1 /* $NetBSD: pmap.c,v 1.98 2007/10/17 19:57:47 garbled Exp $ */ 2 3 /*- 4 * Copyright (c) 1996, 1997 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Jeremy Cooper. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the NetBSD 21 * Foundation, Inc. and its contributors. 22 * 4. Neither the name of The NetBSD Foundation nor the names of its 23 * contributors may be used to endorse or promote products derived 24 * from this software without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 27 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 28 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 29 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 30 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 31 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 32 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 33 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 34 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 35 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 36 * POSSIBILITY OF SUCH DAMAGE. 37 */ 38 39 /* 40 * XXX These comments aren't quite accurate. Need to change. 41 * The sun3x uses the MC68851 Memory Management Unit, which is built 42 * into the CPU. The 68851 maps virtual to physical addresses using 43 * a multi-level table lookup, which is stored in the very memory that 44 * it maps. The number of levels of lookup is configurable from one 45 * to four. In this implementation, we use three, named 'A' through 'C'. 46 * 47 * The MMU translates virtual addresses into physical addresses by 48 * traversing these tables in a process called a 'table walk'. The most 49 * significant 7 bits of the Virtual Address ('VA') being translated are 50 * used as an index into the level A table, whose base in physical memory 51 * is stored in a special MMU register, the 'CPU Root Pointer' or CRP. The 52 * address found at that index in the A table is used as the base 53 * address for the next table, the B table. The next six bits of the VA are 54 * used as an index into the B table, which in turn gives the base address 55 * of the third and final C table. 56 * 57 * The next six bits of the VA are used as an index into the C table to 58 * locate a Page Table Entry (PTE). The PTE is a physical address in memory 59 * to which the remaining 13 bits of the VA are added, producing the 60 * mapped physical address. 61 * 62 * To map the entire memory space in this manner would require 2114296 bytes 63 * of page tables per process - quite expensive. Instead we will 64 * allocate a fixed but considerably smaller space for the page tables at 65 * the time the VM system is initialized. When the pmap code is asked by 66 * the kernel to map a VA to a PA, it allocates tables as needed from this 67 * pool. When there are no more tables in the pool, tables are stolen 68 * from the oldest mapped entries in the tree. This is only possible 69 * because all memory mappings are stored in the kernel memory map 70 * structures, independent of the pmap structures. A VA which references 71 * one of these invalidated maps will cause a page fault. The kernel 72 * will determine that the page fault was caused by a task using a valid 73 * VA, but for some reason (which does not concern it), that address was 74 * not mapped. It will ask the pmap code to re-map the entry and then 75 * it will resume executing the faulting task. 76 * 77 * In this manner the most efficient use of the page table space is 78 * achieved. Tasks which do not execute often will have their tables 79 * stolen and reused by tasks which execute more frequently. The best 80 * size for the page table pool will probably be determined by 81 * experimentation. 82 * 83 * You read all of the comments so far. Good for you. 84 * Now go play! 85 */ 86 87 /*** A Note About the 68851 Address Translation Cache 88 * The MC68851 has a 64 entry cache, called the Address Translation Cache 89 * or 'ATC'. This cache stores the most recently used page descriptors 90 * accessed by the MMU when it does translations. Using a marker called a 91 * 'task alias' the MMU can store the descriptors from 8 different table 92 * spaces concurrently. The task alias is associated with the base 93 * address of the level A table of that address space. When an address 94 * space is currently active (the CRP currently points to its A table) 95 * the only cached descriptors that will be obeyed are ones which have a 96 * matching task alias of the current space associated with them. 97 * 98 * Since the cache is always consulted before any table lookups are done, 99 * it is important that it accurately reflect the state of the MMU tables. 100 * Whenever a change has been made to a table that has been loaded into 101 * the MMU, the code must be sure to flush any cached entries that are 102 * affected by the change. These instances are documented in the code at 103 * various points. 104 */ 105 /*** A Note About the Note About the 68851 Address Translation Cache 106 * 4 months into this code I discovered that the sun3x does not have 107 * a MC68851 chip. Instead, it has a version of this MMU that is part of the 108 * the 68030 CPU. 109 * All though it behaves very similarly to the 68851, it only has 1 task 110 * alias and a 22 entry cache. So sadly (or happily), the first paragraph 111 * of the previous note does not apply to the sun3x pmap. 112 */ 113 114 #include <sys/cdefs.h> 115 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.98 2007/10/17 19:57:47 garbled Exp $"); 116 117 #include "opt_ddb.h" 118 #include "opt_pmap_debug.h" 119 120 #include <sys/param.h> 121 #include <sys/systm.h> 122 #include <sys/proc.h> 123 #include <sys/malloc.h> 124 #include <sys/pool.h> 125 #include <sys/user.h> 126 #include <sys/queue.h> 127 #include <sys/kcore.h> 128 129 #include <uvm/uvm.h> 130 131 #include <machine/cpu.h> 132 #include <machine/kcore.h> 133 #include <machine/mon.h> 134 #include <machine/pmap.h> 135 #include <machine/pte.h> 136 #include <machine/vmparam.h> 137 #include <m68k/cacheops.h> 138 139 #include <sun3/sun3/cache.h> 140 #include <sun3/sun3/machdep.h> 141 142 #include "pmap_pvt.h" 143 144 /* XXX - What headers declare these? */ 145 extern struct pcb *curpcb; 146 extern int physmem; 147 148 /* Defined in locore.s */ 149 extern char kernel_text[]; 150 151 /* Defined by the linker */ 152 extern char etext[], edata[], end[]; 153 extern char *esym; /* DDB */ 154 155 /*************************** DEBUGGING DEFINITIONS *********************** 156 * Macros, preprocessor defines and variables used in debugging can make * 157 * code hard to read. Anything used exclusively for debugging purposes * 158 * is defined here to avoid having such mess scattered around the file. * 159 *************************************************************************/ 160 #ifdef PMAP_DEBUG 161 /* 162 * To aid the debugging process, macros should be expanded into smaller steps 163 * that accomplish the same goal, yet provide convenient places for placing 164 * breakpoints. When this code is compiled with PMAP_DEBUG mode defined, the 165 * 'INLINE' keyword is defined to an empty string. This way, any function 166 * defined to be a 'static INLINE' will become 'outlined' and compiled as 167 * a separate function, which is much easier to debug. 168 */ 169 #define INLINE /* nothing */ 170 171 /* 172 * It is sometimes convenient to watch the activity of a particular table 173 * in the system. The following variables are used for that purpose. 174 */ 175 a_tmgr_t *pmap_watch_atbl = 0; 176 b_tmgr_t *pmap_watch_btbl = 0; 177 c_tmgr_t *pmap_watch_ctbl = 0; 178 179 int pmap_debug = 0; 180 #define DPRINT(args) if (pmap_debug) printf args 181 182 #else /********** Stuff below is defined if NOT debugging **************/ 183 184 #define INLINE inline 185 #define DPRINT(args) /* nada */ 186 187 #endif /* PMAP_DEBUG */ 188 /*********************** END OF DEBUGGING DEFINITIONS ********************/ 189 190 /*** Management Structure - Memory Layout 191 * For every MMU table in the sun3x pmap system there must be a way to 192 * manage it; we must know which process is using it, what other tables 193 * depend on it, and whether or not it contains any locked pages. This 194 * is solved by the creation of 'table management' or 'tmgr' 195 * structures. One for each MMU table in the system. 196 * 197 * MAP OF MEMORY USED BY THE PMAP SYSTEM 198 * 199 * towards lower memory 200 * kernAbase -> +-------------------------------------------------------+ 201 * | Kernel MMU A level table | 202 * kernBbase -> +-------------------------------------------------------+ 203 * | Kernel MMU B level tables | 204 * kernCbase -> +-------------------------------------------------------+ 205 * | | 206 * | Kernel MMU C level tables | 207 * | | 208 * mmuCbase -> +-------------------------------------------------------+ 209 * | User MMU C level tables | 210 * mmuAbase -> +-------------------------------------------------------+ 211 * | | 212 * | User MMU A level tables | 213 * | | 214 * mmuBbase -> +-------------------------------------------------------+ 215 * | User MMU B level tables | 216 * tmgrAbase -> +-------------------------------------------------------+ 217 * | TMGR A level table structures | 218 * tmgrBbase -> +-------------------------------------------------------+ 219 * | TMGR B level table structures | 220 * tmgrCbase -> +-------------------------------------------------------+ 221 * | TMGR C level table structures | 222 * pvbase -> +-------------------------------------------------------+ 223 * | Physical to Virtual mapping table (list heads) | 224 * pvebase -> +-------------------------------------------------------+ 225 * | Physical to Virtual mapping table (list elements) | 226 * | | 227 * +-------------------------------------------------------+ 228 * towards higher memory 229 * 230 * For every A table in the MMU A area, there will be a corresponding 231 * a_tmgr structure in the TMGR A area. The same will be true for 232 * the B and C tables. This arrangement will make it easy to find the 233 * controling tmgr structure for any table in the system by use of 234 * (relatively) simple macros. 235 */ 236 237 /* 238 * Global variables for storing the base addresses for the areas 239 * labeled above. 240 */ 241 static vaddr_t kernAphys; 242 static mmu_long_dte_t *kernAbase; 243 static mmu_short_dte_t *kernBbase; 244 static mmu_short_pte_t *kernCbase; 245 static mmu_short_pte_t *mmuCbase; 246 static mmu_short_dte_t *mmuBbase; 247 static mmu_long_dte_t *mmuAbase; 248 static a_tmgr_t *Atmgrbase; 249 static b_tmgr_t *Btmgrbase; 250 static c_tmgr_t *Ctmgrbase; 251 static pv_t *pvbase; 252 static pv_elem_t *pvebase; 253 struct pmap kernel_pmap; 254 255 /* 256 * This holds the CRP currently loaded into the MMU. 257 */ 258 struct mmu_rootptr kernel_crp; 259 260 /* 261 * Just all around global variables. 262 */ 263 static TAILQ_HEAD(a_pool_head_struct, a_tmgr_struct) a_pool; 264 static TAILQ_HEAD(b_pool_head_struct, b_tmgr_struct) b_pool; 265 static TAILQ_HEAD(c_pool_head_struct, c_tmgr_struct) c_pool; 266 267 268 /* 269 * Flags used to mark the safety/availability of certain operations or 270 * resources. 271 */ 272 /* Safe to use pmap_bootstrap_alloc(). */ 273 static bool bootstrap_alloc_enabled = false; 274 /* Temporary virtual pages are in use */ 275 int tmp_vpages_inuse; 276 277 /* 278 * XXX: For now, retain the traditional variables that were 279 * used in the old pmap/vm interface (without NONCONTIG). 280 */ 281 /* Kernel virtual address space available: */ 282 vaddr_t virtual_avail, virtual_end; 283 /* Physical address space available: */ 284 paddr_t avail_start, avail_end; 285 286 /* This keep track of the end of the contiguously mapped range. */ 287 vaddr_t virtual_contig_end; 288 289 /* Physical address used by pmap_next_page() */ 290 paddr_t avail_next; 291 292 /* These are used by pmap_copy_page(), etc. */ 293 vaddr_t tmp_vpages[2]; 294 295 /* memory pool for pmap structures */ 296 struct pool pmap_pmap_pool; 297 298 /* 299 * The 3/80 is the only member of the sun3x family that has non-contiguous 300 * physical memory. Memory is divided into 4 banks which are physically 301 * locatable on the system board. Although the size of these banks varies 302 * with the size of memory they contain, their base addresses are 303 * permenently fixed. The following structure, which describes these 304 * banks, is initialized by pmap_bootstrap() after it reads from a similar 305 * structure provided by the ROM Monitor. 306 * 307 * For the other machines in the sun3x architecture which do have contiguous 308 * RAM, this list will have only one entry, which will describe the entire 309 * range of available memory. 310 */ 311 struct pmap_physmem_struct avail_mem[SUN3X_NPHYS_RAM_SEGS]; 312 u_int total_phys_mem; 313 314 /*************************************************************************/ 315 316 /* 317 * XXX - Should "tune" these based on statistics. 318 * 319 * My first guess about the relative numbers of these needed is 320 * based on the fact that a "typical" process will have several 321 * pages mapped at low virtual addresses (text, data, bss), then 322 * some mapped shared libraries, and then some stack pages mapped 323 * near the high end of the VA space. Each process can use only 324 * one A table, and most will use only two B tables (maybe three) 325 * and probably about four C tables. Therefore, the first guess 326 * at the relative numbers of these needed is 1:2:4 -gwr 327 * 328 * The number of C tables needed is closely related to the amount 329 * of physical memory available plus a certain amount attributable 330 * to the use of double mappings. With a few simulation statistics 331 * we can find a reasonably good estimation of this unknown value. 332 * Armed with that and the above ratios, we have a good idea of what 333 * is needed at each level. -j 334 * 335 * Note: It is not physical memory memory size, but the total mapped 336 * virtual space required by the combined working sets of all the 337 * currently _runnable_ processes. (Sleeping ones don't count.) 338 * The amount of physical memory should be irrelevant. -gwr 339 */ 340 #ifdef FIXED_NTABLES 341 #define NUM_A_TABLES 16 342 #define NUM_B_TABLES 32 343 #define NUM_C_TABLES 64 344 #else 345 unsigned int NUM_A_TABLES, NUM_B_TABLES, NUM_C_TABLES; 346 #endif /* FIXED_NTABLES */ 347 348 /* 349 * This determines our total virtual mapping capacity. 350 * Yes, it is a FIXED value so we can pre-allocate. 351 */ 352 #define NUM_USER_PTES (NUM_C_TABLES * MMU_C_TBL_SIZE) 353 354 /* 355 * The size of the Kernel Virtual Address Space (KVAS) 356 * for purposes of MMU table allocation is -KERNBASE 357 * (length from KERNBASE to 0xFFFFffff) 358 */ 359 #define KVAS_SIZE (-KERNBASE) 360 361 /* Numbers of kernel MMU tables to support KVAS_SIZE. */ 362 #define KERN_B_TABLES (KVAS_SIZE >> MMU_TIA_SHIFT) 363 #define KERN_C_TABLES (KVAS_SIZE >> MMU_TIB_SHIFT) 364 #define NUM_KERN_PTES (KVAS_SIZE >> MMU_TIC_SHIFT) 365 366 /*************************** MISCELANEOUS MACROS *************************/ 367 #define pmap_lock(pmap) simple_lock(&pmap->pm_lock) 368 #define pmap_unlock(pmap) simple_unlock(&pmap->pm_lock) 369 #define pmap_add_ref(pmap) ++pmap->pm_refcount 370 #define pmap_del_ref(pmap) --pmap->pm_refcount 371 #define pmap_refcount(pmap) pmap->pm_refcount 372 373 void *pmap_bootstrap_alloc(int); 374 375 static INLINE void *mmu_ptov(paddr_t); 376 static INLINE paddr_t mmu_vtop(void *); 377 378 #if 0 379 static INLINE a_tmgr_t *mmuA2tmgr(mmu_long_dte_t *); 380 #endif /* 0 */ 381 static INLINE b_tmgr_t *mmuB2tmgr(mmu_short_dte_t *); 382 static INLINE c_tmgr_t *mmuC2tmgr(mmu_short_pte_t *); 383 384 static INLINE pv_t *pa2pv(paddr_t); 385 static INLINE int pteidx(mmu_short_pte_t *); 386 static INLINE pmap_t current_pmap(void); 387 388 /* 389 * We can always convert between virtual and physical addresses 390 * for anything in the range [KERNBASE ... avail_start] because 391 * that range is GUARANTEED to be mapped linearly. 392 * We rely heavily upon this feature! 393 */ 394 static INLINE void * 395 mmu_ptov(paddr_t pa) 396 { 397 vaddr_t va; 398 399 va = (pa + KERNBASE); 400 #ifdef PMAP_DEBUG 401 if ((va < KERNBASE) || (va >= virtual_contig_end)) 402 panic("mmu_ptov"); 403 #endif 404 return (void *)va; 405 } 406 407 static INLINE paddr_t 408 mmu_vtop(void *vva) 409 { 410 vaddr_t va; 411 412 va = (vaddr_t)vva; 413 #ifdef PMAP_DEBUG 414 if ((va < KERNBASE) || (va >= virtual_contig_end)) 415 panic("mmu_vtop"); 416 #endif 417 return va - KERNBASE; 418 } 419 420 /* 421 * These macros map MMU tables to their corresponding manager structures. 422 * They are needed quite often because many of the pointers in the pmap 423 * system reference MMU tables and not the structures that control them. 424 * There needs to be a way to find one when given the other and these 425 * macros do so by taking advantage of the memory layout described above. 426 * Here's a quick step through the first macro, mmuA2tmgr(): 427 * 428 * 1) find the offset of the given MMU A table from the base of its table 429 * pool (table - mmuAbase). 430 * 2) convert this offset into a table index by dividing it by the 431 * size of one MMU 'A' table. (sizeof(mmu_long_dte_t) * MMU_A_TBL_SIZE) 432 * 3) use this index to select the corresponding 'A' table manager 433 * structure from the 'A' table manager pool (Atmgrbase[index]). 434 */ 435 /* This function is not currently used. */ 436 #if 0 437 static INLINE a_tmgr_t * 438 mmuA2tmgr(mmu_long_dte_t *mmuAtbl) 439 { 440 int idx; 441 442 /* Which table is this in? */ 443 idx = (mmuAtbl - mmuAbase) / MMU_A_TBL_SIZE; 444 #ifdef PMAP_DEBUG 445 if ((idx < 0) || (idx >= NUM_A_TABLES)) 446 panic("mmuA2tmgr"); 447 #endif 448 return &Atmgrbase[idx]; 449 } 450 #endif /* 0 */ 451 452 static INLINE b_tmgr_t * 453 mmuB2tmgr(mmu_short_dte_t *mmuBtbl) 454 { 455 int idx; 456 457 /* Which table is this in? */ 458 idx = (mmuBtbl - mmuBbase) / MMU_B_TBL_SIZE; 459 #ifdef PMAP_DEBUG 460 if ((idx < 0) || (idx >= NUM_B_TABLES)) 461 panic("mmuB2tmgr"); 462 #endif 463 return &Btmgrbase[idx]; 464 } 465 466 /* mmuC2tmgr INTERNAL 467 ** 468 * Given a pte known to belong to a C table, return the address of 469 * that table's management structure. 470 */ 471 static INLINE c_tmgr_t * 472 mmuC2tmgr(mmu_short_pte_t *mmuCtbl) 473 { 474 int idx; 475 476 /* Which table is this in? */ 477 idx = (mmuCtbl - mmuCbase) / MMU_C_TBL_SIZE; 478 #ifdef PMAP_DEBUG 479 if ((idx < 0) || (idx >= NUM_C_TABLES)) 480 panic("mmuC2tmgr"); 481 #endif 482 return &Ctmgrbase[idx]; 483 } 484 485 /* This is now a function call below. 486 * #define pa2pv(pa) \ 487 * (&pvbase[(unsigned long)\ 488 * m68k_btop(pa)\ 489 * ]) 490 */ 491 492 /* pa2pv INTERNAL 493 ** 494 * Return the pv_list_head element which manages the given physical 495 * address. 496 */ 497 static INLINE pv_t * 498 pa2pv(paddr_t pa) 499 { 500 struct pmap_physmem_struct *bank; 501 int idx; 502 503 bank = &avail_mem[0]; 504 while (pa >= bank->pmem_end) 505 bank = bank->pmem_next; 506 507 pa -= bank->pmem_start; 508 idx = bank->pmem_pvbase + m68k_btop(pa); 509 #ifdef PMAP_DEBUG 510 if ((idx < 0) || (idx >= physmem)) 511 panic("pa2pv"); 512 #endif 513 return &pvbase[idx]; 514 } 515 516 /* pteidx INTERNAL 517 ** 518 * Return the index of the given PTE within the entire fixed table of 519 * PTEs. 520 */ 521 static INLINE int 522 pteidx(mmu_short_pte_t *pte) 523 { 524 525 return pte - kernCbase; 526 } 527 528 /* 529 * This just offers a place to put some debugging checks, 530 * and reduces the number of places "curlwp" appears... 531 */ 532 static INLINE pmap_t 533 current_pmap(void) 534 { 535 struct vmspace *vm; 536 struct vm_map *map; 537 pmap_t pmap; 538 539 vm = curproc->p_vmspace; 540 map = &vm->vm_map; 541 pmap = vm_map_pmap(map); 542 543 return pmap; 544 } 545 546 547 /*************************** FUNCTION DEFINITIONS ************************ 548 * These appear here merely for the compiler to enforce type checking on * 549 * all function calls. * 550 *************************************************************************/ 551 552 /* 553 * Internal functions 554 */ 555 a_tmgr_t *get_a_table(void); 556 b_tmgr_t *get_b_table(void); 557 c_tmgr_t *get_c_table(void); 558 int free_a_table(a_tmgr_t *, bool); 559 int free_b_table(b_tmgr_t *, bool); 560 int free_c_table(c_tmgr_t *, bool); 561 562 void pmap_bootstrap_aalign(int); 563 void pmap_alloc_usermmu(void); 564 void pmap_alloc_usertmgr(void); 565 void pmap_alloc_pv(void); 566 void pmap_init_a_tables(void); 567 void pmap_init_b_tables(void); 568 void pmap_init_c_tables(void); 569 void pmap_init_pv(void); 570 void pmap_clear_pv(paddr_t, int); 571 static INLINE bool is_managed(paddr_t); 572 573 bool pmap_remove_a(a_tmgr_t *, vaddr_t, vaddr_t); 574 bool pmap_remove_b(b_tmgr_t *, vaddr_t, vaddr_t); 575 bool pmap_remove_c(c_tmgr_t *, vaddr_t, vaddr_t); 576 void pmap_remove_pte(mmu_short_pte_t *); 577 578 void pmap_enter_kernel(vaddr_t, paddr_t, vm_prot_t); 579 static INLINE void pmap_remove_kernel(vaddr_t, vaddr_t); 580 static INLINE void pmap_protect_kernel(vaddr_t, vaddr_t, vm_prot_t); 581 static INLINE bool pmap_extract_kernel(vaddr_t, paddr_t *); 582 vaddr_t pmap_get_pteinfo(u_int, pmap_t *, c_tmgr_t **); 583 static INLINE int pmap_dereference(pmap_t); 584 585 bool pmap_stroll(pmap_t, vaddr_t, a_tmgr_t **, b_tmgr_t **, c_tmgr_t **, 586 mmu_short_pte_t **, int *, int *, int *); 587 void pmap_bootstrap_copyprom(void); 588 void pmap_takeover_mmu(void); 589 void pmap_bootstrap_setprom(void); 590 static void pmap_page_upload(void); 591 592 #ifdef PMAP_DEBUG 593 /* Debugging function definitions */ 594 void pv_list(paddr_t, int); 595 #endif /* PMAP_DEBUG */ 596 597 /** Interface functions 598 ** - functions required by the Mach VM Pmap interface, with MACHINE_CONTIG 599 ** defined. 600 ** The new UVM doesn't require them so now INTERNAL. 601 **/ 602 static INLINE void pmap_pinit(pmap_t); 603 static INLINE void pmap_release(pmap_t); 604 605 /********************************** CODE ******************************** 606 * Functions that are called from other parts of the kernel are labeled * 607 * as 'INTERFACE' functions. Functions that are only called from * 608 * within the pmap module are labeled as 'INTERNAL' functions. * 609 * Functions that are internal, but are not (currently) used at all are * 610 * labeled 'INTERNAL_X'. * 611 ************************************************************************/ 612 613 /* pmap_bootstrap INTERNAL 614 ** 615 * Initializes the pmap system. Called at boot time from 616 * locore2.c:_vm_init() 617 * 618 * Reminder: having a pmap_bootstrap_alloc() and also having the VM 619 * system implement pmap_steal_memory() is redundant. 620 * Don't release this code without removing one or the other! 621 */ 622 void 623 pmap_bootstrap(vaddr_t nextva) 624 { 625 struct physmemory *membank; 626 struct pmap_physmem_struct *pmap_membank; 627 vaddr_t va, eva; 628 paddr_t pa; 629 int b, c, i, j; /* running table counts */ 630 int size, resvmem; 631 632 /* 633 * This function is called by __bootstrap after it has 634 * determined the type of machine and made the appropriate 635 * patches to the ROM vectors (XXX- I don't quite know what I meant 636 * by that.) It allocates and sets up enough of the pmap system 637 * to manage the kernel's address space. 638 */ 639 640 /* 641 * Determine the range of kernel virtual and physical 642 * space available. Note that we ABSOLUTELY DEPEND on 643 * the fact that the first bank of memory (4MB) is 644 * mapped linearly to KERNBASE (which we guaranteed in 645 * the first instructions of locore.s). 646 * That is plenty for our bootstrap work. 647 */ 648 virtual_avail = m68k_round_page(nextva); 649 virtual_contig_end = KERNBASE + 0x400000; /* +4MB */ 650 virtual_end = VM_MAX_KERNEL_ADDRESS; 651 /* Don't need avail_start til later. */ 652 653 /* We may now call pmap_bootstrap_alloc(). */ 654 bootstrap_alloc_enabled = true; 655 656 /* 657 * This is a somewhat unwrapped loop to deal with 658 * copying the PROM's 'phsymem' banks into the pmap's 659 * banks. The following is always assumed: 660 * 1. There is always at least one bank of memory. 661 * 2. There is always a last bank of memory, and its 662 * pmem_next member must be set to NULL. 663 */ 664 membank = romVectorPtr->v_physmemory; 665 pmap_membank = avail_mem; 666 total_phys_mem = 0; 667 668 for (;;) { /* break on !membank */ 669 pmap_membank->pmem_start = membank->address; 670 pmap_membank->pmem_end = membank->address + membank->size; 671 total_phys_mem += membank->size; 672 membank = membank->next; 673 if (!membank) 674 break; 675 /* This silly syntax arises because pmap_membank 676 * is really a pre-allocated array, but it is put into 677 * use as a linked list. 678 */ 679 pmap_membank->pmem_next = pmap_membank + 1; 680 pmap_membank = pmap_membank->pmem_next; 681 } 682 /* This is the last element. */ 683 pmap_membank->pmem_next = NULL; 684 685 /* 686 * Note: total_phys_mem, physmem represent 687 * actual physical memory, including that 688 * reserved for the PROM monitor. 689 */ 690 physmem = btoc(total_phys_mem); 691 692 /* 693 * Avail_end is set to the first byte of physical memory 694 * after the end of the last bank. We use this only to 695 * determine if a physical address is "managed" memory. 696 * This address range should be reduced to prevent the 697 * physical pages needed by the PROM monitor from being used 698 * in the VM system. 699 */ 700 resvmem = total_phys_mem - *(romVectorPtr->memoryAvail); 701 resvmem = m68k_round_page(resvmem); 702 avail_end = pmap_membank->pmem_end - resvmem; 703 704 /* 705 * First allocate enough kernel MMU tables to map all 706 * of kernel virtual space from KERNBASE to 0xFFFFFFFF. 707 * Note: All must be aligned on 256 byte boundaries. 708 * Start with the level-A table (one of those). 709 */ 710 size = sizeof(mmu_long_dte_t) * MMU_A_TBL_SIZE; 711 kernAbase = pmap_bootstrap_alloc(size); 712 memset(kernAbase, 0, size); 713 714 /* Now the level-B kernel tables... */ 715 size = sizeof(mmu_short_dte_t) * MMU_B_TBL_SIZE * KERN_B_TABLES; 716 kernBbase = pmap_bootstrap_alloc(size); 717 memset(kernBbase, 0, size); 718 719 /* Now the level-C kernel tables... */ 720 size = sizeof(mmu_short_pte_t) * MMU_C_TBL_SIZE * KERN_C_TABLES; 721 kernCbase = pmap_bootstrap_alloc(size); 722 memset(kernCbase, 0, size); 723 /* 724 * Note: In order for the PV system to work correctly, the kernel 725 * and user-level C tables must be allocated contiguously. 726 * Nothing should be allocated between here and the allocation of 727 * mmuCbase below. XXX: Should do this as one allocation, and 728 * then compute a pointer for mmuCbase instead of this... 729 * 730 * Allocate user MMU tables. 731 * These must be contiguous with the preceding. 732 */ 733 734 #ifndef FIXED_NTABLES 735 /* 736 * The number of user-level C tables that should be allocated is 737 * related to the size of physical memory. In general, there should 738 * be enough tables to map four times the amount of available RAM. 739 * The extra amount is needed because some table space is wasted by 740 * fragmentation. 741 */ 742 NUM_C_TABLES = (total_phys_mem * 4) / (MMU_C_TBL_SIZE * MMU_PAGE_SIZE); 743 NUM_B_TABLES = NUM_C_TABLES / 2; 744 NUM_A_TABLES = NUM_B_TABLES / 2; 745 #endif /* !FIXED_NTABLES */ 746 747 size = sizeof(mmu_short_pte_t) * MMU_C_TBL_SIZE * NUM_C_TABLES; 748 mmuCbase = pmap_bootstrap_alloc(size); 749 750 size = sizeof(mmu_short_dte_t) * MMU_B_TBL_SIZE * NUM_B_TABLES; 751 mmuBbase = pmap_bootstrap_alloc(size); 752 753 size = sizeof(mmu_long_dte_t) * MMU_A_TBL_SIZE * NUM_A_TABLES; 754 mmuAbase = pmap_bootstrap_alloc(size); 755 756 /* 757 * Fill in the never-changing part of the kernel tables. 758 * For simplicity, the kernel's mappings will be editable as a 759 * flat array of page table entries at kernCbase. The 760 * higher level 'A' and 'B' tables must be initialized to point 761 * to this lower one. 762 */ 763 b = c = 0; 764 765 /* 766 * Invalidate all mappings below KERNBASE in the A table. 767 * This area has already been zeroed out, but it is good 768 * practice to explicitly show that we are interpreting 769 * it as a list of A table descriptors. 770 */ 771 for (i = 0; i < MMU_TIA(KERNBASE); i++) { 772 kernAbase[i].addr.raw = 0; 773 } 774 775 /* 776 * Set up the kernel A and B tables so that they will reference the 777 * correct spots in the contiguous table of PTEs allocated for the 778 * kernel's virtual memory space. 779 */ 780 for (i = MMU_TIA(KERNBASE); i < MMU_A_TBL_SIZE; i++) { 781 kernAbase[i].attr.raw = 782 MMU_LONG_DTE_LU | MMU_LONG_DTE_SUPV | MMU_DT_SHORT; 783 kernAbase[i].addr.raw = mmu_vtop(&kernBbase[b]); 784 785 for (j = 0; j < MMU_B_TBL_SIZE; j++) { 786 kernBbase[b + j].attr.raw = 787 mmu_vtop(&kernCbase[c]) | MMU_DT_SHORT; 788 c += MMU_C_TBL_SIZE; 789 } 790 b += MMU_B_TBL_SIZE; 791 } 792 793 pmap_alloc_usermmu(); /* Allocate user MMU tables. */ 794 pmap_alloc_usertmgr(); /* Allocate user MMU table managers.*/ 795 pmap_alloc_pv(); /* Allocate physical->virtual map. */ 796 797 /* 798 * We are now done with pmap_bootstrap_alloc(). Round up 799 * `virtual_avail' to the nearest page, and set the flag 800 * to prevent use of pmap_bootstrap_alloc() hereafter. 801 */ 802 pmap_bootstrap_aalign(PAGE_SIZE); 803 bootstrap_alloc_enabled = false; 804 805 /* 806 * Now that we are done with pmap_bootstrap_alloc(), we 807 * must save the virtual and physical addresses of the 808 * end of the linearly mapped range, which are stored in 809 * virtual_contig_end and avail_start, respectively. 810 * These variables will never change after this point. 811 */ 812 virtual_contig_end = virtual_avail; 813 avail_start = virtual_avail - KERNBASE; 814 815 /* 816 * `avail_next' is a running pointer used by pmap_next_page() to 817 * keep track of the next available physical page to be handed 818 * to the VM system during its initialization, in which it 819 * asks for physical pages, one at a time. 820 */ 821 avail_next = avail_start; 822 823 /* 824 * Now allocate some virtual addresses, but not the physical pages 825 * behind them. Note that virtual_avail is already page-aligned. 826 * 827 * tmp_vpages[] is an array of two virtual pages used for temporary 828 * kernel mappings in the pmap module to facilitate various physical 829 * address-oritented operations. 830 */ 831 tmp_vpages[0] = virtual_avail; 832 virtual_avail += PAGE_SIZE; 833 tmp_vpages[1] = virtual_avail; 834 virtual_avail += PAGE_SIZE; 835 836 /** Initialize the PV system **/ 837 pmap_init_pv(); 838 839 /* 840 * Fill in the kernel_pmap structure and kernel_crp. 841 */ 842 kernAphys = mmu_vtop(kernAbase); 843 kernel_pmap.pm_a_tmgr = NULL; 844 kernel_pmap.pm_a_phys = kernAphys; 845 kernel_pmap.pm_refcount = 1; /* always in use */ 846 simple_lock_init(&kernel_pmap.pm_lock); 847 848 kernel_crp.rp_attr = MMU_LONG_DTE_LU | MMU_DT_LONG; 849 kernel_crp.rp_addr = kernAphys; 850 851 /* 852 * Now pmap_enter_kernel() may be used safely and will be 853 * the main interface used hereafter to modify the kernel's 854 * virtual address space. Note that since we are still running 855 * under the PROM's address table, none of these table modifications 856 * actually take effect until pmap_takeover_mmu() is called. 857 * 858 * Note: Our tables do NOT have the PROM linear mappings! 859 * Only the mappings created here exist in our tables, so 860 * remember to map anything we expect to use. 861 */ 862 va = (vaddr_t)KERNBASE; 863 pa = 0; 864 865 /* 866 * The first page of the kernel virtual address space is the msgbuf 867 * page. The page attributes (data, non-cached) are set here, while 868 * the address is assigned to this global pointer in cpu_startup(). 869 * It is non-cached, mostly due to paranoia. 870 */ 871 pmap_enter_kernel(va, pa|PMAP_NC, VM_PROT_ALL); 872 va += PAGE_SIZE; 873 pa += PAGE_SIZE; 874 875 /* Next page is used as the temporary stack. */ 876 pmap_enter_kernel(va, pa, VM_PROT_ALL); 877 va += PAGE_SIZE; 878 pa += PAGE_SIZE; 879 880 /* 881 * Map all of the kernel's text segment as read-only and cacheable. 882 * (Cacheable is implied by default). Unfortunately, the last bytes 883 * of kernel text and the first bytes of kernel data will often be 884 * sharing the same page. Therefore, the last page of kernel text 885 * has to be mapped as read/write, to accommodate the data. 886 */ 887 eva = m68k_trunc_page((vaddr_t)etext); 888 for (; va < eva; va += PAGE_SIZE, pa += PAGE_SIZE) 889 pmap_enter_kernel(va, pa, VM_PROT_READ|VM_PROT_EXECUTE); 890 891 /* 892 * Map all of the kernel's data as read/write and cacheable. 893 * This includes: data, BSS, symbols, and everything in the 894 * contiguous memory used by pmap_bootstrap_alloc() 895 */ 896 for (; pa < avail_start; va += PAGE_SIZE, pa += PAGE_SIZE) 897 pmap_enter_kernel(va, pa, VM_PROT_READ|VM_PROT_WRITE); 898 899 /* 900 * At this point we are almost ready to take over the MMU. But first 901 * we must save the PROM's address space in our map, as we call its 902 * routines and make references to its data later in the kernel. 903 */ 904 pmap_bootstrap_copyprom(); 905 pmap_takeover_mmu(); 906 pmap_bootstrap_setprom(); 907 908 /* Notify the VM system of our page size. */ 909 uvmexp.pagesize = PAGE_SIZE; 910 uvm_setpagesize(); 911 912 pmap_page_upload(); 913 } 914 915 916 /* pmap_alloc_usermmu INTERNAL 917 ** 918 * Called from pmap_bootstrap() to allocate MMU tables that will 919 * eventually be used for user mappings. 920 */ 921 void 922 pmap_alloc_usermmu(void) 923 { 924 925 /* XXX: Moved into caller. */ 926 } 927 928 /* pmap_alloc_pv INTERNAL 929 ** 930 * Called from pmap_bootstrap() to allocate the physical 931 * to virtual mapping list. Each physical page of memory 932 * in the system has a corresponding element in this list. 933 */ 934 void 935 pmap_alloc_pv(void) 936 { 937 int i; 938 unsigned int total_mem; 939 940 /* 941 * Allocate a pv_head structure for every page of physical 942 * memory that will be managed by the system. Since memory on 943 * the 3/80 is non-contiguous, we cannot arrive at a total page 944 * count by subtraction of the lowest available address from the 945 * highest, but rather we have to step through each memory 946 * bank and add the number of pages in each to the total. 947 * 948 * At this time we also initialize the offset of each bank's 949 * starting pv_head within the pv_head list so that the physical 950 * memory state routines (pmap_is_referenced(), 951 * pmap_is_modified(), et al.) can quickly find coresponding 952 * pv_heads in spite of the non-contiguity. 953 */ 954 total_mem = 0; 955 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) { 956 avail_mem[i].pmem_pvbase = m68k_btop(total_mem); 957 total_mem += avail_mem[i].pmem_end - avail_mem[i].pmem_start; 958 if (avail_mem[i].pmem_next == NULL) 959 break; 960 } 961 pvbase = (pv_t *)pmap_bootstrap_alloc(sizeof(pv_t) * 962 m68k_btop(total_phys_mem)); 963 } 964 965 /* pmap_alloc_usertmgr INTERNAL 966 ** 967 * Called from pmap_bootstrap() to allocate the structures which 968 * facilitate management of user MMU tables. Each user MMU table 969 * in the system has one such structure associated with it. 970 */ 971 void 972 pmap_alloc_usertmgr(void) 973 { 974 /* Allocate user MMU table managers */ 975 /* It would be a lot simpler to just make these BSS, but */ 976 /* we may want to change their size at boot time... -j */ 977 Atmgrbase = 978 (a_tmgr_t *)pmap_bootstrap_alloc(sizeof(a_tmgr_t) * NUM_A_TABLES); 979 Btmgrbase = 980 (b_tmgr_t *)pmap_bootstrap_alloc(sizeof(b_tmgr_t) * NUM_B_TABLES); 981 Ctmgrbase = 982 (c_tmgr_t *)pmap_bootstrap_alloc(sizeof(c_tmgr_t) * NUM_C_TABLES); 983 984 /* 985 * Allocate PV list elements for the physical to virtual 986 * mapping system. 987 */ 988 pvebase = (pv_elem_t *)pmap_bootstrap_alloc(sizeof(pv_elem_t) * 989 (NUM_USER_PTES + NUM_KERN_PTES)); 990 } 991 992 /* pmap_bootstrap_copyprom() INTERNAL 993 ** 994 * Copy the PROM mappings into our own tables. Note, we 995 * can use physical addresses until __bootstrap returns. 996 */ 997 void 998 pmap_bootstrap_copyprom(void) 999 { 1000 struct sunromvec *romp; 1001 int *mon_ctbl; 1002 mmu_short_pte_t *kpte; 1003 int i, len; 1004 1005 romp = romVectorPtr; 1006 1007 /* 1008 * Copy the mappings in SUN3X_MON_KDB_BASE...SUN3X_MONEND 1009 * Note: mon_ctbl[0] maps SUN3X_MON_KDB_BASE 1010 */ 1011 mon_ctbl = *romp->monptaddr; 1012 i = m68k_btop(SUN3X_MON_KDB_BASE - KERNBASE); 1013 kpte = &kernCbase[i]; 1014 len = m68k_btop(SUN3X_MONEND - SUN3X_MON_KDB_BASE); 1015 1016 for (i = 0; i < len; i++) { 1017 kpte[i].attr.raw = mon_ctbl[i]; 1018 } 1019 1020 /* 1021 * Copy the mappings at MON_DVMA_BASE (to the end). 1022 * Note, in here, mon_ctbl[0] maps MON_DVMA_BASE. 1023 * Actually, we only want the last page, which the 1024 * PROM has set up for use by the "ie" driver. 1025 * (The i82686 needs its SCP there.) 1026 * If we copy all the mappings, pmap_enter_kernel 1027 * may complain about finding valid PTEs that are 1028 * not recorded in our PV lists... 1029 */ 1030 mon_ctbl = *romp->shadowpteaddr; 1031 i = m68k_btop(SUN3X_MON_DVMA_BASE - KERNBASE); 1032 kpte = &kernCbase[i]; 1033 len = m68k_btop(SUN3X_MON_DVMA_SIZE); 1034 for (i = (len - 1); i < len; i++) { 1035 kpte[i].attr.raw = mon_ctbl[i]; 1036 } 1037 } 1038 1039 /* pmap_takeover_mmu INTERNAL 1040 ** 1041 * Called from pmap_bootstrap() after it has copied enough of the 1042 * PROM mappings into the kernel map so that we can use our own 1043 * MMU table. 1044 */ 1045 void 1046 pmap_takeover_mmu(void) 1047 { 1048 1049 loadcrp(&kernel_crp); 1050 } 1051 1052 /* pmap_bootstrap_setprom() INTERNAL 1053 ** 1054 * Set the PROM mappings so it can see kernel space. 1055 * Note that physical addresses are used here, which 1056 * we can get away with because this runs with the 1057 * low 1GB set for transparent translation. 1058 */ 1059 void 1060 pmap_bootstrap_setprom(void) 1061 { 1062 mmu_long_dte_t *mon_dte; 1063 extern struct mmu_rootptr mon_crp; 1064 int i; 1065 1066 mon_dte = (mmu_long_dte_t *)mon_crp.rp_addr; 1067 for (i = MMU_TIA(KERNBASE); i < MMU_TIA(KERN_END); i++) { 1068 mon_dte[i].attr.raw = kernAbase[i].attr.raw; 1069 mon_dte[i].addr.raw = kernAbase[i].addr.raw; 1070 } 1071 } 1072 1073 1074 /* pmap_init INTERFACE 1075 ** 1076 * Called at the end of vm_init() to set up the pmap system to go 1077 * into full time operation. All initialization of kernel_pmap 1078 * should be already done by now, so this should just do things 1079 * needed for user-level pmaps to work. 1080 */ 1081 void 1082 pmap_init(void) 1083 { 1084 1085 /** Initialize the manager pools **/ 1086 TAILQ_INIT(&a_pool); 1087 TAILQ_INIT(&b_pool); 1088 TAILQ_INIT(&c_pool); 1089 1090 /************************************************************** 1091 * Initialize all tmgr structures and MMU tables they manage. * 1092 **************************************************************/ 1093 /** Initialize A tables **/ 1094 pmap_init_a_tables(); 1095 /** Initialize B tables **/ 1096 pmap_init_b_tables(); 1097 /** Initialize C tables **/ 1098 pmap_init_c_tables(); 1099 1100 /** Initialize the pmap pools **/ 1101 pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl", 1102 &pool_allocator_nointr, IPL_NONE); 1103 } 1104 1105 /* pmap_init_a_tables() INTERNAL 1106 ** 1107 * Initializes all A managers, their MMU A tables, and inserts 1108 * them into the A manager pool for use by the system. 1109 */ 1110 void 1111 pmap_init_a_tables(void) 1112 { 1113 int i; 1114 a_tmgr_t *a_tbl; 1115 1116 for (i = 0; i < NUM_A_TABLES; i++) { 1117 /* Select the next available A manager from the pool */ 1118 a_tbl = &Atmgrbase[i]; 1119 1120 /* 1121 * Clear its parent entry. Set its wired and valid 1122 * entry count to zero. 1123 */ 1124 a_tbl->at_parent = NULL; 1125 a_tbl->at_wcnt = a_tbl->at_ecnt = 0; 1126 1127 /* Assign it the next available MMU A table from the pool */ 1128 a_tbl->at_dtbl = &mmuAbase[i * MMU_A_TBL_SIZE]; 1129 1130 /* 1131 * Initialize the MMU A table with the table in the `proc0', 1132 * or kernel, mapping. This ensures that every process has 1133 * the kernel mapped in the top part of its address space. 1134 */ 1135 memcpy(a_tbl->at_dtbl, kernAbase, 1136 MMU_A_TBL_SIZE * sizeof(mmu_long_dte_t)); 1137 1138 /* 1139 * Finally, insert the manager into the A pool, 1140 * making it ready to be used by the system. 1141 */ 1142 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link); 1143 } 1144 } 1145 1146 /* pmap_init_b_tables() INTERNAL 1147 ** 1148 * Initializes all B table managers, their MMU B tables, and 1149 * inserts them into the B manager pool for use by the system. 1150 */ 1151 void 1152 pmap_init_b_tables(void) 1153 { 1154 int i, j; 1155 b_tmgr_t *b_tbl; 1156 1157 for (i = 0; i < NUM_B_TABLES; i++) { 1158 /* Select the next available B manager from the pool */ 1159 b_tbl = &Btmgrbase[i]; 1160 1161 b_tbl->bt_parent = NULL; /* clear its parent, */ 1162 b_tbl->bt_pidx = 0; /* parent index, */ 1163 b_tbl->bt_wcnt = 0; /* wired entry count, */ 1164 b_tbl->bt_ecnt = 0; /* valid entry count. */ 1165 1166 /* Assign it the next available MMU B table from the pool */ 1167 b_tbl->bt_dtbl = &mmuBbase[i * MMU_B_TBL_SIZE]; 1168 1169 /* Invalidate every descriptor in the table */ 1170 for (j = 0; j < MMU_B_TBL_SIZE; j++) 1171 b_tbl->bt_dtbl[j].attr.raw = MMU_DT_INVALID; 1172 1173 /* Insert the manager into the B pool */ 1174 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link); 1175 } 1176 } 1177 1178 /* pmap_init_c_tables() INTERNAL 1179 ** 1180 * Initializes all C table managers, their MMU C tables, and 1181 * inserts them into the C manager pool for use by the system. 1182 */ 1183 void 1184 pmap_init_c_tables(void) 1185 { 1186 int i, j; 1187 c_tmgr_t *c_tbl; 1188 1189 for (i = 0; i < NUM_C_TABLES; i++) { 1190 /* Select the next available C manager from the pool */ 1191 c_tbl = &Ctmgrbase[i]; 1192 1193 c_tbl->ct_parent = NULL; /* clear its parent, */ 1194 c_tbl->ct_pidx = 0; /* parent index, */ 1195 c_tbl->ct_wcnt = 0; /* wired entry count, */ 1196 c_tbl->ct_ecnt = 0; /* valid entry count, */ 1197 c_tbl->ct_pmap = NULL; /* parent pmap, */ 1198 c_tbl->ct_va = 0; /* base of managed range */ 1199 1200 /* Assign it the next available MMU C table from the pool */ 1201 c_tbl->ct_dtbl = &mmuCbase[i * MMU_C_TBL_SIZE]; 1202 1203 for (j = 0; j < MMU_C_TBL_SIZE; j++) 1204 c_tbl->ct_dtbl[j].attr.raw = MMU_DT_INVALID; 1205 1206 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link); 1207 } 1208 } 1209 1210 /* pmap_init_pv() INTERNAL 1211 ** 1212 * Initializes the Physical to Virtual mapping system. 1213 */ 1214 void 1215 pmap_init_pv(void) 1216 { 1217 int i; 1218 1219 /* Initialize every PV head. */ 1220 for (i = 0; i < m68k_btop(total_phys_mem); i++) { 1221 pvbase[i].pv_idx = PVE_EOL; /* Indicate no mappings */ 1222 pvbase[i].pv_flags = 0; /* Zero out page flags */ 1223 } 1224 } 1225 1226 /* is_managed INTERNAL 1227 ** 1228 * Determine if the given physical address is managed by the PV system. 1229 * Note that this logic assumes that no one will ask for the status of 1230 * addresses which lie in-between the memory banks on the 3/80. If they 1231 * do so, it will falsely report that it is managed. 1232 * 1233 * Note: A "managed" address is one that was reported to the VM system as 1234 * a "usable page" during system startup. As such, the VM system expects the 1235 * pmap module to keep an accurate track of the useage of those pages. 1236 * Any page not given to the VM system at startup does not exist (as far as 1237 * the VM system is concerned) and is therefore "unmanaged." Examples are 1238 * those pages which belong to the ROM monitor and the memory allocated before 1239 * the VM system was started. 1240 */ 1241 static INLINE bool 1242 is_managed(paddr_t pa) 1243 { 1244 if (pa >= avail_start && pa < avail_end) 1245 return true; 1246 else 1247 return false; 1248 } 1249 1250 /* get_a_table INTERNAL 1251 ** 1252 * Retrieve and return a level A table for use in a user map. 1253 */ 1254 a_tmgr_t * 1255 get_a_table(void) 1256 { 1257 a_tmgr_t *tbl; 1258 pmap_t pmap; 1259 1260 /* Get the top A table in the pool */ 1261 tbl = TAILQ_FIRST(&a_pool); 1262 if (tbl == NULL) { 1263 /* 1264 * XXX - Instead of panicking here and in other get_x_table 1265 * functions, we do have the option of sleeping on the head of 1266 * the table pool. Any function which updates the table pool 1267 * would then issue a wakeup() on the head, thus waking up any 1268 * processes waiting for a table. 1269 * 1270 * Actually, the place to sleep would be when some process 1271 * asks for a "wired" mapping that would run us short of 1272 * mapping resources. This design DEPENDS on always having 1273 * some mapping resources in the pool for stealing, so we 1274 * must make sure we NEVER let the pool become empty. -gwr 1275 */ 1276 panic("get_a_table: out of A tables."); 1277 } 1278 1279 TAILQ_REMOVE(&a_pool, tbl, at_link); 1280 /* 1281 * If the table has a non-null parent pointer then it is in use. 1282 * Forcibly abduct it from its parent and clear its entries. 1283 * No re-entrancy worries here. This table would not be in the 1284 * table pool unless it was available for use. 1285 * 1286 * Note that the second argument to free_a_table() is false. This 1287 * indicates that the table should not be relinked into the A table 1288 * pool. That is a job for the function that called us. 1289 */ 1290 if (tbl->at_parent) { 1291 KASSERT(tbl->at_wcnt == 0); 1292 pmap = tbl->at_parent; 1293 free_a_table(tbl, false); 1294 pmap->pm_a_tmgr = NULL; 1295 pmap->pm_a_phys = kernAphys; 1296 } 1297 return tbl; 1298 } 1299 1300 /* get_b_table INTERNAL 1301 ** 1302 * Return a level B table for use. 1303 */ 1304 b_tmgr_t * 1305 get_b_table(void) 1306 { 1307 b_tmgr_t *tbl; 1308 1309 /* See 'get_a_table' for comments. */ 1310 tbl = TAILQ_FIRST(&b_pool); 1311 if (tbl == NULL) 1312 panic("get_b_table: out of B tables."); 1313 TAILQ_REMOVE(&b_pool, tbl, bt_link); 1314 if (tbl->bt_parent) { 1315 KASSERT(tbl->bt_wcnt == 0); 1316 tbl->bt_parent->at_dtbl[tbl->bt_pidx].attr.raw = MMU_DT_INVALID; 1317 tbl->bt_parent->at_ecnt--; 1318 free_b_table(tbl, false); 1319 } 1320 return tbl; 1321 } 1322 1323 /* get_c_table INTERNAL 1324 ** 1325 * Return a level C table for use. 1326 */ 1327 c_tmgr_t * 1328 get_c_table(void) 1329 { 1330 c_tmgr_t *tbl; 1331 1332 /* See 'get_a_table' for comments */ 1333 tbl = TAILQ_FIRST(&c_pool); 1334 if (tbl == NULL) 1335 panic("get_c_table: out of C tables."); 1336 TAILQ_REMOVE(&c_pool, tbl, ct_link); 1337 if (tbl->ct_parent) { 1338 KASSERT(tbl->ct_wcnt == 0); 1339 tbl->ct_parent->bt_dtbl[tbl->ct_pidx].attr.raw = MMU_DT_INVALID; 1340 tbl->ct_parent->bt_ecnt--; 1341 free_c_table(tbl, false); 1342 } 1343 return tbl; 1344 } 1345 1346 /* 1347 * The following 'free_table' and 'steal_table' functions are called to 1348 * detach tables from their current obligations (parents and children) and 1349 * prepare them for reuse in another mapping. 1350 * 1351 * Free_table is used when the calling function will handle the fate 1352 * of the parent table, such as returning it to the free pool when it has 1353 * no valid entries. Functions that do not want to handle this should 1354 * call steal_table, in which the parent table's descriptors and entry 1355 * count are automatically modified when this table is removed. 1356 */ 1357 1358 /* free_a_table INTERNAL 1359 ** 1360 * Unmaps the given A table and all child tables from their current 1361 * mappings. Returns the number of pages that were invalidated. 1362 * If 'relink' is true, the function will return the table to the head 1363 * of the available table pool. 1364 * 1365 * Cache note: The MC68851 will automatically flush all 1366 * descriptors derived from a given A table from its 1367 * Automatic Translation Cache (ATC) if we issue a 1368 * 'PFLUSHR' instruction with the base address of the 1369 * table. This function should do, and does so. 1370 * Note note: We are using an MC68030 - there is no 1371 * PFLUSHR. 1372 */ 1373 int 1374 free_a_table(a_tmgr_t *a_tbl, bool relink) 1375 { 1376 int i, removed_cnt; 1377 mmu_long_dte_t *dte; 1378 mmu_short_dte_t *dtbl; 1379 b_tmgr_t *b_tbl; 1380 uint8_t at_wired, bt_wired; 1381 1382 /* 1383 * Flush the ATC cache of all cached descriptors derived 1384 * from this table. 1385 * Sun3x does not use 68851's cached table feature 1386 * flush_atc_crp(mmu_vtop(a_tbl->dte)); 1387 */ 1388 1389 /* 1390 * Remove any pending cache flushes that were designated 1391 * for the pmap this A table belongs to. 1392 * a_tbl->parent->atc_flushq[0] = 0; 1393 * Not implemented in sun3x. 1394 */ 1395 1396 /* 1397 * All A tables in the system should retain a map for the 1398 * kernel. If the table contains any valid descriptors 1399 * (other than those for the kernel area), invalidate them all, 1400 * stopping short of the kernel's entries. 1401 */ 1402 removed_cnt = 0; 1403 at_wired = a_tbl->at_wcnt; 1404 if (a_tbl->at_ecnt) { 1405 dte = a_tbl->at_dtbl; 1406 for (i = 0; i < MMU_TIA(KERNBASE); i++) { 1407 /* 1408 * If a table entry points to a valid B table, free 1409 * it and its children. 1410 */ 1411 if (MMU_VALID_DT(dte[i])) { 1412 /* 1413 * The following block does several things, 1414 * from innermost expression to the 1415 * outermost: 1416 * 1) It extracts the base (cc 1996) 1417 * address of the B table pointed 1418 * to in the A table entry dte[i]. 1419 * 2) It converts this base address into 1420 * the virtual address it can be 1421 * accessed with. (all MMU tables point 1422 * to physical addresses.) 1423 * 3) It finds the corresponding manager 1424 * structure which manages this MMU table. 1425 * 4) It frees the manager structure. 1426 * (This frees the MMU table and all 1427 * child tables. See 'free_b_table' for 1428 * details.) 1429 */ 1430 dtbl = mmu_ptov(dte[i].addr.raw); 1431 b_tbl = mmuB2tmgr(dtbl); 1432 bt_wired = b_tbl->bt_wcnt; 1433 removed_cnt += free_b_table(b_tbl, true); 1434 if (bt_wired) 1435 a_tbl->at_wcnt--; 1436 dte[i].attr.raw = MMU_DT_INVALID; 1437 } 1438 } 1439 a_tbl->at_ecnt = 0; 1440 } 1441 KASSERT(a_tbl->at_wcnt == 0); 1442 1443 if (relink) { 1444 a_tbl->at_parent = NULL; 1445 if (!at_wired) 1446 TAILQ_REMOVE(&a_pool, a_tbl, at_link); 1447 TAILQ_INSERT_HEAD(&a_pool, a_tbl, at_link); 1448 } 1449 return removed_cnt; 1450 } 1451 1452 /* free_b_table INTERNAL 1453 ** 1454 * Unmaps the given B table and all its children from their current 1455 * mappings. Returns the number of pages that were invalidated. 1456 * (For comments, see 'free_a_table()'). 1457 */ 1458 int 1459 free_b_table(b_tmgr_t *b_tbl, bool relink) 1460 { 1461 int i, removed_cnt; 1462 mmu_short_dte_t *dte; 1463 mmu_short_pte_t *dtbl; 1464 c_tmgr_t *c_tbl; 1465 uint8_t bt_wired, ct_wired; 1466 1467 removed_cnt = 0; 1468 bt_wired = b_tbl->bt_wcnt; 1469 if (b_tbl->bt_ecnt) { 1470 dte = b_tbl->bt_dtbl; 1471 for (i = 0; i < MMU_B_TBL_SIZE; i++) { 1472 if (MMU_VALID_DT(dte[i])) { 1473 dtbl = mmu_ptov(MMU_DTE_PA(dte[i])); 1474 c_tbl = mmuC2tmgr(dtbl); 1475 ct_wired = c_tbl->ct_wcnt; 1476 removed_cnt += free_c_table(c_tbl, true); 1477 if (ct_wired) 1478 b_tbl->bt_wcnt--; 1479 dte[i].attr.raw = MMU_DT_INVALID; 1480 } 1481 } 1482 b_tbl->bt_ecnt = 0; 1483 } 1484 KASSERT(b_tbl->bt_wcnt == 0); 1485 1486 if (relink) { 1487 b_tbl->bt_parent = NULL; 1488 if (!bt_wired) 1489 TAILQ_REMOVE(&b_pool, b_tbl, bt_link); 1490 TAILQ_INSERT_HEAD(&b_pool, b_tbl, bt_link); 1491 } 1492 return removed_cnt; 1493 } 1494 1495 /* free_c_table INTERNAL 1496 ** 1497 * Unmaps the given C table from use and returns it to the pool for 1498 * re-use. Returns the number of pages that were invalidated. 1499 * 1500 * This function preserves any physical page modification information 1501 * contained in the page descriptors within the C table by calling 1502 * 'pmap_remove_pte().' 1503 */ 1504 int 1505 free_c_table(c_tmgr_t *c_tbl, bool relink) 1506 { 1507 mmu_short_pte_t *c_pte; 1508 int i, removed_cnt; 1509 uint8_t ct_wired; 1510 1511 removed_cnt = 0; 1512 ct_wired = c_tbl->ct_wcnt; 1513 if (c_tbl->ct_ecnt) { 1514 for (i = 0; i < MMU_C_TBL_SIZE; i++) { 1515 c_pte = &c_tbl->ct_dtbl[i]; 1516 if (MMU_VALID_DT(*c_pte)) { 1517 if (c_pte->attr.raw & MMU_SHORT_PTE_WIRED) 1518 c_tbl->ct_wcnt--; 1519 pmap_remove_pte(c_pte); 1520 removed_cnt++; 1521 } 1522 } 1523 c_tbl->ct_ecnt = 0; 1524 } 1525 KASSERT(c_tbl->ct_wcnt == 0); 1526 1527 if (relink) { 1528 c_tbl->ct_parent = NULL; 1529 if (!ct_wired) 1530 TAILQ_REMOVE(&c_pool, c_tbl, ct_link); 1531 TAILQ_INSERT_HEAD(&c_pool, c_tbl, ct_link); 1532 } 1533 return removed_cnt; 1534 } 1535 1536 1537 /* pmap_remove_pte INTERNAL 1538 ** 1539 * Unmap the given pte and preserve any page modification 1540 * information by transfering it to the pv head of the 1541 * physical page it maps to. This function does not update 1542 * any reference counts because it is assumed that the calling 1543 * function will do so. 1544 */ 1545 void 1546 pmap_remove_pte(mmu_short_pte_t *pte) 1547 { 1548 u_short pv_idx, targ_idx; 1549 paddr_t pa; 1550 pv_t *pv; 1551 1552 pa = MMU_PTE_PA(*pte); 1553 if (is_managed(pa)) { 1554 pv = pa2pv(pa); 1555 targ_idx = pteidx(pte); /* Index of PTE being removed */ 1556 1557 /* 1558 * If the PTE being removed is the first (or only) PTE in 1559 * the list of PTEs currently mapped to this page, remove the 1560 * PTE by changing the index found on the PV head. Otherwise 1561 * a linear search through the list will have to be executed 1562 * in order to find the PVE which points to the PTE being 1563 * removed, so that it may be modified to point to its new 1564 * neighbor. 1565 */ 1566 1567 pv_idx = pv->pv_idx; /* Index of first PTE in PV list */ 1568 if (pv_idx == targ_idx) { 1569 pv->pv_idx = pvebase[targ_idx].pve_next; 1570 } else { 1571 1572 /* 1573 * Find the PV element pointing to the target 1574 * element. Note: may have pv_idx==PVE_EOL 1575 */ 1576 1577 for (;;) { 1578 if (pv_idx == PVE_EOL) { 1579 goto pv_not_found; 1580 } 1581 if (pvebase[pv_idx].pve_next == targ_idx) 1582 break; 1583 pv_idx = pvebase[pv_idx].pve_next; 1584 } 1585 1586 /* 1587 * At this point, pv_idx is the index of the PV 1588 * element just before the target element in the list. 1589 * Unlink the target. 1590 */ 1591 1592 pvebase[pv_idx].pve_next = pvebase[targ_idx].pve_next; 1593 } 1594 1595 /* 1596 * Save the mod/ref bits of the pte by simply 1597 * ORing the entire pte onto the pv_flags member 1598 * of the pv structure. 1599 * There is no need to use a separate bit pattern 1600 * for usage information on the pv head than that 1601 * which is used on the MMU ptes. 1602 */ 1603 1604 pv_not_found: 1605 pv->pv_flags |= (u_short) pte->attr.raw; 1606 } 1607 pte->attr.raw = MMU_DT_INVALID; 1608 } 1609 1610 /* pmap_stroll INTERNAL 1611 ** 1612 * Retrieve the addresses of all table managers involved in the mapping of 1613 * the given virtual address. If the table walk completed successfully, 1614 * return true. If it was only partially successful, return false. 1615 * The table walk performed by this function is important to many other 1616 * functions in this module. 1617 * 1618 * Note: This function ought to be easier to read. 1619 */ 1620 bool 1621 pmap_stroll(pmap_t pmap, vaddr_t va, a_tmgr_t **a_tbl, b_tmgr_t **b_tbl, 1622 c_tmgr_t **c_tbl, mmu_short_pte_t **pte, int *a_idx, int *b_idx, 1623 int *pte_idx) 1624 { 1625 mmu_long_dte_t *a_dte; /* A: long descriptor table */ 1626 mmu_short_dte_t *b_dte; /* B: short descriptor table */ 1627 1628 if (pmap == pmap_kernel()) 1629 return false; 1630 1631 /* Does the given pmap have its own A table? */ 1632 *a_tbl = pmap->pm_a_tmgr; 1633 if (*a_tbl == NULL) 1634 return false; /* No. Return unknown. */ 1635 /* Does the A table have a valid B table 1636 * under the corresponding table entry? 1637 */ 1638 *a_idx = MMU_TIA(va); 1639 a_dte = &((*a_tbl)->at_dtbl[*a_idx]); 1640 if (!MMU_VALID_DT(*a_dte)) 1641 return false; /* No. Return unknown. */ 1642 /* Yes. Extract B table from the A table. */ 1643 *b_tbl = mmuB2tmgr(mmu_ptov(a_dte->addr.raw)); 1644 /* 1645 * Does the B table have a valid C table 1646 * under the corresponding table entry? 1647 */ 1648 *b_idx = MMU_TIB(va); 1649 b_dte = &((*b_tbl)->bt_dtbl[*b_idx]); 1650 if (!MMU_VALID_DT(*b_dte)) 1651 return false; /* No. Return unknown. */ 1652 /* Yes. Extract C table from the B table. */ 1653 *c_tbl = mmuC2tmgr(mmu_ptov(MMU_DTE_PA(*b_dte))); 1654 *pte_idx = MMU_TIC(va); 1655 *pte = &((*c_tbl)->ct_dtbl[*pte_idx]); 1656 1657 return true; 1658 } 1659 1660 /* pmap_enter INTERFACE 1661 ** 1662 * Called by the kernel to map a virtual address 1663 * to a physical address in the given process map. 1664 * 1665 * Note: this function should apply an exclusive lock 1666 * on the pmap system for its duration. (it certainly 1667 * would save my hair!!) 1668 * This function ought to be easier to read. 1669 */ 1670 int 1671 pmap_enter(pmap_t pmap, vaddr_t va, paddr_t pa, vm_prot_t prot, int flags) 1672 { 1673 bool insert, managed; /* Marks the need for PV insertion.*/ 1674 u_short nidx; /* PV list index */ 1675 int mapflags; /* Flags for the mapping (see NOTE1) */ 1676 u_int a_idx, b_idx, pte_idx; /* table indices */ 1677 a_tmgr_t *a_tbl; /* A: long descriptor table manager */ 1678 b_tmgr_t *b_tbl; /* B: short descriptor table manager */ 1679 c_tmgr_t *c_tbl; /* C: short page table manager */ 1680 mmu_long_dte_t *a_dte; /* A: long descriptor table */ 1681 mmu_short_dte_t *b_dte; /* B: short descriptor table */ 1682 mmu_short_pte_t *c_pte; /* C: short page descriptor table */ 1683 pv_t *pv; /* pv list head */ 1684 bool wired; /* is the mapping to be wired? */ 1685 enum {NONE, NEWA, NEWB, NEWC} llevel; /* used at end */ 1686 1687 if (pmap == pmap_kernel()) { 1688 pmap_enter_kernel(va, pa, prot); 1689 return 0; 1690 } 1691 1692 /* 1693 * Determine if the mapping should be wired. 1694 */ 1695 wired = ((flags & PMAP_WIRED) != 0); 1696 1697 /* 1698 * NOTE1: 1699 * 1700 * On November 13, 1999, someone changed the pmap_enter() API such 1701 * that it now accepts a 'flags' argument. This new argument 1702 * contains bit-flags for the architecture-independent (UVM) system to 1703 * use in signalling certain mapping requirements to the architecture- 1704 * dependent (pmap) system. The argument it replaces, 'wired', is now 1705 * one of the flags within it. 1706 * 1707 * In addition to flags signaled by the architecture-independent 1708 * system, parts of the architecture-dependent section of the sun3x 1709 * kernel pass their own flags in the lower, unused bits of the 1710 * physical address supplied to this function. These flags are 1711 * extracted and stored in the temporary variable 'mapflags'. 1712 * 1713 * Extract sun3x specific flags from the physical address. 1714 */ 1715 mapflags = (pa & ~MMU_PAGE_MASK); 1716 pa &= MMU_PAGE_MASK; 1717 1718 /* 1719 * Determine if the physical address being mapped is on-board RAM. 1720 * Any other area of the address space is likely to belong to a 1721 * device and hence it would be disasterous to cache its contents. 1722 */ 1723 if ((managed = is_managed(pa)) == false) 1724 mapflags |= PMAP_NC; 1725 1726 /* 1727 * For user mappings we walk along the MMU tables of the given 1728 * pmap, reaching a PTE which describes the virtual page being 1729 * mapped or changed. If any level of the walk ends in an invalid 1730 * entry, a table must be allocated and the entry must be updated 1731 * to point to it. 1732 * There is a bit of confusion as to whether this code must be 1733 * re-entrant. For now we will assume it is. To support 1734 * re-entrancy we must unlink tables from the table pool before 1735 * we assume we may use them. Tables are re-linked into the pool 1736 * when we are finished with them at the end of the function. 1737 * But I don't feel like doing that until we have proof that this 1738 * needs to be re-entrant. 1739 * 'llevel' records which tables need to be relinked. 1740 */ 1741 llevel = NONE; 1742 1743 /* 1744 * Step 1 - Retrieve the A table from the pmap. If it has no 1745 * A table, allocate a new one from the available pool. 1746 */ 1747 1748 a_tbl = pmap->pm_a_tmgr; 1749 if (a_tbl == NULL) { 1750 /* 1751 * This pmap does not currently have an A table. Allocate 1752 * a new one. 1753 */ 1754 a_tbl = get_a_table(); 1755 a_tbl->at_parent = pmap; 1756 1757 /* 1758 * Assign this new A table to the pmap, and calculate its 1759 * physical address so that loadcrp() can be used to make 1760 * the table active. 1761 */ 1762 pmap->pm_a_tmgr = a_tbl; 1763 pmap->pm_a_phys = mmu_vtop(a_tbl->at_dtbl); 1764 1765 /* 1766 * If the process receiving a new A table is the current 1767 * process, we are responsible for setting the MMU so that 1768 * it becomes the current address space. This only adds 1769 * new mappings, so no need to flush anything. 1770 */ 1771 if (pmap == current_pmap()) { 1772 kernel_crp.rp_addr = pmap->pm_a_phys; 1773 loadcrp(&kernel_crp); 1774 } 1775 1776 if (!wired) 1777 llevel = NEWA; 1778 } else { 1779 /* 1780 * Use the A table already allocated for this pmap. 1781 * Unlink it from the A table pool if necessary. 1782 */ 1783 if (wired && !a_tbl->at_wcnt) 1784 TAILQ_REMOVE(&a_pool, a_tbl, at_link); 1785 } 1786 1787 /* 1788 * Step 2 - Walk into the B table. If there is no valid B table, 1789 * allocate one. 1790 */ 1791 1792 a_idx = MMU_TIA(va); /* Calculate the TIA of the VA. */ 1793 a_dte = &a_tbl->at_dtbl[a_idx]; /* Retrieve descriptor from table */ 1794 if (MMU_VALID_DT(*a_dte)) { /* Is the descriptor valid? */ 1795 /* The descriptor is valid. Use the B table it points to. */ 1796 /************************************* 1797 * a_idx * 1798 * v * 1799 * a_tbl -> +-+-+-+-+-+-+-+-+-+-+-+- * 1800 * | | | | | | | | | | | | * 1801 * +-+-+-+-+-+-+-+-+-+-+-+- * 1802 * | * 1803 * \- b_tbl -> +-+- * 1804 * | | * 1805 * +-+- * 1806 *************************************/ 1807 b_dte = mmu_ptov(a_dte->addr.raw); 1808 b_tbl = mmuB2tmgr(b_dte); 1809 1810 /* 1811 * If the requested mapping must be wired, but this table 1812 * being used to map it is not, the table must be removed 1813 * from the available pool and its wired entry count 1814 * incremented. 1815 */ 1816 if (wired && !b_tbl->bt_wcnt) { 1817 TAILQ_REMOVE(&b_pool, b_tbl, bt_link); 1818 a_tbl->at_wcnt++; 1819 } 1820 } else { 1821 /* The descriptor is invalid. Allocate a new B table. */ 1822 b_tbl = get_b_table(); 1823 1824 /* Point the parent A table descriptor to this new B table. */ 1825 a_dte->addr.raw = mmu_vtop(b_tbl->bt_dtbl); 1826 a_dte->attr.raw = MMU_LONG_DTE_LU | MMU_DT_SHORT; 1827 a_tbl->at_ecnt++; /* Update parent's valid entry count */ 1828 1829 /* Create the necessary back references to the parent table */ 1830 b_tbl->bt_parent = a_tbl; 1831 b_tbl->bt_pidx = a_idx; 1832 1833 /* 1834 * If this table is to be wired, make sure the parent A table 1835 * wired count is updated to reflect that it has another wired 1836 * entry. 1837 */ 1838 if (wired) 1839 a_tbl->at_wcnt++; 1840 else if (llevel == NONE) 1841 llevel = NEWB; 1842 } 1843 1844 /* 1845 * Step 3 - Walk into the C table, if there is no valid C table, 1846 * allocate one. 1847 */ 1848 1849 b_idx = MMU_TIB(va); /* Calculate the TIB of the VA */ 1850 b_dte = &b_tbl->bt_dtbl[b_idx]; /* Retrieve descriptor from table */ 1851 if (MMU_VALID_DT(*b_dte)) { /* Is the descriptor valid? */ 1852 /* The descriptor is valid. Use the C table it points to. */ 1853 /************************************** 1854 * c_idx * 1855 * | v * 1856 * \- b_tbl -> +-+-+-+-+-+-+-+-+-+-+- * 1857 * | | | | | | | | | | | * 1858 * +-+-+-+-+-+-+-+-+-+-+- * 1859 * | * 1860 * \- c_tbl -> +-+-- * 1861 * | | | * 1862 * +-+-- * 1863 **************************************/ 1864 c_pte = mmu_ptov(MMU_PTE_PA(*b_dte)); 1865 c_tbl = mmuC2tmgr(c_pte); 1866 1867 /* If mapping is wired and table is not */ 1868 if (wired && !c_tbl->ct_wcnt) { 1869 TAILQ_REMOVE(&c_pool, c_tbl, ct_link); 1870 b_tbl->bt_wcnt++; 1871 } 1872 } else { 1873 /* The descriptor is invalid. Allocate a new C table. */ 1874 c_tbl = get_c_table(); 1875 1876 /* Point the parent B table descriptor to this new C table. */ 1877 b_dte->attr.raw = mmu_vtop(c_tbl->ct_dtbl); 1878 b_dte->attr.raw |= MMU_DT_SHORT; 1879 b_tbl->bt_ecnt++; /* Update parent's valid entry count */ 1880 1881 /* Create the necessary back references to the parent table */ 1882 c_tbl->ct_parent = b_tbl; 1883 c_tbl->ct_pidx = b_idx; 1884 /* 1885 * Store the pmap and base virtual managed address for faster 1886 * retrieval in the PV functions. 1887 */ 1888 c_tbl->ct_pmap = pmap; 1889 c_tbl->ct_va = (va & (MMU_TIA_MASK|MMU_TIB_MASK)); 1890 1891 /* 1892 * If this table is to be wired, make sure the parent B table 1893 * wired count is updated to reflect that it has another wired 1894 * entry. 1895 */ 1896 if (wired) 1897 b_tbl->bt_wcnt++; 1898 else if (llevel == NONE) 1899 llevel = NEWC; 1900 } 1901 1902 /* 1903 * Step 4 - Deposit a page descriptor (PTE) into the appropriate 1904 * slot of the C table, describing the PA to which the VA is mapped. 1905 */ 1906 1907 pte_idx = MMU_TIC(va); 1908 c_pte = &c_tbl->ct_dtbl[pte_idx]; 1909 if (MMU_VALID_DT(*c_pte)) { /* Is the entry currently valid? */ 1910 /* 1911 * The PTE is currently valid. This particular call 1912 * is just a synonym for one (or more) of the following 1913 * operations: 1914 * change protection of a page 1915 * change wiring status of a page 1916 * remove the mapping of a page 1917 */ 1918 1919 /* First check if this is a wiring operation. */ 1920 if (c_pte->attr.raw & MMU_SHORT_PTE_WIRED) { 1921 /* 1922 * The existing mapping is wired, so adjust wired 1923 * entry count here. If new mapping is still wired, 1924 * wired entry count will be incremented again later. 1925 */ 1926 c_tbl->ct_wcnt--; 1927 if (!wired) { 1928 /* 1929 * The mapping of this PTE is being changed 1930 * from wired to unwired. 1931 * Adjust wired entry counts in each table and 1932 * set llevel flag to put unwired tables back 1933 * into the active pool. 1934 */ 1935 if (c_tbl->ct_wcnt == 0) { 1936 llevel = NEWC; 1937 if (--b_tbl->bt_wcnt == 0) { 1938 llevel = NEWB; 1939 if (--a_tbl->at_wcnt == 0) { 1940 llevel = NEWA; 1941 } 1942 } 1943 } 1944 } 1945 } 1946 1947 /* Is the new address the same as the old? */ 1948 if (MMU_PTE_PA(*c_pte) == pa) { 1949 /* 1950 * Yes, mark that it does not need to be reinserted 1951 * into the PV list. 1952 */ 1953 insert = false; 1954 1955 /* 1956 * Clear all but the modified, referenced and wired 1957 * bits on the PTE. 1958 */ 1959 c_pte->attr.raw &= (MMU_SHORT_PTE_M 1960 | MMU_SHORT_PTE_USED | MMU_SHORT_PTE_WIRED); 1961 } else { 1962 /* No, remove the old entry */ 1963 pmap_remove_pte(c_pte); 1964 insert = true; 1965 } 1966 1967 /* 1968 * TLB flush is only necessary if modifying current map. 1969 * However, in pmap_enter(), the pmap almost always IS 1970 * the current pmap, so don't even bother to check. 1971 */ 1972 TBIS(va); 1973 } else { 1974 /* 1975 * The PTE is invalid. Increment the valid entry count in 1976 * the C table manager to reflect the addition of a new entry. 1977 */ 1978 c_tbl->ct_ecnt++; 1979 1980 /* XXX - temporarily make sure the PTE is cleared. */ 1981 c_pte->attr.raw = 0; 1982 1983 /* It will also need to be inserted into the PV list. */ 1984 insert = true; 1985 } 1986 1987 /* 1988 * If page is changing from unwired to wired status, set an unused bit 1989 * within the PTE to indicate that it is wired. Also increment the 1990 * wired entry count in the C table manager. 1991 */ 1992 if (wired) { 1993 c_pte->attr.raw |= MMU_SHORT_PTE_WIRED; 1994 c_tbl->ct_wcnt++; 1995 } 1996 1997 /* 1998 * Map the page, being careful to preserve modify/reference/wired 1999 * bits. At this point it is assumed that the PTE either has no bits 2000 * set, or if there are set bits, they are only modified, reference or 2001 * wired bits. If not, the following statement will cause erratic 2002 * behavior. 2003 */ 2004 #ifdef PMAP_DEBUG 2005 if (c_pte->attr.raw & ~(MMU_SHORT_PTE_M | 2006 MMU_SHORT_PTE_USED | MMU_SHORT_PTE_WIRED)) { 2007 printf("pmap_enter: junk left in PTE at %p\n", c_pte); 2008 Debugger(); 2009 } 2010 #endif 2011 c_pte->attr.raw |= ((u_long) pa | MMU_DT_PAGE); 2012 2013 /* 2014 * If the mapping should be read-only, set the write protect 2015 * bit in the PTE. 2016 */ 2017 if (!(prot & VM_PROT_WRITE)) 2018 c_pte->attr.raw |= MMU_SHORT_PTE_WP; 2019 2020 /* 2021 * Mark the PTE as used and/or modified as specified by the flags arg. 2022 */ 2023 if (flags & VM_PROT_ALL) { 2024 c_pte->attr.raw |= MMU_SHORT_PTE_USED; 2025 if (flags & VM_PROT_WRITE) { 2026 c_pte->attr.raw |= MMU_SHORT_PTE_M; 2027 } 2028 } 2029 2030 /* 2031 * If the mapping should be cache inhibited (indicated by the flag 2032 * bits found on the lower order of the physical address.) 2033 * mark the PTE as a cache inhibited page. 2034 */ 2035 if (mapflags & PMAP_NC) 2036 c_pte->attr.raw |= MMU_SHORT_PTE_CI; 2037 2038 /* 2039 * If the physical address being mapped is managed by the PV 2040 * system then link the pte into the list of pages mapped to that 2041 * address. 2042 */ 2043 if (insert && managed) { 2044 pv = pa2pv(pa); 2045 nidx = pteidx(c_pte); 2046 2047 pvebase[nidx].pve_next = pv->pv_idx; 2048 pv->pv_idx = nidx; 2049 } 2050 2051 /* Move any allocated or unwired tables back into the active pool. */ 2052 2053 switch (llevel) { 2054 case NEWA: 2055 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link); 2056 /* FALLTHROUGH */ 2057 case NEWB: 2058 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link); 2059 /* FALLTHROUGH */ 2060 case NEWC: 2061 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link); 2062 /* FALLTHROUGH */ 2063 default: 2064 break; 2065 } 2066 2067 return 0; 2068 } 2069 2070 /* pmap_enter_kernel INTERNAL 2071 ** 2072 * Map the given virtual address to the given physical address within the 2073 * kernel address space. This function exists because the kernel map does 2074 * not do dynamic table allocation. It consists of a contiguous array of ptes 2075 * and can be edited directly without the need to walk through any tables. 2076 * 2077 * XXX: "Danger, Will Robinson!" 2078 * Note that the kernel should never take a fault on any page 2079 * between [ KERNBASE .. virtual_avail ] and this is checked in 2080 * trap.c for kernel-mode MMU faults. This means that mappings 2081 * created in that range must be implicily wired. -gwr 2082 */ 2083 void 2084 pmap_enter_kernel(vaddr_t va, paddr_t pa, vm_prot_t prot) 2085 { 2086 bool was_valid, insert; 2087 u_short pte_idx; 2088 int flags; 2089 mmu_short_pte_t *pte; 2090 pv_t *pv; 2091 paddr_t old_pa; 2092 2093 flags = (pa & ~MMU_PAGE_MASK); 2094 pa &= MMU_PAGE_MASK; 2095 2096 if (is_managed(pa)) 2097 insert = true; 2098 else 2099 insert = false; 2100 2101 /* 2102 * Calculate the index of the PTE being modified. 2103 */ 2104 pte_idx = (u_long)m68k_btop(va - KERNBASE); 2105 2106 /* This array is traditionally named "Sysmap" */ 2107 pte = &kernCbase[pte_idx]; 2108 2109 if (MMU_VALID_DT(*pte)) { 2110 was_valid = true; 2111 /* 2112 * If the PTE already maps a different 2113 * physical address, umap and pv_unlink. 2114 */ 2115 old_pa = MMU_PTE_PA(*pte); 2116 if (pa != old_pa) 2117 pmap_remove_pte(pte); 2118 else { 2119 /* 2120 * Old PA and new PA are the same. No need to 2121 * relink the mapping within the PV list. 2122 */ 2123 insert = false; 2124 2125 /* 2126 * Save any mod/ref bits on the PTE. 2127 */ 2128 pte->attr.raw &= (MMU_SHORT_PTE_USED|MMU_SHORT_PTE_M); 2129 } 2130 } else { 2131 pte->attr.raw = MMU_DT_INVALID; 2132 was_valid = false; 2133 } 2134 2135 /* 2136 * Map the page. Being careful to preserve modified/referenced bits 2137 * on the PTE. 2138 */ 2139 pte->attr.raw |= (pa | MMU_DT_PAGE); 2140 2141 if (!(prot & VM_PROT_WRITE)) /* If access should be read-only */ 2142 pte->attr.raw |= MMU_SHORT_PTE_WP; 2143 if (flags & PMAP_NC) 2144 pte->attr.raw |= MMU_SHORT_PTE_CI; 2145 if (was_valid) 2146 TBIS(va); 2147 2148 /* 2149 * Insert the PTE into the PV system, if need be. 2150 */ 2151 if (insert) { 2152 pv = pa2pv(pa); 2153 pvebase[pte_idx].pve_next = pv->pv_idx; 2154 pv->pv_idx = pte_idx; 2155 } 2156 } 2157 2158 void 2159 pmap_kenter_pa(vaddr_t va, paddr_t pa, vm_prot_t prot) 2160 { 2161 mmu_short_pte_t *pte; 2162 2163 /* This array is traditionally named "Sysmap" */ 2164 pte = &kernCbase[(u_long)m68k_btop(va - KERNBASE)]; 2165 2166 KASSERT(!MMU_VALID_DT(*pte)); 2167 pte->attr.raw = MMU_DT_INVALID | MMU_DT_PAGE | (pa & MMU_PAGE_MASK); 2168 if (!(prot & VM_PROT_WRITE)) 2169 pte->attr.raw |= MMU_SHORT_PTE_WP; 2170 } 2171 2172 void 2173 pmap_kremove(vaddr_t va, vsize_t len) 2174 { 2175 int idx, eidx; 2176 2177 #ifdef PMAP_DEBUG 2178 if ((va & PGOFSET) || (len & PGOFSET)) 2179 panic("pmap_kremove: alignment"); 2180 #endif 2181 2182 idx = m68k_btop(va - KERNBASE); 2183 eidx = m68k_btop(va + len - KERNBASE); 2184 2185 while (idx < eidx) { 2186 kernCbase[idx++].attr.raw = MMU_DT_INVALID; 2187 TBIS(va); 2188 va += PAGE_SIZE; 2189 } 2190 } 2191 2192 /* pmap_map INTERNAL 2193 ** 2194 * Map a contiguous range of physical memory into a contiguous range of 2195 * the kernel virtual address space. 2196 * 2197 * Used for device mappings and early mapping of the kernel text/data/bss. 2198 * Returns the first virtual address beyond the end of the range. 2199 */ 2200 vaddr_t 2201 pmap_map(vaddr_t va, paddr_t pa, paddr_t endpa, int prot) 2202 { 2203 int sz; 2204 2205 sz = endpa - pa; 2206 do { 2207 pmap_enter_kernel(va, pa, prot); 2208 va += PAGE_SIZE; 2209 pa += PAGE_SIZE; 2210 sz -= PAGE_SIZE; 2211 } while (sz > 0); 2212 pmap_update(pmap_kernel()); 2213 return va; 2214 } 2215 2216 /* pmap_protect_kernel INTERNAL 2217 ** 2218 * Apply the given protection code to a kernel address range. 2219 */ 2220 static INLINE void 2221 pmap_protect_kernel(vaddr_t startva, vaddr_t endva, vm_prot_t prot) 2222 { 2223 vaddr_t va; 2224 mmu_short_pte_t *pte; 2225 2226 pte = &kernCbase[(unsigned long) m68k_btop(startva - KERNBASE)]; 2227 for (va = startva; va < endva; va += PAGE_SIZE, pte++) { 2228 if (MMU_VALID_DT(*pte)) { 2229 switch (prot) { 2230 case VM_PROT_ALL: 2231 break; 2232 case VM_PROT_EXECUTE: 2233 case VM_PROT_READ: 2234 case VM_PROT_READ|VM_PROT_EXECUTE: 2235 pte->attr.raw |= MMU_SHORT_PTE_WP; 2236 break; 2237 case VM_PROT_NONE: 2238 /* this is an alias for 'pmap_remove_kernel' */ 2239 pmap_remove_pte(pte); 2240 break; 2241 default: 2242 break; 2243 } 2244 /* 2245 * since this is the kernel, immediately flush any cached 2246 * descriptors for this address. 2247 */ 2248 TBIS(va); 2249 } 2250 } 2251 } 2252 2253 /* pmap_protect INTERFACE 2254 ** 2255 * Apply the given protection to the given virtual address range within 2256 * the given map. 2257 * 2258 * It is ok for the protection applied to be stronger than what is 2259 * specified. We use this to our advantage when the given map has no 2260 * mapping for the virtual address. By skipping a page when this 2261 * is discovered, we are effectively applying a protection of VM_PROT_NONE, 2262 * and therefore do not need to map the page just to apply a protection 2263 * code. Only pmap_enter() needs to create new mappings if they do not exist. 2264 * 2265 * XXX - This function could be speeded up by using pmap_stroll() for inital 2266 * setup, and then manual scrolling in the for() loop. 2267 */ 2268 void 2269 pmap_protect(pmap_t pmap, vaddr_t startva, vaddr_t endva, vm_prot_t prot) 2270 { 2271 bool iscurpmap; 2272 int a_idx, b_idx, c_idx; 2273 a_tmgr_t *a_tbl; 2274 b_tmgr_t *b_tbl; 2275 c_tmgr_t *c_tbl; 2276 mmu_short_pte_t *pte; 2277 2278 if (pmap == pmap_kernel()) { 2279 pmap_protect_kernel(startva, endva, prot); 2280 return; 2281 } 2282 2283 /* 2284 * In this particular pmap implementation, there are only three 2285 * types of memory protection: 'all' (read/write/execute), 2286 * 'read-only' (read/execute) and 'none' (no mapping.) 2287 * It is not possible for us to treat 'executable' as a separate 2288 * protection type. Therefore, protection requests that seek to 2289 * remove execute permission while retaining read or write, and those 2290 * that make little sense (write-only for example) are ignored. 2291 */ 2292 switch (prot) { 2293 case VM_PROT_NONE: 2294 /* 2295 * A request to apply the protection code of 2296 * 'VM_PROT_NONE' is a synonym for pmap_remove(). 2297 */ 2298 pmap_remove(pmap, startva, endva); 2299 return; 2300 case VM_PROT_EXECUTE: 2301 case VM_PROT_READ: 2302 case VM_PROT_READ|VM_PROT_EXECUTE: 2303 /* continue */ 2304 break; 2305 case VM_PROT_WRITE: 2306 case VM_PROT_WRITE|VM_PROT_READ: 2307 case VM_PROT_WRITE|VM_PROT_EXECUTE: 2308 case VM_PROT_ALL: 2309 /* None of these should happen in a sane system. */ 2310 return; 2311 } 2312 2313 /* 2314 * If the pmap has no A table, it has no mappings and therefore 2315 * there is nothing to protect. 2316 */ 2317 if ((a_tbl = pmap->pm_a_tmgr) == NULL) 2318 return; 2319 2320 a_idx = MMU_TIA(startva); 2321 b_idx = MMU_TIB(startva); 2322 c_idx = MMU_TIC(startva); 2323 b_tbl = NULL; 2324 c_tbl = NULL; 2325 2326 iscurpmap = (pmap == current_pmap()); 2327 while (startva < endva) { 2328 if (b_tbl || MMU_VALID_DT(a_tbl->at_dtbl[a_idx])) { 2329 if (b_tbl == NULL) { 2330 b_tbl = (b_tmgr_t *) a_tbl->at_dtbl[a_idx].addr.raw; 2331 b_tbl = mmu_ptov((vaddr_t)b_tbl); 2332 b_tbl = mmuB2tmgr((mmu_short_dte_t *)b_tbl); 2333 } 2334 if (c_tbl || MMU_VALID_DT(b_tbl->bt_dtbl[b_idx])) { 2335 if (c_tbl == NULL) { 2336 c_tbl = (c_tmgr_t *) MMU_DTE_PA(b_tbl->bt_dtbl[b_idx]); 2337 c_tbl = mmu_ptov((vaddr_t)c_tbl); 2338 c_tbl = mmuC2tmgr((mmu_short_pte_t *)c_tbl); 2339 } 2340 if (MMU_VALID_DT(c_tbl->ct_dtbl[c_idx])) { 2341 pte = &c_tbl->ct_dtbl[c_idx]; 2342 /* make the mapping read-only */ 2343 pte->attr.raw |= MMU_SHORT_PTE_WP; 2344 /* 2345 * If we just modified the current address space, 2346 * flush any translations for the modified page from 2347 * the translation cache and any data from it in the 2348 * data cache. 2349 */ 2350 if (iscurpmap) 2351 TBIS(startva); 2352 } 2353 startva += PAGE_SIZE; 2354 2355 if (++c_idx >= MMU_C_TBL_SIZE) { /* exceeded C table? */ 2356 c_tbl = NULL; 2357 c_idx = 0; 2358 if (++b_idx >= MMU_B_TBL_SIZE) { /* exceeded B table? */ 2359 b_tbl = NULL; 2360 b_idx = 0; 2361 } 2362 } 2363 } else { /* C table wasn't valid */ 2364 c_tbl = NULL; 2365 c_idx = 0; 2366 startva += MMU_TIB_RANGE; 2367 if (++b_idx >= MMU_B_TBL_SIZE) { /* exceeded B table? */ 2368 b_tbl = NULL; 2369 b_idx = 0; 2370 } 2371 } /* C table */ 2372 } else { /* B table wasn't valid */ 2373 b_tbl = NULL; 2374 b_idx = 0; 2375 startva += MMU_TIA_RANGE; 2376 a_idx++; 2377 } /* B table */ 2378 } 2379 } 2380 2381 /* pmap_unwire INTERFACE 2382 ** 2383 * Clear the wired attribute of the specified page. 2384 * 2385 * This function is called from vm_fault.c to unwire 2386 * a mapping. 2387 */ 2388 void 2389 pmap_unwire(pmap_t pmap, vaddr_t va) 2390 { 2391 int a_idx, b_idx, c_idx; 2392 a_tmgr_t *a_tbl; 2393 b_tmgr_t *b_tbl; 2394 c_tmgr_t *c_tbl; 2395 mmu_short_pte_t *pte; 2396 2397 /* Kernel mappings always remain wired. */ 2398 if (pmap == pmap_kernel()) 2399 return; 2400 2401 /* 2402 * Walk through the tables. If the walk terminates without 2403 * a valid PTE then the address wasn't wired in the first place. 2404 * Return immediately. 2405 */ 2406 if (pmap_stroll(pmap, va, &a_tbl, &b_tbl, &c_tbl, &pte, &a_idx, 2407 &b_idx, &c_idx) == false) 2408 return; 2409 2410 2411 /* Is the PTE wired? If not, return. */ 2412 if (!(pte->attr.raw & MMU_SHORT_PTE_WIRED)) 2413 return; 2414 2415 /* Remove the wiring bit. */ 2416 pte->attr.raw &= ~(MMU_SHORT_PTE_WIRED); 2417 2418 /* 2419 * Decrement the wired entry count in the C table. 2420 * If it reaches zero the following things happen: 2421 * 1. The table no longer has any wired entries and is considered 2422 * unwired. 2423 * 2. It is placed on the available queue. 2424 * 3. The parent table's wired entry count is decremented. 2425 * 4. If it reaches zero, this process repeats at step 1 and 2426 * stops at after reaching the A table. 2427 */ 2428 if (--c_tbl->ct_wcnt == 0) { 2429 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link); 2430 if (--b_tbl->bt_wcnt == 0) { 2431 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link); 2432 if (--a_tbl->at_wcnt == 0) { 2433 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link); 2434 } 2435 } 2436 } 2437 } 2438 2439 /* pmap_copy INTERFACE 2440 ** 2441 * Copy the mappings of a range of addresses in one pmap, into 2442 * the destination address of another. 2443 * 2444 * This routine is advisory. Should we one day decide that MMU tables 2445 * may be shared by more than one pmap, this function should be used to 2446 * link them together. Until that day however, we do nothing. 2447 */ 2448 void 2449 pmap_copy(pmap_t pmap_a, pmap_t pmap_b, vaddr_t dst, vsize_t len, vaddr_t src) 2450 { 2451 2452 /* not implemented. */ 2453 } 2454 2455 /* pmap_copy_page INTERFACE 2456 ** 2457 * Copy the contents of one physical page into another. 2458 * 2459 * This function makes use of two virtual pages allocated in pmap_bootstrap() 2460 * to map the two specified physical pages into the kernel address space. 2461 * 2462 * Note: We could use the transparent translation registers to make the 2463 * mappings. If we do so, be sure to disable interrupts before using them. 2464 */ 2465 void 2466 pmap_copy_page(paddr_t srcpa, paddr_t dstpa) 2467 { 2468 vaddr_t srcva, dstva; 2469 int s; 2470 2471 srcva = tmp_vpages[0]; 2472 dstva = tmp_vpages[1]; 2473 2474 s = splvm(); 2475 #ifdef DIAGNOSTIC 2476 if (tmp_vpages_inuse++) 2477 panic("pmap_copy_page: temporary vpages are in use."); 2478 #endif 2479 2480 /* Map pages as non-cacheable to avoid cache polution? */ 2481 pmap_kenter_pa(srcva, srcpa, VM_PROT_READ); 2482 pmap_kenter_pa(dstva, dstpa, VM_PROT_READ | VM_PROT_WRITE); 2483 2484 /* Hand-optimized version of bcopy(src, dst, PAGE_SIZE) */ 2485 copypage((char *)srcva, (char *)dstva); 2486 2487 pmap_kremove(srcva, PAGE_SIZE); 2488 pmap_kremove(dstva, PAGE_SIZE); 2489 2490 #ifdef DIAGNOSTIC 2491 --tmp_vpages_inuse; 2492 #endif 2493 splx(s); 2494 } 2495 2496 /* pmap_zero_page INTERFACE 2497 ** 2498 * Zero the contents of the specified physical page. 2499 * 2500 * Uses one of the virtual pages allocated in pmap_boostrap() 2501 * to map the specified page into the kernel address space. 2502 */ 2503 void 2504 pmap_zero_page(paddr_t dstpa) 2505 { 2506 vaddr_t dstva; 2507 int s; 2508 2509 dstva = tmp_vpages[1]; 2510 s = splvm(); 2511 #ifdef DIAGNOSTIC 2512 if (tmp_vpages_inuse++) 2513 panic("pmap_zero_page: temporary vpages are in use."); 2514 #endif 2515 2516 /* The comments in pmap_copy_page() above apply here also. */ 2517 pmap_kenter_pa(dstva, dstpa, VM_PROT_READ | VM_PROT_WRITE); 2518 2519 /* Hand-optimized version of bzero(ptr, PAGE_SIZE) */ 2520 zeropage((char *)dstva); 2521 2522 pmap_kremove(dstva, PAGE_SIZE); 2523 #ifdef DIAGNOSTIC 2524 --tmp_vpages_inuse; 2525 #endif 2526 splx(s); 2527 } 2528 2529 /* pmap_collect INTERFACE 2530 ** 2531 * Called from the VM system when we are about to swap out 2532 * the process using this pmap. This should give up any 2533 * resources held here, including all its MMU tables. 2534 */ 2535 void 2536 pmap_collect(pmap_t pmap) 2537 { 2538 2539 /* XXX - todo... */ 2540 } 2541 2542 /* pmap_pinit INTERNAL 2543 ** 2544 * Initialize a pmap structure. 2545 */ 2546 static INLINE void 2547 pmap_pinit(pmap_t pmap) 2548 { 2549 2550 memset(pmap, 0, sizeof(struct pmap)); 2551 pmap->pm_a_tmgr = NULL; 2552 pmap->pm_a_phys = kernAphys; 2553 pmap->pm_refcount = 1; 2554 simple_lock_init(&pmap->pm_lock); 2555 } 2556 2557 /* pmap_create INTERFACE 2558 ** 2559 * Create and return a pmap structure. 2560 */ 2561 pmap_t 2562 pmap_create(void) 2563 { 2564 pmap_t pmap; 2565 2566 pmap = pool_get(&pmap_pmap_pool, PR_WAITOK); 2567 pmap_pinit(pmap); 2568 return pmap; 2569 } 2570 2571 /* pmap_release INTERNAL 2572 ** 2573 * Release any resources held by the given pmap. 2574 * 2575 * This is the reverse analog to pmap_pinit. It does not 2576 * necessarily mean for the pmap structure to be deallocated, 2577 * as in pmap_destroy. 2578 */ 2579 static INLINE void 2580 pmap_release(pmap_t pmap) 2581 { 2582 2583 /* 2584 * As long as the pmap contains no mappings, 2585 * which always should be the case whenever 2586 * this function is called, there really should 2587 * be nothing to do. 2588 */ 2589 #ifdef PMAP_DEBUG 2590 if (pmap == pmap_kernel()) 2591 panic("pmap_release: kernel pmap"); 2592 #endif 2593 /* 2594 * XXX - If this pmap has an A table, give it back. 2595 * The pmap SHOULD be empty by now, and pmap_remove 2596 * should have already given back the A table... 2597 * However, I see: pmap->pm_a_tmgr->at_ecnt == 1 2598 * at this point, which means some mapping was not 2599 * removed when it should have been. -gwr 2600 */ 2601 if (pmap->pm_a_tmgr != NULL) { 2602 /* First make sure we are not using it! */ 2603 if (kernel_crp.rp_addr == pmap->pm_a_phys) { 2604 kernel_crp.rp_addr = kernAphys; 2605 loadcrp(&kernel_crp); 2606 } 2607 #ifdef PMAP_DEBUG /* XXX - todo! */ 2608 /* XXX - Now complain... */ 2609 printf("pmap_release: still have table\n"); 2610 Debugger(); 2611 #endif 2612 free_a_table(pmap->pm_a_tmgr, true); 2613 pmap->pm_a_tmgr = NULL; 2614 pmap->pm_a_phys = kernAphys; 2615 } 2616 } 2617 2618 /* pmap_reference INTERFACE 2619 ** 2620 * Increment the reference count of a pmap. 2621 */ 2622 void 2623 pmap_reference(pmap_t pmap) 2624 { 2625 pmap_lock(pmap); 2626 pmap_add_ref(pmap); 2627 pmap_unlock(pmap); 2628 } 2629 2630 /* pmap_dereference INTERNAL 2631 ** 2632 * Decrease the reference count on the given pmap 2633 * by one and return the current count. 2634 */ 2635 static INLINE int 2636 pmap_dereference(pmap_t pmap) 2637 { 2638 int rtn; 2639 2640 pmap_lock(pmap); 2641 rtn = pmap_del_ref(pmap); 2642 pmap_unlock(pmap); 2643 2644 return rtn; 2645 } 2646 2647 /* pmap_destroy INTERFACE 2648 ** 2649 * Decrement a pmap's reference count and delete 2650 * the pmap if it becomes zero. Will be called 2651 * only after all mappings have been removed. 2652 */ 2653 void 2654 pmap_destroy(pmap_t pmap) 2655 { 2656 2657 if (pmap_dereference(pmap) == 0) { 2658 pmap_release(pmap); 2659 pool_put(&pmap_pmap_pool, pmap); 2660 } 2661 } 2662 2663 /* pmap_is_referenced INTERFACE 2664 ** 2665 * Determine if the given physical page has been 2666 * referenced (read from [or written to.]) 2667 */ 2668 bool 2669 pmap_is_referenced(struct vm_page *pg) 2670 { 2671 paddr_t pa = VM_PAGE_TO_PHYS(pg); 2672 pv_t *pv; 2673 int idx; 2674 2675 /* 2676 * Check the flags on the pv head. If they are set, 2677 * return immediately. Otherwise a search must be done. 2678 */ 2679 2680 pv = pa2pv(pa); 2681 if (pv->pv_flags & PV_FLAGS_USED) 2682 return true; 2683 2684 /* 2685 * Search through all pv elements pointing 2686 * to this page and query their reference bits 2687 */ 2688 2689 for (idx = pv->pv_idx; idx != PVE_EOL; idx = pvebase[idx].pve_next) { 2690 if (MMU_PTE_USED(kernCbase[idx])) { 2691 return true; 2692 } 2693 } 2694 return false; 2695 } 2696 2697 /* pmap_is_modified INTERFACE 2698 ** 2699 * Determine if the given physical page has been 2700 * modified (written to.) 2701 */ 2702 bool 2703 pmap_is_modified(struct vm_page *pg) 2704 { 2705 paddr_t pa = VM_PAGE_TO_PHYS(pg); 2706 pv_t *pv; 2707 int idx; 2708 2709 /* see comments in pmap_is_referenced() */ 2710 pv = pa2pv(pa); 2711 if (pv->pv_flags & PV_FLAGS_MDFY) 2712 return true; 2713 2714 for (idx = pv->pv_idx; 2715 idx != PVE_EOL; 2716 idx = pvebase[idx].pve_next) { 2717 2718 if (MMU_PTE_MODIFIED(kernCbase[idx])) { 2719 return true; 2720 } 2721 } 2722 2723 return false; 2724 } 2725 2726 /* pmap_page_protect INTERFACE 2727 ** 2728 * Applies the given protection to all mappings to the given 2729 * physical page. 2730 */ 2731 void 2732 pmap_page_protect(struct vm_page *pg, vm_prot_t prot) 2733 { 2734 paddr_t pa = VM_PAGE_TO_PHYS(pg); 2735 pv_t *pv; 2736 int idx; 2737 vaddr_t va; 2738 struct mmu_short_pte_struct *pte; 2739 c_tmgr_t *c_tbl; 2740 pmap_t pmap, curpmap; 2741 2742 curpmap = current_pmap(); 2743 pv = pa2pv(pa); 2744 2745 for (idx = pv->pv_idx; idx != PVE_EOL; idx = pvebase[idx].pve_next) { 2746 pte = &kernCbase[idx]; 2747 switch (prot) { 2748 case VM_PROT_ALL: 2749 /* do nothing */ 2750 break; 2751 case VM_PROT_EXECUTE: 2752 case VM_PROT_READ: 2753 case VM_PROT_READ|VM_PROT_EXECUTE: 2754 /* 2755 * Determine the virtual address mapped by 2756 * the PTE and flush ATC entries if necessary. 2757 */ 2758 va = pmap_get_pteinfo(idx, &pmap, &c_tbl); 2759 pte->attr.raw |= MMU_SHORT_PTE_WP; 2760 if (pmap == curpmap || pmap == pmap_kernel()) 2761 TBIS(va); 2762 break; 2763 case VM_PROT_NONE: 2764 /* Save the mod/ref bits. */ 2765 pv->pv_flags |= pte->attr.raw; 2766 /* Invalidate the PTE. */ 2767 pte->attr.raw = MMU_DT_INVALID; 2768 2769 /* 2770 * Update table counts. And flush ATC entries 2771 * if necessary. 2772 */ 2773 va = pmap_get_pteinfo(idx, &pmap, &c_tbl); 2774 2775 /* 2776 * If the PTE belongs to the kernel map, 2777 * be sure to flush the page it maps. 2778 */ 2779 if (pmap == pmap_kernel()) { 2780 TBIS(va); 2781 } else { 2782 /* 2783 * The PTE belongs to a user map. 2784 * update the entry count in the C 2785 * table to which it belongs and flush 2786 * the ATC if the mapping belongs to 2787 * the current pmap. 2788 */ 2789 c_tbl->ct_ecnt--; 2790 if (pmap == curpmap) 2791 TBIS(va); 2792 } 2793 break; 2794 default: 2795 break; 2796 } 2797 } 2798 2799 /* 2800 * If the protection code indicates that all mappings to the page 2801 * be removed, truncate the PV list to zero entries. 2802 */ 2803 if (prot == VM_PROT_NONE) 2804 pv->pv_idx = PVE_EOL; 2805 } 2806 2807 /* pmap_get_pteinfo INTERNAL 2808 ** 2809 * Called internally to find the pmap and virtual address within that 2810 * map to which the pte at the given index maps. Also includes the PTE's C 2811 * table manager. 2812 * 2813 * Returns the pmap in the argument provided, and the virtual address 2814 * by return value. 2815 */ 2816 vaddr_t 2817 pmap_get_pteinfo(u_int idx, pmap_t *pmap, c_tmgr_t **tbl) 2818 { 2819 vaddr_t va = 0; 2820 2821 /* 2822 * Determine if the PTE is a kernel PTE or a user PTE. 2823 */ 2824 if (idx >= NUM_KERN_PTES) { 2825 /* 2826 * The PTE belongs to a user mapping. 2827 */ 2828 /* XXX: Would like an inline for this to validate idx... */ 2829 *tbl = &Ctmgrbase[(idx - NUM_KERN_PTES) / MMU_C_TBL_SIZE]; 2830 2831 *pmap = (*tbl)->ct_pmap; 2832 /* 2833 * To find the va to which the PTE maps, we first take 2834 * the table's base virtual address mapping which is stored 2835 * in ct_va. We then increment this address by a page for 2836 * every slot skipped until we reach the PTE. 2837 */ 2838 va = (*tbl)->ct_va; 2839 va += m68k_ptob(idx % MMU_C_TBL_SIZE); 2840 } else { 2841 /* 2842 * The PTE belongs to the kernel map. 2843 */ 2844 *pmap = pmap_kernel(); 2845 2846 va = m68k_ptob(idx); 2847 va += KERNBASE; 2848 } 2849 2850 return va; 2851 } 2852 2853 /* pmap_clear_modify INTERFACE 2854 ** 2855 * Clear the modification bit on the page at the specified 2856 * physical address. 2857 * 2858 */ 2859 bool 2860 pmap_clear_modify(struct vm_page *pg) 2861 { 2862 paddr_t pa = VM_PAGE_TO_PHYS(pg); 2863 bool rv; 2864 2865 rv = pmap_is_modified(pg); 2866 pmap_clear_pv(pa, PV_FLAGS_MDFY); 2867 return rv; 2868 } 2869 2870 /* pmap_clear_reference INTERFACE 2871 ** 2872 * Clear the referenced bit on the page at the specified 2873 * physical address. 2874 */ 2875 bool 2876 pmap_clear_reference(struct vm_page *pg) 2877 { 2878 paddr_t pa = VM_PAGE_TO_PHYS(pg); 2879 bool rv; 2880 2881 rv = pmap_is_referenced(pg); 2882 pmap_clear_pv(pa, PV_FLAGS_USED); 2883 return rv; 2884 } 2885 2886 /* pmap_clear_pv INTERNAL 2887 ** 2888 * Clears the specified flag from the specified physical address. 2889 * (Used by pmap_clear_modify() and pmap_clear_reference().) 2890 * 2891 * Flag is one of: 2892 * PV_FLAGS_MDFY - Page modified bit. 2893 * PV_FLAGS_USED - Page used (referenced) bit. 2894 * 2895 * This routine must not only clear the flag on the pv list 2896 * head. It must also clear the bit on every pte in the pv 2897 * list associated with the address. 2898 */ 2899 void 2900 pmap_clear_pv(paddr_t pa, int flag) 2901 { 2902 pv_t *pv; 2903 int idx; 2904 vaddr_t va; 2905 pmap_t pmap; 2906 mmu_short_pte_t *pte; 2907 c_tmgr_t *c_tbl; 2908 2909 pv = pa2pv(pa); 2910 pv->pv_flags &= ~(flag); 2911 for (idx = pv->pv_idx; idx != PVE_EOL; idx = pvebase[idx].pve_next) { 2912 pte = &kernCbase[idx]; 2913 pte->attr.raw &= ~(flag); 2914 2915 /* 2916 * The MC68030 MMU will not set the modified or 2917 * referenced bits on any MMU tables for which it has 2918 * a cached descriptor with its modify bit set. To insure 2919 * that it will modify these bits on the PTE during the next 2920 * time it is written to or read from, we must flush it from 2921 * the ATC. 2922 * 2923 * Ordinarily it is only necessary to flush the descriptor 2924 * if it is used in the current address space. But since I 2925 * am not sure that there will always be a notion of 2926 * 'the current address space' when this function is called, 2927 * I will skip the test and always flush the address. It 2928 * does no harm. 2929 */ 2930 2931 va = pmap_get_pteinfo(idx, &pmap, &c_tbl); 2932 TBIS(va); 2933 } 2934 } 2935 2936 /* pmap_extract_kernel INTERNAL 2937 ** 2938 * Extract a translation from the kernel address space. 2939 */ 2940 static INLINE bool 2941 pmap_extract_kernel(vaddr_t va, paddr_t *pap) 2942 { 2943 mmu_short_pte_t *pte; 2944 2945 pte = &kernCbase[(u_int)m68k_btop(va - KERNBASE)]; 2946 if (!MMU_VALID_DT(*pte)) 2947 return false; 2948 if (pap != NULL) 2949 *pap = MMU_PTE_PA(*pte); 2950 return true; 2951 } 2952 2953 /* pmap_extract INTERFACE 2954 ** 2955 * Return the physical address mapped by the virtual address 2956 * in the specified pmap. 2957 * 2958 * Note: this function should also apply an exclusive lock 2959 * on the pmap system during its duration. 2960 */ 2961 bool 2962 pmap_extract(pmap_t pmap, vaddr_t va, paddr_t *pap) 2963 { 2964 int a_idx, b_idx, pte_idx; 2965 a_tmgr_t *a_tbl; 2966 b_tmgr_t *b_tbl; 2967 c_tmgr_t *c_tbl; 2968 mmu_short_pte_t *c_pte; 2969 2970 if (pmap == pmap_kernel()) 2971 return pmap_extract_kernel(va, pap); 2972 2973 if (pmap_stroll(pmap, va, &a_tbl, &b_tbl, &c_tbl, 2974 &c_pte, &a_idx, &b_idx, &pte_idx) == false) 2975 return false; 2976 2977 if (!MMU_VALID_DT(*c_pte)) 2978 return false; 2979 2980 if (pap != NULL) 2981 *pap = MMU_PTE_PA(*c_pte); 2982 return true; 2983 } 2984 2985 /* pmap_remove_kernel INTERNAL 2986 ** 2987 * Remove the mapping of a range of virtual addresses from the kernel map. 2988 * The arguments are already page-aligned. 2989 */ 2990 static INLINE void 2991 pmap_remove_kernel(vaddr_t sva, vaddr_t eva) 2992 { 2993 int idx, eidx; 2994 2995 #ifdef PMAP_DEBUG 2996 if ((sva & PGOFSET) || (eva & PGOFSET)) 2997 panic("pmap_remove_kernel: alignment"); 2998 #endif 2999 3000 idx = m68k_btop(sva - KERNBASE); 3001 eidx = m68k_btop(eva - KERNBASE); 3002 3003 while (idx < eidx) { 3004 pmap_remove_pte(&kernCbase[idx++]); 3005 TBIS(sva); 3006 sva += PAGE_SIZE; 3007 } 3008 } 3009 3010 /* pmap_remove INTERFACE 3011 ** 3012 * Remove the mapping of a range of virtual addresses from the given pmap. 3013 * 3014 */ 3015 void 3016 pmap_remove(pmap_t pmap, vaddr_t sva, vaddr_t eva) 3017 { 3018 3019 if (pmap == pmap_kernel()) { 3020 pmap_remove_kernel(sva, eva); 3021 return; 3022 } 3023 3024 /* 3025 * If the pmap doesn't have an A table of its own, it has no mappings 3026 * that can be removed. 3027 */ 3028 if (pmap->pm_a_tmgr == NULL) 3029 return; 3030 3031 /* 3032 * Remove the specified range from the pmap. If the function 3033 * returns true, the operation removed all the valid mappings 3034 * in the pmap and freed its A table. If this happened to the 3035 * currently loaded pmap, the MMU root pointer must be reloaded 3036 * with the default 'kernel' map. 3037 */ 3038 if (pmap_remove_a(pmap->pm_a_tmgr, sva, eva)) { 3039 if (kernel_crp.rp_addr == pmap->pm_a_phys) { 3040 kernel_crp.rp_addr = kernAphys; 3041 loadcrp(&kernel_crp); 3042 /* will do TLB flush below */ 3043 } 3044 pmap->pm_a_tmgr = NULL; 3045 pmap->pm_a_phys = kernAphys; 3046 } 3047 3048 /* 3049 * If we just modified the current address space, 3050 * make sure to flush the MMU cache. 3051 * 3052 * XXX - this could be an unecessarily large flush. 3053 * XXX - Could decide, based on the size of the VA range 3054 * to be removed, whether to flush "by pages" or "all". 3055 */ 3056 if (pmap == current_pmap()) 3057 TBIAU(); 3058 } 3059 3060 /* pmap_remove_a INTERNAL 3061 ** 3062 * This is function number one in a set of three that removes a range 3063 * of memory in the most efficient manner by removing the highest possible 3064 * tables from the memory space. This particular function attempts to remove 3065 * as many B tables as it can, delegating the remaining fragmented ranges to 3066 * pmap_remove_b(). 3067 * 3068 * If the removal operation results in an empty A table, the function returns 3069 * true. 3070 * 3071 * It's ugly but will do for now. 3072 */ 3073 bool 3074 pmap_remove_a(a_tmgr_t *a_tbl, vaddr_t sva, vaddr_t eva) 3075 { 3076 bool empty; 3077 int idx; 3078 vaddr_t nstart, nend; 3079 b_tmgr_t *b_tbl; 3080 mmu_long_dte_t *a_dte; 3081 mmu_short_dte_t *b_dte; 3082 uint8_t at_wired, bt_wired; 3083 3084 /* 3085 * The following code works with what I call a 'granularity 3086 * reduction algorithim'. A range of addresses will always have 3087 * the following properties, which are classified according to 3088 * how the range relates to the size of the current granularity 3089 * - an A table entry: 3090 * 3091 * 1 2 3 4 3092 * -+---+---+---+---+---+---+---+- 3093 * -+---+---+---+---+---+---+---+- 3094 * 3095 * A range will always start on a granularity boundary, illustrated 3096 * by '+' signs in the table above, or it will start at some point 3097 * inbetween a granularity boundary, as illustrated by point 1. 3098 * The first step in removing a range of addresses is to remove the 3099 * range between 1 and 2, the nearest granularity boundary. This 3100 * job is handled by the section of code governed by the 3101 * 'if (start < nstart)' statement. 3102 * 3103 * A range will always encompass zero or more intergral granules, 3104 * illustrated by points 2 and 3. Integral granules are easy to 3105 * remove. The removal of these granules is the second step, and 3106 * is handled by the code block 'if (nstart < nend)'. 3107 * 3108 * Lastly, a range will always end on a granularity boundary, 3109 * ill. by point 3, or it will fall just beyond one, ill. by point 3110 * 4. The last step involves removing this range and is handled by 3111 * the code block 'if (nend < end)'. 3112 */ 3113 nstart = MMU_ROUND_UP_A(sva); 3114 nend = MMU_ROUND_A(eva); 3115 3116 at_wired = a_tbl->at_wcnt; 3117 3118 if (sva < nstart) { 3119 /* 3120 * This block is executed if the range starts between 3121 * a granularity boundary. 3122 * 3123 * First find the DTE which is responsible for mapping 3124 * the start of the range. 3125 */ 3126 idx = MMU_TIA(sva); 3127 a_dte = &a_tbl->at_dtbl[idx]; 3128 3129 /* 3130 * If the DTE is valid then delegate the removal of the sub 3131 * range to pmap_remove_b(), which can remove addresses at 3132 * a finer granularity. 3133 */ 3134 if (MMU_VALID_DT(*a_dte)) { 3135 b_dte = mmu_ptov(a_dte->addr.raw); 3136 b_tbl = mmuB2tmgr(b_dte); 3137 bt_wired = b_tbl->bt_wcnt; 3138 3139 /* 3140 * The sub range to be removed starts at the start 3141 * of the full range we were asked to remove, and ends 3142 * at the greater of: 3143 * 1. The end of the full range, -or- 3144 * 2. The end of the full range, rounded down to the 3145 * nearest granularity boundary. 3146 */ 3147 if (eva < nstart) 3148 empty = pmap_remove_b(b_tbl, sva, eva); 3149 else 3150 empty = pmap_remove_b(b_tbl, sva, nstart); 3151 3152 /* 3153 * If the child table no longer has wired entries, 3154 * decrement wired entry count. 3155 */ 3156 if (bt_wired && b_tbl->bt_wcnt == 0) 3157 a_tbl->at_wcnt--; 3158 3159 /* 3160 * If the removal resulted in an empty B table, 3161 * invalidate the DTE that points to it and decrement 3162 * the valid entry count of the A table. 3163 */ 3164 if (empty) { 3165 a_dte->attr.raw = MMU_DT_INVALID; 3166 a_tbl->at_ecnt--; 3167 } 3168 } 3169 /* 3170 * If the DTE is invalid, the address range is already non- 3171 * existent and can simply be skipped. 3172 */ 3173 } 3174 if (nstart < nend) { 3175 /* 3176 * This block is executed if the range spans a whole number 3177 * multiple of granules (A table entries.) 3178 * 3179 * First find the DTE which is responsible for mapping 3180 * the start of the first granule involved. 3181 */ 3182 idx = MMU_TIA(nstart); 3183 a_dte = &a_tbl->at_dtbl[idx]; 3184 3185 /* 3186 * Remove entire sub-granules (B tables) one at a time, 3187 * until reaching the end of the range. 3188 */ 3189 for (; nstart < nend; a_dte++, nstart += MMU_TIA_RANGE) 3190 if (MMU_VALID_DT(*a_dte)) { 3191 /* 3192 * Find the B table manager for the 3193 * entry and free it. 3194 */ 3195 b_dte = mmu_ptov(a_dte->addr.raw); 3196 b_tbl = mmuB2tmgr(b_dte); 3197 bt_wired = b_tbl->bt_wcnt; 3198 3199 free_b_table(b_tbl, true); 3200 3201 /* 3202 * All child entries has been removed. 3203 * If there were any wired entries in it, 3204 * decrement wired entry count. 3205 */ 3206 if (bt_wired) 3207 a_tbl->at_wcnt--; 3208 3209 /* 3210 * Invalidate the DTE that points to the 3211 * B table and decrement the valid entry 3212 * count of the A table. 3213 */ 3214 a_dte->attr.raw = MMU_DT_INVALID; 3215 a_tbl->at_ecnt--; 3216 } 3217 } 3218 if (nend < eva) { 3219 /* 3220 * This block is executed if the range ends beyond a 3221 * granularity boundary. 3222 * 3223 * First find the DTE which is responsible for mapping 3224 * the start of the nearest (rounded down) granularity 3225 * boundary. 3226 */ 3227 idx = MMU_TIA(nend); 3228 a_dte = &a_tbl->at_dtbl[idx]; 3229 3230 /* 3231 * If the DTE is valid then delegate the removal of the sub 3232 * range to pmap_remove_b(), which can remove addresses at 3233 * a finer granularity. 3234 */ 3235 if (MMU_VALID_DT(*a_dte)) { 3236 /* 3237 * Find the B table manager for the entry 3238 * and hand it to pmap_remove_b() along with 3239 * the sub range. 3240 */ 3241 b_dte = mmu_ptov(a_dte->addr.raw); 3242 b_tbl = mmuB2tmgr(b_dte); 3243 bt_wired = b_tbl->bt_wcnt; 3244 3245 empty = pmap_remove_b(b_tbl, nend, eva); 3246 3247 /* 3248 * If the child table no longer has wired entries, 3249 * decrement wired entry count. 3250 */ 3251 if (bt_wired && b_tbl->bt_wcnt == 0) 3252 a_tbl->at_wcnt--; 3253 /* 3254 * If the removal resulted in an empty B table, 3255 * invalidate the DTE that points to it and decrement 3256 * the valid entry count of the A table. 3257 */ 3258 if (empty) { 3259 a_dte->attr.raw = MMU_DT_INVALID; 3260 a_tbl->at_ecnt--; 3261 } 3262 } 3263 } 3264 3265 /* 3266 * If there are no more entries in the A table, release it 3267 * back to the available pool and return true. 3268 */ 3269 if (a_tbl->at_ecnt == 0) { 3270 KASSERT(a_tbl->at_wcnt == 0); 3271 a_tbl->at_parent = NULL; 3272 if (!at_wired) 3273 TAILQ_REMOVE(&a_pool, a_tbl, at_link); 3274 TAILQ_INSERT_HEAD(&a_pool, a_tbl, at_link); 3275 empty = true; 3276 } else { 3277 /* 3278 * If the table doesn't have wired entries any longer 3279 * but still has unwired entries, put it back into 3280 * the available queue. 3281 */ 3282 if (at_wired && a_tbl->at_wcnt == 0) 3283 TAILQ_INSERT_TAIL(&a_pool, a_tbl, at_link); 3284 empty = false; 3285 } 3286 3287 return empty; 3288 } 3289 3290 /* pmap_remove_b INTERNAL 3291 ** 3292 * Remove a range of addresses from an address space, trying to remove entire 3293 * C tables if possible. 3294 * 3295 * If the operation results in an empty B table, the function returns true. 3296 */ 3297 bool 3298 pmap_remove_b(b_tmgr_t *b_tbl, vaddr_t sva, vaddr_t eva) 3299 { 3300 bool empty; 3301 int idx; 3302 vaddr_t nstart, nend, rstart; 3303 c_tmgr_t *c_tbl; 3304 mmu_short_dte_t *b_dte; 3305 mmu_short_pte_t *c_dte; 3306 uint8_t bt_wired, ct_wired; 3307 3308 nstart = MMU_ROUND_UP_B(sva); 3309 nend = MMU_ROUND_B(eva); 3310 3311 bt_wired = b_tbl->bt_wcnt; 3312 3313 if (sva < nstart) { 3314 idx = MMU_TIB(sva); 3315 b_dte = &b_tbl->bt_dtbl[idx]; 3316 if (MMU_VALID_DT(*b_dte)) { 3317 c_dte = mmu_ptov(MMU_DTE_PA(*b_dte)); 3318 c_tbl = mmuC2tmgr(c_dte); 3319 ct_wired = c_tbl->ct_wcnt; 3320 3321 if (eva < nstart) 3322 empty = pmap_remove_c(c_tbl, sva, eva); 3323 else 3324 empty = pmap_remove_c(c_tbl, sva, nstart); 3325 3326 /* 3327 * If the child table no longer has wired entries, 3328 * decrement wired entry count. 3329 */ 3330 if (ct_wired && c_tbl->ct_wcnt == 0) 3331 b_tbl->bt_wcnt--; 3332 3333 if (empty) { 3334 b_dte->attr.raw = MMU_DT_INVALID; 3335 b_tbl->bt_ecnt--; 3336 } 3337 } 3338 } 3339 if (nstart < nend) { 3340 idx = MMU_TIB(nstart); 3341 b_dte = &b_tbl->bt_dtbl[idx]; 3342 rstart = nstart; 3343 while (rstart < nend) { 3344 if (MMU_VALID_DT(*b_dte)) { 3345 c_dte = mmu_ptov(MMU_DTE_PA(*b_dte)); 3346 c_tbl = mmuC2tmgr(c_dte); 3347 ct_wired = c_tbl->ct_wcnt; 3348 3349 free_c_table(c_tbl, true); 3350 3351 /* 3352 * All child entries has been removed. 3353 * If there were any wired entries in it, 3354 * decrement wired entry count. 3355 */ 3356 if (ct_wired) 3357 b_tbl->bt_wcnt--; 3358 3359 b_dte->attr.raw = MMU_DT_INVALID; 3360 b_tbl->bt_ecnt--; 3361 } 3362 b_dte++; 3363 rstart += MMU_TIB_RANGE; 3364 } 3365 } 3366 if (nend < eva) { 3367 idx = MMU_TIB(nend); 3368 b_dte = &b_tbl->bt_dtbl[idx]; 3369 if (MMU_VALID_DT(*b_dte)) { 3370 c_dte = mmu_ptov(MMU_DTE_PA(*b_dte)); 3371 c_tbl = mmuC2tmgr(c_dte); 3372 ct_wired = c_tbl->ct_wcnt; 3373 empty = pmap_remove_c(c_tbl, nend, eva); 3374 3375 /* 3376 * If the child table no longer has wired entries, 3377 * decrement wired entry count. 3378 */ 3379 if (ct_wired && c_tbl->ct_wcnt == 0) 3380 b_tbl->bt_wcnt--; 3381 3382 if (empty) { 3383 b_dte->attr.raw = MMU_DT_INVALID; 3384 b_tbl->bt_ecnt--; 3385 } 3386 } 3387 } 3388 3389 if (b_tbl->bt_ecnt == 0) { 3390 KASSERT(b_tbl->bt_wcnt == 0); 3391 b_tbl->bt_parent = NULL; 3392 if (!bt_wired) 3393 TAILQ_REMOVE(&b_pool, b_tbl, bt_link); 3394 TAILQ_INSERT_HEAD(&b_pool, b_tbl, bt_link); 3395 empty = true; 3396 } else { 3397 /* 3398 * If the table doesn't have wired entries any longer 3399 * but still has unwired entries, put it back into 3400 * the available queue. 3401 */ 3402 if (bt_wired && b_tbl->bt_wcnt == 0) 3403 TAILQ_INSERT_TAIL(&b_pool, b_tbl, bt_link); 3404 3405 empty = false; 3406 } 3407 3408 return empty; 3409 } 3410 3411 /* pmap_remove_c INTERNAL 3412 ** 3413 * Remove a range of addresses from the given C table. 3414 */ 3415 bool 3416 pmap_remove_c(c_tmgr_t *c_tbl, vaddr_t sva, vaddr_t eva) 3417 { 3418 bool empty; 3419 int idx; 3420 mmu_short_pte_t *c_pte; 3421 uint8_t ct_wired; 3422 3423 ct_wired = c_tbl->ct_wcnt; 3424 3425 idx = MMU_TIC(sva); 3426 c_pte = &c_tbl->ct_dtbl[idx]; 3427 for (; sva < eva; sva += MMU_PAGE_SIZE, c_pte++) { 3428 if (MMU_VALID_DT(*c_pte)) { 3429 if (c_pte->attr.raw & MMU_SHORT_PTE_WIRED) 3430 c_tbl->ct_wcnt--; 3431 pmap_remove_pte(c_pte); 3432 c_tbl->ct_ecnt--; 3433 } 3434 } 3435 3436 if (c_tbl->ct_ecnt == 0) { 3437 KASSERT(c_tbl->ct_wcnt == 0); 3438 c_tbl->ct_parent = NULL; 3439 if (!ct_wired) 3440 TAILQ_REMOVE(&c_pool, c_tbl, ct_link); 3441 TAILQ_INSERT_HEAD(&c_pool, c_tbl, ct_link); 3442 empty = true; 3443 } else { 3444 /* 3445 * If the table doesn't have wired entries any longer 3446 * but still has unwired entries, put it back into 3447 * the available queue. 3448 */ 3449 if (ct_wired && c_tbl->ct_wcnt == 0) 3450 TAILQ_INSERT_TAIL(&c_pool, c_tbl, ct_link); 3451 empty = false; 3452 } 3453 3454 return empty; 3455 } 3456 3457 /* pmap_bootstrap_alloc INTERNAL 3458 ** 3459 * Used internally for memory allocation at startup when malloc is not 3460 * available. This code will fail once it crosses the first memory 3461 * bank boundary on the 3/80. Hopefully by then however, the VM system 3462 * will be in charge of allocation. 3463 */ 3464 void * 3465 pmap_bootstrap_alloc(int size) 3466 { 3467 void *rtn; 3468 3469 #ifdef PMAP_DEBUG 3470 if (bootstrap_alloc_enabled == false) { 3471 mon_printf("pmap_bootstrap_alloc: disabled\n"); 3472 sunmon_abort(); 3473 } 3474 #endif 3475 3476 rtn = (void *) virtual_avail; 3477 virtual_avail += size; 3478 3479 #ifdef PMAP_DEBUG 3480 if (virtual_avail > virtual_contig_end) { 3481 mon_printf("pmap_bootstrap_alloc: out of mem\n"); 3482 sunmon_abort(); 3483 } 3484 #endif 3485 3486 return rtn; 3487 } 3488 3489 /* pmap_bootstap_aalign INTERNAL 3490 ** 3491 * Used to insure that the next call to pmap_bootstrap_alloc() will 3492 * return a chunk of memory aligned to the specified size. 3493 * 3494 * Note: This function will only support alignment sizes that are powers 3495 * of two. 3496 */ 3497 void 3498 pmap_bootstrap_aalign(int size) 3499 { 3500 int off; 3501 3502 off = virtual_avail & (size - 1); 3503 if (off) { 3504 (void)pmap_bootstrap_alloc(size - off); 3505 } 3506 } 3507 3508 /* pmap_pa_exists 3509 ** 3510 * Used by the /dev/mem driver to see if a given PA is memory 3511 * that can be mapped. (The PA is not in a hole.) 3512 */ 3513 int 3514 pmap_pa_exists(paddr_t pa) 3515 { 3516 int i; 3517 3518 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) { 3519 if ((pa >= avail_mem[i].pmem_start) && 3520 (pa < avail_mem[i].pmem_end)) 3521 return 1; 3522 if (avail_mem[i].pmem_next == NULL) 3523 break; 3524 } 3525 return 0; 3526 } 3527 3528 /* Called only from locore.s and pmap.c */ 3529 void _pmap_switch(pmap_t pmap); 3530 3531 /* 3532 * _pmap_switch INTERNAL 3533 * 3534 * This is called by locore.s:cpu_switch() when it is 3535 * switching to a new process. Load new translations. 3536 * Note: done in-line by locore.s unless PMAP_DEBUG 3537 * 3538 * Note that we do NOT allocate a context here, but 3539 * share the "kernel only" context until we really 3540 * need our own context for user-space mappings in 3541 * pmap_enter_user(). [ s/context/mmu A table/ ] 3542 */ 3543 void 3544 _pmap_switch(pmap_t pmap) 3545 { 3546 u_long rootpa; 3547 3548 /* 3549 * Only do reload/flush if we have to. 3550 * Note that if the old and new process 3551 * were BOTH using the "null" context, 3552 * then this will NOT flush the TLB. 3553 */ 3554 rootpa = pmap->pm_a_phys; 3555 if (kernel_crp.rp_addr != rootpa) { 3556 DPRINT(("pmap_activate(%p)\n", pmap)); 3557 kernel_crp.rp_addr = rootpa; 3558 loadcrp(&kernel_crp); 3559 TBIAU(); 3560 } 3561 } 3562 3563 /* 3564 * Exported version of pmap_activate(). This is called from the 3565 * machine-independent VM code when a process is given a new pmap. 3566 * If (p == curlwp) do like cpu_switch would do; otherwise just 3567 * take this as notification that the process has a new pmap. 3568 */ 3569 void 3570 pmap_activate(struct lwp *l) 3571 { 3572 3573 if (l->l_proc == curproc) { 3574 _pmap_switch(l->l_proc->p_vmspace->vm_map.pmap); 3575 } 3576 } 3577 3578 /* 3579 * pmap_deactivate INTERFACE 3580 ** 3581 * This is called to deactivate the specified process's address space. 3582 */ 3583 void 3584 pmap_deactivate(struct lwp *l) 3585 { 3586 3587 /* Nothing to do. */ 3588 } 3589 3590 /* 3591 * Fill in the sun3x-specific part of the kernel core header 3592 * for dumpsys(). (See machdep.c for the rest.) 3593 */ 3594 void 3595 pmap_kcore_hdr(struct sun3x_kcore_hdr *sh) 3596 { 3597 u_long spa, len; 3598 int i; 3599 3600 sh->pg_frame = MMU_SHORT_PTE_BASEADDR; 3601 sh->pg_valid = MMU_DT_PAGE; 3602 sh->contig_end = virtual_contig_end; 3603 sh->kernCbase = (u_long)kernCbase; 3604 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) { 3605 spa = avail_mem[i].pmem_start; 3606 spa = m68k_trunc_page(spa); 3607 len = avail_mem[i].pmem_end - spa; 3608 len = m68k_round_page(len); 3609 sh->ram_segs[i].start = spa; 3610 sh->ram_segs[i].size = len; 3611 } 3612 } 3613 3614 3615 /* pmap_virtual_space INTERFACE 3616 ** 3617 * Return the current available range of virtual addresses in the 3618 * arguuments provided. Only really called once. 3619 */ 3620 void 3621 pmap_virtual_space(vaddr_t *vstart, vaddr_t *vend) 3622 { 3623 3624 *vstart = virtual_avail; 3625 *vend = virtual_end; 3626 } 3627 3628 /* 3629 * Provide memory to the VM system. 3630 * 3631 * Assume avail_start is always in the 3632 * first segment as pmap_bootstrap does. 3633 */ 3634 static void 3635 pmap_page_upload(void) 3636 { 3637 paddr_t a, b; /* memory range */ 3638 int i; 3639 3640 /* Supply the memory in segments. */ 3641 for (i = 0; i < SUN3X_NPHYS_RAM_SEGS; i++) { 3642 a = atop(avail_mem[i].pmem_start); 3643 b = atop(avail_mem[i].pmem_end); 3644 if (i == 0) 3645 a = atop(avail_start); 3646 if (avail_mem[i].pmem_end > avail_end) 3647 b = atop(avail_end); 3648 3649 uvm_page_physload(a, b, a, b, VM_FREELIST_DEFAULT); 3650 3651 if (avail_mem[i].pmem_next == NULL) 3652 break; 3653 } 3654 } 3655 3656 /* pmap_count INTERFACE 3657 ** 3658 * Return the number of resident (valid) pages in the given pmap. 3659 * 3660 * Note: If this function is handed the kernel map, it will report 3661 * that it has no mappings. Hopefully the VM system won't ask for kernel 3662 * map statistics. 3663 */ 3664 segsz_t 3665 pmap_count(pmap_t pmap, int type) 3666 { 3667 u_int count; 3668 int a_idx, b_idx; 3669 a_tmgr_t *a_tbl; 3670 b_tmgr_t *b_tbl; 3671 c_tmgr_t *c_tbl; 3672 3673 /* 3674 * If the pmap does not have its own A table manager, it has no 3675 * valid entires. 3676 */ 3677 if (pmap->pm_a_tmgr == NULL) 3678 return 0; 3679 3680 a_tbl = pmap->pm_a_tmgr; 3681 3682 count = 0; 3683 for (a_idx = 0; a_idx < MMU_TIA(KERNBASE); a_idx++) { 3684 if (MMU_VALID_DT(a_tbl->at_dtbl[a_idx])) { 3685 b_tbl = mmuB2tmgr(mmu_ptov(a_tbl->at_dtbl[a_idx].addr.raw)); 3686 for (b_idx = 0; b_idx < MMU_B_TBL_SIZE; b_idx++) { 3687 if (MMU_VALID_DT(b_tbl->bt_dtbl[b_idx])) { 3688 c_tbl = mmuC2tmgr( 3689 mmu_ptov(MMU_DTE_PA(b_tbl->bt_dtbl[b_idx]))); 3690 if (type == 0) 3691 /* 3692 * A resident entry count has been requested. 3693 */ 3694 count += c_tbl->ct_ecnt; 3695 else 3696 /* 3697 * A wired entry count has been requested. 3698 */ 3699 count += c_tbl->ct_wcnt; 3700 } 3701 } 3702 } 3703 } 3704 3705 return count; 3706 } 3707 3708 /************************ SUN3 COMPATIBILITY ROUTINES ******************** 3709 * The following routines are only used by DDB for tricky kernel text * 3710 * text operations in db_memrw.c. They are provided for sun3 * 3711 * compatibility. * 3712 *************************************************************************/ 3713 /* get_pte INTERNAL 3714 ** 3715 * Return the page descriptor the describes the kernel mapping 3716 * of the given virtual address. 3717 */ 3718 extern u_long ptest_addr(u_long); /* XXX: locore.s */ 3719 u_int 3720 get_pte(vaddr_t va) 3721 { 3722 u_long pte_pa; 3723 mmu_short_pte_t *pte; 3724 3725 /* Get the physical address of the PTE */ 3726 pte_pa = ptest_addr(va & ~PGOFSET); 3727 3728 /* Convert to a virtual address... */ 3729 pte = (mmu_short_pte_t *) (KERNBASE + pte_pa); 3730 3731 /* Make sure it is in our level-C tables... */ 3732 if ((pte < kernCbase) || 3733 (pte >= &mmuCbase[NUM_USER_PTES])) 3734 return 0; 3735 3736 /* ... and just return its contents. */ 3737 return (pte->attr.raw); 3738 } 3739 3740 3741 /* set_pte INTERNAL 3742 ** 3743 * Set the page descriptor that describes the kernel mapping 3744 * of the given virtual address. 3745 */ 3746 void 3747 set_pte(vaddr_t va, u_int pte) 3748 { 3749 u_long idx; 3750 3751 if (va < KERNBASE) 3752 return; 3753 3754 idx = (unsigned long) m68k_btop(va - KERNBASE); 3755 kernCbase[idx].attr.raw = pte; 3756 TBIS(va); 3757 } 3758 3759 /* 3760 * Routine: pmap_procwr 3761 * 3762 * Function: 3763 * Synchronize caches corresponding to [addr, addr+len) in p. 3764 */ 3765 void 3766 pmap_procwr(struct proc *p, vaddr_t va, size_t len) 3767 { 3768 3769 (void)cachectl1(0x80000004, va, len, p); 3770 } 3771 3772 3773 #ifdef PMAP_DEBUG 3774 /************************** DEBUGGING ROUTINES ************************** 3775 * The following routines are meant to be an aid to debugging the pmap * 3776 * system. They are callable from the DDB command line and should be * 3777 * prepared to be handed unstable or incomplete states of the system. * 3778 ************************************************************************/ 3779 3780 /* pv_list 3781 ** 3782 * List all pages found on the pv list for the given physical page. 3783 * To avoid endless loops, the listing will stop at the end of the list 3784 * or after 'n' entries - whichever comes first. 3785 */ 3786 void 3787 pv_list(paddr_t pa, int n) 3788 { 3789 int idx; 3790 vaddr_t va; 3791 pv_t *pv; 3792 c_tmgr_t *c_tbl; 3793 pmap_t pmap; 3794 3795 pv = pa2pv(pa); 3796 idx = pv->pv_idx; 3797 for (; idx != PVE_EOL && n > 0; idx = pvebase[idx].pve_next, n--) { 3798 va = pmap_get_pteinfo(idx, &pmap, &c_tbl); 3799 printf("idx %d, pmap 0x%x, va 0x%x, c_tbl %x\n", 3800 idx, (u_int) pmap, (u_int) va, (u_int) c_tbl); 3801 } 3802 } 3803 #endif /* PMAP_DEBUG */ 3804 3805 #ifdef NOT_YET 3806 /* and maybe not ever */ 3807 /************************** LOW-LEVEL ROUTINES ************************** 3808 * These routines will eventually be re-written into assembly and placed* 3809 * in locore.s. They are here now as stubs so that the pmap module can * 3810 * be linked as a standalone user program for testing. * 3811 ************************************************************************/ 3812 /* flush_atc_crp INTERNAL 3813 ** 3814 * Flush all page descriptors derived from the given CPU Root Pointer 3815 * (CRP), or 'A' table as it is known here, from the 68851's automatic 3816 * cache. 3817 */ 3818 void 3819 flush_atc_crp(int a_tbl) 3820 { 3821 mmu_long_rp_t rp; 3822 3823 /* Create a temporary root table pointer that points to the 3824 * given A table. 3825 */ 3826 rp.attr.raw = ~MMU_LONG_RP_LU; 3827 rp.addr.raw = (unsigned int) a_tbl; 3828 3829 mmu_pflushr(&rp); 3830 /* mmu_pflushr: 3831 * movel sp(4)@,a0 3832 * pflushr a0@ 3833 * rts 3834 */ 3835 } 3836 #endif /* NOT_YET */ 3837