xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision cab6a39d7b343596a5823e65c0f7b426551ec22d)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2008-2015 Nathan Whitehorn
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 /*
33  * Manages physical address maps.
34  *
35  * Since the information managed by this module is also stored by the
36  * logical address mapping module, this module may throw away valid virtual
37  * to physical mappings at almost any time.  However, invalidations of
38  * mappings must be done as requested.
39  *
40  * In order to cope with hardware architectures which make virtual to
41  * physical map invalidates expensive, this module may delay invalidate
42  * reduced protection operations until such time as they are actually
43  * necessary.  This module is given full information as to which processors
44  * are currently using which maps, and to when physical maps must be made
45  * correct.
46  */
47 
48 #include "opt_kstack_pages.h"
49 
50 #include <sys/param.h>
51 #include <sys/kernel.h>
52 #include <sys/conf.h>
53 #include <sys/queue.h>
54 #include <sys/cpuset.h>
55 #include <sys/kerneldump.h>
56 #include <sys/ktr.h>
57 #include <sys/lock.h>
58 #include <sys/msgbuf.h>
59 #include <sys/malloc.h>
60 #include <sys/mman.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 #include <sys/reboot.h>
70 
71 #include <sys/kdb.h>
72 
73 #include <dev/ofw/openfirm.h>
74 
75 #include <vm/vm.h>
76 #include <vm/pmap.h>
77 #include <vm/vm_param.h>
78 #include <vm/vm_kern.h>
79 #include <vm/vm_page.h>
80 #include <vm/vm_phys.h>
81 #include <vm/vm_map.h>
82 #include <vm/vm_object.h>
83 #include <vm/vm_extern.h>
84 #include <vm/vm_pageout.h>
85 #include <vm/vm_dumpset.h>
86 #include <vm/vm_reserv.h>
87 #include <vm/uma.h>
88 
89 #include <machine/_inttypes.h>
90 #include <machine/cpu.h>
91 #include <machine/ifunc.h>
92 #include <machine/platform.h>
93 #include <machine/frame.h>
94 #include <machine/md_var.h>
95 #include <machine/psl.h>
96 #include <machine/bat.h>
97 #include <machine/hid.h>
98 #include <machine/pte.h>
99 #include <machine/sr.h>
100 #include <machine/trap.h>
101 #include <machine/mmuvar.h>
102 
103 #include "mmu_oea64.h"
104 
105 void moea64_release_vsid(uint64_t vsid);
106 uintptr_t moea64_get_unique_vsid(void);
107 
108 #define DISABLE_TRANS(msr)	msr = mfmsr(); mtmsr(msr & ~PSL_DR)
109 #define ENABLE_TRANS(msr)	mtmsr(msr)
110 
111 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
112 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
113 #define	VSID_HASH_MASK		0x0000007fffffffffULL
114 
115 /*
116  * Locking semantics:
117  *
118  * There are two locks of interest: the page locks and the pmap locks, which
119  * protect their individual PVO lists and are locked in that order. The contents
120  * of all PVO entries are protected by the locks of their respective pmaps.
121  * The pmap of any PVO is guaranteed not to change so long as the PVO is linked
122  * into any list.
123  *
124  */
125 
126 #define PV_LOCK_COUNT	PA_LOCK_COUNT
127 static struct mtx_padalign pv_lock[PV_LOCK_COUNT];
128 
129 /*
130  * Cheap NUMA-izing of the pv locks, to reduce contention across domains.
131  * NUMA domains on POWER9 appear to be indexed as sparse memory spaces, with the
132  * index at (N << 45).
133  */
134 #ifdef __powerpc64__
135 #define PV_LOCK_IDX(pa)	((pa_index(pa) * (((pa) >> 45) + 1)) % PV_LOCK_COUNT)
136 #else
137 #define PV_LOCK_IDX(pa)	(pa_index(pa) % PV_LOCK_COUNT)
138 #endif
139 #define PV_LOCKPTR(pa)	((struct mtx *)(&pv_lock[PV_LOCK_IDX(pa)]))
140 #define PV_LOCK(pa)		mtx_lock(PV_LOCKPTR(pa))
141 #define PV_UNLOCK(pa)		mtx_unlock(PV_LOCKPTR(pa))
142 #define PV_LOCKASSERT(pa) 	mtx_assert(PV_LOCKPTR(pa), MA_OWNED)
143 #define PV_PAGE_LOCK(m)		PV_LOCK(VM_PAGE_TO_PHYS(m))
144 #define PV_PAGE_UNLOCK(m)	PV_UNLOCK(VM_PAGE_TO_PHYS(m))
145 #define PV_PAGE_LOCKASSERT(m)	PV_LOCKASSERT(VM_PAGE_TO_PHYS(m))
146 
147 /* Superpage PV lock */
148 
149 #define	PV_LOCK_SIZE		(1<<PDRSHIFT)
150 
151 static __always_inline void
152 moea64_sp_pv_lock(vm_paddr_t pa)
153 {
154 	vm_paddr_t pa_end;
155 
156 	/* Note: breaking when pa_end is reached to avoid overflows */
157 	pa_end = pa + (HPT_SP_SIZE - PV_LOCK_SIZE);
158 	for (;;) {
159 		mtx_lock_flags(PV_LOCKPTR(pa), MTX_DUPOK);
160 		if (pa == pa_end)
161 			break;
162 		pa += PV_LOCK_SIZE;
163 	}
164 }
165 
166 static __always_inline void
167 moea64_sp_pv_unlock(vm_paddr_t pa)
168 {
169 	vm_paddr_t pa_end;
170 
171 	/* Note: breaking when pa_end is reached to avoid overflows */
172 	pa_end = pa;
173 	pa += HPT_SP_SIZE - PV_LOCK_SIZE;
174 	for (;;) {
175 		mtx_unlock_flags(PV_LOCKPTR(pa), MTX_DUPOK);
176 		if (pa == pa_end)
177 			break;
178 		pa -= PV_LOCK_SIZE;
179 	}
180 }
181 
182 #define	SP_PV_LOCK_ALIGNED(pa)		moea64_sp_pv_lock(pa)
183 #define	SP_PV_UNLOCK_ALIGNED(pa)	moea64_sp_pv_unlock(pa)
184 #define	SP_PV_LOCK(pa)			moea64_sp_pv_lock((pa) & ~HPT_SP_MASK)
185 #define	SP_PV_UNLOCK(pa)		moea64_sp_pv_unlock((pa) & ~HPT_SP_MASK)
186 #define	SP_PV_PAGE_LOCK(m)		SP_PV_LOCK(VM_PAGE_TO_PHYS(m))
187 #define	SP_PV_PAGE_UNLOCK(m)		SP_PV_UNLOCK(VM_PAGE_TO_PHYS(m))
188 
189 struct ofw_map {
190 	cell_t	om_va;
191 	cell_t	om_len;
192 	uint64_t om_pa;
193 	cell_t	om_mode;
194 };
195 
196 extern unsigned char _etext[];
197 extern unsigned char _end[];
198 
199 extern void *slbtrap, *slbtrapend;
200 
201 /*
202  * Map of physical memory regions.
203  */
204 static struct	mem_region *regions;
205 static struct	mem_region *pregions;
206 static struct	numa_mem_region *numa_pregions;
207 static u_int	phys_avail_count;
208 static int	regions_sz, pregions_sz, numapregions_sz;
209 
210 extern void bs_remap_earlyboot(void);
211 
212 /*
213  * Lock for the SLB tables.
214  */
215 struct mtx	moea64_slb_mutex;
216 
217 /*
218  * PTEG data.
219  */
220 u_long		moea64_pteg_count;
221 u_long		moea64_pteg_mask;
222 
223 /*
224  * PVO data.
225  */
226 
227 uma_zone_t	moea64_pvo_zone; /* zone for pvo entries */
228 
229 static struct	pvo_entry *moea64_bpvo_pool;
230 static int	moea64_bpvo_pool_index = 0;
231 static int	moea64_bpvo_pool_size = 0;
232 SYSCTL_INT(_machdep, OID_AUTO, moea64_allocated_bpvo_entries, CTLFLAG_RD,
233     &moea64_bpvo_pool_index, 0, "");
234 
235 #define	BPVO_POOL_SIZE	327680 /* Sensible historical default value */
236 #define	BPVO_POOL_EXPANSION_FACTOR	3
237 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
238 #ifdef __powerpc64__
239 #define	NVSIDS		(NPMAPS * 16)
240 #define VSID_HASHMASK	0xffffffffUL
241 #else
242 #define NVSIDS		NPMAPS
243 #define VSID_HASHMASK	0xfffffUL
244 #endif
245 static u_int	moea64_vsid_bitmap[NVSIDS / VSID_NBPW];
246 
247 static boolean_t moea64_initialized = FALSE;
248 
249 #ifdef MOEA64_STATS
250 /*
251  * Statistics.
252  */
253 u_int	moea64_pte_valid = 0;
254 u_int	moea64_pte_overflow = 0;
255 u_int	moea64_pvo_entries = 0;
256 u_int	moea64_pvo_enter_calls = 0;
257 u_int	moea64_pvo_remove_calls = 0;
258 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_valid, CTLFLAG_RD,
259     &moea64_pte_valid, 0, "");
260 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_overflow, CTLFLAG_RD,
261     &moea64_pte_overflow, 0, "");
262 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_entries, CTLFLAG_RD,
263     &moea64_pvo_entries, 0, "");
264 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_enter_calls, CTLFLAG_RD,
265     &moea64_pvo_enter_calls, 0, "");
266 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_remove_calls, CTLFLAG_RD,
267     &moea64_pvo_remove_calls, 0, "");
268 #endif
269 
270 vm_offset_t	moea64_scratchpage_va[2];
271 struct pvo_entry *moea64_scratchpage_pvo[2];
272 struct	mtx	moea64_scratchpage_mtx;
273 
274 uint64_t 	moea64_large_page_mask = 0;
275 uint64_t	moea64_large_page_size = 0;
276 int		moea64_large_page_shift = 0;
277 bool		moea64_has_lp_4k_16m = false;
278 
279 /*
280  * PVO calls.
281  */
282 static int	moea64_pvo_enter(struct pvo_entry *pvo,
283 		    struct pvo_head *pvo_head, struct pvo_entry **oldpvo);
284 static void	moea64_pvo_remove_from_pmap(struct pvo_entry *pvo);
285 static void	moea64_pvo_remove_from_page(struct pvo_entry *pvo);
286 static void	moea64_pvo_remove_from_page_locked(
287 		    struct pvo_entry *pvo, vm_page_t m);
288 static struct	pvo_entry *moea64_pvo_find_va(pmap_t, vm_offset_t);
289 
290 /*
291  * Utility routines.
292  */
293 static boolean_t	moea64_query_bit(vm_page_t, uint64_t);
294 static u_int		moea64_clear_bit(vm_page_t, uint64_t);
295 static void		moea64_kremove(vm_offset_t);
296 static void		moea64_syncicache(pmap_t pmap, vm_offset_t va,
297 			    vm_paddr_t pa, vm_size_t sz);
298 static void		moea64_pmap_init_qpages(void);
299 static void		moea64_remove_locked(pmap_t, vm_offset_t,
300 			    vm_offset_t, struct pvo_dlist *);
301 
302 /*
303  * Superpages data and routines.
304  */
305 
306 /*
307  * PVO flags (in vaddr) that must match for promotion to succeed.
308  * Note that protection bits are checked separately, as they reside in
309  * another field.
310  */
311 #define	PVO_FLAGS_PROMOTE	(PVO_WIRED | PVO_MANAGED | PVO_PTEGIDX_VALID)
312 
313 #define	PVO_IS_SP(pvo)		(((pvo)->pvo_vaddr & PVO_LARGE) && \
314 				 (pvo)->pvo_pmap != kernel_pmap)
315 
316 /* Get physical address from PVO. */
317 #define	PVO_PADDR(pvo)		moea64_pvo_paddr(pvo)
318 
319 /* MD page flag indicating that the page is a superpage. */
320 #define	MDPG_ATTR_SP		0x40000000
321 
322 SYSCTL_DECL(_vm_pmap);
323 
324 static SYSCTL_NODE(_vm_pmap, OID_AUTO, sp, CTLFLAG_RD, 0,
325     "SP page mapping counters");
326 
327 static u_long sp_demotions;
328 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, demotions, CTLFLAG_RD,
329     &sp_demotions, 0, "SP page demotions");
330 
331 static u_long sp_mappings;
332 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, mappings, CTLFLAG_RD,
333     &sp_mappings, 0, "SP page mappings");
334 
335 static u_long sp_p_failures;
336 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_failures, CTLFLAG_RD,
337     &sp_p_failures, 0, "SP page promotion failures");
338 
339 static u_long sp_p_fail_pa;
340 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_fail_pa, CTLFLAG_RD,
341     &sp_p_fail_pa, 0, "SP page promotion failure: PAs don't match");
342 
343 static u_long sp_p_fail_flags;
344 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_fail_flags, CTLFLAG_RD,
345     &sp_p_fail_flags, 0, "SP page promotion failure: page flags don't match");
346 
347 static u_long sp_p_fail_prot;
348 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_fail_prot, CTLFLAG_RD,
349     &sp_p_fail_prot, 0,
350     "SP page promotion failure: page protections don't match");
351 
352 static u_long sp_p_fail_wimg;
353 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_fail_wimg, CTLFLAG_RD,
354     &sp_p_fail_wimg, 0, "SP page promotion failure: WIMG bits don't match");
355 
356 static u_long sp_promotions;
357 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, promotions, CTLFLAG_RD,
358     &sp_promotions, 0, "SP page promotions");
359 
360 static bool moea64_ps_enabled(pmap_t);
361 static void moea64_align_superpage(vm_object_t, vm_ooffset_t,
362     vm_offset_t *, vm_size_t);
363 
364 static int moea64_sp_enter(pmap_t pmap, vm_offset_t va,
365     vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind);
366 static struct pvo_entry *moea64_sp_remove(struct pvo_entry *sp,
367     struct pvo_dlist *tofree);
368 
369 static void moea64_sp_promote(pmap_t pmap, vm_offset_t va, vm_page_t m);
370 static void moea64_sp_demote_aligned(struct pvo_entry *sp);
371 static void moea64_sp_demote(struct pvo_entry *pvo);
372 
373 static struct pvo_entry *moea64_sp_unwire(struct pvo_entry *sp);
374 static struct pvo_entry *moea64_sp_protect(struct pvo_entry *sp,
375     vm_prot_t prot);
376 
377 static int64_t moea64_sp_query(struct pvo_entry *pvo, uint64_t ptebit);
378 static int64_t moea64_sp_clear(struct pvo_entry *pvo, vm_page_t m,
379     uint64_t ptebit);
380 
381 static __inline bool moea64_sp_pvo_in_range(struct pvo_entry *pvo,
382     vm_offset_t sva, vm_offset_t eva);
383 
384 /*
385  * Kernel MMU interface
386  */
387 void moea64_clear_modify(vm_page_t);
388 void moea64_copy_page(vm_page_t, vm_page_t);
389 void moea64_copy_page_dmap(vm_page_t, vm_page_t);
390 void moea64_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
391     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
392 void moea64_copy_pages_dmap(vm_page_t *ma, vm_offset_t a_offset,
393     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
394 int moea64_enter(pmap_t, vm_offset_t, vm_page_t, vm_prot_t,
395     u_int flags, int8_t psind);
396 void moea64_enter_object(pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
397     vm_prot_t);
398 void moea64_enter_quick(pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
399 vm_paddr_t moea64_extract(pmap_t, vm_offset_t);
400 vm_page_t moea64_extract_and_hold(pmap_t, vm_offset_t, vm_prot_t);
401 void moea64_init(void);
402 boolean_t moea64_is_modified(vm_page_t);
403 boolean_t moea64_is_prefaultable(pmap_t, vm_offset_t);
404 boolean_t moea64_is_referenced(vm_page_t);
405 int moea64_ts_referenced(vm_page_t);
406 vm_offset_t moea64_map(vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
407 boolean_t moea64_page_exists_quick(pmap_t, vm_page_t);
408 void moea64_page_init(vm_page_t);
409 int moea64_page_wired_mappings(vm_page_t);
410 int moea64_pinit(pmap_t);
411 void moea64_pinit0(pmap_t);
412 void moea64_protect(pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
413 void moea64_qenter(vm_offset_t, vm_page_t *, int);
414 void moea64_qremove(vm_offset_t, int);
415 void moea64_release(pmap_t);
416 void moea64_remove(pmap_t, vm_offset_t, vm_offset_t);
417 void moea64_remove_pages(pmap_t);
418 void moea64_remove_all(vm_page_t);
419 void moea64_remove_write(vm_page_t);
420 void moea64_unwire(pmap_t, vm_offset_t, vm_offset_t);
421 void moea64_zero_page(vm_page_t);
422 void moea64_zero_page_dmap(vm_page_t);
423 void moea64_zero_page_area(vm_page_t, int, int);
424 void moea64_activate(struct thread *);
425 void moea64_deactivate(struct thread *);
426 void *moea64_mapdev(vm_paddr_t, vm_size_t);
427 void *moea64_mapdev_attr(vm_paddr_t, vm_size_t, vm_memattr_t);
428 void moea64_unmapdev(vm_offset_t, vm_size_t);
429 vm_paddr_t moea64_kextract(vm_offset_t);
430 void moea64_page_set_memattr(vm_page_t m, vm_memattr_t ma);
431 void moea64_kenter_attr(vm_offset_t, vm_paddr_t, vm_memattr_t ma);
432 void moea64_kenter(vm_offset_t, vm_paddr_t);
433 boolean_t moea64_dev_direct_mapped(vm_paddr_t, vm_size_t);
434 static void moea64_sync_icache(pmap_t, vm_offset_t, vm_size_t);
435 void moea64_dumpsys_map(vm_paddr_t pa, size_t sz,
436     void **va);
437 void moea64_scan_init(void);
438 vm_offset_t moea64_quick_enter_page(vm_page_t m);
439 vm_offset_t moea64_quick_enter_page_dmap(vm_page_t m);
440 void moea64_quick_remove_page(vm_offset_t addr);
441 boolean_t moea64_page_is_mapped(vm_page_t m);
442 static int moea64_map_user_ptr(pmap_t pm,
443     volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen);
444 static int moea64_decode_kernel_ptr(vm_offset_t addr,
445     int *is_user, vm_offset_t *decoded_addr);
446 static size_t moea64_scan_pmap(void);
447 static void *moea64_dump_pmap_init(unsigned blkpgs);
448 #ifdef __powerpc64__
449 static void moea64_page_array_startup(long);
450 #endif
451 static int moea64_mincore(pmap_t, vm_offset_t, vm_paddr_t *);
452 
453 static struct pmap_funcs moea64_methods = {
454 	.clear_modify = moea64_clear_modify,
455 	.copy_page = moea64_copy_page,
456 	.copy_pages = moea64_copy_pages,
457 	.enter = moea64_enter,
458 	.enter_object = moea64_enter_object,
459 	.enter_quick = moea64_enter_quick,
460 	.extract = moea64_extract,
461 	.extract_and_hold = moea64_extract_and_hold,
462 	.init = moea64_init,
463 	.is_modified = moea64_is_modified,
464 	.is_prefaultable = moea64_is_prefaultable,
465 	.is_referenced = moea64_is_referenced,
466 	.ts_referenced = moea64_ts_referenced,
467 	.map =      		moea64_map,
468 	.mincore = moea64_mincore,
469 	.page_exists_quick = moea64_page_exists_quick,
470 	.page_init = moea64_page_init,
471 	.page_wired_mappings = moea64_page_wired_mappings,
472 	.pinit = moea64_pinit,
473 	.pinit0 = moea64_pinit0,
474 	.protect = moea64_protect,
475 	.qenter = moea64_qenter,
476 	.qremove = moea64_qremove,
477 	.release = moea64_release,
478 	.remove = moea64_remove,
479 	.remove_pages = moea64_remove_pages,
480 	.remove_all =       	moea64_remove_all,
481 	.remove_write = moea64_remove_write,
482 	.sync_icache = moea64_sync_icache,
483 	.unwire = moea64_unwire,
484 	.zero_page =        	moea64_zero_page,
485 	.zero_page_area = moea64_zero_page_area,
486 	.activate = moea64_activate,
487 	.deactivate =       	moea64_deactivate,
488 	.page_set_memattr = moea64_page_set_memattr,
489 	.quick_enter_page =  moea64_quick_enter_page,
490 	.quick_remove_page =  moea64_quick_remove_page,
491 	.page_is_mapped = moea64_page_is_mapped,
492 #ifdef __powerpc64__
493 	.page_array_startup = moea64_page_array_startup,
494 #endif
495 	.ps_enabled = moea64_ps_enabled,
496 	.align_superpage = moea64_align_superpage,
497 
498 	/* Internal interfaces */
499 	.mapdev = moea64_mapdev,
500 	.mapdev_attr = moea64_mapdev_attr,
501 	.unmapdev = moea64_unmapdev,
502 	.kextract = moea64_kextract,
503 	.kenter = moea64_kenter,
504 	.kenter_attr = moea64_kenter_attr,
505 	.dev_direct_mapped = moea64_dev_direct_mapped,
506 	.dumpsys_pa_init = moea64_scan_init,
507 	.dumpsys_scan_pmap = moea64_scan_pmap,
508 	.dumpsys_dump_pmap_init =    moea64_dump_pmap_init,
509 	.dumpsys_map_chunk = moea64_dumpsys_map,
510 	.map_user_ptr = moea64_map_user_ptr,
511 	.decode_kernel_ptr =  moea64_decode_kernel_ptr,
512 };
513 
514 MMU_DEF(oea64_mmu, "mmu_oea64_base", moea64_methods);
515 
516 /*
517  * Get physical address from PVO.
518  *
519  * For superpages, the lower bits are not stored on pvo_pte.pa and must be
520  * obtained from VA.
521  */
522 static __always_inline vm_paddr_t
523 moea64_pvo_paddr(struct pvo_entry *pvo)
524 {
525 	vm_paddr_t pa;
526 
527 	pa = (pvo)->pvo_pte.pa & LPTE_RPGN;
528 
529 	if (PVO_IS_SP(pvo)) {
530 		pa &= ~HPT_SP_MASK; /* This is needed to clear LPTE_LP bits. */
531 		pa |= PVO_VADDR(pvo) & HPT_SP_MASK;
532 	}
533 	return (pa);
534 }
535 
536 static struct pvo_head *
537 vm_page_to_pvoh(vm_page_t m)
538 {
539 
540 	mtx_assert(PV_LOCKPTR(VM_PAGE_TO_PHYS(m)), MA_OWNED);
541 	return (&m->md.mdpg_pvoh);
542 }
543 
544 static struct pvo_entry *
545 alloc_pvo_entry(int bootstrap)
546 {
547 	struct pvo_entry *pvo;
548 
549 	if (!moea64_initialized || bootstrap) {
550 		if (moea64_bpvo_pool_index >= moea64_bpvo_pool_size) {
551 			panic("%s: bpvo pool exhausted, index=%d, size=%d, bytes=%zd."
552 			    "Try setting machdep.moea64_bpvo_pool_size tunable",
553 			    __func__, moea64_bpvo_pool_index,
554 			    moea64_bpvo_pool_size,
555 			    moea64_bpvo_pool_size * sizeof(struct pvo_entry));
556 		}
557 		pvo = &moea64_bpvo_pool[
558 		    atomic_fetchadd_int(&moea64_bpvo_pool_index, 1)];
559 		bzero(pvo, sizeof(*pvo));
560 		pvo->pvo_vaddr = PVO_BOOTSTRAP;
561 	} else
562 		pvo = uma_zalloc(moea64_pvo_zone, M_NOWAIT | M_ZERO);
563 
564 	return (pvo);
565 }
566 
567 static void
568 init_pvo_entry(struct pvo_entry *pvo, pmap_t pmap, vm_offset_t va)
569 {
570 	uint64_t vsid;
571 	uint64_t hash;
572 	int shift;
573 
574 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
575 
576 	pvo->pvo_pmap = pmap;
577 	va &= ~ADDR_POFF;
578 	pvo->pvo_vaddr |= va;
579 	vsid = va_to_vsid(pmap, va);
580 	pvo->pvo_vpn = (uint64_t)((va & ADDR_PIDX) >> ADDR_PIDX_SHFT)
581 	    | (vsid << 16);
582 
583 	if (pmap == kernel_pmap && (pvo->pvo_vaddr & PVO_LARGE) != 0)
584 		shift = moea64_large_page_shift;
585 	else
586 		shift = ADDR_PIDX_SHFT;
587 	hash = (vsid & VSID_HASH_MASK) ^ (((uint64_t)va & ADDR_PIDX) >> shift);
588 	pvo->pvo_pte.slot = (hash & moea64_pteg_mask) << 3;
589 }
590 
591 static void
592 free_pvo_entry(struct pvo_entry *pvo)
593 {
594 
595 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
596 		uma_zfree(moea64_pvo_zone, pvo);
597 }
598 
599 void
600 moea64_pte_from_pvo(const struct pvo_entry *pvo, struct lpte *lpte)
601 {
602 
603 	lpte->pte_hi = moea64_pte_vpn_from_pvo_vpn(pvo);
604 	lpte->pte_hi |= LPTE_VALID;
605 
606 	if (pvo->pvo_vaddr & PVO_LARGE)
607 		lpte->pte_hi |= LPTE_BIG;
608 	if (pvo->pvo_vaddr & PVO_WIRED)
609 		lpte->pte_hi |= LPTE_WIRED;
610 	if (pvo->pvo_vaddr & PVO_HID)
611 		lpte->pte_hi |= LPTE_HID;
612 
613 	lpte->pte_lo = pvo->pvo_pte.pa; /* Includes WIMG bits */
614 	if (pvo->pvo_pte.prot & VM_PROT_WRITE)
615 		lpte->pte_lo |= LPTE_BW;
616 	else
617 		lpte->pte_lo |= LPTE_BR;
618 
619 	if (!(pvo->pvo_pte.prot & VM_PROT_EXECUTE))
620 		lpte->pte_lo |= LPTE_NOEXEC;
621 }
622 
623 static __inline uint64_t
624 moea64_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
625 {
626 	uint64_t pte_lo;
627 	int i;
628 
629 	if (ma != VM_MEMATTR_DEFAULT) {
630 		switch (ma) {
631 		case VM_MEMATTR_UNCACHEABLE:
632 			return (LPTE_I | LPTE_G);
633 		case VM_MEMATTR_CACHEABLE:
634 			return (LPTE_M);
635 		case VM_MEMATTR_WRITE_COMBINING:
636 		case VM_MEMATTR_WRITE_BACK:
637 		case VM_MEMATTR_PREFETCHABLE:
638 			return (LPTE_I);
639 		case VM_MEMATTR_WRITE_THROUGH:
640 			return (LPTE_W | LPTE_M);
641 		}
642 	}
643 
644 	/*
645 	 * Assume the page is cache inhibited and access is guarded unless
646 	 * it's in our available memory array.
647 	 */
648 	pte_lo = LPTE_I | LPTE_G;
649 	for (i = 0; i < pregions_sz; i++) {
650 		if ((pa >= pregions[i].mr_start) &&
651 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
652 			pte_lo &= ~(LPTE_I | LPTE_G);
653 			pte_lo |= LPTE_M;
654 			break;
655 		}
656 	}
657 
658 	return pte_lo;
659 }
660 
661 /*
662  * Quick sort callout for comparing memory regions.
663  */
664 static int	om_cmp(const void *a, const void *b);
665 
666 static int
667 om_cmp(const void *a, const void *b)
668 {
669 	const struct	ofw_map *mapa;
670 	const struct	ofw_map *mapb;
671 
672 	mapa = a;
673 	mapb = b;
674 	if (mapa->om_pa < mapb->om_pa)
675 		return (-1);
676 	else if (mapa->om_pa > mapb->om_pa)
677 		return (1);
678 	else
679 		return (0);
680 }
681 
682 static void
683 moea64_add_ofw_mappings(phandle_t mmu, size_t sz)
684 {
685 	struct ofw_map	translations[sz/(4*sizeof(cell_t))]; /*>= 4 cells per */
686 	pcell_t		acells, trans_cells[sz/sizeof(cell_t)];
687 	struct pvo_entry *pvo;
688 	register_t	msr;
689 	vm_offset_t	off;
690 	vm_paddr_t	pa_base;
691 	int		i, j;
692 
693 	bzero(translations, sz);
694 	OF_getencprop(OF_finddevice("/"), "#address-cells", &acells,
695 	    sizeof(acells));
696 	if (OF_getencprop(mmu, "translations", trans_cells, sz) == -1)
697 		panic("moea64_bootstrap: can't get ofw translations");
698 
699 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
700 	sz /= sizeof(cell_t);
701 	for (i = 0, j = 0; i < sz; j++) {
702 		translations[j].om_va = trans_cells[i++];
703 		translations[j].om_len = trans_cells[i++];
704 		translations[j].om_pa = trans_cells[i++];
705 		if (acells == 2) {
706 			translations[j].om_pa <<= 32;
707 			translations[j].om_pa |= trans_cells[i++];
708 		}
709 		translations[j].om_mode = trans_cells[i++];
710 	}
711 	KASSERT(i == sz, ("Translations map has incorrect cell count (%d/%zd)",
712 	    i, sz));
713 
714 	sz = j;
715 	qsort(translations, sz, sizeof (*translations), om_cmp);
716 
717 	for (i = 0; i < sz; i++) {
718 		pa_base = translations[i].om_pa;
719 	      #ifndef __powerpc64__
720 		if ((translations[i].om_pa >> 32) != 0)
721 			panic("OFW translations above 32-bit boundary!");
722 	      #endif
723 
724 		if (pa_base % PAGE_SIZE)
725 			panic("OFW translation not page-aligned (phys)!");
726 		if (translations[i].om_va % PAGE_SIZE)
727 			panic("OFW translation not page-aligned (virt)!");
728 
729 		CTR3(KTR_PMAP, "translation: pa=%#zx va=%#x len=%#x",
730 		    pa_base, translations[i].om_va, translations[i].om_len);
731 
732 		/* Now enter the pages for this mapping */
733 
734 		DISABLE_TRANS(msr);
735 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
736 			/* If this address is direct-mapped, skip remapping */
737 			if (hw_direct_map &&
738 			    translations[i].om_va == PHYS_TO_DMAP(pa_base) &&
739 			    moea64_calc_wimg(pa_base + off, VM_MEMATTR_DEFAULT)
740  			    == LPTE_M)
741 				continue;
742 
743 			PMAP_LOCK(kernel_pmap);
744 			pvo = moea64_pvo_find_va(kernel_pmap,
745 			    translations[i].om_va + off);
746 			PMAP_UNLOCK(kernel_pmap);
747 			if (pvo != NULL)
748 				continue;
749 
750 			moea64_kenter(translations[i].om_va + off,
751 			    pa_base + off);
752 		}
753 		ENABLE_TRANS(msr);
754 	}
755 }
756 
757 #ifdef __powerpc64__
758 static void
759 moea64_probe_large_page(void)
760 {
761 	uint16_t pvr = mfpvr() >> 16;
762 
763 	switch (pvr) {
764 	case IBM970:
765 	case IBM970FX:
766 	case IBM970MP:
767 		powerpc_sync(); isync();
768 		mtspr(SPR_HID4, mfspr(SPR_HID4) & ~HID4_970_DISABLE_LG_PG);
769 		powerpc_sync(); isync();
770 
771 		/* FALLTHROUGH */
772 	default:
773 		if (moea64_large_page_size == 0) {
774 			moea64_large_page_size = 0x1000000; /* 16 MB */
775 			moea64_large_page_shift = 24;
776 		}
777 	}
778 
779 	moea64_large_page_mask = moea64_large_page_size - 1;
780 }
781 
782 static void
783 moea64_bootstrap_slb_prefault(vm_offset_t va, int large)
784 {
785 	struct slb *cache;
786 	struct slb entry;
787 	uint64_t esid, slbe;
788 	uint64_t i;
789 
790 	cache = PCPU_GET(aim.slb);
791 	esid = va >> ADDR_SR_SHFT;
792 	slbe = (esid << SLBE_ESID_SHIFT) | SLBE_VALID;
793 
794 	for (i = 0; i < 64; i++) {
795 		if (cache[i].slbe == (slbe | i))
796 			return;
797 	}
798 
799 	entry.slbe = slbe;
800 	entry.slbv = KERNEL_VSID(esid) << SLBV_VSID_SHIFT;
801 	if (large)
802 		entry.slbv |= SLBV_L;
803 
804 	slb_insert_kernel(entry.slbe, entry.slbv);
805 }
806 #endif
807 
808 static int
809 moea64_kenter_large(vm_offset_t va, vm_paddr_t pa, uint64_t attr, int bootstrap)
810 {
811 	struct pvo_entry *pvo;
812 	uint64_t pte_lo;
813 	int error;
814 
815 	pte_lo = LPTE_M;
816 	pte_lo |= attr;
817 
818 	pvo = alloc_pvo_entry(bootstrap);
819 	pvo->pvo_vaddr |= PVO_WIRED | PVO_LARGE;
820 	init_pvo_entry(pvo, kernel_pmap, va);
821 
822 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE |
823 	    VM_PROT_EXECUTE;
824 	pvo->pvo_pte.pa = pa | pte_lo;
825 	error = moea64_pvo_enter(pvo, NULL, NULL);
826 	if (error != 0)
827 		panic("Error %d inserting large page\n", error);
828 	return (0);
829 }
830 
831 static void
832 moea64_setup_direct_map(vm_offset_t kernelstart,
833     vm_offset_t kernelend)
834 {
835 	register_t msr;
836 	vm_paddr_t pa, pkernelstart, pkernelend;
837 	vm_offset_t size, off;
838 	uint64_t pte_lo;
839 	int i;
840 
841 	if (moea64_large_page_size == 0)
842 		hw_direct_map = 0;
843 
844 	DISABLE_TRANS(msr);
845 	if (hw_direct_map) {
846 		PMAP_LOCK(kernel_pmap);
847 		for (i = 0; i < pregions_sz; i++) {
848 		  for (pa = pregions[i].mr_start; pa < pregions[i].mr_start +
849 		     pregions[i].mr_size; pa += moea64_large_page_size) {
850 			pte_lo = LPTE_M;
851 			if (pa & moea64_large_page_mask) {
852 				pa &= moea64_large_page_mask;
853 				pte_lo |= LPTE_G;
854 			}
855 			if (pa + moea64_large_page_size >
856 			    pregions[i].mr_start + pregions[i].mr_size)
857 				pte_lo |= LPTE_G;
858 
859 			moea64_kenter_large(PHYS_TO_DMAP(pa), pa, pte_lo, 1);
860 		  }
861 		}
862 		PMAP_UNLOCK(kernel_pmap);
863 	}
864 
865 	/*
866 	 * Make sure the kernel and BPVO pool stay mapped on systems either
867 	 * without a direct map or on which the kernel is not already executing
868 	 * out of the direct-mapped region.
869 	 */
870 	if (kernelstart < DMAP_BASE_ADDRESS) {
871 		/*
872 		 * For pre-dmap execution, we need to use identity mapping
873 		 * because we will be operating with the mmu on but in the
874 		 * wrong address configuration until we __restartkernel().
875 		 */
876 		for (pa = kernelstart & ~PAGE_MASK; pa < kernelend;
877 		    pa += PAGE_SIZE)
878 			moea64_kenter(pa, pa);
879 	} else if (!hw_direct_map) {
880 		pkernelstart = kernelstart & ~DMAP_BASE_ADDRESS;
881 		pkernelend = kernelend & ~DMAP_BASE_ADDRESS;
882 		for (pa = pkernelstart & ~PAGE_MASK; pa < pkernelend;
883 		    pa += PAGE_SIZE)
884 			moea64_kenter(pa | DMAP_BASE_ADDRESS, pa);
885 	}
886 
887 	if (!hw_direct_map) {
888 		size = moea64_bpvo_pool_size*sizeof(struct pvo_entry);
889 		off = (vm_offset_t)(moea64_bpvo_pool);
890 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
891 			moea64_kenter(pa, pa);
892 
893 		/* Map exception vectors */
894 		for (pa = EXC_RSVD; pa < EXC_LAST; pa += PAGE_SIZE)
895 			moea64_kenter(pa | DMAP_BASE_ADDRESS, pa);
896 	}
897 	ENABLE_TRANS(msr);
898 
899 	/*
900 	 * Allow user to override unmapped_buf_allowed for testing.
901 	 * XXXKIB Only direct map implementation was tested.
902 	 */
903 	if (!TUNABLE_INT_FETCH("vfs.unmapped_buf_allowed",
904 	    &unmapped_buf_allowed))
905 		unmapped_buf_allowed = hw_direct_map;
906 }
907 
908 /* Quick sort callout for comparing physical addresses. */
909 static int
910 pa_cmp(const void *a, const void *b)
911 {
912 	const vm_paddr_t *pa = a, *pb = b;
913 
914 	if (*pa < *pb)
915 		return (-1);
916 	else if (*pa > *pb)
917 		return (1);
918 	else
919 		return (0);
920 }
921 
922 void
923 moea64_early_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
924 {
925 	int		i, j;
926 	vm_size_t	physsz, hwphyssz;
927 	vm_paddr_t	kernelphysstart, kernelphysend;
928 	int		rm_pavail;
929 
930 	/* Level 0 reservations consist of 4096 pages (16MB superpage). */
931 	vm_level_0_order = 12;
932 
933 #ifndef __powerpc64__
934 	/* We don't have a direct map since there is no BAT */
935 	hw_direct_map = 0;
936 
937 	/* Make sure battable is zero, since we have no BAT */
938 	for (i = 0; i < 16; i++) {
939 		battable[i].batu = 0;
940 		battable[i].batl = 0;
941 	}
942 #else
943 	/* Install trap handlers for SLBs */
944 	bcopy(&slbtrap, (void *)EXC_DSE,(size_t)&slbtrapend - (size_t)&slbtrap);
945 	bcopy(&slbtrap, (void *)EXC_ISE,(size_t)&slbtrapend - (size_t)&slbtrap);
946 	__syncicache((void *)EXC_DSE, 0x80);
947 	__syncicache((void *)EXC_ISE, 0x80);
948 #endif
949 
950 	kernelphysstart = kernelstart & ~DMAP_BASE_ADDRESS;
951 	kernelphysend = kernelend & ~DMAP_BASE_ADDRESS;
952 
953 	/* Get physical memory regions from firmware */
954 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
955 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
956 
957 	if (PHYS_AVAIL_ENTRIES < regions_sz)
958 		panic("moea64_bootstrap: phys_avail too small");
959 
960 	phys_avail_count = 0;
961 	physsz = 0;
962 	hwphyssz = 0;
963 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
964 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
965 		CTR3(KTR_PMAP, "region: %#zx - %#zx (%#zx)",
966 		    regions[i].mr_start, regions[i].mr_start +
967 		    regions[i].mr_size, regions[i].mr_size);
968 		if (hwphyssz != 0 &&
969 		    (physsz + regions[i].mr_size) >= hwphyssz) {
970 			if (physsz < hwphyssz) {
971 				phys_avail[j] = regions[i].mr_start;
972 				phys_avail[j + 1] = regions[i].mr_start +
973 				    hwphyssz - physsz;
974 				physsz = hwphyssz;
975 				phys_avail_count++;
976 				dump_avail[j] = phys_avail[j];
977 				dump_avail[j + 1] = phys_avail[j + 1];
978 			}
979 			break;
980 		}
981 		phys_avail[j] = regions[i].mr_start;
982 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
983 		phys_avail_count++;
984 		physsz += regions[i].mr_size;
985 		dump_avail[j] = phys_avail[j];
986 		dump_avail[j + 1] = phys_avail[j + 1];
987 	}
988 
989 	/* Check for overlap with the kernel and exception vectors */
990 	rm_pavail = 0;
991 	for (j = 0; j < 2*phys_avail_count; j+=2) {
992 		if (phys_avail[j] < EXC_LAST)
993 			phys_avail[j] += EXC_LAST;
994 
995 		if (phys_avail[j] >= kernelphysstart &&
996 		    phys_avail[j+1] <= kernelphysend) {
997 			phys_avail[j] = phys_avail[j+1] = ~0;
998 			rm_pavail++;
999 			continue;
1000 		}
1001 
1002 		if (kernelphysstart >= phys_avail[j] &&
1003 		    kernelphysstart < phys_avail[j+1]) {
1004 			if (kernelphysend < phys_avail[j+1]) {
1005 				phys_avail[2*phys_avail_count] =
1006 				    (kernelphysend & ~PAGE_MASK) + PAGE_SIZE;
1007 				phys_avail[2*phys_avail_count + 1] =
1008 				    phys_avail[j+1];
1009 				phys_avail_count++;
1010 			}
1011 
1012 			phys_avail[j+1] = kernelphysstart & ~PAGE_MASK;
1013 		}
1014 
1015 		if (kernelphysend >= phys_avail[j] &&
1016 		    kernelphysend < phys_avail[j+1]) {
1017 			if (kernelphysstart > phys_avail[j]) {
1018 				phys_avail[2*phys_avail_count] = phys_avail[j];
1019 				phys_avail[2*phys_avail_count + 1] =
1020 				    kernelphysstart & ~PAGE_MASK;
1021 				phys_avail_count++;
1022 			}
1023 
1024 			phys_avail[j] = (kernelphysend & ~PAGE_MASK) +
1025 			    PAGE_SIZE;
1026 		}
1027 	}
1028 
1029 	/* Remove physical available regions marked for removal (~0) */
1030 	if (rm_pavail) {
1031 		qsort(phys_avail, 2*phys_avail_count, sizeof(phys_avail[0]),
1032 			pa_cmp);
1033 		phys_avail_count -= rm_pavail;
1034 		for (i = 2*phys_avail_count;
1035 		     i < 2*(phys_avail_count + rm_pavail); i+=2)
1036 			phys_avail[i] = phys_avail[i+1] = 0;
1037 	}
1038 
1039 	physmem = btoc(physsz);
1040 
1041 #ifdef PTEGCOUNT
1042 	moea64_pteg_count = PTEGCOUNT;
1043 #else
1044 	moea64_pteg_count = 0x1000;
1045 
1046 	while (moea64_pteg_count < physmem)
1047 		moea64_pteg_count <<= 1;
1048 
1049 	moea64_pteg_count >>= 1;
1050 #endif /* PTEGCOUNT */
1051 }
1052 
1053 void
1054 moea64_mid_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
1055 {
1056 	int		i;
1057 
1058 	/*
1059 	 * Set PTEG mask
1060 	 */
1061 	moea64_pteg_mask = moea64_pteg_count - 1;
1062 
1063 	/*
1064 	 * Initialize SLB table lock and page locks
1065 	 */
1066 	mtx_init(&moea64_slb_mutex, "SLB table", NULL, MTX_DEF);
1067 	for (i = 0; i < PV_LOCK_COUNT; i++)
1068 		mtx_init(&pv_lock[i], "page pv", NULL, MTX_DEF);
1069 
1070 	/*
1071 	 * Initialise the bootstrap pvo pool.
1072 	 */
1073 	TUNABLE_INT_FETCH("machdep.moea64_bpvo_pool_size", &moea64_bpvo_pool_size);
1074 	if (moea64_bpvo_pool_size == 0) {
1075 		if (!hw_direct_map)
1076 			moea64_bpvo_pool_size = ((ptoa((uintmax_t)physmem) * sizeof(struct vm_page)) /
1077 			    (PAGE_SIZE * PAGE_SIZE)) * BPVO_POOL_EXPANSION_FACTOR;
1078 		else
1079 			moea64_bpvo_pool_size = BPVO_POOL_SIZE;
1080 	}
1081 
1082 	if (boothowto & RB_VERBOSE) {
1083 		printf("mmu_oea64: bpvo pool entries = %d, bpvo pool size = %zu MB\n",
1084 		    moea64_bpvo_pool_size,
1085 		    moea64_bpvo_pool_size*sizeof(struct pvo_entry) / 1048576);
1086 	}
1087 
1088 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
1089 		moea64_bpvo_pool_size*sizeof(struct pvo_entry), PAGE_SIZE);
1090 	moea64_bpvo_pool_index = 0;
1091 
1092 	/* Place at address usable through the direct map */
1093 	if (hw_direct_map)
1094 		moea64_bpvo_pool = (struct pvo_entry *)
1095 		    PHYS_TO_DMAP((uintptr_t)moea64_bpvo_pool);
1096 
1097 	/*
1098 	 * Make sure kernel vsid is allocated as well as VSID 0.
1099 	 */
1100 	#ifndef __powerpc64__
1101 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NVSIDS - 1)) / VSID_NBPW]
1102 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
1103 	moea64_vsid_bitmap[0] |= 1;
1104 	#endif
1105 
1106 	/*
1107 	 * Initialize the kernel pmap (which is statically allocated).
1108 	 */
1109 	#ifdef __powerpc64__
1110 	for (i = 0; i < 64; i++) {
1111 		pcpup->pc_aim.slb[i].slbv = 0;
1112 		pcpup->pc_aim.slb[i].slbe = 0;
1113 	}
1114 	#else
1115 	for (i = 0; i < 16; i++)
1116 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
1117 	#endif
1118 
1119 	kernel_pmap->pmap_phys = kernel_pmap;
1120 	CPU_FILL(&kernel_pmap->pm_active);
1121 	RB_INIT(&kernel_pmap->pmap_pvo);
1122 
1123 	PMAP_LOCK_INIT(kernel_pmap);
1124 
1125 	/*
1126 	 * Now map in all the other buffers we allocated earlier
1127 	 */
1128 
1129 	moea64_setup_direct_map(kernelstart, kernelend);
1130 }
1131 
1132 void
1133 moea64_late_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
1134 {
1135 	ihandle_t	mmui;
1136 	phandle_t	chosen;
1137 	phandle_t	mmu;
1138 	ssize_t		sz;
1139 	int		i;
1140 	vm_offset_t	pa, va;
1141 	void		*dpcpu;
1142 
1143 	/*
1144 	 * Set up the Open Firmware pmap and add its mappings if not in real
1145 	 * mode.
1146 	 */
1147 
1148 	chosen = OF_finddevice("/chosen");
1149 	if (chosen != -1 && OF_getencprop(chosen, "mmu", &mmui, 4) != -1) {
1150 		mmu = OF_instance_to_package(mmui);
1151 		if (mmu == -1 ||
1152 		    (sz = OF_getproplen(mmu, "translations")) == -1)
1153 			sz = 0;
1154 		if (sz > 6144 /* tmpstksz - 2 KB headroom */)
1155 			panic("moea64_bootstrap: too many ofw translations");
1156 
1157 		if (sz > 0)
1158 			moea64_add_ofw_mappings(mmu, sz);
1159 	}
1160 
1161 	/*
1162 	 * Calculate the last available physical address.
1163 	 */
1164 	Maxmem = 0;
1165 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
1166 		Maxmem = MAX(Maxmem, powerpc_btop(phys_avail[i + 1]));
1167 
1168 	/*
1169 	 * Initialize MMU.
1170 	 */
1171 	pmap_cpu_bootstrap(0);
1172 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
1173 	pmap_bootstrapped++;
1174 
1175 	/*
1176 	 * Set the start and end of kva.
1177 	 */
1178 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
1179 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
1180 
1181 	/*
1182 	 * Map the entire KVA range into the SLB. We must not fault there.
1183 	 */
1184 	#ifdef __powerpc64__
1185 	for (va = virtual_avail; va < virtual_end; va += SEGMENT_LENGTH)
1186 		moea64_bootstrap_slb_prefault(va, 0);
1187 	#endif
1188 
1189 	/*
1190 	 * Remap any early IO mappings (console framebuffer, etc.)
1191 	 */
1192 	bs_remap_earlyboot();
1193 
1194 	/*
1195 	 * Figure out how far we can extend virtual_end into segment 16
1196 	 * without running into existing mappings. Segment 16 is guaranteed
1197 	 * to contain neither RAM nor devices (at least on Apple hardware),
1198 	 * but will generally contain some OFW mappings we should not
1199 	 * step on.
1200 	 */
1201 
1202 	#ifndef __powerpc64__	/* KVA is in high memory on PPC64 */
1203 	PMAP_LOCK(kernel_pmap);
1204 	while (virtual_end < VM_MAX_KERNEL_ADDRESS &&
1205 	    moea64_pvo_find_va(kernel_pmap, virtual_end+1) == NULL)
1206 		virtual_end += PAGE_SIZE;
1207 	PMAP_UNLOCK(kernel_pmap);
1208 	#endif
1209 
1210 	/*
1211 	 * Allocate a kernel stack with a guard page for thread0 and map it
1212 	 * into the kernel page map.
1213 	 */
1214 	pa = moea64_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
1215 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
1216 	virtual_avail = va + kstack_pages * PAGE_SIZE;
1217 	CTR2(KTR_PMAP, "moea64_bootstrap: kstack0 at %#x (%#x)", pa, va);
1218 	thread0.td_kstack = va;
1219 	thread0.td_kstack_pages = kstack_pages;
1220 	for (i = 0; i < kstack_pages; i++) {
1221 		moea64_kenter(va, pa);
1222 		pa += PAGE_SIZE;
1223 		va += PAGE_SIZE;
1224 	}
1225 
1226 	/*
1227 	 * Allocate virtual address space for the message buffer.
1228 	 */
1229 	pa = msgbuf_phys = moea64_bootstrap_alloc(msgbufsize, PAGE_SIZE);
1230 	msgbufp = (struct msgbuf *)virtual_avail;
1231 	va = virtual_avail;
1232 	virtual_avail += round_page(msgbufsize);
1233 	while (va < virtual_avail) {
1234 		moea64_kenter(va, pa);
1235 		pa += PAGE_SIZE;
1236 		va += PAGE_SIZE;
1237 	}
1238 
1239 	/*
1240 	 * Allocate virtual address space for the dynamic percpu area.
1241 	 */
1242 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
1243 	dpcpu = (void *)virtual_avail;
1244 	va = virtual_avail;
1245 	virtual_avail += DPCPU_SIZE;
1246 	while (va < virtual_avail) {
1247 		moea64_kenter(va, pa);
1248 		pa += PAGE_SIZE;
1249 		va += PAGE_SIZE;
1250 	}
1251 	dpcpu_init(dpcpu, curcpu);
1252 
1253 	crashdumpmap = (caddr_t)virtual_avail;
1254 	virtual_avail += MAXDUMPPGS * PAGE_SIZE;
1255 
1256 	/*
1257 	 * Allocate some things for page zeroing. We put this directly
1258 	 * in the page table and use MOEA64_PTE_REPLACE to avoid any
1259 	 * of the PVO book-keeping or other parts of the VM system
1260 	 * from even knowing that this hack exists.
1261 	 */
1262 
1263 	if (!hw_direct_map) {
1264 		mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL,
1265 		    MTX_DEF);
1266 		for (i = 0; i < 2; i++) {
1267 			moea64_scratchpage_va[i] = (virtual_end+1) - PAGE_SIZE;
1268 			virtual_end -= PAGE_SIZE;
1269 
1270 			moea64_kenter(moea64_scratchpage_va[i], 0);
1271 
1272 			PMAP_LOCK(kernel_pmap);
1273 			moea64_scratchpage_pvo[i] = moea64_pvo_find_va(
1274 			    kernel_pmap, (vm_offset_t)moea64_scratchpage_va[i]);
1275 			PMAP_UNLOCK(kernel_pmap);
1276 		}
1277 	}
1278 
1279 	numa_mem_regions(&numa_pregions, &numapregions_sz);
1280 }
1281 
1282 static void
1283 moea64_pmap_init_qpages(void)
1284 {
1285 	struct pcpu *pc;
1286 	int i;
1287 
1288 	if (hw_direct_map)
1289 		return;
1290 
1291 	CPU_FOREACH(i) {
1292 		pc = pcpu_find(i);
1293 		pc->pc_qmap_addr = kva_alloc(PAGE_SIZE);
1294 		if (pc->pc_qmap_addr == 0)
1295 			panic("pmap_init_qpages: unable to allocate KVA");
1296 		PMAP_LOCK(kernel_pmap);
1297 		pc->pc_aim.qmap_pvo =
1298 		    moea64_pvo_find_va(kernel_pmap, pc->pc_qmap_addr);
1299 		PMAP_UNLOCK(kernel_pmap);
1300 		mtx_init(&pc->pc_aim.qmap_lock, "qmap lock", NULL, MTX_DEF);
1301 	}
1302 }
1303 
1304 SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, moea64_pmap_init_qpages, NULL);
1305 
1306 /*
1307  * Activate a user pmap.  This mostly involves setting some non-CPU
1308  * state.
1309  */
1310 void
1311 moea64_activate(struct thread *td)
1312 {
1313 	pmap_t	pm;
1314 
1315 	pm = &td->td_proc->p_vmspace->vm_pmap;
1316 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1317 
1318 	#ifdef __powerpc64__
1319 	PCPU_SET(aim.userslb, pm->pm_slb);
1320 	__asm __volatile("slbmte %0, %1; isync" ::
1321 	    "r"(td->td_pcb->pcb_cpu.aim.usr_vsid), "r"(USER_SLB_SLBE));
1322 	#else
1323 	PCPU_SET(curpmap, pm->pmap_phys);
1324 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1325 	#endif
1326 }
1327 
1328 void
1329 moea64_deactivate(struct thread *td)
1330 {
1331 	pmap_t	pm;
1332 
1333 	__asm __volatile("isync; slbie %0" :: "r"(USER_ADDR));
1334 
1335 	pm = &td->td_proc->p_vmspace->vm_pmap;
1336 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1337 	#ifdef __powerpc64__
1338 	PCPU_SET(aim.userslb, NULL);
1339 	#else
1340 	PCPU_SET(curpmap, NULL);
1341 	#endif
1342 }
1343 
1344 void
1345 moea64_unwire(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1346 {
1347 	struct	pvo_entry key, *pvo;
1348 	vm_page_t m;
1349 	int64_t	refchg;
1350 
1351 	key.pvo_vaddr = sva;
1352 	PMAP_LOCK(pm);
1353 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1354 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1355 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1356 		if (PVO_IS_SP(pvo)) {
1357 			if (moea64_sp_pvo_in_range(pvo, sva, eva)) {
1358 				pvo = moea64_sp_unwire(pvo);
1359 				continue;
1360 			} else {
1361 				CTR1(KTR_PMAP, "%s: demote before unwire",
1362 				    __func__);
1363 				moea64_sp_demote(pvo);
1364 			}
1365 		}
1366 
1367 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1368 			panic("moea64_unwire: pvo %p is missing PVO_WIRED",
1369 			    pvo);
1370 		pvo->pvo_vaddr &= ~PVO_WIRED;
1371 		refchg = moea64_pte_replace(pvo, 0 /* No invalidation */);
1372 		if ((pvo->pvo_vaddr & PVO_MANAGED) &&
1373 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1374 			if (refchg < 0)
1375 				refchg = LPTE_CHG;
1376 			m = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
1377 
1378 			refchg |= atomic_readandclear_32(&m->md.mdpg_attrs);
1379 			if (refchg & LPTE_CHG)
1380 				vm_page_dirty(m);
1381 			if (refchg & LPTE_REF)
1382 				vm_page_aflag_set(m, PGA_REFERENCED);
1383 		}
1384 		pm->pm_stats.wired_count--;
1385 	}
1386 	PMAP_UNLOCK(pm);
1387 }
1388 
1389 static int
1390 moea64_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *pap)
1391 {
1392 	struct pvo_entry *pvo;
1393 	vm_paddr_t pa;
1394 	vm_page_t m;
1395 	int val;
1396 	bool managed;
1397 
1398 	PMAP_LOCK(pmap);
1399 
1400 	pvo = moea64_pvo_find_va(pmap, addr);
1401 	if (pvo != NULL) {
1402 		pa = PVO_PADDR(pvo);
1403 		m = PHYS_TO_VM_PAGE(pa);
1404 		managed = (pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED;
1405 		if (PVO_IS_SP(pvo))
1406 			val = MINCORE_INCORE | MINCORE_PSIND(1);
1407 		else
1408 			val = MINCORE_INCORE;
1409 	} else {
1410 		PMAP_UNLOCK(pmap);
1411 		return (0);
1412 	}
1413 
1414 	PMAP_UNLOCK(pmap);
1415 
1416 	if (m == NULL)
1417 		return (0);
1418 
1419 	if (managed) {
1420 		if (moea64_is_modified(m))
1421 			val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
1422 
1423 		if (moea64_is_referenced(m))
1424 			val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
1425 	}
1426 
1427 	if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) !=
1428 	    (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) &&
1429 	    managed) {
1430 		*pap = pa;
1431 	}
1432 
1433 	return (val);
1434 }
1435 
1436 /*
1437  * This goes through and sets the physical address of our
1438  * special scratch PTE to the PA we want to zero or copy. Because
1439  * of locking issues (this can get called in pvo_enter() by
1440  * the UMA allocator), we can't use most other utility functions here
1441  */
1442 
1443 static __inline
1444 void moea64_set_scratchpage_pa(int which, vm_paddr_t pa)
1445 {
1446 	struct pvo_entry *pvo;
1447 
1448 	KASSERT(!hw_direct_map, ("Using OEA64 scratchpage with a direct map!"));
1449 	mtx_assert(&moea64_scratchpage_mtx, MA_OWNED);
1450 
1451 	pvo = moea64_scratchpage_pvo[which];
1452 	PMAP_LOCK(pvo->pvo_pmap);
1453 	pvo->pvo_pte.pa =
1454 	    moea64_calc_wimg(pa, VM_MEMATTR_DEFAULT) | (uint64_t)pa;
1455 	moea64_pte_replace(pvo, MOEA64_PTE_INVALIDATE);
1456 	PMAP_UNLOCK(pvo->pvo_pmap);
1457 	isync();
1458 }
1459 
1460 void
1461 moea64_copy_page(vm_page_t msrc, vm_page_t mdst)
1462 {
1463 	mtx_lock(&moea64_scratchpage_mtx);
1464 
1465 	moea64_set_scratchpage_pa(0, VM_PAGE_TO_PHYS(msrc));
1466 	moea64_set_scratchpage_pa(1, VM_PAGE_TO_PHYS(mdst));
1467 
1468 	bcopy((void *)moea64_scratchpage_va[0],
1469 	    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1470 
1471 	mtx_unlock(&moea64_scratchpage_mtx);
1472 }
1473 
1474 void
1475 moea64_copy_page_dmap(vm_page_t msrc, vm_page_t mdst)
1476 {
1477 	vm_offset_t	dst;
1478 	vm_offset_t	src;
1479 
1480 	dst = VM_PAGE_TO_PHYS(mdst);
1481 	src = VM_PAGE_TO_PHYS(msrc);
1482 
1483 	bcopy((void *)PHYS_TO_DMAP(src), (void *)PHYS_TO_DMAP(dst),
1484 	    PAGE_SIZE);
1485 }
1486 
1487 inline void
1488 moea64_copy_pages_dmap(vm_page_t *ma, vm_offset_t a_offset,
1489     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1490 {
1491 	void *a_cp, *b_cp;
1492 	vm_offset_t a_pg_offset, b_pg_offset;
1493 	int cnt;
1494 
1495 	while (xfersize > 0) {
1496 		a_pg_offset = a_offset & PAGE_MASK;
1497 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1498 		a_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1499 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT])) +
1500 		    a_pg_offset;
1501 		b_pg_offset = b_offset & PAGE_MASK;
1502 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1503 		b_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1504 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT])) +
1505 		    b_pg_offset;
1506 		bcopy(a_cp, b_cp, cnt);
1507 		a_offset += cnt;
1508 		b_offset += cnt;
1509 		xfersize -= cnt;
1510 	}
1511 }
1512 
1513 void
1514 moea64_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
1515     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1516 {
1517 	void *a_cp, *b_cp;
1518 	vm_offset_t a_pg_offset, b_pg_offset;
1519 	int cnt;
1520 
1521 	mtx_lock(&moea64_scratchpage_mtx);
1522 	while (xfersize > 0) {
1523 		a_pg_offset = a_offset & PAGE_MASK;
1524 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1525 		moea64_set_scratchpage_pa(0,
1526 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]));
1527 		a_cp = (char *)moea64_scratchpage_va[0] + a_pg_offset;
1528 		b_pg_offset = b_offset & PAGE_MASK;
1529 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1530 		moea64_set_scratchpage_pa(1,
1531 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]));
1532 		b_cp = (char *)moea64_scratchpage_va[1] + b_pg_offset;
1533 		bcopy(a_cp, b_cp, cnt);
1534 		a_offset += cnt;
1535 		b_offset += cnt;
1536 		xfersize -= cnt;
1537 	}
1538 	mtx_unlock(&moea64_scratchpage_mtx);
1539 }
1540 
1541 void
1542 moea64_zero_page_area(vm_page_t m, int off, int size)
1543 {
1544 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1545 
1546 	if (size + off > PAGE_SIZE)
1547 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1548 
1549 	if (hw_direct_map) {
1550 		bzero((caddr_t)(uintptr_t)PHYS_TO_DMAP(pa) + off, size);
1551 	} else {
1552 		mtx_lock(&moea64_scratchpage_mtx);
1553 		moea64_set_scratchpage_pa(0, pa);
1554 		bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1555 		mtx_unlock(&moea64_scratchpage_mtx);
1556 	}
1557 }
1558 
1559 /*
1560  * Zero a page of physical memory by temporarily mapping it
1561  */
1562 void
1563 moea64_zero_page(vm_page_t m)
1564 {
1565 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1566 	vm_offset_t va, off;
1567 
1568 	mtx_lock(&moea64_scratchpage_mtx);
1569 
1570 	moea64_set_scratchpage_pa(0, pa);
1571 	va = moea64_scratchpage_va[0];
1572 
1573 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1574 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1575 
1576 	mtx_unlock(&moea64_scratchpage_mtx);
1577 }
1578 
1579 void
1580 moea64_zero_page_dmap(vm_page_t m)
1581 {
1582 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1583 	vm_offset_t va, off;
1584 
1585 	va = PHYS_TO_DMAP(pa);
1586 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1587 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1588 }
1589 
1590 vm_offset_t
1591 moea64_quick_enter_page(vm_page_t m)
1592 {
1593 	struct pvo_entry *pvo;
1594 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1595 
1596 	/*
1597  	 * MOEA64_PTE_REPLACE does some locking, so we can't just grab
1598 	 * a critical section and access the PCPU data like on i386.
1599 	 * Instead, pin the thread and grab the PCPU lock to prevent
1600 	 * a preempting thread from using the same PCPU data.
1601 	 */
1602 	sched_pin();
1603 
1604 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_NOTOWNED);
1605 	pvo = PCPU_GET(aim.qmap_pvo);
1606 
1607 	mtx_lock(PCPU_PTR(aim.qmap_lock));
1608 	pvo->pvo_pte.pa = moea64_calc_wimg(pa, pmap_page_get_memattr(m)) |
1609 	    (uint64_t)pa;
1610 	moea64_pte_replace(pvo, MOEA64_PTE_INVALIDATE);
1611 	isync();
1612 
1613 	return (PCPU_GET(qmap_addr));
1614 }
1615 
1616 vm_offset_t
1617 moea64_quick_enter_page_dmap(vm_page_t m)
1618 {
1619 
1620 	return (PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)));
1621 }
1622 
1623 void
1624 moea64_quick_remove_page(vm_offset_t addr)
1625 {
1626 
1627 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_OWNED);
1628 	KASSERT(PCPU_GET(qmap_addr) == addr,
1629 	    ("moea64_quick_remove_page: invalid address"));
1630 	mtx_unlock(PCPU_PTR(aim.qmap_lock));
1631 	sched_unpin();
1632 }
1633 
1634 boolean_t
1635 moea64_page_is_mapped(vm_page_t m)
1636 {
1637 	return (!LIST_EMPTY(&(m)->md.mdpg_pvoh));
1638 }
1639 
1640 /*
1641  * Map the given physical page at the specified virtual address in the
1642  * target pmap with the protection requested.  If specified the page
1643  * will be wired down.
1644  */
1645 
1646 int
1647 moea64_enter(pmap_t pmap, vm_offset_t va, vm_page_t m,
1648     vm_prot_t prot, u_int flags, int8_t psind)
1649 {
1650 	struct		pvo_entry *pvo, *oldpvo, *tpvo;
1651 	struct		pvo_head *pvo_head;
1652 	uint64_t	pte_lo;
1653 	int		error;
1654 	vm_paddr_t	pa;
1655 
1656 	if ((m->oflags & VPO_UNMANAGED) == 0) {
1657 		if ((flags & PMAP_ENTER_QUICK_LOCKED) == 0)
1658 			VM_PAGE_OBJECT_BUSY_ASSERT(m);
1659 		else
1660 			VM_OBJECT_ASSERT_LOCKED(m->object);
1661 	}
1662 
1663 	if (psind > 0)
1664 		return (moea64_sp_enter(pmap, va, m, prot, flags, psind));
1665 
1666 	pvo = alloc_pvo_entry(0);
1667 	if (pvo == NULL)
1668 		return (KERN_RESOURCE_SHORTAGE);
1669 	pvo->pvo_pmap = NULL; /* to be filled in later */
1670 	pvo->pvo_pte.prot = prot;
1671 
1672 	pa = VM_PAGE_TO_PHYS(m);
1673 	pte_lo = moea64_calc_wimg(pa, pmap_page_get_memattr(m));
1674 	pvo->pvo_pte.pa = pa | pte_lo;
1675 
1676 	if ((flags & PMAP_ENTER_WIRED) != 0)
1677 		pvo->pvo_vaddr |= PVO_WIRED;
1678 
1679 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea64_initialized) {
1680 		pvo_head = NULL;
1681 	} else {
1682 		pvo_head = &m->md.mdpg_pvoh;
1683 		pvo->pvo_vaddr |= PVO_MANAGED;
1684 	}
1685 
1686 	PV_LOCK(pa);
1687 	PMAP_LOCK(pmap);
1688 	if (pvo->pvo_pmap == NULL)
1689 		init_pvo_entry(pvo, pmap, va);
1690 
1691 	if (moea64_ps_enabled(pmap) &&
1692 	    (tpvo = moea64_pvo_find_va(pmap, va & ~HPT_SP_MASK)) != NULL &&
1693 	    PVO_IS_SP(tpvo)) {
1694 		/* Demote SP before entering a regular page */
1695 		CTR2(KTR_PMAP, "%s: demote before enter: va=%#jx",
1696 		    __func__, (uintmax_t)va);
1697 		moea64_sp_demote_aligned(tpvo);
1698 	}
1699 
1700 	if (prot & VM_PROT_WRITE)
1701 		if (pmap_bootstrapped &&
1702 		    (m->oflags & VPO_UNMANAGED) == 0)
1703 			vm_page_aflag_set(m, PGA_WRITEABLE);
1704 
1705 	error = moea64_pvo_enter(pvo, pvo_head, &oldpvo);
1706 	if (error == EEXIST) {
1707 		if (oldpvo->pvo_vaddr == pvo->pvo_vaddr &&
1708 		    oldpvo->pvo_pte.pa == pvo->pvo_pte.pa &&
1709 		    oldpvo->pvo_pte.prot == prot) {
1710 			/* Identical mapping already exists */
1711 			error = 0;
1712 
1713 			/* If not in page table, reinsert it */
1714 			if (moea64_pte_synch(oldpvo) < 0) {
1715 				STAT_MOEA64(moea64_pte_overflow--);
1716 				moea64_pte_insert(oldpvo);
1717 			}
1718 
1719 			/* Then just clean up and go home */
1720 			PMAP_UNLOCK(pmap);
1721 			PV_UNLOCK(pa);
1722 			free_pvo_entry(pvo);
1723 			pvo = NULL;
1724 			goto out;
1725 		} else {
1726 			/* Otherwise, need to kill it first */
1727 			KASSERT(oldpvo->pvo_pmap == pmap, ("pmap of old "
1728 			    "mapping does not match new mapping"));
1729 			moea64_pvo_remove_from_pmap(oldpvo);
1730 			moea64_pvo_enter(pvo, pvo_head, NULL);
1731 		}
1732 	}
1733 	PMAP_UNLOCK(pmap);
1734 	PV_UNLOCK(pa);
1735 
1736 	/* Free any dead pages */
1737 	if (error == EEXIST) {
1738 		moea64_pvo_remove_from_page(oldpvo);
1739 		free_pvo_entry(oldpvo);
1740 	}
1741 
1742 out:
1743 	/*
1744 	 * Flush the page from the instruction cache if this page is
1745 	 * mapped executable and cacheable.
1746 	 */
1747 	if (pmap != kernel_pmap && (m->a.flags & PGA_EXECUTABLE) == 0 &&
1748 	    (pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1749 		vm_page_aflag_set(m, PGA_EXECUTABLE);
1750 		moea64_syncicache(pmap, va, pa, PAGE_SIZE);
1751 	}
1752 
1753 #if VM_NRESERVLEVEL > 0
1754 	/*
1755 	 * Try to promote pages.
1756 	 *
1757 	 * If the VA of the entered page is not aligned with its PA,
1758 	 * don't try page promotion as it is not possible.
1759 	 * This reduces the number of promotion failures dramatically.
1760 	 */
1761 	if (moea64_ps_enabled(pmap) && pmap != kernel_pmap && pvo != NULL &&
1762 	    (pvo->pvo_vaddr & PVO_MANAGED) != 0 &&
1763 	    (va & HPT_SP_MASK) == (pa & HPT_SP_MASK) &&
1764 	    (m->flags & PG_FICTITIOUS) == 0 &&
1765 	    vm_reserv_level_iffullpop(m) == 0)
1766 		moea64_sp_promote(pmap, va, m);
1767 #endif
1768 
1769 	return (KERN_SUCCESS);
1770 }
1771 
1772 static void
1773 moea64_syncicache(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
1774     vm_size_t sz)
1775 {
1776 
1777 	/*
1778 	 * This is much trickier than on older systems because
1779 	 * we can't sync the icache on physical addresses directly
1780 	 * without a direct map. Instead we check a couple of cases
1781 	 * where the memory is already mapped in and, failing that,
1782 	 * use the same trick we use for page zeroing to create
1783 	 * a temporary mapping for this physical address.
1784 	 */
1785 
1786 	if (!pmap_bootstrapped) {
1787 		/*
1788 		 * If PMAP is not bootstrapped, we are likely to be
1789 		 * in real mode.
1790 		 */
1791 		__syncicache((void *)(uintptr_t)pa, sz);
1792 	} else if (pmap == kernel_pmap) {
1793 		__syncicache((void *)va, sz);
1794 	} else if (hw_direct_map) {
1795 		__syncicache((void *)(uintptr_t)PHYS_TO_DMAP(pa), sz);
1796 	} else {
1797 		/* Use the scratch page to set up a temp mapping */
1798 
1799 		mtx_lock(&moea64_scratchpage_mtx);
1800 
1801 		moea64_set_scratchpage_pa(1, pa & ~ADDR_POFF);
1802 		__syncicache((void *)(moea64_scratchpage_va[1] +
1803 		    (va & ADDR_POFF)), sz);
1804 
1805 		mtx_unlock(&moea64_scratchpage_mtx);
1806 	}
1807 }
1808 
1809 /*
1810  * Maps a sequence of resident pages belonging to the same object.
1811  * The sequence begins with the given page m_start.  This page is
1812  * mapped at the given virtual address start.  Each subsequent page is
1813  * mapped at a virtual address that is offset from start by the same
1814  * amount as the page is offset from m_start within the object.  The
1815  * last page in the sequence is the page with the largest offset from
1816  * m_start that can be mapped at a virtual address less than the given
1817  * virtual address end.  Not every virtual page between start and end
1818  * is mapped; only those for which a resident page exists with the
1819  * corresponding offset from m_start are mapped.
1820  */
1821 void
1822 moea64_enter_object(pmap_t pm, vm_offset_t start, vm_offset_t end,
1823     vm_page_t m_start, vm_prot_t prot)
1824 {
1825 	vm_page_t m;
1826 	vm_pindex_t diff, psize;
1827 	vm_offset_t va;
1828 	int8_t psind;
1829 
1830 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1831 
1832 	psize = atop(end - start);
1833 	m = m_start;
1834 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1835 		va = start + ptoa(diff);
1836 		if ((va & HPT_SP_MASK) == 0 && va + HPT_SP_SIZE <= end &&
1837 		    m->psind == 1 && moea64_ps_enabled(pm))
1838 			psind = 1;
1839 		else
1840 			psind = 0;
1841 		moea64_enter(pm, va, m, prot &
1842 		    (VM_PROT_READ | VM_PROT_EXECUTE),
1843 		    PMAP_ENTER_NOSLEEP | PMAP_ENTER_QUICK_LOCKED, psind);
1844 		if (psind == 1)
1845 			m = &m[HPT_SP_SIZE / PAGE_SIZE - 1];
1846 		m = TAILQ_NEXT(m, listq);
1847 	}
1848 }
1849 
1850 void
1851 moea64_enter_quick(pmap_t pm, vm_offset_t va, vm_page_t m,
1852     vm_prot_t prot)
1853 {
1854 
1855 	moea64_enter(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1856 	    PMAP_ENTER_NOSLEEP | PMAP_ENTER_QUICK_LOCKED, 0);
1857 }
1858 
1859 vm_paddr_t
1860 moea64_extract(pmap_t pm, vm_offset_t va)
1861 {
1862 	struct	pvo_entry *pvo;
1863 	vm_paddr_t pa;
1864 
1865 	PMAP_LOCK(pm);
1866 	pvo = moea64_pvo_find_va(pm, va);
1867 	if (pvo == NULL)
1868 		pa = 0;
1869 	else
1870 		pa = PVO_PADDR(pvo) | (va - PVO_VADDR(pvo));
1871 	PMAP_UNLOCK(pm);
1872 
1873 	return (pa);
1874 }
1875 
1876 /*
1877  * Atomically extract and hold the physical page with the given
1878  * pmap and virtual address pair if that mapping permits the given
1879  * protection.
1880  */
1881 vm_page_t
1882 moea64_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1883 {
1884 	struct	pvo_entry *pvo;
1885 	vm_page_t m;
1886 
1887 	m = NULL;
1888 	PMAP_LOCK(pmap);
1889 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1890 	if (pvo != NULL && (pvo->pvo_pte.prot & prot) == prot) {
1891 		m = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
1892 		if (!vm_page_wire_mapped(m))
1893 			m = NULL;
1894 	}
1895 	PMAP_UNLOCK(pmap);
1896 	return (m);
1897 }
1898 
1899 static void *
1900 moea64_uma_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain,
1901     uint8_t *flags, int wait)
1902 {
1903 	struct pvo_entry *pvo;
1904         vm_offset_t va;
1905         vm_page_t m;
1906         int needed_lock;
1907 
1908 	/*
1909 	 * This entire routine is a horrible hack to avoid bothering kmem
1910 	 * for new KVA addresses. Because this can get called from inside
1911 	 * kmem allocation routines, calling kmem for a new address here
1912 	 * can lead to multiply locking non-recursive mutexes.
1913 	 */
1914 
1915 	*flags = UMA_SLAB_PRIV;
1916 	needed_lock = !PMAP_LOCKED(kernel_pmap);
1917 
1918 	m = vm_page_alloc_domain(NULL, 0, domain,
1919 	    malloc2vm_flags(wait) | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1920 	if (m == NULL)
1921 		return (NULL);
1922 
1923 	va = VM_PAGE_TO_PHYS(m);
1924 
1925 	pvo = alloc_pvo_entry(1 /* bootstrap */);
1926 
1927 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE;
1928 	pvo->pvo_pte.pa = VM_PAGE_TO_PHYS(m) | LPTE_M;
1929 
1930 	if (needed_lock)
1931 		PMAP_LOCK(kernel_pmap);
1932 
1933 	init_pvo_entry(pvo, kernel_pmap, va);
1934 	pvo->pvo_vaddr |= PVO_WIRED;
1935 
1936 	moea64_pvo_enter(pvo, NULL, NULL);
1937 
1938 	if (needed_lock)
1939 		PMAP_UNLOCK(kernel_pmap);
1940 
1941 	if ((wait & M_ZERO) && (m->flags & PG_ZERO) == 0)
1942                 bzero((void *)va, PAGE_SIZE);
1943 
1944 	return (void *)va;
1945 }
1946 
1947 extern int elf32_nxstack;
1948 
1949 void
1950 moea64_init()
1951 {
1952 
1953 	CTR0(KTR_PMAP, "moea64_init");
1954 
1955 	moea64_pvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1956 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1957 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1958 
1959 	/*
1960 	 * Are large page mappings enabled?
1961 	 *
1962 	 * While HPT superpages are not better tested, leave it disabled by
1963 	 * default.
1964 	 */
1965 	superpages_enabled = 0;
1966 	TUNABLE_INT_FETCH("vm.pmap.superpages_enabled", &superpages_enabled);
1967 	if (superpages_enabled) {
1968 		KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0,
1969 		    ("moea64_init: can't assign to pagesizes[1]"));
1970 
1971 		if (moea64_large_page_size == 0) {
1972 			printf("mmu_oea64: HW does not support large pages. "
1973 					"Disabling superpages...\n");
1974 			superpages_enabled = 0;
1975 		} else if (!moea64_has_lp_4k_16m) {
1976 			printf("mmu_oea64: "
1977 			    "HW does not support mixed 4KB/16MB page sizes. "
1978 			    "Disabling superpages...\n");
1979 			superpages_enabled = 0;
1980 		} else
1981 			pagesizes[1] = HPT_SP_SIZE;
1982 	}
1983 
1984 	if (!hw_direct_map) {
1985 		uma_zone_set_allocf(moea64_pvo_zone, moea64_uma_page_alloc);
1986 	}
1987 
1988 #ifdef COMPAT_FREEBSD32
1989 	elf32_nxstack = 1;
1990 #endif
1991 
1992 	moea64_initialized = TRUE;
1993 }
1994 
1995 boolean_t
1996 moea64_is_referenced(vm_page_t m)
1997 {
1998 
1999 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2000 	    ("moea64_is_referenced: page %p is not managed", m));
2001 
2002 	return (moea64_query_bit(m, LPTE_REF));
2003 }
2004 
2005 boolean_t
2006 moea64_is_modified(vm_page_t m)
2007 {
2008 
2009 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2010 	    ("moea64_is_modified: page %p is not managed", m));
2011 
2012 	/*
2013 	 * If the page is not busied then this check is racy.
2014 	 */
2015 	if (!pmap_page_is_write_mapped(m))
2016 		return (FALSE);
2017 
2018 	return (moea64_query_bit(m, LPTE_CHG));
2019 }
2020 
2021 boolean_t
2022 moea64_is_prefaultable(pmap_t pmap, vm_offset_t va)
2023 {
2024 	struct pvo_entry *pvo;
2025 	boolean_t rv = TRUE;
2026 
2027 	PMAP_LOCK(pmap);
2028 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
2029 	if (pvo != NULL)
2030 		rv = FALSE;
2031 	PMAP_UNLOCK(pmap);
2032 	return (rv);
2033 }
2034 
2035 void
2036 moea64_clear_modify(vm_page_t m)
2037 {
2038 
2039 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2040 	    ("moea64_clear_modify: page %p is not managed", m));
2041 	vm_page_assert_busied(m);
2042 
2043 	if (!pmap_page_is_write_mapped(m))
2044 		return;
2045 	moea64_clear_bit(m, LPTE_CHG);
2046 }
2047 
2048 /*
2049  * Clear the write and modified bits in each of the given page's mappings.
2050  */
2051 void
2052 moea64_remove_write(vm_page_t m)
2053 {
2054 	struct	pvo_entry *pvo;
2055 	int64_t	refchg, ret;
2056 	pmap_t	pmap;
2057 
2058 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2059 	    ("moea64_remove_write: page %p is not managed", m));
2060 	vm_page_assert_busied(m);
2061 
2062 	if (!pmap_page_is_write_mapped(m))
2063 		return;
2064 
2065 	powerpc_sync();
2066 	PV_PAGE_LOCK(m);
2067 	refchg = 0;
2068 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2069 		pmap = pvo->pvo_pmap;
2070 		PMAP_LOCK(pmap);
2071 		if (!(pvo->pvo_vaddr & PVO_DEAD) &&
2072 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2073 			if (PVO_IS_SP(pvo)) {
2074 				CTR1(KTR_PMAP, "%s: demote before remwr",
2075 				    __func__);
2076 				moea64_sp_demote(pvo);
2077 			}
2078 			pvo->pvo_pte.prot &= ~VM_PROT_WRITE;
2079 			ret = moea64_pte_replace(pvo, MOEA64_PTE_PROT_UPDATE);
2080 			if (ret < 0)
2081 				ret = LPTE_CHG;
2082 			refchg |= ret;
2083 			if (pvo->pvo_pmap == kernel_pmap)
2084 				isync();
2085 		}
2086 		PMAP_UNLOCK(pmap);
2087 	}
2088 	if ((refchg | atomic_readandclear_32(&m->md.mdpg_attrs)) & LPTE_CHG)
2089 		vm_page_dirty(m);
2090 	vm_page_aflag_clear(m, PGA_WRITEABLE);
2091 	PV_PAGE_UNLOCK(m);
2092 }
2093 
2094 /*
2095  *	moea64_ts_referenced:
2096  *
2097  *	Return a count of reference bits for a page, clearing those bits.
2098  *	It is not necessary for every reference bit to be cleared, but it
2099  *	is necessary that 0 only be returned when there are truly no
2100  *	reference bits set.
2101  *
2102  *	XXX: The exact number of bits to check and clear is a matter that
2103  *	should be tested and standardized at some point in the future for
2104  *	optimal aging of shared pages.
2105  */
2106 int
2107 moea64_ts_referenced(vm_page_t m)
2108 {
2109 
2110 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2111 	    ("moea64_ts_referenced: page %p is not managed", m));
2112 	return (moea64_clear_bit(m, LPTE_REF));
2113 }
2114 
2115 /*
2116  * Modify the WIMG settings of all mappings for a page.
2117  */
2118 void
2119 moea64_page_set_memattr(vm_page_t m, vm_memattr_t ma)
2120 {
2121 	struct	pvo_entry *pvo;
2122 	int64_t	refchg;
2123 	pmap_t	pmap;
2124 	uint64_t lo;
2125 
2126 	CTR3(KTR_PMAP, "%s: pa=%#jx, ma=%#x",
2127 	    __func__, (uintmax_t)VM_PAGE_TO_PHYS(m), ma);
2128 
2129 	if ((m->oflags & VPO_UNMANAGED) != 0) {
2130 		m->md.mdpg_cache_attrs = ma;
2131 		return;
2132 	}
2133 
2134 	lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
2135 
2136 	PV_PAGE_LOCK(m);
2137 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2138 		pmap = pvo->pvo_pmap;
2139 		PMAP_LOCK(pmap);
2140 		if (!(pvo->pvo_vaddr & PVO_DEAD)) {
2141 			if (PVO_IS_SP(pvo)) {
2142 				CTR1(KTR_PMAP,
2143 				    "%s: demote before set_memattr", __func__);
2144 				moea64_sp_demote(pvo);
2145 			}
2146 			pvo->pvo_pte.pa &= ~LPTE_WIMG;
2147 			pvo->pvo_pte.pa |= lo;
2148 			refchg = moea64_pte_replace(pvo, MOEA64_PTE_INVALIDATE);
2149 			if (refchg < 0)
2150 				refchg = (pvo->pvo_pte.prot & VM_PROT_WRITE) ?
2151 				    LPTE_CHG : 0;
2152 			if ((pvo->pvo_vaddr & PVO_MANAGED) &&
2153 			    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2154 				refchg |=
2155 				    atomic_readandclear_32(&m->md.mdpg_attrs);
2156 				if (refchg & LPTE_CHG)
2157 					vm_page_dirty(m);
2158 				if (refchg & LPTE_REF)
2159 					vm_page_aflag_set(m, PGA_REFERENCED);
2160 			}
2161 			if (pvo->pvo_pmap == kernel_pmap)
2162 				isync();
2163 		}
2164 		PMAP_UNLOCK(pmap);
2165 	}
2166 	m->md.mdpg_cache_attrs = ma;
2167 	PV_PAGE_UNLOCK(m);
2168 }
2169 
2170 /*
2171  * Map a wired page into kernel virtual address space.
2172  */
2173 void
2174 moea64_kenter_attr(vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
2175 {
2176 	int		error;
2177 	struct pvo_entry *pvo, *oldpvo;
2178 
2179 	do {
2180 		pvo = alloc_pvo_entry(0);
2181 		if (pvo == NULL)
2182 			vm_wait(NULL);
2183 	} while (pvo == NULL);
2184 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;
2185 	pvo->pvo_pte.pa = (pa & ~ADDR_POFF) | moea64_calc_wimg(pa, ma);
2186 	pvo->pvo_vaddr |= PVO_WIRED;
2187 
2188 	PMAP_LOCK(kernel_pmap);
2189 	oldpvo = moea64_pvo_find_va(kernel_pmap, va);
2190 	if (oldpvo != NULL)
2191 		moea64_pvo_remove_from_pmap(oldpvo);
2192 	init_pvo_entry(pvo, kernel_pmap, va);
2193 	error = moea64_pvo_enter(pvo, NULL, NULL);
2194 	PMAP_UNLOCK(kernel_pmap);
2195 
2196 	/* Free any dead pages */
2197 	if (oldpvo != NULL) {
2198 		moea64_pvo_remove_from_page(oldpvo);
2199 		free_pvo_entry(oldpvo);
2200 	}
2201 
2202 	if (error != 0)
2203 		panic("moea64_kenter: failed to enter va %#zx pa %#jx: %d", va,
2204 		    (uintmax_t)pa, error);
2205 }
2206 
2207 void
2208 moea64_kenter(vm_offset_t va, vm_paddr_t pa)
2209 {
2210 
2211 	moea64_kenter_attr(va, pa, VM_MEMATTR_DEFAULT);
2212 }
2213 
2214 /*
2215  * Extract the physical page address associated with the given kernel virtual
2216  * address.
2217  */
2218 vm_paddr_t
2219 moea64_kextract(vm_offset_t va)
2220 {
2221 	struct		pvo_entry *pvo;
2222 	vm_paddr_t pa;
2223 
2224 	/*
2225 	 * Shortcut the direct-mapped case when applicable.  We never put
2226 	 * anything but 1:1 (or 62-bit aliased) mappings below
2227 	 * VM_MIN_KERNEL_ADDRESS.
2228 	 */
2229 	if (va < VM_MIN_KERNEL_ADDRESS)
2230 		return (va & ~DMAP_BASE_ADDRESS);
2231 
2232 	PMAP_LOCK(kernel_pmap);
2233 	pvo = moea64_pvo_find_va(kernel_pmap, va);
2234 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found for %#" PRIxPTR,
2235 	    va));
2236 	pa = PVO_PADDR(pvo) | (va - PVO_VADDR(pvo));
2237 	PMAP_UNLOCK(kernel_pmap);
2238 	return (pa);
2239 }
2240 
2241 /*
2242  * Remove a wired page from kernel virtual address space.
2243  */
2244 void
2245 moea64_kremove(vm_offset_t va)
2246 {
2247 	moea64_remove(kernel_pmap, va, va + PAGE_SIZE);
2248 }
2249 
2250 /*
2251  * Provide a kernel pointer corresponding to a given userland pointer.
2252  * The returned pointer is valid until the next time this function is
2253  * called in this thread. This is used internally in copyin/copyout.
2254  */
2255 static int
2256 moea64_map_user_ptr(pmap_t pm, volatile const void *uaddr,
2257     void **kaddr, size_t ulen, size_t *klen)
2258 {
2259 	size_t l;
2260 #ifdef __powerpc64__
2261 	struct slb *slb;
2262 #endif
2263 	register_t slbv;
2264 
2265 	*kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
2266 	l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
2267 	if (l > ulen)
2268 		l = ulen;
2269 	if (klen)
2270 		*klen = l;
2271 	else if (l != ulen)
2272 		return (EFAULT);
2273 
2274 #ifdef __powerpc64__
2275 	/* Try lockless look-up first */
2276 	slb = user_va_to_slb_entry(pm, (vm_offset_t)uaddr);
2277 
2278 	if (slb == NULL) {
2279 		/* If it isn't there, we need to pre-fault the VSID */
2280 		PMAP_LOCK(pm);
2281 		slbv = va_to_vsid(pm, (vm_offset_t)uaddr) << SLBV_VSID_SHIFT;
2282 		PMAP_UNLOCK(pm);
2283 	} else {
2284 		slbv = slb->slbv;
2285 	}
2286 
2287 	/* Mark segment no-execute */
2288 	slbv |= SLBV_N;
2289 #else
2290 	slbv = va_to_vsid(pm, (vm_offset_t)uaddr);
2291 
2292 	/* Mark segment no-execute */
2293 	slbv |= SR_N;
2294 #endif
2295 
2296 	/* If we have already set this VSID, we can just return */
2297 	if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == slbv)
2298 		return (0);
2299 
2300 	__asm __volatile("isync");
2301 	curthread->td_pcb->pcb_cpu.aim.usr_segm =
2302 	    (uintptr_t)uaddr >> ADDR_SR_SHFT;
2303 	curthread->td_pcb->pcb_cpu.aim.usr_vsid = slbv;
2304 #ifdef __powerpc64__
2305 	__asm __volatile ("slbie %0; slbmte %1, %2; isync" ::
2306 	    "r"(USER_ADDR), "r"(slbv), "r"(USER_SLB_SLBE));
2307 #else
2308 	__asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(slbv));
2309 #endif
2310 
2311 	return (0);
2312 }
2313 
2314 /*
2315  * Figure out where a given kernel pointer (usually in a fault) points
2316  * to from the VM's perspective, potentially remapping into userland's
2317  * address space.
2318  */
2319 static int
2320 moea64_decode_kernel_ptr(vm_offset_t addr, int *is_user,
2321     vm_offset_t *decoded_addr)
2322 {
2323 	vm_offset_t user_sr;
2324 
2325 	if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
2326 		user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
2327 		addr &= ADDR_PIDX | ADDR_POFF;
2328 		addr |= user_sr << ADDR_SR_SHFT;
2329 		*decoded_addr = addr;
2330 		*is_user = 1;
2331 	} else {
2332 		*decoded_addr = addr;
2333 		*is_user = 0;
2334 	}
2335 
2336 	return (0);
2337 }
2338 
2339 /*
2340  * Map a range of physical addresses into kernel virtual address space.
2341  *
2342  * The value passed in *virt is a suggested virtual address for the mapping.
2343  * Architectures which can support a direct-mapped physical to virtual region
2344  * can return the appropriate address within that region, leaving '*virt'
2345  * unchanged.  Other architectures should map the pages starting at '*virt' and
2346  * update '*virt' with the first usable address after the mapped region.
2347  */
2348 vm_offset_t
2349 moea64_map(vm_offset_t *virt, vm_paddr_t pa_start,
2350     vm_paddr_t pa_end, int prot)
2351 {
2352 	vm_offset_t	sva, va;
2353 
2354 	if (hw_direct_map) {
2355 		/*
2356 		 * Check if every page in the region is covered by the direct
2357 		 * map. The direct map covers all of physical memory. Use
2358 		 * moea64_calc_wimg() as a shortcut to see if the page is in
2359 		 * physical memory as a way to see if the direct map covers it.
2360 		 */
2361 		for (va = pa_start; va < pa_end; va += PAGE_SIZE)
2362 			if (moea64_calc_wimg(va, VM_MEMATTR_DEFAULT) != LPTE_M)
2363 				break;
2364 		if (va == pa_end)
2365 			return (PHYS_TO_DMAP(pa_start));
2366 	}
2367 	sva = *virt;
2368 	va = sva;
2369 	/* XXX respect prot argument */
2370 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
2371 		moea64_kenter(va, pa_start);
2372 	*virt = va;
2373 
2374 	return (sva);
2375 }
2376 
2377 /*
2378  * Returns true if the pmap's pv is one of the first
2379  * 16 pvs linked to from this page.  This count may
2380  * be changed upwards or downwards in the future; it
2381  * is only necessary that true be returned for a small
2382  * subset of pmaps for proper page aging.
2383  */
2384 boolean_t
2385 moea64_page_exists_quick(pmap_t pmap, vm_page_t m)
2386 {
2387         int loops;
2388 	struct pvo_entry *pvo;
2389 	boolean_t rv;
2390 
2391 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2392 	    ("moea64_page_exists_quick: page %p is not managed", m));
2393 	loops = 0;
2394 	rv = FALSE;
2395 	PV_PAGE_LOCK(m);
2396 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2397 		if (!(pvo->pvo_vaddr & PVO_DEAD) && pvo->pvo_pmap == pmap) {
2398 			rv = TRUE;
2399 			break;
2400 		}
2401 		if (++loops >= 16)
2402 			break;
2403 	}
2404 	PV_PAGE_UNLOCK(m);
2405 	return (rv);
2406 }
2407 
2408 void
2409 moea64_page_init(vm_page_t m)
2410 {
2411 
2412 	m->md.mdpg_attrs = 0;
2413 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
2414 	LIST_INIT(&m->md.mdpg_pvoh);
2415 }
2416 
2417 /*
2418  * Return the number of managed mappings to the given physical page
2419  * that are wired.
2420  */
2421 int
2422 moea64_page_wired_mappings(vm_page_t m)
2423 {
2424 	struct pvo_entry *pvo;
2425 	int count;
2426 
2427 	count = 0;
2428 	if ((m->oflags & VPO_UNMANAGED) != 0)
2429 		return (count);
2430 	PV_PAGE_LOCK(m);
2431 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
2432 		if ((pvo->pvo_vaddr & (PVO_DEAD | PVO_WIRED)) == PVO_WIRED)
2433 			count++;
2434 	PV_PAGE_UNLOCK(m);
2435 	return (count);
2436 }
2437 
2438 static uintptr_t	moea64_vsidcontext;
2439 
2440 uintptr_t
2441 moea64_get_unique_vsid(void) {
2442 	u_int entropy;
2443 	register_t hash;
2444 	uint32_t mask;
2445 	int i;
2446 
2447 	entropy = 0;
2448 	__asm __volatile("mftb %0" : "=r"(entropy));
2449 
2450 	mtx_lock(&moea64_slb_mutex);
2451 	for (i = 0; i < NVSIDS; i += VSID_NBPW) {
2452 		u_int	n;
2453 
2454 		/*
2455 		 * Create a new value by mutiplying by a prime and adding in
2456 		 * entropy from the timebase register.  This is to make the
2457 		 * VSID more random so that the PT hash function collides
2458 		 * less often.  (Note that the prime casues gcc to do shifts
2459 		 * instead of a multiply.)
2460 		 */
2461 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
2462 		hash = moea64_vsidcontext & (NVSIDS - 1);
2463 		if (hash == 0)		/* 0 is special, avoid it */
2464 			continue;
2465 		n = hash >> 5;
2466 		mask = 1 << (hash & (VSID_NBPW - 1));
2467 		hash = (moea64_vsidcontext & VSID_HASHMASK);
2468 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
2469 			/* anything free in this bucket? */
2470 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
2471 				entropy = (moea64_vsidcontext >> 20);
2472 				continue;
2473 			}
2474 			i = ffs(~moea64_vsid_bitmap[n]) - 1;
2475 			mask = 1 << i;
2476 			hash &= rounddown2(VSID_HASHMASK, VSID_NBPW);
2477 			hash |= i;
2478 		}
2479 		if (hash == VSID_VRMA)	/* also special, avoid this too */
2480 			continue;
2481 		KASSERT(!(moea64_vsid_bitmap[n] & mask),
2482 		    ("Allocating in-use VSID %#zx\n", hash));
2483 		moea64_vsid_bitmap[n] |= mask;
2484 		mtx_unlock(&moea64_slb_mutex);
2485 		return (hash);
2486 	}
2487 
2488 	mtx_unlock(&moea64_slb_mutex);
2489 	panic("%s: out of segments",__func__);
2490 }
2491 
2492 #ifdef __powerpc64__
2493 int
2494 moea64_pinit(pmap_t pmap)
2495 {
2496 
2497 	RB_INIT(&pmap->pmap_pvo);
2498 
2499 	pmap->pm_slb_tree_root = slb_alloc_tree();
2500 	pmap->pm_slb = slb_alloc_user_cache();
2501 	pmap->pm_slb_len = 0;
2502 
2503 	return (1);
2504 }
2505 #else
2506 int
2507 moea64_pinit(pmap_t pmap)
2508 {
2509 	int	i;
2510 	uint32_t hash;
2511 
2512 	RB_INIT(&pmap->pmap_pvo);
2513 
2514 	if (pmap_bootstrapped)
2515 		pmap->pmap_phys = (pmap_t)moea64_kextract((vm_offset_t)pmap);
2516 	else
2517 		pmap->pmap_phys = pmap;
2518 
2519 	/*
2520 	 * Allocate some segment registers for this pmap.
2521 	 */
2522 	hash = moea64_get_unique_vsid();
2523 
2524 	for (i = 0; i < 16; i++)
2525 		pmap->pm_sr[i] = VSID_MAKE(i, hash);
2526 
2527 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_pinit: pm_sr[0] = 0"));
2528 
2529 	return (1);
2530 }
2531 #endif
2532 
2533 /*
2534  * Initialize the pmap associated with process 0.
2535  */
2536 void
2537 moea64_pinit0(pmap_t pm)
2538 {
2539 
2540 	PMAP_LOCK_INIT(pm);
2541 	moea64_pinit(pm);
2542 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
2543 }
2544 
2545 /*
2546  * Set the physical protection on the specified range of this map as requested.
2547  */
2548 static void
2549 moea64_pvo_protect( pmap_t pm, struct pvo_entry *pvo, vm_prot_t prot)
2550 {
2551 	struct vm_page *pg;
2552 	vm_prot_t oldprot;
2553 	int32_t refchg;
2554 
2555 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2556 
2557 	/*
2558 	 * Change the protection of the page.
2559 	 */
2560 	oldprot = pvo->pvo_pte.prot;
2561 	pvo->pvo_pte.prot = prot;
2562 	pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
2563 
2564 	/*
2565 	 * If the PVO is in the page table, update mapping
2566 	 */
2567 	refchg = moea64_pte_replace(pvo, MOEA64_PTE_PROT_UPDATE);
2568 	if (refchg < 0)
2569 		refchg = (oldprot & VM_PROT_WRITE) ? LPTE_CHG : 0;
2570 
2571 	if (pm != kernel_pmap && pg != NULL &&
2572 	    (pg->a.flags & PGA_EXECUTABLE) == 0 &&
2573 	    (pvo->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
2574 		if ((pg->oflags & VPO_UNMANAGED) == 0)
2575 			vm_page_aflag_set(pg, PGA_EXECUTABLE);
2576 		moea64_syncicache(pm, PVO_VADDR(pvo),
2577 		    PVO_PADDR(pvo), PAGE_SIZE);
2578 	}
2579 
2580 	/*
2581 	 * Update vm about the REF/CHG bits if the page is managed and we have
2582 	 * removed write access.
2583 	 */
2584 	if (pg != NULL && (pvo->pvo_vaddr & PVO_MANAGED) &&
2585 	    (oldprot & VM_PROT_WRITE)) {
2586 		refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2587 		if (refchg & LPTE_CHG)
2588 			vm_page_dirty(pg);
2589 		if (refchg & LPTE_REF)
2590 			vm_page_aflag_set(pg, PGA_REFERENCED);
2591 	}
2592 }
2593 
2594 void
2595 moea64_protect(pmap_t pm, vm_offset_t sva, vm_offset_t eva,
2596     vm_prot_t prot)
2597 {
2598 	struct	pvo_entry *pvo, key;
2599 
2600 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm,
2601 	    sva, eva, prot);
2602 
2603 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
2604 	    ("moea64_protect: non current pmap"));
2605 
2606 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2607 		moea64_remove(pm, sva, eva);
2608 		return;
2609 	}
2610 
2611 	PMAP_LOCK(pm);
2612 	key.pvo_vaddr = sva;
2613 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2614 	    pvo != NULL && PVO_VADDR(pvo) < eva;
2615 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
2616 		if (PVO_IS_SP(pvo)) {
2617 			if (moea64_sp_pvo_in_range(pvo, sva, eva)) {
2618 				pvo = moea64_sp_protect(pvo, prot);
2619 				continue;
2620 			} else {
2621 				CTR1(KTR_PMAP, "%s: demote before protect",
2622 				    __func__);
2623 				moea64_sp_demote(pvo);
2624 			}
2625 		}
2626 		moea64_pvo_protect(pm, pvo, prot);
2627 	}
2628 	PMAP_UNLOCK(pm);
2629 }
2630 
2631 /*
2632  * Map a list of wired pages into kernel virtual address space.  This is
2633  * intended for temporary mappings which do not need page modification or
2634  * references recorded.  Existing mappings in the region are overwritten.
2635  */
2636 void
2637 moea64_qenter(vm_offset_t va, vm_page_t *m, int count)
2638 {
2639 	while (count-- > 0) {
2640 		moea64_kenter(va, VM_PAGE_TO_PHYS(*m));
2641 		va += PAGE_SIZE;
2642 		m++;
2643 	}
2644 }
2645 
2646 /*
2647  * Remove page mappings from kernel virtual address space.  Intended for
2648  * temporary mappings entered by moea64_qenter.
2649  */
2650 void
2651 moea64_qremove(vm_offset_t va, int count)
2652 {
2653 	while (count-- > 0) {
2654 		moea64_kremove(va);
2655 		va += PAGE_SIZE;
2656 	}
2657 }
2658 
2659 void
2660 moea64_release_vsid(uint64_t vsid)
2661 {
2662 	int idx, mask;
2663 
2664 	mtx_lock(&moea64_slb_mutex);
2665 	idx = vsid & (NVSIDS-1);
2666 	mask = 1 << (idx % VSID_NBPW);
2667 	idx /= VSID_NBPW;
2668 	KASSERT(moea64_vsid_bitmap[idx] & mask,
2669 	    ("Freeing unallocated VSID %#jx", vsid));
2670 	moea64_vsid_bitmap[idx] &= ~mask;
2671 	mtx_unlock(&moea64_slb_mutex);
2672 }
2673 
2674 void
2675 moea64_release(pmap_t pmap)
2676 {
2677 
2678 	/*
2679 	 * Free segment registers' VSIDs
2680 	 */
2681     #ifdef __powerpc64__
2682 	slb_free_tree(pmap);
2683 	slb_free_user_cache(pmap->pm_slb);
2684     #else
2685 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_release: pm_sr[0] = 0"));
2686 
2687 	moea64_release_vsid(VSID_TO_HASH(pmap->pm_sr[0]));
2688     #endif
2689 }
2690 
2691 /*
2692  * Remove all pages mapped by the specified pmap
2693  */
2694 void
2695 moea64_remove_pages(pmap_t pm)
2696 {
2697 	struct pvo_entry *pvo, *tpvo;
2698 	struct pvo_dlist tofree;
2699 
2700 	SLIST_INIT(&tofree);
2701 
2702 	PMAP_LOCK(pm);
2703 	RB_FOREACH_SAFE(pvo, pvo_tree, &pm->pmap_pvo, tpvo) {
2704 		if (pvo->pvo_vaddr & PVO_WIRED)
2705 			continue;
2706 
2707 		/*
2708 		 * For locking reasons, remove this from the page table and
2709 		 * pmap, but save delinking from the vm_page for a second
2710 		 * pass
2711 		 */
2712 		moea64_pvo_remove_from_pmap(pvo);
2713 		SLIST_INSERT_HEAD(&tofree, pvo, pvo_dlink);
2714 	}
2715 	PMAP_UNLOCK(pm);
2716 
2717 	while (!SLIST_EMPTY(&tofree)) {
2718 		pvo = SLIST_FIRST(&tofree);
2719 		SLIST_REMOVE_HEAD(&tofree, pvo_dlink);
2720 		moea64_pvo_remove_from_page(pvo);
2721 		free_pvo_entry(pvo);
2722 	}
2723 }
2724 
2725 static void
2726 moea64_remove_locked(pmap_t pm, vm_offset_t sva, vm_offset_t eva,
2727     struct pvo_dlist *tofree)
2728 {
2729 	struct pvo_entry *pvo, *tpvo, key;
2730 
2731 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2732 
2733 	key.pvo_vaddr = sva;
2734 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2735 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2736 		if (PVO_IS_SP(pvo)) {
2737 			if (moea64_sp_pvo_in_range(pvo, sva, eva)) {
2738 				tpvo = moea64_sp_remove(pvo, tofree);
2739 				continue;
2740 			} else {
2741 				CTR1(KTR_PMAP, "%s: demote before remove",
2742 				    __func__);
2743 				moea64_sp_demote(pvo);
2744 			}
2745 		}
2746 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2747 
2748 		/*
2749 		 * For locking reasons, remove this from the page table and
2750 		 * pmap, but save delinking from the vm_page for a second
2751 		 * pass
2752 		 */
2753 		moea64_pvo_remove_from_pmap(pvo);
2754 		SLIST_INSERT_HEAD(tofree, pvo, pvo_dlink);
2755 	}
2756 }
2757 
2758 /*
2759  * Remove the given range of addresses from the specified map.
2760  */
2761 void
2762 moea64_remove(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
2763 {
2764 	struct pvo_entry *pvo;
2765 	struct pvo_dlist tofree;
2766 
2767 	/*
2768 	 * Perform an unsynchronized read.  This is, however, safe.
2769 	 */
2770 	if (pm->pm_stats.resident_count == 0)
2771 		return;
2772 
2773 	SLIST_INIT(&tofree);
2774 	PMAP_LOCK(pm);
2775 	moea64_remove_locked(pm, sva, eva, &tofree);
2776 	PMAP_UNLOCK(pm);
2777 
2778 	while (!SLIST_EMPTY(&tofree)) {
2779 		pvo = SLIST_FIRST(&tofree);
2780 		SLIST_REMOVE_HEAD(&tofree, pvo_dlink);
2781 		moea64_pvo_remove_from_page(pvo);
2782 		free_pvo_entry(pvo);
2783 	}
2784 }
2785 
2786 /*
2787  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
2788  * will reflect changes in pte's back to the vm_page.
2789  */
2790 void
2791 moea64_remove_all(vm_page_t m)
2792 {
2793 	struct	pvo_entry *pvo, *next_pvo;
2794 	struct	pvo_head freequeue;
2795 	int	wasdead;
2796 	pmap_t	pmap;
2797 
2798 	LIST_INIT(&freequeue);
2799 
2800 	PV_PAGE_LOCK(m);
2801 	LIST_FOREACH_SAFE(pvo, vm_page_to_pvoh(m), pvo_vlink, next_pvo) {
2802 		pmap = pvo->pvo_pmap;
2803 		PMAP_LOCK(pmap);
2804 		wasdead = (pvo->pvo_vaddr & PVO_DEAD);
2805 		if (!wasdead) {
2806 			if (PVO_IS_SP(pvo)) {
2807 				CTR1(KTR_PMAP, "%s: demote before remove_all",
2808 				    __func__);
2809 				moea64_sp_demote(pvo);
2810 			}
2811 			moea64_pvo_remove_from_pmap(pvo);
2812 		}
2813 		moea64_pvo_remove_from_page_locked(pvo, m);
2814 		if (!wasdead)
2815 			LIST_INSERT_HEAD(&freequeue, pvo, pvo_vlink);
2816 		PMAP_UNLOCK(pmap);
2817 
2818 	}
2819 	KASSERT(!pmap_page_is_mapped(m), ("Page still has mappings"));
2820 	KASSERT((m->a.flags & PGA_WRITEABLE) == 0, ("Page still writable"));
2821 	PV_PAGE_UNLOCK(m);
2822 
2823 	/* Clean up UMA allocations */
2824 	LIST_FOREACH_SAFE(pvo, &freequeue, pvo_vlink, next_pvo)
2825 		free_pvo_entry(pvo);
2826 }
2827 
2828 /*
2829  * Allocate a physical page of memory directly from the phys_avail map.
2830  * Can only be called from moea64_bootstrap before avail start and end are
2831  * calculated.
2832  */
2833 vm_offset_t
2834 moea64_bootstrap_alloc(vm_size_t size, vm_size_t align)
2835 {
2836 	vm_offset_t	s, e;
2837 	int		i, j;
2838 
2839 	size = round_page(size);
2840 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2841 		if (align != 0)
2842 			s = roundup2(phys_avail[i], align);
2843 		else
2844 			s = phys_avail[i];
2845 		e = s + size;
2846 
2847 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2848 			continue;
2849 
2850 		if (s + size > platform_real_maxaddr())
2851 			continue;
2852 
2853 		if (s == phys_avail[i]) {
2854 			phys_avail[i] += size;
2855 		} else if (e == phys_avail[i + 1]) {
2856 			phys_avail[i + 1] -= size;
2857 		} else {
2858 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2859 				phys_avail[j] = phys_avail[j - 2];
2860 				phys_avail[j + 1] = phys_avail[j - 1];
2861 			}
2862 
2863 			phys_avail[i + 3] = phys_avail[i + 1];
2864 			phys_avail[i + 1] = s;
2865 			phys_avail[i + 2] = e;
2866 			phys_avail_count++;
2867 		}
2868 
2869 		return (s);
2870 	}
2871 	panic("moea64_bootstrap_alloc: could not allocate memory");
2872 }
2873 
2874 static int
2875 moea64_pvo_enter(struct pvo_entry *pvo, struct pvo_head *pvo_head,
2876     struct pvo_entry **oldpvop)
2877 {
2878 	struct pvo_entry *old_pvo;
2879 	int err;
2880 
2881 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2882 
2883 	STAT_MOEA64(moea64_pvo_enter_calls++);
2884 
2885 	/*
2886 	 * Add to pmap list
2887 	 */
2888 	old_pvo = RB_INSERT(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2889 
2890 	if (old_pvo != NULL) {
2891 		if (oldpvop != NULL)
2892 			*oldpvop = old_pvo;
2893 		return (EEXIST);
2894 	}
2895 
2896 	if (pvo_head != NULL) {
2897 		LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2898 	}
2899 
2900 	if (pvo->pvo_vaddr & PVO_WIRED)
2901 		pvo->pvo_pmap->pm_stats.wired_count++;
2902 	pvo->pvo_pmap->pm_stats.resident_count++;
2903 
2904 	/*
2905 	 * Insert it into the hardware page table
2906 	 */
2907 	err = moea64_pte_insert(pvo);
2908 	if (err != 0) {
2909 		panic("moea64_pvo_enter: overflow");
2910 	}
2911 
2912 	STAT_MOEA64(moea64_pvo_entries++);
2913 
2914 	if (pvo->pvo_pmap == kernel_pmap)
2915 		isync();
2916 
2917 #ifdef __powerpc64__
2918 	/*
2919 	 * Make sure all our bootstrap mappings are in the SLB as soon
2920 	 * as virtual memory is switched on.
2921 	 */
2922 	if (!pmap_bootstrapped)
2923 		moea64_bootstrap_slb_prefault(PVO_VADDR(pvo),
2924 		    pvo->pvo_vaddr & PVO_LARGE);
2925 #endif
2926 
2927 	return (0);
2928 }
2929 
2930 static void
2931 moea64_pvo_remove_from_pmap(struct pvo_entry *pvo)
2932 {
2933 	struct	vm_page *pg;
2934 	int32_t refchg;
2935 
2936 	KASSERT(pvo->pvo_pmap != NULL, ("Trying to remove PVO with no pmap"));
2937 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2938 	KASSERT(!(pvo->pvo_vaddr & PVO_DEAD), ("Trying to remove dead PVO"));
2939 
2940 	/*
2941 	 * If there is an active pte entry, we need to deactivate it
2942 	 */
2943 	refchg = moea64_pte_unset(pvo);
2944 	if (refchg < 0) {
2945 		/*
2946 		 * If it was evicted from the page table, be pessimistic and
2947 		 * dirty the page.
2948 		 */
2949 		if (pvo->pvo_pte.prot & VM_PROT_WRITE)
2950 			refchg = LPTE_CHG;
2951 		else
2952 			refchg = 0;
2953 	}
2954 
2955 	/*
2956 	 * Update our statistics.
2957 	 */
2958 	pvo->pvo_pmap->pm_stats.resident_count--;
2959 	if (pvo->pvo_vaddr & PVO_WIRED)
2960 		pvo->pvo_pmap->pm_stats.wired_count--;
2961 
2962 	/*
2963 	 * Remove this PVO from the pmap list.
2964 	 */
2965 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2966 
2967 	/*
2968 	 * Mark this for the next sweep
2969 	 */
2970 	pvo->pvo_vaddr |= PVO_DEAD;
2971 
2972 	/* Send RC bits to VM */
2973 	if ((pvo->pvo_vaddr & PVO_MANAGED) &&
2974 	    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2975 		pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
2976 		if (pg != NULL) {
2977 			refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2978 			if (refchg & LPTE_CHG)
2979 				vm_page_dirty(pg);
2980 			if (refchg & LPTE_REF)
2981 				vm_page_aflag_set(pg, PGA_REFERENCED);
2982 		}
2983 	}
2984 }
2985 
2986 static inline void
2987 moea64_pvo_remove_from_page_locked(struct pvo_entry *pvo,
2988     vm_page_t m)
2989 {
2990 
2991 	KASSERT(pvo->pvo_vaddr & PVO_DEAD, ("Trying to delink live page"));
2992 
2993 	/* Use NULL pmaps as a sentinel for races in page deletion */
2994 	if (pvo->pvo_pmap == NULL)
2995 		return;
2996 	pvo->pvo_pmap = NULL;
2997 
2998 	/*
2999 	 * Update vm about page writeability/executability if managed
3000 	 */
3001 	PV_LOCKASSERT(PVO_PADDR(pvo));
3002 	if (pvo->pvo_vaddr & PVO_MANAGED) {
3003 		if (m != NULL) {
3004 			LIST_REMOVE(pvo, pvo_vlink);
3005 			if (LIST_EMPTY(vm_page_to_pvoh(m)))
3006 				vm_page_aflag_clear(m,
3007 				    PGA_WRITEABLE | PGA_EXECUTABLE);
3008 		}
3009 	}
3010 
3011 	STAT_MOEA64(moea64_pvo_entries--);
3012 	STAT_MOEA64(moea64_pvo_remove_calls++);
3013 }
3014 
3015 static void
3016 moea64_pvo_remove_from_page(struct pvo_entry *pvo)
3017 {
3018 	vm_page_t pg = NULL;
3019 
3020 	if (pvo->pvo_vaddr & PVO_MANAGED)
3021 		pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
3022 
3023 	PV_LOCK(PVO_PADDR(pvo));
3024 	moea64_pvo_remove_from_page_locked(pvo, pg);
3025 	PV_UNLOCK(PVO_PADDR(pvo));
3026 }
3027 
3028 static struct pvo_entry *
3029 moea64_pvo_find_va(pmap_t pm, vm_offset_t va)
3030 {
3031 	struct pvo_entry key;
3032 
3033 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
3034 
3035 	key.pvo_vaddr = va & ~ADDR_POFF;
3036 	return (RB_FIND(pvo_tree, &pm->pmap_pvo, &key));
3037 }
3038 
3039 static boolean_t
3040 moea64_query_bit(vm_page_t m, uint64_t ptebit)
3041 {
3042 	struct	pvo_entry *pvo;
3043 	int64_t ret;
3044 	boolean_t rv;
3045 	vm_page_t sp;
3046 
3047 	/*
3048 	 * See if this bit is stored in the page already.
3049 	 *
3050 	 * For superpages, the bit is stored in the first vm page.
3051 	 */
3052 	if ((m->md.mdpg_attrs & ptebit) != 0 ||
3053 	    ((sp = PHYS_TO_VM_PAGE(VM_PAGE_TO_PHYS(m) & ~HPT_SP_MASK)) != NULL &&
3054 	     (sp->md.mdpg_attrs & (ptebit | MDPG_ATTR_SP)) ==
3055 	     (ptebit | MDPG_ATTR_SP)))
3056 		return (TRUE);
3057 
3058 	/*
3059 	 * Examine each PTE.  Sync so that any pending REF/CHG bits are
3060 	 * flushed to the PTEs.
3061 	 */
3062 	rv = FALSE;
3063 	powerpc_sync();
3064 	PV_PAGE_LOCK(m);
3065 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
3066 		if (PVO_IS_SP(pvo)) {
3067 			ret = moea64_sp_query(pvo, ptebit);
3068 			/*
3069 			 * If SP was not demoted, check its REF/CHG bits here.
3070 			 */
3071 			if (ret != -1) {
3072 				if ((ret & ptebit) != 0) {
3073 					rv = TRUE;
3074 					break;
3075 				}
3076 				continue;
3077 			}
3078 			/* else, fallthrough */
3079 		}
3080 
3081 		ret = 0;
3082 
3083 		/*
3084 		 * See if this pvo has a valid PTE.  if so, fetch the
3085 		 * REF/CHG bits from the valid PTE.  If the appropriate
3086 		 * ptebit is set, return success.
3087 		 */
3088 		PMAP_LOCK(pvo->pvo_pmap);
3089 		if (!(pvo->pvo_vaddr & PVO_DEAD))
3090 			ret = moea64_pte_synch(pvo);
3091 		PMAP_UNLOCK(pvo->pvo_pmap);
3092 
3093 		if (ret > 0) {
3094 			atomic_set_32(&m->md.mdpg_attrs,
3095 			    ret & (LPTE_CHG | LPTE_REF));
3096 			if (ret & ptebit) {
3097 				rv = TRUE;
3098 				break;
3099 			}
3100 		}
3101 	}
3102 	PV_PAGE_UNLOCK(m);
3103 
3104 	return (rv);
3105 }
3106 
3107 static u_int
3108 moea64_clear_bit(vm_page_t m, u_int64_t ptebit)
3109 {
3110 	u_int	count;
3111 	struct	pvo_entry *pvo;
3112 	int64_t ret;
3113 
3114 	/*
3115 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
3116 	 * we can reset the right ones).
3117 	 */
3118 	powerpc_sync();
3119 
3120 	/*
3121 	 * For each pvo entry, clear the pte's ptebit.
3122 	 */
3123 	count = 0;
3124 	PV_PAGE_LOCK(m);
3125 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
3126 		if (PVO_IS_SP(pvo)) {
3127 			if ((ret = moea64_sp_clear(pvo, m, ptebit)) != -1) {
3128 				count += ret;
3129 				continue;
3130 			}
3131 		}
3132 		ret = 0;
3133 
3134 		PMAP_LOCK(pvo->pvo_pmap);
3135 		if (!(pvo->pvo_vaddr & PVO_DEAD))
3136 			ret = moea64_pte_clear(pvo, ptebit);
3137 		PMAP_UNLOCK(pvo->pvo_pmap);
3138 
3139 		if (ret > 0 && (ret & ptebit))
3140 			count++;
3141 	}
3142 	atomic_clear_32(&m->md.mdpg_attrs, ptebit);
3143 	PV_PAGE_UNLOCK(m);
3144 
3145 	return (count);
3146 }
3147 
3148 boolean_t
3149 moea64_dev_direct_mapped(vm_paddr_t pa, vm_size_t size)
3150 {
3151 	struct pvo_entry *pvo, key;
3152 	vm_offset_t ppa;
3153 	int error = 0;
3154 
3155 	if (hw_direct_map && mem_valid(pa, size) == 0)
3156 		return (0);
3157 
3158 	PMAP_LOCK(kernel_pmap);
3159 	ppa = pa & ~ADDR_POFF;
3160 	key.pvo_vaddr = DMAP_BASE_ADDRESS + ppa;
3161 	for (pvo = RB_FIND(pvo_tree, &kernel_pmap->pmap_pvo, &key);
3162 	    ppa < pa + size; ppa += PAGE_SIZE,
3163 	    pvo = RB_NEXT(pvo_tree, &kernel_pmap->pmap_pvo, pvo)) {
3164 		if (pvo == NULL || PVO_PADDR(pvo) != ppa) {
3165 			error = EFAULT;
3166 			break;
3167 		}
3168 	}
3169 	PMAP_UNLOCK(kernel_pmap);
3170 
3171 	return (error);
3172 }
3173 
3174 /*
3175  * Map a set of physical memory pages into the kernel virtual
3176  * address space. Return a pointer to where it is mapped. This
3177  * routine is intended to be used for mapping device memory,
3178  * NOT real memory.
3179  */
3180 void *
3181 moea64_mapdev_attr(vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
3182 {
3183 	vm_offset_t va, tmpva, ppa, offset;
3184 
3185 	ppa = trunc_page(pa);
3186 	offset = pa & PAGE_MASK;
3187 	size = roundup2(offset + size, PAGE_SIZE);
3188 
3189 	va = kva_alloc(size);
3190 
3191 	if (!va)
3192 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
3193 
3194 	for (tmpva = va; size > 0;) {
3195 		moea64_kenter_attr(tmpva, ppa, ma);
3196 		size -= PAGE_SIZE;
3197 		tmpva += PAGE_SIZE;
3198 		ppa += PAGE_SIZE;
3199 	}
3200 
3201 	return ((void *)(va + offset));
3202 }
3203 
3204 void *
3205 moea64_mapdev(vm_paddr_t pa, vm_size_t size)
3206 {
3207 
3208 	return moea64_mapdev_attr(pa, size, VM_MEMATTR_DEFAULT);
3209 }
3210 
3211 void
3212 moea64_unmapdev(vm_offset_t va, vm_size_t size)
3213 {
3214 	vm_offset_t base, offset;
3215 
3216 	base = trunc_page(va);
3217 	offset = va & PAGE_MASK;
3218 	size = roundup2(offset + size, PAGE_SIZE);
3219 
3220 	moea64_qremove(base, atop(size));
3221 	kva_free(base, size);
3222 }
3223 
3224 void
3225 moea64_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz)
3226 {
3227 	struct pvo_entry *pvo;
3228 	vm_offset_t lim;
3229 	vm_paddr_t pa;
3230 	vm_size_t len;
3231 
3232 	if (__predict_false(pm == NULL))
3233 		pm = &curthread->td_proc->p_vmspace->vm_pmap;
3234 
3235 	PMAP_LOCK(pm);
3236 	while (sz > 0) {
3237 		lim = round_page(va+1);
3238 		len = MIN(lim - va, sz);
3239 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
3240 		if (pvo != NULL && !(pvo->pvo_pte.pa & LPTE_I)) {
3241 			pa = PVO_PADDR(pvo) | (va & ADDR_POFF);
3242 			moea64_syncicache(pm, va, pa, len);
3243 		}
3244 		va += len;
3245 		sz -= len;
3246 	}
3247 	PMAP_UNLOCK(pm);
3248 }
3249 
3250 void
3251 moea64_dumpsys_map(vm_paddr_t pa, size_t sz, void **va)
3252 {
3253 
3254 	*va = (void *)(uintptr_t)pa;
3255 }
3256 
3257 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
3258 
3259 void
3260 moea64_scan_init()
3261 {
3262 	struct pvo_entry *pvo;
3263 	vm_offset_t va;
3264 	int i;
3265 
3266 	if (!do_minidump) {
3267 		/* Initialize phys. segments for dumpsys(). */
3268 		memset(&dump_map, 0, sizeof(dump_map));
3269 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
3270 		for (i = 0; i < pregions_sz; i++) {
3271 			dump_map[i].pa_start = pregions[i].mr_start;
3272 			dump_map[i].pa_size = pregions[i].mr_size;
3273 		}
3274 		return;
3275 	}
3276 
3277 	/* Virtual segments for minidumps: */
3278 	memset(&dump_map, 0, sizeof(dump_map));
3279 
3280 	/* 1st: kernel .data and .bss. */
3281 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
3282 	dump_map[0].pa_size = round_page((uintptr_t)_end) -
3283 	    dump_map[0].pa_start;
3284 
3285 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
3286 	dump_map[1].pa_start = (vm_paddr_t)(uintptr_t)msgbufp->msg_ptr;
3287 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
3288 
3289 	/* 3rd: kernel VM. */
3290 	va = dump_map[1].pa_start + dump_map[1].pa_size;
3291 	/* Find start of next chunk (from va). */
3292 	while (va < virtual_end) {
3293 		/* Don't dump the buffer cache. */
3294 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
3295 			va = kmi.buffer_eva;
3296 			continue;
3297 		}
3298 		pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
3299 		if (pvo != NULL && !(pvo->pvo_vaddr & PVO_DEAD))
3300 			break;
3301 		va += PAGE_SIZE;
3302 	}
3303 	if (va < virtual_end) {
3304 		dump_map[2].pa_start = va;
3305 		va += PAGE_SIZE;
3306 		/* Find last page in chunk. */
3307 		while (va < virtual_end) {
3308 			/* Don't run into the buffer cache. */
3309 			if (va == kmi.buffer_sva)
3310 				break;
3311 			pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
3312 			if (pvo == NULL || (pvo->pvo_vaddr & PVO_DEAD))
3313 				break;
3314 			va += PAGE_SIZE;
3315 		}
3316 		dump_map[2].pa_size = va - dump_map[2].pa_start;
3317 	}
3318 }
3319 
3320 #ifdef __powerpc64__
3321 
3322 static size_t
3323 moea64_scan_pmap()
3324 {
3325 	struct pvo_entry *pvo;
3326 	vm_paddr_t pa, pa_end;
3327 	vm_offset_t va, pgva, kstart, kend, kstart_lp, kend_lp;
3328 	uint64_t lpsize;
3329 
3330 	lpsize = moea64_large_page_size;
3331 	kstart = trunc_page((vm_offset_t)_etext);
3332 	kend = round_page((vm_offset_t)_end);
3333 	kstart_lp = kstart & ~moea64_large_page_mask;
3334 	kend_lp = (kend + moea64_large_page_mask) & ~moea64_large_page_mask;
3335 
3336 	CTR4(KTR_PMAP, "moea64_scan_pmap: kstart=0x%016lx, kend=0x%016lx, "
3337 	    "kstart_lp=0x%016lx, kend_lp=0x%016lx",
3338 	    kstart, kend, kstart_lp, kend_lp);
3339 
3340 	PMAP_LOCK(kernel_pmap);
3341 	RB_FOREACH(pvo, pvo_tree, &kernel_pmap->pmap_pvo) {
3342 		va = pvo->pvo_vaddr;
3343 
3344 		if (va & PVO_DEAD)
3345 			continue;
3346 
3347 		/* Skip DMAP (except kernel area) */
3348 		if (va >= DMAP_BASE_ADDRESS && va <= DMAP_MAX_ADDRESS) {
3349 			if (va & PVO_LARGE) {
3350 				pgva = va & ~moea64_large_page_mask;
3351 				if (pgva < kstart_lp || pgva >= kend_lp)
3352 					continue;
3353 			} else {
3354 				pgva = trunc_page(va);
3355 				if (pgva < kstart || pgva >= kend)
3356 					continue;
3357 			}
3358 		}
3359 
3360 		pa = PVO_PADDR(pvo);
3361 
3362 		if (va & PVO_LARGE) {
3363 			pa_end = pa + lpsize;
3364 			for (; pa < pa_end; pa += PAGE_SIZE) {
3365 				if (is_dumpable(pa))
3366 					dump_add_page(pa);
3367 			}
3368 		} else {
3369 			if (is_dumpable(pa))
3370 				dump_add_page(pa);
3371 		}
3372 	}
3373 	PMAP_UNLOCK(kernel_pmap);
3374 
3375 	return (sizeof(struct lpte) * moea64_pteg_count * 8);
3376 }
3377 
3378 static struct dump_context dump_ctx;
3379 
3380 static void *
3381 moea64_dump_pmap_init(unsigned blkpgs)
3382 {
3383 	dump_ctx.ptex = 0;
3384 	dump_ctx.ptex_end = moea64_pteg_count * 8;
3385 	dump_ctx.blksz = blkpgs * PAGE_SIZE;
3386 	return (&dump_ctx);
3387 }
3388 
3389 #else
3390 
3391 static size_t
3392 moea64_scan_pmap()
3393 {
3394 	return (0);
3395 }
3396 
3397 static void *
3398 moea64_dump_pmap_init(unsigned blkpgs)
3399 {
3400 	return (NULL);
3401 }
3402 
3403 #endif
3404 
3405 #ifdef __powerpc64__
3406 static void
3407 moea64_map_range(vm_offset_t va, vm_paddr_t pa, vm_size_t npages)
3408 {
3409 
3410 	for (; npages > 0; --npages) {
3411 		if (moea64_large_page_size != 0 &&
3412 		    (pa & moea64_large_page_mask) == 0 &&
3413 		    (va & moea64_large_page_mask) == 0 &&
3414 		    npages >= (moea64_large_page_size >> PAGE_SHIFT)) {
3415 			PMAP_LOCK(kernel_pmap);
3416 			moea64_kenter_large(va, pa, 0, 0);
3417 			PMAP_UNLOCK(kernel_pmap);
3418 			pa += moea64_large_page_size;
3419 			va += moea64_large_page_size;
3420 			npages -= (moea64_large_page_size >> PAGE_SHIFT) - 1;
3421 		} else {
3422 			moea64_kenter(va, pa);
3423 			pa += PAGE_SIZE;
3424 			va += PAGE_SIZE;
3425 		}
3426 	}
3427 }
3428 
3429 static void
3430 moea64_page_array_startup(long pages)
3431 {
3432 	long dom_pages[MAXMEMDOM];
3433 	vm_paddr_t pa;
3434 	vm_offset_t va, vm_page_base;
3435 	vm_size_t needed, size;
3436 	long page;
3437 	int domain;
3438 	int i;
3439 
3440 	vm_page_base = 0xd000000000000000ULL;
3441 
3442 	/* Short-circuit single-domain systems. */
3443 	if (vm_ndomains == 1) {
3444 		size = round_page(pages * sizeof(struct vm_page));
3445 		pa = vm_phys_early_alloc(0, size);
3446 		vm_page_base = moea64_map(&vm_page_base,
3447 		    pa, pa + size, VM_PROT_READ | VM_PROT_WRITE);
3448 		vm_page_array_size = pages;
3449 		vm_page_array = (vm_page_t)vm_page_base;
3450 		return;
3451 	}
3452 
3453 	page = 0;
3454 	for (i = 0; i < MAXMEMDOM; i++)
3455 		dom_pages[i] = 0;
3456 
3457 	/* Now get the number of pages required per domain. */
3458 	for (i = 0; i < vm_phys_nsegs; i++) {
3459 		domain = vm_phys_segs[i].domain;
3460 		KASSERT(domain < MAXMEMDOM,
3461 		    ("Invalid vm_phys_segs NUMA domain %d!\n", domain));
3462 		/* Get size of vm_page_array needed for this segment. */
3463 		size = btoc(vm_phys_segs[i].end - vm_phys_segs[i].start);
3464 		dom_pages[domain] += size;
3465 	}
3466 
3467 	for (i = 0; phys_avail[i + 1] != 0; i+= 2) {
3468 		domain = vm_phys_domain(phys_avail[i]);
3469 		KASSERT(domain < MAXMEMDOM,
3470 		    ("Invalid phys_avail NUMA domain %d!\n", domain));
3471 		size = btoc(phys_avail[i + 1] - phys_avail[i]);
3472 		dom_pages[domain] += size;
3473 	}
3474 
3475 	/*
3476 	 * Map in chunks that can get us all 16MB pages.  There will be some
3477 	 * overlap between domains, but that's acceptable for now.
3478 	 */
3479 	vm_page_array_size = 0;
3480 	va = vm_page_base;
3481 	for (i = 0; i < MAXMEMDOM && vm_page_array_size < pages; i++) {
3482 		if (dom_pages[i] == 0)
3483 			continue;
3484 		size = ulmin(pages - vm_page_array_size, dom_pages[i]);
3485 		size = round_page(size * sizeof(struct vm_page));
3486 		needed = size;
3487 		size = roundup2(size, moea64_large_page_size);
3488 		pa = vm_phys_early_alloc(i, size);
3489 		vm_page_array_size += size / sizeof(struct vm_page);
3490 		moea64_map_range(va, pa, size >> PAGE_SHIFT);
3491 		/* Scoot up domain 0, to reduce the domain page overlap. */
3492 		if (i == 0)
3493 			vm_page_base += size - needed;
3494 		va += size;
3495 	}
3496 	vm_page_array = (vm_page_t)vm_page_base;
3497 	vm_page_array_size = pages;
3498 }
3499 #endif
3500 
3501 static int64_t
3502 moea64_null_method(void)
3503 {
3504 	return (0);
3505 }
3506 
3507 static int64_t moea64_pte_replace_default(struct pvo_entry *pvo, int flags)
3508 {
3509 	int64_t refchg;
3510 
3511 	refchg = moea64_pte_unset(pvo);
3512 	moea64_pte_insert(pvo);
3513 
3514 	return (refchg);
3515 }
3516 
3517 struct moea64_funcs *moea64_ops;
3518 
3519 #define DEFINE_OEA64_IFUNC(ret, func, args, def)		\
3520 	DEFINE_IFUNC(, ret, moea64_##func, args) {		\
3521 		moea64_##func##_t f;				\
3522 		if (moea64_ops == NULL)				\
3523 			return ((moea64_##func##_t)def);	\
3524 		f = moea64_ops->func;				\
3525 		return (f != NULL ? f : (moea64_##func##_t)def);\
3526 	}
3527 
3528 void
3529 moea64_install(void)
3530 {
3531 #ifdef __powerpc64__
3532 	if (hw_direct_map == -1) {
3533 		moea64_probe_large_page();
3534 
3535 		/* Use a direct map if we have large page support */
3536 		if (moea64_large_page_size > 0)
3537 			hw_direct_map = 1;
3538 		else
3539 			hw_direct_map = 0;
3540 	}
3541 #endif
3542 
3543 	/*
3544 	 * Default to non-DMAP, and switch over to DMAP functions once we know
3545 	 * we have DMAP.
3546 	 */
3547 	if (hw_direct_map) {
3548 		moea64_methods.quick_enter_page = moea64_quick_enter_page_dmap;
3549 		moea64_methods.quick_remove_page = NULL;
3550 		moea64_methods.copy_page = moea64_copy_page_dmap;
3551 		moea64_methods.zero_page = moea64_zero_page_dmap;
3552 		moea64_methods.copy_pages = moea64_copy_pages_dmap;
3553 	}
3554 }
3555 
3556 DEFINE_OEA64_IFUNC(int64_t, pte_replace, (struct pvo_entry *, int),
3557     moea64_pte_replace_default)
3558 DEFINE_OEA64_IFUNC(int64_t, pte_insert, (struct pvo_entry *), moea64_null_method)
3559 DEFINE_OEA64_IFUNC(int64_t, pte_unset, (struct pvo_entry *), moea64_null_method)
3560 DEFINE_OEA64_IFUNC(int64_t, pte_clear, (struct pvo_entry *, uint64_t),
3561     moea64_null_method)
3562 DEFINE_OEA64_IFUNC(int64_t, pte_synch, (struct pvo_entry *), moea64_null_method)
3563 DEFINE_OEA64_IFUNC(int64_t, pte_insert_sp, (struct pvo_entry *), moea64_null_method)
3564 DEFINE_OEA64_IFUNC(int64_t, pte_unset_sp, (struct pvo_entry *), moea64_null_method)
3565 DEFINE_OEA64_IFUNC(int64_t, pte_replace_sp, (struct pvo_entry *), moea64_null_method)
3566 
3567 /* Superpage functions */
3568 
3569 /* MMU interface */
3570 
3571 static bool
3572 moea64_ps_enabled(pmap_t pmap)
3573 {
3574 	return (superpages_enabled);
3575 }
3576 
3577 static void
3578 moea64_align_superpage(vm_object_t object, vm_ooffset_t offset,
3579     vm_offset_t *addr, vm_size_t size)
3580 {
3581 	vm_offset_t sp_offset;
3582 
3583 	if (size < HPT_SP_SIZE)
3584 		return;
3585 
3586 	CTR4(KTR_PMAP, "%s: offs=%#jx, addr=%p, size=%#jx",
3587 	    __func__, (uintmax_t)offset, addr, (uintmax_t)size);
3588 
3589 	if (object != NULL && (object->flags & OBJ_COLORED) != 0)
3590 		offset += ptoa(object->pg_color);
3591 	sp_offset = offset & HPT_SP_MASK;
3592 	if (size - ((HPT_SP_SIZE - sp_offset) & HPT_SP_MASK) < HPT_SP_SIZE ||
3593 	    (*addr & HPT_SP_MASK) == sp_offset)
3594 		return;
3595 	if ((*addr & HPT_SP_MASK) < sp_offset)
3596 		*addr = (*addr & ~HPT_SP_MASK) + sp_offset;
3597 	else
3598 		*addr = ((*addr + HPT_SP_MASK) & ~HPT_SP_MASK) + sp_offset;
3599 }
3600 
3601 /* Helpers */
3602 
3603 static __inline void
3604 moea64_pvo_cleanup(struct pvo_dlist *tofree)
3605 {
3606 	struct pvo_entry *pvo;
3607 
3608 	/* clean up */
3609 	while (!SLIST_EMPTY(tofree)) {
3610 		pvo = SLIST_FIRST(tofree);
3611 		SLIST_REMOVE_HEAD(tofree, pvo_dlink);
3612 		if (pvo->pvo_vaddr & PVO_DEAD)
3613 			moea64_pvo_remove_from_page(pvo);
3614 		free_pvo_entry(pvo);
3615 	}
3616 }
3617 
3618 static __inline uint16_t
3619 pvo_to_vmpage_flags(struct pvo_entry *pvo)
3620 {
3621 	uint16_t flags;
3622 
3623 	flags = 0;
3624 	if ((pvo->pvo_pte.prot & VM_PROT_WRITE) != 0)
3625 		flags |= PGA_WRITEABLE;
3626 	if ((pvo->pvo_pte.prot & VM_PROT_EXECUTE) != 0)
3627 		flags |= PGA_EXECUTABLE;
3628 
3629 	return (flags);
3630 }
3631 
3632 /*
3633  * Check if the given pvo and its superpage are in sva-eva range.
3634  */
3635 static __inline bool
3636 moea64_sp_pvo_in_range(struct pvo_entry *pvo, vm_offset_t sva, vm_offset_t eva)
3637 {
3638 	vm_offset_t spva;
3639 
3640 	spva = PVO_VADDR(pvo) & ~HPT_SP_MASK;
3641 	if (spva >= sva && spva + HPT_SP_SIZE <= eva) {
3642 		/*
3643 		 * Because this function is intended to be called from loops
3644 		 * that iterate over ordered pvo entries, if the condition
3645 		 * above is true then the pvo must be the first of its
3646 		 * superpage.
3647 		 */
3648 		KASSERT(PVO_VADDR(pvo) == spva,
3649 		    ("%s: unexpected unaligned superpage pvo", __func__));
3650 		return (true);
3651 	}
3652 	return (false);
3653 }
3654 
3655 /*
3656  * Update vm about the REF/CHG bits if the superpage is managed and
3657  * has (or had) write access.
3658  */
3659 static void
3660 moea64_sp_refchg_process(struct pvo_entry *sp, vm_page_t m,
3661     int64_t sp_refchg, vm_prot_t prot)
3662 {
3663 	vm_page_t m_end;
3664 	int64_t refchg;
3665 
3666 	if ((sp->pvo_vaddr & PVO_MANAGED) != 0 && (prot & VM_PROT_WRITE) != 0) {
3667 		for (m_end = &m[HPT_SP_PAGES]; m < m_end; m++) {
3668 			refchg = sp_refchg |
3669 			    atomic_readandclear_32(&m->md.mdpg_attrs);
3670 			if (refchg & LPTE_CHG)
3671 				vm_page_dirty(m);
3672 			if (refchg & LPTE_REF)
3673 				vm_page_aflag_set(m, PGA_REFERENCED);
3674 		}
3675 	}
3676 }
3677 
3678 /* Superpage ops */
3679 
3680 static int
3681 moea64_sp_enter(pmap_t pmap, vm_offset_t va, vm_page_t m,
3682     vm_prot_t prot, u_int flags, int8_t psind)
3683 {
3684 	struct pvo_entry *pvo, **pvos;
3685 	struct pvo_head *pvo_head;
3686 	vm_offset_t sva;
3687 	vm_page_t sm;
3688 	vm_paddr_t pa, spa;
3689 	bool sync;
3690 	struct pvo_dlist tofree;
3691 	int error, i;
3692 	uint16_t aflags;
3693 
3694 	KASSERT((va & HPT_SP_MASK) == 0, ("%s: va %#jx unaligned",
3695 	    __func__, (uintmax_t)va));
3696 	KASSERT(psind == 1, ("%s: invalid psind: %d", __func__, psind));
3697 	KASSERT(m->psind == 1, ("%s: invalid m->psind: %d",
3698 	    __func__, m->psind));
3699 	KASSERT(pmap != kernel_pmap,
3700 	    ("%s: function called with kernel pmap", __func__));
3701 
3702 	CTR5(KTR_PMAP, "%s: va=%#jx, pa=%#jx, prot=%#x, flags=%#x, psind=1",
3703 	    __func__, (uintmax_t)va, (uintmax_t)VM_PAGE_TO_PHYS(m),
3704 	    prot, flags);
3705 
3706 	SLIST_INIT(&tofree);
3707 
3708 	sva = va;
3709 	sm = m;
3710 	spa = pa = VM_PAGE_TO_PHYS(sm);
3711 
3712 	/* Try to allocate all PVOs first, to make failure handling easier. */
3713 	pvos = malloc(HPT_SP_PAGES * sizeof(struct pvo_entry *), M_TEMP,
3714 	    M_NOWAIT);
3715 	if (pvos == NULL) {
3716 		CTR1(KTR_PMAP, "%s: failed to alloc pvo array", __func__);
3717 		return (KERN_RESOURCE_SHORTAGE);
3718 	}
3719 
3720 	for (i = 0; i < HPT_SP_PAGES; i++) {
3721 		pvos[i] = alloc_pvo_entry(0);
3722 		if (pvos[i] == NULL) {
3723 			CTR1(KTR_PMAP, "%s: failed to alloc pvo", __func__);
3724 			for (i = i - 1; i >= 0; i--)
3725 				free_pvo_entry(pvos[i]);
3726 			free(pvos, M_TEMP);
3727 			return (KERN_RESOURCE_SHORTAGE);
3728 		}
3729 	}
3730 
3731 	SP_PV_LOCK_ALIGNED(spa);
3732 	PMAP_LOCK(pmap);
3733 
3734 	/* Note: moea64_remove_locked() also clears cached REF/CHG bits. */
3735 	moea64_remove_locked(pmap, va, va + HPT_SP_SIZE, &tofree);
3736 
3737 	/* Enter pages */
3738 	for (i = 0; i < HPT_SP_PAGES;
3739 	    i++, va += PAGE_SIZE, pa += PAGE_SIZE, m++) {
3740 		pvo = pvos[i];
3741 
3742 		pvo->pvo_pte.prot = prot;
3743 		pvo->pvo_pte.pa = (pa & ~HPT_SP_MASK) | LPTE_LP_4K_16M |
3744 		    moea64_calc_wimg(pa, pmap_page_get_memattr(m));
3745 
3746 		if ((flags & PMAP_ENTER_WIRED) != 0)
3747 			pvo->pvo_vaddr |= PVO_WIRED;
3748 		pvo->pvo_vaddr |= PVO_LARGE;
3749 
3750 		if ((m->oflags & VPO_UNMANAGED) != 0)
3751 			pvo_head = NULL;
3752 		else {
3753 			pvo_head = &m->md.mdpg_pvoh;
3754 			pvo->pvo_vaddr |= PVO_MANAGED;
3755 		}
3756 
3757 		init_pvo_entry(pvo, pmap, va);
3758 
3759 		error = moea64_pvo_enter(pvo, pvo_head, NULL);
3760 		/*
3761 		 * All superpage PVOs were previously removed, so no errors
3762 		 * should occur while inserting the new ones.
3763 		 */
3764 		KASSERT(error == 0, ("%s: unexpected error "
3765 			    "when inserting superpage PVO: %d",
3766 			    __func__, error));
3767 	}
3768 
3769 	PMAP_UNLOCK(pmap);
3770 	SP_PV_UNLOCK_ALIGNED(spa);
3771 
3772 	sync = (sm->a.flags & PGA_EXECUTABLE) == 0;
3773 	/* Note: moea64_pvo_cleanup() also clears page prot. flags. */
3774 	moea64_pvo_cleanup(&tofree);
3775 	pvo = pvos[0];
3776 
3777 	/* Set vm page flags */
3778 	aflags = pvo_to_vmpage_flags(pvo);
3779 	if (aflags != 0)
3780 		for (m = sm; m < &sm[HPT_SP_PAGES]; m++)
3781 			vm_page_aflag_set(m, aflags);
3782 
3783 	/*
3784 	 * Flush the page from the instruction cache if this page is
3785 	 * mapped executable and cacheable.
3786 	 */
3787 	if (sync && (pvo->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0)
3788 		moea64_syncicache(pmap, sva, spa, HPT_SP_SIZE);
3789 
3790 	atomic_add_long(&sp_mappings, 1);
3791 	CTR3(KTR_PMAP, "%s: SP success for va %#jx in pmap %p",
3792 	    __func__, (uintmax_t)sva, pmap);
3793 
3794 	free(pvos, M_TEMP);
3795 	return (KERN_SUCCESS);
3796 }
3797 
3798 static void
3799 moea64_sp_promote(pmap_t pmap, vm_offset_t va, vm_page_t m)
3800 {
3801 	struct pvo_entry *first, *pvo;
3802 	vm_paddr_t pa, pa_end;
3803 	vm_offset_t sva, va_end;
3804 	int64_t sp_refchg;
3805 
3806 	/* This CTR may generate a lot of output. */
3807 	/* CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)va); */
3808 
3809 	va &= ~HPT_SP_MASK;
3810 	sva = va;
3811 	/* Get superpage */
3812 	pa = VM_PAGE_TO_PHYS(m) & ~HPT_SP_MASK;
3813 	m = PHYS_TO_VM_PAGE(pa);
3814 
3815 	PMAP_LOCK(pmap);
3816 
3817 	/*
3818 	 * Check if all pages meet promotion criteria.
3819 	 *
3820 	 * XXX In some cases the loop below may be executed for each or most
3821 	 * of the entered pages of a superpage, which can be expensive
3822 	 * (although it was not profiled) and need some optimization.
3823 	 *
3824 	 * Some cases where this seems to happen are:
3825 	 * - When a superpage is first entered read-only and later becomes
3826 	 *   read-write.
3827 	 * - When some of the superpage's virtual addresses map to previously
3828 	 *   wired/cached pages while others map to pages allocated from a
3829 	 *   different physical address range. A common scenario where this
3830 	 *   happens is when mmap'ing a file that is already present in FS
3831 	 *   block cache and doesn't fill a superpage.
3832 	 */
3833 	first = pvo = moea64_pvo_find_va(pmap, sva);
3834 	for (pa_end = pa + HPT_SP_SIZE;
3835 	    pa < pa_end; pa += PAGE_SIZE, va += PAGE_SIZE) {
3836 		if (pvo == NULL || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
3837 			CTR3(KTR_PMAP,
3838 			    "%s: NULL or dead PVO: pmap=%p, va=%#jx",
3839 			    __func__, pmap, (uintmax_t)va);
3840 			goto error;
3841 		}
3842 		if (PVO_PADDR(pvo) != pa) {
3843 			CTR5(KTR_PMAP, "%s: PAs don't match: "
3844 			    "pmap=%p, va=%#jx, pvo_pa=%#jx, exp_pa=%#jx",
3845 			    __func__, pmap, (uintmax_t)va,
3846 			    (uintmax_t)PVO_PADDR(pvo), (uintmax_t)pa);
3847 			atomic_add_long(&sp_p_fail_pa, 1);
3848 			goto error;
3849 		}
3850 		if ((first->pvo_vaddr & PVO_FLAGS_PROMOTE) !=
3851 		    (pvo->pvo_vaddr & PVO_FLAGS_PROMOTE)) {
3852 			CTR5(KTR_PMAP, "%s: PVO flags don't match: "
3853 			    "pmap=%p, va=%#jx, pvo_flags=%#jx, exp_flags=%#jx",
3854 			    __func__, pmap, (uintmax_t)va,
3855 			    (uintmax_t)(pvo->pvo_vaddr & PVO_FLAGS_PROMOTE),
3856 			    (uintmax_t)(first->pvo_vaddr & PVO_FLAGS_PROMOTE));
3857 			atomic_add_long(&sp_p_fail_flags, 1);
3858 			goto error;
3859 		}
3860 		if (first->pvo_pte.prot != pvo->pvo_pte.prot) {
3861 			CTR5(KTR_PMAP, "%s: PVO protections don't match: "
3862 			    "pmap=%p, va=%#jx, pvo_prot=%#x, exp_prot=%#x",
3863 			    __func__, pmap, (uintmax_t)va,
3864 			    pvo->pvo_pte.prot, first->pvo_pte.prot);
3865 			atomic_add_long(&sp_p_fail_prot, 1);
3866 			goto error;
3867 		}
3868 		if ((first->pvo_pte.pa & LPTE_WIMG) !=
3869 		    (pvo->pvo_pte.pa & LPTE_WIMG)) {
3870 			CTR5(KTR_PMAP, "%s: WIMG bits don't match: "
3871 			    "pmap=%p, va=%#jx, pvo_wimg=%#jx, exp_wimg=%#jx",
3872 			    __func__, pmap, (uintmax_t)va,
3873 			    (uintmax_t)(pvo->pvo_pte.pa & LPTE_WIMG),
3874 			    (uintmax_t)(first->pvo_pte.pa & LPTE_WIMG));
3875 			atomic_add_long(&sp_p_fail_wimg, 1);
3876 			goto error;
3877 		}
3878 
3879 		pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo);
3880 	}
3881 
3882 	/* All OK, promote. */
3883 
3884 	/*
3885 	 * Handle superpage REF/CHG bits. If REF or CHG is set in
3886 	 * any page, then it must be set in the superpage.
3887 	 *
3888 	 * Instead of querying each page, we take advantage of two facts:
3889 	 * 1- If a page is being promoted, it was referenced.
3890 	 * 2- If promoted pages are writable, they were modified.
3891 	 */
3892 	sp_refchg = LPTE_REF |
3893 	    ((first->pvo_pte.prot & VM_PROT_WRITE) != 0 ? LPTE_CHG : 0);
3894 
3895 	/* Promote pages */
3896 
3897 	for (pvo = first, va_end = PVO_VADDR(pvo) + HPT_SP_SIZE;
3898 	    pvo != NULL && PVO_VADDR(pvo) < va_end;
3899 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo)) {
3900 		pvo->pvo_pte.pa &= ADDR_POFF | ~HPT_SP_MASK;
3901 		pvo->pvo_pte.pa |= LPTE_LP_4K_16M;
3902 		pvo->pvo_vaddr |= PVO_LARGE;
3903 	}
3904 	moea64_pte_replace_sp(first);
3905 
3906 	/* Send REF/CHG bits to VM */
3907 	moea64_sp_refchg_process(first, m, sp_refchg, first->pvo_pte.prot);
3908 
3909 	/* Use first page to cache REF/CHG bits */
3910 	atomic_set_32(&m->md.mdpg_attrs, sp_refchg | MDPG_ATTR_SP);
3911 
3912 	PMAP_UNLOCK(pmap);
3913 
3914 	atomic_add_long(&sp_mappings, 1);
3915 	atomic_add_long(&sp_promotions, 1);
3916 	CTR3(KTR_PMAP, "%s: success for va %#jx in pmap %p",
3917 	    __func__, (uintmax_t)sva, pmap);
3918 	return;
3919 
3920 error:
3921 	atomic_add_long(&sp_p_failures, 1);
3922 	PMAP_UNLOCK(pmap);
3923 }
3924 
3925 static void
3926 moea64_sp_demote_aligned(struct pvo_entry *sp)
3927 {
3928 	struct pvo_entry *pvo;
3929 	vm_offset_t va, va_end;
3930 	vm_paddr_t pa;
3931 	vm_page_t m;
3932 	pmap_t pmap;
3933 	int64_t refchg;
3934 
3935 	CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)PVO_VADDR(sp));
3936 
3937 	pmap = sp->pvo_pmap;
3938 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3939 
3940 	pvo = sp;
3941 
3942 	/* Demote pages */
3943 
3944 	va = PVO_VADDR(pvo);
3945 	pa = PVO_PADDR(pvo);
3946 	m = PHYS_TO_VM_PAGE(pa);
3947 
3948 	for (pvo = sp, va_end = va + HPT_SP_SIZE;
3949 	    pvo != NULL && PVO_VADDR(pvo) < va_end;
3950 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo),
3951 	    va += PAGE_SIZE, pa += PAGE_SIZE) {
3952 		KASSERT(pvo && PVO_VADDR(pvo) == va,
3953 		    ("%s: missing PVO for va %#jx", __func__, (uintmax_t)va));
3954 
3955 		pvo->pvo_vaddr &= ~PVO_LARGE;
3956 		pvo->pvo_pte.pa &= ~LPTE_RPGN;
3957 		pvo->pvo_pte.pa |= pa;
3958 
3959 	}
3960 	refchg = moea64_pte_replace_sp(sp);
3961 
3962 	/*
3963 	 * Clear SP flag
3964 	 *
3965 	 * XXX It is possible that another pmap has this page mapped as
3966 	 *     part of a superpage, but as the SP flag is used only for
3967 	 *     caching SP REF/CHG bits, that will be queried if not set
3968 	 *     in cache, it should be ok to clear it here.
3969 	 */
3970 	atomic_clear_32(&m->md.mdpg_attrs, MDPG_ATTR_SP);
3971 
3972 	/*
3973 	 * Handle superpage REF/CHG bits. A bit set in the superpage
3974 	 * means all pages should consider it set.
3975 	 */
3976 	moea64_sp_refchg_process(sp, m, refchg, sp->pvo_pte.prot);
3977 
3978 	atomic_add_long(&sp_demotions, 1);
3979 	CTR3(KTR_PMAP, "%s: success for va %#jx in pmap %p",
3980 	    __func__, (uintmax_t)PVO_VADDR(sp), pmap);
3981 }
3982 
3983 static void
3984 moea64_sp_demote(struct pvo_entry *pvo)
3985 {
3986 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
3987 
3988 	if ((PVO_VADDR(pvo) & HPT_SP_MASK) != 0) {
3989 		pvo = moea64_pvo_find_va(pvo->pvo_pmap,
3990 		    PVO_VADDR(pvo) & ~HPT_SP_MASK);
3991 		KASSERT(pvo != NULL, ("%s: missing PVO for va %#jx",
3992 		     __func__, (uintmax_t)(PVO_VADDR(pvo) & ~HPT_SP_MASK)));
3993 	}
3994 	moea64_sp_demote_aligned(pvo);
3995 }
3996 
3997 static struct pvo_entry *
3998 moea64_sp_unwire(struct pvo_entry *sp)
3999 {
4000 	struct pvo_entry *pvo, *prev;
4001 	vm_offset_t eva;
4002 	pmap_t pm;
4003 	int64_t ret, refchg;
4004 
4005 	CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)PVO_VADDR(sp));
4006 
4007 	pm = sp->pvo_pmap;
4008 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
4009 
4010 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4011 	refchg = 0;
4012 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4013 	    prev = pvo, pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
4014 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
4015 			panic("%s: pvo %p is missing PVO_WIRED",
4016 			    __func__, pvo);
4017 		pvo->pvo_vaddr &= ~PVO_WIRED;
4018 
4019 		ret = moea64_pte_replace(pvo, 0 /* No invalidation */);
4020 		if (ret < 0)
4021 			refchg |= LPTE_CHG;
4022 		else
4023 			refchg |= ret;
4024 
4025 		pm->pm_stats.wired_count--;
4026 	}
4027 
4028 	/* Send REF/CHG bits to VM */
4029 	moea64_sp_refchg_process(sp, PHYS_TO_VM_PAGE(PVO_PADDR(sp)),
4030 	    refchg, sp->pvo_pte.prot);
4031 
4032 	return (prev);
4033 }
4034 
4035 static struct pvo_entry *
4036 moea64_sp_protect(struct pvo_entry *sp, vm_prot_t prot)
4037 {
4038 	struct pvo_entry *pvo, *prev;
4039 	vm_offset_t eva;
4040 	pmap_t pm;
4041 	vm_page_t m, m_end;
4042 	int64_t ret, refchg;
4043 	vm_prot_t oldprot;
4044 
4045 	CTR3(KTR_PMAP, "%s: va=%#jx, prot=%x",
4046 	    __func__, (uintmax_t)PVO_VADDR(sp), prot);
4047 
4048 	pm = sp->pvo_pmap;
4049 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
4050 
4051 	oldprot = sp->pvo_pte.prot;
4052 	m = PHYS_TO_VM_PAGE(PVO_PADDR(sp));
4053 	KASSERT(m != NULL, ("%s: missing vm page for pa %#jx",
4054 	    __func__, (uintmax_t)PVO_PADDR(sp)));
4055 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4056 	refchg = 0;
4057 
4058 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4059 	    prev = pvo, pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
4060 		pvo->pvo_pte.prot = prot;
4061 		/*
4062 		 * If the PVO is in the page table, update mapping
4063 		 */
4064 		ret = moea64_pte_replace(pvo, MOEA64_PTE_PROT_UPDATE);
4065 		if (ret < 0)
4066 			refchg |= LPTE_CHG;
4067 		else
4068 			refchg |= ret;
4069 	}
4070 
4071 	/* Send REF/CHG bits to VM */
4072 	moea64_sp_refchg_process(sp, m, refchg, oldprot);
4073 
4074 	/* Handle pages that became executable */
4075 	if ((m->a.flags & PGA_EXECUTABLE) == 0 &&
4076 	    (sp->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
4077 		if ((m->oflags & VPO_UNMANAGED) == 0)
4078 			for (m_end = &m[HPT_SP_PAGES]; m < m_end; m++)
4079 				vm_page_aflag_set(m, PGA_EXECUTABLE);
4080 		moea64_syncicache(pm, PVO_VADDR(sp), PVO_PADDR(sp),
4081 		    HPT_SP_SIZE);
4082 	}
4083 
4084 	return (prev);
4085 }
4086 
4087 static struct pvo_entry *
4088 moea64_sp_remove(struct pvo_entry *sp, struct pvo_dlist *tofree)
4089 {
4090 	struct pvo_entry *pvo, *tpvo;
4091 	vm_offset_t eva;
4092 	pmap_t pm;
4093 
4094 	CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)PVO_VADDR(sp));
4095 
4096 	pm = sp->pvo_pmap;
4097 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
4098 
4099 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4100 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
4101 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
4102 
4103 		/*
4104 		 * For locking reasons, remove this from the page table and
4105 		 * pmap, but save delinking from the vm_page for a second
4106 		 * pass
4107 		 */
4108 		moea64_pvo_remove_from_pmap(pvo);
4109 		SLIST_INSERT_HEAD(tofree, pvo, pvo_dlink);
4110 	}
4111 
4112 	/*
4113 	 * Clear SP bit
4114 	 *
4115 	 * XXX See comment in moea64_sp_demote_aligned() for why it's
4116 	 *     ok to always clear the SP bit on remove/demote.
4117 	 */
4118 	atomic_clear_32(&PHYS_TO_VM_PAGE(PVO_PADDR(sp))->md.mdpg_attrs,
4119 	    MDPG_ATTR_SP);
4120 
4121 	return (tpvo);
4122 }
4123 
4124 static int64_t
4125 moea64_sp_query_locked(struct pvo_entry *pvo, uint64_t ptebit)
4126 {
4127 	int64_t refchg, ret;
4128 	vm_offset_t eva;
4129 	vm_page_t m;
4130 	pmap_t pmap;
4131 	struct pvo_entry *sp;
4132 
4133 	pmap = pvo->pvo_pmap;
4134 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4135 
4136 	/* Get first SP PVO */
4137 	if ((PVO_VADDR(pvo) & HPT_SP_MASK) != 0) {
4138 		sp = moea64_pvo_find_va(pmap, PVO_VADDR(pvo) & ~HPT_SP_MASK);
4139 		KASSERT(sp != NULL, ("%s: missing PVO for va %#jx",
4140 		     __func__, (uintmax_t)(PVO_VADDR(pvo) & ~HPT_SP_MASK)));
4141 	} else
4142 		sp = pvo;
4143 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4144 
4145 	refchg = 0;
4146 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4147 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo)) {
4148 		ret = moea64_pte_synch(pvo);
4149 		if (ret > 0) {
4150 			refchg |= ret & (LPTE_CHG | LPTE_REF);
4151 			if ((refchg & ptebit) != 0)
4152 				break;
4153 		}
4154 	}
4155 
4156 	/* Save results */
4157 	if (refchg != 0) {
4158 		m = PHYS_TO_VM_PAGE(PVO_PADDR(sp));
4159 		atomic_set_32(&m->md.mdpg_attrs, refchg | MDPG_ATTR_SP);
4160 	}
4161 
4162 	return (refchg);
4163 }
4164 
4165 static int64_t
4166 moea64_sp_query(struct pvo_entry *pvo, uint64_t ptebit)
4167 {
4168 	int64_t refchg;
4169 	pmap_t pmap;
4170 
4171 	pmap = pvo->pvo_pmap;
4172 	PMAP_LOCK(pmap);
4173 
4174 	/*
4175 	 * Check if SP was demoted/removed before pmap lock was acquired.
4176 	 */
4177 	if (!PVO_IS_SP(pvo) || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
4178 		CTR2(KTR_PMAP, "%s: demoted/removed: pa=%#jx",
4179 		    __func__, (uintmax_t)PVO_PADDR(pvo));
4180 		PMAP_UNLOCK(pmap);
4181 		return (-1);
4182 	}
4183 
4184 	refchg = moea64_sp_query_locked(pvo, ptebit);
4185 	PMAP_UNLOCK(pmap);
4186 
4187 	CTR4(KTR_PMAP, "%s: va=%#jx, pa=%#jx: refchg=%#jx",
4188 	    __func__, (uintmax_t)PVO_VADDR(pvo),
4189 	    (uintmax_t)PVO_PADDR(pvo), (uintmax_t)refchg);
4190 
4191 	return (refchg);
4192 }
4193 
4194 static int64_t
4195 moea64_sp_pvo_clear(struct pvo_entry *pvo, uint64_t ptebit)
4196 {
4197 	int64_t refchg, ret;
4198 	pmap_t pmap;
4199 	struct pvo_entry *sp;
4200 	vm_offset_t eva;
4201 	vm_page_t m;
4202 
4203 	pmap = pvo->pvo_pmap;
4204 	PMAP_LOCK(pmap);
4205 
4206 	/*
4207 	 * Check if SP was demoted/removed before pmap lock was acquired.
4208 	 */
4209 	if (!PVO_IS_SP(pvo) || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
4210 		CTR2(KTR_PMAP, "%s: demoted/removed: pa=%#jx",
4211 		    __func__, (uintmax_t)PVO_PADDR(pvo));
4212 		PMAP_UNLOCK(pmap);
4213 		return (-1);
4214 	}
4215 
4216 	/* Get first SP PVO */
4217 	if ((PVO_VADDR(pvo) & HPT_SP_MASK) != 0) {
4218 		sp = moea64_pvo_find_va(pmap, PVO_VADDR(pvo) & ~HPT_SP_MASK);
4219 		KASSERT(sp != NULL, ("%s: missing PVO for va %#jx",
4220 		     __func__, (uintmax_t)(PVO_VADDR(pvo) & ~HPT_SP_MASK)));
4221 	} else
4222 		sp = pvo;
4223 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4224 
4225 	refchg = 0;
4226 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4227 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo)) {
4228 		ret = moea64_pte_clear(pvo, ptebit);
4229 		if (ret > 0)
4230 			refchg |= ret & (LPTE_CHG | LPTE_REF);
4231 	}
4232 
4233 	m = PHYS_TO_VM_PAGE(PVO_PADDR(sp));
4234 	atomic_clear_32(&m->md.mdpg_attrs, ptebit);
4235 	PMAP_UNLOCK(pmap);
4236 
4237 	CTR4(KTR_PMAP, "%s: va=%#jx, pa=%#jx: refchg=%#jx",
4238 	    __func__, (uintmax_t)PVO_VADDR(sp),
4239 	    (uintmax_t)PVO_PADDR(sp), (uintmax_t)refchg);
4240 
4241 	return (refchg);
4242 }
4243 
4244 static int64_t
4245 moea64_sp_clear(struct pvo_entry *pvo, vm_page_t m, uint64_t ptebit)
4246 {
4247 	int64_t count, ret;
4248 	pmap_t pmap;
4249 
4250 	count = 0;
4251 	pmap = pvo->pvo_pmap;
4252 
4253 	/*
4254 	 * Since this reference bit is shared by 4096 4KB pages, it
4255 	 * should not be cleared every time it is tested. Apply a
4256 	 * simple "hash" function on the physical page number, the
4257 	 * virtual superpage number, and the pmap address to select
4258 	 * one 4KB page out of the 4096 on which testing the
4259 	 * reference bit will result in clearing that reference bit.
4260 	 * This function is designed to avoid the selection of the
4261 	 * same 4KB page for every 16MB page mapping.
4262 	 *
4263 	 * Always leave the reference bit of a wired mapping set, as
4264 	 * the current state of its reference bit won't affect page
4265 	 * replacement.
4266 	 */
4267 	if (ptebit == LPTE_REF && (((VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) ^
4268 	    (PVO_VADDR(pvo) >> HPT_SP_SHIFT) ^ (uintptr_t)pmap) &
4269 	    (HPT_SP_PAGES - 1)) == 0 && (pvo->pvo_vaddr & PVO_WIRED) == 0) {
4270 		if ((ret = moea64_sp_pvo_clear(pvo, ptebit)) == -1)
4271 			return (-1);
4272 
4273 		if ((ret & ptebit) != 0)
4274 			count++;
4275 
4276 	/*
4277 	 * If this page was not selected by the hash function, then assume
4278 	 * its REF bit was set.
4279 	 */
4280 	} else if (ptebit == LPTE_REF) {
4281 		count++;
4282 
4283 	/*
4284 	 * To clear the CHG bit of a single SP page, first it must be demoted.
4285 	 * But if no CHG bit is set, no bit clear and thus no SP demotion is
4286 	 * needed.
4287 	 */
4288 	} else {
4289 		CTR4(KTR_PMAP, "%s: ptebit=%#jx, va=%#jx, pa=%#jx",
4290 		    __func__, (uintmax_t)ptebit, (uintmax_t)PVO_VADDR(pvo),
4291 		    (uintmax_t)PVO_PADDR(pvo));
4292 
4293 		PMAP_LOCK(pmap);
4294 
4295 		/*
4296 		 * Make sure SP wasn't demoted/removed before pmap lock
4297 		 * was acquired.
4298 		 */
4299 		if (!PVO_IS_SP(pvo) || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
4300 			CTR2(KTR_PMAP, "%s: demoted/removed: pa=%#jx",
4301 			    __func__, (uintmax_t)PVO_PADDR(pvo));
4302 			PMAP_UNLOCK(pmap);
4303 			return (-1);
4304 		}
4305 
4306 		ret = moea64_sp_query_locked(pvo, ptebit);
4307 		if ((ret & ptebit) != 0)
4308 			count++;
4309 		else {
4310 			PMAP_UNLOCK(pmap);
4311 			return (0);
4312 		}
4313 
4314 		moea64_sp_demote(pvo);
4315 		moea64_pte_clear(pvo, ptebit);
4316 
4317 		/*
4318 		 * Write protect the mapping to a single page so that a
4319 		 * subsequent write access may repromote.
4320 		 */
4321 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
4322 			moea64_pvo_protect(pmap, pvo,
4323 			    pvo->pvo_pte.prot & ~VM_PROT_WRITE);
4324 
4325 		PMAP_UNLOCK(pmap);
4326 	}
4327 
4328 	return (count);
4329 }
4330