1 /* 2 * linux/arch/arm/mm/mmu.c 3 * 4 * Copyright (C) 1995-2005 Russell King 5 * 6 * This program is free software; you can redistribute it and/or modify 7 * it under the terms of the GNU General Public License version 2 as 8 * published by the Free Software Foundation. 9 */ 10 #include <linux/module.h> 11 #include <linux/kernel.h> 12 #include <linux/errno.h> 13 #include <linux/init.h> 14 #include <linux/mman.h> 15 #include <linux/nodemask.h> 16 #include <linux/memblock.h> 17 #include <linux/fs.h> 18 #include <linux/vmalloc.h> 19 #include <linux/sizes.h> 20 21 #include <asm/cp15.h> 22 #include <asm/cputype.h> 23 #include <asm/sections.h> 24 #include <asm/cachetype.h> 25 #include <asm/fixmap.h> 26 #include <asm/sections.h> 27 #include <asm/setup.h> 28 #include <asm/smp_plat.h> 29 #include <asm/tlb.h> 30 #include <asm/highmem.h> 31 #include <asm/system_info.h> 32 #include <asm/traps.h> 33 #include <asm/procinfo.h> 34 #include <asm/memory.h> 35 36 #include <asm/mach/arch.h> 37 #include <asm/mach/map.h> 38 #include <asm/mach/pci.h> 39 #include <asm/fixmap.h> 40 41 #include "fault.h" 42 #include "mm.h" 43 #include "tcm.h" 44 45 /* 46 * empty_zero_page is a special page that is used for 47 * zero-initialized data and COW. 48 */ 49 struct page *empty_zero_page; 50 EXPORT_SYMBOL(empty_zero_page); 51 52 /* 53 * The pmd table for the upper-most set of pages. 54 */ 55 pmd_t *top_pmd; 56 57 pmdval_t user_pmd_table = _PAGE_USER_TABLE; 58 59 #define CPOLICY_UNCACHED 0 60 #define CPOLICY_BUFFERED 1 61 #define CPOLICY_WRITETHROUGH 2 62 #define CPOLICY_WRITEBACK 3 63 #define CPOLICY_WRITEALLOC 4 64 65 static unsigned int cachepolicy __initdata = CPOLICY_WRITEBACK; 66 static unsigned int ecc_mask __initdata = 0; 67 pgprot_t pgprot_user; 68 pgprot_t pgprot_kernel; 69 pgprot_t pgprot_hyp_device; 70 pgprot_t pgprot_s2; 71 pgprot_t pgprot_s2_device; 72 73 EXPORT_SYMBOL(pgprot_user); 74 EXPORT_SYMBOL(pgprot_kernel); 75 76 struct cachepolicy { 77 const char policy[16]; 78 unsigned int cr_mask; 79 pmdval_t pmd; 80 pteval_t pte; 81 pteval_t pte_s2; 82 }; 83 84 #ifdef CONFIG_ARM_LPAE 85 #define s2_policy(policy) policy 86 #else 87 #define s2_policy(policy) 0 88 #endif 89 90 static struct cachepolicy cache_policies[] __initdata = { 91 { 92 .policy = "uncached", 93 .cr_mask = CR_W|CR_C, 94 .pmd = PMD_SECT_UNCACHED, 95 .pte = L_PTE_MT_UNCACHED, 96 .pte_s2 = s2_policy(L_PTE_S2_MT_UNCACHED), 97 }, { 98 .policy = "buffered", 99 .cr_mask = CR_C, 100 .pmd = PMD_SECT_BUFFERED, 101 .pte = L_PTE_MT_BUFFERABLE, 102 .pte_s2 = s2_policy(L_PTE_S2_MT_UNCACHED), 103 }, { 104 .policy = "writethrough", 105 .cr_mask = 0, 106 .pmd = PMD_SECT_WT, 107 .pte = L_PTE_MT_WRITETHROUGH, 108 .pte_s2 = s2_policy(L_PTE_S2_MT_WRITETHROUGH), 109 }, { 110 .policy = "writeback", 111 .cr_mask = 0, 112 .pmd = PMD_SECT_WB, 113 .pte = L_PTE_MT_WRITEBACK, 114 .pte_s2 = s2_policy(L_PTE_S2_MT_WRITEBACK), 115 }, { 116 .policy = "writealloc", 117 .cr_mask = 0, 118 .pmd = PMD_SECT_WBWA, 119 .pte = L_PTE_MT_WRITEALLOC, 120 .pte_s2 = s2_policy(L_PTE_S2_MT_WRITEBACK), 121 } 122 }; 123 124 #ifdef CONFIG_CPU_CP15 125 static unsigned long initial_pmd_value __initdata = 0; 126 127 /* 128 * Initialise the cache_policy variable with the initial state specified 129 * via the "pmd" value. This is used to ensure that on ARMv6 and later, 130 * the C code sets the page tables up with the same policy as the head 131 * assembly code, which avoids an illegal state where the TLBs can get 132 * confused. See comments in early_cachepolicy() for more information. 133 */ 134 void __init init_default_cache_policy(unsigned long pmd) 135 { 136 int i; 137 138 initial_pmd_value = pmd; 139 140 pmd &= PMD_SECT_TEX(1) | PMD_SECT_BUFFERABLE | PMD_SECT_CACHEABLE; 141 142 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) 143 if (cache_policies[i].pmd == pmd) { 144 cachepolicy = i; 145 break; 146 } 147 148 if (i == ARRAY_SIZE(cache_policies)) 149 pr_err("ERROR: could not find cache policy\n"); 150 } 151 152 /* 153 * These are useful for identifying cache coherency problems by allowing 154 * the cache or the cache and writebuffer to be turned off. (Note: the 155 * write buffer should not be on and the cache off). 156 */ 157 static int __init early_cachepolicy(char *p) 158 { 159 int i, selected = -1; 160 161 for (i = 0; i < ARRAY_SIZE(cache_policies); i++) { 162 int len = strlen(cache_policies[i].policy); 163 164 if (memcmp(p, cache_policies[i].policy, len) == 0) { 165 selected = i; 166 break; 167 } 168 } 169 170 if (selected == -1) 171 pr_err("ERROR: unknown or unsupported cache policy\n"); 172 173 /* 174 * This restriction is partly to do with the way we boot; it is 175 * unpredictable to have memory mapped using two different sets of 176 * memory attributes (shared, type, and cache attribs). We can not 177 * change these attributes once the initial assembly has setup the 178 * page tables. 179 */ 180 if (cpu_architecture() >= CPU_ARCH_ARMv6 && selected != cachepolicy) { 181 pr_warn("Only cachepolicy=%s supported on ARMv6 and later\n", 182 cache_policies[cachepolicy].policy); 183 return 0; 184 } 185 186 if (selected != cachepolicy) { 187 unsigned long cr = __clear_cr(cache_policies[selected].cr_mask); 188 cachepolicy = selected; 189 flush_cache_all(); 190 set_cr(cr); 191 } 192 return 0; 193 } 194 early_param("cachepolicy", early_cachepolicy); 195 196 static int __init early_nocache(char *__unused) 197 { 198 char *p = "buffered"; 199 pr_warn("nocache is deprecated; use cachepolicy=%s\n", p); 200 early_cachepolicy(p); 201 return 0; 202 } 203 early_param("nocache", early_nocache); 204 205 static int __init early_nowrite(char *__unused) 206 { 207 char *p = "uncached"; 208 pr_warn("nowb is deprecated; use cachepolicy=%s\n", p); 209 early_cachepolicy(p); 210 return 0; 211 } 212 early_param("nowb", early_nowrite); 213 214 #ifndef CONFIG_ARM_LPAE 215 static int __init early_ecc(char *p) 216 { 217 if (memcmp(p, "on", 2) == 0) 218 ecc_mask = PMD_PROTECTION; 219 else if (memcmp(p, "off", 3) == 0) 220 ecc_mask = 0; 221 return 0; 222 } 223 early_param("ecc", early_ecc); 224 #endif 225 226 #else /* ifdef CONFIG_CPU_CP15 */ 227 228 static int __init early_cachepolicy(char *p) 229 { 230 pr_warn("cachepolicy kernel parameter not supported without cp15\n"); 231 } 232 early_param("cachepolicy", early_cachepolicy); 233 234 static int __init noalign_setup(char *__unused) 235 { 236 pr_warn("noalign kernel parameter not supported without cp15\n"); 237 } 238 __setup("noalign", noalign_setup); 239 240 #endif /* ifdef CONFIG_CPU_CP15 / else */ 241 242 #define PROT_PTE_DEVICE L_PTE_PRESENT|L_PTE_YOUNG|L_PTE_DIRTY|L_PTE_XN 243 #define PROT_PTE_S2_DEVICE PROT_PTE_DEVICE 244 #define PROT_SECT_DEVICE PMD_TYPE_SECT|PMD_SECT_AP_WRITE 245 246 static struct mem_type mem_types[] = { 247 [MT_DEVICE] = { /* Strongly ordered / ARMv6 shared device */ 248 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_SHARED | 249 L_PTE_SHARED, 250 .prot_pte_s2 = s2_policy(PROT_PTE_S2_DEVICE) | 251 s2_policy(L_PTE_S2_MT_DEV_SHARED) | 252 L_PTE_SHARED, 253 .prot_l1 = PMD_TYPE_TABLE, 254 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_S, 255 .domain = DOMAIN_IO, 256 }, 257 [MT_DEVICE_NONSHARED] = { /* ARMv6 non-shared device */ 258 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_NONSHARED, 259 .prot_l1 = PMD_TYPE_TABLE, 260 .prot_sect = PROT_SECT_DEVICE, 261 .domain = DOMAIN_IO, 262 }, 263 [MT_DEVICE_CACHED] = { /* ioremap_cached */ 264 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_CACHED, 265 .prot_l1 = PMD_TYPE_TABLE, 266 .prot_sect = PROT_SECT_DEVICE | PMD_SECT_WB, 267 .domain = DOMAIN_IO, 268 }, 269 [MT_DEVICE_WC] = { /* ioremap_wc */ 270 .prot_pte = PROT_PTE_DEVICE | L_PTE_MT_DEV_WC, 271 .prot_l1 = PMD_TYPE_TABLE, 272 .prot_sect = PROT_SECT_DEVICE, 273 .domain = DOMAIN_IO, 274 }, 275 [MT_UNCACHED] = { 276 .prot_pte = PROT_PTE_DEVICE, 277 .prot_l1 = PMD_TYPE_TABLE, 278 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN, 279 .domain = DOMAIN_IO, 280 }, 281 [MT_CACHECLEAN] = { 282 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN, 283 .domain = DOMAIN_KERNEL, 284 }, 285 #ifndef CONFIG_ARM_LPAE 286 [MT_MINICLEAN] = { 287 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN | PMD_SECT_MINICACHE, 288 .domain = DOMAIN_KERNEL, 289 }, 290 #endif 291 [MT_LOW_VECTORS] = { 292 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY | 293 L_PTE_RDONLY, 294 .prot_l1 = PMD_TYPE_TABLE, 295 .domain = DOMAIN_VECTORS, 296 }, 297 [MT_HIGH_VECTORS] = { 298 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY | 299 L_PTE_USER | L_PTE_RDONLY, 300 .prot_l1 = PMD_TYPE_TABLE, 301 .domain = DOMAIN_VECTORS, 302 }, 303 [MT_MEMORY_RWX] = { 304 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY, 305 .prot_l1 = PMD_TYPE_TABLE, 306 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE, 307 .domain = DOMAIN_KERNEL, 308 }, 309 [MT_MEMORY_RW] = { 310 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY | 311 L_PTE_XN, 312 .prot_l1 = PMD_TYPE_TABLE, 313 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE, 314 .domain = DOMAIN_KERNEL, 315 }, 316 [MT_ROM] = { 317 .prot_sect = PMD_TYPE_SECT, 318 .domain = DOMAIN_KERNEL, 319 }, 320 [MT_MEMORY_RWX_NONCACHED] = { 321 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY | 322 L_PTE_MT_BUFFERABLE, 323 .prot_l1 = PMD_TYPE_TABLE, 324 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE, 325 .domain = DOMAIN_KERNEL, 326 }, 327 [MT_MEMORY_RW_DTCM] = { 328 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY | 329 L_PTE_XN, 330 .prot_l1 = PMD_TYPE_TABLE, 331 .prot_sect = PMD_TYPE_SECT | PMD_SECT_XN, 332 .domain = DOMAIN_KERNEL, 333 }, 334 [MT_MEMORY_RWX_ITCM] = { 335 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY, 336 .prot_l1 = PMD_TYPE_TABLE, 337 .domain = DOMAIN_KERNEL, 338 }, 339 [MT_MEMORY_RW_SO] = { 340 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY | 341 L_PTE_MT_UNCACHED | L_PTE_XN, 342 .prot_l1 = PMD_TYPE_TABLE, 343 .prot_sect = PMD_TYPE_SECT | PMD_SECT_AP_WRITE | PMD_SECT_S | 344 PMD_SECT_UNCACHED | PMD_SECT_XN, 345 .domain = DOMAIN_KERNEL, 346 }, 347 [MT_MEMORY_DMA_READY] = { 348 .prot_pte = L_PTE_PRESENT | L_PTE_YOUNG | L_PTE_DIRTY | 349 L_PTE_XN, 350 .prot_l1 = PMD_TYPE_TABLE, 351 .domain = DOMAIN_KERNEL, 352 }, 353 }; 354 355 const struct mem_type *get_mem_type(unsigned int type) 356 { 357 return type < ARRAY_SIZE(mem_types) ? &mem_types[type] : NULL; 358 } 359 EXPORT_SYMBOL(get_mem_type); 360 361 static pte_t *(*pte_offset_fixmap)(pmd_t *dir, unsigned long addr); 362 363 static pte_t bm_pte[PTRS_PER_PTE + PTE_HWTABLE_PTRS] 364 __aligned(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE) __initdata; 365 366 static pte_t * __init pte_offset_early_fixmap(pmd_t *dir, unsigned long addr) 367 { 368 return &bm_pte[pte_index(addr)]; 369 } 370 371 static pte_t *pte_offset_late_fixmap(pmd_t *dir, unsigned long addr) 372 { 373 return pte_offset_kernel(dir, addr); 374 } 375 376 static inline pmd_t * __init fixmap_pmd(unsigned long addr) 377 { 378 pgd_t *pgd = pgd_offset_k(addr); 379 pud_t *pud = pud_offset(pgd, addr); 380 pmd_t *pmd = pmd_offset(pud, addr); 381 382 return pmd; 383 } 384 385 void __init early_fixmap_init(void) 386 { 387 pmd_t *pmd; 388 389 /* 390 * The early fixmap range spans multiple pmds, for which 391 * we are not prepared: 392 */ 393 BUILD_BUG_ON((__fix_to_virt(__end_of_early_ioremap_region) >> PMD_SHIFT) 394 != FIXADDR_TOP >> PMD_SHIFT); 395 396 pmd = fixmap_pmd(FIXADDR_TOP); 397 pmd_populate_kernel(&init_mm, pmd, bm_pte); 398 399 pte_offset_fixmap = pte_offset_early_fixmap; 400 } 401 402 /* 403 * To avoid TLB flush broadcasts, this uses local_flush_tlb_kernel_range(). 404 * As a result, this can only be called with preemption disabled, as under 405 * stop_machine(). 406 */ 407 void __set_fixmap(enum fixed_addresses idx, phys_addr_t phys, pgprot_t prot) 408 { 409 unsigned long vaddr = __fix_to_virt(idx); 410 pte_t *pte = pte_offset_fixmap(pmd_off_k(vaddr), vaddr); 411 412 /* Make sure fixmap region does not exceed available allocation. */ 413 BUILD_BUG_ON(FIXADDR_START + (__end_of_fixed_addresses * PAGE_SIZE) > 414 FIXADDR_END); 415 BUG_ON(idx >= __end_of_fixed_addresses); 416 417 if (pgprot_val(prot)) 418 set_pte_at(NULL, vaddr, pte, 419 pfn_pte(phys >> PAGE_SHIFT, prot)); 420 else 421 pte_clear(NULL, vaddr, pte); 422 local_flush_tlb_kernel_range(vaddr, vaddr + PAGE_SIZE); 423 } 424 425 /* 426 * Adjust the PMD section entries according to the CPU in use. 427 */ 428 static void __init build_mem_type_table(void) 429 { 430 struct cachepolicy *cp; 431 unsigned int cr = get_cr(); 432 pteval_t user_pgprot, kern_pgprot, vecs_pgprot; 433 pteval_t hyp_device_pgprot, s2_pgprot, s2_device_pgprot; 434 int cpu_arch = cpu_architecture(); 435 int i; 436 437 if (cpu_arch < CPU_ARCH_ARMv6) { 438 #if defined(CONFIG_CPU_DCACHE_DISABLE) 439 if (cachepolicy > CPOLICY_BUFFERED) 440 cachepolicy = CPOLICY_BUFFERED; 441 #elif defined(CONFIG_CPU_DCACHE_WRITETHROUGH) 442 if (cachepolicy > CPOLICY_WRITETHROUGH) 443 cachepolicy = CPOLICY_WRITETHROUGH; 444 #endif 445 } 446 if (cpu_arch < CPU_ARCH_ARMv5) { 447 if (cachepolicy >= CPOLICY_WRITEALLOC) 448 cachepolicy = CPOLICY_WRITEBACK; 449 ecc_mask = 0; 450 } 451 452 if (is_smp()) { 453 if (cachepolicy != CPOLICY_WRITEALLOC) { 454 pr_warn("Forcing write-allocate cache policy for SMP\n"); 455 cachepolicy = CPOLICY_WRITEALLOC; 456 } 457 if (!(initial_pmd_value & PMD_SECT_S)) { 458 pr_warn("Forcing shared mappings for SMP\n"); 459 initial_pmd_value |= PMD_SECT_S; 460 } 461 } 462 463 /* 464 * Strip out features not present on earlier architectures. 465 * Pre-ARMv5 CPUs don't have TEX bits. Pre-ARMv6 CPUs or those 466 * without extended page tables don't have the 'Shared' bit. 467 */ 468 if (cpu_arch < CPU_ARCH_ARMv5) 469 for (i = 0; i < ARRAY_SIZE(mem_types); i++) 470 mem_types[i].prot_sect &= ~PMD_SECT_TEX(7); 471 if ((cpu_arch < CPU_ARCH_ARMv6 || !(cr & CR_XP)) && !cpu_is_xsc3()) 472 for (i = 0; i < ARRAY_SIZE(mem_types); i++) 473 mem_types[i].prot_sect &= ~PMD_SECT_S; 474 475 /* 476 * ARMv5 and lower, bit 4 must be set for page tables (was: cache 477 * "update-able on write" bit on ARM610). However, Xscale and 478 * Xscale3 require this bit to be cleared. 479 */ 480 if (cpu_is_xscale() || cpu_is_xsc3()) { 481 for (i = 0; i < ARRAY_SIZE(mem_types); i++) { 482 mem_types[i].prot_sect &= ~PMD_BIT4; 483 mem_types[i].prot_l1 &= ~PMD_BIT4; 484 } 485 } else if (cpu_arch < CPU_ARCH_ARMv6) { 486 for (i = 0; i < ARRAY_SIZE(mem_types); i++) { 487 if (mem_types[i].prot_l1) 488 mem_types[i].prot_l1 |= PMD_BIT4; 489 if (mem_types[i].prot_sect) 490 mem_types[i].prot_sect |= PMD_BIT4; 491 } 492 } 493 494 /* 495 * Mark the device areas according to the CPU/architecture. 496 */ 497 if (cpu_is_xsc3() || (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP))) { 498 if (!cpu_is_xsc3()) { 499 /* 500 * Mark device regions on ARMv6+ as execute-never 501 * to prevent speculative instruction fetches. 502 */ 503 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_XN; 504 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_XN; 505 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_XN; 506 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_XN; 507 508 /* Also setup NX memory mapping */ 509 mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_XN; 510 } 511 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) { 512 /* 513 * For ARMv7 with TEX remapping, 514 * - shared device is SXCB=1100 515 * - nonshared device is SXCB=0100 516 * - write combine device mem is SXCB=0001 517 * (Uncached Normal memory) 518 */ 519 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1); 520 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(1); 521 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE; 522 } else if (cpu_is_xsc3()) { 523 /* 524 * For Xscale3, 525 * - shared device is TEXCB=00101 526 * - nonshared device is TEXCB=01000 527 * - write combine device mem is TEXCB=00100 528 * (Inner/Outer Uncacheable in xsc3 parlance) 529 */ 530 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_TEX(1) | PMD_SECT_BUFFERED; 531 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2); 532 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1); 533 } else { 534 /* 535 * For ARMv6 and ARMv7 without TEX remapping, 536 * - shared device is TEXCB=00001 537 * - nonshared device is TEXCB=01000 538 * - write combine device mem is TEXCB=00100 539 * (Uncached Normal in ARMv6 parlance). 540 */ 541 mem_types[MT_DEVICE].prot_sect |= PMD_SECT_BUFFERED; 542 mem_types[MT_DEVICE_NONSHARED].prot_sect |= PMD_SECT_TEX(2); 543 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_TEX(1); 544 } 545 } else { 546 /* 547 * On others, write combining is "Uncached/Buffered" 548 */ 549 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_BUFFERABLE; 550 } 551 552 /* 553 * Now deal with the memory-type mappings 554 */ 555 cp = &cache_policies[cachepolicy]; 556 vecs_pgprot = kern_pgprot = user_pgprot = cp->pte; 557 s2_pgprot = cp->pte_s2; 558 hyp_device_pgprot = mem_types[MT_DEVICE].prot_pte; 559 s2_device_pgprot = mem_types[MT_DEVICE].prot_pte_s2; 560 561 #ifndef CONFIG_ARM_LPAE 562 /* 563 * We don't use domains on ARMv6 (since this causes problems with 564 * v6/v7 kernels), so we must use a separate memory type for user 565 * r/o, kernel r/w to map the vectors page. 566 */ 567 if (cpu_arch == CPU_ARCH_ARMv6) 568 vecs_pgprot |= L_PTE_MT_VECTORS; 569 570 /* 571 * Check is it with support for the PXN bit 572 * in the Short-descriptor translation table format descriptors. 573 */ 574 if (cpu_arch == CPU_ARCH_ARMv7 && 575 (read_cpuid_ext(CPUID_EXT_MMFR0) & 0xF) == 4) { 576 user_pmd_table |= PMD_PXNTABLE; 577 } 578 #endif 579 580 /* 581 * ARMv6 and above have extended page tables. 582 */ 583 if (cpu_arch >= CPU_ARCH_ARMv6 && (cr & CR_XP)) { 584 #ifndef CONFIG_ARM_LPAE 585 /* 586 * Mark cache clean areas and XIP ROM read only 587 * from SVC mode and no access from userspace. 588 */ 589 mem_types[MT_ROM].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE; 590 mem_types[MT_MINICLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE; 591 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_APX|PMD_SECT_AP_WRITE; 592 #endif 593 594 /* 595 * If the initial page tables were created with the S bit 596 * set, then we need to do the same here for the same 597 * reasons given in early_cachepolicy(). 598 */ 599 if (initial_pmd_value & PMD_SECT_S) { 600 user_pgprot |= L_PTE_SHARED; 601 kern_pgprot |= L_PTE_SHARED; 602 vecs_pgprot |= L_PTE_SHARED; 603 s2_pgprot |= L_PTE_SHARED; 604 mem_types[MT_DEVICE_WC].prot_sect |= PMD_SECT_S; 605 mem_types[MT_DEVICE_WC].prot_pte |= L_PTE_SHARED; 606 mem_types[MT_DEVICE_CACHED].prot_sect |= PMD_SECT_S; 607 mem_types[MT_DEVICE_CACHED].prot_pte |= L_PTE_SHARED; 608 mem_types[MT_MEMORY_RWX].prot_sect |= PMD_SECT_S; 609 mem_types[MT_MEMORY_RWX].prot_pte |= L_PTE_SHARED; 610 mem_types[MT_MEMORY_RW].prot_sect |= PMD_SECT_S; 611 mem_types[MT_MEMORY_RW].prot_pte |= L_PTE_SHARED; 612 mem_types[MT_MEMORY_DMA_READY].prot_pte |= L_PTE_SHARED; 613 mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_S; 614 mem_types[MT_MEMORY_RWX_NONCACHED].prot_pte |= L_PTE_SHARED; 615 } 616 } 617 618 /* 619 * Non-cacheable Normal - intended for memory areas that must 620 * not cause dirty cache line writebacks when used 621 */ 622 if (cpu_arch >= CPU_ARCH_ARMv6) { 623 if (cpu_arch >= CPU_ARCH_ARMv7 && (cr & CR_TRE)) { 624 /* Non-cacheable Normal is XCB = 001 */ 625 mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= 626 PMD_SECT_BUFFERED; 627 } else { 628 /* For both ARMv6 and non-TEX-remapping ARMv7 */ 629 mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= 630 PMD_SECT_TEX(1); 631 } 632 } else { 633 mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= PMD_SECT_BUFFERABLE; 634 } 635 636 #ifdef CONFIG_ARM_LPAE 637 /* 638 * Do not generate access flag faults for the kernel mappings. 639 */ 640 for (i = 0; i < ARRAY_SIZE(mem_types); i++) { 641 mem_types[i].prot_pte |= PTE_EXT_AF; 642 if (mem_types[i].prot_sect) 643 mem_types[i].prot_sect |= PMD_SECT_AF; 644 } 645 kern_pgprot |= PTE_EXT_AF; 646 vecs_pgprot |= PTE_EXT_AF; 647 648 /* 649 * Set PXN for user mappings 650 */ 651 user_pgprot |= PTE_EXT_PXN; 652 #endif 653 654 for (i = 0; i < 16; i++) { 655 pteval_t v = pgprot_val(protection_map[i]); 656 protection_map[i] = __pgprot(v | user_pgprot); 657 } 658 659 mem_types[MT_LOW_VECTORS].prot_pte |= vecs_pgprot; 660 mem_types[MT_HIGH_VECTORS].prot_pte |= vecs_pgprot; 661 662 pgprot_user = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | user_pgprot); 663 pgprot_kernel = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | 664 L_PTE_DIRTY | kern_pgprot); 665 pgprot_s2 = __pgprot(L_PTE_PRESENT | L_PTE_YOUNG | s2_pgprot); 666 pgprot_s2_device = __pgprot(s2_device_pgprot); 667 pgprot_hyp_device = __pgprot(hyp_device_pgprot); 668 669 mem_types[MT_LOW_VECTORS].prot_l1 |= ecc_mask; 670 mem_types[MT_HIGH_VECTORS].prot_l1 |= ecc_mask; 671 mem_types[MT_MEMORY_RWX].prot_sect |= ecc_mask | cp->pmd; 672 mem_types[MT_MEMORY_RWX].prot_pte |= kern_pgprot; 673 mem_types[MT_MEMORY_RW].prot_sect |= ecc_mask | cp->pmd; 674 mem_types[MT_MEMORY_RW].prot_pte |= kern_pgprot; 675 mem_types[MT_MEMORY_DMA_READY].prot_pte |= kern_pgprot; 676 mem_types[MT_MEMORY_RWX_NONCACHED].prot_sect |= ecc_mask; 677 mem_types[MT_ROM].prot_sect |= cp->pmd; 678 679 switch (cp->pmd) { 680 case PMD_SECT_WT: 681 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WT; 682 break; 683 case PMD_SECT_WB: 684 case PMD_SECT_WBWA: 685 mem_types[MT_CACHECLEAN].prot_sect |= PMD_SECT_WB; 686 break; 687 } 688 pr_info("Memory policy: %sData cache %s\n", 689 ecc_mask ? "ECC enabled, " : "", cp->policy); 690 691 for (i = 0; i < ARRAY_SIZE(mem_types); i++) { 692 struct mem_type *t = &mem_types[i]; 693 if (t->prot_l1) 694 t->prot_l1 |= PMD_DOMAIN(t->domain); 695 if (t->prot_sect) 696 t->prot_sect |= PMD_DOMAIN(t->domain); 697 } 698 } 699 700 #ifdef CONFIG_ARM_DMA_MEM_BUFFERABLE 701 pgprot_t phys_mem_access_prot(struct file *file, unsigned long pfn, 702 unsigned long size, pgprot_t vma_prot) 703 { 704 if (!pfn_valid(pfn)) 705 return pgprot_noncached(vma_prot); 706 else if (file->f_flags & O_SYNC) 707 return pgprot_writecombine(vma_prot); 708 return vma_prot; 709 } 710 EXPORT_SYMBOL(phys_mem_access_prot); 711 #endif 712 713 #define vectors_base() (vectors_high() ? 0xffff0000 : 0) 714 715 static void __init *early_alloc_aligned(unsigned long sz, unsigned long align) 716 { 717 void *ptr = __va(memblock_alloc(sz, align)); 718 memset(ptr, 0, sz); 719 return ptr; 720 } 721 722 static void __init *early_alloc(unsigned long sz) 723 { 724 return early_alloc_aligned(sz, sz); 725 } 726 727 static pte_t * __init early_pte_alloc(pmd_t *pmd, unsigned long addr, unsigned long prot) 728 { 729 if (pmd_none(*pmd)) { 730 pte_t *pte = early_alloc(PTE_HWTABLE_OFF + PTE_HWTABLE_SIZE); 731 __pmd_populate(pmd, __pa(pte), prot); 732 } 733 BUG_ON(pmd_bad(*pmd)); 734 return pte_offset_kernel(pmd, addr); 735 } 736 737 static void __init alloc_init_pte(pmd_t *pmd, unsigned long addr, 738 unsigned long end, unsigned long pfn, 739 const struct mem_type *type) 740 { 741 pte_t *pte = early_pte_alloc(pmd, addr, type->prot_l1); 742 do { 743 set_pte_ext(pte, pfn_pte(pfn, __pgprot(type->prot_pte)), 0); 744 pfn++; 745 } while (pte++, addr += PAGE_SIZE, addr != end); 746 } 747 748 static void __init __map_init_section(pmd_t *pmd, unsigned long addr, 749 unsigned long end, phys_addr_t phys, 750 const struct mem_type *type) 751 { 752 pmd_t *p = pmd; 753 754 #ifndef CONFIG_ARM_LPAE 755 /* 756 * In classic MMU format, puds and pmds are folded in to 757 * the pgds. pmd_offset gives the PGD entry. PGDs refer to a 758 * group of L1 entries making up one logical pointer to 759 * an L2 table (2MB), where as PMDs refer to the individual 760 * L1 entries (1MB). Hence increment to get the correct 761 * offset for odd 1MB sections. 762 * (See arch/arm/include/asm/pgtable-2level.h) 763 */ 764 if (addr & SECTION_SIZE) 765 pmd++; 766 #endif 767 do { 768 *pmd = __pmd(phys | type->prot_sect); 769 phys += SECTION_SIZE; 770 } while (pmd++, addr += SECTION_SIZE, addr != end); 771 772 flush_pmd_entry(p); 773 } 774 775 static void __init alloc_init_pmd(pud_t *pud, unsigned long addr, 776 unsigned long end, phys_addr_t phys, 777 const struct mem_type *type) 778 { 779 pmd_t *pmd = pmd_offset(pud, addr); 780 unsigned long next; 781 782 do { 783 /* 784 * With LPAE, we must loop over to map 785 * all the pmds for the given range. 786 */ 787 next = pmd_addr_end(addr, end); 788 789 /* 790 * Try a section mapping - addr, next and phys must all be 791 * aligned to a section boundary. 792 */ 793 if (type->prot_sect && 794 ((addr | next | phys) & ~SECTION_MASK) == 0) { 795 __map_init_section(pmd, addr, next, phys, type); 796 } else { 797 alloc_init_pte(pmd, addr, next, 798 __phys_to_pfn(phys), type); 799 } 800 801 phys += next - addr; 802 803 } while (pmd++, addr = next, addr != end); 804 } 805 806 static void __init alloc_init_pud(pgd_t *pgd, unsigned long addr, 807 unsigned long end, phys_addr_t phys, 808 const struct mem_type *type) 809 { 810 pud_t *pud = pud_offset(pgd, addr); 811 unsigned long next; 812 813 do { 814 next = pud_addr_end(addr, end); 815 alloc_init_pmd(pud, addr, next, phys, type); 816 phys += next - addr; 817 } while (pud++, addr = next, addr != end); 818 } 819 820 #ifndef CONFIG_ARM_LPAE 821 static void __init create_36bit_mapping(struct map_desc *md, 822 const struct mem_type *type) 823 { 824 unsigned long addr, length, end; 825 phys_addr_t phys; 826 pgd_t *pgd; 827 828 addr = md->virtual; 829 phys = __pfn_to_phys(md->pfn); 830 length = PAGE_ALIGN(md->length); 831 832 if (!(cpu_architecture() >= CPU_ARCH_ARMv6 || cpu_is_xsc3())) { 833 pr_err("MM: CPU does not support supersection mapping for 0x%08llx at 0x%08lx\n", 834 (long long)__pfn_to_phys((u64)md->pfn), addr); 835 return; 836 } 837 838 /* N.B. ARMv6 supersections are only defined to work with domain 0. 839 * Since domain assignments can in fact be arbitrary, the 840 * 'domain == 0' check below is required to insure that ARMv6 841 * supersections are only allocated for domain 0 regardless 842 * of the actual domain assignments in use. 843 */ 844 if (type->domain) { 845 pr_err("MM: invalid domain in supersection mapping for 0x%08llx at 0x%08lx\n", 846 (long long)__pfn_to_phys((u64)md->pfn), addr); 847 return; 848 } 849 850 if ((addr | length | __pfn_to_phys(md->pfn)) & ~SUPERSECTION_MASK) { 851 pr_err("MM: cannot create mapping for 0x%08llx at 0x%08lx invalid alignment\n", 852 (long long)__pfn_to_phys((u64)md->pfn), addr); 853 return; 854 } 855 856 /* 857 * Shift bits [35:32] of address into bits [23:20] of PMD 858 * (See ARMv6 spec). 859 */ 860 phys |= (((md->pfn >> (32 - PAGE_SHIFT)) & 0xF) << 20); 861 862 pgd = pgd_offset_k(addr); 863 end = addr + length; 864 do { 865 pud_t *pud = pud_offset(pgd, addr); 866 pmd_t *pmd = pmd_offset(pud, addr); 867 int i; 868 869 for (i = 0; i < 16; i++) 870 *pmd++ = __pmd(phys | type->prot_sect | PMD_SECT_SUPER); 871 872 addr += SUPERSECTION_SIZE; 873 phys += SUPERSECTION_SIZE; 874 pgd += SUPERSECTION_SIZE >> PGDIR_SHIFT; 875 } while (addr != end); 876 } 877 #endif /* !CONFIG_ARM_LPAE */ 878 879 /* 880 * Create the page directory entries and any necessary 881 * page tables for the mapping specified by `md'. We 882 * are able to cope here with varying sizes and address 883 * offsets, and we take full advantage of sections and 884 * supersections. 885 */ 886 static void __init create_mapping(struct map_desc *md) 887 { 888 unsigned long addr, length, end; 889 phys_addr_t phys; 890 const struct mem_type *type; 891 pgd_t *pgd; 892 893 if (md->virtual != vectors_base() && md->virtual < TASK_SIZE) { 894 pr_warn("BUG: not creating mapping for 0x%08llx at 0x%08lx in user region\n", 895 (long long)__pfn_to_phys((u64)md->pfn), md->virtual); 896 return; 897 } 898 899 if ((md->type == MT_DEVICE || md->type == MT_ROM) && 900 md->virtual >= PAGE_OFFSET && md->virtual < FIXADDR_START && 901 (md->virtual < VMALLOC_START || md->virtual >= VMALLOC_END)) { 902 pr_warn("BUG: mapping for 0x%08llx at 0x%08lx out of vmalloc space\n", 903 (long long)__pfn_to_phys((u64)md->pfn), md->virtual); 904 } 905 906 type = &mem_types[md->type]; 907 908 #ifndef CONFIG_ARM_LPAE 909 /* 910 * Catch 36-bit addresses 911 */ 912 if (md->pfn >= 0x100000) { 913 create_36bit_mapping(md, type); 914 return; 915 } 916 #endif 917 918 addr = md->virtual & PAGE_MASK; 919 phys = __pfn_to_phys(md->pfn); 920 length = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK)); 921 922 if (type->prot_l1 == 0 && ((addr | phys | length) & ~SECTION_MASK)) { 923 pr_warn("BUG: map for 0x%08llx at 0x%08lx can not be mapped using pages, ignoring.\n", 924 (long long)__pfn_to_phys(md->pfn), addr); 925 return; 926 } 927 928 pgd = pgd_offset_k(addr); 929 end = addr + length; 930 do { 931 unsigned long next = pgd_addr_end(addr, end); 932 933 alloc_init_pud(pgd, addr, next, phys, type); 934 935 phys += next - addr; 936 addr = next; 937 } while (pgd++, addr != end); 938 } 939 940 /* 941 * Create the architecture specific mappings 942 */ 943 void __init iotable_init(struct map_desc *io_desc, int nr) 944 { 945 struct map_desc *md; 946 struct vm_struct *vm; 947 struct static_vm *svm; 948 949 if (!nr) 950 return; 951 952 svm = early_alloc_aligned(sizeof(*svm) * nr, __alignof__(*svm)); 953 954 for (md = io_desc; nr; md++, nr--) { 955 create_mapping(md); 956 957 vm = &svm->vm; 958 vm->addr = (void *)(md->virtual & PAGE_MASK); 959 vm->size = PAGE_ALIGN(md->length + (md->virtual & ~PAGE_MASK)); 960 vm->phys_addr = __pfn_to_phys(md->pfn); 961 vm->flags = VM_IOREMAP | VM_ARM_STATIC_MAPPING; 962 vm->flags |= VM_ARM_MTYPE(md->type); 963 vm->caller = iotable_init; 964 add_static_vm_early(svm++); 965 } 966 } 967 968 void __init vm_reserve_area_early(unsigned long addr, unsigned long size, 969 void *caller) 970 { 971 struct vm_struct *vm; 972 struct static_vm *svm; 973 974 svm = early_alloc_aligned(sizeof(*svm), __alignof__(*svm)); 975 976 vm = &svm->vm; 977 vm->addr = (void *)addr; 978 vm->size = size; 979 vm->flags = VM_IOREMAP | VM_ARM_EMPTY_MAPPING; 980 vm->caller = caller; 981 add_static_vm_early(svm); 982 } 983 984 #ifndef CONFIG_ARM_LPAE 985 986 /* 987 * The Linux PMD is made of two consecutive section entries covering 2MB 988 * (see definition in include/asm/pgtable-2level.h). However a call to 989 * create_mapping() may optimize static mappings by using individual 990 * 1MB section mappings. This leaves the actual PMD potentially half 991 * initialized if the top or bottom section entry isn't used, leaving it 992 * open to problems if a subsequent ioremap() or vmalloc() tries to use 993 * the virtual space left free by that unused section entry. 994 * 995 * Let's avoid the issue by inserting dummy vm entries covering the unused 996 * PMD halves once the static mappings are in place. 997 */ 998 999 static void __init pmd_empty_section_gap(unsigned long addr) 1000 { 1001 vm_reserve_area_early(addr, SECTION_SIZE, pmd_empty_section_gap); 1002 } 1003 1004 static void __init fill_pmd_gaps(void) 1005 { 1006 struct static_vm *svm; 1007 struct vm_struct *vm; 1008 unsigned long addr, next = 0; 1009 pmd_t *pmd; 1010 1011 list_for_each_entry(svm, &static_vmlist, list) { 1012 vm = &svm->vm; 1013 addr = (unsigned long)vm->addr; 1014 if (addr < next) 1015 continue; 1016 1017 /* 1018 * Check if this vm starts on an odd section boundary. 1019 * If so and the first section entry for this PMD is free 1020 * then we block the corresponding virtual address. 1021 */ 1022 if ((addr & ~PMD_MASK) == SECTION_SIZE) { 1023 pmd = pmd_off_k(addr); 1024 if (pmd_none(*pmd)) 1025 pmd_empty_section_gap(addr & PMD_MASK); 1026 } 1027 1028 /* 1029 * Then check if this vm ends on an odd section boundary. 1030 * If so and the second section entry for this PMD is empty 1031 * then we block the corresponding virtual address. 1032 */ 1033 addr += vm->size; 1034 if ((addr & ~PMD_MASK) == SECTION_SIZE) { 1035 pmd = pmd_off_k(addr) + 1; 1036 if (pmd_none(*pmd)) 1037 pmd_empty_section_gap(addr); 1038 } 1039 1040 /* no need to look at any vm entry until we hit the next PMD */ 1041 next = (addr + PMD_SIZE - 1) & PMD_MASK; 1042 } 1043 } 1044 1045 #else 1046 #define fill_pmd_gaps() do { } while (0) 1047 #endif 1048 1049 #if defined(CONFIG_PCI) && !defined(CONFIG_NEED_MACH_IO_H) 1050 static void __init pci_reserve_io(void) 1051 { 1052 struct static_vm *svm; 1053 1054 svm = find_static_vm_vaddr((void *)PCI_IO_VIRT_BASE); 1055 if (svm) 1056 return; 1057 1058 vm_reserve_area_early(PCI_IO_VIRT_BASE, SZ_2M, pci_reserve_io); 1059 } 1060 #else 1061 #define pci_reserve_io() do { } while (0) 1062 #endif 1063 1064 #ifdef CONFIG_DEBUG_LL 1065 void __init debug_ll_io_init(void) 1066 { 1067 struct map_desc map; 1068 1069 debug_ll_addr(&map.pfn, &map.virtual); 1070 if (!map.pfn || !map.virtual) 1071 return; 1072 map.pfn = __phys_to_pfn(map.pfn); 1073 map.virtual &= PAGE_MASK; 1074 map.length = PAGE_SIZE; 1075 map.type = MT_DEVICE; 1076 iotable_init(&map, 1); 1077 } 1078 #endif 1079 1080 static void * __initdata vmalloc_min = 1081 (void *)(VMALLOC_END - (240 << 20) - VMALLOC_OFFSET); 1082 1083 /* 1084 * vmalloc=size forces the vmalloc area to be exactly 'size' 1085 * bytes. This can be used to increase (or decrease) the vmalloc 1086 * area - the default is 240m. 1087 */ 1088 static int __init early_vmalloc(char *arg) 1089 { 1090 unsigned long vmalloc_reserve = memparse(arg, NULL); 1091 1092 if (vmalloc_reserve < SZ_16M) { 1093 vmalloc_reserve = SZ_16M; 1094 pr_warn("vmalloc area too small, limiting to %luMB\n", 1095 vmalloc_reserve >> 20); 1096 } 1097 1098 if (vmalloc_reserve > VMALLOC_END - (PAGE_OFFSET + SZ_32M)) { 1099 vmalloc_reserve = VMALLOC_END - (PAGE_OFFSET + SZ_32M); 1100 pr_warn("vmalloc area is too big, limiting to %luMB\n", 1101 vmalloc_reserve >> 20); 1102 } 1103 1104 vmalloc_min = (void *)(VMALLOC_END - vmalloc_reserve); 1105 return 0; 1106 } 1107 early_param("vmalloc", early_vmalloc); 1108 1109 phys_addr_t arm_lowmem_limit __initdata = 0; 1110 1111 void __init sanity_check_meminfo(void) 1112 { 1113 phys_addr_t memblock_limit = 0; 1114 int highmem = 0; 1115 phys_addr_t vmalloc_limit = __pa(vmalloc_min - 1) + 1; 1116 struct memblock_region *reg; 1117 bool should_use_highmem = false; 1118 1119 for_each_memblock(memory, reg) { 1120 phys_addr_t block_start = reg->base; 1121 phys_addr_t block_end = reg->base + reg->size; 1122 phys_addr_t size_limit = reg->size; 1123 1124 if (reg->base >= vmalloc_limit) 1125 highmem = 1; 1126 else 1127 size_limit = vmalloc_limit - reg->base; 1128 1129 1130 if (!IS_ENABLED(CONFIG_HIGHMEM) || cache_is_vipt_aliasing()) { 1131 1132 if (highmem) { 1133 pr_notice("Ignoring RAM at %pa-%pa (!CONFIG_HIGHMEM)\n", 1134 &block_start, &block_end); 1135 memblock_remove(reg->base, reg->size); 1136 should_use_highmem = true; 1137 continue; 1138 } 1139 1140 if (reg->size > size_limit) { 1141 phys_addr_t overlap_size = reg->size - size_limit; 1142 1143 pr_notice("Truncating RAM at %pa-%pa to -%pa", 1144 &block_start, &block_end, &vmalloc_limit); 1145 memblock_remove(vmalloc_limit, overlap_size); 1146 block_end = vmalloc_limit; 1147 should_use_highmem = true; 1148 } 1149 } 1150 1151 if (!highmem) { 1152 if (block_end > arm_lowmem_limit) { 1153 if (reg->size > size_limit) 1154 arm_lowmem_limit = vmalloc_limit; 1155 else 1156 arm_lowmem_limit = block_end; 1157 } 1158 1159 /* 1160 * Find the first non-pmd-aligned page, and point 1161 * memblock_limit at it. This relies on rounding the 1162 * limit down to be pmd-aligned, which happens at the 1163 * end of this function. 1164 * 1165 * With this algorithm, the start or end of almost any 1166 * bank can be non-pmd-aligned. The only exception is 1167 * that the start of the bank 0 must be section- 1168 * aligned, since otherwise memory would need to be 1169 * allocated when mapping the start of bank 0, which 1170 * occurs before any free memory is mapped. 1171 */ 1172 if (!memblock_limit) { 1173 if (!IS_ALIGNED(block_start, PMD_SIZE)) 1174 memblock_limit = block_start; 1175 else if (!IS_ALIGNED(block_end, PMD_SIZE)) 1176 memblock_limit = arm_lowmem_limit; 1177 } 1178 1179 } 1180 } 1181 1182 if (should_use_highmem) 1183 pr_notice("Consider using a HIGHMEM enabled kernel.\n"); 1184 1185 high_memory = __va(arm_lowmem_limit - 1) + 1; 1186 1187 /* 1188 * Round the memblock limit down to a pmd size. This 1189 * helps to ensure that we will allocate memory from the 1190 * last full pmd, which should be mapped. 1191 */ 1192 if (memblock_limit) 1193 memblock_limit = round_down(memblock_limit, PMD_SIZE); 1194 if (!memblock_limit) 1195 memblock_limit = arm_lowmem_limit; 1196 1197 memblock_set_current_limit(memblock_limit); 1198 } 1199 1200 static inline void prepare_page_table(void) 1201 { 1202 unsigned long addr; 1203 phys_addr_t end; 1204 1205 /* 1206 * Clear out all the mappings below the kernel image. 1207 */ 1208 for (addr = 0; addr < MODULES_VADDR; addr += PMD_SIZE) 1209 pmd_clear(pmd_off_k(addr)); 1210 1211 #ifdef CONFIG_XIP_KERNEL 1212 /* The XIP kernel is mapped in the module area -- skip over it */ 1213 addr = ((unsigned long)_etext + PMD_SIZE - 1) & PMD_MASK; 1214 #endif 1215 for ( ; addr < PAGE_OFFSET; addr += PMD_SIZE) 1216 pmd_clear(pmd_off_k(addr)); 1217 1218 /* 1219 * Find the end of the first block of lowmem. 1220 */ 1221 end = memblock.memory.regions[0].base + memblock.memory.regions[0].size; 1222 if (end >= arm_lowmem_limit) 1223 end = arm_lowmem_limit; 1224 1225 /* 1226 * Clear out all the kernel space mappings, except for the first 1227 * memory bank, up to the vmalloc region. 1228 */ 1229 for (addr = __phys_to_virt(end); 1230 addr < VMALLOC_START; addr += PMD_SIZE) 1231 pmd_clear(pmd_off_k(addr)); 1232 } 1233 1234 #ifdef CONFIG_ARM_LPAE 1235 /* the first page is reserved for pgd */ 1236 #define SWAPPER_PG_DIR_SIZE (PAGE_SIZE + \ 1237 PTRS_PER_PGD * PTRS_PER_PMD * sizeof(pmd_t)) 1238 #else 1239 #define SWAPPER_PG_DIR_SIZE (PTRS_PER_PGD * sizeof(pgd_t)) 1240 #endif 1241 1242 /* 1243 * Reserve the special regions of memory 1244 */ 1245 void __init arm_mm_memblock_reserve(void) 1246 { 1247 /* 1248 * Reserve the page tables. These are already in use, 1249 * and can only be in node 0. 1250 */ 1251 memblock_reserve(__pa(swapper_pg_dir), SWAPPER_PG_DIR_SIZE); 1252 1253 #ifdef CONFIG_SA1111 1254 /* 1255 * Because of the SA1111 DMA bug, we want to preserve our 1256 * precious DMA-able memory... 1257 */ 1258 memblock_reserve(PHYS_OFFSET, __pa(swapper_pg_dir) - PHYS_OFFSET); 1259 #endif 1260 } 1261 1262 /* 1263 * Set up the device mappings. Since we clear out the page tables for all 1264 * mappings above VMALLOC_START, except early fixmap, we might remove debug 1265 * device mappings. This means earlycon can be used to debug this function 1266 * Any other function or debugging method which may touch any device _will_ 1267 * crash the kernel. 1268 */ 1269 static void __init devicemaps_init(const struct machine_desc *mdesc) 1270 { 1271 struct map_desc map; 1272 unsigned long addr; 1273 void *vectors; 1274 1275 /* 1276 * Allocate the vector page early. 1277 */ 1278 vectors = early_alloc(PAGE_SIZE * 2); 1279 1280 early_trap_init(vectors); 1281 1282 /* 1283 * Clear page table except top pmd used by early fixmaps 1284 */ 1285 for (addr = VMALLOC_START; addr < (FIXADDR_TOP & PMD_MASK); addr += PMD_SIZE) 1286 pmd_clear(pmd_off_k(addr)); 1287 1288 /* 1289 * Map the kernel if it is XIP. 1290 * It is always first in the modulearea. 1291 */ 1292 #ifdef CONFIG_XIP_KERNEL 1293 map.pfn = __phys_to_pfn(CONFIG_XIP_PHYS_ADDR & SECTION_MASK); 1294 map.virtual = MODULES_VADDR; 1295 map.length = ((unsigned long)_etext - map.virtual + ~SECTION_MASK) & SECTION_MASK; 1296 map.type = MT_ROM; 1297 create_mapping(&map); 1298 #endif 1299 1300 /* 1301 * Map the cache flushing regions. 1302 */ 1303 #ifdef FLUSH_BASE 1304 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS); 1305 map.virtual = FLUSH_BASE; 1306 map.length = SZ_1M; 1307 map.type = MT_CACHECLEAN; 1308 create_mapping(&map); 1309 #endif 1310 #ifdef FLUSH_BASE_MINICACHE 1311 map.pfn = __phys_to_pfn(FLUSH_BASE_PHYS + SZ_1M); 1312 map.virtual = FLUSH_BASE_MINICACHE; 1313 map.length = SZ_1M; 1314 map.type = MT_MINICLEAN; 1315 create_mapping(&map); 1316 #endif 1317 1318 /* 1319 * Create a mapping for the machine vectors at the high-vectors 1320 * location (0xffff0000). If we aren't using high-vectors, also 1321 * create a mapping at the low-vectors virtual address. 1322 */ 1323 map.pfn = __phys_to_pfn(virt_to_phys(vectors)); 1324 map.virtual = 0xffff0000; 1325 map.length = PAGE_SIZE; 1326 #ifdef CONFIG_KUSER_HELPERS 1327 map.type = MT_HIGH_VECTORS; 1328 #else 1329 map.type = MT_LOW_VECTORS; 1330 #endif 1331 create_mapping(&map); 1332 1333 if (!vectors_high()) { 1334 map.virtual = 0; 1335 map.length = PAGE_SIZE * 2; 1336 map.type = MT_LOW_VECTORS; 1337 create_mapping(&map); 1338 } 1339 1340 /* Now create a kernel read-only mapping */ 1341 map.pfn += 1; 1342 map.virtual = 0xffff0000 + PAGE_SIZE; 1343 map.length = PAGE_SIZE; 1344 map.type = MT_LOW_VECTORS; 1345 create_mapping(&map); 1346 1347 /* 1348 * Ask the machine support to map in the statically mapped devices. 1349 */ 1350 if (mdesc->map_io) 1351 mdesc->map_io(); 1352 else 1353 debug_ll_io_init(); 1354 fill_pmd_gaps(); 1355 1356 /* Reserve fixed i/o space in VMALLOC region */ 1357 pci_reserve_io(); 1358 1359 /* 1360 * Finally flush the caches and tlb to ensure that we're in a 1361 * consistent state wrt the writebuffer. This also ensures that 1362 * any write-allocated cache lines in the vector page are written 1363 * back. After this point, we can start to touch devices again. 1364 */ 1365 local_flush_tlb_all(); 1366 flush_cache_all(); 1367 1368 /* Enable asynchronous aborts */ 1369 early_abt_enable(); 1370 } 1371 1372 static void __init kmap_init(void) 1373 { 1374 #ifdef CONFIG_HIGHMEM 1375 pkmap_page_table = early_pte_alloc(pmd_off_k(PKMAP_BASE), 1376 PKMAP_BASE, _PAGE_KERNEL_TABLE); 1377 #endif 1378 1379 early_pte_alloc(pmd_off_k(FIXADDR_START), FIXADDR_START, 1380 _PAGE_KERNEL_TABLE); 1381 } 1382 1383 static void __init map_lowmem(void) 1384 { 1385 struct memblock_region *reg; 1386 phys_addr_t kernel_x_start = round_down(__pa(_stext), SECTION_SIZE); 1387 phys_addr_t kernel_x_end = round_up(__pa(__init_end), SECTION_SIZE); 1388 1389 /* Map all the lowmem memory banks. */ 1390 for_each_memblock(memory, reg) { 1391 phys_addr_t start = reg->base; 1392 phys_addr_t end = start + reg->size; 1393 struct map_desc map; 1394 1395 if (end > arm_lowmem_limit) 1396 end = arm_lowmem_limit; 1397 if (start >= end) 1398 break; 1399 1400 if (end < kernel_x_start) { 1401 map.pfn = __phys_to_pfn(start); 1402 map.virtual = __phys_to_virt(start); 1403 map.length = end - start; 1404 map.type = MT_MEMORY_RWX; 1405 1406 create_mapping(&map); 1407 } else if (start >= kernel_x_end) { 1408 map.pfn = __phys_to_pfn(start); 1409 map.virtual = __phys_to_virt(start); 1410 map.length = end - start; 1411 map.type = MT_MEMORY_RW; 1412 1413 create_mapping(&map); 1414 } else { 1415 /* This better cover the entire kernel */ 1416 if (start < kernel_x_start) { 1417 map.pfn = __phys_to_pfn(start); 1418 map.virtual = __phys_to_virt(start); 1419 map.length = kernel_x_start - start; 1420 map.type = MT_MEMORY_RW; 1421 1422 create_mapping(&map); 1423 } 1424 1425 map.pfn = __phys_to_pfn(kernel_x_start); 1426 map.virtual = __phys_to_virt(kernel_x_start); 1427 map.length = kernel_x_end - kernel_x_start; 1428 map.type = MT_MEMORY_RWX; 1429 1430 create_mapping(&map); 1431 1432 if (kernel_x_end < end) { 1433 map.pfn = __phys_to_pfn(kernel_x_end); 1434 map.virtual = __phys_to_virt(kernel_x_end); 1435 map.length = end - kernel_x_end; 1436 map.type = MT_MEMORY_RW; 1437 1438 create_mapping(&map); 1439 } 1440 } 1441 } 1442 } 1443 1444 #ifdef CONFIG_ARM_PV_FIXUP 1445 extern unsigned long __atags_pointer; 1446 typedef void pgtables_remap(long long offset, unsigned long pgd, void *bdata); 1447 pgtables_remap lpae_pgtables_remap_asm; 1448 1449 /* 1450 * early_paging_init() recreates boot time page table setup, allowing machines 1451 * to switch over to a high (>4G) address space on LPAE systems 1452 */ 1453 void __init early_paging_init(const struct machine_desc *mdesc) 1454 { 1455 pgtables_remap *lpae_pgtables_remap; 1456 unsigned long pa_pgd; 1457 unsigned int cr, ttbcr; 1458 long long offset; 1459 void *boot_data; 1460 1461 if (!mdesc->pv_fixup) 1462 return; 1463 1464 offset = mdesc->pv_fixup(); 1465 if (offset == 0) 1466 return; 1467 1468 /* 1469 * Get the address of the remap function in the 1:1 identity 1470 * mapping setup by the early page table assembly code. We 1471 * must get this prior to the pv update. The following barrier 1472 * ensures that this is complete before we fixup any P:V offsets. 1473 */ 1474 lpae_pgtables_remap = (pgtables_remap *)(unsigned long)__pa(lpae_pgtables_remap_asm); 1475 pa_pgd = __pa(swapper_pg_dir); 1476 boot_data = __va(__atags_pointer); 1477 barrier(); 1478 1479 pr_info("Switching physical address space to 0x%08llx\n", 1480 (u64)PHYS_OFFSET + offset); 1481 1482 /* Re-set the phys pfn offset, and the pv offset */ 1483 __pv_offset += offset; 1484 __pv_phys_pfn_offset += PFN_DOWN(offset); 1485 1486 /* Run the patch stub to update the constants */ 1487 fixup_pv_table(&__pv_table_begin, 1488 (&__pv_table_end - &__pv_table_begin) << 2); 1489 1490 /* 1491 * We changing not only the virtual to physical mapping, but also 1492 * the physical addresses used to access memory. We need to flush 1493 * all levels of cache in the system with caching disabled to 1494 * ensure that all data is written back, and nothing is prefetched 1495 * into the caches. We also need to prevent the TLB walkers 1496 * allocating into the caches too. Note that this is ARMv7 LPAE 1497 * specific. 1498 */ 1499 cr = get_cr(); 1500 set_cr(cr & ~(CR_I | CR_C)); 1501 asm("mrc p15, 0, %0, c2, c0, 2" : "=r" (ttbcr)); 1502 asm volatile("mcr p15, 0, %0, c2, c0, 2" 1503 : : "r" (ttbcr & ~(3 << 8 | 3 << 10))); 1504 flush_cache_all(); 1505 1506 /* 1507 * Fixup the page tables - this must be in the idmap region as 1508 * we need to disable the MMU to do this safely, and hence it 1509 * needs to be assembly. It's fairly simple, as we're using the 1510 * temporary tables setup by the initial assembly code. 1511 */ 1512 lpae_pgtables_remap(offset, pa_pgd, boot_data); 1513 1514 /* Re-enable the caches and cacheable TLB walks */ 1515 asm volatile("mcr p15, 0, %0, c2, c0, 2" : : "r" (ttbcr)); 1516 set_cr(cr); 1517 } 1518 1519 #else 1520 1521 void __init early_paging_init(const struct machine_desc *mdesc) 1522 { 1523 long long offset; 1524 1525 if (!mdesc->pv_fixup) 1526 return; 1527 1528 offset = mdesc->pv_fixup(); 1529 if (offset == 0) 1530 return; 1531 1532 pr_crit("Physical address space modification is only to support Keystone2.\n"); 1533 pr_crit("Please enable ARM_LPAE and ARM_PATCH_PHYS_VIRT support to use this\n"); 1534 pr_crit("feature. Your kernel may crash now, have a good day.\n"); 1535 add_taint(TAINT_CPU_OUT_OF_SPEC, LOCKDEP_STILL_OK); 1536 } 1537 1538 #endif 1539 1540 static void __init early_fixmap_shutdown(void) 1541 { 1542 int i; 1543 unsigned long va = fix_to_virt(__end_of_permanent_fixed_addresses - 1); 1544 1545 pte_offset_fixmap = pte_offset_late_fixmap; 1546 pmd_clear(fixmap_pmd(va)); 1547 local_flush_tlb_kernel_page(va); 1548 1549 for (i = 0; i < __end_of_permanent_fixed_addresses; i++) { 1550 pte_t *pte; 1551 struct map_desc map; 1552 1553 map.virtual = fix_to_virt(i); 1554 pte = pte_offset_early_fixmap(pmd_off_k(map.virtual), map.virtual); 1555 1556 /* Only i/o device mappings are supported ATM */ 1557 if (pte_none(*pte) || 1558 (pte_val(*pte) & L_PTE_MT_MASK) != L_PTE_MT_DEV_SHARED) 1559 continue; 1560 1561 map.pfn = pte_pfn(*pte); 1562 map.type = MT_DEVICE; 1563 map.length = PAGE_SIZE; 1564 1565 create_mapping(&map); 1566 } 1567 } 1568 1569 /* 1570 * paging_init() sets up the page tables, initialises the zone memory 1571 * maps, and sets up the zero page, bad page and bad page tables. 1572 */ 1573 void __init paging_init(const struct machine_desc *mdesc) 1574 { 1575 void *zero_page; 1576 1577 build_mem_type_table(); 1578 prepare_page_table(); 1579 map_lowmem(); 1580 memblock_set_current_limit(arm_lowmem_limit); 1581 dma_contiguous_remap(); 1582 early_fixmap_shutdown(); 1583 devicemaps_init(mdesc); 1584 kmap_init(); 1585 tcm_init(); 1586 1587 top_pmd = pmd_off_k(0xffff0000); 1588 1589 /* allocate the zero page. */ 1590 zero_page = early_alloc(PAGE_SIZE); 1591 1592 bootmem_init(); 1593 1594 empty_zero_page = virt_to_page(zero_page); 1595 __flush_dcache_page(NULL, empty_zero_page); 1596 } 1597