pmap.c revision 1.14.2.4
1 /* $NetBSD: pmap.c,v 1.14.2.4 2002/01/10 19:37:49 thorpej Exp $ */ 2 3 /* 4 * Copyright (c) 2001 Richard Earnshaw 5 * Copyright (c) 2001 Christopher Gilbert 6 * All rights reserved. 7 * 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 3. The name of the company nor the name of the author may be used to 14 * endorse or promote products derived from this software without specific 15 * prior written permission. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED 18 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF 19 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 20 * IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, 21 * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 22 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 23 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30 /*- 31 * Copyright (c) 1999 The NetBSD Foundation, Inc. 32 * All rights reserved. 33 * 34 * This code is derived from software contributed to The NetBSD Foundation 35 * by Charles M. Hannum. 36 * 37 * Redistribution and use in source and binary forms, with or without 38 * modification, are permitted provided that the following conditions 39 * are met: 40 * 1. Redistributions of source code must retain the above copyright 41 * notice, this list of conditions and the following disclaimer. 42 * 2. Redistributions in binary form must reproduce the above copyright 43 * notice, this list of conditions and the following disclaimer in the 44 * documentation and/or other materials provided with the distribution. 45 * 3. All advertising materials mentioning features or use of this software 46 * must display the following acknowledgement: 47 * This product includes software developed by the NetBSD 48 * Foundation, Inc. and its contributors. 49 * 4. Neither the name of The NetBSD Foundation nor the names of its 50 * contributors may be used to endorse or promote products derived 51 * from this software without specific prior written permission. 52 * 53 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 54 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 55 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 56 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 57 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 58 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 59 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 60 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 61 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 62 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 63 * POSSIBILITY OF SUCH DAMAGE. 64 */ 65 66 /* 67 * Copyright (c) 1994-1998 Mark Brinicombe. 68 * Copyright (c) 1994 Brini. 69 * All rights reserved. 70 * 71 * This code is derived from software written for Brini by Mark Brinicombe 72 * 73 * Redistribution and use in source and binary forms, with or without 74 * modification, are permitted provided that the following conditions 75 * are met: 76 * 1. Redistributions of source code must retain the above copyright 77 * notice, this list of conditions and the following disclaimer. 78 * 2. Redistributions in binary form must reproduce the above copyright 79 * notice, this list of conditions and the following disclaimer in the 80 * documentation and/or other materials provided with the distribution. 81 * 3. All advertising materials mentioning features or use of this software 82 * must display the following acknowledgement: 83 * This product includes software developed by Mark Brinicombe. 84 * 4. The name of the author may not be used to endorse or promote products 85 * derived from this software without specific prior written permission. 86 * 87 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 88 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 89 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 90 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 91 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 92 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 93 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 94 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 95 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 96 * 97 * RiscBSD kernel project 98 * 99 * pmap.c 100 * 101 * Machine dependant vm stuff 102 * 103 * Created : 20/09/94 104 */ 105 106 /* 107 * Performance improvements, UVM changes, overhauls and part-rewrites 108 * were contributed by Neil A. Carson <neil (at) causality.com>. 109 */ 110 111 /* 112 * The dram block info is currently referenced from the bootconfig. 113 * This should be placed in a separate structure. 114 */ 115 116 /* 117 * Special compilation symbols 118 * PMAP_DEBUG - Build in pmap_debug_level code 119 */ 120 121 /* Include header files */ 122 123 #include "opt_pmap_debug.h" 124 #include "opt_ddb.h" 125 126 #include <sys/types.h> 127 #include <sys/param.h> 128 #include <sys/kernel.h> 129 #include <sys/systm.h> 130 #include <sys/proc.h> 131 #include <sys/malloc.h> 132 #include <sys/user.h> 133 #include <sys/pool.h> 134 #include <sys/cdefs.h> 135 136 #include <uvm/uvm.h> 137 138 #include <machine/bootconfig.h> 139 #include <machine/bus.h> 140 #include <machine/pmap.h> 141 #include <machine/pcb.h> 142 #include <machine/param.h> 143 #include <arm/arm32/katelib.h> 144 145 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.14.2.4 2002/01/10 19:37:49 thorpej Exp $"); 146 147 #ifdef PMAP_DEBUG 148 #define PDEBUG(_lev_,_stat_) \ 149 if (pmap_debug_level >= (_lev_)) \ 150 ((_stat_)) 151 int pmap_debug_level = -2; 152 153 /* 154 * for switching to potentially finer grained debugging 155 */ 156 #define PDB_FOLLOW 0x0001 157 #define PDB_INIT 0x0002 158 #define PDB_ENTER 0x0004 159 #define PDB_REMOVE 0x0008 160 #define PDB_CREATE 0x0010 161 #define PDB_PTPAGE 0x0020 162 #define PDB_ASN 0x0040 163 #define PDB_BITS 0x0080 164 #define PDB_COLLECT 0x0100 165 #define PDB_PROTECT 0x0200 166 #define PDB_BOOTSTRAP 0x1000 167 #define PDB_PARANOIA 0x2000 168 #define PDB_WIRING 0x4000 169 #define PDB_PVDUMP 0x8000 170 171 int debugmap = 0; 172 int pmapdebug = PDB_PARANOIA | PDB_FOLLOW; 173 #define NPDEBUG(_lev_,_stat_) \ 174 if (pmapdebug & (_lev_)) \ 175 ((_stat_)) 176 177 #else /* PMAP_DEBUG */ 178 #define PDEBUG(_lev_,_stat_) /* Nothing */ 179 #define PDEBUG(_lev_,_stat_) /* Nothing */ 180 #endif /* PMAP_DEBUG */ 181 182 struct pmap kernel_pmap_store; 183 184 /* 185 * pool that pmap structures are allocated from 186 */ 187 188 struct pool pmap_pmap_pool; 189 190 pagehook_t page_hook0; 191 pagehook_t page_hook1; 192 char *memhook; 193 pt_entry_t msgbufpte; 194 extern caddr_t msgbufaddr; 195 196 boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 197 /* 198 * locking data structures 199 */ 200 201 static struct lock pmap_main_lock; 202 static struct simplelock pvalloc_lock; 203 #ifdef LOCKDEBUG 204 #define PMAP_MAP_TO_HEAD_LOCK() \ 205 (void) spinlockmgr(&pmap_main_lock, LK_SHARED, NULL) 206 #define PMAP_MAP_TO_HEAD_UNLOCK() \ 207 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL) 208 209 #define PMAP_HEAD_TO_MAP_LOCK() \ 210 (void) spinlockmgr(&pmap_main_lock, LK_EXCLUSIVE, NULL) 211 #define PMAP_HEAD_TO_MAP_UNLOCK() \ 212 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL) 213 #else 214 #define PMAP_MAP_TO_HEAD_LOCK() /* nothing */ 215 #define PMAP_MAP_TO_HEAD_UNLOCK() /* nothing */ 216 #define PMAP_HEAD_TO_MAP_LOCK() /* nothing */ 217 #define PMAP_HEAD_TO_MAP_UNLOCK() /* nothing */ 218 #endif /* LOCKDEBUG */ 219 220 /* 221 * pv_page management structures: locked by pvalloc_lock 222 */ 223 224 TAILQ_HEAD(pv_pagelist, pv_page); 225 static struct pv_pagelist pv_freepages; /* list of pv_pages with free entrys */ 226 static struct pv_pagelist pv_unusedpgs; /* list of unused pv_pages */ 227 static int pv_nfpvents; /* # of free pv entries */ 228 static struct pv_page *pv_initpage; /* bootstrap page from kernel_map */ 229 static vaddr_t pv_cachedva; /* cached VA for later use */ 230 231 #define PVE_LOWAT (PVE_PER_PVPAGE / 2) /* free pv_entry low water mark */ 232 #define PVE_HIWAT (PVE_LOWAT + (PVE_PER_PVPAGE * 2)) 233 /* high water mark */ 234 235 /* 236 * local prototypes 237 */ 238 239 static struct pv_entry *pmap_add_pvpage __P((struct pv_page *, boolean_t)); 240 static struct pv_entry *pmap_alloc_pv __P((struct pmap *, int)); /* see codes below */ 241 #define ALLOCPV_NEED 0 /* need PV now */ 242 #define ALLOCPV_TRY 1 /* just try to allocate, don't steal */ 243 #define ALLOCPV_NONEED 2 /* don't need PV, just growing cache */ 244 static struct pv_entry *pmap_alloc_pvpage __P((struct pmap *, int)); 245 static void pmap_enter_pv __P((struct pv_head *, 246 struct pv_entry *, struct pmap *, 247 vaddr_t, struct vm_page *, int)); 248 static void pmap_free_pv __P((struct pmap *, struct pv_entry *)); 249 static void pmap_free_pvs __P((struct pmap *, struct pv_entry *)); 250 static void pmap_free_pv_doit __P((struct pv_entry *)); 251 static void pmap_free_pvpage __P((void)); 252 static boolean_t pmap_is_curpmap __P((struct pmap *)); 253 static struct pv_entry *pmap_remove_pv __P((struct pv_head *, struct pmap *, 254 vaddr_t)); 255 #define PMAP_REMOVE_ALL 0 /* remove all mappings */ 256 #define PMAP_REMOVE_SKIPWIRED 1 /* skip wired mappings */ 257 258 static u_int pmap_modify_pv __P((struct pmap *, vaddr_t, struct pv_head *, 259 u_int, u_int)); 260 261 static void pmap_free_l1pt __P((struct l1pt *)); 262 static int pmap_allocpagedir __P((struct pmap *)); 263 static int pmap_clean_page __P((struct pv_entry *, boolean_t)); 264 static struct pv_head *pmap_find_pvh __P((paddr_t)); 265 static void pmap_remove_all __P((paddr_t)); 266 267 268 vsize_t npages; 269 270 static struct vm_page *pmap_alloc_ptp __P((struct pmap *, vaddr_t, boolean_t)); 271 static struct vm_page *pmap_get_ptp __P((struct pmap *, vaddr_t, boolean_t)); 272 __inline static void pmap_clearbit __P((paddr_t, unsigned int)); 273 __inline static boolean_t pmap_testbit __P((paddr_t, unsigned int)); 274 275 extern paddr_t physical_start; 276 extern paddr_t physical_freestart; 277 extern paddr_t physical_end; 278 extern paddr_t physical_freeend; 279 extern unsigned int free_pages; 280 extern int max_processes; 281 282 vaddr_t virtual_start; 283 vaddr_t virtual_end; 284 285 vaddr_t avail_start; 286 vaddr_t avail_end; 287 288 extern pv_addr_t systempage; 289 290 #define ALLOC_PAGE_HOOK(x, s) \ 291 x.va = virtual_start; \ 292 x.pte = (pt_entry_t *)pmap_pte(pmap_kernel(), virtual_start); \ 293 virtual_start += s; 294 295 /* Variables used by the L1 page table queue code */ 296 SIMPLEQ_HEAD(l1pt_queue, l1pt); 297 struct l1pt_queue l1pt_static_queue; /* head of our static l1 queue */ 298 int l1pt_static_queue_count; /* items in the static l1 queue */ 299 int l1pt_static_create_count; /* static l1 items created */ 300 struct l1pt_queue l1pt_queue; /* head of our l1 queue */ 301 int l1pt_queue_count; /* items in the l1 queue */ 302 int l1pt_create_count; /* stat - L1's create count */ 303 int l1pt_reuse_count; /* stat - L1's reused count */ 304 305 /* Local function prototypes (not used outside this file) */ 306 pt_entry_t *pmap_pte __P((struct pmap *pmap, vaddr_t va)); 307 void map_pagetable __P((vaddr_t pagetable, vaddr_t va, 308 paddr_t pa, unsigned int flags)); 309 void pmap_copy_on_write __P((paddr_t pa)); 310 void pmap_pinit __P((struct pmap *)); 311 void pmap_freepagedir __P((struct pmap *)); 312 313 /* Other function prototypes */ 314 extern void bzero_page __P((vaddr_t)); 315 extern void bcopy_page __P((vaddr_t, vaddr_t)); 316 317 struct l1pt *pmap_alloc_l1pt __P((void)); 318 static __inline void pmap_map_in_l1 __P((struct pmap *pmap, vaddr_t va, 319 vaddr_t l2pa, boolean_t)); 320 321 static pt_entry_t *pmap_map_ptes __P((struct pmap *)); 322 static void pmap_unmap_ptes __P((struct pmap *)); 323 324 __inline static void pmap_vac_me_harder __P((struct pmap *, struct pv_head *, 325 pt_entry_t *, boolean_t)); 326 static void pmap_vac_me_kpmap __P((struct pmap *, struct pv_head *, 327 pt_entry_t *, boolean_t)); 328 static void pmap_vac_me_user __P((struct pmap *, struct pv_head *, 329 pt_entry_t *, boolean_t)); 330 331 /* 332 * Cache enable bits in PTE to use on pages that are cacheable. 333 * On most machines this is cacheable/bufferable, but on some, eg arm10, we 334 * can chose between write-through and write-back cacheing. 335 */ 336 pt_entry_t pte_cache_mode = (PT_C | PT_B); 337 338 /* 339 * real definition of pv_entry. 340 */ 341 342 struct pv_entry { 343 struct pv_entry *pv_next; /* next pv_entry */ 344 struct pmap *pv_pmap; /* pmap where mapping lies */ 345 vaddr_t pv_va; /* virtual address for mapping */ 346 int pv_flags; /* flags */ 347 struct vm_page *pv_ptp; /* vm_page for the ptp */ 348 }; 349 350 /* 351 * pv_entrys are dynamically allocated in chunks from a single page. 352 * we keep track of how many pv_entrys are in use for each page and 353 * we can free pv_entry pages if needed. there is one lock for the 354 * entire allocation system. 355 */ 356 357 struct pv_page_info { 358 TAILQ_ENTRY(pv_page) pvpi_list; 359 struct pv_entry *pvpi_pvfree; 360 int pvpi_nfree; 361 }; 362 363 /* 364 * number of pv_entry's in a pv_page 365 * (note: won't work on systems where NPBG isn't a constant) 366 */ 367 368 #define PVE_PER_PVPAGE ((NBPG - sizeof(struct pv_page_info)) / \ 369 sizeof(struct pv_entry)) 370 371 /* 372 * a pv_page: where pv_entrys are allocated from 373 */ 374 375 struct pv_page { 376 struct pv_page_info pvinfo; 377 struct pv_entry pvents[PVE_PER_PVPAGE]; 378 }; 379 380 #ifdef MYCROFT_HACK 381 int mycroft_hack = 0; 382 #endif 383 384 /* Function to set the debug level of the pmap code */ 385 386 #ifdef PMAP_DEBUG 387 void 388 pmap_debug(level) 389 int level; 390 { 391 pmap_debug_level = level; 392 printf("pmap_debug: level=%d\n", pmap_debug_level); 393 } 394 #endif /* PMAP_DEBUG */ 395 396 __inline static boolean_t 397 pmap_is_curpmap(struct pmap *pmap) 398 { 399 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap) 400 || (pmap == pmap_kernel())) 401 return (TRUE); 402 return (FALSE); 403 } 404 #include "isadma.h" 405 406 #if NISADMA > 0 407 /* 408 * Used to protect memory for ISA DMA bounce buffers. If, when loading 409 * pages into the system, memory intersects with any of these ranges, 410 * the intersecting memory will be loaded into a lower-priority free list. 411 */ 412 bus_dma_segment_t *pmap_isa_dma_ranges; 413 int pmap_isa_dma_nranges; 414 415 boolean_t pmap_isa_dma_range_intersect __P((paddr_t, psize_t, 416 paddr_t *, psize_t *)); 417 418 /* 419 * Check if a memory range intersects with an ISA DMA range, and 420 * return the page-rounded intersection if it does. The intersection 421 * will be placed on a lower-priority free list. 422 */ 423 boolean_t 424 pmap_isa_dma_range_intersect(pa, size, pap, sizep) 425 paddr_t pa; 426 psize_t size; 427 paddr_t *pap; 428 psize_t *sizep; 429 { 430 bus_dma_segment_t *ds; 431 int i; 432 433 if (pmap_isa_dma_ranges == NULL) 434 return (FALSE); 435 436 for (i = 0, ds = pmap_isa_dma_ranges; 437 i < pmap_isa_dma_nranges; i++, ds++) { 438 if (ds->ds_addr <= pa && pa < (ds->ds_addr + ds->ds_len)) { 439 /* 440 * Beginning of region intersects with this range. 441 */ 442 *pap = trunc_page(pa); 443 *sizep = round_page(min(pa + size, 444 ds->ds_addr + ds->ds_len) - pa); 445 return (TRUE); 446 } 447 if (pa < ds->ds_addr && ds->ds_addr < (pa + size)) { 448 /* 449 * End of region intersects with this range. 450 */ 451 *pap = trunc_page(ds->ds_addr); 452 *sizep = round_page(min((pa + size) - ds->ds_addr, 453 ds->ds_len)); 454 return (TRUE); 455 } 456 } 457 458 /* 459 * No intersection found. 460 */ 461 return (FALSE); 462 } 463 #endif /* NISADMA > 0 */ 464 465 /* 466 * p v _ e n t r y f u n c t i o n s 467 */ 468 469 /* 470 * pv_entry allocation functions: 471 * the main pv_entry allocation functions are: 472 * pmap_alloc_pv: allocate a pv_entry structure 473 * pmap_free_pv: free one pv_entry 474 * pmap_free_pvs: free a list of pv_entrys 475 * 476 * the rest are helper functions 477 */ 478 479 /* 480 * pmap_alloc_pv: inline function to allocate a pv_entry structure 481 * => we lock pvalloc_lock 482 * => if we fail, we call out to pmap_alloc_pvpage 483 * => 3 modes: 484 * ALLOCPV_NEED = we really need a pv_entry, even if we have to steal it 485 * ALLOCPV_TRY = we want a pv_entry, but not enough to steal 486 * ALLOCPV_NONEED = we are trying to grow our free list, don't really need 487 * one now 488 * 489 * "try" is for optional functions like pmap_copy(). 490 */ 491 492 __inline static struct pv_entry * 493 pmap_alloc_pv(pmap, mode) 494 struct pmap *pmap; 495 int mode; 496 { 497 struct pv_page *pvpage; 498 struct pv_entry *pv; 499 500 simple_lock(&pvalloc_lock); 501 502 if (pv_freepages.tqh_first != NULL) { 503 pvpage = pv_freepages.tqh_first; 504 pvpage->pvinfo.pvpi_nfree--; 505 if (pvpage->pvinfo.pvpi_nfree == 0) { 506 /* nothing left in this one? */ 507 TAILQ_REMOVE(&pv_freepages, pvpage, pvinfo.pvpi_list); 508 } 509 pv = pvpage->pvinfo.pvpi_pvfree; 510 #ifdef DIAGNOSTIC 511 if (pv == NULL) 512 panic("pmap_alloc_pv: pvpi_nfree off"); 513 #endif 514 pvpage->pvinfo.pvpi_pvfree = pv->pv_next; 515 pv_nfpvents--; /* took one from pool */ 516 } else { 517 pv = NULL; /* need more of them */ 518 } 519 520 /* 521 * if below low water mark or we didn't get a pv_entry we try and 522 * create more pv_entrys ... 523 */ 524 525 if (pv_nfpvents < PVE_LOWAT || pv == NULL) { 526 if (pv == NULL) 527 pv = pmap_alloc_pvpage(pmap, (mode == ALLOCPV_TRY) ? 528 mode : ALLOCPV_NEED); 529 else 530 (void) pmap_alloc_pvpage(pmap, ALLOCPV_NONEED); 531 } 532 533 simple_unlock(&pvalloc_lock); 534 return(pv); 535 } 536 537 /* 538 * pmap_alloc_pvpage: maybe allocate a new pvpage 539 * 540 * if need_entry is false: try and allocate a new pv_page 541 * if need_entry is true: try and allocate a new pv_page and return a 542 * new pv_entry from it. if we are unable to allocate a pv_page 543 * we make a last ditch effort to steal a pv_page from some other 544 * mapping. if that fails, we panic... 545 * 546 * => we assume that the caller holds pvalloc_lock 547 */ 548 549 static struct pv_entry * 550 pmap_alloc_pvpage(pmap, mode) 551 struct pmap *pmap; 552 int mode; 553 { 554 struct vm_page *pg; 555 struct pv_page *pvpage; 556 struct pv_entry *pv; 557 int s; 558 559 /* 560 * if we need_entry and we've got unused pv_pages, allocate from there 561 */ 562 563 if (mode != ALLOCPV_NONEED && pv_unusedpgs.tqh_first != NULL) { 564 565 /* move it to pv_freepages list */ 566 pvpage = pv_unusedpgs.tqh_first; 567 TAILQ_REMOVE(&pv_unusedpgs, pvpage, pvinfo.pvpi_list); 568 TAILQ_INSERT_HEAD(&pv_freepages, pvpage, pvinfo.pvpi_list); 569 570 /* allocate a pv_entry */ 571 pvpage->pvinfo.pvpi_nfree--; /* can't go to zero */ 572 pv = pvpage->pvinfo.pvpi_pvfree; 573 #ifdef DIAGNOSTIC 574 if (pv == NULL) 575 panic("pmap_alloc_pvpage: pvpi_nfree off"); 576 #endif 577 pvpage->pvinfo.pvpi_pvfree = pv->pv_next; 578 579 pv_nfpvents--; /* took one from pool */ 580 return(pv); 581 } 582 583 /* 584 * see if we've got a cached unmapped VA that we can map a page in. 585 * if not, try to allocate one. 586 */ 587 588 589 if (pv_cachedva == 0) { 590 s = splvm(); 591 pv_cachedva = uvm_km_kmemalloc(kmem_map, NULL, 592 PAGE_SIZE, UVM_KMF_TRYLOCK|UVM_KMF_VALLOC); 593 splx(s); 594 if (pv_cachedva == 0) { 595 return (NULL); 596 } 597 } 598 599 pg = uvm_pagealloc(NULL, pv_cachedva - vm_map_min(kernel_map), NULL, 600 UVM_PGA_USERESERVE); 601 if (pg) 602 pg->flags &= ~PG_BUSY; /* never busy */ 603 604 if (pg == NULL) 605 return (NULL); 606 607 /* 608 * add a mapping for our new pv_page and free its entrys (save one!) 609 * 610 * NOTE: If we are allocating a PV page for the kernel pmap, the 611 * pmap is already locked! (...but entering the mapping is safe...) 612 */ 613 614 pmap_kenter_pa(pv_cachedva, VM_PAGE_TO_PHYS(pg), VM_PROT_ALL); 615 pmap_update(pmap_kernel()); 616 pvpage = (struct pv_page *) pv_cachedva; 617 pv_cachedva = 0; 618 return (pmap_add_pvpage(pvpage, mode != ALLOCPV_NONEED)); 619 } 620 621 /* 622 * pmap_add_pvpage: add a pv_page's pv_entrys to the free list 623 * 624 * => caller must hold pvalloc_lock 625 * => if need_entry is true, we allocate and return one pv_entry 626 */ 627 628 static struct pv_entry * 629 pmap_add_pvpage(pvp, need_entry) 630 struct pv_page *pvp; 631 boolean_t need_entry; 632 { 633 int tofree, lcv; 634 635 /* do we need to return one? */ 636 tofree = (need_entry) ? PVE_PER_PVPAGE - 1 : PVE_PER_PVPAGE; 637 638 pvp->pvinfo.pvpi_pvfree = NULL; 639 pvp->pvinfo.pvpi_nfree = tofree; 640 for (lcv = 0 ; lcv < tofree ; lcv++) { 641 pvp->pvents[lcv].pv_next = pvp->pvinfo.pvpi_pvfree; 642 pvp->pvinfo.pvpi_pvfree = &pvp->pvents[lcv]; 643 } 644 if (need_entry) 645 TAILQ_INSERT_TAIL(&pv_freepages, pvp, pvinfo.pvpi_list); 646 else 647 TAILQ_INSERT_TAIL(&pv_unusedpgs, pvp, pvinfo.pvpi_list); 648 pv_nfpvents += tofree; 649 return((need_entry) ? &pvp->pvents[lcv] : NULL); 650 } 651 652 /* 653 * pmap_free_pv_doit: actually free a pv_entry 654 * 655 * => do not call this directly! instead use either 656 * 1. pmap_free_pv ==> free a single pv_entry 657 * 2. pmap_free_pvs => free a list of pv_entrys 658 * => we must be holding pvalloc_lock 659 */ 660 661 __inline static void 662 pmap_free_pv_doit(pv) 663 struct pv_entry *pv; 664 { 665 struct pv_page *pvp; 666 667 pvp = (struct pv_page *) arm_trunc_page((vaddr_t)pv); 668 pv_nfpvents++; 669 pvp->pvinfo.pvpi_nfree++; 670 671 /* nfree == 1 => fully allocated page just became partly allocated */ 672 if (pvp->pvinfo.pvpi_nfree == 1) { 673 TAILQ_INSERT_HEAD(&pv_freepages, pvp, pvinfo.pvpi_list); 674 } 675 676 /* free it */ 677 pv->pv_next = pvp->pvinfo.pvpi_pvfree; 678 pvp->pvinfo.pvpi_pvfree = pv; 679 680 /* 681 * are all pv_page's pv_entry's free? move it to unused queue. 682 */ 683 684 if (pvp->pvinfo.pvpi_nfree == PVE_PER_PVPAGE) { 685 TAILQ_REMOVE(&pv_freepages, pvp, pvinfo.pvpi_list); 686 TAILQ_INSERT_HEAD(&pv_unusedpgs, pvp, pvinfo.pvpi_list); 687 } 688 } 689 690 /* 691 * pmap_free_pv: free a single pv_entry 692 * 693 * => we gain the pvalloc_lock 694 */ 695 696 __inline static void 697 pmap_free_pv(pmap, pv) 698 struct pmap *pmap; 699 struct pv_entry *pv; 700 { 701 simple_lock(&pvalloc_lock); 702 pmap_free_pv_doit(pv); 703 704 /* 705 * Can't free the PV page if the PV entries were associated with 706 * the kernel pmap; the pmap is already locked. 707 */ 708 if (pv_nfpvents > PVE_HIWAT && pv_unusedpgs.tqh_first != NULL && 709 pmap != pmap_kernel()) 710 pmap_free_pvpage(); 711 712 simple_unlock(&pvalloc_lock); 713 } 714 715 /* 716 * pmap_free_pvs: free a list of pv_entrys 717 * 718 * => we gain the pvalloc_lock 719 */ 720 721 __inline static void 722 pmap_free_pvs(pmap, pvs) 723 struct pmap *pmap; 724 struct pv_entry *pvs; 725 { 726 struct pv_entry *nextpv; 727 728 simple_lock(&pvalloc_lock); 729 730 for ( /* null */ ; pvs != NULL ; pvs = nextpv) { 731 nextpv = pvs->pv_next; 732 pmap_free_pv_doit(pvs); 733 } 734 735 /* 736 * Can't free the PV page if the PV entries were associated with 737 * the kernel pmap; the pmap is already locked. 738 */ 739 if (pv_nfpvents > PVE_HIWAT && pv_unusedpgs.tqh_first != NULL && 740 pmap != pmap_kernel()) 741 pmap_free_pvpage(); 742 743 simple_unlock(&pvalloc_lock); 744 } 745 746 747 /* 748 * pmap_free_pvpage: try and free an unused pv_page structure 749 * 750 * => assume caller is holding the pvalloc_lock and that 751 * there is a page on the pv_unusedpgs list 752 * => if we can't get a lock on the kmem_map we try again later 753 * => note: analysis of MI kmem_map usage [i.e. malloc/free] shows 754 * that if we can lock the kmem_map then we are not already 755 * holding kmem_object's lock. 756 */ 757 758 static void 759 pmap_free_pvpage() 760 { 761 int s; 762 struct vm_map *map; 763 struct vm_map_entry *dead_entries; 764 struct pv_page *pvp; 765 766 s = splvm(); /* protect kmem_map */ 767 768 pvp = pv_unusedpgs.tqh_first; 769 770 /* 771 * note: watch out for pv_initpage which is allocated out of 772 * kernel_map rather than kmem_map. 773 */ 774 if (pvp == pv_initpage) 775 map = kernel_map; 776 else 777 map = kmem_map; 778 779 if (vm_map_lock_try(map)) { 780 781 /* remove pvp from pv_unusedpgs */ 782 TAILQ_REMOVE(&pv_unusedpgs, pvp, pvinfo.pvpi_list); 783 784 /* unmap the page */ 785 dead_entries = NULL; 786 uvm_unmap_remove(map, (vaddr_t)pvp, ((vaddr_t)pvp) + PAGE_SIZE, 787 &dead_entries); 788 vm_map_unlock(map); 789 790 if (dead_entries != NULL) 791 uvm_unmap_detach(dead_entries, 0); 792 793 pv_nfpvents -= PVE_PER_PVPAGE; /* update free count */ 794 } 795 796 if (pvp == pv_initpage) 797 /* no more initpage, we've freed it */ 798 pv_initpage = NULL; 799 800 splx(s); 801 } 802 803 /* 804 * main pv_entry manipulation functions: 805 * pmap_enter_pv: enter a mapping onto a pv_head list 806 * pmap_remove_pv: remove a mappiing from a pv_head list 807 * 808 * NOTE: pmap_enter_pv expects to lock the pvh itself 809 * pmap_remove_pv expects te caller to lock the pvh before calling 810 */ 811 812 /* 813 * pmap_enter_pv: enter a mapping onto a pv_head lst 814 * 815 * => caller should hold the proper lock on pmap_main_lock 816 * => caller should have pmap locked 817 * => we will gain the lock on the pv_head and allocate the new pv_entry 818 * => caller should adjust ptp's wire_count before calling 819 * => caller should not adjust pmap's wire_count 820 */ 821 822 __inline static void 823 pmap_enter_pv(pvh, pve, pmap, va, ptp, flags) 824 struct pv_head *pvh; 825 struct pv_entry *pve; /* preallocated pve for us to use */ 826 struct pmap *pmap; 827 vaddr_t va; 828 struct vm_page *ptp; /* PTP in pmap that maps this VA */ 829 int flags; 830 { 831 pve->pv_pmap = pmap; 832 pve->pv_va = va; 833 pve->pv_ptp = ptp; /* NULL for kernel pmap */ 834 pve->pv_flags = flags; 835 simple_lock(&pvh->pvh_lock); /* lock pv_head */ 836 pve->pv_next = pvh->pvh_list; /* add to ... */ 837 pvh->pvh_list = pve; /* ... locked list */ 838 simple_unlock(&pvh->pvh_lock); /* unlock, done! */ 839 if (pve->pv_flags & PT_W) 840 ++pmap->pm_stats.wired_count; 841 } 842 843 /* 844 * pmap_remove_pv: try to remove a mapping from a pv_list 845 * 846 * => caller should hold proper lock on pmap_main_lock 847 * => pmap should be locked 848 * => caller should hold lock on pv_head [so that attrs can be adjusted] 849 * => caller should adjust ptp's wire_count and free PTP if needed 850 * => caller should NOT adjust pmap's wire_count 851 * => we return the removed pve 852 */ 853 854 __inline static struct pv_entry * 855 pmap_remove_pv(pvh, pmap, va) 856 struct pv_head *pvh; 857 struct pmap *pmap; 858 vaddr_t va; 859 { 860 struct pv_entry *pve, **prevptr; 861 862 prevptr = &pvh->pvh_list; /* previous pv_entry pointer */ 863 pve = *prevptr; 864 while (pve) { 865 if (pve->pv_pmap == pmap && pve->pv_va == va) { /* match? */ 866 *prevptr = pve->pv_next; /* remove it! */ 867 if (pve->pv_flags & PT_W) 868 --pmap->pm_stats.wired_count; 869 break; 870 } 871 prevptr = &pve->pv_next; /* previous pointer */ 872 pve = pve->pv_next; /* advance */ 873 } 874 return(pve); /* return removed pve */ 875 } 876 877 /* 878 * 879 * pmap_modify_pv: Update pv flags 880 * 881 * => caller should hold lock on pv_head [so that attrs can be adjusted] 882 * => caller should NOT adjust pmap's wire_count 883 * => caller must call pmap_vac_me_harder() if writable status of a page 884 * may have changed. 885 * => we return the old flags 886 * 887 * Modify a physical-virtual mapping in the pv table 888 */ 889 890 /*__inline */ 891 static u_int 892 pmap_modify_pv(pmap, va, pvh, bic_mask, eor_mask) 893 struct pmap *pmap; 894 vaddr_t va; 895 struct pv_head *pvh; 896 u_int bic_mask; 897 u_int eor_mask; 898 { 899 struct pv_entry *npv; 900 u_int flags, oflags; 901 902 /* 903 * There is at least one VA mapping this page. 904 */ 905 906 for (npv = pvh->pvh_list; npv; npv = npv->pv_next) { 907 if (pmap == npv->pv_pmap && va == npv->pv_va) { 908 oflags = npv->pv_flags; 909 npv->pv_flags = flags = 910 ((oflags & ~bic_mask) ^ eor_mask); 911 if ((flags ^ oflags) & PT_W) { 912 if (flags & PT_W) 913 ++pmap->pm_stats.wired_count; 914 else 915 --pmap->pm_stats.wired_count; 916 } 917 return (oflags); 918 } 919 } 920 return (0); 921 } 922 923 /* 924 * Map the specified level 2 pagetable into the level 1 page table for 925 * the given pmap to cover a chunk of virtual address space starting from the 926 * address specified. 927 */ 928 static /*__inline*/ void 929 pmap_map_in_l1(pmap, va, l2pa, selfref) 930 struct pmap *pmap; 931 vaddr_t va, l2pa; 932 boolean_t selfref; 933 { 934 vaddr_t ptva; 935 936 /* Calculate the index into the L1 page table. */ 937 ptva = (va >> PDSHIFT) & ~3; 938 939 PDEBUG(0, printf("wiring %08lx in to pd%p pte0x%lx va0x%lx\n", l2pa, 940 pmap->pm_pdir, L1_PTE(l2pa), ptva)); 941 942 /* Map page table into the L1. */ 943 pmap->pm_pdir[ptva + 0] = L1_PTE(l2pa + 0x000); 944 pmap->pm_pdir[ptva + 1] = L1_PTE(l2pa + 0x400); 945 pmap->pm_pdir[ptva + 2] = L1_PTE(l2pa + 0x800); 946 pmap->pm_pdir[ptva + 3] = L1_PTE(l2pa + 0xc00); 947 948 PDEBUG(0, printf("pt self reference %lx in %lx\n", 949 L2_PTE_NC_NB(l2pa, AP_KRW), pmap->pm_vptpt)); 950 951 /* Map the page table into the page table area. */ 952 if (selfref) { 953 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) = 954 L2_PTE_NC_NB(l2pa, AP_KRW); 955 } 956 /* XXX should be a purge */ 957 /* cpu_tlb_flushD();*/ 958 } 959 960 #if 0 961 static /*__inline*/ void 962 pmap_unmap_in_l1(pmap, va) 963 struct pmap *pmap; 964 vaddr_t va; 965 { 966 vaddr_t ptva; 967 968 /* Calculate the index into the L1 page table. */ 969 ptva = (va >> PDSHIFT) & ~3; 970 971 /* Unmap page table from the L1. */ 972 pmap->pm_pdir[ptva + 0] = 0; 973 pmap->pm_pdir[ptva + 1] = 0; 974 pmap->pm_pdir[ptva + 2] = 0; 975 pmap->pm_pdir[ptva + 3] = 0; 976 977 /* Unmap the page table from the page table area. */ 978 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) = 0; 979 980 /* XXX should be a purge */ 981 /* cpu_tlb_flushD();*/ 982 } 983 #endif 984 985 /* 986 * Used to map a range of physical addresses into kernel 987 * virtual address space. 988 * 989 * For now, VM is already on, we only need to map the 990 * specified memory. 991 */ 992 vaddr_t 993 pmap_map(va, spa, epa, prot) 994 vaddr_t va, spa, epa; 995 int prot; 996 { 997 while (spa < epa) { 998 pmap_kenter_pa(va, spa, prot); 999 va += NBPG; 1000 spa += NBPG; 1001 } 1002 pmap_update(pmap_kernel()); 1003 return(va); 1004 } 1005 1006 1007 /* 1008 * void pmap_bootstrap(pd_entry_t *kernel_l1pt, pv_addr_t kernel_ptpt) 1009 * 1010 * bootstrap the pmap system. This is called from initarm and allows 1011 * the pmap system to initailise any structures it requires. 1012 * 1013 * Currently this sets up the kernel_pmap that is statically allocated 1014 * and also allocated virtual addresses for certain page hooks. 1015 * Currently the only one page hook is allocated that is used 1016 * to zero physical pages of memory. 1017 * It also initialises the start and end address of the kernel data space. 1018 */ 1019 extern paddr_t physical_freestart; 1020 extern paddr_t physical_freeend; 1021 1022 char *boot_head; 1023 1024 void 1025 pmap_bootstrap(kernel_l1pt, kernel_ptpt) 1026 pd_entry_t *kernel_l1pt; 1027 pv_addr_t kernel_ptpt; 1028 { 1029 int loop; 1030 paddr_t start, end; 1031 #if NISADMA > 0 1032 paddr_t istart; 1033 psize_t isize; 1034 #endif 1035 1036 pmap_kernel()->pm_pdir = kernel_l1pt; 1037 pmap_kernel()->pm_pptpt = kernel_ptpt.pv_pa; 1038 pmap_kernel()->pm_vptpt = kernel_ptpt.pv_va; 1039 simple_lock_init(&pmap_kernel()->pm_lock); 1040 pmap_kernel()->pm_obj.pgops = NULL; 1041 TAILQ_INIT(&(pmap_kernel()->pm_obj.memq)); 1042 pmap_kernel()->pm_obj.uo_npages = 0; 1043 pmap_kernel()->pm_obj.uo_refs = 1; 1044 1045 /* 1046 * Initialize PAGE_SIZE-dependent variables. 1047 */ 1048 uvm_setpagesize(); 1049 1050 npages = 0; 1051 loop = 0; 1052 while (loop < bootconfig.dramblocks) { 1053 start = (paddr_t)bootconfig.dram[loop].address; 1054 end = start + (bootconfig.dram[loop].pages * NBPG); 1055 if (start < physical_freestart) 1056 start = physical_freestart; 1057 if (end > physical_freeend) 1058 end = physical_freeend; 1059 #if 0 1060 printf("%d: %lx -> %lx\n", loop, start, end - 1); 1061 #endif 1062 #if NISADMA > 0 1063 if (pmap_isa_dma_range_intersect(start, end - start, 1064 &istart, &isize)) { 1065 /* 1066 * Place the pages that intersect with the 1067 * ISA DMA range onto the ISA DMA free list. 1068 */ 1069 #if 0 1070 printf(" ISADMA 0x%lx -> 0x%lx\n", istart, 1071 istart + isize - 1); 1072 #endif 1073 uvm_page_physload(atop(istart), 1074 atop(istart + isize), atop(istart), 1075 atop(istart + isize), VM_FREELIST_ISADMA); 1076 npages += atop(istart + isize) - atop(istart); 1077 1078 /* 1079 * Load the pieces that come before 1080 * the intersection into the default 1081 * free list. 1082 */ 1083 if (start < istart) { 1084 #if 0 1085 printf(" BEFORE 0x%lx -> 0x%lx\n", 1086 start, istart - 1); 1087 #endif 1088 uvm_page_physload(atop(start), 1089 atop(istart), atop(start), 1090 atop(istart), VM_FREELIST_DEFAULT); 1091 npages += atop(istart) - atop(start); 1092 } 1093 1094 /* 1095 * Load the pieces that come after 1096 * the intersection into the default 1097 * free list. 1098 */ 1099 if ((istart + isize) < end) { 1100 #if 0 1101 printf(" AFTER 0x%lx -> 0x%lx\n", 1102 (istart + isize), end - 1); 1103 #endif 1104 uvm_page_physload(atop(istart + isize), 1105 atop(end), atop(istart + isize), 1106 atop(end), VM_FREELIST_DEFAULT); 1107 npages += atop(end) - atop(istart + isize); 1108 } 1109 } else { 1110 uvm_page_physload(atop(start), atop(end), 1111 atop(start), atop(end), VM_FREELIST_DEFAULT); 1112 npages += atop(end) - atop(start); 1113 } 1114 #else /* NISADMA > 0 */ 1115 uvm_page_physload(atop(start), atop(end), 1116 atop(start), atop(end), VM_FREELIST_DEFAULT); 1117 npages += atop(end) - atop(start); 1118 #endif /* NISADMA > 0 */ 1119 ++loop; 1120 } 1121 1122 #ifdef MYCROFT_HACK 1123 printf("npages = %ld\n", npages); 1124 #endif 1125 1126 virtual_start = KERNEL_VM_BASE; 1127 virtual_end = virtual_start + KERNEL_VM_SIZE - 1; 1128 1129 ALLOC_PAGE_HOOK(page_hook0, NBPG); 1130 ALLOC_PAGE_HOOK(page_hook1, NBPG); 1131 1132 /* 1133 * The mem special device needs a virtual hook but we don't 1134 * need a pte 1135 */ 1136 memhook = (char *)virtual_start; 1137 virtual_start += NBPG; 1138 1139 msgbufaddr = (caddr_t)virtual_start; 1140 msgbufpte = (pt_entry_t)pmap_pte(pmap_kernel(), virtual_start); 1141 virtual_start += round_page(MSGBUFSIZE); 1142 1143 /* 1144 * init the static-global locks and global lists. 1145 */ 1146 spinlockinit(&pmap_main_lock, "pmaplk", 0); 1147 simple_lock_init(&pvalloc_lock); 1148 TAILQ_INIT(&pv_freepages); 1149 TAILQ_INIT(&pv_unusedpgs); 1150 1151 /* 1152 * compute the number of pages we have and then allocate RAM 1153 * for each pages' pv_head and saved attributes. 1154 */ 1155 { 1156 int npages, lcv; 1157 vsize_t s; 1158 1159 npages = 0; 1160 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) 1161 npages += (vm_physmem[lcv].end - vm_physmem[lcv].start); 1162 s = (vsize_t) (sizeof(struct pv_head) * npages + 1163 sizeof(char) * npages); 1164 s = round_page(s); /* round up */ 1165 boot_head = (char *)uvm_pageboot_alloc(s); 1166 bzero((char *)boot_head, s); 1167 if (boot_head == 0) 1168 panic("pmap_init: unable to allocate pv_heads"); 1169 } 1170 1171 /* 1172 * initialize the pmap pool. 1173 */ 1174 1175 pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl", 1176 0, pool_page_alloc_nointr, pool_page_free_nointr, M_VMPMAP); 1177 1178 cpu_cache_cleanD(); 1179 } 1180 1181 /* 1182 * void pmap_init(void) 1183 * 1184 * Initialize the pmap module. 1185 * Called by vm_init() in vm/vm_init.c in order to initialise 1186 * any structures that the pmap system needs to map virtual memory. 1187 */ 1188 1189 extern int physmem; 1190 1191 void 1192 pmap_init() 1193 { 1194 int lcv, i; 1195 1196 #ifdef MYCROFT_HACK 1197 printf("physmem = %d\n", physmem); 1198 #endif 1199 1200 /* 1201 * Set the available memory vars - These do not map to real memory 1202 * addresses and cannot as the physical memory is fragmented. 1203 * They are used by ps for %mem calculations. 1204 * One could argue whether this should be the entire memory or just 1205 * the memory that is useable in a user process. 1206 */ 1207 avail_start = 0; 1208 avail_end = physmem * NBPG; 1209 1210 /* allocate pv_head stuff first */ 1211 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) { 1212 vm_physmem[lcv].pmseg.pvhead = (struct pv_head *)boot_head; 1213 boot_head = (char *)(vaddr_t)(vm_physmem[lcv].pmseg.pvhead + 1214 (vm_physmem[lcv].end - vm_physmem[lcv].start)); 1215 for (i = 0; 1216 i < (vm_physmem[lcv].end - vm_physmem[lcv].start); i++) { 1217 simple_lock_init( 1218 &vm_physmem[lcv].pmseg.pvhead[i].pvh_lock); 1219 } 1220 } 1221 1222 /* now allocate attrs */ 1223 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) { 1224 vm_physmem[lcv].pmseg.attrs = (char *) boot_head; 1225 boot_head = (char *)(vaddr_t)(vm_physmem[lcv].pmseg.attrs + 1226 (vm_physmem[lcv].end - vm_physmem[lcv].start)); 1227 } 1228 1229 /* 1230 * now we need to free enough pv_entry structures to allow us to get 1231 * the kmem_map/kmem_object allocated and inited (done after this 1232 * function is finished). to do this we allocate one bootstrap page out 1233 * of kernel_map and use it to provide an initial pool of pv_entry 1234 * structures. we never free this page. 1235 */ 1236 1237 pv_initpage = (struct pv_page *) uvm_km_alloc(kernel_map, PAGE_SIZE); 1238 if (pv_initpage == NULL) 1239 panic("pmap_init: pv_initpage"); 1240 pv_cachedva = 0; /* a VA we have allocated but not used yet */ 1241 pv_nfpvents = 0; 1242 (void) pmap_add_pvpage(pv_initpage, FALSE); 1243 1244 #ifdef MYCROFT_HACK 1245 for (lcv = 0 ; lcv < vm_nphysseg ; lcv++) { 1246 printf("physseg[%d] pvent=%p attrs=%p start=%ld end=%ld\n", 1247 lcv, 1248 vm_physmem[lcv].pmseg.pvent, vm_physmem[lcv].pmseg.attrs, 1249 vm_physmem[lcv].start, vm_physmem[lcv].end); 1250 } 1251 #endif 1252 pmap_initialized = TRUE; 1253 1254 /* Initialise our L1 page table queues and counters */ 1255 SIMPLEQ_INIT(&l1pt_static_queue); 1256 l1pt_static_queue_count = 0; 1257 l1pt_static_create_count = 0; 1258 SIMPLEQ_INIT(&l1pt_queue); 1259 l1pt_queue_count = 0; 1260 l1pt_create_count = 0; 1261 l1pt_reuse_count = 0; 1262 } 1263 1264 /* 1265 * pmap_postinit() 1266 * 1267 * This routine is called after the vm and kmem subsystems have been 1268 * initialised. This allows the pmap code to perform any initialisation 1269 * that can only be done one the memory allocation is in place. 1270 */ 1271 1272 void 1273 pmap_postinit() 1274 { 1275 int loop; 1276 struct l1pt *pt; 1277 1278 #ifdef PMAP_STATIC_L1S 1279 for (loop = 0; loop < PMAP_STATIC_L1S; ++loop) { 1280 #else /* PMAP_STATIC_L1S */ 1281 for (loop = 0; loop < max_processes; ++loop) { 1282 #endif /* PMAP_STATIC_L1S */ 1283 /* Allocate a L1 page table */ 1284 pt = pmap_alloc_l1pt(); 1285 if (!pt) 1286 panic("Cannot allocate static L1 page tables\n"); 1287 1288 /* Clean it */ 1289 bzero((void *)pt->pt_va, PD_SIZE); 1290 pt->pt_flags |= (PTFLAG_STATIC | PTFLAG_CLEAN); 1291 /* Add the page table to the queue */ 1292 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pt, pt_queue); 1293 ++l1pt_static_queue_count; 1294 ++l1pt_static_create_count; 1295 } 1296 } 1297 1298 1299 /* 1300 * Create and return a physical map. 1301 * 1302 * If the size specified for the map is zero, the map is an actual physical 1303 * map, and may be referenced by the hardware. 1304 * 1305 * If the size specified is non-zero, the map will be used in software only, 1306 * and is bounded by that size. 1307 */ 1308 1309 pmap_t 1310 pmap_create() 1311 { 1312 struct pmap *pmap; 1313 1314 /* 1315 * Fetch pmap entry from the pool 1316 */ 1317 1318 pmap = pool_get(&pmap_pmap_pool, PR_WAITOK); 1319 /* XXX is this really needed! */ 1320 memset(pmap, 0, sizeof(*pmap)); 1321 1322 simple_lock_init(&pmap->pm_obj.vmobjlock); 1323 pmap->pm_obj.pgops = NULL; /* currently not a mappable object */ 1324 TAILQ_INIT(&pmap->pm_obj.memq); 1325 pmap->pm_obj.uo_npages = 0; 1326 pmap->pm_obj.uo_refs = 1; 1327 pmap->pm_stats.wired_count = 0; 1328 pmap->pm_stats.resident_count = 1; 1329 1330 /* Now init the machine part of the pmap */ 1331 pmap_pinit(pmap); 1332 return(pmap); 1333 } 1334 1335 /* 1336 * pmap_alloc_l1pt() 1337 * 1338 * This routine allocates physical and virtual memory for a L1 page table 1339 * and wires it. 1340 * A l1pt structure is returned to describe the allocated page table. 1341 * 1342 * This routine is allowed to fail if the required memory cannot be allocated. 1343 * In this case NULL is returned. 1344 */ 1345 1346 struct l1pt * 1347 pmap_alloc_l1pt(void) 1348 { 1349 paddr_t pa; 1350 vaddr_t va; 1351 struct l1pt *pt; 1352 int error; 1353 struct vm_page *m; 1354 pt_entry_t *ptes; 1355 1356 /* Allocate virtual address space for the L1 page table */ 1357 va = uvm_km_valloc(kernel_map, PD_SIZE); 1358 if (va == 0) { 1359 #ifdef DIAGNOSTIC 1360 PDEBUG(0, 1361 printf("pmap: Cannot allocate pageable memory for L1\n")); 1362 #endif /* DIAGNOSTIC */ 1363 return(NULL); 1364 } 1365 1366 /* Allocate memory for the l1pt structure */ 1367 pt = (struct l1pt *)malloc(sizeof(struct l1pt), M_VMPMAP, M_WAITOK); 1368 1369 /* 1370 * Allocate pages from the VM system. 1371 */ 1372 TAILQ_INIT(&pt->pt_plist); 1373 error = uvm_pglistalloc(PD_SIZE, physical_start, physical_end, 1374 PD_SIZE, 0, &pt->pt_plist, 1, M_WAITOK); 1375 if (error) { 1376 #ifdef DIAGNOSTIC 1377 PDEBUG(0, 1378 printf("pmap: Cannot allocate physical mem for L1 (%d)\n", 1379 error)); 1380 #endif /* DIAGNOSTIC */ 1381 /* Release the resources we already have claimed */ 1382 free(pt, M_VMPMAP); 1383 uvm_km_free(kernel_map, va, PD_SIZE); 1384 return(NULL); 1385 } 1386 1387 /* Map our physical pages into our virtual space */ 1388 pt->pt_va = va; 1389 m = pt->pt_plist.tqh_first; 1390 ptes = pmap_map_ptes(pmap_kernel()); 1391 while (m && va < (pt->pt_va + PD_SIZE)) { 1392 pa = VM_PAGE_TO_PHYS(m); 1393 1394 pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE); 1395 1396 /* Revoke cacheability and bufferability */ 1397 /* XXX should be done better than this */ 1398 ptes[arm_byte_to_page(va)] &= ~(PT_C | PT_B); 1399 1400 va += NBPG; 1401 m = m->pageq.tqe_next; 1402 } 1403 pmap_unmap_ptes(pmap_kernel()); 1404 pmap_update(pmap_kernel()); 1405 1406 #ifdef DIAGNOSTIC 1407 if (m) 1408 panic("pmap_alloc_l1pt: pglist not empty\n"); 1409 #endif /* DIAGNOSTIC */ 1410 1411 pt->pt_flags = 0; 1412 return(pt); 1413 } 1414 1415 /* 1416 * Free a L1 page table previously allocated with pmap_alloc_l1pt(). 1417 */ 1418 static void 1419 pmap_free_l1pt(pt) 1420 struct l1pt *pt; 1421 { 1422 /* Separate the physical memory for the virtual space */ 1423 pmap_kremove(pt->pt_va, PD_SIZE); 1424 pmap_update(pmap_kernel()); 1425 1426 /* Return the physical memory */ 1427 uvm_pglistfree(&pt->pt_plist); 1428 1429 /* Free the virtual space */ 1430 uvm_km_free(kernel_map, pt->pt_va, PD_SIZE); 1431 1432 /* Free the l1pt structure */ 1433 free(pt, M_VMPMAP); 1434 } 1435 1436 /* 1437 * Allocate a page directory. 1438 * This routine will either allocate a new page directory from the pool 1439 * of L1 page tables currently held by the kernel or it will allocate 1440 * a new one via pmap_alloc_l1pt(). 1441 * It will then initialise the l1 page table for use. 1442 */ 1443 static int 1444 pmap_allocpagedir(pmap) 1445 struct pmap *pmap; 1446 { 1447 paddr_t pa; 1448 struct l1pt *pt; 1449 pt_entry_t *pte; 1450 1451 PDEBUG(0, printf("pmap_allocpagedir(%p)\n", pmap)); 1452 1453 /* Do we have any spare L1's lying around ? */ 1454 if (l1pt_static_queue_count) { 1455 --l1pt_static_queue_count; 1456 pt = l1pt_static_queue.sqh_first; 1457 SIMPLEQ_REMOVE_HEAD(&l1pt_static_queue, pt, pt_queue); 1458 } else if (l1pt_queue_count) { 1459 --l1pt_queue_count; 1460 pt = l1pt_queue.sqh_first; 1461 SIMPLEQ_REMOVE_HEAD(&l1pt_queue, pt, pt_queue); 1462 ++l1pt_reuse_count; 1463 } else { 1464 pt = pmap_alloc_l1pt(); 1465 if (!pt) 1466 return(ENOMEM); 1467 ++l1pt_create_count; 1468 } 1469 1470 /* Store the pointer to the l1 descriptor in the pmap. */ 1471 pmap->pm_l1pt = pt; 1472 1473 /* Get the physical address of the start of the l1 */ 1474 pa = VM_PAGE_TO_PHYS(pt->pt_plist.tqh_first); 1475 1476 /* Store the virtual address of the l1 in the pmap. */ 1477 pmap->pm_pdir = (pd_entry_t *)pt->pt_va; 1478 1479 /* Clean the L1 if it is dirty */ 1480 if (!(pt->pt_flags & PTFLAG_CLEAN)) 1481 bzero((void *)pmap->pm_pdir, (PD_SIZE - KERNEL_PD_SIZE)); 1482 1483 /* Do we already have the kernel mappings ? */ 1484 if (!(pt->pt_flags & PTFLAG_KPT)) { 1485 /* Duplicate the kernel mapping i.e. all mappings 0xf0000000+ */ 1486 1487 bcopy((char *)pmap_kernel()->pm_pdir + (PD_SIZE - KERNEL_PD_SIZE), 1488 (char *)pmap->pm_pdir + (PD_SIZE - KERNEL_PD_SIZE), 1489 KERNEL_PD_SIZE); 1490 pt->pt_flags |= PTFLAG_KPT; 1491 } 1492 1493 /* Allocate a page table to map all the page tables for this pmap */ 1494 1495 #ifdef DIAGNOSTIC 1496 if (pmap->pm_vptpt) { 1497 /* XXX What if we have one already ? */ 1498 panic("pmap_allocpagedir: have pt already\n"); 1499 } 1500 #endif /* DIAGNOSTIC */ 1501 pmap->pm_vptpt = uvm_km_zalloc(kernel_map, NBPG); 1502 if (pmap->pm_vptpt == 0) { 1503 pmap_freepagedir(pmap); 1504 return(ENOMEM); 1505 } 1506 1507 (void) pmap_extract(pmap_kernel(), pmap->pm_vptpt, &pmap->pm_pptpt); 1508 pmap->pm_pptpt &= PG_FRAME; 1509 /* Revoke cacheability and bufferability */ 1510 /* XXX should be done better than this */ 1511 pte = pmap_pte(pmap_kernel(), pmap->pm_vptpt); 1512 *pte = *pte & ~(PT_C | PT_B); 1513 1514 /* Wire in this page table */ 1515 pmap_map_in_l1(pmap, PROCESS_PAGE_TBLS_BASE, pmap->pm_pptpt, TRUE); 1516 1517 pt->pt_flags &= ~PTFLAG_CLEAN; /* L1 is dirty now */ 1518 1519 /* 1520 * Map the kernel page tables for 0xf0000000 + 1521 * into the page table used to map the 1522 * pmap's page tables 1523 */ 1524 bcopy((char *)(PROCESS_PAGE_TBLS_BASE 1525 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) 1526 + ((PD_SIZE - KERNEL_PD_SIZE) >> 2)), 1527 (char *)pmap->pm_vptpt + ((PD_SIZE - KERNEL_PD_SIZE) >> 2), 1528 (KERNEL_PD_SIZE >> 2)); 1529 1530 return(0); 1531 } 1532 1533 1534 /* 1535 * Initialize a preallocated and zeroed pmap structure, 1536 * such as one in a vmspace structure. 1537 */ 1538 1539 void 1540 pmap_pinit(pmap) 1541 struct pmap *pmap; 1542 { 1543 int backoff = 6; 1544 int retry = 10; 1545 1546 PDEBUG(0, printf("pmap_pinit(%p)\n", pmap)); 1547 1548 /* Keep looping until we succeed in allocating a page directory */ 1549 while (pmap_allocpagedir(pmap) != 0) { 1550 /* 1551 * Ok we failed to allocate a suitable block of memory for an 1552 * L1 page table. This means that either: 1553 * 1. 16KB of virtual address space could not be allocated 1554 * 2. 16KB of physically contiguous memory on a 16KB boundary 1555 * could not be allocated. 1556 * 1557 * Since we cannot fail we will sleep for a while and try 1558 * again. 1559 * 1560 * Searching for a suitable L1 PT is expensive: 1561 * to avoid hogging the system when memory is really 1562 * scarce, use an exponential back-off so that 1563 * eventually we won't retry more than once every 8 1564 * seconds. This should allow other processes to run 1565 * to completion and free up resources. 1566 */ 1567 (void) ltsleep(&lbolt, PVM, "l1ptwait", (hz << 3) >> backoff, 1568 NULL); 1569 if (--retry == 0) { 1570 retry = 10; 1571 if (backoff) 1572 --backoff; 1573 } 1574 } 1575 1576 /* Map zero page for the pmap. This will also map the L2 for it */ 1577 pmap_enter(pmap, 0x00000000, systempage.pv_pa, 1578 VM_PROT_READ, VM_PROT_READ | PMAP_WIRED); 1579 pmap_update(pmap); 1580 } 1581 1582 1583 void 1584 pmap_freepagedir(pmap) 1585 struct pmap *pmap; 1586 { 1587 /* Free the memory used for the page table mapping */ 1588 if (pmap->pm_vptpt != 0) 1589 uvm_km_free(kernel_map, (vaddr_t)pmap->pm_vptpt, NBPG); 1590 1591 /* junk the L1 page table */ 1592 if (pmap->pm_l1pt->pt_flags & PTFLAG_STATIC) { 1593 /* Add the page table to the queue */ 1594 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pmap->pm_l1pt, pt_queue); 1595 ++l1pt_static_queue_count; 1596 } else if (l1pt_queue_count < 8) { 1597 /* Add the page table to the queue */ 1598 SIMPLEQ_INSERT_TAIL(&l1pt_queue, pmap->pm_l1pt, pt_queue); 1599 ++l1pt_queue_count; 1600 } else 1601 pmap_free_l1pt(pmap->pm_l1pt); 1602 } 1603 1604 1605 /* 1606 * Retire the given physical map from service. 1607 * Should only be called if the map contains no valid mappings. 1608 */ 1609 1610 void 1611 pmap_destroy(pmap) 1612 struct pmap *pmap; 1613 { 1614 struct vm_page *page; 1615 int count; 1616 1617 if (pmap == NULL) 1618 return; 1619 1620 PDEBUG(0, printf("pmap_destroy(%p)\n", pmap)); 1621 1622 /* 1623 * Drop reference count 1624 */ 1625 simple_lock(&pmap->pm_obj.vmobjlock); 1626 count = --pmap->pm_obj.uo_refs; 1627 simple_unlock(&pmap->pm_obj.vmobjlock); 1628 if (count > 0) { 1629 return; 1630 } 1631 1632 /* 1633 * reference count is zero, free pmap resources and then free pmap. 1634 */ 1635 1636 /* Remove the zero page mapping */ 1637 pmap_remove(pmap, 0x00000000, 0x00000000 + NBPG); 1638 pmap_update(pmap); 1639 1640 /* 1641 * Free any page tables still mapped 1642 * This is only temporay until pmap_enter can count the number 1643 * of mappings made in a page table. Then pmap_remove() can 1644 * reduce the count and free the pagetable when the count 1645 * reaches zero. Note that entries in this list should match the 1646 * contents of the ptpt, however this is faster than walking a 1024 1647 * entries looking for pt's 1648 * taken from i386 pmap.c 1649 */ 1650 while (pmap->pm_obj.memq.tqh_first != NULL) { 1651 page = pmap->pm_obj.memq.tqh_first; 1652 #ifdef DIAGNOSTIC 1653 if (page->flags & PG_BUSY) 1654 panic("pmap_release: busy page table page"); 1655 #endif 1656 /* pmap_page_protect? currently no need for it. */ 1657 1658 page->wire_count = 0; 1659 uvm_pagefree(page); 1660 } 1661 1662 /* Free the page dir */ 1663 pmap_freepagedir(pmap); 1664 1665 /* return the pmap to the pool */ 1666 pool_put(&pmap_pmap_pool, pmap); 1667 } 1668 1669 1670 /* 1671 * void pmap_reference(struct pmap *pmap) 1672 * 1673 * Add a reference to the specified pmap. 1674 */ 1675 1676 void 1677 pmap_reference(pmap) 1678 struct pmap *pmap; 1679 { 1680 if (pmap == NULL) 1681 return; 1682 1683 simple_lock(&pmap->pm_lock); 1684 pmap->pm_obj.uo_refs++; 1685 simple_unlock(&pmap->pm_lock); 1686 } 1687 1688 /* 1689 * void pmap_virtual_space(vaddr_t *start, vaddr_t *end) 1690 * 1691 * Return the start and end addresses of the kernel's virtual space. 1692 * These values are setup in pmap_bootstrap and are updated as pages 1693 * are allocated. 1694 */ 1695 1696 void 1697 pmap_virtual_space(start, end) 1698 vaddr_t *start; 1699 vaddr_t *end; 1700 { 1701 *start = virtual_start; 1702 *end = virtual_end; 1703 } 1704 1705 1706 /* 1707 * Activate the address space for the specified process. If the process 1708 * is the current process, load the new MMU context. 1709 */ 1710 void 1711 pmap_activate(p) 1712 struct proc *p; 1713 { 1714 struct pmap *pmap = p->p_vmspace->vm_map.pmap; 1715 struct pcb *pcb = &p->p_addr->u_pcb; 1716 1717 (void) pmap_extract(pmap_kernel(), (vaddr_t)pmap->pm_pdir, 1718 (paddr_t *)&pcb->pcb_pagedir); 1719 1720 PDEBUG(0, printf("pmap_activate: p=%p pmap=%p pcb=%p pdir=%p l1=%p\n", 1721 p, pmap, pcb, pmap->pm_pdir, pcb->pcb_pagedir)); 1722 1723 if (p == curproc) { 1724 PDEBUG(0, printf("pmap_activate: setting TTB\n")); 1725 setttb((u_int)pcb->pcb_pagedir); 1726 } 1727 #if 0 1728 pmap->pm_pdchanged = FALSE; 1729 #endif 1730 } 1731 1732 1733 /* 1734 * Deactivate the address space of the specified process. 1735 */ 1736 void 1737 pmap_deactivate(p) 1738 struct proc *p; 1739 { 1740 } 1741 1742 /* 1743 * Perform any deferred pmap operations. 1744 */ 1745 void 1746 pmap_update(struct pmap *pmap) 1747 { 1748 1749 /* 1750 * We haven't deferred any pmap operations, but we do need to 1751 * make sure TLB/cache operations have completed. 1752 */ 1753 cpu_cpwait(); 1754 } 1755 1756 /* 1757 * pmap_clean_page() 1758 * 1759 * This is a local function used to work out the best strategy to clean 1760 * a single page referenced by its entry in the PV table. It's used by 1761 * pmap_copy_page, pmap_zero page and maybe some others later on. 1762 * 1763 * Its policy is effectively: 1764 * o If there are no mappings, we don't bother doing anything with the cache. 1765 * o If there is one mapping, we clean just that page. 1766 * o If there are multiple mappings, we clean the entire cache. 1767 * 1768 * So that some functions can be further optimised, it returns 0 if it didn't 1769 * clean the entire cache, or 1 if it did. 1770 * 1771 * XXX One bug in this routine is that if the pv_entry has a single page 1772 * mapped at 0x00000000 a whole cache clean will be performed rather than 1773 * just the 1 page. Since this should not occur in everyday use and if it does 1774 * it will just result in not the most efficient clean for the page. 1775 */ 1776 static int 1777 pmap_clean_page(pv, is_src) 1778 struct pv_entry *pv; 1779 boolean_t is_src; 1780 { 1781 struct pmap *pmap; 1782 struct pv_entry *npv; 1783 int cache_needs_cleaning = 0; 1784 vaddr_t page_to_clean = 0; 1785 1786 if (pv == NULL) 1787 /* nothing mapped in so nothing to flush */ 1788 return (0); 1789 1790 /* Since we flush the cache each time we change curproc, we 1791 * only need to flush the page if it is in the current pmap. 1792 */ 1793 if (curproc) 1794 pmap = curproc->p_vmspace->vm_map.pmap; 1795 else 1796 pmap = pmap_kernel(); 1797 1798 for (npv = pv; npv; npv = npv->pv_next) { 1799 if (npv->pv_pmap == pmap) { 1800 /* The page is mapped non-cacheable in 1801 * this map. No need to flush the cache. 1802 */ 1803 if (npv->pv_flags & PT_NC) { 1804 #ifdef DIAGNOSTIC 1805 if (cache_needs_cleaning) 1806 panic("pmap_clean_page: " 1807 "cache inconsistency"); 1808 #endif 1809 break; 1810 } 1811 #if 0 1812 /* This doesn't work, because pmap_protect 1813 doesn't flush changes on pages that it 1814 has write-protected. */ 1815 1816 /* If the page is not writable and this 1817 is the source, then there is no need 1818 to flush it from the cache. */ 1819 else if (is_src && ! (npv->pv_flags & PT_Wr)) 1820 continue; 1821 #endif 1822 if (cache_needs_cleaning){ 1823 page_to_clean = 0; 1824 break; 1825 } 1826 else 1827 page_to_clean = npv->pv_va; 1828 cache_needs_cleaning = 1; 1829 } 1830 } 1831 1832 if (page_to_clean) 1833 cpu_cache_purgeID_rng(page_to_clean, NBPG); 1834 else if (cache_needs_cleaning) { 1835 cpu_cache_purgeID(); 1836 return (1); 1837 } 1838 return (0); 1839 } 1840 1841 /* 1842 * pmap_find_pv() 1843 * 1844 * This is a local function that finds a PV head for a given physical page. 1845 * This is a common op, and this function removes loads of ifdefs in the code. 1846 */ 1847 static __inline struct pv_head * 1848 pmap_find_pvh(phys) 1849 paddr_t phys; 1850 { 1851 int bank, off; 1852 struct pv_head *pvh; 1853 1854 if ((bank = vm_physseg_find(atop(phys), &off)) == -1) 1855 panic("pmap_find_pv: not a real page, phys=%lx\n", phys); 1856 pvh = &vm_physmem[bank].pmseg.pvhead[off]; 1857 return (pvh); 1858 } 1859 1860 /* 1861 * pmap_zero_page() 1862 * 1863 * Zero a given physical page by mapping it at a page hook point. 1864 * In doing the zero page op, the page we zero is mapped cachable, as with 1865 * StrongARM accesses to non-cached pages are non-burst making writing 1866 * _any_ bulk data very slow. 1867 */ 1868 void 1869 pmap_zero_page(phys) 1870 paddr_t phys; 1871 { 1872 struct pv_head *pvh; 1873 1874 /* Get an entry for this page, and clean it it. */ 1875 pvh = pmap_find_pvh(phys); 1876 simple_lock(&pvh->pvh_lock); 1877 pmap_clean_page(pvh->pvh_list, FALSE); 1878 simple_unlock(&pvh->pvh_lock); 1879 1880 /* 1881 * Hook in the page, zero it, and purge the cache for that 1882 * zeroed page. Invalidate the TLB as needed. 1883 */ 1884 *page_hook0.pte = L2_PTE(phys & PG_FRAME, AP_KRW); 1885 cpu_tlb_flushD_SE(page_hook0.va); 1886 cpu_cpwait(); 1887 bzero_page(page_hook0.va); 1888 cpu_cache_purgeD_rng(page_hook0.va, NBPG); 1889 } 1890 1891 /* pmap_pageidlezero() 1892 * 1893 * The same as above, except that we assume that the page is not 1894 * mapped. This means we never have to flush the cache first. Called 1895 * from the idle loop. 1896 */ 1897 boolean_t 1898 pmap_pageidlezero(phys) 1899 paddr_t phys; 1900 { 1901 int i, *ptr; 1902 boolean_t rv = TRUE; 1903 1904 #ifdef DIAGNOSTIC 1905 struct pv_head *pvh; 1906 1907 pvh = pmap_find_pvh(phys); 1908 if (pvh->pvh_list != NULL) 1909 panic("pmap_pageidlezero: zeroing mapped page\n"); 1910 #endif 1911 1912 /* 1913 * Hook in the page, zero it, and purge the cache for that 1914 * zeroed page. Invalidate the TLB as needed. 1915 */ 1916 *page_hook0.pte = L2_PTE(phys & PG_FRAME, AP_KRW); 1917 cpu_tlb_flushD_SE(page_hook0.va); 1918 cpu_cpwait(); 1919 1920 for (i = 0, ptr = (int *)page_hook0.va; 1921 i < (NBPG / sizeof(int)); i++) { 1922 if (sched_whichqs != 0) { 1923 /* 1924 * A process has become ready. Abort now, 1925 * so we don't keep it waiting while we 1926 * do slow memory access to finish this 1927 * page. 1928 */ 1929 rv = FALSE; 1930 break; 1931 } 1932 *ptr++ = 0; 1933 } 1934 1935 if (rv) 1936 /* 1937 * if we aborted we'll rezero this page again later so don't 1938 * purge it unless we finished it 1939 */ 1940 cpu_cache_purgeD_rng(page_hook0.va, NBPG); 1941 return (rv); 1942 } 1943 1944 /* 1945 * pmap_copy_page() 1946 * 1947 * Copy one physical page into another, by mapping the pages into 1948 * hook points. The same comment regarding cachability as in 1949 * pmap_zero_page also applies here. 1950 */ 1951 void 1952 pmap_copy_page(src, dest) 1953 paddr_t src; 1954 paddr_t dest; 1955 { 1956 struct pv_head *src_pvh, *dest_pvh; 1957 boolean_t cleanedcache; 1958 1959 /* Get PV entries for the pages, and clean them if needed. */ 1960 src_pvh = pmap_find_pvh(src); 1961 1962 simple_lock(&src_pvh->pvh_lock); 1963 cleanedcache = pmap_clean_page(src_pvh->pvh_list, TRUE); 1964 simple_unlock(&src_pvh->pvh_lock); 1965 1966 if (cleanedcache == 0) { 1967 dest_pvh = pmap_find_pvh(dest); 1968 simple_lock(&dest_pvh->pvh_lock); 1969 pmap_clean_page(dest_pvh->pvh_list, FALSE); 1970 simple_unlock(&dest_pvh->pvh_lock); 1971 } 1972 /* 1973 * Map the pages into the page hook points, copy them, and purge 1974 * the cache for the appropriate page. Invalidate the TLB 1975 * as required. 1976 */ 1977 *page_hook0.pte = L2_PTE(src & PG_FRAME, AP_KRW); 1978 *page_hook1.pte = L2_PTE(dest & PG_FRAME, AP_KRW); 1979 cpu_tlb_flushD_SE(page_hook0.va); 1980 cpu_tlb_flushD_SE(page_hook1.va); 1981 cpu_cpwait(); 1982 bcopy_page(page_hook0.va, page_hook1.va); 1983 cpu_cache_purgeD_rng(page_hook0.va, NBPG); 1984 cpu_cache_purgeD_rng(page_hook1.va, NBPG); 1985 } 1986 1987 #if 0 1988 void 1989 pmap_pte_addref(pmap, va) 1990 struct pmap *pmap; 1991 vaddr_t va; 1992 { 1993 pd_entry_t *pde; 1994 paddr_t pa; 1995 struct vm_page *m; 1996 1997 if (pmap == pmap_kernel()) 1998 return; 1999 2000 pde = pmap_pde(pmap, va & ~(3 << PDSHIFT)); 2001 pa = pmap_pte_pa(pde); 2002 m = PHYS_TO_VM_PAGE(pa); 2003 ++m->wire_count; 2004 #ifdef MYCROFT_HACK 2005 printf("addref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n", 2006 pmap, va, pde, pa, m, m->wire_count); 2007 #endif 2008 } 2009 2010 void 2011 pmap_pte_delref(pmap, va) 2012 struct pmap *pmap; 2013 vaddr_t va; 2014 { 2015 pd_entry_t *pde; 2016 paddr_t pa; 2017 struct vm_page *m; 2018 2019 if (pmap == pmap_kernel()) 2020 return; 2021 2022 pde = pmap_pde(pmap, va & ~(3 << PDSHIFT)); 2023 pa = pmap_pte_pa(pde); 2024 m = PHYS_TO_VM_PAGE(pa); 2025 --m->wire_count; 2026 #ifdef MYCROFT_HACK 2027 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n", 2028 pmap, va, pde, pa, m, m->wire_count); 2029 #endif 2030 if (m->wire_count == 0) { 2031 #ifdef MYCROFT_HACK 2032 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p\n", 2033 pmap, va, pde, pa, m); 2034 #endif 2035 pmap_unmap_in_l1(pmap, va); 2036 uvm_pagefree(m); 2037 --pmap->pm_stats.resident_count; 2038 } 2039 } 2040 #else 2041 #define pmap_pte_addref(pmap, va) 2042 #define pmap_pte_delref(pmap, va) 2043 #endif 2044 2045 /* 2046 * Since we have a virtually indexed cache, we may need to inhibit caching if 2047 * there is more than one mapping and at least one of them is writable. 2048 * Since we purge the cache on every context switch, we only need to check for 2049 * other mappings within the same pmap, or kernel_pmap. 2050 * This function is also called when a page is unmapped, to possibly reenable 2051 * caching on any remaining mappings. 2052 * 2053 * The code implements the following logic, where: 2054 * 2055 * KW = # of kernel read/write pages 2056 * KR = # of kernel read only pages 2057 * UW = # of user read/write pages 2058 * UR = # of user read only pages 2059 * OW = # of user read/write pages in another pmap, then 2060 * 2061 * KC = kernel mapping is cacheable 2062 * UC = user mapping is cacheable 2063 * 2064 * KW=0,KR=0 KW=0,KR>0 KW=1,KR=0 KW>1,KR>=0 2065 * +--------------------------------------------- 2066 * UW=0,UR=0,OW=0 | --- KC=1 KC=1 KC=0 2067 * UW=0,UR>0,OW=0 | UC=1 KC=1,UC=1 KC=0,UC=0 KC=0,UC=0 2068 * UW=0,UR>0,OW>0 | UC=1 KC=0,UC=1 KC=0,UC=0 KC=0,UC=0 2069 * UW=1,UR=0,OW=0 | UC=1 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0 2070 * UW>1,UR>=0,OW>=0 | UC=0 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0 2071 * 2072 * Note that the pmap must have it's ptes mapped in, and passed with ptes. 2073 */ 2074 __inline static void 2075 pmap_vac_me_harder(struct pmap *pmap, struct pv_head *pvh, pt_entry_t *ptes, 2076 boolean_t clear_cache) 2077 { 2078 if (pmap == pmap_kernel()) 2079 pmap_vac_me_kpmap(pmap, pvh, ptes, clear_cache); 2080 else 2081 pmap_vac_me_user(pmap, pvh, ptes, clear_cache); 2082 } 2083 2084 static void 2085 pmap_vac_me_kpmap(struct pmap *pmap, struct pv_head *pvh, pt_entry_t *ptes, 2086 boolean_t clear_cache) 2087 { 2088 int user_entries = 0; 2089 int user_writable = 0; 2090 int user_cacheable = 0; 2091 int kernel_entries = 0; 2092 int kernel_writable = 0; 2093 int kernel_cacheable = 0; 2094 struct pv_entry *pv; 2095 struct pmap *last_pmap = pmap; 2096 2097 #ifdef DIAGNOSTIC 2098 if (pmap != pmap_kernel()) 2099 panic("pmap_vac_me_kpmap: pmap != pmap_kernel()"); 2100 #endif 2101 2102 /* 2103 * Pass one, see if there are both kernel and user pmaps for 2104 * this page. Calculate whether there are user-writable or 2105 * kernel-writable pages. 2106 */ 2107 for (pv = pvh->pvh_list; pv != NULL; pv = pv->pv_next) { 2108 if (pv->pv_pmap != pmap) { 2109 user_entries++; 2110 if (pv->pv_flags & PT_Wr) 2111 user_writable++; 2112 if ((pv->pv_flags & PT_NC) == 0) 2113 user_cacheable++; 2114 } else { 2115 kernel_entries++; 2116 if (pv->pv_flags & PT_Wr) 2117 kernel_writable++; 2118 if ((pv->pv_flags & PT_NC) == 0) 2119 kernel_cacheable++; 2120 } 2121 } 2122 2123 /* 2124 * We know we have just been updating a kernel entry, so if 2125 * all user pages are already cacheable, then there is nothing 2126 * further to do. 2127 */ 2128 if (kernel_entries == 0 && 2129 user_cacheable == user_entries) 2130 return; 2131 2132 if (user_entries) { 2133 /* 2134 * Scan over the list again, for each entry, if it 2135 * might not be set correctly, call pmap_vac_me_user 2136 * to recalculate the settings. 2137 */ 2138 for (pv = pvh->pvh_list; pv; pv = pv->pv_next) { 2139 /* 2140 * We know kernel mappings will get set 2141 * correctly in other calls. We also know 2142 * that if the pmap is the same as last_pmap 2143 * then we've just handled this entry. 2144 */ 2145 if (pv->pv_pmap == pmap || pv->pv_pmap == last_pmap) 2146 continue; 2147 /* 2148 * If there are kernel entries and this page 2149 * is writable but non-cacheable, then we can 2150 * skip this entry also. 2151 */ 2152 if (kernel_entries > 0 && 2153 (pv->pv_flags & (PT_NC | PT_Wr)) == 2154 (PT_NC | PT_Wr)) 2155 continue; 2156 /* 2157 * Similarly if there are no kernel-writable 2158 * entries and the page is already 2159 * read-only/cacheable. 2160 */ 2161 if (kernel_writable == 0 && 2162 (pv->pv_flags & (PT_NC | PT_Wr)) == 0) 2163 continue; 2164 /* 2165 * For some of the remaining cases, we know 2166 * that we must recalculate, but for others we 2167 * can't tell if they are correct or not, so 2168 * we recalculate anyway. 2169 */ 2170 pmap_unmap_ptes(last_pmap); 2171 last_pmap = pv->pv_pmap; 2172 ptes = pmap_map_ptes(last_pmap); 2173 pmap_vac_me_user(last_pmap, pvh, ptes, 2174 pmap_is_curpmap(last_pmap)); 2175 } 2176 /* Restore the pte mapping that was passed to us. */ 2177 if (last_pmap != pmap) { 2178 pmap_unmap_ptes(last_pmap); 2179 ptes = pmap_map_ptes(pmap); 2180 } 2181 if (kernel_entries == 0) 2182 return; 2183 } 2184 2185 pmap_vac_me_user(pmap, pvh, ptes, clear_cache); 2186 return; 2187 } 2188 2189 static void 2190 pmap_vac_me_user(struct pmap *pmap, struct pv_head *pvh, pt_entry_t *ptes, 2191 boolean_t clear_cache) 2192 { 2193 struct pmap *kpmap = pmap_kernel(); 2194 struct pv_entry *pv, *npv; 2195 int entries = 0; 2196 int writable = 0; 2197 int cacheable_entries = 0; 2198 int kern_cacheable = 0; 2199 int other_writable = 0; 2200 2201 pv = pvh->pvh_list; 2202 KASSERT(ptes != NULL); 2203 2204 /* 2205 * Count mappings and writable mappings in this pmap. 2206 * Include kernel mappings as part of our own. 2207 * Keep a pointer to the first one. 2208 */ 2209 for (npv = pv; npv; npv = npv->pv_next) { 2210 /* Count mappings in the same pmap */ 2211 if (pmap == npv->pv_pmap || 2212 kpmap == npv->pv_pmap) { 2213 if (entries++ == 0) 2214 pv = npv; 2215 /* Cacheable mappings */ 2216 if ((npv->pv_flags & PT_NC) == 0) { 2217 cacheable_entries++; 2218 if (kpmap == npv->pv_pmap) 2219 kern_cacheable++; 2220 } 2221 /* Writable mappings */ 2222 if (npv->pv_flags & PT_Wr) 2223 ++writable; 2224 } else if (npv->pv_flags & PT_Wr) 2225 other_writable = 1; 2226 } 2227 2228 PDEBUG(3,printf("pmap_vac_me_harder: pmap %p Entries %d, " 2229 "writable %d cacheable %d %s\n", pmap, entries, writable, 2230 cacheable_entries, clear_cache ? "clean" : "no clean")); 2231 2232 /* 2233 * Enable or disable caching as necessary. 2234 * Note: the first entry might be part of the kernel pmap, 2235 * so we can't assume this is indicative of the state of the 2236 * other (maybe non-kpmap) entries. 2237 */ 2238 if ((entries > 1 && writable) || 2239 (entries > 0 && pmap == kpmap && other_writable)) { 2240 if (cacheable_entries == 0) 2241 return; 2242 for (npv = pv; npv; npv = npv->pv_next) { 2243 if ((pmap == npv->pv_pmap 2244 || kpmap == npv->pv_pmap) && 2245 (npv->pv_flags & PT_NC) == 0) { 2246 ptes[arm_byte_to_page(npv->pv_va)] &= 2247 ~(PT_C | PT_B); 2248 npv->pv_flags |= PT_NC; 2249 /* 2250 * If this page needs flushing from the 2251 * cache, and we aren't going to do it 2252 * below, do it now. 2253 */ 2254 if ((cacheable_entries < 4 && 2255 (clear_cache || npv->pv_pmap == kpmap)) || 2256 (npv->pv_pmap == kpmap && 2257 !clear_cache && kern_cacheable < 4)) { 2258 cpu_cache_purgeID_rng(npv->pv_va, 2259 NBPG); 2260 cpu_tlb_flushID_SE(npv->pv_va); 2261 } 2262 } 2263 } 2264 if ((clear_cache && cacheable_entries >= 4) || 2265 kern_cacheable >= 4) { 2266 cpu_cache_purgeID(); 2267 cpu_tlb_flushID(); 2268 } 2269 cpu_cpwait(); 2270 } else if (entries > 0) { 2271 /* 2272 * Turn cacheing back on for some pages. If it is a kernel 2273 * page, only do so if there are no other writable pages. 2274 */ 2275 for (npv = pv; npv; npv = npv->pv_next) { 2276 if ((pmap == npv->pv_pmap || 2277 (kpmap == npv->pv_pmap && other_writable == 0)) && 2278 (npv->pv_flags & PT_NC)) { 2279 ptes[arm_byte_to_page(npv->pv_va)] |= 2280 pte_cache_mode; 2281 npv->pv_flags &= ~PT_NC; 2282 } 2283 } 2284 } 2285 } 2286 2287 /* 2288 * pmap_remove() 2289 * 2290 * pmap_remove is responsible for nuking a number of mappings for a range 2291 * of virtual address space in the current pmap. To do this efficiently 2292 * is interesting, because in a number of cases a wide virtual address 2293 * range may be supplied that contains few actual mappings. So, the 2294 * optimisations are: 2295 * 1. Try and skip over hunks of address space for which an L1 entry 2296 * does not exist. 2297 * 2. Build up a list of pages we've hit, up to a maximum, so we can 2298 * maybe do just a partial cache clean. This path of execution is 2299 * complicated by the fact that the cache must be flushed _before_ 2300 * the PTE is nuked, being a VAC :-) 2301 * 3. Maybe later fast-case a single page, but I don't think this is 2302 * going to make _that_ much difference overall. 2303 */ 2304 2305 #define PMAP_REMOVE_CLEAN_LIST_SIZE 3 2306 2307 void 2308 pmap_remove(pmap, sva, eva) 2309 struct pmap *pmap; 2310 vaddr_t sva; 2311 vaddr_t eva; 2312 { 2313 int cleanlist_idx = 0; 2314 struct pagelist { 2315 vaddr_t va; 2316 pt_entry_t *pte; 2317 } cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE]; 2318 pt_entry_t *pte = 0, *ptes; 2319 paddr_t pa; 2320 int pmap_active; 2321 struct pv_head *pvh; 2322 2323 /* Exit quick if there is no pmap */ 2324 if (!pmap) 2325 return; 2326 2327 PDEBUG(0, printf("pmap_remove: pmap=%p sva=%08lx eva=%08lx\n", pmap, sva, eva)); 2328 2329 sva &= PG_FRAME; 2330 eva &= PG_FRAME; 2331 2332 /* 2333 * we lock in the pmap => pv_head direction 2334 */ 2335 PMAP_MAP_TO_HEAD_LOCK(); 2336 2337 ptes = pmap_map_ptes(pmap); 2338 /* Get a page table pointer */ 2339 while (sva < eva) { 2340 if (pmap_pde_page(pmap_pde(pmap, sva))) 2341 break; 2342 sva = (sva & PD_MASK) + NBPD; 2343 } 2344 2345 pte = &ptes[arm_byte_to_page(sva)]; 2346 /* Note if the pmap is active thus require cache and tlb cleans */ 2347 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap) 2348 || (pmap == pmap_kernel())) 2349 pmap_active = 1; 2350 else 2351 pmap_active = 0; 2352 2353 /* Now loop along */ 2354 while (sva < eva) { 2355 /* Check if we can move to the next PDE (l1 chunk) */ 2356 if (!(sva & PT_MASK)) 2357 if (!pmap_pde_page(pmap_pde(pmap, sva))) { 2358 sva += NBPD; 2359 pte += arm_byte_to_page(NBPD); 2360 continue; 2361 } 2362 2363 /* We've found a valid PTE, so this page of PTEs has to go. */ 2364 if (pmap_pte_v(pte)) { 2365 int bank, off; 2366 2367 /* Update statistics */ 2368 --pmap->pm_stats.resident_count; 2369 2370 /* 2371 * Add this page to our cache remove list, if we can. 2372 * If, however the cache remove list is totally full, 2373 * then do a complete cache invalidation taking note 2374 * to backtrack the PTE table beforehand, and ignore 2375 * the lists in future because there's no longer any 2376 * point in bothering with them (we've paid the 2377 * penalty, so will carry on unhindered). Otherwise, 2378 * when we fall out, we just clean the list. 2379 */ 2380 PDEBUG(10, printf("remove: inv pte at %p(%x) ", pte, *pte)); 2381 pa = pmap_pte_pa(pte); 2382 2383 if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) { 2384 /* Add to the clean list. */ 2385 cleanlist[cleanlist_idx].pte = pte; 2386 cleanlist[cleanlist_idx].va = sva; 2387 cleanlist_idx++; 2388 } else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) { 2389 int cnt; 2390 2391 /* Nuke everything if needed. */ 2392 if (pmap_active) { 2393 cpu_cache_purgeID(); 2394 cpu_tlb_flushID(); 2395 } 2396 2397 /* 2398 * Roll back the previous PTE list, 2399 * and zero out the current PTE. 2400 */ 2401 for (cnt = 0; cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) { 2402 *cleanlist[cnt].pte = 0; 2403 pmap_pte_delref(pmap, cleanlist[cnt].va); 2404 } 2405 *pte = 0; 2406 pmap_pte_delref(pmap, sva); 2407 cleanlist_idx++; 2408 } else { 2409 /* 2410 * We've already nuked the cache and 2411 * TLB, so just carry on regardless, 2412 * and we won't need to do it again 2413 */ 2414 *pte = 0; 2415 pmap_pte_delref(pmap, sva); 2416 } 2417 2418 /* 2419 * Update flags. In a number of circumstances, 2420 * we could cluster a lot of these and do a 2421 * number of sequential pages in one go. 2422 */ 2423 if ((bank = vm_physseg_find(atop(pa), &off)) != -1) { 2424 struct pv_entry *pve; 2425 pvh = &vm_physmem[bank].pmseg.pvhead[off]; 2426 simple_lock(&pvh->pvh_lock); 2427 pve = pmap_remove_pv(pvh, pmap, sva); 2428 pmap_free_pv(pmap, pve); 2429 pmap_vac_me_harder(pmap, pvh, ptes, FALSE); 2430 simple_unlock(&pvh->pvh_lock); 2431 } 2432 } 2433 sva += NBPG; 2434 pte++; 2435 } 2436 2437 pmap_unmap_ptes(pmap); 2438 /* 2439 * Now, if we've fallen through down to here, chances are that there 2440 * are less than PMAP_REMOVE_CLEAN_LIST_SIZE mappings left. 2441 */ 2442 if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) { 2443 u_int cnt; 2444 2445 for (cnt = 0; cnt < cleanlist_idx; cnt++) { 2446 if (pmap_active) { 2447 cpu_cache_purgeID_rng(cleanlist[cnt].va, NBPG); 2448 *cleanlist[cnt].pte = 0; 2449 cpu_tlb_flushID_SE(cleanlist[cnt].va); 2450 } else 2451 *cleanlist[cnt].pte = 0; 2452 pmap_pte_delref(pmap, cleanlist[cnt].va); 2453 } 2454 } 2455 PMAP_MAP_TO_HEAD_UNLOCK(); 2456 } 2457 2458 /* 2459 * Routine: pmap_remove_all 2460 * Function: 2461 * Removes this physical page from 2462 * all physical maps in which it resides. 2463 * Reflects back modify bits to the pager. 2464 */ 2465 2466 static void 2467 pmap_remove_all(pa) 2468 paddr_t pa; 2469 { 2470 struct pv_entry *pv, *npv; 2471 struct pv_head *pvh; 2472 struct pmap *pmap; 2473 pt_entry_t *pte, *ptes; 2474 2475 PDEBUG(0, printf("pmap_remove_all: pa=%lx ", pa)); 2476 2477 /* set pv_head => pmap locking */ 2478 PMAP_HEAD_TO_MAP_LOCK(); 2479 2480 pvh = pmap_find_pvh(pa); 2481 simple_lock(&pvh->pvh_lock); 2482 2483 pv = pvh->pvh_list; 2484 if (pv == NULL) 2485 { 2486 PDEBUG(0, printf("free page\n")); 2487 simple_unlock(&pvh->pvh_lock); 2488 PMAP_HEAD_TO_MAP_UNLOCK(); 2489 return; 2490 } 2491 pmap_clean_page(pv, FALSE); 2492 2493 while (pv) { 2494 pmap = pv->pv_pmap; 2495 ptes = pmap_map_ptes(pmap); 2496 pte = &ptes[arm_byte_to_page(pv->pv_va)]; 2497 2498 PDEBUG(0, printf("[%p,%08x,%08lx,%08x] ", pmap, *pte, 2499 pv->pv_va, pv->pv_flags)); 2500 #ifdef DEBUG 2501 if (!pmap_pde_page(pmap_pde(pmap, pv->pv_va)) || 2502 !pmap_pte_v(pte) || pmap_pte_pa(pte) != pa) 2503 panic("pmap_remove_all: bad mapping"); 2504 #endif /* DEBUG */ 2505 2506 /* 2507 * Update statistics 2508 */ 2509 --pmap->pm_stats.resident_count; 2510 2511 /* Wired bit */ 2512 if (pv->pv_flags & PT_W) 2513 --pmap->pm_stats.wired_count; 2514 2515 /* 2516 * Invalidate the PTEs. 2517 * XXX: should cluster them up and invalidate as many 2518 * as possible at once. 2519 */ 2520 2521 #ifdef needednotdone 2522 reduce wiring count on page table pages as references drop 2523 #endif 2524 2525 *pte = 0; 2526 pmap_pte_delref(pmap, pv->pv_va); 2527 2528 npv = pv->pv_next; 2529 pmap_free_pv(pmap, pv); 2530 pv = npv; 2531 pmap_unmap_ptes(pmap); 2532 } 2533 pvh->pvh_list = NULL; 2534 simple_unlock(&pvh->pvh_lock); 2535 PMAP_HEAD_TO_MAP_UNLOCK(); 2536 2537 PDEBUG(0, printf("done\n")); 2538 cpu_tlb_flushID(); 2539 cpu_cpwait(); 2540 } 2541 2542 2543 /* 2544 * Set the physical protection on the specified range of this map as requested. 2545 */ 2546 2547 void 2548 pmap_protect(pmap, sva, eva, prot) 2549 struct pmap *pmap; 2550 vaddr_t sva; 2551 vaddr_t eva; 2552 vm_prot_t prot; 2553 { 2554 pt_entry_t *pte = NULL, *ptes; 2555 int armprot; 2556 int flush = 0; 2557 paddr_t pa; 2558 int bank, off; 2559 struct pv_head *pvh; 2560 2561 PDEBUG(0, printf("pmap_protect: pmap=%p %08lx->%08lx %x\n", 2562 pmap, sva, eva, prot)); 2563 2564 if (~prot & VM_PROT_READ) { 2565 /* Just remove the mappings. */ 2566 pmap_remove(pmap, sva, eva); 2567 /* pmap_update not needed as it should be called by the caller 2568 * of pmap_protect */ 2569 return; 2570 } 2571 if (prot & VM_PROT_WRITE) { 2572 /* 2573 * If this is a read->write transition, just ignore it and let 2574 * uvm_fault() take care of it later. 2575 */ 2576 return; 2577 } 2578 2579 sva &= PG_FRAME; 2580 eva &= PG_FRAME; 2581 2582 /* Need to lock map->head */ 2583 PMAP_MAP_TO_HEAD_LOCK(); 2584 2585 ptes = pmap_map_ptes(pmap); 2586 /* 2587 * We need to acquire a pointer to a page table page before entering 2588 * the following loop. 2589 */ 2590 while (sva < eva) { 2591 if (pmap_pde_page(pmap_pde(pmap, sva))) 2592 break; 2593 sva = (sva & PD_MASK) + NBPD; 2594 } 2595 2596 pte = &ptes[arm_byte_to_page(sva)]; 2597 2598 while (sva < eva) { 2599 /* only check once in a while */ 2600 if ((sva & PT_MASK) == 0) { 2601 if (!pmap_pde_page(pmap_pde(pmap, sva))) { 2602 /* We can race ahead here, to the next pde. */ 2603 sva += NBPD; 2604 pte += arm_byte_to_page(NBPD); 2605 continue; 2606 } 2607 } 2608 2609 if (!pmap_pte_v(pte)) 2610 goto next; 2611 2612 flush = 1; 2613 2614 armprot = 0; 2615 if (sva < VM_MAXUSER_ADDRESS) 2616 armprot |= PT_AP(AP_U); 2617 else if (sva < VM_MAX_ADDRESS) 2618 armprot |= PT_AP(AP_W); /* XXX Ekk what is this ? */ 2619 *pte = (*pte & 0xfffff00f) | armprot; 2620 2621 pa = pmap_pte_pa(pte); 2622 2623 /* Get the physical page index */ 2624 2625 /* Clear write flag */ 2626 if ((bank = vm_physseg_find(atop(pa), &off)) != -1) { 2627 pvh = &vm_physmem[bank].pmseg.pvhead[off]; 2628 simple_lock(&pvh->pvh_lock); 2629 (void) pmap_modify_pv(pmap, sva, pvh, PT_Wr, 0); 2630 pmap_vac_me_harder(pmap, pvh, ptes, FALSE); 2631 simple_unlock(&pvh->pvh_lock); 2632 } 2633 2634 next: 2635 sva += NBPG; 2636 pte++; 2637 } 2638 pmap_unmap_ptes(pmap); 2639 PMAP_MAP_TO_HEAD_UNLOCK(); 2640 if (flush) 2641 cpu_tlb_flushID(); 2642 } 2643 2644 /* 2645 * void pmap_enter(struct pmap *pmap, vaddr_t va, paddr_t pa, vm_prot_t prot, 2646 * int flags) 2647 * 2648 * Insert the given physical page (p) at 2649 * the specified virtual address (v) in the 2650 * target physical map with the protection requested. 2651 * 2652 * If specified, the page will be wired down, meaning 2653 * that the related pte can not be reclaimed. 2654 * 2655 * NB: This is the only routine which MAY NOT lazy-evaluate 2656 * or lose information. That is, this routine must actually 2657 * insert this page into the given map NOW. 2658 */ 2659 2660 int 2661 pmap_enter(pmap, va, pa, prot, flags) 2662 struct pmap *pmap; 2663 vaddr_t va; 2664 paddr_t pa; 2665 vm_prot_t prot; 2666 int flags; 2667 { 2668 pt_entry_t *pte, *ptes; 2669 u_int npte; 2670 int bank, off; 2671 paddr_t opa; 2672 int nflags; 2673 boolean_t wired = (flags & PMAP_WIRED) != 0; 2674 struct pv_entry *pve; 2675 struct pv_head *pvh; 2676 int error; 2677 2678 PDEBUG(5, printf("pmap_enter: V%08lx P%08lx in pmap %p prot=%08x, wired = %d\n", 2679 va, pa, pmap, prot, wired)); 2680 2681 #ifdef DIAGNOSTIC 2682 /* Valid address ? */ 2683 if (va >= (KERNEL_VM_BASE + KERNEL_VM_SIZE)) 2684 panic("pmap_enter: too big"); 2685 if (pmap != pmap_kernel() && va != 0) { 2686 if (va < VM_MIN_ADDRESS || va >= VM_MAXUSER_ADDRESS) 2687 panic("pmap_enter: kernel page in user map"); 2688 } else { 2689 if (va >= VM_MIN_ADDRESS && va < VM_MAXUSER_ADDRESS) 2690 panic("pmap_enter: user page in kernel map"); 2691 if (va >= VM_MAXUSER_ADDRESS && va < VM_MAX_ADDRESS) 2692 panic("pmap_enter: entering PT page"); 2693 } 2694 #endif 2695 /* get lock */ 2696 PMAP_MAP_TO_HEAD_LOCK(); 2697 /* 2698 * Get a pointer to the pte for this virtual address. If the 2699 * pte pointer is NULL then we are missing the L2 page table 2700 * so we need to create one. 2701 */ 2702 /* XXX horrible hack to get us working with lockdebug */ 2703 simple_lock(&pmap->pm_obj.vmobjlock); 2704 pte = pmap_pte(pmap, va); 2705 if (!pte) { 2706 struct vm_page *ptp; 2707 2708 /* if failure is allowed then don't try too hard */ 2709 ptp = pmap_get_ptp(pmap, va, flags & PMAP_CANFAIL); 2710 if (ptp == NULL) { 2711 if (flags & PMAP_CANFAIL) { 2712 error = ENOMEM; 2713 goto out; 2714 } 2715 panic("pmap_enter: get ptp failed"); 2716 } 2717 2718 pte = pmap_pte(pmap, va); 2719 #ifdef DIAGNOSTIC 2720 if (!pte) 2721 panic("pmap_enter: no pte"); 2722 #endif 2723 } 2724 2725 nflags = 0; 2726 if (prot & VM_PROT_WRITE) 2727 nflags |= PT_Wr; 2728 if (wired) 2729 nflags |= PT_W; 2730 2731 /* More debugging info */ 2732 PDEBUG(5, printf("pmap_enter: pte for V%08lx = V%p (%08x)\n", va, pte, 2733 *pte)); 2734 2735 /* Is the pte valid ? If so then this page is already mapped */ 2736 if (pmap_pte_v(pte)) { 2737 /* Get the physical address of the current page mapped */ 2738 opa = pmap_pte_pa(pte); 2739 2740 #ifdef MYCROFT_HACK 2741 printf("pmap_enter: pmap=%p va=%lx pa=%lx opa=%lx\n", pmap, va, pa, opa); 2742 #endif 2743 2744 /* Are we mapping the same page ? */ 2745 if (opa == pa) { 2746 /* All we must be doing is changing the protection */ 2747 PDEBUG(0, printf("Case 02 in pmap_enter (V%08lx P%08lx)\n", 2748 va, pa)); 2749 2750 /* Has the wiring changed ? */ 2751 if ((bank = vm_physseg_find(atop(pa), &off)) != -1) { 2752 pvh = &vm_physmem[bank].pmseg.pvhead[off]; 2753 simple_lock(&pvh->pvh_lock); 2754 (void) pmap_modify_pv(pmap, va, pvh, 2755 PT_Wr | PT_W, nflags); 2756 simple_unlock(&pvh->pvh_lock); 2757 } else { 2758 pvh = NULL; 2759 } 2760 } else { 2761 /* We are replacing the page with a new one. */ 2762 cpu_cache_purgeID_rng(va, NBPG); 2763 2764 PDEBUG(0, printf("Case 03 in pmap_enter (V%08lx P%08lx P%08lx)\n", 2765 va, pa, opa)); 2766 2767 /* 2768 * If it is part of our managed memory then we 2769 * must remove it from the PV list 2770 */ 2771 if ((bank = vm_physseg_find(atop(opa), &off)) != -1) { 2772 pvh = &vm_physmem[bank].pmseg.pvhead[off]; 2773 simple_lock(&pvh->pvh_lock); 2774 pve = pmap_remove_pv(pvh, pmap, va); 2775 simple_unlock(&pvh->pvh_lock); 2776 } else { 2777 pve = NULL; 2778 } 2779 2780 goto enter; 2781 } 2782 } else { 2783 opa = 0; 2784 pve = NULL; 2785 pmap_pte_addref(pmap, va); 2786 2787 /* pte is not valid so we must be hooking in a new page */ 2788 ++pmap->pm_stats.resident_count; 2789 2790 enter: 2791 /* 2792 * Enter on the PV list if part of our managed memory 2793 */ 2794 bank = vm_physseg_find(atop(pa), &off); 2795 2796 if (pmap_initialized && (bank != -1)) { 2797 pvh = &vm_physmem[bank].pmseg.pvhead[off]; 2798 if (pve == NULL) { 2799 pve = pmap_alloc_pv(pmap, ALLOCPV_NEED); 2800 if (pve == NULL) { 2801 if (flags & PMAP_CANFAIL) { 2802 error = ENOMEM; 2803 goto out; 2804 } 2805 panic("pmap_enter: no pv entries available"); 2806 } 2807 } 2808 /* enter_pv locks pvh when adding */ 2809 pmap_enter_pv(pvh, pve, pmap, va, NULL, nflags); 2810 } else { 2811 pvh = NULL; 2812 if (pve != NULL) 2813 pmap_free_pv(pmap, pve); 2814 } 2815 } 2816 2817 #ifdef MYCROFT_HACK 2818 if (mycroft_hack) 2819 printf("pmap_enter: pmap=%p va=%lx pa=%lx opa=%lx bank=%d off=%d pv=%p\n", pmap, va, pa, opa, bank, off, pv); 2820 #endif 2821 2822 /* Construct the pte, giving the correct access. */ 2823 npte = (pa & PG_FRAME); 2824 2825 /* VA 0 is magic. */ 2826 if (pmap != pmap_kernel() && va != 0) 2827 npte |= PT_AP(AP_U); 2828 2829 if (pmap_initialized && bank != -1) { 2830 #ifdef DIAGNOSTIC 2831 if ((flags & VM_PROT_ALL) & ~prot) 2832 panic("pmap_enter: access_type exceeds prot"); 2833 #endif 2834 npte |= pte_cache_mode; 2835 if (flags & VM_PROT_WRITE) { 2836 npte |= L2_SPAGE | PT_AP(AP_W); 2837 vm_physmem[bank].pmseg.attrs[off] |= PT_H | PT_M; 2838 } else if (flags & VM_PROT_ALL) { 2839 npte |= L2_SPAGE; 2840 vm_physmem[bank].pmseg.attrs[off] |= PT_H; 2841 } else 2842 npte |= L2_INVAL; 2843 } else { 2844 if (prot & VM_PROT_WRITE) 2845 npte |= L2_SPAGE | PT_AP(AP_W); 2846 else if (prot & VM_PROT_ALL) 2847 npte |= L2_SPAGE; 2848 else 2849 npte |= L2_INVAL; 2850 } 2851 2852 #ifdef MYCROFT_HACK 2853 if (mycroft_hack) 2854 printf("pmap_enter: pmap=%p va=%lx pa=%lx prot=%x wired=%d access_type=%x npte=%08x\n", pmap, va, pa, prot, wired, flags & VM_PROT_ALL, npte); 2855 #endif 2856 2857 *pte = npte; 2858 2859 if (pmap_initialized && bank != -1) 2860 { 2861 boolean_t pmap_active = FALSE; 2862 /* XXX this will change once the whole of pmap_enter uses 2863 * map_ptes 2864 */ 2865 ptes = pmap_map_ptes(pmap); 2866 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap) 2867 || (pmap == pmap_kernel())) 2868 pmap_active = TRUE; 2869 simple_lock(&pvh->pvh_lock); 2870 pmap_vac_me_harder(pmap, pvh, ptes, pmap_active); 2871 simple_unlock(&pvh->pvh_lock); 2872 pmap_unmap_ptes(pmap); 2873 } 2874 2875 /* Better flush the TLB ... */ 2876 cpu_tlb_flushID_SE(va); 2877 error = 0; 2878 out: 2879 simple_unlock(&pmap->pm_obj.vmobjlock); 2880 PMAP_MAP_TO_HEAD_UNLOCK(); 2881 PDEBUG(5, printf("pmap_enter: pte = V%p %08x\n", pte, *pte)); 2882 2883 return error; 2884 } 2885 2886 void 2887 pmap_kenter_pa(va, pa, prot) 2888 vaddr_t va; 2889 paddr_t pa; 2890 vm_prot_t prot; 2891 { 2892 struct pmap *pmap = pmap_kernel(); 2893 pt_entry_t *pte; 2894 struct vm_page *pg; 2895 2896 if (!pmap_pde_page(pmap_pde(pmap, va))) { 2897 2898 #ifdef DIAGNOSTIC 2899 if (pmap_pde_v(pmap_pde(pmap, va))) 2900 panic("Trying to map kernel page into section mapping" 2901 " VA=%lx PA=%lx", va, pa); 2902 #endif 2903 /* 2904 * For the kernel pmaps it would be better to ensure 2905 * that they are always present, and to grow the 2906 * kernel as required. 2907 */ 2908 2909 /* must lock the pmap */ 2910 simple_lock(&(pmap_kernel()->pm_obj.vmobjlock)); 2911 /* Allocate a page table */ 2912 pg = uvm_pagealloc(&(pmap_kernel()->pm_obj), 0, NULL, 2913 UVM_PGA_USERESERVE | UVM_PGA_ZERO); 2914 if (pg == NULL) { 2915 panic("pmap_kenter_pa: no free pages"); 2916 } 2917 pg->flags &= ~PG_BUSY; /* never busy */ 2918 2919 /* Wire this page table into the L1. */ 2920 pmap_map_in_l1(pmap, va, VM_PAGE_TO_PHYS(pg), TRUE); 2921 simple_unlock(&(pmap_kernel()->pm_obj.vmobjlock)); 2922 } 2923 pte = vtopte(va); 2924 KASSERT(!pmap_pte_v(pte)); 2925 *pte = L2_PTE(pa, AP_KRW); 2926 } 2927 2928 void 2929 pmap_kremove(va, len) 2930 vaddr_t va; 2931 vsize_t len; 2932 { 2933 pt_entry_t *pte; 2934 2935 for (len >>= PAGE_SHIFT; len > 0; len--, va += PAGE_SIZE) { 2936 2937 /* 2938 * We assume that we will only be called with small 2939 * regions of memory. 2940 */ 2941 2942 KASSERT(pmap_pde_page(pmap_pde(pmap_kernel(), va))); 2943 pte = vtopte(va); 2944 cpu_cache_purgeID_rng(va, PAGE_SIZE); 2945 *pte = 0; 2946 cpu_tlb_flushID_SE(va); 2947 } 2948 } 2949 2950 /* 2951 * pmap_page_protect: 2952 * 2953 * Lower the permission for all mappings to a given page. 2954 */ 2955 2956 void 2957 pmap_page_protect(pg, prot) 2958 struct vm_page *pg; 2959 vm_prot_t prot; 2960 { 2961 paddr_t pa = VM_PAGE_TO_PHYS(pg); 2962 2963 PDEBUG(0, printf("pmap_page_protect(pa=%lx, prot=%d)\n", pa, prot)); 2964 2965 switch(prot) { 2966 case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE: 2967 case VM_PROT_READ|VM_PROT_WRITE: 2968 return; 2969 2970 case VM_PROT_READ: 2971 case VM_PROT_READ|VM_PROT_EXECUTE: 2972 pmap_copy_on_write(pa); 2973 break; 2974 2975 default: 2976 pmap_remove_all(pa); 2977 break; 2978 } 2979 } 2980 2981 2982 /* 2983 * Routine: pmap_unwire 2984 * Function: Clear the wired attribute for a map/virtual-address 2985 * pair. 2986 * In/out conditions: 2987 * The mapping must already exist in the pmap. 2988 */ 2989 2990 void 2991 pmap_unwire(pmap, va) 2992 struct pmap *pmap; 2993 vaddr_t va; 2994 { 2995 pt_entry_t *pte; 2996 paddr_t pa; 2997 int bank, off; 2998 struct pv_head *pvh; 2999 3000 /* 3001 * Make sure pmap is valid. -dct 3002 */ 3003 if (pmap == NULL) 3004 return; 3005 3006 /* Get the pte */ 3007 pte = pmap_pte(pmap, va); 3008 if (!pte) 3009 return; 3010 3011 /* Extract the physical address of the page */ 3012 pa = pmap_pte_pa(pte); 3013 3014 if ((bank = vm_physseg_find(atop(pa), &off)) == -1) 3015 return; 3016 pvh = &vm_physmem[bank].pmseg.pvhead[off]; 3017 simple_lock(&pvh->pvh_lock); 3018 /* Update the wired bit in the pv entry for this page. */ 3019 (void) pmap_modify_pv(pmap, va, pvh, PT_W, 0); 3020 simple_unlock(&pvh->pvh_lock); 3021 } 3022 3023 /* 3024 * pt_entry_t *pmap_pte(struct pmap *pmap, vaddr_t va) 3025 * 3026 * Return the pointer to a page table entry corresponding to the supplied 3027 * virtual address. 3028 * 3029 * The page directory is first checked to make sure that a page table 3030 * for the address in question exists and if it does a pointer to the 3031 * entry is returned. 3032 * 3033 * The way this works is that that the kernel page tables are mapped 3034 * into the memory map at ALT_PAGE_TBLS_BASE to ALT_PAGE_TBLS_BASE+4MB. 3035 * This allows page tables to be located quickly. 3036 */ 3037 pt_entry_t * 3038 pmap_pte(pmap, va) 3039 struct pmap *pmap; 3040 vaddr_t va; 3041 { 3042 pt_entry_t *ptp; 3043 pt_entry_t *result; 3044 3045 /* The pmap must be valid */ 3046 if (!pmap) 3047 return(NULL); 3048 3049 /* Return the address of the pte */ 3050 PDEBUG(10, printf("pmap_pte: pmap=%p va=V%08lx pde = V%p (%08X)\n", 3051 pmap, va, pmap_pde(pmap, va), *(pmap_pde(pmap, va)))); 3052 3053 /* Do we have a valid pde ? If not we don't have a page table */ 3054 if (!pmap_pde_page(pmap_pde(pmap, va))) { 3055 PDEBUG(0, printf("pmap_pte: failed - pde = %p\n", 3056 pmap_pde(pmap, va))); 3057 return(NULL); 3058 } 3059 3060 PDEBUG(10, printf("pmap pagetable = P%08lx current = P%08x\n", 3061 pmap->pm_pptpt, (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE 3062 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) + 3063 (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & PG_FRAME))); 3064 3065 /* 3066 * If the pmap is the kernel pmap or the pmap is the active one 3067 * then we can just return a pointer to entry relative to 3068 * PROCESS_PAGE_TBLS_BASE. 3069 * Otherwise we need to map the page tables to an alternative 3070 * address and reference them there. 3071 */ 3072 if (pmap == pmap_kernel() || pmap->pm_pptpt 3073 == (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE 3074 + ((PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) & 3075 ~3) + (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & PG_FRAME)) { 3076 ptp = (pt_entry_t *)PROCESS_PAGE_TBLS_BASE; 3077 } else { 3078 struct proc *p = curproc; 3079 3080 /* If we don't have a valid curproc use proc0 */ 3081 /* Perhaps we should just use kernel_pmap instead */ 3082 if (p == NULL) 3083 p = &proc0; 3084 #ifdef DIAGNOSTIC 3085 /* 3086 * The pmap should always be valid for the process so 3087 * panic if it is not. 3088 */ 3089 if (!p->p_vmspace || !p->p_vmspace->vm_map.pmap) { 3090 printf("pmap_pte: va=%08lx p=%p vm=%p\n", 3091 va, p, p->p_vmspace); 3092 console_debugger(); 3093 } 3094 /* 3095 * The pmap for the current process should be mapped. If it 3096 * is not then we have a problem. 3097 */ 3098 if (p->p_vmspace->vm_map.pmap->pm_pptpt != 3099 (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE 3100 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) + 3101 (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & PG_FRAME)) { 3102 printf("pmap pagetable = P%08lx current = P%08x ", 3103 pmap->pm_pptpt, (*((pt_entry_t *)(PROCESS_PAGE_TBLS_BASE 3104 + (PROCESS_PAGE_TBLS_BASE >> (PGSHIFT - 2)) + 3105 (PROCESS_PAGE_TBLS_BASE >> PDSHIFT))) & 3106 PG_FRAME)); 3107 printf("pptpt=%lx\n", p->p_vmspace->vm_map.pmap->pm_pptpt); 3108 panic("pmap_pte: current and pmap mismatch\n"); 3109 } 3110 #endif 3111 3112 ptp = (pt_entry_t *)ALT_PAGE_TBLS_BASE; 3113 pmap_map_in_l1(p->p_vmspace->vm_map.pmap, ALT_PAGE_TBLS_BASE, 3114 pmap->pm_pptpt, FALSE); 3115 cpu_tlb_flushD(); 3116 cpu_cpwait(); 3117 } 3118 PDEBUG(10, printf("page tables base = %p offset=%lx\n", ptp, 3119 ((va >> (PGSHIFT-2)) & ~3))); 3120 result = (pt_entry_t *)((char *)ptp + ((va >> (PGSHIFT-2)) & ~3)); 3121 return(result); 3122 } 3123 3124 /* 3125 * Routine: pmap_extract 3126 * Function: 3127 * Extract the physical page address associated 3128 * with the given map/virtual_address pair. 3129 */ 3130 boolean_t 3131 pmap_extract(pmap, va, pap) 3132 struct pmap *pmap; 3133 vaddr_t va; 3134 paddr_t *pap; 3135 { 3136 pt_entry_t *pte, *ptes; 3137 paddr_t pa; 3138 3139 PDEBUG(5, printf("pmap_extract: pmap=%p, va=V%08lx\n", pmap, va)); 3140 3141 /* 3142 * Get the pte for this virtual address. 3143 */ 3144 ptes = pmap_map_ptes(pmap); 3145 pte = &ptes[arm_byte_to_page(va)]; 3146 3147 /* 3148 * If there is no pte then there is no page table etc. 3149 * Is the pte valid ? If not then no paged is actually mapped here 3150 * XXX Should we handle section mappings? 3151 */ 3152 if (!pmap_pde_page(pmap_pde(pmap, va)) || !pmap_pte_v(pte)){ 3153 pmap_unmap_ptes(pmap); 3154 return (FALSE); 3155 } 3156 3157 /* Return the physical address depending on the PTE type */ 3158 /* XXX What about L1 section mappings ? */ 3159 if ((*(pte) & L2_MASK) == L2_LPAGE) { 3160 /* Extract the physical address from the pte */ 3161 pa = (*(pte)) & ~(L2_LPAGE_SIZE - 1); 3162 3163 PDEBUG(5, printf("pmap_extract: LPAGE pa = P%08lx\n", 3164 (pa | (va & (L2_LPAGE_SIZE - 1))))); 3165 3166 if (pap != NULL) 3167 *pap = pa | (va & (L2_LPAGE_SIZE - 1)); 3168 } else { 3169 /* Extract the physical address from the pte */ 3170 pa = pmap_pte_pa(pte); 3171 3172 PDEBUG(5, printf("pmap_extract: SPAGE pa = P%08lx\n", 3173 (pa | (va & ~PG_FRAME)))); 3174 3175 if (pap != NULL) 3176 *pap = pa | (va & ~PG_FRAME); 3177 } 3178 pmap_unmap_ptes(pmap); 3179 return (TRUE); 3180 } 3181 3182 3183 /* 3184 * Copy the range specified by src_addr/len from the source map to the 3185 * range dst_addr/len in the destination map. 3186 * 3187 * This routine is only advisory and need not do anything. 3188 */ 3189 3190 void 3191 pmap_copy(dst_pmap, src_pmap, dst_addr, len, src_addr) 3192 struct pmap *dst_pmap; 3193 struct pmap *src_pmap; 3194 vaddr_t dst_addr; 3195 vsize_t len; 3196 vaddr_t src_addr; 3197 { 3198 PDEBUG(0, printf("pmap_copy(%p, %p, %lx, %lx, %lx)\n", 3199 dst_pmap, src_pmap, dst_addr, len, src_addr)); 3200 } 3201 3202 #if defined(PMAP_DEBUG) 3203 void 3204 pmap_dump_pvlist(phys, m) 3205 vaddr_t phys; 3206 char *m; 3207 { 3208 struct pv_head *pvh; 3209 struct pv_entry *pv; 3210 int bank, off; 3211 3212 if ((bank = vm_physseg_find(atop(phys), &off)) == -1) { 3213 printf("INVALID PA\n"); 3214 return; 3215 } 3216 pvh = &vm_physmem[bank].pmseg.pvhead[off]; 3217 simple_lock(&pvh->pvh_lock); 3218 printf("%s %08lx:", m, phys); 3219 if (pvh->pvh_list == NULL) { 3220 printf(" no mappings\n"); 3221 return; 3222 } 3223 3224 for (pv = pvh->pvh_list; pv; pv = pv->pv_next) 3225 printf(" pmap %p va %08lx flags %08x", pv->pv_pmap, 3226 pv->pv_va, pv->pv_flags); 3227 3228 printf("\n"); 3229 simple_unlock(&pvh->pvh_lock); 3230 } 3231 3232 #endif /* PMAP_DEBUG */ 3233 3234 __inline static boolean_t 3235 pmap_testbit(pa, setbits) 3236 paddr_t pa; 3237 unsigned int setbits; 3238 { 3239 int bank, off; 3240 3241 PDEBUG(1, printf("pmap_testbit: pa=%08lx set=%08x\n", pa, setbits)); 3242 3243 if ((bank = vm_physseg_find(atop(pa), &off)) == -1) 3244 return(FALSE); 3245 3246 /* 3247 * Check saved info only 3248 */ 3249 if (vm_physmem[bank].pmseg.attrs[off] & setbits) { 3250 PDEBUG(0, printf("pmap_attributes = %02x\n", 3251 vm_physmem[bank].pmseg.attrs[off])); 3252 return(TRUE); 3253 } 3254 3255 return(FALSE); 3256 } 3257 3258 static pt_entry_t * 3259 pmap_map_ptes(struct pmap *pmap) 3260 { 3261 struct proc *p; 3262 3263 /* the kernel's pmap is always accessible */ 3264 if (pmap == pmap_kernel()) { 3265 return (pt_entry_t *)PROCESS_PAGE_TBLS_BASE ; 3266 } 3267 3268 if (pmap_is_curpmap(pmap)) { 3269 simple_lock(&pmap->pm_obj.vmobjlock); 3270 return (pt_entry_t *)PROCESS_PAGE_TBLS_BASE; 3271 } 3272 3273 p = curproc; 3274 3275 if (p == NULL) 3276 p = &proc0; 3277 3278 /* need to lock both curpmap and pmap: use ordered locking */ 3279 if ((unsigned) pmap < (unsigned) curproc->p_vmspace->vm_map.pmap) { 3280 simple_lock(&pmap->pm_obj.vmobjlock); 3281 simple_lock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock); 3282 } else { 3283 simple_lock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock); 3284 simple_lock(&pmap->pm_obj.vmobjlock); 3285 } 3286 3287 pmap_map_in_l1(p->p_vmspace->vm_map.pmap, ALT_PAGE_TBLS_BASE, 3288 pmap->pm_pptpt, FALSE); 3289 cpu_tlb_flushD(); 3290 cpu_cpwait(); 3291 return (pt_entry_t *)ALT_PAGE_TBLS_BASE; 3292 } 3293 3294 /* 3295 * pmap_unmap_ptes: unlock the PTE mapping of "pmap" 3296 */ 3297 3298 static void 3299 pmap_unmap_ptes(pmap) 3300 struct pmap *pmap; 3301 { 3302 if (pmap == pmap_kernel()) { 3303 return; 3304 } 3305 if (pmap_is_curpmap(pmap)) { 3306 simple_unlock(&pmap->pm_obj.vmobjlock); 3307 } else { 3308 simple_unlock(&pmap->pm_obj.vmobjlock); 3309 simple_unlock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock); 3310 } 3311 } 3312 3313 /* 3314 * Modify pte bits for all ptes corresponding to the given physical address. 3315 * We use `maskbits' rather than `clearbits' because we're always passing 3316 * constants and the latter would require an extra inversion at run-time. 3317 */ 3318 3319 static void 3320 pmap_clearbit(pa, maskbits) 3321 paddr_t pa; 3322 unsigned int maskbits; 3323 { 3324 struct pv_entry *pv; 3325 struct pv_head *pvh; 3326 pt_entry_t *pte; 3327 vaddr_t va; 3328 int bank, off, tlbentry; 3329 3330 PDEBUG(1, printf("pmap_clearbit: pa=%08lx mask=%08x\n", 3331 pa, maskbits)); 3332 3333 tlbentry = 0; 3334 3335 if ((bank = vm_physseg_find(atop(pa), &off)) == -1) 3336 return; 3337 PMAP_HEAD_TO_MAP_LOCK(); 3338 pvh = &vm_physmem[bank].pmseg.pvhead[off]; 3339 simple_lock(&pvh->pvh_lock); 3340 3341 /* 3342 * Clear saved attributes (modify, reference) 3343 */ 3344 vm_physmem[bank].pmseg.attrs[off] &= ~maskbits; 3345 3346 if (pvh->pvh_list == NULL) { 3347 simple_unlock(&pvh->pvh_lock); 3348 PMAP_HEAD_TO_MAP_UNLOCK(); 3349 return; 3350 } 3351 3352 /* 3353 * Loop over all current mappings setting/clearing as appropos 3354 */ 3355 for (pv = pvh->pvh_list; pv; pv = pv->pv_next) { 3356 va = pv->pv_va; 3357 pv->pv_flags &= ~maskbits; 3358 pte = pmap_pte(pv->pv_pmap, va); 3359 KASSERT(pte != NULL); 3360 if (maskbits & (PT_Wr|PT_M)) { 3361 if ((pv->pv_flags & PT_NC)) { 3362 /* 3363 * Entry is not cacheable: reenable 3364 * the cache, nothing to flush 3365 * 3366 * Don't turn caching on again if this 3367 * is a modified emulation. This 3368 * would be inconsitent with the 3369 * settings created by 3370 * pmap_vac_me_harder(). 3371 * 3372 * There's no need to call 3373 * pmap_vac_me_harder() here: all 3374 * pages are loosing their write 3375 * permission. 3376 * 3377 */ 3378 if (maskbits & PT_Wr) { 3379 *pte |= pte_cache_mode; 3380 pv->pv_flags &= ~PT_NC; 3381 } 3382 } else if (pmap_is_curpmap(pv->pv_pmap)) 3383 /* 3384 * Entry is cacheable: check if pmap is 3385 * current if it is flush it, 3386 * otherwise it won't be in the cache 3387 */ 3388 cpu_cache_purgeID_rng(pv->pv_va, NBPG); 3389 3390 /* make the pte read only */ 3391 *pte &= ~PT_AP(AP_W); 3392 } 3393 3394 if (maskbits & PT_H) 3395 *pte = (*pte & ~L2_MASK) | L2_INVAL; 3396 3397 if (pmap_is_curpmap(pv->pv_pmap)) 3398 /* 3399 * if we had cacheable pte's we'd clean the 3400 * pte out to memory here 3401 * 3402 * flush tlb entry as it's in the current pmap 3403 */ 3404 cpu_tlb_flushID_SE(pv->pv_va); 3405 } 3406 cpu_cpwait(); 3407 3408 simple_unlock(&pvh->pvh_lock); 3409 PMAP_HEAD_TO_MAP_UNLOCK(); 3410 } 3411 3412 3413 boolean_t 3414 pmap_clear_modify(pg) 3415 struct vm_page *pg; 3416 { 3417 paddr_t pa = VM_PAGE_TO_PHYS(pg); 3418 boolean_t rv; 3419 3420 PDEBUG(0, printf("pmap_clear_modify pa=%08lx\n", pa)); 3421 rv = pmap_testbit(pa, PT_M); 3422 pmap_clearbit(pa, PT_M); 3423 return rv; 3424 } 3425 3426 3427 boolean_t 3428 pmap_clear_reference(pg) 3429 struct vm_page *pg; 3430 { 3431 paddr_t pa = VM_PAGE_TO_PHYS(pg); 3432 boolean_t rv; 3433 3434 PDEBUG(0, printf("pmap_clear_reference pa=%08lx\n", pa)); 3435 rv = pmap_testbit(pa, PT_H); 3436 pmap_clearbit(pa, PT_H); 3437 return rv; 3438 } 3439 3440 3441 void 3442 pmap_copy_on_write(pa) 3443 paddr_t pa; 3444 { 3445 PDEBUG(0, printf("pmap_copy_on_write pa=%08lx\n", pa)); 3446 pmap_clearbit(pa, PT_Wr); 3447 } 3448 3449 3450 boolean_t 3451 pmap_is_modified(pg) 3452 struct vm_page *pg; 3453 { 3454 paddr_t pa = VM_PAGE_TO_PHYS(pg); 3455 boolean_t result; 3456 3457 result = pmap_testbit(pa, PT_M); 3458 PDEBUG(1, printf("pmap_is_modified pa=%08lx %x\n", pa, result)); 3459 return (result); 3460 } 3461 3462 3463 boolean_t 3464 pmap_is_referenced(pg) 3465 struct vm_page *pg; 3466 { 3467 paddr_t pa = VM_PAGE_TO_PHYS(pg); 3468 boolean_t result; 3469 3470 result = pmap_testbit(pa, PT_H); 3471 PDEBUG(0, printf("pmap_is_referenced pa=%08lx %x\n", pa, result)); 3472 return (result); 3473 } 3474 3475 3476 int 3477 pmap_modified_emulation(pmap, va) 3478 struct pmap *pmap; 3479 vaddr_t va; 3480 { 3481 pt_entry_t *pte; 3482 paddr_t pa; 3483 int bank, off; 3484 struct pv_head *pvh; 3485 u_int flags; 3486 3487 PDEBUG(2, printf("pmap_modified_emulation\n")); 3488 3489 /* Get the pte */ 3490 pte = pmap_pte(pmap, va); 3491 if (!pte) { 3492 PDEBUG(2, printf("no pte\n")); 3493 return(0); 3494 } 3495 3496 PDEBUG(1, printf("*pte=%08x\n", *pte)); 3497 3498 /* Check for a zero pte */ 3499 if (*pte == 0) 3500 return(0); 3501 3502 /* This can happen if user code tries to access kernel memory. */ 3503 if ((*pte & PT_AP(AP_W)) != 0) 3504 return (0); 3505 3506 /* Extract the physical address of the page */ 3507 pa = pmap_pte_pa(pte); 3508 if ((bank = vm_physseg_find(atop(pa), &off)) == -1) 3509 return(0); 3510 3511 PMAP_HEAD_TO_MAP_LOCK(); 3512 /* Get the current flags for this page. */ 3513 pvh = &vm_physmem[bank].pmseg.pvhead[off]; 3514 /* XXX: needed if we hold head->map lock? */ 3515 simple_lock(&pvh->pvh_lock); 3516 3517 flags = pmap_modify_pv(pmap, va, pvh, 0, 0); 3518 PDEBUG(2, printf("pmap_modified_emulation: flags = %08x\n", flags)); 3519 3520 /* 3521 * Do the flags say this page is writable ? If not then it is a 3522 * genuine write fault. If yes then the write fault is our fault 3523 * as we did not reflect the write access in the PTE. Now we know 3524 * a write has occurred we can correct this and also set the 3525 * modified bit 3526 */ 3527 if (~flags & PT_Wr) { 3528 simple_unlock(&pvh->pvh_lock); 3529 PMAP_HEAD_TO_MAP_UNLOCK(); 3530 return(0); 3531 } 3532 3533 PDEBUG(0, printf("pmap_modified_emulation: Got a hit va=%08lx, pte = %p (%08x)\n", 3534 va, pte, *pte)); 3535 vm_physmem[bank].pmseg.attrs[off] |= PT_H | PT_M; 3536 3537 /* 3538 * Re-enable write permissions for the page. No need to call 3539 * pmap_vac_me_harder(), since this is just a 3540 * modified-emulation fault, and the PT_Wr bit isn't changing. We've 3541 * already set the cacheable bits based on the assumption that we 3542 * can write to this page. 3543 */ 3544 *pte = (*pte & ~L2_MASK) | L2_SPAGE | PT_AP(AP_W); 3545 PDEBUG(0, printf("->(%08x)\n", *pte)); 3546 3547 simple_unlock(&pvh->pvh_lock); 3548 PMAP_HEAD_TO_MAP_UNLOCK(); 3549 /* Return, indicating the problem has been dealt with */ 3550 cpu_tlb_flushID_SE(va); 3551 cpu_cpwait(); 3552 return(1); 3553 } 3554 3555 3556 int 3557 pmap_handled_emulation(pmap, va) 3558 struct pmap *pmap; 3559 vaddr_t va; 3560 { 3561 pt_entry_t *pte; 3562 paddr_t pa; 3563 int bank, off; 3564 3565 PDEBUG(2, printf("pmap_handled_emulation\n")); 3566 3567 /* Get the pte */ 3568 pte = pmap_pte(pmap, va); 3569 if (!pte) { 3570 PDEBUG(2, printf("no pte\n")); 3571 return(0); 3572 } 3573 3574 PDEBUG(1, printf("*pte=%08x\n", *pte)); 3575 3576 /* Check for a zero pte */ 3577 if (*pte == 0) 3578 return(0); 3579 3580 /* This can happen if user code tries to access kernel memory. */ 3581 if ((*pte & L2_MASK) != L2_INVAL) 3582 return (0); 3583 3584 /* Extract the physical address of the page */ 3585 pa = pmap_pte_pa(pte); 3586 if ((bank = vm_physseg_find(atop(pa), &off)) == -1) 3587 return(0); 3588 3589 /* 3590 * Ok we just enable the pte and mark the attibs as handled 3591 */ 3592 PDEBUG(0, printf("pmap_handled_emulation: Got a hit va=%08lx pte = %p (%08x)\n", 3593 va, pte, *pte)); 3594 vm_physmem[bank].pmseg.attrs[off] |= PT_H; 3595 *pte = (*pte & ~L2_MASK) | L2_SPAGE; 3596 PDEBUG(0, printf("->(%08x)\n", *pte)); 3597 3598 /* Return, indicating the problem has been dealt with */ 3599 cpu_tlb_flushID_SE(va); 3600 cpu_cpwait(); 3601 return(1); 3602 } 3603 3604 3605 3606 3607 /* 3608 * pmap_collect: free resources held by a pmap 3609 * 3610 * => optional function. 3611 * => called when a process is swapped out to free memory. 3612 */ 3613 3614 void 3615 pmap_collect(pmap) 3616 struct pmap *pmap; 3617 { 3618 } 3619 3620 /* 3621 * Routine: pmap_procwr 3622 * 3623 * Function: 3624 * Synchronize caches corresponding to [addr, addr+len) in p. 3625 * 3626 */ 3627 void 3628 pmap_procwr(p, va, len) 3629 struct proc *p; 3630 vaddr_t va; 3631 int len; 3632 { 3633 /* We only need to do anything if it is the current process. */ 3634 if (p == curproc) 3635 cpu_cache_syncI_rng(va, len); 3636 } 3637 /* 3638 * PTP functions 3639 */ 3640 3641 /* 3642 * pmap_steal_ptp: Steal a PTP from somewhere else. 3643 * 3644 * This is just a placeholder, for now we never steal. 3645 */ 3646 3647 static struct vm_page * 3648 pmap_steal_ptp(struct pmap *pmap, vaddr_t va) 3649 { 3650 return (NULL); 3651 } 3652 3653 /* 3654 * pmap_get_ptp: get a PTP (if there isn't one, allocate a new one) 3655 * 3656 * => pmap should NOT be pmap_kernel() 3657 * => pmap should be locked 3658 */ 3659 3660 static struct vm_page * 3661 pmap_get_ptp(struct pmap *pmap, vaddr_t va, boolean_t just_try) 3662 { 3663 struct vm_page *ptp; 3664 3665 if (pmap_pde_page(pmap_pde(pmap, va))) { 3666 3667 /* valid... check hint (saves us a PA->PG lookup) */ 3668 #if 0 3669 if (pmap->pm_ptphint && 3670 ((unsigned)pmap_pde(pmap, va) & PG_FRAME) == 3671 VM_PAGE_TO_PHYS(pmap->pm_ptphint)) 3672 return (pmap->pm_ptphint); 3673 #endif 3674 ptp = uvm_pagelookup(&pmap->pm_obj, va); 3675 #ifdef DIAGNOSTIC 3676 if (ptp == NULL) 3677 panic("pmap_get_ptp: unmanaged user PTP"); 3678 #endif 3679 // pmap->pm_ptphint = ptp; 3680 return(ptp); 3681 } 3682 3683 /* allocate a new PTP (updates ptphint) */ 3684 return(pmap_alloc_ptp(pmap, va, just_try)); 3685 } 3686 3687 /* 3688 * pmap_alloc_ptp: allocate a PTP for a PMAP 3689 * 3690 * => pmap should already be locked by caller 3691 * => we use the ptp's wire_count to count the number of active mappings 3692 * in the PTP (we start it at one to prevent any chance this PTP 3693 * will ever leak onto the active/inactive queues) 3694 */ 3695 3696 /*__inline */ static struct vm_page * 3697 pmap_alloc_ptp(struct pmap *pmap, vaddr_t va, boolean_t just_try) 3698 { 3699 struct vm_page *ptp; 3700 3701 ptp = uvm_pagealloc(&pmap->pm_obj, va, NULL, 3702 UVM_PGA_USERESERVE|UVM_PGA_ZERO); 3703 if (ptp == NULL) { 3704 if (just_try) 3705 return (NULL); 3706 3707 ptp = pmap_steal_ptp(pmap, va); 3708 3709 if (ptp == NULL) 3710 return (NULL); 3711 /* Stole a page, zero it. */ 3712 pmap_zero_page(VM_PAGE_TO_PHYS(ptp)); 3713 } 3714 3715 /* got one! */ 3716 ptp->flags &= ~PG_BUSY; /* never busy */ 3717 ptp->wire_count = 1; /* no mappings yet */ 3718 pmap_map_in_l1(pmap, va, VM_PAGE_TO_PHYS(ptp), TRUE); 3719 pmap->pm_stats.resident_count++; /* count PTP as resident */ 3720 // pmap->pm_ptphint = ptp; 3721 return (ptp); 3722 } 3723 3724 /* End of pmap.c */ 3725
Indexes created Mon Oct 27 20:09:55 GMT 2025