xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision e9ac41698b2f322d55ccf9da50a3596edb2c1800)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 /*
31  * Manages physical address maps.
32  *
33  * Since the information managed by this module is also stored by the
34  * logical address mapping module, this module may throw away valid virtual
35  * to physical mappings at almost any time.  However, invalidations of
36  * mappings must be done as requested.
37  *
38  * In order to cope with hardware architectures which make virtual to
39  * physical map invalidates expensive, this module may delay invalidate
40  * reduced protection operations until such time as they are actually
41  * necessary.  This module is given full information as to which processors
42  * are currently using which maps, and to when physical maps must be made
43  * correct.
44  */
45 
46 #include "opt_kstack_pages.h"
47 
48 #include <sys/param.h>
49 #include <sys/kernel.h>
50 #include <sys/conf.h>
51 #include <sys/queue.h>
52 #include <sys/cpuset.h>
53 #include <sys/kerneldump.h>
54 #include <sys/ktr.h>
55 #include <sys/lock.h>
56 #include <sys/msgbuf.h>
57 #include <sys/malloc.h>
58 #include <sys/mman.h>
59 #include <sys/mutex.h>
60 #include <sys/proc.h>
61 #include <sys/rwlock.h>
62 #include <sys/sched.h>
63 #include <sys/sysctl.h>
64 #include <sys/systm.h>
65 #include <sys/vmmeter.h>
66 #include <sys/smp.h>
67 #include <sys/reboot.h>
68 
69 #include <sys/kdb.h>
70 
71 #include <dev/ofw/openfirm.h>
72 
73 #include <vm/vm.h>
74 #include <vm/pmap.h>
75 #include <vm/vm_param.h>
76 #include <vm/vm_kern.h>
77 #include <vm/vm_page.h>
78 #include <vm/vm_phys.h>
79 #include <vm/vm_map.h>
80 #include <vm/vm_object.h>
81 #include <vm/vm_extern.h>
82 #include <vm/vm_pageout.h>
83 #include <vm/vm_dumpset.h>
84 #include <vm/vm_reserv.h>
85 #include <vm/uma.h>
86 
87 #include <machine/_inttypes.h>
88 #include <machine/cpu.h>
89 #include <machine/ifunc.h>
90 #include <machine/platform.h>
91 #include <machine/frame.h>
92 #include <machine/md_var.h>
93 #include <machine/psl.h>
94 #include <machine/bat.h>
95 #include <machine/hid.h>
96 #include <machine/pte.h>
97 #include <machine/sr.h>
98 #include <machine/trap.h>
99 #include <machine/mmuvar.h>
100 
101 #include "mmu_oea64.h"
102 
103 void moea64_release_vsid(uint64_t vsid);
104 uintptr_t moea64_get_unique_vsid(void);
105 
106 #define DISABLE_TRANS(msr)	msr = mfmsr(); mtmsr(msr & ~PSL_DR)
107 #define ENABLE_TRANS(msr)	mtmsr(msr)
108 
109 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
110 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
111 #define	VSID_HASH_MASK		0x0000007fffffffffULL
112 
113 /*
114  * Locking semantics:
115  *
116  * There are two locks of interest: the page locks and the pmap locks, which
117  * protect their individual PVO lists and are locked in that order. The contents
118  * of all PVO entries are protected by the locks of their respective pmaps.
119  * The pmap of any PVO is guaranteed not to change so long as the PVO is linked
120  * into any list.
121  *
122  */
123 
124 #define PV_LOCK_COUNT	PA_LOCK_COUNT
125 static struct mtx_padalign pv_lock[PV_LOCK_COUNT];
126 
127 /*
128  * Cheap NUMA-izing of the pv locks, to reduce contention across domains.
129  * NUMA domains on POWER9 appear to be indexed as sparse memory spaces, with the
130  * index at (N << 45).
131  */
132 #ifdef __powerpc64__
133 #define PV_LOCK_IDX(pa)	((pa_index(pa) * (((pa) >> 45) + 1)) % PV_LOCK_COUNT)
134 #else
135 #define PV_LOCK_IDX(pa)	(pa_index(pa) % PV_LOCK_COUNT)
136 #endif
137 #define PV_LOCKPTR(pa)	((struct mtx *)(&pv_lock[PV_LOCK_IDX(pa)]))
138 #define PV_LOCK(pa)		mtx_lock(PV_LOCKPTR(pa))
139 #define PV_UNLOCK(pa)		mtx_unlock(PV_LOCKPTR(pa))
140 #define PV_LOCKASSERT(pa) 	mtx_assert(PV_LOCKPTR(pa), MA_OWNED)
141 #define PV_PAGE_LOCK(m)		PV_LOCK(VM_PAGE_TO_PHYS(m))
142 #define PV_PAGE_UNLOCK(m)	PV_UNLOCK(VM_PAGE_TO_PHYS(m))
143 #define PV_PAGE_LOCKASSERT(m)	PV_LOCKASSERT(VM_PAGE_TO_PHYS(m))
144 
145 /* Superpage PV lock */
146 
147 #define	PV_LOCK_SIZE		(1<<PDRSHIFT)
148 
149 static __always_inline void
150 moea64_sp_pv_lock(vm_paddr_t pa)
151 {
152 	vm_paddr_t pa_end;
153 
154 	/* Note: breaking when pa_end is reached to avoid overflows */
155 	pa_end = pa + (HPT_SP_SIZE - PV_LOCK_SIZE);
156 	for (;;) {
157 		mtx_lock_flags(PV_LOCKPTR(pa), MTX_DUPOK);
158 		if (pa == pa_end)
159 			break;
160 		pa += PV_LOCK_SIZE;
161 	}
162 }
163 
164 static __always_inline void
165 moea64_sp_pv_unlock(vm_paddr_t pa)
166 {
167 	vm_paddr_t pa_end;
168 
169 	/* Note: breaking when pa_end is reached to avoid overflows */
170 	pa_end = pa;
171 	pa += HPT_SP_SIZE - PV_LOCK_SIZE;
172 	for (;;) {
173 		mtx_unlock_flags(PV_LOCKPTR(pa), MTX_DUPOK);
174 		if (pa == pa_end)
175 			break;
176 		pa -= PV_LOCK_SIZE;
177 	}
178 }
179 
180 #define	SP_PV_LOCK_ALIGNED(pa)		moea64_sp_pv_lock(pa)
181 #define	SP_PV_UNLOCK_ALIGNED(pa)	moea64_sp_pv_unlock(pa)
182 #define	SP_PV_LOCK(pa)			moea64_sp_pv_lock((pa) & ~HPT_SP_MASK)
183 #define	SP_PV_UNLOCK(pa)		moea64_sp_pv_unlock((pa) & ~HPT_SP_MASK)
184 #define	SP_PV_PAGE_LOCK(m)		SP_PV_LOCK(VM_PAGE_TO_PHYS(m))
185 #define	SP_PV_PAGE_UNLOCK(m)		SP_PV_UNLOCK(VM_PAGE_TO_PHYS(m))
186 
187 struct ofw_map {
188 	cell_t	om_va;
189 	cell_t	om_len;
190 	uint64_t om_pa;
191 	cell_t	om_mode;
192 };
193 
194 extern unsigned char _etext[];
195 extern unsigned char _end[];
196 
197 extern void *slbtrap, *slbtrapend;
198 
199 /*
200  * Map of physical memory regions.
201  */
202 static struct	mem_region *regions;
203 static struct	mem_region *pregions;
204 static struct	numa_mem_region *numa_pregions;
205 static u_int	phys_avail_count;
206 static int	regions_sz, pregions_sz, numapregions_sz;
207 
208 extern void bs_remap_earlyboot(void);
209 
210 /*
211  * Lock for the SLB tables.
212  */
213 struct mtx	moea64_slb_mutex;
214 
215 /*
216  * PTEG data.
217  */
218 u_long		moea64_pteg_count;
219 u_long		moea64_pteg_mask;
220 
221 /*
222  * PVO data.
223  */
224 
225 uma_zone_t	moea64_pvo_zone; /* zone for pvo entries */
226 
227 static struct	pvo_entry *moea64_bpvo_pool;
228 static int	moea64_bpvo_pool_index = 0;
229 static int	moea64_bpvo_pool_size = 0;
230 SYSCTL_INT(_machdep, OID_AUTO, moea64_allocated_bpvo_entries, CTLFLAG_RD,
231     &moea64_bpvo_pool_index, 0, "");
232 
233 #define	BPVO_POOL_SIZE	327680 /* Sensible historical default value */
234 #define	BPVO_POOL_EXPANSION_FACTOR	3
235 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
236 #ifdef __powerpc64__
237 #define	NVSIDS		(NPMAPS * 16)
238 #define VSID_HASHMASK	0xffffffffUL
239 #else
240 #define NVSIDS		NPMAPS
241 #define VSID_HASHMASK	0xfffffUL
242 #endif
243 static u_int	moea64_vsid_bitmap[NVSIDS / VSID_NBPW];
244 
245 static bool	moea64_initialized = false;
246 
247 #ifdef MOEA64_STATS
248 /*
249  * Statistics.
250  */
251 u_int	moea64_pte_valid = 0;
252 u_int	moea64_pte_overflow = 0;
253 u_int	moea64_pvo_entries = 0;
254 u_int	moea64_pvo_enter_calls = 0;
255 u_int	moea64_pvo_remove_calls = 0;
256 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_valid, CTLFLAG_RD,
257     &moea64_pte_valid, 0, "");
258 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_overflow, CTLFLAG_RD,
259     &moea64_pte_overflow, 0, "");
260 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_entries, CTLFLAG_RD,
261     &moea64_pvo_entries, 0, "");
262 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_enter_calls, CTLFLAG_RD,
263     &moea64_pvo_enter_calls, 0, "");
264 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_remove_calls, CTLFLAG_RD,
265     &moea64_pvo_remove_calls, 0, "");
266 #endif
267 
268 vm_offset_t	moea64_scratchpage_va[2];
269 struct pvo_entry *moea64_scratchpage_pvo[2];
270 struct	mtx	moea64_scratchpage_mtx;
271 
272 uint64_t 	moea64_large_page_mask = 0;
273 uint64_t	moea64_large_page_size = 0;
274 int		moea64_large_page_shift = 0;
275 bool		moea64_has_lp_4k_16m = false;
276 
277 /*
278  * PVO calls.
279  */
280 static int	moea64_pvo_enter(struct pvo_entry *pvo,
281 		    struct pvo_head *pvo_head, struct pvo_entry **oldpvo);
282 static void	moea64_pvo_remove_from_pmap(struct pvo_entry *pvo);
283 static void	moea64_pvo_remove_from_page(struct pvo_entry *pvo);
284 static void	moea64_pvo_remove_from_page_locked(
285 		    struct pvo_entry *pvo, vm_page_t m);
286 static struct	pvo_entry *moea64_pvo_find_va(pmap_t, vm_offset_t);
287 
288 /*
289  * Utility routines.
290  */
291 static bool		moea64_query_bit(vm_page_t, uint64_t);
292 static u_int		moea64_clear_bit(vm_page_t, uint64_t);
293 static void		moea64_kremove(vm_offset_t);
294 static void		moea64_syncicache(pmap_t pmap, vm_offset_t va,
295 			    vm_paddr_t pa, vm_size_t sz);
296 static void		moea64_pmap_init_qpages(void);
297 static void		moea64_remove_locked(pmap_t, vm_offset_t,
298 			    vm_offset_t, struct pvo_dlist *);
299 
300 /*
301  * Superpages data and routines.
302  */
303 
304 /*
305  * PVO flags (in vaddr) that must match for promotion to succeed.
306  * Note that protection bits are checked separately, as they reside in
307  * another field.
308  */
309 #define	PVO_FLAGS_PROMOTE	(PVO_WIRED | PVO_MANAGED | PVO_PTEGIDX_VALID)
310 
311 #define	PVO_IS_SP(pvo)		(((pvo)->pvo_vaddr & PVO_LARGE) && \
312 				 (pvo)->pvo_pmap != kernel_pmap)
313 
314 /* Get physical address from PVO. */
315 #define	PVO_PADDR(pvo)		moea64_pvo_paddr(pvo)
316 
317 /* MD page flag indicating that the page is a superpage. */
318 #define	MDPG_ATTR_SP		0x40000000
319 
320 SYSCTL_DECL(_vm_pmap);
321 
322 static SYSCTL_NODE(_vm_pmap, OID_AUTO, sp, CTLFLAG_RD, 0,
323     "SP page mapping counters");
324 
325 static u_long sp_demotions;
326 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, demotions, CTLFLAG_RD,
327     &sp_demotions, 0, "SP page demotions");
328 
329 static u_long sp_mappings;
330 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, mappings, CTLFLAG_RD,
331     &sp_mappings, 0, "SP page mappings");
332 
333 static u_long sp_p_failures;
334 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_failures, CTLFLAG_RD,
335     &sp_p_failures, 0, "SP page promotion failures");
336 
337 static u_long sp_p_fail_pa;
338 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_fail_pa, CTLFLAG_RD,
339     &sp_p_fail_pa, 0, "SP page promotion failure: PAs don't match");
340 
341 static u_long sp_p_fail_flags;
342 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_fail_flags, CTLFLAG_RD,
343     &sp_p_fail_flags, 0, "SP page promotion failure: page flags don't match");
344 
345 static u_long sp_p_fail_prot;
346 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_fail_prot, CTLFLAG_RD,
347     &sp_p_fail_prot, 0,
348     "SP page promotion failure: page protections don't match");
349 
350 static u_long sp_p_fail_wimg;
351 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_fail_wimg, CTLFLAG_RD,
352     &sp_p_fail_wimg, 0, "SP page promotion failure: WIMG bits don't match");
353 
354 static u_long sp_promotions;
355 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, promotions, CTLFLAG_RD,
356     &sp_promotions, 0, "SP page promotions");
357 
358 static bool moea64_ps_enabled(pmap_t);
359 static void moea64_align_superpage(vm_object_t, vm_ooffset_t,
360     vm_offset_t *, vm_size_t);
361 
362 static int moea64_sp_enter(pmap_t pmap, vm_offset_t va,
363     vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind);
364 static struct pvo_entry *moea64_sp_remove(struct pvo_entry *sp,
365     struct pvo_dlist *tofree);
366 
367 static void moea64_sp_promote(pmap_t pmap, vm_offset_t va, vm_page_t m);
368 static void moea64_sp_demote_aligned(struct pvo_entry *sp);
369 static void moea64_sp_demote(struct pvo_entry *pvo);
370 
371 static struct pvo_entry *moea64_sp_unwire(struct pvo_entry *sp);
372 static struct pvo_entry *moea64_sp_protect(struct pvo_entry *sp,
373     vm_prot_t prot);
374 
375 static int64_t moea64_sp_query(struct pvo_entry *pvo, uint64_t ptebit);
376 static int64_t moea64_sp_clear(struct pvo_entry *pvo, vm_page_t m,
377     uint64_t ptebit);
378 
379 static __inline bool moea64_sp_pvo_in_range(struct pvo_entry *pvo,
380     vm_offset_t sva, vm_offset_t eva);
381 
382 /*
383  * Kernel MMU interface
384  */
385 void moea64_clear_modify(vm_page_t);
386 void moea64_copy_page(vm_page_t, vm_page_t);
387 void moea64_copy_page_dmap(vm_page_t, vm_page_t);
388 void moea64_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
389     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
390 void moea64_copy_pages_dmap(vm_page_t *ma, vm_offset_t a_offset,
391     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
392 int moea64_enter(pmap_t, vm_offset_t, vm_page_t, vm_prot_t,
393     u_int flags, int8_t psind);
394 void moea64_enter_object(pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
395     vm_prot_t);
396 void moea64_enter_quick(pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
397 vm_paddr_t moea64_extract(pmap_t, vm_offset_t);
398 vm_page_t moea64_extract_and_hold(pmap_t, vm_offset_t, vm_prot_t);
399 void moea64_init(void);
400 bool moea64_is_modified(vm_page_t);
401 bool moea64_is_prefaultable(pmap_t, vm_offset_t);
402 bool moea64_is_referenced(vm_page_t);
403 int moea64_ts_referenced(vm_page_t);
404 vm_offset_t moea64_map(vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
405 bool moea64_page_exists_quick(pmap_t, vm_page_t);
406 void moea64_page_init(vm_page_t);
407 int moea64_page_wired_mappings(vm_page_t);
408 int moea64_pinit(pmap_t);
409 void moea64_pinit0(pmap_t);
410 void moea64_protect(pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
411 void moea64_qenter(vm_offset_t, vm_page_t *, int);
412 void moea64_qremove(vm_offset_t, int);
413 void moea64_release(pmap_t);
414 void moea64_remove(pmap_t, vm_offset_t, vm_offset_t);
415 void moea64_remove_pages(pmap_t);
416 void moea64_remove_all(vm_page_t);
417 void moea64_remove_write(vm_page_t);
418 void moea64_unwire(pmap_t, vm_offset_t, vm_offset_t);
419 void moea64_zero_page(vm_page_t);
420 void moea64_zero_page_dmap(vm_page_t);
421 void moea64_zero_page_area(vm_page_t, int, int);
422 void moea64_activate(struct thread *);
423 void moea64_deactivate(struct thread *);
424 void *moea64_mapdev(vm_paddr_t, vm_size_t);
425 void *moea64_mapdev_attr(vm_paddr_t, vm_size_t, vm_memattr_t);
426 void moea64_unmapdev(void *, vm_size_t);
427 vm_paddr_t moea64_kextract(vm_offset_t);
428 void moea64_page_set_memattr(vm_page_t m, vm_memattr_t ma);
429 void moea64_kenter_attr(vm_offset_t, vm_paddr_t, vm_memattr_t ma);
430 void moea64_kenter(vm_offset_t, vm_paddr_t);
431 int moea64_dev_direct_mapped(vm_paddr_t, vm_size_t);
432 static void moea64_sync_icache(pmap_t, vm_offset_t, vm_size_t);
433 void moea64_dumpsys_map(vm_paddr_t pa, size_t sz,
434     void **va);
435 void moea64_scan_init(void);
436 vm_offset_t moea64_quick_enter_page(vm_page_t m);
437 vm_offset_t moea64_quick_enter_page_dmap(vm_page_t m);
438 void moea64_quick_remove_page(vm_offset_t addr);
439 bool moea64_page_is_mapped(vm_page_t m);
440 static int moea64_map_user_ptr(pmap_t pm,
441     volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen);
442 static int moea64_decode_kernel_ptr(vm_offset_t addr,
443     int *is_user, vm_offset_t *decoded_addr);
444 static size_t moea64_scan_pmap(struct bitset *dump_bitset);
445 static void *moea64_dump_pmap_init(unsigned blkpgs);
446 #ifdef __powerpc64__
447 static void moea64_page_array_startup(long);
448 #endif
449 static int moea64_mincore(pmap_t, vm_offset_t, vm_paddr_t *);
450 
451 static struct pmap_funcs moea64_methods = {
452 	.clear_modify = moea64_clear_modify,
453 	.copy_page = moea64_copy_page,
454 	.copy_pages = moea64_copy_pages,
455 	.enter = moea64_enter,
456 	.enter_object = moea64_enter_object,
457 	.enter_quick = moea64_enter_quick,
458 	.extract = moea64_extract,
459 	.extract_and_hold = moea64_extract_and_hold,
460 	.init = moea64_init,
461 	.is_modified = moea64_is_modified,
462 	.is_prefaultable = moea64_is_prefaultable,
463 	.is_referenced = moea64_is_referenced,
464 	.ts_referenced = moea64_ts_referenced,
465 	.map =      		moea64_map,
466 	.mincore = moea64_mincore,
467 	.page_exists_quick = moea64_page_exists_quick,
468 	.page_init = moea64_page_init,
469 	.page_wired_mappings = moea64_page_wired_mappings,
470 	.pinit = moea64_pinit,
471 	.pinit0 = moea64_pinit0,
472 	.protect = moea64_protect,
473 	.qenter = moea64_qenter,
474 	.qremove = moea64_qremove,
475 	.release = moea64_release,
476 	.remove = moea64_remove,
477 	.remove_pages = moea64_remove_pages,
478 	.remove_all =       	moea64_remove_all,
479 	.remove_write = moea64_remove_write,
480 	.sync_icache = moea64_sync_icache,
481 	.unwire = moea64_unwire,
482 	.zero_page =        	moea64_zero_page,
483 	.zero_page_area = moea64_zero_page_area,
484 	.activate = moea64_activate,
485 	.deactivate =       	moea64_deactivate,
486 	.page_set_memattr = moea64_page_set_memattr,
487 	.quick_enter_page =  moea64_quick_enter_page,
488 	.quick_remove_page =  moea64_quick_remove_page,
489 	.page_is_mapped = moea64_page_is_mapped,
490 #ifdef __powerpc64__
491 	.page_array_startup = moea64_page_array_startup,
492 #endif
493 	.ps_enabled = moea64_ps_enabled,
494 	.align_superpage = moea64_align_superpage,
495 
496 	/* Internal interfaces */
497 	.mapdev = moea64_mapdev,
498 	.mapdev_attr = moea64_mapdev_attr,
499 	.unmapdev = moea64_unmapdev,
500 	.kextract = moea64_kextract,
501 	.kenter = moea64_kenter,
502 	.kenter_attr = moea64_kenter_attr,
503 	.dev_direct_mapped = moea64_dev_direct_mapped,
504 	.dumpsys_pa_init = moea64_scan_init,
505 	.dumpsys_scan_pmap = moea64_scan_pmap,
506 	.dumpsys_dump_pmap_init =    moea64_dump_pmap_init,
507 	.dumpsys_map_chunk = moea64_dumpsys_map,
508 	.map_user_ptr = moea64_map_user_ptr,
509 	.decode_kernel_ptr =  moea64_decode_kernel_ptr,
510 };
511 
512 MMU_DEF(oea64_mmu, "mmu_oea64_base", moea64_methods);
513 
514 /*
515  * Get physical address from PVO.
516  *
517  * For superpages, the lower bits are not stored on pvo_pte.pa and must be
518  * obtained from VA.
519  */
520 static __always_inline vm_paddr_t
521 moea64_pvo_paddr(struct pvo_entry *pvo)
522 {
523 	vm_paddr_t pa;
524 
525 	pa = (pvo)->pvo_pte.pa & LPTE_RPGN;
526 
527 	if (PVO_IS_SP(pvo)) {
528 		pa &= ~HPT_SP_MASK; /* This is needed to clear LPTE_LP bits. */
529 		pa |= PVO_VADDR(pvo) & HPT_SP_MASK;
530 	}
531 	return (pa);
532 }
533 
534 static struct pvo_head *
535 vm_page_to_pvoh(vm_page_t m)
536 {
537 
538 	mtx_assert(PV_LOCKPTR(VM_PAGE_TO_PHYS(m)), MA_OWNED);
539 	return (&m->md.mdpg_pvoh);
540 }
541 
542 static struct pvo_entry *
543 alloc_pvo_entry(int bootstrap)
544 {
545 	struct pvo_entry *pvo;
546 
547 	if (!moea64_initialized || bootstrap) {
548 		if (moea64_bpvo_pool_index >= moea64_bpvo_pool_size) {
549 			panic("%s: bpvo pool exhausted, index=%d, size=%d, bytes=%zd."
550 			    "Try setting machdep.moea64_bpvo_pool_size tunable",
551 			    __func__, moea64_bpvo_pool_index,
552 			    moea64_bpvo_pool_size,
553 			    moea64_bpvo_pool_size * sizeof(struct pvo_entry));
554 		}
555 		pvo = &moea64_bpvo_pool[
556 		    atomic_fetchadd_int(&moea64_bpvo_pool_index, 1)];
557 		bzero(pvo, sizeof(*pvo));
558 		pvo->pvo_vaddr = PVO_BOOTSTRAP;
559 	} else
560 		pvo = uma_zalloc(moea64_pvo_zone, M_NOWAIT | M_ZERO);
561 
562 	return (pvo);
563 }
564 
565 static void
566 init_pvo_entry(struct pvo_entry *pvo, pmap_t pmap, vm_offset_t va)
567 {
568 	uint64_t vsid;
569 	uint64_t hash;
570 	int shift;
571 
572 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
573 
574 	pvo->pvo_pmap = pmap;
575 	va &= ~ADDR_POFF;
576 	pvo->pvo_vaddr |= va;
577 	vsid = va_to_vsid(pmap, va);
578 	pvo->pvo_vpn = (uint64_t)((va & ADDR_PIDX) >> ADDR_PIDX_SHFT)
579 	    | (vsid << 16);
580 
581 	if (pmap == kernel_pmap && (pvo->pvo_vaddr & PVO_LARGE) != 0)
582 		shift = moea64_large_page_shift;
583 	else
584 		shift = ADDR_PIDX_SHFT;
585 	hash = (vsid & VSID_HASH_MASK) ^ (((uint64_t)va & ADDR_PIDX) >> shift);
586 	pvo->pvo_pte.slot = (hash & moea64_pteg_mask) << 3;
587 }
588 
589 static void
590 free_pvo_entry(struct pvo_entry *pvo)
591 {
592 
593 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
594 		uma_zfree(moea64_pvo_zone, pvo);
595 }
596 
597 void
598 moea64_pte_from_pvo(const struct pvo_entry *pvo, struct lpte *lpte)
599 {
600 
601 	lpte->pte_hi = moea64_pte_vpn_from_pvo_vpn(pvo);
602 	lpte->pte_hi |= LPTE_VALID;
603 
604 	if (pvo->pvo_vaddr & PVO_LARGE)
605 		lpte->pte_hi |= LPTE_BIG;
606 	if (pvo->pvo_vaddr & PVO_WIRED)
607 		lpte->pte_hi |= LPTE_WIRED;
608 	if (pvo->pvo_vaddr & PVO_HID)
609 		lpte->pte_hi |= LPTE_HID;
610 
611 	lpte->pte_lo = pvo->pvo_pte.pa; /* Includes WIMG bits */
612 	if (pvo->pvo_pte.prot & VM_PROT_WRITE)
613 		lpte->pte_lo |= LPTE_BW;
614 	else
615 		lpte->pte_lo |= LPTE_BR;
616 
617 	if (!(pvo->pvo_pte.prot & VM_PROT_EXECUTE))
618 		lpte->pte_lo |= LPTE_NOEXEC;
619 }
620 
621 static __inline uint64_t
622 moea64_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
623 {
624 	uint64_t pte_lo;
625 	int i;
626 
627 	if (ma != VM_MEMATTR_DEFAULT) {
628 		switch (ma) {
629 		case VM_MEMATTR_UNCACHEABLE:
630 			return (LPTE_I | LPTE_G);
631 		case VM_MEMATTR_CACHEABLE:
632 			return (LPTE_M);
633 		case VM_MEMATTR_WRITE_COMBINING:
634 		case VM_MEMATTR_WRITE_BACK:
635 		case VM_MEMATTR_PREFETCHABLE:
636 			return (LPTE_I);
637 		case VM_MEMATTR_WRITE_THROUGH:
638 			return (LPTE_W | LPTE_M);
639 		}
640 	}
641 
642 	/*
643 	 * Assume the page is cache inhibited and access is guarded unless
644 	 * it's in our available memory array.
645 	 */
646 	pte_lo = LPTE_I | LPTE_G;
647 	for (i = 0; i < pregions_sz; i++) {
648 		if ((pa >= pregions[i].mr_start) &&
649 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
650 			pte_lo &= ~(LPTE_I | LPTE_G);
651 			pte_lo |= LPTE_M;
652 			break;
653 		}
654 	}
655 
656 	return pte_lo;
657 }
658 
659 /*
660  * Quick sort callout for comparing memory regions.
661  */
662 static int	om_cmp(const void *a, const void *b);
663 
664 static int
665 om_cmp(const void *a, const void *b)
666 {
667 	const struct	ofw_map *mapa;
668 	const struct	ofw_map *mapb;
669 
670 	mapa = a;
671 	mapb = b;
672 	if (mapa->om_pa < mapb->om_pa)
673 		return (-1);
674 	else if (mapa->om_pa > mapb->om_pa)
675 		return (1);
676 	else
677 		return (0);
678 }
679 
680 static void
681 moea64_add_ofw_mappings(phandle_t mmu, size_t sz)
682 {
683 	struct ofw_map	translations[sz/(4*sizeof(cell_t))]; /*>= 4 cells per */
684 	pcell_t		acells, trans_cells[sz/sizeof(cell_t)];
685 	struct pvo_entry *pvo;
686 	register_t	msr;
687 	vm_offset_t	off;
688 	vm_paddr_t	pa_base;
689 	int		i, j;
690 
691 	bzero(translations, sz);
692 	OF_getencprop(OF_finddevice("/"), "#address-cells", &acells,
693 	    sizeof(acells));
694 	if (OF_getencprop(mmu, "translations", trans_cells, sz) == -1)
695 		panic("moea64_bootstrap: can't get ofw translations");
696 
697 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
698 	sz /= sizeof(cell_t);
699 	for (i = 0, j = 0; i < sz; j++) {
700 		translations[j].om_va = trans_cells[i++];
701 		translations[j].om_len = trans_cells[i++];
702 		translations[j].om_pa = trans_cells[i++];
703 		if (acells == 2) {
704 			translations[j].om_pa <<= 32;
705 			translations[j].om_pa |= trans_cells[i++];
706 		}
707 		translations[j].om_mode = trans_cells[i++];
708 	}
709 	KASSERT(i == sz, ("Translations map has incorrect cell count (%d/%zd)",
710 	    i, sz));
711 
712 	sz = j;
713 	qsort(translations, sz, sizeof (*translations), om_cmp);
714 
715 	for (i = 0; i < sz; i++) {
716 		pa_base = translations[i].om_pa;
717 	      #ifndef __powerpc64__
718 		if ((translations[i].om_pa >> 32) != 0)
719 			panic("OFW translations above 32-bit boundary!");
720 	      #endif
721 
722 		if (pa_base % PAGE_SIZE)
723 			panic("OFW translation not page-aligned (phys)!");
724 		if (translations[i].om_va % PAGE_SIZE)
725 			panic("OFW translation not page-aligned (virt)!");
726 
727 		CTR3(KTR_PMAP, "translation: pa=%#zx va=%#x len=%#x",
728 		    pa_base, translations[i].om_va, translations[i].om_len);
729 
730 		/* Now enter the pages for this mapping */
731 
732 		DISABLE_TRANS(msr);
733 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
734 			/* If this address is direct-mapped, skip remapping */
735 			if (hw_direct_map &&
736 			    translations[i].om_va == PHYS_TO_DMAP(pa_base) &&
737 			    moea64_calc_wimg(pa_base + off, VM_MEMATTR_DEFAULT)
738  			    == LPTE_M)
739 				continue;
740 
741 			PMAP_LOCK(kernel_pmap);
742 			pvo = moea64_pvo_find_va(kernel_pmap,
743 			    translations[i].om_va + off);
744 			PMAP_UNLOCK(kernel_pmap);
745 			if (pvo != NULL)
746 				continue;
747 
748 			moea64_kenter(translations[i].om_va + off,
749 			    pa_base + off);
750 		}
751 		ENABLE_TRANS(msr);
752 	}
753 }
754 
755 #ifdef __powerpc64__
756 static void
757 moea64_probe_large_page(void)
758 {
759 	uint16_t pvr = mfpvr() >> 16;
760 
761 	switch (pvr) {
762 	case IBM970:
763 	case IBM970FX:
764 	case IBM970MP:
765 		powerpc_sync(); isync();
766 		mtspr(SPR_HID4, mfspr(SPR_HID4) & ~HID4_970_DISABLE_LG_PG);
767 		powerpc_sync(); isync();
768 
769 		/* FALLTHROUGH */
770 	default:
771 		if (moea64_large_page_size == 0) {
772 			moea64_large_page_size = 0x1000000; /* 16 MB */
773 			moea64_large_page_shift = 24;
774 		}
775 	}
776 
777 	moea64_large_page_mask = moea64_large_page_size - 1;
778 }
779 
780 static void
781 moea64_bootstrap_slb_prefault(vm_offset_t va, int large)
782 {
783 	struct slb *cache;
784 	struct slb entry;
785 	uint64_t esid, slbe;
786 	uint64_t i;
787 
788 	cache = PCPU_GET(aim.slb);
789 	esid = va >> ADDR_SR_SHFT;
790 	slbe = (esid << SLBE_ESID_SHIFT) | SLBE_VALID;
791 
792 	for (i = 0; i < 64; i++) {
793 		if (cache[i].slbe == (slbe | i))
794 			return;
795 	}
796 
797 	entry.slbe = slbe;
798 	entry.slbv = KERNEL_VSID(esid) << SLBV_VSID_SHIFT;
799 	if (large)
800 		entry.slbv |= SLBV_L;
801 
802 	slb_insert_kernel(entry.slbe, entry.slbv);
803 }
804 #endif
805 
806 static int
807 moea64_kenter_large(vm_offset_t va, vm_paddr_t pa, uint64_t attr, int bootstrap)
808 {
809 	struct pvo_entry *pvo;
810 	uint64_t pte_lo;
811 	int error;
812 
813 	pte_lo = LPTE_M;
814 	pte_lo |= attr;
815 
816 	pvo = alloc_pvo_entry(bootstrap);
817 	pvo->pvo_vaddr |= PVO_WIRED | PVO_LARGE;
818 	init_pvo_entry(pvo, kernel_pmap, va);
819 
820 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE |
821 	    VM_PROT_EXECUTE;
822 	pvo->pvo_pte.pa = pa | pte_lo;
823 	error = moea64_pvo_enter(pvo, NULL, NULL);
824 	if (error != 0)
825 		panic("Error %d inserting large page\n", error);
826 	return (0);
827 }
828 
829 static void
830 moea64_setup_direct_map(vm_offset_t kernelstart,
831     vm_offset_t kernelend)
832 {
833 	register_t msr;
834 	vm_paddr_t pa, pkernelstart, pkernelend;
835 	vm_offset_t size, off;
836 	uint64_t pte_lo;
837 	int i;
838 
839 	if (moea64_large_page_size == 0)
840 		hw_direct_map = 0;
841 
842 	DISABLE_TRANS(msr);
843 	if (hw_direct_map) {
844 		PMAP_LOCK(kernel_pmap);
845 		for (i = 0; i < pregions_sz; i++) {
846 		  for (pa = pregions[i].mr_start; pa < pregions[i].mr_start +
847 		     pregions[i].mr_size; pa += moea64_large_page_size) {
848 			pte_lo = LPTE_M;
849 			if (pa & moea64_large_page_mask) {
850 				pa &= moea64_large_page_mask;
851 				pte_lo |= LPTE_G;
852 			}
853 			if (pa + moea64_large_page_size >
854 			    pregions[i].mr_start + pregions[i].mr_size)
855 				pte_lo |= LPTE_G;
856 
857 			moea64_kenter_large(PHYS_TO_DMAP(pa), pa, pte_lo, 1);
858 		  }
859 		}
860 		PMAP_UNLOCK(kernel_pmap);
861 	}
862 
863 	/*
864 	 * Make sure the kernel and BPVO pool stay mapped on systems either
865 	 * without a direct map or on which the kernel is not already executing
866 	 * out of the direct-mapped region.
867 	 */
868 	if (kernelstart < DMAP_BASE_ADDRESS) {
869 		/*
870 		 * For pre-dmap execution, we need to use identity mapping
871 		 * because we will be operating with the mmu on but in the
872 		 * wrong address configuration until we __restartkernel().
873 		 */
874 		for (pa = kernelstart & ~PAGE_MASK; pa < kernelend;
875 		    pa += PAGE_SIZE)
876 			moea64_kenter(pa, pa);
877 	} else if (!hw_direct_map) {
878 		pkernelstart = kernelstart & ~DMAP_BASE_ADDRESS;
879 		pkernelend = kernelend & ~DMAP_BASE_ADDRESS;
880 		for (pa = pkernelstart & ~PAGE_MASK; pa < pkernelend;
881 		    pa += PAGE_SIZE)
882 			moea64_kenter(pa | DMAP_BASE_ADDRESS, pa);
883 	}
884 
885 	if (!hw_direct_map) {
886 		size = moea64_bpvo_pool_size*sizeof(struct pvo_entry);
887 		off = (vm_offset_t)(moea64_bpvo_pool);
888 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
889 			moea64_kenter(pa, pa);
890 
891 		/* Map exception vectors */
892 		for (pa = EXC_RSVD; pa < EXC_LAST; pa += PAGE_SIZE)
893 			moea64_kenter(pa | DMAP_BASE_ADDRESS, pa);
894 	}
895 	ENABLE_TRANS(msr);
896 
897 	/*
898 	 * Allow user to override unmapped_buf_allowed for testing.
899 	 * XXXKIB Only direct map implementation was tested.
900 	 */
901 	if (!TUNABLE_INT_FETCH("vfs.unmapped_buf_allowed",
902 	    &unmapped_buf_allowed))
903 		unmapped_buf_allowed = hw_direct_map;
904 }
905 
906 /* Quick sort callout for comparing physical addresses. */
907 static int
908 pa_cmp(const void *a, const void *b)
909 {
910 	const vm_paddr_t *pa = a, *pb = b;
911 
912 	if (*pa < *pb)
913 		return (-1);
914 	else if (*pa > *pb)
915 		return (1);
916 	else
917 		return (0);
918 }
919 
920 void
921 moea64_early_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
922 {
923 	int		i, j;
924 	vm_size_t	physsz, hwphyssz;
925 	vm_paddr_t	kernelphysstart, kernelphysend;
926 	int		rm_pavail;
927 
928 	/* Level 0 reservations consist of 4096 pages (16MB superpage). */
929 	vm_level_0_order = 12;
930 
931 #ifndef __powerpc64__
932 	/* We don't have a direct map since there is no BAT */
933 	hw_direct_map = 0;
934 
935 	/* Make sure battable is zero, since we have no BAT */
936 	for (i = 0; i < 16; i++) {
937 		battable[i].batu = 0;
938 		battable[i].batl = 0;
939 	}
940 #else
941 	/* Install trap handlers for SLBs */
942 	bcopy(&slbtrap, (void *)EXC_DSE,(size_t)&slbtrapend - (size_t)&slbtrap);
943 	bcopy(&slbtrap, (void *)EXC_ISE,(size_t)&slbtrapend - (size_t)&slbtrap);
944 	__syncicache((void *)EXC_DSE, 0x80);
945 	__syncicache((void *)EXC_ISE, 0x80);
946 #endif
947 
948 	kernelphysstart = kernelstart & ~DMAP_BASE_ADDRESS;
949 	kernelphysend = kernelend & ~DMAP_BASE_ADDRESS;
950 
951 	/* Get physical memory regions from firmware */
952 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
953 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
954 
955 	if (PHYS_AVAIL_ENTRIES < regions_sz)
956 		panic("moea64_bootstrap: phys_avail too small");
957 
958 	phys_avail_count = 0;
959 	physsz = 0;
960 	hwphyssz = 0;
961 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
962 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
963 		CTR3(KTR_PMAP, "region: %#zx - %#zx (%#zx)",
964 		    regions[i].mr_start, regions[i].mr_start +
965 		    regions[i].mr_size, regions[i].mr_size);
966 		if (hwphyssz != 0 &&
967 		    (physsz + regions[i].mr_size) >= hwphyssz) {
968 			if (physsz < hwphyssz) {
969 				phys_avail[j] = regions[i].mr_start;
970 				phys_avail[j + 1] = regions[i].mr_start +
971 				    hwphyssz - physsz;
972 				physsz = hwphyssz;
973 				phys_avail_count++;
974 				dump_avail[j] = phys_avail[j];
975 				dump_avail[j + 1] = phys_avail[j + 1];
976 			}
977 			break;
978 		}
979 		phys_avail[j] = regions[i].mr_start;
980 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
981 		phys_avail_count++;
982 		physsz += regions[i].mr_size;
983 		dump_avail[j] = phys_avail[j];
984 		dump_avail[j + 1] = phys_avail[j + 1];
985 	}
986 
987 	/* Check for overlap with the kernel and exception vectors */
988 	rm_pavail = 0;
989 	for (j = 0; j < 2*phys_avail_count; j+=2) {
990 		if (phys_avail[j] < EXC_LAST)
991 			phys_avail[j] += EXC_LAST;
992 
993 		if (phys_avail[j] >= kernelphysstart &&
994 		    phys_avail[j+1] <= kernelphysend) {
995 			phys_avail[j] = phys_avail[j+1] = ~0;
996 			rm_pavail++;
997 			continue;
998 		}
999 
1000 		if (kernelphysstart >= phys_avail[j] &&
1001 		    kernelphysstart < phys_avail[j+1]) {
1002 			if (kernelphysend < phys_avail[j+1]) {
1003 				phys_avail[2*phys_avail_count] =
1004 				    (kernelphysend & ~PAGE_MASK) + PAGE_SIZE;
1005 				phys_avail[2*phys_avail_count + 1] =
1006 				    phys_avail[j+1];
1007 				phys_avail_count++;
1008 			}
1009 
1010 			phys_avail[j+1] = kernelphysstart & ~PAGE_MASK;
1011 		}
1012 
1013 		if (kernelphysend >= phys_avail[j] &&
1014 		    kernelphysend < phys_avail[j+1]) {
1015 			if (kernelphysstart > phys_avail[j]) {
1016 				phys_avail[2*phys_avail_count] = phys_avail[j];
1017 				phys_avail[2*phys_avail_count + 1] =
1018 				    kernelphysstart & ~PAGE_MASK;
1019 				phys_avail_count++;
1020 			}
1021 
1022 			phys_avail[j] = (kernelphysend & ~PAGE_MASK) +
1023 			    PAGE_SIZE;
1024 		}
1025 	}
1026 
1027 	/* Remove physical available regions marked for removal (~0) */
1028 	if (rm_pavail) {
1029 		qsort(phys_avail, 2*phys_avail_count, sizeof(phys_avail[0]),
1030 			pa_cmp);
1031 		phys_avail_count -= rm_pavail;
1032 		for (i = 2*phys_avail_count;
1033 		     i < 2*(phys_avail_count + rm_pavail); i+=2)
1034 			phys_avail[i] = phys_avail[i+1] = 0;
1035 	}
1036 
1037 	physmem = btoc(physsz);
1038 
1039 #ifdef PTEGCOUNT
1040 	moea64_pteg_count = PTEGCOUNT;
1041 #else
1042 	moea64_pteg_count = 0x1000;
1043 
1044 	while (moea64_pteg_count < physmem)
1045 		moea64_pteg_count <<= 1;
1046 
1047 	moea64_pteg_count >>= 1;
1048 #endif /* PTEGCOUNT */
1049 }
1050 
1051 void
1052 moea64_mid_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
1053 {
1054 	int		i;
1055 
1056 	/*
1057 	 * Set PTEG mask
1058 	 */
1059 	moea64_pteg_mask = moea64_pteg_count - 1;
1060 
1061 	/*
1062 	 * Initialize SLB table lock and page locks
1063 	 */
1064 	mtx_init(&moea64_slb_mutex, "SLB table", NULL, MTX_DEF);
1065 	for (i = 0; i < PV_LOCK_COUNT; i++)
1066 		mtx_init(&pv_lock[i], "page pv", NULL, MTX_DEF);
1067 
1068 	/*
1069 	 * Initialise the bootstrap pvo pool.
1070 	 */
1071 	TUNABLE_INT_FETCH("machdep.moea64_bpvo_pool_size", &moea64_bpvo_pool_size);
1072 	if (moea64_bpvo_pool_size == 0) {
1073 		if (!hw_direct_map)
1074 			moea64_bpvo_pool_size = ((ptoa((uintmax_t)physmem) * sizeof(struct vm_page)) /
1075 			    (PAGE_SIZE * PAGE_SIZE)) * BPVO_POOL_EXPANSION_FACTOR;
1076 		else
1077 			moea64_bpvo_pool_size = BPVO_POOL_SIZE;
1078 	}
1079 
1080 	if (boothowto & RB_VERBOSE) {
1081 		printf("mmu_oea64: bpvo pool entries = %d, bpvo pool size = %zu MB\n",
1082 		    moea64_bpvo_pool_size,
1083 		    moea64_bpvo_pool_size*sizeof(struct pvo_entry) / 1048576);
1084 	}
1085 
1086 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
1087 		moea64_bpvo_pool_size*sizeof(struct pvo_entry), PAGE_SIZE);
1088 	moea64_bpvo_pool_index = 0;
1089 
1090 	/* Place at address usable through the direct map */
1091 	if (hw_direct_map)
1092 		moea64_bpvo_pool = (struct pvo_entry *)
1093 		    PHYS_TO_DMAP((uintptr_t)moea64_bpvo_pool);
1094 
1095 	/*
1096 	 * Make sure kernel vsid is allocated as well as VSID 0.
1097 	 */
1098 	#ifndef __powerpc64__
1099 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NVSIDS - 1)) / VSID_NBPW]
1100 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
1101 	moea64_vsid_bitmap[0] |= 1;
1102 	#endif
1103 
1104 	/*
1105 	 * Initialize the kernel pmap (which is statically allocated).
1106 	 */
1107 	#ifdef __powerpc64__
1108 	for (i = 0; i < 64; i++) {
1109 		pcpup->pc_aim.slb[i].slbv = 0;
1110 		pcpup->pc_aim.slb[i].slbe = 0;
1111 	}
1112 	#else
1113 	for (i = 0; i < 16; i++)
1114 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
1115 	#endif
1116 
1117 	kernel_pmap->pmap_phys = kernel_pmap;
1118 	CPU_FILL(&kernel_pmap->pm_active);
1119 	RB_INIT(&kernel_pmap->pmap_pvo);
1120 
1121 	PMAP_LOCK_INIT(kernel_pmap);
1122 
1123 	/*
1124 	 * Now map in all the other buffers we allocated earlier
1125 	 */
1126 
1127 	moea64_setup_direct_map(kernelstart, kernelend);
1128 }
1129 
1130 void
1131 moea64_late_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
1132 {
1133 	ihandle_t	mmui;
1134 	phandle_t	chosen;
1135 	phandle_t	mmu;
1136 	ssize_t		sz;
1137 	int		i;
1138 	vm_offset_t	pa, va;
1139 	void		*dpcpu;
1140 
1141 	/*
1142 	 * Set up the Open Firmware pmap and add its mappings if not in real
1143 	 * mode.
1144 	 */
1145 
1146 	chosen = OF_finddevice("/chosen");
1147 	if (chosen != -1 && OF_getencprop(chosen, "mmu", &mmui, 4) != -1) {
1148 		mmu = OF_instance_to_package(mmui);
1149 		if (mmu == -1 ||
1150 		    (sz = OF_getproplen(mmu, "translations")) == -1)
1151 			sz = 0;
1152 		if (sz > 6144 /* tmpstksz - 2 KB headroom */)
1153 			panic("moea64_bootstrap: too many ofw translations");
1154 
1155 		if (sz > 0)
1156 			moea64_add_ofw_mappings(mmu, sz);
1157 	}
1158 
1159 	/*
1160 	 * Calculate the last available physical address.
1161 	 */
1162 	Maxmem = 0;
1163 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
1164 		Maxmem = MAX(Maxmem, powerpc_btop(phys_avail[i + 1]));
1165 
1166 	/*
1167 	 * Initialize MMU.
1168 	 */
1169 	pmap_cpu_bootstrap(0);
1170 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
1171 	pmap_bootstrapped++;
1172 
1173 	/*
1174 	 * Set the start and end of kva.
1175 	 */
1176 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
1177 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
1178 
1179 	/*
1180 	 * Map the entire KVA range into the SLB. We must not fault there.
1181 	 */
1182 	#ifdef __powerpc64__
1183 	for (va = virtual_avail; va < virtual_end; va += SEGMENT_LENGTH)
1184 		moea64_bootstrap_slb_prefault(va, 0);
1185 	#endif
1186 
1187 	/*
1188 	 * Remap any early IO mappings (console framebuffer, etc.)
1189 	 */
1190 	bs_remap_earlyboot();
1191 
1192 	/*
1193 	 * Figure out how far we can extend virtual_end into segment 16
1194 	 * without running into existing mappings. Segment 16 is guaranteed
1195 	 * to contain neither RAM nor devices (at least on Apple hardware),
1196 	 * but will generally contain some OFW mappings we should not
1197 	 * step on.
1198 	 */
1199 
1200 	#ifndef __powerpc64__	/* KVA is in high memory on PPC64 */
1201 	PMAP_LOCK(kernel_pmap);
1202 	while (virtual_end < VM_MAX_KERNEL_ADDRESS &&
1203 	    moea64_pvo_find_va(kernel_pmap, virtual_end+1) == NULL)
1204 		virtual_end += PAGE_SIZE;
1205 	PMAP_UNLOCK(kernel_pmap);
1206 	#endif
1207 
1208 	/*
1209 	 * Allocate a kernel stack with a guard page for thread0 and map it
1210 	 * into the kernel page map.
1211 	 */
1212 	pa = moea64_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
1213 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
1214 	virtual_avail = va + kstack_pages * PAGE_SIZE;
1215 	CTR2(KTR_PMAP, "moea64_bootstrap: kstack0 at %#x (%#x)", pa, va);
1216 	thread0.td_kstack = va;
1217 	thread0.td_kstack_pages = kstack_pages;
1218 	for (i = 0; i < kstack_pages; i++) {
1219 		moea64_kenter(va, pa);
1220 		pa += PAGE_SIZE;
1221 		va += PAGE_SIZE;
1222 	}
1223 
1224 	/*
1225 	 * Allocate virtual address space for the message buffer.
1226 	 */
1227 	pa = msgbuf_phys = moea64_bootstrap_alloc(msgbufsize, PAGE_SIZE);
1228 	msgbufp = (struct msgbuf *)virtual_avail;
1229 	va = virtual_avail;
1230 	virtual_avail += round_page(msgbufsize);
1231 	while (va < virtual_avail) {
1232 		moea64_kenter(va, pa);
1233 		pa += PAGE_SIZE;
1234 		va += PAGE_SIZE;
1235 	}
1236 
1237 	/*
1238 	 * Allocate virtual address space for the dynamic percpu area.
1239 	 */
1240 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
1241 	dpcpu = (void *)virtual_avail;
1242 	va = virtual_avail;
1243 	virtual_avail += DPCPU_SIZE;
1244 	while (va < virtual_avail) {
1245 		moea64_kenter(va, pa);
1246 		pa += PAGE_SIZE;
1247 		va += PAGE_SIZE;
1248 	}
1249 	dpcpu_init(dpcpu, curcpu);
1250 
1251 	crashdumpmap = (caddr_t)virtual_avail;
1252 	virtual_avail += MAXDUMPPGS * PAGE_SIZE;
1253 
1254 	/*
1255 	 * Allocate some things for page zeroing. We put this directly
1256 	 * in the page table and use MOEA64_PTE_REPLACE to avoid any
1257 	 * of the PVO book-keeping or other parts of the VM system
1258 	 * from even knowing that this hack exists.
1259 	 */
1260 
1261 	if (!hw_direct_map) {
1262 		mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL,
1263 		    MTX_DEF);
1264 		for (i = 0; i < 2; i++) {
1265 			moea64_scratchpage_va[i] = (virtual_end+1) - PAGE_SIZE;
1266 			virtual_end -= PAGE_SIZE;
1267 
1268 			moea64_kenter(moea64_scratchpage_va[i], 0);
1269 
1270 			PMAP_LOCK(kernel_pmap);
1271 			moea64_scratchpage_pvo[i] = moea64_pvo_find_va(
1272 			    kernel_pmap, (vm_offset_t)moea64_scratchpage_va[i]);
1273 			PMAP_UNLOCK(kernel_pmap);
1274 		}
1275 	}
1276 
1277 	numa_mem_regions(&numa_pregions, &numapregions_sz);
1278 }
1279 
1280 static void
1281 moea64_pmap_init_qpages(void)
1282 {
1283 	struct pcpu *pc;
1284 	int i;
1285 
1286 	if (hw_direct_map)
1287 		return;
1288 
1289 	CPU_FOREACH(i) {
1290 		pc = pcpu_find(i);
1291 		pc->pc_qmap_addr = kva_alloc(PAGE_SIZE);
1292 		if (pc->pc_qmap_addr == 0)
1293 			panic("pmap_init_qpages: unable to allocate KVA");
1294 		PMAP_LOCK(kernel_pmap);
1295 		pc->pc_aim.qmap_pvo =
1296 		    moea64_pvo_find_va(kernel_pmap, pc->pc_qmap_addr);
1297 		PMAP_UNLOCK(kernel_pmap);
1298 		mtx_init(&pc->pc_aim.qmap_lock, "qmap lock", NULL, MTX_DEF);
1299 	}
1300 }
1301 
1302 SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, moea64_pmap_init_qpages, NULL);
1303 
1304 /*
1305  * Activate a user pmap.  This mostly involves setting some non-CPU
1306  * state.
1307  */
1308 void
1309 moea64_activate(struct thread *td)
1310 {
1311 	pmap_t	pm;
1312 
1313 	pm = &td->td_proc->p_vmspace->vm_pmap;
1314 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1315 
1316 	#ifdef __powerpc64__
1317 	PCPU_SET(aim.userslb, pm->pm_slb);
1318 	__asm __volatile("slbmte %0, %1; isync" ::
1319 	    "r"(td->td_pcb->pcb_cpu.aim.usr_vsid), "r"(USER_SLB_SLBE));
1320 	#else
1321 	PCPU_SET(curpmap, pm->pmap_phys);
1322 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1323 	#endif
1324 }
1325 
1326 void
1327 moea64_deactivate(struct thread *td)
1328 {
1329 	pmap_t	pm;
1330 
1331 	__asm __volatile("isync; slbie %0" :: "r"(USER_ADDR));
1332 
1333 	pm = &td->td_proc->p_vmspace->vm_pmap;
1334 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1335 	#ifdef __powerpc64__
1336 	PCPU_SET(aim.userslb, NULL);
1337 	#else
1338 	PCPU_SET(curpmap, NULL);
1339 	#endif
1340 }
1341 
1342 void
1343 moea64_unwire(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1344 {
1345 	struct	pvo_entry key, *pvo;
1346 	vm_page_t m;
1347 	int64_t	refchg;
1348 
1349 	key.pvo_vaddr = sva;
1350 	PMAP_LOCK(pm);
1351 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1352 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1353 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1354 		if (PVO_IS_SP(pvo)) {
1355 			if (moea64_sp_pvo_in_range(pvo, sva, eva)) {
1356 				pvo = moea64_sp_unwire(pvo);
1357 				continue;
1358 			} else {
1359 				CTR1(KTR_PMAP, "%s: demote before unwire",
1360 				    __func__);
1361 				moea64_sp_demote(pvo);
1362 			}
1363 		}
1364 
1365 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1366 			panic("moea64_unwire: pvo %p is missing PVO_WIRED",
1367 			    pvo);
1368 		pvo->pvo_vaddr &= ~PVO_WIRED;
1369 		refchg = moea64_pte_replace(pvo, 0 /* No invalidation */);
1370 		if ((pvo->pvo_vaddr & PVO_MANAGED) &&
1371 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1372 			if (refchg < 0)
1373 				refchg = LPTE_CHG;
1374 			m = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
1375 
1376 			refchg |= atomic_readandclear_32(&m->md.mdpg_attrs);
1377 			if (refchg & LPTE_CHG)
1378 				vm_page_dirty(m);
1379 			if (refchg & LPTE_REF)
1380 				vm_page_aflag_set(m, PGA_REFERENCED);
1381 		}
1382 		pm->pm_stats.wired_count--;
1383 	}
1384 	PMAP_UNLOCK(pm);
1385 }
1386 
1387 static int
1388 moea64_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *pap)
1389 {
1390 	struct pvo_entry *pvo;
1391 	vm_paddr_t pa;
1392 	vm_page_t m;
1393 	int val;
1394 	bool managed;
1395 
1396 	PMAP_LOCK(pmap);
1397 
1398 	pvo = moea64_pvo_find_va(pmap, addr);
1399 	if (pvo != NULL) {
1400 		pa = PVO_PADDR(pvo);
1401 		m = PHYS_TO_VM_PAGE(pa);
1402 		managed = (pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED;
1403 		if (PVO_IS_SP(pvo))
1404 			val = MINCORE_INCORE | MINCORE_PSIND(1);
1405 		else
1406 			val = MINCORE_INCORE;
1407 	} else {
1408 		PMAP_UNLOCK(pmap);
1409 		return (0);
1410 	}
1411 
1412 	PMAP_UNLOCK(pmap);
1413 
1414 	if (m == NULL)
1415 		return (0);
1416 
1417 	if (managed) {
1418 		if (moea64_is_modified(m))
1419 			val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
1420 
1421 		if (moea64_is_referenced(m))
1422 			val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
1423 	}
1424 
1425 	if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) !=
1426 	    (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) &&
1427 	    managed) {
1428 		*pap = pa;
1429 	}
1430 
1431 	return (val);
1432 }
1433 
1434 /*
1435  * This goes through and sets the physical address of our
1436  * special scratch PTE to the PA we want to zero or copy. Because
1437  * of locking issues (this can get called in pvo_enter() by
1438  * the UMA allocator), we can't use most other utility functions here
1439  */
1440 
1441 static __inline
1442 void moea64_set_scratchpage_pa(int which, vm_paddr_t pa)
1443 {
1444 	struct pvo_entry *pvo;
1445 
1446 	KASSERT(!hw_direct_map, ("Using OEA64 scratchpage with a direct map!"));
1447 	mtx_assert(&moea64_scratchpage_mtx, MA_OWNED);
1448 
1449 	pvo = moea64_scratchpage_pvo[which];
1450 	PMAP_LOCK(pvo->pvo_pmap);
1451 	pvo->pvo_pte.pa =
1452 	    moea64_calc_wimg(pa, VM_MEMATTR_DEFAULT) | (uint64_t)pa;
1453 	moea64_pte_replace(pvo, MOEA64_PTE_INVALIDATE);
1454 	PMAP_UNLOCK(pvo->pvo_pmap);
1455 	isync();
1456 }
1457 
1458 void
1459 moea64_copy_page(vm_page_t msrc, vm_page_t mdst)
1460 {
1461 	mtx_lock(&moea64_scratchpage_mtx);
1462 
1463 	moea64_set_scratchpage_pa(0, VM_PAGE_TO_PHYS(msrc));
1464 	moea64_set_scratchpage_pa(1, VM_PAGE_TO_PHYS(mdst));
1465 
1466 	bcopy((void *)moea64_scratchpage_va[0],
1467 	    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1468 
1469 	mtx_unlock(&moea64_scratchpage_mtx);
1470 }
1471 
1472 void
1473 moea64_copy_page_dmap(vm_page_t msrc, vm_page_t mdst)
1474 {
1475 	vm_offset_t	dst;
1476 	vm_offset_t	src;
1477 
1478 	dst = VM_PAGE_TO_PHYS(mdst);
1479 	src = VM_PAGE_TO_PHYS(msrc);
1480 
1481 	bcopy((void *)PHYS_TO_DMAP(src), (void *)PHYS_TO_DMAP(dst),
1482 	    PAGE_SIZE);
1483 }
1484 
1485 inline void
1486 moea64_copy_pages_dmap(vm_page_t *ma, vm_offset_t a_offset,
1487     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1488 {
1489 	void *a_cp, *b_cp;
1490 	vm_offset_t a_pg_offset, b_pg_offset;
1491 	int cnt;
1492 
1493 	while (xfersize > 0) {
1494 		a_pg_offset = a_offset & PAGE_MASK;
1495 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1496 		a_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1497 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT])) +
1498 		    a_pg_offset;
1499 		b_pg_offset = b_offset & PAGE_MASK;
1500 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1501 		b_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1502 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT])) +
1503 		    b_pg_offset;
1504 		bcopy(a_cp, b_cp, cnt);
1505 		a_offset += cnt;
1506 		b_offset += cnt;
1507 		xfersize -= cnt;
1508 	}
1509 }
1510 
1511 void
1512 moea64_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
1513     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1514 {
1515 	void *a_cp, *b_cp;
1516 	vm_offset_t a_pg_offset, b_pg_offset;
1517 	int cnt;
1518 
1519 	mtx_lock(&moea64_scratchpage_mtx);
1520 	while (xfersize > 0) {
1521 		a_pg_offset = a_offset & PAGE_MASK;
1522 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1523 		moea64_set_scratchpage_pa(0,
1524 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]));
1525 		a_cp = (char *)moea64_scratchpage_va[0] + a_pg_offset;
1526 		b_pg_offset = b_offset & PAGE_MASK;
1527 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1528 		moea64_set_scratchpage_pa(1,
1529 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]));
1530 		b_cp = (char *)moea64_scratchpage_va[1] + b_pg_offset;
1531 		bcopy(a_cp, b_cp, cnt);
1532 		a_offset += cnt;
1533 		b_offset += cnt;
1534 		xfersize -= cnt;
1535 	}
1536 	mtx_unlock(&moea64_scratchpage_mtx);
1537 }
1538 
1539 void
1540 moea64_zero_page_area(vm_page_t m, int off, int size)
1541 {
1542 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1543 
1544 	if (size + off > PAGE_SIZE)
1545 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1546 
1547 	if (hw_direct_map) {
1548 		bzero((caddr_t)(uintptr_t)PHYS_TO_DMAP(pa) + off, size);
1549 	} else {
1550 		mtx_lock(&moea64_scratchpage_mtx);
1551 		moea64_set_scratchpage_pa(0, pa);
1552 		bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1553 		mtx_unlock(&moea64_scratchpage_mtx);
1554 	}
1555 }
1556 
1557 /*
1558  * Zero a page of physical memory by temporarily mapping it
1559  */
1560 void
1561 moea64_zero_page(vm_page_t m)
1562 {
1563 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1564 	vm_offset_t va, off;
1565 
1566 	mtx_lock(&moea64_scratchpage_mtx);
1567 
1568 	moea64_set_scratchpage_pa(0, pa);
1569 	va = moea64_scratchpage_va[0];
1570 
1571 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1572 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1573 
1574 	mtx_unlock(&moea64_scratchpage_mtx);
1575 }
1576 
1577 void
1578 moea64_zero_page_dmap(vm_page_t m)
1579 {
1580 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1581 	vm_offset_t va, off;
1582 
1583 	va = PHYS_TO_DMAP(pa);
1584 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1585 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1586 }
1587 
1588 vm_offset_t
1589 moea64_quick_enter_page(vm_page_t m)
1590 {
1591 	struct pvo_entry *pvo;
1592 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1593 
1594 	/*
1595  	 * MOEA64_PTE_REPLACE does some locking, so we can't just grab
1596 	 * a critical section and access the PCPU data like on i386.
1597 	 * Instead, pin the thread and grab the PCPU lock to prevent
1598 	 * a preempting thread from using the same PCPU data.
1599 	 */
1600 	sched_pin();
1601 
1602 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_NOTOWNED);
1603 	pvo = PCPU_GET(aim.qmap_pvo);
1604 
1605 	mtx_lock(PCPU_PTR(aim.qmap_lock));
1606 	pvo->pvo_pte.pa = moea64_calc_wimg(pa, pmap_page_get_memattr(m)) |
1607 	    (uint64_t)pa;
1608 	moea64_pte_replace(pvo, MOEA64_PTE_INVALIDATE);
1609 	isync();
1610 
1611 	return (PCPU_GET(qmap_addr));
1612 }
1613 
1614 vm_offset_t
1615 moea64_quick_enter_page_dmap(vm_page_t m)
1616 {
1617 
1618 	return (PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)));
1619 }
1620 
1621 void
1622 moea64_quick_remove_page(vm_offset_t addr)
1623 {
1624 
1625 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_OWNED);
1626 	KASSERT(PCPU_GET(qmap_addr) == addr,
1627 	    ("moea64_quick_remove_page: invalid address"));
1628 	mtx_unlock(PCPU_PTR(aim.qmap_lock));
1629 	sched_unpin();
1630 }
1631 
1632 bool
1633 moea64_page_is_mapped(vm_page_t m)
1634 {
1635 	return (!LIST_EMPTY(&(m)->md.mdpg_pvoh));
1636 }
1637 
1638 /*
1639  * Map the given physical page at the specified virtual address in the
1640  * target pmap with the protection requested.  If specified the page
1641  * will be wired down.
1642  */
1643 
1644 int
1645 moea64_enter(pmap_t pmap, vm_offset_t va, vm_page_t m,
1646     vm_prot_t prot, u_int flags, int8_t psind)
1647 {
1648 	struct		pvo_entry *pvo, *oldpvo, *tpvo;
1649 	struct		pvo_head *pvo_head;
1650 	uint64_t	pte_lo;
1651 	int		error;
1652 	vm_paddr_t	pa;
1653 
1654 	if ((m->oflags & VPO_UNMANAGED) == 0) {
1655 		if ((flags & PMAP_ENTER_QUICK_LOCKED) == 0)
1656 			VM_PAGE_OBJECT_BUSY_ASSERT(m);
1657 		else
1658 			VM_OBJECT_ASSERT_LOCKED(m->object);
1659 	}
1660 
1661 	if (psind > 0)
1662 		return (moea64_sp_enter(pmap, va, m, prot, flags, psind));
1663 
1664 	pvo = alloc_pvo_entry(0);
1665 	if (pvo == NULL)
1666 		return (KERN_RESOURCE_SHORTAGE);
1667 	pvo->pvo_pmap = NULL; /* to be filled in later */
1668 	pvo->pvo_pte.prot = prot;
1669 
1670 	pa = VM_PAGE_TO_PHYS(m);
1671 	pte_lo = moea64_calc_wimg(pa, pmap_page_get_memattr(m));
1672 	pvo->pvo_pte.pa = pa | pte_lo;
1673 
1674 	if ((flags & PMAP_ENTER_WIRED) != 0)
1675 		pvo->pvo_vaddr |= PVO_WIRED;
1676 
1677 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea64_initialized) {
1678 		pvo_head = NULL;
1679 	} else {
1680 		pvo_head = &m->md.mdpg_pvoh;
1681 		pvo->pvo_vaddr |= PVO_MANAGED;
1682 	}
1683 
1684 	PV_LOCK(pa);
1685 	PMAP_LOCK(pmap);
1686 	if (pvo->pvo_pmap == NULL)
1687 		init_pvo_entry(pvo, pmap, va);
1688 
1689 	if (moea64_ps_enabled(pmap) &&
1690 	    (tpvo = moea64_pvo_find_va(pmap, va & ~HPT_SP_MASK)) != NULL &&
1691 	    PVO_IS_SP(tpvo)) {
1692 		/* Demote SP before entering a regular page */
1693 		CTR2(KTR_PMAP, "%s: demote before enter: va=%#jx",
1694 		    __func__, (uintmax_t)va);
1695 		moea64_sp_demote_aligned(tpvo);
1696 	}
1697 
1698 	if (prot & VM_PROT_WRITE)
1699 		if (pmap_bootstrapped &&
1700 		    (m->oflags & VPO_UNMANAGED) == 0)
1701 			vm_page_aflag_set(m, PGA_WRITEABLE);
1702 
1703 	error = moea64_pvo_enter(pvo, pvo_head, &oldpvo);
1704 	if (error == EEXIST) {
1705 		if (oldpvo->pvo_vaddr == pvo->pvo_vaddr &&
1706 		    oldpvo->pvo_pte.pa == pvo->pvo_pte.pa &&
1707 		    oldpvo->pvo_pte.prot == prot) {
1708 			/* Identical mapping already exists */
1709 			error = 0;
1710 
1711 			/* If not in page table, reinsert it */
1712 			if (moea64_pte_synch(oldpvo) < 0) {
1713 				STAT_MOEA64(moea64_pte_overflow--);
1714 				moea64_pte_insert(oldpvo);
1715 			}
1716 
1717 			/* Then just clean up and go home */
1718 			PMAP_UNLOCK(pmap);
1719 			PV_UNLOCK(pa);
1720 			free_pvo_entry(pvo);
1721 			pvo = NULL;
1722 			goto out;
1723 		} else {
1724 			/* Otherwise, need to kill it first */
1725 			KASSERT(oldpvo->pvo_pmap == pmap, ("pmap of old "
1726 			    "mapping does not match new mapping"));
1727 			moea64_pvo_remove_from_pmap(oldpvo);
1728 			moea64_pvo_enter(pvo, pvo_head, NULL);
1729 		}
1730 	}
1731 	PMAP_UNLOCK(pmap);
1732 	PV_UNLOCK(pa);
1733 
1734 	/* Free any dead pages */
1735 	if (error == EEXIST) {
1736 		moea64_pvo_remove_from_page(oldpvo);
1737 		free_pvo_entry(oldpvo);
1738 	}
1739 
1740 out:
1741 	/*
1742 	 * Flush the page from the instruction cache if this page is
1743 	 * mapped executable and cacheable.
1744 	 */
1745 	if (pmap != kernel_pmap && (m->a.flags & PGA_EXECUTABLE) == 0 &&
1746 	    (pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1747 		vm_page_aflag_set(m, PGA_EXECUTABLE);
1748 		moea64_syncicache(pmap, va, pa, PAGE_SIZE);
1749 	}
1750 
1751 #if VM_NRESERVLEVEL > 0
1752 	/*
1753 	 * Try to promote pages.
1754 	 *
1755 	 * If the VA of the entered page is not aligned with its PA,
1756 	 * don't try page promotion as it is not possible.
1757 	 * This reduces the number of promotion failures dramatically.
1758 	 *
1759 	 * Ignore VM_PROT_NO_PROMOTE unless PMAP_ENTER_QUICK_LOCKED.
1760 	 */
1761 	if (moea64_ps_enabled(pmap) && pmap != kernel_pmap && pvo != NULL &&
1762 	    (pvo->pvo_vaddr & PVO_MANAGED) != 0 &&
1763 	    (va & HPT_SP_MASK) == (pa & HPT_SP_MASK) &&
1764 	    ((prot & VM_PROT_NO_PROMOTE) == 0 ||
1765 	    (flags & PMAP_ENTER_QUICK_LOCKED) == 0) &&
1766 	    (m->flags & PG_FICTITIOUS) == 0 &&
1767 	    vm_reserv_level_iffullpop(m) == 0)
1768 		moea64_sp_promote(pmap, va, m);
1769 #endif
1770 
1771 	return (KERN_SUCCESS);
1772 }
1773 
1774 static void
1775 moea64_syncicache(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
1776     vm_size_t sz)
1777 {
1778 
1779 	/*
1780 	 * This is much trickier than on older systems because
1781 	 * we can't sync the icache on physical addresses directly
1782 	 * without a direct map. Instead we check a couple of cases
1783 	 * where the memory is already mapped in and, failing that,
1784 	 * use the same trick we use for page zeroing to create
1785 	 * a temporary mapping for this physical address.
1786 	 */
1787 
1788 	if (!pmap_bootstrapped) {
1789 		/*
1790 		 * If PMAP is not bootstrapped, we are likely to be
1791 		 * in real mode.
1792 		 */
1793 		__syncicache((void *)(uintptr_t)pa, sz);
1794 	} else if (pmap == kernel_pmap) {
1795 		__syncicache((void *)va, sz);
1796 	} else if (hw_direct_map) {
1797 		__syncicache((void *)(uintptr_t)PHYS_TO_DMAP(pa), sz);
1798 	} else {
1799 		/* Use the scratch page to set up a temp mapping */
1800 
1801 		mtx_lock(&moea64_scratchpage_mtx);
1802 
1803 		moea64_set_scratchpage_pa(1, pa & ~ADDR_POFF);
1804 		__syncicache((void *)(moea64_scratchpage_va[1] +
1805 		    (va & ADDR_POFF)), sz);
1806 
1807 		mtx_unlock(&moea64_scratchpage_mtx);
1808 	}
1809 }
1810 
1811 /*
1812  * Maps a sequence of resident pages belonging to the same object.
1813  * The sequence begins with the given page m_start.  This page is
1814  * mapped at the given virtual address start.  Each subsequent page is
1815  * mapped at a virtual address that is offset from start by the same
1816  * amount as the page is offset from m_start within the object.  The
1817  * last page in the sequence is the page with the largest offset from
1818  * m_start that can be mapped at a virtual address less than the given
1819  * virtual address end.  Not every virtual page between start and end
1820  * is mapped; only those for which a resident page exists with the
1821  * corresponding offset from m_start are mapped.
1822  */
1823 void
1824 moea64_enter_object(pmap_t pm, vm_offset_t start, vm_offset_t end,
1825     vm_page_t m_start, vm_prot_t prot)
1826 {
1827 	vm_page_t m;
1828 	vm_pindex_t diff, psize;
1829 	vm_offset_t va;
1830 	int8_t psind;
1831 
1832 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1833 
1834 	psize = atop(end - start);
1835 	m = m_start;
1836 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1837 		va = start + ptoa(diff);
1838 		if ((va & HPT_SP_MASK) == 0 && va + HPT_SP_SIZE <= end &&
1839 		    m->psind == 1 && moea64_ps_enabled(pm))
1840 			psind = 1;
1841 		else
1842 			psind = 0;
1843 		moea64_enter(pm, va, m, prot &
1844 		    (VM_PROT_READ | VM_PROT_EXECUTE),
1845 		    PMAP_ENTER_NOSLEEP | PMAP_ENTER_QUICK_LOCKED, psind);
1846 		if (psind == 1)
1847 			m = &m[HPT_SP_SIZE / PAGE_SIZE - 1];
1848 		m = TAILQ_NEXT(m, listq);
1849 	}
1850 }
1851 
1852 void
1853 moea64_enter_quick(pmap_t pm, vm_offset_t va, vm_page_t m,
1854     vm_prot_t prot)
1855 {
1856 
1857 	moea64_enter(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE |
1858 	    VM_PROT_NO_PROMOTE), PMAP_ENTER_NOSLEEP | PMAP_ENTER_QUICK_LOCKED,
1859 	    0);
1860 }
1861 
1862 vm_paddr_t
1863 moea64_extract(pmap_t pm, vm_offset_t va)
1864 {
1865 	struct	pvo_entry *pvo;
1866 	vm_paddr_t pa;
1867 
1868 	PMAP_LOCK(pm);
1869 	pvo = moea64_pvo_find_va(pm, va);
1870 	if (pvo == NULL)
1871 		pa = 0;
1872 	else
1873 		pa = PVO_PADDR(pvo) | (va - PVO_VADDR(pvo));
1874 	PMAP_UNLOCK(pm);
1875 
1876 	return (pa);
1877 }
1878 
1879 /*
1880  * Atomically extract and hold the physical page with the given
1881  * pmap and virtual address pair if that mapping permits the given
1882  * protection.
1883  */
1884 vm_page_t
1885 moea64_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1886 {
1887 	struct	pvo_entry *pvo;
1888 	vm_page_t m;
1889 
1890 	m = NULL;
1891 	PMAP_LOCK(pmap);
1892 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1893 	if (pvo != NULL && (pvo->pvo_pte.prot & prot) == prot) {
1894 		m = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
1895 		if (!vm_page_wire_mapped(m))
1896 			m = NULL;
1897 	}
1898 	PMAP_UNLOCK(pmap);
1899 	return (m);
1900 }
1901 
1902 static void *
1903 moea64_uma_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain,
1904     uint8_t *flags, int wait)
1905 {
1906 	struct pvo_entry *pvo;
1907         vm_offset_t va;
1908         vm_page_t m;
1909         int needed_lock;
1910 
1911 	/*
1912 	 * This entire routine is a horrible hack to avoid bothering kmem
1913 	 * for new KVA addresses. Because this can get called from inside
1914 	 * kmem allocation routines, calling kmem for a new address here
1915 	 * can lead to multiply locking non-recursive mutexes.
1916 	 */
1917 
1918 	*flags = UMA_SLAB_PRIV;
1919 	needed_lock = !PMAP_LOCKED(kernel_pmap);
1920 
1921 	m = vm_page_alloc_noobj_domain(domain, malloc2vm_flags(wait) |
1922 	    VM_ALLOC_WIRED);
1923 	if (m == NULL)
1924 		return (NULL);
1925 
1926 	va = VM_PAGE_TO_PHYS(m);
1927 
1928 	pvo = alloc_pvo_entry(1 /* bootstrap */);
1929 
1930 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE;
1931 	pvo->pvo_pte.pa = VM_PAGE_TO_PHYS(m) | LPTE_M;
1932 
1933 	if (needed_lock)
1934 		PMAP_LOCK(kernel_pmap);
1935 
1936 	init_pvo_entry(pvo, kernel_pmap, va);
1937 	pvo->pvo_vaddr |= PVO_WIRED;
1938 
1939 	moea64_pvo_enter(pvo, NULL, NULL);
1940 
1941 	if (needed_lock)
1942 		PMAP_UNLOCK(kernel_pmap);
1943 
1944 	return (void *)va;
1945 }
1946 
1947 extern int elf32_nxstack;
1948 
1949 void
1950 moea64_init(void)
1951 {
1952 
1953 	CTR0(KTR_PMAP, "moea64_init");
1954 
1955 	moea64_pvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1956 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1957 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1958 
1959 	/*
1960 	 * Are large page mappings enabled?
1961 	 *
1962 	 * While HPT superpages are not better tested, leave it disabled by
1963 	 * default.
1964 	 */
1965 	superpages_enabled = 0;
1966 	TUNABLE_INT_FETCH("vm.pmap.superpages_enabled", &superpages_enabled);
1967 	if (superpages_enabled) {
1968 		KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0,
1969 		    ("moea64_init: can't assign to pagesizes[1]"));
1970 
1971 		if (moea64_large_page_size == 0) {
1972 			printf("mmu_oea64: HW does not support large pages. "
1973 					"Disabling superpages...\n");
1974 			superpages_enabled = 0;
1975 		} else if (!moea64_has_lp_4k_16m) {
1976 			printf("mmu_oea64: "
1977 			    "HW does not support mixed 4KB/16MB page sizes. "
1978 			    "Disabling superpages...\n");
1979 			superpages_enabled = 0;
1980 		} else
1981 			pagesizes[1] = HPT_SP_SIZE;
1982 	}
1983 
1984 	if (!hw_direct_map) {
1985 		uma_zone_set_allocf(moea64_pvo_zone, moea64_uma_page_alloc);
1986 	}
1987 
1988 #ifdef COMPAT_FREEBSD32
1989 	elf32_nxstack = 1;
1990 #endif
1991 
1992 	moea64_initialized = true;
1993 }
1994 
1995 bool
1996 moea64_is_referenced(vm_page_t m)
1997 {
1998 
1999 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2000 	    ("moea64_is_referenced: page %p is not managed", m));
2001 
2002 	return (moea64_query_bit(m, LPTE_REF));
2003 }
2004 
2005 bool
2006 moea64_is_modified(vm_page_t m)
2007 {
2008 
2009 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2010 	    ("moea64_is_modified: page %p is not managed", m));
2011 
2012 	/*
2013 	 * If the page is not busied then this check is racy.
2014 	 */
2015 	if (!pmap_page_is_write_mapped(m))
2016 		return (false);
2017 
2018 	return (moea64_query_bit(m, LPTE_CHG));
2019 }
2020 
2021 bool
2022 moea64_is_prefaultable(pmap_t pmap, vm_offset_t va)
2023 {
2024 	struct pvo_entry *pvo;
2025 	bool rv = true;
2026 
2027 	PMAP_LOCK(pmap);
2028 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
2029 	if (pvo != NULL)
2030 		rv = false;
2031 	PMAP_UNLOCK(pmap);
2032 	return (rv);
2033 }
2034 
2035 void
2036 moea64_clear_modify(vm_page_t m)
2037 {
2038 
2039 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2040 	    ("moea64_clear_modify: page %p is not managed", m));
2041 	vm_page_assert_busied(m);
2042 
2043 	if (!pmap_page_is_write_mapped(m))
2044 		return;
2045 	moea64_clear_bit(m, LPTE_CHG);
2046 }
2047 
2048 /*
2049  * Clear the write and modified bits in each of the given page's mappings.
2050  */
2051 void
2052 moea64_remove_write(vm_page_t m)
2053 {
2054 	struct	pvo_entry *pvo;
2055 	int64_t	refchg, ret;
2056 	pmap_t	pmap;
2057 
2058 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2059 	    ("moea64_remove_write: page %p is not managed", m));
2060 	vm_page_assert_busied(m);
2061 
2062 	if (!pmap_page_is_write_mapped(m))
2063 		return;
2064 
2065 	powerpc_sync();
2066 	PV_PAGE_LOCK(m);
2067 	refchg = 0;
2068 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2069 		pmap = pvo->pvo_pmap;
2070 		PMAP_LOCK(pmap);
2071 		if (!(pvo->pvo_vaddr & PVO_DEAD) &&
2072 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2073 			if (PVO_IS_SP(pvo)) {
2074 				CTR1(KTR_PMAP, "%s: demote before remwr",
2075 				    __func__);
2076 				moea64_sp_demote(pvo);
2077 			}
2078 			pvo->pvo_pte.prot &= ~VM_PROT_WRITE;
2079 			ret = moea64_pte_replace(pvo, MOEA64_PTE_PROT_UPDATE);
2080 			if (ret < 0)
2081 				ret = LPTE_CHG;
2082 			refchg |= ret;
2083 			if (pvo->pvo_pmap == kernel_pmap)
2084 				isync();
2085 		}
2086 		PMAP_UNLOCK(pmap);
2087 	}
2088 	if ((refchg | atomic_readandclear_32(&m->md.mdpg_attrs)) & LPTE_CHG)
2089 		vm_page_dirty(m);
2090 	vm_page_aflag_clear(m, PGA_WRITEABLE);
2091 	PV_PAGE_UNLOCK(m);
2092 }
2093 
2094 /*
2095  *	moea64_ts_referenced:
2096  *
2097  *	Return a count of reference bits for a page, clearing those bits.
2098  *	It is not necessary for every reference bit to be cleared, but it
2099  *	is necessary that 0 only be returned when there are truly no
2100  *	reference bits set.
2101  *
2102  *	XXX: The exact number of bits to check and clear is a matter that
2103  *	should be tested and standardized at some point in the future for
2104  *	optimal aging of shared pages.
2105  */
2106 int
2107 moea64_ts_referenced(vm_page_t m)
2108 {
2109 
2110 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2111 	    ("moea64_ts_referenced: page %p is not managed", m));
2112 	return (moea64_clear_bit(m, LPTE_REF));
2113 }
2114 
2115 /*
2116  * Modify the WIMG settings of all mappings for a page.
2117  */
2118 void
2119 moea64_page_set_memattr(vm_page_t m, vm_memattr_t ma)
2120 {
2121 	struct	pvo_entry *pvo;
2122 	int64_t	refchg;
2123 	pmap_t	pmap;
2124 	uint64_t lo;
2125 
2126 	CTR3(KTR_PMAP, "%s: pa=%#jx, ma=%#x",
2127 	    __func__, (uintmax_t)VM_PAGE_TO_PHYS(m), ma);
2128 
2129 	if ((m->oflags & VPO_UNMANAGED) != 0) {
2130 		m->md.mdpg_cache_attrs = ma;
2131 		return;
2132 	}
2133 
2134 	lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
2135 
2136 	PV_PAGE_LOCK(m);
2137 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2138 		pmap = pvo->pvo_pmap;
2139 		PMAP_LOCK(pmap);
2140 		if (!(pvo->pvo_vaddr & PVO_DEAD)) {
2141 			if (PVO_IS_SP(pvo)) {
2142 				CTR1(KTR_PMAP,
2143 				    "%s: demote before set_memattr", __func__);
2144 				moea64_sp_demote(pvo);
2145 			}
2146 			pvo->pvo_pte.pa &= ~LPTE_WIMG;
2147 			pvo->pvo_pte.pa |= lo;
2148 			refchg = moea64_pte_replace(pvo, MOEA64_PTE_INVALIDATE);
2149 			if (refchg < 0)
2150 				refchg = (pvo->pvo_pte.prot & VM_PROT_WRITE) ?
2151 				    LPTE_CHG : 0;
2152 			if ((pvo->pvo_vaddr & PVO_MANAGED) &&
2153 			    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2154 				refchg |=
2155 				    atomic_readandclear_32(&m->md.mdpg_attrs);
2156 				if (refchg & LPTE_CHG)
2157 					vm_page_dirty(m);
2158 				if (refchg & LPTE_REF)
2159 					vm_page_aflag_set(m, PGA_REFERENCED);
2160 			}
2161 			if (pvo->pvo_pmap == kernel_pmap)
2162 				isync();
2163 		}
2164 		PMAP_UNLOCK(pmap);
2165 	}
2166 	m->md.mdpg_cache_attrs = ma;
2167 	PV_PAGE_UNLOCK(m);
2168 }
2169 
2170 /*
2171  * Map a wired page into kernel virtual address space.
2172  */
2173 void
2174 moea64_kenter_attr(vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
2175 {
2176 	int		error;
2177 	struct pvo_entry *pvo, *oldpvo;
2178 
2179 	do {
2180 		pvo = alloc_pvo_entry(0);
2181 		if (pvo == NULL)
2182 			vm_wait(NULL);
2183 	} while (pvo == NULL);
2184 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;
2185 	pvo->pvo_pte.pa = (pa & ~ADDR_POFF) | moea64_calc_wimg(pa, ma);
2186 	pvo->pvo_vaddr |= PVO_WIRED;
2187 
2188 	PMAP_LOCK(kernel_pmap);
2189 	oldpvo = moea64_pvo_find_va(kernel_pmap, va);
2190 	if (oldpvo != NULL)
2191 		moea64_pvo_remove_from_pmap(oldpvo);
2192 	init_pvo_entry(pvo, kernel_pmap, va);
2193 	error = moea64_pvo_enter(pvo, NULL, NULL);
2194 	PMAP_UNLOCK(kernel_pmap);
2195 
2196 	/* Free any dead pages */
2197 	if (oldpvo != NULL) {
2198 		moea64_pvo_remove_from_page(oldpvo);
2199 		free_pvo_entry(oldpvo);
2200 	}
2201 
2202 	if (error != 0)
2203 		panic("moea64_kenter: failed to enter va %#zx pa %#jx: %d", va,
2204 		    (uintmax_t)pa, error);
2205 }
2206 
2207 void
2208 moea64_kenter(vm_offset_t va, vm_paddr_t pa)
2209 {
2210 
2211 	moea64_kenter_attr(va, pa, VM_MEMATTR_DEFAULT);
2212 }
2213 
2214 /*
2215  * Extract the physical page address associated with the given kernel virtual
2216  * address.
2217  */
2218 vm_paddr_t
2219 moea64_kextract(vm_offset_t va)
2220 {
2221 	struct		pvo_entry *pvo;
2222 	vm_paddr_t pa;
2223 
2224 	/*
2225 	 * Shortcut the direct-mapped case when applicable.  We never put
2226 	 * anything but 1:1 (or 62-bit aliased) mappings below
2227 	 * VM_MIN_KERNEL_ADDRESS.
2228 	 */
2229 	if (va < VM_MIN_KERNEL_ADDRESS)
2230 		return (va & ~DMAP_BASE_ADDRESS);
2231 
2232 	PMAP_LOCK(kernel_pmap);
2233 	pvo = moea64_pvo_find_va(kernel_pmap, va);
2234 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found for %#" PRIxPTR,
2235 	    va));
2236 	pa = PVO_PADDR(pvo) | (va - PVO_VADDR(pvo));
2237 	PMAP_UNLOCK(kernel_pmap);
2238 	return (pa);
2239 }
2240 
2241 /*
2242  * Remove a wired page from kernel virtual address space.
2243  */
2244 void
2245 moea64_kremove(vm_offset_t va)
2246 {
2247 	moea64_remove(kernel_pmap, va, va + PAGE_SIZE);
2248 }
2249 
2250 /*
2251  * Provide a kernel pointer corresponding to a given userland pointer.
2252  * The returned pointer is valid until the next time this function is
2253  * called in this thread. This is used internally in copyin/copyout.
2254  */
2255 static int
2256 moea64_map_user_ptr(pmap_t pm, volatile const void *uaddr,
2257     void **kaddr, size_t ulen, size_t *klen)
2258 {
2259 	size_t l;
2260 #ifdef __powerpc64__
2261 	struct slb *slb;
2262 #endif
2263 	register_t slbv;
2264 
2265 	*kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
2266 	l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
2267 	if (l > ulen)
2268 		l = ulen;
2269 	if (klen)
2270 		*klen = l;
2271 	else if (l != ulen)
2272 		return (EFAULT);
2273 
2274 #ifdef __powerpc64__
2275 	/* Try lockless look-up first */
2276 	slb = user_va_to_slb_entry(pm, (vm_offset_t)uaddr);
2277 
2278 	if (slb == NULL) {
2279 		/* If it isn't there, we need to pre-fault the VSID */
2280 		PMAP_LOCK(pm);
2281 		slbv = va_to_vsid(pm, (vm_offset_t)uaddr) << SLBV_VSID_SHIFT;
2282 		PMAP_UNLOCK(pm);
2283 	} else {
2284 		slbv = slb->slbv;
2285 	}
2286 
2287 	/* Mark segment no-execute */
2288 	slbv |= SLBV_N;
2289 #else
2290 	slbv = va_to_vsid(pm, (vm_offset_t)uaddr);
2291 
2292 	/* Mark segment no-execute */
2293 	slbv |= SR_N;
2294 #endif
2295 
2296 	/* If we have already set this VSID, we can just return */
2297 	if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == slbv)
2298 		return (0);
2299 
2300 	__asm __volatile("isync");
2301 	curthread->td_pcb->pcb_cpu.aim.usr_segm =
2302 	    (uintptr_t)uaddr >> ADDR_SR_SHFT;
2303 	curthread->td_pcb->pcb_cpu.aim.usr_vsid = slbv;
2304 #ifdef __powerpc64__
2305 	__asm __volatile ("slbie %0; slbmte %1, %2; isync" ::
2306 	    "r"(USER_ADDR), "r"(slbv), "r"(USER_SLB_SLBE));
2307 #else
2308 	__asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(slbv));
2309 #endif
2310 
2311 	return (0);
2312 }
2313 
2314 /*
2315  * Figure out where a given kernel pointer (usually in a fault) points
2316  * to from the VM's perspective, potentially remapping into userland's
2317  * address space.
2318  */
2319 static int
2320 moea64_decode_kernel_ptr(vm_offset_t addr, int *is_user,
2321     vm_offset_t *decoded_addr)
2322 {
2323 	vm_offset_t user_sr;
2324 
2325 	if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
2326 		user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
2327 		addr &= ADDR_PIDX | ADDR_POFF;
2328 		addr |= user_sr << ADDR_SR_SHFT;
2329 		*decoded_addr = addr;
2330 		*is_user = 1;
2331 	} else {
2332 		*decoded_addr = addr;
2333 		*is_user = 0;
2334 	}
2335 
2336 	return (0);
2337 }
2338 
2339 /*
2340  * Map a range of physical addresses into kernel virtual address space.
2341  *
2342  * The value passed in *virt is a suggested virtual address for the mapping.
2343  * Architectures which can support a direct-mapped physical to virtual region
2344  * can return the appropriate address within that region, leaving '*virt'
2345  * unchanged.  Other architectures should map the pages starting at '*virt' and
2346  * update '*virt' with the first usable address after the mapped region.
2347  */
2348 vm_offset_t
2349 moea64_map(vm_offset_t *virt, vm_paddr_t pa_start,
2350     vm_paddr_t pa_end, int prot)
2351 {
2352 	vm_offset_t	sva, va;
2353 
2354 	if (hw_direct_map) {
2355 		/*
2356 		 * Check if every page in the region is covered by the direct
2357 		 * map. The direct map covers all of physical memory. Use
2358 		 * moea64_calc_wimg() as a shortcut to see if the page is in
2359 		 * physical memory as a way to see if the direct map covers it.
2360 		 */
2361 		for (va = pa_start; va < pa_end; va += PAGE_SIZE)
2362 			if (moea64_calc_wimg(va, VM_MEMATTR_DEFAULT) != LPTE_M)
2363 				break;
2364 		if (va == pa_end)
2365 			return (PHYS_TO_DMAP(pa_start));
2366 	}
2367 	sva = *virt;
2368 	va = sva;
2369 	/* XXX respect prot argument */
2370 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
2371 		moea64_kenter(va, pa_start);
2372 	*virt = va;
2373 
2374 	return (sva);
2375 }
2376 
2377 /*
2378  * Returns true if the pmap's pv is one of the first
2379  * 16 pvs linked to from this page.  This count may
2380  * be changed upwards or downwards in the future; it
2381  * is only necessary that true be returned for a small
2382  * subset of pmaps for proper page aging.
2383  */
2384 bool
2385 moea64_page_exists_quick(pmap_t pmap, vm_page_t m)
2386 {
2387         int loops;
2388 	struct pvo_entry *pvo;
2389 	bool rv;
2390 
2391 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2392 	    ("moea64_page_exists_quick: page %p is not managed", m));
2393 	loops = 0;
2394 	rv = false;
2395 	PV_PAGE_LOCK(m);
2396 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2397 		if (!(pvo->pvo_vaddr & PVO_DEAD) && pvo->pvo_pmap == pmap) {
2398 			rv = true;
2399 			break;
2400 		}
2401 		if (++loops >= 16)
2402 			break;
2403 	}
2404 	PV_PAGE_UNLOCK(m);
2405 	return (rv);
2406 }
2407 
2408 void
2409 moea64_page_init(vm_page_t m)
2410 {
2411 
2412 	m->md.mdpg_attrs = 0;
2413 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
2414 	LIST_INIT(&m->md.mdpg_pvoh);
2415 }
2416 
2417 /*
2418  * Return the number of managed mappings to the given physical page
2419  * that are wired.
2420  */
2421 int
2422 moea64_page_wired_mappings(vm_page_t m)
2423 {
2424 	struct pvo_entry *pvo;
2425 	int count;
2426 
2427 	count = 0;
2428 	if ((m->oflags & VPO_UNMANAGED) != 0)
2429 		return (count);
2430 	PV_PAGE_LOCK(m);
2431 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
2432 		if ((pvo->pvo_vaddr & (PVO_DEAD | PVO_WIRED)) == PVO_WIRED)
2433 			count++;
2434 	PV_PAGE_UNLOCK(m);
2435 	return (count);
2436 }
2437 
2438 static uintptr_t	moea64_vsidcontext;
2439 
2440 uintptr_t
2441 moea64_get_unique_vsid(void) {
2442 	u_int entropy;
2443 	register_t hash;
2444 	uint32_t mask;
2445 	int i;
2446 
2447 	entropy = 0;
2448 	__asm __volatile("mftb %0" : "=r"(entropy));
2449 
2450 	mtx_lock(&moea64_slb_mutex);
2451 	for (i = 0; i < NVSIDS; i += VSID_NBPW) {
2452 		u_int	n;
2453 
2454 		/*
2455 		 * Create a new value by multiplying by a prime and adding in
2456 		 * entropy from the timebase register.  This is to make the
2457 		 * VSID more random so that the PT hash function collides
2458 		 * less often.  (Note that the prime casues gcc to do shifts
2459 		 * instead of a multiply.)
2460 		 */
2461 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
2462 		hash = moea64_vsidcontext & (NVSIDS - 1);
2463 		if (hash == 0)		/* 0 is special, avoid it */
2464 			continue;
2465 		n = hash >> 5;
2466 		mask = 1 << (hash & (VSID_NBPW - 1));
2467 		hash = (moea64_vsidcontext & VSID_HASHMASK);
2468 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
2469 			/* anything free in this bucket? */
2470 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
2471 				entropy = (moea64_vsidcontext >> 20);
2472 				continue;
2473 			}
2474 			i = ffs(~moea64_vsid_bitmap[n]) - 1;
2475 			mask = 1 << i;
2476 			hash &= rounddown2(VSID_HASHMASK, VSID_NBPW);
2477 			hash |= i;
2478 		}
2479 		if (hash == VSID_VRMA)	/* also special, avoid this too */
2480 			continue;
2481 		KASSERT(!(moea64_vsid_bitmap[n] & mask),
2482 		    ("Allocating in-use VSID %#zx\n", hash));
2483 		moea64_vsid_bitmap[n] |= mask;
2484 		mtx_unlock(&moea64_slb_mutex);
2485 		return (hash);
2486 	}
2487 
2488 	mtx_unlock(&moea64_slb_mutex);
2489 	panic("%s: out of segments",__func__);
2490 }
2491 
2492 #ifdef __powerpc64__
2493 int
2494 moea64_pinit(pmap_t pmap)
2495 {
2496 
2497 	RB_INIT(&pmap->pmap_pvo);
2498 
2499 	pmap->pm_slb_tree_root = slb_alloc_tree();
2500 	pmap->pm_slb = slb_alloc_user_cache();
2501 	pmap->pm_slb_len = 0;
2502 
2503 	return (1);
2504 }
2505 #else
2506 int
2507 moea64_pinit(pmap_t pmap)
2508 {
2509 	int	i;
2510 	uint32_t hash;
2511 
2512 	RB_INIT(&pmap->pmap_pvo);
2513 
2514 	if (pmap_bootstrapped)
2515 		pmap->pmap_phys = (pmap_t)moea64_kextract((vm_offset_t)pmap);
2516 	else
2517 		pmap->pmap_phys = pmap;
2518 
2519 	/*
2520 	 * Allocate some segment registers for this pmap.
2521 	 */
2522 	hash = moea64_get_unique_vsid();
2523 
2524 	for (i = 0; i < 16; i++)
2525 		pmap->pm_sr[i] = VSID_MAKE(i, hash);
2526 
2527 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_pinit: pm_sr[0] = 0"));
2528 
2529 	return (1);
2530 }
2531 #endif
2532 
2533 /*
2534  * Initialize the pmap associated with process 0.
2535  */
2536 void
2537 moea64_pinit0(pmap_t pm)
2538 {
2539 
2540 	PMAP_LOCK_INIT(pm);
2541 	moea64_pinit(pm);
2542 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
2543 }
2544 
2545 /*
2546  * Set the physical protection on the specified range of this map as requested.
2547  */
2548 static void
2549 moea64_pvo_protect( pmap_t pm, struct pvo_entry *pvo, vm_prot_t prot)
2550 {
2551 	struct vm_page *pg;
2552 	vm_prot_t oldprot;
2553 	int32_t refchg;
2554 
2555 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2556 
2557 	/*
2558 	 * Change the protection of the page.
2559 	 */
2560 	oldprot = pvo->pvo_pte.prot;
2561 	pvo->pvo_pte.prot = prot;
2562 	pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
2563 
2564 	/*
2565 	 * If the PVO is in the page table, update mapping
2566 	 */
2567 	refchg = moea64_pte_replace(pvo, MOEA64_PTE_PROT_UPDATE);
2568 	if (refchg < 0)
2569 		refchg = (oldprot & VM_PROT_WRITE) ? LPTE_CHG : 0;
2570 
2571 	if (pm != kernel_pmap && pg != NULL &&
2572 	    (pg->a.flags & PGA_EXECUTABLE) == 0 &&
2573 	    (pvo->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
2574 		if ((pg->oflags & VPO_UNMANAGED) == 0)
2575 			vm_page_aflag_set(pg, PGA_EXECUTABLE);
2576 		moea64_syncicache(pm, PVO_VADDR(pvo),
2577 		    PVO_PADDR(pvo), PAGE_SIZE);
2578 	}
2579 
2580 	/*
2581 	 * Update vm about the REF/CHG bits if the page is managed and we have
2582 	 * removed write access.
2583 	 */
2584 	if (pg != NULL && (pvo->pvo_vaddr & PVO_MANAGED) &&
2585 	    (oldprot & VM_PROT_WRITE)) {
2586 		refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2587 		if (refchg & LPTE_CHG)
2588 			vm_page_dirty(pg);
2589 		if (refchg & LPTE_REF)
2590 			vm_page_aflag_set(pg, PGA_REFERENCED);
2591 	}
2592 }
2593 
2594 void
2595 moea64_protect(pmap_t pm, vm_offset_t sva, vm_offset_t eva,
2596     vm_prot_t prot)
2597 {
2598 	struct	pvo_entry *pvo, key;
2599 
2600 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm,
2601 	    sva, eva, prot);
2602 
2603 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
2604 	    ("moea64_protect: non current pmap"));
2605 
2606 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2607 		moea64_remove(pm, sva, eva);
2608 		return;
2609 	}
2610 
2611 	PMAP_LOCK(pm);
2612 	key.pvo_vaddr = sva;
2613 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2614 	    pvo != NULL && PVO_VADDR(pvo) < eva;
2615 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
2616 		if (PVO_IS_SP(pvo)) {
2617 			if (moea64_sp_pvo_in_range(pvo, sva, eva)) {
2618 				pvo = moea64_sp_protect(pvo, prot);
2619 				continue;
2620 			} else {
2621 				CTR1(KTR_PMAP, "%s: demote before protect",
2622 				    __func__);
2623 				moea64_sp_demote(pvo);
2624 			}
2625 		}
2626 		moea64_pvo_protect(pm, pvo, prot);
2627 	}
2628 	PMAP_UNLOCK(pm);
2629 }
2630 
2631 /*
2632  * Map a list of wired pages into kernel virtual address space.  This is
2633  * intended for temporary mappings which do not need page modification or
2634  * references recorded.  Existing mappings in the region are overwritten.
2635  */
2636 void
2637 moea64_qenter(vm_offset_t va, vm_page_t *m, int count)
2638 {
2639 	while (count-- > 0) {
2640 		moea64_kenter(va, VM_PAGE_TO_PHYS(*m));
2641 		va += PAGE_SIZE;
2642 		m++;
2643 	}
2644 }
2645 
2646 /*
2647  * Remove page mappings from kernel virtual address space.  Intended for
2648  * temporary mappings entered by moea64_qenter.
2649  */
2650 void
2651 moea64_qremove(vm_offset_t va, int count)
2652 {
2653 	while (count-- > 0) {
2654 		moea64_kremove(va);
2655 		va += PAGE_SIZE;
2656 	}
2657 }
2658 
2659 void
2660 moea64_release_vsid(uint64_t vsid)
2661 {
2662 	int idx, mask;
2663 
2664 	mtx_lock(&moea64_slb_mutex);
2665 	idx = vsid & (NVSIDS-1);
2666 	mask = 1 << (idx % VSID_NBPW);
2667 	idx /= VSID_NBPW;
2668 	KASSERT(moea64_vsid_bitmap[idx] & mask,
2669 	    ("Freeing unallocated VSID %#jx", vsid));
2670 	moea64_vsid_bitmap[idx] &= ~mask;
2671 	mtx_unlock(&moea64_slb_mutex);
2672 }
2673 
2674 void
2675 moea64_release(pmap_t pmap)
2676 {
2677 
2678 	/*
2679 	 * Free segment registers' VSIDs
2680 	 */
2681     #ifdef __powerpc64__
2682 	slb_free_tree(pmap);
2683 	slb_free_user_cache(pmap->pm_slb);
2684     #else
2685 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_release: pm_sr[0] = 0"));
2686 
2687 	moea64_release_vsid(VSID_TO_HASH(pmap->pm_sr[0]));
2688     #endif
2689 }
2690 
2691 /*
2692  * Remove all pages mapped by the specified pmap
2693  */
2694 void
2695 moea64_remove_pages(pmap_t pm)
2696 {
2697 	struct pvo_entry *pvo, *tpvo;
2698 	struct pvo_dlist tofree;
2699 
2700 	SLIST_INIT(&tofree);
2701 
2702 	PMAP_LOCK(pm);
2703 	RB_FOREACH_SAFE(pvo, pvo_tree, &pm->pmap_pvo, tpvo) {
2704 		if (pvo->pvo_vaddr & PVO_WIRED)
2705 			continue;
2706 
2707 		/*
2708 		 * For locking reasons, remove this from the page table and
2709 		 * pmap, but save delinking from the vm_page for a second
2710 		 * pass
2711 		 */
2712 		moea64_pvo_remove_from_pmap(pvo);
2713 		SLIST_INSERT_HEAD(&tofree, pvo, pvo_dlink);
2714 	}
2715 	PMAP_UNLOCK(pm);
2716 
2717 	while (!SLIST_EMPTY(&tofree)) {
2718 		pvo = SLIST_FIRST(&tofree);
2719 		SLIST_REMOVE_HEAD(&tofree, pvo_dlink);
2720 		moea64_pvo_remove_from_page(pvo);
2721 		free_pvo_entry(pvo);
2722 	}
2723 }
2724 
2725 static void
2726 moea64_remove_locked(pmap_t pm, vm_offset_t sva, vm_offset_t eva,
2727     struct pvo_dlist *tofree)
2728 {
2729 	struct pvo_entry *pvo, *tpvo, key;
2730 
2731 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2732 
2733 	key.pvo_vaddr = sva;
2734 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2735 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2736 		if (PVO_IS_SP(pvo)) {
2737 			if (moea64_sp_pvo_in_range(pvo, sva, eva)) {
2738 				tpvo = moea64_sp_remove(pvo, tofree);
2739 				continue;
2740 			} else {
2741 				CTR1(KTR_PMAP, "%s: demote before remove",
2742 				    __func__);
2743 				moea64_sp_demote(pvo);
2744 			}
2745 		}
2746 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2747 
2748 		/*
2749 		 * For locking reasons, remove this from the page table and
2750 		 * pmap, but save delinking from the vm_page for a second
2751 		 * pass
2752 		 */
2753 		moea64_pvo_remove_from_pmap(pvo);
2754 		SLIST_INSERT_HEAD(tofree, pvo, pvo_dlink);
2755 	}
2756 }
2757 
2758 /*
2759  * Remove the given range of addresses from the specified map.
2760  */
2761 void
2762 moea64_remove(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
2763 {
2764 	struct pvo_entry *pvo;
2765 	struct pvo_dlist tofree;
2766 
2767 	/*
2768 	 * Perform an unsynchronized read.  This is, however, safe.
2769 	 */
2770 	if (pm->pm_stats.resident_count == 0)
2771 		return;
2772 
2773 	SLIST_INIT(&tofree);
2774 	PMAP_LOCK(pm);
2775 	moea64_remove_locked(pm, sva, eva, &tofree);
2776 	PMAP_UNLOCK(pm);
2777 
2778 	while (!SLIST_EMPTY(&tofree)) {
2779 		pvo = SLIST_FIRST(&tofree);
2780 		SLIST_REMOVE_HEAD(&tofree, pvo_dlink);
2781 		moea64_pvo_remove_from_page(pvo);
2782 		free_pvo_entry(pvo);
2783 	}
2784 }
2785 
2786 /*
2787  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
2788  * will reflect changes in pte's back to the vm_page.
2789  */
2790 void
2791 moea64_remove_all(vm_page_t m)
2792 {
2793 	struct	pvo_entry *pvo, *next_pvo;
2794 	struct	pvo_head freequeue;
2795 	int	wasdead;
2796 	pmap_t	pmap;
2797 
2798 	LIST_INIT(&freequeue);
2799 
2800 	PV_PAGE_LOCK(m);
2801 	LIST_FOREACH_SAFE(pvo, vm_page_to_pvoh(m), pvo_vlink, next_pvo) {
2802 		pmap = pvo->pvo_pmap;
2803 		PMAP_LOCK(pmap);
2804 		wasdead = (pvo->pvo_vaddr & PVO_DEAD);
2805 		if (!wasdead) {
2806 			if (PVO_IS_SP(pvo)) {
2807 				CTR1(KTR_PMAP, "%s: demote before remove_all",
2808 				    __func__);
2809 				moea64_sp_demote(pvo);
2810 			}
2811 			moea64_pvo_remove_from_pmap(pvo);
2812 		}
2813 		moea64_pvo_remove_from_page_locked(pvo, m);
2814 		if (!wasdead)
2815 			LIST_INSERT_HEAD(&freequeue, pvo, pvo_vlink);
2816 		PMAP_UNLOCK(pmap);
2817 
2818 	}
2819 	KASSERT(!pmap_page_is_mapped(m), ("Page still has mappings"));
2820 	KASSERT((m->a.flags & PGA_WRITEABLE) == 0, ("Page still writable"));
2821 	PV_PAGE_UNLOCK(m);
2822 
2823 	/* Clean up UMA allocations */
2824 	LIST_FOREACH_SAFE(pvo, &freequeue, pvo_vlink, next_pvo)
2825 		free_pvo_entry(pvo);
2826 }
2827 
2828 /*
2829  * Allocate a physical page of memory directly from the phys_avail map.
2830  * Can only be called from moea64_bootstrap before avail start and end are
2831  * calculated.
2832  */
2833 vm_offset_t
2834 moea64_bootstrap_alloc(vm_size_t size, vm_size_t align)
2835 {
2836 	vm_offset_t	s, e;
2837 	int		i, j;
2838 
2839 	size = round_page(size);
2840 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2841 		if (align != 0)
2842 			s = roundup2(phys_avail[i], align);
2843 		else
2844 			s = phys_avail[i];
2845 		e = s + size;
2846 
2847 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2848 			continue;
2849 
2850 		if (s + size > platform_real_maxaddr())
2851 			continue;
2852 
2853 		if (s == phys_avail[i]) {
2854 			phys_avail[i] += size;
2855 		} else if (e == phys_avail[i + 1]) {
2856 			phys_avail[i + 1] -= size;
2857 		} else {
2858 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2859 				phys_avail[j] = phys_avail[j - 2];
2860 				phys_avail[j + 1] = phys_avail[j - 1];
2861 			}
2862 
2863 			phys_avail[i + 3] = phys_avail[i + 1];
2864 			phys_avail[i + 1] = s;
2865 			phys_avail[i + 2] = e;
2866 			phys_avail_count++;
2867 		}
2868 
2869 		return (s);
2870 	}
2871 	panic("moea64_bootstrap_alloc: could not allocate memory");
2872 }
2873 
2874 static int
2875 moea64_pvo_enter(struct pvo_entry *pvo, struct pvo_head *pvo_head,
2876     struct pvo_entry **oldpvop)
2877 {
2878 	struct pvo_entry *old_pvo;
2879 	int err;
2880 
2881 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2882 
2883 	STAT_MOEA64(moea64_pvo_enter_calls++);
2884 
2885 	/*
2886 	 * Add to pmap list
2887 	 */
2888 	old_pvo = RB_INSERT(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2889 
2890 	if (old_pvo != NULL) {
2891 		if (oldpvop != NULL)
2892 			*oldpvop = old_pvo;
2893 		return (EEXIST);
2894 	}
2895 
2896 	if (pvo_head != NULL) {
2897 		LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2898 	}
2899 
2900 	if (pvo->pvo_vaddr & PVO_WIRED)
2901 		pvo->pvo_pmap->pm_stats.wired_count++;
2902 	pvo->pvo_pmap->pm_stats.resident_count++;
2903 
2904 	/*
2905 	 * Insert it into the hardware page table
2906 	 */
2907 	err = moea64_pte_insert(pvo);
2908 	if (err != 0) {
2909 		panic("moea64_pvo_enter: overflow");
2910 	}
2911 
2912 	STAT_MOEA64(moea64_pvo_entries++);
2913 
2914 	if (pvo->pvo_pmap == kernel_pmap)
2915 		isync();
2916 
2917 #ifdef __powerpc64__
2918 	/*
2919 	 * Make sure all our bootstrap mappings are in the SLB as soon
2920 	 * as virtual memory is switched on.
2921 	 */
2922 	if (!pmap_bootstrapped)
2923 		moea64_bootstrap_slb_prefault(PVO_VADDR(pvo),
2924 		    pvo->pvo_vaddr & PVO_LARGE);
2925 #endif
2926 
2927 	return (0);
2928 }
2929 
2930 static void
2931 moea64_pvo_remove_from_pmap(struct pvo_entry *pvo)
2932 {
2933 	struct	vm_page *pg;
2934 	int32_t refchg;
2935 
2936 	KASSERT(pvo->pvo_pmap != NULL, ("Trying to remove PVO with no pmap"));
2937 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2938 	KASSERT(!(pvo->pvo_vaddr & PVO_DEAD), ("Trying to remove dead PVO"));
2939 
2940 	/*
2941 	 * If there is an active pte entry, we need to deactivate it
2942 	 */
2943 	refchg = moea64_pte_unset(pvo);
2944 	if (refchg < 0) {
2945 		/*
2946 		 * If it was evicted from the page table, be pessimistic and
2947 		 * dirty the page.
2948 		 */
2949 		if (pvo->pvo_pte.prot & VM_PROT_WRITE)
2950 			refchg = LPTE_CHG;
2951 		else
2952 			refchg = 0;
2953 	}
2954 
2955 	/*
2956 	 * Update our statistics.
2957 	 */
2958 	pvo->pvo_pmap->pm_stats.resident_count--;
2959 	if (pvo->pvo_vaddr & PVO_WIRED)
2960 		pvo->pvo_pmap->pm_stats.wired_count--;
2961 
2962 	/*
2963 	 * Remove this PVO from the pmap list.
2964 	 */
2965 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2966 
2967 	/*
2968 	 * Mark this for the next sweep
2969 	 */
2970 	pvo->pvo_vaddr |= PVO_DEAD;
2971 
2972 	/* Send RC bits to VM */
2973 	if ((pvo->pvo_vaddr & PVO_MANAGED) &&
2974 	    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2975 		pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
2976 		if (pg != NULL) {
2977 			refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2978 			if (refchg & LPTE_CHG)
2979 				vm_page_dirty(pg);
2980 			if (refchg & LPTE_REF)
2981 				vm_page_aflag_set(pg, PGA_REFERENCED);
2982 		}
2983 	}
2984 }
2985 
2986 static inline void
2987 moea64_pvo_remove_from_page_locked(struct pvo_entry *pvo,
2988     vm_page_t m)
2989 {
2990 
2991 	KASSERT(pvo->pvo_vaddr & PVO_DEAD, ("Trying to delink live page"));
2992 
2993 	/* Use NULL pmaps as a sentinel for races in page deletion */
2994 	if (pvo->pvo_pmap == NULL)
2995 		return;
2996 	pvo->pvo_pmap = NULL;
2997 
2998 	/*
2999 	 * Update vm about page writeability/executability if managed
3000 	 */
3001 	PV_LOCKASSERT(PVO_PADDR(pvo));
3002 	if (pvo->pvo_vaddr & PVO_MANAGED) {
3003 		if (m != NULL) {
3004 			LIST_REMOVE(pvo, pvo_vlink);
3005 			if (LIST_EMPTY(vm_page_to_pvoh(m)))
3006 				vm_page_aflag_clear(m,
3007 				    PGA_WRITEABLE | PGA_EXECUTABLE);
3008 		}
3009 	}
3010 
3011 	STAT_MOEA64(moea64_pvo_entries--);
3012 	STAT_MOEA64(moea64_pvo_remove_calls++);
3013 }
3014 
3015 static void
3016 moea64_pvo_remove_from_page(struct pvo_entry *pvo)
3017 {
3018 	vm_page_t pg = NULL;
3019 
3020 	if (pvo->pvo_vaddr & PVO_MANAGED)
3021 		pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
3022 
3023 	PV_LOCK(PVO_PADDR(pvo));
3024 	moea64_pvo_remove_from_page_locked(pvo, pg);
3025 	PV_UNLOCK(PVO_PADDR(pvo));
3026 }
3027 
3028 static struct pvo_entry *
3029 moea64_pvo_find_va(pmap_t pm, vm_offset_t va)
3030 {
3031 	struct pvo_entry key;
3032 
3033 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
3034 
3035 	key.pvo_vaddr = va & ~ADDR_POFF;
3036 	return (RB_FIND(pvo_tree, &pm->pmap_pvo, &key));
3037 }
3038 
3039 static bool
3040 moea64_query_bit(vm_page_t m, uint64_t ptebit)
3041 {
3042 	struct	pvo_entry *pvo;
3043 	int64_t ret;
3044 	bool rv;
3045 	vm_page_t sp;
3046 
3047 	/*
3048 	 * See if this bit is stored in the page already.
3049 	 *
3050 	 * For superpages, the bit is stored in the first vm page.
3051 	 */
3052 	if ((m->md.mdpg_attrs & ptebit) != 0 ||
3053 	    ((sp = PHYS_TO_VM_PAGE(VM_PAGE_TO_PHYS(m) & ~HPT_SP_MASK)) != NULL &&
3054 	     (sp->md.mdpg_attrs & (ptebit | MDPG_ATTR_SP)) ==
3055 	     (ptebit | MDPG_ATTR_SP)))
3056 		return (true);
3057 
3058 	/*
3059 	 * Examine each PTE.  Sync so that any pending REF/CHG bits are
3060 	 * flushed to the PTEs.
3061 	 */
3062 	rv = false;
3063 	powerpc_sync();
3064 	PV_PAGE_LOCK(m);
3065 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
3066 		if (PVO_IS_SP(pvo)) {
3067 			ret = moea64_sp_query(pvo, ptebit);
3068 			/*
3069 			 * If SP was not demoted, check its REF/CHG bits here.
3070 			 */
3071 			if (ret != -1) {
3072 				if ((ret & ptebit) != 0) {
3073 					rv = true;
3074 					break;
3075 				}
3076 				continue;
3077 			}
3078 			/* else, fallthrough */
3079 		}
3080 
3081 		ret = 0;
3082 
3083 		/*
3084 		 * See if this pvo has a valid PTE.  if so, fetch the
3085 		 * REF/CHG bits from the valid PTE.  If the appropriate
3086 		 * ptebit is set, return success.
3087 		 */
3088 		PMAP_LOCK(pvo->pvo_pmap);
3089 		if (!(pvo->pvo_vaddr & PVO_DEAD))
3090 			ret = moea64_pte_synch(pvo);
3091 		PMAP_UNLOCK(pvo->pvo_pmap);
3092 
3093 		if (ret > 0) {
3094 			atomic_set_32(&m->md.mdpg_attrs,
3095 			    ret & (LPTE_CHG | LPTE_REF));
3096 			if (ret & ptebit) {
3097 				rv = true;
3098 				break;
3099 			}
3100 		}
3101 	}
3102 	PV_PAGE_UNLOCK(m);
3103 
3104 	return (rv);
3105 }
3106 
3107 static u_int
3108 moea64_clear_bit(vm_page_t m, u_int64_t ptebit)
3109 {
3110 	u_int	count;
3111 	struct	pvo_entry *pvo;
3112 	int64_t ret;
3113 
3114 	/*
3115 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
3116 	 * we can reset the right ones).
3117 	 */
3118 	powerpc_sync();
3119 
3120 	/*
3121 	 * For each pvo entry, clear the pte's ptebit.
3122 	 */
3123 	count = 0;
3124 	PV_PAGE_LOCK(m);
3125 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
3126 		if (PVO_IS_SP(pvo)) {
3127 			if ((ret = moea64_sp_clear(pvo, m, ptebit)) != -1) {
3128 				count += ret;
3129 				continue;
3130 			}
3131 		}
3132 		ret = 0;
3133 
3134 		PMAP_LOCK(pvo->pvo_pmap);
3135 		if (!(pvo->pvo_vaddr & PVO_DEAD))
3136 			ret = moea64_pte_clear(pvo, ptebit);
3137 		PMAP_UNLOCK(pvo->pvo_pmap);
3138 
3139 		if (ret > 0 && (ret & ptebit))
3140 			count++;
3141 	}
3142 	atomic_clear_32(&m->md.mdpg_attrs, ptebit);
3143 	PV_PAGE_UNLOCK(m);
3144 
3145 	return (count);
3146 }
3147 
3148 int
3149 moea64_dev_direct_mapped(vm_paddr_t pa, vm_size_t size)
3150 {
3151 	struct pvo_entry *pvo, key;
3152 	vm_offset_t ppa;
3153 	int error = 0;
3154 
3155 	if (hw_direct_map && mem_valid(pa, size) == 0)
3156 		return (0);
3157 
3158 	PMAP_LOCK(kernel_pmap);
3159 	ppa = pa & ~ADDR_POFF;
3160 	key.pvo_vaddr = DMAP_BASE_ADDRESS + ppa;
3161 	for (pvo = RB_FIND(pvo_tree, &kernel_pmap->pmap_pvo, &key);
3162 	    ppa < pa + size; ppa += PAGE_SIZE,
3163 	    pvo = RB_NEXT(pvo_tree, &kernel_pmap->pmap_pvo, pvo)) {
3164 		if (pvo == NULL || PVO_PADDR(pvo) != ppa) {
3165 			error = EFAULT;
3166 			break;
3167 		}
3168 	}
3169 	PMAP_UNLOCK(kernel_pmap);
3170 
3171 	return (error);
3172 }
3173 
3174 /*
3175  * Map a set of physical memory pages into the kernel virtual
3176  * address space. Return a pointer to where it is mapped. This
3177  * routine is intended to be used for mapping device memory,
3178  * NOT real memory.
3179  */
3180 void *
3181 moea64_mapdev_attr(vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
3182 {
3183 	vm_offset_t va, tmpva, ppa, offset;
3184 
3185 	ppa = trunc_page(pa);
3186 	offset = pa & PAGE_MASK;
3187 	size = roundup2(offset + size, PAGE_SIZE);
3188 
3189 	va = kva_alloc(size);
3190 
3191 	if (!va)
3192 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
3193 
3194 	for (tmpva = va; size > 0;) {
3195 		moea64_kenter_attr(tmpva, ppa, ma);
3196 		size -= PAGE_SIZE;
3197 		tmpva += PAGE_SIZE;
3198 		ppa += PAGE_SIZE;
3199 	}
3200 
3201 	return ((void *)(va + offset));
3202 }
3203 
3204 void *
3205 moea64_mapdev(vm_paddr_t pa, vm_size_t size)
3206 {
3207 
3208 	return moea64_mapdev_attr(pa, size, VM_MEMATTR_DEFAULT);
3209 }
3210 
3211 void
3212 moea64_unmapdev(void *p, vm_size_t size)
3213 {
3214 	vm_offset_t base, offset, va;
3215 
3216 	va = (vm_offset_t)p;
3217 	base = trunc_page(va);
3218 	offset = va & PAGE_MASK;
3219 	size = roundup2(offset + size, PAGE_SIZE);
3220 
3221 	moea64_qremove(base, atop(size));
3222 	kva_free(base, size);
3223 }
3224 
3225 void
3226 moea64_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz)
3227 {
3228 	struct pvo_entry *pvo;
3229 	vm_offset_t lim;
3230 	vm_paddr_t pa;
3231 	vm_size_t len;
3232 
3233 	if (__predict_false(pm == NULL))
3234 		pm = &curthread->td_proc->p_vmspace->vm_pmap;
3235 
3236 	PMAP_LOCK(pm);
3237 	while (sz > 0) {
3238 		lim = round_page(va+1);
3239 		len = MIN(lim - va, sz);
3240 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
3241 		if (pvo != NULL && !(pvo->pvo_pte.pa & LPTE_I)) {
3242 			pa = PVO_PADDR(pvo) | (va & ADDR_POFF);
3243 			moea64_syncicache(pm, va, pa, len);
3244 		}
3245 		va += len;
3246 		sz -= len;
3247 	}
3248 	PMAP_UNLOCK(pm);
3249 }
3250 
3251 void
3252 moea64_dumpsys_map(vm_paddr_t pa, size_t sz, void **va)
3253 {
3254 
3255 	*va = (void *)(uintptr_t)pa;
3256 }
3257 
3258 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
3259 
3260 void
3261 moea64_scan_init(void)
3262 {
3263 	struct pvo_entry *pvo;
3264 	vm_offset_t va;
3265 	int i;
3266 
3267 	if (!do_minidump) {
3268 		/* Initialize phys. segments for dumpsys(). */
3269 		memset(&dump_map, 0, sizeof(dump_map));
3270 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
3271 		for (i = 0; i < pregions_sz; i++) {
3272 			dump_map[i].pa_start = pregions[i].mr_start;
3273 			dump_map[i].pa_size = pregions[i].mr_size;
3274 		}
3275 		return;
3276 	}
3277 
3278 	/* Virtual segments for minidumps: */
3279 	memset(&dump_map, 0, sizeof(dump_map));
3280 
3281 	/* 1st: kernel .data and .bss. */
3282 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
3283 	dump_map[0].pa_size = round_page((uintptr_t)_end) -
3284 	    dump_map[0].pa_start;
3285 
3286 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
3287 	dump_map[1].pa_start = (vm_paddr_t)(uintptr_t)msgbufp->msg_ptr;
3288 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
3289 
3290 	/* 3rd: kernel VM. */
3291 	va = dump_map[1].pa_start + dump_map[1].pa_size;
3292 	/* Find start of next chunk (from va). */
3293 	while (va < virtual_end) {
3294 		/* Don't dump the buffer cache. */
3295 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
3296 			va = kmi.buffer_eva;
3297 			continue;
3298 		}
3299 		pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
3300 		if (pvo != NULL && !(pvo->pvo_vaddr & PVO_DEAD))
3301 			break;
3302 		va += PAGE_SIZE;
3303 	}
3304 	if (va < virtual_end) {
3305 		dump_map[2].pa_start = va;
3306 		va += PAGE_SIZE;
3307 		/* Find last page in chunk. */
3308 		while (va < virtual_end) {
3309 			/* Don't run into the buffer cache. */
3310 			if (va == kmi.buffer_sva)
3311 				break;
3312 			pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
3313 			if (pvo == NULL || (pvo->pvo_vaddr & PVO_DEAD))
3314 				break;
3315 			va += PAGE_SIZE;
3316 		}
3317 		dump_map[2].pa_size = va - dump_map[2].pa_start;
3318 	}
3319 }
3320 
3321 #ifdef __powerpc64__
3322 
3323 static size_t
3324 moea64_scan_pmap(struct bitset *dump_bitset)
3325 {
3326 	struct pvo_entry *pvo;
3327 	vm_paddr_t pa, pa_end;
3328 	vm_offset_t va, pgva, kstart, kend, kstart_lp, kend_lp;
3329 	uint64_t lpsize;
3330 
3331 	lpsize = moea64_large_page_size;
3332 	kstart = trunc_page((vm_offset_t)_etext);
3333 	kend = round_page((vm_offset_t)_end);
3334 	kstart_lp = kstart & ~moea64_large_page_mask;
3335 	kend_lp = (kend + moea64_large_page_mask) & ~moea64_large_page_mask;
3336 
3337 	CTR4(KTR_PMAP, "moea64_scan_pmap: kstart=0x%016lx, kend=0x%016lx, "
3338 	    "kstart_lp=0x%016lx, kend_lp=0x%016lx",
3339 	    kstart, kend, kstart_lp, kend_lp);
3340 
3341 	PMAP_LOCK(kernel_pmap);
3342 	RB_FOREACH(pvo, pvo_tree, &kernel_pmap->pmap_pvo) {
3343 		va = pvo->pvo_vaddr;
3344 
3345 		if (va & PVO_DEAD)
3346 			continue;
3347 
3348 		/* Skip DMAP (except kernel area) */
3349 		if (va >= DMAP_BASE_ADDRESS && va <= DMAP_MAX_ADDRESS) {
3350 			if (va & PVO_LARGE) {
3351 				pgva = va & ~moea64_large_page_mask;
3352 				if (pgva < kstart_lp || pgva >= kend_lp)
3353 					continue;
3354 			} else {
3355 				pgva = trunc_page(va);
3356 				if (pgva < kstart || pgva >= kend)
3357 					continue;
3358 			}
3359 		}
3360 
3361 		pa = PVO_PADDR(pvo);
3362 
3363 		if (va & PVO_LARGE) {
3364 			pa_end = pa + lpsize;
3365 			for (; pa < pa_end; pa += PAGE_SIZE) {
3366 				if (vm_phys_is_dumpable(pa))
3367 					vm_page_dump_add(dump_bitset, pa);
3368 			}
3369 		} else {
3370 			if (vm_phys_is_dumpable(pa))
3371 				vm_page_dump_add(dump_bitset, pa);
3372 		}
3373 	}
3374 	PMAP_UNLOCK(kernel_pmap);
3375 
3376 	return (sizeof(struct lpte) * moea64_pteg_count * 8);
3377 }
3378 
3379 static struct dump_context dump_ctx;
3380 
3381 static void *
3382 moea64_dump_pmap_init(unsigned blkpgs)
3383 {
3384 	dump_ctx.ptex = 0;
3385 	dump_ctx.ptex_end = moea64_pteg_count * 8;
3386 	dump_ctx.blksz = blkpgs * PAGE_SIZE;
3387 	return (&dump_ctx);
3388 }
3389 
3390 #else
3391 
3392 static size_t
3393 moea64_scan_pmap(struct bitset *dump_bitset __unused)
3394 {
3395 	return (0);
3396 }
3397 
3398 static void *
3399 moea64_dump_pmap_init(unsigned blkpgs)
3400 {
3401 	return (NULL);
3402 }
3403 
3404 #endif
3405 
3406 #ifdef __powerpc64__
3407 static void
3408 moea64_map_range(vm_offset_t va, vm_paddr_t pa, vm_size_t npages)
3409 {
3410 
3411 	for (; npages > 0; --npages) {
3412 		if (moea64_large_page_size != 0 &&
3413 		    (pa & moea64_large_page_mask) == 0 &&
3414 		    (va & moea64_large_page_mask) == 0 &&
3415 		    npages >= (moea64_large_page_size >> PAGE_SHIFT)) {
3416 			PMAP_LOCK(kernel_pmap);
3417 			moea64_kenter_large(va, pa, 0, 0);
3418 			PMAP_UNLOCK(kernel_pmap);
3419 			pa += moea64_large_page_size;
3420 			va += moea64_large_page_size;
3421 			npages -= (moea64_large_page_size >> PAGE_SHIFT) - 1;
3422 		} else {
3423 			moea64_kenter(va, pa);
3424 			pa += PAGE_SIZE;
3425 			va += PAGE_SIZE;
3426 		}
3427 	}
3428 }
3429 
3430 static void
3431 moea64_page_array_startup(long pages)
3432 {
3433 	long dom_pages[MAXMEMDOM];
3434 	vm_paddr_t pa;
3435 	vm_offset_t va, vm_page_base;
3436 	vm_size_t needed, size;
3437 	int domain;
3438 	int i;
3439 
3440 	vm_page_base = 0xd000000000000000ULL;
3441 
3442 	/* Short-circuit single-domain systems. */
3443 	if (vm_ndomains == 1) {
3444 		size = round_page(pages * sizeof(struct vm_page));
3445 		pa = vm_phys_early_alloc(0, size);
3446 		vm_page_base = moea64_map(&vm_page_base,
3447 		    pa, pa + size, VM_PROT_READ | VM_PROT_WRITE);
3448 		vm_page_array_size = pages;
3449 		vm_page_array = (vm_page_t)vm_page_base;
3450 		return;
3451 	}
3452 
3453 	for (i = 0; i < MAXMEMDOM; i++)
3454 		dom_pages[i] = 0;
3455 
3456 	/* Now get the number of pages required per domain. */
3457 	for (i = 0; i < vm_phys_nsegs; i++) {
3458 		domain = vm_phys_segs[i].domain;
3459 		KASSERT(domain < MAXMEMDOM,
3460 		    ("Invalid vm_phys_segs NUMA domain %d!\n", domain));
3461 		/* Get size of vm_page_array needed for this segment. */
3462 		size = btoc(vm_phys_segs[i].end - vm_phys_segs[i].start);
3463 		dom_pages[domain] += size;
3464 	}
3465 
3466 	for (i = 0; phys_avail[i + 1] != 0; i+= 2) {
3467 		domain = vm_phys_domain(phys_avail[i]);
3468 		KASSERT(domain < MAXMEMDOM,
3469 		    ("Invalid phys_avail NUMA domain %d!\n", domain));
3470 		size = btoc(phys_avail[i + 1] - phys_avail[i]);
3471 		dom_pages[domain] += size;
3472 	}
3473 
3474 	/*
3475 	 * Map in chunks that can get us all 16MB pages.  There will be some
3476 	 * overlap between domains, but that's acceptable for now.
3477 	 */
3478 	vm_page_array_size = 0;
3479 	va = vm_page_base;
3480 	for (i = 0; i < MAXMEMDOM && vm_page_array_size < pages; i++) {
3481 		if (dom_pages[i] == 0)
3482 			continue;
3483 		size = ulmin(pages - vm_page_array_size, dom_pages[i]);
3484 		size = round_page(size * sizeof(struct vm_page));
3485 		needed = size;
3486 		size = roundup2(size, moea64_large_page_size);
3487 		pa = vm_phys_early_alloc(i, size);
3488 		vm_page_array_size += size / sizeof(struct vm_page);
3489 		moea64_map_range(va, pa, size >> PAGE_SHIFT);
3490 		/* Scoot up domain 0, to reduce the domain page overlap. */
3491 		if (i == 0)
3492 			vm_page_base += size - needed;
3493 		va += size;
3494 	}
3495 	vm_page_array = (vm_page_t)vm_page_base;
3496 	vm_page_array_size = pages;
3497 }
3498 #endif
3499 
3500 static int64_t
3501 moea64_null_method(void)
3502 {
3503 	return (0);
3504 }
3505 
3506 static int64_t moea64_pte_replace_default(struct pvo_entry *pvo, int flags)
3507 {
3508 	int64_t refchg;
3509 
3510 	refchg = moea64_pte_unset(pvo);
3511 	moea64_pte_insert(pvo);
3512 
3513 	return (refchg);
3514 }
3515 
3516 struct moea64_funcs *moea64_ops;
3517 
3518 #define DEFINE_OEA64_IFUNC(ret, func, args, def)		\
3519 	DEFINE_IFUNC(, ret, moea64_##func, args) {		\
3520 		moea64_##func##_t f;				\
3521 		if (moea64_ops == NULL)				\
3522 			return ((moea64_##func##_t)def);	\
3523 		f = moea64_ops->func;				\
3524 		return (f != NULL ? f : (moea64_##func##_t)def);\
3525 	}
3526 
3527 void
3528 moea64_install(void)
3529 {
3530 #ifdef __powerpc64__
3531 	if (hw_direct_map == -1) {
3532 		moea64_probe_large_page();
3533 
3534 		/* Use a direct map if we have large page support */
3535 		if (moea64_large_page_size > 0)
3536 			hw_direct_map = 1;
3537 		else
3538 			hw_direct_map = 0;
3539 	}
3540 #endif
3541 
3542 	/*
3543 	 * Default to non-DMAP, and switch over to DMAP functions once we know
3544 	 * we have DMAP.
3545 	 */
3546 	if (hw_direct_map) {
3547 		moea64_methods.quick_enter_page = moea64_quick_enter_page_dmap;
3548 		moea64_methods.quick_remove_page = NULL;
3549 		moea64_methods.copy_page = moea64_copy_page_dmap;
3550 		moea64_methods.zero_page = moea64_zero_page_dmap;
3551 		moea64_methods.copy_pages = moea64_copy_pages_dmap;
3552 	}
3553 }
3554 
3555 DEFINE_OEA64_IFUNC(int64_t, pte_replace, (struct pvo_entry *, int),
3556     moea64_pte_replace_default)
3557 DEFINE_OEA64_IFUNC(int64_t, pte_insert, (struct pvo_entry *), moea64_null_method)
3558 DEFINE_OEA64_IFUNC(int64_t, pte_unset, (struct pvo_entry *), moea64_null_method)
3559 DEFINE_OEA64_IFUNC(int64_t, pte_clear, (struct pvo_entry *, uint64_t),
3560     moea64_null_method)
3561 DEFINE_OEA64_IFUNC(int64_t, pte_synch, (struct pvo_entry *), moea64_null_method)
3562 DEFINE_OEA64_IFUNC(int64_t, pte_insert_sp, (struct pvo_entry *), moea64_null_method)
3563 DEFINE_OEA64_IFUNC(int64_t, pte_unset_sp, (struct pvo_entry *), moea64_null_method)
3564 DEFINE_OEA64_IFUNC(int64_t, pte_replace_sp, (struct pvo_entry *), moea64_null_method)
3565 
3566 /* Superpage functions */
3567 
3568 /* MMU interface */
3569 
3570 static bool
3571 moea64_ps_enabled(pmap_t pmap)
3572 {
3573 	return (superpages_enabled);
3574 }
3575 
3576 static void
3577 moea64_align_superpage(vm_object_t object, vm_ooffset_t offset,
3578     vm_offset_t *addr, vm_size_t size)
3579 {
3580 	vm_offset_t sp_offset;
3581 
3582 	if (size < HPT_SP_SIZE)
3583 		return;
3584 
3585 	CTR4(KTR_PMAP, "%s: offs=%#jx, addr=%p, size=%#jx",
3586 	    __func__, (uintmax_t)offset, addr, (uintmax_t)size);
3587 
3588 	if (object != NULL && (object->flags & OBJ_COLORED) != 0)
3589 		offset += ptoa(object->pg_color);
3590 	sp_offset = offset & HPT_SP_MASK;
3591 	if (size - ((HPT_SP_SIZE - sp_offset) & HPT_SP_MASK) < HPT_SP_SIZE ||
3592 	    (*addr & HPT_SP_MASK) == sp_offset)
3593 		return;
3594 	if ((*addr & HPT_SP_MASK) < sp_offset)
3595 		*addr = (*addr & ~HPT_SP_MASK) + sp_offset;
3596 	else
3597 		*addr = ((*addr + HPT_SP_MASK) & ~HPT_SP_MASK) + sp_offset;
3598 }
3599 
3600 /* Helpers */
3601 
3602 static __inline void
3603 moea64_pvo_cleanup(struct pvo_dlist *tofree)
3604 {
3605 	struct pvo_entry *pvo;
3606 
3607 	/* clean up */
3608 	while (!SLIST_EMPTY(tofree)) {
3609 		pvo = SLIST_FIRST(tofree);
3610 		SLIST_REMOVE_HEAD(tofree, pvo_dlink);
3611 		if (pvo->pvo_vaddr & PVO_DEAD)
3612 			moea64_pvo_remove_from_page(pvo);
3613 		free_pvo_entry(pvo);
3614 	}
3615 }
3616 
3617 static __inline uint16_t
3618 pvo_to_vmpage_flags(struct pvo_entry *pvo)
3619 {
3620 	uint16_t flags;
3621 
3622 	flags = 0;
3623 	if ((pvo->pvo_pte.prot & VM_PROT_WRITE) != 0)
3624 		flags |= PGA_WRITEABLE;
3625 	if ((pvo->pvo_pte.prot & VM_PROT_EXECUTE) != 0)
3626 		flags |= PGA_EXECUTABLE;
3627 
3628 	return (flags);
3629 }
3630 
3631 /*
3632  * Check if the given pvo and its superpage are in sva-eva range.
3633  */
3634 static __inline bool
3635 moea64_sp_pvo_in_range(struct pvo_entry *pvo, vm_offset_t sva, vm_offset_t eva)
3636 {
3637 	vm_offset_t spva;
3638 
3639 	spva = PVO_VADDR(pvo) & ~HPT_SP_MASK;
3640 	if (spva >= sva && spva + HPT_SP_SIZE <= eva) {
3641 		/*
3642 		 * Because this function is intended to be called from loops
3643 		 * that iterate over ordered pvo entries, if the condition
3644 		 * above is true then the pvo must be the first of its
3645 		 * superpage.
3646 		 */
3647 		KASSERT(PVO_VADDR(pvo) == spva,
3648 		    ("%s: unexpected unaligned superpage pvo", __func__));
3649 		return (true);
3650 	}
3651 	return (false);
3652 }
3653 
3654 /*
3655  * Update vm about the REF/CHG bits if the superpage is managed and
3656  * has (or had) write access.
3657  */
3658 static void
3659 moea64_sp_refchg_process(struct pvo_entry *sp, vm_page_t m,
3660     int64_t sp_refchg, vm_prot_t prot)
3661 {
3662 	vm_page_t m_end;
3663 	int64_t refchg;
3664 
3665 	if ((sp->pvo_vaddr & PVO_MANAGED) != 0 && (prot & VM_PROT_WRITE) != 0) {
3666 		for (m_end = &m[HPT_SP_PAGES]; m < m_end; m++) {
3667 			refchg = sp_refchg |
3668 			    atomic_readandclear_32(&m->md.mdpg_attrs);
3669 			if (refchg & LPTE_CHG)
3670 				vm_page_dirty(m);
3671 			if (refchg & LPTE_REF)
3672 				vm_page_aflag_set(m, PGA_REFERENCED);
3673 		}
3674 	}
3675 }
3676 
3677 /* Superpage ops */
3678 
3679 static int
3680 moea64_sp_enter(pmap_t pmap, vm_offset_t va, vm_page_t m,
3681     vm_prot_t prot, u_int flags, int8_t psind)
3682 {
3683 	struct pvo_entry *pvo, **pvos;
3684 	struct pvo_head *pvo_head;
3685 	vm_offset_t sva;
3686 	vm_page_t sm;
3687 	vm_paddr_t pa, spa;
3688 	bool sync;
3689 	struct pvo_dlist tofree;
3690 	int error __diagused, i;
3691 	uint16_t aflags;
3692 
3693 	KASSERT((va & HPT_SP_MASK) == 0, ("%s: va %#jx unaligned",
3694 	    __func__, (uintmax_t)va));
3695 	KASSERT(psind == 1, ("%s: invalid psind: %d", __func__, psind));
3696 	KASSERT(m->psind == 1, ("%s: invalid m->psind: %d",
3697 	    __func__, m->psind));
3698 	KASSERT(pmap != kernel_pmap,
3699 	    ("%s: function called with kernel pmap", __func__));
3700 
3701 	CTR5(KTR_PMAP, "%s: va=%#jx, pa=%#jx, prot=%#x, flags=%#x, psind=1",
3702 	    __func__, (uintmax_t)va, (uintmax_t)VM_PAGE_TO_PHYS(m),
3703 	    prot, flags);
3704 
3705 	SLIST_INIT(&tofree);
3706 
3707 	sva = va;
3708 	sm = m;
3709 	spa = pa = VM_PAGE_TO_PHYS(sm);
3710 
3711 	/* Try to allocate all PVOs first, to make failure handling easier. */
3712 	pvos = malloc(HPT_SP_PAGES * sizeof(struct pvo_entry *), M_TEMP,
3713 	    M_NOWAIT);
3714 	if (pvos == NULL) {
3715 		CTR1(KTR_PMAP, "%s: failed to alloc pvo array", __func__);
3716 		return (KERN_RESOURCE_SHORTAGE);
3717 	}
3718 
3719 	for (i = 0; i < HPT_SP_PAGES; i++) {
3720 		pvos[i] = alloc_pvo_entry(0);
3721 		if (pvos[i] == NULL) {
3722 			CTR1(KTR_PMAP, "%s: failed to alloc pvo", __func__);
3723 			for (i = i - 1; i >= 0; i--)
3724 				free_pvo_entry(pvos[i]);
3725 			free(pvos, M_TEMP);
3726 			return (KERN_RESOURCE_SHORTAGE);
3727 		}
3728 	}
3729 
3730 	SP_PV_LOCK_ALIGNED(spa);
3731 	PMAP_LOCK(pmap);
3732 
3733 	/* Note: moea64_remove_locked() also clears cached REF/CHG bits. */
3734 	moea64_remove_locked(pmap, va, va + HPT_SP_SIZE, &tofree);
3735 
3736 	/* Enter pages */
3737 	for (i = 0; i < HPT_SP_PAGES;
3738 	    i++, va += PAGE_SIZE, pa += PAGE_SIZE, m++) {
3739 		pvo = pvos[i];
3740 
3741 		pvo->pvo_pte.prot = prot;
3742 		pvo->pvo_pte.pa = (pa & ~HPT_SP_MASK) | LPTE_LP_4K_16M |
3743 		    moea64_calc_wimg(pa, pmap_page_get_memattr(m));
3744 
3745 		if ((flags & PMAP_ENTER_WIRED) != 0)
3746 			pvo->pvo_vaddr |= PVO_WIRED;
3747 		pvo->pvo_vaddr |= PVO_LARGE;
3748 
3749 		if ((m->oflags & VPO_UNMANAGED) != 0)
3750 			pvo_head = NULL;
3751 		else {
3752 			pvo_head = &m->md.mdpg_pvoh;
3753 			pvo->pvo_vaddr |= PVO_MANAGED;
3754 		}
3755 
3756 		init_pvo_entry(pvo, pmap, va);
3757 
3758 		error = moea64_pvo_enter(pvo, pvo_head, NULL);
3759 		/*
3760 		 * All superpage PVOs were previously removed, so no errors
3761 		 * should occur while inserting the new ones.
3762 		 */
3763 		KASSERT(error == 0, ("%s: unexpected error "
3764 			    "when inserting superpage PVO: %d",
3765 			    __func__, error));
3766 	}
3767 
3768 	PMAP_UNLOCK(pmap);
3769 	SP_PV_UNLOCK_ALIGNED(spa);
3770 
3771 	sync = (sm->a.flags & PGA_EXECUTABLE) == 0;
3772 	/* Note: moea64_pvo_cleanup() also clears page prot. flags. */
3773 	moea64_pvo_cleanup(&tofree);
3774 	pvo = pvos[0];
3775 
3776 	/* Set vm page flags */
3777 	aflags = pvo_to_vmpage_flags(pvo);
3778 	if (aflags != 0)
3779 		for (m = sm; m < &sm[HPT_SP_PAGES]; m++)
3780 			vm_page_aflag_set(m, aflags);
3781 
3782 	/*
3783 	 * Flush the page from the instruction cache if this page is
3784 	 * mapped executable and cacheable.
3785 	 */
3786 	if (sync && (pvo->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0)
3787 		moea64_syncicache(pmap, sva, spa, HPT_SP_SIZE);
3788 
3789 	atomic_add_long(&sp_mappings, 1);
3790 	CTR3(KTR_PMAP, "%s: SP success for va %#jx in pmap %p",
3791 	    __func__, (uintmax_t)sva, pmap);
3792 
3793 	free(pvos, M_TEMP);
3794 	return (KERN_SUCCESS);
3795 }
3796 
3797 static void
3798 moea64_sp_promote(pmap_t pmap, vm_offset_t va, vm_page_t m)
3799 {
3800 	struct pvo_entry *first, *pvo;
3801 	vm_paddr_t pa, pa_end;
3802 	vm_offset_t sva, va_end;
3803 	int64_t sp_refchg;
3804 
3805 	/* This CTR may generate a lot of output. */
3806 	/* CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)va); */
3807 
3808 	va &= ~HPT_SP_MASK;
3809 	sva = va;
3810 	/* Get superpage */
3811 	pa = VM_PAGE_TO_PHYS(m) & ~HPT_SP_MASK;
3812 	m = PHYS_TO_VM_PAGE(pa);
3813 
3814 	PMAP_LOCK(pmap);
3815 
3816 	/*
3817 	 * Check if all pages meet promotion criteria.
3818 	 *
3819 	 * XXX In some cases the loop below may be executed for each or most
3820 	 * of the entered pages of a superpage, which can be expensive
3821 	 * (although it was not profiled) and need some optimization.
3822 	 *
3823 	 * Some cases where this seems to happen are:
3824 	 * - When a superpage is first entered read-only and later becomes
3825 	 *   read-write.
3826 	 * - When some of the superpage's virtual addresses map to previously
3827 	 *   wired/cached pages while others map to pages allocated from a
3828 	 *   different physical address range. A common scenario where this
3829 	 *   happens is when mmap'ing a file that is already present in FS
3830 	 *   block cache and doesn't fill a superpage.
3831 	 */
3832 	first = pvo = moea64_pvo_find_va(pmap, sva);
3833 	for (pa_end = pa + HPT_SP_SIZE;
3834 	    pa < pa_end; pa += PAGE_SIZE, va += PAGE_SIZE) {
3835 		if (pvo == NULL || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
3836 			CTR3(KTR_PMAP,
3837 			    "%s: NULL or dead PVO: pmap=%p, va=%#jx",
3838 			    __func__, pmap, (uintmax_t)va);
3839 			goto error;
3840 		}
3841 		if (PVO_PADDR(pvo) != pa) {
3842 			CTR5(KTR_PMAP, "%s: PAs don't match: "
3843 			    "pmap=%p, va=%#jx, pvo_pa=%#jx, exp_pa=%#jx",
3844 			    __func__, pmap, (uintmax_t)va,
3845 			    (uintmax_t)PVO_PADDR(pvo), (uintmax_t)pa);
3846 			atomic_add_long(&sp_p_fail_pa, 1);
3847 			goto error;
3848 		}
3849 		if ((first->pvo_vaddr & PVO_FLAGS_PROMOTE) !=
3850 		    (pvo->pvo_vaddr & PVO_FLAGS_PROMOTE)) {
3851 			CTR5(KTR_PMAP, "%s: PVO flags don't match: "
3852 			    "pmap=%p, va=%#jx, pvo_flags=%#jx, exp_flags=%#jx",
3853 			    __func__, pmap, (uintmax_t)va,
3854 			    (uintmax_t)(pvo->pvo_vaddr & PVO_FLAGS_PROMOTE),
3855 			    (uintmax_t)(first->pvo_vaddr & PVO_FLAGS_PROMOTE));
3856 			atomic_add_long(&sp_p_fail_flags, 1);
3857 			goto error;
3858 		}
3859 		if (first->pvo_pte.prot != pvo->pvo_pte.prot) {
3860 			CTR5(KTR_PMAP, "%s: PVO protections don't match: "
3861 			    "pmap=%p, va=%#jx, pvo_prot=%#x, exp_prot=%#x",
3862 			    __func__, pmap, (uintmax_t)va,
3863 			    pvo->pvo_pte.prot, first->pvo_pte.prot);
3864 			atomic_add_long(&sp_p_fail_prot, 1);
3865 			goto error;
3866 		}
3867 		if ((first->pvo_pte.pa & LPTE_WIMG) !=
3868 		    (pvo->pvo_pte.pa & LPTE_WIMG)) {
3869 			CTR5(KTR_PMAP, "%s: WIMG bits don't match: "
3870 			    "pmap=%p, va=%#jx, pvo_wimg=%#jx, exp_wimg=%#jx",
3871 			    __func__, pmap, (uintmax_t)va,
3872 			    (uintmax_t)(pvo->pvo_pte.pa & LPTE_WIMG),
3873 			    (uintmax_t)(first->pvo_pte.pa & LPTE_WIMG));
3874 			atomic_add_long(&sp_p_fail_wimg, 1);
3875 			goto error;
3876 		}
3877 
3878 		pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo);
3879 	}
3880 
3881 	/* All OK, promote. */
3882 
3883 	/*
3884 	 * Handle superpage REF/CHG bits. If REF or CHG is set in
3885 	 * any page, then it must be set in the superpage.
3886 	 *
3887 	 * Instead of querying each page, we take advantage of two facts:
3888 	 * 1- If a page is being promoted, it was referenced.
3889 	 * 2- If promoted pages are writable, they were modified.
3890 	 */
3891 	sp_refchg = LPTE_REF |
3892 	    ((first->pvo_pte.prot & VM_PROT_WRITE) != 0 ? LPTE_CHG : 0);
3893 
3894 	/* Promote pages */
3895 
3896 	for (pvo = first, va_end = PVO_VADDR(pvo) + HPT_SP_SIZE;
3897 	    pvo != NULL && PVO_VADDR(pvo) < va_end;
3898 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo)) {
3899 		pvo->pvo_pte.pa &= ADDR_POFF | ~HPT_SP_MASK;
3900 		pvo->pvo_pte.pa |= LPTE_LP_4K_16M;
3901 		pvo->pvo_vaddr |= PVO_LARGE;
3902 	}
3903 	moea64_pte_replace_sp(first);
3904 
3905 	/* Send REF/CHG bits to VM */
3906 	moea64_sp_refchg_process(first, m, sp_refchg, first->pvo_pte.prot);
3907 
3908 	/* Use first page to cache REF/CHG bits */
3909 	atomic_set_32(&m->md.mdpg_attrs, sp_refchg | MDPG_ATTR_SP);
3910 
3911 	PMAP_UNLOCK(pmap);
3912 
3913 	atomic_add_long(&sp_mappings, 1);
3914 	atomic_add_long(&sp_promotions, 1);
3915 	CTR3(KTR_PMAP, "%s: success for va %#jx in pmap %p",
3916 	    __func__, (uintmax_t)sva, pmap);
3917 	return;
3918 
3919 error:
3920 	atomic_add_long(&sp_p_failures, 1);
3921 	PMAP_UNLOCK(pmap);
3922 }
3923 
3924 static void
3925 moea64_sp_demote_aligned(struct pvo_entry *sp)
3926 {
3927 	struct pvo_entry *pvo;
3928 	vm_offset_t va, va_end;
3929 	vm_paddr_t pa;
3930 	vm_page_t m;
3931 	pmap_t pmap __diagused;
3932 	int64_t refchg;
3933 
3934 	CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)PVO_VADDR(sp));
3935 
3936 	pmap = sp->pvo_pmap;
3937 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3938 
3939 	pvo = sp;
3940 
3941 	/* Demote pages */
3942 
3943 	va = PVO_VADDR(pvo);
3944 	pa = PVO_PADDR(pvo);
3945 	m = PHYS_TO_VM_PAGE(pa);
3946 
3947 	for (pvo = sp, va_end = va + HPT_SP_SIZE;
3948 	    pvo != NULL && PVO_VADDR(pvo) < va_end;
3949 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo),
3950 	    va += PAGE_SIZE, pa += PAGE_SIZE) {
3951 		KASSERT(pvo && PVO_VADDR(pvo) == va,
3952 		    ("%s: missing PVO for va %#jx", __func__, (uintmax_t)va));
3953 
3954 		pvo->pvo_vaddr &= ~PVO_LARGE;
3955 		pvo->pvo_pte.pa &= ~LPTE_RPGN;
3956 		pvo->pvo_pte.pa |= pa;
3957 
3958 	}
3959 	refchg = moea64_pte_replace_sp(sp);
3960 
3961 	/*
3962 	 * Clear SP flag
3963 	 *
3964 	 * XXX It is possible that another pmap has this page mapped as
3965 	 *     part of a superpage, but as the SP flag is used only for
3966 	 *     caching SP REF/CHG bits, that will be queried if not set
3967 	 *     in cache, it should be ok to clear it here.
3968 	 */
3969 	atomic_clear_32(&m->md.mdpg_attrs, MDPG_ATTR_SP);
3970 
3971 	/*
3972 	 * Handle superpage REF/CHG bits. A bit set in the superpage
3973 	 * means all pages should consider it set.
3974 	 */
3975 	moea64_sp_refchg_process(sp, m, refchg, sp->pvo_pte.prot);
3976 
3977 	atomic_add_long(&sp_demotions, 1);
3978 	CTR3(KTR_PMAP, "%s: success for va %#jx in pmap %p",
3979 	    __func__, (uintmax_t)PVO_VADDR(sp), pmap);
3980 }
3981 
3982 static void
3983 moea64_sp_demote(struct pvo_entry *pvo)
3984 {
3985 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
3986 
3987 	if ((PVO_VADDR(pvo) & HPT_SP_MASK) != 0) {
3988 		pvo = moea64_pvo_find_va(pvo->pvo_pmap,
3989 		    PVO_VADDR(pvo) & ~HPT_SP_MASK);
3990 		KASSERT(pvo != NULL, ("%s: missing PVO for va %#jx",
3991 		     __func__, (uintmax_t)(PVO_VADDR(pvo) & ~HPT_SP_MASK)));
3992 	}
3993 	moea64_sp_demote_aligned(pvo);
3994 }
3995 
3996 static struct pvo_entry *
3997 moea64_sp_unwire(struct pvo_entry *sp)
3998 {
3999 	struct pvo_entry *pvo, *prev;
4000 	vm_offset_t eva;
4001 	pmap_t pm;
4002 	int64_t ret, refchg;
4003 
4004 	CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)PVO_VADDR(sp));
4005 
4006 	pm = sp->pvo_pmap;
4007 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
4008 
4009 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4010 	refchg = 0;
4011 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4012 	    prev = pvo, pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
4013 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
4014 			panic("%s: pvo %p is missing PVO_WIRED",
4015 			    __func__, pvo);
4016 		pvo->pvo_vaddr &= ~PVO_WIRED;
4017 
4018 		ret = moea64_pte_replace(pvo, 0 /* No invalidation */);
4019 		if (ret < 0)
4020 			refchg |= LPTE_CHG;
4021 		else
4022 			refchg |= ret;
4023 
4024 		pm->pm_stats.wired_count--;
4025 	}
4026 
4027 	/* Send REF/CHG bits to VM */
4028 	moea64_sp_refchg_process(sp, PHYS_TO_VM_PAGE(PVO_PADDR(sp)),
4029 	    refchg, sp->pvo_pte.prot);
4030 
4031 	return (prev);
4032 }
4033 
4034 static struct pvo_entry *
4035 moea64_sp_protect(struct pvo_entry *sp, vm_prot_t prot)
4036 {
4037 	struct pvo_entry *pvo, *prev;
4038 	vm_offset_t eva;
4039 	pmap_t pm;
4040 	vm_page_t m, m_end;
4041 	int64_t ret, refchg;
4042 	vm_prot_t oldprot;
4043 
4044 	CTR3(KTR_PMAP, "%s: va=%#jx, prot=%x",
4045 	    __func__, (uintmax_t)PVO_VADDR(sp), prot);
4046 
4047 	pm = sp->pvo_pmap;
4048 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
4049 
4050 	oldprot = sp->pvo_pte.prot;
4051 	m = PHYS_TO_VM_PAGE(PVO_PADDR(sp));
4052 	KASSERT(m != NULL, ("%s: missing vm page for pa %#jx",
4053 	    __func__, (uintmax_t)PVO_PADDR(sp)));
4054 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4055 	refchg = 0;
4056 
4057 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4058 	    prev = pvo, pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
4059 		pvo->pvo_pte.prot = prot;
4060 		/*
4061 		 * If the PVO is in the page table, update mapping
4062 		 */
4063 		ret = moea64_pte_replace(pvo, MOEA64_PTE_PROT_UPDATE);
4064 		if (ret < 0)
4065 			refchg |= LPTE_CHG;
4066 		else
4067 			refchg |= ret;
4068 	}
4069 
4070 	/* Send REF/CHG bits to VM */
4071 	moea64_sp_refchg_process(sp, m, refchg, oldprot);
4072 
4073 	/* Handle pages that became executable */
4074 	if ((m->a.flags & PGA_EXECUTABLE) == 0 &&
4075 	    (sp->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
4076 		if ((m->oflags & VPO_UNMANAGED) == 0)
4077 			for (m_end = &m[HPT_SP_PAGES]; m < m_end; m++)
4078 				vm_page_aflag_set(m, PGA_EXECUTABLE);
4079 		moea64_syncicache(pm, PVO_VADDR(sp), PVO_PADDR(sp),
4080 		    HPT_SP_SIZE);
4081 	}
4082 
4083 	return (prev);
4084 }
4085 
4086 static struct pvo_entry *
4087 moea64_sp_remove(struct pvo_entry *sp, struct pvo_dlist *tofree)
4088 {
4089 	struct pvo_entry *pvo, *tpvo;
4090 	vm_offset_t eva;
4091 	pmap_t pm __diagused;
4092 
4093 	CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)PVO_VADDR(sp));
4094 
4095 	pm = sp->pvo_pmap;
4096 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
4097 
4098 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4099 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
4100 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
4101 
4102 		/*
4103 		 * For locking reasons, remove this from the page table and
4104 		 * pmap, but save delinking from the vm_page for a second
4105 		 * pass
4106 		 */
4107 		moea64_pvo_remove_from_pmap(pvo);
4108 		SLIST_INSERT_HEAD(tofree, pvo, pvo_dlink);
4109 	}
4110 
4111 	/*
4112 	 * Clear SP bit
4113 	 *
4114 	 * XXX See comment in moea64_sp_demote_aligned() for why it's
4115 	 *     ok to always clear the SP bit on remove/demote.
4116 	 */
4117 	atomic_clear_32(&PHYS_TO_VM_PAGE(PVO_PADDR(sp))->md.mdpg_attrs,
4118 	    MDPG_ATTR_SP);
4119 
4120 	return (tpvo);
4121 }
4122 
4123 static int64_t
4124 moea64_sp_query_locked(struct pvo_entry *pvo, uint64_t ptebit)
4125 {
4126 	int64_t refchg, ret;
4127 	vm_offset_t eva;
4128 	vm_page_t m;
4129 	pmap_t pmap;
4130 	struct pvo_entry *sp;
4131 
4132 	pmap = pvo->pvo_pmap;
4133 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4134 
4135 	/* Get first SP PVO */
4136 	if ((PVO_VADDR(pvo) & HPT_SP_MASK) != 0) {
4137 		sp = moea64_pvo_find_va(pmap, PVO_VADDR(pvo) & ~HPT_SP_MASK);
4138 		KASSERT(sp != NULL, ("%s: missing PVO for va %#jx",
4139 		     __func__, (uintmax_t)(PVO_VADDR(pvo) & ~HPT_SP_MASK)));
4140 	} else
4141 		sp = pvo;
4142 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4143 
4144 	refchg = 0;
4145 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4146 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo)) {
4147 		ret = moea64_pte_synch(pvo);
4148 		if (ret > 0) {
4149 			refchg |= ret & (LPTE_CHG | LPTE_REF);
4150 			if ((refchg & ptebit) != 0)
4151 				break;
4152 		}
4153 	}
4154 
4155 	/* Save results */
4156 	if (refchg != 0) {
4157 		m = PHYS_TO_VM_PAGE(PVO_PADDR(sp));
4158 		atomic_set_32(&m->md.mdpg_attrs, refchg | MDPG_ATTR_SP);
4159 	}
4160 
4161 	return (refchg);
4162 }
4163 
4164 static int64_t
4165 moea64_sp_query(struct pvo_entry *pvo, uint64_t ptebit)
4166 {
4167 	int64_t refchg;
4168 	pmap_t pmap;
4169 
4170 	pmap = pvo->pvo_pmap;
4171 	PMAP_LOCK(pmap);
4172 
4173 	/*
4174 	 * Check if SP was demoted/removed before pmap lock was acquired.
4175 	 */
4176 	if (!PVO_IS_SP(pvo) || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
4177 		CTR2(KTR_PMAP, "%s: demoted/removed: pa=%#jx",
4178 		    __func__, (uintmax_t)PVO_PADDR(pvo));
4179 		PMAP_UNLOCK(pmap);
4180 		return (-1);
4181 	}
4182 
4183 	refchg = moea64_sp_query_locked(pvo, ptebit);
4184 	PMAP_UNLOCK(pmap);
4185 
4186 	CTR4(KTR_PMAP, "%s: va=%#jx, pa=%#jx: refchg=%#jx",
4187 	    __func__, (uintmax_t)PVO_VADDR(pvo),
4188 	    (uintmax_t)PVO_PADDR(pvo), (uintmax_t)refchg);
4189 
4190 	return (refchg);
4191 }
4192 
4193 static int64_t
4194 moea64_sp_pvo_clear(struct pvo_entry *pvo, uint64_t ptebit)
4195 {
4196 	int64_t refchg, ret;
4197 	pmap_t pmap;
4198 	struct pvo_entry *sp;
4199 	vm_offset_t eva;
4200 	vm_page_t m;
4201 
4202 	pmap = pvo->pvo_pmap;
4203 	PMAP_LOCK(pmap);
4204 
4205 	/*
4206 	 * Check if SP was demoted/removed before pmap lock was acquired.
4207 	 */
4208 	if (!PVO_IS_SP(pvo) || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
4209 		CTR2(KTR_PMAP, "%s: demoted/removed: pa=%#jx",
4210 		    __func__, (uintmax_t)PVO_PADDR(pvo));
4211 		PMAP_UNLOCK(pmap);
4212 		return (-1);
4213 	}
4214 
4215 	/* Get first SP PVO */
4216 	if ((PVO_VADDR(pvo) & HPT_SP_MASK) != 0) {
4217 		sp = moea64_pvo_find_va(pmap, PVO_VADDR(pvo) & ~HPT_SP_MASK);
4218 		KASSERT(sp != NULL, ("%s: missing PVO for va %#jx",
4219 		     __func__, (uintmax_t)(PVO_VADDR(pvo) & ~HPT_SP_MASK)));
4220 	} else
4221 		sp = pvo;
4222 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4223 
4224 	refchg = 0;
4225 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4226 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo)) {
4227 		ret = moea64_pte_clear(pvo, ptebit);
4228 		if (ret > 0)
4229 			refchg |= ret & (LPTE_CHG | LPTE_REF);
4230 	}
4231 
4232 	m = PHYS_TO_VM_PAGE(PVO_PADDR(sp));
4233 	atomic_clear_32(&m->md.mdpg_attrs, ptebit);
4234 	PMAP_UNLOCK(pmap);
4235 
4236 	CTR4(KTR_PMAP, "%s: va=%#jx, pa=%#jx: refchg=%#jx",
4237 	    __func__, (uintmax_t)PVO_VADDR(sp),
4238 	    (uintmax_t)PVO_PADDR(sp), (uintmax_t)refchg);
4239 
4240 	return (refchg);
4241 }
4242 
4243 static int64_t
4244 moea64_sp_clear(struct pvo_entry *pvo, vm_page_t m, uint64_t ptebit)
4245 {
4246 	int64_t count, ret;
4247 	pmap_t pmap;
4248 
4249 	count = 0;
4250 	pmap = pvo->pvo_pmap;
4251 
4252 	/*
4253 	 * Since this reference bit is shared by 4096 4KB pages, it
4254 	 * should not be cleared every time it is tested. Apply a
4255 	 * simple "hash" function on the physical page number, the
4256 	 * virtual superpage number, and the pmap address to select
4257 	 * one 4KB page out of the 4096 on which testing the
4258 	 * reference bit will result in clearing that reference bit.
4259 	 * This function is designed to avoid the selection of the
4260 	 * same 4KB page for every 16MB page mapping.
4261 	 *
4262 	 * Always leave the reference bit of a wired mapping set, as
4263 	 * the current state of its reference bit won't affect page
4264 	 * replacement.
4265 	 */
4266 	if (ptebit == LPTE_REF && (((VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) ^
4267 	    (PVO_VADDR(pvo) >> HPT_SP_SHIFT) ^ (uintptr_t)pmap) &
4268 	    (HPT_SP_PAGES - 1)) == 0 && (pvo->pvo_vaddr & PVO_WIRED) == 0) {
4269 		if ((ret = moea64_sp_pvo_clear(pvo, ptebit)) == -1)
4270 			return (-1);
4271 
4272 		if ((ret & ptebit) != 0)
4273 			count++;
4274 
4275 	/*
4276 	 * If this page was not selected by the hash function, then assume
4277 	 * its REF bit was set.
4278 	 */
4279 	} else if (ptebit == LPTE_REF) {
4280 		count++;
4281 
4282 	/*
4283 	 * To clear the CHG bit of a single SP page, first it must be demoted.
4284 	 * But if no CHG bit is set, no bit clear and thus no SP demotion is
4285 	 * needed.
4286 	 */
4287 	} else {
4288 		CTR4(KTR_PMAP, "%s: ptebit=%#jx, va=%#jx, pa=%#jx",
4289 		    __func__, (uintmax_t)ptebit, (uintmax_t)PVO_VADDR(pvo),
4290 		    (uintmax_t)PVO_PADDR(pvo));
4291 
4292 		PMAP_LOCK(pmap);
4293 
4294 		/*
4295 		 * Make sure SP wasn't demoted/removed before pmap lock
4296 		 * was acquired.
4297 		 */
4298 		if (!PVO_IS_SP(pvo) || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
4299 			CTR2(KTR_PMAP, "%s: demoted/removed: pa=%#jx",
4300 			    __func__, (uintmax_t)PVO_PADDR(pvo));
4301 			PMAP_UNLOCK(pmap);
4302 			return (-1);
4303 		}
4304 
4305 		ret = moea64_sp_query_locked(pvo, ptebit);
4306 		if ((ret & ptebit) != 0)
4307 			count++;
4308 		else {
4309 			PMAP_UNLOCK(pmap);
4310 			return (0);
4311 		}
4312 
4313 		moea64_sp_demote(pvo);
4314 		moea64_pte_clear(pvo, ptebit);
4315 
4316 		/*
4317 		 * Write protect the mapping to a single page so that a
4318 		 * subsequent write access may repromote.
4319 		 */
4320 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
4321 			moea64_pvo_protect(pmap, pvo,
4322 			    pvo->pvo_pte.prot & ~VM_PROT_WRITE);
4323 
4324 		PMAP_UNLOCK(pmap);
4325 	}
4326 
4327 	return (count);
4328 }
4329