1 #include <linux/gfp.h> 2 #include <linux/initrd.h> 3 #include <linux/ioport.h> 4 #include <linux/swap.h> 5 #include <linux/memblock.h> 6 #include <linux/bootmem.h> /* for max_low_pfn */ 7 8 #include <asm/set_memory.h> 9 #include <asm/e820/api.h> 10 #include <asm/init.h> 11 #include <asm/page.h> 12 #include <asm/page_types.h> 13 #include <asm/sections.h> 14 #include <asm/setup.h> 15 #include <asm/tlbflush.h> 16 #include <asm/tlb.h> 17 #include <asm/proto.h> 18 #include <asm/dma.h> /* for MAX_DMA_PFN */ 19 #include <asm/microcode.h> 20 #include <asm/kaslr.h> 21 #include <asm/hypervisor.h> 22 23 /* 24 * We need to define the tracepoints somewhere, and tlb.c 25 * is only compied when SMP=y. 26 */ 27 #define CREATE_TRACE_POINTS 28 #include <trace/events/tlb.h> 29 30 #include "mm_internal.h" 31 32 /* 33 * Tables translating between page_cache_type_t and pte encoding. 34 * 35 * The default values are defined statically as minimal supported mode; 36 * WC and WT fall back to UC-. pat_init() updates these values to support 37 * more cache modes, WC and WT, when it is safe to do so. See pat_init() 38 * for the details. Note, __early_ioremap() used during early boot-time 39 * takes pgprot_t (pte encoding) and does not use these tables. 40 * 41 * Index into __cachemode2pte_tbl[] is the cachemode. 42 * 43 * Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte 44 * (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2. 45 */ 46 uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = { 47 [_PAGE_CACHE_MODE_WB ] = 0 | 0 , 48 [_PAGE_CACHE_MODE_WC ] = 0 | _PAGE_PCD, 49 [_PAGE_CACHE_MODE_UC_MINUS] = 0 | _PAGE_PCD, 50 [_PAGE_CACHE_MODE_UC ] = _PAGE_PWT | _PAGE_PCD, 51 [_PAGE_CACHE_MODE_WT ] = 0 | _PAGE_PCD, 52 [_PAGE_CACHE_MODE_WP ] = 0 | _PAGE_PCD, 53 }; 54 EXPORT_SYMBOL(__cachemode2pte_tbl); 55 56 uint8_t __pte2cachemode_tbl[8] = { 57 [__pte2cm_idx( 0 | 0 | 0 )] = _PAGE_CACHE_MODE_WB, 58 [__pte2cm_idx(_PAGE_PWT | 0 | 0 )] = _PAGE_CACHE_MODE_UC_MINUS, 59 [__pte2cm_idx( 0 | _PAGE_PCD | 0 )] = _PAGE_CACHE_MODE_UC_MINUS, 60 [__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0 )] = _PAGE_CACHE_MODE_UC, 61 [__pte2cm_idx( 0 | 0 | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB, 62 [__pte2cm_idx(_PAGE_PWT | 0 | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS, 63 [__pte2cm_idx(0 | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS, 64 [__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC, 65 }; 66 EXPORT_SYMBOL(__pte2cachemode_tbl); 67 68 static unsigned long __initdata pgt_buf_start; 69 static unsigned long __initdata pgt_buf_end; 70 static unsigned long __initdata pgt_buf_top; 71 72 static unsigned long min_pfn_mapped; 73 74 static bool __initdata can_use_brk_pgt = true; 75 76 /* 77 * Pages returned are already directly mapped. 78 * 79 * Changing that is likely to break Xen, see commit: 80 * 81 * 279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve 82 * 83 * for detailed information. 84 */ 85 __ref void *alloc_low_pages(unsigned int num) 86 { 87 unsigned long pfn; 88 int i; 89 90 if (after_bootmem) { 91 unsigned int order; 92 93 order = get_order((unsigned long)num << PAGE_SHIFT); 94 return (void *)__get_free_pages(GFP_ATOMIC | __GFP_NOTRACK | 95 __GFP_ZERO, order); 96 } 97 98 if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) { 99 unsigned long ret; 100 if (min_pfn_mapped >= max_pfn_mapped) 101 panic("alloc_low_pages: ran out of memory"); 102 ret = memblock_find_in_range(min_pfn_mapped << PAGE_SHIFT, 103 max_pfn_mapped << PAGE_SHIFT, 104 PAGE_SIZE * num , PAGE_SIZE); 105 if (!ret) 106 panic("alloc_low_pages: can not alloc memory"); 107 memblock_reserve(ret, PAGE_SIZE * num); 108 pfn = ret >> PAGE_SHIFT; 109 } else { 110 pfn = pgt_buf_end; 111 pgt_buf_end += num; 112 printk(KERN_DEBUG "BRK [%#010lx, %#010lx] PGTABLE\n", 113 pfn << PAGE_SHIFT, (pgt_buf_end << PAGE_SHIFT) - 1); 114 } 115 116 for (i = 0; i < num; i++) { 117 void *adr; 118 119 adr = __va((pfn + i) << PAGE_SHIFT); 120 clear_page(adr); 121 } 122 123 return __va(pfn << PAGE_SHIFT); 124 } 125 126 /* 127 * By default need 3 4k for initial PMD_SIZE, 3 4k for 0-ISA_END_ADDRESS. 128 * With KASLR memory randomization, depending on the machine e820 memory 129 * and the PUD alignment. We may need twice more pages when KASLR memory 130 * randomization is enabled. 131 */ 132 #ifndef CONFIG_RANDOMIZE_MEMORY 133 #define INIT_PGD_PAGE_COUNT 6 134 #else 135 #define INIT_PGD_PAGE_COUNT 12 136 #endif 137 #define INIT_PGT_BUF_SIZE (INIT_PGD_PAGE_COUNT * PAGE_SIZE) 138 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE); 139 void __init early_alloc_pgt_buf(void) 140 { 141 unsigned long tables = INIT_PGT_BUF_SIZE; 142 phys_addr_t base; 143 144 base = __pa(extend_brk(tables, PAGE_SIZE)); 145 146 pgt_buf_start = base >> PAGE_SHIFT; 147 pgt_buf_end = pgt_buf_start; 148 pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT); 149 } 150 151 int after_bootmem; 152 153 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES); 154 155 struct map_range { 156 unsigned long start; 157 unsigned long end; 158 unsigned page_size_mask; 159 }; 160 161 static int page_size_mask; 162 163 static void __init probe_page_size_mask(void) 164 { 165 /* 166 * For CONFIG_KMEMCHECK or pagealloc debugging, identity mapping will 167 * use small pages. 168 * This will simplify cpa(), which otherwise needs to support splitting 169 * large pages into small in interrupt context, etc. 170 */ 171 if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled() && !IS_ENABLED(CONFIG_KMEMCHECK)) 172 page_size_mask |= 1 << PG_LEVEL_2M; 173 else 174 direct_gbpages = 0; 175 176 /* Enable PSE if available */ 177 if (boot_cpu_has(X86_FEATURE_PSE)) 178 cr4_set_bits_and_update_boot(X86_CR4_PSE); 179 180 /* Enable PGE if available */ 181 if (boot_cpu_has(X86_FEATURE_PGE)) { 182 cr4_set_bits_and_update_boot(X86_CR4_PGE); 183 __supported_pte_mask |= _PAGE_GLOBAL; 184 } else 185 __supported_pte_mask &= ~_PAGE_GLOBAL; 186 187 /* Enable 1 GB linear kernel mappings if available: */ 188 if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) { 189 printk(KERN_INFO "Using GB pages for direct mapping\n"); 190 page_size_mask |= 1 << PG_LEVEL_1G; 191 } else { 192 direct_gbpages = 0; 193 } 194 } 195 196 #ifdef CONFIG_X86_32 197 #define NR_RANGE_MR 3 198 #else /* CONFIG_X86_64 */ 199 #define NR_RANGE_MR 5 200 #endif 201 202 static int __meminit save_mr(struct map_range *mr, int nr_range, 203 unsigned long start_pfn, unsigned long end_pfn, 204 unsigned long page_size_mask) 205 { 206 if (start_pfn < end_pfn) { 207 if (nr_range >= NR_RANGE_MR) 208 panic("run out of range for init_memory_mapping\n"); 209 mr[nr_range].start = start_pfn<<PAGE_SHIFT; 210 mr[nr_range].end = end_pfn<<PAGE_SHIFT; 211 mr[nr_range].page_size_mask = page_size_mask; 212 nr_range++; 213 } 214 215 return nr_range; 216 } 217 218 /* 219 * adjust the page_size_mask for small range to go with 220 * big page size instead small one if nearby are ram too. 221 */ 222 static void __ref adjust_range_page_size_mask(struct map_range *mr, 223 int nr_range) 224 { 225 int i; 226 227 for (i = 0; i < nr_range; i++) { 228 if ((page_size_mask & (1<<PG_LEVEL_2M)) && 229 !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) { 230 unsigned long start = round_down(mr[i].start, PMD_SIZE); 231 unsigned long end = round_up(mr[i].end, PMD_SIZE); 232 233 #ifdef CONFIG_X86_32 234 if ((end >> PAGE_SHIFT) > max_low_pfn) 235 continue; 236 #endif 237 238 if (memblock_is_region_memory(start, end - start)) 239 mr[i].page_size_mask |= 1<<PG_LEVEL_2M; 240 } 241 if ((page_size_mask & (1<<PG_LEVEL_1G)) && 242 !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) { 243 unsigned long start = round_down(mr[i].start, PUD_SIZE); 244 unsigned long end = round_up(mr[i].end, PUD_SIZE); 245 246 if (memblock_is_region_memory(start, end - start)) 247 mr[i].page_size_mask |= 1<<PG_LEVEL_1G; 248 } 249 } 250 } 251 252 static const char *page_size_string(struct map_range *mr) 253 { 254 static const char str_1g[] = "1G"; 255 static const char str_2m[] = "2M"; 256 static const char str_4m[] = "4M"; 257 static const char str_4k[] = "4k"; 258 259 if (mr->page_size_mask & (1<<PG_LEVEL_1G)) 260 return str_1g; 261 /* 262 * 32-bit without PAE has a 4M large page size. 263 * PG_LEVEL_2M is misnamed, but we can at least 264 * print out the right size in the string. 265 */ 266 if (IS_ENABLED(CONFIG_X86_32) && 267 !IS_ENABLED(CONFIG_X86_PAE) && 268 mr->page_size_mask & (1<<PG_LEVEL_2M)) 269 return str_4m; 270 271 if (mr->page_size_mask & (1<<PG_LEVEL_2M)) 272 return str_2m; 273 274 return str_4k; 275 } 276 277 static int __meminit split_mem_range(struct map_range *mr, int nr_range, 278 unsigned long start, 279 unsigned long end) 280 { 281 unsigned long start_pfn, end_pfn, limit_pfn; 282 unsigned long pfn; 283 int i; 284 285 limit_pfn = PFN_DOWN(end); 286 287 /* head if not big page alignment ? */ 288 pfn = start_pfn = PFN_DOWN(start); 289 #ifdef CONFIG_X86_32 290 /* 291 * Don't use a large page for the first 2/4MB of memory 292 * because there are often fixed size MTRRs in there 293 * and overlapping MTRRs into large pages can cause 294 * slowdowns. 295 */ 296 if (pfn == 0) 297 end_pfn = PFN_DOWN(PMD_SIZE); 298 else 299 end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 300 #else /* CONFIG_X86_64 */ 301 end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 302 #endif 303 if (end_pfn > limit_pfn) 304 end_pfn = limit_pfn; 305 if (start_pfn < end_pfn) { 306 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0); 307 pfn = end_pfn; 308 } 309 310 /* big page (2M) range */ 311 start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 312 #ifdef CONFIG_X86_32 313 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE)); 314 #else /* CONFIG_X86_64 */ 315 end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE)); 316 if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE))) 317 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE)); 318 #endif 319 320 if (start_pfn < end_pfn) { 321 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 322 page_size_mask & (1<<PG_LEVEL_2M)); 323 pfn = end_pfn; 324 } 325 326 #ifdef CONFIG_X86_64 327 /* big page (1G) range */ 328 start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE)); 329 end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE)); 330 if (start_pfn < end_pfn) { 331 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 332 page_size_mask & 333 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G))); 334 pfn = end_pfn; 335 } 336 337 /* tail is not big page (1G) alignment */ 338 start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 339 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE)); 340 if (start_pfn < end_pfn) { 341 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 342 page_size_mask & (1<<PG_LEVEL_2M)); 343 pfn = end_pfn; 344 } 345 #endif 346 347 /* tail is not big page (2M) alignment */ 348 start_pfn = pfn; 349 end_pfn = limit_pfn; 350 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0); 351 352 if (!after_bootmem) 353 adjust_range_page_size_mask(mr, nr_range); 354 355 /* try to merge same page size and continuous */ 356 for (i = 0; nr_range > 1 && i < nr_range - 1; i++) { 357 unsigned long old_start; 358 if (mr[i].end != mr[i+1].start || 359 mr[i].page_size_mask != mr[i+1].page_size_mask) 360 continue; 361 /* move it */ 362 old_start = mr[i].start; 363 memmove(&mr[i], &mr[i+1], 364 (nr_range - 1 - i) * sizeof(struct map_range)); 365 mr[i--].start = old_start; 366 nr_range--; 367 } 368 369 for (i = 0; i < nr_range; i++) 370 pr_debug(" [mem %#010lx-%#010lx] page %s\n", 371 mr[i].start, mr[i].end - 1, 372 page_size_string(&mr[i])); 373 374 return nr_range; 375 } 376 377 struct range pfn_mapped[E820_MAX_ENTRIES]; 378 int nr_pfn_mapped; 379 380 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn) 381 { 382 nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES, 383 nr_pfn_mapped, start_pfn, end_pfn); 384 nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES); 385 386 max_pfn_mapped = max(max_pfn_mapped, end_pfn); 387 388 if (start_pfn < (1UL<<(32-PAGE_SHIFT))) 389 max_low_pfn_mapped = max(max_low_pfn_mapped, 390 min(end_pfn, 1UL<<(32-PAGE_SHIFT))); 391 } 392 393 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn) 394 { 395 int i; 396 397 for (i = 0; i < nr_pfn_mapped; i++) 398 if ((start_pfn >= pfn_mapped[i].start) && 399 (end_pfn <= pfn_mapped[i].end)) 400 return true; 401 402 return false; 403 } 404 405 /* 406 * Setup the direct mapping of the physical memory at PAGE_OFFSET. 407 * This runs before bootmem is initialized and gets pages directly from 408 * the physical memory. To access them they are temporarily mapped. 409 */ 410 unsigned long __ref init_memory_mapping(unsigned long start, 411 unsigned long end) 412 { 413 struct map_range mr[NR_RANGE_MR]; 414 unsigned long ret = 0; 415 int nr_range, i; 416 417 pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n", 418 start, end - 1); 419 420 memset(mr, 0, sizeof(mr)); 421 nr_range = split_mem_range(mr, 0, start, end); 422 423 for (i = 0; i < nr_range; i++) 424 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end, 425 mr[i].page_size_mask); 426 427 add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT); 428 429 return ret >> PAGE_SHIFT; 430 } 431 432 /* 433 * We need to iterate through the E820 memory map and create direct mappings 434 * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply 435 * create direct mappings for all pfns from [0 to max_low_pfn) and 436 * [4GB to max_pfn) because of possible memory holes in high addresses 437 * that cannot be marked as UC by fixed/variable range MTRRs. 438 * Depending on the alignment of E820 ranges, this may possibly result 439 * in using smaller size (i.e. 4K instead of 2M or 1G) page tables. 440 * 441 * init_mem_mapping() calls init_range_memory_mapping() with big range. 442 * That range would have hole in the middle or ends, and only ram parts 443 * will be mapped in init_range_memory_mapping(). 444 */ 445 static unsigned long __init init_range_memory_mapping( 446 unsigned long r_start, 447 unsigned long r_end) 448 { 449 unsigned long start_pfn, end_pfn; 450 unsigned long mapped_ram_size = 0; 451 int i; 452 453 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) { 454 u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end); 455 u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end); 456 if (start >= end) 457 continue; 458 459 /* 460 * if it is overlapping with brk pgt, we need to 461 * alloc pgt buf from memblock instead. 462 */ 463 can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >= 464 min(end, (u64)pgt_buf_top<<PAGE_SHIFT); 465 init_memory_mapping(start, end); 466 mapped_ram_size += end - start; 467 can_use_brk_pgt = true; 468 } 469 470 return mapped_ram_size; 471 } 472 473 static unsigned long __init get_new_step_size(unsigned long step_size) 474 { 475 /* 476 * Initial mapped size is PMD_SIZE (2M). 477 * We can not set step_size to be PUD_SIZE (1G) yet. 478 * In worse case, when we cross the 1G boundary, and 479 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k) 480 * to map 1G range with PTE. Hence we use one less than the 481 * difference of page table level shifts. 482 * 483 * Don't need to worry about overflow in the top-down case, on 32bit, 484 * when step_size is 0, round_down() returns 0 for start, and that 485 * turns it into 0x100000000ULL. 486 * In the bottom-up case, round_up(x, 0) returns 0 though too, which 487 * needs to be taken into consideration by the code below. 488 */ 489 return step_size << (PMD_SHIFT - PAGE_SHIFT - 1); 490 } 491 492 /** 493 * memory_map_top_down - Map [map_start, map_end) top down 494 * @map_start: start address of the target memory range 495 * @map_end: end address of the target memory range 496 * 497 * This function will setup direct mapping for memory range 498 * [map_start, map_end) in top-down. That said, the page tables 499 * will be allocated at the end of the memory, and we map the 500 * memory in top-down. 501 */ 502 static void __init memory_map_top_down(unsigned long map_start, 503 unsigned long map_end) 504 { 505 unsigned long real_end, start, last_start; 506 unsigned long step_size; 507 unsigned long addr; 508 unsigned long mapped_ram_size = 0; 509 510 /* xen has big range in reserved near end of ram, skip it at first.*/ 511 addr = memblock_find_in_range(map_start, map_end, PMD_SIZE, PMD_SIZE); 512 real_end = addr + PMD_SIZE; 513 514 /* step_size need to be small so pgt_buf from BRK could cover it */ 515 step_size = PMD_SIZE; 516 max_pfn_mapped = 0; /* will get exact value next */ 517 min_pfn_mapped = real_end >> PAGE_SHIFT; 518 last_start = start = real_end; 519 520 /* 521 * We start from the top (end of memory) and go to the bottom. 522 * The memblock_find_in_range() gets us a block of RAM from the 523 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages 524 * for page table. 525 */ 526 while (last_start > map_start) { 527 if (last_start > step_size) { 528 start = round_down(last_start - 1, step_size); 529 if (start < map_start) 530 start = map_start; 531 } else 532 start = map_start; 533 mapped_ram_size += init_range_memory_mapping(start, 534 last_start); 535 last_start = start; 536 min_pfn_mapped = last_start >> PAGE_SHIFT; 537 if (mapped_ram_size >= step_size) 538 step_size = get_new_step_size(step_size); 539 } 540 541 if (real_end < map_end) 542 init_range_memory_mapping(real_end, map_end); 543 } 544 545 /** 546 * memory_map_bottom_up - Map [map_start, map_end) bottom up 547 * @map_start: start address of the target memory range 548 * @map_end: end address of the target memory range 549 * 550 * This function will setup direct mapping for memory range 551 * [map_start, map_end) in bottom-up. Since we have limited the 552 * bottom-up allocation above the kernel, the page tables will 553 * be allocated just above the kernel and we map the memory 554 * in [map_start, map_end) in bottom-up. 555 */ 556 static void __init memory_map_bottom_up(unsigned long map_start, 557 unsigned long map_end) 558 { 559 unsigned long next, start; 560 unsigned long mapped_ram_size = 0; 561 /* step_size need to be small so pgt_buf from BRK could cover it */ 562 unsigned long step_size = PMD_SIZE; 563 564 start = map_start; 565 min_pfn_mapped = start >> PAGE_SHIFT; 566 567 /* 568 * We start from the bottom (@map_start) and go to the top (@map_end). 569 * The memblock_find_in_range() gets us a block of RAM from the 570 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages 571 * for page table. 572 */ 573 while (start < map_end) { 574 if (step_size && map_end - start > step_size) { 575 next = round_up(start + 1, step_size); 576 if (next > map_end) 577 next = map_end; 578 } else { 579 next = map_end; 580 } 581 582 mapped_ram_size += init_range_memory_mapping(start, next); 583 start = next; 584 585 if (mapped_ram_size >= step_size) 586 step_size = get_new_step_size(step_size); 587 } 588 } 589 590 void __init init_mem_mapping(void) 591 { 592 unsigned long end; 593 594 probe_page_size_mask(); 595 596 #ifdef CONFIG_X86_64 597 end = max_pfn << PAGE_SHIFT; 598 #else 599 end = max_low_pfn << PAGE_SHIFT; 600 #endif 601 602 /* the ISA range is always mapped regardless of memory holes */ 603 init_memory_mapping(0, ISA_END_ADDRESS); 604 605 /* Init the trampoline, possibly with KASLR memory offset */ 606 init_trampoline(); 607 608 /* 609 * If the allocation is in bottom-up direction, we setup direct mapping 610 * in bottom-up, otherwise we setup direct mapping in top-down. 611 */ 612 if (memblock_bottom_up()) { 613 unsigned long kernel_end = __pa_symbol(_end); 614 615 /* 616 * we need two separate calls here. This is because we want to 617 * allocate page tables above the kernel. So we first map 618 * [kernel_end, end) to make memory above the kernel be mapped 619 * as soon as possible. And then use page tables allocated above 620 * the kernel to map [ISA_END_ADDRESS, kernel_end). 621 */ 622 memory_map_bottom_up(kernel_end, end); 623 memory_map_bottom_up(ISA_END_ADDRESS, kernel_end); 624 } else { 625 memory_map_top_down(ISA_END_ADDRESS, end); 626 } 627 628 #ifdef CONFIG_X86_64 629 if (max_pfn > max_low_pfn) { 630 /* can we preseve max_low_pfn ?*/ 631 max_low_pfn = max_pfn; 632 } 633 #else 634 early_ioremap_page_table_range_init(); 635 #endif 636 637 load_cr3(swapper_pg_dir); 638 __flush_tlb_all(); 639 640 hypervisor_init_mem_mapping(); 641 642 early_memtest(0, max_pfn_mapped << PAGE_SHIFT); 643 } 644 645 /* 646 * devmem_is_allowed() checks to see if /dev/mem access to a certain address 647 * is valid. The argument is a physical page number. 648 * 649 * On x86, access has to be given to the first megabyte of RAM because that 650 * area traditionally contains BIOS code and data regions used by X, dosemu, 651 * and similar apps. Since they map the entire memory range, the whole range 652 * must be allowed (for mapping), but any areas that would otherwise be 653 * disallowed are flagged as being "zero filled" instead of rejected. 654 * Access has to be given to non-kernel-ram areas as well, these contain the 655 * PCI mmio resources as well as potential bios/acpi data regions. 656 */ 657 int devmem_is_allowed(unsigned long pagenr) 658 { 659 if (page_is_ram(pagenr)) { 660 /* 661 * For disallowed memory regions in the low 1MB range, 662 * request that the page be shown as all zeros. 663 */ 664 if (pagenr < 256) 665 return 2; 666 667 return 0; 668 } 669 670 /* 671 * This must follow RAM test, since System RAM is considered a 672 * restricted resource under CONFIG_STRICT_IOMEM. 673 */ 674 if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) { 675 /* Low 1MB bypasses iomem restrictions. */ 676 if (pagenr < 256) 677 return 1; 678 679 return 0; 680 } 681 682 return 1; 683 } 684 685 void free_init_pages(char *what, unsigned long begin, unsigned long end) 686 { 687 unsigned long begin_aligned, end_aligned; 688 689 /* Make sure boundaries are page aligned */ 690 begin_aligned = PAGE_ALIGN(begin); 691 end_aligned = end & PAGE_MASK; 692 693 if (WARN_ON(begin_aligned != begin || end_aligned != end)) { 694 begin = begin_aligned; 695 end = end_aligned; 696 } 697 698 if (begin >= end) 699 return; 700 701 /* 702 * If debugging page accesses then do not free this memory but 703 * mark them not present - any buggy init-section access will 704 * create a kernel page fault: 705 */ 706 if (debug_pagealloc_enabled()) { 707 pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n", 708 begin, end - 1); 709 set_memory_np(begin, (end - begin) >> PAGE_SHIFT); 710 } else { 711 /* 712 * We just marked the kernel text read only above, now that 713 * we are going to free part of that, we need to make that 714 * writeable and non-executable first. 715 */ 716 set_memory_nx(begin, (end - begin) >> PAGE_SHIFT); 717 set_memory_rw(begin, (end - begin) >> PAGE_SHIFT); 718 719 free_reserved_area((void *)begin, (void *)end, 720 POISON_FREE_INITMEM, what); 721 } 722 } 723 724 void __ref free_initmem(void) 725 { 726 e820__reallocate_tables(); 727 728 free_init_pages("unused kernel", 729 (unsigned long)(&__init_begin), 730 (unsigned long)(&__init_end)); 731 } 732 733 #ifdef CONFIG_BLK_DEV_INITRD 734 void __init free_initrd_mem(unsigned long start, unsigned long end) 735 { 736 /* 737 * end could be not aligned, and We can not align that, 738 * decompresser could be confused by aligned initrd_end 739 * We already reserve the end partial page before in 740 * - i386_start_kernel() 741 * - x86_64_start_kernel() 742 * - relocate_initrd() 743 * So here We can do PAGE_ALIGN() safely to get partial page to be freed 744 */ 745 free_init_pages("initrd", start, PAGE_ALIGN(end)); 746 } 747 #endif 748 749 /* 750 * Calculate the precise size of the DMA zone (first 16 MB of RAM), 751 * and pass it to the MM layer - to help it set zone watermarks more 752 * accurately. 753 * 754 * Done on 64-bit systems only for the time being, although 32-bit systems 755 * might benefit from this as well. 756 */ 757 void __init memblock_find_dma_reserve(void) 758 { 759 #ifdef CONFIG_X86_64 760 u64 nr_pages = 0, nr_free_pages = 0; 761 unsigned long start_pfn, end_pfn; 762 phys_addr_t start_addr, end_addr; 763 int i; 764 u64 u; 765 766 /* 767 * Iterate over all memory ranges (free and reserved ones alike), 768 * to calculate the total number of pages in the first 16 MB of RAM: 769 */ 770 nr_pages = 0; 771 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) { 772 start_pfn = min(start_pfn, MAX_DMA_PFN); 773 end_pfn = min(end_pfn, MAX_DMA_PFN); 774 775 nr_pages += end_pfn - start_pfn; 776 } 777 778 /* 779 * Iterate over free memory ranges to calculate the number of free 780 * pages in the DMA zone, while not counting potential partial 781 * pages at the beginning or the end of the range: 782 */ 783 nr_free_pages = 0; 784 for_each_free_mem_range(u, NUMA_NO_NODE, MEMBLOCK_NONE, &start_addr, &end_addr, NULL) { 785 start_pfn = min_t(unsigned long, PFN_UP(start_addr), MAX_DMA_PFN); 786 end_pfn = min_t(unsigned long, PFN_DOWN(end_addr), MAX_DMA_PFN); 787 788 if (start_pfn < end_pfn) 789 nr_free_pages += end_pfn - start_pfn; 790 } 791 792 set_dma_reserve(nr_pages - nr_free_pages); 793 #endif 794 } 795 796 void __init zone_sizes_init(void) 797 { 798 unsigned long max_zone_pfns[MAX_NR_ZONES]; 799 800 memset(max_zone_pfns, 0, sizeof(max_zone_pfns)); 801 802 #ifdef CONFIG_ZONE_DMA 803 max_zone_pfns[ZONE_DMA] = min(MAX_DMA_PFN, max_low_pfn); 804 #endif 805 #ifdef CONFIG_ZONE_DMA32 806 max_zone_pfns[ZONE_DMA32] = min(MAX_DMA32_PFN, max_low_pfn); 807 #endif 808 max_zone_pfns[ZONE_NORMAL] = max_low_pfn; 809 #ifdef CONFIG_HIGHMEM 810 max_zone_pfns[ZONE_HIGHMEM] = max_pfn; 811 #endif 812 813 free_area_init_nodes(max_zone_pfns); 814 } 815 816 DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate) = { 817 .loaded_mm = &init_mm, 818 .state = 0, 819 .cr4 = ~0UL, /* fail hard if we screw up cr4 shadow initialization */ 820 }; 821 EXPORT_SYMBOL_GPL(cpu_tlbstate); 822 823 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache) 824 { 825 /* entry 0 MUST be WB (hardwired to speed up translations) */ 826 BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB); 827 828 __cachemode2pte_tbl[cache] = __cm_idx2pte(entry); 829 __pte2cachemode_tbl[entry] = cache; 830 } 831