xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision 8fc257994d0ce2396196d7a06d50d20c8015f4b7)
1 /*-
2  * Copyright (c) 2001 The NetBSD Foundation, Inc.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to The NetBSD Foundation
6  * by Matt Thomas <matt@3am-software.com> of Allegro Networks, Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *        This product includes software developed by the NetBSD
19  *        Foundation, Inc. and its contributors.
20  * 4. Neither the name of The NetBSD Foundation nor the names of its
21  *    contributors may be used to endorse or promote products derived
22  *    from this software without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
25  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
26  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
27  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
28  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
29  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
30  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
31  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
32  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
33  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
34  * POSSIBILITY OF SUCH DAMAGE.
35  */
36 /*-
37  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
38  * Copyright (C) 1995, 1996 TooLs GmbH.
39  * All rights reserved.
40  *
41  * Redistribution and use in source and binary forms, with or without
42  * modification, are permitted provided that the following conditions
43  * are met:
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. All advertising materials mentioning features or use of this software
50  *    must display the following acknowledgement:
51  *	This product includes software developed by TooLs GmbH.
52  * 4. The name of TooLs GmbH may not be used to endorse or promote products
53  *    derived from this software without specific prior written permission.
54  *
55  * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
56  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
57  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
58  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
59  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
60  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
61  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
62  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
63  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
64  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
65  *
66  * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $
67  */
68 /*-
69  * Copyright (C) 2001 Benno Rice.
70  * All rights reserved.
71  *
72  * Redistribution and use in source and binary forms, with or without
73  * modification, are permitted provided that the following conditions
74  * are met:
75  * 1. Redistributions of source code must retain the above copyright
76  *    notice, this list of conditions and the following disclaimer.
77  * 2. Redistributions in binary form must reproduce the above copyright
78  *    notice, this list of conditions and the following disclaimer in the
79  *    documentation and/or other materials provided with the distribution.
80  *
81  * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR
82  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
83  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
84  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
85  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
86  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
87  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
88  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
89  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
90  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
91  */
92 
93 #include <sys/cdefs.h>
94 __FBSDID("$FreeBSD$");
95 
96 /*
97  * Manages physical address maps.
98  *
99  * In addition to hardware address maps, this module is called upon to
100  * provide software-use-only maps which may or may not be stored in the
101  * same form as hardware maps.  These pseudo-maps are used to store
102  * intermediate results from copy operations to and from address spaces.
103  *
104  * Since the information managed by this module is also stored by the
105  * logical address mapping module, this module may throw away valid virtual
106  * to physical mappings at almost any time.  However, invalidations of
107  * mappings must be done as requested.
108  *
109  * In order to cope with hardware architectures which make virtual to
110  * physical map invalidates expensive, this module may delay invalidate
111  * reduced protection operations until such time as they are actually
112  * necessary.  This module is given full information as to which processors
113  * are currently using which maps, and to when physical maps must be made
114  * correct.
115  */
116 
117 #include "opt_kstack_pages.h"
118 
119 #include <sys/param.h>
120 #include <sys/kernel.h>
121 #include <sys/ktr.h>
122 #include <sys/lock.h>
123 #include <sys/msgbuf.h>
124 #include <sys/mutex.h>
125 #include <sys/proc.h>
126 #include <sys/sysctl.h>
127 #include <sys/systm.h>
128 #include <sys/vmmeter.h>
129 
130 #include <sys/kdb.h>
131 
132 #include <dev/ofw/openfirm.h>
133 
134 #include <vm/vm.h>
135 #include <vm/vm_param.h>
136 #include <vm/vm_kern.h>
137 #include <vm/vm_page.h>
138 #include <vm/vm_map.h>
139 #include <vm/vm_object.h>
140 #include <vm/vm_extern.h>
141 #include <vm/vm_pageout.h>
142 #include <vm/vm_pager.h>
143 #include <vm/uma.h>
144 
145 #include <machine/cpu.h>
146 #include <machine/platform.h>
147 #include <machine/frame.h>
148 #include <machine/md_var.h>
149 #include <machine/psl.h>
150 #include <machine/bat.h>
151 #include <machine/pte.h>
152 #include <machine/sr.h>
153 #include <machine/trap.h>
154 #include <machine/mmuvar.h>
155 
156 #include "mmu_if.h"
157 
158 #define	MOEA_DEBUG
159 
160 #define TODO	panic("%s: not implemented", __func__);
161 
162 static __inline u_int32_t
163 cntlzw(volatile u_int32_t a) {
164 	u_int32_t b;
165 	__asm ("cntlzw %0, %1" : "=r"(b) : "r"(a));
166 	return b;
167 }
168 
169 static __inline uint64_t
170 va_to_vsid(pmap_t pm, vm_offset_t va)
171 {
172 	return ((pm->pm_sr[(uintptr_t)va >> ADDR_SR_SHFT]) & SR_VSID_MASK);
173 }
174 
175 #define	PTESYNC()	__asm __volatile("ptesync");
176 #define	TLBSYNC()	__asm __volatile("tlbsync; ptesync");
177 #define	SYNC()		__asm __volatile("sync");
178 #define	EIEIO()		__asm __volatile("eieio");
179 
180 /*
181  * The tlbie instruction must be executed in 64-bit mode
182  * so we have to twiddle MSR[SF] around every invocation.
183  * Just to add to the fun, exceptions must be off as well
184  * so that we can't trap in 64-bit mode. What a pain.
185  */
186 struct mtx	tlbie_mutex;
187 
188 static __inline void
189 TLBIE(pmap_t pmap, vm_offset_t va) {
190 	uint64_t vpn;
191 	register_t vpn_hi, vpn_lo;
192 	register_t msr;
193 	register_t scratch;
194 
195 	vpn = (uint64_t)(va & ADDR_PIDX);
196 	if (pmap != NULL)
197 		vpn |= (va_to_vsid(pmap,va) << 28);
198 	vpn &= ~(0xffffULL << 48);
199 
200 	vpn_hi = (uint32_t)(vpn >> 32);
201 	vpn_lo = (uint32_t)vpn;
202 
203 	mtx_lock_spin(&tlbie_mutex);
204 	__asm __volatile("\
205 	    mfmsr %0; \
206 	    mr %1, %0; \
207 	    insrdi %1,%5,1,0; \
208 	    mtmsrd %1; \
209 	    ptesync; \
210 	    \
211 	    sld %1,%2,%4; \
212 	    or %1,%1,%3; \
213 	    tlbie %1; \
214 	    \
215 	    mtmsrd %0; \
216 	    eieio; \
217 	    tlbsync; \
218 	    ptesync;"
219 	: "=r"(msr), "=r"(scratch) : "r"(vpn_hi), "r"(vpn_lo), "r"(32), "r"(1)
220 	    : "memory");
221 	mtx_unlock_spin(&tlbie_mutex);
222 }
223 
224 #define DISABLE_TRANS(msr)	msr = mfmsr(); mtmsr(msr & ~PSL_DR); isync()
225 #define ENABLE_TRANS(msr)	mtmsr(msr); isync()
226 
227 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
228 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
229 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
230 #define	VSID_HASH_MASK		0x0000007fffffffffULL
231 
232 #define	PVO_PTEGIDX_MASK	0x007UL		/* which PTEG slot */
233 #define	PVO_PTEGIDX_VALID	0x008UL		/* slot is valid */
234 #define	PVO_WIRED		0x010UL		/* PVO entry is wired */
235 #define	PVO_MANAGED		0x020UL		/* PVO entry is managed */
236 #define	PVO_BOOTSTRAP		0x080UL		/* PVO entry allocated during
237 						   bootstrap */
238 #define PVO_FAKE		0x100UL		/* fictitious phys page */
239 #define	PVO_VADDR(pvo)		((pvo)->pvo_vaddr & ~ADDR_POFF)
240 #define PVO_ISFAKE(pvo)		((pvo)->pvo_vaddr & PVO_FAKE)
241 #define	PVO_PTEGIDX_GET(pvo)	((pvo)->pvo_vaddr & PVO_PTEGIDX_MASK)
242 #define	PVO_PTEGIDX_ISSET(pvo)	((pvo)->pvo_vaddr & PVO_PTEGIDX_VALID)
243 #define	PVO_PTEGIDX_CLR(pvo)	\
244 	((void)((pvo)->pvo_vaddr &= ~(PVO_PTEGIDX_VALID|PVO_PTEGIDX_MASK)))
245 #define	PVO_PTEGIDX_SET(pvo, i)	\
246 	((void)((pvo)->pvo_vaddr |= (i)|PVO_PTEGIDX_VALID))
247 
248 #define	MOEA_PVO_CHECK(pvo)
249 
250 #define LOCK_TABLE() mtx_lock(&moea64_table_mutex)
251 #define UNLOCK_TABLE() mtx_unlock(&moea64_table_mutex);
252 #define ASSERT_TABLE_LOCK() mtx_assert(&moea64_table_mutex, MA_OWNED)
253 
254 struct ofw_map {
255 	vm_offset_t	om_va;
256 	vm_size_t	om_len;
257 	vm_offset_t	om_pa_hi;
258 	vm_offset_t	om_pa_lo;
259 	u_int		om_mode;
260 };
261 
262 /*
263  * Map of physical memory regions.
264  */
265 static struct	mem_region *regions;
266 static struct	mem_region *pregions;
267 extern u_int	phys_avail_count;
268 extern int	regions_sz, pregions_sz;
269 extern int	ofw_real_mode;
270 
271 extern struct pmap ofw_pmap;
272 
273 extern void bs_remap_earlyboot(void);
274 
275 
276 /*
277  * Lock for the pteg and pvo tables.
278  */
279 struct mtx	moea64_table_mutex;
280 
281 /*
282  * PTEG data.
283  */
284 static struct	lpteg *moea64_pteg_table;
285 u_int		moea64_pteg_count;
286 u_int		moea64_pteg_mask;
287 
288 /*
289  * PVO data.
290  */
291 struct	pvo_head *moea64_pvo_table;		/* pvo entries by pteg index */
292 /* lists of unmanaged pages */
293 struct	pvo_head moea64_pvo_kunmanaged =
294     LIST_HEAD_INITIALIZER(moea64_pvo_kunmanaged);
295 struct	pvo_head moea64_pvo_unmanaged =
296     LIST_HEAD_INITIALIZER(moea64_pvo_unmanaged);
297 
298 uma_zone_t	moea64_upvo_zone; /* zone for pvo entries for unmanaged pages */
299 uma_zone_t	moea64_mpvo_zone; /* zone for pvo entries for managed pages */
300 
301 #define	BPVO_POOL_SIZE	327680
302 static struct	pvo_entry *moea64_bpvo_pool;
303 static int	moea64_bpvo_pool_index = 0;
304 
305 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
306 static u_int	moea64_vsid_bitmap[NPMAPS / VSID_NBPW];
307 
308 static boolean_t moea64_initialized = FALSE;
309 
310 /*
311  * Statistics.
312  */
313 u_int	moea64_pte_valid = 0;
314 u_int	moea64_pte_overflow = 0;
315 u_int	moea64_pvo_entries = 0;
316 u_int	moea64_pvo_enter_calls = 0;
317 u_int	moea64_pvo_remove_calls = 0;
318 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_valid, CTLFLAG_RD,
319     &moea64_pte_valid, 0, "");
320 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_overflow, CTLFLAG_RD,
321     &moea64_pte_overflow, 0, "");
322 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_entries, CTLFLAG_RD,
323     &moea64_pvo_entries, 0, "");
324 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_enter_calls, CTLFLAG_RD,
325     &moea64_pvo_enter_calls, 0, "");
326 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_remove_calls, CTLFLAG_RD,
327     &moea64_pvo_remove_calls, 0, "");
328 
329 vm_offset_t	moea64_scratchpage_va[2];
330 struct	lpte 	*moea64_scratchpage_pte[2];
331 struct	mtx	moea64_scratchpage_mtx;
332 
333 /*
334  * Allocate physical memory for use in moea64_bootstrap.
335  */
336 static vm_offset_t	moea64_bootstrap_alloc(vm_size_t, u_int);
337 
338 /*
339  * PTE calls.
340  */
341 static int		moea64_pte_insert(u_int, struct lpte *);
342 
343 /*
344  * PVO calls.
345  */
346 static int	moea64_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *,
347 		    vm_offset_t, vm_offset_t, uint64_t, int);
348 static void	moea64_pvo_remove(struct pvo_entry *, int);
349 static struct	pvo_entry *moea64_pvo_find_va(pmap_t, vm_offset_t, int *);
350 static struct	lpte *moea64_pvo_to_pte(const struct pvo_entry *, int);
351 
352 /*
353  * Utility routines.
354  */
355 static void		moea64_bridge_bootstrap(mmu_t mmup,
356 			    vm_offset_t kernelstart, vm_offset_t kernelend);
357 static void		moea64_bridge_cpu_bootstrap(mmu_t, int ap);
358 static void		moea64_enter_locked(pmap_t, vm_offset_t, vm_page_t,
359 			    vm_prot_t, boolean_t);
360 static boolean_t	moea64_query_bit(vm_page_t, u_int64_t);
361 static u_int		moea64_clear_bit(vm_page_t, u_int64_t);
362 static void		moea64_kremove(mmu_t, vm_offset_t);
363 static void		moea64_syncicache(pmap_t pmap, vm_offset_t va,
364 			    vm_offset_t pa, vm_size_t sz);
365 static void		tlbia(void);
366 
367 /*
368  * Kernel MMU interface
369  */
370 void moea64_change_wiring(mmu_t, pmap_t, vm_offset_t, boolean_t);
371 void moea64_clear_modify(mmu_t, vm_page_t);
372 void moea64_clear_reference(mmu_t, vm_page_t);
373 void moea64_copy_page(mmu_t, vm_page_t, vm_page_t);
374 void moea64_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, boolean_t);
375 void moea64_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
376     vm_prot_t);
377 void moea64_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
378 vm_paddr_t moea64_extract(mmu_t, pmap_t, vm_offset_t);
379 vm_page_t moea64_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
380 void moea64_init(mmu_t);
381 boolean_t moea64_is_modified(mmu_t, vm_page_t);
382 boolean_t moea64_is_referenced(mmu_t, vm_page_t);
383 boolean_t moea64_ts_referenced(mmu_t, vm_page_t);
384 vm_offset_t moea64_map(mmu_t, vm_offset_t *, vm_offset_t, vm_offset_t, int);
385 boolean_t moea64_page_exists_quick(mmu_t, pmap_t, vm_page_t);
386 int moea64_page_wired_mappings(mmu_t, vm_page_t);
387 void moea64_pinit(mmu_t, pmap_t);
388 void moea64_pinit0(mmu_t, pmap_t);
389 void moea64_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
390 void moea64_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
391 void moea64_qremove(mmu_t, vm_offset_t, int);
392 void moea64_release(mmu_t, pmap_t);
393 void moea64_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
394 void moea64_remove_all(mmu_t, vm_page_t);
395 void moea64_remove_write(mmu_t, vm_page_t);
396 void moea64_zero_page(mmu_t, vm_page_t);
397 void moea64_zero_page_area(mmu_t, vm_page_t, int, int);
398 void moea64_zero_page_idle(mmu_t, vm_page_t);
399 void moea64_activate(mmu_t, struct thread *);
400 void moea64_deactivate(mmu_t, struct thread *);
401 void *moea64_mapdev(mmu_t, vm_offset_t, vm_size_t);
402 void moea64_unmapdev(mmu_t, vm_offset_t, vm_size_t);
403 vm_offset_t moea64_kextract(mmu_t, vm_offset_t);
404 void moea64_kenter(mmu_t, vm_offset_t, vm_offset_t);
405 boolean_t moea64_dev_direct_mapped(mmu_t, vm_offset_t, vm_size_t);
406 static void moea64_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
407 
408 static mmu_method_t moea64_bridge_methods[] = {
409 	MMUMETHOD(mmu_change_wiring,	moea64_change_wiring),
410 	MMUMETHOD(mmu_clear_modify,	moea64_clear_modify),
411 	MMUMETHOD(mmu_clear_reference,	moea64_clear_reference),
412 	MMUMETHOD(mmu_copy_page,	moea64_copy_page),
413 	MMUMETHOD(mmu_enter,		moea64_enter),
414 	MMUMETHOD(mmu_enter_object,	moea64_enter_object),
415 	MMUMETHOD(mmu_enter_quick,	moea64_enter_quick),
416 	MMUMETHOD(mmu_extract,		moea64_extract),
417 	MMUMETHOD(mmu_extract_and_hold,	moea64_extract_and_hold),
418 	MMUMETHOD(mmu_init,		moea64_init),
419 	MMUMETHOD(mmu_is_modified,	moea64_is_modified),
420 	MMUMETHOD(mmu_is_referenced,	moea64_is_referenced),
421 	MMUMETHOD(mmu_ts_referenced,	moea64_ts_referenced),
422 	MMUMETHOD(mmu_map,     		moea64_map),
423 	MMUMETHOD(mmu_page_exists_quick,moea64_page_exists_quick),
424 	MMUMETHOD(mmu_page_wired_mappings,moea64_page_wired_mappings),
425 	MMUMETHOD(mmu_pinit,		moea64_pinit),
426 	MMUMETHOD(mmu_pinit0,		moea64_pinit0),
427 	MMUMETHOD(mmu_protect,		moea64_protect),
428 	MMUMETHOD(mmu_qenter,		moea64_qenter),
429 	MMUMETHOD(mmu_qremove,		moea64_qremove),
430 	MMUMETHOD(mmu_release,		moea64_release),
431 	MMUMETHOD(mmu_remove,		moea64_remove),
432 	MMUMETHOD(mmu_remove_all,      	moea64_remove_all),
433 	MMUMETHOD(mmu_remove_write,	moea64_remove_write),
434 	MMUMETHOD(mmu_sync_icache,	moea64_sync_icache),
435 	MMUMETHOD(mmu_zero_page,       	moea64_zero_page),
436 	MMUMETHOD(mmu_zero_page_area,	moea64_zero_page_area),
437 	MMUMETHOD(mmu_zero_page_idle,	moea64_zero_page_idle),
438 	MMUMETHOD(mmu_activate,		moea64_activate),
439 	MMUMETHOD(mmu_deactivate,      	moea64_deactivate),
440 
441 	/* Internal interfaces */
442 	MMUMETHOD(mmu_bootstrap,       	moea64_bridge_bootstrap),
443 	MMUMETHOD(mmu_cpu_bootstrap,   	moea64_bridge_cpu_bootstrap),
444 	MMUMETHOD(mmu_mapdev,		moea64_mapdev),
445 	MMUMETHOD(mmu_unmapdev,		moea64_unmapdev),
446 	MMUMETHOD(mmu_kextract,		moea64_kextract),
447 	MMUMETHOD(mmu_kenter,		moea64_kenter),
448 	MMUMETHOD(mmu_dev_direct_mapped,moea64_dev_direct_mapped),
449 
450 	{ 0, 0 }
451 };
452 
453 static mmu_def_t oea64_bridge_mmu = {
454 	MMU_TYPE_G5,
455 	moea64_bridge_methods,
456 	0
457 };
458 MMU_DEF(oea64_bridge_mmu);
459 
460 static __inline u_int
461 va_to_pteg(uint64_t vsid, vm_offset_t addr)
462 {
463 	uint64_t hash;
464 
465 	hash = (vsid & VSID_HASH_MASK) ^ (((uint64_t)addr & ADDR_PIDX) >>
466 	    ADDR_PIDX_SHFT);
467 	return (hash & moea64_pteg_mask);
468 }
469 
470 static __inline struct pvo_head *
471 pa_to_pvoh(vm_offset_t pa, vm_page_t *pg_p)
472 {
473 	struct	vm_page *pg;
474 
475 	pg = PHYS_TO_VM_PAGE(pa);
476 
477 	if (pg_p != NULL)
478 		*pg_p = pg;
479 
480 	if (pg == NULL)
481 		return (&moea64_pvo_unmanaged);
482 
483 	return (&pg->md.mdpg_pvoh);
484 }
485 
486 static __inline struct pvo_head *
487 vm_page_to_pvoh(vm_page_t m)
488 {
489 
490 	return (&m->md.mdpg_pvoh);
491 }
492 
493 static __inline void
494 moea64_attr_clear(vm_page_t m, u_int64_t ptebit)
495 {
496 
497 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
498 	m->md.mdpg_attrs &= ~ptebit;
499 }
500 
501 static __inline u_int64_t
502 moea64_attr_fetch(vm_page_t m)
503 {
504 
505 	return (m->md.mdpg_attrs);
506 }
507 
508 static __inline void
509 moea64_attr_save(vm_page_t m, u_int64_t ptebit)
510 {
511 
512 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
513 	m->md.mdpg_attrs |= ptebit;
514 }
515 
516 static __inline void
517 moea64_pte_create(struct lpte *pt, uint64_t vsid, vm_offset_t va,
518     uint64_t pte_lo)
519 {
520 	ASSERT_TABLE_LOCK();
521 
522 	/*
523 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
524 	 * set when the real pte is set in memory.
525 	 *
526 	 * Note: Don't set the valid bit for correct operation of tlb update.
527 	 */
528 	pt->pte_hi = (vsid << LPTE_VSID_SHIFT) |
529 	    (((uint64_t)(va & ADDR_PIDX) >> ADDR_API_SHFT64) & LPTE_API);
530 
531 	pt->pte_lo = pte_lo;
532 }
533 
534 static __inline void
535 moea64_pte_synch(struct lpte *pt, struct lpte *pvo_pt)
536 {
537 
538 	ASSERT_TABLE_LOCK();
539 
540 	pvo_pt->pte_lo |= pt->pte_lo & (LPTE_REF | LPTE_CHG);
541 }
542 
543 static __inline void
544 moea64_pte_clear(struct lpte *pt, pmap_t pmap, vm_offset_t va, u_int64_t ptebit)
545 {
546 	ASSERT_TABLE_LOCK();
547 
548 	/*
549 	 * As shown in Section 7.6.3.2.3
550 	 */
551 	pt->pte_lo &= ~ptebit;
552 	TLBIE(pmap,va);
553 }
554 
555 static __inline void
556 moea64_pte_set(struct lpte *pt, struct lpte *pvo_pt)
557 {
558 
559 	ASSERT_TABLE_LOCK();
560 	pvo_pt->pte_hi |= LPTE_VALID;
561 
562 	/*
563 	 * Update the PTE as defined in section 7.6.3.1.
564 	 * Note that the REF/CHG bits are from pvo_pt and thus should have
565 	 * been saved so this routine can restore them (if desired).
566 	 */
567 	pt->pte_lo = pvo_pt->pte_lo;
568 	EIEIO();
569 	pt->pte_hi = pvo_pt->pte_hi;
570 	PTESYNC();
571 	moea64_pte_valid++;
572 }
573 
574 static __inline void
575 moea64_pte_unset(struct lpte *pt, struct lpte *pvo_pt, pmap_t pmap, vm_offset_t va)
576 {
577 	ASSERT_TABLE_LOCK();
578 	pvo_pt->pte_hi &= ~LPTE_VALID;
579 
580 	/*
581 	 * Force the reg & chg bits back into the PTEs.
582 	 */
583 	SYNC();
584 
585 	/*
586 	 * Invalidate the pte.
587 	 */
588 	pt->pte_hi &= ~LPTE_VALID;
589 	TLBIE(pmap,va);
590 
591 	/*
592 	 * Save the reg & chg bits.
593 	 */
594 	moea64_pte_synch(pt, pvo_pt);
595 	moea64_pte_valid--;
596 }
597 
598 static __inline void
599 moea64_pte_change(struct lpte *pt, struct lpte *pvo_pt, pmap_t pmap, vm_offset_t va)
600 {
601 
602 	/*
603 	 * Invalidate the PTE
604 	 */
605 	moea64_pte_unset(pt, pvo_pt, pmap, va);
606 	moea64_pte_set(pt, pvo_pt);
607 	if (pmap == kernel_pmap)
608 		isync();
609 }
610 
611 static __inline uint64_t
612 moea64_calc_wimg(vm_offset_t pa)
613 {
614 	uint64_t pte_lo;
615 	int i;
616 
617 	/*
618 	 * Assume the page is cache inhibited and access is guarded unless
619 	 * it's in our available memory array.
620 	 */
621 	pte_lo = LPTE_I | LPTE_G;
622 	for (i = 0; i < pregions_sz; i++) {
623 		if ((pa >= pregions[i].mr_start) &&
624 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
625 			pte_lo &= ~(LPTE_I | LPTE_G);
626 			pte_lo |= LPTE_M;
627 			break;
628 		}
629 	}
630 
631 	return pte_lo;
632 }
633 
634 /*
635  * Quick sort callout for comparing memory regions.
636  */
637 static int	mr_cmp(const void *a, const void *b);
638 static int	om_cmp(const void *a, const void *b);
639 
640 static int
641 mr_cmp(const void *a, const void *b)
642 {
643 	const struct	mem_region *regiona;
644 	const struct	mem_region *regionb;
645 
646 	regiona = a;
647 	regionb = b;
648 	if (regiona->mr_start < regionb->mr_start)
649 		return (-1);
650 	else if (regiona->mr_start > regionb->mr_start)
651 		return (1);
652 	else
653 		return (0);
654 }
655 
656 static int
657 om_cmp(const void *a, const void *b)
658 {
659 	const struct	ofw_map *mapa;
660 	const struct	ofw_map *mapb;
661 
662 	mapa = a;
663 	mapb = b;
664 	if (mapa->om_pa_hi < mapb->om_pa_hi)
665 		return (-1);
666 	else if (mapa->om_pa_hi > mapb->om_pa_hi)
667 		return (1);
668 	else if (mapa->om_pa_lo < mapb->om_pa_lo)
669 		return (-1);
670 	else if (mapa->om_pa_lo > mapb->om_pa_lo)
671 		return (1);
672 	else
673 		return (0);
674 }
675 
676 static void
677 moea64_bridge_cpu_bootstrap(mmu_t mmup, int ap)
678 {
679 	int i = 0;
680 
681 	/*
682 	 * Initialize segment registers and MMU
683 	 */
684 
685 	mtmsr(mfmsr() & ~PSL_DR & ~PSL_IR); isync();
686 	for (i = 0; i < 16; i++) {
687 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
688 	}
689 	__asm __volatile ("ptesync; mtsdr1 %0; isync"
690 	    :: "r"((u_int)moea64_pteg_table
691 		     | (32 - cntlzw(moea64_pteg_mask >> 11))));
692 	tlbia();
693 }
694 
695 static void
696 moea64_add_ofw_mappings(mmu_t mmup, phandle_t mmu, size_t sz)
697 {
698 	struct ofw_map	translations[sz/sizeof(struct ofw_map)];
699 	register_t	msr;
700 	vm_offset_t	off;
701 	vm_paddr_t	pa_base;
702 	int		i, ofw_mappings;
703 
704 	bzero(translations, sz);
705 	if (OF_getprop(mmu, "translations", translations, sz) == -1)
706 		panic("moea64_bootstrap: can't get ofw translations");
707 
708 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
709 	sz /= sizeof(*translations);
710 	qsort(translations, sz, sizeof (*translations), om_cmp);
711 
712 	for (i = 0, ofw_mappings = 0; i < sz; i++) {
713 		CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
714 		    (uint32_t)(translations[i].om_pa_lo), translations[i].om_va,
715 		    translations[i].om_len);
716 
717 		if (translations[i].om_pa_lo % PAGE_SIZE)
718 			panic("OFW translation not page-aligned!");
719 
720 		if (translations[i].om_pa_hi)
721 			panic("OFW translations above 32-bit boundary!");
722 
723 		pa_base = translations[i].om_pa_lo;
724 
725 		/* Now enter the pages for this mapping */
726 
727 		DISABLE_TRANS(msr);
728 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
729 			moea64_kenter(mmup, translations[i].om_va + off,
730 			    pa_base + off);
731 
732 			ofw_mappings++;
733 		}
734 		ENABLE_TRANS(msr);
735 	}
736 }
737 
738 static void
739 moea64_bridge_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
740 {
741 	ihandle_t	mmui;
742 	phandle_t	chosen;
743 	phandle_t	mmu;
744 	size_t		sz;
745 	int		i, j;
746 	vm_size_t	size, physsz, hwphyssz;
747 	vm_offset_t	pa, va, off;
748 	register_t	msr;
749 	void		*dpcpu;
750 
751 	/* We don't have a direct map since there is no BAT */
752 	hw_direct_map = 0;
753 
754 	/* Make sure battable is zero, since we have no BAT */
755 	for (i = 0; i < 16; i++) {
756 		battable[i].batu = 0;
757 		battable[i].batl = 0;
758 	}
759 
760 	/* Get physical memory regions from firmware */
761 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
762 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
763 
764 	qsort(pregions, pregions_sz, sizeof(*pregions), mr_cmp);
765 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
766 		panic("moea64_bootstrap: phys_avail too small");
767 	qsort(regions, regions_sz, sizeof(*regions), mr_cmp);
768 	phys_avail_count = 0;
769 	physsz = 0;
770 	hwphyssz = 0;
771 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
772 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
773 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
774 		    regions[i].mr_start + regions[i].mr_size,
775 		    regions[i].mr_size);
776 		if (hwphyssz != 0 &&
777 		    (physsz + regions[i].mr_size) >= hwphyssz) {
778 			if (physsz < hwphyssz) {
779 				phys_avail[j] = regions[i].mr_start;
780 				phys_avail[j + 1] = regions[i].mr_start +
781 				    hwphyssz - physsz;
782 				physsz = hwphyssz;
783 				phys_avail_count++;
784 			}
785 			break;
786 		}
787 		phys_avail[j] = regions[i].mr_start;
788 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
789 		phys_avail_count++;
790 		physsz += regions[i].mr_size;
791 	}
792 	physmem = btoc(physsz);
793 
794 	/*
795 	 * Allocate PTEG table.
796 	 */
797 #ifdef PTEGCOUNT
798 	moea64_pteg_count = PTEGCOUNT;
799 #else
800 	moea64_pteg_count = 0x1000;
801 
802 	while (moea64_pteg_count < physmem)
803 		moea64_pteg_count <<= 1;
804 #endif /* PTEGCOUNT */
805 
806 	size = moea64_pteg_count * sizeof(struct lpteg);
807 	CTR2(KTR_PMAP, "moea64_bootstrap: %d PTEGs, %d bytes",
808 	    moea64_pteg_count, size);
809 
810 	/*
811 	 * We now need to allocate memory. This memory, to be allocated,
812 	 * has to reside in a page table. The page table we are about to
813 	 * allocate. We don't have BAT. So drop to data real mode for a minute
814 	 * as a measure of last resort. We do this a couple times.
815 	 */
816 
817 	moea64_pteg_table = (struct lpteg *)moea64_bootstrap_alloc(size, size);
818 	DISABLE_TRANS(msr);
819 	bzero((void *)moea64_pteg_table, moea64_pteg_count * sizeof(struct lpteg));
820 	ENABLE_TRANS(msr);
821 
822 	moea64_pteg_mask = moea64_pteg_count - 1;
823 
824 	CTR1(KTR_PMAP, "moea64_bootstrap: PTEG table at %p", moea64_pteg_table);
825 
826 	/*
827 	 * Allocate pv/overflow lists.
828 	 */
829 	size = sizeof(struct pvo_head) * moea64_pteg_count;
830 
831 	moea64_pvo_table = (struct pvo_head *)moea64_bootstrap_alloc(size,
832 	    PAGE_SIZE);
833 	CTR1(KTR_PMAP, "moea64_bootstrap: PVO table at %p", moea64_pvo_table);
834 
835 	DISABLE_TRANS(msr);
836 	for (i = 0; i < moea64_pteg_count; i++)
837 		LIST_INIT(&moea64_pvo_table[i]);
838 	ENABLE_TRANS(msr);
839 
840 	/*
841 	 * Initialize the lock that synchronizes access to the pteg and pvo
842 	 * tables.
843 	 */
844 	mtx_init(&moea64_table_mutex, "pmap table", NULL, MTX_DEF |
845 	    MTX_RECURSE);
846 
847 	/*
848 	 * Initialize the TLBIE lock. TLBIE can only be executed by one CPU.
849 	 */
850 	mtx_init(&tlbie_mutex, "tlbie mutex", NULL, MTX_SPIN);
851 
852 	/*
853 	 * Initialise the unmanaged pvo pool.
854 	 */
855 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
856 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
857 	moea64_bpvo_pool_index = 0;
858 
859 	/*
860 	 * Make sure kernel vsid is allocated as well as VSID 0.
861 	 */
862 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
863 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
864 	moea64_vsid_bitmap[0] |= 1;
865 
866 	/*
867 	 * Initialize the kernel pmap (which is statically allocated).
868 	 */
869 	for (i = 0; i < 16; i++)
870 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
871 
872 	kernel_pmap->pmap_phys = kernel_pmap;
873 	kernel_pmap->pm_active = ~0;
874 
875 	PMAP_LOCK_INIT(kernel_pmap);
876 
877 	/*
878 	 * Now map in all the other buffers we allocated earlier
879 	 */
880 
881 	DISABLE_TRANS(msr);
882 	size = moea64_pteg_count * sizeof(struct lpteg);
883 	off = (vm_offset_t)(moea64_pteg_table);
884 	for (pa = off; pa < off + size; pa += PAGE_SIZE)
885 		moea64_kenter(mmup, pa, pa);
886 	size = sizeof(struct pvo_head) * moea64_pteg_count;
887 	off = (vm_offset_t)(moea64_pvo_table);
888 	for (pa = off; pa < off + size; pa += PAGE_SIZE)
889 		moea64_kenter(mmup, pa, pa);
890 	size = BPVO_POOL_SIZE*sizeof(struct pvo_entry);
891 	off = (vm_offset_t)(moea64_bpvo_pool);
892 	for (pa = off; pa < off + size; pa += PAGE_SIZE)
893 		moea64_kenter(mmup, pa, pa);
894 
895 	/*
896 	 * Map certain important things, like ourselves.
897 	 *
898 	 * NOTE: We do not map the exception vector space. That code is
899 	 * used only in real mode, and leaving it unmapped allows us to
900 	 * catch NULL pointer deferences, instead of making NULL a valid
901 	 * address.
902 	 */
903 
904 	for (pa = kernelstart & ~PAGE_MASK; pa < kernelend; pa += PAGE_SIZE)
905 		moea64_kenter(mmup, pa, pa);
906 	ENABLE_TRANS(msr);
907 
908 	if (!ofw_real_mode) {
909 	    /*
910 	     * Set up the Open Firmware pmap and add its mappings.
911 	     */
912 
913 	    moea64_pinit(mmup, &ofw_pmap);
914 	    for (i = 0; i < 16; i++)
915 		ofw_pmap.pm_sr[i] = kernel_pmap->pm_sr[i];
916 
917 	    if ((chosen = OF_finddevice("/chosen")) == -1)
918 		panic("moea64_bootstrap: can't find /chosen");
919 	    OF_getprop(chosen, "mmu", &mmui, 4);
920 	    if ((mmu = OF_instance_to_package(mmui)) == -1)
921 		panic("moea64_bootstrap: can't get mmu package");
922 	    if ((sz = OF_getproplen(mmu, "translations")) == -1)
923 		panic("moea64_bootstrap: can't get ofw translation count");
924 	    if (sz > 6144 /* tmpstksz - 2 KB headroom */)
925 		panic("moea64_bootstrap: too many ofw translations");
926 
927 	    moea64_add_ofw_mappings(mmup, mmu, sz);
928 	}
929 
930 #ifdef SMP
931 	TLBSYNC();
932 #endif
933 
934 	/*
935 	 * Calculate the last available physical address.
936 	 */
937 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
938 		;
939 	Maxmem = powerpc_btop(phys_avail[i + 1]);
940 
941 	/*
942 	 * Initialize MMU and remap early physical mappings
943 	 */
944 	moea64_bridge_cpu_bootstrap(mmup,0);
945 	mtmsr(mfmsr() | PSL_DR | PSL_IR); isync();
946 	pmap_bootstrapped++;
947 	bs_remap_earlyboot();
948 
949 	/*
950 	 * Set the start and end of kva.
951 	 */
952 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
953 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
954 
955 	/*
956 	 * Figure out how far we can extend virtual_end into segment 16
957 	 * without running into existing mappings. Segment 16 is guaranteed
958 	 * to contain neither RAM nor devices (at least on Apple hardware),
959 	 * but will generally contain some OFW mappings we should not
960 	 * step on.
961 	 */
962 
963 	PMAP_LOCK(kernel_pmap);
964 	while (moea64_pvo_find_va(kernel_pmap, virtual_end+1, NULL) == NULL)
965 		virtual_end += PAGE_SIZE;
966 	PMAP_UNLOCK(kernel_pmap);
967 
968 	/*
969 	 * Allocate some things for page zeroing. We put this directly
970 	 * in the page table, marked with LPTE_LOCKED, to avoid any
971 	 * of the PVO book-keeping or other parts of the VM system
972 	 * from even knowing that this hack exists.
973 	 */
974 
975 	mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL, MTX_DEF);
976 	for (i = 0; i < 2; i++) {
977 		struct lpte pt;
978 		uint64_t vsid;
979 		int pteidx, ptegidx;
980 
981 		moea64_scratchpage_va[i] = (virtual_end+1) - PAGE_SIZE;
982 		virtual_end -= PAGE_SIZE;
983 
984 		LOCK_TABLE();
985 
986 		vsid = va_to_vsid(kernel_pmap, moea64_scratchpage_va[i]);
987 		moea64_pte_create(&pt, vsid, moea64_scratchpage_va[i],
988 		    LPTE_NOEXEC);
989 		pt.pte_hi |= LPTE_LOCKED;
990 
991 		ptegidx = va_to_pteg(vsid, moea64_scratchpage_va[i]);
992 		pteidx = moea64_pte_insert(ptegidx, &pt);
993 		if (pt.pte_hi & LPTE_HID)
994 			ptegidx ^= moea64_pteg_mask;
995 
996 		moea64_scratchpage_pte[i] =
997 		    &moea64_pteg_table[ptegidx].pt[pteidx];
998 
999 		UNLOCK_TABLE();
1000 	}
1001 
1002 	/*
1003 	 * Allocate a kernel stack with a guard page for thread0 and map it
1004 	 * into the kernel page map.
1005 	 */
1006 	pa = moea64_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, PAGE_SIZE);
1007 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
1008 	virtual_avail = va + KSTACK_PAGES * PAGE_SIZE;
1009 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
1010 	thread0.td_kstack = va;
1011 	thread0.td_kstack_pages = KSTACK_PAGES;
1012 	for (i = 0; i < KSTACK_PAGES; i++) {
1013 		moea64_kenter(mmup, va, pa);
1014 		pa += PAGE_SIZE;
1015 		va += PAGE_SIZE;
1016 	}
1017 
1018 	/*
1019 	 * Allocate virtual address space for the message buffer.
1020 	 */
1021 	pa = msgbuf_phys = moea64_bootstrap_alloc(MSGBUF_SIZE, PAGE_SIZE);
1022 	msgbufp = (struct msgbuf *)virtual_avail;
1023 	va = virtual_avail;
1024 	virtual_avail += round_page(MSGBUF_SIZE);
1025 	while (va < virtual_avail) {
1026 		moea64_kenter(mmup, va, pa);
1027 		pa += PAGE_SIZE;
1028 		va += PAGE_SIZE;
1029 	}
1030 
1031 	/*
1032 	 * Allocate virtual address space for the dynamic percpu area.
1033 	 */
1034 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
1035 	dpcpu = (void *)virtual_avail;
1036 	virtual_avail += DPCPU_SIZE;
1037 	while (va < virtual_avail) {
1038 		moea64_kenter(mmup, va, pa);
1039 		pa += PAGE_SIZE;
1040 		va += PAGE_SIZE;
1041 	}
1042 	dpcpu_init(dpcpu, 0);
1043 }
1044 
1045 /*
1046  * Activate a user pmap.  The pmap must be activated before it's address
1047  * space can be accessed in any way.
1048  */
1049 void
1050 moea64_activate(mmu_t mmu, struct thread *td)
1051 {
1052 	pmap_t	pm, pmr;
1053 
1054 	/*
1055 	 * Load all the data we need up front to encourage the compiler to
1056 	 * not issue any loads while we have interrupts disabled below.
1057 	 */
1058 	pm = &td->td_proc->p_vmspace->vm_pmap;
1059 	pmr = pm->pmap_phys;
1060 
1061 	pm->pm_active |= PCPU_GET(cpumask);
1062 	PCPU_SET(curpmap, pmr);
1063 }
1064 
1065 void
1066 moea64_deactivate(mmu_t mmu, struct thread *td)
1067 {
1068 	pmap_t	pm;
1069 
1070 	pm = &td->td_proc->p_vmspace->vm_pmap;
1071 	pm->pm_active &= ~(PCPU_GET(cpumask));
1072 	PCPU_SET(curpmap, NULL);
1073 }
1074 
1075 void
1076 moea64_change_wiring(mmu_t mmu, pmap_t pm, vm_offset_t va, boolean_t wired)
1077 {
1078 	struct	pvo_entry *pvo;
1079 
1080 	PMAP_LOCK(pm);
1081 	pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1082 
1083 	if (pvo != NULL) {
1084 		if (wired) {
1085 			if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1086 				pm->pm_stats.wired_count++;
1087 			pvo->pvo_vaddr |= PVO_WIRED;
1088 		} else {
1089 			if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1090 				pm->pm_stats.wired_count--;
1091 			pvo->pvo_vaddr &= ~PVO_WIRED;
1092 		}
1093 	}
1094 	PMAP_UNLOCK(pm);
1095 }
1096 
1097 /*
1098  * This goes through and sets the physical address of our
1099  * special scratch PTE to the PA we want to zero or copy. Because
1100  * of locking issues (this can get called in pvo_enter() by
1101  * the UMA allocator), we can't use most other utility functions here
1102  */
1103 
1104 static __inline
1105 void moea64_set_scratchpage_pa(int which, vm_offset_t pa) {
1106 
1107 	mtx_assert(&moea64_scratchpage_mtx, MA_OWNED);
1108 
1109 	moea64_scratchpage_pte[which]->pte_hi &= ~LPTE_VALID;
1110 	TLBIE(kernel_pmap, moea64_scratchpage_va[which]);
1111 
1112 	moea64_scratchpage_pte[which]->pte_lo &=
1113 	    ~(LPTE_WIMG | LPTE_RPGN);
1114 	moea64_scratchpage_pte[which]->pte_lo |=
1115 	    moea64_calc_wimg(pa) | (uint64_t)pa;
1116 	EIEIO();
1117 
1118 	moea64_scratchpage_pte[which]->pte_hi |= LPTE_VALID;
1119 	PTESYNC(); isync();
1120 }
1121 
1122 void
1123 moea64_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1124 {
1125 	vm_offset_t	dst;
1126 	vm_offset_t	src;
1127 
1128 	dst = VM_PAGE_TO_PHYS(mdst);
1129 	src = VM_PAGE_TO_PHYS(msrc);
1130 
1131 	mtx_lock(&moea64_scratchpage_mtx);
1132 
1133 	moea64_set_scratchpage_pa(0,src);
1134 	moea64_set_scratchpage_pa(1,dst);
1135 
1136 	kcopy((void *)moea64_scratchpage_va[0],
1137 	    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1138 
1139 	mtx_unlock(&moea64_scratchpage_mtx);
1140 }
1141 
1142 void
1143 moea64_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1144 {
1145 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1146 
1147 	if (!moea64_initialized)
1148 		panic("moea64_zero_page: can't zero pa %#x", pa);
1149 	if (size + off > PAGE_SIZE)
1150 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1151 
1152 	mtx_lock(&moea64_scratchpage_mtx);
1153 
1154 	moea64_set_scratchpage_pa(0,pa);
1155 	bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1156 	mtx_unlock(&moea64_scratchpage_mtx);
1157 }
1158 
1159 /*
1160  * Zero a page of physical memory by temporarily mapping it
1161  */
1162 void
1163 moea64_zero_page(mmu_t mmu, vm_page_t m)
1164 {
1165 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1166 	vm_offset_t off;
1167 
1168 	if (!moea64_initialized)
1169 		panic("moea64_zero_page: can't zero pa %#x", pa);
1170 
1171 	mtx_lock(&moea64_scratchpage_mtx);
1172 
1173 	moea64_set_scratchpage_pa(0,pa);
1174 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1175 		__asm __volatile("dcbz 0,%0" ::
1176 		    "r"(moea64_scratchpage_va[0] + off));
1177 	mtx_unlock(&moea64_scratchpage_mtx);
1178 }
1179 
1180 void
1181 moea64_zero_page_idle(mmu_t mmu, vm_page_t m)
1182 {
1183 
1184 	moea64_zero_page(mmu, m);
1185 }
1186 
1187 /*
1188  * Map the given physical page at the specified virtual address in the
1189  * target pmap with the protection requested.  If specified the page
1190  * will be wired down.
1191  */
1192 void
1193 moea64_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m,
1194     vm_prot_t prot, boolean_t wired)
1195 {
1196 
1197 	vm_page_lock_queues();
1198 	PMAP_LOCK(pmap);
1199 	moea64_enter_locked(pmap, va, m, prot, wired);
1200 	vm_page_unlock_queues();
1201 	PMAP_UNLOCK(pmap);
1202 }
1203 
1204 /*
1205  * Map the given physical page at the specified virtual address in the
1206  * target pmap with the protection requested.  If specified the page
1207  * will be wired down.
1208  *
1209  * The page queues and pmap must be locked.
1210  */
1211 
1212 static void
1213 moea64_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1214     boolean_t wired)
1215 {
1216 	struct		pvo_head *pvo_head;
1217 	uma_zone_t	zone;
1218 	vm_page_t	pg;
1219 	uint64_t	pte_lo;
1220 	u_int		pvo_flags;
1221 	int		error;
1222 
1223 	if (!moea64_initialized) {
1224 		pvo_head = &moea64_pvo_kunmanaged;
1225 		pg = NULL;
1226 		zone = moea64_upvo_zone;
1227 		pvo_flags = 0;
1228 	} else {
1229 		pvo_head = vm_page_to_pvoh(m);
1230 		pg = m;
1231 		zone = moea64_mpvo_zone;
1232 		pvo_flags = PVO_MANAGED;
1233 	}
1234 
1235 	if (pmap_bootstrapped)
1236 		mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1237 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1238 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0 ||
1239 	    (m->oflags & VPO_BUSY) != 0 || VM_OBJECT_LOCKED(m->object),
1240 	    ("moea64_enter_locked: page %p is not busy", m));
1241 
1242 	/* XXX change the pvo head for fake pages */
1243 	if ((m->flags & PG_FICTITIOUS) == PG_FICTITIOUS) {
1244 		pvo_flags &= ~PVO_MANAGED;
1245 		pvo_head = &moea64_pvo_kunmanaged;
1246 		zone = moea64_upvo_zone;
1247 	}
1248 
1249 	pte_lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m));
1250 
1251 	if (prot & VM_PROT_WRITE) {
1252 		pte_lo |= LPTE_BW;
1253 		if (pmap_bootstrapped &&
1254 		    (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0)
1255 			vm_page_flag_set(m, PG_WRITEABLE);
1256 	} else
1257 		pte_lo |= LPTE_BR;
1258 
1259 	if (prot & VM_PROT_EXECUTE)
1260 		pvo_flags |= VM_PROT_EXECUTE;
1261 
1262 	if (wired)
1263 		pvo_flags |= PVO_WIRED;
1264 
1265 	if ((m->flags & PG_FICTITIOUS) != 0)
1266 		pvo_flags |= PVO_FAKE;
1267 
1268 	error = moea64_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1269 	    pte_lo, pvo_flags);
1270 
1271 	/*
1272 	 * Flush the page from the instruction cache if this page is
1273 	 * mapped executable and cacheable.
1274 	 */
1275 	if ((pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1276 		moea64_syncicache(pmap, va, VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1277 	}
1278 }
1279 
1280 static void
1281 moea64_syncicache(pmap_t pmap, vm_offset_t va, vm_offset_t pa, vm_size_t sz)
1282 {
1283 
1284 	/*
1285 	 * This is much trickier than on older systems because
1286 	 * we can't sync the icache on physical addresses directly
1287 	 * without a direct map. Instead we check a couple of cases
1288 	 * where the memory is already mapped in and, failing that,
1289 	 * use the same trick we use for page zeroing to create
1290 	 * a temporary mapping for this physical address.
1291 	 */
1292 
1293 	if (!pmap_bootstrapped) {
1294 		/*
1295 		 * If PMAP is not bootstrapped, we are likely to be
1296 		 * in real mode.
1297 		 */
1298 		__syncicache((void *)pa, sz);
1299 	} else if (pmap == kernel_pmap) {
1300 		__syncicache((void *)va, sz);
1301 	} else {
1302 		/* Use the scratch page to set up a temp mapping */
1303 
1304 		mtx_lock(&moea64_scratchpage_mtx);
1305 
1306 		moea64_set_scratchpage_pa(1,pa & ~ADDR_POFF);
1307 		__syncicache((void *)(moea64_scratchpage_va[1] +
1308 		    (va & ADDR_POFF)), sz);
1309 
1310 		mtx_unlock(&moea64_scratchpage_mtx);
1311 	}
1312 }
1313 
1314 /*
1315  * Maps a sequence of resident pages belonging to the same object.
1316  * The sequence begins with the given page m_start.  This page is
1317  * mapped at the given virtual address start.  Each subsequent page is
1318  * mapped at a virtual address that is offset from start by the same
1319  * amount as the page is offset from m_start within the object.  The
1320  * last page in the sequence is the page with the largest offset from
1321  * m_start that can be mapped at a virtual address less than the given
1322  * virtual address end.  Not every virtual page between start and end
1323  * is mapped; only those for which a resident page exists with the
1324  * corresponding offset from m_start are mapped.
1325  */
1326 void
1327 moea64_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1328     vm_page_t m_start, vm_prot_t prot)
1329 {
1330 	vm_page_t m;
1331 	vm_pindex_t diff, psize;
1332 
1333 	psize = atop(end - start);
1334 	m = m_start;
1335 	vm_page_lock_queues();
1336 	PMAP_LOCK(pm);
1337 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1338 		moea64_enter_locked(pm, start + ptoa(diff), m, prot &
1339 		    (VM_PROT_READ | VM_PROT_EXECUTE), FALSE);
1340 		m = TAILQ_NEXT(m, listq);
1341 	}
1342 	vm_page_unlock_queues();
1343 	PMAP_UNLOCK(pm);
1344 }
1345 
1346 void
1347 moea64_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1348     vm_prot_t prot)
1349 {
1350 
1351 	vm_page_lock_queues();
1352 	PMAP_LOCK(pm);
1353 	moea64_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1354 	    FALSE);
1355 	vm_page_unlock_queues();
1356 	PMAP_UNLOCK(pm);
1357 }
1358 
1359 vm_paddr_t
1360 moea64_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1361 {
1362 	struct	pvo_entry *pvo;
1363 	vm_paddr_t pa;
1364 
1365 	PMAP_LOCK(pm);
1366 	pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1367 	if (pvo == NULL)
1368 		pa = 0;
1369 	else
1370 		pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) | (va & ADDR_POFF);
1371 	PMAP_UNLOCK(pm);
1372 	return (pa);
1373 }
1374 
1375 /*
1376  * Atomically extract and hold the physical page with the given
1377  * pmap and virtual address pair if that mapping permits the given
1378  * protection.
1379  */
1380 vm_page_t
1381 moea64_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1382 {
1383 	struct	pvo_entry *pvo;
1384 	vm_page_t m;
1385         vm_paddr_t pa;
1386 
1387 	m = NULL;
1388 	pa = 0;
1389 	PMAP_LOCK(pmap);
1390 retry:
1391 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1392 	if (pvo != NULL && (pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) &&
1393 	    ((pvo->pvo_pte.lpte.pte_lo & LPTE_PP) == LPTE_RW ||
1394 	     (prot & VM_PROT_WRITE) == 0)) {
1395 		if (vm_page_pa_tryrelock(pmap,
1396 			pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN, &pa))
1397 			goto retry;
1398 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN);
1399 		vm_page_hold(m);
1400 	}
1401 	PA_UNLOCK_COND(pa);
1402 	PMAP_UNLOCK(pmap);
1403 	return (m);
1404 }
1405 
1406 static void *
1407 moea64_uma_page_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
1408 {
1409 	/*
1410 	 * This entire routine is a horrible hack to avoid bothering kmem
1411 	 * for new KVA addresses. Because this can get called from inside
1412 	 * kmem allocation routines, calling kmem for a new address here
1413 	 * can lead to multiply locking non-recursive mutexes.
1414 	 */
1415 	static vm_pindex_t color;
1416         vm_offset_t va;
1417 
1418         vm_page_t m;
1419         int pflags, needed_lock;
1420 
1421 	*flags = UMA_SLAB_PRIV;
1422 	needed_lock = !PMAP_LOCKED(kernel_pmap);
1423 
1424 	if (needed_lock)
1425 		PMAP_LOCK(kernel_pmap);
1426 
1427         if ((wait & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT)
1428                 pflags = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
1429         else
1430                 pflags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED;
1431         if (wait & M_ZERO)
1432                 pflags |= VM_ALLOC_ZERO;
1433 
1434         for (;;) {
1435                 m = vm_page_alloc(NULL, color++, pflags | VM_ALLOC_NOOBJ);
1436                 if (m == NULL) {
1437                         if (wait & M_NOWAIT)
1438                                 return (NULL);
1439                         VM_WAIT;
1440                 } else
1441                         break;
1442         }
1443 
1444 	va = VM_PAGE_TO_PHYS(m);
1445 
1446 	moea64_pvo_enter(kernel_pmap, moea64_upvo_zone,
1447 	    &moea64_pvo_kunmanaged, va, VM_PAGE_TO_PHYS(m), LPTE_M,
1448 	    PVO_WIRED | PVO_BOOTSTRAP);
1449 
1450 	if (needed_lock)
1451 		PMAP_UNLOCK(kernel_pmap);
1452 
1453 	if ((wait & M_ZERO) && (m->flags & PG_ZERO) == 0)
1454                 bzero((void *)va, PAGE_SIZE);
1455 
1456 	return (void *)va;
1457 }
1458 
1459 void
1460 moea64_init(mmu_t mmu)
1461 {
1462 
1463 	CTR0(KTR_PMAP, "moea64_init");
1464 
1465 	moea64_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1466 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1467 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1468 	moea64_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1469 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1470 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1471 
1472 	if (!hw_direct_map) {
1473 		uma_zone_set_allocf(moea64_upvo_zone,moea64_uma_page_alloc);
1474 		uma_zone_set_allocf(moea64_mpvo_zone,moea64_uma_page_alloc);
1475 	}
1476 
1477 	moea64_initialized = TRUE;
1478 }
1479 
1480 boolean_t
1481 moea64_is_referenced(mmu_t mmu, vm_page_t m)
1482 {
1483 
1484 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1485 	    ("moea64_is_referenced: page %p is not managed", m));
1486 	return (moea64_query_bit(m, PTE_REF));
1487 }
1488 
1489 boolean_t
1490 moea64_is_modified(mmu_t mmu, vm_page_t m)
1491 {
1492 
1493 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1494 	    ("moea64_is_modified: page %p is not managed", m));
1495 
1496 	/*
1497 	 * If the page is not VPO_BUSY, then PG_WRITEABLE cannot be
1498 	 * concurrently set while the object is locked.  Thus, if PG_WRITEABLE
1499 	 * is clear, no PTEs can have LPTE_CHG set.
1500 	 */
1501 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1502 	if ((m->oflags & VPO_BUSY) == 0 &&
1503 	    (m->flags & PG_WRITEABLE) == 0)
1504 		return (FALSE);
1505 	return (moea64_query_bit(m, LPTE_CHG));
1506 }
1507 
1508 void
1509 moea64_clear_reference(mmu_t mmu, vm_page_t m)
1510 {
1511 
1512 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1513 	    ("moea64_clear_reference: page %p is not managed", m));
1514 	moea64_clear_bit(m, LPTE_REF);
1515 }
1516 
1517 void
1518 moea64_clear_modify(mmu_t mmu, vm_page_t m)
1519 {
1520 
1521 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1522 	    ("moea64_clear_modify: page %p is not managed", m));
1523 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1524 	KASSERT((m->oflags & VPO_BUSY) == 0,
1525 	    ("moea64_clear_modify: page %p is busy", m));
1526 
1527 	/*
1528 	 * If the page is not PG_WRITEABLE, then no PTEs can have LPTE_CHG
1529 	 * set.  If the object containing the page is locked and the page is
1530 	 * not VPO_BUSY, then PG_WRITEABLE cannot be concurrently set.
1531 	 */
1532 	if ((m->flags & PG_WRITEABLE) == 0)
1533 		return;
1534 	moea64_clear_bit(m, LPTE_CHG);
1535 }
1536 
1537 /*
1538  * Clear the write and modified bits in each of the given page's mappings.
1539  */
1540 void
1541 moea64_remove_write(mmu_t mmu, vm_page_t m)
1542 {
1543 	struct	pvo_entry *pvo;
1544 	struct	lpte *pt;
1545 	pmap_t	pmap;
1546 	uint64_t lo;
1547 
1548 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1549 	    ("moea64_remove_write: page %p is not managed", m));
1550 
1551 	/*
1552 	 * If the page is not VPO_BUSY, then PG_WRITEABLE cannot be set by
1553 	 * another thread while the object is locked.  Thus, if PG_WRITEABLE
1554 	 * is clear, no page table entries need updating.
1555 	 */
1556 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1557 	if ((m->oflags & VPO_BUSY) == 0 &&
1558 	    (m->flags & PG_WRITEABLE) == 0)
1559 		return;
1560 	vm_page_lock_queues();
1561 	lo = moea64_attr_fetch(m);
1562 	SYNC();
1563 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1564 		pmap = pvo->pvo_pmap;
1565 		PMAP_LOCK(pmap);
1566 		LOCK_TABLE();
1567 		if ((pvo->pvo_pte.lpte.pte_lo & LPTE_PP) != LPTE_BR) {
1568 			pt = moea64_pvo_to_pte(pvo, -1);
1569 			pvo->pvo_pte.lpte.pte_lo &= ~LPTE_PP;
1570 			pvo->pvo_pte.lpte.pte_lo |= LPTE_BR;
1571 			if (pt != NULL) {
1572 				moea64_pte_synch(pt, &pvo->pvo_pte.lpte);
1573 				lo |= pvo->pvo_pte.lpte.pte_lo;
1574 				pvo->pvo_pte.lpte.pte_lo &= ~LPTE_CHG;
1575 				moea64_pte_change(pt, &pvo->pvo_pte.lpte,
1576 				    pvo->pvo_pmap, PVO_VADDR(pvo));
1577 			}
1578 		}
1579 		UNLOCK_TABLE();
1580 		PMAP_UNLOCK(pmap);
1581 	}
1582 	if ((lo & LPTE_CHG) != 0) {
1583 		moea64_attr_clear(m, LPTE_CHG);
1584 		vm_page_dirty(m);
1585 	}
1586 	vm_page_flag_clear(m, PG_WRITEABLE);
1587 	vm_page_unlock_queues();
1588 }
1589 
1590 /*
1591  *	moea64_ts_referenced:
1592  *
1593  *	Return a count of reference bits for a page, clearing those bits.
1594  *	It is not necessary for every reference bit to be cleared, but it
1595  *	is necessary that 0 only be returned when there are truly no
1596  *	reference bits set.
1597  *
1598  *	XXX: The exact number of bits to check and clear is a matter that
1599  *	should be tested and standardized at some point in the future for
1600  *	optimal aging of shared pages.
1601  */
1602 boolean_t
1603 moea64_ts_referenced(mmu_t mmu, vm_page_t m)
1604 {
1605 
1606 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1607 	    ("moea64_ts_referenced: page %p is not managed", m));
1608 	return (moea64_clear_bit(m, LPTE_REF));
1609 }
1610 
1611 /*
1612  * Map a wired page into kernel virtual address space.
1613  */
1614 void
1615 moea64_kenter(mmu_t mmu, vm_offset_t va, vm_offset_t pa)
1616 {
1617 	uint64_t	pte_lo;
1618 	int		error;
1619 
1620 #if 0
1621 	if (!pmap_bootstrapped) {
1622 		if (va >= VM_MIN_KERNEL_ADDRESS && va < virtual_end)
1623 			panic("Trying to enter an address in KVA -- %#x!\n",pa);
1624 	}
1625 #endif
1626 
1627 	pte_lo = moea64_calc_wimg(pa);
1628 
1629 	PMAP_LOCK(kernel_pmap);
1630 	error = moea64_pvo_enter(kernel_pmap, moea64_upvo_zone,
1631 	    &moea64_pvo_kunmanaged, va, pa, pte_lo,
1632 	    PVO_WIRED | VM_PROT_EXECUTE);
1633 
1634 	if (error != 0 && error != ENOENT)
1635 		panic("moea64_kenter: failed to enter va %#x pa %#x: %d", va,
1636 		    pa, error);
1637 
1638 	/*
1639 	 * Flush the memory from the instruction cache.
1640 	 */
1641 	if ((pte_lo & (LPTE_I | LPTE_G)) == 0) {
1642 		__syncicache((void *)va, PAGE_SIZE);
1643 	}
1644 	PMAP_UNLOCK(kernel_pmap);
1645 }
1646 
1647 /*
1648  * Extract the physical page address associated with the given kernel virtual
1649  * address.
1650  */
1651 vm_offset_t
1652 moea64_kextract(mmu_t mmu, vm_offset_t va)
1653 {
1654 	struct		pvo_entry *pvo;
1655 	vm_paddr_t pa;
1656 
1657 	/*
1658 	 * Shortcut the direct-mapped case when applicable.  We never put
1659 	 * anything but 1:1 mappings below VM_MIN_KERNEL_ADDRESS.
1660 	 */
1661 	if (va < VM_MIN_KERNEL_ADDRESS)
1662 		return (va);
1663 
1664 	PMAP_LOCK(kernel_pmap);
1665 	pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1666 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found"));
1667 	pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) | (va & ADDR_POFF);
1668 	PMAP_UNLOCK(kernel_pmap);
1669 	return (pa);
1670 }
1671 
1672 /*
1673  * Remove a wired page from kernel virtual address space.
1674  */
1675 void
1676 moea64_kremove(mmu_t mmu, vm_offset_t va)
1677 {
1678 	moea64_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1679 }
1680 
1681 /*
1682  * Map a range of physical addresses into kernel virtual address space.
1683  *
1684  * The value passed in *virt is a suggested virtual address for the mapping.
1685  * Architectures which can support a direct-mapped physical to virtual region
1686  * can return the appropriate address within that region, leaving '*virt'
1687  * unchanged.  We cannot and therefore do not; *virt is updated with the
1688  * first usable address after the mapped region.
1689  */
1690 vm_offset_t
1691 moea64_map(mmu_t mmu, vm_offset_t *virt, vm_offset_t pa_start,
1692     vm_offset_t pa_end, int prot)
1693 {
1694 	vm_offset_t	sva, va;
1695 
1696 	sva = *virt;
1697 	va = sva;
1698 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1699 		moea64_kenter(mmu, va, pa_start);
1700 	*virt = va;
1701 
1702 	return (sva);
1703 }
1704 
1705 /*
1706  * Returns true if the pmap's pv is one of the first
1707  * 16 pvs linked to from this page.  This count may
1708  * be changed upwards or downwards in the future; it
1709  * is only necessary that true be returned for a small
1710  * subset of pmaps for proper page aging.
1711  */
1712 boolean_t
1713 moea64_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1714 {
1715         int loops;
1716 	struct pvo_entry *pvo;
1717 	boolean_t rv;
1718 
1719 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1720 	    ("moea64_page_exists_quick: page %p is not managed", m));
1721 	loops = 0;
1722 	rv = FALSE;
1723 	vm_page_lock_queues();
1724 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1725 		if (pvo->pvo_pmap == pmap) {
1726 			rv = TRUE;
1727 			break;
1728 		}
1729 		if (++loops >= 16)
1730 			break;
1731 	}
1732 	vm_page_unlock_queues();
1733 	return (rv);
1734 }
1735 
1736 /*
1737  * Return the number of managed mappings to the given physical page
1738  * that are wired.
1739  */
1740 int
1741 moea64_page_wired_mappings(mmu_t mmu, vm_page_t m)
1742 {
1743 	struct pvo_entry *pvo;
1744 	int count;
1745 
1746 	count = 0;
1747 	if ((m->flags & PG_FICTITIOUS) != 0)
1748 		return (count);
1749 	vm_page_lock_queues();
1750 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1751 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1752 			count++;
1753 	vm_page_unlock_queues();
1754 	return (count);
1755 }
1756 
1757 static u_int	moea64_vsidcontext;
1758 
1759 void
1760 moea64_pinit(mmu_t mmu, pmap_t pmap)
1761 {
1762 	int	i, mask;
1763 	u_int	entropy;
1764 
1765 	PMAP_LOCK_INIT(pmap);
1766 
1767 	entropy = 0;
1768 	__asm __volatile("mftb %0" : "=r"(entropy));
1769 
1770 	if (pmap_bootstrapped)
1771 		pmap->pmap_phys = (pmap_t)moea64_kextract(mmu, (vm_offset_t)pmap);
1772 	else
1773 		pmap->pmap_phys = pmap;
1774 
1775 	/*
1776 	 * Allocate some segment registers for this pmap.
1777 	 */
1778 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1779 		u_int	hash, n;
1780 
1781 		/*
1782 		 * Create a new value by mutiplying by a prime and adding in
1783 		 * entropy from the timebase register.  This is to make the
1784 		 * VSID more random so that the PT hash function collides
1785 		 * less often.  (Note that the prime casues gcc to do shifts
1786 		 * instead of a multiply.)
1787 		 */
1788 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
1789 		hash = moea64_vsidcontext & (NPMAPS - 1);
1790 		if (hash == 0)		/* 0 is special, avoid it */
1791 			continue;
1792 		n = hash >> 5;
1793 		mask = 1 << (hash & (VSID_NBPW - 1));
1794 		hash = (moea64_vsidcontext & 0xfffff);
1795 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
1796 			/* anything free in this bucket? */
1797 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
1798 				entropy = (moea64_vsidcontext >> 20);
1799 				continue;
1800 			}
1801 			i = ffs(~moea64_vsid_bitmap[i]) - 1;
1802 			mask = 1 << i;
1803 			hash &= 0xfffff & ~(VSID_NBPW - 1);
1804 			hash |= i;
1805 		}
1806 		moea64_vsid_bitmap[n] |= mask;
1807 		for (i = 0; i < 16; i++) {
1808 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1809 		}
1810 		return;
1811 	}
1812 
1813 	panic("moea64_pinit: out of segments");
1814 }
1815 
1816 /*
1817  * Initialize the pmap associated with process 0.
1818  */
1819 void
1820 moea64_pinit0(mmu_t mmu, pmap_t pm)
1821 {
1822 	moea64_pinit(mmu, pm);
1823 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1824 }
1825 
1826 /*
1827  * Set the physical protection on the specified range of this map as requested.
1828  */
1829 void
1830 moea64_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1831     vm_prot_t prot)
1832 {
1833 	struct	pvo_entry *pvo;
1834 	struct	lpte *pt;
1835 	int	pteidx;
1836 
1837 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm, sva,
1838 	    eva, prot);
1839 
1840 
1841 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1842 	    ("moea64_protect: non current pmap"));
1843 
1844 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1845 		moea64_remove(mmu, pm, sva, eva);
1846 		return;
1847 	}
1848 
1849 	vm_page_lock_queues();
1850 	PMAP_LOCK(pm);
1851 	for (; sva < eva; sva += PAGE_SIZE) {
1852 		pvo = moea64_pvo_find_va(pm, sva, &pteidx);
1853 		if (pvo == NULL)
1854 			continue;
1855 
1856 		/*
1857 		 * Grab the PTE pointer before we diddle with the cached PTE
1858 		 * copy.
1859 		 */
1860 		LOCK_TABLE();
1861 		pt = moea64_pvo_to_pte(pvo, pteidx);
1862 
1863 		/*
1864 		 * Change the protection of the page.
1865 		 */
1866 		pvo->pvo_pte.lpte.pte_lo &= ~LPTE_PP;
1867 		pvo->pvo_pte.lpte.pte_lo |= LPTE_BR;
1868 		pvo->pvo_pte.lpte.pte_lo &= ~LPTE_NOEXEC;
1869 		if ((prot & VM_PROT_EXECUTE) == 0)
1870 			pvo->pvo_pte.lpte.pte_lo |= LPTE_NOEXEC;
1871 
1872 		/*
1873 		 * If the PVO is in the page table, update that pte as well.
1874 		 */
1875 		if (pt != NULL) {
1876 			moea64_pte_change(pt, &pvo->pvo_pte.lpte,
1877 			    pvo->pvo_pmap, PVO_VADDR(pvo));
1878 			if ((pvo->pvo_pte.lpte.pte_lo &
1879 			    (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1880 				moea64_syncicache(pm, sva,
1881 				    pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN,
1882 				    PAGE_SIZE);
1883 			}
1884 		}
1885 		UNLOCK_TABLE();
1886 	}
1887 	vm_page_unlock_queues();
1888 	PMAP_UNLOCK(pm);
1889 }
1890 
1891 /*
1892  * Map a list of wired pages into kernel virtual address space.  This is
1893  * intended for temporary mappings which do not need page modification or
1894  * references recorded.  Existing mappings in the region are overwritten.
1895  */
1896 void
1897 moea64_qenter(mmu_t mmu, vm_offset_t va, vm_page_t *m, int count)
1898 {
1899 	while (count-- > 0) {
1900 		moea64_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1901 		va += PAGE_SIZE;
1902 		m++;
1903 	}
1904 }
1905 
1906 /*
1907  * Remove page mappings from kernel virtual address space.  Intended for
1908  * temporary mappings entered by moea64_qenter.
1909  */
1910 void
1911 moea64_qremove(mmu_t mmu, vm_offset_t va, int count)
1912 {
1913 	while (count-- > 0) {
1914 		moea64_kremove(mmu, va);
1915 		va += PAGE_SIZE;
1916 	}
1917 }
1918 
1919 void
1920 moea64_release(mmu_t mmu, pmap_t pmap)
1921 {
1922         int idx, mask;
1923 
1924 	/*
1925 	 * Free segment register's VSID
1926 	 */
1927         if (pmap->pm_sr[0] == 0)
1928                 panic("moea64_release");
1929 
1930         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1931         mask = 1 << (idx % VSID_NBPW);
1932         idx /= VSID_NBPW;
1933         moea64_vsid_bitmap[idx] &= ~mask;
1934 	PMAP_LOCK_DESTROY(pmap);
1935 }
1936 
1937 /*
1938  * Remove the given range of addresses from the specified map.
1939  */
1940 void
1941 moea64_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1942 {
1943 	struct	pvo_entry *pvo;
1944 	int	pteidx;
1945 
1946 	vm_page_lock_queues();
1947 	PMAP_LOCK(pm);
1948 	for (; sva < eva; sva += PAGE_SIZE) {
1949 		pvo = moea64_pvo_find_va(pm, sva, &pteidx);
1950 		if (pvo != NULL) {
1951 			moea64_pvo_remove(pvo, pteidx);
1952 		}
1953 	}
1954 	vm_page_unlock_queues();
1955 	PMAP_UNLOCK(pm);
1956 }
1957 
1958 /*
1959  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
1960  * will reflect changes in pte's back to the vm_page.
1961  */
1962 void
1963 moea64_remove_all(mmu_t mmu, vm_page_t m)
1964 {
1965 	struct  pvo_head *pvo_head;
1966 	struct	pvo_entry *pvo, *next_pvo;
1967 	pmap_t	pmap;
1968 
1969 	vm_page_lock_queues();
1970 	pvo_head = vm_page_to_pvoh(m);
1971 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1972 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1973 
1974 		MOEA_PVO_CHECK(pvo);	/* sanity check */
1975 		pmap = pvo->pvo_pmap;
1976 		PMAP_LOCK(pmap);
1977 		moea64_pvo_remove(pvo, -1);
1978 		PMAP_UNLOCK(pmap);
1979 	}
1980 	if ((m->flags & PG_WRITEABLE) && moea64_is_modified(mmu, m)) {
1981 		moea64_attr_clear(m, LPTE_CHG);
1982 		vm_page_dirty(m);
1983 	}
1984 	vm_page_flag_clear(m, PG_WRITEABLE);
1985 	vm_page_unlock_queues();
1986 }
1987 
1988 /*
1989  * Allocate a physical page of memory directly from the phys_avail map.
1990  * Can only be called from moea64_bootstrap before avail start and end are
1991  * calculated.
1992  */
1993 static vm_offset_t
1994 moea64_bootstrap_alloc(vm_size_t size, u_int align)
1995 {
1996 	vm_offset_t	s, e;
1997 	int		i, j;
1998 
1999 	size = round_page(size);
2000 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2001 		if (align != 0)
2002 			s = (phys_avail[i] + align - 1) & ~(align - 1);
2003 		else
2004 			s = phys_avail[i];
2005 		e = s + size;
2006 
2007 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2008 			continue;
2009 
2010 		if (s == phys_avail[i]) {
2011 			phys_avail[i] += size;
2012 		} else if (e == phys_avail[i + 1]) {
2013 			phys_avail[i + 1] -= size;
2014 		} else {
2015 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2016 				phys_avail[j] = phys_avail[j - 2];
2017 				phys_avail[j + 1] = phys_avail[j - 1];
2018 			}
2019 
2020 			phys_avail[i + 3] = phys_avail[i + 1];
2021 			phys_avail[i + 1] = s;
2022 			phys_avail[i + 2] = e;
2023 			phys_avail_count++;
2024 		}
2025 
2026 		return (s);
2027 	}
2028 	panic("moea64_bootstrap_alloc: could not allocate memory");
2029 }
2030 
2031 static void
2032 tlbia(void)
2033 {
2034 	vm_offset_t i;
2035 	register_t msr, scratch;
2036 
2037 	for (i = 0; i < 0xFF000; i += 0x00001000) {
2038 		__asm __volatile("\
2039 		    mfmsr %0; \
2040 		    mr %1, %0; \
2041 		    insrdi %1,%3,1,0; \
2042 		    mtmsrd %1; \
2043 		    ptesync; \
2044 		    \
2045 		    tlbiel %2; \
2046 		    \
2047 		    mtmsrd %0; \
2048 		    eieio; \
2049 		    tlbsync; \
2050 		    ptesync;"
2051 		: "=r"(msr), "=r"(scratch) : "r"(i), "r"(1));
2052 	}
2053 }
2054 
2055 static int
2056 moea64_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
2057     vm_offset_t va, vm_offset_t pa, uint64_t pte_lo, int flags)
2058 {
2059 	struct	 pvo_entry *pvo;
2060 	uint64_t vsid;
2061 	int	 first;
2062 	u_int	 ptegidx;
2063 	int	 i;
2064 	int      bootstrap;
2065 
2066 	/*
2067 	 * One nasty thing that can happen here is that the UMA calls to
2068 	 * allocate new PVOs need to map more memory, which calls pvo_enter(),
2069 	 * which calls UMA...
2070 	 *
2071 	 * We break the loop by detecting recursion and allocating out of
2072 	 * the bootstrap pool.
2073 	 */
2074 
2075 	moea64_pvo_enter_calls++;
2076 	first = 0;
2077 	bootstrap = (flags & PVO_BOOTSTRAP);
2078 
2079 	if (!moea64_initialized)
2080 		bootstrap = 1;
2081 
2082 	/*
2083 	 * Compute the PTE Group index.
2084 	 */
2085 	va &= ~ADDR_POFF;
2086 	vsid = va_to_vsid(pm, va);
2087 	ptegidx = va_to_pteg(vsid, va);
2088 
2089 	/*
2090 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
2091 	 * there is a mapping.
2092 	 */
2093 	LOCK_TABLE();
2094 
2095 	LIST_FOREACH(pvo, &moea64_pvo_table[ptegidx], pvo_olink) {
2096 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2097 			if ((pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) == pa &&
2098 			    (pvo->pvo_pte.lpte.pte_lo & LPTE_PP) ==
2099 			    (pte_lo & LPTE_PP)) {
2100 				UNLOCK_TABLE();
2101 				return (0);
2102 			}
2103 			moea64_pvo_remove(pvo, -1);
2104 			break;
2105 		}
2106 	}
2107 
2108 	/*
2109 	 * If we aren't overwriting a mapping, try to allocate.
2110 	 */
2111 	if (bootstrap) {
2112 		if (moea64_bpvo_pool_index >= BPVO_POOL_SIZE) {
2113 			panic("moea64_enter: bpvo pool exhausted, %d, %d, %d",
2114 			      moea64_bpvo_pool_index, BPVO_POOL_SIZE,
2115 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
2116 		}
2117 		pvo = &moea64_bpvo_pool[moea64_bpvo_pool_index];
2118 		moea64_bpvo_pool_index++;
2119 		bootstrap = 1;
2120 	} else {
2121 		/*
2122 		 * Note: drop the table lock around the UMA allocation in
2123 		 * case the UMA allocator needs to manipulate the page
2124 		 * table. The mapping we are working with is already
2125 		 * protected by the PMAP lock.
2126 		 */
2127 		UNLOCK_TABLE();
2128 		pvo = uma_zalloc(zone, M_NOWAIT);
2129 		LOCK_TABLE();
2130 	}
2131 
2132 	if (pvo == NULL) {
2133 		UNLOCK_TABLE();
2134 		return (ENOMEM);
2135 	}
2136 
2137 	moea64_pvo_entries++;
2138 	pvo->pvo_vaddr = va;
2139 	pvo->pvo_pmap = pm;
2140 	LIST_INSERT_HEAD(&moea64_pvo_table[ptegidx], pvo, pvo_olink);
2141 	pvo->pvo_vaddr &= ~ADDR_POFF;
2142 
2143 	if (!(flags & VM_PROT_EXECUTE))
2144 		pte_lo |= LPTE_NOEXEC;
2145 	if (flags & PVO_WIRED)
2146 		pvo->pvo_vaddr |= PVO_WIRED;
2147 	if (pvo_head != &moea64_pvo_kunmanaged)
2148 		pvo->pvo_vaddr |= PVO_MANAGED;
2149 	if (bootstrap)
2150 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
2151 	if (flags & PVO_FAKE)
2152 		pvo->pvo_vaddr |= PVO_FAKE;
2153 
2154 	moea64_pte_create(&pvo->pvo_pte.lpte, vsid, va,
2155 	    (uint64_t)(pa) | pte_lo);
2156 
2157 	/*
2158 	 * Remember if the list was empty and therefore will be the first
2159 	 * item.
2160 	 */
2161 	if (LIST_FIRST(pvo_head) == NULL)
2162 		first = 1;
2163 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2164 
2165 	if (pvo->pvo_vaddr & PVO_WIRED)
2166 		pm->pm_stats.wired_count++;
2167 	pm->pm_stats.resident_count++;
2168 
2169 	/*
2170 	 * We hope this succeeds but it isn't required.
2171 	 */
2172 	i = moea64_pte_insert(ptegidx, &pvo->pvo_pte.lpte);
2173 	if (i >= 0) {
2174 		PVO_PTEGIDX_SET(pvo, i);
2175 	} else {
2176 		panic("moea64_pvo_enter: overflow");
2177 		moea64_pte_overflow++;
2178 	}
2179 
2180 	if (pm == kernel_pmap)
2181 		isync();
2182 
2183 	UNLOCK_TABLE();
2184 
2185 	return (first ? ENOENT : 0);
2186 }
2187 
2188 static void
2189 moea64_pvo_remove(struct pvo_entry *pvo, int pteidx)
2190 {
2191 	struct	lpte *pt;
2192 
2193 	/*
2194 	 * If there is an active pte entry, we need to deactivate it (and
2195 	 * save the ref & cfg bits).
2196 	 */
2197 	LOCK_TABLE();
2198 	pt = moea64_pvo_to_pte(pvo, pteidx);
2199 	if (pt != NULL) {
2200 		moea64_pte_unset(pt, &pvo->pvo_pte.lpte, pvo->pvo_pmap,
2201 		    PVO_VADDR(pvo));
2202 		PVO_PTEGIDX_CLR(pvo);
2203 	} else {
2204 		moea64_pte_overflow--;
2205 	}
2206 
2207 	/*
2208 	 * Update our statistics.
2209 	 */
2210 	pvo->pvo_pmap->pm_stats.resident_count--;
2211 	if (pvo->pvo_vaddr & PVO_WIRED)
2212 		pvo->pvo_pmap->pm_stats.wired_count--;
2213 
2214 	/*
2215 	 * Save the REF/CHG bits into their cache if the page is managed.
2216 	 */
2217 	if ((pvo->pvo_vaddr & (PVO_MANAGED|PVO_FAKE)) == PVO_MANAGED) {
2218 		struct	vm_page *pg;
2219 
2220 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN);
2221 		if (pg != NULL) {
2222 			moea64_attr_save(pg, pvo->pvo_pte.lpte.pte_lo &
2223 			    (LPTE_REF | LPTE_CHG));
2224 		}
2225 	}
2226 
2227 	/*
2228 	 * Remove this PVO from the PV list.
2229 	 */
2230 	LIST_REMOVE(pvo, pvo_vlink);
2231 
2232 	/*
2233 	 * Remove this from the overflow list and return it to the pool
2234 	 * if we aren't going to reuse it.
2235 	 */
2236 	LIST_REMOVE(pvo, pvo_olink);
2237 	UNLOCK_TABLE();
2238 
2239 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2240 		uma_zfree((pvo->pvo_vaddr & PVO_MANAGED) ? moea64_mpvo_zone :
2241 		    moea64_upvo_zone, pvo);
2242 
2243 	moea64_pvo_entries--;
2244 	moea64_pvo_remove_calls++;
2245 }
2246 
2247 static __inline int
2248 moea64_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2249 {
2250 
2251 	/*
2252 	 * We can find the actual pte entry without searching by grabbing
2253 	 * the PTEG index from 3 unused bits in pvo_vaddr and by
2254 	 * noticing the HID bit.
2255 	 */
2256 	if (pvo->pvo_pte.lpte.pte_hi & LPTE_HID)
2257 		ptegidx ^= moea64_pteg_mask;
2258 
2259 	return ((ptegidx << 3) | PVO_PTEGIDX_GET(pvo));
2260 }
2261 
2262 static struct pvo_entry *
2263 moea64_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2264 {
2265 	struct		pvo_entry *pvo;
2266 	int		ptegidx;
2267 	uint64_t	vsid;
2268 
2269 	va &= ~ADDR_POFF;
2270 	vsid = va_to_vsid(pm, va);
2271 	ptegidx = va_to_pteg(vsid, va);
2272 
2273 	LOCK_TABLE();
2274 	LIST_FOREACH(pvo, &moea64_pvo_table[ptegidx], pvo_olink) {
2275 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2276 			if (pteidx_p)
2277 				*pteidx_p = moea64_pvo_pte_index(pvo, ptegidx);
2278 			break;
2279 		}
2280 	}
2281 	UNLOCK_TABLE();
2282 
2283 	return (pvo);
2284 }
2285 
2286 static struct lpte *
2287 moea64_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2288 {
2289 	struct lpte *pt;
2290 
2291 	/*
2292 	 * If we haven't been supplied the ptegidx, calculate it.
2293 	 */
2294 	if (pteidx == -1) {
2295 		int		ptegidx;
2296 		uint64_t	vsid;
2297 
2298 		vsid = va_to_vsid(pvo->pvo_pmap, PVO_VADDR(pvo));
2299 		ptegidx = va_to_pteg(vsid, PVO_VADDR(pvo));
2300 		pteidx = moea64_pvo_pte_index(pvo, ptegidx);
2301 	}
2302 
2303 	pt = &moea64_pteg_table[pteidx >> 3].pt[pteidx & 7];
2304 
2305 	if ((pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) &&
2306 	    !PVO_PTEGIDX_ISSET(pvo)) {
2307 		panic("moea64_pvo_to_pte: pvo %p has valid pte in pvo but no "
2308 		    "valid pte index", pvo);
2309 	}
2310 
2311 	if ((pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) == 0 &&
2312 	    PVO_PTEGIDX_ISSET(pvo)) {
2313 		panic("moea64_pvo_to_pte: pvo %p has valid pte index in pvo "
2314 		    "pvo but no valid pte", pvo);
2315 	}
2316 
2317 	if ((pt->pte_hi ^ (pvo->pvo_pte.lpte.pte_hi & ~LPTE_VALID)) ==
2318 	    LPTE_VALID) {
2319 		if ((pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) == 0) {
2320 			panic("moea64_pvo_to_pte: pvo %p has valid pte in "
2321 			    "moea64_pteg_table %p but invalid in pvo", pvo, pt);
2322 		}
2323 
2324 		if (((pt->pte_lo ^ pvo->pvo_pte.lpte.pte_lo) &
2325 		    ~(LPTE_M|LPTE_CHG|LPTE_REF)) != 0) {
2326 			panic("moea64_pvo_to_pte: pvo %p pte does not match "
2327 			    "pte %p in moea64_pteg_table difference is %#x",
2328 			    pvo, pt,
2329 			    (uint32_t)(pt->pte_lo ^ pvo->pvo_pte.lpte.pte_lo));
2330 		}
2331 
2332 		ASSERT_TABLE_LOCK();
2333 		return (pt);
2334 	}
2335 
2336 	if (pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) {
2337 		panic("moea64_pvo_to_pte: pvo %p has invalid pte %p in "
2338 		    "moea64_pteg_table but valid in pvo", pvo, pt);
2339 	}
2340 
2341 	return (NULL);
2342 }
2343 
2344 static int
2345 moea64_pte_insert(u_int ptegidx, struct lpte *pvo_pt)
2346 {
2347 	struct	lpte *pt;
2348 	int	i;
2349 
2350 	ASSERT_TABLE_LOCK();
2351 
2352 	/*
2353 	 * First try primary hash.
2354 	 */
2355 	for (pt = moea64_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2356 		if ((pt->pte_hi & LPTE_VALID) == 0 &&
2357 		    (pt->pte_hi & LPTE_LOCKED) == 0) {
2358 			pvo_pt->pte_hi &= ~LPTE_HID;
2359 			moea64_pte_set(pt, pvo_pt);
2360 			return (i);
2361 		}
2362 	}
2363 
2364 	/*
2365 	 * Now try secondary hash.
2366 	 */
2367 	ptegidx ^= moea64_pteg_mask;
2368 
2369 	for (pt = moea64_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2370 		if ((pt->pte_hi & LPTE_VALID) == 0 &&
2371 		    (pt->pte_hi & LPTE_LOCKED) == 0) {
2372 			pvo_pt->pte_hi |= LPTE_HID;
2373 			moea64_pte_set(pt, pvo_pt);
2374 			return (i);
2375 		}
2376 	}
2377 
2378 	panic("moea64_pte_insert: overflow");
2379 	return (-1);
2380 }
2381 
2382 static boolean_t
2383 moea64_query_bit(vm_page_t m, u_int64_t ptebit)
2384 {
2385 	struct	pvo_entry *pvo;
2386 	struct	lpte *pt;
2387 
2388 	if (moea64_attr_fetch(m) & ptebit)
2389 		return (TRUE);
2390 
2391 	vm_page_lock_queues();
2392 
2393 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2394 		MOEA_PVO_CHECK(pvo);	/* sanity check */
2395 
2396 		/*
2397 		 * See if we saved the bit off.  If so, cache it and return
2398 		 * success.
2399 		 */
2400 		if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2401 			moea64_attr_save(m, ptebit);
2402 			MOEA_PVO_CHECK(pvo);	/* sanity check */
2403 			vm_page_unlock_queues();
2404 			return (TRUE);
2405 		}
2406 	}
2407 
2408 	/*
2409 	 * No luck, now go through the hard part of looking at the PTEs
2410 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2411 	 * the PTEs.
2412 	 */
2413 	SYNC();
2414 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2415 		MOEA_PVO_CHECK(pvo);	/* sanity check */
2416 
2417 		/*
2418 		 * See if this pvo has a valid PTE.  if so, fetch the
2419 		 * REF/CHG bits from the valid PTE.  If the appropriate
2420 		 * ptebit is set, cache it and return success.
2421 		 */
2422 		LOCK_TABLE();
2423 		pt = moea64_pvo_to_pte(pvo, -1);
2424 		if (pt != NULL) {
2425 			moea64_pte_synch(pt, &pvo->pvo_pte.lpte);
2426 			if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2427 				UNLOCK_TABLE();
2428 
2429 				moea64_attr_save(m, ptebit);
2430 				MOEA_PVO_CHECK(pvo);	/* sanity check */
2431 				vm_page_unlock_queues();
2432 				return (TRUE);
2433 			}
2434 		}
2435 		UNLOCK_TABLE();
2436 	}
2437 
2438 	vm_page_unlock_queues();
2439 	return (FALSE);
2440 }
2441 
2442 static u_int
2443 moea64_clear_bit(vm_page_t m, u_int64_t ptebit)
2444 {
2445 	u_int	count;
2446 	struct	pvo_entry *pvo;
2447 	struct	lpte *pt;
2448 
2449 	vm_page_lock_queues();
2450 
2451 	/*
2452 	 * Clear the cached value.
2453 	 */
2454 	moea64_attr_clear(m, ptebit);
2455 
2456 	/*
2457 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2458 	 * we can reset the right ones).  note that since the pvo entries and
2459 	 * list heads are accessed via BAT0 and are never placed in the page
2460 	 * table, we don't have to worry about further accesses setting the
2461 	 * REF/CHG bits.
2462 	 */
2463 	SYNC();
2464 
2465 	/*
2466 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2467 	 * valid pte clear the ptebit from the valid pte.
2468 	 */
2469 	count = 0;
2470 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2471 		MOEA_PVO_CHECK(pvo);	/* sanity check */
2472 
2473 		LOCK_TABLE();
2474 		pt = moea64_pvo_to_pte(pvo, -1);
2475 		if (pt != NULL) {
2476 			moea64_pte_synch(pt, &pvo->pvo_pte.lpte);
2477 			if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2478 				count++;
2479 				moea64_pte_clear(pt, pvo->pvo_pmap, PVO_VADDR(pvo), ptebit);
2480 			}
2481 		}
2482 		pvo->pvo_pte.lpte.pte_lo &= ~ptebit;
2483 		MOEA_PVO_CHECK(pvo);	/* sanity check */
2484 		UNLOCK_TABLE();
2485 	}
2486 
2487 	vm_page_unlock_queues();
2488 	return (count);
2489 }
2490 
2491 boolean_t
2492 moea64_dev_direct_mapped(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2493 {
2494 	struct pvo_entry *pvo;
2495 	vm_offset_t ppa;
2496 	int error = 0;
2497 
2498 	PMAP_LOCK(kernel_pmap);
2499 	for (ppa = pa & ~ADDR_POFF; ppa < pa + size; ppa += PAGE_SIZE) {
2500 		pvo = moea64_pvo_find_va(kernel_pmap, ppa, NULL);
2501 		if (pvo == NULL ||
2502 		    (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) != ppa) {
2503 			error = EFAULT;
2504 			break;
2505 		}
2506 	}
2507 	PMAP_UNLOCK(kernel_pmap);
2508 
2509 	return (error);
2510 }
2511 
2512 /*
2513  * Map a set of physical memory pages into the kernel virtual
2514  * address space. Return a pointer to where it is mapped. This
2515  * routine is intended to be used for mapping device memory,
2516  * NOT real memory.
2517  */
2518 void *
2519 moea64_mapdev(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2520 {
2521 	vm_offset_t va, tmpva, ppa, offset;
2522 
2523 	ppa = trunc_page(pa);
2524 	offset = pa & PAGE_MASK;
2525 	size = roundup(offset + size, PAGE_SIZE);
2526 
2527 	va = kmem_alloc_nofault(kernel_map, size);
2528 
2529 	if (!va)
2530 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
2531 
2532 	for (tmpva = va; size > 0;) {
2533 		moea64_kenter(mmu, tmpva, ppa);
2534 		size -= PAGE_SIZE;
2535 		tmpva += PAGE_SIZE;
2536 		ppa += PAGE_SIZE;
2537 	}
2538 
2539 	return ((void *)(va + offset));
2540 }
2541 
2542 void
2543 moea64_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2544 {
2545 	vm_offset_t base, offset;
2546 
2547 	base = trunc_page(va);
2548 	offset = va & PAGE_MASK;
2549 	size = roundup(offset + size, PAGE_SIZE);
2550 
2551 	kmem_free(kernel_map, base, size);
2552 }
2553 
2554 static void
2555 moea64_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2556 {
2557 	struct pvo_entry *pvo;
2558 	vm_offset_t lim;
2559 	vm_paddr_t pa;
2560 	vm_size_t len;
2561 
2562 	PMAP_LOCK(pm);
2563 	while (sz > 0) {
2564 		lim = round_page(va);
2565 		len = MIN(lim - va, sz);
2566 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2567 		if (pvo != NULL) {
2568 			pa = (pvo->pvo_pte.pte.pte_lo & LPTE_RPGN) |
2569 			    (va & ADDR_POFF);
2570 			moea64_syncicache(pm, va, pa, len);
2571 		}
2572 		va += len;
2573 		sz -= len;
2574 	}
2575 	PMAP_UNLOCK(pm);
2576 }
2577