1 /* $NetBSD: pmap.c,v 1.156 2005/12/10 21:19:57 scw Exp $ */ 2 3 /* 4 * Copyright 2003 Wasabi Systems, Inc. 5 * All rights reserved. 6 * 7 * Written by Steve C. Woodford for Wasabi Systems, Inc. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 3. All advertising materials mentioning features or use of this software 18 * must display the following acknowledgement: 19 * This product includes software developed for the NetBSD Project by 20 * Wasabi Systems, Inc. 21 * 4. The name of Wasabi Systems, Inc. may not be used to endorse 22 * or promote products derived from this software without specific prior 23 * written permission. 24 * 25 * THIS SOFTWARE IS PROVIDED BY WASABI SYSTEMS, INC. ``AS IS'' AND 26 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 27 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 28 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL WASABI SYSTEMS, INC 29 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 30 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 31 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 32 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 33 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 34 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 35 * POSSIBILITY OF SUCH DAMAGE. 36 */ 37 38 /* 39 * Copyright (c) 2002-2003 Wasabi Systems, Inc. 40 * Copyright (c) 2001 Richard Earnshaw 41 * Copyright (c) 2001-2002 Christopher Gilbert 42 * All rights reserved. 43 * 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. The name of the company nor the name of the author may be used to 50 * endorse or promote products derived from this software without specific 51 * prior written permission. 52 * 53 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED 54 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF 55 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 56 * IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, 57 * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 58 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 59 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 60 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 61 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 62 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 63 * SUCH DAMAGE. 64 */ 65 66 /*- 67 * Copyright (c) 1999 The NetBSD Foundation, Inc. 68 * All rights reserved. 69 * 70 * This code is derived from software contributed to The NetBSD Foundation 71 * by Charles M. Hannum. 72 * 73 * Redistribution and use in source and binary forms, with or without 74 * modification, are permitted provided that the following conditions 75 * are met: 76 * 1. Redistributions of source code must retain the above copyright 77 * notice, this list of conditions and the following disclaimer. 78 * 2. Redistributions in binary form must reproduce the above copyright 79 * notice, this list of conditions and the following disclaimer in the 80 * documentation and/or other materials provided with the distribution. 81 * 3. All advertising materials mentioning features or use of this software 82 * must display the following acknowledgement: 83 * This product includes software developed by the NetBSD 84 * Foundation, Inc. and its contributors. 85 * 4. Neither the name of The NetBSD Foundation nor the names of its 86 * contributors may be used to endorse or promote products derived 87 * from this software without specific prior written permission. 88 * 89 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 90 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 91 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 92 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 93 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 94 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 95 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 96 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 97 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 98 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 99 * POSSIBILITY OF SUCH DAMAGE. 100 */ 101 102 /* 103 * Copyright (c) 1994-1998 Mark Brinicombe. 104 * Copyright (c) 1994 Brini. 105 * All rights reserved. 106 * 107 * This code is derived from software written for Brini by Mark Brinicombe 108 * 109 * Redistribution and use in source and binary forms, with or without 110 * modification, are permitted provided that the following conditions 111 * are met: 112 * 1. Redistributions of source code must retain the above copyright 113 * notice, this list of conditions and the following disclaimer. 114 * 2. Redistributions in binary form must reproduce the above copyright 115 * notice, this list of conditions and the following disclaimer in the 116 * documentation and/or other materials provided with the distribution. 117 * 3. All advertising materials mentioning features or use of this software 118 * must display the following acknowledgement: 119 * This product includes software developed by Mark Brinicombe. 120 * 4. The name of the author may not be used to endorse or promote products 121 * derived from this software without specific prior written permission. 122 * 123 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 124 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 125 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 126 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 127 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 128 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 129 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 130 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 131 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 132 * 133 * RiscBSD kernel project 134 * 135 * pmap.c 136 * 137 * Machine dependant vm stuff 138 * 139 * Created : 20/09/94 140 */ 141 142 /* 143 * Performance improvements, UVM changes, overhauls and part-rewrites 144 * were contributed by Neil A. Carson <neil@causality.com>. 145 */ 146 147 /* 148 * Overhauled again to speedup the pmap, use MMU Domains so that L1 tables 149 * can be shared, and re-work the KVM layout, by Steve Woodford of Wasabi 150 * Systems, Inc. 151 * 152 * There are still a few things outstanding at this time: 153 * 154 * - There are some unresolved issues for MP systems: 155 * 156 * o The L1 metadata needs a lock, or more specifically, some places 157 * need to acquire an exclusive lock when modifying L1 translation 158 * table entries. 159 * 160 * o When one cpu modifies an L1 entry, and that L1 table is also 161 * being used by another cpu, then the latter will need to be told 162 * that a tlb invalidation may be necessary. (But only if the old 163 * domain number in the L1 entry being over-written is currently 164 * the active domain on that cpu). I guess there are lots more tlb 165 * shootdown issues too... 166 * 167 * o If the vector_page is at 0x00000000 instead of 0xffff0000, then 168 * MP systems will lose big-time because of the MMU domain hack. 169 * The only way this can be solved (apart from moving the vector 170 * page to 0xffff0000) is to reserve the first 1MB of user address 171 * space for kernel use only. This would require re-linking all 172 * applications so that the text section starts above this 1MB 173 * boundary. 174 * 175 * o Tracking which VM space is resident in the cache/tlb has not yet 176 * been implemented for MP systems. 177 * 178 * o Finally, there is a pathological condition where two cpus running 179 * two separate processes (not lwps) which happen to share an L1 180 * can get into a fight over one or more L1 entries. This will result 181 * in a significant slow-down if both processes are in tight loops. 182 */ 183 184 /* 185 * Special compilation symbols 186 * PMAP_DEBUG - Build in pmap_debug_level code 187 */ 188 189 /* Include header files */ 190 191 #include "opt_cpuoptions.h" 192 #include "opt_pmap_debug.h" 193 #include "opt_ddb.h" 194 #include "opt_lockdebug.h" 195 #include "opt_multiprocessor.h" 196 197 #include <sys/types.h> 198 #include <sys/param.h> 199 #include <sys/kernel.h> 200 #include <sys/systm.h> 201 #include <sys/proc.h> 202 #include <sys/malloc.h> 203 #include <sys/user.h> 204 #include <sys/pool.h> 205 #include <sys/cdefs.h> 206 207 #include <uvm/uvm.h> 208 209 #include <machine/bus.h> 210 #include <machine/pmap.h> 211 #include <machine/pcb.h> 212 #include <machine/param.h> 213 #include <arm/arm32/katelib.h> 214 215 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.156 2005/12/10 21:19:57 scw Exp $"); 216 217 #ifdef PMAP_DEBUG 218 219 /* XXX need to get rid of all refs to this */ 220 int pmap_debug_level = 0; 221 222 /* 223 * for switching to potentially finer grained debugging 224 */ 225 #define PDB_FOLLOW 0x0001 226 #define PDB_INIT 0x0002 227 #define PDB_ENTER 0x0004 228 #define PDB_REMOVE 0x0008 229 #define PDB_CREATE 0x0010 230 #define PDB_PTPAGE 0x0020 231 #define PDB_GROWKERN 0x0040 232 #define PDB_BITS 0x0080 233 #define PDB_COLLECT 0x0100 234 #define PDB_PROTECT 0x0200 235 #define PDB_MAP_L1 0x0400 236 #define PDB_BOOTSTRAP 0x1000 237 #define PDB_PARANOIA 0x2000 238 #define PDB_WIRING 0x4000 239 #define PDB_PVDUMP 0x8000 240 #define PDB_VAC 0x10000 241 #define PDB_KENTER 0x20000 242 #define PDB_KREMOVE 0x40000 243 244 int debugmap = 1; 245 int pmapdebug = 0; 246 #define NPDEBUG(_lev_,_stat_) \ 247 if (pmapdebug & (_lev_)) \ 248 ((_stat_)) 249 250 #else /* PMAP_DEBUG */ 251 #define NPDEBUG(_lev_,_stat_) /* Nothing */ 252 #endif /* PMAP_DEBUG */ 253 254 /* 255 * pmap_kernel() points here 256 */ 257 struct pmap kernel_pmap_store; 258 259 /* 260 * Which pmap is currently 'live' in the cache 261 * 262 * XXXSCW: Fix for SMP ... 263 */ 264 union pmap_cache_state *pmap_cache_state; 265 266 /* 267 * Pool and cache that pmap structures are allocated from. 268 * We use a cache to avoid clearing the pm_l2[] array (1KB) 269 * in pmap_create(). 270 */ 271 static struct pool pmap_pmap_pool; 272 static struct pool_cache pmap_pmap_cache; 273 static LIST_HEAD(, pmap) pmap_pmaps; 274 275 /* 276 * Pool of PV structures 277 */ 278 static struct pool pmap_pv_pool; 279 static void *pmap_bootstrap_pv_page_alloc(struct pool *, int); 280 static void pmap_bootstrap_pv_page_free(struct pool *, void *); 281 static struct pool_allocator pmap_bootstrap_pv_allocator = { 282 pmap_bootstrap_pv_page_alloc, pmap_bootstrap_pv_page_free 283 }; 284 285 /* 286 * Pool and cache of l2_dtable structures. 287 * We use a cache to avoid clearing the structures when they're 288 * allocated. (196 bytes) 289 */ 290 static struct pool pmap_l2dtable_pool; 291 static struct pool_cache pmap_l2dtable_cache; 292 static vaddr_t pmap_kernel_l2dtable_kva; 293 294 /* 295 * Pool and cache of L2 page descriptors. 296 * We use a cache to avoid clearing the descriptor table 297 * when they're allocated. (1KB) 298 */ 299 static struct pool pmap_l2ptp_pool; 300 static struct pool_cache pmap_l2ptp_cache; 301 static vaddr_t pmap_kernel_l2ptp_kva; 302 static paddr_t pmap_kernel_l2ptp_phys; 303 304 /* 305 * pmap copy/zero page, and mem(5) hook point 306 */ 307 static pt_entry_t *csrc_pte, *cdst_pte; 308 static vaddr_t csrcp, cdstp; 309 char *memhook; 310 extern caddr_t msgbufaddr; 311 312 /* 313 * Flag to indicate if pmap_init() has done its thing 314 */ 315 boolean_t pmap_initialized; 316 317 /* 318 * Misc. locking data structures 319 */ 320 321 #if defined(MULTIPROCESSOR) || defined(LOCKDEBUG) 322 static struct lock pmap_main_lock; 323 324 #define PMAP_MAP_TO_HEAD_LOCK() \ 325 (void) spinlockmgr(&pmap_main_lock, LK_SHARED, NULL) 326 #define PMAP_MAP_TO_HEAD_UNLOCK() \ 327 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL) 328 #define PMAP_HEAD_TO_MAP_LOCK() \ 329 (void) spinlockmgr(&pmap_main_lock, LK_EXCLUSIVE, NULL) 330 #define PMAP_HEAD_TO_MAP_UNLOCK() \ 331 spinlockmgr(&pmap_main_lock, LK_RELEASE, (void *) 0) 332 #else 333 #define PMAP_MAP_TO_HEAD_LOCK() /* null */ 334 #define PMAP_MAP_TO_HEAD_UNLOCK() /* null */ 335 #define PMAP_HEAD_TO_MAP_LOCK() /* null */ 336 #define PMAP_HEAD_TO_MAP_UNLOCK() /* null */ 337 #endif 338 339 #define pmap_acquire_pmap_lock(pm) \ 340 do { \ 341 if ((pm) != pmap_kernel()) \ 342 simple_lock(&(pm)->pm_lock); \ 343 } while (/*CONSTCOND*/0) 344 345 #define pmap_release_pmap_lock(pm) \ 346 do { \ 347 if ((pm) != pmap_kernel()) \ 348 simple_unlock(&(pm)->pm_lock); \ 349 } while (/*CONSTCOND*/0) 350 351 352 /* 353 * Metadata for L1 translation tables. 354 */ 355 struct l1_ttable { 356 /* Entry on the L1 Table list */ 357 SLIST_ENTRY(l1_ttable) l1_link; 358 359 /* Entry on the L1 Least Recently Used list */ 360 TAILQ_ENTRY(l1_ttable) l1_lru; 361 362 /* Track how many domains are allocated from this L1 */ 363 volatile u_int l1_domain_use_count; 364 365 /* 366 * A free-list of domain numbers for this L1. 367 * We avoid using ffs() and a bitmap to track domains since ffs() 368 * is slow on ARM. 369 */ 370 u_int8_t l1_domain_first; 371 u_int8_t l1_domain_free[PMAP_DOMAINS]; 372 373 /* Physical address of this L1 page table */ 374 paddr_t l1_physaddr; 375 376 /* KVA of this L1 page table */ 377 pd_entry_t *l1_kva; 378 }; 379 380 /* 381 * Convert a virtual address into its L1 table index. That is, the 382 * index used to locate the L2 descriptor table pointer in an L1 table. 383 * This is basically used to index l1->l1_kva[]. 384 * 385 * Each L2 descriptor table represents 1MB of VA space. 386 */ 387 #define L1_IDX(va) (((vaddr_t)(va)) >> L1_S_SHIFT) 388 389 /* 390 * L1 Page Tables are tracked using a Least Recently Used list. 391 * - New L1s are allocated from the HEAD. 392 * - Freed L1s are added to the TAIl. 393 * - Recently accessed L1s (where an 'access' is some change to one of 394 * the userland pmaps which owns this L1) are moved to the TAIL. 395 */ 396 static TAILQ_HEAD(, l1_ttable) l1_lru_list; 397 static struct simplelock l1_lru_lock; 398 399 /* 400 * A list of all L1 tables 401 */ 402 static SLIST_HEAD(, l1_ttable) l1_list; 403 404 /* 405 * The l2_dtable tracks L2_BUCKET_SIZE worth of L1 slots. 406 * 407 * This is normally 16MB worth L2 page descriptors for any given pmap. 408 * Reference counts are maintained for L2 descriptors so they can be 409 * freed when empty. 410 */ 411 struct l2_dtable { 412 /* The number of L2 page descriptors allocated to this l2_dtable */ 413 u_int l2_occupancy; 414 415 /* List of L2 page descriptors */ 416 struct l2_bucket { 417 pt_entry_t *l2b_kva; /* KVA of L2 Descriptor Table */ 418 paddr_t l2b_phys; /* Physical address of same */ 419 u_short l2b_l1idx; /* This L2 table's L1 index */ 420 u_short l2b_occupancy; /* How many active descriptors */ 421 } l2_bucket[L2_BUCKET_SIZE]; 422 }; 423 424 /* 425 * Given an L1 table index, calculate the corresponding l2_dtable index 426 * and bucket index within the l2_dtable. 427 */ 428 #define L2_IDX(l1idx) (((l1idx) >> L2_BUCKET_LOG2) & \ 429 (L2_SIZE - 1)) 430 #define L2_BUCKET(l1idx) ((l1idx) & (L2_BUCKET_SIZE - 1)) 431 432 /* 433 * Given a virtual address, this macro returns the 434 * virtual address required to drop into the next L2 bucket. 435 */ 436 #define L2_NEXT_BUCKET(va) (((va) & L1_S_FRAME) + L1_S_SIZE) 437 438 /* 439 * L2 allocation. 440 */ 441 #define pmap_alloc_l2_dtable() \ 442 pool_cache_get(&pmap_l2dtable_cache, PR_NOWAIT) 443 #define pmap_free_l2_dtable(l2) \ 444 pool_cache_put(&pmap_l2dtable_cache, (l2)) 445 #define pmap_alloc_l2_ptp(pap) \ 446 ((pt_entry_t *)pool_cache_get_paddr(&pmap_l2ptp_cache,\ 447 PR_NOWAIT, (pap))) 448 449 /* 450 * We try to map the page tables write-through, if possible. However, not 451 * all CPUs have a write-through cache mode, so on those we have to sync 452 * the cache when we frob page tables. 453 * 454 * We try to evaluate this at compile time, if possible. However, it's 455 * not always possible to do that, hence this run-time var. 456 */ 457 int pmap_needs_pte_sync; 458 459 /* 460 * Real definition of pv_entry. 461 */ 462 struct pv_entry { 463 struct pv_entry *pv_next; /* next pv_entry */ 464 pmap_t pv_pmap; /* pmap where mapping lies */ 465 vaddr_t pv_va; /* virtual address for mapping */ 466 u_int pv_flags; /* flags */ 467 }; 468 469 /* 470 * Macro to determine if a mapping might be resident in the 471 * instruction cache and/or TLB 472 */ 473 #define PV_BEEN_EXECD(f) (((f) & (PVF_REF | PVF_EXEC)) == (PVF_REF | PVF_EXEC)) 474 475 /* 476 * Macro to determine if a mapping might be resident in the 477 * data cache and/or TLB 478 */ 479 #define PV_BEEN_REFD(f) (((f) & PVF_REF) != 0) 480 481 /* 482 * Local prototypes 483 */ 484 static int pmap_set_pt_cache_mode(pd_entry_t *, vaddr_t); 485 static void pmap_alloc_specials(vaddr_t *, int, vaddr_t *, 486 pt_entry_t **); 487 static boolean_t pmap_is_current(pmap_t); 488 static boolean_t pmap_is_cached(pmap_t); 489 static void pmap_enter_pv(struct vm_page *, struct pv_entry *, 490 pmap_t, vaddr_t, u_int); 491 static struct pv_entry *pmap_find_pv(struct vm_page *, pmap_t, vaddr_t); 492 static struct pv_entry *pmap_remove_pv(struct vm_page *, pmap_t, vaddr_t, int); 493 static u_int pmap_modify_pv(struct vm_page *, pmap_t, vaddr_t, 494 u_int, u_int); 495 496 static void pmap_pinit(pmap_t); 497 static int pmap_pmap_ctor(void *, void *, int); 498 499 static void pmap_alloc_l1(pmap_t); 500 static void pmap_free_l1(pmap_t); 501 static void pmap_use_l1(pmap_t); 502 503 static struct l2_bucket *pmap_get_l2_bucket(pmap_t, vaddr_t); 504 static struct l2_bucket *pmap_alloc_l2_bucket(pmap_t, vaddr_t); 505 static void pmap_free_l2_bucket(pmap_t, struct l2_bucket *, u_int); 506 static int pmap_l2ptp_ctor(void *, void *, int); 507 static int pmap_l2dtable_ctor(void *, void *, int); 508 509 static void pmap_vac_me_harder(struct vm_page *, pmap_t, vaddr_t); 510 static void pmap_vac_me_kpmap(struct vm_page *, pmap_t, vaddr_t); 511 static void pmap_vac_me_user(struct vm_page *, pmap_t, vaddr_t); 512 513 static void pmap_clearbit(struct vm_page *, u_int); 514 static int pmap_clean_page(struct pv_entry *, boolean_t); 515 static void pmap_page_remove(struct vm_page *); 516 517 static void pmap_init_l1(struct l1_ttable *, pd_entry_t *); 518 static vaddr_t kernel_pt_lookup(paddr_t); 519 520 521 /* 522 * External function prototypes 523 */ 524 extern void bzero_page(vaddr_t); 525 extern void bcopy_page(vaddr_t, vaddr_t); 526 527 /* 528 * Misc variables 529 */ 530 vaddr_t virtual_avail; 531 vaddr_t virtual_end; 532 vaddr_t pmap_curmaxkvaddr; 533 534 vaddr_t avail_start; 535 vaddr_t avail_end; 536 537 extern pv_addr_t systempage; 538 539 /* Function to set the debug level of the pmap code */ 540 541 #ifdef PMAP_DEBUG 542 void 543 pmap_debug(int level) 544 { 545 pmap_debug_level = level; 546 printf("pmap_debug: level=%d\n", pmap_debug_level); 547 } 548 #endif /* PMAP_DEBUG */ 549 550 /* 551 * A bunch of routines to conditionally flush the caches/TLB depending 552 * on whether the specified pmap actually needs to be flushed at any 553 * given time. 554 */ 555 static __inline void 556 pmap_tlb_flushID_SE(pmap_t pm, vaddr_t va) 557 { 558 559 if (pm->pm_cstate.cs_tlb_id) 560 cpu_tlb_flushID_SE(va); 561 } 562 563 static __inline void 564 pmap_tlb_flushD_SE(pmap_t pm, vaddr_t va) 565 { 566 567 if (pm->pm_cstate.cs_tlb_d) 568 cpu_tlb_flushD_SE(va); 569 } 570 571 static __inline void 572 pmap_tlb_flushID(pmap_t pm) 573 { 574 575 if (pm->pm_cstate.cs_tlb_id) { 576 cpu_tlb_flushID(); 577 pm->pm_cstate.cs_tlb = 0; 578 } 579 } 580 581 static __inline void 582 pmap_tlb_flushD(pmap_t pm) 583 { 584 585 if (pm->pm_cstate.cs_tlb_d) { 586 cpu_tlb_flushD(); 587 pm->pm_cstate.cs_tlb_d = 0; 588 } 589 } 590 591 static __inline void 592 pmap_idcache_wbinv_range(pmap_t pm, vaddr_t va, vsize_t len) 593 { 594 595 if (pm->pm_cstate.cs_cache_id) 596 cpu_idcache_wbinv_range(va, len); 597 } 598 599 static __inline void 600 pmap_dcache_wb_range(pmap_t pm, vaddr_t va, vsize_t len, 601 boolean_t do_inv, boolean_t rd_only) 602 { 603 604 if (pm->pm_cstate.cs_cache_d) { 605 if (do_inv) { 606 if (rd_only) 607 cpu_dcache_inv_range(va, len); 608 else 609 cpu_dcache_wbinv_range(va, len); 610 } else 611 if (!rd_only) 612 cpu_dcache_wb_range(va, len); 613 } 614 } 615 616 static __inline void 617 pmap_idcache_wbinv_all(pmap_t pm) 618 { 619 620 if (pm->pm_cstate.cs_cache_id) { 621 cpu_idcache_wbinv_all(); 622 pm->pm_cstate.cs_cache = 0; 623 } 624 } 625 626 static __inline void 627 pmap_dcache_wbinv_all(pmap_t pm) 628 { 629 630 if (pm->pm_cstate.cs_cache_d) { 631 cpu_dcache_wbinv_all(); 632 pm->pm_cstate.cs_cache_d = 0; 633 } 634 } 635 636 static __inline boolean_t 637 pmap_is_current(pmap_t pm) 638 { 639 640 if (pm == pmap_kernel() || 641 (curproc && curproc->p_vmspace->vm_map.pmap == pm)) 642 return (TRUE); 643 644 return (FALSE); 645 } 646 647 static __inline boolean_t 648 pmap_is_cached(pmap_t pm) 649 { 650 651 if (pm == pmap_kernel() || pmap_cache_state == NULL || 652 pmap_cache_state == &pm->pm_cstate) 653 return (TRUE); 654 655 return (FALSE); 656 } 657 658 /* 659 * PTE_SYNC_CURRENT: 660 * 661 * Make sure the pte is written out to RAM. 662 * We need to do this for one of two cases: 663 * - We're dealing with the kernel pmap 664 * - There is no pmap active in the cache/tlb. 665 * - The specified pmap is 'active' in the cache/tlb. 666 */ 667 #ifdef PMAP_INCLUDE_PTE_SYNC 668 #define PTE_SYNC_CURRENT(pm, ptep) \ 669 do { \ 670 if (PMAP_NEEDS_PTE_SYNC && \ 671 pmap_is_cached(pm)) \ 672 PTE_SYNC(ptep); \ 673 } while (/*CONSTCOND*/0) 674 #else 675 #define PTE_SYNC_CURRENT(pm, ptep) /* nothing */ 676 #endif 677 678 /* 679 * main pv_entry manipulation functions: 680 * pmap_enter_pv: enter a mapping onto a vm_page list 681 * pmap_remove_pv: remove a mappiing from a vm_page list 682 * 683 * NOTE: pmap_enter_pv expects to lock the pvh itself 684 * pmap_remove_pv expects te caller to lock the pvh before calling 685 */ 686 687 /* 688 * pmap_enter_pv: enter a mapping onto a vm_page lst 689 * 690 * => caller should hold the proper lock on pmap_main_lock 691 * => caller should have pmap locked 692 * => we will gain the lock on the vm_page and allocate the new pv_entry 693 * => caller should adjust ptp's wire_count before calling 694 * => caller should not adjust pmap's wire_count 695 */ 696 static void 697 pmap_enter_pv(struct vm_page *pg, struct pv_entry *pve, pmap_t pm, 698 vaddr_t va, u_int flags) 699 { 700 701 NPDEBUG(PDB_PVDUMP, 702 printf("pmap_enter_pv: pm %p, pg %p, flags 0x%x\n", pm, pg, flags)); 703 704 pve->pv_pmap = pm; 705 pve->pv_va = va; 706 pve->pv_flags = flags; 707 708 simple_lock(&pg->mdpage.pvh_slock); /* lock vm_page */ 709 pve->pv_next = pg->mdpage.pvh_list; /* add to ... */ 710 pg->mdpage.pvh_list = pve; /* ... locked list */ 711 pg->mdpage.pvh_attrs |= flags & (PVF_REF | PVF_MOD); 712 if (pm == pmap_kernel()) { 713 if (flags & PVF_WRITE) 714 pg->mdpage.krw_mappings++; 715 else 716 pg->mdpage.kro_mappings++; 717 } else 718 if (flags & PVF_WRITE) 719 pg->mdpage.urw_mappings++; 720 else 721 pg->mdpage.uro_mappings++; 722 simple_unlock(&pg->mdpage.pvh_slock); /* unlock, done! */ 723 724 if (pve->pv_flags & PVF_WIRED) 725 ++pm->pm_stats.wired_count; 726 } 727 728 /* 729 * 730 * pmap_find_pv: Find a pv entry 731 * 732 * => caller should hold lock on vm_page 733 */ 734 static __inline struct pv_entry * 735 pmap_find_pv(struct vm_page *pg, pmap_t pm, vaddr_t va) 736 { 737 struct pv_entry *pv; 738 739 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) { 740 if (pm == pv->pv_pmap && va == pv->pv_va) 741 break; 742 } 743 744 return (pv); 745 } 746 747 /* 748 * pmap_remove_pv: try to remove a mapping from a pv_list 749 * 750 * => caller should hold proper lock on pmap_main_lock 751 * => pmap should be locked 752 * => caller should hold lock on vm_page [so that attrs can be adjusted] 753 * => caller should adjust ptp's wire_count and free PTP if needed 754 * => caller should NOT adjust pmap's wire_count 755 * => we return the removed pve 756 */ 757 static struct pv_entry * 758 pmap_remove_pv(struct vm_page *pg, pmap_t pm, vaddr_t va, int skip_wired) 759 { 760 struct pv_entry *pve, **prevptr; 761 762 NPDEBUG(PDB_PVDUMP, 763 printf("pmap_remove_pv: pm %p, pg %p, va 0x%08lx\n", pm, pg, va)); 764 765 prevptr = &pg->mdpage.pvh_list; /* previous pv_entry pointer */ 766 pve = *prevptr; 767 768 while (pve) { 769 if (pve->pv_pmap == pm && pve->pv_va == va) { /* match? */ 770 NPDEBUG(PDB_PVDUMP, printf("pmap_remove_pv: pm %p, pg " 771 "%p, flags 0x%x\n", pm, pg, pve->pv_flags)); 772 if (pve->pv_flags & PVF_WIRED) { 773 if (skip_wired) 774 return (NULL); 775 --pm->pm_stats.wired_count; 776 } 777 *prevptr = pve->pv_next; /* remove it! */ 778 if (pm == pmap_kernel()) { 779 if (pve->pv_flags & PVF_WRITE) 780 pg->mdpage.krw_mappings--; 781 else 782 pg->mdpage.kro_mappings--; 783 } else 784 if (pve->pv_flags & PVF_WRITE) 785 pg->mdpage.urw_mappings--; 786 else 787 pg->mdpage.uro_mappings--; 788 break; 789 } 790 prevptr = &pve->pv_next; /* previous pointer */ 791 pve = pve->pv_next; /* advance */ 792 } 793 794 return(pve); /* return removed pve */ 795 } 796 797 /* 798 * 799 * pmap_modify_pv: Update pv flags 800 * 801 * => caller should hold lock on vm_page [so that attrs can be adjusted] 802 * => caller should NOT adjust pmap's wire_count 803 * => caller must call pmap_vac_me_harder() if writable status of a page 804 * may have changed. 805 * => we return the old flags 806 * 807 * Modify a physical-virtual mapping in the pv table 808 */ 809 static u_int 810 pmap_modify_pv(struct vm_page *pg, pmap_t pm, vaddr_t va, 811 u_int clr_mask, u_int set_mask) 812 { 813 struct pv_entry *npv; 814 u_int flags, oflags; 815 816 if ((npv = pmap_find_pv(pg, pm, va)) == NULL) 817 return (0); 818 819 NPDEBUG(PDB_PVDUMP, 820 printf("pmap_modify_pv: pm %p, pg %p, clr 0x%x, set 0x%x, flags 0x%x\n", pm, pg, clr_mask, set_mask, npv->pv_flags)); 821 822 /* 823 * There is at least one VA mapping this page. 824 */ 825 826 if (clr_mask & (PVF_REF | PVF_MOD)) 827 pg->mdpage.pvh_attrs |= set_mask & (PVF_REF | PVF_MOD); 828 829 oflags = npv->pv_flags; 830 npv->pv_flags = flags = (oflags & ~clr_mask) | set_mask; 831 832 if ((flags ^ oflags) & PVF_WIRED) { 833 if (flags & PVF_WIRED) 834 ++pm->pm_stats.wired_count; 835 else 836 --pm->pm_stats.wired_count; 837 } 838 839 if ((flags ^ oflags) & PVF_WRITE) { 840 if (pm == pmap_kernel()) { 841 if (flags & PVF_WRITE) { 842 pg->mdpage.krw_mappings++; 843 pg->mdpage.kro_mappings--; 844 } else { 845 pg->mdpage.kro_mappings++; 846 pg->mdpage.krw_mappings--; 847 } 848 } else 849 if (flags & PVF_WRITE) { 850 pg->mdpage.urw_mappings++; 851 pg->mdpage.uro_mappings--; 852 } else { 853 pg->mdpage.uro_mappings++; 854 pg->mdpage.urw_mappings--; 855 } 856 } 857 858 return (oflags); 859 } 860 861 static void 862 pmap_pinit(pmap_t pm) 863 { 864 865 if (vector_page < KERNEL_BASE) { 866 /* 867 * Map the vector page. 868 */ 869 pmap_enter(pm, vector_page, systempage.pv_pa, 870 VM_PROT_READ, VM_PROT_READ | PMAP_WIRED); 871 pmap_update(pm); 872 } 873 } 874 875 /* 876 * Allocate an L1 translation table for the specified pmap. 877 * This is called at pmap creation time. 878 */ 879 static void 880 pmap_alloc_l1(pmap_t pm) 881 { 882 struct l1_ttable *l1; 883 u_int8_t domain; 884 885 /* 886 * Remove the L1 at the head of the LRU list 887 */ 888 simple_lock(&l1_lru_lock); 889 l1 = TAILQ_FIRST(&l1_lru_list); 890 KDASSERT(l1 != NULL); 891 TAILQ_REMOVE(&l1_lru_list, l1, l1_lru); 892 893 /* 894 * Pick the first available domain number, and update 895 * the link to the next number. 896 */ 897 domain = l1->l1_domain_first; 898 l1->l1_domain_first = l1->l1_domain_free[domain]; 899 900 /* 901 * If there are still free domain numbers in this L1, 902 * put it back on the TAIL of the LRU list. 903 */ 904 if (++l1->l1_domain_use_count < PMAP_DOMAINS) 905 TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru); 906 907 simple_unlock(&l1_lru_lock); 908 909 /* 910 * Fix up the relevant bits in the pmap structure 911 */ 912 pm->pm_l1 = l1; 913 pm->pm_domain = domain; 914 } 915 916 /* 917 * Free an L1 translation table. 918 * This is called at pmap destruction time. 919 */ 920 static void 921 pmap_free_l1(pmap_t pm) 922 { 923 struct l1_ttable *l1 = pm->pm_l1; 924 925 simple_lock(&l1_lru_lock); 926 927 /* 928 * If this L1 is currently on the LRU list, remove it. 929 */ 930 if (l1->l1_domain_use_count < PMAP_DOMAINS) 931 TAILQ_REMOVE(&l1_lru_list, l1, l1_lru); 932 933 /* 934 * Free up the domain number which was allocated to the pmap 935 */ 936 l1->l1_domain_free[pm->pm_domain] = l1->l1_domain_first; 937 l1->l1_domain_first = pm->pm_domain; 938 l1->l1_domain_use_count--; 939 940 /* 941 * The L1 now must have at least 1 free domain, so add 942 * it back to the LRU list. If the use count is zero, 943 * put it at the head of the list, otherwise it goes 944 * to the tail. 945 */ 946 if (l1->l1_domain_use_count == 0) 947 TAILQ_INSERT_HEAD(&l1_lru_list, l1, l1_lru); 948 else 949 TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru); 950 951 simple_unlock(&l1_lru_lock); 952 } 953 954 static __inline void 955 pmap_use_l1(pmap_t pm) 956 { 957 struct l1_ttable *l1; 958 959 /* 960 * Do nothing if we're in interrupt context. 961 * Access to an L1 by the kernel pmap must not affect 962 * the LRU list. 963 */ 964 if (current_intr_depth || pm == pmap_kernel()) 965 return; 966 967 l1 = pm->pm_l1; 968 969 /* 970 * If the L1 is not currently on the LRU list, just return 971 */ 972 if (l1->l1_domain_use_count == PMAP_DOMAINS) 973 return; 974 975 simple_lock(&l1_lru_lock); 976 977 /* 978 * Check the use count again, now that we've acquired the lock 979 */ 980 if (l1->l1_domain_use_count == PMAP_DOMAINS) { 981 simple_unlock(&l1_lru_lock); 982 return; 983 } 984 985 /* 986 * Move the L1 to the back of the LRU list 987 */ 988 TAILQ_REMOVE(&l1_lru_list, l1, l1_lru); 989 TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru); 990 991 simple_unlock(&l1_lru_lock); 992 } 993 994 /* 995 * void pmap_free_l2_ptp(pt_entry_t *, paddr_t *) 996 * 997 * Free an L2 descriptor table. 998 */ 999 static __inline void 1000 #ifndef PMAP_INCLUDE_PTE_SYNC 1001 pmap_free_l2_ptp(pt_entry_t *l2, paddr_t pa) 1002 #else 1003 pmap_free_l2_ptp(boolean_t need_sync, pt_entry_t *l2, paddr_t pa) 1004 #endif 1005 { 1006 #ifdef PMAP_INCLUDE_PTE_SYNC 1007 /* 1008 * Note: With a write-back cache, we may need to sync this 1009 * L2 table before re-using it. 1010 * This is because it may have belonged to a non-current 1011 * pmap, in which case the cache syncs would have been 1012 * skipped when the pages were being unmapped. If the 1013 * L2 table were then to be immediately re-allocated to 1014 * the *current* pmap, it may well contain stale mappings 1015 * which have not yet been cleared by a cache write-back 1016 * and so would still be visible to the mmu. 1017 */ 1018 if (need_sync) 1019 PTE_SYNC_RANGE(l2, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t)); 1020 #endif 1021 pool_cache_put_paddr(&pmap_l2ptp_cache, (void *)l2, pa); 1022 } 1023 1024 /* 1025 * Returns a pointer to the L2 bucket associated with the specified pmap 1026 * and VA, or NULL if no L2 bucket exists for the address. 1027 */ 1028 static __inline struct l2_bucket * 1029 pmap_get_l2_bucket(pmap_t pm, vaddr_t va) 1030 { 1031 struct l2_dtable *l2; 1032 struct l2_bucket *l2b; 1033 u_short l1idx; 1034 1035 l1idx = L1_IDX(va); 1036 1037 if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL || 1038 (l2b = &l2->l2_bucket[L2_BUCKET(l1idx)])->l2b_kva == NULL) 1039 return (NULL); 1040 1041 return (l2b); 1042 } 1043 1044 /* 1045 * Returns a pointer to the L2 bucket associated with the specified pmap 1046 * and VA. 1047 * 1048 * If no L2 bucket exists, perform the necessary allocations to put an L2 1049 * bucket/page table in place. 1050 * 1051 * Note that if a new L2 bucket/page was allocated, the caller *must* 1052 * increment the bucket occupancy counter appropriately *before* 1053 * releasing the pmap's lock to ensure no other thread or cpu deallocates 1054 * the bucket/page in the meantime. 1055 */ 1056 static struct l2_bucket * 1057 pmap_alloc_l2_bucket(pmap_t pm, vaddr_t va) 1058 { 1059 struct l2_dtable *l2; 1060 struct l2_bucket *l2b; 1061 u_short l1idx; 1062 1063 l1idx = L1_IDX(va); 1064 1065 if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) { 1066 /* 1067 * No mapping at this address, as there is 1068 * no entry in the L1 table. 1069 * Need to allocate a new l2_dtable. 1070 */ 1071 if ((l2 = pmap_alloc_l2_dtable()) == NULL) 1072 return (NULL); 1073 1074 /* 1075 * Link it into the parent pmap 1076 */ 1077 pm->pm_l2[L2_IDX(l1idx)] = l2; 1078 } 1079 1080 l2b = &l2->l2_bucket[L2_BUCKET(l1idx)]; 1081 1082 /* 1083 * Fetch pointer to the L2 page table associated with the address. 1084 */ 1085 if (l2b->l2b_kva == NULL) { 1086 pt_entry_t *ptep; 1087 1088 /* 1089 * No L2 page table has been allocated. Chances are, this 1090 * is because we just allocated the l2_dtable, above. 1091 */ 1092 if ((ptep = pmap_alloc_l2_ptp(&l2b->l2b_phys)) == NULL) { 1093 /* 1094 * Oops, no more L2 page tables available at this 1095 * time. We may need to deallocate the l2_dtable 1096 * if we allocated a new one above. 1097 */ 1098 if (l2->l2_occupancy == 0) { 1099 pm->pm_l2[L2_IDX(l1idx)] = NULL; 1100 pmap_free_l2_dtable(l2); 1101 } 1102 return (NULL); 1103 } 1104 1105 l2->l2_occupancy++; 1106 l2b->l2b_kva = ptep; 1107 l2b->l2b_l1idx = l1idx; 1108 } 1109 1110 return (l2b); 1111 } 1112 1113 /* 1114 * One or more mappings in the specified L2 descriptor table have just been 1115 * invalidated. 1116 * 1117 * Garbage collect the metadata and descriptor table itself if necessary. 1118 * 1119 * The pmap lock must be acquired when this is called (not necessary 1120 * for the kernel pmap). 1121 */ 1122 static void 1123 pmap_free_l2_bucket(pmap_t pm, struct l2_bucket *l2b, u_int count) 1124 { 1125 struct l2_dtable *l2; 1126 pd_entry_t *pl1pd, l1pd; 1127 pt_entry_t *ptep; 1128 u_short l1idx; 1129 1130 KDASSERT(count <= l2b->l2b_occupancy); 1131 1132 /* 1133 * Update the bucket's reference count according to how many 1134 * PTEs the caller has just invalidated. 1135 */ 1136 l2b->l2b_occupancy -= count; 1137 1138 /* 1139 * Note: 1140 * 1141 * Level 2 page tables allocated to the kernel pmap are never freed 1142 * as that would require checking all Level 1 page tables and 1143 * removing any references to the Level 2 page table. See also the 1144 * comment elsewhere about never freeing bootstrap L2 descriptors. 1145 * 1146 * We make do with just invalidating the mapping in the L2 table. 1147 * 1148 * This isn't really a big deal in practice and, in fact, leads 1149 * to a performance win over time as we don't need to continually 1150 * alloc/free. 1151 */ 1152 if (l2b->l2b_occupancy > 0 || pm == pmap_kernel()) 1153 return; 1154 1155 /* 1156 * There are no more valid mappings in this level 2 page table. 1157 * Go ahead and NULL-out the pointer in the bucket, then 1158 * free the page table. 1159 */ 1160 l1idx = l2b->l2b_l1idx; 1161 ptep = l2b->l2b_kva; 1162 l2b->l2b_kva = NULL; 1163 1164 pl1pd = &pm->pm_l1->l1_kva[l1idx]; 1165 1166 /* 1167 * If the L1 slot matches the pmap's domain 1168 * number, then invalidate it. 1169 */ 1170 l1pd = *pl1pd & (L1_TYPE_MASK | L1_C_DOM_MASK); 1171 if (l1pd == (L1_C_DOM(pm->pm_domain) | L1_TYPE_C)) { 1172 *pl1pd = 0; 1173 PTE_SYNC(pl1pd); 1174 } 1175 1176 /* 1177 * Release the L2 descriptor table back to the pool cache. 1178 */ 1179 #ifndef PMAP_INCLUDE_PTE_SYNC 1180 pmap_free_l2_ptp(ptep, l2b->l2b_phys); 1181 #else 1182 pmap_free_l2_ptp(!pmap_is_cached(pm), ptep, l2b->l2b_phys); 1183 #endif 1184 1185 /* 1186 * Update the reference count in the associated l2_dtable 1187 */ 1188 l2 = pm->pm_l2[L2_IDX(l1idx)]; 1189 if (--l2->l2_occupancy > 0) 1190 return; 1191 1192 /* 1193 * There are no more valid mappings in any of the Level 1 1194 * slots managed by this l2_dtable. Go ahead and NULL-out 1195 * the pointer in the parent pmap and free the l2_dtable. 1196 */ 1197 pm->pm_l2[L2_IDX(l1idx)] = NULL; 1198 pmap_free_l2_dtable(l2); 1199 } 1200 1201 /* 1202 * Pool cache constructors for L2 descriptor tables, metadata and pmap 1203 * structures. 1204 */ 1205 static int 1206 pmap_l2ptp_ctor(void *arg, void *v, int flags) 1207 { 1208 #ifndef PMAP_INCLUDE_PTE_SYNC 1209 struct l2_bucket *l2b; 1210 pt_entry_t *ptep, pte; 1211 vaddr_t va = (vaddr_t)v & ~PGOFSET; 1212 1213 /* 1214 * The mappings for these page tables were initially made using 1215 * pmap_kenter_pa() by the pool subsystem. Therefore, the cache- 1216 * mode will not be right for page table mappings. To avoid 1217 * polluting the pmap_kenter_pa() code with a special case for 1218 * page tables, we simply fix up the cache-mode here if it's not 1219 * correct. 1220 */ 1221 l2b = pmap_get_l2_bucket(pmap_kernel(), va); 1222 KDASSERT(l2b != NULL); 1223 ptep = &l2b->l2b_kva[l2pte_index(va)]; 1224 pte = *ptep; 1225 1226 if ((pte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) { 1227 /* 1228 * Page tables must have the cache-mode set to Write-Thru. 1229 */ 1230 *ptep = (pte & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode_pt; 1231 PTE_SYNC(ptep); 1232 cpu_tlb_flushD_SE(va); 1233 cpu_cpwait(); 1234 } 1235 #endif 1236 1237 memset(v, 0, L2_TABLE_SIZE_REAL); 1238 PTE_SYNC_RANGE(v, L2_TABLE_SIZE_REAL / sizeof(pt_entry_t)); 1239 return (0); 1240 } 1241 1242 static int 1243 pmap_l2dtable_ctor(void *arg, void *v, int flags) 1244 { 1245 1246 memset(v, 0, sizeof(struct l2_dtable)); 1247 return (0); 1248 } 1249 1250 static int 1251 pmap_pmap_ctor(void *arg, void *v, int flags) 1252 { 1253 1254 memset(v, 0, sizeof(struct pmap)); 1255 return (0); 1256 } 1257 1258 /* 1259 * Since we have a virtually indexed cache, we may need to inhibit caching if 1260 * there is more than one mapping and at least one of them is writable. 1261 * Since we purge the cache on every context switch, we only need to check for 1262 * other mappings within the same pmap, or kernel_pmap. 1263 * This function is also called when a page is unmapped, to possibly reenable 1264 * caching on any remaining mappings. 1265 * 1266 * The code implements the following logic, where: 1267 * 1268 * KW = # of kernel read/write pages 1269 * KR = # of kernel read only pages 1270 * UW = # of user read/write pages 1271 * UR = # of user read only pages 1272 * 1273 * KC = kernel mapping is cacheable 1274 * UC = user mapping is cacheable 1275 * 1276 * KW=0,KR=0 KW=0,KR>0 KW=1,KR=0 KW>1,KR>=0 1277 * +--------------------------------------------- 1278 * UW=0,UR=0 | --- KC=1 KC=1 KC=0 1279 * UW=0,UR>0 | UC=1 KC=1,UC=1 KC=0,UC=0 KC=0,UC=0 1280 * UW=1,UR=0 | UC=1 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0 1281 * UW>1,UR>=0 | UC=0 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0 1282 */ 1283 1284 static const int pmap_vac_flags[4][4] = { 1285 {-1, 0, 0, PVF_KNC}, 1286 {0, 0, PVF_NC, PVF_NC}, 1287 {0, PVF_NC, PVF_NC, PVF_NC}, 1288 {PVF_UNC, PVF_NC, PVF_NC, PVF_NC} 1289 }; 1290 1291 static __inline int 1292 pmap_get_vac_flags(const struct vm_page *pg) 1293 { 1294 int kidx, uidx; 1295 1296 kidx = 0; 1297 if (pg->mdpage.kro_mappings || pg->mdpage.krw_mappings > 1) 1298 kidx |= 1; 1299 if (pg->mdpage.krw_mappings) 1300 kidx |= 2; 1301 1302 uidx = 0; 1303 if (pg->mdpage.uro_mappings || pg->mdpage.urw_mappings > 1) 1304 uidx |= 1; 1305 if (pg->mdpage.urw_mappings) 1306 uidx |= 2; 1307 1308 return (pmap_vac_flags[uidx][kidx]); 1309 } 1310 1311 static __inline void 1312 pmap_vac_me_harder(struct vm_page *pg, pmap_t pm, vaddr_t va) 1313 { 1314 int nattr; 1315 1316 nattr = pmap_get_vac_flags(pg); 1317 1318 if (nattr < 0) { 1319 pg->mdpage.pvh_attrs &= ~PVF_NC; 1320 return; 1321 } 1322 1323 if (nattr == 0 && (pg->mdpage.pvh_attrs & PVF_NC) == 0) 1324 return; 1325 1326 if (pm == pmap_kernel()) 1327 pmap_vac_me_kpmap(pg, pm, va); 1328 else 1329 pmap_vac_me_user(pg, pm, va); 1330 1331 pg->mdpage.pvh_attrs = (pg->mdpage.pvh_attrs & ~PVF_NC) | nattr; 1332 } 1333 1334 static void 1335 pmap_vac_me_kpmap(struct vm_page *pg, pmap_t pm, vaddr_t va) 1336 { 1337 u_int u_cacheable, u_entries; 1338 struct pv_entry *pv; 1339 pmap_t last_pmap = pm; 1340 1341 /* 1342 * Pass one, see if there are both kernel and user pmaps for 1343 * this page. Calculate whether there are user-writable or 1344 * kernel-writable pages. 1345 */ 1346 u_cacheable = 0; 1347 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) { 1348 if (pv->pv_pmap != pm && (pv->pv_flags & PVF_NC) == 0) 1349 u_cacheable++; 1350 } 1351 1352 u_entries = pg->mdpage.urw_mappings + pg->mdpage.uro_mappings; 1353 1354 /* 1355 * We know we have just been updating a kernel entry, so if 1356 * all user pages are already cacheable, then there is nothing 1357 * further to do. 1358 */ 1359 if (pg->mdpage.k_mappings == 0 && u_cacheable == u_entries) 1360 return; 1361 1362 if (u_entries) { 1363 /* 1364 * Scan over the list again, for each entry, if it 1365 * might not be set correctly, call pmap_vac_me_user 1366 * to recalculate the settings. 1367 */ 1368 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) { 1369 /* 1370 * We know kernel mappings will get set 1371 * correctly in other calls. We also know 1372 * that if the pmap is the same as last_pmap 1373 * then we've just handled this entry. 1374 */ 1375 if (pv->pv_pmap == pm || pv->pv_pmap == last_pmap) 1376 continue; 1377 1378 /* 1379 * If there are kernel entries and this page 1380 * is writable but non-cacheable, then we can 1381 * skip this entry also. 1382 */ 1383 if (pg->mdpage.k_mappings && 1384 (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 1385 (PVF_NC | PVF_WRITE)) 1386 continue; 1387 1388 /* 1389 * Similarly if there are no kernel-writable 1390 * entries and the page is already 1391 * read-only/cacheable. 1392 */ 1393 if (pg->mdpage.krw_mappings == 0 && 1394 (pv->pv_flags & (PVF_NC | PVF_WRITE)) == 0) 1395 continue; 1396 1397 /* 1398 * For some of the remaining cases, we know 1399 * that we must recalculate, but for others we 1400 * can't tell if they are correct or not, so 1401 * we recalculate anyway. 1402 */ 1403 pmap_vac_me_user(pg, (last_pmap = pv->pv_pmap), 0); 1404 } 1405 1406 if (pg->mdpage.k_mappings == 0) 1407 return; 1408 } 1409 1410 pmap_vac_me_user(pg, pm, va); 1411 } 1412 1413 static void 1414 pmap_vac_me_user(struct vm_page *pg, pmap_t pm, vaddr_t va) 1415 { 1416 pmap_t kpmap = pmap_kernel(); 1417 struct pv_entry *pv, *npv; 1418 struct l2_bucket *l2b; 1419 pt_entry_t *ptep, pte; 1420 u_int entries = 0; 1421 u_int writable = 0; 1422 u_int cacheable_entries = 0; 1423 u_int kern_cacheable = 0; 1424 u_int other_writable = 0; 1425 1426 /* 1427 * Count mappings and writable mappings in this pmap. 1428 * Include kernel mappings as part of our own. 1429 * Keep a pointer to the first one. 1430 */ 1431 for (pv = npv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) { 1432 /* Count mappings in the same pmap */ 1433 if (pm == pv->pv_pmap || kpmap == pv->pv_pmap) { 1434 if (entries++ == 0) 1435 npv = pv; 1436 1437 /* Cacheable mappings */ 1438 if ((pv->pv_flags & PVF_NC) == 0) { 1439 cacheable_entries++; 1440 if (kpmap == pv->pv_pmap) 1441 kern_cacheable++; 1442 } 1443 1444 /* Writable mappings */ 1445 if (pv->pv_flags & PVF_WRITE) 1446 ++writable; 1447 } else 1448 if (pv->pv_flags & PVF_WRITE) 1449 other_writable = 1; 1450 } 1451 1452 /* 1453 * Enable or disable caching as necessary. 1454 * Note: the first entry might be part of the kernel pmap, 1455 * so we can't assume this is indicative of the state of the 1456 * other (maybe non-kpmap) entries. 1457 */ 1458 if ((entries > 1 && writable) || 1459 (entries > 0 && pm == kpmap && other_writable)) { 1460 if (cacheable_entries == 0) 1461 return; 1462 1463 for (pv = npv; pv; pv = pv->pv_next) { 1464 if ((pm != pv->pv_pmap && kpmap != pv->pv_pmap) || 1465 (pv->pv_flags & PVF_NC)) 1466 continue; 1467 1468 pv->pv_flags |= PVF_NC; 1469 1470 l2b = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va); 1471 ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)]; 1472 pte = *ptep & ~L2_S_CACHE_MASK; 1473 1474 if ((va != pv->pv_va || pm != pv->pv_pmap) && 1475 l2pte_valid(pte)) { 1476 if (PV_BEEN_EXECD(pv->pv_flags)) { 1477 pmap_idcache_wbinv_range(pv->pv_pmap, 1478 pv->pv_va, PAGE_SIZE); 1479 pmap_tlb_flushID_SE(pv->pv_pmap, 1480 pv->pv_va); 1481 } else 1482 if (PV_BEEN_REFD(pv->pv_flags)) { 1483 pmap_dcache_wb_range(pv->pv_pmap, 1484 pv->pv_va, PAGE_SIZE, TRUE, 1485 (pv->pv_flags & PVF_WRITE) == 0); 1486 pmap_tlb_flushD_SE(pv->pv_pmap, 1487 pv->pv_va); 1488 } 1489 } 1490 1491 *ptep = pte; 1492 PTE_SYNC_CURRENT(pv->pv_pmap, ptep); 1493 } 1494 cpu_cpwait(); 1495 } else 1496 if (entries > cacheable_entries) { 1497 /* 1498 * Turn cacheing back on for some pages. If it is a kernel 1499 * page, only do so if there are no other writable pages. 1500 */ 1501 for (pv = npv; pv; pv = pv->pv_next) { 1502 if (!(pv->pv_flags & PVF_NC) || (pm != pv->pv_pmap && 1503 (kpmap != pv->pv_pmap || other_writable))) 1504 continue; 1505 1506 pv->pv_flags &= ~PVF_NC; 1507 1508 l2b = pmap_get_l2_bucket(pv->pv_pmap, pv->pv_va); 1509 ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)]; 1510 pte = (*ptep & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode; 1511 1512 if (l2pte_valid(pte)) { 1513 if (PV_BEEN_EXECD(pv->pv_flags)) { 1514 pmap_tlb_flushID_SE(pv->pv_pmap, 1515 pv->pv_va); 1516 } else 1517 if (PV_BEEN_REFD(pv->pv_flags)) { 1518 pmap_tlb_flushD_SE(pv->pv_pmap, 1519 pv->pv_va); 1520 } 1521 } 1522 1523 *ptep = pte; 1524 PTE_SYNC_CURRENT(pv->pv_pmap, ptep); 1525 } 1526 } 1527 } 1528 1529 /* 1530 * Modify pte bits for all ptes corresponding to the given physical address. 1531 * We use `maskbits' rather than `clearbits' because we're always passing 1532 * constants and the latter would require an extra inversion at run-time. 1533 */ 1534 static void 1535 pmap_clearbit(struct vm_page *pg, u_int maskbits) 1536 { 1537 struct l2_bucket *l2b; 1538 struct pv_entry *pv; 1539 pt_entry_t *ptep, npte, opte; 1540 pmap_t pm; 1541 vaddr_t va; 1542 u_int oflags; 1543 1544 NPDEBUG(PDB_BITS, 1545 printf("pmap_clearbit: pg %p (0x%08lx) mask 0x%x\n", 1546 pg, VM_PAGE_TO_PHYS(pg), maskbits)); 1547 1548 PMAP_HEAD_TO_MAP_LOCK(); 1549 simple_lock(&pg->mdpage.pvh_slock); 1550 1551 /* 1552 * Clear saved attributes (modify, reference) 1553 */ 1554 pg->mdpage.pvh_attrs &= ~(maskbits & (PVF_MOD | PVF_REF)); 1555 1556 if (pg->mdpage.pvh_list == NULL) { 1557 simple_unlock(&pg->mdpage.pvh_slock); 1558 PMAP_HEAD_TO_MAP_UNLOCK(); 1559 return; 1560 } 1561 1562 /* 1563 * Loop over all current mappings setting/clearing as appropos 1564 */ 1565 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) { 1566 va = pv->pv_va; 1567 pm = pv->pv_pmap; 1568 oflags = pv->pv_flags; 1569 pv->pv_flags &= ~maskbits; 1570 1571 pmap_acquire_pmap_lock(pm); 1572 1573 l2b = pmap_get_l2_bucket(pm, va); 1574 KDASSERT(l2b != NULL); 1575 1576 ptep = &l2b->l2b_kva[l2pte_index(va)]; 1577 npte = opte = *ptep; 1578 1579 NPDEBUG(PDB_BITS, 1580 printf( 1581 "pmap_clearbit: pv %p, pm %p, va 0x%08lx, flag 0x%x\n", 1582 pv, pv->pv_pmap, pv->pv_va, oflags)); 1583 1584 if (maskbits & (PVF_WRITE|PVF_MOD)) { 1585 if ((pv->pv_flags & PVF_NC)) { 1586 /* 1587 * Entry is not cacheable: 1588 * 1589 * Don't turn caching on again if this is a 1590 * modified emulation. This would be 1591 * inconsitent with the settings created by 1592 * pmap_vac_me_harder(). Otherwise, it's safe 1593 * to re-enable cacheing. 1594 * 1595 * There's no need to call pmap_vac_me_harder() 1596 * here: all pages are losing their write 1597 * permission. 1598 */ 1599 if (maskbits & PVF_WRITE) { 1600 npte |= pte_l2_s_cache_mode; 1601 pv->pv_flags &= ~PVF_NC; 1602 } 1603 } else 1604 if (opte & L2_S_PROT_W) { 1605 /* 1606 * Entry is writable/cacheable: check if pmap 1607 * is current if it is flush it, otherwise it 1608 * won't be in the cache 1609 */ 1610 if (PV_BEEN_EXECD(oflags)) 1611 pmap_idcache_wbinv_range(pm, pv->pv_va, 1612 PAGE_SIZE); 1613 else 1614 if (PV_BEEN_REFD(oflags)) 1615 pmap_dcache_wb_range(pm, pv->pv_va, 1616 PAGE_SIZE, 1617 (maskbits & PVF_REF) ? TRUE : FALSE, 1618 FALSE); 1619 } 1620 1621 /* make the pte read only */ 1622 npte &= ~L2_S_PROT_W; 1623 1624 if (maskbits & PVF_WRITE) { 1625 /* 1626 * Keep alias accounting up to date 1627 */ 1628 if (pv->pv_pmap == pmap_kernel()) { 1629 if (oflags & PVF_WRITE) { 1630 pg->mdpage.krw_mappings--; 1631 pg->mdpage.kro_mappings++; 1632 } 1633 } else 1634 if (oflags & PVF_WRITE) { 1635 pg->mdpage.urw_mappings--; 1636 pg->mdpage.uro_mappings++; 1637 } 1638 } 1639 } 1640 1641 if (maskbits & PVF_REF) { 1642 if ((pv->pv_flags & PVF_NC) == 0 && 1643 (maskbits & (PVF_WRITE|PVF_MOD)) == 0) { 1644 /* 1645 * Check npte here; we may have already 1646 * done the wbinv above, and the validity 1647 * of the PTE is the same for opte and 1648 * npte. 1649 */ 1650 if (npte & L2_S_PROT_W) { 1651 if (PV_BEEN_EXECD(oflags)) 1652 pmap_idcache_wbinv_range(pm, 1653 pv->pv_va, PAGE_SIZE); 1654 else 1655 if (PV_BEEN_REFD(oflags)) 1656 pmap_dcache_wb_range(pm, 1657 pv->pv_va, PAGE_SIZE, 1658 TRUE, FALSE); 1659 } else 1660 if ((npte & L2_TYPE_MASK) != L2_TYPE_INV) { 1661 /* XXXJRT need idcache_inv_range */ 1662 if (PV_BEEN_EXECD(oflags)) 1663 pmap_idcache_wbinv_range(pm, 1664 pv->pv_va, PAGE_SIZE); 1665 else 1666 if (PV_BEEN_REFD(oflags)) 1667 pmap_dcache_wb_range(pm, 1668 pv->pv_va, PAGE_SIZE, 1669 TRUE, TRUE); 1670 } 1671 } 1672 1673 /* 1674 * Make the PTE invalid so that we will take a 1675 * page fault the next time the mapping is 1676 * referenced. 1677 */ 1678 npte &= ~L2_TYPE_MASK; 1679 npte |= L2_TYPE_INV; 1680 } 1681 1682 if (npte != opte) { 1683 *ptep = npte; 1684 PTE_SYNC(ptep); 1685 /* Flush the TLB entry if a current pmap. */ 1686 if (PV_BEEN_EXECD(oflags)) 1687 pmap_tlb_flushID_SE(pm, pv->pv_va); 1688 else 1689 if (PV_BEEN_REFD(oflags)) 1690 pmap_tlb_flushD_SE(pm, pv->pv_va); 1691 } 1692 1693 pmap_release_pmap_lock(pm); 1694 1695 NPDEBUG(PDB_BITS, 1696 printf("pmap_clearbit: pm %p va 0x%lx opte 0x%08x npte 0x%08x\n", 1697 pm, va, opte, npte)); 1698 } 1699 1700 simple_unlock(&pg->mdpage.pvh_slock); 1701 PMAP_HEAD_TO_MAP_UNLOCK(); 1702 } 1703 1704 /* 1705 * pmap_clean_page() 1706 * 1707 * This is a local function used to work out the best strategy to clean 1708 * a single page referenced by its entry in the PV table. It's used by 1709 * pmap_copy_page, pmap_zero page and maybe some others later on. 1710 * 1711 * Its policy is effectively: 1712 * o If there are no mappings, we don't bother doing anything with the cache. 1713 * o If there is one mapping, we clean just that page. 1714 * o If there are multiple mappings, we clean the entire cache. 1715 * 1716 * So that some functions can be further optimised, it returns 0 if it didn't 1717 * clean the entire cache, or 1 if it did. 1718 * 1719 * XXX One bug in this routine is that if the pv_entry has a single page 1720 * mapped at 0x00000000 a whole cache clean will be performed rather than 1721 * just the 1 page. Since this should not occur in everyday use and if it does 1722 * it will just result in not the most efficient clean for the page. 1723 */ 1724 static int 1725 pmap_clean_page(struct pv_entry *pv, boolean_t is_src) 1726 { 1727 pmap_t pm, pm_to_clean = NULL; 1728 struct pv_entry *npv; 1729 u_int cache_needs_cleaning = 0; 1730 u_int flags = 0; 1731 vaddr_t page_to_clean = 0; 1732 1733 if (pv == NULL) { 1734 /* nothing mapped in so nothing to flush */ 1735 return (0); 1736 } 1737 1738 /* 1739 * Since we flush the cache each time we change to a different 1740 * user vmspace, we only need to flush the page if it is in the 1741 * current pmap. 1742 */ 1743 if (curproc) 1744 pm = curproc->p_vmspace->vm_map.pmap; 1745 else 1746 pm = pmap_kernel(); 1747 1748 for (npv = pv; npv; npv = npv->pv_next) { 1749 if (npv->pv_pmap == pmap_kernel() || npv->pv_pmap == pm) { 1750 flags |= npv->pv_flags; 1751 /* 1752 * The page is mapped non-cacheable in 1753 * this map. No need to flush the cache. 1754 */ 1755 if (npv->pv_flags & PVF_NC) { 1756 #ifdef DIAGNOSTIC 1757 if (cache_needs_cleaning) 1758 panic("pmap_clean_page: " 1759 "cache inconsistency"); 1760 #endif 1761 break; 1762 } else if (is_src && (npv->pv_flags & PVF_WRITE) == 0) 1763 continue; 1764 if (cache_needs_cleaning) { 1765 page_to_clean = 0; 1766 break; 1767 } else { 1768 page_to_clean = npv->pv_va; 1769 pm_to_clean = npv->pv_pmap; 1770 } 1771 cache_needs_cleaning = 1; 1772 } 1773 } 1774 1775 if (page_to_clean) { 1776 if (PV_BEEN_EXECD(flags)) 1777 pmap_idcache_wbinv_range(pm_to_clean, page_to_clean, 1778 PAGE_SIZE); 1779 else 1780 pmap_dcache_wb_range(pm_to_clean, page_to_clean, 1781 PAGE_SIZE, !is_src, (flags & PVF_WRITE) == 0); 1782 } else if (cache_needs_cleaning) { 1783 if (PV_BEEN_EXECD(flags)) 1784 pmap_idcache_wbinv_all(pm); 1785 else 1786 pmap_dcache_wbinv_all(pm); 1787 return (1); 1788 } 1789 return (0); 1790 } 1791 1792 /* 1793 * Routine: pmap_page_remove 1794 * Function: 1795 * Removes this physical page from 1796 * all physical maps in which it resides. 1797 * Reflects back modify bits to the pager. 1798 */ 1799 static void 1800 pmap_page_remove(struct vm_page *pg) 1801 { 1802 struct l2_bucket *l2b; 1803 struct pv_entry *pv, *npv; 1804 pmap_t pm, curpm; 1805 pt_entry_t *ptep, pte; 1806 boolean_t flush; 1807 u_int flags; 1808 1809 NPDEBUG(PDB_FOLLOW, 1810 printf("pmap_page_remove: pg %p (0x%08lx)\n", pg, 1811 VM_PAGE_TO_PHYS(pg))); 1812 1813 PMAP_HEAD_TO_MAP_LOCK(); 1814 simple_lock(&pg->mdpage.pvh_slock); 1815 1816 pv = pg->mdpage.pvh_list; 1817 if (pv == NULL) { 1818 simple_unlock(&pg->mdpage.pvh_slock); 1819 PMAP_HEAD_TO_MAP_UNLOCK(); 1820 return; 1821 } 1822 1823 /* 1824 * Clear alias counts 1825 */ 1826 pg->mdpage.k_mappings = 0; 1827 pg->mdpage.urw_mappings = pg->mdpage.uro_mappings = 0; 1828 1829 flush = FALSE; 1830 flags = 0; 1831 if (curproc) 1832 curpm = curproc->p_vmspace->vm_map.pmap; 1833 else 1834 curpm = pmap_kernel(); 1835 1836 pmap_clean_page(pv, FALSE); 1837 1838 while (pv) { 1839 pm = pv->pv_pmap; 1840 if (flush == FALSE && (pm == curpm || pm == pmap_kernel())) 1841 flush = TRUE; 1842 1843 pmap_acquire_pmap_lock(pm); 1844 1845 l2b = pmap_get_l2_bucket(pm, pv->pv_va); 1846 KDASSERT(l2b != NULL); 1847 1848 ptep = &l2b->l2b_kva[l2pte_index(pv->pv_va)]; 1849 pte = *ptep; 1850 1851 /* 1852 * Update statistics 1853 */ 1854 --pm->pm_stats.resident_count; 1855 1856 /* Wired bit */ 1857 if (pv->pv_flags & PVF_WIRED) 1858 --pm->pm_stats.wired_count; 1859 1860 flags |= pv->pv_flags; 1861 1862 /* 1863 * Invalidate the PTEs. 1864 */ 1865 *ptep = 0; 1866 PTE_SYNC_CURRENT(pm, ptep); 1867 pmap_free_l2_bucket(pm, l2b, 1); 1868 1869 npv = pv->pv_next; 1870 pool_put(&pmap_pv_pool, pv); 1871 pv = npv; 1872 pmap_release_pmap_lock(pm); 1873 } 1874 pg->mdpage.pvh_list = NULL; 1875 simple_unlock(&pg->mdpage.pvh_slock); 1876 PMAP_HEAD_TO_MAP_UNLOCK(); 1877 1878 if (flush) { 1879 /* 1880 * Note: We can't use pmap_tlb_flush{I,}D() here since that 1881 * would need a subsequent call to pmap_update() to ensure 1882 * curpm->pm_cstate.cs_all is reset. Our callers are not 1883 * required to do that (see pmap(9)), so we can't modify 1884 * the current pmap's state. 1885 */ 1886 if (PV_BEEN_EXECD(flags)) 1887 cpu_tlb_flushID(); 1888 else 1889 cpu_tlb_flushD(); 1890 } 1891 cpu_cpwait(); 1892 } 1893 1894 /* 1895 * pmap_t pmap_create(void) 1896 * 1897 * Create a new pmap structure from scratch. 1898 */ 1899 pmap_t 1900 pmap_create(void) 1901 { 1902 pmap_t pm; 1903 1904 pm = pool_cache_get(&pmap_pmap_cache, PR_WAITOK); 1905 1906 simple_lock_init(&pm->pm_lock); 1907 pm->pm_obj.pgops = NULL; /* currently not a mappable object */ 1908 TAILQ_INIT(&pm->pm_obj.memq); 1909 pm->pm_obj.uo_npages = 0; 1910 pm->pm_obj.uo_refs = 1; 1911 pm->pm_stats.wired_count = 0; 1912 pm->pm_stats.resident_count = 1; 1913 pm->pm_cstate.cs_all = 0; 1914 pmap_alloc_l1(pm); 1915 1916 /* 1917 * Note: The pool cache ensures that the pm_l2[] array is already 1918 * initialised to zero. 1919 */ 1920 1921 pmap_pinit(pm); 1922 1923 LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list); 1924 1925 return (pm); 1926 } 1927 1928 /* 1929 * void pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot, 1930 * int flags) 1931 * 1932 * Insert the given physical page (p) at 1933 * the specified virtual address (v) in the 1934 * target physical map with the protection requested. 1935 * 1936 * NB: This is the only routine which MAY NOT lazy-evaluate 1937 * or lose information. That is, this routine must actually 1938 * insert this page into the given map NOW. 1939 */ 1940 int 1941 pmap_enter(pmap_t pm, vaddr_t va, paddr_t pa, vm_prot_t prot, int flags) 1942 { 1943 struct l2_bucket *l2b; 1944 struct vm_page *pg, *opg; 1945 struct pv_entry *pve; 1946 pt_entry_t *ptep, npte, opte; 1947 u_int nflags; 1948 u_int oflags; 1949 1950 NPDEBUG(PDB_ENTER, printf("pmap_enter: pm %p va 0x%lx pa 0x%lx prot %x flag %x\n", pm, va, pa, prot, flags)); 1951 1952 KDASSERT((flags & PMAP_WIRED) == 0 || (flags & VM_PROT_ALL) != 0); 1953 KDASSERT(((va | pa) & PGOFSET) == 0); 1954 1955 /* 1956 * Get a pointer to the page. Later on in this function, we 1957 * test for a managed page by checking pg != NULL. 1958 */ 1959 pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL; 1960 1961 nflags = 0; 1962 if (prot & VM_PROT_WRITE) 1963 nflags |= PVF_WRITE; 1964 if (prot & VM_PROT_EXECUTE) 1965 nflags |= PVF_EXEC; 1966 if (flags & PMAP_WIRED) 1967 nflags |= PVF_WIRED; 1968 1969 PMAP_MAP_TO_HEAD_LOCK(); 1970 pmap_acquire_pmap_lock(pm); 1971 1972 /* 1973 * Fetch the L2 bucket which maps this page, allocating one if 1974 * necessary for user pmaps. 1975 */ 1976 if (pm == pmap_kernel()) 1977 l2b = pmap_get_l2_bucket(pm, va); 1978 else 1979 l2b = pmap_alloc_l2_bucket(pm, va); 1980 if (l2b == NULL) { 1981 if (flags & PMAP_CANFAIL) { 1982 pmap_release_pmap_lock(pm); 1983 PMAP_MAP_TO_HEAD_UNLOCK(); 1984 return (ENOMEM); 1985 } 1986 panic("pmap_enter: failed to allocate L2 bucket"); 1987 } 1988 ptep = &l2b->l2b_kva[l2pte_index(va)]; 1989 opte = *ptep; 1990 npte = pa; 1991 oflags = 0; 1992 1993 if (opte) { 1994 /* 1995 * There is already a mapping at this address. 1996 * If the physical address is different, lookup the 1997 * vm_page. 1998 */ 1999 if (l2pte_pa(opte) != pa) 2000 opg = PHYS_TO_VM_PAGE(l2pte_pa(opte)); 2001 else 2002 opg = pg; 2003 } else 2004 opg = NULL; 2005 2006 if (pg) { 2007 /* 2008 * This is to be a managed mapping. 2009 */ 2010 if ((flags & VM_PROT_ALL) || 2011 (pg->mdpage.pvh_attrs & PVF_REF)) { 2012 /* 2013 * - The access type indicates that we don't need 2014 * to do referenced emulation. 2015 * OR 2016 * - The physical page has already been referenced 2017 * so no need to re-do referenced emulation here. 2018 */ 2019 npte |= L2_S_PROTO; 2020 2021 nflags |= PVF_REF; 2022 2023 if ((prot & VM_PROT_WRITE) != 0 && 2024 ((flags & VM_PROT_WRITE) != 0 || 2025 (pg->mdpage.pvh_attrs & PVF_MOD) != 0)) { 2026 /* 2027 * This is a writable mapping, and the 2028 * page's mod state indicates it has 2029 * already been modified. Make it 2030 * writable from the outset. 2031 */ 2032 npte |= L2_S_PROT_W; 2033 nflags |= PVF_MOD; 2034 } 2035 } else { 2036 /* 2037 * Need to do page referenced emulation. 2038 */ 2039 npte |= L2_TYPE_INV; 2040 } 2041 2042 npte |= pte_l2_s_cache_mode; 2043 2044 if (pg == opg) { 2045 /* 2046 * We're changing the attrs of an existing mapping. 2047 */ 2048 simple_lock(&pg->mdpage.pvh_slock); 2049 oflags = pmap_modify_pv(pg, pm, va, 2050 PVF_WRITE | PVF_EXEC | PVF_WIRED | 2051 PVF_MOD | PVF_REF, nflags); 2052 simple_unlock(&pg->mdpage.pvh_slock); 2053 2054 /* 2055 * We may need to flush the cache if we're 2056 * doing rw-ro... 2057 */ 2058 if (pm->pm_cstate.cs_cache_d && 2059 (oflags & PVF_NC) == 0 && 2060 (opte & L2_S_PROT_W) != 0 && 2061 (prot & VM_PROT_WRITE) == 0) 2062 cpu_dcache_wb_range(va, PAGE_SIZE); 2063 } else { 2064 /* 2065 * New mapping, or changing the backing page 2066 * of an existing mapping. 2067 */ 2068 if (opg) { 2069 /* 2070 * Replacing an existing mapping with a new one. 2071 * It is part of our managed memory so we 2072 * must remove it from the PV list 2073 */ 2074 simple_lock(&opg->mdpage.pvh_slock); 2075 pve = pmap_remove_pv(opg, pm, va, 0); 2076 pmap_vac_me_harder(opg, pm, 0); 2077 simple_unlock(&opg->mdpage.pvh_slock); 2078 oflags = pve->pv_flags; 2079 2080 /* 2081 * If the old mapping was valid (ref/mod 2082 * emulation creates 'invalid' mappings 2083 * initially) then make sure to frob 2084 * the cache. 2085 */ 2086 if ((oflags & PVF_NC) == 0 && 2087 l2pte_valid(opte)) { 2088 if (PV_BEEN_EXECD(oflags)) { 2089 pmap_idcache_wbinv_range(pm, va, 2090 PAGE_SIZE); 2091 } else 2092 if (PV_BEEN_REFD(oflags)) { 2093 pmap_dcache_wb_range(pm, va, 2094 PAGE_SIZE, TRUE, 2095 (oflags & PVF_WRITE) == 0); 2096 } 2097 } 2098 } else 2099 if ((pve = pool_get(&pmap_pv_pool, PR_NOWAIT)) == NULL){ 2100 if ((flags & PMAP_CANFAIL) == 0) 2101 panic("pmap_enter: no pv entries"); 2102 2103 if (pm != pmap_kernel()) 2104 pmap_free_l2_bucket(pm, l2b, 0); 2105 pmap_release_pmap_lock(pm); 2106 PMAP_MAP_TO_HEAD_UNLOCK(); 2107 NPDEBUG(PDB_ENTER, 2108 printf("pmap_enter: ENOMEM\n")); 2109 return (ENOMEM); 2110 } 2111 2112 pmap_enter_pv(pg, pve, pm, va, nflags); 2113 } 2114 } else { 2115 /* 2116 * We're mapping an unmanaged page. 2117 * These are always readable, and possibly writable, from 2118 * the get go as we don't need to track ref/mod status. 2119 */ 2120 npte |= L2_S_PROTO; 2121 if (prot & VM_PROT_WRITE) 2122 npte |= L2_S_PROT_W; 2123 2124 /* 2125 * Make sure the vector table is mapped cacheable 2126 */ 2127 if (pm != pmap_kernel() && va == vector_page) 2128 npte |= pte_l2_s_cache_mode; 2129 2130 if (opg) { 2131 /* 2132 * Looks like there's an existing 'managed' mapping 2133 * at this address. 2134 */ 2135 simple_lock(&opg->mdpage.pvh_slock); 2136 pve = pmap_remove_pv(opg, pm, va, 0); 2137 pmap_vac_me_harder(opg, pm, 0); 2138 simple_unlock(&opg->mdpage.pvh_slock); 2139 oflags = pve->pv_flags; 2140 2141 if ((oflags & PVF_NC) == 0 && l2pte_valid(opte)) { 2142 if (PV_BEEN_EXECD(oflags)) 2143 pmap_idcache_wbinv_range(pm, va, 2144 PAGE_SIZE); 2145 else 2146 if (PV_BEEN_REFD(oflags)) 2147 pmap_dcache_wb_range(pm, va, PAGE_SIZE, 2148 TRUE, (oflags & PVF_WRITE) == 0); 2149 } 2150 pool_put(&pmap_pv_pool, pve); 2151 } 2152 } 2153 2154 /* 2155 * Make sure userland mappings get the right permissions 2156 */ 2157 if (pm != pmap_kernel() && va != vector_page) 2158 npte |= L2_S_PROT_U; 2159 2160 /* 2161 * Keep the stats up to date 2162 */ 2163 if (opte == 0) { 2164 l2b->l2b_occupancy++; 2165 pm->pm_stats.resident_count++; 2166 } 2167 2168 NPDEBUG(PDB_ENTER, 2169 printf("pmap_enter: opte 0x%08x npte 0x%08x\n", opte, npte)); 2170 2171 /* 2172 * If this is just a wiring change, the two PTEs will be 2173 * identical, so there's no need to update the page table. 2174 */ 2175 if (npte != opte) { 2176 boolean_t is_cached = pmap_is_cached(pm); 2177 2178 *ptep = npte; 2179 if (is_cached) { 2180 /* 2181 * We only need to frob the cache/tlb if this pmap 2182 * is current 2183 */ 2184 PTE_SYNC(ptep); 2185 if (va != vector_page && l2pte_valid(npte)) { 2186 /* 2187 * This mapping is likely to be accessed as 2188 * soon as we return to userland. Fix up the 2189 * L1 entry to avoid taking another 2190 * page/domain fault. 2191 */ 2192 pd_entry_t *pl1pd, l1pd; 2193 2194 pl1pd = &pm->pm_l1->l1_kva[L1_IDX(va)]; 2195 l1pd = l2b->l2b_phys | L1_C_DOM(pm->pm_domain) | 2196 L1_C_PROTO; 2197 if (*pl1pd != l1pd) { 2198 *pl1pd = l1pd; 2199 PTE_SYNC(pl1pd); 2200 } 2201 } 2202 } 2203 2204 if (PV_BEEN_EXECD(oflags)) 2205 pmap_tlb_flushID_SE(pm, va); 2206 else 2207 if (PV_BEEN_REFD(oflags)) 2208 pmap_tlb_flushD_SE(pm, va); 2209 2210 NPDEBUG(PDB_ENTER, 2211 printf("pmap_enter: is_cached %d cs 0x%08x\n", 2212 is_cached, pm->pm_cstate.cs_all)); 2213 2214 if (pg != NULL) { 2215 simple_lock(&pg->mdpage.pvh_slock); 2216 pmap_vac_me_harder(pg, pm, va); 2217 simple_unlock(&pg->mdpage.pvh_slock); 2218 } 2219 } 2220 2221 pmap_release_pmap_lock(pm); 2222 PMAP_MAP_TO_HEAD_UNLOCK(); 2223 2224 return (0); 2225 } 2226 2227 /* 2228 * pmap_remove() 2229 * 2230 * pmap_remove is responsible for nuking a number of mappings for a range 2231 * of virtual address space in the current pmap. To do this efficiently 2232 * is interesting, because in a number of cases a wide virtual address 2233 * range may be supplied that contains few actual mappings. So, the 2234 * optimisations are: 2235 * 1. Skip over hunks of address space for which no L1 or L2 entry exists. 2236 * 2. Build up a list of pages we've hit, up to a maximum, so we can 2237 * maybe do just a partial cache clean. This path of execution is 2238 * complicated by the fact that the cache must be flushed _before_ 2239 * the PTE is nuked, being a VAC :-) 2240 * 3. If we're called after UVM calls pmap_remove_all(), we can defer 2241 * all invalidations until pmap_update(), since pmap_remove_all() has 2242 * already flushed the cache. 2243 * 4. Maybe later fast-case a single page, but I don't think this is 2244 * going to make _that_ much difference overall. 2245 */ 2246 2247 #define PMAP_REMOVE_CLEAN_LIST_SIZE 3 2248 2249 void 2250 pmap_do_remove(pmap_t pm, vaddr_t sva, vaddr_t eva, int skip_wired) 2251 { 2252 struct l2_bucket *l2b; 2253 vaddr_t next_bucket; 2254 pt_entry_t *ptep; 2255 u_int cleanlist_idx, total, cnt; 2256 struct { 2257 vaddr_t va; 2258 pt_entry_t *pte; 2259 } cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE]; 2260 u_int mappings, is_exec, is_refd; 2261 2262 NPDEBUG(PDB_REMOVE, printf("pmap_do_remove: pmap=%p sva=%08lx " 2263 "eva=%08lx\n", pm, sva, eva)); 2264 2265 /* 2266 * we lock in the pmap => pv_head direction 2267 */ 2268 PMAP_MAP_TO_HEAD_LOCK(); 2269 pmap_acquire_pmap_lock(pm); 2270 2271 if (pm->pm_remove_all || !pmap_is_cached(pm)) { 2272 cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1; 2273 if (pm->pm_cstate.cs_tlb == 0) 2274 pm->pm_remove_all = TRUE; 2275 } else 2276 cleanlist_idx = 0; 2277 2278 total = 0; 2279 2280 while (sva < eva) { 2281 /* 2282 * Do one L2 bucket's worth at a time. 2283 */ 2284 next_bucket = L2_NEXT_BUCKET(sva); 2285 if (next_bucket > eva) 2286 next_bucket = eva; 2287 2288 l2b = pmap_get_l2_bucket(pm, sva); 2289 if (l2b == NULL) { 2290 sva = next_bucket; 2291 continue; 2292 } 2293 2294 ptep = &l2b->l2b_kva[l2pte_index(sva)]; 2295 2296 for (mappings = 0; sva < next_bucket; sva += PAGE_SIZE, ptep++){ 2297 struct vm_page *pg; 2298 pt_entry_t pte; 2299 paddr_t pa; 2300 2301 pte = *ptep; 2302 2303 if (pte == 0) { 2304 /* Nothing here, move along */ 2305 continue; 2306 } 2307 2308 pa = l2pte_pa(pte); 2309 is_exec = 0; 2310 is_refd = 1; 2311 2312 /* 2313 * Update flags. In a number of circumstances, 2314 * we could cluster a lot of these and do a 2315 * number of sequential pages in one go. 2316 */ 2317 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) { 2318 struct pv_entry *pve; 2319 simple_lock(&pg->mdpage.pvh_slock); 2320 pve = pmap_remove_pv(pg, pm, sva, skip_wired); 2321 pmap_vac_me_harder(pg, pm, 0); 2322 simple_unlock(&pg->mdpage.pvh_slock); 2323 if (pve != NULL) { 2324 if (pm->pm_remove_all == FALSE) { 2325 is_exec = 2326 PV_BEEN_EXECD(pve->pv_flags); 2327 is_refd = 2328 PV_BEEN_REFD(pve->pv_flags); 2329 } 2330 pool_put(&pmap_pv_pool, pve); 2331 } else 2332 if (skip_wired) { 2333 /* The mapping is wired. Skip it */ 2334 continue; 2335 } 2336 } else 2337 if (skip_wired) { 2338 /* Unmanaged pages are always wired. */ 2339 continue; 2340 } 2341 2342 mappings++; 2343 2344 if (!l2pte_valid(pte)) { 2345 /* 2346 * Ref/Mod emulation is still active for this 2347 * mapping, therefore it is has not yet been 2348 * accessed. No need to frob the cache/tlb. 2349 */ 2350 *ptep = 0; 2351 PTE_SYNC_CURRENT(pm, ptep); 2352 continue; 2353 } 2354 2355 if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) { 2356 /* Add to the clean list. */ 2357 cleanlist[cleanlist_idx].pte = ptep; 2358 cleanlist[cleanlist_idx].va = 2359 sva | (is_exec & 1); 2360 cleanlist_idx++; 2361 } else 2362 if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) { 2363 /* Nuke everything if needed. */ 2364 pmap_idcache_wbinv_all(pm); 2365 pmap_tlb_flushID(pm); 2366 2367 /* 2368 * Roll back the previous PTE list, 2369 * and zero out the current PTE. 2370 */ 2371 for (cnt = 0; 2372 cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) { 2373 *cleanlist[cnt].pte = 0; 2374 } 2375 *ptep = 0; 2376 PTE_SYNC(ptep); 2377 cleanlist_idx++; 2378 pm->pm_remove_all = TRUE; 2379 } else { 2380 *ptep = 0; 2381 PTE_SYNC(ptep); 2382 if (pm->pm_remove_all == FALSE) { 2383 if (is_exec) 2384 pmap_tlb_flushID_SE(pm, sva); 2385 else 2386 if (is_refd) 2387 pmap_tlb_flushD_SE(pm, sva); 2388 } 2389 } 2390 } 2391 2392 /* 2393 * Deal with any left overs 2394 */ 2395 if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) { 2396 total += cleanlist_idx; 2397 for (cnt = 0; cnt < cleanlist_idx; cnt++) { 2398 if (pm->pm_cstate.cs_all != 0) { 2399 vaddr_t clva = cleanlist[cnt].va & ~1; 2400 if (cleanlist[cnt].va & 1) { 2401 pmap_idcache_wbinv_range(pm, 2402 clva, PAGE_SIZE); 2403 pmap_tlb_flushID_SE(pm, clva); 2404 } else { 2405 pmap_dcache_wb_range(pm, 2406 clva, PAGE_SIZE, TRUE, 2407 FALSE); 2408 pmap_tlb_flushD_SE(pm, clva); 2409 } 2410 } 2411 *cleanlist[cnt].pte = 0; 2412 PTE_SYNC_CURRENT(pm, cleanlist[cnt].pte); 2413 } 2414 2415 /* 2416 * If it looks like we're removing a whole bunch 2417 * of mappings, it's faster to just write-back 2418 * the whole cache now and defer TLB flushes until 2419 * pmap_update() is called. 2420 */ 2421 if (total <= PMAP_REMOVE_CLEAN_LIST_SIZE) 2422 cleanlist_idx = 0; 2423 else { 2424 cleanlist_idx = PMAP_REMOVE_CLEAN_LIST_SIZE + 1; 2425 pmap_idcache_wbinv_all(pm); 2426 pm->pm_remove_all = TRUE; 2427 } 2428 } 2429 2430 pmap_free_l2_bucket(pm, l2b, mappings); 2431 pm->pm_stats.resident_count -= mappings; 2432 } 2433 2434 pmap_release_pmap_lock(pm); 2435 PMAP_MAP_TO_HEAD_UNLOCK(); 2436 } 2437 2438 /* 2439 * pmap_kenter_pa: enter an unmanaged, wired kernel mapping 2440 * 2441 * We assume there is already sufficient KVM space available 2442 * to do this, as we can't allocate L2 descriptor tables/metadata 2443 * from here. 2444 */ 2445 void 2446 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot) 2447 { 2448 struct l2_bucket *l2b; 2449 pt_entry_t *ptep, opte; 2450 2451 NPDEBUG(PDB_KENTER, 2452 printf("pmap_kenter_pa: va 0x%08lx, pa 0x%08lx, prot 0x%x\n", 2453 va, pa, prot)); 2454 2455 l2b = pmap_get_l2_bucket(pmap_kernel(), va); 2456 KDASSERT(l2b != NULL); 2457 2458 ptep = &l2b->l2b_kva[l2pte_index(va)]; 2459 opte = *ptep; 2460 2461 if (l2pte_valid(opte)) { 2462 cpu_dcache_wbinv_range(va, PAGE_SIZE); 2463 cpu_tlb_flushD_SE(va); 2464 cpu_cpwait(); 2465 } else 2466 if (opte == 0) 2467 l2b->l2b_occupancy++; 2468 2469 *ptep = L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | 2470 pte_l2_s_cache_mode; 2471 PTE_SYNC(ptep); 2472 } 2473 2474 void 2475 pmap_kremove(vaddr_t va, vsize_t len) 2476 { 2477 struct l2_bucket *l2b; 2478 pt_entry_t *ptep, *sptep, opte; 2479 vaddr_t next_bucket, eva; 2480 u_int mappings; 2481 2482 NPDEBUG(PDB_KREMOVE, printf("pmap_kremove: va 0x%08lx, len 0x%08lx\n", 2483 va, len)); 2484 2485 eva = va + len; 2486 2487 while (va < eva) { 2488 next_bucket = L2_NEXT_BUCKET(va); 2489 if (next_bucket > eva) 2490 next_bucket = eva; 2491 2492 l2b = pmap_get_l2_bucket(pmap_kernel(), va); 2493 KDASSERT(l2b != NULL); 2494 2495 sptep = ptep = &l2b->l2b_kva[l2pte_index(va)]; 2496 mappings = 0; 2497 2498 while (va < next_bucket) { 2499 opte = *ptep; 2500 if (l2pte_valid(opte)) { 2501 cpu_dcache_wbinv_range(va, PAGE_SIZE); 2502 cpu_tlb_flushD_SE(va); 2503 } 2504 if (opte) { 2505 *ptep = 0; 2506 mappings++; 2507 } 2508 va += PAGE_SIZE; 2509 ptep++; 2510 } 2511 KDASSERT(mappings <= l2b->l2b_occupancy); 2512 l2b->l2b_occupancy -= mappings; 2513 PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep)); 2514 } 2515 cpu_cpwait(); 2516 } 2517 2518 boolean_t 2519 pmap_extract(pmap_t pm, vaddr_t va, paddr_t *pap) 2520 { 2521 struct l2_dtable *l2; 2522 pd_entry_t *pl1pd, l1pd; 2523 pt_entry_t *ptep, pte; 2524 paddr_t pa; 2525 u_int l1idx; 2526 2527 pmap_acquire_pmap_lock(pm); 2528 2529 l1idx = L1_IDX(va); 2530 pl1pd = &pm->pm_l1->l1_kva[l1idx]; 2531 l1pd = *pl1pd; 2532 2533 if (l1pte_section_p(l1pd)) { 2534 /* 2535 * These should only happen for pmap_kernel() 2536 */ 2537 KDASSERT(pm == pmap_kernel()); 2538 pmap_release_pmap_lock(pm); 2539 pa = (l1pd & L1_S_FRAME) | (va & L1_S_OFFSET); 2540 } else { 2541 /* 2542 * Note that we can't rely on the validity of the L1 2543 * descriptor as an indication that a mapping exists. 2544 * We have to look it up in the L2 dtable. 2545 */ 2546 l2 = pm->pm_l2[L2_IDX(l1idx)]; 2547 2548 if (l2 == NULL || 2549 (ptep = l2->l2_bucket[L2_BUCKET(l1idx)].l2b_kva) == NULL) { 2550 pmap_release_pmap_lock(pm); 2551 return (FALSE); 2552 } 2553 2554 ptep = &ptep[l2pte_index(va)]; 2555 pte = *ptep; 2556 pmap_release_pmap_lock(pm); 2557 2558 if (pte == 0) 2559 return (FALSE); 2560 2561 switch (pte & L2_TYPE_MASK) { 2562 case L2_TYPE_L: 2563 pa = (pte & L2_L_FRAME) | (va & L2_L_OFFSET); 2564 break; 2565 2566 default: 2567 pa = (pte & L2_S_FRAME) | (va & L2_S_OFFSET); 2568 break; 2569 } 2570 } 2571 2572 if (pap != NULL) 2573 *pap = pa; 2574 2575 return (TRUE); 2576 } 2577 2578 void 2579 pmap_protect(pmap_t pm, vaddr_t sva, vaddr_t eva, vm_prot_t prot) 2580 { 2581 struct l2_bucket *l2b; 2582 pt_entry_t *ptep, pte; 2583 vaddr_t next_bucket; 2584 u_int flags; 2585 int flush; 2586 2587 NPDEBUG(PDB_PROTECT, 2588 printf("pmap_protect: pm %p sva 0x%lx eva 0x%lx prot 0x%x\n", 2589 pm, sva, eva, prot)); 2590 2591 if ((prot & VM_PROT_READ) == 0) { 2592 pmap_remove(pm, sva, eva); 2593 return; 2594 } 2595 2596 if (prot & VM_PROT_WRITE) { 2597 /* 2598 * If this is a read->write transition, just ignore it and let 2599 * uvm_fault() take care of it later. 2600 */ 2601 return; 2602 } 2603 2604 PMAP_MAP_TO_HEAD_LOCK(); 2605 pmap_acquire_pmap_lock(pm); 2606 2607 /* 2608 * OK, at this point, we know we're doing write-protect operation. 2609 * If the pmap is active, write-back the range. 2610 */ 2611 pmap_dcache_wb_range(pm, sva, eva - sva, FALSE, FALSE); 2612 2613 flush = ((eva - sva) >= (PAGE_SIZE * 4)) ? 0 : -1; 2614 flags = 0; 2615 2616 while (sva < eva) { 2617 next_bucket = L2_NEXT_BUCKET(sva); 2618 if (next_bucket > eva) 2619 next_bucket = eva; 2620 2621 l2b = pmap_get_l2_bucket(pm, sva); 2622 if (l2b == NULL) { 2623 sva = next_bucket; 2624 continue; 2625 } 2626 2627 ptep = &l2b->l2b_kva[l2pte_index(sva)]; 2628 2629 while (sva < next_bucket) { 2630 if ((pte = *ptep) != 0 && (pte & L2_S_PROT_W) != 0) { 2631 struct vm_page *pg; 2632 u_int f; 2633 2634 pg = PHYS_TO_VM_PAGE(l2pte_pa(pte)); 2635 pte &= ~L2_S_PROT_W; 2636 *ptep = pte; 2637 PTE_SYNC(ptep); 2638 2639 if (pg != NULL) { 2640 simple_lock(&pg->mdpage.pvh_slock); 2641 f = pmap_modify_pv(pg, pm, sva, 2642 PVF_WRITE, 0); 2643 pmap_vac_me_harder(pg, pm, sva); 2644 simple_unlock(&pg->mdpage.pvh_slock); 2645 } else 2646 f = PVF_REF | PVF_EXEC; 2647 2648 if (flush >= 0) { 2649 flush++; 2650 flags |= f; 2651 } else 2652 if (PV_BEEN_EXECD(f)) 2653 pmap_tlb_flushID_SE(pm, sva); 2654 else 2655 if (PV_BEEN_REFD(f)) 2656 pmap_tlb_flushD_SE(pm, sva); 2657 } 2658 2659 sva += PAGE_SIZE; 2660 ptep++; 2661 } 2662 } 2663 2664 pmap_release_pmap_lock(pm); 2665 PMAP_MAP_TO_HEAD_UNLOCK(); 2666 2667 if (flush) { 2668 if (PV_BEEN_EXECD(flags)) 2669 pmap_tlb_flushID(pm); 2670 else 2671 if (PV_BEEN_REFD(flags)) 2672 pmap_tlb_flushD(pm); 2673 } 2674 } 2675 2676 void 2677 pmap_page_protect(struct vm_page *pg, vm_prot_t prot) 2678 { 2679 2680 NPDEBUG(PDB_PROTECT, 2681 printf("pmap_page_protect: pg %p (0x%08lx), prot 0x%x\n", 2682 pg, VM_PAGE_TO_PHYS(pg), prot)); 2683 2684 switch(prot) { 2685 case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE: 2686 case VM_PROT_READ|VM_PROT_WRITE: 2687 return; 2688 2689 case VM_PROT_READ: 2690 case VM_PROT_READ|VM_PROT_EXECUTE: 2691 pmap_clearbit(pg, PVF_WRITE); 2692 break; 2693 2694 default: 2695 pmap_page_remove(pg); 2696 break; 2697 } 2698 } 2699 2700 /* 2701 * pmap_clear_modify: 2702 * 2703 * Clear the "modified" attribute for a page. 2704 */ 2705 boolean_t 2706 pmap_clear_modify(struct vm_page *pg) 2707 { 2708 boolean_t rv; 2709 2710 if (pg->mdpage.pvh_attrs & PVF_MOD) { 2711 rv = TRUE; 2712 pmap_clearbit(pg, PVF_MOD); 2713 } else 2714 rv = FALSE; 2715 2716 return (rv); 2717 } 2718 2719 /* 2720 * pmap_clear_reference: 2721 * 2722 * Clear the "referenced" attribute for a page. 2723 */ 2724 boolean_t 2725 pmap_clear_reference(struct vm_page *pg) 2726 { 2727 boolean_t rv; 2728 2729 if (pg->mdpage.pvh_attrs & PVF_REF) { 2730 rv = TRUE; 2731 pmap_clearbit(pg, PVF_REF); 2732 } else 2733 rv = FALSE; 2734 2735 return (rv); 2736 } 2737 2738 /* 2739 * pmap_is_modified: 2740 * 2741 * Test if a page has the "modified" attribute. 2742 */ 2743 /* See <arm/arm32/pmap.h> */ 2744 2745 /* 2746 * pmap_is_referenced: 2747 * 2748 * Test if a page has the "referenced" attribute. 2749 */ 2750 /* See <arm/arm32/pmap.h> */ 2751 2752 int 2753 pmap_fault_fixup(pmap_t pm, vaddr_t va, vm_prot_t ftype, int user) 2754 { 2755 struct l2_dtable *l2; 2756 struct l2_bucket *l2b; 2757 pd_entry_t *pl1pd, l1pd; 2758 pt_entry_t *ptep, pte; 2759 paddr_t pa; 2760 u_int l1idx; 2761 int rv = 0; 2762 2763 PMAP_MAP_TO_HEAD_LOCK(); 2764 pmap_acquire_pmap_lock(pm); 2765 2766 l1idx = L1_IDX(va); 2767 2768 /* 2769 * If there is no l2_dtable for this address, then the process 2770 * has no business accessing it. 2771 * 2772 * Note: This will catch userland processes trying to access 2773 * kernel addresses. 2774 */ 2775 l2 = pm->pm_l2[L2_IDX(l1idx)]; 2776 if (l2 == NULL) 2777 goto out; 2778 2779 /* 2780 * Likewise if there is no L2 descriptor table 2781 */ 2782 l2b = &l2->l2_bucket[L2_BUCKET(l1idx)]; 2783 if (l2b->l2b_kva == NULL) 2784 goto out; 2785 2786 /* 2787 * Check the PTE itself. 2788 */ 2789 ptep = &l2b->l2b_kva[l2pte_index(va)]; 2790 pte = *ptep; 2791 if (pte == 0) 2792 goto out; 2793 2794 /* 2795 * Catch a userland access to the vector page mapped at 0x0 2796 */ 2797 if (user && (pte & L2_S_PROT_U) == 0) 2798 goto out; 2799 2800 pa = l2pte_pa(pte); 2801 2802 if ((ftype & VM_PROT_WRITE) && (pte & L2_S_PROT_W) == 0) { 2803 /* 2804 * This looks like a good candidate for "page modified" 2805 * emulation... 2806 */ 2807 struct pv_entry *pv; 2808 struct vm_page *pg; 2809 2810 /* Extract the physical address of the page */ 2811 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) 2812 goto out; 2813 2814 /* Get the current flags for this page. */ 2815 simple_lock(&pg->mdpage.pvh_slock); 2816 2817 pv = pmap_find_pv(pg, pm, va); 2818 if (pv == NULL) { 2819 simple_unlock(&pg->mdpage.pvh_slock); 2820 goto out; 2821 } 2822 2823 /* 2824 * Do the flags say this page is writable? If not then it 2825 * is a genuine write fault. If yes then the write fault is 2826 * our fault as we did not reflect the write access in the 2827 * PTE. Now we know a write has occurred we can correct this 2828 * and also set the modified bit 2829 */ 2830 if ((pv->pv_flags & PVF_WRITE) == 0) { 2831 simple_unlock(&pg->mdpage.pvh_slock); 2832 goto out; 2833 } 2834 2835 NPDEBUG(PDB_FOLLOW, 2836 printf("pmap_fault_fixup: mod emul. pm %p, va 0x%08lx, pa 0x%08lx\n", 2837 pm, va, VM_PAGE_TO_PHYS(pg))); 2838 2839 pg->mdpage.pvh_attrs |= PVF_REF | PVF_MOD; 2840 pv->pv_flags |= PVF_REF | PVF_MOD; 2841 simple_unlock(&pg->mdpage.pvh_slock); 2842 2843 /* 2844 * Re-enable write permissions for the page. No need to call 2845 * pmap_vac_me_harder(), since this is just a 2846 * modified-emulation fault, and the PVF_WRITE bit isn't 2847 * changing. We've already set the cacheable bits based on 2848 * the assumption that we can write to this page. 2849 */ 2850 *ptep = (pte & ~L2_TYPE_MASK) | L2_S_PROTO | L2_S_PROT_W; 2851 PTE_SYNC(ptep); 2852 rv = 1; 2853 } else 2854 if ((pte & L2_TYPE_MASK) == L2_TYPE_INV) { 2855 /* 2856 * This looks like a good candidate for "page referenced" 2857 * emulation. 2858 */ 2859 struct pv_entry *pv; 2860 struct vm_page *pg; 2861 2862 /* Extract the physical address of the page */ 2863 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) 2864 goto out; 2865 2866 /* Get the current flags for this page. */ 2867 simple_lock(&pg->mdpage.pvh_slock); 2868 2869 pv = pmap_find_pv(pg, pm, va); 2870 if (pv == NULL) { 2871 simple_unlock(&pg->mdpage.pvh_slock); 2872 goto out; 2873 } 2874 2875 pg->mdpage.pvh_attrs |= PVF_REF; 2876 pv->pv_flags |= PVF_REF; 2877 simple_unlock(&pg->mdpage.pvh_slock); 2878 2879 NPDEBUG(PDB_FOLLOW, 2880 printf("pmap_fault_fixup: ref emul. pm %p, va 0x%08lx, pa 0x%08lx\n", 2881 pm, va, VM_PAGE_TO_PHYS(pg))); 2882 2883 *ptep = (pte & ~L2_TYPE_MASK) | L2_S_PROTO; 2884 PTE_SYNC(ptep); 2885 rv = 1; 2886 } 2887 2888 /* 2889 * We know there is a valid mapping here, so simply 2890 * fix up the L1 if necessary. 2891 */ 2892 pl1pd = &pm->pm_l1->l1_kva[l1idx]; 2893 l1pd = l2b->l2b_phys | L1_C_DOM(pm->pm_domain) | L1_C_PROTO; 2894 if (*pl1pd != l1pd) { 2895 *pl1pd = l1pd; 2896 PTE_SYNC(pl1pd); 2897 rv = 1; 2898 } 2899 2900 #ifdef CPU_SA110 2901 /* 2902 * There are bugs in the rev K SA110. This is a check for one 2903 * of them. 2904 */ 2905 if (rv == 0 && curcpu()->ci_arm_cputype == CPU_ID_SA110 && 2906 curcpu()->ci_arm_cpurev < 3) { 2907 /* Always current pmap */ 2908 if (l2pte_valid(pte)) { 2909 extern int kernel_debug; 2910 if (kernel_debug & 1) { 2911 struct proc *p = curlwp->l_proc; 2912 printf("prefetch_abort: page is already " 2913 "mapped - pte=%p *pte=%08x\n", ptep, pte); 2914 printf("prefetch_abort: pc=%08lx proc=%p " 2915 "process=%s\n", va, p, p->p_comm); 2916 printf("prefetch_abort: far=%08x fs=%x\n", 2917 cpu_faultaddress(), cpu_faultstatus()); 2918 } 2919 #ifdef DDB 2920 if (kernel_debug & 2) 2921 Debugger(); 2922 #endif 2923 rv = 1; 2924 } 2925 } 2926 #endif /* CPU_SA110 */ 2927 2928 #ifdef DEBUG 2929 /* 2930 * If 'rv == 0' at this point, it generally indicates that there is a 2931 * stale TLB entry for the faulting address. This happens when two or 2932 * more processes are sharing an L1. Since we don't flush the TLB on 2933 * a context switch between such processes, we can take domain faults 2934 * for mappings which exist at the same VA in both processes. EVEN IF 2935 * WE'VE RECENTLY FIXED UP THE CORRESPONDING L1 in pmap_enter(), for 2936 * example. 2937 * 2938 * This is extremely likely to happen if pmap_enter() updated the L1 2939 * entry for a recently entered mapping. In this case, the TLB is 2940 * flushed for the new mapping, but there may still be TLB entries for 2941 * other mappings belonging to other processes in the 1MB range 2942 * covered by the L1 entry. 2943 * 2944 * Since 'rv == 0', we know that the L1 already contains the correct 2945 * value, so the fault must be due to a stale TLB entry. 2946 * 2947 * Since we always need to flush the TLB anyway in the case where we 2948 * fixed up the L1, or frobbed the L2 PTE, we effectively deal with 2949 * stale TLB entries dynamically. 2950 * 2951 * However, the above condition can ONLY happen if the current L1 is 2952 * being shared. If it happens when the L1 is unshared, it indicates 2953 * that other parts of the pmap are not doing their job WRT managing 2954 * the TLB. 2955 */ 2956 if (rv == 0 && pm->pm_l1->l1_domain_use_count == 1) { 2957 extern int last_fault_code; 2958 printf("fixup: pm %p, va 0x%lx, ftype %d - nothing to do!\n", 2959 pm, va, ftype); 2960 printf("fixup: l2 %p, l2b %p, ptep %p, pl1pd %p\n", 2961 l2, l2b, ptep, pl1pd); 2962 printf("fixup: pte 0x%x, l1pd 0x%x, last code 0x%x\n", 2963 pte, l1pd, last_fault_code); 2964 #ifdef DDB 2965 Debugger(); 2966 #endif 2967 } 2968 #endif 2969 2970 cpu_tlb_flushID_SE(va); 2971 cpu_cpwait(); 2972 2973 rv = 1; 2974 2975 out: 2976 pmap_release_pmap_lock(pm); 2977 PMAP_MAP_TO_HEAD_UNLOCK(); 2978 2979 return (rv); 2980 } 2981 2982 /* 2983 * pmap_collect: free resources held by a pmap 2984 * 2985 * => optional function. 2986 * => called when a process is swapped out to free memory. 2987 */ 2988 void 2989 pmap_collect(pmap_t pm) 2990 { 2991 2992 pmap_idcache_wbinv_all(pm); 2993 pm->pm_remove_all = TRUE; 2994 pmap_do_remove(pm, VM_MIN_ADDRESS, VM_MAX_ADDRESS, 1); 2995 pmap_update(pm); 2996 } 2997 2998 /* 2999 * Routine: pmap_procwr 3000 * 3001 * Function: 3002 * Synchronize caches corresponding to [addr, addr+len) in p. 3003 * 3004 */ 3005 void 3006 pmap_procwr(struct proc *p, vaddr_t va, int len) 3007 { 3008 /* We only need to do anything if it is the current process. */ 3009 if (p == curproc) 3010 cpu_icache_sync_range(va, len); 3011 } 3012 3013 /* 3014 * Routine: pmap_unwire 3015 * Function: Clear the wired attribute for a map/virtual-address pair. 3016 * 3017 * In/out conditions: 3018 * The mapping must already exist in the pmap. 3019 */ 3020 void 3021 pmap_unwire(pmap_t pm, vaddr_t va) 3022 { 3023 struct l2_bucket *l2b; 3024 pt_entry_t *ptep, pte; 3025 struct vm_page *pg; 3026 paddr_t pa; 3027 3028 NPDEBUG(PDB_WIRING, printf("pmap_unwire: pm %p, va 0x%08lx\n", pm, va)); 3029 3030 PMAP_MAP_TO_HEAD_LOCK(); 3031 pmap_acquire_pmap_lock(pm); 3032 3033 l2b = pmap_get_l2_bucket(pm, va); 3034 KDASSERT(l2b != NULL); 3035 3036 ptep = &l2b->l2b_kva[l2pte_index(va)]; 3037 pte = *ptep; 3038 3039 /* Extract the physical address of the page */ 3040 pa = l2pte_pa(pte); 3041 3042 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) { 3043 /* Update the wired bit in the pv entry for this page. */ 3044 simple_lock(&pg->mdpage.pvh_slock); 3045 (void) pmap_modify_pv(pg, pm, va, PVF_WIRED, 0); 3046 simple_unlock(&pg->mdpage.pvh_slock); 3047 } 3048 3049 pmap_release_pmap_lock(pm); 3050 PMAP_MAP_TO_HEAD_UNLOCK(); 3051 } 3052 3053 void 3054 pmap_activate(struct lwp *l) 3055 { 3056 pmap_t pm; 3057 struct pcb *pcb; 3058 int s; 3059 3060 pm = l->l_proc->p_vmspace->vm_map.pmap; 3061 pcb = &l->l_addr->u_pcb; 3062 3063 pmap_set_pcb_pagedir(pm, pcb); 3064 3065 if (l == curlwp) { 3066 u_int cur_dacr, cur_ttb; 3067 3068 __asm __volatile("mrc p15, 0, %0, c2, c0, 0" : "=r"(cur_ttb)); 3069 __asm __volatile("mrc p15, 0, %0, c3, c0, 0" : "=r"(cur_dacr)); 3070 3071 cur_ttb &= ~(L1_TABLE_SIZE - 1); 3072 3073 if (cur_ttb == (u_int)pcb->pcb_pagedir && 3074 cur_dacr == pcb->pcb_dacr) { 3075 /* 3076 * No need to switch address spaces. 3077 */ 3078 return; 3079 } 3080 3081 s = splhigh(); 3082 pmap_acquire_pmap_lock(pm); 3083 disable_interrupts(I32_bit | F32_bit); 3084 3085 /* 3086 * We MUST, I repeat, MUST fix up the L1 entry corresponding 3087 * to 'vector_page' in the incoming L1 table before switching 3088 * to it otherwise subsequent interrupts/exceptions (including 3089 * domain faults!) will jump into hyperspace. 3090 */ 3091 if (pcb->pcb_pl1vec) { 3092 *pcb->pcb_pl1vec = pcb->pcb_l1vec; 3093 /* 3094 * Don't need to PTE_SYNC() at this point since 3095 * cpu_setttb() is about to flush both the cache 3096 * and the TLB. 3097 */ 3098 } 3099 3100 cpu_domains(pcb->pcb_dacr); 3101 cpu_setttb(pcb->pcb_pagedir); 3102 3103 enable_interrupts(I32_bit | F32_bit); 3104 3105 /* 3106 * Flag any previous userland pmap as being NOT 3107 * resident in the cache/tlb. 3108 */ 3109 if (pmap_cache_state && pmap_cache_state != &pm->pm_cstate) 3110 pmap_cache_state->cs_all = 0; 3111 3112 /* 3113 * The new pmap, however, IS resident. 3114 */ 3115 pmap_cache_state = &pm->pm_cstate; 3116 pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL; 3117 pmap_release_pmap_lock(pm); 3118 splx(s); 3119 } 3120 } 3121 3122 void 3123 pmap_deactivate(struct lwp *l) 3124 { 3125 } 3126 3127 void 3128 pmap_update(pmap_t pm) 3129 { 3130 3131 if (pm->pm_remove_all) { 3132 /* 3133 * Finish up the pmap_remove_all() optimisation by flushing 3134 * the TLB. 3135 */ 3136 pmap_tlb_flushID(pm); 3137 pm->pm_remove_all = FALSE; 3138 } 3139 3140 if (pmap_is_current(pm)) { 3141 /* 3142 * If we're dealing with a current userland pmap, move its L1 3143 * to the end of the LRU. 3144 */ 3145 if (pm != pmap_kernel()) 3146 pmap_use_l1(pm); 3147 3148 /* 3149 * We can assume we're done with frobbing the cache/tlb for 3150 * now. Make sure any future pmap ops don't skip cache/tlb 3151 * flushes. 3152 */ 3153 pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL; 3154 } 3155 3156 /* 3157 * make sure TLB/cache operations have completed. 3158 */ 3159 cpu_cpwait(); 3160 } 3161 3162 void 3163 pmap_remove_all(pmap_t pm) 3164 { 3165 3166 /* 3167 * The vmspace described by this pmap is about to be torn down. 3168 * Until pmap_update() is called, UVM will only make calls 3169 * to pmap_remove(). We can make life much simpler by flushing 3170 * the cache now, and deferring TLB invalidation to pmap_update(). 3171 */ 3172 pmap_idcache_wbinv_all(pm); 3173 pm->pm_remove_all = TRUE; 3174 } 3175 3176 /* 3177 * Retire the given physical map from service. 3178 * Should only be called if the map contains no valid mappings. 3179 */ 3180 void 3181 pmap_destroy(pmap_t pm) 3182 { 3183 u_int count; 3184 3185 if (pm == NULL) 3186 return; 3187 3188 if (pm->pm_remove_all) { 3189 pmap_tlb_flushID(pm); 3190 pm->pm_remove_all = FALSE; 3191 } 3192 3193 /* 3194 * Drop reference count 3195 */ 3196 simple_lock(&pm->pm_lock); 3197 count = --pm->pm_obj.uo_refs; 3198 simple_unlock(&pm->pm_lock); 3199 if (count > 0) { 3200 if (pmap_is_current(pm)) { 3201 if (pm != pmap_kernel()) 3202 pmap_use_l1(pm); 3203 pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL; 3204 } 3205 return; 3206 } 3207 3208 /* 3209 * reference count is zero, free pmap resources and then free pmap. 3210 */ 3211 3212 if (vector_page < KERNEL_BASE) { 3213 struct pcb *pcb = &lwp0.l_addr->u_pcb; 3214 3215 if (pmap_is_current(pm)) { 3216 /* 3217 * Frob the L1 entry corresponding to the vector 3218 * page so that it contains the kernel pmap's domain 3219 * number. This will ensure pmap_remove() does not 3220 * pull the current vector page out from under us. 3221 */ 3222 disable_interrupts(I32_bit | F32_bit); 3223 *pcb->pcb_pl1vec = pcb->pcb_l1vec; 3224 cpu_domains(pcb->pcb_dacr); 3225 cpu_setttb(pcb->pcb_pagedir); 3226 enable_interrupts(I32_bit | F32_bit); 3227 } 3228 3229 /* Remove the vector page mapping */ 3230 pmap_remove(pm, vector_page, vector_page + PAGE_SIZE); 3231 pmap_update(pm); 3232 3233 /* 3234 * Make sure cpu_switch(), et al, DTRT. This is safe to do 3235 * since this process has no remaining mappings of its own. 3236 */ 3237 curpcb->pcb_pl1vec = pcb->pcb_pl1vec; 3238 curpcb->pcb_l1vec = pcb->pcb_l1vec; 3239 curpcb->pcb_dacr = pcb->pcb_dacr; 3240 curpcb->pcb_pagedir = pcb->pcb_pagedir; 3241 } 3242 3243 LIST_REMOVE(pm, pm_list); 3244 3245 pmap_free_l1(pm); 3246 3247 /* return the pmap to the pool */ 3248 pool_cache_put(&pmap_pmap_cache, pm); 3249 } 3250 3251 3252 /* 3253 * void pmap_reference(pmap_t pm) 3254 * 3255 * Add a reference to the specified pmap. 3256 */ 3257 void 3258 pmap_reference(pmap_t pm) 3259 { 3260 3261 if (pm == NULL) 3262 return; 3263 3264 pmap_use_l1(pm); 3265 3266 simple_lock(&pm->pm_lock); 3267 pm->pm_obj.uo_refs++; 3268 simple_unlock(&pm->pm_lock); 3269 } 3270 3271 /* 3272 * pmap_zero_page() 3273 * 3274 * Zero a given physical page by mapping it at a page hook point. 3275 * In doing the zero page op, the page we zero is mapped cachable, as with 3276 * StrongARM accesses to non-cached pages are non-burst making writing 3277 * _any_ bulk data very slow. 3278 */ 3279 #if (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 3280 void 3281 pmap_zero_page_generic(paddr_t phys) 3282 { 3283 #ifdef DEBUG 3284 struct vm_page *pg = PHYS_TO_VM_PAGE(phys); 3285 3286 if (pg->mdpage.pvh_list != NULL) 3287 panic("pmap_zero_page: page has mappings"); 3288 #endif 3289 3290 KDASSERT((phys & PGOFSET) == 0); 3291 3292 /* 3293 * Hook in the page, zero it, and purge the cache for that 3294 * zeroed page. Invalidate the TLB as needed. 3295 */ 3296 *cdst_pte = L2_S_PROTO | phys | 3297 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode; 3298 PTE_SYNC(cdst_pte); 3299 cpu_tlb_flushD_SE(cdstp); 3300 cpu_cpwait(); 3301 bzero_page(cdstp); 3302 cpu_dcache_wbinv_range(cdstp, PAGE_SIZE); 3303 } 3304 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 */ 3305 3306 #if ARM_MMU_XSCALE == 1 3307 void 3308 pmap_zero_page_xscale(paddr_t phys) 3309 { 3310 #ifdef DEBUG 3311 struct vm_page *pg = PHYS_TO_VM_PAGE(phys); 3312 3313 if (pg->mdpage.pvh_list != NULL) 3314 panic("pmap_zero_page: page has mappings"); 3315 #endif 3316 3317 KDASSERT((phys & PGOFSET) == 0); 3318 3319 /* 3320 * Hook in the page, zero it, and purge the cache for that 3321 * zeroed page. Invalidate the TLB as needed. 3322 */ 3323 *cdst_pte = L2_S_PROTO | phys | 3324 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | 3325 L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X); /* mini-data */ 3326 PTE_SYNC(cdst_pte); 3327 cpu_tlb_flushD_SE(cdstp); 3328 cpu_cpwait(); 3329 bzero_page(cdstp); 3330 xscale_cache_clean_minidata(); 3331 } 3332 #endif /* ARM_MMU_XSCALE == 1 */ 3333 3334 /* pmap_pageidlezero() 3335 * 3336 * The same as above, except that we assume that the page is not 3337 * mapped. This means we never have to flush the cache first. Called 3338 * from the idle loop. 3339 */ 3340 boolean_t 3341 pmap_pageidlezero(paddr_t phys) 3342 { 3343 unsigned int i; 3344 int *ptr; 3345 boolean_t rv = TRUE; 3346 #ifdef DEBUG 3347 struct vm_page *pg; 3348 3349 pg = PHYS_TO_VM_PAGE(phys); 3350 if (pg->mdpage.pvh_list != NULL) 3351 panic("pmap_pageidlezero: page has mappings"); 3352 #endif 3353 3354 KDASSERT((phys & PGOFSET) == 0); 3355 3356 /* 3357 * Hook in the page, zero it, and purge the cache for that 3358 * zeroed page. Invalidate the TLB as needed. 3359 */ 3360 *cdst_pte = L2_S_PROTO | phys | 3361 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode; 3362 PTE_SYNC(cdst_pte); 3363 cpu_tlb_flushD_SE(cdstp); 3364 cpu_cpwait(); 3365 3366 for (i = 0, ptr = (int *)cdstp; 3367 i < (PAGE_SIZE / sizeof(int)); i++) { 3368 if (sched_whichqs != 0) { 3369 /* 3370 * A process has become ready. Abort now, 3371 * so we don't keep it waiting while we 3372 * do slow memory access to finish this 3373 * page. 3374 */ 3375 rv = FALSE; 3376 break; 3377 } 3378 *ptr++ = 0; 3379 } 3380 3381 if (rv) 3382 /* 3383 * if we aborted we'll rezero this page again later so don't 3384 * purge it unless we finished it 3385 */ 3386 cpu_dcache_wbinv_range(cdstp, PAGE_SIZE); 3387 3388 return (rv); 3389 } 3390 3391 /* 3392 * pmap_copy_page() 3393 * 3394 * Copy one physical page into another, by mapping the pages into 3395 * hook points. The same comment regarding cachability as in 3396 * pmap_zero_page also applies here. 3397 */ 3398 #if (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 3399 void 3400 pmap_copy_page_generic(paddr_t src, paddr_t dst) 3401 { 3402 struct vm_page *src_pg = PHYS_TO_VM_PAGE(src); 3403 #ifdef DEBUG 3404 struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst); 3405 3406 if (dst_pg->mdpage.pvh_list != NULL) 3407 panic("pmap_copy_page: dst page has mappings"); 3408 #endif 3409 3410 KDASSERT((src & PGOFSET) == 0); 3411 KDASSERT((dst & PGOFSET) == 0); 3412 3413 /* 3414 * Clean the source page. Hold the source page's lock for 3415 * the duration of the copy so that no other mappings can 3416 * be created while we have a potentially aliased mapping. 3417 */ 3418 simple_lock(&src_pg->mdpage.pvh_slock); 3419 (void) pmap_clean_page(src_pg->mdpage.pvh_list, TRUE); 3420 3421 /* 3422 * Map the pages into the page hook points, copy them, and purge 3423 * the cache for the appropriate page. Invalidate the TLB 3424 * as required. 3425 */ 3426 *csrc_pte = L2_S_PROTO | src | 3427 L2_S_PROT(PTE_KERNEL, VM_PROT_READ) | pte_l2_s_cache_mode; 3428 PTE_SYNC(csrc_pte); 3429 *cdst_pte = L2_S_PROTO | dst | 3430 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | pte_l2_s_cache_mode; 3431 PTE_SYNC(cdst_pte); 3432 cpu_tlb_flushD_SE(csrcp); 3433 cpu_tlb_flushD_SE(cdstp); 3434 cpu_cpwait(); 3435 bcopy_page(csrcp, cdstp); 3436 cpu_dcache_inv_range(csrcp, PAGE_SIZE); 3437 simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */ 3438 cpu_dcache_wbinv_range(cdstp, PAGE_SIZE); 3439 } 3440 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 */ 3441 3442 #if ARM_MMU_XSCALE == 1 3443 void 3444 pmap_copy_page_xscale(paddr_t src, paddr_t dst) 3445 { 3446 struct vm_page *src_pg = PHYS_TO_VM_PAGE(src); 3447 #ifdef DEBUG 3448 struct vm_page *dst_pg = PHYS_TO_VM_PAGE(dst); 3449 3450 if (dst_pg->mdpage.pvh_list != NULL) 3451 panic("pmap_copy_page: dst page has mappings"); 3452 #endif 3453 3454 KDASSERT((src & PGOFSET) == 0); 3455 KDASSERT((dst & PGOFSET) == 0); 3456 3457 /* 3458 * Clean the source page. Hold the source page's lock for 3459 * the duration of the copy so that no other mappings can 3460 * be created while we have a potentially aliased mapping. 3461 */ 3462 simple_lock(&src_pg->mdpage.pvh_slock); 3463 (void) pmap_clean_page(src_pg->mdpage.pvh_list, TRUE); 3464 3465 /* 3466 * Map the pages into the page hook points, copy them, and purge 3467 * the cache for the appropriate page. Invalidate the TLB 3468 * as required. 3469 */ 3470 *csrc_pte = L2_S_PROTO | src | 3471 L2_S_PROT(PTE_KERNEL, VM_PROT_READ) | 3472 L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X); /* mini-data */ 3473 PTE_SYNC(csrc_pte); 3474 *cdst_pte = L2_S_PROTO | dst | 3475 L2_S_PROT(PTE_KERNEL, VM_PROT_WRITE) | 3476 L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X); /* mini-data */ 3477 PTE_SYNC(cdst_pte); 3478 cpu_tlb_flushD_SE(csrcp); 3479 cpu_tlb_flushD_SE(cdstp); 3480 cpu_cpwait(); 3481 bcopy_page(csrcp, cdstp); 3482 simple_unlock(&src_pg->mdpage.pvh_slock); /* cache is safe again */ 3483 xscale_cache_clean_minidata(); 3484 } 3485 #endif /* ARM_MMU_XSCALE == 1 */ 3486 3487 /* 3488 * void pmap_virtual_space(vaddr_t *start, vaddr_t *end) 3489 * 3490 * Return the start and end addresses of the kernel's virtual space. 3491 * These values are setup in pmap_bootstrap and are updated as pages 3492 * are allocated. 3493 */ 3494 void 3495 pmap_virtual_space(vaddr_t *start, vaddr_t *end) 3496 { 3497 *start = virtual_avail; 3498 *end = virtual_end; 3499 } 3500 3501 /* 3502 * Helper function for pmap_grow_l2_bucket() 3503 */ 3504 static __inline int 3505 pmap_grow_map(vaddr_t va, pt_entry_t cache_mode, paddr_t *pap) 3506 { 3507 struct l2_bucket *l2b; 3508 pt_entry_t *ptep; 3509 paddr_t pa; 3510 3511 if (uvm.page_init_done == FALSE) { 3512 if (uvm_page_physget(&pa) == FALSE) 3513 return (1); 3514 } else { 3515 struct vm_page *pg; 3516 pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_USERESERVE); 3517 if (pg == NULL) 3518 return (1); 3519 pa = VM_PAGE_TO_PHYS(pg); 3520 } 3521 3522 if (pap) 3523 *pap = pa; 3524 3525 l2b = pmap_get_l2_bucket(pmap_kernel(), va); 3526 KDASSERT(l2b != NULL); 3527 3528 ptep = &l2b->l2b_kva[l2pte_index(va)]; 3529 *ptep = L2_S_PROTO | pa | cache_mode | 3530 L2_S_PROT(PTE_KERNEL, VM_PROT_READ | VM_PROT_WRITE); 3531 PTE_SYNC(ptep); 3532 memset((void *)va, 0, PAGE_SIZE); 3533 return (0); 3534 } 3535 3536 /* 3537 * This is the same as pmap_alloc_l2_bucket(), except that it is only 3538 * used by pmap_growkernel(). 3539 */ 3540 static __inline struct l2_bucket * 3541 pmap_grow_l2_bucket(pmap_t pm, vaddr_t va) 3542 { 3543 struct l2_dtable *l2; 3544 struct l2_bucket *l2b; 3545 u_short l1idx; 3546 vaddr_t nva; 3547 3548 l1idx = L1_IDX(va); 3549 3550 if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) { 3551 /* 3552 * No mapping at this address, as there is 3553 * no entry in the L1 table. 3554 * Need to allocate a new l2_dtable. 3555 */ 3556 nva = pmap_kernel_l2dtable_kva; 3557 if ((nva & PGOFSET) == 0) { 3558 /* 3559 * Need to allocate a backing page 3560 */ 3561 if (pmap_grow_map(nva, pte_l2_s_cache_mode, NULL)) 3562 return (NULL); 3563 } 3564 3565 l2 = (struct l2_dtable *)nva; 3566 nva += sizeof(struct l2_dtable); 3567 3568 if ((nva & PGOFSET) < (pmap_kernel_l2dtable_kva & PGOFSET)) { 3569 /* 3570 * The new l2_dtable straddles a page boundary. 3571 * Map in another page to cover it. 3572 */ 3573 if (pmap_grow_map(nva, pte_l2_s_cache_mode, NULL)) 3574 return (NULL); 3575 } 3576 3577 pmap_kernel_l2dtable_kva = nva; 3578 3579 /* 3580 * Link it into the parent pmap 3581 */ 3582 pm->pm_l2[L2_IDX(l1idx)] = l2; 3583 } 3584 3585 l2b = &l2->l2_bucket[L2_BUCKET(l1idx)]; 3586 3587 /* 3588 * Fetch pointer to the L2 page table associated with the address. 3589 */ 3590 if (l2b->l2b_kva == NULL) { 3591 pt_entry_t *ptep; 3592 3593 /* 3594 * No L2 page table has been allocated. Chances are, this 3595 * is because we just allocated the l2_dtable, above. 3596 */ 3597 nva = pmap_kernel_l2ptp_kva; 3598 ptep = (pt_entry_t *)nva; 3599 if ((nva & PGOFSET) == 0) { 3600 /* 3601 * Need to allocate a backing page 3602 */ 3603 if (pmap_grow_map(nva, pte_l2_s_cache_mode_pt, 3604 &pmap_kernel_l2ptp_phys)) 3605 return (NULL); 3606 PTE_SYNC_RANGE(ptep, PAGE_SIZE / sizeof(pt_entry_t)); 3607 } 3608 3609 l2->l2_occupancy++; 3610 l2b->l2b_kva = ptep; 3611 l2b->l2b_l1idx = l1idx; 3612 l2b->l2b_phys = pmap_kernel_l2ptp_phys; 3613 3614 pmap_kernel_l2ptp_kva += L2_TABLE_SIZE_REAL; 3615 pmap_kernel_l2ptp_phys += L2_TABLE_SIZE_REAL; 3616 } 3617 3618 return (l2b); 3619 } 3620 3621 vaddr_t 3622 pmap_growkernel(vaddr_t maxkvaddr) 3623 { 3624 pmap_t kpm = pmap_kernel(); 3625 struct l1_ttable *l1; 3626 struct l2_bucket *l2b; 3627 pd_entry_t *pl1pd; 3628 int s; 3629 3630 if (maxkvaddr <= pmap_curmaxkvaddr) 3631 goto out; /* we are OK */ 3632 3633 NPDEBUG(PDB_GROWKERN, 3634 printf("pmap_growkernel: growing kernel from 0x%lx to 0x%lx\n", 3635 pmap_curmaxkvaddr, maxkvaddr)); 3636 3637 KDASSERT(maxkvaddr <= virtual_end); 3638 3639 /* 3640 * whoops! we need to add kernel PTPs 3641 */ 3642 3643 s = splhigh(); /* to be safe */ 3644 simple_lock(&kpm->pm_lock); 3645 3646 /* Map 1MB at a time */ 3647 for (; pmap_curmaxkvaddr < maxkvaddr; pmap_curmaxkvaddr += L1_S_SIZE) { 3648 3649 l2b = pmap_grow_l2_bucket(kpm, pmap_curmaxkvaddr); 3650 KDASSERT(l2b != NULL); 3651 3652 /* Distribute new L1 entry to all other L1s */ 3653 SLIST_FOREACH(l1, &l1_list, l1_link) { 3654 pl1pd = &l1->l1_kva[L1_IDX(pmap_curmaxkvaddr)]; 3655 *pl1pd = l2b->l2b_phys | L1_C_DOM(PMAP_DOMAIN_KERNEL) | 3656 L1_C_PROTO; 3657 PTE_SYNC(pl1pd); 3658 } 3659 } 3660 3661 /* 3662 * flush out the cache, expensive but growkernel will happen so 3663 * rarely 3664 */ 3665 cpu_dcache_wbinv_all(); 3666 cpu_tlb_flushD(); 3667 cpu_cpwait(); 3668 3669 simple_unlock(&kpm->pm_lock); 3670 splx(s); 3671 3672 out: 3673 return (pmap_curmaxkvaddr); 3674 } 3675 3676 /************************ Utility routines ****************************/ 3677 3678 /* 3679 * vector_page_setprot: 3680 * 3681 * Manipulate the protection of the vector page. 3682 */ 3683 void 3684 vector_page_setprot(int prot) 3685 { 3686 struct l2_bucket *l2b; 3687 pt_entry_t *ptep; 3688 3689 l2b = pmap_get_l2_bucket(pmap_kernel(), vector_page); 3690 KDASSERT(l2b != NULL); 3691 3692 ptep = &l2b->l2b_kva[l2pte_index(vector_page)]; 3693 3694 *ptep = (*ptep & ~L1_S_PROT_MASK) | L2_S_PROT(PTE_KERNEL, prot); 3695 PTE_SYNC(ptep); 3696 cpu_tlb_flushD_SE(vector_page); 3697 cpu_cpwait(); 3698 } 3699 3700 /* 3701 * This is used to stuff certain critical values into the PCB where they 3702 * can be accessed quickly from cpu_switch() et al. 3703 */ 3704 void 3705 pmap_set_pcb_pagedir(pmap_t pm, struct pcb *pcb) 3706 { 3707 struct l2_bucket *l2b; 3708 3709 KDASSERT(pm->pm_l1); 3710 3711 pcb->pcb_pagedir = pm->pm_l1->l1_physaddr; 3712 pcb->pcb_dacr = (DOMAIN_CLIENT << (PMAP_DOMAIN_KERNEL * 2)) | 3713 (DOMAIN_CLIENT << (pm->pm_domain * 2)); 3714 pcb->pcb_cstate = (void *)&pm->pm_cstate; 3715 3716 if (vector_page < KERNEL_BASE) { 3717 pcb->pcb_pl1vec = &pm->pm_l1->l1_kva[L1_IDX(vector_page)]; 3718 l2b = pmap_get_l2_bucket(pm, vector_page); 3719 pcb->pcb_l1vec = l2b->l2b_phys | L1_C_PROTO | 3720 L1_C_DOM(pm->pm_domain); 3721 } else 3722 pcb->pcb_pl1vec = NULL; 3723 } 3724 3725 /* 3726 * Fetch pointers to the PDE/PTE for the given pmap/VA pair. 3727 * Returns TRUE if the mapping exists, else FALSE. 3728 * 3729 * NOTE: This function is only used by a couple of arm-specific modules. 3730 * It is not safe to take any pmap locks here, since we could be right 3731 * in the middle of debugging the pmap anyway... 3732 * 3733 * It is possible for this routine to return FALSE even though a valid 3734 * mapping does exist. This is because we don't lock, so the metadata 3735 * state may be inconsistent. 3736 * 3737 * NOTE: We can return a NULL *ptp in the case where the L1 pde is 3738 * a "section" mapping. 3739 */ 3740 boolean_t 3741 pmap_get_pde_pte(pmap_t pm, vaddr_t va, pd_entry_t **pdp, pt_entry_t **ptp) 3742 { 3743 struct l2_dtable *l2; 3744 pd_entry_t *pl1pd, l1pd; 3745 pt_entry_t *ptep; 3746 u_short l1idx; 3747 3748 if (pm->pm_l1 == NULL) 3749 return (FALSE); 3750 3751 l1idx = L1_IDX(va); 3752 *pdp = pl1pd = &pm->pm_l1->l1_kva[l1idx]; 3753 l1pd = *pl1pd; 3754 3755 if (l1pte_section_p(l1pd)) { 3756 *ptp = NULL; 3757 return (TRUE); 3758 } 3759 3760 if (pm->pm_l2 == NULL) 3761 return (FALSE); 3762 3763 l2 = pm->pm_l2[L2_IDX(l1idx)]; 3764 3765 if (l2 == NULL || 3766 (ptep = l2->l2_bucket[L2_BUCKET(l1idx)].l2b_kva) == NULL) { 3767 return (FALSE); 3768 } 3769 3770 *ptp = &ptep[l2pte_index(va)]; 3771 return (TRUE); 3772 } 3773 3774 boolean_t 3775 pmap_get_pde(pmap_t pm, vaddr_t va, pd_entry_t **pdp) 3776 { 3777 u_short l1idx; 3778 3779 if (pm->pm_l1 == NULL) 3780 return (FALSE); 3781 3782 l1idx = L1_IDX(va); 3783 *pdp = &pm->pm_l1->l1_kva[l1idx]; 3784 3785 return (TRUE); 3786 } 3787 3788 /************************ Bootstrapping routines ****************************/ 3789 3790 static void 3791 pmap_init_l1(struct l1_ttable *l1, pd_entry_t *l1pt) 3792 { 3793 int i; 3794 3795 l1->l1_kva = l1pt; 3796 l1->l1_domain_use_count = 0; 3797 l1->l1_domain_first = 0; 3798 3799 for (i = 0; i < PMAP_DOMAINS; i++) 3800 l1->l1_domain_free[i] = i + 1; 3801 3802 /* 3803 * Copy the kernel's L1 entries to each new L1. 3804 */ 3805 if (pmap_initialized) 3806 memcpy(l1pt, pmap_kernel()->pm_l1->l1_kva, L1_TABLE_SIZE); 3807 3808 if (pmap_extract(pmap_kernel(), (vaddr_t)l1pt, 3809 &l1->l1_physaddr) == FALSE) 3810 panic("pmap_init_l1: can't get PA of L1 at %p", l1pt); 3811 3812 SLIST_INSERT_HEAD(&l1_list, l1, l1_link); 3813 TAILQ_INSERT_TAIL(&l1_lru_list, l1, l1_lru); 3814 } 3815 3816 /* 3817 * pmap_bootstrap() is called from the board-specific initarm() routine 3818 * once the kernel L1/L2 descriptors tables have been set up. 3819 * 3820 * This is a somewhat convoluted process since pmap bootstrap is, effectively, 3821 * spread over a number of disparate files/functions. 3822 * 3823 * We are passed the following parameters 3824 * - kernel_l1pt 3825 * This is a pointer to the base of the kernel's L1 translation table. 3826 * - vstart 3827 * 1MB-aligned start of managed kernel virtual memory. 3828 * - vend 3829 * 1MB-aligned end of managed kernel virtual memory. 3830 * 3831 * We use the first parameter to build the metadata (struct l1_ttable and 3832 * struct l2_dtable) necessary to track kernel mappings. 3833 */ 3834 #define PMAP_STATIC_L2_SIZE 16 3835 void 3836 pmap_bootstrap(pd_entry_t *kernel_l1pt, vaddr_t vstart, vaddr_t vend) 3837 { 3838 static struct l1_ttable static_l1; 3839 static struct l2_dtable static_l2[PMAP_STATIC_L2_SIZE]; 3840 struct l1_ttable *l1 = &static_l1; 3841 struct l2_dtable *l2; 3842 struct l2_bucket *l2b; 3843 pmap_t pm = pmap_kernel(); 3844 pd_entry_t pde; 3845 pt_entry_t *ptep; 3846 paddr_t pa; 3847 vaddr_t va; 3848 vsize_t size; 3849 int l1idx, l2idx, l2next = 0; 3850 3851 /* 3852 * Initialise the kernel pmap object 3853 */ 3854 pm->pm_l1 = l1; 3855 pm->pm_domain = PMAP_DOMAIN_KERNEL; 3856 pm->pm_cstate.cs_all = PMAP_CACHE_STATE_ALL; 3857 simple_lock_init(&pm->pm_lock); 3858 pm->pm_obj.pgops = NULL; 3859 TAILQ_INIT(&pm->pm_obj.memq); 3860 pm->pm_obj.uo_npages = 0; 3861 pm->pm_obj.uo_refs = 1; 3862 3863 /* 3864 * Scan the L1 translation table created by initarm() and create 3865 * the required metadata for all valid mappings found in it. 3866 */ 3867 for (l1idx = 0; l1idx < (L1_TABLE_SIZE / sizeof(pd_entry_t)); l1idx++) { 3868 pde = kernel_l1pt[l1idx]; 3869 3870 /* 3871 * We're only interested in Coarse mappings. 3872 * pmap_extract() can deal with section mappings without 3873 * recourse to checking L2 metadata. 3874 */ 3875 if ((pde & L1_TYPE_MASK) != L1_TYPE_C) 3876 continue; 3877 3878 /* 3879 * Lookup the KVA of this L2 descriptor table 3880 */ 3881 pa = (paddr_t)(pde & L1_C_ADDR_MASK); 3882 ptep = (pt_entry_t *)kernel_pt_lookup(pa); 3883 if (ptep == NULL) { 3884 panic("pmap_bootstrap: No L2 for va 0x%x, pa 0x%lx", 3885 (u_int)l1idx << L1_S_SHIFT, pa); 3886 } 3887 3888 /* 3889 * Fetch the associated L2 metadata structure. 3890 * Allocate a new one if necessary. 3891 */ 3892 if ((l2 = pm->pm_l2[L2_IDX(l1idx)]) == NULL) { 3893 if (l2next == PMAP_STATIC_L2_SIZE) 3894 panic("pmap_bootstrap: out of static L2s"); 3895 pm->pm_l2[L2_IDX(l1idx)] = l2 = &static_l2[l2next++]; 3896 } 3897 3898 /* 3899 * One more L1 slot tracked... 3900 */ 3901 l2->l2_occupancy++; 3902 3903 /* 3904 * Fill in the details of the L2 descriptor in the 3905 * appropriate bucket. 3906 */ 3907 l2b = &l2->l2_bucket[L2_BUCKET(l1idx)]; 3908 l2b->l2b_kva = ptep; 3909 l2b->l2b_phys = pa; 3910 l2b->l2b_l1idx = l1idx; 3911 3912 /* 3913 * Establish an initial occupancy count for this descriptor 3914 */ 3915 for (l2idx = 0; 3916 l2idx < (L2_TABLE_SIZE_REAL / sizeof(pt_entry_t)); 3917 l2idx++) { 3918 if ((ptep[l2idx] & L2_TYPE_MASK) != L2_TYPE_INV) { 3919 l2b->l2b_occupancy++; 3920 } 3921 } 3922 3923 /* 3924 * Make sure the descriptor itself has the correct cache mode. 3925 * If not, fix it, but whine about the problem. Port-meisters 3926 * should consider this a clue to fix up their initarm() 3927 * function. :) 3928 */ 3929 if (pmap_set_pt_cache_mode(kernel_l1pt, (vaddr_t)ptep)) { 3930 printf("pmap_bootstrap: WARNING! wrong cache mode for " 3931 "L2 pte @ %p\n", ptep); 3932 } 3933 } 3934 3935 /* 3936 * Ensure the primary (kernel) L1 has the correct cache mode for 3937 * a page table. Bitch if it is not correctly set. 3938 */ 3939 for (va = (vaddr_t)kernel_l1pt; 3940 va < ((vaddr_t)kernel_l1pt + L1_TABLE_SIZE); va += PAGE_SIZE) { 3941 if (pmap_set_pt_cache_mode(kernel_l1pt, va)) 3942 printf("pmap_bootstrap: WARNING! wrong cache mode for " 3943 "primary L1 @ 0x%lx\n", va); 3944 } 3945 3946 cpu_dcache_wbinv_all(); 3947 cpu_tlb_flushID(); 3948 cpu_cpwait(); 3949 3950 /* 3951 * now we allocate the "special" VAs which are used for tmp mappings 3952 * by the pmap (and other modules). we allocate the VAs by advancing 3953 * virtual_avail (note that there are no pages mapped at these VAs). 3954 * 3955 * Managed KVM space start from wherever initarm() tells us. 3956 */ 3957 virtual_avail = vstart; 3958 virtual_end = vend; 3959 3960 pmap_alloc_specials(&virtual_avail, 1, &csrcp, &csrc_pte); 3961 pmap_set_pt_cache_mode(kernel_l1pt, (vaddr_t)csrc_pte); 3962 pmap_alloc_specials(&virtual_avail, 1, &cdstp, &cdst_pte); 3963 pmap_set_pt_cache_mode(kernel_l1pt, (vaddr_t)cdst_pte); 3964 pmap_alloc_specials(&virtual_avail, 1, (void *)&memhook, NULL); 3965 pmap_alloc_specials(&virtual_avail, round_page(MSGBUFSIZE) / PAGE_SIZE, 3966 (void *)&msgbufaddr, NULL); 3967 3968 /* 3969 * Allocate a range of kernel virtual address space to be used 3970 * for L2 descriptor tables and metadata allocation in 3971 * pmap_growkernel(). 3972 */ 3973 size = ((virtual_end - pmap_curmaxkvaddr) + L1_S_OFFSET) / L1_S_SIZE; 3974 pmap_alloc_specials(&virtual_avail, 3975 round_page(size * L2_TABLE_SIZE_REAL) / PAGE_SIZE, 3976 &pmap_kernel_l2ptp_kva, NULL); 3977 3978 size = (size + (L2_BUCKET_SIZE - 1)) / L2_BUCKET_SIZE; 3979 pmap_alloc_specials(&virtual_avail, 3980 round_page(size * sizeof(struct l2_dtable)) / PAGE_SIZE, 3981 &pmap_kernel_l2dtable_kva, NULL); 3982 3983 /* 3984 * init the static-global locks and global pmap list. 3985 */ 3986 #if defined(MULTIPROCESSOR) || defined(LOCKDEBUG) 3987 spinlockinit(&pmap_main_lock, "pmaplk", 0); 3988 #endif 3989 3990 /* 3991 * We can now initialise the first L1's metadata. 3992 */ 3993 SLIST_INIT(&l1_list); 3994 TAILQ_INIT(&l1_lru_list); 3995 simple_lock_init(&l1_lru_lock); 3996 pmap_init_l1(l1, kernel_l1pt); 3997 3998 /* 3999 * Initialize the pmap pool and cache 4000 */ 4001 pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl", 4002 &pool_allocator_nointr); 4003 pool_cache_init(&pmap_pmap_cache, &pmap_pmap_pool, 4004 pmap_pmap_ctor, NULL, NULL); 4005 LIST_INIT(&pmap_pmaps); 4006 LIST_INSERT_HEAD(&pmap_pmaps, pm, pm_list); 4007 4008 /* 4009 * Initialize the pv pool. 4010 */ 4011 pool_init(&pmap_pv_pool, sizeof(struct pv_entry), 0, 0, 0, "pvepl", 4012 &pmap_bootstrap_pv_allocator); 4013 4014 /* 4015 * Initialize the L2 dtable pool and cache. 4016 */ 4017 pool_init(&pmap_l2dtable_pool, sizeof(struct l2_dtable), 0, 0, 0, 4018 "l2dtblpl", NULL); 4019 pool_cache_init(&pmap_l2dtable_cache, &pmap_l2dtable_pool, 4020 pmap_l2dtable_ctor, NULL, NULL); 4021 4022 /* 4023 * Initialise the L2 descriptor table pool and cache 4024 */ 4025 pool_init(&pmap_l2ptp_pool, L2_TABLE_SIZE_REAL, 0, L2_TABLE_SIZE_REAL, 4026 0, "l2ptppl", NULL); 4027 pool_cache_init(&pmap_l2ptp_cache, &pmap_l2ptp_pool, 4028 pmap_l2ptp_ctor, NULL, NULL); 4029 4030 cpu_dcache_wbinv_all(); 4031 } 4032 4033 static int 4034 pmap_set_pt_cache_mode(pd_entry_t *kl1, vaddr_t va) 4035 { 4036 pd_entry_t *pdep, pde; 4037 pt_entry_t *ptep, pte; 4038 vaddr_t pa; 4039 int rv = 0; 4040 4041 /* 4042 * Make sure the descriptor itself has the correct cache mode 4043 */ 4044 pdep = &kl1[L1_IDX(va)]; 4045 pde = *pdep; 4046 4047 if (l1pte_section_p(pde)) { 4048 if ((pde & L1_S_CACHE_MASK) != pte_l1_s_cache_mode_pt) { 4049 *pdep = (pde & ~L1_S_CACHE_MASK) | 4050 pte_l1_s_cache_mode_pt; 4051 PTE_SYNC(pdep); 4052 cpu_dcache_wbinv_range((vaddr_t)pdep, sizeof(*pdep)); 4053 rv = 1; 4054 } 4055 } else { 4056 pa = (paddr_t)(pde & L1_C_ADDR_MASK); 4057 ptep = (pt_entry_t *)kernel_pt_lookup(pa); 4058 if (ptep == NULL) 4059 panic("pmap_bootstrap: No L2 for L2 @ va %p\n", ptep); 4060 4061 ptep = &ptep[l2pte_index(va)]; 4062 pte = *ptep; 4063 if ((pte & L2_S_CACHE_MASK) != pte_l2_s_cache_mode_pt) { 4064 *ptep = (pte & ~L2_S_CACHE_MASK) | 4065 pte_l2_s_cache_mode_pt; 4066 PTE_SYNC(ptep); 4067 cpu_dcache_wbinv_range((vaddr_t)ptep, sizeof(*ptep)); 4068 rv = 1; 4069 } 4070 } 4071 4072 return (rv); 4073 } 4074 4075 static void 4076 pmap_alloc_specials(vaddr_t *availp, int pages, vaddr_t *vap, pt_entry_t **ptep) 4077 { 4078 vaddr_t va = *availp; 4079 struct l2_bucket *l2b; 4080 4081 if (ptep) { 4082 l2b = pmap_get_l2_bucket(pmap_kernel(), va); 4083 if (l2b == NULL) 4084 panic("pmap_alloc_specials: no l2b for 0x%lx", va); 4085 4086 if (ptep) 4087 *ptep = &l2b->l2b_kva[l2pte_index(va)]; 4088 } 4089 4090 *vap = va; 4091 *availp = va + (PAGE_SIZE * pages); 4092 } 4093 4094 void 4095 pmap_init(void) 4096 { 4097 extern int physmem; 4098 4099 /* 4100 * Set the available memory vars - These do not map to real memory 4101 * addresses and cannot as the physical memory is fragmented. 4102 * They are used by ps for %mem calculations. 4103 * One could argue whether this should be the entire memory or just 4104 * the memory that is useable in a user process. 4105 */ 4106 avail_start = 0; 4107 avail_end = physmem * PAGE_SIZE; 4108 4109 /* 4110 * Now we need to free enough pv_entry structures to allow us to get 4111 * the kmem_map/kmem_object allocated and inited (done after this 4112 * function is finished). to do this we allocate one bootstrap page out 4113 * of kernel_map and use it to provide an initial pool of pv_entry 4114 * structures. we never free this page. 4115 */ 4116 pool_setlowat(&pmap_pv_pool, 4117 (PAGE_SIZE / sizeof(struct pv_entry)) * 2); 4118 4119 pmap_initialized = TRUE; 4120 } 4121 4122 static vaddr_t last_bootstrap_page = 0; 4123 static void *free_bootstrap_pages = NULL; 4124 4125 static void * 4126 pmap_bootstrap_pv_page_alloc(struct pool *pp, int flags) 4127 { 4128 extern void *pool_page_alloc(struct pool *, int); 4129 vaddr_t new_page; 4130 void *rv; 4131 4132 if (pmap_initialized) 4133 return (pool_page_alloc(pp, flags)); 4134 4135 if (free_bootstrap_pages) { 4136 rv = free_bootstrap_pages; 4137 free_bootstrap_pages = *((void **)rv); 4138 return (rv); 4139 } 4140 4141 new_page = uvm_km_alloc(kernel_map, PAGE_SIZE, 0, 4142 UVM_KMF_WIRED | ((flags & PR_WAITOK) ? 0 : UVM_KMF_NOWAIT)); 4143 4144 KASSERT(new_page > last_bootstrap_page); 4145 last_bootstrap_page = new_page; 4146 return ((void *)new_page); 4147 } 4148 4149 static void 4150 pmap_bootstrap_pv_page_free(struct pool *pp, void *v) 4151 { 4152 extern void pool_page_free(struct pool *, void *); 4153 4154 if ((vaddr_t)v <= last_bootstrap_page) { 4155 *((void **)v) = free_bootstrap_pages; 4156 free_bootstrap_pages = v; 4157 return; 4158 } 4159 4160 if (pmap_initialized) { 4161 pool_page_free(pp, v); 4162 return; 4163 } 4164 } 4165 4166 /* 4167 * pmap_postinit() 4168 * 4169 * This routine is called after the vm and kmem subsystems have been 4170 * initialised. This allows the pmap code to perform any initialisation 4171 * that can only be done one the memory allocation is in place. 4172 */ 4173 void 4174 pmap_postinit(void) 4175 { 4176 extern paddr_t physical_start, physical_end; 4177 struct l2_bucket *l2b; 4178 struct l1_ttable *l1; 4179 struct pglist plist; 4180 struct vm_page *m; 4181 pd_entry_t *pl1pt; 4182 pt_entry_t *ptep, pte; 4183 vaddr_t va, eva; 4184 u_int loop, needed; 4185 int error; 4186 4187 pool_setlowat(&pmap_l2ptp_pool, 4188 (PAGE_SIZE / L2_TABLE_SIZE_REAL) * 4); 4189 pool_setlowat(&pmap_l2dtable_pool, 4190 (PAGE_SIZE / sizeof(struct l2_dtable)) * 2); 4191 4192 needed = (maxproc / PMAP_DOMAINS) + ((maxproc % PMAP_DOMAINS) ? 1 : 0); 4193 needed -= 1; 4194 4195 l1 = malloc(sizeof(*l1) * needed, M_VMPMAP, M_WAITOK); 4196 4197 for (loop = 0; loop < needed; loop++, l1++) { 4198 /* Allocate a L1 page table */ 4199 va = uvm_km_alloc(kernel_map, L1_TABLE_SIZE, 0, UVM_KMF_VAONLY); 4200 if (va == 0) 4201 panic("Cannot allocate L1 KVM"); 4202 4203 error = uvm_pglistalloc(L1_TABLE_SIZE, physical_start, 4204 physical_end, L1_TABLE_SIZE, 0, &plist, 1, M_WAITOK); 4205 if (error) 4206 panic("Cannot allocate L1 physical pages"); 4207 4208 m = TAILQ_FIRST(&plist); 4209 eva = va + L1_TABLE_SIZE; 4210 pl1pt = (pd_entry_t *)va; 4211 4212 while (m && va < eva) { 4213 paddr_t pa = VM_PAGE_TO_PHYS(m); 4214 4215 pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE); 4216 4217 /* 4218 * Make sure the L1 descriptor table is mapped 4219 * with the cache-mode set to write-through. 4220 */ 4221 l2b = pmap_get_l2_bucket(pmap_kernel(), va); 4222 ptep = &l2b->l2b_kva[l2pte_index(va)]; 4223 pte = *ptep; 4224 pte = (pte & ~L2_S_CACHE_MASK) | pte_l2_s_cache_mode_pt; 4225 *ptep = pte; 4226 PTE_SYNC(ptep); 4227 cpu_tlb_flushD_SE(va); 4228 4229 va += PAGE_SIZE; 4230 m = TAILQ_NEXT(m, pageq); 4231 } 4232 4233 #ifdef DIAGNOSTIC 4234 if (m) 4235 panic("pmap_alloc_l1pt: pglist not empty"); 4236 #endif /* DIAGNOSTIC */ 4237 4238 pmap_init_l1(l1, pl1pt); 4239 } 4240 4241 #ifdef DEBUG 4242 printf("pmap_postinit: Allocated %d static L1 descriptor tables\n", 4243 needed); 4244 #endif 4245 } 4246 4247 /* 4248 * Note that the following routines are used by board-specific initialisation 4249 * code to configure the initial kernel page tables. 4250 * 4251 * If ARM32_NEW_VM_LAYOUT is *not* defined, they operate on the assumption that 4252 * L2 page-table pages are 4KB in size and use 4 L1 slots. This mimics the 4253 * behaviour of the old pmap, and provides an easy migration path for 4254 * initial bring-up of the new pmap on existing ports. Fortunately, 4255 * pmap_bootstrap() compensates for this hackery. This is only a stop-gap and 4256 * will be deprecated. 4257 * 4258 * If ARM32_NEW_VM_LAYOUT *is* defined, these functions deal with 1KB L2 page 4259 * tables. 4260 */ 4261 4262 /* 4263 * This list exists for the benefit of pmap_map_chunk(). It keeps track 4264 * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can 4265 * find them as necessary. 4266 * 4267 * Note that the data on this list MUST remain valid after initarm() returns, 4268 * as pmap_bootstrap() uses it to contruct L2 table metadata. 4269 */ 4270 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list); 4271 4272 static vaddr_t 4273 kernel_pt_lookup(paddr_t pa) 4274 { 4275 pv_addr_t *pv; 4276 4277 SLIST_FOREACH(pv, &kernel_pt_list, pv_list) { 4278 #ifndef ARM32_NEW_VM_LAYOUT 4279 if (pv->pv_pa == (pa & ~PGOFSET)) 4280 return (pv->pv_va | (pa & PGOFSET)); 4281 #else 4282 if (pv->pv_pa == pa) 4283 return (pv->pv_va); 4284 #endif 4285 } 4286 return (0); 4287 } 4288 4289 /* 4290 * pmap_map_section: 4291 * 4292 * Create a single section mapping. 4293 */ 4294 void 4295 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache) 4296 { 4297 pd_entry_t *pde = (pd_entry_t *) l1pt; 4298 pd_entry_t fl; 4299 4300 KASSERT(((va | pa) & L1_S_OFFSET) == 0); 4301 4302 switch (cache) { 4303 case PTE_NOCACHE: 4304 default: 4305 fl = 0; 4306 break; 4307 4308 case PTE_CACHE: 4309 fl = pte_l1_s_cache_mode; 4310 break; 4311 4312 case PTE_PAGETABLE: 4313 fl = pte_l1_s_cache_mode_pt; 4314 break; 4315 } 4316 4317 pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa | 4318 L1_S_PROT(PTE_KERNEL, prot) | fl | L1_S_DOM(PMAP_DOMAIN_KERNEL); 4319 PTE_SYNC(&pde[va >> L1_S_SHIFT]); 4320 } 4321 4322 /* 4323 * pmap_map_entry: 4324 * 4325 * Create a single page mapping. 4326 */ 4327 void 4328 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache) 4329 { 4330 pd_entry_t *pde = (pd_entry_t *) l1pt; 4331 pt_entry_t fl; 4332 pt_entry_t *pte; 4333 4334 KASSERT(((va | pa) & PGOFSET) == 0); 4335 4336 switch (cache) { 4337 case PTE_NOCACHE: 4338 default: 4339 fl = 0; 4340 break; 4341 4342 case PTE_CACHE: 4343 fl = pte_l2_s_cache_mode; 4344 break; 4345 4346 case PTE_PAGETABLE: 4347 fl = pte_l2_s_cache_mode_pt; 4348 break; 4349 } 4350 4351 if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C) 4352 panic("pmap_map_entry: no L2 table for VA 0x%08lx", va); 4353 4354 #ifndef ARM32_NEW_VM_LAYOUT 4355 pte = (pt_entry_t *) 4356 kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME); 4357 #else 4358 pte = (pt_entry_t *) kernel_pt_lookup(pde[L1_IDX(va)] & L1_C_ADDR_MASK); 4359 #endif 4360 if (pte == NULL) 4361 panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va); 4362 4363 #ifndef ARM32_NEW_VM_LAYOUT 4364 pte[(va >> PGSHIFT) & 0x3ff] = 4365 L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | fl; 4366 PTE_SYNC(&pte[(va >> PGSHIFT) & 0x3ff]); 4367 #else 4368 pte[l2pte_index(va)] = 4369 L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | fl; 4370 PTE_SYNC(&pte[l2pte_index(va)]); 4371 #endif 4372 } 4373 4374 /* 4375 * pmap_link_l2pt: 4376 * 4377 * Link the L2 page table specified by "l2pv" into the L1 4378 * page table at the slot for "va". 4379 */ 4380 void 4381 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv) 4382 { 4383 pd_entry_t *pde = (pd_entry_t *) l1pt, proto; 4384 u_int slot = va >> L1_S_SHIFT; 4385 4386 #ifndef ARM32_NEW_VM_LAYOUT 4387 KASSERT((va & ((L1_S_SIZE * 4) - 1)) == 0); 4388 KASSERT((l2pv->pv_pa & PGOFSET) == 0); 4389 #endif 4390 4391 proto = L1_S_DOM(PMAP_DOMAIN_KERNEL) | L1_C_PROTO; 4392 4393 pde[slot + 0] = proto | (l2pv->pv_pa + 0x000); 4394 #ifdef ARM32_NEW_VM_LAYOUT 4395 PTE_SYNC(&pde[slot]); 4396 #else 4397 pde[slot + 1] = proto | (l2pv->pv_pa + 0x400); 4398 pde[slot + 2] = proto | (l2pv->pv_pa + 0x800); 4399 pde[slot + 3] = proto | (l2pv->pv_pa + 0xc00); 4400 PTE_SYNC_RANGE(&pde[slot + 0], 4); 4401 #endif 4402 4403 SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list); 4404 } 4405 4406 /* 4407 * pmap_map_chunk: 4408 * 4409 * Map a chunk of memory using the most efficient mappings 4410 * possible (section, large page, small page) into the 4411 * provided L1 and L2 tables at the specified virtual address. 4412 */ 4413 vsize_t 4414 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size, 4415 int prot, int cache) 4416 { 4417 pd_entry_t *pde = (pd_entry_t *) l1pt; 4418 pt_entry_t *pte, f1, f2s, f2l; 4419 vsize_t resid; 4420 int i; 4421 4422 resid = (size + (PAGE_SIZE - 1)) & ~(PAGE_SIZE - 1); 4423 4424 if (l1pt == 0) 4425 panic("pmap_map_chunk: no L1 table provided"); 4426 4427 #ifdef VERBOSE_INIT_ARM 4428 printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx " 4429 "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache); 4430 #endif 4431 4432 switch (cache) { 4433 case PTE_NOCACHE: 4434 default: 4435 f1 = 0; 4436 f2l = 0; 4437 f2s = 0; 4438 break; 4439 4440 case PTE_CACHE: 4441 f1 = pte_l1_s_cache_mode; 4442 f2l = pte_l2_l_cache_mode; 4443 f2s = pte_l2_s_cache_mode; 4444 break; 4445 4446 case PTE_PAGETABLE: 4447 f1 = pte_l1_s_cache_mode_pt; 4448 f2l = pte_l2_l_cache_mode_pt; 4449 f2s = pte_l2_s_cache_mode_pt; 4450 break; 4451 } 4452 4453 size = resid; 4454 4455 while (resid > 0) { 4456 /* See if we can use a section mapping. */ 4457 if (L1_S_MAPPABLE_P(va, pa, resid)) { 4458 #ifdef VERBOSE_INIT_ARM 4459 printf("S"); 4460 #endif 4461 pde[va >> L1_S_SHIFT] = L1_S_PROTO | pa | 4462 L1_S_PROT(PTE_KERNEL, prot) | f1 | 4463 L1_S_DOM(PMAP_DOMAIN_KERNEL); 4464 PTE_SYNC(&pde[va >> L1_S_SHIFT]); 4465 va += L1_S_SIZE; 4466 pa += L1_S_SIZE; 4467 resid -= L1_S_SIZE; 4468 continue; 4469 } 4470 4471 /* 4472 * Ok, we're going to use an L2 table. Make sure 4473 * one is actually in the corresponding L1 slot 4474 * for the current VA. 4475 */ 4476 if ((pde[va >> L1_S_SHIFT] & L1_TYPE_MASK) != L1_TYPE_C) 4477 panic("pmap_map_chunk: no L2 table for VA 0x%08lx", va); 4478 4479 #ifndef ARM32_NEW_VM_LAYOUT 4480 pte = (pt_entry_t *) 4481 kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME); 4482 #else 4483 pte = (pt_entry_t *) kernel_pt_lookup( 4484 pde[L1_IDX(va)] & L1_C_ADDR_MASK); 4485 #endif 4486 if (pte == NULL) 4487 panic("pmap_map_chunk: can't find L2 table for VA" 4488 "0x%08lx", va); 4489 4490 /* See if we can use a L2 large page mapping. */ 4491 if (L2_L_MAPPABLE_P(va, pa, resid)) { 4492 #ifdef VERBOSE_INIT_ARM 4493 printf("L"); 4494 #endif 4495 for (i = 0; i < 16; i++) { 4496 #ifndef ARM32_NEW_VM_LAYOUT 4497 pte[((va >> PGSHIFT) & 0x3f0) + i] = 4498 L2_L_PROTO | pa | 4499 L2_L_PROT(PTE_KERNEL, prot) | f2l; 4500 PTE_SYNC(&pte[((va >> PGSHIFT) & 0x3f0) + i]); 4501 #else 4502 pte[l2pte_index(va) + i] = 4503 L2_L_PROTO | pa | 4504 L2_L_PROT(PTE_KERNEL, prot) | f2l; 4505 PTE_SYNC(&pte[l2pte_index(va) + i]); 4506 #endif 4507 } 4508 va += L2_L_SIZE; 4509 pa += L2_L_SIZE; 4510 resid -= L2_L_SIZE; 4511 continue; 4512 } 4513 4514 /* Use a small page mapping. */ 4515 #ifdef VERBOSE_INIT_ARM 4516 printf("P"); 4517 #endif 4518 #ifndef ARM32_NEW_VM_LAYOUT 4519 pte[(va >> PGSHIFT) & 0x3ff] = 4520 L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | f2s; 4521 PTE_SYNC(&pte[(va >> PGSHIFT) & 0x3ff]); 4522 #else 4523 pte[l2pte_index(va)] = 4524 L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, prot) | f2s; 4525 PTE_SYNC(&pte[l2pte_index(va)]); 4526 #endif 4527 va += PAGE_SIZE; 4528 pa += PAGE_SIZE; 4529 resid -= PAGE_SIZE; 4530 } 4531 #ifdef VERBOSE_INIT_ARM 4532 printf("\n"); 4533 #endif 4534 return (size); 4535 } 4536 4537 /********************** Static device map routines ***************************/ 4538 4539 static const struct pmap_devmap *pmap_devmap_table; 4540 4541 /* 4542 * Register the devmap table. This is provided in case early console 4543 * initialization needs to register mappings created by bootstrap code 4544 * before pmap_devmap_bootstrap() is called. 4545 */ 4546 void 4547 pmap_devmap_register(const struct pmap_devmap *table) 4548 { 4549 4550 pmap_devmap_table = table; 4551 } 4552 4553 /* 4554 * Map all of the static regions in the devmap table, and remember 4555 * the devmap table so other parts of the kernel can look up entries 4556 * later. 4557 */ 4558 void 4559 pmap_devmap_bootstrap(vaddr_t l1pt, const struct pmap_devmap *table) 4560 { 4561 int i; 4562 4563 pmap_devmap_table = table; 4564 4565 for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) { 4566 #ifdef VERBOSE_INIT_ARM 4567 printf("devmap: %08lx -> %08lx @ %08lx\n", 4568 pmap_devmap_table[i].pd_pa, 4569 pmap_devmap_table[i].pd_pa + 4570 pmap_devmap_table[i].pd_size - 1, 4571 pmap_devmap_table[i].pd_va); 4572 #endif 4573 pmap_map_chunk(l1pt, pmap_devmap_table[i].pd_va, 4574 pmap_devmap_table[i].pd_pa, 4575 pmap_devmap_table[i].pd_size, 4576 pmap_devmap_table[i].pd_prot, 4577 pmap_devmap_table[i].pd_cache); 4578 } 4579 } 4580 4581 const struct pmap_devmap * 4582 pmap_devmap_find_pa(paddr_t pa, psize_t size) 4583 { 4584 uint64_t endpa; 4585 int i; 4586 4587 if (pmap_devmap_table == NULL) 4588 return (NULL); 4589 4590 endpa = (uint64_t)pa + (uint64_t)size; 4591 4592 for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) { 4593 if (pa >= pmap_devmap_table[i].pd_pa && 4594 endpa <= (uint64_t)pmap_devmap_table[i].pd_pa + 4595 (uint64_t)pmap_devmap_table[i].pd_size) 4596 return (&pmap_devmap_table[i]); 4597 } 4598 4599 return (NULL); 4600 } 4601 4602 const struct pmap_devmap * 4603 pmap_devmap_find_va(vaddr_t va, vsize_t size) 4604 { 4605 int i; 4606 4607 if (pmap_devmap_table == NULL) 4608 return (NULL); 4609 4610 for (i = 0; pmap_devmap_table[i].pd_size != 0; i++) { 4611 if (va >= pmap_devmap_table[i].pd_va && 4612 va + size <= pmap_devmap_table[i].pd_va + 4613 pmap_devmap_table[i].pd_size) 4614 return (&pmap_devmap_table[i]); 4615 } 4616 4617 return (NULL); 4618 } 4619 4620 /********************** PTE initialization routines **************************/ 4621 4622 /* 4623 * These routines are called when the CPU type is identified to set up 4624 * the PTE prototypes, cache modes, etc. 4625 * 4626 * The variables are always here, just in case LKMs need to reference 4627 * them (though, they shouldn't). 4628 */ 4629 4630 pt_entry_t pte_l1_s_cache_mode; 4631 pt_entry_t pte_l1_s_cache_mode_pt; 4632 pt_entry_t pte_l1_s_cache_mask; 4633 4634 pt_entry_t pte_l2_l_cache_mode; 4635 pt_entry_t pte_l2_l_cache_mode_pt; 4636 pt_entry_t pte_l2_l_cache_mask; 4637 4638 pt_entry_t pte_l2_s_cache_mode; 4639 pt_entry_t pte_l2_s_cache_mode_pt; 4640 pt_entry_t pte_l2_s_cache_mask; 4641 4642 pt_entry_t pte_l2_s_prot_u; 4643 pt_entry_t pte_l2_s_prot_w; 4644 pt_entry_t pte_l2_s_prot_mask; 4645 4646 pt_entry_t pte_l1_s_proto; 4647 pt_entry_t pte_l1_c_proto; 4648 pt_entry_t pte_l2_s_proto; 4649 4650 void (*pmap_copy_page_func)(paddr_t, paddr_t); 4651 void (*pmap_zero_page_func)(paddr_t); 4652 4653 #if (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 4654 void 4655 pmap_pte_init_generic(void) 4656 { 4657 4658 pte_l1_s_cache_mode = L1_S_B|L1_S_C; 4659 pte_l1_s_cache_mask = L1_S_CACHE_MASK_generic; 4660 4661 pte_l2_l_cache_mode = L2_B|L2_C; 4662 pte_l2_l_cache_mask = L2_L_CACHE_MASK_generic; 4663 4664 pte_l2_s_cache_mode = L2_B|L2_C; 4665 pte_l2_s_cache_mask = L2_S_CACHE_MASK_generic; 4666 4667 /* 4668 * If we have a write-through cache, set B and C. If 4669 * we have a write-back cache, then we assume setting 4670 * only C will make those pages write-through. 4671 */ 4672 if (cpufuncs.cf_dcache_wb_range == (void *) cpufunc_nullop) { 4673 pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C; 4674 pte_l2_l_cache_mode_pt = L2_B|L2_C; 4675 pte_l2_s_cache_mode_pt = L2_B|L2_C; 4676 } else { 4677 pte_l1_s_cache_mode_pt = L1_S_C; 4678 pte_l2_l_cache_mode_pt = L2_C; 4679 pte_l2_s_cache_mode_pt = L2_C; 4680 } 4681 4682 pte_l2_s_prot_u = L2_S_PROT_U_generic; 4683 pte_l2_s_prot_w = L2_S_PROT_W_generic; 4684 pte_l2_s_prot_mask = L2_S_PROT_MASK_generic; 4685 4686 pte_l1_s_proto = L1_S_PROTO_generic; 4687 pte_l1_c_proto = L1_C_PROTO_generic; 4688 pte_l2_s_proto = L2_S_PROTO_generic; 4689 4690 pmap_copy_page_func = pmap_copy_page_generic; 4691 pmap_zero_page_func = pmap_zero_page_generic; 4692 } 4693 4694 #if defined(CPU_ARM8) 4695 void 4696 pmap_pte_init_arm8(void) 4697 { 4698 4699 /* 4700 * ARM8 is compatible with generic, but we need to use 4701 * the page tables uncached. 4702 */ 4703 pmap_pte_init_generic(); 4704 4705 pte_l1_s_cache_mode_pt = 0; 4706 pte_l2_l_cache_mode_pt = 0; 4707 pte_l2_s_cache_mode_pt = 0; 4708 } 4709 #endif /* CPU_ARM8 */ 4710 4711 #if defined(CPU_ARM9) && defined(ARM9_CACHE_WRITE_THROUGH) 4712 void 4713 pmap_pte_init_arm9(void) 4714 { 4715 4716 /* 4717 * ARM9 is compatible with generic, but we want to use 4718 * write-through caching for now. 4719 */ 4720 pmap_pte_init_generic(); 4721 4722 pte_l1_s_cache_mode = L1_S_C; 4723 pte_l2_l_cache_mode = L2_C; 4724 pte_l2_s_cache_mode = L2_C; 4725 4726 pte_l1_s_cache_mode_pt = L1_S_C; 4727 pte_l2_l_cache_mode_pt = L2_C; 4728 pte_l2_s_cache_mode_pt = L2_C; 4729 } 4730 #endif /* CPU_ARM9 */ 4731 #endif /* (ARM_MMU_GENERIC + ARM_MMU_SA1) != 0 */ 4732 4733 #if defined(CPU_ARM10) 4734 void 4735 pmap_pte_init_arm10(void) 4736 { 4737 4738 /* 4739 * ARM10 is compatible with generic, but we want to use 4740 * write-through caching for now. 4741 */ 4742 pmap_pte_init_generic(); 4743 4744 pte_l1_s_cache_mode = L1_S_B | L1_S_C; 4745 pte_l2_l_cache_mode = L2_B | L2_C; 4746 pte_l2_s_cache_mode = L2_B | L2_C; 4747 4748 pte_l1_s_cache_mode_pt = L1_S_C; 4749 pte_l2_l_cache_mode_pt = L2_C; 4750 pte_l2_s_cache_mode_pt = L2_C; 4751 4752 } 4753 #endif /* CPU_ARM10 */ 4754 4755 #if ARM_MMU_SA1 == 1 4756 void 4757 pmap_pte_init_sa1(void) 4758 { 4759 4760 /* 4761 * The StrongARM SA-1 cache does not have a write-through 4762 * mode. So, do the generic initialization, then reset 4763 * the page table cache mode to B=1,C=1, and note that 4764 * the PTEs need to be sync'd. 4765 */ 4766 pmap_pte_init_generic(); 4767 4768 pte_l1_s_cache_mode_pt = L1_S_B|L1_S_C; 4769 pte_l2_l_cache_mode_pt = L2_B|L2_C; 4770 pte_l2_s_cache_mode_pt = L2_B|L2_C; 4771 4772 pmap_needs_pte_sync = 1; 4773 } 4774 #endif /* ARM_MMU_SA1 == 1*/ 4775 4776 #if ARM_MMU_XSCALE == 1 4777 #if (ARM_NMMUS > 1) 4778 static u_int xscale_use_minidata; 4779 #endif 4780 4781 void 4782 pmap_pte_init_xscale(void) 4783 { 4784 uint32_t auxctl; 4785 int write_through = 0; 4786 4787 pte_l1_s_cache_mode = L1_S_B|L1_S_C; 4788 pte_l1_s_cache_mask = L1_S_CACHE_MASK_xscale; 4789 4790 pte_l2_l_cache_mode = L2_B|L2_C; 4791 pte_l2_l_cache_mask = L2_L_CACHE_MASK_xscale; 4792 4793 pte_l2_s_cache_mode = L2_B|L2_C; 4794 pte_l2_s_cache_mask = L2_S_CACHE_MASK_xscale; 4795 4796 pte_l1_s_cache_mode_pt = L1_S_C; 4797 pte_l2_l_cache_mode_pt = L2_C; 4798 pte_l2_s_cache_mode_pt = L2_C; 4799 4800 #ifdef XSCALE_CACHE_READ_WRITE_ALLOCATE 4801 /* 4802 * The XScale core has an enhanced mode where writes that 4803 * miss the cache cause a cache line to be allocated. This 4804 * is significantly faster than the traditional, write-through 4805 * behavior of this case. 4806 */ 4807 pte_l1_s_cache_mode |= L1_S_XSCALE_TEX(TEX_XSCALE_X); 4808 pte_l2_l_cache_mode |= L2_XSCALE_L_TEX(TEX_XSCALE_X); 4809 pte_l2_s_cache_mode |= L2_XSCALE_T_TEX(TEX_XSCALE_X); 4810 #endif /* XSCALE_CACHE_READ_WRITE_ALLOCATE */ 4811 4812 #ifdef XSCALE_CACHE_WRITE_THROUGH 4813 /* 4814 * Some versions of the XScale core have various bugs in 4815 * their cache units, the work-around for which is to run 4816 * the cache in write-through mode. Unfortunately, this 4817 * has a major (negative) impact on performance. So, we 4818 * go ahead and run fast-and-loose, in the hopes that we 4819 * don't line up the planets in a way that will trip the 4820 * bugs. 4821 * 4822 * However, we give you the option to be slow-but-correct. 4823 */ 4824 write_through = 1; 4825 #elif defined(XSCALE_CACHE_WRITE_BACK) 4826 /* force write back cache mode */ 4827 write_through = 0; 4828 #elif defined(CPU_XSCALE_PXA250) || defined(CPU_XSCALE_PXA270) 4829 /* 4830 * Intel PXA2[15]0 processors are known to have a bug in 4831 * write-back cache on revision 4 and earlier (stepping 4832 * A[01] and B[012]). Fixed for C0 and later. 4833 */ 4834 { 4835 uint32_t id, type; 4836 4837 id = cpufunc_id(); 4838 type = id & ~(CPU_ID_XSCALE_COREREV_MASK|CPU_ID_REVISION_MASK); 4839 4840 if (type == CPU_ID_PXA250 || type == CPU_ID_PXA210) { 4841 if ((id & CPU_ID_REVISION_MASK) < 5) { 4842 /* write through for stepping A0-1 and B0-2 */ 4843 write_through = 1; 4844 } 4845 } 4846 } 4847 #endif /* XSCALE_CACHE_WRITE_THROUGH */ 4848 4849 if (write_through) { 4850 pte_l1_s_cache_mode = L1_S_C; 4851 pte_l2_l_cache_mode = L2_C; 4852 pte_l2_s_cache_mode = L2_C; 4853 } 4854 4855 #if (ARM_NMMUS > 1) 4856 xscale_use_minidata = 1; 4857 #endif 4858 4859 pte_l2_s_prot_u = L2_S_PROT_U_xscale; 4860 pte_l2_s_prot_w = L2_S_PROT_W_xscale; 4861 pte_l2_s_prot_mask = L2_S_PROT_MASK_xscale; 4862 4863 pte_l1_s_proto = L1_S_PROTO_xscale; 4864 pte_l1_c_proto = L1_C_PROTO_xscale; 4865 pte_l2_s_proto = L2_S_PROTO_xscale; 4866 4867 pmap_copy_page_func = pmap_copy_page_xscale; 4868 pmap_zero_page_func = pmap_zero_page_xscale; 4869 4870 /* 4871 * Disable ECC protection of page table access, for now. 4872 */ 4873 __asm __volatile("mrc p15, 0, %0, c1, c0, 1" : "=r" (auxctl)); 4874 auxctl &= ~XSCALE_AUXCTL_P; 4875 __asm __volatile("mcr p15, 0, %0, c1, c0, 1" : : "r" (auxctl)); 4876 } 4877 4878 /* 4879 * xscale_setup_minidata: 4880 * 4881 * Set up the mini-data cache clean area. We require the 4882 * caller to allocate the right amount of physically and 4883 * virtually contiguous space. 4884 */ 4885 void 4886 xscale_setup_minidata(vaddr_t l1pt, vaddr_t va, paddr_t pa) 4887 { 4888 extern vaddr_t xscale_minidata_clean_addr; 4889 extern vsize_t xscale_minidata_clean_size; /* already initialized */ 4890 pd_entry_t *pde = (pd_entry_t *) l1pt; 4891 pt_entry_t *pte; 4892 vsize_t size; 4893 uint32_t auxctl; 4894 4895 xscale_minidata_clean_addr = va; 4896 4897 /* Round it to page size. */ 4898 size = (xscale_minidata_clean_size + L2_S_OFFSET) & L2_S_FRAME; 4899 4900 for (; size != 0; 4901 va += L2_S_SIZE, pa += L2_S_SIZE, size -= L2_S_SIZE) { 4902 #ifndef ARM32_NEW_VM_LAYOUT 4903 pte = (pt_entry_t *) 4904 kernel_pt_lookup(pde[va >> L1_S_SHIFT] & L2_S_FRAME); 4905 #else 4906 pte = (pt_entry_t *) kernel_pt_lookup( 4907 pde[L1_IDX(va)] & L1_C_ADDR_MASK); 4908 #endif 4909 if (pte == NULL) 4910 panic("xscale_setup_minidata: can't find L2 table for " 4911 "VA 0x%08lx", va); 4912 #ifndef ARM32_NEW_VM_LAYOUT 4913 pte[(va >> PGSHIFT) & 0x3ff] = 4914 #else 4915 pte[l2pte_index(va)] = 4916 #endif 4917 L2_S_PROTO | pa | L2_S_PROT(PTE_KERNEL, VM_PROT_READ) | 4918 L2_C | L2_XSCALE_T_TEX(TEX_XSCALE_X); 4919 } 4920 4921 /* 4922 * Configure the mini-data cache for write-back with 4923 * read/write-allocate. 4924 * 4925 * NOTE: In order to reconfigure the mini-data cache, we must 4926 * make sure it contains no valid data! In order to do that, 4927 * we must issue a global data cache invalidate command! 4928 * 4929 * WE ASSUME WE ARE RUNNING UN-CACHED WHEN THIS ROUTINE IS CALLED! 4930 * THIS IS VERY IMPORTANT! 4931 */ 4932 4933 /* Invalidate data and mini-data. */ 4934 __asm __volatile("mcr p15, 0, %0, c7, c6, 0" : : "r" (0)); 4935 __asm __volatile("mrc p15, 0, %0, c1, c0, 1" : "=r" (auxctl)); 4936 auxctl = (auxctl & ~XSCALE_AUXCTL_MD_MASK) | XSCALE_AUXCTL_MD_WB_RWA; 4937 __asm __volatile("mcr p15, 0, %0, c1, c0, 1" : : "r" (auxctl)); 4938 } 4939 4940 /* 4941 * Change the PTEs for the specified kernel mappings such that they 4942 * will use the mini data cache instead of the main data cache. 4943 */ 4944 void 4945 pmap_uarea(vaddr_t va) 4946 { 4947 struct l2_bucket *l2b; 4948 pt_entry_t *ptep, *sptep, pte; 4949 vaddr_t next_bucket, eva; 4950 4951 #if (ARM_NMMUS > 1) 4952 if (xscale_use_minidata == 0) 4953 return; 4954 #endif 4955 4956 eva = va + USPACE; 4957 4958 while (va < eva) { 4959 next_bucket = L2_NEXT_BUCKET(va); 4960 if (next_bucket > eva) 4961 next_bucket = eva; 4962 4963 l2b = pmap_get_l2_bucket(pmap_kernel(), va); 4964 KDASSERT(l2b != NULL); 4965 4966 sptep = ptep = &l2b->l2b_kva[l2pte_index(va)]; 4967 4968 while (va < next_bucket) { 4969 pte = *ptep; 4970 if (!l2pte_minidata(pte)) { 4971 cpu_dcache_wbinv_range(va, PAGE_SIZE); 4972 cpu_tlb_flushD_SE(va); 4973 *ptep = pte & ~L2_B; 4974 } 4975 ptep++; 4976 va += PAGE_SIZE; 4977 } 4978 PTE_SYNC_RANGE(sptep, (u_int)(ptep - sptep)); 4979 } 4980 cpu_cpwait(); 4981 } 4982 #endif /* ARM_MMU_XSCALE == 1 */ 4983 4984 #if defined(DDB) 4985 /* 4986 * A couple of ddb-callable functions for dumping pmaps 4987 */ 4988 void pmap_dump_all(void); 4989 void pmap_dump(pmap_t); 4990 4991 void 4992 pmap_dump_all(void) 4993 { 4994 pmap_t pm; 4995 4996 LIST_FOREACH(pm, &pmap_pmaps, pm_list) { 4997 if (pm == pmap_kernel()) 4998 continue; 4999 pmap_dump(pm); 5000 printf("\n"); 5001 } 5002 } 5003 5004 static pt_entry_t ncptes[64]; 5005 static void pmap_dump_ncpg(pmap_t); 5006 5007 void 5008 pmap_dump(pmap_t pm) 5009 { 5010 struct l2_dtable *l2; 5011 struct l2_bucket *l2b; 5012 pt_entry_t *ptep, pte; 5013 vaddr_t l2_va, l2b_va, va; 5014 int i, j, k, occ, rows = 0; 5015 5016 if (pm == pmap_kernel()) 5017 printf("pmap_kernel (%p): ", pm); 5018 else 5019 printf("user pmap (%p): ", pm); 5020 5021 printf("domain %d, l1 at %p\n", pm->pm_domain, pm->pm_l1->l1_kva); 5022 5023 l2_va = 0; 5024 for (i = 0; i < L2_SIZE; i++, l2_va += 0x01000000) { 5025 l2 = pm->pm_l2[i]; 5026 5027 if (l2 == NULL || l2->l2_occupancy == 0) 5028 continue; 5029 5030 l2b_va = l2_va; 5031 for (j = 0; j < L2_BUCKET_SIZE; j++, l2b_va += 0x00100000) { 5032 l2b = &l2->l2_bucket[j]; 5033 5034 if (l2b->l2b_occupancy == 0 || l2b->l2b_kva == NULL) 5035 continue; 5036 5037 ptep = l2b->l2b_kva; 5038 5039 for (k = 0; k < 256 && ptep[k] == 0; k++) 5040 ; 5041 5042 k &= ~63; 5043 occ = l2b->l2b_occupancy; 5044 va = l2b_va + (k * 4096); 5045 for (; k < 256; k++, va += 0x1000) { 5046 char ch = ' '; 5047 if ((k % 64) == 0) { 5048 if ((rows % 8) == 0) { 5049 printf( 5050 " |0000 |8000 |10000 |18000 |20000 |28000 |30000 |38000\n"); 5051 } 5052 printf("%08lx: ", va); 5053 } 5054 5055 ncptes[k & 63] = 0; 5056 pte = ptep[k]; 5057 if (pte == 0) { 5058 ch = '.'; 5059 } else { 5060 occ--; 5061 switch (pte & 0x0c) { 5062 case 0x00: 5063 ch = 'D'; /* No cache No buff */ 5064 break; 5065 case 0x04: 5066 ch = 'B'; /* No cache buff */ 5067 break; 5068 case 0x08: 5069 if (pte & 0x40) 5070 ch = 'm'; 5071 else 5072 ch = 'C'; /* Cache No buff */ 5073 break; 5074 case 0x0c: 5075 ch = 'F'; /* Cache Buff */ 5076 break; 5077 } 5078 5079 if ((pte & L2_S_PROT_U) == L2_S_PROT_U) 5080 ch += 0x20; 5081 5082 if ((pte & 0xc) == 0) 5083 ncptes[k & 63] = pte; 5084 } 5085 5086 if ((k % 64) == 63) { 5087 rows++; 5088 printf("%c\n", ch); 5089 pmap_dump_ncpg(pm); 5090 if (occ == 0) 5091 break; 5092 } else 5093 printf("%c", ch); 5094 } 5095 } 5096 } 5097 } 5098 5099 static void 5100 pmap_dump_ncpg(pmap_t pm) 5101 { 5102 struct vm_page *pg; 5103 struct pv_entry *pv; 5104 int i; 5105 5106 for (i = 0; i < 63; i++) { 5107 if (ncptes[i] == 0) 5108 continue; 5109 5110 pg = PHYS_TO_VM_PAGE(l2pte_pa(ncptes[i])); 5111 if (pg == NULL) 5112 continue; 5113 5114 printf(" pa 0x%08lx: krw %d kro %d urw %d uro %d\n", 5115 VM_PAGE_TO_PHYS(pg), 5116 pg->mdpage.krw_mappings, pg->mdpage.kro_mappings, 5117 pg->mdpage.urw_mappings, pg->mdpage.uro_mappings); 5118 5119 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) { 5120 printf(" %c va 0x%08lx, flags 0x%x\n", 5121 (pm == pv->pv_pmap) ? '*' : ' ', 5122 pv->pv_va, pv->pv_flags); 5123 } 5124 } 5125 } 5126 #endif 5127