xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision eac7052fdebb90caf2f653e06187bdbca837b9c7)
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 static struct pmap_funcs moea_methods = {
330 	.clear_modify = moea_clear_modify,
331 	.copy_page = moea_copy_page,
332 	.copy_pages = moea_copy_pages,
333 	.enter = moea_enter,
334 	.enter_object = moea_enter_object,
335 	.enter_quick = moea_enter_quick,
336 	.extract = moea_extract,
337 	.extract_and_hold = moea_extract_and_hold,
338 	.init = moea_init,
339 	.is_modified = moea_is_modified,
340 	.is_prefaultable = moea_is_prefaultable,
341 	.is_referenced = moea_is_referenced,
342 	.ts_referenced = moea_ts_referenced,
343 	.map =      		moea_map,
344 	.page_exists_quick = moea_page_exists_quick,
345 	.page_init = moea_page_init,
346 	.page_wired_mappings = moea_page_wired_mappings,
347 	.pinit = moea_pinit,
348 	.pinit0 = moea_pinit0,
349 	.protect = moea_protect,
350 	.qenter = moea_qenter,
351 	.qremove = moea_qremove,
352 	.release = moea_release,
353 	.remove = moea_remove,
354 	.remove_all =       	moea_remove_all,
355 	.remove_write = moea_remove_write,
356 	.sync_icache = moea_sync_icache,
357 	.unwire = moea_unwire,
358 	.zero_page =        	moea_zero_page,
359 	.zero_page_area = moea_zero_page_area,
360 	.activate = moea_activate,
361 	.deactivate =       	moea_deactivate,
362 	.page_set_memattr = moea_page_set_memattr,
363 	.quick_enter_page =  moea_quick_enter_page,
364 	.quick_remove_page =  moea_quick_remove_page,
365 	.page_is_mapped = moea_page_is_mapped,
366 
367 	/* Internal interfaces */
368 	.bootstrap =        	moea_bootstrap,
369 	.cpu_bootstrap =    	moea_cpu_bootstrap,
370 	.mapdev_attr = moea_mapdev_attr,
371 	.mapdev = moea_mapdev,
372 	.unmapdev = moea_unmapdev,
373 	.kextract = moea_kextract,
374 	.kenter = moea_kenter,
375 	.kenter_attr = moea_kenter_attr,
376 	.dev_direct_mapped = moea_dev_direct_mapped,
377 	.dumpsys_pa_init = moea_scan_init,
378 	.dumpsys_map_chunk = moea_dumpsys_map,
379 	.map_user_ptr = moea_map_user_ptr,
380 	.decode_kernel_ptr =  moea_decode_kernel_ptr,
381 };
382 
383 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods);
384 
385 static __inline uint32_t
386 moea_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
387 {
388 	uint32_t pte_lo;
389 	int i;
390 
391 	if (ma != VM_MEMATTR_DEFAULT) {
392 		switch (ma) {
393 		case VM_MEMATTR_UNCACHEABLE:
394 			return (PTE_I | PTE_G);
395 		case VM_MEMATTR_CACHEABLE:
396 			return (PTE_M);
397 		case VM_MEMATTR_WRITE_COMBINING:
398 		case VM_MEMATTR_WRITE_BACK:
399 		case VM_MEMATTR_PREFETCHABLE:
400 			return (PTE_I);
401 		case VM_MEMATTR_WRITE_THROUGH:
402 			return (PTE_W | PTE_M);
403 		}
404 	}
405 
406 	/*
407 	 * Assume the page is cache inhibited and access is guarded unless
408 	 * it's in our available memory array.
409 	 */
410 	pte_lo = PTE_I | PTE_G;
411 	for (i = 0; i < pregions_sz; i++) {
412 		if ((pa >= pregions[i].mr_start) &&
413 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
414 			pte_lo = PTE_M;
415 			break;
416 		}
417 	}
418 
419 	return pte_lo;
420 }
421 
422 /*
423  * Translate OFW translations into VM attributes.
424  */
425 static __inline vm_memattr_t
426 moea_bootstrap_convert_wimg(uint32_t mode)
427 {
428 
429 	switch (mode) {
430 	case (PTE_I | PTE_G):
431 		/* PCI device memory */
432 		return VM_MEMATTR_UNCACHEABLE;
433 	case (PTE_M):
434 		/* Explicitly coherent */
435 		return VM_MEMATTR_CACHEABLE;
436 	case 0: /* Default claim */
437 	case 2: /* Alternate PP bits set by OF for the original payload */
438 		/* "Normal" memory. */
439 		return VM_MEMATTR_DEFAULT;
440 
441 	default:
442 		/* Err on the side of caution for unknowns */
443 		/* XXX should we panic instead? */
444 		return VM_MEMATTR_UNCACHEABLE;
445 	}
446 }
447 
448 static void
449 tlbie(vm_offset_t va)
450 {
451 
452 	mtx_lock_spin(&tlbie_mtx);
453 	__asm __volatile("ptesync");
454 	__asm __volatile("tlbie %0" :: "r"(va));
455 	__asm __volatile("eieio; tlbsync; ptesync");
456 	mtx_unlock_spin(&tlbie_mtx);
457 }
458 
459 static void
460 tlbia(void)
461 {
462 	vm_offset_t va;
463 
464 	for (va = 0; va < 0x00040000; va += 0x00001000) {
465 		__asm __volatile("tlbie %0" :: "r"(va));
466 		powerpc_sync();
467 	}
468 	__asm __volatile("tlbsync");
469 	powerpc_sync();
470 }
471 
472 static __inline int
473 va_to_sr(u_int *sr, vm_offset_t va)
474 {
475 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
476 }
477 
478 static __inline u_int
479 va_to_pteg(u_int sr, vm_offset_t addr)
480 {
481 	u_int hash;
482 
483 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
484 	    ADDR_PIDX_SHFT);
485 	return (hash & moea_pteg_mask);
486 }
487 
488 static __inline struct pvo_head *
489 vm_page_to_pvoh(vm_page_t m)
490 {
491 
492 	return (&m->md.mdpg_pvoh);
493 }
494 
495 static __inline void
496 moea_attr_clear(vm_page_t m, int ptebit)
497 {
498 
499 	rw_assert(&pvh_global_lock, RA_WLOCKED);
500 	m->md.mdpg_attrs &= ~ptebit;
501 }
502 
503 static __inline int
504 moea_attr_fetch(vm_page_t m)
505 {
506 
507 	return (m->md.mdpg_attrs);
508 }
509 
510 static __inline void
511 moea_attr_save(vm_page_t m, int ptebit)
512 {
513 
514 	rw_assert(&pvh_global_lock, RA_WLOCKED);
515 	m->md.mdpg_attrs |= ptebit;
516 }
517 
518 static __inline int
519 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
520 {
521 	if (pt->pte_hi == pvo_pt->pte_hi)
522 		return (1);
523 
524 	return (0);
525 }
526 
527 static __inline int
528 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
529 {
530 	return (pt->pte_hi & ~PTE_VALID) ==
531 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
532 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
533 }
534 
535 static __inline void
536 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
537 {
538 
539 	mtx_assert(&moea_table_mutex, MA_OWNED);
540 
541 	/*
542 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
543 	 * set when the real pte is set in memory.
544 	 *
545 	 * Note: Don't set the valid bit for correct operation of tlb update.
546 	 */
547 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
548 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
549 	pt->pte_lo = pte_lo;
550 }
551 
552 static __inline void
553 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
554 {
555 
556 	mtx_assert(&moea_table_mutex, MA_OWNED);
557 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
558 }
559 
560 static __inline void
561 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
562 {
563 
564 	mtx_assert(&moea_table_mutex, MA_OWNED);
565 
566 	/*
567 	 * As shown in Section 7.6.3.2.3
568 	 */
569 	pt->pte_lo &= ~ptebit;
570 	tlbie(va);
571 }
572 
573 static __inline void
574 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
575 {
576 
577 	mtx_assert(&moea_table_mutex, MA_OWNED);
578 	pvo_pt->pte_hi |= PTE_VALID;
579 
580 	/*
581 	 * Update the PTE as defined in section 7.6.3.1.
582 	 * Note that the REF/CHG bits are from pvo_pt and thus should have
583 	 * been saved so this routine can restore them (if desired).
584 	 */
585 	pt->pte_lo = pvo_pt->pte_lo;
586 	powerpc_sync();
587 	pt->pte_hi = pvo_pt->pte_hi;
588 	powerpc_sync();
589 	moea_pte_valid++;
590 }
591 
592 static __inline void
593 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
594 {
595 
596 	mtx_assert(&moea_table_mutex, MA_OWNED);
597 	pvo_pt->pte_hi &= ~PTE_VALID;
598 
599 	/*
600 	 * Force the reg & chg bits back into the PTEs.
601 	 */
602 	powerpc_sync();
603 
604 	/*
605 	 * Invalidate the pte.
606 	 */
607 	pt->pte_hi &= ~PTE_VALID;
608 
609 	tlbie(va);
610 
611 	/*
612 	 * Save the reg & chg bits.
613 	 */
614 	moea_pte_synch(pt, pvo_pt);
615 	moea_pte_valid--;
616 }
617 
618 static __inline void
619 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
620 {
621 
622 	/*
623 	 * Invalidate the PTE
624 	 */
625 	moea_pte_unset(pt, pvo_pt, va);
626 	moea_pte_set(pt, pvo_pt);
627 }
628 
629 /*
630  * Quick sort callout for comparing memory regions.
631  */
632 static int	om_cmp(const void *a, const void *b);
633 
634 static int
635 om_cmp(const void *a, const void *b)
636 {
637 	const struct	ofw_map *mapa;
638 	const struct	ofw_map *mapb;
639 
640 	mapa = a;
641 	mapb = b;
642 	if (mapa->om_pa < mapb->om_pa)
643 		return (-1);
644 	else if (mapa->om_pa > mapb->om_pa)
645 		return (1);
646 	else
647 		return (0);
648 }
649 
650 void
651 moea_cpu_bootstrap(int ap)
652 {
653 	u_int sdr;
654 	int i;
655 
656 	if (ap) {
657 		powerpc_sync();
658 		__asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu));
659 		__asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl));
660 		isync();
661 		__asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu));
662 		__asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl));
663 		isync();
664 	}
665 
666 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
667 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
668 	isync();
669 
670 	__asm __volatile("mtibatu 1,%0" :: "r"(0));
671 	__asm __volatile("mtdbatu 2,%0" :: "r"(0));
672 	__asm __volatile("mtibatu 2,%0" :: "r"(0));
673 	__asm __volatile("mtdbatu 3,%0" :: "r"(0));
674 	__asm __volatile("mtibatu 3,%0" :: "r"(0));
675 	isync();
676 
677 	for (i = 0; i < 16; i++)
678 		mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
679 	powerpc_sync();
680 
681 	sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10);
682 	__asm __volatile("mtsdr1 %0" :: "r"(sdr));
683 	isync();
684 
685 	tlbia();
686 }
687 
688 void
689 moea_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
690 {
691 	ihandle_t	mmui;
692 	phandle_t	chosen, mmu;
693 	int		sz;
694 	int		i, j;
695 	vm_size_t	size, physsz, hwphyssz;
696 	vm_offset_t	pa, va, off;
697 	void		*dpcpu;
698 
699 	/*
700 	 * Map PCI memory space.
701 	 */
702 	battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
703 	battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
704 
705 	battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
706 	battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
707 
708 	battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
709 	battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
710 
711 	battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
712 	battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
713 
714 	powerpc_sync();
715 
716 	/* map pci space */
717 	__asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
718 	__asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
719 	isync();
720 
721 	/* set global direct map flag */
722 	hw_direct_map = 1;
723 
724 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
725 	CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
726 
727 	for (i = 0; i < pregions_sz; i++) {
728 		vm_offset_t pa;
729 		vm_offset_t end;
730 
731 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
732 			pregions[i].mr_start,
733 			pregions[i].mr_start + pregions[i].mr_size,
734 			pregions[i].mr_size);
735 		/*
736 		 * Install entries into the BAT table to allow all
737 		 * of physmem to be convered by on-demand BAT entries.
738 		 * The loop will sometimes set the same battable element
739 		 * twice, but that's fine since they won't be used for
740 		 * a while yet.
741 		 */
742 		pa = pregions[i].mr_start & 0xf0000000;
743 		end = pregions[i].mr_start + pregions[i].mr_size;
744 		do {
745                         u_int n = pa >> ADDR_SR_SHFT;
746 
747 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
748 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
749 			pa += SEGMENT_LENGTH;
750 		} while (pa < end);
751 	}
752 
753 	if (PHYS_AVAIL_ENTRIES < regions_sz)
754 		panic("moea_bootstrap: phys_avail too small");
755 
756 	phys_avail_count = 0;
757 	physsz = 0;
758 	hwphyssz = 0;
759 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
760 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
761 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
762 		    regions[i].mr_start + regions[i].mr_size,
763 		    regions[i].mr_size);
764 		if (hwphyssz != 0 &&
765 		    (physsz + regions[i].mr_size) >= hwphyssz) {
766 			if (physsz < hwphyssz) {
767 				phys_avail[j] = regions[i].mr_start;
768 				phys_avail[j + 1] = regions[i].mr_start +
769 				    hwphyssz - physsz;
770 				physsz = hwphyssz;
771 				phys_avail_count++;
772 			}
773 			break;
774 		}
775 		phys_avail[j] = regions[i].mr_start;
776 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
777 		phys_avail_count++;
778 		physsz += regions[i].mr_size;
779 	}
780 
781 	/* Check for overlap with the kernel and exception vectors */
782 	for (j = 0; j < 2*phys_avail_count; j+=2) {
783 		if (phys_avail[j] < EXC_LAST)
784 			phys_avail[j] += EXC_LAST;
785 
786 		if (kernelstart >= phys_avail[j] &&
787 		    kernelstart < phys_avail[j+1]) {
788 			if (kernelend < phys_avail[j+1]) {
789 				phys_avail[2*phys_avail_count] =
790 				    (kernelend & ~PAGE_MASK) + PAGE_SIZE;
791 				phys_avail[2*phys_avail_count + 1] =
792 				    phys_avail[j+1];
793 				phys_avail_count++;
794 			}
795 
796 			phys_avail[j+1] = kernelstart & ~PAGE_MASK;
797 		}
798 
799 		if (kernelend >= phys_avail[j] &&
800 		    kernelend < phys_avail[j+1]) {
801 			if (kernelstart > phys_avail[j]) {
802 				phys_avail[2*phys_avail_count] = phys_avail[j];
803 				phys_avail[2*phys_avail_count + 1] =
804 				    kernelstart & ~PAGE_MASK;
805 				phys_avail_count++;
806 			}
807 
808 			phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
809 		}
810 	}
811 
812 	physmem = btoc(physsz);
813 
814 	/*
815 	 * Allocate PTEG table.
816 	 */
817 #ifdef PTEGCOUNT
818 	moea_pteg_count = PTEGCOUNT;
819 #else
820 	moea_pteg_count = 0x1000;
821 
822 	while (moea_pteg_count < physmem)
823 		moea_pteg_count <<= 1;
824 
825 	moea_pteg_count >>= 1;
826 #endif /* PTEGCOUNT */
827 
828 	size = moea_pteg_count * sizeof(struct pteg);
829 	CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
830 	    size);
831 	moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
832 	CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
833 	bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
834 	moea_pteg_mask = moea_pteg_count - 1;
835 
836 	/*
837 	 * Allocate pv/overflow lists.
838 	 */
839 	size = sizeof(struct pvo_head) * moea_pteg_count;
840 	moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
841 	    PAGE_SIZE);
842 	CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
843 	for (i = 0; i < moea_pteg_count; i++)
844 		LIST_INIT(&moea_pvo_table[i]);
845 
846 	/*
847 	 * Initialize the lock that synchronizes access to the pteg and pvo
848 	 * tables.
849 	 */
850 	mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF |
851 	    MTX_RECURSE);
852 	mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF);
853 
854 	mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN);
855 
856 	/*
857 	 * Initialise the unmanaged pvo pool.
858 	 */
859 	moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
860 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
861 	moea_bpvo_pool_index = 0;
862 
863 	/*
864 	 * Make sure kernel vsid is allocated as well as VSID 0.
865 	 */
866 	moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
867 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
868 	moea_vsid_bitmap[0] |= 1;
869 
870 	/*
871 	 * Initialize the kernel pmap (which is statically allocated).
872 	 */
873 	PMAP_LOCK_INIT(kernel_pmap);
874 	for (i = 0; i < 16; i++)
875 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
876 	CPU_FILL(&kernel_pmap->pm_active);
877 	RB_INIT(&kernel_pmap->pmap_pvo);
878 
879  	/*
880 	 * Initialize the global pv list lock.
881 	 */
882 	rw_init(&pvh_global_lock, "pmap pv global");
883 
884 	/*
885 	 * Set up the Open Firmware mappings
886 	 */
887 	chosen = OF_finddevice("/chosen");
888 	if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1 &&
889 	    (mmu = OF_instance_to_package(mmui)) != -1 &&
890 	    (sz = OF_getproplen(mmu, "translations")) != -1) {
891 		translations = NULL;
892 		for (i = 0; phys_avail[i] != 0; i += 2) {
893 			if (phys_avail[i + 1] >= sz) {
894 				translations = (struct ofw_map *)phys_avail[i];
895 				break;
896 			}
897 		}
898 		if (translations == NULL)
899 			panic("moea_bootstrap: no space to copy translations");
900 		bzero(translations, sz);
901 		if (OF_getprop(mmu, "translations", translations, sz) == -1)
902 			panic("moea_bootstrap: can't get ofw translations");
903 		CTR0(KTR_PMAP, "moea_bootstrap: translations");
904 		sz /= sizeof(*translations);
905 		qsort(translations, sz, sizeof (*translations), om_cmp);
906 		for (i = 0; i < sz; i++) {
907 			CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
908 			    translations[i].om_pa, translations[i].om_va,
909 			    translations[i].om_len);
910 
911 			/*
912 			 * If the mapping is 1:1, let the RAM and device
913 			 * on-demand BAT tables take care of the translation.
914 			 *
915 			 * However, always enter mappings for segment 16,
916 			 * which is mixed-protection and therefore not
917 			 * compatible with a BAT entry.
918 			 */
919 			if ((translations[i].om_va >> ADDR_SR_SHFT) != 0xf &&
920 				translations[i].om_va == translations[i].om_pa)
921 					continue;
922 
923 			/* Enter the pages */
924 			for (off = 0; off < translations[i].om_len;
925 			    off += PAGE_SIZE)
926 				moea_kenter_attr(translations[i].om_va + off,
927 				    translations[i].om_pa + off,
928 				    moea_bootstrap_convert_wimg(translations[i].om_mode));
929 		}
930 	}
931 
932 	/*
933 	 * Calculate the last available physical address.
934 	 */
935 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
936 		;
937 	Maxmem = powerpc_btop(phys_avail[i + 1]);
938 
939 	moea_cpu_bootstrap(0);
940 	mtmsr(mfmsr() | PSL_DR | PSL_IR);
941 	pmap_bootstrapped++;
942 
943 	/*
944 	 * Set the start and end of kva.
945 	 */
946 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
947 	virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
948 
949 	/*
950 	 * Allocate a kernel stack with a guard page for thread0 and map it
951 	 * into the kernel page map.
952 	 */
953 	pa = moea_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
954 	va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
955 	virtual_avail = va + kstack_pages * PAGE_SIZE;
956 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
957 	thread0.td_kstack = va;
958 	thread0.td_kstack_pages = kstack_pages;
959 	for (i = 0; i < kstack_pages; i++) {
960 		moea_kenter(va, pa);
961 		pa += PAGE_SIZE;
962 		va += PAGE_SIZE;
963 	}
964 
965 	/*
966 	 * Allocate virtual address space for the message buffer.
967 	 */
968 	pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE);
969 	msgbufp = (struct msgbuf *)virtual_avail;
970 	va = virtual_avail;
971 	virtual_avail += round_page(msgbufsize);
972 	while (va < virtual_avail) {
973 		moea_kenter(va, pa);
974 		pa += PAGE_SIZE;
975 		va += PAGE_SIZE;
976 	}
977 
978 	/*
979 	 * Allocate virtual address space for the dynamic percpu area.
980 	 */
981 	pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
982 	dpcpu = (void *)virtual_avail;
983 	va = virtual_avail;
984 	virtual_avail += DPCPU_SIZE;
985 	while (va < virtual_avail) {
986 		moea_kenter(va, pa);
987 		pa += PAGE_SIZE;
988 		va += PAGE_SIZE;
989 	}
990 	dpcpu_init(dpcpu, 0);
991 }
992 
993 /*
994  * Activate a user pmap.  The pmap must be activated before it's address
995  * space can be accessed in any way.
996  */
997 void
998 moea_activate(struct thread *td)
999 {
1000 	pmap_t	pm, pmr;
1001 
1002 	/*
1003 	 * Load all the data we need up front to encourage the compiler to
1004 	 * not issue any loads while we have interrupts disabled below.
1005 	 */
1006 	pm = &td->td_proc->p_vmspace->vm_pmap;
1007 	pmr = pm->pmap_phys;
1008 
1009 	CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1010 	PCPU_SET(curpmap, pmr);
1011 
1012 	mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1013 }
1014 
1015 void
1016 moea_deactivate(struct thread *td)
1017 {
1018 	pmap_t	pm;
1019 
1020 	pm = &td->td_proc->p_vmspace->vm_pmap;
1021 	CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1022 	PCPU_SET(curpmap, NULL);
1023 }
1024 
1025 void
1026 moea_unwire(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1027 {
1028 	struct	pvo_entry key, *pvo;
1029 
1030 	PMAP_LOCK(pm);
1031 	key.pvo_vaddr = sva;
1032 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1033 	    pvo != NULL && PVO_VADDR(pvo) < eva;
1034 	    pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1035 		if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1036 			panic("moea_unwire: pvo %p is missing PVO_WIRED", pvo);
1037 		pvo->pvo_vaddr &= ~PVO_WIRED;
1038 		pm->pm_stats.wired_count--;
1039 	}
1040 	PMAP_UNLOCK(pm);
1041 }
1042 
1043 void
1044 moea_copy_page(vm_page_t msrc, vm_page_t mdst)
1045 {
1046 	vm_offset_t	dst;
1047 	vm_offset_t	src;
1048 
1049 	dst = VM_PAGE_TO_PHYS(mdst);
1050 	src = VM_PAGE_TO_PHYS(msrc);
1051 
1052 	bcopy((void *)src, (void *)dst, PAGE_SIZE);
1053 }
1054 
1055 void
1056 moea_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
1057     vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1058 {
1059 	void *a_cp, *b_cp;
1060 	vm_offset_t a_pg_offset, b_pg_offset;
1061 	int cnt;
1062 
1063 	while (xfersize > 0) {
1064 		a_pg_offset = a_offset & PAGE_MASK;
1065 		cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1066 		a_cp = (char *)VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]) +
1067 		    a_pg_offset;
1068 		b_pg_offset = b_offset & PAGE_MASK;
1069 		cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1070 		b_cp = (char *)VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]) +
1071 		    b_pg_offset;
1072 		bcopy(a_cp, b_cp, cnt);
1073 		a_offset += cnt;
1074 		b_offset += cnt;
1075 		xfersize -= cnt;
1076 	}
1077 }
1078 
1079 /*
1080  * Zero a page of physical memory by temporarily mapping it into the tlb.
1081  */
1082 void
1083 moea_zero_page(vm_page_t m)
1084 {
1085 	vm_offset_t off, pa = VM_PAGE_TO_PHYS(m);
1086 
1087 	for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1088 		__asm __volatile("dcbz 0,%0" :: "r"(pa + off));
1089 }
1090 
1091 void
1092 moea_zero_page_area(vm_page_t m, int off, int size)
1093 {
1094 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1095 	void *va = (void *)(pa + off);
1096 
1097 	bzero(va, size);
1098 }
1099 
1100 vm_offset_t
1101 moea_quick_enter_page(vm_page_t m)
1102 {
1103 
1104 	return (VM_PAGE_TO_PHYS(m));
1105 }
1106 
1107 void
1108 moea_quick_remove_page(vm_offset_t addr)
1109 {
1110 }
1111 
1112 boolean_t
1113 moea_page_is_mapped(vm_page_t m)
1114 {
1115 	return (!LIST_EMPTY(&(m)->md.mdpg_pvoh));
1116 }
1117 
1118 /*
1119  * Map the given physical page at the specified virtual address in the
1120  * target pmap with the protection requested.  If specified the page
1121  * will be wired down.
1122  */
1123 int
1124 moea_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1125     u_int flags, int8_t psind)
1126 {
1127 	int error;
1128 
1129 	for (;;) {
1130 		rw_wlock(&pvh_global_lock);
1131 		PMAP_LOCK(pmap);
1132 		error = moea_enter_locked(pmap, va, m, prot, flags, psind);
1133 		rw_wunlock(&pvh_global_lock);
1134 		PMAP_UNLOCK(pmap);
1135 		if (error != ENOMEM)
1136 			return (KERN_SUCCESS);
1137 		if ((flags & PMAP_ENTER_NOSLEEP) != 0)
1138 			return (KERN_RESOURCE_SHORTAGE);
1139 		VM_OBJECT_ASSERT_UNLOCKED(m->object);
1140 		vm_wait(NULL);
1141 	}
1142 }
1143 
1144 /*
1145  * Map the given physical page at the specified virtual address in the
1146  * target pmap with the protection requested.  If specified the page
1147  * will be wired down.
1148  *
1149  * The global pvh and pmap must be locked.
1150  */
1151 static int
1152 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1153     u_int flags, int8_t psind __unused)
1154 {
1155 	struct		pvo_head *pvo_head;
1156 	uma_zone_t	zone;
1157 	u_int		pte_lo, pvo_flags;
1158 	int		error;
1159 
1160 	if (pmap_bootstrapped)
1161 		rw_assert(&pvh_global_lock, RA_WLOCKED);
1162 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1163 	if ((m->oflags & VPO_UNMANAGED) == 0) {
1164 		if ((flags & PMAP_ENTER_QUICK_LOCKED) == 0)
1165 			VM_PAGE_OBJECT_BUSY_ASSERT(m);
1166 		else
1167 			VM_OBJECT_ASSERT_LOCKED(m->object);
1168 	}
1169 
1170 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea_initialized) {
1171 		pvo_head = &moea_pvo_kunmanaged;
1172 		zone = moea_upvo_zone;
1173 		pvo_flags = 0;
1174 	} else {
1175 		pvo_head = vm_page_to_pvoh(m);
1176 		zone = moea_mpvo_zone;
1177 		pvo_flags = PVO_MANAGED;
1178 	}
1179 
1180 	pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1181 
1182 	if (prot & VM_PROT_WRITE) {
1183 		pte_lo |= PTE_BW;
1184 		if (pmap_bootstrapped &&
1185 		    (m->oflags & VPO_UNMANAGED) == 0)
1186 			vm_page_aflag_set(m, PGA_WRITEABLE);
1187 	} else
1188 		pte_lo |= PTE_BR;
1189 
1190 	if ((flags & PMAP_ENTER_WIRED) != 0)
1191 		pvo_flags |= PVO_WIRED;
1192 
1193 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1194 	    pte_lo, pvo_flags);
1195 
1196 	/*
1197 	 * Flush the real page from the instruction cache. This has be done
1198 	 * for all user mappings to prevent information leakage via the
1199 	 * instruction cache. moea_pvo_enter() returns ENOENT for the first
1200 	 * mapping for a page.
1201 	 */
1202 	if (pmap != kernel_pmap && error == ENOENT &&
1203 	    (pte_lo & (PTE_I | PTE_G)) == 0)
1204 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1205 
1206 	return (error);
1207 }
1208 
1209 /*
1210  * Maps a sequence of resident pages belonging to the same object.
1211  * The sequence begins with the given page m_start.  This page is
1212  * mapped at the given virtual address start.  Each subsequent page is
1213  * mapped at a virtual address that is offset from start by the same
1214  * amount as the page is offset from m_start within the object.  The
1215  * last page in the sequence is the page with the largest offset from
1216  * m_start that can be mapped at a virtual address less than the given
1217  * virtual address end.  Not every virtual page between start and end
1218  * is mapped; only those for which a resident page exists with the
1219  * corresponding offset from m_start are mapped.
1220  */
1221 void
1222 moea_enter_object(pmap_t pm, vm_offset_t start, vm_offset_t end,
1223     vm_page_t m_start, vm_prot_t prot)
1224 {
1225 	vm_page_t m;
1226 	vm_pindex_t diff, psize;
1227 
1228 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1229 
1230 	psize = atop(end - start);
1231 	m = m_start;
1232 	rw_wlock(&pvh_global_lock);
1233 	PMAP_LOCK(pm);
1234 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1235 		moea_enter_locked(pm, start + ptoa(diff), m, prot &
1236 		    (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_QUICK_LOCKED,
1237 		    0);
1238 		m = TAILQ_NEXT(m, listq);
1239 	}
1240 	rw_wunlock(&pvh_global_lock);
1241 	PMAP_UNLOCK(pm);
1242 }
1243 
1244 void
1245 moea_enter_quick(pmap_t pm, vm_offset_t va, vm_page_t m,
1246     vm_prot_t prot)
1247 {
1248 
1249 	rw_wlock(&pvh_global_lock);
1250 	PMAP_LOCK(pm);
1251 	moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1252 	    PMAP_ENTER_QUICK_LOCKED, 0);
1253 	rw_wunlock(&pvh_global_lock);
1254 	PMAP_UNLOCK(pm);
1255 }
1256 
1257 vm_paddr_t
1258 moea_extract(pmap_t pm, vm_offset_t va)
1259 {
1260 	struct	pvo_entry *pvo;
1261 	vm_paddr_t pa;
1262 
1263 	PMAP_LOCK(pm);
1264 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1265 	if (pvo == NULL)
1266 		pa = 0;
1267 	else
1268 		pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1269 	PMAP_UNLOCK(pm);
1270 	return (pa);
1271 }
1272 
1273 /*
1274  * Atomically extract and hold the physical page with the given
1275  * pmap and virtual address pair if that mapping permits the given
1276  * protection.
1277  */
1278 vm_page_t
1279 moea_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1280 {
1281 	struct	pvo_entry *pvo;
1282 	vm_page_t m;
1283 
1284 	m = NULL;
1285 	PMAP_LOCK(pmap);
1286 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1287 	if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1288 	    ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1289 	     (prot & VM_PROT_WRITE) == 0)) {
1290 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
1291 		if (!vm_page_wire_mapped(m))
1292 			m = NULL;
1293 	}
1294 	PMAP_UNLOCK(pmap);
1295 	return (m);
1296 }
1297 
1298 void
1299 moea_init()
1300 {
1301 
1302 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1303 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1304 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1305 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1306 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1307 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1308 	moea_initialized = TRUE;
1309 }
1310 
1311 boolean_t
1312 moea_is_referenced(vm_page_t m)
1313 {
1314 	boolean_t rv;
1315 
1316 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1317 	    ("moea_is_referenced: page %p is not managed", m));
1318 	rw_wlock(&pvh_global_lock);
1319 	rv = moea_query_bit(m, PTE_REF);
1320 	rw_wunlock(&pvh_global_lock);
1321 	return (rv);
1322 }
1323 
1324 boolean_t
1325 moea_is_modified(vm_page_t m)
1326 {
1327 	boolean_t rv;
1328 
1329 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1330 	    ("moea_is_modified: page %p is not managed", m));
1331 
1332 	/*
1333 	 * If the page is not busied then this check is racy.
1334 	 */
1335 	if (!pmap_page_is_write_mapped(m))
1336 		return (FALSE);
1337 
1338 	rw_wlock(&pvh_global_lock);
1339 	rv = moea_query_bit(m, PTE_CHG);
1340 	rw_wunlock(&pvh_global_lock);
1341 	return (rv);
1342 }
1343 
1344 boolean_t
1345 moea_is_prefaultable(pmap_t pmap, vm_offset_t va)
1346 {
1347 	struct pvo_entry *pvo;
1348 	boolean_t rv;
1349 
1350 	PMAP_LOCK(pmap);
1351 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1352 	rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1353 	PMAP_UNLOCK(pmap);
1354 	return (rv);
1355 }
1356 
1357 void
1358 moea_clear_modify(vm_page_t m)
1359 {
1360 
1361 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1362 	    ("moea_clear_modify: page %p is not managed", m));
1363 	vm_page_assert_busied(m);
1364 
1365 	if (!pmap_page_is_write_mapped(m))
1366 		return;
1367 	rw_wlock(&pvh_global_lock);
1368 	moea_clear_bit(m, PTE_CHG);
1369 	rw_wunlock(&pvh_global_lock);
1370 }
1371 
1372 /*
1373  * Clear the write and modified bits in each of the given page's mappings.
1374  */
1375 void
1376 moea_remove_write(vm_page_t m)
1377 {
1378 	struct	pvo_entry *pvo;
1379 	struct	pte *pt;
1380 	pmap_t	pmap;
1381 	u_int	lo;
1382 
1383 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1384 	    ("moea_remove_write: page %p is not managed", m));
1385 	vm_page_assert_busied(m);
1386 
1387 	if (!pmap_page_is_write_mapped(m))
1388 		return;
1389 	rw_wlock(&pvh_global_lock);
1390 	lo = moea_attr_fetch(m);
1391 	powerpc_sync();
1392 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1393 		pmap = pvo->pvo_pmap;
1394 		PMAP_LOCK(pmap);
1395 		if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1396 			pt = moea_pvo_to_pte(pvo, -1);
1397 			pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1398 			pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1399 			if (pt != NULL) {
1400 				moea_pte_synch(pt, &pvo->pvo_pte.pte);
1401 				lo |= pvo->pvo_pte.pte.pte_lo;
1402 				pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1403 				moea_pte_change(pt, &pvo->pvo_pte.pte,
1404 				    pvo->pvo_vaddr);
1405 				mtx_unlock(&moea_table_mutex);
1406 			}
1407 		}
1408 		PMAP_UNLOCK(pmap);
1409 	}
1410 	if ((lo & PTE_CHG) != 0) {
1411 		moea_attr_clear(m, PTE_CHG);
1412 		vm_page_dirty(m);
1413 	}
1414 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1415 	rw_wunlock(&pvh_global_lock);
1416 }
1417 
1418 /*
1419  *	moea_ts_referenced:
1420  *
1421  *	Return a count of reference bits for a page, clearing those bits.
1422  *	It is not necessary for every reference bit to be cleared, but it
1423  *	is necessary that 0 only be returned when there are truly no
1424  *	reference bits set.
1425  *
1426  *	XXX: The exact number of bits to check and clear is a matter that
1427  *	should be tested and standardized at some point in the future for
1428  *	optimal aging of shared pages.
1429  */
1430 int
1431 moea_ts_referenced(vm_page_t m)
1432 {
1433 	int count;
1434 
1435 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1436 	    ("moea_ts_referenced: page %p is not managed", m));
1437 	rw_wlock(&pvh_global_lock);
1438 	count = moea_clear_bit(m, PTE_REF);
1439 	rw_wunlock(&pvh_global_lock);
1440 	return (count);
1441 }
1442 
1443 /*
1444  * Modify the WIMG settings of all mappings for a page.
1445  */
1446 void
1447 moea_page_set_memattr(vm_page_t m, vm_memattr_t ma)
1448 {
1449 	struct	pvo_entry *pvo;
1450 	struct	pvo_head *pvo_head;
1451 	struct	pte *pt;
1452 	pmap_t	pmap;
1453 	u_int	lo;
1454 
1455 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1456 		m->md.mdpg_cache_attrs = ma;
1457 		return;
1458 	}
1459 
1460 	rw_wlock(&pvh_global_lock);
1461 	pvo_head = vm_page_to_pvoh(m);
1462 	lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1463 
1464 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1465 		pmap = pvo->pvo_pmap;
1466 		PMAP_LOCK(pmap);
1467 		pt = moea_pvo_to_pte(pvo, -1);
1468 		pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1469 		pvo->pvo_pte.pte.pte_lo |= lo;
1470 		if (pt != NULL) {
1471 			moea_pte_change(pt, &pvo->pvo_pte.pte,
1472 			    pvo->pvo_vaddr);
1473 			if (pvo->pvo_pmap == kernel_pmap)
1474 				isync();
1475 		}
1476 		mtx_unlock(&moea_table_mutex);
1477 		PMAP_UNLOCK(pmap);
1478 	}
1479 	m->md.mdpg_cache_attrs = ma;
1480 	rw_wunlock(&pvh_global_lock);
1481 }
1482 
1483 /*
1484  * Map a wired page into kernel virtual address space.
1485  */
1486 void
1487 moea_kenter(vm_offset_t va, vm_paddr_t pa)
1488 {
1489 
1490 	moea_kenter_attr(va, pa, VM_MEMATTR_DEFAULT);
1491 }
1492 
1493 void
1494 moea_kenter_attr(vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
1495 {
1496 	u_int		pte_lo;
1497 	int		error;
1498 
1499 #if 0
1500 	if (va < VM_MIN_KERNEL_ADDRESS)
1501 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1502 		    va);
1503 #endif
1504 
1505 	pte_lo = moea_calc_wimg(pa, ma);
1506 
1507 	PMAP_LOCK(kernel_pmap);
1508 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1509 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1510 
1511 	if (error != 0 && error != ENOENT)
1512 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1513 		    pa, error);
1514 
1515 	PMAP_UNLOCK(kernel_pmap);
1516 }
1517 
1518 /*
1519  * Extract the physical page address associated with the given kernel virtual
1520  * address.
1521  */
1522 vm_paddr_t
1523 moea_kextract(vm_offset_t va)
1524 {
1525 	struct		pvo_entry *pvo;
1526 	vm_paddr_t pa;
1527 
1528 	/*
1529 	 * Allow direct mappings on 32-bit OEA
1530 	 */
1531 	if (va < VM_MIN_KERNEL_ADDRESS) {
1532 		return (va);
1533 	}
1534 
1535 	PMAP_LOCK(kernel_pmap);
1536 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1537 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1538 	pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1539 	PMAP_UNLOCK(kernel_pmap);
1540 	return (pa);
1541 }
1542 
1543 /*
1544  * Remove a wired page from kernel virtual address space.
1545  */
1546 void
1547 moea_kremove(vm_offset_t va)
1548 {
1549 
1550 	moea_remove(kernel_pmap, va, va + PAGE_SIZE);
1551 }
1552 
1553 /*
1554  * Provide a kernel pointer corresponding to a given userland pointer.
1555  * The returned pointer is valid until the next time this function is
1556  * called in this thread. This is used internally in copyin/copyout.
1557  */
1558 int
1559 moea_map_user_ptr(pmap_t pm, volatile const void *uaddr,
1560     void **kaddr, size_t ulen, size_t *klen)
1561 {
1562 	size_t l;
1563 	register_t vsid;
1564 
1565 	*kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
1566 	l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
1567 	if (l > ulen)
1568 		l = ulen;
1569 	if (klen)
1570 		*klen = l;
1571 	else if (l != ulen)
1572 		return (EFAULT);
1573 
1574 	vsid = va_to_vsid(pm, (vm_offset_t)uaddr);
1575 
1576 	/* Mark segment no-execute */
1577 	vsid |= SR_N;
1578 
1579 	/* If we have already set this VSID, we can just return */
1580 	if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == vsid)
1581 		return (0);
1582 
1583 	__asm __volatile("isync");
1584 	curthread->td_pcb->pcb_cpu.aim.usr_segm =
1585 	    (uintptr_t)uaddr >> ADDR_SR_SHFT;
1586 	curthread->td_pcb->pcb_cpu.aim.usr_vsid = vsid;
1587 	__asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(vsid));
1588 
1589 	return (0);
1590 }
1591 
1592 /*
1593  * Figure out where a given kernel pointer (usually in a fault) points
1594  * to from the VM's perspective, potentially remapping into userland's
1595  * address space.
1596  */
1597 static int
1598 moea_decode_kernel_ptr(vm_offset_t addr, int *is_user,
1599     vm_offset_t *decoded_addr)
1600 {
1601 	vm_offset_t user_sr;
1602 
1603 	if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
1604 		user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
1605 		addr &= ADDR_PIDX | ADDR_POFF;
1606 		addr |= user_sr << ADDR_SR_SHFT;
1607 		*decoded_addr = addr;
1608 		*is_user = 1;
1609 	} else {
1610 		*decoded_addr = addr;
1611 		*is_user = 0;
1612 	}
1613 
1614 	return (0);
1615 }
1616 
1617 /*
1618  * Map a range of physical addresses into kernel virtual address space.
1619  *
1620  * The value passed in *virt is a suggested virtual address for the mapping.
1621  * Architectures which can support a direct-mapped physical to virtual region
1622  * can return the appropriate address within that region, leaving '*virt'
1623  * unchanged.  We cannot and therefore do not; *virt is updated with the
1624  * first usable address after the mapped region.
1625  */
1626 vm_offset_t
1627 moea_map(vm_offset_t *virt, vm_paddr_t pa_start,
1628     vm_paddr_t pa_end, int prot)
1629 {
1630 	vm_offset_t	sva, va;
1631 
1632 	sva = *virt;
1633 	va = sva;
1634 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1635 		moea_kenter(va, pa_start);
1636 	*virt = va;
1637 	return (sva);
1638 }
1639 
1640 /*
1641  * Returns true if the pmap's pv is one of the first
1642  * 16 pvs linked to from this page.  This count may
1643  * be changed upwards or downwards in the future; it
1644  * is only necessary that true be returned for a small
1645  * subset of pmaps for proper page aging.
1646  */
1647 boolean_t
1648 moea_page_exists_quick(pmap_t pmap, vm_page_t m)
1649 {
1650         int loops;
1651 	struct pvo_entry *pvo;
1652 	boolean_t rv;
1653 
1654 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1655 	    ("moea_page_exists_quick: page %p is not managed", m));
1656 	loops = 0;
1657 	rv = FALSE;
1658 	rw_wlock(&pvh_global_lock);
1659 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1660 		if (pvo->pvo_pmap == pmap) {
1661 			rv = TRUE;
1662 			break;
1663 		}
1664 		if (++loops >= 16)
1665 			break;
1666 	}
1667 	rw_wunlock(&pvh_global_lock);
1668 	return (rv);
1669 }
1670 
1671 void
1672 moea_page_init(vm_page_t m)
1673 {
1674 
1675 	m->md.mdpg_attrs = 0;
1676 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
1677 	LIST_INIT(&m->md.mdpg_pvoh);
1678 }
1679 
1680 /*
1681  * Return the number of managed mappings to the given physical page
1682  * that are wired.
1683  */
1684 int
1685 moea_page_wired_mappings(vm_page_t m)
1686 {
1687 	struct pvo_entry *pvo;
1688 	int count;
1689 
1690 	count = 0;
1691 	if ((m->oflags & VPO_UNMANAGED) != 0)
1692 		return (count);
1693 	rw_wlock(&pvh_global_lock);
1694 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1695 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1696 			count++;
1697 	rw_wunlock(&pvh_global_lock);
1698 	return (count);
1699 }
1700 
1701 static u_int	moea_vsidcontext;
1702 
1703 int
1704 moea_pinit(pmap_t pmap)
1705 {
1706 	int	i, mask;
1707 	u_int	entropy;
1708 
1709 	RB_INIT(&pmap->pmap_pvo);
1710 
1711 	entropy = 0;
1712 	__asm __volatile("mftb %0" : "=r"(entropy));
1713 
1714 	if ((pmap->pmap_phys = (pmap_t)moea_kextract((vm_offset_t)pmap))
1715 	    == NULL) {
1716 		pmap->pmap_phys = pmap;
1717 	}
1718 
1719 	mtx_lock(&moea_vsid_mutex);
1720 	/*
1721 	 * Allocate some segment registers for this pmap.
1722 	 */
1723 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1724 		u_int	hash, n;
1725 
1726 		/*
1727 		 * Create a new value by mutiplying by a prime and adding in
1728 		 * entropy from the timebase register.  This is to make the
1729 		 * VSID more random so that the PT hash function collides
1730 		 * less often.  (Note that the prime casues gcc to do shifts
1731 		 * instead of a multiply.)
1732 		 */
1733 		moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1734 		hash = moea_vsidcontext & (NPMAPS - 1);
1735 		if (hash == 0)		/* 0 is special, avoid it */
1736 			continue;
1737 		n = hash >> 5;
1738 		mask = 1 << (hash & (VSID_NBPW - 1));
1739 		hash = (moea_vsidcontext & 0xfffff);
1740 		if (moea_vsid_bitmap[n] & mask) {	/* collision? */
1741 			/* anything free in this bucket? */
1742 			if (moea_vsid_bitmap[n] == 0xffffffff) {
1743 				entropy = (moea_vsidcontext >> 20);
1744 				continue;
1745 			}
1746 			i = ffs(~moea_vsid_bitmap[n]) - 1;
1747 			mask = 1 << i;
1748 			hash &= rounddown2(0xfffff, VSID_NBPW);
1749 			hash |= i;
1750 		}
1751 		KASSERT(!(moea_vsid_bitmap[n] & mask),
1752 		    ("Allocating in-use VSID group %#x\n", hash));
1753 		moea_vsid_bitmap[n] |= mask;
1754 		for (i = 0; i < 16; i++)
1755 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1756 		mtx_unlock(&moea_vsid_mutex);
1757 		return (1);
1758 	}
1759 
1760 	mtx_unlock(&moea_vsid_mutex);
1761 	panic("moea_pinit: out of segments");
1762 }
1763 
1764 /*
1765  * Initialize the pmap associated with process 0.
1766  */
1767 void
1768 moea_pinit0(pmap_t pm)
1769 {
1770 
1771 	PMAP_LOCK_INIT(pm);
1772 	moea_pinit(pm);
1773 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1774 }
1775 
1776 /*
1777  * Set the physical protection on the specified range of this map as requested.
1778  */
1779 void
1780 moea_protect(pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1781     vm_prot_t prot)
1782 {
1783 	struct	pvo_entry *pvo, *tpvo, key;
1784 	struct	pte *pt;
1785 
1786 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1787 	    ("moea_protect: non current pmap"));
1788 
1789 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1790 		moea_remove(pm, sva, eva);
1791 		return;
1792 	}
1793 
1794 	rw_wlock(&pvh_global_lock);
1795 	PMAP_LOCK(pm);
1796 	key.pvo_vaddr = sva;
1797 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1798 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1799 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1800 
1801 		/*
1802 		 * Grab the PTE pointer before we diddle with the cached PTE
1803 		 * copy.
1804 		 */
1805 		pt = moea_pvo_to_pte(pvo, -1);
1806 		/*
1807 		 * Change the protection of the page.
1808 		 */
1809 		pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1810 		pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1811 
1812 		/*
1813 		 * If the PVO is in the page table, update that pte as well.
1814 		 */
1815 		if (pt != NULL) {
1816 			moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1817 			mtx_unlock(&moea_table_mutex);
1818 		}
1819 	}
1820 	rw_wunlock(&pvh_global_lock);
1821 	PMAP_UNLOCK(pm);
1822 }
1823 
1824 /*
1825  * Map a list of wired pages into kernel virtual address space.  This is
1826  * intended for temporary mappings which do not need page modification or
1827  * references recorded.  Existing mappings in the region are overwritten.
1828  */
1829 void
1830 moea_qenter(vm_offset_t sva, vm_page_t *m, int count)
1831 {
1832 	vm_offset_t va;
1833 
1834 	va = sva;
1835 	while (count-- > 0) {
1836 		moea_kenter(va, VM_PAGE_TO_PHYS(*m));
1837 		va += PAGE_SIZE;
1838 		m++;
1839 	}
1840 }
1841 
1842 /*
1843  * Remove page mappings from kernel virtual address space.  Intended for
1844  * temporary mappings entered by moea_qenter.
1845  */
1846 void
1847 moea_qremove(vm_offset_t sva, int count)
1848 {
1849 	vm_offset_t va;
1850 
1851 	va = sva;
1852 	while (count-- > 0) {
1853 		moea_kremove(va);
1854 		va += PAGE_SIZE;
1855 	}
1856 }
1857 
1858 void
1859 moea_release(pmap_t pmap)
1860 {
1861         int idx, mask;
1862 
1863 	/*
1864 	 * Free segment register's VSID
1865 	 */
1866         if (pmap->pm_sr[0] == 0)
1867                 panic("moea_release");
1868 
1869 	mtx_lock(&moea_vsid_mutex);
1870         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1871         mask = 1 << (idx % VSID_NBPW);
1872         idx /= VSID_NBPW;
1873         moea_vsid_bitmap[idx] &= ~mask;
1874 	mtx_unlock(&moea_vsid_mutex);
1875 }
1876 
1877 /*
1878  * Remove the given range of addresses from the specified map.
1879  */
1880 void
1881 moea_remove(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1882 {
1883 	struct	pvo_entry *pvo, *tpvo, key;
1884 
1885 	rw_wlock(&pvh_global_lock);
1886 	PMAP_LOCK(pm);
1887 	key.pvo_vaddr = sva;
1888 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1889 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1890 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1891 		moea_pvo_remove(pvo, -1);
1892 	}
1893 	PMAP_UNLOCK(pm);
1894 	rw_wunlock(&pvh_global_lock);
1895 }
1896 
1897 /*
1898  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1899  * will reflect changes in pte's back to the vm_page.
1900  */
1901 void
1902 moea_remove_all(vm_page_t m)
1903 {
1904 	struct  pvo_head *pvo_head;
1905 	struct	pvo_entry *pvo, *next_pvo;
1906 	pmap_t	pmap;
1907 
1908 	rw_wlock(&pvh_global_lock);
1909 	pvo_head = vm_page_to_pvoh(m);
1910 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1911 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1912 
1913 		pmap = pvo->pvo_pmap;
1914 		PMAP_LOCK(pmap);
1915 		moea_pvo_remove(pvo, -1);
1916 		PMAP_UNLOCK(pmap);
1917 	}
1918 	if ((m->a.flags & PGA_WRITEABLE) && moea_query_bit(m, PTE_CHG)) {
1919 		moea_attr_clear(m, PTE_CHG);
1920 		vm_page_dirty(m);
1921 	}
1922 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1923 	rw_wunlock(&pvh_global_lock);
1924 }
1925 
1926 /*
1927  * Allocate a physical page of memory directly from the phys_avail map.
1928  * Can only be called from moea_bootstrap before avail start and end are
1929  * calculated.
1930  */
1931 static vm_offset_t
1932 moea_bootstrap_alloc(vm_size_t size, u_int align)
1933 {
1934 	vm_offset_t	s, e;
1935 	int		i, j;
1936 
1937 	size = round_page(size);
1938 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1939 		if (align != 0)
1940 			s = roundup2(phys_avail[i], align);
1941 		else
1942 			s = phys_avail[i];
1943 		e = s + size;
1944 
1945 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1946 			continue;
1947 
1948 		if (s == phys_avail[i]) {
1949 			phys_avail[i] += size;
1950 		} else if (e == phys_avail[i + 1]) {
1951 			phys_avail[i + 1] -= size;
1952 		} else {
1953 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1954 				phys_avail[j] = phys_avail[j - 2];
1955 				phys_avail[j + 1] = phys_avail[j - 1];
1956 			}
1957 
1958 			phys_avail[i + 3] = phys_avail[i + 1];
1959 			phys_avail[i + 1] = s;
1960 			phys_avail[i + 2] = e;
1961 			phys_avail_count++;
1962 		}
1963 
1964 		return (s);
1965 	}
1966 	panic("moea_bootstrap_alloc: could not allocate memory");
1967 }
1968 
1969 static void
1970 moea_syncicache(vm_paddr_t pa, vm_size_t len)
1971 {
1972 	__syncicache((void *)pa, len);
1973 }
1974 
1975 static int
1976 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1977     vm_offset_t va, vm_paddr_t pa, u_int pte_lo, int flags)
1978 {
1979 	struct	pvo_entry *pvo;
1980 	u_int	sr;
1981 	int	first;
1982 	u_int	ptegidx;
1983 	int	i;
1984 	int     bootstrap;
1985 
1986 	moea_pvo_enter_calls++;
1987 	first = 0;
1988 	bootstrap = 0;
1989 
1990 	/*
1991 	 * Compute the PTE Group index.
1992 	 */
1993 	va &= ~ADDR_POFF;
1994 	sr = va_to_sr(pm->pm_sr, va);
1995 	ptegidx = va_to_pteg(sr, va);
1996 
1997 	/*
1998 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
1999 	 * there is a mapping.
2000 	 */
2001 	mtx_lock(&moea_table_mutex);
2002 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2003 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2004 			if ((pvo->pvo_pte.pte.pte_lo & PTE_RPGN) == pa &&
2005 			    (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
2006 			    (pte_lo & PTE_PP)) {
2007 				/*
2008 				 * The PTE is not changing.  Instead, this may
2009 				 * be a request to change the mapping's wired
2010 				 * attribute.
2011 				 */
2012 				mtx_unlock(&moea_table_mutex);
2013 				if ((flags & PVO_WIRED) != 0 &&
2014 				    (pvo->pvo_vaddr & PVO_WIRED) == 0) {
2015 					pvo->pvo_vaddr |= PVO_WIRED;
2016 					pm->pm_stats.wired_count++;
2017 				} else if ((flags & PVO_WIRED) == 0 &&
2018 				    (pvo->pvo_vaddr & PVO_WIRED) != 0) {
2019 					pvo->pvo_vaddr &= ~PVO_WIRED;
2020 					pm->pm_stats.wired_count--;
2021 				}
2022 				return (0);
2023 			}
2024 			moea_pvo_remove(pvo, -1);
2025 			break;
2026 		}
2027 	}
2028 
2029 	/*
2030 	 * If we aren't overwriting a mapping, try to allocate.
2031 	 */
2032 	if (moea_initialized) {
2033 		pvo = uma_zalloc(zone, M_NOWAIT);
2034 	} else {
2035 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
2036 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
2037 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
2038 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
2039 		}
2040 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
2041 		moea_bpvo_pool_index++;
2042 		bootstrap = 1;
2043 	}
2044 
2045 	if (pvo == NULL) {
2046 		mtx_unlock(&moea_table_mutex);
2047 		return (ENOMEM);
2048 	}
2049 
2050 	moea_pvo_entries++;
2051 	pvo->pvo_vaddr = va;
2052 	pvo->pvo_pmap = pm;
2053 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
2054 	pvo->pvo_vaddr &= ~ADDR_POFF;
2055 	if (flags & PVO_WIRED)
2056 		pvo->pvo_vaddr |= PVO_WIRED;
2057 	if (pvo_head != &moea_pvo_kunmanaged)
2058 		pvo->pvo_vaddr |= PVO_MANAGED;
2059 	if (bootstrap)
2060 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
2061 
2062 	moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo);
2063 
2064 	/*
2065 	 * Add to pmap list
2066 	 */
2067 	RB_INSERT(pvo_tree, &pm->pmap_pvo, pvo);
2068 
2069 	/*
2070 	 * Remember if the list was empty and therefore will be the first
2071 	 * item.
2072 	 */
2073 	if (LIST_FIRST(pvo_head) == NULL)
2074 		first = 1;
2075 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2076 
2077 	if (pvo->pvo_vaddr & PVO_WIRED)
2078 		pm->pm_stats.wired_count++;
2079 	pm->pm_stats.resident_count++;
2080 
2081 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2082 	KASSERT(i < 8, ("Invalid PTE index"));
2083 	if (i >= 0) {
2084 		PVO_PTEGIDX_SET(pvo, i);
2085 	} else {
2086 		panic("moea_pvo_enter: overflow");
2087 		moea_pte_overflow++;
2088 	}
2089 	mtx_unlock(&moea_table_mutex);
2090 
2091 	return (first ? ENOENT : 0);
2092 }
2093 
2094 static void
2095 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
2096 {
2097 	struct	pte *pt;
2098 
2099 	/*
2100 	 * If there is an active pte entry, we need to deactivate it (and
2101 	 * save the ref & cfg bits).
2102 	 */
2103 	pt = moea_pvo_to_pte(pvo, pteidx);
2104 	if (pt != NULL) {
2105 		moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
2106 		mtx_unlock(&moea_table_mutex);
2107 		PVO_PTEGIDX_CLR(pvo);
2108 	} else {
2109 		moea_pte_overflow--;
2110 	}
2111 
2112 	/*
2113 	 * Update our statistics.
2114 	 */
2115 	pvo->pvo_pmap->pm_stats.resident_count--;
2116 	if (pvo->pvo_vaddr & PVO_WIRED)
2117 		pvo->pvo_pmap->pm_stats.wired_count--;
2118 
2119 	/*
2120 	 * Remove this PVO from the PV and pmap lists.
2121 	 */
2122 	LIST_REMOVE(pvo, pvo_vlink);
2123 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2124 
2125 	/*
2126 	 * Save the REF/CHG bits into their cache if the page is managed.
2127 	 * Clear PGA_WRITEABLE if all mappings of the page have been removed.
2128 	 */
2129 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED) {
2130 		struct vm_page *pg;
2131 
2132 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
2133 		if (pg != NULL) {
2134 			moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2135 			    (PTE_REF | PTE_CHG));
2136 			if (LIST_EMPTY(&pg->md.mdpg_pvoh))
2137 				vm_page_aflag_clear(pg, PGA_WRITEABLE);
2138 		}
2139 	}
2140 
2141 	/*
2142 	 * Remove this from the overflow list and return it to the pool
2143 	 * if we aren't going to reuse it.
2144 	 */
2145 	LIST_REMOVE(pvo, pvo_olink);
2146 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2147 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
2148 		    moea_upvo_zone, pvo);
2149 	moea_pvo_entries--;
2150 	moea_pvo_remove_calls++;
2151 }
2152 
2153 static __inline int
2154 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2155 {
2156 	int	pteidx;
2157 
2158 	/*
2159 	 * We can find the actual pte entry without searching by grabbing
2160 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2161 	 * noticing the HID bit.
2162 	 */
2163 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2164 	if (pvo->pvo_pte.pte.pte_hi & PTE_HID)
2165 		pteidx ^= moea_pteg_mask * 8;
2166 
2167 	return (pteidx);
2168 }
2169 
2170 static struct pvo_entry *
2171 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2172 {
2173 	struct	pvo_entry *pvo;
2174 	int	ptegidx;
2175 	u_int	sr;
2176 
2177 	va &= ~ADDR_POFF;
2178 	sr = va_to_sr(pm->pm_sr, va);
2179 	ptegidx = va_to_pteg(sr, va);
2180 
2181 	mtx_lock(&moea_table_mutex);
2182 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2183 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2184 			if (pteidx_p)
2185 				*pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2186 			break;
2187 		}
2188 	}
2189 	mtx_unlock(&moea_table_mutex);
2190 
2191 	return (pvo);
2192 }
2193 
2194 static struct pte *
2195 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2196 {
2197 	struct	pte *pt;
2198 
2199 	/*
2200 	 * If we haven't been supplied the ptegidx, calculate it.
2201 	 */
2202 	if (pteidx == -1) {
2203 		int	ptegidx;
2204 		u_int	sr;
2205 
2206 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2207 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2208 		pteidx = moea_pvo_pte_index(pvo, ptegidx);
2209 	}
2210 
2211 	pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2212 	mtx_lock(&moea_table_mutex);
2213 
2214 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2215 		panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2216 		    "valid pte index", pvo);
2217 	}
2218 
2219 	if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2220 		panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2221 		    "pvo but no valid pte", pvo);
2222 	}
2223 
2224 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2225 		if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) {
2226 			panic("moea_pvo_to_pte: pvo %p has valid pte in "
2227 			    "moea_pteg_table %p but invalid in pvo", pvo, pt);
2228 		}
2229 
2230 		if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2231 		    != 0) {
2232 			panic("moea_pvo_to_pte: pvo %p pte does not match "
2233 			    "pte %p in moea_pteg_table", pvo, pt);
2234 		}
2235 
2236 		mtx_assert(&moea_table_mutex, MA_OWNED);
2237 		return (pt);
2238 	}
2239 
2240 	if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) {
2241 		panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2242 		    "moea_pteg_table but valid in pvo: %8x, %8x", pvo, pt, pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2243 	}
2244 
2245 	mtx_unlock(&moea_table_mutex);
2246 	return (NULL);
2247 }
2248 
2249 /*
2250  * XXX: THIS STUFF SHOULD BE IN pte.c?
2251  */
2252 int
2253 moea_pte_spill(vm_offset_t addr)
2254 {
2255 	struct	pvo_entry *source_pvo, *victim_pvo;
2256 	struct	pvo_entry *pvo;
2257 	int	ptegidx, i, j;
2258 	u_int	sr;
2259 	struct	pteg *pteg;
2260 	struct	pte *pt;
2261 
2262 	moea_pte_spills++;
2263 
2264 	sr = mfsrin(addr);
2265 	ptegidx = va_to_pteg(sr, addr);
2266 
2267 	/*
2268 	 * Have to substitute some entry.  Use the primary hash for this.
2269 	 * Use low bits of timebase as random generator.
2270 	 */
2271 	pteg = &moea_pteg_table[ptegidx];
2272 	mtx_lock(&moea_table_mutex);
2273 	__asm __volatile("mftb %0" : "=r"(i));
2274 	i &= 7;
2275 	pt = &pteg->pt[i];
2276 
2277 	source_pvo = NULL;
2278 	victim_pvo = NULL;
2279 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2280 		/*
2281 		 * We need to find a pvo entry for this address.
2282 		 */
2283 		if (source_pvo == NULL &&
2284 		    moea_pte_match(&pvo->pvo_pte.pte, sr, addr,
2285 		    pvo->pvo_pte.pte.pte_hi & PTE_HID)) {
2286 			/*
2287 			 * Now found an entry to be spilled into the pteg.
2288 			 * The PTE is now valid, so we know it's active.
2289 			 */
2290 			j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2291 
2292 			if (j >= 0) {
2293 				PVO_PTEGIDX_SET(pvo, j);
2294 				moea_pte_overflow--;
2295 				mtx_unlock(&moea_table_mutex);
2296 				return (1);
2297 			}
2298 
2299 			source_pvo = pvo;
2300 
2301 			if (victim_pvo != NULL)
2302 				break;
2303 		}
2304 
2305 		/*
2306 		 * We also need the pvo entry of the victim we are replacing
2307 		 * so save the R & C bits of the PTE.
2308 		 */
2309 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2310 		    moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2311 			victim_pvo = pvo;
2312 			if (source_pvo != NULL)
2313 				break;
2314 		}
2315 	}
2316 
2317 	if (source_pvo == NULL) {
2318 		mtx_unlock(&moea_table_mutex);
2319 		return (0);
2320 	}
2321 
2322 	if (victim_pvo == NULL) {
2323 		if ((pt->pte_hi & PTE_HID) == 0)
2324 			panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2325 			    "entry", pt);
2326 
2327 		/*
2328 		 * If this is a secondary PTE, we need to search it's primary
2329 		 * pvo bucket for the matching PVO.
2330 		 */
2331 		LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2332 		    pvo_olink) {
2333 			/*
2334 			 * We also need the pvo entry of the victim we are
2335 			 * replacing so save the R & C bits of the PTE.
2336 			 */
2337 			if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2338 				victim_pvo = pvo;
2339 				break;
2340 			}
2341 		}
2342 
2343 		if (victim_pvo == NULL)
2344 			panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2345 			    "entry", pt);
2346 	}
2347 
2348 	/*
2349 	 * We are invalidating the TLB entry for the EA we are replacing even
2350 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2351 	 * contained in the TLB entry.
2352 	 */
2353 	source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID;
2354 
2355 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2356 	moea_pte_set(pt, &source_pvo->pvo_pte.pte);
2357 
2358 	PVO_PTEGIDX_CLR(victim_pvo);
2359 	PVO_PTEGIDX_SET(source_pvo, i);
2360 	moea_pte_replacements++;
2361 
2362 	mtx_unlock(&moea_table_mutex);
2363 	return (1);
2364 }
2365 
2366 static __inline struct pvo_entry *
2367 moea_pte_spillable_ident(u_int ptegidx)
2368 {
2369 	struct	pte *pt;
2370 	struct	pvo_entry *pvo_walk, *pvo = NULL;
2371 
2372 	LIST_FOREACH(pvo_walk, &moea_pvo_table[ptegidx], pvo_olink) {
2373 		if (pvo_walk->pvo_vaddr & PVO_WIRED)
2374 			continue;
2375 
2376 		if (!(pvo_walk->pvo_pte.pte.pte_hi & PTE_VALID))
2377 			continue;
2378 
2379 		pt = moea_pvo_to_pte(pvo_walk, -1);
2380 
2381 		if (pt == NULL)
2382 			continue;
2383 
2384 		pvo = pvo_walk;
2385 
2386 		mtx_unlock(&moea_table_mutex);
2387 		if (!(pt->pte_lo & PTE_REF))
2388 			return (pvo_walk);
2389 	}
2390 
2391 	return (pvo);
2392 }
2393 
2394 static int
2395 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2396 {
2397 	struct	pte *pt;
2398 	struct	pvo_entry *victim_pvo;
2399 	int	i;
2400 	int	victim_idx;
2401 	u_int	pteg_bkpidx = ptegidx;
2402 
2403 	mtx_assert(&moea_table_mutex, MA_OWNED);
2404 
2405 	/*
2406 	 * First try primary hash.
2407 	 */
2408 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2409 		if ((pt->pte_hi & PTE_VALID) == 0) {
2410 			pvo_pt->pte_hi &= ~PTE_HID;
2411 			moea_pte_set(pt, pvo_pt);
2412 			return (i);
2413 		}
2414 	}
2415 
2416 	/*
2417 	 * Now try secondary hash.
2418 	 */
2419 	ptegidx ^= moea_pteg_mask;
2420 
2421 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2422 		if ((pt->pte_hi & PTE_VALID) == 0) {
2423 			pvo_pt->pte_hi |= PTE_HID;
2424 			moea_pte_set(pt, pvo_pt);
2425 			return (i);
2426 		}
2427 	}
2428 
2429 	/* Try again, but this time try to force a PTE out. */
2430 	ptegidx = pteg_bkpidx;
2431 
2432 	victim_pvo = moea_pte_spillable_ident(ptegidx);
2433 	if (victim_pvo == NULL) {
2434 		ptegidx ^= moea_pteg_mask;
2435 		victim_pvo = moea_pte_spillable_ident(ptegidx);
2436 	}
2437 
2438 	if (victim_pvo == NULL) {
2439 		panic("moea_pte_insert: overflow");
2440 		return (-1);
2441 	}
2442 
2443 	victim_idx = moea_pvo_pte_index(victim_pvo, ptegidx);
2444 
2445 	if (pteg_bkpidx == ptegidx)
2446 		pvo_pt->pte_hi &= ~PTE_HID;
2447 	else
2448 		pvo_pt->pte_hi |= PTE_HID;
2449 
2450 	/*
2451 	 * Synchronize the sacrifice PTE with its PVO, then mark both
2452 	 * invalid. The PVO will be reused when/if the VM system comes
2453 	 * here after a fault.
2454 	 */
2455 	pt = &moea_pteg_table[victim_idx >> 3].pt[victim_idx & 7];
2456 
2457 	if (pt->pte_hi != victim_pvo->pvo_pte.pte.pte_hi)
2458 	    panic("Victim PVO doesn't match PTE! PVO: %8x, PTE: %8x", victim_pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2459 
2460 	/*
2461 	 * Set the new PTE.
2462 	 */
2463 	moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2464 	PVO_PTEGIDX_CLR(victim_pvo);
2465 	moea_pte_overflow++;
2466 	moea_pte_set(pt, pvo_pt);
2467 
2468 	return (victim_idx & 7);
2469 }
2470 
2471 static boolean_t
2472 moea_query_bit(vm_page_t m, int ptebit)
2473 {
2474 	struct	pvo_entry *pvo;
2475 	struct	pte *pt;
2476 
2477 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2478 	if (moea_attr_fetch(m) & ptebit)
2479 		return (TRUE);
2480 
2481 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2482 		/*
2483 		 * See if we saved the bit off.  If so, cache it and return
2484 		 * success.
2485 		 */
2486 		if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2487 			moea_attr_save(m, ptebit);
2488 			return (TRUE);
2489 		}
2490 	}
2491 
2492 	/*
2493 	 * No luck, now go through the hard part of looking at the PTEs
2494 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2495 	 * the PTEs.
2496 	 */
2497 	powerpc_sync();
2498 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2499 		/*
2500 		 * See if this pvo has a valid PTE.  if so, fetch the
2501 		 * REF/CHG bits from the valid PTE.  If the appropriate
2502 		 * ptebit is set, cache it and return success.
2503 		 */
2504 		pt = moea_pvo_to_pte(pvo, -1);
2505 		if (pt != NULL) {
2506 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2507 			mtx_unlock(&moea_table_mutex);
2508 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2509 				moea_attr_save(m, ptebit);
2510 				return (TRUE);
2511 			}
2512 		}
2513 	}
2514 
2515 	return (FALSE);
2516 }
2517 
2518 static u_int
2519 moea_clear_bit(vm_page_t m, int ptebit)
2520 {
2521 	u_int	count;
2522 	struct	pvo_entry *pvo;
2523 	struct	pte *pt;
2524 
2525 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2526 
2527 	/*
2528 	 * Clear the cached value.
2529 	 */
2530 	moea_attr_clear(m, ptebit);
2531 
2532 	/*
2533 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2534 	 * we can reset the right ones).  note that since the pvo entries and
2535 	 * list heads are accessed via BAT0 and are never placed in the page
2536 	 * table, we don't have to worry about further accesses setting the
2537 	 * REF/CHG bits.
2538 	 */
2539 	powerpc_sync();
2540 
2541 	/*
2542 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2543 	 * valid pte clear the ptebit from the valid pte.
2544 	 */
2545 	count = 0;
2546 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2547 		pt = moea_pvo_to_pte(pvo, -1);
2548 		if (pt != NULL) {
2549 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2550 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2551 				count++;
2552 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2553 			}
2554 			mtx_unlock(&moea_table_mutex);
2555 		}
2556 		pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2557 	}
2558 
2559 	return (count);
2560 }
2561 
2562 /*
2563  * Return true if the physical range is encompassed by the battable[idx]
2564  */
2565 static int
2566 moea_bat_mapped(int idx, vm_paddr_t pa, vm_size_t size)
2567 {
2568 	u_int prot;
2569 	u_int32_t start;
2570 	u_int32_t end;
2571 	u_int32_t bat_ble;
2572 
2573 	/*
2574 	 * Return immediately if not a valid mapping
2575 	 */
2576 	if (!(battable[idx].batu & BAT_Vs))
2577 		return (EINVAL);
2578 
2579 	/*
2580 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2581 	 * so it can function as an i/o page
2582 	 */
2583 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2584 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2585 		return (EPERM);
2586 
2587 	/*
2588 	 * The address should be within the BAT range. Assume that the
2589 	 * start address in the BAT has the correct alignment (thus
2590 	 * not requiring masking)
2591 	 */
2592 	start = battable[idx].batl & BAT_PBS;
2593 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2594 	end = start | (bat_ble << 15) | 0x7fff;
2595 
2596 	if ((pa < start) || ((pa + size) > end))
2597 		return (ERANGE);
2598 
2599 	return (0);
2600 }
2601 
2602 boolean_t
2603 moea_dev_direct_mapped(vm_paddr_t pa, vm_size_t size)
2604 {
2605 	int i;
2606 
2607 	/*
2608 	 * This currently does not work for entries that
2609 	 * overlap 256M BAT segments.
2610 	 */
2611 
2612 	for(i = 0; i < 16; i++)
2613 		if (moea_bat_mapped(i, pa, size) == 0)
2614 			return (0);
2615 
2616 	return (EFAULT);
2617 }
2618 
2619 /*
2620  * Map a set of physical memory pages into the kernel virtual
2621  * address space. Return a pointer to where it is mapped. This
2622  * routine is intended to be used for mapping device memory,
2623  * NOT real memory.
2624  */
2625 void *
2626 moea_mapdev(vm_paddr_t pa, vm_size_t size)
2627 {
2628 
2629 	return (moea_mapdev_attr(pa, size, VM_MEMATTR_DEFAULT));
2630 }
2631 
2632 void *
2633 moea_mapdev_attr(vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
2634 {
2635 	vm_offset_t va, tmpva, ppa, offset;
2636 	int i;
2637 
2638 	ppa = trunc_page(pa);
2639 	offset = pa & PAGE_MASK;
2640 	size = roundup(offset + size, PAGE_SIZE);
2641 
2642 	/*
2643 	 * If the physical address lies within a valid BAT table entry,
2644 	 * return the 1:1 mapping. This currently doesn't work
2645 	 * for regions that overlap 256M BAT segments.
2646 	 */
2647 	for (i = 0; i < 16; i++) {
2648 		if (moea_bat_mapped(i, pa, size) == 0)
2649 			return ((void *) pa);
2650 	}
2651 
2652 	va = kva_alloc(size);
2653 	if (!va)
2654 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2655 
2656 	for (tmpva = va; size > 0;) {
2657 		moea_kenter_attr(tmpva, ppa, ma);
2658 		tlbie(tmpva);
2659 		size -= PAGE_SIZE;
2660 		tmpva += PAGE_SIZE;
2661 		ppa += PAGE_SIZE;
2662 	}
2663 
2664 	return ((void *)(va + offset));
2665 }
2666 
2667 void
2668 moea_unmapdev(vm_offset_t va, vm_size_t size)
2669 {
2670 	vm_offset_t base, offset;
2671 
2672 	/*
2673 	 * If this is outside kernel virtual space, then it's a
2674 	 * battable entry and doesn't require unmapping
2675 	 */
2676 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2677 		base = trunc_page(va);
2678 		offset = va & PAGE_MASK;
2679 		size = roundup(offset + size, PAGE_SIZE);
2680 		moea_qremove(base, atop(size));
2681 		kva_free(base, size);
2682 	}
2683 }
2684 
2685 static void
2686 moea_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz)
2687 {
2688 	struct pvo_entry *pvo;
2689 	vm_offset_t lim;
2690 	vm_paddr_t pa;
2691 	vm_size_t len;
2692 
2693 	PMAP_LOCK(pm);
2694 	while (sz > 0) {
2695 		lim = round_page(va + 1);
2696 		len = MIN(lim - va, sz);
2697 		pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2698 		if (pvo != NULL) {
2699 			pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) |
2700 			    (va & ADDR_POFF);
2701 			moea_syncicache(pa, len);
2702 		}
2703 		va += len;
2704 		sz -= len;
2705 	}
2706 	PMAP_UNLOCK(pm);
2707 }
2708 
2709 void
2710 moea_dumpsys_map(vm_paddr_t pa, size_t sz, void **va)
2711 {
2712 
2713 	*va = (void *)pa;
2714 }
2715 
2716 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
2717 
2718 void
2719 moea_scan_init()
2720 {
2721 	struct pvo_entry *pvo;
2722 	vm_offset_t va;
2723 	int i;
2724 
2725 	if (!do_minidump) {
2726 		/* Initialize phys. segments for dumpsys(). */
2727 		memset(&dump_map, 0, sizeof(dump_map));
2728 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
2729 		for (i = 0; i < pregions_sz; i++) {
2730 			dump_map[i].pa_start = pregions[i].mr_start;
2731 			dump_map[i].pa_size = pregions[i].mr_size;
2732 		}
2733 		return;
2734 	}
2735 
2736 	/* Virtual segments for minidumps: */
2737 	memset(&dump_map, 0, sizeof(dump_map));
2738 
2739 	/* 1st: kernel .data and .bss. */
2740 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
2741 	dump_map[0].pa_size =
2742 	    round_page((uintptr_t)_end) - dump_map[0].pa_start;
2743 
2744 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2745 	dump_map[1].pa_start = (vm_paddr_t)msgbufp->msg_ptr;
2746 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
2747 
2748 	/* 3rd: kernel VM. */
2749 	va = dump_map[1].pa_start + dump_map[1].pa_size;
2750 	/* Find start of next chunk (from va). */
2751 	while (va < virtual_end) {
2752 		/* Don't dump the buffer cache. */
2753 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
2754 			va = kmi.buffer_eva;
2755 			continue;
2756 		}
2757 		pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
2758 		if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2759 			break;
2760 		va += PAGE_SIZE;
2761 	}
2762 	if (va < virtual_end) {
2763 		dump_map[2].pa_start = va;
2764 		va += PAGE_SIZE;
2765 		/* Find last page in chunk. */
2766 		while (va < virtual_end) {
2767 			/* Don't run into the buffer cache. */
2768 			if (va == kmi.buffer_sva)
2769 				break;
2770 			pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF,
2771 			    NULL);
2772 			if (pvo == NULL ||
2773 			    !(pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2774 				break;
2775 			va += PAGE_SIZE;
2776 		}
2777 		dump_map[2].pa_size = va - dump_map[2].pa_start;
2778 	}
2779 }
2780