1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause AND BSD-4-Clause
3 *
4 * Copyright (c) 2001 The NetBSD Foundation, Inc.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The NetBSD Foundation
8 * by Matt Thomas <matt@3am-software.com> of Allegro Networks, Inc.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29 * POSSIBILITY OF SUCH DAMAGE.
30 */
31 /*-
32 * Copyright (C) 1995, 1996 Wolfgang Solfrank.
33 * Copyright (C) 1995, 1996 TooLs GmbH.
34 * All rights reserved.
35 *
36 * Redistribution and use in source and binary forms, with or without
37 * modification, are permitted provided that the following conditions
38 * are met:
39 * 1. Redistributions of source code must retain the above copyright
40 * notice, this list of conditions and the following disclaimer.
41 * 2. Redistributions in binary form must reproduce the above copyright
42 * notice, this list of conditions and the following disclaimer in the
43 * documentation and/or other materials provided with the distribution.
44 * 3. All advertising materials mentioning features or use of this software
45 * must display the following acknowledgement:
46 * This product includes software developed by TooLs GmbH.
47 * 4. The name of TooLs GmbH may not be used to endorse or promote products
48 * derived from this software without specific prior written permission.
49 *
50 * THIS SOFTWARE IS PROVIDED BY TOOLS GMBH ``AS IS'' AND ANY EXPRESS OR
51 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
52 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
53 * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
54 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
55 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
56 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
57 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
58 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
59 * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
60 *
61 * $NetBSD: pmap.c,v 1.28 2000/03/26 20:42:36 kleink Exp $
62 */
63 /*-
64 * Copyright (C) 2001 Benno Rice.
65 * All rights reserved.
66 *
67 * Redistribution and use in source and binary forms, with or without
68 * modification, are permitted provided that the following conditions
69 * are met:
70 * 1. Redistributions of source code must retain the above copyright
71 * notice, this list of conditions and the following disclaimer.
72 * 2. Redistributions in binary form must reproduce the above copyright
73 * notice, this list of conditions and the following disclaimer in the
74 * documentation and/or other materials provided with the distribution.
75 *
76 * THIS SOFTWARE IS PROVIDED BY Benno Rice ``AS IS'' AND ANY EXPRESS OR
77 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
78 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
79 * IN NO EVENT SHALL TOOLS GMBH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
80 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
81 * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
82 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
83 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
84 * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF
85 * ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
86 */
87
88 #include <sys/cdefs.h>
89 /*
90 * Manages physical address maps.
91 *
92 * Since the information managed by this module is also stored by the
93 * logical address mapping module, this module may throw away valid virtual
94 * to physical mappings at almost any time. However, invalidations of
95 * mappings must be done as requested.
96 *
97 * In order to cope with hardware architectures which make virtual to
98 * physical map invalidates expensive, this module may delay invalidate
99 * reduced protection operations until such time as they are actually
100 * necessary. This module is given full information as to which processors
101 * are currently using which maps, and to when physical maps must be made
102 * correct.
103 */
104
105 #include "opt_kstack_pages.h"
106
107 #include <sys/param.h>
108 #include <sys/kernel.h>
109 #include <sys/conf.h>
110 #include <sys/queue.h>
111 #include <sys/cpuset.h>
112 #include <sys/kerneldump.h>
113 #include <sys/ktr.h>
114 #include <sys/lock.h>
115 #include <sys/mman.h>
116 #include <sys/msgbuf.h>
117 #include <sys/mutex.h>
118 #include <sys/proc.h>
119 #include <sys/rwlock.h>
120 #include <sys/sched.h>
121 #include <sys/sysctl.h>
122 #include <sys/systm.h>
123 #include <sys/vmmeter.h>
124
125 #include <dev/ofw/openfirm.h>
126
127 #include <vm/vm.h>
128 #include <vm/pmap.h>
129 #include <vm/vm_param.h>
130 #include <vm/vm_kern.h>
131 #include <vm/vm_page.h>
132 #include <vm/vm_map.h>
133 #include <vm/vm_object.h>
134 #include <vm/vm_extern.h>
135 #include <vm/vm_page.h>
136 #include <vm/vm_phys.h>
137 #include <vm/vm_pageout.h>
138 #include <vm/vm_radix.h>
139 #include <vm/uma.h>
140
141 #include <machine/cpu.h>
142 #include <machine/platform.h>
143 #include <machine/bat.h>
144 #include <machine/frame.h>
145 #include <machine/md_var.h>
146 #include <machine/psl.h>
147 #include <machine/pte.h>
148 #include <machine/smp.h>
149 #include <machine/sr.h>
150 #include <machine/mmuvar.h>
151 #include <machine/trap.h>
152
153 #define MOEA_DEBUG
154
155 #define TODO panic("%s: not implemented", __func__);
156
157 #define VSID_MAKE(sr, hash) ((sr) | (((hash) & 0xfffff) << 4))
158 #define VSID_TO_SR(vsid) ((vsid) & 0xf)
159 #define VSID_TO_HASH(vsid) (((vsid) >> 4) & 0xfffff)
160
161 /* Get physical address from PVO. */
162 #define PVO_PADDR(pvo) ((pvo)->pvo_pte.pte.pte_lo & PTE_RPGN)
163
164 struct ofw_map {
165 vm_offset_t om_va;
166 vm_size_t om_len;
167 vm_offset_t om_pa;
168 u_int om_mode;
169 };
170
171 extern unsigned char _etext[];
172 extern unsigned char _end[];
173
174 /*
175 * Map of physical memory regions.
176 */
177 static struct mem_region *regions;
178 static struct mem_region *pregions;
179 static u_int phys_avail_count;
180 static int regions_sz, pregions_sz;
181 static struct ofw_map *translations;
182
183 /*
184 * Lock for the pteg and pvo tables.
185 */
186 struct mtx moea_table_mutex;
187 struct mtx moea_vsid_mutex;
188
189 /* tlbie instruction synchronization */
190 static struct mtx tlbie_mtx;
191
192 /*
193 * PTEG data.
194 */
195 static struct pteg *moea_pteg_table;
196 u_int moea_pteg_count;
197 u_int moea_pteg_mask;
198
199 /*
200 * PVO data.
201 */
202 struct pvo_head *moea_pvo_table; /* pvo entries by pteg index */
203 struct pvo_head moea_pvo_kunmanaged =
204 LIST_HEAD_INITIALIZER(moea_pvo_kunmanaged); /* list of unmanaged pages */
205
206 static struct rwlock_padalign pvh_global_lock;
207
208 uma_zone_t moea_upvo_zone; /* zone for pvo entries for unmanaged pages */
209 uma_zone_t moea_mpvo_zone; /* zone for pvo entries for managed pages */
210
211 #define BPVO_POOL_SIZE 32768
212 static struct pvo_entry *moea_bpvo_pool;
213 static int moea_bpvo_pool_index = 0;
214
215 #define VSID_NBPW (sizeof(u_int32_t) * 8)
216 static u_int moea_vsid_bitmap[NPMAPS / VSID_NBPW];
217
218 static bool moea_initialized = false;
219
220 /*
221 * Statistics.
222 */
223 u_int moea_pte_valid = 0;
224 u_int moea_pte_overflow = 0;
225 u_int moea_pte_replacements = 0;
226 u_int moea_pvo_entries = 0;
227 u_int moea_pvo_enter_calls = 0;
228 u_int moea_pvo_remove_calls = 0;
229 u_int moea_pte_spills = 0;
230 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_valid, CTLFLAG_RD, &moea_pte_valid,
231 0, "");
232 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_overflow, CTLFLAG_RD,
233 &moea_pte_overflow, 0, "");
234 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_replacements, CTLFLAG_RD,
235 &moea_pte_replacements, 0, "");
236 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_entries, CTLFLAG_RD, &moea_pvo_entries,
237 0, "");
238 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_enter_calls, CTLFLAG_RD,
239 &moea_pvo_enter_calls, 0, "");
240 SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_remove_calls, CTLFLAG_RD,
241 &moea_pvo_remove_calls, 0, "");
242 SYSCTL_INT(_machdep, OID_AUTO, moea_pte_spills, CTLFLAG_RD,
243 &moea_pte_spills, 0, "");
244
245 /*
246 * Allocate physical memory for use in moea_bootstrap.
247 */
248 static vm_offset_t moea_bootstrap_alloc(vm_size_t, u_int);
249
250 /*
251 * PTE calls.
252 */
253 static int moea_pte_insert(u_int, struct pte *);
254
255 /*
256 * PVO calls.
257 */
258 static int moea_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *,
259 vm_offset_t, vm_paddr_t, u_int, int);
260 static void moea_pvo_remove(struct pvo_entry *, int);
261 static struct pvo_entry *moea_pvo_find_va(pmap_t, vm_offset_t, int *);
262 static struct pte *moea_pvo_to_pte(const struct pvo_entry *, int);
263
264 /*
265 * Utility routines.
266 */
267 static int moea_enter_locked(pmap_t, vm_offset_t, vm_page_t,
268 vm_prot_t, u_int, int8_t);
269 static void moea_syncicache(vm_paddr_t, vm_size_t);
270 static bool moea_query_bit(vm_page_t, int);
271 static u_int moea_clear_bit(vm_page_t, int);
272 static void moea_kremove(vm_offset_t);
273 int moea_pte_spill(vm_offset_t);
274
275 /*
276 * Kernel MMU interface
277 */
278 void moea_clear_modify(vm_page_t);
279 void moea_copy_page(vm_page_t, vm_page_t);
280 void moea_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
281 vm_page_t *mb, vm_offset_t b_offset, int xfersize);
282 int moea_enter(pmap_t, vm_offset_t, vm_page_t, vm_prot_t, u_int,
283 int8_t);
284 void moea_enter_object(pmap_t, vm_offset_t, vm_offset_t, vm_page_t,
285 vm_prot_t);
286 void moea_enter_quick(pmap_t, vm_offset_t, vm_page_t, vm_prot_t);
287 vm_paddr_t moea_extract(pmap_t, vm_offset_t);
288 vm_page_t moea_extract_and_hold(pmap_t, vm_offset_t, vm_prot_t);
289 void moea_init(void);
290 bool moea_is_modified(vm_page_t);
291 bool moea_is_prefaultable(pmap_t, vm_offset_t);
292 bool moea_is_referenced(vm_page_t);
293 int moea_ts_referenced(vm_page_t);
294 vm_offset_t moea_map(vm_offset_t *, vm_paddr_t, vm_paddr_t, int);
295 static int moea_mincore(pmap_t, vm_offset_t, vm_paddr_t *);
296 bool moea_page_exists_quick(pmap_t, vm_page_t);
297 void moea_page_init(vm_page_t);
298 int moea_page_wired_mappings(vm_page_t);
299 int moea_pinit(pmap_t);
300 void moea_pinit0(pmap_t);
301 void moea_protect(pmap_t, vm_offset_t, vm_offset_t, vm_prot_t);
302 void moea_qenter(vm_offset_t, vm_page_t *, int);
303 void moea_qremove(vm_offset_t, int);
304 void moea_release(pmap_t);
305 void moea_remove(pmap_t, vm_offset_t, vm_offset_t);
306 void moea_remove_all(vm_page_t);
307 void moea_remove_write(vm_page_t);
308 void moea_unwire(pmap_t, vm_offset_t, vm_offset_t);
309 void moea_zero_page(vm_page_t);
310 void moea_zero_page_area(vm_page_t, int, int);
311 void moea_activate(struct thread *);
312 void moea_deactivate(struct thread *);
313 void moea_cpu_bootstrap(int);
314 void moea_bootstrap(vm_offset_t, vm_offset_t);
315 void *moea_mapdev(vm_paddr_t, vm_size_t);
316 void *moea_mapdev_attr(vm_paddr_t, vm_size_t, vm_memattr_t);
317 void moea_unmapdev(void *, vm_size_t);
318 vm_paddr_t moea_kextract(vm_offset_t);
319 void moea_kenter_attr(vm_offset_t, vm_paddr_t, vm_memattr_t);
320 void moea_kenter(vm_offset_t, vm_paddr_t);
321 void moea_page_set_memattr(vm_page_t m, vm_memattr_t ma);
322 int moea_dev_direct_mapped(vm_paddr_t, vm_size_t);
323 static void moea_sync_icache(pmap_t, vm_offset_t, vm_size_t);
324 void moea_dumpsys_map(vm_paddr_t pa, size_t sz, void **va);
325 void moea_scan_init(void);
326 vm_offset_t moea_quick_enter_page(vm_page_t m);
327 void moea_quick_remove_page(vm_offset_t addr);
328 bool moea_page_is_mapped(vm_page_t m);
329 bool moea_ps_enabled(pmap_t pmap);
330 static int moea_map_user_ptr(pmap_t pm,
331 volatile const void *uaddr, void **kaddr, size_t ulen, size_t *klen);
332 static int moea_decode_kernel_ptr(vm_offset_t addr,
333 int *is_user, vm_offset_t *decoded_addr);
334
335 static struct pmap_funcs moea_methods = {
336 .clear_modify = moea_clear_modify,
337 .copy_page = moea_copy_page,
338 .copy_pages = moea_copy_pages,
339 .enter = moea_enter,
340 .enter_object = moea_enter_object,
341 .enter_quick = moea_enter_quick,
342 .extract = moea_extract,
343 .extract_and_hold = moea_extract_and_hold,
344 .init = moea_init,
345 .is_modified = moea_is_modified,
346 .is_prefaultable = moea_is_prefaultable,
347 .is_referenced = moea_is_referenced,
348 .ts_referenced = moea_ts_referenced,
349 .map = moea_map,
350 .page_exists_quick = moea_page_exists_quick,
351 .page_init = moea_page_init,
352 .page_wired_mappings = moea_page_wired_mappings,
353 .pinit = moea_pinit,
354 .pinit0 = moea_pinit0,
355 .protect = moea_protect,
356 .qenter = moea_qenter,
357 .qremove = moea_qremove,
358 .release = moea_release,
359 .remove = moea_remove,
360 .remove_all = moea_remove_all,
361 .mincore = moea_mincore,
362 .remove_write = moea_remove_write,
363 .sync_icache = moea_sync_icache,
364 .unwire = moea_unwire,
365 .zero_page = moea_zero_page,
366 .zero_page_area = moea_zero_page_area,
367 .activate = moea_activate,
368 .deactivate = moea_deactivate,
369 .page_set_memattr = moea_page_set_memattr,
370 .quick_enter_page = moea_quick_enter_page,
371 .quick_remove_page = moea_quick_remove_page,
372 .page_is_mapped = moea_page_is_mapped,
373 .ps_enabled = moea_ps_enabled,
374
375 /* Internal interfaces */
376 .bootstrap = moea_bootstrap,
377 .cpu_bootstrap = moea_cpu_bootstrap,
378 .mapdev_attr = moea_mapdev_attr,
379 .mapdev = moea_mapdev,
380 .unmapdev = moea_unmapdev,
381 .kextract = moea_kextract,
382 .kenter = moea_kenter,
383 .kenter_attr = moea_kenter_attr,
384 .dev_direct_mapped = moea_dev_direct_mapped,
385 .dumpsys_pa_init = moea_scan_init,
386 .dumpsys_map_chunk = moea_dumpsys_map,
387 .map_user_ptr = moea_map_user_ptr,
388 .decode_kernel_ptr = moea_decode_kernel_ptr,
389 };
390
391 MMU_DEF(oea_mmu, MMU_TYPE_OEA, moea_methods);
392
393 static __inline uint32_t
moea_calc_wimg(vm_paddr_t pa,vm_memattr_t ma)394 moea_calc_wimg(vm_paddr_t pa, vm_memattr_t ma)
395 {
396 uint32_t pte_lo;
397 int i;
398
399 if (ma != VM_MEMATTR_DEFAULT) {
400 switch (ma) {
401 case VM_MEMATTR_UNCACHEABLE:
402 return (PTE_I | PTE_G);
403 case VM_MEMATTR_CACHEABLE:
404 return (PTE_M);
405 case VM_MEMATTR_WRITE_COMBINING:
406 case VM_MEMATTR_WRITE_BACK:
407 case VM_MEMATTR_PREFETCHABLE:
408 return (PTE_I);
409 case VM_MEMATTR_WRITE_THROUGH:
410 return (PTE_W | PTE_M);
411 }
412 }
413
414 /*
415 * Assume the page is cache inhibited and access is guarded unless
416 * it's in our available memory array.
417 */
418 pte_lo = PTE_I | PTE_G;
419 for (i = 0; i < pregions_sz; i++) {
420 if ((pa >= pregions[i].mr_start) &&
421 (pa < (pregions[i].mr_start + pregions[i].mr_size))) {
422 pte_lo = PTE_M;
423 break;
424 }
425 }
426
427 return pte_lo;
428 }
429
430 /*
431 * Translate OFW translations into VM attributes.
432 */
433 static __inline vm_memattr_t
moea_bootstrap_convert_wimg(uint32_t mode)434 moea_bootstrap_convert_wimg(uint32_t mode)
435 {
436
437 switch (mode) {
438 case (PTE_I | PTE_G):
439 /* PCI device memory */
440 return VM_MEMATTR_UNCACHEABLE;
441 case (PTE_M):
442 /* Explicitly coherent */
443 return VM_MEMATTR_CACHEABLE;
444 case 0: /* Default claim */
445 case 2: /* Alternate PP bits set by OF for the original payload */
446 /* "Normal" memory. */
447 return VM_MEMATTR_DEFAULT;
448
449 default:
450 /* Err on the side of caution for unknowns */
451 /* XXX should we panic instead? */
452 return VM_MEMATTR_UNCACHEABLE;
453 }
454 }
455
456 static void
tlbie(vm_offset_t va)457 tlbie(vm_offset_t va)
458 {
459
460 mtx_lock_spin(&tlbie_mtx);
461 __asm __volatile("ptesync");
462 __asm __volatile("tlbie %0" :: "r"(va));
463 __asm __volatile("eieio; tlbsync; ptesync");
464 mtx_unlock_spin(&tlbie_mtx);
465 }
466
467 static void
tlbia(void)468 tlbia(void)
469 {
470 vm_offset_t va;
471
472 for (va = 0; va < 0x00040000; va += 0x00001000) {
473 __asm __volatile("tlbie %0" :: "r"(va));
474 powerpc_sync();
475 }
476 __asm __volatile("tlbsync");
477 powerpc_sync();
478 }
479
480 static __inline int
va_to_sr(u_int * sr,vm_offset_t va)481 va_to_sr(u_int *sr, vm_offset_t va)
482 {
483 return (sr[(uintptr_t)va >> ADDR_SR_SHFT]);
484 }
485
486 static __inline u_int
va_to_pteg(u_int sr,vm_offset_t addr)487 va_to_pteg(u_int sr, vm_offset_t addr)
488 {
489 u_int hash;
490
491 hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >>
492 ADDR_PIDX_SHFT);
493 return (hash & moea_pteg_mask);
494 }
495
496 static __inline struct pvo_head *
vm_page_to_pvoh(vm_page_t m)497 vm_page_to_pvoh(vm_page_t m)
498 {
499
500 return (&m->md.mdpg_pvoh);
501 }
502
503 static __inline void
moea_attr_clear(vm_page_t m,int ptebit)504 moea_attr_clear(vm_page_t m, int ptebit)
505 {
506
507 rw_assert(&pvh_global_lock, RA_WLOCKED);
508 m->md.mdpg_attrs &= ~ptebit;
509 }
510
511 static __inline int
moea_attr_fetch(vm_page_t m)512 moea_attr_fetch(vm_page_t m)
513 {
514
515 return (m->md.mdpg_attrs);
516 }
517
518 static __inline void
moea_attr_save(vm_page_t m,int ptebit)519 moea_attr_save(vm_page_t m, int ptebit)
520 {
521
522 rw_assert(&pvh_global_lock, RA_WLOCKED);
523 m->md.mdpg_attrs |= ptebit;
524 }
525
526 static __inline int
moea_pte_compare(const struct pte * pt,const struct pte * pvo_pt)527 moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt)
528 {
529 if (pt->pte_hi == pvo_pt->pte_hi)
530 return (1);
531
532 return (0);
533 }
534
535 static __inline int
moea_pte_match(struct pte * pt,u_int sr,vm_offset_t va,int which)536 moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which)
537 {
538 return (pt->pte_hi & ~PTE_VALID) ==
539 (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
540 ((va >> ADDR_API_SHFT) & PTE_API) | which);
541 }
542
543 static __inline void
moea_pte_create(struct pte * pt,u_int sr,vm_offset_t va,u_int pte_lo)544 moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo)
545 {
546
547 mtx_assert(&moea_table_mutex, MA_OWNED);
548
549 /*
550 * Construct a PTE. Default to IMB initially. Valid bit only gets
551 * set when the real pte is set in memory.
552 *
553 * Note: Don't set the valid bit for correct operation of tlb update.
554 */
555 pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) |
556 (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API);
557 pt->pte_lo = pte_lo;
558 }
559
560 static __inline void
moea_pte_synch(struct pte * pt,struct pte * pvo_pt)561 moea_pte_synch(struct pte *pt, struct pte *pvo_pt)
562 {
563
564 mtx_assert(&moea_table_mutex, MA_OWNED);
565 pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG);
566 }
567
568 static __inline void
moea_pte_clear(struct pte * pt,vm_offset_t va,int ptebit)569 moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit)
570 {
571
572 mtx_assert(&moea_table_mutex, MA_OWNED);
573
574 /*
575 * As shown in Section 7.6.3.2.3
576 */
577 pt->pte_lo &= ~ptebit;
578 tlbie(va);
579 }
580
581 static __inline void
moea_pte_set(struct pte * pt,struct pte * pvo_pt)582 moea_pte_set(struct pte *pt, struct pte *pvo_pt)
583 {
584
585 mtx_assert(&moea_table_mutex, MA_OWNED);
586 pvo_pt->pte_hi |= PTE_VALID;
587
588 /*
589 * Update the PTE as defined in section 7.6.3.1.
590 * Note that the REF/CHG bits are from pvo_pt and thus should have
591 * been saved so this routine can restore them (if desired).
592 */
593 pt->pte_lo = pvo_pt->pte_lo;
594 powerpc_sync();
595 pt->pte_hi = pvo_pt->pte_hi;
596 powerpc_sync();
597 moea_pte_valid++;
598 }
599
600 static __inline void
moea_pte_unset(struct pte * pt,struct pte * pvo_pt,vm_offset_t va)601 moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
602 {
603
604 mtx_assert(&moea_table_mutex, MA_OWNED);
605 pvo_pt->pte_hi &= ~PTE_VALID;
606
607 /*
608 * Force the reg & chg bits back into the PTEs.
609 */
610 powerpc_sync();
611
612 /*
613 * Invalidate the pte.
614 */
615 pt->pte_hi &= ~PTE_VALID;
616
617 tlbie(va);
618
619 /*
620 * Save the reg & chg bits.
621 */
622 moea_pte_synch(pt, pvo_pt);
623 moea_pte_valid--;
624 }
625
626 static __inline void
moea_pte_change(struct pte * pt,struct pte * pvo_pt,vm_offset_t va)627 moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va)
628 {
629
630 /*
631 * Invalidate the PTE
632 */
633 moea_pte_unset(pt, pvo_pt, va);
634 moea_pte_set(pt, pvo_pt);
635 }
636
637 /*
638 * Quick sort callout for comparing memory regions.
639 */
640 static int om_cmp(const void *a, const void *b);
641
642 static int
om_cmp(const void * a,const void * b)643 om_cmp(const void *a, const void *b)
644 {
645 const struct ofw_map *mapa;
646 const struct ofw_map *mapb;
647
648 mapa = a;
649 mapb = b;
650 if (mapa->om_pa < mapb->om_pa)
651 return (-1);
652 else if (mapa->om_pa > mapb->om_pa)
653 return (1);
654 else
655 return (0);
656 }
657
658 void
moea_cpu_bootstrap(int ap)659 moea_cpu_bootstrap(int ap)
660 {
661 u_int sdr;
662 int i;
663
664 if (ap) {
665 powerpc_sync();
666 __asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu));
667 __asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl));
668 isync();
669 __asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu));
670 __asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl));
671 isync();
672 }
673
674 __asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
675 __asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
676 isync();
677
678 __asm __volatile("mtibatu 1,%0" :: "r"(0));
679 __asm __volatile("mtdbatu 2,%0" :: "r"(0));
680 __asm __volatile("mtibatu 2,%0" :: "r"(0));
681 __asm __volatile("mtdbatu 3,%0" :: "r"(0));
682 __asm __volatile("mtibatu 3,%0" :: "r"(0));
683 isync();
684
685 for (i = 0; i < 16; i++)
686 mtsrin(i << ADDR_SR_SHFT, kernel_pmap->pm_sr[i]);
687 powerpc_sync();
688
689 sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10);
690 __asm __volatile("mtsdr1 %0" :: "r"(sdr));
691 isync();
692
693 tlbia();
694 }
695
696 void
moea_bootstrap(vm_offset_t kernelstart,vm_offset_t kernelend)697 moea_bootstrap(vm_offset_t kernelstart, vm_offset_t kernelend)
698 {
699 ihandle_t mmui;
700 phandle_t chosen, mmu;
701 int sz;
702 int i, j;
703 vm_size_t size, physsz, hwphyssz;
704 vm_offset_t pa, va, off;
705 void *dpcpu;
706
707 /*
708 * Map PCI memory space.
709 */
710 battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW);
711 battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs);
712
713 battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW);
714 battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs);
715
716 battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW);
717 battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs);
718
719 battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW);
720 battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs);
721
722 powerpc_sync();
723
724 /* map pci space */
725 __asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu));
726 __asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl));
727 isync();
728
729 /* set global direct map flag */
730 hw_direct_map = 1;
731
732 mem_regions(&pregions, &pregions_sz, ®ions, ®ions_sz);
733 CTR0(KTR_PMAP, "moea_bootstrap: physical memory");
734
735 for (i = 0; i < pregions_sz; i++) {
736 vm_offset_t pa;
737 vm_offset_t end;
738
739 CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)",
740 pregions[i].mr_start,
741 pregions[i].mr_start + pregions[i].mr_size,
742 pregions[i].mr_size);
743 /*
744 * Install entries into the BAT table to allow all
745 * of physmem to be convered by on-demand BAT entries.
746 * The loop will sometimes set the same battable element
747 * twice, but that's fine since they won't be used for
748 * a while yet.
749 */
750 pa = pregions[i].mr_start & 0xf0000000;
751 end = pregions[i].mr_start + pregions[i].mr_size;
752 do {
753 u_int n = pa >> ADDR_SR_SHFT;
754
755 battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW);
756 battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs);
757 pa += SEGMENT_LENGTH;
758 } while (pa < end);
759 }
760
761 if (PHYS_AVAIL_ENTRIES < regions_sz)
762 panic("moea_bootstrap: phys_avail too small");
763
764 phys_avail_count = 0;
765 physsz = 0;
766 hwphyssz = 0;
767 TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz);
768 for (i = 0, j = 0; i < regions_sz; i++, j += 2) {
769 CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start,
770 regions[i].mr_start + regions[i].mr_size,
771 regions[i].mr_size);
772 if (hwphyssz != 0 &&
773 (physsz + regions[i].mr_size) >= hwphyssz) {
774 if (physsz < hwphyssz) {
775 phys_avail[j] = regions[i].mr_start;
776 phys_avail[j + 1] = regions[i].mr_start +
777 hwphyssz - physsz;
778 physsz = hwphyssz;
779 phys_avail_count++;
780 }
781 break;
782 }
783 phys_avail[j] = regions[i].mr_start;
784 phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size;
785 phys_avail_count++;
786 physsz += regions[i].mr_size;
787 }
788
789 /* Check for overlap with the kernel and exception vectors */
790 for (j = 0; j < 2*phys_avail_count; j+=2) {
791 if (phys_avail[j] < EXC_LAST)
792 phys_avail[j] += EXC_LAST;
793
794 if (kernelstart >= phys_avail[j] &&
795 kernelstart < phys_avail[j+1]) {
796 if (kernelend < phys_avail[j+1]) {
797 phys_avail[2*phys_avail_count] =
798 (kernelend & ~PAGE_MASK) + PAGE_SIZE;
799 phys_avail[2*phys_avail_count + 1] =
800 phys_avail[j+1];
801 phys_avail_count++;
802 }
803
804 phys_avail[j+1] = kernelstart & ~PAGE_MASK;
805 }
806
807 if (kernelend >= phys_avail[j] &&
808 kernelend < phys_avail[j+1]) {
809 if (kernelstart > phys_avail[j]) {
810 phys_avail[2*phys_avail_count] = phys_avail[j];
811 phys_avail[2*phys_avail_count + 1] =
812 kernelstart & ~PAGE_MASK;
813 phys_avail_count++;
814 }
815
816 phys_avail[j] = (kernelend & ~PAGE_MASK) + PAGE_SIZE;
817 }
818 }
819
820 physmem = btoc(physsz);
821
822 /*
823 * Allocate PTEG table.
824 */
825 #ifdef PTEGCOUNT
826 moea_pteg_count = PTEGCOUNT;
827 #else
828 moea_pteg_count = 0x1000;
829
830 while (moea_pteg_count < physmem)
831 moea_pteg_count <<= 1;
832
833 moea_pteg_count >>= 1;
834 #endif /* PTEGCOUNT */
835
836 size = moea_pteg_count * sizeof(struct pteg);
837 CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count,
838 size);
839 moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size);
840 CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table);
841 bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg));
842 moea_pteg_mask = moea_pteg_count - 1;
843
844 /*
845 * Allocate pv/overflow lists.
846 */
847 size = sizeof(struct pvo_head) * moea_pteg_count;
848 moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size,
849 PAGE_SIZE);
850 CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table);
851 for (i = 0; i < moea_pteg_count; i++)
852 LIST_INIT(&moea_pvo_table[i]);
853
854 /*
855 * Initialize the lock that synchronizes access to the pteg and pvo
856 * tables.
857 */
858 mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF |
859 MTX_RECURSE);
860 mtx_init(&moea_vsid_mutex, "VSID table", NULL, MTX_DEF);
861
862 mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN);
863
864 /*
865 * Initialise the unmanaged pvo pool.
866 */
867 moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc(
868 BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0);
869 moea_bpvo_pool_index = 0;
870
871 /*
872 * Make sure kernel vsid is allocated as well as VSID 0.
873 */
874 moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW]
875 |= 1 << (KERNEL_VSIDBITS % VSID_NBPW);
876 moea_vsid_bitmap[0] |= 1;
877
878 /*
879 * Initialize the kernel pmap (which is statically allocated).
880 */
881 PMAP_LOCK_INIT(kernel_pmap);
882 for (i = 0; i < 16; i++)
883 kernel_pmap->pm_sr[i] = EMPTY_SEGMENT + i;
884 CPU_FILL(&kernel_pmap->pm_active);
885 RB_INIT(&kernel_pmap->pmap_pvo);
886
887 /*
888 * Initialize the global pv list lock.
889 */
890 rw_init(&pvh_global_lock, "pmap pv global");
891
892 /*
893 * Set up the Open Firmware mappings
894 */
895 chosen = OF_finddevice("/chosen");
896 if (chosen != -1 && OF_getprop(chosen, "mmu", &mmui, 4) != -1 &&
897 (mmu = OF_instance_to_package(mmui)) != -1 &&
898 (sz = OF_getproplen(mmu, "translations")) != -1) {
899 translations = NULL;
900 for (i = 0; phys_avail[i] != 0; i += 2) {
901 if (phys_avail[i + 1] >= sz) {
902 translations = (struct ofw_map *)phys_avail[i];
903 break;
904 }
905 }
906 if (translations == NULL)
907 panic("moea_bootstrap: no space to copy translations");
908 bzero(translations, sz);
909 if (OF_getprop(mmu, "translations", translations, sz) == -1)
910 panic("moea_bootstrap: can't get ofw translations");
911 CTR0(KTR_PMAP, "moea_bootstrap: translations");
912 sz /= sizeof(*translations);
913 qsort(translations, sz, sizeof (*translations), om_cmp);
914 for (i = 0; i < sz; i++) {
915 CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x",
916 translations[i].om_pa, translations[i].om_va,
917 translations[i].om_len);
918
919 /*
920 * If the mapping is 1:1, let the RAM and device
921 * on-demand BAT tables take care of the translation.
922 *
923 * However, always enter mappings for segment 16,
924 * which is mixed-protection and therefore not
925 * compatible with a BAT entry.
926 */
927 if ((translations[i].om_va >> ADDR_SR_SHFT) != 0xf &&
928 translations[i].om_va == translations[i].om_pa)
929 continue;
930
931 /* Enter the pages */
932 for (off = 0; off < translations[i].om_len;
933 off += PAGE_SIZE)
934 moea_kenter_attr(translations[i].om_va + off,
935 translations[i].om_pa + off,
936 moea_bootstrap_convert_wimg(translations[i].om_mode));
937 }
938 }
939
940 /*
941 * Calculate the last available physical address.
942 */
943 for (i = 0; phys_avail[i + 2] != 0; i += 2)
944 ;
945 Maxmem = powerpc_btop(phys_avail[i + 1]);
946
947 moea_cpu_bootstrap(0);
948 mtmsr(mfmsr() | PSL_DR | PSL_IR);
949 pmap_bootstrapped++;
950
951 /*
952 * Set the start and end of kva.
953 */
954 virtual_avail = VM_MIN_KERNEL_ADDRESS;
955 virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS;
956
957 /*
958 * Allocate a kernel stack with a guard page for thread0 and map it
959 * into the kernel page map.
960 */
961 pa = moea_bootstrap_alloc(kstack_pages * PAGE_SIZE, PAGE_SIZE);
962 va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE;
963 virtual_avail = va + kstack_pages * PAGE_SIZE;
964 CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va);
965 thread0.td_kstack = va;
966 thread0.td_kstack_pages = kstack_pages;
967 for (i = 0; i < kstack_pages; i++) {
968 moea_kenter(va, pa);
969 pa += PAGE_SIZE;
970 va += PAGE_SIZE;
971 }
972
973 /*
974 * Allocate virtual address space for the message buffer.
975 */
976 pa = msgbuf_phys = moea_bootstrap_alloc(msgbufsize, PAGE_SIZE);
977 msgbufp = (struct msgbuf *)virtual_avail;
978 va = virtual_avail;
979 virtual_avail += round_page(msgbufsize);
980 while (va < virtual_avail) {
981 moea_kenter(va, pa);
982 pa += PAGE_SIZE;
983 va += PAGE_SIZE;
984 }
985
986 /*
987 * Allocate virtual address space for the dynamic percpu area.
988 */
989 pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE);
990 dpcpu = (void *)virtual_avail;
991 va = virtual_avail;
992 virtual_avail += DPCPU_SIZE;
993 while (va < virtual_avail) {
994 moea_kenter(va, pa);
995 pa += PAGE_SIZE;
996 va += PAGE_SIZE;
997 }
998 dpcpu_init(dpcpu, 0);
999 }
1000
1001 /*
1002 * Activate a user pmap. The pmap must be activated before it's address
1003 * space can be accessed in any way.
1004 */
1005 void
moea_activate(struct thread * td)1006 moea_activate(struct thread *td)
1007 {
1008 pmap_t pm, pmr;
1009
1010 /*
1011 * Load all the data we need up front to encourage the compiler to
1012 * not issue any loads while we have interrupts disabled below.
1013 */
1014 pm = &td->td_proc->p_vmspace->vm_pmap;
1015 pmr = pm->pmap_phys;
1016
1017 CPU_SET(PCPU_GET(cpuid), &pm->pm_active);
1018 PCPU_SET(curpmap, pmr);
1019
1020 mtsrin(USER_SR << ADDR_SR_SHFT, td->td_pcb->pcb_cpu.aim.usr_vsid);
1021 }
1022
1023 void
moea_deactivate(struct thread * td)1024 moea_deactivate(struct thread *td)
1025 {
1026 pmap_t pm;
1027
1028 pm = &td->td_proc->p_vmspace->vm_pmap;
1029 CPU_CLR(PCPU_GET(cpuid), &pm->pm_active);
1030 PCPU_SET(curpmap, NULL);
1031 }
1032
1033 void
moea_unwire(pmap_t pm,vm_offset_t sva,vm_offset_t eva)1034 moea_unwire(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1035 {
1036 struct pvo_entry key, *pvo;
1037
1038 PMAP_LOCK(pm);
1039 key.pvo_vaddr = sva;
1040 for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1041 pvo != NULL && PVO_VADDR(pvo) < eva;
1042 pvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo)) {
1043 if ((pvo->pvo_vaddr & PVO_WIRED) == 0)
1044 panic("moea_unwire: pvo %p is missing PVO_WIRED", pvo);
1045 pvo->pvo_vaddr &= ~PVO_WIRED;
1046 pm->pm_stats.wired_count--;
1047 }
1048 PMAP_UNLOCK(pm);
1049 }
1050
1051 void
moea_copy_page(vm_page_t msrc,vm_page_t mdst)1052 moea_copy_page(vm_page_t msrc, vm_page_t mdst)
1053 {
1054 vm_offset_t dst;
1055 vm_offset_t src;
1056
1057 dst = VM_PAGE_TO_PHYS(mdst);
1058 src = VM_PAGE_TO_PHYS(msrc);
1059
1060 bcopy((void *)src, (void *)dst, PAGE_SIZE);
1061 }
1062
1063 void
moea_copy_pages(vm_page_t * ma,vm_offset_t a_offset,vm_page_t * mb,vm_offset_t b_offset,int xfersize)1064 moea_copy_pages(vm_page_t *ma, vm_offset_t a_offset,
1065 vm_page_t *mb, vm_offset_t b_offset, int xfersize)
1066 {
1067 void *a_cp, *b_cp;
1068 vm_offset_t a_pg_offset, b_pg_offset;
1069 int cnt;
1070
1071 while (xfersize > 0) {
1072 a_pg_offset = a_offset & PAGE_MASK;
1073 cnt = min(xfersize, PAGE_SIZE - a_pg_offset);
1074 a_cp = (char *)VM_PAGE_TO_PHYS(ma[a_offset >> PAGE_SHIFT]) +
1075 a_pg_offset;
1076 b_pg_offset = b_offset & PAGE_MASK;
1077 cnt = min(cnt, PAGE_SIZE - b_pg_offset);
1078 b_cp = (char *)VM_PAGE_TO_PHYS(mb[b_offset >> PAGE_SHIFT]) +
1079 b_pg_offset;
1080 bcopy(a_cp, b_cp, cnt);
1081 a_offset += cnt;
1082 b_offset += cnt;
1083 xfersize -= cnt;
1084 }
1085 }
1086
1087 /*
1088 * Zero a page of physical memory by temporarily mapping it into the tlb.
1089 */
1090 void
moea_zero_page(vm_page_t m)1091 moea_zero_page(vm_page_t m)
1092 {
1093 vm_offset_t off, pa = VM_PAGE_TO_PHYS(m);
1094
1095 for (off = 0; off < PAGE_SIZE; off += cacheline_size)
1096 __asm __volatile("dcbz 0,%0" :: "r"(pa + off));
1097 }
1098
1099 void
moea_zero_page_area(vm_page_t m,int off,int size)1100 moea_zero_page_area(vm_page_t m, int off, int size)
1101 {
1102 vm_offset_t pa = VM_PAGE_TO_PHYS(m);
1103 void *va = (void *)(pa + off);
1104
1105 bzero(va, size);
1106 }
1107
1108 vm_offset_t
moea_quick_enter_page(vm_page_t m)1109 moea_quick_enter_page(vm_page_t m)
1110 {
1111
1112 return (VM_PAGE_TO_PHYS(m));
1113 }
1114
1115 void
moea_quick_remove_page(vm_offset_t addr)1116 moea_quick_remove_page(vm_offset_t addr)
1117 {
1118 }
1119
1120 bool
moea_page_is_mapped(vm_page_t m)1121 moea_page_is_mapped(vm_page_t m)
1122 {
1123 return (!LIST_EMPTY(&(m)->md.mdpg_pvoh));
1124 }
1125
1126 bool
moea_ps_enabled(pmap_t pmap __unused)1127 moea_ps_enabled(pmap_t pmap __unused)
1128 {
1129 return (false);
1130 }
1131
1132 /*
1133 * Map the given physical page at the specified virtual address in the
1134 * target pmap with the protection requested. If specified the page
1135 * will be wired down.
1136 */
1137 int
moea_enter(pmap_t pmap,vm_offset_t va,vm_page_t m,vm_prot_t prot,u_int flags,int8_t psind)1138 moea_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1139 u_int flags, int8_t psind)
1140 {
1141 int error;
1142
1143 for (;;) {
1144 rw_wlock(&pvh_global_lock);
1145 PMAP_LOCK(pmap);
1146 error = moea_enter_locked(pmap, va, m, prot, flags, psind);
1147 rw_wunlock(&pvh_global_lock);
1148 PMAP_UNLOCK(pmap);
1149 if (error != ENOMEM)
1150 return (KERN_SUCCESS);
1151 if ((flags & PMAP_ENTER_NOSLEEP) != 0)
1152 return (KERN_RESOURCE_SHORTAGE);
1153 VM_OBJECT_ASSERT_UNLOCKED(m->object);
1154 vm_wait(NULL);
1155 }
1156 }
1157
1158 /*
1159 * Map the given physical page at the specified virtual address in the
1160 * target pmap with the protection requested. If specified the page
1161 * will be wired down.
1162 *
1163 * The global pvh and pmap must be locked.
1164 */
1165 static int
moea_enter_locked(pmap_t pmap,vm_offset_t va,vm_page_t m,vm_prot_t prot,u_int flags,int8_t psind __unused)1166 moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot,
1167 u_int flags, int8_t psind __unused)
1168 {
1169 struct pvo_head *pvo_head;
1170 uma_zone_t zone;
1171 u_int pte_lo, pvo_flags;
1172 int error;
1173
1174 if (pmap_bootstrapped)
1175 rw_assert(&pvh_global_lock, RA_WLOCKED);
1176 PMAP_LOCK_ASSERT(pmap, MA_OWNED);
1177 if ((m->oflags & VPO_UNMANAGED) == 0) {
1178 if ((flags & PMAP_ENTER_QUICK_LOCKED) == 0)
1179 VM_PAGE_OBJECT_BUSY_ASSERT(m);
1180 else
1181 VM_OBJECT_ASSERT_LOCKED(m->object);
1182 }
1183
1184 if ((m->oflags & VPO_UNMANAGED) != 0 || !moea_initialized) {
1185 pvo_head = &moea_pvo_kunmanaged;
1186 zone = moea_upvo_zone;
1187 pvo_flags = 0;
1188 } else {
1189 pvo_head = vm_page_to_pvoh(m);
1190 zone = moea_mpvo_zone;
1191 pvo_flags = PVO_MANAGED;
1192 }
1193
1194 pte_lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), pmap_page_get_memattr(m));
1195
1196 if (prot & VM_PROT_WRITE) {
1197 pte_lo |= PTE_BW;
1198 if (pmap_bootstrapped &&
1199 (m->oflags & VPO_UNMANAGED) == 0)
1200 vm_page_aflag_set(m, PGA_WRITEABLE);
1201 } else
1202 pte_lo |= PTE_BR;
1203
1204 if ((flags & PMAP_ENTER_WIRED) != 0)
1205 pvo_flags |= PVO_WIRED;
1206
1207 error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m),
1208 pte_lo, pvo_flags);
1209
1210 /*
1211 * Flush the real page from the instruction cache. This has be done
1212 * for all user mappings to prevent information leakage via the
1213 * instruction cache. moea_pvo_enter() returns ENOENT for the first
1214 * mapping for a page.
1215 */
1216 if (pmap != kernel_pmap && error == ENOENT &&
1217 (pte_lo & (PTE_I | PTE_G)) == 0)
1218 moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE);
1219
1220 return (error);
1221 }
1222
1223 /*
1224 * Maps a sequence of resident pages belonging to the same object.
1225 * The sequence begins with the given page m_start. This page is
1226 * mapped at the given virtual address start. Each subsequent page is
1227 * mapped at a virtual address that is offset from start by the same
1228 * amount as the page is offset from m_start within the object. The
1229 * last page in the sequence is the page with the largest offset from
1230 * m_start that can be mapped at a virtual address less than the given
1231 * virtual address end. Not every virtual page between start and end
1232 * is mapped; only those for which a resident page exists with the
1233 * corresponding offset from m_start are mapped.
1234 */
1235 void
moea_enter_object(pmap_t pm,vm_offset_t start,vm_offset_t end,vm_page_t m_start,vm_prot_t prot)1236 moea_enter_object(pmap_t pm, vm_offset_t start, vm_offset_t end,
1237 vm_page_t m_start, vm_prot_t prot)
1238 {
1239 struct pctrie_iter pages;
1240 vm_offset_t va;
1241 vm_page_t m;
1242
1243 VM_OBJECT_ASSERT_LOCKED(m_start->object);
1244
1245 vm_page_iter_limit_init(&pages, m_start->object,
1246 m_start->pindex + atop(end - start));
1247 m = vm_radix_iter_lookup(&pages, m_start->pindex);
1248 rw_wlock(&pvh_global_lock);
1249 PMAP_LOCK(pm);
1250 while (m != NULL) {
1251 va = start + ptoa(m->pindex - m_start->pindex);
1252 moea_enter_locked(pm, va, m, prot &
1253 (VM_PROT_READ | VM_PROT_EXECUTE), PMAP_ENTER_QUICK_LOCKED,
1254 0);
1255 m = vm_radix_iter_step(&pages);
1256 }
1257 rw_wunlock(&pvh_global_lock);
1258 PMAP_UNLOCK(pm);
1259 }
1260
1261 void
moea_enter_quick(pmap_t pm,vm_offset_t va,vm_page_t m,vm_prot_t prot)1262 moea_enter_quick(pmap_t pm, vm_offset_t va, vm_page_t m,
1263 vm_prot_t prot)
1264 {
1265
1266 rw_wlock(&pvh_global_lock);
1267 PMAP_LOCK(pm);
1268 moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE),
1269 PMAP_ENTER_QUICK_LOCKED, 0);
1270 rw_wunlock(&pvh_global_lock);
1271 PMAP_UNLOCK(pm);
1272 }
1273
1274 vm_paddr_t
moea_extract(pmap_t pm,vm_offset_t va)1275 moea_extract(pmap_t pm, vm_offset_t va)
1276 {
1277 struct pvo_entry *pvo;
1278 vm_paddr_t pa;
1279
1280 PMAP_LOCK(pm);
1281 pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1282 if (pvo == NULL)
1283 pa = 0;
1284 else
1285 pa = PVO_PADDR(pvo) | (va & ADDR_POFF);
1286 PMAP_UNLOCK(pm);
1287 return (pa);
1288 }
1289
1290 /*
1291 * Atomically extract and hold the physical page with the given
1292 * pmap and virtual address pair if that mapping permits the given
1293 * protection.
1294 */
1295 vm_page_t
moea_extract_and_hold(pmap_t pmap,vm_offset_t va,vm_prot_t prot)1296 moea_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot)
1297 {
1298 struct pvo_entry *pvo;
1299 vm_page_t m;
1300
1301 m = NULL;
1302 PMAP_LOCK(pmap);
1303 pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1304 if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) &&
1305 ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW ||
1306 (prot & VM_PROT_WRITE) == 0)) {
1307 m = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
1308 if (!vm_page_wire_mapped(m))
1309 m = NULL;
1310 }
1311 PMAP_UNLOCK(pmap);
1312 return (m);
1313 }
1314
1315 void
moea_init(void)1316 moea_init(void)
1317 {
1318
1319 moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry),
1320 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1321 UMA_ZONE_VM | UMA_ZONE_NOFREE);
1322 moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry),
1323 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR,
1324 UMA_ZONE_VM | UMA_ZONE_NOFREE);
1325 moea_initialized = true;
1326 }
1327
1328 bool
moea_is_referenced(vm_page_t m)1329 moea_is_referenced(vm_page_t m)
1330 {
1331 bool rv;
1332
1333 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1334 ("moea_is_referenced: page %p is not managed", m));
1335 rw_wlock(&pvh_global_lock);
1336 rv = moea_query_bit(m, PTE_REF);
1337 rw_wunlock(&pvh_global_lock);
1338 return (rv);
1339 }
1340
1341 bool
moea_is_modified(vm_page_t m)1342 moea_is_modified(vm_page_t m)
1343 {
1344 bool rv;
1345
1346 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1347 ("moea_is_modified: page %p is not managed", m));
1348
1349 /*
1350 * If the page is not busied then this check is racy.
1351 */
1352 if (!pmap_page_is_write_mapped(m))
1353 return (false);
1354
1355 rw_wlock(&pvh_global_lock);
1356 rv = moea_query_bit(m, PTE_CHG);
1357 rw_wunlock(&pvh_global_lock);
1358 return (rv);
1359 }
1360
1361 bool
moea_is_prefaultable(pmap_t pmap,vm_offset_t va)1362 moea_is_prefaultable(pmap_t pmap, vm_offset_t va)
1363 {
1364 struct pvo_entry *pvo;
1365 bool rv;
1366
1367 PMAP_LOCK(pmap);
1368 pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL);
1369 rv = pvo == NULL || (pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0;
1370 PMAP_UNLOCK(pmap);
1371 return (rv);
1372 }
1373
1374 void
moea_clear_modify(vm_page_t m)1375 moea_clear_modify(vm_page_t m)
1376 {
1377
1378 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1379 ("moea_clear_modify: page %p is not managed", m));
1380 vm_page_assert_busied(m);
1381
1382 if (!pmap_page_is_write_mapped(m))
1383 return;
1384 rw_wlock(&pvh_global_lock);
1385 moea_clear_bit(m, PTE_CHG);
1386 rw_wunlock(&pvh_global_lock);
1387 }
1388
1389 /*
1390 * Clear the write and modified bits in each of the given page's mappings.
1391 */
1392 void
moea_remove_write(vm_page_t m)1393 moea_remove_write(vm_page_t m)
1394 {
1395 struct pvo_entry *pvo;
1396 struct pte *pt;
1397 pmap_t pmap;
1398 u_int lo;
1399
1400 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1401 ("moea_remove_write: page %p is not managed", m));
1402 vm_page_assert_busied(m);
1403
1404 if (!pmap_page_is_write_mapped(m))
1405 return;
1406 rw_wlock(&pvh_global_lock);
1407 lo = moea_attr_fetch(m);
1408 powerpc_sync();
1409 LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1410 pmap = pvo->pvo_pmap;
1411 PMAP_LOCK(pmap);
1412 if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) {
1413 pt = moea_pvo_to_pte(pvo, -1);
1414 pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1415 pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1416 if (pt != NULL) {
1417 moea_pte_synch(pt, &pvo->pvo_pte.pte);
1418 lo |= pvo->pvo_pte.pte.pte_lo;
1419 pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG;
1420 moea_pte_change(pt, &pvo->pvo_pte.pte,
1421 pvo->pvo_vaddr);
1422 mtx_unlock(&moea_table_mutex);
1423 }
1424 }
1425 PMAP_UNLOCK(pmap);
1426 }
1427 if ((lo & PTE_CHG) != 0) {
1428 moea_attr_clear(m, PTE_CHG);
1429 vm_page_dirty(m);
1430 }
1431 vm_page_aflag_clear(m, PGA_WRITEABLE);
1432 rw_wunlock(&pvh_global_lock);
1433 }
1434
1435 /*
1436 * moea_ts_referenced:
1437 *
1438 * Return a count of reference bits for a page, clearing those bits.
1439 * It is not necessary for every reference bit to be cleared, but it
1440 * is necessary that 0 only be returned when there are truly no
1441 * reference bits set.
1442 *
1443 * XXX: The exact number of bits to check and clear is a matter that
1444 * should be tested and standardized at some point in the future for
1445 * optimal aging of shared pages.
1446 */
1447 int
moea_ts_referenced(vm_page_t m)1448 moea_ts_referenced(vm_page_t m)
1449 {
1450 int count;
1451
1452 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1453 ("moea_ts_referenced: page %p is not managed", m));
1454 rw_wlock(&pvh_global_lock);
1455 count = moea_clear_bit(m, PTE_REF);
1456 rw_wunlock(&pvh_global_lock);
1457 return (count);
1458 }
1459
1460 /*
1461 * Modify the WIMG settings of all mappings for a page.
1462 */
1463 void
moea_page_set_memattr(vm_page_t m,vm_memattr_t ma)1464 moea_page_set_memattr(vm_page_t m, vm_memattr_t ma)
1465 {
1466 struct pvo_entry *pvo;
1467 struct pvo_head *pvo_head;
1468 struct pte *pt;
1469 pmap_t pmap;
1470 u_int lo;
1471
1472 if ((m->oflags & VPO_UNMANAGED) != 0) {
1473 m->md.mdpg_cache_attrs = ma;
1474 return;
1475 }
1476
1477 rw_wlock(&pvh_global_lock);
1478 pvo_head = vm_page_to_pvoh(m);
1479 lo = moea_calc_wimg(VM_PAGE_TO_PHYS(m), ma);
1480
1481 LIST_FOREACH(pvo, pvo_head, pvo_vlink) {
1482 pmap = pvo->pvo_pmap;
1483 PMAP_LOCK(pmap);
1484 pt = moea_pvo_to_pte(pvo, -1);
1485 pvo->pvo_pte.pte.pte_lo &= ~PTE_WIMG;
1486 pvo->pvo_pte.pte.pte_lo |= lo;
1487 if (pt != NULL) {
1488 moea_pte_change(pt, &pvo->pvo_pte.pte,
1489 pvo->pvo_vaddr);
1490 if (pvo->pvo_pmap == kernel_pmap)
1491 isync();
1492 }
1493 mtx_unlock(&moea_table_mutex);
1494 PMAP_UNLOCK(pmap);
1495 }
1496 m->md.mdpg_cache_attrs = ma;
1497 rw_wunlock(&pvh_global_lock);
1498 }
1499
1500 /*
1501 * Map a wired page into kernel virtual address space.
1502 */
1503 void
moea_kenter(vm_offset_t va,vm_paddr_t pa)1504 moea_kenter(vm_offset_t va, vm_paddr_t pa)
1505 {
1506
1507 moea_kenter_attr(va, pa, VM_MEMATTR_DEFAULT);
1508 }
1509
1510 void
moea_kenter_attr(vm_offset_t va,vm_paddr_t pa,vm_memattr_t ma)1511 moea_kenter_attr(vm_offset_t va, vm_paddr_t pa, vm_memattr_t ma)
1512 {
1513 u_int pte_lo;
1514 int error;
1515
1516 #if 0
1517 if (va < VM_MIN_KERNEL_ADDRESS)
1518 panic("moea_kenter: attempt to enter non-kernel address %#x",
1519 va);
1520 #endif
1521
1522 pte_lo = moea_calc_wimg(pa, ma);
1523
1524 PMAP_LOCK(kernel_pmap);
1525 error = moea_pvo_enter(kernel_pmap, moea_upvo_zone,
1526 &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED);
1527
1528 if (error != 0 && error != ENOENT)
1529 panic("moea_kenter: failed to enter va %#x pa %#x: %d", va,
1530 pa, error);
1531
1532 PMAP_UNLOCK(kernel_pmap);
1533 }
1534
1535 /*
1536 * Extract the physical page address associated with the given kernel virtual
1537 * address.
1538 */
1539 vm_paddr_t
moea_kextract(vm_offset_t va)1540 moea_kextract(vm_offset_t va)
1541 {
1542 struct pvo_entry *pvo;
1543 vm_paddr_t pa;
1544
1545 /*
1546 * Allow direct mappings on 32-bit OEA
1547 */
1548 if (va < VM_MIN_KERNEL_ADDRESS) {
1549 return (va);
1550 }
1551
1552 PMAP_LOCK(kernel_pmap);
1553 pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
1554 KASSERT(pvo != NULL, ("moea_kextract: no addr found"));
1555 pa = PVO_PADDR(pvo) | (va & ADDR_POFF);
1556 PMAP_UNLOCK(kernel_pmap);
1557 return (pa);
1558 }
1559
1560 /*
1561 * Remove a wired page from kernel virtual address space.
1562 */
1563 void
moea_kremove(vm_offset_t va)1564 moea_kremove(vm_offset_t va)
1565 {
1566
1567 moea_remove(kernel_pmap, va, va + PAGE_SIZE);
1568 }
1569
1570 /*
1571 * Provide a kernel pointer corresponding to a given userland pointer.
1572 * The returned pointer is valid until the next time this function is
1573 * called in this thread. This is used internally in copyin/copyout.
1574 */
1575 int
moea_map_user_ptr(pmap_t pm,volatile const void * uaddr,void ** kaddr,size_t ulen,size_t * klen)1576 moea_map_user_ptr(pmap_t pm, volatile const void *uaddr,
1577 void **kaddr, size_t ulen, size_t *klen)
1578 {
1579 size_t l;
1580 register_t vsid;
1581
1582 *kaddr = (char *)USER_ADDR + ((uintptr_t)uaddr & ~SEGMENT_MASK);
1583 l = ((char *)USER_ADDR + SEGMENT_LENGTH) - (char *)(*kaddr);
1584 if (l > ulen)
1585 l = ulen;
1586 if (klen)
1587 *klen = l;
1588 else if (l != ulen)
1589 return (EFAULT);
1590
1591 vsid = va_to_vsid(pm, (vm_offset_t)uaddr);
1592
1593 /* Mark segment no-execute */
1594 vsid |= SR_N;
1595
1596 /* If we have already set this VSID, we can just return */
1597 if (curthread->td_pcb->pcb_cpu.aim.usr_vsid == vsid)
1598 return (0);
1599
1600 __asm __volatile("isync");
1601 curthread->td_pcb->pcb_cpu.aim.usr_segm =
1602 (uintptr_t)uaddr >> ADDR_SR_SHFT;
1603 curthread->td_pcb->pcb_cpu.aim.usr_vsid = vsid;
1604 __asm __volatile("mtsr %0,%1; isync" :: "n"(USER_SR), "r"(vsid));
1605
1606 return (0);
1607 }
1608
1609 /*
1610 * Figure out where a given kernel pointer (usually in a fault) points
1611 * to from the VM's perspective, potentially remapping into userland's
1612 * address space.
1613 */
1614 static int
moea_decode_kernel_ptr(vm_offset_t addr,int * is_user,vm_offset_t * decoded_addr)1615 moea_decode_kernel_ptr(vm_offset_t addr, int *is_user,
1616 vm_offset_t *decoded_addr)
1617 {
1618 vm_offset_t user_sr;
1619
1620 if ((addr >> ADDR_SR_SHFT) == (USER_ADDR >> ADDR_SR_SHFT)) {
1621 user_sr = curthread->td_pcb->pcb_cpu.aim.usr_segm;
1622 addr &= ADDR_PIDX | ADDR_POFF;
1623 addr |= user_sr << ADDR_SR_SHFT;
1624 *decoded_addr = addr;
1625 *is_user = 1;
1626 } else {
1627 *decoded_addr = addr;
1628 *is_user = 0;
1629 }
1630
1631 return (0);
1632 }
1633
1634 /*
1635 * Map a range of physical addresses into kernel virtual address space.
1636 *
1637 * The value passed in *virt is a suggested virtual address for the mapping.
1638 * Architectures which can support a direct-mapped physical to virtual region
1639 * can return the appropriate address within that region, leaving '*virt'
1640 * unchanged. We cannot and therefore do not; *virt is updated with the
1641 * first usable address after the mapped region.
1642 */
1643 vm_offset_t
moea_map(vm_offset_t * virt,vm_paddr_t pa_start,vm_paddr_t pa_end,int prot)1644 moea_map(vm_offset_t *virt, vm_paddr_t pa_start,
1645 vm_paddr_t pa_end, int prot)
1646 {
1647 vm_offset_t sva, va;
1648
1649 sva = *virt;
1650 va = sva;
1651 for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE)
1652 moea_kenter(va, pa_start);
1653 *virt = va;
1654 return (sva);
1655 }
1656
1657 /*
1658 * Returns true if the pmap's pv is one of the first
1659 * 16 pvs linked to from this page. This count may
1660 * be changed upwards or downwards in the future; it
1661 * is only necessary that true be returned for a small
1662 * subset of pmaps for proper page aging.
1663 */
1664 bool
moea_page_exists_quick(pmap_t pmap,vm_page_t m)1665 moea_page_exists_quick(pmap_t pmap, vm_page_t m)
1666 {
1667 int loops;
1668 struct pvo_entry *pvo;
1669 bool rv;
1670
1671 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
1672 ("moea_page_exists_quick: page %p is not managed", m));
1673 loops = 0;
1674 rv = false;
1675 rw_wlock(&pvh_global_lock);
1676 LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
1677 if (pvo->pvo_pmap == pmap) {
1678 rv = true;
1679 break;
1680 }
1681 if (++loops >= 16)
1682 break;
1683 }
1684 rw_wunlock(&pvh_global_lock);
1685 return (rv);
1686 }
1687
1688 void
moea_page_init(vm_page_t m)1689 moea_page_init(vm_page_t m)
1690 {
1691
1692 m->md.mdpg_attrs = 0;
1693 m->md.mdpg_cache_attrs = VM_MEMATTR_DEFAULT;
1694 LIST_INIT(&m->md.mdpg_pvoh);
1695 }
1696
1697 /*
1698 * Return the number of managed mappings to the given physical page
1699 * that are wired.
1700 */
1701 int
moea_page_wired_mappings(vm_page_t m)1702 moea_page_wired_mappings(vm_page_t m)
1703 {
1704 struct pvo_entry *pvo;
1705 int count;
1706
1707 count = 0;
1708 if ((m->oflags & VPO_UNMANAGED) != 0)
1709 return (count);
1710 rw_wlock(&pvh_global_lock);
1711 LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink)
1712 if ((pvo->pvo_vaddr & PVO_WIRED) != 0)
1713 count++;
1714 rw_wunlock(&pvh_global_lock);
1715 return (count);
1716 }
1717
1718 static u_int moea_vsidcontext;
1719
1720 int
moea_pinit(pmap_t pmap)1721 moea_pinit(pmap_t pmap)
1722 {
1723 int i, mask;
1724 u_int entropy;
1725
1726 RB_INIT(&pmap->pmap_pvo);
1727
1728 entropy = 0;
1729 __asm __volatile("mftb %0" : "=r"(entropy));
1730
1731 if ((pmap->pmap_phys = (pmap_t)moea_kextract((vm_offset_t)pmap))
1732 == NULL) {
1733 pmap->pmap_phys = pmap;
1734 }
1735
1736 mtx_lock(&moea_vsid_mutex);
1737 /*
1738 * Allocate some segment registers for this pmap.
1739 */
1740 for (i = 0; i < NPMAPS; i += VSID_NBPW) {
1741 u_int hash, n;
1742
1743 /*
1744 * Create a new value by multiplying by a prime and adding in
1745 * entropy from the timebase register. This is to make the
1746 * VSID more random so that the PT hash function collides
1747 * less often. (Note that the prime casues gcc to do shifts
1748 * instead of a multiply.)
1749 */
1750 moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy;
1751 hash = moea_vsidcontext & (NPMAPS - 1);
1752 if (hash == 0) /* 0 is special, avoid it */
1753 continue;
1754 n = hash >> 5;
1755 mask = 1 << (hash & (VSID_NBPW - 1));
1756 hash = (moea_vsidcontext & 0xfffff);
1757 if (moea_vsid_bitmap[n] & mask) { /* collision? */
1758 /* anything free in this bucket? */
1759 if (moea_vsid_bitmap[n] == 0xffffffff) {
1760 entropy = (moea_vsidcontext >> 20);
1761 continue;
1762 }
1763 i = ffs(~moea_vsid_bitmap[n]) - 1;
1764 mask = 1 << i;
1765 hash &= rounddown2(0xfffff, VSID_NBPW);
1766 hash |= i;
1767 }
1768 KASSERT(!(moea_vsid_bitmap[n] & mask),
1769 ("Allocating in-use VSID group %#x\n", hash));
1770 moea_vsid_bitmap[n] |= mask;
1771 for (i = 0; i < 16; i++)
1772 pmap->pm_sr[i] = VSID_MAKE(i, hash);
1773 mtx_unlock(&moea_vsid_mutex);
1774 return (1);
1775 }
1776
1777 mtx_unlock(&moea_vsid_mutex);
1778 panic("moea_pinit: out of segments");
1779 }
1780
1781 /*
1782 * Initialize the pmap associated with process 0.
1783 */
1784 void
moea_pinit0(pmap_t pm)1785 moea_pinit0(pmap_t pm)
1786 {
1787
1788 PMAP_LOCK_INIT(pm);
1789 moea_pinit(pm);
1790 bzero(&pm->pm_stats, sizeof(pm->pm_stats));
1791 }
1792
1793 /*
1794 * Set the physical protection on the specified range of this map as requested.
1795 */
1796 void
moea_protect(pmap_t pm,vm_offset_t sva,vm_offset_t eva,vm_prot_t prot)1797 moea_protect(pmap_t pm, vm_offset_t sva, vm_offset_t eva,
1798 vm_prot_t prot)
1799 {
1800 struct pvo_entry *pvo, *tpvo, key;
1801 struct pte *pt;
1802
1803 KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap,
1804 ("moea_protect: non current pmap"));
1805
1806 if ((prot & VM_PROT_READ) == VM_PROT_NONE) {
1807 moea_remove(pm, sva, eva);
1808 return;
1809 }
1810
1811 rw_wlock(&pvh_global_lock);
1812 PMAP_LOCK(pm);
1813 key.pvo_vaddr = sva;
1814 for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1815 pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1816 tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1817
1818 /*
1819 * Grab the PTE pointer before we diddle with the cached PTE
1820 * copy.
1821 */
1822 pt = moea_pvo_to_pte(pvo, -1);
1823 /*
1824 * Change the protection of the page.
1825 */
1826 pvo->pvo_pte.pte.pte_lo &= ~PTE_PP;
1827 pvo->pvo_pte.pte.pte_lo |= PTE_BR;
1828
1829 /*
1830 * If the PVO is in the page table, update that pte as well.
1831 */
1832 if (pt != NULL) {
1833 moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
1834 mtx_unlock(&moea_table_mutex);
1835 }
1836 }
1837 rw_wunlock(&pvh_global_lock);
1838 PMAP_UNLOCK(pm);
1839 }
1840
1841 /*
1842 * Map a list of wired pages into kernel virtual address space. This is
1843 * intended for temporary mappings which do not need page modification or
1844 * references recorded. Existing mappings in the region are overwritten.
1845 */
1846 void
moea_qenter(vm_offset_t sva,vm_page_t * m,int count)1847 moea_qenter(vm_offset_t sva, vm_page_t *m, int count)
1848 {
1849 vm_offset_t va;
1850
1851 va = sva;
1852 while (count-- > 0) {
1853 moea_kenter(va, VM_PAGE_TO_PHYS(*m));
1854 va += PAGE_SIZE;
1855 m++;
1856 }
1857 }
1858
1859 /*
1860 * Remove page mappings from kernel virtual address space. Intended for
1861 * temporary mappings entered by moea_qenter.
1862 */
1863 void
moea_qremove(vm_offset_t sva,int count)1864 moea_qremove(vm_offset_t sva, int count)
1865 {
1866 vm_offset_t va;
1867
1868 va = sva;
1869 while (count-- > 0) {
1870 moea_kremove(va);
1871 va += PAGE_SIZE;
1872 }
1873 }
1874
1875 void
moea_release(pmap_t pmap)1876 moea_release(pmap_t pmap)
1877 {
1878 int idx, mask;
1879
1880 /*
1881 * Free segment register's VSID
1882 */
1883 if (pmap->pm_sr[0] == 0)
1884 panic("moea_release");
1885
1886 mtx_lock(&moea_vsid_mutex);
1887 idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1);
1888 mask = 1 << (idx % VSID_NBPW);
1889 idx /= VSID_NBPW;
1890 moea_vsid_bitmap[idx] &= ~mask;
1891 mtx_unlock(&moea_vsid_mutex);
1892 }
1893
1894 /*
1895 * Remove the given range of addresses from the specified map.
1896 */
1897 void
moea_remove(pmap_t pm,vm_offset_t sva,vm_offset_t eva)1898 moea_remove(pmap_t pm, vm_offset_t sva, vm_offset_t eva)
1899 {
1900 struct pvo_entry *pvo, *tpvo, key;
1901
1902 rw_wlock(&pvh_global_lock);
1903 PMAP_LOCK(pm);
1904 key.pvo_vaddr = sva;
1905 for (pvo = RB_NFIND(pvo_tree, &pm->pmap_pvo, &key);
1906 pvo != NULL && PVO_VADDR(pvo) < eva; pvo = tpvo) {
1907 tpvo = RB_NEXT(pvo_tree, &pm->pmap_pvo, pvo);
1908 moea_pvo_remove(pvo, -1);
1909 }
1910 PMAP_UNLOCK(pm);
1911 rw_wunlock(&pvh_global_lock);
1912 }
1913
1914 /*
1915 * Remove physical page from all pmaps in which it resides. moea_pvo_remove()
1916 * will reflect changes in pte's back to the vm_page.
1917 */
1918 void
moea_remove_all(vm_page_t m)1919 moea_remove_all(vm_page_t m)
1920 {
1921 struct pvo_head *pvo_head;
1922 struct pvo_entry *pvo, *next_pvo;
1923 pmap_t pmap;
1924
1925 rw_wlock(&pvh_global_lock);
1926 pvo_head = vm_page_to_pvoh(m);
1927 for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) {
1928 next_pvo = LIST_NEXT(pvo, pvo_vlink);
1929
1930 pmap = pvo->pvo_pmap;
1931 PMAP_LOCK(pmap);
1932 moea_pvo_remove(pvo, -1);
1933 PMAP_UNLOCK(pmap);
1934 }
1935 if ((m->a.flags & PGA_WRITEABLE) && moea_query_bit(m, PTE_CHG)) {
1936 moea_attr_clear(m, PTE_CHG);
1937 vm_page_dirty(m);
1938 }
1939 vm_page_aflag_clear(m, PGA_WRITEABLE);
1940 rw_wunlock(&pvh_global_lock);
1941 }
1942
1943 static int
moea_mincore(pmap_t pm,vm_offset_t va,vm_paddr_t * pap)1944 moea_mincore(pmap_t pm, vm_offset_t va, vm_paddr_t *pap)
1945 {
1946 struct pvo_entry *pvo;
1947 vm_paddr_t pa;
1948 vm_page_t m;
1949 int val;
1950 bool managed;
1951
1952 PMAP_LOCK(pm);
1953
1954 pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
1955 if (pvo != NULL) {
1956 pa = PVO_PADDR(pvo);
1957 m = PHYS_TO_VM_PAGE(pa);
1958 managed = (pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED;
1959 val = MINCORE_INCORE;
1960 } else {
1961 PMAP_UNLOCK(pm);
1962 return (0);
1963 }
1964
1965 PMAP_UNLOCK(pm);
1966
1967 if (m == NULL)
1968 return (0);
1969
1970 if (managed) {
1971 if (moea_is_modified(m))
1972 val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER;
1973
1974 if (moea_is_referenced(m))
1975 val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER;
1976 }
1977
1978 if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) !=
1979 (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) &&
1980 managed) {
1981 *pap = pa;
1982 }
1983
1984 return (val);
1985 }
1986
1987 /*
1988 * Allocate a physical page of memory directly from the phys_avail map.
1989 * Can only be called from moea_bootstrap before avail start and end are
1990 * calculated.
1991 */
1992 static vm_offset_t
moea_bootstrap_alloc(vm_size_t size,u_int align)1993 moea_bootstrap_alloc(vm_size_t size, u_int align)
1994 {
1995 vm_offset_t s, e;
1996 int i, j;
1997
1998 size = round_page(size);
1999 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
2000 if (align != 0)
2001 s = roundup2(phys_avail[i], align);
2002 else
2003 s = phys_avail[i];
2004 e = s + size;
2005
2006 if (s < phys_avail[i] || e > phys_avail[i + 1])
2007 continue;
2008
2009 if (s == phys_avail[i]) {
2010 phys_avail[i] += size;
2011 } else if (e == phys_avail[i + 1]) {
2012 phys_avail[i + 1] -= size;
2013 } else {
2014 for (j = phys_avail_count * 2; j > i; j -= 2) {
2015 phys_avail[j] = phys_avail[j - 2];
2016 phys_avail[j + 1] = phys_avail[j - 1];
2017 }
2018
2019 phys_avail[i + 3] = phys_avail[i + 1];
2020 phys_avail[i + 1] = s;
2021 phys_avail[i + 2] = e;
2022 phys_avail_count++;
2023 }
2024
2025 return (s);
2026 }
2027 panic("moea_bootstrap_alloc: could not allocate memory");
2028 }
2029
2030 static void
moea_syncicache(vm_paddr_t pa,vm_size_t len)2031 moea_syncicache(vm_paddr_t pa, vm_size_t len)
2032 {
2033 __syncicache((void *)pa, len);
2034 }
2035
2036 static int
moea_pvo_enter(pmap_t pm,uma_zone_t zone,struct pvo_head * pvo_head,vm_offset_t va,vm_paddr_t pa,u_int pte_lo,int flags)2037 moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head,
2038 vm_offset_t va, vm_paddr_t pa, u_int pte_lo, int flags)
2039 {
2040 struct pvo_entry *pvo;
2041 u_int sr;
2042 int first;
2043 u_int ptegidx;
2044 int i;
2045 int bootstrap;
2046
2047 moea_pvo_enter_calls++;
2048 first = 0;
2049 bootstrap = 0;
2050
2051 /*
2052 * Compute the PTE Group index.
2053 */
2054 va &= ~ADDR_POFF;
2055 sr = va_to_sr(pm->pm_sr, va);
2056 ptegidx = va_to_pteg(sr, va);
2057
2058 /*
2059 * Remove any existing mapping for this page. Reuse the pvo entry if
2060 * there is a mapping.
2061 */
2062 mtx_lock(&moea_table_mutex);
2063 LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2064 if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2065 if (PVO_PADDR(pvo) == pa &&
2066 (pvo->pvo_pte.pte.pte_lo & PTE_PP) ==
2067 (pte_lo & PTE_PP)) {
2068 /*
2069 * The PTE is not changing. Instead, this may
2070 * be a request to change the mapping's wired
2071 * attribute.
2072 */
2073 mtx_unlock(&moea_table_mutex);
2074 if ((flags & PVO_WIRED) != 0 &&
2075 (pvo->pvo_vaddr & PVO_WIRED) == 0) {
2076 pvo->pvo_vaddr |= PVO_WIRED;
2077 pm->pm_stats.wired_count++;
2078 } else if ((flags & PVO_WIRED) == 0 &&
2079 (pvo->pvo_vaddr & PVO_WIRED) != 0) {
2080 pvo->pvo_vaddr &= ~PVO_WIRED;
2081 pm->pm_stats.wired_count--;
2082 }
2083 return (0);
2084 }
2085 moea_pvo_remove(pvo, -1);
2086 break;
2087 }
2088 }
2089
2090 /*
2091 * If we aren't overwriting a mapping, try to allocate.
2092 */
2093 if (moea_initialized) {
2094 pvo = uma_zalloc(zone, M_NOWAIT);
2095 } else {
2096 if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) {
2097 panic("moea_enter: bpvo pool exhausted, %d, %d, %d",
2098 moea_bpvo_pool_index, BPVO_POOL_SIZE,
2099 BPVO_POOL_SIZE * sizeof(struct pvo_entry));
2100 }
2101 pvo = &moea_bpvo_pool[moea_bpvo_pool_index];
2102 moea_bpvo_pool_index++;
2103 bootstrap = 1;
2104 }
2105
2106 if (pvo == NULL) {
2107 mtx_unlock(&moea_table_mutex);
2108 return (ENOMEM);
2109 }
2110
2111 moea_pvo_entries++;
2112 pvo->pvo_vaddr = va;
2113 pvo->pvo_pmap = pm;
2114 LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink);
2115 pvo->pvo_vaddr &= ~ADDR_POFF;
2116 if (flags & PVO_WIRED)
2117 pvo->pvo_vaddr |= PVO_WIRED;
2118 if (pvo_head != &moea_pvo_kunmanaged)
2119 pvo->pvo_vaddr |= PVO_MANAGED;
2120 if (bootstrap)
2121 pvo->pvo_vaddr |= PVO_BOOTSTRAP;
2122
2123 moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo);
2124
2125 /*
2126 * Add to pmap list
2127 */
2128 RB_INSERT(pvo_tree, &pm->pmap_pvo, pvo);
2129
2130 /*
2131 * Remember if the list was empty and therefore will be the first
2132 * item.
2133 */
2134 if (LIST_FIRST(pvo_head) == NULL)
2135 first = 1;
2136 LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink);
2137
2138 if (pvo->pvo_vaddr & PVO_WIRED)
2139 pm->pm_stats.wired_count++;
2140 pm->pm_stats.resident_count++;
2141
2142 i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2143 KASSERT(i < 8, ("Invalid PTE index"));
2144 if (i >= 0) {
2145 PVO_PTEGIDX_SET(pvo, i);
2146 } else {
2147 panic("moea_pvo_enter: overflow");
2148 moea_pte_overflow++;
2149 }
2150 mtx_unlock(&moea_table_mutex);
2151
2152 return (first ? ENOENT : 0);
2153 }
2154
2155 static void
moea_pvo_remove(struct pvo_entry * pvo,int pteidx)2156 moea_pvo_remove(struct pvo_entry *pvo, int pteidx)
2157 {
2158 struct pte *pt;
2159
2160 /*
2161 * If there is an active pte entry, we need to deactivate it (and
2162 * save the ref & cfg bits).
2163 */
2164 pt = moea_pvo_to_pte(pvo, pteidx);
2165 if (pt != NULL) {
2166 moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr);
2167 mtx_unlock(&moea_table_mutex);
2168 PVO_PTEGIDX_CLR(pvo);
2169 } else {
2170 moea_pte_overflow--;
2171 }
2172
2173 /*
2174 * Update our statistics.
2175 */
2176 pvo->pvo_pmap->pm_stats.resident_count--;
2177 if (pvo->pvo_vaddr & PVO_WIRED)
2178 pvo->pvo_pmap->pm_stats.wired_count--;
2179
2180 /*
2181 * Remove this PVO from the PV and pmap lists.
2182 */
2183 LIST_REMOVE(pvo, pvo_vlink);
2184 RB_REMOVE(pvo_tree, &pvo->pvo_pmap->pmap_pvo, pvo);
2185
2186 /*
2187 * Save the REF/CHG bits into their cache if the page is managed.
2188 * Clear PGA_WRITEABLE if all mappings of the page have been removed.
2189 */
2190 if ((pvo->pvo_vaddr & PVO_MANAGED) == PVO_MANAGED) {
2191 struct vm_page *pg;
2192
2193 pg = PHYS_TO_VM_PAGE(PVO_PADDR(pvo));
2194 if (pg != NULL) {
2195 moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo &
2196 (PTE_REF | PTE_CHG));
2197 if (LIST_EMPTY(&pg->md.mdpg_pvoh))
2198 vm_page_aflag_clear(pg, PGA_WRITEABLE);
2199 }
2200 }
2201
2202 /*
2203 * Remove this from the overflow list and return it to the pool
2204 * if we aren't going to reuse it.
2205 */
2206 LIST_REMOVE(pvo, pvo_olink);
2207 if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP))
2208 uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone :
2209 moea_upvo_zone, pvo);
2210 moea_pvo_entries--;
2211 moea_pvo_remove_calls++;
2212 }
2213
2214 static __inline int
moea_pvo_pte_index(const struct pvo_entry * pvo,int ptegidx)2215 moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx)
2216 {
2217 int pteidx;
2218
2219 /*
2220 * We can find the actual pte entry without searching by grabbing
2221 * the PTEG index from 3 unused bits in pte_lo[11:9] and by
2222 * noticing the HID bit.
2223 */
2224 pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo);
2225 if (pvo->pvo_pte.pte.pte_hi & PTE_HID)
2226 pteidx ^= moea_pteg_mask * 8;
2227
2228 return (pteidx);
2229 }
2230
2231 static struct pvo_entry *
moea_pvo_find_va(pmap_t pm,vm_offset_t va,int * pteidx_p)2232 moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p)
2233 {
2234 struct pvo_entry *pvo;
2235 int ptegidx;
2236 u_int sr;
2237
2238 va &= ~ADDR_POFF;
2239 sr = va_to_sr(pm->pm_sr, va);
2240 ptegidx = va_to_pteg(sr, va);
2241
2242 mtx_lock(&moea_table_mutex);
2243 LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2244 if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) {
2245 if (pteidx_p)
2246 *pteidx_p = moea_pvo_pte_index(pvo, ptegidx);
2247 break;
2248 }
2249 }
2250 mtx_unlock(&moea_table_mutex);
2251
2252 return (pvo);
2253 }
2254
2255 static struct pte *
moea_pvo_to_pte(const struct pvo_entry * pvo,int pteidx)2256 moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx)
2257 {
2258 struct pte *pt;
2259
2260 /*
2261 * If we haven't been supplied the ptegidx, calculate it.
2262 */
2263 if (pteidx == -1) {
2264 int ptegidx;
2265 u_int sr;
2266
2267 sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr);
2268 ptegidx = va_to_pteg(sr, pvo->pvo_vaddr);
2269 pteidx = moea_pvo_pte_index(pvo, ptegidx);
2270 }
2271
2272 pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7];
2273 mtx_lock(&moea_table_mutex);
2274
2275 if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) {
2276 panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no "
2277 "valid pte index", pvo);
2278 }
2279
2280 if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) {
2281 panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo "
2282 "pvo but no valid pte", pvo);
2283 }
2284
2285 if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) {
2286 if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) {
2287 panic("moea_pvo_to_pte: pvo %p has valid pte in "
2288 "moea_pteg_table %p but invalid in pvo", pvo, pt);
2289 }
2290
2291 if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF))
2292 != 0) {
2293 panic("moea_pvo_to_pte: pvo %p pte does not match "
2294 "pte %p in moea_pteg_table", pvo, pt);
2295 }
2296
2297 mtx_assert(&moea_table_mutex, MA_OWNED);
2298 return (pt);
2299 }
2300
2301 if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) {
2302 panic("moea_pvo_to_pte: pvo %p has invalid pte %p in "
2303 "moea_pteg_table but valid in pvo: %8x, %8x", pvo, pt, pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2304 }
2305
2306 mtx_unlock(&moea_table_mutex);
2307 return (NULL);
2308 }
2309
2310 /*
2311 * XXX: THIS STUFF SHOULD BE IN pte.c?
2312 */
2313 int
moea_pte_spill(vm_offset_t addr)2314 moea_pte_spill(vm_offset_t addr)
2315 {
2316 struct pvo_entry *source_pvo, *victim_pvo;
2317 struct pvo_entry *pvo;
2318 int ptegidx, i, j;
2319 u_int sr;
2320 struct pteg *pteg;
2321 struct pte *pt;
2322
2323 moea_pte_spills++;
2324
2325 sr = mfsrin(addr);
2326 ptegidx = va_to_pteg(sr, addr);
2327
2328 /*
2329 * Have to substitute some entry. Use the primary hash for this.
2330 * Use low bits of timebase as random generator.
2331 */
2332 pteg = &moea_pteg_table[ptegidx];
2333 mtx_lock(&moea_table_mutex);
2334 __asm __volatile("mftb %0" : "=r"(i));
2335 i &= 7;
2336 pt = &pteg->pt[i];
2337
2338 source_pvo = NULL;
2339 victim_pvo = NULL;
2340 LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) {
2341 /*
2342 * We need to find a pvo entry for this address.
2343 */
2344 if (source_pvo == NULL &&
2345 moea_pte_match(&pvo->pvo_pte.pte, sr, addr,
2346 pvo->pvo_pte.pte.pte_hi & PTE_HID)) {
2347 /*
2348 * Now found an entry to be spilled into the pteg.
2349 * The PTE is now valid, so we know it's active.
2350 */
2351 j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte);
2352
2353 if (j >= 0) {
2354 PVO_PTEGIDX_SET(pvo, j);
2355 moea_pte_overflow--;
2356 mtx_unlock(&moea_table_mutex);
2357 return (1);
2358 }
2359
2360 source_pvo = pvo;
2361
2362 if (victim_pvo != NULL)
2363 break;
2364 }
2365
2366 /*
2367 * We also need the pvo entry of the victim we are replacing
2368 * so save the R & C bits of the PTE.
2369 */
2370 if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL &&
2371 moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2372 victim_pvo = pvo;
2373 if (source_pvo != NULL)
2374 break;
2375 }
2376 }
2377
2378 if (source_pvo == NULL) {
2379 mtx_unlock(&moea_table_mutex);
2380 return (0);
2381 }
2382
2383 if (victim_pvo == NULL) {
2384 if ((pt->pte_hi & PTE_HID) == 0)
2385 panic("moea_pte_spill: victim p-pte (%p) has no pvo"
2386 "entry", pt);
2387
2388 /*
2389 * If this is a secondary PTE, we need to search it's primary
2390 * pvo bucket for the matching PVO.
2391 */
2392 LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask],
2393 pvo_olink) {
2394 /*
2395 * We also need the pvo entry of the victim we are
2396 * replacing so save the R & C bits of the PTE.
2397 */
2398 if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) {
2399 victim_pvo = pvo;
2400 break;
2401 }
2402 }
2403
2404 if (victim_pvo == NULL)
2405 panic("moea_pte_spill: victim s-pte (%p) has no pvo"
2406 "entry", pt);
2407 }
2408
2409 /*
2410 * We are invalidating the TLB entry for the EA we are replacing even
2411 * though it's valid. If we don't, we lose any ref/chg bit changes
2412 * contained in the TLB entry.
2413 */
2414 source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID;
2415
2416 moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2417 moea_pte_set(pt, &source_pvo->pvo_pte.pte);
2418
2419 PVO_PTEGIDX_CLR(victim_pvo);
2420 PVO_PTEGIDX_SET(source_pvo, i);
2421 moea_pte_replacements++;
2422
2423 mtx_unlock(&moea_table_mutex);
2424 return (1);
2425 }
2426
2427 static __inline struct pvo_entry *
moea_pte_spillable_ident(u_int ptegidx)2428 moea_pte_spillable_ident(u_int ptegidx)
2429 {
2430 struct pte *pt;
2431 struct pvo_entry *pvo_walk, *pvo = NULL;
2432
2433 LIST_FOREACH(pvo_walk, &moea_pvo_table[ptegidx], pvo_olink) {
2434 if (pvo_walk->pvo_vaddr & PVO_WIRED)
2435 continue;
2436
2437 if (!(pvo_walk->pvo_pte.pte.pte_hi & PTE_VALID))
2438 continue;
2439
2440 pt = moea_pvo_to_pte(pvo_walk, -1);
2441
2442 if (pt == NULL)
2443 continue;
2444
2445 pvo = pvo_walk;
2446
2447 mtx_unlock(&moea_table_mutex);
2448 if (!(pt->pte_lo & PTE_REF))
2449 return (pvo_walk);
2450 }
2451
2452 return (pvo);
2453 }
2454
2455 static int
moea_pte_insert(u_int ptegidx,struct pte * pvo_pt)2456 moea_pte_insert(u_int ptegidx, struct pte *pvo_pt)
2457 {
2458 struct pte *pt;
2459 struct pvo_entry *victim_pvo;
2460 int i;
2461 int victim_idx;
2462 u_int pteg_bkpidx = ptegidx;
2463
2464 mtx_assert(&moea_table_mutex, MA_OWNED);
2465
2466 /*
2467 * First try primary hash.
2468 */
2469 for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2470 if ((pt->pte_hi & PTE_VALID) == 0) {
2471 pvo_pt->pte_hi &= ~PTE_HID;
2472 moea_pte_set(pt, pvo_pt);
2473 return (i);
2474 }
2475 }
2476
2477 /*
2478 * Now try secondary hash.
2479 */
2480 ptegidx ^= moea_pteg_mask;
2481
2482 for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) {
2483 if ((pt->pte_hi & PTE_VALID) == 0) {
2484 pvo_pt->pte_hi |= PTE_HID;
2485 moea_pte_set(pt, pvo_pt);
2486 return (i);
2487 }
2488 }
2489
2490 /* Try again, but this time try to force a PTE out. */
2491 ptegidx = pteg_bkpidx;
2492
2493 victim_pvo = moea_pte_spillable_ident(ptegidx);
2494 if (victim_pvo == NULL) {
2495 ptegidx ^= moea_pteg_mask;
2496 victim_pvo = moea_pte_spillable_ident(ptegidx);
2497 }
2498
2499 if (victim_pvo == NULL) {
2500 panic("moea_pte_insert: overflow");
2501 return (-1);
2502 }
2503
2504 victim_idx = moea_pvo_pte_index(victim_pvo, ptegidx);
2505
2506 if (pteg_bkpidx == ptegidx)
2507 pvo_pt->pte_hi &= ~PTE_HID;
2508 else
2509 pvo_pt->pte_hi |= PTE_HID;
2510
2511 /*
2512 * Synchronize the sacrifice PTE with its PVO, then mark both
2513 * invalid. The PVO will be reused when/if the VM system comes
2514 * here after a fault.
2515 */
2516 pt = &moea_pteg_table[victim_idx >> 3].pt[victim_idx & 7];
2517
2518 if (pt->pte_hi != victim_pvo->pvo_pte.pte.pte_hi)
2519 panic("Victim PVO doesn't match PTE! PVO: %8x, PTE: %8x", victim_pvo->pvo_pte.pte.pte_hi, pt->pte_hi);
2520
2521 /*
2522 * Set the new PTE.
2523 */
2524 moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr);
2525 PVO_PTEGIDX_CLR(victim_pvo);
2526 moea_pte_overflow++;
2527 moea_pte_set(pt, pvo_pt);
2528
2529 return (victim_idx & 7);
2530 }
2531
2532 static bool
moea_query_bit(vm_page_t m,int ptebit)2533 moea_query_bit(vm_page_t m, int ptebit)
2534 {
2535 struct pvo_entry *pvo;
2536 struct pte *pt;
2537
2538 rw_assert(&pvh_global_lock, RA_WLOCKED);
2539 if (moea_attr_fetch(m) & ptebit)
2540 return (true);
2541
2542 LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2543 /*
2544 * See if we saved the bit off. If so, cache it and return
2545 * success.
2546 */
2547 if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2548 moea_attr_save(m, ptebit);
2549 return (true);
2550 }
2551 }
2552
2553 /*
2554 * No luck, now go through the hard part of looking at the PTEs
2555 * themselves. Sync so that any pending REF/CHG bits are flushed to
2556 * the PTEs.
2557 */
2558 powerpc_sync();
2559 LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2560 /*
2561 * See if this pvo has a valid PTE. if so, fetch the
2562 * REF/CHG bits from the valid PTE. If the appropriate
2563 * ptebit is set, cache it and return success.
2564 */
2565 pt = moea_pvo_to_pte(pvo, -1);
2566 if (pt != NULL) {
2567 moea_pte_synch(pt, &pvo->pvo_pte.pte);
2568 mtx_unlock(&moea_table_mutex);
2569 if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2570 moea_attr_save(m, ptebit);
2571 return (true);
2572 }
2573 }
2574 }
2575
2576 return (false);
2577 }
2578
2579 static u_int
moea_clear_bit(vm_page_t m,int ptebit)2580 moea_clear_bit(vm_page_t m, int ptebit)
2581 {
2582 u_int count;
2583 struct pvo_entry *pvo;
2584 struct pte *pt;
2585
2586 rw_assert(&pvh_global_lock, RA_WLOCKED);
2587
2588 /*
2589 * Clear the cached value.
2590 */
2591 moea_attr_clear(m, ptebit);
2592
2593 /*
2594 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so
2595 * we can reset the right ones). note that since the pvo entries and
2596 * list heads are accessed via BAT0 and are never placed in the page
2597 * table, we don't have to worry about further accesses setting the
2598 * REF/CHG bits.
2599 */
2600 powerpc_sync();
2601
2602 /*
2603 * For each pvo entry, clear the pvo's ptebit. If this pvo has a
2604 * valid pte clear the ptebit from the valid pte.
2605 */
2606 count = 0;
2607 LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) {
2608 pt = moea_pvo_to_pte(pvo, -1);
2609 if (pt != NULL) {
2610 moea_pte_synch(pt, &pvo->pvo_pte.pte);
2611 if (pvo->pvo_pte.pte.pte_lo & ptebit) {
2612 count++;
2613 moea_pte_clear(pt, PVO_VADDR(pvo), ptebit);
2614 }
2615 mtx_unlock(&moea_table_mutex);
2616 }
2617 pvo->pvo_pte.pte.pte_lo &= ~ptebit;
2618 }
2619
2620 return (count);
2621 }
2622
2623 /*
2624 * Return true if the physical range is encompassed by the battable[idx]
2625 */
2626 static int
moea_bat_mapped(int idx,vm_paddr_t pa,vm_size_t size)2627 moea_bat_mapped(int idx, vm_paddr_t pa, vm_size_t size)
2628 {
2629 u_int prot;
2630 u_int32_t start;
2631 u_int32_t end;
2632 u_int32_t bat_ble;
2633
2634 /*
2635 * Return immediately if not a valid mapping
2636 */
2637 if (!(battable[idx].batu & BAT_Vs))
2638 return (EINVAL);
2639
2640 /*
2641 * The BAT entry must be cache-inhibited, guarded, and r/w
2642 * so it can function as an i/o page
2643 */
2644 prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW);
2645 if (prot != (BAT_I|BAT_G|BAT_PP_RW))
2646 return (EPERM);
2647
2648 /*
2649 * The address should be within the BAT range. Assume that the
2650 * start address in the BAT has the correct alignment (thus
2651 * not requiring masking)
2652 */
2653 start = battable[idx].batl & BAT_PBS;
2654 bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03;
2655 end = start | (bat_ble << 15) | 0x7fff;
2656
2657 if ((pa < start) || ((pa + size) > end))
2658 return (ERANGE);
2659
2660 return (0);
2661 }
2662
2663 int
moea_dev_direct_mapped(vm_paddr_t pa,vm_size_t size)2664 moea_dev_direct_mapped(vm_paddr_t pa, vm_size_t size)
2665 {
2666 int i;
2667
2668 /*
2669 * This currently does not work for entries that
2670 * overlap 256M BAT segments.
2671 */
2672
2673 for(i = 0; i < 16; i++)
2674 if (moea_bat_mapped(i, pa, size) == 0)
2675 return (0);
2676
2677 return (EFAULT);
2678 }
2679
2680 /*
2681 * Map a set of physical memory pages into the kernel virtual
2682 * address space. Return a pointer to where it is mapped. This
2683 * routine is intended to be used for mapping device memory,
2684 * NOT real memory.
2685 */
2686 void *
moea_mapdev(vm_paddr_t pa,vm_size_t size)2687 moea_mapdev(vm_paddr_t pa, vm_size_t size)
2688 {
2689
2690 return (moea_mapdev_attr(pa, size, VM_MEMATTR_DEFAULT));
2691 }
2692
2693 void *
moea_mapdev_attr(vm_paddr_t pa,vm_size_t size,vm_memattr_t ma)2694 moea_mapdev_attr(vm_paddr_t pa, vm_size_t size, vm_memattr_t ma)
2695 {
2696 vm_offset_t va, tmpva, ppa, offset;
2697 int i;
2698
2699 ppa = trunc_page(pa);
2700 offset = pa & PAGE_MASK;
2701 size = roundup(offset + size, PAGE_SIZE);
2702
2703 /*
2704 * If the physical address lies within a valid BAT table entry,
2705 * return the 1:1 mapping. This currently doesn't work
2706 * for regions that overlap 256M BAT segments.
2707 */
2708 for (i = 0; i < 16; i++) {
2709 if (moea_bat_mapped(i, pa, size) == 0)
2710 return ((void *) pa);
2711 }
2712
2713 va = kva_alloc(size);
2714 if (!va)
2715 panic("moea_mapdev: Couldn't alloc kernel virtual memory");
2716
2717 for (tmpva = va; size > 0;) {
2718 moea_kenter_attr(tmpva, ppa, ma);
2719 tlbie(tmpva);
2720 size -= PAGE_SIZE;
2721 tmpva += PAGE_SIZE;
2722 ppa += PAGE_SIZE;
2723 }
2724
2725 return ((void *)(va + offset));
2726 }
2727
2728 void
moea_unmapdev(void * p,vm_size_t size)2729 moea_unmapdev(void *p, vm_size_t size)
2730 {
2731 vm_offset_t base, offset, va;
2732
2733 /*
2734 * If this is outside kernel virtual space, then it's a
2735 * battable entry and doesn't require unmapping
2736 */
2737 va = (vm_offset_t)p;
2738 if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) {
2739 base = trunc_page(va);
2740 offset = va & PAGE_MASK;
2741 size = roundup(offset + size, PAGE_SIZE);
2742 moea_qremove(base, atop(size));
2743 kva_free(base, size);
2744 }
2745 }
2746
2747 static void
moea_sync_icache(pmap_t pm,vm_offset_t va,vm_size_t sz)2748 moea_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz)
2749 {
2750 struct pvo_entry *pvo;
2751 vm_offset_t lim;
2752 vm_paddr_t pa;
2753 vm_size_t len;
2754
2755 PMAP_LOCK(pm);
2756 while (sz > 0) {
2757 lim = round_page(va + 1);
2758 len = MIN(lim - va, sz);
2759 pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL);
2760 if (pvo != NULL) {
2761 pa = PVO_PADDR(pvo) | (va & ADDR_POFF);
2762 moea_syncicache(pa, len);
2763 }
2764 va += len;
2765 sz -= len;
2766 }
2767 PMAP_UNLOCK(pm);
2768 }
2769
2770 void
moea_dumpsys_map(vm_paddr_t pa,size_t sz,void ** va)2771 moea_dumpsys_map(vm_paddr_t pa, size_t sz, void **va)
2772 {
2773
2774 *va = (void *)pa;
2775 }
2776
2777 extern struct dump_pa dump_map[PHYS_AVAIL_SZ + 1];
2778
2779 void
moea_scan_init(void)2780 moea_scan_init(void)
2781 {
2782 struct pvo_entry *pvo;
2783 vm_offset_t va;
2784 int i;
2785
2786 if (!do_minidump) {
2787 /* Initialize phys. segments for dumpsys(). */
2788 memset(&dump_map, 0, sizeof(dump_map));
2789 mem_regions(&pregions, &pregions_sz, ®ions, ®ions_sz);
2790 for (i = 0; i < pregions_sz; i++) {
2791 dump_map[i].pa_start = pregions[i].mr_start;
2792 dump_map[i].pa_size = pregions[i].mr_size;
2793 }
2794 return;
2795 }
2796
2797 /* Virtual segments for minidumps: */
2798 memset(&dump_map, 0, sizeof(dump_map));
2799
2800 /* 1st: kernel .data and .bss. */
2801 dump_map[0].pa_start = trunc_page((uintptr_t)_etext);
2802 dump_map[0].pa_size =
2803 round_page((uintptr_t)_end) - dump_map[0].pa_start;
2804
2805 /* 2nd: msgbuf and tables (see pmap_bootstrap()). */
2806 dump_map[1].pa_start = (vm_paddr_t)msgbufp->msg_ptr;
2807 dump_map[1].pa_size = round_page(msgbufp->msg_size);
2808
2809 /* 3rd: kernel VM. */
2810 va = dump_map[1].pa_start + dump_map[1].pa_size;
2811 /* Find start of next chunk (from va). */
2812 while (va < virtual_end) {
2813 /* Don't dump the buffer cache. */
2814 if (va >= kmi.buffer_sva && va < kmi.buffer_eva) {
2815 va = kmi.buffer_eva;
2816 continue;
2817 }
2818 pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL);
2819 if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2820 break;
2821 va += PAGE_SIZE;
2822 }
2823 if (va < virtual_end) {
2824 dump_map[2].pa_start = va;
2825 va += PAGE_SIZE;
2826 /* Find last page in chunk. */
2827 while (va < virtual_end) {
2828 /* Don't run into the buffer cache. */
2829 if (va == kmi.buffer_sva)
2830 break;
2831 pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF,
2832 NULL);
2833 if (pvo == NULL ||
2834 !(pvo->pvo_pte.pte.pte_hi & PTE_VALID))
2835 break;
2836 va += PAGE_SIZE;
2837 }
2838 dump_map[2].pa_size = va - dump_map[2].pa_start;
2839 }
2840 }
2841