva += PAGE_SIZE;
2507 		ppa += PAGE_SIZE;
2508 	}
2509 
2510 	return ((void *)(va + offset));
2511 }
2512 
2513 void
2514 moea_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2515 {
2516 	vm_offset_t base, offset;
2517 
2518 	/*
2519 	 * If this is outside kernel virtual space, then it's a
2520 	 * battable entry and doesn't require unmapping
2521 	 */
2522 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2523 		base = trunc_page(va);
2524 		offset = va & PAGE_MASK;
2525 		size = roundup(offset + size, PAGE_SIZE);
2526 		kmem_free(kernel_map, base, size);
2527 	}
2528 }
2529 
2530 static void
2531 moea_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2532 {
2533 	struct pvo_entry *pvo;
2534 	vm_offset_t lim;
2535 	vm_paddr_t pa;
2536 	vm_size_t len;
2537 
2538 	PMAP_LOCK(pm);
2539 	while (sz > 0) {
2540 		lim = round_page(va);
2541 		len = MIN(lim - va, sz);
2542 		pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2543 		if (pvo != NULL) {
2544 			pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) |
2545 			    (va & ADDR_POFF);
2546 			moea_syncicache(pa, len);
2547 		}
2548 		va += len;
2549 		sz -= len;
2550 	}
2551 	PMAP_UNLOCK(pm);
2552 }
2553