xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision 5ffd83dbcc34f10e07f6d3e968ae6365869615f4)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD AND BSD-4-Clause
3  *
4  * Copyright (c) 2001 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Matt Thomas <matt@3am-software.com> of Allegro Networks, Inc.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 /*-
32  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
33  * Copyright (C) 1995, 1996 TooLs GmbH.
34  * All rights reserved.
35  *
36  * Redistribution and use in source and binary forms, with or without
37  * modification, are permitted provided that the following conditions
38  * are met:
39  * 1. Redistributions of source code must retain the above copyright
40  *    notice, this list of conditions and the following disclaimer.
41  * 2. Redistributions in binary form must reproduce the above copyright
42  *    notice, this list of conditions and the following disclaimer in the
43  *    documentation and/or other materials provided with the distribution.
44  * 3. All advertising materials mentioning features or use of this software
45  *    must display the following acknowledgement:
46  *	This product includes software developed by TooLs GmbH.
47  * 4. The name of TooLs GmbH may not be used to endorse or promote products
48  *    derived from this software without specific prior written permission.
49  *
50  * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
51  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
52  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
53  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
54  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
55  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
56  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
57  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
58  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
59  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
60  *
61  * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $
62  */
63 /*-
64  * Copyright (C) 2001 Benno Rice.
65  * All rights reserved.
66  *
67  * Redistribution and use in source and binary forms, with or without
68  * modification, are permitted provided that the following conditions
69  * are met:
70  * 1. Redistributions of source code must retain the above copyright
71  *    notice, this list of conditions and the following disclaimer.
72  * 2. Redistributions in binary form must reproduce the above copyright
73  *    notice, this list of conditions and the following disclaimer in the
74  *    documentation and/or other materials provided with the distribution.
75  *
76  * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR
77  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
78  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
79  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
80  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
81  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
82  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
83  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
84  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
85  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
86  */
87 
88 #include <sys/cdefs.h>
89 __FBSDID("$FreeBSD$");
90 
91 /*
92  * Manages physical address maps.
93  *
94  * Since the information managed by this module is also stored by the
95  * logical address mapping module, this module may throw away valid virtual
96  * to physical mappings at almost any time.  However, invalidations of
97  * mappings must be done as requested.
98  *
99  * In order to cope with hardware architectures which make virtual to
100  * physical map invalidates expensive, this module may delay invalidate
101  * reduced protection operations until such time as they are actually
102  * necessary.  This module is given full information as to which processors
103  * are currently using which maps, and to when physical maps must be made
104  * correct.
105  */
106 
107 #include "opt_kstack_pages.h"
108 
109 #include <sys/param.h>
110 #include <sys/kernel.h>
111 #include <sys/conf.h>
112 #include <sys/queue.h>
113 #include <sys/cpuset.h>
114 #include <sys/kerneldump.h>
115 #include <sys/ktr.h>
116 #include <sys/lock.h>
117 #include <sys/msgbuf.h>
118 #include <sys/mutex.h>
119 #include <sys/proc.h>
120 #include <sys/rwlock.h>
121 #include <sys/sched.h>
122 #include <sys/sysctl.h>
123 #include <sys/systm.h>
124 #include <sys/vmmeter.h>
125 
126 #include <dev/ofw/openfirm.h>
127 
128 #include <vm/vm.h>
129 #include <vm/vm_param.h>
130 #include <vm/vm_kern.h>
131 #include <vm/vm_page.h>
132 #include <vm/vm_map.h>
133 #include <vm/vm_object.h>
134 #include <vm/vm_extern.h>
135 #include <vm/vm_page.h>
136 #include <vm/vm_phys.h>
137 #include <vm/vm_pageout.h>
138 #include <vm/uma.h>
139 
140 #include <machine/cpu.h>
141 #include <machine/platform.h>
142 #include <machine/bat.h>
143 #include <machine/frame.h>
144 #include <machine/md_var.h>
145 #include <machine/psl.h>
146 #include <machine/pte.h>
147 #include <machine/smp.h>
148 #include <machine/sr.h>
149 #include <machine/mmuvar.h>
150 #include <machine/trap.h>
151 
152 #define	MOEA_DEBUG
153 
154 #define TODO	panic("%s: not implemented", __func__);
155 
156 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
157 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
158 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
159 
160 struct ofw_map {
161 	vm_offset_t	om_va;
162 	vm_size_t	om_len;
163 	vm_offset_t	om_pa;
164 	u_int		om_mode;
165 };
166 
167 extern unsigned char _etext[];
168 extern unsigned char _end[];
169 
170 /*
171  * Map of physical memory regions.
172  */
173 static struct	mem_region *regions;
174 static struct	mem_region *pregions;
175 static u_int    phys_avail_count;
176 static int	regions_sz, pregions_sz;
177 static struct	ofw_map *translations;
178 
179 /*
180  * Lock for the pteg and pvo tables.
181  */
182 struct mtx	moea_table_mutex;
183 struct mtx	moea_vsid_mutex;
184 
185 /* tlbie instruction synchronization */
186 static struct mtx tlbie_mtx;
187 
188 /*
189  * PTEG data.
190  */
191 static struct	pteg *moea_pteg_table;
192 u_int		moea_pteg_count;
193 u_int		moea_pteg_mask;
194 
195 /*
196  * PVO data.
197  */
198 struct	pvo_head *moea_pvo_table;		/* pvo entries by pteg index */
199 struct	pvo_head moea_pvo_kunmanaged =
200     LIST_HEAD_INITIALIZER(moea_pvo_kunmanaged);	/* list of unmanaged pages */
201 
202 static struct rwlock_padalign pvh_global_lock;
203 
204 uma_zone_t	moea_upvo_zone;	/* zone for pvo entries for unmanaged pages */
205 uma_zone_t	moea_mpvo_zone;	/* zone for pvo entries for managed pages */
206 
207 #define	BPVO_POOL_SIZE	32768
208 static struct	pvo_entry *moea_bpvo_pool;
209 static int	moea_bpvo_pool_index = 0;
210 
211 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
212 static u_int	moea_vsid_bitmap[NPMAPS / VSID_NBPW];
213 
214 static boolean_t moea_initialized = FALSE;
215 
216 /*
217  * Statistics.
218  */
219 u_int	moea_pte_valid = 0;
220 u_int	moea_pte_overflow = 0;
221 u_int	moea_pte_replacements = 0;
222 u_int	moea_pvo_entries = 0;
223 u_int	moea_pvo_enter_calls = 0;
224 u_int	moea_pvo_remove_calls = 0;
225 u_int	moea_pte_spills = 0;
226 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_valid, CTLFLAG_RD, &moea_pte_valid,
227     0, "");
228 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_overflow, CTLFLAG_RD,
229     &moea_pte_overflow, 0, "");
230 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_replacements, CTLFLAG_RD,
231     &moea_pte_replacements, 0, "");
232 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_entries, CTLFLAG_RD, &moea_pvo_entries,
233     0, "");
234 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_enter_calls, CTLFLAG_RD,
235     &moea_pvo_enter_calls, 0, "");
236 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_remove_calls, CTLFLAG_RD,
237     &moea_pvo_remove_calls, 0, "");
238 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_spills, CTLFLAG_RD,
239     &moea_pte_spills, 0, "");
240 
241 /*
242  * Allocate physical memory for use in moea_bootstrap.
243  */
244 static vm_offset_t	moea_bootstrap_alloc(vm_size_t, u_int);
245 
246 /*
247  * PTE calls.
248  */
249 static int		moea_pte_insert(u_int, struct pte *);
250 
251 /*
252  * PVO calls.
253  */
254 static int	moea_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *,
255 		    vm_offset_t, vm_paddr_t, u_int, int);
256 static void	moea_pvo_remove(struct pvo_entry *, int);
257 static struct	pvo_entry *moea_pvo_find_va(pmap_t, vm_offset_t, int *);
258 static struct	pte *moea_pvo_to_pte(const struct pvo_entry *, int);
259 
260 /*
261  * Utility routines.
262  */
263 static int		moea_enter_locked(pmap_t, vm_offset_t, vm_page_t,
264 			    vm_prot_t, u_int, int8_t);
265 static void		moea_syncicache(vm_paddr_t, vm_size_t);
266 static boolean_t	moea_query_bit(vm_page_t, int);
267 static u_int		moea_clear_bit(vm_page_t, int);
268 static void		moea_kremove(vm_offset_t);
269 int		moea_pte_spill(vm_offset_t);
270 
271 /*
272  * Kernel MMU interface
273  */
274 void moea_clear_modify(vm_page_t);
275 void moea_copy_page(vm_page_t, vm_page_t);
276 void moea_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
277     vm_page_t *mb, vm_offset_t b_offset, int xfersize);
278 int moea_enter(pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int,
279     int8_t);
280 void moea_enter_object(pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
281     vm_prot_t);
282 void moea_enter_quick(pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
283 vm_paddr_t moea_extract(pmap_t, vm_offset_t);
284 vm_page_t moea_extract_and_hold(pmap_t, vm_offset_t, vm_prot_t);
285 void moea_init(void);
286 boolean_t moea_is_modified(vm_page_t);
287 boolean_t moea_is_prefaultable(pmap_t, vm_offset_t);
288 boolean_t moea_is_referenced(vm_page_t);
289 int moea_ts_referenced(vm_page_t);
290 vm_offset_t moea_map(vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
291 boolean_t moea_page_exists_quick(pmap_t, vm_page_t);
292 void moea_page_init(vm_page_t);
293 int moea_page_wired_mappings(vm_page_t);
294 int moea_pinit(pmap_t);
295 void moea_pinit0(pmap_t);
296 void moea_protect(pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
297 void moea_qenter(vm_offset_t, vm_page_t *, int);
298 void moea_qremove(vm_offset_t, int);
299 void moea_release(pmap_t);
300 void moea_remove(pmap_t, vm_offset_t, vm_offset_t);
301 void moea_remove_all(vm_page_t);
302 void moea_remove_write(vm_page_t);
303 void moea_unwire(pmap_t, vm_offset_t, vm_offset_t);
304 void moea_zero_page(vm_page_t);
305 void moea_zero_page_area(vm_page_t, int, int);
306 void moea_activate(struct thread *);
307 void moea_deactivate(struct thread *);
308 void moea_cpu_bootstrap(int);
309 void moea_bootstrap(vm_offset_t, vm_offset_t);
310 void *moea_mapdev(vm_paddr_t, vm_size_t);
311 void *moea_mapdev_attr(vm_paddr_t, vm_size_t, vm_memattr_t);
312 void moea_unmapdev(vm_offset_t, vm_size_t);
313 vm_paddr_t moea_kextract(vm_offset_t);
314 void moea_kenter_attr(vm_offset_t, vm_paddr_t, vm_memattr_t);
315 void moea_kenter(vm_offset_t, vm_paddr_t);
316 void moea_page_set_memattr(vm_page_t m, vm_memattr_t ma);
317 boolean_t moea_dev_direct_mapped(vm_paddr_t, vm_size_t);
318 static void moea_sync_icache(pmap_t, vm_offset_t, vm_size_t);
319 void moea_dumpsys_map(vm_paddr_t pa, size_t sz, void **va);
320 void moea_scan_init(void);
321 vm_offset_t moea_quick_enter_page(vm_page_t m);
322 void moea_quick_remove_page(vm_offset_t addr);
323 boolean_t moea_page_is_mapped(vm_page_t m);
324 static int moea_map_user_ptr(pmap_t pm,
325     volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen);
326 static int moea_decode_kernel_ptr(vm_offset_t addr,
327     int *is_user, vm_offset_t *decoded_addr);
328 
329 
330 static struct pmap_funcs moea_methods = {
331 	.clear_modify = moea_clear_modify,
332 	.copy_page = moea_copy_page,
333 	.copy_pages = moea_copy_pages,
334 	.enter = moea_enter,
335 	.enter_object = moea_enter_object,
336 	.enter_quick = moea_enter_quick,
337 	.extract = moea_extract,
338 	.extract_and_hold = moea_extract_and_hold,
339 	.init = moea_init,
340 	.is_modified = moea_is_modified,
341 	.is_prefaultable = moea_is_prefaultable,
342 	.is_referenced = moea_is_referenced,
343 	.ts_referenced = moea_ts_referenced,
344 	.map =      		moea_map,
345 	.page_exists_quick = moea_page_exists_quick,
346 	.page_init = moea_page_init,
347 	.page_wired_mappings = moea_page_wired_mappings,
348 	.pinit = moea_pinit,
349 	.pinit0 = moea_pinit0,
350 	.protect = moea_protect,
351 	.qenter = moea_qenter,
352 	.qremove = moea_qremove,
353 	.release = moea_release,
354 	.remove = moea_remove,
355 	.remove_all =       	moea_remove_all,
356 	.remove_write = moea_remove_write,
357 	.sync_icache = moea_sync_icache,
358 	.unwire = moea_unwire,
359 	.zero_page =        	moea_zero_page,
360 	.zero_page_area = moea_zero_page_area,
361 	.activate = moea_activate,
362 	.deactivate =       	moea_deactivate,
363 	.page_set_memattr = moea_page_set_memattr,
364 	.quick_enter_page =  moea_quick_enter_page,
365 	.quick_remove_page =  moea_quick_remove_page,
366 	.page_is_mapped = moea_page_is_mapped,
367 
368 	/* Internal interfaces */
369 	.bootstrap =        	moea_bootstrap,
370 	.cpu_bootstrap =    	moea_cpu_bootstrap,
371 	.mapdev_attr = moea_mapdev_attr,
372 	.mapdev = moea_mapdev,
373 	.unmapdev = moea_unmapdev,
374 	.kextract = moea_kextract,
375 	.kenter = moea_kenter,
376 	.kenter_attr = moea_kenter_attr,
377 	.dev_direct_mapped = moea_dev_direct_mapped,
378 	.dumpsys_pa_init = moea_scan_init,
379 	.dumpsys_map_chunk = moea_dumpsys_map,
380 	.map_user_ptr = moea_map_user_ptr,
381 	.decode_kernel_ptr =  moea_decode_kernel_ptr,
382 };
383 
384 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods);
385 
386 static __inline uint32_t
387 moea_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
388 {
389 	uint32_t pte_lo;
390 	int i;
391 
392 	if (ma != VM_MEMATTR_DEFAULT) {
393 		switch (ma) {
394 		case VM_MEMATTR_UNCACHEABLE:
395 			return (PTE_I | PTE_G);
396 		case VM_MEMATTR_CACHEABLE:
397 			return (PTE_M);
398 		case VM_MEMATTR_WRITE_COMBINING:
399 		case VM_MEMATTR_WRITE_BACK:
400 		case VM_MEMATTR_PREFETCHABLE:
401 			return (PTE_I);
402 		case VM_MEMATTR_WRITE_THROUGH:
403 			return (PTE_W | PTE_M);
404 		}
405 	}
406 
407 	/*
408 	 * Assume the page is cache inhibited and access is guarded unless
409 	 * it's in our available memory array.
410 	 */
411 	pte_lo = PTE_I | PTE_G;
412 	for (i = 0; i < pregions_sz; i++) {
413 		if ((pa >= pregions[i].mr_start) &&
414 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
415 			pte_lo = PTE_M;
416 			break;
417 		}
418 	}
419 
420 	return pte_lo;
421 }
422 
423 static void
424 tlbie(vm_offset_t va)
425 {
426 
427 	mtx_lock_spin(&tlbie_mtx);
428 	__asm __volatile("ptesync");
429 	__asm __volatile("tlbie %0" :: "r"(va));
430 	__asm __volatile("eieio; tlbsync; ptesync");
431 	mtx_unlock_spin(&tlbie_mtx);
432 }
433 
434 static void
435 tlbia(void)
436 {
437 	vm_offset_t va;
438 
439 	for (va = 0; va < 0x00040000; va += 0x00001000) {
440 		__asm __volatile("tlbie %0" :: "r"(va));
441 		powerpc_sync();
442 	}
443 	__asm __volatile("tlbsync");
444 	powerpc_sync();
445 }
446 
447 static __inline int
448 va_to_sr(u_int *sr, vm_offset_t va)
449 {
450 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
451 }
452 
453 static __inline u_int
454 va_to_pteg(u_int sr, vm_offset_t addr)
455 {
456 	u_int hash;
457 
458 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
459 	    ADDR_PIDX_SHFT);
460 	return (hash & moea_pteg_mask);
461 }
462 
463 static __inline struct pvo_head *
464 vm_page_to_pvoh(vm_page_t m)
465 {
466 
467 	return (&m->md.mdpg_pvoh);
468 }
469 
470 static __inline void
471 moea_attr_clear(vm_page_t m, int ptebit)
472 {
473 
474 	rw_assert(&pvh_global_lock, RA_WLOCKED);
475 	m->md.mdpg_attrs &= ~ptebit;
476 }
477 
478 static __inline int
479 moea_attr_fetch(vm_page_t m)
480 {
481 
482 	return (m->md.mdpg_attrs);
483 }
484 
485 static __inline void
486 moea_attr_save(vm_page_t m, int ptebit)
487 {
488 
489 	rw_assert(&pvh_global_lock, RA_WLOCKED);
490 	m->md.mdpg_attrs |= ptebit;
491 }
492 
493 static __inline int
494 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
495 {
496 	if (pt->pte_hi == pvo_pt->pte_hi)
497 		return (1);
498 
499 	return (0);
500 }
501 
502 static __inline int
503 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
504 {
505 	return (pt->pte_hi & ~PTE_VALID) ==
506 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
507 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
508 }
509 
510 static __inline void
511 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
512 {
513 
514 	mtx_assert(&moea_table_mutex, MA_OWNED);
515 
516 	/*
517 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
518 	 * set when the real pte is set in memory.
519 	 *
520 	 * Note: Don't set the valid bit for correct operation of tlb update.
521 	 */
522 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
523 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
524 	pt->pte_lo = pte_lo;
525 }
526 
527 static __inline void
528 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
529 {
530 
531 	mtx_assert(&moea_table_mutex, MA_OWNED);
532 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
533 }
534 
535 static __inline void
536 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
537 {
538 
539 	mtx_assert(&moea_table_mutex, MA_OWNED);
540 
541 	/*
542 	 * As shown in Section 7.6.3.2.3
543 	 */
544 	pt->pte_lo &= ~ptebit;
545 	tlbie(va);
546 }
547 
548 static __inline void
549 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
550 {
551 
552 	mtx_assert(&moea_table_mutex, MA_OWNED);
553 	pvo_pt->pte_hi |= PTE_VALID;
554 
555 	/*
556 	 * Update the PTE as defined in section 7.6.3.1.
557 	 * Note that the REF/CHG bits are from pvo_pt and thus should have
558 	 * been saved so this routine can restore them (if desired).
559 	 */
560 	pt->pte_lo = pvo_pt->pte_lo;
561 	powerpc_sync();
562 	pt->pte_hi = pvo_pt->pte_hi;
563 	powerpc_sync();
564 	moea_pte_valid++;
565 }
566 
567 static __inline void
568 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
569 {
570 
571 	mtx_assert(&moea_table_mutex, MA_OWNED);
572 	pvo_pt->pte_hi &= ~PTE_VALID;
573 
574 	/*
575 	 * Force the reg & chg bits back into the PTEs.
576 	 */
577 	powerpc_sync();
578 
579 	/*
580 	 * Invalidate the pte.
581 	 */
582 	pt->pte_hi &= ~PTE_VALID;
583 
584 	tlbie(va);
585 
586 	/*
587 	 * Save the reg & chg bits.
588 	 */
589 	moea_pte_synch(pt, pvo_pt);
590 	moea_pte_valid--;
591 }
592 
593 static __inline void
594 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
595 {
596 
597 	/*
598 	 * Invalidate the PTE
599 	 */
600 	moea_pte_unset(pt, pvo_pt, va);
601 	moea_pte_set(pt, pvo_pt);
602 }
603 
604 /*
605  * Quick sort callout for comparing memory regions.
606  */
607 static int	om_cmp(const void *a, const void *b);
608 
609 static int
610 om_cmp(const void *a, const void *b)
611 {
612 	const struct	ofw_map *mapa;
613 	const struct	ofw_map *mapb;
614 
615 	mapa = a;
616 	mapb = b;
617 	if (mapa->om_pa < mapb->om_pa)
618 		return (-1);
619 	else if (mapa->om_pa > mapb->om_pa)
620 		return (1);
621 	else
622 		return (0);
623 }
624 
625 void
626 moea_cpu_bootstrap(int ap)
627 {
628 	u_int sdr;
629 	int i;
630 
631 	if (ap) {
632 		powerpc_sync();
633 		__asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu));
634 		__asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl));
635 		isync();
636 		__asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu));
637 		__asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl));
638 		isync();
639 	}
640 
641 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
642 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
643 	isync();
644 
645 	__asm __volatile("mtibatu 1,%0" :: "r"(0));
646 	__asm __volatile("mtdbatu 2,%0" :: "r"(0));
647 	__asm __volatile("mtibatu 2,%0" :: "r"(0));
648 	__asm __volatile("mtdbatu 3,%0" :: "r"(0));
649 	__asm __volatile("mtibatu 3,%0" :: "r"(0));
650 	isync();
651 
652 	for (i = 0; i < 16; i++)
653 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
654 	powerpc_sync();
655 
656 	sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10);
657 	__asm __volatile("mtsdr1 %0" :: "r"(sdr));
658 	isync();
659 
660 	tlbia();
661 }
662 
663 void
664 moea_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
665 {
666 	ihandle_t	mmui;
667 	phandle_t	chosen, mmu;
668 	int		sz;
669 	int		i, j;
670 	vm_size_t	size, physsz, hwphyssz;
671 	vm_offset_t	pa, va, off;
672 	void		*dpcpu;
673 	register_t	msr;
674 
675         /*
676          * Set up BAT0 to map the lowest 256 MB area
677          */
678         battable[0x0].batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
679         battable[0x0].batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
680 
681 	/*
682 	 * Map PCI memory space.
683 	 */
684 	battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
685 	battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
686 
687 	battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
688 	battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
689 
690 	battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
691 	battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
692 
693 	battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
694 	battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
695 
696 	/*
697 	 * Map obio devices.
698 	 */
699 	battable[0xf].batl = BATL(0xf0000000, BAT_I|BAT_G, BAT_PP_RW);
700 	battable[0xf].batu = BATU(0xf0000000, BAT_BL_256M, BAT_Vs);
701 
702 	/*
703 	 * Use an IBAT and a DBAT to map the bottom segment of memory
704 	 * where we are. Turn off instruction relocation temporarily
705 	 * to prevent faults while reprogramming the IBAT.
706 	 */
707 	msr = mfmsr();
708 	mtmsr(msr & ~PSL_IR);
709 	__asm (".balign 32; \n"
710 	       "mtibatu 0,%0; mtibatl 0,%1; isync; \n"
711 	       "mtdbatu 0,%0; mtdbatl 0,%1; isync"
712 	    :: "r"(battable[0].batu), "r"(battable[0].batl));
713 	mtmsr(msr);
714 
715 	/* map pci space */
716 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
717 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
718 	isync();
719 
720 	/* set global direct map flag */
721 	hw_direct_map = 1;
722 
723 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
724 	CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
725 
726 	for (i = 0; i < pregions_sz; i++) {
727 		vm_offset_t pa;
728 		vm_offset_t end;
729 
730 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
731 			pregions[i].mr_start,
732 			pregions[i].mr_start + pregions[i].mr_size,
733 			pregions[i].mr_size);
734 		/*
735 		 * Install entries into the BAT table to allow all
736 		 * of physmem to be convered by on-demand BAT entries.
737 		 * The loop will sometimes set the same battable element
738 		 * twice, but that's fine since they won't be used for
739 		 * a while yet.
740 		 */
741 		pa = pregions[i].mr_start & 0xf0000000;
742 		end = pregions[i].mr_start + pregions[i].mr_size;
743 		do {
744                         u_int n = pa >> ADDR_SR_SHFT;
745 
746 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
747 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
748 			pa += SEGMENT_LENGTH;
749 		} while (pa < end);
750 	}
751 
752 	if (PHYS_AVAIL_ENTRIES < regions_sz)
753 		panic("moea_bootstrap: phys_avail too small");
754 
755 	phys_avail_count = 0;
756 	physsz = 0;
757 	hwphyssz = 0;
758 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
759 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
760 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
761 		    regions[i].mr_start + regions[i].mr_size,
762 		    regions[i].mr_size);
763 		if (hwphyssz != 0 &&
764 		    (physsz + regions[i].mr_size) >= hwphyssz) {
765 			if (physsz < hwphyssz) {
766 				phys_avail[j] = regions[i].mr_start;
767 				phys_avail[j + 1] = regions[i].mr_start +
768 				    hwphyssz - physsz;
769 				physsz = hwphyssz;
770 				phys_avail_count++;
771 			}
772 			break;
773 		}
774 		phys_avail[j] = regions[i].mr_start;
775 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
776 		phys_avail_count++;
777 		physsz += regions[i].mr_size;
778 	}
779 
780 	/* Check for overlap with the kernel and exception vectors */
781 	for (j = 0; j < 2*phys_avail_count; j+=2) {
782 		if (phys_avail[j] < EXC_LAST)
783 			phys_avail[j] += EXC_LAST;
784 
785 		if (kernelstart >= phys_avail[j] &&
786 		    kernelstart < phys_avail[j+1]) {
787 			if (kernelend < phys_avail[j+1]) {
788 				phys_avail[2*phys_avail_count] =
789 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
790 				phys_avail[2*phys_avail_count + 1] =
791 				    phys_avail[j+1];
792 				phys_avail_count++;
793 			}
794 
795 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
796 		}
797 
798 		if (kernelend >= phys_avail[j] &&
799 		    kernelend < phys_avail[j+1]) {
800 			if (kernelstart > phys_avail[j]) {
801 				phys_avail[2*phys_avail_count] = phys_avail[j];
802 				phys_avail[2*phys_avail_count + 1] =
803 				    kernelstart & ~PAGE_MASK;
804 				phys_avail_count++;
805 			}
806 
807 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
808 		}
809 	}
810 
811 	physmem = btoc(physsz);
812 
813 	/*
814 	 * Allocate PTEG table.
815 	 */
816 #ifdef PTEGCOUNT
817 	moea_pteg_count = PTEGCOUNT;
818 #else
819 	moea_pteg_count = 0x1000;
820 
821 	while (moea_pteg_count < physmem)
822 		moea_pteg_count <<= 1;
823 
824 	moea_pteg_count >>= 1;
825 #endif /* PTEGCOUNT */
826 
827 	size = moea_pteg_count * sizeof(struct pteg);
828 	CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
829 	    size);
830 	moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
831 	CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
832 	bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
833 	moea_pteg_mask = moea_pteg_count - 1;
834 
835 	/*
836 	 * Allocate pv/overflow lists.
837 	 */
838 	size = sizeof(struct pvo_head) * moea_pteg_count;
839 	moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
840 	    PAGE_SIZE);
841 	CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
842 	for (i = 0; i < moea_pteg_count; i++)
843 		LIST_INIT(&moea_pvo_table[i]);
844 
845 	/*
846 	 * Initialize the lock that synchronizes access to the pteg and pvo
847 	 * tables.
848 	 */
849 	mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF |
850 	    MTX_RECURSE);
851 	mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF);
852 
853 	mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN);
854 
855 	/*
856 	 * Initialise the unmanaged pvo pool.
857 	 */
858 	moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
859 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
860 	moea_bpvo_pool_index = 0;
861 
862 	/*
863 	 * Make sure kernel vsid is allocated as well as VSID 0.
864 	 */
865 	moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
866 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
867 	moea_vsid_bitmap[0] |= 1;
868 
869 	/*
870 	 * Initialize the kernel pmap (which is statically allocated).
871 	 */
872 	PMAP_LOCK_INIT(kernel_pmap);
873 	for (i = 0; i < 16; i++)
874 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
875 	CPU_FILL(&kernel_pmap->pm_active);
876 	RB_INIT(&kernel_pmap->pmap_pvo);
877 
878  	/*
879 	 * Initialize the global pv list lock.
880 	 */
881 	rw_init(&pvh_global_lock, "pmap pv global");
882 
883 	/*
884 	 * Set up the Open Firmware mappings
885 	 */
886 	chosen = OF_finddevice("/chosen");
887 	if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1 &&
888 	    (mmu = OF_instance_to_package(mmui)) != -1 &&
889 	    (sz = OF_getproplen(mmu, "translations")) != -1) {
890 		translations = NULL;
891 		for (i = 0; phys_avail[i] != 0; i += 2) {
892 			if (phys_avail[i + 1] >= sz) {
893 				translations = (struct ofw_map *)phys_avail[i];
894 				break;
895 			}
896 		}
897 		if (translations == NULL)
898 			panic("moea_bootstrap: no space to copy translations");
899 		bzero(translations, sz);
900 		if (OF_getprop(mmu, "translations", translations, sz) == -1)
901 			panic("moea_bootstrap: can't get ofw translations");
902 		CTR0(KTR_PMAP, "moea_bootstrap: translations");
903 		sz /= sizeof(*translations);
904 		qsort(translations, sz, sizeof (*translations), om_cmp);
905 		for (i = 0; i < sz; i++) {
906 			CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
907 			    translations[i].om_pa, translations[i].om_va,
908 			    translations[i].om_len);
909 
910 			/*
911 			 * If the mapping is 1:1, let the RAM and device
912 			 * on-demand BAT tables take care of the translation.
913 			 */
914 			if (translations[i].om_va == translations[i].om_pa)
915 				continue;
916 
917 			/* Enter the pages */
918 			for (off = 0; off < translations[i].om_len;
919 			    off += PAGE_SIZE)
920 				moea_kenter(translations[i].om_va + off,
921 					    translations[i].om_pa + off);
922 		}
923 	}
924 
925 	/*
926 	 * Calculate the last available physical address.
927 	 */
928 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
929 		;
930 	Maxmem = powerpc_btop(phys_avail[i + 1]);
931 
932 	moea_cpu_bootstrap(0);
933 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
934 	pmap_bootstrapped++;
935 
936 	/*
937 	 * Set the start and end of kva.
938 	 */
939 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
940 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
941 
942 	/*
943 	 * Allocate a kernel stack with a guard page for thread0 and map it
944 	 * into the kernel page map.
945 	 */
946 	pa = moea_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
947 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
948 	virtual_avail = va + kstack_pages * PAGE_SIZE;
949 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
950 	thread0.td_kstack = va;
951 	thread0.td_kstack_pages = kstack_pages;
952 	for (i = 0; i < kstack_pages; i++) {
953 		moea_kenter(va, pa);
954 		pa += PAGE_SIZE;
955 		va += PAGE_SIZE;
956 	}
957 
958 	/*
959 	 * Allocate virtual address space for the message buffer.
960 	 */
961 	pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE);
962 	msgbufp = (struct msgbuf *)virtual_avail;
963 	va = virtual_avail;
964 	virtual_avail += round_page(msgbufsize);
965 	while (va < virtual_avail) {
966 		moea_kenter(va, pa);
967 		pa += PAGE_SIZE;
968 		va += PAGE_SIZE;
969 	}
970 
971 	/*
972 	 * Allocate virtual address space for the dynamic percpu area.
973 	 */
974 	pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
975 	dpcpu = (void *)virtual_avail;
976 	va = virtual_avail;
977 	virtual_avail += DPCPU_SIZE;
978 	while (va < virtual_avail) {
979 		moea_kenter(va, pa);
980 		pa += PAGE_SIZE;
981 		va += PAGE_SIZE;
982 	}
983 	dpcpu_init(dpcpu, 0);
984 }
985 
986 /*
987  * Activate a user pmap.  The pmap must be activated before it's address
988  * space can be accessed in any way.
989  */
990 void
991 moea_activate(struct thread *td)
992 {
993 	pmap_t	pm, pmr;
994 
995 	/*
996 	 * Load all the data we need up front to encourage the compiler to
997 	 * not issue any loads while we have interrupts disabled below.
998 	 */
999 	pm = &td->td_proc->p_vmspace->vm_pmap;
1000 	pmr = pm->pmap_phys;
1001 
1002 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1003 	PCPU_SET(curpmap, pmr);
1004 
1005 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1006 }
1007 
1008 void
1009 moea_deactivate(struct thread *td)
1010 {
1011 	pmap_t	pm;
1012 
1013 	pm = &td->td_proc->p_vmspace->vm_pmap;
1014 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1015 	PCPU_SET(curpmap, NULL);
1016 }
1017 
1018 void
1019 moea_unwire(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1020 {
1021 	struct	pvo_entry key, *pvo;
1022 
1023 	PMAP_LOCK(pm);
1024 	key.pvo_vaddr = sva;
1025 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1026 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1027 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1028 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1029 			panic("moea_unwire: pvo %p is missing PVO_WIRED", pvo);
1030 		pvo->pvo_vaddr &= ~PVO_WIRED;
1031 		pm->pm_stats.wired_count--;
1032 	}
1033 	PMAP_UNLOCK(pm);
1034 }
1035 
1036 void
1037 moea_copy_page(vm_page_t msrc, vm_page_t mdst)
1038 {
1039 	vm_offset_t	dst;
1040 	vm_offset_t	src;
1041 
1042 	dst = VM_PAGE_TO_PHYS(mdst);
1043 	src = VM_PAGE_TO_PHYS(msrc);
1044 
1045 	bcopy((void *)src, (void *)dst, PAGE_SIZE);
1046 }
1047 
1048 void
1049 moea_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
1050     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1051 {
1052 	void *a_cp, *b_cp;
1053 	vm_offset_t a_pg_offset, b_pg_offset;
1054 	int cnt;
1055 
1056 	while (xfersize > 0) {
1057 		a_pg_offset = a_offset & PAGE_MASK;
1058 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1059 		a_cp = (char *)VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]) +
1060 		    a_pg_offset;
1061 		b_pg_offset = b_offset & PAGE_MASK;
1062 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1063 		b_cp = (char *)VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]) +
1064 		    b_pg_offset;
1065 		bcopy(a_cp, b_cp, cnt);
1066 		a_offset += cnt;
1067 		b_offset += cnt;
1068 		xfersize -= cnt;
1069 	}
1070 }
1071 
1072 /*
1073  * Zero a page of physical memory by temporarily mapping it into the tlb.
1074  */
1075 void
1076 moea_zero_page(vm_page_t m)
1077 {
1078 	vm_offset_t off, pa = VM_PAGE_TO_PHYS(m);
1079 
1080 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1081 		__asm __volatile("dcbz 0,%0" :: "r"(pa + off));
1082 }
1083 
1084 void
1085 moea_zero_page_area(vm_page_t m, int off, int size)
1086 {
1087 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1088 	void *va = (void *)(pa + off);
1089 
1090 	bzero(va, size);
1091 }
1092 
1093 vm_offset_t
1094 moea_quick_enter_page(vm_page_t m)
1095 {
1096 
1097 	return (VM_PAGE_TO_PHYS(m));
1098 }
1099 
1100 void
1101 moea_quick_remove_page(vm_offset_t addr)
1102 {
1103 }
1104 
1105 boolean_t
1106 moea_page_is_mapped(vm_page_t m)
1107 {
1108 	return (!LIST_EMPTY(&(m)->md.mdpg_pvoh));
1109 }
1110 
1111 /*
1112  * Map the given physical page at the specified virtual address in the
1113  * target pmap with the protection requested.  If specified the page
1114  * will be wired down.
1115  */
1116 int
1117 moea_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1118     u_int flags, int8_t psind)
1119 {
1120 	int error;
1121 
1122 	for (;;) {
1123 		rw_wlock(&pvh_global_lock);
1124 		PMAP_LOCK(pmap);
1125 		error = moea_enter_locked(pmap, va, m, prot, flags, psind);
1126 		rw_wunlock(&pvh_global_lock);
1127 		PMAP_UNLOCK(pmap);
1128 		if (error != ENOMEM)
1129 			return (KERN_SUCCESS);
1130 		if ((flags & PMAP_ENTER_NOSLEEP) != 0)
1131 			return (KERN_RESOURCE_SHORTAGE);
1132 		VM_OBJECT_ASSERT_UNLOCKED(m->object);
1133 		vm_wait(NULL);
1134 	}
1135 }
1136 
1137 /*
1138  * Map the given physical page at the specified virtual address in the
1139  * target pmap with the protection requested.  If specified the page
1140  * will be wired down.
1141  *
1142  * The global pvh and pmap must be locked.
1143  */
1144 static int
1145 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1146     u_int flags, int8_t psind __unused)
1147 {
1148 	struct		pvo_head *pvo_head;
1149 	uma_zone_t	zone;
1150 	u_int		pte_lo, pvo_flags;
1151 	int		error;
1152 
1153 	if (pmap_bootstrapped)
1154 		rw_assert(&pvh_global_lock, RA_WLOCKED);
1155 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1156 	if ((m->oflags & VPO_UNMANAGED) == 0) {
1157 		if ((flags & PMAP_ENTER_QUICK_LOCKED) == 0)
1158 			VM_PAGE_OBJECT_BUSY_ASSERT(m);
1159 		else
1160 			VM_OBJECT_ASSERT_LOCKED(m->object);
1161 	}
1162 
1163 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea_initialized) {
1164 		pvo_head = &moea_pvo_kunmanaged;
1165 		zone = moea_upvo_zone;
1166 		pvo_flags = 0;
1167 	} else {
1168 		pvo_head = vm_page_to_pvoh(m);
1169 		zone = moea_mpvo_zone;
1170 		pvo_flags = PVO_MANAGED;
1171 	}
1172 
1173 	pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1174 
1175 	if (prot & VM_PROT_WRITE) {
1176 		pte_lo |= PTE_BW;
1177 		if (pmap_bootstrapped &&
1178 		    (m->oflags & VPO_UNMANAGED) == 0)
1179 			vm_page_aflag_set(m, PGA_WRITEABLE);
1180 	} else
1181 		pte_lo |= PTE_BR;
1182 
1183 	if ((flags & PMAP_ENTER_WIRED) != 0)
1184 		pvo_flags |= PVO_WIRED;
1185 
1186 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1187 	    pte_lo, pvo_flags);
1188 
1189 	/*
1190 	 * Flush the real page from the instruction cache. This has be done
1191 	 * for all user mappings to prevent information leakage via the
1192 	 * instruction cache. moea_pvo_enter() returns ENOENT for the first
1193 	 * mapping for a page.
1194 	 */
1195 	if (pmap != kernel_pmap && error == ENOENT &&
1196 	    (pte_lo & (PTE_I | PTE_G)) == 0)
1197 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1198 
1199 	return (error);
1200 }
1201 
1202 /*
1203  * Maps a sequence of resident pages belonging to the same object.
1204  * The sequence begins with the given page m_start.  This page is
1205  * mapped at the given virtual address start.  Each subsequent page is
1206  * mapped at a virtual address that is offset from start by the same
1207  * amount as the page is offset from m_start within the object.  The
1208  * last page in the sequence is the page with the largest offset from
1209  * m_start that can be mapped at a virtual address less than the given
1210  * virtual address end.  Not every virtual page between start and end
1211  * is mapped; only those for which a resident page exists with the
1212  * corresponding offset from m_start are mapped.
1213  */
1214 void
1215 moea_enter_object(pmap_t pm, vm_offset_t start, vm_offset_t end,
1216     vm_page_t m_start, vm_prot_t prot)
1217 {
1218 	vm_page_t m;
1219 	vm_pindex_t diff, psize;
1220 
1221 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1222 
1223 	psize = atop(end - start);
1224 	m = m_start;
1225 	rw_wlock(&pvh_global_lock);
1226 	PMAP_LOCK(pm);
1227 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1228 		moea_enter_locked(pm, start + ptoa(diff), m, prot &
1229 		    (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_QUICK_LOCKED,
1230 		    0);
1231 		m = TAILQ_NEXT(m, listq);
1232 	}
1233 	rw_wunlock(&pvh_global_lock);
1234 	PMAP_UNLOCK(pm);
1235 }
1236 
1237 void
1238 moea_enter_quick(pmap_t pm, vm_offset_t va, vm_page_t m,
1239     vm_prot_t prot)
1240 {
1241 
1242 	rw_wlock(&pvh_global_lock);
1243 	PMAP_LOCK(pm);
1244 	moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1245 	    PMAP_ENTER_QUICK_LOCKED, 0);
1246 	rw_wunlock(&pvh_global_lock);
1247 	PMAP_UNLOCK(pm);
1248 }
1249 
1250 vm_paddr_t
1251 moea_extract(pmap_t pm, vm_offset_t va)
1252 {
1253 	struct	pvo_entry *pvo;
1254 	vm_paddr_t pa;
1255 
1256 	PMAP_LOCK(pm);
1257 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1258 	if (pvo == NULL)
1259 		pa = 0;
1260 	else
1261 		pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1262 	PMAP_UNLOCK(pm);
1263 	return (pa);
1264 }
1265 
1266 /*
1267  * Atomically extract and hold the physical page with the given
1268  * pmap and virtual address pair if that mapping permits the given
1269  * protection.
1270  */
1271 vm_page_t
1272 moea_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1273 {
1274 	struct	pvo_entry *pvo;
1275 	vm_page_t m;
1276 
1277 	m = NULL;
1278 	PMAP_LOCK(pmap);
1279 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1280 	if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1281 	    ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1282 	     (prot & VM_PROT_WRITE) == 0)) {
1283 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
1284 		if (!vm_page_wire_mapped(m))
1285 			m = NULL;
1286 	}
1287 	PMAP_UNLOCK(pmap);
1288 	return (m);
1289 }
1290 
1291 void
1292 moea_init()
1293 {
1294 
1295 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1296 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1297 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1298 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1299 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1300 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1301 	moea_initialized = TRUE;
1302 }
1303 
1304 boolean_t
1305 moea_is_referenced(vm_page_t m)
1306 {
1307 	boolean_t rv;
1308 
1309 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1310 	    ("moea_is_referenced: page %p is not managed", m));
1311 	rw_wlock(&pvh_global_lock);
1312 	rv = moea_query_bit(m, PTE_REF);
1313 	rw_wunlock(&pvh_global_lock);
1314 	return (rv);
1315 }
1316 
1317 boolean_t
1318 moea_is_modified(vm_page_t m)
1319 {
1320 	boolean_t rv;
1321 
1322 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1323 	    ("moea_is_modified: page %p is not managed", m));
1324 
1325 	/*
1326 	 * If the page is not busied then this check is racy.
1327 	 */
1328 	if (!pmap_page_is_write_mapped(m))
1329 		return (FALSE);
1330 
1331 	rw_wlock(&pvh_global_lock);
1332 	rv = moea_query_bit(m, PTE_CHG);
1333 	rw_wunlock(&pvh_global_lock);
1334 	return (rv);
1335 }
1336 
1337 boolean_t
1338 moea_is_prefaultable(pmap_t pmap, vm_offset_t va)
1339 {
1340 	struct pvo_entry *pvo;
1341 	boolean_t rv;
1342 
1343 	PMAP_LOCK(pmap);
1344 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1345 	rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1346 	PMAP_UNLOCK(pmap);
1347 	return (rv);
1348 }
1349 
1350 void
1351 moea_clear_modify(vm_page_t m)
1352 {
1353 
1354 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1355 	    ("moea_clear_modify: page %p is not managed", m));
1356 	vm_page_assert_busied(m);
1357 
1358 	if (!pmap_page_is_write_mapped(m))
1359 		return;
1360 	rw_wlock(&pvh_global_lock);
1361 	moea_clear_bit(m, PTE_CHG);
1362 	rw_wunlock(&pvh_global_lock);
1363 }
1364 
1365 /*
1366  * Clear the write and modified bits in each of the given page's mappings.
1367  */
1368 void
1369 moea_remove_write(vm_page_t m)
1370 {
1371 	struct	pvo_entry *pvo;
1372 	struct	pte *pt;
1373 	pmap_t	pmap;
1374 	u_int	lo;
1375 
1376 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1377 	    ("moea_remove_write: page %p is not managed", m));
1378 	vm_page_assert_busied(m);
1379 
1380 	if (!pmap_page_is_write_mapped(m))
1381 		return;
1382 	rw_wlock(&pvh_global_lock);
1383 	lo = moea_attr_fetch(m);
1384 	powerpc_sync();
1385 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1386 		pmap = pvo->pvo_pmap;
1387 		PMAP_LOCK(pmap);
1388 		if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1389 			pt = moea_pvo_to_pte(pvo, -1);
1390 			pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1391 			pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1392 			if (pt != NULL) {
1393 				moea_pte_synch(pt, &pvo->pvo_pte.pte);
1394 				lo |= pvo->pvo_pte.pte.pte_lo;
1395 				pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1396 				moea_pte_change(pt, &pvo->pvo_pte.pte,
1397 				    pvo->pvo_vaddr);
1398 				mtx_unlock(&moea_table_mutex);
1399 			}
1400 		}
1401 		PMAP_UNLOCK(pmap);
1402 	}
1403 	if ((lo & PTE_CHG) != 0) {
1404 		moea_attr_clear(m, PTE_CHG);
1405 		vm_page_dirty(m);
1406 	}
1407 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1408 	rw_wunlock(&pvh_global_lock);
1409 }
1410 
1411 /*
1412  *	moea_ts_referenced:
1413  *
1414  *	Return a count of reference bits for a page, clearing those bits.
1415  *	It is not necessary for every reference bit to be cleared, but it
1416  *	is necessary that 0 only be returned when there are truly no
1417  *	reference bits set.
1418  *
1419  *	XXX: The exact number of bits to check and clear is a matter that
1420  *	should be tested and standardized at some point in the future for
1421  *	optimal aging of shared pages.
1422  */
1423 int
1424 moea_ts_referenced(vm_page_t m)
1425 {
1426 	int count;
1427 
1428 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1429 	    ("moea_ts_referenced: page %p is not managed", m));
1430 	rw_wlock(&pvh_global_lock);
1431 	count = moea_clear_bit(m, PTE_REF);
1432 	rw_wunlock(&pvh_global_lock);
1433 	return (count);
1434 }
1435 
1436 /*
1437  * Modify the WIMG settings of all mappings for a page.
1438  */
1439 void
1440 moea_page_set_memattr(vm_page_t m, vm_memattr_t ma)
1441 {
1442 	struct	pvo_entry *pvo;
1443 	struct	pvo_head *pvo_head;
1444 	struct	pte *pt;
1445 	pmap_t	pmap;
1446 	u_int	lo;
1447 
1448 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1449 		m->md.mdpg_cache_attrs = ma;
1450 		return;
1451 	}
1452 
1453 	rw_wlock(&pvh_global_lock);
1454 	pvo_head = vm_page_to_pvoh(m);
1455 	lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1456 
1457 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1458 		pmap = pvo->pvo_pmap;
1459 		PMAP_LOCK(pmap);
1460 		pt = moea_pvo_to_pte(pvo, -1);
1461 		pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1462 		pvo->pvo_pte.pte.pte_lo |= lo;
1463 		if (pt != NULL) {
1464 			moea_pte_change(pt, &pvo->pvo_pte.pte,
1465 			    pvo->pvo_vaddr);
1466 			if (pvo->pvo_pmap == kernel_pmap)
1467 				isync();
1468 		}
1469 		mtx_unlock(&moea_table_mutex);
1470 		PMAP_UNLOCK(pmap);
1471 	}
1472 	m->md.mdpg_cache_attrs = ma;
1473 	rw_wunlock(&pvh_global_lock);
1474 }
1475 
1476 /*
1477  * Map a wired page into kernel virtual address space.
1478  */
1479 void
1480 moea_kenter(vm_offset_t va, vm_paddr_t pa)
1481 {
1482 
1483 	moea_kenter_attr(va, pa, VM_MEMATTR_DEFAULT);
1484 }
1485 
1486 void
1487 moea_kenter_attr(vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
1488 {
1489 	u_int		pte_lo;
1490 	int		error;
1491 
1492 #if 0
1493 	if (va < VM_MIN_KERNEL_ADDRESS)
1494 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1495 		    va);
1496 #endif
1497 
1498 	pte_lo = moea_calc_wimg(pa, ma);
1499 
1500 	PMAP_LOCK(kernel_pmap);
1501 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1502 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1503 
1504 	if (error != 0 && error != ENOENT)
1505 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1506 		    pa, error);
1507 
1508 	PMAP_UNLOCK(kernel_pmap);
1509 }
1510 
1511 /*
1512  * Extract the physical page address associated with the given kernel virtual
1513  * address.
1514  */
1515 vm_paddr_t
1516 moea_kextract(vm_offset_t va)
1517 {
1518 	struct		pvo_entry *pvo;
1519 	vm_paddr_t pa;
1520 
1521 	/*
1522 	 * Allow direct mappings on 32-bit OEA
1523 	 */
1524 	if (va < VM_MIN_KERNEL_ADDRESS) {
1525 		return (va);
1526 	}
1527 
1528 	PMAP_LOCK(kernel_pmap);
1529 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1530 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1531 	pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1532 	PMAP_UNLOCK(kernel_pmap);
1533 	return (pa);
1534 }
1535 
1536 /*
1537  * Remove a wired page from kernel virtual address space.
1538  */
1539 void
1540 moea_kremove(vm_offset_t va)
1541 {
1542 
1543 	moea_remove(kernel_pmap, va, va + PAGE_SIZE);
1544 }
1545 
1546 /*
1547  * Provide a kernel pointer corresponding to a given userland pointer.
1548  * The returned pointer is valid until the next time this function is
1549  * called in this thread. This is used internally in copyin/copyout.
1550  */
1551 int
1552 moea_map_user_ptr(pmap_t pm, volatile const void *uaddr,
1553     void **kaddr, size_t ulen, size_t *klen)
1554 {
1555 	size_t l;
1556 	register_t vsid;
1557 
1558 	*kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
1559 	l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
1560 	if (l > ulen)
1561 		l = ulen;
1562 	if (klen)
1563 		*klen = l;
1564 	else if (l != ulen)
1565 		return (EFAULT);
1566 
1567 	vsid = va_to_vsid(pm, (vm_offset_t)uaddr);
1568 
1569 	/* Mark segment no-execute */
1570 	vsid |= SR_N;
1571 
1572 	/* If we have already set this VSID, we can just return */
1573 	if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == vsid)
1574 		return (0);
1575 
1576 	__asm __volatile("isync");
1577 	curthread->td_pcb->pcb_cpu.aim.usr_segm =
1578 	    (uintptr_t)uaddr >> ADDR_SR_SHFT;
1579 	curthread->td_pcb->pcb_cpu.aim.usr_vsid = vsid;
1580 	__asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(vsid));
1581 
1582 	return (0);
1583 }
1584 
1585 /*
1586  * Figure out where a given kernel pointer (usually in a fault) points
1587  * to from the VM's perspective, potentially remapping into userland's
1588  * address space.
1589  */
1590 static int
1591 moea_decode_kernel_ptr(vm_offset_t addr, int *is_user,
1592     vm_offset_t *decoded_addr)
1593 {
1594 	vm_offset_t user_sr;
1595 
1596 	if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
1597 		user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
1598 		addr &= ADDR_PIDX | ADDR_POFF;
1599 		addr |= user_sr << ADDR_SR_SHFT;
1600 		*decoded_addr = addr;
1601 		*is_user = 1;
1602 	} else {
1603 		*decoded_addr = addr;
1604 		*is_user = 0;
1605 	}
1606 
1607 	return (0);
1608 }
1609 
1610 /*
1611  * Map a range of physical addresses into kernel virtual address space.
1612  *
1613  * The value passed in *virt is a suggested virtual address for the mapping.
1614  * Architectures which can support a direct-mapped physical to virtual region
1615  * can return the appropriate address within that region, leaving '*virt'
1616  * unchanged.  We cannot and therefore do not; *virt is updated with the
1617  * first usable address after the mapped region.
1618  */
1619 vm_offset_t
1620 moea_map(vm_offset_t *virt, vm_paddr_t pa_start,
1621     vm_paddr_t pa_end, int prot)
1622 {
1623 	vm_offset_t	sva, va;
1624 
1625 	sva = *virt;
1626 	va = sva;
1627 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1628 		moea_kenter(va, pa_start);
1629 	*virt = va;
1630 	return (sva);
1631 }
1632 
1633 /*
1634  * Returns true if the pmap's pv is one of the first
1635  * 16 pvs linked to from this page.  This count may
1636  * be changed upwards or downwards in the future; it
1637  * is only necessary that true be returned for a small
1638  * subset of pmaps for proper page aging.
1639  */
1640 boolean_t
1641 moea_page_exists_quick(pmap_t pmap, vm_page_t m)
1642 {
1643         int loops;
1644 	struct pvo_entry *pvo;
1645 	boolean_t rv;
1646 
1647 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1648 	    ("moea_page_exists_quick: page %p is not managed", m));
1649 	loops = 0;
1650 	rv = FALSE;
1651 	rw_wlock(&pvh_global_lock);
1652 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1653 		if (pvo->pvo_pmap == pmap) {
1654 			rv = TRUE;
1655 			break;
1656 		}
1657 		if (++loops >= 16)
1658 			break;
1659 	}
1660 	rw_wunlock(&pvh_global_lock);
1661 	return (rv);
1662 }
1663 
1664 void
1665 moea_page_init(vm_page_t m)
1666 {
1667 
1668 	m->md.mdpg_attrs = 0;
1669 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
1670 	LIST_INIT(&m->md.mdpg_pvoh);
1671 }
1672 
1673 /*
1674  * Return the number of managed mappings to the given physical page
1675  * that are wired.
1676  */
1677 int
1678 moea_page_wired_mappings(vm_page_t m)
1679 {
1680 	struct pvo_entry *pvo;
1681 	int count;
1682 
1683 	count = 0;
1684 	if ((m->oflags & VPO_UNMANAGED) != 0)
1685 		return (count);
1686 	rw_wlock(&pvh_global_lock);
1687 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1688 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1689 			count++;
1690 	rw_wunlock(&pvh_global_lock);
1691 	return (count);
1692 }
1693 
1694 static u_int	moea_vsidcontext;
1695 
1696 int
1697 moea_pinit(pmap_t pmap)
1698 {
1699 	int	i, mask;
1700 	u_int	entropy;
1701 
1702 	RB_INIT(&pmap->pmap_pvo);
1703 
1704 	entropy = 0;
1705 	__asm __volatile("mftb %0" : "=r"(entropy));
1706 
1707 	if ((pmap->pmap_phys = (pmap_t)moea_kextract((vm_offset_t)pmap))
1708 	    == NULL) {
1709 		pmap->pmap_phys = pmap;
1710 	}
1711 
1712 
1713 	mtx_lock(&moea_vsid_mutex);
1714 	/*
1715 	 * Allocate some segment registers for this pmap.
1716 	 */
1717 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1718 		u_int	hash, n;
1719 
1720 		/*
1721 		 * Create a new value by mutiplying by a prime and adding in
1722 		 * entropy from the timebase register.  This is to make the
1723 		 * VSID more random so that the PT hash function collides
1724 		 * less often.  (Note that the prime casues gcc to do shifts
1725 		 * instead of a multiply.)
1726 		 */
1727 		moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1728 		hash = moea_vsidcontext & (NPMAPS - 1);
1729 		if (hash == 0)		/* 0 is special, avoid it */
1730 			continue;
1731 		n = hash >> 5;
1732 		mask = 1 << (hash & (VSID_NBPW - 1));
1733 		hash = (moea_vsidcontext & 0xfffff);
1734 		if (moea_vsid_bitmap[n] & mask) {	/* collision? */
1735 			/* anything free in this bucket? */
1736 			if (moea_vsid_bitmap[n] == 0xffffffff) {
1737 				entropy = (moea_vsidcontext >> 20);
1738 				continue;
1739 			}
1740 			i = ffs(~moea_vsid_bitmap[n]) - 1;
1741 			mask = 1 << i;
1742 			hash &= rounddown2(0xfffff, VSID_NBPW);
1743 			hash |= i;
1744 		}
1745 		KASSERT(!(moea_vsid_bitmap[n] & mask),
1746 		    ("Allocating in-use VSID group %#x\n", hash));
1747 		moea_vsid_bitmap[n] |= mask;
1748 		for (i = 0; i < 16; i++)
1749 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1750 		mtx_unlock(&moea_vsid_mutex);
1751 		return (1);
1752 	}
1753 
1754 	mtx_unlock(&moea_vsid_mutex);
1755 	panic("moea_pinit: out of segments");
1756 }
1757 
1758 /*
1759  * Initialize the pmap associated with process 0.
1760  */
1761 void
1762 moea_pinit0(pmap_t pm)
1763 {
1764 
1765 	PMAP_LOCK_INIT(pm);
1766 	moea_pinit(pm);
1767 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1768 }
1769 
1770 /*
1771  * Set the physical protection on the specified range of this map as requested.
1772  */
1773 void
1774 moea_protect(pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1775     vm_prot_t prot)
1776 {
1777 	struct	pvo_entry *pvo, *tpvo, key;
1778 	struct	pte *pt;
1779 
1780 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1781 	    ("moea_protect: non current pmap"));
1782 
1783 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1784 		moea_remove(pm, sva, eva);
1785 		return;
1786 	}
1787 
1788 	rw_wlock(&pvh_global_lock);
1789 	PMAP_LOCK(pm);
1790 	key.pvo_vaddr = sva;
1791 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1792 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1793 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1794 
1795 		/*
1796 		 * Grab the PTE pointer before we diddle with the cached PTE
1797 		 * copy.
1798 		 */
1799 		pt = moea_pvo_to_pte(pvo, -1);
1800 		/*
1801 		 * Change the protection of the page.
1802 		 */
1803 		pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1804 		pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1805 
1806 		/*
1807 		 * If the PVO is in the page table, update that pte as well.
1808 		 */
1809 		if (pt != NULL) {
1810 			moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1811 			mtx_unlock(&moea_table_mutex);
1812 		}
1813 	}
1814 	rw_wunlock(&pvh_global_lock);
1815 	PMAP_UNLOCK(pm);
1816 }
1817 
1818 /*
1819  * Map a list of wired pages into kernel virtual address space.  This is
1820  * intended for temporary mappings which do not need page modification or
1821  * references recorded.  Existing mappings in the region are overwritten.
1822  */
1823 void
1824 moea_qenter(vm_offset_t sva, vm_page_t *m, int count)
1825 {
1826 	vm_offset_t va;
1827 
1828 	va = sva;
1829 	while (count-- > 0) {
1830 		moea_kenter(va, VM_PAGE_TO_PHYS(*m));
1831 		va += PAGE_SIZE;
1832 		m++;
1833 	}
1834 }
1835 
1836 /*
1837  * Remove page mappings from kernel virtual address space.  Intended for
1838  * temporary mappings entered by moea_qenter.
1839  */
1840 void
1841 moea_qremove(vm_offset_t sva, int count)
1842 {
1843 	vm_offset_t va;
1844 
1845 	va = sva;
1846 	while (count-- > 0) {
1847 		moea_kremove(va);
1848 		va += PAGE_SIZE;
1849 	}
1850 }
1851 
1852 void
1853 moea_release(pmap_t pmap)
1854 {
1855         int idx, mask;
1856 
1857 	/*
1858 	 * Free segment register's VSID
1859 	 */
1860         if (pmap->pm_sr[0] == 0)
1861                 panic("moea_release");
1862 
1863 	mtx_lock(&moea_vsid_mutex);
1864         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1865         mask = 1 << (idx % VSID_NBPW);
1866         idx /= VSID_NBPW;
1867         moea_vsid_bitmap[idx] &= ~mask;
1868 	mtx_unlock(&moea_vsid_mutex);
1869 }
1870 
1871 /*
1872  * Remove the given range of addresses from the specified map.
1873  */
1874 void
1875 moea_remove(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1876 {
1877 	struct	pvo_entry *pvo, *tpvo, key;
1878 
1879 	rw_wlock(&pvh_global_lock);
1880 	PMAP_LOCK(pm);
1881 	key.pvo_vaddr = sva;
1882 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1883 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1884 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1885 		moea_pvo_remove(pvo, -1);
1886 	}
1887 	PMAP_UNLOCK(pm);
1888 	rw_wunlock(&pvh_global_lock);
1889 }
1890 
1891 /*
1892  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1893  * will reflect changes in pte's back to the vm_page.
1894  */
1895 void
1896 moea_remove_all(vm_page_t m)
1897 {
1898 	struct  pvo_head *pvo_head;
1899 	struct	pvo_entry *pvo, *next_pvo;
1900 	pmap_t	pmap;
1901 
1902 	rw_wlock(&pvh_global_lock);
1903 	pvo_head = vm_page_to_pvoh(m);
1904 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1905 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1906 
1907 		pmap = pvo->pvo_pmap;
1908 		PMAP_LOCK(pmap);
1909 		moea_pvo_remove(pvo, -1);
1910 		PMAP_UNLOCK(pmap);
1911 	}
1912 	if ((m->a.flags & PGA_WRITEABLE) && moea_query_bit(m, PTE_CHG)) {
1913 		moea_attr_clear(m, PTE_CHG);
1914 		vm_page_dirty(m);
1915 	}
1916 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1917 	rw_wunlock(&pvh_global_lock);
1918 }
1919 
1920 /*
1921  * Allocate a physical page of memory directly from the phys_avail map.
1922  * Can only be called from moea_bootstrap before avail start and end are
1923  * calculated.
1924  */
1925 static vm_offset_t
1926 moea_bootstrap_alloc(vm_size_t size, u_int align)
1927 {
1928 	vm_offset_t	s, e;
1929 	int		i, j;
1930 
1931 	size = round_page(size);
1932 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1933 		if (align != 0)
1934 			s = roundup2(phys_avail[i], align);
1935 		else
1936 			s = phys_avail[i];
1937 		e = s + size;
1938 
1939 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1940 			continue;
1941 
1942 		if (s == phys_avail[i]) {
1943 			phys_avail[i] += size;
1944 		} else if (e == phys_avail[i + 1]) {
1945 			phys_avail[i + 1] -= size;
1946 		} else {
1947 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1948 				phys_avail[j] = phys_avail[j - 2];
1949 				phys_avail[j + 1] = phys_avail[j - 1];
1950 			}
1951 
1952 			phys_avail[i + 3] = phys_avail[i + 1];
1953 			phys_avail[i + 1] = s;
1954 			phys_avail[i + 2] = e;
1955 			phys_avail_count++;
1956 		}
1957 
1958 		return (s);
1959 	}
1960 	panic("moea_bootstrap_alloc: could not allocate memory");
1961 }
1962 
1963 static void
1964 moea_syncicache(vm_paddr_t pa, vm_size_t len)
1965 {
1966 	__syncicache((void *)pa, len);
1967 }
1968 
1969 static int
1970 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1971     vm_offset_t va, vm_paddr_t pa, u_int pte_lo, int flags)
1972 {
1973 	struct	pvo_entry *pvo;
1974 	u_int	sr;
1975 	int	first;
1976 	u_int	ptegidx;
1977 	int	i;
1978 	int     bootstrap;
1979 
1980 	moea_pvo_enter_calls++;
1981 	first = 0;
1982 	bootstrap = 0;
1983 
1984 	/*
1985 	 * Compute the PTE Group index.
1986 	 */
1987 	va &= ~ADDR_POFF;
1988 	sr = va_to_sr(pm->pm_sr, va);
1989 	ptegidx = va_to_pteg(sr, va);
1990 
1991 	/*
1992 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
1993 	 * there is a mapping.
1994 	 */
1995 	mtx_lock(&moea_table_mutex);
1996 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
1997 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
1998 			if ((pvo->pvo_pte.pte.pte_lo & PTE_RPGN) == pa &&
1999 			    (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
2000 			    (pte_lo & PTE_PP)) {
2001 				/*
2002 				 * The PTE is not changing.  Instead, this may
2003 				 * be a request to change the mapping's wired
2004 				 * attribute.
2005 				 */
2006 				mtx_unlock(&moea_table_mutex);
2007 				if ((flags & PVO_WIRED) != 0 &&
2008 				    (pvo->pvo_vaddr & PVO_WIRED) == 0) {
2009 					pvo->pvo_vaddr |= PVO_WIRED;
2010 					pm->pm_stats.wired_count++;
2011 				} else if ((flags & PVO_WIRED) == 0 &&
2012 				    (pvo->pvo_vaddr & PVO_WIRED) != 0) {
2013 					pvo->pvo_vaddr &= ~PVO_WIRED;
2014 					pm->pm_stats.wired_count--;
2015 				}
2016 				return (0);
2017 			}
2018 			moea_pvo_remove(pvo, -1);
2019 			break;
2020 		}
2021 	}
2022 
2023 	/*
2024 	 * If we aren't overwriting a mapping, try to allocate.
2025 	 */
2026 	if (moea_initialized) {
2027 		pvo = uma_zalloc(zone, M_NOWAIT);
2028 	} else {
2029 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
2030 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
2031 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
2032 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
2033 		}
2034 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
2035 		moea_bpvo_pool_index++;
2036 		bootstrap = 1;
2037 	}
2038 
2039 	if (pvo == NULL) {
2040 		mtx_unlock(&moea_table_mutex);
2041 		return (ENOMEM);
2042 	}
2043 
2044 	moea_pvo_entries++;
2045 	pvo->pvo_vaddr = va;
2046 	pvo->pvo_pmap = pm;
2047 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
2048 	pvo->pvo_vaddr &= ~ADDR_POFF;
2049 	if (flags & PVO_WIRED)
2050 		pvo->pvo_vaddr |= PVO_WIRED;
2051 	if (pvo_head != &moea_pvo_kunmanaged)
2052 		pvo->pvo_vaddr |= PVO_MANAGED;
2053 	if (bootstrap)
2054 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
2055 
2056 	moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo);
2057 
2058 	/*
2059 	 * Add to pmap list
2060 	 */
2061 	RB_INSERT(pvo_tree, &pm->pmap_pvo, pvo);
2062 
2063 	/*
2064 	 * Remember if the list was empty and therefore will be the first
2065 	 * item.
2066 	 */
2067 	if (LIST_FIRST(pvo_head) == NULL)
2068 		first = 1;
2069 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2070 
2071 	if (pvo->pvo_vaddr & PVO_WIRED)
2072 		pm->pm_stats.wired_count++;
2073 	pm->pm_stats.resident_count++;
2074 
2075 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2076 	KASSERT(i < 8, ("Invalid PTE index"));
2077 	if (i >= 0) {
2078 		PVO_PTEGIDX_SET(pvo, i);
2079 	} else {
2080 		panic("moea_pvo_enter: overflow");
2081 		moea_pte_overflow++;
2082 	}
2083 	mtx_unlock(&moea_table_mutex);
2084 
2085 	return (first ? ENOENT : 0);
2086 }
2087 
2088 static void
2089 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
2090 {
2091 	struct	pte *pt;
2092 
2093 	/*
2094 	 * If there is an active pte entry, we need to deactivate it (and
2095 	 * save the ref & cfg bits).
2096 	 */
2097 	pt = moea_pvo_to_pte(pvo, pteidx);
2098 	if (pt != NULL) {
2099 		moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
2100 		mtx_unlock(&moea_table_mutex);
2101 		PVO_PTEGIDX_CLR(pvo);
2102 	} else {
2103 		moea_pte_overflow--;
2104 	}
2105 
2106 	/*
2107 	 * Update our statistics.
2108 	 */
2109 	pvo->pvo_pmap->pm_stats.resident_count--;
2110 	if (pvo->pvo_vaddr & PVO_WIRED)
2111 		pvo->pvo_pmap->pm_stats.wired_count--;
2112 
2113 	/*
2114 	 * Remove this PVO from the PV and pmap lists.
2115 	 */
2116 	LIST_REMOVE(pvo, pvo_vlink);
2117 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2118 
2119 	/*
2120 	 * Save the REF/CHG bits into their cache if the page is managed.
2121 	 * Clear PGA_WRITEABLE if all mappings of the page have been removed.
2122 	 */
2123 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED) {
2124 		struct vm_page *pg;
2125 
2126 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
2127 		if (pg != NULL) {
2128 			moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2129 			    (PTE_REF | PTE_CHG));
2130 			if (LIST_EMPTY(&pg->md.mdpg_pvoh))
2131 				vm_page_aflag_clear(pg, PGA_WRITEABLE);
2132 		}
2133 	}
2134 
2135 	/*
2136 	 * Remove this from the overflow list and return it to the pool
2137 	 * if we aren't going to reuse it.
2138 	 */
2139 	LIST_REMOVE(pvo, pvo_olink);
2140 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2141 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
2142 		    moea_upvo_zone, pvo);
2143 	moea_pvo_entries--;
2144 	moea_pvo_remove_calls++;
2145 }
2146 
2147 static __inline int
2148 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2149 {
2150 	int	pteidx;
2151 
2152 	/*
2153 	 * We can find the actual pte entry without searching by grabbing
2154 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2155 	 * noticing the HID bit.
2156 	 */
2157 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2158 	if (pvo->pvo_pte.pte.pte_hi & PTE_HID)
2159 		pteidx ^= moea_pteg_mask * 8;
2160 
2161 	return (pteidx);
2162 }
2163 
2164 static struct pvo_entry *
2165 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2166 {
2167 	struct	pvo_entry *pvo;
2168 	int	ptegidx;
2169 	u_int	sr;
2170 
2171 	va &= ~ADDR_POFF;
2172 	sr = va_to_sr(pm->pm_sr, va);
2173 	ptegidx = va_to_pteg(sr, va);
2174 
2175 	mtx_lock(&moea_table_mutex);
2176 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2177 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2178 			if (pteidx_p)
2179 				*pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2180 			break;
2181 		}
2182 	}
2183 	mtx_unlock(&moea_table_mutex);
2184 
2185 	return (pvo);
2186 }
2187 
2188 static struct pte *
2189 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2190 {
2191 	struct	pte *pt;
2192 
2193 	/*
2194 	 * If we haven't been supplied the ptegidx, calculate it.
2195 	 */
2196 	if (pteidx == -1) {
2197 		int	ptegidx;
2198 		u_int	sr;
2199 
2200 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2201 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2202 		pteidx = moea_pvo_pte_index(pvo, ptegidx);
2203 	}
2204 
2205 	pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2206 	mtx_lock(&moea_table_mutex);
2207 
2208 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2209 		panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2210 		    "valid pte index", pvo);
2211 	}
2212 
2213 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2214 		panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2215 		    "pvo but no valid pte", pvo);
2216 	}
2217 
2218 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2219 		if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) {
2220 			panic("moea_pvo_to_pte: pvo %p has valid pte in "
2221 			    "moea_pteg_table %p but invalid in pvo", pvo, pt);
2222 		}
2223 
2224 		if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2225 		    != 0) {
2226 			panic("moea_pvo_to_pte: pvo %p pte does not match "
2227 			    "pte %p in moea_pteg_table", pvo, pt);
2228 		}
2229 
2230 		mtx_assert(&moea_table_mutex, MA_OWNED);
2231 		return (pt);
2232 	}
2233 
2234 	if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) {
2235 		panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2236 		    "moea_pteg_table but valid in pvo: %8x, %8x", pvo, pt, pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2237 	}
2238 
2239 	mtx_unlock(&moea_table_mutex);
2240 	return (NULL);
2241 }
2242 
2243 /*
2244  * XXX: THIS STUFF SHOULD BE IN pte.c?
2245  */
2246 int
2247 moea_pte_spill(vm_offset_t addr)
2248 {
2249 	struct	pvo_entry *source_pvo, *victim_pvo;
2250 	struct	pvo_entry *pvo;
2251 	int	ptegidx, i, j;
2252 	u_int	sr;
2253 	struct	pteg *pteg;
2254 	struct	pte *pt;
2255 
2256 	moea_pte_spills++;
2257 
2258 	sr = mfsrin(addr);
2259 	ptegidx = va_to_pteg(sr, addr);
2260 
2261 	/*
2262 	 * Have to substitute some entry.  Use the primary hash for this.
2263 	 * Use low bits of timebase as random generator.
2264 	 */
2265 	pteg = &moea_pteg_table[ptegidx];
2266 	mtx_lock(&moea_table_mutex);
2267 	__asm __volatile("mftb %0" : "=r"(i));
2268 	i &= 7;
2269 	pt = &pteg->pt[i];
2270 
2271 	source_pvo = NULL;
2272 	victim_pvo = NULL;
2273 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2274 		/*
2275 		 * We need to find a pvo entry for this address.
2276 		 */
2277 		if (source_pvo == NULL &&
2278 		    moea_pte_match(&pvo->pvo_pte.pte, sr, addr,
2279 		    pvo->pvo_pte.pte.pte_hi & PTE_HID)) {
2280 			/*
2281 			 * Now found an entry to be spilled into the pteg.
2282 			 * The PTE is now valid, so we know it's active.
2283 			 */
2284 			j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2285 
2286 			if (j >= 0) {
2287 				PVO_PTEGIDX_SET(pvo, j);
2288 				moea_pte_overflow--;
2289 				mtx_unlock(&moea_table_mutex);
2290 				return (1);
2291 			}
2292 
2293 			source_pvo = pvo;
2294 
2295 			if (victim_pvo != NULL)
2296 				break;
2297 		}
2298 
2299 		/*
2300 		 * We also need the pvo entry of the victim we are replacing
2301 		 * so save the R & C bits of the PTE.
2302 		 */
2303 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2304 		    moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2305 			victim_pvo = pvo;
2306 			if (source_pvo != NULL)
2307 				break;
2308 		}
2309 	}
2310 
2311 	if (source_pvo == NULL) {
2312 		mtx_unlock(&moea_table_mutex);
2313 		return (0);
2314 	}
2315 
2316 	if (victim_pvo == NULL) {
2317 		if ((pt->pte_hi & PTE_HID) == 0)
2318 			panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2319 			    "entry", pt);
2320 
2321 		/*
2322 		 * If this is a secondary PTE, we need to search it's primary
2323 		 * pvo bucket for the matching PVO.
2324 		 */
2325 		LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2326 		    pvo_olink) {
2327 			/*
2328 			 * We also need the pvo entry of the victim we are
2329 			 * replacing so save the R & C bits of the PTE.
2330 			 */
2331 			if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2332 				victim_pvo = pvo;
2333 				break;
2334 			}
2335 		}
2336 
2337 		if (victim_pvo == NULL)
2338 			panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2339 			    "entry", pt);
2340 	}
2341 
2342 	/*
2343 	 * We are invalidating the TLB entry for the EA we are replacing even
2344 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2345 	 * contained in the TLB entry.
2346 	 */
2347 	source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID;
2348 
2349 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2350 	moea_pte_set(pt, &source_pvo->pvo_pte.pte);
2351 
2352 	PVO_PTEGIDX_CLR(victim_pvo);
2353 	PVO_PTEGIDX_SET(source_pvo, i);
2354 	moea_pte_replacements++;
2355 
2356 	mtx_unlock(&moea_table_mutex);
2357 	return (1);
2358 }
2359 
2360 static __inline struct pvo_entry *
2361 moea_pte_spillable_ident(u_int ptegidx)
2362 {
2363 	struct	pte *pt;
2364 	struct	pvo_entry *pvo_walk, *pvo = NULL;
2365 
2366 	LIST_FOREACH(pvo_walk, &moea_pvo_table[ptegidx], pvo_olink) {
2367 		if (pvo_walk->pvo_vaddr & PVO_WIRED)
2368 			continue;
2369 
2370 		if (!(pvo_walk->pvo_pte.pte.pte_hi & PTE_VALID))
2371 			continue;
2372 
2373 		pt = moea_pvo_to_pte(pvo_walk, -1);
2374 
2375 		if (pt == NULL)
2376 			continue;
2377 
2378 		pvo = pvo_walk;
2379 
2380 		mtx_unlock(&moea_table_mutex);
2381 		if (!(pt->pte_lo & PTE_REF))
2382 			return (pvo_walk);
2383 	}
2384 
2385 	return (pvo);
2386 }
2387 
2388 static int
2389 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2390 {
2391 	struct	pte *pt;
2392 	struct	pvo_entry *victim_pvo;
2393 	int	i;
2394 	int	victim_idx;
2395 	u_int	pteg_bkpidx = ptegidx;
2396 
2397 	mtx_assert(&moea_table_mutex, MA_OWNED);
2398 
2399 	/*
2400 	 * First try primary hash.
2401 	 */
2402 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2403 		if ((pt->pte_hi & PTE_VALID) == 0) {
2404 			pvo_pt->pte_hi &= ~PTE_HID;
2405 			moea_pte_set(pt, pvo_pt);
2406 			return (i);
2407 		}
2408 	}
2409 
2410 	/*
2411 	 * Now try secondary hash.
2412 	 */
2413 	ptegidx ^= moea_pteg_mask;
2414 
2415 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2416 		if ((pt->pte_hi & PTE_VALID) == 0) {
2417 			pvo_pt->pte_hi |= PTE_HID;
2418 			moea_pte_set(pt, pvo_pt);
2419 			return (i);
2420 		}
2421 	}
2422 
2423 	/* Try again, but this time try to force a PTE out. */
2424 	ptegidx = pteg_bkpidx;
2425 
2426 	victim_pvo = moea_pte_spillable_ident(ptegidx);
2427 	if (victim_pvo == NULL) {
2428 		ptegidx ^= moea_pteg_mask;
2429 		victim_pvo = moea_pte_spillable_ident(ptegidx);
2430 	}
2431 
2432 	if (victim_pvo == NULL) {
2433 		panic("moea_pte_insert: overflow");
2434 		return (-1);
2435 	}
2436 
2437 	victim_idx = moea_pvo_pte_index(victim_pvo, ptegidx);
2438 
2439 	if (pteg_bkpidx == ptegidx)
2440 		pvo_pt->pte_hi &= ~PTE_HID;
2441 	else
2442 		pvo_pt->pte_hi |= PTE_HID;
2443 
2444 	/*
2445 	 * Synchronize the sacrifice PTE with its PVO, then mark both
2446 	 * invalid. The PVO will be reused when/if the VM system comes
2447 	 * here after a fault.
2448 	 */
2449 	pt = &moea_pteg_table[victim_idx >> 3].pt[victim_idx & 7];
2450 
2451 	if (pt->pte_hi != victim_pvo->pvo_pte.pte.pte_hi)
2452 	    panic("Victim PVO doesn't match PTE! PVO: %8x, PTE: %8x", victim_pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2453 
2454 	/*
2455 	 * Set the new PTE.
2456 	 */
2457 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2458 	PVO_PTEGIDX_CLR(victim_pvo);
2459 	moea_pte_overflow++;
2460 	moea_pte_set(pt, pvo_pt);
2461 
2462 	return (victim_idx & 7);
2463 }
2464 
2465 static boolean_t
2466 moea_query_bit(vm_page_t m, int ptebit)
2467 {
2468 	struct	pvo_entry *pvo;
2469 	struct	pte *pt;
2470 
2471 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2472 	if (moea_attr_fetch(m) & ptebit)
2473 		return (TRUE);
2474 
2475 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2476 
2477 		/*
2478 		 * See if we saved the bit off.  If so, cache it and return
2479 		 * success.
2480 		 */
2481 		if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2482 			moea_attr_save(m, ptebit);
2483 			return (TRUE);
2484 		}
2485 	}
2486 
2487 	/*
2488 	 * No luck, now go through the hard part of looking at the PTEs
2489 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2490 	 * the PTEs.
2491 	 */
2492 	powerpc_sync();
2493 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2494 
2495 		/*
2496 		 * See if this pvo has a valid PTE.  if so, fetch the
2497 		 * REF/CHG bits from the valid PTE.  If the appropriate
2498 		 * ptebit is set, cache it and return success.
2499 		 */
2500 		pt = moea_pvo_to_pte(pvo, -1);
2501 		if (pt != NULL) {
2502 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2503 			mtx_unlock(&moea_table_mutex);
2504 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2505 				moea_attr_save(m, ptebit);
2506 				return (TRUE);
2507 			}
2508 		}
2509 	}
2510 
2511 	return (FALSE);
2512 }
2513 
2514 static u_int
2515 moea_clear_bit(vm_page_t m, int ptebit)
2516 {
2517 	u_int	count;
2518 	struct	pvo_entry *pvo;
2519 	struct	pte *pt;
2520 
2521 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2522 
2523 	/*
2524 	 * Clear the cached value.
2525 	 */
2526 	moea_attr_clear(m, ptebit);
2527 
2528 	/*
2529 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2530 	 * we can reset the right ones).  note that since the pvo entries and
2531 	 * list heads are accessed via BAT0 and are never placed in the page
2532 	 * table, we don't have to worry about further accesses setting the
2533 	 * REF/CHG bits.
2534 	 */
2535 	powerpc_sync();
2536 
2537 	/*
2538 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2539 	 * valid pte clear the ptebit from the valid pte.
2540 	 */
2541 	count = 0;
2542 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2543 		pt = moea_pvo_to_pte(pvo, -1);
2544 		if (pt != NULL) {
2545 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2546 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2547 				count++;
2548 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2549 			}
2550 			mtx_unlock(&moea_table_mutex);
2551 		}
2552 		pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2553 	}
2554 
2555 	return (count);
2556 }
2557 
2558 /*
2559  * Return true if the physical range is encompassed by the battable[idx]
2560  */
2561 static int
2562 moea_bat_mapped(int idx, vm_paddr_t pa, vm_size_t size)
2563 {
2564 	u_int prot;
2565 	u_int32_t start;
2566 	u_int32_t end;
2567 	u_int32_t bat_ble;
2568 
2569 	/*
2570 	 * Return immediately if not a valid mapping
2571 	 */
2572 	if (!(battable[idx].batu & BAT_Vs))
2573 		return (EINVAL);
2574 
2575 	/*
2576 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2577 	 * so it can function as an i/o page
2578 	 */
2579 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2580 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2581 		return (EPERM);
2582 
2583 	/*
2584 	 * The address should be within the BAT range. Assume that the
2585 	 * start address in the BAT has the correct alignment (thus
2586 	 * not requiring masking)
2587 	 */
2588 	start = battable[idx].batl & BAT_PBS;
2589 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2590 	end = start | (bat_ble << 15) | 0x7fff;
2591 
2592 	if ((pa < start) || ((pa + size) > end))
2593 		return (ERANGE);
2594 
2595 	return (0);
2596 }
2597 
2598 boolean_t
2599 moea_dev_direct_mapped(vm_paddr_t pa, vm_size_t size)
2600 {
2601 	int i;
2602 
2603 	/*
2604 	 * This currently does not work for entries that
2605 	 * overlap 256M BAT segments.
2606 	 */
2607 
2608 	for(i = 0; i < 16; i++)
2609 		if (moea_bat_mapped(i, pa, size) == 0)
2610 			return (0);
2611 
2612 	return (EFAULT);
2613 }
2614 
2615 /*
2616  * Map a set of physical memory pages into the kernel virtual
2617  * address space. Return a pointer to where it is mapped. This
2618  * routine is intended to be used for mapping device memory,
2619  * NOT real memory.
2620  */
2621 void *
2622 moea_mapdev(vm_paddr_t pa, vm_size_t size)
2623 {
2624 
2625 	return (moea_mapdev_attr(pa, size, VM_MEMATTR_DEFAULT));
2626 }
2627 
2628 void *
2629 moea_mapdev_attr(vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
2630 {
2631 	vm_offset_t va, tmpva, ppa, offset;
2632 	int i;
2633 
2634 	ppa = trunc_page(pa);
2635 	offset = pa & PAGE_MASK;
2636 	size = roundup(offset + size, PAGE_SIZE);
2637 
2638 	/*
2639 	 * If the physical address lies within a valid BAT table entry,
2640 	 * return the 1:1 mapping. This currently doesn't work
2641 	 * for regions that overlap 256M BAT segments.
2642 	 */
2643 	for (i = 0; i < 16; i++) {
2644 		if (moea_bat_mapped(i, pa, size) == 0)
2645 			return ((void *) pa);
2646 	}
2647 
2648 	va = kva_alloc(size);
2649 	if (!va)
2650 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2651 
2652 	for (tmpva = va; size > 0;) {
2653 		moea_kenter_attr(tmpva, ppa, ma);
2654 		tlbie(tmpva);
2655 		size -= PAGE_SIZE;
2656 		tmpva += PAGE_SIZE;
2657 		ppa += PAGE_SIZE;
2658 	}
2659 
2660 	return ((void *)(va + offset));
2661 }
2662 
2663 void
2664 moea_unmapdev(vm_offset_t va, vm_size_t size)
2665 {
2666 	vm_offset_t base, offset;
2667 
2668 	/*
2669 	 * If this is outside kernel virtual space, then it's a
2670 	 * battable entry and doesn't require unmapping
2671 	 */
2672 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2673 		base = trunc_page(va);
2674 		offset = va & PAGE_MASK;
2675 		size = roundup(offset + size, PAGE_SIZE);
2676 		moea_qremove(base, atop(size));
2677 		kva_free(base, size);
2678 	}
2679 }
2680 
2681 static void
2682 moea_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz)
2683 {
2684 	struct pvo_entry *pvo;
2685 	vm_offset_t lim;
2686 	vm_paddr_t pa;
2687 	vm_size_t len;
2688 
2689 	PMAP_LOCK(pm);
2690 	while (sz > 0) {
2691 		lim = round_page(va + 1);
2692 		len = MIN(lim - va, sz);
2693 		pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2694 		if (pvo != NULL) {
2695 			pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) |
2696 			    (va & ADDR_POFF);
2697 			moea_syncicache(pa, len);
2698 		}
2699 		va += len;
2700 		sz -= len;
2701 	}
2702 	PMAP_UNLOCK(pm);
2703 }
2704 
2705 void
2706 moea_dumpsys_map(vm_paddr_t pa, size_t sz, void **va)
2707 {
2708 
2709 	*va = (void *)pa;
2710 }
2711 
2712 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
2713 
2714 void
2715 moea_scan_init()
2716 {
2717 	struct pvo_entry *pvo;
2718 	vm_offset_t va;
2719 	int i;
2720 
2721 	if (!do_minidump) {
2722 		/* Initialize phys. segments for dumpsys(). */
2723 		memset(&dump_map, 0, sizeof(dump_map));
2724 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
2725 		for (i = 0; i < pregions_sz; i++) {
2726 			dump_map[i].pa_start = pregions[i].mr_start;
2727 			dump_map[i].pa_size = pregions[i].mr_size;
2728 		}
2729 		return;
2730 	}
2731 
2732 	/* Virtual segments for minidumps: */
2733 	memset(&dump_map, 0, sizeof(dump_map));
2734 
2735 	/* 1st: kernel .data and .bss. */
2736 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
2737 	dump_map[0].pa_size =
2738 	    round_page((uintptr_t)_end) - dump_map[0].pa_start;
2739 
2740 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2741 	dump_map[1].pa_start = (vm_paddr_t)msgbufp->msg_ptr;
2742 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
2743 
2744 	/* 3rd: kernel VM. */
2745 	va = dump_map[1].pa_start + dump_map[1].pa_size;
2746 	/* Find start of next chunk (from va). */
2747 	while (va < virtual_end) {
2748 		/* Don't dump the buffer cache. */
2749 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
2750 			va = kmi.buffer_eva;
2751 			continue;
2752 		}
2753 		pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
2754 		if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2755 			break;
2756 		va += PAGE_SIZE;
2757 	}
2758 	if (va < virtual_end) {
2759 		dump_map[2].pa_start = va;
2760 		va += PAGE_SIZE;
2761 		/* Find last page in chunk. */
2762 		while (va < virtual_end) {
2763 			/* Don't run into the buffer cache. */
2764 			if (va == kmi.buffer_sva)
2765 				break;
2766 			pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF,
2767 			    NULL);
2768 			if (pvo == NULL ||
2769 			    !(pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2770 				break;
2771 			va += PAGE_SIZE;
2772 		}
2773 		dump_map[2].pa_size = va - dump_map[2].pa_start;
2774 	}
2775 }
2776