xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision d8a0fe102c0cfdfcd5b818f850eff09d8536c9bc)
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 /*
144  * Map of physical memory regions.
145  */
146 static struct	mem_region *regions;
147 static struct	mem_region *pregions;
148 static u_int	phys_avail_count;
149 static int	regions_sz, pregions_sz;
150 
151 extern void bs_remap_earlyboot(void);
152 
153 /*
154  * Lock for the SLB tables.
155  */
156 struct mtx	moea64_slb_mutex;
157 
158 /*
159  * PTEG data.
160  */
161 u_int		moea64_pteg_count;
162 u_int		moea64_pteg_mask;
163 
164 /*
165  * PVO data.
166  */
167 
168 uma_zone_t	moea64_pvo_zone; /* zone for pvo entries */
169 
170 static struct	pvo_entry *moea64_bpvo_pool;
171 static int	moea64_bpvo_pool_index = 0;
172 static int	moea64_bpvo_pool_size = 327680;
173 TUNABLE_INT("machdep.moea64_bpvo_pool_size", &moea64_bpvo_pool_size);
174 SYSCTL_INT(_machdep, OID_AUTO, moea64_allocated_bpvo_entries, CTLFLAG_RD,
175     &moea64_bpvo_pool_index, 0, "");
176 
177 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
178 #ifdef __powerpc64__
179 #define	NVSIDS		(NPMAPS * 16)
180 #define VSID_HASHMASK	0xffffffffUL
181 #else
182 #define NVSIDS		NPMAPS
183 #define VSID_HASHMASK	0xfffffUL
184 #endif
185 static u_int	moea64_vsid_bitmap[NVSIDS / VSID_NBPW];
186 
187 static boolean_t moea64_initialized = FALSE;
188 
189 /*
190  * Statistics.
191  */
192 u_int	moea64_pte_valid = 0;
193 u_int	moea64_pte_overflow = 0;
194 u_int	moea64_pvo_entries = 0;
195 u_int	moea64_pvo_enter_calls = 0;
196 u_int	moea64_pvo_remove_calls = 0;
197 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_valid, CTLFLAG_RD,
198     &moea64_pte_valid, 0, "");
199 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_overflow, CTLFLAG_RD,
200     &moea64_pte_overflow, 0, "");
201 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_entries, CTLFLAG_RD,
202     &moea64_pvo_entries, 0, "");
203 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_enter_calls, CTLFLAG_RD,
204     &moea64_pvo_enter_calls, 0, "");
205 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_remove_calls, CTLFLAG_RD,
206     &moea64_pvo_remove_calls, 0, "");
207 
208 vm_offset_t	moea64_scratchpage_va[2];
209 struct pvo_entry *moea64_scratchpage_pvo[2];
210 struct	mtx	moea64_scratchpage_mtx;
211 
212 uint64_t 	moea64_large_page_mask = 0;
213 uint64_t	moea64_large_page_size = 0;
214 int		moea64_large_page_shift = 0;
215 
216 /*
217  * PVO calls.
218  */
219 static int	moea64_pvo_enter(mmu_t mmu, struct pvo_entry *pvo,
220 		    struct pvo_head *pvo_head);
221 static void	moea64_pvo_remove_from_pmap(mmu_t mmu, struct pvo_entry *pvo);
222 static void	moea64_pvo_remove_from_page(mmu_t mmu, struct pvo_entry *pvo);
223 static struct	pvo_entry *moea64_pvo_find_va(pmap_t, vm_offset_t);
224 
225 /*
226  * Utility routines.
227  */
228 static boolean_t	moea64_query_bit(mmu_t, vm_page_t, uint64_t);
229 static u_int		moea64_clear_bit(mmu_t, vm_page_t, uint64_t);
230 static void		moea64_kremove(mmu_t, vm_offset_t);
231 static void		moea64_syncicache(mmu_t, pmap_t pmap, vm_offset_t va,
232 			    vm_paddr_t pa, vm_size_t sz);
233 static void		moea64_pmap_init_qpages(void);
234 
235 /*
236  * Kernel MMU interface
237  */
238 void moea64_clear_modify(mmu_t, vm_page_t);
239 void moea64_copy_page(mmu_t, vm_page_t, vm_page_t);
240 void moea64_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
241     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
242 int moea64_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t,
243     u_int flags, int8_t psind);
244 void moea64_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
245     vm_prot_t);
246 void moea64_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
247 vm_paddr_t moea64_extract(mmu_t, pmap_t, vm_offset_t);
248 vm_page_t moea64_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
249 void moea64_init(mmu_t);
250 boolean_t moea64_is_modified(mmu_t, vm_page_t);
251 boolean_t moea64_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
252 boolean_t moea64_is_referenced(mmu_t, vm_page_t);
253 int moea64_ts_referenced(mmu_t, vm_page_t);
254 vm_offset_t moea64_map(mmu_t, vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
255 boolean_t moea64_page_exists_quick(mmu_t, pmap_t, vm_page_t);
256 void moea64_page_init(mmu_t, vm_page_t);
257 int moea64_page_wired_mappings(mmu_t, vm_page_t);
258 void moea64_pinit(mmu_t, pmap_t);
259 void moea64_pinit0(mmu_t, pmap_t);
260 void moea64_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
261 void moea64_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
262 void moea64_qremove(mmu_t, vm_offset_t, int);
263 void moea64_release(mmu_t, pmap_t);
264 void moea64_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
265 void moea64_remove_pages(mmu_t, pmap_t);
266 void moea64_remove_all(mmu_t, vm_page_t);
267 void moea64_remove_write(mmu_t, vm_page_t);
268 void moea64_unwire(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
269 void moea64_zero_page(mmu_t, vm_page_t);
270 void moea64_zero_page_area(mmu_t, vm_page_t, int, int);
271 void moea64_activate(mmu_t, struct thread *);
272 void moea64_deactivate(mmu_t, struct thread *);
273 void *moea64_mapdev(mmu_t, vm_paddr_t, vm_size_t);
274 void *moea64_mapdev_attr(mmu_t, vm_paddr_t, vm_size_t, vm_memattr_t);
275 void moea64_unmapdev(mmu_t, vm_offset_t, vm_size_t);
276 vm_paddr_t moea64_kextract(mmu_t, vm_offset_t);
277 void moea64_page_set_memattr(mmu_t, vm_page_t m, vm_memattr_t ma);
278 void moea64_kenter_attr(mmu_t, vm_offset_t, vm_paddr_t, vm_memattr_t ma);
279 void moea64_kenter(mmu_t, vm_offset_t, vm_paddr_t);
280 boolean_t moea64_dev_direct_mapped(mmu_t, vm_paddr_t, vm_size_t);
281 static void moea64_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
282 void moea64_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz,
283     void **va);
284 void moea64_scan_init(mmu_t mmu);
285 vm_offset_t moea64_quick_enter_page(mmu_t mmu, vm_page_t m);
286 void moea64_quick_remove_page(mmu_t mmu, vm_offset_t addr);
287 
288 static mmu_method_t moea64_methods[] = {
289 	MMUMETHOD(mmu_clear_modify,	moea64_clear_modify),
290 	MMUMETHOD(mmu_copy_page,	moea64_copy_page),
291 	MMUMETHOD(mmu_copy_pages,	moea64_copy_pages),
292 	MMUMETHOD(mmu_enter,		moea64_enter),
293 	MMUMETHOD(mmu_enter_object,	moea64_enter_object),
294 	MMUMETHOD(mmu_enter_quick,	moea64_enter_quick),
295 	MMUMETHOD(mmu_extract,		moea64_extract),
296 	MMUMETHOD(mmu_extract_and_hold,	moea64_extract_and_hold),
297 	MMUMETHOD(mmu_init,		moea64_init),
298 	MMUMETHOD(mmu_is_modified,	moea64_is_modified),
299 	MMUMETHOD(mmu_is_prefaultable,	moea64_is_prefaultable),
300 	MMUMETHOD(mmu_is_referenced,	moea64_is_referenced),
301 	MMUMETHOD(mmu_ts_referenced,	moea64_ts_referenced),
302 	MMUMETHOD(mmu_map,     		moea64_map),
303 	MMUMETHOD(mmu_page_exists_quick,moea64_page_exists_quick),
304 	MMUMETHOD(mmu_page_init,	moea64_page_init),
305 	MMUMETHOD(mmu_page_wired_mappings,moea64_page_wired_mappings),
306 	MMUMETHOD(mmu_pinit,		moea64_pinit),
307 	MMUMETHOD(mmu_pinit0,		moea64_pinit0),
308 	MMUMETHOD(mmu_protect,		moea64_protect),
309 	MMUMETHOD(mmu_qenter,		moea64_qenter),
310 	MMUMETHOD(mmu_qremove,		moea64_qremove),
311 	MMUMETHOD(mmu_release,		moea64_release),
312 	MMUMETHOD(mmu_remove,		moea64_remove),
313 	MMUMETHOD(mmu_remove_pages,	moea64_remove_pages),
314 	MMUMETHOD(mmu_remove_all,      	moea64_remove_all),
315 	MMUMETHOD(mmu_remove_write,	moea64_remove_write),
316 	MMUMETHOD(mmu_sync_icache,	moea64_sync_icache),
317 	MMUMETHOD(mmu_unwire,		moea64_unwire),
318 	MMUMETHOD(mmu_zero_page,       	moea64_zero_page),
319 	MMUMETHOD(mmu_zero_page_area,	moea64_zero_page_area),
320 	MMUMETHOD(mmu_activate,		moea64_activate),
321 	MMUMETHOD(mmu_deactivate,      	moea64_deactivate),
322 	MMUMETHOD(mmu_page_set_memattr,	moea64_page_set_memattr),
323 	MMUMETHOD(mmu_quick_enter_page, moea64_quick_enter_page),
324 	MMUMETHOD(mmu_quick_remove_page, moea64_quick_remove_page),
325 
326 	/* Internal interfaces */
327 	MMUMETHOD(mmu_mapdev,		moea64_mapdev),
328 	MMUMETHOD(mmu_mapdev_attr,	moea64_mapdev_attr),
329 	MMUMETHOD(mmu_unmapdev,		moea64_unmapdev),
330 	MMUMETHOD(mmu_kextract,		moea64_kextract),
331 	MMUMETHOD(mmu_kenter,		moea64_kenter),
332 	MMUMETHOD(mmu_kenter_attr,	moea64_kenter_attr),
333 	MMUMETHOD(mmu_dev_direct_mapped,moea64_dev_direct_mapped),
334 	MMUMETHOD(mmu_scan_init,	moea64_scan_init),
335 	MMUMETHOD(mmu_dumpsys_map,	moea64_dumpsys_map),
336 
337 	{ 0, 0 }
338 };
339 
340 MMU_DEF(oea64_mmu, "mmu_oea64_base", moea64_methods, 0);
341 
342 static struct pvo_head *
343 vm_page_to_pvoh(vm_page_t m)
344 {
345 
346 	mtx_assert(PV_LOCKPTR(VM_PAGE_TO_PHYS(m)), MA_OWNED);
347 	return (&m->md.mdpg_pvoh);
348 }
349 
350 static struct pvo_entry *
351 alloc_pvo_entry(int bootstrap)
352 {
353 	struct pvo_entry *pvo;
354 
355 	if (!moea64_initialized || bootstrap) {
356 		if (moea64_bpvo_pool_index >= moea64_bpvo_pool_size) {
357 			panic("moea64_enter: bpvo pool exhausted, %d, %d, %zd",
358 			      moea64_bpvo_pool_index, moea64_bpvo_pool_size,
359 			      moea64_bpvo_pool_size * sizeof(struct pvo_entry));
360 		}
361 		pvo = &moea64_bpvo_pool[
362 		    atomic_fetchadd_int(&moea64_bpvo_pool_index, 1)];
363 		bzero(pvo, sizeof(*pvo));
364 		pvo->pvo_vaddr = PVO_BOOTSTRAP;
365 	} else {
366 		pvo = uma_zalloc(moea64_pvo_zone, M_NOWAIT);
367 		bzero(pvo, sizeof(*pvo));
368 	}
369 
370 	return (pvo);
371 }
372 
373 
374 static void
375 init_pvo_entry(struct pvo_entry *pvo, pmap_t pmap, vm_offset_t va)
376 {
377 	uint64_t vsid;
378 	uint64_t hash;
379 	int shift;
380 
381 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
382 
383 	pvo->pvo_pmap = pmap;
384 	va &= ~ADDR_POFF;
385 	pvo->pvo_vaddr |= va;
386 	vsid = va_to_vsid(pmap, va);
387 	pvo->pvo_vpn = (uint64_t)((va & ADDR_PIDX) >> ADDR_PIDX_SHFT)
388 	    | (vsid << 16);
389 
390 	shift = (pvo->pvo_vaddr & PVO_LARGE) ? moea64_large_page_shift :
391 	    ADDR_PIDX_SHFT;
392 	hash = (vsid & VSID_HASH_MASK) ^ (((uint64_t)va & ADDR_PIDX) >> shift);
393 	pvo->pvo_pte.slot = (hash & moea64_pteg_mask) << 3;
394 }
395 
396 static void
397 free_pvo_entry(struct pvo_entry *pvo)
398 {
399 
400 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
401 		uma_zfree(moea64_pvo_zone, pvo);
402 }
403 
404 void
405 moea64_pte_from_pvo(const struct pvo_entry *pvo, struct lpte *lpte)
406 {
407 
408 	lpte->pte_hi = (pvo->pvo_vpn >> (ADDR_API_SHFT64 - ADDR_PIDX_SHFT)) &
409 	    LPTE_AVPN_MASK;
410 	lpte->pte_hi |= LPTE_VALID;
411 
412 	if (pvo->pvo_vaddr & PVO_LARGE)
413 		lpte->pte_hi |= LPTE_BIG;
414 	if (pvo->pvo_vaddr & PVO_WIRED)
415 		lpte->pte_hi |= LPTE_WIRED;
416 	if (pvo->pvo_vaddr & PVO_HID)
417 		lpte->pte_hi |= LPTE_HID;
418 
419 	lpte->pte_lo = pvo->pvo_pte.pa; /* Includes WIMG bits */
420 	if (pvo->pvo_pte.prot & VM_PROT_WRITE)
421 		lpte->pte_lo |= LPTE_BW;
422 	else
423 		lpte->pte_lo |= LPTE_BR;
424 
425 	if (!(pvo->pvo_pte.prot & VM_PROT_EXECUTE))
426 		lpte->pte_lo |= LPTE_NOEXEC;
427 }
428 
429 static __inline uint64_t
430 moea64_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
431 {
432 	uint64_t pte_lo;
433 	int i;
434 
435 	if (ma != VM_MEMATTR_DEFAULT) {
436 		switch (ma) {
437 		case VM_MEMATTR_UNCACHEABLE:
438 			return (LPTE_I | LPTE_G);
439 		case VM_MEMATTR_CACHEABLE:
440 			return (LPTE_M);
441 		case VM_MEMATTR_WRITE_COMBINING:
442 		case VM_MEMATTR_WRITE_BACK:
443 		case VM_MEMATTR_PREFETCHABLE:
444 			return (LPTE_I);
445 		case VM_MEMATTR_WRITE_THROUGH:
446 			return (LPTE_W | LPTE_M);
447 		}
448 	}
449 
450 	/*
451 	 * Assume the page is cache inhibited and access is guarded unless
452 	 * it's in our available memory array.
453 	 */
454 	pte_lo = LPTE_I | LPTE_G;
455 	for (i = 0; i < pregions_sz; i++) {
456 		if ((pa >= pregions[i].mr_start) &&
457 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
458 			pte_lo &= ~(LPTE_I | LPTE_G);
459 			pte_lo |= LPTE_M;
460 			break;
461 		}
462 	}
463 
464 	return pte_lo;
465 }
466 
467 /*
468  * Quick sort callout for comparing memory regions.
469  */
470 static int	om_cmp(const void *a, const void *b);
471 
472 static int
473 om_cmp(const void *a, const void *b)
474 {
475 	const struct	ofw_map *mapa;
476 	const struct	ofw_map *mapb;
477 
478 	mapa = a;
479 	mapb = b;
480 	if (mapa->om_pa < mapb->om_pa)
481 		return (-1);
482 	else if (mapa->om_pa > mapb->om_pa)
483 		return (1);
484 	else
485 		return (0);
486 }
487 
488 static void
489 moea64_add_ofw_mappings(mmu_t mmup, phandle_t mmu, size_t sz)
490 {
491 	struct ofw_map	translations[sz/(4*sizeof(cell_t))]; /*>= 4 cells per */
492 	pcell_t		acells, trans_cells[sz/sizeof(cell_t)];
493 	struct pvo_entry *pvo;
494 	register_t	msr;
495 	vm_offset_t	off;
496 	vm_paddr_t	pa_base;
497 	int		i, j;
498 
499 	bzero(translations, sz);
500 	OF_getencprop(OF_finddevice("/"), "#address-cells", &acells,
501 	    sizeof(acells));
502 	if (OF_getencprop(mmu, "translations", trans_cells, sz) == -1)
503 		panic("moea64_bootstrap: can't get ofw translations");
504 
505 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
506 	sz /= sizeof(cell_t);
507 	for (i = 0, j = 0; i < sz; j++) {
508 		translations[j].om_va = trans_cells[i++];
509 		translations[j].om_len = trans_cells[i++];
510 		translations[j].om_pa = trans_cells[i++];
511 		if (acells == 2) {
512 			translations[j].om_pa <<= 32;
513 			translations[j].om_pa |= trans_cells[i++];
514 		}
515 		translations[j].om_mode = trans_cells[i++];
516 	}
517 	KASSERT(i == sz, ("Translations map has incorrect cell count (%d/%zd)",
518 	    i, sz));
519 
520 	sz = j;
521 	qsort(translations, sz, sizeof (*translations), om_cmp);
522 
523 	for (i = 0; i < sz; i++) {
524 		pa_base = translations[i].om_pa;
525 	      #ifndef __powerpc64__
526 		if ((translations[i].om_pa >> 32) != 0)
527 			panic("OFW translations above 32-bit boundary!");
528 	      #endif
529 
530 		if (pa_base % PAGE_SIZE)
531 			panic("OFW translation not page-aligned (phys)!");
532 		if (translations[i].om_va % PAGE_SIZE)
533 			panic("OFW translation not page-aligned (virt)!");
534 
535 		CTR3(KTR_PMAP, "translation: pa=%#zx va=%#x len=%#x",
536 		    pa_base, translations[i].om_va, translations[i].om_len);
537 
538 		/* Now enter the pages for this mapping */
539 
540 		DISABLE_TRANS(msr);
541 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
542 			/* If this address is direct-mapped, skip remapping */
543 			if (hw_direct_map && translations[i].om_va == pa_base &&
544 			    moea64_calc_wimg(pa_base + off, VM_MEMATTR_DEFAULT) 			    == LPTE_M)
545 				continue;
546 
547 			PMAP_LOCK(kernel_pmap);
548 			pvo = moea64_pvo_find_va(kernel_pmap,
549 			    translations[i].om_va + off);
550 			PMAP_UNLOCK(kernel_pmap);
551 			if (pvo != NULL)
552 				continue;
553 
554 			moea64_kenter(mmup, translations[i].om_va + off,
555 			    pa_base + off);
556 		}
557 		ENABLE_TRANS(msr);
558 	}
559 }
560 
561 #ifdef __powerpc64__
562 static void
563 moea64_probe_large_page(void)
564 {
565 	uint16_t pvr = mfpvr() >> 16;
566 
567 	switch (pvr) {
568 	case IBM970:
569 	case IBM970FX:
570 	case IBM970MP:
571 		powerpc_sync(); isync();
572 		mtspr(SPR_HID4, mfspr(SPR_HID4) & ~HID4_970_DISABLE_LG_PG);
573 		powerpc_sync(); isync();
574 
575 		/* FALLTHROUGH */
576 	default:
577 		if (moea64_large_page_size == 0) {
578 			moea64_large_page_size = 0x1000000; /* 16 MB */
579 			moea64_large_page_shift = 24;
580 		}
581 	}
582 
583 	moea64_large_page_mask = moea64_large_page_size - 1;
584 }
585 
586 static void
587 moea64_bootstrap_slb_prefault(vm_offset_t va, int large)
588 {
589 	struct slb *cache;
590 	struct slb entry;
591 	uint64_t esid, slbe;
592 	uint64_t i;
593 
594 	cache = PCPU_GET(slb);
595 	esid = va >> ADDR_SR_SHFT;
596 	slbe = (esid << SLBE_ESID_SHIFT) | SLBE_VALID;
597 
598 	for (i = 0; i < 64; i++) {
599 		if (cache[i].slbe == (slbe | i))
600 			return;
601 	}
602 
603 	entry.slbe = slbe;
604 	entry.slbv = KERNEL_VSID(esid) << SLBV_VSID_SHIFT;
605 	if (large)
606 		entry.slbv |= SLBV_L;
607 
608 	slb_insert_kernel(entry.slbe, entry.slbv);
609 }
610 #endif
611 
612 static void
613 moea64_setup_direct_map(mmu_t mmup, vm_offset_t kernelstart,
614     vm_offset_t kernelend)
615 {
616 	struct pvo_entry *pvo;
617 	register_t msr;
618 	vm_paddr_t pa;
619 	vm_offset_t size, off;
620 	uint64_t pte_lo;
621 	int i;
622 
623 	if (moea64_large_page_size == 0)
624 		hw_direct_map = 0;
625 
626 	DISABLE_TRANS(msr);
627 	if (hw_direct_map) {
628 		PMAP_LOCK(kernel_pmap);
629 		for (i = 0; i < pregions_sz; i++) {
630 		  for (pa = pregions[i].mr_start; pa < pregions[i].mr_start +
631 		     pregions[i].mr_size; pa += moea64_large_page_size) {
632 			pte_lo = LPTE_M;
633 
634 			pvo = alloc_pvo_entry(1 /* bootstrap */);
635 			pvo->pvo_vaddr |= PVO_WIRED | PVO_LARGE;
636 			init_pvo_entry(pvo, kernel_pmap, pa);
637 
638 			/*
639 			 * Set memory access as guarded if prefetch within
640 			 * the page could exit the available physmem area.
641 			 */
642 			if (pa & moea64_large_page_mask) {
643 				pa &= moea64_large_page_mask;
644 				pte_lo |= LPTE_G;
645 			}
646 			if (pa + moea64_large_page_size >
647 			    pregions[i].mr_start + pregions[i].mr_size)
648 				pte_lo |= LPTE_G;
649 
650 			pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE |
651 			    VM_PROT_EXECUTE;
652 			pvo->pvo_pte.pa = pa | pte_lo;
653 			moea64_pvo_enter(mmup, pvo, NULL);
654 		  }
655 		}
656 		PMAP_UNLOCK(kernel_pmap);
657 	} else {
658 		size = moea64_bpvo_pool_size*sizeof(struct pvo_entry);
659 		off = (vm_offset_t)(moea64_bpvo_pool);
660 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
661 		moea64_kenter(mmup, pa, pa);
662 
663 		/*
664 		 * Map certain important things, like ourselves.
665 		 *
666 		 * NOTE: We do not map the exception vector space. That code is
667 		 * used only in real mode, and leaving it unmapped allows us to
668 		 * catch NULL pointer deferences, instead of making NULL a valid
669 		 * address.
670 		 */
671 
672 		for (pa = kernelstart & ~PAGE_MASK; pa < kernelend;
673 		    pa += PAGE_SIZE)
674 			moea64_kenter(mmup, pa, pa);
675 	}
676 	ENABLE_TRANS(msr);
677 
678 	/*
679 	 * Allow user to override unmapped_buf_allowed for testing.
680 	 * XXXKIB Only direct map implementation was tested.
681 	 */
682 	if (!TUNABLE_INT_FETCH("vfs.unmapped_buf_allowed",
683 	    &unmapped_buf_allowed))
684 		unmapped_buf_allowed = hw_direct_map;
685 }
686 
687 void
688 moea64_early_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
689 {
690 	int		i, j;
691 	vm_size_t	physsz, hwphyssz;
692 
693 #ifndef __powerpc64__
694 	/* We don't have a direct map since there is no BAT */
695 	hw_direct_map = 0;
696 
697 	/* Make sure battable is zero, since we have no BAT */
698 	for (i = 0; i < 16; i++) {
699 		battable[i].batu = 0;
700 		battable[i].batl = 0;
701 	}
702 #else
703 	moea64_probe_large_page();
704 
705 	/* Use a direct map if we have large page support */
706 	if (moea64_large_page_size > 0)
707 		hw_direct_map = 1;
708 	else
709 		hw_direct_map = 0;
710 #endif
711 
712 	/* Get physical memory regions from firmware */
713 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
714 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
715 
716 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
717 		panic("moea64_bootstrap: phys_avail too small");
718 
719 	phys_avail_count = 0;
720 	physsz = 0;
721 	hwphyssz = 0;
722 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
723 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
724 		CTR3(KTR_PMAP, "region: %#zx - %#zx (%#zx)",
725 		    regions[i].mr_start, regions[i].mr_start +
726 		    regions[i].mr_size, regions[i].mr_size);
727 		if (hwphyssz != 0 &&
728 		    (physsz + regions[i].mr_size) >= hwphyssz) {
729 			if (physsz < hwphyssz) {
730 				phys_avail[j] = regions[i].mr_start;
731 				phys_avail[j + 1] = regions[i].mr_start +
732 				    hwphyssz - physsz;
733 				physsz = hwphyssz;
734 				phys_avail_count++;
735 			}
736 			break;
737 		}
738 		phys_avail[j] = regions[i].mr_start;
739 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
740 		phys_avail_count++;
741 		physsz += regions[i].mr_size;
742 	}
743 
744 	/* Check for overlap with the kernel and exception vectors */
745 	for (j = 0; j < 2*phys_avail_count; j+=2) {
746 		if (phys_avail[j] < EXC_LAST)
747 			phys_avail[j] += EXC_LAST;
748 
749 		if (kernelstart >= phys_avail[j] &&
750 		    kernelstart < phys_avail[j+1]) {
751 			if (kernelend < phys_avail[j+1]) {
752 				phys_avail[2*phys_avail_count] =
753 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
754 				phys_avail[2*phys_avail_count + 1] =
755 				    phys_avail[j+1];
756 				phys_avail_count++;
757 			}
758 
759 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
760 		}
761 
762 		if (kernelend >= phys_avail[j] &&
763 		    kernelend < phys_avail[j+1]) {
764 			if (kernelstart > phys_avail[j]) {
765 				phys_avail[2*phys_avail_count] = phys_avail[j];
766 				phys_avail[2*phys_avail_count + 1] =
767 				    kernelstart & ~PAGE_MASK;
768 				phys_avail_count++;
769 			}
770 
771 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
772 		}
773 	}
774 
775 	physmem = btoc(physsz);
776 
777 #ifdef PTEGCOUNT
778 	moea64_pteg_count = PTEGCOUNT;
779 #else
780 	moea64_pteg_count = 0x1000;
781 
782 	while (moea64_pteg_count < physmem)
783 		moea64_pteg_count <<= 1;
784 
785 	moea64_pteg_count >>= 1;
786 #endif /* PTEGCOUNT */
787 }
788 
789 void
790 moea64_mid_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
791 {
792 	int		i;
793 
794 	/*
795 	 * Set PTEG mask
796 	 */
797 	moea64_pteg_mask = moea64_pteg_count - 1;
798 
799 	/*
800 	 * Initialize SLB table lock and page locks
801 	 */
802 	mtx_init(&moea64_slb_mutex, "SLB table", NULL, MTX_DEF);
803 	for (i = 0; i < PV_LOCK_COUNT; i++)
804 		mtx_init(&pv_lock[i], "page pv", NULL, MTX_DEF);
805 
806 	/*
807 	 * Initialise the bootstrap pvo pool.
808 	 */
809 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
810 		moea64_bpvo_pool_size*sizeof(struct pvo_entry), 0);
811 	moea64_bpvo_pool_index = 0;
812 
813 	/*
814 	 * Make sure kernel vsid is allocated as well as VSID 0.
815 	 */
816 	#ifndef __powerpc64__
817 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NVSIDS - 1)) / VSID_NBPW]
818 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
819 	moea64_vsid_bitmap[0] |= 1;
820 	#endif
821 
822 	/*
823 	 * Initialize the kernel pmap (which is statically allocated).
824 	 */
825 	#ifdef __powerpc64__
826 	for (i = 0; i < 64; i++) {
827 		pcpup->pc_slb[i].slbv = 0;
828 		pcpup->pc_slb[i].slbe = 0;
829 	}
830 	#else
831 	for (i = 0; i < 16; i++)
832 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
833 	#endif
834 
835 	kernel_pmap->pmap_phys = kernel_pmap;
836 	CPU_FILL(&kernel_pmap->pm_active);
837 	RB_INIT(&kernel_pmap->pmap_pvo);
838 
839 	PMAP_LOCK_INIT(kernel_pmap);
840 
841 	/*
842 	 * Now map in all the other buffers we allocated earlier
843 	 */
844 
845 	moea64_setup_direct_map(mmup, kernelstart, kernelend);
846 }
847 
848 void
849 moea64_late_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
850 {
851 	ihandle_t	mmui;
852 	phandle_t	chosen;
853 	phandle_t	mmu;
854 	ssize_t		sz;
855 	int		i;
856 	vm_offset_t	pa, va;
857 	void		*dpcpu;
858 
859 	/*
860 	 * Set up the Open Firmware pmap and add its mappings if not in real
861 	 * mode.
862 	 */
863 
864 	chosen = OF_finddevice("/chosen");
865 	if (chosen != -1 && OF_getencprop(chosen, "mmu", &mmui, 4) != -1) {
866 		mmu = OF_instance_to_package(mmui);
867 		if (mmu == -1 ||
868 		    (sz = OF_getproplen(mmu, "translations")) == -1)
869 			sz = 0;
870 		if (sz > 6144 /* tmpstksz - 2 KB headroom */)
871 			panic("moea64_bootstrap: too many ofw translations");
872 
873 		if (sz > 0)
874 			moea64_add_ofw_mappings(mmup, mmu, sz);
875 	}
876 
877 	/*
878 	 * Calculate the last available physical address.
879 	 */
880 	Maxmem = 0;
881 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
882 		Maxmem = max(Maxmem, powerpc_btop(phys_avail[i + 1]));
883 
884 	/*
885 	 * Initialize MMU and remap early physical mappings
886 	 */
887 	MMU_CPU_BOOTSTRAP(mmup,0);
888 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
889 	pmap_bootstrapped++;
890 	bs_remap_earlyboot();
891 
892 	/*
893 	 * Set the start and end of kva.
894 	 */
895 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
896 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
897 
898 	/*
899 	 * Map the entire KVA range into the SLB. We must not fault there.
900 	 */
901 	#ifdef __powerpc64__
902 	for (va = virtual_avail; va < virtual_end; va += SEGMENT_LENGTH)
903 		moea64_bootstrap_slb_prefault(va, 0);
904 	#endif
905 
906 	/*
907 	 * Figure out how far we can extend virtual_end into segment 16
908 	 * without running into existing mappings. Segment 16 is guaranteed
909 	 * to contain neither RAM nor devices (at least on Apple hardware),
910 	 * but will generally contain some OFW mappings we should not
911 	 * step on.
912 	 */
913 
914 	#ifndef __powerpc64__	/* KVA is in high memory on PPC64 */
915 	PMAP_LOCK(kernel_pmap);
916 	while (virtual_end < VM_MAX_KERNEL_ADDRESS &&
917 	    moea64_pvo_find_va(kernel_pmap, virtual_end+1) == NULL)
918 		virtual_end += PAGE_SIZE;
919 	PMAP_UNLOCK(kernel_pmap);
920 	#endif
921 
922 	/*
923 	 * Allocate a kernel stack with a guard page for thread0 and map it
924 	 * into the kernel page map.
925 	 */
926 	pa = moea64_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
927 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
928 	virtual_avail = va + kstack_pages * PAGE_SIZE;
929 	CTR2(KTR_PMAP, "moea64_bootstrap: kstack0 at %#x (%#x)", pa, va);
930 	thread0.td_kstack = va;
931 	thread0.td_kstack_pages = kstack_pages;
932 	for (i = 0; i < kstack_pages; i++) {
933 		moea64_kenter(mmup, va, pa);
934 		pa += PAGE_SIZE;
935 		va += PAGE_SIZE;
936 	}
937 
938 	/*
939 	 * Allocate virtual address space for the message buffer.
940 	 */
941 	pa = msgbuf_phys = moea64_bootstrap_alloc(msgbufsize, PAGE_SIZE);
942 	msgbufp = (struct msgbuf *)virtual_avail;
943 	va = virtual_avail;
944 	virtual_avail += round_page(msgbufsize);
945 	while (va < virtual_avail) {
946 		moea64_kenter(mmup, va, pa);
947 		pa += PAGE_SIZE;
948 		va += PAGE_SIZE;
949 	}
950 
951 	/*
952 	 * Allocate virtual address space for the dynamic percpu area.
953 	 */
954 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
955 	dpcpu = (void *)virtual_avail;
956 	va = virtual_avail;
957 	virtual_avail += DPCPU_SIZE;
958 	while (va < virtual_avail) {
959 		moea64_kenter(mmup, va, pa);
960 		pa += PAGE_SIZE;
961 		va += PAGE_SIZE;
962 	}
963 	dpcpu_init(dpcpu, curcpu);
964 
965 	/*
966 	 * Allocate some things for page zeroing. We put this directly
967 	 * in the page table and use MOEA64_PTE_REPLACE to avoid any
968 	 * of the PVO book-keeping or other parts of the VM system
969 	 * from even knowing that this hack exists.
970 	 */
971 
972 	if (!hw_direct_map) {
973 		mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL,
974 		    MTX_DEF);
975 		for (i = 0; i < 2; i++) {
976 			moea64_scratchpage_va[i] = (virtual_end+1) - PAGE_SIZE;
977 			virtual_end -= PAGE_SIZE;
978 
979 			moea64_kenter(mmup, moea64_scratchpage_va[i], 0);
980 
981 			PMAP_LOCK(kernel_pmap);
982 			moea64_scratchpage_pvo[i] = moea64_pvo_find_va(
983 			    kernel_pmap, (vm_offset_t)moea64_scratchpage_va[i]);
984 			PMAP_UNLOCK(kernel_pmap);
985 		}
986 	}
987 }
988 
989 static void
990 moea64_pmap_init_qpages(void)
991 {
992 	struct pcpu *pc;
993 	int i;
994 
995 	if (hw_direct_map)
996 		return;
997 
998 	CPU_FOREACH(i) {
999 		pc = pcpu_find(i);
1000 		pc->pc_qmap_addr = kva_alloc(PAGE_SIZE);
1001 		if (pc->pc_qmap_addr == 0)
1002 			panic("pmap_init_qpages: unable to allocate KVA");
1003 		PMAP_LOCK(kernel_pmap);
1004 		pc->pc_qmap_pvo = moea64_pvo_find_va(kernel_pmap, pc->pc_qmap_addr);
1005 		PMAP_UNLOCK(kernel_pmap);
1006 		mtx_init(&pc->pc_qmap_lock, "qmap lock", NULL, MTX_DEF);
1007 	}
1008 }
1009 
1010 SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, moea64_pmap_init_qpages, NULL);
1011 
1012 /*
1013  * Activate a user pmap.  This mostly involves setting some non-CPU
1014  * state.
1015  */
1016 void
1017 moea64_activate(mmu_t mmu, struct thread *td)
1018 {
1019 	pmap_t	pm;
1020 
1021 	pm = &td->td_proc->p_vmspace->vm_pmap;
1022 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1023 
1024 	#ifdef __powerpc64__
1025 	PCPU_SET(userslb, pm->pm_slb);
1026 	__asm __volatile("slbmte %0, %1; isync" ::
1027 	    "r"(td->td_pcb->pcb_cpu.aim.usr_vsid), "r"(USER_SLB_SLBE));
1028 	#else
1029 	PCPU_SET(curpmap, pm->pmap_phys);
1030 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1031 	#endif
1032 }
1033 
1034 void
1035 moea64_deactivate(mmu_t mmu, struct thread *td)
1036 {
1037 	pmap_t	pm;
1038 
1039 	__asm __volatile("isync; slbie %0" :: "r"(USER_ADDR));
1040 
1041 	pm = &td->td_proc->p_vmspace->vm_pmap;
1042 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1043 	#ifdef __powerpc64__
1044 	PCPU_SET(userslb, NULL);
1045 	#else
1046 	PCPU_SET(curpmap, NULL);
1047 	#endif
1048 }
1049 
1050 void
1051 moea64_unwire(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1052 {
1053 	struct	pvo_entry key, *pvo;
1054 	vm_page_t m;
1055 	int64_t	refchg;
1056 
1057 	key.pvo_vaddr = sva;
1058 	PMAP_LOCK(pm);
1059 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1060 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1061 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1062 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1063 			panic("moea64_unwire: pvo %p is missing PVO_WIRED",
1064 			    pvo);
1065 		pvo->pvo_vaddr &= ~PVO_WIRED;
1066 		refchg = MOEA64_PTE_REPLACE(mmu, pvo, 0 /* No invalidation */);
1067 		if ((pvo->pvo_vaddr & PVO_MANAGED) &&
1068 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1069 			if (refchg < 0)
1070 				refchg = LPTE_CHG;
1071 			m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
1072 
1073 			refchg |= atomic_readandclear_32(&m->md.mdpg_attrs);
1074 			if (refchg & LPTE_CHG)
1075 				vm_page_dirty(m);
1076 			if (refchg & LPTE_REF)
1077 				vm_page_aflag_set(m, PGA_REFERENCED);
1078 		}
1079 		pm->pm_stats.wired_count--;
1080 	}
1081 	PMAP_UNLOCK(pm);
1082 }
1083 
1084 /*
1085  * This goes through and sets the physical address of our
1086  * special scratch PTE to the PA we want to zero or copy. Because
1087  * of locking issues (this can get called in pvo_enter() by
1088  * the UMA allocator), we can't use most other utility functions here
1089  */
1090 
1091 static __inline
1092 void moea64_set_scratchpage_pa(mmu_t mmup, int which, vm_paddr_t pa) {
1093 
1094 	KASSERT(!hw_direct_map, ("Using OEA64 scratchpage with a direct map!"));
1095 	mtx_assert(&moea64_scratchpage_mtx, MA_OWNED);
1096 
1097 	moea64_scratchpage_pvo[which]->pvo_pte.pa =
1098 	    moea64_calc_wimg(pa, VM_MEMATTR_DEFAULT) | (uint64_t)pa;
1099 	MOEA64_PTE_REPLACE(mmup, moea64_scratchpage_pvo[which],
1100 	    MOEA64_PTE_INVALIDATE);
1101 	isync();
1102 }
1103 
1104 void
1105 moea64_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
1106 {
1107 	vm_offset_t	dst;
1108 	vm_offset_t	src;
1109 
1110 	dst = VM_PAGE_TO_PHYS(mdst);
1111 	src = VM_PAGE_TO_PHYS(msrc);
1112 
1113 	if (hw_direct_map) {
1114 		bcopy((void *)src, (void *)dst, PAGE_SIZE);
1115 	} else {
1116 		mtx_lock(&moea64_scratchpage_mtx);
1117 
1118 		moea64_set_scratchpage_pa(mmu, 0, src);
1119 		moea64_set_scratchpage_pa(mmu, 1, dst);
1120 
1121 		bcopy((void *)moea64_scratchpage_va[0],
1122 		    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1123 
1124 		mtx_unlock(&moea64_scratchpage_mtx);
1125 	}
1126 }
1127 
1128 static inline void
1129 moea64_copy_pages_dmap(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1130     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1131 {
1132 	void *a_cp, *b_cp;
1133 	vm_offset_t a_pg_offset, b_pg_offset;
1134 	int cnt;
1135 
1136 	while (xfersize > 0) {
1137 		a_pg_offset = a_offset & PAGE_MASK;
1138 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1139 		a_cp = (char *)VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]) +
1140 		    a_pg_offset;
1141 		b_pg_offset = b_offset & PAGE_MASK;
1142 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1143 		b_cp = (char *)VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]) +
1144 		    b_pg_offset;
1145 		bcopy(a_cp, b_cp, cnt);
1146 		a_offset += cnt;
1147 		b_offset += cnt;
1148 		xfersize -= cnt;
1149 	}
1150 }
1151 
1152 static inline void
1153 moea64_copy_pages_nodmap(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1154     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1155 {
1156 	void *a_cp, *b_cp;
1157 	vm_offset_t a_pg_offset, b_pg_offset;
1158 	int cnt;
1159 
1160 	mtx_lock(&moea64_scratchpage_mtx);
1161 	while (xfersize > 0) {
1162 		a_pg_offset = a_offset & PAGE_MASK;
1163 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1164 		moea64_set_scratchpage_pa(mmu, 0,
1165 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]));
1166 		a_cp = (char *)moea64_scratchpage_va[0] + a_pg_offset;
1167 		b_pg_offset = b_offset & PAGE_MASK;
1168 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1169 		moea64_set_scratchpage_pa(mmu, 1,
1170 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]));
1171 		b_cp = (char *)moea64_scratchpage_va[1] + b_pg_offset;
1172 		bcopy(a_cp, b_cp, cnt);
1173 		a_offset += cnt;
1174 		b_offset += cnt;
1175 		xfersize -= cnt;
1176 	}
1177 	mtx_unlock(&moea64_scratchpage_mtx);
1178 }
1179 
1180 void
1181 moea64_copy_pages(mmu_t mmu, vm_page_t *ma, vm_offset_t a_offset,
1182     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1183 {
1184 
1185 	if (hw_direct_map) {
1186 		moea64_copy_pages_dmap(mmu, ma, a_offset, mb, b_offset,
1187 		    xfersize);
1188 	} else {
1189 		moea64_copy_pages_nodmap(mmu, ma, a_offset, mb, b_offset,
1190 		    xfersize);
1191 	}
1192 }
1193 
1194 void
1195 moea64_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
1196 {
1197 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1198 
1199 	if (size + off > PAGE_SIZE)
1200 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1201 
1202 	if (hw_direct_map) {
1203 		bzero((caddr_t)pa + off, size);
1204 	} else {
1205 		mtx_lock(&moea64_scratchpage_mtx);
1206 		moea64_set_scratchpage_pa(mmu, 0, pa);
1207 		bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1208 		mtx_unlock(&moea64_scratchpage_mtx);
1209 	}
1210 }
1211 
1212 /*
1213  * Zero a page of physical memory by temporarily mapping it
1214  */
1215 void
1216 moea64_zero_page(mmu_t mmu, vm_page_t m)
1217 {
1218 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1219 	vm_offset_t va, off;
1220 
1221 	if (!hw_direct_map) {
1222 		mtx_lock(&moea64_scratchpage_mtx);
1223 
1224 		moea64_set_scratchpage_pa(mmu, 0, pa);
1225 		va = moea64_scratchpage_va[0];
1226 	} else {
1227 		va = pa;
1228 	}
1229 
1230 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1231 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1232 
1233 	if (!hw_direct_map)
1234 		mtx_unlock(&moea64_scratchpage_mtx);
1235 }
1236 
1237 vm_offset_t
1238 moea64_quick_enter_page(mmu_t mmu, vm_page_t m)
1239 {
1240 	struct pvo_entry *pvo;
1241 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1242 
1243 	if (hw_direct_map)
1244 		return (pa);
1245 
1246 	/*
1247  	 * MOEA64_PTE_REPLACE does some locking, so we can't just grab
1248 	 * a critical section and access the PCPU data like on i386.
1249 	 * Instead, pin the thread and grab the PCPU lock to prevent
1250 	 * a preempting thread from using the same PCPU data.
1251 	 */
1252 	sched_pin();
1253 
1254 	mtx_assert(PCPU_PTR(qmap_lock), MA_NOTOWNED);
1255 	pvo = PCPU_GET(qmap_pvo);
1256 
1257 	mtx_lock(PCPU_PTR(qmap_lock));
1258 	pvo->pvo_pte.pa = moea64_calc_wimg(pa, pmap_page_get_memattr(m)) |
1259 	    (uint64_t)pa;
1260 	MOEA64_PTE_REPLACE(mmu, pvo, MOEA64_PTE_INVALIDATE);
1261 	isync();
1262 
1263 	return (PCPU_GET(qmap_addr));
1264 }
1265 
1266 void
1267 moea64_quick_remove_page(mmu_t mmu, vm_offset_t addr)
1268 {
1269 	if (hw_direct_map)
1270 		return;
1271 
1272 	mtx_assert(PCPU_PTR(qmap_lock), MA_OWNED);
1273 	KASSERT(PCPU_GET(qmap_addr) == addr,
1274 	    ("moea64_quick_remove_page: invalid address"));
1275 	mtx_unlock(PCPU_PTR(qmap_lock));
1276 	sched_unpin();
1277 }
1278 
1279 /*
1280  * Map the given physical page at the specified virtual address in the
1281  * target pmap with the protection requested.  If specified the page
1282  * will be wired down.
1283  */
1284 
1285 int
1286 moea64_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m,
1287     vm_prot_t prot, u_int flags, int8_t psind)
1288 {
1289 	struct		pvo_entry *pvo, *oldpvo;
1290 	struct		pvo_head *pvo_head;
1291 	uint64_t	pte_lo;
1292 	int		error;
1293 
1294 	if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m))
1295 		VM_OBJECT_ASSERT_LOCKED(m->object);
1296 
1297 	pvo = alloc_pvo_entry(0);
1298 	pvo->pvo_pmap = NULL; /* to be filled in later */
1299 	pvo->pvo_pte.prot = prot;
1300 
1301 	pte_lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1302 	pvo->pvo_pte.pa = VM_PAGE_TO_PHYS(m) | pte_lo;
1303 
1304 	if ((flags & PMAP_ENTER_WIRED) != 0)
1305 		pvo->pvo_vaddr |= PVO_WIRED;
1306 
1307 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea64_initialized) {
1308 		pvo_head = NULL;
1309 	} else {
1310 		pvo_head = &m->md.mdpg_pvoh;
1311 		pvo->pvo_vaddr |= PVO_MANAGED;
1312 	}
1313 
1314 	for (;;) {
1315 		PV_PAGE_LOCK(m);
1316 		PMAP_LOCK(pmap);
1317 		if (pvo->pvo_pmap == NULL)
1318 			init_pvo_entry(pvo, pmap, va);
1319 		if (prot & VM_PROT_WRITE)
1320 			if (pmap_bootstrapped &&
1321 			    (m->oflags & VPO_UNMANAGED) == 0)
1322 				vm_page_aflag_set(m, PGA_WRITEABLE);
1323 
1324 		oldpvo = moea64_pvo_find_va(pmap, va);
1325 		if (oldpvo != NULL) {
1326 			if (oldpvo->pvo_vaddr == pvo->pvo_vaddr &&
1327 			    oldpvo->pvo_pte.pa == pvo->pvo_pte.pa &&
1328 			    oldpvo->pvo_pte.prot == prot) {
1329 				/* Identical mapping already exists */
1330 				error = 0;
1331 
1332 				/* If not in page table, reinsert it */
1333 				if (MOEA64_PTE_SYNCH(mmu, oldpvo) < 0) {
1334 					moea64_pte_overflow--;
1335 					MOEA64_PTE_INSERT(mmu, oldpvo);
1336 				}
1337 
1338 				/* Then just clean up and go home */
1339 				PV_PAGE_UNLOCK(m);
1340 				PMAP_UNLOCK(pmap);
1341 				free_pvo_entry(pvo);
1342 				break;
1343 			}
1344 
1345 			/* Otherwise, need to kill it first */
1346 			KASSERT(oldpvo->pvo_pmap == pmap, ("pmap of old "
1347 			    "mapping does not match new mapping"));
1348 			moea64_pvo_remove_from_pmap(mmu, oldpvo);
1349 		}
1350 		error = moea64_pvo_enter(mmu, pvo, pvo_head);
1351 		PV_PAGE_UNLOCK(m);
1352 		PMAP_UNLOCK(pmap);
1353 
1354 		/* Free any dead pages */
1355 		if (oldpvo != NULL) {
1356 			PV_LOCK(oldpvo->pvo_pte.pa & LPTE_RPGN);
1357 			moea64_pvo_remove_from_page(mmu, oldpvo);
1358 			PV_UNLOCK(oldpvo->pvo_pte.pa & LPTE_RPGN);
1359 			free_pvo_entry(oldpvo);
1360 		}
1361 
1362 		if (error != ENOMEM)
1363 			break;
1364 		if ((flags & PMAP_ENTER_NOSLEEP) != 0)
1365 			return (KERN_RESOURCE_SHORTAGE);
1366 		VM_OBJECT_ASSERT_UNLOCKED(m->object);
1367 		VM_WAIT;
1368 	}
1369 
1370 	/*
1371 	 * Flush the page from the instruction cache if this page is
1372 	 * mapped executable and cacheable.
1373 	 */
1374 	if (pmap != kernel_pmap && !(m->aflags & PGA_EXECUTABLE) &&
1375 	    (pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1376 		vm_page_aflag_set(m, PGA_EXECUTABLE);
1377 		moea64_syncicache(mmu, pmap, va, VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1378 	}
1379 	return (KERN_SUCCESS);
1380 }
1381 
1382 static void
1383 moea64_syncicache(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
1384     vm_size_t sz)
1385 {
1386 
1387 	/*
1388 	 * This is much trickier than on older systems because
1389 	 * we can't sync the icache on physical addresses directly
1390 	 * without a direct map. Instead we check a couple of cases
1391 	 * where the memory is already mapped in and, failing that,
1392 	 * use the same trick we use for page zeroing to create
1393 	 * a temporary mapping for this physical address.
1394 	 */
1395 
1396 	if (!pmap_bootstrapped) {
1397 		/*
1398 		 * If PMAP is not bootstrapped, we are likely to be
1399 		 * in real mode.
1400 		 */
1401 		__syncicache((void *)pa, sz);
1402 	} else if (pmap == kernel_pmap) {
1403 		__syncicache((void *)va, sz);
1404 	} else if (hw_direct_map) {
1405 		__syncicache((void *)pa, sz);
1406 	} else {
1407 		/* Use the scratch page to set up a temp mapping */
1408 
1409 		mtx_lock(&moea64_scratchpage_mtx);
1410 
1411 		moea64_set_scratchpage_pa(mmu, 1, pa & ~ADDR_POFF);
1412 		__syncicache((void *)(moea64_scratchpage_va[1] +
1413 		    (va & ADDR_POFF)), sz);
1414 
1415 		mtx_unlock(&moea64_scratchpage_mtx);
1416 	}
1417 }
1418 
1419 /*
1420  * Maps a sequence of resident pages belonging to the same object.
1421  * The sequence begins with the given page m_start.  This page is
1422  * mapped at the given virtual address start.  Each subsequent page is
1423  * mapped at a virtual address that is offset from start by the same
1424  * amount as the page is offset from m_start within the object.  The
1425  * last page in the sequence is the page with the largest offset from
1426  * m_start that can be mapped at a virtual address less than the given
1427  * virtual address end.  Not every virtual page between start and end
1428  * is mapped; only those for which a resident page exists with the
1429  * corresponding offset from m_start are mapped.
1430  */
1431 void
1432 moea64_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1433     vm_page_t m_start, vm_prot_t prot)
1434 {
1435 	vm_page_t m;
1436 	vm_pindex_t diff, psize;
1437 
1438 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1439 
1440 	psize = atop(end - start);
1441 	m = m_start;
1442 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1443 		moea64_enter(mmu, pm, start + ptoa(diff), m, prot &
1444 		    (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_NOSLEEP, 0);
1445 		m = TAILQ_NEXT(m, listq);
1446 	}
1447 }
1448 
1449 void
1450 moea64_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1451     vm_prot_t prot)
1452 {
1453 
1454 	moea64_enter(mmu, pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1455 	    PMAP_ENTER_NOSLEEP, 0);
1456 }
1457 
1458 vm_paddr_t
1459 moea64_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1460 {
1461 	struct	pvo_entry *pvo;
1462 	vm_paddr_t pa;
1463 
1464 	PMAP_LOCK(pm);
1465 	pvo = moea64_pvo_find_va(pm, va);
1466 	if (pvo == NULL)
1467 		pa = 0;
1468 	else
1469 		pa = (pvo->pvo_pte.pa & LPTE_RPGN) | (va - PVO_VADDR(pvo));
1470 	PMAP_UNLOCK(pm);
1471 
1472 	return (pa);
1473 }
1474 
1475 /*
1476  * Atomically extract and hold the physical page with the given
1477  * pmap and virtual address pair if that mapping permits the given
1478  * protection.
1479  */
1480 vm_page_t
1481 moea64_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1482 {
1483 	struct	pvo_entry *pvo;
1484 	vm_page_t m;
1485         vm_paddr_t pa;
1486 
1487 	m = NULL;
1488 	pa = 0;
1489 	PMAP_LOCK(pmap);
1490 retry:
1491 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1492 	if (pvo != NULL && (pvo->pvo_pte.prot & prot) == prot) {
1493 		if (vm_page_pa_tryrelock(pmap,
1494 		    pvo->pvo_pte.pa & LPTE_RPGN, &pa))
1495 			goto retry;
1496 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
1497 		vm_page_hold(m);
1498 	}
1499 	PA_UNLOCK_COND(pa);
1500 	PMAP_UNLOCK(pmap);
1501 	return (m);
1502 }
1503 
1504 static mmu_t installed_mmu;
1505 
1506 static void *
1507 moea64_uma_page_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *flags,
1508     int wait)
1509 {
1510 	struct pvo_entry *pvo;
1511         vm_offset_t va;
1512         vm_page_t m;
1513         int needed_lock;
1514 
1515 	/*
1516 	 * This entire routine is a horrible hack to avoid bothering kmem
1517 	 * for new KVA addresses. Because this can get called from inside
1518 	 * kmem allocation routines, calling kmem for a new address here
1519 	 * can lead to multiply locking non-recursive mutexes.
1520 	 */
1521 
1522 	*flags = UMA_SLAB_PRIV;
1523 	needed_lock = !PMAP_LOCKED(kernel_pmap);
1524 
1525 	m = vm_page_alloc(NULL, 0,
1526 	    malloc2vm_flags(wait) | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1527 	if (m == NULL)
1528 		return (NULL);
1529 
1530 	va = VM_PAGE_TO_PHYS(m);
1531 
1532 	pvo = alloc_pvo_entry(1 /* bootstrap */);
1533 
1534 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE;
1535 	pvo->pvo_pte.pa = VM_PAGE_TO_PHYS(m) | LPTE_M;
1536 
1537 	if (needed_lock)
1538 		PMAP_LOCK(kernel_pmap);
1539 
1540 	init_pvo_entry(pvo, kernel_pmap, va);
1541 	pvo->pvo_vaddr |= PVO_WIRED;
1542 
1543 	moea64_pvo_enter(installed_mmu, pvo, NULL);
1544 
1545 	if (needed_lock)
1546 		PMAP_UNLOCK(kernel_pmap);
1547 
1548 	if ((wait & M_ZERO) && (m->flags & PG_ZERO) == 0)
1549                 bzero((void *)va, PAGE_SIZE);
1550 
1551 	return (void *)va;
1552 }
1553 
1554 extern int elf32_nxstack;
1555 
1556 void
1557 moea64_init(mmu_t mmu)
1558 {
1559 
1560 	CTR0(KTR_PMAP, "moea64_init");
1561 
1562 	moea64_pvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1563 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1564 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1565 
1566 	if (!hw_direct_map) {
1567 		installed_mmu = mmu;
1568 		uma_zone_set_allocf(moea64_pvo_zone,moea64_uma_page_alloc);
1569 	}
1570 
1571 #ifdef COMPAT_FREEBSD32
1572 	elf32_nxstack = 1;
1573 #endif
1574 
1575 	moea64_initialized = TRUE;
1576 }
1577 
1578 boolean_t
1579 moea64_is_referenced(mmu_t mmu, vm_page_t m)
1580 {
1581 
1582 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1583 	    ("moea64_is_referenced: page %p is not managed", m));
1584 
1585 	return (moea64_query_bit(mmu, m, LPTE_REF));
1586 }
1587 
1588 boolean_t
1589 moea64_is_modified(mmu_t mmu, vm_page_t m)
1590 {
1591 
1592 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1593 	    ("moea64_is_modified: page %p is not managed", m));
1594 
1595 	/*
1596 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1597 	 * concurrently set while the object is locked.  Thus, if PGA_WRITEABLE
1598 	 * is clear, no PTEs can have LPTE_CHG set.
1599 	 */
1600 	VM_OBJECT_ASSERT_LOCKED(m->object);
1601 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1602 		return (FALSE);
1603 	return (moea64_query_bit(mmu, m, LPTE_CHG));
1604 }
1605 
1606 boolean_t
1607 moea64_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1608 {
1609 	struct pvo_entry *pvo;
1610 	boolean_t rv = TRUE;
1611 
1612 	PMAP_LOCK(pmap);
1613 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1614 	if (pvo != NULL)
1615 		rv = FALSE;
1616 	PMAP_UNLOCK(pmap);
1617 	return (rv);
1618 }
1619 
1620 void
1621 moea64_clear_modify(mmu_t mmu, vm_page_t m)
1622 {
1623 
1624 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1625 	    ("moea64_clear_modify: page %p is not managed", m));
1626 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1627 	KASSERT(!vm_page_xbusied(m),
1628 	    ("moea64_clear_modify: page %p is exclusive busied", m));
1629 
1630 	/*
1631 	 * If the page is not PGA_WRITEABLE, then no PTEs can have LPTE_CHG
1632 	 * set.  If the object containing the page is locked and the page is
1633 	 * not exclusive busied, then PGA_WRITEABLE cannot be concurrently set.
1634 	 */
1635 	if ((m->aflags & PGA_WRITEABLE) == 0)
1636 		return;
1637 	moea64_clear_bit(mmu, m, LPTE_CHG);
1638 }
1639 
1640 /*
1641  * Clear the write and modified bits in each of the given page's mappings.
1642  */
1643 void
1644 moea64_remove_write(mmu_t mmu, vm_page_t m)
1645 {
1646 	struct	pvo_entry *pvo;
1647 	int64_t	refchg, ret;
1648 	pmap_t	pmap;
1649 
1650 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1651 	    ("moea64_remove_write: page %p is not managed", m));
1652 
1653 	/*
1654 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1655 	 * set by another thread while the object is locked.  Thus,
1656 	 * if PGA_WRITEABLE is clear, no page table entries need updating.
1657 	 */
1658 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1659 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1660 		return;
1661 	powerpc_sync();
1662 	PV_PAGE_LOCK(m);
1663 	refchg = 0;
1664 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1665 		pmap = pvo->pvo_pmap;
1666 		PMAP_LOCK(pmap);
1667 		if (!(pvo->pvo_vaddr & PVO_DEAD) &&
1668 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1669 			pvo->pvo_pte.prot &= ~VM_PROT_WRITE;
1670 			ret = MOEA64_PTE_REPLACE(mmu, pvo,
1671 			    MOEA64_PTE_PROT_UPDATE);
1672 			if (ret < 0)
1673 				ret = LPTE_CHG;
1674 			refchg |= ret;
1675 			if (pvo->pvo_pmap == kernel_pmap)
1676 				isync();
1677 		}
1678 		PMAP_UNLOCK(pmap);
1679 	}
1680 	if ((refchg | atomic_readandclear_32(&m->md.mdpg_attrs)) & LPTE_CHG)
1681 		vm_page_dirty(m);
1682 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1683 	PV_PAGE_UNLOCK(m);
1684 }
1685 
1686 /*
1687  *	moea64_ts_referenced:
1688  *
1689  *	Return a count of reference bits for a page, clearing those bits.
1690  *	It is not necessary for every reference bit to be cleared, but it
1691  *	is necessary that 0 only be returned when there are truly no
1692  *	reference bits set.
1693  *
1694  *	XXX: The exact number of bits to check and clear is a matter that
1695  *	should be tested and standardized at some point in the future for
1696  *	optimal aging of shared pages.
1697  */
1698 int
1699 moea64_ts_referenced(mmu_t mmu, vm_page_t m)
1700 {
1701 
1702 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1703 	    ("moea64_ts_referenced: page %p is not managed", m));
1704 	return (moea64_clear_bit(mmu, m, LPTE_REF));
1705 }
1706 
1707 /*
1708  * Modify the WIMG settings of all mappings for a page.
1709  */
1710 void
1711 moea64_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1712 {
1713 	struct	pvo_entry *pvo;
1714 	int64_t	refchg;
1715 	pmap_t	pmap;
1716 	uint64_t lo;
1717 
1718 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1719 		m->md.mdpg_cache_attrs = ma;
1720 		return;
1721 	}
1722 
1723 	lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1724 
1725 	PV_PAGE_LOCK(m);
1726 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1727 		pmap = pvo->pvo_pmap;
1728 		PMAP_LOCK(pmap);
1729 		if (!(pvo->pvo_vaddr & PVO_DEAD)) {
1730 			pvo->pvo_pte.pa &= ~LPTE_WIMG;
1731 			pvo->pvo_pte.pa |= lo;
1732 			refchg = MOEA64_PTE_REPLACE(mmu, pvo,
1733 			    MOEA64_PTE_INVALIDATE);
1734 			if (refchg < 0)
1735 				refchg = (pvo->pvo_pte.prot & VM_PROT_WRITE) ?
1736 				    LPTE_CHG : 0;
1737 			if ((pvo->pvo_vaddr & PVO_MANAGED) &&
1738 			    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1739 				refchg |=
1740 				    atomic_readandclear_32(&m->md.mdpg_attrs);
1741 				if (refchg & LPTE_CHG)
1742 					vm_page_dirty(m);
1743 				if (refchg & LPTE_REF)
1744 					vm_page_aflag_set(m, PGA_REFERENCED);
1745 			}
1746 			if (pvo->pvo_pmap == kernel_pmap)
1747 				isync();
1748 		}
1749 		PMAP_UNLOCK(pmap);
1750 	}
1751 	m->md.mdpg_cache_attrs = ma;
1752 	PV_PAGE_UNLOCK(m);
1753 }
1754 
1755 /*
1756  * Map a wired page into kernel virtual address space.
1757  */
1758 void
1759 moea64_kenter_attr(mmu_t mmu, vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
1760 {
1761 	int		error;
1762 	struct pvo_entry *pvo, *oldpvo;
1763 
1764 	pvo = alloc_pvo_entry(0);
1765 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;
1766 	pvo->pvo_pte.pa = (pa & ~ADDR_POFF) | moea64_calc_wimg(pa, ma);
1767 	pvo->pvo_vaddr |= PVO_WIRED;
1768 
1769 	PMAP_LOCK(kernel_pmap);
1770 	oldpvo = moea64_pvo_find_va(kernel_pmap, va);
1771 	if (oldpvo != NULL)
1772 		moea64_pvo_remove_from_pmap(mmu, oldpvo);
1773 	init_pvo_entry(pvo, kernel_pmap, va);
1774 	error = moea64_pvo_enter(mmu, pvo, NULL);
1775 	PMAP_UNLOCK(kernel_pmap);
1776 
1777 	/* Free any dead pages */
1778 	if (oldpvo != NULL) {
1779 		PV_LOCK(oldpvo->pvo_pte.pa & LPTE_RPGN);
1780 		moea64_pvo_remove_from_page(mmu, oldpvo);
1781 		PV_UNLOCK(oldpvo->pvo_pte.pa & LPTE_RPGN);
1782 		free_pvo_entry(oldpvo);
1783 	}
1784 
1785 	if (error != 0 && error != ENOENT)
1786 		panic("moea64_kenter: failed to enter va %#zx pa %#zx: %d", va,
1787 		    pa, error);
1788 }
1789 
1790 void
1791 moea64_kenter(mmu_t mmu, vm_offset_t va, vm_paddr_t pa)
1792 {
1793 
1794 	moea64_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1795 }
1796 
1797 /*
1798  * Extract the physical page address associated with the given kernel virtual
1799  * address.
1800  */
1801 vm_paddr_t
1802 moea64_kextract(mmu_t mmu, vm_offset_t va)
1803 {
1804 	struct		pvo_entry *pvo;
1805 	vm_paddr_t pa;
1806 
1807 	/*
1808 	 * Shortcut the direct-mapped case when applicable.  We never put
1809 	 * anything but 1:1 mappings below VM_MIN_KERNEL_ADDRESS.
1810 	 */
1811 	if (va < VM_MIN_KERNEL_ADDRESS)
1812 		return (va);
1813 
1814 	PMAP_LOCK(kernel_pmap);
1815 	pvo = moea64_pvo_find_va(kernel_pmap, va);
1816 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found for %#" PRIxPTR,
1817 	    va));
1818 	pa = (pvo->pvo_pte.pa & LPTE_RPGN) | (va - PVO_VADDR(pvo));
1819 	PMAP_UNLOCK(kernel_pmap);
1820 	return (pa);
1821 }
1822 
1823 /*
1824  * Remove a wired page from kernel virtual address space.
1825  */
1826 void
1827 moea64_kremove(mmu_t mmu, vm_offset_t va)
1828 {
1829 	moea64_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1830 }
1831 
1832 /*
1833  * Map a range of physical addresses into kernel virtual address space.
1834  *
1835  * The value passed in *virt is a suggested virtual address for the mapping.
1836  * Architectures which can support a direct-mapped physical to virtual region
1837  * can return the appropriate address within that region, leaving '*virt'
1838  * unchanged.  Other architectures should map the pages starting at '*virt' and
1839  * update '*virt' with the first usable address after the mapped region.
1840  */
1841 vm_offset_t
1842 moea64_map(mmu_t mmu, vm_offset_t *virt, vm_paddr_t pa_start,
1843     vm_paddr_t pa_end, int prot)
1844 {
1845 	vm_offset_t	sva, va;
1846 
1847 	if (hw_direct_map) {
1848 		/*
1849 		 * Check if every page in the region is covered by the direct
1850 		 * map. The direct map covers all of physical memory. Use
1851 		 * moea64_calc_wimg() as a shortcut to see if the page is in
1852 		 * physical memory as a way to see if the direct map covers it.
1853 		 */
1854 		for (va = pa_start; va < pa_end; va += PAGE_SIZE)
1855 			if (moea64_calc_wimg(va, VM_MEMATTR_DEFAULT) != LPTE_M)
1856 				break;
1857 		if (va == pa_end)
1858 			return (pa_start);
1859 	}
1860 	sva = *virt;
1861 	va = sva;
1862 	/* XXX respect prot argument */
1863 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1864 		moea64_kenter(mmu, va, pa_start);
1865 	*virt = va;
1866 
1867 	return (sva);
1868 }
1869 
1870 /*
1871  * Returns true if the pmap's pv is one of the first
1872  * 16 pvs linked to from this page.  This count may
1873  * be changed upwards or downwards in the future; it
1874  * is only necessary that true be returned for a small
1875  * subset of pmaps for proper page aging.
1876  */
1877 boolean_t
1878 moea64_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1879 {
1880         int loops;
1881 	struct pvo_entry *pvo;
1882 	boolean_t rv;
1883 
1884 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1885 	    ("moea64_page_exists_quick: page %p is not managed", m));
1886 	loops = 0;
1887 	rv = FALSE;
1888 	PV_PAGE_LOCK(m);
1889 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1890 		if (!(pvo->pvo_vaddr & PVO_DEAD) && pvo->pvo_pmap == pmap) {
1891 			rv = TRUE;
1892 			break;
1893 		}
1894 		if (++loops >= 16)
1895 			break;
1896 	}
1897 	PV_PAGE_UNLOCK(m);
1898 	return (rv);
1899 }
1900 
1901 void
1902 moea64_page_init(mmu_t mmu __unused, vm_page_t m)
1903 {
1904 
1905 	m->md.mdpg_attrs = 0;
1906 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
1907 	LIST_INIT(&m->md.mdpg_pvoh);
1908 }
1909 
1910 /*
1911  * Return the number of managed mappings to the given physical page
1912  * that are wired.
1913  */
1914 int
1915 moea64_page_wired_mappings(mmu_t mmu, vm_page_t m)
1916 {
1917 	struct pvo_entry *pvo;
1918 	int count;
1919 
1920 	count = 0;
1921 	if ((m->oflags & VPO_UNMANAGED) != 0)
1922 		return (count);
1923 	PV_PAGE_LOCK(m);
1924 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1925 		if ((pvo->pvo_vaddr & (PVO_DEAD | PVO_WIRED)) == PVO_WIRED)
1926 			count++;
1927 	PV_PAGE_UNLOCK(m);
1928 	return (count);
1929 }
1930 
1931 static uintptr_t	moea64_vsidcontext;
1932 
1933 uintptr_t
1934 moea64_get_unique_vsid(void) {
1935 	u_int entropy;
1936 	register_t hash;
1937 	uint32_t mask;
1938 	int i;
1939 
1940 	entropy = 0;
1941 	__asm __volatile("mftb %0" : "=r"(entropy));
1942 
1943 	mtx_lock(&moea64_slb_mutex);
1944 	for (i = 0; i < NVSIDS; i += VSID_NBPW) {
1945 		u_int	n;
1946 
1947 		/*
1948 		 * Create a new value by mutiplying by a prime and adding in
1949 		 * entropy from the timebase register.  This is to make the
1950 		 * VSID more random so that the PT hash function collides
1951 		 * less often.  (Note that the prime casues gcc to do shifts
1952 		 * instead of a multiply.)
1953 		 */
1954 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
1955 		hash = moea64_vsidcontext & (NVSIDS - 1);
1956 		if (hash == 0)		/* 0 is special, avoid it */
1957 			continue;
1958 		n = hash >> 5;
1959 		mask = 1 << (hash & (VSID_NBPW - 1));
1960 		hash = (moea64_vsidcontext & VSID_HASHMASK);
1961 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
1962 			/* anything free in this bucket? */
1963 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
1964 				entropy = (moea64_vsidcontext >> 20);
1965 				continue;
1966 			}
1967 			i = ffs(~moea64_vsid_bitmap[n]) - 1;
1968 			mask = 1 << i;
1969 			hash &= rounddown2(VSID_HASHMASK, VSID_NBPW);
1970 			hash |= i;
1971 		}
1972 		if (hash == VSID_VRMA)	/* also special, avoid this too */
1973 			continue;
1974 		KASSERT(!(moea64_vsid_bitmap[n] & mask),
1975 		    ("Allocating in-use VSID %#zx\n", hash));
1976 		moea64_vsid_bitmap[n] |= mask;
1977 		mtx_unlock(&moea64_slb_mutex);
1978 		return (hash);
1979 	}
1980 
1981 	mtx_unlock(&moea64_slb_mutex);
1982 	panic("%s: out of segments",__func__);
1983 }
1984 
1985 #ifdef __powerpc64__
1986 void
1987 moea64_pinit(mmu_t mmu, pmap_t pmap)
1988 {
1989 
1990 	RB_INIT(&pmap->pmap_pvo);
1991 
1992 	pmap->pm_slb_tree_root = slb_alloc_tree();
1993 	pmap->pm_slb = slb_alloc_user_cache();
1994 	pmap->pm_slb_len = 0;
1995 }
1996 #else
1997 void
1998 moea64_pinit(mmu_t mmu, pmap_t pmap)
1999 {
2000 	int	i;
2001 	uint32_t hash;
2002 
2003 	RB_INIT(&pmap->pmap_pvo);
2004 
2005 	if (pmap_bootstrapped)
2006 		pmap->pmap_phys = (pmap_t)moea64_kextract(mmu,
2007 		    (vm_offset_t)pmap);
2008 	else
2009 		pmap->pmap_phys = pmap;
2010 
2011 	/*
2012 	 * Allocate some segment registers for this pmap.
2013 	 */
2014 	hash = moea64_get_unique_vsid();
2015 
2016 	for (i = 0; i < 16; i++)
2017 		pmap->pm_sr[i] = VSID_MAKE(i, hash);
2018 
2019 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_pinit: pm_sr[0] = 0"));
2020 }
2021 #endif
2022 
2023 /*
2024  * Initialize the pmap associated with process 0.
2025  */
2026 void
2027 moea64_pinit0(mmu_t mmu, pmap_t pm)
2028 {
2029 
2030 	PMAP_LOCK_INIT(pm);
2031 	moea64_pinit(mmu, pm);
2032 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
2033 }
2034 
2035 /*
2036  * Set the physical protection on the specified range of this map as requested.
2037  */
2038 static void
2039 moea64_pvo_protect(mmu_t mmu,  pmap_t pm, struct pvo_entry *pvo, vm_prot_t prot)
2040 {
2041 	struct vm_page *pg;
2042 	vm_prot_t oldprot;
2043 	int32_t refchg;
2044 
2045 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2046 
2047 	/*
2048 	 * Change the protection of the page.
2049 	 */
2050 	oldprot = pvo->pvo_pte.prot;
2051 	pvo->pvo_pte.prot = prot;
2052 	pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
2053 
2054 	/*
2055 	 * If the PVO is in the page table, update mapping
2056 	 */
2057 	refchg = MOEA64_PTE_REPLACE(mmu, pvo, MOEA64_PTE_PROT_UPDATE);
2058 	if (refchg < 0)
2059 		refchg = (oldprot & VM_PROT_WRITE) ? LPTE_CHG : 0;
2060 
2061 	if (pm != kernel_pmap && pg != NULL && !(pg->aflags & PGA_EXECUTABLE) &&
2062 	    (pvo->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
2063 		if ((pg->oflags & VPO_UNMANAGED) == 0)
2064 			vm_page_aflag_set(pg, PGA_EXECUTABLE);
2065 		moea64_syncicache(mmu, pm, PVO_VADDR(pvo),
2066 		    pvo->pvo_pte.pa & LPTE_RPGN, PAGE_SIZE);
2067 	}
2068 
2069 	/*
2070 	 * Update vm about the REF/CHG bits if the page is managed and we have
2071 	 * removed write access.
2072 	 */
2073 	if (pg != NULL && (pvo->pvo_vaddr & PVO_MANAGED) &&
2074 	    (oldprot & VM_PROT_WRITE)) {
2075 		refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2076 		if (refchg & LPTE_CHG)
2077 			vm_page_dirty(pg);
2078 		if (refchg & LPTE_REF)
2079 			vm_page_aflag_set(pg, PGA_REFERENCED);
2080 	}
2081 }
2082 
2083 void
2084 moea64_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
2085     vm_prot_t prot)
2086 {
2087 	struct	pvo_entry *pvo, *tpvo, key;
2088 
2089 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm,
2090 	    sva, eva, prot);
2091 
2092 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
2093 	    ("moea64_protect: non current pmap"));
2094 
2095 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2096 		moea64_remove(mmu, pm, sva, eva);
2097 		return;
2098 	}
2099 
2100 	PMAP_LOCK(pm);
2101 	key.pvo_vaddr = sva;
2102 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2103 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2104 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2105 		moea64_pvo_protect(mmu, pm, pvo, prot);
2106 	}
2107 	PMAP_UNLOCK(pm);
2108 }
2109 
2110 /*
2111  * Map a list of wired pages into kernel virtual address space.  This is
2112  * intended for temporary mappings which do not need page modification or
2113  * references recorded.  Existing mappings in the region are overwritten.
2114  */
2115 void
2116 moea64_qenter(mmu_t mmu, vm_offset_t va, vm_page_t *m, int count)
2117 {
2118 	while (count-- > 0) {
2119 		moea64_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
2120 		va += PAGE_SIZE;
2121 		m++;
2122 	}
2123 }
2124 
2125 /*
2126  * Remove page mappings from kernel virtual address space.  Intended for
2127  * temporary mappings entered by moea64_qenter.
2128  */
2129 void
2130 moea64_qremove(mmu_t mmu, vm_offset_t va, int count)
2131 {
2132 	while (count-- > 0) {
2133 		moea64_kremove(mmu, va);
2134 		va += PAGE_SIZE;
2135 	}
2136 }
2137 
2138 void
2139 moea64_release_vsid(uint64_t vsid)
2140 {
2141 	int idx, mask;
2142 
2143 	mtx_lock(&moea64_slb_mutex);
2144 	idx = vsid & (NVSIDS-1);
2145 	mask = 1 << (idx % VSID_NBPW);
2146 	idx /= VSID_NBPW;
2147 	KASSERT(moea64_vsid_bitmap[idx] & mask,
2148 	    ("Freeing unallocated VSID %#jx", vsid));
2149 	moea64_vsid_bitmap[idx] &= ~mask;
2150 	mtx_unlock(&moea64_slb_mutex);
2151 }
2152 
2153 
2154 void
2155 moea64_release(mmu_t mmu, pmap_t pmap)
2156 {
2157 
2158 	/*
2159 	 * Free segment registers' VSIDs
2160 	 */
2161     #ifdef __powerpc64__
2162 	slb_free_tree(pmap);
2163 	slb_free_user_cache(pmap->pm_slb);
2164     #else
2165 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_release: pm_sr[0] = 0"));
2166 
2167 	moea64_release_vsid(VSID_TO_HASH(pmap->pm_sr[0]));
2168     #endif
2169 }
2170 
2171 /*
2172  * Remove all pages mapped by the specified pmap
2173  */
2174 void
2175 moea64_remove_pages(mmu_t mmu, pmap_t pm)
2176 {
2177 	struct pvo_entry *pvo, *tpvo;
2178 	struct pvo_tree tofree;
2179 
2180 	RB_INIT(&tofree);
2181 
2182 	PMAP_LOCK(pm);
2183 	RB_FOREACH_SAFE(pvo, pvo_tree, &pm->pmap_pvo, tpvo) {
2184 		if (pvo->pvo_vaddr & PVO_WIRED)
2185 			continue;
2186 
2187 		/*
2188 		 * For locking reasons, remove this from the page table and
2189 		 * pmap, but save delinking from the vm_page for a second
2190 		 * pass
2191 		 */
2192 		moea64_pvo_remove_from_pmap(mmu, pvo);
2193 		RB_INSERT(pvo_tree, &tofree, pvo);
2194 	}
2195 	PMAP_UNLOCK(pm);
2196 
2197 	RB_FOREACH_SAFE(pvo, pvo_tree, &tofree, tpvo) {
2198 		PV_LOCK(pvo->pvo_pte.pa & LPTE_RPGN);
2199 		moea64_pvo_remove_from_page(mmu, pvo);
2200 		PV_UNLOCK(pvo->pvo_pte.pa & LPTE_RPGN);
2201 		RB_REMOVE(pvo_tree, &tofree, pvo);
2202 		free_pvo_entry(pvo);
2203 	}
2204 }
2205 
2206 /*
2207  * Remove the given range of addresses from the specified map.
2208  */
2209 void
2210 moea64_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
2211 {
2212 	struct  pvo_entry *pvo, *tpvo, key;
2213 	struct pvo_tree tofree;
2214 
2215 	/*
2216 	 * Perform an unsynchronized read.  This is, however, safe.
2217 	 */
2218 	if (pm->pm_stats.resident_count == 0)
2219 		return;
2220 
2221 	key.pvo_vaddr = sva;
2222 
2223 	RB_INIT(&tofree);
2224 
2225 	PMAP_LOCK(pm);
2226 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2227 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2228 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2229 
2230 		/*
2231 		 * For locking reasons, remove this from the page table and
2232 		 * pmap, but save delinking from the vm_page for a second
2233 		 * pass
2234 		 */
2235 		moea64_pvo_remove_from_pmap(mmu, pvo);
2236 		RB_INSERT(pvo_tree, &tofree, pvo);
2237 	}
2238 	PMAP_UNLOCK(pm);
2239 
2240 	RB_FOREACH_SAFE(pvo, pvo_tree, &tofree, tpvo) {
2241 		PV_LOCK(pvo->pvo_pte.pa & LPTE_RPGN);
2242 		moea64_pvo_remove_from_page(mmu, pvo);
2243 		PV_UNLOCK(pvo->pvo_pte.pa & LPTE_RPGN);
2244 		RB_REMOVE(pvo_tree, &tofree, pvo);
2245 		free_pvo_entry(pvo);
2246 	}
2247 }
2248 
2249 /*
2250  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
2251  * will reflect changes in pte's back to the vm_page.
2252  */
2253 void
2254 moea64_remove_all(mmu_t mmu, vm_page_t m)
2255 {
2256 	struct	pvo_entry *pvo, *next_pvo;
2257 	struct	pvo_head freequeue;
2258 	int	wasdead;
2259 	pmap_t	pmap;
2260 
2261 	LIST_INIT(&freequeue);
2262 
2263 	PV_PAGE_LOCK(m);
2264 	LIST_FOREACH_SAFE(pvo, vm_page_to_pvoh(m), pvo_vlink, next_pvo) {
2265 		pmap = pvo->pvo_pmap;
2266 		PMAP_LOCK(pmap);
2267 		wasdead = (pvo->pvo_vaddr & PVO_DEAD);
2268 		if (!wasdead)
2269 			moea64_pvo_remove_from_pmap(mmu, pvo);
2270 		moea64_pvo_remove_from_page(mmu, pvo);
2271 		if (!wasdead)
2272 			LIST_INSERT_HEAD(&freequeue, pvo, pvo_vlink);
2273 		PMAP_UNLOCK(pmap);
2274 
2275 	}
2276 	KASSERT(!pmap_page_is_mapped(m), ("Page still has mappings"));
2277 	KASSERT(!(m->aflags & PGA_WRITEABLE), ("Page still writable"));
2278 	PV_PAGE_UNLOCK(m);
2279 
2280 	/* Clean up UMA allocations */
2281 	LIST_FOREACH_SAFE(pvo, &freequeue, pvo_vlink, next_pvo)
2282 		free_pvo_entry(pvo);
2283 }
2284 
2285 /*
2286  * Allocate a physical page of memory directly from the phys_avail map.
2287  * Can only be called from moea64_bootstrap before avail start and end are
2288  * calculated.
2289  */
2290 vm_offset_t
2291 moea64_bootstrap_alloc(vm_size_t size, u_int align)
2292 {
2293 	vm_offset_t	s, e;
2294 	int		i, j;
2295 
2296 	size = round_page(size);
2297 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2298 		if (align != 0)
2299 			s = roundup2(phys_avail[i], align);
2300 		else
2301 			s = phys_avail[i];
2302 		e = s + size;
2303 
2304 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2305 			continue;
2306 
2307 		if (s + size > platform_real_maxaddr())
2308 			continue;
2309 
2310 		if (s == phys_avail[i]) {
2311 			phys_avail[i] += size;
2312 		} else if (e == phys_avail[i + 1]) {
2313 			phys_avail[i + 1] -= size;
2314 		} else {
2315 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2316 				phys_avail[j] = phys_avail[j - 2];
2317 				phys_avail[j + 1] = phys_avail[j - 1];
2318 			}
2319 
2320 			phys_avail[i + 3] = phys_avail[i + 1];
2321 			phys_avail[i + 1] = s;
2322 			phys_avail[i + 2] = e;
2323 			phys_avail_count++;
2324 		}
2325 
2326 		return (s);
2327 	}
2328 	panic("moea64_bootstrap_alloc: could not allocate memory");
2329 }
2330 
2331 static int
2332 moea64_pvo_enter(mmu_t mmu, struct pvo_entry *pvo, struct pvo_head *pvo_head)
2333 {
2334 	int first, err;
2335 
2336 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2337 	KASSERT(moea64_pvo_find_va(pvo->pvo_pmap, PVO_VADDR(pvo)) == NULL,
2338 	    ("Existing mapping for VA %#jx", (uintmax_t)PVO_VADDR(pvo)));
2339 
2340 	moea64_pvo_enter_calls++;
2341 
2342 	/*
2343 	 * Add to pmap list
2344 	 */
2345 	RB_INSERT(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2346 
2347 	/*
2348 	 * Remember if the list was empty and therefore will be the first
2349 	 * item.
2350 	 */
2351 	if (pvo_head != NULL) {
2352 		if (LIST_FIRST(pvo_head) == NULL)
2353 			first = 1;
2354 		LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2355 	}
2356 
2357 	if (pvo->pvo_vaddr & PVO_WIRED)
2358 		pvo->pvo_pmap->pm_stats.wired_count++;
2359 	pvo->pvo_pmap->pm_stats.resident_count++;
2360 
2361 	/*
2362 	 * Insert it into the hardware page table
2363 	 */
2364 	err = MOEA64_PTE_INSERT(mmu, pvo);
2365 	if (err != 0) {
2366 		panic("moea64_pvo_enter: overflow");
2367 	}
2368 
2369 	moea64_pvo_entries++;
2370 
2371 	if (pvo->pvo_pmap == kernel_pmap)
2372 		isync();
2373 
2374 #ifdef __powerpc64__
2375 	/*
2376 	 * Make sure all our bootstrap mappings are in the SLB as soon
2377 	 * as virtual memory is switched on.
2378 	 */
2379 	if (!pmap_bootstrapped)
2380 		moea64_bootstrap_slb_prefault(PVO_VADDR(pvo),
2381 		    pvo->pvo_vaddr & PVO_LARGE);
2382 #endif
2383 
2384 	return (first ? ENOENT : 0);
2385 }
2386 
2387 static void
2388 moea64_pvo_remove_from_pmap(mmu_t mmu, struct pvo_entry *pvo)
2389 {
2390 	struct	vm_page *pg;
2391 	int32_t refchg;
2392 
2393 	KASSERT(pvo->pvo_pmap != NULL, ("Trying to remove PVO with no pmap"));
2394 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2395 	KASSERT(!(pvo->pvo_vaddr & PVO_DEAD), ("Trying to remove dead PVO"));
2396 
2397 	/*
2398 	 * If there is an active pte entry, we need to deactivate it
2399 	 */
2400 	refchg = MOEA64_PTE_UNSET(mmu, pvo);
2401 	if (refchg < 0) {
2402 		/*
2403 		 * If it was evicted from the page table, be pessimistic and
2404 		 * dirty the page.
2405 		 */
2406 		if (pvo->pvo_pte.prot & VM_PROT_WRITE)
2407 			refchg = LPTE_CHG;
2408 		else
2409 			refchg = 0;
2410 	}
2411 
2412 	/*
2413 	 * Update our statistics.
2414 	 */
2415 	pvo->pvo_pmap->pm_stats.resident_count--;
2416 	if (pvo->pvo_vaddr & PVO_WIRED)
2417 		pvo->pvo_pmap->pm_stats.wired_count--;
2418 
2419 	/*
2420 	 * Remove this PVO from the pmap list.
2421 	 */
2422 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2423 
2424 	/*
2425 	 * Mark this for the next sweep
2426 	 */
2427 	pvo->pvo_vaddr |= PVO_DEAD;
2428 
2429 	/* Send RC bits to VM */
2430 	if ((pvo->pvo_vaddr & PVO_MANAGED) &&
2431 	    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2432 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
2433 		if (pg != NULL) {
2434 			refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2435 			if (refchg & LPTE_CHG)
2436 				vm_page_dirty(pg);
2437 			if (refchg & LPTE_REF)
2438 				vm_page_aflag_set(pg, PGA_REFERENCED);
2439 		}
2440 	}
2441 }
2442 
2443 static void
2444 moea64_pvo_remove_from_page(mmu_t mmu, struct pvo_entry *pvo)
2445 {
2446 	struct	vm_page *pg;
2447 
2448 	KASSERT(pvo->pvo_vaddr & PVO_DEAD, ("Trying to delink live page"));
2449 
2450 	/* Use NULL pmaps as a sentinel for races in page deletion */
2451 	if (pvo->pvo_pmap == NULL)
2452 		return;
2453 	pvo->pvo_pmap = NULL;
2454 
2455 	/*
2456 	 * Update vm about page writeability/executability if managed
2457 	 */
2458 	PV_LOCKASSERT(pvo->pvo_pte.pa & LPTE_RPGN);
2459 	pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pa & LPTE_RPGN);
2460 
2461 	if ((pvo->pvo_vaddr & PVO_MANAGED) && pg != NULL) {
2462 		LIST_REMOVE(pvo, pvo_vlink);
2463 		if (LIST_EMPTY(vm_page_to_pvoh(pg)))
2464 			vm_page_aflag_clear(pg, PGA_WRITEABLE | PGA_EXECUTABLE);
2465 	}
2466 
2467 	moea64_pvo_entries--;
2468 	moea64_pvo_remove_calls++;
2469 }
2470 
2471 static struct pvo_entry *
2472 moea64_pvo_find_va(pmap_t pm, vm_offset_t va)
2473 {
2474 	struct pvo_entry key;
2475 
2476 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2477 
2478 	key.pvo_vaddr = va & ~ADDR_POFF;
2479 	return (RB_FIND(pvo_tree, &pm->pmap_pvo, &key));
2480 }
2481 
2482 static boolean_t
2483 moea64_query_bit(mmu_t mmu, vm_page_t m, uint64_t ptebit)
2484 {
2485 	struct	pvo_entry *pvo;
2486 	int64_t ret;
2487 	boolean_t rv;
2488 
2489 	/*
2490 	 * See if this bit is stored in the page already.
2491 	 */
2492 	if (m->md.mdpg_attrs & ptebit)
2493 		return (TRUE);
2494 
2495 	/*
2496 	 * Examine each PTE.  Sync so that any pending REF/CHG bits are
2497 	 * flushed to the PTEs.
2498 	 */
2499 	rv = FALSE;
2500 	powerpc_sync();
2501 	PV_PAGE_LOCK(m);
2502 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2503 		ret = 0;
2504 
2505 		/*
2506 		 * See if this pvo has a valid PTE.  if so, fetch the
2507 		 * REF/CHG bits from the valid PTE.  If the appropriate
2508 		 * ptebit is set, return success.
2509 		 */
2510 		PMAP_LOCK(pvo->pvo_pmap);
2511 		if (!(pvo->pvo_vaddr & PVO_DEAD))
2512 			ret = MOEA64_PTE_SYNCH(mmu, pvo);
2513 		PMAP_UNLOCK(pvo->pvo_pmap);
2514 
2515 		if (ret > 0) {
2516 			atomic_set_32(&m->md.mdpg_attrs,
2517 			    ret & (LPTE_CHG | LPTE_REF));
2518 			if (ret & ptebit) {
2519 				rv = TRUE;
2520 				break;
2521 			}
2522 		}
2523 	}
2524 	PV_PAGE_UNLOCK(m);
2525 
2526 	return (rv);
2527 }
2528 
2529 static u_int
2530 moea64_clear_bit(mmu_t mmu, vm_page_t m, u_int64_t ptebit)
2531 {
2532 	u_int	count;
2533 	struct	pvo_entry *pvo;
2534 	int64_t ret;
2535 
2536 	/*
2537 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2538 	 * we can reset the right ones).
2539 	 */
2540 	powerpc_sync();
2541 
2542 	/*
2543 	 * For each pvo entry, clear the pte's ptebit.
2544 	 */
2545 	count = 0;
2546 	PV_PAGE_LOCK(m);
2547 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2548 		ret = 0;
2549 
2550 		PMAP_LOCK(pvo->pvo_pmap);
2551 		if (!(pvo->pvo_vaddr & PVO_DEAD))
2552 			ret = MOEA64_PTE_CLEAR(mmu, pvo, ptebit);
2553 		PMAP_UNLOCK(pvo->pvo_pmap);
2554 
2555 		if (ret > 0 && (ret & ptebit))
2556 			count++;
2557 	}
2558 	atomic_clear_32(&m->md.mdpg_attrs, ptebit);
2559 	PV_PAGE_UNLOCK(m);
2560 
2561 	return (count);
2562 }
2563 
2564 boolean_t
2565 moea64_dev_direct_mapped(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2566 {
2567 	struct pvo_entry *pvo, key;
2568 	vm_offset_t ppa;
2569 	int error = 0;
2570 
2571 	PMAP_LOCK(kernel_pmap);
2572 	key.pvo_vaddr = ppa = pa & ~ADDR_POFF;
2573 	for (pvo = RB_FIND(pvo_tree, &kernel_pmap->pmap_pvo, &key);
2574 	    ppa < pa + size; ppa += PAGE_SIZE,
2575 	    pvo = RB_NEXT(pvo_tree, &kernel_pmap->pmap_pvo, pvo)) {
2576 		if (pvo == NULL || (pvo->pvo_pte.pa & LPTE_RPGN) != ppa) {
2577 			error = EFAULT;
2578 			break;
2579 		}
2580 	}
2581 	PMAP_UNLOCK(kernel_pmap);
2582 
2583 	return (error);
2584 }
2585 
2586 /*
2587  * Map a set of physical memory pages into the kernel virtual
2588  * address space. Return a pointer to where it is mapped. This
2589  * routine is intended to be used for mapping device memory,
2590  * NOT real memory.
2591  */
2592 void *
2593 moea64_mapdev_attr(mmu_t mmu, vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
2594 {
2595 	vm_offset_t va, tmpva, ppa, offset;
2596 
2597 	ppa = trunc_page(pa);
2598 	offset = pa & PAGE_MASK;
2599 	size = roundup2(offset + size, PAGE_SIZE);
2600 
2601 	va = kva_alloc(size);
2602 
2603 	if (!va)
2604 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
2605 
2606 	for (tmpva = va; size > 0;) {
2607 		moea64_kenter_attr(mmu, tmpva, ppa, ma);
2608 		size -= PAGE_SIZE;
2609 		tmpva += PAGE_SIZE;
2610 		ppa += PAGE_SIZE;
2611 	}
2612 
2613 	return ((void *)(va + offset));
2614 }
2615 
2616 void *
2617 moea64_mapdev(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2618 {
2619 
2620 	return moea64_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT);
2621 }
2622 
2623 void
2624 moea64_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2625 {
2626 	vm_offset_t base, offset;
2627 
2628 	base = trunc_page(va);
2629 	offset = va & PAGE_MASK;
2630 	size = roundup2(offset + size, PAGE_SIZE);
2631 
2632 	kva_free(base, size);
2633 }
2634 
2635 void
2636 moea64_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2637 {
2638 	struct pvo_entry *pvo;
2639 	vm_offset_t lim;
2640 	vm_paddr_t pa;
2641 	vm_size_t len;
2642 
2643 	PMAP_LOCK(pm);
2644 	while (sz > 0) {
2645 		lim = round_page(va);
2646 		len = MIN(lim - va, sz);
2647 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
2648 		if (pvo != NULL && !(pvo->pvo_pte.pa & LPTE_I)) {
2649 			pa = (pvo->pvo_pte.pa & LPTE_RPGN) | (va & ADDR_POFF);
2650 			moea64_syncicache(mmu, pm, va, pa, len);
2651 		}
2652 		va += len;
2653 		sz -= len;
2654 	}
2655 	PMAP_UNLOCK(pm);
2656 }
2657 
2658 void
2659 moea64_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va)
2660 {
2661 
2662 	*va = (void *)pa;
2663 }
2664 
2665 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
2666 
2667 void
2668 moea64_scan_init(mmu_t mmu)
2669 {
2670 	struct pvo_entry *pvo;
2671 	vm_offset_t va;
2672 	int i;
2673 
2674 	if (!do_minidump) {
2675 		/* Initialize phys. segments for dumpsys(). */
2676 		memset(&dump_map, 0, sizeof(dump_map));
2677 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
2678 		for (i = 0; i < pregions_sz; i++) {
2679 			dump_map[i].pa_start = pregions[i].mr_start;
2680 			dump_map[i].pa_size = pregions[i].mr_size;
2681 		}
2682 		return;
2683 	}
2684 
2685 	/* Virtual segments for minidumps: */
2686 	memset(&dump_map, 0, sizeof(dump_map));
2687 
2688 	/* 1st: kernel .data and .bss. */
2689 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
2690 	dump_map[0].pa_size = round_page((uintptr_t)_end) -
2691 	    dump_map[0].pa_start;
2692 
2693 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2694 	dump_map[1].pa_start = (vm_paddr_t)msgbufp->msg_ptr;
2695 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
2696 
2697 	/* 3rd: kernel VM. */
2698 	va = dump_map[1].pa_start + dump_map[1].pa_size;
2699 	/* Find start of next chunk (from va). */
2700 	while (va < virtual_end) {
2701 		/* Don't dump the buffer cache. */
2702 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
2703 			va = kmi.buffer_eva;
2704 			continue;
2705 		}
2706 		pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
2707 		if (pvo != NULL && !(pvo->pvo_vaddr & PVO_DEAD))
2708 			break;
2709 		va += PAGE_SIZE;
2710 	}
2711 	if (va < virtual_end) {
2712 		dump_map[2].pa_start = va;
2713 		va += PAGE_SIZE;
2714 		/* Find last page in chunk. */
2715 		while (va < virtual_end) {
2716 			/* Don't run into the buffer cache. */
2717 			if (va == kmi.buffer_sva)
2718 				break;
2719 			pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
2720 			if (pvo != NULL && !(pvo->pvo_vaddr & PVO_DEAD))
2721 				break;
2722 			va += PAGE_SIZE;
2723 		}
2724 		dump_map[2].pa_size = va - dump_map[2].pa_start;
2725 	}
2726 }
2727 
2728