xref: /freebsd/sys/powerpc/aim/mmu_oea64.c (revision cc68614da8232d8baaca0ae0d0dd8f890f06623e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2008-2015 Nathan Whitehorn
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  *
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 #include <sys/cdefs.h>
30 __FBSDID("$FreeBSD$");
31 
32 /*
33  * Manages physical address maps.
34  *
35  * Since the information managed by this module is also stored by the
36  * logical address mapping module, this module may throw away valid virtual
37  * to physical mappings at almost any time.  However, invalidations of
38  * mappings must be done as requested.
39  *
40  * In order to cope with hardware architectures which make virtual to
41  * physical map invalidates expensive, this module may delay invalidate
42  * reduced protection operations until such time as they are actually
43  * necessary.  This module is given full information as to which processors
44  * are currently using which maps, and to when physical maps must be made
45  * correct.
46  */
47 
48 #include "opt_kstack_pages.h"
49 
50 #include <sys/param.h>
51 #include <sys/kernel.h>
52 #include <sys/conf.h>
53 #include <sys/queue.h>
54 #include <sys/cpuset.h>
55 #include <sys/kerneldump.h>
56 #include <sys/ktr.h>
57 #include <sys/lock.h>
58 #include <sys/msgbuf.h>
59 #include <sys/malloc.h>
60 #include <sys/mman.h>
61 #include <sys/mutex.h>
62 #include <sys/proc.h>
63 #include <sys/rwlock.h>
64 #include <sys/sched.h>
65 #include <sys/sysctl.h>
66 #include <sys/systm.h>
67 #include <sys/vmmeter.h>
68 #include <sys/smp.h>
69 #include <sys/reboot.h>
70 
71 #include <sys/kdb.h>
72 
73 #include <dev/ofw/openfirm.h>
74 
75 #include <vm/vm.h>
76 #include <vm/pmap.h>
77 #include <vm/vm_param.h>
78 #include <vm/vm_kern.h>
79 #include <vm/vm_page.h>
80 #include <vm/vm_phys.h>
81 #include <vm/vm_map.h>
82 #include <vm/vm_object.h>
83 #include <vm/vm_extern.h>
84 #include <vm/vm_pageout.h>
85 #include <vm/vm_dumpset.h>
86 #include <vm/vm_reserv.h>
87 #include <vm/uma.h>
88 
89 #include <machine/_inttypes.h>
90 #include <machine/cpu.h>
91 #include <machine/ifunc.h>
92 #include <machine/platform.h>
93 #include <machine/frame.h>
94 #include <machine/md_var.h>
95 #include <machine/psl.h>
96 #include <machine/bat.h>
97 #include <machine/hid.h>
98 #include <machine/pte.h>
99 #include <machine/sr.h>
100 #include <machine/trap.h>
101 #include <machine/mmuvar.h>
102 
103 #include "mmu_oea64.h"
104 
105 void moea64_release_vsid(uint64_t vsid);
106 uintptr_t moea64_get_unique_vsid(void);
107 
108 #define DISABLE_TRANS(msr)	msr = mfmsr(); mtmsr(msr & ~PSL_DR)
109 #define ENABLE_TRANS(msr)	mtmsr(msr)
110 
111 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
112 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
113 #define	VSID_HASH_MASK		0x0000007fffffffffULL
114 
115 /*
116  * Locking semantics:
117  *
118  * There are two locks of interest: the page locks and the pmap locks, which
119  * protect their individual PVO lists and are locked in that order. The contents
120  * of all PVO entries are protected by the locks of their respective pmaps.
121  * The pmap of any PVO is guaranteed not to change so long as the PVO is linked
122  * into any list.
123  *
124  */
125 
126 #define PV_LOCK_COUNT	PA_LOCK_COUNT
127 static struct mtx_padalign pv_lock[PV_LOCK_COUNT];
128 
129 /*
130  * Cheap NUMA-izing of the pv locks, to reduce contention across domains.
131  * NUMA domains on POWER9 appear to be indexed as sparse memory spaces, with the
132  * index at (N << 45).
133  */
134 #ifdef __powerpc64__
135 #define PV_LOCK_IDX(pa)	((pa_index(pa) * (((pa) >> 45) + 1)) % PV_LOCK_COUNT)
136 #else
137 #define PV_LOCK_IDX(pa)	(pa_index(pa) % PV_LOCK_COUNT)
138 #endif
139 #define PV_LOCKPTR(pa)	((struct mtx *)(&pv_lock[PV_LOCK_IDX(pa)]))
140 #define PV_LOCK(pa)		mtx_lock(PV_LOCKPTR(pa))
141 #define PV_UNLOCK(pa)		mtx_unlock(PV_LOCKPTR(pa))
142 #define PV_LOCKASSERT(pa) 	mtx_assert(PV_LOCKPTR(pa), MA_OWNED)
143 #define PV_PAGE_LOCK(m)		PV_LOCK(VM_PAGE_TO_PHYS(m))
144 #define PV_PAGE_UNLOCK(m)	PV_UNLOCK(VM_PAGE_TO_PHYS(m))
145 #define PV_PAGE_LOCKASSERT(m)	PV_LOCKASSERT(VM_PAGE_TO_PHYS(m))
146 
147 /* Superpage PV lock */
148 
149 #define	PV_LOCK_SIZE		(1<<PDRSHIFT)
150 
151 static __always_inline void
152 moea64_sp_pv_lock(vm_paddr_t pa)
153 {
154 	vm_paddr_t pa_end;
155 
156 	/* Note: breaking when pa_end is reached to avoid overflows */
157 	pa_end = pa + (HPT_SP_SIZE - PV_LOCK_SIZE);
158 	for (;;) {
159 		mtx_lock_flags(PV_LOCKPTR(pa), MTX_DUPOK);
160 		if (pa == pa_end)
161 			break;
162 		pa += PV_LOCK_SIZE;
163 	}
164 }
165 
166 static __always_inline void
167 moea64_sp_pv_unlock(vm_paddr_t pa)
168 {
169 	vm_paddr_t pa_end;
170 
171 	/* Note: breaking when pa_end is reached to avoid overflows */
172 	pa_end = pa;
173 	pa += HPT_SP_SIZE - PV_LOCK_SIZE;
174 	for (;;) {
175 		mtx_unlock_flags(PV_LOCKPTR(pa), MTX_DUPOK);
176 		if (pa == pa_end)
177 			break;
178 		pa -= PV_LOCK_SIZE;
179 	}
180 }
181 
182 #define	SP_PV_LOCK_ALIGNED(pa)		moea64_sp_pv_lock(pa)
183 #define	SP_PV_UNLOCK_ALIGNED(pa)	moea64_sp_pv_unlock(pa)
184 #define	SP_PV_LOCK(pa)			moea64_sp_pv_lock((pa) & ~HPT_SP_MASK)
185 #define	SP_PV_UNLOCK(pa)		moea64_sp_pv_unlock((pa) & ~HPT_SP_MASK)
186 #define	SP_PV_PAGE_LOCK(m)		SP_PV_LOCK(VM_PAGE_TO_PHYS(m))
187 #define	SP_PV_PAGE_UNLOCK(m)		SP_PV_UNLOCK(VM_PAGE_TO_PHYS(m))
188 
189 struct ofw_map {
190 	cell_t	om_va;
191 	cell_t	om_len;
192 	uint64_t om_pa;
193 	cell_t	om_mode;
194 };
195 
196 extern unsigned char _etext[];
197 extern unsigned char _end[];
198 
199 extern void *slbtrap, *slbtrapend;
200 
201 /*
202  * Map of physical memory regions.
203  */
204 static struct	mem_region *regions;
205 static struct	mem_region *pregions;
206 static struct	numa_mem_region *numa_pregions;
207 static u_int	phys_avail_count;
208 static int	regions_sz, pregions_sz, numapregions_sz;
209 
210 extern void bs_remap_earlyboot(void);
211 
212 /*
213  * Lock for the SLB tables.
214  */
215 struct mtx	moea64_slb_mutex;
216 
217 /*
218  * PTEG data.
219  */
220 u_long		moea64_pteg_count;
221 u_long		moea64_pteg_mask;
222 
223 /*
224  * PVO data.
225  */
226 
227 uma_zone_t	moea64_pvo_zone; /* zone for pvo entries */
228 
229 static struct	pvo_entry *moea64_bpvo_pool;
230 static int	moea64_bpvo_pool_index = 0;
231 static int	moea64_bpvo_pool_size = 0;
232 SYSCTL_INT(_machdep, OID_AUTO, moea64_allocated_bpvo_entries, CTLFLAG_RD,
233     &moea64_bpvo_pool_index, 0, "");
234 
235 #define	BPVO_POOL_SIZE	327680 /* Sensible historical default value */
236 #define	BPVO_POOL_EXPANSION_FACTOR	3
237 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
238 #ifdef __powerpc64__
239 #define	NVSIDS		(NPMAPS * 16)
240 #define VSID_HASHMASK	0xffffffffUL
241 #else
242 #define NVSIDS		NPMAPS
243 #define VSID_HASHMASK	0xfffffUL
244 #endif
245 static u_int	moea64_vsid_bitmap[NVSIDS / VSID_NBPW];
246 
247 static boolean_t moea64_initialized = FALSE;
248 
249 #ifdef MOEA64_STATS
250 /*
251  * Statistics.
252  */
253 u_int	moea64_pte_valid = 0;
254 u_int	moea64_pte_overflow = 0;
255 u_int	moea64_pvo_entries = 0;
256 u_int	moea64_pvo_enter_calls = 0;
257 u_int	moea64_pvo_remove_calls = 0;
258 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_valid, CTLFLAG_RD,
259     &moea64_pte_valid, 0, "");
260 SYSCTL_INT(_machdep, OID_AUTO, moea64_pte_overflow, CTLFLAG_RD,
261     &moea64_pte_overflow, 0, "");
262 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_entries, CTLFLAG_RD,
263     &moea64_pvo_entries, 0, "");
264 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_enter_calls, CTLFLAG_RD,
265     &moea64_pvo_enter_calls, 0, "");
266 SYSCTL_INT(_machdep, OID_AUTO, moea64_pvo_remove_calls, CTLFLAG_RD,
267     &moea64_pvo_remove_calls, 0, "");
268 #endif
269 
270 vm_offset_t	moea64_scratchpage_va[2];
271 struct pvo_entry *moea64_scratchpage_pvo[2];
272 struct	mtx	moea64_scratchpage_mtx;
273 
274 uint64_t 	moea64_large_page_mask = 0;
275 uint64_t	moea64_large_page_size = 0;
276 int		moea64_large_page_shift = 0;
277 bool		moea64_has_lp_4k_16m = false;
278 
279 /*
280  * PVO calls.
281  */
282 static int	moea64_pvo_enter(struct pvo_entry *pvo,
283 		    struct pvo_head *pvo_head, struct pvo_entry **oldpvo);
284 static void	moea64_pvo_remove_from_pmap(struct pvo_entry *pvo);
285 static void	moea64_pvo_remove_from_page(struct pvo_entry *pvo);
286 static void	moea64_pvo_remove_from_page_locked(
287 		    struct pvo_entry *pvo, vm_page_t m);
288 static struct	pvo_entry *moea64_pvo_find_va(pmap_t, vm_offset_t);
289 
290 /*
291  * Utility routines.
292  */
293 static boolean_t	moea64_query_bit(vm_page_t, uint64_t);
294 static u_int		moea64_clear_bit(vm_page_t, uint64_t);
295 static void		moea64_kremove(vm_offset_t);
296 static void		moea64_syncicache(pmap_t pmap, vm_offset_t va,
297 			    vm_paddr_t pa, vm_size_t sz);
298 static void		moea64_pmap_init_qpages(void);
299 static void		moea64_remove_locked(pmap_t, vm_offset_t,
300 			    vm_offset_t, struct pvo_dlist *);
301 
302 /*
303  * Superpages data and routines.
304  */
305 
306 /*
307  * PVO flags (in vaddr) that must match for promotion to succeed.
308  * Note that protection bits are checked separately, as they reside in
309  * another field.
310  */
311 #define	PVO_FLAGS_PROMOTE	(PVO_WIRED | PVO_MANAGED | PVO_PTEGIDX_VALID)
312 
313 #define	PVO_IS_SP(pvo)		(((pvo)->pvo_vaddr & PVO_LARGE) && \
314 				 (pvo)->pvo_pmap != kernel_pmap)
315 
316 /* Get physical address from PVO. */
317 #define	PVO_PADDR(pvo)		moea64_pvo_paddr(pvo)
318 
319 /* MD page flag indicating that the page is a superpage. */
320 #define	MDPG_ATTR_SP		0x40000000
321 
322 SYSCTL_DECL(_vm_pmap);
323 
324 static SYSCTL_NODE(_vm_pmap, OID_AUTO, sp, CTLFLAG_RD, 0,
325     "SP page mapping counters");
326 
327 static u_long sp_demotions;
328 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, demotions, CTLFLAG_RD,
329     &sp_demotions, 0, "SP page demotions");
330 
331 static u_long sp_mappings;
332 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, mappings, CTLFLAG_RD,
333     &sp_mappings, 0, "SP page mappings");
334 
335 static u_long sp_p_failures;
336 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_failures, CTLFLAG_RD,
337     &sp_p_failures, 0, "SP page promotion failures");
338 
339 static u_long sp_p_fail_pa;
340 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_fail_pa, CTLFLAG_RD,
341     &sp_p_fail_pa, 0, "SP page promotion failure: PAs don't match");
342 
343 static u_long sp_p_fail_flags;
344 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_fail_flags, CTLFLAG_RD,
345     &sp_p_fail_flags, 0, "SP page promotion failure: page flags don't match");
346 
347 static u_long sp_p_fail_prot;
348 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_fail_prot, CTLFLAG_RD,
349     &sp_p_fail_prot, 0,
350     "SP page promotion failure: page protections don't match");
351 
352 static u_long sp_p_fail_wimg;
353 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, p_fail_wimg, CTLFLAG_RD,
354     &sp_p_fail_wimg, 0, "SP page promotion failure: WIMG bits don't match");
355 
356 static u_long sp_promotions;
357 SYSCTL_ULONG(_vm_pmap_sp, OID_AUTO, promotions, CTLFLAG_RD,
358     &sp_promotions, 0, "SP page promotions");
359 
360 static bool moea64_ps_enabled(pmap_t);
361 static void moea64_align_superpage(vm_object_t, vm_ooffset_t,
362     vm_offset_t *, vm_size_t);
363 
364 static int moea64_sp_enter(pmap_t pmap, vm_offset_t va,
365     vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind);
366 static struct pvo_entry *moea64_sp_remove(struct pvo_entry *sp,
367     struct pvo_dlist *tofree);
368 
369 static void moea64_sp_promote(pmap_t pmap, vm_offset_t va, vm_page_t m);
370 static void moea64_sp_demote_aligned(struct pvo_entry *sp);
371 static void moea64_sp_demote(struct pvo_entry *pvo);
372 
373 static struct pvo_entry *moea64_sp_unwire(struct pvo_entry *sp);
374 static struct pvo_entry *moea64_sp_protect(struct pvo_entry *sp,
375     vm_prot_t prot);
376 
377 static int64_t moea64_sp_query(struct pvo_entry *pvo, uint64_t ptebit);
378 static int64_t moea64_sp_clear(struct pvo_entry *pvo, vm_page_t m,
379     uint64_t ptebit);
380 
381 static __inline bool moea64_sp_pvo_in_range(struct pvo_entry *pvo,
382     vm_offset_t sva, vm_offset_t eva);
383 
384 /*
385  * Kernel MMU interface
386  */
387 void moea64_clear_modify(vm_page_t);
388 void moea64_copy_page(vm_page_t, vm_page_t);
389 void moea64_copy_page_dmap(vm_page_t, vm_page_t);
390 void moea64_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
391     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
392 void moea64_copy_pages_dmap(vm_page_t *ma, vm_offset_t a_offset,
393     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
394 int moea64_enter(pmap_t, vm_offset_t, vm_page_t, vm_prot_t,
395     u_int flags, int8_t psind);
396 void moea64_enter_object(pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
397     vm_prot_t);
398 void moea64_enter_quick(pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
399 vm_paddr_t moea64_extract(pmap_t, vm_offset_t);
400 vm_page_t moea64_extract_and_hold(pmap_t, vm_offset_t, vm_prot_t);
401 void moea64_init(void);
402 boolean_t moea64_is_modified(vm_page_t);
403 boolean_t moea64_is_prefaultable(pmap_t, vm_offset_t);
404 boolean_t moea64_is_referenced(vm_page_t);
405 int moea64_ts_referenced(vm_page_t);
406 vm_offset_t moea64_map(vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
407 boolean_t moea64_page_exists_quick(pmap_t, vm_page_t);
408 void moea64_page_init(vm_page_t);
409 int moea64_page_wired_mappings(vm_page_t);
410 int moea64_pinit(pmap_t);
411 void moea64_pinit0(pmap_t);
412 void moea64_protect(pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
413 void moea64_qenter(vm_offset_t, vm_page_t *, int);
414 void moea64_qremove(vm_offset_t, int);
415 void moea64_release(pmap_t);
416 void moea64_remove(pmap_t, vm_offset_t, vm_offset_t);
417 void moea64_remove_pages(pmap_t);
418 void moea64_remove_all(vm_page_t);
419 void moea64_remove_write(vm_page_t);
420 void moea64_unwire(pmap_t, vm_offset_t, vm_offset_t);
421 void moea64_zero_page(vm_page_t);
422 void moea64_zero_page_dmap(vm_page_t);
423 void moea64_zero_page_area(vm_page_t, int, int);
424 void moea64_activate(struct thread *);
425 void moea64_deactivate(struct thread *);
426 void *moea64_mapdev(vm_paddr_t, vm_size_t);
427 void *moea64_mapdev_attr(vm_paddr_t, vm_size_t, vm_memattr_t);
428 void moea64_unmapdev(vm_offset_t, vm_size_t);
429 vm_paddr_t moea64_kextract(vm_offset_t);
430 void moea64_page_set_memattr(vm_page_t m, vm_memattr_t ma);
431 void moea64_kenter_attr(vm_offset_t, vm_paddr_t, vm_memattr_t ma);
432 void moea64_kenter(vm_offset_t, vm_paddr_t);
433 boolean_t moea64_dev_direct_mapped(vm_paddr_t, vm_size_t);
434 static void moea64_sync_icache(pmap_t, vm_offset_t, vm_size_t);
435 void moea64_dumpsys_map(vm_paddr_t pa, size_t sz,
436     void **va);
437 void moea64_scan_init(void);
438 vm_offset_t moea64_quick_enter_page(vm_page_t m);
439 vm_offset_t moea64_quick_enter_page_dmap(vm_page_t m);
440 void moea64_quick_remove_page(vm_offset_t addr);
441 boolean_t moea64_page_is_mapped(vm_page_t m);
442 static int moea64_map_user_ptr(pmap_t pm,
443     volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen);
444 static int moea64_decode_kernel_ptr(vm_offset_t addr,
445     int *is_user, vm_offset_t *decoded_addr);
446 static size_t moea64_scan_pmap(struct bitset *dump_bitset);
447 static void *moea64_dump_pmap_init(unsigned blkpgs);
448 #ifdef __powerpc64__
449 static void moea64_page_array_startup(long);
450 #endif
451 static int moea64_mincore(pmap_t, vm_offset_t, vm_paddr_t *);
452 
453 static struct pmap_funcs moea64_methods = {
454 	.clear_modify = moea64_clear_modify,
455 	.copy_page = moea64_copy_page,
456 	.copy_pages = moea64_copy_pages,
457 	.enter = moea64_enter,
458 	.enter_object = moea64_enter_object,
459 	.enter_quick = moea64_enter_quick,
460 	.extract = moea64_extract,
461 	.extract_and_hold = moea64_extract_and_hold,
462 	.init = moea64_init,
463 	.is_modified = moea64_is_modified,
464 	.is_prefaultable = moea64_is_prefaultable,
465 	.is_referenced = moea64_is_referenced,
466 	.ts_referenced = moea64_ts_referenced,
467 	.map =      		moea64_map,
468 	.mincore = moea64_mincore,
469 	.page_exists_quick = moea64_page_exists_quick,
470 	.page_init = moea64_page_init,
471 	.page_wired_mappings = moea64_page_wired_mappings,
472 	.pinit = moea64_pinit,
473 	.pinit0 = moea64_pinit0,
474 	.protect = moea64_protect,
475 	.qenter = moea64_qenter,
476 	.qremove = moea64_qremove,
477 	.release = moea64_release,
478 	.remove = moea64_remove,
479 	.remove_pages = moea64_remove_pages,
480 	.remove_all =       	moea64_remove_all,
481 	.remove_write = moea64_remove_write,
482 	.sync_icache = moea64_sync_icache,
483 	.unwire = moea64_unwire,
484 	.zero_page =        	moea64_zero_page,
485 	.zero_page_area = moea64_zero_page_area,
486 	.activate = moea64_activate,
487 	.deactivate =       	moea64_deactivate,
488 	.page_set_memattr = moea64_page_set_memattr,
489 	.quick_enter_page =  moea64_quick_enter_page,
490 	.quick_remove_page =  moea64_quick_remove_page,
491 	.page_is_mapped = moea64_page_is_mapped,
492 #ifdef __powerpc64__
493 	.page_array_startup = moea64_page_array_startup,
494 #endif
495 	.ps_enabled = moea64_ps_enabled,
496 	.align_superpage = moea64_align_superpage,
497 
498 	/* Internal interfaces */
499 	.mapdev = moea64_mapdev,
500 	.mapdev_attr = moea64_mapdev_attr,
501 	.unmapdev = moea64_unmapdev,
502 	.kextract = moea64_kextract,
503 	.kenter = moea64_kenter,
504 	.kenter_attr = moea64_kenter_attr,
505 	.dev_direct_mapped = moea64_dev_direct_mapped,
506 	.dumpsys_pa_init = moea64_scan_init,
507 	.dumpsys_scan_pmap = moea64_scan_pmap,
508 	.dumpsys_dump_pmap_init =    moea64_dump_pmap_init,
509 	.dumpsys_map_chunk = moea64_dumpsys_map,
510 	.map_user_ptr = moea64_map_user_ptr,
511 	.decode_kernel_ptr =  moea64_decode_kernel_ptr,
512 };
513 
514 MMU_DEF(oea64_mmu, "mmu_oea64_base", moea64_methods);
515 
516 /*
517  * Get physical address from PVO.
518  *
519  * For superpages, the lower bits are not stored on pvo_pte.pa and must be
520  * obtained from VA.
521  */
522 static __always_inline vm_paddr_t
523 moea64_pvo_paddr(struct pvo_entry *pvo)
524 {
525 	vm_paddr_t pa;
526 
527 	pa = (pvo)->pvo_pte.pa & LPTE_RPGN;
528 
529 	if (PVO_IS_SP(pvo)) {
530 		pa &= ~HPT_SP_MASK; /* This is needed to clear LPTE_LP bits. */
531 		pa |= PVO_VADDR(pvo) & HPT_SP_MASK;
532 	}
533 	return (pa);
534 }
535 
536 static struct pvo_head *
537 vm_page_to_pvoh(vm_page_t m)
538 {
539 
540 	mtx_assert(PV_LOCKPTR(VM_PAGE_TO_PHYS(m)), MA_OWNED);
541 	return (&m->md.mdpg_pvoh);
542 }
543 
544 static struct pvo_entry *
545 alloc_pvo_entry(int bootstrap)
546 {
547 	struct pvo_entry *pvo;
548 
549 	if (!moea64_initialized || bootstrap) {
550 		if (moea64_bpvo_pool_index >= moea64_bpvo_pool_size) {
551 			panic("%s: bpvo pool exhausted, index=%d, size=%d, bytes=%zd."
552 			    "Try setting machdep.moea64_bpvo_pool_size tunable",
553 			    __func__, moea64_bpvo_pool_index,
554 			    moea64_bpvo_pool_size,
555 			    moea64_bpvo_pool_size * sizeof(struct pvo_entry));
556 		}
557 		pvo = &moea64_bpvo_pool[
558 		    atomic_fetchadd_int(&moea64_bpvo_pool_index, 1)];
559 		bzero(pvo, sizeof(*pvo));
560 		pvo->pvo_vaddr = PVO_BOOTSTRAP;
561 	} else
562 		pvo = uma_zalloc(moea64_pvo_zone, M_NOWAIT | M_ZERO);
563 
564 	return (pvo);
565 }
566 
567 static void
568 init_pvo_entry(struct pvo_entry *pvo, pmap_t pmap, vm_offset_t va)
569 {
570 	uint64_t vsid;
571 	uint64_t hash;
572 	int shift;
573 
574 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
575 
576 	pvo->pvo_pmap = pmap;
577 	va &= ~ADDR_POFF;
578 	pvo->pvo_vaddr |= va;
579 	vsid = va_to_vsid(pmap, va);
580 	pvo->pvo_vpn = (uint64_t)((va & ADDR_PIDX) >> ADDR_PIDX_SHFT)
581 	    | (vsid << 16);
582 
583 	if (pmap == kernel_pmap && (pvo->pvo_vaddr & PVO_LARGE) != 0)
584 		shift = moea64_large_page_shift;
585 	else
586 		shift = ADDR_PIDX_SHFT;
587 	hash = (vsid & VSID_HASH_MASK) ^ (((uint64_t)va & ADDR_PIDX) >> shift);
588 	pvo->pvo_pte.slot = (hash & moea64_pteg_mask) << 3;
589 }
590 
591 static void
592 free_pvo_entry(struct pvo_entry *pvo)
593 {
594 
595 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
596 		uma_zfree(moea64_pvo_zone, pvo);
597 }
598 
599 void
600 moea64_pte_from_pvo(const struct pvo_entry *pvo, struct lpte *lpte)
601 {
602 
603 	lpte->pte_hi = moea64_pte_vpn_from_pvo_vpn(pvo);
604 	lpte->pte_hi |= LPTE_VALID;
605 
606 	if (pvo->pvo_vaddr & PVO_LARGE)
607 		lpte->pte_hi |= LPTE_BIG;
608 	if (pvo->pvo_vaddr & PVO_WIRED)
609 		lpte->pte_hi |= LPTE_WIRED;
610 	if (pvo->pvo_vaddr & PVO_HID)
611 		lpte->pte_hi |= LPTE_HID;
612 
613 	lpte->pte_lo = pvo->pvo_pte.pa; /* Includes WIMG bits */
614 	if (pvo->pvo_pte.prot & VM_PROT_WRITE)
615 		lpte->pte_lo |= LPTE_BW;
616 	else
617 		lpte->pte_lo |= LPTE_BR;
618 
619 	if (!(pvo->pvo_pte.prot & VM_PROT_EXECUTE))
620 		lpte->pte_lo |= LPTE_NOEXEC;
621 }
622 
623 static __inline uint64_t
624 moea64_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
625 {
626 	uint64_t pte_lo;
627 	int i;
628 
629 	if (ma != VM_MEMATTR_DEFAULT) {
630 		switch (ma) {
631 		case VM_MEMATTR_UNCACHEABLE:
632 			return (LPTE_I | LPTE_G);
633 		case VM_MEMATTR_CACHEABLE:
634 			return (LPTE_M);
635 		case VM_MEMATTR_WRITE_COMBINING:
636 		case VM_MEMATTR_WRITE_BACK:
637 		case VM_MEMATTR_PREFETCHABLE:
638 			return (LPTE_I);
639 		case VM_MEMATTR_WRITE_THROUGH:
640 			return (LPTE_W | LPTE_M);
641 		}
642 	}
643 
644 	/*
645 	 * Assume the page is cache inhibited and access is guarded unless
646 	 * it's in our available memory array.
647 	 */
648 	pte_lo = LPTE_I | LPTE_G;
649 	for (i = 0; i < pregions_sz; i++) {
650 		if ((pa >= pregions[i].mr_start) &&
651 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
652 			pte_lo &= ~(LPTE_I | LPTE_G);
653 			pte_lo |= LPTE_M;
654 			break;
655 		}
656 	}
657 
658 	return pte_lo;
659 }
660 
661 /*
662  * Quick sort callout for comparing memory regions.
663  */
664 static int	om_cmp(const void *a, const void *b);
665 
666 static int
667 om_cmp(const void *a, const void *b)
668 {
669 	const struct	ofw_map *mapa;
670 	const struct	ofw_map *mapb;
671 
672 	mapa = a;
673 	mapb = b;
674 	if (mapa->om_pa < mapb->om_pa)
675 		return (-1);
676 	else if (mapa->om_pa > mapb->om_pa)
677 		return (1);
678 	else
679 		return (0);
680 }
681 
682 static void
683 moea64_add_ofw_mappings(phandle_t mmu, size_t sz)
684 {
685 	struct ofw_map	translations[sz/(4*sizeof(cell_t))]; /*>= 4 cells per */
686 	pcell_t		acells, trans_cells[sz/sizeof(cell_t)];
687 	struct pvo_entry *pvo;
688 	register_t	msr;
689 	vm_offset_t	off;
690 	vm_paddr_t	pa_base;
691 	int		i, j;
692 
693 	bzero(translations, sz);
694 	OF_getencprop(OF_finddevice("/"), "#address-cells", &acells,
695 	    sizeof(acells));
696 	if (OF_getencprop(mmu, "translations", trans_cells, sz) == -1)
697 		panic("moea64_bootstrap: can't get ofw translations");
698 
699 	CTR0(KTR_PMAP, "moea64_add_ofw_mappings: translations");
700 	sz /= sizeof(cell_t);
701 	for (i = 0, j = 0; i < sz; j++) {
702 		translations[j].om_va = trans_cells[i++];
703 		translations[j].om_len = trans_cells[i++];
704 		translations[j].om_pa = trans_cells[i++];
705 		if (acells == 2) {
706 			translations[j].om_pa <<= 32;
707 			translations[j].om_pa |= trans_cells[i++];
708 		}
709 		translations[j].om_mode = trans_cells[i++];
710 	}
711 	KASSERT(i == sz, ("Translations map has incorrect cell count (%d/%zd)",
712 	    i, sz));
713 
714 	sz = j;
715 	qsort(translations, sz, sizeof (*translations), om_cmp);
716 
717 	for (i = 0; i < sz; i++) {
718 		pa_base = translations[i].om_pa;
719 	      #ifndef __powerpc64__
720 		if ((translations[i].om_pa >> 32) != 0)
721 			panic("OFW translations above 32-bit boundary!");
722 	      #endif
723 
724 		if (pa_base % PAGE_SIZE)
725 			panic("OFW translation not page-aligned (phys)!");
726 		if (translations[i].om_va % PAGE_SIZE)
727 			panic("OFW translation not page-aligned (virt)!");
728 
729 		CTR3(KTR_PMAP, "translation: pa=%#zx va=%#x len=%#x",
730 		    pa_base, translations[i].om_va, translations[i].om_len);
731 
732 		/* Now enter the pages for this mapping */
733 
734 		DISABLE_TRANS(msr);
735 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
736 			/* If this address is direct-mapped, skip remapping */
737 			if (hw_direct_map &&
738 			    translations[i].om_va == PHYS_TO_DMAP(pa_base) &&
739 			    moea64_calc_wimg(pa_base + off, VM_MEMATTR_DEFAULT)
740  			    == LPTE_M)
741 				continue;
742 
743 			PMAP_LOCK(kernel_pmap);
744 			pvo = moea64_pvo_find_va(kernel_pmap,
745 			    translations[i].om_va + off);
746 			PMAP_UNLOCK(kernel_pmap);
747 			if (pvo != NULL)
748 				continue;
749 
750 			moea64_kenter(translations[i].om_va + off,
751 			    pa_base + off);
752 		}
753 		ENABLE_TRANS(msr);
754 	}
755 }
756 
757 #ifdef __powerpc64__
758 static void
759 moea64_probe_large_page(void)
760 {
761 	uint16_t pvr = mfpvr() >> 16;
762 
763 	switch (pvr) {
764 	case IBM970:
765 	case IBM970FX:
766 	case IBM970MP:
767 		powerpc_sync(); isync();
768 		mtspr(SPR_HID4, mfspr(SPR_HID4) & ~HID4_970_DISABLE_LG_PG);
769 		powerpc_sync(); isync();
770 
771 		/* FALLTHROUGH */
772 	default:
773 		if (moea64_large_page_size == 0) {
774 			moea64_large_page_size = 0x1000000; /* 16 MB */
775 			moea64_large_page_shift = 24;
776 		}
777 	}
778 
779 	moea64_large_page_mask = moea64_large_page_size - 1;
780 }
781 
782 static void
783 moea64_bootstrap_slb_prefault(vm_offset_t va, int large)
784 {
785 	struct slb *cache;
786 	struct slb entry;
787 	uint64_t esid, slbe;
788 	uint64_t i;
789 
790 	cache = PCPU_GET(aim.slb);
791 	esid = va >> ADDR_SR_SHFT;
792 	slbe = (esid << SLBE_ESID_SHIFT) | SLBE_VALID;
793 
794 	for (i = 0; i < 64; i++) {
795 		if (cache[i].slbe == (slbe | i))
796 			return;
797 	}
798 
799 	entry.slbe = slbe;
800 	entry.slbv = KERNEL_VSID(esid) << SLBV_VSID_SHIFT;
801 	if (large)
802 		entry.slbv |= SLBV_L;
803 
804 	slb_insert_kernel(entry.slbe, entry.slbv);
805 }
806 #endif
807 
808 static int
809 moea64_kenter_large(vm_offset_t va, vm_paddr_t pa, uint64_t attr, int bootstrap)
810 {
811 	struct pvo_entry *pvo;
812 	uint64_t pte_lo;
813 	int error;
814 
815 	pte_lo = LPTE_M;
816 	pte_lo |= attr;
817 
818 	pvo = alloc_pvo_entry(bootstrap);
819 	pvo->pvo_vaddr |= PVO_WIRED | PVO_LARGE;
820 	init_pvo_entry(pvo, kernel_pmap, va);
821 
822 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE |
823 	    VM_PROT_EXECUTE;
824 	pvo->pvo_pte.pa = pa | pte_lo;
825 	error = moea64_pvo_enter(pvo, NULL, NULL);
826 	if (error != 0)
827 		panic("Error %d inserting large page\n", error);
828 	return (0);
829 }
830 
831 static void
832 moea64_setup_direct_map(vm_offset_t kernelstart,
833     vm_offset_t kernelend)
834 {
835 	register_t msr;
836 	vm_paddr_t pa, pkernelstart, pkernelend;
837 	vm_offset_t size, off;
838 	uint64_t pte_lo;
839 	int i;
840 
841 	if (moea64_large_page_size == 0)
842 		hw_direct_map = 0;
843 
844 	DISABLE_TRANS(msr);
845 	if (hw_direct_map) {
846 		PMAP_LOCK(kernel_pmap);
847 		for (i = 0; i < pregions_sz; i++) {
848 		  for (pa = pregions[i].mr_start; pa < pregions[i].mr_start +
849 		     pregions[i].mr_size; pa += moea64_large_page_size) {
850 			pte_lo = LPTE_M;
851 			if (pa & moea64_large_page_mask) {
852 				pa &= moea64_large_page_mask;
853 				pte_lo |= LPTE_G;
854 			}
855 			if (pa + moea64_large_page_size >
856 			    pregions[i].mr_start + pregions[i].mr_size)
857 				pte_lo |= LPTE_G;
858 
859 			moea64_kenter_large(PHYS_TO_DMAP(pa), pa, pte_lo, 1);
860 		  }
861 		}
862 		PMAP_UNLOCK(kernel_pmap);
863 	}
864 
865 	/*
866 	 * Make sure the kernel and BPVO pool stay mapped on systems either
867 	 * without a direct map or on which the kernel is not already executing
868 	 * out of the direct-mapped region.
869 	 */
870 	if (kernelstart < DMAP_BASE_ADDRESS) {
871 		/*
872 		 * For pre-dmap execution, we need to use identity mapping
873 		 * because we will be operating with the mmu on but in the
874 		 * wrong address configuration until we __restartkernel().
875 		 */
876 		for (pa = kernelstart & ~PAGE_MASK; pa < kernelend;
877 		    pa += PAGE_SIZE)
878 			moea64_kenter(pa, pa);
879 	} else if (!hw_direct_map) {
880 		pkernelstart = kernelstart & ~DMAP_BASE_ADDRESS;
881 		pkernelend = kernelend & ~DMAP_BASE_ADDRESS;
882 		for (pa = pkernelstart & ~PAGE_MASK; pa < pkernelend;
883 		    pa += PAGE_SIZE)
884 			moea64_kenter(pa | DMAP_BASE_ADDRESS, pa);
885 	}
886 
887 	if (!hw_direct_map) {
888 		size = moea64_bpvo_pool_size*sizeof(struct pvo_entry);
889 		off = (vm_offset_t)(moea64_bpvo_pool);
890 		for (pa = off; pa < off + size; pa += PAGE_SIZE)
891 			moea64_kenter(pa, pa);
892 
893 		/* Map exception vectors */
894 		for (pa = EXC_RSVD; pa < EXC_LAST; pa += PAGE_SIZE)
895 			moea64_kenter(pa | DMAP_BASE_ADDRESS, pa);
896 	}
897 	ENABLE_TRANS(msr);
898 
899 	/*
900 	 * Allow user to override unmapped_buf_allowed for testing.
901 	 * XXXKIB Only direct map implementation was tested.
902 	 */
903 	if (!TUNABLE_INT_FETCH("vfs.unmapped_buf_allowed",
904 	    &unmapped_buf_allowed))
905 		unmapped_buf_allowed = hw_direct_map;
906 }
907 
908 /* Quick sort callout for comparing physical addresses. */
909 static int
910 pa_cmp(const void *a, const void *b)
911 {
912 	const vm_paddr_t *pa = a, *pb = b;
913 
914 	if (*pa < *pb)
915 		return (-1);
916 	else if (*pa > *pb)
917 		return (1);
918 	else
919 		return (0);
920 }
921 
922 void
923 moea64_early_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
924 {
925 	int		i, j;
926 	vm_size_t	physsz, hwphyssz;
927 	vm_paddr_t	kernelphysstart, kernelphysend;
928 	int		rm_pavail;
929 
930 	/* Level 0 reservations consist of 4096 pages (16MB superpage). */
931 	vm_level_0_order = 12;
932 
933 #ifndef __powerpc64__
934 	/* We don't have a direct map since there is no BAT */
935 	hw_direct_map = 0;
936 
937 	/* Make sure battable is zero, since we have no BAT */
938 	for (i = 0; i < 16; i++) {
939 		battable[i].batu = 0;
940 		battable[i].batl = 0;
941 	}
942 #else
943 	/* Install trap handlers for SLBs */
944 	bcopy(&slbtrap, (void *)EXC_DSE,(size_t)&slbtrapend - (size_t)&slbtrap);
945 	bcopy(&slbtrap, (void *)EXC_ISE,(size_t)&slbtrapend - (size_t)&slbtrap);
946 	__syncicache((void *)EXC_DSE, 0x80);
947 	__syncicache((void *)EXC_ISE, 0x80);
948 #endif
949 
950 	kernelphysstart = kernelstart & ~DMAP_BASE_ADDRESS;
951 	kernelphysend = kernelend & ~DMAP_BASE_ADDRESS;
952 
953 	/* Get physical memory regions from firmware */
954 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
955 	CTR0(KTR_PMAP, "moea64_bootstrap: physical memory");
956 
957 	if (PHYS_AVAIL_ENTRIES < regions_sz)
958 		panic("moea64_bootstrap: phys_avail too small");
959 
960 	phys_avail_count = 0;
961 	physsz = 0;
962 	hwphyssz = 0;
963 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
964 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
965 		CTR3(KTR_PMAP, "region: %#zx - %#zx (%#zx)",
966 		    regions[i].mr_start, regions[i].mr_start +
967 		    regions[i].mr_size, regions[i].mr_size);
968 		if (hwphyssz != 0 &&
969 		    (physsz + regions[i].mr_size) >= hwphyssz) {
970 			if (physsz < hwphyssz) {
971 				phys_avail[j] = regions[i].mr_start;
972 				phys_avail[j + 1] = regions[i].mr_start +
973 				    hwphyssz - physsz;
974 				physsz = hwphyssz;
975 				phys_avail_count++;
976 				dump_avail[j] = phys_avail[j];
977 				dump_avail[j + 1] = phys_avail[j + 1];
978 			}
979 			break;
980 		}
981 		phys_avail[j] = regions[i].mr_start;
982 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
983 		phys_avail_count++;
984 		physsz += regions[i].mr_size;
985 		dump_avail[j] = phys_avail[j];
986 		dump_avail[j + 1] = phys_avail[j + 1];
987 	}
988 
989 	/* Check for overlap with the kernel and exception vectors */
990 	rm_pavail = 0;
991 	for (j = 0; j < 2*phys_avail_count; j+=2) {
992 		if (phys_avail[j] < EXC_LAST)
993 			phys_avail[j] += EXC_LAST;
994 
995 		if (phys_avail[j] >= kernelphysstart &&
996 		    phys_avail[j+1] <= kernelphysend) {
997 			phys_avail[j] = phys_avail[j+1] = ~0;
998 			rm_pavail++;
999 			continue;
1000 		}
1001 
1002 		if (kernelphysstart >= phys_avail[j] &&
1003 		    kernelphysstart < phys_avail[j+1]) {
1004 			if (kernelphysend < phys_avail[j+1]) {
1005 				phys_avail[2*phys_avail_count] =
1006 				    (kernelphysend & ~PAGE_MASK) + PAGE_SIZE;
1007 				phys_avail[2*phys_avail_count + 1] =
1008 				    phys_avail[j+1];
1009 				phys_avail_count++;
1010 			}
1011 
1012 			phys_avail[j+1] = kernelphysstart & ~PAGE_MASK;
1013 		}
1014 
1015 		if (kernelphysend >= phys_avail[j] &&
1016 		    kernelphysend < phys_avail[j+1]) {
1017 			if (kernelphysstart > phys_avail[j]) {
1018 				phys_avail[2*phys_avail_count] = phys_avail[j];
1019 				phys_avail[2*phys_avail_count + 1] =
1020 				    kernelphysstart & ~PAGE_MASK;
1021 				phys_avail_count++;
1022 			}
1023 
1024 			phys_avail[j] = (kernelphysend & ~PAGE_MASK) +
1025 			    PAGE_SIZE;
1026 		}
1027 	}
1028 
1029 	/* Remove physical available regions marked for removal (~0) */
1030 	if (rm_pavail) {
1031 		qsort(phys_avail, 2*phys_avail_count, sizeof(phys_avail[0]),
1032 			pa_cmp);
1033 		phys_avail_count -= rm_pavail;
1034 		for (i = 2*phys_avail_count;
1035 		     i < 2*(phys_avail_count + rm_pavail); i+=2)
1036 			phys_avail[i] = phys_avail[i+1] = 0;
1037 	}
1038 
1039 	physmem = btoc(physsz);
1040 
1041 #ifdef PTEGCOUNT
1042 	moea64_pteg_count = PTEGCOUNT;
1043 #else
1044 	moea64_pteg_count = 0x1000;
1045 
1046 	while (moea64_pteg_count < physmem)
1047 		moea64_pteg_count <<= 1;
1048 
1049 	moea64_pteg_count >>= 1;
1050 #endif /* PTEGCOUNT */
1051 }
1052 
1053 void
1054 moea64_mid_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
1055 {
1056 	int		i;
1057 
1058 	/*
1059 	 * Set PTEG mask
1060 	 */
1061 	moea64_pteg_mask = moea64_pteg_count - 1;
1062 
1063 	/*
1064 	 * Initialize SLB table lock and page locks
1065 	 */
1066 	mtx_init(&moea64_slb_mutex, "SLB table", NULL, MTX_DEF);
1067 	for (i = 0; i < PV_LOCK_COUNT; i++)
1068 		mtx_init(&pv_lock[i], "page pv", NULL, MTX_DEF);
1069 
1070 	/*
1071 	 * Initialise the bootstrap pvo pool.
1072 	 */
1073 	TUNABLE_INT_FETCH("machdep.moea64_bpvo_pool_size", &moea64_bpvo_pool_size);
1074 	if (moea64_bpvo_pool_size == 0) {
1075 		if (!hw_direct_map)
1076 			moea64_bpvo_pool_size = ((ptoa((uintmax_t)physmem) * sizeof(struct vm_page)) /
1077 			    (PAGE_SIZE * PAGE_SIZE)) * BPVO_POOL_EXPANSION_FACTOR;
1078 		else
1079 			moea64_bpvo_pool_size = BPVO_POOL_SIZE;
1080 	}
1081 
1082 	if (boothowto & RB_VERBOSE) {
1083 		printf("mmu_oea64: bpvo pool entries = %d, bpvo pool size = %zu MB\n",
1084 		    moea64_bpvo_pool_size,
1085 		    moea64_bpvo_pool_size*sizeof(struct pvo_entry) / 1048576);
1086 	}
1087 
1088 	moea64_bpvo_pool = (struct pvo_entry *)moea64_bootstrap_alloc(
1089 		moea64_bpvo_pool_size*sizeof(struct pvo_entry), PAGE_SIZE);
1090 	moea64_bpvo_pool_index = 0;
1091 
1092 	/* Place at address usable through the direct map */
1093 	if (hw_direct_map)
1094 		moea64_bpvo_pool = (struct pvo_entry *)
1095 		    PHYS_TO_DMAP((uintptr_t)moea64_bpvo_pool);
1096 
1097 	/*
1098 	 * Make sure kernel vsid is allocated as well as VSID 0.
1099 	 */
1100 	#ifndef __powerpc64__
1101 	moea64_vsid_bitmap[(KERNEL_VSIDBITS & (NVSIDS - 1)) / VSID_NBPW]
1102 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
1103 	moea64_vsid_bitmap[0] |= 1;
1104 	#endif
1105 
1106 	/*
1107 	 * Initialize the kernel pmap (which is statically allocated).
1108 	 */
1109 	#ifdef __powerpc64__
1110 	for (i = 0; i < 64; i++) {
1111 		pcpup->pc_aim.slb[i].slbv = 0;
1112 		pcpup->pc_aim.slb[i].slbe = 0;
1113 	}
1114 	#else
1115 	for (i = 0; i < 16; i++)
1116 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
1117 	#endif
1118 
1119 	kernel_pmap->pmap_phys = kernel_pmap;
1120 	CPU_FILL(&kernel_pmap->pm_active);
1121 	RB_INIT(&kernel_pmap->pmap_pvo);
1122 
1123 	PMAP_LOCK_INIT(kernel_pmap);
1124 
1125 	/*
1126 	 * Now map in all the other buffers we allocated earlier
1127 	 */
1128 
1129 	moea64_setup_direct_map(kernelstart, kernelend);
1130 }
1131 
1132 void
1133 moea64_late_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
1134 {
1135 	ihandle_t	mmui;
1136 	phandle_t	chosen;
1137 	phandle_t	mmu;
1138 	ssize_t		sz;
1139 	int		i;
1140 	vm_offset_t	pa, va;
1141 	void		*dpcpu;
1142 
1143 	/*
1144 	 * Set up the Open Firmware pmap and add its mappings if not in real
1145 	 * mode.
1146 	 */
1147 
1148 	chosen = OF_finddevice("/chosen");
1149 	if (chosen != -1 && OF_getencprop(chosen, "mmu", &mmui, 4) != -1) {
1150 		mmu = OF_instance_to_package(mmui);
1151 		if (mmu == -1 ||
1152 		    (sz = OF_getproplen(mmu, "translations")) == -1)
1153 			sz = 0;
1154 		if (sz > 6144 /* tmpstksz - 2 KB headroom */)
1155 			panic("moea64_bootstrap: too many ofw translations");
1156 
1157 		if (sz > 0)
1158 			moea64_add_ofw_mappings(mmu, sz);
1159 	}
1160 
1161 	/*
1162 	 * Calculate the last available physical address.
1163 	 */
1164 	Maxmem = 0;
1165 	for (i = 0; phys_avail[i + 1] != 0; i += 2)
1166 		Maxmem = MAX(Maxmem, powerpc_btop(phys_avail[i + 1]));
1167 
1168 	/*
1169 	 * Initialize MMU.
1170 	 */
1171 	pmap_cpu_bootstrap(0);
1172 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
1173 	pmap_bootstrapped++;
1174 
1175 	/*
1176 	 * Set the start and end of kva.
1177 	 */
1178 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
1179 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
1180 
1181 	/*
1182 	 * Map the entire KVA range into the SLB. We must not fault there.
1183 	 */
1184 	#ifdef __powerpc64__
1185 	for (va = virtual_avail; va < virtual_end; va += SEGMENT_LENGTH)
1186 		moea64_bootstrap_slb_prefault(va, 0);
1187 	#endif
1188 
1189 	/*
1190 	 * Remap any early IO mappings (console framebuffer, etc.)
1191 	 */
1192 	bs_remap_earlyboot();
1193 
1194 	/*
1195 	 * Figure out how far we can extend virtual_end into segment 16
1196 	 * without running into existing mappings. Segment 16 is guaranteed
1197 	 * to contain neither RAM nor devices (at least on Apple hardware),
1198 	 * but will generally contain some OFW mappings we should not
1199 	 * step on.
1200 	 */
1201 
1202 	#ifndef __powerpc64__	/* KVA is in high memory on PPC64 */
1203 	PMAP_LOCK(kernel_pmap);
1204 	while (virtual_end < VM_MAX_KERNEL_ADDRESS &&
1205 	    moea64_pvo_find_va(kernel_pmap, virtual_end+1) == NULL)
1206 		virtual_end += PAGE_SIZE;
1207 	PMAP_UNLOCK(kernel_pmap);
1208 	#endif
1209 
1210 	/*
1211 	 * Allocate a kernel stack with a guard page for thread0 and map it
1212 	 * into the kernel page map.
1213 	 */
1214 	pa = moea64_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
1215 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
1216 	virtual_avail = va + kstack_pages * PAGE_SIZE;
1217 	CTR2(KTR_PMAP, "moea64_bootstrap: kstack0 at %#x (%#x)", pa, va);
1218 	thread0.td_kstack = va;
1219 	thread0.td_kstack_pages = kstack_pages;
1220 	for (i = 0; i < kstack_pages; i++) {
1221 		moea64_kenter(va, pa);
1222 		pa += PAGE_SIZE;
1223 		va += PAGE_SIZE;
1224 	}
1225 
1226 	/*
1227 	 * Allocate virtual address space for the message buffer.
1228 	 */
1229 	pa = msgbuf_phys = moea64_bootstrap_alloc(msgbufsize, PAGE_SIZE);
1230 	msgbufp = (struct msgbuf *)virtual_avail;
1231 	va = virtual_avail;
1232 	virtual_avail += round_page(msgbufsize);
1233 	while (va < virtual_avail) {
1234 		moea64_kenter(va, pa);
1235 		pa += PAGE_SIZE;
1236 		va += PAGE_SIZE;
1237 	}
1238 
1239 	/*
1240 	 * Allocate virtual address space for the dynamic percpu area.
1241 	 */
1242 	pa = moea64_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
1243 	dpcpu = (void *)virtual_avail;
1244 	va = virtual_avail;
1245 	virtual_avail += DPCPU_SIZE;
1246 	while (va < virtual_avail) {
1247 		moea64_kenter(va, pa);
1248 		pa += PAGE_SIZE;
1249 		va += PAGE_SIZE;
1250 	}
1251 	dpcpu_init(dpcpu, curcpu);
1252 
1253 	crashdumpmap = (caddr_t)virtual_avail;
1254 	virtual_avail += MAXDUMPPGS * PAGE_SIZE;
1255 
1256 	/*
1257 	 * Allocate some things for page zeroing. We put this directly
1258 	 * in the page table and use MOEA64_PTE_REPLACE to avoid any
1259 	 * of the PVO book-keeping or other parts of the VM system
1260 	 * from even knowing that this hack exists.
1261 	 */
1262 
1263 	if (!hw_direct_map) {
1264 		mtx_init(&moea64_scratchpage_mtx, "pvo zero page", NULL,
1265 		    MTX_DEF);
1266 		for (i = 0; i < 2; i++) {
1267 			moea64_scratchpage_va[i] = (virtual_end+1) - PAGE_SIZE;
1268 			virtual_end -= PAGE_SIZE;
1269 
1270 			moea64_kenter(moea64_scratchpage_va[i], 0);
1271 
1272 			PMAP_LOCK(kernel_pmap);
1273 			moea64_scratchpage_pvo[i] = moea64_pvo_find_va(
1274 			    kernel_pmap, (vm_offset_t)moea64_scratchpage_va[i]);
1275 			PMAP_UNLOCK(kernel_pmap);
1276 		}
1277 	}
1278 
1279 	numa_mem_regions(&numa_pregions, &numapregions_sz);
1280 }
1281 
1282 static void
1283 moea64_pmap_init_qpages(void)
1284 {
1285 	struct pcpu *pc;
1286 	int i;
1287 
1288 	if (hw_direct_map)
1289 		return;
1290 
1291 	CPU_FOREACH(i) {
1292 		pc = pcpu_find(i);
1293 		pc->pc_qmap_addr = kva_alloc(PAGE_SIZE);
1294 		if (pc->pc_qmap_addr == 0)
1295 			panic("pmap_init_qpages: unable to allocate KVA");
1296 		PMAP_LOCK(kernel_pmap);
1297 		pc->pc_aim.qmap_pvo =
1298 		    moea64_pvo_find_va(kernel_pmap, pc->pc_qmap_addr);
1299 		PMAP_UNLOCK(kernel_pmap);
1300 		mtx_init(&pc->pc_aim.qmap_lock, "qmap lock", NULL, MTX_DEF);
1301 	}
1302 }
1303 
1304 SYSINIT(qpages_init, SI_SUB_CPU, SI_ORDER_ANY, moea64_pmap_init_qpages, NULL);
1305 
1306 /*
1307  * Activate a user pmap.  This mostly involves setting some non-CPU
1308  * state.
1309  */
1310 void
1311 moea64_activate(struct thread *td)
1312 {
1313 	pmap_t	pm;
1314 
1315 	pm = &td->td_proc->p_vmspace->vm_pmap;
1316 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1317 
1318 	#ifdef __powerpc64__
1319 	PCPU_SET(aim.userslb, pm->pm_slb);
1320 	__asm __volatile("slbmte %0, %1; isync" ::
1321 	    "r"(td->td_pcb->pcb_cpu.aim.usr_vsid), "r"(USER_SLB_SLBE));
1322 	#else
1323 	PCPU_SET(curpmap, pm->pmap_phys);
1324 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1325 	#endif
1326 }
1327 
1328 void
1329 moea64_deactivate(struct thread *td)
1330 {
1331 	pmap_t	pm;
1332 
1333 	__asm __volatile("isync; slbie %0" :: "r"(USER_ADDR));
1334 
1335 	pm = &td->td_proc->p_vmspace->vm_pmap;
1336 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1337 	#ifdef __powerpc64__
1338 	PCPU_SET(aim.userslb, NULL);
1339 	#else
1340 	PCPU_SET(curpmap, NULL);
1341 	#endif
1342 }
1343 
1344 void
1345 moea64_unwire(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1346 {
1347 	struct	pvo_entry key, *pvo;
1348 	vm_page_t m;
1349 	int64_t	refchg;
1350 
1351 	key.pvo_vaddr = sva;
1352 	PMAP_LOCK(pm);
1353 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1354 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1355 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1356 		if (PVO_IS_SP(pvo)) {
1357 			if (moea64_sp_pvo_in_range(pvo, sva, eva)) {
1358 				pvo = moea64_sp_unwire(pvo);
1359 				continue;
1360 			} else {
1361 				CTR1(KTR_PMAP, "%s: demote before unwire",
1362 				    __func__);
1363 				moea64_sp_demote(pvo);
1364 			}
1365 		}
1366 
1367 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1368 			panic("moea64_unwire: pvo %p is missing PVO_WIRED",
1369 			    pvo);
1370 		pvo->pvo_vaddr &= ~PVO_WIRED;
1371 		refchg = moea64_pte_replace(pvo, 0 /* No invalidation */);
1372 		if ((pvo->pvo_vaddr & PVO_MANAGED) &&
1373 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
1374 			if (refchg < 0)
1375 				refchg = LPTE_CHG;
1376 			m = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
1377 
1378 			refchg |= atomic_readandclear_32(&m->md.mdpg_attrs);
1379 			if (refchg & LPTE_CHG)
1380 				vm_page_dirty(m);
1381 			if (refchg & LPTE_REF)
1382 				vm_page_aflag_set(m, PGA_REFERENCED);
1383 		}
1384 		pm->pm_stats.wired_count--;
1385 	}
1386 	PMAP_UNLOCK(pm);
1387 }
1388 
1389 static int
1390 moea64_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *pap)
1391 {
1392 	struct pvo_entry *pvo;
1393 	vm_paddr_t pa;
1394 	vm_page_t m;
1395 	int val;
1396 	bool managed;
1397 
1398 	PMAP_LOCK(pmap);
1399 
1400 	pvo = moea64_pvo_find_va(pmap, addr);
1401 	if (pvo != NULL) {
1402 		pa = PVO_PADDR(pvo);
1403 		m = PHYS_TO_VM_PAGE(pa);
1404 		managed = (pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED;
1405 		if (PVO_IS_SP(pvo))
1406 			val = MINCORE_INCORE | MINCORE_PSIND(1);
1407 		else
1408 			val = MINCORE_INCORE;
1409 	} else {
1410 		PMAP_UNLOCK(pmap);
1411 		return (0);
1412 	}
1413 
1414 	PMAP_UNLOCK(pmap);
1415 
1416 	if (m == NULL)
1417 		return (0);
1418 
1419 	if (managed) {
1420 		if (moea64_is_modified(m))
1421 			val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
1422 
1423 		if (moea64_is_referenced(m))
1424 			val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
1425 	}
1426 
1427 	if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) !=
1428 	    (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) &&
1429 	    managed) {
1430 		*pap = pa;
1431 	}
1432 
1433 	return (val);
1434 }
1435 
1436 /*
1437  * This goes through and sets the physical address of our
1438  * special scratch PTE to the PA we want to zero or copy. Because
1439  * of locking issues (this can get called in pvo_enter() by
1440  * the UMA allocator), we can't use most other utility functions here
1441  */
1442 
1443 static __inline
1444 void moea64_set_scratchpage_pa(int which, vm_paddr_t pa)
1445 {
1446 	struct pvo_entry *pvo;
1447 
1448 	KASSERT(!hw_direct_map, ("Using OEA64 scratchpage with a direct map!"));
1449 	mtx_assert(&moea64_scratchpage_mtx, MA_OWNED);
1450 
1451 	pvo = moea64_scratchpage_pvo[which];
1452 	PMAP_LOCK(pvo->pvo_pmap);
1453 	pvo->pvo_pte.pa =
1454 	    moea64_calc_wimg(pa, VM_MEMATTR_DEFAULT) | (uint64_t)pa;
1455 	moea64_pte_replace(pvo, MOEA64_PTE_INVALIDATE);
1456 	PMAP_UNLOCK(pvo->pvo_pmap);
1457 	isync();
1458 }
1459 
1460 void
1461 moea64_copy_page(vm_page_t msrc, vm_page_t mdst)
1462 {
1463 	mtx_lock(&moea64_scratchpage_mtx);
1464 
1465 	moea64_set_scratchpage_pa(0, VM_PAGE_TO_PHYS(msrc));
1466 	moea64_set_scratchpage_pa(1, VM_PAGE_TO_PHYS(mdst));
1467 
1468 	bcopy((void *)moea64_scratchpage_va[0],
1469 	    (void *)moea64_scratchpage_va[1], PAGE_SIZE);
1470 
1471 	mtx_unlock(&moea64_scratchpage_mtx);
1472 }
1473 
1474 void
1475 moea64_copy_page_dmap(vm_page_t msrc, vm_page_t mdst)
1476 {
1477 	vm_offset_t	dst;
1478 	vm_offset_t	src;
1479 
1480 	dst = VM_PAGE_TO_PHYS(mdst);
1481 	src = VM_PAGE_TO_PHYS(msrc);
1482 
1483 	bcopy((void *)PHYS_TO_DMAP(src), (void *)PHYS_TO_DMAP(dst),
1484 	    PAGE_SIZE);
1485 }
1486 
1487 inline void
1488 moea64_copy_pages_dmap(vm_page_t *ma, vm_offset_t a_offset,
1489     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1490 {
1491 	void *a_cp, *b_cp;
1492 	vm_offset_t a_pg_offset, b_pg_offset;
1493 	int cnt;
1494 
1495 	while (xfersize > 0) {
1496 		a_pg_offset = a_offset & PAGE_MASK;
1497 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1498 		a_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1499 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT])) +
1500 		    a_pg_offset;
1501 		b_pg_offset = b_offset & PAGE_MASK;
1502 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1503 		b_cp = (char *)(uintptr_t)PHYS_TO_DMAP(
1504 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT])) +
1505 		    b_pg_offset;
1506 		bcopy(a_cp, b_cp, cnt);
1507 		a_offset += cnt;
1508 		b_offset += cnt;
1509 		xfersize -= cnt;
1510 	}
1511 }
1512 
1513 void
1514 moea64_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
1515     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1516 {
1517 	void *a_cp, *b_cp;
1518 	vm_offset_t a_pg_offset, b_pg_offset;
1519 	int cnt;
1520 
1521 	mtx_lock(&moea64_scratchpage_mtx);
1522 	while (xfersize > 0) {
1523 		a_pg_offset = a_offset & PAGE_MASK;
1524 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1525 		moea64_set_scratchpage_pa(0,
1526 		    VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]));
1527 		a_cp = (char *)moea64_scratchpage_va[0] + a_pg_offset;
1528 		b_pg_offset = b_offset & PAGE_MASK;
1529 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1530 		moea64_set_scratchpage_pa(1,
1531 		    VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]));
1532 		b_cp = (char *)moea64_scratchpage_va[1] + b_pg_offset;
1533 		bcopy(a_cp, b_cp, cnt);
1534 		a_offset += cnt;
1535 		b_offset += cnt;
1536 		xfersize -= cnt;
1537 	}
1538 	mtx_unlock(&moea64_scratchpage_mtx);
1539 }
1540 
1541 void
1542 moea64_zero_page_area(vm_page_t m, int off, int size)
1543 {
1544 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1545 
1546 	if (size + off > PAGE_SIZE)
1547 		panic("moea64_zero_page: size + off > PAGE_SIZE");
1548 
1549 	if (hw_direct_map) {
1550 		bzero((caddr_t)(uintptr_t)PHYS_TO_DMAP(pa) + off, size);
1551 	} else {
1552 		mtx_lock(&moea64_scratchpage_mtx);
1553 		moea64_set_scratchpage_pa(0, pa);
1554 		bzero((caddr_t)moea64_scratchpage_va[0] + off, size);
1555 		mtx_unlock(&moea64_scratchpage_mtx);
1556 	}
1557 }
1558 
1559 /*
1560  * Zero a page of physical memory by temporarily mapping it
1561  */
1562 void
1563 moea64_zero_page(vm_page_t m)
1564 {
1565 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1566 	vm_offset_t va, off;
1567 
1568 	mtx_lock(&moea64_scratchpage_mtx);
1569 
1570 	moea64_set_scratchpage_pa(0, pa);
1571 	va = moea64_scratchpage_va[0];
1572 
1573 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1574 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1575 
1576 	mtx_unlock(&moea64_scratchpage_mtx);
1577 }
1578 
1579 void
1580 moea64_zero_page_dmap(vm_page_t m)
1581 {
1582 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1583 	vm_offset_t va, off;
1584 
1585 	va = PHYS_TO_DMAP(pa);
1586 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1587 		__asm __volatile("dcbz 0,%0" :: "r"(va + off));
1588 }
1589 
1590 vm_offset_t
1591 moea64_quick_enter_page(vm_page_t m)
1592 {
1593 	struct pvo_entry *pvo;
1594 	vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1595 
1596 	/*
1597  	 * MOEA64_PTE_REPLACE does some locking, so we can't just grab
1598 	 * a critical section and access the PCPU data like on i386.
1599 	 * Instead, pin the thread and grab the PCPU lock to prevent
1600 	 * a preempting thread from using the same PCPU data.
1601 	 */
1602 	sched_pin();
1603 
1604 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_NOTOWNED);
1605 	pvo = PCPU_GET(aim.qmap_pvo);
1606 
1607 	mtx_lock(PCPU_PTR(aim.qmap_lock));
1608 	pvo->pvo_pte.pa = moea64_calc_wimg(pa, pmap_page_get_memattr(m)) |
1609 	    (uint64_t)pa;
1610 	moea64_pte_replace(pvo, MOEA64_PTE_INVALIDATE);
1611 	isync();
1612 
1613 	return (PCPU_GET(qmap_addr));
1614 }
1615 
1616 vm_offset_t
1617 moea64_quick_enter_page_dmap(vm_page_t m)
1618 {
1619 
1620 	return (PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)));
1621 }
1622 
1623 void
1624 moea64_quick_remove_page(vm_offset_t addr)
1625 {
1626 
1627 	mtx_assert(PCPU_PTR(aim.qmap_lock), MA_OWNED);
1628 	KASSERT(PCPU_GET(qmap_addr) == addr,
1629 	    ("moea64_quick_remove_page: invalid address"));
1630 	mtx_unlock(PCPU_PTR(aim.qmap_lock));
1631 	sched_unpin();
1632 }
1633 
1634 boolean_t
1635 moea64_page_is_mapped(vm_page_t m)
1636 {
1637 	return (!LIST_EMPTY(&(m)->md.mdpg_pvoh));
1638 }
1639 
1640 /*
1641  * Map the given physical page at the specified virtual address in the
1642  * target pmap with the protection requested.  If specified the page
1643  * will be wired down.
1644  */
1645 
1646 int
1647 moea64_enter(pmap_t pmap, vm_offset_t va, vm_page_t m,
1648     vm_prot_t prot, u_int flags, int8_t psind)
1649 {
1650 	struct		pvo_entry *pvo, *oldpvo, *tpvo;
1651 	struct		pvo_head *pvo_head;
1652 	uint64_t	pte_lo;
1653 	int		error;
1654 	vm_paddr_t	pa;
1655 
1656 	if ((m->oflags & VPO_UNMANAGED) == 0) {
1657 		if ((flags & PMAP_ENTER_QUICK_LOCKED) == 0)
1658 			VM_PAGE_OBJECT_BUSY_ASSERT(m);
1659 		else
1660 			VM_OBJECT_ASSERT_LOCKED(m->object);
1661 	}
1662 
1663 	if (psind > 0)
1664 		return (moea64_sp_enter(pmap, va, m, prot, flags, psind));
1665 
1666 	pvo = alloc_pvo_entry(0);
1667 	if (pvo == NULL)
1668 		return (KERN_RESOURCE_SHORTAGE);
1669 	pvo->pvo_pmap = NULL; /* to be filled in later */
1670 	pvo->pvo_pte.prot = prot;
1671 
1672 	pa = VM_PAGE_TO_PHYS(m);
1673 	pte_lo = moea64_calc_wimg(pa, pmap_page_get_memattr(m));
1674 	pvo->pvo_pte.pa = pa | pte_lo;
1675 
1676 	if ((flags & PMAP_ENTER_WIRED) != 0)
1677 		pvo->pvo_vaddr |= PVO_WIRED;
1678 
1679 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea64_initialized) {
1680 		pvo_head = NULL;
1681 	} else {
1682 		pvo_head = &m->md.mdpg_pvoh;
1683 		pvo->pvo_vaddr |= PVO_MANAGED;
1684 	}
1685 
1686 	PV_LOCK(pa);
1687 	PMAP_LOCK(pmap);
1688 	if (pvo->pvo_pmap == NULL)
1689 		init_pvo_entry(pvo, pmap, va);
1690 
1691 	if (moea64_ps_enabled(pmap) &&
1692 	    (tpvo = moea64_pvo_find_va(pmap, va & ~HPT_SP_MASK)) != NULL &&
1693 	    PVO_IS_SP(tpvo)) {
1694 		/* Demote SP before entering a regular page */
1695 		CTR2(KTR_PMAP, "%s: demote before enter: va=%#jx",
1696 		    __func__, (uintmax_t)va);
1697 		moea64_sp_demote_aligned(tpvo);
1698 	}
1699 
1700 	if (prot & VM_PROT_WRITE)
1701 		if (pmap_bootstrapped &&
1702 		    (m->oflags & VPO_UNMANAGED) == 0)
1703 			vm_page_aflag_set(m, PGA_WRITEABLE);
1704 
1705 	error = moea64_pvo_enter(pvo, pvo_head, &oldpvo);
1706 	if (error == EEXIST) {
1707 		if (oldpvo->pvo_vaddr == pvo->pvo_vaddr &&
1708 		    oldpvo->pvo_pte.pa == pvo->pvo_pte.pa &&
1709 		    oldpvo->pvo_pte.prot == prot) {
1710 			/* Identical mapping already exists */
1711 			error = 0;
1712 
1713 			/* If not in page table, reinsert it */
1714 			if (moea64_pte_synch(oldpvo) < 0) {
1715 				STAT_MOEA64(moea64_pte_overflow--);
1716 				moea64_pte_insert(oldpvo);
1717 			}
1718 
1719 			/* Then just clean up and go home */
1720 			PMAP_UNLOCK(pmap);
1721 			PV_UNLOCK(pa);
1722 			free_pvo_entry(pvo);
1723 			pvo = NULL;
1724 			goto out;
1725 		} else {
1726 			/* Otherwise, need to kill it first */
1727 			KASSERT(oldpvo->pvo_pmap == pmap, ("pmap of old "
1728 			    "mapping does not match new mapping"));
1729 			moea64_pvo_remove_from_pmap(oldpvo);
1730 			moea64_pvo_enter(pvo, pvo_head, NULL);
1731 		}
1732 	}
1733 	PMAP_UNLOCK(pmap);
1734 	PV_UNLOCK(pa);
1735 
1736 	/* Free any dead pages */
1737 	if (error == EEXIST) {
1738 		moea64_pvo_remove_from_page(oldpvo);
1739 		free_pvo_entry(oldpvo);
1740 	}
1741 
1742 out:
1743 	/*
1744 	 * Flush the page from the instruction cache if this page is
1745 	 * mapped executable and cacheable.
1746 	 */
1747 	if (pmap != kernel_pmap && (m->a.flags & PGA_EXECUTABLE) == 0 &&
1748 	    (pte_lo & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
1749 		vm_page_aflag_set(m, PGA_EXECUTABLE);
1750 		moea64_syncicache(pmap, va, pa, PAGE_SIZE);
1751 	}
1752 
1753 #if VM_NRESERVLEVEL > 0
1754 	/*
1755 	 * Try to promote pages.
1756 	 *
1757 	 * If the VA of the entered page is not aligned with its PA,
1758 	 * don't try page promotion as it is not possible.
1759 	 * This reduces the number of promotion failures dramatically.
1760 	 */
1761 	if (moea64_ps_enabled(pmap) && pmap != kernel_pmap && pvo != NULL &&
1762 	    (pvo->pvo_vaddr & PVO_MANAGED) != 0 &&
1763 	    (va & HPT_SP_MASK) == (pa & HPT_SP_MASK) &&
1764 	    (m->flags & PG_FICTITIOUS) == 0 &&
1765 	    vm_reserv_level_iffullpop(m) == 0)
1766 		moea64_sp_promote(pmap, va, m);
1767 #endif
1768 
1769 	return (KERN_SUCCESS);
1770 }
1771 
1772 static void
1773 moea64_syncicache(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
1774     vm_size_t sz)
1775 {
1776 
1777 	/*
1778 	 * This is much trickier than on older systems because
1779 	 * we can't sync the icache on physical addresses directly
1780 	 * without a direct map. Instead we check a couple of cases
1781 	 * where the memory is already mapped in and, failing that,
1782 	 * use the same trick we use for page zeroing to create
1783 	 * a temporary mapping for this physical address.
1784 	 */
1785 
1786 	if (!pmap_bootstrapped) {
1787 		/*
1788 		 * If PMAP is not bootstrapped, we are likely to be
1789 		 * in real mode.
1790 		 */
1791 		__syncicache((void *)(uintptr_t)pa, sz);
1792 	} else if (pmap == kernel_pmap) {
1793 		__syncicache((void *)va, sz);
1794 	} else if (hw_direct_map) {
1795 		__syncicache((void *)(uintptr_t)PHYS_TO_DMAP(pa), sz);
1796 	} else {
1797 		/* Use the scratch page to set up a temp mapping */
1798 
1799 		mtx_lock(&moea64_scratchpage_mtx);
1800 
1801 		moea64_set_scratchpage_pa(1, pa & ~ADDR_POFF);
1802 		__syncicache((void *)(moea64_scratchpage_va[1] +
1803 		    (va & ADDR_POFF)), sz);
1804 
1805 		mtx_unlock(&moea64_scratchpage_mtx);
1806 	}
1807 }
1808 
1809 /*
1810  * Maps a sequence of resident pages belonging to the same object.
1811  * The sequence begins with the given page m_start.  This page is
1812  * mapped at the given virtual address start.  Each subsequent page is
1813  * mapped at a virtual address that is offset from start by the same
1814  * amount as the page is offset from m_start within the object.  The
1815  * last page in the sequence is the page with the largest offset from
1816  * m_start that can be mapped at a virtual address less than the given
1817  * virtual address end.  Not every virtual page between start and end
1818  * is mapped; only those for which a resident page exists with the
1819  * corresponding offset from m_start are mapped.
1820  */
1821 void
1822 moea64_enter_object(pmap_t pm, vm_offset_t start, vm_offset_t end,
1823     vm_page_t m_start, vm_prot_t prot)
1824 {
1825 	vm_page_t m;
1826 	vm_pindex_t diff, psize;
1827 	vm_offset_t va;
1828 	int8_t psind;
1829 
1830 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1831 
1832 	psize = atop(end - start);
1833 	m = m_start;
1834 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1835 		va = start + ptoa(diff);
1836 		if ((va & HPT_SP_MASK) == 0 && va + HPT_SP_SIZE <= end &&
1837 		    m->psind == 1 && moea64_ps_enabled(pm))
1838 			psind = 1;
1839 		else
1840 			psind = 0;
1841 		moea64_enter(pm, va, m, prot &
1842 		    (VM_PROT_READ | VM_PROT_EXECUTE),
1843 		    PMAP_ENTER_NOSLEEP | PMAP_ENTER_QUICK_LOCKED, psind);
1844 		if (psind == 1)
1845 			m = &m[HPT_SP_SIZE / PAGE_SIZE - 1];
1846 		m = TAILQ_NEXT(m, listq);
1847 	}
1848 }
1849 
1850 void
1851 moea64_enter_quick(pmap_t pm, vm_offset_t va, vm_page_t m,
1852     vm_prot_t prot)
1853 {
1854 
1855 	moea64_enter(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1856 	    PMAP_ENTER_NOSLEEP | PMAP_ENTER_QUICK_LOCKED, 0);
1857 }
1858 
1859 vm_paddr_t
1860 moea64_extract(pmap_t pm, vm_offset_t va)
1861 {
1862 	struct	pvo_entry *pvo;
1863 	vm_paddr_t pa;
1864 
1865 	PMAP_LOCK(pm);
1866 	pvo = moea64_pvo_find_va(pm, va);
1867 	if (pvo == NULL)
1868 		pa = 0;
1869 	else
1870 		pa = PVO_PADDR(pvo) | (va - PVO_VADDR(pvo));
1871 	PMAP_UNLOCK(pm);
1872 
1873 	return (pa);
1874 }
1875 
1876 /*
1877  * Atomically extract and hold the physical page with the given
1878  * pmap and virtual address pair if that mapping permits the given
1879  * protection.
1880  */
1881 vm_page_t
1882 moea64_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1883 {
1884 	struct	pvo_entry *pvo;
1885 	vm_page_t m;
1886 
1887 	m = NULL;
1888 	PMAP_LOCK(pmap);
1889 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
1890 	if (pvo != NULL && (pvo->pvo_pte.prot & prot) == prot) {
1891 		m = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
1892 		if (!vm_page_wire_mapped(m))
1893 			m = NULL;
1894 	}
1895 	PMAP_UNLOCK(pmap);
1896 	return (m);
1897 }
1898 
1899 static void *
1900 moea64_uma_page_alloc(uma_zone_t zone, vm_size_t bytes, int domain,
1901     uint8_t *flags, int wait)
1902 {
1903 	struct pvo_entry *pvo;
1904         vm_offset_t va;
1905         vm_page_t m;
1906         int needed_lock;
1907 
1908 	/*
1909 	 * This entire routine is a horrible hack to avoid bothering kmem
1910 	 * for new KVA addresses. Because this can get called from inside
1911 	 * kmem allocation routines, calling kmem for a new address here
1912 	 * can lead to multiply locking non-recursive mutexes.
1913 	 */
1914 
1915 	*flags = UMA_SLAB_PRIV;
1916 	needed_lock = !PMAP_LOCKED(kernel_pmap);
1917 
1918 	m = vm_page_alloc_noobj_domain(domain, malloc2vm_flags(wait) |
1919 	    VM_ALLOC_WIRED);
1920 	if (m == NULL)
1921 		return (NULL);
1922 
1923 	va = VM_PAGE_TO_PHYS(m);
1924 
1925 	pvo = alloc_pvo_entry(1 /* bootstrap */);
1926 
1927 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE;
1928 	pvo->pvo_pte.pa = VM_PAGE_TO_PHYS(m) | LPTE_M;
1929 
1930 	if (needed_lock)
1931 		PMAP_LOCK(kernel_pmap);
1932 
1933 	init_pvo_entry(pvo, kernel_pmap, va);
1934 	pvo->pvo_vaddr |= PVO_WIRED;
1935 
1936 	moea64_pvo_enter(pvo, NULL, NULL);
1937 
1938 	if (needed_lock)
1939 		PMAP_UNLOCK(kernel_pmap);
1940 
1941 	return (void *)va;
1942 }
1943 
1944 extern int elf32_nxstack;
1945 
1946 void
1947 moea64_init()
1948 {
1949 
1950 	CTR0(KTR_PMAP, "moea64_init");
1951 
1952 	moea64_pvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1953 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1954 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1955 
1956 	/*
1957 	 * Are large page mappings enabled?
1958 	 *
1959 	 * While HPT superpages are not better tested, leave it disabled by
1960 	 * default.
1961 	 */
1962 	superpages_enabled = 0;
1963 	TUNABLE_INT_FETCH("vm.pmap.superpages_enabled", &superpages_enabled);
1964 	if (superpages_enabled) {
1965 		KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0,
1966 		    ("moea64_init: can't assign to pagesizes[1]"));
1967 
1968 		if (moea64_large_page_size == 0) {
1969 			printf("mmu_oea64: HW does not support large pages. "
1970 					"Disabling superpages...\n");
1971 			superpages_enabled = 0;
1972 		} else if (!moea64_has_lp_4k_16m) {
1973 			printf("mmu_oea64: "
1974 			    "HW does not support mixed 4KB/16MB page sizes. "
1975 			    "Disabling superpages...\n");
1976 			superpages_enabled = 0;
1977 		} else
1978 			pagesizes[1] = HPT_SP_SIZE;
1979 	}
1980 
1981 	if (!hw_direct_map) {
1982 		uma_zone_set_allocf(moea64_pvo_zone, moea64_uma_page_alloc);
1983 	}
1984 
1985 #ifdef COMPAT_FREEBSD32
1986 	elf32_nxstack = 1;
1987 #endif
1988 
1989 	moea64_initialized = TRUE;
1990 }
1991 
1992 boolean_t
1993 moea64_is_referenced(vm_page_t m)
1994 {
1995 
1996 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1997 	    ("moea64_is_referenced: page %p is not managed", m));
1998 
1999 	return (moea64_query_bit(m, LPTE_REF));
2000 }
2001 
2002 boolean_t
2003 moea64_is_modified(vm_page_t m)
2004 {
2005 
2006 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2007 	    ("moea64_is_modified: page %p is not managed", m));
2008 
2009 	/*
2010 	 * If the page is not busied then this check is racy.
2011 	 */
2012 	if (!pmap_page_is_write_mapped(m))
2013 		return (FALSE);
2014 
2015 	return (moea64_query_bit(m, LPTE_CHG));
2016 }
2017 
2018 boolean_t
2019 moea64_is_prefaultable(pmap_t pmap, vm_offset_t va)
2020 {
2021 	struct pvo_entry *pvo;
2022 	boolean_t rv = TRUE;
2023 
2024 	PMAP_LOCK(pmap);
2025 	pvo = moea64_pvo_find_va(pmap, va & ~ADDR_POFF);
2026 	if (pvo != NULL)
2027 		rv = FALSE;
2028 	PMAP_UNLOCK(pmap);
2029 	return (rv);
2030 }
2031 
2032 void
2033 moea64_clear_modify(vm_page_t m)
2034 {
2035 
2036 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2037 	    ("moea64_clear_modify: page %p is not managed", m));
2038 	vm_page_assert_busied(m);
2039 
2040 	if (!pmap_page_is_write_mapped(m))
2041 		return;
2042 	moea64_clear_bit(m, LPTE_CHG);
2043 }
2044 
2045 /*
2046  * Clear the write and modified bits in each of the given page's mappings.
2047  */
2048 void
2049 moea64_remove_write(vm_page_t m)
2050 {
2051 	struct	pvo_entry *pvo;
2052 	int64_t	refchg, ret;
2053 	pmap_t	pmap;
2054 
2055 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2056 	    ("moea64_remove_write: page %p is not managed", m));
2057 	vm_page_assert_busied(m);
2058 
2059 	if (!pmap_page_is_write_mapped(m))
2060 		return;
2061 
2062 	powerpc_sync();
2063 	PV_PAGE_LOCK(m);
2064 	refchg = 0;
2065 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2066 		pmap = pvo->pvo_pmap;
2067 		PMAP_LOCK(pmap);
2068 		if (!(pvo->pvo_vaddr & PVO_DEAD) &&
2069 		    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2070 			if (PVO_IS_SP(pvo)) {
2071 				CTR1(KTR_PMAP, "%s: demote before remwr",
2072 				    __func__);
2073 				moea64_sp_demote(pvo);
2074 			}
2075 			pvo->pvo_pte.prot &= ~VM_PROT_WRITE;
2076 			ret = moea64_pte_replace(pvo, MOEA64_PTE_PROT_UPDATE);
2077 			if (ret < 0)
2078 				ret = LPTE_CHG;
2079 			refchg |= ret;
2080 			if (pvo->pvo_pmap == kernel_pmap)
2081 				isync();
2082 		}
2083 		PMAP_UNLOCK(pmap);
2084 	}
2085 	if ((refchg | atomic_readandclear_32(&m->md.mdpg_attrs)) & LPTE_CHG)
2086 		vm_page_dirty(m);
2087 	vm_page_aflag_clear(m, PGA_WRITEABLE);
2088 	PV_PAGE_UNLOCK(m);
2089 }
2090 
2091 /*
2092  *	moea64_ts_referenced:
2093  *
2094  *	Return a count of reference bits for a page, clearing those bits.
2095  *	It is not necessary for every reference bit to be cleared, but it
2096  *	is necessary that 0 only be returned when there are truly no
2097  *	reference bits set.
2098  *
2099  *	XXX: The exact number of bits to check and clear is a matter that
2100  *	should be tested and standardized at some point in the future for
2101  *	optimal aging of shared pages.
2102  */
2103 int
2104 moea64_ts_referenced(vm_page_t m)
2105 {
2106 
2107 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2108 	    ("moea64_ts_referenced: page %p is not managed", m));
2109 	return (moea64_clear_bit(m, LPTE_REF));
2110 }
2111 
2112 /*
2113  * Modify the WIMG settings of all mappings for a page.
2114  */
2115 void
2116 moea64_page_set_memattr(vm_page_t m, vm_memattr_t ma)
2117 {
2118 	struct	pvo_entry *pvo;
2119 	int64_t	refchg;
2120 	pmap_t	pmap;
2121 	uint64_t lo;
2122 
2123 	CTR3(KTR_PMAP, "%s: pa=%#jx, ma=%#x",
2124 	    __func__, (uintmax_t)VM_PAGE_TO_PHYS(m), ma);
2125 
2126 	if ((m->oflags & VPO_UNMANAGED) != 0) {
2127 		m->md.mdpg_cache_attrs = ma;
2128 		return;
2129 	}
2130 
2131 	lo = moea64_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
2132 
2133 	PV_PAGE_LOCK(m);
2134 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2135 		pmap = pvo->pvo_pmap;
2136 		PMAP_LOCK(pmap);
2137 		if (!(pvo->pvo_vaddr & PVO_DEAD)) {
2138 			if (PVO_IS_SP(pvo)) {
2139 				CTR1(KTR_PMAP,
2140 				    "%s: demote before set_memattr", __func__);
2141 				moea64_sp_demote(pvo);
2142 			}
2143 			pvo->pvo_pte.pa &= ~LPTE_WIMG;
2144 			pvo->pvo_pte.pa |= lo;
2145 			refchg = moea64_pte_replace(pvo, MOEA64_PTE_INVALIDATE);
2146 			if (refchg < 0)
2147 				refchg = (pvo->pvo_pte.prot & VM_PROT_WRITE) ?
2148 				    LPTE_CHG : 0;
2149 			if ((pvo->pvo_vaddr & PVO_MANAGED) &&
2150 			    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2151 				refchg |=
2152 				    atomic_readandclear_32(&m->md.mdpg_attrs);
2153 				if (refchg & LPTE_CHG)
2154 					vm_page_dirty(m);
2155 				if (refchg & LPTE_REF)
2156 					vm_page_aflag_set(m, PGA_REFERENCED);
2157 			}
2158 			if (pvo->pvo_pmap == kernel_pmap)
2159 				isync();
2160 		}
2161 		PMAP_UNLOCK(pmap);
2162 	}
2163 	m->md.mdpg_cache_attrs = ma;
2164 	PV_PAGE_UNLOCK(m);
2165 }
2166 
2167 /*
2168  * Map a wired page into kernel virtual address space.
2169  */
2170 void
2171 moea64_kenter_attr(vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
2172 {
2173 	int		error;
2174 	struct pvo_entry *pvo, *oldpvo;
2175 
2176 	do {
2177 		pvo = alloc_pvo_entry(0);
2178 		if (pvo == NULL)
2179 			vm_wait(NULL);
2180 	} while (pvo == NULL);
2181 	pvo->pvo_pte.prot = VM_PROT_READ | VM_PROT_WRITE | VM_PROT_EXECUTE;
2182 	pvo->pvo_pte.pa = (pa & ~ADDR_POFF) | moea64_calc_wimg(pa, ma);
2183 	pvo->pvo_vaddr |= PVO_WIRED;
2184 
2185 	PMAP_LOCK(kernel_pmap);
2186 	oldpvo = moea64_pvo_find_va(kernel_pmap, va);
2187 	if (oldpvo != NULL)
2188 		moea64_pvo_remove_from_pmap(oldpvo);
2189 	init_pvo_entry(pvo, kernel_pmap, va);
2190 	error = moea64_pvo_enter(pvo, NULL, NULL);
2191 	PMAP_UNLOCK(kernel_pmap);
2192 
2193 	/* Free any dead pages */
2194 	if (oldpvo != NULL) {
2195 		moea64_pvo_remove_from_page(oldpvo);
2196 		free_pvo_entry(oldpvo);
2197 	}
2198 
2199 	if (error != 0)
2200 		panic("moea64_kenter: failed to enter va %#zx pa %#jx: %d", va,
2201 		    (uintmax_t)pa, error);
2202 }
2203 
2204 void
2205 moea64_kenter(vm_offset_t va, vm_paddr_t pa)
2206 {
2207 
2208 	moea64_kenter_attr(va, pa, VM_MEMATTR_DEFAULT);
2209 }
2210 
2211 /*
2212  * Extract the physical page address associated with the given kernel virtual
2213  * address.
2214  */
2215 vm_paddr_t
2216 moea64_kextract(vm_offset_t va)
2217 {
2218 	struct		pvo_entry *pvo;
2219 	vm_paddr_t pa;
2220 
2221 	/*
2222 	 * Shortcut the direct-mapped case when applicable.  We never put
2223 	 * anything but 1:1 (or 62-bit aliased) mappings below
2224 	 * VM_MIN_KERNEL_ADDRESS.
2225 	 */
2226 	if (va < VM_MIN_KERNEL_ADDRESS)
2227 		return (va & ~DMAP_BASE_ADDRESS);
2228 
2229 	PMAP_LOCK(kernel_pmap);
2230 	pvo = moea64_pvo_find_va(kernel_pmap, va);
2231 	KASSERT(pvo != NULL, ("moea64_kextract: no addr found for %#" PRIxPTR,
2232 	    va));
2233 	pa = PVO_PADDR(pvo) | (va - PVO_VADDR(pvo));
2234 	PMAP_UNLOCK(kernel_pmap);
2235 	return (pa);
2236 }
2237 
2238 /*
2239  * Remove a wired page from kernel virtual address space.
2240  */
2241 void
2242 moea64_kremove(vm_offset_t va)
2243 {
2244 	moea64_remove(kernel_pmap, va, va + PAGE_SIZE);
2245 }
2246 
2247 /*
2248  * Provide a kernel pointer corresponding to a given userland pointer.
2249  * The returned pointer is valid until the next time this function is
2250  * called in this thread. This is used internally in copyin/copyout.
2251  */
2252 static int
2253 moea64_map_user_ptr(pmap_t pm, volatile const void *uaddr,
2254     void **kaddr, size_t ulen, size_t *klen)
2255 {
2256 	size_t l;
2257 #ifdef __powerpc64__
2258 	struct slb *slb;
2259 #endif
2260 	register_t slbv;
2261 
2262 	*kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
2263 	l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
2264 	if (l > ulen)
2265 		l = ulen;
2266 	if (klen)
2267 		*klen = l;
2268 	else if (l != ulen)
2269 		return (EFAULT);
2270 
2271 #ifdef __powerpc64__
2272 	/* Try lockless look-up first */
2273 	slb = user_va_to_slb_entry(pm, (vm_offset_t)uaddr);
2274 
2275 	if (slb == NULL) {
2276 		/* If it isn't there, we need to pre-fault the VSID */
2277 		PMAP_LOCK(pm);
2278 		slbv = va_to_vsid(pm, (vm_offset_t)uaddr) << SLBV_VSID_SHIFT;
2279 		PMAP_UNLOCK(pm);
2280 	} else {
2281 		slbv = slb->slbv;
2282 	}
2283 
2284 	/* Mark segment no-execute */
2285 	slbv |= SLBV_N;
2286 #else
2287 	slbv = va_to_vsid(pm, (vm_offset_t)uaddr);
2288 
2289 	/* Mark segment no-execute */
2290 	slbv |= SR_N;
2291 #endif
2292 
2293 	/* If we have already set this VSID, we can just return */
2294 	if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == slbv)
2295 		return (0);
2296 
2297 	__asm __volatile("isync");
2298 	curthread->td_pcb->pcb_cpu.aim.usr_segm =
2299 	    (uintptr_t)uaddr >> ADDR_SR_SHFT;
2300 	curthread->td_pcb->pcb_cpu.aim.usr_vsid = slbv;
2301 #ifdef __powerpc64__
2302 	__asm __volatile ("slbie %0; slbmte %1, %2; isync" ::
2303 	    "r"(USER_ADDR), "r"(slbv), "r"(USER_SLB_SLBE));
2304 #else
2305 	__asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(slbv));
2306 #endif
2307 
2308 	return (0);
2309 }
2310 
2311 /*
2312  * Figure out where a given kernel pointer (usually in a fault) points
2313  * to from the VM's perspective, potentially remapping into userland's
2314  * address space.
2315  */
2316 static int
2317 moea64_decode_kernel_ptr(vm_offset_t addr, int *is_user,
2318     vm_offset_t *decoded_addr)
2319 {
2320 	vm_offset_t user_sr;
2321 
2322 	if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
2323 		user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
2324 		addr &= ADDR_PIDX | ADDR_POFF;
2325 		addr |= user_sr << ADDR_SR_SHFT;
2326 		*decoded_addr = addr;
2327 		*is_user = 1;
2328 	} else {
2329 		*decoded_addr = addr;
2330 		*is_user = 0;
2331 	}
2332 
2333 	return (0);
2334 }
2335 
2336 /*
2337  * Map a range of physical addresses into kernel virtual address space.
2338  *
2339  * The value passed in *virt is a suggested virtual address for the mapping.
2340  * Architectures which can support a direct-mapped physical to virtual region
2341  * can return the appropriate address within that region, leaving '*virt'
2342  * unchanged.  Other architectures should map the pages starting at '*virt' and
2343  * update '*virt' with the first usable address after the mapped region.
2344  */
2345 vm_offset_t
2346 moea64_map(vm_offset_t *virt, vm_paddr_t pa_start,
2347     vm_paddr_t pa_end, int prot)
2348 {
2349 	vm_offset_t	sva, va;
2350 
2351 	if (hw_direct_map) {
2352 		/*
2353 		 * Check if every page in the region is covered by the direct
2354 		 * map. The direct map covers all of physical memory. Use
2355 		 * moea64_calc_wimg() as a shortcut to see if the page is in
2356 		 * physical memory as a way to see if the direct map covers it.
2357 		 */
2358 		for (va = pa_start; va < pa_end; va += PAGE_SIZE)
2359 			if (moea64_calc_wimg(va, VM_MEMATTR_DEFAULT) != LPTE_M)
2360 				break;
2361 		if (va == pa_end)
2362 			return (PHYS_TO_DMAP(pa_start));
2363 	}
2364 	sva = *virt;
2365 	va = sva;
2366 	/* XXX respect prot argument */
2367 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
2368 		moea64_kenter(va, pa_start);
2369 	*virt = va;
2370 
2371 	return (sva);
2372 }
2373 
2374 /*
2375  * Returns true if the pmap's pv is one of the first
2376  * 16 pvs linked to from this page.  This count may
2377  * be changed upwards or downwards in the future; it
2378  * is only necessary that true be returned for a small
2379  * subset of pmaps for proper page aging.
2380  */
2381 boolean_t
2382 moea64_page_exists_quick(pmap_t pmap, vm_page_t m)
2383 {
2384         int loops;
2385 	struct pvo_entry *pvo;
2386 	boolean_t rv;
2387 
2388 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
2389 	    ("moea64_page_exists_quick: page %p is not managed", m));
2390 	loops = 0;
2391 	rv = FALSE;
2392 	PV_PAGE_LOCK(m);
2393 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2394 		if (!(pvo->pvo_vaddr & PVO_DEAD) && pvo->pvo_pmap == pmap) {
2395 			rv = TRUE;
2396 			break;
2397 		}
2398 		if (++loops >= 16)
2399 			break;
2400 	}
2401 	PV_PAGE_UNLOCK(m);
2402 	return (rv);
2403 }
2404 
2405 void
2406 moea64_page_init(vm_page_t m)
2407 {
2408 
2409 	m->md.mdpg_attrs = 0;
2410 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
2411 	LIST_INIT(&m->md.mdpg_pvoh);
2412 }
2413 
2414 /*
2415  * Return the number of managed mappings to the given physical page
2416  * that are wired.
2417  */
2418 int
2419 moea64_page_wired_mappings(vm_page_t m)
2420 {
2421 	struct pvo_entry *pvo;
2422 	int count;
2423 
2424 	count = 0;
2425 	if ((m->oflags & VPO_UNMANAGED) != 0)
2426 		return (count);
2427 	PV_PAGE_LOCK(m);
2428 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
2429 		if ((pvo->pvo_vaddr & (PVO_DEAD | PVO_WIRED)) == PVO_WIRED)
2430 			count++;
2431 	PV_PAGE_UNLOCK(m);
2432 	return (count);
2433 }
2434 
2435 static uintptr_t	moea64_vsidcontext;
2436 
2437 uintptr_t
2438 moea64_get_unique_vsid(void) {
2439 	u_int entropy;
2440 	register_t hash;
2441 	uint32_t mask;
2442 	int i;
2443 
2444 	entropy = 0;
2445 	__asm __volatile("mftb %0" : "=r"(entropy));
2446 
2447 	mtx_lock(&moea64_slb_mutex);
2448 	for (i = 0; i < NVSIDS; i += VSID_NBPW) {
2449 		u_int	n;
2450 
2451 		/*
2452 		 * Create a new value by mutiplying by a prime and adding in
2453 		 * entropy from the timebase register.  This is to make the
2454 		 * VSID more random so that the PT hash function collides
2455 		 * less often.  (Note that the prime casues gcc to do shifts
2456 		 * instead of a multiply.)
2457 		 */
2458 		moea64_vsidcontext = (moea64_vsidcontext * 0x1105) + entropy;
2459 		hash = moea64_vsidcontext & (NVSIDS - 1);
2460 		if (hash == 0)		/* 0 is special, avoid it */
2461 			continue;
2462 		n = hash >> 5;
2463 		mask = 1 << (hash & (VSID_NBPW - 1));
2464 		hash = (moea64_vsidcontext & VSID_HASHMASK);
2465 		if (moea64_vsid_bitmap[n] & mask) {	/* collision? */
2466 			/* anything free in this bucket? */
2467 			if (moea64_vsid_bitmap[n] == 0xffffffff) {
2468 				entropy = (moea64_vsidcontext >> 20);
2469 				continue;
2470 			}
2471 			i = ffs(~moea64_vsid_bitmap[n]) - 1;
2472 			mask = 1 << i;
2473 			hash &= rounddown2(VSID_HASHMASK, VSID_NBPW);
2474 			hash |= i;
2475 		}
2476 		if (hash == VSID_VRMA)	/* also special, avoid this too */
2477 			continue;
2478 		KASSERT(!(moea64_vsid_bitmap[n] & mask),
2479 		    ("Allocating in-use VSID %#zx\n", hash));
2480 		moea64_vsid_bitmap[n] |= mask;
2481 		mtx_unlock(&moea64_slb_mutex);
2482 		return (hash);
2483 	}
2484 
2485 	mtx_unlock(&moea64_slb_mutex);
2486 	panic("%s: out of segments",__func__);
2487 }
2488 
2489 #ifdef __powerpc64__
2490 int
2491 moea64_pinit(pmap_t pmap)
2492 {
2493 
2494 	RB_INIT(&pmap->pmap_pvo);
2495 
2496 	pmap->pm_slb_tree_root = slb_alloc_tree();
2497 	pmap->pm_slb = slb_alloc_user_cache();
2498 	pmap->pm_slb_len = 0;
2499 
2500 	return (1);
2501 }
2502 #else
2503 int
2504 moea64_pinit(pmap_t pmap)
2505 {
2506 	int	i;
2507 	uint32_t hash;
2508 
2509 	RB_INIT(&pmap->pmap_pvo);
2510 
2511 	if (pmap_bootstrapped)
2512 		pmap->pmap_phys = (pmap_t)moea64_kextract((vm_offset_t)pmap);
2513 	else
2514 		pmap->pmap_phys = pmap;
2515 
2516 	/*
2517 	 * Allocate some segment registers for this pmap.
2518 	 */
2519 	hash = moea64_get_unique_vsid();
2520 
2521 	for (i = 0; i < 16; i++)
2522 		pmap->pm_sr[i] = VSID_MAKE(i, hash);
2523 
2524 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_pinit: pm_sr[0] = 0"));
2525 
2526 	return (1);
2527 }
2528 #endif
2529 
2530 /*
2531  * Initialize the pmap associated with process 0.
2532  */
2533 void
2534 moea64_pinit0(pmap_t pm)
2535 {
2536 
2537 	PMAP_LOCK_INIT(pm);
2538 	moea64_pinit(pm);
2539 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
2540 }
2541 
2542 /*
2543  * Set the physical protection on the specified range of this map as requested.
2544  */
2545 static void
2546 moea64_pvo_protect( pmap_t pm, struct pvo_entry *pvo, vm_prot_t prot)
2547 {
2548 	struct vm_page *pg;
2549 	vm_prot_t oldprot;
2550 	int32_t refchg;
2551 
2552 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2553 
2554 	/*
2555 	 * Change the protection of the page.
2556 	 */
2557 	oldprot = pvo->pvo_pte.prot;
2558 	pvo->pvo_pte.prot = prot;
2559 	pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
2560 
2561 	/*
2562 	 * If the PVO is in the page table, update mapping
2563 	 */
2564 	refchg = moea64_pte_replace(pvo, MOEA64_PTE_PROT_UPDATE);
2565 	if (refchg < 0)
2566 		refchg = (oldprot & VM_PROT_WRITE) ? LPTE_CHG : 0;
2567 
2568 	if (pm != kernel_pmap && pg != NULL &&
2569 	    (pg->a.flags & PGA_EXECUTABLE) == 0 &&
2570 	    (pvo->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
2571 		if ((pg->oflags & VPO_UNMANAGED) == 0)
2572 			vm_page_aflag_set(pg, PGA_EXECUTABLE);
2573 		moea64_syncicache(pm, PVO_VADDR(pvo),
2574 		    PVO_PADDR(pvo), PAGE_SIZE);
2575 	}
2576 
2577 	/*
2578 	 * Update vm about the REF/CHG bits if the page is managed and we have
2579 	 * removed write access.
2580 	 */
2581 	if (pg != NULL && (pvo->pvo_vaddr & PVO_MANAGED) &&
2582 	    (oldprot & VM_PROT_WRITE)) {
2583 		refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2584 		if (refchg & LPTE_CHG)
2585 			vm_page_dirty(pg);
2586 		if (refchg & LPTE_REF)
2587 			vm_page_aflag_set(pg, PGA_REFERENCED);
2588 	}
2589 }
2590 
2591 void
2592 moea64_protect(pmap_t pm, vm_offset_t sva, vm_offset_t eva,
2593     vm_prot_t prot)
2594 {
2595 	struct	pvo_entry *pvo, key;
2596 
2597 	CTR4(KTR_PMAP, "moea64_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm,
2598 	    sva, eva, prot);
2599 
2600 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
2601 	    ("moea64_protect: non current pmap"));
2602 
2603 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
2604 		moea64_remove(pm, sva, eva);
2605 		return;
2606 	}
2607 
2608 	PMAP_LOCK(pm);
2609 	key.pvo_vaddr = sva;
2610 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2611 	    pvo != NULL && PVO_VADDR(pvo) < eva;
2612 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
2613 		if (PVO_IS_SP(pvo)) {
2614 			if (moea64_sp_pvo_in_range(pvo, sva, eva)) {
2615 				pvo = moea64_sp_protect(pvo, prot);
2616 				continue;
2617 			} else {
2618 				CTR1(KTR_PMAP, "%s: demote before protect",
2619 				    __func__);
2620 				moea64_sp_demote(pvo);
2621 			}
2622 		}
2623 		moea64_pvo_protect(pm, pvo, prot);
2624 	}
2625 	PMAP_UNLOCK(pm);
2626 }
2627 
2628 /*
2629  * Map a list of wired pages into kernel virtual address space.  This is
2630  * intended for temporary mappings which do not need page modification or
2631  * references recorded.  Existing mappings in the region are overwritten.
2632  */
2633 void
2634 moea64_qenter(vm_offset_t va, vm_page_t *m, int count)
2635 {
2636 	while (count-- > 0) {
2637 		moea64_kenter(va, VM_PAGE_TO_PHYS(*m));
2638 		va += PAGE_SIZE;
2639 		m++;
2640 	}
2641 }
2642 
2643 /*
2644  * Remove page mappings from kernel virtual address space.  Intended for
2645  * temporary mappings entered by moea64_qenter.
2646  */
2647 void
2648 moea64_qremove(vm_offset_t va, int count)
2649 {
2650 	while (count-- > 0) {
2651 		moea64_kremove(va);
2652 		va += PAGE_SIZE;
2653 	}
2654 }
2655 
2656 void
2657 moea64_release_vsid(uint64_t vsid)
2658 {
2659 	int idx, mask;
2660 
2661 	mtx_lock(&moea64_slb_mutex);
2662 	idx = vsid & (NVSIDS-1);
2663 	mask = 1 << (idx % VSID_NBPW);
2664 	idx /= VSID_NBPW;
2665 	KASSERT(moea64_vsid_bitmap[idx] & mask,
2666 	    ("Freeing unallocated VSID %#jx", vsid));
2667 	moea64_vsid_bitmap[idx] &= ~mask;
2668 	mtx_unlock(&moea64_slb_mutex);
2669 }
2670 
2671 void
2672 moea64_release(pmap_t pmap)
2673 {
2674 
2675 	/*
2676 	 * Free segment registers' VSIDs
2677 	 */
2678     #ifdef __powerpc64__
2679 	slb_free_tree(pmap);
2680 	slb_free_user_cache(pmap->pm_slb);
2681     #else
2682 	KASSERT(pmap->pm_sr[0] != 0, ("moea64_release: pm_sr[0] = 0"));
2683 
2684 	moea64_release_vsid(VSID_TO_HASH(pmap->pm_sr[0]));
2685     #endif
2686 }
2687 
2688 /*
2689  * Remove all pages mapped by the specified pmap
2690  */
2691 void
2692 moea64_remove_pages(pmap_t pm)
2693 {
2694 	struct pvo_entry *pvo, *tpvo;
2695 	struct pvo_dlist tofree;
2696 
2697 	SLIST_INIT(&tofree);
2698 
2699 	PMAP_LOCK(pm);
2700 	RB_FOREACH_SAFE(pvo, pvo_tree, &pm->pmap_pvo, tpvo) {
2701 		if (pvo->pvo_vaddr & PVO_WIRED)
2702 			continue;
2703 
2704 		/*
2705 		 * For locking reasons, remove this from the page table and
2706 		 * pmap, but save delinking from the vm_page for a second
2707 		 * pass
2708 		 */
2709 		moea64_pvo_remove_from_pmap(pvo);
2710 		SLIST_INSERT_HEAD(&tofree, pvo, pvo_dlink);
2711 	}
2712 	PMAP_UNLOCK(pm);
2713 
2714 	while (!SLIST_EMPTY(&tofree)) {
2715 		pvo = SLIST_FIRST(&tofree);
2716 		SLIST_REMOVE_HEAD(&tofree, pvo_dlink);
2717 		moea64_pvo_remove_from_page(pvo);
2718 		free_pvo_entry(pvo);
2719 	}
2720 }
2721 
2722 static void
2723 moea64_remove_locked(pmap_t pm, vm_offset_t sva, vm_offset_t eva,
2724     struct pvo_dlist *tofree)
2725 {
2726 	struct pvo_entry *pvo, *tpvo, key;
2727 
2728 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
2729 
2730 	key.pvo_vaddr = sva;
2731 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
2732 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
2733 		if (PVO_IS_SP(pvo)) {
2734 			if (moea64_sp_pvo_in_range(pvo, sva, eva)) {
2735 				tpvo = moea64_sp_remove(pvo, tofree);
2736 				continue;
2737 			} else {
2738 				CTR1(KTR_PMAP, "%s: demote before remove",
2739 				    __func__);
2740 				moea64_sp_demote(pvo);
2741 			}
2742 		}
2743 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
2744 
2745 		/*
2746 		 * For locking reasons, remove this from the page table and
2747 		 * pmap, but save delinking from the vm_page for a second
2748 		 * pass
2749 		 */
2750 		moea64_pvo_remove_from_pmap(pvo);
2751 		SLIST_INSERT_HEAD(tofree, pvo, pvo_dlink);
2752 	}
2753 }
2754 
2755 /*
2756  * Remove the given range of addresses from the specified map.
2757  */
2758 void
2759 moea64_remove(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
2760 {
2761 	struct pvo_entry *pvo;
2762 	struct pvo_dlist tofree;
2763 
2764 	/*
2765 	 * Perform an unsynchronized read.  This is, however, safe.
2766 	 */
2767 	if (pm->pm_stats.resident_count == 0)
2768 		return;
2769 
2770 	SLIST_INIT(&tofree);
2771 	PMAP_LOCK(pm);
2772 	moea64_remove_locked(pm, sva, eva, &tofree);
2773 	PMAP_UNLOCK(pm);
2774 
2775 	while (!SLIST_EMPTY(&tofree)) {
2776 		pvo = SLIST_FIRST(&tofree);
2777 		SLIST_REMOVE_HEAD(&tofree, pvo_dlink);
2778 		moea64_pvo_remove_from_page(pvo);
2779 		free_pvo_entry(pvo);
2780 	}
2781 }
2782 
2783 /*
2784  * Remove physical page from all pmaps in which it resides. moea64_pvo_remove()
2785  * will reflect changes in pte's back to the vm_page.
2786  */
2787 void
2788 moea64_remove_all(vm_page_t m)
2789 {
2790 	struct	pvo_entry *pvo, *next_pvo;
2791 	struct	pvo_head freequeue;
2792 	int	wasdead;
2793 	pmap_t	pmap;
2794 
2795 	LIST_INIT(&freequeue);
2796 
2797 	PV_PAGE_LOCK(m);
2798 	LIST_FOREACH_SAFE(pvo, vm_page_to_pvoh(m), pvo_vlink, next_pvo) {
2799 		pmap = pvo->pvo_pmap;
2800 		PMAP_LOCK(pmap);
2801 		wasdead = (pvo->pvo_vaddr & PVO_DEAD);
2802 		if (!wasdead) {
2803 			if (PVO_IS_SP(pvo)) {
2804 				CTR1(KTR_PMAP, "%s: demote before remove_all",
2805 				    __func__);
2806 				moea64_sp_demote(pvo);
2807 			}
2808 			moea64_pvo_remove_from_pmap(pvo);
2809 		}
2810 		moea64_pvo_remove_from_page_locked(pvo, m);
2811 		if (!wasdead)
2812 			LIST_INSERT_HEAD(&freequeue, pvo, pvo_vlink);
2813 		PMAP_UNLOCK(pmap);
2814 
2815 	}
2816 	KASSERT(!pmap_page_is_mapped(m), ("Page still has mappings"));
2817 	KASSERT((m->a.flags & PGA_WRITEABLE) == 0, ("Page still writable"));
2818 	PV_PAGE_UNLOCK(m);
2819 
2820 	/* Clean up UMA allocations */
2821 	LIST_FOREACH_SAFE(pvo, &freequeue, pvo_vlink, next_pvo)
2822 		free_pvo_entry(pvo);
2823 }
2824 
2825 /*
2826  * Allocate a physical page of memory directly from the phys_avail map.
2827  * Can only be called from moea64_bootstrap before avail start and end are
2828  * calculated.
2829  */
2830 vm_offset_t
2831 moea64_bootstrap_alloc(vm_size_t size, vm_size_t align)
2832 {
2833 	vm_offset_t	s, e;
2834 	int		i, j;
2835 
2836 	size = round_page(size);
2837 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2838 		if (align != 0)
2839 			s = roundup2(phys_avail[i], align);
2840 		else
2841 			s = phys_avail[i];
2842 		e = s + size;
2843 
2844 		if (s < phys_avail[i] || e > phys_avail[i + 1])
2845 			continue;
2846 
2847 		if (s + size > platform_real_maxaddr())
2848 			continue;
2849 
2850 		if (s == phys_avail[i]) {
2851 			phys_avail[i] += size;
2852 		} else if (e == phys_avail[i + 1]) {
2853 			phys_avail[i + 1] -= size;
2854 		} else {
2855 			for (j = phys_avail_count * 2; j > i; j -= 2) {
2856 				phys_avail[j] = phys_avail[j - 2];
2857 				phys_avail[j + 1] = phys_avail[j - 1];
2858 			}
2859 
2860 			phys_avail[i + 3] = phys_avail[i + 1];
2861 			phys_avail[i + 1] = s;
2862 			phys_avail[i + 2] = e;
2863 			phys_avail_count++;
2864 		}
2865 
2866 		return (s);
2867 	}
2868 	panic("moea64_bootstrap_alloc: could not allocate memory");
2869 }
2870 
2871 static int
2872 moea64_pvo_enter(struct pvo_entry *pvo, struct pvo_head *pvo_head,
2873     struct pvo_entry **oldpvop)
2874 {
2875 	struct pvo_entry *old_pvo;
2876 	int err;
2877 
2878 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2879 
2880 	STAT_MOEA64(moea64_pvo_enter_calls++);
2881 
2882 	/*
2883 	 * Add to pmap list
2884 	 */
2885 	old_pvo = RB_INSERT(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2886 
2887 	if (old_pvo != NULL) {
2888 		if (oldpvop != NULL)
2889 			*oldpvop = old_pvo;
2890 		return (EEXIST);
2891 	}
2892 
2893 	if (pvo_head != NULL) {
2894 		LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2895 	}
2896 
2897 	if (pvo->pvo_vaddr & PVO_WIRED)
2898 		pvo->pvo_pmap->pm_stats.wired_count++;
2899 	pvo->pvo_pmap->pm_stats.resident_count++;
2900 
2901 	/*
2902 	 * Insert it into the hardware page table
2903 	 */
2904 	err = moea64_pte_insert(pvo);
2905 	if (err != 0) {
2906 		panic("moea64_pvo_enter: overflow");
2907 	}
2908 
2909 	STAT_MOEA64(moea64_pvo_entries++);
2910 
2911 	if (pvo->pvo_pmap == kernel_pmap)
2912 		isync();
2913 
2914 #ifdef __powerpc64__
2915 	/*
2916 	 * Make sure all our bootstrap mappings are in the SLB as soon
2917 	 * as virtual memory is switched on.
2918 	 */
2919 	if (!pmap_bootstrapped)
2920 		moea64_bootstrap_slb_prefault(PVO_VADDR(pvo),
2921 		    pvo->pvo_vaddr & PVO_LARGE);
2922 #endif
2923 
2924 	return (0);
2925 }
2926 
2927 static void
2928 moea64_pvo_remove_from_pmap(struct pvo_entry *pvo)
2929 {
2930 	struct	vm_page *pg;
2931 	int32_t refchg;
2932 
2933 	KASSERT(pvo->pvo_pmap != NULL, ("Trying to remove PVO with no pmap"));
2934 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
2935 	KASSERT(!(pvo->pvo_vaddr & PVO_DEAD), ("Trying to remove dead PVO"));
2936 
2937 	/*
2938 	 * If there is an active pte entry, we need to deactivate it
2939 	 */
2940 	refchg = moea64_pte_unset(pvo);
2941 	if (refchg < 0) {
2942 		/*
2943 		 * If it was evicted from the page table, be pessimistic and
2944 		 * dirty the page.
2945 		 */
2946 		if (pvo->pvo_pte.prot & VM_PROT_WRITE)
2947 			refchg = LPTE_CHG;
2948 		else
2949 			refchg = 0;
2950 	}
2951 
2952 	/*
2953 	 * Update our statistics.
2954 	 */
2955 	pvo->pvo_pmap->pm_stats.resident_count--;
2956 	if (pvo->pvo_vaddr & PVO_WIRED)
2957 		pvo->pvo_pmap->pm_stats.wired_count--;
2958 
2959 	/*
2960 	 * Remove this PVO from the pmap list.
2961 	 */
2962 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2963 
2964 	/*
2965 	 * Mark this for the next sweep
2966 	 */
2967 	pvo->pvo_vaddr |= PVO_DEAD;
2968 
2969 	/* Send RC bits to VM */
2970 	if ((pvo->pvo_vaddr & PVO_MANAGED) &&
2971 	    (pvo->pvo_pte.prot & VM_PROT_WRITE)) {
2972 		pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
2973 		if (pg != NULL) {
2974 			refchg |= atomic_readandclear_32(&pg->md.mdpg_attrs);
2975 			if (refchg & LPTE_CHG)
2976 				vm_page_dirty(pg);
2977 			if (refchg & LPTE_REF)
2978 				vm_page_aflag_set(pg, PGA_REFERENCED);
2979 		}
2980 	}
2981 }
2982 
2983 static inline void
2984 moea64_pvo_remove_from_page_locked(struct pvo_entry *pvo,
2985     vm_page_t m)
2986 {
2987 
2988 	KASSERT(pvo->pvo_vaddr & PVO_DEAD, ("Trying to delink live page"));
2989 
2990 	/* Use NULL pmaps as a sentinel for races in page deletion */
2991 	if (pvo->pvo_pmap == NULL)
2992 		return;
2993 	pvo->pvo_pmap = NULL;
2994 
2995 	/*
2996 	 * Update vm about page writeability/executability if managed
2997 	 */
2998 	PV_LOCKASSERT(PVO_PADDR(pvo));
2999 	if (pvo->pvo_vaddr & PVO_MANAGED) {
3000 		if (m != NULL) {
3001 			LIST_REMOVE(pvo, pvo_vlink);
3002 			if (LIST_EMPTY(vm_page_to_pvoh(m)))
3003 				vm_page_aflag_clear(m,
3004 				    PGA_WRITEABLE | PGA_EXECUTABLE);
3005 		}
3006 	}
3007 
3008 	STAT_MOEA64(moea64_pvo_entries--);
3009 	STAT_MOEA64(moea64_pvo_remove_calls++);
3010 }
3011 
3012 static void
3013 moea64_pvo_remove_from_page(struct pvo_entry *pvo)
3014 {
3015 	vm_page_t pg = NULL;
3016 
3017 	if (pvo->pvo_vaddr & PVO_MANAGED)
3018 		pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
3019 
3020 	PV_LOCK(PVO_PADDR(pvo));
3021 	moea64_pvo_remove_from_page_locked(pvo, pg);
3022 	PV_UNLOCK(PVO_PADDR(pvo));
3023 }
3024 
3025 static struct pvo_entry *
3026 moea64_pvo_find_va(pmap_t pm, vm_offset_t va)
3027 {
3028 	struct pvo_entry key;
3029 
3030 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
3031 
3032 	key.pvo_vaddr = va & ~ADDR_POFF;
3033 	return (RB_FIND(pvo_tree, &pm->pmap_pvo, &key));
3034 }
3035 
3036 static boolean_t
3037 moea64_query_bit(vm_page_t m, uint64_t ptebit)
3038 {
3039 	struct	pvo_entry *pvo;
3040 	int64_t ret;
3041 	boolean_t rv;
3042 	vm_page_t sp;
3043 
3044 	/*
3045 	 * See if this bit is stored in the page already.
3046 	 *
3047 	 * For superpages, the bit is stored in the first vm page.
3048 	 */
3049 	if ((m->md.mdpg_attrs & ptebit) != 0 ||
3050 	    ((sp = PHYS_TO_VM_PAGE(VM_PAGE_TO_PHYS(m) & ~HPT_SP_MASK)) != NULL &&
3051 	     (sp->md.mdpg_attrs & (ptebit | MDPG_ATTR_SP)) ==
3052 	     (ptebit | MDPG_ATTR_SP)))
3053 		return (TRUE);
3054 
3055 	/*
3056 	 * Examine each PTE.  Sync so that any pending REF/CHG bits are
3057 	 * flushed to the PTEs.
3058 	 */
3059 	rv = FALSE;
3060 	powerpc_sync();
3061 	PV_PAGE_LOCK(m);
3062 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
3063 		if (PVO_IS_SP(pvo)) {
3064 			ret = moea64_sp_query(pvo, ptebit);
3065 			/*
3066 			 * If SP was not demoted, check its REF/CHG bits here.
3067 			 */
3068 			if (ret != -1) {
3069 				if ((ret & ptebit) != 0) {
3070 					rv = TRUE;
3071 					break;
3072 				}
3073 				continue;
3074 			}
3075 			/* else, fallthrough */
3076 		}
3077 
3078 		ret = 0;
3079 
3080 		/*
3081 		 * See if this pvo has a valid PTE.  if so, fetch the
3082 		 * REF/CHG bits from the valid PTE.  If the appropriate
3083 		 * ptebit is set, return success.
3084 		 */
3085 		PMAP_LOCK(pvo->pvo_pmap);
3086 		if (!(pvo->pvo_vaddr & PVO_DEAD))
3087 			ret = moea64_pte_synch(pvo);
3088 		PMAP_UNLOCK(pvo->pvo_pmap);
3089 
3090 		if (ret > 0) {
3091 			atomic_set_32(&m->md.mdpg_attrs,
3092 			    ret & (LPTE_CHG | LPTE_REF));
3093 			if (ret & ptebit) {
3094 				rv = TRUE;
3095 				break;
3096 			}
3097 		}
3098 	}
3099 	PV_PAGE_UNLOCK(m);
3100 
3101 	return (rv);
3102 }
3103 
3104 static u_int
3105 moea64_clear_bit(vm_page_t m, u_int64_t ptebit)
3106 {
3107 	u_int	count;
3108 	struct	pvo_entry *pvo;
3109 	int64_t ret;
3110 
3111 	/*
3112 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
3113 	 * we can reset the right ones).
3114 	 */
3115 	powerpc_sync();
3116 
3117 	/*
3118 	 * For each pvo entry, clear the pte's ptebit.
3119 	 */
3120 	count = 0;
3121 	PV_PAGE_LOCK(m);
3122 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
3123 		if (PVO_IS_SP(pvo)) {
3124 			if ((ret = moea64_sp_clear(pvo, m, ptebit)) != -1) {
3125 				count += ret;
3126 				continue;
3127 			}
3128 		}
3129 		ret = 0;
3130 
3131 		PMAP_LOCK(pvo->pvo_pmap);
3132 		if (!(pvo->pvo_vaddr & PVO_DEAD))
3133 			ret = moea64_pte_clear(pvo, ptebit);
3134 		PMAP_UNLOCK(pvo->pvo_pmap);
3135 
3136 		if (ret > 0 && (ret & ptebit))
3137 			count++;
3138 	}
3139 	atomic_clear_32(&m->md.mdpg_attrs, ptebit);
3140 	PV_PAGE_UNLOCK(m);
3141 
3142 	return (count);
3143 }
3144 
3145 boolean_t
3146 moea64_dev_direct_mapped(vm_paddr_t pa, vm_size_t size)
3147 {
3148 	struct pvo_entry *pvo, key;
3149 	vm_offset_t ppa;
3150 	int error = 0;
3151 
3152 	if (hw_direct_map && mem_valid(pa, size) == 0)
3153 		return (0);
3154 
3155 	PMAP_LOCK(kernel_pmap);
3156 	ppa = pa & ~ADDR_POFF;
3157 	key.pvo_vaddr = DMAP_BASE_ADDRESS + ppa;
3158 	for (pvo = RB_FIND(pvo_tree, &kernel_pmap->pmap_pvo, &key);
3159 	    ppa < pa + size; ppa += PAGE_SIZE,
3160 	    pvo = RB_NEXT(pvo_tree, &kernel_pmap->pmap_pvo, pvo)) {
3161 		if (pvo == NULL || PVO_PADDR(pvo) != ppa) {
3162 			error = EFAULT;
3163 			break;
3164 		}
3165 	}
3166 	PMAP_UNLOCK(kernel_pmap);
3167 
3168 	return (error);
3169 }
3170 
3171 /*
3172  * Map a set of physical memory pages into the kernel virtual
3173  * address space. Return a pointer to where it is mapped. This
3174  * routine is intended to be used for mapping device memory,
3175  * NOT real memory.
3176  */
3177 void *
3178 moea64_mapdev_attr(vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
3179 {
3180 	vm_offset_t va, tmpva, ppa, offset;
3181 
3182 	ppa = trunc_page(pa);
3183 	offset = pa & PAGE_MASK;
3184 	size = roundup2(offset + size, PAGE_SIZE);
3185 
3186 	va = kva_alloc(size);
3187 
3188 	if (!va)
3189 		panic("moea64_mapdev: Couldn't alloc kernel virtual memory");
3190 
3191 	for (tmpva = va; size > 0;) {
3192 		moea64_kenter_attr(tmpva, ppa, ma);
3193 		size -= PAGE_SIZE;
3194 		tmpva += PAGE_SIZE;
3195 		ppa += PAGE_SIZE;
3196 	}
3197 
3198 	return ((void *)(va + offset));
3199 }
3200 
3201 void *
3202 moea64_mapdev(vm_paddr_t pa, vm_size_t size)
3203 {
3204 
3205 	return moea64_mapdev_attr(pa, size, VM_MEMATTR_DEFAULT);
3206 }
3207 
3208 void
3209 moea64_unmapdev(vm_offset_t va, vm_size_t size)
3210 {
3211 	vm_offset_t base, offset;
3212 
3213 	base = trunc_page(va);
3214 	offset = va & PAGE_MASK;
3215 	size = roundup2(offset + size, PAGE_SIZE);
3216 
3217 	moea64_qremove(base, atop(size));
3218 	kva_free(base, size);
3219 }
3220 
3221 void
3222 moea64_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz)
3223 {
3224 	struct pvo_entry *pvo;
3225 	vm_offset_t lim;
3226 	vm_paddr_t pa;
3227 	vm_size_t len;
3228 
3229 	if (__predict_false(pm == NULL))
3230 		pm = &curthread->td_proc->p_vmspace->vm_pmap;
3231 
3232 	PMAP_LOCK(pm);
3233 	while (sz > 0) {
3234 		lim = round_page(va+1);
3235 		len = MIN(lim - va, sz);
3236 		pvo = moea64_pvo_find_va(pm, va & ~ADDR_POFF);
3237 		if (pvo != NULL && !(pvo->pvo_pte.pa & LPTE_I)) {
3238 			pa = PVO_PADDR(pvo) | (va & ADDR_POFF);
3239 			moea64_syncicache(pm, va, pa, len);
3240 		}
3241 		va += len;
3242 		sz -= len;
3243 	}
3244 	PMAP_UNLOCK(pm);
3245 }
3246 
3247 void
3248 moea64_dumpsys_map(vm_paddr_t pa, size_t sz, void **va)
3249 {
3250 
3251 	*va = (void *)(uintptr_t)pa;
3252 }
3253 
3254 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
3255 
3256 void
3257 moea64_scan_init()
3258 {
3259 	struct pvo_entry *pvo;
3260 	vm_offset_t va;
3261 	int i;
3262 
3263 	if (!do_minidump) {
3264 		/* Initialize phys. segments for dumpsys(). */
3265 		memset(&dump_map, 0, sizeof(dump_map));
3266 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
3267 		for (i = 0; i < pregions_sz; i++) {
3268 			dump_map[i].pa_start = pregions[i].mr_start;
3269 			dump_map[i].pa_size = pregions[i].mr_size;
3270 		}
3271 		return;
3272 	}
3273 
3274 	/* Virtual segments for minidumps: */
3275 	memset(&dump_map, 0, sizeof(dump_map));
3276 
3277 	/* 1st: kernel .data and .bss. */
3278 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
3279 	dump_map[0].pa_size = round_page((uintptr_t)_end) -
3280 	    dump_map[0].pa_start;
3281 
3282 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
3283 	dump_map[1].pa_start = (vm_paddr_t)(uintptr_t)msgbufp->msg_ptr;
3284 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
3285 
3286 	/* 3rd: kernel VM. */
3287 	va = dump_map[1].pa_start + dump_map[1].pa_size;
3288 	/* Find start of next chunk (from va). */
3289 	while (va < virtual_end) {
3290 		/* Don't dump the buffer cache. */
3291 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
3292 			va = kmi.buffer_eva;
3293 			continue;
3294 		}
3295 		pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
3296 		if (pvo != NULL && !(pvo->pvo_vaddr & PVO_DEAD))
3297 			break;
3298 		va += PAGE_SIZE;
3299 	}
3300 	if (va < virtual_end) {
3301 		dump_map[2].pa_start = va;
3302 		va += PAGE_SIZE;
3303 		/* Find last page in chunk. */
3304 		while (va < virtual_end) {
3305 			/* Don't run into the buffer cache. */
3306 			if (va == kmi.buffer_sva)
3307 				break;
3308 			pvo = moea64_pvo_find_va(kernel_pmap, va & ~ADDR_POFF);
3309 			if (pvo == NULL || (pvo->pvo_vaddr & PVO_DEAD))
3310 				break;
3311 			va += PAGE_SIZE;
3312 		}
3313 		dump_map[2].pa_size = va - dump_map[2].pa_start;
3314 	}
3315 }
3316 
3317 #ifdef __powerpc64__
3318 
3319 static size_t
3320 moea64_scan_pmap(struct bitset *dump_bitset)
3321 {
3322 	struct pvo_entry *pvo;
3323 	vm_paddr_t pa, pa_end;
3324 	vm_offset_t va, pgva, kstart, kend, kstart_lp, kend_lp;
3325 	uint64_t lpsize;
3326 
3327 	lpsize = moea64_large_page_size;
3328 	kstart = trunc_page((vm_offset_t)_etext);
3329 	kend = round_page((vm_offset_t)_end);
3330 	kstart_lp = kstart & ~moea64_large_page_mask;
3331 	kend_lp = (kend + moea64_large_page_mask) & ~moea64_large_page_mask;
3332 
3333 	CTR4(KTR_PMAP, "moea64_scan_pmap: kstart=0x%016lx, kend=0x%016lx, "
3334 	    "kstart_lp=0x%016lx, kend_lp=0x%016lx",
3335 	    kstart, kend, kstart_lp, kend_lp);
3336 
3337 	PMAP_LOCK(kernel_pmap);
3338 	RB_FOREACH(pvo, pvo_tree, &kernel_pmap->pmap_pvo) {
3339 		va = pvo->pvo_vaddr;
3340 
3341 		if (va & PVO_DEAD)
3342 			continue;
3343 
3344 		/* Skip DMAP (except kernel area) */
3345 		if (va >= DMAP_BASE_ADDRESS && va <= DMAP_MAX_ADDRESS) {
3346 			if (va & PVO_LARGE) {
3347 				pgva = va & ~moea64_large_page_mask;
3348 				if (pgva < kstart_lp || pgva >= kend_lp)
3349 					continue;
3350 			} else {
3351 				pgva = trunc_page(va);
3352 				if (pgva < kstart || pgva >= kend)
3353 					continue;
3354 			}
3355 		}
3356 
3357 		pa = PVO_PADDR(pvo);
3358 
3359 		if (va & PVO_LARGE) {
3360 			pa_end = pa + lpsize;
3361 			for (; pa < pa_end; pa += PAGE_SIZE) {
3362 				if (vm_phys_is_dumpable(pa))
3363 					vm_page_dump_add(dump_bitset, pa);
3364 			}
3365 		} else {
3366 			if (vm_phys_is_dumpable(pa))
3367 				vm_page_dump_add(dump_bitset, pa);
3368 		}
3369 	}
3370 	PMAP_UNLOCK(kernel_pmap);
3371 
3372 	return (sizeof(struct lpte) * moea64_pteg_count * 8);
3373 }
3374 
3375 static struct dump_context dump_ctx;
3376 
3377 static void *
3378 moea64_dump_pmap_init(unsigned blkpgs)
3379 {
3380 	dump_ctx.ptex = 0;
3381 	dump_ctx.ptex_end = moea64_pteg_count * 8;
3382 	dump_ctx.blksz = blkpgs * PAGE_SIZE;
3383 	return (&dump_ctx);
3384 }
3385 
3386 #else
3387 
3388 static size_t
3389 moea64_scan_pmap(struct bitset *dump_bitset __unused)
3390 {
3391 	return (0);
3392 }
3393 
3394 static void *
3395 moea64_dump_pmap_init(unsigned blkpgs)
3396 {
3397 	return (NULL);
3398 }
3399 
3400 #endif
3401 
3402 #ifdef __powerpc64__
3403 static void
3404 moea64_map_range(vm_offset_t va, vm_paddr_t pa, vm_size_t npages)
3405 {
3406 
3407 	for (; npages > 0; --npages) {
3408 		if (moea64_large_page_size != 0 &&
3409 		    (pa & moea64_large_page_mask) == 0 &&
3410 		    (va & moea64_large_page_mask) == 0 &&
3411 		    npages >= (moea64_large_page_size >> PAGE_SHIFT)) {
3412 			PMAP_LOCK(kernel_pmap);
3413 			moea64_kenter_large(va, pa, 0, 0);
3414 			PMAP_UNLOCK(kernel_pmap);
3415 			pa += moea64_large_page_size;
3416 			va += moea64_large_page_size;
3417 			npages -= (moea64_large_page_size >> PAGE_SHIFT) - 1;
3418 		} else {
3419 			moea64_kenter(va, pa);
3420 			pa += PAGE_SIZE;
3421 			va += PAGE_SIZE;
3422 		}
3423 	}
3424 }
3425 
3426 static void
3427 moea64_page_array_startup(long pages)
3428 {
3429 	long dom_pages[MAXMEMDOM];
3430 	vm_paddr_t pa;
3431 	vm_offset_t va, vm_page_base;
3432 	vm_size_t needed, size;
3433 	long page;
3434 	int domain;
3435 	int i;
3436 
3437 	vm_page_base = 0xd000000000000000ULL;
3438 
3439 	/* Short-circuit single-domain systems. */
3440 	if (vm_ndomains == 1) {
3441 		size = round_page(pages * sizeof(struct vm_page));
3442 		pa = vm_phys_early_alloc(0, size);
3443 		vm_page_base = moea64_map(&vm_page_base,
3444 		    pa, pa + size, VM_PROT_READ | VM_PROT_WRITE);
3445 		vm_page_array_size = pages;
3446 		vm_page_array = (vm_page_t)vm_page_base;
3447 		return;
3448 	}
3449 
3450 	page = 0;
3451 	for (i = 0; i < MAXMEMDOM; i++)
3452 		dom_pages[i] = 0;
3453 
3454 	/* Now get the number of pages required per domain. */
3455 	for (i = 0; i < vm_phys_nsegs; i++) {
3456 		domain = vm_phys_segs[i].domain;
3457 		KASSERT(domain < MAXMEMDOM,
3458 		    ("Invalid vm_phys_segs NUMA domain %d!\n", domain));
3459 		/* Get size of vm_page_array needed for this segment. */
3460 		size = btoc(vm_phys_segs[i].end - vm_phys_segs[i].start);
3461 		dom_pages[domain] += size;
3462 	}
3463 
3464 	for (i = 0; phys_avail[i + 1] != 0; i+= 2) {
3465 		domain = vm_phys_domain(phys_avail[i]);
3466 		KASSERT(domain < MAXMEMDOM,
3467 		    ("Invalid phys_avail NUMA domain %d!\n", domain));
3468 		size = btoc(phys_avail[i + 1] - phys_avail[i]);
3469 		dom_pages[domain] += size;
3470 	}
3471 
3472 	/*
3473 	 * Map in chunks that can get us all 16MB pages.  There will be some
3474 	 * overlap between domains, but that's acceptable for now.
3475 	 */
3476 	vm_page_array_size = 0;
3477 	va = vm_page_base;
3478 	for (i = 0; i < MAXMEMDOM && vm_page_array_size < pages; i++) {
3479 		if (dom_pages[i] == 0)
3480 			continue;
3481 		size = ulmin(pages - vm_page_array_size, dom_pages[i]);
3482 		size = round_page(size * sizeof(struct vm_page));
3483 		needed = size;
3484 		size = roundup2(size, moea64_large_page_size);
3485 		pa = vm_phys_early_alloc(i, size);
3486 		vm_page_array_size += size / sizeof(struct vm_page);
3487 		moea64_map_range(va, pa, size >> PAGE_SHIFT);
3488 		/* Scoot up domain 0, to reduce the domain page overlap. */
3489 		if (i == 0)
3490 			vm_page_base += size - needed;
3491 		va += size;
3492 	}
3493 	vm_page_array = (vm_page_t)vm_page_base;
3494 	vm_page_array_size = pages;
3495 }
3496 #endif
3497 
3498 static int64_t
3499 moea64_null_method(void)
3500 {
3501 	return (0);
3502 }
3503 
3504 static int64_t moea64_pte_replace_default(struct pvo_entry *pvo, int flags)
3505 {
3506 	int64_t refchg;
3507 
3508 	refchg = moea64_pte_unset(pvo);
3509 	moea64_pte_insert(pvo);
3510 
3511 	return (refchg);
3512 }
3513 
3514 struct moea64_funcs *moea64_ops;
3515 
3516 #define DEFINE_OEA64_IFUNC(ret, func, args, def)		\
3517 	DEFINE_IFUNC(, ret, moea64_##func, args) {		\
3518 		moea64_##func##_t f;				\
3519 		if (moea64_ops == NULL)				\
3520 			return ((moea64_##func##_t)def);	\
3521 		f = moea64_ops->func;				\
3522 		return (f != NULL ? f : (moea64_##func##_t)def);\
3523 	}
3524 
3525 void
3526 moea64_install(void)
3527 {
3528 #ifdef __powerpc64__
3529 	if (hw_direct_map == -1) {
3530 		moea64_probe_large_page();
3531 
3532 		/* Use a direct map if we have large page support */
3533 		if (moea64_large_page_size > 0)
3534 			hw_direct_map = 1;
3535 		else
3536 			hw_direct_map = 0;
3537 	}
3538 #endif
3539 
3540 	/*
3541 	 * Default to non-DMAP, and switch over to DMAP functions once we know
3542 	 * we have DMAP.
3543 	 */
3544 	if (hw_direct_map) {
3545 		moea64_methods.quick_enter_page = moea64_quick_enter_page_dmap;
3546 		moea64_methods.quick_remove_page = NULL;
3547 		moea64_methods.copy_page = moea64_copy_page_dmap;
3548 		moea64_methods.zero_page = moea64_zero_page_dmap;
3549 		moea64_methods.copy_pages = moea64_copy_pages_dmap;
3550 	}
3551 }
3552 
3553 DEFINE_OEA64_IFUNC(int64_t, pte_replace, (struct pvo_entry *, int),
3554     moea64_pte_replace_default)
3555 DEFINE_OEA64_IFUNC(int64_t, pte_insert, (struct pvo_entry *), moea64_null_method)
3556 DEFINE_OEA64_IFUNC(int64_t, pte_unset, (struct pvo_entry *), moea64_null_method)
3557 DEFINE_OEA64_IFUNC(int64_t, pte_clear, (struct pvo_entry *, uint64_t),
3558     moea64_null_method)
3559 DEFINE_OEA64_IFUNC(int64_t, pte_synch, (struct pvo_entry *), moea64_null_method)
3560 DEFINE_OEA64_IFUNC(int64_t, pte_insert_sp, (struct pvo_entry *), moea64_null_method)
3561 DEFINE_OEA64_IFUNC(int64_t, pte_unset_sp, (struct pvo_entry *), moea64_null_method)
3562 DEFINE_OEA64_IFUNC(int64_t, pte_replace_sp, (struct pvo_entry *), moea64_null_method)
3563 
3564 /* Superpage functions */
3565 
3566 /* MMU interface */
3567 
3568 static bool
3569 moea64_ps_enabled(pmap_t pmap)
3570 {
3571 	return (superpages_enabled);
3572 }
3573 
3574 static void
3575 moea64_align_superpage(vm_object_t object, vm_ooffset_t offset,
3576     vm_offset_t *addr, vm_size_t size)
3577 {
3578 	vm_offset_t sp_offset;
3579 
3580 	if (size < HPT_SP_SIZE)
3581 		return;
3582 
3583 	CTR4(KTR_PMAP, "%s: offs=%#jx, addr=%p, size=%#jx",
3584 	    __func__, (uintmax_t)offset, addr, (uintmax_t)size);
3585 
3586 	if (object != NULL && (object->flags & OBJ_COLORED) != 0)
3587 		offset += ptoa(object->pg_color);
3588 	sp_offset = offset & HPT_SP_MASK;
3589 	if (size - ((HPT_SP_SIZE - sp_offset) & HPT_SP_MASK) < HPT_SP_SIZE ||
3590 	    (*addr & HPT_SP_MASK) == sp_offset)
3591 		return;
3592 	if ((*addr & HPT_SP_MASK) < sp_offset)
3593 		*addr = (*addr & ~HPT_SP_MASK) + sp_offset;
3594 	else
3595 		*addr = ((*addr + HPT_SP_MASK) & ~HPT_SP_MASK) + sp_offset;
3596 }
3597 
3598 /* Helpers */
3599 
3600 static __inline void
3601 moea64_pvo_cleanup(struct pvo_dlist *tofree)
3602 {
3603 	struct pvo_entry *pvo;
3604 
3605 	/* clean up */
3606 	while (!SLIST_EMPTY(tofree)) {
3607 		pvo = SLIST_FIRST(tofree);
3608 		SLIST_REMOVE_HEAD(tofree, pvo_dlink);
3609 		if (pvo->pvo_vaddr & PVO_DEAD)
3610 			moea64_pvo_remove_from_page(pvo);
3611 		free_pvo_entry(pvo);
3612 	}
3613 }
3614 
3615 static __inline uint16_t
3616 pvo_to_vmpage_flags(struct pvo_entry *pvo)
3617 {
3618 	uint16_t flags;
3619 
3620 	flags = 0;
3621 	if ((pvo->pvo_pte.prot & VM_PROT_WRITE) != 0)
3622 		flags |= PGA_WRITEABLE;
3623 	if ((pvo->pvo_pte.prot & VM_PROT_EXECUTE) != 0)
3624 		flags |= PGA_EXECUTABLE;
3625 
3626 	return (flags);
3627 }
3628 
3629 /*
3630  * Check if the given pvo and its superpage are in sva-eva range.
3631  */
3632 static __inline bool
3633 moea64_sp_pvo_in_range(struct pvo_entry *pvo, vm_offset_t sva, vm_offset_t eva)
3634 {
3635 	vm_offset_t spva;
3636 
3637 	spva = PVO_VADDR(pvo) & ~HPT_SP_MASK;
3638 	if (spva >= sva && spva + HPT_SP_SIZE <= eva) {
3639 		/*
3640 		 * Because this function is intended to be called from loops
3641 		 * that iterate over ordered pvo entries, if the condition
3642 		 * above is true then the pvo must be the first of its
3643 		 * superpage.
3644 		 */
3645 		KASSERT(PVO_VADDR(pvo) == spva,
3646 		    ("%s: unexpected unaligned superpage pvo", __func__));
3647 		return (true);
3648 	}
3649 	return (false);
3650 }
3651 
3652 /*
3653  * Update vm about the REF/CHG bits if the superpage is managed and
3654  * has (or had) write access.
3655  */
3656 static void
3657 moea64_sp_refchg_process(struct pvo_entry *sp, vm_page_t m,
3658     int64_t sp_refchg, vm_prot_t prot)
3659 {
3660 	vm_page_t m_end;
3661 	int64_t refchg;
3662 
3663 	if ((sp->pvo_vaddr & PVO_MANAGED) != 0 && (prot & VM_PROT_WRITE) != 0) {
3664 		for (m_end = &m[HPT_SP_PAGES]; m < m_end; m++) {
3665 			refchg = sp_refchg |
3666 			    atomic_readandclear_32(&m->md.mdpg_attrs);
3667 			if (refchg & LPTE_CHG)
3668 				vm_page_dirty(m);
3669 			if (refchg & LPTE_REF)
3670 				vm_page_aflag_set(m, PGA_REFERENCED);
3671 		}
3672 	}
3673 }
3674 
3675 /* Superpage ops */
3676 
3677 static int
3678 moea64_sp_enter(pmap_t pmap, vm_offset_t va, vm_page_t m,
3679     vm_prot_t prot, u_int flags, int8_t psind)
3680 {
3681 	struct pvo_entry *pvo, **pvos;
3682 	struct pvo_head *pvo_head;
3683 	vm_offset_t sva;
3684 	vm_page_t sm;
3685 	vm_paddr_t pa, spa;
3686 	bool sync;
3687 	struct pvo_dlist tofree;
3688 	int error, i;
3689 	uint16_t aflags;
3690 
3691 	KASSERT((va & HPT_SP_MASK) == 0, ("%s: va %#jx unaligned",
3692 	    __func__, (uintmax_t)va));
3693 	KASSERT(psind == 1, ("%s: invalid psind: %d", __func__, psind));
3694 	KASSERT(m->psind == 1, ("%s: invalid m->psind: %d",
3695 	    __func__, m->psind));
3696 	KASSERT(pmap != kernel_pmap,
3697 	    ("%s: function called with kernel pmap", __func__));
3698 
3699 	CTR5(KTR_PMAP, "%s: va=%#jx, pa=%#jx, prot=%#x, flags=%#x, psind=1",
3700 	    __func__, (uintmax_t)va, (uintmax_t)VM_PAGE_TO_PHYS(m),
3701 	    prot, flags);
3702 
3703 	SLIST_INIT(&tofree);
3704 
3705 	sva = va;
3706 	sm = m;
3707 	spa = pa = VM_PAGE_TO_PHYS(sm);
3708 
3709 	/* Try to allocate all PVOs first, to make failure handling easier. */
3710 	pvos = malloc(HPT_SP_PAGES * sizeof(struct pvo_entry *), M_TEMP,
3711 	    M_NOWAIT);
3712 	if (pvos == NULL) {
3713 		CTR1(KTR_PMAP, "%s: failed to alloc pvo array", __func__);
3714 		return (KERN_RESOURCE_SHORTAGE);
3715 	}
3716 
3717 	for (i = 0; i < HPT_SP_PAGES; i++) {
3718 		pvos[i] = alloc_pvo_entry(0);
3719 		if (pvos[i] == NULL) {
3720 			CTR1(KTR_PMAP, "%s: failed to alloc pvo", __func__);
3721 			for (i = i - 1; i >= 0; i--)
3722 				free_pvo_entry(pvos[i]);
3723 			free(pvos, M_TEMP);
3724 			return (KERN_RESOURCE_SHORTAGE);
3725 		}
3726 	}
3727 
3728 	SP_PV_LOCK_ALIGNED(spa);
3729 	PMAP_LOCK(pmap);
3730 
3731 	/* Note: moea64_remove_locked() also clears cached REF/CHG bits. */
3732 	moea64_remove_locked(pmap, va, va + HPT_SP_SIZE, &tofree);
3733 
3734 	/* Enter pages */
3735 	for (i = 0; i < HPT_SP_PAGES;
3736 	    i++, va += PAGE_SIZE, pa += PAGE_SIZE, m++) {
3737 		pvo = pvos[i];
3738 
3739 		pvo->pvo_pte.prot = prot;
3740 		pvo->pvo_pte.pa = (pa & ~HPT_SP_MASK) | LPTE_LP_4K_16M |
3741 		    moea64_calc_wimg(pa, pmap_page_get_memattr(m));
3742 
3743 		if ((flags & PMAP_ENTER_WIRED) != 0)
3744 			pvo->pvo_vaddr |= PVO_WIRED;
3745 		pvo->pvo_vaddr |= PVO_LARGE;
3746 
3747 		if ((m->oflags & VPO_UNMANAGED) != 0)
3748 			pvo_head = NULL;
3749 		else {
3750 			pvo_head = &m->md.mdpg_pvoh;
3751 			pvo->pvo_vaddr |= PVO_MANAGED;
3752 		}
3753 
3754 		init_pvo_entry(pvo, pmap, va);
3755 
3756 		error = moea64_pvo_enter(pvo, pvo_head, NULL);
3757 		/*
3758 		 * All superpage PVOs were previously removed, so no errors
3759 		 * should occur while inserting the new ones.
3760 		 */
3761 		KASSERT(error == 0, ("%s: unexpected error "
3762 			    "when inserting superpage PVO: %d",
3763 			    __func__, error));
3764 	}
3765 
3766 	PMAP_UNLOCK(pmap);
3767 	SP_PV_UNLOCK_ALIGNED(spa);
3768 
3769 	sync = (sm->a.flags & PGA_EXECUTABLE) == 0;
3770 	/* Note: moea64_pvo_cleanup() also clears page prot. flags. */
3771 	moea64_pvo_cleanup(&tofree);
3772 	pvo = pvos[0];
3773 
3774 	/* Set vm page flags */
3775 	aflags = pvo_to_vmpage_flags(pvo);
3776 	if (aflags != 0)
3777 		for (m = sm; m < &sm[HPT_SP_PAGES]; m++)
3778 			vm_page_aflag_set(m, aflags);
3779 
3780 	/*
3781 	 * Flush the page from the instruction cache if this page is
3782 	 * mapped executable and cacheable.
3783 	 */
3784 	if (sync && (pvo->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0)
3785 		moea64_syncicache(pmap, sva, spa, HPT_SP_SIZE);
3786 
3787 	atomic_add_long(&sp_mappings, 1);
3788 	CTR3(KTR_PMAP, "%s: SP success for va %#jx in pmap %p",
3789 	    __func__, (uintmax_t)sva, pmap);
3790 
3791 	free(pvos, M_TEMP);
3792 	return (KERN_SUCCESS);
3793 }
3794 
3795 static void
3796 moea64_sp_promote(pmap_t pmap, vm_offset_t va, vm_page_t m)
3797 {
3798 	struct pvo_entry *first, *pvo;
3799 	vm_paddr_t pa, pa_end;
3800 	vm_offset_t sva, va_end;
3801 	int64_t sp_refchg;
3802 
3803 	/* This CTR may generate a lot of output. */
3804 	/* CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)va); */
3805 
3806 	va &= ~HPT_SP_MASK;
3807 	sva = va;
3808 	/* Get superpage */
3809 	pa = VM_PAGE_TO_PHYS(m) & ~HPT_SP_MASK;
3810 	m = PHYS_TO_VM_PAGE(pa);
3811 
3812 	PMAP_LOCK(pmap);
3813 
3814 	/*
3815 	 * Check if all pages meet promotion criteria.
3816 	 *
3817 	 * XXX In some cases the loop below may be executed for each or most
3818 	 * of the entered pages of a superpage, which can be expensive
3819 	 * (although it was not profiled) and need some optimization.
3820 	 *
3821 	 * Some cases where this seems to happen are:
3822 	 * - When a superpage is first entered read-only and later becomes
3823 	 *   read-write.
3824 	 * - When some of the superpage's virtual addresses map to previously
3825 	 *   wired/cached pages while others map to pages allocated from a
3826 	 *   different physical address range. A common scenario where this
3827 	 *   happens is when mmap'ing a file that is already present in FS
3828 	 *   block cache and doesn't fill a superpage.
3829 	 */
3830 	first = pvo = moea64_pvo_find_va(pmap, sva);
3831 	for (pa_end = pa + HPT_SP_SIZE;
3832 	    pa < pa_end; pa += PAGE_SIZE, va += PAGE_SIZE) {
3833 		if (pvo == NULL || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
3834 			CTR3(KTR_PMAP,
3835 			    "%s: NULL or dead PVO: pmap=%p, va=%#jx",
3836 			    __func__, pmap, (uintmax_t)va);
3837 			goto error;
3838 		}
3839 		if (PVO_PADDR(pvo) != pa) {
3840 			CTR5(KTR_PMAP, "%s: PAs don't match: "
3841 			    "pmap=%p, va=%#jx, pvo_pa=%#jx, exp_pa=%#jx",
3842 			    __func__, pmap, (uintmax_t)va,
3843 			    (uintmax_t)PVO_PADDR(pvo), (uintmax_t)pa);
3844 			atomic_add_long(&sp_p_fail_pa, 1);
3845 			goto error;
3846 		}
3847 		if ((first->pvo_vaddr & PVO_FLAGS_PROMOTE) !=
3848 		    (pvo->pvo_vaddr & PVO_FLAGS_PROMOTE)) {
3849 			CTR5(KTR_PMAP, "%s: PVO flags don't match: "
3850 			    "pmap=%p, va=%#jx, pvo_flags=%#jx, exp_flags=%#jx",
3851 			    __func__, pmap, (uintmax_t)va,
3852 			    (uintmax_t)(pvo->pvo_vaddr & PVO_FLAGS_PROMOTE),
3853 			    (uintmax_t)(first->pvo_vaddr & PVO_FLAGS_PROMOTE));
3854 			atomic_add_long(&sp_p_fail_flags, 1);
3855 			goto error;
3856 		}
3857 		if (first->pvo_pte.prot != pvo->pvo_pte.prot) {
3858 			CTR5(KTR_PMAP, "%s: PVO protections don't match: "
3859 			    "pmap=%p, va=%#jx, pvo_prot=%#x, exp_prot=%#x",
3860 			    __func__, pmap, (uintmax_t)va,
3861 			    pvo->pvo_pte.prot, first->pvo_pte.prot);
3862 			atomic_add_long(&sp_p_fail_prot, 1);
3863 			goto error;
3864 		}
3865 		if ((first->pvo_pte.pa & LPTE_WIMG) !=
3866 		    (pvo->pvo_pte.pa & LPTE_WIMG)) {
3867 			CTR5(KTR_PMAP, "%s: WIMG bits don't match: "
3868 			    "pmap=%p, va=%#jx, pvo_wimg=%#jx, exp_wimg=%#jx",
3869 			    __func__, pmap, (uintmax_t)va,
3870 			    (uintmax_t)(pvo->pvo_pte.pa & LPTE_WIMG),
3871 			    (uintmax_t)(first->pvo_pte.pa & LPTE_WIMG));
3872 			atomic_add_long(&sp_p_fail_wimg, 1);
3873 			goto error;
3874 		}
3875 
3876 		pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo);
3877 	}
3878 
3879 	/* All OK, promote. */
3880 
3881 	/*
3882 	 * Handle superpage REF/CHG bits. If REF or CHG is set in
3883 	 * any page, then it must be set in the superpage.
3884 	 *
3885 	 * Instead of querying each page, we take advantage of two facts:
3886 	 * 1- If a page is being promoted, it was referenced.
3887 	 * 2- If promoted pages are writable, they were modified.
3888 	 */
3889 	sp_refchg = LPTE_REF |
3890 	    ((first->pvo_pte.prot & VM_PROT_WRITE) != 0 ? LPTE_CHG : 0);
3891 
3892 	/* Promote pages */
3893 
3894 	for (pvo = first, va_end = PVO_VADDR(pvo) + HPT_SP_SIZE;
3895 	    pvo != NULL && PVO_VADDR(pvo) < va_end;
3896 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo)) {
3897 		pvo->pvo_pte.pa &= ADDR_POFF | ~HPT_SP_MASK;
3898 		pvo->pvo_pte.pa |= LPTE_LP_4K_16M;
3899 		pvo->pvo_vaddr |= PVO_LARGE;
3900 	}
3901 	moea64_pte_replace_sp(first);
3902 
3903 	/* Send REF/CHG bits to VM */
3904 	moea64_sp_refchg_process(first, m, sp_refchg, first->pvo_pte.prot);
3905 
3906 	/* Use first page to cache REF/CHG bits */
3907 	atomic_set_32(&m->md.mdpg_attrs, sp_refchg | MDPG_ATTR_SP);
3908 
3909 	PMAP_UNLOCK(pmap);
3910 
3911 	atomic_add_long(&sp_mappings, 1);
3912 	atomic_add_long(&sp_promotions, 1);
3913 	CTR3(KTR_PMAP, "%s: success for va %#jx in pmap %p",
3914 	    __func__, (uintmax_t)sva, pmap);
3915 	return;
3916 
3917 error:
3918 	atomic_add_long(&sp_p_failures, 1);
3919 	PMAP_UNLOCK(pmap);
3920 }
3921 
3922 static void
3923 moea64_sp_demote_aligned(struct pvo_entry *sp)
3924 {
3925 	struct pvo_entry *pvo;
3926 	vm_offset_t va, va_end;
3927 	vm_paddr_t pa;
3928 	vm_page_t m;
3929 	pmap_t pmap;
3930 	int64_t refchg;
3931 
3932 	CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)PVO_VADDR(sp));
3933 
3934 	pmap = sp->pvo_pmap;
3935 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
3936 
3937 	pvo = sp;
3938 
3939 	/* Demote pages */
3940 
3941 	va = PVO_VADDR(pvo);
3942 	pa = PVO_PADDR(pvo);
3943 	m = PHYS_TO_VM_PAGE(pa);
3944 
3945 	for (pvo = sp, va_end = va + HPT_SP_SIZE;
3946 	    pvo != NULL && PVO_VADDR(pvo) < va_end;
3947 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo),
3948 	    va += PAGE_SIZE, pa += PAGE_SIZE) {
3949 		KASSERT(pvo && PVO_VADDR(pvo) == va,
3950 		    ("%s: missing PVO for va %#jx", __func__, (uintmax_t)va));
3951 
3952 		pvo->pvo_vaddr &= ~PVO_LARGE;
3953 		pvo->pvo_pte.pa &= ~LPTE_RPGN;
3954 		pvo->pvo_pte.pa |= pa;
3955 
3956 	}
3957 	refchg = moea64_pte_replace_sp(sp);
3958 
3959 	/*
3960 	 * Clear SP flag
3961 	 *
3962 	 * XXX It is possible that another pmap has this page mapped as
3963 	 *     part of a superpage, but as the SP flag is used only for
3964 	 *     caching SP REF/CHG bits, that will be queried if not set
3965 	 *     in cache, it should be ok to clear it here.
3966 	 */
3967 	atomic_clear_32(&m->md.mdpg_attrs, MDPG_ATTR_SP);
3968 
3969 	/*
3970 	 * Handle superpage REF/CHG bits. A bit set in the superpage
3971 	 * means all pages should consider it set.
3972 	 */
3973 	moea64_sp_refchg_process(sp, m, refchg, sp->pvo_pte.prot);
3974 
3975 	atomic_add_long(&sp_demotions, 1);
3976 	CTR3(KTR_PMAP, "%s: success for va %#jx in pmap %p",
3977 	    __func__, (uintmax_t)PVO_VADDR(sp), pmap);
3978 }
3979 
3980 static void
3981 moea64_sp_demote(struct pvo_entry *pvo)
3982 {
3983 	PMAP_LOCK_ASSERT(pvo->pvo_pmap, MA_OWNED);
3984 
3985 	if ((PVO_VADDR(pvo) & HPT_SP_MASK) != 0) {
3986 		pvo = moea64_pvo_find_va(pvo->pvo_pmap,
3987 		    PVO_VADDR(pvo) & ~HPT_SP_MASK);
3988 		KASSERT(pvo != NULL, ("%s: missing PVO for va %#jx",
3989 		     __func__, (uintmax_t)(PVO_VADDR(pvo) & ~HPT_SP_MASK)));
3990 	}
3991 	moea64_sp_demote_aligned(pvo);
3992 }
3993 
3994 static struct pvo_entry *
3995 moea64_sp_unwire(struct pvo_entry *sp)
3996 {
3997 	struct pvo_entry *pvo, *prev;
3998 	vm_offset_t eva;
3999 	pmap_t pm;
4000 	int64_t ret, refchg;
4001 
4002 	CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)PVO_VADDR(sp));
4003 
4004 	pm = sp->pvo_pmap;
4005 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
4006 
4007 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4008 	refchg = 0;
4009 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4010 	    prev = pvo, pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
4011 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
4012 			panic("%s: pvo %p is missing PVO_WIRED",
4013 			    __func__, pvo);
4014 		pvo->pvo_vaddr &= ~PVO_WIRED;
4015 
4016 		ret = moea64_pte_replace(pvo, 0 /* No invalidation */);
4017 		if (ret < 0)
4018 			refchg |= LPTE_CHG;
4019 		else
4020 			refchg |= ret;
4021 
4022 		pm->pm_stats.wired_count--;
4023 	}
4024 
4025 	/* Send REF/CHG bits to VM */
4026 	moea64_sp_refchg_process(sp, PHYS_TO_VM_PAGE(PVO_PADDR(sp)),
4027 	    refchg, sp->pvo_pte.prot);
4028 
4029 	return (prev);
4030 }
4031 
4032 static struct pvo_entry *
4033 moea64_sp_protect(struct pvo_entry *sp, vm_prot_t prot)
4034 {
4035 	struct pvo_entry *pvo, *prev;
4036 	vm_offset_t eva;
4037 	pmap_t pm;
4038 	vm_page_t m, m_end;
4039 	int64_t ret, refchg;
4040 	vm_prot_t oldprot;
4041 
4042 	CTR3(KTR_PMAP, "%s: va=%#jx, prot=%x",
4043 	    __func__, (uintmax_t)PVO_VADDR(sp), prot);
4044 
4045 	pm = sp->pvo_pmap;
4046 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
4047 
4048 	oldprot = sp->pvo_pte.prot;
4049 	m = PHYS_TO_VM_PAGE(PVO_PADDR(sp));
4050 	KASSERT(m != NULL, ("%s: missing vm page for pa %#jx",
4051 	    __func__, (uintmax_t)PVO_PADDR(sp)));
4052 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4053 	refchg = 0;
4054 
4055 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4056 	    prev = pvo, pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
4057 		pvo->pvo_pte.prot = prot;
4058 		/*
4059 		 * If the PVO is in the page table, update mapping
4060 		 */
4061 		ret = moea64_pte_replace(pvo, MOEA64_PTE_PROT_UPDATE);
4062 		if (ret < 0)
4063 			refchg |= LPTE_CHG;
4064 		else
4065 			refchg |= ret;
4066 	}
4067 
4068 	/* Send REF/CHG bits to VM */
4069 	moea64_sp_refchg_process(sp, m, refchg, oldprot);
4070 
4071 	/* Handle pages that became executable */
4072 	if ((m->a.flags & PGA_EXECUTABLE) == 0 &&
4073 	    (sp->pvo_pte.pa & (LPTE_I | LPTE_G | LPTE_NOEXEC)) == 0) {
4074 		if ((m->oflags & VPO_UNMANAGED) == 0)
4075 			for (m_end = &m[HPT_SP_PAGES]; m < m_end; m++)
4076 				vm_page_aflag_set(m, PGA_EXECUTABLE);
4077 		moea64_syncicache(pm, PVO_VADDR(sp), PVO_PADDR(sp),
4078 		    HPT_SP_SIZE);
4079 	}
4080 
4081 	return (prev);
4082 }
4083 
4084 static struct pvo_entry *
4085 moea64_sp_remove(struct pvo_entry *sp, struct pvo_dlist *tofree)
4086 {
4087 	struct pvo_entry *pvo, *tpvo;
4088 	vm_offset_t eva;
4089 	pmap_t pm;
4090 
4091 	CTR2(KTR_PMAP, "%s: va=%#jx", __func__, (uintmax_t)PVO_VADDR(sp));
4092 
4093 	pm = sp->pvo_pmap;
4094 	PMAP_LOCK_ASSERT(pm, MA_OWNED);
4095 
4096 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4097 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
4098 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
4099 
4100 		/*
4101 		 * For locking reasons, remove this from the page table and
4102 		 * pmap, but save delinking from the vm_page for a second
4103 		 * pass
4104 		 */
4105 		moea64_pvo_remove_from_pmap(pvo);
4106 		SLIST_INSERT_HEAD(tofree, pvo, pvo_dlink);
4107 	}
4108 
4109 	/*
4110 	 * Clear SP bit
4111 	 *
4112 	 * XXX See comment in moea64_sp_demote_aligned() for why it's
4113 	 *     ok to always clear the SP bit on remove/demote.
4114 	 */
4115 	atomic_clear_32(&PHYS_TO_VM_PAGE(PVO_PADDR(sp))->md.mdpg_attrs,
4116 	    MDPG_ATTR_SP);
4117 
4118 	return (tpvo);
4119 }
4120 
4121 static int64_t
4122 moea64_sp_query_locked(struct pvo_entry *pvo, uint64_t ptebit)
4123 {
4124 	int64_t refchg, ret;
4125 	vm_offset_t eva;
4126 	vm_page_t m;
4127 	pmap_t pmap;
4128 	struct pvo_entry *sp;
4129 
4130 	pmap = pvo->pvo_pmap;
4131 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
4132 
4133 	/* Get first SP PVO */
4134 	if ((PVO_VADDR(pvo) & HPT_SP_MASK) != 0) {
4135 		sp = moea64_pvo_find_va(pmap, PVO_VADDR(pvo) & ~HPT_SP_MASK);
4136 		KASSERT(sp != NULL, ("%s: missing PVO for va %#jx",
4137 		     __func__, (uintmax_t)(PVO_VADDR(pvo) & ~HPT_SP_MASK)));
4138 	} else
4139 		sp = pvo;
4140 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4141 
4142 	refchg = 0;
4143 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4144 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo)) {
4145 		ret = moea64_pte_synch(pvo);
4146 		if (ret > 0) {
4147 			refchg |= ret & (LPTE_CHG | LPTE_REF);
4148 			if ((refchg & ptebit) != 0)
4149 				break;
4150 		}
4151 	}
4152 
4153 	/* Save results */
4154 	if (refchg != 0) {
4155 		m = PHYS_TO_VM_PAGE(PVO_PADDR(sp));
4156 		atomic_set_32(&m->md.mdpg_attrs, refchg | MDPG_ATTR_SP);
4157 	}
4158 
4159 	return (refchg);
4160 }
4161 
4162 static int64_t
4163 moea64_sp_query(struct pvo_entry *pvo, uint64_t ptebit)
4164 {
4165 	int64_t refchg;
4166 	pmap_t pmap;
4167 
4168 	pmap = pvo->pvo_pmap;
4169 	PMAP_LOCK(pmap);
4170 
4171 	/*
4172 	 * Check if SP was demoted/removed before pmap lock was acquired.
4173 	 */
4174 	if (!PVO_IS_SP(pvo) || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
4175 		CTR2(KTR_PMAP, "%s: demoted/removed: pa=%#jx",
4176 		    __func__, (uintmax_t)PVO_PADDR(pvo));
4177 		PMAP_UNLOCK(pmap);
4178 		return (-1);
4179 	}
4180 
4181 	refchg = moea64_sp_query_locked(pvo, ptebit);
4182 	PMAP_UNLOCK(pmap);
4183 
4184 	CTR4(KTR_PMAP, "%s: va=%#jx, pa=%#jx: refchg=%#jx",
4185 	    __func__, (uintmax_t)PVO_VADDR(pvo),
4186 	    (uintmax_t)PVO_PADDR(pvo), (uintmax_t)refchg);
4187 
4188 	return (refchg);
4189 }
4190 
4191 static int64_t
4192 moea64_sp_pvo_clear(struct pvo_entry *pvo, uint64_t ptebit)
4193 {
4194 	int64_t refchg, ret;
4195 	pmap_t pmap;
4196 	struct pvo_entry *sp;
4197 	vm_offset_t eva;
4198 	vm_page_t m;
4199 
4200 	pmap = pvo->pvo_pmap;
4201 	PMAP_LOCK(pmap);
4202 
4203 	/*
4204 	 * Check if SP was demoted/removed before pmap lock was acquired.
4205 	 */
4206 	if (!PVO_IS_SP(pvo) || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
4207 		CTR2(KTR_PMAP, "%s: demoted/removed: pa=%#jx",
4208 		    __func__, (uintmax_t)PVO_PADDR(pvo));
4209 		PMAP_UNLOCK(pmap);
4210 		return (-1);
4211 	}
4212 
4213 	/* Get first SP PVO */
4214 	if ((PVO_VADDR(pvo) & HPT_SP_MASK) != 0) {
4215 		sp = moea64_pvo_find_va(pmap, PVO_VADDR(pvo) & ~HPT_SP_MASK);
4216 		KASSERT(sp != NULL, ("%s: missing PVO for va %#jx",
4217 		     __func__, (uintmax_t)(PVO_VADDR(pvo) & ~HPT_SP_MASK)));
4218 	} else
4219 		sp = pvo;
4220 	eva = PVO_VADDR(sp) + HPT_SP_SIZE;
4221 
4222 	refchg = 0;
4223 	for (pvo = sp; pvo != NULL && PVO_VADDR(pvo) < eva;
4224 	    pvo = RB_NEXT(pvo_tree, &pmap->pmap_pvo, pvo)) {
4225 		ret = moea64_pte_clear(pvo, ptebit);
4226 		if (ret > 0)
4227 			refchg |= ret & (LPTE_CHG | LPTE_REF);
4228 	}
4229 
4230 	m = PHYS_TO_VM_PAGE(PVO_PADDR(sp));
4231 	atomic_clear_32(&m->md.mdpg_attrs, ptebit);
4232 	PMAP_UNLOCK(pmap);
4233 
4234 	CTR4(KTR_PMAP, "%s: va=%#jx, pa=%#jx: refchg=%#jx",
4235 	    __func__, (uintmax_t)PVO_VADDR(sp),
4236 	    (uintmax_t)PVO_PADDR(sp), (uintmax_t)refchg);
4237 
4238 	return (refchg);
4239 }
4240 
4241 static int64_t
4242 moea64_sp_clear(struct pvo_entry *pvo, vm_page_t m, uint64_t ptebit)
4243 {
4244 	int64_t count, ret;
4245 	pmap_t pmap;
4246 
4247 	count = 0;
4248 	pmap = pvo->pvo_pmap;
4249 
4250 	/*
4251 	 * Since this reference bit is shared by 4096 4KB pages, it
4252 	 * should not be cleared every time it is tested. Apply a
4253 	 * simple "hash" function on the physical page number, the
4254 	 * virtual superpage number, and the pmap address to select
4255 	 * one 4KB page out of the 4096 on which testing the
4256 	 * reference bit will result in clearing that reference bit.
4257 	 * This function is designed to avoid the selection of the
4258 	 * same 4KB page for every 16MB page mapping.
4259 	 *
4260 	 * Always leave the reference bit of a wired mapping set, as
4261 	 * the current state of its reference bit won't affect page
4262 	 * replacement.
4263 	 */
4264 	if (ptebit == LPTE_REF && (((VM_PAGE_TO_PHYS(m) >> PAGE_SHIFT) ^
4265 	    (PVO_VADDR(pvo) >> HPT_SP_SHIFT) ^ (uintptr_t)pmap) &
4266 	    (HPT_SP_PAGES - 1)) == 0 && (pvo->pvo_vaddr & PVO_WIRED) == 0) {
4267 		if ((ret = moea64_sp_pvo_clear(pvo, ptebit)) == -1)
4268 			return (-1);
4269 
4270 		if ((ret & ptebit) != 0)
4271 			count++;
4272 
4273 	/*
4274 	 * If this page was not selected by the hash function, then assume
4275 	 * its REF bit was set.
4276 	 */
4277 	} else if (ptebit == LPTE_REF) {
4278 		count++;
4279 
4280 	/*
4281 	 * To clear the CHG bit of a single SP page, first it must be demoted.
4282 	 * But if no CHG bit is set, no bit clear and thus no SP demotion is
4283 	 * needed.
4284 	 */
4285 	} else {
4286 		CTR4(KTR_PMAP, "%s: ptebit=%#jx, va=%#jx, pa=%#jx",
4287 		    __func__, (uintmax_t)ptebit, (uintmax_t)PVO_VADDR(pvo),
4288 		    (uintmax_t)PVO_PADDR(pvo));
4289 
4290 		PMAP_LOCK(pmap);
4291 
4292 		/*
4293 		 * Make sure SP wasn't demoted/removed before pmap lock
4294 		 * was acquired.
4295 		 */
4296 		if (!PVO_IS_SP(pvo) || (pvo->pvo_vaddr & PVO_DEAD) != 0) {
4297 			CTR2(KTR_PMAP, "%s: demoted/removed: pa=%#jx",
4298 			    __func__, (uintmax_t)PVO_PADDR(pvo));
4299 			PMAP_UNLOCK(pmap);
4300 			return (-1);
4301 		}
4302 
4303 		ret = moea64_sp_query_locked(pvo, ptebit);
4304 		if ((ret & ptebit) != 0)
4305 			count++;
4306 		else {
4307 			PMAP_UNLOCK(pmap);
4308 			return (0);
4309 		}
4310 
4311 		moea64_sp_demote(pvo);
4312 		moea64_pte_clear(pvo, ptebit);
4313 
4314 		/*
4315 		 * Write protect the mapping to a single page so that a
4316 		 * subsequent write access may repromote.
4317 		 */
4318 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
4319 			moea64_pvo_protect(pmap, pvo,
4320 			    pvo->pvo_pte.prot & ~VM_PROT_WRITE);
4321 
4322 		PMAP_UNLOCK(pmap);
4323 	}
4324 
4325 	return (count);
4326 }
4327