xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision 389e4940069316fe667ffa263fa7d6390d0a960f)
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_pageout.h>
136 #include <vm/uma.h>
137 
138 #include <machine/cpu.h>
139 #include <machine/platform.h>
140 #include <machine/bat.h>
141 #include <machine/frame.h>
142 #include <machine/md_var.h>
143 #include <machine/psl.h>
144 #include <machine/pte.h>
145 #include <machine/smp.h>
146 #include <machine/sr.h>
147 #include <machine/mmuvar.h>
148 #include <machine/trap.h>
149 
150 #include "mmu_if.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(mmu_t, vm_offset_t);
269 int		moea_pte_spill(vm_offset_t);
270 
271 /*
272  * Kernel MMU interface
273  */
274 void moea_clear_modify(mmu_t, vm_page_t);
275 void moea_copy_page(mmu_t, vm_page_t, vm_page_t);
276 void moea_copy_pages(mmu_t mmu, 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(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int,
279     int8_t);
280 void moea_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
281     vm_prot_t);
282 void moea_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
283 vm_paddr_t moea_extract(mmu_t, pmap_t, vm_offset_t);
284 vm_page_t moea_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
285 void moea_init(mmu_t);
286 boolean_t moea_is_modified(mmu_t, vm_page_t);
287 boolean_t moea_is_prefaultable(mmu_t, pmap_t, vm_offset_t);
288 boolean_t moea_is_referenced(mmu_t, vm_page_t);
289 int moea_ts_referenced(mmu_t, vm_page_t);
290 vm_offset_t moea_map(mmu_t, vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
291 boolean_t moea_page_exists_quick(mmu_t, pmap_t, vm_page_t);
292 void moea_page_init(mmu_t, vm_page_t);
293 int moea_page_wired_mappings(mmu_t, vm_page_t);
294 void moea_pinit(mmu_t, pmap_t);
295 void moea_pinit0(mmu_t, pmap_t);
296 void moea_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
297 void moea_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
298 void moea_qremove(mmu_t, vm_offset_t, int);
299 void moea_release(mmu_t, pmap_t);
300 void moea_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
301 void moea_remove_all(mmu_t, vm_page_t);
302 void moea_remove_write(mmu_t, vm_page_t);
303 void moea_unwire(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
304 void moea_zero_page(mmu_t, vm_page_t);
305 void moea_zero_page_area(mmu_t, vm_page_t, int, int);
306 void moea_activate(mmu_t, struct thread *);
307 void moea_deactivate(mmu_t, struct thread *);
308 void moea_cpu_bootstrap(mmu_t, int);
309 void moea_bootstrap(mmu_t, vm_offset_t, vm_offset_t);
310 void *moea_mapdev(mmu_t, vm_paddr_t, vm_size_t);
311 void *moea_mapdev_attr(mmu_t, vm_paddr_t, vm_size_t, vm_memattr_t);
312 void moea_unmapdev(mmu_t, vm_offset_t, vm_size_t);
313 vm_paddr_t moea_kextract(mmu_t, vm_offset_t);
314 void moea_kenter_attr(mmu_t, vm_offset_t, vm_paddr_t, vm_memattr_t);
315 void moea_kenter(mmu_t, vm_offset_t, vm_paddr_t);
316 void moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma);
317 boolean_t moea_dev_direct_mapped(mmu_t, vm_paddr_t, vm_size_t);
318 static void moea_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t);
319 void moea_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va);
320 void moea_scan_init(mmu_t mmu);
321 vm_offset_t moea_quick_enter_page(mmu_t mmu, vm_page_t m);
322 void moea_quick_remove_page(mmu_t mmu, vm_offset_t addr);
323 static int moea_map_user_ptr(mmu_t mmu, pmap_t pm,
324     volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen);
325 static int moea_decode_kernel_ptr(mmu_t mmu, vm_offset_t addr,
326     int *is_user, vm_offset_t *decoded_addr);
327 
328 
329 static mmu_method_t moea_methods[] = {
330 	MMUMETHOD(mmu_clear_modify,	moea_clear_modify),
331 	MMUMETHOD(mmu_copy_page,	moea_copy_page),
332 	MMUMETHOD(mmu_copy_pages,	moea_copy_pages),
333 	MMUMETHOD(mmu_enter,		moea_enter),
334 	MMUMETHOD(mmu_enter_object,	moea_enter_object),
335 	MMUMETHOD(mmu_enter_quick,	moea_enter_quick),
336 	MMUMETHOD(mmu_extract,		moea_extract),
337 	MMUMETHOD(mmu_extract_and_hold,	moea_extract_and_hold),
338 	MMUMETHOD(mmu_init,		moea_init),
339 	MMUMETHOD(mmu_is_modified,	moea_is_modified),
340 	MMUMETHOD(mmu_is_prefaultable,	moea_is_prefaultable),
341 	MMUMETHOD(mmu_is_referenced,	moea_is_referenced),
342 	MMUMETHOD(mmu_ts_referenced,	moea_ts_referenced),
343 	MMUMETHOD(mmu_map,     		moea_map),
344 	MMUMETHOD(mmu_page_exists_quick,moea_page_exists_quick),
345 	MMUMETHOD(mmu_page_init,	moea_page_init),
346 	MMUMETHOD(mmu_page_wired_mappings,moea_page_wired_mappings),
347 	MMUMETHOD(mmu_pinit,		moea_pinit),
348 	MMUMETHOD(mmu_pinit0,		moea_pinit0),
349 	MMUMETHOD(mmu_protect,		moea_protect),
350 	MMUMETHOD(mmu_qenter,		moea_qenter),
351 	MMUMETHOD(mmu_qremove,		moea_qremove),
352 	MMUMETHOD(mmu_release,		moea_release),
353 	MMUMETHOD(mmu_remove,		moea_remove),
354 	MMUMETHOD(mmu_remove_all,      	moea_remove_all),
355 	MMUMETHOD(mmu_remove_write,	moea_remove_write),
356 	MMUMETHOD(mmu_sync_icache,	moea_sync_icache),
357 	MMUMETHOD(mmu_unwire,		moea_unwire),
358 	MMUMETHOD(mmu_zero_page,       	moea_zero_page),
359 	MMUMETHOD(mmu_zero_page_area,	moea_zero_page_area),
360 	MMUMETHOD(mmu_activate,		moea_activate),
361 	MMUMETHOD(mmu_deactivate,      	moea_deactivate),
362 	MMUMETHOD(mmu_page_set_memattr,	moea_page_set_memattr),
363 	MMUMETHOD(mmu_quick_enter_page, moea_quick_enter_page),
364 	MMUMETHOD(mmu_quick_remove_page, moea_quick_remove_page),
365 
366 	/* Internal interfaces */
367 	MMUMETHOD(mmu_bootstrap,       	moea_bootstrap),
368 	MMUMETHOD(mmu_cpu_bootstrap,   	moea_cpu_bootstrap),
369 	MMUMETHOD(mmu_mapdev_attr,	moea_mapdev_attr),
370 	MMUMETHOD(mmu_mapdev,		moea_mapdev),
371 	MMUMETHOD(mmu_unmapdev,		moea_unmapdev),
372 	MMUMETHOD(mmu_kextract,		moea_kextract),
373 	MMUMETHOD(mmu_kenter,		moea_kenter),
374 	MMUMETHOD(mmu_kenter_attr,	moea_kenter_attr),
375 	MMUMETHOD(mmu_dev_direct_mapped,moea_dev_direct_mapped),
376 	MMUMETHOD(mmu_scan_init,	moea_scan_init),
377 	MMUMETHOD(mmu_dumpsys_map,	moea_dumpsys_map),
378 	MMUMETHOD(mmu_map_user_ptr,	moea_map_user_ptr),
379 	MMUMETHOD(mmu_decode_kernel_ptr, moea_decode_kernel_ptr),
380 
381 	{ 0, 0 }
382 };
383 
384 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods, 0);
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(mmu_t mmup, 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(mmu_t mmup, 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 (sizeof(phys_avail)/sizeof(phys_avail[0]) < 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(mmup, 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(mmup,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(mmup, 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(mmup, 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(mmup, 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(mmu_t mmu, 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(mmu_t mmu, 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(mmu_t mmu, 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(mmu_t mmu, 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(mmu_t mmu, 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(mmu_t mmu, 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(mmu_t mmu, 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(mmu_t mmu, vm_page_t m)
1095 {
1096 
1097 	return (VM_PAGE_TO_PHYS(m));
1098 }
1099 
1100 void
1101 moea_quick_remove_page(mmu_t mmu, vm_offset_t addr)
1102 {
1103 }
1104 
1105 /*
1106  * Map the given physical page at the specified virtual address in the
1107  * target pmap with the protection requested.  If specified the page
1108  * will be wired down.
1109  */
1110 int
1111 moea_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1112     u_int flags, int8_t psind)
1113 {
1114 	int error;
1115 
1116 	for (;;) {
1117 		rw_wlock(&pvh_global_lock);
1118 		PMAP_LOCK(pmap);
1119 		error = moea_enter_locked(pmap, va, m, prot, flags, psind);
1120 		rw_wunlock(&pvh_global_lock);
1121 		PMAP_UNLOCK(pmap);
1122 		if (error != ENOMEM)
1123 			return (KERN_SUCCESS);
1124 		if ((flags & PMAP_ENTER_NOSLEEP) != 0)
1125 			return (KERN_RESOURCE_SHORTAGE);
1126 		VM_OBJECT_ASSERT_UNLOCKED(m->object);
1127 		vm_wait(NULL);
1128 	}
1129 }
1130 
1131 /*
1132  * Map the given physical page at the specified virtual address in the
1133  * target pmap with the protection requested.  If specified the page
1134  * will be wired down.
1135  *
1136  * The global pvh and pmap must be locked.
1137  */
1138 static int
1139 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1140     u_int flags, int8_t psind __unused)
1141 {
1142 	struct		pvo_head *pvo_head;
1143 	uma_zone_t	zone;
1144 	u_int		pte_lo, pvo_flags;
1145 	int		error;
1146 
1147 	if (pmap_bootstrapped)
1148 		rw_assert(&pvh_global_lock, RA_WLOCKED);
1149 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1150 	if ((m->oflags & VPO_UNMANAGED) == 0 && !vm_page_xbusied(m))
1151 		VM_OBJECT_ASSERT_LOCKED(m->object);
1152 
1153 	if ((m->oflags & VPO_UNMANAGED) != 0 || !moea_initialized) {
1154 		pvo_head = &moea_pvo_kunmanaged;
1155 		zone = moea_upvo_zone;
1156 		pvo_flags = 0;
1157 	} else {
1158 		pvo_head = vm_page_to_pvoh(m);
1159 		zone = moea_mpvo_zone;
1160 		pvo_flags = PVO_MANAGED;
1161 	}
1162 
1163 	pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1164 
1165 	if (prot & VM_PROT_WRITE) {
1166 		pte_lo |= PTE_BW;
1167 		if (pmap_bootstrapped &&
1168 		    (m->oflags & VPO_UNMANAGED) == 0)
1169 			vm_page_aflag_set(m, PGA_WRITEABLE);
1170 	} else
1171 		pte_lo |= PTE_BR;
1172 
1173 	if ((flags & PMAP_ENTER_WIRED) != 0)
1174 		pvo_flags |= PVO_WIRED;
1175 
1176 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1177 	    pte_lo, pvo_flags);
1178 
1179 	/*
1180 	 * Flush the real page from the instruction cache. This has be done
1181 	 * for all user mappings to prevent information leakage via the
1182 	 * instruction cache. moea_pvo_enter() returns ENOENT for the first
1183 	 * mapping for a page.
1184 	 */
1185 	if (pmap != kernel_pmap && error == ENOENT &&
1186 	    (pte_lo & (PTE_I | PTE_G)) == 0)
1187 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1188 
1189 	return (error);
1190 }
1191 
1192 /*
1193  * Maps a sequence of resident pages belonging to the same object.
1194  * The sequence begins with the given page m_start.  This page is
1195  * mapped at the given virtual address start.  Each subsequent page is
1196  * mapped at a virtual address that is offset from start by the same
1197  * amount as the page is offset from m_start within the object.  The
1198  * last page in the sequence is the page with the largest offset from
1199  * m_start that can be mapped at a virtual address less than the given
1200  * virtual address end.  Not every virtual page between start and end
1201  * is mapped; only those for which a resident page exists with the
1202  * corresponding offset from m_start are mapped.
1203  */
1204 void
1205 moea_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end,
1206     vm_page_t m_start, vm_prot_t prot)
1207 {
1208 	vm_page_t m;
1209 	vm_pindex_t diff, psize;
1210 
1211 	VM_OBJECT_ASSERT_LOCKED(m_start->object);
1212 
1213 	psize = atop(end - start);
1214 	m = m_start;
1215 	rw_wlock(&pvh_global_lock);
1216 	PMAP_LOCK(pm);
1217 	while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) {
1218 		moea_enter_locked(pm, start + ptoa(diff), m, prot &
1219 		    (VM_PROT_READ | VM_PROT_EXECUTE), 0, 0);
1220 		m = TAILQ_NEXT(m, listq);
1221 	}
1222 	rw_wunlock(&pvh_global_lock);
1223 	PMAP_UNLOCK(pm);
1224 }
1225 
1226 void
1227 moea_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1228     vm_prot_t prot)
1229 {
1230 
1231 	rw_wlock(&pvh_global_lock);
1232 	PMAP_LOCK(pm);
1233 	moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1234 	    0, 0);
1235 	rw_wunlock(&pvh_global_lock);
1236 	PMAP_UNLOCK(pm);
1237 }
1238 
1239 vm_paddr_t
1240 moea_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1241 {
1242 	struct	pvo_entry *pvo;
1243 	vm_paddr_t pa;
1244 
1245 	PMAP_LOCK(pm);
1246 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1247 	if (pvo == NULL)
1248 		pa = 0;
1249 	else
1250 		pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1251 	PMAP_UNLOCK(pm);
1252 	return (pa);
1253 }
1254 
1255 /*
1256  * Atomically extract and hold the physical page with the given
1257  * pmap and virtual address pair if that mapping permits the given
1258  * protection.
1259  */
1260 vm_page_t
1261 moea_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1262 {
1263 	struct	pvo_entry *pvo;
1264 	vm_page_t m;
1265         vm_paddr_t pa;
1266 
1267 	m = NULL;
1268 	pa = 0;
1269 	PMAP_LOCK(pmap);
1270 retry:
1271 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1272 	if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1273 	    ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1274 	     (prot & VM_PROT_WRITE) == 0)) {
1275 		if (vm_page_pa_tryrelock(pmap, pvo->pvo_pte.pte.pte_lo & PTE_RPGN, &pa))
1276 			goto retry;
1277 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
1278 		vm_page_hold(m);
1279 	}
1280 	PA_UNLOCK_COND(pa);
1281 	PMAP_UNLOCK(pmap);
1282 	return (m);
1283 }
1284 
1285 void
1286 moea_init(mmu_t mmu)
1287 {
1288 
1289 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1290 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1291 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1292 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1293 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1294 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1295 	moea_initialized = TRUE;
1296 }
1297 
1298 boolean_t
1299 moea_is_referenced(mmu_t mmu, vm_page_t m)
1300 {
1301 	boolean_t rv;
1302 
1303 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1304 	    ("moea_is_referenced: page %p is not managed", m));
1305 	rw_wlock(&pvh_global_lock);
1306 	rv = moea_query_bit(m, PTE_REF);
1307 	rw_wunlock(&pvh_global_lock);
1308 	return (rv);
1309 }
1310 
1311 boolean_t
1312 moea_is_modified(mmu_t mmu, vm_page_t m)
1313 {
1314 	boolean_t rv;
1315 
1316 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1317 	    ("moea_is_modified: page %p is not managed", m));
1318 
1319 	/*
1320 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1321 	 * concurrently set while the object is locked.  Thus, if PGA_WRITEABLE
1322 	 * is clear, no PTEs can have PTE_CHG set.
1323 	 */
1324 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1325 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1326 		return (FALSE);
1327 	rw_wlock(&pvh_global_lock);
1328 	rv = moea_query_bit(m, PTE_CHG);
1329 	rw_wunlock(&pvh_global_lock);
1330 	return (rv);
1331 }
1332 
1333 boolean_t
1334 moea_is_prefaultable(mmu_t mmu, pmap_t pmap, vm_offset_t va)
1335 {
1336 	struct pvo_entry *pvo;
1337 	boolean_t rv;
1338 
1339 	PMAP_LOCK(pmap);
1340 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1341 	rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1342 	PMAP_UNLOCK(pmap);
1343 	return (rv);
1344 }
1345 
1346 void
1347 moea_clear_modify(mmu_t mmu, vm_page_t m)
1348 {
1349 
1350 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1351 	    ("moea_clear_modify: page %p is not managed", m));
1352 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1353 	KASSERT(!vm_page_xbusied(m),
1354 	    ("moea_clear_modify: page %p is exclusive busy", m));
1355 
1356 	/*
1357 	 * If the page is not PGA_WRITEABLE, then no PTEs can have PTE_CHG
1358 	 * set.  If the object containing the page is locked and the page is
1359 	 * not exclusive busied, then PGA_WRITEABLE cannot be concurrently set.
1360 	 */
1361 	if ((m->aflags & PGA_WRITEABLE) == 0)
1362 		return;
1363 	rw_wlock(&pvh_global_lock);
1364 	moea_clear_bit(m, PTE_CHG);
1365 	rw_wunlock(&pvh_global_lock);
1366 }
1367 
1368 /*
1369  * Clear the write and modified bits in each of the given page's mappings.
1370  */
1371 void
1372 moea_remove_write(mmu_t mmu, vm_page_t m)
1373 {
1374 	struct	pvo_entry *pvo;
1375 	struct	pte *pt;
1376 	pmap_t	pmap;
1377 	u_int	lo;
1378 
1379 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1380 	    ("moea_remove_write: page %p is not managed", m));
1381 
1382 	/*
1383 	 * If the page is not exclusive busied, then PGA_WRITEABLE cannot be
1384 	 * set by another thread while the object is locked.  Thus,
1385 	 * if PGA_WRITEABLE is clear, no page table entries need updating.
1386 	 */
1387 	VM_OBJECT_ASSERT_WLOCKED(m->object);
1388 	if (!vm_page_xbusied(m) && (m->aflags & PGA_WRITEABLE) == 0)
1389 		return;
1390 	rw_wlock(&pvh_global_lock);
1391 	lo = moea_attr_fetch(m);
1392 	powerpc_sync();
1393 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1394 		pmap = pvo->pvo_pmap;
1395 		PMAP_LOCK(pmap);
1396 		if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1397 			pt = moea_pvo_to_pte(pvo, -1);
1398 			pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1399 			pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1400 			if (pt != NULL) {
1401 				moea_pte_synch(pt, &pvo->pvo_pte.pte);
1402 				lo |= pvo->pvo_pte.pte.pte_lo;
1403 				pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1404 				moea_pte_change(pt, &pvo->pvo_pte.pte,
1405 				    pvo->pvo_vaddr);
1406 				mtx_unlock(&moea_table_mutex);
1407 			}
1408 		}
1409 		PMAP_UNLOCK(pmap);
1410 	}
1411 	if ((lo & PTE_CHG) != 0) {
1412 		moea_attr_clear(m, PTE_CHG);
1413 		vm_page_dirty(m);
1414 	}
1415 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1416 	rw_wunlock(&pvh_global_lock);
1417 }
1418 
1419 /*
1420  *	moea_ts_referenced:
1421  *
1422  *	Return a count of reference bits for a page, clearing those bits.
1423  *	It is not necessary for every reference bit to be cleared, but it
1424  *	is necessary that 0 only be returned when there are truly no
1425  *	reference bits set.
1426  *
1427  *	XXX: The exact number of bits to check and clear is a matter that
1428  *	should be tested and standardized at some point in the future for
1429  *	optimal aging of shared pages.
1430  */
1431 int
1432 moea_ts_referenced(mmu_t mmu, vm_page_t m)
1433 {
1434 	int count;
1435 
1436 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1437 	    ("moea_ts_referenced: page %p is not managed", m));
1438 	rw_wlock(&pvh_global_lock);
1439 	count = moea_clear_bit(m, PTE_REF);
1440 	rw_wunlock(&pvh_global_lock);
1441 	return (count);
1442 }
1443 
1444 /*
1445  * Modify the WIMG settings of all mappings for a page.
1446  */
1447 void
1448 moea_page_set_memattr(mmu_t mmu, vm_page_t m, vm_memattr_t ma)
1449 {
1450 	struct	pvo_entry *pvo;
1451 	struct	pvo_head *pvo_head;
1452 	struct	pte *pt;
1453 	pmap_t	pmap;
1454 	u_int	lo;
1455 
1456 	if ((m->oflags & VPO_UNMANAGED) != 0) {
1457 		m->md.mdpg_cache_attrs = ma;
1458 		return;
1459 	}
1460 
1461 	rw_wlock(&pvh_global_lock);
1462 	pvo_head = vm_page_to_pvoh(m);
1463 	lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1464 
1465 	LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1466 		pmap = pvo->pvo_pmap;
1467 		PMAP_LOCK(pmap);
1468 		pt = moea_pvo_to_pte(pvo, -1);
1469 		pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1470 		pvo->pvo_pte.pte.pte_lo |= lo;
1471 		if (pt != NULL) {
1472 			moea_pte_change(pt, &pvo->pvo_pte.pte,
1473 			    pvo->pvo_vaddr);
1474 			if (pvo->pvo_pmap == kernel_pmap)
1475 				isync();
1476 		}
1477 		mtx_unlock(&moea_table_mutex);
1478 		PMAP_UNLOCK(pmap);
1479 	}
1480 	m->md.mdpg_cache_attrs = ma;
1481 	rw_wunlock(&pvh_global_lock);
1482 }
1483 
1484 /*
1485  * Map a wired page into kernel virtual address space.
1486  */
1487 void
1488 moea_kenter(mmu_t mmu, vm_offset_t va, vm_paddr_t pa)
1489 {
1490 
1491 	moea_kenter_attr(mmu, va, pa, VM_MEMATTR_DEFAULT);
1492 }
1493 
1494 void
1495 moea_kenter_attr(mmu_t mmu, vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
1496 {
1497 	u_int		pte_lo;
1498 	int		error;
1499 
1500 #if 0
1501 	if (va < VM_MIN_KERNEL_ADDRESS)
1502 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1503 		    va);
1504 #endif
1505 
1506 	pte_lo = moea_calc_wimg(pa, ma);
1507 
1508 	PMAP_LOCK(kernel_pmap);
1509 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1510 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1511 
1512 	if (error != 0 && error != ENOENT)
1513 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1514 		    pa, error);
1515 
1516 	PMAP_UNLOCK(kernel_pmap);
1517 }
1518 
1519 /*
1520  * Extract the physical page address associated with the given kernel virtual
1521  * address.
1522  */
1523 vm_paddr_t
1524 moea_kextract(mmu_t mmu, vm_offset_t va)
1525 {
1526 	struct		pvo_entry *pvo;
1527 	vm_paddr_t pa;
1528 
1529 	/*
1530 	 * Allow direct mappings on 32-bit OEA
1531 	 */
1532 	if (va < VM_MIN_KERNEL_ADDRESS) {
1533 		return (va);
1534 	}
1535 
1536 	PMAP_LOCK(kernel_pmap);
1537 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1538 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1539 	pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1540 	PMAP_UNLOCK(kernel_pmap);
1541 	return (pa);
1542 }
1543 
1544 /*
1545  * Remove a wired page from kernel virtual address space.
1546  */
1547 void
1548 moea_kremove(mmu_t mmu, vm_offset_t va)
1549 {
1550 
1551 	moea_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1552 }
1553 
1554 /*
1555  * Provide a kernel pointer corresponding to a given userland pointer.
1556  * The returned pointer is valid until the next time this function is
1557  * called in this thread. This is used internally in copyin/copyout.
1558  */
1559 int
1560 moea_map_user_ptr(mmu_t mmu, pmap_t pm, volatile const void *uaddr,
1561     void **kaddr, size_t ulen, size_t *klen)
1562 {
1563 	size_t l;
1564 	register_t vsid;
1565 
1566 	*kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
1567 	l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
1568 	if (l > ulen)
1569 		l = ulen;
1570 	if (klen)
1571 		*klen = l;
1572 	else if (l != ulen)
1573 		return (EFAULT);
1574 
1575 	vsid = va_to_vsid(pm, (vm_offset_t)uaddr);
1576 
1577 	/* Mark segment no-execute */
1578 	vsid |= SR_N;
1579 
1580 	/* If we have already set this VSID, we can just return */
1581 	if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == vsid)
1582 		return (0);
1583 
1584 	__asm __volatile("isync");
1585 	curthread->td_pcb->pcb_cpu.aim.usr_segm =
1586 	    (uintptr_t)uaddr >> ADDR_SR_SHFT;
1587 	curthread->td_pcb->pcb_cpu.aim.usr_vsid = vsid;
1588 	__asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(vsid));
1589 
1590 	return (0);
1591 }
1592 
1593 /*
1594  * Figure out where a given kernel pointer (usually in a fault) points
1595  * to from the VM's perspective, potentially remapping into userland's
1596  * address space.
1597  */
1598 static int
1599 moea_decode_kernel_ptr(mmu_t mmu, vm_offset_t addr, int *is_user,
1600     vm_offset_t *decoded_addr)
1601 {
1602 	vm_offset_t user_sr;
1603 
1604 	if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
1605 		user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
1606 		addr &= ADDR_PIDX | ADDR_POFF;
1607 		addr |= user_sr << ADDR_SR_SHFT;
1608 		*decoded_addr = addr;
1609 		*is_user = 1;
1610 	} else {
1611 		*decoded_addr = addr;
1612 		*is_user = 0;
1613 	}
1614 
1615 	return (0);
1616 }
1617 
1618 /*
1619  * Map a range of physical addresses into kernel virtual address space.
1620  *
1621  * The value passed in *virt is a suggested virtual address for the mapping.
1622  * Architectures which can support a direct-mapped physical to virtual region
1623  * can return the appropriate address within that region, leaving '*virt'
1624  * unchanged.  We cannot and therefore do not; *virt is updated with the
1625  * first usable address after the mapped region.
1626  */
1627 vm_offset_t
1628 moea_map(mmu_t mmu, vm_offset_t *virt, vm_paddr_t pa_start,
1629     vm_paddr_t pa_end, int prot)
1630 {
1631 	vm_offset_t	sva, va;
1632 
1633 	sva = *virt;
1634 	va = sva;
1635 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1636 		moea_kenter(mmu, va, pa_start);
1637 	*virt = va;
1638 	return (sva);
1639 }
1640 
1641 /*
1642  * Returns true if the pmap's pv is one of the first
1643  * 16 pvs linked to from this page.  This count may
1644  * be changed upwards or downwards in the future; it
1645  * is only necessary that true be returned for a small
1646  * subset of pmaps for proper page aging.
1647  */
1648 boolean_t
1649 moea_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1650 {
1651         int loops;
1652 	struct pvo_entry *pvo;
1653 	boolean_t rv;
1654 
1655 	KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1656 	    ("moea_page_exists_quick: page %p is not managed", m));
1657 	loops = 0;
1658 	rv = FALSE;
1659 	rw_wlock(&pvh_global_lock);
1660 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1661 		if (pvo->pvo_pmap == pmap) {
1662 			rv = TRUE;
1663 			break;
1664 		}
1665 		if (++loops >= 16)
1666 			break;
1667 	}
1668 	rw_wunlock(&pvh_global_lock);
1669 	return (rv);
1670 }
1671 
1672 void
1673 moea_page_init(mmu_t mmu __unused, vm_page_t m)
1674 {
1675 
1676 	m->md.mdpg_attrs = 0;
1677 	m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
1678 	LIST_INIT(&m->md.mdpg_pvoh);
1679 }
1680 
1681 /*
1682  * Return the number of managed mappings to the given physical page
1683  * that are wired.
1684  */
1685 int
1686 moea_page_wired_mappings(mmu_t mmu, vm_page_t m)
1687 {
1688 	struct pvo_entry *pvo;
1689 	int count;
1690 
1691 	count = 0;
1692 	if ((m->oflags & VPO_UNMANAGED) != 0)
1693 		return (count);
1694 	rw_wlock(&pvh_global_lock);
1695 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1696 		if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1697 			count++;
1698 	rw_wunlock(&pvh_global_lock);
1699 	return (count);
1700 }
1701 
1702 static u_int	moea_vsidcontext;
1703 
1704 void
1705 moea_pinit(mmu_t mmu, pmap_t pmap)
1706 {
1707 	int	i, mask;
1708 	u_int	entropy;
1709 
1710 	KASSERT((int)pmap < VM_MIN_KERNEL_ADDRESS, ("moea_pinit: virt pmap"));
1711 	RB_INIT(&pmap->pmap_pvo);
1712 
1713 	entropy = 0;
1714 	__asm __volatile("mftb %0" : "=r"(entropy));
1715 
1716 	if ((pmap->pmap_phys = (pmap_t)moea_kextract(mmu, (vm_offset_t)pmap))
1717 	    == NULL) {
1718 		pmap->pmap_phys = pmap;
1719 	}
1720 
1721 
1722 	mtx_lock(&moea_vsid_mutex);
1723 	/*
1724 	 * Allocate some segment registers for this pmap.
1725 	 */
1726 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1727 		u_int	hash, n;
1728 
1729 		/*
1730 		 * Create a new value by mutiplying by a prime and adding in
1731 		 * entropy from the timebase register.  This is to make the
1732 		 * VSID more random so that the PT hash function collides
1733 		 * less often.  (Note that the prime casues gcc to do shifts
1734 		 * instead of a multiply.)
1735 		 */
1736 		moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1737 		hash = moea_vsidcontext & (NPMAPS - 1);
1738 		if (hash == 0)		/* 0 is special, avoid it */
1739 			continue;
1740 		n = hash >> 5;
1741 		mask = 1 << (hash & (VSID_NBPW - 1));
1742 		hash = (moea_vsidcontext & 0xfffff);
1743 		if (moea_vsid_bitmap[n] & mask) {	/* collision? */
1744 			/* anything free in this bucket? */
1745 			if (moea_vsid_bitmap[n] == 0xffffffff) {
1746 				entropy = (moea_vsidcontext >> 20);
1747 				continue;
1748 			}
1749 			i = ffs(~moea_vsid_bitmap[n]) - 1;
1750 			mask = 1 << i;
1751 			hash &= rounddown2(0xfffff, VSID_NBPW);
1752 			hash |= i;
1753 		}
1754 		KASSERT(!(moea_vsid_bitmap[n] & mask),
1755 		    ("Allocating in-use VSID group %#x\n", hash));
1756 		moea_vsid_bitmap[n] |= mask;
1757 		for (i = 0; i < 16; i++)
1758 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1759 		mtx_unlock(&moea_vsid_mutex);
1760 		return;
1761 	}
1762 
1763 	mtx_unlock(&moea_vsid_mutex);
1764 	panic("moea_pinit: out of segments");
1765 }
1766 
1767 /*
1768  * Initialize the pmap associated with process 0.
1769  */
1770 void
1771 moea_pinit0(mmu_t mmu, pmap_t pm)
1772 {
1773 
1774 	PMAP_LOCK_INIT(pm);
1775 	moea_pinit(mmu, pm);
1776 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1777 }
1778 
1779 /*
1780  * Set the physical protection on the specified range of this map as requested.
1781  */
1782 void
1783 moea_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1784     vm_prot_t prot)
1785 {
1786 	struct	pvo_entry *pvo, *tpvo, key;
1787 	struct	pte *pt;
1788 
1789 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1790 	    ("moea_protect: non current pmap"));
1791 
1792 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1793 		moea_remove(mmu, pm, sva, eva);
1794 		return;
1795 	}
1796 
1797 	rw_wlock(&pvh_global_lock);
1798 	PMAP_LOCK(pm);
1799 	key.pvo_vaddr = sva;
1800 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1801 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1802 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1803 
1804 		/*
1805 		 * Grab the PTE pointer before we diddle with the cached PTE
1806 		 * copy.
1807 		 */
1808 		pt = moea_pvo_to_pte(pvo, -1);
1809 		/*
1810 		 * Change the protection of the page.
1811 		 */
1812 		pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1813 		pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1814 
1815 		/*
1816 		 * If the PVO is in the page table, update that pte as well.
1817 		 */
1818 		if (pt != NULL) {
1819 			moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1820 			mtx_unlock(&moea_table_mutex);
1821 		}
1822 	}
1823 	rw_wunlock(&pvh_global_lock);
1824 	PMAP_UNLOCK(pm);
1825 }
1826 
1827 /*
1828  * Map a list of wired pages into kernel virtual address space.  This is
1829  * intended for temporary mappings which do not need page modification or
1830  * references recorded.  Existing mappings in the region are overwritten.
1831  */
1832 void
1833 moea_qenter(mmu_t mmu, vm_offset_t sva, vm_page_t *m, int count)
1834 {
1835 	vm_offset_t va;
1836 
1837 	va = sva;
1838 	while (count-- > 0) {
1839 		moea_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1840 		va += PAGE_SIZE;
1841 		m++;
1842 	}
1843 }
1844 
1845 /*
1846  * Remove page mappings from kernel virtual address space.  Intended for
1847  * temporary mappings entered by moea_qenter.
1848  */
1849 void
1850 moea_qremove(mmu_t mmu, vm_offset_t sva, int count)
1851 {
1852 	vm_offset_t va;
1853 
1854 	va = sva;
1855 	while (count-- > 0) {
1856 		moea_kremove(mmu, va);
1857 		va += PAGE_SIZE;
1858 	}
1859 }
1860 
1861 void
1862 moea_release(mmu_t mmu, pmap_t pmap)
1863 {
1864         int idx, mask;
1865 
1866 	/*
1867 	 * Free segment register's VSID
1868 	 */
1869         if (pmap->pm_sr[0] == 0)
1870                 panic("moea_release");
1871 
1872 	mtx_lock(&moea_vsid_mutex);
1873         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1874         mask = 1 << (idx % VSID_NBPW);
1875         idx /= VSID_NBPW;
1876         moea_vsid_bitmap[idx] &= ~mask;
1877 	mtx_unlock(&moea_vsid_mutex);
1878 }
1879 
1880 /*
1881  * Remove the given range of addresses from the specified map.
1882  */
1883 void
1884 moea_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1885 {
1886 	struct	pvo_entry *pvo, *tpvo, key;
1887 
1888 	rw_wlock(&pvh_global_lock);
1889 	PMAP_LOCK(pm);
1890 	key.pvo_vaddr = sva;
1891 	for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1892 	    pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1893 		tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1894 		moea_pvo_remove(pvo, -1);
1895 	}
1896 	PMAP_UNLOCK(pm);
1897 	rw_wunlock(&pvh_global_lock);
1898 }
1899 
1900 /*
1901  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1902  * will reflect changes in pte's back to the vm_page.
1903  */
1904 void
1905 moea_remove_all(mmu_t mmu, vm_page_t m)
1906 {
1907 	struct  pvo_head *pvo_head;
1908 	struct	pvo_entry *pvo, *next_pvo;
1909 	pmap_t	pmap;
1910 
1911 	rw_wlock(&pvh_global_lock);
1912 	pvo_head = vm_page_to_pvoh(m);
1913 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1914 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1915 
1916 		pmap = pvo->pvo_pmap;
1917 		PMAP_LOCK(pmap);
1918 		moea_pvo_remove(pvo, -1);
1919 		PMAP_UNLOCK(pmap);
1920 	}
1921 	if ((m->aflags & PGA_WRITEABLE) && moea_query_bit(m, PTE_CHG)) {
1922 		moea_attr_clear(m, PTE_CHG);
1923 		vm_page_dirty(m);
1924 	}
1925 	vm_page_aflag_clear(m, PGA_WRITEABLE);
1926 	rw_wunlock(&pvh_global_lock);
1927 }
1928 
1929 /*
1930  * Allocate a physical page of memory directly from the phys_avail map.
1931  * Can only be called from moea_bootstrap before avail start and end are
1932  * calculated.
1933  */
1934 static vm_offset_t
1935 moea_bootstrap_alloc(vm_size_t size, u_int align)
1936 {
1937 	vm_offset_t	s, e;
1938 	int		i, j;
1939 
1940 	size = round_page(size);
1941 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1942 		if (align != 0)
1943 			s = roundup2(phys_avail[i], align);
1944 		else
1945 			s = phys_avail[i];
1946 		e = s + size;
1947 
1948 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1949 			continue;
1950 
1951 		if (s == phys_avail[i]) {
1952 			phys_avail[i] += size;
1953 		} else if (e == phys_avail[i + 1]) {
1954 			phys_avail[i + 1] -= size;
1955 		} else {
1956 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1957 				phys_avail[j] = phys_avail[j - 2];
1958 				phys_avail[j + 1] = phys_avail[j - 1];
1959 			}
1960 
1961 			phys_avail[i + 3] = phys_avail[i + 1];
1962 			phys_avail[i + 1] = s;
1963 			phys_avail[i + 2] = e;
1964 			phys_avail_count++;
1965 		}
1966 
1967 		return (s);
1968 	}
1969 	panic("moea_bootstrap_alloc: could not allocate memory");
1970 }
1971 
1972 static void
1973 moea_syncicache(vm_paddr_t pa, vm_size_t len)
1974 {
1975 	__syncicache((void *)pa, len);
1976 }
1977 
1978 static int
1979 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1980     vm_offset_t va, vm_paddr_t pa, u_int pte_lo, int flags)
1981 {
1982 	struct	pvo_entry *pvo;
1983 	u_int	sr;
1984 	int	first;
1985 	u_int	ptegidx;
1986 	int	i;
1987 	int     bootstrap;
1988 
1989 	moea_pvo_enter_calls++;
1990 	first = 0;
1991 	bootstrap = 0;
1992 
1993 	/*
1994 	 * Compute the PTE Group index.
1995 	 */
1996 	va &= ~ADDR_POFF;
1997 	sr = va_to_sr(pm->pm_sr, va);
1998 	ptegidx = va_to_pteg(sr, va);
1999 
2000 	/*
2001 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
2002 	 * there is a mapping.
2003 	 */
2004 	mtx_lock(&moea_table_mutex);
2005 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2006 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2007 			if ((pvo->pvo_pte.pte.pte_lo & PTE_RPGN) == pa &&
2008 			    (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
2009 			    (pte_lo & PTE_PP)) {
2010 				/*
2011 				 * The PTE is not changing.  Instead, this may
2012 				 * be a request to change the mapping's wired
2013 				 * attribute.
2014 				 */
2015 				mtx_unlock(&moea_table_mutex);
2016 				if ((flags & PVO_WIRED) != 0 &&
2017 				    (pvo->pvo_vaddr & PVO_WIRED) == 0) {
2018 					pvo->pvo_vaddr |= PVO_WIRED;
2019 					pm->pm_stats.wired_count++;
2020 				} else if ((flags & PVO_WIRED) == 0 &&
2021 				    (pvo->pvo_vaddr & PVO_WIRED) != 0) {
2022 					pvo->pvo_vaddr &= ~PVO_WIRED;
2023 					pm->pm_stats.wired_count--;
2024 				}
2025 				return (0);
2026 			}
2027 			moea_pvo_remove(pvo, -1);
2028 			break;
2029 		}
2030 	}
2031 
2032 	/*
2033 	 * If we aren't overwriting a mapping, try to allocate.
2034 	 */
2035 	if (moea_initialized) {
2036 		pvo = uma_zalloc(zone, M_NOWAIT);
2037 	} else {
2038 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
2039 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
2040 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
2041 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
2042 		}
2043 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
2044 		moea_bpvo_pool_index++;
2045 		bootstrap = 1;
2046 	}
2047 
2048 	if (pvo == NULL) {
2049 		mtx_unlock(&moea_table_mutex);
2050 		return (ENOMEM);
2051 	}
2052 
2053 	moea_pvo_entries++;
2054 	pvo->pvo_vaddr = va;
2055 	pvo->pvo_pmap = pm;
2056 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
2057 	pvo->pvo_vaddr &= ~ADDR_POFF;
2058 	if (flags & PVO_WIRED)
2059 		pvo->pvo_vaddr |= PVO_WIRED;
2060 	if (pvo_head != &moea_pvo_kunmanaged)
2061 		pvo->pvo_vaddr |= PVO_MANAGED;
2062 	if (bootstrap)
2063 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
2064 
2065 	moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo);
2066 
2067 	/*
2068 	 * Add to pmap list
2069 	 */
2070 	RB_INSERT(pvo_tree, &pm->pmap_pvo, pvo);
2071 
2072 	/*
2073 	 * Remember if the list was empty and therefore will be the first
2074 	 * item.
2075 	 */
2076 	if (LIST_FIRST(pvo_head) == NULL)
2077 		first = 1;
2078 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2079 
2080 	if (pvo->pvo_vaddr & PVO_WIRED)
2081 		pm->pm_stats.wired_count++;
2082 	pm->pm_stats.resident_count++;
2083 
2084 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2085 	KASSERT(i < 8, ("Invalid PTE index"));
2086 	if (i >= 0) {
2087 		PVO_PTEGIDX_SET(pvo, i);
2088 	} else {
2089 		panic("moea_pvo_enter: overflow");
2090 		moea_pte_overflow++;
2091 	}
2092 	mtx_unlock(&moea_table_mutex);
2093 
2094 	return (first ? ENOENT : 0);
2095 }
2096 
2097 static void
2098 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
2099 {
2100 	struct	pte *pt;
2101 
2102 	/*
2103 	 * If there is an active pte entry, we need to deactivate it (and
2104 	 * save the ref & cfg bits).
2105 	 */
2106 	pt = moea_pvo_to_pte(pvo, pteidx);
2107 	if (pt != NULL) {
2108 		moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
2109 		mtx_unlock(&moea_table_mutex);
2110 		PVO_PTEGIDX_CLR(pvo);
2111 	} else {
2112 		moea_pte_overflow--;
2113 	}
2114 
2115 	/*
2116 	 * Update our statistics.
2117 	 */
2118 	pvo->pvo_pmap->pm_stats.resident_count--;
2119 	if (pvo->pvo_vaddr & PVO_WIRED)
2120 		pvo->pvo_pmap->pm_stats.wired_count--;
2121 
2122 	/*
2123 	 * Save the REF/CHG bits into their cache if the page is managed.
2124 	 */
2125 	if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED) {
2126 		struct	vm_page *pg;
2127 
2128 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN);
2129 		if (pg != NULL) {
2130 			moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2131 			    (PTE_REF | PTE_CHG));
2132 		}
2133 	}
2134 
2135 	/*
2136 	 * Remove this PVO from the PV and pmap lists.
2137 	 */
2138 	LIST_REMOVE(pvo, pvo_vlink);
2139 	RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
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 		/*
2484 		 * See if we saved the bit off.  If so, cache it and return
2485 		 * success.
2486 		 */
2487 		if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2488 			moea_attr_save(m, ptebit);
2489 			return (TRUE);
2490 		}
2491 	}
2492 
2493 	/*
2494 	 * No luck, now go through the hard part of looking at the PTEs
2495 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2496 	 * the PTEs.
2497 	 */
2498 	powerpc_sync();
2499 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2500 
2501 		/*
2502 		 * See if this pvo has a valid PTE.  if so, fetch the
2503 		 * REF/CHG bits from the valid PTE.  If the appropriate
2504 		 * ptebit is set, cache it and return success.
2505 		 */
2506 		pt = moea_pvo_to_pte(pvo, -1);
2507 		if (pt != NULL) {
2508 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2509 			mtx_unlock(&moea_table_mutex);
2510 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2511 				moea_attr_save(m, ptebit);
2512 				return (TRUE);
2513 			}
2514 		}
2515 	}
2516 
2517 	return (FALSE);
2518 }
2519 
2520 static u_int
2521 moea_clear_bit(vm_page_t m, int ptebit)
2522 {
2523 	u_int	count;
2524 	struct	pvo_entry *pvo;
2525 	struct	pte *pt;
2526 
2527 	rw_assert(&pvh_global_lock, RA_WLOCKED);
2528 
2529 	/*
2530 	 * Clear the cached value.
2531 	 */
2532 	moea_attr_clear(m, ptebit);
2533 
2534 	/*
2535 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2536 	 * we can reset the right ones).  note that since the pvo entries and
2537 	 * list heads are accessed via BAT0 and are never placed in the page
2538 	 * table, we don't have to worry about further accesses setting the
2539 	 * REF/CHG bits.
2540 	 */
2541 	powerpc_sync();
2542 
2543 	/*
2544 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2545 	 * valid pte clear the ptebit from the valid pte.
2546 	 */
2547 	count = 0;
2548 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2549 		pt = moea_pvo_to_pte(pvo, -1);
2550 		if (pt != NULL) {
2551 			moea_pte_synch(pt, &pvo->pvo_pte.pte);
2552 			if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2553 				count++;
2554 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2555 			}
2556 			mtx_unlock(&moea_table_mutex);
2557 		}
2558 		pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2559 	}
2560 
2561 	return (count);
2562 }
2563 
2564 /*
2565  * Return true if the physical range is encompassed by the battable[idx]
2566  */
2567 static int
2568 moea_bat_mapped(int idx, vm_paddr_t pa, vm_size_t size)
2569 {
2570 	u_int prot;
2571 	u_int32_t start;
2572 	u_int32_t end;
2573 	u_int32_t bat_ble;
2574 
2575 	/*
2576 	 * Return immediately if not a valid mapping
2577 	 */
2578 	if (!(battable[idx].batu & BAT_Vs))
2579 		return (EINVAL);
2580 
2581 	/*
2582 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2583 	 * so it can function as an i/o page
2584 	 */
2585 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2586 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2587 		return (EPERM);
2588 
2589 	/*
2590 	 * The address should be within the BAT range. Assume that the
2591 	 * start address in the BAT has the correct alignment (thus
2592 	 * not requiring masking)
2593 	 */
2594 	start = battable[idx].batl & BAT_PBS;
2595 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2596 	end = start | (bat_ble << 15) | 0x7fff;
2597 
2598 	if ((pa < start) || ((pa + size) > end))
2599 		return (ERANGE);
2600 
2601 	return (0);
2602 }
2603 
2604 boolean_t
2605 moea_dev_direct_mapped(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2606 {
2607 	int i;
2608 
2609 	/*
2610 	 * This currently does not work for entries that
2611 	 * overlap 256M BAT segments.
2612 	 */
2613 
2614 	for(i = 0; i < 16; i++)
2615 		if (moea_bat_mapped(i, pa, size) == 0)
2616 			return (0);
2617 
2618 	return (EFAULT);
2619 }
2620 
2621 /*
2622  * Map a set of physical memory pages into the kernel virtual
2623  * address space. Return a pointer to where it is mapped. This
2624  * routine is intended to be used for mapping device memory,
2625  * NOT real memory.
2626  */
2627 void *
2628 moea_mapdev(mmu_t mmu, vm_paddr_t pa, vm_size_t size)
2629 {
2630 
2631 	return (moea_mapdev_attr(mmu, pa, size, VM_MEMATTR_DEFAULT));
2632 }
2633 
2634 void *
2635 moea_mapdev_attr(mmu_t mmu, vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
2636 {
2637 	vm_offset_t va, tmpva, ppa, offset;
2638 	int i;
2639 
2640 	ppa = trunc_page(pa);
2641 	offset = pa & PAGE_MASK;
2642 	size = roundup(offset + size, PAGE_SIZE);
2643 
2644 	/*
2645 	 * If the physical address lies within a valid BAT table entry,
2646 	 * return the 1:1 mapping. This currently doesn't work
2647 	 * for regions that overlap 256M BAT segments.
2648 	 */
2649 	for (i = 0; i < 16; i++) {
2650 		if (moea_bat_mapped(i, pa, size) == 0)
2651 			return ((void *) pa);
2652 	}
2653 
2654 	va = kva_alloc(size);
2655 	if (!va)
2656 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2657 
2658 	for (tmpva = va; size > 0;) {
2659 		moea_kenter_attr(mmu, tmpva, ppa, ma);
2660 		tlbie(tmpva);
2661 		size -= PAGE_SIZE;
2662 		tmpva += PAGE_SIZE;
2663 		ppa += PAGE_SIZE;
2664 	}
2665 
2666 	return ((void *)(va + offset));
2667 }
2668 
2669 void
2670 moea_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2671 {
2672 	vm_offset_t base, offset;
2673 
2674 	/*
2675 	 * If this is outside kernel virtual space, then it's a
2676 	 * battable entry and doesn't require unmapping
2677 	 */
2678 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2679 		base = trunc_page(va);
2680 		offset = va & PAGE_MASK;
2681 		size = roundup(offset + size, PAGE_SIZE);
2682 		kva_free(base, size);
2683 	}
2684 }
2685 
2686 static void
2687 moea_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz)
2688 {
2689 	struct pvo_entry *pvo;
2690 	vm_offset_t lim;
2691 	vm_paddr_t pa;
2692 	vm_size_t len;
2693 
2694 	PMAP_LOCK(pm);
2695 	while (sz > 0) {
2696 		lim = round_page(va);
2697 		len = MIN(lim - va, sz);
2698 		pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2699 		if (pvo != NULL) {
2700 			pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) |
2701 			    (va & ADDR_POFF);
2702 			moea_syncicache(pa, len);
2703 		}
2704 		va += len;
2705 		sz -= len;
2706 	}
2707 	PMAP_UNLOCK(pm);
2708 }
2709 
2710 void
2711 moea_dumpsys_map(mmu_t mmu, vm_paddr_t pa, size_t sz, void **va)
2712 {
2713 
2714 	*va = (void *)pa;
2715 }
2716 
2717 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
2718 
2719 void
2720 moea_scan_init(mmu_t mmu)
2721 {
2722 	struct pvo_entry *pvo;
2723 	vm_offset_t va;
2724 	int i;
2725 
2726 	if (!do_minidump) {
2727 		/* Initialize phys. segments for dumpsys(). */
2728 		memset(&dump_map, 0, sizeof(dump_map));
2729 		mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
2730 		for (i = 0; i < pregions_sz; i++) {
2731 			dump_map[i].pa_start = pregions[i].mr_start;
2732 			dump_map[i].pa_size = pregions[i].mr_size;
2733 		}
2734 		return;
2735 	}
2736 
2737 	/* Virtual segments for minidumps: */
2738 	memset(&dump_map, 0, sizeof(dump_map));
2739 
2740 	/* 1st: kernel .data and .bss. */
2741 	dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
2742 	dump_map[0].pa_size =
2743 	    round_page((uintptr_t)_end) - dump_map[0].pa_start;
2744 
2745 	/* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2746 	dump_map[1].pa_start = (vm_paddr_t)msgbufp->msg_ptr;
2747 	dump_map[1].pa_size = round_page(msgbufp->msg_size);
2748 
2749 	/* 3rd: kernel VM. */
2750 	va = dump_map[1].pa_start + dump_map[1].pa_size;
2751 	/* Find start of next chunk (from va). */
2752 	while (va < virtual_end) {
2753 		/* Don't dump the buffer cache. */
2754 		if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
2755 			va = kmi.buffer_eva;
2756 			continue;
2757 		}
2758 		pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
2759 		if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2760 			break;
2761 		va += PAGE_SIZE;
2762 	}
2763 	if (va < virtual_end) {
2764 		dump_map[2].pa_start = va;
2765 		va += PAGE_SIZE;
2766 		/* Find last page in chunk. */
2767 		while (va < virtual_end) {
2768 			/* Don't run into the buffer cache. */
2769 			if (va == kmi.buffer_sva)
2770 				break;
2771 			pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF,
2772 			    NULL);
2773 			if (pvo == NULL ||
2774 			    !(pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2775 				break;
2776 			va += PAGE_SIZE;
2777 		}
2778 		dump_map[2].pa_size = va - dump_map[2].pa_start;
2779 	}
2780 }
2781