xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision a98ff317388a00b992f1bf8404dee596f9383f5e)
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  * Since the information managed by this module is also stored by the
100  * logical address mapping module, this module may throw away valid virtual
101  * to physical mappings at almost any time.  However, invalidations of
102  * mappings must be done as requested.
103  *
104  * In order to cope with hardware architectures which make virtual to
105  * physical map invalidates expensive, this module may delay invalidate
106  * reduced protection operations until such time as they are actually
107  * necessary.  This module is given full information as to which processors
108  * are currently using which maps, and to when physical maps must be made
109  * correct.
110  */
111 
112 #include "opt_compat.h"
113 #include "opt_kstack_pages.h"
114 
115 #include <sys/param.h>
116 #include <sys/kernel.h>
117 #include <sys/queue.h>
118 #include <sys/cpuset.h>
119 #include <sys/ktr.h>
120 #include <sys/lock.h>
121 #include <sys/msgbuf.h>
122 #include <sys/malloc.h>
123 #include <sys/mutex.h>
124 #include <sys/proc.h>
125 #include <sys/rwlock.h>
126 #include <sys/sched.h>
127 #include <sys/sysctl.h>
128 #include <sys/systm.h>
129 #include <sys/vmmeter.h>
130 
131 #include <sys/kdb.h>
132 
133 #include <dev/ofw/openfirm.h>
134 
135 #include <vm/vm.h>
136 #include <vm/vm_param.h>
137 #include <vm/vm_kern.h>
138 #include <vm/vm_page.h>
139 #include <vm/vm_map.h>
140 #include <vm/vm_object.h>
141 #include <vm/vm_extern.h>
142 #include <vm/vm_pageout.h>
143 #include <vm/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 /*
173  * Locking semantics:
174  * -- Read lock: if no modifications are being made to either the PVO lists
175  *    or page table or if any modifications being made result in internal
176  *    changes (e.g. wiring, protection) such that the existence of the PVOs
177  *    is unchanged and they remain associated with the same pmap (in which
178  *    case the changes should be protected by the pmap lock)
179  * -- Write lock: required if PTEs/PVOs are being inserted or removed.
180  */
181 
182 #define LOCK_TABLE_RD() rw_rlock(&moea64_table_lock)
183 #define UNLOCK_TABLE_RD() rw_runlock(&moea64_table_lock)
184 #define LOCK_TABLE_WR() rw_wlock(&moea64_table_lock)
185 #define UNLOCK_TABLE_WR() rw_wunlock(&moea64_table_lock)
186 
187 struct ofw_map {
188 	cell_t	om_va;
189 	cell_t	om_len;
190 	cell_t	om_pa_hi;
191 	cell_t	om_pa_lo;
192 	cell_t	om_mode;
193 };
194 
195 /*
196  * Map of physical memory regions.
197  */
198 static struct	mem_region *regions;
199 static struct	mem_region *pregions;
200 static u_int	phys_avail_count;
201 static int	regions_sz, pregions_sz;
202 
203 extern void bs_remap_earlyboot(void);
204 
205 /*
206  * Lock for the pteg and pvo tables.
207  */
208 struct rwlock	moea64_table_lock;
209 struct mtx	moea64_slb_mutex;
210 
211 /*
212  * PTEG data.
213  */
214 u_int		moea64_pteg_count;
215 u_int		moea64_pteg_mask;
216 
217 /*
218  * PVO data.
219  */
220 struct	pvo_head *moea64_pvo_table;		/* pvo entries by pteg index */
221 
222 uma_zone_t	moea64_upvo_zone; /* zone for pvo entries for unmanaged pages */
223 uma_zone_t	moea64_mpvo_zone; /* zone for pvo entries for managed pages */
224 
225 #define	BPVO_POOL_SIZE	327680
226 static struct	pvo_entry *moea64_bpvo_pool;
227 static int	moea64_bpvo_pool_index = 0;
228 
229 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
230 #ifdef __powerpc64__
231 #define	NVSIDS		(NPMAPS * 16)
232 #define VSID_HASHMASK	0xffffffffUL
233 #else
234 #define NVSIDS		NPMAPS
235 #define VSID_HASHMASK	0xfffffUL
236 #endif
237 static u_int	moea64_vsid_bitmap[NVSIDS / VSID_NBPW];
238 
239 static boolean_t moea64_initialized = FALSE;
240 
241 /*
242  * Statistics.
243  */
244 u_int	moea64_pte_valid = 0;
245 u_int	moea64_pte_overflow = 0;
246 u_int	moea64_pvo_entries = 0;
247 u_int	moea64_pvo_enter_calls = 0;
248 u_int	moea64_pvo_remove_calls = 0;
249 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_valid, CTLFLAG_RD,
250     &moea64_pte_valid, 0, "");
251 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_overflow, CTLFLAG_RD,
252     &moea64_pte_overflow, 0, "");
253 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_entries, CTLFLAG_RD,
254     &moea64_pvo_entries, 0, "");
255 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_enter_calls, CTLFLAG_RD,
256     &moea64_pvo_enter_calls, 0, "");
257 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_remove_calls, CTLFLAG_RD,
258     &moea64_pvo_remove_calls, 0, "");
259 
260 vm_offset_t	moea64_scratchpage_va[2];
261 struct pvo_entry *moea64_scratchpage_pvo[2];
262 uintptr_t	moea64_scratchpage_pte[2];
263 struct	mtx	moea64_scratchpage_mtx;
264 
265 uint64_t 	moea64_large_page_mask = 0;
266 int		moea64_large_page_size = 0;
267 int		moea64_large_page_shift = 0;
268 
269 /*
270  * PVO calls.
271  */
272 static int	moea64_pvo_enter(mmu_t, pmap_t, uma_zone_t, struct pvo_head *,
273 		    vm_offset_t, vm_offset_t, uint64_t, int);
274 static void	moea64_pvo_remove(mmu_t, struct pvo_entry *);
275 static struct	pvo_entry *moea64_pvo_find_va(pmap_t, vm_offset_t);
276 
277 /*
278  * Utility routines.
279  */
280 static boolean_t	moea64_query_bit(mmu_t, vm_page_t, u_int64_t);
281 static u_int		moea64_clear_bit(mmu_t, vm_page_t, u_int64_t);
282 static void		moea64_kremove(mmu_t, vm_offset_t);
283 static void		moea64_syncicache(mmu_t, pmap_t pmap, vm_offset_t va,
284 			    vm_offset_t pa, vm_size_t sz);
285 
286 /*
287  * Kernel MMU interface
288  */
289 void moea64_change_wiring(mmu_t, pmap_t, vm_offset_t, boolean_t);
290 void moea64_clear_modify(mmu_t, vm_page_t);
291 void moea64_clear_reference(mmu_t, vm_page_t);
292 void moea64_copy_page(mmu_t, vm_page_t, vm_page_t);
293 void moea64_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
294     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
295 void moea64_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, boolean_t);
296 void moea64_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
297     vm_prot_t);
298 void moea64_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
299 vm_paddr_t moea64_extract(mmu_t, pmap_t, vm_offset_t);
300 vm_page_t moea64_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
301 void moea64_init(mmu_t);
302 boolean_t moea64_is_modified(mmu_t, vm_page_t);
303 boolean_t moea64_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
304 boolean_t moea64_is_referenced(mmu_t, vm_page_t);
305 int moea64_ts_referenced(mmu_t, vm_page_t);
306 vm_offset_t moea64_map(mmu_t, vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
307 boolean_t moea64_page_exists_quick(mmu_t, pmap_t, vm_page_t);
308 int moea64_page_wired_mappings(mmu_t, vm_page_t);
309 void moea64_pinit(mmu_t, pmap_t);
310 void moea64_pinit0(mmu_t, pmap_t);
311 void moea64_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
312 void moea64_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
313 void moea64_qremove(mmu_t, vm_offset_t, int);
314 void moea64_release(mmu_t, pmap_t);
315 void moea64_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
316 void moea64_remove_pages(mmu_t, pmap_t);
317 void moea64_remove_all(mmu_t, vm_page_t);
318 void moea64_remove_write(mmu_t, vm_page_t);
319 void moea64_zero_page(mmu_t, vm_page_t);
320 void moea64_zero_page_area(mmu_t, vm_page_t, int, int);
321 void moea64_zero_page_idle(mmu_t, vm_page_t);
322 void moea64_activate(mmu_t, struct thread *);
323 void moea64_deactivate(mmu_t, struct thread *);
324 void *moea64_mapdev(mmu_t, vm_paddr_t, vm_size_t);
325 void *moea64_mapdev_attr(mmu_t, vm_offset_t, vm_size_t, vm_memattr_t);
326 void moea64_unmapdev(mmu_t, vm_offset_t, vm_size_t);
327 vm_paddr_t moea64_kextract(mmu_t, vm_offset_t);
328 void moea64_page_set_memattr(mmu_t, vm_page_t m, vm_memattr_t ma);
329 void moea64_kenter_attr(mmu_t, vm_offset_t, vm_offset_t, vm_memattr_t ma);
330 void moea64_kenter(mmu_t, vm_offset_t, vm_paddr_t);
331 boolean_t moea64_dev_direct_mapped(mmu_t, vm_paddr_t, vm_size_t);
332 static void moea64_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
333 
334 static mmu_method_t moea64_methods[] = {
335 	MMUMETHOD(mmu_change_wiring,	moea64_change_wiring),
336 	MMUMETHOD(mmu_clear_modify,	moea64_clear_modify),
337 	MMUMETHOD(mmu_clear_reference,	moea64_clear_reference),
338 	MMUMETHOD(mmu_copy_page,	moea64_copy_page),
339 	MMUMETHOD(mmu_copy_pages,	moea64_copy_pages),
340 	MMUMETHOD(mmu_enter,		moea64_enter),
341 	MMUMETHOD(mmu_enter_object,	moea64_enter_object),
342 	MMUMETHOD(mmu_enter_quick,	moea64_enter_quick),
343 	MMUMETHOD(mmu_extract,		moea64_extract),
344 	MMUMETHOD(mmu_extract_and_hold,	moea64_extract_and_hold),
345 	MMUMETHOD(mmu_init,		moea64_init),
346 	MMUMETHOD(mmu_is_modified,	moea64_is_modified),
347 	MMUMETHOD(mmu_is_prefaultable,	moea64_is_prefaultable),
348 	MMUMETHOD(mmu_is_referenced,	moea64_is_referenced),
349 	MMUMETHOD(mmu_ts_referenced,	moea64_ts_referenced),
350 	MMUMETHOD(mmu_map,     		moea64_map),
351 	MMUMETHOD(mmu_page_exists_quick,moea64_page_exists_quick),
352 	MMUMETHOD(mmu_page_wired_mappings,moea64_page_wired_mappings),
353 	MMUMETHOD(mmu_pinit,		moea64_pinit),
354 	MMUMETHOD(mmu_pinit0,		moea64_pinit0),
355 	MMUMETHOD(mmu_protect,		moea64_protect),
356 	MMUMETHOD(mmu_qenter,		moea64_qenter),
357 	MMUMETHOD(mmu_qremove,		moea64_qremove),
358 	MMUMETHOD(mmu_release,		moea64_release),
359 	MMUMETHOD(mmu_remove,		moea64_remove),
360 	MMUMETHOD(mmu_remove_pages,	moea64_remove_pages),
361 	MMUMETHOD(mmu_remove_all,      	moea64_remove_all),
362 	MMUMETHOD(mmu_remove_write,	moea64_remove_write),
363 	MMUMETHOD(mmu_sync_icache,	moea64_sync_icache),
364 	MMUMETHOD(mmu_zero_page,       	moea64_zero_page),
365 	MMUMETHOD(mmu_zero_page_area,	moea64_zero_page_area),
366 	MMUMETHOD(mmu_zero_page_idle,	moea64_zero_page_idle),
367 	MMUMETHOD(mmu_activate,		moea64_activate),
368 	MMUMETHOD(mmu_deactivate,      	moea64_deactivate),
369 	MMUMETHOD(mmu_page_set_memattr,	moea64_page_set_memattr),
370 
371 	/* Internal interfaces */
372 	MMUMETHOD(mmu_mapdev,		moea64_mapdev),
373 	MMUMETHOD(mmu_mapdev_attr,	moea64_mapdev_attr),
374 	MMUMETHOD(mmu_unmapdev,		moea64_unmapdev),
375 	MMUMETHOD(mmu_kextract,		moea64_kextract),
376 	MMUMETHOD(mmu_kenter,		moea64_kenter),
377 	MMUMETHOD(mmu_kenter_attr,	moea64_kenter_attr),
378 	MMUMETHOD(mmu_dev_direct_mapped,moea64_dev_direct_mapped),
379 
380 	{ 0, 0 }
381 };
382 
383 MMU_DEF(oea64_mmu, "mmu_oea64_base", moea64_methods, 0);
384 
385 static __inline u_int
386 va_to_pteg(uint64_t vsid, vm_offset_t addr, int large)
387 {
388 	uint64_t hash;
389 	int shift;
390 
391 	shift = large ? moea64_large_page_shift : ADDR_PIDX_SHFT;
392 	hash = (vsid & VSID_HASH_MASK) ^ (((uint64_t)addr & ADDR_PIDX) >>
393 	    shift);
394 	return (hash & moea64_pteg_mask);
395 }
396 
397 static __inline struct pvo_head *
398 vm_page_to_pvoh(vm_page_t m)
399 {
400 
401 	return (&m->md.mdpg_pvoh);
402 }
403 
404 static __inline void
405 moea64_pte_create(struct lpte *pt, uint64_t vsid, vm_offset_t va,
406     uint64_t pte_lo, int flags)
407 {
408 
409 	/*
410 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
411 	 * set when the real pte is set in memory.
412 	 *
413 	 * Note: Don't set the valid bit for correct operation of tlb update.
414 	 */
415 	pt->pte_hi = (vsid << LPTE_VSID_SHIFT) |
416 	    (((uint64_t)(va & ADDR_PIDX) >> ADDR_API_SHFT64) & LPTE_API);
417 
418 	if (flags & PVO_LARGE)
419 		pt->pte_hi |= LPTE_BIG;
420 
421 	pt->pte_lo = pte_lo;
422 }
423 
424 static __inline uint64_t
425 moea64_calc_wimg(vm_offset_t pa, vm_memattr_t ma)
426 {
427 	uint64_t pte_lo;
428 	int i;
429 
430 	if (ma != VM_MEMATTR_DEFAULT) {
431 		switch (ma) {
432 		case VM_MEMATTR_UNCACHEABLE:
433 			return (LPTE_I | LPTE_G);
434 		case VM_MEMATTR_WRITE_COMBINING:
435 		case VM_MEMATTR_WRITE_BACK:
436 		case VM_MEMATTR_PREFETCHABLE:
437 			return (LPTE_I);
438 		case VM_MEMATTR_WRITE_THROUGH:
439 			return (LPTE_W | LPTE_M);
440 		}
441 	}
442 
443 	/*
444 	 * Assume the page is cache inhibited and access is guarded unless
445 	 * it's in our available memory array.
446 	 */
447 	pte_lo = LPTE_I | LPTE_G;
448 	for (i = 0; i < pregions_sz; i++) {
449 		if ((pa >= pregions[i].mr_start) &&
450 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
451 			pte_lo &= ~(LPTE_I | LPTE_G);
452 			pte_lo |= LPTE_M;
453 			break;
454 		}
455 	}
456 
457 	return pte_lo;
458 }
459 
460 /*
461  * Quick sort callout for comparing memory regions.
462  */
463 static int	om_cmp(const void *a, const void *b);
464 
465 static int
466 om_cmp(const void *a, const void *b)
467 {
468 	const struct	ofw_map *mapa;
469 	const struct	ofw_map *mapb;
470 
471 	mapa = a;
472 	mapb = b;
473 	if (mapa->om_pa_hi < mapb->om_pa_hi)
474 		return (-1);
475 	else if (mapa->om_pa_hi > mapb->om_pa_hi)
476 		return (1);
477 	else if (mapa->om_pa_lo < mapb->om_pa_lo)
478 		return (-1);
479 	else if (mapa->om_pa_lo > mapb->om_pa_lo)
480 		return (1);
481 	else
482 		return (0);
483 }
484 
485 static void
486 moea64_add_ofw_mappings(mmu_t mmup, phandle_t mmu, size_t sz)
487 {
488 	struct ofw_map	translations[sz/sizeof(struct ofw_map)];
489 	register_t	msr;
490 	vm_offset_t	off;
491 	vm_paddr_t	pa_base;
492 	int		i;
493 
494 	bzero(translations, sz);
495 	if (OF_getprop(mmu, "translations", translations, sz) == -1)
496 		panic("moea64_bootstrap: can't get ofw translations");
497 
498 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
499 	sz /= sizeof(*translations);
500 	qsort(translations, sz, sizeof (*translations), om_cmp);
501 
502 	for (i = 0; i < sz; i++) {
503 		CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
504 		    (uint32_t)(translations[i].om_pa_lo), translations[i].om_va,
505 		    translations[i].om_len);
506 
507 		if (translations[i].om_pa_lo % PAGE_SIZE)
508 			panic("OFW translation not page-aligned!");
509 
510 		pa_base = translations[i].om_pa_lo;
511 
512 	      #ifdef __powerpc64__
513 		pa_base += (vm_offset_t)translations[i].om_pa_hi << 32;
514 	      #else
515 		if (translations[i].om_pa_hi)
516 			panic("OFW translations above 32-bit boundary!");
517 	      #endif
518 
519 		/* Now enter the pages for this mapping */
520 
521 		DISABLE_TRANS(msr);
522 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
523 			if (moea64_pvo_find_va(kernel_pmap,
524 			    translations[i].om_va + off) != NULL)
525 				continue;
526 
527 			moea64_kenter(mmup, translations[i].om_va + off,
528 			    pa_base + off);
529 		}
530 		ENABLE_TRANS(msr);
531 	}
532 }
533 
534 #ifdef __powerpc64__
535 static void
536 moea64_probe_large_page(void)
537 {
538 	uint16_t pvr = mfpvr() >> 16;
539 
540 	switch (pvr) {
541 	case IBM970:
542 	case IBM970FX:
543 	case IBM970MP:
544 		powerpc_sync(); isync();
545 		mtspr(SPR_HID4, mfspr(SPR_HID4) & ~HID4_970_DISABLE_LG_PG);
546 		powerpc_sync(); isync();
547 
548 		/* FALLTHROUGH */
549 	case IBMCELLBE:
550 		moea64_large_page_size = 0x1000000; /* 16 MB */
551 		moea64_large_page_shift = 24;
552 		break;
553 	default:
554 		moea64_large_page_size = 0;
555 	}
556 
557 	moea64_large_page_mask = moea64_large_page_size - 1;
558 }
559 
560 static void
561 moea64_bootstrap_slb_prefault(vm_offset_t va, int large)
562 {
563 	struct slb *cache;
564 	struct slb entry;
565 	uint64_t esid, slbe;
566 	uint64_t i;
567 
568 	cache = PCPU_GET(slb);
569 	esid = va >> ADDR_SR_SHFT;
570 	slbe = (esid << SLBE_ESID_SHIFT) | SLBE_VALID;
571 
572 	for (i = 0; i < 64; i++) {
573 		if (cache[i].slbe == (slbe | i))
574 			return;
575 	}
576 
577 	entry.slbe = slbe;
578 	entry.slbv = KERNEL_VSID(esid) << SLBV_VSID_SHIFT;
579 	if (large)
580 		entry.slbv |= SLBV_L;
581 
582 	slb_insert_kernel(entry.slbe, entry.slbv);
583 }
584 #endif
585 
586 static void
587 moea64_setup_direct_map(mmu_t mmup, vm_offset_t kernelstart,
588     vm_offset_t kernelend)
589 {
590 	register_t msr;
591 	vm_paddr_t pa;
592 	vm_offset_t size, off;
593 	uint64_t pte_lo;
594 	int i;
595 
596 	if (moea64_large_page_size == 0)
597 		hw_direct_map = 0;
598 
599 	DISABLE_TRANS(msr);
600 	if (hw_direct_map) {
601 		LOCK_TABLE_WR();
602 		PMAP_LOCK(kernel_pmap);
603 		for (i = 0; i < pregions_sz; i++) {
604 		  for (pa = pregions[i].mr_start; pa < pregions[i].mr_start +
605 		     pregions[i].mr_size; pa += moea64_large_page_size) {
606 			pte_lo = LPTE_M;
607 
608 			/*
609 			 * Set memory access as guarded if prefetch within
610 			 * the page could exit the available physmem area.
611 			 */
612 			if (pa & moea64_large_page_mask) {
613 				pa &= moea64_large_page_mask;
614 				pte_lo |= LPTE_G;
615 			}
616 			if (pa + moea64_large_page_size >
617 			    pregions[i].mr_start + pregions[i].mr_size)
618 				pte_lo |= LPTE_G;
619 
620 			moea64_pvo_enter(mmup, kernel_pmap, moea64_upvo_zone,
621 				    NULL, pa, pa, pte_lo,
622 				    PVO_WIRED | PVO_LARGE);
623 		  }
624 		}
625 		PMAP_UNLOCK(kernel_pmap);
626 		UNLOCK_TABLE_WR();
627 	} else {
628 		size = sizeof(struct pvo_head) * moea64_pteg_count;
629 		off = (vm_offset_t)(moea64_pvo_table);
630 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
631 			moea64_kenter(mmup, pa, pa);
632 		size = BPVO_POOL_SIZE*sizeof(struct pvo_entry);
633 		off = (vm_offset_t)(moea64_bpvo_pool);
634 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
635 		moea64_kenter(mmup, pa, pa);
636 
637 		/*
638 		 * Map certain important things, like ourselves.
639 		 *
640 		 * NOTE: We do not map the exception vector space. That code is
641 		 * used only in real mode, and leaving it unmapped allows us to
642 		 * catch NULL pointer deferences, instead of making NULL a valid
643 		 * address.
644 		 */
645 
646 		for (pa = kernelstart & ~PAGE_MASK; pa < kernelend;
647 		    pa += PAGE_SIZE)
648 			moea64_kenter(mmup, pa, pa);
649 	}
650 	ENABLE_TRANS(msr);
651 
652 	/*
653 	 * Allow user to override unmapped_buf_allowed for testing.
654 	 * XXXKIB Only direct map implementation was tested.
655 	 */
656 	if (!TUNABLE_INT_FETCH("vfs.unmapped_buf_allowed",
657 	    &unmapped_buf_allowed))
658 		unmapped_buf_allowed = hw_direct_map;
659 }
660 
661 void
662 moea64_early_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
663 {
664 	int		i, j;
665 	vm_size_t	physsz, hwphyssz;
666 
667 #ifndef __powerpc64__
668 	/* We don't have a direct map since there is no BAT */
669 	hw_direct_map = 0;
670 
671 	/* Make sure battable is zero, since we have no BAT */
672 	for (i = 0; i < 16; i++) {
673 		battable[i].batu = 0;
674 		battable[i].batl = 0;
675 	}
676 #else
677 	moea64_probe_large_page();
678 
679 	/* Use a direct map if we have large page support */
680 	if (moea64_large_page_size > 0)
681 		hw_direct_map = 1;
682 	else
683 		hw_direct_map = 0;
684 #endif
685 
686 	/* Get physical memory regions from firmware */
687 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
688 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
689 
690 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
691 		panic("moea64_bootstrap: phys_avail too small");
692 
693 	phys_avail_count = 0;
694 	physsz = 0;
695 	hwphyssz = 0;
696 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
697 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
698 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
699 		    regions[i].mr_start + regions[i].mr_size,
700 		    regions[i].mr_size);
701 		if (hwphyssz != 0 &&
702 		    (physsz + regions[i].mr_size) >= hwphyssz) {
703 			if (physsz < hwphyssz) {
704 				phys_avail[j] = regions[i].mr_start;
705 				phys_avail[j + 1] = regions[i].mr_start +
706 				    hwphyssz - physsz;
707 				physsz = hwphyssz;
708 				phys_avail_count++;
709 			}
710 			break;
711 		}
712 		phys_avail[j] = regions[i].mr_start;
713 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
714 		phys_avail_count++;
715 		physsz += regions[i].mr_size;
716 	}
717 
718 	/* Check for overlap with the kernel and exception vectors */
719 	for (j = 0; j < 2*phys_avail_count; j+=2) {
720 		if (phys_avail[j] < EXC_LAST)
721 			phys_avail[j] += EXC_LAST;
722 
723 		if (kernelstart >= phys_avail[j] &&
724 		    kernelstart < phys_avail[j+1]) {
725 			if (kernelend < phys_avail[j+1]) {
726 				phys_avail[2*phys_avail_count] =
727 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
728 				phys_avail[2*phys_avail_count + 1] =
729 				    phys_avail[j+1];
730 				phys_avail_count++;
731 			}
732 
733 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
734 		}
735 
736 		if (kernelend >= phys_avail[j] &&
737 		    kernelend < phys_avail[j+1]) {
738 			if (kernelstart > phys_avail[j]) {
739 				phys_avail[2*phys_avail_count] = phys_avail[j];
740 				phys_avail[2*phys_avail_count + 1] =
741 				    kernelstart & ~PAGE_MASK;
742 				phys_avail_count++;
743 			}
744 
745 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
746 		}
747 	}
748 
749 	physmem = btoc(physsz);
750 
751 #ifdef PTEGCOUNT
752 	moea64_pteg_count = PTEGCOUNT;
753 #else
754 	moea64_pteg_count = 0x1000;
755 
756 	while (moea64_pteg_count < physmem)
757 		moea64_pteg_count <<= 1;
758 
759 	moea64_pteg_count >>= 1;
760 #endif /* PTEGCOUNT */
761 }
762 
763 void
764 moea64_mid_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
765 {
766 	vm_size_t	size;
767 	register_t	msr;
768 	int		i;
769 
770 	/*
771 	 * Set PTEG mask
772 	 */
773 	moea64_pteg_mask = moea64_pteg_count - 1;
774 
775 	/*
776 	 * Allocate pv/overflow lists.
777 	 */
778 	size = sizeof(struct pvo_head) * moea64_pteg_count;
779 
780 	moea64_pvo_table = (struct pvo_head *)moea64_bootstrap_alloc(size,
781 	    PAGE_SIZE);
782 	CTR1(KTR_PMAP, "moea64_bootstrap: PVO table at %p", moea64_pvo_table);
783 
784 	DISABLE_TRANS(msr);
785 	for (i = 0; i < moea64_pteg_count; i++)
786 		LIST_INIT(&moea64_pvo_table[i]);
787 	ENABLE_TRANS(msr);
788 
789 	/*
790 	 * Initialize the lock that synchronizes access to the pteg and pvo
791 	 * tables.
792 	 */
793 	rw_init_flags(&moea64_table_lock, "pmap tables", RW_RECURSE);
794 	mtx_init(&moea64_slb_mutex, "SLB table", NULL, MTX_DEF);
795 
796 	/*
797 	 * Initialise the unmanaged pvo pool.
798 	 */
799 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
800 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
801 	moea64_bpvo_pool_index = 0;
802 
803 	/*
804 	 * Make sure kernel vsid is allocated as well as VSID 0.
805 	 */
806 	#ifndef __powerpc64__
807 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NVSIDS - 1)) / VSID_NBPW]
808 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
809 	moea64_vsid_bitmap[0] |= 1;
810 	#endif
811 
812 	/*
813 	 * Initialize the kernel pmap (which is statically allocated).
814 	 */
815 	#ifdef __powerpc64__
816 	for (i = 0; i < 64; i++) {
817 		pcpup->pc_slb[i].slbv = 0;
818 		pcpup->pc_slb[i].slbe = 0;
819 	}
820 	#else
821 	for (i = 0; i < 16; i++)
822 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
823 	#endif
824 
825 	kernel_pmap->pmap_phys = kernel_pmap;
826 	CPU_FILL(&kernel_pmap->pm_active);
827 	RB_INIT(&kernel_pmap->pmap_pvo);
828 
829 	PMAP_LOCK_INIT(kernel_pmap);
830 
831 	/*
832 	 * Now map in all the other buffers we allocated earlier
833 	 */
834 
835 	moea64_setup_direct_map(mmup, kernelstart, kernelend);
836 }
837 
838 void
839 moea64_late_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
840 {
841 	ihandle_t	mmui;
842 	phandle_t	chosen;
843 	phandle_t	mmu;
844 	size_t		sz;
845 	int		i;
846 	vm_offset_t	pa, va;
847 	void		*dpcpu;
848 
849 	/*
850 	 * Set up the Open Firmware pmap and add its mappings if not in real
851 	 * mode.
852 	 */
853 
854 	chosen = OF_finddevice("/chosen");
855 	if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1) {
856 	    mmu = OF_instance_to_package(mmui);
857 	    if (mmu == -1 || (sz = OF_getproplen(mmu, "translations")) == -1)
858 		sz = 0;
859 	    if (sz > 6144 /* tmpstksz - 2 KB headroom */)
860 		panic("moea64_bootstrap: too many ofw translations");
861 
862 	    if (sz > 0)
863 		moea64_add_ofw_mappings(mmup, mmu, sz);
864 	}
865 
866 	/*
867 	 * Calculate the last available physical address.
868 	 */
869 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
870 		;
871 	Maxmem = powerpc_btop(phys_avail[i + 1]);
872 
873 	/*
874 	 * Initialize MMU and remap early physical mappings
875 	 */
876 	MMU_CPU_BOOTSTRAP(mmup,0);
877 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
878 	pmap_bootstrapped++;
879 	bs_remap_earlyboot();
880 
881 	/*
882 	 * Set the start and end of kva.
883 	 */
884 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
885 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
886 
887 	/*
888 	 * Map the entire KVA range into the SLB. We must not fault there.
889 	 */
890 	#ifdef __powerpc64__
891 	for (va = virtual_avail; va < virtual_end; va += SEGMENT_LENGTH)
892 		moea64_bootstrap_slb_prefault(va, 0);
893 	#endif
894 
895 	/*
896 	 * Figure out how far we can extend virtual_end into segment 16
897 	 * without running into existing mappings. Segment 16 is guaranteed
898 	 * to contain neither RAM nor devices (at least on Apple hardware),
899 	 * but will generally contain some OFW mappings we should not
900 	 * step on.
901 	 */
902 
903 	#ifndef __powerpc64__	/* KVA is in high memory on PPC64 */
904 	PMAP_LOCK(kernel_pmap);
905 	while (virtual_end < VM_MAX_KERNEL_ADDRESS &&
906 	    moea64_pvo_find_va(kernel_pmap, virtual_end+1) == NULL)
907 		virtual_end += PAGE_SIZE;
908 	PMAP_UNLOCK(kernel_pmap);
909 	#endif
910 
911 	/*
912 	 * Allocate a kernel stack with a guard page for thread0 and map it
913 	 * into the kernel page map.
914 	 */
915 	pa = moea64_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, PAGE_SIZE);
916 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
917 	virtual_avail = va + KSTACK_PAGES * PAGE_SIZE;
918 	CTR2(KTR_PMAP, "moea64_bootstrap: kstack0 at %#x (%#x)", pa, va);
919 	thread0.td_kstack = va;
920 	thread0.td_kstack_pages = KSTACK_PAGES;
921 	for (i = 0; i < KSTACK_PAGES; i++) {
922 		moea64_kenter(mmup, va, pa);
923 		pa += PAGE_SIZE;
924 		va += PAGE_SIZE;
925 	}
926 
927 	/*
928 	 * Allocate virtual address space for the message buffer.
929 	 */
930 	pa = msgbuf_phys = moea64_bootstrap_alloc(msgbufsize, PAGE_SIZE);
931 	msgbufp = (struct msgbuf *)virtual_avail;
932 	va = virtual_avail;
933 	virtual_avail += round_page(msgbufsize);
934 	while (va < virtual_avail) {
935 		moea64_kenter(mmup, va, pa);
936 		pa += PAGE_SIZE;
937 		va += PAGE_SIZE;
938 	}
939 
940 	/*
941 	 * Allocate virtual address space for the dynamic percpu area.
942 	 */
943 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
944 	dpcpu = (void *)virtual_avail;
945 	va = virtual_avail;
946 	virtual_avail += DPCPU_SIZE;
947 	while (va < virtual_avail) {
948 		moea64_kenter(mmup, va, pa);
949 		pa += PAGE_SIZE;
950 		va += PAGE_SIZE;
951 	}
952 	dpcpu_init(dpcpu, 0);
953 
954 	/*
955 	 * Allocate some things for page zeroing. We put this directly
956 	 * in the page table, marked with LPTE_LOCKED, to avoid any
957 	 * of the PVO book-keeping or other parts of the VM system
958 	 * from even knowing that this hack exists.
959 	 */
960 
961 	if (!hw_direct_map) {
962 		mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL,
963 		    MTX_DEF);
964 		for (i = 0; i < 2; i++) {
965 			moea64_scratchpage_va[i] = (virtual_end+1) - PAGE_SIZE;
966 			virtual_end -= PAGE_SIZE;
967 
968 			moea64_kenter(mmup, moea64_scratchpage_va[i], 0);
969 
970 			moea64_scratchpage_pvo[i] = moea64_pvo_find_va(
971 			    kernel_pmap, (vm_offset_t)moea64_scratchpage_va[i]);
972 			LOCK_TABLE_RD();
973 			moea64_scratchpage_pte[i] = MOEA64_PVO_TO_PTE(
974 			    mmup, moea64_scratchpage_pvo[i]);
975 			moea64_scratchpage_pvo[i]->pvo_pte.lpte.pte_hi
976 			    |= LPTE_LOCKED;
977 			MOEA64_PTE_CHANGE(mmup, moea64_scratchpage_pte[i],
978 			    &moea64_scratchpage_pvo[i]->pvo_pte.lpte,
979 			    moea64_scratchpage_pvo[i]->pvo_vpn);
980 			UNLOCK_TABLE_RD();
981 		}
982 	}
983 }
984 
985 /*
986  * Activate a user pmap.  The pmap must be activated before its address
987  * space can be accessed in any way.
988  */
989 void
990 moea64_activate(mmu_t mmu, struct thread *td)
991 {
992 	pmap_t	pm;
993 
994 	pm = &td->td_proc->p_vmspace->vm_pmap;
995 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
996 
997 	#ifdef __powerpc64__
998 	PCPU_SET(userslb, pm->pm_slb);
999 	#else
1000 	PCPU_SET(curpmap, pm->pmap_phys);
1001 	#endif
1002 }
1003 
1004 void
1005 moea64_deactivate(mmu_t mmu, struct thread *td)
1006 {
1007 	pmap_t	pm;
1008 
1009 	pm = &td->td_proc->p_vmspace->vm_pmap;
1010 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1011 	#ifdef __powerpc64__
1012 	PCPU_SET(userslb, NULL);
1013 	#else
1014 	PCPU_SET(curpmap, NULL);
1015 	#endif
1016 }
1017 
1018 void
1019 moea64_change_wiring(mmu_t mmu, pmap_t pm, vm_offset_t va, boolean_t wired)
1020 {
1021 	struct	pvo_entry *pvo;
1022 	uintptr_t pt;
1023 	uint64_t vsid;
1024 	int	i, ptegidx;
1025 
1026 	LOCK_TABLE_WR();
1027 	PMAP_LOCK(pm);
1028 	pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
1029 
1030 	if (pvo != NULL) {
1031 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
1032 
1033 		if (wired) {
1034 			if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1035 				pm->pm_stats.wired_count++;
1036 			pvo->pvo_vaddr |= PVO_WIRED;
1037 			pvo->pvo_pte.lpte.pte_hi |= LPTE_WIRED;
1038 		} else {
1039 			if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1040 				pm->pm_stats.wired_count--;
1041 			pvo->pvo_vaddr &= ~PVO_WIRED;
1042 			pvo->pvo_pte.lpte.pte_hi &= ~LPTE_WIRED;
1043 		}
1044 
1045 		if (pt != -1) {
1046 			/* Update wiring flag in page table. */
1047 			MOEA64_PTE_CHANGE(mmu, pt, &pvo->pvo_pte.lpte,
1048 			    pvo->pvo_vpn);
1049 		} else if (wired) {
1050 			/*
1051 			 * If we are wiring the page, and it wasn't in the
1052 			 * page table before, add it.
1053 			 */
1054 			vsid = PVO_VSID(pvo);
1055 			ptegidx = va_to_pteg(vsid, PVO_VADDR(pvo),
1056 			    pvo->pvo_vaddr & PVO_LARGE);
1057 
1058 			i = MOEA64_PTE_INSERT(mmu, ptegidx, &pvo->pvo_pte.lpte);
1059 
1060 			if (i >= 0) {
1061 				PVO_PTEGIDX_CLR(pvo);
1062 				PVO_PTEGIDX_SET(pvo, i);
1063 			}
1064 		}
1065 
1066 	}
1067 	UNLOCK_TABLE_WR();
1068 	PMAP_UNLOCK(pm);
1069 }
1070 
1071 /*
1072  * This goes through and sets the physical address of our
1073  * special scratch PTE to the PA we want to zero or copy. Because
1074  * of locking issues (this can get called in pvo_enter() by
1075  * the UMA allocator), we can't use most other utility functions here
1076  */
1077 
1078 static __inline
1079 void moea64_set_scratchpage_pa(mmu_t mmup, int which, vm_offset_t pa) {
1080 
1081 	KASSERT(!hw_direct_map, ("Using OEA64 scratchpage with a direct map!"));
1082 	mtx_assert(&moea64_scratchpage_mtx, MA_OWNED);
1083 
1084 	moea64_scratchpage_pvo[which]->pvo_pte.lpte.pte_lo &=
1085 	    ~(LPTE_WIMG | LPTE_RPGN);
1086 	moea64_scratchpage_pvo[which]->pvo_pte.lpte.pte_lo |=
1087 	    moea64_calc_wimg(pa, VM_MEMATTR_DEFAULT) | (uint64_t)pa;
1088 	MOEA64_PTE_CHANGE(mmup, moea64_scratchpage_pte[which],
1089 	    &moea64_scratchpage_pvo[which]->pvo_pte.lpte,
1090 	    moea64_scratchpage_pvo[which]->pvo_vpn);
1091 	isync();
1092 }
1093 
1094 void
1095 moea64_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1096 {
1097 	vm_offset_t	dst;
1098 	vm_offset_t	src;
1099 
1100 	dst = VM_PAGE_TO_PHYS(mdst);
1101 	src = VM_PAGE_TO_PHYS(msrc);
1102 
1103 	if (hw_direct_map) {
1104 		bcopy((void *)src, (void *)dst, PAGE_SIZE);
1105 	} else {
1106 		mtx_lock(&moea64_scratchpage_mtx);
1107 
1108 		moea64_set_scratchpage_pa(mmu, 0, src);
1109 		moea64_set_scratchpage_pa(mmu, 1, dst);
1110 
1111 		bcopy((void *)moea64_scratchpage_va[0],
1112 		    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1113 
1114 		mtx_unlock(&moea64_scratchpage_mtx);
1115 	}
1116 }
1117 
1118 static inline void
1119 moea64_copy_pages_dmap(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1120     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1121 {
1122 	void *a_cp, *b_cp;
1123 	vm_offset_t a_pg_offset, b_pg_offset;
1124 	int cnt;
1125 
1126 	while (xfersize > 0) {
1127 		a_pg_offset = a_offset & PAGE_MASK;
1128 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1129 		a_cp = (char *)VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]) +
1130 		    a_pg_offset;
1131 		b_pg_offset = b_offset & PAGE_MASK;
1132 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1133 		b_cp = (char *)VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]) +
1134 		    b_pg_offset;
1135 		bcopy(a_cp, b_cp, cnt);
1136 		a_offset += cnt;
1137 		b_offset += cnt;
1138 		xfersize -= cnt;
1139 	}
1140 }
1141 
1142 static inline void
1143 moea64_copy_pages_nodmap(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1144     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1145 {
1146 	void *a_cp, *b_cp;
1147 	vm_offset_t a_pg_offset, b_pg_offset;
1148 	int cnt;
1149 
1150 	mtx_lock(&moea64_scratchpage_mtx);
1151 	while (xfersize > 0) {
1152 		a_pg_offset = a_offset & PAGE_MASK;
1153 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1154 		moea64_set_scratchpage_pa(mmu, 0,
1155 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]));
1156 		a_cp = (char *)moea64_scratchpage_va[0] + a_pg_offset;
1157 		b_pg_offset = b_offset & PAGE_MASK;
1158 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1159 		moea64_set_scratchpage_pa(mmu, 1,
1160 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]));
1161 		b_cp = (char *)moea64_scratchpage_va[1] + b_pg_offset;
1162 		bcopy(a_cp, b_cp, cnt);
1163 		a_offset += cnt;
1164 		b_offset += cnt;
1165 		xfersize -= cnt;
1166 	}
1167 	mtx_unlock(&moea64_scratchpage_mtx);
1168 }
1169 
1170 void
1171 moea64_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1172     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1173 {
1174 
1175 	if (hw_direct_map) {
1176 		moea64_copy_pages_dmap(mmu, ma, a_offset, mb, b_offset,
1177 		    xfersize);
1178 	} else {
1179 		moea64_copy_pages_nodmap(mmu, ma, a_offset, mb, b_offset,
1180 		    xfersize);
1181 	}
1182 }
1183 
1184 void
1185 moea64_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1186 {
1187 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1188 
1189 	if (size + off > PAGE_SIZE)
1190 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1191 
1192 	if (hw_direct_map) {
1193 		bzero((caddr_t)pa + off, size);
1194 	} else {
1195 		mtx_lock(&moea64_scratchpage_mtx);
1196 		moea64_set_scratchpage_pa(mmu, 0, pa);
1197 		bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1198 		mtx_unlock(&moea64_scratchpage_mtx);
1199 	}
1200 }
1201 
1202 /*
1203  * Zero a page of physical memory by temporarily mapping it
1204  */
1205 void
1206 moea64_zero_page(mmu_t mmu, vm_page_t m)
1207 {
1208 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1209 	vm_offset_t va, off;
1210 
1211 	if (!hw_direct_map) {
1212 		mtx_lock(&moea64_scratchpage_mtx);
1213 
1214 		moea64_set_scratchpage_pa(mmu, 0, pa);
1215 		va = moea64_scratchpage_va[0];
1216 	} else {
1217 		va = pa;
1218 	}
1219 
1220 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1221 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1222 
1223 	if (!hw_direct_map)
1224 		mtx_unlock(&moea64_scratchpage_mtx);
1225 }
1226 
1227 void
1228 moea64_zero_page_idle(mmu_t mmu, vm_page_t m)
1229 {
1230 
1231 	moea64_zero_page(mmu, m);
1232 }
1233 
1234 /*
1235  * Map the given physical page at the specified virtual address in the
1236  * target pmap with the protection requested.  If specified the page
1237  * will be wired down.
1238  */
1239 
1240 void
1241 moea64_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m,
1242     vm_prot_t prot, boolean_t wired)
1243 {
1244 	struct		pvo_head *pvo_head;
1245 	uma_zone_t	zone;
1246 	vm_page_t	pg;
1247 	uint64_t	pte_lo;
1248 	u_int		pvo_flags;
1249 	int		error;
1250 
1251 	if (!moea64_initialized) {
1252 		pvo_head = NULL;
1253 		pg = NULL;
1254 		zone = moea64_upvo_zone;
1255 		pvo_flags = 0;
1256 	} else {
1257 		pvo_head = vm_page_to_pvoh(m);
1258 		pg = m;
1259 		zone = moea64_mpvo_zone;
1260 		pvo_flags = PVO_MANAGED;
1261 	}
1262 
1263 	if ((m->oflags & (VPO_UNMANAGED | VPO_BUSY)) == 0)
1264 		VM_OBJECT_ASSERT_LOCKED(m->object);
1265 
1266 	/* XXX change the pvo head for fake pages */
1267 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1268 		pvo_flags &= ~PVO_MANAGED;
1269 		pvo_head = NULL;
1270 		zone = moea64_upvo_zone;
1271 	}
1272 
1273 	pte_lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1274 
1275 	if (prot & VM_PROT_WRITE) {
1276 		pte_lo |= LPTE_BW;
1277 		if (pmap_bootstrapped &&
1278 		    (m->oflags & VPO_UNMANAGED) == 0)
1279 			vm_page_aflag_set(m, PGA_WRITEABLE);
1280 	} else
1281 		pte_lo |= LPTE_BR;
1282 
1283 	if ((prot & VM_PROT_EXECUTE) == 0)
1284 		pte_lo |= LPTE_NOEXEC;
1285 
1286 	if (wired)
1287 		pvo_flags |= PVO_WIRED;
1288 
1289 	LOCK_TABLE_WR();
1290 	PMAP_LOCK(pmap);
1291 	error = moea64_pvo_enter(mmu, pmap, zone, pvo_head, va,
1292 	    VM_PAGE_TO_PHYS(m), pte_lo, pvo_flags);
1293 	PMAP_UNLOCK(pmap);
1294 	UNLOCK_TABLE_WR();
1295 
1296 	/*
1297 	 * Flush the page from the instruction cache if this page is
1298 	 * mapped executable and cacheable.
1299 	 */
1300 	if (pmap != kernel_pmap && !(m->aflags & PGA_EXECUTABLE) &&
1301 	    (pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1302 		vm_page_aflag_set(m, PGA_EXECUTABLE);
1303 		moea64_syncicache(mmu, pmap, va, VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1304 	}
1305 }
1306 
1307 static void
1308 moea64_syncicache(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_offset_t pa,
1309     vm_size_t sz)
1310 {
1311 
1312 	/*
1313 	 * This is much trickier than on older systems because
1314 	 * we can't sync the icache on physical addresses directly
1315 	 * without a direct map. Instead we check a couple of cases
1316 	 * where the memory is already mapped in and, failing that,
1317 	 * use the same trick we use for page zeroing to create
1318 	 * a temporary mapping for this physical address.
1319 	 */
1320 
1321 	if (!pmap_bootstrapped) {
1322 		/*
1323 		 * If PMAP is not bootstrapped, we are likely to be
1324 		 * in real mode.
1325 		 */
1326 		__syncicache((void *)pa, sz);
1327 	} else if (pmap == kernel_pmap) {
1328 		__syncicache((void *)va, sz);
1329 	} else if (hw_direct_map) {
1330 		__syncicache((void *)pa, sz);
1331 	} else {
1332 		/* Use the scratch page to set up a temp mapping */
1333 
1334 		mtx_lock(&moea64_scratchpage_mtx);
1335 
1336 		moea64_set_scratchpage_pa(mmu, 1, pa & ~ADDR_POFF);
1337 		__syncicache((void *)(moea64_scratchpage_va[1] +
1338 		    (va & ADDR_POFF)), sz);
1339 
1340 		mtx_unlock(&moea64_scratchpage_mtx);
1341 	}
1342 }
1343 
1344 /*
1345  * Maps a sequence of resident pages belonging to the same object.
1346  * The sequence begins with the given page m_start.  This page is
1347  * mapped at the given virtual address start.  Each subsequent page is
1348  * mapped at a virtual address that is offset from start by the same
1349  * amount as the page is offset from m_start within the object.  The
1350  * last page in the sequence is the page with the largest offset from
1351  * m_start that can be mapped at a virtual address less than the given
1352  * virtual address end.  Not every virtual page between start and end
1353  * is mapped; only those for which a resident page exists with the
1354  * corresponding offset from m_start are mapped.
1355  */
1356 void
1357 moea64_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1358     vm_page_t m_start, vm_prot_t prot)
1359 {
1360 	vm_page_t m;
1361 	vm_pindex_t diff, psize;
1362 
1363 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1364 
1365 	psize = atop(end - start);
1366 	m = m_start;
1367 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1368 		moea64_enter(mmu, pm, start + ptoa(diff), m, prot &
1369 		    (VM_PROT_READ | VM_PROT_EXECUTE), FALSE);
1370 		m = TAILQ_NEXT(m, listq);
1371 	}
1372 }
1373 
1374 void
1375 moea64_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1376     vm_prot_t prot)
1377 {
1378 
1379 	moea64_enter(mmu, pm, va, m,
1380 	    prot & (VM_PROT_READ | VM_PROT_EXECUTE), FALSE);
1381 }
1382 
1383 vm_paddr_t
1384 moea64_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1385 {
1386 	struct	pvo_entry *pvo;
1387 	vm_paddr_t pa;
1388 
1389 	PMAP_LOCK(pm);
1390 	pvo = moea64_pvo_find_va(pm, va);
1391 	if (pvo == NULL)
1392 		pa = 0;
1393 	else
1394 		pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) |
1395 		    (va - PVO_VADDR(pvo));
1396 	PMAP_UNLOCK(pm);
1397 	return (pa);
1398 }
1399 
1400 /*
1401  * Atomically extract and hold the physical page with the given
1402  * pmap and virtual address pair if that mapping permits the given
1403  * protection.
1404  */
1405 vm_page_t
1406 moea64_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1407 {
1408 	struct	pvo_entry *pvo;
1409 	vm_page_t m;
1410         vm_paddr_t pa;
1411 
1412 	m = NULL;
1413 	pa = 0;
1414 	PMAP_LOCK(pmap);
1415 retry:
1416 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1417 	if (pvo != NULL && (pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) &&
1418 	    ((pvo->pvo_pte.lpte.pte_lo & LPTE_PP) == LPTE_RW ||
1419 	     (prot & VM_PROT_WRITE) == 0)) {
1420 		if (vm_page_pa_tryrelock(pmap,
1421 			pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN, &pa))
1422 			goto retry;
1423 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN);
1424 		vm_page_hold(m);
1425 	}
1426 	PA_UNLOCK_COND(pa);
1427 	PMAP_UNLOCK(pmap);
1428 	return (m);
1429 }
1430 
1431 static mmu_t installed_mmu;
1432 
1433 static void *
1434 moea64_uma_page_alloc(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
1435 {
1436 	/*
1437 	 * This entire routine is a horrible hack to avoid bothering kmem
1438 	 * for new KVA addresses. Because this can get called from inside
1439 	 * kmem allocation routines, calling kmem for a new address here
1440 	 * can lead to multiply locking non-recursive mutexes.
1441 	 */
1442         vm_offset_t va;
1443 
1444         vm_page_t m;
1445         int pflags, needed_lock;
1446 
1447 	*flags = UMA_SLAB_PRIV;
1448 	needed_lock = !PMAP_LOCKED(kernel_pmap);
1449 	pflags = malloc2vm_flags(wait) | VM_ALLOC_WIRED;
1450 
1451         for (;;) {
1452                 m = vm_page_alloc(NULL, 0, pflags | VM_ALLOC_NOOBJ);
1453                 if (m == NULL) {
1454                         if (wait & M_NOWAIT)
1455                                 return (NULL);
1456                         VM_WAIT;
1457                 } else
1458                         break;
1459         }
1460 
1461 	va = VM_PAGE_TO_PHYS(m);
1462 
1463 	LOCK_TABLE_WR();
1464 	if (needed_lock)
1465 		PMAP_LOCK(kernel_pmap);
1466 
1467 	moea64_pvo_enter(installed_mmu, kernel_pmap, moea64_upvo_zone,
1468 	    NULL, va, VM_PAGE_TO_PHYS(m), LPTE_M, PVO_WIRED | PVO_BOOTSTRAP);
1469 
1470 	if (needed_lock)
1471 		PMAP_UNLOCK(kernel_pmap);
1472 	UNLOCK_TABLE_WR();
1473 
1474 	if ((wait & M_ZERO) && (m->flags & PG_ZERO) == 0)
1475                 bzero((void *)va, PAGE_SIZE);
1476 
1477 	return (void *)va;
1478 }
1479 
1480 extern int elf32_nxstack;
1481 
1482 void
1483 moea64_init(mmu_t mmu)
1484 {
1485 
1486 	CTR0(KTR_PMAP, "moea64_init");
1487 
1488 	moea64_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1489 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1490 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1491 	moea64_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1492 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1493 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1494 
1495 	if (!hw_direct_map) {
1496 		installed_mmu = mmu;
1497 		uma_zone_set_allocf(moea64_upvo_zone,moea64_uma_page_alloc);
1498 		uma_zone_set_allocf(moea64_mpvo_zone,moea64_uma_page_alloc);
1499 	}
1500 
1501 #ifdef COMPAT_FREEBSD32
1502 	elf32_nxstack = 1;
1503 #endif
1504 
1505 	moea64_initialized = TRUE;
1506 }
1507 
1508 boolean_t
1509 moea64_is_referenced(mmu_t mmu, vm_page_t m)
1510 {
1511 
1512 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1513 	    ("moea64_is_referenced: page %p is not managed", m));
1514 	return (moea64_query_bit(mmu, m, PTE_REF));
1515 }
1516 
1517 boolean_t
1518 moea64_is_modified(mmu_t mmu, vm_page_t m)
1519 {
1520 
1521 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1522 	    ("moea64_is_modified: page %p is not managed", m));
1523 
1524 	/*
1525 	 * If the page is not VPO_BUSY, then PGA_WRITEABLE cannot be
1526 	 * concurrently set while the object is locked.  Thus, if PGA_WRITEABLE
1527 	 * is clear, no PTEs can have LPTE_CHG set.
1528 	 */
1529 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1530 	if ((m->oflags & VPO_BUSY) == 0 &&
1531 	    (m->aflags & PGA_WRITEABLE) == 0)
1532 		return (FALSE);
1533 	return (moea64_query_bit(mmu, m, LPTE_CHG));
1534 }
1535 
1536 boolean_t
1537 moea64_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1538 {
1539 	struct pvo_entry *pvo;
1540 	boolean_t rv;
1541 
1542 	PMAP_LOCK(pmap);
1543 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1544 	rv = pvo == NULL || (pvo->pvo_pte.lpte.pte_hi & LPTE_VALID) == 0;
1545 	PMAP_UNLOCK(pmap);
1546 	return (rv);
1547 }
1548 
1549 void
1550 moea64_clear_reference(mmu_t mmu, vm_page_t m)
1551 {
1552 
1553 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1554 	    ("moea64_clear_reference: page %p is not managed", m));
1555 	moea64_clear_bit(mmu, m, LPTE_REF);
1556 }
1557 
1558 void
1559 moea64_clear_modify(mmu_t mmu, vm_page_t m)
1560 {
1561 
1562 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1563 	    ("moea64_clear_modify: page %p is not managed", m));
1564 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1565 	KASSERT((m->oflags & VPO_BUSY) == 0,
1566 	    ("moea64_clear_modify: page %p is busy", m));
1567 
1568 	/*
1569 	 * If the page is not PGA_WRITEABLE, then no PTEs can have LPTE_CHG
1570 	 * set.  If the object containing the page is locked and the page is
1571 	 * not VPO_BUSY, then PGA_WRITEABLE cannot be concurrently set.
1572 	 */
1573 	if ((m->aflags & PGA_WRITEABLE) == 0)
1574 		return;
1575 	moea64_clear_bit(mmu, m, LPTE_CHG);
1576 }
1577 
1578 /*
1579  * Clear the write and modified bits in each of the given page's mappings.
1580  */
1581 void
1582 moea64_remove_write(mmu_t mmu, vm_page_t m)
1583 {
1584 	struct	pvo_entry *pvo;
1585 	uintptr_t pt;
1586 	pmap_t	pmap;
1587 	uint64_t lo = 0;
1588 
1589 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1590 	    ("moea64_remove_write: page %p is not managed", m));
1591 
1592 	/*
1593 	 * If the page is not VPO_BUSY, then PGA_WRITEABLE cannot be set by
1594 	 * another thread while the object is locked.  Thus, if PGA_WRITEABLE
1595 	 * is clear, no page table entries need updating.
1596 	 */
1597 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1598 	if ((m->oflags & VPO_BUSY) == 0 &&
1599 	    (m->aflags & PGA_WRITEABLE) == 0)
1600 		return;
1601 	powerpc_sync();
1602 	LOCK_TABLE_RD();
1603 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1604 		pmap = pvo->pvo_pmap;
1605 		PMAP_LOCK(pmap);
1606 		if ((pvo->pvo_pte.lpte.pte_lo & LPTE_PP) != LPTE_BR) {
1607 			pt = MOEA64_PVO_TO_PTE(mmu, pvo);
1608 			pvo->pvo_pte.lpte.pte_lo &= ~LPTE_PP;
1609 			pvo->pvo_pte.lpte.pte_lo |= LPTE_BR;
1610 			if (pt != -1) {
1611 				MOEA64_PTE_SYNCH(mmu, pt, &pvo->pvo_pte.lpte);
1612 				lo |= pvo->pvo_pte.lpte.pte_lo;
1613 				pvo->pvo_pte.lpte.pte_lo &= ~LPTE_CHG;
1614 				MOEA64_PTE_CHANGE(mmu, pt,
1615 				    &pvo->pvo_pte.lpte, pvo->pvo_vpn);
1616 				if (pvo->pvo_pmap == kernel_pmap)
1617 					isync();
1618 			}
1619 		}
1620 		if ((lo & LPTE_CHG) != 0)
1621 			vm_page_dirty(m);
1622 		PMAP_UNLOCK(pmap);
1623 	}
1624 	UNLOCK_TABLE_RD();
1625 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1626 }
1627 
1628 /*
1629  *	moea64_ts_referenced:
1630  *
1631  *	Return a count of reference bits for a page, clearing those bits.
1632  *	It is not necessary for every reference bit to be cleared, but it
1633  *	is necessary that 0 only be returned when there are truly no
1634  *	reference bits set.
1635  *
1636  *	XXX: The exact number of bits to check and clear is a matter that
1637  *	should be tested and standardized at some point in the future for
1638  *	optimal aging of shared pages.
1639  */
1640 int
1641 moea64_ts_referenced(mmu_t mmu, vm_page_t m)
1642 {
1643 
1644 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1645 	    ("moea64_ts_referenced: page %p is not managed", m));
1646 	return (moea64_clear_bit(mmu, m, LPTE_REF));
1647 }
1648 
1649 /*
1650  * Modify the WIMG settings of all mappings for a page.
1651  */
1652 void
1653 moea64_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1654 {
1655 	struct	pvo_entry *pvo;
1656 	struct  pvo_head *pvo_head;
1657 	uintptr_t pt;
1658 	pmap_t	pmap;
1659 	uint64_t lo;
1660 
1661 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1662 		m->md.mdpg_cache_attrs = ma;
1663 		return;
1664 	}
1665 
1666 	pvo_head = vm_page_to_pvoh(m);
1667 	lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1668 	LOCK_TABLE_RD();
1669 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1670 		pmap = pvo->pvo_pmap;
1671 		PMAP_LOCK(pmap);
1672 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
1673 		pvo->pvo_pte.lpte.pte_lo &= ~LPTE_WIMG;
1674 		pvo->pvo_pte.lpte.pte_lo |= lo;
1675 		if (pt != -1) {
1676 			MOEA64_PTE_CHANGE(mmu, pt, &pvo->pvo_pte.lpte,
1677 			    pvo->pvo_vpn);
1678 			if (pvo->pvo_pmap == kernel_pmap)
1679 				isync();
1680 		}
1681 		PMAP_UNLOCK(pmap);
1682 	}
1683 	UNLOCK_TABLE_RD();
1684 	m->md.mdpg_cache_attrs = ma;
1685 }
1686 
1687 /*
1688  * Map a wired page into kernel virtual address space.
1689  */
1690 void
1691 moea64_kenter_attr(mmu_t mmu, vm_offset_t va, vm_offset_t pa, vm_memattr_t ma)
1692 {
1693 	uint64_t	pte_lo;
1694 	int		error;
1695 
1696 	pte_lo = moea64_calc_wimg(pa, ma);
1697 
1698 	LOCK_TABLE_WR();
1699 	PMAP_LOCK(kernel_pmap);
1700 	error = moea64_pvo_enter(mmu, kernel_pmap, moea64_upvo_zone,
1701 	    NULL, va, pa, pte_lo, PVO_WIRED);
1702 	PMAP_UNLOCK(kernel_pmap);
1703 	UNLOCK_TABLE_WR();
1704 
1705 	if (error != 0 && error != ENOENT)
1706 		panic("moea64_kenter: failed to enter va %#zx pa %#zx: %d", va,
1707 		    pa, error);
1708 }
1709 
1710 void
1711 moea64_kenter(mmu_t mmu, vm_offset_t va, vm_paddr_t pa)
1712 {
1713 
1714 	moea64_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1715 }
1716 
1717 /*
1718  * Extract the physical page address associated with the given kernel virtual
1719  * address.
1720  */
1721 vm_paddr_t
1722 moea64_kextract(mmu_t mmu, vm_offset_t va)
1723 {
1724 	struct		pvo_entry *pvo;
1725 	vm_paddr_t pa;
1726 
1727 	/*
1728 	 * Shortcut the direct-mapped case when applicable.  We never put
1729 	 * anything but 1:1 mappings below VM_MIN_KERNEL_ADDRESS.
1730 	 */
1731 	if (va < VM_MIN_KERNEL_ADDRESS)
1732 		return (va);
1733 
1734 	PMAP_LOCK(kernel_pmap);
1735 	pvo = moea64_pvo_find_va(kernel_pmap, va);
1736 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found for %#" PRIxPTR,
1737 	    va));
1738 	pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) | (va - PVO_VADDR(pvo));
1739 	PMAP_UNLOCK(kernel_pmap);
1740 	return (pa);
1741 }
1742 
1743 /*
1744  * Remove a wired page from kernel virtual address space.
1745  */
1746 void
1747 moea64_kremove(mmu_t mmu, vm_offset_t va)
1748 {
1749 	moea64_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1750 }
1751 
1752 /*
1753  * Map a range of physical addresses into kernel virtual address space.
1754  *
1755  * The value passed in *virt is a suggested virtual address for the mapping.
1756  * Architectures which can support a direct-mapped physical to virtual region
1757  * can return the appropriate address within that region, leaving '*virt'
1758  * unchanged.  We cannot and therefore do not; *virt is updated with the
1759  * first usable address after the mapped region.
1760  */
1761 vm_offset_t
1762 moea64_map(mmu_t mmu, vm_offset_t *virt, vm_paddr_t pa_start,
1763     vm_paddr_t pa_end, int prot)
1764 {
1765 	vm_offset_t	sva, va;
1766 
1767 	sva = *virt;
1768 	va = sva;
1769 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1770 		moea64_kenter(mmu, va, pa_start);
1771 	*virt = va;
1772 
1773 	return (sva);
1774 }
1775 
1776 /*
1777  * Returns true if the pmap's pv is one of the first
1778  * 16 pvs linked to from this page.  This count may
1779  * be changed upwards or downwards in the future; it
1780  * is only necessary that true be returned for a small
1781  * subset of pmaps for proper page aging.
1782  */
1783 boolean_t
1784 moea64_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1785 {
1786         int loops;
1787 	struct pvo_entry *pvo;
1788 	boolean_t rv;
1789 
1790 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1791 	    ("moea64_page_exists_quick: page %p is not managed", m));
1792 	loops = 0;
1793 	rv = FALSE;
1794 	LOCK_TABLE_RD();
1795 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1796 		if (pvo->pvo_pmap == pmap) {
1797 			rv = TRUE;
1798 			break;
1799 		}
1800 		if (++loops >= 16)
1801 			break;
1802 	}
1803 	UNLOCK_TABLE_RD();
1804 	return (rv);
1805 }
1806 
1807 /*
1808  * Return the number of managed mappings to the given physical page
1809  * that are wired.
1810  */
1811 int
1812 moea64_page_wired_mappings(mmu_t mmu, vm_page_t m)
1813 {
1814 	struct pvo_entry *pvo;
1815 	int count;
1816 
1817 	count = 0;
1818 	if ((m->oflags & VPO_UNMANAGED) != 0)
1819 		return (count);
1820 	LOCK_TABLE_RD();
1821 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1822 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1823 			count++;
1824 	UNLOCK_TABLE_RD();
1825 	return (count);
1826 }
1827 
1828 static uintptr_t	moea64_vsidcontext;
1829 
1830 uintptr_t
1831 moea64_get_unique_vsid(void) {
1832 	u_int entropy;
1833 	register_t hash;
1834 	uint32_t mask;
1835 	int i;
1836 
1837 	entropy = 0;
1838 	__asm __volatile("mftb %0" : "=r"(entropy));
1839 
1840 	mtx_lock(&moea64_slb_mutex);
1841 	for (i = 0; i < NVSIDS; i += VSID_NBPW) {
1842 		u_int	n;
1843 
1844 		/*
1845 		 * Create a new value by mutiplying by a prime and adding in
1846 		 * entropy from the timebase register.  This is to make the
1847 		 * VSID more random so that the PT hash function collides
1848 		 * less often.  (Note that the prime casues gcc to do shifts
1849 		 * instead of a multiply.)
1850 		 */
1851 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
1852 		hash = moea64_vsidcontext & (NVSIDS - 1);
1853 		if (hash == 0)		/* 0 is special, avoid it */
1854 			continue;
1855 		n = hash >> 5;
1856 		mask = 1 << (hash & (VSID_NBPW - 1));
1857 		hash = (moea64_vsidcontext & VSID_HASHMASK);
1858 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
1859 			/* anything free in this bucket? */
1860 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
1861 				entropy = (moea64_vsidcontext >> 20);
1862 				continue;
1863 			}
1864 			i = ffs(~moea64_vsid_bitmap[n]) - 1;
1865 			mask = 1 << i;
1866 			hash &= VSID_HASHMASK & ~(VSID_NBPW - 1);
1867 			hash |= i;
1868 		}
1869 		KASSERT(!(moea64_vsid_bitmap[n] & mask),
1870 		    ("Allocating in-use VSID %#zx\n", hash));
1871 		moea64_vsid_bitmap[n] |= mask;
1872 		mtx_unlock(&moea64_slb_mutex);
1873 		return (hash);
1874 	}
1875 
1876 	mtx_unlock(&moea64_slb_mutex);
1877 	panic("%s: out of segments",__func__);
1878 }
1879 
1880 #ifdef __powerpc64__
1881 void
1882 moea64_pinit(mmu_t mmu, pmap_t pmap)
1883 {
1884 	PMAP_LOCK_INIT(pmap);
1885 	RB_INIT(&pmap->pmap_pvo);
1886 
1887 	pmap->pm_slb_tree_root = slb_alloc_tree();
1888 	pmap->pm_slb = slb_alloc_user_cache();
1889 	pmap->pm_slb_len = 0;
1890 }
1891 #else
1892 void
1893 moea64_pinit(mmu_t mmu, pmap_t pmap)
1894 {
1895 	int	i;
1896 	uint32_t hash;
1897 
1898 	PMAP_LOCK_INIT(pmap);
1899 	RB_INIT(&pmap->pmap_pvo);
1900 
1901 	if (pmap_bootstrapped)
1902 		pmap->pmap_phys = (pmap_t)moea64_kextract(mmu,
1903 		    (vm_offset_t)pmap);
1904 	else
1905 		pmap->pmap_phys = pmap;
1906 
1907 	/*
1908 	 * Allocate some segment registers for this pmap.
1909 	 */
1910 	hash = moea64_get_unique_vsid();
1911 
1912 	for (i = 0; i < 16; i++)
1913 		pmap->pm_sr[i] = VSID_MAKE(i, hash);
1914 
1915 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_pinit: pm_sr[0] = 0"));
1916 }
1917 #endif
1918 
1919 /*
1920  * Initialize the pmap associated with process 0.
1921  */
1922 void
1923 moea64_pinit0(mmu_t mmu, pmap_t pm)
1924 {
1925 	moea64_pinit(mmu, pm);
1926 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1927 }
1928 
1929 /*
1930  * Set the physical protection on the specified range of this map as requested.
1931  */
1932 static void
1933 moea64_pvo_protect(mmu_t mmu,  pmap_t pm, struct pvo_entry *pvo, vm_prot_t prot)
1934 {
1935 	uintptr_t pt;
1936 	struct	vm_page *pg;
1937 	uint64_t oldlo;
1938 
1939 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
1940 
1941 	/*
1942 	 * Grab the PTE pointer before we diddle with the cached PTE
1943 	 * copy.
1944 	 */
1945 	pt = MOEA64_PVO_TO_PTE(mmu, pvo);
1946 
1947 	/*
1948 	 * Change the protection of the page.
1949 	 */
1950 	oldlo = pvo->pvo_pte.lpte.pte_lo;
1951 	pvo->pvo_pte.lpte.pte_lo &= ~LPTE_PP;
1952 	pvo->pvo_pte.lpte.pte_lo &= ~LPTE_NOEXEC;
1953 	if ((prot & VM_PROT_EXECUTE) == 0)
1954 		pvo->pvo_pte.lpte.pte_lo |= LPTE_NOEXEC;
1955 	if (prot & VM_PROT_WRITE)
1956 		pvo->pvo_pte.lpte.pte_lo |= LPTE_BW;
1957 	else
1958 		pvo->pvo_pte.lpte.pte_lo |= LPTE_BR;
1959 
1960 	pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN);
1961 
1962 	/*
1963 	 * If the PVO is in the page table, update that pte as well.
1964 	 */
1965 	if (pt != -1)
1966 		MOEA64_PTE_CHANGE(mmu, pt, &pvo->pvo_pte.lpte,
1967 		    pvo->pvo_vpn);
1968 	if (pm != kernel_pmap && pg != NULL && !(pg->aflags & PGA_EXECUTABLE) &&
1969 	    (pvo->pvo_pte.lpte.pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1970 		if ((pg->oflags & VPO_UNMANAGED) == 0)
1971 			vm_page_aflag_set(pg, PGA_EXECUTABLE);
1972 		moea64_syncicache(mmu, pm, PVO_VADDR(pvo),
1973 		    pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN, PAGE_SIZE);
1974 	}
1975 
1976 	/*
1977 	 * Update vm about the REF/CHG bits if the page is managed and we have
1978 	 * removed write access.
1979 	 */
1980 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED &&
1981 	    (oldlo & LPTE_PP) != LPTE_BR && !(prot & VM_PROT_WRITE)) {
1982 		if (pg != NULL) {
1983 			if (pvo->pvo_pte.lpte.pte_lo & LPTE_CHG)
1984 				vm_page_dirty(pg);
1985 			if (pvo->pvo_pte.lpte.pte_lo & LPTE_REF)
1986 				vm_page_aflag_set(pg, PGA_REFERENCED);
1987 		}
1988 	}
1989 }
1990 
1991 void
1992 moea64_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1993     vm_prot_t prot)
1994 {
1995 	struct	pvo_entry *pvo, *tpvo, key;
1996 
1997 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm,
1998 	    sva, eva, prot);
1999 
2000 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
2001 	    ("moea64_protect: non current pmap"));
2002 
2003 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2004 		moea64_remove(mmu, pm, sva, eva);
2005 		return;
2006 	}
2007 
2008 	LOCK_TABLE_RD();
2009 	PMAP_LOCK(pm);
2010 	key.pvo_vaddr = sva;
2011 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2012 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2013 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2014 		moea64_pvo_protect(mmu, pm, pvo, prot);
2015 	}
2016 	UNLOCK_TABLE_RD();
2017 	PMAP_UNLOCK(pm);
2018 }
2019 
2020 /*
2021  * Map a list of wired pages into kernel virtual address space.  This is
2022  * intended for temporary mappings which do not need page modification or
2023  * references recorded.  Existing mappings in the region are overwritten.
2024  */
2025 void
2026 moea64_qenter(mmu_t mmu, vm_offset_t va, vm_page_t *m, int count)
2027 {
2028 	while (count-- > 0) {
2029 		moea64_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
2030 		va += PAGE_SIZE;
2031 		m++;
2032 	}
2033 }
2034 
2035 /*
2036  * Remove page mappings from kernel virtual address space.  Intended for
2037  * temporary mappings entered by moea64_qenter.
2038  */
2039 void
2040 moea64_qremove(mmu_t mmu, vm_offset_t va, int count)
2041 {
2042 	while (count-- > 0) {
2043 		moea64_kremove(mmu, va);
2044 		va += PAGE_SIZE;
2045 	}
2046 }
2047 
2048 void
2049 moea64_release_vsid(uint64_t vsid)
2050 {
2051 	int idx, mask;
2052 
2053 	mtx_lock(&moea64_slb_mutex);
2054 	idx = vsid & (NVSIDS-1);
2055 	mask = 1 << (idx % VSID_NBPW);
2056 	idx /= VSID_NBPW;
2057 	KASSERT(moea64_vsid_bitmap[idx] & mask,
2058 	    ("Freeing unallocated VSID %#jx", vsid));
2059 	moea64_vsid_bitmap[idx] &= ~mask;
2060 	mtx_unlock(&moea64_slb_mutex);
2061 }
2062 
2063 
2064 void
2065 moea64_release(mmu_t mmu, pmap_t pmap)
2066 {
2067 
2068 	/*
2069 	 * Free segment registers' VSIDs
2070 	 */
2071     #ifdef __powerpc64__
2072 	slb_free_tree(pmap);
2073 	slb_free_user_cache(pmap->pm_slb);
2074     #else
2075 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_release: pm_sr[0] = 0"));
2076 
2077 	moea64_release_vsid(VSID_TO_HASH(pmap->pm_sr[0]));
2078     #endif
2079 
2080 	PMAP_LOCK_DESTROY(pmap);
2081 }
2082 
2083 /*
2084  * Remove all pages mapped by the specified pmap
2085  */
2086 void
2087 moea64_remove_pages(mmu_t mmu, pmap_t pm)
2088 {
2089 	struct	pvo_entry *pvo, *tpvo;
2090 
2091 	LOCK_TABLE_WR();
2092 	PMAP_LOCK(pm);
2093 	RB_FOREACH_SAFE(pvo, pvo_tree, &pm->pmap_pvo, tpvo) {
2094 		if (!(pvo->pvo_vaddr & PVO_WIRED))
2095 			moea64_pvo_remove(mmu, pvo);
2096 	}
2097 	UNLOCK_TABLE_WR();
2098 	PMAP_UNLOCK(pm);
2099 }
2100 
2101 /*
2102  * Remove the given range of addresses from the specified map.
2103  */
2104 void
2105 moea64_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
2106 {
2107 	struct	pvo_entry *pvo, *tpvo, key;
2108 
2109 	/*
2110 	 * Perform an unsynchronized read.  This is, however, safe.
2111 	 */
2112 	if (pm->pm_stats.resident_count == 0)
2113 		return;
2114 
2115 	LOCK_TABLE_WR();
2116 	PMAP_LOCK(pm);
2117 	key.pvo_vaddr = sva;
2118 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2119 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2120 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2121 		moea64_pvo_remove(mmu, pvo);
2122 	}
2123 	UNLOCK_TABLE_WR();
2124 	PMAP_UNLOCK(pm);
2125 }
2126 
2127 /*
2128  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
2129  * will reflect changes in pte's back to the vm_page.
2130  */
2131 void
2132 moea64_remove_all(mmu_t mmu, vm_page_t m)
2133 {
2134 	struct	pvo_entry *pvo, *next_pvo;
2135 	pmap_t	pmap;
2136 
2137 	LOCK_TABLE_WR();
2138 	LIST_FOREACH_SAFE(pvo, vm_page_to_pvoh(m), pvo_vlink, next_pvo) {
2139 		pmap = pvo->pvo_pmap;
2140 		PMAP_LOCK(pmap);
2141 		moea64_pvo_remove(mmu, pvo);
2142 		PMAP_UNLOCK(pmap);
2143 	}
2144 	UNLOCK_TABLE_WR();
2145 	if ((m->aflags & PGA_WRITEABLE) && moea64_is_modified(mmu, m))
2146 		vm_page_dirty(m);
2147 	vm_page_aflag_clear(m, PGA_WRITEABLE);
2148 	vm_page_aflag_clear(m, PGA_EXECUTABLE);
2149 }
2150 
2151 /*
2152  * Allocate a physical page of memory directly from the phys_avail map.
2153  * Can only be called from moea64_bootstrap before avail start and end are
2154  * calculated.
2155  */
2156 vm_offset_t
2157 moea64_bootstrap_alloc(vm_size_t size, u_int align)
2158 {
2159 	vm_offset_t	s, e;
2160 	int		i, j;
2161 
2162 	size = round_page(size);
2163 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2164 		if (align != 0)
2165 			s = (phys_avail[i] + align - 1) & ~(align - 1);
2166 		else
2167 			s = phys_avail[i];
2168 		e = s + size;
2169 
2170 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2171 			continue;
2172 
2173 		if (s + size > platform_real_maxaddr())
2174 			continue;
2175 
2176 		if (s == phys_avail[i]) {
2177 			phys_avail[i] += size;
2178 		} else if (e == phys_avail[i + 1]) {
2179 			phys_avail[i + 1] -= size;
2180 		} else {
2181 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2182 				phys_avail[j] = phys_avail[j - 2];
2183 				phys_avail[j + 1] = phys_avail[j - 1];
2184 			}
2185 
2186 			phys_avail[i + 3] = phys_avail[i + 1];
2187 			phys_avail[i + 1] = s;
2188 			phys_avail[i + 2] = e;
2189 			phys_avail_count++;
2190 		}
2191 
2192 		return (s);
2193 	}
2194 	panic("moea64_bootstrap_alloc: could not allocate memory");
2195 }
2196 
2197 static int
2198 moea64_pvo_enter(mmu_t mmu, pmap_t pm, uma_zone_t zone,
2199     struct pvo_head *pvo_head, vm_offset_t va, vm_offset_t pa,
2200     uint64_t pte_lo, int flags)
2201 {
2202 	struct	 pvo_entry *pvo;
2203 	uint64_t vsid;
2204 	int	 first;
2205 	u_int	 ptegidx;
2206 	int	 i;
2207 	int      bootstrap;
2208 
2209 	/*
2210 	 * One nasty thing that can happen here is that the UMA calls to
2211 	 * allocate new PVOs need to map more memory, which calls pvo_enter(),
2212 	 * which calls UMA...
2213 	 *
2214 	 * We break the loop by detecting recursion and allocating out of
2215 	 * the bootstrap pool.
2216 	 */
2217 
2218 	first = 0;
2219 	bootstrap = (flags & PVO_BOOTSTRAP);
2220 
2221 	if (!moea64_initialized)
2222 		bootstrap = 1;
2223 
2224 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2225 	rw_assert(&moea64_table_lock, RA_WLOCKED);
2226 
2227 	/*
2228 	 * Compute the PTE Group index.
2229 	 */
2230 	va &= ~ADDR_POFF;
2231 	vsid = va_to_vsid(pm, va);
2232 	ptegidx = va_to_pteg(vsid, va, flags & PVO_LARGE);
2233 
2234 	/*
2235 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
2236 	 * there is a mapping.
2237 	 */
2238 	moea64_pvo_enter_calls++;
2239 
2240 	LIST_FOREACH(pvo, &moea64_pvo_table[ptegidx], pvo_olink) {
2241 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2242 			if ((pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) == pa &&
2243 			    (pvo->pvo_pte.lpte.pte_lo & (LPTE_NOEXEC | LPTE_PP))
2244 			    == (pte_lo & (LPTE_NOEXEC | LPTE_PP))) {
2245 			    	if (!(pvo->pvo_pte.lpte.pte_hi & LPTE_VALID)) {
2246 					/* Re-insert if spilled */
2247 					i = MOEA64_PTE_INSERT(mmu, ptegidx,
2248 					    &pvo->pvo_pte.lpte);
2249 					if (i >= 0)
2250 						PVO_PTEGIDX_SET(pvo, i);
2251 					moea64_pte_overflow--;
2252 				}
2253 				return (0);
2254 			}
2255 			moea64_pvo_remove(mmu, pvo);
2256 			break;
2257 		}
2258 	}
2259 
2260 	/*
2261 	 * If we aren't overwriting a mapping, try to allocate.
2262 	 */
2263 	if (bootstrap) {
2264 		if (moea64_bpvo_pool_index >= BPVO_POOL_SIZE) {
2265 			panic("moea64_enter: bpvo pool exhausted, %d, %d, %zd",
2266 			      moea64_bpvo_pool_index, BPVO_POOL_SIZE,
2267 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
2268 		}
2269 		pvo = &moea64_bpvo_pool[moea64_bpvo_pool_index];
2270 		moea64_bpvo_pool_index++;
2271 		bootstrap = 1;
2272 	} else {
2273 		pvo = uma_zalloc(zone, M_NOWAIT);
2274 	}
2275 
2276 	if (pvo == NULL)
2277 		return (ENOMEM);
2278 
2279 	moea64_pvo_entries++;
2280 	pvo->pvo_vaddr = va;
2281 	pvo->pvo_vpn = (uint64_t)((va & ADDR_PIDX) >> ADDR_PIDX_SHFT)
2282 	    | (vsid << 16);
2283 	pvo->pvo_pmap = pm;
2284 	LIST_INSERT_HEAD(&moea64_pvo_table[ptegidx], pvo, pvo_olink);
2285 	pvo->pvo_vaddr &= ~ADDR_POFF;
2286 
2287 	if (flags & PVO_WIRED)
2288 		pvo->pvo_vaddr |= PVO_WIRED;
2289 	if (pvo_head != NULL)
2290 		pvo->pvo_vaddr |= PVO_MANAGED;
2291 	if (bootstrap)
2292 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
2293 	if (flags & PVO_LARGE)
2294 		pvo->pvo_vaddr |= PVO_LARGE;
2295 
2296 	moea64_pte_create(&pvo->pvo_pte.lpte, vsid, va,
2297 	    (uint64_t)(pa) | pte_lo, flags);
2298 
2299 	/*
2300 	 * Add to pmap list
2301 	 */
2302 	RB_INSERT(pvo_tree, &pm->pmap_pvo, pvo);
2303 
2304 	/*
2305 	 * Remember if the list was empty and therefore will be the first
2306 	 * item.
2307 	 */
2308 	if (pvo_head != NULL) {
2309 		if (LIST_FIRST(pvo_head) == NULL)
2310 			first = 1;
2311 		LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2312 	}
2313 
2314 	if (pvo->pvo_vaddr & PVO_WIRED) {
2315 		pvo->pvo_pte.lpte.pte_hi |= LPTE_WIRED;
2316 		pm->pm_stats.wired_count++;
2317 	}
2318 	pm->pm_stats.resident_count++;
2319 
2320 	/*
2321 	 * We hope this succeeds but it isn't required.
2322 	 */
2323 	i = MOEA64_PTE_INSERT(mmu, ptegidx, &pvo->pvo_pte.lpte);
2324 	if (i >= 0) {
2325 		PVO_PTEGIDX_SET(pvo, i);
2326 	} else {
2327 		panic("moea64_pvo_enter: overflow");
2328 		moea64_pte_overflow++;
2329 	}
2330 
2331 	if (pm == kernel_pmap)
2332 		isync();
2333 
2334 #ifdef __powerpc64__
2335 	/*
2336 	 * Make sure all our bootstrap mappings are in the SLB as soon
2337 	 * as virtual memory is switched on.
2338 	 */
2339 	if (!pmap_bootstrapped)
2340 		moea64_bootstrap_slb_prefault(va, flags & PVO_LARGE);
2341 #endif
2342 
2343 	return (first ? ENOENT : 0);
2344 }
2345 
2346 static void
2347 moea64_pvo_remove(mmu_t mmu, struct pvo_entry *pvo)
2348 {
2349 	struct	vm_page *pg;
2350 	uintptr_t pt;
2351 
2352 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2353 	rw_assert(&moea64_table_lock, RA_WLOCKED);
2354 
2355 	/*
2356 	 * If there is an active pte entry, we need to deactivate it (and
2357 	 * save the ref & cfg bits).
2358 	 */
2359 	pt = MOEA64_PVO_TO_PTE(mmu, pvo);
2360 	if (pt != -1) {
2361 		MOEA64_PTE_UNSET(mmu, pt, &pvo->pvo_pte.lpte, pvo->pvo_vpn);
2362 		PVO_PTEGIDX_CLR(pvo);
2363 	} else {
2364 		moea64_pte_overflow--;
2365 	}
2366 
2367 	/*
2368 	 * Update our statistics.
2369 	 */
2370 	pvo->pvo_pmap->pm_stats.resident_count--;
2371 	if (pvo->pvo_vaddr & PVO_WIRED)
2372 		pvo->pvo_pmap->pm_stats.wired_count--;
2373 
2374 	/*
2375 	 * Remove this PVO from the pmap list.
2376 	 */
2377 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2378 
2379 	/*
2380 	 * Remove this from the overflow list and return it to the pool
2381 	 * if we aren't going to reuse it.
2382 	 */
2383 	LIST_REMOVE(pvo, pvo_olink);
2384 
2385 	/*
2386 	 * Update vm about the REF/CHG bits if the page is managed.
2387 	 */
2388 	pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN);
2389 
2390 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED && pg != NULL) {
2391 		LIST_REMOVE(pvo, pvo_vlink);
2392 		if ((pvo->pvo_pte.lpte.pte_lo & LPTE_PP) != LPTE_BR) {
2393 			if (pvo->pvo_pte.lpte.pte_lo & LPTE_CHG)
2394 				vm_page_dirty(pg);
2395 			if (pvo->pvo_pte.lpte.pte_lo & LPTE_REF)
2396 				vm_page_aflag_set(pg, PGA_REFERENCED);
2397 			if (LIST_EMPTY(vm_page_to_pvoh(pg)))
2398 				vm_page_aflag_clear(pg, PGA_WRITEABLE);
2399 		}
2400 		if (LIST_EMPTY(vm_page_to_pvoh(pg)))
2401 			vm_page_aflag_clear(pg, PGA_EXECUTABLE);
2402 	}
2403 
2404 	moea64_pvo_entries--;
2405 	moea64_pvo_remove_calls++;
2406 
2407 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2408 		uma_zfree((pvo->pvo_vaddr & PVO_MANAGED) ? moea64_mpvo_zone :
2409 		    moea64_upvo_zone, pvo);
2410 }
2411 
2412 static struct pvo_entry *
2413 moea64_pvo_find_va(pmap_t pm, vm_offset_t va)
2414 {
2415 	struct pvo_entry key;
2416 
2417 	key.pvo_vaddr = va & ~ADDR_POFF;
2418 	return (RB_FIND(pvo_tree, &pm->pmap_pvo, &key));
2419 }
2420 
2421 static boolean_t
2422 moea64_query_bit(mmu_t mmu, vm_page_t m, u_int64_t ptebit)
2423 {
2424 	struct	pvo_entry *pvo;
2425 	uintptr_t pt;
2426 
2427 	LOCK_TABLE_RD();
2428 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2429 		/*
2430 		 * See if we saved the bit off.  If so, return success.
2431 		 */
2432 		if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2433 			UNLOCK_TABLE_RD();
2434 			return (TRUE);
2435 		}
2436 	}
2437 
2438 	/*
2439 	 * No luck, now go through the hard part of looking at the PTEs
2440 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2441 	 * the PTEs.
2442 	 */
2443 	powerpc_sync();
2444 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2445 
2446 		/*
2447 		 * See if this pvo has a valid PTE.  if so, fetch the
2448 		 * REF/CHG bits from the valid PTE.  If the appropriate
2449 		 * ptebit is set, return success.
2450 		 */
2451 		PMAP_LOCK(pvo->pvo_pmap);
2452 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
2453 		if (pt != -1) {
2454 			MOEA64_PTE_SYNCH(mmu, pt, &pvo->pvo_pte.lpte);
2455 			if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2456 				PMAP_UNLOCK(pvo->pvo_pmap);
2457 				UNLOCK_TABLE_RD();
2458 				return (TRUE);
2459 			}
2460 		}
2461 		PMAP_UNLOCK(pvo->pvo_pmap);
2462 	}
2463 
2464 	UNLOCK_TABLE_RD();
2465 	return (FALSE);
2466 }
2467 
2468 static u_int
2469 moea64_clear_bit(mmu_t mmu, vm_page_t m, u_int64_t ptebit)
2470 {
2471 	u_int	count;
2472 	struct	pvo_entry *pvo;
2473 	uintptr_t pt;
2474 
2475 	/*
2476 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2477 	 * we can reset the right ones).  note that since the pvo entries and
2478 	 * list heads are accessed via BAT0 and are never placed in the page
2479 	 * table, we don't have to worry about further accesses setting the
2480 	 * REF/CHG bits.
2481 	 */
2482 	powerpc_sync();
2483 
2484 	/*
2485 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2486 	 * valid pte clear the ptebit from the valid pte.
2487 	 */
2488 	count = 0;
2489 	LOCK_TABLE_RD();
2490 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2491 		PMAP_LOCK(pvo->pvo_pmap);
2492 		pt = MOEA64_PVO_TO_PTE(mmu, pvo);
2493 		if (pt != -1) {
2494 			MOEA64_PTE_SYNCH(mmu, pt, &pvo->pvo_pte.lpte);
2495 			if (pvo->pvo_pte.lpte.pte_lo & ptebit) {
2496 				count++;
2497 				MOEA64_PTE_CLEAR(mmu, pt, &pvo->pvo_pte.lpte,
2498 				    pvo->pvo_vpn, ptebit);
2499 			}
2500 		}
2501 		pvo->pvo_pte.lpte.pte_lo &= ~ptebit;
2502 		PMAP_UNLOCK(pvo->pvo_pmap);
2503 	}
2504 
2505 	UNLOCK_TABLE_RD();
2506 	return (count);
2507 }
2508 
2509 boolean_t
2510 moea64_dev_direct_mapped(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2511 {
2512 	struct pvo_entry *pvo, key;
2513 	vm_offset_t ppa;
2514 	int error = 0;
2515 
2516 	PMAP_LOCK(kernel_pmap);
2517 	key.pvo_vaddr = ppa = pa & ~ADDR_POFF;
2518 	for (pvo = RB_FIND(pvo_tree, &kernel_pmap->pmap_pvo, &key);
2519 	    ppa < pa + size; ppa += PAGE_SIZE,
2520 	    pvo = RB_NEXT(pvo_tree, &kernel_pmap->pmap_pvo, pvo)) {
2521 		if (pvo == NULL ||
2522 		    (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) != ppa) {
2523 			error = EFAULT;
2524 			break;
2525 		}
2526 	}
2527 	PMAP_UNLOCK(kernel_pmap);
2528 
2529 	return (error);
2530 }
2531 
2532 /*
2533  * Map a set of physical memory pages into the kernel virtual
2534  * address space. Return a pointer to where it is mapped. This
2535  * routine is intended to be used for mapping device memory,
2536  * NOT real memory.
2537  */
2538 void *
2539 moea64_mapdev_attr(mmu_t mmu, vm_offset_t pa, vm_size_t size, vm_memattr_t ma)
2540 {
2541 	vm_offset_t va, tmpva, ppa, offset;
2542 
2543 	ppa = trunc_page(pa);
2544 	offset = pa & PAGE_MASK;
2545 	size = roundup2(offset + size, PAGE_SIZE);
2546 
2547 	va = kmem_alloc_nofault(kernel_map, size);
2548 
2549 	if (!va)
2550 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
2551 
2552 	for (tmpva = va; size > 0;) {
2553 		moea64_kenter_attr(mmu, tmpva, ppa, ma);
2554 		size -= PAGE_SIZE;
2555 		tmpva += PAGE_SIZE;
2556 		ppa += PAGE_SIZE;
2557 	}
2558 
2559 	return ((void *)(va + offset));
2560 }
2561 
2562 void *
2563 moea64_mapdev(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2564 {
2565 
2566 	return moea64_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT);
2567 }
2568 
2569 void
2570 moea64_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2571 {
2572 	vm_offset_t base, offset;
2573 
2574 	base = trunc_page(va);
2575 	offset = va & PAGE_MASK;
2576 	size = roundup2(offset + size, PAGE_SIZE);
2577 
2578 	kmem_free(kernel_map, base, size);
2579 }
2580 
2581 void
2582 moea64_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2583 {
2584 	struct pvo_entry *pvo;
2585 	vm_offset_t lim;
2586 	vm_paddr_t pa;
2587 	vm_size_t len;
2588 
2589 	PMAP_LOCK(pm);
2590 	while (sz > 0) {
2591 		lim = round_page(va);
2592 		len = MIN(lim - va, sz);
2593 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
2594 		if (pvo != NULL && !(pvo->pvo_pte.lpte.pte_lo & LPTE_I)) {
2595 			pa = (pvo->pvo_pte.lpte.pte_lo & LPTE_RPGN) |
2596 			    (va & ADDR_POFF);
2597 			moea64_syncicache(mmu, pm, va, pa, len);
2598 		}
2599 		va += len;
2600 		sz -= len;
2601 	}
2602 	PMAP_UNLOCK(pm);
2603 }
2604