pmap.c revision 1.68
1 /* $NetBSD: pmap.c,v 1.68 2002/03/24 21:32:18 thorpej Exp $ */ 2 3 /* 4 * Copyright (c) 2002 Wasabi Systems, Inc. 5 * Copyright (c) 2001 Richard Earnshaw 6 * Copyright (c) 2001 Christopher Gilbert 7 * All rights reserved. 8 * 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 3. The name of the company nor the name of the author may be used to 15 * endorse or promote products derived from this software without specific 16 * prior written permission. 17 * 18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR IMPLIED 19 * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF 20 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 21 * IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, 22 * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 23 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 24 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 28 * SUCH DAMAGE. 29 */ 30 31 /*- 32 * Copyright (c) 1999 The NetBSD Foundation, Inc. 33 * All rights reserved. 34 * 35 * This code is derived from software contributed to The NetBSD Foundation 36 * by Charles M. Hannum. 37 * 38 * Redistribution and use in source and binary forms, with or without 39 * modification, are permitted provided that the following conditions 40 * are met: 41 * 1. Redistributions of source code must retain the above copyright 42 * notice, this list of conditions and the following disclaimer. 43 * 2. Redistributions in binary form must reproduce the above copyright 44 * notice, this list of conditions and the following disclaimer in the 45 * documentation and/or other materials provided with the distribution. 46 * 3. All advertising materials mentioning features or use of this software 47 * must display the following acknowledgement: 48 * This product includes software developed by the NetBSD 49 * Foundation, Inc. and its contributors. 50 * 4. Neither the name of The NetBSD Foundation nor the names of its 51 * contributors may be used to endorse or promote products derived 52 * from this software without specific prior written permission. 53 * 54 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 55 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 56 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 57 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 58 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 59 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 60 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 61 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 62 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 63 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 64 * POSSIBILITY OF SUCH DAMAGE. 65 */ 66 67 /* 68 * Copyright (c) 1994-1998 Mark Brinicombe. 69 * Copyright (c) 1994 Brini. 70 * All rights reserved. 71 * 72 * This code is derived from software written for Brini by Mark Brinicombe 73 * 74 * Redistribution and use in source and binary forms, with or without 75 * modification, are permitted provided that the following conditions 76 * are met: 77 * 1. Redistributions of source code must retain the above copyright 78 * notice, this list of conditions and the following disclaimer. 79 * 2. Redistributions in binary form must reproduce the above copyright 80 * notice, this list of conditions and the following disclaimer in the 81 * documentation and/or other materials provided with the distribution. 82 * 3. All advertising materials mentioning features or use of this software 83 * must display the following acknowledgement: 84 * This product includes software developed by Mark Brinicombe. 85 * 4. The name of the author may not be used to endorse or promote products 86 * derived from this software without specific prior written permission. 87 * 88 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR 89 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES 90 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. 91 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, 92 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT 93 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 94 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 95 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 96 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF 97 * 98 * RiscBSD kernel project 99 * 100 * pmap.c 101 * 102 * Machine dependant vm stuff 103 * 104 * Created : 20/09/94 105 */ 106 107 /* 108 * Performance improvements, UVM changes, overhauls and part-rewrites 109 * were contributed by Neil A. Carson <neil (at) causality.com>. 110 */ 111 112 /* 113 * The dram block info is currently referenced from the bootconfig. 114 * This should be placed in a separate structure. 115 */ 116 117 /* 118 * Special compilation symbols 119 * PMAP_DEBUG - Build in pmap_debug_level code 120 */ 121 122 /* Include header files */ 123 124 #include "opt_pmap_debug.h" 125 #include "opt_ddb.h" 126 127 #include <sys/types.h> 128 #include <sys/param.h> 129 #include <sys/kernel.h> 130 #include <sys/systm.h> 131 #include <sys/proc.h> 132 #include <sys/malloc.h> 133 #include <sys/user.h> 134 #include <sys/pool.h> 135 #include <sys/cdefs.h> 136 137 #include <uvm/uvm.h> 138 139 #include <machine/bootconfig.h> 140 #include <machine/bus.h> 141 #include <machine/pmap.h> 142 #include <machine/pcb.h> 143 #include <machine/param.h> 144 #include <arm/arm32/katelib.h> 145 146 __KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.68 2002/03/24 21:32:18 thorpej Exp $"); 147 #ifdef PMAP_DEBUG 148 #define PDEBUG(_lev_,_stat_) \ 149 if (pmap_debug_level >= (_lev_)) \ 150 ((_stat_)) 151 int pmap_debug_level = -2; 152 void pmap_dump_pvlist(vaddr_t phys, char *m); 153 154 /* 155 * for switching to potentially finer grained debugging 156 */ 157 #define PDB_FOLLOW 0x0001 158 #define PDB_INIT 0x0002 159 #define PDB_ENTER 0x0004 160 #define PDB_REMOVE 0x0008 161 #define PDB_CREATE 0x0010 162 #define PDB_PTPAGE 0x0020 163 #define PDB_GROWKERN 0x0040 164 #define PDB_BITS 0x0080 165 #define PDB_COLLECT 0x0100 166 #define PDB_PROTECT 0x0200 167 #define PDB_MAP_L1 0x0400 168 #define PDB_BOOTSTRAP 0x1000 169 #define PDB_PARANOIA 0x2000 170 #define PDB_WIRING 0x4000 171 #define PDB_PVDUMP 0x8000 172 173 int debugmap = 0; 174 int pmapdebug = PDB_PARANOIA | PDB_FOLLOW; 175 #define NPDEBUG(_lev_,_stat_) \ 176 if (pmapdebug & (_lev_)) \ 177 ((_stat_)) 178 179 #else /* PMAP_DEBUG */ 180 #define PDEBUG(_lev_,_stat_) /* Nothing */ 181 #define NPDEBUG(_lev_,_stat_) /* Nothing */ 182 #endif /* PMAP_DEBUG */ 183 184 struct pmap kernel_pmap_store; 185 186 /* 187 * linked list of all non-kernel pmaps 188 */ 189 190 static struct pmap_head pmaps; 191 192 /* 193 * pool that pmap structures are allocated from 194 */ 195 196 struct pool pmap_pmap_pool; 197 198 static pt_entry_t *csrc_pte, *cdst_pte; 199 static vaddr_t csrcp, cdstp; 200 201 char *memhook; 202 extern caddr_t msgbufaddr; 203 204 boolean_t pmap_initialized = FALSE; /* Has pmap_init completed? */ 205 /* 206 * locking data structures 207 */ 208 209 static struct lock pmap_main_lock; 210 static struct simplelock pvalloc_lock; 211 static struct simplelock pmaps_lock; 212 #ifdef LOCKDEBUG 213 #define PMAP_MAP_TO_HEAD_LOCK() \ 214 (void) spinlockmgr(&pmap_main_lock, LK_SHARED, NULL) 215 #define PMAP_MAP_TO_HEAD_UNLOCK() \ 216 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL) 217 218 #define PMAP_HEAD_TO_MAP_LOCK() \ 219 (void) spinlockmgr(&pmap_main_lock, LK_EXCLUSIVE, NULL) 220 #define PMAP_HEAD_TO_MAP_UNLOCK() \ 221 (void) spinlockmgr(&pmap_main_lock, LK_RELEASE, NULL) 222 #else 223 #define PMAP_MAP_TO_HEAD_LOCK() /* nothing */ 224 #define PMAP_MAP_TO_HEAD_UNLOCK() /* nothing */ 225 #define PMAP_HEAD_TO_MAP_LOCK() /* nothing */ 226 #define PMAP_HEAD_TO_MAP_UNLOCK() /* nothing */ 227 #endif /* LOCKDEBUG */ 228 229 /* 230 * pv_page management structures: locked by pvalloc_lock 231 */ 232 233 TAILQ_HEAD(pv_pagelist, pv_page); 234 static struct pv_pagelist pv_freepages; /* list of pv_pages with free entrys */ 235 static struct pv_pagelist pv_unusedpgs; /* list of unused pv_pages */ 236 static int pv_nfpvents; /* # of free pv entries */ 237 static struct pv_page *pv_initpage; /* bootstrap page from kernel_map */ 238 static vaddr_t pv_cachedva; /* cached VA for later use */ 239 240 #define PVE_LOWAT (PVE_PER_PVPAGE / 2) /* free pv_entry low water mark */ 241 #define PVE_HIWAT (PVE_LOWAT + (PVE_PER_PVPAGE * 2)) 242 /* high water mark */ 243 244 /* 245 * local prototypes 246 */ 247 248 static struct pv_entry *pmap_add_pvpage __P((struct pv_page *, boolean_t)); 249 static struct pv_entry *pmap_alloc_pv __P((struct pmap *, int)); /* see codes below */ 250 #define ALLOCPV_NEED 0 /* need PV now */ 251 #define ALLOCPV_TRY 1 /* just try to allocate, don't steal */ 252 #define ALLOCPV_NONEED 2 /* don't need PV, just growing cache */ 253 static struct pv_entry *pmap_alloc_pvpage __P((struct pmap *, int)); 254 static void pmap_enter_pv __P((struct vm_page *, 255 struct pv_entry *, struct pmap *, 256 vaddr_t, struct vm_page *, int)); 257 static void pmap_free_pv __P((struct pmap *, struct pv_entry *)); 258 static void pmap_free_pvs __P((struct pmap *, struct pv_entry *)); 259 static void pmap_free_pv_doit __P((struct pv_entry *)); 260 static void pmap_free_pvpage __P((void)); 261 static boolean_t pmap_is_curpmap __P((struct pmap *)); 262 static struct pv_entry *pmap_remove_pv __P((struct vm_page *, struct pmap *, 263 vaddr_t)); 264 #define PMAP_REMOVE_ALL 0 /* remove all mappings */ 265 #define PMAP_REMOVE_SKIPWIRED 1 /* skip wired mappings */ 266 267 static u_int pmap_modify_pv __P((struct pmap *, vaddr_t, struct vm_page *, 268 u_int, u_int)); 269 270 static void pmap_free_l1pt __P((struct l1pt *)); 271 static int pmap_allocpagedir __P((struct pmap *)); 272 static int pmap_clean_page __P((struct pv_entry *, boolean_t)); 273 static void pmap_remove_all __P((struct vm_page *)); 274 275 276 vsize_t npages; 277 278 static struct vm_page *pmap_alloc_ptp __P((struct pmap *, vaddr_t)); 279 static struct vm_page *pmap_get_ptp __P((struct pmap *, vaddr_t)); 280 __inline static void pmap_clearbit __P((struct vm_page *, unsigned int)); 281 282 extern paddr_t physical_start; 283 extern paddr_t physical_freestart; 284 extern paddr_t physical_end; 285 extern paddr_t physical_freeend; 286 extern unsigned int free_pages; 287 extern int max_processes; 288 289 vaddr_t virtual_avail; 290 vaddr_t virtual_end; 291 vaddr_t pmap_curmaxkvaddr; 292 293 vaddr_t avail_start; 294 vaddr_t avail_end; 295 296 extern pv_addr_t systempage; 297 298 /* Variables used by the L1 page table queue code */ 299 SIMPLEQ_HEAD(l1pt_queue, l1pt); 300 struct l1pt_queue l1pt_static_queue; /* head of our static l1 queue */ 301 int l1pt_static_queue_count; /* items in the static l1 queue */ 302 int l1pt_static_create_count; /* static l1 items created */ 303 struct l1pt_queue l1pt_queue; /* head of our l1 queue */ 304 int l1pt_queue_count; /* items in the l1 queue */ 305 int l1pt_create_count; /* stat - L1's create count */ 306 int l1pt_reuse_count; /* stat - L1's reused count */ 307 308 /* Local function prototypes (not used outside this file) */ 309 void pmap_copy_on_write __P((struct vm_page *)); 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 vm_page *, 325 pt_entry_t *, boolean_t)); 326 static void pmap_vac_me_kpmap __P((struct pmap *, struct vm_page *, 327 pt_entry_t *, boolean_t)); 328 static void pmap_vac_me_user __P((struct pmap *, struct vm_page *, 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 { 399 400 if ((curproc && curproc->p_vmspace->vm_map.pmap == pmap) || 401 pmap == pmap_kernel()) 402 return (TRUE); 403 404 return (FALSE); 405 } 406 407 #include "isadma.h" 408 409 #if NISADMA > 0 410 /* 411 * Used to protect memory for ISA DMA bounce buffers. If, when loading 412 * pages into the system, memory intersects with any of these ranges, 413 * the intersecting memory will be loaded into a lower-priority free list. 414 */ 415 bus_dma_segment_t *pmap_isa_dma_ranges; 416 int pmap_isa_dma_nranges; 417 418 boolean_t pmap_isa_dma_range_intersect __P((paddr_t, psize_t, 419 paddr_t *, psize_t *)); 420 421 /* 422 * Check if a memory range intersects with an ISA DMA range, and 423 * return the page-rounded intersection if it does. The intersection 424 * will be placed on a lower-priority free list. 425 */ 426 boolean_t 427 pmap_isa_dma_range_intersect(pa, size, pap, sizep) 428 paddr_t pa; 429 psize_t size; 430 paddr_t *pap; 431 psize_t *sizep; 432 { 433 bus_dma_segment_t *ds; 434 int i; 435 436 if (pmap_isa_dma_ranges == NULL) 437 return (FALSE); 438 439 for (i = 0, ds = pmap_isa_dma_ranges; 440 i < pmap_isa_dma_nranges; i++, ds++) { 441 if (ds->ds_addr <= pa && pa < (ds->ds_addr + ds->ds_len)) { 442 /* 443 * Beginning of region intersects with this range. 444 */ 445 *pap = trunc_page(pa); 446 *sizep = round_page(min(pa + size, 447 ds->ds_addr + ds->ds_len) - pa); 448 return (TRUE); 449 } 450 if (pa < ds->ds_addr && ds->ds_addr < (pa + size)) { 451 /* 452 * End of region intersects with this range. 453 */ 454 *pap = trunc_page(ds->ds_addr); 455 *sizep = round_page(min((pa + size) - ds->ds_addr, 456 ds->ds_len)); 457 return (TRUE); 458 } 459 } 460 461 /* 462 * No intersection found. 463 */ 464 return (FALSE); 465 } 466 #endif /* NISADMA > 0 */ 467 468 /* 469 * p v _ e n t r y f u n c t i o n s 470 */ 471 472 /* 473 * pv_entry allocation functions: 474 * the main pv_entry allocation functions are: 475 * pmap_alloc_pv: allocate a pv_entry structure 476 * pmap_free_pv: free one pv_entry 477 * pmap_free_pvs: free a list of pv_entrys 478 * 479 * the rest are helper functions 480 */ 481 482 /* 483 * pmap_alloc_pv: inline function to allocate a pv_entry structure 484 * => we lock pvalloc_lock 485 * => if we fail, we call out to pmap_alloc_pvpage 486 * => 3 modes: 487 * ALLOCPV_NEED = we really need a pv_entry, even if we have to steal it 488 * ALLOCPV_TRY = we want a pv_entry, but not enough to steal 489 * ALLOCPV_NONEED = we are trying to grow our free list, don't really need 490 * one now 491 * 492 * "try" is for optional functions like pmap_copy(). 493 */ 494 495 __inline static struct pv_entry * 496 pmap_alloc_pv(pmap, mode) 497 struct pmap *pmap; 498 int mode; 499 { 500 struct pv_page *pvpage; 501 struct pv_entry *pv; 502 503 simple_lock(&pvalloc_lock); 504 505 pvpage = TAILQ_FIRST(&pv_freepages); 506 507 if (pvpage != NULL) { 508 pvpage->pvinfo.pvpi_nfree--; 509 if (pvpage->pvinfo.pvpi_nfree == 0) { 510 /* nothing left in this one? */ 511 TAILQ_REMOVE(&pv_freepages, pvpage, pvinfo.pvpi_list); 512 } 513 pv = pvpage->pvinfo.pvpi_pvfree; 514 KASSERT(pv); 515 pvpage->pvinfo.pvpi_pvfree = pv->pv_next; 516 pv_nfpvents--; /* took one from pool */ 517 } else { 518 pv = NULL; /* need more of them */ 519 } 520 521 /* 522 * if below low water mark or we didn't get a pv_entry we try and 523 * create more pv_entrys ... 524 */ 525 526 if (pv_nfpvents < PVE_LOWAT || pv == NULL) { 527 if (pv == NULL) 528 pv = pmap_alloc_pvpage(pmap, (mode == ALLOCPV_TRY) ? 529 mode : ALLOCPV_NEED); 530 else 531 (void) pmap_alloc_pvpage(pmap, ALLOCPV_NONEED); 532 } 533 534 simple_unlock(&pvalloc_lock); 535 return(pv); 536 } 537 538 /* 539 * pmap_alloc_pvpage: maybe allocate a new pvpage 540 * 541 * if need_entry is false: try and allocate a new pv_page 542 * if need_entry is true: try and allocate a new pv_page and return a 543 * new pv_entry from it. if we are unable to allocate a pv_page 544 * we make a last ditch effort to steal a pv_page from some other 545 * mapping. if that fails, we panic... 546 * 547 * => we assume that the caller holds pvalloc_lock 548 */ 549 550 static struct pv_entry * 551 pmap_alloc_pvpage(pmap, mode) 552 struct pmap *pmap; 553 int mode; 554 { 555 struct vm_page *pg; 556 struct pv_page *pvpage; 557 struct pv_entry *pv; 558 int s; 559 560 /* 561 * if we need_entry and we've got unused pv_pages, allocate from there 562 */ 563 564 pvpage = TAILQ_FIRST(&pv_unusedpgs); 565 if (mode != ALLOCPV_NONEED && pvpage != NULL) { 566 567 /* move it to pv_freepages list */ 568 TAILQ_REMOVE(&pv_unusedpgs, pvpage, pvinfo.pvpi_list); 569 TAILQ_INSERT_HEAD(&pv_freepages, pvpage, pvinfo.pvpi_list); 570 571 /* allocate a pv_entry */ 572 pvpage->pvinfo.pvpi_nfree--; /* can't go to zero */ 573 pv = pvpage->pvinfo.pvpi_pvfree; 574 KASSERT(pv); 575 pvpage->pvinfo.pvpi_pvfree = pv->pv_next; 576 577 pv_nfpvents--; /* took one from pool */ 578 return(pv); 579 } 580 581 /* 582 * see if we've got a cached unmapped VA that we can map a page in. 583 * if not, try to allocate one. 584 */ 585 586 587 if (pv_cachedva == 0) { 588 s = splvm(); 589 pv_cachedva = uvm_km_kmemalloc(kmem_map, NULL, 590 PAGE_SIZE, UVM_KMF_TRYLOCK|UVM_KMF_VALLOC); 591 splx(s); 592 if (pv_cachedva == 0) { 593 return (NULL); 594 } 595 } 596 597 pg = uvm_pagealloc(NULL, pv_cachedva - vm_map_min(kernel_map), NULL, 598 UVM_PGA_USERESERVE); 599 600 if (pg == NULL) 601 return (NULL); 602 pg->flags &= ~PG_BUSY; /* never busy */ 603 604 /* 605 * add a mapping for our new pv_page and free its entrys (save one!) 606 * 607 * NOTE: If we are allocating a PV page for the kernel pmap, the 608 * pmap is already locked! (...but entering the mapping is safe...) 609 */ 610 611 pmap_kenter_pa(pv_cachedva, VM_PAGE_TO_PHYS(pg), 612 VM_PROT_READ|VM_PROT_WRITE); 613 pmap_update(pmap_kernel()); 614 pvpage = (struct pv_page *) pv_cachedva; 615 pv_cachedva = 0; 616 return (pmap_add_pvpage(pvpage, mode != ALLOCPV_NONEED)); 617 } 618 619 /* 620 * pmap_add_pvpage: add a pv_page's pv_entrys to the free list 621 * 622 * => caller must hold pvalloc_lock 623 * => if need_entry is true, we allocate and return one pv_entry 624 */ 625 626 static struct pv_entry * 627 pmap_add_pvpage(pvp, need_entry) 628 struct pv_page *pvp; 629 boolean_t need_entry; 630 { 631 int tofree, lcv; 632 633 /* do we need to return one? */ 634 tofree = (need_entry) ? PVE_PER_PVPAGE - 1 : PVE_PER_PVPAGE; 635 636 pvp->pvinfo.pvpi_pvfree = NULL; 637 pvp->pvinfo.pvpi_nfree = tofree; 638 for (lcv = 0 ; lcv < tofree ; lcv++) { 639 pvp->pvents[lcv].pv_next = pvp->pvinfo.pvpi_pvfree; 640 pvp->pvinfo.pvpi_pvfree = &pvp->pvents[lcv]; 641 } 642 if (need_entry) 643 TAILQ_INSERT_TAIL(&pv_freepages, pvp, pvinfo.pvpi_list); 644 else 645 TAILQ_INSERT_TAIL(&pv_unusedpgs, pvp, pvinfo.pvpi_list); 646 pv_nfpvents += tofree; 647 return((need_entry) ? &pvp->pvents[lcv] : NULL); 648 } 649 650 /* 651 * pmap_free_pv_doit: actually free a pv_entry 652 * 653 * => do not call this directly! instead use either 654 * 1. pmap_free_pv ==> free a single pv_entry 655 * 2. pmap_free_pvs => free a list of pv_entrys 656 * => we must be holding pvalloc_lock 657 */ 658 659 __inline static void 660 pmap_free_pv_doit(pv) 661 struct pv_entry *pv; 662 { 663 struct pv_page *pvp; 664 665 pvp = (struct pv_page *) arm_trunc_page((vaddr_t)pv); 666 pv_nfpvents++; 667 pvp->pvinfo.pvpi_nfree++; 668 669 /* nfree == 1 => fully allocated page just became partly allocated */ 670 if (pvp->pvinfo.pvpi_nfree == 1) { 671 TAILQ_INSERT_HEAD(&pv_freepages, pvp, pvinfo.pvpi_list); 672 } 673 674 /* free it */ 675 pv->pv_next = pvp->pvinfo.pvpi_pvfree; 676 pvp->pvinfo.pvpi_pvfree = pv; 677 678 /* 679 * are all pv_page's pv_entry's free? move it to unused queue. 680 */ 681 682 if (pvp->pvinfo.pvpi_nfree == PVE_PER_PVPAGE) { 683 TAILQ_REMOVE(&pv_freepages, pvp, pvinfo.pvpi_list); 684 TAILQ_INSERT_HEAD(&pv_unusedpgs, pvp, pvinfo.pvpi_list); 685 } 686 } 687 688 /* 689 * pmap_free_pv: free a single pv_entry 690 * 691 * => we gain the pvalloc_lock 692 */ 693 694 __inline static void 695 pmap_free_pv(pmap, pv) 696 struct pmap *pmap; 697 struct pv_entry *pv; 698 { 699 simple_lock(&pvalloc_lock); 700 pmap_free_pv_doit(pv); 701 702 /* 703 * Can't free the PV page if the PV entries were associated with 704 * the kernel pmap; the pmap is already locked. 705 */ 706 if (pv_nfpvents > PVE_HIWAT && TAILQ_FIRST(&pv_unusedpgs) != NULL && 707 pmap != pmap_kernel()) 708 pmap_free_pvpage(); 709 710 simple_unlock(&pvalloc_lock); 711 } 712 713 /* 714 * pmap_free_pvs: free a list of pv_entrys 715 * 716 * => we gain the pvalloc_lock 717 */ 718 719 __inline static void 720 pmap_free_pvs(pmap, pvs) 721 struct pmap *pmap; 722 struct pv_entry *pvs; 723 { 724 struct pv_entry *nextpv; 725 726 simple_lock(&pvalloc_lock); 727 728 for ( /* null */ ; pvs != NULL ; pvs = nextpv) { 729 nextpv = pvs->pv_next; 730 pmap_free_pv_doit(pvs); 731 } 732 733 /* 734 * Can't free the PV page if the PV entries were associated with 735 * the kernel pmap; the pmap is already locked. 736 */ 737 if (pv_nfpvents > PVE_HIWAT && TAILQ_FIRST(&pv_unusedpgs) != NULL && 738 pmap != pmap_kernel()) 739 pmap_free_pvpage(); 740 741 simple_unlock(&pvalloc_lock); 742 } 743 744 745 /* 746 * pmap_free_pvpage: try and free an unused pv_page structure 747 * 748 * => assume caller is holding the pvalloc_lock and that 749 * there is a page on the pv_unusedpgs list 750 * => if we can't get a lock on the kmem_map we try again later 751 */ 752 753 static void 754 pmap_free_pvpage() 755 { 756 int s; 757 struct vm_map *map; 758 struct vm_map_entry *dead_entries; 759 struct pv_page *pvp; 760 761 s = splvm(); /* protect kmem_map */ 762 763 pvp = TAILQ_FIRST(&pv_unusedpgs); 764 765 /* 766 * note: watch out for pv_initpage which is allocated out of 767 * kernel_map rather than kmem_map. 768 */ 769 if (pvp == pv_initpage) 770 map = kernel_map; 771 else 772 map = kmem_map; 773 if (vm_map_lock_try(map)) { 774 775 /* remove pvp from pv_unusedpgs */ 776 TAILQ_REMOVE(&pv_unusedpgs, pvp, pvinfo.pvpi_list); 777 778 /* unmap the page */ 779 dead_entries = NULL; 780 uvm_unmap_remove(map, (vaddr_t)pvp, ((vaddr_t)pvp) + PAGE_SIZE, 781 &dead_entries); 782 vm_map_unlock(map); 783 784 if (dead_entries != NULL) 785 uvm_unmap_detach(dead_entries, 0); 786 787 pv_nfpvents -= PVE_PER_PVPAGE; /* update free count */ 788 } 789 if (pvp == pv_initpage) 790 /* no more initpage, we've freed it */ 791 pv_initpage = NULL; 792 793 splx(s); 794 } 795 796 /* 797 * main pv_entry manipulation functions: 798 * pmap_enter_pv: enter a mapping onto a vm_page list 799 * pmap_remove_pv: remove a mappiing from a vm_page list 800 * 801 * NOTE: pmap_enter_pv expects to lock the pvh itself 802 * pmap_remove_pv expects te caller to lock the pvh before calling 803 */ 804 805 /* 806 * pmap_enter_pv: enter a mapping onto a vm_page lst 807 * 808 * => caller should hold the proper lock on pmap_main_lock 809 * => caller should have pmap locked 810 * => we will gain the lock on the vm_page and allocate the new pv_entry 811 * => caller should adjust ptp's wire_count before calling 812 * => caller should not adjust pmap's wire_count 813 */ 814 815 __inline static void 816 pmap_enter_pv(pg, pve, pmap, va, ptp, flags) 817 struct vm_page *pg; 818 struct pv_entry *pve; /* preallocated pve for us to use */ 819 struct pmap *pmap; 820 vaddr_t va; 821 struct vm_page *ptp; /* PTP in pmap that maps this VA */ 822 int flags; 823 { 824 pve->pv_pmap = pmap; 825 pve->pv_va = va; 826 pve->pv_ptp = ptp; /* NULL for kernel pmap */ 827 pve->pv_flags = flags; 828 simple_lock(&pg->mdpage.pvh_slock); /* lock vm_page */ 829 pve->pv_next = pg->mdpage.pvh_list; /* add to ... */ 830 pg->mdpage.pvh_list = pve; /* ... locked list */ 831 simple_unlock(&pg->mdpage.pvh_slock); /* unlock, done! */ 832 if (pve->pv_flags & PT_W) 833 ++pmap->pm_stats.wired_count; 834 } 835 836 /* 837 * pmap_remove_pv: try to remove a mapping from a pv_list 838 * 839 * => caller should hold proper lock on pmap_main_lock 840 * => pmap should be locked 841 * => caller should hold lock on vm_page [so that attrs can be adjusted] 842 * => caller should adjust ptp's wire_count and free PTP if needed 843 * => caller should NOT adjust pmap's wire_count 844 * => we return the removed pve 845 */ 846 847 __inline static struct pv_entry * 848 pmap_remove_pv(pg, pmap, va) 849 struct vm_page *pg; 850 struct pmap *pmap; 851 vaddr_t va; 852 { 853 struct pv_entry *pve, **prevptr; 854 855 prevptr = &pg->mdpage.pvh_list; /* previous pv_entry pointer */ 856 pve = *prevptr; 857 while (pve) { 858 if (pve->pv_pmap == pmap && pve->pv_va == va) { /* match? */ 859 *prevptr = pve->pv_next; /* remove it! */ 860 if (pve->pv_flags & PT_W) 861 --pmap->pm_stats.wired_count; 862 break; 863 } 864 prevptr = &pve->pv_next; /* previous pointer */ 865 pve = pve->pv_next; /* advance */ 866 } 867 return(pve); /* return removed pve */ 868 } 869 870 /* 871 * 872 * pmap_modify_pv: Update pv flags 873 * 874 * => caller should hold lock on vm_page [so that attrs can be adjusted] 875 * => caller should NOT adjust pmap's wire_count 876 * => caller must call pmap_vac_me_harder() if writable status of a page 877 * may have changed. 878 * => we return the old flags 879 * 880 * Modify a physical-virtual mapping in the pv table 881 */ 882 883 /*__inline */ 884 static u_int 885 pmap_modify_pv(pmap, va, pg, bic_mask, eor_mask) 886 struct pmap *pmap; 887 vaddr_t va; 888 struct vm_page *pg; 889 u_int bic_mask; 890 u_int eor_mask; 891 { 892 struct pv_entry *npv; 893 u_int flags, oflags; 894 895 /* 896 * There is at least one VA mapping this page. 897 */ 898 899 for (npv = pg->mdpage.pvh_list; npv; npv = npv->pv_next) { 900 if (pmap == npv->pv_pmap && va == npv->pv_va) { 901 oflags = npv->pv_flags; 902 npv->pv_flags = flags = 903 ((oflags & ~bic_mask) ^ eor_mask); 904 if ((flags ^ oflags) & PT_W) { 905 if (flags & PT_W) 906 ++pmap->pm_stats.wired_count; 907 else 908 --pmap->pm_stats.wired_count; 909 } 910 return (oflags); 911 } 912 } 913 return (0); 914 } 915 916 /* 917 * Map the specified level 2 pagetable into the level 1 page table for 918 * the given pmap to cover a chunk of virtual address space starting from the 919 * address specified. 920 */ 921 static /*__inline*/ void 922 pmap_map_in_l1(pmap, va, l2pa, selfref) 923 struct pmap *pmap; 924 vaddr_t va, l2pa; 925 boolean_t selfref; 926 { 927 vaddr_t ptva; 928 929 /* Calculate the index into the L1 page table. */ 930 ptva = (va >> PDSHIFT) & ~3; 931 932 NPDEBUG(PDB_MAP_L1, printf("wiring %08lx in to pd%p pte0x%lx va0x%lx\n", l2pa, 933 pmap->pm_pdir, L1_PTE(l2pa), ptva)); 934 935 /* Map page table into the L1. */ 936 pmap->pm_pdir[ptva + 0] = L1_PTE(l2pa + 0x000); 937 pmap->pm_pdir[ptva + 1] = L1_PTE(l2pa + 0x400); 938 pmap->pm_pdir[ptva + 2] = L1_PTE(l2pa + 0x800); 939 pmap->pm_pdir[ptva + 3] = L1_PTE(l2pa + 0xc00); 940 941 /* Map the page table into the page table area. */ 942 if (selfref) { 943 NPDEBUG(PDB_MAP_L1, printf("pt self reference %lx in %lx\n", 944 L2_PTE_NC_NB(l2pa, AP_KRW), pmap->pm_vptpt)); 945 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) = 946 L2_PTE_NC_NB(l2pa, AP_KRW); 947 } 948 /* XXX should be a purge */ 949 /* cpu_tlb_flushD();*/ 950 } 951 952 #if 0 953 static /*__inline*/ void 954 pmap_unmap_in_l1(pmap, va) 955 struct pmap *pmap; 956 vaddr_t va; 957 { 958 vaddr_t ptva; 959 960 /* Calculate the index into the L1 page table. */ 961 ptva = (va >> PDSHIFT) & ~3; 962 963 /* Unmap page table from the L1. */ 964 pmap->pm_pdir[ptva + 0] = 0; 965 pmap->pm_pdir[ptva + 1] = 0; 966 pmap->pm_pdir[ptva + 2] = 0; 967 pmap->pm_pdir[ptva + 3] = 0; 968 969 /* Unmap the page table from the page table area. */ 970 *((pt_entry_t *)(pmap->pm_vptpt + ptva)) = 0; 971 972 /* XXX should be a purge */ 973 /* cpu_tlb_flushD();*/ 974 } 975 #endif 976 977 /* 978 * Used to map a range of physical addresses into kernel 979 * virtual address space. 980 * 981 * For now, VM is already on, we only need to map the 982 * specified memory. 983 */ 984 vaddr_t 985 pmap_map(va, spa, epa, prot) 986 vaddr_t va, spa, epa; 987 int prot; 988 { 989 while (spa < epa) { 990 pmap_kenter_pa(va, spa, prot); 991 va += NBPG; 992 spa += NBPG; 993 } 994 pmap_update(pmap_kernel()); 995 return(va); 996 } 997 998 999 /* 1000 * void pmap_bootstrap(pd_entry_t *kernel_l1pt, pv_addr_t kernel_ptpt) 1001 * 1002 * bootstrap the pmap system. This is called from initarm and allows 1003 * the pmap system to initailise any structures it requires. 1004 * 1005 * Currently this sets up the kernel_pmap that is statically allocated 1006 * and also allocated virtual addresses for certain page hooks. 1007 * Currently the only one page hook is allocated that is used 1008 * to zero physical pages of memory. 1009 * It also initialises the start and end address of the kernel data space. 1010 */ 1011 extern paddr_t physical_freestart; 1012 extern paddr_t physical_freeend; 1013 1014 char *boot_head; 1015 1016 void 1017 pmap_bootstrap(kernel_l1pt, kernel_ptpt) 1018 pd_entry_t *kernel_l1pt; 1019 pv_addr_t kernel_ptpt; 1020 { 1021 pt_entry_t *pte; 1022 int loop; 1023 paddr_t start, end; 1024 #if NISADMA > 0 1025 paddr_t istart; 1026 psize_t isize; 1027 #endif 1028 1029 pmap_kernel()->pm_pdir = kernel_l1pt; 1030 pmap_kernel()->pm_pptpt = kernel_ptpt.pv_pa; 1031 pmap_kernel()->pm_vptpt = kernel_ptpt.pv_va; 1032 simple_lock_init(&pmap_kernel()->pm_lock); 1033 pmap_kernel()->pm_obj.pgops = NULL; 1034 TAILQ_INIT(&(pmap_kernel()->pm_obj.memq)); 1035 pmap_kernel()->pm_obj.uo_npages = 0; 1036 pmap_kernel()->pm_obj.uo_refs = 1; 1037 1038 /* 1039 * Initialize PAGE_SIZE-dependent variables. 1040 */ 1041 uvm_setpagesize(); 1042 1043 npages = 0; 1044 loop = 0; 1045 while (loop < bootconfig.dramblocks) { 1046 start = (paddr_t)bootconfig.dram[loop].address; 1047 end = start + (bootconfig.dram[loop].pages * NBPG); 1048 if (start < physical_freestart) 1049 start = physical_freestart; 1050 if (end > physical_freeend) 1051 end = physical_freeend; 1052 #if 0 1053 printf("%d: %lx -> %lx\n", loop, start, end - 1); 1054 #endif 1055 #if NISADMA > 0 1056 if (pmap_isa_dma_range_intersect(start, end - start, 1057 &istart, &isize)) { 1058 /* 1059 * Place the pages that intersect with the 1060 * ISA DMA range onto the ISA DMA free list. 1061 */ 1062 #if 0 1063 printf(" ISADMA 0x%lx -> 0x%lx\n", istart, 1064 istart + isize - 1); 1065 #endif 1066 uvm_page_physload(atop(istart), 1067 atop(istart + isize), atop(istart), 1068 atop(istart + isize), VM_FREELIST_ISADMA); 1069 npages += atop(istart + isize) - atop(istart); 1070 1071 /* 1072 * Load the pieces that come before 1073 * the intersection into the default 1074 * free list. 1075 */ 1076 if (start < istart) { 1077 #if 0 1078 printf(" BEFORE 0x%lx -> 0x%lx\n", 1079 start, istart - 1); 1080 #endif 1081 uvm_page_physload(atop(start), 1082 atop(istart), atop(start), 1083 atop(istart), VM_FREELIST_DEFAULT); 1084 npages += atop(istart) - atop(start); 1085 } 1086 1087 /* 1088 * Load the pieces that come after 1089 * the intersection into the default 1090 * free list. 1091 */ 1092 if ((istart + isize) < end) { 1093 #if 0 1094 printf(" AFTER 0x%lx -> 0x%lx\n", 1095 (istart + isize), end - 1); 1096 #endif 1097 uvm_page_physload(atop(istart + isize), 1098 atop(end), atop(istart + isize), 1099 atop(end), VM_FREELIST_DEFAULT); 1100 npages += atop(end) - atop(istart + isize); 1101 } 1102 } else { 1103 uvm_page_physload(atop(start), atop(end), 1104 atop(start), atop(end), VM_FREELIST_DEFAULT); 1105 npages += atop(end) - atop(start); 1106 } 1107 #else /* NISADMA > 0 */ 1108 uvm_page_physload(atop(start), atop(end), 1109 atop(start), atop(end), VM_FREELIST_DEFAULT); 1110 npages += atop(end) - atop(start); 1111 #endif /* NISADMA > 0 */ 1112 ++loop; 1113 } 1114 1115 #ifdef MYCROFT_HACK 1116 printf("npages = %ld\n", npages); 1117 #endif 1118 1119 virtual_avail = KERNEL_VM_BASE; 1120 virtual_end = KERNEL_VM_BASE + KERNEL_VM_SIZE - 1; 1121 1122 /* 1123 * now we allocate the "special" VAs which are used for tmp mappings 1124 * by the pmap (and other modules). we allocate the VAs by advancing 1125 * virtual_avail (note that there are no pages mapped at these VAs). 1126 * we find the PTE that maps the allocated VA via the linear PTE 1127 * mapping. 1128 */ 1129 1130 pte = ((pt_entry_t *) PTE_BASE) + atop(virtual_avail); 1131 1132 csrcp = virtual_avail; csrc_pte = pte; 1133 virtual_avail += PAGE_SIZE; pte++; 1134 1135 cdstp = virtual_avail; cdst_pte = pte; 1136 virtual_avail += PAGE_SIZE; pte++; 1137 1138 memhook = (char *) virtual_avail; /* don't need pte */ 1139 virtual_avail += PAGE_SIZE; pte++; 1140 1141 msgbufaddr = (caddr_t) virtual_avail; /* don't need pte */ 1142 virtual_avail += round_page(MSGBUFSIZE); 1143 pte += atop(round_page(MSGBUFSIZE)); 1144 1145 /* 1146 * init the static-global locks and global lists. 1147 */ 1148 spinlockinit(&pmap_main_lock, "pmaplk", 0); 1149 simple_lock_init(&pvalloc_lock); 1150 simple_lock_init(&pmaps_lock); 1151 LIST_INIT(&pmaps); 1152 TAILQ_INIT(&pv_freepages); 1153 TAILQ_INIT(&pv_unusedpgs); 1154 1155 /* 1156 * initialize the pmap pool. 1157 */ 1158 1159 pool_init(&pmap_pmap_pool, sizeof(struct pmap), 0, 0, 0, "pmappl", 1160 &pool_allocator_nointr); 1161 1162 cpu_dcache_wbinv_all(); 1163 } 1164 1165 /* 1166 * void pmap_init(void) 1167 * 1168 * Initialize the pmap module. 1169 * Called by vm_init() in vm/vm_init.c in order to initialise 1170 * any structures that the pmap system needs to map virtual memory. 1171 */ 1172 1173 extern int physmem; 1174 1175 void 1176 pmap_init() 1177 { 1178 1179 /* 1180 * Set the available memory vars - These do not map to real memory 1181 * addresses and cannot as the physical memory is fragmented. 1182 * They are used by ps for %mem calculations. 1183 * One could argue whether this should be the entire memory or just 1184 * the memory that is useable in a user process. 1185 */ 1186 avail_start = 0; 1187 avail_end = physmem * NBPG; 1188 1189 /* 1190 * now we need to free enough pv_entry structures to allow us to get 1191 * the kmem_map/kmem_object allocated and inited (done after this 1192 * function is finished). to do this we allocate one bootstrap page out 1193 * of kernel_map and use it to provide an initial pool of pv_entry 1194 * structures. we never free this page. 1195 */ 1196 1197 pv_initpage = (struct pv_page *) uvm_km_alloc(kernel_map, PAGE_SIZE); 1198 if (pv_initpage == NULL) 1199 panic("pmap_init: pv_initpage"); 1200 pv_cachedva = 0; /* a VA we have allocated but not used yet */ 1201 pv_nfpvents = 0; 1202 (void) pmap_add_pvpage(pv_initpage, FALSE); 1203 1204 pmap_initialized = TRUE; 1205 1206 /* Initialise our L1 page table queues and counters */ 1207 SIMPLEQ_INIT(&l1pt_static_queue); 1208 l1pt_static_queue_count = 0; 1209 l1pt_static_create_count = 0; 1210 SIMPLEQ_INIT(&l1pt_queue); 1211 l1pt_queue_count = 0; 1212 l1pt_create_count = 0; 1213 l1pt_reuse_count = 0; 1214 } 1215 1216 /* 1217 * pmap_postinit() 1218 * 1219 * This routine is called after the vm and kmem subsystems have been 1220 * initialised. This allows the pmap code to perform any initialisation 1221 * that can only be done one the memory allocation is in place. 1222 */ 1223 1224 void 1225 pmap_postinit() 1226 { 1227 int loop; 1228 struct l1pt *pt; 1229 1230 #ifdef PMAP_STATIC_L1S 1231 for (loop = 0; loop < PMAP_STATIC_L1S; ++loop) { 1232 #else /* PMAP_STATIC_L1S */ 1233 for (loop = 0; loop < max_processes; ++loop) { 1234 #endif /* PMAP_STATIC_L1S */ 1235 /* Allocate a L1 page table */ 1236 pt = pmap_alloc_l1pt(); 1237 if (!pt) 1238 panic("Cannot allocate static L1 page tables\n"); 1239 1240 /* Clean it */ 1241 bzero((void *)pt->pt_va, PD_SIZE); 1242 pt->pt_flags |= (PTFLAG_STATIC | PTFLAG_CLEAN); 1243 /* Add the page table to the queue */ 1244 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pt, pt_queue); 1245 ++l1pt_static_queue_count; 1246 ++l1pt_static_create_count; 1247 } 1248 } 1249 1250 1251 /* 1252 * Create and return a physical map. 1253 * 1254 * If the size specified for the map is zero, the map is an actual physical 1255 * map, and may be referenced by the hardware. 1256 * 1257 * If the size specified is non-zero, the map will be used in software only, 1258 * and is bounded by that size. 1259 */ 1260 1261 pmap_t 1262 pmap_create() 1263 { 1264 struct pmap *pmap; 1265 1266 /* 1267 * Fetch pmap entry from the pool 1268 */ 1269 1270 pmap = pool_get(&pmap_pmap_pool, PR_WAITOK); 1271 /* XXX is this really needed! */ 1272 memset(pmap, 0, sizeof(*pmap)); 1273 1274 simple_lock_init(&pmap->pm_obj.vmobjlock); 1275 pmap->pm_obj.pgops = NULL; /* currently not a mappable object */ 1276 TAILQ_INIT(&pmap->pm_obj.memq); 1277 pmap->pm_obj.uo_npages = 0; 1278 pmap->pm_obj.uo_refs = 1; 1279 pmap->pm_stats.wired_count = 0; 1280 pmap->pm_stats.resident_count = 1; 1281 1282 /* Now init the machine part of the pmap */ 1283 pmap_pinit(pmap); 1284 return(pmap); 1285 } 1286 1287 /* 1288 * pmap_alloc_l1pt() 1289 * 1290 * This routine allocates physical and virtual memory for a L1 page table 1291 * and wires it. 1292 * A l1pt structure is returned to describe the allocated page table. 1293 * 1294 * This routine is allowed to fail if the required memory cannot be allocated. 1295 * In this case NULL is returned. 1296 */ 1297 1298 struct l1pt * 1299 pmap_alloc_l1pt(void) 1300 { 1301 paddr_t pa; 1302 vaddr_t va; 1303 struct l1pt *pt; 1304 int error; 1305 struct vm_page *m; 1306 pt_entry_t *ptes; 1307 1308 /* Allocate virtual address space for the L1 page table */ 1309 va = uvm_km_valloc(kernel_map, PD_SIZE); 1310 if (va == 0) { 1311 #ifdef DIAGNOSTIC 1312 PDEBUG(0, 1313 printf("pmap: Cannot allocate pageable memory for L1\n")); 1314 #endif /* DIAGNOSTIC */ 1315 return(NULL); 1316 } 1317 1318 /* Allocate memory for the l1pt structure */ 1319 pt = (struct l1pt *)malloc(sizeof(struct l1pt), M_VMPMAP, M_WAITOK); 1320 1321 /* 1322 * Allocate pages from the VM system. 1323 */ 1324 TAILQ_INIT(&pt->pt_plist); 1325 error = uvm_pglistalloc(PD_SIZE, physical_start, physical_end, 1326 PD_SIZE, 0, &pt->pt_plist, 1, M_WAITOK); 1327 if (error) { 1328 #ifdef DIAGNOSTIC 1329 PDEBUG(0, 1330 printf("pmap: Cannot allocate physical mem for L1 (%d)\n", 1331 error)); 1332 #endif /* DIAGNOSTIC */ 1333 /* Release the resources we already have claimed */ 1334 free(pt, M_VMPMAP); 1335 uvm_km_free(kernel_map, va, PD_SIZE); 1336 return(NULL); 1337 } 1338 1339 /* Map our physical pages into our virtual space */ 1340 pt->pt_va = va; 1341 m = TAILQ_FIRST(&pt->pt_plist); 1342 ptes = pmap_map_ptes(pmap_kernel()); 1343 while (m && va < (pt->pt_va + PD_SIZE)) { 1344 pa = VM_PAGE_TO_PHYS(m); 1345 1346 pmap_kenter_pa(va, pa, VM_PROT_READ | VM_PROT_WRITE); 1347 1348 /* Revoke cacheability and bufferability */ 1349 /* XXX should be done better than this */ 1350 ptes[arm_btop(va)] &= ~(PT_C | PT_B); 1351 1352 va += NBPG; 1353 m = m->pageq.tqe_next; 1354 } 1355 pmap_unmap_ptes(pmap_kernel()); 1356 pmap_update(pmap_kernel()); 1357 1358 #ifdef DIAGNOSTIC 1359 if (m) 1360 panic("pmap_alloc_l1pt: pglist not empty\n"); 1361 #endif /* DIAGNOSTIC */ 1362 1363 pt->pt_flags = 0; 1364 return(pt); 1365 } 1366 1367 /* 1368 * Free a L1 page table previously allocated with pmap_alloc_l1pt(). 1369 */ 1370 static void 1371 pmap_free_l1pt(pt) 1372 struct l1pt *pt; 1373 { 1374 /* Separate the physical memory for the virtual space */ 1375 pmap_kremove(pt->pt_va, PD_SIZE); 1376 pmap_update(pmap_kernel()); 1377 1378 /* Return the physical memory */ 1379 uvm_pglistfree(&pt->pt_plist); 1380 1381 /* Free the virtual space */ 1382 uvm_km_free(kernel_map, pt->pt_va, PD_SIZE); 1383 1384 /* Free the l1pt structure */ 1385 free(pt, M_VMPMAP); 1386 } 1387 1388 /* 1389 * Allocate a page directory. 1390 * This routine will either allocate a new page directory from the pool 1391 * of L1 page tables currently held by the kernel or it will allocate 1392 * a new one via pmap_alloc_l1pt(). 1393 * It will then initialise the l1 page table for use. 1394 * 1395 * XXX must tidy up and fix this code, not happy about how it does the pmaps_locking 1396 */ 1397 static int 1398 pmap_allocpagedir(pmap) 1399 struct pmap *pmap; 1400 { 1401 paddr_t pa; 1402 struct l1pt *pt; 1403 pt_entry_t *pte; 1404 1405 PDEBUG(0, printf("pmap_allocpagedir(%p)\n", pmap)); 1406 1407 /* Do we have any spare L1's lying around ? */ 1408 if (l1pt_static_queue_count) { 1409 --l1pt_static_queue_count; 1410 pt = l1pt_static_queue.sqh_first; 1411 SIMPLEQ_REMOVE_HEAD(&l1pt_static_queue, pt, pt_queue); 1412 } else if (l1pt_queue_count) { 1413 --l1pt_queue_count; 1414 pt = l1pt_queue.sqh_first; 1415 SIMPLEQ_REMOVE_HEAD(&l1pt_queue, pt, pt_queue); 1416 ++l1pt_reuse_count; 1417 } else { 1418 pt = pmap_alloc_l1pt(); 1419 if (!pt) 1420 return(ENOMEM); 1421 ++l1pt_create_count; 1422 } 1423 1424 /* Store the pointer to the l1 descriptor in the pmap. */ 1425 pmap->pm_l1pt = pt; 1426 1427 /* Get the physical address of the start of the l1 */ 1428 pa = VM_PAGE_TO_PHYS(TAILQ_FIRST(&pt->pt_plist)); 1429 1430 /* Store the virtual address of the l1 in the pmap. */ 1431 pmap->pm_pdir = (pd_entry_t *)pt->pt_va; 1432 1433 /* Clean the L1 if it is dirty */ 1434 if (!(pt->pt_flags & PTFLAG_CLEAN)) 1435 bzero((void *)pmap->pm_pdir, (PD_SIZE - KERNEL_PD_SIZE)); 1436 1437 /* Allocate a page table to map all the page tables for this pmap */ 1438 1439 #ifdef DIAGNOSTIC 1440 if (pmap->pm_vptpt) { 1441 /* XXX What if we have one already ? */ 1442 panic("pmap_allocpagedir: have pt already\n"); 1443 } 1444 #endif /* DIAGNOSTIC */ 1445 pmap->pm_vptpt = uvm_km_zalloc(kernel_map, NBPG); 1446 if (pmap->pm_vptpt == 0) { 1447 pmap_freepagedir(pmap); 1448 return(ENOMEM); 1449 } 1450 1451 /* need to lock this all up for growkernel */ 1452 simple_lock(&pmaps_lock); 1453 /* wish we didn't have to keep this locked... */ 1454 1455 /* Duplicate the kernel mappings. */ 1456 bcopy((char *)pmap_kernel()->pm_pdir + (PD_SIZE - KERNEL_PD_SIZE), 1457 (char *)pmap->pm_pdir + (PD_SIZE - KERNEL_PD_SIZE), 1458 KERNEL_PD_SIZE); 1459 1460 pte = vtopte(pmap->pm_vptpt); 1461 pmap->pm_pptpt = l2pte_pa(*pte); 1462 1463 /* Revoke cacheability and bufferability */ 1464 /* XXX should be done better than this */ 1465 *pte &= ~(PT_C | PT_B); 1466 1467 /* Wire in this page table */ 1468 pmap_map_in_l1(pmap, PTE_BASE, pmap->pm_pptpt, TRUE); 1469 1470 pt->pt_flags &= ~PTFLAG_CLEAN; /* L1 is dirty now */ 1471 1472 /* 1473 * Map the kernel page tables into the new PT map. 1474 */ 1475 bcopy((char *)(PTE_BASE 1476 + (PTE_BASE >> (PGSHIFT - 2)) 1477 + ((PD_SIZE - KERNEL_PD_SIZE) >> 2)), 1478 (char *)pmap->pm_vptpt + ((PD_SIZE - KERNEL_PD_SIZE) >> 2), 1479 (KERNEL_PD_SIZE >> 2)); 1480 1481 LIST_INSERT_HEAD(&pmaps, pmap, pm_list); 1482 simple_unlock(&pmaps_lock); 1483 1484 return(0); 1485 } 1486 1487 1488 /* 1489 * Initialize a preallocated and zeroed pmap structure, 1490 * such as one in a vmspace structure. 1491 */ 1492 1493 void 1494 pmap_pinit(pmap) 1495 struct pmap *pmap; 1496 { 1497 int backoff = 6; 1498 int retry = 10; 1499 1500 PDEBUG(0, printf("pmap_pinit(%p)\n", pmap)); 1501 1502 /* Keep looping until we succeed in allocating a page directory */ 1503 while (pmap_allocpagedir(pmap) != 0) { 1504 /* 1505 * Ok we failed to allocate a suitable block of memory for an 1506 * L1 page table. This means that either: 1507 * 1. 16KB of virtual address space could not be allocated 1508 * 2. 16KB of physically contiguous memory on a 16KB boundary 1509 * could not be allocated. 1510 * 1511 * Since we cannot fail we will sleep for a while and try 1512 * again. 1513 * 1514 * Searching for a suitable L1 PT is expensive: 1515 * to avoid hogging the system when memory is really 1516 * scarce, use an exponential back-off so that 1517 * eventually we won't retry more than once every 8 1518 * seconds. This should allow other processes to run 1519 * to completion and free up resources. 1520 */ 1521 (void) ltsleep(&lbolt, PVM, "l1ptwait", (hz << 3) >> backoff, 1522 NULL); 1523 if (--retry == 0) { 1524 retry = 10; 1525 if (backoff) 1526 --backoff; 1527 } 1528 } 1529 1530 /* Map zero page for the pmap. This will also map the L2 for it */ 1531 pmap_enter(pmap, 0x00000000, systempage.pv_pa, 1532 VM_PROT_READ, VM_PROT_READ | PMAP_WIRED); 1533 pmap_update(pmap); 1534 } 1535 1536 1537 void 1538 pmap_freepagedir(pmap) 1539 struct pmap *pmap; 1540 { 1541 /* Free the memory used for the page table mapping */ 1542 if (pmap->pm_vptpt != 0) 1543 uvm_km_free(kernel_map, (vaddr_t)pmap->pm_vptpt, NBPG); 1544 1545 /* junk the L1 page table */ 1546 if (pmap->pm_l1pt->pt_flags & PTFLAG_STATIC) { 1547 /* Add the page table to the queue */ 1548 SIMPLEQ_INSERT_TAIL(&l1pt_static_queue, pmap->pm_l1pt, pt_queue); 1549 ++l1pt_static_queue_count; 1550 } else if (l1pt_queue_count < 8) { 1551 /* Add the page table to the queue */ 1552 SIMPLEQ_INSERT_TAIL(&l1pt_queue, pmap->pm_l1pt, pt_queue); 1553 ++l1pt_queue_count; 1554 } else 1555 pmap_free_l1pt(pmap->pm_l1pt); 1556 } 1557 1558 1559 /* 1560 * Retire the given physical map from service. 1561 * Should only be called if the map contains no valid mappings. 1562 */ 1563 1564 void 1565 pmap_destroy(pmap) 1566 struct pmap *pmap; 1567 { 1568 struct vm_page *page; 1569 int count; 1570 1571 if (pmap == NULL) 1572 return; 1573 1574 PDEBUG(0, printf("pmap_destroy(%p)\n", pmap)); 1575 1576 /* 1577 * Drop reference count 1578 */ 1579 simple_lock(&pmap->pm_obj.vmobjlock); 1580 count = --pmap->pm_obj.uo_refs; 1581 simple_unlock(&pmap->pm_obj.vmobjlock); 1582 if (count > 0) { 1583 return; 1584 } 1585 1586 /* 1587 * reference count is zero, free pmap resources and then free pmap. 1588 */ 1589 1590 /* 1591 * remove it from global list of pmaps 1592 */ 1593 1594 simple_lock(&pmaps_lock); 1595 LIST_REMOVE(pmap, pm_list); 1596 simple_unlock(&pmaps_lock); 1597 1598 /* Remove the zero page mapping */ 1599 pmap_remove(pmap, 0x00000000, 0x00000000 + NBPG); 1600 pmap_update(pmap); 1601 1602 /* 1603 * Free any page tables still mapped 1604 * This is only temporay until pmap_enter can count the number 1605 * of mappings made in a page table. Then pmap_remove() can 1606 * reduce the count and free the pagetable when the count 1607 * reaches zero. Note that entries in this list should match the 1608 * contents of the ptpt, however this is faster than walking a 1024 1609 * entries looking for pt's 1610 * taken from i386 pmap.c 1611 */ 1612 while ((page = TAILQ_FIRST(&pmap->pm_obj.memq)) != NULL) { 1613 KASSERT((page->flags & PG_BUSY) == 0); 1614 page->wire_count = 0; 1615 uvm_pagefree(page); 1616 } 1617 1618 /* Free the page dir */ 1619 pmap_freepagedir(pmap); 1620 1621 /* return the pmap to the pool */ 1622 pool_put(&pmap_pmap_pool, pmap); 1623 } 1624 1625 1626 /* 1627 * void pmap_reference(struct pmap *pmap) 1628 * 1629 * Add a reference to the specified pmap. 1630 */ 1631 1632 void 1633 pmap_reference(pmap) 1634 struct pmap *pmap; 1635 { 1636 if (pmap == NULL) 1637 return; 1638 1639 simple_lock(&pmap->pm_lock); 1640 pmap->pm_obj.uo_refs++; 1641 simple_unlock(&pmap->pm_lock); 1642 } 1643 1644 /* 1645 * void pmap_virtual_space(vaddr_t *start, vaddr_t *end) 1646 * 1647 * Return the start and end addresses of the kernel's virtual space. 1648 * These values are setup in pmap_bootstrap and are updated as pages 1649 * are allocated. 1650 */ 1651 1652 void 1653 pmap_virtual_space(start, end) 1654 vaddr_t *start; 1655 vaddr_t *end; 1656 { 1657 *start = virtual_avail; 1658 *end = virtual_end; 1659 } 1660 1661 1662 /* 1663 * Activate the address space for the specified process. If the process 1664 * is the current process, load the new MMU context. 1665 */ 1666 void 1667 pmap_activate(p) 1668 struct proc *p; 1669 { 1670 struct pmap *pmap = p->p_vmspace->vm_map.pmap; 1671 struct pcb *pcb = &p->p_addr->u_pcb; 1672 1673 (void) pmap_extract(pmap_kernel(), (vaddr_t)pmap->pm_pdir, 1674 (paddr_t *)&pcb->pcb_pagedir); 1675 1676 PDEBUG(0, printf("pmap_activate: p=%p pmap=%p pcb=%p pdir=%p l1=%p\n", 1677 p, pmap, pcb, pmap->pm_pdir, pcb->pcb_pagedir)); 1678 1679 if (p == curproc) { 1680 PDEBUG(0, printf("pmap_activate: setting TTB\n")); 1681 setttb((u_int)pcb->pcb_pagedir); 1682 } 1683 #if 0 1684 pmap->pm_pdchanged = FALSE; 1685 #endif 1686 } 1687 1688 1689 /* 1690 * Deactivate the address space of the specified process. 1691 */ 1692 void 1693 pmap_deactivate(p) 1694 struct proc *p; 1695 { 1696 } 1697 1698 /* 1699 * Perform any deferred pmap operations. 1700 */ 1701 void 1702 pmap_update(struct pmap *pmap) 1703 { 1704 1705 /* 1706 * We haven't deferred any pmap operations, but we do need to 1707 * make sure TLB/cache operations have completed. 1708 */ 1709 cpu_cpwait(); 1710 } 1711 1712 /* 1713 * pmap_clean_page() 1714 * 1715 * This is a local function used to work out the best strategy to clean 1716 * a single page referenced by its entry in the PV table. It's used by 1717 * pmap_copy_page, pmap_zero page and maybe some others later on. 1718 * 1719 * Its policy is effectively: 1720 * o If there are no mappings, we don't bother doing anything with the cache. 1721 * o If there is one mapping, we clean just that page. 1722 * o If there are multiple mappings, we clean the entire cache. 1723 * 1724 * So that some functions can be further optimised, it returns 0 if it didn't 1725 * clean the entire cache, or 1 if it did. 1726 * 1727 * XXX One bug in this routine is that if the pv_entry has a single page 1728 * mapped at 0x00000000 a whole cache clean will be performed rather than 1729 * just the 1 page. Since this should not occur in everyday use and if it does 1730 * it will just result in not the most efficient clean for the page. 1731 */ 1732 static int 1733 pmap_clean_page(pv, is_src) 1734 struct pv_entry *pv; 1735 boolean_t is_src; 1736 { 1737 struct pmap *pmap; 1738 struct pv_entry *npv; 1739 int cache_needs_cleaning = 0; 1740 vaddr_t page_to_clean = 0; 1741 1742 if (pv == NULL) 1743 /* nothing mapped in so nothing to flush */ 1744 return (0); 1745 1746 /* Since we flush the cache each time we change curproc, we 1747 * only need to flush the page if it is in the current pmap. 1748 */ 1749 if (curproc) 1750 pmap = curproc->p_vmspace->vm_map.pmap; 1751 else 1752 pmap = pmap_kernel(); 1753 1754 for (npv = pv; npv; npv = npv->pv_next) { 1755 if (npv->pv_pmap == pmap) { 1756 /* The page is mapped non-cacheable in 1757 * this map. No need to flush the cache. 1758 */ 1759 if (npv->pv_flags & PT_NC) { 1760 #ifdef DIAGNOSTIC 1761 if (cache_needs_cleaning) 1762 panic("pmap_clean_page: " 1763 "cache inconsistency"); 1764 #endif 1765 break; 1766 } 1767 #if 0 1768 /* This doesn't work, because pmap_protect 1769 doesn't flush changes on pages that it 1770 has write-protected. */ 1771 1772 /* If the page is not writable and this 1773 is the source, then there is no need 1774 to flush it from the cache. */ 1775 else if (is_src && ! (npv->pv_flags & PT_Wr)) 1776 continue; 1777 #endif 1778 if (cache_needs_cleaning){ 1779 page_to_clean = 0; 1780 break; 1781 } 1782 else 1783 page_to_clean = npv->pv_va; 1784 cache_needs_cleaning = 1; 1785 } 1786 } 1787 1788 if (page_to_clean) 1789 cpu_idcache_wbinv_range(page_to_clean, NBPG); 1790 else if (cache_needs_cleaning) { 1791 cpu_idcache_wbinv_all(); 1792 return (1); 1793 } 1794 return (0); 1795 } 1796 1797 /* 1798 * pmap_zero_page() 1799 * 1800 * Zero a given physical page by mapping it at a page hook point. 1801 * In doing the zero page op, the page we zero is mapped cachable, as with 1802 * StrongARM accesses to non-cached pages are non-burst making writing 1803 * _any_ bulk data very slow. 1804 */ 1805 void 1806 pmap_zero_page(phys) 1807 paddr_t phys; 1808 { 1809 struct vm_page *pg; 1810 1811 /* Get an entry for this page, and clean it it. */ 1812 pg = PHYS_TO_VM_PAGE(phys); 1813 simple_lock(&pg->mdpage.pvh_slock); 1814 pmap_clean_page(pg->mdpage.pvh_list, FALSE); 1815 simple_unlock(&pg->mdpage.pvh_slock); 1816 1817 /* 1818 * Hook in the page, zero it, and purge the cache for that 1819 * zeroed page. Invalidate the TLB as needed. 1820 */ 1821 *cdst_pte = L2_PTE(phys & PG_FRAME, AP_KRW); 1822 cpu_tlb_flushD_SE(cdstp); 1823 cpu_cpwait(); 1824 bzero_page(cdstp); 1825 cpu_dcache_wbinv_range(cdstp, NBPG); 1826 } 1827 1828 /* pmap_pageidlezero() 1829 * 1830 * The same as above, except that we assume that the page is not 1831 * mapped. This means we never have to flush the cache first. Called 1832 * from the idle loop. 1833 */ 1834 boolean_t 1835 pmap_pageidlezero(phys) 1836 paddr_t phys; 1837 { 1838 int i, *ptr; 1839 boolean_t rv = TRUE; 1840 1841 #ifdef DIAGNOSTIC 1842 struct vm_page *pg; 1843 1844 pg = PHYS_TO_VM_PAGE(phys); 1845 if (pg->mdpage.pvh_list != NULL) 1846 panic("pmap_pageidlezero: zeroing mapped page\n"); 1847 #endif 1848 1849 /* 1850 * Hook in the page, zero it, and purge the cache for that 1851 * zeroed page. Invalidate the TLB as needed. 1852 */ 1853 *cdst_pte = L2_PTE(phys & PG_FRAME, AP_KRW); 1854 cpu_tlb_flushD_SE(cdstp); 1855 cpu_cpwait(); 1856 1857 for (i = 0, ptr = (int *)cdstp; 1858 i < (NBPG / sizeof(int)); i++) { 1859 if (sched_whichqs != 0) { 1860 /* 1861 * A process has become ready. Abort now, 1862 * so we don't keep it waiting while we 1863 * do slow memory access to finish this 1864 * page. 1865 */ 1866 rv = FALSE; 1867 break; 1868 } 1869 *ptr++ = 0; 1870 } 1871 1872 if (rv) 1873 /* 1874 * if we aborted we'll rezero this page again later so don't 1875 * purge it unless we finished it 1876 */ 1877 cpu_dcache_wbinv_range(cdstp, NBPG); 1878 return (rv); 1879 } 1880 1881 /* 1882 * pmap_copy_page() 1883 * 1884 * Copy one physical page into another, by mapping the pages into 1885 * hook points. The same comment regarding cachability as in 1886 * pmap_zero_page also applies here. 1887 */ 1888 void 1889 pmap_copy_page(src, dest) 1890 paddr_t src; 1891 paddr_t dest; 1892 { 1893 struct vm_page *src_pg, *dest_pg; 1894 boolean_t cleanedcache; 1895 1896 /* Get PV entries for the pages, and clean them if needed. */ 1897 src_pg = PHYS_TO_VM_PAGE(src); 1898 1899 simple_lock(&src_pg->mdpage.pvh_slock); 1900 cleanedcache = pmap_clean_page(src_pg->mdpage.pvh_list, TRUE); 1901 simple_unlock(&src_pg->mdpage.pvh_slock); 1902 1903 if (cleanedcache == 0) { 1904 dest_pg = PHYS_TO_VM_PAGE(dest); 1905 simple_lock(&dest_pg->mdpage.pvh_slock); 1906 pmap_clean_page(dest_pg->mdpage.pvh_list, FALSE); 1907 simple_unlock(&dest_pg->mdpage.pvh_slock); 1908 } 1909 /* 1910 * Map the pages into the page hook points, copy them, and purge 1911 * the cache for the appropriate page. Invalidate the TLB 1912 * as required. 1913 */ 1914 *csrc_pte = L2_PTE(src & PG_FRAME, AP_KR); 1915 *cdst_pte = L2_PTE(dest & PG_FRAME, AP_KRW); 1916 cpu_tlb_flushD_SE(csrcp); 1917 cpu_tlb_flushD_SE(cdstp); 1918 cpu_cpwait(); 1919 bcopy_page(csrcp, cdstp); 1920 cpu_dcache_inv_range(csrcp, NBPG); 1921 cpu_dcache_wbinv_range(cdstp, NBPG); 1922 } 1923 1924 #if 0 1925 void 1926 pmap_pte_addref(pmap, va) 1927 struct pmap *pmap; 1928 vaddr_t va; 1929 { 1930 pd_entry_t *pde; 1931 paddr_t pa; 1932 struct vm_page *m; 1933 1934 if (pmap == pmap_kernel()) 1935 return; 1936 1937 pde = pmap_pde(pmap, va & ~(3 << PDSHIFT)); 1938 pa = pmap_pte_pa(pde); 1939 m = PHYS_TO_VM_PAGE(pa); 1940 ++m->wire_count; 1941 #ifdef MYCROFT_HACK 1942 printf("addref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n", 1943 pmap, va, pde, pa, m, m->wire_count); 1944 #endif 1945 } 1946 1947 void 1948 pmap_pte_delref(pmap, va) 1949 struct pmap *pmap; 1950 vaddr_t va; 1951 { 1952 pd_entry_t *pde; 1953 paddr_t pa; 1954 struct vm_page *m; 1955 1956 if (pmap == pmap_kernel()) 1957 return; 1958 1959 pde = pmap_pde(pmap, va & ~(3 << PDSHIFT)); 1960 pa = pmap_pte_pa(pde); 1961 m = PHYS_TO_VM_PAGE(pa); 1962 --m->wire_count; 1963 #ifdef MYCROFT_HACK 1964 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p wire=%d\n", 1965 pmap, va, pde, pa, m, m->wire_count); 1966 #endif 1967 if (m->wire_count == 0) { 1968 #ifdef MYCROFT_HACK 1969 printf("delref pmap=%p va=%08lx pde=%p pa=%08lx m=%p\n", 1970 pmap, va, pde, pa, m); 1971 #endif 1972 pmap_unmap_in_l1(pmap, va); 1973 uvm_pagefree(m); 1974 --pmap->pm_stats.resident_count; 1975 } 1976 } 1977 #else 1978 #define pmap_pte_addref(pmap, va) 1979 #define pmap_pte_delref(pmap, va) 1980 #endif 1981 1982 /* 1983 * Since we have a virtually indexed cache, we may need to inhibit caching if 1984 * there is more than one mapping and at least one of them is writable. 1985 * Since we purge the cache on every context switch, we only need to check for 1986 * other mappings within the same pmap, or kernel_pmap. 1987 * This function is also called when a page is unmapped, to possibly reenable 1988 * caching on any remaining mappings. 1989 * 1990 * The code implements the following logic, where: 1991 * 1992 * KW = # of kernel read/write pages 1993 * KR = # of kernel read only pages 1994 * UW = # of user read/write pages 1995 * UR = # of user read only pages 1996 * OW = # of user read/write pages in another pmap, then 1997 * 1998 * KC = kernel mapping is cacheable 1999 * UC = user mapping is cacheable 2000 * 2001 * KW=0,KR=0 KW=0,KR>0 KW=1,KR=0 KW>1,KR>=0 2002 * +--------------------------------------------- 2003 * UW=0,UR=0,OW=0 | --- KC=1 KC=1 KC=0 2004 * UW=0,UR>0,OW=0 | UC=1 KC=1,UC=1 KC=0,UC=0 KC=0,UC=0 2005 * UW=0,UR>0,OW>0 | UC=1 KC=0,UC=1 KC=0,UC=0 KC=0,UC=0 2006 * UW=1,UR=0,OW=0 | UC=1 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0 2007 * UW>1,UR>=0,OW>=0 | UC=0 KC=0,UC=0 KC=0,UC=0 KC=0,UC=0 2008 * 2009 * Note that the pmap must have it's ptes mapped in, and passed with ptes. 2010 */ 2011 __inline static void 2012 pmap_vac_me_harder(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes, 2013 boolean_t clear_cache) 2014 { 2015 if (pmap == pmap_kernel()) 2016 pmap_vac_me_kpmap(pmap, pg, ptes, clear_cache); 2017 else 2018 pmap_vac_me_user(pmap, pg, ptes, clear_cache); 2019 } 2020 2021 static void 2022 pmap_vac_me_kpmap(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes, 2023 boolean_t clear_cache) 2024 { 2025 int user_entries = 0; 2026 int user_writable = 0; 2027 int user_cacheable = 0; 2028 int kernel_entries = 0; 2029 int kernel_writable = 0; 2030 int kernel_cacheable = 0; 2031 struct pv_entry *pv; 2032 struct pmap *last_pmap = pmap; 2033 2034 #ifdef DIAGNOSTIC 2035 if (pmap != pmap_kernel()) 2036 panic("pmap_vac_me_kpmap: pmap != pmap_kernel()"); 2037 #endif 2038 2039 /* 2040 * Pass one, see if there are both kernel and user pmaps for 2041 * this page. Calculate whether there are user-writable or 2042 * kernel-writable pages. 2043 */ 2044 for (pv = pg->mdpage.pvh_list; pv != NULL; pv = pv->pv_next) { 2045 if (pv->pv_pmap != pmap) { 2046 user_entries++; 2047 if (pv->pv_flags & PT_Wr) 2048 user_writable++; 2049 if ((pv->pv_flags & PT_NC) == 0) 2050 user_cacheable++; 2051 } else { 2052 kernel_entries++; 2053 if (pv->pv_flags & PT_Wr) 2054 kernel_writable++; 2055 if ((pv->pv_flags & PT_NC) == 0) 2056 kernel_cacheable++; 2057 } 2058 } 2059 2060 /* 2061 * We know we have just been updating a kernel entry, so if 2062 * all user pages are already cacheable, then there is nothing 2063 * further to do. 2064 */ 2065 if (kernel_entries == 0 && 2066 user_cacheable == user_entries) 2067 return; 2068 2069 if (user_entries) { 2070 /* 2071 * Scan over the list again, for each entry, if it 2072 * might not be set correctly, call pmap_vac_me_user 2073 * to recalculate the settings. 2074 */ 2075 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) { 2076 /* 2077 * We know kernel mappings will get set 2078 * correctly in other calls. We also know 2079 * that if the pmap is the same as last_pmap 2080 * then we've just handled this entry. 2081 */ 2082 if (pv->pv_pmap == pmap || pv->pv_pmap == last_pmap) 2083 continue; 2084 /* 2085 * If there are kernel entries and this page 2086 * is writable but non-cacheable, then we can 2087 * skip this entry also. 2088 */ 2089 if (kernel_entries > 0 && 2090 (pv->pv_flags & (PT_NC | PT_Wr)) == 2091 (PT_NC | PT_Wr)) 2092 continue; 2093 /* 2094 * Similarly if there are no kernel-writable 2095 * entries and the page is already 2096 * read-only/cacheable. 2097 */ 2098 if (kernel_writable == 0 && 2099 (pv->pv_flags & (PT_NC | PT_Wr)) == 0) 2100 continue; 2101 /* 2102 * For some of the remaining cases, we know 2103 * that we must recalculate, but for others we 2104 * can't tell if they are correct or not, so 2105 * we recalculate anyway. 2106 */ 2107 pmap_unmap_ptes(last_pmap); 2108 last_pmap = pv->pv_pmap; 2109 ptes = pmap_map_ptes(last_pmap); 2110 pmap_vac_me_user(last_pmap, pg, ptes, 2111 pmap_is_curpmap(last_pmap)); 2112 } 2113 /* Restore the pte mapping that was passed to us. */ 2114 if (last_pmap != pmap) { 2115 pmap_unmap_ptes(last_pmap); 2116 ptes = pmap_map_ptes(pmap); 2117 } 2118 if (kernel_entries == 0) 2119 return; 2120 } 2121 2122 pmap_vac_me_user(pmap, pg, ptes, clear_cache); 2123 return; 2124 } 2125 2126 static void 2127 pmap_vac_me_user(struct pmap *pmap, struct vm_page *pg, pt_entry_t *ptes, 2128 boolean_t clear_cache) 2129 { 2130 struct pmap *kpmap = pmap_kernel(); 2131 struct pv_entry *pv, *npv; 2132 int entries = 0; 2133 int writable = 0; 2134 int cacheable_entries = 0; 2135 int kern_cacheable = 0; 2136 int other_writable = 0; 2137 2138 pv = pg->mdpage.pvh_list; 2139 KASSERT(ptes != NULL); 2140 2141 /* 2142 * Count mappings and writable mappings in this pmap. 2143 * Include kernel mappings as part of our own. 2144 * Keep a pointer to the first one. 2145 */ 2146 for (npv = pv; npv; npv = npv->pv_next) { 2147 /* Count mappings in the same pmap */ 2148 if (pmap == npv->pv_pmap || 2149 kpmap == npv->pv_pmap) { 2150 if (entries++ == 0) 2151 pv = npv; 2152 /* Cacheable mappings */ 2153 if ((npv->pv_flags & PT_NC) == 0) { 2154 cacheable_entries++; 2155 if (kpmap == npv->pv_pmap) 2156 kern_cacheable++; 2157 } 2158 /* Writable mappings */ 2159 if (npv->pv_flags & PT_Wr) 2160 ++writable; 2161 } else if (npv->pv_flags & PT_Wr) 2162 other_writable = 1; 2163 } 2164 2165 PDEBUG(3,printf("pmap_vac_me_harder: pmap %p Entries %d, " 2166 "writable %d cacheable %d %s\n", pmap, entries, writable, 2167 cacheable_entries, clear_cache ? "clean" : "no clean")); 2168 2169 /* 2170 * Enable or disable caching as necessary. 2171 * Note: the first entry might be part of the kernel pmap, 2172 * so we can't assume this is indicative of the state of the 2173 * other (maybe non-kpmap) entries. 2174 */ 2175 if ((entries > 1 && writable) || 2176 (entries > 0 && pmap == kpmap && other_writable)) { 2177 if (cacheable_entries == 0) 2178 return; 2179 for (npv = pv; npv; npv = npv->pv_next) { 2180 if ((pmap == npv->pv_pmap 2181 || kpmap == npv->pv_pmap) && 2182 (npv->pv_flags & PT_NC) == 0) { 2183 ptes[arm_btop(npv->pv_va)] &= ~(PT_C | PT_B); 2184 npv->pv_flags |= PT_NC; 2185 /* 2186 * If this page needs flushing from the 2187 * cache, and we aren't going to do it 2188 * below, do it now. 2189 */ 2190 if ((cacheable_entries < 4 && 2191 (clear_cache || npv->pv_pmap == kpmap)) || 2192 (npv->pv_pmap == kpmap && 2193 !clear_cache && kern_cacheable < 4)) { 2194 cpu_idcache_wbinv_range(npv->pv_va, 2195 NBPG); 2196 cpu_tlb_flushID_SE(npv->pv_va); 2197 } 2198 } 2199 } 2200 if ((clear_cache && cacheable_entries >= 4) || 2201 kern_cacheable >= 4) { 2202 cpu_idcache_wbinv_all(); 2203 cpu_tlb_flushID(); 2204 } 2205 cpu_cpwait(); 2206 } else if (entries > 0) { 2207 /* 2208 * Turn cacheing back on for some pages. If it is a kernel 2209 * page, only do so if there are no other writable pages. 2210 */ 2211 for (npv = pv; npv; npv = npv->pv_next) { 2212 if ((pmap == npv->pv_pmap || 2213 (kpmap == npv->pv_pmap && other_writable == 0)) && 2214 (npv->pv_flags & PT_NC)) { 2215 ptes[arm_btop(npv->pv_va)] |= pte_cache_mode; 2216 npv->pv_flags &= ~PT_NC; 2217 } 2218 } 2219 } 2220 } 2221 2222 /* 2223 * pmap_remove() 2224 * 2225 * pmap_remove is responsible for nuking a number of mappings for a range 2226 * of virtual address space in the current pmap. To do this efficiently 2227 * is interesting, because in a number of cases a wide virtual address 2228 * range may be supplied that contains few actual mappings. So, the 2229 * optimisations are: 2230 * 1. Try and skip over hunks of address space for which an L1 entry 2231 * does not exist. 2232 * 2. Build up a list of pages we've hit, up to a maximum, so we can 2233 * maybe do just a partial cache clean. This path of execution is 2234 * complicated by the fact that the cache must be flushed _before_ 2235 * the PTE is nuked, being a VAC :-) 2236 * 3. Maybe later fast-case a single page, but I don't think this is 2237 * going to make _that_ much difference overall. 2238 */ 2239 2240 #define PMAP_REMOVE_CLEAN_LIST_SIZE 3 2241 2242 void 2243 pmap_remove(pmap, sva, eva) 2244 struct pmap *pmap; 2245 vaddr_t sva; 2246 vaddr_t eva; 2247 { 2248 int cleanlist_idx = 0; 2249 struct pagelist { 2250 vaddr_t va; 2251 pt_entry_t *pte; 2252 } cleanlist[PMAP_REMOVE_CLEAN_LIST_SIZE]; 2253 pt_entry_t *pte = 0, *ptes; 2254 paddr_t pa; 2255 int pmap_active; 2256 struct vm_page *pg; 2257 2258 /* Exit quick if there is no pmap */ 2259 if (!pmap) 2260 return; 2261 2262 PDEBUG(0, printf("pmap_remove: pmap=%p sva=%08lx eva=%08lx\n", pmap, sva, eva)); 2263 2264 sva &= PG_FRAME; 2265 eva &= PG_FRAME; 2266 2267 /* 2268 * we lock in the pmap => vm_page direction 2269 */ 2270 PMAP_MAP_TO_HEAD_LOCK(); 2271 2272 ptes = pmap_map_ptes(pmap); 2273 /* Get a page table pointer */ 2274 while (sva < eva) { 2275 if (pmap_pde_page(pmap_pde(pmap, sva))) 2276 break; 2277 sva = (sva & PD_MASK) + NBPD; 2278 } 2279 2280 pte = &ptes[arm_btop(sva)]; 2281 /* Note if the pmap is active thus require cache and tlb cleans */ 2282 pmap_active = pmap_is_curpmap(pmap); 2283 2284 /* Now loop along */ 2285 while (sva < eva) { 2286 /* Check if we can move to the next PDE (l1 chunk) */ 2287 if (!(sva & PT_MASK)) 2288 if (!pmap_pde_page(pmap_pde(pmap, sva))) { 2289 sva += NBPD; 2290 pte += arm_btop(NBPD); 2291 continue; 2292 } 2293 2294 /* We've found a valid PTE, so this page of PTEs has to go. */ 2295 if (pmap_pte_v(pte)) { 2296 /* Update statistics */ 2297 --pmap->pm_stats.resident_count; 2298 2299 /* 2300 * Add this page to our cache remove list, if we can. 2301 * If, however the cache remove list is totally full, 2302 * then do a complete cache invalidation taking note 2303 * to backtrack the PTE table beforehand, and ignore 2304 * the lists in future because there's no longer any 2305 * point in bothering with them (we've paid the 2306 * penalty, so will carry on unhindered). Otherwise, 2307 * when we fall out, we just clean the list. 2308 */ 2309 PDEBUG(10, printf("remove: inv pte at %p(%x) ", pte, *pte)); 2310 pa = pmap_pte_pa(pte); 2311 2312 if (cleanlist_idx < PMAP_REMOVE_CLEAN_LIST_SIZE) { 2313 /* Add to the clean list. */ 2314 cleanlist[cleanlist_idx].pte = pte; 2315 cleanlist[cleanlist_idx].va = sva; 2316 cleanlist_idx++; 2317 } else if (cleanlist_idx == PMAP_REMOVE_CLEAN_LIST_SIZE) { 2318 int cnt; 2319 2320 /* Nuke everything if needed. */ 2321 if (pmap_active) { 2322 cpu_idcache_wbinv_all(); 2323 cpu_tlb_flushID(); 2324 } 2325 2326 /* 2327 * Roll back the previous PTE list, 2328 * and zero out the current PTE. 2329 */ 2330 for (cnt = 0; cnt < PMAP_REMOVE_CLEAN_LIST_SIZE; cnt++) { 2331 *cleanlist[cnt].pte = 0; 2332 pmap_pte_delref(pmap, cleanlist[cnt].va); 2333 } 2334 *pte = 0; 2335 pmap_pte_delref(pmap, sva); 2336 cleanlist_idx++; 2337 } else { 2338 /* 2339 * We've already nuked the cache and 2340 * TLB, so just carry on regardless, 2341 * and we won't need to do it again 2342 */ 2343 *pte = 0; 2344 pmap_pte_delref(pmap, sva); 2345 } 2346 2347 /* 2348 * Update flags. In a number of circumstances, 2349 * we could cluster a lot of these and do a 2350 * number of sequential pages in one go. 2351 */ 2352 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) { 2353 struct pv_entry *pve; 2354 simple_lock(&pg->mdpage.pvh_slock); 2355 pve = pmap_remove_pv(pg, pmap, sva); 2356 pmap_free_pv(pmap, pve); 2357 pmap_vac_me_harder(pmap, pg, ptes, FALSE); 2358 simple_unlock(&pg->mdpage.pvh_slock); 2359 } 2360 } 2361 sva += NBPG; 2362 pte++; 2363 } 2364 2365 pmap_unmap_ptes(pmap); 2366 /* 2367 * Now, if we've fallen through down to here, chances are that there 2368 * are less than PMAP_REMOVE_CLEAN_LIST_SIZE mappings left. 2369 */ 2370 if (cleanlist_idx <= PMAP_REMOVE_CLEAN_LIST_SIZE) { 2371 u_int cnt; 2372 2373 for (cnt = 0; cnt < cleanlist_idx; cnt++) { 2374 if (pmap_active) { 2375 cpu_idcache_wbinv_range(cleanlist[cnt].va, 2376 NBPG); 2377 *cleanlist[cnt].pte = 0; 2378 cpu_tlb_flushID_SE(cleanlist[cnt].va); 2379 } else 2380 *cleanlist[cnt].pte = 0; 2381 pmap_pte_delref(pmap, cleanlist[cnt].va); 2382 } 2383 } 2384 PMAP_MAP_TO_HEAD_UNLOCK(); 2385 } 2386 2387 /* 2388 * Routine: pmap_remove_all 2389 * Function: 2390 * Removes this physical page from 2391 * all physical maps in which it resides. 2392 * Reflects back modify bits to the pager. 2393 */ 2394 2395 static void 2396 pmap_remove_all(pg) 2397 struct vm_page *pg; 2398 { 2399 struct pv_entry *pv, *npv; 2400 struct pmap *pmap; 2401 pt_entry_t *pte, *ptes; 2402 2403 PDEBUG(0, printf("pmap_remove_all: pa=%lx ", VM_PAGE_TO_PHYS(pg))); 2404 2405 /* set vm_page => pmap locking */ 2406 PMAP_HEAD_TO_MAP_LOCK(); 2407 2408 simple_lock(&pg->mdpage.pvh_slock); 2409 2410 pv = pg->mdpage.pvh_list; 2411 if (pv == NULL) { 2412 PDEBUG(0, printf("free page\n")); 2413 simple_unlock(&pg->mdpage.pvh_slock); 2414 PMAP_HEAD_TO_MAP_UNLOCK(); 2415 return; 2416 } 2417 pmap_clean_page(pv, FALSE); 2418 2419 while (pv) { 2420 pmap = pv->pv_pmap; 2421 ptes = pmap_map_ptes(pmap); 2422 pte = &ptes[arm_btop(pv->pv_va)]; 2423 2424 PDEBUG(0, printf("[%p,%08x,%08lx,%08x] ", pmap, *pte, 2425 pv->pv_va, pv->pv_flags)); 2426 #ifdef DEBUG 2427 if (!pmap_pde_page(pmap_pde(pmap, pv->pv_va)) || 2428 !pmap_pte_v(pte) || pmap_pte_pa(pte) != pa) 2429 panic("pmap_remove_all: bad mapping"); 2430 #endif /* DEBUG */ 2431 2432 /* 2433 * Update statistics 2434 */ 2435 --pmap->pm_stats.resident_count; 2436 2437 /* Wired bit */ 2438 if (pv->pv_flags & PT_W) 2439 --pmap->pm_stats.wired_count; 2440 2441 /* 2442 * Invalidate the PTEs. 2443 * XXX: should cluster them up and invalidate as many 2444 * as possible at once. 2445 */ 2446 2447 #ifdef needednotdone 2448 reduce wiring count on page table pages as references drop 2449 #endif 2450 2451 *pte = 0; 2452 pmap_pte_delref(pmap, pv->pv_va); 2453 2454 npv = pv->pv_next; 2455 pmap_free_pv(pmap, pv); 2456 pv = npv; 2457 pmap_unmap_ptes(pmap); 2458 } 2459 pg->mdpage.pvh_list = NULL; 2460 simple_unlock(&pg->mdpage.pvh_slock); 2461 PMAP_HEAD_TO_MAP_UNLOCK(); 2462 2463 PDEBUG(0, printf("done\n")); 2464 cpu_tlb_flushID(); 2465 cpu_cpwait(); 2466 } 2467 2468 2469 /* 2470 * Set the physical protection on the specified range of this map as requested. 2471 */ 2472 2473 void 2474 pmap_protect(pmap, sva, eva, prot) 2475 struct pmap *pmap; 2476 vaddr_t sva; 2477 vaddr_t eva; 2478 vm_prot_t prot; 2479 { 2480 pt_entry_t *pte = NULL, *ptes; 2481 struct vm_page *pg; 2482 int armprot; 2483 int flush = 0; 2484 paddr_t pa; 2485 2486 PDEBUG(0, printf("pmap_protect: pmap=%p %08lx->%08lx %x\n", 2487 pmap, sva, eva, prot)); 2488 2489 if (~prot & VM_PROT_READ) { 2490 /* Just remove the mappings. */ 2491 pmap_remove(pmap, sva, eva); 2492 /* pmap_update not needed as it should be called by the caller 2493 * of pmap_protect */ 2494 return; 2495 } 2496 if (prot & VM_PROT_WRITE) { 2497 /* 2498 * If this is a read->write transition, just ignore it and let 2499 * uvm_fault() take care of it later. 2500 */ 2501 return; 2502 } 2503 2504 sva &= PG_FRAME; 2505 eva &= PG_FRAME; 2506 2507 /* Need to lock map->head */ 2508 PMAP_MAP_TO_HEAD_LOCK(); 2509 2510 ptes = pmap_map_ptes(pmap); 2511 /* 2512 * We need to acquire a pointer to a page table page before entering 2513 * the following loop. 2514 */ 2515 while (sva < eva) { 2516 if (pmap_pde_page(pmap_pde(pmap, sva))) 2517 break; 2518 sva = (sva & PD_MASK) + NBPD; 2519 } 2520 2521 pte = &ptes[arm_btop(sva)]; 2522 2523 while (sva < eva) { 2524 /* only check once in a while */ 2525 if ((sva & PT_MASK) == 0) { 2526 if (!pmap_pde_page(pmap_pde(pmap, sva))) { 2527 /* We can race ahead here, to the next pde. */ 2528 sva += NBPD; 2529 pte += arm_btop(NBPD); 2530 continue; 2531 } 2532 } 2533 2534 if (!pmap_pte_v(pte)) 2535 goto next; 2536 2537 flush = 1; 2538 2539 armprot = 0; 2540 if (sva < VM_MAXUSER_ADDRESS) 2541 armprot |= PT_AP(AP_U); 2542 else if (sva < VM_MAX_ADDRESS) 2543 armprot |= PT_AP(AP_W); /* XXX Ekk what is this ? */ 2544 *pte = (*pte & 0xfffff00f) | armprot; 2545 2546 pa = pmap_pte_pa(pte); 2547 2548 /* Get the physical page index */ 2549 2550 /* Clear write flag */ 2551 if ((pg = PHYS_TO_VM_PAGE(pa)) != NULL) { 2552 simple_lock(&pg->mdpage.pvh_slock); 2553 (void) pmap_modify_pv(pmap, sva, pg, PT_Wr, 0); 2554 pmap_vac_me_harder(pmap, pg, ptes, FALSE); 2555 simple_unlock(&pg->mdpage.pvh_slock); 2556 } 2557 2558 next: 2559 sva += NBPG; 2560 pte++; 2561 } 2562 pmap_unmap_ptes(pmap); 2563 PMAP_MAP_TO_HEAD_UNLOCK(); 2564 if (flush) 2565 cpu_tlb_flushID(); 2566 } 2567 2568 /* 2569 * void pmap_enter(struct pmap *pmap, vaddr_t va, paddr_t pa, vm_prot_t prot, 2570 * int flags) 2571 * 2572 * Insert the given physical page (p) at 2573 * the specified virtual address (v) in the 2574 * target physical map with the protection requested. 2575 * 2576 * If specified, the page will be wired down, meaning 2577 * that the related pte can not be reclaimed. 2578 * 2579 * NB: This is the only routine which MAY NOT lazy-evaluate 2580 * or lose information. That is, this routine must actually 2581 * insert this page into the given map NOW. 2582 */ 2583 2584 int 2585 pmap_enter(pmap, va, pa, prot, flags) 2586 struct pmap *pmap; 2587 vaddr_t va; 2588 paddr_t pa; 2589 vm_prot_t prot; 2590 int flags; 2591 { 2592 pt_entry_t *ptes, opte, npte; 2593 paddr_t opa; 2594 boolean_t wired = (flags & PMAP_WIRED) != 0; 2595 struct vm_page *pg; 2596 struct pv_entry *pve; 2597 int error, nflags; 2598 2599 PDEBUG(5, printf("pmap_enter: V%08lx P%08lx in pmap %p prot=%08x, wired = %d\n", 2600 va, pa, pmap, prot, wired)); 2601 2602 #ifdef DIAGNOSTIC 2603 /* Valid address ? */ 2604 if (va >= (pmap_curmaxkvaddr)) 2605 panic("pmap_enter: too big"); 2606 if (pmap != pmap_kernel() && va != 0) { 2607 if (va < VM_MIN_ADDRESS || va >= VM_MAXUSER_ADDRESS) 2608 panic("pmap_enter: kernel page in user map"); 2609 } else { 2610 if (va >= VM_MIN_ADDRESS && va < VM_MAXUSER_ADDRESS) 2611 panic("pmap_enter: user page in kernel map"); 2612 if (va >= VM_MAXUSER_ADDRESS && va < VM_MAX_ADDRESS) 2613 panic("pmap_enter: entering PT page"); 2614 } 2615 #endif 2616 /* 2617 * Get a pointer to the page. Later on in this function, we 2618 * test for a managed page by checking pg != NULL. 2619 */ 2620 pg = pmap_initialized ? PHYS_TO_VM_PAGE(pa) : NULL; 2621 2622 /* get lock */ 2623 PMAP_MAP_TO_HEAD_LOCK(); 2624 2625 /* 2626 * map the ptes. If there's not already an L2 table for this 2627 * address, allocate one. 2628 */ 2629 ptes = pmap_map_ptes(pmap); /* locks pmap */ 2630 if (pmap_pde_v(pmap_pde(pmap, va)) == 0) { 2631 struct vm_page *ptp; 2632 2633 /* kernel should be pre-grown */ 2634 KASSERT(pmap != pmap_kernel()); 2635 2636 /* if failure is allowed then don't try too hard */ 2637 ptp = pmap_get_ptp(pmap, va); 2638 if (ptp == NULL) { 2639 if (flags & PMAP_CANFAIL) { 2640 error = ENOMEM; 2641 goto out; 2642 } 2643 panic("pmap_enter: get ptp failed"); 2644 } 2645 } 2646 opte = ptes[arm_btop(va)]; 2647 2648 nflags = 0; 2649 if (prot & VM_PROT_WRITE) 2650 nflags |= PT_Wr; 2651 if (wired) 2652 nflags |= PT_W; 2653 2654 /* Is the pte valid ? If so then this page is already mapped */ 2655 if (l2pte_valid(opte)) { 2656 /* Get the physical address of the current page mapped */ 2657 opa = l2pte_pa(opte); 2658 2659 /* Are we mapping the same page ? */ 2660 if (opa == pa) { 2661 /* Has the wiring changed ? */ 2662 if (pg != NULL) { 2663 simple_lock(&pg->mdpage.pvh_slock); 2664 (void) pmap_modify_pv(pmap, va, pg, 2665 PT_Wr | PT_W, nflags); 2666 simple_unlock(&pg->mdpage.pvh_slock); 2667 } 2668 } else { 2669 struct vm_page *opg; 2670 2671 /* We are replacing the page with a new one. */ 2672 cpu_idcache_wbinv_range(va, NBPG); 2673 2674 /* 2675 * If it is part of our managed memory then we 2676 * must remove it from the PV list 2677 */ 2678 if ((opg = PHYS_TO_VM_PAGE(opa)) != NULL) { 2679 simple_lock(&opg->mdpage.pvh_slock); 2680 pve = pmap_remove_pv(opg, pmap, va); 2681 simple_unlock(&opg->mdpage.pvh_slock); 2682 } else { 2683 pve = NULL; 2684 } 2685 2686 goto enter; 2687 } 2688 } else { 2689 opa = 0; 2690 pve = NULL; 2691 pmap_pte_addref(pmap, va); 2692 2693 /* pte is not valid so we must be hooking in a new page */ 2694 ++pmap->pm_stats.resident_count; 2695 2696 enter: 2697 /* 2698 * Enter on the PV list if part of our managed memory 2699 */ 2700 if (pg != NULL) { 2701 if (pve == NULL) { 2702 pve = pmap_alloc_pv(pmap, ALLOCPV_NEED); 2703 if (pve == NULL) { 2704 if (flags & PMAP_CANFAIL) { 2705 error = ENOMEM; 2706 goto out; 2707 } 2708 panic("pmap_enter: no pv entries " 2709 "available"); 2710 } 2711 } 2712 /* enter_pv locks pvh when adding */ 2713 pmap_enter_pv(pg, pve, pmap, va, NULL, nflags); 2714 } else { 2715 if (pve != NULL) 2716 pmap_free_pv(pmap, pve); 2717 } 2718 } 2719 2720 /* Construct the pte, giving the correct access. */ 2721 npte = (pa & PG_FRAME); 2722 2723 /* VA 0 is magic. */ 2724 if (pmap != pmap_kernel() && va != 0) 2725 npte |= PT_AP(AP_U); 2726 2727 if (pg != NULL) { 2728 #ifdef DIAGNOSTIC 2729 if ((flags & VM_PROT_ALL) & ~prot) 2730 panic("pmap_enter: access_type exceeds prot"); 2731 #endif 2732 npte |= pte_cache_mode; 2733 if (flags & VM_PROT_WRITE) { 2734 npte |= L2_SPAGE | PT_AP(AP_W); 2735 pg->mdpage.pvh_attrs |= PT_H | PT_M; 2736 } else if (flags & VM_PROT_ALL) { 2737 npte |= L2_SPAGE; 2738 pg->mdpage.pvh_attrs |= PT_H; 2739 } else 2740 npte |= L2_INVAL; 2741 } else { 2742 if (prot & VM_PROT_WRITE) 2743 npte |= L2_SPAGE | PT_AP(AP_W); 2744 else if (prot & VM_PROT_ALL) 2745 npte |= L2_SPAGE; 2746 else 2747 npte |= L2_INVAL; 2748 } 2749 2750 ptes[arm_btop(va)] = npte; 2751 2752 if (pg != NULL) { 2753 simple_lock(&pg->mdpage.pvh_slock); 2754 pmap_vac_me_harder(pmap, pg, ptes, pmap_is_curpmap(pmap)); 2755 simple_unlock(&pg->mdpage.pvh_slock); 2756 } 2757 2758 /* Better flush the TLB ... */ 2759 cpu_tlb_flushID_SE(va); 2760 error = 0; 2761 out: 2762 pmap_unmap_ptes(pmap); /* unlocks pmap */ 2763 PMAP_MAP_TO_HEAD_UNLOCK(); 2764 2765 return error; 2766 } 2767 2768 /* 2769 * pmap_kenter_pa: enter a kernel mapping 2770 * 2771 * => no need to lock anything assume va is already allocated 2772 * => should be faster than normal pmap enter function 2773 */ 2774 void 2775 pmap_kenter_pa(va, pa, prot) 2776 vaddr_t va; 2777 paddr_t pa; 2778 vm_prot_t prot; 2779 { 2780 pt_entry_t *pte; 2781 2782 pte = vtopte(va); 2783 KASSERT(!pmap_pte_v(pte)); 2784 *pte = L2_PTE(pa, AP_KRW); 2785 } 2786 2787 void 2788 pmap_kremove(va, len) 2789 vaddr_t va; 2790 vsize_t len; 2791 { 2792 pt_entry_t *pte; 2793 2794 for (len >>= PAGE_SHIFT; len > 0; len--, va += PAGE_SIZE) { 2795 2796 /* 2797 * We assume that we will only be called with small 2798 * regions of memory. 2799 */ 2800 2801 KASSERT(pmap_pde_page(pmap_pde(pmap_kernel(), va))); 2802 pte = vtopte(va); 2803 cpu_idcache_wbinv_range(va, PAGE_SIZE); 2804 *pte = 0; 2805 cpu_tlb_flushID_SE(va); 2806 } 2807 } 2808 2809 /* 2810 * pmap_page_protect: 2811 * 2812 * Lower the permission for all mappings to a given page. 2813 */ 2814 2815 void 2816 pmap_page_protect(pg, prot) 2817 struct vm_page *pg; 2818 vm_prot_t prot; 2819 { 2820 2821 PDEBUG(0, printf("pmap_page_protect(pa=%lx, prot=%d)\n", 2822 VM_PAGE_TO_PHYS(pg), prot)); 2823 2824 switch(prot) { 2825 case VM_PROT_READ|VM_PROT_WRITE|VM_PROT_EXECUTE: 2826 case VM_PROT_READ|VM_PROT_WRITE: 2827 return; 2828 2829 case VM_PROT_READ: 2830 case VM_PROT_READ|VM_PROT_EXECUTE: 2831 pmap_copy_on_write(pg); 2832 break; 2833 2834 default: 2835 pmap_remove_all(pg); 2836 break; 2837 } 2838 } 2839 2840 2841 /* 2842 * Routine: pmap_unwire 2843 * Function: Clear the wired attribute for a map/virtual-address 2844 * pair. 2845 * In/out conditions: 2846 * The mapping must already exist in the pmap. 2847 */ 2848 2849 void 2850 pmap_unwire(pmap, va) 2851 struct pmap *pmap; 2852 vaddr_t va; 2853 { 2854 pt_entry_t *ptes; 2855 struct vm_page *pg; 2856 paddr_t pa; 2857 2858 PMAP_MAP_TO_HEAD_LOCK(); 2859 ptes = pmap_map_ptes(pmap); /* locks pmap */ 2860 2861 if (pmap_pde_v(pmap_pde(pmap, va))) { 2862 #ifdef DIAGNOSTIC 2863 if (l2pte_valid(ptes[arm_btop(va)]) == 0) 2864 panic("pmap_unwire: invalid L2 PTE"); 2865 #endif 2866 /* Extract the physical address of the page */ 2867 pa = l2pte_pa(ptes[arm_btop(va)]); 2868 2869 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) 2870 goto out; 2871 2872 /* Update the wired bit in the pv entry for this page. */ 2873 simple_lock(&pg->mdpage.pvh_slock); 2874 (void) pmap_modify_pv(pmap, va, pg, PT_W, 0); 2875 simple_unlock(&pg->mdpage.pvh_slock); 2876 } 2877 #ifdef DIAGNOSTIC 2878 else { 2879 panic("pmap_unwire: invalid L1 PTE"); 2880 } 2881 #endif 2882 out: 2883 pmap_unmap_ptes(pmap); /* unlocks pmap */ 2884 PMAP_MAP_TO_HEAD_UNLOCK(); 2885 } 2886 2887 /* 2888 * Routine: pmap_extract 2889 * Function: 2890 * Extract the physical page address associated 2891 * with the given map/virtual_address pair. 2892 */ 2893 boolean_t 2894 pmap_extract(pmap, va, pap) 2895 struct pmap *pmap; 2896 vaddr_t va; 2897 paddr_t *pap; 2898 { 2899 pd_entry_t *pde; 2900 pt_entry_t *pte, *ptes; 2901 paddr_t pa; 2902 boolean_t rv = TRUE; 2903 2904 PDEBUG(5, printf("pmap_extract: pmap=%p, va=V%08lx\n", pmap, va)); 2905 2906 /* 2907 * Get the pte for this virtual address. 2908 */ 2909 pde = pmap_pde(pmap, va); 2910 ptes = pmap_map_ptes(pmap); 2911 pte = &ptes[arm_btop(va)]; 2912 2913 if (pmap_pde_section(pde)) { 2914 pa = (*pde & PD_MASK) | (va & (L1_SEC_SIZE - 1)); 2915 goto out; 2916 } else if (pmap_pde_page(pde) == 0 || pmap_pte_v(pte) == 0) { 2917 rv = FALSE; 2918 goto out; 2919 } 2920 2921 if ((*pte & L2_MASK) == L2_LPAGE) { 2922 /* Extract the physical address from the pte */ 2923 pa = *pte & ~(L2_LPAGE_SIZE - 1); 2924 2925 PDEBUG(5, printf("pmap_extract: LPAGE pa = P%08lx\n", 2926 (pa | (va & (L2_LPAGE_SIZE - 1))))); 2927 2928 if (pap != NULL) 2929 *pap = pa | (va & (L2_LPAGE_SIZE - 1)); 2930 goto out; 2931 } 2932 2933 /* Extract the physical address from the pte */ 2934 pa = pmap_pte_pa(pte); 2935 2936 PDEBUG(5, printf("pmap_extract: SPAGE pa = P%08lx\n", 2937 (pa | (va & ~PG_FRAME)))); 2938 2939 if (pap != NULL) 2940 *pap = pa | (va & ~PG_FRAME); 2941 out: 2942 pmap_unmap_ptes(pmap); 2943 return (rv); 2944 } 2945 2946 2947 /* 2948 * Copy the range specified by src_addr/len from the source map to the 2949 * range dst_addr/len in the destination map. 2950 * 2951 * This routine is only advisory and need not do anything. 2952 */ 2953 2954 void 2955 pmap_copy(dst_pmap, src_pmap, dst_addr, len, src_addr) 2956 struct pmap *dst_pmap; 2957 struct pmap *src_pmap; 2958 vaddr_t dst_addr; 2959 vsize_t len; 2960 vaddr_t src_addr; 2961 { 2962 PDEBUG(0, printf("pmap_copy(%p, %p, %lx, %lx, %lx)\n", 2963 dst_pmap, src_pmap, dst_addr, len, src_addr)); 2964 } 2965 2966 #if defined(PMAP_DEBUG) 2967 void 2968 pmap_dump_pvlist(phys, m) 2969 vaddr_t phys; 2970 char *m; 2971 { 2972 struct vm_page *pg; 2973 struct pv_entry *pv; 2974 2975 if ((pg = PHYS_TO_VM_PAGE(phys)) == NULL) { 2976 printf("INVALID PA\n"); 2977 return; 2978 } 2979 simple_lock(&pg->mdpage.pvh_slock); 2980 printf("%s %08lx:", m, phys); 2981 if (pg->mdpage.pvh_list == NULL) { 2982 printf(" no mappings\n"); 2983 return; 2984 } 2985 2986 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) 2987 printf(" pmap %p va %08lx flags %08x", pv->pv_pmap, 2988 pv->pv_va, pv->pv_flags); 2989 2990 printf("\n"); 2991 simple_unlock(&pg->mdpage.pvh_slock); 2992 } 2993 2994 #endif /* PMAP_DEBUG */ 2995 2996 static pt_entry_t * 2997 pmap_map_ptes(struct pmap *pmap) 2998 { 2999 struct proc *p; 3000 3001 /* the kernel's pmap is always accessible */ 3002 if (pmap == pmap_kernel()) { 3003 return (pt_entry_t *)PTE_BASE ; 3004 } 3005 3006 if (pmap_is_curpmap(pmap)) { 3007 simple_lock(&pmap->pm_obj.vmobjlock); 3008 return (pt_entry_t *)PTE_BASE; 3009 } 3010 3011 p = curproc; 3012 3013 if (p == NULL) 3014 p = &proc0; 3015 3016 /* need to lock both curpmap and pmap: use ordered locking */ 3017 if ((unsigned) pmap < (unsigned) curproc->p_vmspace->vm_map.pmap) { 3018 simple_lock(&pmap->pm_obj.vmobjlock); 3019 simple_lock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock); 3020 } else { 3021 simple_lock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock); 3022 simple_lock(&pmap->pm_obj.vmobjlock); 3023 } 3024 3025 pmap_map_in_l1(p->p_vmspace->vm_map.pmap, APTE_BASE, 3026 pmap->pm_pptpt, FALSE); 3027 cpu_tlb_flushD(); 3028 cpu_cpwait(); 3029 return (pt_entry_t *)APTE_BASE; 3030 } 3031 3032 /* 3033 * pmap_unmap_ptes: unlock the PTE mapping of "pmap" 3034 */ 3035 3036 static void 3037 pmap_unmap_ptes(pmap) 3038 struct pmap *pmap; 3039 { 3040 if (pmap == pmap_kernel()) { 3041 return; 3042 } 3043 if (pmap_is_curpmap(pmap)) { 3044 simple_unlock(&pmap->pm_obj.vmobjlock); 3045 } else { 3046 simple_unlock(&pmap->pm_obj.vmobjlock); 3047 simple_unlock(&curproc->p_vmspace->vm_map.pmap->pm_obj.vmobjlock); 3048 } 3049 } 3050 3051 /* 3052 * Modify pte bits for all ptes corresponding to the given physical address. 3053 * We use `maskbits' rather than `clearbits' because we're always passing 3054 * constants and the latter would require an extra inversion at run-time. 3055 */ 3056 3057 static void 3058 pmap_clearbit(pg, maskbits) 3059 struct vm_page *pg; 3060 unsigned int maskbits; 3061 { 3062 struct pv_entry *pv; 3063 pt_entry_t *ptes; 3064 vaddr_t va; 3065 int tlbentry; 3066 3067 PDEBUG(1, printf("pmap_clearbit: pa=%08lx mask=%08x\n", 3068 VM_PAGE_TO_PHYS(pg), maskbits)); 3069 3070 tlbentry = 0; 3071 3072 PMAP_HEAD_TO_MAP_LOCK(); 3073 simple_lock(&pg->mdpage.pvh_slock); 3074 3075 /* 3076 * Clear saved attributes (modify, reference) 3077 */ 3078 pg->mdpage.pvh_attrs &= ~maskbits; 3079 3080 if (pg->mdpage.pvh_list == NULL) { 3081 simple_unlock(&pg->mdpage.pvh_slock); 3082 PMAP_HEAD_TO_MAP_UNLOCK(); 3083 return; 3084 } 3085 3086 /* 3087 * Loop over all current mappings setting/clearing as appropos 3088 */ 3089 for (pv = pg->mdpage.pvh_list; pv; pv = pv->pv_next) { 3090 va = pv->pv_va; 3091 pv->pv_flags &= ~maskbits; 3092 ptes = pmap_map_ptes(pv->pv_pmap); /* locks pmap */ 3093 KASSERT(pmap_pde_v(pmap_pde(pv->pv_pmap, va))); 3094 if (maskbits & (PT_Wr|PT_M)) { 3095 if ((pv->pv_flags & PT_NC)) { 3096 /* 3097 * Entry is not cacheable: reenable 3098 * the cache, nothing to flush 3099 * 3100 * Don't turn caching on again if this 3101 * is a modified emulation. This 3102 * would be inconsitent with the 3103 * settings created by 3104 * pmap_vac_me_harder(). 3105 * 3106 * There's no need to call 3107 * pmap_vac_me_harder() here: all 3108 * pages are loosing their write 3109 * permission. 3110 * 3111 */ 3112 if (maskbits & PT_Wr) { 3113 ptes[arm_btop(va)] |= pte_cache_mode; 3114 pv->pv_flags &= ~PT_NC; 3115 } 3116 } else if (pmap_is_curpmap(pv->pv_pmap)) { 3117 /* 3118 * Entry is cacheable: check if pmap is 3119 * current if it is flush it, 3120 * otherwise it won't be in the cache 3121 */ 3122 cpu_idcache_wbinv_range(pv->pv_va, NBPG); 3123 } 3124 3125 /* make the pte read only */ 3126 ptes[arm_btop(va)] &= ~PT_AP(AP_W); 3127 } 3128 3129 if (maskbits & PT_H) 3130 ptes[arm_btop(va)] = 3131 (ptes[arm_btop(va)] & ~L2_MASK) | L2_INVAL; 3132 3133 if (pmap_is_curpmap(pv->pv_pmap)) { 3134 /* 3135 * if we had cacheable pte's we'd clean the 3136 * pte out to memory here 3137 * 3138 * flush tlb entry as it's in the current pmap 3139 */ 3140 cpu_tlb_flushID_SE(pv->pv_va); 3141 } 3142 pmap_unmap_ptes(pv->pv_pmap); /* unlocks pmap */ 3143 } 3144 cpu_cpwait(); 3145 3146 simple_unlock(&pg->mdpage.pvh_slock); 3147 PMAP_HEAD_TO_MAP_UNLOCK(); 3148 } 3149 3150 /* 3151 * pmap_clear_modify: 3152 * 3153 * Clear the "modified" attribute for a page. 3154 */ 3155 boolean_t 3156 pmap_clear_modify(pg) 3157 struct vm_page *pg; 3158 { 3159 boolean_t rv; 3160 3161 if (pg->mdpage.pvh_attrs & PT_M) { 3162 rv = TRUE; 3163 pmap_clearbit(pg, PT_M); 3164 } else 3165 rv = FALSE; 3166 3167 PDEBUG(0, printf("pmap_clear_modify pa=%08lx -> %d\n", 3168 VM_PAGE_TO_PHYS(pg), rv)); 3169 3170 return (rv); 3171 } 3172 3173 /* 3174 * pmap_clear_reference: 3175 * 3176 * Clear the "referenced" attribute for a page. 3177 */ 3178 boolean_t 3179 pmap_clear_reference(pg) 3180 struct vm_page *pg; 3181 { 3182 boolean_t rv; 3183 3184 if (pg->mdpage.pvh_attrs & PT_H) { 3185 rv = TRUE; 3186 pmap_clearbit(pg, PT_H); 3187 } else 3188 rv = FALSE; 3189 3190 PDEBUG(0, printf("pmap_clear_reference pa=%08lx -> %d\n", 3191 VM_PAGE_TO_PHYS(pg), rv)); 3192 3193 return (rv); 3194 } 3195 3196 3197 void 3198 pmap_copy_on_write(pg) 3199 struct vm_page *pg; 3200 { 3201 PDEBUG(0, printf("pmap_copy_on_write pa=%08lx\n", VM_PAGE_TO_PHYS(pg))); 3202 pmap_clearbit(pg, PT_Wr); 3203 } 3204 3205 /* 3206 * pmap_is_modified: 3207 * 3208 * Test if a page has the "modified" attribute. 3209 */ 3210 /* See <arm/arm32/pmap.h> */ 3211 3212 /* 3213 * pmap_is_referenced: 3214 * 3215 * Test if a page has the "referenced" attribute. 3216 */ 3217 /* See <arm/arm32/pmap.h> */ 3218 3219 int 3220 pmap_modified_emulation(pmap, va) 3221 struct pmap *pmap; 3222 vaddr_t va; 3223 { 3224 pt_entry_t *ptes; 3225 struct vm_page *pg; 3226 paddr_t pa; 3227 u_int flags; 3228 int rv = 0; 3229 3230 PDEBUG(2, printf("pmap_modified_emulation\n")); 3231 3232 PMAP_MAP_TO_HEAD_LOCK(); 3233 ptes = pmap_map_ptes(pmap); /* locks pmap */ 3234 3235 if (pmap_pde_v(pmap_pde(pmap, va)) == 0) { 3236 PDEBUG(2, printf("L1 PTE invalid\n")); 3237 goto out; 3238 } 3239 3240 PDEBUG(1, printf("pte=%08x\n", ptes[arm_btop(va)])); 3241 3242 /* Check for a invalid pte */ 3243 if (l2pte_valid(ptes[arm_btop(va)]) == 0) 3244 goto out; 3245 3246 /* This can happen if user code tries to access kernel memory. */ 3247 if ((ptes[arm_btop(va)] & PT_AP(AP_W)) != 0) 3248 goto out; 3249 3250 /* Extract the physical address of the page */ 3251 pa = l2pte_pa(ptes[arm_btop(va)]); 3252 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) 3253 goto out; 3254 3255 /* Get the current flags for this page. */ 3256 simple_lock(&pg->mdpage.pvh_slock); 3257 3258 flags = pmap_modify_pv(pmap, va, pg, 0, 0); 3259 PDEBUG(2, printf("pmap_modified_emulation: flags = %08x\n", flags)); 3260 3261 /* 3262 * Do the flags say this page is writable ? If not then it is a 3263 * genuine write fault. If yes then the write fault is our fault 3264 * as we did not reflect the write access in the PTE. Now we know 3265 * a write has occurred we can correct this and also set the 3266 * modified bit 3267 */ 3268 if (~flags & PT_Wr) { 3269 simple_unlock(&pg->mdpage.pvh_slock); 3270 goto out; 3271 } 3272 3273 PDEBUG(0, 3274 printf("pmap_modified_emulation: Got a hit va=%08lx, pte = %08x\n", 3275 va, ptes[arm_btop(va)])); 3276 pg->mdpage.pvh_attrs |= PT_H | PT_M; 3277 3278 /* 3279 * Re-enable write permissions for the page. No need to call 3280 * pmap_vac_me_harder(), since this is just a 3281 * modified-emulation fault, and the PT_Wr bit isn't changing. We've 3282 * already set the cacheable bits based on the assumption that we 3283 * can write to this page. 3284 */ 3285 ptes[arm_btop(va)] = 3286 (ptes[arm_btop(va)] & ~L2_MASK) | L2_SPAGE | PT_AP(AP_W); 3287 PDEBUG(0, printf("->(%08x)\n", ptes[arm_btop(va)])); 3288 3289 simple_unlock(&pg->mdpage.pvh_slock); 3290 3291 cpu_tlb_flushID_SE(va); 3292 cpu_cpwait(); 3293 rv = 1; 3294 out: 3295 pmap_unmap_ptes(pmap); /* unlocks pmap */ 3296 PMAP_MAP_TO_HEAD_UNLOCK(); 3297 return (rv); 3298 } 3299 3300 int 3301 pmap_handled_emulation(pmap, va) 3302 struct pmap *pmap; 3303 vaddr_t va; 3304 { 3305 pt_entry_t *ptes; 3306 struct vm_page *pg; 3307 paddr_t pa; 3308 int rv = 0; 3309 3310 PDEBUG(2, printf("pmap_handled_emulation\n")); 3311 3312 PMAP_MAP_TO_HEAD_LOCK(); 3313 ptes = pmap_map_ptes(pmap); /* locks pmap */ 3314 3315 if (pmap_pde_v(pmap_pde(pmap, va)) == 0) { 3316 PDEBUG(2, printf("L1 PTE invalid\n")); 3317 goto out; 3318 } 3319 3320 PDEBUG(1, printf("pte=%08x\n", ptes[arm_btop(va)])); 3321 3322 /* Check for invalid pte */ 3323 if (l2pte_valid(ptes[arm_btop(va)]) == 0) 3324 goto out; 3325 3326 /* This can happen if user code tries to access kernel memory. */ 3327 if ((ptes[arm_btop(va)] & L2_MASK) != L2_INVAL) 3328 goto out; 3329 3330 /* Extract the physical address of the page */ 3331 pa = l2pte_pa(ptes[arm_btop(va)]); 3332 if ((pg = PHYS_TO_VM_PAGE(pa)) == NULL) 3333 goto out; 3334 3335 simple_lock(&pg->mdpage.pvh_slock); 3336 3337 /* 3338 * Ok we just enable the pte and mark the attibs as handled 3339 * XXX Should we traverse the PV list and enable all PTEs? 3340 */ 3341 PDEBUG(0, 3342 printf("pmap_handled_emulation: Got a hit va=%08lx pte = %08x\n", 3343 va, ptes[arm_btop(va)])); 3344 pg->mdpage.pvh_attrs |= PT_H; 3345 3346 ptes[arm_btop(va)] = (ptes[arm_btop(va)] & ~L2_MASK) | L2_SPAGE; 3347 PDEBUG(0, printf("->(%08x)\n", ptes[arm_btop(va)])); 3348 3349 simple_unlock(&pg->mdpage.pvh_slock); 3350 3351 cpu_tlb_flushID_SE(va); 3352 cpu_cpwait(); 3353 rv = 1; 3354 out: 3355 pmap_unmap_ptes(pmap); /* unlocks pmap */ 3356 PMAP_MAP_TO_HEAD_UNLOCK(); 3357 return (rv); 3358 } 3359 3360 /* 3361 * pmap_collect: free resources held by a pmap 3362 * 3363 * => optional function. 3364 * => called when a process is swapped out to free memory. 3365 */ 3366 3367 void 3368 pmap_collect(pmap) 3369 struct pmap *pmap; 3370 { 3371 } 3372 3373 /* 3374 * Routine: pmap_procwr 3375 * 3376 * Function: 3377 * Synchronize caches corresponding to [addr, addr+len) in p. 3378 * 3379 */ 3380 void 3381 pmap_procwr(p, va, len) 3382 struct proc *p; 3383 vaddr_t va; 3384 int len; 3385 { 3386 /* We only need to do anything if it is the current process. */ 3387 if (p == curproc) 3388 cpu_icache_sync_range(va, len); 3389 } 3390 /* 3391 * PTP functions 3392 */ 3393 3394 /* 3395 * pmap_get_ptp: get a PTP (if there isn't one, allocate a new one) 3396 * 3397 * => pmap should NOT be pmap_kernel() 3398 * => pmap should be locked 3399 */ 3400 3401 static struct vm_page * 3402 pmap_get_ptp(struct pmap *pmap, vaddr_t va) 3403 { 3404 struct vm_page *ptp; 3405 3406 if (pmap_pde_page(pmap_pde(pmap, va))) { 3407 3408 /* valid... check hint (saves us a PA->PG lookup) */ 3409 #if 0 3410 if (pmap->pm_ptphint && 3411 ((unsigned)pmap_pde(pmap, va) & PG_FRAME) == 3412 VM_PAGE_TO_PHYS(pmap->pm_ptphint)) 3413 return (pmap->pm_ptphint); 3414 #endif 3415 ptp = uvm_pagelookup(&pmap->pm_obj, va); 3416 #ifdef DIAGNOSTIC 3417 if (ptp == NULL) 3418 panic("pmap_get_ptp: unmanaged user PTP"); 3419 #endif 3420 // pmap->pm_ptphint = ptp; 3421 return(ptp); 3422 } 3423 3424 /* allocate a new PTP (updates ptphint) */ 3425 return(pmap_alloc_ptp(pmap, va)); 3426 } 3427 3428 /* 3429 * pmap_alloc_ptp: allocate a PTP for a PMAP 3430 * 3431 * => pmap should already be locked by caller 3432 * => we use the ptp's wire_count to count the number of active mappings 3433 * in the PTP (we start it at one to prevent any chance this PTP 3434 * will ever leak onto the active/inactive queues) 3435 */ 3436 3437 /*__inline */ static struct vm_page * 3438 pmap_alloc_ptp(struct pmap *pmap, vaddr_t va) 3439 { 3440 struct vm_page *ptp; 3441 3442 ptp = uvm_pagealloc(&pmap->pm_obj, va, NULL, 3443 UVM_PGA_USERESERVE|UVM_PGA_ZERO); 3444 if (ptp == NULL) 3445 return (NULL); 3446 3447 /* got one! */ 3448 ptp->flags &= ~PG_BUSY; /* never busy */ 3449 ptp->wire_count = 1; /* no mappings yet */ 3450 pmap_map_in_l1(pmap, va, VM_PAGE_TO_PHYS(ptp), TRUE); 3451 pmap->pm_stats.resident_count++; /* count PTP as resident */ 3452 // pmap->pm_ptphint = ptp; 3453 return (ptp); 3454 } 3455 3456 vaddr_t 3457 pmap_growkernel(maxkvaddr) 3458 vaddr_t maxkvaddr; 3459 { 3460 struct pmap *kpm = pmap_kernel(), *pm; 3461 int s; 3462 paddr_t ptaddr; 3463 struct vm_page *ptp; 3464 3465 if (maxkvaddr <= pmap_curmaxkvaddr) 3466 goto out; /* we are OK */ 3467 NPDEBUG(PDB_GROWKERN, printf("pmap_growkernel: growing kernel from %lx to %lx\n", 3468 pmap_curmaxkvaddr, maxkvaddr)); 3469 3470 /* 3471 * whoops! we need to add kernel PTPs 3472 */ 3473 3474 s = splhigh(); /* to be safe */ 3475 simple_lock(&kpm->pm_obj.vmobjlock); 3476 /* due to the way the arm pmap works we map 4MB at a time */ 3477 for (/*null*/ ; pmap_curmaxkvaddr < maxkvaddr ; pmap_curmaxkvaddr += 4 * NBPD) { 3478 3479 if (uvm.page_init_done == FALSE) { 3480 3481 /* 3482 * we're growing the kernel pmap early (from 3483 * uvm_pageboot_alloc()). this case must be 3484 * handled a little differently. 3485 */ 3486 3487 if (uvm_page_physget(&ptaddr) == FALSE) 3488 panic("pmap_growkernel: out of memory"); 3489 pmap_zero_page(ptaddr); 3490 3491 /* map this page in */ 3492 pmap_map_in_l1(kpm, (pmap_curmaxkvaddr + 1), ptaddr, TRUE); 3493 3494 /* count PTP as resident */ 3495 kpm->pm_stats.resident_count++; 3496 continue; 3497 } 3498 3499 /* 3500 * THIS *MUST* BE CODED SO AS TO WORK IN THE 3501 * pmap_initialized == FALSE CASE! WE MAY BE 3502 * INVOKED WHILE pmap_init() IS RUNNING! 3503 */ 3504 3505 if ((ptp = pmap_alloc_ptp(kpm, (pmap_curmaxkvaddr + 1))) == NULL) { 3506 panic("pmap_growkernel: alloc ptp failed"); 3507 } 3508 3509 /* distribute new kernel PTP to all active pmaps */ 3510 simple_lock(&pmaps_lock); 3511 LIST_FOREACH(pm, &pmaps, pm_list) { 3512 pmap_map_in_l1(pm, (pmap_curmaxkvaddr + 1), VM_PAGE_TO_PHYS(ptp), TRUE); 3513 } 3514 3515 simple_unlock(&pmaps_lock); 3516 } 3517 3518 /* 3519 * flush out the cache, expensive but growkernel will happen so 3520 * rarely 3521 */ 3522 cpu_tlb_flushD(); 3523 cpu_cpwait(); 3524 3525 simple_unlock(&kpm->pm_obj.vmobjlock); 3526 splx(s); 3527 3528 out: 3529 return (pmap_curmaxkvaddr); 3530 } 3531 3532 3533 3534 /************************ Bootstrapping routines ****************************/ 3535 3536 /* 3537 * This list exists for the benefit of pmap_map_chunk(). It keeps track 3538 * of the kernel L2 tables during bootstrap, so that pmap_map_chunk() can 3539 * find them as necessary. 3540 * 3541 * Note that the data on this list is not valid after initarm() returns. 3542 */ 3543 SLIST_HEAD(, pv_addr) kernel_pt_list = SLIST_HEAD_INITIALIZER(kernel_pt_list); 3544 3545 static vaddr_t 3546 kernel_pt_lookup(paddr_t pa) 3547 { 3548 pv_addr_t *pv; 3549 3550 SLIST_FOREACH(pv, &kernel_pt_list, pv_list) { 3551 if (pv->pv_pa == pa) 3552 return (pv->pv_va); 3553 } 3554 return (0); 3555 } 3556 3557 /* 3558 * pmap_map_section: 3559 * 3560 * Create a single section mapping. 3561 */ 3562 void 3563 pmap_map_section(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache) 3564 { 3565 pd_entry_t *pde = (pd_entry_t *) l1pt; 3566 pd_entry_t ap = (prot & VM_PROT_WRITE) ? AP_KRW : AP_KR; 3567 pd_entry_t fl = (cache == PTE_CACHE) ? pte_cache_mode : 0; 3568 3569 KASSERT(((va | pa) & (L1_SEC_SIZE - 1)) == 0); 3570 3571 pde[va >> PDSHIFT] = L1_SECPTE(pa & PD_MASK, ap, fl); 3572 } 3573 3574 /* 3575 * pmap_map_entry: 3576 * 3577 * Create a single page mapping. 3578 */ 3579 void 3580 pmap_map_entry(vaddr_t l1pt, vaddr_t va, paddr_t pa, int prot, int cache) 3581 { 3582 pd_entry_t *pde = (pd_entry_t *) l1pt; 3583 pt_entry_t ap = (prot & VM_PROT_WRITE) ? AP_KRW : AP_KR; 3584 pt_entry_t fl = (cache == PTE_CACHE) ? pte_cache_mode : 0; 3585 pt_entry_t *pte; 3586 3587 KASSERT(((va | pa) & PGOFSET) == 0); 3588 3589 if ((pde[va >> PDSHIFT] & L1_MASK) != L1_PAGE) 3590 panic("pmap_map_entry: no L2 table for VA 0x%08lx", va); 3591 3592 pte = (pt_entry_t *) 3593 kernel_pt_lookup(pde[va >> PDSHIFT] & PG_FRAME); 3594 if (pte == NULL) 3595 panic("pmap_map_entry: can't find L2 table for VA 0x%08lx", va); 3596 3597 pte[(va >> PGSHIFT) & 0x3ff] = L2_SPTE(pa & PG_FRAME, ap, fl); 3598 } 3599 3600 /* 3601 * pmap_link_l2pt: 3602 * 3603 * Link the L2 page table specified by "pa" into the L1 3604 * page table at the slot for "va". 3605 */ 3606 void 3607 pmap_link_l2pt(vaddr_t l1pt, vaddr_t va, pv_addr_t *l2pv) 3608 { 3609 pd_entry_t *pde = (pd_entry_t *) l1pt; 3610 u_int slot = va >> PDSHIFT; 3611 3612 KASSERT((l2pv->pv_pa & PGOFSET) == 0); 3613 3614 pde[slot + 0] = L1_PTE(l2pv->pv_pa + 0x000); 3615 pde[slot + 1] = L1_PTE(l2pv->pv_pa + 0x400); 3616 pde[slot + 2] = L1_PTE(l2pv->pv_pa + 0x800); 3617 pde[slot + 3] = L1_PTE(l2pv->pv_pa + 0xc00); 3618 3619 SLIST_INSERT_HEAD(&kernel_pt_list, l2pv, pv_list); 3620 } 3621 3622 /* 3623 * pmap_map_chunk: 3624 * 3625 * Map a chunk of memory using the most efficient mappings 3626 * possible (section, large page, small page) into the 3627 * provided L1 and L2 tables at the specified virtual address. 3628 */ 3629 vsize_t 3630 pmap_map_chunk(vaddr_t l1pt, vaddr_t va, paddr_t pa, vsize_t size, 3631 int prot, int cache) 3632 { 3633 pd_entry_t *pde = (pd_entry_t *) l1pt; 3634 pt_entry_t ap = (prot & VM_PROT_WRITE) ? AP_KRW : AP_KR; 3635 pt_entry_t fl = (cache == PTE_CACHE) ? pte_cache_mode : 0; 3636 pt_entry_t *pte; 3637 vsize_t resid; 3638 int i; 3639 3640 resid = (size + (NBPG - 1)) & ~(NBPG - 1); 3641 3642 if (l1pt == 0) 3643 panic("pmap_map_chunk: no L1 table provided"); 3644 3645 #ifdef VERBOSE_INIT_ARM 3646 printf("pmap_map_chunk: pa=0x%lx va=0x%lx size=0x%lx resid=0x%lx " 3647 "prot=0x%x cache=%d\n", pa, va, size, resid, prot, cache); 3648 #endif 3649 3650 size = resid; 3651 3652 while (resid > 0) { 3653 /* See if we can use a section mapping. */ 3654 if (((pa | va) & (L1_SEC_SIZE - 1)) == 0 && 3655 resid >= L1_SEC_SIZE) { 3656 #ifdef VERBOSE_INIT_ARM 3657 printf("S"); 3658 #endif 3659 pde[va >> PDSHIFT] = L1_SECPTE(pa, ap, fl); 3660 va += L1_SEC_SIZE; 3661 pa += L1_SEC_SIZE; 3662 resid -= L1_SEC_SIZE; 3663 continue; 3664 } 3665 3666 /* 3667 * Ok, we're going to use an L2 table. Make sure 3668 * one is actually in the corresponding L1 slot 3669 * for the current VA. 3670 */ 3671 if ((pde[va >> PDSHIFT] & L1_MASK) != L1_PAGE) 3672 panic("pmap_map_chunk: no L2 table for VA 0x%08lx", va); 3673 3674 pte = (pt_entry_t *) 3675 kernel_pt_lookup(pde[va >> PDSHIFT] & PG_FRAME); 3676 if (pte == NULL) 3677 panic("pmap_map_chunk: can't find L2 table for VA" 3678 "0x%08lx", va); 3679 3680 /* See if we can use a L2 large page mapping. */ 3681 if (((pa | va) & (L2_LPAGE_SIZE - 1)) == 0 && 3682 resid >= L2_LPAGE_SIZE) { 3683 #ifdef VERBOSE_INIT_ARM 3684 printf("L"); 3685 #endif 3686 for (i = 0; i < 16; i++) { 3687 pte[((va >> PGSHIFT) & 0x3f0) + i] = 3688 L2_LPTE(pa, ap, fl); 3689 } 3690 va += L2_LPAGE_SIZE; 3691 pa += L2_LPAGE_SIZE; 3692 resid -= L2_LPAGE_SIZE; 3693 continue; 3694 } 3695 3696 /* Use a small page mapping. */ 3697 #ifdef VERBOSE_INIT_ARM 3698 printf("P"); 3699 #endif 3700 pte[(va >> PGSHIFT) & 0x3ff] = L2_SPTE(pa, ap, fl); 3701 va += NBPG; 3702 pa += NBPG; 3703 resid -= NBPG; 3704 } 3705 #ifdef VERBOSE_INIT_ARM 3706 printf("\n"); 3707 #endif 3708 return (size); 3709 } 3710
Indexes created Sat Oct 25 16:10:12 GMT 2025