xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision ac099daf6742ead81ea7ea86351a8ef4e783041b)
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_pages(vm_page_t *ma, vm_offset_t a_offset,
390     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
391 int moea64_enter(pmap_t, vm_offset_t, vm_page_t, vm_prot_t,
392     u_int flags, int8_t psind);
393 void moea64_enter_object(pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
394     vm_prot_t);
395 void moea64_enter_quick(pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
396 vm_paddr_t moea64_extract(pmap_t, vm_offset_t);
397 vm_page_t moea64_extract_and_hold(pmap_t, vm_offset_t, vm_prot_t);
398 void moea64_init(void);
399 boolean_t moea64_is_modified(vm_page_t);
400 boolean_t moea64_is_prefaultable(pmap_t, vm_offset_t);
401 boolean_t moea64_is_referenced(vm_page_t);
402 int moea64_ts_referenced(vm_page_t);
403 vm_offset_t moea64_map(vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
404 boolean_t moea64_page_exists_quick(pmap_t, vm_page_t);
405 void moea64_page_init(vm_page_t);
406 int moea64_page_wired_mappings(vm_page_t);
407 int moea64_pinit(pmap_t);
408 void moea64_pinit0(pmap_t);
409 void moea64_protect(pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
410 void moea64_qenter(vm_offset_t, vm_page_t *, int);
411 void moea64_qremove(vm_offset_t, int);
412 void moea64_release(pmap_t);
413 void moea64_remove(pmap_t, vm_offset_t, vm_offset_t);
414 void moea64_remove_pages(pmap_t);
415 void moea64_remove_all(vm_page_t);
416 void moea64_remove_write(vm_page_t);
417 void moea64_unwire(pmap_t, vm_offset_t, vm_offset_t);
418 void moea64_zero_page(vm_page_t);
419 void moea64_zero_page_area(vm_page_t, int, int);
420 void moea64_activate(struct thread *);
421 void moea64_deactivate(struct thread *);
422 void *moea64_mapdev(vm_paddr_t, vm_size_t);
423 void *moea64_mapdev_attr(vm_paddr_t, vm_size_t, vm_memattr_t);
424 void moea64_unmapdev(vm_offset_t, vm_size_t);
425 vm_paddr_t moea64_kextract(vm_offset_t);
426 void moea64_page_set_memattr(vm_page_t m, vm_memattr_t ma);
427 void moea64_kenter_attr(vm_offset_t, vm_paddr_t, vm_memattr_t ma);
428 void moea64_kenter(vm_offset_t, vm_paddr_t);
429 boolean_t moea64_dev_direct_mapped(vm_paddr_t, vm_size_t);
430 static void moea64_sync_icache(pmap_t, vm_offset_t, vm_size_t);
431 void moea64_dumpsys_map(vm_paddr_t pa, size_t sz,
432     void **va);
433 void moea64_scan_init(void);
434 vm_offset_t moea64_quick_enter_page(vm_page_t m);
435 void moea64_quick_remove_page(vm_offset_t addr);
436 boolean_t moea64_page_is_mapped(vm_page_t m);
437 static int moea64_map_user_ptr(pmap_t pm,
438     volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen);
439 static int moea64_decode_kernel_ptr(vm_offset_t addr,
440     int *is_user, vm_offset_t *decoded_addr);
441 static size_t moea64_scan_pmap(void);
442 static void *moea64_dump_pmap_init(unsigned blkpgs);
443 #ifdef __powerpc64__
444 static void moea64_page_array_startup(long);
445 #endif
446 static int moea64_mincore(pmap_t, vm_offset_t, vm_paddr_t *);
447 
448 static struct pmap_funcs moea64_methods = {
449 	.clear_modify = moea64_clear_modify,
450 	.copy_page = moea64_copy_page,
451 	.copy_pages = moea64_copy_pages,
452 	.enter = moea64_enter,
453 	.enter_object = moea64_enter_object,
454 	.enter_quick = moea64_enter_quick,
455 	.extract = moea64_extract,
456 	.extract_and_hold = moea64_extract_and_hold,
457 	.init = moea64_init,
458 	.is_modified = moea64_is_modified,
459 	.is_prefaultable = moea64_is_prefaultable,
460 	.is_referenced = moea64_is_referenced,
461 	.ts_referenced = moea64_ts_referenced,
462 	.map =      		moea64_map,
463 	.mincore = moea64_mincore,
464 	.page_exists_quick = moea64_page_exists_quick,
465 	.page_init = moea64_page_init,
466 	.page_wired_mappings = moea64_page_wired_mappings,
467 	.pinit = moea64_pinit,
468 	.pinit0 = moea64_pinit0,
469 	.protect = moea64_protect,
470 	.qenter = moea64_qenter,
471 	.qremove = moea64_qremove,
472 	.release = moea64_release,
473 	.remove = moea64_remove,
474 	.remove_pages = moea64_remove_pages,
475 	.remove_all =       	moea64_remove_all,
476 	.remove_write = moea64_remove_write,
477 	.sync_icache = moea64_sync_icache,
478 	.unwire = moea64_unwire,
479 	.zero_page =        	moea64_zero_page,
480 	.zero_page_area = moea64_zero_page_area,
481 	.activate = moea64_activate,
482 	.deactivate =       	moea64_deactivate,
483 	.page_set_memattr = moea64_page_set_memattr,
484 	.quick_enter_page =  moea64_quick_enter_page,
485 	.quick_remove_page =  moea64_quick_remove_page,
486 	.page_is_mapped = moea64_page_is_mapped,
487 #ifdef __powerpc64__
488 	.page_array_startup = moea64_page_array_startup,
489 #endif
490 	.ps_enabled = moea64_ps_enabled,
491 	.align_superpage = moea64_align_superpage,
492 
493 	/* Internal interfaces */
494 	.mapdev = moea64_mapdev,
495 	.mapdev_attr = moea64_mapdev_attr,
496 	.unmapdev = moea64_unmapdev,
497 	.kextract = moea64_kextract,
498 	.kenter = moea64_kenter,
499 	.kenter_attr = moea64_kenter_attr,
500 	.dev_direct_mapped = moea64_dev_direct_mapped,
501 	.dumpsys_pa_init = moea64_scan_init,
502 	.dumpsys_scan_pmap = moea64_scan_pmap,
503 	.dumpsys_dump_pmap_init =    moea64_dump_pmap_init,
504 	.dumpsys_map_chunk = moea64_dumpsys_map,
505 	.map_user_ptr = moea64_map_user_ptr,
506 	.decode_kernel_ptr =  moea64_decode_kernel_ptr,
507 };
508 
509 MMU_DEF(oea64_mmu, "mmu_oea64_base", moea64_methods);
510 
511 /*
512  * Get physical address from PVO.
513  *
514  * For superpages, the lower bits are not stored on pvo_pte.pa and must be
515  * obtained from VA.
516  */
517 static __always_inline vm_paddr_t
518 moea64_pvo_paddr(struct pvo_entry *pvo)
519 {
520 	vm_paddr_t pa;
521 
522 	pa = (pvo)->pvo_pte.pa & LPTE_RPGN;
523 
524 	if (PVO_IS_SP(pvo)) {
525 		pa &= ~HPT_SP_MASK; /* This is needed to clear LPTE_LP bits. */
526 		pa |= PVO_VADDR(pvo) & HPT_SP_MASK;
527 	}
528 	return (pa);
529 }
530 
531 static struct pvo_head *
532 vm_page_to_pvoh(vm_page_t m)
533 {
534 
535 	mtx_assert(PV_LOCKPTR(VM_PAGE_TO_PHYS(m)), MA_OWNED);
536 	return (&m->md.mdpg_pvoh);
537 }
538 
539 static struct pvo_entry *
540 alloc_pvo_entry(int bootstrap)
541 {
542 	struct pvo_entry *pvo;
543 
544 	if (!moea64_initialized || bootstrap) {
545 		if (moea64_bpvo_pool_index >= moea64_bpvo_pool_size) {
546 			panic("%s: bpvo pool exhausted, index=%d, size=%d, bytes=%zd."
547 			    "Try setting machdep.moea64_bpvo_pool_size tunable",
548 			    __func__, moea64_bpvo_pool_index,
549 			    moea64_bpvo_pool_size,
550 			    moea64_bpvo_pool_size * sizeof(struct pvo_entry));
551 		}
552 		pvo = &moea64_bpvo_pool[
553 		    atomic_fetchadd_int(&moea64_bpvo_pool_index, 1)];
554 		bzero(pvo, sizeof(*pvo));
555 		pvo->pvo_vaddr = PVO_BOOTSTRAP;
556 	} else
557 		pvo = uma_zalloc(moea64_pvo_zone, M_NOWAIT | M_ZERO);
558 
559 	return (pvo);
560 }
561 
562 static void
563 init_pvo_entry(struct pvo_entry *pvo, pmap_t pmap, vm_offset_t va)
564 {
565 	uint64_t vsid;
566 	uint64_t hash;
567 	int shift;
568 
569 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
570 
571 	pvo->pvo_pmap = pmap;
572 	va &= ~ADDR_POFF;
573 	pvo->pvo_vaddr |= va;
574 	vsid = va_to_vsid(pmap, va);
575 	pvo->pvo_vpn = (uint64_t)((va & ADDR_PIDX) >> ADDR_PIDX_SHFT)
576 	    | (vsid << 16);
577 
578 	if (pmap == kernel_pmap && (pvo->pvo_vaddr & PVO_LARGE) != 0)
579 		shift = moea64_large_page_shift;
580 	else
581 		shift = ADDR_PIDX_SHFT;
582 	hash = (vsid & VSID_HASH_MASK) ^ (((uint64_t)va & ADDR_PIDX) >> shift);
583 	pvo->pvo_pte.slot = (hash & moea64_pteg_mask) << 3;
584 }
585 
586 static void
587 free_pvo_entry(struct pvo_entry *pvo)
588 {
589 
590 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
591 		uma_zfree(moea64_pvo_zone, pvo);
592 }
593 
594 void
595 moea64_pte_from_pvo(const struct pvo_entry *pvo, struct lpte *lpte)
596 {
597 
598 	lpte->pte_hi = moea64_pte_vpn_from_pvo_vpn(pvo);
599 	lpte->pte_hi |= LPTE_VALID;
600 
601 	if (pvo->pvo_vaddr & PVO_LARGE)
602 		lpte->pte_hi |= LPTE_BIG;
603 	if (pvo->pvo_vaddr & PVO_WIRED)
604 		lpte->pte_hi |= LPTE_WIRED;
605 	if (pvo->pvo_vaddr & PVO_HID)
606 		lpte->pte_hi |= LPTE_HID;
607 
608 	lpte->pte_lo = pvo->pvo_pte.pa; /* Includes WIMG bits */
609 	if (pvo->pvo_pte.prot & VM_PROT_WRITE)
610 		lpte->pte_lo |= LPTE_BW;
611 	else
612 		lpte->pte_lo |= LPTE_BR;
613 
614 	if (!(pvo->pvo_pte.prot & VM_PROT_EXECUTE))
615 		lpte->pte_lo |= LPTE_NOEXEC;
616 }
617 
618 static __inline uint64_t
619 moea64_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
620 {
621 	uint64_t pte_lo;
622 	int i;
623 
624 	if (ma != VM_MEMATTR_DEFAULT) {
625 		switch (ma) {
626 		case VM_MEMATTR_UNCACHEABLE:
627 			return (LPTE_I | LPTE_G);
628 		case VM_MEMATTR_CACHEABLE:
629 			return (LPTE_M);
630 		case VM_MEMATTR_WRITE_COMBINING:
631 		case VM_MEMATTR_WRITE_BACK:
632 		case VM_MEMATTR_PREFETCHABLE:
633 			return (LPTE_I);
634 		case VM_MEMATTR_WRITE_THROUGH:
635 			return (LPTE_W | LPTE_M);
636 		}
637 	}
638 
639 	/*
640 	 * Assume the page is cache inhibited and access is guarded unless
641 	 * it's in our available memory array.
642 	 */
643 	pte_lo = LPTE_I | LPTE_G;
644 	for (i = 0; i < pregions_sz; i++) {
645 		if ((pa >= pregions[i].mr_start) &&
646 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
647 			pte_lo &= ~(LPTE_I | LPTE_G);
648 			pte_lo |= LPTE_M;
649 			break;
650 		}
651 	}
652 
653 	return pte_lo;
654 }
655 
656 /*
657  * Quick sort callout for comparing memory regions.
658  */
659 static int	om_cmp(const void *a, const void *b);
660 
661 static int
662 om_cmp(const void *a, const void *b)
663 {
664 	const struct	ofw_map *mapa;
665 	const struct	ofw_map *mapb;
666 
667 	mapa = a;
668 	mapb = b;
669 	if (mapa->om_pa < mapb->om_pa)
670 		return (-1);
671 	else if (mapa->om_pa > mapb->om_pa)
672 		return (1);
673 	else
674 		return (0);
675 }
676 
677 static void
678 moea64_add_ofw_mappings(phandle_t mmu, size_t sz)
679 {
680 	struct ofw_map	translations[sz/(4*sizeof(cell_t))]; /*>= 4 cells per */
681 	pcell_t		acells, trans_cells[sz/sizeof(cell_t)];
682 	struct pvo_entry *pvo;
683 	register_t	msr;
684 	vm_offset_t	off;
685 	vm_paddr_t	pa_base;
686 	int		i, j;
687 
688 	bzero(translations, sz);
689 	OF_getencprop(OF_finddevice("/"), "#address-cells", &acells,
690 	    sizeof(acells));
691 	if (OF_getencprop(mmu, "translations", trans_cells, sz) == -1)
692 		panic("moea64_bootstrap: can't get ofw translations");
693 
694 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
695 	sz /= sizeof(cell_t);
696 	for (i = 0, j = 0; i < sz; j++) {
697 		translations[j].om_va = trans_cells[i++];
698 		translations[j].om_len = trans_cells[i++];
699 		translations[j].om_pa = trans_cells[i++];
700 		if (acells == 2) {
701 			translations[j].om_pa <<= 32;
702 			translations[j].om_pa |= trans_cells[i++];
703 		}
704 		translations[j].om_mode = trans_cells[i++];
705 	}
706 	KASSERT(i == sz, ("Translations map has incorrect cell count (%d/%zd)",
707 	    i, sz));
708 
709 	sz = j;
710 	qsort(translations, sz, sizeof (*translations), om_cmp);
711 
712 	for (i = 0; i < sz; i++) {
713 		pa_base = translations[i].om_pa;
714 	      #ifndef __powerpc64__
715 		if ((translations[i].om_pa >> 32) != 0)
716 			panic("OFW translations above 32-bit boundary!");
717 	      #endif
718 
719 		if (pa_base % PAGE_SIZE)
720 			panic("OFW translation not page-aligned (phys)!");
721 		if (translations[i].om_va % PAGE_SIZE)
722 			panic("OFW translation not page-aligned (virt)!");
723 
724 		CTR3(KTR_PMAP, "translation: pa=%#zx va=%#x len=%#x",
725 		    pa_base, translations[i].om_va, translations[i].om_len);
726 
727 		/* Now enter the pages for this mapping */
728 
729 		DISABLE_TRANS(msr);
730 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
731 			/* If this address is direct-mapped, skip remapping */
732 			if (hw_direct_map &&
733 			    translations[i].om_va == PHYS_TO_DMAP(pa_base) &&
734 			    moea64_calc_wimg(pa_base + off, VM_MEMATTR_DEFAULT)
735  			    == LPTE_M)
736 				continue;
737 
738 			PMAP_LOCK(kernel_pmap);
739 			pvo = moea64_pvo_find_va(kernel_pmap,
740 			    translations[i].om_va + off);
741 			PMAP_UNLOCK(kernel_pmap);
742 			if (pvo != NULL)
743 				continue;
744 
745 			moea64_kenter(translations[i].om_va + off,
746 			    pa_base + off);
747 		}
748 		ENABLE_TRANS(msr);
749 	}
750 }
751 
752 #ifdef __powerpc64__
753 static void
754 moea64_probe_large_page(void)
755 {
756 	uint16_t pvr = mfpvr() >> 16;
757 
758 	switch (pvr) {
759 	case IBM970:
760 	case IBM970FX:
761 	case IBM970MP:
762 		powerpc_sync(); isync();
763 		mtspr(SPR_HID4, mfspr(SPR_HID4) & ~HID4_970_DISABLE_LG_PG);
764 		powerpc_sync(); isync();
765 
766 		/* FALLTHROUGH */
767 	default:
768 		if (moea64_large_page_size == 0) {
769 			moea64_large_page_size = 0x1000000; /* 16 MB */
770 			moea64_large_page_shift = 24;
771 		}
772 	}
773 
774 	moea64_large_page_mask = moea64_large_page_size - 1;
775 }
776 
777 static void
778 moea64_bootstrap_slb_prefault(vm_offset_t va, int large)
779 {
780 	struct slb *cache;
781 	struct slb entry;
782 	uint64_t esid, slbe;
783 	uint64_t i;
784 
785 	cache = PCPU_GET(aim.slb);
786 	esid = va >> ADDR_SR_SHFT;
787 	slbe = (esid << SLBE_ESID_SHIFT) | SLBE_VALID;
788 
789 	for (i = 0; i < 64; i++) {
790 		if (cache[i].slbe == (slbe | i))
791 			return;
792 	}
793 
794 	entry.slbe = slbe;
795 	entry.slbv = KERNEL_VSID(esid) << SLBV_VSID_SHIFT;
796 	if (large)
797 		entry.slbv |= SLBV_L;
798 
799 	slb_insert_kernel(entry.slbe, entry.slbv);
800 }
801 #endif
802 
803 static int
804 moea64_kenter_large(vm_offset_t va, vm_paddr_t pa, uint64_t attr, int bootstrap)
805 {
806 	struct pvo_entry *pvo;
807 	uint64_t pte_lo;
808 	int error;
809 
810 	pte_lo = LPTE_M;
811 	pte_lo |= attr;
812 
813 	pvo = alloc_pvo_entry(bootstrap);
814 	pvo->pvo_vaddr |= PVO_WIRED | PVO_LARGE;
815 	init_pvo_entry(pvo, kernel_pmap, va);
816 
817 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE |
818 	    VM_PROT_EXECUTE;
819 	pvo->pvo_pte.pa = pa | pte_lo;
820 	error = moea64_pvo_enter(pvo, NULL, NULL);
821 	if (error != 0)
822 		panic("Error %d inserting large page\n", error);
823 	return (0);
824 }
825 
826 static void
827 moea64_setup_direct_map(vm_offset_t kernelstart,
828     vm_offset_t kernelend)
829 {
830 	register_t msr;
831 	vm_paddr_t pa, pkernelstart, pkernelend;
832 	vm_offset_t size, off;
833 	uint64_t pte_lo;
834 	int i;
835 
836 	if (moea64_large_page_size == 0)
837 		hw_direct_map = 0;
838 
839 	DISABLE_TRANS(msr);
840 	if (hw_direct_map) {
841 		PMAP_LOCK(kernel_pmap);
842 		for (i = 0; i < pregions_sz; i++) {
843 		  for (pa = pregions[i].mr_start; pa < pregions[i].mr_start +
844 		     pregions[i].mr_size; pa += moea64_large_page_size) {
845 			pte_lo = LPTE_M;
846 			if (pa & moea64_large_page_mask) {
847 				pa &= moea64_large_page_mask;
848 				pte_lo |= LPTE_G;
849 			}
850 			if (pa + moea64_large_page_size >
851 			    pregions[i].mr_start + pregions[i].mr_size)
852 				pte_lo |= LPTE_G;
853 
854 			moea64_kenter_large(PHYS_TO_DMAP(pa), pa, pte_lo, 1);
855 		  }
856 		}
857 		PMAP_UNLOCK(kernel_pmap);
858 	}
859 
860 	/*
861 	 * Make sure the kernel and BPVO pool stay mapped on systems either
862 	 * without a direct map or on which the kernel is not already executing
863 	 * out of the direct-mapped region.
864 	 */
865 	if (kernelstart < DMAP_BASE_ADDRESS) {
866 		/*
867 		 * For pre-dmap execution, we need to use identity mapping
868 		 * because we will be operating with the mmu on but in the
869 		 * wrong address configuration until we __restartkernel().
870 		 */
871 		for (pa = kernelstart & ~PAGE_MASK; pa < kernelend;
872 		    pa += PAGE_SIZE)
873 			moea64_kenter(pa, pa);
874 	} else if (!hw_direct_map) {
875 		pkernelstart = kernelstart & ~DMAP_BASE_ADDRESS;
876 		pkernelend = kernelend & ~DMAP_BASE_ADDRESS;
877 		for (pa = pkernelstart & ~PAGE_MASK; pa < pkernelend;
878 		    pa += PAGE_SIZE)
879 			moea64_kenter(pa | DMAP_BASE_ADDRESS, pa);
880 	}
881 
882 	if (!hw_direct_map) {
883 		size = moea64_bpvo_pool_size*sizeof(struct pvo_entry);
884 		off = (vm_offset_t)(moea64_bpvo_pool);
885 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
886 			moea64_kenter(pa, pa);
887 
888 		/* Map exception vectors */
889 		for (pa = EXC_RSVD; pa < EXC_LAST; pa += PAGE_SIZE)
890 			moea64_kenter(pa | DMAP_BASE_ADDRESS, pa);
891 	}
892 	ENABLE_TRANS(msr);
893 
894 	/*
895 	 * Allow user to override unmapped_buf_allowed for testing.
896 	 * XXXKIB Only direct map implementation was tested.
897 	 */
898 	if (!TUNABLE_INT_FETCH("vfs.unmapped_buf_allowed",
899 	    &unmapped_buf_allowed))
900 		unmapped_buf_allowed = hw_direct_map;
901 }
902 
903 /* Quick sort callout for comparing physical addresses. */
904 static int
905 pa_cmp(const void *a, const void *b)
906 {
907 	const vm_paddr_t *pa = a, *pb = b;
908 
909 	if (*pa < *pb)
910 		return (-1);
911 	else if (*pa > *pb)
912 		return (1);
913 	else
914 		return (0);
915 }
916 
917 void
918 moea64_early_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
919 {
920 	int		i, j;
921 	vm_size_t	physsz, hwphyssz;
922 	vm_paddr_t	kernelphysstart, kernelphysend;
923 	int		rm_pavail;
924 
925 	/* Level 0 reservations consist of 4096 pages (16MB superpage). */
926 	vm_level_0_order = 12;
927 
928 #ifndef __powerpc64__
929 	/* We don't have a direct map since there is no BAT */
930 	hw_direct_map = 0;
931 
932 	/* Make sure battable is zero, since we have no BAT */
933 	for (i = 0; i < 16; i++) {
934 		battable[i].batu = 0;
935 		battable[i].batl = 0;
936 	}
937 #else
938 	moea64_probe_large_page();
939 
940 	/* Use a direct map if we have large page support */
941 	if (moea64_large_page_size > 0)
942 		hw_direct_map = 1;
943 	else
944 		hw_direct_map = 0;
945 
946 	/* Install trap handlers for SLBs */
947 	bcopy(&slbtrap, (void *)EXC_DSE,(size_t)&slbtrapend - (size_t)&slbtrap);
948 	bcopy(&slbtrap, (void *)EXC_ISE,(size_t)&slbtrapend - (size_t)&slbtrap);
949 	__syncicache((void *)EXC_DSE, 0x80);
950 	__syncicache((void *)EXC_ISE, 0x80);
951 #endif
952 
953 	kernelphysstart = kernelstart & ~DMAP_BASE_ADDRESS;
954 	kernelphysend = kernelend & ~DMAP_BASE_ADDRESS;
955 
956 	/* Get physical memory regions from firmware */
957 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
958 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
959 
960 	if (PHYS_AVAIL_ENTRIES < regions_sz)
961 		panic("moea64_bootstrap: phys_avail too small");
962 
963 	phys_avail_count = 0;
964 	physsz = 0;
965 	hwphyssz = 0;
966 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
967 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
968 		CTR3(KTR_PMAP, "region: %#zx - %#zx (%#zx)",
969 		    regions[i].mr_start, regions[i].mr_start +
970 		    regions[i].mr_size, regions[i].mr_size);
971 		if (hwphyssz != 0 &&
972 		    (physsz + regions[i].mr_size) >= hwphyssz) {
973 			if (physsz < hwphyssz) {
974 				phys_avail[j] = regions[i].mr_start;
975 				phys_avail[j + 1] = regions[i].mr_start +
976 				    hwphyssz - physsz;
977 				physsz = hwphyssz;
978 				phys_avail_count++;
979 				dump_avail[j] = phys_avail[j];
980 				dump_avail[j + 1] = phys_avail[j + 1];
981 			}
982 			break;
983 		}
984 		phys_avail[j] = regions[i].mr_start;
985 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
986 		phys_avail_count++;
987 		physsz += regions[i].mr_size;
988 		dump_avail[j] = phys_avail[j];
989 		dump_avail[j + 1] = phys_avail[j + 1];
990 	}
991 
992 	/* Check for overlap with the kernel and exception vectors */
993 	rm_pavail = 0;
994 	for (j = 0; j < 2*phys_avail_count; j+=2) {
995 		if (phys_avail[j] < EXC_LAST)
996 			phys_avail[j] += EXC_LAST;
997 
998 		if (phys_avail[j] >= kernelphysstart &&
999 		    phys_avail[j+1] <= kernelphysend) {
1000 			phys_avail[j] = phys_avail[j+1] = ~0;
1001 			rm_pavail++;
1002 			continue;
1003 		}
1004 
1005 		if (kernelphysstart >= phys_avail[j] &&
1006 		    kernelphysstart < phys_avail[j+1]) {
1007 			if (kernelphysend < phys_avail[j+1]) {
1008 				phys_avail[2*phys_avail_count] =
1009 				    (kernelphysend & ~PAGE_MASK) + PAGE_SIZE;
1010 				phys_avail[2*phys_avail_count + 1] =
1011 				    phys_avail[j+1];
1012 				phys_avail_count++;
1013 			}
1014 
1015 			phys_avail[j+1] = kernelphysstart & ~PAGE_MASK;
1016 		}
1017 
1018 		if (kernelphysend >= phys_avail[j] &&
1019 		    kernelphysend < phys_avail[j+1]) {
1020 			if (kernelphysstart > phys_avail[j]) {
1021 				phys_avail[2*phys_avail_count] = phys_avail[j];
1022 				phys_avail[2*phys_avail_count + 1] =
1023 				    kernelphysstart & ~PAGE_MASK;
1024 				phys_avail_count++;
1025 			}
1026 
1027 			phys_avail[j] = (kernelphysend & ~PAGE_MASK) +
1028 			    PAGE_SIZE;
1029 		}
1030 	}
1031 
1032 	/* Remove physical available regions marked for removal (~0) */
1033 	if (rm_pavail) {
1034 		qsort(phys_avail, 2*phys_avail_count, sizeof(phys_avail[0]),
1035 			pa_cmp);
1036 		phys_avail_count -= rm_pavail;
1037 		for (i = 2*phys_avail_count;
1038 		     i < 2*(phys_avail_count + rm_pavail); i+=2)
1039 			phys_avail[i] = phys_avail[i+1] = 0;
1040 	}
1041 
1042 	physmem = btoc(physsz);
1043 
1044 #ifdef PTEGCOUNT
1045 	moea64_pteg_count = PTEGCOUNT;
1046 #else
1047 	moea64_pteg_count = 0x1000;
1048 
1049 	while (moea64_pteg_count < physmem)
1050 		moea64_pteg_count <<= 1;
1051 
1052 	moea64_pteg_count >>= 1;
1053 #endif /* PTEGCOUNT */
1054 }
1055 
1056 void
1057 moea64_mid_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
1058 {
1059 	int		i;
1060 
1061 	/*
1062 	 * Set PTEG mask
1063 	 */
1064 	moea64_pteg_mask = moea64_pteg_count - 1;
1065 
1066 	/*
1067 	 * Initialize SLB table lock and page locks
1068 	 */
1069 	mtx_init(&moea64_slb_mutex, "SLB table", NULL, MTX_DEF);
1070 	for (i = 0; i < PV_LOCK_COUNT; i++)
1071 		mtx_init(&pv_lock[i], "page pv", NULL, MTX_DEF);
1072 
1073 	/*
1074 	 * Initialise the bootstrap pvo pool.
1075 	 */
1076 	TUNABLE_INT_FETCH("machdep.moea64_bpvo_pool_size", &moea64_bpvo_pool_size);
1077 	if (moea64_bpvo_pool_size == 0) {
1078 		if (!hw_direct_map)
1079 			moea64_bpvo_pool_size = ((ptoa((uintmax_t)physmem) * sizeof(struct vm_page)) /
1080 			    (PAGE_SIZE * PAGE_SIZE)) * BPVO_POOL_EXPANSION_FACTOR;
1081 		else
1082 			moea64_bpvo_pool_size = BPVO_POOL_SIZE;
1083 	}
1084 
1085 	if (boothowto & RB_VERBOSE) {
1086 		printf("mmu_oea64: bpvo pool entries = %d, bpvo pool size = %zu MB\n",
1087 		    moea64_bpvo_pool_size,
1088 		    moea64_bpvo_pool_size*sizeof(struct pvo_entry) / 1048576);
1089 	}
1090 
1091 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
1092 		moea64_bpvo_pool_size*sizeof(struct pvo_entry), PAGE_SIZE);
1093 	moea64_bpvo_pool_index = 0;
1094 
1095 	/* Place at address usable through the direct map */
1096 	if (hw_direct_map)
1097 		moea64_bpvo_pool = (struct pvo_entry *)
1098 		    PHYS_TO_DMAP((uintptr_t)moea64_bpvo_pool);
1099 
1100 	/*
1101 	 * Make sure kernel vsid is allocated as well as VSID 0.
1102 	 */
1103 	#ifndef __powerpc64__
1104 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NVSIDS - 1)) / VSID_NBPW]
1105 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
1106 	moea64_vsid_bitmap[0] |= 1;
1107 	#endif
1108 
1109 	/*
1110 	 * Initialize the kernel pmap (which is statically allocated).
1111 	 */
1112 	#ifdef __powerpc64__
1113 	for (i = 0; i < 64; i++) {
1114 		pcpup->pc_aim.slb[i].slbv = 0;
1115 		pcpup->pc_aim.slb[i].slbe = 0;
1116 	}
1117 	#else
1118 	for (i = 0; i < 16; i++)
1119 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
1120 	#endif
1121 
1122 	kernel_pmap->pmap_phys = kernel_pmap;
1123 	CPU_FILL(&kernel_pmap->pm_active);
1124 	RB_INIT(&kernel_pmap->pmap_pvo);
1125 
1126 	PMAP_LOCK_INIT(kernel_pmap);
1127 
1128 	/*
1129 	 * Now map in all the other buffers we allocated earlier
1130 	 */
1131 
1132 	moea64_setup_direct_map(kernelstart, kernelend);
1133 }
1134 
1135 void
1136 moea64_late_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
1137 {
1138 	ihandle_t	mmui;
1139 	phandle_t	chosen;
1140 	phandle_t	mmu;
1141 	ssize_t		sz;
1142 	int		i;
1143 	vm_offset_t	pa, va;
1144 	void		*dpcpu;
1145 
1146 	/*
1147 	 * Set up the Open Firmware pmap and add its mappings if not in real
1148 	 * mode.
1149 	 */
1150 
1151 	chosen = OF_finddevice("/chosen");
1152 	if (chosen != -1 && OF_getencprop(chosen, "mmu", &mmui, 4) != -1) {
1153 		mmu = OF_instance_to_package(mmui);
1154 		if (mmu == -1 ||
1155 		    (sz = OF_getproplen(mmu, "translations")) == -1)
1156 			sz = 0;
1157 		if (sz > 6144 /* tmpstksz - 2 KB headroom */)
1158 			panic("moea64_bootstrap: too many ofw translations");
1159 
1160 		if (sz > 0)
1161 			moea64_add_ofw_mappings(mmu, sz);
1162 	}
1163 
1164 	/*
1165 	 * Calculate the last available physical address.
1166 	 */
1167 	Maxmem = 0;
1168 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
1169 		Maxmem = MAX(Maxmem, powerpc_btop(phys_avail[i + 1]));
1170 
1171 	/*
1172 	 * Initialize MMU.
1173 	 */
1174 	pmap_cpu_bootstrap(0);
1175 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
1176 	pmap_bootstrapped++;
1177 
1178 	/*
1179 	 * Set the start and end of kva.
1180 	 */
1181 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
1182 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
1183 
1184 	/*
1185 	 * Map the entire KVA range into the SLB. We must not fault there.
1186 	 */
1187 	#ifdef __powerpc64__
1188 	for (va = virtual_avail; va < virtual_end; va += SEGMENT_LENGTH)
1189 		moea64_bootstrap_slb_prefault(va, 0);
1190 	#endif
1191 
1192 	/*
1193 	 * Remap any early IO mappings (console framebuffer, etc.)
1194 	 */
1195 	bs_remap_earlyboot();
1196 
1197 	/*
1198 	 * Figure out how far we can extend virtual_end into segment 16
1199 	 * without running into existing mappings. Segment 16 is guaranteed
1200 	 * to contain neither RAM nor devices (at least on Apple hardware),
1201 	 * but will generally contain some OFW mappings we should not
1202 	 * step on.
1203 	 */
1204 
1205 	#ifndef __powerpc64__	/* KVA is in high memory on PPC64 */
1206 	PMAP_LOCK(kernel_pmap);
1207 	while (virtual_end < VM_MAX_KERNEL_ADDRESS &&
1208 	    moea64_pvo_find_va(kernel_pmap, virtual_end+1) == NULL)
1209 		virtual_end += PAGE_SIZE;
1210 	PMAP_UNLOCK(kernel_pmap);
1211 	#endif
1212 
1213 	/*
1214 	 * Allocate a kernel stack with a guard page for thread0 and map it
1215 	 * into the kernel page map.
1216 	 */
1217 	pa = moea64_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
1218 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
1219 	virtual_avail = va + kstack_pages * PAGE_SIZE;
1220 	CTR2(KTR_PMAP, "moea64_bootstrap: kstack0 at %#x (%#x)", pa, va);
1221 	thread0.td_kstack = va;
1222 	thread0.td_kstack_pages = kstack_pages;
1223 	for (i = 0; i < kstack_pages; i++) {
1224 		moea64_kenter(va, pa);
1225 		pa += PAGE_SIZE;
1226 		va += PAGE_SIZE;
1227 	}
1228 
1229 	/*
1230 	 * Allocate virtual address space for the message buffer.
1231 	 */
1232 	pa = msgbuf_phys = moea64_bootstrap_alloc(msgbufsize, PAGE_SIZE);
1233 	msgbufp = (struct msgbuf *)virtual_avail;
1234 	va = virtual_avail;
1235 	virtual_avail += round_page(msgbufsize);
1236 	while (va < virtual_avail) {
1237 		moea64_kenter(va, pa);
1238 		pa += PAGE_SIZE;
1239 		va += PAGE_SIZE;
1240 	}
1241 
1242 	/*
1243 	 * Allocate virtual address space for the dynamic percpu area.
1244 	 */
1245 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
1246 	dpcpu = (void *)virtual_avail;
1247 	va = virtual_avail;
1248 	virtual_avail += DPCPU_SIZE;
1249 	while (va < virtual_avail) {
1250 		moea64_kenter(va, pa);
1251 		pa += PAGE_SIZE;
1252 		va += PAGE_SIZE;
1253 	}
1254 	dpcpu_init(dpcpu, curcpu);
1255 
1256 	crashdumpmap = (caddr_t)virtual_avail;
1257 	virtual_avail += MAXDUMPPGS * PAGE_SIZE;
1258 
1259 	/*
1260 	 * Allocate some things for page zeroing. We put this directly
1261 	 * in the page table and use MOEA64_PTE_REPLACE to avoid any
1262 	 * of the PVO book-keeping or other parts of the VM system
1263 	 * from even knowing that this hack exists.
1264 	 */
1265 
1266 	if (!hw_direct_map) {
1267 		mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL,
1268 		    MTX_DEF);
1269 		for (i = 0; i < 2; i++) {
1270 			moea64_scratchpage_va[i] = (virtual_end+1) - PAGE_SIZE;
1271 			virtual_end -= PAGE_SIZE;
1272 
1273 			moea64_kenter(moea64_scratchpage_va[i], 0);
1274 
1275 			PMAP_LOCK(kernel_pmap);
1276 			moea64_scratchpage_pvo[i] = moea64_pvo_find_va(
1277 			    kernel_pmap, (vm_offset_t)moea64_scratchpage_va[i]);
1278 			PMAP_UNLOCK(kernel_pmap);
1279 		}
1280 	}
1281 
1282 	numa_mem_regions(&numa_pregions, &numapregions_sz);
1283 }
1284 
1285 static void
1286 moea64_pmap_init_qpages(void)
1287 {
1288 	struct pcpu *pc;
1289 	int i;
1290 
1291 	if (hw_direct_map)
1292 		return;
1293 
1294 	CPU_FOREACH(i) {
1295 		pc = pcpu_find(i);
1296 		pc->pc_qmap_addr = kva_alloc(PAGE_SIZE);
1297 		if (pc->pc_qmap_addr == 0)
1298 			panic("pmap_init_qpages: unable to allocate KVA");
1299 		PMAP_LOCK(kernel_pmap);
1300 		pc->pc_aim.qmap_pvo =
1301 		    moea64_pvo_find_va(kernel_pmap, pc->pc_qmap_addr);
1302 		PMAP_UNLOCK(kernel_pmap);
1303 		mtx_init(&pc->pc_aim.qmap_lock, "qmap lock", NULL, MTX_DEF);
1304 	}
1305 }
1306 
1307 SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, moea64_pmap_init_qpages, NULL);
1308 
1309 /*
1310  * Activate a user pmap.  This mostly involves setting some non-CPU
1311  * state.
1312  */
1313 void
1314 moea64_activate(struct thread *td)
1315 {
1316 	pmap_t	pm;
1317 
1318 	pm = &td->td_proc->p_vmspace->vm_pmap;
1319 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1320 
1321 	#ifdef __powerpc64__
1322 	PCPU_SET(aim.userslb, pm->pm_slb);
1323 	__asm __volatile("slbmte %0, %1; isync" ::
1324 	    "r"(td->td_pcb->pcb_cpu.aim.usr_vsid), "r"(USER_SLB_SLBE));
1325 	#else
1326 	PCPU_SET(curpmap, pm->pmap_phys);
1327 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1328 	#endif
1329 }
1330 
1331 void
1332 moea64_deactivate(struct thread *td)
1333 {
1334 	pmap_t	pm;
1335 
1336 	__asm __volatile("isync; slbie %0" :: "r"(USER_ADDR));
1337 
1338 	pm = &td->td_proc->p_vmspace->vm_pmap;
1339 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1340 	#ifdef __powerpc64__
1341 	PCPU_SET(aim.userslb, NULL);
1342 	#else
1343 	PCPU_SET(curpmap, NULL);
1344 	#endif
1345 }
1346 
1347 void
1348 moea64_unwire(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1349 {
1350 	struct	pvo_entry key, *pvo;
1351 	vm_page_t m;
1352 	int64_t	refchg;
1353 
1354 	key.pvo_vaddr = sva;
1355 	PMAP_LOCK(pm);
1356 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1357 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1358 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1359 		if (PVO_IS_SP(pvo)) {
1360 			if (moea64_sp_pvo_in_range(pvo, sva, eva)) {
1361 				pvo = moea64_sp_unwire(pvo);
1362 				continue;
1363 			} else {
1364 				CTR1(KTR_PMAP, "%s: demote before unwire",
1365 				    __func__);
1366 				moea64_sp_demote(pvo);
1367 			}
1368 		}
1369 
1370 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1371 			panic("moea64_unwire: pvo %p is missing PVO_WIRED",
1372 			    pvo);
1373 		pvo->pvo_vaddr &= ~PVO_WIRED;
1374 		refchg = moea64_pte_replace(pvo, 0 /* No invalidation */);
1375 		if ((pvo->pvo_vaddr & PVO_MANAGED) &&
1376 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1377 			if (refchg < 0)
1378 				refchg = LPTE_CHG;
1379 			m = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
1380 
1381 			refchg |= atomic_readandclear_32(&m->md.mdpg_attrs);
1382 			if (refchg & LPTE_CHG)
1383 				vm_page_dirty(m);
1384 			if (refchg & LPTE_REF)
1385 				vm_page_aflag_set(m, PGA_REFERENCED);
1386 		}
1387 		pm->pm_stats.wired_count--;
1388 	}
1389 	PMAP_UNLOCK(pm);
1390 }
1391 
1392 static int
1393 moea64_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *pap)
1394 {
1395 	struct pvo_entry *pvo;
1396 	vm_paddr_t pa;
1397 	vm_page_t m;
1398 	int val;
1399 	bool managed;
1400 
1401 	PMAP_LOCK(pmap);
1402 
1403 	pvo = moea64_pvo_find_va(pmap, addr);
1404 	if (pvo != NULL) {
1405 		pa = PVO_PADDR(pvo);
1406 		m = PHYS_TO_VM_PAGE(pa);
1407 		managed = (pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED;
1408 		if (PVO_IS_SP(pvo))
1409 			val = MINCORE_INCORE | MINCORE_PSIND(1);
1410 		else
1411 			val = MINCORE_INCORE;
1412 	} else {
1413 		PMAP_UNLOCK(pmap);
1414 		return (0);
1415 	}
1416 
1417 	PMAP_UNLOCK(pmap);
1418 
1419 	if (m == NULL)
1420 		return (0);
1421 
1422 	if (managed) {
1423 		if (moea64_is_modified(m))
1424 			val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
1425 
1426 		if (moea64_is_referenced(m))
1427 			val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
1428 	}
1429 
1430 	if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) !=
1431 	    (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) &&
1432 	    managed) {
1433 		*pap = pa;
1434 	}
1435 
1436 	return (val);
1437 }
1438 
1439 /*
1440  * This goes through and sets the physical address of our
1441  * special scratch PTE to the PA we want to zero or copy. Because
1442  * of locking issues (this can get called in pvo_enter() by
1443  * the UMA allocator), we can't use most other utility functions here
1444  */
1445 
1446 static __inline
1447 void moea64_set_scratchpage_pa(int which, vm_paddr_t pa)
1448 {
1449 	struct pvo_entry *pvo;
1450 
1451 	KASSERT(!hw_direct_map, ("Using OEA64 scratchpage with a direct map!"));
1452 	mtx_assert(&moea64_scratchpage_mtx, MA_OWNED);
1453 
1454 	pvo = moea64_scratchpage_pvo[which];
1455 	PMAP_LOCK(pvo->pvo_pmap);
1456 	pvo->pvo_pte.pa =
1457 	    moea64_calc_wimg(pa, VM_MEMATTR_DEFAULT) | (uint64_t)pa;
1458 	moea64_pte_replace(pvo, MOEA64_PTE_INVALIDATE);
1459 	PMAP_UNLOCK(pvo->pvo_pmap);
1460 	isync();
1461 }
1462 
1463 void
1464 moea64_copy_page(vm_page_t msrc, vm_page_t mdst)
1465 {
1466 	vm_offset_t	dst;
1467 	vm_offset_t	src;
1468 
1469 	dst = VM_PAGE_TO_PHYS(mdst);
1470 	src = VM_PAGE_TO_PHYS(msrc);
1471 
1472 	if (hw_direct_map) {
1473 		bcopy((void *)PHYS_TO_DMAP(src), (void *)PHYS_TO_DMAP(dst),
1474 		    PAGE_SIZE);
1475 	} else {
1476 		mtx_lock(&moea64_scratchpage_mtx);
1477 
1478 		moea64_set_scratchpage_pa(0, src);
1479 		moea64_set_scratchpage_pa(1, dst);
1480 
1481 		bcopy((void *)moea64_scratchpage_va[0],
1482 		    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1483 
1484 		mtx_unlock(&moea64_scratchpage_mtx);
1485 	}
1486 }
1487 
1488 static inline void
1489 moea64_copy_pages_dmap(vm_page_t *ma, vm_offset_t a_offset,
1490     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1491 {
1492 	void *a_cp, *b_cp;
1493 	vm_offset_t a_pg_offset, b_pg_offset;
1494 	int cnt;
1495 
1496 	while (xfersize > 0) {
1497 		a_pg_offset = a_offset & PAGE_MASK;
1498 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1499 		a_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1500 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT])) +
1501 		    a_pg_offset;
1502 		b_pg_offset = b_offset & PAGE_MASK;
1503 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1504 		b_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1505 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT])) +
1506 		    b_pg_offset;
1507 		bcopy(a_cp, b_cp, cnt);
1508 		a_offset += cnt;
1509 		b_offset += cnt;
1510 		xfersize -= cnt;
1511 	}
1512 }
1513 
1514 static inline void
1515 moea64_copy_pages_nodmap(vm_page_t *ma, vm_offset_t a_offset,
1516     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1517 {
1518 	void *a_cp, *b_cp;
1519 	vm_offset_t a_pg_offset, b_pg_offset;
1520 	int cnt;
1521 
1522 	mtx_lock(&moea64_scratchpage_mtx);
1523 	while (xfersize > 0) {
1524 		a_pg_offset = a_offset & PAGE_MASK;
1525 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1526 		moea64_set_scratchpage_pa(0,
1527 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]));
1528 		a_cp = (char *)moea64_scratchpage_va[0] + a_pg_offset;
1529 		b_pg_offset = b_offset & PAGE_MASK;
1530 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1531 		moea64_set_scratchpage_pa(1,
1532 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]));
1533 		b_cp = (char *)moea64_scratchpage_va[1] + b_pg_offset;
1534 		bcopy(a_cp, b_cp, cnt);
1535 		a_offset += cnt;
1536 		b_offset += cnt;
1537 		xfersize -= cnt;
1538 	}
1539 	mtx_unlock(&moea64_scratchpage_mtx);
1540 }
1541 
1542 void
1543 moea64_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
1544     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1545 {
1546 
1547 	if (hw_direct_map) {
1548 		moea64_copy_pages_dmap(ma, a_offset, mb, b_offset,
1549 		    xfersize);
1550 	} else {
1551 		moea64_copy_pages_nodmap(ma, a_offset, mb, b_offset,
1552 		    xfersize);
1553 	}
1554 }
1555 
1556 void
1557 moea64_zero_page_area(vm_page_t m, int off, int size)
1558 {
1559 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1560 
1561 	if (size + off > PAGE_SIZE)
1562 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1563 
1564 	if (hw_direct_map) {
1565 		bzero((caddr_t)(uintptr_t)PHYS_TO_DMAP(pa) + off, size);
1566 	} else {
1567 		mtx_lock(&moea64_scratchpage_mtx);
1568 		moea64_set_scratchpage_pa(0, pa);
1569 		bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1570 		mtx_unlock(&moea64_scratchpage_mtx);
1571 	}
1572 }
1573 
1574 /*
1575  * Zero a page of physical memory by temporarily mapping it
1576  */
1577 void
1578 moea64_zero_page(vm_page_t m)
1579 {
1580 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1581 	vm_offset_t va, off;
1582 
1583 	if (!hw_direct_map) {
1584 		mtx_lock(&moea64_scratchpage_mtx);
1585 
1586 		moea64_set_scratchpage_pa(0, pa);
1587 		va = moea64_scratchpage_va[0];
1588 	} else {
1589 		va = PHYS_TO_DMAP(pa);
1590 	}
1591 
1592 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1593 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1594 
1595 	if (!hw_direct_map)
1596 		mtx_unlock(&moea64_scratchpage_mtx);
1597 }
1598 
1599 vm_offset_t
1600 moea64_quick_enter_page(vm_page_t m)
1601 {
1602 	struct pvo_entry *pvo;
1603 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1604 
1605 	if (hw_direct_map)
1606 		return (PHYS_TO_DMAP(pa));
1607 
1608 	/*
1609  	 * MOEA64_PTE_REPLACE does some locking, so we can't just grab
1610 	 * a critical section and access the PCPU data like on i386.
1611 	 * Instead, pin the thread and grab the PCPU lock to prevent
1612 	 * a preempting thread from using the same PCPU data.
1613 	 */
1614 	sched_pin();
1615 
1616 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_NOTOWNED);
1617 	pvo = PCPU_GET(aim.qmap_pvo);
1618 
1619 	mtx_lock(PCPU_PTR(aim.qmap_lock));
1620 	pvo->pvo_pte.pa = moea64_calc_wimg(pa, pmap_page_get_memattr(m)) |
1621 	    (uint64_t)pa;
1622 	moea64_pte_replace(pvo, MOEA64_PTE_INVALIDATE);
1623 	isync();
1624 
1625 	return (PCPU_GET(qmap_addr));
1626 }
1627 
1628 void
1629 moea64_quick_remove_page(vm_offset_t addr)
1630 {
1631 	if (hw_direct_map)
1632 		return;
1633 
1634 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_OWNED);
1635 	KASSERT(PCPU_GET(qmap_addr) == addr,
1636 	    ("moea64_quick_remove_page: invalid address"));
1637 	mtx_unlock(PCPU_PTR(aim.qmap_lock));
1638 	sched_unpin();
1639 }
1640 
1641 boolean_t
1642 moea64_page_is_mapped(vm_page_t m)
1643 {
1644 	return (!LIST_EMPTY(&(m)->md.mdpg_pvoh));
1645 }
1646 
1647 /*
1648  * Map the given physical page at the specified virtual address in the
1649  * target pmap with the protection requested.  If specified the page
1650  * will be wired down.
1651  */
1652 
1653 int
1654 moea64_enter(pmap_t pmap, vm_offset_t va, vm_page_t m,
1655     vm_prot_t prot, u_int flags, int8_t psind)
1656 {
1657 	struct		pvo_entry *pvo, *oldpvo, *tpvo;
1658 	struct		pvo_head *pvo_head;
1659 	uint64_t	pte_lo;
1660 	int		error;
1661 	vm_paddr_t	pa;
1662 
1663 	if ((m->oflags & VPO_UNMANAGED) == 0) {
1664 		if ((flags & PMAP_ENTER_QUICK_LOCKED) == 0)
1665 			VM_PAGE_OBJECT_BUSY_ASSERT(m);
1666 		else
1667 			VM_OBJECT_ASSERT_LOCKED(m->object);
1668 	}
1669 
1670 	if (psind > 0)
1671 		return (moea64_sp_enter(pmap, va, m, prot, flags, psind));
1672 
1673 	pvo = alloc_pvo_entry(0);
1674 	if (pvo == NULL)
1675 		return (KERN_RESOURCE_SHORTAGE);
1676 	pvo->pvo_pmap = NULL; /* to be filled in later */
1677 	pvo->pvo_pte.prot = prot;
1678 
1679 	pa = VM_PAGE_TO_PHYS(m);
1680 	pte_lo = moea64_calc_wimg(pa, pmap_page_get_memattr(m));
1681 	pvo->pvo_pte.pa = pa | pte_lo;
1682 
1683 	if ((flags & PMAP_ENTER_WIRED) != 0)
1684 		pvo->pvo_vaddr |= PVO_WIRED;
1685 
1686 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea64_initialized) {
1687 		pvo_head = NULL;
1688 	} else {
1689 		pvo_head = &m->md.mdpg_pvoh;
1690 		pvo->pvo_vaddr |= PVO_MANAGED;
1691 	}
1692 
1693 	PV_LOCK(pa);
1694 	PMAP_LOCK(pmap);
1695 	if (pvo->pvo_pmap == NULL)
1696 		init_pvo_entry(pvo, pmap, va);
1697 
1698 	if (moea64_ps_enabled(pmap) &&
1699 	    (tpvo = moea64_pvo_find_va(pmap, va & ~HPT_SP_MASK)) != NULL &&
1700 	    PVO_IS_SP(tpvo)) {
1701 		/* Demote SP before entering a regular page */
1702 		CTR2(KTR_PMAP, "%s: demote before enter: va=%#jx",
1703 		    __func__, (uintmax_t)va);
1704 		moea64_sp_demote_aligned(tpvo);
1705 	}
1706 
1707 	if (prot & VM_PROT_WRITE)
1708 		if (pmap_bootstrapped &&
1709 		    (m->oflags & VPO_UNMANAGED) == 0)
1710 			vm_page_aflag_set(m, PGA_WRITEABLE);
1711 
1712 	error = moea64_pvo_enter(pvo, pvo_head, &oldpvo);
1713 	if (error == EEXIST) {
1714 		if (oldpvo->pvo_vaddr == pvo->pvo_vaddr &&
1715 		    oldpvo->pvo_pte.pa == pvo->pvo_pte.pa &&
1716 		    oldpvo->pvo_pte.prot == prot) {
1717 			/* Identical mapping already exists */
1718 			error = 0;
1719 
1720 			/* If not in page table, reinsert it */
1721 			if (moea64_pte_synch(oldpvo) < 0) {
1722 				STAT_MOEA64(moea64_pte_overflow--);
1723 				moea64_pte_insert(oldpvo);
1724 			}
1725 
1726 			/* Then just clean up and go home */
1727 			PMAP_UNLOCK(pmap);
1728 			PV_UNLOCK(pa);
1729 			free_pvo_entry(pvo);
1730 			pvo = NULL;
1731 			goto out;
1732 		} else {
1733 			/* Otherwise, need to kill it first */
1734 			KASSERT(oldpvo->pvo_pmap == pmap, ("pmap of old "
1735 			    "mapping does not match new mapping"));
1736 			moea64_pvo_remove_from_pmap(oldpvo);
1737 			moea64_pvo_enter(pvo, pvo_head, NULL);
1738 		}
1739 	}
1740 	PMAP_UNLOCK(pmap);
1741 	PV_UNLOCK(pa);
1742 
1743 	/* Free any dead pages */
1744 	if (error == EEXIST) {
1745 		moea64_pvo_remove_from_page(oldpvo);
1746 		free_pvo_entry(oldpvo);
1747 	}
1748 
1749 out:
1750 	/*
1751 	 * Flush the page from the instruction cache if this page is
1752 	 * mapped executable and cacheable.
1753 	 */
1754 	if (pmap != kernel_pmap && (m->a.flags & PGA_EXECUTABLE) == 0 &&
1755 	    (pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1756 		vm_page_aflag_set(m, PGA_EXECUTABLE);
1757 		moea64_syncicache(pmap, va, pa, PAGE_SIZE);
1758 	}
1759 
1760 #if VM_NRESERVLEVEL > 0
1761 	/*
1762 	 * Try to promote pages.
1763 	 *
1764 	 * If the VA of the entered page is not aligned with its PA,
1765 	 * don't try page promotion as it is not possible.
1766 	 * This reduces the number of promotion failures dramatically.
1767 	 */
1768 	if (moea64_ps_enabled(pmap) && pmap != kernel_pmap && pvo != NULL &&
1769 	    (pvo->pvo_vaddr & PVO_MANAGED) != 0 &&
1770 	    (va & HPT_SP_MASK) == (pa & HPT_SP_MASK) &&
1771 	    (m->flags & PG_FICTITIOUS) == 0 &&
1772 	    vm_reserv_level_iffullpop(m) == 0)
1773 		moea64_sp_promote(pmap, va, m);
1774 #endif
1775 
1776 	return (KERN_SUCCESS);
1777 }
1778 
1779 static void
1780 moea64_syncicache(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
1781     vm_size_t sz)
1782 {
1783 
1784 	/*
1785 	 * This is much trickier than on older systems because
1786 	 * we can't sync the icache on physical addresses directly
1787 	 * without a direct map. Instead we check a couple of cases
1788 	 * where the memory is already mapped in and, failing that,
1789 	 * use the same trick we use for page zeroing to create
1790 	 * a temporary mapping for this physical address.
1791 	 */
1792 
1793 	if (!pmap_bootstrapped) {
1794 		/*
1795 		 * If PMAP is not bootstrapped, we are likely to be
1796 		 * in real mode.
1797 		 */
1798 		__syncicache((void *)(uintptr_t)pa, sz);
1799 	} else if (pmap == kernel_pmap) {
1800 		__syncicache((void *)va, sz);
1801 	} else if (hw_direct_map) {
1802 		__syncicache((void *)(uintptr_t)PHYS_TO_DMAP(pa), sz);
1803 	} else {
1804 		/* Use the scratch page to set up a temp mapping */
1805 
1806 		mtx_lock(&moea64_scratchpage_mtx);
1807 
1808 		moea64_set_scratchpage_pa(1, pa & ~ADDR_POFF);
1809 		__syncicache((void *)(moea64_scratchpage_va[1] +
1810 		    (va & ADDR_POFF)), sz);
1811 
1812 		mtx_unlock(&moea64_scratchpage_mtx);
1813 	}
1814 }
1815 
1816 /*
1817  * Maps a sequence of resident pages belonging to the same object.
1818  * The sequence begins with the given page m_start.  This page is
1819  * mapped at the given virtual address start.  Each subsequent page is
1820  * mapped at a virtual address that is offset from start by the same
1821  * amount as the page is offset from m_start within the object.  The
1822  * last page in the sequence is the page with the largest offset from
1823  * m_start that can be mapped at a virtual address less than the given
1824  * virtual address end.  Not every virtual page between start and end
1825  * is mapped; only those for which a resident page exists with the
1826  * corresponding offset from m_start are mapped.
1827  */
1828 void
1829 moea64_enter_object(pmap_t pm, vm_offset_t start, vm_offset_t end,
1830     vm_page_t m_start, vm_prot_t prot)
1831 {
1832 	vm_page_t m;
1833 	vm_pindex_t diff, psize;
1834 	vm_offset_t va;
1835 	int8_t psind;
1836 
1837 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1838 
1839 	psize = atop(end - start);
1840 	m = m_start;
1841 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1842 		va = start + ptoa(diff);
1843 		if ((va & HPT_SP_MASK) == 0 && va + HPT_SP_SIZE <= end &&
1844 		    m->psind == 1 && moea64_ps_enabled(pm))
1845 			psind = 1;
1846 		else
1847 			psind = 0;
1848 		moea64_enter(pm, va, m, prot &
1849 		    (VM_PROT_READ | VM_PROT_EXECUTE),
1850 		    PMAP_ENTER_NOSLEEP | PMAP_ENTER_QUICK_LOCKED, psind);
1851 		if (psind == 1)
1852 			m = &m[HPT_SP_SIZE / PAGE_SIZE - 1];
1853 		m = TAILQ_NEXT(m, listq);
1854 	}
1855 }
1856 
1857 void
1858 moea64_enter_quick(pmap_t pm, vm_offset_t va, vm_page_t m,
1859     vm_prot_t prot)
1860 {
1861 
1862 	moea64_enter(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1863 	    PMAP_ENTER_NOSLEEP | PMAP_ENTER_QUICK_LOCKED, 0);
1864 }
1865 
1866 vm_paddr_t
1867 moea64_extract(pmap_t pm, vm_offset_t va)
1868 {
1869 	struct	pvo_entry *pvo;
1870 	vm_paddr_t pa;
1871 
1872 	PMAP_LOCK(pm);
1873 	pvo = moea64_pvo_find_va(pm, va);
1874 	if (pvo == NULL)
1875 		pa = 0;
1876 	else
1877 		pa = PVO_PADDR(pvo) | (va - PVO_VADDR(pvo));
1878 	PMAP_UNLOCK(pm);
1879 
1880 	return (pa);
1881 }
1882 
1883 /*
1884  * Atomically extract and hold the physical page with the given
1885  * pmap and virtual address pair if that mapping permits the given
1886  * protection.
1887  */
1888 vm_page_t
1889 moea64_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1890 {
1891 	struct	pvo_entry *pvo;
1892 	vm_page_t m;
1893 
1894 	m = NULL;
1895 	PMAP_LOCK(pmap);
1896 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1897 	if (pvo != NULL && (pvo->pvo_pte.prot & prot) == prot) {
1898 		m = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
1899 		if (!vm_page_wire_mapped(m))
1900 			m = NULL;
1901 	}
1902 	PMAP_UNLOCK(pmap);
1903 	return (m);
1904 }
1905 
1906 static void *
1907 moea64_uma_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain,
1908     uint8_t *flags, int wait)
1909 {
1910 	struct pvo_entry *pvo;
1911         vm_offset_t va;
1912         vm_page_t m;
1913         int needed_lock;
1914 
1915 	/*
1916 	 * This entire routine is a horrible hack to avoid bothering kmem
1917 	 * for new KVA addresses. Because this can get called from inside
1918 	 * kmem allocation routines, calling kmem for a new address here
1919 	 * can lead to multiply locking non-recursive mutexes.
1920 	 */
1921 
1922 	*flags = UMA_SLAB_PRIV;
1923 	needed_lock = !PMAP_LOCKED(kernel_pmap);
1924 
1925 	m = vm_page_alloc_domain(NULL, 0, domain,
1926 	    malloc2vm_flags(wait) | VM_ALLOC_WIRED | VM_ALLOC_NOOBJ);
1927 	if (m == NULL)
1928 		return (NULL);
1929 
1930 	va = VM_PAGE_TO_PHYS(m);
1931 
1932 	pvo = alloc_pvo_entry(1 /* bootstrap */);
1933 
1934 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE;
1935 	pvo->pvo_pte.pa = VM_PAGE_TO_PHYS(m) | LPTE_M;
1936 
1937 	if (needed_lock)
1938 		PMAP_LOCK(kernel_pmap);
1939 
1940 	init_pvo_entry(pvo, kernel_pmap, va);
1941 	pvo->pvo_vaddr |= PVO_WIRED;
1942 
1943 	moea64_pvo_enter(pvo, NULL, NULL);
1944 
1945 	if (needed_lock)
1946 		PMAP_UNLOCK(kernel_pmap);
1947 
1948 	if ((wait & M_ZERO) && (m->flags & PG_ZERO) == 0)
1949                 bzero((void *)va, PAGE_SIZE);
1950 
1951 	return (void *)va;
1952 }
1953 
1954 extern int elf32_nxstack;
1955 
1956 void
1957 moea64_init()
1958 {
1959 
1960 	CTR0(KTR_PMAP, "moea64_init");
1961 
1962 	moea64_pvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1963 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1964 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1965 
1966 	/*
1967 	 * Are large page mappings enabled?
1968 	 *
1969 	 * While HPT superpages are not better tested, leave it disabled by
1970 	 * default.
1971 	 */
1972 	superpages_enabled = 0;
1973 	TUNABLE_INT_FETCH("vm.pmap.superpages_enabled", &superpages_enabled);
1974 	if (superpages_enabled) {
1975 		KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0,
1976 		    ("moea64_init: can't assign to pagesizes[1]"));
1977 
1978 		if (moea64_large_page_size == 0) {
1979 			printf("mmu_oea64: HW does not support large pages. "
1980 					"Disabling superpages...\n");
1981 			superpages_enabled = 0;
1982 		} else if (!moea64_has_lp_4k_16m) {
1983 			printf("mmu_oea64: "
1984 			    "HW does not support mixed 4KB/16MB page sizes. "
1985 			    "Disabling superpages...\n");
1986 			superpages_enabled = 0;
1987 		} else
1988 			pagesizes[1] = HPT_SP_SIZE;
1989 	}
1990 
1991 	if (!hw_direct_map) {
1992 		uma_zone_set_allocf(moea64_pvo_zone, moea64_uma_page_alloc);
1993 	}
1994 
1995 #ifdef COMPAT_FREEBSD32
1996 	elf32_nxstack = 1;
1997 #endif
1998 
1999 	moea64_initialized = TRUE;
2000 }
2001 
2002 boolean_t
2003 moea64_is_referenced(vm_page_t m)
2004 {
2005 
2006 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2007 	    ("moea64_is_referenced: page %p is not managed", m));
2008 
2009 	return (moea64_query_bit(m, LPTE_REF));
2010 }
2011 
2012 boolean_t
2013 moea64_is_modified(vm_page_t m)
2014 {
2015 
2016 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2017 	    ("moea64_is_modified: page %p is not managed", m));
2018 
2019 	/*
2020 	 * If the page is not busied then this check is racy.
2021 	 */
2022 	if (!pmap_page_is_write_mapped(m))
2023 		return (FALSE);
2024 
2025 	return (moea64_query_bit(m, LPTE_CHG));
2026 }
2027 
2028 boolean_t
2029 moea64_is_prefaultable(pmap_t pmap, vm_offset_t va)
2030 {
2031 	struct pvo_entry *pvo;
2032 	boolean_t rv = TRUE;
2033 
2034 	PMAP_LOCK(pmap);
2035 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
2036 	if (pvo != NULL)
2037 		rv = FALSE;
2038 	PMAP_UNLOCK(pmap);
2039 	return (rv);
2040 }
2041 
2042 void
2043 moea64_clear_modify(vm_page_t m)
2044 {
2045 
2046 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2047 	    ("moea64_clear_modify: page %p is not managed", m));
2048 	vm_page_assert_busied(m);
2049 
2050 	if (!pmap_page_is_write_mapped(m))
2051 		return;
2052 	moea64_clear_bit(m, LPTE_CHG);
2053 }
2054 
2055 /*
2056  * Clear the write and modified bits in each of the given page's mappings.
2057  */
2058 void
2059 moea64_remove_write(vm_page_t m)
2060 {
2061 	struct	pvo_entry *pvo;
2062 	int64_t	refchg, ret;
2063 	pmap_t	pmap;
2064 
2065 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2066 	    ("moea64_remove_write: page %p is not managed", m));
2067 	vm_page_assert_busied(m);
2068 
2069 	if (!pmap_page_is_write_mapped(m))
2070 		return;
2071 
2072 	powerpc_sync();
2073 	PV_PAGE_LOCK(m);
2074 	refchg = 0;
2075 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2076 		pmap = pvo->pvo_pmap;
2077 		PMAP_LOCK(pmap);
2078 		if (!(pvo->pvo_vaddr & PVO_DEAD) &&
2079 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2080 			if (PVO_IS_SP(pvo)) {
2081 				CTR1(KTR_PMAP, "%s: demote before remwr",
2082 				    __func__);
2083 				moea64_sp_demote(pvo);
2084 			}
2085 			pvo->pvo_pte.prot &= ~VM_PROT_WRITE;
2086 			ret = moea64_pte_replace(pvo, MOEA64_PTE_PROT_UPDATE);
2087 			if (ret < 0)
2088 				ret = LPTE_CHG;
2089 			refchg |= ret;
2090 			if (pvo->pvo_pmap == kernel_pmap)
2091 				isync();
2092 		}
2093 		PMAP_UNLOCK(pmap);
2094 	}
2095 	if ((refchg | atomic_readandclear_32(&m->md.mdpg_attrs)) & LPTE_CHG)
2096 		vm_page_dirty(m);
2097 	vm_page_aflag_clear(m, PGA_WRITEABLE);
2098 	PV_PAGE_UNLOCK(m);
2099 }
2100 
2101 /*
2102  *	moea64_ts_referenced:
2103  *
2104  *	Return a count of reference bits for a page, clearing those bits.
2105  *	It is not necessary for every reference bit to be cleared, but it
2106  *	is necessary that 0 only be returned when there are truly no
2107  *	reference bits set.
2108  *
2109  *	XXX: The exact number of bits to check and clear is a matter that
2110  *	should be tested and standardized at some point in the future for
2111  *	optimal aging of shared pages.
2112  */
2113 int
2114 moea64_ts_referenced(vm_page_t m)
2115 {
2116 
2117 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2118 	    ("moea64_ts_referenced: page %p is not managed", m));
2119 	return (moea64_clear_bit(m, LPTE_REF));
2120 }
2121 
2122 /*
2123  * Modify the WIMG settings of all mappings for a page.
2124  */
2125 void
2126 moea64_page_set_memattr(vm_page_t m, vm_memattr_t ma)
2127 {
2128 	struct	pvo_entry *pvo;
2129 	int64_t	refchg;
2130 	pmap_t	pmap;
2131 	uint64_t lo;
2132 
2133 	CTR3(KTR_PMAP, "%s: pa=%#jx, ma=%#x",
2134 	    __func__, (uintmax_t)VM_PAGE_TO_PHYS(m), ma);
2135 
2136 	if ((m->oflags & VPO_UNMANAGED) != 0) {
2137 		m->md.mdpg_cache_attrs = ma;
2138 		return;
2139 	}
2140 
2141 	lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
2142 
2143 	PV_PAGE_LOCK(m);
2144 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2145 		pmap = pvo->pvo_pmap;
2146 		PMAP_LOCK(pmap);
2147 		if (!(pvo->pvo_vaddr & PVO_DEAD)) {
2148 			if (PVO_IS_SP(pvo)) {
2149 				CTR1(KTR_PMAP,
2150 				    "%s: demote before set_memattr", __func__);
2151 				moea64_sp_demote(pvo);
2152 			}
2153 			pvo->pvo_pte.pa &= ~LPTE_WIMG;
2154 			pvo->pvo_pte.pa |= lo;
2155 			refchg = moea64_pte_replace(pvo, MOEA64_PTE_INVALIDATE);
2156 			if (refchg < 0)
2157 				refchg = (pvo->pvo_pte.prot & VM_PROT_WRITE) ?
2158 				    LPTE_CHG : 0;
2159 			if ((pvo->pvo_vaddr & PVO_MANAGED) &&
2160 			    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2161 				refchg |=
2162 				    atomic_readandclear_32(&m->md.mdpg_attrs);
2163 				if (refchg & LPTE_CHG)
2164 					vm_page_dirty(m);
2165 				if (refchg & LPTE_REF)
2166 					vm_page_aflag_set(m, PGA_REFERENCED);
2167 			}
2168 			if (pvo->pvo_pmap == kernel_pmap)
2169 				isync();
2170 		}
2171 		PMAP_UNLOCK(pmap);
2172 	}
2173 	m->md.mdpg_cache_attrs = ma;
2174 	PV_PAGE_UNLOCK(m);
2175 }
2176 
2177 /*
2178  * Map a wired page into kernel virtual address space.
2179  */
2180 void
2181 moea64_kenter_attr(vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
2182 {
2183 	int		error;
2184 	struct pvo_entry *pvo, *oldpvo;
2185 
2186 	do {
2187 		pvo = alloc_pvo_entry(0);
2188 		if (pvo == NULL)
2189 			vm_wait(NULL);
2190 	} while (pvo == NULL);
2191 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;
2192 	pvo->pvo_pte.pa = (pa & ~ADDR_POFF) | moea64_calc_wimg(pa, ma);
2193 	pvo->pvo_vaddr |= PVO_WIRED;
2194 
2195 	PMAP_LOCK(kernel_pmap);
2196 	oldpvo = moea64_pvo_find_va(kernel_pmap, va);
2197 	if (oldpvo != NULL)
2198 		moea64_pvo_remove_from_pmap(oldpvo);
2199 	init_pvo_entry(pvo, kernel_pmap, va);
2200 	error = moea64_pvo_enter(pvo, NULL, NULL);
2201 	PMAP_UNLOCK(kernel_pmap);
2202 
2203 	/* Free any dead pages */
2204 	if (oldpvo != NULL) {
2205 		moea64_pvo_remove_from_page(oldpvo);
2206 		free_pvo_entry(oldpvo);
2207 	}
2208 
2209 	if (error != 0)
2210 		panic("moea64_kenter: failed to enter va %#zx pa %#jx: %d", va,
2211 		    (uintmax_t)pa, error);
2212 }
2213 
2214 void
2215 moea64_kenter(vm_offset_t va, vm_paddr_t pa)
2216 {
2217 
2218 	moea64_kenter_attr(va, pa, VM_MEMATTR_DEFAULT);
2219 }
2220 
2221 /*
2222  * Extract the physical page address associated with the given kernel virtual
2223  * address.
2224  */
2225 vm_paddr_t
2226 moea64_kextract(vm_offset_t va)
2227 {
2228 	struct		pvo_entry *pvo;
2229 	vm_paddr_t pa;
2230 
2231 	/*
2232 	 * Shortcut the direct-mapped case when applicable.  We never put
2233 	 * anything but 1:1 (or 62-bit aliased) mappings below
2234 	 * VM_MIN_KERNEL_ADDRESS.
2235 	 */
2236 	if (va < VM_MIN_KERNEL_ADDRESS)
2237 		return (va & ~DMAP_BASE_ADDRESS);
2238 
2239 	PMAP_LOCK(kernel_pmap);
2240 	pvo = moea64_pvo_find_va(kernel_pmap, va);
2241 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found for %#" PRIxPTR,
2242 	    va));
2243 	pa = PVO_PADDR(pvo) | (va - PVO_VADDR(pvo));
2244 	PMAP_UNLOCK(kernel_pmap);
2245 	return (pa);
2246 }
2247 
2248 /*
2249  * Remove a wired page from kernel virtual address space.
2250  */
2251 void
2252 moea64_kremove(vm_offset_t va)
2253 {
2254 	moea64_remove(kernel_pmap, va, va + PAGE_SIZE);
2255 }
2256 
2257 /*
2258  * Provide a kernel pointer corresponding to a given userland pointer.
2259  * The returned pointer is valid until the next time this function is
2260  * called in this thread. This is used internally in copyin/copyout.
2261  */
2262 static int
2263 moea64_map_user_ptr(pmap_t pm, volatile const void *uaddr,
2264     void **kaddr, size_t ulen, size_t *klen)
2265 {
2266 	size_t l;
2267 #ifdef __powerpc64__
2268 	struct slb *slb;
2269 #endif
2270 	register_t slbv;
2271 
2272 	*kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
2273 	l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
2274 	if (l > ulen)
2275 		l = ulen;
2276 	if (klen)
2277 		*klen = l;
2278 	else if (l != ulen)
2279 		return (EFAULT);
2280 
2281 #ifdef __powerpc64__
2282 	/* Try lockless look-up first */
2283 	slb = user_va_to_slb_entry(pm, (vm_offset_t)uaddr);
2284 
2285 	if (slb == NULL) {
2286 		/* If it isn't there, we need to pre-fault the VSID */
2287 		PMAP_LOCK(pm);
2288 		slbv = va_to_vsid(pm, (vm_offset_t)uaddr) << SLBV_VSID_SHIFT;
2289 		PMAP_UNLOCK(pm);
2290 	} else {
2291 		slbv = slb->slbv;
2292 	}
2293 
2294 	/* Mark segment no-execute */
2295 	slbv |= SLBV_N;
2296 #else
2297 	slbv = va_to_vsid(pm, (vm_offset_t)uaddr);
2298 
2299 	/* Mark segment no-execute */
2300 	slbv |= SR_N;
2301 #endif
2302 
2303 	/* If we have already set this VSID, we can just return */
2304 	if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == slbv)
2305 		return (0);
2306 
2307 	__asm __volatile("isync");
2308 	curthread->td_pcb->pcb_cpu.aim.usr_segm =
2309 	    (uintptr_t)uaddr >> ADDR_SR_SHFT;
2310 	curthread->td_pcb->pcb_cpu.aim.usr_vsid = slbv;
2311 #ifdef __powerpc64__
2312 	__asm __volatile ("slbie %0; slbmte %1, %2; isync" ::
2313 	    "r"(USER_ADDR), "r"(slbv), "r"(USER_SLB_SLBE));
2314 #else
2315 	__asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(slbv));
2316 #endif
2317 
2318 	return (0);
2319 }
2320 
2321 /*
2322  * Figure out where a given kernel pointer (usually in a fault) points
2323  * to from the VM's perspective, potentially remapping into userland's
2324  * address space.
2325  */
2326 static int
2327 moea64_decode_kernel_ptr(vm_offset_t addr, int *is_user,
2328     vm_offset_t *decoded_addr)
2329 {
2330 	vm_offset_t user_sr;
2331 
2332 	if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
2333 		user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
2334 		addr &= ADDR_PIDX | ADDR_POFF;
2335 		addr |= user_sr << ADDR_SR_SHFT;
2336 		*decoded_addr = addr;
2337 		*is_user = 1;
2338 	} else {
2339 		*decoded_addr = addr;
2340 		*is_user = 0;
2341 	}
2342 
2343 	return (0);
2344 }
2345 
2346 /*
2347  * Map a range of physical addresses into kernel virtual address space.
2348  *
2349  * The value passed in *virt is a suggested virtual address for the mapping.
2350  * Architectures which can support a direct-mapped physical to virtual region
2351  * can return the appropriate address within that region, leaving '*virt'
2352  * unchanged.  Other architectures should map the pages starting at '*virt' and
2353  * update '*virt' with the first usable address after the mapped region.
2354  */
2355 vm_offset_t
2356 moea64_map(vm_offset_t *virt, vm_paddr_t pa_start,
2357     vm_paddr_t pa_end, int prot)
2358 {
2359 	vm_offset_t	sva, va;
2360 
2361 	if (hw_direct_map) {
2362 		/*
2363 		 * Check if every page in the region is covered by the direct
2364 		 * map. The direct map covers all of physical memory. Use
2365 		 * moea64_calc_wimg() as a shortcut to see if the page is in
2366 		 * physical memory as a way to see if the direct map covers it.
2367 		 */
2368 		for (va = pa_start; va < pa_end; va += PAGE_SIZE)
2369 			if (moea64_calc_wimg(va, VM_MEMATTR_DEFAULT) != LPTE_M)
2370 				break;
2371 		if (va == pa_end)
2372 			return (PHYS_TO_DMAP(pa_start));
2373 	}
2374 	sva = *virt;
2375 	va = sva;
2376 	/* XXX respect prot argument */
2377 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
2378 		moea64_kenter(va, pa_start);
2379 	*virt = va;
2380 
2381 	return (sva);
2382 }
2383 
2384 /*
2385  * Returns true if the pmap's pv is one of the first
2386  * 16 pvs linked to from this page.  This count may
2387  * be changed upwards or downwards in the future; it
2388  * is only necessary that true be returned for a small
2389  * subset of pmaps for proper page aging.
2390  */
2391 boolean_t
2392 moea64_page_exists_quick(pmap_t pmap, vm_page_t m)
2393 {
2394         int loops;
2395 	struct pvo_entry *pvo;
2396 	boolean_t rv;
2397 
2398 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2399 	    ("moea64_page_exists_quick: page %p is not managed", m));
2400 	loops = 0;
2401 	rv = FALSE;
2402 	PV_PAGE_LOCK(m);
2403 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2404 		if (!(pvo->pvo_vaddr & PVO_DEAD) && pvo->pvo_pmap == pmap) {
2405 			rv = TRUE;
2406 			break;
2407 		}
2408 		if (++loops >= 16)
2409 			break;
2410 	}
2411 	PV_PAGE_UNLOCK(m);
2412 	return (rv);
2413 }
2414 
2415 void
2416 moea64_page_init(vm_page_t m)
2417 {
2418 
2419 	m->md.mdpg_attrs = 0;
2420 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
2421 	LIST_INIT(&m->md.mdpg_pvoh);
2422 }
2423 
2424 /*
2425  * Return the number of managed mappings to the given physical page
2426  * that are wired.
2427  */
2428 int
2429 moea64_page_wired_mappings(vm_page_t m)
2430 {
2431 	struct pvo_entry *pvo;
2432 	int count;
2433 
2434 	count = 0;
2435 	if ((m->oflags & VPO_UNMANAGED) != 0)
2436 		return (count);
2437 	PV_PAGE_LOCK(m);
2438 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
2439 		if ((pvo->pvo_vaddr & (PVO_DEAD | PVO_WIRED)) == PVO_WIRED)
2440 			count++;
2441 	PV_PAGE_UNLOCK(m);
2442 	return (count);
2443 }
2444 
2445 static uintptr_t	moea64_vsidcontext;
2446 
2447 uintptr_t
2448 moea64_get_unique_vsid(void) {
2449 	u_int entropy;
2450 	register_t hash;
2451 	uint32_t mask;
2452 	int i;
2453 
2454 	entropy = 0;
2455 	__asm __volatile("mftb %0" : "=r"(entropy));
2456 
2457 	mtx_lock(&moea64_slb_mutex);
2458 	for (i = 0; i < NVSIDS; i += VSID_NBPW) {
2459 		u_int	n;
2460 
2461 		/*
2462 		 * Create a new value by mutiplying by a prime and adding in
2463 		 * entropy from the timebase register.  This is to make the
2464 		 * VSID more random so that the PT hash function collides
2465 		 * less often.  (Note that the prime casues gcc to do shifts
2466 		 * instead of a multiply.)
2467 		 */
2468 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
2469 		hash = moea64_vsidcontext & (NVSIDS - 1);
2470 		if (hash == 0)		/* 0 is special, avoid it */
2471 			continue;
2472 		n = hash >> 5;
2473 		mask = 1 << (hash & (VSID_NBPW - 1));
2474 		hash = (moea64_vsidcontext & VSID_HASHMASK);
2475 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
2476 			/* anything free in this bucket? */
2477 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
2478 				entropy = (moea64_vsidcontext >> 20);
2479 				continue;
2480 			}
2481 			i = ffs(~moea64_vsid_bitmap[n]) - 1;
2482 			mask = 1 << i;
2483 			hash &= rounddown2(VSID_HASHMASK, VSID_NBPW);
2484 			hash |= i;
2485 		}
2486 		if (hash == VSID_VRMA)	/* also special, avoid this too */
2487 			continue;
2488 		KASSERT(!(moea64_vsid_bitmap[n] & mask),
2489 		    ("Allocating in-use VSID %#zx\n", hash));
2490 		moea64_vsid_bitmap[n] |= mask;
2491 		mtx_unlock(&moea64_slb_mutex);
2492 		return (hash);
2493 	}
2494 
2495 	mtx_unlock(&moea64_slb_mutex);
2496 	panic("%s: out of segments",__func__);
2497 }
2498 
2499 #ifdef __powerpc64__
2500 int
2501 moea64_pinit(pmap_t pmap)
2502 {
2503 
2504 	RB_INIT(&pmap->pmap_pvo);
2505 
2506 	pmap->pm_slb_tree_root = slb_alloc_tree();
2507 	pmap->pm_slb = slb_alloc_user_cache();
2508 	pmap->pm_slb_len = 0;
2509 
2510 	return (1);
2511 }
2512 #else
2513 int
2514 moea64_pinit(pmap_t pmap)
2515 {
2516 	int	i;
2517 	uint32_t hash;
2518 
2519 	RB_INIT(&pmap->pmap_pvo);
2520 
2521 	if (pmap_bootstrapped)
2522 		pmap->pmap_phys = (pmap_t)moea64_kextract((vm_offset_t)pmap);
2523 	else
2524 		pmap->pmap_phys = pmap;
2525 
2526 	/*
2527 	 * Allocate some segment registers for this pmap.
2528 	 */
2529 	hash = moea64_get_unique_vsid();
2530 
2531 	for (i = 0; i < 16; i++)
2532 		pmap->pm_sr[i] = VSID_MAKE(i, hash);
2533 
2534 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_pinit: pm_sr[0] = 0"));
2535 
2536 	return (1);
2537 }
2538 #endif
2539 
2540 /*
2541  * Initialize the pmap associated with process 0.
2542  */
2543 void
2544 moea64_pinit0(pmap_t pm)
2545 {
2546 
2547 	PMAP_LOCK_INIT(pm);
2548 	moea64_pinit(pm);
2549 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
2550 }
2551 
2552 /*
2553  * Set the physical protection on the specified range of this map as requested.
2554  */
2555 static void
2556 moea64_pvo_protect( pmap_t pm, struct pvo_entry *pvo, vm_prot_t prot)
2557 {
2558 	struct vm_page *pg;
2559 	vm_prot_t oldprot;
2560 	int32_t refchg;
2561 
2562 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2563 
2564 	/*
2565 	 * Change the protection of the page.
2566 	 */
2567 	oldprot = pvo->pvo_pte.prot;
2568 	pvo->pvo_pte.prot = prot;
2569 	pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
2570 
2571 	/*
2572 	 * If the PVO is in the page table, update mapping
2573 	 */
2574 	refchg = moea64_pte_replace(pvo, MOEA64_PTE_PROT_UPDATE);
2575 	if (refchg < 0)
2576 		refchg = (oldprot & VM_PROT_WRITE) ? LPTE_CHG : 0;
2577 
2578 	if (pm != kernel_pmap && pg != NULL &&
2579 	    (pg->a.flags & PGA_EXECUTABLE) == 0 &&
2580 	    (pvo->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
2581 		if ((pg->oflags & VPO_UNMANAGED) == 0)
2582 			vm_page_aflag_set(pg, PGA_EXECUTABLE);
2583 		moea64_syncicache(pm, PVO_VADDR(pvo),
2584 		    PVO_PADDR(pvo), PAGE_SIZE);
2585 	}
2586 
2587 	/*
2588 	 * Update vm about the REF/CHG bits if the page is managed and we have
2589 	 * removed write access.
2590 	 */
2591 	if (pg != NULL && (pvo->pvo_vaddr & PVO_MANAGED) &&
2592 	    (oldprot & VM_PROT_WRITE)) {
2593 		refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2594 		if (refchg & LPTE_CHG)
2595 			vm_page_dirty(pg);
2596 		if (refchg & LPTE_REF)
2597 			vm_page_aflag_set(pg, PGA_REFERENCED);
2598 	}
2599 }
2600 
2601 void
2602 moea64_protect(pmap_t pm, vm_offset_t sva, vm_offset_t eva,
2603     vm_prot_t prot)
2604 {
2605 	struct	pvo_entry *pvo, key;
2606 
2607 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm,
2608 	    sva, eva, prot);
2609 
2610 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
2611 	    ("moea64_protect: non current pmap"));
2612 
2613 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2614 		moea64_remove(pm, sva, eva);
2615 		return;
2616 	}
2617 
2618 	PMAP_LOCK(pm);
2619 	key.pvo_vaddr = sva;
2620 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2621 	    pvo != NULL && PVO_VADDR(pvo) < eva;
2622 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
2623 		if (PVO_IS_SP(pvo)) {
2624 			if (moea64_sp_pvo_in_range(pvo, sva, eva)) {
2625 				pvo = moea64_sp_protect(pvo, prot);
2626 				continue;
2627 			} else {
2628 				CTR1(KTR_PMAP, "%s: demote before protect",
2629 				    __func__);
2630 				moea64_sp_demote(pvo);
2631 			}
2632 		}
2633 		moea64_pvo_protect(pm, pvo, prot);
2634 	}
2635 	PMAP_UNLOCK(pm);
2636 }
2637 
2638 /*
2639  * Map a list of wired pages into kernel virtual address space.  This is
2640  * intended for temporary mappings which do not need page modification or
2641  * references recorded.  Existing mappings in the region are overwritten.
2642  */
2643 void
2644 moea64_qenter(vm_offset_t va, vm_page_t *m, int count)
2645 {
2646 	while (count-- > 0) {
2647 		moea64_kenter(va, VM_PAGE_TO_PHYS(*m));
2648 		va += PAGE_SIZE;
2649 		m++;
2650 	}
2651 }
2652 
2653 /*
2654  * Remove page mappings from kernel virtual address space.  Intended for
2655  * temporary mappings entered by moea64_qenter.
2656  */
2657 void
2658 moea64_qremove(vm_offset_t va, int count)
2659 {
2660 	while (count-- > 0) {
2661 		moea64_kremove(va);
2662 		va += PAGE_SIZE;
2663 	}
2664 }
2665 
2666 void
2667 moea64_release_vsid(uint64_t vsid)
2668 {
2669 	int idx, mask;
2670 
2671 	mtx_lock(&moea64_slb_mutex);
2672 	idx = vsid & (NVSIDS-1);
2673 	mask = 1 << (idx % VSID_NBPW);
2674 	idx /= VSID_NBPW;
2675 	KASSERT(moea64_vsid_bitmap[idx] & mask,
2676 	    ("Freeing unallocated VSID %#jx", vsid));
2677 	moea64_vsid_bitmap[idx] &= ~mask;
2678 	mtx_unlock(&moea64_slb_mutex);
2679 }
2680 
2681 void
2682 moea64_release(pmap_t pmap)
2683 {
2684 
2685 	/*
2686 	 * Free segment registers' VSIDs
2687 	 */
2688     #ifdef __powerpc64__
2689 	slb_free_tree(pmap);
2690 	slb_free_user_cache(pmap->pm_slb);
2691     #else
2692 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_release: pm_sr[0] = 0"));
2693 
2694 	moea64_release_vsid(VSID_TO_HASH(pmap->pm_sr[0]));
2695     #endif
2696 }
2697 
2698 /*
2699  * Remove all pages mapped by the specified pmap
2700  */
2701 void
2702 moea64_remove_pages(pmap_t pm)
2703 {
2704 	struct pvo_entry *pvo, *tpvo;
2705 	struct pvo_dlist tofree;
2706 
2707 	SLIST_INIT(&tofree);
2708 
2709 	PMAP_LOCK(pm);
2710 	RB_FOREACH_SAFE(pvo, pvo_tree, &pm->pmap_pvo, tpvo) {
2711 		if (pvo->pvo_vaddr & PVO_WIRED)
2712 			continue;
2713 
2714 		/*
2715 		 * For locking reasons, remove this from the page table and
2716 		 * pmap, but save delinking from the vm_page for a second
2717 		 * pass
2718 		 */
2719 		moea64_pvo_remove_from_pmap(pvo);
2720 		SLIST_INSERT_HEAD(&tofree, pvo, pvo_dlink);
2721 	}
2722 	PMAP_UNLOCK(pm);
2723 
2724 	while (!SLIST_EMPTY(&tofree)) {
2725 		pvo = SLIST_FIRST(&tofree);
2726 		SLIST_REMOVE_HEAD(&tofree, pvo_dlink);
2727 		moea64_pvo_remove_from_page(pvo);
2728 		free_pvo_entry(pvo);
2729 	}
2730 }
2731 
2732 static void
2733 moea64_remove_locked(pmap_t pm, vm_offset_t sva, vm_offset_t eva,
2734     struct pvo_dlist *tofree)
2735 {
2736 	struct pvo_entry *pvo, *tpvo, key;
2737 
2738 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2739 
2740 	key.pvo_vaddr = sva;
2741 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2742 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2743 		if (PVO_IS_SP(pvo)) {
2744 			if (moea64_sp_pvo_in_range(pvo, sva, eva)) {
2745 				tpvo = moea64_sp_remove(pvo, tofree);
2746 				continue;
2747 			} else {
2748 				CTR1(KTR_PMAP, "%s: demote before remove",
2749 				    __func__);
2750 				moea64_sp_demote(pvo);
2751 			}
2752 		}
2753 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2754 
2755 		/*
2756 		 * For locking reasons, remove this from the page table and
2757 		 * pmap, but save delinking from the vm_page for a second
2758 		 * pass
2759 		 */
2760 		moea64_pvo_remove_from_pmap(pvo);
2761 		SLIST_INSERT_HEAD(tofree, pvo, pvo_dlink);
2762 	}
2763 }
2764 
2765 /*
2766  * Remove the given range of addresses from the specified map.
2767  */
2768 void
2769 moea64_remove(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
2770 {
2771 	struct pvo_entry *pvo;
2772 	struct pvo_dlist tofree;
2773 
2774 	/*
2775 	 * Perform an unsynchronized read.  This is, however, safe.
2776 	 */
2777 	if (pm->pm_stats.resident_count == 0)
2778 		return;
2779 
2780 	SLIST_INIT(&tofree);
2781 	PMAP_LOCK(pm);
2782 	moea64_remove_locked(pm, sva, eva, &tofree);
2783 	PMAP_UNLOCK(pm);
2784 
2785 	while (!SLIST_EMPTY(&tofree)) {
2786 		pvo = SLIST_FIRST(&tofree);
2787 		SLIST_REMOVE_HEAD(&tofree, pvo_dlink);
2788 		moea64_pvo_remove_from_page(pvo);
2789 		free_pvo_entry(pvo);
2790 	}
2791 }
2792 
2793 /*
2794  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
2795  * will reflect changes in pte's back to the vm_page.
2796  */
2797 void
2798 moea64_remove_all(vm_page_t m)
2799 {
2800 	struct	pvo_entry *pvo, *next_pvo;
2801 	struct	pvo_head freequeue;
2802 	int	wasdead;
2803 	pmap_t	pmap;
2804 
2805 	LIST_INIT(&freequeue);
2806 
2807 	PV_PAGE_LOCK(m);
2808 	LIST_FOREACH_SAFE(pvo, vm_page_to_pvoh(m), pvo_vlink, next_pvo) {
2809 		pmap = pvo->pvo_pmap;
2810 		PMAP_LOCK(pmap);
2811 		wasdead = (pvo->pvo_vaddr & PVO_DEAD);
2812 		if (!wasdead) {
2813 			if (PVO_IS_SP(pvo)) {
2814 				CTR1(KTR_PMAP, "%s: demote before remove_all",
2815 				    __func__);
2816 				moea64_sp_demote(pvo);
2817 			}
2818 			moea64_pvo_remove_from_pmap(pvo);
2819 		}
2820 		moea64_pvo_remove_from_page_locked(pvo, m);
2821 		if (!wasdead)
2822 			LIST_INSERT_HEAD(&freequeue, pvo, pvo_vlink);
2823 		PMAP_UNLOCK(pmap);
2824 
2825 	}
2826 	KASSERT(!pmap_page_is_mapped(m), ("Page still has mappings"));
2827 	KASSERT((m->a.flags & PGA_WRITEABLE) == 0, ("Page still writable"));
2828 	PV_PAGE_UNLOCK(m);
2829 
2830 	/* Clean up UMA allocations */
2831 	LIST_FOREACH_SAFE(pvo, &freequeue, pvo_vlink, next_pvo)
2832 		free_pvo_entry(pvo);
2833 }
2834 
2835 /*
2836  * Allocate a physical page of memory directly from the phys_avail map.
2837  * Can only be called from moea64_bootstrap before avail start and end are
2838  * calculated.
2839  */
2840 vm_offset_t
2841 moea64_bootstrap_alloc(vm_size_t size, vm_size_t align)
2842 {
2843 	vm_offset_t	s, e;
2844 	int		i, j;
2845 
2846 	size = round_page(size);
2847 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2848 		if (align != 0)
2849 			s = roundup2(phys_avail[i], align);
2850 		else
2851 			s = phys_avail[i];
2852 		e = s + size;
2853 
2854 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2855 			continue;
2856 
2857 		if (s + size > platform_real_maxaddr())
2858 			continue;
2859 
2860 		if (s == phys_avail[i]) {
2861 			phys_avail[i] += size;
2862 		} else if (e == phys_avail[i + 1]) {
2863 			phys_avail[i + 1] -= size;
2864 		} else {
2865 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2866 				phys_avail[j] = phys_avail[j - 2];
2867 				phys_avail[j + 1] = phys_avail[j - 1];
2868 			}
2869 
2870 			phys_avail[i + 3] = phys_avail[i + 1];
2871 			phys_avail[i + 1] = s;
2872 			phys_avail[i + 2] = e;
2873 			phys_avail_count++;
2874 		}
2875 
2876 		return (s);
2877 	}
2878 	panic("moea64_bootstrap_alloc: could not allocate memory");
2879 }
2880 
2881 static int
2882 moea64_pvo_enter(struct pvo_entry *pvo, struct pvo_head *pvo_head,
2883     struct pvo_entry **oldpvop)
2884 {
2885 	struct pvo_entry *old_pvo;
2886 	int err;
2887 
2888 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2889 
2890 	STAT_MOEA64(moea64_pvo_enter_calls++);
2891 
2892 	/*
2893 	 * Add to pmap list
2894 	 */
2895 	old_pvo = RB_INSERT(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2896 
2897 	if (old_pvo != NULL) {
2898 		if (oldpvop != NULL)
2899 			*oldpvop = old_pvo;
2900 		return (EEXIST);
2901 	}
2902 
2903 	if (pvo_head != NULL) {
2904 		LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2905 	}
2906 
2907 	if (pvo->pvo_vaddr & PVO_WIRED)
2908 		pvo->pvo_pmap->pm_stats.wired_count++;
2909 	pvo->pvo_pmap->pm_stats.resident_count++;
2910 
2911 	/*
2912 	 * Insert it into the hardware page table
2913 	 */
2914 	err = moea64_pte_insert(pvo);
2915 	if (err != 0) {
2916 		panic("moea64_pvo_enter: overflow");
2917 	}
2918 
2919 	STAT_MOEA64(moea64_pvo_entries++);
2920 
2921 	if (pvo->pvo_pmap == kernel_pmap)
2922 		isync();
2923 
2924 #ifdef __powerpc64__
2925 	/*
2926 	 * Make sure all our bootstrap mappings are in the SLB as soon
2927 	 * as virtual memory is switched on.
2928 	 */
2929 	if (!pmap_bootstrapped)
2930 		moea64_bootstrap_slb_prefault(PVO_VADDR(pvo),
2931 		    pvo->pvo_vaddr & PVO_LARGE);
2932 #endif
2933 
2934 	return (0);
2935 }
2936 
2937 static void
2938 moea64_pvo_remove_from_pmap(struct pvo_entry *pvo)
2939 {
2940 	struct	vm_page *pg;
2941 	int32_t refchg;
2942 
2943 	KASSERT(pvo->pvo_pmap != NULL, ("Trying to remove PVO with no pmap"));
2944 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2945 	KASSERT(!(pvo->pvo_vaddr & PVO_DEAD), ("Trying to remove dead PVO"));
2946 
2947 	/*
2948 	 * If there is an active pte entry, we need to deactivate it
2949 	 */
2950 	refchg = moea64_pte_unset(pvo);
2951 	if (refchg < 0) {
2952 		/*
2953 		 * If it was evicted from the page table, be pessimistic and
2954 		 * dirty the page.
2955 		 */
2956 		if (pvo->pvo_pte.prot & VM_PROT_WRITE)
2957 			refchg = LPTE_CHG;
2958 		else
2959 			refchg = 0;
2960 	}
2961 
2962 	/*
2963 	 * Update our statistics.
2964 	 */
2965 	pvo->pvo_pmap->pm_stats.resident_count--;
2966 	if (pvo->pvo_vaddr & PVO_WIRED)
2967 		pvo->pvo_pmap->pm_stats.wired_count--;
2968 
2969 	/*
2970 	 * Remove this PVO from the pmap list.
2971 	 */
2972 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2973 
2974 	/*
2975 	 * Mark this for the next sweep
2976 	 */
2977 	pvo->pvo_vaddr |= PVO_DEAD;
2978 
2979 	/* Send RC bits to VM */
2980 	if ((pvo->pvo_vaddr & PVO_MANAGED) &&
2981 	    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2982 		pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
2983 		if (pg != NULL) {
2984 			refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2985 			if (refchg & LPTE_CHG)
2986 				vm_page_dirty(pg);
2987 			if (refchg & LPTE_REF)
2988 				vm_page_aflag_set(pg, PGA_REFERENCED);
2989 		}
2990 	}
2991 }
2992 
2993 static inline void
2994 moea64_pvo_remove_from_page_locked(struct pvo_entry *pvo,
2995     vm_page_t m)
2996 {
2997 
2998 	KASSERT(pvo->pvo_vaddr & PVO_DEAD, ("Trying to delink live page"));
2999 
3000 	/* Use NULL pmaps as a sentinel for races in page deletion */
3001 	if (pvo->pvo_pmap == NULL)
3002 		return;
3003 	pvo->pvo_pmap = NULL;
3004 
3005 	/*
3006 	 * Update vm about page writeability/executability if managed
3007 	 */
3008 	PV_LOCKASSERT(PVO_PADDR(pvo));
3009 	if (pvo->pvo_vaddr & PVO_MANAGED) {
3010 		if (m != NULL) {
3011 			LIST_REMOVE(pvo, pvo_vlink);
3012 			if (LIST_EMPTY(vm_page_to_pvoh(m)))
3013 				vm_page_aflag_clear(m,
3014 				    PGA_WRITEABLE | PGA_EXECUTABLE);
3015 		}
3016 	}
3017 
3018 	STAT_MOEA64(moea64_pvo_entries--);
3019 	STAT_MOEA64(moea64_pvo_remove_calls++);
3020 }
3021 
3022 static void
3023 moea64_pvo_remove_from_page(struct pvo_entry *pvo)
3024 {
3025 	vm_page_t pg = NULL;
3026 
3027 	if (pvo->pvo_vaddr & PVO_MANAGED)
3028 		pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
3029 
3030 	PV_LOCK(PVO_PADDR(pvo));
3031 	moea64_pvo_remove_from_page_locked(pvo, pg);
3032 	PV_UNLOCK(PVO_PADDR(pvo));
3033 }
3034 
3035 static struct pvo_entry *
3036 moea64_pvo_find_va(pmap_t pm, vm_offset_t va)
3037 {
3038 	struct pvo_entry key;
3039 
3040 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
3041 
3042 	key.pvo_vaddr = va & ~ADDR_POFF;
3043 	return (RB_FIND(pvo_tree, &pm->pmap_pvo, &key));
3044 }
3045 
3046 static boolean_t
3047 moea64_query_bit(vm_page_t m, uint64_t ptebit)
3048 {
3049 	struct	pvo_entry *pvo;
3050 	int64_t ret;
3051 	boolean_t rv;
3052 	vm_page_t sp;
3053 
3054 	/*
3055 	 * See if this bit is stored in the page already.
3056 	 *
3057 	 * For superpages, the bit is stored in the first vm page.
3058 	 */
3059 	if ((m->md.mdpg_attrs & ptebit) != 0 ||
3060 	    ((sp = PHYS_TO_VM_PAGE(VM_PAGE_TO_PHYS(m) & ~HPT_SP_MASK)) != NULL &&
3061 	     (sp->md.mdpg_attrs & (ptebit | MDPG_ATTR_SP)) ==
3062 	     (ptebit | MDPG_ATTR_SP)))
3063 		return (TRUE);
3064 
3065 	/*
3066 	 * Examine each PTE.  Sync so that any pending REF/CHG bits are
3067 	 * flushed to the PTEs.
3068 	 */
3069 	rv = FALSE;
3070 	powerpc_sync();
3071 	PV_PAGE_LOCK(m);
3072 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
3073 		if (PVO_IS_SP(pvo)) {
3074 			ret = moea64_sp_query(pvo, ptebit);
3075 			/*
3076 			 * If SP was not demoted, check its REF/CHG bits here.
3077 			 */
3078 			if (ret != -1) {
3079 				if ((ret & ptebit) != 0) {
3080 					rv = TRUE;
3081 					break;
3082 				}
3083 				continue;
3084 			}
3085 			/* else, fallthrough */
3086 		}
3087 
3088 		ret = 0;
3089 
3090 		/*
3091 		 * See if this pvo has a valid PTE.  if so, fetch the
3092 		 * REF/CHG bits from the valid PTE.  If the appropriate
3093 		 * ptebit is set, return success.
3094 		 */
3095 		PMAP_LOCK(pvo->pvo_pmap);
3096 		if (!(pvo->pvo_vaddr & PVO_DEAD))
3097 			ret = moea64_pte_synch(pvo);
3098 		PMAP_UNLOCK(pvo->pvo_pmap);
3099 
3100 		if (ret > 0) {
3101 			atomic_set_32(&m->md.mdpg_attrs,
3102 			    ret & (LPTE_CHG | LPTE_REF));
3103 			if (ret & ptebit) {
3104 				rv = TRUE;
3105 				break;
3106 			}
3107 		}
3108 	}
3109 	PV_PAGE_UNLOCK(m);
3110 
3111 	return (rv);
3112 }
3113 
3114 static u_int
3115 moea64_clear_bit(vm_page_t m, u_int64_t ptebit)
3116 {
3117 	u_int	count;
3118 	struct	pvo_entry *pvo;
3119 	int64_t ret;
3120 
3121 	/*
3122 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
3123 	 * we can reset the right ones).
3124 	 */
3125 	powerpc_sync();
3126 
3127 	/*
3128 	 * For each pvo entry, clear the pte's ptebit.
3129 	 */
3130 	count = 0;
3131 	PV_PAGE_LOCK(m);
3132 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
3133 		if (PVO_IS_SP(pvo)) {
3134 			if ((ret = moea64_sp_clear(pvo, m, ptebit)) != -1) {
3135 				count += ret;
3136 				continue;
3137 			}
3138 		}
3139 		ret = 0;
3140 
3141 		PMAP_LOCK(pvo->pvo_pmap);
3142 		if (!(pvo->pvo_vaddr & PVO_DEAD))
3143 			ret = moea64_pte_clear(pvo, ptebit);
3144 		PMAP_UNLOCK(pvo->pvo_pmap);
3145 
3146 		if (ret > 0 && (ret & ptebit))
3147 			count++;
3148 	}
3149 	atomic_clear_32(&m->md.mdpg_attrs, ptebit);
3150 	PV_PAGE_UNLOCK(m);
3151 
3152 	return (count);
3153 }
3154 
3155 boolean_t
3156 moea64_dev_direct_mapped(vm_paddr_t pa, vm_size_t size)
3157 {
3158 	struct pvo_entry *pvo, key;
3159 	vm_offset_t ppa;
3160 	int error = 0;
3161 
3162 	if (hw_direct_map && mem_valid(pa, size) == 0)
3163 		return (0);
3164 
3165 	PMAP_LOCK(kernel_pmap);
3166 	ppa = pa & ~ADDR_POFF;
3167 	key.pvo_vaddr = DMAP_BASE_ADDRESS + ppa;
3168 	for (pvo = RB_FIND(pvo_tree, &kernel_pmap->pmap_pvo, &key);
3169 	    ppa < pa + size; ppa += PAGE_SIZE,
3170 	    pvo = RB_NEXT(pvo_tree, &kernel_pmap->pmap_pvo, pvo)) {
3171 		if (pvo == NULL || PVO_PADDR(pvo) != ppa) {
3172 			error = EFAULT;
3173 			break;
3174 		}
3175 	}
3176 	PMAP_UNLOCK(kernel_pmap);
3177 
3178 	return (error);
3179 }
3180 
3181 /*
3182  * Map a set of physical memory pages into the kernel virtual
3183  * address space. Return a pointer to where it is mapped. This
3184  * routine is intended to be used for mapping device memory,
3185  * NOT real memory.
3186  */
3187 void *
3188 moea64_mapdev_attr(vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
3189 {
3190 	vm_offset_t va, tmpva, ppa, offset;
3191 
3192 	ppa = trunc_page(pa);
3193 	offset = pa & PAGE_MASK;
3194 	size = roundup2(offset + size, PAGE_SIZE);
3195 
3196 	va = kva_alloc(size);
3197 
3198 	if (!va)
3199 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
3200 
3201 	for (tmpva = va; size > 0;) {
3202 		moea64_kenter_attr(tmpva, ppa, ma);
3203 		size -= PAGE_SIZE;
3204 		tmpva += PAGE_SIZE;
3205 		ppa += PAGE_SIZE;
3206 	}
3207 
3208 	return ((void *)(va + offset));
3209 }
3210 
3211 void *
3212 moea64_mapdev(vm_paddr_t pa, vm_size_t size)
3213 {
3214 
3215 	return moea64_mapdev_attr(pa, size, VM_MEMATTR_DEFAULT);
3216 }
3217 
3218 void
3219 moea64_unmapdev(vm_offset_t va, vm_size_t size)
3220 {
3221 	vm_offset_t base, offset;
3222 
3223 	base = trunc_page(va);
3224 	offset = va & PAGE_MASK;
3225 	size = roundup2(offset + size, PAGE_SIZE);
3226 
3227 	moea64_qremove(base, atop(size));
3228 	kva_free(base, size);
3229 }
3230 
3231 void
3232 moea64_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz)
3233 {
3234 	struct pvo_entry *pvo;
3235 	vm_offset_t lim;
3236 	vm_paddr_t pa;
3237 	vm_size_t len;
3238 
3239 	if (__predict_false(pm == NULL))
3240 		pm = &curthread->td_proc->p_vmspace->vm_pmap;
3241 
3242 	PMAP_LOCK(pm);
3243 	while (sz > 0) {
3244 		lim = round_page(va+1);
3245 		len = MIN(lim - va, sz);
3246 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
3247 		if (pvo != NULL && !(pvo->pvo_pte.pa & LPTE_I)) {
3248 			pa = PVO_PADDR(pvo) | (va & ADDR_POFF);
3249 			moea64_syncicache(pm, va, pa, len);
3250 		}
3251 		va += len;
3252 		sz -= len;
3253 	}
3254 	PMAP_UNLOCK(pm);
3255 }
3256 
3257 void
3258 moea64_dumpsys_map(vm_paddr_t pa, size_t sz, void **va)
3259 {
3260 
3261 	*va = (void *)(uintptr_t)pa;
3262 }
3263 
3264 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
3265 
3266 void
3267 moea64_scan_init()
3268 {
3269 	struct pvo_entry *pvo;
3270 	vm_offset_t va;
3271 	int i;
3272 
3273 	if (!do_minidump) {
3274 		/* Initialize phys. segments for dumpsys(). */
3275 		memset(&dump_map, 0, sizeof(dump_map));
3276 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
3277 		for (i = 0; i < pregions_sz; i++) {
3278 			dump_map[i].pa_start = pregions[i].mr_start;
3279 			dump_map[i].pa_size = pregions[i].mr_size;
3280 		}
3281 		return;
3282 	}
3283 
3284 	/* Virtual segments for minidumps: */
3285 	memset(&dump_map, 0, sizeof(dump_map));
3286 
3287 	/* 1st: kernel .data and .bss. */
3288 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
3289 	dump_map[0].pa_size = round_page((uintptr_t)_end) -
3290 	    dump_map[0].pa_start;
3291 
3292 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
3293 	dump_map[1].pa_start = (vm_paddr_t)(uintptr_t)msgbufp->msg_ptr;
3294 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
3295 
3296 	/* 3rd: kernel VM. */
3297 	va = dump_map[1].pa_start + dump_map[1].pa_size;
3298 	/* Find start of next chunk (from va). */
3299 	while (va < virtual_end) {
3300 		/* Don't dump the buffer cache. */
3301 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
3302 			va = kmi.buffer_eva;
3303 			continue;
3304 		}
3305 		pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
3306 		if (pvo != NULL && !(pvo->pvo_vaddr & PVO_DEAD))
3307 			break;
3308 		va += PAGE_SIZE;
3309 	}
3310 	if (va < virtual_end) {
3311 		dump_map[2].pa_start = va;
3312 		va += PAGE_SIZE;
3313 		/* Find last page in chunk. */
3314 		while (va < virtual_end) {
3315 			/* Don't run into the buffer cache. */
3316 			if (va == kmi.buffer_sva)
3317 				break;
3318 			pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
3319 			if (pvo == NULL || (pvo->pvo_vaddr & PVO_DEAD))
3320 				break;
3321 			va += PAGE_SIZE;
3322 		}
3323 		dump_map[2].pa_size = va - dump_map[2].pa_start;
3324 	}
3325 }
3326 
3327 #ifdef __powerpc64__
3328 
3329 static size_t
3330 moea64_scan_pmap()
3331 {
3332 	struct pvo_entry *pvo;
3333 	vm_paddr_t pa, pa_end;
3334 	vm_offset_t va, pgva, kstart, kend, kstart_lp, kend_lp;
3335 	uint64_t lpsize;
3336 
3337 	lpsize = moea64_large_page_size;
3338 	kstart = trunc_page((vm_offset_t)_etext);
3339 	kend = round_page((vm_offset_t)_end);
3340 	kstart_lp = kstart & ~moea64_large_page_mask;
3341 	kend_lp = (kend + moea64_large_page_mask) & ~moea64_large_page_mask;
3342 
3343 	CTR4(KTR_PMAP, "moea64_scan_pmap: kstart=0x%016lx, kend=0x%016lx, "
3344 	    "kstart_lp=0x%016lx, kend_lp=0x%016lx",
3345 	    kstart, kend, kstart_lp, kend_lp);
3346 
3347 	PMAP_LOCK(kernel_pmap);
3348 	RB_FOREACH(pvo, pvo_tree, &kernel_pmap->pmap_pvo) {
3349 		va = pvo->pvo_vaddr;
3350 
3351 		if (va & PVO_DEAD)
3352 			continue;
3353 
3354 		/* Skip DMAP (except kernel area) */
3355 		if (va >= DMAP_BASE_ADDRESS && va <= DMAP_MAX_ADDRESS) {
3356 			if (va & PVO_LARGE) {
3357 				pgva = va & ~moea64_large_page_mask;
3358 				if (pgva < kstart_lp || pgva >= kend_lp)
3359 					continue;
3360 			} else {
3361 				pgva = trunc_page(va);
3362 				if (pgva < kstart || pgva >= kend)
3363 					continue;
3364 			}
3365 		}
3366 
3367 		pa = PVO_PADDR(pvo);
3368 
3369 		if (va & PVO_LARGE) {
3370 			pa_end = pa + lpsize;
3371 			for (; pa < pa_end; pa += PAGE_SIZE) {
3372 				if (is_dumpable(pa))
3373 					dump_add_page(pa);
3374 			}
3375 		} else {
3376 			if (is_dumpable(pa))
3377 				dump_add_page(pa);
3378 		}
3379 	}
3380 	PMAP_UNLOCK(kernel_pmap);
3381 
3382 	return (sizeof(struct lpte) * moea64_pteg_count * 8);
3383 }
3384 
3385 static struct dump_context dump_ctx;
3386 
3387 static void *
3388 moea64_dump_pmap_init(unsigned blkpgs)
3389 {
3390 	dump_ctx.ptex = 0;
3391 	dump_ctx.ptex_end = moea64_pteg_count * 8;
3392 	dump_ctx.blksz = blkpgs * PAGE_SIZE;
3393 	return (&dump_ctx);
3394 }
3395 
3396 #else
3397 
3398 static size_t
3399 moea64_scan_pmap()
3400 {
3401 	return (0);
3402 }
3403 
3404 static void *
3405 moea64_dump_pmap_init(unsigned blkpgs)
3406 {
3407 	return (NULL);
3408 }
3409 
3410 #endif
3411 
3412 #ifdef __powerpc64__
3413 static void
3414 moea64_map_range(vm_offset_t va, vm_paddr_t pa, vm_size_t npages)
3415 {
3416 
3417 	for (; npages > 0; --npages) {
3418 		if (moea64_large_page_size != 0 &&
3419 		    (pa & moea64_large_page_mask) == 0 &&
3420 		    (va & moea64_large_page_mask) == 0 &&
3421 		    npages >= (moea64_large_page_size >> PAGE_SHIFT)) {
3422 			PMAP_LOCK(kernel_pmap);
3423 			moea64_kenter_large(va, pa, 0, 0);
3424 			PMAP_UNLOCK(kernel_pmap);
3425 			pa += moea64_large_page_size;
3426 			va += moea64_large_page_size;
3427 			npages -= (moea64_large_page_size >> PAGE_SHIFT) - 1;
3428 		} else {
3429 			moea64_kenter(va, pa);
3430 			pa += PAGE_SIZE;
3431 			va += PAGE_SIZE;
3432 		}
3433 	}
3434 }
3435 
3436 static void
3437 moea64_page_array_startup(long pages)
3438 {
3439 	long dom_pages[MAXMEMDOM];
3440 	vm_paddr_t pa;
3441 	vm_offset_t va, vm_page_base;
3442 	vm_size_t needed, size;
3443 	long page;
3444 	int domain;
3445 	int i;
3446 
3447 	vm_page_base = 0xd000000000000000ULL;
3448 
3449 	/* Short-circuit single-domain systems. */
3450 	if (vm_ndomains == 1) {
3451 		size = round_page(pages * sizeof(struct vm_page));
3452 		pa = vm_phys_early_alloc(0, size);
3453 		vm_page_base = moea64_map(&vm_page_base,
3454 		    pa, pa + size, VM_PROT_READ | VM_PROT_WRITE);
3455 		vm_page_array_size = pages;
3456 		vm_page_array = (vm_page_t)vm_page_base;
3457 		return;
3458 	}
3459 
3460 	page = 0;
3461 	for (i = 0; i < MAXMEMDOM; i++)
3462 		dom_pages[i] = 0;
3463 
3464 	/* Now get the number of pages required per domain. */
3465 	for (i = 0; i < vm_phys_nsegs; i++) {
3466 		domain = vm_phys_segs[i].domain;
3467 		KASSERT(domain < MAXMEMDOM,
3468 		    ("Invalid vm_phys_segs NUMA domain %d!\n", domain));
3469 		/* Get size of vm_page_array needed for this segment. */
3470 		size = btoc(vm_phys_segs[i].end - vm_phys_segs[i].start);
3471 		dom_pages[domain] += size;
3472 	}
3473 
3474 	for (i = 0; phys_avail[i + 1] != 0; i+= 2) {
3475 		domain = vm_phys_domain(phys_avail[i]);
3476 		KASSERT(domain < MAXMEMDOM,
3477 		    ("Invalid phys_avail NUMA domain %d!\n", domain));
3478 		size = btoc(phys_avail[i + 1] - phys_avail[i]);
3479 		dom_pages[domain] += size;
3480 	}
3481 
3482 	/*
3483 	 * Map in chunks that can get us all 16MB pages.  There will be some
3484 	 * overlap between domains, but that's acceptable for now.
3485 	 */
3486 	vm_page_array_size = 0;
3487 	va = vm_page_base;
3488 	for (i = 0; i < MAXMEMDOM && vm_page_array_size < pages; i++) {
3489 		if (dom_pages[i] == 0)
3490 			continue;
3491 		size = ulmin(pages - vm_page_array_size, dom_pages[i]);
3492 		size = round_page(size * sizeof(struct vm_page));
3493 		needed = size;
3494 		size = roundup2(size, moea64_large_page_size);
3495 		pa = vm_phys_early_alloc(i, size);
3496 		vm_page_array_size += size / sizeof(struct vm_page);
3497 		moea64_map_range(va, pa, size >> PAGE_SHIFT);
3498 		/* Scoot up domain 0, to reduce the domain page overlap. */
3499 		if (i == 0)
3500 			vm_page_base += size - needed;
3501 		va += size;
3502 	}
3503 	vm_page_array = (vm_page_t)vm_page_base;
3504 	vm_page_array_size = pages;
3505 }
3506 #endif
3507 
3508 static int64_t
3509 moea64_null_method(void)
3510 {
3511 	return (0);
3512 }
3513 
3514 static int64_t moea64_pte_replace_default(struct pvo_entry *pvo, int flags)
3515 {
3516 	int64_t refchg;
3517 
3518 	refchg = moea64_pte_unset(pvo);
3519 	moea64_pte_insert(pvo);
3520 
3521 	return (refchg);
3522 }
3523 
3524 struct moea64_funcs *moea64_ops;
3525 
3526 #define DEFINE_OEA64_IFUNC(ret, func, args, def)		\
3527 	DEFINE_IFUNC(, ret, moea64_##func, args) {		\
3528 		moea64_##func##_t f;				\
3529 		if (moea64_ops == NULL)				\
3530 			return ((moea64_##func##_t)def);	\
3531 		f = moea64_ops->func;				\
3532 		return (f != NULL ? f : (moea64_##func##_t)def);\
3533 	}
3534 
3535 DEFINE_OEA64_IFUNC(int64_t, pte_replace, (struct pvo_entry *, int),
3536     moea64_pte_replace_default)
3537 DEFINE_OEA64_IFUNC(int64_t, pte_insert, (struct pvo_entry *), moea64_null_method)
3538 DEFINE_OEA64_IFUNC(int64_t, pte_unset, (struct pvo_entry *), moea64_null_method)
3539 DEFINE_OEA64_IFUNC(int64_t, pte_clear, (struct pvo_entry *, uint64_t),
3540     moea64_null_method)
3541 DEFINE_OEA64_IFUNC(int64_t, pte_synch, (struct pvo_entry *), moea64_null_method)
3542 DEFINE_OEA64_IFUNC(int64_t, pte_insert_sp, (struct pvo_entry *), moea64_null_method)
3543 DEFINE_OEA64_IFUNC(int64_t, pte_unset_sp, (struct pvo_entry *), moea64_null_method)
3544 DEFINE_OEA64_IFUNC(int64_t, pte_replace_sp, (struct pvo_entry *), moea64_null_method)
3545 
3546 /* Superpage functions */
3547 
3548 /* MMU interface */
3549 
3550 static bool
3551 moea64_ps_enabled(pmap_t pmap)
3552 {
3553 	return (superpages_enabled);
3554 }
3555 
3556 static void
3557 moea64_align_superpage(vm_object_t object, vm_ooffset_t offset,
3558     vm_offset_t *addr, vm_size_t size)
3559 {
3560 	vm_offset_t sp_offset;
3561 
3562 	if (size < HPT_SP_SIZE)
3563 		return;
3564 
3565 	CTR4(KTR_PMAP, "%s: offs=%#jx, addr=%p, size=%#jx",
3566 	    __func__, (uintmax_t)offset, addr, (uintmax_t)size);
3567 
3568 	if (object != NULL && (object->flags & OBJ_COLORED) != 0)
3569 		offset += ptoa(object->pg_color);
3570 	sp_offset = offset & HPT_SP_MASK;
3571 	if (size - ((HPT_SP_SIZE - sp_offset) & HPT_SP_MASK) < HPT_SP_SIZE ||
3572 	    (*addr & HPT_SP_MASK) == sp_offset)
3573 		return;
3574 	if ((*addr & HPT_SP_MASK) < sp_offset)
3575 		*addr = (*addr & ~HPT_SP_MASK) + sp_offset;
3576 	else
3577 		*addr = ((*addr + HPT_SP_MASK) & ~HPT_SP_MASK) + sp_offset;
3578 }
3579 
3580 /* Helpers */
3581 
3582 static __inline void
3583 moea64_pvo_cleanup(struct pvo_dlist *tofree)
3584 {
3585 	struct pvo_entry *pvo;
3586 
3587 	/* clean up */
3588 	while (!SLIST_EMPTY(tofree)) {
3589 		pvo = SLIST_FIRST(tofree);
3590 		SLIST_REMOVE_HEAD(tofree, pvo_dlink);
3591 		if (pvo->pvo_vaddr & PVO_DEAD)
3592 			moea64_pvo_remove_from_page(pvo);
3593 		free_pvo_entry(pvo);
3594 	}
3595 }
3596 
3597 static __inline uint16_t
3598 pvo_to_vmpage_flags(struct pvo_entry *pvo)
3599 {
3600 	uint16_t flags;
3601 
3602 	flags = 0;
3603 	if ((pvo->pvo_pte.prot & VM_PROT_WRITE) != 0)
3604 		flags |= PGA_WRITEABLE;
3605 	if ((pvo->pvo_pte.prot & VM_PROT_EXECUTE) != 0)
3606 		flags |= PGA_EXECUTABLE;
3607 
3608 	return (flags);
3609 }
3610 
3611 /*
3612  * Check if the given pvo and its superpage are in sva-eva range.
3613  */
3614 static __inline bool
3615 moea64_sp_pvo_in_range(struct pvo_entry *pvo, vm_offset_t sva, vm_offset_t eva)
3616 {
3617 	vm_offset_t spva;
3618 
3619 	spva = PVO_VADDR(pvo) & ~HPT_SP_MASK;
3620 	if (spva >= sva && spva + HPT_SP_SIZE <= eva) {
3621 		/*
3622 		 * Because this function is intended to be called from loops
3623 		 * that iterate over ordered pvo entries, if the condition
3624 		 * above is true then the pvo must be the first of its
3625 		 * superpage.
3626 		 */
3627 		KASSERT(PVO_VADDR(pvo) == spva,
3628 		    ("%s: unexpected unaligned superpage pvo", __func__));
3629 		return (true);
3630 	}
3631 	return (false);
3632 }
3633 
3634 /*
3635  * Update vm about the REF/CHG bits if the superpage is managed and
3636  * has (or had) write access.
3637  */
3638 static void
3639 moea64_sp_refchg_process(struct pvo_entry *sp, vm_page_t m,
3640     int64_t sp_refchg, vm_prot_t prot)
3641 {
3642 	vm_page_t m_end;
3643 	int64_t refchg;
3644 
3645 	if ((sp->pvo_vaddr & PVO_MANAGED) != 0 && (prot & VM_PROT_WRITE) != 0) {
3646 		for (m_end = &m[HPT_SP_PAGES]; m < m_end; m++) {
3647 			refchg = sp_refchg |
3648 			    atomic_readandclear_32(&m->md.mdpg_attrs);
3649 			if (refchg & LPTE_CHG)
3650 				vm_page_dirty(m);
3651 			if (refchg & LPTE_REF)
3652 				vm_page_aflag_set(m, PGA_REFERENCED);
3653 		}
3654 	}
3655 }
3656 
3657 /* Superpage ops */
3658 
3659 static int
3660 moea64_sp_enter(pmap_t pmap, vm_offset_t va, vm_page_t m,
3661     vm_prot_t prot, u_int flags, int8_t psind)
3662 {
3663 	struct pvo_entry *pvo, **pvos;
3664 	struct pvo_head *pvo_head;
3665 	vm_offset_t sva;
3666 	vm_page_t sm;
3667 	vm_paddr_t pa, spa;
3668 	bool sync;
3669 	struct pvo_dlist tofree;
3670 	int error, i;
3671 	uint16_t aflags;
3672 
3673 	KASSERT((va & HPT_SP_MASK) == 0, ("%s: va %#jx unaligned",
3674 	    __func__, (uintmax_t)va));
3675 	KASSERT(psind == 1, ("%s: invalid psind: %d", __func__, psind));
3676 	KASSERT(m->psind == 1, ("%s: invalid m->psind: %d",
3677 	    __func__, m->psind));
3678 	KASSERT(pmap != kernel_pmap,
3679 	    ("%s: function called with kernel pmap", __func__));
3680 
3681 	CTR5(KTR_PMAP, "%s: va=%#jx, pa=%#jx, prot=%#x, flags=%#x, psind=1",
3682 	    __func__, (uintmax_t)va, (uintmax_t)VM_PAGE_TO_PHYS(m),
3683 	    prot, flags);
3684 
3685 	SLIST_INIT(&tofree);
3686 
3687 	sva = va;
3688 	sm = m;
3689 	spa = pa = VM_PAGE_TO_PHYS(sm);
3690 
3691 	/* Try to allocate all PVOs first, to make failure handling easier. */
3692 	pvos = malloc(HPT_SP_PAGES * sizeof(struct pvo_entry *), M_TEMP,
3693 	    M_NOWAIT);
3694 	if (pvos == NULL) {
3695 		CTR1(KTR_PMAP, "%s: failed to alloc pvo array", __func__);
3696 		return (KERN_RESOURCE_SHORTAGE);
3697 	}
3698 
3699 	for (i = 0; i < HPT_SP_PAGES; i++) {
3700 		pvos[i] = alloc_pvo_entry(0);
3701 		if (pvos[i] == NULL) {
3702 			CTR1(KTR_PMAP, "%s: failed to alloc pvo", __func__);
3703 			for (i = i - 1; i >= 0; i--)
3704 				free_pvo_entry(pvos[i]);
3705 			free(pvos, M_TEMP);
3706 			return (KERN_RESOURCE_SHORTAGE);
3707 		}
3708 	}
3709 
3710 	SP_PV_LOCK_ALIGNED(spa);
3711 	PMAP_LOCK(pmap);
3712 
3713 	/* Note: moea64_remove_locked() also clears cached REF/CHG bits. */
3714 	moea64_remove_locked(pmap, va, va + HPT_SP_SIZE, &tofree);
3715 
3716 	/* Enter pages */
3717 	for (i = 0; i < HPT_SP_PAGES;
3718 	    i++, va += PAGE_SIZE, pa += PAGE_SIZE, m++) {
3719 		pvo = pvos[i];
3720 
3721 		pvo->pvo_pte.prot = prot;
3722 		pvo->pvo_pte.pa = (pa & ~HPT_SP_MASK) | LPTE_LP_4K_16M |
3723 		    moea64_calc_wimg(pa, pmap_page_get_memattr(m));
3724 
3725 		if ((flags & PMAP_ENTER_WIRED) != 0)
3726 			pvo->pvo_vaddr |= PVO_WIRED;
3727 		pvo->pvo_vaddr |= PVO_LARGE;
3728 
3729 		if ((m->oflags & VPO_UNMANAGED) != 0)
3730 			pvo_head = NULL;
3731 		else {
3732 			pvo_head = &m->md.mdpg_pvoh;
3733 			pvo->pvo_vaddr |= PVO_MANAGED;
3734 		}
3735 
3736 		init_pvo_entry(pvo, pmap, va);
3737 
3738 		error = moea64_pvo_enter(pvo, pvo_head, NULL);
3739 		/*
3740 		 * All superpage PVOs were previously removed, so no errors
3741 		 * should occur while inserting the new ones.
3742 		 */
3743 		KASSERT(error == 0, ("%s: unexpected error "
3744 			    "when inserting superpage PVO: %d",
3745 			    __func__, error));
3746 	}
3747 
3748 	PMAP_UNLOCK(pmap);
3749 	SP_PV_UNLOCK_ALIGNED(spa);
3750 
3751 	sync = (sm->a.flags & PGA_EXECUTABLE) == 0;
3752 	/* Note: moea64_pvo_cleanup() also clears page prot. flags. */
3753 	moea64_pvo_cleanup(&tofree);
3754 	pvo = pvos[0];
3755 
3756 	/* Set vm page flags */
3757 	aflags = pvo_to_vmpage_flags(pvo);
3758 	if (aflags != 0)
3759 		for (m = sm; m < &sm[HPT_SP_PAGES]; m++)
3760 			vm_page_aflag_set(m, aflags);
3761 
3762 	/*
3763 	 * Flush the page from the instruction cache if this page is
3764 	 * mapped executable and cacheable.
3765 	 */
3766 	if (sync && (pvo->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0)
3767 		moea64_syncicache(pmap, sva, spa, HPT_SP_SIZE);
3768 
3769 	atomic_add_long(&sp_mappings, 1);
3770 	CTR3(KTR_PMAP, "%s: SP success for va %#jx in pmap %p",
3771 	    __func__, (uintmax_t)sva, pmap);
3772 
3773 	free(pvos, M_TEMP);
3774 	return (KERN_SUCCESS);
3775 }
3776 
3777 static void
3778 moea64_sp_promote(pmap_t pmap, vm_offset_t va, vm_page_t m)
3779 {
3780 	struct pvo_entry *first, *pvo;
3781 	vm_paddr_t pa, pa_end;
3782 	vm_offset_t sva, va_end;
3783 	int64_t sp_refchg;
3784 
3785 	/* This CTR may generate a lot of output. */
3786 	/* CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)va); */
3787 
3788 	va &= ~HPT_SP_MASK;
3789 	sva = va;
3790 	/* Get superpage */
3791 	pa = VM_PAGE_TO_PHYS(m) & ~HPT_SP_MASK;
3792 	m = PHYS_TO_VM_PAGE(pa);
3793 
3794 	PMAP_LOCK(pmap);
3795 
3796 	/*
3797 	 * Check if all pages meet promotion criteria.
3798 	 *
3799 	 * XXX In some cases the loop below may be executed for each or most
3800 	 * of the entered pages of a superpage, which can be expensive
3801 	 * (although it was not profiled) and need some optimization.
3802 	 *
3803 	 * Some cases where this seems to happen are:
3804 	 * - When a superpage is first entered read-only and later becomes
3805 	 *   read-write.
3806 	 * - When some of the superpage's virtual addresses map to previously
3807 	 *   wired/cached pages while others map to pages allocated from a
3808 	 *   different physical address range. A common scenario where this
3809 	 *   happens is when mmap'ing a file that is already present in FS
3810 	 *   block cache and doesn't fill a superpage.
3811 	 */
3812 	first = pvo = moea64_pvo_find_va(pmap, sva);
3813 	for (pa_end = pa + HPT_SP_SIZE;
3814 	    pa < pa_end; pa += PAGE_SIZE, va += PAGE_SIZE) {
3815 		if (pvo == NULL || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
3816 			CTR3(KTR_PMAP,
3817 			    "%s: NULL or dead PVO: pmap=%p, va=%#jx",
3818 			    __func__, pmap, (uintmax_t)va);
3819 			goto error;
3820 		}
3821 		if (PVO_PADDR(pvo) != pa) {
3822 			CTR5(KTR_PMAP, "%s: PAs don't match: "
3823 			    "pmap=%p, va=%#jx, pvo_pa=%#jx, exp_pa=%#jx",
3824 			    __func__, pmap, (uintmax_t)va,
3825 			    (uintmax_t)PVO_PADDR(pvo), (uintmax_t)pa);
3826 			atomic_add_long(&sp_p_fail_pa, 1);
3827 			goto error;
3828 		}
3829 		if ((first->pvo_vaddr & PVO_FLAGS_PROMOTE) !=
3830 		    (pvo->pvo_vaddr & PVO_FLAGS_PROMOTE)) {
3831 			CTR5(KTR_PMAP, "%s: PVO flags don't match: "
3832 			    "pmap=%p, va=%#jx, pvo_flags=%#jx, exp_flags=%#jx",
3833 			    __func__, pmap, (uintmax_t)va,
3834 			    (uintmax_t)(pvo->pvo_vaddr & PVO_FLAGS_PROMOTE),
3835 			    (uintmax_t)(first->pvo_vaddr & PVO_FLAGS_PROMOTE));
3836 			atomic_add_long(&sp_p_fail_flags, 1);
3837 			goto error;
3838 		}
3839 		if (first->pvo_pte.prot != pvo->pvo_pte.prot) {
3840 			CTR5(KTR_PMAP, "%s: PVO protections don't match: "
3841 			    "pmap=%p, va=%#jx, pvo_prot=%#x, exp_prot=%#x",
3842 			    __func__, pmap, (uintmax_t)va,
3843 			    pvo->pvo_pte.prot, first->pvo_pte.prot);
3844 			atomic_add_long(&sp_p_fail_prot, 1);
3845 			goto error;
3846 		}
3847 		if ((first->pvo_pte.pa & LPTE_WIMG) !=
3848 		    (pvo->pvo_pte.pa & LPTE_WIMG)) {
3849 			CTR5(KTR_PMAP, "%s: WIMG bits don't match: "
3850 			    "pmap=%p, va=%#jx, pvo_wimg=%#jx, exp_wimg=%#jx",
3851 			    __func__, pmap, (uintmax_t)va,
3852 			    (uintmax_t)(pvo->pvo_pte.pa & LPTE_WIMG),
3853 			    (uintmax_t)(first->pvo_pte.pa & LPTE_WIMG));
3854 			atomic_add_long(&sp_p_fail_wimg, 1);
3855 			goto error;
3856 		}
3857 
3858 		pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo);
3859 	}
3860 
3861 	/* All OK, promote. */
3862 
3863 	/*
3864 	 * Handle superpage REF/CHG bits. If REF or CHG is set in
3865 	 * any page, then it must be set in the superpage.
3866 	 *
3867 	 * Instead of querying each page, we take advantage of two facts:
3868 	 * 1- If a page is being promoted, it was referenced.
3869 	 * 2- If promoted pages are writable, they were modified.
3870 	 */
3871 	sp_refchg = LPTE_REF |
3872 	    ((first->pvo_pte.prot & VM_PROT_WRITE) != 0 ? LPTE_CHG : 0);
3873 
3874 	/* Promote pages */
3875 
3876 	for (pvo = first, va_end = PVO_VADDR(pvo) + HPT_SP_SIZE;
3877 	    pvo != NULL && PVO_VADDR(pvo) < va_end;
3878 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo)) {
3879 		pvo->pvo_pte.pa &= ADDR_POFF | ~HPT_SP_MASK;
3880 		pvo->pvo_pte.pa |= LPTE_LP_4K_16M;
3881 		pvo->pvo_vaddr |= PVO_LARGE;
3882 	}
3883 	moea64_pte_replace_sp(first);
3884 
3885 	/* Send REF/CHG bits to VM */
3886 	moea64_sp_refchg_process(first, m, sp_refchg, first->pvo_pte.prot);
3887 
3888 	/* Use first page to cache REF/CHG bits */
3889 	atomic_set_32(&m->md.mdpg_attrs, sp_refchg | MDPG_ATTR_SP);
3890 
3891 	PMAP_UNLOCK(pmap);
3892 
3893 	atomic_add_long(&sp_mappings, 1);
3894 	atomic_add_long(&sp_promotions, 1);
3895 	CTR3(KTR_PMAP, "%s: success for va %#jx in pmap %p",
3896 	    __func__, (uintmax_t)sva, pmap);
3897 	return;
3898 
3899 error:
3900 	atomic_add_long(&sp_p_failures, 1);
3901 	PMAP_UNLOCK(pmap);
3902 }
3903 
3904 static void
3905 moea64_sp_demote_aligned(struct pvo_entry *sp)
3906 {
3907 	struct pvo_entry *pvo;
3908 	vm_offset_t va, va_end;
3909 	vm_paddr_t pa;
3910 	vm_page_t m;
3911 	pmap_t pmap;
3912 	int64_t refchg;
3913 
3914 	CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)PVO_VADDR(sp));
3915 
3916 	pmap = sp->pvo_pmap;
3917 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3918 
3919 	pvo = sp;
3920 
3921 	/* Demote pages */
3922 
3923 	va = PVO_VADDR(pvo);
3924 	pa = PVO_PADDR(pvo);
3925 	m = PHYS_TO_VM_PAGE(pa);
3926 
3927 	for (pvo = sp, va_end = va + HPT_SP_SIZE;
3928 	    pvo != NULL && PVO_VADDR(pvo) < va_end;
3929 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo),
3930 	    va += PAGE_SIZE, pa += PAGE_SIZE) {
3931 		KASSERT(pvo && PVO_VADDR(pvo) == va,
3932 		    ("%s: missing PVO for va %#jx", __func__, (uintmax_t)va));
3933 
3934 		pvo->pvo_vaddr &= ~PVO_LARGE;
3935 		pvo->pvo_pte.pa &= ~LPTE_RPGN;
3936 		pvo->pvo_pte.pa |= pa;
3937 
3938 	}
3939 	refchg = moea64_pte_replace_sp(sp);
3940 
3941 	/*
3942 	 * Clear SP flag
3943 	 *
3944 	 * XXX It is possible that another pmap has this page mapped as
3945 	 *     part of a superpage, but as the SP flag is used only for
3946 	 *     caching SP REF/CHG bits, that will be queried if not set
3947 	 *     in cache, it should be ok to clear it here.
3948 	 */
3949 	atomic_clear_32(&m->md.mdpg_attrs, MDPG_ATTR_SP);
3950 
3951 	/*
3952 	 * Handle superpage REF/CHG bits. A bit set in the superpage
3953 	 * means all pages should consider it set.
3954 	 */
3955 	moea64_sp_refchg_process(sp, m, refchg, sp->pvo_pte.prot);
3956 
3957 	atomic_add_long(&sp_demotions, 1);
3958 	CTR3(KTR_PMAP, "%s: success for va %#jx in pmap %p",
3959 	    __func__, (uintmax_t)PVO_VADDR(sp), pmap);
3960 }
3961 
3962 static void
3963 moea64_sp_demote(struct pvo_entry *pvo)
3964 {
3965 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
3966 
3967 	if ((PVO_VADDR(pvo) & HPT_SP_MASK) != 0) {
3968 		pvo = moea64_pvo_find_va(pvo->pvo_pmap,
3969 		    PVO_VADDR(pvo) & ~HPT_SP_MASK);
3970 		KASSERT(pvo != NULL, ("%s: missing PVO for va %#jx",
3971 		     __func__, (uintmax_t)(PVO_VADDR(pvo) & ~HPT_SP_MASK)));
3972 	}
3973 	moea64_sp_demote_aligned(pvo);
3974 }
3975 
3976 static struct pvo_entry *
3977 moea64_sp_unwire(struct pvo_entry *sp)
3978 {
3979 	struct pvo_entry *pvo, *prev;
3980 	vm_offset_t eva;
3981 	pmap_t pm;
3982 	int64_t ret, refchg;
3983 
3984 	CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)PVO_VADDR(sp));
3985 
3986 	pm = sp->pvo_pmap;
3987 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
3988 
3989 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
3990 	refchg = 0;
3991 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
3992 	    prev = pvo, pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
3993 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
3994 			panic("%s: pvo %p is missing PVO_WIRED",
3995 			    __func__, pvo);
3996 		pvo->pvo_vaddr &= ~PVO_WIRED;
3997 
3998 		ret = moea64_pte_replace(pvo, 0 /* No invalidation */);
3999 		if (ret < 0)
4000 			refchg |= LPTE_CHG;
4001 		else
4002 			refchg |= ret;
4003 
4004 		pm->pm_stats.wired_count--;
4005 	}
4006 
4007 	/* Send REF/CHG bits to VM */
4008 	moea64_sp_refchg_process(sp, PHYS_TO_VM_PAGE(PVO_PADDR(sp)),
4009 	    refchg, sp->pvo_pte.prot);
4010 
4011 	return (prev);
4012 }
4013 
4014 static struct pvo_entry *
4015 moea64_sp_protect(struct pvo_entry *sp, vm_prot_t prot)
4016 {
4017 	struct pvo_entry *pvo, *prev;
4018 	vm_offset_t eva;
4019 	pmap_t pm;
4020 	vm_page_t m, m_end;
4021 	int64_t ret, refchg;
4022 	vm_prot_t oldprot;
4023 
4024 	CTR3(KTR_PMAP, "%s: va=%#jx, prot=%x",
4025 	    __func__, (uintmax_t)PVO_VADDR(sp), prot);
4026 
4027 	pm = sp->pvo_pmap;
4028 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
4029 
4030 	oldprot = sp->pvo_pte.prot;
4031 	m = PHYS_TO_VM_PAGE(PVO_PADDR(sp));
4032 	KASSERT(m != NULL, ("%s: missing vm page for pa %#jx",
4033 	    __func__, (uintmax_t)PVO_PADDR(sp)));
4034 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4035 	refchg = 0;
4036 
4037 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4038 	    prev = pvo, pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
4039 		pvo->pvo_pte.prot = prot;
4040 		/*
4041 		 * If the PVO is in the page table, update mapping
4042 		 */
4043 		ret = moea64_pte_replace(pvo, MOEA64_PTE_PROT_UPDATE);
4044 		if (ret < 0)
4045 			refchg |= LPTE_CHG;
4046 		else
4047 			refchg |= ret;
4048 	}
4049 
4050 	/* Send REF/CHG bits to VM */
4051 	moea64_sp_refchg_process(sp, m, refchg, oldprot);
4052 
4053 	/* Handle pages that became executable */
4054 	if ((m->a.flags & PGA_EXECUTABLE) == 0 &&
4055 	    (sp->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
4056 		if ((m->oflags & VPO_UNMANAGED) == 0)
4057 			for (m_end = &m[HPT_SP_PAGES]; m < m_end; m++)
4058 				vm_page_aflag_set(m, PGA_EXECUTABLE);
4059 		moea64_syncicache(pm, PVO_VADDR(sp), PVO_PADDR(sp),
4060 		    HPT_SP_SIZE);
4061 	}
4062 
4063 	return (prev);
4064 }
4065 
4066 static struct pvo_entry *
4067 moea64_sp_remove(struct pvo_entry *sp, struct pvo_dlist *tofree)
4068 {
4069 	struct pvo_entry *pvo, *tpvo;
4070 	vm_offset_t eva;
4071 	pmap_t pm;
4072 
4073 	CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)PVO_VADDR(sp));
4074 
4075 	pm = sp->pvo_pmap;
4076 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
4077 
4078 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4079 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
4080 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
4081 
4082 		/*
4083 		 * For locking reasons, remove this from the page table and
4084 		 * pmap, but save delinking from the vm_page for a second
4085 		 * pass
4086 		 */
4087 		moea64_pvo_remove_from_pmap(pvo);
4088 		SLIST_INSERT_HEAD(tofree, pvo, pvo_dlink);
4089 	}
4090 
4091 	/*
4092 	 * Clear SP bit
4093 	 *
4094 	 * XXX See comment in moea64_sp_demote_aligned() for why it's
4095 	 *     ok to always clear the SP bit on remove/demote.
4096 	 */
4097 	atomic_clear_32(&PHYS_TO_VM_PAGE(PVO_PADDR(sp))->md.mdpg_attrs,
4098 	    MDPG_ATTR_SP);
4099 
4100 	return (tpvo);
4101 }
4102 
4103 static int64_t
4104 moea64_sp_query_locked(struct pvo_entry *pvo, uint64_t ptebit)
4105 {
4106 	int64_t refchg, ret;
4107 	vm_offset_t eva;
4108 	vm_page_t m;
4109 	pmap_t pmap;
4110 	struct pvo_entry *sp;
4111 
4112 	pmap = pvo->pvo_pmap;
4113 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4114 
4115 	/* Get first SP PVO */
4116 	if ((PVO_VADDR(pvo) & HPT_SP_MASK) != 0) {
4117 		sp = moea64_pvo_find_va(pmap, PVO_VADDR(pvo) & ~HPT_SP_MASK);
4118 		KASSERT(sp != NULL, ("%s: missing PVO for va %#jx",
4119 		     __func__, (uintmax_t)(PVO_VADDR(pvo) & ~HPT_SP_MASK)));
4120 	} else
4121 		sp = pvo;
4122 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4123 
4124 	refchg = 0;
4125 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4126 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo)) {
4127 		ret = moea64_pte_synch(pvo);
4128 		if (ret > 0) {
4129 			refchg |= ret & (LPTE_CHG | LPTE_REF);
4130 			if ((refchg & ptebit) != 0)
4131 				break;
4132 		}
4133 	}
4134 
4135 	/* Save results */
4136 	if (refchg != 0) {
4137 		m = PHYS_TO_VM_PAGE(PVO_PADDR(sp));
4138 		atomic_set_32(&m->md.mdpg_attrs, refchg | MDPG_ATTR_SP);
4139 	}
4140 
4141 	return (refchg);
4142 }
4143 
4144 static int64_t
4145 moea64_sp_query(struct pvo_entry *pvo, uint64_t ptebit)
4146 {
4147 	int64_t refchg;
4148 	pmap_t pmap;
4149 
4150 	pmap = pvo->pvo_pmap;
4151 	PMAP_LOCK(pmap);
4152 
4153 	/*
4154 	 * Check if SP was demoted/removed before pmap lock was acquired.
4155 	 */
4156 	if (!PVO_IS_SP(pvo) || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
4157 		CTR2(KTR_PMAP, "%s: demoted/removed: pa=%#jx",
4158 		    __func__, (uintmax_t)PVO_PADDR(pvo));
4159 		PMAP_UNLOCK(pmap);
4160 		return (-1);
4161 	}
4162 
4163 	refchg = moea64_sp_query_locked(pvo, ptebit);
4164 	PMAP_UNLOCK(pmap);
4165 
4166 	CTR4(KTR_PMAP, "%s: va=%#jx, pa=%#jx: refchg=%#jx",
4167 	    __func__, (uintmax_t)PVO_VADDR(pvo),
4168 	    (uintmax_t)PVO_PADDR(pvo), (uintmax_t)refchg);
4169 
4170 	return (refchg);
4171 }
4172 
4173 static int64_t
4174 moea64_sp_pvo_clear(struct pvo_entry *pvo, uint64_t ptebit)
4175 {
4176 	int64_t refchg, ret;
4177 	pmap_t pmap;
4178 	struct pvo_entry *sp;
4179 	vm_offset_t eva;
4180 	vm_page_t m;
4181 
4182 	pmap = pvo->pvo_pmap;
4183 	PMAP_LOCK(pmap);
4184 
4185 	/*
4186 	 * Check if SP was demoted/removed before pmap lock was acquired.
4187 	 */
4188 	if (!PVO_IS_SP(pvo) || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
4189 		CTR2(KTR_PMAP, "%s: demoted/removed: pa=%#jx",
4190 		    __func__, (uintmax_t)PVO_PADDR(pvo));
4191 		PMAP_UNLOCK(pmap);
4192 		return (-1);
4193 	}
4194 
4195 	/* Get first SP PVO */
4196 	if ((PVO_VADDR(pvo) & HPT_SP_MASK) != 0) {
4197 		sp = moea64_pvo_find_va(pmap, PVO_VADDR(pvo) & ~HPT_SP_MASK);
4198 		KASSERT(sp != NULL, ("%s: missing PVO for va %#jx",
4199 		     __func__, (uintmax_t)(PVO_VADDR(pvo) & ~HPT_SP_MASK)));
4200 	} else
4201 		sp = pvo;
4202 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4203 
4204 	refchg = 0;
4205 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4206 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo)) {
4207 		ret = moea64_pte_clear(pvo, ptebit);
4208 		if (ret > 0)
4209 			refchg |= ret & (LPTE_CHG | LPTE_REF);
4210 	}
4211 
4212 	m = PHYS_TO_VM_PAGE(PVO_PADDR(sp));
4213 	atomic_clear_32(&m->md.mdpg_attrs, ptebit);
4214 	PMAP_UNLOCK(pmap);
4215 
4216 	CTR4(KTR_PMAP, "%s: va=%#jx, pa=%#jx: refchg=%#jx",
4217 	    __func__, (uintmax_t)PVO_VADDR(sp),
4218 	    (uintmax_t)PVO_PADDR(sp), (uintmax_t)refchg);
4219 
4220 	return (refchg);
4221 }
4222 
4223 static int64_t
4224 moea64_sp_clear(struct pvo_entry *pvo, vm_page_t m, uint64_t ptebit)
4225 {
4226 	int64_t count, ret;
4227 	pmap_t pmap;
4228 
4229 	count = 0;
4230 	pmap = pvo->pvo_pmap;
4231 
4232 	/*
4233 	 * Since this reference bit is shared by 4096 4KB pages, it
4234 	 * should not be cleared every time it is tested. Apply a
4235 	 * simple "hash" function on the physical page number, the
4236 	 * virtual superpage number, and the pmap address to select
4237 	 * one 4KB page out of the 4096 on which testing the
4238 	 * reference bit will result in clearing that reference bit.
4239 	 * This function is designed to avoid the selection of the
4240 	 * same 4KB page for every 16MB page mapping.
4241 	 *
4242 	 * Always leave the reference bit of a wired mapping set, as
4243 	 * the current state of its reference bit won't affect page
4244 	 * replacement.
4245 	 */
4246 	if (ptebit == LPTE_REF && (((VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) ^
4247 	    (PVO_VADDR(pvo) >> HPT_SP_SHIFT) ^ (uintptr_t)pmap) &
4248 	    (HPT_SP_PAGES - 1)) == 0 && (pvo->pvo_vaddr & PVO_WIRED) == 0) {
4249 		if ((ret = moea64_sp_pvo_clear(pvo, ptebit)) == -1)
4250 			return (-1);
4251 
4252 		if ((ret & ptebit) != 0)
4253 			count++;
4254 
4255 	/*
4256 	 * If this page was not selected by the hash function, then assume
4257 	 * its REF bit was set.
4258 	 */
4259 	} else if (ptebit == LPTE_REF) {
4260 		count++;
4261 
4262 	/*
4263 	 * To clear the CHG bit of a single SP page, first it must be demoted.
4264 	 * But if no CHG bit is set, no bit clear and thus no SP demotion is
4265 	 * needed.
4266 	 */
4267 	} else {
4268 		CTR4(KTR_PMAP, "%s: ptebit=%#jx, va=%#jx, pa=%#jx",
4269 		    __func__, (uintmax_t)ptebit, (uintmax_t)PVO_VADDR(pvo),
4270 		    (uintmax_t)PVO_PADDR(pvo));
4271 
4272 		PMAP_LOCK(pmap);
4273 
4274 		/*
4275 		 * Make sure SP wasn't demoted/removed before pmap lock
4276 		 * was acquired.
4277 		 */
4278 		if (!PVO_IS_SP(pvo) || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
4279 			CTR2(KTR_PMAP, "%s: demoted/removed: pa=%#jx",
4280 			    __func__, (uintmax_t)PVO_PADDR(pvo));
4281 			PMAP_UNLOCK(pmap);
4282 			return (-1);
4283 		}
4284 
4285 		ret = moea64_sp_query_locked(pvo, ptebit);
4286 		if ((ret & ptebit) != 0)
4287 			count++;
4288 		else {
4289 			PMAP_UNLOCK(pmap);
4290 			return (0);
4291 		}
4292 
4293 		moea64_sp_demote(pvo);
4294 		moea64_pte_clear(pvo, ptebit);
4295 
4296 		/*
4297 		 * Write protect the mapping to a single page so that a
4298 		 * subsequent write access may repromote.
4299 		 */
4300 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
4301 			moea64_pvo_protect(pmap, pvo,
4302 			    pvo->pvo_pte.prot & ~VM_PROT_WRITE);
4303 
4304 		PMAP_UNLOCK(pmap);
4305 	}
4306 
4307 	return (count);
4308 }
4309