xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision 4f29da19bd44f0e99f021510460a81bf754c21d2)
1 /*-
2  * Copyright (c) 2001 The NetBSD Foundation, Inc.
3  * All rights reserved.
4  *
5  * This code is derived from software contributed to The NetBSD Foundation
6  * by Matt Thomas <matt@3am-software.com> of Allegro Networks, Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. All advertising materials mentioning features or use of this software
17  *    must display the following acknowledgement:
18  *        This product includes software developed by the NetBSD
19  *        Foundation, Inc. and its contributors.
20  * 4. Neither the name of The NetBSD Foundation nor the names of its
21  *    contributors may be used to endorse or promote products derived
22  *    from this software without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
25  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
26  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
27  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
28  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
29  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
30  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
31  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
32  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
33  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
34  * POSSIBILITY OF SUCH DAMAGE.
35  */
36 /*-
37  * Copyright (C) 1995, 1996 Wolfgang Solfrank.
38  * Copyright (C) 1995, 1996 TooLs GmbH.
39  * All rights reserved.
40  *
41  * Redistribution and use in source and binary forms, with or without
42  * modification, are permitted provided that the following conditions
43  * are met:
44  * 1. Redistributions of source code must retain the above copyright
45  *    notice, this list of conditions and the following disclaimer.
46  * 2. Redistributions in binary form must reproduce the above copyright
47  *    notice, this list of conditions and the following disclaimer in the
48  *    documentation and/or other materials provided with the distribution.
49  * 3. All advertising materials mentioning features or use of this software
50  *    must display the following acknowledgement:
51  *	This product includes software developed by TooLs GmbH.
52  * 4. The name of TooLs GmbH may not be used to endorse or promote products
53  *    derived from this software without specific prior written permission.
54  *
55  * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
56  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
57  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
58  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
59  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
60  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
61  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
62  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
63  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
64  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
65  *
66  * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $
67  */
68 /*-
69  * Copyright (C) 2001 Benno Rice.
70  * All rights reserved.
71  *
72  * Redistribution and use in source and binary forms, with or without
73  * modification, are permitted provided that the following conditions
74  * are met:
75  * 1. Redistributions of source code must retain the above copyright
76  *    notice, this list of conditions and the following disclaimer.
77  * 2. Redistributions in binary form must reproduce the above copyright
78  *    notice, this list of conditions and the following disclaimer in the
79  *    documentation and/or other materials provided with the distribution.
80  *
81  * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR
82  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
83  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
84  * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
85  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
86  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
87  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
88  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
89  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
90  * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
91  */
92 
93 #include <sys/cdefs.h>
94 __FBSDID("$FreeBSD$");
95 
96 /*
97  * Manages physical address maps.
98  *
99  * In addition to hardware address maps, this module is called upon to
100  * provide software-use-only maps which may or may not be stored in the
101  * same form as hardware maps.  These pseudo-maps are used to store
102  * intermediate results from copy operations to and from address spaces.
103  *
104  * Since the information managed by this module is also stored by the
105  * logical address mapping module, this module may throw away valid virtual
106  * to physical mappings at almost any time.  However, invalidations of
107  * mappings must be done as requested.
108  *
109  * In order to cope with hardware architectures which make virtual to
110  * physical map invalidates expensive, this module may delay invalidate
111  * reduced protection operations until such time as they are actually
112  * necessary.  This module is given full information as to which processors
113  * are currently using which maps, and to when physical maps must be made
114  * correct.
115  */
116 
117 #include "opt_kstack_pages.h"
118 
119 #include <sys/param.h>
120 #include <sys/kernel.h>
121 #include <sys/ktr.h>
122 #include <sys/lock.h>
123 #include <sys/msgbuf.h>
124 #include <sys/mutex.h>
125 #include <sys/proc.h>
126 #include <sys/sysctl.h>
127 #include <sys/systm.h>
128 #include <sys/vmmeter.h>
129 
130 #include <dev/ofw/openfirm.h>
131 
132 #include <vm/vm.h>
133 #include <vm/vm_param.h>
134 #include <vm/vm_kern.h>
135 #include <vm/vm_page.h>
136 #include <vm/vm_map.h>
137 #include <vm/vm_object.h>
138 #include <vm/vm_extern.h>
139 #include <vm/vm_pageout.h>
140 #include <vm/vm_pager.h>
141 #include <vm/uma.h>
142 
143 #include <machine/cpu.h>
144 #include <machine/powerpc.h>
145 #include <machine/bat.h>
146 #include <machine/frame.h>
147 #include <machine/md_var.h>
148 #include <machine/psl.h>
149 #include <machine/pte.h>
150 #include <machine/sr.h>
151 #include <machine/mmuvar.h>
152 
153 #include "mmu_if.h"
154 
155 #define	MOEA_DEBUG
156 
157 #define TODO	panic("%s: not implemented", __func__);
158 
159 #define	TLBIE(va)	__asm __volatile("tlbie %0" :: "r"(va))
160 #define	TLBSYNC()	__asm __volatile("tlbsync");
161 #define	SYNC()		__asm __volatile("sync");
162 #define	EIEIO()		__asm __volatile("eieio");
163 
164 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
165 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
166 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
167 
168 #define	PVO_PTEGIDX_MASK	0x007		/* which PTEG slot */
169 #define	PVO_PTEGIDX_VALID	0x008		/* slot is valid */
170 #define	PVO_WIRED		0x010		/* PVO entry is wired */
171 #define	PVO_MANAGED		0x020		/* PVO entry is managed */
172 #define	PVO_EXECUTABLE		0x040		/* PVO entry is executable */
173 #define	PVO_BOOTSTRAP		0x080		/* PVO entry allocated during
174 						   bootstrap */
175 #define PVO_FAKE		0x100		/* fictitious phys page */
176 #define	PVO_VADDR(pvo)		((pvo)->pvo_vaddr & ~ADDR_POFF)
177 #define	PVO_ISEXECUTABLE(pvo)	((pvo)->pvo_vaddr & PVO_EXECUTABLE)
178 #define PVO_ISFAKE(pvo)		((pvo)->pvo_vaddr & PVO_FAKE)
179 #define	PVO_PTEGIDX_GET(pvo)	((pvo)->pvo_vaddr & PVO_PTEGIDX_MASK)
180 #define	PVO_PTEGIDX_ISSET(pvo)	((pvo)->pvo_vaddr & PVO_PTEGIDX_VALID)
181 #define	PVO_PTEGIDX_CLR(pvo)	\
182 	((void)((pvo)->pvo_vaddr &= ~(PVO_PTEGIDX_VALID|PVO_PTEGIDX_MASK)))
183 #define	PVO_PTEGIDX_SET(pvo, i)	\
184 	((void)((pvo)->pvo_vaddr |= (i)|PVO_PTEGIDX_VALID))
185 
186 #define	MOEA_PVO_CHECK(pvo)
187 
188 struct ofw_map {
189 	vm_offset_t	om_va;
190 	vm_size_t	om_len;
191 	vm_offset_t	om_pa;
192 	u_int		om_mode;
193 };
194 
195 /*
196  * Map of physical memory regions.
197  */
198 static struct	mem_region *regions;
199 static struct	mem_region *pregions;
200 u_int           phys_avail_count;
201 int		regions_sz, pregions_sz;
202 static struct	ofw_map *translations;
203 
204 extern struct pmap ofw_pmap;
205 
206 
207 
208 /*
209  * Lock for the pteg and pvo tables.
210  */
211 struct mtx	moea_table_mutex;
212 
213 /*
214  * PTEG data.
215  */
216 static struct	pteg *moea_pteg_table;
217 u_int		moea_pteg_count;
218 u_int		moea_pteg_mask;
219 
220 /*
221  * PVO data.
222  */
223 struct	pvo_head *moea_pvo_table;		/* pvo entries by pteg index */
224 struct	pvo_head moea_pvo_kunmanaged =
225     LIST_HEAD_INITIALIZER(moea_pvo_kunmanaged);	/* list of unmanaged pages */
226 struct	pvo_head moea_pvo_unmanaged =
227     LIST_HEAD_INITIALIZER(moea_pvo_unmanaged);	/* list of unmanaged pages */
228 
229 uma_zone_t	moea_upvo_zone;	/* zone for pvo entries for unmanaged pages */
230 uma_zone_t	moea_mpvo_zone;	/* zone for pvo entries for managed pages */
231 
232 #define	BPVO_POOL_SIZE	32768
233 static struct	pvo_entry *moea_bpvo_pool;
234 static int	moea_bpvo_pool_index = 0;
235 
236 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
237 static u_int	moea_vsid_bitmap[NPMAPS / VSID_NBPW];
238 
239 static boolean_t moea_initialized = FALSE;
240 
241 /*
242  * Statistics.
243  */
244 u_int	moea_pte_valid = 0;
245 u_int	moea_pte_overflow = 0;
246 u_int	moea_pte_replacements = 0;
247 u_int	moea_pvo_entries = 0;
248 u_int	moea_pvo_enter_calls = 0;
249 u_int	moea_pvo_remove_calls = 0;
250 u_int	moea_pte_spills = 0;
251 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_valid, CTLFLAG_RD, &moea_pte_valid,
252     0, "");
253 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_overflow, CTLFLAG_RD,
254     &moea_pte_overflow, 0, "");
255 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_replacements, CTLFLAG_RD,
256     &moea_pte_replacements, 0, "");
257 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_entries, CTLFLAG_RD, &moea_pvo_entries,
258     0, "");
259 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_enter_calls, CTLFLAG_RD,
260     &moea_pvo_enter_calls, 0, "");
261 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_remove_calls, CTLFLAG_RD,
262     &moea_pvo_remove_calls, 0, "");
263 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_spills, CTLFLAG_RD,
264     &moea_pte_spills, 0, "");
265 
266 struct	pvo_entry *moea_pvo_zeropage;
267 
268 vm_offset_t	moea_rkva_start = VM_MIN_KERNEL_ADDRESS;
269 u_int		moea_rkva_count = 4;
270 
271 /*
272  * Allocate physical memory for use in moea_bootstrap.
273  */
274 static vm_offset_t	moea_bootstrap_alloc(vm_size_t, u_int);
275 
276 /*
277  * PTE calls.
278  */
279 static int		moea_pte_insert(u_int, struct pte *);
280 
281 /*
282  * PVO calls.
283  */
284 static int	moea_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *,
285 		    vm_offset_t, vm_offset_t, u_int, int);
286 static void	moea_pvo_remove(struct pvo_entry *, int);
287 static struct	pvo_entry *moea_pvo_find_va(pmap_t, vm_offset_t, int *);
288 static struct	pte *moea_pvo_to_pte(const struct pvo_entry *, int);
289 
290 /*
291  * Utility routines.
292  */
293 static struct		pvo_entry *moea_rkva_alloc(mmu_t);
294 static void		moea_pa_map(struct pvo_entry *, vm_offset_t,
295 			    struct pte *, int *);
296 static void		moea_pa_unmap(struct pvo_entry *, struct pte *, int *);
297 static void		moea_syncicache(vm_offset_t, vm_size_t);
298 static boolean_t	moea_query_bit(vm_page_t, int);
299 static u_int		moea_clear_bit(vm_page_t, int, int *);
300 static void		moea_kremove(mmu_t, vm_offset_t);
301 static void		tlbia(void);
302 int		moea_pte_spill(vm_offset_t);
303 
304 /*
305  * Kernel MMU interface
306  */
307 void moea_change_wiring(mmu_t, pmap_t, vm_offset_t, boolean_t);
308 void moea_clear_modify(mmu_t, vm_page_t);
309 void moea_clear_reference(mmu_t, vm_page_t);
310 void moea_copy_page(mmu_t, vm_page_t, vm_page_t);
311 void moea_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, boolean_t);
312 vm_page_t moea_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t,
313     vm_page_t);
314 vm_paddr_t moea_extract(mmu_t, pmap_t, vm_offset_t);
315 vm_page_t moea_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t);
316 void moea_init(mmu_t);
317 boolean_t moea_is_modified(mmu_t, vm_page_t);
318 boolean_t moea_ts_referenced(mmu_t, vm_page_t);
319 vm_offset_t moea_map(mmu_t, vm_offset_t *, vm_offset_t, vm_offset_t, int);
320 boolean_t moea_page_exists_quick(mmu_t, pmap_t, vm_page_t);
321 void moea_page_protect(mmu_t, vm_page_t, vm_prot_t);
322 void moea_pinit(mmu_t, pmap_t);
323 void moea_pinit0(mmu_t, pmap_t);
324 void moea_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
325 void moea_qenter(mmu_t, vm_offset_t, vm_page_t *, int);
326 void moea_qremove(mmu_t, vm_offset_t, int);
327 void moea_release(mmu_t, pmap_t);
328 void moea_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t);
329 void moea_remove_all(mmu_t, vm_page_t);
330 void moea_zero_page(mmu_t, vm_page_t);
331 void moea_zero_page_area(mmu_t, vm_page_t, int, int);
332 void moea_zero_page_idle(mmu_t, vm_page_t);
333 void moea_activate(mmu_t, struct thread *);
334 void moea_deactivate(mmu_t, struct thread *);
335 void moea_bootstrap(mmu_t, vm_offset_t, vm_offset_t);
336 void *moea_mapdev(mmu_t, vm_offset_t, vm_size_t);
337 void moea_unmapdev(mmu_t, vm_offset_t, vm_size_t);
338 vm_offset_t moea_kextract(mmu_t, vm_offset_t);
339 void moea_kenter(mmu_t, vm_offset_t, vm_offset_t);
340 boolean_t moea_dev_direct_mapped(mmu_t, vm_offset_t, vm_size_t);
341 
342 static mmu_method_t moea_methods[] = {
343 	MMUMETHOD(mmu_change_wiring,	moea_change_wiring),
344 	MMUMETHOD(mmu_clear_modify,	moea_clear_modify),
345 	MMUMETHOD(mmu_clear_reference,	moea_clear_reference),
346 	MMUMETHOD(mmu_copy_page,	moea_copy_page),
347 	MMUMETHOD(mmu_enter,		moea_enter),
348 	MMUMETHOD(mmu_enter_quick,	moea_enter_quick),
349 	MMUMETHOD(mmu_extract,		moea_extract),
350 	MMUMETHOD(mmu_extract_and_hold,	moea_extract_and_hold),
351 	MMUMETHOD(mmu_init,		moea_init),
352 	MMUMETHOD(mmu_is_modified,	moea_is_modified),
353 	MMUMETHOD(mmu_ts_referenced,	moea_ts_referenced),
354 	MMUMETHOD(mmu_map,     		moea_map),
355 	MMUMETHOD(mmu_page_exists_quick,moea_page_exists_quick),
356 	MMUMETHOD(mmu_page_protect,    	moea_page_protect),
357 	MMUMETHOD(mmu_pinit,		moea_pinit),
358 	MMUMETHOD(mmu_pinit0,		moea_pinit0),
359 	MMUMETHOD(mmu_protect,		moea_protect),
360 	MMUMETHOD(mmu_qenter,		moea_qenter),
361 	MMUMETHOD(mmu_qremove,		moea_qremove),
362 	MMUMETHOD(mmu_release,		moea_release),
363 	MMUMETHOD(mmu_remove,		moea_remove),
364 	MMUMETHOD(mmu_remove_all,      	moea_remove_all),
365 	MMUMETHOD(mmu_zero_page,       	moea_zero_page),
366 	MMUMETHOD(mmu_zero_page_area,	moea_zero_page_area),
367 	MMUMETHOD(mmu_zero_page_idle,	moea_zero_page_idle),
368 	MMUMETHOD(mmu_activate,		moea_activate),
369 	MMUMETHOD(mmu_deactivate,      	moea_deactivate),
370 
371 	/* Internal interfaces */
372 	MMUMETHOD(mmu_bootstrap,       	moea_bootstrap),
373 	MMUMETHOD(mmu_mapdev,		moea_mapdev),
374 	MMUMETHOD(mmu_unmapdev,		moea_unmapdev),
375 	MMUMETHOD(mmu_kextract,		moea_kextract),
376 	MMUMETHOD(mmu_kenter,		moea_kenter),
377 	MMUMETHOD(mmu_dev_direct_mapped,moea_dev_direct_mapped),
378 
379 	{ 0, 0 }
380 };
381 
382 static mmu_def_t oea_mmu = {
383 	MMU_TYPE_OEA,
384 	moea_methods,
385 	0
386 };
387 MMU_DEF(oea_mmu);
388 
389 
390 static __inline int
391 va_to_sr(u_int *sr, vm_offset_t va)
392 {
393 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
394 }
395 
396 static __inline u_int
397 va_to_pteg(u_int sr, vm_offset_t addr)
398 {
399 	u_int hash;
400 
401 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
402 	    ADDR_PIDX_SHFT);
403 	return (hash & moea_pteg_mask);
404 }
405 
406 static __inline struct pvo_head *
407 pa_to_pvoh(vm_offset_t pa, vm_page_t *pg_p)
408 {
409 	struct	vm_page *pg;
410 
411 	pg = PHYS_TO_VM_PAGE(pa);
412 
413 	if (pg_p != NULL)
414 		*pg_p = pg;
415 
416 	if (pg == NULL)
417 		return (&moea_pvo_unmanaged);
418 
419 	return (&pg->md.mdpg_pvoh);
420 }
421 
422 static __inline struct pvo_head *
423 vm_page_to_pvoh(vm_page_t m)
424 {
425 
426 	return (&m->md.mdpg_pvoh);
427 }
428 
429 static __inline void
430 moea_attr_clear(vm_page_t m, int ptebit)
431 {
432 
433 	m->md.mdpg_attrs &= ~ptebit;
434 }
435 
436 static __inline int
437 moea_attr_fetch(vm_page_t m)
438 {
439 
440 	return (m->md.mdpg_attrs);
441 }
442 
443 static __inline void
444 moea_attr_save(vm_page_t m, int ptebit)
445 {
446 
447 	m->md.mdpg_attrs |= ptebit;
448 }
449 
450 static __inline int
451 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
452 {
453 	if (pt->pte_hi == pvo_pt->pte_hi)
454 		return (1);
455 
456 	return (0);
457 }
458 
459 static __inline int
460 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
461 {
462 	return (pt->pte_hi & ~PTE_VALID) ==
463 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
464 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
465 }
466 
467 static __inline void
468 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
469 {
470 	/*
471 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
472 	 * set when the real pte is set in memory.
473 	 *
474 	 * Note: Don't set the valid bit for correct operation of tlb update.
475 	 */
476 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
477 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
478 	pt->pte_lo = pte_lo;
479 }
480 
481 static __inline void
482 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
483 {
484 
485 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
486 }
487 
488 static __inline void
489 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
490 {
491 
492 	/*
493 	 * As shown in Section 7.6.3.2.3
494 	 */
495 	pt->pte_lo &= ~ptebit;
496 	TLBIE(va);
497 	EIEIO();
498 	TLBSYNC();
499 	SYNC();
500 }
501 
502 static __inline void
503 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
504 {
505 
506 	pvo_pt->pte_hi |= PTE_VALID;
507 
508 	/*
509 	 * Update the PTE as defined in section 7.6.3.1.
510 	 * Note that the REF/CHG bits are from pvo_pt and thus should havce
511 	 * been saved so this routine can restore them (if desired).
512 	 */
513 	pt->pte_lo = pvo_pt->pte_lo;
514 	EIEIO();
515 	pt->pte_hi = pvo_pt->pte_hi;
516 	SYNC();
517 	moea_pte_valid++;
518 }
519 
520 static __inline void
521 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
522 {
523 
524 	pvo_pt->pte_hi &= ~PTE_VALID;
525 
526 	/*
527 	 * Force the reg & chg bits back into the PTEs.
528 	 */
529 	SYNC();
530 
531 	/*
532 	 * Invalidate the pte.
533 	 */
534 	pt->pte_hi &= ~PTE_VALID;
535 
536 	SYNC();
537 	TLBIE(va);
538 	EIEIO();
539 	TLBSYNC();
540 	SYNC();
541 
542 	/*
543 	 * Save the reg & chg bits.
544 	 */
545 	moea_pte_synch(pt, pvo_pt);
546 	moea_pte_valid--;
547 }
548 
549 static __inline void
550 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
551 {
552 
553 	/*
554 	 * Invalidate the PTE
555 	 */
556 	moea_pte_unset(pt, pvo_pt, va);
557 	moea_pte_set(pt, pvo_pt);
558 }
559 
560 /*
561  * Quick sort callout for comparing memory regions.
562  */
563 static int	mr_cmp(const void *a, const void *b);
564 static int	om_cmp(const void *a, const void *b);
565 
566 static int
567 mr_cmp(const void *a, const void *b)
568 {
569 	const struct	mem_region *regiona;
570 	const struct	mem_region *regionb;
571 
572 	regiona = a;
573 	regionb = b;
574 	if (regiona->mr_start < regionb->mr_start)
575 		return (-1);
576 	else if (regiona->mr_start > regionb->mr_start)
577 		return (1);
578 	else
579 		return (0);
580 }
581 
582 static int
583 om_cmp(const void *a, const void *b)
584 {
585 	const struct	ofw_map *mapa;
586 	const struct	ofw_map *mapb;
587 
588 	mapa = a;
589 	mapb = b;
590 	if (mapa->om_pa < mapb->om_pa)
591 		return (-1);
592 	else if (mapa->om_pa > mapb->om_pa)
593 		return (1);
594 	else
595 		return (0);
596 }
597 
598 void
599 moea_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend)
600 {
601 	ihandle_t	mmui;
602 	phandle_t	chosen, mmu;
603 	int		sz;
604 	int		i, j;
605 	int		ofw_mappings;
606 	vm_size_t	size, physsz, hwphyssz;
607 	vm_offset_t	pa, va, off;
608 	u_int		batl, batu;
609 
610         /*
611          * Set up BAT0 to map the lowest 256 MB area
612          */
613         battable[0x0].batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
614         battable[0x0].batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
615 
616         /*
617          * Map PCI memory space.
618          */
619         battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
620         battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
621 
622         battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
623         battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
624 
625         battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
626         battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
627 
628         battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
629         battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
630 
631         /*
632          * Map obio devices.
633          */
634         battable[0xf].batl = BATL(0xf0000000, BAT_I|BAT_G, BAT_PP_RW);
635         battable[0xf].batu = BATU(0xf0000000, BAT_BL_256M, BAT_Vs);
636 
637 	/*
638 	 * Use an IBAT and a DBAT to map the bottom segment of memory
639 	 * where we are.
640 	 */
641 	batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
642 	batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
643 	__asm (".balign 32; \n"
644 	       "mtibatu 0,%0; mtibatl 0,%1; isync; \n"
645 	       "mtdbatu 0,%0; mtdbatl 0,%1; isync"
646 	    :: "r"(batu), "r"(batl));
647 
648 #if 0
649 	/* map frame buffer */
650 	batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
651 	batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
652 	__asm ("mtdbatu 1,%0; mtdbatl 1,%1; isync"
653 	    :: "r"(batu), "r"(batl));
654 #endif
655 
656 #if 1
657 	/* map pci space */
658 	batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
659 	batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
660 	__asm ("mtdbatu 1,%0; mtdbatl 1,%1; isync"
661 	    :: "r"(batu), "r"(batl));
662 #endif
663 
664 	/*
665 	 * Set the start and end of kva.
666 	 */
667 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
668 	virtual_end = VM_MAX_KERNEL_ADDRESS;
669 
670 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
671 	CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
672 
673 	qsort(pregions, pregions_sz, sizeof(*pregions), mr_cmp);
674 	for (i = 0; i < pregions_sz; i++) {
675 		vm_offset_t pa;
676 		vm_offset_t end;
677 
678 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
679 			pregions[i].mr_start,
680 			pregions[i].mr_start + pregions[i].mr_size,
681 			pregions[i].mr_size);
682 		/*
683 		 * Install entries into the BAT table to allow all
684 		 * of physmem to be convered by on-demand BAT entries.
685 		 * The loop will sometimes set the same battable element
686 		 * twice, but that's fine since they won't be used for
687 		 * a while yet.
688 		 */
689 		pa = pregions[i].mr_start & 0xf0000000;
690 		end = pregions[i].mr_start + pregions[i].mr_size;
691 		do {
692                         u_int n = pa >> ADDR_SR_SHFT;
693 
694 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
695 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
696 			pa += SEGMENT_LENGTH;
697 		} while (pa < end);
698 	}
699 
700 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
701 		panic("moea_bootstrap: phys_avail too small");
702 	qsort(regions, regions_sz, sizeof(*regions), mr_cmp);
703 	phys_avail_count = 0;
704 	physsz = 0;
705 	hwphyssz = 0;
706 	TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
707 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
708 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
709 		    regions[i].mr_start + regions[i].mr_size,
710 		    regions[i].mr_size);
711 		if (hwphyssz != 0 &&
712 		    (physsz + regions[i].mr_size) >= hwphyssz) {
713 			if (physsz < hwphyssz) {
714 				phys_avail[j] = regions[i].mr_start;
715 				phys_avail[j + 1] = regions[i].mr_start +
716 				    hwphyssz - physsz;
717 				physsz = hwphyssz;
718 				phys_avail_count++;
719 			}
720 			break;
721 		}
722 		phys_avail[j] = regions[i].mr_start;
723 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
724 		phys_avail_count++;
725 		physsz += regions[i].mr_size;
726 	}
727 	physmem = btoc(physsz);
728 
729 	/*
730 	 * Allocate PTEG table.
731 	 */
732 #ifdef PTEGCOUNT
733 	moea_pteg_count = PTEGCOUNT;
734 #else
735 	moea_pteg_count = 0x1000;
736 
737 	while (moea_pteg_count < physmem)
738 		moea_pteg_count <<= 1;
739 
740 	moea_pteg_count >>= 1;
741 #endif /* PTEGCOUNT */
742 
743 	size = moea_pteg_count * sizeof(struct pteg);
744 	CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
745 	    size);
746 	moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
747 	CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
748 	bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
749 	moea_pteg_mask = moea_pteg_count - 1;
750 
751 	/*
752 	 * Allocate pv/overflow lists.
753 	 */
754 	size = sizeof(struct pvo_head) * moea_pteg_count;
755 	moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
756 	    PAGE_SIZE);
757 	CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
758 	for (i = 0; i < moea_pteg_count; i++)
759 		LIST_INIT(&moea_pvo_table[i]);
760 
761 	/*
762 	 * Initialize the lock that synchronizes access to the pteg and pvo
763 	 * tables.
764 	 */
765 	mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF);
766 
767 	/*
768 	 * Allocate the message buffer.
769 	 */
770 	msgbuf_phys = moea_bootstrap_alloc(MSGBUF_SIZE, 0);
771 
772 	/*
773 	 * Initialise the unmanaged pvo pool.
774 	 */
775 	moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
776 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
777 	moea_bpvo_pool_index = 0;
778 
779 	/*
780 	 * Make sure kernel vsid is allocated as well as VSID 0.
781 	 */
782 	moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
783 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
784 	moea_vsid_bitmap[0] |= 1;
785 
786 	/*
787 	 * Set up the Open Firmware pmap and add it's mappings.
788 	 */
789 	moea_pinit(mmup, &ofw_pmap);
790 	ofw_pmap.pm_sr[KERNEL_SR] = KERNEL_SEGMENT;
791 	ofw_pmap.pm_sr[KERNEL2_SR] = KERNEL2_SEGMENT;
792 	if ((chosen = OF_finddevice("/chosen")) == -1)
793 		panic("moea_bootstrap: can't find /chosen");
794 	OF_getprop(chosen, "mmu", &mmui, 4);
795 	if ((mmu = OF_instance_to_package(mmui)) == -1)
796 		panic("moea_bootstrap: can't get mmu package");
797 	if ((sz = OF_getproplen(mmu, "translations")) == -1)
798 		panic("moea_bootstrap: can't get ofw translation count");
799 	translations = NULL;
800 	for (i = 0; phys_avail[i] != 0; i += 2) {
801 		if (phys_avail[i + 1] >= sz) {
802 			translations = (struct ofw_map *)phys_avail[i];
803 			break;
804 		}
805 	}
806 	if (translations == NULL)
807 		panic("moea_bootstrap: no space to copy translations");
808 	bzero(translations, sz);
809 	if (OF_getprop(mmu, "translations", translations, sz) == -1)
810 		panic("moea_bootstrap: can't get ofw translations");
811 	CTR0(KTR_PMAP, "moea_bootstrap: translations");
812 	sz /= sizeof(*translations);
813 	qsort(translations, sz, sizeof (*translations), om_cmp);
814 	for (i = 0, ofw_mappings = 0; i < sz; i++) {
815 		CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
816 		    translations[i].om_pa, translations[i].om_va,
817 		    translations[i].om_len);
818 
819 		/*
820 		 * If the mapping is 1:1, let the RAM and device on-demand
821 		 * BAT tables take care of the translation.
822 		 */
823 		if (translations[i].om_va == translations[i].om_pa)
824 			continue;
825 
826 		/* Enter the pages */
827 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
828 			struct	vm_page m;
829 
830 			m.phys_addr = translations[i].om_pa + off;
831 			moea_enter(mmup, &ofw_pmap,
832 				   translations[i].om_va + off, &m,
833 				   VM_PROT_ALL, 1);
834 			ofw_mappings++;
835 		}
836 	}
837 #ifdef SMP
838 	TLBSYNC();
839 #endif
840 
841 	/*
842 	 * Initialize the kernel pmap (which is statically allocated).
843 	 */
844 	PMAP_LOCK_INIT(kernel_pmap);
845 	for (i = 0; i < 16; i++) {
846 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT;
847 	}
848 	kernel_pmap->pm_sr[KERNEL_SR] = KERNEL_SEGMENT;
849 	kernel_pmap->pm_sr[KERNEL2_SR] = KERNEL2_SEGMENT;
850 	kernel_pmap->pm_active = ~0;
851 
852 	/*
853 	 * Allocate a kernel stack with a guard page for thread0 and map it
854 	 * into the kernel page map.
855 	 */
856 	pa = moea_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, 0);
857 	kstack0_phys = pa;
858 	kstack0 = virtual_avail + (KSTACK_GUARD_PAGES * PAGE_SIZE);
859 	CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", kstack0_phys,
860 	    kstack0);
861 	virtual_avail += (KSTACK_PAGES + KSTACK_GUARD_PAGES) * PAGE_SIZE;
862 	for (i = 0; i < KSTACK_PAGES; i++) {
863 		pa = kstack0_phys + i * PAGE_SIZE;
864 		va = kstack0 + i * PAGE_SIZE;
865 		moea_kenter(mmup, va, pa);
866 		TLBIE(va);
867 	}
868 
869 	/*
870 	 * Calculate the last available physical address.
871 	 */
872 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
873 		;
874 	Maxmem = powerpc_btop(phys_avail[i + 1]);
875 
876 	/*
877 	 * Allocate virtual address space for the message buffer.
878 	 */
879 	msgbufp = (struct msgbuf *)virtual_avail;
880 	virtual_avail += round_page(MSGBUF_SIZE);
881 
882 	/*
883 	 * Initialize hardware.
884 	 */
885 	for (i = 0; i < 16; i++) {
886 		mtsrin(i << ADDR_SR_SHFT, EMPTY_SEGMENT);
887 	}
888 	__asm __volatile ("mtsr %0,%1"
889 	    :: "n"(KERNEL_SR), "r"(KERNEL_SEGMENT));
890 	__asm __volatile ("mtsr %0,%1"
891 	    :: "n"(KERNEL2_SR), "r"(KERNEL2_SEGMENT));
892 	__asm __volatile ("sync; mtsdr1 %0; isync"
893 	    :: "r"((u_int)moea_pteg_table | (moea_pteg_mask >> 10)));
894 	tlbia();
895 
896 	pmap_bootstrapped++;
897 }
898 
899 /*
900  * Activate a user pmap.  The pmap must be activated before it's address
901  * space can be accessed in any way.
902  */
903 void
904 moea_activate(mmu_t mmu, struct thread *td)
905 {
906 	pmap_t	pm, pmr;
907 
908 	/*
909 	 * Load all the data we need up front to encourage the compiler to
910 	 * not issue any loads while we have interrupts disabled below.
911 	 */
912 	pm = &td->td_proc->p_vmspace->vm_pmap;
913 
914 	if ((pmr = (pmap_t)moea_kextract(mmu, (vm_offset_t)pm)) == NULL)
915 		pmr = pm;
916 
917 	pm->pm_active |= PCPU_GET(cpumask);
918 	PCPU_SET(curpmap, pmr);
919 }
920 
921 void
922 moea_deactivate(mmu_t mmu, struct thread *td)
923 {
924 	pmap_t	pm;
925 
926 	pm = &td->td_proc->p_vmspace->vm_pmap;
927 	pm->pm_active &= ~(PCPU_GET(cpumask));
928 	PCPU_SET(curpmap, NULL);
929 }
930 
931 void
932 moea_change_wiring(mmu_t mmu, pmap_t pm, vm_offset_t va, boolean_t wired)
933 {
934 	struct	pvo_entry *pvo;
935 
936 	PMAP_LOCK(pm);
937 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
938 
939 	if (pvo != NULL) {
940 		if (wired) {
941 			if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
942 				pm->pm_stats.wired_count++;
943 			pvo->pvo_vaddr |= PVO_WIRED;
944 		} else {
945 			if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
946 				pm->pm_stats.wired_count--;
947 			pvo->pvo_vaddr &= ~PVO_WIRED;
948 		}
949 	}
950 	PMAP_UNLOCK(pm);
951 }
952 
953 void
954 moea_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst)
955 {
956 	vm_offset_t	dst;
957 	vm_offset_t	src;
958 
959 	dst = VM_PAGE_TO_PHYS(mdst);
960 	src = VM_PAGE_TO_PHYS(msrc);
961 
962 	kcopy((void *)src, (void *)dst, PAGE_SIZE);
963 }
964 
965 /*
966  * Zero a page of physical memory by temporarily mapping it into the tlb.
967  */
968 void
969 moea_zero_page(mmu_t mmu, vm_page_t m)
970 {
971 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
972 	caddr_t va;
973 
974 	if (pa < SEGMENT_LENGTH) {
975 		va = (caddr_t) pa;
976 	} else if (moea_initialized) {
977 		if (moea_pvo_zeropage == NULL)
978 			moea_pvo_zeropage = moea_rkva_alloc(mmu);
979 		moea_pa_map(moea_pvo_zeropage, pa, NULL, NULL);
980 		va = (caddr_t)PVO_VADDR(moea_pvo_zeropage);
981 	} else {
982 		panic("moea_zero_page: can't zero pa %#x", pa);
983 	}
984 
985 	bzero(va, PAGE_SIZE);
986 
987 	if (pa >= SEGMENT_LENGTH)
988 		moea_pa_unmap(moea_pvo_zeropage, NULL, NULL);
989 }
990 
991 void
992 moea_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size)
993 {
994 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
995 	caddr_t va;
996 
997 	if (pa < SEGMENT_LENGTH) {
998 		va = (caddr_t) pa;
999 	} else if (moea_initialized) {
1000 		if (moea_pvo_zeropage == NULL)
1001 			moea_pvo_zeropage = moea_rkva_alloc(mmu);
1002 		moea_pa_map(moea_pvo_zeropage, pa, NULL, NULL);
1003 		va = (caddr_t)PVO_VADDR(moea_pvo_zeropage);
1004 	} else {
1005 		panic("moea_zero_page: can't zero pa %#x", pa);
1006 	}
1007 
1008 	bzero(va + off, size);
1009 
1010 	if (pa >= SEGMENT_LENGTH)
1011 		moea_pa_unmap(moea_pvo_zeropage, NULL, NULL);
1012 }
1013 
1014 void
1015 moea_zero_page_idle(mmu_t mmu, vm_page_t m)
1016 {
1017 
1018 	/* XXX this is called outside of Giant, is moea_zero_page safe? */
1019 	/* XXX maybe have a dedicated mapping for this to avoid the problem? */
1020 	mtx_lock(&Giant);
1021 	moea_zero_page(mmu, m);
1022 	mtx_unlock(&Giant);
1023 }
1024 
1025 /*
1026  * Map the given physical page at the specified virtual address in the
1027  * target pmap with the protection requested.  If specified the page
1028  * will be wired down.
1029  */
1030 void
1031 moea_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1032 	   boolean_t wired)
1033 {
1034 	struct		pvo_head *pvo_head;
1035 	uma_zone_t	zone;
1036 	vm_page_t	pg;
1037 	u_int		pte_lo, pvo_flags, was_exec, i;
1038 	int		error;
1039 
1040 	if (!moea_initialized) {
1041 		pvo_head = &moea_pvo_kunmanaged;
1042 		zone = moea_upvo_zone;
1043 		pvo_flags = 0;
1044 		pg = NULL;
1045 		was_exec = PTE_EXEC;
1046 	} else {
1047 		pvo_head = vm_page_to_pvoh(m);
1048 		pg = m;
1049 		zone = moea_mpvo_zone;
1050 		pvo_flags = PVO_MANAGED;
1051 		was_exec = 0;
1052 	}
1053 	if (pmap_bootstrapped)
1054 		vm_page_lock_queues();
1055 	PMAP_LOCK(pmap);
1056 
1057 	/* XXX change the pvo head for fake pages */
1058 	if ((m->flags & PG_FICTITIOUS) == PG_FICTITIOUS)
1059 		pvo_head = &moea_pvo_kunmanaged;
1060 
1061 	/*
1062 	 * If this is a managed page, and it's the first reference to the page,
1063 	 * clear the execness of the page.  Otherwise fetch the execness.
1064 	 */
1065 	if ((pg != NULL) && ((m->flags & PG_FICTITIOUS) == 0)) {
1066 		if (LIST_EMPTY(pvo_head)) {
1067 			moea_attr_clear(pg, PTE_EXEC);
1068 		} else {
1069 			was_exec = moea_attr_fetch(pg) & PTE_EXEC;
1070 		}
1071 	}
1072 
1073 	/*
1074 	 * Assume the page is cache inhibited and access is guarded unless
1075 	 * it's in our available memory array.
1076 	 */
1077 	pte_lo = PTE_I | PTE_G;
1078 	for (i = 0; i < pregions_sz; i++) {
1079 		if ((VM_PAGE_TO_PHYS(m) >= pregions[i].mr_start) &&
1080 		    (VM_PAGE_TO_PHYS(m) <
1081 			(pregions[i].mr_start + pregions[i].mr_size))) {
1082 			pte_lo &= ~(PTE_I | PTE_G);
1083 			break;
1084 		}
1085 	}
1086 
1087 	if (prot & VM_PROT_WRITE)
1088 		pte_lo |= PTE_BW;
1089 	else
1090 		pte_lo |= PTE_BR;
1091 
1092 	if (prot & VM_PROT_EXECUTE)
1093 		pvo_flags |= PVO_EXECUTABLE;
1094 
1095 	if (wired)
1096 		pvo_flags |= PVO_WIRED;
1097 
1098 	if ((m->flags & PG_FICTITIOUS) != 0)
1099 		pvo_flags |= PVO_FAKE;
1100 
1101 	error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1102 	    pte_lo, pvo_flags);
1103 
1104 	/*
1105 	 * Flush the real page from the instruction cache if this page is
1106 	 * mapped executable and cacheable and was not previously mapped (or
1107 	 * was not mapped executable).
1108 	 */
1109 	if (error == 0 && (pvo_flags & PVO_EXECUTABLE) &&
1110 	    (pte_lo & PTE_I) == 0 && was_exec == 0) {
1111 		/*
1112 		 * Flush the real memory from the cache.
1113 		 */
1114 		moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1115 		if (pg != NULL)
1116 			moea_attr_save(pg, PTE_EXEC);
1117 	}
1118 	if (pmap_bootstrapped)
1119 		vm_page_unlock_queues();
1120 
1121 	/* XXX syncicache always until problems are sorted */
1122 	moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1123 	PMAP_UNLOCK(pmap);
1124 }
1125 
1126 vm_page_t
1127 moea_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m,
1128     vm_prot_t prot, vm_page_t mpte)
1129 {
1130 
1131 	vm_page_busy(m);
1132 	vm_page_unlock_queues();
1133 	VM_OBJECT_UNLOCK(m->object);
1134 	mtx_lock(&Giant);
1135 	moea_enter(mmu, pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1136 	    FALSE);
1137 	mtx_unlock(&Giant);
1138 	VM_OBJECT_LOCK(m->object);
1139 	vm_page_lock_queues();
1140 	vm_page_wakeup(m);
1141 	return (NULL);
1142 }
1143 
1144 vm_paddr_t
1145 moea_extract(mmu_t mmu, pmap_t pm, vm_offset_t va)
1146 {
1147 	struct	pvo_entry *pvo;
1148 	vm_paddr_t pa;
1149 
1150 	PMAP_LOCK(pm);
1151 	pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1152 	if (pvo == NULL)
1153 		pa = 0;
1154 	else
1155 		pa = (pvo->pvo_pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1156 	PMAP_UNLOCK(pm);
1157 	return (pa);
1158 }
1159 
1160 /*
1161  * Atomically extract and hold the physical page with the given
1162  * pmap and virtual address pair if that mapping permits the given
1163  * protection.
1164  */
1165 vm_page_t
1166 moea_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1167 {
1168 	struct	pvo_entry *pvo;
1169 	vm_page_t m;
1170 
1171 	m = NULL;
1172 	mtx_lock(&Giant);
1173 	vm_page_lock_queues();
1174 	PMAP_LOCK(pmap);
1175 	pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1176 	if (pvo != NULL && (pvo->pvo_pte.pte_hi & PTE_VALID) &&
1177 	    ((pvo->pvo_pte.pte_lo & PTE_PP) == PTE_RW ||
1178 	     (prot & VM_PROT_WRITE) == 0)) {
1179 		m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte_lo & PTE_RPGN);
1180 		vm_page_hold(m);
1181 	}
1182 	vm_page_unlock_queues();
1183 	PMAP_UNLOCK(pmap);
1184 	mtx_unlock(&Giant);
1185 	return (m);
1186 }
1187 
1188 void
1189 moea_init(mmu_t mmu)
1190 {
1191 
1192 	CTR0(KTR_PMAP, "moea_init");
1193 
1194 	moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1195 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1196 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1197 	moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1198 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1199 	    UMA_ZONE_VM | UMA_ZONE_NOFREE);
1200 	moea_initialized = TRUE;
1201 }
1202 
1203 boolean_t
1204 moea_is_modified(mmu_t mmu, vm_page_t m)
1205 {
1206 
1207 	if ((m->flags & (PG_FICTITIOUS |PG_UNMANAGED)) != 0)
1208 		return (FALSE);
1209 
1210 	return (moea_query_bit(m, PTE_CHG));
1211 }
1212 
1213 void
1214 moea_clear_reference(mmu_t mmu, vm_page_t m)
1215 {
1216 
1217 	if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0)
1218 		return;
1219 	moea_clear_bit(m, PTE_REF, NULL);
1220 }
1221 
1222 void
1223 moea_clear_modify(mmu_t mmu, vm_page_t m)
1224 {
1225 
1226 	if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0)
1227 		return;
1228 	moea_clear_bit(m, PTE_CHG, NULL);
1229 }
1230 
1231 /*
1232  *	moea_ts_referenced:
1233  *
1234  *	Return a count of reference bits for a page, clearing those bits.
1235  *	It is not necessary for every reference bit to be cleared, but it
1236  *	is necessary that 0 only be returned when there are truly no
1237  *	reference bits set.
1238  *
1239  *	XXX: The exact number of bits to check and clear is a matter that
1240  *	should be tested and standardized at some point in the future for
1241  *	optimal aging of shared pages.
1242  */
1243 boolean_t
1244 moea_ts_referenced(mmu_t mmu, vm_page_t m)
1245 {
1246 	int count;
1247 
1248 	if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0)
1249 		return (0);
1250 
1251 	count = moea_clear_bit(m, PTE_REF, NULL);
1252 
1253 	return (count);
1254 }
1255 
1256 /*
1257  * Map a wired page into kernel virtual address space.
1258  */
1259 void
1260 moea_kenter(mmu_t mmu, vm_offset_t va, vm_offset_t pa)
1261 {
1262 	u_int		pte_lo;
1263 	int		error;
1264 	int		i;
1265 
1266 #if 0
1267 	if (va < VM_MIN_KERNEL_ADDRESS)
1268 		panic("moea_kenter: attempt to enter non-kernel address %#x",
1269 		    va);
1270 #endif
1271 
1272 	pte_lo = PTE_I | PTE_G;
1273 	for (i = 0; i < pregions_sz; i++) {
1274 		if ((pa >= pregions[i].mr_start) &&
1275 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
1276 			pte_lo &= ~(PTE_I | PTE_G);
1277 			break;
1278 		}
1279 	}
1280 
1281 	PMAP_LOCK(kernel_pmap);
1282 	error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1283 	    &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1284 
1285 	if (error != 0 && error != ENOENT)
1286 		panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1287 		    pa, error);
1288 
1289 	/*
1290 	 * Flush the real memory from the instruction cache.
1291 	 */
1292 	if ((pte_lo & (PTE_I | PTE_G)) == 0) {
1293 		moea_syncicache(pa, PAGE_SIZE);
1294 	}
1295 	PMAP_UNLOCK(kernel_pmap);
1296 }
1297 
1298 /*
1299  * Extract the physical page address associated with the given kernel virtual
1300  * address.
1301  */
1302 vm_offset_t
1303 moea_kextract(mmu_t mmu, vm_offset_t va)
1304 {
1305 	struct		pvo_entry *pvo;
1306 	vm_paddr_t pa;
1307 
1308 #ifdef UMA_MD_SMALL_ALLOC
1309 	/*
1310 	 * Allow direct mappings
1311 	 */
1312 	if (va < VM_MIN_KERNEL_ADDRESS) {
1313 		return (va);
1314 	}
1315 #endif
1316 
1317 	PMAP_LOCK(kernel_pmap);
1318 	pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1319 	KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1320 	pa = (pvo->pvo_pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF);
1321 	PMAP_UNLOCK(kernel_pmap);
1322 	return (pa);
1323 }
1324 
1325 /*
1326  * Remove a wired page from kernel virtual address space.
1327  */
1328 void
1329 moea_kremove(mmu_t mmu, vm_offset_t va)
1330 {
1331 
1332 	moea_remove(mmu, kernel_pmap, va, va + PAGE_SIZE);
1333 }
1334 
1335 /*
1336  * Map a range of physical addresses into kernel virtual address space.
1337  *
1338  * The value passed in *virt is a suggested virtual address for the mapping.
1339  * Architectures which can support a direct-mapped physical to virtual region
1340  * can return the appropriate address within that region, leaving '*virt'
1341  * unchanged.  We cannot and therefore do not; *virt is updated with the
1342  * first usable address after the mapped region.
1343  */
1344 vm_offset_t
1345 moea_map(mmu_t mmu, vm_offset_t *virt, vm_offset_t pa_start,
1346     vm_offset_t pa_end, int prot)
1347 {
1348 	vm_offset_t	sva, va;
1349 
1350 	sva = *virt;
1351 	va = sva;
1352 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1353 		moea_kenter(mmu, va, pa_start);
1354 	*virt = va;
1355 	return (sva);
1356 }
1357 
1358 /*
1359  * Lower the permission for all mappings to a given page.
1360  */
1361 void
1362 moea_page_protect(mmu_t mmu, vm_page_t m, vm_prot_t prot)
1363 {
1364 	struct	pvo_head *pvo_head;
1365 	struct	pvo_entry *pvo, *next_pvo;
1366 	struct	pte *pt;
1367 	pmap_t	pmap;
1368 
1369 	/*
1370 	 * Since the routine only downgrades protection, if the
1371 	 * maximal protection is desired, there isn't any change
1372 	 * to be made.
1373 	 */
1374 	if ((prot & (VM_PROT_READ|VM_PROT_WRITE)) ==
1375 	    (VM_PROT_READ|VM_PROT_WRITE))
1376 		return;
1377 
1378 	pvo_head = vm_page_to_pvoh(m);
1379 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1380 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1381 		MOEA_PVO_CHECK(pvo);	/* sanity check */
1382 		pmap = pvo->pvo_pmap;
1383 		PMAP_LOCK(pmap);
1384 
1385 		/*
1386 		 * Downgrading to no mapping at all, we just remove the entry.
1387 		 */
1388 		if ((prot & VM_PROT_READ) == 0) {
1389 			moea_pvo_remove(pvo, -1);
1390 			PMAP_UNLOCK(pmap);
1391 			continue;
1392 		}
1393 
1394 		/*
1395 		 * If EXEC permission is being revoked, just clear the flag
1396 		 * in the PVO.
1397 		 */
1398 		if ((prot & VM_PROT_EXECUTE) == 0)
1399 			pvo->pvo_vaddr &= ~PVO_EXECUTABLE;
1400 
1401 		/*
1402 		 * If this entry is already RO, don't diddle with the page
1403 		 * table.
1404 		 */
1405 		if ((pvo->pvo_pte.pte_lo & PTE_PP) == PTE_BR) {
1406 			PMAP_UNLOCK(pmap);
1407 			MOEA_PVO_CHECK(pvo);
1408 			continue;
1409 		}
1410 
1411 		/*
1412 		 * Grab the PTE before we diddle the bits so pvo_to_pte can
1413 		 * verify the pte contents are as expected.
1414 		 */
1415 		pt = moea_pvo_to_pte(pvo, -1);
1416 		pvo->pvo_pte.pte_lo &= ~PTE_PP;
1417 		pvo->pvo_pte.pte_lo |= PTE_BR;
1418 		if (pt != NULL)
1419 			moea_pte_change(pt, &pvo->pvo_pte, pvo->pvo_vaddr);
1420 		PMAP_UNLOCK(pmap);
1421 		MOEA_PVO_CHECK(pvo);	/* sanity check */
1422 	}
1423 
1424 	/*
1425 	 * Downgrading from writeable: clear the VM page flag
1426 	 */
1427 	if ((prot & VM_PROT_WRITE) != VM_PROT_WRITE)
1428 		vm_page_flag_clear(m, PG_WRITEABLE);
1429 }
1430 
1431 /*
1432  * Returns true if the pmap's pv is one of the first
1433  * 16 pvs linked to from this page.  This count may
1434  * be changed upwards or downwards in the future; it
1435  * is only necessary that true be returned for a small
1436  * subset of pmaps for proper page aging.
1437  */
1438 boolean_t
1439 moea_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m)
1440 {
1441         int loops;
1442 	struct pvo_entry *pvo;
1443 
1444         if (!moea_initialized || (m->flags & PG_FICTITIOUS))
1445                 return FALSE;
1446 
1447 	loops = 0;
1448 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1449 		if (pvo->pvo_pmap == pmap)
1450 			return (TRUE);
1451 		if (++loops >= 16)
1452 			break;
1453 	}
1454 
1455 	return (FALSE);
1456 }
1457 
1458 static u_int	moea_vsidcontext;
1459 
1460 void
1461 moea_pinit(mmu_t mmu, pmap_t pmap)
1462 {
1463 	int	i, mask;
1464 	u_int	entropy;
1465 
1466 	KASSERT((int)pmap < VM_MIN_KERNEL_ADDRESS, ("moea_pinit: virt pmap"));
1467 	PMAP_LOCK_INIT(pmap);
1468 
1469 	entropy = 0;
1470 	__asm __volatile("mftb %0" : "=r"(entropy));
1471 
1472 	/*
1473 	 * Allocate some segment registers for this pmap.
1474 	 */
1475 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1476 		u_int	hash, n;
1477 
1478 		/*
1479 		 * Create a new value by mutiplying by a prime and adding in
1480 		 * entropy from the timebase register.  This is to make the
1481 		 * VSID more random so that the PT hash function collides
1482 		 * less often.  (Note that the prime casues gcc to do shifts
1483 		 * instead of a multiply.)
1484 		 */
1485 		moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1486 		hash = moea_vsidcontext & (NPMAPS - 1);
1487 		if (hash == 0)		/* 0 is special, avoid it */
1488 			continue;
1489 		n = hash >> 5;
1490 		mask = 1 << (hash & (VSID_NBPW - 1));
1491 		hash = (moea_vsidcontext & 0xfffff);
1492 		if (moea_vsid_bitmap[n] & mask) {	/* collision? */
1493 			/* anything free in this bucket? */
1494 			if (moea_vsid_bitmap[n] == 0xffffffff) {
1495 				entropy = (moea_vsidcontext >> 20);
1496 				continue;
1497 			}
1498 			i = ffs(~moea_vsid_bitmap[i]) - 1;
1499 			mask = 1 << i;
1500 			hash &= 0xfffff & ~(VSID_NBPW - 1);
1501 			hash |= i;
1502 		}
1503 		moea_vsid_bitmap[n] |= mask;
1504 		for (i = 0; i < 16; i++)
1505 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1506 		return;
1507 	}
1508 
1509 	panic("moea_pinit: out of segments");
1510 }
1511 
1512 /*
1513  * Initialize the pmap associated with process 0.
1514  */
1515 void
1516 moea_pinit0(mmu_t mmu, pmap_t pm)
1517 {
1518 
1519 	moea_pinit(mmu, pm);
1520 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1521 }
1522 
1523 /*
1524  * Set the physical protection on the specified range of this map as requested.
1525  */
1526 void
1527 moea_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1528     vm_prot_t prot)
1529 {
1530 	struct	pvo_entry *pvo;
1531 	struct	pte *pt;
1532 	int	pteidx;
1533 
1534 	CTR4(KTR_PMAP, "moea_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm, sva,
1535 	    eva, prot);
1536 
1537 
1538 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1539 	    ("moea_protect: non current pmap"));
1540 
1541 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1542 		mtx_lock(&Giant);
1543 		moea_remove(mmu, pm, sva, eva);
1544 		mtx_unlock(&Giant);
1545 		return;
1546 	}
1547 
1548 	mtx_lock(&Giant);
1549 	vm_page_lock_queues();
1550 	PMAP_LOCK(pm);
1551 	for (; sva < eva; sva += PAGE_SIZE) {
1552 		pvo = moea_pvo_find_va(pm, sva, &pteidx);
1553 		if (pvo == NULL)
1554 			continue;
1555 
1556 		if ((prot & VM_PROT_EXECUTE) == 0)
1557 			pvo->pvo_vaddr &= ~PVO_EXECUTABLE;
1558 
1559 		/*
1560 		 * Grab the PTE pointer before we diddle with the cached PTE
1561 		 * copy.
1562 		 */
1563 		pt = moea_pvo_to_pte(pvo, pteidx);
1564 		/*
1565 		 * Change the protection of the page.
1566 		 */
1567 		pvo->pvo_pte.pte_lo &= ~PTE_PP;
1568 		pvo->pvo_pte.pte_lo |= PTE_BR;
1569 
1570 		/*
1571 		 * If the PVO is in the page table, update that pte as well.
1572 		 */
1573 		if (pt != NULL)
1574 			moea_pte_change(pt, &pvo->pvo_pte, pvo->pvo_vaddr);
1575 	}
1576 	vm_page_unlock_queues();
1577 	PMAP_UNLOCK(pm);
1578 	mtx_unlock(&Giant);
1579 }
1580 
1581 /*
1582  * Map a list of wired pages into kernel virtual address space.  This is
1583  * intended for temporary mappings which do not need page modification or
1584  * references recorded.  Existing mappings in the region are overwritten.
1585  */
1586 void
1587 moea_qenter(mmu_t mmu, vm_offset_t sva, vm_page_t *m, int count)
1588 {
1589 	vm_offset_t va;
1590 
1591 	va = sva;
1592 	while (count-- > 0) {
1593 		moea_kenter(mmu, va, VM_PAGE_TO_PHYS(*m));
1594 		va += PAGE_SIZE;
1595 		m++;
1596 	}
1597 }
1598 
1599 /*
1600  * Remove page mappings from kernel virtual address space.  Intended for
1601  * temporary mappings entered by moea_qenter.
1602  */
1603 void
1604 moea_qremove(mmu_t mmu, vm_offset_t sva, int count)
1605 {
1606 	vm_offset_t va;
1607 
1608 	va = sva;
1609 	while (count-- > 0) {
1610 		moea_kremove(mmu, va);
1611 		va += PAGE_SIZE;
1612 	}
1613 }
1614 
1615 void
1616 moea_release(mmu_t mmu, pmap_t pmap)
1617 {
1618         int idx, mask;
1619 
1620 	/*
1621 	 * Free segment register's VSID
1622 	 */
1623         if (pmap->pm_sr[0] == 0)
1624                 panic("moea_release");
1625 
1626         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1627         mask = 1 << (idx % VSID_NBPW);
1628         idx /= VSID_NBPW;
1629         moea_vsid_bitmap[idx] &= ~mask;
1630 	PMAP_LOCK_DESTROY(pmap);
1631 }
1632 
1633 /*
1634  * Remove the given range of addresses from the specified map.
1635  */
1636 void
1637 moea_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1638 {
1639 	struct	pvo_entry *pvo;
1640 	int	pteidx;
1641 
1642 	vm_page_lock_queues();
1643 	PMAP_LOCK(pm);
1644 	for (; sva < eva; sva += PAGE_SIZE) {
1645 		pvo = moea_pvo_find_va(pm, sva, &pteidx);
1646 		if (pvo != NULL) {
1647 			moea_pvo_remove(pvo, pteidx);
1648 		}
1649 	}
1650 	PMAP_UNLOCK(pm);
1651 	vm_page_unlock_queues();
1652 }
1653 
1654 /*
1655  * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1656  * will reflect changes in pte's back to the vm_page.
1657  */
1658 void
1659 moea_remove_all(mmu_t mmu, vm_page_t m)
1660 {
1661 	struct  pvo_head *pvo_head;
1662 	struct	pvo_entry *pvo, *next_pvo;
1663 	pmap_t	pmap;
1664 
1665 	mtx_assert(&vm_page_queue_mtx, MA_OWNED);
1666 
1667 	pvo_head = vm_page_to_pvoh(m);
1668 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1669 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1670 
1671 		MOEA_PVO_CHECK(pvo);	/* sanity check */
1672 		pmap = pvo->pvo_pmap;
1673 		PMAP_LOCK(pmap);
1674 		moea_pvo_remove(pvo, -1);
1675 		PMAP_UNLOCK(pmap);
1676 	}
1677 	vm_page_flag_clear(m, PG_WRITEABLE);
1678 }
1679 
1680 /*
1681  * Allocate a physical page of memory directly from the phys_avail map.
1682  * Can only be called from moea_bootstrap before avail start and end are
1683  * calculated.
1684  */
1685 static vm_offset_t
1686 moea_bootstrap_alloc(vm_size_t size, u_int align)
1687 {
1688 	vm_offset_t	s, e;
1689 	int		i, j;
1690 
1691 	size = round_page(size);
1692 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1693 		if (align != 0)
1694 			s = (phys_avail[i] + align - 1) & ~(align - 1);
1695 		else
1696 			s = phys_avail[i];
1697 		e = s + size;
1698 
1699 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1700 			continue;
1701 
1702 		if (s == phys_avail[i]) {
1703 			phys_avail[i] += size;
1704 		} else if (e == phys_avail[i + 1]) {
1705 			phys_avail[i + 1] -= size;
1706 		} else {
1707 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1708 				phys_avail[j] = phys_avail[j - 2];
1709 				phys_avail[j + 1] = phys_avail[j - 1];
1710 			}
1711 
1712 			phys_avail[i + 3] = phys_avail[i + 1];
1713 			phys_avail[i + 1] = s;
1714 			phys_avail[i + 2] = e;
1715 			phys_avail_count++;
1716 		}
1717 
1718 		return (s);
1719 	}
1720 	panic("moea_bootstrap_alloc: could not allocate memory");
1721 }
1722 
1723 /*
1724  * Return an unmapped pvo for a kernel virtual address.
1725  * Used by pmap functions that operate on physical pages.
1726  */
1727 static struct pvo_entry *
1728 moea_rkva_alloc(mmu_t mmu)
1729 {
1730 	struct		pvo_entry *pvo;
1731 	struct		pte *pt;
1732 	vm_offset_t	kva;
1733 	int		pteidx;
1734 
1735 	if (moea_rkva_count == 0)
1736 		panic("moea_rkva_alloc: no more reserved KVAs");
1737 
1738 	kva = moea_rkva_start + (PAGE_SIZE * --moea_rkva_count);
1739 	moea_kenter(mmu, kva, 0);
1740 
1741 	pvo = moea_pvo_find_va(kernel_pmap, kva, &pteidx);
1742 
1743 	if (pvo == NULL)
1744 		panic("moea_kva_alloc: moea_pvo_find_va failed");
1745 
1746 	pt = moea_pvo_to_pte(pvo, pteidx);
1747 
1748 	if (pt == NULL)
1749 		panic("moea_kva_alloc: moea_pvo_to_pte failed");
1750 
1751 	moea_pte_unset(pt, &pvo->pvo_pte, pvo->pvo_vaddr);
1752 	PVO_PTEGIDX_CLR(pvo);
1753 
1754 	moea_pte_overflow++;
1755 
1756 	return (pvo);
1757 }
1758 
1759 static void
1760 moea_pa_map(struct pvo_entry *pvo, vm_offset_t pa, struct pte *saved_pt,
1761     int *depth_p)
1762 {
1763 	struct	pte *pt;
1764 
1765 	/*
1766 	 * If this pvo already has a valid pte, we need to save it so it can
1767 	 * be restored later.  We then just reload the new PTE over the old
1768 	 * slot.
1769 	 */
1770 	if (saved_pt != NULL) {
1771 		pt = moea_pvo_to_pte(pvo, -1);
1772 
1773 		if (pt != NULL) {
1774 			moea_pte_unset(pt, &pvo->pvo_pte, pvo->pvo_vaddr);
1775 			PVO_PTEGIDX_CLR(pvo);
1776 			moea_pte_overflow++;
1777 		}
1778 
1779 		*saved_pt = pvo->pvo_pte;
1780 
1781 		pvo->pvo_pte.pte_lo &= ~PTE_RPGN;
1782 	}
1783 
1784 	pvo->pvo_pte.pte_lo |= pa;
1785 
1786 	if (!moea_pte_spill(pvo->pvo_vaddr))
1787 		panic("moea_pa_map: could not spill pvo %p", pvo);
1788 
1789 	if (depth_p != NULL)
1790 		(*depth_p)++;
1791 }
1792 
1793 static void
1794 moea_pa_unmap(struct pvo_entry *pvo, struct pte *saved_pt, int *depth_p)
1795 {
1796 	struct	pte *pt;
1797 
1798 	pt = moea_pvo_to_pte(pvo, -1);
1799 
1800 	if (pt != NULL) {
1801 		moea_pte_unset(pt, &pvo->pvo_pte, pvo->pvo_vaddr);
1802 		PVO_PTEGIDX_CLR(pvo);
1803 		moea_pte_overflow++;
1804 	}
1805 
1806 	pvo->pvo_pte.pte_lo &= ~PTE_RPGN;
1807 
1808 	/*
1809 	 * If there is a saved PTE and it's valid, restore it and return.
1810 	 */
1811 	if (saved_pt != NULL && (saved_pt->pte_lo & PTE_RPGN) != 0) {
1812 		if (depth_p != NULL && --(*depth_p) == 0)
1813 			panic("moea_pa_unmap: restoring but depth == 0");
1814 
1815 		pvo->pvo_pte = *saved_pt;
1816 
1817 		if (!moea_pte_spill(pvo->pvo_vaddr))
1818 			panic("moea_pa_unmap: could not spill pvo %p", pvo);
1819 	}
1820 }
1821 
1822 static void
1823 moea_syncicache(vm_offset_t pa, vm_size_t len)
1824 {
1825 	__syncicache((void *)pa, len);
1826 }
1827 
1828 static void
1829 tlbia(void)
1830 {
1831 	caddr_t	i;
1832 
1833 	SYNC();
1834 	for (i = 0; i < (caddr_t)0x00040000; i += 0x00001000) {
1835 		TLBIE(i);
1836 		EIEIO();
1837 	}
1838 	TLBSYNC();
1839 	SYNC();
1840 }
1841 
1842 static int
1843 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1844     vm_offset_t va, vm_offset_t pa, u_int pte_lo, int flags)
1845 {
1846 	struct	pvo_entry *pvo;
1847 	u_int	sr;
1848 	int	first;
1849 	u_int	ptegidx;
1850 	int	i;
1851 	int     bootstrap;
1852 
1853 	moea_pvo_enter_calls++;
1854 	first = 0;
1855 	bootstrap = 0;
1856 
1857 	/*
1858 	 * Compute the PTE Group index.
1859 	 */
1860 	va &= ~ADDR_POFF;
1861 	sr = va_to_sr(pm->pm_sr, va);
1862 	ptegidx = va_to_pteg(sr, va);
1863 
1864 	/*
1865 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
1866 	 * there is a mapping.
1867 	 */
1868 	mtx_lock(&moea_table_mutex);
1869 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
1870 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
1871 			if ((pvo->pvo_pte.pte_lo & PTE_RPGN) == pa &&
1872 			    (pvo->pvo_pte.pte_lo & PTE_PP) ==
1873 			    (pte_lo & PTE_PP)) {
1874 				mtx_unlock(&moea_table_mutex);
1875 				return (0);
1876 			}
1877 			moea_pvo_remove(pvo, -1);
1878 			break;
1879 		}
1880 	}
1881 
1882 	/*
1883 	 * If we aren't overwriting a mapping, try to allocate.
1884 	 */
1885 	if (moea_initialized) {
1886 		pvo = uma_zalloc(zone, M_NOWAIT);
1887 	} else {
1888 		if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
1889 			panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
1890 			      moea_bpvo_pool_index, BPVO_POOL_SIZE,
1891 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
1892 		}
1893 		pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
1894 		moea_bpvo_pool_index++;
1895 		bootstrap = 1;
1896 	}
1897 
1898 	if (pvo == NULL) {
1899 		mtx_unlock(&moea_table_mutex);
1900 		return (ENOMEM);
1901 	}
1902 
1903 	moea_pvo_entries++;
1904 	pvo->pvo_vaddr = va;
1905 	pvo->pvo_pmap = pm;
1906 	LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
1907 	pvo->pvo_vaddr &= ~ADDR_POFF;
1908 	if (flags & VM_PROT_EXECUTE)
1909 		pvo->pvo_vaddr |= PVO_EXECUTABLE;
1910 	if (flags & PVO_WIRED)
1911 		pvo->pvo_vaddr |= PVO_WIRED;
1912 	if (pvo_head != &moea_pvo_kunmanaged)
1913 		pvo->pvo_vaddr |= PVO_MANAGED;
1914 	if (bootstrap)
1915 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
1916 	if (flags & PVO_FAKE)
1917 		pvo->pvo_vaddr |= PVO_FAKE;
1918 
1919 	moea_pte_create(&pvo->pvo_pte, sr, va, pa | pte_lo);
1920 
1921 	/*
1922 	 * Remember if the list was empty and therefore will be the first
1923 	 * item.
1924 	 */
1925 	if (LIST_FIRST(pvo_head) == NULL)
1926 		first = 1;
1927 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
1928 
1929 	if (pvo->pvo_pte.pte_lo & PVO_WIRED)
1930 		pm->pm_stats.wired_count++;
1931 	pm->pm_stats.resident_count++;
1932 
1933 	/*
1934 	 * We hope this succeeds but it isn't required.
1935 	 */
1936 	i = moea_pte_insert(ptegidx, &pvo->pvo_pte);
1937 	if (i >= 0) {
1938 		PVO_PTEGIDX_SET(pvo, i);
1939 	} else {
1940 		panic("moea_pvo_enter: overflow");
1941 		moea_pte_overflow++;
1942 	}
1943 	mtx_unlock(&moea_table_mutex);
1944 
1945 	return (first ? ENOENT : 0);
1946 }
1947 
1948 static void
1949 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
1950 {
1951 	struct	pte *pt;
1952 
1953 	/*
1954 	 * If there is an active pte entry, we need to deactivate it (and
1955 	 * save the ref & cfg bits).
1956 	 */
1957 	pt = moea_pvo_to_pte(pvo, pteidx);
1958 	if (pt != NULL) {
1959 		moea_pte_unset(pt, &pvo->pvo_pte, pvo->pvo_vaddr);
1960 		PVO_PTEGIDX_CLR(pvo);
1961 	} else {
1962 		moea_pte_overflow--;
1963 	}
1964 
1965 	/*
1966 	 * Update our statistics.
1967 	 */
1968 	pvo->pvo_pmap->pm_stats.resident_count--;
1969 	if (pvo->pvo_pte.pte_lo & PVO_WIRED)
1970 		pvo->pvo_pmap->pm_stats.wired_count--;
1971 
1972 	/*
1973 	 * Save the REF/CHG bits into their cache if the page is managed.
1974 	 */
1975 	if ((pvo->pvo_vaddr & (PVO_MANAGED|PVO_FAKE)) == PVO_MANAGED) {
1976 		struct	vm_page *pg;
1977 
1978 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte_lo & PTE_RPGN);
1979 		if (pg != NULL) {
1980 			moea_attr_save(pg, pvo->pvo_pte.pte_lo &
1981 			    (PTE_REF | PTE_CHG));
1982 		}
1983 	}
1984 
1985 	/*
1986 	 * Remove this PVO from the PV list.
1987 	 */
1988 	LIST_REMOVE(pvo, pvo_vlink);
1989 
1990 	/*
1991 	 * Remove this from the overflow list and return it to the pool
1992 	 * if we aren't going to reuse it.
1993 	 */
1994 	LIST_REMOVE(pvo, pvo_olink);
1995 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
1996 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
1997 		    moea_upvo_zone, pvo);
1998 	moea_pvo_entries--;
1999 	moea_pvo_remove_calls++;
2000 }
2001 
2002 static __inline int
2003 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2004 {
2005 	int	pteidx;
2006 
2007 	/*
2008 	 * We can find the actual pte entry without searching by grabbing
2009 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2010 	 * noticing the HID bit.
2011 	 */
2012 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2013 	if (pvo->pvo_pte.pte_hi & PTE_HID)
2014 		pteidx ^= moea_pteg_mask * 8;
2015 
2016 	return (pteidx);
2017 }
2018 
2019 static struct pvo_entry *
2020 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2021 {
2022 	struct	pvo_entry *pvo;
2023 	int	ptegidx;
2024 	u_int	sr;
2025 
2026 	va &= ~ADDR_POFF;
2027 	sr = va_to_sr(pm->pm_sr, va);
2028 	ptegidx = va_to_pteg(sr, va);
2029 
2030 	mtx_lock(&moea_table_mutex);
2031 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2032 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2033 			if (pteidx_p)
2034 				*pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2035 			break;
2036 		}
2037 	}
2038 	mtx_unlock(&moea_table_mutex);
2039 
2040 	return (pvo);
2041 }
2042 
2043 static struct pte *
2044 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2045 {
2046 	struct	pte *pt;
2047 
2048 	/*
2049 	 * If we haven't been supplied the ptegidx, calculate it.
2050 	 */
2051 	if (pteidx == -1) {
2052 		int	ptegidx;
2053 		u_int	sr;
2054 
2055 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2056 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2057 		pteidx = moea_pvo_pte_index(pvo, ptegidx);
2058 	}
2059 
2060 	pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2061 
2062 	if ((pvo->pvo_pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2063 		panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2064 		    "valid pte index", pvo);
2065 	}
2066 
2067 	if ((pvo->pvo_pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2068 		panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2069 		    "pvo but no valid pte", pvo);
2070 	}
2071 
2072 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2073 		if ((pvo->pvo_pte.pte_hi & PTE_VALID) == 0) {
2074 			panic("moea_pvo_to_pte: pvo %p has valid pte in "
2075 			    "moea_pteg_table %p but invalid in pvo", pvo, pt);
2076 		}
2077 
2078 		if (((pt->pte_lo ^ pvo->pvo_pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2079 		    != 0) {
2080 			panic("moea_pvo_to_pte: pvo %p pte does not match "
2081 			    "pte %p in moea_pteg_table", pvo, pt);
2082 		}
2083 
2084 		return (pt);
2085 	}
2086 
2087 	if (pvo->pvo_pte.pte_hi & PTE_VALID) {
2088 		panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2089 		    "moea_pteg_table but valid in pvo", pvo, pt);
2090 	}
2091 
2092 	return (NULL);
2093 }
2094 
2095 /*
2096  * XXX: THIS STUFF SHOULD BE IN pte.c?
2097  */
2098 int
2099 moea_pte_spill(vm_offset_t addr)
2100 {
2101 	struct	pvo_entry *source_pvo, *victim_pvo;
2102 	struct	pvo_entry *pvo;
2103 	int	ptegidx, i, j;
2104 	u_int	sr;
2105 	struct	pteg *pteg;
2106 	struct	pte *pt;
2107 
2108 	moea_pte_spills++;
2109 
2110 	sr = mfsrin(addr);
2111 	ptegidx = va_to_pteg(sr, addr);
2112 
2113 	/*
2114 	 * Have to substitute some entry.  Use the primary hash for this.
2115 	 * Use low bits of timebase as random generator.
2116 	 */
2117 	pteg = &moea_pteg_table[ptegidx];
2118 	mtx_lock(&moea_table_mutex);
2119 	__asm __volatile("mftb %0" : "=r"(i));
2120 	i &= 7;
2121 	pt = &pteg->pt[i];
2122 
2123 	source_pvo = NULL;
2124 	victim_pvo = NULL;
2125 	LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2126 		/*
2127 		 * We need to find a pvo entry for this address.
2128 		 */
2129 		MOEA_PVO_CHECK(pvo);
2130 		if (source_pvo == NULL &&
2131 		    moea_pte_match(&pvo->pvo_pte, sr, addr,
2132 		    pvo->pvo_pte.pte_hi & PTE_HID)) {
2133 			/*
2134 			 * Now found an entry to be spilled into the pteg.
2135 			 * The PTE is now valid, so we know it's active.
2136 			 */
2137 			j = moea_pte_insert(ptegidx, &pvo->pvo_pte);
2138 
2139 			if (j >= 0) {
2140 				PVO_PTEGIDX_SET(pvo, j);
2141 				moea_pte_overflow--;
2142 				MOEA_PVO_CHECK(pvo);
2143 				mtx_unlock(&moea_table_mutex);
2144 				return (1);
2145 			}
2146 
2147 			source_pvo = pvo;
2148 
2149 			if (victim_pvo != NULL)
2150 				break;
2151 		}
2152 
2153 		/*
2154 		 * We also need the pvo entry of the victim we are replacing
2155 		 * so save the R & C bits of the PTE.
2156 		 */
2157 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2158 		    moea_pte_compare(pt, &pvo->pvo_pte)) {
2159 			victim_pvo = pvo;
2160 			if (source_pvo != NULL)
2161 				break;
2162 		}
2163 	}
2164 
2165 	if (source_pvo == NULL) {
2166 		mtx_unlock(&moea_table_mutex);
2167 		return (0);
2168 	}
2169 
2170 	if (victim_pvo == NULL) {
2171 		if ((pt->pte_hi & PTE_HID) == 0)
2172 			panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2173 			    "entry", pt);
2174 
2175 		/*
2176 		 * If this is a secondary PTE, we need to search it's primary
2177 		 * pvo bucket for the matching PVO.
2178 		 */
2179 		LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2180 		    pvo_olink) {
2181 			MOEA_PVO_CHECK(pvo);
2182 			/*
2183 			 * We also need the pvo entry of the victim we are
2184 			 * replacing so save the R & C bits of the PTE.
2185 			 */
2186 			if (moea_pte_compare(pt, &pvo->pvo_pte)) {
2187 				victim_pvo = pvo;
2188 				break;
2189 			}
2190 		}
2191 
2192 		if (victim_pvo == NULL)
2193 			panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2194 			    "entry", pt);
2195 	}
2196 
2197 	/*
2198 	 * We are invalidating the TLB entry for the EA we are replacing even
2199 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2200 	 * contained in the TLB entry.
2201 	 */
2202 	source_pvo->pvo_pte.pte_hi &= ~PTE_HID;
2203 
2204 	moea_pte_unset(pt, &victim_pvo->pvo_pte, victim_pvo->pvo_vaddr);
2205 	moea_pte_set(pt, &source_pvo->pvo_pte);
2206 
2207 	PVO_PTEGIDX_CLR(victim_pvo);
2208 	PVO_PTEGIDX_SET(source_pvo, i);
2209 	moea_pte_replacements++;
2210 
2211 	MOEA_PVO_CHECK(victim_pvo);
2212 	MOEA_PVO_CHECK(source_pvo);
2213 
2214 	mtx_unlock(&moea_table_mutex);
2215 	return (1);
2216 }
2217 
2218 static int
2219 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2220 {
2221 	struct	pte *pt;
2222 	int	i;
2223 
2224 	/*
2225 	 * First try primary hash.
2226 	 */
2227 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2228 		if ((pt->pte_hi & PTE_VALID) == 0) {
2229 			pvo_pt->pte_hi &= ~PTE_HID;
2230 			moea_pte_set(pt, pvo_pt);
2231 			return (i);
2232 		}
2233 	}
2234 
2235 	/*
2236 	 * Now try secondary hash.
2237 	 */
2238 	ptegidx ^= moea_pteg_mask;
2239 	ptegidx++;
2240 	for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2241 		if ((pt->pte_hi & PTE_VALID) == 0) {
2242 			pvo_pt->pte_hi |= PTE_HID;
2243 			moea_pte_set(pt, pvo_pt);
2244 			return (i);
2245 		}
2246 	}
2247 
2248 	panic("moea_pte_insert: overflow");
2249 	return (-1);
2250 }
2251 
2252 static boolean_t
2253 moea_query_bit(vm_page_t m, int ptebit)
2254 {
2255 	struct	pvo_entry *pvo;
2256 	struct	pte *pt;
2257 
2258 #if 0
2259 	if (moea_attr_fetch(m) & ptebit)
2260 		return (TRUE);
2261 #endif
2262 
2263 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2264 		MOEA_PVO_CHECK(pvo);	/* sanity check */
2265 
2266 		/*
2267 		 * See if we saved the bit off.  If so, cache it and return
2268 		 * success.
2269 		 */
2270 		if (pvo->pvo_pte.pte_lo & ptebit) {
2271 			moea_attr_save(m, ptebit);
2272 			MOEA_PVO_CHECK(pvo);	/* sanity check */
2273 			return (TRUE);
2274 		}
2275 	}
2276 
2277 	/*
2278 	 * No luck, now go through the hard part of looking at the PTEs
2279 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2280 	 * the PTEs.
2281 	 */
2282 	SYNC();
2283 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2284 		MOEA_PVO_CHECK(pvo);	/* sanity check */
2285 
2286 		/*
2287 		 * See if this pvo has a valid PTE.  if so, fetch the
2288 		 * REF/CHG bits from the valid PTE.  If the appropriate
2289 		 * ptebit is set, cache it and return success.
2290 		 */
2291 		pt = moea_pvo_to_pte(pvo, -1);
2292 		if (pt != NULL) {
2293 			moea_pte_synch(pt, &pvo->pvo_pte);
2294 			if (pvo->pvo_pte.pte_lo & ptebit) {
2295 				moea_attr_save(m, ptebit);
2296 				MOEA_PVO_CHECK(pvo);	/* sanity check */
2297 				return (TRUE);
2298 			}
2299 		}
2300 	}
2301 
2302 	return (FALSE);
2303 }
2304 
2305 static u_int
2306 moea_clear_bit(vm_page_t m, int ptebit, int *origbit)
2307 {
2308 	u_int	count;
2309 	struct	pvo_entry *pvo;
2310 	struct	pte *pt;
2311 	int	rv;
2312 
2313 	/*
2314 	 * Clear the cached value.
2315 	 */
2316 	rv = moea_attr_fetch(m);
2317 	moea_attr_clear(m, ptebit);
2318 
2319 	/*
2320 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2321 	 * we can reset the right ones).  note that since the pvo entries and
2322 	 * list heads are accessed via BAT0 and are never placed in the page
2323 	 * table, we don't have to worry about further accesses setting the
2324 	 * REF/CHG bits.
2325 	 */
2326 	SYNC();
2327 
2328 	/*
2329 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2330 	 * valid pte clear the ptebit from the valid pte.
2331 	 */
2332 	count = 0;
2333 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2334 		MOEA_PVO_CHECK(pvo);	/* sanity check */
2335 		pt = moea_pvo_to_pte(pvo, -1);
2336 		if (pt != NULL) {
2337 			moea_pte_synch(pt, &pvo->pvo_pte);
2338 			if (pvo->pvo_pte.pte_lo & ptebit) {
2339 				count++;
2340 				moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2341 			}
2342 		}
2343 		rv |= pvo->pvo_pte.pte_lo;
2344 		pvo->pvo_pte.pte_lo &= ~ptebit;
2345 		MOEA_PVO_CHECK(pvo);	/* sanity check */
2346 	}
2347 
2348 	if (origbit != NULL) {
2349 		*origbit = rv;
2350 	}
2351 
2352 	return (count);
2353 }
2354 
2355 /*
2356  * Return true if the physical range is encompassed by the battable[idx]
2357  */
2358 static int
2359 moea_bat_mapped(int idx, vm_offset_t pa, vm_size_t size)
2360 {
2361 	u_int prot;
2362 	u_int32_t start;
2363 	u_int32_t end;
2364 	u_int32_t bat_ble;
2365 
2366 	/*
2367 	 * Return immediately if not a valid mapping
2368 	 */
2369 	if (!battable[idx].batu & BAT_Vs)
2370 		return (EINVAL);
2371 
2372 	/*
2373 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2374 	 * so it can function as an i/o page
2375 	 */
2376 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2377 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2378 		return (EPERM);
2379 
2380 	/*
2381 	 * The address should be within the BAT range. Assume that the
2382 	 * start address in the BAT has the correct alignment (thus
2383 	 * not requiring masking)
2384 	 */
2385 	start = battable[idx].batl & BAT_PBS;
2386 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2387 	end = start | (bat_ble << 15) | 0x7fff;
2388 
2389 	if ((pa < start) || ((pa + size) > end))
2390 		return (ERANGE);
2391 
2392 	return (0);
2393 }
2394 
2395 boolean_t
2396 moea_dev_direct_mapped(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2397 {
2398 	int i;
2399 
2400 	/*
2401 	 * This currently does not work for entries that
2402 	 * overlap 256M BAT segments.
2403 	 */
2404 
2405 	for(i = 0; i < 16; i++)
2406 		if (moea_bat_mapped(i, pa, size) == 0)
2407 			return (0);
2408 
2409 	return (EFAULT);
2410 }
2411 
2412 /*
2413  * Map a set of physical memory pages into the kernel virtual
2414  * address space. Return a pointer to where it is mapped. This
2415  * routine is intended to be used for mapping device memory,
2416  * NOT real memory.
2417  */
2418 void *
2419 moea_mapdev(mmu_t mmu, vm_offset_t pa, vm_size_t size)
2420 {
2421 	vm_offset_t va, tmpva, ppa, offset;
2422 	int i;
2423 
2424 	ppa = trunc_page(pa);
2425 	offset = pa & PAGE_MASK;
2426 	size = roundup(offset + size, PAGE_SIZE);
2427 
2428 	GIANT_REQUIRED;
2429 
2430 	/*
2431 	 * If the physical address lies within a valid BAT table entry,
2432 	 * return the 1:1 mapping. This currently doesn't work
2433 	 * for regions that overlap 256M BAT segments.
2434 	 */
2435 	for (i = 0; i < 16; i++) {
2436 		if (moea_bat_mapped(i, pa, size) == 0)
2437 			return ((void *) pa);
2438 	}
2439 
2440 	va = kmem_alloc_nofault(kernel_map, size);
2441 	if (!va)
2442 		panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2443 
2444 	for (tmpva = va; size > 0;) {
2445 		moea_kenter(mmu, tmpva, ppa);
2446 		TLBIE(tmpva); /* XXX or should it be invalidate-all ? */
2447 		size -= PAGE_SIZE;
2448 		tmpva += PAGE_SIZE;
2449 		ppa += PAGE_SIZE;
2450 	}
2451 
2452 	return ((void *)(va + offset));
2453 }
2454 
2455 void
2456 moea_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size)
2457 {
2458 	vm_offset_t base, offset;
2459 
2460 	/*
2461 	 * If this is outside kernel virtual space, then it's a
2462 	 * battable entry and doesn't require unmapping
2463 	 */
2464 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= VM_MAX_KERNEL_ADDRESS)) {
2465 		base = trunc_page(va);
2466 		offset = va & PAGE_MASK;
2467 		size = roundup(offset + size, PAGE_SIZE);
2468 		kmem_free(kernel_map, base, size);
2469 	}
2470 }
2471