xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision 884a2a699669ec61e2366e3e358342dbc94be24a)
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/_inttypes.h>
146 #include <machine/cpu.h>
147 #include <machine/platform.h>
148 #include <machine/frame.h>
149 #include <machine/md_var.h>
150 #include <machine/psl.h>
151 #include <machine/bat.h>
152 #include <machine/hid.h>
153 #include <machine/pte.h>
154 #include <machine/sr.h>
155 #include <machine/trap.h>
156 #include <machine/mmuvar.h>
157 
158 #include "mmu_oea64.h"
159 #include "mmu_if.h"
160 #include "moea64_if.h"
161 
162 void moea64_release_vsid(uint64_t vsid);
163 uintptr_t moea64_get_unique_vsid(void);
164 
165 #define DISABLE_TRANS(msr)	msr = mfmsr(); mtmsr(msr & ~PSL_DR)
166 #define ENABLE_TRANS(msr)	mtmsr(msr)
167 
168 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
169 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
170 #define	VSID_HASH_MASK		0x0000007fffffffffULL
171 
172 #define LOCK_TABLE() mtx_lock(&moea64_table_mutex)
173 #define UNLOCK_TABLE() mtx_unlock(&moea64_table_mutex);
174 #define ASSERT_TABLE_LOCK() mtx_assert(&moea64_table_mutex, MA_OWNED)
175 
176 struct ofw_map {
177 	cell_t	om_va;
178 	cell_t	om_len;
179 	cell_t	om_pa_hi;
180 	cell_t	om_pa_lo;
181 	cell_t	om_mode;
182 };
183 
184 /*
185  * Map of physical memory regions.
186  */
187 static struct	mem_region *regions;
188 static struct	mem_region *pregions;
189 static u_int	phys_avail_count;
190 static int	regions_sz, pregions_sz;
191 
192 extern void bs_remap_earlyboot(void);
193 
194 /*
195  * Lock for the pteg and pvo tables.
196  */
197 struct mtx	moea64_table_mutex;
198 struct mtx	moea64_slb_mutex;
199 
200 /*
201  * PTEG data.
202  */
203 u_int		moea64_pteg_count;
204 u_int		moea64_pteg_mask;
205 
206 /*
207  * PVO data.
208  */
209 struct	pvo_head *moea64_pvo_table;		/* pvo entries by pteg index */
210 struct	pvo_head moea64_pvo_kunmanaged =	/* list of unmanaged pages */
211     LIST_HEAD_INITIALIZER(moea64_pvo_kunmanaged);
212 
213 uma_zone_t	moea64_upvo_zone; /* zone for pvo entries for unmanaged pages */
214 uma_zone_t	moea64_mpvo_zone; /* zone for pvo entries for managed pages */
215 
216 #define	BPVO_POOL_SIZE	327680
217 static struct	pvo_entry *moea64_bpvo_pool;
218 static int	moea64_bpvo_pool_index = 0;
219 
220 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
221 #ifdef __powerpc64__
222 #define	NVSIDS		(NPMAPS * 16)
223 #define VSID_HASHMASK	0xffffffffUL
224 #else
225 #define NVSIDS		NPMAPS
226 #define VSID_HASHMASK	0xfffffUL
227 #endif
228 static u_int	moea64_vsid_bitmap[NVSIDS / VSID_NBPW];
229 
230 static boolean_t moea64_initialized = FALSE;
231 
232 /*
233  * Statistics.
234  */
235 u_int	moea64_pte_valid = 0;
236 u_int	moea64_pte_overflow = 0;
237 u_int	moea64_pvo_entries = 0;
238 u_int	moea64_pvo_enter_calls = 0;
239 u_int	moea64_pvo_remove_calls = 0;
240 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_valid, CTLFLAG_RD,
241     &moea64_pte_valid, 0, "");
242 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_overflow, CTLFLAG_RD,
243     &moea64_pte_overflow, 0, "");
244 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_entries, CTLFLAG_RD,
245     &moea64_pvo_entries, 0, "");
246 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_enter_calls, CTLFLAG_RD,
247     &moea64_pvo_enter_calls, 0, "");
248 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_remove_calls, CTLFLAG_RD,
249     &moea64_pvo_remove_calls, 0, "");
250 
251 vm_offset_t	moea64_scratchpage_va[2];
252 struct pvo_entry *moea64_scratchpage_pvo[2];
253 uintptr_t	moea64_scratchpage_pte[2];
254 struct	mtx	moea64_scratchpage_mtx;
255 
256 uint64_t 	moea64_large_page_mask = 0;
257 int		moea64_large_page_size = 0;
258 int		moea64_large_page_shift = 0;
259 
260 /*
261  * PVO calls.
262  */
263 static int	moea64_pvo_enter(mmu_t, pmap_t, uma_zone_t, struct pvo_head *,
264 		    vm_offset_t, vm_offset_t, uint64_t, int);
265 static void	moea64_pvo_remove(mmu_t, struct pvo_entry *);
266 static struct	pvo_entry *moea64_pvo_find_va(pmap_t, vm_offset_t);
267 
268 /*
269  * Utility routines.
270  */
271 static void		moea64_enter_locked(mmu_t, pmap_t, vm_offset_t,
272 			    vm_page_t, vm_prot_t, boolean_t);
273 static boolean_t	moea64_query_bit(mmu_t, vm_page_t, u_int64_t);
274 static u_int		moea64_clear_bit(mmu_t, vm_page_t, u_int64_t);
275 static void		moea64_kremove(mmu_t, vm_offset_t);
276 static void		moea64_syncicache(mmu_t, pmap_t pmap, vm_offset_t va,
277 			    vm_offset_t pa, vm_size_t sz);
278 
279 /*
280  * Kernel MMU interface
281  */
282 void moea64_change_wiring(mmu_t, pmap_t, vm_offset_t, boolean_t);
283 void moea64_clear_modify(mmu_t, vm_page_t);
284 void moea64_clear_reference(mmu_t, vm_page_t);
285 void moea64_copy_page(mmu_t, vm_page_t, vm_page_t);
286 void moea64_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, boolean_t);
287 void moea64_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
288     vm_prot_t);
289 void moea64_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
290 vm_paddr_t moea64_extract(mmu_t, pmap_t, vm_offset_t);
291 vm_page_t moea64_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
292 void moea64_init(mmu_t);
293 boolean_t moea64_is_modified(mmu_t, vm_page_t);
294 boolean_t moea64_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
295 boolean_t moea64_is_referenced(mmu_t, vm_page_t);
296 boolean_t moea64_ts_referenced(mmu_t, vm_page_t);
297 vm_offset_t moea64_map(mmu_t, vm_offset_t *, vm_offset_t, vm_offset_t, int);
298 boolean_t moea64_page_exists_quick(mmu_t, pmap_t, vm_page_t);
299 int moea64_page_wired_mappings(mmu_t, vm_page_t);
300 void moea64_pinit(mmu_t, pmap_t);
301 void moea64_pinit0(mmu_t, pmap_t);
302 void moea64_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
303 void moea64_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
304 void moea64_qremove(mmu_t, vm_offset_t, int);
305 void moea64_release(mmu_t, pmap_t);
306 void moea64_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
307 void moea64_remove_all(mmu_t, vm_page_t);
308 void moea64_remove_write(mmu_t, vm_page_t);
309 void moea64_zero_page(mmu_t, vm_page_t);
310 void moea64_zero_page_area(mmu_t, vm_page_t, int, int);
311 void moea64_zero_page_idle(mmu_t, vm_page_t);
312 void moea64_activate(mmu_t, struct thread *);
313 void moea64_deactivate(mmu_t, struct thread *);
314 void *moea64_mapdev(mmu_t, vm_offset_t, vm_size_t);
315 void *moea64_mapdev_attr(mmu_t, vm_offset_t, vm_size_t, vm_memattr_t);
316 void moea64_unmapdev(mmu_t, vm_offset_t, vm_size_t);
317 vm_offset_t moea64_kextract(mmu_t, vm_offset_t);
318 void moea64_page_set_memattr(mmu_t, vm_page_t m, vm_memattr_t ma);
319 void moea64_kenter_attr(mmu_t, vm_offset_t, vm_offset_t, vm_memattr_t ma);
320 void moea64_kenter(mmu_t, vm_offset_t, vm_offset_t);
321 boolean_t moea64_dev_direct_mapped(mmu_t, vm_offset_t, vm_size_t);
322 static void moea64_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
323 
324 static mmu_method_t moea64_methods[] = {
325 	MMUMETHOD(mmu_change_wiring,	moea64_change_wiring),
326 	MMUMETHOD(mmu_clear_modify,	moea64_clear_modify),
327 	MMUMETHOD(mmu_clear_reference,	moea64_clear_reference),
328 	MMUMETHOD(mmu_copy_page,	moea64_copy_page),
329 	MMUMETHOD(mmu_enter,		moea64_enter),
330 	MMUMETHOD(mmu_enter_object,	moea64_enter_object),
331 	MMUMETHOD(mmu_enter_quick,	moea64_enter_quick),
332 	MMUMETHOD(mmu_extract,		moea64_extract),
333 	MMUMETHOD(mmu_extract_and_hold,	moea64_extract_and_hold),
334 	MMUMETHOD(mmu_init,		moea64_init),
335 	MMUMETHOD(mmu_is_modified,	moea64_is_modified),
336 	MMUMETHOD(mmu_is_prefaultable,	moea64_is_prefaultable),
337 	MMUMETHOD(mmu_is_referenced,	moea64_is_referenced),
338 	MMUMETHOD(mmu_ts_referenced,	moea64_ts_referenced),
339 	MMUMETHOD(mmu_map,     		moea64_map),
340 	MMUMETHOD(mmu_page_exists_quick,moea64_page_exists_quick),
341 	MMUMETHOD(mmu_page_wired_mappings,moea64_page_wired_mappings),
342 	MMUMETHOD(mmu_pinit,		moea64_pinit),
343 	MMUMETHOD(mmu_pinit0,		moea64_pinit0),
344 	MMUMETHOD(mmu_protect,		moea64_protect),
345 	MMUMETHOD(mmu_qenter,		moea64_qenter),
346 	MMUMETHOD(mmu_qremove,		moea64_qremove),
347 	MMUMETHOD(mmu_release,		moea64_release),
348 	MMUMETHOD(mmu_remove,		moea64_remove),
349 	MMUMETHOD(mmu_remove_all,      	moea64_remove_all),
350 	MMUMETHOD(mmu_remove_write,	moea64_remove_write),
351 	MMUMETHOD(mmu_sync_icache,	moea64_sync_icache),
352 	MMUMETHOD(mmu_zero_page,       	moea64_zero_page),
353 	MMUMETHOD(mmu_zero_page_area,	moea64_zero_page_area),
354 	MMUMETHOD(mmu_zero_page_idle,	moea64_zero_page_idle),
355 	MMUMETHOD(mmu_activate,		moea64_activate),
356 	MMUMETHOD(mmu_deactivate,      	moea64_deactivate),
357 	MMUMETHOD(mmu_page_set_memattr,	moea64_page_set_memattr),
358 
359 	/* Internal interfaces */
360 	MMUMETHOD(mmu_mapdev,		moea64_mapdev),
361 	MMUMETHOD(mmu_mapdev_attr,	moea64_mapdev_attr),
362 	MMUMETHOD(mmu_unmapdev,		moea64_unmapdev),
363 	MMUMETHOD(mmu_kextract,		moea64_kextract),
364 	MMUMETHOD(mmu_kenter,		moea64_kenter),
365 	MMUMETHOD(mmu_kenter_attr,	moea64_kenter_attr),
366 	MMUMETHOD(mmu_dev_direct_mapped,moea64_dev_direct_mapped),
367 
368 	{ 0, 0 }
369 };
370 
371 MMU_DEF(oea64_mmu, "mmu_oea64_base", moea64_methods, 0);
372 
373 static __inline u_int
374 va_to_pteg(uint64_t vsid, vm_offset_t addr, int large)
375 {
376 	uint64_t hash;
377 	int shift;
378 
379 	shift = large ? moea64_large_page_shift : ADDR_PIDX_SHFT;
380 	hash = (vsid & VSID_HASH_MASK) ^ (((uint64_t)addr & ADDR_PIDX) >>
381 	    shift);
382 	return (hash & moea64_pteg_mask);
383 }
384 
385 static __inline struct pvo_head *
386 vm_page_to_pvoh(vm_page_t m)
387 {
388 
389 	return (&m->md.mdpg_pvoh);
390 }
391 
392 static __inline void
393 moea64_attr_clear(vm_page_t m, u_int64_t ptebit)
394 {
395 
396 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
397 	m->md.mdpg_attrs &= ~ptebit;
398 }
399 
400 static __inline u_int64_t
401 moea64_attr_fetch(vm_page_t m)
402 {
403 
404 	return (m->md.mdpg_attrs);
405 }
406 
407 static __inline void
408 moea64_attr_save(vm_page_t m, u_int64_t ptebit)
409 {
410 
411 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
412 	m->md.mdpg_attrs |= ptebit;
413 }
414 
415 static __inline void
416 moea64_pte_create(struct lpte *pt, uint64_t vsid, vm_offset_t va,
417     uint64_t pte_lo, int flags)
418 {
419 
420 	ASSERT_TABLE_LOCK();
421 
422 	/*
423 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
424 	 * set when the real pte is set in memory.
425 	 *
426 	 * Note: Don't set the valid bit for correct operation of tlb update.
427 	 */
428 	pt->pte_hi = (vsid << LPTE_VSID_SHIFT) |
429 	    (((uint64_t)(va & ADDR_PIDX) >> ADDR_API_SHFT64) & LPTE_API);
430 
431 	if (flags & PVO_LARGE)
432 		pt->pte_hi |= LPTE_BIG;
433 
434 	pt->pte_lo = pte_lo;
435 }
436 
437 static __inline uint64_t
438 moea64_calc_wimg(vm_offset_t pa, vm_memattr_t ma)
439 {
440 	uint64_t pte_lo;
441 	int i;
442 
443 	if (ma != VM_MEMATTR_DEFAULT) {
444 		switch (ma) {
445 		case VM_MEMATTR_UNCACHEABLE:
446 			return (LPTE_I | LPTE_G);
447 		case VM_MEMATTR_WRITE_COMBINING:
448 		case VM_MEMATTR_WRITE_BACK:
449 		case VM_MEMATTR_PREFETCHABLE:
450 			return (LPTE_I);
451 		case VM_MEMATTR_WRITE_THROUGH:
452 			return (LPTE_W | LPTE_M);
453 		}
454 	}
455 
456 	/*
457 	 * Assume the page is cache inhibited and access is guarded unless
458 	 * it's in our available memory array.
459 	 */
460 	pte_lo = LPTE_I | LPTE_G;
461 	for (i = 0; i < pregions_sz; i++) {
462 		if ((pa >= pregions[i].mr_start) &&
463 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
464 			pte_lo &= ~(LPTE_I | LPTE_G);
465 			pte_lo |= LPTE_M;
466 			break;
467 		}
468 	}
469 
470 	return pte_lo;
471 }
472 
473 /*
474  * Quick sort callout for comparing memory regions.
475  */
476 static int	om_cmp(const void *a, const void *b);
477 
478 static int
479 om_cmp(const void *a, const void *b)
480 {
481 	const struct	ofw_map *mapa;
482 	const struct	ofw_map *mapb;
483 
484 	mapa = a;
485 	mapb = b;
486 	if (mapa->om_pa_hi < mapb->om_pa_hi)
487 		return (-1);
488 	else if (mapa->om_pa_hi > mapb->om_pa_hi)
489 		return (1);
490 	else if (mapa->om_pa_lo < mapb->om_pa_lo)
491 		return (-1);
492 	else if (mapa->om_pa_lo > mapb->om_pa_lo)
493 		return (1);
494 	else
495 		return (0);
496 }
497 
498 static void
499 moea64_add_ofw_mappings(mmu_t mmup, phandle_t mmu, size_t sz)
500 {
501 	struct ofw_map	translations[sz/sizeof(struct ofw_map)];
502 	register_t	msr;
503 	vm_offset_t	off;
504 	vm_paddr_t	pa_base;
505 	int		i;
506 
507 	bzero(translations, sz);
508 	if (OF_getprop(mmu, "translations", translations, sz) == -1)
509 		panic("moea64_bootstrap: can't get ofw translations");
510 
511 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
512 	sz /= sizeof(*translations);
513 	qsort(translations, sz, sizeof (*translations), om_cmp);
514 
515 	for (i = 0; i < sz; i++) {
516 		CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
517 		    (uint32_t)(translations[i].om_pa_lo), translations[i].om_va,
518 		    translations[i].om_len);
519 
520 		if (translations[i].om_pa_lo % PAGE_SIZE)
521 			panic("OFW translation not page-aligned!");
522 
523 		pa_base = translations[i].om_pa_lo;
524 
525 	      #ifdef __powerpc64__
526 		pa_base += (vm_offset_t)translations[i].om_pa_hi << 32;
527 	      #else
528 		if (translations[i].om_pa_hi)
529 			panic("OFW translations above 32-bit boundary!");
530 	      #endif
531 
532 		/* Now enter the pages for this mapping */
533 
534 		DISABLE_TRANS(msr);
535 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
536 			if (moea64_pvo_find_va(kernel_pmap,
537 			    translations[i].om_va + off) != NULL)
538 				continue;
539 
540 			moea64_kenter(mmup, translations[i].om_va + off,
541 			    pa_base + off);
542 		}
543 		ENABLE_TRANS(msr);
544 	}
545 }
546 
547 #ifdef __powerpc64__
548 static void
549 moea64_probe_large_page(void)
550 {
551 	uint16_t pvr = mfpvr() >> 16;
552 
553 	switch (pvr) {
554 	case IBM970:
555 	case IBM970FX:
556 	case IBM970MP:
557 		powerpc_sync(); isync();
558 		mtspr(SPR_HID4, mfspr(SPR_HID4) & ~HID4_970_DISABLE_LG_PG);
559 		powerpc_sync(); isync();
560 
561 		/* FALLTHROUGH */
562 	case IBMCELLBE:
563 		moea64_large_page_size = 0x1000000; /* 16 MB */
564 		moea64_large_page_shift = 24;
565 		break;
566 	default:
567 		moea64_large_page_size = 0;
568 	}
569 
570 	moea64_large_page_mask = moea64_large_page_size - 1;
571 }
572 
573 static void
574 moea64_bootstrap_slb_prefault(vm_offset_t va, int large)
575 {
576 	struct slb *cache;
577 	struct slb entry;
578 	uint64_t esid, slbe;
579 	uint64_t i;
580 
581 	cache = PCPU_GET(slb);
582 	esid = va >> ADDR_SR_SHFT;
583 	slbe = (esid << SLBE_ESID_SHIFT) | SLBE_VALID;
584 
585 	for (i = 0; i < 64; i++) {
586 		if (cache[i].slbe == (slbe | i))
587 			return;
588 	}
589 
590 	entry.slbe = slbe;
591 	entry.slbv = KERNEL_VSID(esid) << SLBV_VSID_SHIFT;
592 	if (large)
593 		entry.slbv |= SLBV_L;
594 
595 	slb_insert_kernel(entry.slbe, entry.slbv);
596 }
597 #endif
598 
599 static void
600 moea64_setup_direct_map(mmu_t mmup, vm_offset_t kernelstart,
601     vm_offset_t kernelend)
602 {
603 	register_t msr;
604 	vm_paddr_t pa;
605 	vm_offset_t size, off;
606 	uint64_t pte_lo;
607 	int i;
608 
609 	if (moea64_large_page_size == 0)
610 		hw_direct_map = 0;
611 
612 	DISABLE_TRANS(msr);
613 	if (hw_direct_map) {
614 		PMAP_LOCK(kernel_pmap);
615 		for (i = 0; i < pregions_sz; i++) {
616 		  for (pa = pregions[i].mr_start; pa < pregions[i].mr_start +
617 		     pregions[i].mr_size; pa += moea64_large_page_size) {
618 			pte_lo = LPTE_M;
619 
620 			/*
621 			 * Set memory access as guarded if prefetch within
622 			 * the page could exit the available physmem area.
623 			 */
624 			if (pa & moea64_large_page_mask) {
625 				pa &= moea64_large_page_mask;
626 				pte_lo |= LPTE_G;
627 			}
628 			if (pa + moea64_large_page_size >
629 			    pregions[i].mr_start + pregions[i].mr_size)
630 				pte_lo |= LPTE_G;
631 
632 			moea64_pvo_enter(mmup, kernel_pmap, moea64_upvo_zone,
633 				    &moea64_pvo_kunmanaged, pa, pa,
634 				    pte_lo, PVO_WIRED | PVO_LARGE);
635 		  }
636 		}
637 		PMAP_UNLOCK(kernel_pmap);
638 	} else {
639 		size = sizeof(struct pvo_head) * moea64_pteg_count;
640 		off = (vm_offset_t)(moea64_pvo_table);
641 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
642 			moea64_kenter(mmup, pa, pa);
643 		size = BPVO_POOL_SIZE*sizeof(struct pvo_entry);
644 		off = (vm_offset_t)(moea64_bpvo_pool);
645 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
646 		moea64_kenter(mmup, pa, pa);
647 
648 		/*
649 		 * Map certain important things, like ourselves.
650 		 *
651 		 * NOTE: We do not map the exception vector space. That code is
652 		 * used only in real mode, and leaving it unmapped allows us to
653 		 * catch NULL pointer deferences, instead of making NULL a valid
654 		 * address.
655 		 */
656 
657 		for (pa = kernelstart & ~PAGE_MASK; pa < kernelend;
658 		    pa += PAGE_SIZE)
659 			moea64_kenter(mmup, pa, pa);
660 	}
661 	ENABLE_TRANS(msr);
662 }
663 
664 void
665 moea64_early_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
666 {
667 	int		i, j;
668 	vm_size_t	physsz, hwphyssz;
669 
670 #ifndef __powerpc64__
671 	/* We don't have a direct map since there is no BAT */
672 	hw_direct_map = 0;
673 
674 	/* Make sure battable is zero, since we have no BAT */
675 	for (i = 0; i < 16; i++) {
676 		battable[i].batu = 0;
677 		battable[i].batl = 0;
678 	}
679 #else
680 	moea64_probe_large_page();
681 
682 	/* Use a direct map if we have large page support */
683 	if (moea64_large_page_size > 0)
684 		hw_direct_map = 1;
685 	else
686 		hw_direct_map = 0;
687 #endif
688 
689 	/* Get physical memory regions from firmware */
690 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
691 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
692 
693 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
694 		panic("moea64_bootstrap: phys_avail too small");
695 
696 	phys_avail_count = 0;
697 	physsz = 0;
698 	hwphyssz = 0;
699 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
700 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
701 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
702 		    regions[i].mr_start + regions[i].mr_size,
703 		    regions[i].mr_size);
704 		if (hwphyssz != 0 &&
705 		    (physsz + regions[i].mr_size) >= hwphyssz) {
706 			if (physsz < hwphyssz) {
707 				phys_avail[j] = regions[i].mr_start;
708 				phys_avail[j + 1] = regions[i].mr_start +
709 				    hwphyssz - physsz;
710 				physsz = hwphyssz;
711 				phys_avail_count++;
712 			}
713 			break;
714 		}
715 		phys_avail[j] = regions[i].mr_start;
716 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
717 		phys_avail_count++;
718 		physsz += regions[i].mr_size;
719 	}
720 
721 	/* Check for overlap with the kernel and exception vectors */
722 	for (j = 0; j < 2*phys_avail_count; j+=2) {
723 		if (phys_avail[j] < EXC_LAST)
724 			phys_avail[j] += EXC_LAST;
725 
726 		if (kernelstart >= phys_avail[j] &&
727 		    kernelstart < phys_avail[j+1]) {
728 			if (kernelend < phys_avail[j+1]) {
729 				phys_avail[2*phys_avail_count] =
730 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
731 				phys_avail[2*phys_avail_count + 1] =
732 				    phys_avail[j+1];
733 				phys_avail_count++;
734 			}
735 
736 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
737 		}
738 
739 		if (kernelend >= phys_avail[j] &&
740 		    kernelend < phys_avail[j+1]) {
741 			if (kernelstart > phys_avail[j]) {
742 				phys_avail[2*phys_avail_count] = phys_avail[j];
743 				phys_avail[2*phys_avail_count + 1] =
744 				    kernelstart & ~PAGE_MASK;
745 				phys_avail_count++;
746 			}
747 
748 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
749 		}
750 	}
751 
752 	physmem = btoc(physsz);
753 
754 #ifdef PTEGCOUNT
755 	moea64_pteg_count = PTEGCOUNT;
756 #else
757 	moea64_pteg_count = 0x1000;
758 
759 	while (moea64_pteg_count < physmem)
760 		moea64_pteg_count <<= 1;
761 
762 	moea64_pteg_count >>= 1;
763 #endif /* PTEGCOUNT */
764 }
765 
766 void
767 moea64_mid_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
768 {
769 	vm_size_t	size;
770 	register_t	msr;
771 	int		i;
772 
773 	/*
774 	 * Set PTEG mask
775 	 */
776 	moea64_pteg_mask = moea64_pteg_count - 1;
777 
778 	/*
779 	 * Allocate pv/overflow lists.
780 	 */
781 	size = sizeof(struct pvo_head) * moea64_pteg_count;
782 
783 	moea64_pvo_table = (struct pvo_head *)moea64_bootstrap_alloc(size,
784 	    PAGE_SIZE);
785 	CTR1(KTR_PMAP, "moea64_bootstrap: PVO table at %p", moea64_pvo_table);
786 
787 	DISABLE_TRANS(msr);
788 	for (i = 0; i < moea64_pteg_count; i++)
789 		LIST_INIT(&moea64_pvo_table[i]);
790 	ENABLE_TRANS(msr);
791 
792 	/*
793 	 * Initialize the lock that synchronizes access to the pteg and pvo
794 	 * tables.
795 	 */
796 	mtx_init(&moea64_table_mutex, "pmap table", NULL, MTX_DEF |
797 	    MTX_RECURSE);
798 	mtx_init(&moea64_slb_mutex, "SLB table", NULL, MTX_DEF);
799 
800 	/*
801 	 * Initialise the unmanaged pvo pool.
802 	 */
803 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
804 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
805 	moea64_bpvo_pool_index = 0;
806 
807 	/*
808 	 * Make sure kernel vsid is allocated as well as VSID 0.
809 	 */
810 	#ifndef __powerpc64__
811 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NVSIDS - 1)) / VSID_NBPW]
812 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
813 	moea64_vsid_bitmap[0] |= 1;
814 	#endif
815 
816 	/*
817 	 * Initialize the kernel pmap (which is statically allocated).
818 	 */
819 	#ifdef __powerpc64__
820 	for (i = 0; i < 64; i++) {
821 		pcpup->pc_slb[i].slbv = 0;
822 		pcpup->pc_slb[i].slbe = 0;
823 	}
824 	#else
825 	for (i = 0; i < 16; i++)
826 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
827 	#endif
828 
829 	kernel_pmap->pmap_phys = kernel_pmap;
830 	kernel_pmap->pm_active = ~0;
831 
832 	PMAP_LOCK_INIT(kernel_pmap);
833 
834 	/*
835 	 * Now map in all the other buffers we allocated earlier
836 	 */
837 
838 	moea64_setup_direct_map(mmup, kernelstart, kernelend);
839 }
840 
841 void
842 moea64_late_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
843 {
844 	ihandle_t	mmui;
845 	phandle_t	chosen;
846 	phandle_t	mmu;
847 	size_t		sz;
848 	int		i;
849 	vm_offset_t	pa, va;
850 	void		*dpcpu;
851 
852 	/*
853 	 * Set up the Open Firmware pmap and add its mappings if not in real
854 	 * mode.
855 	 */
856 
857 	chosen = OF_finddevice("/chosen");
858 	if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1) {
859 	    mmu = OF_instance_to_package(mmui);
860 	    if (mmu == -1 || (sz = OF_getproplen(mmu, "translations")) == -1)
861 		sz = 0;
862 	    if (sz > 6144 /* tmpstksz - 2 KB headroom */)
863 		panic("moea64_bootstrap: too many ofw translations");
864 
865 	    if (sz > 0)
866 		moea64_add_ofw_mappings(mmup, mmu, sz);
867 	}
868 
869 	/*
870 	 * Calculate the last available physical address.
871 	 */
872 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
873 		;
874 	Maxmem = powerpc_btop(phys_avail[i + 1]);
875 
876 	/*
877 	 * Initialize MMU and remap early physical mappings
878 	 */
879 	MMU_CPU_BOOTSTRAP(mmup,0);
880 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
881 	pmap_bootstrapped++;
882 	bs_remap_earlyboot();
883 
884 	/*
885 	 * Set the start and end of kva.
886 	 */
887 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
888 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
889 
890 	/*
891 	 * Map the entire KVA range into the SLB. We must not fault there.
892 	 */
893 	#ifdef __powerpc64__
894 	for (va = virtual_avail; va < virtual_end; va += SEGMENT_LENGTH)
895 		moea64_bootstrap_slb_prefault(va, 0);
896 	#endif
897 
898 	/*
899 	 * Figure out how far we can extend virtual_end into segment 16
900 	 * without running into existing mappings. Segment 16 is guaranteed
901 	 * to contain neither RAM nor devices (at least on Apple hardware),
902 	 * but will generally contain some OFW mappings we should not
903 	 * step on.
904 	 */
905 
906 	#ifndef __powerpc64__	/* KVA is in high memory on PPC64 */
907 	PMAP_LOCK(kernel_pmap);
908 	while (virtual_end < VM_MAX_KERNEL_ADDRESS &&
909 	    moea64_pvo_find_va(kernel_pmap, virtual_end+1) == NULL)
910 		virtual_end += PAGE_SIZE;
911 	PMAP_UNLOCK(kernel_pmap);
912 	#endif
913 
914 	/*
915 	 * Allocate a kernel stack with a guard page for thread0 and map it
916 	 * into the kernel page map.
917 	 */
918 	pa = moea64_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, PAGE_SIZE);
919 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
920 	virtual_avail = va + KSTACK_PAGES * PAGE_SIZE;
921 	CTR2(KTR_PMAP, "moea64_bootstrap: kstack0 at %#x (%#x)", pa, va);
922 	thread0.td_kstack = va;
923 	thread0.td_kstack_pages = KSTACK_PAGES;
924 	for (i = 0; i < KSTACK_PAGES; i++) {
925 		moea64_kenter(mmup, va, pa);
926 		pa += PAGE_SIZE;
927 		va += PAGE_SIZE;
928 	}
929 
930 	/*
931 	 * Allocate virtual address space for the message buffer.
932 	 */
933 	pa = msgbuf_phys = moea64_bootstrap_alloc(msgbufsize, PAGE_SIZE);
934 	msgbufp = (struct msgbuf *)virtual_avail;
935 	va = virtual_avail;
936 	virtual_avail += round_page(msgbufsize);
937 	while (va < virtual_avail) {
938 		moea64_kenter(mmup, va, pa);
939 		pa += PAGE_SIZE;
940 		va += PAGE_SIZE;
941 	}
942 
943 	/*
944 	 * Allocate virtual address space for the dynamic percpu area.
945 	 */
946 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
947 	dpcpu = (void *)virtual_avail;
948 	va = virtual_avail;
949 	virtual_avail += DPCPU_SIZE;
950 	while (va < virtual_avail) {
951 		moea64_kenter(mmup, va, pa);
952 		pa += PAGE_SIZE;
953 		va += PAGE_SIZE;
954 	}
955 	dpcpu_init(dpcpu, 0);
956 
957 	/*
958 	 * Allocate some things for page zeroing. We put this directly
959 	 * in the page table, marked with LPTE_LOCKED, to avoid any
960 	 * of the PVO book-keeping or other parts of the VM system
961 	 * from even knowing that this hack exists.
962 	 */
963 
964 	if (!hw_direct_map) {
965 		mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL,
966 		    MTX_DEF);
967 		for (i = 0; i < 2; i++) {
968 			moea64_scratchpage_va[i] = (virtual_end+1) - PAGE_SIZE;
969 			virtual_end -= PAGE_SIZE;
970 
971 			moea64_kenter(mmup, moea64_scratchpage_va[i], 0);
972 
973 			moea64_scratchpage_pvo[i] = moea64_pvo_find_va(
974 			    kernel_pmap, (vm_offset_t)moea64_scratchpage_va[i]);
975 			LOCK_TABLE();
976 			moea64_scratchpage_pte[i] = MOEA64_PVO_TO_PTE(
977 			    mmup, moea64_scratchpage_pvo[i]);
978 			moea64_scratchpage_pvo[i]->pvo_pte.lpte.pte_hi
979 			    |= LPTE_LOCKED;
980 			MOEA64_PTE_CHANGE(mmup, moea64_scratchpage_pte[i],
981 			    &moea64_scratchpage_pvo[i]->pvo_pte.lpte,
982 			    moea64_scratchpage_pvo[i]->pvo_vpn);
983 			UNLOCK_TABLE();
984 		}
985 	}
986 }
987 
988 /*
989  * Activate a user pmap.  The pmap must be activated before its address
990  * space can be accessed in any way.
991  */
992 void
993 moea64_activate(mmu_t mmu, struct thread *td)
994 {
995 	pmap_t	pm;
996 
997 	pm = &td->td_proc->p_vmspace->vm_pmap;
998 	pm->pm_active |= PCPU_GET(cpumask);
999 
1000 	#ifdef __powerpc64__
1001 	PCPU_SET(userslb, pm->pm_slb);
1002 	#else
1003 	PCPU_SET(curpmap, pm->pmap_phys);
1004 	#endif
1005 }
1006 
1007 void
1008 moea64_deactivate(mmu_t mmu, struct thread *td)
1009 {
1010 	pmap_t	pm;
1011 
1012 	pm = &td->td_proc->p_vmspace->vm_pmap;
1013 	pm->pm_active &= ~(PCPU_GET(cpumask));
1014 	#ifdef __powerpc64__
1015 	PCPU_SET(userslb, NULL);
1016 	#else
1017 	PCPU_SET(curpmap, NULL);
1018 	#endif
1019 }
1020 
1021 void
1022 moea64_change_wiring(mmu_t mmu, pmap_t pm, vm_offset_t va, boolean_t wired)
1023 {
1024 	struct	pvo_entry *pvo;
1025 	uintptr_t pt;
1026 	uint64_t vsid;
1027 	int	i, ptegidx;
1028 
1029 	PMAP_LOCK(pm);
1030 	pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
1031 
1032 	if (pvo != NULL) {
1033 		LOCK_TABLE();
1034 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
1035 
1036 		if (wired) {
1037 			if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1038 				pm->pm_stats.wired_count++;
1039 			pvo->pvo_vaddr |= PVO_WIRED;
1040 			pvo->pvo_pte.lpte.pte_hi |= LPTE_WIRED;
1041 		} else {
1042 			if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1043 				pm->pm_stats.wired_count--;
1044 			pvo->pvo_vaddr &= ~PVO_WIRED;
1045 			pvo->pvo_pte.lpte.pte_hi &= ~LPTE_WIRED;
1046 		}
1047 
1048 		if (pt != -1) {
1049 			/* Update wiring flag in page table. */
1050 			MOEA64_PTE_CHANGE(mmu, pt, &pvo->pvo_pte.lpte,
1051 			    pvo->pvo_vpn);
1052 		} else if (wired) {
1053 			/*
1054 			 * If we are wiring the page, and it wasn't in the
1055 			 * page table before, add it.
1056 			 */
1057 			vsid = PVO_VSID(pvo);
1058 			ptegidx = va_to_pteg(vsid, PVO_VADDR(pvo),
1059 			    pvo->pvo_vaddr & PVO_LARGE);
1060 
1061 			i = MOEA64_PTE_INSERT(mmu, ptegidx, &pvo->pvo_pte.lpte);
1062 
1063 			if (i >= 0) {
1064 				PVO_PTEGIDX_CLR(pvo);
1065 				PVO_PTEGIDX_SET(pvo, i);
1066 			}
1067 		}
1068 
1069 		UNLOCK_TABLE();
1070 	}
1071 	PMAP_UNLOCK(pm);
1072 }
1073 
1074 /*
1075  * This goes through and sets the physical address of our
1076  * special scratch PTE to the PA we want to zero or copy. Because
1077  * of locking issues (this can get called in pvo_enter() by
1078  * the UMA allocator), we can't use most other utility functions here
1079  */
1080 
1081 static __inline
1082 void moea64_set_scratchpage_pa(mmu_t mmup, int which, vm_offset_t pa) {
1083 
1084 	KASSERT(!hw_direct_map, ("Using OEA64 scratchpage with a direct map!"));
1085 	mtx_assert(&moea64_scratchpage_mtx, MA_OWNED);
1086 
1087 	moea64_scratchpage_pvo[which]->pvo_pte.lpte.pte_lo &=
1088 	    ~(LPTE_WIMG | LPTE_RPGN);
1089 	moea64_scratchpage_pvo[which]->pvo_pte.lpte.pte_lo |=
1090 	    moea64_calc_wimg(pa, VM_MEMATTR_DEFAULT) | (uint64_t)pa;
1091 	MOEA64_PTE_CHANGE(mmup, moea64_scratchpage_pte[which],
1092 	    &moea64_scratchpage_pvo[which]->pvo_pte.lpte,
1093 	    moea64_scratchpage_pvo[which]->pvo_vpn);
1094 	isync();
1095 }
1096 
1097 void
1098 moea64_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1099 {
1100 	vm_offset_t	dst;
1101 	vm_offset_t	src;
1102 
1103 	dst = VM_PAGE_TO_PHYS(mdst);
1104 	src = VM_PAGE_TO_PHYS(msrc);
1105 
1106 	if (hw_direct_map) {
1107 		kcopy((void *)src, (void *)dst, PAGE_SIZE);
1108 	} else {
1109 		mtx_lock(&moea64_scratchpage_mtx);
1110 
1111 		moea64_set_scratchpage_pa(mmu, 0, src);
1112 		moea64_set_scratchpage_pa(mmu, 1, dst);
1113 
1114 		kcopy((void *)moea64_scratchpage_va[0],
1115 		    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1116 
1117 		mtx_unlock(&moea64_scratchpage_mtx);
1118 	}
1119 }
1120 
1121 void
1122 moea64_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1123 {
1124 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1125 
1126 	if (size + off > PAGE_SIZE)
1127 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1128 
1129 	if (hw_direct_map) {
1130 		bzero((caddr_t)pa + off, size);
1131 	} else {
1132 		mtx_lock(&moea64_scratchpage_mtx);
1133 		moea64_set_scratchpage_pa(mmu, 0, pa);
1134 		bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1135 		mtx_unlock(&moea64_scratchpage_mtx);
1136 	}
1137 }
1138 
1139 /*
1140  * Zero a page of physical memory by temporarily mapping it
1141  */
1142 void
1143 moea64_zero_page(mmu_t mmu, vm_page_t m)
1144 {
1145 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1146 	vm_offset_t va, off;
1147 
1148 	if (!hw_direct_map) {
1149 		mtx_lock(&moea64_scratchpage_mtx);
1150 
1151 		moea64_set_scratchpage_pa(mmu, 0, pa);
1152 		va = moea64_scratchpage_va[0];
1153 	} else {
1154 		va = pa;
1155 	}
1156 
1157 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1158 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1159 
1160 	if (!hw_direct_map)
1161 		mtx_unlock(&moea64_scratchpage_mtx);
1162 }
1163 
1164 void
1165 moea64_zero_page_idle(mmu_t mmu, vm_page_t m)
1166 {
1167 
1168 	moea64_zero_page(mmu, m);
1169 }
1170 
1171 /*
1172  * Map the given physical page at the specified virtual address in the
1173  * target pmap with the protection requested.  If specified the page
1174  * will be wired down.
1175  */
1176 void
1177 moea64_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m,
1178     vm_prot_t prot, boolean_t wired)
1179 {
1180 
1181 	vm_page_lock_queues();
1182 	PMAP_LOCK(pmap);
1183 	moea64_enter_locked(mmu, pmap, va, m, prot, wired);
1184 	vm_page_unlock_queues();
1185 	PMAP_UNLOCK(pmap);
1186 }
1187 
1188 /*
1189  * Map the given physical page at the specified virtual address in the
1190  * target pmap with the protection requested.  If specified the page
1191  * will be wired down.
1192  *
1193  * The page queues and pmap must be locked.
1194  */
1195 
1196 static void
1197 moea64_enter_locked(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m,
1198     vm_prot_t prot, boolean_t wired)
1199 {
1200 	struct		pvo_head *pvo_head;
1201 	uma_zone_t	zone;
1202 	vm_page_t	pg;
1203 	uint64_t	pte_lo;
1204 	u_int		pvo_flags;
1205 	int		error;
1206 
1207 	if (!moea64_initialized) {
1208 		pvo_head = &moea64_pvo_kunmanaged;
1209 		pg = NULL;
1210 		zone = moea64_upvo_zone;
1211 		pvo_flags = 0;
1212 	} else {
1213 		pvo_head = vm_page_to_pvoh(m);
1214 		pg = m;
1215 		zone = moea64_mpvo_zone;
1216 		pvo_flags = PVO_MANAGED;
1217 	}
1218 
1219 	if (pmap_bootstrapped)
1220 		mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1221 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1222 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0 ||
1223 	    (m->oflags & VPO_BUSY) != 0 || VM_OBJECT_LOCKED(m->object),
1224 	    ("moea64_enter_locked: page %p is not busy", m));
1225 
1226 	/* XXX change the pvo head for fake pages */
1227 	if ((m->flags & PG_FICTITIOUS) == PG_FICTITIOUS) {
1228 		pvo_flags &= ~PVO_MANAGED;
1229 		pvo_head = &moea64_pvo_kunmanaged;
1230 		zone = moea64_upvo_zone;
1231 	}
1232 
1233 	pte_lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1234 
1235 	if (prot & VM_PROT_WRITE) {
1236 		pte_lo |= LPTE_BW;
1237 		if (pmap_bootstrapped &&
1238 		    (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0)
1239 			vm_page_flag_set(m, PG_WRITEABLE);
1240 	} else
1241 		pte_lo |= LPTE_BR;
1242 
1243 	if ((prot & VM_PROT_EXECUTE) == 0)
1244 		pte_lo |= LPTE_NOEXEC;
1245 
1246 	if (wired)
1247 		pvo_flags |= PVO_WIRED;
1248 
1249 	if ((m->flags & PG_FICTITIOUS) != 0)
1250 		pvo_flags |= PVO_FAKE;
1251 
1252 	error = moea64_pvo_enter(mmu, pmap, zone, pvo_head, va,
1253 	    VM_PAGE_TO_PHYS(m), pte_lo, pvo_flags);
1254 
1255 	/*
1256 	 * Flush the page from the instruction cache if this page is
1257 	 * mapped executable and cacheable.
1258 	 */
1259 	if ((pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0)
1260 		moea64_syncicache(mmu, pmap, va, VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1261 }
1262 
1263 static void
1264 moea64_syncicache(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_offset_t pa,
1265     vm_size_t sz)
1266 {
1267 
1268 	/*
1269 	 * This is much trickier than on older systems because
1270 	 * we can't sync the icache on physical addresses directly
1271 	 * without a direct map. Instead we check a couple of cases
1272 	 * where the memory is already mapped in and, failing that,
1273 	 * use the same trick we use for page zeroing to create
1274 	 * a temporary mapping for this physical address.
1275 	 */
1276 
1277 	if (!pmap_bootstrapped) {
1278 		/*
1279 		 * If PMAP is not bootstrapped, we are likely to be
1280 		 * in real mode.
1281 		 */
1282 		__syncicache((void *)pa, sz);
1283 	} else if (pmap == kernel_pmap) {
1284 		__syncicache((void *)va, sz);
1285 	} else if (hw_direct_map) {
1286 		__syncicache((void *)pa, sz);
1287 	} else {
1288 		/* Use the scratch page to set up a temp mapping */
1289 
1290 		mtx_lock(&moea64_scratchpage_mtx);
1291 
1292 		moea64_set_scratchpage_pa(mmu, 1, pa & ~ADDR_POFF);
1293 		__syncicache((void *)(moea64_scratchpage_va[1] +
1294 		    (va & ADDR_POFF)), sz);
1295 
1296 		mtx_unlock(&moea64_scratchpage_mtx);
1297 	}
1298 }
1299 
1300 /*
1301  * Maps a sequence of resident pages belonging to the same object.
1302  * The sequence begins with the given page m_start.  This page is
1303  * mapped at the given virtual address start.  Each subsequent page is
1304  * mapped at a virtual address that is offset from start by the same
1305  * amount as the page is offset from m_start within the object.  The
1306  * last page in the sequence is the page with the largest offset from
1307  * m_start that can be mapped at a virtual address less than the given
1308  * virtual address end.  Not every virtual page between start and end
1309  * is mapped; only those for which a resident page exists with the
1310  * corresponding offset from m_start are mapped.
1311  */
1312 void
1313 moea64_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1314     vm_page_t m_start, vm_prot_t prot)
1315 {
1316 	vm_page_t m;
1317 	vm_pindex_t diff, psize;
1318 
1319 	psize = atop(end - start);
1320 	m = m_start;
1321 	vm_page_lock_queues();
1322 	PMAP_LOCK(pm);
1323 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1324 		moea64_enter_locked(mmu, pm, start + ptoa(diff), m, prot &
1325 		    (VM_PROT_READ | VM_PROT_EXECUTE), FALSE);
1326 		m = TAILQ_NEXT(m, listq);
1327 	}
1328 	vm_page_unlock_queues();
1329 	PMAP_UNLOCK(pm);
1330 }
1331 
1332 void
1333 moea64_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1334     vm_prot_t prot)
1335 {
1336 
1337 	vm_page_lock_queues();
1338 	PMAP_LOCK(pm);
1339 	moea64_enter_locked(mmu, pm, va, m,
1340 	    prot & (VM_PROT_READ | VM_PROT_EXECUTE), FALSE);
1341 	vm_page_unlock_queues();
1342 	PMAP_UNLOCK(pm);
1343 }
1344 
1345 vm_paddr_t
1346 moea64_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1347 {
1348 	struct	pvo_entry *pvo;
1349 	vm_paddr_t pa;
1350 
1351 	PMAP_LOCK(pm);
1352 	pvo = moea64_pvo_find_va(pm, va);
1353 	if (pvo == NULL)
1354 		pa = 0;
1355 	else
1356 		pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) |
1357 		    (va - PVO_VADDR(pvo));
1358 	PMAP_UNLOCK(pm);
1359 	return (pa);
1360 }
1361 
1362 /*
1363  * Atomically extract and hold the physical page with the given
1364  * pmap and virtual address pair if that mapping permits the given
1365  * protection.
1366  */
1367 vm_page_t
1368 moea64_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1369 {
1370 	struct	pvo_entry *pvo;
1371 	vm_page_t m;
1372         vm_paddr_t pa;
1373 
1374 	m = NULL;
1375 	pa = 0;
1376 	PMAP_LOCK(pmap);
1377 retry:
1378 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1379 	if (pvo != NULL && (pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) &&
1380 	    ((pvo->pvo_pte.lpte.pte_lo & LPTE_PP) == LPTE_RW ||
1381 	     (prot & VM_PROT_WRITE) == 0)) {
1382 		if (vm_page_pa_tryrelock(pmap,
1383 			pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN, &pa))
1384 			goto retry;
1385 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN);
1386 		vm_page_hold(m);
1387 	}
1388 	PA_UNLOCK_COND(pa);
1389 	PMAP_UNLOCK(pmap);
1390 	return (m);
1391 }
1392 
1393 static mmu_t installed_mmu;
1394 
1395 static void *
1396 moea64_uma_page_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
1397 {
1398 	/*
1399 	 * This entire routine is a horrible hack to avoid bothering kmem
1400 	 * for new KVA addresses. Because this can get called from inside
1401 	 * kmem allocation routines, calling kmem for a new address here
1402 	 * can lead to multiply locking non-recursive mutexes.
1403 	 */
1404 	static vm_pindex_t color;
1405         vm_offset_t va;
1406 
1407         vm_page_t m;
1408         int pflags, needed_lock;
1409 
1410 	*flags = UMA_SLAB_PRIV;
1411 	needed_lock = !PMAP_LOCKED(kernel_pmap);
1412 
1413 	if (needed_lock)
1414 		PMAP_LOCK(kernel_pmap);
1415 
1416         if ((wait & (M_NOWAIT|M_USE_RESERVE)) == M_NOWAIT)
1417                 pflags = VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED;
1418         else
1419                 pflags = VM_ALLOC_SYSTEM | VM_ALLOC_WIRED;
1420         if (wait & M_ZERO)
1421                 pflags |= VM_ALLOC_ZERO;
1422 
1423         for (;;) {
1424                 m = vm_page_alloc(NULL, color++, pflags | VM_ALLOC_NOOBJ);
1425                 if (m == NULL) {
1426                         if (wait & M_NOWAIT)
1427                                 return (NULL);
1428                         VM_WAIT;
1429                 } else
1430                         break;
1431         }
1432 
1433 	va = VM_PAGE_TO_PHYS(m);
1434 
1435 	moea64_pvo_enter(installed_mmu, kernel_pmap, moea64_upvo_zone,
1436 	    &moea64_pvo_kunmanaged, va, VM_PAGE_TO_PHYS(m), LPTE_M,
1437 	    PVO_WIRED | PVO_BOOTSTRAP);
1438 
1439 	if (needed_lock)
1440 		PMAP_UNLOCK(kernel_pmap);
1441 
1442 	if ((wait & M_ZERO) && (m->flags & PG_ZERO) == 0)
1443                 bzero((void *)va, PAGE_SIZE);
1444 
1445 	return (void *)va;
1446 }
1447 
1448 void
1449 moea64_init(mmu_t mmu)
1450 {
1451 
1452 	CTR0(KTR_PMAP, "moea64_init");
1453 
1454 	moea64_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1455 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1456 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1457 	moea64_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1458 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1459 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1460 
1461 	if (!hw_direct_map) {
1462 		installed_mmu = mmu;
1463 		uma_zone_set_allocf(moea64_upvo_zone,moea64_uma_page_alloc);
1464 		uma_zone_set_allocf(moea64_mpvo_zone,moea64_uma_page_alloc);
1465 	}
1466 
1467 	moea64_initialized = TRUE;
1468 }
1469 
1470 boolean_t
1471 moea64_is_referenced(mmu_t mmu, vm_page_t m)
1472 {
1473 
1474 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1475 	    ("moea64_is_referenced: page %p is not managed", m));
1476 	return (moea64_query_bit(mmu, m, PTE_REF));
1477 }
1478 
1479 boolean_t
1480 moea64_is_modified(mmu_t mmu, vm_page_t m)
1481 {
1482 
1483 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1484 	    ("moea64_is_modified: page %p is not managed", m));
1485 
1486 	/*
1487 	 * If the page is not VPO_BUSY, then PG_WRITEABLE cannot be
1488 	 * concurrently set while the object is locked.  Thus, if PG_WRITEABLE
1489 	 * is clear, no PTEs can have LPTE_CHG set.
1490 	 */
1491 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1492 	if ((m->oflags & VPO_BUSY) == 0 &&
1493 	    (m->flags & PG_WRITEABLE) == 0)
1494 		return (FALSE);
1495 	return (moea64_query_bit(mmu, m, LPTE_CHG));
1496 }
1497 
1498 boolean_t
1499 moea64_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1500 {
1501 	struct pvo_entry *pvo;
1502 	boolean_t rv;
1503 
1504 	PMAP_LOCK(pmap);
1505 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1506 	rv = pvo == NULL || (pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) == 0;
1507 	PMAP_UNLOCK(pmap);
1508 	return (rv);
1509 }
1510 
1511 void
1512 moea64_clear_reference(mmu_t mmu, vm_page_t m)
1513 {
1514 
1515 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1516 	    ("moea64_clear_reference: page %p is not managed", m));
1517 	moea64_clear_bit(mmu, m, LPTE_REF);
1518 }
1519 
1520 void
1521 moea64_clear_modify(mmu_t mmu, vm_page_t m)
1522 {
1523 
1524 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1525 	    ("moea64_clear_modify: page %p is not managed", m));
1526 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1527 	KASSERT((m->oflags & VPO_BUSY) == 0,
1528 	    ("moea64_clear_modify: page %p is busy", m));
1529 
1530 	/*
1531 	 * If the page is not PG_WRITEABLE, then no PTEs can have LPTE_CHG
1532 	 * set.  If the object containing the page is locked and the page is
1533 	 * not VPO_BUSY, then PG_WRITEABLE cannot be concurrently set.
1534 	 */
1535 	if ((m->flags & PG_WRITEABLE) == 0)
1536 		return;
1537 	moea64_clear_bit(mmu, m, LPTE_CHG);
1538 }
1539 
1540 /*
1541  * Clear the write and modified bits in each of the given page's mappings.
1542  */
1543 void
1544 moea64_remove_write(mmu_t mmu, vm_page_t m)
1545 {
1546 	struct	pvo_entry *pvo;
1547 	uintptr_t pt;
1548 	pmap_t	pmap;
1549 	uint64_t lo;
1550 
1551 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1552 	    ("moea64_remove_write: page %p is not managed", m));
1553 
1554 	/*
1555 	 * If the page is not VPO_BUSY, then PG_WRITEABLE cannot be set by
1556 	 * another thread while the object is locked.  Thus, if PG_WRITEABLE
1557 	 * is clear, no page table entries need updating.
1558 	 */
1559 	VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED);
1560 	if ((m->oflags & VPO_BUSY) == 0 &&
1561 	    (m->flags & PG_WRITEABLE) == 0)
1562 		return;
1563 	vm_page_lock_queues();
1564 	lo = moea64_attr_fetch(m);
1565 	powerpc_sync();
1566 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1567 		pmap = pvo->pvo_pmap;
1568 		PMAP_LOCK(pmap);
1569 		LOCK_TABLE();
1570 		if ((pvo->pvo_pte.lpte.pte_lo & LPTE_PP) != LPTE_BR) {
1571 			pt = MOEA64_PVO_TO_PTE(mmu, pvo);
1572 			pvo->pvo_pte.lpte.pte_lo &= ~LPTE_PP;
1573 			pvo->pvo_pte.lpte.pte_lo |= LPTE_BR;
1574 			if (pt != -1) {
1575 				MOEA64_PTE_SYNCH(mmu, pt, &pvo->pvo_pte.lpte);
1576 				lo |= pvo->pvo_pte.lpte.pte_lo;
1577 				pvo->pvo_pte.lpte.pte_lo &= ~LPTE_CHG;
1578 				MOEA64_PTE_CHANGE(mmu, pt,
1579 				    &pvo->pvo_pte.lpte, pvo->pvo_vpn);
1580 				if (pvo->pvo_pmap == kernel_pmap)
1581 					isync();
1582 			}
1583 		}
1584 		UNLOCK_TABLE();
1585 		PMAP_UNLOCK(pmap);
1586 	}
1587 	if ((lo & LPTE_CHG) != 0) {
1588 		moea64_attr_clear(m, LPTE_CHG);
1589 		vm_page_dirty(m);
1590 	}
1591 	vm_page_flag_clear(m, PG_WRITEABLE);
1592 	vm_page_unlock_queues();
1593 }
1594 
1595 /*
1596  *	moea64_ts_referenced:
1597  *
1598  *	Return a count of reference bits for a page, clearing those bits.
1599  *	It is not necessary for every reference bit to be cleared, but it
1600  *	is necessary that 0 only be returned when there are truly no
1601  *	reference bits set.
1602  *
1603  *	XXX: The exact number of bits to check and clear is a matter that
1604  *	should be tested and standardized at some point in the future for
1605  *	optimal aging of shared pages.
1606  */
1607 boolean_t
1608 moea64_ts_referenced(mmu_t mmu, vm_page_t m)
1609 {
1610 
1611 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1612 	    ("moea64_ts_referenced: page %p is not managed", m));
1613 	return (moea64_clear_bit(mmu, m, LPTE_REF));
1614 }
1615 
1616 /*
1617  * Modify the WIMG settings of all mappings for a page.
1618  */
1619 void
1620 moea64_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1621 {
1622 	struct	pvo_entry *pvo;
1623 	struct  pvo_head *pvo_head;
1624 	uintptr_t pt;
1625 	pmap_t	pmap;
1626 	uint64_t lo;
1627 
1628 	if (m->flags & PG_FICTITIOUS) {
1629 		m->md.mdpg_cache_attrs = ma;
1630 		return;
1631 	}
1632 
1633 	vm_page_lock_queues();
1634 	pvo_head = vm_page_to_pvoh(m);
1635 	lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1636 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1637 		pmap = pvo->pvo_pmap;
1638 		PMAP_LOCK(pmap);
1639 		LOCK_TABLE();
1640 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
1641 		pvo->pvo_pte.lpte.pte_lo &= ~LPTE_WIMG;
1642 		pvo->pvo_pte.lpte.pte_lo |= lo;
1643 		if (pt != -1) {
1644 			MOEA64_PTE_CHANGE(mmu, pt, &pvo->pvo_pte.lpte,
1645 			    pvo->pvo_vpn);
1646 			if (pvo->pvo_pmap == kernel_pmap)
1647 				isync();
1648 		}
1649 		UNLOCK_TABLE();
1650 		PMAP_UNLOCK(pmap);
1651 	}
1652 	m->md.mdpg_cache_attrs = ma;
1653 	vm_page_unlock_queues();
1654 }
1655 
1656 /*
1657  * Map a wired page into kernel virtual address space.
1658  */
1659 void
1660 moea64_kenter_attr(mmu_t mmu, vm_offset_t va, vm_offset_t pa, vm_memattr_t ma)
1661 {
1662 	uint64_t	pte_lo;
1663 	int		error;
1664 
1665 	pte_lo = moea64_calc_wimg(pa, ma);
1666 
1667 	PMAP_LOCK(kernel_pmap);
1668 	error = moea64_pvo_enter(mmu, kernel_pmap, moea64_upvo_zone,
1669 	    &moea64_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1670 
1671 	if (error != 0 && error != ENOENT)
1672 		panic("moea64_kenter: failed to enter va %#zx pa %#zx: %d", va,
1673 		    pa, error);
1674 
1675 	/*
1676 	 * Flush the memory from the instruction cache.
1677 	 */
1678 	if ((pte_lo & (LPTE_I | LPTE_G)) == 0)
1679 		__syncicache((void *)va, PAGE_SIZE);
1680 	PMAP_UNLOCK(kernel_pmap);
1681 }
1682 
1683 void
1684 moea64_kenter(mmu_t mmu, vm_offset_t va, vm_offset_t pa)
1685 {
1686 
1687 	moea64_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1688 }
1689 
1690 /*
1691  * Extract the physical page address associated with the given kernel virtual
1692  * address.
1693  */
1694 vm_offset_t
1695 moea64_kextract(mmu_t mmu, vm_offset_t va)
1696 {
1697 	struct		pvo_entry *pvo;
1698 	vm_paddr_t pa;
1699 
1700 	/*
1701 	 * Shortcut the direct-mapped case when applicable.  We never put
1702 	 * anything but 1:1 mappings below VM_MIN_KERNEL_ADDRESS.
1703 	 */
1704 	if (va < VM_MIN_KERNEL_ADDRESS)
1705 		return (va);
1706 
1707 	PMAP_LOCK(kernel_pmap);
1708 	pvo = moea64_pvo_find_va(kernel_pmap, va);
1709 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found for %#" PRIxPTR,
1710 	    va));
1711 	pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) + (va - PVO_VADDR(pvo));
1712 	PMAP_UNLOCK(kernel_pmap);
1713 	return (pa);
1714 }
1715 
1716 /*
1717  * Remove a wired page from kernel virtual address space.
1718  */
1719 void
1720 moea64_kremove(mmu_t mmu, vm_offset_t va)
1721 {
1722 	moea64_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1723 }
1724 
1725 /*
1726  * Map a range of physical addresses into kernel virtual address space.
1727  *
1728  * The value passed in *virt is a suggested virtual address for the mapping.
1729  * Architectures which can support a direct-mapped physical to virtual region
1730  * can return the appropriate address within that region, leaving '*virt'
1731  * unchanged.  We cannot and therefore do not; *virt is updated with the
1732  * first usable address after the mapped region.
1733  */
1734 vm_offset_t
1735 moea64_map(mmu_t mmu, vm_offset_t *virt, vm_offset_t pa_start,
1736     vm_offset_t pa_end, int prot)
1737 {
1738 	vm_offset_t	sva, va;
1739 
1740 	sva = *virt;
1741 	va = sva;
1742 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1743 		moea64_kenter(mmu, va, pa_start);
1744 	*virt = va;
1745 
1746 	return (sva);
1747 }
1748 
1749 /*
1750  * Returns true if the pmap's pv is one of the first
1751  * 16 pvs linked to from this page.  This count may
1752  * be changed upwards or downwards in the future; it
1753  * is only necessary that true be returned for a small
1754  * subset of pmaps for proper page aging.
1755  */
1756 boolean_t
1757 moea64_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1758 {
1759         int loops;
1760 	struct pvo_entry *pvo;
1761 	boolean_t rv;
1762 
1763 	KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0,
1764 	    ("moea64_page_exists_quick: page %p is not managed", m));
1765 	loops = 0;
1766 	rv = FALSE;
1767 	vm_page_lock_queues();
1768 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1769 		if (pvo->pvo_pmap == pmap) {
1770 			rv = TRUE;
1771 			break;
1772 		}
1773 		if (++loops >= 16)
1774 			break;
1775 	}
1776 	vm_page_unlock_queues();
1777 	return (rv);
1778 }
1779 
1780 /*
1781  * Return the number of managed mappings to the given physical page
1782  * that are wired.
1783  */
1784 int
1785 moea64_page_wired_mappings(mmu_t mmu, vm_page_t m)
1786 {
1787 	struct pvo_entry *pvo;
1788 	int count;
1789 
1790 	count = 0;
1791 	if ((m->flags & PG_FICTITIOUS) != 0)
1792 		return (count);
1793 	vm_page_lock_queues();
1794 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1795 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1796 			count++;
1797 	vm_page_unlock_queues();
1798 	return (count);
1799 }
1800 
1801 static uintptr_t	moea64_vsidcontext;
1802 
1803 uintptr_t
1804 moea64_get_unique_vsid(void) {
1805 	u_int entropy;
1806 	register_t hash;
1807 	uint32_t mask;
1808 	int i;
1809 
1810 	entropy = 0;
1811 	__asm __volatile("mftb %0" : "=r"(entropy));
1812 
1813 	mtx_lock(&moea64_slb_mutex);
1814 	for (i = 0; i < NVSIDS; i += VSID_NBPW) {
1815 		u_int	n;
1816 
1817 		/*
1818 		 * Create a new value by mutiplying by a prime and adding in
1819 		 * entropy from the timebase register.  This is to make the
1820 		 * VSID more random so that the PT hash function collides
1821 		 * less often.  (Note that the prime casues gcc to do shifts
1822 		 * instead of a multiply.)
1823 		 */
1824 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
1825 		hash = moea64_vsidcontext & (NVSIDS - 1);
1826 		if (hash == 0)		/* 0 is special, avoid it */
1827 			continue;
1828 		n = hash >> 5;
1829 		mask = 1 << (hash & (VSID_NBPW - 1));
1830 		hash = (moea64_vsidcontext & VSID_HASHMASK);
1831 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
1832 			/* anything free in this bucket? */
1833 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
1834 				entropy = (moea64_vsidcontext >> 20);
1835 				continue;
1836 			}
1837 			i = ffs(~moea64_vsid_bitmap[n]) - 1;
1838 			mask = 1 << i;
1839 			hash &= VSID_HASHMASK & ~(VSID_NBPW - 1);
1840 			hash |= i;
1841 		}
1842 		KASSERT(!(moea64_vsid_bitmap[n] & mask),
1843 		    ("Allocating in-use VSID %#zx\n", hash));
1844 		moea64_vsid_bitmap[n] |= mask;
1845 		mtx_unlock(&moea64_slb_mutex);
1846 		return (hash);
1847 	}
1848 
1849 	mtx_unlock(&moea64_slb_mutex);
1850 	panic("%s: out of segments",__func__);
1851 }
1852 
1853 #ifdef __powerpc64__
1854 void
1855 moea64_pinit(mmu_t mmu, pmap_t pmap)
1856 {
1857 	PMAP_LOCK_INIT(pmap);
1858 
1859 	pmap->pm_slb_tree_root = slb_alloc_tree();
1860 	pmap->pm_slb = slb_alloc_user_cache();
1861 	pmap->pm_slb_len = 0;
1862 }
1863 #else
1864 void
1865 moea64_pinit(mmu_t mmu, pmap_t pmap)
1866 {
1867 	int	i;
1868 	uint32_t hash;
1869 
1870 	PMAP_LOCK_INIT(pmap);
1871 
1872 	if (pmap_bootstrapped)
1873 		pmap->pmap_phys = (pmap_t)moea64_kextract(mmu,
1874 		    (vm_offset_t)pmap);
1875 	else
1876 		pmap->pmap_phys = pmap;
1877 
1878 	/*
1879 	 * Allocate some segment registers for this pmap.
1880 	 */
1881 	hash = moea64_get_unique_vsid();
1882 
1883 	for (i = 0; i < 16; i++)
1884 		pmap->pm_sr[i] = VSID_MAKE(i, hash);
1885 
1886 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_pinit: pm_sr[0] = 0"));
1887 }
1888 #endif
1889 
1890 /*
1891  * Initialize the pmap associated with process 0.
1892  */
1893 void
1894 moea64_pinit0(mmu_t mmu, pmap_t pm)
1895 {
1896 	moea64_pinit(mmu, pm);
1897 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1898 }
1899 
1900 /*
1901  * Set the physical protection on the specified range of this map as requested.
1902  */
1903 void
1904 moea64_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1905     vm_prot_t prot)
1906 {
1907 	struct	pvo_entry *pvo;
1908 	uintptr_t pt;
1909 
1910 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm, sva,
1911 	    eva, prot);
1912 
1913 
1914 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1915 	    ("moea64_protect: non current pmap"));
1916 
1917 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1918 		moea64_remove(mmu, pm, sva, eva);
1919 		return;
1920 	}
1921 
1922 	vm_page_lock_queues();
1923 	PMAP_LOCK(pm);
1924 	for (; sva < eva; sva += PAGE_SIZE) {
1925 		pvo = moea64_pvo_find_va(pm, sva);
1926 		if (pvo == NULL)
1927 			continue;
1928 
1929 		/*
1930 		 * Grab the PTE pointer before we diddle with the cached PTE
1931 		 * copy.
1932 		 */
1933 		LOCK_TABLE();
1934 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
1935 
1936 		/*
1937 		 * Change the protection of the page.
1938 		 */
1939 		pvo->pvo_pte.lpte.pte_lo &= ~LPTE_PP;
1940 		pvo->pvo_pte.lpte.pte_lo |= LPTE_BR;
1941 		pvo->pvo_pte.lpte.pte_lo &= ~LPTE_NOEXEC;
1942 		if ((prot & VM_PROT_EXECUTE) == 0)
1943 			pvo->pvo_pte.lpte.pte_lo |= LPTE_NOEXEC;
1944 
1945 		/*
1946 		 * If the PVO is in the page table, update that pte as well.
1947 		 */
1948 		if (pt != -1) {
1949 			MOEA64_PTE_CHANGE(mmu, pt, &pvo->pvo_pte.lpte,
1950 			    pvo->pvo_vpn);
1951 			if ((pvo->pvo_pte.lpte.pte_lo &
1952 			    (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1953 				moea64_syncicache(mmu, pm, sva,
1954 				    pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN,
1955 				    PAGE_SIZE);
1956 			}
1957 		}
1958 		UNLOCK_TABLE();
1959 	}
1960 	vm_page_unlock_queues();
1961 	PMAP_UNLOCK(pm);
1962 }
1963 
1964 /*
1965  * Map a list of wired pages into kernel virtual address space.  This is
1966  * intended for temporary mappings which do not need page modification or
1967  * references recorded.  Existing mappings in the region are overwritten.
1968  */
1969 void
1970 moea64_qenter(mmu_t mmu, vm_offset_t va, vm_page_t *m, int count)
1971 {
1972 	while (count-- > 0) {
1973 		moea64_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1974 		va += PAGE_SIZE;
1975 		m++;
1976 	}
1977 }
1978 
1979 /*
1980  * Remove page mappings from kernel virtual address space.  Intended for
1981  * temporary mappings entered by moea64_qenter.
1982  */
1983 void
1984 moea64_qremove(mmu_t mmu, vm_offset_t va, int count)
1985 {
1986 	while (count-- > 0) {
1987 		moea64_kremove(mmu, va);
1988 		va += PAGE_SIZE;
1989 	}
1990 }
1991 
1992 void
1993 moea64_release_vsid(uint64_t vsid)
1994 {
1995 	int idx, mask;
1996 
1997 	mtx_lock(&moea64_slb_mutex);
1998 	idx = vsid & (NVSIDS-1);
1999 	mask = 1 << (idx % VSID_NBPW);
2000 	idx /= VSID_NBPW;
2001 	KASSERT(moea64_vsid_bitmap[idx] & mask,
2002 	    ("Freeing unallocated VSID %#jx", vsid));
2003 	moea64_vsid_bitmap[idx] &= ~mask;
2004 	mtx_unlock(&moea64_slb_mutex);
2005 }
2006 
2007 
2008 void
2009 moea64_release(mmu_t mmu, pmap_t pmap)
2010 {
2011 
2012 	/*
2013 	 * Free segment registers' VSIDs
2014 	 */
2015     #ifdef __powerpc64__
2016 	slb_free_tree(pmap);
2017 	slb_free_user_cache(pmap->pm_slb);
2018     #else
2019 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_release: pm_sr[0] = 0"));
2020 
2021 	moea64_release_vsid(VSID_TO_HASH(pmap->pm_sr[0]));
2022     #endif
2023 
2024 	PMAP_LOCK_DESTROY(pmap);
2025 }
2026 
2027 /*
2028  * Remove the given range of addresses from the specified map.
2029  */
2030 void
2031 moea64_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
2032 {
2033 	struct	pvo_entry *pvo;
2034 
2035 	vm_page_lock_queues();
2036 	PMAP_LOCK(pm);
2037 	for (; sva < eva; sva += PAGE_SIZE) {
2038 		pvo = moea64_pvo_find_va(pm, sva);
2039 		if (pvo != NULL)
2040 			moea64_pvo_remove(mmu, pvo);
2041 	}
2042 	vm_page_unlock_queues();
2043 	PMAP_UNLOCK(pm);
2044 }
2045 
2046 /*
2047  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
2048  * will reflect changes in pte's back to the vm_page.
2049  */
2050 void
2051 moea64_remove_all(mmu_t mmu, vm_page_t m)
2052 {
2053 	struct  pvo_head *pvo_head;
2054 	struct	pvo_entry *pvo, *next_pvo;
2055 	pmap_t	pmap;
2056 
2057 	vm_page_lock_queues();
2058 	pvo_head = vm_page_to_pvoh(m);
2059 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
2060 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
2061 
2062 		pmap = pvo->pvo_pmap;
2063 		PMAP_LOCK(pmap);
2064 		moea64_pvo_remove(mmu, pvo);
2065 		PMAP_UNLOCK(pmap);
2066 	}
2067 	if ((m->flags & PG_WRITEABLE) && moea64_is_modified(mmu, m)) {
2068 		moea64_attr_clear(m, LPTE_CHG);
2069 		vm_page_dirty(m);
2070 	}
2071 	vm_page_flag_clear(m, PG_WRITEABLE);
2072 	vm_page_unlock_queues();
2073 }
2074 
2075 /*
2076  * Allocate a physical page of memory directly from the phys_avail map.
2077  * Can only be called from moea64_bootstrap before avail start and end are
2078  * calculated.
2079  */
2080 vm_offset_t
2081 moea64_bootstrap_alloc(vm_size_t size, u_int align)
2082 {
2083 	vm_offset_t	s, e;
2084 	int		i, j;
2085 
2086 	size = round_page(size);
2087 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2088 		if (align != 0)
2089 			s = (phys_avail[i] + align - 1) & ~(align - 1);
2090 		else
2091 			s = phys_avail[i];
2092 		e = s + size;
2093 
2094 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2095 			continue;
2096 
2097 		if (s + size > platform_real_maxaddr())
2098 			continue;
2099 
2100 		if (s == phys_avail[i]) {
2101 			phys_avail[i] += size;
2102 		} else if (e == phys_avail[i + 1]) {
2103 			phys_avail[i + 1] -= size;
2104 		} else {
2105 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2106 				phys_avail[j] = phys_avail[j - 2];
2107 				phys_avail[j + 1] = phys_avail[j - 1];
2108 			}
2109 
2110 			phys_avail[i + 3] = phys_avail[i + 1];
2111 			phys_avail[i + 1] = s;
2112 			phys_avail[i + 2] = e;
2113 			phys_avail_count++;
2114 		}
2115 
2116 		return (s);
2117 	}
2118 	panic("moea64_bootstrap_alloc: could not allocate memory");
2119 }
2120 
2121 static int
2122 moea64_pvo_enter(mmu_t mmu, pmap_t pm, uma_zone_t zone,
2123     struct pvo_head *pvo_head, vm_offset_t va, vm_offset_t pa,
2124     uint64_t pte_lo, int flags)
2125 {
2126 	struct	 pvo_entry *pvo;
2127 	uint64_t vsid;
2128 	int	 first;
2129 	u_int	 ptegidx;
2130 	int	 i;
2131 	int      bootstrap;
2132 
2133 	/*
2134 	 * One nasty thing that can happen here is that the UMA calls to
2135 	 * allocate new PVOs need to map more memory, which calls pvo_enter(),
2136 	 * which calls UMA...
2137 	 *
2138 	 * We break the loop by detecting recursion and allocating out of
2139 	 * the bootstrap pool.
2140 	 */
2141 
2142 	first = 0;
2143 	bootstrap = (flags & PVO_BOOTSTRAP);
2144 
2145 	if (!moea64_initialized)
2146 		bootstrap = 1;
2147 
2148 	/*
2149 	 * Compute the PTE Group index.
2150 	 */
2151 	va &= ~ADDR_POFF;
2152 	vsid = va_to_vsid(pm, va);
2153 	ptegidx = va_to_pteg(vsid, va, flags & PVO_LARGE);
2154 
2155 	/*
2156 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
2157 	 * there is a mapping.
2158 	 */
2159 	LOCK_TABLE();
2160 
2161 	moea64_pvo_enter_calls++;
2162 
2163 	LIST_FOREACH(pvo, &moea64_pvo_table[ptegidx], pvo_olink) {
2164 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2165 			if ((pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) == pa &&
2166 			    (pvo->pvo_pte.lpte.pte_lo & (LPTE_NOEXEC | LPTE_PP))
2167 			    == (pte_lo & (LPTE_NOEXEC | LPTE_PP))) {
2168 			    	if (!(pvo->pvo_pte.lpte.pte_hi & LPTE_VALID)) {
2169 					/* Re-insert if spilled */
2170 					i = MOEA64_PTE_INSERT(mmu, ptegidx,
2171 					    &pvo->pvo_pte.lpte);
2172 					if (i >= 0)
2173 						PVO_PTEGIDX_SET(pvo, i);
2174 					moea64_pte_overflow--;
2175 				}
2176 				UNLOCK_TABLE();
2177 				return (0);
2178 			}
2179 			moea64_pvo_remove(mmu, pvo);
2180 			break;
2181 		}
2182 	}
2183 
2184 	/*
2185 	 * If we aren't overwriting a mapping, try to allocate.
2186 	 */
2187 	if (bootstrap) {
2188 		if (moea64_bpvo_pool_index >= BPVO_POOL_SIZE) {
2189 			panic("moea64_enter: bpvo pool exhausted, %d, %d, %zd",
2190 			      moea64_bpvo_pool_index, BPVO_POOL_SIZE,
2191 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
2192 		}
2193 		pvo = &moea64_bpvo_pool[moea64_bpvo_pool_index];
2194 		moea64_bpvo_pool_index++;
2195 		bootstrap = 1;
2196 	} else {
2197 		/*
2198 		 * Note: drop the table lock around the UMA allocation in
2199 		 * case the UMA allocator needs to manipulate the page
2200 		 * table. The mapping we are working with is already
2201 		 * protected by the PMAP lock.
2202 		 */
2203 		UNLOCK_TABLE();
2204 		pvo = uma_zalloc(zone, M_NOWAIT);
2205 		LOCK_TABLE();
2206 	}
2207 
2208 	if (pvo == NULL) {
2209 		UNLOCK_TABLE();
2210 		return (ENOMEM);
2211 	}
2212 
2213 	moea64_pvo_entries++;
2214 	pvo->pvo_vaddr = va;
2215 	pvo->pvo_vpn = (uint64_t)((va & ADDR_PIDX) >> ADDR_PIDX_SHFT)
2216 	    | (vsid << 16);
2217 	pvo->pvo_pmap = pm;
2218 	LIST_INSERT_HEAD(&moea64_pvo_table[ptegidx], pvo, pvo_olink);
2219 	pvo->pvo_vaddr &= ~ADDR_POFF;
2220 
2221 	if (flags & PVO_WIRED)
2222 		pvo->pvo_vaddr |= PVO_WIRED;
2223 	if (pvo_head != &moea64_pvo_kunmanaged)
2224 		pvo->pvo_vaddr |= PVO_MANAGED;
2225 	if (bootstrap)
2226 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
2227 	if (flags & PVO_FAKE)
2228 		pvo->pvo_vaddr |= PVO_FAKE;
2229 	if (flags & PVO_LARGE)
2230 		pvo->pvo_vaddr |= PVO_LARGE;
2231 
2232 	moea64_pte_create(&pvo->pvo_pte.lpte, vsid, va,
2233 	    (uint64_t)(pa) | pte_lo, flags);
2234 
2235 	/*
2236 	 * Remember if the list was empty and therefore will be the first
2237 	 * item.
2238 	 */
2239 	if (LIST_FIRST(pvo_head) == NULL)
2240 		first = 1;
2241 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2242 
2243 	if (pvo->pvo_vaddr & PVO_WIRED) {
2244 		pvo->pvo_pte.lpte.pte_hi |= LPTE_WIRED;
2245 		pm->pm_stats.wired_count++;
2246 	}
2247 	pm->pm_stats.resident_count++;
2248 
2249 	/*
2250 	 * We hope this succeeds but it isn't required.
2251 	 */
2252 	i = MOEA64_PTE_INSERT(mmu, ptegidx, &pvo->pvo_pte.lpte);
2253 	if (i >= 0) {
2254 		PVO_PTEGIDX_SET(pvo, i);
2255 	} else {
2256 		panic("moea64_pvo_enter: overflow");
2257 		moea64_pte_overflow++;
2258 	}
2259 
2260 	if (pm == kernel_pmap)
2261 		isync();
2262 
2263 	UNLOCK_TABLE();
2264 
2265 #ifdef __powerpc64__
2266 	/*
2267 	 * Make sure all our bootstrap mappings are in the SLB as soon
2268 	 * as virtual memory is switched on.
2269 	 */
2270 	if (!pmap_bootstrapped)
2271 		moea64_bootstrap_slb_prefault(va, flags & PVO_LARGE);
2272 #endif
2273 
2274 	return (first ? ENOENT : 0);
2275 }
2276 
2277 static void
2278 moea64_pvo_remove(mmu_t mmu, struct pvo_entry *pvo)
2279 {
2280 	uintptr_t pt;
2281 
2282 	/*
2283 	 * If there is an active pte entry, we need to deactivate it (and
2284 	 * save the ref & cfg bits).
2285 	 */
2286 	LOCK_TABLE();
2287 	pt = MOEA64_PVO_TO_PTE(mmu, pvo);
2288 	if (pt != -1) {
2289 		MOEA64_PTE_UNSET(mmu, pt, &pvo->pvo_pte.lpte, pvo->pvo_vpn);
2290 		PVO_PTEGIDX_CLR(pvo);
2291 	} else {
2292 		moea64_pte_overflow--;
2293 	}
2294 
2295 	/*
2296 	 * Update our statistics.
2297 	 */
2298 	pvo->pvo_pmap->pm_stats.resident_count--;
2299 	if (pvo->pvo_vaddr & PVO_WIRED)
2300 		pvo->pvo_pmap->pm_stats.wired_count--;
2301 
2302 	/*
2303 	 * Save the REF/CHG bits into their cache if the page is managed.
2304 	 */
2305 	if ((pvo->pvo_vaddr & (PVO_MANAGED|PVO_FAKE)) == PVO_MANAGED) {
2306 		struct	vm_page *pg;
2307 
2308 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN);
2309 		if (pg != NULL) {
2310 			moea64_attr_save(pg, pvo->pvo_pte.lpte.pte_lo &
2311 			    (LPTE_REF | LPTE_CHG));
2312 		}
2313 	}
2314 
2315 	/*
2316 	 * Remove this PVO from the PV list.
2317 	 */
2318 	LIST_REMOVE(pvo, pvo_vlink);
2319 
2320 	/*
2321 	 * Remove this from the overflow list and return it to the pool
2322 	 * if we aren't going to reuse it.
2323 	 */
2324 	LIST_REMOVE(pvo, pvo_olink);
2325 
2326 	moea64_pvo_entries--;
2327 	moea64_pvo_remove_calls++;
2328 
2329 	UNLOCK_TABLE();
2330 
2331 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2332 		uma_zfree((pvo->pvo_vaddr & PVO_MANAGED) ? moea64_mpvo_zone :
2333 		    moea64_upvo_zone, pvo);
2334 }
2335 
2336 static struct pvo_entry *
2337 moea64_pvo_find_va(pmap_t pm, vm_offset_t va)
2338 {
2339 	struct		pvo_entry *pvo;
2340 	int		ptegidx;
2341 	uint64_t	vsid;
2342 	#ifdef __powerpc64__
2343 	uint64_t	slbv;
2344 
2345 	if (pm == kernel_pmap) {
2346 		slbv = kernel_va_to_slbv(va);
2347 	} else {
2348 		struct slb *slb;
2349 		slb = user_va_to_slb_entry(pm, va);
2350 		/* The page is not mapped if the segment isn't */
2351 		if (slb == NULL)
2352 			return NULL;
2353 		slbv = slb->slbv;
2354 	}
2355 
2356 	vsid = (slbv & SLBV_VSID_MASK) >> SLBV_VSID_SHIFT;
2357 	if (slbv & SLBV_L)
2358 		va &= ~moea64_large_page_mask;
2359 	else
2360 		va &= ~ADDR_POFF;
2361 	ptegidx = va_to_pteg(vsid, va, slbv & SLBV_L);
2362 	#else
2363 	va &= ~ADDR_POFF;
2364 	vsid = va_to_vsid(pm, va);
2365 	ptegidx = va_to_pteg(vsid, va, 0);
2366 	#endif
2367 
2368 	LOCK_TABLE();
2369 	LIST_FOREACH(pvo, &moea64_pvo_table[ptegidx], pvo_olink) {
2370 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va)
2371 			break;
2372 	}
2373 	UNLOCK_TABLE();
2374 
2375 	return (pvo);
2376 }
2377 
2378 static boolean_t
2379 moea64_query_bit(mmu_t mmu, vm_page_t m, u_int64_t ptebit)
2380 {
2381 	struct	pvo_entry *pvo;
2382 	uintptr_t pt;
2383 
2384 	if (moea64_attr_fetch(m) & ptebit)
2385 		return (TRUE);
2386 
2387 	vm_page_lock_queues();
2388 
2389 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2390 
2391 		/*
2392 		 * See if we saved the bit off.  If so, cache it and return
2393 		 * success.
2394 		 */
2395 		if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2396 			moea64_attr_save(m, ptebit);
2397 			vm_page_unlock_queues();
2398 			return (TRUE);
2399 		}
2400 	}
2401 
2402 	/*
2403 	 * No luck, now go through the hard part of looking at the PTEs
2404 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2405 	 * the PTEs.
2406 	 */
2407 	powerpc_sync();
2408 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2409 
2410 		/*
2411 		 * See if this pvo has a valid PTE.  if so, fetch the
2412 		 * REF/CHG bits from the valid PTE.  If the appropriate
2413 		 * ptebit is set, cache it and return success.
2414 		 */
2415 		LOCK_TABLE();
2416 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
2417 		if (pt != -1) {
2418 			MOEA64_PTE_SYNCH(mmu, pt, &pvo->pvo_pte.lpte);
2419 			if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2420 				UNLOCK_TABLE();
2421 
2422 				moea64_attr_save(m, ptebit);
2423 				vm_page_unlock_queues();
2424 				return (TRUE);
2425 			}
2426 		}
2427 		UNLOCK_TABLE();
2428 	}
2429 
2430 	vm_page_unlock_queues();
2431 	return (FALSE);
2432 }
2433 
2434 static u_int
2435 moea64_clear_bit(mmu_t mmu, vm_page_t m, u_int64_t ptebit)
2436 {
2437 	u_int	count;
2438 	struct	pvo_entry *pvo;
2439 	uintptr_t pt;
2440 
2441 	vm_page_lock_queues();
2442 
2443 	/*
2444 	 * Clear the cached value.
2445 	 */
2446 	moea64_attr_clear(m, ptebit);
2447 
2448 	/*
2449 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2450 	 * we can reset the right ones).  note that since the pvo entries and
2451 	 * list heads are accessed via BAT0 and are never placed in the page
2452 	 * table, we don't have to worry about further accesses setting the
2453 	 * REF/CHG bits.
2454 	 */
2455 	powerpc_sync();
2456 
2457 	/*
2458 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2459 	 * valid pte clear the ptebit from the valid pte.
2460 	 */
2461 	count = 0;
2462 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2463 
2464 		LOCK_TABLE();
2465 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
2466 		if (pt != -1) {
2467 			MOEA64_PTE_SYNCH(mmu, pt, &pvo->pvo_pte.lpte);
2468 			if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2469 				count++;
2470 				MOEA64_PTE_CLEAR(mmu, pt, &pvo->pvo_pte.lpte,
2471 				    pvo->pvo_vpn, ptebit);
2472 			}
2473 		}
2474 		pvo->pvo_pte.lpte.pte_lo &= ~ptebit;
2475 		UNLOCK_TABLE();
2476 	}
2477 
2478 	vm_page_unlock_queues();
2479 	return (count);
2480 }
2481 
2482 boolean_t
2483 moea64_dev_direct_mapped(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2484 {
2485 	struct pvo_entry *pvo;
2486 	vm_offset_t ppa;
2487 	int error = 0;
2488 
2489 	PMAP_LOCK(kernel_pmap);
2490 	for (ppa = pa & ~ADDR_POFF; ppa < pa + size; ppa += PAGE_SIZE) {
2491 		pvo = moea64_pvo_find_va(kernel_pmap, ppa);
2492 		if (pvo == NULL ||
2493 		    (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) != ppa) {
2494 			error = EFAULT;
2495 			break;
2496 		}
2497 	}
2498 	PMAP_UNLOCK(kernel_pmap);
2499 
2500 	return (error);
2501 }
2502 
2503 /*
2504  * Map a set of physical memory pages into the kernel virtual
2505  * address space. Return a pointer to where it is mapped. This
2506  * routine is intended to be used for mapping device memory,
2507  * NOT real memory.
2508  */
2509 void *
2510 moea64_mapdev_attr(mmu_t mmu, vm_offset_t pa, vm_size_t size, vm_memattr_t ma)
2511 {
2512 	vm_offset_t va, tmpva, ppa, offset;
2513 
2514 	ppa = trunc_page(pa);
2515 	offset = pa & PAGE_MASK;
2516 	size = roundup(offset + size, PAGE_SIZE);
2517 
2518 	va = kmem_alloc_nofault(kernel_map, size);
2519 
2520 	if (!va)
2521 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
2522 
2523 	for (tmpva = va; size > 0;) {
2524 		moea64_kenter_attr(mmu, tmpva, ppa, ma);
2525 		size -= PAGE_SIZE;
2526 		tmpva += PAGE_SIZE;
2527 		ppa += PAGE_SIZE;
2528 	}
2529 
2530 	return ((void *)(va + offset));
2531 }
2532 
2533 void *
2534 moea64_mapdev(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2535 {
2536 
2537 	return moea64_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT);
2538 }
2539 
2540 void
2541 moea64_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2542 {
2543 	vm_offset_t base, offset;
2544 
2545 	base = trunc_page(va);
2546 	offset = va & PAGE_MASK;
2547 	size = roundup(offset + size, PAGE_SIZE);
2548 
2549 	kmem_free(kernel_map, base, size);
2550 }
2551 
2552 void
2553 moea64_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2554 {
2555 	struct pvo_entry *pvo;
2556 	vm_offset_t lim;
2557 	vm_paddr_t pa;
2558 	vm_size_t len;
2559 
2560 	PMAP_LOCK(pm);
2561 	while (sz > 0) {
2562 		lim = round_page(va);
2563 		len = MIN(lim - va, sz);
2564 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
2565 		if (pvo != NULL && !(pvo->pvo_pte.lpte.pte_lo & LPTE_I)) {
2566 			pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) |
2567 			    (va & ADDR_POFF);
2568 			moea64_syncicache(mmu, pm, va, pa, len);
2569 		}
2570 		va += len;
2571 		sz -= len;
2572 	}
2573 	PMAP_UNLOCK(pm);
2574 }
2575