xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision ce6a89e27cd190313be39bb479880aeda4778436)
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 SYSCTL_INT(_machdep, OID_AUTO, moea64_allocated_bpvo_entries, CTLFLAG_RD,
189     &moea64_bpvo_pool_index, 0, "");
190 
191 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
192 #ifdef __powerpc64__
193 #define	NVSIDS		(NPMAPS * 16)
194 #define VSID_HASHMASK	0xffffffffUL
195 #else
196 #define NVSIDS		NPMAPS
197 #define VSID_HASHMASK	0xfffffUL
198 #endif
199 static u_int	moea64_vsid_bitmap[NVSIDS / VSID_NBPW];
200 
201 static boolean_t moea64_initialized = FALSE;
202 
203 #ifdef MOEA64_STATS
204 /*
205  * Statistics.
206  */
207 u_int	moea64_pte_valid = 0;
208 u_int	moea64_pte_overflow = 0;
209 u_int	moea64_pvo_entries = 0;
210 u_int	moea64_pvo_enter_calls = 0;
211 u_int	moea64_pvo_remove_calls = 0;
212 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_valid, CTLFLAG_RD,
213     &moea64_pte_valid, 0, "");
214 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_overflow, CTLFLAG_RD,
215     &moea64_pte_overflow, 0, "");
216 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_entries, CTLFLAG_RD,
217     &moea64_pvo_entries, 0, "");
218 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_enter_calls, CTLFLAG_RD,
219     &moea64_pvo_enter_calls, 0, "");
220 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_remove_calls, CTLFLAG_RD,
221     &moea64_pvo_remove_calls, 0, "");
222 #endif
223 
224 vm_offset_t	moea64_scratchpage_va[2];
225 struct pvo_entry *moea64_scratchpage_pvo[2];
226 struct	mtx	moea64_scratchpage_mtx;
227 
228 uint64_t 	moea64_large_page_mask = 0;
229 uint64_t	moea64_large_page_size = 0;
230 int		moea64_large_page_shift = 0;
231 
232 /*
233  * PVO calls.
234  */
235 static int	moea64_pvo_enter(mmu_t mmu, struct pvo_entry *pvo,
236 		    struct pvo_head *pvo_head, struct pvo_entry **oldpvo);
237 static void	moea64_pvo_remove_from_pmap(mmu_t mmu, struct pvo_entry *pvo);
238 static void	moea64_pvo_remove_from_page(mmu_t mmu, struct pvo_entry *pvo);
239 static void	moea64_pvo_remove_from_page_locked(mmu_t mmu,
240 		    struct pvo_entry *pvo, vm_page_t m);
241 static struct	pvo_entry *moea64_pvo_find_va(pmap_t, vm_offset_t);
242 
243 /*
244  * Utility routines.
245  */
246 static boolean_t	moea64_query_bit(mmu_t, vm_page_t, uint64_t);
247 static u_int		moea64_clear_bit(mmu_t, vm_page_t, uint64_t);
248 static void		moea64_kremove(mmu_t, vm_offset_t);
249 static void		moea64_syncicache(mmu_t, pmap_t pmap, vm_offset_t va,
250 			    vm_paddr_t pa, vm_size_t sz);
251 static void		moea64_pmap_init_qpages(void);
252 
253 /*
254  * Kernel MMU interface
255  */
256 void moea64_clear_modify(mmu_t, vm_page_t);
257 void moea64_copy_page(mmu_t, vm_page_t, vm_page_t);
258 void moea64_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
259     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
260 int moea64_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t,
261     u_int flags, int8_t psind);
262 void moea64_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
263     vm_prot_t);
264 void moea64_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
265 vm_paddr_t moea64_extract(mmu_t, pmap_t, vm_offset_t);
266 vm_page_t moea64_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
267 void moea64_init(mmu_t);
268 boolean_t moea64_is_modified(mmu_t, vm_page_t);
269 boolean_t moea64_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
270 boolean_t moea64_is_referenced(mmu_t, vm_page_t);
271 int moea64_ts_referenced(mmu_t, vm_page_t);
272 vm_offset_t moea64_map(mmu_t, vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
273 boolean_t moea64_page_exists_quick(mmu_t, pmap_t, vm_page_t);
274 void moea64_page_init(mmu_t, vm_page_t);
275 int moea64_page_wired_mappings(mmu_t, vm_page_t);
276 void moea64_pinit(mmu_t, pmap_t);
277 void moea64_pinit0(mmu_t, pmap_t);
278 void moea64_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
279 void moea64_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
280 void moea64_qremove(mmu_t, vm_offset_t, int);
281 void moea64_release(mmu_t, pmap_t);
282 void moea64_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
283 void moea64_remove_pages(mmu_t, pmap_t);
284 void moea64_remove_all(mmu_t, vm_page_t);
285 void moea64_remove_write(mmu_t, vm_page_t);
286 void moea64_unwire(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
287 void moea64_zero_page(mmu_t, vm_page_t);
288 void moea64_zero_page_area(mmu_t, vm_page_t, int, int);
289 void moea64_activate(mmu_t, struct thread *);
290 void moea64_deactivate(mmu_t, struct thread *);
291 void *moea64_mapdev(mmu_t, vm_paddr_t, vm_size_t);
292 void *moea64_mapdev_attr(mmu_t, vm_paddr_t, vm_size_t, vm_memattr_t);
293 void moea64_unmapdev(mmu_t, vm_offset_t, vm_size_t);
294 vm_paddr_t moea64_kextract(mmu_t, vm_offset_t);
295 void moea64_page_set_memattr(mmu_t, vm_page_t m, vm_memattr_t ma);
296 void moea64_kenter_attr(mmu_t, vm_offset_t, vm_paddr_t, vm_memattr_t ma);
297 void moea64_kenter(mmu_t, vm_offset_t, vm_paddr_t);
298 boolean_t moea64_dev_direct_mapped(mmu_t, vm_paddr_t, vm_size_t);
299 static void moea64_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
300 void moea64_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz,
301     void **va);
302 void moea64_scan_init(mmu_t mmu);
303 vm_offset_t moea64_quick_enter_page(mmu_t mmu, vm_page_t m);
304 void moea64_quick_remove_page(mmu_t mmu, vm_offset_t addr);
305 static int moea64_map_user_ptr(mmu_t mmu, pmap_t pm,
306     volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen);
307 static int moea64_decode_kernel_ptr(mmu_t mmu, vm_offset_t addr,
308     int *is_user, vm_offset_t *decoded_addr);
309 static size_t moea64_scan_pmap(mmu_t mmu);
310 static void *moea64_dump_pmap_init(mmu_t mmu, unsigned blkpgs);
311 #ifdef __powerpc64__
312 static void moea64_page_array_startup(mmu_t, long);
313 #endif
314 
315 
316 static mmu_method_t moea64_methods[] = {
317 	MMUMETHOD(mmu_clear_modify,	moea64_clear_modify),
318 	MMUMETHOD(mmu_copy_page,	moea64_copy_page),
319 	MMUMETHOD(mmu_copy_pages,	moea64_copy_pages),
320 	MMUMETHOD(mmu_enter,		moea64_enter),
321 	MMUMETHOD(mmu_enter_object,	moea64_enter_object),
322 	MMUMETHOD(mmu_enter_quick,	moea64_enter_quick),
323 	MMUMETHOD(mmu_extract,		moea64_extract),
324 	MMUMETHOD(mmu_extract_and_hold,	moea64_extract_and_hold),
325 	MMUMETHOD(mmu_init,		moea64_init),
326 	MMUMETHOD(mmu_is_modified,	moea64_is_modified),
327 	MMUMETHOD(mmu_is_prefaultable,	moea64_is_prefaultable),
328 	MMUMETHOD(mmu_is_referenced,	moea64_is_referenced),
329 	MMUMETHOD(mmu_ts_referenced,	moea64_ts_referenced),
330 	MMUMETHOD(mmu_map,     		moea64_map),
331 	MMUMETHOD(mmu_page_exists_quick,moea64_page_exists_quick),
332 	MMUMETHOD(mmu_page_init,	moea64_page_init),
333 	MMUMETHOD(mmu_page_wired_mappings,moea64_page_wired_mappings),
334 	MMUMETHOD(mmu_pinit,		moea64_pinit),
335 	MMUMETHOD(mmu_pinit0,		moea64_pinit0),
336 	MMUMETHOD(mmu_protect,		moea64_protect),
337 	MMUMETHOD(mmu_qenter,		moea64_qenter),
338 	MMUMETHOD(mmu_qremove,		moea64_qremove),
339 	MMUMETHOD(mmu_release,		moea64_release),
340 	MMUMETHOD(mmu_remove,		moea64_remove),
341 	MMUMETHOD(mmu_remove_pages,	moea64_remove_pages),
342 	MMUMETHOD(mmu_remove_all,      	moea64_remove_all),
343 	MMUMETHOD(mmu_remove_write,	moea64_remove_write),
344 	MMUMETHOD(mmu_sync_icache,	moea64_sync_icache),
345 	MMUMETHOD(mmu_unwire,		moea64_unwire),
346 	MMUMETHOD(mmu_zero_page,       	moea64_zero_page),
347 	MMUMETHOD(mmu_zero_page_area,	moea64_zero_page_area),
348 	MMUMETHOD(mmu_activate,		moea64_activate),
349 	MMUMETHOD(mmu_deactivate,      	moea64_deactivate),
350 	MMUMETHOD(mmu_page_set_memattr,	moea64_page_set_memattr),
351 	MMUMETHOD(mmu_quick_enter_page, moea64_quick_enter_page),
352 	MMUMETHOD(mmu_quick_remove_page, moea64_quick_remove_page),
353 #ifdef __powerpc64__
354 	MMUMETHOD(mmu_page_array_startup,	moea64_page_array_startup),
355 #endif
356 
357 	/* Internal interfaces */
358 	MMUMETHOD(mmu_mapdev,		moea64_mapdev),
359 	MMUMETHOD(mmu_mapdev_attr,	moea64_mapdev_attr),
360 	MMUMETHOD(mmu_unmapdev,		moea64_unmapdev),
361 	MMUMETHOD(mmu_kextract,		moea64_kextract),
362 	MMUMETHOD(mmu_kenter,		moea64_kenter),
363 	MMUMETHOD(mmu_kenter_attr,	moea64_kenter_attr),
364 	MMUMETHOD(mmu_dev_direct_mapped,moea64_dev_direct_mapped),
365 	MMUMETHOD(mmu_scan_init,	moea64_scan_init),
366 	MMUMETHOD(mmu_scan_pmap,	moea64_scan_pmap),
367 	MMUMETHOD(mmu_dump_pmap_init,   moea64_dump_pmap_init),
368 	MMUMETHOD(mmu_dumpsys_map,	moea64_dumpsys_map),
369 	MMUMETHOD(mmu_map_user_ptr,	moea64_map_user_ptr),
370 	MMUMETHOD(mmu_decode_kernel_ptr, moea64_decode_kernel_ptr),
371 
372 	{ 0, 0 }
373 };
374 
375 MMU_DEF(oea64_mmu, "mmu_oea64_base", moea64_methods, 0);
376 
377 static struct pvo_head *
378 vm_page_to_pvoh(vm_page_t m)
379 {
380 
381 	mtx_assert(PV_LOCKPTR(VM_PAGE_TO_PHYS(m)), MA_OWNED);
382 	return (&m->md.mdpg_pvoh);
383 }
384 
385 static struct pvo_entry *
386 alloc_pvo_entry(int bootstrap)
387 {
388 	struct pvo_entry *pvo;
389 
390 	if (!moea64_initialized || bootstrap) {
391 		if (moea64_bpvo_pool_index >= moea64_bpvo_pool_size) {
392 			panic("%s: bpvo pool exhausted, index=%d, size=%d, bytes=%zd."
393 			    "Try setting machdep.moea64_bpvo_pool_size tunable",
394 			    __func__, moea64_bpvo_pool_index,
395 			    moea64_bpvo_pool_size,
396 			    moea64_bpvo_pool_size * sizeof(struct pvo_entry));
397 		}
398 		pvo = &moea64_bpvo_pool[
399 		    atomic_fetchadd_int(&moea64_bpvo_pool_index, 1)];
400 		bzero(pvo, sizeof(*pvo));
401 		pvo->pvo_vaddr = PVO_BOOTSTRAP;
402 	} else
403 		pvo = uma_zalloc(moea64_pvo_zone, M_NOWAIT | M_ZERO);
404 
405 	return (pvo);
406 }
407 
408 
409 static void
410 init_pvo_entry(struct pvo_entry *pvo, pmap_t pmap, vm_offset_t va)
411 {
412 	uint64_t vsid;
413 	uint64_t hash;
414 	int shift;
415 
416 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
417 
418 	pvo->pvo_pmap = pmap;
419 	va &= ~ADDR_POFF;
420 	pvo->pvo_vaddr |= va;
421 	vsid = va_to_vsid(pmap, va);
422 	pvo->pvo_vpn = (uint64_t)((va & ADDR_PIDX) >> ADDR_PIDX_SHFT)
423 	    | (vsid << 16);
424 
425 	shift = (pvo->pvo_vaddr & PVO_LARGE) ? moea64_large_page_shift :
426 	    ADDR_PIDX_SHFT;
427 	hash = (vsid & VSID_HASH_MASK) ^ (((uint64_t)va & ADDR_PIDX) >> shift);
428 	pvo->pvo_pte.slot = (hash & moea64_pteg_mask) << 3;
429 }
430 
431 static void
432 free_pvo_entry(struct pvo_entry *pvo)
433 {
434 
435 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
436 		uma_zfree(moea64_pvo_zone, pvo);
437 }
438 
439 void
440 moea64_pte_from_pvo(const struct pvo_entry *pvo, struct lpte *lpte)
441 {
442 
443 	lpte->pte_hi = moea64_pte_vpn_from_pvo_vpn(pvo);
444 	lpte->pte_hi |= LPTE_VALID;
445 
446 	if (pvo->pvo_vaddr & PVO_LARGE)
447 		lpte->pte_hi |= LPTE_BIG;
448 	if (pvo->pvo_vaddr & PVO_WIRED)
449 		lpte->pte_hi |= LPTE_WIRED;
450 	if (pvo->pvo_vaddr & PVO_HID)
451 		lpte->pte_hi |= LPTE_HID;
452 
453 	lpte->pte_lo = pvo->pvo_pte.pa; /* Includes WIMG bits */
454 	if (pvo->pvo_pte.prot & VM_PROT_WRITE)
455 		lpte->pte_lo |= LPTE_BW;
456 	else
457 		lpte->pte_lo |= LPTE_BR;
458 
459 	if (!(pvo->pvo_pte.prot & VM_PROT_EXECUTE))
460 		lpte->pte_lo |= LPTE_NOEXEC;
461 }
462 
463 static __inline uint64_t
464 moea64_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
465 {
466 	uint64_t pte_lo;
467 	int i;
468 
469 	if (ma != VM_MEMATTR_DEFAULT) {
470 		switch (ma) {
471 		case VM_MEMATTR_UNCACHEABLE:
472 			return (LPTE_I | LPTE_G);
473 		case VM_MEMATTR_CACHEABLE:
474 			return (LPTE_M);
475 		case VM_MEMATTR_WRITE_COMBINING:
476 		case VM_MEMATTR_WRITE_BACK:
477 		case VM_MEMATTR_PREFETCHABLE:
478 			return (LPTE_I);
479 		case VM_MEMATTR_WRITE_THROUGH:
480 			return (LPTE_W | LPTE_M);
481 		}
482 	}
483 
484 	/*
485 	 * Assume the page is cache inhibited and access is guarded unless
486 	 * it's in our available memory array.
487 	 */
488 	pte_lo = LPTE_I | LPTE_G;
489 	for (i = 0; i < pregions_sz; i++) {
490 		if ((pa >= pregions[i].mr_start) &&
491 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
492 			pte_lo &= ~(LPTE_I | LPTE_G);
493 			pte_lo |= LPTE_M;
494 			break;
495 		}
496 	}
497 
498 	return pte_lo;
499 }
500 
501 /*
502  * Quick sort callout for comparing memory regions.
503  */
504 static int	om_cmp(const void *a, const void *b);
505 
506 static int
507 om_cmp(const void *a, const void *b)
508 {
509 	const struct	ofw_map *mapa;
510 	const struct	ofw_map *mapb;
511 
512 	mapa = a;
513 	mapb = b;
514 	if (mapa->om_pa < mapb->om_pa)
515 		return (-1);
516 	else if (mapa->om_pa > mapb->om_pa)
517 		return (1);
518 	else
519 		return (0);
520 }
521 
522 static void
523 moea64_add_ofw_mappings(mmu_t mmup, phandle_t mmu, size_t sz)
524 {
525 	struct ofw_map	translations[sz/(4*sizeof(cell_t))]; /*>= 4 cells per */
526 	pcell_t		acells, trans_cells[sz/sizeof(cell_t)];
527 	struct pvo_entry *pvo;
528 	register_t	msr;
529 	vm_offset_t	off;
530 	vm_paddr_t	pa_base;
531 	int		i, j;
532 
533 	bzero(translations, sz);
534 	OF_getencprop(OF_finddevice("/"), "#address-cells", &acells,
535 	    sizeof(acells));
536 	if (OF_getencprop(mmu, "translations", trans_cells, sz) == -1)
537 		panic("moea64_bootstrap: can't get ofw translations");
538 
539 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
540 	sz /= sizeof(cell_t);
541 	for (i = 0, j = 0; i < sz; j++) {
542 		translations[j].om_va = trans_cells[i++];
543 		translations[j].om_len = trans_cells[i++];
544 		translations[j].om_pa = trans_cells[i++];
545 		if (acells == 2) {
546 			translations[j].om_pa <<= 32;
547 			translations[j].om_pa |= trans_cells[i++];
548 		}
549 		translations[j].om_mode = trans_cells[i++];
550 	}
551 	KASSERT(i == sz, ("Translations map has incorrect cell count (%d/%zd)",
552 	    i, sz));
553 
554 	sz = j;
555 	qsort(translations, sz, sizeof (*translations), om_cmp);
556 
557 	for (i = 0; i < sz; i++) {
558 		pa_base = translations[i].om_pa;
559 	      #ifndef __powerpc64__
560 		if ((translations[i].om_pa >> 32) != 0)
561 			panic("OFW translations above 32-bit boundary!");
562 	      #endif
563 
564 		if (pa_base % PAGE_SIZE)
565 			panic("OFW translation not page-aligned (phys)!");
566 		if (translations[i].om_va % PAGE_SIZE)
567 			panic("OFW translation not page-aligned (virt)!");
568 
569 		CTR3(KTR_PMAP, "translation: pa=%#zx va=%#x len=%#x",
570 		    pa_base, translations[i].om_va, translations[i].om_len);
571 
572 		/* Now enter the pages for this mapping */
573 
574 		DISABLE_TRANS(msr);
575 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
576 			/* If this address is direct-mapped, skip remapping */
577 			if (hw_direct_map &&
578 			    translations[i].om_va == PHYS_TO_DMAP(pa_base) &&
579 			    moea64_calc_wimg(pa_base + off, VM_MEMATTR_DEFAULT)
580  			    == LPTE_M)
581 				continue;
582 
583 			PMAP_LOCK(kernel_pmap);
584 			pvo = moea64_pvo_find_va(kernel_pmap,
585 			    translations[i].om_va + off);
586 			PMAP_UNLOCK(kernel_pmap);
587 			if (pvo != NULL)
588 				continue;
589 
590 			moea64_kenter(mmup, translations[i].om_va + off,
591 			    pa_base + off);
592 		}
593 		ENABLE_TRANS(msr);
594 	}
595 }
596 
597 #ifdef __powerpc64__
598 static void
599 moea64_probe_large_page(void)
600 {
601 	uint16_t pvr = mfpvr() >> 16;
602 
603 	switch (pvr) {
604 	case IBM970:
605 	case IBM970FX:
606 	case IBM970MP:
607 		powerpc_sync(); isync();
608 		mtspr(SPR_HID4, mfspr(SPR_HID4) & ~HID4_970_DISABLE_LG_PG);
609 		powerpc_sync(); isync();
610 
611 		/* FALLTHROUGH */
612 	default:
613 		if (moea64_large_page_size == 0) {
614 			moea64_large_page_size = 0x1000000; /* 16 MB */
615 			moea64_large_page_shift = 24;
616 		}
617 	}
618 
619 	moea64_large_page_mask = moea64_large_page_size - 1;
620 }
621 
622 static void
623 moea64_bootstrap_slb_prefault(vm_offset_t va, int large)
624 {
625 	struct slb *cache;
626 	struct slb entry;
627 	uint64_t esid, slbe;
628 	uint64_t i;
629 
630 	cache = PCPU_GET(aim.slb);
631 	esid = va >> ADDR_SR_SHFT;
632 	slbe = (esid << SLBE_ESID_SHIFT) | SLBE_VALID;
633 
634 	for (i = 0; i < 64; i++) {
635 		if (cache[i].slbe == (slbe | i))
636 			return;
637 	}
638 
639 	entry.slbe = slbe;
640 	entry.slbv = KERNEL_VSID(esid) << SLBV_VSID_SHIFT;
641 	if (large)
642 		entry.slbv |= SLBV_L;
643 
644 	slb_insert_kernel(entry.slbe, entry.slbv);
645 }
646 #endif
647 
648 static int
649 moea64_kenter_large(mmu_t mmup, vm_offset_t va, vm_paddr_t pa, uint64_t attr, int bootstrap)
650 {
651 	struct pvo_entry *pvo;
652 	uint64_t pte_lo;
653 	int error;
654 
655 	pte_lo = LPTE_M;
656 	pte_lo |= attr;
657 
658 	pvo = alloc_pvo_entry(bootstrap);
659 	pvo->pvo_vaddr |= PVO_WIRED | PVO_LARGE;
660 	init_pvo_entry(pvo, kernel_pmap, va);
661 
662 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE |
663 	    VM_PROT_EXECUTE;
664 	pvo->pvo_pte.pa = pa | pte_lo;
665 	error = moea64_pvo_enter(mmup, pvo, NULL, NULL);
666 	if (error != 0)
667 		panic("Error %d inserting large page\n", error);
668 	return (0);
669 }
670 
671 static void
672 moea64_setup_direct_map(mmu_t mmup, vm_offset_t kernelstart,
673     vm_offset_t kernelend)
674 {
675 	register_t msr;
676 	vm_paddr_t pa, pkernelstart, pkernelend;
677 	vm_offset_t size, off;
678 	uint64_t pte_lo;
679 	int i;
680 
681 	if (moea64_large_page_size == 0)
682 		hw_direct_map = 0;
683 
684 	DISABLE_TRANS(msr);
685 	if (hw_direct_map) {
686 		PMAP_LOCK(kernel_pmap);
687 		for (i = 0; i < pregions_sz; i++) {
688 		  for (pa = pregions[i].mr_start; pa < pregions[i].mr_start +
689 		     pregions[i].mr_size; pa += moea64_large_page_size) {
690 			pte_lo = LPTE_M;
691 			if (pa & moea64_large_page_mask) {
692 				pa &= moea64_large_page_mask;
693 				pte_lo |= LPTE_G;
694 			}
695 			if (pa + moea64_large_page_size >
696 			    pregions[i].mr_start + pregions[i].mr_size)
697 				pte_lo |= LPTE_G;
698 
699 			moea64_kenter_large(mmup, PHYS_TO_DMAP(pa), pa, pte_lo, 1);
700 		  }
701 		}
702 		PMAP_UNLOCK(kernel_pmap);
703 	}
704 
705 	/*
706 	 * Make sure the kernel and BPVO pool stay mapped on systems either
707 	 * without a direct map or on which the kernel is not already executing
708 	 * out of the direct-mapped region.
709 	 */
710 	if (kernelstart < DMAP_BASE_ADDRESS) {
711 		/*
712 		 * For pre-dmap execution, we need to use identity mapping
713 		 * because we will be operating with the mmu on but in the
714 		 * wrong address configuration until we __restartkernel().
715 		 */
716 		for (pa = kernelstart & ~PAGE_MASK; pa < kernelend;
717 		    pa += PAGE_SIZE)
718 			moea64_kenter(mmup, pa, pa);
719 	} else if (!hw_direct_map) {
720 		pkernelstart = kernelstart & ~DMAP_BASE_ADDRESS;
721 		pkernelend = kernelend & ~DMAP_BASE_ADDRESS;
722 		for (pa = pkernelstart & ~PAGE_MASK; pa < pkernelend;
723 		    pa += PAGE_SIZE)
724 			moea64_kenter(mmup, pa | DMAP_BASE_ADDRESS, pa);
725 	}
726 
727 	if (!hw_direct_map) {
728 		size = moea64_bpvo_pool_size*sizeof(struct pvo_entry);
729 		off = (vm_offset_t)(moea64_bpvo_pool);
730 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
731 			moea64_kenter(mmup, pa, pa);
732 
733 		/* Map exception vectors */
734 		for (pa = EXC_RSVD; pa < EXC_LAST; pa += PAGE_SIZE)
735 			moea64_kenter(mmup, pa | DMAP_BASE_ADDRESS, pa);
736 	}
737 	ENABLE_TRANS(msr);
738 
739 	/*
740 	 * Allow user to override unmapped_buf_allowed for testing.
741 	 * XXXKIB Only direct map implementation was tested.
742 	 */
743 	if (!TUNABLE_INT_FETCH("vfs.unmapped_buf_allowed",
744 	    &unmapped_buf_allowed))
745 		unmapped_buf_allowed = hw_direct_map;
746 }
747 
748 /* Quick sort callout for comparing physical addresses. */
749 static int
750 pa_cmp(const void *a, const void *b)
751 {
752 	const vm_paddr_t *pa = a, *pb = b;
753 
754 	if (*pa < *pb)
755 		return (-1);
756 	else if (*pa > *pb)
757 		return (1);
758 	else
759 		return (0);
760 }
761 
762 void
763 moea64_early_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
764 {
765 	int		i, j;
766 	vm_size_t	physsz, hwphyssz;
767 	vm_paddr_t	kernelphysstart, kernelphysend;
768 	int		rm_pavail;
769 
770 #ifndef __powerpc64__
771 	/* We don't have a direct map since there is no BAT */
772 	hw_direct_map = 0;
773 
774 	/* Make sure battable is zero, since we have no BAT */
775 	for (i = 0; i < 16; i++) {
776 		battable[i].batu = 0;
777 		battable[i].batl = 0;
778 	}
779 #else
780 	moea64_probe_large_page();
781 
782 	/* Use a direct map if we have large page support */
783 	if (moea64_large_page_size > 0)
784 		hw_direct_map = 1;
785 	else
786 		hw_direct_map = 0;
787 
788 	/* Install trap handlers for SLBs */
789 	bcopy(&slbtrap, (void *)EXC_DSE,(size_t)&slbtrapend - (size_t)&slbtrap);
790 	bcopy(&slbtrap, (void *)EXC_ISE,(size_t)&slbtrapend - (size_t)&slbtrap);
791 	__syncicache((void *)EXC_DSE, 0x80);
792 	__syncicache((void *)EXC_ISE, 0x80);
793 #endif
794 
795 	kernelphysstart = kernelstart & ~DMAP_BASE_ADDRESS;
796 	kernelphysend = kernelend & ~DMAP_BASE_ADDRESS;
797 
798 	/* Get physical memory regions from firmware */
799 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
800 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
801 
802 	if (PHYS_AVAIL_ENTRIES < regions_sz)
803 		panic("moea64_bootstrap: phys_avail too small");
804 
805 	phys_avail_count = 0;
806 	physsz = 0;
807 	hwphyssz = 0;
808 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
809 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
810 		CTR3(KTR_PMAP, "region: %#zx - %#zx (%#zx)",
811 		    regions[i].mr_start, regions[i].mr_start +
812 		    regions[i].mr_size, regions[i].mr_size);
813 		if (hwphyssz != 0 &&
814 		    (physsz + regions[i].mr_size) >= hwphyssz) {
815 			if (physsz < hwphyssz) {
816 				phys_avail[j] = regions[i].mr_start;
817 				phys_avail[j + 1] = regions[i].mr_start +
818 				    hwphyssz - physsz;
819 				physsz = hwphyssz;
820 				phys_avail_count++;
821 				dump_avail[j] = phys_avail[j];
822 				dump_avail[j + 1] = phys_avail[j + 1];
823 			}
824 			break;
825 		}
826 		phys_avail[j] = regions[i].mr_start;
827 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
828 		phys_avail_count++;
829 		physsz += regions[i].mr_size;
830 		dump_avail[j] = phys_avail[j];
831 		dump_avail[j + 1] = phys_avail[j + 1];
832 	}
833 
834 	/* Check for overlap with the kernel and exception vectors */
835 	rm_pavail = 0;
836 	for (j = 0; j < 2*phys_avail_count; j+=2) {
837 		if (phys_avail[j] < EXC_LAST)
838 			phys_avail[j] += EXC_LAST;
839 
840 		if (phys_avail[j] >= kernelphysstart &&
841 		    phys_avail[j+1] <= kernelphysend) {
842 			phys_avail[j] = phys_avail[j+1] = ~0;
843 			rm_pavail++;
844 			continue;
845 		}
846 
847 		if (kernelphysstart >= phys_avail[j] &&
848 		    kernelphysstart < phys_avail[j+1]) {
849 			if (kernelphysend < phys_avail[j+1]) {
850 				phys_avail[2*phys_avail_count] =
851 				    (kernelphysend & ~PAGE_MASK) + PAGE_SIZE;
852 				phys_avail[2*phys_avail_count + 1] =
853 				    phys_avail[j+1];
854 				phys_avail_count++;
855 			}
856 
857 			phys_avail[j+1] = kernelphysstart & ~PAGE_MASK;
858 		}
859 
860 		if (kernelphysend >= phys_avail[j] &&
861 		    kernelphysend < phys_avail[j+1]) {
862 			if (kernelphysstart > phys_avail[j]) {
863 				phys_avail[2*phys_avail_count] = phys_avail[j];
864 				phys_avail[2*phys_avail_count + 1] =
865 				    kernelphysstart & ~PAGE_MASK;
866 				phys_avail_count++;
867 			}
868 
869 			phys_avail[j] = (kernelphysend & ~PAGE_MASK) +
870 			    PAGE_SIZE;
871 		}
872 	}
873 
874 	/* Remove physical available regions marked for removal (~0) */
875 	if (rm_pavail) {
876 		qsort(phys_avail, 2*phys_avail_count, sizeof(phys_avail[0]),
877 			pa_cmp);
878 		phys_avail_count -= rm_pavail;
879 		for (i = 2*phys_avail_count;
880 		     i < 2*(phys_avail_count + rm_pavail); i+=2)
881 			phys_avail[i] = phys_avail[i+1] = 0;
882 	}
883 
884 	physmem = btoc(physsz);
885 
886 #ifdef PTEGCOUNT
887 	moea64_pteg_count = PTEGCOUNT;
888 #else
889 	moea64_pteg_count = 0x1000;
890 
891 	while (moea64_pteg_count < physmem)
892 		moea64_pteg_count <<= 1;
893 
894 	moea64_pteg_count >>= 1;
895 #endif /* PTEGCOUNT */
896 }
897 
898 void
899 moea64_mid_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
900 {
901 	int		i;
902 
903 	/*
904 	 * Set PTEG mask
905 	 */
906 	moea64_pteg_mask = moea64_pteg_count - 1;
907 
908 	/*
909 	 * Initialize SLB table lock and page locks
910 	 */
911 	mtx_init(&moea64_slb_mutex, "SLB table", NULL, MTX_DEF);
912 	for (i = 0; i < PV_LOCK_COUNT; i++)
913 		mtx_init(&pv_lock[i], "page pv", NULL, MTX_DEF);
914 
915 	/*
916 	 * Initialise the bootstrap pvo pool.
917 	 */
918 	TUNABLE_INT_FETCH("machdep.moea64_bpvo_pool_size", &moea64_bpvo_pool_size);
919 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
920 		moea64_bpvo_pool_size*sizeof(struct pvo_entry), PAGE_SIZE);
921 	moea64_bpvo_pool_index = 0;
922 
923 	/* Place at address usable through the direct map */
924 	if (hw_direct_map)
925 		moea64_bpvo_pool = (struct pvo_entry *)
926 		    PHYS_TO_DMAP((uintptr_t)moea64_bpvo_pool);
927 
928 	/*
929 	 * Make sure kernel vsid is allocated as well as VSID 0.
930 	 */
931 	#ifndef __powerpc64__
932 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NVSIDS - 1)) / VSID_NBPW]
933 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
934 	moea64_vsid_bitmap[0] |= 1;
935 	#endif
936 
937 	/*
938 	 * Initialize the kernel pmap (which is statically allocated).
939 	 */
940 	#ifdef __powerpc64__
941 	for (i = 0; i < 64; i++) {
942 		pcpup->pc_aim.slb[i].slbv = 0;
943 		pcpup->pc_aim.slb[i].slbe = 0;
944 	}
945 	#else
946 	for (i = 0; i < 16; i++)
947 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
948 	#endif
949 
950 	kernel_pmap->pmap_phys = kernel_pmap;
951 	CPU_FILL(&kernel_pmap->pm_active);
952 	RB_INIT(&kernel_pmap->pmap_pvo);
953 
954 	PMAP_LOCK_INIT(kernel_pmap);
955 
956 	/*
957 	 * Now map in all the other buffers we allocated earlier
958 	 */
959 
960 	moea64_setup_direct_map(mmup, kernelstart, kernelend);
961 }
962 
963 void
964 moea64_late_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
965 {
966 	ihandle_t	mmui;
967 	phandle_t	chosen;
968 	phandle_t	mmu;
969 	ssize_t		sz;
970 	int		i;
971 	vm_offset_t	pa, va;
972 	void		*dpcpu;
973 
974 	/*
975 	 * Set up the Open Firmware pmap and add its mappings if not in real
976 	 * mode.
977 	 */
978 
979 	chosen = OF_finddevice("/chosen");
980 	if (chosen != -1 && OF_getencprop(chosen, "mmu", &mmui, 4) != -1) {
981 		mmu = OF_instance_to_package(mmui);
982 		if (mmu == -1 ||
983 		    (sz = OF_getproplen(mmu, "translations")) == -1)
984 			sz = 0;
985 		if (sz > 6144 /* tmpstksz - 2 KB headroom */)
986 			panic("moea64_bootstrap: too many ofw translations");
987 
988 		if (sz > 0)
989 			moea64_add_ofw_mappings(mmup, mmu, sz);
990 	}
991 
992 	/*
993 	 * Calculate the last available physical address.
994 	 */
995 	Maxmem = 0;
996 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
997 		Maxmem = MAX(Maxmem, powerpc_btop(phys_avail[i + 1]));
998 
999 	/*
1000 	 * Initialize MMU.
1001 	 */
1002 	MMU_CPU_BOOTSTRAP(mmup,0);
1003 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
1004 	pmap_bootstrapped++;
1005 
1006 	/*
1007 	 * Set the start and end of kva.
1008 	 */
1009 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
1010 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
1011 
1012 	/*
1013 	 * Map the entire KVA range into the SLB. We must not fault there.
1014 	 */
1015 	#ifdef __powerpc64__
1016 	for (va = virtual_avail; va < virtual_end; va += SEGMENT_LENGTH)
1017 		moea64_bootstrap_slb_prefault(va, 0);
1018 	#endif
1019 
1020 	/*
1021 	 * Remap any early IO mappings (console framebuffer, etc.)
1022 	 */
1023 	bs_remap_earlyboot();
1024 
1025 	/*
1026 	 * Figure out how far we can extend virtual_end into segment 16
1027 	 * without running into existing mappings. Segment 16 is guaranteed
1028 	 * to contain neither RAM nor devices (at least on Apple hardware),
1029 	 * but will generally contain some OFW mappings we should not
1030 	 * step on.
1031 	 */
1032 
1033 	#ifndef __powerpc64__	/* KVA is in high memory on PPC64 */
1034 	PMAP_LOCK(kernel_pmap);
1035 	while (virtual_end < VM_MAX_KERNEL_ADDRESS &&
1036 	    moea64_pvo_find_va(kernel_pmap, virtual_end+1) == NULL)
1037 		virtual_end += PAGE_SIZE;
1038 	PMAP_UNLOCK(kernel_pmap);
1039 	#endif
1040 
1041 	/*
1042 	 * Allocate a kernel stack with a guard page for thread0 and map it
1043 	 * into the kernel page map.
1044 	 */
1045 	pa = moea64_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
1046 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
1047 	virtual_avail = va + kstack_pages * PAGE_SIZE;
1048 	CTR2(KTR_PMAP, "moea64_bootstrap: kstack0 at %#x (%#x)", pa, va);
1049 	thread0.td_kstack = va;
1050 	thread0.td_kstack_pages = kstack_pages;
1051 	for (i = 0; i < kstack_pages; i++) {
1052 		moea64_kenter(mmup, va, pa);
1053 		pa += PAGE_SIZE;
1054 		va += PAGE_SIZE;
1055 	}
1056 
1057 	/*
1058 	 * Allocate virtual address space for the message buffer.
1059 	 */
1060 	pa = msgbuf_phys = moea64_bootstrap_alloc(msgbufsize, PAGE_SIZE);
1061 	msgbufp = (struct msgbuf *)virtual_avail;
1062 	va = virtual_avail;
1063 	virtual_avail += round_page(msgbufsize);
1064 	while (va < virtual_avail) {
1065 		moea64_kenter(mmup, va, pa);
1066 		pa += PAGE_SIZE;
1067 		va += PAGE_SIZE;
1068 	}
1069 
1070 	/*
1071 	 * Allocate virtual address space for the dynamic percpu area.
1072 	 */
1073 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
1074 	dpcpu = (void *)virtual_avail;
1075 	va = virtual_avail;
1076 	virtual_avail += DPCPU_SIZE;
1077 	while (va < virtual_avail) {
1078 		moea64_kenter(mmup, va, pa);
1079 		pa += PAGE_SIZE;
1080 		va += PAGE_SIZE;
1081 	}
1082 	dpcpu_init(dpcpu, curcpu);
1083 
1084 	crashdumpmap = (caddr_t)virtual_avail;
1085 	virtual_avail += MAXDUMPPGS * PAGE_SIZE;
1086 
1087 	/*
1088 	 * Allocate some things for page zeroing. We put this directly
1089 	 * in the page table and use MOEA64_PTE_REPLACE to avoid any
1090 	 * of the PVO book-keeping or other parts of the VM system
1091 	 * from even knowing that this hack exists.
1092 	 */
1093 
1094 	if (!hw_direct_map) {
1095 		mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL,
1096 		    MTX_DEF);
1097 		for (i = 0; i < 2; i++) {
1098 			moea64_scratchpage_va[i] = (virtual_end+1) - PAGE_SIZE;
1099 			virtual_end -= PAGE_SIZE;
1100 
1101 			moea64_kenter(mmup, moea64_scratchpage_va[i], 0);
1102 
1103 			PMAP_LOCK(kernel_pmap);
1104 			moea64_scratchpage_pvo[i] = moea64_pvo_find_va(
1105 			    kernel_pmap, (vm_offset_t)moea64_scratchpage_va[i]);
1106 			PMAP_UNLOCK(kernel_pmap);
1107 		}
1108 	}
1109 
1110 	numa_mem_regions(&numa_pregions, &numapregions_sz);
1111 }
1112 
1113 static void
1114 moea64_pmap_init_qpages(void)
1115 {
1116 	struct pcpu *pc;
1117 	int i;
1118 
1119 	if (hw_direct_map)
1120 		return;
1121 
1122 	CPU_FOREACH(i) {
1123 		pc = pcpu_find(i);
1124 		pc->pc_qmap_addr = kva_alloc(PAGE_SIZE);
1125 		if (pc->pc_qmap_addr == 0)
1126 			panic("pmap_init_qpages: unable to allocate KVA");
1127 		PMAP_LOCK(kernel_pmap);
1128 		pc->pc_aim.qmap_pvo =
1129 		    moea64_pvo_find_va(kernel_pmap, pc->pc_qmap_addr);
1130 		PMAP_UNLOCK(kernel_pmap);
1131 		mtx_init(&pc->pc_aim.qmap_lock, "qmap lock", NULL, MTX_DEF);
1132 	}
1133 }
1134 
1135 SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, moea64_pmap_init_qpages, NULL);
1136 
1137 /*
1138  * Activate a user pmap.  This mostly involves setting some non-CPU
1139  * state.
1140  */
1141 void
1142 moea64_activate(mmu_t mmu, struct thread *td)
1143 {
1144 	pmap_t	pm;
1145 
1146 	pm = &td->td_proc->p_vmspace->vm_pmap;
1147 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1148 
1149 	#ifdef __powerpc64__
1150 	PCPU_SET(aim.userslb, pm->pm_slb);
1151 	__asm __volatile("slbmte %0, %1; isync" ::
1152 	    "r"(td->td_pcb->pcb_cpu.aim.usr_vsid), "r"(USER_SLB_SLBE));
1153 	#else
1154 	PCPU_SET(curpmap, pm->pmap_phys);
1155 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1156 	#endif
1157 }
1158 
1159 void
1160 moea64_deactivate(mmu_t mmu, struct thread *td)
1161 {
1162 	pmap_t	pm;
1163 
1164 	__asm __volatile("isync; slbie %0" :: "r"(USER_ADDR));
1165 
1166 	pm = &td->td_proc->p_vmspace->vm_pmap;
1167 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1168 	#ifdef __powerpc64__
1169 	PCPU_SET(aim.userslb, NULL);
1170 	#else
1171 	PCPU_SET(curpmap, NULL);
1172 	#endif
1173 }
1174 
1175 void
1176 moea64_unwire(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1177 {
1178 	struct	pvo_entry key, *pvo;
1179 	vm_page_t m;
1180 	int64_t	refchg;
1181 
1182 	key.pvo_vaddr = sva;
1183 	PMAP_LOCK(pm);
1184 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1185 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1186 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1187 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1188 			panic("moea64_unwire: pvo %p is missing PVO_WIRED",
1189 			    pvo);
1190 		pvo->pvo_vaddr &= ~PVO_WIRED;
1191 		refchg = MOEA64_PTE_REPLACE(mmu, pvo, 0 /* No invalidation */);
1192 		if ((pvo->pvo_vaddr & PVO_MANAGED) &&
1193 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1194 			if (refchg < 0)
1195 				refchg = LPTE_CHG;
1196 			m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
1197 
1198 			refchg |= atomic_readandclear_32(&m->md.mdpg_attrs);
1199 			if (refchg & LPTE_CHG)
1200 				vm_page_dirty(m);
1201 			if (refchg & LPTE_REF)
1202 				vm_page_aflag_set(m, PGA_REFERENCED);
1203 		}
1204 		pm->pm_stats.wired_count--;
1205 	}
1206 	PMAP_UNLOCK(pm);
1207 }
1208 
1209 /*
1210  * This goes through and sets the physical address of our
1211  * special scratch PTE to the PA we want to zero or copy. Because
1212  * of locking issues (this can get called in pvo_enter() by
1213  * the UMA allocator), we can't use most other utility functions here
1214  */
1215 
1216 static __inline
1217 void moea64_set_scratchpage_pa(mmu_t mmup, int which, vm_paddr_t pa)
1218 {
1219 	struct pvo_entry *pvo;
1220 
1221 	KASSERT(!hw_direct_map, ("Using OEA64 scratchpage with a direct map!"));
1222 	mtx_assert(&moea64_scratchpage_mtx, MA_OWNED);
1223 
1224 	pvo = moea64_scratchpage_pvo[which];
1225 	PMAP_LOCK(pvo->pvo_pmap);
1226 	pvo->pvo_pte.pa =
1227 	    moea64_calc_wimg(pa, VM_MEMATTR_DEFAULT) | (uint64_t)pa;
1228 	MOEA64_PTE_REPLACE(mmup, pvo, MOEA64_PTE_INVALIDATE);
1229 	PMAP_UNLOCK(pvo->pvo_pmap);
1230 	isync();
1231 }
1232 
1233 void
1234 moea64_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1235 {
1236 	vm_offset_t	dst;
1237 	vm_offset_t	src;
1238 
1239 	dst = VM_PAGE_TO_PHYS(mdst);
1240 	src = VM_PAGE_TO_PHYS(msrc);
1241 
1242 	if (hw_direct_map) {
1243 		bcopy((void *)PHYS_TO_DMAP(src), (void *)PHYS_TO_DMAP(dst),
1244 		    PAGE_SIZE);
1245 	} else {
1246 		mtx_lock(&moea64_scratchpage_mtx);
1247 
1248 		moea64_set_scratchpage_pa(mmu, 0, src);
1249 		moea64_set_scratchpage_pa(mmu, 1, dst);
1250 
1251 		bcopy((void *)moea64_scratchpage_va[0],
1252 		    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1253 
1254 		mtx_unlock(&moea64_scratchpage_mtx);
1255 	}
1256 }
1257 
1258 static inline void
1259 moea64_copy_pages_dmap(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1260     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1261 {
1262 	void *a_cp, *b_cp;
1263 	vm_offset_t a_pg_offset, b_pg_offset;
1264 	int cnt;
1265 
1266 	while (xfersize > 0) {
1267 		a_pg_offset = a_offset & PAGE_MASK;
1268 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1269 		a_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1270 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT])) +
1271 		    a_pg_offset;
1272 		b_pg_offset = b_offset & PAGE_MASK;
1273 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1274 		b_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1275 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT])) +
1276 		    b_pg_offset;
1277 		bcopy(a_cp, b_cp, cnt);
1278 		a_offset += cnt;
1279 		b_offset += cnt;
1280 		xfersize -= cnt;
1281 	}
1282 }
1283 
1284 static inline void
1285 moea64_copy_pages_nodmap(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 	void *a_cp, *b_cp;
1289 	vm_offset_t a_pg_offset, b_pg_offset;
1290 	int cnt;
1291 
1292 	mtx_lock(&moea64_scratchpage_mtx);
1293 	while (xfersize > 0) {
1294 		a_pg_offset = a_offset & PAGE_MASK;
1295 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1296 		moea64_set_scratchpage_pa(mmu, 0,
1297 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]));
1298 		a_cp = (char *)moea64_scratchpage_va[0] + a_pg_offset;
1299 		b_pg_offset = b_offset & PAGE_MASK;
1300 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1301 		moea64_set_scratchpage_pa(mmu, 1,
1302 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]));
1303 		b_cp = (char *)moea64_scratchpage_va[1] + b_pg_offset;
1304 		bcopy(a_cp, b_cp, cnt);
1305 		a_offset += cnt;
1306 		b_offset += cnt;
1307 		xfersize -= cnt;
1308 	}
1309 	mtx_unlock(&moea64_scratchpage_mtx);
1310 }
1311 
1312 void
1313 moea64_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1314     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1315 {
1316 
1317 	if (hw_direct_map) {
1318 		moea64_copy_pages_dmap(mmu, ma, a_offset, mb, b_offset,
1319 		    xfersize);
1320 	} else {
1321 		moea64_copy_pages_nodmap(mmu, ma, a_offset, mb, b_offset,
1322 		    xfersize);
1323 	}
1324 }
1325 
1326 void
1327 moea64_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1328 {
1329 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1330 
1331 	if (size + off > PAGE_SIZE)
1332 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1333 
1334 	if (hw_direct_map) {
1335 		bzero((caddr_t)(uintptr_t)PHYS_TO_DMAP(pa) + off, size);
1336 	} else {
1337 		mtx_lock(&moea64_scratchpage_mtx);
1338 		moea64_set_scratchpage_pa(mmu, 0, pa);
1339 		bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1340 		mtx_unlock(&moea64_scratchpage_mtx);
1341 	}
1342 }
1343 
1344 /*
1345  * Zero a page of physical memory by temporarily mapping it
1346  */
1347 void
1348 moea64_zero_page(mmu_t mmu, vm_page_t m)
1349 {
1350 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1351 	vm_offset_t va, off;
1352 
1353 	if (!hw_direct_map) {
1354 		mtx_lock(&moea64_scratchpage_mtx);
1355 
1356 		moea64_set_scratchpage_pa(mmu, 0, pa);
1357 		va = moea64_scratchpage_va[0];
1358 	} else {
1359 		va = PHYS_TO_DMAP(pa);
1360 	}
1361 
1362 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1363 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1364 
1365 	if (!hw_direct_map)
1366 		mtx_unlock(&moea64_scratchpage_mtx);
1367 }
1368 
1369 vm_offset_t
1370 moea64_quick_enter_page(mmu_t mmu, vm_page_t m)
1371 {
1372 	struct pvo_entry *pvo;
1373 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1374 
1375 	if (hw_direct_map)
1376 		return (PHYS_TO_DMAP(pa));
1377 
1378 	/*
1379  	 * MOEA64_PTE_REPLACE does some locking, so we can't just grab
1380 	 * a critical section and access the PCPU data like on i386.
1381 	 * Instead, pin the thread and grab the PCPU lock to prevent
1382 	 * a preempting thread from using the same PCPU data.
1383 	 */
1384 	sched_pin();
1385 
1386 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_NOTOWNED);
1387 	pvo = PCPU_GET(aim.qmap_pvo);
1388 
1389 	mtx_lock(PCPU_PTR(aim.qmap_lock));
1390 	pvo->pvo_pte.pa = moea64_calc_wimg(pa, pmap_page_get_memattr(m)) |
1391 	    (uint64_t)pa;
1392 	MOEA64_PTE_REPLACE(mmu, pvo, MOEA64_PTE_INVALIDATE);
1393 	isync();
1394 
1395 	return (PCPU_GET(qmap_addr));
1396 }
1397 
1398 void
1399 moea64_quick_remove_page(mmu_t mmu, vm_offset_t addr)
1400 {
1401 	if (hw_direct_map)
1402 		return;
1403 
1404 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_OWNED);
1405 	KASSERT(PCPU_GET(qmap_addr) == addr,
1406 	    ("moea64_quick_remove_page: invalid address"));
1407 	mtx_unlock(PCPU_PTR(aim.qmap_lock));
1408 	sched_unpin();
1409 }
1410 
1411 /*
1412  * Map the given physical page at the specified virtual address in the
1413  * target pmap with the protection requested.  If specified the page
1414  * will be wired down.
1415  */
1416 
1417 int
1418 moea64_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m,
1419     vm_prot_t prot, u_int flags, int8_t psind)
1420 {
1421 	struct		pvo_entry *pvo, *oldpvo;
1422 	struct		pvo_head *pvo_head;
1423 	uint64_t	pte_lo;
1424 	int		error;
1425 
1426 	if ((m->oflags & VPO_UNMANAGED) == 0) {
1427 		if ((flags & PMAP_ENTER_QUICK_LOCKED) == 0)
1428 			VM_PAGE_OBJECT_BUSY_ASSERT(m);
1429 		else
1430 			VM_OBJECT_ASSERT_LOCKED(m->object);
1431 	}
1432 
1433 	pvo = alloc_pvo_entry(0);
1434 	if (pvo == NULL)
1435 		return (KERN_RESOURCE_SHORTAGE);
1436 	pvo->pvo_pmap = NULL; /* to be filled in later */
1437 	pvo->pvo_pte.prot = prot;
1438 
1439 	pte_lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1440 	pvo->pvo_pte.pa = VM_PAGE_TO_PHYS(m) | pte_lo;
1441 
1442 	if ((flags & PMAP_ENTER_WIRED) != 0)
1443 		pvo->pvo_vaddr |= PVO_WIRED;
1444 
1445 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea64_initialized) {
1446 		pvo_head = NULL;
1447 	} else {
1448 		pvo_head = &m->md.mdpg_pvoh;
1449 		pvo->pvo_vaddr |= PVO_MANAGED;
1450 	}
1451 
1452 	PV_PAGE_LOCK(m);
1453 	PMAP_LOCK(pmap);
1454 	if (pvo->pvo_pmap == NULL)
1455 		init_pvo_entry(pvo, pmap, va);
1456 	if (prot & VM_PROT_WRITE)
1457 		if (pmap_bootstrapped &&
1458 		    (m->oflags & VPO_UNMANAGED) == 0)
1459 			vm_page_aflag_set(m, PGA_WRITEABLE);
1460 
1461 	error = moea64_pvo_enter(mmu, pvo, pvo_head, &oldpvo);
1462 	if (error == EEXIST) {
1463 		if (oldpvo->pvo_vaddr == pvo->pvo_vaddr &&
1464 		    oldpvo->pvo_pte.pa == pvo->pvo_pte.pa &&
1465 		    oldpvo->pvo_pte.prot == prot) {
1466 			/* Identical mapping already exists */
1467 			error = 0;
1468 
1469 			/* If not in page table, reinsert it */
1470 			if (MOEA64_PTE_SYNCH(mmu, oldpvo) < 0) {
1471 				STAT_MOEA64(moea64_pte_overflow--);
1472 				MOEA64_PTE_INSERT(mmu, oldpvo);
1473 			}
1474 
1475 			/* Then just clean up and go home */
1476 			PV_PAGE_UNLOCK(m);
1477 			PMAP_UNLOCK(pmap);
1478 			free_pvo_entry(pvo);
1479 			goto out;
1480 		} else {
1481 			/* Otherwise, need to kill it first */
1482 			KASSERT(oldpvo->pvo_pmap == pmap, ("pmap of old "
1483 			    "mapping does not match new mapping"));
1484 			moea64_pvo_remove_from_pmap(mmu, oldpvo);
1485 			moea64_pvo_enter(mmu, pvo, pvo_head, NULL);
1486 		}
1487 	}
1488 	PMAP_UNLOCK(pmap);
1489 	PV_PAGE_UNLOCK(m);
1490 
1491 	/* Free any dead pages */
1492 	if (error == EEXIST) {
1493 		moea64_pvo_remove_from_page(mmu, oldpvo);
1494 		free_pvo_entry(oldpvo);
1495 	}
1496 
1497 out:
1498 	/*
1499 	 * Flush the page from the instruction cache if this page is
1500 	 * mapped executable and cacheable.
1501 	 */
1502 	if (pmap != kernel_pmap && (m->a.flags & PGA_EXECUTABLE) == 0 &&
1503 	    (pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1504 		vm_page_aflag_set(m, PGA_EXECUTABLE);
1505 		moea64_syncicache(mmu, pmap, va, VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1506 	}
1507 	return (KERN_SUCCESS);
1508 }
1509 
1510 static void
1511 moea64_syncicache(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
1512     vm_size_t sz)
1513 {
1514 
1515 	/*
1516 	 * This is much trickier than on older systems because
1517 	 * we can't sync the icache on physical addresses directly
1518 	 * without a direct map. Instead we check a couple of cases
1519 	 * where the memory is already mapped in and, failing that,
1520 	 * use the same trick we use for page zeroing to create
1521 	 * a temporary mapping for this physical address.
1522 	 */
1523 
1524 	if (!pmap_bootstrapped) {
1525 		/*
1526 		 * If PMAP is not bootstrapped, we are likely to be
1527 		 * in real mode.
1528 		 */
1529 		__syncicache((void *)(uintptr_t)pa, sz);
1530 	} else if (pmap == kernel_pmap) {
1531 		__syncicache((void *)va, sz);
1532 	} else if (hw_direct_map) {
1533 		__syncicache((void *)(uintptr_t)PHYS_TO_DMAP(pa), sz);
1534 	} else {
1535 		/* Use the scratch page to set up a temp mapping */
1536 
1537 		mtx_lock(&moea64_scratchpage_mtx);
1538 
1539 		moea64_set_scratchpage_pa(mmu, 1, pa & ~ADDR_POFF);
1540 		__syncicache((void *)(moea64_scratchpage_va[1] +
1541 		    (va & ADDR_POFF)), sz);
1542 
1543 		mtx_unlock(&moea64_scratchpage_mtx);
1544 	}
1545 }
1546 
1547 /*
1548  * Maps a sequence of resident pages belonging to the same object.
1549  * The sequence begins with the given page m_start.  This page is
1550  * mapped at the given virtual address start.  Each subsequent page is
1551  * mapped at a virtual address that is offset from start by the same
1552  * amount as the page is offset from m_start within the object.  The
1553  * last page in the sequence is the page with the largest offset from
1554  * m_start that can be mapped at a virtual address less than the given
1555  * virtual address end.  Not every virtual page between start and end
1556  * is mapped; only those for which a resident page exists with the
1557  * corresponding offset from m_start are mapped.
1558  */
1559 void
1560 moea64_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1561     vm_page_t m_start, vm_prot_t prot)
1562 {
1563 	vm_page_t m;
1564 	vm_pindex_t diff, psize;
1565 
1566 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1567 
1568 	psize = atop(end - start);
1569 	m = m_start;
1570 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1571 		moea64_enter(mmu, pm, start + ptoa(diff), m, prot &
1572 		    (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_NOSLEEP |
1573 		    PMAP_ENTER_QUICK_LOCKED, 0);
1574 		m = TAILQ_NEXT(m, listq);
1575 	}
1576 }
1577 
1578 void
1579 moea64_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1580     vm_prot_t prot)
1581 {
1582 
1583 	moea64_enter(mmu, pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1584 	    PMAP_ENTER_NOSLEEP | PMAP_ENTER_QUICK_LOCKED, 0);
1585 }
1586 
1587 vm_paddr_t
1588 moea64_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1589 {
1590 	struct	pvo_entry *pvo;
1591 	vm_paddr_t pa;
1592 
1593 	PMAP_LOCK(pm);
1594 	pvo = moea64_pvo_find_va(pm, va);
1595 	if (pvo == NULL)
1596 		pa = 0;
1597 	else
1598 		pa = (pvo->pvo_pte.pa & LPTE_RPGN) | (va - PVO_VADDR(pvo));
1599 	PMAP_UNLOCK(pm);
1600 
1601 	return (pa);
1602 }
1603 
1604 /*
1605  * Atomically extract and hold the physical page with the given
1606  * pmap and virtual address pair if that mapping permits the given
1607  * protection.
1608  */
1609 vm_page_t
1610 moea64_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1611 {
1612 	struct	pvo_entry *pvo;
1613 	vm_page_t m;
1614 
1615 	m = NULL;
1616 	PMAP_LOCK(pmap);
1617 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1618 	if (pvo != NULL && (pvo->pvo_pte.prot & prot) == prot) {
1619 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
1620 		if (!vm_page_wire_mapped(m))
1621 			m = NULL;
1622 	}
1623 	PMAP_UNLOCK(pmap);
1624 	return (m);
1625 }
1626 
1627 static mmu_t installed_mmu;
1628 
1629 static void *
1630 moea64_uma_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain,
1631     uint8_t *flags, int wait)
1632 {
1633 	struct pvo_entry *pvo;
1634         vm_offset_t va;
1635         vm_page_t m;
1636         int needed_lock;
1637 
1638 	/*
1639 	 * This entire routine is a horrible hack to avoid bothering kmem
1640 	 * for new KVA addresses. Because this can get called from inside
1641 	 * kmem allocation routines, calling kmem for a new address here
1642 	 * can lead to multiply locking non-recursive mutexes.
1643 	 */
1644 
1645 	*flags = UMA_SLAB_PRIV;
1646 	needed_lock = !PMAP_LOCKED(kernel_pmap);
1647 
1648 	m = vm_page_alloc_domain(NULL, 0, domain,
1649 	    malloc2vm_flags(wait) | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1650 	if (m == NULL)
1651 		return (NULL);
1652 
1653 	va = VM_PAGE_TO_PHYS(m);
1654 
1655 	pvo = alloc_pvo_entry(1 /* bootstrap */);
1656 
1657 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE;
1658 	pvo->pvo_pte.pa = VM_PAGE_TO_PHYS(m) | LPTE_M;
1659 
1660 	if (needed_lock)
1661 		PMAP_LOCK(kernel_pmap);
1662 
1663 	init_pvo_entry(pvo, kernel_pmap, va);
1664 	pvo->pvo_vaddr |= PVO_WIRED;
1665 
1666 	moea64_pvo_enter(installed_mmu, pvo, NULL, NULL);
1667 
1668 	if (needed_lock)
1669 		PMAP_UNLOCK(kernel_pmap);
1670 
1671 	if ((wait & M_ZERO) && (m->flags & PG_ZERO) == 0)
1672                 bzero((void *)va, PAGE_SIZE);
1673 
1674 	return (void *)va;
1675 }
1676 
1677 extern int elf32_nxstack;
1678 
1679 void
1680 moea64_init(mmu_t mmu)
1681 {
1682 
1683 	CTR0(KTR_PMAP, "moea64_init");
1684 
1685 	moea64_pvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1686 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1687 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1688 
1689 	if (!hw_direct_map) {
1690 		installed_mmu = mmu;
1691 		uma_zone_set_allocf(moea64_pvo_zone, moea64_uma_page_alloc);
1692 	}
1693 
1694 #ifdef COMPAT_FREEBSD32
1695 	elf32_nxstack = 1;
1696 #endif
1697 
1698 	moea64_initialized = TRUE;
1699 }
1700 
1701 boolean_t
1702 moea64_is_referenced(mmu_t mmu, vm_page_t m)
1703 {
1704 
1705 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1706 	    ("moea64_is_referenced: page %p is not managed", m));
1707 
1708 	return (moea64_query_bit(mmu, m, LPTE_REF));
1709 }
1710 
1711 boolean_t
1712 moea64_is_modified(mmu_t mmu, vm_page_t m)
1713 {
1714 
1715 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1716 	    ("moea64_is_modified: page %p is not managed", m));
1717 
1718 	/*
1719 	 * If the page is not busied then this check is racy.
1720 	 */
1721 	if (!pmap_page_is_write_mapped(m))
1722 		return (FALSE);
1723 
1724 	return (moea64_query_bit(mmu, m, LPTE_CHG));
1725 }
1726 
1727 boolean_t
1728 moea64_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1729 {
1730 	struct pvo_entry *pvo;
1731 	boolean_t rv = TRUE;
1732 
1733 	PMAP_LOCK(pmap);
1734 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1735 	if (pvo != NULL)
1736 		rv = FALSE;
1737 	PMAP_UNLOCK(pmap);
1738 	return (rv);
1739 }
1740 
1741 void
1742 moea64_clear_modify(mmu_t mmu, vm_page_t m)
1743 {
1744 
1745 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1746 	    ("moea64_clear_modify: page %p is not managed", m));
1747 	vm_page_assert_busied(m);
1748 
1749 	if (!pmap_page_is_write_mapped(m))
1750 		return;
1751 	moea64_clear_bit(mmu, m, LPTE_CHG);
1752 }
1753 
1754 /*
1755  * Clear the write and modified bits in each of the given page's mappings.
1756  */
1757 void
1758 moea64_remove_write(mmu_t mmu, vm_page_t m)
1759 {
1760 	struct	pvo_entry *pvo;
1761 	int64_t	refchg, ret;
1762 	pmap_t	pmap;
1763 
1764 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1765 	    ("moea64_remove_write: page %p is not managed", m));
1766 	vm_page_assert_busied(m);
1767 
1768 	if (!pmap_page_is_write_mapped(m))
1769 		return
1770 
1771 	powerpc_sync();
1772 	PV_PAGE_LOCK(m);
1773 	refchg = 0;
1774 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1775 		pmap = pvo->pvo_pmap;
1776 		PMAP_LOCK(pmap);
1777 		if (!(pvo->pvo_vaddr & PVO_DEAD) &&
1778 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1779 			pvo->pvo_pte.prot &= ~VM_PROT_WRITE;
1780 			ret = MOEA64_PTE_REPLACE(mmu, pvo,
1781 			    MOEA64_PTE_PROT_UPDATE);
1782 			if (ret < 0)
1783 				ret = LPTE_CHG;
1784 			refchg |= ret;
1785 			if (pvo->pvo_pmap == kernel_pmap)
1786 				isync();
1787 		}
1788 		PMAP_UNLOCK(pmap);
1789 	}
1790 	if ((refchg | atomic_readandclear_32(&m->md.mdpg_attrs)) & LPTE_CHG)
1791 		vm_page_dirty(m);
1792 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1793 	PV_PAGE_UNLOCK(m);
1794 }
1795 
1796 /*
1797  *	moea64_ts_referenced:
1798  *
1799  *	Return a count of reference bits for a page, clearing those bits.
1800  *	It is not necessary for every reference bit to be cleared, but it
1801  *	is necessary that 0 only be returned when there are truly no
1802  *	reference bits set.
1803  *
1804  *	XXX: The exact number of bits to check and clear is a matter that
1805  *	should be tested and standardized at some point in the future for
1806  *	optimal aging of shared pages.
1807  */
1808 int
1809 moea64_ts_referenced(mmu_t mmu, vm_page_t m)
1810 {
1811 
1812 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1813 	    ("moea64_ts_referenced: page %p is not managed", m));
1814 	return (moea64_clear_bit(mmu, m, LPTE_REF));
1815 }
1816 
1817 /*
1818  * Modify the WIMG settings of all mappings for a page.
1819  */
1820 void
1821 moea64_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1822 {
1823 	struct	pvo_entry *pvo;
1824 	int64_t	refchg;
1825 	pmap_t	pmap;
1826 	uint64_t lo;
1827 
1828 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1829 		m->md.mdpg_cache_attrs = ma;
1830 		return;
1831 	}
1832 
1833 	lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1834 
1835 	PV_PAGE_LOCK(m);
1836 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1837 		pmap = pvo->pvo_pmap;
1838 		PMAP_LOCK(pmap);
1839 		if (!(pvo->pvo_vaddr & PVO_DEAD)) {
1840 			pvo->pvo_pte.pa &= ~LPTE_WIMG;
1841 			pvo->pvo_pte.pa |= lo;
1842 			refchg = MOEA64_PTE_REPLACE(mmu, pvo,
1843 			    MOEA64_PTE_INVALIDATE);
1844 			if (refchg < 0)
1845 				refchg = (pvo->pvo_pte.prot & VM_PROT_WRITE) ?
1846 				    LPTE_CHG : 0;
1847 			if ((pvo->pvo_vaddr & PVO_MANAGED) &&
1848 			    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1849 				refchg |=
1850 				    atomic_readandclear_32(&m->md.mdpg_attrs);
1851 				if (refchg & LPTE_CHG)
1852 					vm_page_dirty(m);
1853 				if (refchg & LPTE_REF)
1854 					vm_page_aflag_set(m, PGA_REFERENCED);
1855 			}
1856 			if (pvo->pvo_pmap == kernel_pmap)
1857 				isync();
1858 		}
1859 		PMAP_UNLOCK(pmap);
1860 	}
1861 	m->md.mdpg_cache_attrs = ma;
1862 	PV_PAGE_UNLOCK(m);
1863 }
1864 
1865 /*
1866  * Map a wired page into kernel virtual address space.
1867  */
1868 void
1869 moea64_kenter_attr(mmu_t mmu, vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
1870 {
1871 	int		error;
1872 	struct pvo_entry *pvo, *oldpvo;
1873 
1874 	do {
1875 		pvo = alloc_pvo_entry(0);
1876 		if (pvo == NULL)
1877 			vm_wait(NULL);
1878 	} while (pvo == NULL);
1879 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;
1880 	pvo->pvo_pte.pa = (pa & ~ADDR_POFF) | moea64_calc_wimg(pa, ma);
1881 	pvo->pvo_vaddr |= PVO_WIRED;
1882 
1883 	PMAP_LOCK(kernel_pmap);
1884 	oldpvo = moea64_pvo_find_va(kernel_pmap, va);
1885 	if (oldpvo != NULL)
1886 		moea64_pvo_remove_from_pmap(mmu, oldpvo);
1887 	init_pvo_entry(pvo, kernel_pmap, va);
1888 	error = moea64_pvo_enter(mmu, pvo, NULL, NULL);
1889 	PMAP_UNLOCK(kernel_pmap);
1890 
1891 	/* Free any dead pages */
1892 	if (oldpvo != NULL) {
1893 		moea64_pvo_remove_from_page(mmu, oldpvo);
1894 		free_pvo_entry(oldpvo);
1895 	}
1896 
1897 	if (error != 0)
1898 		panic("moea64_kenter: failed to enter va %#zx pa %#jx: %d", va,
1899 		    (uintmax_t)pa, error);
1900 }
1901 
1902 void
1903 moea64_kenter(mmu_t mmu, vm_offset_t va, vm_paddr_t pa)
1904 {
1905 
1906 	moea64_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1907 }
1908 
1909 /*
1910  * Extract the physical page address associated with the given kernel virtual
1911  * address.
1912  */
1913 vm_paddr_t
1914 moea64_kextract(mmu_t mmu, vm_offset_t va)
1915 {
1916 	struct		pvo_entry *pvo;
1917 	vm_paddr_t pa;
1918 
1919 	/*
1920 	 * Shortcut the direct-mapped case when applicable.  We never put
1921 	 * anything but 1:1 (or 62-bit aliased) mappings below
1922 	 * VM_MIN_KERNEL_ADDRESS.
1923 	 */
1924 	if (va < VM_MIN_KERNEL_ADDRESS)
1925 		return (va & ~DMAP_BASE_ADDRESS);
1926 
1927 	PMAP_LOCK(kernel_pmap);
1928 	pvo = moea64_pvo_find_va(kernel_pmap, va);
1929 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found for %#" PRIxPTR,
1930 	    va));
1931 	pa = (pvo->pvo_pte.pa & LPTE_RPGN) | (va - PVO_VADDR(pvo));
1932 	PMAP_UNLOCK(kernel_pmap);
1933 	return (pa);
1934 }
1935 
1936 /*
1937  * Remove a wired page from kernel virtual address space.
1938  */
1939 void
1940 moea64_kremove(mmu_t mmu, vm_offset_t va)
1941 {
1942 	moea64_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1943 }
1944 
1945 /*
1946  * Provide a kernel pointer corresponding to a given userland pointer.
1947  * The returned pointer is valid until the next time this function is
1948  * called in this thread. This is used internally in copyin/copyout.
1949  */
1950 static int
1951 moea64_map_user_ptr(mmu_t mmu, pmap_t pm, volatile const void *uaddr,
1952     void **kaddr, size_t ulen, size_t *klen)
1953 {
1954 	size_t l;
1955 #ifdef __powerpc64__
1956 	struct slb *slb;
1957 #endif
1958 	register_t slbv;
1959 
1960 	*kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
1961 	l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
1962 	if (l > ulen)
1963 		l = ulen;
1964 	if (klen)
1965 		*klen = l;
1966 	else if (l != ulen)
1967 		return (EFAULT);
1968 
1969 #ifdef __powerpc64__
1970 	/* Try lockless look-up first */
1971 	slb = user_va_to_slb_entry(pm, (vm_offset_t)uaddr);
1972 
1973 	if (slb == NULL) {
1974 		/* If it isn't there, we need to pre-fault the VSID */
1975 		PMAP_LOCK(pm);
1976 		slbv = va_to_vsid(pm, (vm_offset_t)uaddr) << SLBV_VSID_SHIFT;
1977 		PMAP_UNLOCK(pm);
1978 	} else {
1979 		slbv = slb->slbv;
1980 	}
1981 
1982 	/* Mark segment no-execute */
1983 	slbv |= SLBV_N;
1984 #else
1985 	slbv = va_to_vsid(pm, (vm_offset_t)uaddr);
1986 
1987 	/* Mark segment no-execute */
1988 	slbv |= SR_N;
1989 #endif
1990 
1991 	/* If we have already set this VSID, we can just return */
1992 	if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == slbv)
1993 		return (0);
1994 
1995 	__asm __volatile("isync");
1996 	curthread->td_pcb->pcb_cpu.aim.usr_segm =
1997 	    (uintptr_t)uaddr >> ADDR_SR_SHFT;
1998 	curthread->td_pcb->pcb_cpu.aim.usr_vsid = slbv;
1999 #ifdef __powerpc64__
2000 	__asm __volatile ("slbie %0; slbmte %1, %2; isync" ::
2001 	    "r"(USER_ADDR), "r"(slbv), "r"(USER_SLB_SLBE));
2002 #else
2003 	__asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(slbv));
2004 #endif
2005 
2006 	return (0);
2007 }
2008 
2009 /*
2010  * Figure out where a given kernel pointer (usually in a fault) points
2011  * to from the VM's perspective, potentially remapping into userland's
2012  * address space.
2013  */
2014 static int
2015 moea64_decode_kernel_ptr(mmu_t mmu, vm_offset_t addr, int *is_user,
2016     vm_offset_t *decoded_addr)
2017 {
2018 	vm_offset_t user_sr;
2019 
2020 	if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
2021 		user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
2022 		addr &= ADDR_PIDX | ADDR_POFF;
2023 		addr |= user_sr << ADDR_SR_SHFT;
2024 		*decoded_addr = addr;
2025 		*is_user = 1;
2026 	} else {
2027 		*decoded_addr = addr;
2028 		*is_user = 0;
2029 	}
2030 
2031 	return (0);
2032 }
2033 
2034 /*
2035  * Map a range of physical addresses into kernel virtual address space.
2036  *
2037  * The value passed in *virt is a suggested virtual address for the mapping.
2038  * Architectures which can support a direct-mapped physical to virtual region
2039  * can return the appropriate address within that region, leaving '*virt'
2040  * unchanged.  Other architectures should map the pages starting at '*virt' and
2041  * update '*virt' with the first usable address after the mapped region.
2042  */
2043 vm_offset_t
2044 moea64_map(mmu_t mmu, vm_offset_t *virt, vm_paddr_t pa_start,
2045     vm_paddr_t pa_end, int prot)
2046 {
2047 	vm_offset_t	sva, va;
2048 
2049 	if (hw_direct_map) {
2050 		/*
2051 		 * Check if every page in the region is covered by the direct
2052 		 * map. The direct map covers all of physical memory. Use
2053 		 * moea64_calc_wimg() as a shortcut to see if the page is in
2054 		 * physical memory as a way to see if the direct map covers it.
2055 		 */
2056 		for (va = pa_start; va < pa_end; va += PAGE_SIZE)
2057 			if (moea64_calc_wimg(va, VM_MEMATTR_DEFAULT) != LPTE_M)
2058 				break;
2059 		if (va == pa_end)
2060 			return (PHYS_TO_DMAP(pa_start));
2061 	}
2062 	sva = *virt;
2063 	va = sva;
2064 	/* XXX respect prot argument */
2065 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
2066 		moea64_kenter(mmu, va, pa_start);
2067 	*virt = va;
2068 
2069 	return (sva);
2070 }
2071 
2072 /*
2073  * Returns true if the pmap's pv is one of the first
2074  * 16 pvs linked to from this page.  This count may
2075  * be changed upwards or downwards in the future; it
2076  * is only necessary that true be returned for a small
2077  * subset of pmaps for proper page aging.
2078  */
2079 boolean_t
2080 moea64_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
2081 {
2082         int loops;
2083 	struct pvo_entry *pvo;
2084 	boolean_t rv;
2085 
2086 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2087 	    ("moea64_page_exists_quick: page %p is not managed", m));
2088 	loops = 0;
2089 	rv = FALSE;
2090 	PV_PAGE_LOCK(m);
2091 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2092 		if (!(pvo->pvo_vaddr & PVO_DEAD) && pvo->pvo_pmap == pmap) {
2093 			rv = TRUE;
2094 			break;
2095 		}
2096 		if (++loops >= 16)
2097 			break;
2098 	}
2099 	PV_PAGE_UNLOCK(m);
2100 	return (rv);
2101 }
2102 
2103 void
2104 moea64_page_init(mmu_t mmu __unused, vm_page_t m)
2105 {
2106 
2107 	m->md.mdpg_attrs = 0;
2108 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
2109 	LIST_INIT(&m->md.mdpg_pvoh);
2110 }
2111 
2112 /*
2113  * Return the number of managed mappings to the given physical page
2114  * that are wired.
2115  */
2116 int
2117 moea64_page_wired_mappings(mmu_t mmu, vm_page_t m)
2118 {
2119 	struct pvo_entry *pvo;
2120 	int count;
2121 
2122 	count = 0;
2123 	if ((m->oflags & VPO_UNMANAGED) != 0)
2124 		return (count);
2125 	PV_PAGE_LOCK(m);
2126 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
2127 		if ((pvo->pvo_vaddr & (PVO_DEAD | PVO_WIRED)) == PVO_WIRED)
2128 			count++;
2129 	PV_PAGE_UNLOCK(m);
2130 	return (count);
2131 }
2132 
2133 static uintptr_t	moea64_vsidcontext;
2134 
2135 uintptr_t
2136 moea64_get_unique_vsid(void) {
2137 	u_int entropy;
2138 	register_t hash;
2139 	uint32_t mask;
2140 	int i;
2141 
2142 	entropy = 0;
2143 	__asm __volatile("mftb %0" : "=r"(entropy));
2144 
2145 	mtx_lock(&moea64_slb_mutex);
2146 	for (i = 0; i < NVSIDS; i += VSID_NBPW) {
2147 		u_int	n;
2148 
2149 		/*
2150 		 * Create a new value by mutiplying by a prime and adding in
2151 		 * entropy from the timebase register.  This is to make the
2152 		 * VSID more random so that the PT hash function collides
2153 		 * less often.  (Note that the prime casues gcc to do shifts
2154 		 * instead of a multiply.)
2155 		 */
2156 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
2157 		hash = moea64_vsidcontext & (NVSIDS - 1);
2158 		if (hash == 0)		/* 0 is special, avoid it */
2159 			continue;
2160 		n = hash >> 5;
2161 		mask = 1 << (hash & (VSID_NBPW - 1));
2162 		hash = (moea64_vsidcontext & VSID_HASHMASK);
2163 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
2164 			/* anything free in this bucket? */
2165 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
2166 				entropy = (moea64_vsidcontext >> 20);
2167 				continue;
2168 			}
2169 			i = ffs(~moea64_vsid_bitmap[n]) - 1;
2170 			mask = 1 << i;
2171 			hash &= rounddown2(VSID_HASHMASK, VSID_NBPW);
2172 			hash |= i;
2173 		}
2174 		if (hash == VSID_VRMA)	/* also special, avoid this too */
2175 			continue;
2176 		KASSERT(!(moea64_vsid_bitmap[n] & mask),
2177 		    ("Allocating in-use VSID %#zx\n", hash));
2178 		moea64_vsid_bitmap[n] |= mask;
2179 		mtx_unlock(&moea64_slb_mutex);
2180 		return (hash);
2181 	}
2182 
2183 	mtx_unlock(&moea64_slb_mutex);
2184 	panic("%s: out of segments",__func__);
2185 }
2186 
2187 #ifdef __powerpc64__
2188 void
2189 moea64_pinit(mmu_t mmu, pmap_t pmap)
2190 {
2191 
2192 	RB_INIT(&pmap->pmap_pvo);
2193 
2194 	pmap->pm_slb_tree_root = slb_alloc_tree();
2195 	pmap->pm_slb = slb_alloc_user_cache();
2196 	pmap->pm_slb_len = 0;
2197 }
2198 #else
2199 void
2200 moea64_pinit(mmu_t mmu, pmap_t pmap)
2201 {
2202 	int	i;
2203 	uint32_t hash;
2204 
2205 	RB_INIT(&pmap->pmap_pvo);
2206 
2207 	if (pmap_bootstrapped)
2208 		pmap->pmap_phys = (pmap_t)moea64_kextract(mmu,
2209 		    (vm_offset_t)pmap);
2210 	else
2211 		pmap->pmap_phys = pmap;
2212 
2213 	/*
2214 	 * Allocate some segment registers for this pmap.
2215 	 */
2216 	hash = moea64_get_unique_vsid();
2217 
2218 	for (i = 0; i < 16; i++)
2219 		pmap->pm_sr[i] = VSID_MAKE(i, hash);
2220 
2221 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_pinit: pm_sr[0] = 0"));
2222 }
2223 #endif
2224 
2225 /*
2226  * Initialize the pmap associated with process 0.
2227  */
2228 void
2229 moea64_pinit0(mmu_t mmu, pmap_t pm)
2230 {
2231 
2232 	PMAP_LOCK_INIT(pm);
2233 	moea64_pinit(mmu, pm);
2234 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
2235 }
2236 
2237 /*
2238  * Set the physical protection on the specified range of this map as requested.
2239  */
2240 static void
2241 moea64_pvo_protect(mmu_t mmu,  pmap_t pm, struct pvo_entry *pvo, vm_prot_t prot)
2242 {
2243 	struct vm_page *pg;
2244 	vm_prot_t oldprot;
2245 	int32_t refchg;
2246 
2247 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2248 
2249 	/*
2250 	 * Change the protection of the page.
2251 	 */
2252 	oldprot = pvo->pvo_pte.prot;
2253 	pvo->pvo_pte.prot = prot;
2254 	pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
2255 
2256 	/*
2257 	 * If the PVO is in the page table, update mapping
2258 	 */
2259 	refchg = MOEA64_PTE_REPLACE(mmu, pvo, MOEA64_PTE_PROT_UPDATE);
2260 	if (refchg < 0)
2261 		refchg = (oldprot & VM_PROT_WRITE) ? LPTE_CHG : 0;
2262 
2263 	if (pm != kernel_pmap && pg != NULL &&
2264 	    (pg->a.flags & PGA_EXECUTABLE) == 0 &&
2265 	    (pvo->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
2266 		if ((pg->oflags & VPO_UNMANAGED) == 0)
2267 			vm_page_aflag_set(pg, PGA_EXECUTABLE);
2268 		moea64_syncicache(mmu, pm, PVO_VADDR(pvo),
2269 		    pvo->pvo_pte.pa & LPTE_RPGN, PAGE_SIZE);
2270 	}
2271 
2272 	/*
2273 	 * Update vm about the REF/CHG bits if the page is managed and we have
2274 	 * removed write access.
2275 	 */
2276 	if (pg != NULL && (pvo->pvo_vaddr & PVO_MANAGED) &&
2277 	    (oldprot & VM_PROT_WRITE)) {
2278 		refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2279 		if (refchg & LPTE_CHG)
2280 			vm_page_dirty(pg);
2281 		if (refchg & LPTE_REF)
2282 			vm_page_aflag_set(pg, PGA_REFERENCED);
2283 	}
2284 }
2285 
2286 void
2287 moea64_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
2288     vm_prot_t prot)
2289 {
2290 	struct	pvo_entry *pvo, *tpvo, key;
2291 
2292 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm,
2293 	    sva, eva, prot);
2294 
2295 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
2296 	    ("moea64_protect: non current pmap"));
2297 
2298 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2299 		moea64_remove(mmu, pm, sva, eva);
2300 		return;
2301 	}
2302 
2303 	PMAP_LOCK(pm);
2304 	key.pvo_vaddr = sva;
2305 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2306 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2307 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2308 		moea64_pvo_protect(mmu, pm, pvo, prot);
2309 	}
2310 	PMAP_UNLOCK(pm);
2311 }
2312 
2313 /*
2314  * Map a list of wired pages into kernel virtual address space.  This is
2315  * intended for temporary mappings which do not need page modification or
2316  * references recorded.  Existing mappings in the region are overwritten.
2317  */
2318 void
2319 moea64_qenter(mmu_t mmu, vm_offset_t va, vm_page_t *m, int count)
2320 {
2321 	while (count-- > 0) {
2322 		moea64_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
2323 		va += PAGE_SIZE;
2324 		m++;
2325 	}
2326 }
2327 
2328 /*
2329  * Remove page mappings from kernel virtual address space.  Intended for
2330  * temporary mappings entered by moea64_qenter.
2331  */
2332 void
2333 moea64_qremove(mmu_t mmu, vm_offset_t va, int count)
2334 {
2335 	while (count-- > 0) {
2336 		moea64_kremove(mmu, va);
2337 		va += PAGE_SIZE;
2338 	}
2339 }
2340 
2341 void
2342 moea64_release_vsid(uint64_t vsid)
2343 {
2344 	int idx, mask;
2345 
2346 	mtx_lock(&moea64_slb_mutex);
2347 	idx = vsid & (NVSIDS-1);
2348 	mask = 1 << (idx % VSID_NBPW);
2349 	idx /= VSID_NBPW;
2350 	KASSERT(moea64_vsid_bitmap[idx] & mask,
2351 	    ("Freeing unallocated VSID %#jx", vsid));
2352 	moea64_vsid_bitmap[idx] &= ~mask;
2353 	mtx_unlock(&moea64_slb_mutex);
2354 }
2355 
2356 
2357 void
2358 moea64_release(mmu_t mmu, pmap_t pmap)
2359 {
2360 
2361 	/*
2362 	 * Free segment registers' VSIDs
2363 	 */
2364     #ifdef __powerpc64__
2365 	slb_free_tree(pmap);
2366 	slb_free_user_cache(pmap->pm_slb);
2367     #else
2368 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_release: pm_sr[0] = 0"));
2369 
2370 	moea64_release_vsid(VSID_TO_HASH(pmap->pm_sr[0]));
2371     #endif
2372 }
2373 
2374 /*
2375  * Remove all pages mapped by the specified pmap
2376  */
2377 void
2378 moea64_remove_pages(mmu_t mmu, pmap_t pm)
2379 {
2380 	struct pvo_entry *pvo, *tpvo;
2381 	struct pvo_dlist tofree;
2382 
2383 	SLIST_INIT(&tofree);
2384 
2385 	PMAP_LOCK(pm);
2386 	RB_FOREACH_SAFE(pvo, pvo_tree, &pm->pmap_pvo, tpvo) {
2387 		if (pvo->pvo_vaddr & PVO_WIRED)
2388 			continue;
2389 
2390 		/*
2391 		 * For locking reasons, remove this from the page table and
2392 		 * pmap, but save delinking from the vm_page for a second
2393 		 * pass
2394 		 */
2395 		moea64_pvo_remove_from_pmap(mmu, pvo);
2396 		SLIST_INSERT_HEAD(&tofree, pvo, pvo_dlink);
2397 	}
2398 	PMAP_UNLOCK(pm);
2399 
2400 	while (!SLIST_EMPTY(&tofree)) {
2401 		pvo = SLIST_FIRST(&tofree);
2402 		SLIST_REMOVE_HEAD(&tofree, pvo_dlink);
2403 		moea64_pvo_remove_from_page(mmu, pvo);
2404 		free_pvo_entry(pvo);
2405 	}
2406 }
2407 
2408 /*
2409  * Remove the given range of addresses from the specified map.
2410  */
2411 void
2412 moea64_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
2413 {
2414 	struct  pvo_entry *pvo, *tpvo, key;
2415 	struct pvo_dlist tofree;
2416 
2417 	/*
2418 	 * Perform an unsynchronized read.  This is, however, safe.
2419 	 */
2420 	if (pm->pm_stats.resident_count == 0)
2421 		return;
2422 
2423 	key.pvo_vaddr = sva;
2424 
2425 	SLIST_INIT(&tofree);
2426 
2427 	PMAP_LOCK(pm);
2428 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2429 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2430 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2431 
2432 		/*
2433 		 * For locking reasons, remove this from the page table and
2434 		 * pmap, but save delinking from the vm_page for a second
2435 		 * pass
2436 		 */
2437 		moea64_pvo_remove_from_pmap(mmu, pvo);
2438 		SLIST_INSERT_HEAD(&tofree, pvo, pvo_dlink);
2439 	}
2440 	PMAP_UNLOCK(pm);
2441 
2442 	while (!SLIST_EMPTY(&tofree)) {
2443 		pvo = SLIST_FIRST(&tofree);
2444 		SLIST_REMOVE_HEAD(&tofree, pvo_dlink);
2445 		moea64_pvo_remove_from_page(mmu, pvo);
2446 		free_pvo_entry(pvo);
2447 	}
2448 }
2449 
2450 /*
2451  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
2452  * will reflect changes in pte's back to the vm_page.
2453  */
2454 void
2455 moea64_remove_all(mmu_t mmu, vm_page_t m)
2456 {
2457 	struct	pvo_entry *pvo, *next_pvo;
2458 	struct	pvo_head freequeue;
2459 	int	wasdead;
2460 	pmap_t	pmap;
2461 
2462 	LIST_INIT(&freequeue);
2463 
2464 	PV_PAGE_LOCK(m);
2465 	LIST_FOREACH_SAFE(pvo, vm_page_to_pvoh(m), pvo_vlink, next_pvo) {
2466 		pmap = pvo->pvo_pmap;
2467 		PMAP_LOCK(pmap);
2468 		wasdead = (pvo->pvo_vaddr & PVO_DEAD);
2469 		if (!wasdead)
2470 			moea64_pvo_remove_from_pmap(mmu, pvo);
2471 		moea64_pvo_remove_from_page_locked(mmu, pvo, m);
2472 		if (!wasdead)
2473 			LIST_INSERT_HEAD(&freequeue, pvo, pvo_vlink);
2474 		PMAP_UNLOCK(pmap);
2475 
2476 	}
2477 	KASSERT(!pmap_page_is_mapped(m), ("Page still has mappings"));
2478 	KASSERT((m->a.flags & PGA_WRITEABLE) == 0, ("Page still writable"));
2479 	PV_PAGE_UNLOCK(m);
2480 
2481 	/* Clean up UMA allocations */
2482 	LIST_FOREACH_SAFE(pvo, &freequeue, pvo_vlink, next_pvo)
2483 		free_pvo_entry(pvo);
2484 }
2485 
2486 /*
2487  * Allocate a physical page of memory directly from the phys_avail map.
2488  * Can only be called from moea64_bootstrap before avail start and end are
2489  * calculated.
2490  */
2491 vm_offset_t
2492 moea64_bootstrap_alloc(vm_size_t size, vm_size_t align)
2493 {
2494 	vm_offset_t	s, e;
2495 	int		i, j;
2496 
2497 	size = round_page(size);
2498 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2499 		if (align != 0)
2500 			s = roundup2(phys_avail[i], align);
2501 		else
2502 			s = phys_avail[i];
2503 		e = s + size;
2504 
2505 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2506 			continue;
2507 
2508 		if (s + size > platform_real_maxaddr())
2509 			continue;
2510 
2511 		if (s == phys_avail[i]) {
2512 			phys_avail[i] += size;
2513 		} else if (e == phys_avail[i + 1]) {
2514 			phys_avail[i + 1] -= size;
2515 		} else {
2516 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2517 				phys_avail[j] = phys_avail[j - 2];
2518 				phys_avail[j + 1] = phys_avail[j - 1];
2519 			}
2520 
2521 			phys_avail[i + 3] = phys_avail[i + 1];
2522 			phys_avail[i + 1] = s;
2523 			phys_avail[i + 2] = e;
2524 			phys_avail_count++;
2525 		}
2526 
2527 		return (s);
2528 	}
2529 	panic("moea64_bootstrap_alloc: could not allocate memory");
2530 }
2531 
2532 static int
2533 moea64_pvo_enter(mmu_t mmu, struct pvo_entry *pvo, struct pvo_head *pvo_head,
2534     struct pvo_entry **oldpvop)
2535 {
2536 	struct pvo_entry *old_pvo;
2537 	int err;
2538 
2539 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2540 
2541 	STAT_MOEA64(moea64_pvo_enter_calls++);
2542 
2543 	/*
2544 	 * Add to pmap list
2545 	 */
2546 	old_pvo = RB_INSERT(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2547 
2548 	if (old_pvo != NULL) {
2549 		if (oldpvop != NULL)
2550 			*oldpvop = old_pvo;
2551 		return (EEXIST);
2552 	}
2553 
2554 	if (pvo_head != NULL) {
2555 		LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2556 	}
2557 
2558 	if (pvo->pvo_vaddr & PVO_WIRED)
2559 		pvo->pvo_pmap->pm_stats.wired_count++;
2560 	pvo->pvo_pmap->pm_stats.resident_count++;
2561 
2562 	/*
2563 	 * Insert it into the hardware page table
2564 	 */
2565 	err = MOEA64_PTE_INSERT(mmu, pvo);
2566 	if (err != 0) {
2567 		panic("moea64_pvo_enter: overflow");
2568 	}
2569 
2570 	STAT_MOEA64(moea64_pvo_entries++);
2571 
2572 	if (pvo->pvo_pmap == kernel_pmap)
2573 		isync();
2574 
2575 #ifdef __powerpc64__
2576 	/*
2577 	 * Make sure all our bootstrap mappings are in the SLB as soon
2578 	 * as virtual memory is switched on.
2579 	 */
2580 	if (!pmap_bootstrapped)
2581 		moea64_bootstrap_slb_prefault(PVO_VADDR(pvo),
2582 		    pvo->pvo_vaddr & PVO_LARGE);
2583 #endif
2584 
2585 	return (0);
2586 }
2587 
2588 static void
2589 moea64_pvo_remove_from_pmap(mmu_t mmu, struct pvo_entry *pvo)
2590 {
2591 	struct	vm_page *pg;
2592 	int32_t refchg;
2593 
2594 	KASSERT(pvo->pvo_pmap != NULL, ("Trying to remove PVO with no pmap"));
2595 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2596 	KASSERT(!(pvo->pvo_vaddr & PVO_DEAD), ("Trying to remove dead PVO"));
2597 
2598 	/*
2599 	 * If there is an active pte entry, we need to deactivate it
2600 	 */
2601 	refchg = MOEA64_PTE_UNSET(mmu, pvo);
2602 	if (refchg < 0) {
2603 		/*
2604 		 * If it was evicted from the page table, be pessimistic and
2605 		 * dirty the page.
2606 		 */
2607 		if (pvo->pvo_pte.prot & VM_PROT_WRITE)
2608 			refchg = LPTE_CHG;
2609 		else
2610 			refchg = 0;
2611 	}
2612 
2613 	/*
2614 	 * Update our statistics.
2615 	 */
2616 	pvo->pvo_pmap->pm_stats.resident_count--;
2617 	if (pvo->pvo_vaddr & PVO_WIRED)
2618 		pvo->pvo_pmap->pm_stats.wired_count--;
2619 
2620 	/*
2621 	 * Remove this PVO from the pmap list.
2622 	 */
2623 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2624 
2625 	/*
2626 	 * Mark this for the next sweep
2627 	 */
2628 	pvo->pvo_vaddr |= PVO_DEAD;
2629 
2630 	/* Send RC bits to VM */
2631 	if ((pvo->pvo_vaddr & PVO_MANAGED) &&
2632 	    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2633 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
2634 		if (pg != NULL) {
2635 			refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2636 			if (refchg & LPTE_CHG)
2637 				vm_page_dirty(pg);
2638 			if (refchg & LPTE_REF)
2639 				vm_page_aflag_set(pg, PGA_REFERENCED);
2640 		}
2641 	}
2642 }
2643 
2644 static inline void
2645 moea64_pvo_remove_from_page_locked(mmu_t mmu, struct pvo_entry *pvo,
2646     vm_page_t m)
2647 {
2648 
2649 	KASSERT(pvo->pvo_vaddr & PVO_DEAD, ("Trying to delink live page"));
2650 
2651 	/* Use NULL pmaps as a sentinel for races in page deletion */
2652 	if (pvo->pvo_pmap == NULL)
2653 		return;
2654 	pvo->pvo_pmap = NULL;
2655 
2656 	/*
2657 	 * Update vm about page writeability/executability if managed
2658 	 */
2659 	PV_LOCKASSERT(pvo->pvo_pte.pa & LPTE_RPGN);
2660 	if (pvo->pvo_vaddr & PVO_MANAGED) {
2661 		if (m != NULL) {
2662 			LIST_REMOVE(pvo, pvo_vlink);
2663 			if (LIST_EMPTY(vm_page_to_pvoh(m)))
2664 				vm_page_aflag_clear(m,
2665 				    PGA_WRITEABLE | PGA_EXECUTABLE);
2666 		}
2667 	}
2668 
2669 	STAT_MOEA64(moea64_pvo_entries--);
2670 	STAT_MOEA64(moea64_pvo_remove_calls++);
2671 }
2672 
2673 static void
2674 moea64_pvo_remove_from_page(mmu_t mmu, struct pvo_entry *pvo)
2675 {
2676 	vm_page_t pg = NULL;
2677 
2678 	if (pvo->pvo_vaddr & PVO_MANAGED)
2679 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
2680 
2681 	PV_LOCK(pvo->pvo_pte.pa & LPTE_RPGN);
2682 	moea64_pvo_remove_from_page_locked(mmu, pvo, pg);
2683 	PV_UNLOCK(pvo->pvo_pte.pa & LPTE_RPGN);
2684 }
2685 
2686 static struct pvo_entry *
2687 moea64_pvo_find_va(pmap_t pm, vm_offset_t va)
2688 {
2689 	struct pvo_entry key;
2690 
2691 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2692 
2693 	key.pvo_vaddr = va & ~ADDR_POFF;
2694 	return (RB_FIND(pvo_tree, &pm->pmap_pvo, &key));
2695 }
2696 
2697 static boolean_t
2698 moea64_query_bit(mmu_t mmu, vm_page_t m, uint64_t ptebit)
2699 {
2700 	struct	pvo_entry *pvo;
2701 	int64_t ret;
2702 	boolean_t rv;
2703 
2704 	/*
2705 	 * See if this bit is stored in the page already.
2706 	 */
2707 	if (m->md.mdpg_attrs & ptebit)
2708 		return (TRUE);
2709 
2710 	/*
2711 	 * Examine each PTE.  Sync so that any pending REF/CHG bits are
2712 	 * flushed to the PTEs.
2713 	 */
2714 	rv = FALSE;
2715 	powerpc_sync();
2716 	PV_PAGE_LOCK(m);
2717 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2718 		ret = 0;
2719 
2720 		/*
2721 		 * See if this pvo has a valid PTE.  if so, fetch the
2722 		 * REF/CHG bits from the valid PTE.  If the appropriate
2723 		 * ptebit is set, return success.
2724 		 */
2725 		PMAP_LOCK(pvo->pvo_pmap);
2726 		if (!(pvo->pvo_vaddr & PVO_DEAD))
2727 			ret = MOEA64_PTE_SYNCH(mmu, pvo);
2728 		PMAP_UNLOCK(pvo->pvo_pmap);
2729 
2730 		if (ret > 0) {
2731 			atomic_set_32(&m->md.mdpg_attrs,
2732 			    ret & (LPTE_CHG | LPTE_REF));
2733 			if (ret & ptebit) {
2734 				rv = TRUE;
2735 				break;
2736 			}
2737 		}
2738 	}
2739 	PV_PAGE_UNLOCK(m);
2740 
2741 	return (rv);
2742 }
2743 
2744 static u_int
2745 moea64_clear_bit(mmu_t mmu, vm_page_t m, u_int64_t ptebit)
2746 {
2747 	u_int	count;
2748 	struct	pvo_entry *pvo;
2749 	int64_t ret;
2750 
2751 	/*
2752 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2753 	 * we can reset the right ones).
2754 	 */
2755 	powerpc_sync();
2756 
2757 	/*
2758 	 * For each pvo entry, clear the pte's ptebit.
2759 	 */
2760 	count = 0;
2761 	PV_PAGE_LOCK(m);
2762 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2763 		ret = 0;
2764 
2765 		PMAP_LOCK(pvo->pvo_pmap);
2766 		if (!(pvo->pvo_vaddr & PVO_DEAD))
2767 			ret = MOEA64_PTE_CLEAR(mmu, pvo, ptebit);
2768 		PMAP_UNLOCK(pvo->pvo_pmap);
2769 
2770 		if (ret > 0 && (ret & ptebit))
2771 			count++;
2772 	}
2773 	atomic_clear_32(&m->md.mdpg_attrs, ptebit);
2774 	PV_PAGE_UNLOCK(m);
2775 
2776 	return (count);
2777 }
2778 
2779 boolean_t
2780 moea64_dev_direct_mapped(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2781 {
2782 	struct pvo_entry *pvo, key;
2783 	vm_offset_t ppa;
2784 	int error = 0;
2785 
2786 	if (hw_direct_map && mem_valid(pa, size) == 0)
2787 		return (0);
2788 
2789 	PMAP_LOCK(kernel_pmap);
2790 	ppa = pa & ~ADDR_POFF;
2791 	key.pvo_vaddr = DMAP_BASE_ADDRESS + ppa;
2792 	for (pvo = RB_FIND(pvo_tree, &kernel_pmap->pmap_pvo, &key);
2793 	    ppa < pa + size; ppa += PAGE_SIZE,
2794 	    pvo = RB_NEXT(pvo_tree, &kernel_pmap->pmap_pvo, pvo)) {
2795 		if (pvo == NULL || (pvo->pvo_pte.pa & LPTE_RPGN) != ppa) {
2796 			error = EFAULT;
2797 			break;
2798 		}
2799 	}
2800 	PMAP_UNLOCK(kernel_pmap);
2801 
2802 	return (error);
2803 }
2804 
2805 /*
2806  * Map a set of physical memory pages into the kernel virtual
2807  * address space. Return a pointer to where it is mapped. This
2808  * routine is intended to be used for mapping device memory,
2809  * NOT real memory.
2810  */
2811 void *
2812 moea64_mapdev_attr(mmu_t mmu, vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
2813 {
2814 	vm_offset_t va, tmpva, ppa, offset;
2815 
2816 	ppa = trunc_page(pa);
2817 	offset = pa & PAGE_MASK;
2818 	size = roundup2(offset + size, PAGE_SIZE);
2819 
2820 	va = kva_alloc(size);
2821 
2822 	if (!va)
2823 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
2824 
2825 	for (tmpva = va; size > 0;) {
2826 		moea64_kenter_attr(mmu, tmpva, ppa, ma);
2827 		size -= PAGE_SIZE;
2828 		tmpva += PAGE_SIZE;
2829 		ppa += PAGE_SIZE;
2830 	}
2831 
2832 	return ((void *)(va + offset));
2833 }
2834 
2835 void *
2836 moea64_mapdev(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2837 {
2838 
2839 	return moea64_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT);
2840 }
2841 
2842 void
2843 moea64_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2844 {
2845 	vm_offset_t base, offset;
2846 
2847 	base = trunc_page(va);
2848 	offset = va & PAGE_MASK;
2849 	size = roundup2(offset + size, PAGE_SIZE);
2850 
2851 	kva_free(base, size);
2852 }
2853 
2854 void
2855 moea64_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2856 {
2857 	struct pvo_entry *pvo;
2858 	vm_offset_t lim;
2859 	vm_paddr_t pa;
2860 	vm_size_t len;
2861 
2862 	if (__predict_false(pm == NULL))
2863 		pm = &curthread->td_proc->p_vmspace->vm_pmap;
2864 
2865 	PMAP_LOCK(pm);
2866 	while (sz > 0) {
2867 		lim = round_page(va+1);
2868 		len = MIN(lim - va, sz);
2869 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
2870 		if (pvo != NULL && !(pvo->pvo_pte.pa & LPTE_I)) {
2871 			pa = (pvo->pvo_pte.pa & LPTE_RPGN) | (va & ADDR_POFF);
2872 			moea64_syncicache(mmu, pm, va, pa, len);
2873 		}
2874 		va += len;
2875 		sz -= len;
2876 	}
2877 	PMAP_UNLOCK(pm);
2878 }
2879 
2880 void
2881 moea64_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va)
2882 {
2883 
2884 	*va = (void *)(uintptr_t)pa;
2885 }
2886 
2887 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
2888 
2889 void
2890 moea64_scan_init(mmu_t mmu)
2891 {
2892 	struct pvo_entry *pvo;
2893 	vm_offset_t va;
2894 	int i;
2895 
2896 	if (!do_minidump) {
2897 		/* Initialize phys. segments for dumpsys(). */
2898 		memset(&dump_map, 0, sizeof(dump_map));
2899 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
2900 		for (i = 0; i < pregions_sz; i++) {
2901 			dump_map[i].pa_start = pregions[i].mr_start;
2902 			dump_map[i].pa_size = pregions[i].mr_size;
2903 		}
2904 		return;
2905 	}
2906 
2907 	/* Virtual segments for minidumps: */
2908 	memset(&dump_map, 0, sizeof(dump_map));
2909 
2910 	/* 1st: kernel .data and .bss. */
2911 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
2912 	dump_map[0].pa_size = round_page((uintptr_t)_end) -
2913 	    dump_map[0].pa_start;
2914 
2915 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2916 	dump_map[1].pa_start = (vm_paddr_t)(uintptr_t)msgbufp->msg_ptr;
2917 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
2918 
2919 	/* 3rd: kernel VM. */
2920 	va = dump_map[1].pa_start + dump_map[1].pa_size;
2921 	/* Find start of next chunk (from va). */
2922 	while (va < virtual_end) {
2923 		/* Don't dump the buffer cache. */
2924 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
2925 			va = kmi.buffer_eva;
2926 			continue;
2927 		}
2928 		pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
2929 		if (pvo != NULL && !(pvo->pvo_vaddr & PVO_DEAD))
2930 			break;
2931 		va += PAGE_SIZE;
2932 	}
2933 	if (va < virtual_end) {
2934 		dump_map[2].pa_start = va;
2935 		va += PAGE_SIZE;
2936 		/* Find last page in chunk. */
2937 		while (va < virtual_end) {
2938 			/* Don't run into the buffer cache. */
2939 			if (va == kmi.buffer_sva)
2940 				break;
2941 			pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
2942 			if (pvo == NULL || (pvo->pvo_vaddr & PVO_DEAD))
2943 				break;
2944 			va += PAGE_SIZE;
2945 		}
2946 		dump_map[2].pa_size = va - dump_map[2].pa_start;
2947 	}
2948 }
2949 
2950 #ifdef __powerpc64__
2951 
2952 static size_t
2953 moea64_scan_pmap(mmu_t mmu)
2954 {
2955 	struct pvo_entry *pvo;
2956 	vm_paddr_t pa, pa_end;
2957 	vm_offset_t va, pgva, kstart, kend, kstart_lp, kend_lp;
2958 	uint64_t lpsize;
2959 
2960 	lpsize = moea64_large_page_size;
2961 	kstart = trunc_page((vm_offset_t)_etext);
2962 	kend = round_page((vm_offset_t)_end);
2963 	kstart_lp = kstart & ~moea64_large_page_mask;
2964 	kend_lp = (kend + moea64_large_page_mask) & ~moea64_large_page_mask;
2965 
2966 	CTR4(KTR_PMAP, "moea64_scan_pmap: kstart=0x%016lx, kend=0x%016lx, "
2967 	    "kstart_lp=0x%016lx, kend_lp=0x%016lx",
2968 	    kstart, kend, kstart_lp, kend_lp);
2969 
2970 	PMAP_LOCK(kernel_pmap);
2971 	RB_FOREACH(pvo, pvo_tree, &kernel_pmap->pmap_pvo) {
2972 		va = pvo->pvo_vaddr;
2973 
2974 		if (va & PVO_DEAD)
2975 			continue;
2976 
2977 		/* Skip DMAP (except kernel area) */
2978 		if (va >= DMAP_BASE_ADDRESS && va <= DMAP_MAX_ADDRESS) {
2979 			if (va & PVO_LARGE) {
2980 				pgva = va & ~moea64_large_page_mask;
2981 				if (pgva < kstart_lp || pgva >= kend_lp)
2982 					continue;
2983 			} else {
2984 				pgva = trunc_page(va);
2985 				if (pgva < kstart || pgva >= kend)
2986 					continue;
2987 			}
2988 		}
2989 
2990 		pa = pvo->pvo_pte.pa & LPTE_RPGN;
2991 
2992 		if (va & PVO_LARGE) {
2993 			pa_end = pa + lpsize;
2994 			for (; pa < pa_end; pa += PAGE_SIZE) {
2995 				if (is_dumpable(pa))
2996 					dump_add_page(pa);
2997 			}
2998 		} else {
2999 			if (is_dumpable(pa))
3000 				dump_add_page(pa);
3001 		}
3002 	}
3003 	PMAP_UNLOCK(kernel_pmap);
3004 
3005 	return (sizeof(struct lpte) * moea64_pteg_count * 8);
3006 }
3007 
3008 static struct dump_context dump_ctx;
3009 
3010 static void *
3011 moea64_dump_pmap_init(mmu_t mmu, unsigned blkpgs)
3012 {
3013 	dump_ctx.ptex = 0;
3014 	dump_ctx.ptex_end = moea64_pteg_count * 8;
3015 	dump_ctx.blksz = blkpgs * PAGE_SIZE;
3016 	return (&dump_ctx);
3017 }
3018 
3019 #else
3020 
3021 static size_t
3022 moea64_scan_pmap(mmu_t mmu)
3023 {
3024 	return (0);
3025 }
3026 
3027 static void *
3028 moea64_dump_pmap_init(mmu_t mmu, unsigned blkpgs)
3029 {
3030 	return (NULL);
3031 }
3032 
3033 #endif
3034 
3035 #ifdef __powerpc64__
3036 static void
3037 moea64_map_range(mmu_t mmu, vm_offset_t va, vm_paddr_t pa, vm_size_t npages)
3038 {
3039 
3040 	for (; npages > 0; --npages) {
3041 		if (moea64_large_page_size != 0 &&
3042 		    (pa & moea64_large_page_mask) == 0 &&
3043 		    (va & moea64_large_page_mask) == 0 &&
3044 		    npages >= (moea64_large_page_size >> PAGE_SHIFT)) {
3045 			PMAP_LOCK(kernel_pmap);
3046 			moea64_kenter_large(mmu, va, pa, 0, 0);
3047 			PMAP_UNLOCK(kernel_pmap);
3048 			pa += moea64_large_page_size;
3049 			va += moea64_large_page_size;
3050 			npages -= (moea64_large_page_size >> PAGE_SHIFT) - 1;
3051 		} else {
3052 			moea64_kenter(mmu, va, pa);
3053 			pa += PAGE_SIZE;
3054 			va += PAGE_SIZE;
3055 		}
3056 	}
3057 }
3058 
3059 static void
3060 moea64_page_array_startup(mmu_t mmu, long pages)
3061 {
3062 	long dom_pages[MAXMEMDOM];
3063 	vm_paddr_t pa;
3064 	vm_offset_t va, vm_page_base;
3065 	vm_size_t needed, size;
3066 	long page;
3067 	int domain;
3068 	int i;
3069 
3070 	vm_page_base = 0xd000000000000000ULL;
3071 
3072 	/* Short-circuit single-domain systems. */
3073 	if (vm_ndomains == 1) {
3074 		size = round_page(pages * sizeof(struct vm_page));
3075 		pa = vm_phys_early_alloc(0, size);
3076 		vm_page_base = moea64_map(mmu, &vm_page_base,
3077 		    pa, pa + size, VM_PROT_READ | VM_PROT_WRITE);
3078 		vm_page_array_size = pages;
3079 		vm_page_array = (vm_page_t)vm_page_base;
3080 		return;
3081 	}
3082 
3083 	page = 0;
3084 	for (i = 0; i < MAXMEMDOM; i++)
3085 		dom_pages[i] = 0;
3086 
3087 	/* Now get the number of pages required per domain. */
3088 	for (i = 0; i < vm_phys_nsegs; i++) {
3089 		domain = vm_phys_segs[i].domain;
3090 		KASSERT(domain < MAXMEMDOM,
3091 		    ("Invalid vm_phys_segs NUMA domain %d!\n", domain));
3092 		/* Get size of vm_page_array needed for this segment. */
3093 		size = btoc(vm_phys_segs[i].end - vm_phys_segs[i].start);
3094 		dom_pages[domain] += size;
3095 	}
3096 
3097 	for (i = 0; phys_avail[i + 1] != 0; i+= 2) {
3098 		domain = _vm_phys_domain(phys_avail[i]);
3099 		KASSERT(domain < MAXMEMDOM,
3100 		    ("Invalid phys_avail NUMA domain %d!\n", domain));
3101 		size = btoc(phys_avail[i + 1] - phys_avail[i]);
3102 		dom_pages[domain] += size;
3103 	}
3104 
3105 	/*
3106 	 * Map in chunks that can get us all 16MB pages.  There will be some
3107 	 * overlap between domains, but that's acceptable for now.
3108 	 */
3109 	vm_page_array_size = 0;
3110 	va = vm_page_base;
3111 	for (i = 0; i < MAXMEMDOM && vm_page_array_size < pages; i++) {
3112 		if (dom_pages[i] == 0)
3113 			continue;
3114 		size = ulmin(pages - vm_page_array_size, dom_pages[i]);
3115 		size = round_page(size * sizeof(struct vm_page));
3116 		needed = size;
3117 		size = roundup2(size, moea64_large_page_size);
3118 		pa = vm_phys_early_alloc(i, size);
3119 		vm_page_array_size += size / sizeof(struct vm_page);
3120 		moea64_map_range(mmu, va, pa, size >> PAGE_SHIFT);
3121 		/* Scoot up domain 0, to reduce the domain page overlap. */
3122 		if (i == 0)
3123 			vm_page_base += size - needed;
3124 		va += size;
3125 	}
3126 	vm_page_array = (vm_page_t)vm_page_base;
3127 	vm_page_array_size = pages;
3128 }
3129 #endif
3130