xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision eb6d21b4ca6d668cf89afd99eef7baeafa712197)
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	TLBSYNC()	__asm __volatile("tlbsync; ptesync");
176 #define	SYNC()		__asm __volatile("sync");
177 #define	EIEIO()		__asm __volatile("eieio");
178 
179 /*
180  * The tlbie instruction must be executed in 64-bit mode
181  * so we have to twiddle MSR[SF] around every invocation.
182  * Just to add to the fun, exceptions must be off as well
183  * so that we can't trap in 64-bit mode. What a pain.
184  */
185 struct mtx	tlbie_mutex;
186 
187 static __inline void
188 TLBIE(pmap_t pmap, vm_offset_t va) {
189 	uint64_t vpn;
190 	register_t vpn_hi, vpn_lo;
191 	register_t msr;
192 	register_t scratch;
193 
194 	vpn = (uint64_t)(va & ADDR_PIDX);
195 	if (pmap != NULL)
196 		vpn |= (va_to_vsid(pmap,va) << 28);
197 
198 	vpn_hi = (uint32_t)(vpn >> 32);
199 	vpn_lo = (uint32_t)vpn;
200 
201 	mtx_lock_spin(&tlbie_mutex);
202 	__asm __volatile("\
203 	    mfmsr %0; \
204 	    clrldi %1,%0,49; \
205 	    mtmsr %1; \
206 	    insrdi %1,%5,1,0; \
207 	    mtmsrd %1; \
208 	    ptesync; \
209 	    \
210 	    sld %1,%2,%4; \
211 	    or %1,%1,%3; \
212 	    tlbie %1; \
213 	    \
214 	    mtmsrd %0; \
215 	    eieio; \
216 	    tlbsync; \
217 	    ptesync;"
218 	: "=r"(msr), "=r"(scratch) : "r"(vpn_hi), "r"(vpn_lo), "r"(32), "r"(1));
219 	mtx_unlock_spin(&tlbie_mutex);
220 }
221 
222 #define DISABLE_TRANS(msr)	msr = mfmsr(); mtmsr(msr & ~PSL_DR); isync()
223 #define ENABLE_TRANS(msr)	mtmsr(msr); isync()
224 
225 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
226 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
227 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
228 
229 #define	PVO_PTEGIDX_MASK	0x007		/* which PTEG slot */
230 #define	PVO_PTEGIDX_VALID	0x008		/* slot is valid */
231 #define	PVO_WIRED		0x010		/* PVO entry is wired */
232 #define	PVO_MANAGED		0x020		/* PVO entry is managed */
233 #define	PVO_BOOTSTRAP		0x080		/* PVO entry allocated during
234 						   bootstrap */
235 #define PVO_FAKE		0x100		/* fictitious phys page */
236 #define	PVO_VADDR(pvo)		((pvo)->pvo_vaddr & ~ADDR_POFF)
237 #define PVO_ISFAKE(pvo)		((pvo)->pvo_vaddr & PVO_FAKE)
238 #define	PVO_PTEGIDX_GET(pvo)	((pvo)->pvo_vaddr & PVO_PTEGIDX_MASK)
239 #define	PVO_PTEGIDX_ISSET(pvo)	((pvo)->pvo_vaddr & PVO_PTEGIDX_VALID)
240 #define	PVO_PTEGIDX_CLR(pvo)	\
241 	((void)((pvo)->pvo_vaddr &= ~(PVO_PTEGIDX_VALID|PVO_PTEGIDX_MASK)))
242 #define	PVO_PTEGIDX_SET(pvo, i)	\
243 	((void)((pvo)->pvo_vaddr |= (i)|PVO_PTEGIDX_VALID))
244 
245 #define	MOEA_PVO_CHECK(pvo)
246 
247 #define LOCK_TABLE() mtx_lock(&moea64_table_mutex)
248 #define UNLOCK_TABLE() mtx_unlock(&moea64_table_mutex);
249 #define ASSERT_TABLE_LOCK() mtx_assert(&moea64_table_mutex, MA_OWNED)
250 
251 struct ofw_map {
252 	vm_offset_t	om_va;
253 	vm_size_t	om_len;
254 	vm_offset_t	om_pa_hi;
255 	vm_offset_t	om_pa_lo;
256 	u_int		om_mode;
257 };
258 
259 /*
260  * Map of physical memory regions.
261  */
262 static struct	mem_region *regions;
263 static struct	mem_region *pregions;
264 extern u_int	phys_avail_count;
265 extern int	regions_sz, pregions_sz;
266 extern int	ofw_real_mode;
267 
268 extern struct pmap ofw_pmap;
269 
270 extern void bs_remap_earlyboot(void);
271 
272 
273 /*
274  * Lock for the pteg and pvo tables.
275  */
276 struct mtx	moea64_table_mutex;
277 
278 /*
279  * PTEG data.
280  */
281 static struct	lpteg *moea64_pteg_table;
282 u_int		moea64_pteg_count;
283 u_int		moea64_pteg_mask;
284 
285 /*
286  * PVO data.
287  */
288 struct	pvo_head *moea64_pvo_table;		/* pvo entries by pteg index */
289 /* lists of unmanaged pages */
290 struct	pvo_head moea64_pvo_kunmanaged =
291     LIST_HEAD_INITIALIZER(moea64_pvo_kunmanaged);
292 struct	pvo_head moea64_pvo_unmanaged =
293     LIST_HEAD_INITIALIZER(moea64_pvo_unmanaged);
294 
295 uma_zone_t	moea64_upvo_zone; /* zone for pvo entries for unmanaged pages */
296 uma_zone_t	moea64_mpvo_zone; /* zone for pvo entries for managed pages */
297 
298 vm_offset_t	pvo_allocator_start;
299 vm_offset_t	pvo_allocator_end;
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	pvo_entry *moea64_scratchpage_pvo[2];
331 struct	lpte 	*moea64_scratchpage_pte[2];
332 struct	mtx	moea64_scratchpage_mtx;
333 
334 /*
335  * Allocate physical memory for use in moea64_bootstrap.
336  */
337 static vm_offset_t	moea64_bootstrap_alloc(vm_size_t, u_int);
338 
339 /*
340  * PTE calls.
341  */
342 static int		moea64_pte_insert(u_int, struct lpte *);
343 
344 /*
345  * PVO calls.
346  */
347 static int	moea64_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *,
348 		    vm_offset_t, vm_offset_t, uint64_t, int);
349 static void	moea64_pvo_remove(struct pvo_entry *, int);
350 static struct	pvo_entry *moea64_pvo_find_va(pmap_t, vm_offset_t, int *);
351 static struct	lpte *moea64_pvo_to_pte(const struct pvo_entry *, int);
352 
353 /*
354  * Utility routines.
355  */
356 static void		moea64_bridge_bootstrap(mmu_t mmup,
357 			    vm_offset_t kernelstart, vm_offset_t kernelend);
358 static void		moea64_bridge_cpu_bootstrap(mmu_t, int ap);
359 static void		moea64_enter_locked(pmap_t, vm_offset_t, vm_page_t,
360 			    vm_prot_t, boolean_t);
361 static boolean_t	moea64_query_bit(vm_page_t, u_int64_t);
362 static u_int		moea64_clear_bit(vm_page_t, u_int64_t, u_int64_t *);
363 static void		moea64_kremove(mmu_t, vm_offset_t);
364 static void		moea64_syncicache(pmap_t pmap, vm_offset_t va,
365 			    vm_offset_t pa, vm_size_t sz);
366 static void		tlbia(void);
367 
368 /*
369  * Kernel MMU interface
370  */
371 void moea64_change_wiring(mmu_t, pmap_t, vm_offset_t, boolean_t);
372 void moea64_clear_modify(mmu_t, vm_page_t);
373 void moea64_clear_reference(mmu_t, vm_page_t);
374 void moea64_copy_page(mmu_t, vm_page_t, vm_page_t);
375 void moea64_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, boolean_t);
376 void moea64_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
377     vm_prot_t);
378 void moea64_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
379 vm_paddr_t moea64_extract(mmu_t, pmap_t, vm_offset_t);
380 vm_page_t moea64_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
381 void moea64_init(mmu_t);
382 boolean_t moea64_is_modified(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_ts_referenced,	moea64_ts_referenced),
421 	MMUMETHOD(mmu_map,     		moea64_map),
422 	MMUMETHOD(mmu_page_exists_quick,moea64_page_exists_quick),
423 	MMUMETHOD(mmu_page_wired_mappings,moea64_page_wired_mappings),
424 	MMUMETHOD(mmu_pinit,		moea64_pinit),
425 	MMUMETHOD(mmu_pinit0,		moea64_pinit0),
426 	MMUMETHOD(mmu_protect,		moea64_protect),
427 	MMUMETHOD(mmu_qenter,		moea64_qenter),
428 	MMUMETHOD(mmu_qremove,		moea64_qremove),
429 	MMUMETHOD(mmu_release,		moea64_release),
430 	MMUMETHOD(mmu_remove,		moea64_remove),
431 	MMUMETHOD(mmu_remove_all,      	moea64_remove_all),
432 	MMUMETHOD(mmu_remove_write,	moea64_remove_write),
433 	MMUMETHOD(mmu_sync_icache,	moea64_sync_icache),
434 	MMUMETHOD(mmu_zero_page,       	moea64_zero_page),
435 	MMUMETHOD(mmu_zero_page_area,	moea64_zero_page_area),
436 	MMUMETHOD(mmu_zero_page_idle,	moea64_zero_page_idle),
437 	MMUMETHOD(mmu_activate,		moea64_activate),
438 	MMUMETHOD(mmu_deactivate,      	moea64_deactivate),
439 
440 	/* Internal interfaces */
441 	MMUMETHOD(mmu_bootstrap,       	moea64_bridge_bootstrap),
442 	MMUMETHOD(mmu_cpu_bootstrap,   	moea64_bridge_cpu_bootstrap),
443 	MMUMETHOD(mmu_mapdev,		moea64_mapdev),
444 	MMUMETHOD(mmu_unmapdev,		moea64_unmapdev),
445 	MMUMETHOD(mmu_kextract,		moea64_kextract),
446 	MMUMETHOD(mmu_kenter,		moea64_kenter),
447 	MMUMETHOD(mmu_dev_direct_mapped,moea64_dev_direct_mapped),
448 
449 	{ 0, 0 }
450 };
451 
452 static mmu_def_t oea64_bridge_mmu = {
453 	MMU_TYPE_G5,
454 	moea64_bridge_methods,
455 	0
456 };
457 MMU_DEF(oea64_bridge_mmu);
458 
459 static __inline u_int
460 va_to_pteg(uint64_t vsid, vm_offset_t addr)
461 {
462 	u_int hash;
463 
464 	hash = vsid ^ (((uint64_t)addr & ADDR_PIDX) >>
465 	    ADDR_PIDX_SHFT);
466 	return (hash & moea64_pteg_mask);
467 }
468 
469 static __inline struct pvo_head *
470 pa_to_pvoh(vm_offset_t pa, vm_page_t *pg_p)
471 {
472 	struct	vm_page *pg;
473 
474 	pg = PHYS_TO_VM_PAGE(pa);
475 
476 	if (pg_p != NULL)
477 		*pg_p = pg;
478 
479 	if (pg == NULL)
480 		return (&moea64_pvo_unmanaged);
481 
482 	return (&pg->md.mdpg_pvoh);
483 }
484 
485 static __inline struct pvo_head *
486 vm_page_to_pvoh(vm_page_t m)
487 {
488 
489 	return (&m->md.mdpg_pvoh);
490 }
491 
492 static __inline void
493 moea64_attr_clear(vm_page_t m, u_int64_t ptebit)
494 {
495 
496 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
497 	m->md.mdpg_attrs &= ~ptebit;
498 }
499 
500 static __inline u_int64_t
501 moea64_attr_fetch(vm_page_t m)
502 {
503 
504 	return (m->md.mdpg_attrs);
505 }
506 
507 static __inline void
508 moea64_attr_save(vm_page_t m, u_int64_t ptebit)
509 {
510 
511 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
512 	m->md.mdpg_attrs |= ptebit;
513 }
514 
515 static __inline int
516 moea64_pte_compare(const struct lpte *pt, const struct lpte *pvo_pt)
517 {
518 	if (pt->pte_hi == pvo_pt->pte_hi)
519 		return (1);
520 
521 	return (0);
522 }
523 
524 static __inline int
525 moea64_pte_match(struct lpte *pt, uint64_t vsid, vm_offset_t va, int which)
526 {
527 	return (pt->pte_hi & ~LPTE_VALID) ==
528 	    ((vsid << LPTE_VSID_SHIFT) |
529 	    ((uint64_t)(va >> ADDR_API_SHFT64) & LPTE_API) | which);
530 }
531 
532 static __inline void
533 moea64_pte_create(struct lpte *pt, uint64_t vsid, vm_offset_t va,
534     uint64_t pte_lo)
535 {
536 	ASSERT_TABLE_LOCK();
537 
538 	/*
539 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
540 	 * set when the real pte is set in memory.
541 	 *
542 	 * Note: Don't set the valid bit for correct operation of tlb update.
543 	 */
544 	pt->pte_hi = (vsid << LPTE_VSID_SHIFT) |
545 	    (((uint64_t)(va & ADDR_PIDX) >> ADDR_API_SHFT64) & LPTE_API);
546 
547 	pt->pte_lo = pte_lo;
548 }
549 
550 static __inline void
551 moea64_pte_synch(struct lpte *pt, struct lpte *pvo_pt)
552 {
553 
554 	ASSERT_TABLE_LOCK();
555 
556 	pvo_pt->pte_lo |= pt->pte_lo & (LPTE_REF | LPTE_CHG);
557 }
558 
559 static __inline void
560 moea64_pte_clear(struct lpte *pt, pmap_t pmap, vm_offset_t va, u_int64_t ptebit)
561 {
562 	ASSERT_TABLE_LOCK();
563 
564 	/*
565 	 * As shown in Section 7.6.3.2.3
566 	 */
567 	pt->pte_lo &= ~ptebit;
568 	TLBIE(pmap,va);
569 }
570 
571 static __inline void
572 moea64_pte_set(struct lpte *pt, struct lpte *pvo_pt)
573 {
574 
575 	ASSERT_TABLE_LOCK();
576 	pvo_pt->pte_hi |= LPTE_VALID;
577 
578 	/*
579 	 * Update the PTE as defined in section 7.6.3.1.
580 	 * Note that the REF/CHG bits are from pvo_pt and thus should have
581 	 * been saved so this routine can restore them (if desired).
582 	 */
583 	pt->pte_lo = pvo_pt->pte_lo;
584 	EIEIO();
585 	pt->pte_hi = pvo_pt->pte_hi;
586 	SYNC();
587 	moea64_pte_valid++;
588 }
589 
590 static __inline void
591 moea64_pte_unset(struct lpte *pt, struct lpte *pvo_pt, pmap_t pmap, vm_offset_t va)
592 {
593 	ASSERT_TABLE_LOCK();
594 	pvo_pt->pte_hi &= ~LPTE_VALID;
595 
596 	/*
597 	 * Force the reg & chg bits back into the PTEs.
598 	 */
599 	SYNC();
600 
601 	/*
602 	 * Invalidate the pte.
603 	 */
604 	pt->pte_hi &= ~LPTE_VALID;
605 
606 	TLBIE(pmap,va);
607 
608 	/*
609 	 * Save the reg & chg bits.
610 	 */
611 	moea64_pte_synch(pt, pvo_pt);
612 	moea64_pte_valid--;
613 }
614 
615 static __inline void
616 moea64_pte_change(struct lpte *pt, struct lpte *pvo_pt, pmap_t pmap, vm_offset_t va)
617 {
618 
619 	/*
620 	 * Invalidate the PTE
621 	 */
622 	moea64_pte_unset(pt, pvo_pt, pmap, va);
623 	moea64_pte_set(pt, pvo_pt);
624 }
625 
626 static __inline uint64_t
627 moea64_calc_wimg(vm_offset_t pa)
628 {
629 	uint64_t pte_lo;
630 	int i;
631 
632 	/*
633 	 * Assume the page is cache inhibited and access is guarded unless
634 	 * it's in our available memory array.
635 	 */
636 	pte_lo = LPTE_I | LPTE_G;
637 	for (i = 0; i < pregions_sz; i++) {
638 		if ((pa >= pregions[i].mr_start) &&
639 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
640 			pte_lo &= ~(LPTE_I | LPTE_G);
641 			pte_lo |= LPTE_M;
642 			break;
643 		}
644 	}
645 
646 	return pte_lo;
647 }
648 
649 /*
650  * Quick sort callout for comparing memory regions.
651  */
652 static int	mr_cmp(const void *a, const void *b);
653 static int	om_cmp(const void *a, const void *b);
654 
655 static int
656 mr_cmp(const void *a, const void *b)
657 {
658 	const struct	mem_region *regiona;
659 	const struct	mem_region *regionb;
660 
661 	regiona = a;
662 	regionb = b;
663 	if (regiona->mr_start < regionb->mr_start)
664 		return (-1);
665 	else if (regiona->mr_start > regionb->mr_start)
666 		return (1);
667 	else
668 		return (0);
669 }
670 
671 static int
672 om_cmp(const void *a, const void *b)
673 {
674 	const struct	ofw_map *mapa;
675 	const struct	ofw_map *mapb;
676 
677 	mapa = a;
678 	mapb = b;
679 	if (mapa->om_pa_hi < mapb->om_pa_hi)
680 		return (-1);
681 	else if (mapa->om_pa_hi > mapb->om_pa_hi)
682 		return (1);
683 	else if (mapa->om_pa_lo < mapb->om_pa_lo)
684 		return (-1);
685 	else if (mapa->om_pa_lo > mapb->om_pa_lo)
686 		return (1);
687 	else
688 		return (0);
689 }
690 
691 static void
692 moea64_bridge_cpu_bootstrap(mmu_t mmup, int ap)
693 {
694 	int i = 0;
695 
696 	/*
697 	 * Initialize segment registers and MMU
698 	 */
699 
700 	mtmsr(mfmsr() & ~PSL_DR & ~PSL_IR); isync();
701 	for (i = 0; i < 16; i++) {
702 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
703 	}
704 	__asm __volatile ("sync; mtsdr1 %0; isync"
705 	    :: "r"((u_int)moea64_pteg_table
706 		     | (32 - cntlzw(moea64_pteg_mask >> 11))));
707 	tlbia();
708 }
709 
710 static void
711 moea64_add_ofw_mappings(mmu_t mmup, phandle_t mmu, size_t sz)
712 {
713 	struct ofw_map	translations[sz/sizeof(struct ofw_map)];
714 	register_t	msr;
715 	vm_offset_t	off;
716 	int		i, ofw_mappings;
717 
718 	bzero(translations, sz);
719 	if (OF_getprop(mmu, "translations", translations, sz) == -1)
720 		panic("moea64_bootstrap: can't get ofw translations");
721 
722 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
723 	sz /= sizeof(*translations);
724 	qsort(translations, sz, sizeof (*translations), om_cmp);
725 
726 	for (i = 0, ofw_mappings = 0; i < sz; i++) {
727 		CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
728 		    (uint32_t)(translations[i].om_pa_lo), translations[i].om_va,
729 		    translations[i].om_len);
730 
731 		if (translations[i].om_pa_lo % PAGE_SIZE)
732 			panic("OFW translation not page-aligned!");
733 
734 		if (translations[i].om_pa_hi)
735 			panic("OFW translations above 32-bit boundary!");
736 
737 		/* Now enter the pages for this mapping */
738 
739 		/*
740 		 * Lock the ofw pmap. pmap_kenter(), which we use for the
741 		 * pages the kernel also needs, does its own locking.
742 		 */
743 		PMAP_LOCK(&ofw_pmap);
744 		DISABLE_TRANS(msr);
745 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
746 			struct vm_page m;
747 
748 			/* Map low memory mappings into the kernel pmap, too.
749 			 * These are typically mappings made by the loader,
750 			 * so we need them if we want to keep executing. */
751 
752 			if (translations[i].om_va + off < SEGMENT_LENGTH)
753 				moea64_kenter(mmup, translations[i].om_va + off,
754 				    translations[i].om_va + off);
755 
756 			m.phys_addr = translations[i].om_pa_lo + off;
757 			moea64_enter_locked(&ofw_pmap,
758 			    translations[i].om_va + off, &m, VM_PROT_ALL, 1);
759 
760 			ofw_mappings++;
761 		}
762 		ENABLE_TRANS(msr);
763 		PMAP_UNLOCK(&ofw_pmap);
764 	}
765 }
766 
767 static void
768 moea64_bridge_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
769 {
770 	ihandle_t	mmui;
771 	phandle_t	chosen;
772 	phandle_t	mmu;
773 	size_t		sz;
774 	int		i, j;
775 	vm_size_t	size, physsz, hwphyssz;
776 	vm_offset_t	pa, va, off;
777 	register_t	msr;
778 	void		*dpcpu;
779 
780 	/* We don't have a direct map since there is no BAT */
781 	hw_direct_map = 0;
782 
783 	/* Make sure battable is zero, since we have no BAT */
784 	for (i = 0; i < 16; i++) {
785 		battable[i].batu = 0;
786 		battable[i].batl = 0;
787 	}
788 
789 	/* Get physical memory regions from firmware */
790 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
791 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
792 
793 	qsort(pregions, pregions_sz, sizeof(*pregions), mr_cmp);
794 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
795 		panic("moea64_bootstrap: phys_avail too small");
796 	qsort(regions, regions_sz, sizeof(*regions), mr_cmp);
797 	phys_avail_count = 0;
798 	physsz = 0;
799 	hwphyssz = 0;
800 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
801 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
802 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
803 		    regions[i].mr_start + regions[i].mr_size,
804 		    regions[i].mr_size);
805 		if (hwphyssz != 0 &&
806 		    (physsz + regions[i].mr_size) >= hwphyssz) {
807 			if (physsz < hwphyssz) {
808 				phys_avail[j] = regions[i].mr_start;
809 				phys_avail[j + 1] = regions[i].mr_start +
810 				    hwphyssz - physsz;
811 				physsz = hwphyssz;
812 				phys_avail_count++;
813 			}
814 			break;
815 		}
816 		phys_avail[j] = regions[i].mr_start;
817 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
818 		phys_avail_count++;
819 		physsz += regions[i].mr_size;
820 	}
821 	physmem = btoc(physsz);
822 
823 	/*
824 	 * Allocate PTEG table.
825 	 */
826 #ifdef PTEGCOUNT
827 	moea64_pteg_count = PTEGCOUNT;
828 #else
829 	moea64_pteg_count = 0x1000;
830 
831 	while (moea64_pteg_count < physmem)
832 		moea64_pteg_count <<= 1;
833 #endif /* PTEGCOUNT */
834 
835 	size = moea64_pteg_count * sizeof(struct lpteg);
836 	CTR2(KTR_PMAP, "moea64_bootstrap: %d PTEGs, %d bytes",
837 	    moea64_pteg_count, size);
838 
839 	/*
840 	 * We now need to allocate memory. This memory, to be allocated,
841 	 * has to reside in a page table. The page table we are about to
842 	 * allocate. We don't have BAT. So drop to data real mode for a minute
843 	 * as a measure of last resort. We do this a couple times.
844 	 */
845 
846 	moea64_pteg_table = (struct lpteg *)moea64_bootstrap_alloc(size, size);
847 	DISABLE_TRANS(msr);
848 	bzero((void *)moea64_pteg_table, moea64_pteg_count * sizeof(struct lpteg));
849 	ENABLE_TRANS(msr);
850 
851 	moea64_pteg_mask = moea64_pteg_count - 1;
852 
853 	CTR1(KTR_PMAP, "moea64_bootstrap: PTEG table at %p", moea64_pteg_table);
854 
855 	/*
856 	 * Allocate pv/overflow lists.
857 	 */
858 	size = sizeof(struct pvo_head) * moea64_pteg_count;
859 
860 	moea64_pvo_table = (struct pvo_head *)moea64_bootstrap_alloc(size,
861 	    PAGE_SIZE);
862 	CTR1(KTR_PMAP, "moea64_bootstrap: PVO table at %p", moea64_pvo_table);
863 
864 	DISABLE_TRANS(msr);
865 	for (i = 0; i < moea64_pteg_count; i++)
866 		LIST_INIT(&moea64_pvo_table[i]);
867 	ENABLE_TRANS(msr);
868 
869 	/*
870 	 * Initialize the lock that synchronizes access to the pteg and pvo
871 	 * tables.
872 	 */
873 	mtx_init(&moea64_table_mutex, "pmap table", NULL, MTX_DEF |
874 	    MTX_RECURSE);
875 
876 	/*
877 	 * Initialize the TLBIE lock. TLBIE can only be executed by one CPU.
878 	 */
879 	mtx_init(&tlbie_mutex, "tlbie mutex", NULL, MTX_SPIN);
880 
881 	/*
882 	 * Initialise the unmanaged pvo pool.
883 	 */
884 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
885 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
886 	moea64_bpvo_pool_index = 0;
887 
888 	/*
889 	 * Make sure kernel vsid is allocated as well as VSID 0.
890 	 */
891 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
892 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
893 	moea64_vsid_bitmap[0] |= 1;
894 
895 	/*
896 	 * Initialize the kernel pmap (which is statically allocated).
897 	 */
898 	for (i = 0; i < 16; i++)
899 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
900 
901 	kernel_pmap->pmap_phys = kernel_pmap;
902 	kernel_pmap->pm_active = ~0;
903 
904 	PMAP_LOCK_INIT(kernel_pmap);
905 
906 	/*
907 	 * Now map in all the other buffers we allocated earlier
908 	 */
909 
910 	DISABLE_TRANS(msr);
911 	size = moea64_pteg_count * sizeof(struct lpteg);
912 	off = (vm_offset_t)(moea64_pteg_table);
913 	for (pa = off; pa < off + size; pa += PAGE_SIZE)
914 		moea64_kenter(mmup, pa, pa);
915 	size = sizeof(struct pvo_head) * moea64_pteg_count;
916 	off = (vm_offset_t)(moea64_pvo_table);
917 	for (pa = off; pa < off + size; pa += PAGE_SIZE)
918 		moea64_kenter(mmup, pa, pa);
919 	size = BPVO_POOL_SIZE*sizeof(struct pvo_entry);
920 	off = (vm_offset_t)(moea64_bpvo_pool);
921 	for (pa = off; pa < off + size; pa += PAGE_SIZE)
922 		moea64_kenter(mmup, pa, pa);
923 
924 	/*
925 	 * Map certain important things, like ourselves.
926 	 *
927 	 * NOTE: We do not map the exception vector space. That code is
928 	 * used only in real mode, and leaving it unmapped allows us to
929 	 * catch NULL pointer deferences, instead of making NULL a valid
930 	 * address.
931 	 */
932 
933 	for (pa = kernelstart & ~PAGE_MASK; pa < kernelend; pa += PAGE_SIZE)
934 		moea64_kenter(mmup, pa, pa);
935 	ENABLE_TRANS(msr);
936 
937 	if (!ofw_real_mode) {
938 	    /*
939 	     * Set up the Open Firmware pmap and add its mappings.
940 	     */
941 
942 	    moea64_pinit(mmup, &ofw_pmap);
943 	    ofw_pmap.pm_sr[KERNEL_SR] = kernel_pmap->pm_sr[KERNEL_SR];
944 	    ofw_pmap.pm_sr[KERNEL2_SR] = kernel_pmap->pm_sr[KERNEL2_SR];
945 
946 	    if ((chosen = OF_finddevice("/chosen")) == -1)
947 		panic("moea64_bootstrap: can't find /chosen");
948 	    OF_getprop(chosen, "mmu", &mmui, 4);
949 	    if ((mmu = OF_instance_to_package(mmui)) == -1)
950 		panic("moea64_bootstrap: can't get mmu package");
951 	    if ((sz = OF_getproplen(mmu, "translations")) == -1)
952 		panic("moea64_bootstrap: can't get ofw translation count");
953 	    if (sz > 6144 /* tmpstksz - 2 KB headroom */)
954 		panic("moea64_bootstrap: too many ofw translations");
955 
956 	    moea64_add_ofw_mappings(mmup, mmu, sz);
957 	}
958 
959 #ifdef SMP
960 	TLBSYNC();
961 #endif
962 
963 	/*
964 	 * Calculate the last available physical address.
965 	 */
966 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
967 		;
968 	Maxmem = powerpc_btop(phys_avail[i + 1]);
969 
970 	/*
971 	 * Initialize MMU and remap early physical mappings
972 	 */
973 	moea64_bridge_cpu_bootstrap(mmup,0);
974 	mtmsr(mfmsr() | PSL_DR | PSL_IR); isync();
975 	pmap_bootstrapped++;
976 	bs_remap_earlyboot();
977 
978 	/*
979 	 * Set the start and end of kva.
980 	 */
981 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
982 	virtual_end = VM_MAX_KERNEL_ADDRESS;
983 
984 	/*
985 	 * Allocate some stupid buffer regions.
986 	 */
987 
988 	pvo_allocator_start = virtual_avail;
989 	virtual_avail += SEGMENT_LENGTH/4;
990 	pvo_allocator_end = virtual_avail;
991 
992 	/*
993 	 * Allocate some things for page zeroing
994 	 */
995 
996 	mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL, MTX_DEF);
997 	for (i = 0; i < 2; i++) {
998 		moea64_scratchpage_va[i] = virtual_avail;
999 		virtual_avail += PAGE_SIZE;
1000 
1001 		moea64_kenter(mmup,moea64_scratchpage_va[i],kernelstart);
1002 
1003 		LOCK_TABLE();
1004 		moea64_scratchpage_pvo[i] = moea64_pvo_find_va(kernel_pmap,
1005 		    moea64_scratchpage_va[i],&j);
1006 		moea64_scratchpage_pte[i] = moea64_pvo_to_pte(
1007 		    moea64_scratchpage_pvo[i],j);
1008 		UNLOCK_TABLE();
1009 	}
1010 
1011 	/*
1012 	 * Allocate a kernel stack with a guard page for thread0 and map it
1013 	 * into the kernel page map.
1014 	 */
1015 	pa = moea64_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, PAGE_SIZE);
1016 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
1017 	virtual_avail = va + KSTACK_PAGES * PAGE_SIZE;
1018 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
1019 	thread0.td_kstack = va;
1020 	thread0.td_kstack_pages = KSTACK_PAGES;
1021 	for (i = 0; i < KSTACK_PAGES; i++) {
1022 		moea64_kenter(mmup, va, pa);;
1023 		pa += PAGE_SIZE;
1024 		va += PAGE_SIZE;
1025 	}
1026 
1027 	/*
1028 	 * Allocate virtual address space for the message buffer.
1029 	 */
1030 	pa = msgbuf_phys = moea64_bootstrap_alloc(MSGBUF_SIZE, PAGE_SIZE);
1031 	msgbufp = (struct msgbuf *)virtual_avail;
1032 	va = virtual_avail;
1033 	virtual_avail += round_page(MSGBUF_SIZE);
1034 	while (va < virtual_avail) {
1035 		moea64_kenter(mmup, va, pa);;
1036 		pa += PAGE_SIZE;
1037 		va += PAGE_SIZE;
1038 	}
1039 
1040 	/*
1041 	 * Allocate virtual address space for the dynamic percpu area.
1042 	 */
1043 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
1044 	dpcpu = (void *)virtual_avail;
1045 	va = virtual_avail;
1046 	virtual_avail += DPCPU_SIZE;
1047 	while (va < virtual_avail) {
1048 		moea64_kenter(mmup, va, pa);;
1049 		pa += PAGE_SIZE;
1050 		va += PAGE_SIZE;
1051 	}
1052 	dpcpu_init(dpcpu, 0);
1053 }
1054 
1055 /*
1056  * Activate a user pmap.  The pmap must be activated before it's address
1057  * space can be accessed in any way.
1058  */
1059 void
1060 moea64_activate(mmu_t mmu, struct thread *td)
1061 {
1062 	pmap_t	pm, pmr;
1063 
1064 	/*
1065 	 * Load all the data we need up front to encourage the compiler to
1066 	 * not issue any loads while we have interrupts disabled below.
1067 	 */
1068 	pm = &td->td_proc->p_vmspace->vm_pmap;
1069 	pmr = pm->pmap_phys;
1070 
1071 	pm->pm_active |= PCPU_GET(cpumask);
1072 	PCPU_SET(curpmap, pmr);
1073 }
1074 
1075 void
1076 moea64_deactivate(mmu_t mmu, struct thread *td)
1077 {
1078 	pmap_t	pm;
1079 
1080 	pm = &td->td_proc->p_vmspace->vm_pmap;
1081 	pm->pm_active &= ~(PCPU_GET(cpumask));
1082 	PCPU_SET(curpmap, NULL);
1083 }
1084 
1085 void
1086 moea64_change_wiring(mmu_t mmu, pmap_t pm, vm_offset_t va, boolean_t wired)
1087 {
1088 	struct	pvo_entry *pvo;
1089 
1090 	PMAP_LOCK(pm);
1091 	pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1092 
1093 	if (pvo != NULL) {
1094 		if (wired) {
1095 			if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1096 				pm->pm_stats.wired_count++;
1097 			pvo->pvo_vaddr |= PVO_WIRED;
1098 		} else {
1099 			if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1100 				pm->pm_stats.wired_count--;
1101 			pvo->pvo_vaddr &= ~PVO_WIRED;
1102 		}
1103 	}
1104 	PMAP_UNLOCK(pm);
1105 }
1106 
1107 /*
1108  * Zero a page of physical memory by temporarily mapping it into the tlb.
1109  */
1110 void
1111 moea64_zero_page(mmu_t mmu, vm_page_t m)
1112 {
1113 	moea64_zero_page_area(mmu,m,0,PAGE_SIZE);
1114 }
1115 
1116 /*
1117  * This goes through and sets the physical address of our
1118  * special scratch PTE to the PA we want to zero or copy. Because
1119  * of locking issues (this can get called in pvo_enter() by
1120  * the UMA allocator), we can't use most other utility functions here
1121  */
1122 
1123 static __inline
1124 void moea64_set_scratchpage_pa(int which, vm_offset_t pa) {
1125 	moea64_scratchpage_pvo[which]->pvo_pte.lpte.pte_lo &=
1126 	    (~LPTE_WIMG & ~LPTE_RPGN);
1127 	moea64_scratchpage_pvo[which]->pvo_pte.lpte.pte_lo |=
1128 	    moea64_calc_wimg(pa) | (uint64_t)pa;
1129 
1130 	moea64_scratchpage_pte[which]->pte_hi &= ~LPTE_VALID;
1131 	TLBIE(kernel_pmap, moea64_scratchpage_va[which]);
1132 
1133 	moea64_scratchpage_pte[which]->pte_lo =
1134 	    moea64_scratchpage_pvo[which]->pvo_pte.lpte.pte_lo;
1135 	EIEIO();
1136 
1137 	moea64_scratchpage_pte[which]->pte_hi |= LPTE_VALID;
1138 	TLBIE(kernel_pmap, moea64_scratchpage_va[which]);
1139 }
1140 
1141 void
1142 moea64_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1143 {
1144 	vm_offset_t	dst;
1145 	vm_offset_t	src;
1146 
1147 	dst = VM_PAGE_TO_PHYS(mdst);
1148 	src = VM_PAGE_TO_PHYS(msrc);
1149 
1150 	mtx_lock(&moea64_scratchpage_mtx);
1151 
1152 	moea64_set_scratchpage_pa(0,src);
1153 	moea64_set_scratchpage_pa(1,dst);
1154 
1155 	kcopy((void *)moea64_scratchpage_va[0],
1156 	    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1157 
1158 	__syncicache((void *)moea64_scratchpage_va[1],PAGE_SIZE);
1159 
1160 	mtx_unlock(&moea64_scratchpage_mtx);
1161 }
1162 
1163 void
1164 moea64_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1165 {
1166 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1167 
1168 	if (!moea64_initialized)
1169 		panic("moea64_zero_page: can't zero pa %#x", pa);
1170 	if (size + off > PAGE_SIZE)
1171 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1172 
1173 	mtx_lock(&moea64_scratchpage_mtx);
1174 
1175 	moea64_set_scratchpage_pa(0,pa);
1176 	bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1177 	__syncicache((void *)moea64_scratchpage_va[0],PAGE_SIZE);
1178 
1179 	mtx_unlock(&moea64_scratchpage_mtx);
1180 }
1181 
1182 void
1183 moea64_zero_page_idle(mmu_t mmu, vm_page_t m)
1184 {
1185 
1186 	moea64_zero_page(mmu, m);
1187 }
1188 
1189 /*
1190  * Map the given physical page at the specified virtual address in the
1191  * target pmap with the protection requested.  If specified the page
1192  * will be wired down.
1193  */
1194 void
1195 moea64_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m,
1196     vm_prot_t prot, boolean_t wired)
1197 {
1198 
1199 	vm_page_lock_queues();
1200 	PMAP_LOCK(pmap);
1201 	moea64_enter_locked(pmap, va, m, prot, wired);
1202 	vm_page_unlock_queues();
1203 	PMAP_UNLOCK(pmap);
1204 }
1205 
1206 /*
1207  * Map the given physical page at the specified virtual address in the
1208  * target pmap with the protection requested.  If specified the page
1209  * will be wired down.
1210  *
1211  * The page queues and pmap must be locked.
1212  */
1213 
1214 static void
1215 moea64_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1216     boolean_t wired)
1217 {
1218 	struct		pvo_head *pvo_head;
1219 	uma_zone_t	zone;
1220 	vm_page_t	pg;
1221 	uint64_t	pte_lo;
1222 	u_int		pvo_flags;
1223 	int		error;
1224 
1225 	if (!moea64_initialized) {
1226 		pvo_head = &moea64_pvo_kunmanaged;
1227 		pg = NULL;
1228 		zone = moea64_upvo_zone;
1229 		pvo_flags = 0;
1230 	} else {
1231 		pvo_head = vm_page_to_pvoh(m);
1232 		pg = m;
1233 		zone = moea64_mpvo_zone;
1234 		pvo_flags = PVO_MANAGED;
1235 	}
1236 
1237 	if (pmap_bootstrapped)
1238 		mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1239 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1240 
1241 	/* XXX change the pvo head for fake pages */
1242 	if ((m->flags & PG_FICTITIOUS) == PG_FICTITIOUS) {
1243 		pvo_flags &= ~PVO_MANAGED;
1244 		pvo_head = &moea64_pvo_kunmanaged;
1245 		zone = moea64_upvo_zone;
1246 	}
1247 
1248 	pte_lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m));
1249 
1250 	if (prot & VM_PROT_WRITE) {
1251 		pte_lo |= LPTE_BW;
1252 		if (pmap_bootstrapped)
1253 			vm_page_flag_set(m, PG_WRITEABLE);
1254 	} else
1255 		pte_lo |= LPTE_BR;
1256 
1257 	if (prot & VM_PROT_EXECUTE)
1258 		pvo_flags |= VM_PROT_EXECUTE;
1259 
1260 	if (wired)
1261 		pvo_flags |= PVO_WIRED;
1262 
1263 	if ((m->flags & PG_FICTITIOUS) != 0)
1264 		pvo_flags |= PVO_FAKE;
1265 
1266 	error = moea64_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1267 	    pte_lo, pvo_flags);
1268 
1269 	if (pmap == kernel_pmap)
1270 		TLBIE(pmap, va);
1271 
1272 	/*
1273 	 * Flush the page from the instruction cache if this page is
1274 	 * mapped executable and cacheable.
1275 	 */
1276 	if ((pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1277 		moea64_syncicache(pmap, va, VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1278 	}
1279 }
1280 
1281 static void
1282 moea64_syncicache(pmap_t pmap, vm_offset_t va, vm_offset_t pa, vm_size_t sz)
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);
1307 		__syncicache((void *)moea64_scratchpage_va[1], sz);
1308 
1309 		mtx_unlock(&moea64_scratchpage_mtx);
1310 	}
1311 }
1312 
1313 /*
1314  * Maps a sequence of resident pages belonging to the same object.
1315  * The sequence begins with the given page m_start.  This page is
1316  * mapped at the given virtual address start.  Each subsequent page is
1317  * mapped at a virtual address that is offset from start by the same
1318  * amount as the page is offset from m_start within the object.  The
1319  * last page in the sequence is the page with the largest offset from
1320  * m_start that can be mapped at a virtual address less than the given
1321  * virtual address end.  Not every virtual page between start and end
1322  * is mapped; only those for which a resident page exists with the
1323  * corresponding offset from m_start are mapped.
1324  */
1325 void
1326 moea64_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1327     vm_page_t m_start, vm_prot_t prot)
1328 {
1329 	vm_page_t m;
1330 	vm_pindex_t diff, psize;
1331 
1332 	psize = atop(end - start);
1333 	m = m_start;
1334 	PMAP_LOCK(pm);
1335 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1336 		moea64_enter_locked(pm, start + ptoa(diff), m, prot &
1337 		    (VM_PROT_READ | VM_PROT_EXECUTE), FALSE);
1338 		m = TAILQ_NEXT(m, listq);
1339 	}
1340 	PMAP_UNLOCK(pm);
1341 }
1342 
1343 void
1344 moea64_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1345     vm_prot_t prot)
1346 {
1347 	PMAP_LOCK(pm);
1348 	moea64_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1349 	    FALSE);
1350 	PMAP_UNLOCK(pm);
1351 
1352 }
1353 
1354 vm_paddr_t
1355 moea64_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1356 {
1357 	struct	pvo_entry *pvo;
1358 	vm_paddr_t pa;
1359 
1360 	PMAP_LOCK(pm);
1361 	pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1362 	if (pvo == NULL)
1363 		pa = 0;
1364 	else
1365 		pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) | (va & ADDR_POFF);
1366 	PMAP_UNLOCK(pm);
1367 	return (pa);
1368 }
1369 
1370 /*
1371  * Atomically extract and hold the physical page with the given
1372  * pmap and virtual address pair if that mapping permits the given
1373  * protection.
1374  */
1375 vm_page_t
1376 moea64_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1377 {
1378 	struct	pvo_entry *pvo;
1379 	vm_page_t m;
1380 
1381 	m = NULL;
1382 	vm_page_lock_queues();
1383 	PMAP_LOCK(pmap);
1384 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1385 	if (pvo != NULL && (pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) &&
1386 	    ((pvo->pvo_pte.lpte.pte_lo & LPTE_PP) == LPTE_RW ||
1387 	     (prot & VM_PROT_WRITE) == 0)) {
1388 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN);
1389 		vm_page_hold(m);
1390 	}
1391 	vm_page_unlock_queues();
1392 	PMAP_UNLOCK(pmap);
1393 	return (m);
1394 }
1395 
1396 static void *
1397 moea64_uma_page_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
1398 {
1399 	/*
1400 	 * This entire routine is a horrible hack to avoid bothering kmem
1401 	 * for new KVA addresses. Because this can get called from inside
1402 	 * kmem allocation routines, calling kmem for a new address here
1403 	 * can lead to multiply locking non-recursive mutexes.
1404 	 */
1405 	static vm_pindex_t color;
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 (needed_lock)
1415 		PMAP_LOCK(kernel_pmap);
1416 
1417         if ((wait & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT)
1418                 pflags = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
1419         else
1420                 pflags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED;
1421         if (wait & M_ZERO)
1422                 pflags |= VM_ALLOC_ZERO;
1423 
1424         for (;;) {
1425                 m = vm_page_alloc(NULL, color++, pflags | VM_ALLOC_NOOBJ);
1426                 if (m == NULL) {
1427                         if (wait & M_NOWAIT)
1428                                 return (NULL);
1429                         VM_WAIT;
1430                 } else
1431                         break;
1432         }
1433 
1434 	va = pvo_allocator_start;
1435 	pvo_allocator_start += PAGE_SIZE;
1436 
1437 	if (pvo_allocator_start >= pvo_allocator_end)
1438 		panic("Ran out of PVO allocator buffer space!");
1439 
1440 	moea64_pvo_enter(kernel_pmap, moea64_upvo_zone,
1441 	    &moea64_pvo_kunmanaged, va,  VM_PAGE_TO_PHYS(m), LPTE_M,
1442 	    PVO_WIRED | PVO_BOOTSTRAP);
1443 
1444 	TLBIE(kernel_pmap, va);
1445 
1446 	if (needed_lock)
1447 		PMAP_UNLOCK(kernel_pmap);
1448 
1449 	if ((wait & M_ZERO) && (m->flags & PG_ZERO) == 0)
1450                 bzero((void *)va, PAGE_SIZE);
1451 
1452 	return (void *)va;
1453 }
1454 
1455 void
1456 moea64_init(mmu_t mmu)
1457 {
1458 
1459 	CTR0(KTR_PMAP, "moea64_init");
1460 
1461 	moea64_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1462 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1463 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1464 	moea64_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1465 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1466 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1467 
1468 	if (!hw_direct_map) {
1469 		uma_zone_set_allocf(moea64_upvo_zone,moea64_uma_page_alloc);
1470 		uma_zone_set_allocf(moea64_mpvo_zone,moea64_uma_page_alloc);
1471 	}
1472 
1473 	moea64_initialized = TRUE;
1474 }
1475 
1476 boolean_t
1477 moea64_is_modified(mmu_t mmu, vm_page_t m)
1478 {
1479 
1480 	if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0)
1481 		return (FALSE);
1482 
1483 	return (moea64_query_bit(m, LPTE_CHG));
1484 }
1485 
1486 void
1487 moea64_clear_reference(mmu_t mmu, vm_page_t m)
1488 {
1489 
1490 	if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0)
1491 		return;
1492 	moea64_clear_bit(m, LPTE_REF, NULL);
1493 }
1494 
1495 void
1496 moea64_clear_modify(mmu_t mmu, vm_page_t m)
1497 {
1498 
1499 	if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0)
1500 		return;
1501 	moea64_clear_bit(m, LPTE_CHG, NULL);
1502 }
1503 
1504 /*
1505  * Clear the write and modified bits in each of the given page's mappings.
1506  */
1507 void
1508 moea64_remove_write(mmu_t mmu, vm_page_t m)
1509 {
1510 	struct	pvo_entry *pvo;
1511 	struct	lpte *pt;
1512 	pmap_t	pmap;
1513 	uint64_t lo;
1514 
1515 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1516 	if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0 ||
1517 	    (m->flags & PG_WRITEABLE) == 0)
1518 		return;
1519 	lo = moea64_attr_fetch(m);
1520 	SYNC();
1521 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1522 		pmap = pvo->pvo_pmap;
1523 		PMAP_LOCK(pmap);
1524 		if ((pvo->pvo_pte.lpte.pte_lo & LPTE_PP) != LPTE_BR) {
1525 			LOCK_TABLE();
1526 			pt = moea64_pvo_to_pte(pvo, -1);
1527 			pvo->pvo_pte.lpte.pte_lo &= ~LPTE_PP;
1528 			pvo->pvo_pte.lpte.pte_lo |= LPTE_BR;
1529 			if (pt != NULL) {
1530 				moea64_pte_synch(pt, &pvo->pvo_pte.lpte);
1531 				lo |= pvo->pvo_pte.lpte.pte_lo;
1532 				pvo->pvo_pte.lpte.pte_lo &= ~LPTE_CHG;
1533 				moea64_pte_change(pt, &pvo->pvo_pte.lpte,
1534 				    pvo->pvo_pmap, pvo->pvo_vaddr);
1535 			}
1536 			UNLOCK_TABLE();
1537 		}
1538 		PMAP_UNLOCK(pmap);
1539 	}
1540 	if ((lo & LPTE_CHG) != 0) {
1541 		moea64_attr_clear(m, LPTE_CHG);
1542 		vm_page_dirty(m);
1543 	}
1544 	vm_page_flag_clear(m, PG_WRITEABLE);
1545 }
1546 
1547 /*
1548  *	moea64_ts_referenced:
1549  *
1550  *	Return a count of reference bits for a page, clearing those bits.
1551  *	It is not necessary for every reference bit to be cleared, but it
1552  *	is necessary that 0 only be returned when there are truly no
1553  *	reference bits set.
1554  *
1555  *	XXX: The exact number of bits to check and clear is a matter that
1556  *	should be tested and standardized at some point in the future for
1557  *	optimal aging of shared pages.
1558  */
1559 boolean_t
1560 moea64_ts_referenced(mmu_t mmu, vm_page_t m)
1561 {
1562 	int count;
1563 
1564 	if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0)
1565 		return (0);
1566 
1567 	count = moea64_clear_bit(m, LPTE_REF, NULL);
1568 
1569 	return (count);
1570 }
1571 
1572 /*
1573  * Map a wired page into kernel virtual address space.
1574  */
1575 void
1576 moea64_kenter(mmu_t mmu, vm_offset_t va, vm_offset_t pa)
1577 {
1578 	uint64_t	pte_lo;
1579 	int		error;
1580 
1581 	if (!pmap_bootstrapped) {
1582 		if (va >= VM_MIN_KERNEL_ADDRESS && va < VM_MAX_KERNEL_ADDRESS)
1583 			panic("Trying to enter an address in KVA -- %#x!\n",pa);
1584 	}
1585 
1586 	pte_lo = moea64_calc_wimg(pa);
1587 
1588 	PMAP_LOCK(kernel_pmap);
1589 	error = moea64_pvo_enter(kernel_pmap, moea64_upvo_zone,
1590 	    &moea64_pvo_kunmanaged, va, pa, pte_lo,
1591 	    PVO_WIRED | VM_PROT_EXECUTE);
1592 
1593 	TLBIE(kernel_pmap, va);
1594 
1595 	if (error != 0 && error != ENOENT)
1596 		panic("moea64_kenter: failed to enter va %#x pa %#x: %d", va,
1597 		    pa, error);
1598 
1599 	/*
1600 	 * Flush the memory from the instruction cache.
1601 	 */
1602 	if ((pte_lo & (LPTE_I | LPTE_G)) == 0) {
1603 		__syncicache((void *)va, PAGE_SIZE);
1604 	}
1605 	PMAP_UNLOCK(kernel_pmap);
1606 }
1607 
1608 /*
1609  * Extract the physical page address associated with the given kernel virtual
1610  * address.
1611  */
1612 vm_offset_t
1613 moea64_kextract(mmu_t mmu, vm_offset_t va)
1614 {
1615 	struct		pvo_entry *pvo;
1616 	vm_paddr_t pa;
1617 
1618 	PMAP_LOCK(kernel_pmap);
1619 	pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1620 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found"));
1621 	pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) | (va & ADDR_POFF);
1622 	PMAP_UNLOCK(kernel_pmap);
1623 	return (pa);
1624 }
1625 
1626 /*
1627  * Remove a wired page from kernel virtual address space.
1628  */
1629 void
1630 moea64_kremove(mmu_t mmu, vm_offset_t va)
1631 {
1632 	moea64_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1633 }
1634 
1635 /*
1636  * Map a range of physical addresses into kernel virtual address space.
1637  *
1638  * The value passed in *virt is a suggested virtual address for the mapping.
1639  * Architectures which can support a direct-mapped physical to virtual region
1640  * can return the appropriate address within that region, leaving '*virt'
1641  * unchanged.  We cannot and therefore do not; *virt is updated with the
1642  * first usable address after the mapped region.
1643  */
1644 vm_offset_t
1645 moea64_map(mmu_t mmu, vm_offset_t *virt, vm_offset_t pa_start,
1646     vm_offset_t pa_end, int prot)
1647 {
1648 	vm_offset_t	sva, va;
1649 
1650 	sva = *virt;
1651 	va = sva;
1652 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1653 		moea64_kenter(mmu, va, pa_start);
1654 	*virt = va;
1655 
1656 	return (sva);
1657 }
1658 
1659 /*
1660  * Returns true if the pmap's pv is one of the first
1661  * 16 pvs linked to from this page.  This count may
1662  * be changed upwards or downwards in the future; it
1663  * is only necessary that true be returned for a small
1664  * subset of pmaps for proper page aging.
1665  */
1666 boolean_t
1667 moea64_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1668 {
1669         int loops;
1670 	struct pvo_entry *pvo;
1671 
1672         if (!moea64_initialized || (m->flags & PG_FICTITIOUS))
1673                 return FALSE;
1674 
1675 	loops = 0;
1676 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1677 		if (pvo->pvo_pmap == pmap)
1678 			return (TRUE);
1679 		if (++loops >= 16)
1680 			break;
1681 	}
1682 
1683 	return (FALSE);
1684 }
1685 
1686 /*
1687  * Return the number of managed mappings to the given physical page
1688  * that are wired.
1689  */
1690 int
1691 moea64_page_wired_mappings(mmu_t mmu, vm_page_t m)
1692 {
1693 	struct pvo_entry *pvo;
1694 	int count;
1695 
1696 	count = 0;
1697 	if (!moea64_initialized || (m->flags & PG_FICTITIOUS) != 0)
1698 		return (count);
1699 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1700 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1701 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1702 			count++;
1703 	return (count);
1704 }
1705 
1706 static u_int	moea64_vsidcontext;
1707 
1708 void
1709 moea64_pinit(mmu_t mmu, pmap_t pmap)
1710 {
1711 	int	i, mask;
1712 	u_int	entropy;
1713 
1714 	PMAP_LOCK_INIT(pmap);
1715 
1716 	entropy = 0;
1717 	__asm __volatile("mftb %0" : "=r"(entropy));
1718 
1719 	if (pmap_bootstrapped)
1720 		pmap->pmap_phys = (pmap_t)moea64_kextract(mmu, (vm_offset_t)pmap);
1721 	else
1722 		pmap->pmap_phys = pmap;
1723 
1724 	/*
1725 	 * Allocate some segment registers for this pmap.
1726 	 */
1727 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1728 		u_int	hash, n;
1729 
1730 		/*
1731 		 * Create a new value by mutiplying by a prime and adding in
1732 		 * entropy from the timebase register.  This is to make the
1733 		 * VSID more random so that the PT hash function collides
1734 		 * less often.  (Note that the prime casues gcc to do shifts
1735 		 * instead of a multiply.)
1736 		 */
1737 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
1738 		hash = moea64_vsidcontext & (NPMAPS - 1);
1739 		if (hash == 0)		/* 0 is special, avoid it */
1740 			continue;
1741 		n = hash >> 5;
1742 		mask = 1 << (hash & (VSID_NBPW - 1));
1743 		hash = (moea64_vsidcontext & 0xfffff);
1744 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
1745 			/* anything free in this bucket? */
1746 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
1747 				entropy = (moea64_vsidcontext >> 20);
1748 				continue;
1749 			}
1750 			i = ffs(~moea64_vsid_bitmap[i]) - 1;
1751 			mask = 1 << i;
1752 			hash &= 0xfffff & ~(VSID_NBPW - 1);
1753 			hash |= i;
1754 		}
1755 		moea64_vsid_bitmap[n] |= mask;
1756 		for (i = 0; i < 16; i++) {
1757 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1758 		}
1759 		return;
1760 	}
1761 
1762 	panic("moea64_pinit: out of segments");
1763 }
1764 
1765 /*
1766  * Initialize the pmap associated with process 0.
1767  */
1768 void
1769 moea64_pinit0(mmu_t mmu, pmap_t pm)
1770 {
1771 	moea64_pinit(mmu, pm);
1772 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1773 }
1774 
1775 /*
1776  * Set the physical protection on the specified range of this map as requested.
1777  */
1778 void
1779 moea64_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1780     vm_prot_t prot)
1781 {
1782 	struct	pvo_entry *pvo;
1783 	struct	lpte *pt;
1784 	int	pteidx;
1785 
1786 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm, sva,
1787 	    eva, prot);
1788 
1789 
1790 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1791 	    ("moea64_protect: non current pmap"));
1792 
1793 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1794 		moea64_remove(mmu, pm, sva, eva);
1795 		return;
1796 	}
1797 
1798 	vm_page_lock_queues();
1799 	PMAP_LOCK(pm);
1800 	for (; sva < eva; sva += PAGE_SIZE) {
1801 		pvo = moea64_pvo_find_va(pm, sva, &pteidx);
1802 		if (pvo == NULL)
1803 			continue;
1804 
1805 		/*
1806 		 * Grab the PTE pointer before we diddle with the cached PTE
1807 		 * copy.
1808 		 */
1809 		LOCK_TABLE();
1810 		pt = moea64_pvo_to_pte(pvo, pteidx);
1811 
1812 		/*
1813 		 * Change the protection of the page.
1814 		 */
1815 		pvo->pvo_pte.lpte.pte_lo &= ~LPTE_PP;
1816 		pvo->pvo_pte.lpte.pte_lo |= LPTE_BR;
1817 		pvo->pvo_pte.lpte.pte_lo &= ~LPTE_NOEXEC;
1818 		if ((prot & VM_PROT_EXECUTE) == 0)
1819 			pvo->pvo_pte.lpte.pte_lo |= LPTE_NOEXEC;
1820 
1821 		/*
1822 		 * If the PVO is in the page table, update that pte as well.
1823 		 */
1824 		if (pt != NULL) {
1825 			moea64_pte_change(pt, &pvo->pvo_pte.lpte,
1826 			    pvo->pvo_pmap, pvo->pvo_vaddr);
1827 			if ((pvo->pvo_pte.lpte.pte_lo &
1828 			    (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1829 				moea64_syncicache(pm, sva,
1830 				    pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN,
1831 				    PAGE_SIZE);
1832 			}
1833 		}
1834 		UNLOCK_TABLE();
1835 	}
1836 	vm_page_unlock_queues();
1837 	PMAP_UNLOCK(pm);
1838 }
1839 
1840 /*
1841  * Map a list of wired pages into kernel virtual address space.  This is
1842  * intended for temporary mappings which do not need page modification or
1843  * references recorded.  Existing mappings in the region are overwritten.
1844  */
1845 void
1846 moea64_qenter(mmu_t mmu, vm_offset_t va, vm_page_t *m, int count)
1847 {
1848 	while (count-- > 0) {
1849 		moea64_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1850 		va += PAGE_SIZE;
1851 		m++;
1852 	}
1853 }
1854 
1855 /*
1856  * Remove page mappings from kernel virtual address space.  Intended for
1857  * temporary mappings entered by moea64_qenter.
1858  */
1859 void
1860 moea64_qremove(mmu_t mmu, vm_offset_t va, int count)
1861 {
1862 	while (count-- > 0) {
1863 		moea64_kremove(mmu, va);
1864 		va += PAGE_SIZE;
1865 	}
1866 }
1867 
1868 void
1869 moea64_release(mmu_t mmu, pmap_t pmap)
1870 {
1871         int idx, mask;
1872 
1873 	/*
1874 	 * Free segment register's VSID
1875 	 */
1876         if (pmap->pm_sr[0] == 0)
1877                 panic("moea64_release");
1878 
1879         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1880         mask = 1 << (idx % VSID_NBPW);
1881         idx /= VSID_NBPW;
1882         moea64_vsid_bitmap[idx] &= ~mask;
1883 	PMAP_LOCK_DESTROY(pmap);
1884 }
1885 
1886 /*
1887  * Remove the given range of addresses from the specified map.
1888  */
1889 void
1890 moea64_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1891 {
1892 	struct	pvo_entry *pvo;
1893 	int	pteidx;
1894 
1895 	vm_page_lock_queues();
1896 	PMAP_LOCK(pm);
1897 	for (; sva < eva; sva += PAGE_SIZE) {
1898 		pvo = moea64_pvo_find_va(pm, sva, &pteidx);
1899 		if (pvo != NULL) {
1900 			moea64_pvo_remove(pvo, pteidx);
1901 		}
1902 	}
1903 	vm_page_unlock_queues();
1904 	PMAP_UNLOCK(pm);
1905 }
1906 
1907 /*
1908  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
1909  * will reflect changes in pte's back to the vm_page.
1910  */
1911 void
1912 moea64_remove_all(mmu_t mmu, vm_page_t m)
1913 {
1914 	struct  pvo_head *pvo_head;
1915 	struct	pvo_entry *pvo, *next_pvo;
1916 	pmap_t	pmap;
1917 
1918 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1919 
1920 	pvo_head = vm_page_to_pvoh(m);
1921 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1922 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1923 
1924 		MOEA_PVO_CHECK(pvo);	/* sanity check */
1925 		pmap = pvo->pvo_pmap;
1926 		PMAP_LOCK(pmap);
1927 		moea64_pvo_remove(pvo, -1);
1928 		PMAP_UNLOCK(pmap);
1929 	}
1930 	vm_page_flag_clear(m, PG_WRITEABLE);
1931 }
1932 
1933 /*
1934  * Allocate a physical page of memory directly from the phys_avail map.
1935  * Can only be called from moea64_bootstrap before avail start and end are
1936  * calculated.
1937  */
1938 static vm_offset_t
1939 moea64_bootstrap_alloc(vm_size_t size, u_int align)
1940 {
1941 	vm_offset_t	s, e;
1942 	int		i, j;
1943 
1944 	size = round_page(size);
1945 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1946 		if (align != 0)
1947 			s = (phys_avail[i] + align - 1) & ~(align - 1);
1948 		else
1949 			s = phys_avail[i];
1950 		e = s + size;
1951 
1952 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1953 			continue;
1954 
1955 		if (s == phys_avail[i]) {
1956 			phys_avail[i] += size;
1957 		} else if (e == phys_avail[i + 1]) {
1958 			phys_avail[i + 1] -= size;
1959 		} else {
1960 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1961 				phys_avail[j] = phys_avail[j - 2];
1962 				phys_avail[j + 1] = phys_avail[j - 1];
1963 			}
1964 
1965 			phys_avail[i + 3] = phys_avail[i + 1];
1966 			phys_avail[i + 1] = s;
1967 			phys_avail[i + 2] = e;
1968 			phys_avail_count++;
1969 		}
1970 
1971 		return (s);
1972 	}
1973 	panic("moea64_bootstrap_alloc: could not allocate memory");
1974 }
1975 
1976 static void
1977 tlbia(void)
1978 {
1979 	vm_offset_t i;
1980 	register_t msr, scratch;
1981 
1982 	for (i = 0; i < 0xFF000; i += 0x00001000) {
1983 		__asm __volatile("\
1984 		    mfmsr %0; \
1985 		    mr %1, %0; \
1986 		    insrdi %1,%3,1,0; \
1987 		    mtmsrd %1; \
1988 		    ptesync; \
1989 		    \
1990 		    tlbiel %2; \
1991 		    \
1992 		    mtmsrd %0; \
1993 		    eieio; \
1994 		    tlbsync; \
1995 		    ptesync;"
1996 		: "=r"(msr), "=r"(scratch) : "r"(i), "r"(1));
1997 	}
1998 }
1999 
2000 static int
2001 moea64_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
2002     vm_offset_t va, vm_offset_t pa, uint64_t pte_lo, int flags)
2003 {
2004 	struct	 pvo_entry *pvo;
2005 	uint64_t vsid;
2006 	int	 first;
2007 	u_int	 ptegidx;
2008 	int	 i;
2009 	int      bootstrap;
2010 
2011 	/*
2012 	 * One nasty thing that can happen here is that the UMA calls to
2013 	 * allocate new PVOs need to map more memory, which calls pvo_enter(),
2014 	 * which calls UMA...
2015 	 *
2016 	 * We break the loop by detecting recursion and allocating out of
2017 	 * the bootstrap pool.
2018 	 */
2019 
2020 	moea64_pvo_enter_calls++;
2021 	first = 0;
2022 	bootstrap = (flags & PVO_BOOTSTRAP);
2023 
2024 	if (!moea64_initialized)
2025 		bootstrap = 1;
2026 
2027 	/*
2028 	 * Compute the PTE Group index.
2029 	 */
2030 	va &= ~ADDR_POFF;
2031 	vsid = va_to_vsid(pm, va);
2032 	ptegidx = va_to_pteg(vsid, va);
2033 
2034 	/*
2035 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
2036 	 * there is a mapping.
2037 	 */
2038 	LOCK_TABLE();
2039 
2040 	LIST_FOREACH(pvo, &moea64_pvo_table[ptegidx], pvo_olink) {
2041 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2042 			if ((pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) == pa &&
2043 			    (pvo->pvo_pte.lpte.pte_lo & LPTE_PP) ==
2044 			    (pte_lo & LPTE_PP)) {
2045 				UNLOCK_TABLE();
2046 				return (0);
2047 			}
2048 			moea64_pvo_remove(pvo, -1);
2049 			break;
2050 		}
2051 	}
2052 
2053 	/*
2054 	 * If we aren't overwriting a mapping, try to allocate.
2055 	 */
2056 	if (bootstrap) {
2057 		if (moea64_bpvo_pool_index >= BPVO_POOL_SIZE) {
2058 			panic("moea64_enter: bpvo pool exhausted, %d, %d, %d",
2059 			      moea64_bpvo_pool_index, BPVO_POOL_SIZE,
2060 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
2061 		}
2062 		pvo = &moea64_bpvo_pool[moea64_bpvo_pool_index];
2063 		moea64_bpvo_pool_index++;
2064 		bootstrap = 1;
2065 	} else {
2066 		/*
2067 		 * Note: drop the table around the UMA allocation in
2068 		 * case the UMA allocator needs to manipulate the page
2069 		 * table. The mapping we are working with is already
2070 		 * protected by the PMAP lock.
2071 		 */
2072 		UNLOCK_TABLE();
2073 		pvo = uma_zalloc(zone, M_NOWAIT);
2074 		LOCK_TABLE();
2075 	}
2076 
2077 	if (pvo == NULL) {
2078 		UNLOCK_TABLE();
2079 		return (ENOMEM);
2080 	}
2081 
2082 	moea64_pvo_entries++;
2083 	pvo->pvo_vaddr = va;
2084 	pvo->pvo_pmap = pm;
2085 	LIST_INSERT_HEAD(&moea64_pvo_table[ptegidx], pvo, pvo_olink);
2086 	pvo->pvo_vaddr &= ~ADDR_POFF;
2087 
2088 	if (!(flags & VM_PROT_EXECUTE))
2089 		pte_lo |= LPTE_NOEXEC;
2090 	if (flags & PVO_WIRED)
2091 		pvo->pvo_vaddr |= PVO_WIRED;
2092 	if (pvo_head != &moea64_pvo_kunmanaged)
2093 		pvo->pvo_vaddr |= PVO_MANAGED;
2094 	if (bootstrap)
2095 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
2096 	if (flags & PVO_FAKE)
2097 		pvo->pvo_vaddr |= PVO_FAKE;
2098 
2099 	moea64_pte_create(&pvo->pvo_pte.lpte, vsid, va,
2100 	    (uint64_t)(pa) | pte_lo);
2101 
2102 	/*
2103 	 * Remember if the list was empty and therefore will be the first
2104 	 * item.
2105 	 */
2106 	if (LIST_FIRST(pvo_head) == NULL)
2107 		first = 1;
2108 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2109 
2110 	if (pvo->pvo_pte.lpte.pte_lo & PVO_WIRED)
2111 		pm->pm_stats.wired_count++;
2112 	pm->pm_stats.resident_count++;
2113 
2114 	/*
2115 	 * We hope this succeeds but it isn't required.
2116 	 */
2117 	i = moea64_pte_insert(ptegidx, &pvo->pvo_pte.lpte);
2118 	if (i >= 0) {
2119 		PVO_PTEGIDX_SET(pvo, i);
2120 	} else {
2121 		panic("moea64_pvo_enter: overflow");
2122 		moea64_pte_overflow++;
2123 	}
2124 
2125 	UNLOCK_TABLE();
2126 
2127 	return (first ? ENOENT : 0);
2128 }
2129 
2130 static void
2131 moea64_pvo_remove(struct pvo_entry *pvo, int pteidx)
2132 {
2133 	struct	lpte *pt;
2134 
2135 	/*
2136 	 * If there is an active pte entry, we need to deactivate it (and
2137 	 * save the ref & cfg bits).
2138 	 */
2139 	LOCK_TABLE();
2140 	pt = moea64_pvo_to_pte(pvo, pteidx);
2141 	if (pt != NULL) {
2142 		moea64_pte_unset(pt, &pvo->pvo_pte.lpte, pvo->pvo_pmap,
2143 		    pvo->pvo_vaddr);
2144 		PVO_PTEGIDX_CLR(pvo);
2145 	} else {
2146 		moea64_pte_overflow--;
2147 	}
2148 	UNLOCK_TABLE();
2149 
2150 	/*
2151 	 * Update our statistics.
2152 	 */
2153 	pvo->pvo_pmap->pm_stats.resident_count--;
2154 	if (pvo->pvo_pte.lpte.pte_lo & PVO_WIRED)
2155 		pvo->pvo_pmap->pm_stats.wired_count--;
2156 
2157 	/*
2158 	 * Save the REF/CHG bits into their cache if the page is managed.
2159 	 */
2160 	if ((pvo->pvo_vaddr & (PVO_MANAGED|PVO_FAKE)) == PVO_MANAGED) {
2161 		struct	vm_page *pg;
2162 
2163 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN);
2164 		if (pg != NULL) {
2165 			moea64_attr_save(pg, pvo->pvo_pte.lpte.pte_lo &
2166 			    (LPTE_REF | LPTE_CHG));
2167 		}
2168 	}
2169 
2170 	/*
2171 	 * Remove this PVO from the PV list.
2172 	 */
2173 	LIST_REMOVE(pvo, pvo_vlink);
2174 
2175 	/*
2176 	 * Remove this from the overflow list and return it to the pool
2177 	 * if we aren't going to reuse it.
2178 	 */
2179 	LIST_REMOVE(pvo, pvo_olink);
2180 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2181 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea64_mpvo_zone :
2182 		    moea64_upvo_zone, pvo);
2183 	moea64_pvo_entries--;
2184 	moea64_pvo_remove_calls++;
2185 }
2186 
2187 static __inline int
2188 moea64_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2189 {
2190 	int	pteidx;
2191 
2192 	/*
2193 	 * We can find the actual pte entry without searching by grabbing
2194 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2195 	 * noticing the HID bit.
2196 	 */
2197 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2198 	if (pvo->pvo_pte.lpte.pte_hi & LPTE_HID)
2199 		pteidx ^= moea64_pteg_mask * 8;
2200 
2201 	return (pteidx);
2202 }
2203 
2204 static struct pvo_entry *
2205 moea64_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2206 {
2207 	struct		pvo_entry *pvo;
2208 	int		ptegidx;
2209 	uint64_t	vsid;
2210 
2211 	va &= ~ADDR_POFF;
2212 	vsid = va_to_vsid(pm, va);
2213 	ptegidx = va_to_pteg(vsid, va);
2214 
2215 	LOCK_TABLE();
2216 	LIST_FOREACH(pvo, &moea64_pvo_table[ptegidx], pvo_olink) {
2217 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2218 			if (pteidx_p)
2219 				*pteidx_p = moea64_pvo_pte_index(pvo, ptegidx);
2220 			break;
2221 		}
2222 	}
2223 	UNLOCK_TABLE();
2224 
2225 	return (pvo);
2226 }
2227 
2228 static struct lpte *
2229 moea64_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2230 {
2231 	struct lpte *pt;
2232 
2233 	/*
2234 	 * If we haven't been supplied the ptegidx, calculate it.
2235 	 */
2236 	if (pteidx == -1) {
2237 		int		ptegidx;
2238 		uint64_t	vsid;
2239 
2240 		vsid = va_to_vsid(pvo->pvo_pmap, pvo->pvo_vaddr);
2241 		ptegidx = va_to_pteg(vsid, pvo->pvo_vaddr);
2242 		pteidx = moea64_pvo_pte_index(pvo, ptegidx);
2243 	}
2244 
2245 	pt = &moea64_pteg_table[pteidx >> 3].pt[pteidx & 7];
2246 
2247 	if ((pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) &&
2248 	    !PVO_PTEGIDX_ISSET(pvo)) {
2249 		panic("moea64_pvo_to_pte: pvo %p has valid pte in pvo but no "
2250 		    "valid pte index", pvo);
2251 	}
2252 
2253 	if ((pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) == 0 &&
2254 	    PVO_PTEGIDX_ISSET(pvo)) {
2255 		panic("moea64_pvo_to_pte: pvo %p has valid pte index in pvo "
2256 		    "pvo but no valid pte", pvo);
2257 	}
2258 
2259 	if ((pt->pte_hi ^ (pvo->pvo_pte.lpte.pte_hi & ~LPTE_VALID)) ==
2260 	    LPTE_VALID) {
2261 		if ((pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) == 0) {
2262 			panic("moea64_pvo_to_pte: pvo %p has valid pte in "
2263 			    "moea64_pteg_table %p but invalid in pvo", pvo, pt);
2264 		}
2265 
2266 		if (((pt->pte_lo ^ pvo->pvo_pte.lpte.pte_lo) &
2267 		    ~(LPTE_CHG|LPTE_REF)) != 0) {
2268 			panic("moea64_pvo_to_pte: pvo %p pte does not match "
2269 			    "pte %p in moea64_pteg_table difference is %#x",
2270 			    pvo, pt,
2271 			    (uint32_t)(pt->pte_lo ^ pvo->pvo_pte.lpte.pte_lo));
2272 		}
2273 
2274 		ASSERT_TABLE_LOCK();
2275 		return (pt);
2276 	}
2277 
2278 	if (pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) {
2279 		panic("moea64_pvo_to_pte: pvo %p has invalid pte %p in "
2280 		    "moea64_pteg_table but valid in pvo", pvo, pt);
2281 	}
2282 
2283 	return (NULL);
2284 }
2285 
2286 static int
2287 moea64_pte_insert(u_int ptegidx, struct lpte *pvo_pt)
2288 {
2289 	struct	lpte *pt;
2290 	int	i;
2291 
2292 	ASSERT_TABLE_LOCK();
2293 
2294 	/*
2295 	 * First try primary hash.
2296 	 */
2297 	for (pt = moea64_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2298 		if ((pt->pte_hi & LPTE_VALID) == 0) {
2299 			pvo_pt->pte_hi &= ~LPTE_HID;
2300 			moea64_pte_set(pt, pvo_pt);
2301 			return (i);
2302 		}
2303 	}
2304 
2305 	/*
2306 	 * Now try secondary hash.
2307 	 */
2308 	ptegidx ^= moea64_pteg_mask;
2309 
2310 	for (pt = moea64_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2311 		if ((pt->pte_hi & LPTE_VALID) == 0) {
2312 			pvo_pt->pte_hi |= LPTE_HID;
2313 			moea64_pte_set(pt, pvo_pt);
2314 			return (i);
2315 		}
2316 	}
2317 
2318 	panic("moea64_pte_insert: overflow");
2319 	return (-1);
2320 }
2321 
2322 static boolean_t
2323 moea64_query_bit(vm_page_t m, u_int64_t ptebit)
2324 {
2325 	struct	pvo_entry *pvo;
2326 	struct	lpte *pt;
2327 
2328 #if 0
2329 	if (moea64_attr_fetch(m) & ptebit)
2330 		return (TRUE);
2331 #endif
2332 
2333 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2334 		MOEA_PVO_CHECK(pvo);	/* sanity check */
2335 
2336 		/*
2337 		 * See if we saved the bit off.  If so, cache it and return
2338 		 * success.
2339 		 */
2340 		if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2341 			moea64_attr_save(m, ptebit);
2342 			MOEA_PVO_CHECK(pvo);	/* sanity check */
2343 			return (TRUE);
2344 		}
2345 	}
2346 
2347 	/*
2348 	 * No luck, now go through the hard part of looking at the PTEs
2349 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2350 	 * the PTEs.
2351 	 */
2352 	SYNC();
2353 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2354 		MOEA_PVO_CHECK(pvo);	/* sanity check */
2355 
2356 		/*
2357 		 * See if this pvo has a valid PTE.  if so, fetch the
2358 		 * REF/CHG bits from the valid PTE.  If the appropriate
2359 		 * ptebit is set, cache it and return success.
2360 		 */
2361 		LOCK_TABLE();
2362 		pt = moea64_pvo_to_pte(pvo, -1);
2363 		if (pt != NULL) {
2364 			moea64_pte_synch(pt, &pvo->pvo_pte.lpte);
2365 			if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2366 				UNLOCK_TABLE();
2367 
2368 				moea64_attr_save(m, ptebit);
2369 				MOEA_PVO_CHECK(pvo);	/* sanity check */
2370 				return (TRUE);
2371 			}
2372 		}
2373 		UNLOCK_TABLE();
2374 	}
2375 
2376 	return (FALSE);
2377 }
2378 
2379 static u_int
2380 moea64_clear_bit(vm_page_t m, u_int64_t ptebit, u_int64_t *origbit)
2381 {
2382 	u_int	count;
2383 	struct	pvo_entry *pvo;
2384 	struct	lpte *pt;
2385 	uint64_t rv;
2386 
2387 	/*
2388 	 * Clear the cached value.
2389 	 */
2390 	rv = moea64_attr_fetch(m);
2391 	moea64_attr_clear(m, ptebit);
2392 
2393 	/*
2394 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2395 	 * we can reset the right ones).  note that since the pvo entries and
2396 	 * list heads are accessed via BAT0 and are never placed in the page
2397 	 * table, we don't have to worry about further accesses setting the
2398 	 * REF/CHG bits.
2399 	 */
2400 	SYNC();
2401 
2402 	/*
2403 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2404 	 * valid pte clear the ptebit from the valid pte.
2405 	 */
2406 	count = 0;
2407 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2408 		MOEA_PVO_CHECK(pvo);	/* sanity check */
2409 
2410 		LOCK_TABLE();
2411 		pt = moea64_pvo_to_pte(pvo, -1);
2412 		if (pt != NULL) {
2413 			moea64_pte_synch(pt, &pvo->pvo_pte.lpte);
2414 			if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2415 				count++;
2416 				moea64_pte_clear(pt, pvo->pvo_pmap, PVO_VADDR(pvo), ptebit);
2417 			}
2418 		}
2419 		UNLOCK_TABLE();
2420 		rv |= pvo->pvo_pte.lpte.pte_lo;
2421 		pvo->pvo_pte.lpte.pte_lo &= ~ptebit;
2422 		MOEA_PVO_CHECK(pvo);	/* sanity check */
2423 	}
2424 
2425 	if (origbit != NULL) {
2426 		*origbit = rv;
2427 	}
2428 
2429 	return (count);
2430 }
2431 
2432 boolean_t
2433 moea64_dev_direct_mapped(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2434 {
2435 	return (EFAULT);
2436 }
2437 
2438 /*
2439  * Map a set of physical memory pages into the kernel virtual
2440  * address space. Return a pointer to where it is mapped. This
2441  * routine is intended to be used for mapping device memory,
2442  * NOT real memory.
2443  */
2444 void *
2445 moea64_mapdev(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2446 {
2447 	vm_offset_t va, tmpva, ppa, offset;
2448 
2449 	ppa = trunc_page(pa);
2450 	offset = pa & PAGE_MASK;
2451 	size = roundup(offset + size, PAGE_SIZE);
2452 
2453 	va = kmem_alloc_nofault(kernel_map, size);
2454 
2455 	if (!va)
2456 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
2457 
2458 	for (tmpva = va; size > 0;) {
2459 		moea64_kenter(mmu, tmpva, ppa);
2460 		size -= PAGE_SIZE;
2461 		tmpva += PAGE_SIZE;
2462 		ppa += PAGE_SIZE;
2463 	}
2464 
2465 	return ((void *)(va + offset));
2466 }
2467 
2468 void
2469 moea64_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2470 {
2471 	vm_offset_t base, offset;
2472 
2473 	base = trunc_page(va);
2474 	offset = va & PAGE_MASK;
2475 	size = roundup(offset + size, PAGE_SIZE);
2476 
2477 	kmem_free(kernel_map, base, size);
2478 }
2479 
2480 static void
2481 moea64_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2482 {
2483 	struct pvo_entry *pvo;
2484 	vm_offset_t lim;
2485 	vm_paddr_t pa;
2486 	vm_size_t len;
2487 
2488 	PMAP_LOCK(pm);
2489 	while (sz > 0) {
2490 		lim = round_page(va);
2491 		len = MIN(lim - va, sz);
2492 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2493 		if (pvo != NULL) {
2494 			pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) |
2495 			    (va & ADDR_POFF);
2496 			moea64_syncicache(pm, va, pa, len);
2497 		}
2498 		va += len;
2499 		sz -= len;
2500 	}
2501 	PMAP_UNLOCK(pm);
2502 }
2503