xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision 4310d6deb27da04d3fe079a0584edd557a764e21)
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_compat.h"
118 #include "opt_kstack_pages.h"
119 
120 #include <sys/param.h>
121 #include <sys/kernel.h>
122 #include <sys/queue.h>
123 #include <sys/cpuset.h>
124 #include <sys/ktr.h>
125 #include <sys/lock.h>
126 #include <sys/msgbuf.h>
127 #include <sys/mutex.h>
128 #include <sys/proc.h>
129 #include <sys/rwlock.h>
130 #include <sys/sched.h>
131 #include <sys/sysctl.h>
132 #include <sys/systm.h>
133 #include <sys/vmmeter.h>
134 
135 #include <sys/kdb.h>
136 
137 #include <dev/ofw/openfirm.h>
138 
139 #include <vm/vm.h>
140 #include <vm/vm_param.h>
141 #include <vm/vm_kern.h>
142 #include <vm/vm_page.h>
143 #include <vm/vm_map.h>
144 #include <vm/vm_object.h>
145 #include <vm/vm_extern.h>
146 #include <vm/vm_pageout.h>
147 #include <vm/vm_pager.h>
148 #include <vm/uma.h>
149 
150 #include <machine/_inttypes.h>
151 #include <machine/cpu.h>
152 #include <machine/platform.h>
153 #include <machine/frame.h>
154 #include <machine/md_var.h>
155 #include <machine/psl.h>
156 #include <machine/bat.h>
157 #include <machine/hid.h>
158 #include <machine/pte.h>
159 #include <machine/sr.h>
160 #include <machine/trap.h>
161 #include <machine/mmuvar.h>
162 
163 #include "mmu_oea64.h"
164 #include "mmu_if.h"
165 #include "moea64_if.h"
166 
167 void moea64_release_vsid(uint64_t vsid);
168 uintptr_t moea64_get_unique_vsid(void);
169 
170 #define DISABLE_TRANS(msr)	msr = mfmsr(); mtmsr(msr & ~PSL_DR)
171 #define ENABLE_TRANS(msr)	mtmsr(msr)
172 
173 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
174 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
175 #define	VSID_HASH_MASK		0x0000007fffffffffULL
176 
177 /*
178  * Locking semantics:
179  * -- Read lock: if no modifications are being made to either the PVO lists
180  *    or page table or if any modifications being made result in internal
181  *    changes (e.g. wiring, protection) such that the existence of the PVOs
182  *    is unchanged and they remain associated with the same pmap (in which
183  *    case the changes should be protected by the pmap lock)
184  * -- Write lock: required if PTEs/PVOs are being inserted or removed.
185  */
186 
187 #define LOCK_TABLE_RD() rw_rlock(&moea64_table_lock)
188 #define UNLOCK_TABLE_RD() rw_runlock(&moea64_table_lock)
189 #define LOCK_TABLE_WR() rw_wlock(&moea64_table_lock)
190 #define UNLOCK_TABLE_WR() rw_wunlock(&moea64_table_lock)
191 
192 struct ofw_map {
193 	cell_t	om_va;
194 	cell_t	om_len;
195 	cell_t	om_pa_hi;
196 	cell_t	om_pa_lo;
197 	cell_t	om_mode;
198 };
199 
200 /*
201  * Map of physical memory regions.
202  */
203 static struct	mem_region *regions;
204 static struct	mem_region *pregions;
205 static u_int	phys_avail_count;
206 static int	regions_sz, pregions_sz;
207 
208 extern void bs_remap_earlyboot(void);
209 
210 /*
211  * Lock for the pteg and pvo tables.
212  */
213 struct rwlock	moea64_table_lock;
214 struct mtx	moea64_slb_mutex;
215 
216 /*
217  * PTEG data.
218  */
219 u_int		moea64_pteg_count;
220 u_int		moea64_pteg_mask;
221 
222 /*
223  * PVO data.
224  */
225 struct	pvo_head *moea64_pvo_table;		/* pvo entries by pteg index */
226 struct	pvo_head moea64_pvo_kunmanaged =	/* list of unmanaged pages */
227     LIST_HEAD_INITIALIZER(moea64_pvo_kunmanaged);
228 
229 uma_zone_t	moea64_upvo_zone; /* zone for pvo entries for unmanaged pages */
230 uma_zone_t	moea64_mpvo_zone; /* zone for pvo entries for managed pages */
231 
232 #define	BPVO_POOL_SIZE	327680
233 static struct	pvo_entry *moea64_bpvo_pool;
234 static int	moea64_bpvo_pool_index = 0;
235 
236 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
237 #ifdef __powerpc64__
238 #define	NVSIDS		(NPMAPS * 16)
239 #define VSID_HASHMASK	0xffffffffUL
240 #else
241 #define NVSIDS		NPMAPS
242 #define VSID_HASHMASK	0xfffffUL
243 #endif
244 static u_int	moea64_vsid_bitmap[NVSIDS / VSID_NBPW];
245 
246 static boolean_t moea64_initialized = FALSE;
247 
248 /*
249  * Statistics.
250  */
251 u_int	moea64_pte_valid = 0;
252 u_int	moea64_pte_overflow = 0;
253 u_int	moea64_pvo_entries = 0;
254 u_int	moea64_pvo_enter_calls = 0;
255 u_int	moea64_pvo_remove_calls = 0;
256 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_valid, CTLFLAG_RD,
257     &moea64_pte_valid, 0, "");
258 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_overflow, CTLFLAG_RD,
259     &moea64_pte_overflow, 0, "");
260 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_entries, CTLFLAG_RD,
261     &moea64_pvo_entries, 0, "");
262 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_enter_calls, CTLFLAG_RD,
263     &moea64_pvo_enter_calls, 0, "");
264 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_remove_calls, CTLFLAG_RD,
265     &moea64_pvo_remove_calls, 0, "");
266 
267 vm_offset_t	moea64_scratchpage_va[2];
268 struct pvo_entry *moea64_scratchpage_pvo[2];
269 uintptr_t	moea64_scratchpage_pte[2];
270 struct	mtx	moea64_scratchpage_mtx;
271 
272 uint64_t 	moea64_large_page_mask = 0;
273 int		moea64_large_page_size = 0;
274 int		moea64_large_page_shift = 0;
275 
276 /*
277  * PVO calls.
278  */
279 static int	moea64_pvo_enter(mmu_t, pmap_t, uma_zone_t, struct pvo_head *,
280 		    vm_offset_t, vm_offset_t, uint64_t, int);
281 static void	moea64_pvo_remove(mmu_t, struct pvo_entry *);
282 static struct	pvo_entry *moea64_pvo_find_va(pmap_t, vm_offset_t);
283 
284 /*
285  * Utility routines.
286  */
287 static boolean_t	moea64_query_bit(mmu_t, vm_page_t, u_int64_t);
288 static u_int		moea64_clear_bit(mmu_t, vm_page_t, u_int64_t);
289 static void		moea64_kremove(mmu_t, vm_offset_t);
290 static void		moea64_syncicache(mmu_t, pmap_t pmap, vm_offset_t va,
291 			    vm_offset_t pa, vm_size_t sz);
292 
293 /*
294  * Kernel MMU interface
295  */
296 void moea64_change_wiring(mmu_t, pmap_t, vm_offset_t, boolean_t);
297 void moea64_clear_modify(mmu_t, vm_page_t);
298 void moea64_clear_reference(mmu_t, vm_page_t);
299 void moea64_copy_page(mmu_t, vm_page_t, vm_page_t);
300 void moea64_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, boolean_t);
301 void moea64_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
302     vm_prot_t);
303 void moea64_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
304 vm_paddr_t moea64_extract(mmu_t, pmap_t, vm_offset_t);
305 vm_page_t moea64_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
306 void moea64_init(mmu_t);
307 boolean_t moea64_is_modified(mmu_t, vm_page_t);
308 boolean_t moea64_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
309 boolean_t moea64_is_referenced(mmu_t, vm_page_t);
310 boolean_t moea64_ts_referenced(mmu_t, vm_page_t);
311 vm_offset_t moea64_map(mmu_t, vm_offset_t *, vm_offset_t, vm_offset_t, int);
312 boolean_t moea64_page_exists_quick(mmu_t, pmap_t, vm_page_t);
313 int moea64_page_wired_mappings(mmu_t, vm_page_t);
314 void moea64_pinit(mmu_t, pmap_t);
315 void moea64_pinit0(mmu_t, pmap_t);
316 void moea64_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
317 void moea64_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
318 void moea64_qremove(mmu_t, vm_offset_t, int);
319 void moea64_release(mmu_t, pmap_t);
320 void moea64_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
321 void moea64_remove_pages(mmu_t, pmap_t);
322 void moea64_remove_all(mmu_t, vm_page_t);
323 void moea64_remove_write(mmu_t, vm_page_t);
324 void moea64_zero_page(mmu_t, vm_page_t);
325 void moea64_zero_page_area(mmu_t, vm_page_t, int, int);
326 void moea64_zero_page_idle(mmu_t, vm_page_t);
327 void moea64_activate(mmu_t, struct thread *);
328 void moea64_deactivate(mmu_t, struct thread *);
329 void *moea64_mapdev(mmu_t, vm_offset_t, vm_size_t);
330 void *moea64_mapdev_attr(mmu_t, vm_offset_t, vm_size_t, vm_memattr_t);
331 void moea64_unmapdev(mmu_t, vm_offset_t, vm_size_t);
332 vm_offset_t moea64_kextract(mmu_t, vm_offset_t);
333 void moea64_page_set_memattr(mmu_t, vm_page_t m, vm_memattr_t ma);
334 void moea64_kenter_attr(mmu_t, vm_offset_t, vm_offset_t, vm_memattr_t ma);
335 void moea64_kenter(mmu_t, vm_offset_t, vm_offset_t);
336 boolean_t moea64_dev_direct_mapped(mmu_t, vm_offset_t, vm_size_t);
337 static void moea64_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
338 
339 static mmu_method_t moea64_methods[] = {
340 	MMUMETHOD(mmu_change_wiring,	moea64_change_wiring),
341 	MMUMETHOD(mmu_clear_modify,	moea64_clear_modify),
342 	MMUMETHOD(mmu_clear_reference,	moea64_clear_reference),
343 	MMUMETHOD(mmu_copy_page,	moea64_copy_page),
344 	MMUMETHOD(mmu_enter,		moea64_enter),
345 	MMUMETHOD(mmu_enter_object,	moea64_enter_object),
346 	MMUMETHOD(mmu_enter_quick,	moea64_enter_quick),
347 	MMUMETHOD(mmu_extract,		moea64_extract),
348 	MMUMETHOD(mmu_extract_and_hold,	moea64_extract_and_hold),
349 	MMUMETHOD(mmu_init,		moea64_init),
350 	MMUMETHOD(mmu_is_modified,	moea64_is_modified),
351 	MMUMETHOD(mmu_is_prefaultable,	moea64_is_prefaultable),
352 	MMUMETHOD(mmu_is_referenced,	moea64_is_referenced),
353 	MMUMETHOD(mmu_ts_referenced,	moea64_ts_referenced),
354 	MMUMETHOD(mmu_map,     		moea64_map),
355 	MMUMETHOD(mmu_page_exists_quick,moea64_page_exists_quick),
356 	MMUMETHOD(mmu_page_wired_mappings,moea64_page_wired_mappings),
357 	MMUMETHOD(mmu_pinit,		moea64_pinit),
358 	MMUMETHOD(mmu_pinit0,		moea64_pinit0),
359 	MMUMETHOD(mmu_protect,		moea64_protect),
360 	MMUMETHOD(mmu_qenter,		moea64_qenter),
361 	MMUMETHOD(mmu_qremove,		moea64_qremove),
362 	MMUMETHOD(mmu_release,		moea64_release),
363 	MMUMETHOD(mmu_remove,		moea64_remove),
364 	MMUMETHOD(mmu_remove_pages,	moea64_remove_pages),
365 	MMUMETHOD(mmu_remove_all,      	moea64_remove_all),
366 	MMUMETHOD(mmu_remove_write,	moea64_remove_write),
367 	MMUMETHOD(mmu_sync_icache,	moea64_sync_icache),
368 	MMUMETHOD(mmu_zero_page,       	moea64_zero_page),
369 	MMUMETHOD(mmu_zero_page_area,	moea64_zero_page_area),
370 	MMUMETHOD(mmu_zero_page_idle,	moea64_zero_page_idle),
371 	MMUMETHOD(mmu_activate,		moea64_activate),
372 	MMUMETHOD(mmu_deactivate,      	moea64_deactivate),
373 	MMUMETHOD(mmu_page_set_memattr,	moea64_page_set_memattr),
374 
375 	/* Internal interfaces */
376 	MMUMETHOD(mmu_mapdev,		moea64_mapdev),
377 	MMUMETHOD(mmu_mapdev_attr,	moea64_mapdev_attr),
378 	MMUMETHOD(mmu_unmapdev,		moea64_unmapdev),
379 	MMUMETHOD(mmu_kextract,		moea64_kextract),
380 	MMUMETHOD(mmu_kenter,		moea64_kenter),
381 	MMUMETHOD(mmu_kenter_attr,	moea64_kenter_attr),
382 	MMUMETHOD(mmu_dev_direct_mapped,moea64_dev_direct_mapped),
383 
384 	{ 0, 0 }
385 };
386 
387 MMU_DEF(oea64_mmu, "mmu_oea64_base", moea64_methods, 0);
388 
389 static __inline u_int
390 va_to_pteg(uint64_t vsid, vm_offset_t addr, int large)
391 {
392 	uint64_t hash;
393 	int shift;
394 
395 	shift = large ? moea64_large_page_shift : ADDR_PIDX_SHFT;
396 	hash = (vsid & VSID_HASH_MASK) ^ (((uint64_t)addr & ADDR_PIDX) >>
397 	    shift);
398 	return (hash & moea64_pteg_mask);
399 }
400 
401 static __inline struct pvo_head *
402 vm_page_to_pvoh(vm_page_t m)
403 {
404 
405 	return (&m->md.mdpg_pvoh);
406 }
407 
408 static __inline void
409 moea64_pte_create(struct lpte *pt, uint64_t vsid, vm_offset_t va,
410     uint64_t pte_lo, int flags)
411 {
412 
413 	/*
414 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
415 	 * set when the real pte is set in memory.
416 	 *
417 	 * Note: Don't set the valid bit for correct operation of tlb update.
418 	 */
419 	pt->pte_hi = (vsid << LPTE_VSID_SHIFT) |
420 	    (((uint64_t)(va & ADDR_PIDX) >> ADDR_API_SHFT64) & LPTE_API);
421 
422 	if (flags & PVO_LARGE)
423 		pt->pte_hi |= LPTE_BIG;
424 
425 	pt->pte_lo = pte_lo;
426 }
427 
428 static __inline uint64_t
429 moea64_calc_wimg(vm_offset_t pa, vm_memattr_t ma)
430 {
431 	uint64_t pte_lo;
432 	int i;
433 
434 	if (ma != VM_MEMATTR_DEFAULT) {
435 		switch (ma) {
436 		case VM_MEMATTR_UNCACHEABLE:
437 			return (LPTE_I | LPTE_G);
438 		case VM_MEMATTR_WRITE_COMBINING:
439 		case VM_MEMATTR_WRITE_BACK:
440 		case VM_MEMATTR_PREFETCHABLE:
441 			return (LPTE_I);
442 		case VM_MEMATTR_WRITE_THROUGH:
443 			return (LPTE_W | LPTE_M);
444 		}
445 	}
446 
447 	/*
448 	 * Assume the page is cache inhibited and access is guarded unless
449 	 * it's in our available memory array.
450 	 */
451 	pte_lo = LPTE_I | LPTE_G;
452 	for (i = 0; i < pregions_sz; i++) {
453 		if ((pa >= pregions[i].mr_start) &&
454 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
455 			pte_lo &= ~(LPTE_I | LPTE_G);
456 			pte_lo |= LPTE_M;
457 			break;
458 		}
459 	}
460 
461 	return pte_lo;
462 }
463 
464 /*
465  * Quick sort callout for comparing memory regions.
466  */
467 static int	om_cmp(const void *a, const void *b);
468 
469 static int
470 om_cmp(const void *a, const void *b)
471 {
472 	const struct	ofw_map *mapa;
473 	const struct	ofw_map *mapb;
474 
475 	mapa = a;
476 	mapb = b;
477 	if (mapa->om_pa_hi < mapb->om_pa_hi)
478 		return (-1);
479 	else if (mapa->om_pa_hi > mapb->om_pa_hi)
480 		return (1);
481 	else if (mapa->om_pa_lo < mapb->om_pa_lo)
482 		return (-1);
483 	else if (mapa->om_pa_lo > mapb->om_pa_lo)
484 		return (1);
485 	else
486 		return (0);
487 }
488 
489 static void
490 moea64_add_ofw_mappings(mmu_t mmup, phandle_t mmu, size_t sz)
491 {
492 	struct ofw_map	translations[sz/sizeof(struct ofw_map)];
493 	register_t	msr;
494 	vm_offset_t	off;
495 	vm_paddr_t	pa_base;
496 	int		i;
497 
498 	bzero(translations, sz);
499 	if (OF_getprop(mmu, "translations", translations, sz) == -1)
500 		panic("moea64_bootstrap: can't get ofw translations");
501 
502 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
503 	sz /= sizeof(*translations);
504 	qsort(translations, sz, sizeof (*translations), om_cmp);
505 
506 	for (i = 0; i < sz; i++) {
507 		CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
508 		    (uint32_t)(translations[i].om_pa_lo), translations[i].om_va,
509 		    translations[i].om_len);
510 
511 		if (translations[i].om_pa_lo % PAGE_SIZE)
512 			panic("OFW translation not page-aligned!");
513 
514 		pa_base = translations[i].om_pa_lo;
515 
516 	      #ifdef __powerpc64__
517 		pa_base += (vm_offset_t)translations[i].om_pa_hi << 32;
518 	      #else
519 		if (translations[i].om_pa_hi)
520 			panic("OFW translations above 32-bit boundary!");
521 	      #endif
522 
523 		/* Now enter the pages for this mapping */
524 
525 		DISABLE_TRANS(msr);
526 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
527 			if (moea64_pvo_find_va(kernel_pmap,
528 			    translations[i].om_va + off) != NULL)
529 				continue;
530 
531 			moea64_kenter(mmup, translations[i].om_va + off,
532 			    pa_base + off);
533 		}
534 		ENABLE_TRANS(msr);
535 	}
536 }
537 
538 #ifdef __powerpc64__
539 static void
540 moea64_probe_large_page(void)
541 {
542 	uint16_t pvr = mfpvr() >> 16;
543 
544 	switch (pvr) {
545 	case IBM970:
546 	case IBM970FX:
547 	case IBM970MP:
548 		powerpc_sync(); isync();
549 		mtspr(SPR_HID4, mfspr(SPR_HID4) & ~HID4_970_DISABLE_LG_PG);
550 		powerpc_sync(); isync();
551 
552 		/* FALLTHROUGH */
553 	case IBMCELLBE:
554 		moea64_large_page_size = 0x1000000; /* 16 MB */
555 		moea64_large_page_shift = 24;
556 		break;
557 	default:
558 		moea64_large_page_size = 0;
559 	}
560 
561 	moea64_large_page_mask = moea64_large_page_size - 1;
562 }
563 
564 static void
565 moea64_bootstrap_slb_prefault(vm_offset_t va, int large)
566 {
567 	struct slb *cache;
568 	struct slb entry;
569 	uint64_t esid, slbe;
570 	uint64_t i;
571 
572 	cache = PCPU_GET(slb);
573 	esid = va >> ADDR_SR_SHFT;
574 	slbe = (esid << SLBE_ESID_SHIFT) | SLBE_VALID;
575 
576 	for (i = 0; i < 64; i++) {
577 		if (cache[i].slbe == (slbe | i))
578 			return;
579 	}
580 
581 	entry.slbe = slbe;
582 	entry.slbv = KERNEL_VSID(esid) << SLBV_VSID_SHIFT;
583 	if (large)
584 		entry.slbv |= SLBV_L;
585 
586 	slb_insert_kernel(entry.slbe, entry.slbv);
587 }
588 #endif
589 
590 static void
591 moea64_setup_direct_map(mmu_t mmup, vm_offset_t kernelstart,
592     vm_offset_t kernelend)
593 {
594 	register_t msr;
595 	vm_paddr_t pa;
596 	vm_offset_t size, off;
597 	uint64_t pte_lo;
598 	int i;
599 
600 	if (moea64_large_page_size == 0)
601 		hw_direct_map = 0;
602 
603 	DISABLE_TRANS(msr);
604 	if (hw_direct_map) {
605 		LOCK_TABLE_WR();
606 		PMAP_LOCK(kernel_pmap);
607 		for (i = 0; i < pregions_sz; i++) {
608 		  for (pa = pregions[i].mr_start; pa < pregions[i].mr_start +
609 		     pregions[i].mr_size; pa += moea64_large_page_size) {
610 			pte_lo = LPTE_M;
611 
612 			/*
613 			 * Set memory access as guarded if prefetch within
614 			 * the page could exit the available physmem area.
615 			 */
616 			if (pa & moea64_large_page_mask) {
617 				pa &= moea64_large_page_mask;
618 				pte_lo |= LPTE_G;
619 			}
620 			if (pa + moea64_large_page_size >
621 			    pregions[i].mr_start + pregions[i].mr_size)
622 				pte_lo |= LPTE_G;
623 
624 			moea64_pvo_enter(mmup, kernel_pmap, moea64_upvo_zone,
625 				    &moea64_pvo_kunmanaged, pa, pa,
626 				    pte_lo, PVO_WIRED | PVO_LARGE);
627 		  }
628 		}
629 		PMAP_UNLOCK(kernel_pmap);
630 		UNLOCK_TABLE_WR();
631 	} else {
632 		size = sizeof(struct pvo_head) * moea64_pteg_count;
633 		off = (vm_offset_t)(moea64_pvo_table);
634 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
635 			moea64_kenter(mmup, pa, pa);
636 		size = BPVO_POOL_SIZE*sizeof(struct pvo_entry);
637 		off = (vm_offset_t)(moea64_bpvo_pool);
638 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
639 		moea64_kenter(mmup, pa, pa);
640 
641 		/*
642 		 * Map certain important things, like ourselves.
643 		 *
644 		 * NOTE: We do not map the exception vector space. That code is
645 		 * used only in real mode, and leaving it unmapped allows us to
646 		 * catch NULL pointer deferences, instead of making NULL a valid
647 		 * address.
648 		 */
649 
650 		for (pa = kernelstart & ~PAGE_MASK; pa < kernelend;
651 		    pa += PAGE_SIZE)
652 			moea64_kenter(mmup, pa, pa);
653 	}
654 	ENABLE_TRANS(msr);
655 }
656 
657 void
658 moea64_early_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
659 {
660 	int		i, j;
661 	vm_size_t	physsz, hwphyssz;
662 
663 #ifndef __powerpc64__
664 	/* We don't have a direct map since there is no BAT */
665 	hw_direct_map = 0;
666 
667 	/* Make sure battable is zero, since we have no BAT */
668 	for (i = 0; i < 16; i++) {
669 		battable[i].batu = 0;
670 		battable[i].batl = 0;
671 	}
672 #else
673 	moea64_probe_large_page();
674 
675 	/* Use a direct map if we have large page support */
676 	if (moea64_large_page_size > 0)
677 		hw_direct_map = 1;
678 	else
679 		hw_direct_map = 0;
680 #endif
681 
682 	/* Get physical memory regions from firmware */
683 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
684 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
685 
686 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
687 		panic("moea64_bootstrap: phys_avail too small");
688 
689 	phys_avail_count = 0;
690 	physsz = 0;
691 	hwphyssz = 0;
692 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
693 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
694 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
695 		    regions[i].mr_start + regions[i].mr_size,
696 		    regions[i].mr_size);
697 		if (hwphyssz != 0 &&
698 		    (physsz + regions[i].mr_size) >= hwphyssz) {
699 			if (physsz < hwphyssz) {
700 				phys_avail[j] = regions[i].mr_start;
701 				phys_avail[j + 1] = regions[i].mr_start +
702 				    hwphyssz - physsz;
703 				physsz = hwphyssz;
704 				phys_avail_count++;
705 			}
706 			break;
707 		}
708 		phys_avail[j] = regions[i].mr_start;
709 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
710 		phys_avail_count++;
711 		physsz += regions[i].mr_size;
712 	}
713 
714 	/* Check for overlap with the kernel and exception vectors */
715 	for (j = 0; j < 2*phys_avail_count; j+=2) {
716 		if (phys_avail[j] < EXC_LAST)
717 			phys_avail[j] += EXC_LAST;
718 
719 		if (kernelstart >= phys_avail[j] &&
720 		    kernelstart < phys_avail[j+1]) {
721 			if (kernelend < phys_avail[j+1]) {
722 				phys_avail[2*phys_avail_count] =
723 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
724 				phys_avail[2*phys_avail_count + 1] =
725 				    phys_avail[j+1];
726 				phys_avail_count++;
727 			}
728 
729 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
730 		}
731 
732 		if (kernelend >= phys_avail[j] &&
733 		    kernelend < phys_avail[j+1]) {
734 			if (kernelstart > phys_avail[j]) {
735 				phys_avail[2*phys_avail_count] = phys_avail[j];
736 				phys_avail[2*phys_avail_count + 1] =
737 				    kernelstart & ~PAGE_MASK;
738 				phys_avail_count++;
739 			}
740 
741 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
742 		}
743 	}
744 
745 	physmem = btoc(physsz);
746 
747 #ifdef PTEGCOUNT
748 	moea64_pteg_count = PTEGCOUNT;
749 #else
750 	moea64_pteg_count = 0x1000;
751 
752 	while (moea64_pteg_count < physmem)
753 		moea64_pteg_count <<= 1;
754 
755 	moea64_pteg_count >>= 1;
756 #endif /* PTEGCOUNT */
757 }
758 
759 void
760 moea64_mid_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
761 {
762 	vm_size_t	size;
763 	register_t	msr;
764 	int		i;
765 
766 	/*
767 	 * Set PTEG mask
768 	 */
769 	moea64_pteg_mask = moea64_pteg_count - 1;
770 
771 	/*
772 	 * Allocate pv/overflow lists.
773 	 */
774 	size = sizeof(struct pvo_head) * moea64_pteg_count;
775 
776 	moea64_pvo_table = (struct pvo_head *)moea64_bootstrap_alloc(size,
777 	    PAGE_SIZE);
778 	CTR1(KTR_PMAP, "moea64_bootstrap: PVO table at %p", moea64_pvo_table);
779 
780 	DISABLE_TRANS(msr);
781 	for (i = 0; i < moea64_pteg_count; i++)
782 		LIST_INIT(&moea64_pvo_table[i]);
783 	ENABLE_TRANS(msr);
784 
785 	/*
786 	 * Initialize the lock that synchronizes access to the pteg and pvo
787 	 * tables.
788 	 */
789 	rw_init_flags(&moea64_table_lock, "pmap tables", RW_RECURSE);
790 	mtx_init(&moea64_slb_mutex, "SLB table", NULL, MTX_DEF);
791 
792 	/*
793 	 * Initialise the unmanaged pvo pool.
794 	 */
795 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
796 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
797 	moea64_bpvo_pool_index = 0;
798 
799 	/*
800 	 * Make sure kernel vsid is allocated as well as VSID 0.
801 	 */
802 	#ifndef __powerpc64__
803 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NVSIDS - 1)) / VSID_NBPW]
804 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
805 	moea64_vsid_bitmap[0] |= 1;
806 	#endif
807 
808 	/*
809 	 * Initialize the kernel pmap (which is statically allocated).
810 	 */
811 	#ifdef __powerpc64__
812 	for (i = 0; i < 64; i++) {
813 		pcpup->pc_slb[i].slbv = 0;
814 		pcpup->pc_slb[i].slbe = 0;
815 	}
816 	#else
817 	for (i = 0; i < 16; i++)
818 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
819 	#endif
820 
821 	kernel_pmap->pmap_phys = kernel_pmap;
822 	CPU_FILL(&kernel_pmap->pm_active);
823 	LIST_INIT(&kernel_pmap->pmap_pvo);
824 
825 	PMAP_LOCK_INIT(kernel_pmap);
826 
827 	/*
828 	 * Now map in all the other buffers we allocated earlier
829 	 */
830 
831 	moea64_setup_direct_map(mmup, kernelstart, kernelend);
832 }
833 
834 void
835 moea64_late_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
836 {
837 	ihandle_t	mmui;
838 	phandle_t	chosen;
839 	phandle_t	mmu;
840 	size_t		sz;
841 	int		i;
842 	vm_offset_t	pa, va;
843 	void		*dpcpu;
844 
845 	/*
846 	 * Set up the Open Firmware pmap and add its mappings if not in real
847 	 * mode.
848 	 */
849 
850 	chosen = OF_finddevice("/chosen");
851 	if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1) {
852 	    mmu = OF_instance_to_package(mmui);
853 	    if (mmu == -1 || (sz = OF_getproplen(mmu, "translations")) == -1)
854 		sz = 0;
855 	    if (sz > 6144 /* tmpstksz - 2 KB headroom */)
856 		panic("moea64_bootstrap: too many ofw translations");
857 
858 	    if (sz > 0)
859 		moea64_add_ofw_mappings(mmup, mmu, sz);
860 	}
861 
862 	/*
863 	 * Calculate the last available physical address.
864 	 */
865 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
866 		;
867 	Maxmem = powerpc_btop(phys_avail[i + 1]);
868 
869 	/*
870 	 * Initialize MMU and remap early physical mappings
871 	 */
872 	MMU_CPU_BOOTSTRAP(mmup,0);
873 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
874 	pmap_bootstrapped++;
875 	bs_remap_earlyboot();
876 
877 	/*
878 	 * Set the start and end of kva.
879 	 */
880 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
881 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
882 
883 	/*
884 	 * Map the entire KVA range into the SLB. We must not fault there.
885 	 */
886 	#ifdef __powerpc64__
887 	for (va = virtual_avail; va < virtual_end; va += SEGMENT_LENGTH)
888 		moea64_bootstrap_slb_prefault(va, 0);
889 	#endif
890 
891 	/*
892 	 * Figure out how far we can extend virtual_end into segment 16
893 	 * without running into existing mappings. Segment 16 is guaranteed
894 	 * to contain neither RAM nor devices (at least on Apple hardware),
895 	 * but will generally contain some OFW mappings we should not
896 	 * step on.
897 	 */
898 
899 	#ifndef __powerpc64__	/* KVA is in high memory on PPC64 */
900 	PMAP_LOCK(kernel_pmap);
901 	while (virtual_end < VM_MAX_KERNEL_ADDRESS &&
902 	    moea64_pvo_find_va(kernel_pmap, virtual_end+1) == NULL)
903 		virtual_end += PAGE_SIZE;
904 	PMAP_UNLOCK(kernel_pmap);
905 	#endif
906 
907 	/*
908 	 * Allocate a kernel stack with a guard page for thread0 and map it
909 	 * into the kernel page map.
910 	 */
911 	pa = moea64_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, PAGE_SIZE);
912 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
913 	virtual_avail = va + KSTACK_PAGES * PAGE_SIZE;
914 	CTR2(KTR_PMAP, "moea64_bootstrap: kstack0 at %#x (%#x)", pa, va);
915 	thread0.td_kstack = va;
916 	thread0.td_kstack_pages = KSTACK_PAGES;
917 	for (i = 0; i < KSTACK_PAGES; i++) {
918 		moea64_kenter(mmup, va, pa);
919 		pa += PAGE_SIZE;
920 		va += PAGE_SIZE;
921 	}
922 
923 	/*
924 	 * Allocate virtual address space for the message buffer.
925 	 */
926 	pa = msgbuf_phys = moea64_bootstrap_alloc(msgbufsize, PAGE_SIZE);
927 	msgbufp = (struct msgbuf *)virtual_avail;
928 	va = virtual_avail;
929 	virtual_avail += round_page(msgbufsize);
930 	while (va < virtual_avail) {
931 		moea64_kenter(mmup, va, pa);
932 		pa += PAGE_SIZE;
933 		va += PAGE_SIZE;
934 	}
935 
936 	/*
937 	 * Allocate virtual address space for the dynamic percpu area.
938 	 */
939 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
940 	dpcpu = (void *)virtual_avail;
941 	va = virtual_avail;
942 	virtual_avail += DPCPU_SIZE;
943 	while (va < virtual_avail) {
944 		moea64_kenter(mmup, va, pa);
945 		pa += PAGE_SIZE;
946 		va += PAGE_SIZE;
947 	}
948 	dpcpu_init(dpcpu, 0);
949 
950 	/*
951 	 * Allocate some things for page zeroing. We put this directly
952 	 * in the page table, marked with LPTE_LOCKED, to avoid any
953 	 * of the PVO book-keeping or other parts of the VM system
954 	 * from even knowing that this hack exists.
955 	 */
956 
957 	if (!hw_direct_map) {
958 		mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL,
959 		    MTX_DEF);
960 		for (i = 0; i < 2; i++) {
961 			moea64_scratchpage_va[i] = (virtual_end+1) - PAGE_SIZE;
962 			virtual_end -= PAGE_SIZE;
963 
964 			moea64_kenter(mmup, moea64_scratchpage_va[i], 0);
965 
966 			moea64_scratchpage_pvo[i] = moea64_pvo_find_va(
967 			    kernel_pmap, (vm_offset_t)moea64_scratchpage_va[i]);
968 			LOCK_TABLE_RD();
969 			moea64_scratchpage_pte[i] = MOEA64_PVO_TO_PTE(
970 			    mmup, moea64_scratchpage_pvo[i]);
971 			moea64_scratchpage_pvo[i]->pvo_pte.lpte.pte_hi
972 			    |= LPTE_LOCKED;
973 			MOEA64_PTE_CHANGE(mmup, moea64_scratchpage_pte[i],
974 			    &moea64_scratchpage_pvo[i]->pvo_pte.lpte,
975 			    moea64_scratchpage_pvo[i]->pvo_vpn);
976 			UNLOCK_TABLE_RD();
977 		}
978 	}
979 }
980 
981 /*
982  * Activate a user pmap.  The pmap must be activated before its address
983  * space can be accessed in any way.
984  */
985 void
986 moea64_activate(mmu_t mmu, struct thread *td)
987 {
988 	pmap_t	pm;
989 
990 	pm = &td->td_proc->p_vmspace->vm_pmap;
991 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
992 
993 	#ifdef __powerpc64__
994 	PCPU_SET(userslb, pm->pm_slb);
995 	#else
996 	PCPU_SET(curpmap, pm->pmap_phys);
997 	#endif
998 }
999 
1000 void
1001 moea64_deactivate(mmu_t mmu, struct thread *td)
1002 {
1003 	pmap_t	pm;
1004 
1005 	pm = &td->td_proc->p_vmspace->vm_pmap;
1006 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1007 	#ifdef __powerpc64__
1008 	PCPU_SET(userslb, NULL);
1009 	#else
1010 	PCPU_SET(curpmap, NULL);
1011 	#endif
1012 }
1013 
1014 void
1015 moea64_change_wiring(mmu_t mmu, pmap_t pm, vm_offset_t va, boolean_t wired)
1016 {
1017 	struct	pvo_entry *pvo;
1018 	uintptr_t pt;
1019 	uint64_t vsid;
1020 	int	i, ptegidx;
1021 
1022 	LOCK_TABLE_WR();
1023 	PMAP_LOCK(pm);
1024 	pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
1025 
1026 	if (pvo != NULL) {
1027 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
1028 
1029 		if (wired) {
1030 			if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1031 				pm->pm_stats.wired_count++;
1032 			pvo->pvo_vaddr |= PVO_WIRED;
1033 			pvo->pvo_pte.lpte.pte_hi |= LPTE_WIRED;
1034 		} else {
1035 			if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1036 				pm->pm_stats.wired_count--;
1037 			pvo->pvo_vaddr &= ~PVO_WIRED;
1038 			pvo->pvo_pte.lpte.pte_hi &= ~LPTE_WIRED;
1039 		}
1040 
1041 		if (pt != -1) {
1042 			/* Update wiring flag in page table. */
1043 			MOEA64_PTE_CHANGE(mmu, pt, &pvo->pvo_pte.lpte,
1044 			    pvo->pvo_vpn);
1045 		} else if (wired) {
1046 			/*
1047 			 * If we are wiring the page, and it wasn't in the
1048 			 * page table before, add it.
1049 			 */
1050 			vsid = PVO_VSID(pvo);
1051 			ptegidx = va_to_pteg(vsid, PVO_VADDR(pvo),
1052 			    pvo->pvo_vaddr & PVO_LARGE);
1053 
1054 			i = MOEA64_PTE_INSERT(mmu, ptegidx, &pvo->pvo_pte.lpte);
1055 
1056 			if (i >= 0) {
1057 				PVO_PTEGIDX_CLR(pvo);
1058 				PVO_PTEGIDX_SET(pvo, i);
1059 			}
1060 		}
1061 
1062 	}
1063 	UNLOCK_TABLE_WR();
1064 	PMAP_UNLOCK(pm);
1065 }
1066 
1067 /*
1068  * This goes through and sets the physical address of our
1069  * special scratch PTE to the PA we want to zero or copy. Because
1070  * of locking issues (this can get called in pvo_enter() by
1071  * the UMA allocator), we can't use most other utility functions here
1072  */
1073 
1074 static __inline
1075 void moea64_set_scratchpage_pa(mmu_t mmup, int which, vm_offset_t pa) {
1076 
1077 	KASSERT(!hw_direct_map, ("Using OEA64 scratchpage with a direct map!"));
1078 	mtx_assert(&moea64_scratchpage_mtx, MA_OWNED);
1079 
1080 	moea64_scratchpage_pvo[which]->pvo_pte.lpte.pte_lo &=
1081 	    ~(LPTE_WIMG | LPTE_RPGN);
1082 	moea64_scratchpage_pvo[which]->pvo_pte.lpte.pte_lo |=
1083 	    moea64_calc_wimg(pa, VM_MEMATTR_DEFAULT) | (uint64_t)pa;
1084 	MOEA64_PTE_CHANGE(mmup, moea64_scratchpage_pte[which],
1085 	    &moea64_scratchpage_pvo[which]->pvo_pte.lpte,
1086 	    moea64_scratchpage_pvo[which]->pvo_vpn);
1087 	isync();
1088 }
1089 
1090 void
1091 moea64_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1092 {
1093 	vm_offset_t	dst;
1094 	vm_offset_t	src;
1095 
1096 	dst = VM_PAGE_TO_PHYS(mdst);
1097 	src = VM_PAGE_TO_PHYS(msrc);
1098 
1099 	if (hw_direct_map) {
1100 		bcopy((void *)src, (void *)dst, PAGE_SIZE);
1101 	} else {
1102 		mtx_lock(&moea64_scratchpage_mtx);
1103 
1104 		moea64_set_scratchpage_pa(mmu, 0, src);
1105 		moea64_set_scratchpage_pa(mmu, 1, dst);
1106 
1107 		bcopy((void *)moea64_scratchpage_va[0],
1108 		    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1109 
1110 		mtx_unlock(&moea64_scratchpage_mtx);
1111 	}
1112 }
1113 
1114 void
1115 moea64_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1116 {
1117 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1118 
1119 	if (size + off > PAGE_SIZE)
1120 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1121 
1122 	if (hw_direct_map) {
1123 		bzero((caddr_t)pa + off, size);
1124 	} else {
1125 		mtx_lock(&moea64_scratchpage_mtx);
1126 		moea64_set_scratchpage_pa(mmu, 0, pa);
1127 		bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1128 		mtx_unlock(&moea64_scratchpage_mtx);
1129 	}
1130 }
1131 
1132 /*
1133  * Zero a page of physical memory by temporarily mapping it
1134  */
1135 void
1136 moea64_zero_page(mmu_t mmu, vm_page_t m)
1137 {
1138 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1139 	vm_offset_t va, off;
1140 
1141 	if (!hw_direct_map) {
1142 		mtx_lock(&moea64_scratchpage_mtx);
1143 
1144 		moea64_set_scratchpage_pa(mmu, 0, pa);
1145 		va = moea64_scratchpage_va[0];
1146 	} else {
1147 		va = pa;
1148 	}
1149 
1150 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1151 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1152 
1153 	if (!hw_direct_map)
1154 		mtx_unlock(&moea64_scratchpage_mtx);
1155 }
1156 
1157 void
1158 moea64_zero_page_idle(mmu_t mmu, vm_page_t m)
1159 {
1160 
1161 	moea64_zero_page(mmu, m);
1162 }
1163 
1164 /*
1165  * Map the given physical page at the specified virtual address in the
1166  * target pmap with the protection requested.  If specified the page
1167  * will be wired down.
1168  */
1169 
1170 void
1171 moea64_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m,
1172     vm_prot_t prot, boolean_t wired)
1173 {
1174 	struct		pvo_head *pvo_head;
1175 	uma_zone_t	zone;
1176 	vm_page_t	pg;
1177 	uint64_t	pte_lo;
1178 	u_int		pvo_flags;
1179 	int		error;
1180 
1181 	if (!moea64_initialized) {
1182 		pvo_head = &moea64_pvo_kunmanaged;
1183 		pg = NULL;
1184 		zone = moea64_upvo_zone;
1185 		pvo_flags = 0;
1186 	} else {
1187 		pvo_head = vm_page_to_pvoh(m);
1188 		pg = m;
1189 		zone = moea64_mpvo_zone;
1190 		pvo_flags = PVO_MANAGED;
1191 	}
1192 
1193 	KASSERT((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) != 0 ||
1194 	    VM_OBJECT_LOCKED(m->object),
1195 	    ("moea64_enter: page %p is not busy", m));
1196 
1197 	/* XXX change the pvo head for fake pages */
1198 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1199 		pvo_flags &= ~PVO_MANAGED;
1200 		pvo_head = &moea64_pvo_kunmanaged;
1201 		zone = moea64_upvo_zone;
1202 	}
1203 
1204 	pte_lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1205 
1206 	if (prot & VM_PROT_WRITE) {
1207 		pte_lo |= LPTE_BW;
1208 		if (pmap_bootstrapped &&
1209 		    (m->oflags & VPO_UNMANAGED) == 0)
1210 			vm_page_aflag_set(m, PGA_WRITEABLE);
1211 	} else
1212 		pte_lo |= LPTE_BR;
1213 
1214 	if ((prot & VM_PROT_EXECUTE) == 0)
1215 		pte_lo |= LPTE_NOEXEC;
1216 
1217 	if (wired)
1218 		pvo_flags |= PVO_WIRED;
1219 
1220 	LOCK_TABLE_WR();
1221 	PMAP_LOCK(pmap);
1222 	error = moea64_pvo_enter(mmu, pmap, zone, pvo_head, va,
1223 	    VM_PAGE_TO_PHYS(m), pte_lo, pvo_flags);
1224 	PMAP_UNLOCK(pmap);
1225 	UNLOCK_TABLE_WR();
1226 
1227 	/*
1228 	 * Flush the page from the instruction cache if this page is
1229 	 * mapped executable and cacheable.
1230 	 */
1231 	if (pmap != kernel_pmap && !(m->aflags & PGA_EXECUTABLE) &&
1232 	    (pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1233 		vm_page_aflag_set(m, PGA_EXECUTABLE);
1234 		moea64_syncicache(mmu, pmap, va, VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1235 	}
1236 }
1237 
1238 static void
1239 moea64_syncicache(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_offset_t pa,
1240     vm_size_t sz)
1241 {
1242 
1243 	/*
1244 	 * This is much trickier than on older systems because
1245 	 * we can't sync the icache on physical addresses directly
1246 	 * without a direct map. Instead we check a couple of cases
1247 	 * where the memory is already mapped in and, failing that,
1248 	 * use the same trick we use for page zeroing to create
1249 	 * a temporary mapping for this physical address.
1250 	 */
1251 
1252 	if (!pmap_bootstrapped) {
1253 		/*
1254 		 * If PMAP is not bootstrapped, we are likely to be
1255 		 * in real mode.
1256 		 */
1257 		__syncicache((void *)pa, sz);
1258 	} else if (pmap == kernel_pmap) {
1259 		__syncicache((void *)va, sz);
1260 	} else if (hw_direct_map) {
1261 		__syncicache((void *)pa, sz);
1262 	} else {
1263 		/* Use the scratch page to set up a temp mapping */
1264 
1265 		mtx_lock(&moea64_scratchpage_mtx);
1266 
1267 		moea64_set_scratchpage_pa(mmu, 1, pa & ~ADDR_POFF);
1268 		__syncicache((void *)(moea64_scratchpage_va[1] +
1269 		    (va & ADDR_POFF)), sz);
1270 
1271 		mtx_unlock(&moea64_scratchpage_mtx);
1272 	}
1273 }
1274 
1275 /*
1276  * Maps a sequence of resident pages belonging to the same object.
1277  * The sequence begins with the given page m_start.  This page is
1278  * mapped at the given virtual address start.  Each subsequent page is
1279  * mapped at a virtual address that is offset from start by the same
1280  * amount as the page is offset from m_start within the object.  The
1281  * last page in the sequence is the page with the largest offset from
1282  * m_start that can be mapped at a virtual address less than the given
1283  * virtual address end.  Not every virtual page between start and end
1284  * is mapped; only those for which a resident page exists with the
1285  * corresponding offset from m_start are mapped.
1286  */
1287 void
1288 moea64_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1289     vm_page_t m_start, vm_prot_t prot)
1290 {
1291 	vm_page_t m;
1292 	vm_pindex_t diff, psize;
1293 
1294 	psize = atop(end - start);
1295 	m = m_start;
1296 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1297 		moea64_enter(mmu, pm, start + ptoa(diff), m, prot &
1298 		    (VM_PROT_READ | VM_PROT_EXECUTE), FALSE);
1299 		m = TAILQ_NEXT(m, listq);
1300 	}
1301 }
1302 
1303 void
1304 moea64_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1305     vm_prot_t prot)
1306 {
1307 
1308 	moea64_enter(mmu, pm, va, m,
1309 	    prot & (VM_PROT_READ | VM_PROT_EXECUTE), FALSE);
1310 }
1311 
1312 vm_paddr_t
1313 moea64_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1314 {
1315 	struct	pvo_entry *pvo;
1316 	vm_paddr_t pa;
1317 
1318 	LOCK_TABLE_RD();
1319 	PMAP_LOCK(pm);
1320 	pvo = moea64_pvo_find_va(pm, va);
1321 	if (pvo == NULL)
1322 		pa = 0;
1323 	else
1324 		pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) |
1325 		    (va - PVO_VADDR(pvo));
1326 	UNLOCK_TABLE_RD();
1327 	PMAP_UNLOCK(pm);
1328 	return (pa);
1329 }
1330 
1331 /*
1332  * Atomically extract and hold the physical page with the given
1333  * pmap and virtual address pair if that mapping permits the given
1334  * protection.
1335  */
1336 
1337 extern int pa_tryrelock_restart;
1338 
1339 static int
1340 vm_page_pa_tryrelock_moea64(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
1341 {
1342 	/*
1343 	 * This is a duplicate of vm_page_pa_tryrelock(), but with proper
1344 	 * handling of the table lock
1345 	 */
1346 	vm_paddr_t lockpa;
1347 
1348 	lockpa = *locked;
1349 	*locked = pa;
1350 	if (lockpa) {
1351 		PA_LOCK_ASSERT(lockpa, MA_OWNED);
1352 		if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
1353 			return (0);
1354 		PA_UNLOCK(lockpa);
1355 	}
1356 	if (PA_TRYLOCK(pa))
1357 		return (0);
1358 	UNLOCK_TABLE_RD();
1359 	PMAP_UNLOCK(pmap);
1360 	atomic_add_int(&pa_tryrelock_restart, 1);
1361 	PA_LOCK(pa);
1362 	LOCK_TABLE_RD();
1363 	PMAP_LOCK(pmap);
1364 	return (EAGAIN);
1365 }
1366 
1367 vm_page_t
1368 moea64_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1369 {
1370 	struct	pvo_entry *pvo;
1371 	vm_page_t m;
1372         vm_paddr_t pa;
1373 
1374 	m = NULL;
1375 	pa = 0;
1376 	LOCK_TABLE_RD();
1377 	PMAP_LOCK(pmap);
1378 retry:
1379 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1380 	if (pvo != NULL && (pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) &&
1381 	    ((pvo->pvo_pte.lpte.pte_lo & LPTE_PP) == LPTE_RW ||
1382 	     (prot & VM_PROT_WRITE) == 0)) {
1383 		if (vm_page_pa_tryrelock_moea64(pmap,
1384 			pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN, &pa))
1385 			goto retry;
1386 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN);
1387 		vm_page_hold(m);
1388 	}
1389 	PA_UNLOCK_COND(pa);
1390 	UNLOCK_TABLE_RD();
1391 	PMAP_UNLOCK(pmap);
1392 	return (m);
1393 }
1394 
1395 static mmu_t installed_mmu;
1396 
1397 static void *
1398 moea64_uma_page_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
1399 {
1400 	/*
1401 	 * This entire routine is a horrible hack to avoid bothering kmem
1402 	 * for new KVA addresses. Because this can get called from inside
1403 	 * kmem allocation routines, calling kmem for a new address here
1404 	 * can lead to multiply locking non-recursive mutexes.
1405 	 */
1406         vm_offset_t va;
1407 
1408         vm_page_t m;
1409         int pflags, needed_lock;
1410 
1411 	*flags = UMA_SLAB_PRIV;
1412 	needed_lock = !PMAP_LOCKED(kernel_pmap);
1413 
1414         if ((wait & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT)
1415                 pflags = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
1416         else
1417                 pflags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED;
1418         if (wait & M_ZERO)
1419                 pflags |= VM_ALLOC_ZERO;
1420 
1421         for (;;) {
1422                 m = vm_page_alloc(NULL, 0, pflags | VM_ALLOC_NOOBJ);
1423                 if (m == NULL) {
1424                         if (wait & M_NOWAIT)
1425                                 return (NULL);
1426                         VM_WAIT;
1427                 } else
1428                         break;
1429         }
1430 
1431 	va = VM_PAGE_TO_PHYS(m);
1432 
1433 	LOCK_TABLE_WR();
1434 	if (needed_lock)
1435 		PMAP_LOCK(kernel_pmap);
1436 
1437 	moea64_pvo_enter(installed_mmu, kernel_pmap, moea64_upvo_zone,
1438 	    &moea64_pvo_kunmanaged, va, VM_PAGE_TO_PHYS(m), LPTE_M,
1439 	    PVO_WIRED | PVO_BOOTSTRAP);
1440 
1441 	if (needed_lock)
1442 		PMAP_UNLOCK(kernel_pmap);
1443 	UNLOCK_TABLE_WR();
1444 
1445 	if ((wait & M_ZERO) && (m->flags & PG_ZERO) == 0)
1446                 bzero((void *)va, PAGE_SIZE);
1447 
1448 	return (void *)va;
1449 }
1450 
1451 extern int elf32_nxstack;
1452 
1453 void
1454 moea64_init(mmu_t mmu)
1455 {
1456 
1457 	CTR0(KTR_PMAP, "moea64_init");
1458 
1459 	moea64_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1460 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1461 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1462 	moea64_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1463 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1464 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1465 
1466 	if (!hw_direct_map) {
1467 		installed_mmu = mmu;
1468 		uma_zone_set_allocf(moea64_upvo_zone,moea64_uma_page_alloc);
1469 		uma_zone_set_allocf(moea64_mpvo_zone,moea64_uma_page_alloc);
1470 	}
1471 
1472 #ifdef COMPAT_FREEBSD32
1473 	elf32_nxstack = 1;
1474 #endif
1475 
1476 	moea64_initialized = TRUE;
1477 }
1478 
1479 boolean_t
1480 moea64_is_referenced(mmu_t mmu, vm_page_t m)
1481 {
1482 
1483 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1484 	    ("moea64_is_referenced: page %p is not managed", m));
1485 	return (moea64_query_bit(mmu, m, PTE_REF));
1486 }
1487 
1488 boolean_t
1489 moea64_is_modified(mmu_t mmu, vm_page_t m)
1490 {
1491 
1492 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1493 	    ("moea64_is_modified: page %p is not managed", m));
1494 
1495 	/*
1496 	 * If the page is not VPO_BUSY, then PGA_WRITEABLE cannot be
1497 	 * concurrently set while the object is locked.  Thus, if PGA_WRITEABLE
1498 	 * is clear, no PTEs can have LPTE_CHG set.
1499 	 */
1500 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1501 	if ((m->oflags & VPO_BUSY) == 0 &&
1502 	    (m->aflags & PGA_WRITEABLE) == 0)
1503 		return (FALSE);
1504 	return (moea64_query_bit(mmu, m, LPTE_CHG));
1505 }
1506 
1507 boolean_t
1508 moea64_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1509 {
1510 	struct pvo_entry *pvo;
1511 	boolean_t rv;
1512 
1513 	LOCK_TABLE_RD();
1514 	PMAP_LOCK(pmap);
1515 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1516 	rv = pvo == NULL || (pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) == 0;
1517 	PMAP_UNLOCK(pmap);
1518 	UNLOCK_TABLE_RD();
1519 	return (rv);
1520 }
1521 
1522 void
1523 moea64_clear_reference(mmu_t mmu, vm_page_t m)
1524 {
1525 
1526 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1527 	    ("moea64_clear_reference: page %p is not managed", m));
1528 	moea64_clear_bit(mmu, m, LPTE_REF);
1529 }
1530 
1531 void
1532 moea64_clear_modify(mmu_t mmu, vm_page_t m)
1533 {
1534 
1535 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1536 	    ("moea64_clear_modify: page %p is not managed", m));
1537 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1538 	KASSERT((m->oflags & VPO_BUSY) == 0,
1539 	    ("moea64_clear_modify: page %p is busy", m));
1540 
1541 	/*
1542 	 * If the page is not PGA_WRITEABLE, then no PTEs can have LPTE_CHG
1543 	 * set.  If the object containing the page is locked and the page is
1544 	 * not VPO_BUSY, then PGA_WRITEABLE cannot be concurrently set.
1545 	 */
1546 	if ((m->aflags & PGA_WRITEABLE) == 0)
1547 		return;
1548 	moea64_clear_bit(mmu, m, LPTE_CHG);
1549 }
1550 
1551 /*
1552  * Clear the write and modified bits in each of the given page's mappings.
1553  */
1554 void
1555 moea64_remove_write(mmu_t mmu, vm_page_t m)
1556 {
1557 	struct	pvo_entry *pvo;
1558 	uintptr_t pt;
1559 	pmap_t	pmap;
1560 	uint64_t lo = 0;
1561 
1562 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1563 	    ("moea64_remove_write: page %p is not managed", m));
1564 
1565 	/*
1566 	 * If the page is not VPO_BUSY, then PGA_WRITEABLE cannot be set by
1567 	 * another thread while the object is locked.  Thus, if PGA_WRITEABLE
1568 	 * is clear, no page table entries need updating.
1569 	 */
1570 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1571 	if ((m->oflags & VPO_BUSY) == 0 &&
1572 	    (m->aflags & PGA_WRITEABLE) == 0)
1573 		return;
1574 	powerpc_sync();
1575 	LOCK_TABLE_RD();
1576 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1577 		pmap = pvo->pvo_pmap;
1578 		PMAP_LOCK(pmap);
1579 		if ((pvo->pvo_pte.lpte.pte_lo & LPTE_PP) != LPTE_BR) {
1580 			pt = MOEA64_PVO_TO_PTE(mmu, pvo);
1581 			pvo->pvo_pte.lpte.pte_lo &= ~LPTE_PP;
1582 			pvo->pvo_pte.lpte.pte_lo |= LPTE_BR;
1583 			if (pt != -1) {
1584 				MOEA64_PTE_SYNCH(mmu, pt, &pvo->pvo_pte.lpte);
1585 				lo |= pvo->pvo_pte.lpte.pte_lo;
1586 				pvo->pvo_pte.lpte.pte_lo &= ~LPTE_CHG;
1587 				MOEA64_PTE_CHANGE(mmu, pt,
1588 				    &pvo->pvo_pte.lpte, pvo->pvo_vpn);
1589 				if (pvo->pvo_pmap == kernel_pmap)
1590 					isync();
1591 			}
1592 		}
1593 		if ((lo & LPTE_CHG) != 0)
1594 			vm_page_dirty(m);
1595 		PMAP_UNLOCK(pmap);
1596 	}
1597 	UNLOCK_TABLE_RD();
1598 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1599 }
1600 
1601 /*
1602  *	moea64_ts_referenced:
1603  *
1604  *	Return a count of reference bits for a page, clearing those bits.
1605  *	It is not necessary for every reference bit to be cleared, but it
1606  *	is necessary that 0 only be returned when there are truly no
1607  *	reference bits set.
1608  *
1609  *	XXX: The exact number of bits to check and clear is a matter that
1610  *	should be tested and standardized at some point in the future for
1611  *	optimal aging of shared pages.
1612  */
1613 boolean_t
1614 moea64_ts_referenced(mmu_t mmu, vm_page_t m)
1615 {
1616 
1617 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1618 	    ("moea64_ts_referenced: page %p is not managed", m));
1619 	return (moea64_clear_bit(mmu, m, LPTE_REF));
1620 }
1621 
1622 /*
1623  * Modify the WIMG settings of all mappings for a page.
1624  */
1625 void
1626 moea64_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1627 {
1628 	struct	pvo_entry *pvo;
1629 	struct  pvo_head *pvo_head;
1630 	uintptr_t pt;
1631 	pmap_t	pmap;
1632 	uint64_t lo;
1633 
1634 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1635 		m->md.mdpg_cache_attrs = ma;
1636 		return;
1637 	}
1638 
1639 	pvo_head = vm_page_to_pvoh(m);
1640 	lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1641 	LOCK_TABLE_RD();
1642 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1643 		pmap = pvo->pvo_pmap;
1644 		PMAP_LOCK(pmap);
1645 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
1646 		pvo->pvo_pte.lpte.pte_lo &= ~LPTE_WIMG;
1647 		pvo->pvo_pte.lpte.pte_lo |= lo;
1648 		if (pt != -1) {
1649 			MOEA64_PTE_CHANGE(mmu, pt, &pvo->pvo_pte.lpte,
1650 			    pvo->pvo_vpn);
1651 			if (pvo->pvo_pmap == kernel_pmap)
1652 				isync();
1653 		}
1654 		PMAP_UNLOCK(pmap);
1655 	}
1656 	UNLOCK_TABLE_RD();
1657 	m->md.mdpg_cache_attrs = ma;
1658 }
1659 
1660 /*
1661  * Map a wired page into kernel virtual address space.
1662  */
1663 void
1664 moea64_kenter_attr(mmu_t mmu, vm_offset_t va, vm_offset_t pa, vm_memattr_t ma)
1665 {
1666 	uint64_t	pte_lo;
1667 	int		error;
1668 
1669 	pte_lo = moea64_calc_wimg(pa, ma);
1670 
1671 	LOCK_TABLE_WR();
1672 	PMAP_LOCK(kernel_pmap);
1673 	error = moea64_pvo_enter(mmu, kernel_pmap, moea64_upvo_zone,
1674 	    &moea64_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1675 	PMAP_UNLOCK(kernel_pmap);
1676 	UNLOCK_TABLE_WR();
1677 
1678 	if (error != 0 && error != ENOENT)
1679 		panic("moea64_kenter: failed to enter va %#zx pa %#zx: %d", va,
1680 		    pa, error);
1681 }
1682 
1683 void
1684 moea64_kenter(mmu_t mmu, vm_offset_t va, vm_offset_t pa)
1685 {
1686 
1687 	moea64_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1688 }
1689 
1690 /*
1691  * Extract the physical page address associated with the given kernel virtual
1692  * address.
1693  */
1694 vm_offset_t
1695 moea64_kextract(mmu_t mmu, vm_offset_t va)
1696 {
1697 	struct		pvo_entry *pvo;
1698 	vm_paddr_t pa;
1699 
1700 	/*
1701 	 * Shortcut the direct-mapped case when applicable.  We never put
1702 	 * anything but 1:1 mappings below VM_MIN_KERNEL_ADDRESS.
1703 	 */
1704 	if (va < VM_MIN_KERNEL_ADDRESS)
1705 		return (va);
1706 
1707 	LOCK_TABLE_RD();
1708 	PMAP_LOCK(kernel_pmap);
1709 	pvo = moea64_pvo_find_va(kernel_pmap, va);
1710 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found for %#" PRIxPTR,
1711 	    va));
1712 	pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) | (va - PVO_VADDR(pvo));
1713 	UNLOCK_TABLE_RD();
1714 	PMAP_UNLOCK(kernel_pmap);
1715 	return (pa);
1716 }
1717 
1718 /*
1719  * Remove a wired page from kernel virtual address space.
1720  */
1721 void
1722 moea64_kremove(mmu_t mmu, vm_offset_t va)
1723 {
1724 	moea64_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1725 }
1726 
1727 /*
1728  * Map a range of physical addresses into kernel virtual address space.
1729  *
1730  * The value passed in *virt is a suggested virtual address for the mapping.
1731  * Architectures which can support a direct-mapped physical to virtual region
1732  * can return the appropriate address within that region, leaving '*virt'
1733  * unchanged.  We cannot and therefore do not; *virt is updated with the
1734  * first usable address after the mapped region.
1735  */
1736 vm_offset_t
1737 moea64_map(mmu_t mmu, vm_offset_t *virt, vm_offset_t pa_start,
1738     vm_offset_t pa_end, int prot)
1739 {
1740 	vm_offset_t	sva, va;
1741 
1742 	sva = *virt;
1743 	va = sva;
1744 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1745 		moea64_kenter(mmu, va, pa_start);
1746 	*virt = va;
1747 
1748 	return (sva);
1749 }
1750 
1751 /*
1752  * Returns true if the pmap's pv is one of the first
1753  * 16 pvs linked to from this page.  This count may
1754  * be changed upwards or downwards in the future; it
1755  * is only necessary that true be returned for a small
1756  * subset of pmaps for proper page aging.
1757  */
1758 boolean_t
1759 moea64_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1760 {
1761         int loops;
1762 	struct pvo_entry *pvo;
1763 	boolean_t rv;
1764 
1765 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1766 	    ("moea64_page_exists_quick: page %p is not managed", m));
1767 	loops = 0;
1768 	rv = FALSE;
1769 	LOCK_TABLE_RD();
1770 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1771 		if (pvo->pvo_pmap == pmap) {
1772 			rv = TRUE;
1773 			break;
1774 		}
1775 		if (++loops >= 16)
1776 			break;
1777 	}
1778 	UNLOCK_TABLE_RD();
1779 	return (rv);
1780 }
1781 
1782 /*
1783  * Return the number of managed mappings to the given physical page
1784  * that are wired.
1785  */
1786 int
1787 moea64_page_wired_mappings(mmu_t mmu, vm_page_t m)
1788 {
1789 	struct pvo_entry *pvo;
1790 	int count;
1791 
1792 	count = 0;
1793 	if ((m->oflags & VPO_UNMANAGED) != 0)
1794 		return (count);
1795 	LOCK_TABLE_RD();
1796 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1797 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1798 			count++;
1799 	UNLOCK_TABLE_RD();
1800 	return (count);
1801 }
1802 
1803 static uintptr_t	moea64_vsidcontext;
1804 
1805 uintptr_t
1806 moea64_get_unique_vsid(void) {
1807 	u_int entropy;
1808 	register_t hash;
1809 	uint32_t mask;
1810 	int i;
1811 
1812 	entropy = 0;
1813 	__asm __volatile("mftb %0" : "=r"(entropy));
1814 
1815 	mtx_lock(&moea64_slb_mutex);
1816 	for (i = 0; i < NVSIDS; i += VSID_NBPW) {
1817 		u_int	n;
1818 
1819 		/*
1820 		 * Create a new value by mutiplying by a prime and adding in
1821 		 * entropy from the timebase register.  This is to make the
1822 		 * VSID more random so that the PT hash function collides
1823 		 * less often.  (Note that the prime casues gcc to do shifts
1824 		 * instead of a multiply.)
1825 		 */
1826 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
1827 		hash = moea64_vsidcontext & (NVSIDS - 1);
1828 		if (hash == 0)		/* 0 is special, avoid it */
1829 			continue;
1830 		n = hash >> 5;
1831 		mask = 1 << (hash & (VSID_NBPW - 1));
1832 		hash = (moea64_vsidcontext & VSID_HASHMASK);
1833 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
1834 			/* anything free in this bucket? */
1835 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
1836 				entropy = (moea64_vsidcontext >> 20);
1837 				continue;
1838 			}
1839 			i = ffs(~moea64_vsid_bitmap[n]) - 1;
1840 			mask = 1 << i;
1841 			hash &= VSID_HASHMASK & ~(VSID_NBPW - 1);
1842 			hash |= i;
1843 		}
1844 		KASSERT(!(moea64_vsid_bitmap[n] & mask),
1845 		    ("Allocating in-use VSID %#zx\n", hash));
1846 		moea64_vsid_bitmap[n] |= mask;
1847 		mtx_unlock(&moea64_slb_mutex);
1848 		return (hash);
1849 	}
1850 
1851 	mtx_unlock(&moea64_slb_mutex);
1852 	panic("%s: out of segments",__func__);
1853 }
1854 
1855 #ifdef __powerpc64__
1856 void
1857 moea64_pinit(mmu_t mmu, pmap_t pmap)
1858 {
1859 	PMAP_LOCK_INIT(pmap);
1860 	LIST_INIT(&pmap->pmap_pvo);
1861 
1862 	pmap->pm_slb_tree_root = slb_alloc_tree();
1863 	pmap->pm_slb = slb_alloc_user_cache();
1864 	pmap->pm_slb_len = 0;
1865 }
1866 #else
1867 void
1868 moea64_pinit(mmu_t mmu, pmap_t pmap)
1869 {
1870 	int	i;
1871 	uint32_t hash;
1872 
1873 	PMAP_LOCK_INIT(pmap);
1874 	LIST_INIT(&pmap->pmap_pvo);
1875 
1876 	if (pmap_bootstrapped)
1877 		pmap->pmap_phys = (pmap_t)moea64_kextract(mmu,
1878 		    (vm_offset_t)pmap);
1879 	else
1880 		pmap->pmap_phys = pmap;
1881 
1882 	/*
1883 	 * Allocate some segment registers for this pmap.
1884 	 */
1885 	hash = moea64_get_unique_vsid();
1886 
1887 	for (i = 0; i < 16; i++)
1888 		pmap->pm_sr[i] = VSID_MAKE(i, hash);
1889 
1890 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_pinit: pm_sr[0] = 0"));
1891 }
1892 #endif
1893 
1894 /*
1895  * Initialize the pmap associated with process 0.
1896  */
1897 void
1898 moea64_pinit0(mmu_t mmu, pmap_t pm)
1899 {
1900 	moea64_pinit(mmu, pm);
1901 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1902 }
1903 
1904 /*
1905  * Set the physical protection on the specified range of this map as requested.
1906  */
1907 static void
1908 moea64_pvo_protect(mmu_t mmu,  pmap_t pm, struct pvo_entry *pvo, vm_prot_t prot)
1909 {
1910 	uintptr_t pt;
1911 	struct	vm_page *pg;
1912 	uint64_t oldlo;
1913 
1914 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
1915 
1916 	/*
1917 	 * Grab the PTE pointer before we diddle with the cached PTE
1918 	 * copy.
1919 	 */
1920 	pt = MOEA64_PVO_TO_PTE(mmu, pvo);
1921 
1922 	/*
1923 	 * Change the protection of the page.
1924 	 */
1925 	oldlo = pvo->pvo_pte.lpte.pte_lo;
1926 	pvo->pvo_pte.lpte.pte_lo &= ~LPTE_PP;
1927 	pvo->pvo_pte.lpte.pte_lo &= ~LPTE_NOEXEC;
1928 	if ((prot & VM_PROT_EXECUTE) == 0)
1929 		pvo->pvo_pte.lpte.pte_lo |= LPTE_NOEXEC;
1930 	if (prot & VM_PROT_WRITE)
1931 		pvo->pvo_pte.lpte.pte_lo |= LPTE_BW;
1932 	else
1933 		pvo->pvo_pte.lpte.pte_lo |= LPTE_BR;
1934 
1935 	pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN);
1936 
1937 	/*
1938 	 * If the PVO is in the page table, update that pte as well.
1939 	 */
1940 	if (pt != -1)
1941 		MOEA64_PTE_CHANGE(mmu, pt, &pvo->pvo_pte.lpte,
1942 		    pvo->pvo_vpn);
1943 	if (pm != kernel_pmap && pg != NULL && !(pg->aflags & PGA_EXECUTABLE) &&
1944 	    (pvo->pvo_pte.lpte.pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1945 		if ((pg->oflags & VPO_UNMANAGED) == 0)
1946 			vm_page_aflag_set(pg, PGA_EXECUTABLE);
1947 		moea64_syncicache(mmu, pm, PVO_VADDR(pvo),
1948 		    pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN, PAGE_SIZE);
1949 	}
1950 
1951 	/*
1952 	 * Update vm about the REF/CHG bits if the page is managed and we have
1953 	 * removed write access.
1954 	 */
1955 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED &&
1956 	    (oldlo & LPTE_PP) != LPTE_BR && !(prot && VM_PROT_WRITE)) {
1957 		if (pg != NULL) {
1958 			if (pvo->pvo_pte.lpte.pte_lo & LPTE_CHG)
1959 				vm_page_dirty(pg);
1960 			if (pvo->pvo_pte.lpte.pte_lo & LPTE_REF)
1961 				vm_page_aflag_set(pg, PGA_REFERENCED);
1962 		}
1963 	}
1964 }
1965 
1966 void
1967 moea64_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1968     vm_prot_t prot)
1969 {
1970 	struct	pvo_entry *pvo, *tpvo;
1971 
1972 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm,
1973 	    sva, eva, prot);
1974 
1975 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1976 	    ("moea64_protect: non current pmap"));
1977 
1978 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1979 		moea64_remove(mmu, pm, sva, eva);
1980 		return;
1981 	}
1982 
1983 	LOCK_TABLE_RD();
1984 	PMAP_LOCK(pm);
1985 	if ((eva - sva)/PAGE_SIZE < pm->pm_stats.resident_count) {
1986 		while (sva < eva) {
1987 			#ifdef __powerpc64__
1988 			if (pm != kernel_pmap &&
1989 			    user_va_to_slb_entry(pm, sva) == NULL) {
1990 				sva = roundup2(sva + 1, SEGMENT_LENGTH);
1991 				continue;
1992 			}
1993 			#endif
1994 			pvo = moea64_pvo_find_va(pm, sva);
1995 			if (pvo != NULL)
1996 				moea64_pvo_protect(mmu, pm, pvo, prot);
1997 			sva += PAGE_SIZE;
1998 		}
1999 	} else {
2000 		LIST_FOREACH_SAFE(pvo, &pm->pmap_pvo, pvo_plink, tpvo) {
2001 			if (PVO_VADDR(pvo) < sva || PVO_VADDR(pvo) >= eva)
2002 				continue;
2003 			moea64_pvo_protect(mmu, pm, pvo, prot);
2004 		}
2005 	}
2006 	UNLOCK_TABLE_RD();
2007 	PMAP_UNLOCK(pm);
2008 }
2009 
2010 /*
2011  * Map a list of wired pages into kernel virtual address space.  This is
2012  * intended for temporary mappings which do not need page modification or
2013  * references recorded.  Existing mappings in the region are overwritten.
2014  */
2015 void
2016 moea64_qenter(mmu_t mmu, vm_offset_t va, vm_page_t *m, int count)
2017 {
2018 	while (count-- > 0) {
2019 		moea64_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
2020 		va += PAGE_SIZE;
2021 		m++;
2022 	}
2023 }
2024 
2025 /*
2026  * Remove page mappings from kernel virtual address space.  Intended for
2027  * temporary mappings entered by moea64_qenter.
2028  */
2029 void
2030 moea64_qremove(mmu_t mmu, vm_offset_t va, int count)
2031 {
2032 	while (count-- > 0) {
2033 		moea64_kremove(mmu, va);
2034 		va += PAGE_SIZE;
2035 	}
2036 }
2037 
2038 void
2039 moea64_release_vsid(uint64_t vsid)
2040 {
2041 	int idx, mask;
2042 
2043 	mtx_lock(&moea64_slb_mutex);
2044 	idx = vsid & (NVSIDS-1);
2045 	mask = 1 << (idx % VSID_NBPW);
2046 	idx /= VSID_NBPW;
2047 	KASSERT(moea64_vsid_bitmap[idx] & mask,
2048 	    ("Freeing unallocated VSID %#jx", vsid));
2049 	moea64_vsid_bitmap[idx] &= ~mask;
2050 	mtx_unlock(&moea64_slb_mutex);
2051 }
2052 
2053 
2054 void
2055 moea64_release(mmu_t mmu, pmap_t pmap)
2056 {
2057 
2058 	/*
2059 	 * Free segment registers' VSIDs
2060 	 */
2061     #ifdef __powerpc64__
2062 	slb_free_tree(pmap);
2063 	slb_free_user_cache(pmap->pm_slb);
2064     #else
2065 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_release: pm_sr[0] = 0"));
2066 
2067 	moea64_release_vsid(VSID_TO_HASH(pmap->pm_sr[0]));
2068     #endif
2069 
2070 	PMAP_LOCK_DESTROY(pmap);
2071 }
2072 
2073 /*
2074  * Remove all pages mapped by the specified pmap
2075  */
2076 void
2077 moea64_remove_pages(mmu_t mmu, pmap_t pm)
2078 {
2079 	struct	pvo_entry *pvo, *tpvo;
2080 
2081 	LOCK_TABLE_WR();
2082 	PMAP_LOCK(pm);
2083 	LIST_FOREACH_SAFE(pvo, &pm->pmap_pvo, pvo_plink, tpvo) {
2084 		if (!(pvo->pvo_vaddr & PVO_WIRED))
2085 			moea64_pvo_remove(mmu, pvo);
2086 	}
2087 	UNLOCK_TABLE_WR();
2088 	PMAP_UNLOCK(pm);
2089 }
2090 
2091 /*
2092  * Remove the given range of addresses from the specified map.
2093  */
2094 void
2095 moea64_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
2096 {
2097 	struct	pvo_entry *pvo, *tpvo;
2098 
2099 	/*
2100 	 * Perform an unsynchronized read.  This is, however, safe.
2101 	 */
2102 	if (pm->pm_stats.resident_count == 0)
2103 		return;
2104 
2105 	LOCK_TABLE_WR();
2106 	PMAP_LOCK(pm);
2107 	if ((eva - sva)/PAGE_SIZE < pm->pm_stats.resident_count) {
2108 		while (sva < eva) {
2109 			#ifdef __powerpc64__
2110 			if (pm != kernel_pmap &&
2111 			    user_va_to_slb_entry(pm, sva) == NULL) {
2112 				sva = roundup2(sva + 1, SEGMENT_LENGTH);
2113 				continue;
2114 			}
2115 			#endif
2116 			pvo = moea64_pvo_find_va(pm, sva);
2117 			if (pvo != NULL)
2118 				moea64_pvo_remove(mmu, pvo);
2119 			sva += PAGE_SIZE;
2120 		}
2121 	} else {
2122 		LIST_FOREACH_SAFE(pvo, &pm->pmap_pvo, pvo_plink, tpvo) {
2123 			if (PVO_VADDR(pvo) < sva || PVO_VADDR(pvo) >= eva)
2124 				continue;
2125 			moea64_pvo_remove(mmu, pvo);
2126 		}
2127 	}
2128 	UNLOCK_TABLE_WR();
2129 	PMAP_UNLOCK(pm);
2130 }
2131 
2132 /*
2133  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
2134  * will reflect changes in pte's back to the vm_page.
2135  */
2136 void
2137 moea64_remove_all(mmu_t mmu, vm_page_t m)
2138 {
2139 	struct	pvo_entry *pvo, *next_pvo;
2140 	pmap_t	pmap;
2141 
2142 	LOCK_TABLE_WR();
2143 	LIST_FOREACH_SAFE(pvo, vm_page_to_pvoh(m), pvo_vlink, next_pvo) {
2144 		pmap = pvo->pvo_pmap;
2145 		PMAP_LOCK(pmap);
2146 		moea64_pvo_remove(mmu, pvo);
2147 		PMAP_UNLOCK(pmap);
2148 	}
2149 	UNLOCK_TABLE_WR();
2150 	if ((m->aflags & PGA_WRITEABLE) && moea64_is_modified(mmu, m))
2151 		vm_page_dirty(m);
2152 	vm_page_aflag_clear(m, PGA_WRITEABLE);
2153 	vm_page_aflag_clear(m, PGA_EXECUTABLE);
2154 }
2155 
2156 /*
2157  * Allocate a physical page of memory directly from the phys_avail map.
2158  * Can only be called from moea64_bootstrap before avail start and end are
2159  * calculated.
2160  */
2161 vm_offset_t
2162 moea64_bootstrap_alloc(vm_size_t size, u_int align)
2163 {
2164 	vm_offset_t	s, e;
2165 	int		i, j;
2166 
2167 	size = round_page(size);
2168 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2169 		if (align != 0)
2170 			s = (phys_avail[i] + align - 1) & ~(align - 1);
2171 		else
2172 			s = phys_avail[i];
2173 		e = s + size;
2174 
2175 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2176 			continue;
2177 
2178 		if (s + size > platform_real_maxaddr())
2179 			continue;
2180 
2181 		if (s == phys_avail[i]) {
2182 			phys_avail[i] += size;
2183 		} else if (e == phys_avail[i + 1]) {
2184 			phys_avail[i + 1] -= size;
2185 		} else {
2186 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2187 				phys_avail[j] = phys_avail[j - 2];
2188 				phys_avail[j + 1] = phys_avail[j - 1];
2189 			}
2190 
2191 			phys_avail[i + 3] = phys_avail[i + 1];
2192 			phys_avail[i + 1] = s;
2193 			phys_avail[i + 2] = e;
2194 			phys_avail_count++;
2195 		}
2196 
2197 		return (s);
2198 	}
2199 	panic("moea64_bootstrap_alloc: could not allocate memory");
2200 }
2201 
2202 static int
2203 moea64_pvo_enter(mmu_t mmu, pmap_t pm, uma_zone_t zone,
2204     struct pvo_head *pvo_head, vm_offset_t va, vm_offset_t pa,
2205     uint64_t pte_lo, int flags)
2206 {
2207 	struct	 pvo_entry *pvo;
2208 	uint64_t vsid;
2209 	int	 first;
2210 	u_int	 ptegidx;
2211 	int	 i;
2212 	int      bootstrap;
2213 
2214 	/*
2215 	 * One nasty thing that can happen here is that the UMA calls to
2216 	 * allocate new PVOs need to map more memory, which calls pvo_enter(),
2217 	 * which calls UMA...
2218 	 *
2219 	 * We break the loop by detecting recursion and allocating out of
2220 	 * the bootstrap pool.
2221 	 */
2222 
2223 	first = 0;
2224 	bootstrap = (flags & PVO_BOOTSTRAP);
2225 
2226 	if (!moea64_initialized)
2227 		bootstrap = 1;
2228 
2229 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2230 	rw_assert(&moea64_table_lock, RA_WLOCKED);
2231 
2232 	/*
2233 	 * Compute the PTE Group index.
2234 	 */
2235 	va &= ~ADDR_POFF;
2236 	vsid = va_to_vsid(pm, va);
2237 	ptegidx = va_to_pteg(vsid, va, flags & PVO_LARGE);
2238 
2239 	/*
2240 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
2241 	 * there is a mapping.
2242 	 */
2243 	moea64_pvo_enter_calls++;
2244 
2245 	LIST_FOREACH(pvo, &moea64_pvo_table[ptegidx], pvo_olink) {
2246 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2247 			if ((pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) == pa &&
2248 			    (pvo->pvo_pte.lpte.pte_lo & (LPTE_NOEXEC | LPTE_PP))
2249 			    == (pte_lo & (LPTE_NOEXEC | LPTE_PP))) {
2250 			    	if (!(pvo->pvo_pte.lpte.pte_hi & LPTE_VALID)) {
2251 					/* Re-insert if spilled */
2252 					i = MOEA64_PTE_INSERT(mmu, ptegidx,
2253 					    &pvo->pvo_pte.lpte);
2254 					if (i >= 0)
2255 						PVO_PTEGIDX_SET(pvo, i);
2256 					moea64_pte_overflow--;
2257 				}
2258 				return (0);
2259 			}
2260 			moea64_pvo_remove(mmu, pvo);
2261 			break;
2262 		}
2263 	}
2264 
2265 	/*
2266 	 * If we aren't overwriting a mapping, try to allocate.
2267 	 */
2268 	if (bootstrap) {
2269 		if (moea64_bpvo_pool_index >= BPVO_POOL_SIZE) {
2270 			panic("moea64_enter: bpvo pool exhausted, %d, %d, %zd",
2271 			      moea64_bpvo_pool_index, BPVO_POOL_SIZE,
2272 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
2273 		}
2274 		pvo = &moea64_bpvo_pool[moea64_bpvo_pool_index];
2275 		moea64_bpvo_pool_index++;
2276 		bootstrap = 1;
2277 	} else {
2278 		/*
2279 		 * Note: drop the table lock around the UMA allocation in
2280 		 * case the UMA allocator needs to manipulate the page
2281 		 * table. The mapping we are working with is already
2282 		 * protected by the PMAP lock.
2283 		 */
2284 		pvo = uma_zalloc(zone, M_NOWAIT);
2285 	}
2286 
2287 	if (pvo == NULL)
2288 		return (ENOMEM);
2289 
2290 	moea64_pvo_entries++;
2291 	pvo->pvo_vaddr = va;
2292 	pvo->pvo_vpn = (uint64_t)((va & ADDR_PIDX) >> ADDR_PIDX_SHFT)
2293 	    | (vsid << 16);
2294 	pvo->pvo_pmap = pm;
2295 	LIST_INSERT_HEAD(&moea64_pvo_table[ptegidx], pvo, pvo_olink);
2296 	pvo->pvo_vaddr &= ~ADDR_POFF;
2297 
2298 	if (flags & PVO_WIRED)
2299 		pvo->pvo_vaddr |= PVO_WIRED;
2300 	if (pvo_head != &moea64_pvo_kunmanaged)
2301 		pvo->pvo_vaddr |= PVO_MANAGED;
2302 	if (bootstrap)
2303 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
2304 	if (flags & PVO_LARGE)
2305 		pvo->pvo_vaddr |= PVO_LARGE;
2306 
2307 	moea64_pte_create(&pvo->pvo_pte.lpte, vsid, va,
2308 	    (uint64_t)(pa) | pte_lo, flags);
2309 
2310 	/*
2311 	 * Add to pmap list
2312 	 */
2313 	LIST_INSERT_HEAD(&pm->pmap_pvo, pvo, pvo_plink);
2314 
2315 	/*
2316 	 * Remember if the list was empty and therefore will be the first
2317 	 * item.
2318 	 */
2319 	if (LIST_FIRST(pvo_head) == NULL)
2320 		first = 1;
2321 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2322 
2323 	if (pvo->pvo_vaddr & PVO_WIRED) {
2324 		pvo->pvo_pte.lpte.pte_hi |= LPTE_WIRED;
2325 		pm->pm_stats.wired_count++;
2326 	}
2327 	pm->pm_stats.resident_count++;
2328 
2329 	/*
2330 	 * We hope this succeeds but it isn't required.
2331 	 */
2332 	i = MOEA64_PTE_INSERT(mmu, ptegidx, &pvo->pvo_pte.lpte);
2333 	if (i >= 0) {
2334 		PVO_PTEGIDX_SET(pvo, i);
2335 	} else {
2336 		panic("moea64_pvo_enter: overflow");
2337 		moea64_pte_overflow++;
2338 	}
2339 
2340 	if (pm == kernel_pmap)
2341 		isync();
2342 
2343 #ifdef __powerpc64__
2344 	/*
2345 	 * Make sure all our bootstrap mappings are in the SLB as soon
2346 	 * as virtual memory is switched on.
2347 	 */
2348 	if (!pmap_bootstrapped)
2349 		moea64_bootstrap_slb_prefault(va, flags & PVO_LARGE);
2350 #endif
2351 
2352 	return (first ? ENOENT : 0);
2353 }
2354 
2355 static void
2356 moea64_pvo_remove(mmu_t mmu, struct pvo_entry *pvo)
2357 {
2358 	struct	vm_page *pg;
2359 	uintptr_t pt;
2360 
2361 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2362 	rw_assert(&moea64_table_lock, RA_WLOCKED);
2363 
2364 	/*
2365 	 * If there is an active pte entry, we need to deactivate it (and
2366 	 * save the ref & cfg bits).
2367 	 */
2368 	pt = MOEA64_PVO_TO_PTE(mmu, pvo);
2369 	if (pt != -1) {
2370 		MOEA64_PTE_UNSET(mmu, pt, &pvo->pvo_pte.lpte, pvo->pvo_vpn);
2371 		PVO_PTEGIDX_CLR(pvo);
2372 	} else {
2373 		moea64_pte_overflow--;
2374 	}
2375 
2376 	/*
2377 	 * Update our statistics.
2378 	 */
2379 	pvo->pvo_pmap->pm_stats.resident_count--;
2380 	if (pvo->pvo_vaddr & PVO_WIRED)
2381 		pvo->pvo_pmap->pm_stats.wired_count--;
2382 
2383 	/*
2384 	 * Remove this PVO from the PV and pmap lists.
2385 	 */
2386 	LIST_REMOVE(pvo, pvo_vlink);
2387 	LIST_REMOVE(pvo, pvo_plink);
2388 
2389 	/*
2390 	 * Remove this from the overflow list and return it to the pool
2391 	 * if we aren't going to reuse it.
2392 	 */
2393 	LIST_REMOVE(pvo, pvo_olink);
2394 
2395 	/*
2396 	 * Update vm about the REF/CHG bits if the page is managed.
2397 	 */
2398 	pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN);
2399 
2400 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED && pg != NULL) {
2401 		if ((pvo->pvo_pte.lpte.pte_lo & LPTE_PP) != LPTE_BR) {
2402 			if (pvo->pvo_pte.lpte.pte_lo & LPTE_CHG)
2403 				vm_page_dirty(pg);
2404 			if (pvo->pvo_pte.lpte.pte_lo & LPTE_REF)
2405 				vm_page_aflag_set(pg, PGA_REFERENCED);
2406 			if (LIST_EMPTY(vm_page_to_pvoh(pg)))
2407 				vm_page_aflag_clear(pg, PGA_WRITEABLE);
2408 		}
2409 		if (LIST_EMPTY(vm_page_to_pvoh(pg)))
2410 			vm_page_aflag_clear(pg, PGA_EXECUTABLE);
2411 	}
2412 
2413 	moea64_pvo_entries--;
2414 	moea64_pvo_remove_calls++;
2415 
2416 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2417 		uma_zfree((pvo->pvo_vaddr & PVO_MANAGED) ? moea64_mpvo_zone :
2418 		    moea64_upvo_zone, pvo);
2419 }
2420 
2421 static struct pvo_entry *
2422 moea64_pvo_find_va(pmap_t pm, vm_offset_t va)
2423 {
2424 	struct		pvo_entry *pvo;
2425 	int		ptegidx;
2426 	uint64_t	vsid;
2427 	#ifdef __powerpc64__
2428 	uint64_t	slbv;
2429 
2430 	if (pm == kernel_pmap) {
2431 		slbv = kernel_va_to_slbv(va);
2432 	} else {
2433 		struct slb *slb;
2434 		slb = user_va_to_slb_entry(pm, va);
2435 		/* The page is not mapped if the segment isn't */
2436 		if (slb == NULL)
2437 			return NULL;
2438 		slbv = slb->slbv;
2439 	}
2440 
2441 	vsid = (slbv & SLBV_VSID_MASK) >> SLBV_VSID_SHIFT;
2442 	if (slbv & SLBV_L)
2443 		va &= ~moea64_large_page_mask;
2444 	else
2445 		va &= ~ADDR_POFF;
2446 	ptegidx = va_to_pteg(vsid, va, slbv & SLBV_L);
2447 	#else
2448 	va &= ~ADDR_POFF;
2449 	vsid = va_to_vsid(pm, va);
2450 	ptegidx = va_to_pteg(vsid, va, 0);
2451 	#endif
2452 
2453 	LIST_FOREACH(pvo, &moea64_pvo_table[ptegidx], pvo_olink) {
2454 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va)
2455 			break;
2456 	}
2457 
2458 	return (pvo);
2459 }
2460 
2461 static boolean_t
2462 moea64_query_bit(mmu_t mmu, vm_page_t m, u_int64_t ptebit)
2463 {
2464 	struct	pvo_entry *pvo;
2465 	uintptr_t pt;
2466 
2467 	LOCK_TABLE_RD();
2468 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2469 		/*
2470 		 * See if we saved the bit off.  If so, return success.
2471 		 */
2472 		if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2473 			UNLOCK_TABLE_RD();
2474 			return (TRUE);
2475 		}
2476 	}
2477 
2478 	/*
2479 	 * No luck, now go through the hard part of looking at the PTEs
2480 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2481 	 * the PTEs.
2482 	 */
2483 	powerpc_sync();
2484 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2485 
2486 		/*
2487 		 * See if this pvo has a valid PTE.  if so, fetch the
2488 		 * REF/CHG bits from the valid PTE.  If the appropriate
2489 		 * ptebit is set, return success.
2490 		 */
2491 		PMAP_LOCK(pvo->pvo_pmap);
2492 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
2493 		if (pt != -1) {
2494 			MOEA64_PTE_SYNCH(mmu, pt, &pvo->pvo_pte.lpte);
2495 			if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2496 				PMAP_UNLOCK(pvo->pvo_pmap);
2497 				UNLOCK_TABLE_RD();
2498 				return (TRUE);
2499 			}
2500 		}
2501 		PMAP_UNLOCK(pvo->pvo_pmap);
2502 	}
2503 
2504 	UNLOCK_TABLE_RD();
2505 	return (FALSE);
2506 }
2507 
2508 static u_int
2509 moea64_clear_bit(mmu_t mmu, vm_page_t m, u_int64_t ptebit)
2510 {
2511 	u_int	count;
2512 	struct	pvo_entry *pvo;
2513 	uintptr_t pt;
2514 
2515 	/*
2516 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2517 	 * we can reset the right ones).  note that since the pvo entries and
2518 	 * list heads are accessed via BAT0 and are never placed in the page
2519 	 * table, we don't have to worry about further accesses setting the
2520 	 * REF/CHG bits.
2521 	 */
2522 	powerpc_sync();
2523 
2524 	/*
2525 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2526 	 * valid pte clear the ptebit from the valid pte.
2527 	 */
2528 	count = 0;
2529 	LOCK_TABLE_RD();
2530 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2531 		PMAP_LOCK(pvo->pvo_pmap);
2532 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
2533 		if (pt != -1) {
2534 			MOEA64_PTE_SYNCH(mmu, pt, &pvo->pvo_pte.lpte);
2535 			if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2536 				count++;
2537 				MOEA64_PTE_CLEAR(mmu, pt, &pvo->pvo_pte.lpte,
2538 				    pvo->pvo_vpn, ptebit);
2539 			}
2540 		}
2541 		pvo->pvo_pte.lpte.pte_lo &= ~ptebit;
2542 		PMAP_UNLOCK(pvo->pvo_pmap);
2543 	}
2544 
2545 	UNLOCK_TABLE_RD();
2546 	return (count);
2547 }
2548 
2549 boolean_t
2550 moea64_dev_direct_mapped(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2551 {
2552 	struct pvo_entry *pvo;
2553 	vm_offset_t ppa;
2554 	int error = 0;
2555 
2556 	LOCK_TABLE_RD();
2557 	PMAP_LOCK(kernel_pmap);
2558 	for (ppa = pa & ~ADDR_POFF; ppa < pa + size; ppa += PAGE_SIZE) {
2559 		pvo = moea64_pvo_find_va(kernel_pmap, ppa);
2560 		if (pvo == NULL ||
2561 		    (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) != ppa) {
2562 			error = EFAULT;
2563 			break;
2564 		}
2565 	}
2566 	UNLOCK_TABLE_RD();
2567 	PMAP_UNLOCK(kernel_pmap);
2568 
2569 	return (error);
2570 }
2571 
2572 /*
2573  * Map a set of physical memory pages into the kernel virtual
2574  * address space. Return a pointer to where it is mapped. This
2575  * routine is intended to be used for mapping device memory,
2576  * NOT real memory.
2577  */
2578 void *
2579 moea64_mapdev_attr(mmu_t mmu, vm_offset_t pa, vm_size_t size, vm_memattr_t ma)
2580 {
2581 	vm_offset_t va, tmpva, ppa, offset;
2582 
2583 	ppa = trunc_page(pa);
2584 	offset = pa & PAGE_MASK;
2585 	size = roundup2(offset + size, PAGE_SIZE);
2586 
2587 	va = kmem_alloc_nofault(kernel_map, size);
2588 
2589 	if (!va)
2590 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
2591 
2592 	for (tmpva = va; size > 0;) {
2593 		moea64_kenter_attr(mmu, tmpva, ppa, ma);
2594 		size -= PAGE_SIZE;
2595 		tmpva += PAGE_SIZE;
2596 		ppa += PAGE_SIZE;
2597 	}
2598 
2599 	return ((void *)(va + offset));
2600 }
2601 
2602 void *
2603 moea64_mapdev(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2604 {
2605 
2606 	return moea64_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT);
2607 }
2608 
2609 void
2610 moea64_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2611 {
2612 	vm_offset_t base, offset;
2613 
2614 	base = trunc_page(va);
2615 	offset = va & PAGE_MASK;
2616 	size = roundup2(offset + size, PAGE_SIZE);
2617 
2618 	kmem_free(kernel_map, base, size);
2619 }
2620 
2621 void
2622 moea64_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2623 {
2624 	struct pvo_entry *pvo;
2625 	vm_offset_t lim;
2626 	vm_paddr_t pa;
2627 	vm_size_t len;
2628 
2629 	LOCK_TABLE_RD();
2630 	PMAP_LOCK(pm);
2631 	while (sz > 0) {
2632 		lim = round_page(va);
2633 		len = MIN(lim - va, sz);
2634 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
2635 		if (pvo != NULL && !(pvo->pvo_pte.lpte.pte_lo & LPTE_I)) {
2636 			pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) |
2637 			    (va & ADDR_POFF);
2638 			moea64_syncicache(mmu, pm, va, pa, len);
2639 		}
2640 		va += len;
2641 		sz -= len;
2642 	}
2643 	UNLOCK_TABLE_RD();
2644 	PMAP_UNLOCK(pm);
2645 }
2646