xref: /freebsd/sys/powerpc/aim/mmu_oea.c (revision 7660b554bc59a07be0431c17e0e33815818baa69)
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/powerpc.h>
144 #include <machine/bat.h>
145 #include <machine/frame.h>
146 #include <machine/md_var.h>
147 #include <machine/psl.h>
148 #include <machine/pte.h>
149 #include <machine/sr.h>
150 
151 #define	PMAP_DEBUG
152 
153 #define TODO	panic("%s: not implemented", __func__);
154 
155 #define	PMAP_LOCK(pm)
156 #define	PMAP_UNLOCK(pm)
157 
158 #define	TLBIE(va)	__asm __volatile("tlbie %0" :: "r"(va))
159 #define	TLBSYNC()	__asm __volatile("tlbsync");
160 #define	SYNC()		__asm __volatile("sync");
161 #define	EIEIO()		__asm __volatile("eieio");
162 
163 #define	VSID_MAKE(sr, hash)	((sr) | (((hash) & 0xfffff) << 4))
164 #define	VSID_TO_SR(vsid)	((vsid) & 0xf)
165 #define	VSID_TO_HASH(vsid)	(((vsid) >> 4) & 0xfffff)
166 
167 #define	PVO_PTEGIDX_MASK	0x0007		/* which PTEG slot */
168 #define	PVO_PTEGIDX_VALID	0x0008		/* slot is valid */
169 #define	PVO_WIRED		0x0010		/* PVO entry is wired */
170 #define	PVO_MANAGED		0x0020		/* PVO entry is managed */
171 #define	PVO_EXECUTABLE		0x0040		/* PVO entry is executable */
172 #define	PVO_BOOTSTRAP		0x0080		/* PVO entry allocated during
173 						   bootstrap */
174 #define	PVO_VADDR(pvo)		((pvo)->pvo_vaddr & ~ADDR_POFF)
175 #define	PVO_ISEXECUTABLE(pvo)	((pvo)->pvo_vaddr & PVO_EXECUTABLE)
176 #define	PVO_PTEGIDX_GET(pvo)	((pvo)->pvo_vaddr & PVO_PTEGIDX_MASK)
177 #define	PVO_PTEGIDX_ISSET(pvo)	((pvo)->pvo_vaddr & PVO_PTEGIDX_VALID)
178 #define	PVO_PTEGIDX_CLR(pvo)	\
179 	((void)((pvo)->pvo_vaddr &= ~(PVO_PTEGIDX_VALID|PVO_PTEGIDX_MASK)))
180 #define	PVO_PTEGIDX_SET(pvo, i)	\
181 	((void)((pvo)->pvo_vaddr |= (i)|PVO_PTEGIDX_VALID))
182 
183 #define	PMAP_PVO_CHECK(pvo)
184 
185 struct ofw_map {
186 	vm_offset_t	om_va;
187 	vm_size_t	om_len;
188 	vm_offset_t	om_pa;
189 	u_int		om_mode;
190 };
191 
192 int	pmap_bootstrapped = 0;
193 
194 /*
195  * Virtual and physical address of message buffer.
196  */
197 struct		msgbuf *msgbufp;
198 vm_offset_t	msgbuf_phys;
199 
200 /*
201  * Physical addresses of first and last available physical page.
202  */
203 vm_offset_t avail_start;
204 vm_offset_t avail_end;
205 
206 int pmap_pagedaemon_waken;
207 
208 /*
209  * Map of physical memory regions.
210  */
211 vm_offset_t	phys_avail[128];
212 u_int		phys_avail_count;
213 static struct	mem_region *regions;
214 static struct	mem_region *pregions;
215 int		regions_sz, pregions_sz;
216 static struct	ofw_map *translations;
217 
218 /*
219  * First and last available kernel virtual addresses.
220  */
221 vm_offset_t virtual_avail;
222 vm_offset_t virtual_end;
223 vm_offset_t kernel_vm_end;
224 
225 /*
226  * Kernel pmap.
227  */
228 struct pmap kernel_pmap_store;
229 extern struct pmap ofw_pmap;
230 
231 /*
232  * PTEG data.
233  */
234 static struct	pteg *pmap_pteg_table;
235 u_int		pmap_pteg_count;
236 u_int		pmap_pteg_mask;
237 
238 /*
239  * PVO data.
240  */
241 struct	pvo_head *pmap_pvo_table;		/* pvo entries by pteg index */
242 struct	pvo_head pmap_pvo_kunmanaged =
243     LIST_HEAD_INITIALIZER(pmap_pvo_kunmanaged);	/* list of unmanaged pages */
244 struct	pvo_head pmap_pvo_unmanaged =
245     LIST_HEAD_INITIALIZER(pmap_pvo_unmanaged);	/* list of unmanaged pages */
246 
247 uma_zone_t	pmap_upvo_zone;	/* zone for pvo entries for unmanaged pages */
248 uma_zone_t	pmap_mpvo_zone;	/* zone for pvo entries for managed pages */
249 struct		vm_object pmap_upvo_zone_obj;
250 struct		vm_object pmap_mpvo_zone_obj;
251 
252 #define	BPVO_POOL_SIZE	32768
253 static struct	pvo_entry *pmap_bpvo_pool;
254 static int	pmap_bpvo_pool_index = 0;
255 
256 #define	VSID_NBPW	(sizeof(u_int32_t) * 8)
257 static u_int	pmap_vsid_bitmap[NPMAPS / VSID_NBPW];
258 
259 static boolean_t pmap_initialized = FALSE;
260 
261 /*
262  * Statistics.
263  */
264 u_int	pmap_pte_valid = 0;
265 u_int	pmap_pte_overflow = 0;
266 u_int	pmap_pte_replacements = 0;
267 u_int	pmap_pvo_entries = 0;
268 u_int	pmap_pvo_enter_calls = 0;
269 u_int	pmap_pvo_remove_calls = 0;
270 u_int	pmap_pte_spills = 0;
271 SYSCTL_INT(_machdep, OID_AUTO, pmap_pte_valid, CTLFLAG_RD, &pmap_pte_valid,
272     0, "");
273 SYSCTL_INT(_machdep, OID_AUTO, pmap_pte_overflow, CTLFLAG_RD,
274     &pmap_pte_overflow, 0, "");
275 SYSCTL_INT(_machdep, OID_AUTO, pmap_pte_replacements, CTLFLAG_RD,
276     &pmap_pte_replacements, 0, "");
277 SYSCTL_INT(_machdep, OID_AUTO, pmap_pvo_entries, CTLFLAG_RD, &pmap_pvo_entries,
278     0, "");
279 SYSCTL_INT(_machdep, OID_AUTO, pmap_pvo_enter_calls, CTLFLAG_RD,
280     &pmap_pvo_enter_calls, 0, "");
281 SYSCTL_INT(_machdep, OID_AUTO, pmap_pvo_remove_calls, CTLFLAG_RD,
282     &pmap_pvo_remove_calls, 0, "");
283 SYSCTL_INT(_machdep, OID_AUTO, pmap_pte_spills, CTLFLAG_RD,
284     &pmap_pte_spills, 0, "");
285 
286 struct	pvo_entry *pmap_pvo_zeropage;
287 
288 vm_offset_t	pmap_rkva_start = VM_MIN_KERNEL_ADDRESS;
289 u_int		pmap_rkva_count = 4;
290 
291 /*
292  * Allocate physical memory for use in pmap_bootstrap.
293  */
294 static vm_offset_t	pmap_bootstrap_alloc(vm_size_t, u_int);
295 
296 /*
297  * PTE calls.
298  */
299 static int		pmap_pte_insert(u_int, struct pte *);
300 
301 /*
302  * PVO calls.
303  */
304 static int	pmap_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *,
305 		    vm_offset_t, vm_offset_t, u_int, int);
306 static void	pmap_pvo_remove(struct pvo_entry *, int);
307 static struct	pvo_entry *pmap_pvo_find_va(pmap_t, vm_offset_t, int *);
308 static struct	pte *pmap_pvo_to_pte(const struct pvo_entry *, int);
309 
310 /*
311  * Utility routines.
312  */
313 static void *		pmap_pvo_allocf(uma_zone_t, int, u_int8_t *, int);
314 static struct		pvo_entry *pmap_rkva_alloc(void);
315 static void		pmap_pa_map(struct pvo_entry *, vm_offset_t,
316 			    struct pte *, int *);
317 static void		pmap_pa_unmap(struct pvo_entry *, struct pte *, int *);
318 static void		pmap_syncicache(vm_offset_t, vm_size_t);
319 static boolean_t	pmap_query_bit(vm_page_t, int);
320 static u_int		pmap_clear_bit(vm_page_t, int, int *);
321 static void		tlbia(void);
322 
323 static __inline int
324 va_to_sr(u_int *sr, vm_offset_t va)
325 {
326 	return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
327 }
328 
329 static __inline u_int
330 va_to_pteg(u_int sr, vm_offset_t addr)
331 {
332 	u_int hash;
333 
334 	hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
335 	    ADDR_PIDX_SHFT);
336 	return (hash & pmap_pteg_mask);
337 }
338 
339 static __inline struct pvo_head *
340 pa_to_pvoh(vm_offset_t pa, vm_page_t *pg_p)
341 {
342 	struct	vm_page *pg;
343 
344 	pg = PHYS_TO_VM_PAGE(pa);
345 
346 	if (pg_p != NULL)
347 		*pg_p = pg;
348 
349 	if (pg == NULL)
350 		return (&pmap_pvo_unmanaged);
351 
352 	return (&pg->md.mdpg_pvoh);
353 }
354 
355 static __inline struct pvo_head *
356 vm_page_to_pvoh(vm_page_t m)
357 {
358 
359 	return (&m->md.mdpg_pvoh);
360 }
361 
362 static __inline void
363 pmap_attr_clear(vm_page_t m, int ptebit)
364 {
365 
366 	m->md.mdpg_attrs &= ~ptebit;
367 }
368 
369 static __inline int
370 pmap_attr_fetch(vm_page_t m)
371 {
372 
373 	return (m->md.mdpg_attrs);
374 }
375 
376 static __inline void
377 pmap_attr_save(vm_page_t m, int ptebit)
378 {
379 
380 	m->md.mdpg_attrs |= ptebit;
381 }
382 
383 static __inline int
384 pmap_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
385 {
386 	if (pt->pte_hi == pvo_pt->pte_hi)
387 		return (1);
388 
389 	return (0);
390 }
391 
392 static __inline int
393 pmap_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
394 {
395 	return (pt->pte_hi & ~PTE_VALID) ==
396 	    (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
397 	    ((va >> ADDR_API_SHFT) & PTE_API) | which);
398 }
399 
400 static __inline void
401 pmap_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
402 {
403 	/*
404 	 * Construct a PTE.  Default to IMB initially.  Valid bit only gets
405 	 * set when the real pte is set in memory.
406 	 *
407 	 * Note: Don't set the valid bit for correct operation of tlb update.
408 	 */
409 	pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
410 	    (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
411 	pt->pte_lo = pte_lo;
412 }
413 
414 static __inline void
415 pmap_pte_synch(struct pte *pt, struct pte *pvo_pt)
416 {
417 
418 	pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
419 }
420 
421 static __inline void
422 pmap_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
423 {
424 
425 	/*
426 	 * As shown in Section 7.6.3.2.3
427 	 */
428 	pt->pte_lo &= ~ptebit;
429 	TLBIE(va);
430 	EIEIO();
431 	TLBSYNC();
432 	SYNC();
433 }
434 
435 static __inline void
436 pmap_pte_set(struct pte *pt, struct pte *pvo_pt)
437 {
438 
439 	pvo_pt->pte_hi |= PTE_VALID;
440 
441 	/*
442 	 * Update the PTE as defined in section 7.6.3.1.
443 	 * Note that the REF/CHG bits are from pvo_pt and thus should havce
444 	 * been saved so this routine can restore them (if desired).
445 	 */
446 	pt->pte_lo = pvo_pt->pte_lo;
447 	EIEIO();
448 	pt->pte_hi = pvo_pt->pte_hi;
449 	SYNC();
450 	pmap_pte_valid++;
451 }
452 
453 static __inline void
454 pmap_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
455 {
456 
457 	pvo_pt->pte_hi &= ~PTE_VALID;
458 
459 	/*
460 	 * Force the reg & chg bits back into the PTEs.
461 	 */
462 	SYNC();
463 
464 	/*
465 	 * Invalidate the pte.
466 	 */
467 	pt->pte_hi &= ~PTE_VALID;
468 
469 	SYNC();
470 	TLBIE(va);
471 	EIEIO();
472 	TLBSYNC();
473 	SYNC();
474 
475 	/*
476 	 * Save the reg & chg bits.
477 	 */
478 	pmap_pte_synch(pt, pvo_pt);
479 	pmap_pte_valid--;
480 }
481 
482 static __inline void
483 pmap_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
484 {
485 
486 	/*
487 	 * Invalidate the PTE
488 	 */
489 	pmap_pte_unset(pt, pvo_pt, va);
490 	pmap_pte_set(pt, pvo_pt);
491 }
492 
493 /*
494  * Quick sort callout for comparing memory regions.
495  */
496 static int	mr_cmp(const void *a, const void *b);
497 static int	om_cmp(const void *a, const void *b);
498 
499 static int
500 mr_cmp(const void *a, const void *b)
501 {
502 	const struct	mem_region *regiona;
503 	const struct	mem_region *regionb;
504 
505 	regiona = a;
506 	regionb = b;
507 	if (regiona->mr_start < regionb->mr_start)
508 		return (-1);
509 	else if (regiona->mr_start > regionb->mr_start)
510 		return (1);
511 	else
512 		return (0);
513 }
514 
515 static int
516 om_cmp(const void *a, const void *b)
517 {
518 	const struct	ofw_map *mapa;
519 	const struct	ofw_map *mapb;
520 
521 	mapa = a;
522 	mapb = b;
523 	if (mapa->om_pa < mapb->om_pa)
524 		return (-1);
525 	else if (mapa->om_pa > mapb->om_pa)
526 		return (1);
527 	else
528 		return (0);
529 }
530 
531 void
532 pmap_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
533 {
534 	ihandle_t	mmui;
535 	phandle_t	chosen, mmu;
536 	int		sz;
537 	int		i, j;
538 	int		ofw_mappings;
539 	vm_size_t	size, physsz;
540 	vm_offset_t	pa, va, off;
541 	u_int		batl, batu;
542 
543         /*
544          * Set up BAT0 to map the lowest 256 MB area
545          */
546         battable[0x0].batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
547         battable[0x0].batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
548 
549         /*
550          * Map PCI memory space.
551          */
552         battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
553         battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
554 
555         battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
556         battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
557 
558         battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
559         battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
560 
561         battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
562         battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
563 
564         /*
565          * Map obio devices.
566          */
567         battable[0xf].batl = BATL(0xf0000000, BAT_I|BAT_G, BAT_PP_RW);
568         battable[0xf].batu = BATU(0xf0000000, BAT_BL_256M, BAT_Vs);
569 
570 	/*
571 	 * Use an IBAT and a DBAT to map the bottom segment of memory
572 	 * where we are.
573 	 */
574 	batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs);
575 	batl = BATL(0x00000000, BAT_M, BAT_PP_RW);
576 	__asm ("mtibatu 0,%0; mtibatl 0,%1; mtdbatu 0,%0; mtdbatl 0,%1"
577 	    :: "r"(batu), "r"(batl));
578 
579 #if 0
580 	/* map frame buffer */
581 	batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
582 	batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
583 	__asm ("mtdbatu 1,%0; mtdbatl 1,%1"
584 	    :: "r"(batu), "r"(batl));
585 #endif
586 
587 #if 1
588 	/* map pci space */
589 	batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
590 	batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
591 	__asm ("mtdbatu 1,%0; mtdbatl 1,%1"
592 	    :: "r"(batu), "r"(batl));
593 #endif
594 
595 	/*
596 	 * Set the start and end of kva.
597 	 */
598 	virtual_avail = VM_MIN_KERNEL_ADDRESS;
599 	virtual_end = VM_MAX_KERNEL_ADDRESS;
600 
601 	mem_regions(&pregions, &pregions_sz, &regions, &regions_sz);
602 	CTR0(KTR_PMAP, "pmap_bootstrap: physical memory");
603 
604 	qsort(pregions, pregions_sz, sizeof(*pregions), mr_cmp);
605 	for (i = 0; i < pregions_sz; i++) {
606 		vm_offset_t pa;
607 		vm_offset_t end;
608 
609 		CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
610 			pregions[i].mr_start,
611 			pregions[i].mr_start + pregions[i].mr_size,
612 			pregions[i].mr_size);
613 		/*
614 		 * Install entries into the BAT table to allow all
615 		 * of physmem to be convered by on-demand BAT entries.
616 		 * The loop will sometimes set the same battable element
617 		 * twice, but that's fine since they won't be used for
618 		 * a while yet.
619 		 */
620 		pa = pregions[i].mr_start & 0xf0000000;
621 		end = pregions[i].mr_start + pregions[i].mr_size;
622 		do {
623                         u_int n = pa >> ADDR_SR_SHFT;
624 
625 			battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
626 			battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
627 			pa += SEGMENT_LENGTH;
628 		} while (pa < end);
629 	}
630 
631 	if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz)
632 		panic("pmap_bootstrap: phys_avail too small");
633 	qsort(regions, regions_sz, sizeof(*regions), mr_cmp);
634 	phys_avail_count = 0;
635 	physsz = 0;
636 	for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
637 		CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
638 		    regions[i].mr_start + regions[i].mr_size,
639 		    regions[i].mr_size);
640 		phys_avail[j] = regions[i].mr_start;
641 		phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
642 		phys_avail_count++;
643 		physsz += regions[i].mr_size;
644 	}
645 	physmem = btoc(physsz);
646 
647 	/*
648 	 * Allocate PTEG table.
649 	 */
650 #ifdef PTEGCOUNT
651 	pmap_pteg_count = PTEGCOUNT;
652 #else
653 	pmap_pteg_count = 0x1000;
654 
655 	while (pmap_pteg_count < physmem)
656 		pmap_pteg_count <<= 1;
657 
658 	pmap_pteg_count >>= 1;
659 #endif /* PTEGCOUNT */
660 
661 	size = pmap_pteg_count * sizeof(struct pteg);
662 	CTR2(KTR_PMAP, "pmap_bootstrap: %d PTEGs, %d bytes", pmap_pteg_count,
663 	    size);
664 	pmap_pteg_table = (struct pteg *)pmap_bootstrap_alloc(size, size);
665 	CTR1(KTR_PMAP, "pmap_bootstrap: PTEG table at %p", pmap_pteg_table);
666 	bzero((void *)pmap_pteg_table, pmap_pteg_count * sizeof(struct pteg));
667 	pmap_pteg_mask = pmap_pteg_count - 1;
668 
669 	/*
670 	 * Allocate pv/overflow lists.
671 	 */
672 	size = sizeof(struct pvo_head) * pmap_pteg_count;
673 	pmap_pvo_table = (struct pvo_head *)pmap_bootstrap_alloc(size,
674 	    PAGE_SIZE);
675 	CTR1(KTR_PMAP, "pmap_bootstrap: PVO table at %p", pmap_pvo_table);
676 	for (i = 0; i < pmap_pteg_count; i++)
677 		LIST_INIT(&pmap_pvo_table[i]);
678 
679 	/*
680 	 * Allocate the message buffer.
681 	 */
682 	msgbuf_phys = pmap_bootstrap_alloc(MSGBUF_SIZE, 0);
683 
684 	/*
685 	 * Initialise the unmanaged pvo pool.
686 	 */
687 	pmap_bpvo_pool = (struct pvo_entry *)pmap_bootstrap_alloc(
688 		BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
689 	pmap_bpvo_pool_index = 0;
690 
691 	/*
692 	 * Make sure kernel vsid is allocated as well as VSID 0.
693 	 */
694 	pmap_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
695 		|= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
696 	pmap_vsid_bitmap[0] |= 1;
697 
698 	/*
699 	 * Set up the OpenFirmware pmap and add it's mappings.
700 	 */
701 	pmap_pinit(&ofw_pmap);
702 	ofw_pmap.pm_sr[KERNEL_SR] = KERNEL_SEGMENT;
703 	if ((chosen = OF_finddevice("/chosen")) == -1)
704 		panic("pmap_bootstrap: can't find /chosen");
705 	OF_getprop(chosen, "mmu", &mmui, 4);
706 	if ((mmu = OF_instance_to_package(mmui)) == -1)
707 		panic("pmap_bootstrap: can't get mmu package");
708 	if ((sz = OF_getproplen(mmu, "translations")) == -1)
709 		panic("pmap_bootstrap: can't get ofw translation count");
710 	translations = NULL;
711 	for (i = 0; phys_avail[i + 2] != 0; i += 2) {
712 		if (phys_avail[i + 1] >= sz)
713 			translations = (struct ofw_map *)phys_avail[i];
714 	}
715 	if (translations == NULL)
716 		panic("pmap_bootstrap: no space to copy translations");
717 	bzero(translations, sz);
718 	if (OF_getprop(mmu, "translations", translations, sz) == -1)
719 		panic("pmap_bootstrap: can't get ofw translations");
720 	CTR0(KTR_PMAP, "pmap_bootstrap: translations");
721 	sz /= sizeof(*translations);
722 	qsort(translations, sz, sizeof (*translations), om_cmp);
723 	for (i = 0, ofw_mappings = 0; i < sz; i++) {
724 		CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
725 		    translations[i].om_pa, translations[i].om_va,
726 		    translations[i].om_len);
727 
728 		/*
729 		 * If the mapping is 1:1, let the RAM and device on-demand
730 		 * BAT tables take care of the translation.
731 		 */
732 		if (translations[i].om_va == translations[i].om_pa)
733 			continue;
734 
735 		/* Enter the pages */
736 		for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) {
737 			struct	vm_page m;
738 
739 			m.phys_addr = translations[i].om_pa + off;
740 			pmap_enter(&ofw_pmap, translations[i].om_va + off, &m,
741 				   VM_PROT_ALL, 1);
742 			ofw_mappings++;
743 		}
744 	}
745 #ifdef SMP
746 	TLBSYNC();
747 #endif
748 
749 	/*
750 	 * Initialize the kernel pmap (which is statically allocated).
751 	 */
752 	for (i = 0; i < 16; i++) {
753 		kernel_pmap->pm_sr[i] = EMPTY_SEGMENT;
754 	}
755 	kernel_pmap->pm_sr[KERNEL_SR] = KERNEL_SEGMENT;
756 	kernel_pmap->pm_active = ~0;
757 
758 	/*
759 	 * Allocate a kernel stack with a guard page for thread0 and map it
760 	 * into the kernel page map.
761 	 */
762 	pa = pmap_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, 0);
763 	kstack0_phys = pa;
764 	kstack0 = virtual_avail + (KSTACK_GUARD_PAGES * PAGE_SIZE);
765 	CTR2(KTR_PMAP, "pmap_bootstrap: kstack0 at %#x (%#x)", kstack0_phys,
766 	    kstack0);
767 	virtual_avail += (KSTACK_PAGES + KSTACK_GUARD_PAGES) * PAGE_SIZE;
768 	for (i = 0; i < KSTACK_PAGES; i++) {
769 		pa = kstack0_phys + i * PAGE_SIZE;
770 		va = kstack0 + i * PAGE_SIZE;
771 		pmap_kenter(va, pa);
772 		TLBIE(va);
773 	}
774 
775 	/*
776 	 * Calculate the first and last available physical addresses.
777 	 */
778 	avail_start = phys_avail[0];
779 	for (i = 0; phys_avail[i + 2] != 0; i += 2)
780 		;
781 	avail_end = phys_avail[i + 1];
782 	Maxmem = powerpc_btop(avail_end);
783 
784 	/*
785 	 * Allocate virtual address space for the message buffer.
786 	 */
787 	msgbufp = (struct msgbuf *)virtual_avail;
788 	virtual_avail += round_page(MSGBUF_SIZE);
789 
790 	/*
791 	 * Initialize hardware.
792 	 */
793 	for (i = 0; i < 16; i++) {
794 		mtsrin(i << ADDR_SR_SHFT, EMPTY_SEGMENT);
795 	}
796 	__asm __volatile ("mtsr %0,%1"
797 	    :: "n"(KERNEL_SR), "r"(KERNEL_SEGMENT));
798 	__asm __volatile ("sync; mtsdr1 %0; isync"
799 	    :: "r"((u_int)pmap_pteg_table | (pmap_pteg_mask >> 10)));
800 	tlbia();
801 
802 	pmap_bootstrapped++;
803 }
804 
805 /*
806  * Activate a user pmap.  The pmap must be activated before it's address
807  * space can be accessed in any way.
808  */
809 void
810 pmap_activate(struct thread *td)
811 {
812 	pmap_t	pm, pmr;
813 
814 	/*
815 	 * Load all the data we need up front to encourage the compiler to
816 	 * not issue any loads while we have interrupts disabled below.
817 	 */
818 	pm = &td->td_proc->p_vmspace->vm_pmap;
819 
820 	if ((pmr = (pmap_t)pmap_kextract((vm_offset_t)pm)) == NULL)
821 		pmr = pm;
822 
823 	pm->pm_active |= PCPU_GET(cpumask);
824 	PCPU_SET(curpmap, pmr);
825 }
826 
827 void
828 pmap_deactivate(struct thread *td)
829 {
830 	pmap_t	pm;
831 
832 	pm = &td->td_proc->p_vmspace->vm_pmap;
833 	pm->pm_active &= ~(PCPU_GET(cpumask));
834 	PCPU_SET(curpmap, NULL);
835 }
836 
837 vm_offset_t
838 pmap_addr_hint(vm_object_t object, vm_offset_t va, vm_size_t size)
839 {
840 
841 	return (va);
842 }
843 
844 void
845 pmap_change_wiring(pmap_t pm, vm_offset_t va, boolean_t wired)
846 {
847 	struct	pvo_entry *pvo;
848 
849 	pvo = pmap_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
850 
851 	if (pvo != NULL) {
852 		if (wired) {
853 			if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
854 				pm->pm_stats.wired_count++;
855 			pvo->pvo_vaddr |= PVO_WIRED;
856 		} else {
857 			if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
858 				pm->pm_stats.wired_count--;
859 			pvo->pvo_vaddr &= ~PVO_WIRED;
860 		}
861 	}
862 }
863 
864 void
865 pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr,
866 	  vm_size_t len, vm_offset_t src_addr)
867 {
868 
869 	/*
870 	 * This is not needed as it's mainly an optimisation.
871 	 * It may want to be implemented later though.
872 	 */
873 }
874 
875 void
876 pmap_copy_page(vm_page_t msrc, vm_page_t mdst)
877 {
878 	vm_offset_t	dst;
879 	vm_offset_t	src;
880 
881 	dst = VM_PAGE_TO_PHYS(mdst);
882 	src = VM_PAGE_TO_PHYS(msrc);
883 
884 	kcopy((void *)src, (void *)dst, PAGE_SIZE);
885 }
886 
887 /*
888  * Zero a page of physical memory by temporarily mapping it into the tlb.
889  */
890 void
891 pmap_zero_page(vm_page_t m)
892 {
893 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
894 	caddr_t va;
895 
896 	if (pa < SEGMENT_LENGTH) {
897 		va = (caddr_t) pa;
898 	} else if (pmap_initialized) {
899 		if (pmap_pvo_zeropage == NULL)
900 			pmap_pvo_zeropage = pmap_rkva_alloc();
901 		pmap_pa_map(pmap_pvo_zeropage, pa, NULL, NULL);
902 		va = (caddr_t)PVO_VADDR(pmap_pvo_zeropage);
903 	} else {
904 		panic("pmap_zero_page: can't zero pa %#x", pa);
905 	}
906 
907 	bzero(va, PAGE_SIZE);
908 
909 	if (pa >= SEGMENT_LENGTH)
910 		pmap_pa_unmap(pmap_pvo_zeropage, NULL, NULL);
911 }
912 
913 void
914 pmap_zero_page_area(vm_page_t m, int off, int size)
915 {
916 	vm_offset_t pa = VM_PAGE_TO_PHYS(m);
917 	caddr_t va;
918 
919 	if (pa < SEGMENT_LENGTH) {
920 		va = (caddr_t) pa;
921 	} else if (pmap_initialized) {
922 		if (pmap_pvo_zeropage == NULL)
923 			pmap_pvo_zeropage = pmap_rkva_alloc();
924 		pmap_pa_map(pmap_pvo_zeropage, pa, NULL, NULL);
925 		va = (caddr_t)PVO_VADDR(pmap_pvo_zeropage);
926 	} else {
927 		panic("pmap_zero_page: can't zero pa %#x", pa);
928 	}
929 
930 	bzero(va + off, size);
931 
932 	if (pa >= SEGMENT_LENGTH)
933 		pmap_pa_unmap(pmap_pvo_zeropage, NULL, NULL);
934 }
935 
936 void
937 pmap_zero_page_idle(vm_page_t m)
938 {
939 
940 	/* XXX this is called outside of Giant, is pmap_zero_page safe? */
941 	/* XXX maybe have a dedicated mapping for this to avoid the problem? */
942 	mtx_lock(&Giant);
943 	pmap_zero_page(m);
944 	mtx_unlock(&Giant);
945 }
946 
947 /*
948  * Map the given physical page at the specified virtual address in the
949  * target pmap with the protection requested.  If specified the page
950  * will be wired down.
951  */
952 void
953 pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
954 	   boolean_t wired)
955 {
956 	struct		pvo_head *pvo_head;
957 	uma_zone_t	zone;
958 	vm_page_t	pg;
959 	u_int		pte_lo, pvo_flags, was_exec, i;
960 	int		error;
961 
962 	if (!pmap_initialized) {
963 		pvo_head = &pmap_pvo_kunmanaged;
964 		zone = pmap_upvo_zone;
965 		pvo_flags = 0;
966 		pg = NULL;
967 		was_exec = PTE_EXEC;
968 	} else {
969 		pvo_head = vm_page_to_pvoh(m);
970 		pg = m;
971 		zone = pmap_mpvo_zone;
972 		pvo_flags = PVO_MANAGED;
973 		was_exec = 0;
974 	}
975 
976 	/*
977 	 * If this is a managed page, and it's the first reference to the page,
978 	 * clear the execness of the page.  Otherwise fetch the execness.
979 	 */
980 	if (pg != NULL) {
981 		if (LIST_EMPTY(pvo_head)) {
982 			pmap_attr_clear(pg, PTE_EXEC);
983 		} else {
984 			was_exec = pmap_attr_fetch(pg) & PTE_EXEC;
985 		}
986 	}
987 
988 
989 	/*
990 	 * Assume the page is cache inhibited and access is guarded unless
991 	 * it's in our available memory array.
992 	 */
993 	pte_lo = PTE_I | PTE_G;
994 	for (i = 0; i < pregions_sz; i++) {
995 		if ((VM_PAGE_TO_PHYS(m) >= pregions[i].mr_start) &&
996 		    (VM_PAGE_TO_PHYS(m) <
997 			(pregions[i].mr_start + pregions[i].mr_size))) {
998 			pte_lo &= ~(PTE_I | PTE_G);
999 			break;
1000 		}
1001 	}
1002 
1003 	if (prot & VM_PROT_WRITE)
1004 		pte_lo |= PTE_BW;
1005 	else
1006 		pte_lo |= PTE_BR;
1007 
1008 	pvo_flags |= (prot & VM_PROT_EXECUTE);
1009 
1010 	if (wired)
1011 		pvo_flags |= PVO_WIRED;
1012 
1013 	error = pmap_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1014 	    pte_lo, pvo_flags);
1015 
1016 	/*
1017 	 * Flush the real page from the instruction cache if this page is
1018 	 * mapped executable and cacheable and was not previously mapped (or
1019 	 * was not mapped executable).
1020 	 */
1021 	if (error == 0 && (pvo_flags & PVO_EXECUTABLE) &&
1022 	    (pte_lo & PTE_I) == 0 && was_exec == 0) {
1023 		/*
1024 		 * Flush the real memory from the cache.
1025 		 */
1026 		pmap_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1027 		if (pg != NULL)
1028 			pmap_attr_save(pg, PTE_EXEC);
1029 	}
1030 
1031 	/* XXX syncicache always until problems are sorted */
1032 	pmap_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1033 }
1034 
1035 vm_page_t
1036 pmap_enter_quick(pmap_t pm, vm_offset_t va, vm_page_t m, vm_page_t mpte)
1037 {
1038 
1039 	pmap_enter(pm, va, m, VM_PROT_READ | VM_PROT_EXECUTE, FALSE);
1040 	return (NULL);
1041 }
1042 
1043 vm_offset_t
1044 pmap_extract(pmap_t pm, vm_offset_t va)
1045 {
1046 	struct	pvo_entry *pvo;
1047 
1048 	pvo = pmap_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1049 
1050 	if (pvo != NULL) {
1051 		return ((pvo->pvo_pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF));
1052 	}
1053 
1054 	return (0);
1055 }
1056 
1057 /*
1058  * Grow the number of kernel page table entries.  Unneeded.
1059  */
1060 void
1061 pmap_growkernel(vm_offset_t addr)
1062 {
1063 }
1064 
1065 void
1066 pmap_init(vm_offset_t phys_start, vm_offset_t phys_end)
1067 {
1068 
1069 	CTR0(KTR_PMAP, "pmap_init");
1070 
1071 	pmap_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1072 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE);
1073 	uma_zone_set_allocf(pmap_upvo_zone, pmap_pvo_allocf);
1074 	pmap_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1075 	    NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE);
1076 	uma_zone_set_allocf(pmap_mpvo_zone, pmap_pvo_allocf);
1077 	pmap_initialized = TRUE;
1078 }
1079 
1080 void
1081 pmap_init2(void)
1082 {
1083 
1084 	CTR0(KTR_PMAP, "pmap_init2");
1085 }
1086 
1087 boolean_t
1088 pmap_is_modified(vm_page_t m)
1089 {
1090 
1091 	if ((m->flags & (PG_FICTITIOUS |PG_UNMANAGED)) != 0)
1092 		return (FALSE);
1093 
1094 	return (pmap_query_bit(m, PTE_CHG));
1095 }
1096 
1097 void
1098 pmap_clear_reference(vm_page_t m)
1099 {
1100 
1101 	if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0)
1102 		return;
1103 	pmap_clear_bit(m, PTE_REF, NULL);
1104 }
1105 
1106 void
1107 pmap_clear_modify(vm_page_t m)
1108 {
1109 
1110 	if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0)
1111 		return;
1112 	pmap_clear_bit(m, PTE_CHG, NULL);
1113 }
1114 
1115 /*
1116  *	pmap_ts_referenced:
1117  *
1118  *	Return a count of reference bits for a page, clearing those bits.
1119  *	It is not necessary for every reference bit to be cleared, but it
1120  *	is necessary that 0 only be returned when there are truly no
1121  *	reference bits set.
1122  *
1123  *	XXX: The exact number of bits to check and clear is a matter that
1124  *	should be tested and standardized at some point in the future for
1125  *	optimal aging of shared pages.
1126  */
1127 int
1128 pmap_ts_referenced(vm_page_t m)
1129 {
1130 	int count;
1131 
1132 	if ((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0)
1133 		return (0);
1134 
1135 	count = pmap_clear_bit(m, PTE_REF, NULL);
1136 
1137 	return (count);
1138 }
1139 
1140 /*
1141  * Map a wired page into kernel virtual address space.
1142  */
1143 void
1144 pmap_kenter(vm_offset_t va, vm_offset_t pa)
1145 {
1146 	u_int		pte_lo;
1147 	int		error;
1148 	int		i;
1149 
1150 #if 0
1151 	if (va < VM_MIN_KERNEL_ADDRESS)
1152 		panic("pmap_kenter: attempt to enter non-kernel address %#x",
1153 		    va);
1154 #endif
1155 
1156 	pte_lo = PTE_I | PTE_G;
1157 	for (i = 0; i < pregions_sz; i++) {
1158 		if ((pa >= pregions[i].mr_start) &&
1159 		    (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
1160 			pte_lo &= ~(PTE_I | PTE_G);
1161 			break;
1162 		}
1163 	}
1164 
1165 	error = pmap_pvo_enter(kernel_pmap, pmap_upvo_zone,
1166 	    &pmap_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1167 
1168 	if (error != 0 && error != ENOENT)
1169 		panic("pmap_kenter: failed to enter va %#x pa %#x: %d", va,
1170 		    pa, error);
1171 
1172 	/*
1173 	 * Flush the real memory from the instruction cache.
1174 	 */
1175 	if ((pte_lo & (PTE_I | PTE_G)) == 0) {
1176 		pmap_syncicache(pa, PAGE_SIZE);
1177 	}
1178 }
1179 
1180 /*
1181  * Extract the physical page address associated with the given kernel virtual
1182  * address.
1183  */
1184 vm_offset_t
1185 pmap_kextract(vm_offset_t va)
1186 {
1187 	struct		pvo_entry *pvo;
1188 
1189 	pvo = pmap_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1190 	if (pvo == NULL) {
1191 		return (0);
1192 	}
1193 
1194 	return ((pvo->pvo_pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF));
1195 }
1196 
1197 /*
1198  * Remove a wired page from kernel virtual address space.
1199  */
1200 void
1201 pmap_kremove(vm_offset_t va)
1202 {
1203 
1204 	pmap_remove(kernel_pmap, va, va + PAGE_SIZE);
1205 }
1206 
1207 /*
1208  * Map a range of physical addresses into kernel virtual address space.
1209  *
1210  * The value passed in *virt is a suggested virtual address for the mapping.
1211  * Architectures which can support a direct-mapped physical to virtual region
1212  * can return the appropriate address within that region, leaving '*virt'
1213  * unchanged.  We cannot and therefore do not; *virt is updated with the
1214  * first usable address after the mapped region.
1215  */
1216 vm_offset_t
1217 pmap_map(vm_offset_t *virt, vm_offset_t pa_start, vm_offset_t pa_end, int prot)
1218 {
1219 	vm_offset_t	sva, va;
1220 
1221 	sva = *virt;
1222 	va = sva;
1223 	for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1224 		pmap_kenter(va, pa_start);
1225 	*virt = va;
1226 	return (sva);
1227 }
1228 
1229 int
1230 pmap_mincore(pmap_t pmap, vm_offset_t addr)
1231 {
1232 	TODO;
1233 	return (0);
1234 }
1235 
1236 void
1237 pmap_object_init_pt(pmap_t pm, vm_offset_t addr, vm_object_t object,
1238 		    vm_pindex_t pindex, vm_size_t size)
1239 {
1240 
1241 	VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
1242 	KASSERT(object->type == OBJT_DEVICE,
1243 	    ("pmap_object_init_pt: non-device object"));
1244 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1245 	    ("pmap_object_init_pt: non current pmap"));
1246 }
1247 
1248 /*
1249  * Lower the permission for all mappings to a given page.
1250  */
1251 void
1252 pmap_page_protect(vm_page_t m, vm_prot_t prot)
1253 {
1254 	struct	pvo_head *pvo_head;
1255 	struct	pvo_entry *pvo, *next_pvo;
1256 	struct	pte *pt;
1257 
1258 	/*
1259 	 * Since the routine only downgrades protection, if the
1260 	 * maximal protection is desired, there isn't any change
1261 	 * to be made.
1262 	 */
1263 	if ((prot & (VM_PROT_READ|VM_PROT_WRITE)) ==
1264 	    (VM_PROT_READ|VM_PROT_WRITE))
1265 		return;
1266 
1267 	pvo_head = vm_page_to_pvoh(m);
1268 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1269 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1270 		PMAP_PVO_CHECK(pvo);	/* sanity check */
1271 
1272 		/*
1273 		 * Downgrading to no mapping at all, we just remove the entry.
1274 		 */
1275 		if ((prot & VM_PROT_READ) == 0) {
1276 			pmap_pvo_remove(pvo, -1);
1277 			continue;
1278 		}
1279 
1280 		/*
1281 		 * If EXEC permission is being revoked, just clear the flag
1282 		 * in the PVO.
1283 		 */
1284 		if ((prot & VM_PROT_EXECUTE) == 0)
1285 			pvo->pvo_vaddr &= ~PVO_EXECUTABLE;
1286 
1287 		/*
1288 		 * If this entry is already RO, don't diddle with the page
1289 		 * table.
1290 		 */
1291 		if ((pvo->pvo_pte.pte_lo & PTE_PP) == PTE_BR) {
1292 			PMAP_PVO_CHECK(pvo);
1293 			continue;
1294 		}
1295 
1296 		/*
1297 		 * Grab the PTE before we diddle the bits so pvo_to_pte can
1298 		 * verify the pte contents are as expected.
1299 		 */
1300 		pt = pmap_pvo_to_pte(pvo, -1);
1301 		pvo->pvo_pte.pte_lo &= ~PTE_PP;
1302 		pvo->pvo_pte.pte_lo |= PTE_BR;
1303 		if (pt != NULL)
1304 			pmap_pte_change(pt, &pvo->pvo_pte, pvo->pvo_vaddr);
1305 		PMAP_PVO_CHECK(pvo);	/* sanity check */
1306 	}
1307 }
1308 
1309 /*
1310  * Returns true if the pmap's pv is one of the first
1311  * 16 pvs linked to from this page.  This count may
1312  * be changed upwards or downwards in the future; it
1313  * is only necessary that true be returned for a small
1314  * subset of pmaps for proper page aging.
1315  */
1316 boolean_t
1317 pmap_page_exists_quick(pmap_t pmap, vm_page_t m)
1318 {
1319         int loops;
1320 	struct pvo_entry *pvo;
1321 
1322         if (!pmap_initialized || (m->flags & PG_FICTITIOUS))
1323                 return FALSE;
1324 
1325 	loops = 0;
1326 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1327 		if (pvo->pvo_pmap == pmap)
1328 			return (TRUE);
1329 		if (++loops >= 16)
1330 			break;
1331 	}
1332 
1333 	return (FALSE);
1334 }
1335 
1336 static u_int	pmap_vsidcontext;
1337 
1338 void
1339 pmap_pinit(pmap_t pmap)
1340 {
1341 	int	i, mask;
1342 	u_int	entropy;
1343 
1344 	entropy = 0;
1345 	__asm __volatile("mftb %0" : "=r"(entropy));
1346 
1347 	/*
1348 	 * Allocate some segment registers for this pmap.
1349 	 */
1350 	for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1351 		u_int	hash, n;
1352 
1353 		/*
1354 		 * Create a new value by mutiplying by a prime and adding in
1355 		 * entropy from the timebase register.  This is to make the
1356 		 * VSID more random so that the PT hash function collides
1357 		 * less often.  (Note that the prime casues gcc to do shifts
1358 		 * instead of a multiply.)
1359 		 */
1360 		pmap_vsidcontext = (pmap_vsidcontext * 0x1105) + entropy;
1361 		hash = pmap_vsidcontext & (NPMAPS - 1);
1362 		if (hash == 0)		/* 0 is special, avoid it */
1363 			continue;
1364 		n = hash >> 5;
1365 		mask = 1 << (hash & (VSID_NBPW - 1));
1366 		hash = (pmap_vsidcontext & 0xfffff);
1367 		if (pmap_vsid_bitmap[n] & mask) {	/* collision? */
1368 			/* anything free in this bucket? */
1369 			if (pmap_vsid_bitmap[n] == 0xffffffff) {
1370 				entropy = (pmap_vsidcontext >> 20);
1371 				continue;
1372 			}
1373 			i = ffs(~pmap_vsid_bitmap[i]) - 1;
1374 			mask = 1 << i;
1375 			hash &= 0xfffff & ~(VSID_NBPW - 1);
1376 			hash |= i;
1377 		}
1378 		pmap_vsid_bitmap[n] |= mask;
1379 		for (i = 0; i < 16; i++)
1380 			pmap->pm_sr[i] = VSID_MAKE(i, hash);
1381 		return;
1382 	}
1383 
1384 	panic("pmap_pinit: out of segments");
1385 }
1386 
1387 /*
1388  * Initialize the pmap associated with process 0.
1389  */
1390 void
1391 pmap_pinit0(pmap_t pm)
1392 {
1393 
1394 	pmap_pinit(pm);
1395 	bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1396 }
1397 
1398 void
1399 pmap_pinit2(pmap_t pmap)
1400 {
1401 	/* XXX: Remove this stub when no longer called */
1402 }
1403 
1404 void
1405 pmap_prefault(pmap_t pm, vm_offset_t va, vm_map_entry_t entry)
1406 {
1407 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1408 	    ("pmap_prefault: non current pmap"));
1409 	/* XXX */
1410 }
1411 
1412 /*
1413  * Set the physical protection on the specified range of this map as requested.
1414  */
1415 void
1416 pmap_protect(pmap_t pm, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot)
1417 {
1418 	struct	pvo_entry *pvo;
1419 	struct	pte *pt;
1420 	int	pteidx;
1421 
1422 	CTR4(KTR_PMAP, "pmap_protect: pm=%p sva=%#x eva=%#x prot=%#x", pm, sva,
1423 	    eva, prot);
1424 
1425 
1426 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1427 	    ("pmap_protect: non current pmap"));
1428 
1429 	if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1430 		pmap_remove(pm, sva, eva);
1431 		return;
1432 	}
1433 
1434 	for (; sva < eva; sva += PAGE_SIZE) {
1435 		pvo = pmap_pvo_find_va(pm, sva, &pteidx);
1436 		if (pvo == NULL)
1437 			continue;
1438 
1439 		if ((prot & VM_PROT_EXECUTE) == 0)
1440 			pvo->pvo_vaddr &= ~PVO_EXECUTABLE;
1441 
1442 		/*
1443 		 * Grab the PTE pointer before we diddle with the cached PTE
1444 		 * copy.
1445 		 */
1446 		pt = pmap_pvo_to_pte(pvo, pteidx);
1447 		/*
1448 		 * Change the protection of the page.
1449 		 */
1450 		pvo->pvo_pte.pte_lo &= ~PTE_PP;
1451 		pvo->pvo_pte.pte_lo |= PTE_BR;
1452 
1453 		/*
1454 		 * If the PVO is in the page table, update that pte as well.
1455 		 */
1456 		if (pt != NULL)
1457 			pmap_pte_change(pt, &pvo->pvo_pte, pvo->pvo_vaddr);
1458 	}
1459 }
1460 
1461 /*
1462  * Map a list of wired pages into kernel virtual address space.  This is
1463  * intended for temporary mappings which do not need page modification or
1464  * references recorded.  Existing mappings in the region are overwritten.
1465  */
1466 void
1467 pmap_qenter(vm_offset_t sva, vm_page_t *m, int count)
1468 {
1469 	vm_offset_t va;
1470 
1471 	va = sva;
1472 	while (count-- > 0) {
1473 		pmap_kenter(va, VM_PAGE_TO_PHYS(*m));
1474 		va += PAGE_SIZE;
1475 		m++;
1476 	}
1477 }
1478 
1479 /*
1480  * Remove page mappings from kernel virtual address space.  Intended for
1481  * temporary mappings entered by pmap_qenter.
1482  */
1483 void
1484 pmap_qremove(vm_offset_t sva, int count)
1485 {
1486 	vm_offset_t va;
1487 
1488 	va = sva;
1489 	while (count-- > 0) {
1490 		pmap_kremove(va);
1491 		va += PAGE_SIZE;
1492 	}
1493 }
1494 
1495 void
1496 pmap_release(pmap_t pmap)
1497 {
1498         int idx, mask;
1499 
1500 	/*
1501 	 * Free segment register's VSID
1502 	 */
1503         if (pmap->pm_sr[0] == 0)
1504                 panic("pmap_release");
1505 
1506         idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1507         mask = 1 << (idx % VSID_NBPW);
1508         idx /= VSID_NBPW;
1509         pmap_vsid_bitmap[idx] &= ~mask;
1510 }
1511 
1512 /*
1513  * Remove the given range of addresses from the specified map.
1514  */
1515 void
1516 pmap_remove(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1517 {
1518 	struct	pvo_entry *pvo;
1519 	int	pteidx;
1520 
1521 	for (; sva < eva; sva += PAGE_SIZE) {
1522 		pvo = pmap_pvo_find_va(pm, sva, &pteidx);
1523 		if (pvo != NULL) {
1524 			pmap_pvo_remove(pvo, pteidx);
1525 		}
1526 	}
1527 }
1528 
1529 /*
1530  * Remove physical page from all pmaps in which it resides. pmap_pvo_remove()
1531  * will reflect changes in pte's back to the vm_page.
1532  */
1533 void
1534 pmap_remove_all(vm_page_t m)
1535 {
1536 	struct  pvo_head *pvo_head;
1537 	struct	pvo_entry *pvo, *next_pvo;
1538 
1539 	KASSERT((m->flags & (PG_FICTITIOUS|PG_UNMANAGED)) == 0,
1540 	    ("pv_remove_all: illegal for unmanaged page %#x",
1541 	    VM_PAGE_TO_PHYS(m)));
1542 
1543 	pvo_head = vm_page_to_pvoh(m);
1544 	for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1545 		next_pvo = LIST_NEXT(pvo, pvo_vlink);
1546 
1547 		PMAP_PVO_CHECK(pvo);	/* sanity check */
1548 		pmap_pvo_remove(pvo, -1);
1549 	}
1550 	vm_page_flag_clear(m, PG_WRITEABLE);
1551 }
1552 
1553 /*
1554  * Remove all pages from specified address space, this aids process exit
1555  * speeds.  This is much faster than pmap_remove in the case of running down
1556  * an entire address space.  Only works for the current pmap.
1557  */
1558 void
1559 pmap_remove_pages(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1560 {
1561 
1562 	KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1563 	    ("pmap_remove_pages: non current pmap"));
1564 	pmap_remove(pm, sva, eva);
1565 }
1566 
1567 /*
1568  * Allocate a physical page of memory directly from the phys_avail map.
1569  * Can only be called from pmap_bootstrap before avail start and end are
1570  * calculated.
1571  */
1572 static vm_offset_t
1573 pmap_bootstrap_alloc(vm_size_t size, u_int align)
1574 {
1575 	vm_offset_t	s, e;
1576 	int		i, j;
1577 
1578 	size = round_page(size);
1579 	for (i = 0; phys_avail[i + 1] != 0; i += 2) {
1580 		if (align != 0)
1581 			s = (phys_avail[i] + align - 1) & ~(align - 1);
1582 		else
1583 			s = phys_avail[i];
1584 		e = s + size;
1585 
1586 		if (s < phys_avail[i] || e > phys_avail[i + 1])
1587 			continue;
1588 
1589 		if (s == phys_avail[i]) {
1590 			phys_avail[i] += size;
1591 		} else if (e == phys_avail[i + 1]) {
1592 			phys_avail[i + 1] -= size;
1593 		} else {
1594 			for (j = phys_avail_count * 2; j > i; j -= 2) {
1595 				phys_avail[j] = phys_avail[j - 2];
1596 				phys_avail[j + 1] = phys_avail[j - 1];
1597 			}
1598 
1599 			phys_avail[i + 3] = phys_avail[i + 1];
1600 			phys_avail[i + 1] = s;
1601 			phys_avail[i + 2] = e;
1602 			phys_avail_count++;
1603 		}
1604 
1605 		return (s);
1606 	}
1607 	panic("pmap_bootstrap_alloc: could not allocate memory");
1608 }
1609 
1610 /*
1611  * Return an unmapped pvo for a kernel virtual address.
1612  * Used by pmap functions that operate on physical pages.
1613  */
1614 static struct pvo_entry *
1615 pmap_rkva_alloc(void)
1616 {
1617 	struct		pvo_entry *pvo;
1618 	struct		pte *pt;
1619 	vm_offset_t	kva;
1620 	int		pteidx;
1621 
1622 	if (pmap_rkva_count == 0)
1623 		panic("pmap_rkva_alloc: no more reserved KVAs");
1624 
1625 	kva = pmap_rkva_start + (PAGE_SIZE * --pmap_rkva_count);
1626 	pmap_kenter(kva, 0);
1627 
1628 	pvo = pmap_pvo_find_va(kernel_pmap, kva, &pteidx);
1629 
1630 	if (pvo == NULL)
1631 		panic("pmap_kva_alloc: pmap_pvo_find_va failed");
1632 
1633 	pt = pmap_pvo_to_pte(pvo, pteidx);
1634 
1635 	if (pt == NULL)
1636 		panic("pmap_kva_alloc: pmap_pvo_to_pte failed");
1637 
1638 	pmap_pte_unset(pt, &pvo->pvo_pte, pvo->pvo_vaddr);
1639 	PVO_PTEGIDX_CLR(pvo);
1640 
1641 	pmap_pte_overflow++;
1642 
1643 	return (pvo);
1644 }
1645 
1646 static void
1647 pmap_pa_map(struct pvo_entry *pvo, vm_offset_t pa, struct pte *saved_pt,
1648     int *depth_p)
1649 {
1650 	struct	pte *pt;
1651 
1652 	/*
1653 	 * If this pvo already has a valid pte, we need to save it so it can
1654 	 * be restored later.  We then just reload the new PTE over the old
1655 	 * slot.
1656 	 */
1657 	if (saved_pt != NULL) {
1658 		pt = pmap_pvo_to_pte(pvo, -1);
1659 
1660 		if (pt != NULL) {
1661 			pmap_pte_unset(pt, &pvo->pvo_pte, pvo->pvo_vaddr);
1662 			PVO_PTEGIDX_CLR(pvo);
1663 			pmap_pte_overflow++;
1664 		}
1665 
1666 		*saved_pt = pvo->pvo_pte;
1667 
1668 		pvo->pvo_pte.pte_lo &= ~PTE_RPGN;
1669 	}
1670 
1671 	pvo->pvo_pte.pte_lo |= pa;
1672 
1673 	if (!pmap_pte_spill(pvo->pvo_vaddr))
1674 		panic("pmap_pa_map: could not spill pvo %p", pvo);
1675 
1676 	if (depth_p != NULL)
1677 		(*depth_p)++;
1678 }
1679 
1680 static void
1681 pmap_pa_unmap(struct pvo_entry *pvo, struct pte *saved_pt, int *depth_p)
1682 {
1683 	struct	pte *pt;
1684 
1685 	pt = pmap_pvo_to_pte(pvo, -1);
1686 
1687 	if (pt != NULL) {
1688 		pmap_pte_unset(pt, &pvo->pvo_pte, pvo->pvo_vaddr);
1689 		PVO_PTEGIDX_CLR(pvo);
1690 		pmap_pte_overflow++;
1691 	}
1692 
1693 	pvo->pvo_pte.pte_lo &= ~PTE_RPGN;
1694 
1695 	/*
1696 	 * If there is a saved PTE and it's valid, restore it and return.
1697 	 */
1698 	if (saved_pt != NULL && (saved_pt->pte_lo & PTE_RPGN) != 0) {
1699 		if (depth_p != NULL && --(*depth_p) == 0)
1700 			panic("pmap_pa_unmap: restoring but depth == 0");
1701 
1702 		pvo->pvo_pte = *saved_pt;
1703 
1704 		if (!pmap_pte_spill(pvo->pvo_vaddr))
1705 			panic("pmap_pa_unmap: could not spill pvo %p", pvo);
1706 	}
1707 }
1708 
1709 static void
1710 pmap_syncicache(vm_offset_t pa, vm_size_t len)
1711 {
1712 	__syncicache((void *)pa, len);
1713 }
1714 
1715 static void
1716 tlbia(void)
1717 {
1718 	caddr_t	i;
1719 
1720 	SYNC();
1721 	for (i = 0; i < (caddr_t)0x00040000; i += 0x00001000) {
1722 		TLBIE(i);
1723 		EIEIO();
1724 	}
1725 	TLBSYNC();
1726 	SYNC();
1727 }
1728 
1729 static int
1730 pmap_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
1731     vm_offset_t va, vm_offset_t pa, u_int pte_lo, int flags)
1732 {
1733 	struct	pvo_entry *pvo;
1734 	u_int	sr;
1735 	int	first;
1736 	u_int	ptegidx;
1737 	int	i;
1738 	int     bootstrap;
1739 
1740 	pmap_pvo_enter_calls++;
1741 	first = 0;
1742 
1743 	bootstrap = 0;
1744 
1745 	/*
1746 	 * Compute the PTE Group index.
1747 	 */
1748 	va &= ~ADDR_POFF;
1749 	sr = va_to_sr(pm->pm_sr, va);
1750 	ptegidx = va_to_pteg(sr, va);
1751 
1752 	/*
1753 	 * Remove any existing mapping for this page.  Reuse the pvo entry if
1754 	 * there is a mapping.
1755 	 */
1756 	LIST_FOREACH(pvo, &pmap_pvo_table[ptegidx], pvo_olink) {
1757 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
1758 			if ((pvo->pvo_pte.pte_lo & PTE_RPGN) == pa &&
1759 			    (pvo->pvo_pte.pte_lo & PTE_PP) ==
1760 			    (pte_lo & PTE_PP)) {
1761 				return (0);
1762 			}
1763 			pmap_pvo_remove(pvo, -1);
1764 			break;
1765 		}
1766 	}
1767 
1768 	/*
1769 	 * If we aren't overwriting a mapping, try to allocate.
1770 	 */
1771 	if (pmap_initialized) {
1772 		pvo = uma_zalloc(zone, M_NOWAIT);
1773 	} else {
1774 		if (pmap_bpvo_pool_index >= BPVO_POOL_SIZE) {
1775 			panic("pmap_enter: bpvo pool exhausted, %d, %d, %d",
1776 			      pmap_bpvo_pool_index, BPVO_POOL_SIZE,
1777 			      BPVO_POOL_SIZE * sizeof(struct pvo_entry));
1778 		}
1779 		pvo = &pmap_bpvo_pool[pmap_bpvo_pool_index];
1780 		pmap_bpvo_pool_index++;
1781 		bootstrap = 1;
1782 	}
1783 
1784 	if (pvo == NULL) {
1785 		return (ENOMEM);
1786 	}
1787 
1788 	pmap_pvo_entries++;
1789 	pvo->pvo_vaddr = va;
1790 	pvo->pvo_pmap = pm;
1791 	LIST_INSERT_HEAD(&pmap_pvo_table[ptegidx], pvo, pvo_olink);
1792 	pvo->pvo_vaddr &= ~ADDR_POFF;
1793 	if (flags & VM_PROT_EXECUTE)
1794 		pvo->pvo_vaddr |= PVO_EXECUTABLE;
1795 	if (flags & PVO_WIRED)
1796 		pvo->pvo_vaddr |= PVO_WIRED;
1797 	if (pvo_head != &pmap_pvo_kunmanaged)
1798 		pvo->pvo_vaddr |= PVO_MANAGED;
1799 	if (bootstrap)
1800 		pvo->pvo_vaddr |= PVO_BOOTSTRAP;
1801 	pmap_pte_create(&pvo->pvo_pte, sr, va, pa | pte_lo);
1802 
1803 	/*
1804 	 * Remember if the list was empty and therefore will be the first
1805 	 * item.
1806 	 */
1807 	if (LIST_FIRST(pvo_head) == NULL)
1808 		first = 1;
1809 
1810 	LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
1811 	if (pvo->pvo_pte.pte_lo & PVO_WIRED)
1812 		pvo->pvo_pmap->pm_stats.wired_count++;
1813 	pvo->pvo_pmap->pm_stats.resident_count++;
1814 
1815 	/*
1816 	 * We hope this succeeds but it isn't required.
1817 	 */
1818 	i = pmap_pte_insert(ptegidx, &pvo->pvo_pte);
1819 	if (i >= 0) {
1820 		PVO_PTEGIDX_SET(pvo, i);
1821 	} else {
1822 		panic("pmap_pvo_enter: overflow");
1823 		pmap_pte_overflow++;
1824 	}
1825 
1826 	return (first ? ENOENT : 0);
1827 }
1828 
1829 static void
1830 pmap_pvo_remove(struct pvo_entry *pvo, int pteidx)
1831 {
1832 	struct	pte *pt;
1833 
1834 	/*
1835 	 * If there is an active pte entry, we need to deactivate it (and
1836 	 * save the ref & cfg bits).
1837 	 */
1838 	pt = pmap_pvo_to_pte(pvo, pteidx);
1839 	if (pt != NULL) {
1840 		pmap_pte_unset(pt, &pvo->pvo_pte, pvo->pvo_vaddr);
1841 		PVO_PTEGIDX_CLR(pvo);
1842 	} else {
1843 		pmap_pte_overflow--;
1844 	}
1845 
1846 	/*
1847 	 * Update our statistics.
1848 	 */
1849 	pvo->pvo_pmap->pm_stats.resident_count--;
1850 	if (pvo->pvo_pte.pte_lo & PVO_WIRED)
1851 		pvo->pvo_pmap->pm_stats.wired_count--;
1852 
1853 	/*
1854 	 * Save the REF/CHG bits into their cache if the page is managed.
1855 	 */
1856 	if (pvo->pvo_vaddr & PVO_MANAGED) {
1857 		struct	vm_page *pg;
1858 
1859 		pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte_lo & PTE_RPGN);
1860 		if (pg != NULL) {
1861 			pmap_attr_save(pg, pvo->pvo_pte.pte_lo &
1862 			    (PTE_REF | PTE_CHG));
1863 		}
1864 	}
1865 
1866 	/*
1867 	 * Remove this PVO from the PV list.
1868 	 */
1869 	LIST_REMOVE(pvo, pvo_vlink);
1870 
1871 	/*
1872 	 * Remove this from the overflow list and return it to the pool
1873 	 * if we aren't going to reuse it.
1874 	 */
1875 	LIST_REMOVE(pvo, pvo_olink);
1876 	if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
1877 		uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? pmap_mpvo_zone :
1878 		    pmap_upvo_zone, pvo);
1879 	pmap_pvo_entries--;
1880 	pmap_pvo_remove_calls++;
1881 }
1882 
1883 static __inline int
1884 pmap_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
1885 {
1886 	int	pteidx;
1887 
1888 	/*
1889 	 * We can find the actual pte entry without searching by grabbing
1890 	 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
1891 	 * noticing the HID bit.
1892 	 */
1893 	pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
1894 	if (pvo->pvo_pte.pte_hi & PTE_HID)
1895 		pteidx ^= pmap_pteg_mask * 8;
1896 
1897 	return (pteidx);
1898 }
1899 
1900 static struct pvo_entry *
1901 pmap_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
1902 {
1903 	struct	pvo_entry *pvo;
1904 	int	ptegidx;
1905 	u_int	sr;
1906 
1907 	va &= ~ADDR_POFF;
1908 	sr = va_to_sr(pm->pm_sr, va);
1909 	ptegidx = va_to_pteg(sr, va);
1910 
1911 	LIST_FOREACH(pvo, &pmap_pvo_table[ptegidx], pvo_olink) {
1912 		if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
1913 			if (pteidx_p)
1914 				*pteidx_p = pmap_pvo_pte_index(pvo, ptegidx);
1915 			return (pvo);
1916 		}
1917 	}
1918 
1919 	return (NULL);
1920 }
1921 
1922 static struct pte *
1923 pmap_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
1924 {
1925 	struct	pte *pt;
1926 
1927 	/*
1928 	 * If we haven't been supplied the ptegidx, calculate it.
1929 	 */
1930 	if (pteidx == -1) {
1931 		int	ptegidx;
1932 		u_int	sr;
1933 
1934 		sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
1935 		ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
1936 		pteidx = pmap_pvo_pte_index(pvo, ptegidx);
1937 	}
1938 
1939 	pt = &pmap_pteg_table[pteidx >> 3].pt[pteidx & 7];
1940 
1941 	if ((pvo->pvo_pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
1942 		panic("pmap_pvo_to_pte: pvo %p has valid pte in pvo but no "
1943 		    "valid pte index", pvo);
1944 	}
1945 
1946 	if ((pvo->pvo_pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
1947 		panic("pmap_pvo_to_pte: pvo %p has valid pte index in pvo "
1948 		    "pvo but no valid pte", pvo);
1949 	}
1950 
1951 	if ((pt->pte_hi ^ (pvo->pvo_pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
1952 		if ((pvo->pvo_pte.pte_hi & PTE_VALID) == 0) {
1953 			panic("pmap_pvo_to_pte: pvo %p has valid pte in "
1954 			    "pmap_pteg_table %p but invalid in pvo", pvo, pt);
1955 		}
1956 
1957 		if (((pt->pte_lo ^ pvo->pvo_pte.pte_lo) & ~(PTE_CHG|PTE_REF))
1958 		    != 0) {
1959 			panic("pmap_pvo_to_pte: pvo %p pte does not match "
1960 			    "pte %p in pmap_pteg_table", pvo, pt);
1961 		}
1962 
1963 		return (pt);
1964 	}
1965 
1966 	if (pvo->pvo_pte.pte_hi & PTE_VALID) {
1967 		panic("pmap_pvo_to_pte: pvo %p has invalid pte %p in "
1968 		    "pmap_pteg_table but valid in pvo", pvo, pt);
1969 	}
1970 
1971 	return (NULL);
1972 }
1973 
1974 static void *
1975 pmap_pvo_allocf(uma_zone_t zone, int bytes, u_int8_t *flags, int wait)
1976 {
1977 	static vm_pindex_t color;
1978 	vm_page_t	m;
1979 
1980 	if (bytes != PAGE_SIZE)
1981 		panic("pmap_pvo_allocf: benno was shortsighted.  hit him.");
1982 
1983 	*flags = UMA_SLAB_PRIV;
1984 	/*
1985 	 * The color is only a hint.  Thus, a data race in the read-
1986 	 * modify-write operation below isn't a catastrophe.
1987 	 */
1988 	m = vm_page_alloc(NULL, color++, VM_ALLOC_NOOBJ | VM_ALLOC_SYSTEM);
1989 	if (m == NULL)
1990 		return (NULL);
1991 	return ((void *)VM_PAGE_TO_PHYS(m));
1992 }
1993 
1994 /*
1995  * XXX: THIS STUFF SHOULD BE IN pte.c?
1996  */
1997 int
1998 pmap_pte_spill(vm_offset_t addr)
1999 {
2000 	struct	pvo_entry *source_pvo, *victim_pvo;
2001 	struct	pvo_entry *pvo;
2002 	int	ptegidx, i, j;
2003 	u_int	sr;
2004 	struct	pteg *pteg;
2005 	struct	pte *pt;
2006 
2007 	pmap_pte_spills++;
2008 
2009 	sr = mfsrin(addr);
2010 	ptegidx = va_to_pteg(sr, addr);
2011 
2012 	/*
2013 	 * Have to substitute some entry.  Use the primary hash for this.
2014 	 * Use low bits of timebase as random generator.
2015 	 */
2016 	pteg = &pmap_pteg_table[ptegidx];
2017 	__asm __volatile("mftb %0" : "=r"(i));
2018 	i &= 7;
2019 	pt = &pteg->pt[i];
2020 
2021 	source_pvo = NULL;
2022 	victim_pvo = NULL;
2023 	LIST_FOREACH(pvo, &pmap_pvo_table[ptegidx], pvo_olink) {
2024 		/*
2025 		 * We need to find a pvo entry for this address.
2026 		 */
2027 		PMAP_PVO_CHECK(pvo);
2028 		if (source_pvo == NULL &&
2029 		    pmap_pte_match(&pvo->pvo_pte, sr, addr,
2030 		    pvo->pvo_pte.pte_hi & PTE_HID)) {
2031 			/*
2032 			 * Now found an entry to be spilled into the pteg.
2033 			 * The PTE is now valid, so we know it's active.
2034 			 */
2035 			j = pmap_pte_insert(ptegidx, &pvo->pvo_pte);
2036 
2037 			if (j >= 0) {
2038 				PVO_PTEGIDX_SET(pvo, j);
2039 				pmap_pte_overflow--;
2040 				PMAP_PVO_CHECK(pvo);
2041 				return (1);
2042 			}
2043 
2044 			source_pvo = pvo;
2045 
2046 			if (victim_pvo != NULL)
2047 				break;
2048 		}
2049 
2050 		/*
2051 		 * We also need the pvo entry of the victim we are replacing
2052 		 * so save the R & C bits of the PTE.
2053 		 */
2054 		if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2055 		    pmap_pte_compare(pt, &pvo->pvo_pte)) {
2056 			victim_pvo = pvo;
2057 			if (source_pvo != NULL)
2058 				break;
2059 		}
2060 	}
2061 
2062 	if (source_pvo == NULL)
2063 		return (0);
2064 
2065 	if (victim_pvo == NULL) {
2066 		if ((pt->pte_hi & PTE_HID) == 0)
2067 			panic("pmap_pte_spill: victim p-pte (%p) has no pvo"
2068 			    "entry", pt);
2069 
2070 		/*
2071 		 * If this is a secondary PTE, we need to search it's primary
2072 		 * pvo bucket for the matching PVO.
2073 		 */
2074 		LIST_FOREACH(pvo, &pmap_pvo_table[ptegidx ^ pmap_pteg_mask],
2075 		    pvo_olink) {
2076 			PMAP_PVO_CHECK(pvo);
2077 			/*
2078 			 * We also need the pvo entry of the victim we are
2079 			 * replacing so save the R & C bits of the PTE.
2080 			 */
2081 			if (pmap_pte_compare(pt, &pvo->pvo_pte)) {
2082 				victim_pvo = pvo;
2083 				break;
2084 			}
2085 		}
2086 
2087 		if (victim_pvo == NULL)
2088 			panic("pmap_pte_spill: victim s-pte (%p) has no pvo"
2089 			    "entry", pt);
2090 	}
2091 
2092 	/*
2093 	 * We are invalidating the TLB entry for the EA we are replacing even
2094 	 * though it's valid.  If we don't, we lose any ref/chg bit changes
2095 	 * contained in the TLB entry.
2096 	 */
2097 	source_pvo->pvo_pte.pte_hi &= ~PTE_HID;
2098 
2099 	pmap_pte_unset(pt, &victim_pvo->pvo_pte, victim_pvo->pvo_vaddr);
2100 	pmap_pte_set(pt, &source_pvo->pvo_pte);
2101 
2102 	PVO_PTEGIDX_CLR(victim_pvo);
2103 	PVO_PTEGIDX_SET(source_pvo, i);
2104 	pmap_pte_replacements++;
2105 
2106 	PMAP_PVO_CHECK(victim_pvo);
2107 	PMAP_PVO_CHECK(source_pvo);
2108 
2109 	return (1);
2110 }
2111 
2112 static int
2113 pmap_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2114 {
2115 	struct	pte *pt;
2116 	int	i;
2117 
2118 	/*
2119 	 * First try primary hash.
2120 	 */
2121 	for (pt = pmap_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2122 		if ((pt->pte_hi & PTE_VALID) == 0) {
2123 			pvo_pt->pte_hi &= ~PTE_HID;
2124 			pmap_pte_set(pt, pvo_pt);
2125 			return (i);
2126 		}
2127 	}
2128 
2129 	/*
2130 	 * Now try secondary hash.
2131 	 */
2132 	ptegidx ^= pmap_pteg_mask;
2133 	ptegidx++;
2134 	for (pt = pmap_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2135 		if ((pt->pte_hi & PTE_VALID) == 0) {
2136 			pvo_pt->pte_hi |= PTE_HID;
2137 			pmap_pte_set(pt, pvo_pt);
2138 			return (i);
2139 		}
2140 	}
2141 
2142 	panic("pmap_pte_insert: overflow");
2143 	return (-1);
2144 }
2145 
2146 static boolean_t
2147 pmap_query_bit(vm_page_t m, int ptebit)
2148 {
2149 	struct	pvo_entry *pvo;
2150 	struct	pte *pt;
2151 
2152 	if (pmap_attr_fetch(m) & ptebit)
2153 		return (TRUE);
2154 
2155 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2156 		PMAP_PVO_CHECK(pvo);	/* sanity check */
2157 
2158 		/*
2159 		 * See if we saved the bit off.  If so, cache it and return
2160 		 * success.
2161 		 */
2162 		if (pvo->pvo_pte.pte_lo & ptebit) {
2163 			pmap_attr_save(m, ptebit);
2164 			PMAP_PVO_CHECK(pvo);	/* sanity check */
2165 			return (TRUE);
2166 		}
2167 	}
2168 
2169 	/*
2170 	 * No luck, now go through the hard part of looking at the PTEs
2171 	 * themselves.  Sync so that any pending REF/CHG bits are flushed to
2172 	 * the PTEs.
2173 	 */
2174 	SYNC();
2175 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2176 		PMAP_PVO_CHECK(pvo);	/* sanity check */
2177 
2178 		/*
2179 		 * See if this pvo has a valid PTE.  if so, fetch the
2180 		 * REF/CHG bits from the valid PTE.  If the appropriate
2181 		 * ptebit is set, cache it and return success.
2182 		 */
2183 		pt = pmap_pvo_to_pte(pvo, -1);
2184 		if (pt != NULL) {
2185 			pmap_pte_synch(pt, &pvo->pvo_pte);
2186 			if (pvo->pvo_pte.pte_lo & ptebit) {
2187 				pmap_attr_save(m, ptebit);
2188 				PMAP_PVO_CHECK(pvo);	/* sanity check */
2189 				return (TRUE);
2190 			}
2191 		}
2192 	}
2193 
2194 	return (TRUE);
2195 }
2196 
2197 static u_int
2198 pmap_clear_bit(vm_page_t m, int ptebit, int *origbit)
2199 {
2200 	u_int	count;
2201 	struct	pvo_entry *pvo;
2202 	struct	pte *pt;
2203 	int	rv;
2204 
2205 	/*
2206 	 * Clear the cached value.
2207 	 */
2208 	rv = pmap_attr_fetch(m);
2209 	pmap_attr_clear(m, ptebit);
2210 
2211 	/*
2212 	 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2213 	 * we can reset the right ones).  note that since the pvo entries and
2214 	 * list heads are accessed via BAT0 and are never placed in the page
2215 	 * table, we don't have to worry about further accesses setting the
2216 	 * REF/CHG bits.
2217 	 */
2218 	SYNC();
2219 
2220 	/*
2221 	 * For each pvo entry, clear the pvo's ptebit.  If this pvo has a
2222 	 * valid pte clear the ptebit from the valid pte.
2223 	 */
2224 	count = 0;
2225 	LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2226 		PMAP_PVO_CHECK(pvo);	/* sanity check */
2227 		pt = pmap_pvo_to_pte(pvo, -1);
2228 		if (pt != NULL) {
2229 			pmap_pte_synch(pt, &pvo->pvo_pte);
2230 			if (pvo->pvo_pte.pte_lo & ptebit) {
2231 				count++;
2232 				pmap_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2233 			}
2234 		}
2235 		rv |= pvo->pvo_pte.pte_lo;
2236 		pvo->pvo_pte.pte_lo &= ~ptebit;
2237 		PMAP_PVO_CHECK(pvo);	/* sanity check */
2238 	}
2239 
2240 	if (origbit != NULL) {
2241 		*origbit = rv;
2242 	}
2243 
2244 	return (count);
2245 }
2246 
2247 /*
2248  * Return true if the physical range is encompassed by the battable[idx]
2249  */
2250 static int
2251 pmap_bat_mapped(int idx, vm_offset_t pa, vm_size_t size)
2252 {
2253 	u_int prot;
2254 	u_int32_t start;
2255 	u_int32_t end;
2256 	u_int32_t bat_ble;
2257 
2258 	/*
2259 	 * Return immediately if not a valid mapping
2260 	 */
2261 	if (!battable[idx].batu & BAT_Vs)
2262 		return (EINVAL);
2263 
2264 	/*
2265 	 * The BAT entry must be cache-inhibited, guarded, and r/w
2266 	 * so it can function as an i/o page
2267 	 */
2268 	prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2269 	if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2270 		return (EPERM);
2271 
2272 	/*
2273 	 * The address should be within the BAT range. Assume that the
2274 	 * start address in the BAT has the correct alignment (thus
2275 	 * not requiring masking)
2276 	 */
2277 	start = battable[idx].batl & BAT_PBS;
2278 	bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2279 	end = start | (bat_ble << 15) | 0x7fff;
2280 
2281 	if ((pa < start) || ((pa + size) > end))
2282 		return (ERANGE);
2283 
2284 	return (0);
2285 }
2286 
2287 
2288 /*
2289  * Map a set of physical memory pages into the kernel virtual
2290  * address space. Return a pointer to where it is mapped. This
2291  * routine is intended to be used for mapping device memory,
2292  * NOT real memory.
2293  */
2294 void *
2295 pmap_mapdev(vm_offset_t pa, vm_size_t size)
2296 {
2297 	vm_offset_t va, tmpva, ppa, offset;
2298 	int i;
2299 
2300 	ppa = trunc_page(pa);
2301 	offset = pa & PAGE_MASK;
2302 	size = roundup(offset + size, PAGE_SIZE);
2303 
2304 	GIANT_REQUIRED;
2305 
2306 	/*
2307 	 * If the physical address lies within a valid BAT table entry,
2308 	 * return the 1:1 mapping. This currently doesn't work
2309 	 * for regions that overlap 256M BAT segments.
2310 	 */
2311 	for (i = 0; i < 16; i++) {
2312 		if (pmap_bat_mapped(i, pa, size) == 0)
2313 			return ((void *) pa);
2314 	}
2315 
2316 	va = kmem_alloc_nofault(kernel_map, size);
2317 	if (!va)
2318 		panic("pmap_mapdev: Couldn't alloc kernel virtual memory");
2319 
2320 	for (tmpva = va; size > 0;) {
2321 		pmap_kenter(tmpva, ppa);
2322 		TLBIE(tmpva); /* XXX or should it be invalidate-all ? */
2323 		size -= PAGE_SIZE;
2324 		tmpva += PAGE_SIZE;
2325 		ppa += PAGE_SIZE;
2326 	}
2327 
2328 	return ((void *)(va + offset));
2329 }
2330 
2331 void
2332 pmap_unmapdev(vm_offset_t va, vm_size_t size)
2333 {
2334 	vm_offset_t base, offset;
2335 
2336 	/*
2337 	 * If this is outside kernel virtual space, then it's a
2338 	 * battable entry and doesn't require unmapping
2339 	 */
2340 	if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= VM_MAX_KERNEL_ADDRESS)) {
2341 		base = trunc_page(va);
2342 		offset = va & PAGE_MASK;
2343 		size = roundup(offset + size, PAGE_SIZE);
2344 		kmem_free(kernel_map, base, size);
2345 	}
2346 }
2347