xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision 52f72944b8f5abb2386eae924357dee8aea17d5b)
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_compat.h"
49 #include "opt_kstack_pages.h"
50 
51 #include <sys/param.h>
52 #include <sys/kernel.h>
53 #include <sys/conf.h>
54 #include <sys/queue.h>
55 #include <sys/cpuset.h>
56 #include <sys/kerneldump.h>
57 #include <sys/ktr.h>
58 #include <sys/lock.h>
59 #include <sys/msgbuf.h>
60 #include <sys/malloc.h>
61 #include <sys/mutex.h>
62 #include <sys/proc.h>
63 #include <sys/rwlock.h>
64 #include <sys/sched.h>
65 #include <sys/sysctl.h>
66 #include <sys/systm.h>
67 #include <sys/vmmeter.h>
68 #include <sys/smp.h>
69 
70 #include <sys/kdb.h>
71 
72 #include <dev/ofw/openfirm.h>
73 
74 #include <vm/vm.h>
75 #include <vm/vm_param.h>
76 #include <vm/vm_kern.h>
77 #include <vm/vm_page.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_COUNT	PA_LOCK_COUNT*3
123 static struct mtx_padalign pv_lock[PV_LOCK_COUNT];
124 
125 #define PV_LOCKPTR(pa)	((struct mtx *)(&pv_lock[pa_index(pa) % PV_LOCK_COUNT]))
126 #define PV_LOCK(pa)		mtx_lock(PV_LOCKPTR(pa))
127 #define PV_UNLOCK(pa)		mtx_unlock(PV_LOCKPTR(pa))
128 #define PV_LOCKASSERT(pa) 	mtx_assert(PV_LOCKPTR(pa), MA_OWNED)
129 #define PV_PAGE_LOCK(m)		PV_LOCK(VM_PAGE_TO_PHYS(m))
130 #define PV_PAGE_UNLOCK(m)	PV_UNLOCK(VM_PAGE_TO_PHYS(m))
131 #define PV_PAGE_LOCKASSERT(m)	PV_LOCKASSERT(VM_PAGE_TO_PHYS(m))
132 
133 struct ofw_map {
134 	cell_t	om_va;
135 	cell_t	om_len;
136 	uint64_t om_pa;
137 	cell_t	om_mode;
138 };
139 
140 extern unsigned char _etext[];
141 extern unsigned char _end[];
142 
143 extern void *slbtrap, *slbtrapend;
144 
145 /*
146  * Map of physical memory regions.
147  */
148 static struct	mem_region *regions;
149 static struct	mem_region *pregions;
150 static u_int	phys_avail_count;
151 static int	regions_sz, pregions_sz;
152 
153 extern void bs_remap_earlyboot(void);
154 
155 /*
156  * Lock for the SLB tables.
157  */
158 struct mtx	moea64_slb_mutex;
159 
160 /*
161  * PTEG data.
162  */
163 u_int		moea64_pteg_count;
164 u_int		moea64_pteg_mask;
165 
166 /*
167  * PVO data.
168  */
169 
170 uma_zone_t	moea64_pvo_zone; /* zone for pvo entries */
171 
172 static struct	pvo_entry *moea64_bpvo_pool;
173 static int	moea64_bpvo_pool_index = 0;
174 static int	moea64_bpvo_pool_size = 327680;
175 TUNABLE_INT("machdep.moea64_bpvo_pool_size", &moea64_bpvo_pool_size);
176 SYSCTL_INT(_machdep, OID_AUTO, moea64_allocated_bpvo_entries, CTLFLAG_RD,
177     &moea64_bpvo_pool_index, 0, "");
178 
179 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
180 #ifdef __powerpc64__
181 #define	NVSIDS		(NPMAPS * 16)
182 #define VSID_HASHMASK	0xffffffffUL
183 #else
184 #define NVSIDS		NPMAPS
185 #define VSID_HASHMASK	0xfffffUL
186 #endif
187 static u_int	moea64_vsid_bitmap[NVSIDS / VSID_NBPW];
188 
189 static boolean_t moea64_initialized = FALSE;
190 
191 /*
192  * Statistics.
193  */
194 u_int	moea64_pte_valid = 0;
195 u_int	moea64_pte_overflow = 0;
196 u_int	moea64_pvo_entries = 0;
197 u_int	moea64_pvo_enter_calls = 0;
198 u_int	moea64_pvo_remove_calls = 0;
199 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_valid, CTLFLAG_RD,
200     &moea64_pte_valid, 0, "");
201 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_overflow, CTLFLAG_RD,
202     &moea64_pte_overflow, 0, "");
203 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_entries, CTLFLAG_RD,
204     &moea64_pvo_entries, 0, "");
205 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_enter_calls, CTLFLAG_RD,
206     &moea64_pvo_enter_calls, 0, "");
207 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_remove_calls, CTLFLAG_RD,
208     &moea64_pvo_remove_calls, 0, "");
209 
210 vm_offset_t	moea64_scratchpage_va[2];
211 struct pvo_entry *moea64_scratchpage_pvo[2];
212 struct	mtx	moea64_scratchpage_mtx;
213 
214 uint64_t 	moea64_large_page_mask = 0;
215 uint64_t	moea64_large_page_size = 0;
216 int		moea64_large_page_shift = 0;
217 
218 /*
219  * PVO calls.
220  */
221 static int	moea64_pvo_enter(mmu_t mmu, struct pvo_entry *pvo,
222 		    struct pvo_head *pvo_head);
223 static void	moea64_pvo_remove_from_pmap(mmu_t mmu, struct pvo_entry *pvo);
224 static void	moea64_pvo_remove_from_page(mmu_t mmu, struct pvo_entry *pvo);
225 static struct	pvo_entry *moea64_pvo_find_va(pmap_t, vm_offset_t);
226 
227 /*
228  * Utility routines.
229  */
230 static boolean_t	moea64_query_bit(mmu_t, vm_page_t, uint64_t);
231 static u_int		moea64_clear_bit(mmu_t, vm_page_t, uint64_t);
232 static void		moea64_kremove(mmu_t, vm_offset_t);
233 static void		moea64_syncicache(mmu_t, pmap_t pmap, vm_offset_t va,
234 			    vm_paddr_t pa, vm_size_t sz);
235 static void		moea64_pmap_init_qpages(void);
236 
237 /*
238  * Kernel MMU interface
239  */
240 void moea64_clear_modify(mmu_t, vm_page_t);
241 void moea64_copy_page(mmu_t, vm_page_t, vm_page_t);
242 void moea64_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
243     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
244 int moea64_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t,
245     u_int flags, int8_t psind);
246 void moea64_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
247     vm_prot_t);
248 void moea64_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
249 vm_paddr_t moea64_extract(mmu_t, pmap_t, vm_offset_t);
250 vm_page_t moea64_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
251 void moea64_init(mmu_t);
252 boolean_t moea64_is_modified(mmu_t, vm_page_t);
253 boolean_t moea64_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
254 boolean_t moea64_is_referenced(mmu_t, vm_page_t);
255 int moea64_ts_referenced(mmu_t, vm_page_t);
256 vm_offset_t moea64_map(mmu_t, vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
257 boolean_t moea64_page_exists_quick(mmu_t, pmap_t, vm_page_t);
258 void moea64_page_init(mmu_t, vm_page_t);
259 int moea64_page_wired_mappings(mmu_t, vm_page_t);
260 void moea64_pinit(mmu_t, pmap_t);
261 void moea64_pinit0(mmu_t, pmap_t);
262 void moea64_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
263 void moea64_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
264 void moea64_qremove(mmu_t, vm_offset_t, int);
265 void moea64_release(mmu_t, pmap_t);
266 void moea64_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
267 void moea64_remove_pages(mmu_t, pmap_t);
268 void moea64_remove_all(mmu_t, vm_page_t);
269 void moea64_remove_write(mmu_t, vm_page_t);
270 void moea64_unwire(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
271 void moea64_zero_page(mmu_t, vm_page_t);
272 void moea64_zero_page_area(mmu_t, vm_page_t, int, int);
273 void moea64_activate(mmu_t, struct thread *);
274 void moea64_deactivate(mmu_t, struct thread *);
275 void *moea64_mapdev(mmu_t, vm_paddr_t, vm_size_t);
276 void *moea64_mapdev_attr(mmu_t, vm_paddr_t, vm_size_t, vm_memattr_t);
277 void moea64_unmapdev(mmu_t, vm_offset_t, vm_size_t);
278 vm_paddr_t moea64_kextract(mmu_t, vm_offset_t);
279 void moea64_page_set_memattr(mmu_t, vm_page_t m, vm_memattr_t ma);
280 void moea64_kenter_attr(mmu_t, vm_offset_t, vm_paddr_t, vm_memattr_t ma);
281 void moea64_kenter(mmu_t, vm_offset_t, vm_paddr_t);
282 boolean_t moea64_dev_direct_mapped(mmu_t, vm_paddr_t, vm_size_t);
283 static void moea64_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
284 void moea64_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz,
285     void **va);
286 void moea64_scan_init(mmu_t mmu);
287 vm_offset_t moea64_quick_enter_page(mmu_t mmu, vm_page_t m);
288 void moea64_quick_remove_page(mmu_t mmu, vm_offset_t addr);
289 static int moea64_map_user_ptr(mmu_t mmu, pmap_t pm,
290     volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen);
291 static int moea64_decode_kernel_ptr(mmu_t mmu, vm_offset_t addr,
292     int *is_user, vm_offset_t *decoded_addr);
293 
294 
295 static mmu_method_t moea64_methods[] = {
296 	MMUMETHOD(mmu_clear_modify,	moea64_clear_modify),
297 	MMUMETHOD(mmu_copy_page,	moea64_copy_page),
298 	MMUMETHOD(mmu_copy_pages,	moea64_copy_pages),
299 	MMUMETHOD(mmu_enter,		moea64_enter),
300 	MMUMETHOD(mmu_enter_object,	moea64_enter_object),
301 	MMUMETHOD(mmu_enter_quick,	moea64_enter_quick),
302 	MMUMETHOD(mmu_extract,		moea64_extract),
303 	MMUMETHOD(mmu_extract_and_hold,	moea64_extract_and_hold),
304 	MMUMETHOD(mmu_init,		moea64_init),
305 	MMUMETHOD(mmu_is_modified,	moea64_is_modified),
306 	MMUMETHOD(mmu_is_prefaultable,	moea64_is_prefaultable),
307 	MMUMETHOD(mmu_is_referenced,	moea64_is_referenced),
308 	MMUMETHOD(mmu_ts_referenced,	moea64_ts_referenced),
309 	MMUMETHOD(mmu_map,     		moea64_map),
310 	MMUMETHOD(mmu_page_exists_quick,moea64_page_exists_quick),
311 	MMUMETHOD(mmu_page_init,	moea64_page_init),
312 	MMUMETHOD(mmu_page_wired_mappings,moea64_page_wired_mappings),
313 	MMUMETHOD(mmu_pinit,		moea64_pinit),
314 	MMUMETHOD(mmu_pinit0,		moea64_pinit0),
315 	MMUMETHOD(mmu_protect,		moea64_protect),
316 	MMUMETHOD(mmu_qenter,		moea64_qenter),
317 	MMUMETHOD(mmu_qremove,		moea64_qremove),
318 	MMUMETHOD(mmu_release,		moea64_release),
319 	MMUMETHOD(mmu_remove,		moea64_remove),
320 	MMUMETHOD(mmu_remove_pages,	moea64_remove_pages),
321 	MMUMETHOD(mmu_remove_all,      	moea64_remove_all),
322 	MMUMETHOD(mmu_remove_write,	moea64_remove_write),
323 	MMUMETHOD(mmu_sync_icache,	moea64_sync_icache),
324 	MMUMETHOD(mmu_unwire,		moea64_unwire),
325 	MMUMETHOD(mmu_zero_page,       	moea64_zero_page),
326 	MMUMETHOD(mmu_zero_page_area,	moea64_zero_page_area),
327 	MMUMETHOD(mmu_activate,		moea64_activate),
328 	MMUMETHOD(mmu_deactivate,      	moea64_deactivate),
329 	MMUMETHOD(mmu_page_set_memattr,	moea64_page_set_memattr),
330 	MMUMETHOD(mmu_quick_enter_page, moea64_quick_enter_page),
331 	MMUMETHOD(mmu_quick_remove_page, moea64_quick_remove_page),
332 
333 	/* Internal interfaces */
334 	MMUMETHOD(mmu_mapdev,		moea64_mapdev),
335 	MMUMETHOD(mmu_mapdev_attr,	moea64_mapdev_attr),
336 	MMUMETHOD(mmu_unmapdev,		moea64_unmapdev),
337 	MMUMETHOD(mmu_kextract,		moea64_kextract),
338 	MMUMETHOD(mmu_kenter,		moea64_kenter),
339 	MMUMETHOD(mmu_kenter_attr,	moea64_kenter_attr),
340 	MMUMETHOD(mmu_dev_direct_mapped,moea64_dev_direct_mapped),
341 	MMUMETHOD(mmu_scan_init,	moea64_scan_init),
342 	MMUMETHOD(mmu_dumpsys_map,	moea64_dumpsys_map),
343 	MMUMETHOD(mmu_map_user_ptr,	moea64_map_user_ptr),
344 	MMUMETHOD(mmu_decode_kernel_ptr, moea64_decode_kernel_ptr),
345 
346 	{ 0, 0 }
347 };
348 
349 MMU_DEF(oea64_mmu, "mmu_oea64_base", moea64_methods, 0);
350 
351 static struct pvo_head *
352 vm_page_to_pvoh(vm_page_t m)
353 {
354 
355 	mtx_assert(PV_LOCKPTR(VM_PAGE_TO_PHYS(m)), MA_OWNED);
356 	return (&m->md.mdpg_pvoh);
357 }
358 
359 static struct pvo_entry *
360 alloc_pvo_entry(int bootstrap)
361 {
362 	struct pvo_entry *pvo;
363 
364 	if (!moea64_initialized || bootstrap) {
365 		if (moea64_bpvo_pool_index >= moea64_bpvo_pool_size) {
366 			panic("moea64_enter: bpvo pool exhausted, %d, %d, %zd",
367 			      moea64_bpvo_pool_index, moea64_bpvo_pool_size,
368 			      moea64_bpvo_pool_size * sizeof(struct pvo_entry));
369 		}
370 		pvo = &moea64_bpvo_pool[
371 		    atomic_fetchadd_int(&moea64_bpvo_pool_index, 1)];
372 		bzero(pvo, sizeof(*pvo));
373 		pvo->pvo_vaddr = PVO_BOOTSTRAP;
374 	} else {
375 		pvo = uma_zalloc(moea64_pvo_zone, M_NOWAIT);
376 		bzero(pvo, sizeof(*pvo));
377 	}
378 
379 	return (pvo);
380 }
381 
382 
383 static void
384 init_pvo_entry(struct pvo_entry *pvo, pmap_t pmap, vm_offset_t va)
385 {
386 	uint64_t vsid;
387 	uint64_t hash;
388 	int shift;
389 
390 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
391 
392 	pvo->pvo_pmap = pmap;
393 	va &= ~ADDR_POFF;
394 	pvo->pvo_vaddr |= va;
395 	vsid = va_to_vsid(pmap, va);
396 	pvo->pvo_vpn = (uint64_t)((va & ADDR_PIDX) >> ADDR_PIDX_SHFT)
397 	    | (vsid << 16);
398 
399 	shift = (pvo->pvo_vaddr & PVO_LARGE) ? moea64_large_page_shift :
400 	    ADDR_PIDX_SHFT;
401 	hash = (vsid & VSID_HASH_MASK) ^ (((uint64_t)va & ADDR_PIDX) >> shift);
402 	pvo->pvo_pte.slot = (hash & moea64_pteg_mask) << 3;
403 }
404 
405 static void
406 free_pvo_entry(struct pvo_entry *pvo)
407 {
408 
409 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
410 		uma_zfree(moea64_pvo_zone, pvo);
411 }
412 
413 void
414 moea64_pte_from_pvo(const struct pvo_entry *pvo, struct lpte *lpte)
415 {
416 
417 	lpte->pte_hi = (pvo->pvo_vpn >> (ADDR_API_SHFT64 - ADDR_PIDX_SHFT)) &
418 	    LPTE_AVPN_MASK;
419 	lpte->pte_hi |= LPTE_VALID;
420 
421 	if (pvo->pvo_vaddr & PVO_LARGE)
422 		lpte->pte_hi |= LPTE_BIG;
423 	if (pvo->pvo_vaddr & PVO_WIRED)
424 		lpte->pte_hi |= LPTE_WIRED;
425 	if (pvo->pvo_vaddr & PVO_HID)
426 		lpte->pte_hi |= LPTE_HID;
427 
428 	lpte->pte_lo = pvo->pvo_pte.pa; /* Includes WIMG bits */
429 	if (pvo->pvo_pte.prot & VM_PROT_WRITE)
430 		lpte->pte_lo |= LPTE_BW;
431 	else
432 		lpte->pte_lo |= LPTE_BR;
433 
434 	if (!(pvo->pvo_pte.prot & VM_PROT_EXECUTE))
435 		lpte->pte_lo |= LPTE_NOEXEC;
436 }
437 
438 static __inline uint64_t
439 moea64_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
440 {
441 	uint64_t pte_lo;
442 	int i;
443 
444 	if (ma != VM_MEMATTR_DEFAULT) {
445 		switch (ma) {
446 		case VM_MEMATTR_UNCACHEABLE:
447 			return (LPTE_I | LPTE_G);
448 		case VM_MEMATTR_CACHEABLE:
449 			return (LPTE_M);
450 		case VM_MEMATTR_WRITE_COMBINING:
451 		case VM_MEMATTR_WRITE_BACK:
452 		case VM_MEMATTR_PREFETCHABLE:
453 			return (LPTE_I);
454 		case VM_MEMATTR_WRITE_THROUGH:
455 			return (LPTE_W | LPTE_M);
456 		}
457 	}
458 
459 	/*
460 	 * Assume the page is cache inhibited and access is guarded unless
461 	 * it's in our available memory array.
462 	 */
463 	pte_lo = LPTE_I | LPTE_G;
464 	for (i = 0; i < pregions_sz; i++) {
465 		if ((pa >= pregions[i].mr_start) &&
466 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
467 			pte_lo &= ~(LPTE_I | LPTE_G);
468 			pte_lo |= LPTE_M;
469 			break;
470 		}
471 	}
472 
473 	return pte_lo;
474 }
475 
476 /*
477  * Quick sort callout for comparing memory regions.
478  */
479 static int	om_cmp(const void *a, const void *b);
480 
481 static int
482 om_cmp(const void *a, const void *b)
483 {
484 	const struct	ofw_map *mapa;
485 	const struct	ofw_map *mapb;
486 
487 	mapa = a;
488 	mapb = b;
489 	if (mapa->om_pa < mapb->om_pa)
490 		return (-1);
491 	else if (mapa->om_pa > mapb->om_pa)
492 		return (1);
493 	else
494 		return (0);
495 }
496 
497 static void
498 moea64_add_ofw_mappings(mmu_t mmup, phandle_t mmu, size_t sz)
499 {
500 	struct ofw_map	translations[sz/(4*sizeof(cell_t))]; /*>= 4 cells per */
501 	pcell_t		acells, trans_cells[sz/sizeof(cell_t)];
502 	struct pvo_entry *pvo;
503 	register_t	msr;
504 	vm_offset_t	off;
505 	vm_paddr_t	pa_base;
506 	int		i, j;
507 
508 	bzero(translations, sz);
509 	OF_getencprop(OF_finddevice("/"), "#address-cells", &acells,
510 	    sizeof(acells));
511 	if (OF_getencprop(mmu, "translations", trans_cells, sz) == -1)
512 		panic("moea64_bootstrap: can't get ofw translations");
513 
514 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
515 	sz /= sizeof(cell_t);
516 	for (i = 0, j = 0; i < sz; j++) {
517 		translations[j].om_va = trans_cells[i++];
518 		translations[j].om_len = trans_cells[i++];
519 		translations[j].om_pa = trans_cells[i++];
520 		if (acells == 2) {
521 			translations[j].om_pa <<= 32;
522 			translations[j].om_pa |= trans_cells[i++];
523 		}
524 		translations[j].om_mode = trans_cells[i++];
525 	}
526 	KASSERT(i == sz, ("Translations map has incorrect cell count (%d/%zd)",
527 	    i, sz));
528 
529 	sz = j;
530 	qsort(translations, sz, sizeof (*translations), om_cmp);
531 
532 	for (i = 0; i < sz; i++) {
533 		pa_base = translations[i].om_pa;
534 	      #ifndef __powerpc64__
535 		if ((translations[i].om_pa >> 32) != 0)
536 			panic("OFW translations above 32-bit boundary!");
537 	      #endif
538 
539 		if (pa_base % PAGE_SIZE)
540 			panic("OFW translation not page-aligned (phys)!");
541 		if (translations[i].om_va % PAGE_SIZE)
542 			panic("OFW translation not page-aligned (virt)!");
543 
544 		CTR3(KTR_PMAP, "translation: pa=%#zx va=%#x len=%#x",
545 		    pa_base, translations[i].om_va, translations[i].om_len);
546 
547 		/* Now enter the pages for this mapping */
548 
549 		DISABLE_TRANS(msr);
550 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
551 			/* If this address is direct-mapped, skip remapping */
552 			if (hw_direct_map &&
553 			    translations[i].om_va == PHYS_TO_DMAP(pa_base) &&
554 			    moea64_calc_wimg(pa_base + off, VM_MEMATTR_DEFAULT)
555  			    == LPTE_M)
556 				continue;
557 
558 			PMAP_LOCK(kernel_pmap);
559 			pvo = moea64_pvo_find_va(kernel_pmap,
560 			    translations[i].om_va + off);
561 			PMAP_UNLOCK(kernel_pmap);
562 			if (pvo != NULL)
563 				continue;
564 
565 			moea64_kenter(mmup, translations[i].om_va + off,
566 			    pa_base + off);
567 		}
568 		ENABLE_TRANS(msr);
569 	}
570 }
571 
572 #ifdef __powerpc64__
573 static void
574 moea64_probe_large_page(void)
575 {
576 	uint16_t pvr = mfpvr() >> 16;
577 
578 	switch (pvr) {
579 	case IBM970:
580 	case IBM970FX:
581 	case IBM970MP:
582 		powerpc_sync(); isync();
583 		mtspr(SPR_HID4, mfspr(SPR_HID4) & ~HID4_970_DISABLE_LG_PG);
584 		powerpc_sync(); isync();
585 
586 		/* FALLTHROUGH */
587 	default:
588 		if (moea64_large_page_size == 0) {
589 			moea64_large_page_size = 0x1000000; /* 16 MB */
590 			moea64_large_page_shift = 24;
591 		}
592 	}
593 
594 	moea64_large_page_mask = moea64_large_page_size - 1;
595 }
596 
597 static void
598 moea64_bootstrap_slb_prefault(vm_offset_t va, int large)
599 {
600 	struct slb *cache;
601 	struct slb entry;
602 	uint64_t esid, slbe;
603 	uint64_t i;
604 
605 	cache = PCPU_GET(aim.slb);
606 	esid = va >> ADDR_SR_SHFT;
607 	slbe = (esid << SLBE_ESID_SHIFT) | SLBE_VALID;
608 
609 	for (i = 0; i < 64; i++) {
610 		if (cache[i].slbe == (slbe | i))
611 			return;
612 	}
613 
614 	entry.slbe = slbe;
615 	entry.slbv = KERNEL_VSID(esid) << SLBV_VSID_SHIFT;
616 	if (large)
617 		entry.slbv |= SLBV_L;
618 
619 	slb_insert_kernel(entry.slbe, entry.slbv);
620 }
621 #endif
622 
623 static void
624 moea64_setup_direct_map(mmu_t mmup, vm_offset_t kernelstart,
625     vm_offset_t kernelend)
626 {
627 	struct pvo_entry *pvo;
628 	register_t msr;
629 	vm_paddr_t pa;
630 	vm_offset_t size, off;
631 	uint64_t pte_lo;
632 	int i;
633 
634 	if (moea64_large_page_size == 0)
635 		hw_direct_map = 0;
636 
637 	DISABLE_TRANS(msr);
638 	if (hw_direct_map) {
639 		PMAP_LOCK(kernel_pmap);
640 		for (i = 0; i < pregions_sz; i++) {
641 		  for (pa = pregions[i].mr_start; pa < pregions[i].mr_start +
642 		     pregions[i].mr_size; pa += moea64_large_page_size) {
643 			pte_lo = LPTE_M;
644 
645 			pvo = alloc_pvo_entry(1 /* bootstrap */);
646 			pvo->pvo_vaddr |= PVO_WIRED | PVO_LARGE;
647 			init_pvo_entry(pvo, kernel_pmap, PHYS_TO_DMAP(pa));
648 
649 			/*
650 			 * Set memory access as guarded if prefetch within
651 			 * the page could exit the available physmem area.
652 			 */
653 			if (pa & moea64_large_page_mask) {
654 				pa &= moea64_large_page_mask;
655 				pte_lo |= LPTE_G;
656 			}
657 			if (pa + moea64_large_page_size >
658 			    pregions[i].mr_start + pregions[i].mr_size)
659 				pte_lo |= LPTE_G;
660 
661 			pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE |
662 			    VM_PROT_EXECUTE;
663 			pvo->pvo_pte.pa = pa | pte_lo;
664 			moea64_pvo_enter(mmup, pvo, NULL);
665 		  }
666 		}
667 		PMAP_UNLOCK(kernel_pmap);
668 	}
669 
670 	/*
671 	 * Make sure the kernel and BPVO pool stay mapped on systems either
672 	 * without a direct map or on which the kernel is not already executing
673 	 * out of the direct-mapped region.
674 	 */
675 
676 	if (!hw_direct_map || kernelstart < DMAP_BASE_ADDRESS) {
677 		for (pa = kernelstart & ~PAGE_MASK; pa < kernelend;
678 		    pa += PAGE_SIZE)
679 			moea64_kenter(mmup, pa, pa);
680 	}
681 
682 	if (!hw_direct_map) {
683 		size = moea64_bpvo_pool_size*sizeof(struct pvo_entry);
684 		off = (vm_offset_t)(moea64_bpvo_pool);
685 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
686 			moea64_kenter(mmup, pa, pa);
687 	}
688 	ENABLE_TRANS(msr);
689 
690 	/*
691 	 * Allow user to override unmapped_buf_allowed for testing.
692 	 * XXXKIB Only direct map implementation was tested.
693 	 */
694 	if (!TUNABLE_INT_FETCH("vfs.unmapped_buf_allowed",
695 	    &unmapped_buf_allowed))
696 		unmapped_buf_allowed = hw_direct_map;
697 }
698 
699 void
700 moea64_early_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
701 {
702 	int		i, j;
703 	vm_size_t	physsz, hwphyssz;
704 	vm_paddr_t	kernelphysstart, kernelphysend;
705 
706 #ifndef __powerpc64__
707 	/* We don't have a direct map since there is no BAT */
708 	hw_direct_map = 0;
709 
710 	/* Make sure battable is zero, since we have no BAT */
711 	for (i = 0; i < 16; i++) {
712 		battable[i].batu = 0;
713 		battable[i].batl = 0;
714 	}
715 #else
716 	moea64_probe_large_page();
717 
718 	/* Use a direct map if we have large page support */
719 	if (moea64_large_page_size > 0)
720 		hw_direct_map = 1;
721 	else
722 		hw_direct_map = 0;
723 
724 	/* Install trap handlers for SLBs */
725 	bcopy(&slbtrap, (void *)EXC_DSE,(size_t)&slbtrapend - (size_t)&slbtrap);
726 	bcopy(&slbtrap, (void *)EXC_ISE,(size_t)&slbtrapend - (size_t)&slbtrap);
727 	__syncicache((void *)EXC_DSE, 0x80);
728 	__syncicache((void *)EXC_ISE, 0x80);
729 #endif
730 
731 	kernelphysstart = kernelstart & ~DMAP_BASE_ADDRESS;
732 	kernelphysend = kernelend & ~DMAP_BASE_ADDRESS;
733 
734 	/* Get physical memory regions from firmware */
735 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
736 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
737 
738 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
739 		panic("moea64_bootstrap: phys_avail too small");
740 
741 	phys_avail_count = 0;
742 	physsz = 0;
743 	hwphyssz = 0;
744 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
745 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
746 		CTR3(KTR_PMAP, "region: %#zx - %#zx (%#zx)",
747 		    regions[i].mr_start, regions[i].mr_start +
748 		    regions[i].mr_size, regions[i].mr_size);
749 		if (hwphyssz != 0 &&
750 		    (physsz + regions[i].mr_size) >= hwphyssz) {
751 			if (physsz < hwphyssz) {
752 				phys_avail[j] = regions[i].mr_start;
753 				phys_avail[j + 1] = regions[i].mr_start +
754 				    hwphyssz - physsz;
755 				physsz = hwphyssz;
756 				phys_avail_count++;
757 			}
758 			break;
759 		}
760 		phys_avail[j] = regions[i].mr_start;
761 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
762 		phys_avail_count++;
763 		physsz += regions[i].mr_size;
764 	}
765 
766 	/* Check for overlap with the kernel and exception vectors */
767 	for (j = 0; j < 2*phys_avail_count; j+=2) {
768 		if (phys_avail[j] < EXC_LAST)
769 			phys_avail[j] += EXC_LAST;
770 
771 		if (kernelphysstart >= phys_avail[j] &&
772 		    kernelphysstart < phys_avail[j+1]) {
773 			if (kernelphysend < phys_avail[j+1]) {
774 				phys_avail[2*phys_avail_count] =
775 				    (kernelphysend & ~PAGE_MASK) + PAGE_SIZE;
776 				phys_avail[2*phys_avail_count + 1] =
777 				    phys_avail[j+1];
778 				phys_avail_count++;
779 			}
780 
781 			phys_avail[j+1] = kernelphysstart & ~PAGE_MASK;
782 		}
783 
784 		if (kernelphysend >= phys_avail[j] &&
785 		    kernelphysend < phys_avail[j+1]) {
786 			if (kernelphysstart > phys_avail[j]) {
787 				phys_avail[2*phys_avail_count] = phys_avail[j];
788 				phys_avail[2*phys_avail_count + 1] =
789 				    kernelphysstart & ~PAGE_MASK;
790 				phys_avail_count++;
791 			}
792 
793 			phys_avail[j] = (kernelphysend & ~PAGE_MASK) +
794 			    PAGE_SIZE;
795 		}
796 	}
797 
798 	physmem = btoc(physsz);
799 
800 #ifdef PTEGCOUNT
801 	moea64_pteg_count = PTEGCOUNT;
802 #else
803 	moea64_pteg_count = 0x1000;
804 
805 	while (moea64_pteg_count < physmem)
806 		moea64_pteg_count <<= 1;
807 
808 	moea64_pteg_count >>= 1;
809 #endif /* PTEGCOUNT */
810 }
811 
812 void
813 moea64_mid_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
814 {
815 	int		i;
816 
817 	/*
818 	 * Set PTEG mask
819 	 */
820 	moea64_pteg_mask = moea64_pteg_count - 1;
821 
822 	/*
823 	 * Initialize SLB table lock and page locks
824 	 */
825 	mtx_init(&moea64_slb_mutex, "SLB table", NULL, MTX_DEF);
826 	for (i = 0; i < PV_LOCK_COUNT; i++)
827 		mtx_init(&pv_lock[i], "page pv", NULL, MTX_DEF);
828 
829 	/*
830 	 * Initialise the bootstrap pvo pool.
831 	 */
832 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
833 		moea64_bpvo_pool_size*sizeof(struct pvo_entry), 0);
834 	moea64_bpvo_pool_index = 0;
835 
836 	/* Place at address usable through the direct map */
837 	if (hw_direct_map)
838 		moea64_bpvo_pool = (struct pvo_entry *)
839 		    PHYS_TO_DMAP((uintptr_t)moea64_bpvo_pool);
840 
841 	/*
842 	 * Make sure kernel vsid is allocated as well as VSID 0.
843 	 */
844 	#ifndef __powerpc64__
845 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NVSIDS - 1)) / VSID_NBPW]
846 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
847 	moea64_vsid_bitmap[0] |= 1;
848 	#endif
849 
850 	/*
851 	 * Initialize the kernel pmap (which is statically allocated).
852 	 */
853 	#ifdef __powerpc64__
854 	for (i = 0; i < 64; i++) {
855 		pcpup->pc_aim.slb[i].slbv = 0;
856 		pcpup->pc_aim.slb[i].slbe = 0;
857 	}
858 	#else
859 	for (i = 0; i < 16; i++)
860 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
861 	#endif
862 
863 	kernel_pmap->pmap_phys = kernel_pmap;
864 	CPU_FILL(&kernel_pmap->pm_active);
865 	RB_INIT(&kernel_pmap->pmap_pvo);
866 
867 	PMAP_LOCK_INIT(kernel_pmap);
868 
869 	/*
870 	 * Now map in all the other buffers we allocated earlier
871 	 */
872 
873 	moea64_setup_direct_map(mmup, kernelstart, kernelend);
874 }
875 
876 void
877 moea64_late_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
878 {
879 	ihandle_t	mmui;
880 	phandle_t	chosen;
881 	phandle_t	mmu;
882 	ssize_t		sz;
883 	int		i;
884 	vm_offset_t	pa, va;
885 	void		*dpcpu;
886 
887 	/*
888 	 * Set up the Open Firmware pmap and add its mappings if not in real
889 	 * mode.
890 	 */
891 
892 	chosen = OF_finddevice("/chosen");
893 	if (chosen != -1 && OF_getencprop(chosen, "mmu", &mmui, 4) != -1) {
894 		mmu = OF_instance_to_package(mmui);
895 		if (mmu == -1 ||
896 		    (sz = OF_getproplen(mmu, "translations")) == -1)
897 			sz = 0;
898 		if (sz > 6144 /* tmpstksz - 2 KB headroom */)
899 			panic("moea64_bootstrap: too many ofw translations");
900 
901 		if (sz > 0)
902 			moea64_add_ofw_mappings(mmup, mmu, sz);
903 	}
904 
905 	/*
906 	 * Calculate the last available physical address.
907 	 */
908 	Maxmem = 0;
909 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
910 		Maxmem = max(Maxmem, powerpc_btop(phys_avail[i + 1]));
911 
912 	/*
913 	 * Initialize MMU.
914 	 */
915 	MMU_CPU_BOOTSTRAP(mmup,0);
916 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
917 	pmap_bootstrapped++;
918 
919 	/*
920 	 * Set the start and end of kva.
921 	 */
922 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
923 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
924 
925 	/*
926 	 * Map the entire KVA range into the SLB. We must not fault there.
927 	 */
928 	#ifdef __powerpc64__
929 	for (va = virtual_avail; va < virtual_end; va += SEGMENT_LENGTH)
930 		moea64_bootstrap_slb_prefault(va, 0);
931 	#endif
932 
933 	/*
934 	 * Remap any early IO mappings (console framebuffer, etc.)
935 	 */
936 	bs_remap_earlyboot();
937 
938 	/*
939 	 * Figure out how far we can extend virtual_end into segment 16
940 	 * without running into existing mappings. Segment 16 is guaranteed
941 	 * to contain neither RAM nor devices (at least on Apple hardware),
942 	 * but will generally contain some OFW mappings we should not
943 	 * step on.
944 	 */
945 
946 	#ifndef __powerpc64__	/* KVA is in high memory on PPC64 */
947 	PMAP_LOCK(kernel_pmap);
948 	while (virtual_end < VM_MAX_KERNEL_ADDRESS &&
949 	    moea64_pvo_find_va(kernel_pmap, virtual_end+1) == NULL)
950 		virtual_end += PAGE_SIZE;
951 	PMAP_UNLOCK(kernel_pmap);
952 	#endif
953 
954 	/*
955 	 * Allocate a kernel stack with a guard page for thread0 and map it
956 	 * into the kernel page map.
957 	 */
958 	pa = moea64_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
959 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
960 	virtual_avail = va + kstack_pages * PAGE_SIZE;
961 	CTR2(KTR_PMAP, "moea64_bootstrap: kstack0 at %#x (%#x)", pa, va);
962 	thread0.td_kstack = va;
963 	thread0.td_kstack_pages = kstack_pages;
964 	for (i = 0; i < kstack_pages; i++) {
965 		moea64_kenter(mmup, va, pa);
966 		pa += PAGE_SIZE;
967 		va += PAGE_SIZE;
968 	}
969 
970 	/*
971 	 * Allocate virtual address space for the message buffer.
972 	 */
973 	pa = msgbuf_phys = moea64_bootstrap_alloc(msgbufsize, PAGE_SIZE);
974 	msgbufp = (struct msgbuf *)virtual_avail;
975 	va = virtual_avail;
976 	virtual_avail += round_page(msgbufsize);
977 	while (va < virtual_avail) {
978 		moea64_kenter(mmup, va, pa);
979 		pa += PAGE_SIZE;
980 		va += PAGE_SIZE;
981 	}
982 
983 	/*
984 	 * Allocate virtual address space for the dynamic percpu area.
985 	 */
986 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
987 	dpcpu = (void *)virtual_avail;
988 	va = virtual_avail;
989 	virtual_avail += DPCPU_SIZE;
990 	while (va < virtual_avail) {
991 		moea64_kenter(mmup, va, pa);
992 		pa += PAGE_SIZE;
993 		va += PAGE_SIZE;
994 	}
995 	dpcpu_init(dpcpu, curcpu);
996 
997 	/*
998 	 * Allocate some things for page zeroing. We put this directly
999 	 * in the page table and use MOEA64_PTE_REPLACE to avoid any
1000 	 * of the PVO book-keeping or other parts of the VM system
1001 	 * from even knowing that this hack exists.
1002 	 */
1003 
1004 	if (!hw_direct_map) {
1005 		mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL,
1006 		    MTX_DEF);
1007 		for (i = 0; i < 2; i++) {
1008 			moea64_scratchpage_va[i] = (virtual_end+1) - PAGE_SIZE;
1009 			virtual_end -= PAGE_SIZE;
1010 
1011 			moea64_kenter(mmup, moea64_scratchpage_va[i], 0);
1012 
1013 			PMAP_LOCK(kernel_pmap);
1014 			moea64_scratchpage_pvo[i] = moea64_pvo_find_va(
1015 			    kernel_pmap, (vm_offset_t)moea64_scratchpage_va[i]);
1016 			PMAP_UNLOCK(kernel_pmap);
1017 		}
1018 	}
1019 }
1020 
1021 static void
1022 moea64_pmap_init_qpages(void)
1023 {
1024 	struct pcpu *pc;
1025 	int i;
1026 
1027 	if (hw_direct_map)
1028 		return;
1029 
1030 	CPU_FOREACH(i) {
1031 		pc = pcpu_find(i);
1032 		pc->pc_qmap_addr = kva_alloc(PAGE_SIZE);
1033 		if (pc->pc_qmap_addr == 0)
1034 			panic("pmap_init_qpages: unable to allocate KVA");
1035 		PMAP_LOCK(kernel_pmap);
1036 		pc->pc_aim.qmap_pvo =
1037 		    moea64_pvo_find_va(kernel_pmap, pc->pc_qmap_addr);
1038 		PMAP_UNLOCK(kernel_pmap);
1039 		mtx_init(&pc->pc_aim.qmap_lock, "qmap lock", NULL, MTX_DEF);
1040 	}
1041 }
1042 
1043 SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, moea64_pmap_init_qpages, NULL);
1044 
1045 /*
1046  * Activate a user pmap.  This mostly involves setting some non-CPU
1047  * state.
1048  */
1049 void
1050 moea64_activate(mmu_t mmu, struct thread *td)
1051 {
1052 	pmap_t	pm;
1053 
1054 	pm = &td->td_proc->p_vmspace->vm_pmap;
1055 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1056 
1057 	#ifdef __powerpc64__
1058 	PCPU_SET(aim.userslb, pm->pm_slb);
1059 	__asm __volatile("slbmte %0, %1; isync" ::
1060 	    "r"(td->td_pcb->pcb_cpu.aim.usr_vsid), "r"(USER_SLB_SLBE));
1061 	#else
1062 	PCPU_SET(curpmap, pm->pmap_phys);
1063 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1064 	#endif
1065 }
1066 
1067 void
1068 moea64_deactivate(mmu_t mmu, struct thread *td)
1069 {
1070 	pmap_t	pm;
1071 
1072 	__asm __volatile("isync; slbie %0" :: "r"(USER_ADDR));
1073 
1074 	pm = &td->td_proc->p_vmspace->vm_pmap;
1075 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1076 	#ifdef __powerpc64__
1077 	PCPU_SET(aim.userslb, NULL);
1078 	#else
1079 	PCPU_SET(curpmap, NULL);
1080 	#endif
1081 }
1082 
1083 void
1084 moea64_unwire(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1085 {
1086 	struct	pvo_entry key, *pvo;
1087 	vm_page_t m;
1088 	int64_t	refchg;
1089 
1090 	key.pvo_vaddr = sva;
1091 	PMAP_LOCK(pm);
1092 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1093 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1094 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1095 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1096 			panic("moea64_unwire: pvo %p is missing PVO_WIRED",
1097 			    pvo);
1098 		pvo->pvo_vaddr &= ~PVO_WIRED;
1099 		refchg = MOEA64_PTE_REPLACE(mmu, pvo, 0 /* No invalidation */);
1100 		if ((pvo->pvo_vaddr & PVO_MANAGED) &&
1101 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1102 			if (refchg < 0)
1103 				refchg = LPTE_CHG;
1104 			m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
1105 
1106 			refchg |= atomic_readandclear_32(&m->md.mdpg_attrs);
1107 			if (refchg & LPTE_CHG)
1108 				vm_page_dirty(m);
1109 			if (refchg & LPTE_REF)
1110 				vm_page_aflag_set(m, PGA_REFERENCED);
1111 		}
1112 		pm->pm_stats.wired_count--;
1113 	}
1114 	PMAP_UNLOCK(pm);
1115 }
1116 
1117 /*
1118  * This goes through and sets the physical address of our
1119  * special scratch PTE to the PA we want to zero or copy. Because
1120  * of locking issues (this can get called in pvo_enter() by
1121  * the UMA allocator), we can't use most other utility functions here
1122  */
1123 
1124 static __inline
1125 void moea64_set_scratchpage_pa(mmu_t mmup, int which, vm_paddr_t pa) {
1126 
1127 	KASSERT(!hw_direct_map, ("Using OEA64 scratchpage with a direct map!"));
1128 	mtx_assert(&moea64_scratchpage_mtx, MA_OWNED);
1129 
1130 	moea64_scratchpage_pvo[which]->pvo_pte.pa =
1131 	    moea64_calc_wimg(pa, VM_MEMATTR_DEFAULT) | (uint64_t)pa;
1132 	MOEA64_PTE_REPLACE(mmup, moea64_scratchpage_pvo[which],
1133 	    MOEA64_PTE_INVALIDATE);
1134 	isync();
1135 }
1136 
1137 void
1138 moea64_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1139 {
1140 	vm_offset_t	dst;
1141 	vm_offset_t	src;
1142 
1143 	dst = VM_PAGE_TO_PHYS(mdst);
1144 	src = VM_PAGE_TO_PHYS(msrc);
1145 
1146 	if (hw_direct_map) {
1147 		bcopy((void *)PHYS_TO_DMAP(src), (void *)PHYS_TO_DMAP(dst),
1148 		    PAGE_SIZE);
1149 	} else {
1150 		mtx_lock(&moea64_scratchpage_mtx);
1151 
1152 		moea64_set_scratchpage_pa(mmu, 0, src);
1153 		moea64_set_scratchpage_pa(mmu, 1, dst);
1154 
1155 		bcopy((void *)moea64_scratchpage_va[0],
1156 		    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1157 
1158 		mtx_unlock(&moea64_scratchpage_mtx);
1159 	}
1160 }
1161 
1162 static inline void
1163 moea64_copy_pages_dmap(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1164     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1165 {
1166 	void *a_cp, *b_cp;
1167 	vm_offset_t a_pg_offset, b_pg_offset;
1168 	int cnt;
1169 
1170 	while (xfersize > 0) {
1171 		a_pg_offset = a_offset & PAGE_MASK;
1172 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1173 		a_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1174 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT])) +
1175 		    a_pg_offset;
1176 		b_pg_offset = b_offset & PAGE_MASK;
1177 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1178 		b_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1179 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT])) +
1180 		    b_pg_offset;
1181 		bcopy(a_cp, b_cp, cnt);
1182 		a_offset += cnt;
1183 		b_offset += cnt;
1184 		xfersize -= cnt;
1185 	}
1186 }
1187 
1188 static inline void
1189 moea64_copy_pages_nodmap(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1190     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1191 {
1192 	void *a_cp, *b_cp;
1193 	vm_offset_t a_pg_offset, b_pg_offset;
1194 	int cnt;
1195 
1196 	mtx_lock(&moea64_scratchpage_mtx);
1197 	while (xfersize > 0) {
1198 		a_pg_offset = a_offset & PAGE_MASK;
1199 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1200 		moea64_set_scratchpage_pa(mmu, 0,
1201 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]));
1202 		a_cp = (char *)moea64_scratchpage_va[0] + a_pg_offset;
1203 		b_pg_offset = b_offset & PAGE_MASK;
1204 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1205 		moea64_set_scratchpage_pa(mmu, 1,
1206 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]));
1207 		b_cp = (char *)moea64_scratchpage_va[1] + b_pg_offset;
1208 		bcopy(a_cp, b_cp, cnt);
1209 		a_offset += cnt;
1210 		b_offset += cnt;
1211 		xfersize -= cnt;
1212 	}
1213 	mtx_unlock(&moea64_scratchpage_mtx);
1214 }
1215 
1216 void
1217 moea64_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1218     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1219 {
1220 
1221 	if (hw_direct_map) {
1222 		moea64_copy_pages_dmap(mmu, ma, a_offset, mb, b_offset,
1223 		    xfersize);
1224 	} else {
1225 		moea64_copy_pages_nodmap(mmu, ma, a_offset, mb, b_offset,
1226 		    xfersize);
1227 	}
1228 }
1229 
1230 void
1231 moea64_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1232 {
1233 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1234 
1235 	if (size + off > PAGE_SIZE)
1236 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1237 
1238 	if (hw_direct_map) {
1239 		bzero((caddr_t)(uintptr_t)PHYS_TO_DMAP(pa) + off, size);
1240 	} else {
1241 		mtx_lock(&moea64_scratchpage_mtx);
1242 		moea64_set_scratchpage_pa(mmu, 0, pa);
1243 		bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1244 		mtx_unlock(&moea64_scratchpage_mtx);
1245 	}
1246 }
1247 
1248 /*
1249  * Zero a page of physical memory by temporarily mapping it
1250  */
1251 void
1252 moea64_zero_page(mmu_t mmu, vm_page_t m)
1253 {
1254 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1255 	vm_offset_t va, off;
1256 
1257 	if (!hw_direct_map) {
1258 		mtx_lock(&moea64_scratchpage_mtx);
1259 
1260 		moea64_set_scratchpage_pa(mmu, 0, pa);
1261 		va = moea64_scratchpage_va[0];
1262 	} else {
1263 		va = PHYS_TO_DMAP(pa);
1264 	}
1265 
1266 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1267 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1268 
1269 	if (!hw_direct_map)
1270 		mtx_unlock(&moea64_scratchpage_mtx);
1271 }
1272 
1273 vm_offset_t
1274 moea64_quick_enter_page(mmu_t mmu, vm_page_t m)
1275 {
1276 	struct pvo_entry *pvo;
1277 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1278 
1279 	if (hw_direct_map)
1280 		return (PHYS_TO_DMAP(pa));
1281 
1282 	/*
1283  	 * MOEA64_PTE_REPLACE does some locking, so we can't just grab
1284 	 * a critical section and access the PCPU data like on i386.
1285 	 * Instead, pin the thread and grab the PCPU lock to prevent
1286 	 * a preempting thread from using the same PCPU data.
1287 	 */
1288 	sched_pin();
1289 
1290 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_NOTOWNED);
1291 	pvo = PCPU_GET(aim.qmap_pvo);
1292 
1293 	mtx_lock(PCPU_PTR(aim.qmap_lock));
1294 	pvo->pvo_pte.pa = moea64_calc_wimg(pa, pmap_page_get_memattr(m)) |
1295 	    (uint64_t)pa;
1296 	MOEA64_PTE_REPLACE(mmu, pvo, MOEA64_PTE_INVALIDATE);
1297 	isync();
1298 
1299 	return (PCPU_GET(qmap_addr));
1300 }
1301 
1302 void
1303 moea64_quick_remove_page(mmu_t mmu, vm_offset_t addr)
1304 {
1305 	if (hw_direct_map)
1306 		return;
1307 
1308 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_OWNED);
1309 	KASSERT(PCPU_GET(qmap_addr) == addr,
1310 	    ("moea64_quick_remove_page: invalid address"));
1311 	mtx_unlock(PCPU_PTR(aim.qmap_lock));
1312 	sched_unpin();
1313 }
1314 
1315 /*
1316  * Map the given physical page at the specified virtual address in the
1317  * target pmap with the protection requested.  If specified the page
1318  * will be wired down.
1319  */
1320 
1321 int
1322 moea64_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m,
1323     vm_prot_t prot, u_int flags, int8_t psind)
1324 {
1325 	struct		pvo_entry *pvo, *oldpvo;
1326 	struct		pvo_head *pvo_head;
1327 	uint64_t	pte_lo;
1328 	int		error;
1329 
1330 	if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m))
1331 		VM_OBJECT_ASSERT_LOCKED(m->object);
1332 
1333 	pvo = alloc_pvo_entry(0);
1334 	pvo->pvo_pmap = NULL; /* to be filled in later */
1335 	pvo->pvo_pte.prot = prot;
1336 
1337 	pte_lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1338 	pvo->pvo_pte.pa = VM_PAGE_TO_PHYS(m) | pte_lo;
1339 
1340 	if ((flags & PMAP_ENTER_WIRED) != 0)
1341 		pvo->pvo_vaddr |= PVO_WIRED;
1342 
1343 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea64_initialized) {
1344 		pvo_head = NULL;
1345 	} else {
1346 		pvo_head = &m->md.mdpg_pvoh;
1347 		pvo->pvo_vaddr |= PVO_MANAGED;
1348 	}
1349 
1350 	for (;;) {
1351 		PV_PAGE_LOCK(m);
1352 		PMAP_LOCK(pmap);
1353 		if (pvo->pvo_pmap == NULL)
1354 			init_pvo_entry(pvo, pmap, va);
1355 		if (prot & VM_PROT_WRITE)
1356 			if (pmap_bootstrapped &&
1357 			    (m->oflags & VPO_UNMANAGED) == 0)
1358 				vm_page_aflag_set(m, PGA_WRITEABLE);
1359 
1360 		oldpvo = moea64_pvo_find_va(pmap, va);
1361 		if (oldpvo != NULL) {
1362 			if (oldpvo->pvo_vaddr == pvo->pvo_vaddr &&
1363 			    oldpvo->pvo_pte.pa == pvo->pvo_pte.pa &&
1364 			    oldpvo->pvo_pte.prot == prot) {
1365 				/* Identical mapping already exists */
1366 				error = 0;
1367 
1368 				/* If not in page table, reinsert it */
1369 				if (MOEA64_PTE_SYNCH(mmu, oldpvo) < 0) {
1370 					moea64_pte_overflow--;
1371 					MOEA64_PTE_INSERT(mmu, oldpvo);
1372 				}
1373 
1374 				/* Then just clean up and go home */
1375 				PV_PAGE_UNLOCK(m);
1376 				PMAP_UNLOCK(pmap);
1377 				free_pvo_entry(pvo);
1378 				break;
1379 			}
1380 
1381 			/* Otherwise, need to kill it first */
1382 			KASSERT(oldpvo->pvo_pmap == pmap, ("pmap of old "
1383 			    "mapping does not match new mapping"));
1384 			moea64_pvo_remove_from_pmap(mmu, oldpvo);
1385 		}
1386 		error = moea64_pvo_enter(mmu, pvo, pvo_head);
1387 		PV_PAGE_UNLOCK(m);
1388 		PMAP_UNLOCK(pmap);
1389 
1390 		/* Free any dead pages */
1391 		if (oldpvo != NULL) {
1392 			PV_LOCK(oldpvo->pvo_pte.pa & LPTE_RPGN);
1393 			moea64_pvo_remove_from_page(mmu, oldpvo);
1394 			PV_UNLOCK(oldpvo->pvo_pte.pa & LPTE_RPGN);
1395 			free_pvo_entry(oldpvo);
1396 		}
1397 
1398 		if (error != ENOMEM)
1399 			break;
1400 		if ((flags & PMAP_ENTER_NOSLEEP) != 0)
1401 			return (KERN_RESOURCE_SHORTAGE);
1402 		VM_OBJECT_ASSERT_UNLOCKED(m->object);
1403 		vm_wait(NULL);
1404 	}
1405 
1406 	/*
1407 	 * Flush the page from the instruction cache if this page is
1408 	 * mapped executable and cacheable.
1409 	 */
1410 	if (pmap != kernel_pmap && !(m->aflags & PGA_EXECUTABLE) &&
1411 	    (pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1412 		vm_page_aflag_set(m, PGA_EXECUTABLE);
1413 		moea64_syncicache(mmu, pmap, va, VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1414 	}
1415 	return (KERN_SUCCESS);
1416 }
1417 
1418 static void
1419 moea64_syncicache(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
1420     vm_size_t sz)
1421 {
1422 
1423 	/*
1424 	 * This is much trickier than on older systems because
1425 	 * we can't sync the icache on physical addresses directly
1426 	 * without a direct map. Instead we check a couple of cases
1427 	 * where the memory is already mapped in and, failing that,
1428 	 * use the same trick we use for page zeroing to create
1429 	 * a temporary mapping for this physical address.
1430 	 */
1431 
1432 	if (!pmap_bootstrapped) {
1433 		/*
1434 		 * If PMAP is not bootstrapped, we are likely to be
1435 		 * in real mode.
1436 		 */
1437 		__syncicache((void *)(uintptr_t)pa, sz);
1438 	} else if (pmap == kernel_pmap) {
1439 		__syncicache((void *)va, sz);
1440 	} else if (hw_direct_map) {
1441 		__syncicache((void *)(uintptr_t)PHYS_TO_DMAP(pa), sz);
1442 	} else {
1443 		/* Use the scratch page to set up a temp mapping */
1444 
1445 		mtx_lock(&moea64_scratchpage_mtx);
1446 
1447 		moea64_set_scratchpage_pa(mmu, 1, pa & ~ADDR_POFF);
1448 		__syncicache((void *)(moea64_scratchpage_va[1] +
1449 		    (va & ADDR_POFF)), sz);
1450 
1451 		mtx_unlock(&moea64_scratchpage_mtx);
1452 	}
1453 }
1454 
1455 /*
1456  * Maps a sequence of resident pages belonging to the same object.
1457  * The sequence begins with the given page m_start.  This page is
1458  * mapped at the given virtual address start.  Each subsequent page is
1459  * mapped at a virtual address that is offset from start by the same
1460  * amount as the page is offset from m_start within the object.  The
1461  * last page in the sequence is the page with the largest offset from
1462  * m_start that can be mapped at a virtual address less than the given
1463  * virtual address end.  Not every virtual page between start and end
1464  * is mapped; only those for which a resident page exists with the
1465  * corresponding offset from m_start are mapped.
1466  */
1467 void
1468 moea64_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1469     vm_page_t m_start, vm_prot_t prot)
1470 {
1471 	vm_page_t m;
1472 	vm_pindex_t diff, psize;
1473 
1474 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1475 
1476 	psize = atop(end - start);
1477 	m = m_start;
1478 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1479 		moea64_enter(mmu, pm, start + ptoa(diff), m, prot &
1480 		    (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_NOSLEEP, 0);
1481 		m = TAILQ_NEXT(m, listq);
1482 	}
1483 }
1484 
1485 void
1486 moea64_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1487     vm_prot_t prot)
1488 {
1489 
1490 	moea64_enter(mmu, pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1491 	    PMAP_ENTER_NOSLEEP, 0);
1492 }
1493 
1494 vm_paddr_t
1495 moea64_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1496 {
1497 	struct	pvo_entry *pvo;
1498 	vm_paddr_t pa;
1499 
1500 	PMAP_LOCK(pm);
1501 	pvo = moea64_pvo_find_va(pm, va);
1502 	if (pvo == NULL)
1503 		pa = 0;
1504 	else
1505 		pa = (pvo->pvo_pte.pa & LPTE_RPGN) | (va - PVO_VADDR(pvo));
1506 	PMAP_UNLOCK(pm);
1507 
1508 	return (pa);
1509 }
1510 
1511 /*
1512  * Atomically extract and hold the physical page with the given
1513  * pmap and virtual address pair if that mapping permits the given
1514  * protection.
1515  */
1516 vm_page_t
1517 moea64_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1518 {
1519 	struct	pvo_entry *pvo;
1520 	vm_page_t m;
1521         vm_paddr_t pa;
1522 
1523 	m = NULL;
1524 	pa = 0;
1525 	PMAP_LOCK(pmap);
1526 retry:
1527 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1528 	if (pvo != NULL && (pvo->pvo_pte.prot & prot) == prot) {
1529 		if (vm_page_pa_tryrelock(pmap,
1530 		    pvo->pvo_pte.pa & LPTE_RPGN, &pa))
1531 			goto retry;
1532 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
1533 		vm_page_hold(m);
1534 	}
1535 	PA_UNLOCK_COND(pa);
1536 	PMAP_UNLOCK(pmap);
1537 	return (m);
1538 }
1539 
1540 static mmu_t installed_mmu;
1541 
1542 static void *
1543 moea64_uma_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain,
1544     uint8_t *flags, int wait)
1545 {
1546 	struct pvo_entry *pvo;
1547         vm_offset_t va;
1548         vm_page_t m;
1549         int needed_lock;
1550 
1551 	/*
1552 	 * This entire routine is a horrible hack to avoid bothering kmem
1553 	 * for new KVA addresses. Because this can get called from inside
1554 	 * kmem allocation routines, calling kmem for a new address here
1555 	 * can lead to multiply locking non-recursive mutexes.
1556 	 */
1557 
1558 	*flags = UMA_SLAB_PRIV;
1559 	needed_lock = !PMAP_LOCKED(kernel_pmap);
1560 
1561 	m = vm_page_alloc_domain(NULL, 0, domain,
1562 	    malloc2vm_flags(wait) | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1563 	if (m == NULL)
1564 		return (NULL);
1565 
1566 	va = VM_PAGE_TO_PHYS(m);
1567 
1568 	pvo = alloc_pvo_entry(1 /* bootstrap */);
1569 
1570 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE;
1571 	pvo->pvo_pte.pa = VM_PAGE_TO_PHYS(m) | LPTE_M;
1572 
1573 	if (needed_lock)
1574 		PMAP_LOCK(kernel_pmap);
1575 
1576 	init_pvo_entry(pvo, kernel_pmap, va);
1577 	pvo->pvo_vaddr |= PVO_WIRED;
1578 
1579 	moea64_pvo_enter(installed_mmu, pvo, NULL);
1580 
1581 	if (needed_lock)
1582 		PMAP_UNLOCK(kernel_pmap);
1583 
1584 	if ((wait & M_ZERO) && (m->flags & PG_ZERO) == 0)
1585                 bzero((void *)va, PAGE_SIZE);
1586 
1587 	return (void *)va;
1588 }
1589 
1590 extern int elf32_nxstack;
1591 
1592 void
1593 moea64_init(mmu_t mmu)
1594 {
1595 
1596 	CTR0(KTR_PMAP, "moea64_init");
1597 
1598 	moea64_pvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1599 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1600 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1601 
1602 	if (!hw_direct_map) {
1603 		installed_mmu = mmu;
1604 		uma_zone_set_allocf(moea64_pvo_zone, moea64_uma_page_alloc);
1605 	}
1606 
1607 #ifdef COMPAT_FREEBSD32
1608 	elf32_nxstack = 1;
1609 #endif
1610 
1611 	moea64_initialized = TRUE;
1612 }
1613 
1614 boolean_t
1615 moea64_is_referenced(mmu_t mmu, vm_page_t m)
1616 {
1617 
1618 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1619 	    ("moea64_is_referenced: page %p is not managed", m));
1620 
1621 	return (moea64_query_bit(mmu, m, LPTE_REF));
1622 }
1623 
1624 boolean_t
1625 moea64_is_modified(mmu_t mmu, vm_page_t m)
1626 {
1627 
1628 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1629 	    ("moea64_is_modified: page %p is not managed", m));
1630 
1631 	/*
1632 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1633 	 * concurrently set while the object is locked.  Thus, if PGA_WRITEABLE
1634 	 * is clear, no PTEs can have LPTE_CHG set.
1635 	 */
1636 	VM_OBJECT_ASSERT_LOCKED(m->object);
1637 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1638 		return (FALSE);
1639 	return (moea64_query_bit(mmu, m, LPTE_CHG));
1640 }
1641 
1642 boolean_t
1643 moea64_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1644 {
1645 	struct pvo_entry *pvo;
1646 	boolean_t rv = TRUE;
1647 
1648 	PMAP_LOCK(pmap);
1649 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1650 	if (pvo != NULL)
1651 		rv = FALSE;
1652 	PMAP_UNLOCK(pmap);
1653 	return (rv);
1654 }
1655 
1656 void
1657 moea64_clear_modify(mmu_t mmu, vm_page_t m)
1658 {
1659 
1660 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1661 	    ("moea64_clear_modify: page %p is not managed", m));
1662 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1663 	KASSERT(!vm_page_xbusied(m),
1664 	    ("moea64_clear_modify: page %p is exclusive busied", m));
1665 
1666 	/*
1667 	 * If the page is not PGA_WRITEABLE, then no PTEs can have LPTE_CHG
1668 	 * set.  If the object containing the page is locked and the page is
1669 	 * not exclusive busied, then PGA_WRITEABLE cannot be concurrently set.
1670 	 */
1671 	if ((m->aflags & PGA_WRITEABLE) == 0)
1672 		return;
1673 	moea64_clear_bit(mmu, m, LPTE_CHG);
1674 }
1675 
1676 /*
1677  * Clear the write and modified bits in each of the given page's mappings.
1678  */
1679 void
1680 moea64_remove_write(mmu_t mmu, vm_page_t m)
1681 {
1682 	struct	pvo_entry *pvo;
1683 	int64_t	refchg, ret;
1684 	pmap_t	pmap;
1685 
1686 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1687 	    ("moea64_remove_write: page %p is not managed", m));
1688 
1689 	/*
1690 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1691 	 * set by another thread while the object is locked.  Thus,
1692 	 * if PGA_WRITEABLE is clear, no page table entries need updating.
1693 	 */
1694 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1695 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1696 		return;
1697 	powerpc_sync();
1698 	PV_PAGE_LOCK(m);
1699 	refchg = 0;
1700 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1701 		pmap = pvo->pvo_pmap;
1702 		PMAP_LOCK(pmap);
1703 		if (!(pvo->pvo_vaddr & PVO_DEAD) &&
1704 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1705 			pvo->pvo_pte.prot &= ~VM_PROT_WRITE;
1706 			ret = MOEA64_PTE_REPLACE(mmu, pvo,
1707 			    MOEA64_PTE_PROT_UPDATE);
1708 			if (ret < 0)
1709 				ret = LPTE_CHG;
1710 			refchg |= ret;
1711 			if (pvo->pvo_pmap == kernel_pmap)
1712 				isync();
1713 		}
1714 		PMAP_UNLOCK(pmap);
1715 	}
1716 	if ((refchg | atomic_readandclear_32(&m->md.mdpg_attrs)) & LPTE_CHG)
1717 		vm_page_dirty(m);
1718 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1719 	PV_PAGE_UNLOCK(m);
1720 }
1721 
1722 /*
1723  *	moea64_ts_referenced:
1724  *
1725  *	Return a count of reference bits for a page, clearing those bits.
1726  *	It is not necessary for every reference bit to be cleared, but it
1727  *	is necessary that 0 only be returned when there are truly no
1728  *	reference bits set.
1729  *
1730  *	XXX: The exact number of bits to check and clear is a matter that
1731  *	should be tested and standardized at some point in the future for
1732  *	optimal aging of shared pages.
1733  */
1734 int
1735 moea64_ts_referenced(mmu_t mmu, vm_page_t m)
1736 {
1737 
1738 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1739 	    ("moea64_ts_referenced: page %p is not managed", m));
1740 	return (moea64_clear_bit(mmu, m, LPTE_REF));
1741 }
1742 
1743 /*
1744  * Modify the WIMG settings of all mappings for a page.
1745  */
1746 void
1747 moea64_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1748 {
1749 	struct	pvo_entry *pvo;
1750 	int64_t	refchg;
1751 	pmap_t	pmap;
1752 	uint64_t lo;
1753 
1754 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1755 		m->md.mdpg_cache_attrs = ma;
1756 		return;
1757 	}
1758 
1759 	lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1760 
1761 	PV_PAGE_LOCK(m);
1762 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1763 		pmap = pvo->pvo_pmap;
1764 		PMAP_LOCK(pmap);
1765 		if (!(pvo->pvo_vaddr & PVO_DEAD)) {
1766 			pvo->pvo_pte.pa &= ~LPTE_WIMG;
1767 			pvo->pvo_pte.pa |= lo;
1768 			refchg = MOEA64_PTE_REPLACE(mmu, pvo,
1769 			    MOEA64_PTE_INVALIDATE);
1770 			if (refchg < 0)
1771 				refchg = (pvo->pvo_pte.prot & VM_PROT_WRITE) ?
1772 				    LPTE_CHG : 0;
1773 			if ((pvo->pvo_vaddr & PVO_MANAGED) &&
1774 			    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1775 				refchg |=
1776 				    atomic_readandclear_32(&m->md.mdpg_attrs);
1777 				if (refchg & LPTE_CHG)
1778 					vm_page_dirty(m);
1779 				if (refchg & LPTE_REF)
1780 					vm_page_aflag_set(m, PGA_REFERENCED);
1781 			}
1782 			if (pvo->pvo_pmap == kernel_pmap)
1783 				isync();
1784 		}
1785 		PMAP_UNLOCK(pmap);
1786 	}
1787 	m->md.mdpg_cache_attrs = ma;
1788 	PV_PAGE_UNLOCK(m);
1789 }
1790 
1791 /*
1792  * Map a wired page into kernel virtual address space.
1793  */
1794 void
1795 moea64_kenter_attr(mmu_t mmu, vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
1796 {
1797 	int		error;
1798 	struct pvo_entry *pvo, *oldpvo;
1799 
1800 	pvo = alloc_pvo_entry(0);
1801 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;
1802 	pvo->pvo_pte.pa = (pa & ~ADDR_POFF) | moea64_calc_wimg(pa, ma);
1803 	pvo->pvo_vaddr |= PVO_WIRED;
1804 
1805 	PMAP_LOCK(kernel_pmap);
1806 	oldpvo = moea64_pvo_find_va(kernel_pmap, va);
1807 	if (oldpvo != NULL)
1808 		moea64_pvo_remove_from_pmap(mmu, oldpvo);
1809 	init_pvo_entry(pvo, kernel_pmap, va);
1810 	error = moea64_pvo_enter(mmu, pvo, NULL);
1811 	PMAP_UNLOCK(kernel_pmap);
1812 
1813 	/* Free any dead pages */
1814 	if (oldpvo != NULL) {
1815 		PV_LOCK(oldpvo->pvo_pte.pa & LPTE_RPGN);
1816 		moea64_pvo_remove_from_page(mmu, oldpvo);
1817 		PV_UNLOCK(oldpvo->pvo_pte.pa & LPTE_RPGN);
1818 		free_pvo_entry(oldpvo);
1819 	}
1820 
1821 	if (error != 0 && error != ENOENT)
1822 		panic("moea64_kenter: failed to enter va %#zx pa %#jx: %d", va,
1823 		    (uintmax_t)pa, error);
1824 }
1825 
1826 void
1827 moea64_kenter(mmu_t mmu, vm_offset_t va, vm_paddr_t pa)
1828 {
1829 
1830 	moea64_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1831 }
1832 
1833 /*
1834  * Extract the physical page address associated with the given kernel virtual
1835  * address.
1836  */
1837 vm_paddr_t
1838 moea64_kextract(mmu_t mmu, vm_offset_t va)
1839 {
1840 	struct		pvo_entry *pvo;
1841 	vm_paddr_t pa;
1842 
1843 	/*
1844 	 * Shortcut the direct-mapped case when applicable.  We never put
1845 	 * anything but 1:1 (or 62-bit aliased) mappings below
1846 	 * VM_MIN_KERNEL_ADDRESS.
1847 	 */
1848 	if (va < VM_MIN_KERNEL_ADDRESS)
1849 		return (va & ~DMAP_BASE_ADDRESS);
1850 
1851 	PMAP_LOCK(kernel_pmap);
1852 	pvo = moea64_pvo_find_va(kernel_pmap, va);
1853 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found for %#" PRIxPTR,
1854 	    va));
1855 	pa = (pvo->pvo_pte.pa & LPTE_RPGN) | (va - PVO_VADDR(pvo));
1856 	PMAP_UNLOCK(kernel_pmap);
1857 	return (pa);
1858 }
1859 
1860 /*
1861  * Remove a wired page from kernel virtual address space.
1862  */
1863 void
1864 moea64_kremove(mmu_t mmu, vm_offset_t va)
1865 {
1866 	moea64_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1867 }
1868 
1869 /*
1870  * Provide a kernel pointer corresponding to a given userland pointer.
1871  * The returned pointer is valid until the next time this function is
1872  * called in this thread. This is used internally in copyin/copyout.
1873  */
1874 static int
1875 moea64_map_user_ptr(mmu_t mmu, pmap_t pm, volatile const void *uaddr,
1876     void **kaddr, size_t ulen, size_t *klen)
1877 {
1878 	size_t l;
1879 #ifdef __powerpc64__
1880 	struct slb *slb;
1881 #endif
1882 	register_t slbv;
1883 
1884 	*kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
1885 	l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
1886 	if (l > ulen)
1887 		l = ulen;
1888 	if (klen)
1889 		*klen = l;
1890 	else if (l != ulen)
1891 		return (EFAULT);
1892 
1893 #ifdef __powerpc64__
1894 	/* Try lockless look-up first */
1895 	slb = user_va_to_slb_entry(pm, (vm_offset_t)uaddr);
1896 
1897 	if (slb == NULL) {
1898 		/* If it isn't there, we need to pre-fault the VSID */
1899 		PMAP_LOCK(pm);
1900 		slbv = va_to_vsid(pm, (vm_offset_t)uaddr) << SLBV_VSID_SHIFT;
1901 		PMAP_UNLOCK(pm);
1902 	} else {
1903 		slbv = slb->slbv;
1904 	}
1905 
1906 	/* Mark segment no-execute */
1907 	slbv |= SLBV_N;
1908 #else
1909 	slbv = va_to_vsid(pm, (vm_offset_t)uaddr);
1910 
1911 	/* Mark segment no-execute */
1912 	slbv |= SR_N;
1913 #endif
1914 
1915 	/* If we have already set this VSID, we can just return */
1916 	if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == slbv)
1917 		return (0);
1918 
1919 	__asm __volatile("isync");
1920 	curthread->td_pcb->pcb_cpu.aim.usr_segm =
1921 	    (uintptr_t)uaddr >> ADDR_SR_SHFT;
1922 	curthread->td_pcb->pcb_cpu.aim.usr_vsid = slbv;
1923 #ifdef __powerpc64__
1924 	__asm __volatile ("slbie %0; slbmte %1, %2; isync" ::
1925 	    "r"(USER_ADDR), "r"(slbv), "r"(USER_SLB_SLBE));
1926 #else
1927 	__asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(slbv));
1928 #endif
1929 
1930 	return (0);
1931 }
1932 
1933 /*
1934  * Figure out where a given kernel pointer (usually in a fault) points
1935  * to from the VM's perspective, potentially remapping into userland's
1936  * address space.
1937  */
1938 static int
1939 moea64_decode_kernel_ptr(mmu_t mmu, vm_offset_t addr, int *is_user,
1940     vm_offset_t *decoded_addr)
1941 {
1942 	vm_offset_t user_sr;
1943 
1944 	if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
1945 		user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
1946 		addr &= ADDR_PIDX | ADDR_POFF;
1947 		addr |= user_sr << ADDR_SR_SHFT;
1948 		*decoded_addr = addr;
1949 		*is_user = 1;
1950 	} else {
1951 		*decoded_addr = addr;
1952 		*is_user = 0;
1953 	}
1954 
1955 	return (0);
1956 }
1957 
1958 /*
1959  * Map a range of physical addresses into kernel virtual address space.
1960  *
1961  * The value passed in *virt is a suggested virtual address for the mapping.
1962  * Architectures which can support a direct-mapped physical to virtual region
1963  * can return the appropriate address within that region, leaving '*virt'
1964  * unchanged.  Other architectures should map the pages starting at '*virt' and
1965  * update '*virt' with the first usable address after the mapped region.
1966  */
1967 vm_offset_t
1968 moea64_map(mmu_t mmu, vm_offset_t *virt, vm_paddr_t pa_start,
1969     vm_paddr_t pa_end, int prot)
1970 {
1971 	vm_offset_t	sva, va;
1972 
1973 	if (hw_direct_map) {
1974 		/*
1975 		 * Check if every page in the region is covered by the direct
1976 		 * map. The direct map covers all of physical memory. Use
1977 		 * moea64_calc_wimg() as a shortcut to see if the page is in
1978 		 * physical memory as a way to see if the direct map covers it.
1979 		 */
1980 		for (va = pa_start; va < pa_end; va += PAGE_SIZE)
1981 			if (moea64_calc_wimg(va, VM_MEMATTR_DEFAULT) != LPTE_M)
1982 				break;
1983 		if (va == pa_end)
1984 			return (PHYS_TO_DMAP(pa_start));
1985 	}
1986 	sva = *virt;
1987 	va = sva;
1988 	/* XXX respect prot argument */
1989 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1990 		moea64_kenter(mmu, va, pa_start);
1991 	*virt = va;
1992 
1993 	return (sva);
1994 }
1995 
1996 /*
1997  * Returns true if the pmap's pv is one of the first
1998  * 16 pvs linked to from this page.  This count may
1999  * be changed upwards or downwards in the future; it
2000  * is only necessary that true be returned for a small
2001  * subset of pmaps for proper page aging.
2002  */
2003 boolean_t
2004 moea64_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
2005 {
2006         int loops;
2007 	struct pvo_entry *pvo;
2008 	boolean_t rv;
2009 
2010 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2011 	    ("moea64_page_exists_quick: page %p is not managed", m));
2012 	loops = 0;
2013 	rv = FALSE;
2014 	PV_PAGE_LOCK(m);
2015 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2016 		if (!(pvo->pvo_vaddr & PVO_DEAD) && pvo->pvo_pmap == pmap) {
2017 			rv = TRUE;
2018 			break;
2019 		}
2020 		if (++loops >= 16)
2021 			break;
2022 	}
2023 	PV_PAGE_UNLOCK(m);
2024 	return (rv);
2025 }
2026 
2027 void
2028 moea64_page_init(mmu_t mmu __unused, vm_page_t m)
2029 {
2030 
2031 	m->md.mdpg_attrs = 0;
2032 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
2033 	LIST_INIT(&m->md.mdpg_pvoh);
2034 }
2035 
2036 /*
2037  * Return the number of managed mappings to the given physical page
2038  * that are wired.
2039  */
2040 int
2041 moea64_page_wired_mappings(mmu_t mmu, vm_page_t m)
2042 {
2043 	struct pvo_entry *pvo;
2044 	int count;
2045 
2046 	count = 0;
2047 	if ((m->oflags & VPO_UNMANAGED) != 0)
2048 		return (count);
2049 	PV_PAGE_LOCK(m);
2050 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
2051 		if ((pvo->pvo_vaddr & (PVO_DEAD | PVO_WIRED)) == PVO_WIRED)
2052 			count++;
2053 	PV_PAGE_UNLOCK(m);
2054 	return (count);
2055 }
2056 
2057 static uintptr_t	moea64_vsidcontext;
2058 
2059 uintptr_t
2060 moea64_get_unique_vsid(void) {
2061 	u_int entropy;
2062 	register_t hash;
2063 	uint32_t mask;
2064 	int i;
2065 
2066 	entropy = 0;
2067 	__asm __volatile("mftb %0" : "=r"(entropy));
2068 
2069 	mtx_lock(&moea64_slb_mutex);
2070 	for (i = 0; i < NVSIDS; i += VSID_NBPW) {
2071 		u_int	n;
2072 
2073 		/*
2074 		 * Create a new value by mutiplying by a prime and adding in
2075 		 * entropy from the timebase register.  This is to make the
2076 		 * VSID more random so that the PT hash function collides
2077 		 * less often.  (Note that the prime casues gcc to do shifts
2078 		 * instead of a multiply.)
2079 		 */
2080 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
2081 		hash = moea64_vsidcontext & (NVSIDS - 1);
2082 		if (hash == 0)		/* 0 is special, avoid it */
2083 			continue;
2084 		n = hash >> 5;
2085 		mask = 1 << (hash & (VSID_NBPW - 1));
2086 		hash = (moea64_vsidcontext & VSID_HASHMASK);
2087 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
2088 			/* anything free in this bucket? */
2089 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
2090 				entropy = (moea64_vsidcontext >> 20);
2091 				continue;
2092 			}
2093 			i = ffs(~moea64_vsid_bitmap[n]) - 1;
2094 			mask = 1 << i;
2095 			hash &= rounddown2(VSID_HASHMASK, VSID_NBPW);
2096 			hash |= i;
2097 		}
2098 		if (hash == VSID_VRMA)	/* also special, avoid this too */
2099 			continue;
2100 		KASSERT(!(moea64_vsid_bitmap[n] & mask),
2101 		    ("Allocating in-use VSID %#zx\n", hash));
2102 		moea64_vsid_bitmap[n] |= mask;
2103 		mtx_unlock(&moea64_slb_mutex);
2104 		return (hash);
2105 	}
2106 
2107 	mtx_unlock(&moea64_slb_mutex);
2108 	panic("%s: out of segments",__func__);
2109 }
2110 
2111 #ifdef __powerpc64__
2112 void
2113 moea64_pinit(mmu_t mmu, pmap_t pmap)
2114 {
2115 
2116 	RB_INIT(&pmap->pmap_pvo);
2117 
2118 	pmap->pm_slb_tree_root = slb_alloc_tree();
2119 	pmap->pm_slb = slb_alloc_user_cache();
2120 	pmap->pm_slb_len = 0;
2121 }
2122 #else
2123 void
2124 moea64_pinit(mmu_t mmu, pmap_t pmap)
2125 {
2126 	int	i;
2127 	uint32_t hash;
2128 
2129 	RB_INIT(&pmap->pmap_pvo);
2130 
2131 	if (pmap_bootstrapped)
2132 		pmap->pmap_phys = (pmap_t)moea64_kextract(mmu,
2133 		    (vm_offset_t)pmap);
2134 	else
2135 		pmap->pmap_phys = pmap;
2136 
2137 	/*
2138 	 * Allocate some segment registers for this pmap.
2139 	 */
2140 	hash = moea64_get_unique_vsid();
2141 
2142 	for (i = 0; i < 16; i++)
2143 		pmap->pm_sr[i] = VSID_MAKE(i, hash);
2144 
2145 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_pinit: pm_sr[0] = 0"));
2146 }
2147 #endif
2148 
2149 /*
2150  * Initialize the pmap associated with process 0.
2151  */
2152 void
2153 moea64_pinit0(mmu_t mmu, pmap_t pm)
2154 {
2155 
2156 	PMAP_LOCK_INIT(pm);
2157 	moea64_pinit(mmu, pm);
2158 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
2159 }
2160 
2161 /*
2162  * Set the physical protection on the specified range of this map as requested.
2163  */
2164 static void
2165 moea64_pvo_protect(mmu_t mmu,  pmap_t pm, struct pvo_entry *pvo, vm_prot_t prot)
2166 {
2167 	struct vm_page *pg;
2168 	vm_prot_t oldprot;
2169 	int32_t refchg;
2170 
2171 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2172 
2173 	/*
2174 	 * Change the protection of the page.
2175 	 */
2176 	oldprot = pvo->pvo_pte.prot;
2177 	pvo->pvo_pte.prot = prot;
2178 	pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
2179 
2180 	/*
2181 	 * If the PVO is in the page table, update mapping
2182 	 */
2183 	refchg = MOEA64_PTE_REPLACE(mmu, pvo, MOEA64_PTE_PROT_UPDATE);
2184 	if (refchg < 0)
2185 		refchg = (oldprot & VM_PROT_WRITE) ? LPTE_CHG : 0;
2186 
2187 	if (pm != kernel_pmap && pg != NULL && !(pg->aflags & PGA_EXECUTABLE) &&
2188 	    (pvo->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
2189 		if ((pg->oflags & VPO_UNMANAGED) == 0)
2190 			vm_page_aflag_set(pg, PGA_EXECUTABLE);
2191 		moea64_syncicache(mmu, pm, PVO_VADDR(pvo),
2192 		    pvo->pvo_pte.pa & LPTE_RPGN, PAGE_SIZE);
2193 	}
2194 
2195 	/*
2196 	 * Update vm about the REF/CHG bits if the page is managed and we have
2197 	 * removed write access.
2198 	 */
2199 	if (pg != NULL && (pvo->pvo_vaddr & PVO_MANAGED) &&
2200 	    (oldprot & VM_PROT_WRITE)) {
2201 		refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2202 		if (refchg & LPTE_CHG)
2203 			vm_page_dirty(pg);
2204 		if (refchg & LPTE_REF)
2205 			vm_page_aflag_set(pg, PGA_REFERENCED);
2206 	}
2207 }
2208 
2209 void
2210 moea64_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
2211     vm_prot_t prot)
2212 {
2213 	struct	pvo_entry *pvo, *tpvo, key;
2214 
2215 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm,
2216 	    sva, eva, prot);
2217 
2218 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
2219 	    ("moea64_protect: non current pmap"));
2220 
2221 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2222 		moea64_remove(mmu, pm, sva, eva);
2223 		return;
2224 	}
2225 
2226 	PMAP_LOCK(pm);
2227 	key.pvo_vaddr = sva;
2228 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2229 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2230 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2231 		moea64_pvo_protect(mmu, pm, pvo, prot);
2232 	}
2233 	PMAP_UNLOCK(pm);
2234 }
2235 
2236 /*
2237  * Map a list of wired pages into kernel virtual address space.  This is
2238  * intended for temporary mappings which do not need page modification or
2239  * references recorded.  Existing mappings in the region are overwritten.
2240  */
2241 void
2242 moea64_qenter(mmu_t mmu, vm_offset_t va, vm_page_t *m, int count)
2243 {
2244 	while (count-- > 0) {
2245 		moea64_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
2246 		va += PAGE_SIZE;
2247 		m++;
2248 	}
2249 }
2250 
2251 /*
2252  * Remove page mappings from kernel virtual address space.  Intended for
2253  * temporary mappings entered by moea64_qenter.
2254  */
2255 void
2256 moea64_qremove(mmu_t mmu, vm_offset_t va, int count)
2257 {
2258 	while (count-- > 0) {
2259 		moea64_kremove(mmu, va);
2260 		va += PAGE_SIZE;
2261 	}
2262 }
2263 
2264 void
2265 moea64_release_vsid(uint64_t vsid)
2266 {
2267 	int idx, mask;
2268 
2269 	mtx_lock(&moea64_slb_mutex);
2270 	idx = vsid & (NVSIDS-1);
2271 	mask = 1 << (idx % VSID_NBPW);
2272 	idx /= VSID_NBPW;
2273 	KASSERT(moea64_vsid_bitmap[idx] & mask,
2274 	    ("Freeing unallocated VSID %#jx", vsid));
2275 	moea64_vsid_bitmap[idx] &= ~mask;
2276 	mtx_unlock(&moea64_slb_mutex);
2277 }
2278 
2279 
2280 void
2281 moea64_release(mmu_t mmu, pmap_t pmap)
2282 {
2283 
2284 	/*
2285 	 * Free segment registers' VSIDs
2286 	 */
2287     #ifdef __powerpc64__
2288 	slb_free_tree(pmap);
2289 	slb_free_user_cache(pmap->pm_slb);
2290     #else
2291 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_release: pm_sr[0] = 0"));
2292 
2293 	moea64_release_vsid(VSID_TO_HASH(pmap->pm_sr[0]));
2294     #endif
2295 }
2296 
2297 /*
2298  * Remove all pages mapped by the specified pmap
2299  */
2300 void
2301 moea64_remove_pages(mmu_t mmu, pmap_t pm)
2302 {
2303 	struct pvo_entry *pvo, *tpvo;
2304 	struct pvo_tree tofree;
2305 
2306 	RB_INIT(&tofree);
2307 
2308 	PMAP_LOCK(pm);
2309 	RB_FOREACH_SAFE(pvo, pvo_tree, &pm->pmap_pvo, tpvo) {
2310 		if (pvo->pvo_vaddr & PVO_WIRED)
2311 			continue;
2312 
2313 		/*
2314 		 * For locking reasons, remove this from the page table and
2315 		 * pmap, but save delinking from the vm_page for a second
2316 		 * pass
2317 		 */
2318 		moea64_pvo_remove_from_pmap(mmu, pvo);
2319 		RB_INSERT(pvo_tree, &tofree, pvo);
2320 	}
2321 	PMAP_UNLOCK(pm);
2322 
2323 	RB_FOREACH_SAFE(pvo, pvo_tree, &tofree, tpvo) {
2324 		PV_LOCK(pvo->pvo_pte.pa & LPTE_RPGN);
2325 		moea64_pvo_remove_from_page(mmu, pvo);
2326 		PV_UNLOCK(pvo->pvo_pte.pa & LPTE_RPGN);
2327 		RB_REMOVE(pvo_tree, &tofree, pvo);
2328 		free_pvo_entry(pvo);
2329 	}
2330 }
2331 
2332 /*
2333  * Remove the given range of addresses from the specified map.
2334  */
2335 void
2336 moea64_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
2337 {
2338 	struct  pvo_entry *pvo, *tpvo, key;
2339 	struct pvo_tree tofree;
2340 
2341 	/*
2342 	 * Perform an unsynchronized read.  This is, however, safe.
2343 	 */
2344 	if (pm->pm_stats.resident_count == 0)
2345 		return;
2346 
2347 	key.pvo_vaddr = sva;
2348 
2349 	RB_INIT(&tofree);
2350 
2351 	PMAP_LOCK(pm);
2352 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2353 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2354 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2355 
2356 		/*
2357 		 * For locking reasons, remove this from the page table and
2358 		 * pmap, but save delinking from the vm_page for a second
2359 		 * pass
2360 		 */
2361 		moea64_pvo_remove_from_pmap(mmu, pvo);
2362 		RB_INSERT(pvo_tree, &tofree, pvo);
2363 	}
2364 	PMAP_UNLOCK(pm);
2365 
2366 	RB_FOREACH_SAFE(pvo, pvo_tree, &tofree, tpvo) {
2367 		PV_LOCK(pvo->pvo_pte.pa & LPTE_RPGN);
2368 		moea64_pvo_remove_from_page(mmu, pvo);
2369 		PV_UNLOCK(pvo->pvo_pte.pa & LPTE_RPGN);
2370 		RB_REMOVE(pvo_tree, &tofree, pvo);
2371 		free_pvo_entry(pvo);
2372 	}
2373 }
2374 
2375 /*
2376  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
2377  * will reflect changes in pte's back to the vm_page.
2378  */
2379 void
2380 moea64_remove_all(mmu_t mmu, vm_page_t m)
2381 {
2382 	struct	pvo_entry *pvo, *next_pvo;
2383 	struct	pvo_head freequeue;
2384 	int	wasdead;
2385 	pmap_t	pmap;
2386 
2387 	LIST_INIT(&freequeue);
2388 
2389 	PV_PAGE_LOCK(m);
2390 	LIST_FOREACH_SAFE(pvo, vm_page_to_pvoh(m), pvo_vlink, next_pvo) {
2391 		pmap = pvo->pvo_pmap;
2392 		PMAP_LOCK(pmap);
2393 		wasdead = (pvo->pvo_vaddr & PVO_DEAD);
2394 		if (!wasdead)
2395 			moea64_pvo_remove_from_pmap(mmu, pvo);
2396 		moea64_pvo_remove_from_page(mmu, pvo);
2397 		if (!wasdead)
2398 			LIST_INSERT_HEAD(&freequeue, pvo, pvo_vlink);
2399 		PMAP_UNLOCK(pmap);
2400 
2401 	}
2402 	KASSERT(!pmap_page_is_mapped(m), ("Page still has mappings"));
2403 	KASSERT(!(m->aflags & PGA_WRITEABLE), ("Page still writable"));
2404 	PV_PAGE_UNLOCK(m);
2405 
2406 	/* Clean up UMA allocations */
2407 	LIST_FOREACH_SAFE(pvo, &freequeue, pvo_vlink, next_pvo)
2408 		free_pvo_entry(pvo);
2409 }
2410 
2411 /*
2412  * Allocate a physical page of memory directly from the phys_avail map.
2413  * Can only be called from moea64_bootstrap before avail start and end are
2414  * calculated.
2415  */
2416 vm_offset_t
2417 moea64_bootstrap_alloc(vm_size_t size, u_int align)
2418 {
2419 	vm_offset_t	s, e;
2420 	int		i, j;
2421 
2422 	size = round_page(size);
2423 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2424 		if (align != 0)
2425 			s = roundup2(phys_avail[i], align);
2426 		else
2427 			s = phys_avail[i];
2428 		e = s + size;
2429 
2430 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2431 			continue;
2432 
2433 		if (s + size > platform_real_maxaddr())
2434 			continue;
2435 
2436 		if (s == phys_avail[i]) {
2437 			phys_avail[i] += size;
2438 		} else if (e == phys_avail[i + 1]) {
2439 			phys_avail[i + 1] -= size;
2440 		} else {
2441 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2442 				phys_avail[j] = phys_avail[j - 2];
2443 				phys_avail[j + 1] = phys_avail[j - 1];
2444 			}
2445 
2446 			phys_avail[i + 3] = phys_avail[i + 1];
2447 			phys_avail[i + 1] = s;
2448 			phys_avail[i + 2] = e;
2449 			phys_avail_count++;
2450 		}
2451 
2452 		return (s);
2453 	}
2454 	panic("moea64_bootstrap_alloc: could not allocate memory");
2455 }
2456 
2457 static int
2458 moea64_pvo_enter(mmu_t mmu, struct pvo_entry *pvo, struct pvo_head *pvo_head)
2459 {
2460 	int first, err;
2461 
2462 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2463 	KASSERT(moea64_pvo_find_va(pvo->pvo_pmap, PVO_VADDR(pvo)) == NULL,
2464 	    ("Existing mapping for VA %#jx", (uintmax_t)PVO_VADDR(pvo)));
2465 
2466 	moea64_pvo_enter_calls++;
2467 
2468 	/*
2469 	 * Add to pmap list
2470 	 */
2471 	RB_INSERT(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2472 
2473 	/*
2474 	 * Remember if the list was empty and therefore will be the first
2475 	 * item.
2476 	 */
2477 	if (pvo_head != NULL) {
2478 		if (LIST_FIRST(pvo_head) == NULL)
2479 			first = 1;
2480 		LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2481 	}
2482 
2483 	if (pvo->pvo_vaddr & PVO_WIRED)
2484 		pvo->pvo_pmap->pm_stats.wired_count++;
2485 	pvo->pvo_pmap->pm_stats.resident_count++;
2486 
2487 	/*
2488 	 * Insert it into the hardware page table
2489 	 */
2490 	err = MOEA64_PTE_INSERT(mmu, pvo);
2491 	if (err != 0) {
2492 		panic("moea64_pvo_enter: overflow");
2493 	}
2494 
2495 	moea64_pvo_entries++;
2496 
2497 	if (pvo->pvo_pmap == kernel_pmap)
2498 		isync();
2499 
2500 #ifdef __powerpc64__
2501 	/*
2502 	 * Make sure all our bootstrap mappings are in the SLB as soon
2503 	 * as virtual memory is switched on.
2504 	 */
2505 	if (!pmap_bootstrapped)
2506 		moea64_bootstrap_slb_prefault(PVO_VADDR(pvo),
2507 		    pvo->pvo_vaddr & PVO_LARGE);
2508 #endif
2509 
2510 	return (first ? ENOENT : 0);
2511 }
2512 
2513 static void
2514 moea64_pvo_remove_from_pmap(mmu_t mmu, struct pvo_entry *pvo)
2515 {
2516 	struct	vm_page *pg;
2517 	int32_t refchg;
2518 
2519 	KASSERT(pvo->pvo_pmap != NULL, ("Trying to remove PVO with no pmap"));
2520 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2521 	KASSERT(!(pvo->pvo_vaddr & PVO_DEAD), ("Trying to remove dead PVO"));
2522 
2523 	/*
2524 	 * If there is an active pte entry, we need to deactivate it
2525 	 */
2526 	refchg = MOEA64_PTE_UNSET(mmu, pvo);
2527 	if (refchg < 0) {
2528 		/*
2529 		 * If it was evicted from the page table, be pessimistic and
2530 		 * dirty the page.
2531 		 */
2532 		if (pvo->pvo_pte.prot & VM_PROT_WRITE)
2533 			refchg = LPTE_CHG;
2534 		else
2535 			refchg = 0;
2536 	}
2537 
2538 	/*
2539 	 * Update our statistics.
2540 	 */
2541 	pvo->pvo_pmap->pm_stats.resident_count--;
2542 	if (pvo->pvo_vaddr & PVO_WIRED)
2543 		pvo->pvo_pmap->pm_stats.wired_count--;
2544 
2545 	/*
2546 	 * Remove this PVO from the pmap list.
2547 	 */
2548 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2549 
2550 	/*
2551 	 * Mark this for the next sweep
2552 	 */
2553 	pvo->pvo_vaddr |= PVO_DEAD;
2554 
2555 	/* Send RC bits to VM */
2556 	if ((pvo->pvo_vaddr & PVO_MANAGED) &&
2557 	    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2558 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
2559 		if (pg != NULL) {
2560 			refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2561 			if (refchg & LPTE_CHG)
2562 				vm_page_dirty(pg);
2563 			if (refchg & LPTE_REF)
2564 				vm_page_aflag_set(pg, PGA_REFERENCED);
2565 		}
2566 	}
2567 }
2568 
2569 static void
2570 moea64_pvo_remove_from_page(mmu_t mmu, struct pvo_entry *pvo)
2571 {
2572 	struct	vm_page *pg;
2573 
2574 	KASSERT(pvo->pvo_vaddr & PVO_DEAD, ("Trying to delink live page"));
2575 
2576 	/* Use NULL pmaps as a sentinel for races in page deletion */
2577 	if (pvo->pvo_pmap == NULL)
2578 		return;
2579 	pvo->pvo_pmap = NULL;
2580 
2581 	/*
2582 	 * Update vm about page writeability/executability if managed
2583 	 */
2584 	PV_LOCKASSERT(pvo->pvo_pte.pa & LPTE_RPGN);
2585 	if (pvo->pvo_vaddr & PVO_MANAGED) {
2586 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
2587 
2588 		if (pg != NULL) {
2589 			LIST_REMOVE(pvo, pvo_vlink);
2590 			if (LIST_EMPTY(vm_page_to_pvoh(pg)))
2591 				vm_page_aflag_clear(pg,
2592 				    PGA_WRITEABLE | PGA_EXECUTABLE);
2593 		}
2594 	}
2595 
2596 	moea64_pvo_entries--;
2597 	moea64_pvo_remove_calls++;
2598 }
2599 
2600 static struct pvo_entry *
2601 moea64_pvo_find_va(pmap_t pm, vm_offset_t va)
2602 {
2603 	struct pvo_entry key;
2604 
2605 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2606 
2607 	key.pvo_vaddr = va & ~ADDR_POFF;
2608 	return (RB_FIND(pvo_tree, &pm->pmap_pvo, &key));
2609 }
2610 
2611 static boolean_t
2612 moea64_query_bit(mmu_t mmu, vm_page_t m, uint64_t ptebit)
2613 {
2614 	struct	pvo_entry *pvo;
2615 	int64_t ret;
2616 	boolean_t rv;
2617 
2618 	/*
2619 	 * See if this bit is stored in the page already.
2620 	 */
2621 	if (m->md.mdpg_attrs & ptebit)
2622 		return (TRUE);
2623 
2624 	/*
2625 	 * Examine each PTE.  Sync so that any pending REF/CHG bits are
2626 	 * flushed to the PTEs.
2627 	 */
2628 	rv = FALSE;
2629 	powerpc_sync();
2630 	PV_PAGE_LOCK(m);
2631 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2632 		ret = 0;
2633 
2634 		/*
2635 		 * See if this pvo has a valid PTE.  if so, fetch the
2636 		 * REF/CHG bits from the valid PTE.  If the appropriate
2637 		 * ptebit is set, return success.
2638 		 */
2639 		PMAP_LOCK(pvo->pvo_pmap);
2640 		if (!(pvo->pvo_vaddr & PVO_DEAD))
2641 			ret = MOEA64_PTE_SYNCH(mmu, pvo);
2642 		PMAP_UNLOCK(pvo->pvo_pmap);
2643 
2644 		if (ret > 0) {
2645 			atomic_set_32(&m->md.mdpg_attrs,
2646 			    ret & (LPTE_CHG | LPTE_REF));
2647 			if (ret & ptebit) {
2648 				rv = TRUE;
2649 				break;
2650 			}
2651 		}
2652 	}
2653 	PV_PAGE_UNLOCK(m);
2654 
2655 	return (rv);
2656 }
2657 
2658 static u_int
2659 moea64_clear_bit(mmu_t mmu, vm_page_t m, u_int64_t ptebit)
2660 {
2661 	u_int	count;
2662 	struct	pvo_entry *pvo;
2663 	int64_t ret;
2664 
2665 	/*
2666 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2667 	 * we can reset the right ones).
2668 	 */
2669 	powerpc_sync();
2670 
2671 	/*
2672 	 * For each pvo entry, clear the pte's ptebit.
2673 	 */
2674 	count = 0;
2675 	PV_PAGE_LOCK(m);
2676 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2677 		ret = 0;
2678 
2679 		PMAP_LOCK(pvo->pvo_pmap);
2680 		if (!(pvo->pvo_vaddr & PVO_DEAD))
2681 			ret = MOEA64_PTE_CLEAR(mmu, pvo, ptebit);
2682 		PMAP_UNLOCK(pvo->pvo_pmap);
2683 
2684 		if (ret > 0 && (ret & ptebit))
2685 			count++;
2686 	}
2687 	atomic_clear_32(&m->md.mdpg_attrs, ptebit);
2688 	PV_PAGE_UNLOCK(m);
2689 
2690 	return (count);
2691 }
2692 
2693 boolean_t
2694 moea64_dev_direct_mapped(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2695 {
2696 	struct pvo_entry *pvo, key;
2697 	vm_offset_t ppa;
2698 	int error = 0;
2699 
2700 	if (hw_direct_map && mem_valid(pa, size) == 0)
2701 		return (0);
2702 
2703 	PMAP_LOCK(kernel_pmap);
2704 	ppa = pa & ~ADDR_POFF;
2705 	key.pvo_vaddr = DMAP_BASE_ADDRESS + ppa;
2706 	for (pvo = RB_FIND(pvo_tree, &kernel_pmap->pmap_pvo, &key);
2707 	    ppa < pa + size; ppa += PAGE_SIZE,
2708 	    pvo = RB_NEXT(pvo_tree, &kernel_pmap->pmap_pvo, pvo)) {
2709 		if (pvo == NULL || (pvo->pvo_pte.pa & LPTE_RPGN) != ppa) {
2710 			error = EFAULT;
2711 			break;
2712 		}
2713 	}
2714 	PMAP_UNLOCK(kernel_pmap);
2715 
2716 	return (error);
2717 }
2718 
2719 /*
2720  * Map a set of physical memory pages into the kernel virtual
2721  * address space. Return a pointer to where it is mapped. This
2722  * routine is intended to be used for mapping device memory,
2723  * NOT real memory.
2724  */
2725 void *
2726 moea64_mapdev_attr(mmu_t mmu, vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
2727 {
2728 	vm_offset_t va, tmpva, ppa, offset;
2729 
2730 	ppa = trunc_page(pa);
2731 	offset = pa & PAGE_MASK;
2732 	size = roundup2(offset + size, PAGE_SIZE);
2733 
2734 	va = kva_alloc(size);
2735 
2736 	if (!va)
2737 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
2738 
2739 	for (tmpva = va; size > 0;) {
2740 		moea64_kenter_attr(mmu, tmpva, ppa, ma);
2741 		size -= PAGE_SIZE;
2742 		tmpva += PAGE_SIZE;
2743 		ppa += PAGE_SIZE;
2744 	}
2745 
2746 	return ((void *)(va + offset));
2747 }
2748 
2749 void *
2750 moea64_mapdev(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2751 {
2752 
2753 	return moea64_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT);
2754 }
2755 
2756 void
2757 moea64_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2758 {
2759 	vm_offset_t base, offset;
2760 
2761 	base = trunc_page(va);
2762 	offset = va & PAGE_MASK;
2763 	size = roundup2(offset + size, PAGE_SIZE);
2764 
2765 	kva_free(base, size);
2766 }
2767 
2768 void
2769 moea64_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2770 {
2771 	struct pvo_entry *pvo;
2772 	vm_offset_t lim;
2773 	vm_paddr_t pa;
2774 	vm_size_t len;
2775 
2776 	PMAP_LOCK(pm);
2777 	while (sz > 0) {
2778 		lim = round_page(va);
2779 		len = MIN(lim - va, sz);
2780 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
2781 		if (pvo != NULL && !(pvo->pvo_pte.pa & LPTE_I)) {
2782 			pa = (pvo->pvo_pte.pa & LPTE_RPGN) | (va & ADDR_POFF);
2783 			moea64_syncicache(mmu, pm, va, pa, len);
2784 		}
2785 		va += len;
2786 		sz -= len;
2787 	}
2788 	PMAP_UNLOCK(pm);
2789 }
2790 
2791 void
2792 moea64_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va)
2793 {
2794 
2795 	*va = (void *)(uintptr_t)pa;
2796 }
2797 
2798 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
2799 
2800 void
2801 moea64_scan_init(mmu_t mmu)
2802 {
2803 	struct pvo_entry *pvo;
2804 	vm_offset_t va;
2805 	int i;
2806 
2807 	if (!do_minidump) {
2808 		/* Initialize phys. segments for dumpsys(). */
2809 		memset(&dump_map, 0, sizeof(dump_map));
2810 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
2811 		for (i = 0; i < pregions_sz; i++) {
2812 			dump_map[i].pa_start = pregions[i].mr_start;
2813 			dump_map[i].pa_size = pregions[i].mr_size;
2814 		}
2815 		return;
2816 	}
2817 
2818 	/* Virtual segments for minidumps: */
2819 	memset(&dump_map, 0, sizeof(dump_map));
2820 
2821 	/* 1st: kernel .data and .bss. */
2822 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
2823 	dump_map[0].pa_size = round_page((uintptr_t)_end) -
2824 	    dump_map[0].pa_start;
2825 
2826 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2827 	dump_map[1].pa_start = (vm_paddr_t)(uintptr_t)msgbufp->msg_ptr;
2828 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
2829 
2830 	/* 3rd: kernel VM. */
2831 	va = dump_map[1].pa_start + dump_map[1].pa_size;
2832 	/* Find start of next chunk (from va). */
2833 	while (va < virtual_end) {
2834 		/* Don't dump the buffer cache. */
2835 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
2836 			va = kmi.buffer_eva;
2837 			continue;
2838 		}
2839 		pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
2840 		if (pvo != NULL && !(pvo->pvo_vaddr & PVO_DEAD))
2841 			break;
2842 		va += PAGE_SIZE;
2843 	}
2844 	if (va < virtual_end) {
2845 		dump_map[2].pa_start = va;
2846 		va += PAGE_SIZE;
2847 		/* Find last page in chunk. */
2848 		while (va < virtual_end) {
2849 			/* Don't run into the buffer cache. */
2850 			if (va == kmi.buffer_sva)
2851 				break;
2852 			pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
2853 			if (pvo != NULL && !(pvo->pvo_vaddr & PVO_DEAD))
2854 				break;
2855 			va += PAGE_SIZE;
2856 		}
2857 		dump_map[2].pa_size = va - dump_map[2].pa_start;
2858 	}
2859 }
2860 
2861