xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision 036d2e814bf0f5d88ffb4b24c159320894541757)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2008-2015 Nathan Whitehorn
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
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  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 /*
33  * Manages physical address maps.
34  *
35  * Since the information managed by this module is also stored by the
36  * logical address mapping module, this module may throw away valid virtual
37  * to physical mappings at almost any time.  However, invalidations of
38  * mappings must be done as requested.
39  *
40  * In order to cope with hardware architectures which make virtual to
41  * physical map invalidates expensive, this module may delay invalidate
42  * reduced protection operations until such time as they are actually
43  * necessary.  This module is given full information as to which processors
44  * are currently using which maps, and to when physical maps must be made
45  * correct.
46  */
47 
48 #include "opt_kstack_pages.h"
49 
50 #include <sys/param.h>
51 #include <sys/kernel.h>
52 #include <sys/conf.h>
53 #include <sys/queue.h>
54 #include <sys/cpuset.h>
55 #include <sys/kerneldump.h>
56 #include <sys/ktr.h>
57 #include <sys/lock.h>
58 #include <sys/msgbuf.h>
59 #include <sys/malloc.h>
60 #include <sys/mutex.h>
61 #include <sys/proc.h>
62 #include <sys/rwlock.h>
63 #include <sys/sched.h>
64 #include <sys/sysctl.h>
65 #include <sys/systm.h>
66 #include <sys/vmmeter.h>
67 #include <sys/smp.h>
68 
69 #include <sys/kdb.h>
70 
71 #include <dev/ofw/openfirm.h>
72 
73 #include <vm/vm.h>
74 #include <vm/vm_param.h>
75 #include <vm/vm_kern.h>
76 #include <vm/vm_page.h>
77 #include <vm/vm_phys.h>
78 #include <vm/vm_map.h>
79 #include <vm/vm_object.h>
80 #include <vm/vm_extern.h>
81 #include <vm/vm_pageout.h>
82 #include <vm/uma.h>
83 
84 #include <machine/_inttypes.h>
85 #include <machine/cpu.h>
86 #include <machine/platform.h>
87 #include <machine/frame.h>
88 #include <machine/md_var.h>
89 #include <machine/psl.h>
90 #include <machine/bat.h>
91 #include <machine/hid.h>
92 #include <machine/pte.h>
93 #include <machine/sr.h>
94 #include <machine/trap.h>
95 #include <machine/mmuvar.h>
96 
97 #include "mmu_oea64.h"
98 #include "mmu_if.h"
99 #include "moea64_if.h"
100 
101 void moea64_release_vsid(uint64_t vsid);
102 uintptr_t moea64_get_unique_vsid(void);
103 
104 #define DISABLE_TRANS(msr)	msr = mfmsr(); mtmsr(msr & ~PSL_DR)
105 #define ENABLE_TRANS(msr)	mtmsr(msr)
106 
107 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
108 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
109 #define	VSID_HASH_MASK		0x0000007fffffffffULL
110 
111 /*
112  * Locking semantics:
113  *
114  * There are two locks of interest: the page locks and the pmap locks, which
115  * protect their individual PVO lists and are locked in that order. The contents
116  * of all PVO entries are protected by the locks of their respective pmaps.
117  * The pmap of any PVO is guaranteed not to change so long as the PVO is linked
118  * into any list.
119  *
120  */
121 
122 #define PV_LOCK_PER_DOM	PA_LOCK_COUNT*3
123 #define PV_LOCK_COUNT	PV_LOCK_PER_DOM*MAXMEMDOM
124 static struct mtx_padalign pv_lock[PV_LOCK_COUNT];
125 
126 /*
127  * Cheap NUMA-izing of the pv locks, to reduce contention across domains.
128  * NUMA domains on POWER9 appear to be indexed as sparse memory spaces, with the
129  * index at (N << 45).
130  */
131 #ifdef __powerpc64__
132 #define PV_LOCK_IDX(pa)	(pa_index(pa) % PV_LOCK_PER_DOM + \
133 			(((pa) >> 45) % MAXMEMDOM) * PV_LOCK_PER_DOM)
134 #else
135 #define PV_LOCK_IDX(pa)	(pa_index(pa) % PV_LOCK_COUNT)
136 #endif
137 #define PV_LOCKPTR(pa)	((struct mtx *)(&pv_lock[PV_LOCK_IDX(pa)]))
138 #define PV_LOCK(pa)		mtx_lock(PV_LOCKPTR(pa))
139 #define PV_UNLOCK(pa)		mtx_unlock(PV_LOCKPTR(pa))
140 #define PV_LOCKASSERT(pa) 	mtx_assert(PV_LOCKPTR(pa), MA_OWNED)
141 #define PV_PAGE_LOCK(m)		PV_LOCK(VM_PAGE_TO_PHYS(m))
142 #define PV_PAGE_UNLOCK(m)	PV_UNLOCK(VM_PAGE_TO_PHYS(m))
143 #define PV_PAGE_LOCKASSERT(m)	PV_LOCKASSERT(VM_PAGE_TO_PHYS(m))
144 
145 struct ofw_map {
146 	cell_t	om_va;
147 	cell_t	om_len;
148 	uint64_t om_pa;
149 	cell_t	om_mode;
150 };
151 
152 extern unsigned char _etext[];
153 extern unsigned char _end[];
154 
155 extern void *slbtrap, *slbtrapend;
156 
157 /*
158  * Map of physical memory regions.
159  */
160 static struct	mem_region *regions;
161 static struct	mem_region *pregions;
162 static struct	numa_mem_region *numa_pregions;
163 static u_int	phys_avail_count;
164 static int	regions_sz, pregions_sz, numapregions_sz;
165 
166 extern void bs_remap_earlyboot(void);
167 
168 /*
169  * Lock for the SLB tables.
170  */
171 struct mtx	moea64_slb_mutex;
172 
173 /*
174  * PTEG data.
175  */
176 u_long		moea64_pteg_count;
177 u_long		moea64_pteg_mask;
178 
179 /*
180  * PVO data.
181  */
182 
183 uma_zone_t	moea64_pvo_zone; /* zone for pvo entries */
184 
185 static struct	pvo_entry *moea64_bpvo_pool;
186 static int	moea64_bpvo_pool_index = 0;
187 static int	moea64_bpvo_pool_size = 327680;
188 TUNABLE_INT("machdep.moea64_bpvo_pool_size", &moea64_bpvo_pool_size);
189 SYSCTL_INT(_machdep, OID_AUTO, moea64_allocated_bpvo_entries, CTLFLAG_RD,
190     &moea64_bpvo_pool_index, 0, "");
191 
192 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
193 #ifdef __powerpc64__
194 #define	NVSIDS		(NPMAPS * 16)
195 #define VSID_HASHMASK	0xffffffffUL
196 #else
197 #define NVSIDS		NPMAPS
198 #define VSID_HASHMASK	0xfffffUL
199 #endif
200 static u_int	moea64_vsid_bitmap[NVSIDS / VSID_NBPW];
201 
202 static boolean_t moea64_initialized = FALSE;
203 
204 #ifdef MOEA64_STATS
205 /*
206  * Statistics.
207  */
208 u_int	moea64_pte_valid = 0;
209 u_int	moea64_pte_overflow = 0;
210 u_int	moea64_pvo_entries = 0;
211 u_int	moea64_pvo_enter_calls = 0;
212 u_int	moea64_pvo_remove_calls = 0;
213 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_valid, CTLFLAG_RD,
214     &moea64_pte_valid, 0, "");
215 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_overflow, CTLFLAG_RD,
216     &moea64_pte_overflow, 0, "");
217 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_entries, CTLFLAG_RD,
218     &moea64_pvo_entries, 0, "");
219 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_enter_calls, CTLFLAG_RD,
220     &moea64_pvo_enter_calls, 0, "");
221 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_remove_calls, CTLFLAG_RD,
222     &moea64_pvo_remove_calls, 0, "");
223 #endif
224 
225 vm_offset_t	moea64_scratchpage_va[2];
226 struct pvo_entry *moea64_scratchpage_pvo[2];
227 struct	mtx	moea64_scratchpage_mtx;
228 
229 uint64_t 	moea64_large_page_mask = 0;
230 uint64_t	moea64_large_page_size = 0;
231 int		moea64_large_page_shift = 0;
232 
233 /*
234  * PVO calls.
235  */
236 static int	moea64_pvo_enter(mmu_t mmu, struct pvo_entry *pvo,
237 		    struct pvo_head *pvo_head, struct pvo_entry **oldpvo);
238 static void	moea64_pvo_remove_from_pmap(mmu_t mmu, struct pvo_entry *pvo);
239 static void	moea64_pvo_remove_from_page(mmu_t mmu, struct pvo_entry *pvo);
240 static void	moea64_pvo_remove_from_page_locked(mmu_t mmu,
241 		    struct pvo_entry *pvo, vm_page_t m);
242 static struct	pvo_entry *moea64_pvo_find_va(pmap_t, vm_offset_t);
243 
244 /*
245  * Utility routines.
246  */
247 static boolean_t	moea64_query_bit(mmu_t, vm_page_t, uint64_t);
248 static u_int		moea64_clear_bit(mmu_t, vm_page_t, uint64_t);
249 static void		moea64_kremove(mmu_t, vm_offset_t);
250 static void		moea64_syncicache(mmu_t, pmap_t pmap, vm_offset_t va,
251 			    vm_paddr_t pa, vm_size_t sz);
252 static void		moea64_pmap_init_qpages(void);
253 
254 /*
255  * Kernel MMU interface
256  */
257 void moea64_clear_modify(mmu_t, vm_page_t);
258 void moea64_copy_page(mmu_t, vm_page_t, vm_page_t);
259 void moea64_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
260     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
261 int moea64_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t,
262     u_int flags, int8_t psind);
263 void moea64_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
264     vm_prot_t);
265 void moea64_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
266 vm_paddr_t moea64_extract(mmu_t, pmap_t, vm_offset_t);
267 vm_page_t moea64_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
268 void moea64_init(mmu_t);
269 boolean_t moea64_is_modified(mmu_t, vm_page_t);
270 boolean_t moea64_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
271 boolean_t moea64_is_referenced(mmu_t, vm_page_t);
272 int moea64_ts_referenced(mmu_t, vm_page_t);
273 vm_offset_t moea64_map(mmu_t, vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
274 boolean_t moea64_page_exists_quick(mmu_t, pmap_t, vm_page_t);
275 void moea64_page_init(mmu_t, vm_page_t);
276 int moea64_page_wired_mappings(mmu_t, vm_page_t);
277 void moea64_pinit(mmu_t, pmap_t);
278 void moea64_pinit0(mmu_t, pmap_t);
279 void moea64_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
280 void moea64_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
281 void moea64_qremove(mmu_t, vm_offset_t, int);
282 void moea64_release(mmu_t, pmap_t);
283 void moea64_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
284 void moea64_remove_pages(mmu_t, pmap_t);
285 void moea64_remove_all(mmu_t, vm_page_t);
286 void moea64_remove_write(mmu_t, vm_page_t);
287 void moea64_unwire(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
288 void moea64_zero_page(mmu_t, vm_page_t);
289 void moea64_zero_page_area(mmu_t, vm_page_t, int, int);
290 void moea64_activate(mmu_t, struct thread *);
291 void moea64_deactivate(mmu_t, struct thread *);
292 void *moea64_mapdev(mmu_t, vm_paddr_t, vm_size_t);
293 void *moea64_mapdev_attr(mmu_t, vm_paddr_t, vm_size_t, vm_memattr_t);
294 void moea64_unmapdev(mmu_t, vm_offset_t, vm_size_t);
295 vm_paddr_t moea64_kextract(mmu_t, vm_offset_t);
296 void moea64_page_set_memattr(mmu_t, vm_page_t m, vm_memattr_t ma);
297 void moea64_kenter_attr(mmu_t, vm_offset_t, vm_paddr_t, vm_memattr_t ma);
298 void moea64_kenter(mmu_t, vm_offset_t, vm_paddr_t);
299 boolean_t moea64_dev_direct_mapped(mmu_t, vm_paddr_t, vm_size_t);
300 static void moea64_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
301 void moea64_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz,
302     void **va);
303 void moea64_scan_init(mmu_t mmu);
304 vm_offset_t moea64_quick_enter_page(mmu_t mmu, vm_page_t m);
305 void moea64_quick_remove_page(mmu_t mmu, vm_offset_t addr);
306 static int moea64_map_user_ptr(mmu_t mmu, pmap_t pm,
307     volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen);
308 static int moea64_decode_kernel_ptr(mmu_t mmu, vm_offset_t addr,
309     int *is_user, vm_offset_t *decoded_addr);
310 
311 
312 static mmu_method_t moea64_methods[] = {
313 	MMUMETHOD(mmu_clear_modify,	moea64_clear_modify),
314 	MMUMETHOD(mmu_copy_page,	moea64_copy_page),
315 	MMUMETHOD(mmu_copy_pages,	moea64_copy_pages),
316 	MMUMETHOD(mmu_enter,		moea64_enter),
317 	MMUMETHOD(mmu_enter_object,	moea64_enter_object),
318 	MMUMETHOD(mmu_enter_quick,	moea64_enter_quick),
319 	MMUMETHOD(mmu_extract,		moea64_extract),
320 	MMUMETHOD(mmu_extract_and_hold,	moea64_extract_and_hold),
321 	MMUMETHOD(mmu_init,		moea64_init),
322 	MMUMETHOD(mmu_is_modified,	moea64_is_modified),
323 	MMUMETHOD(mmu_is_prefaultable,	moea64_is_prefaultable),
324 	MMUMETHOD(mmu_is_referenced,	moea64_is_referenced),
325 	MMUMETHOD(mmu_ts_referenced,	moea64_ts_referenced),
326 	MMUMETHOD(mmu_map,     		moea64_map),
327 	MMUMETHOD(mmu_page_exists_quick,moea64_page_exists_quick),
328 	MMUMETHOD(mmu_page_init,	moea64_page_init),
329 	MMUMETHOD(mmu_page_wired_mappings,moea64_page_wired_mappings),
330 	MMUMETHOD(mmu_pinit,		moea64_pinit),
331 	MMUMETHOD(mmu_pinit0,		moea64_pinit0),
332 	MMUMETHOD(mmu_protect,		moea64_protect),
333 	MMUMETHOD(mmu_qenter,		moea64_qenter),
334 	MMUMETHOD(mmu_qremove,		moea64_qremove),
335 	MMUMETHOD(mmu_release,		moea64_release),
336 	MMUMETHOD(mmu_remove,		moea64_remove),
337 	MMUMETHOD(mmu_remove_pages,	moea64_remove_pages),
338 	MMUMETHOD(mmu_remove_all,      	moea64_remove_all),
339 	MMUMETHOD(mmu_remove_write,	moea64_remove_write),
340 	MMUMETHOD(mmu_sync_icache,	moea64_sync_icache),
341 	MMUMETHOD(mmu_unwire,		moea64_unwire),
342 	MMUMETHOD(mmu_zero_page,       	moea64_zero_page),
343 	MMUMETHOD(mmu_zero_page_area,	moea64_zero_page_area),
344 	MMUMETHOD(mmu_activate,		moea64_activate),
345 	MMUMETHOD(mmu_deactivate,      	moea64_deactivate),
346 	MMUMETHOD(mmu_page_set_memattr,	moea64_page_set_memattr),
347 	MMUMETHOD(mmu_quick_enter_page, moea64_quick_enter_page),
348 	MMUMETHOD(mmu_quick_remove_page, moea64_quick_remove_page),
349 
350 	/* Internal interfaces */
351 	MMUMETHOD(mmu_mapdev,		moea64_mapdev),
352 	MMUMETHOD(mmu_mapdev_attr,	moea64_mapdev_attr),
353 	MMUMETHOD(mmu_unmapdev,		moea64_unmapdev),
354 	MMUMETHOD(mmu_kextract,		moea64_kextract),
355 	MMUMETHOD(mmu_kenter,		moea64_kenter),
356 	MMUMETHOD(mmu_kenter_attr,	moea64_kenter_attr),
357 	MMUMETHOD(mmu_dev_direct_mapped,moea64_dev_direct_mapped),
358 	MMUMETHOD(mmu_scan_init,	moea64_scan_init),
359 	MMUMETHOD(mmu_dumpsys_map,	moea64_dumpsys_map),
360 	MMUMETHOD(mmu_map_user_ptr,	moea64_map_user_ptr),
361 	MMUMETHOD(mmu_decode_kernel_ptr, moea64_decode_kernel_ptr),
362 
363 	{ 0, 0 }
364 };
365 
366 MMU_DEF(oea64_mmu, "mmu_oea64_base", moea64_methods, 0);
367 
368 static struct pvo_head *
369 vm_page_to_pvoh(vm_page_t m)
370 {
371 
372 	mtx_assert(PV_LOCKPTR(VM_PAGE_TO_PHYS(m)), MA_OWNED);
373 	return (&m->md.mdpg_pvoh);
374 }
375 
376 static struct pvo_entry *
377 alloc_pvo_entry(int bootstrap)
378 {
379 	struct pvo_entry *pvo;
380 
381 	if (!moea64_initialized || bootstrap) {
382 		if (moea64_bpvo_pool_index >= moea64_bpvo_pool_size) {
383 			panic("moea64_enter: bpvo pool exhausted, %d, %d, %zd",
384 			      moea64_bpvo_pool_index, moea64_bpvo_pool_size,
385 			      moea64_bpvo_pool_size * sizeof(struct pvo_entry));
386 		}
387 		pvo = &moea64_bpvo_pool[
388 		    atomic_fetchadd_int(&moea64_bpvo_pool_index, 1)];
389 		bzero(pvo, sizeof(*pvo));
390 		pvo->pvo_vaddr = PVO_BOOTSTRAP;
391 	} else
392 		pvo = uma_zalloc(moea64_pvo_zone, M_NOWAIT | M_ZERO);
393 
394 	return (pvo);
395 }
396 
397 
398 static void
399 init_pvo_entry(struct pvo_entry *pvo, pmap_t pmap, vm_offset_t va)
400 {
401 	uint64_t vsid;
402 	uint64_t hash;
403 	int shift;
404 
405 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
406 
407 	pvo->pvo_pmap = pmap;
408 	va &= ~ADDR_POFF;
409 	pvo->pvo_vaddr |= va;
410 	vsid = va_to_vsid(pmap, va);
411 	pvo->pvo_vpn = (uint64_t)((va & ADDR_PIDX) >> ADDR_PIDX_SHFT)
412 	    | (vsid << 16);
413 
414 	shift = (pvo->pvo_vaddr & PVO_LARGE) ? moea64_large_page_shift :
415 	    ADDR_PIDX_SHFT;
416 	hash = (vsid & VSID_HASH_MASK) ^ (((uint64_t)va & ADDR_PIDX) >> shift);
417 	pvo->pvo_pte.slot = (hash & moea64_pteg_mask) << 3;
418 }
419 
420 static void
421 free_pvo_entry(struct pvo_entry *pvo)
422 {
423 
424 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
425 		uma_zfree(moea64_pvo_zone, pvo);
426 }
427 
428 void
429 moea64_pte_from_pvo(const struct pvo_entry *pvo, struct lpte *lpte)
430 {
431 
432 	lpte->pte_hi = (pvo->pvo_vpn >> (ADDR_API_SHFT64 - ADDR_PIDX_SHFT)) &
433 	    LPTE_AVPN_MASK;
434 	lpte->pte_hi |= LPTE_VALID;
435 
436 	if (pvo->pvo_vaddr & PVO_LARGE)
437 		lpte->pte_hi |= LPTE_BIG;
438 	if (pvo->pvo_vaddr & PVO_WIRED)
439 		lpte->pte_hi |= LPTE_WIRED;
440 	if (pvo->pvo_vaddr & PVO_HID)
441 		lpte->pte_hi |= LPTE_HID;
442 
443 	lpte->pte_lo = pvo->pvo_pte.pa; /* Includes WIMG bits */
444 	if (pvo->pvo_pte.prot & VM_PROT_WRITE)
445 		lpte->pte_lo |= LPTE_BW;
446 	else
447 		lpte->pte_lo |= LPTE_BR;
448 
449 	if (!(pvo->pvo_pte.prot & VM_PROT_EXECUTE))
450 		lpte->pte_lo |= LPTE_NOEXEC;
451 }
452 
453 static __inline uint64_t
454 moea64_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
455 {
456 	uint64_t pte_lo;
457 	int i;
458 
459 	if (ma != VM_MEMATTR_DEFAULT) {
460 		switch (ma) {
461 		case VM_MEMATTR_UNCACHEABLE:
462 			return (LPTE_I | LPTE_G);
463 		case VM_MEMATTR_CACHEABLE:
464 			return (LPTE_M);
465 		case VM_MEMATTR_WRITE_COMBINING:
466 		case VM_MEMATTR_WRITE_BACK:
467 		case VM_MEMATTR_PREFETCHABLE:
468 			return (LPTE_I);
469 		case VM_MEMATTR_WRITE_THROUGH:
470 			return (LPTE_W | LPTE_M);
471 		}
472 	}
473 
474 	/*
475 	 * Assume the page is cache inhibited and access is guarded unless
476 	 * it's in our available memory array.
477 	 */
478 	pte_lo = LPTE_I | LPTE_G;
479 	for (i = 0; i < pregions_sz; i++) {
480 		if ((pa >= pregions[i].mr_start) &&
481 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
482 			pte_lo &= ~(LPTE_I | LPTE_G);
483 			pte_lo |= LPTE_M;
484 			break;
485 		}
486 	}
487 
488 	return pte_lo;
489 }
490 
491 /*
492  * Quick sort callout for comparing memory regions.
493  */
494 static int	om_cmp(const void *a, const void *b);
495 
496 static int
497 om_cmp(const void *a, const void *b)
498 {
499 	const struct	ofw_map *mapa;
500 	const struct	ofw_map *mapb;
501 
502 	mapa = a;
503 	mapb = b;
504 	if (mapa->om_pa < mapb->om_pa)
505 		return (-1);
506 	else if (mapa->om_pa > mapb->om_pa)
507 		return (1);
508 	else
509 		return (0);
510 }
511 
512 static void
513 moea64_add_ofw_mappings(mmu_t mmup, phandle_t mmu, size_t sz)
514 {
515 	struct ofw_map	translations[sz/(4*sizeof(cell_t))]; /*>= 4 cells per */
516 	pcell_t		acells, trans_cells[sz/sizeof(cell_t)];
517 	struct pvo_entry *pvo;
518 	register_t	msr;
519 	vm_offset_t	off;
520 	vm_paddr_t	pa_base;
521 	int		i, j;
522 
523 	bzero(translations, sz);
524 	OF_getencprop(OF_finddevice("/"), "#address-cells", &acells,
525 	    sizeof(acells));
526 	if (OF_getencprop(mmu, "translations", trans_cells, sz) == -1)
527 		panic("moea64_bootstrap: can't get ofw translations");
528 
529 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
530 	sz /= sizeof(cell_t);
531 	for (i = 0, j = 0; i < sz; j++) {
532 		translations[j].om_va = trans_cells[i++];
533 		translations[j].om_len = trans_cells[i++];
534 		translations[j].om_pa = trans_cells[i++];
535 		if (acells == 2) {
536 			translations[j].om_pa <<= 32;
537 			translations[j].om_pa |= trans_cells[i++];
538 		}
539 		translations[j].om_mode = trans_cells[i++];
540 	}
541 	KASSERT(i == sz, ("Translations map has incorrect cell count (%d/%zd)",
542 	    i, sz));
543 
544 	sz = j;
545 	qsort(translations, sz, sizeof (*translations), om_cmp);
546 
547 	for (i = 0; i < sz; i++) {
548 		pa_base = translations[i].om_pa;
549 	      #ifndef __powerpc64__
550 		if ((translations[i].om_pa >> 32) != 0)
551 			panic("OFW translations above 32-bit boundary!");
552 	      #endif
553 
554 		if (pa_base % PAGE_SIZE)
555 			panic("OFW translation not page-aligned (phys)!");
556 		if (translations[i].om_va % PAGE_SIZE)
557 			panic("OFW translation not page-aligned (virt)!");
558 
559 		CTR3(KTR_PMAP, "translation: pa=%#zx va=%#x len=%#x",
560 		    pa_base, translations[i].om_va, translations[i].om_len);
561 
562 		/* Now enter the pages for this mapping */
563 
564 		DISABLE_TRANS(msr);
565 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
566 			/* If this address is direct-mapped, skip remapping */
567 			if (hw_direct_map &&
568 			    translations[i].om_va == PHYS_TO_DMAP(pa_base) &&
569 			    moea64_calc_wimg(pa_base + off, VM_MEMATTR_DEFAULT)
570  			    == LPTE_M)
571 				continue;
572 
573 			PMAP_LOCK(kernel_pmap);
574 			pvo = moea64_pvo_find_va(kernel_pmap,
575 			    translations[i].om_va + off);
576 			PMAP_UNLOCK(kernel_pmap);
577 			if (pvo != NULL)
578 				continue;
579 
580 			moea64_kenter(mmup, translations[i].om_va + off,
581 			    pa_base + off);
582 		}
583 		ENABLE_TRANS(msr);
584 	}
585 }
586 
587 #ifdef __powerpc64__
588 static void
589 moea64_probe_large_page(void)
590 {
591 	uint16_t pvr = mfpvr() >> 16;
592 
593 	switch (pvr) {
594 	case IBM970:
595 	case IBM970FX:
596 	case IBM970MP:
597 		powerpc_sync(); isync();
598 		mtspr(SPR_HID4, mfspr(SPR_HID4) & ~HID4_970_DISABLE_LG_PG);
599 		powerpc_sync(); isync();
600 
601 		/* FALLTHROUGH */
602 	default:
603 		if (moea64_large_page_size == 0) {
604 			moea64_large_page_size = 0x1000000; /* 16 MB */
605 			moea64_large_page_shift = 24;
606 		}
607 	}
608 
609 	moea64_large_page_mask = moea64_large_page_size - 1;
610 }
611 
612 static void
613 moea64_bootstrap_slb_prefault(vm_offset_t va, int large)
614 {
615 	struct slb *cache;
616 	struct slb entry;
617 	uint64_t esid, slbe;
618 	uint64_t i;
619 
620 	cache = PCPU_GET(aim.slb);
621 	esid = va >> ADDR_SR_SHFT;
622 	slbe = (esid << SLBE_ESID_SHIFT) | SLBE_VALID;
623 
624 	for (i = 0; i < 64; i++) {
625 		if (cache[i].slbe == (slbe | i))
626 			return;
627 	}
628 
629 	entry.slbe = slbe;
630 	entry.slbv = KERNEL_VSID(esid) << SLBV_VSID_SHIFT;
631 	if (large)
632 		entry.slbv |= SLBV_L;
633 
634 	slb_insert_kernel(entry.slbe, entry.slbv);
635 }
636 #endif
637 
638 static void
639 moea64_setup_direct_map(mmu_t mmup, vm_offset_t kernelstart,
640     vm_offset_t kernelend)
641 {
642 	struct pvo_entry *pvo;
643 	register_t msr;
644 	vm_paddr_t pa, pkernelstart, pkernelend;
645 	vm_offset_t size, off;
646 	uint64_t pte_lo;
647 	int i;
648 
649 	if (moea64_large_page_size == 0)
650 		hw_direct_map = 0;
651 
652 	DISABLE_TRANS(msr);
653 	if (hw_direct_map) {
654 		PMAP_LOCK(kernel_pmap);
655 		for (i = 0; i < pregions_sz; i++) {
656 		  for (pa = pregions[i].mr_start; pa < pregions[i].mr_start +
657 		     pregions[i].mr_size; pa += moea64_large_page_size) {
658 			pte_lo = LPTE_M;
659 
660 			pvo = alloc_pvo_entry(1 /* bootstrap */);
661 			pvo->pvo_vaddr |= PVO_WIRED | PVO_LARGE;
662 			init_pvo_entry(pvo, kernel_pmap, PHYS_TO_DMAP(pa));
663 
664 			/*
665 			 * Set memory access as guarded if prefetch within
666 			 * the page could exit the available physmem area.
667 			 */
668 			if (pa & moea64_large_page_mask) {
669 				pa &= moea64_large_page_mask;
670 				pte_lo |= LPTE_G;
671 			}
672 			if (pa + moea64_large_page_size >
673 			    pregions[i].mr_start + pregions[i].mr_size)
674 				pte_lo |= LPTE_G;
675 
676 			pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE |
677 			    VM_PROT_EXECUTE;
678 			pvo->pvo_pte.pa = pa | pte_lo;
679 			moea64_pvo_enter(mmup, pvo, NULL, NULL);
680 		  }
681 		}
682 		PMAP_UNLOCK(kernel_pmap);
683 	}
684 
685 	/*
686 	 * Make sure the kernel and BPVO pool stay mapped on systems either
687 	 * without a direct map or on which the kernel is not already executing
688 	 * out of the direct-mapped region.
689 	 */
690 	if (kernelstart < DMAP_BASE_ADDRESS) {
691 		/*
692 		 * For pre-dmap execution, we need to use identity mapping
693 		 * because we will be operating with the mmu on but in the
694 		 * wrong address configuration until we __restartkernel().
695 		 */
696 		for (pa = kernelstart & ~PAGE_MASK; pa < kernelend;
697 		    pa += PAGE_SIZE)
698 			moea64_kenter(mmup, pa, pa);
699 	} else if (!hw_direct_map) {
700 		pkernelstart = kernelstart & ~DMAP_BASE_ADDRESS;
701 		pkernelend = kernelend & ~DMAP_BASE_ADDRESS;
702 		for (pa = pkernelstart & ~PAGE_MASK; pa < pkernelend;
703 		    pa += PAGE_SIZE)
704 			moea64_kenter(mmup, pa | DMAP_BASE_ADDRESS, pa);
705 	}
706 
707 	if (!hw_direct_map) {
708 		size = moea64_bpvo_pool_size*sizeof(struct pvo_entry);
709 		off = (vm_offset_t)(moea64_bpvo_pool);
710 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
711 			moea64_kenter(mmup, pa, pa);
712 
713 		/* Map exception vectors */
714 		for (pa = EXC_RSVD; pa < EXC_LAST; pa += PAGE_SIZE)
715 			moea64_kenter(mmup, pa | DMAP_BASE_ADDRESS, pa);
716 	}
717 	ENABLE_TRANS(msr);
718 
719 	/*
720 	 * Allow user to override unmapped_buf_allowed for testing.
721 	 * XXXKIB Only direct map implementation was tested.
722 	 */
723 	if (!TUNABLE_INT_FETCH("vfs.unmapped_buf_allowed",
724 	    &unmapped_buf_allowed))
725 		unmapped_buf_allowed = hw_direct_map;
726 }
727 
728 /* Quick sort callout for comparing physical addresses. */
729 static int
730 pa_cmp(const void *a, const void *b)
731 {
732 	const vm_paddr_t *pa = a, *pb = b;
733 
734 	if (*pa < *pb)
735 		return (-1);
736 	else if (*pa > *pb)
737 		return (1);
738 	else
739 		return (0);
740 }
741 
742 void
743 moea64_early_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
744 {
745 	int		i, j;
746 	vm_size_t	physsz, hwphyssz;
747 	vm_paddr_t	kernelphysstart, kernelphysend;
748 	int		rm_pavail;
749 
750 #ifndef __powerpc64__
751 	/* We don't have a direct map since there is no BAT */
752 	hw_direct_map = 0;
753 
754 	/* Make sure battable is zero, since we have no BAT */
755 	for (i = 0; i < 16; i++) {
756 		battable[i].batu = 0;
757 		battable[i].batl = 0;
758 	}
759 #else
760 	moea64_probe_large_page();
761 
762 	/* Use a direct map if we have large page support */
763 	if (moea64_large_page_size > 0)
764 		hw_direct_map = 1;
765 	else
766 		hw_direct_map = 0;
767 
768 	/* Install trap handlers for SLBs */
769 	bcopy(&slbtrap, (void *)EXC_DSE,(size_t)&slbtrapend - (size_t)&slbtrap);
770 	bcopy(&slbtrap, (void *)EXC_ISE,(size_t)&slbtrapend - (size_t)&slbtrap);
771 	__syncicache((void *)EXC_DSE, 0x80);
772 	__syncicache((void *)EXC_ISE, 0x80);
773 #endif
774 
775 	kernelphysstart = kernelstart & ~DMAP_BASE_ADDRESS;
776 	kernelphysend = kernelend & ~DMAP_BASE_ADDRESS;
777 
778 	/* Get physical memory regions from firmware */
779 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
780 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
781 
782 	if (PHYS_AVAIL_ENTRIES < regions_sz)
783 		panic("moea64_bootstrap: phys_avail too small");
784 
785 	phys_avail_count = 0;
786 	physsz = 0;
787 	hwphyssz = 0;
788 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
789 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
790 		CTR3(KTR_PMAP, "region: %#zx - %#zx (%#zx)",
791 		    regions[i].mr_start, regions[i].mr_start +
792 		    regions[i].mr_size, regions[i].mr_size);
793 		if (hwphyssz != 0 &&
794 		    (physsz + regions[i].mr_size) >= hwphyssz) {
795 			if (physsz < hwphyssz) {
796 				phys_avail[j] = regions[i].mr_start;
797 				phys_avail[j + 1] = regions[i].mr_start +
798 				    hwphyssz - physsz;
799 				physsz = hwphyssz;
800 				phys_avail_count++;
801 			}
802 			break;
803 		}
804 		phys_avail[j] = regions[i].mr_start;
805 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
806 		phys_avail_count++;
807 		physsz += regions[i].mr_size;
808 	}
809 
810 	/* Check for overlap with the kernel and exception vectors */
811 	rm_pavail = 0;
812 	for (j = 0; j < 2*phys_avail_count; j+=2) {
813 		if (phys_avail[j] < EXC_LAST)
814 			phys_avail[j] += EXC_LAST;
815 
816 		if (phys_avail[j] >= kernelphysstart &&
817 		    phys_avail[j+1] <= kernelphysend) {
818 			phys_avail[j] = phys_avail[j+1] = ~0;
819 			rm_pavail++;
820 			continue;
821 		}
822 
823 		if (kernelphysstart >= phys_avail[j] &&
824 		    kernelphysstart < phys_avail[j+1]) {
825 			if (kernelphysend < phys_avail[j+1]) {
826 				phys_avail[2*phys_avail_count] =
827 				    (kernelphysend & ~PAGE_MASK) + PAGE_SIZE;
828 				phys_avail[2*phys_avail_count + 1] =
829 				    phys_avail[j+1];
830 				phys_avail_count++;
831 			}
832 
833 			phys_avail[j+1] = kernelphysstart & ~PAGE_MASK;
834 		}
835 
836 		if (kernelphysend >= phys_avail[j] &&
837 		    kernelphysend < phys_avail[j+1]) {
838 			if (kernelphysstart > phys_avail[j]) {
839 				phys_avail[2*phys_avail_count] = phys_avail[j];
840 				phys_avail[2*phys_avail_count + 1] =
841 				    kernelphysstart & ~PAGE_MASK;
842 				phys_avail_count++;
843 			}
844 
845 			phys_avail[j] = (kernelphysend & ~PAGE_MASK) +
846 			    PAGE_SIZE;
847 		}
848 	}
849 
850 	/* Remove physical available regions marked for removal (~0) */
851 	if (rm_pavail) {
852 		qsort(phys_avail, 2*phys_avail_count, sizeof(phys_avail[0]),
853 			pa_cmp);
854 		phys_avail_count -= rm_pavail;
855 		for (i = 2*phys_avail_count;
856 		     i < 2*(phys_avail_count + rm_pavail); i+=2)
857 			phys_avail[i] = phys_avail[i+1] = 0;
858 	}
859 
860 	physmem = btoc(physsz);
861 
862 #ifdef PTEGCOUNT
863 	moea64_pteg_count = PTEGCOUNT;
864 #else
865 	moea64_pteg_count = 0x1000;
866 
867 	while (moea64_pteg_count < physmem)
868 		moea64_pteg_count <<= 1;
869 
870 	moea64_pteg_count >>= 1;
871 #endif /* PTEGCOUNT */
872 }
873 
874 void
875 moea64_mid_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
876 {
877 	int		i;
878 
879 	/*
880 	 * Set PTEG mask
881 	 */
882 	moea64_pteg_mask = moea64_pteg_count - 1;
883 
884 	/*
885 	 * Initialize SLB table lock and page locks
886 	 */
887 	mtx_init(&moea64_slb_mutex, "SLB table", NULL, MTX_DEF);
888 	for (i = 0; i < PV_LOCK_COUNT; i++)
889 		mtx_init(&pv_lock[i], "page pv", NULL, MTX_DEF);
890 
891 	/*
892 	 * Initialise the bootstrap pvo pool.
893 	 */
894 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
895 		moea64_bpvo_pool_size*sizeof(struct pvo_entry), PAGE_SIZE);
896 	moea64_bpvo_pool_index = 0;
897 
898 	/* Place at address usable through the direct map */
899 	if (hw_direct_map)
900 		moea64_bpvo_pool = (struct pvo_entry *)
901 		    PHYS_TO_DMAP((uintptr_t)moea64_bpvo_pool);
902 
903 	/*
904 	 * Make sure kernel vsid is allocated as well as VSID 0.
905 	 */
906 	#ifndef __powerpc64__
907 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NVSIDS - 1)) / VSID_NBPW]
908 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
909 	moea64_vsid_bitmap[0] |= 1;
910 	#endif
911 
912 	/*
913 	 * Initialize the kernel pmap (which is statically allocated).
914 	 */
915 	#ifdef __powerpc64__
916 	for (i = 0; i < 64; i++) {
917 		pcpup->pc_aim.slb[i].slbv = 0;
918 		pcpup->pc_aim.slb[i].slbe = 0;
919 	}
920 	#else
921 	for (i = 0; i < 16; i++)
922 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
923 	#endif
924 
925 	kernel_pmap->pmap_phys = kernel_pmap;
926 	CPU_FILL(&kernel_pmap->pm_active);
927 	RB_INIT(&kernel_pmap->pmap_pvo);
928 
929 	PMAP_LOCK_INIT(kernel_pmap);
930 
931 	/*
932 	 * Now map in all the other buffers we allocated earlier
933 	 */
934 
935 	moea64_setup_direct_map(mmup, kernelstart, kernelend);
936 }
937 
938 void
939 moea64_late_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
940 {
941 	ihandle_t	mmui;
942 	phandle_t	chosen;
943 	phandle_t	mmu;
944 	ssize_t		sz;
945 	int		i;
946 	vm_offset_t	pa, va;
947 	void		*dpcpu;
948 
949 	/*
950 	 * Set up the Open Firmware pmap and add its mappings if not in real
951 	 * mode.
952 	 */
953 
954 	chosen = OF_finddevice("/chosen");
955 	if (chosen != -1 && OF_getencprop(chosen, "mmu", &mmui, 4) != -1) {
956 		mmu = OF_instance_to_package(mmui);
957 		if (mmu == -1 ||
958 		    (sz = OF_getproplen(mmu, "translations")) == -1)
959 			sz = 0;
960 		if (sz > 6144 /* tmpstksz - 2 KB headroom */)
961 			panic("moea64_bootstrap: too many ofw translations");
962 
963 		if (sz > 0)
964 			moea64_add_ofw_mappings(mmup, mmu, sz);
965 	}
966 
967 	/*
968 	 * Calculate the last available physical address.
969 	 */
970 	Maxmem = 0;
971 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
972 		Maxmem = MAX(Maxmem, powerpc_btop(phys_avail[i + 1]));
973 
974 	/*
975 	 * Initialize MMU.
976 	 */
977 	MMU_CPU_BOOTSTRAP(mmup,0);
978 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
979 	pmap_bootstrapped++;
980 
981 	/*
982 	 * Set the start and end of kva.
983 	 */
984 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
985 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
986 
987 	/*
988 	 * Map the entire KVA range into the SLB. We must not fault there.
989 	 */
990 	#ifdef __powerpc64__
991 	for (va = virtual_avail; va < virtual_end; va += SEGMENT_LENGTH)
992 		moea64_bootstrap_slb_prefault(va, 0);
993 	#endif
994 
995 	/*
996 	 * Remap any early IO mappings (console framebuffer, etc.)
997 	 */
998 	bs_remap_earlyboot();
999 
1000 	/*
1001 	 * Figure out how far we can extend virtual_end into segment 16
1002 	 * without running into existing mappings. Segment 16 is guaranteed
1003 	 * to contain neither RAM nor devices (at least on Apple hardware),
1004 	 * but will generally contain some OFW mappings we should not
1005 	 * step on.
1006 	 */
1007 
1008 	#ifndef __powerpc64__	/* KVA is in high memory on PPC64 */
1009 	PMAP_LOCK(kernel_pmap);
1010 	while (virtual_end < VM_MAX_KERNEL_ADDRESS &&
1011 	    moea64_pvo_find_va(kernel_pmap, virtual_end+1) == NULL)
1012 		virtual_end += PAGE_SIZE;
1013 	PMAP_UNLOCK(kernel_pmap);
1014 	#endif
1015 
1016 	/*
1017 	 * Allocate a kernel stack with a guard page for thread0 and map it
1018 	 * into the kernel page map.
1019 	 */
1020 	pa = moea64_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
1021 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
1022 	virtual_avail = va + kstack_pages * PAGE_SIZE;
1023 	CTR2(KTR_PMAP, "moea64_bootstrap: kstack0 at %#x (%#x)", pa, va);
1024 	thread0.td_kstack = va;
1025 	thread0.td_kstack_pages = kstack_pages;
1026 	for (i = 0; i < kstack_pages; i++) {
1027 		moea64_kenter(mmup, va, pa);
1028 		pa += PAGE_SIZE;
1029 		va += PAGE_SIZE;
1030 	}
1031 
1032 	/*
1033 	 * Allocate virtual address space for the message buffer.
1034 	 */
1035 	pa = msgbuf_phys = moea64_bootstrap_alloc(msgbufsize, PAGE_SIZE);
1036 	msgbufp = (struct msgbuf *)virtual_avail;
1037 	va = virtual_avail;
1038 	virtual_avail += round_page(msgbufsize);
1039 	while (va < virtual_avail) {
1040 		moea64_kenter(mmup, va, pa);
1041 		pa += PAGE_SIZE;
1042 		va += PAGE_SIZE;
1043 	}
1044 
1045 	/*
1046 	 * Allocate virtual address space for the dynamic percpu area.
1047 	 */
1048 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
1049 	dpcpu = (void *)virtual_avail;
1050 	va = virtual_avail;
1051 	virtual_avail += DPCPU_SIZE;
1052 	while (va < virtual_avail) {
1053 		moea64_kenter(mmup, va, pa);
1054 		pa += PAGE_SIZE;
1055 		va += PAGE_SIZE;
1056 	}
1057 	dpcpu_init(dpcpu, curcpu);
1058 
1059 	/*
1060 	 * Allocate some things for page zeroing. We put this directly
1061 	 * in the page table and use MOEA64_PTE_REPLACE to avoid any
1062 	 * of the PVO book-keeping or other parts of the VM system
1063 	 * from even knowing that this hack exists.
1064 	 */
1065 
1066 	if (!hw_direct_map) {
1067 		mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL,
1068 		    MTX_DEF);
1069 		for (i = 0; i < 2; i++) {
1070 			moea64_scratchpage_va[i] = (virtual_end+1) - PAGE_SIZE;
1071 			virtual_end -= PAGE_SIZE;
1072 
1073 			moea64_kenter(mmup, moea64_scratchpage_va[i], 0);
1074 
1075 			PMAP_LOCK(kernel_pmap);
1076 			moea64_scratchpage_pvo[i] = moea64_pvo_find_va(
1077 			    kernel_pmap, (vm_offset_t)moea64_scratchpage_va[i]);
1078 			PMAP_UNLOCK(kernel_pmap);
1079 		}
1080 	}
1081 
1082 	numa_mem_regions(&numa_pregions, &numapregions_sz);
1083 }
1084 
1085 static void
1086 moea64_pmap_init_qpages(void)
1087 {
1088 	struct pcpu *pc;
1089 	int i;
1090 
1091 	if (hw_direct_map)
1092 		return;
1093 
1094 	CPU_FOREACH(i) {
1095 		pc = pcpu_find(i);
1096 		pc->pc_qmap_addr = kva_alloc(PAGE_SIZE);
1097 		if (pc->pc_qmap_addr == 0)
1098 			panic("pmap_init_qpages: unable to allocate KVA");
1099 		PMAP_LOCK(kernel_pmap);
1100 		pc->pc_aim.qmap_pvo =
1101 		    moea64_pvo_find_va(kernel_pmap, pc->pc_qmap_addr);
1102 		PMAP_UNLOCK(kernel_pmap);
1103 		mtx_init(&pc->pc_aim.qmap_lock, "qmap lock", NULL, MTX_DEF);
1104 	}
1105 }
1106 
1107 SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, moea64_pmap_init_qpages, NULL);
1108 
1109 /*
1110  * Activate a user pmap.  This mostly involves setting some non-CPU
1111  * state.
1112  */
1113 void
1114 moea64_activate(mmu_t mmu, struct thread *td)
1115 {
1116 	pmap_t	pm;
1117 
1118 	pm = &td->td_proc->p_vmspace->vm_pmap;
1119 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1120 
1121 	#ifdef __powerpc64__
1122 	PCPU_SET(aim.userslb, pm->pm_slb);
1123 	__asm __volatile("slbmte %0, %1; isync" ::
1124 	    "r"(td->td_pcb->pcb_cpu.aim.usr_vsid), "r"(USER_SLB_SLBE));
1125 	#else
1126 	PCPU_SET(curpmap, pm->pmap_phys);
1127 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1128 	#endif
1129 }
1130 
1131 void
1132 moea64_deactivate(mmu_t mmu, struct thread *td)
1133 {
1134 	pmap_t	pm;
1135 
1136 	__asm __volatile("isync; slbie %0" :: "r"(USER_ADDR));
1137 
1138 	pm = &td->td_proc->p_vmspace->vm_pmap;
1139 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1140 	#ifdef __powerpc64__
1141 	PCPU_SET(aim.userslb, NULL);
1142 	#else
1143 	PCPU_SET(curpmap, NULL);
1144 	#endif
1145 }
1146 
1147 void
1148 moea64_unwire(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1149 {
1150 	struct	pvo_entry key, *pvo;
1151 	vm_page_t m;
1152 	int64_t	refchg;
1153 
1154 	key.pvo_vaddr = sva;
1155 	PMAP_LOCK(pm);
1156 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1157 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1158 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1159 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1160 			panic("moea64_unwire: pvo %p is missing PVO_WIRED",
1161 			    pvo);
1162 		pvo->pvo_vaddr &= ~PVO_WIRED;
1163 		refchg = MOEA64_PTE_REPLACE(mmu, pvo, 0 /* No invalidation */);
1164 		if ((pvo->pvo_vaddr & PVO_MANAGED) &&
1165 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1166 			if (refchg < 0)
1167 				refchg = LPTE_CHG;
1168 			m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
1169 
1170 			refchg |= atomic_readandclear_32(&m->md.mdpg_attrs);
1171 			if (refchg & LPTE_CHG)
1172 				vm_page_dirty(m);
1173 			if (refchg & LPTE_REF)
1174 				vm_page_aflag_set(m, PGA_REFERENCED);
1175 		}
1176 		pm->pm_stats.wired_count--;
1177 	}
1178 	PMAP_UNLOCK(pm);
1179 }
1180 
1181 /*
1182  * This goes through and sets the physical address of our
1183  * special scratch PTE to the PA we want to zero or copy. Because
1184  * of locking issues (this can get called in pvo_enter() by
1185  * the UMA allocator), we can't use most other utility functions here
1186  */
1187 
1188 static __inline
1189 void moea64_set_scratchpage_pa(mmu_t mmup, int which, vm_paddr_t pa)
1190 {
1191 	struct pvo_entry *pvo;
1192 
1193 	KASSERT(!hw_direct_map, ("Using OEA64 scratchpage with a direct map!"));
1194 	mtx_assert(&moea64_scratchpage_mtx, MA_OWNED);
1195 
1196 	pvo = moea64_scratchpage_pvo[which];
1197 	PMAP_LOCK(pvo->pvo_pmap);
1198 	pvo->pvo_pte.pa =
1199 	    moea64_calc_wimg(pa, VM_MEMATTR_DEFAULT) | (uint64_t)pa;
1200 	MOEA64_PTE_REPLACE(mmup, pvo, MOEA64_PTE_INVALIDATE);
1201 	PMAP_UNLOCK(pvo->pvo_pmap);
1202 	isync();
1203 }
1204 
1205 void
1206 moea64_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1207 {
1208 	vm_offset_t	dst;
1209 	vm_offset_t	src;
1210 
1211 	dst = VM_PAGE_TO_PHYS(mdst);
1212 	src = VM_PAGE_TO_PHYS(msrc);
1213 
1214 	if (hw_direct_map) {
1215 		bcopy((void *)PHYS_TO_DMAP(src), (void *)PHYS_TO_DMAP(dst),
1216 		    PAGE_SIZE);
1217 	} else {
1218 		mtx_lock(&moea64_scratchpage_mtx);
1219 
1220 		moea64_set_scratchpage_pa(mmu, 0, src);
1221 		moea64_set_scratchpage_pa(mmu, 1, dst);
1222 
1223 		bcopy((void *)moea64_scratchpage_va[0],
1224 		    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1225 
1226 		mtx_unlock(&moea64_scratchpage_mtx);
1227 	}
1228 }
1229 
1230 static inline void
1231 moea64_copy_pages_dmap(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1232     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1233 {
1234 	void *a_cp, *b_cp;
1235 	vm_offset_t a_pg_offset, b_pg_offset;
1236 	int cnt;
1237 
1238 	while (xfersize > 0) {
1239 		a_pg_offset = a_offset & PAGE_MASK;
1240 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1241 		a_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1242 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT])) +
1243 		    a_pg_offset;
1244 		b_pg_offset = b_offset & PAGE_MASK;
1245 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1246 		b_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1247 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT])) +
1248 		    b_pg_offset;
1249 		bcopy(a_cp, b_cp, cnt);
1250 		a_offset += cnt;
1251 		b_offset += cnt;
1252 		xfersize -= cnt;
1253 	}
1254 }
1255 
1256 static inline void
1257 moea64_copy_pages_nodmap(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1258     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1259 {
1260 	void *a_cp, *b_cp;
1261 	vm_offset_t a_pg_offset, b_pg_offset;
1262 	int cnt;
1263 
1264 	mtx_lock(&moea64_scratchpage_mtx);
1265 	while (xfersize > 0) {
1266 		a_pg_offset = a_offset & PAGE_MASK;
1267 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1268 		moea64_set_scratchpage_pa(mmu, 0,
1269 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]));
1270 		a_cp = (char *)moea64_scratchpage_va[0] + a_pg_offset;
1271 		b_pg_offset = b_offset & PAGE_MASK;
1272 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1273 		moea64_set_scratchpage_pa(mmu, 1,
1274 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]));
1275 		b_cp = (char *)moea64_scratchpage_va[1] + b_pg_offset;
1276 		bcopy(a_cp, b_cp, cnt);
1277 		a_offset += cnt;
1278 		b_offset += cnt;
1279 		xfersize -= cnt;
1280 	}
1281 	mtx_unlock(&moea64_scratchpage_mtx);
1282 }
1283 
1284 void
1285 moea64_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1286     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1287 {
1288 
1289 	if (hw_direct_map) {
1290 		moea64_copy_pages_dmap(mmu, ma, a_offset, mb, b_offset,
1291 		    xfersize);
1292 	} else {
1293 		moea64_copy_pages_nodmap(mmu, ma, a_offset, mb, b_offset,
1294 		    xfersize);
1295 	}
1296 }
1297 
1298 void
1299 moea64_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1300 {
1301 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1302 
1303 	if (size + off > PAGE_SIZE)
1304 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1305 
1306 	if (hw_direct_map) {
1307 		bzero((caddr_t)(uintptr_t)PHYS_TO_DMAP(pa) + off, size);
1308 	} else {
1309 		mtx_lock(&moea64_scratchpage_mtx);
1310 		moea64_set_scratchpage_pa(mmu, 0, pa);
1311 		bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1312 		mtx_unlock(&moea64_scratchpage_mtx);
1313 	}
1314 }
1315 
1316 /*
1317  * Zero a page of physical memory by temporarily mapping it
1318  */
1319 void
1320 moea64_zero_page(mmu_t mmu, vm_page_t m)
1321 {
1322 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1323 	vm_offset_t va, off;
1324 
1325 	if (!hw_direct_map) {
1326 		mtx_lock(&moea64_scratchpage_mtx);
1327 
1328 		moea64_set_scratchpage_pa(mmu, 0, pa);
1329 		va = moea64_scratchpage_va[0];
1330 	} else {
1331 		va = PHYS_TO_DMAP(pa);
1332 	}
1333 
1334 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1335 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1336 
1337 	if (!hw_direct_map)
1338 		mtx_unlock(&moea64_scratchpage_mtx);
1339 }
1340 
1341 vm_offset_t
1342 moea64_quick_enter_page(mmu_t mmu, vm_page_t m)
1343 {
1344 	struct pvo_entry *pvo;
1345 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1346 
1347 	if (hw_direct_map)
1348 		return (PHYS_TO_DMAP(pa));
1349 
1350 	/*
1351  	 * MOEA64_PTE_REPLACE does some locking, so we can't just grab
1352 	 * a critical section and access the PCPU data like on i386.
1353 	 * Instead, pin the thread and grab the PCPU lock to prevent
1354 	 * a preempting thread from using the same PCPU data.
1355 	 */
1356 	sched_pin();
1357 
1358 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_NOTOWNED);
1359 	pvo = PCPU_GET(aim.qmap_pvo);
1360 
1361 	mtx_lock(PCPU_PTR(aim.qmap_lock));
1362 	pvo->pvo_pte.pa = moea64_calc_wimg(pa, pmap_page_get_memattr(m)) |
1363 	    (uint64_t)pa;
1364 	MOEA64_PTE_REPLACE(mmu, pvo, MOEA64_PTE_INVALIDATE);
1365 	isync();
1366 
1367 	return (PCPU_GET(qmap_addr));
1368 }
1369 
1370 void
1371 moea64_quick_remove_page(mmu_t mmu, vm_offset_t addr)
1372 {
1373 	if (hw_direct_map)
1374 		return;
1375 
1376 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_OWNED);
1377 	KASSERT(PCPU_GET(qmap_addr) == addr,
1378 	    ("moea64_quick_remove_page: invalid address"));
1379 	mtx_unlock(PCPU_PTR(aim.qmap_lock));
1380 	sched_unpin();
1381 }
1382 
1383 /*
1384  * Map the given physical page at the specified virtual address in the
1385  * target pmap with the protection requested.  If specified the page
1386  * will be wired down.
1387  */
1388 
1389 int
1390 moea64_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m,
1391     vm_prot_t prot, u_int flags, int8_t psind)
1392 {
1393 	struct		pvo_entry *pvo, *oldpvo;
1394 	struct		pvo_head *pvo_head;
1395 	uint64_t	pte_lo;
1396 	int		error;
1397 
1398 	if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m))
1399 		VM_OBJECT_ASSERT_LOCKED(m->object);
1400 
1401 	pvo = alloc_pvo_entry(0);
1402 	if (pvo == NULL)
1403 		return (KERN_RESOURCE_SHORTAGE);
1404 	pvo->pvo_pmap = NULL; /* to be filled in later */
1405 	pvo->pvo_pte.prot = prot;
1406 
1407 	pte_lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1408 	pvo->pvo_pte.pa = VM_PAGE_TO_PHYS(m) | pte_lo;
1409 
1410 	if ((flags & PMAP_ENTER_WIRED) != 0)
1411 		pvo->pvo_vaddr |= PVO_WIRED;
1412 
1413 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea64_initialized) {
1414 		pvo_head = NULL;
1415 	} else {
1416 		pvo_head = &m->md.mdpg_pvoh;
1417 		pvo->pvo_vaddr |= PVO_MANAGED;
1418 	}
1419 
1420 	PV_PAGE_LOCK(m);
1421 	PMAP_LOCK(pmap);
1422 	if (pvo->pvo_pmap == NULL)
1423 		init_pvo_entry(pvo, pmap, va);
1424 	if (prot & VM_PROT_WRITE)
1425 		if (pmap_bootstrapped &&
1426 		    (m->oflags & VPO_UNMANAGED) == 0)
1427 			vm_page_aflag_set(m, PGA_WRITEABLE);
1428 
1429 	error = moea64_pvo_enter(mmu, pvo, pvo_head, &oldpvo);
1430 	if (error == EEXIST) {
1431 		if (oldpvo->pvo_vaddr == pvo->pvo_vaddr &&
1432 		    oldpvo->pvo_pte.pa == pvo->pvo_pte.pa &&
1433 		    oldpvo->pvo_pte.prot == prot) {
1434 			/* Identical mapping already exists */
1435 			error = 0;
1436 
1437 			/* If not in page table, reinsert it */
1438 			if (MOEA64_PTE_SYNCH(mmu, oldpvo) < 0) {
1439 				STAT_MOEA64(moea64_pte_overflow--);
1440 				MOEA64_PTE_INSERT(mmu, oldpvo);
1441 			}
1442 
1443 			/* Then just clean up and go home */
1444 			PV_PAGE_UNLOCK(m);
1445 			PMAP_UNLOCK(pmap);
1446 			free_pvo_entry(pvo);
1447 			goto out;
1448 		} else {
1449 			/* Otherwise, need to kill it first */
1450 			KASSERT(oldpvo->pvo_pmap == pmap, ("pmap of old "
1451 			    "mapping does not match new mapping"));
1452 			moea64_pvo_remove_from_pmap(mmu, oldpvo);
1453 			moea64_pvo_enter(mmu, pvo, pvo_head, NULL);
1454 		}
1455 	}
1456 	PV_PAGE_UNLOCK(m);
1457 	PMAP_UNLOCK(pmap);
1458 
1459 	/* Free any dead pages */
1460 	if (error == EEXIST) {
1461 		moea64_pvo_remove_from_page(mmu, oldpvo);
1462 		free_pvo_entry(oldpvo);
1463 	}
1464 
1465 out:
1466 	/*
1467 	 * Flush the page from the instruction cache if this page is
1468 	 * mapped executable and cacheable.
1469 	 */
1470 	if (pmap != kernel_pmap && !(m->aflags & PGA_EXECUTABLE) &&
1471 	    (pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1472 		vm_page_aflag_set(m, PGA_EXECUTABLE);
1473 		moea64_syncicache(mmu, pmap, va, VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1474 	}
1475 	return (KERN_SUCCESS);
1476 }
1477 
1478 static void
1479 moea64_syncicache(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
1480     vm_size_t sz)
1481 {
1482 
1483 	/*
1484 	 * This is much trickier than on older systems because
1485 	 * we can't sync the icache on physical addresses directly
1486 	 * without a direct map. Instead we check a couple of cases
1487 	 * where the memory is already mapped in and, failing that,
1488 	 * use the same trick we use for page zeroing to create
1489 	 * a temporary mapping for this physical address.
1490 	 */
1491 
1492 	if (!pmap_bootstrapped) {
1493 		/*
1494 		 * If PMAP is not bootstrapped, we are likely to be
1495 		 * in real mode.
1496 		 */
1497 		__syncicache((void *)(uintptr_t)pa, sz);
1498 	} else if (pmap == kernel_pmap) {
1499 		__syncicache((void *)va, sz);
1500 	} else if (hw_direct_map) {
1501 		__syncicache((void *)(uintptr_t)PHYS_TO_DMAP(pa), sz);
1502 	} else {
1503 		/* Use the scratch page to set up a temp mapping */
1504 
1505 		mtx_lock(&moea64_scratchpage_mtx);
1506 
1507 		moea64_set_scratchpage_pa(mmu, 1, pa & ~ADDR_POFF);
1508 		__syncicache((void *)(moea64_scratchpage_va[1] +
1509 		    (va & ADDR_POFF)), sz);
1510 
1511 		mtx_unlock(&moea64_scratchpage_mtx);
1512 	}
1513 }
1514 
1515 /*
1516  * Maps a sequence of resident pages belonging to the same object.
1517  * The sequence begins with the given page m_start.  This page is
1518  * mapped at the given virtual address start.  Each subsequent page is
1519  * mapped at a virtual address that is offset from start by the same
1520  * amount as the page is offset from m_start within the object.  The
1521  * last page in the sequence is the page with the largest offset from
1522  * m_start that can be mapped at a virtual address less than the given
1523  * virtual address end.  Not every virtual page between start and end
1524  * is mapped; only those for which a resident page exists with the
1525  * corresponding offset from m_start are mapped.
1526  */
1527 void
1528 moea64_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1529     vm_page_t m_start, vm_prot_t prot)
1530 {
1531 	vm_page_t m;
1532 	vm_pindex_t diff, psize;
1533 
1534 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1535 
1536 	psize = atop(end - start);
1537 	m = m_start;
1538 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1539 		moea64_enter(mmu, pm, start + ptoa(diff), m, prot &
1540 		    (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_NOSLEEP, 0);
1541 		m = TAILQ_NEXT(m, listq);
1542 	}
1543 }
1544 
1545 void
1546 moea64_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1547     vm_prot_t prot)
1548 {
1549 
1550 	moea64_enter(mmu, pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1551 	    PMAP_ENTER_NOSLEEP, 0);
1552 }
1553 
1554 vm_paddr_t
1555 moea64_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1556 {
1557 	struct	pvo_entry *pvo;
1558 	vm_paddr_t pa;
1559 
1560 	PMAP_LOCK(pm);
1561 	pvo = moea64_pvo_find_va(pm, va);
1562 	if (pvo == NULL)
1563 		pa = 0;
1564 	else
1565 		pa = (pvo->pvo_pte.pa & LPTE_RPGN) | (va - PVO_VADDR(pvo));
1566 	PMAP_UNLOCK(pm);
1567 
1568 	return (pa);
1569 }
1570 
1571 /*
1572  * Atomically extract and hold the physical page with the given
1573  * pmap and virtual address pair if that mapping permits the given
1574  * protection.
1575  */
1576 vm_page_t
1577 moea64_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1578 {
1579 	struct	pvo_entry *pvo;
1580 	vm_page_t m;
1581 
1582 	m = NULL;
1583 	PMAP_LOCK(pmap);
1584 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1585 	if (pvo != NULL && (pvo->pvo_pte.prot & prot) == prot) {
1586 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
1587 		if (!vm_page_wire_mapped(m))
1588 			m = NULL;
1589 	}
1590 	PMAP_UNLOCK(pmap);
1591 	return (m);
1592 }
1593 
1594 static mmu_t installed_mmu;
1595 
1596 static void *
1597 moea64_uma_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain,
1598     uint8_t *flags, int wait)
1599 {
1600 	struct pvo_entry *pvo;
1601         vm_offset_t va;
1602         vm_page_t m;
1603         int needed_lock;
1604 
1605 	/*
1606 	 * This entire routine is a horrible hack to avoid bothering kmem
1607 	 * for new KVA addresses. Because this can get called from inside
1608 	 * kmem allocation routines, calling kmem for a new address here
1609 	 * can lead to multiply locking non-recursive mutexes.
1610 	 */
1611 
1612 	*flags = UMA_SLAB_PRIV;
1613 	needed_lock = !PMAP_LOCKED(kernel_pmap);
1614 
1615 	m = vm_page_alloc_domain(NULL, 0, domain,
1616 	    malloc2vm_flags(wait) | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1617 	if (m == NULL)
1618 		return (NULL);
1619 
1620 	va = VM_PAGE_TO_PHYS(m);
1621 
1622 	pvo = alloc_pvo_entry(1 /* bootstrap */);
1623 
1624 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE;
1625 	pvo->pvo_pte.pa = VM_PAGE_TO_PHYS(m) | LPTE_M;
1626 
1627 	if (needed_lock)
1628 		PMAP_LOCK(kernel_pmap);
1629 
1630 	init_pvo_entry(pvo, kernel_pmap, va);
1631 	pvo->pvo_vaddr |= PVO_WIRED;
1632 
1633 	moea64_pvo_enter(installed_mmu, pvo, NULL, NULL);
1634 
1635 	if (needed_lock)
1636 		PMAP_UNLOCK(kernel_pmap);
1637 
1638 	if ((wait & M_ZERO) && (m->flags & PG_ZERO) == 0)
1639                 bzero((void *)va, PAGE_SIZE);
1640 
1641 	return (void *)va;
1642 }
1643 
1644 extern int elf32_nxstack;
1645 
1646 void
1647 moea64_init(mmu_t mmu)
1648 {
1649 
1650 	CTR0(KTR_PMAP, "moea64_init");
1651 
1652 	moea64_pvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1653 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1654 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1655 
1656 	if (!hw_direct_map) {
1657 		installed_mmu = mmu;
1658 		uma_zone_set_allocf(moea64_pvo_zone, moea64_uma_page_alloc);
1659 	}
1660 
1661 #ifdef COMPAT_FREEBSD32
1662 	elf32_nxstack = 1;
1663 #endif
1664 
1665 	moea64_initialized = TRUE;
1666 }
1667 
1668 boolean_t
1669 moea64_is_referenced(mmu_t mmu, vm_page_t m)
1670 {
1671 
1672 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1673 	    ("moea64_is_referenced: page %p is not managed", m));
1674 
1675 	return (moea64_query_bit(mmu, m, LPTE_REF));
1676 }
1677 
1678 boolean_t
1679 moea64_is_modified(mmu_t mmu, vm_page_t m)
1680 {
1681 
1682 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1683 	    ("moea64_is_modified: page %p is not managed", m));
1684 
1685 	/*
1686 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1687 	 * concurrently set while the object is locked.  Thus, if PGA_WRITEABLE
1688 	 * is clear, no PTEs can have LPTE_CHG set.
1689 	 */
1690 	VM_OBJECT_ASSERT_LOCKED(m->object);
1691 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1692 		return (FALSE);
1693 	return (moea64_query_bit(mmu, m, LPTE_CHG));
1694 }
1695 
1696 boolean_t
1697 moea64_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1698 {
1699 	struct pvo_entry *pvo;
1700 	boolean_t rv = TRUE;
1701 
1702 	PMAP_LOCK(pmap);
1703 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1704 	if (pvo != NULL)
1705 		rv = FALSE;
1706 	PMAP_UNLOCK(pmap);
1707 	return (rv);
1708 }
1709 
1710 void
1711 moea64_clear_modify(mmu_t mmu, vm_page_t m)
1712 {
1713 
1714 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1715 	    ("moea64_clear_modify: page %p is not managed", m));
1716 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1717 	KASSERT(!vm_page_xbusied(m),
1718 	    ("moea64_clear_modify: page %p is exclusive busied", m));
1719 
1720 	/*
1721 	 * If the page is not PGA_WRITEABLE, then no PTEs can have LPTE_CHG
1722 	 * set.  If the object containing the page is locked and the page is
1723 	 * not exclusive busied, then PGA_WRITEABLE cannot be concurrently set.
1724 	 */
1725 	if ((m->aflags & PGA_WRITEABLE) == 0)
1726 		return;
1727 	moea64_clear_bit(mmu, m, LPTE_CHG);
1728 }
1729 
1730 /*
1731  * Clear the write and modified bits in each of the given page's mappings.
1732  */
1733 void
1734 moea64_remove_write(mmu_t mmu, vm_page_t m)
1735 {
1736 	struct	pvo_entry *pvo;
1737 	int64_t	refchg, ret;
1738 	pmap_t	pmap;
1739 
1740 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1741 	    ("moea64_remove_write: page %p is not managed", m));
1742 
1743 	/*
1744 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1745 	 * set by another thread while the object is locked.  Thus,
1746 	 * if PGA_WRITEABLE is clear, no page table entries need updating.
1747 	 */
1748 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1749 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1750 		return;
1751 	powerpc_sync();
1752 	PV_PAGE_LOCK(m);
1753 	refchg = 0;
1754 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1755 		pmap = pvo->pvo_pmap;
1756 		PMAP_LOCK(pmap);
1757 		if (!(pvo->pvo_vaddr & PVO_DEAD) &&
1758 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1759 			pvo->pvo_pte.prot &= ~VM_PROT_WRITE;
1760 			ret = MOEA64_PTE_REPLACE(mmu, pvo,
1761 			    MOEA64_PTE_PROT_UPDATE);
1762 			if (ret < 0)
1763 				ret = LPTE_CHG;
1764 			refchg |= ret;
1765 			if (pvo->pvo_pmap == kernel_pmap)
1766 				isync();
1767 		}
1768 		PMAP_UNLOCK(pmap);
1769 	}
1770 	if ((refchg | atomic_readandclear_32(&m->md.mdpg_attrs)) & LPTE_CHG)
1771 		vm_page_dirty(m);
1772 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1773 	PV_PAGE_UNLOCK(m);
1774 }
1775 
1776 /*
1777  *	moea64_ts_referenced:
1778  *
1779  *	Return a count of reference bits for a page, clearing those bits.
1780  *	It is not necessary for every reference bit to be cleared, but it
1781  *	is necessary that 0 only be returned when there are truly no
1782  *	reference bits set.
1783  *
1784  *	XXX: The exact number of bits to check and clear is a matter that
1785  *	should be tested and standardized at some point in the future for
1786  *	optimal aging of shared pages.
1787  */
1788 int
1789 moea64_ts_referenced(mmu_t mmu, vm_page_t m)
1790 {
1791 
1792 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1793 	    ("moea64_ts_referenced: page %p is not managed", m));
1794 	return (moea64_clear_bit(mmu, m, LPTE_REF));
1795 }
1796 
1797 /*
1798  * Modify the WIMG settings of all mappings for a page.
1799  */
1800 void
1801 moea64_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1802 {
1803 	struct	pvo_entry *pvo;
1804 	int64_t	refchg;
1805 	pmap_t	pmap;
1806 	uint64_t lo;
1807 
1808 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1809 		m->md.mdpg_cache_attrs = ma;
1810 		return;
1811 	}
1812 
1813 	lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1814 
1815 	PV_PAGE_LOCK(m);
1816 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1817 		pmap = pvo->pvo_pmap;
1818 		PMAP_LOCK(pmap);
1819 		if (!(pvo->pvo_vaddr & PVO_DEAD)) {
1820 			pvo->pvo_pte.pa &= ~LPTE_WIMG;
1821 			pvo->pvo_pte.pa |= lo;
1822 			refchg = MOEA64_PTE_REPLACE(mmu, pvo,
1823 			    MOEA64_PTE_INVALIDATE);
1824 			if (refchg < 0)
1825 				refchg = (pvo->pvo_pte.prot & VM_PROT_WRITE) ?
1826 				    LPTE_CHG : 0;
1827 			if ((pvo->pvo_vaddr & PVO_MANAGED) &&
1828 			    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1829 				refchg |=
1830 				    atomic_readandclear_32(&m->md.mdpg_attrs);
1831 				if (refchg & LPTE_CHG)
1832 					vm_page_dirty(m);
1833 				if (refchg & LPTE_REF)
1834 					vm_page_aflag_set(m, PGA_REFERENCED);
1835 			}
1836 			if (pvo->pvo_pmap == kernel_pmap)
1837 				isync();
1838 		}
1839 		PMAP_UNLOCK(pmap);
1840 	}
1841 	m->md.mdpg_cache_attrs = ma;
1842 	PV_PAGE_UNLOCK(m);
1843 }
1844 
1845 /*
1846  * Map a wired page into kernel virtual address space.
1847  */
1848 void
1849 moea64_kenter_attr(mmu_t mmu, vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
1850 {
1851 	int		error;
1852 	struct pvo_entry *pvo, *oldpvo;
1853 
1854 	do {
1855 		pvo = alloc_pvo_entry(0);
1856 		if (pvo == NULL)
1857 			vm_wait(NULL);
1858 	} while (pvo == NULL);
1859 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;
1860 	pvo->pvo_pte.pa = (pa & ~ADDR_POFF) | moea64_calc_wimg(pa, ma);
1861 	pvo->pvo_vaddr |= PVO_WIRED;
1862 
1863 	PMAP_LOCK(kernel_pmap);
1864 	oldpvo = moea64_pvo_find_va(kernel_pmap, va);
1865 	if (oldpvo != NULL)
1866 		moea64_pvo_remove_from_pmap(mmu, oldpvo);
1867 	init_pvo_entry(pvo, kernel_pmap, va);
1868 	error = moea64_pvo_enter(mmu, pvo, NULL, NULL);
1869 	PMAP_UNLOCK(kernel_pmap);
1870 
1871 	/* Free any dead pages */
1872 	if (oldpvo != NULL) {
1873 		moea64_pvo_remove_from_page(mmu, oldpvo);
1874 		free_pvo_entry(oldpvo);
1875 	}
1876 
1877 	if (error != 0 && error != ENOENT)
1878 		panic("moea64_kenter: failed to enter va %#zx pa %#jx: %d", va,
1879 		    (uintmax_t)pa, error);
1880 }
1881 
1882 void
1883 moea64_kenter(mmu_t mmu, vm_offset_t va, vm_paddr_t pa)
1884 {
1885 
1886 	moea64_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1887 }
1888 
1889 /*
1890  * Extract the physical page address associated with the given kernel virtual
1891  * address.
1892  */
1893 vm_paddr_t
1894 moea64_kextract(mmu_t mmu, vm_offset_t va)
1895 {
1896 	struct		pvo_entry *pvo;
1897 	vm_paddr_t pa;
1898 
1899 	/*
1900 	 * Shortcut the direct-mapped case when applicable.  We never put
1901 	 * anything but 1:1 (or 62-bit aliased) mappings below
1902 	 * VM_MIN_KERNEL_ADDRESS.
1903 	 */
1904 	if (va < VM_MIN_KERNEL_ADDRESS)
1905 		return (va & ~DMAP_BASE_ADDRESS);
1906 
1907 	PMAP_LOCK(kernel_pmap);
1908 	pvo = moea64_pvo_find_va(kernel_pmap, va);
1909 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found for %#" PRIxPTR,
1910 	    va));
1911 	pa = (pvo->pvo_pte.pa & LPTE_RPGN) | (va - PVO_VADDR(pvo));
1912 	PMAP_UNLOCK(kernel_pmap);
1913 	return (pa);
1914 }
1915 
1916 /*
1917  * Remove a wired page from kernel virtual address space.
1918  */
1919 void
1920 moea64_kremove(mmu_t mmu, vm_offset_t va)
1921 {
1922 	moea64_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1923 }
1924 
1925 /*
1926  * Provide a kernel pointer corresponding to a given userland pointer.
1927  * The returned pointer is valid until the next time this function is
1928  * called in this thread. This is used internally in copyin/copyout.
1929  */
1930 static int
1931 moea64_map_user_ptr(mmu_t mmu, pmap_t pm, volatile const void *uaddr,
1932     void **kaddr, size_t ulen, size_t *klen)
1933 {
1934 	size_t l;
1935 #ifdef __powerpc64__
1936 	struct slb *slb;
1937 #endif
1938 	register_t slbv;
1939 
1940 	*kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
1941 	l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
1942 	if (l > ulen)
1943 		l = ulen;
1944 	if (klen)
1945 		*klen = l;
1946 	else if (l != ulen)
1947 		return (EFAULT);
1948 
1949 #ifdef __powerpc64__
1950 	/* Try lockless look-up first */
1951 	slb = user_va_to_slb_entry(pm, (vm_offset_t)uaddr);
1952 
1953 	if (slb == NULL) {
1954 		/* If it isn't there, we need to pre-fault the VSID */
1955 		PMAP_LOCK(pm);
1956 		slbv = va_to_vsid(pm, (vm_offset_t)uaddr) << SLBV_VSID_SHIFT;
1957 		PMAP_UNLOCK(pm);
1958 	} else {
1959 		slbv = slb->slbv;
1960 	}
1961 
1962 	/* Mark segment no-execute */
1963 	slbv |= SLBV_N;
1964 #else
1965 	slbv = va_to_vsid(pm, (vm_offset_t)uaddr);
1966 
1967 	/* Mark segment no-execute */
1968 	slbv |= SR_N;
1969 #endif
1970 
1971 	/* If we have already set this VSID, we can just return */
1972 	if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == slbv)
1973 		return (0);
1974 
1975 	__asm __volatile("isync");
1976 	curthread->td_pcb->pcb_cpu.aim.usr_segm =
1977 	    (uintptr_t)uaddr >> ADDR_SR_SHFT;
1978 	curthread->td_pcb->pcb_cpu.aim.usr_vsid = slbv;
1979 #ifdef __powerpc64__
1980 	__asm __volatile ("slbie %0; slbmte %1, %2; isync" ::
1981 	    "r"(USER_ADDR), "r"(slbv), "r"(USER_SLB_SLBE));
1982 #else
1983 	__asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(slbv));
1984 #endif
1985 
1986 	return (0);
1987 }
1988 
1989 /*
1990  * Figure out where a given kernel pointer (usually in a fault) points
1991  * to from the VM's perspective, potentially remapping into userland's
1992  * address space.
1993  */
1994 static int
1995 moea64_decode_kernel_ptr(mmu_t mmu, vm_offset_t addr, int *is_user,
1996     vm_offset_t *decoded_addr)
1997 {
1998 	vm_offset_t user_sr;
1999 
2000 	if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
2001 		user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
2002 		addr &= ADDR_PIDX | ADDR_POFF;
2003 		addr |= user_sr << ADDR_SR_SHFT;
2004 		*decoded_addr = addr;
2005 		*is_user = 1;
2006 	} else {
2007 		*decoded_addr = addr;
2008 		*is_user = 0;
2009 	}
2010 
2011 	return (0);
2012 }
2013 
2014 /*
2015  * Map a range of physical addresses into kernel virtual address space.
2016  *
2017  * The value passed in *virt is a suggested virtual address for the mapping.
2018  * Architectures which can support a direct-mapped physical to virtual region
2019  * can return the appropriate address within that region, leaving '*virt'
2020  * unchanged.  Other architectures should map the pages starting at '*virt' and
2021  * update '*virt' with the first usable address after the mapped region.
2022  */
2023 vm_offset_t
2024 moea64_map(mmu_t mmu, vm_offset_t *virt, vm_paddr_t pa_start,
2025     vm_paddr_t pa_end, int prot)
2026 {
2027 	vm_offset_t	sva, va;
2028 
2029 	if (hw_direct_map) {
2030 		/*
2031 		 * Check if every page in the region is covered by the direct
2032 		 * map. The direct map covers all of physical memory. Use
2033 		 * moea64_calc_wimg() as a shortcut to see if the page is in
2034 		 * physical memory as a way to see if the direct map covers it.
2035 		 */
2036 		for (va = pa_start; va < pa_end; va += PAGE_SIZE)
2037 			if (moea64_calc_wimg(va, VM_MEMATTR_DEFAULT) != LPTE_M)
2038 				break;
2039 		if (va == pa_end)
2040 			return (PHYS_TO_DMAP(pa_start));
2041 	}
2042 	sva = *virt;
2043 	va = sva;
2044 	/* XXX respect prot argument */
2045 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
2046 		moea64_kenter(mmu, va, pa_start);
2047 	*virt = va;
2048 
2049 	return (sva);
2050 }
2051 
2052 /*
2053  * Returns true if the pmap's pv is one of the first
2054  * 16 pvs linked to from this page.  This count may
2055  * be changed upwards or downwards in the future; it
2056  * is only necessary that true be returned for a small
2057  * subset of pmaps for proper page aging.
2058  */
2059 boolean_t
2060 moea64_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
2061 {
2062         int loops;
2063 	struct pvo_entry *pvo;
2064 	boolean_t rv;
2065 
2066 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2067 	    ("moea64_page_exists_quick: page %p is not managed", m));
2068 	loops = 0;
2069 	rv = FALSE;
2070 	PV_PAGE_LOCK(m);
2071 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2072 		if (!(pvo->pvo_vaddr & PVO_DEAD) && pvo->pvo_pmap == pmap) {
2073 			rv = TRUE;
2074 			break;
2075 		}
2076 		if (++loops >= 16)
2077 			break;
2078 	}
2079 	PV_PAGE_UNLOCK(m);
2080 	return (rv);
2081 }
2082 
2083 void
2084 moea64_page_init(mmu_t mmu __unused, vm_page_t m)
2085 {
2086 
2087 	m->md.mdpg_attrs = 0;
2088 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
2089 	LIST_INIT(&m->md.mdpg_pvoh);
2090 }
2091 
2092 /*
2093  * Return the number of managed mappings to the given physical page
2094  * that are wired.
2095  */
2096 int
2097 moea64_page_wired_mappings(mmu_t mmu, vm_page_t m)
2098 {
2099 	struct pvo_entry *pvo;
2100 	int count;
2101 
2102 	count = 0;
2103 	if ((m->oflags & VPO_UNMANAGED) != 0)
2104 		return (count);
2105 	PV_PAGE_LOCK(m);
2106 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
2107 		if ((pvo->pvo_vaddr & (PVO_DEAD | PVO_WIRED)) == PVO_WIRED)
2108 			count++;
2109 	PV_PAGE_UNLOCK(m);
2110 	return (count);
2111 }
2112 
2113 static uintptr_t	moea64_vsidcontext;
2114 
2115 uintptr_t
2116 moea64_get_unique_vsid(void) {
2117 	u_int entropy;
2118 	register_t hash;
2119 	uint32_t mask;
2120 	int i;
2121 
2122 	entropy = 0;
2123 	__asm __volatile("mftb %0" : "=r"(entropy));
2124 
2125 	mtx_lock(&moea64_slb_mutex);
2126 	for (i = 0; i < NVSIDS; i += VSID_NBPW) {
2127 		u_int	n;
2128 
2129 		/*
2130 		 * Create a new value by mutiplying by a prime and adding in
2131 		 * entropy from the timebase register.  This is to make the
2132 		 * VSID more random so that the PT hash function collides
2133 		 * less often.  (Note that the prime casues gcc to do shifts
2134 		 * instead of a multiply.)
2135 		 */
2136 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
2137 		hash = moea64_vsidcontext & (NVSIDS - 1);
2138 		if (hash == 0)		/* 0 is special, avoid it */
2139 			continue;
2140 		n = hash >> 5;
2141 		mask = 1 << (hash & (VSID_NBPW - 1));
2142 		hash = (moea64_vsidcontext & VSID_HASHMASK);
2143 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
2144 			/* anything free in this bucket? */
2145 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
2146 				entropy = (moea64_vsidcontext >> 20);
2147 				continue;
2148 			}
2149 			i = ffs(~moea64_vsid_bitmap[n]) - 1;
2150 			mask = 1 << i;
2151 			hash &= rounddown2(VSID_HASHMASK, VSID_NBPW);
2152 			hash |= i;
2153 		}
2154 		if (hash == VSID_VRMA)	/* also special, avoid this too */
2155 			continue;
2156 		KASSERT(!(moea64_vsid_bitmap[n] & mask),
2157 		    ("Allocating in-use VSID %#zx\n", hash));
2158 		moea64_vsid_bitmap[n] |= mask;
2159 		mtx_unlock(&moea64_slb_mutex);
2160 		return (hash);
2161 	}
2162 
2163 	mtx_unlock(&moea64_slb_mutex);
2164 	panic("%s: out of segments",__func__);
2165 }
2166 
2167 #ifdef __powerpc64__
2168 void
2169 moea64_pinit(mmu_t mmu, pmap_t pmap)
2170 {
2171 
2172 	RB_INIT(&pmap->pmap_pvo);
2173 
2174 	pmap->pm_slb_tree_root = slb_alloc_tree();
2175 	pmap->pm_slb = slb_alloc_user_cache();
2176 	pmap->pm_slb_len = 0;
2177 }
2178 #else
2179 void
2180 moea64_pinit(mmu_t mmu, pmap_t pmap)
2181 {
2182 	int	i;
2183 	uint32_t hash;
2184 
2185 	RB_INIT(&pmap->pmap_pvo);
2186 
2187 	if (pmap_bootstrapped)
2188 		pmap->pmap_phys = (pmap_t)moea64_kextract(mmu,
2189 		    (vm_offset_t)pmap);
2190 	else
2191 		pmap->pmap_phys = pmap;
2192 
2193 	/*
2194 	 * Allocate some segment registers for this pmap.
2195 	 */
2196 	hash = moea64_get_unique_vsid();
2197 
2198 	for (i = 0; i < 16; i++)
2199 		pmap->pm_sr[i] = VSID_MAKE(i, hash);
2200 
2201 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_pinit: pm_sr[0] = 0"));
2202 }
2203 #endif
2204 
2205 /*
2206  * Initialize the pmap associated with process 0.
2207  */
2208 void
2209 moea64_pinit0(mmu_t mmu, pmap_t pm)
2210 {
2211 
2212 	PMAP_LOCK_INIT(pm);
2213 	moea64_pinit(mmu, pm);
2214 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
2215 }
2216 
2217 /*
2218  * Set the physical protection on the specified range of this map as requested.
2219  */
2220 static void
2221 moea64_pvo_protect(mmu_t mmu,  pmap_t pm, struct pvo_entry *pvo, vm_prot_t prot)
2222 {
2223 	struct vm_page *pg;
2224 	vm_prot_t oldprot;
2225 	int32_t refchg;
2226 
2227 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2228 
2229 	/*
2230 	 * Change the protection of the page.
2231 	 */
2232 	oldprot = pvo->pvo_pte.prot;
2233 	pvo->pvo_pte.prot = prot;
2234 	pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
2235 
2236 	/*
2237 	 * If the PVO is in the page table, update mapping
2238 	 */
2239 	refchg = MOEA64_PTE_REPLACE(mmu, pvo, MOEA64_PTE_PROT_UPDATE);
2240 	if (refchg < 0)
2241 		refchg = (oldprot & VM_PROT_WRITE) ? LPTE_CHG : 0;
2242 
2243 	if (pm != kernel_pmap && pg != NULL && !(pg->aflags & PGA_EXECUTABLE) &&
2244 	    (pvo->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
2245 		if ((pg->oflags & VPO_UNMANAGED) == 0)
2246 			vm_page_aflag_set(pg, PGA_EXECUTABLE);
2247 		moea64_syncicache(mmu, pm, PVO_VADDR(pvo),
2248 		    pvo->pvo_pte.pa & LPTE_RPGN, PAGE_SIZE);
2249 	}
2250 
2251 	/*
2252 	 * Update vm about the REF/CHG bits if the page is managed and we have
2253 	 * removed write access.
2254 	 */
2255 	if (pg != NULL && (pvo->pvo_vaddr & PVO_MANAGED) &&
2256 	    (oldprot & VM_PROT_WRITE)) {
2257 		refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2258 		if (refchg & LPTE_CHG)
2259 			vm_page_dirty(pg);
2260 		if (refchg & LPTE_REF)
2261 			vm_page_aflag_set(pg, PGA_REFERENCED);
2262 	}
2263 }
2264 
2265 void
2266 moea64_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
2267     vm_prot_t prot)
2268 {
2269 	struct	pvo_entry *pvo, *tpvo, key;
2270 
2271 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm,
2272 	    sva, eva, prot);
2273 
2274 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
2275 	    ("moea64_protect: non current pmap"));
2276 
2277 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2278 		moea64_remove(mmu, pm, sva, eva);
2279 		return;
2280 	}
2281 
2282 	PMAP_LOCK(pm);
2283 	key.pvo_vaddr = sva;
2284 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2285 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2286 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2287 		moea64_pvo_protect(mmu, pm, pvo, prot);
2288 	}
2289 	PMAP_UNLOCK(pm);
2290 }
2291 
2292 /*
2293  * Map a list of wired pages into kernel virtual address space.  This is
2294  * intended for temporary mappings which do not need page modification or
2295  * references recorded.  Existing mappings in the region are overwritten.
2296  */
2297 void
2298 moea64_qenter(mmu_t mmu, vm_offset_t va, vm_page_t *m, int count)
2299 {
2300 	while (count-- > 0) {
2301 		moea64_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
2302 		va += PAGE_SIZE;
2303 		m++;
2304 	}
2305 }
2306 
2307 /*
2308  * Remove page mappings from kernel virtual address space.  Intended for
2309  * temporary mappings entered by moea64_qenter.
2310  */
2311 void
2312 moea64_qremove(mmu_t mmu, vm_offset_t va, int count)
2313 {
2314 	while (count-- > 0) {
2315 		moea64_kremove(mmu, va);
2316 		va += PAGE_SIZE;
2317 	}
2318 }
2319 
2320 void
2321 moea64_release_vsid(uint64_t vsid)
2322 {
2323 	int idx, mask;
2324 
2325 	mtx_lock(&moea64_slb_mutex);
2326 	idx = vsid & (NVSIDS-1);
2327 	mask = 1 << (idx % VSID_NBPW);
2328 	idx /= VSID_NBPW;
2329 	KASSERT(moea64_vsid_bitmap[idx] & mask,
2330 	    ("Freeing unallocated VSID %#jx", vsid));
2331 	moea64_vsid_bitmap[idx] &= ~mask;
2332 	mtx_unlock(&moea64_slb_mutex);
2333 }
2334 
2335 
2336 void
2337 moea64_release(mmu_t mmu, pmap_t pmap)
2338 {
2339 
2340 	/*
2341 	 * Free segment registers' VSIDs
2342 	 */
2343     #ifdef __powerpc64__
2344 	slb_free_tree(pmap);
2345 	slb_free_user_cache(pmap->pm_slb);
2346     #else
2347 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_release: pm_sr[0] = 0"));
2348 
2349 	moea64_release_vsid(VSID_TO_HASH(pmap->pm_sr[0]));
2350     #endif
2351 }
2352 
2353 /*
2354  * Remove all pages mapped by the specified pmap
2355  */
2356 void
2357 moea64_remove_pages(mmu_t mmu, pmap_t pm)
2358 {
2359 	struct pvo_entry *pvo, *tpvo;
2360 	struct pvo_dlist tofree;
2361 
2362 	SLIST_INIT(&tofree);
2363 
2364 	PMAP_LOCK(pm);
2365 	RB_FOREACH_SAFE(pvo, pvo_tree, &pm->pmap_pvo, tpvo) {
2366 		if (pvo->pvo_vaddr & PVO_WIRED)
2367 			continue;
2368 
2369 		/*
2370 		 * For locking reasons, remove this from the page table and
2371 		 * pmap, but save delinking from the vm_page for a second
2372 		 * pass
2373 		 */
2374 		moea64_pvo_remove_from_pmap(mmu, pvo);
2375 		SLIST_INSERT_HEAD(&tofree, pvo, pvo_dlink);
2376 	}
2377 	PMAP_UNLOCK(pm);
2378 
2379 	while (!SLIST_EMPTY(&tofree)) {
2380 		pvo = SLIST_FIRST(&tofree);
2381 		SLIST_REMOVE_HEAD(&tofree, pvo_dlink);
2382 		moea64_pvo_remove_from_page(mmu, pvo);
2383 		free_pvo_entry(pvo);
2384 	}
2385 }
2386 
2387 /*
2388  * Remove the given range of addresses from the specified map.
2389  */
2390 void
2391 moea64_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
2392 {
2393 	struct  pvo_entry *pvo, *tpvo, key;
2394 	struct pvo_dlist tofree;
2395 
2396 	/*
2397 	 * Perform an unsynchronized read.  This is, however, safe.
2398 	 */
2399 	if (pm->pm_stats.resident_count == 0)
2400 		return;
2401 
2402 	key.pvo_vaddr = sva;
2403 
2404 	SLIST_INIT(&tofree);
2405 
2406 	PMAP_LOCK(pm);
2407 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2408 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2409 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2410 
2411 		/*
2412 		 * For locking reasons, remove this from the page table and
2413 		 * pmap, but save delinking from the vm_page for a second
2414 		 * pass
2415 		 */
2416 		moea64_pvo_remove_from_pmap(mmu, pvo);
2417 		SLIST_INSERT_HEAD(&tofree, pvo, pvo_dlink);
2418 	}
2419 	PMAP_UNLOCK(pm);
2420 
2421 	while (!SLIST_EMPTY(&tofree)) {
2422 		pvo = SLIST_FIRST(&tofree);
2423 		SLIST_REMOVE_HEAD(&tofree, pvo_dlink);
2424 		moea64_pvo_remove_from_page(mmu, pvo);
2425 		free_pvo_entry(pvo);
2426 	}
2427 }
2428 
2429 /*
2430  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
2431  * will reflect changes in pte's back to the vm_page.
2432  */
2433 void
2434 moea64_remove_all(mmu_t mmu, vm_page_t m)
2435 {
2436 	struct	pvo_entry *pvo, *next_pvo;
2437 	struct	pvo_head freequeue;
2438 	int	wasdead;
2439 	pmap_t	pmap;
2440 
2441 	LIST_INIT(&freequeue);
2442 
2443 	PV_PAGE_LOCK(m);
2444 	LIST_FOREACH_SAFE(pvo, vm_page_to_pvoh(m), pvo_vlink, next_pvo) {
2445 		pmap = pvo->pvo_pmap;
2446 		PMAP_LOCK(pmap);
2447 		wasdead = (pvo->pvo_vaddr & PVO_DEAD);
2448 		if (!wasdead)
2449 			moea64_pvo_remove_from_pmap(mmu, pvo);
2450 		moea64_pvo_remove_from_page_locked(mmu, pvo, m);
2451 		if (!wasdead)
2452 			LIST_INSERT_HEAD(&freequeue, pvo, pvo_vlink);
2453 		PMAP_UNLOCK(pmap);
2454 
2455 	}
2456 	KASSERT(!pmap_page_is_mapped(m), ("Page still has mappings"));
2457 	KASSERT(!(m->aflags & PGA_WRITEABLE), ("Page still writable"));
2458 	PV_PAGE_UNLOCK(m);
2459 
2460 	/* Clean up UMA allocations */
2461 	LIST_FOREACH_SAFE(pvo, &freequeue, pvo_vlink, next_pvo)
2462 		free_pvo_entry(pvo);
2463 }
2464 
2465 /*
2466  * Allocate a physical page of memory directly from the phys_avail map.
2467  * Can only be called from moea64_bootstrap before avail start and end are
2468  * calculated.
2469  */
2470 vm_offset_t
2471 moea64_bootstrap_alloc(vm_size_t size, vm_size_t align)
2472 {
2473 	vm_offset_t	s, e;
2474 	int		i, j;
2475 
2476 	size = round_page(size);
2477 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2478 		if (align != 0)
2479 			s = roundup2(phys_avail[i], align);
2480 		else
2481 			s = phys_avail[i];
2482 		e = s + size;
2483 
2484 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2485 			continue;
2486 
2487 		if (s + size > platform_real_maxaddr())
2488 			continue;
2489 
2490 		if (s == phys_avail[i]) {
2491 			phys_avail[i] += size;
2492 		} else if (e == phys_avail[i + 1]) {
2493 			phys_avail[i + 1] -= size;
2494 		} else {
2495 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2496 				phys_avail[j] = phys_avail[j - 2];
2497 				phys_avail[j + 1] = phys_avail[j - 1];
2498 			}
2499 
2500 			phys_avail[i + 3] = phys_avail[i + 1];
2501 			phys_avail[i + 1] = s;
2502 			phys_avail[i + 2] = e;
2503 			phys_avail_count++;
2504 		}
2505 
2506 		return (s);
2507 	}
2508 	panic("moea64_bootstrap_alloc: could not allocate memory");
2509 }
2510 
2511 static int
2512 moea64_pvo_enter(mmu_t mmu, struct pvo_entry *pvo, struct pvo_head *pvo_head,
2513     struct pvo_entry **oldpvop)
2514 {
2515 	int first, err;
2516 	struct pvo_entry *old_pvo;
2517 
2518 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2519 
2520 	STAT_MOEA64(moea64_pvo_enter_calls++);
2521 
2522 	/*
2523 	 * Add to pmap list
2524 	 */
2525 	old_pvo = RB_INSERT(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2526 
2527 	if (old_pvo != NULL) {
2528 		if (oldpvop != NULL)
2529 			*oldpvop = old_pvo;
2530 		return (EEXIST);
2531 	}
2532 
2533 	/*
2534 	 * Remember if the list was empty and therefore will be the first
2535 	 * item.
2536 	 */
2537 	if (pvo_head != NULL) {
2538 		if (LIST_FIRST(pvo_head) == NULL)
2539 			first = 1;
2540 		LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2541 	}
2542 
2543 	if (pvo->pvo_vaddr & PVO_WIRED)
2544 		pvo->pvo_pmap->pm_stats.wired_count++;
2545 	pvo->pvo_pmap->pm_stats.resident_count++;
2546 
2547 	/*
2548 	 * Insert it into the hardware page table
2549 	 */
2550 	err = MOEA64_PTE_INSERT(mmu, pvo);
2551 	if (err != 0) {
2552 		panic("moea64_pvo_enter: overflow");
2553 	}
2554 
2555 	STAT_MOEA64(moea64_pvo_entries++);
2556 
2557 	if (pvo->pvo_pmap == kernel_pmap)
2558 		isync();
2559 
2560 #ifdef __powerpc64__
2561 	/*
2562 	 * Make sure all our bootstrap mappings are in the SLB as soon
2563 	 * as virtual memory is switched on.
2564 	 */
2565 	if (!pmap_bootstrapped)
2566 		moea64_bootstrap_slb_prefault(PVO_VADDR(pvo),
2567 		    pvo->pvo_vaddr & PVO_LARGE);
2568 #endif
2569 
2570 	return (first ? ENOENT : 0);
2571 }
2572 
2573 static void
2574 moea64_pvo_remove_from_pmap(mmu_t mmu, struct pvo_entry *pvo)
2575 {
2576 	struct	vm_page *pg;
2577 	int32_t refchg;
2578 
2579 	KASSERT(pvo->pvo_pmap != NULL, ("Trying to remove PVO with no pmap"));
2580 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2581 	KASSERT(!(pvo->pvo_vaddr & PVO_DEAD), ("Trying to remove dead PVO"));
2582 
2583 	/*
2584 	 * If there is an active pte entry, we need to deactivate it
2585 	 */
2586 	refchg = MOEA64_PTE_UNSET(mmu, pvo);
2587 	if (refchg < 0) {
2588 		/*
2589 		 * If it was evicted from the page table, be pessimistic and
2590 		 * dirty the page.
2591 		 */
2592 		if (pvo->pvo_pte.prot & VM_PROT_WRITE)
2593 			refchg = LPTE_CHG;
2594 		else
2595 			refchg = 0;
2596 	}
2597 
2598 	/*
2599 	 * Update our statistics.
2600 	 */
2601 	pvo->pvo_pmap->pm_stats.resident_count--;
2602 	if (pvo->pvo_vaddr & PVO_WIRED)
2603 		pvo->pvo_pmap->pm_stats.wired_count--;
2604 
2605 	/*
2606 	 * Remove this PVO from the pmap list.
2607 	 */
2608 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2609 
2610 	/*
2611 	 * Mark this for the next sweep
2612 	 */
2613 	pvo->pvo_vaddr |= PVO_DEAD;
2614 
2615 	/* Send RC bits to VM */
2616 	if ((pvo->pvo_vaddr & PVO_MANAGED) &&
2617 	    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2618 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
2619 		if (pg != NULL) {
2620 			refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2621 			if (refchg & LPTE_CHG)
2622 				vm_page_dirty(pg);
2623 			if (refchg & LPTE_REF)
2624 				vm_page_aflag_set(pg, PGA_REFERENCED);
2625 		}
2626 	}
2627 }
2628 
2629 static inline void
2630 moea64_pvo_remove_from_page_locked(mmu_t mmu, struct pvo_entry *pvo,
2631     vm_page_t m)
2632 {
2633 
2634 	KASSERT(pvo->pvo_vaddr & PVO_DEAD, ("Trying to delink live page"));
2635 
2636 	/* Use NULL pmaps as a sentinel for races in page deletion */
2637 	if (pvo->pvo_pmap == NULL)
2638 		return;
2639 	pvo->pvo_pmap = NULL;
2640 
2641 	/*
2642 	 * Update vm about page writeability/executability if managed
2643 	 */
2644 	PV_LOCKASSERT(pvo->pvo_pte.pa & LPTE_RPGN);
2645 	if (pvo->pvo_vaddr & PVO_MANAGED) {
2646 		if (m != NULL) {
2647 			LIST_REMOVE(pvo, pvo_vlink);
2648 			if (LIST_EMPTY(vm_page_to_pvoh(m)))
2649 				vm_page_aflag_clear(m,
2650 				    PGA_WRITEABLE | PGA_EXECUTABLE);
2651 		}
2652 	}
2653 
2654 	STAT_MOEA64(moea64_pvo_entries--);
2655 	STAT_MOEA64(moea64_pvo_remove_calls++);
2656 }
2657 
2658 static void
2659 moea64_pvo_remove_from_page(mmu_t mmu, struct pvo_entry *pvo)
2660 {
2661 	vm_page_t pg = NULL;
2662 
2663 	if (pvo->pvo_vaddr & PVO_MANAGED)
2664 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
2665 
2666 	PV_LOCK(pvo->pvo_pte.pa & LPTE_RPGN);
2667 	moea64_pvo_remove_from_page_locked(mmu, pvo, pg);
2668 	PV_UNLOCK(pvo->pvo_pte.pa & LPTE_RPGN);
2669 }
2670 
2671 static struct pvo_entry *
2672 moea64_pvo_find_va(pmap_t pm, vm_offset_t va)
2673 {
2674 	struct pvo_entry key;
2675 
2676 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2677 
2678 	key.pvo_vaddr = va & ~ADDR_POFF;
2679 	return (RB_FIND(pvo_tree, &pm->pmap_pvo, &key));
2680 }
2681 
2682 static boolean_t
2683 moea64_query_bit(mmu_t mmu, vm_page_t m, uint64_t ptebit)
2684 {
2685 	struct	pvo_entry *pvo;
2686 	int64_t ret;
2687 	boolean_t rv;
2688 
2689 	/*
2690 	 * See if this bit is stored in the page already.
2691 	 */
2692 	if (m->md.mdpg_attrs & ptebit)
2693 		return (TRUE);
2694 
2695 	/*
2696 	 * Examine each PTE.  Sync so that any pending REF/CHG bits are
2697 	 * flushed to the PTEs.
2698 	 */
2699 	rv = FALSE;
2700 	powerpc_sync();
2701 	PV_PAGE_LOCK(m);
2702 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2703 		ret = 0;
2704 
2705 		/*
2706 		 * See if this pvo has a valid PTE.  if so, fetch the
2707 		 * REF/CHG bits from the valid PTE.  If the appropriate
2708 		 * ptebit is set, return success.
2709 		 */
2710 		PMAP_LOCK(pvo->pvo_pmap);
2711 		if (!(pvo->pvo_vaddr & PVO_DEAD))
2712 			ret = MOEA64_PTE_SYNCH(mmu, pvo);
2713 		PMAP_UNLOCK(pvo->pvo_pmap);
2714 
2715 		if (ret > 0) {
2716 			atomic_set_32(&m->md.mdpg_attrs,
2717 			    ret & (LPTE_CHG | LPTE_REF));
2718 			if (ret & ptebit) {
2719 				rv = TRUE;
2720 				break;
2721 			}
2722 		}
2723 	}
2724 	PV_PAGE_UNLOCK(m);
2725 
2726 	return (rv);
2727 }
2728 
2729 static u_int
2730 moea64_clear_bit(mmu_t mmu, vm_page_t m, u_int64_t ptebit)
2731 {
2732 	u_int	count;
2733 	struct	pvo_entry *pvo;
2734 	int64_t ret;
2735 
2736 	/*
2737 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2738 	 * we can reset the right ones).
2739 	 */
2740 	powerpc_sync();
2741 
2742 	/*
2743 	 * For each pvo entry, clear the pte's ptebit.
2744 	 */
2745 	count = 0;
2746 	PV_PAGE_LOCK(m);
2747 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2748 		ret = 0;
2749 
2750 		PMAP_LOCK(pvo->pvo_pmap);
2751 		if (!(pvo->pvo_vaddr & PVO_DEAD))
2752 			ret = MOEA64_PTE_CLEAR(mmu, pvo, ptebit);
2753 		PMAP_UNLOCK(pvo->pvo_pmap);
2754 
2755 		if (ret > 0 && (ret & ptebit))
2756 			count++;
2757 	}
2758 	atomic_clear_32(&m->md.mdpg_attrs, ptebit);
2759 	PV_PAGE_UNLOCK(m);
2760 
2761 	return (count);
2762 }
2763 
2764 boolean_t
2765 moea64_dev_direct_mapped(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2766 {
2767 	struct pvo_entry *pvo, key;
2768 	vm_offset_t ppa;
2769 	int error = 0;
2770 
2771 	if (hw_direct_map && mem_valid(pa, size) == 0)
2772 		return (0);
2773 
2774 	PMAP_LOCK(kernel_pmap);
2775 	ppa = pa & ~ADDR_POFF;
2776 	key.pvo_vaddr = DMAP_BASE_ADDRESS + ppa;
2777 	for (pvo = RB_FIND(pvo_tree, &kernel_pmap->pmap_pvo, &key);
2778 	    ppa < pa + size; ppa += PAGE_SIZE,
2779 	    pvo = RB_NEXT(pvo_tree, &kernel_pmap->pmap_pvo, pvo)) {
2780 		if (pvo == NULL || (pvo->pvo_pte.pa & LPTE_RPGN) != ppa) {
2781 			error = EFAULT;
2782 			break;
2783 		}
2784 	}
2785 	PMAP_UNLOCK(kernel_pmap);
2786 
2787 	return (error);
2788 }
2789 
2790 /*
2791  * Map a set of physical memory pages into the kernel virtual
2792  * address space. Return a pointer to where it is mapped. This
2793  * routine is intended to be used for mapping device memory,
2794  * NOT real memory.
2795  */
2796 void *
2797 moea64_mapdev_attr(mmu_t mmu, vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
2798 {
2799 	vm_offset_t va, tmpva, ppa, offset;
2800 
2801 	ppa = trunc_page(pa);
2802 	offset = pa & PAGE_MASK;
2803 	size = roundup2(offset + size, PAGE_SIZE);
2804 
2805 	va = kva_alloc(size);
2806 
2807 	if (!va)
2808 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
2809 
2810 	for (tmpva = va; size > 0;) {
2811 		moea64_kenter_attr(mmu, tmpva, ppa, ma);
2812 		size -= PAGE_SIZE;
2813 		tmpva += PAGE_SIZE;
2814 		ppa += PAGE_SIZE;
2815 	}
2816 
2817 	return ((void *)(va + offset));
2818 }
2819 
2820 void *
2821 moea64_mapdev(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2822 {
2823 
2824 	return moea64_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT);
2825 }
2826 
2827 void
2828 moea64_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2829 {
2830 	vm_offset_t base, offset;
2831 
2832 	base = trunc_page(va);
2833 	offset = va & PAGE_MASK;
2834 	size = roundup2(offset + size, PAGE_SIZE);
2835 
2836 	kva_free(base, size);
2837 }
2838 
2839 void
2840 moea64_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2841 {
2842 	struct pvo_entry *pvo;
2843 	vm_offset_t lim;
2844 	vm_paddr_t pa;
2845 	vm_size_t len;
2846 
2847 	if (__predict_false(pm == NULL))
2848 		pm = &curthread->td_proc->p_vmspace->vm_pmap;
2849 
2850 	PMAP_LOCK(pm);
2851 	while (sz > 0) {
2852 		lim = round_page(va+1);
2853 		len = MIN(lim - va, sz);
2854 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
2855 		if (pvo != NULL && !(pvo->pvo_pte.pa & LPTE_I)) {
2856 			pa = (pvo->pvo_pte.pa & LPTE_RPGN) | (va & ADDR_POFF);
2857 			moea64_syncicache(mmu, pm, va, pa, len);
2858 		}
2859 		va += len;
2860 		sz -= len;
2861 	}
2862 	PMAP_UNLOCK(pm);
2863 }
2864 
2865 void
2866 moea64_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va)
2867 {
2868 
2869 	*va = (void *)(uintptr_t)pa;
2870 }
2871 
2872 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
2873 
2874 void
2875 moea64_scan_init(mmu_t mmu)
2876 {
2877 	struct pvo_entry *pvo;
2878 	vm_offset_t va;
2879 	int i;
2880 
2881 	if (!do_minidump) {
2882 		/* Initialize phys. segments for dumpsys(). */
2883 		memset(&dump_map, 0, sizeof(dump_map));
2884 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
2885 		for (i = 0; i < pregions_sz; i++) {
2886 			dump_map[i].pa_start = pregions[i].mr_start;
2887 			dump_map[i].pa_size = pregions[i].mr_size;
2888 		}
2889 		return;
2890 	}
2891 
2892 	/* Virtual segments for minidumps: */
2893 	memset(&dump_map, 0, sizeof(dump_map));
2894 
2895 	/* 1st: kernel .data and .bss. */
2896 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
2897 	dump_map[0].pa_size = round_page((uintptr_t)_end) -
2898 	    dump_map[0].pa_start;
2899 
2900 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2901 	dump_map[1].pa_start = (vm_paddr_t)(uintptr_t)msgbufp->msg_ptr;
2902 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
2903 
2904 	/* 3rd: kernel VM. */
2905 	va = dump_map[1].pa_start + dump_map[1].pa_size;
2906 	/* Find start of next chunk (from va). */
2907 	while (va < virtual_end) {
2908 		/* Don't dump the buffer cache. */
2909 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
2910 			va = kmi.buffer_eva;
2911 			continue;
2912 		}
2913 		pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
2914 		if (pvo != NULL && !(pvo->pvo_vaddr & PVO_DEAD))
2915 			break;
2916 		va += PAGE_SIZE;
2917 	}
2918 	if (va < virtual_end) {
2919 		dump_map[2].pa_start = va;
2920 		va += PAGE_SIZE;
2921 		/* Find last page in chunk. */
2922 		while (va < virtual_end) {
2923 			/* Don't run into the buffer cache. */
2924 			if (va == kmi.buffer_sva)
2925 				break;
2926 			pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
2927 			if (pvo == NULL || (pvo->pvo_vaddr & PVO_DEAD))
2928 				break;
2929 			va += PAGE_SIZE;
2930 		}
2931 		dump_map[2].pa_size = va - dump_map[2].pa_start;
2932 	}
2933 }
2934 
2935