mmu_oea.c revision 209369
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: head/sys/powerpc/aim/mmu_oea.c 209369 2010-06-20 16:56:48Z nwhitehorn $"); 95 96/* 97 * Manages physical address maps. 98 * 99 * In addition to hardware address maps, this module is called upon to 100 * provide software-use-only maps which may or may not be stored in the 101 * same form as hardware maps. These pseudo-maps are used to store 102 * intermediate results from copy operations to and from address spaces. 103 * 104 * Since the information managed by this module is also stored by the 105 * logical address mapping module, this module may throw away valid virtual 106 * to physical mappings at almost any time. However, invalidations of 107 * mappings must be done as requested. 108 * 109 * In order to cope with hardware architectures which make virtual to 110 * physical map invalidates expensive, this module may delay invalidate 111 * reduced protection operations until such time as they are actually 112 * necessary. This module is given full information as to which processors 113 * are currently using which maps, and to when physical maps must be made 114 * correct. 115 */ 116 117#include "opt_kstack_pages.h" 118 119#include <sys/param.h> 120#include <sys/kernel.h> 121#include <sys/ktr.h> 122#include <sys/lock.h> 123#include <sys/msgbuf.h> 124#include <sys/mutex.h> 125#include <sys/proc.h> 126#include <sys/sysctl.h> 127#include <sys/systm.h> 128#include <sys/vmmeter.h> 129 130#include <dev/ofw/openfirm.h> 131 132#include <vm/vm.h> 133#include <vm/vm_param.h> 134#include <vm/vm_kern.h> 135#include <vm/vm_page.h> 136#include <vm/vm_map.h> 137#include <vm/vm_object.h> 138#include <vm/vm_extern.h> 139#include <vm/vm_pageout.h> 140#include <vm/vm_pager.h> 141#include <vm/uma.h> 142 143#include <machine/cpu.h> 144#include <machine/platform.h> 145#include <machine/bat.h> 146#include <machine/frame.h> 147#include <machine/md_var.h> 148#include <machine/psl.h> 149#include <machine/pte.h> 150#include <machine/smp.h> 151#include <machine/sr.h> 152#include <machine/mmuvar.h> 153 154#include "mmu_if.h" 155 156#define MOEA_DEBUG 157 158#define TODO panic("%s: not implemented", __func__); 159 160#define VSID_MAKE(sr, hash) ((sr) | (((hash) & 0xfffff) << 4)) 161#define VSID_TO_SR(vsid) ((vsid) & 0xf) 162#define VSID_TO_HASH(vsid) (((vsid) >> 4) & 0xfffff) 163 164#define PVO_PTEGIDX_MASK 0x007 /* which PTEG slot */ 165#define PVO_PTEGIDX_VALID 0x008 /* slot is valid */ 166#define PVO_WIRED 0x010 /* PVO entry is wired */ 167#define PVO_MANAGED 0x020 /* PVO entry is managed */ 168#define PVO_EXECUTABLE 0x040 /* PVO entry is executable */ 169#define PVO_BOOTSTRAP 0x080 /* PVO entry allocated during 170 bootstrap */ 171#define PVO_FAKE 0x100 /* fictitious phys page */ 172#define PVO_VADDR(pvo) ((pvo)->pvo_vaddr & ~ADDR_POFF) 173#define PVO_ISEXECUTABLE(pvo) ((pvo)->pvo_vaddr & PVO_EXECUTABLE) 174#define PVO_ISFAKE(pvo) ((pvo)->pvo_vaddr & PVO_FAKE) 175#define PVO_PTEGIDX_GET(pvo) ((pvo)->pvo_vaddr & PVO_PTEGIDX_MASK) 176#define PVO_PTEGIDX_ISSET(pvo) ((pvo)->pvo_vaddr & PVO_PTEGIDX_VALID) 177#define PVO_PTEGIDX_CLR(pvo) \ 178 ((void)((pvo)->pvo_vaddr &= ~(PVO_PTEGIDX_VALID|PVO_PTEGIDX_MASK))) 179#define PVO_PTEGIDX_SET(pvo, i) \ 180 ((void)((pvo)->pvo_vaddr |= (i)|PVO_PTEGIDX_VALID)) 181 182#define MOEA_PVO_CHECK(pvo) 183 184struct ofw_map { 185 vm_offset_t om_va; 186 vm_size_t om_len; 187 vm_offset_t om_pa; 188 u_int om_mode; 189}; 190 191/* 192 * Map of physical memory regions. 193 */ 194static struct mem_region *regions; 195static struct mem_region *pregions; 196u_int phys_avail_count; 197int regions_sz, pregions_sz; 198static struct ofw_map *translations; 199 200extern struct pmap ofw_pmap; 201 202/* 203 * Lock for the pteg and pvo tables. 204 */ 205struct mtx moea_table_mutex; 206 207/* tlbie instruction synchronization */ 208static struct mtx tlbie_mtx; 209 210/* 211 * PTEG data. 212 */ 213static struct pteg *moea_pteg_table; 214u_int moea_pteg_count; 215u_int moea_pteg_mask; 216 217/* 218 * PVO data. 219 */ 220struct pvo_head *moea_pvo_table; /* pvo entries by pteg index */ 221struct pvo_head moea_pvo_kunmanaged = 222 LIST_HEAD_INITIALIZER(moea_pvo_kunmanaged); /* list of unmanaged pages */ 223struct pvo_head moea_pvo_unmanaged = 224 LIST_HEAD_INITIALIZER(moea_pvo_unmanaged); /* list of unmanaged pages */ 225 226uma_zone_t moea_upvo_zone; /* zone for pvo entries for unmanaged pages */ 227uma_zone_t moea_mpvo_zone; /* zone for pvo entries for managed pages */ 228 229#define BPVO_POOL_SIZE 32768 230static struct pvo_entry *moea_bpvo_pool; 231static int moea_bpvo_pool_index = 0; 232 233#define VSID_NBPW (sizeof(u_int32_t) * 8) 234static u_int moea_vsid_bitmap[NPMAPS / VSID_NBPW]; 235 236static boolean_t moea_initialized = FALSE; 237 238/* 239 * Statistics. 240 */ 241u_int moea_pte_valid = 0; 242u_int moea_pte_overflow = 0; 243u_int moea_pte_replacements = 0; 244u_int moea_pvo_entries = 0; 245u_int moea_pvo_enter_calls = 0; 246u_int moea_pvo_remove_calls = 0; 247u_int moea_pte_spills = 0; 248SYSCTL_INT(_machdep, OID_AUTO, moea_pte_valid, CTLFLAG_RD, &moea_pte_valid, 249 0, ""); 250SYSCTL_INT(_machdep, OID_AUTO, moea_pte_overflow, CTLFLAG_RD, 251 &moea_pte_overflow, 0, ""); 252SYSCTL_INT(_machdep, OID_AUTO, moea_pte_replacements, CTLFLAG_RD, 253 &moea_pte_replacements, 0, ""); 254SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_entries, CTLFLAG_RD, &moea_pvo_entries, 255 0, ""); 256SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_enter_calls, CTLFLAG_RD, 257 &moea_pvo_enter_calls, 0, ""); 258SYSCTL_INT(_machdep, OID_AUTO, moea_pvo_remove_calls, CTLFLAG_RD, 259 &moea_pvo_remove_calls, 0, ""); 260SYSCTL_INT(_machdep, OID_AUTO, moea_pte_spills, CTLFLAG_RD, 261 &moea_pte_spills, 0, ""); 262 263/* 264 * Allocate physical memory for use in moea_bootstrap. 265 */ 266static vm_offset_t moea_bootstrap_alloc(vm_size_t, u_int); 267 268/* 269 * PTE calls. 270 */ 271static int moea_pte_insert(u_int, struct pte *); 272 273/* 274 * PVO calls. 275 */ 276static int moea_pvo_enter(pmap_t, uma_zone_t, struct pvo_head *, 277 vm_offset_t, vm_offset_t, u_int, int); 278static void moea_pvo_remove(struct pvo_entry *, int); 279static struct pvo_entry *moea_pvo_find_va(pmap_t, vm_offset_t, int *); 280static struct pte *moea_pvo_to_pte(const struct pvo_entry *, int); 281 282/* 283 * Utility routines. 284 */ 285static void moea_enter_locked(pmap_t, vm_offset_t, vm_page_t, 286 vm_prot_t, boolean_t); 287static void moea_syncicache(vm_offset_t, vm_size_t); 288static boolean_t moea_query_bit(vm_page_t, int); 289static u_int moea_clear_bit(vm_page_t, int); 290static void moea_kremove(mmu_t, vm_offset_t); 291int moea_pte_spill(vm_offset_t); 292 293/* 294 * Kernel MMU interface 295 */ 296void moea_change_wiring(mmu_t, pmap_t, vm_offset_t, boolean_t); 297void moea_clear_modify(mmu_t, vm_page_t); 298void moea_clear_reference(mmu_t, vm_page_t); 299void moea_copy_page(mmu_t, vm_page_t, vm_page_t); 300void moea_enter(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t, boolean_t); 301void moea_enter_object(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_page_t, 302 vm_prot_t); 303void moea_enter_quick(mmu_t, pmap_t, vm_offset_t, vm_page_t, vm_prot_t); 304vm_paddr_t moea_extract(mmu_t, pmap_t, vm_offset_t); 305vm_page_t moea_extract_and_hold(mmu_t, pmap_t, vm_offset_t, vm_prot_t); 306void moea_init(mmu_t); 307boolean_t moea_is_modified(mmu_t, vm_page_t); 308boolean_t moea_is_referenced(mmu_t, vm_page_t); 309boolean_t moea_ts_referenced(mmu_t, vm_page_t); 310vm_offset_t moea_map(mmu_t, vm_offset_t *, vm_offset_t, vm_offset_t, int); 311boolean_t moea_page_exists_quick(mmu_t, pmap_t, vm_page_t); 312int moea_page_wired_mappings(mmu_t, vm_page_t); 313void moea_pinit(mmu_t, pmap_t); 314void moea_pinit0(mmu_t, pmap_t); 315void moea_protect(mmu_t, pmap_t, vm_offset_t, vm_offset_t, vm_prot_t); 316void moea_qenter(mmu_t, vm_offset_t, vm_page_t *, int); 317void moea_qremove(mmu_t, vm_offset_t, int); 318void moea_release(mmu_t, pmap_t); 319void moea_remove(mmu_t, pmap_t, vm_offset_t, vm_offset_t); 320void moea_remove_all(mmu_t, vm_page_t); 321void moea_remove_write(mmu_t, vm_page_t); 322void moea_zero_page(mmu_t, vm_page_t); 323void moea_zero_page_area(mmu_t, vm_page_t, int, int); 324void moea_zero_page_idle(mmu_t, vm_page_t); 325void moea_activate(mmu_t, struct thread *); 326void moea_deactivate(mmu_t, struct thread *); 327void moea_cpu_bootstrap(mmu_t, int); 328void moea_bootstrap(mmu_t, vm_offset_t, vm_offset_t); 329void *moea_mapdev(mmu_t, vm_offset_t, vm_size_t); 330void moea_unmapdev(mmu_t, vm_offset_t, vm_size_t); 331vm_offset_t moea_kextract(mmu_t, vm_offset_t); 332void moea_kenter(mmu_t, vm_offset_t, vm_offset_t); 333boolean_t moea_dev_direct_mapped(mmu_t, vm_offset_t, vm_size_t); 334static void moea_sync_icache(mmu_t, pmap_t, vm_offset_t, vm_size_t); 335 336static mmu_method_t moea_methods[] = { 337 MMUMETHOD(mmu_change_wiring, moea_change_wiring), 338 MMUMETHOD(mmu_clear_modify, moea_clear_modify), 339 MMUMETHOD(mmu_clear_reference, moea_clear_reference), 340 MMUMETHOD(mmu_copy_page, moea_copy_page), 341 MMUMETHOD(mmu_enter, moea_enter), 342 MMUMETHOD(mmu_enter_object, moea_enter_object), 343 MMUMETHOD(mmu_enter_quick, moea_enter_quick), 344 MMUMETHOD(mmu_extract, moea_extract), 345 MMUMETHOD(mmu_extract_and_hold, moea_extract_and_hold), 346 MMUMETHOD(mmu_init, moea_init), 347 MMUMETHOD(mmu_is_modified, moea_is_modified), 348 MMUMETHOD(mmu_is_referenced, moea_is_referenced), 349 MMUMETHOD(mmu_ts_referenced, moea_ts_referenced), 350 MMUMETHOD(mmu_map, moea_map), 351 MMUMETHOD(mmu_page_exists_quick,moea_page_exists_quick), 352 MMUMETHOD(mmu_page_wired_mappings,moea_page_wired_mappings), 353 MMUMETHOD(mmu_pinit, moea_pinit), 354 MMUMETHOD(mmu_pinit0, moea_pinit0), 355 MMUMETHOD(mmu_protect, moea_protect), 356 MMUMETHOD(mmu_qenter, moea_qenter), 357 MMUMETHOD(mmu_qremove, moea_qremove), 358 MMUMETHOD(mmu_release, moea_release), 359 MMUMETHOD(mmu_remove, moea_remove), 360 MMUMETHOD(mmu_remove_all, moea_remove_all), 361 MMUMETHOD(mmu_remove_write, moea_remove_write), 362 MMUMETHOD(mmu_sync_icache, moea_sync_icache), 363 MMUMETHOD(mmu_zero_page, moea_zero_page), 364 MMUMETHOD(mmu_zero_page_area, moea_zero_page_area), 365 MMUMETHOD(mmu_zero_page_idle, moea_zero_page_idle), 366 MMUMETHOD(mmu_activate, moea_activate), 367 MMUMETHOD(mmu_deactivate, moea_deactivate), 368 369 /* Internal interfaces */ 370 MMUMETHOD(mmu_bootstrap, moea_bootstrap), 371 MMUMETHOD(mmu_cpu_bootstrap, moea_cpu_bootstrap), 372 MMUMETHOD(mmu_mapdev, moea_mapdev), 373 MMUMETHOD(mmu_unmapdev, moea_unmapdev), 374 MMUMETHOD(mmu_kextract, moea_kextract), 375 MMUMETHOD(mmu_kenter, moea_kenter), 376 MMUMETHOD(mmu_dev_direct_mapped,moea_dev_direct_mapped), 377 378 { 0, 0 } 379}; 380 381static mmu_def_t oea_mmu = { 382 MMU_TYPE_OEA, 383 moea_methods, 384 0 385}; 386MMU_DEF(oea_mmu); 387 388static void 389tlbie(vm_offset_t va) 390{ 391 392 mtx_lock_spin(&tlbie_mtx); 393 __asm __volatile("tlbie %0" :: "r"(va)); 394 __asm __volatile("tlbsync"); 395 powerpc_sync(); 396 mtx_unlock_spin(&tlbie_mtx); 397} 398 399static void 400tlbia(void) 401{ 402 vm_offset_t va; 403 404 for (va = 0; va < 0x00040000; va += 0x00001000) { 405 __asm __volatile("tlbie %0" :: "r"(va)); 406 powerpc_sync(); 407 } 408 __asm __volatile("tlbsync"); 409 powerpc_sync(); 410} 411 412static __inline int 413va_to_sr(u_int *sr, vm_offset_t va) 414{ 415 return (sr[(uintptr_t)va >> ADDR_SR_SHFT]); 416} 417 418static __inline u_int 419va_to_pteg(u_int sr, vm_offset_t addr) 420{ 421 u_int hash; 422 423 hash = (sr & SR_VSID_MASK) ^ (((u_int)addr & ADDR_PIDX) >> 424 ADDR_PIDX_SHFT); 425 return (hash & moea_pteg_mask); 426} 427 428static __inline struct pvo_head * 429pa_to_pvoh(vm_offset_t pa, vm_page_t *pg_p) 430{ 431 struct vm_page *pg; 432 433 pg = PHYS_TO_VM_PAGE(pa); 434 435 if (pg_p != NULL) 436 *pg_p = pg; 437 438 if (pg == NULL) 439 return (&moea_pvo_unmanaged); 440 441 return (&pg->md.mdpg_pvoh); 442} 443 444static __inline struct pvo_head * 445vm_page_to_pvoh(vm_page_t m) 446{ 447 448 return (&m->md.mdpg_pvoh); 449} 450 451static __inline void 452moea_attr_clear(vm_page_t m, int ptebit) 453{ 454 455 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 456 m->md.mdpg_attrs &= ~ptebit; 457} 458 459static __inline int 460moea_attr_fetch(vm_page_t m) 461{ 462 463 return (m->md.mdpg_attrs); 464} 465 466static __inline void 467moea_attr_save(vm_page_t m, int ptebit) 468{ 469 470 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 471 m->md.mdpg_attrs |= ptebit; 472} 473 474static __inline int 475moea_pte_compare(const struct pte *pt, const struct pte *pvo_pt) 476{ 477 if (pt->pte_hi == pvo_pt->pte_hi) 478 return (1); 479 480 return (0); 481} 482 483static __inline int 484moea_pte_match(struct pte *pt, u_int sr, vm_offset_t va, int which) 485{ 486 return (pt->pte_hi & ~PTE_VALID) == 487 (((sr & SR_VSID_MASK) << PTE_VSID_SHFT) | 488 ((va >> ADDR_API_SHFT) & PTE_API) | which); 489} 490 491static __inline void 492moea_pte_create(struct pte *pt, u_int sr, vm_offset_t va, u_int pte_lo) 493{ 494 495 mtx_assert(&moea_table_mutex, MA_OWNED); 496 497 /* 498 * Construct a PTE. Default to IMB initially. Valid bit only gets 499 * set when the real pte is set in memory. 500 * 501 * Note: Don't set the valid bit for correct operation of tlb update. 502 */ 503 pt->pte_hi = ((sr & SR_VSID_MASK) << PTE_VSID_SHFT) | 504 (((va & ADDR_PIDX) >> ADDR_API_SHFT) & PTE_API); 505 pt->pte_lo = pte_lo; 506} 507 508static __inline void 509moea_pte_synch(struct pte *pt, struct pte *pvo_pt) 510{ 511 512 mtx_assert(&moea_table_mutex, MA_OWNED); 513 pvo_pt->pte_lo |= pt->pte_lo & (PTE_REF | PTE_CHG); 514} 515 516static __inline void 517moea_pte_clear(struct pte *pt, vm_offset_t va, int ptebit) 518{ 519 520 mtx_assert(&moea_table_mutex, MA_OWNED); 521 522 /* 523 * As shown in Section 7.6.3.2.3 524 */ 525 pt->pte_lo &= ~ptebit; 526 tlbie(va); 527} 528 529static __inline void 530moea_pte_set(struct pte *pt, struct pte *pvo_pt) 531{ 532 533 mtx_assert(&moea_table_mutex, MA_OWNED); 534 pvo_pt->pte_hi |= PTE_VALID; 535 536 /* 537 * Update the PTE as defined in section 7.6.3.1. 538 * Note that the REF/CHG bits are from pvo_pt and thus should havce 539 * been saved so this routine can restore them (if desired). 540 */ 541 pt->pte_lo = pvo_pt->pte_lo; 542 powerpc_sync(); 543 pt->pte_hi = pvo_pt->pte_hi; 544 powerpc_sync(); 545 moea_pte_valid++; 546} 547 548static __inline void 549moea_pte_unset(struct pte *pt, struct pte *pvo_pt, vm_offset_t va) 550{ 551 552 mtx_assert(&moea_table_mutex, MA_OWNED); 553 pvo_pt->pte_hi &= ~PTE_VALID; 554 555 /* 556 * Force the reg & chg bits back into the PTEs. 557 */ 558 powerpc_sync(); 559 560 /* 561 * Invalidate the pte. 562 */ 563 pt->pte_hi &= ~PTE_VALID; 564 565 tlbie(va); 566 567 /* 568 * Save the reg & chg bits. 569 */ 570 moea_pte_synch(pt, pvo_pt); 571 moea_pte_valid--; 572} 573 574static __inline void 575moea_pte_change(struct pte *pt, struct pte *pvo_pt, vm_offset_t va) 576{ 577 578 /* 579 * Invalidate the PTE 580 */ 581 moea_pte_unset(pt, pvo_pt, va); 582 moea_pte_set(pt, pvo_pt); 583} 584 585/* 586 * Quick sort callout for comparing memory regions. 587 */ 588static int mr_cmp(const void *a, const void *b); 589static int om_cmp(const void *a, const void *b); 590 591static int 592mr_cmp(const void *a, const void *b) 593{ 594 const struct mem_region *regiona; 595 const struct mem_region *regionb; 596 597 regiona = a; 598 regionb = b; 599 if (regiona->mr_start < regionb->mr_start) 600 return (-1); 601 else if (regiona->mr_start > regionb->mr_start) 602 return (1); 603 else 604 return (0); 605} 606 607static int 608om_cmp(const void *a, const void *b) 609{ 610 const struct ofw_map *mapa; 611 const struct ofw_map *mapb; 612 613 mapa = a; 614 mapb = b; 615 if (mapa->om_pa < mapb->om_pa) 616 return (-1); 617 else if (mapa->om_pa > mapb->om_pa) 618 return (1); 619 else 620 return (0); 621} 622 623void 624moea_cpu_bootstrap(mmu_t mmup, int ap) 625{ 626 u_int sdr; 627 int i; 628 629 if (ap) { 630 powerpc_sync(); 631 __asm __volatile("mtdbatu 0,%0" :: "r"(battable[0].batu)); 632 __asm __volatile("mtdbatl 0,%0" :: "r"(battable[0].batl)); 633 isync(); 634 __asm __volatile("mtibatu 0,%0" :: "r"(battable[0].batu)); 635 __asm __volatile("mtibatl 0,%0" :: "r"(battable[0].batl)); 636 isync(); 637 } 638 639 __asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu)); 640 __asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl)); 641 isync(); 642 643 __asm __volatile("mtibatu 1,%0" :: "r"(0)); 644 __asm __volatile("mtdbatu 2,%0" :: "r"(0)); 645 __asm __volatile("mtibatu 2,%0" :: "r"(0)); 646 __asm __volatile("mtdbatu 3,%0" :: "r"(0)); 647 __asm __volatile("mtibatu 3,%0" :: "r"(0)); 648 isync(); 649 650 for (i = 0; i < 16; i++) 651 mtsrin(i << ADDR_SR_SHFT, EMPTY_SEGMENT); 652 653 __asm __volatile("mtsr %0,%1" :: "n"(KERNEL_SR), "r"(KERNEL_SEGMENT)); 654 __asm __volatile("mtsr %0,%1" :: "n"(KERNEL2_SR), "r"(KERNEL2_SEGMENT)); 655 powerpc_sync(); 656 657 sdr = (u_int)moea_pteg_table | (moea_pteg_mask >> 10); 658 __asm __volatile("mtsdr1 %0" :: "r"(sdr)); 659 isync(); 660 661 tlbia(); 662} 663 664void 665moea_bootstrap(mmu_t mmup, vm_offset_t kernelstart, vm_offset_t kernelend) 666{ 667 ihandle_t mmui; 668 phandle_t chosen, mmu; 669 int sz; 670 int i, j; 671 int ofw_mappings; 672 vm_size_t size, physsz, hwphyssz; 673 vm_offset_t pa, va, off; 674 void *dpcpu; 675 register_t msr; 676 677 /* 678 * Set up BAT0 to map the lowest 256 MB area 679 */ 680 battable[0x0].batl = BATL(0x00000000, BAT_M, BAT_PP_RW); 681 battable[0x0].batu = BATU(0x00000000, BAT_BL_256M, BAT_Vs); 682 683 /* 684 * Map PCI memory space. 685 */ 686 battable[0x8].batl = BATL(0x80000000, BAT_I|BAT_G, BAT_PP_RW); 687 battable[0x8].batu = BATU(0x80000000, BAT_BL_256M, BAT_Vs); 688 689 battable[0x9].batl = BATL(0x90000000, BAT_I|BAT_G, BAT_PP_RW); 690 battable[0x9].batu = BATU(0x90000000, BAT_BL_256M, BAT_Vs); 691 692 battable[0xa].batl = BATL(0xa0000000, BAT_I|BAT_G, BAT_PP_RW); 693 battable[0xa].batu = BATU(0xa0000000, BAT_BL_256M, BAT_Vs); 694 695 battable[0xb].batl = BATL(0xb0000000, BAT_I|BAT_G, BAT_PP_RW); 696 battable[0xb].batu = BATU(0xb0000000, BAT_BL_256M, BAT_Vs); 697 698 /* 699 * Map obio devices. 700 */ 701 battable[0xf].batl = BATL(0xf0000000, BAT_I|BAT_G, BAT_PP_RW); 702 battable[0xf].batu = BATU(0xf0000000, BAT_BL_256M, BAT_Vs); 703 704 /* 705 * Use an IBAT and a DBAT to map the bottom segment of memory 706 * where we are. Turn off instruction relocation temporarily 707 * to prevent faults while reprogramming the IBAT. 708 */ 709 msr = mfmsr(); 710 mtmsr(msr & ~PSL_IR); 711 __asm (".balign 32; \n" 712 "mtibatu 0,%0; mtibatl 0,%1; isync; \n" 713 "mtdbatu 0,%0; mtdbatl 0,%1; isync" 714 :: "r"(battable[0].batu), "r"(battable[0].batl)); 715 mtmsr(msr); 716 717 /* map pci space */ 718 __asm __volatile("mtdbatu 1,%0" :: "r"(battable[8].batu)); 719 __asm __volatile("mtdbatl 1,%0" :: "r"(battable[8].batl)); 720 isync(); 721 722 /* set global direct map flag */ 723 hw_direct_map = 1; 724 725 mem_regions(&pregions, &pregions_sz, ®ions, ®ions_sz); 726 CTR0(KTR_PMAP, "moea_bootstrap: physical memory"); 727 728 qsort(pregions, pregions_sz, sizeof(*pregions), mr_cmp); 729 for (i = 0; i < pregions_sz; i++) { 730 vm_offset_t pa; 731 vm_offset_t end; 732 733 CTR3(KTR_PMAP, "physregion: %#x - %#x (%#x)", 734 pregions[i].mr_start, 735 pregions[i].mr_start + pregions[i].mr_size, 736 pregions[i].mr_size); 737 /* 738 * Install entries into the BAT table to allow all 739 * of physmem to be convered by on-demand BAT entries. 740 * The loop will sometimes set the same battable element 741 * twice, but that's fine since they won't be used for 742 * a while yet. 743 */ 744 pa = pregions[i].mr_start & 0xf0000000; 745 end = pregions[i].mr_start + pregions[i].mr_size; 746 do { 747 u_int n = pa >> ADDR_SR_SHFT; 748 749 battable[n].batl = BATL(pa, BAT_M, BAT_PP_RW); 750 battable[n].batu = BATU(pa, BAT_BL_256M, BAT_Vs); 751 pa += SEGMENT_LENGTH; 752 } while (pa < end); 753 } 754 755 if (sizeof(phys_avail)/sizeof(phys_avail[0]) < regions_sz) 756 panic("moea_bootstrap: phys_avail too small"); 757 qsort(regions, regions_sz, sizeof(*regions), mr_cmp); 758 phys_avail_count = 0; 759 physsz = 0; 760 hwphyssz = 0; 761 TUNABLE_ULONG_FETCH("hw.physmem", (u_long *) &hwphyssz); 762 for (i = 0, j = 0; i < regions_sz; i++, j += 2) { 763 CTR3(KTR_PMAP, "region: %#x - %#x (%#x)", regions[i].mr_start, 764 regions[i].mr_start + regions[i].mr_size, 765 regions[i].mr_size); 766 if (hwphyssz != 0 && 767 (physsz + regions[i].mr_size) >= hwphyssz) { 768 if (physsz < hwphyssz) { 769 phys_avail[j] = regions[i].mr_start; 770 phys_avail[j + 1] = regions[i].mr_start + 771 hwphyssz - physsz; 772 physsz = hwphyssz; 773 phys_avail_count++; 774 } 775 break; 776 } 777 phys_avail[j] = regions[i].mr_start; 778 phys_avail[j + 1] = regions[i].mr_start + regions[i].mr_size; 779 phys_avail_count++; 780 physsz += regions[i].mr_size; 781 } 782 physmem = btoc(physsz); 783 784 /* 785 * Allocate PTEG table. 786 */ 787#ifdef PTEGCOUNT 788 moea_pteg_count = PTEGCOUNT; 789#else 790 moea_pteg_count = 0x1000; 791 792 while (moea_pteg_count < physmem) 793 moea_pteg_count <<= 1; 794 795 moea_pteg_count >>= 1; 796#endif /* PTEGCOUNT */ 797 798 size = moea_pteg_count * sizeof(struct pteg); 799 CTR2(KTR_PMAP, "moea_bootstrap: %d PTEGs, %d bytes", moea_pteg_count, 800 size); 801 moea_pteg_table = (struct pteg *)moea_bootstrap_alloc(size, size); 802 CTR1(KTR_PMAP, "moea_bootstrap: PTEG table at %p", moea_pteg_table); 803 bzero((void *)moea_pteg_table, moea_pteg_count * sizeof(struct pteg)); 804 moea_pteg_mask = moea_pteg_count - 1; 805 806 /* 807 * Allocate pv/overflow lists. 808 */ 809 size = sizeof(struct pvo_head) * moea_pteg_count; 810 moea_pvo_table = (struct pvo_head *)moea_bootstrap_alloc(size, 811 PAGE_SIZE); 812 CTR1(KTR_PMAP, "moea_bootstrap: PVO table at %p", moea_pvo_table); 813 for (i = 0; i < moea_pteg_count; i++) 814 LIST_INIT(&moea_pvo_table[i]); 815 816 /* 817 * Initialize the lock that synchronizes access to the pteg and pvo 818 * tables. 819 */ 820 mtx_init(&moea_table_mutex, "pmap table", NULL, MTX_DEF | 821 MTX_RECURSE); 822 823 mtx_init(&tlbie_mtx, "tlbie", NULL, MTX_SPIN); 824 825 /* 826 * Initialise the unmanaged pvo pool. 827 */ 828 moea_bpvo_pool = (struct pvo_entry *)moea_bootstrap_alloc( 829 BPVO_POOL_SIZE*sizeof(struct pvo_entry), 0); 830 moea_bpvo_pool_index = 0; 831 832 /* 833 * Make sure kernel vsid is allocated as well as VSID 0. 834 */ 835 moea_vsid_bitmap[(KERNEL_VSIDBITS & (NPMAPS - 1)) / VSID_NBPW] 836 |= 1 << (KERNEL_VSIDBITS % VSID_NBPW); 837 moea_vsid_bitmap[0] |= 1; 838 839 /* 840 * Set up the Open Firmware pmap and add it's mappings. 841 */ 842 moea_pinit(mmup, &ofw_pmap); 843 ofw_pmap.pm_sr[KERNEL_SR] = KERNEL_SEGMENT; 844 ofw_pmap.pm_sr[KERNEL2_SR] = KERNEL2_SEGMENT; 845 if ((chosen = OF_finddevice("/chosen")) == -1) 846 panic("moea_bootstrap: can't find /chosen"); 847 OF_getprop(chosen, "mmu", &mmui, 4); 848 if ((mmu = OF_instance_to_package(mmui)) == -1) 849 panic("moea_bootstrap: can't get mmu package"); 850 if ((sz = OF_getproplen(mmu, "translations")) == -1) 851 panic("moea_bootstrap: can't get ofw translation count"); 852 translations = NULL; 853 for (i = 0; phys_avail[i] != 0; i += 2) { 854 if (phys_avail[i + 1] >= sz) { 855 translations = (struct ofw_map *)phys_avail[i]; 856 break; 857 } 858 } 859 if (translations == NULL) 860 panic("moea_bootstrap: no space to copy translations"); 861 bzero(translations, sz); 862 if (OF_getprop(mmu, "translations", translations, sz) == -1) 863 panic("moea_bootstrap: can't get ofw translations"); 864 CTR0(KTR_PMAP, "moea_bootstrap: translations"); 865 sz /= sizeof(*translations); 866 qsort(translations, sz, sizeof (*translations), om_cmp); 867 for (i = 0, ofw_mappings = 0; i < sz; i++) { 868 CTR3(KTR_PMAP, "translation: pa=%#x va=%#x len=%#x", 869 translations[i].om_pa, translations[i].om_va, 870 translations[i].om_len); 871 872 /* 873 * If the mapping is 1:1, let the RAM and device on-demand 874 * BAT tables take care of the translation. 875 */ 876 if (translations[i].om_va == translations[i].om_pa) 877 continue; 878 879 /* Enter the pages */ 880 for (off = 0; off < translations[i].om_len; off += PAGE_SIZE) { 881 struct vm_page m; 882 883 m.phys_addr = translations[i].om_pa + off; 884 m.oflags = VPO_BUSY; 885 PMAP_LOCK(&ofw_pmap); 886 moea_enter_locked(&ofw_pmap, 887 translations[i].om_va + off, &m, 888 VM_PROT_ALL, 1); 889 PMAP_UNLOCK(&ofw_pmap); 890 ofw_mappings++; 891 } 892 } 893 894 /* 895 * Calculate the last available physical address. 896 */ 897 for (i = 0; phys_avail[i + 2] != 0; i += 2) 898 ; 899 Maxmem = powerpc_btop(phys_avail[i + 1]); 900 901 /* 902 * Initialize the kernel pmap (which is statically allocated). 903 */ 904 PMAP_LOCK_INIT(kernel_pmap); 905 for (i = 0; i < 16; i++) { 906 kernel_pmap->pm_sr[i] = EMPTY_SEGMENT; 907 } 908 kernel_pmap->pm_sr[KERNEL_SR] = KERNEL_SEGMENT; 909 kernel_pmap->pm_sr[KERNEL2_SR] = KERNEL2_SEGMENT; 910 kernel_pmap->pm_active = ~0; 911 912 moea_cpu_bootstrap(mmup,0); 913 914 pmap_bootstrapped++; 915 916 /* 917 * Set the start and end of kva. 918 */ 919 virtual_avail = VM_MIN_KERNEL_ADDRESS; 920 virtual_end = VM_MAX_SAFE_KERNEL_ADDRESS; 921 922 /* 923 * Allocate a kernel stack with a guard page for thread0 and map it 924 * into the kernel page map. 925 */ 926 pa = moea_bootstrap_alloc(KSTACK_PAGES * PAGE_SIZE, PAGE_SIZE); 927 va = virtual_avail + KSTACK_GUARD_PAGES * PAGE_SIZE; 928 virtual_avail = va + KSTACK_PAGES * PAGE_SIZE; 929 CTR2(KTR_PMAP, "moea_bootstrap: kstack0 at %#x (%#x)", pa, va); 930 thread0.td_kstack = va; 931 thread0.td_kstack_pages = KSTACK_PAGES; 932 for (i = 0; i < KSTACK_PAGES; i++) { 933 moea_kenter(mmup, va, pa); 934 pa += PAGE_SIZE; 935 va += PAGE_SIZE; 936 } 937 938 /* 939 * Allocate virtual address space for the message buffer. 940 */ 941 pa = msgbuf_phys = moea_bootstrap_alloc(MSGBUF_SIZE, PAGE_SIZE); 942 msgbufp = (struct msgbuf *)virtual_avail; 943 va = virtual_avail; 944 virtual_avail += round_page(MSGBUF_SIZE); 945 while (va < virtual_avail) { 946 moea_kenter(mmup, va, pa); 947 pa += PAGE_SIZE; 948 va += PAGE_SIZE; 949 } 950 951 /* 952 * Allocate virtual address space for the dynamic percpu area. 953 */ 954 pa = moea_bootstrap_alloc(DPCPU_SIZE, PAGE_SIZE); 955 dpcpu = (void *)virtual_avail; 956 va = virtual_avail; 957 virtual_avail += DPCPU_SIZE; 958 while (va < virtual_avail) { 959 moea_kenter(mmup, va, pa); 960 pa += PAGE_SIZE; 961 va += PAGE_SIZE; 962 } 963 dpcpu_init(dpcpu, 0); 964} 965 966/* 967 * Activate a user pmap. The pmap must be activated before it's address 968 * space can be accessed in any way. 969 */ 970void 971moea_activate(mmu_t mmu, struct thread *td) 972{ 973 pmap_t pm, pmr; 974 975 /* 976 * Load all the data we need up front to encourage the compiler to 977 * not issue any loads while we have interrupts disabled below. 978 */ 979 pm = &td->td_proc->p_vmspace->vm_pmap; 980 pmr = pm->pmap_phys; 981 982 pm->pm_active |= PCPU_GET(cpumask); 983 PCPU_SET(curpmap, pmr); 984} 985 986void 987moea_deactivate(mmu_t mmu, struct thread *td) 988{ 989 pmap_t pm; 990 991 pm = &td->td_proc->p_vmspace->vm_pmap; 992 pm->pm_active &= ~PCPU_GET(cpumask); 993 PCPU_SET(curpmap, NULL); 994} 995 996void 997moea_change_wiring(mmu_t mmu, pmap_t pm, vm_offset_t va, boolean_t wired) 998{ 999 struct pvo_entry *pvo; 1000 1001 PMAP_LOCK(pm); 1002 pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL); 1003 1004 if (pvo != NULL) { 1005 if (wired) { 1006 if ((pvo->pvo_vaddr & PVO_WIRED) == 0) 1007 pm->pm_stats.wired_count++; 1008 pvo->pvo_vaddr |= PVO_WIRED; 1009 } else { 1010 if ((pvo->pvo_vaddr & PVO_WIRED) != 0) 1011 pm->pm_stats.wired_count--; 1012 pvo->pvo_vaddr &= ~PVO_WIRED; 1013 } 1014 } 1015 PMAP_UNLOCK(pm); 1016} 1017 1018void 1019moea_copy_page(mmu_t mmu, vm_page_t msrc, vm_page_t mdst) 1020{ 1021 vm_offset_t dst; 1022 vm_offset_t src; 1023 1024 dst = VM_PAGE_TO_PHYS(mdst); 1025 src = VM_PAGE_TO_PHYS(msrc); 1026 1027 kcopy((void *)src, (void *)dst, PAGE_SIZE); 1028} 1029 1030/* 1031 * Zero a page of physical memory by temporarily mapping it into the tlb. 1032 */ 1033void 1034moea_zero_page(mmu_t mmu, vm_page_t m) 1035{ 1036 vm_offset_t pa = VM_PAGE_TO_PHYS(m); 1037 void *va = (void *)pa; 1038 1039 bzero(va, PAGE_SIZE); 1040} 1041 1042void 1043moea_zero_page_area(mmu_t mmu, vm_page_t m, int off, int size) 1044{ 1045 vm_offset_t pa = VM_PAGE_TO_PHYS(m); 1046 void *va = (void *)(pa + off); 1047 1048 bzero(va, size); 1049} 1050 1051void 1052moea_zero_page_idle(mmu_t mmu, vm_page_t m) 1053{ 1054 vm_offset_t pa = VM_PAGE_TO_PHYS(m); 1055 void *va = (void *)pa; 1056 1057 bzero(va, PAGE_SIZE); 1058} 1059 1060/* 1061 * Map the given physical page at the specified virtual address in the 1062 * target pmap with the protection requested. If specified the page 1063 * will be wired down. 1064 */ 1065void 1066moea_enter(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1067 boolean_t wired) 1068{ 1069 1070 vm_page_lock_queues(); 1071 PMAP_LOCK(pmap); 1072 moea_enter_locked(pmap, va, m, prot, wired); 1073 vm_page_unlock_queues(); 1074 PMAP_UNLOCK(pmap); 1075} 1076 1077/* 1078 * Map the given physical page at the specified virtual address in the 1079 * target pmap with the protection requested. If specified the page 1080 * will be wired down. 1081 * 1082 * The page queues and pmap must be locked. 1083 */ 1084static void 1085moea_enter_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, 1086 boolean_t wired) 1087{ 1088 struct pvo_head *pvo_head; 1089 uma_zone_t zone; 1090 vm_page_t pg; 1091 u_int pte_lo, pvo_flags, was_exec, i; 1092 int error; 1093 1094 if (!moea_initialized) { 1095 pvo_head = &moea_pvo_kunmanaged; 1096 zone = moea_upvo_zone; 1097 pvo_flags = 0; 1098 pg = NULL; 1099 was_exec = PTE_EXEC; 1100 } else { 1101 pvo_head = vm_page_to_pvoh(m); 1102 pg = m; 1103 zone = moea_mpvo_zone; 1104 pvo_flags = PVO_MANAGED; 1105 was_exec = 0; 1106 } 1107 if (pmap_bootstrapped) 1108 mtx_assert(&vm_page_queue_mtx, MA_OWNED); 1109 PMAP_LOCK_ASSERT(pmap, MA_OWNED); 1110 KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) != 0 || 1111 (m->oflags & VPO_BUSY) != 0 || VM_OBJECT_LOCKED(m->object), 1112 ("moea_enter_locked: page %p is not busy", m)); 1113 1114 /* XXX change the pvo head for fake pages */ 1115 if ((m->flags & PG_FICTITIOUS) == PG_FICTITIOUS) { 1116 pvo_flags &= ~PVO_MANAGED; 1117 pvo_head = &moea_pvo_kunmanaged; 1118 zone = moea_upvo_zone; 1119 } 1120 1121 /* 1122 * If this is a managed page, and it's the first reference to the page, 1123 * clear the execness of the page. Otherwise fetch the execness. 1124 */ 1125 if ((pg != NULL) && ((m->flags & PG_FICTITIOUS) == 0)) { 1126 if (LIST_EMPTY(pvo_head)) { 1127 moea_attr_clear(pg, PTE_EXEC); 1128 } else { 1129 was_exec = moea_attr_fetch(pg) & PTE_EXEC; 1130 } 1131 } 1132 1133 /* 1134 * Assume the page is cache inhibited and access is guarded unless 1135 * it's in our available memory array. 1136 */ 1137 pte_lo = PTE_I | PTE_G; 1138 for (i = 0; i < pregions_sz; i++) { 1139 if ((VM_PAGE_TO_PHYS(m) >= pregions[i].mr_start) && 1140 (VM_PAGE_TO_PHYS(m) < 1141 (pregions[i].mr_start + pregions[i].mr_size))) { 1142 pte_lo = PTE_M; 1143 break; 1144 } 1145 } 1146 1147 if (prot & VM_PROT_WRITE) { 1148 pte_lo |= PTE_BW; 1149 if (pmap_bootstrapped && 1150 (m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0) 1151 vm_page_flag_set(m, PG_WRITEABLE); 1152 } else 1153 pte_lo |= PTE_BR; 1154 1155 if (prot & VM_PROT_EXECUTE) 1156 pvo_flags |= PVO_EXECUTABLE; 1157 1158 if (wired) 1159 pvo_flags |= PVO_WIRED; 1160 1161 if ((m->flags & PG_FICTITIOUS) != 0) 1162 pvo_flags |= PVO_FAKE; 1163 1164 error = moea_pvo_enter(pmap, zone, pvo_head, va, VM_PAGE_TO_PHYS(m), 1165 pte_lo, pvo_flags); 1166 1167 /* 1168 * Flush the real page from the instruction cache if this page is 1169 * mapped executable and cacheable and was not previously mapped (or 1170 * was not mapped executable). 1171 */ 1172 if (error == 0 && (pvo_flags & PVO_EXECUTABLE) && 1173 (pte_lo & PTE_I) == 0 && was_exec == 0) { 1174 /* 1175 * Flush the real memory from the cache. 1176 */ 1177 moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE); 1178 if (pg != NULL) 1179 moea_attr_save(pg, PTE_EXEC); 1180 } 1181 1182 /* XXX syncicache always until problems are sorted */ 1183 moea_syncicache(VM_PAGE_TO_PHYS(m), PAGE_SIZE); 1184} 1185 1186/* 1187 * Maps a sequence of resident pages belonging to the same object. 1188 * The sequence begins with the given page m_start. This page is 1189 * mapped at the given virtual address start. Each subsequent page is 1190 * mapped at a virtual address that is offset from start by the same 1191 * amount as the page is offset from m_start within the object. The 1192 * last page in the sequence is the page with the largest offset from 1193 * m_start that can be mapped at a virtual address less than the given 1194 * virtual address end. Not every virtual page between start and end 1195 * is mapped; only those for which a resident page exists with the 1196 * corresponding offset from m_start are mapped. 1197 */ 1198void 1199moea_enter_object(mmu_t mmu, pmap_t pm, vm_offset_t start, vm_offset_t end, 1200 vm_page_t m_start, vm_prot_t prot) 1201{ 1202 vm_page_t m; 1203 vm_pindex_t diff, psize; 1204 1205 psize = atop(end - start); 1206 m = m_start; 1207 vm_page_lock_queues(); 1208 PMAP_LOCK(pm); 1209 while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { 1210 moea_enter_locked(pm, start + ptoa(diff), m, prot & 1211 (VM_PROT_READ | VM_PROT_EXECUTE), FALSE); 1212 m = TAILQ_NEXT(m, listq); 1213 } 1214 vm_page_unlock_queues(); 1215 PMAP_UNLOCK(pm); 1216} 1217 1218void 1219moea_enter_quick(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_page_t m, 1220 vm_prot_t prot) 1221{ 1222 1223 vm_page_lock_queues(); 1224 PMAP_LOCK(pm); 1225 moea_enter_locked(pm, va, m, prot & (VM_PROT_READ | VM_PROT_EXECUTE), 1226 FALSE); 1227 vm_page_unlock_queues(); 1228 PMAP_UNLOCK(pm); 1229} 1230 1231vm_paddr_t 1232moea_extract(mmu_t mmu, pmap_t pm, vm_offset_t va) 1233{ 1234 struct pvo_entry *pvo; 1235 vm_paddr_t pa; 1236 1237 PMAP_LOCK(pm); 1238 pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL); 1239 if (pvo == NULL) 1240 pa = 0; 1241 else 1242 pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF); 1243 PMAP_UNLOCK(pm); 1244 return (pa); 1245} 1246 1247/* 1248 * Atomically extract and hold the physical page with the given 1249 * pmap and virtual address pair if that mapping permits the given 1250 * protection. 1251 */ 1252vm_page_t 1253moea_extract_and_hold(mmu_t mmu, pmap_t pmap, vm_offset_t va, vm_prot_t prot) 1254{ 1255 struct pvo_entry *pvo; 1256 vm_page_t m; 1257 vm_paddr_t pa; 1258 1259 m = NULL; 1260 pa = 0; 1261 PMAP_LOCK(pmap); 1262retry: 1263 pvo = moea_pvo_find_va(pmap, va & ~ADDR_POFF, NULL); 1264 if (pvo != NULL && (pvo->pvo_pte.pte.pte_hi & PTE_VALID) && 1265 ((pvo->pvo_pte.pte.pte_lo & PTE_PP) == PTE_RW || 1266 (prot & VM_PROT_WRITE) == 0)) { 1267 if (vm_page_pa_tryrelock(pmap, pvo->pvo_pte.pte.pte_lo & PTE_RPGN, &pa)) 1268 goto retry; 1269 m = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN); 1270 vm_page_hold(m); 1271 } 1272 PA_UNLOCK_COND(pa); 1273 PMAP_UNLOCK(pmap); 1274 return (m); 1275} 1276 1277void 1278moea_init(mmu_t mmu) 1279{ 1280 1281 moea_upvo_zone = uma_zcreate("UPVO entry", sizeof (struct pvo_entry), 1282 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 1283 UMA_ZONE_VM | UMA_ZONE_NOFREE); 1284 moea_mpvo_zone = uma_zcreate("MPVO entry", sizeof(struct pvo_entry), 1285 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 1286 UMA_ZONE_VM | UMA_ZONE_NOFREE); 1287 moea_initialized = TRUE; 1288} 1289 1290boolean_t 1291moea_is_referenced(mmu_t mmu, vm_page_t m) 1292{ 1293 1294 KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0, 1295 ("moea_is_referenced: page %p is not managed", m)); 1296 return (moea_query_bit(m, PTE_REF)); 1297} 1298 1299boolean_t 1300moea_is_modified(mmu_t mmu, vm_page_t m) 1301{ 1302 1303 KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0, 1304 ("moea_is_modified: page %p is not managed", m)); 1305 1306 /* 1307 * If the page is not VPO_BUSY, then PG_WRITEABLE cannot be 1308 * concurrently set while the object is locked. Thus, if PG_WRITEABLE 1309 * is clear, no PTEs can have PTE_CHG set. 1310 */ 1311 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1312 if ((m->oflags & VPO_BUSY) == 0 && 1313 (m->flags & PG_WRITEABLE) == 0) 1314 return (FALSE); 1315 return (moea_query_bit(m, PTE_CHG)); 1316} 1317 1318void 1319moea_clear_reference(mmu_t mmu, vm_page_t m) 1320{ 1321 1322 KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0, 1323 ("moea_clear_reference: page %p is not managed", m)); 1324 moea_clear_bit(m, PTE_REF); 1325} 1326 1327void 1328moea_clear_modify(mmu_t mmu, vm_page_t m) 1329{ 1330 1331 KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0, 1332 ("moea_clear_modify: page %p is not managed", m)); 1333 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1334 KASSERT((m->oflags & VPO_BUSY) == 0, 1335 ("moea_clear_modify: page %p is busy", m)); 1336 1337 /* 1338 * If the page is not PG_WRITEABLE, then no PTEs can have PTE_CHG 1339 * set. If the object containing the page is locked and the page is 1340 * not VPO_BUSY, then PG_WRITEABLE cannot be concurrently set. 1341 */ 1342 if ((m->flags & PG_WRITEABLE) == 0) 1343 return; 1344 moea_clear_bit(m, PTE_CHG); 1345} 1346 1347/* 1348 * Clear the write and modified bits in each of the given page's mappings. 1349 */ 1350void 1351moea_remove_write(mmu_t mmu, vm_page_t m) 1352{ 1353 struct pvo_entry *pvo; 1354 struct pte *pt; 1355 pmap_t pmap; 1356 u_int lo; 1357 1358 KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0, 1359 ("moea_remove_write: page %p is not managed", m)); 1360 1361 /* 1362 * If the page is not VPO_BUSY, then PG_WRITEABLE cannot be set by 1363 * another thread while the object is locked. Thus, if PG_WRITEABLE 1364 * is clear, no page table entries need updating. 1365 */ 1366 VM_OBJECT_LOCK_ASSERT(m->object, MA_OWNED); 1367 if ((m->oflags & VPO_BUSY) == 0 && 1368 (m->flags & PG_WRITEABLE) == 0) 1369 return; 1370 vm_page_lock_queues(); 1371 lo = moea_attr_fetch(m); 1372 powerpc_sync(); 1373 LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) { 1374 pmap = pvo->pvo_pmap; 1375 PMAP_LOCK(pmap); 1376 if ((pvo->pvo_pte.pte.pte_lo & PTE_PP) != PTE_BR) { 1377 pt = moea_pvo_to_pte(pvo, -1); 1378 pvo->pvo_pte.pte.pte_lo &= ~PTE_PP; 1379 pvo->pvo_pte.pte.pte_lo |= PTE_BR; 1380 if (pt != NULL) { 1381 moea_pte_synch(pt, &pvo->pvo_pte.pte); 1382 lo |= pvo->pvo_pte.pte.pte_lo; 1383 pvo->pvo_pte.pte.pte_lo &= ~PTE_CHG; 1384 moea_pte_change(pt, &pvo->pvo_pte.pte, 1385 pvo->pvo_vaddr); 1386 mtx_unlock(&moea_table_mutex); 1387 } 1388 } 1389 PMAP_UNLOCK(pmap); 1390 } 1391 if ((lo & PTE_CHG) != 0) { 1392 moea_attr_clear(m, PTE_CHG); 1393 vm_page_dirty(m); 1394 } 1395 vm_page_flag_clear(m, PG_WRITEABLE); 1396 vm_page_unlock_queues(); 1397} 1398 1399/* 1400 * moea_ts_referenced: 1401 * 1402 * Return a count of reference bits for a page, clearing those bits. 1403 * It is not necessary for every reference bit to be cleared, but it 1404 * is necessary that 0 only be returned when there are truly no 1405 * reference bits set. 1406 * 1407 * XXX: The exact number of bits to check and clear is a matter that 1408 * should be tested and standardized at some point in the future for 1409 * optimal aging of shared pages. 1410 */ 1411boolean_t 1412moea_ts_referenced(mmu_t mmu, vm_page_t m) 1413{ 1414 1415 KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0, 1416 ("moea_ts_referenced: page %p is not managed", m)); 1417 return (moea_clear_bit(m, PTE_REF)); 1418} 1419 1420/* 1421 * Map a wired page into kernel virtual address space. 1422 */ 1423void 1424moea_kenter(mmu_t mmu, vm_offset_t va, vm_offset_t pa) 1425{ 1426 u_int pte_lo; 1427 int error; 1428 int i; 1429 1430#if 0 1431 if (va < VM_MIN_KERNEL_ADDRESS) 1432 panic("moea_kenter: attempt to enter non-kernel address %#x", 1433 va); 1434#endif 1435 1436 pte_lo = PTE_I | PTE_G; 1437 for (i = 0; i < pregions_sz; i++) { 1438 if ((pa >= pregions[i].mr_start) && 1439 (pa < (pregions[i].mr_start + pregions[i].mr_size))) { 1440 pte_lo = PTE_M; 1441 break; 1442 } 1443 } 1444 1445 PMAP_LOCK(kernel_pmap); 1446 error = moea_pvo_enter(kernel_pmap, moea_upvo_zone, 1447 &moea_pvo_kunmanaged, va, pa, pte_lo, PVO_WIRED); 1448 1449 if (error != 0 && error != ENOENT) 1450 panic("moea_kenter: failed to enter va %#x pa %#x: %d", va, 1451 pa, error); 1452 1453 /* 1454 * Flush the real memory from the instruction cache. 1455 */ 1456 if ((pte_lo & (PTE_I | PTE_G)) == 0) { 1457 moea_syncicache(pa, PAGE_SIZE); 1458 } 1459 PMAP_UNLOCK(kernel_pmap); 1460} 1461 1462/* 1463 * Extract the physical page address associated with the given kernel virtual 1464 * address. 1465 */ 1466vm_offset_t 1467moea_kextract(mmu_t mmu, vm_offset_t va) 1468{ 1469 struct pvo_entry *pvo; 1470 vm_paddr_t pa; 1471 1472 /* 1473 * Allow direct mappings on 32-bit OEA 1474 */ 1475 if (va < VM_MIN_KERNEL_ADDRESS) { 1476 return (va); 1477 } 1478 1479 PMAP_LOCK(kernel_pmap); 1480 pvo = moea_pvo_find_va(kernel_pmap, va & ~ADDR_POFF, NULL); 1481 KASSERT(pvo != NULL, ("moea_kextract: no addr found")); 1482 pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | (va & ADDR_POFF); 1483 PMAP_UNLOCK(kernel_pmap); 1484 return (pa); 1485} 1486 1487/* 1488 * Remove a wired page from kernel virtual address space. 1489 */ 1490void 1491moea_kremove(mmu_t mmu, vm_offset_t va) 1492{ 1493 1494 moea_remove(mmu, kernel_pmap, va, va + PAGE_SIZE); 1495} 1496 1497/* 1498 * Map a range of physical addresses into kernel virtual address space. 1499 * 1500 * The value passed in *virt is a suggested virtual address for the mapping. 1501 * Architectures which can support a direct-mapped physical to virtual region 1502 * can return the appropriate address within that region, leaving '*virt' 1503 * unchanged. We cannot and therefore do not; *virt is updated with the 1504 * first usable address after the mapped region. 1505 */ 1506vm_offset_t 1507moea_map(mmu_t mmu, vm_offset_t *virt, vm_offset_t pa_start, 1508 vm_offset_t pa_end, int prot) 1509{ 1510 vm_offset_t sva, va; 1511 1512 sva = *virt; 1513 va = sva; 1514 for (; pa_start < pa_end; pa_start += PAGE_SIZE, va += PAGE_SIZE) 1515 moea_kenter(mmu, va, pa_start); 1516 *virt = va; 1517 return (sva); 1518} 1519 1520/* 1521 * Returns true if the pmap's pv is one of the first 1522 * 16 pvs linked to from this page. This count may 1523 * be changed upwards or downwards in the future; it 1524 * is only necessary that true be returned for a small 1525 * subset of pmaps for proper page aging. 1526 */ 1527boolean_t 1528moea_page_exists_quick(mmu_t mmu, pmap_t pmap, vm_page_t m) 1529{ 1530 int loops; 1531 struct pvo_entry *pvo; 1532 boolean_t rv; 1533 1534 KASSERT((m->flags & (PG_FICTITIOUS | PG_UNMANAGED)) == 0, 1535 ("moea_page_exists_quick: page %p is not managed", m)); 1536 loops = 0; 1537 rv = FALSE; 1538 vm_page_lock_queues(); 1539 LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) { 1540 if (pvo->pvo_pmap == pmap) { 1541 rv = TRUE; 1542 break; 1543 } 1544 if (++loops >= 16) 1545 break; 1546 } 1547 vm_page_unlock_queues(); 1548 return (rv); 1549} 1550 1551/* 1552 * Return the number of managed mappings to the given physical page 1553 * that are wired. 1554 */ 1555int 1556moea_page_wired_mappings(mmu_t mmu, vm_page_t m) 1557{ 1558 struct pvo_entry *pvo; 1559 int count; 1560 1561 count = 0; 1562 if ((m->flags & PG_FICTITIOUS) != 0) 1563 return (count); 1564 vm_page_lock_queues(); 1565 LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) 1566 if ((pvo->pvo_vaddr & PVO_WIRED) != 0) 1567 count++; 1568 vm_page_unlock_queues(); 1569 return (count); 1570} 1571 1572static u_int moea_vsidcontext; 1573 1574void 1575moea_pinit(mmu_t mmu, pmap_t pmap) 1576{ 1577 int i, mask; 1578 u_int entropy; 1579 1580 KASSERT((int)pmap < VM_MIN_KERNEL_ADDRESS, ("moea_pinit: virt pmap")); 1581 PMAP_LOCK_INIT(pmap); 1582 1583 entropy = 0; 1584 __asm __volatile("mftb %0" : "=r"(entropy)); 1585 1586 if ((pmap->pmap_phys = (pmap_t)moea_kextract(mmu, (vm_offset_t)pmap)) 1587 == NULL) { 1588 pmap->pmap_phys = pmap; 1589 } 1590 1591 1592 /* 1593 * Allocate some segment registers for this pmap. 1594 */ 1595 for (i = 0; i < NPMAPS; i += VSID_NBPW) { 1596 u_int hash, n; 1597 1598 /* 1599 * Create a new value by mutiplying by a prime and adding in 1600 * entropy from the timebase register. This is to make the 1601 * VSID more random so that the PT hash function collides 1602 * less often. (Note that the prime casues gcc to do shifts 1603 * instead of a multiply.) 1604 */ 1605 moea_vsidcontext = (moea_vsidcontext * 0x1105) + entropy; 1606 hash = moea_vsidcontext & (NPMAPS - 1); 1607 if (hash == 0) /* 0 is special, avoid it */ 1608 continue; 1609 n = hash >> 5; 1610 mask = 1 << (hash & (VSID_NBPW - 1)); 1611 hash = (moea_vsidcontext & 0xfffff); 1612 if (moea_vsid_bitmap[n] & mask) { /* collision? */ 1613 /* anything free in this bucket? */ 1614 if (moea_vsid_bitmap[n] == 0xffffffff) { 1615 entropy = (moea_vsidcontext >> 20); 1616 continue; 1617 } 1618 i = ffs(~moea_vsid_bitmap[i]) - 1; 1619 mask = 1 << i; 1620 hash &= 0xfffff & ~(VSID_NBPW - 1); 1621 hash |= i; 1622 } 1623 moea_vsid_bitmap[n] |= mask; 1624 for (i = 0; i < 16; i++) 1625 pmap->pm_sr[i] = VSID_MAKE(i, hash); 1626 return; 1627 } 1628 1629 panic("moea_pinit: out of segments"); 1630} 1631 1632/* 1633 * Initialize the pmap associated with process 0. 1634 */ 1635void 1636moea_pinit0(mmu_t mmu, pmap_t pm) 1637{ 1638 1639 moea_pinit(mmu, pm); 1640 bzero(&pm->pm_stats, sizeof(pm->pm_stats)); 1641} 1642 1643/* 1644 * Set the physical protection on the specified range of this map as requested. 1645 */ 1646void 1647moea_protect(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva, 1648 vm_prot_t prot) 1649{ 1650 struct pvo_entry *pvo; 1651 struct pte *pt; 1652 int pteidx; 1653 1654 KASSERT(pm == &curproc->p_vmspace->vm_pmap || pm == kernel_pmap, 1655 ("moea_protect: non current pmap")); 1656 1657 if ((prot & VM_PROT_READ) == VM_PROT_NONE) { 1658 moea_remove(mmu, pm, sva, eva); 1659 return; 1660 } 1661 1662 vm_page_lock_queues(); 1663 PMAP_LOCK(pm); 1664 for (; sva < eva; sva += PAGE_SIZE) { 1665 pvo = moea_pvo_find_va(pm, sva, &pteidx); 1666 if (pvo == NULL) 1667 continue; 1668 1669 if ((prot & VM_PROT_EXECUTE) == 0) 1670 pvo->pvo_vaddr &= ~PVO_EXECUTABLE; 1671 1672 /* 1673 * Grab the PTE pointer before we diddle with the cached PTE 1674 * copy. 1675 */ 1676 pt = moea_pvo_to_pte(pvo, pteidx); 1677 /* 1678 * Change the protection of the page. 1679 */ 1680 pvo->pvo_pte.pte.pte_lo &= ~PTE_PP; 1681 pvo->pvo_pte.pte.pte_lo |= PTE_BR; 1682 1683 /* 1684 * If the PVO is in the page table, update that pte as well. 1685 */ 1686 if (pt != NULL) { 1687 moea_pte_change(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr); 1688 mtx_unlock(&moea_table_mutex); 1689 } 1690 } 1691 vm_page_unlock_queues(); 1692 PMAP_UNLOCK(pm); 1693} 1694 1695/* 1696 * Map a list of wired pages into kernel virtual address space. This is 1697 * intended for temporary mappings which do not need page modification or 1698 * references recorded. Existing mappings in the region are overwritten. 1699 */ 1700void 1701moea_qenter(mmu_t mmu, vm_offset_t sva, vm_page_t *m, int count) 1702{ 1703 vm_offset_t va; 1704 1705 va = sva; 1706 while (count-- > 0) { 1707 moea_kenter(mmu, va, VM_PAGE_TO_PHYS(*m)); 1708 va += PAGE_SIZE; 1709 m++; 1710 } 1711} 1712 1713/* 1714 * Remove page mappings from kernel virtual address space. Intended for 1715 * temporary mappings entered by moea_qenter. 1716 */ 1717void 1718moea_qremove(mmu_t mmu, vm_offset_t sva, int count) 1719{ 1720 vm_offset_t va; 1721 1722 va = sva; 1723 while (count-- > 0) { 1724 moea_kremove(mmu, va); 1725 va += PAGE_SIZE; 1726 } 1727} 1728 1729void 1730moea_release(mmu_t mmu, pmap_t pmap) 1731{ 1732 int idx, mask; 1733 1734 /* 1735 * Free segment register's VSID 1736 */ 1737 if (pmap->pm_sr[0] == 0) 1738 panic("moea_release"); 1739 1740 idx = VSID_TO_HASH(pmap->pm_sr[0]) & (NPMAPS-1); 1741 mask = 1 << (idx % VSID_NBPW); 1742 idx /= VSID_NBPW; 1743 moea_vsid_bitmap[idx] &= ~mask; 1744 PMAP_LOCK_DESTROY(pmap); 1745} 1746 1747/* 1748 * Remove the given range of addresses from the specified map. 1749 */ 1750void 1751moea_remove(mmu_t mmu, pmap_t pm, vm_offset_t sva, vm_offset_t eva) 1752{ 1753 struct pvo_entry *pvo; 1754 int pteidx; 1755 1756 vm_page_lock_queues(); 1757 PMAP_LOCK(pm); 1758 for (; sva < eva; sva += PAGE_SIZE) { 1759 pvo = moea_pvo_find_va(pm, sva, &pteidx); 1760 if (pvo != NULL) { 1761 moea_pvo_remove(pvo, pteidx); 1762 } 1763 } 1764 PMAP_UNLOCK(pm); 1765 vm_page_unlock_queues(); 1766} 1767 1768/* 1769 * Remove physical page from all pmaps in which it resides. moea_pvo_remove() 1770 * will reflect changes in pte's back to the vm_page. 1771 */ 1772void 1773moea_remove_all(mmu_t mmu, vm_page_t m) 1774{ 1775 struct pvo_head *pvo_head; 1776 struct pvo_entry *pvo, *next_pvo; 1777 pmap_t pmap; 1778 1779 vm_page_lock_queues(); 1780 pvo_head = vm_page_to_pvoh(m); 1781 for (pvo = LIST_FIRST(pvo_head); pvo != NULL; pvo = next_pvo) { 1782 next_pvo = LIST_NEXT(pvo, pvo_vlink); 1783 1784 MOEA_PVO_CHECK(pvo); /* sanity check */ 1785 pmap = pvo->pvo_pmap; 1786 PMAP_LOCK(pmap); 1787 moea_pvo_remove(pvo, -1); 1788 PMAP_UNLOCK(pmap); 1789 } 1790 if ((m->flags & PG_WRITEABLE) && moea_is_modified(mmu, m)) { 1791 moea_attr_clear(m, PTE_CHG); 1792 vm_page_dirty(m); 1793 } 1794 vm_page_flag_clear(m, PG_WRITEABLE); 1795 vm_page_unlock_queues(); 1796} 1797 1798/* 1799 * Allocate a physical page of memory directly from the phys_avail map. 1800 * Can only be called from moea_bootstrap before avail start and end are 1801 * calculated. 1802 */ 1803static vm_offset_t 1804moea_bootstrap_alloc(vm_size_t size, u_int align) 1805{ 1806 vm_offset_t s, e; 1807 int i, j; 1808 1809 size = round_page(size); 1810 for (i = 0; phys_avail[i + 1] != 0; i += 2) { 1811 if (align != 0) 1812 s = (phys_avail[i] + align - 1) & ~(align - 1); 1813 else 1814 s = phys_avail[i]; 1815 e = s + size; 1816 1817 if (s < phys_avail[i] || e > phys_avail[i + 1]) 1818 continue; 1819 1820 if (s == phys_avail[i]) { 1821 phys_avail[i] += size; 1822 } else if (e == phys_avail[i + 1]) { 1823 phys_avail[i + 1] -= size; 1824 } else { 1825 for (j = phys_avail_count * 2; j > i; j -= 2) { 1826 phys_avail[j] = phys_avail[j - 2]; 1827 phys_avail[j + 1] = phys_avail[j - 1]; 1828 } 1829 1830 phys_avail[i + 3] = phys_avail[i + 1]; 1831 phys_avail[i + 1] = s; 1832 phys_avail[i + 2] = e; 1833 phys_avail_count++; 1834 } 1835 1836 return (s); 1837 } 1838 panic("moea_bootstrap_alloc: could not allocate memory"); 1839} 1840 1841static void 1842moea_syncicache(vm_offset_t pa, vm_size_t len) 1843{ 1844 __syncicache((void *)pa, len); 1845} 1846 1847static int 1848moea_pvo_enter(pmap_t pm, uma_zone_t zone, struct pvo_head *pvo_head, 1849 vm_offset_t va, vm_offset_t pa, u_int pte_lo, int flags) 1850{ 1851 struct pvo_entry *pvo; 1852 u_int sr; 1853 int first; 1854 u_int ptegidx; 1855 int i; 1856 int bootstrap; 1857 1858 moea_pvo_enter_calls++; 1859 first = 0; 1860 bootstrap = 0; 1861 1862 /* 1863 * Compute the PTE Group index. 1864 */ 1865 va &= ~ADDR_POFF; 1866 sr = va_to_sr(pm->pm_sr, va); 1867 ptegidx = va_to_pteg(sr, va); 1868 1869 /* 1870 * Remove any existing mapping for this page. Reuse the pvo entry if 1871 * there is a mapping. 1872 */ 1873 mtx_lock(&moea_table_mutex); 1874 LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) { 1875 if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) { 1876 if ((pvo->pvo_pte.pte.pte_lo & PTE_RPGN) == pa && 1877 (pvo->pvo_pte.pte.pte_lo & PTE_PP) == 1878 (pte_lo & PTE_PP)) { 1879 mtx_unlock(&moea_table_mutex); 1880 return (0); 1881 } 1882 moea_pvo_remove(pvo, -1); 1883 break; 1884 } 1885 } 1886 1887 /* 1888 * If we aren't overwriting a mapping, try to allocate. 1889 */ 1890 if (moea_initialized) { 1891 pvo = uma_zalloc(zone, M_NOWAIT); 1892 } else { 1893 if (moea_bpvo_pool_index >= BPVO_POOL_SIZE) { 1894 panic("moea_enter: bpvo pool exhausted, %d, %d, %d", 1895 moea_bpvo_pool_index, BPVO_POOL_SIZE, 1896 BPVO_POOL_SIZE * sizeof(struct pvo_entry)); 1897 } 1898 pvo = &moea_bpvo_pool[moea_bpvo_pool_index]; 1899 moea_bpvo_pool_index++; 1900 bootstrap = 1; 1901 } 1902 1903 if (pvo == NULL) { 1904 mtx_unlock(&moea_table_mutex); 1905 return (ENOMEM); 1906 } 1907 1908 moea_pvo_entries++; 1909 pvo->pvo_vaddr = va; 1910 pvo->pvo_pmap = pm; 1911 LIST_INSERT_HEAD(&moea_pvo_table[ptegidx], pvo, pvo_olink); 1912 pvo->pvo_vaddr &= ~ADDR_POFF; 1913 if (flags & VM_PROT_EXECUTE) 1914 pvo->pvo_vaddr |= PVO_EXECUTABLE; 1915 if (flags & PVO_WIRED) 1916 pvo->pvo_vaddr |= PVO_WIRED; 1917 if (pvo_head != &moea_pvo_kunmanaged) 1918 pvo->pvo_vaddr |= PVO_MANAGED; 1919 if (bootstrap) 1920 pvo->pvo_vaddr |= PVO_BOOTSTRAP; 1921 if (flags & PVO_FAKE) 1922 pvo->pvo_vaddr |= PVO_FAKE; 1923 1924 moea_pte_create(&pvo->pvo_pte.pte, sr, va, pa | pte_lo); 1925 1926 /* 1927 * Remember if the list was empty and therefore will be the first 1928 * item. 1929 */ 1930 if (LIST_FIRST(pvo_head) == NULL) 1931 first = 1; 1932 LIST_INSERT_HEAD(pvo_head, pvo, pvo_vlink); 1933 1934 if (pvo->pvo_pte.pte.pte_lo & PVO_WIRED) 1935 pm->pm_stats.wired_count++; 1936 pm->pm_stats.resident_count++; 1937 1938 /* 1939 * We hope this succeeds but it isn't required. 1940 */ 1941 i = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte); 1942 if (i >= 0) { 1943 PVO_PTEGIDX_SET(pvo, i); 1944 } else { 1945 panic("moea_pvo_enter: overflow"); 1946 moea_pte_overflow++; 1947 } 1948 mtx_unlock(&moea_table_mutex); 1949 1950 return (first ? ENOENT : 0); 1951} 1952 1953static void 1954moea_pvo_remove(struct pvo_entry *pvo, int pteidx) 1955{ 1956 struct pte *pt; 1957 1958 /* 1959 * If there is an active pte entry, we need to deactivate it (and 1960 * save the ref & cfg bits). 1961 */ 1962 pt = moea_pvo_to_pte(pvo, pteidx); 1963 if (pt != NULL) { 1964 moea_pte_unset(pt, &pvo->pvo_pte.pte, pvo->pvo_vaddr); 1965 mtx_unlock(&moea_table_mutex); 1966 PVO_PTEGIDX_CLR(pvo); 1967 } else { 1968 moea_pte_overflow--; 1969 } 1970 1971 /* 1972 * Update our statistics. 1973 */ 1974 pvo->pvo_pmap->pm_stats.resident_count--; 1975 if (pvo->pvo_pte.pte.pte_lo & PVO_WIRED) 1976 pvo->pvo_pmap->pm_stats.wired_count--; 1977 1978 /* 1979 * Save the REF/CHG bits into their cache if the page is managed. 1980 */ 1981 if ((pvo->pvo_vaddr & (PVO_MANAGED|PVO_FAKE)) == PVO_MANAGED) { 1982 struct vm_page *pg; 1983 1984 pg = PHYS_TO_VM_PAGE(pvo->pvo_pte.pte.pte_lo & PTE_RPGN); 1985 if (pg != NULL) { 1986 moea_attr_save(pg, pvo->pvo_pte.pte.pte_lo & 1987 (PTE_REF | PTE_CHG)); 1988 } 1989 } 1990 1991 /* 1992 * Remove this PVO from the PV list. 1993 */ 1994 LIST_REMOVE(pvo, pvo_vlink); 1995 1996 /* 1997 * Remove this from the overflow list and return it to the pool 1998 * if we aren't going to reuse it. 1999 */ 2000 LIST_REMOVE(pvo, pvo_olink); 2001 if (!(pvo->pvo_vaddr & PVO_BOOTSTRAP)) 2002 uma_zfree(pvo->pvo_vaddr & PVO_MANAGED ? moea_mpvo_zone : 2003 moea_upvo_zone, pvo); 2004 moea_pvo_entries--; 2005 moea_pvo_remove_calls++; 2006} 2007 2008static __inline int 2009moea_pvo_pte_index(const struct pvo_entry *pvo, int ptegidx) 2010{ 2011 int pteidx; 2012 2013 /* 2014 * We can find the actual pte entry without searching by grabbing 2015 * the PTEG index from 3 unused bits in pte_lo[11:9] and by 2016 * noticing the HID bit. 2017 */ 2018 pteidx = ptegidx * 8 + PVO_PTEGIDX_GET(pvo); 2019 if (pvo->pvo_pte.pte.pte_hi & PTE_HID) 2020 pteidx ^= moea_pteg_mask * 8; 2021 2022 return (pteidx); 2023} 2024 2025static struct pvo_entry * 2026moea_pvo_find_va(pmap_t pm, vm_offset_t va, int *pteidx_p) 2027{ 2028 struct pvo_entry *pvo; 2029 int ptegidx; 2030 u_int sr; 2031 2032 va &= ~ADDR_POFF; 2033 sr = va_to_sr(pm->pm_sr, va); 2034 ptegidx = va_to_pteg(sr, va); 2035 2036 mtx_lock(&moea_table_mutex); 2037 LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) { 2038 if (pvo->pvo_pmap == pm && PVO_VADDR(pvo) == va) { 2039 if (pteidx_p) 2040 *pteidx_p = moea_pvo_pte_index(pvo, ptegidx); 2041 break; 2042 } 2043 } 2044 mtx_unlock(&moea_table_mutex); 2045 2046 return (pvo); 2047} 2048 2049static struct pte * 2050moea_pvo_to_pte(const struct pvo_entry *pvo, int pteidx) 2051{ 2052 struct pte *pt; 2053 2054 /* 2055 * If we haven't been supplied the ptegidx, calculate it. 2056 */ 2057 if (pteidx == -1) { 2058 int ptegidx; 2059 u_int sr; 2060 2061 sr = va_to_sr(pvo->pvo_pmap->pm_sr, pvo->pvo_vaddr); 2062 ptegidx = va_to_pteg(sr, pvo->pvo_vaddr); 2063 pteidx = moea_pvo_pte_index(pvo, ptegidx); 2064 } 2065 2066 pt = &moea_pteg_table[pteidx >> 3].pt[pteidx & 7]; 2067 mtx_lock(&moea_table_mutex); 2068 2069 if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) && !PVO_PTEGIDX_ISSET(pvo)) { 2070 panic("moea_pvo_to_pte: pvo %p has valid pte in pvo but no " 2071 "valid pte index", pvo); 2072 } 2073 2074 if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0 && PVO_PTEGIDX_ISSET(pvo)) { 2075 panic("moea_pvo_to_pte: pvo %p has valid pte index in pvo " 2076 "pvo but no valid pte", pvo); 2077 } 2078 2079 if ((pt->pte_hi ^ (pvo->pvo_pte.pte.pte_hi & ~PTE_VALID)) == PTE_VALID) { 2080 if ((pvo->pvo_pte.pte.pte_hi & PTE_VALID) == 0) { 2081 panic("moea_pvo_to_pte: pvo %p has valid pte in " 2082 "moea_pteg_table %p but invalid in pvo", pvo, pt); 2083 } 2084 2085 if (((pt->pte_lo ^ pvo->pvo_pte.pte.pte_lo) & ~(PTE_CHG|PTE_REF)) 2086 != 0) { 2087 panic("moea_pvo_to_pte: pvo %p pte does not match " 2088 "pte %p in moea_pteg_table", pvo, pt); 2089 } 2090 2091 mtx_assert(&moea_table_mutex, MA_OWNED); 2092 return (pt); 2093 } 2094 2095 if (pvo->pvo_pte.pte.pte_hi & PTE_VALID) { 2096 panic("moea_pvo_to_pte: pvo %p has invalid pte %p in " 2097 "moea_pteg_table but valid in pvo", pvo, pt); 2098 } 2099 2100 mtx_unlock(&moea_table_mutex); 2101 return (NULL); 2102} 2103 2104/* 2105 * XXX: THIS STUFF SHOULD BE IN pte.c? 2106 */ 2107int 2108moea_pte_spill(vm_offset_t addr) 2109{ 2110 struct pvo_entry *source_pvo, *victim_pvo; 2111 struct pvo_entry *pvo; 2112 int ptegidx, i, j; 2113 u_int sr; 2114 struct pteg *pteg; 2115 struct pte *pt; 2116 2117 moea_pte_spills++; 2118 2119 sr = mfsrin(addr); 2120 ptegidx = va_to_pteg(sr, addr); 2121 2122 /* 2123 * Have to substitute some entry. Use the primary hash for this. 2124 * Use low bits of timebase as random generator. 2125 */ 2126 pteg = &moea_pteg_table[ptegidx]; 2127 mtx_lock(&moea_table_mutex); 2128 __asm __volatile("mftb %0" : "=r"(i)); 2129 i &= 7; 2130 pt = &pteg->pt[i]; 2131 2132 source_pvo = NULL; 2133 victim_pvo = NULL; 2134 LIST_FOREACH(pvo, &moea_pvo_table[ptegidx], pvo_olink) { 2135 /* 2136 * We need to find a pvo entry for this address. 2137 */ 2138 MOEA_PVO_CHECK(pvo); 2139 if (source_pvo == NULL && 2140 moea_pte_match(&pvo->pvo_pte.pte, sr, addr, 2141 pvo->pvo_pte.pte.pte_hi & PTE_HID)) { 2142 /* 2143 * Now found an entry to be spilled into the pteg. 2144 * The PTE is now valid, so we know it's active. 2145 */ 2146 j = moea_pte_insert(ptegidx, &pvo->pvo_pte.pte); 2147 2148 if (j >= 0) { 2149 PVO_PTEGIDX_SET(pvo, j); 2150 moea_pte_overflow--; 2151 MOEA_PVO_CHECK(pvo); 2152 mtx_unlock(&moea_table_mutex); 2153 return (1); 2154 } 2155 2156 source_pvo = pvo; 2157 2158 if (victim_pvo != NULL) 2159 break; 2160 } 2161 2162 /* 2163 * We also need the pvo entry of the victim we are replacing 2164 * so save the R & C bits of the PTE. 2165 */ 2166 if ((pt->pte_hi & PTE_HID) == 0 && victim_pvo == NULL && 2167 moea_pte_compare(pt, &pvo->pvo_pte.pte)) { 2168 victim_pvo = pvo; 2169 if (source_pvo != NULL) 2170 break; 2171 } 2172 } 2173 2174 if (source_pvo == NULL) { 2175 mtx_unlock(&moea_table_mutex); 2176 return (0); 2177 } 2178 2179 if (victim_pvo == NULL) { 2180 if ((pt->pte_hi & PTE_HID) == 0) 2181 panic("moea_pte_spill: victim p-pte (%p) has no pvo" 2182 "entry", pt); 2183 2184 /* 2185 * If this is a secondary PTE, we need to search it's primary 2186 * pvo bucket for the matching PVO. 2187 */ 2188 LIST_FOREACH(pvo, &moea_pvo_table[ptegidx ^ moea_pteg_mask], 2189 pvo_olink) { 2190 MOEA_PVO_CHECK(pvo); 2191 /* 2192 * We also need the pvo entry of the victim we are 2193 * replacing so save the R & C bits of the PTE. 2194 */ 2195 if (moea_pte_compare(pt, &pvo->pvo_pte.pte)) { 2196 victim_pvo = pvo; 2197 break; 2198 } 2199 } 2200 2201 if (victim_pvo == NULL) 2202 panic("moea_pte_spill: victim s-pte (%p) has no pvo" 2203 "entry", pt); 2204 } 2205 2206 /* 2207 * We are invalidating the TLB entry for the EA we are replacing even 2208 * though it's valid. If we don't, we lose any ref/chg bit changes 2209 * contained in the TLB entry. 2210 */ 2211 source_pvo->pvo_pte.pte.pte_hi &= ~PTE_HID; 2212 2213 moea_pte_unset(pt, &victim_pvo->pvo_pte.pte, victim_pvo->pvo_vaddr); 2214 moea_pte_set(pt, &source_pvo->pvo_pte.pte); 2215 2216 PVO_PTEGIDX_CLR(victim_pvo); 2217 PVO_PTEGIDX_SET(source_pvo, i); 2218 moea_pte_replacements++; 2219 2220 MOEA_PVO_CHECK(victim_pvo); 2221 MOEA_PVO_CHECK(source_pvo); 2222 2223 mtx_unlock(&moea_table_mutex); 2224 return (1); 2225} 2226 2227static int 2228moea_pte_insert(u_int ptegidx, struct pte *pvo_pt) 2229{ 2230 struct pte *pt; 2231 int i; 2232 2233 mtx_assert(&moea_table_mutex, MA_OWNED); 2234 2235 /* 2236 * First try primary hash. 2237 */ 2238 for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) { 2239 if ((pt->pte_hi & PTE_VALID) == 0) { 2240 pvo_pt->pte_hi &= ~PTE_HID; 2241 moea_pte_set(pt, pvo_pt); 2242 return (i); 2243 } 2244 } 2245 2246 /* 2247 * Now try secondary hash. 2248 */ 2249 ptegidx ^= moea_pteg_mask; 2250 2251 for (pt = moea_pteg_table[ptegidx].pt, i = 0; i < 8; i++, pt++) { 2252 if ((pt->pte_hi & PTE_VALID) == 0) { 2253 pvo_pt->pte_hi |= PTE_HID; 2254 moea_pte_set(pt, pvo_pt); 2255 return (i); 2256 } 2257 } 2258 2259 panic("moea_pte_insert: overflow"); 2260 return (-1); 2261} 2262 2263static boolean_t 2264moea_query_bit(vm_page_t m, int ptebit) 2265{ 2266 struct pvo_entry *pvo; 2267 struct pte *pt; 2268 2269 if (moea_attr_fetch(m) & ptebit) 2270 return (TRUE); 2271 2272 vm_page_lock_queues(); 2273 LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) { 2274 MOEA_PVO_CHECK(pvo); /* sanity check */ 2275 2276 /* 2277 * See if we saved the bit off. If so, cache it and return 2278 * success. 2279 */ 2280 if (pvo->pvo_pte.pte.pte_lo & ptebit) { 2281 moea_attr_save(m, ptebit); 2282 MOEA_PVO_CHECK(pvo); /* sanity check */ 2283 vm_page_unlock_queues(); 2284 return (TRUE); 2285 } 2286 } 2287 2288 /* 2289 * No luck, now go through the hard part of looking at the PTEs 2290 * themselves. Sync so that any pending REF/CHG bits are flushed to 2291 * the PTEs. 2292 */ 2293 powerpc_sync(); 2294 LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) { 2295 MOEA_PVO_CHECK(pvo); /* sanity check */ 2296 2297 /* 2298 * See if this pvo has a valid PTE. if so, fetch the 2299 * REF/CHG bits from the valid PTE. If the appropriate 2300 * ptebit is set, cache it and return success. 2301 */ 2302 pt = moea_pvo_to_pte(pvo, -1); 2303 if (pt != NULL) { 2304 moea_pte_synch(pt, &pvo->pvo_pte.pte); 2305 mtx_unlock(&moea_table_mutex); 2306 if (pvo->pvo_pte.pte.pte_lo & ptebit) { 2307 moea_attr_save(m, ptebit); 2308 MOEA_PVO_CHECK(pvo); /* sanity check */ 2309 vm_page_unlock_queues(); 2310 return (TRUE); 2311 } 2312 } 2313 } 2314 2315 vm_page_unlock_queues(); 2316 return (FALSE); 2317} 2318 2319static u_int 2320moea_clear_bit(vm_page_t m, int ptebit) 2321{ 2322 u_int count; 2323 struct pvo_entry *pvo; 2324 struct pte *pt; 2325 2326 vm_page_lock_queues(); 2327 2328 /* 2329 * Clear the cached value. 2330 */ 2331 moea_attr_clear(m, ptebit); 2332 2333 /* 2334 * Sync so that any pending REF/CHG bits are flushed to the PTEs (so 2335 * we can reset the right ones). note that since the pvo entries and 2336 * list heads are accessed via BAT0 and are never placed in the page 2337 * table, we don't have to worry about further accesses setting the 2338 * REF/CHG bits. 2339 */ 2340 powerpc_sync(); 2341 2342 /* 2343 * For each pvo entry, clear the pvo's ptebit. If this pvo has a 2344 * valid pte clear the ptebit from the valid pte. 2345 */ 2346 count = 0; 2347 LIST_FOREACH(pvo, vm_page_to_pvoh(m), pvo_vlink) { 2348 MOEA_PVO_CHECK(pvo); /* sanity check */ 2349 pt = moea_pvo_to_pte(pvo, -1); 2350 if (pt != NULL) { 2351 moea_pte_synch(pt, &pvo->pvo_pte.pte); 2352 if (pvo->pvo_pte.pte.pte_lo & ptebit) { 2353 count++; 2354 moea_pte_clear(pt, PVO_VADDR(pvo), ptebit); 2355 } 2356 mtx_unlock(&moea_table_mutex); 2357 } 2358 pvo->pvo_pte.pte.pte_lo &= ~ptebit; 2359 MOEA_PVO_CHECK(pvo); /* sanity check */ 2360 } 2361 2362 vm_page_unlock_queues(); 2363 return (count); 2364} 2365 2366/* 2367 * Return true if the physical range is encompassed by the battable[idx] 2368 */ 2369static int 2370moea_bat_mapped(int idx, vm_offset_t pa, vm_size_t size) 2371{ 2372 u_int prot; 2373 u_int32_t start; 2374 u_int32_t end; 2375 u_int32_t bat_ble; 2376 2377 /* 2378 * Return immediately if not a valid mapping 2379 */ 2380 if (!battable[idx].batu & BAT_Vs) 2381 return (EINVAL); 2382 2383 /* 2384 * The BAT entry must be cache-inhibited, guarded, and r/w 2385 * so it can function as an i/o page 2386 */ 2387 prot = battable[idx].batl & (BAT_I|BAT_G|BAT_PP_RW); 2388 if (prot != (BAT_I|BAT_G|BAT_PP_RW)) 2389 return (EPERM); 2390 2391 /* 2392 * The address should be within the BAT range. Assume that the 2393 * start address in the BAT has the correct alignment (thus 2394 * not requiring masking) 2395 */ 2396 start = battable[idx].batl & BAT_PBS; 2397 bat_ble = (battable[idx].batu & ~(BAT_EBS)) | 0x03; 2398 end = start | (bat_ble << 15) | 0x7fff; 2399 2400 if ((pa < start) || ((pa + size) > end)) 2401 return (ERANGE); 2402 2403 return (0); 2404} 2405 2406boolean_t 2407moea_dev_direct_mapped(mmu_t mmu, vm_offset_t pa, vm_size_t size) 2408{ 2409 int i; 2410 2411 /* 2412 * This currently does not work for entries that 2413 * overlap 256M BAT segments. 2414 */ 2415 2416 for(i = 0; i < 16; i++) 2417 if (moea_bat_mapped(i, pa, size) == 0) 2418 return (0); 2419 2420 return (EFAULT); 2421} 2422 2423/* 2424 * Map a set of physical memory pages into the kernel virtual 2425 * address space. Return a pointer to where it is mapped. This 2426 * routine is intended to be used for mapping device memory, 2427 * NOT real memory. 2428 */ 2429void * 2430moea_mapdev(mmu_t mmu, vm_offset_t pa, vm_size_t size) 2431{ 2432 vm_offset_t va, tmpva, ppa, offset; 2433 int i; 2434 2435 ppa = trunc_page(pa); 2436 offset = pa & PAGE_MASK; 2437 size = roundup(offset + size, PAGE_SIZE); 2438 2439 GIANT_REQUIRED; 2440 2441 /* 2442 * If the physical address lies within a valid BAT table entry, 2443 * return the 1:1 mapping. This currently doesn't work 2444 * for regions that overlap 256M BAT segments. 2445 */ 2446 for (i = 0; i < 16; i++) { 2447 if (moea_bat_mapped(i, pa, size) == 0) 2448 return ((void *) pa); 2449 } 2450 2451 va = kmem_alloc_nofault(kernel_map, size); 2452 if (!va) 2453 panic("moea_mapdev: Couldn't alloc kernel virtual memory"); 2454 2455 for (tmpva = va; size > 0;) { 2456 moea_kenter(mmu, tmpva, ppa); 2457 tlbie(tmpva); 2458 size -= PAGE_SIZE; 2459 tmpva += PAGE_SIZE; 2460 ppa += PAGE_SIZE; 2461 } 2462 2463 return ((void *)(va + offset)); 2464} 2465 2466void 2467moea_unmapdev(mmu_t mmu, vm_offset_t va, vm_size_t size) 2468{ 2469 vm_offset_t base, offset; 2470 2471 /* 2472 * If this is outside kernel virtual space, then it's a 2473 * battable entry and doesn't require unmapping 2474 */ 2475 if ((va >= VM_MIN_KERNEL_ADDRESS) && (va <= virtual_end)) { 2476 base = trunc_page(va); 2477 offset = va & PAGE_MASK; 2478 size = roundup(offset + size, PAGE_SIZE); 2479 kmem_free(kernel_map, base, size); 2480 } 2481} 2482 2483static void 2484moea_sync_icache(mmu_t mmu, pmap_t pm, vm_offset_t va, vm_size_t sz) 2485{ 2486 struct pvo_entry *pvo; 2487 vm_offset_t lim; 2488 vm_paddr_t pa; 2489 vm_size_t len; 2490 2491 PMAP_LOCK(pm); 2492 while (sz > 0) { 2493 lim = round_page(va); 2494 len = MIN(lim - va, sz); 2495 pvo = moea_pvo_find_va(pm, va & ~ADDR_POFF, NULL); 2496 if (pvo != NULL) { 2497 pa = (pvo->pvo_pte.pte.pte_lo & PTE_RPGN) | 2498 (va & ADDR_POFF); 2499 moea_syncicache(pa, len); 2500 } 2501 va += len; 2502 sz -= len; 2503 } 2504 PMAP_UNLOCK(pm); 2505} 2506