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 #include <linux/swapfile.h> 8 #include <linux/swapops.h> 9 10 #include <asm/set_memory.h> 11 #include <asm/e820/api.h> 12 #include <asm/init.h> 13 #include <asm/page.h> 14 #include <asm/page_types.h> 15 #include <asm/sections.h> 16 #include <asm/setup.h> 17 #include <asm/tlbflush.h> 18 #include <asm/tlb.h> 19 #include <asm/proto.h> 20 #include <asm/dma.h> /* for MAX_DMA_PFN */ 21 #include <asm/microcode.h> 22 #include <asm/kaslr.h> 23 #include <asm/hypervisor.h> 24 #include <asm/cpufeature.h> 25 #include <asm/pti.h> 26 27 /* 28 * We need to define the tracepoints somewhere, and tlb.c 29 * is only compied when SMP=y. 30 */ 31 #define CREATE_TRACE_POINTS 32 #include <trace/events/tlb.h> 33 34 #include "mm_internal.h" 35 36 /* 37 * Tables translating between page_cache_type_t and pte encoding. 38 * 39 * The default values are defined statically as minimal supported mode; 40 * WC and WT fall back to UC-. pat_init() updates these values to support 41 * more cache modes, WC and WT, when it is safe to do so. See pat_init() 42 * for the details. Note, __early_ioremap() used during early boot-time 43 * takes pgprot_t (pte encoding) and does not use these tables. 44 * 45 * Index into __cachemode2pte_tbl[] is the cachemode. 46 * 47 * Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte 48 * (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2. 49 */ 50 uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = { 51 [_PAGE_CACHE_MODE_WB ] = 0 | 0 , 52 [_PAGE_CACHE_MODE_WC ] = 0 | _PAGE_PCD, 53 [_PAGE_CACHE_MODE_UC_MINUS] = 0 | _PAGE_PCD, 54 [_PAGE_CACHE_MODE_UC ] = _PAGE_PWT | _PAGE_PCD, 55 [_PAGE_CACHE_MODE_WT ] = 0 | _PAGE_PCD, 56 [_PAGE_CACHE_MODE_WP ] = 0 | _PAGE_PCD, 57 }; 58 EXPORT_SYMBOL(__cachemode2pte_tbl); 59 60 uint8_t __pte2cachemode_tbl[8] = { 61 [__pte2cm_idx( 0 | 0 | 0 )] = _PAGE_CACHE_MODE_WB, 62 [__pte2cm_idx(_PAGE_PWT | 0 | 0 )] = _PAGE_CACHE_MODE_UC_MINUS, 63 [__pte2cm_idx( 0 | _PAGE_PCD | 0 )] = _PAGE_CACHE_MODE_UC_MINUS, 64 [__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0 )] = _PAGE_CACHE_MODE_UC, 65 [__pte2cm_idx( 0 | 0 | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB, 66 [__pte2cm_idx(_PAGE_PWT | 0 | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS, 67 [__pte2cm_idx(0 | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS, 68 [__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC, 69 }; 70 EXPORT_SYMBOL(__pte2cachemode_tbl); 71 72 static unsigned long __initdata pgt_buf_start; 73 static unsigned long __initdata pgt_buf_end; 74 static unsigned long __initdata pgt_buf_top; 75 76 static unsigned long min_pfn_mapped; 77 78 static bool __initdata can_use_brk_pgt = true; 79 80 /* 81 * Pages returned are already directly mapped. 82 * 83 * Changing that is likely to break Xen, see commit: 84 * 85 * 279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve 86 * 87 * for detailed information. 88 */ 89 __ref void *alloc_low_pages(unsigned int num) 90 { 91 unsigned long pfn; 92 int i; 93 94 if (after_bootmem) { 95 unsigned int order; 96 97 order = get_order((unsigned long)num << PAGE_SHIFT); 98 return (void *)__get_free_pages(GFP_ATOMIC | __GFP_ZERO, order); 99 } 100 101 if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) { 102 unsigned long ret; 103 if (min_pfn_mapped >= max_pfn_mapped) 104 panic("alloc_low_pages: ran out of memory"); 105 ret = memblock_find_in_range(min_pfn_mapped << PAGE_SHIFT, 106 max_pfn_mapped << PAGE_SHIFT, 107 PAGE_SIZE * num , PAGE_SIZE); 108 if (!ret) 109 panic("alloc_low_pages: can not alloc memory"); 110 memblock_reserve(ret, PAGE_SIZE * num); 111 pfn = ret >> PAGE_SHIFT; 112 } else { 113 pfn = pgt_buf_end; 114 pgt_buf_end += num; 115 printk(KERN_DEBUG "BRK [%#010lx, %#010lx] PGTABLE\n", 116 pfn << PAGE_SHIFT, (pgt_buf_end << PAGE_SHIFT) - 1); 117 } 118 119 for (i = 0; i < num; i++) { 120 void *adr; 121 122 adr = __va((pfn + i) << PAGE_SHIFT); 123 clear_page(adr); 124 } 125 126 return __va(pfn << PAGE_SHIFT); 127 } 128 129 /* 130 * By default need 3 4k for initial PMD_SIZE, 3 4k for 0-ISA_END_ADDRESS. 131 * With KASLR memory randomization, depending on the machine e820 memory 132 * and the PUD alignment. We may need twice more pages when KASLR memory 133 * randomization is enabled. 134 */ 135 #ifndef CONFIG_RANDOMIZE_MEMORY 136 #define INIT_PGD_PAGE_COUNT 6 137 #else 138 #define INIT_PGD_PAGE_COUNT 12 139 #endif 140 #define INIT_PGT_BUF_SIZE (INIT_PGD_PAGE_COUNT * PAGE_SIZE) 141 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE); 142 void __init early_alloc_pgt_buf(void) 143 { 144 unsigned long tables = INIT_PGT_BUF_SIZE; 145 phys_addr_t base; 146 147 base = __pa(extend_brk(tables, PAGE_SIZE)); 148 149 pgt_buf_start = base >> PAGE_SHIFT; 150 pgt_buf_end = pgt_buf_start; 151 pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT); 152 } 153 154 int after_bootmem; 155 156 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES); 157 158 struct map_range { 159 unsigned long start; 160 unsigned long end; 161 unsigned page_size_mask; 162 }; 163 164 static int page_size_mask; 165 166 static void __init probe_page_size_mask(void) 167 { 168 /* 169 * For pagealloc debugging, identity mapping will use small pages. 170 * This will simplify cpa(), which otherwise needs to support splitting 171 * large pages into small in interrupt context, etc. 172 */ 173 if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled()) 174 page_size_mask |= 1 << PG_LEVEL_2M; 175 else 176 direct_gbpages = 0; 177 178 /* Enable PSE if available */ 179 if (boot_cpu_has(X86_FEATURE_PSE)) 180 cr4_set_bits_and_update_boot(X86_CR4_PSE); 181 182 /* Enable PGE if available */ 183 __supported_pte_mask &= ~_PAGE_GLOBAL; 184 if (boot_cpu_has(X86_FEATURE_PGE)) { 185 cr4_set_bits_and_update_boot(X86_CR4_PGE); 186 __supported_pte_mask |= _PAGE_GLOBAL; 187 } 188 189 /* By the default is everything supported: */ 190 __default_kernel_pte_mask = __supported_pte_mask; 191 /* Except when with PTI where the kernel is mostly non-Global: */ 192 if (cpu_feature_enabled(X86_FEATURE_PTI)) 193 __default_kernel_pte_mask &= ~_PAGE_GLOBAL; 194 195 /* Enable 1 GB linear kernel mappings if available: */ 196 if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) { 197 printk(KERN_INFO "Using GB pages for direct mapping\n"); 198 page_size_mask |= 1 << PG_LEVEL_1G; 199 } else { 200 direct_gbpages = 0; 201 } 202 } 203 204 static void setup_pcid(void) 205 { 206 if (!IS_ENABLED(CONFIG_X86_64)) 207 return; 208 209 if (!boot_cpu_has(X86_FEATURE_PCID)) 210 return; 211 212 if (boot_cpu_has(X86_FEATURE_PGE)) { 213 /* 214 * This can't be cr4_set_bits_and_update_boot() -- the 215 * trampoline code can't handle CR4.PCIDE and it wouldn't 216 * do any good anyway. Despite the name, 217 * cr4_set_bits_and_update_boot() doesn't actually cause 218 * the bits in question to remain set all the way through 219 * the secondary boot asm. 220 * 221 * Instead, we brute-force it and set CR4.PCIDE manually in 222 * start_secondary(). 223 */ 224 cr4_set_bits(X86_CR4_PCIDE); 225 226 /* 227 * INVPCID's single-context modes (2/3) only work if we set 228 * X86_CR4_PCIDE, *and* we INVPCID support. It's unusable 229 * on systems that have X86_CR4_PCIDE clear, or that have 230 * no INVPCID support at all. 231 */ 232 if (boot_cpu_has(X86_FEATURE_INVPCID)) 233 setup_force_cpu_cap(X86_FEATURE_INVPCID_SINGLE); 234 } else { 235 /* 236 * flush_tlb_all(), as currently implemented, won't work if 237 * PCID is on but PGE is not. Since that combination 238 * doesn't exist on real hardware, there's no reason to try 239 * to fully support it, but it's polite to avoid corrupting 240 * data if we're on an improperly configured VM. 241 */ 242 setup_clear_cpu_cap(X86_FEATURE_PCID); 243 } 244 } 245 246 #ifdef CONFIG_X86_32 247 #define NR_RANGE_MR 3 248 #else /* CONFIG_X86_64 */ 249 #define NR_RANGE_MR 5 250 #endif 251 252 static int __meminit save_mr(struct map_range *mr, int nr_range, 253 unsigned long start_pfn, unsigned long end_pfn, 254 unsigned long page_size_mask) 255 { 256 if (start_pfn < end_pfn) { 257 if (nr_range >= NR_RANGE_MR) 258 panic("run out of range for init_memory_mapping\n"); 259 mr[nr_range].start = start_pfn<<PAGE_SHIFT; 260 mr[nr_range].end = end_pfn<<PAGE_SHIFT; 261 mr[nr_range].page_size_mask = page_size_mask; 262 nr_range++; 263 } 264 265 return nr_range; 266 } 267 268 /* 269 * adjust the page_size_mask for small range to go with 270 * big page size instead small one if nearby are ram too. 271 */ 272 static void __ref adjust_range_page_size_mask(struct map_range *mr, 273 int nr_range) 274 { 275 int i; 276 277 for (i = 0; i < nr_range; i++) { 278 if ((page_size_mask & (1<<PG_LEVEL_2M)) && 279 !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) { 280 unsigned long start = round_down(mr[i].start, PMD_SIZE); 281 unsigned long end = round_up(mr[i].end, PMD_SIZE); 282 283 #ifdef CONFIG_X86_32 284 if ((end >> PAGE_SHIFT) > max_low_pfn) 285 continue; 286 #endif 287 288 if (memblock_is_region_memory(start, end - start)) 289 mr[i].page_size_mask |= 1<<PG_LEVEL_2M; 290 } 291 if ((page_size_mask & (1<<PG_LEVEL_1G)) && 292 !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) { 293 unsigned long start = round_down(mr[i].start, PUD_SIZE); 294 unsigned long end = round_up(mr[i].end, PUD_SIZE); 295 296 if (memblock_is_region_memory(start, end - start)) 297 mr[i].page_size_mask |= 1<<PG_LEVEL_1G; 298 } 299 } 300 } 301 302 static const char *page_size_string(struct map_range *mr) 303 { 304 static const char str_1g[] = "1G"; 305 static const char str_2m[] = "2M"; 306 static const char str_4m[] = "4M"; 307 static const char str_4k[] = "4k"; 308 309 if (mr->page_size_mask & (1<<PG_LEVEL_1G)) 310 return str_1g; 311 /* 312 * 32-bit without PAE has a 4M large page size. 313 * PG_LEVEL_2M is misnamed, but we can at least 314 * print out the right size in the string. 315 */ 316 if (IS_ENABLED(CONFIG_X86_32) && 317 !IS_ENABLED(CONFIG_X86_PAE) && 318 mr->page_size_mask & (1<<PG_LEVEL_2M)) 319 return str_4m; 320 321 if (mr->page_size_mask & (1<<PG_LEVEL_2M)) 322 return str_2m; 323 324 return str_4k; 325 } 326 327 static int __meminit split_mem_range(struct map_range *mr, int nr_range, 328 unsigned long start, 329 unsigned long end) 330 { 331 unsigned long start_pfn, end_pfn, limit_pfn; 332 unsigned long pfn; 333 int i; 334 335 limit_pfn = PFN_DOWN(end); 336 337 /* head if not big page alignment ? */ 338 pfn = start_pfn = PFN_DOWN(start); 339 #ifdef CONFIG_X86_32 340 /* 341 * Don't use a large page for the first 2/4MB of memory 342 * because there are often fixed size MTRRs in there 343 * and overlapping MTRRs into large pages can cause 344 * slowdowns. 345 */ 346 if (pfn == 0) 347 end_pfn = PFN_DOWN(PMD_SIZE); 348 else 349 end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 350 #else /* CONFIG_X86_64 */ 351 end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 352 #endif 353 if (end_pfn > limit_pfn) 354 end_pfn = limit_pfn; 355 if (start_pfn < end_pfn) { 356 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0); 357 pfn = end_pfn; 358 } 359 360 /* big page (2M) range */ 361 start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 362 #ifdef CONFIG_X86_32 363 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE)); 364 #else /* CONFIG_X86_64 */ 365 end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE)); 366 if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE))) 367 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE)); 368 #endif 369 370 if (start_pfn < end_pfn) { 371 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 372 page_size_mask & (1<<PG_LEVEL_2M)); 373 pfn = end_pfn; 374 } 375 376 #ifdef CONFIG_X86_64 377 /* big page (1G) range */ 378 start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE)); 379 end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE)); 380 if (start_pfn < end_pfn) { 381 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 382 page_size_mask & 383 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G))); 384 pfn = end_pfn; 385 } 386 387 /* tail is not big page (1G) alignment */ 388 start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 389 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE)); 390 if (start_pfn < end_pfn) { 391 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 392 page_size_mask & (1<<PG_LEVEL_2M)); 393 pfn = end_pfn; 394 } 395 #endif 396 397 /* tail is not big page (2M) alignment */ 398 start_pfn = pfn; 399 end_pfn = limit_pfn; 400 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0); 401 402 if (!after_bootmem) 403 adjust_range_page_size_mask(mr, nr_range); 404 405 /* try to merge same page size and continuous */ 406 for (i = 0; nr_range > 1 && i < nr_range - 1; i++) { 407 unsigned long old_start; 408 if (mr[i].end != mr[i+1].start || 409 mr[i].page_size_mask != mr[i+1].page_size_mask) 410 continue; 411 /* move it */ 412 old_start = mr[i].start; 413 memmove(&mr[i], &mr[i+1], 414 (nr_range - 1 - i) * sizeof(struct map_range)); 415 mr[i--].start = old_start; 416 nr_range--; 417 } 418 419 for (i = 0; i < nr_range; i++) 420 pr_debug(" [mem %#010lx-%#010lx] page %s\n", 421 mr[i].start, mr[i].end - 1, 422 page_size_string(&mr[i])); 423 424 return nr_range; 425 } 426 427 struct range pfn_mapped[E820_MAX_ENTRIES]; 428 int nr_pfn_mapped; 429 430 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn) 431 { 432 nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES, 433 nr_pfn_mapped, start_pfn, end_pfn); 434 nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES); 435 436 max_pfn_mapped = max(max_pfn_mapped, end_pfn); 437 438 if (start_pfn < (1UL<<(32-PAGE_SHIFT))) 439 max_low_pfn_mapped = max(max_low_pfn_mapped, 440 min(end_pfn, 1UL<<(32-PAGE_SHIFT))); 441 } 442 443 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn) 444 { 445 int i; 446 447 for (i = 0; i < nr_pfn_mapped; i++) 448 if ((start_pfn >= pfn_mapped[i].start) && 449 (end_pfn <= pfn_mapped[i].end)) 450 return true; 451 452 return false; 453 } 454 455 /* 456 * Setup the direct mapping of the physical memory at PAGE_OFFSET. 457 * This runs before bootmem is initialized and gets pages directly from 458 * the physical memory. To access them they are temporarily mapped. 459 */ 460 unsigned long __ref init_memory_mapping(unsigned long start, 461 unsigned long end) 462 { 463 struct map_range mr[NR_RANGE_MR]; 464 unsigned long ret = 0; 465 int nr_range, i; 466 467 pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n", 468 start, end - 1); 469 470 memset(mr, 0, sizeof(mr)); 471 nr_range = split_mem_range(mr, 0, start, end); 472 473 for (i = 0; i < nr_range; i++) 474 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end, 475 mr[i].page_size_mask); 476 477 add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT); 478 479 return ret >> PAGE_SHIFT; 480 } 481 482 /* 483 * We need to iterate through the E820 memory map and create direct mappings 484 * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply 485 * create direct mappings for all pfns from [0 to max_low_pfn) and 486 * [4GB to max_pfn) because of possible memory holes in high addresses 487 * that cannot be marked as UC by fixed/variable range MTRRs. 488 * Depending on the alignment of E820 ranges, this may possibly result 489 * in using smaller size (i.e. 4K instead of 2M or 1G) page tables. 490 * 491 * init_mem_mapping() calls init_range_memory_mapping() with big range. 492 * That range would have hole in the middle or ends, and only ram parts 493 * will be mapped in init_range_memory_mapping(). 494 */ 495 static unsigned long __init init_range_memory_mapping( 496 unsigned long r_start, 497 unsigned long r_end) 498 { 499 unsigned long start_pfn, end_pfn; 500 unsigned long mapped_ram_size = 0; 501 int i; 502 503 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) { 504 u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end); 505 u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end); 506 if (start >= end) 507 continue; 508 509 /* 510 * if it is overlapping with brk pgt, we need to 511 * alloc pgt buf from memblock instead. 512 */ 513 can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >= 514 min(end, (u64)pgt_buf_top<<PAGE_SHIFT); 515 init_memory_mapping(start, end); 516 mapped_ram_size += end - start; 517 can_use_brk_pgt = true; 518 } 519 520 return mapped_ram_size; 521 } 522 523 static unsigned long __init get_new_step_size(unsigned long step_size) 524 { 525 /* 526 * Initial mapped size is PMD_SIZE (2M). 527 * We can not set step_size to be PUD_SIZE (1G) yet. 528 * In worse case, when we cross the 1G boundary, and 529 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k) 530 * to map 1G range with PTE. Hence we use one less than the 531 * difference of page table level shifts. 532 * 533 * Don't need to worry about overflow in the top-down case, on 32bit, 534 * when step_size is 0, round_down() returns 0 for start, and that 535 * turns it into 0x100000000ULL. 536 * In the bottom-up case, round_up(x, 0) returns 0 though too, which 537 * needs to be taken into consideration by the code below. 538 */ 539 return step_size << (PMD_SHIFT - PAGE_SHIFT - 1); 540 } 541 542 /** 543 * memory_map_top_down - Map [map_start, map_end) top down 544 * @map_start: start address of the target memory range 545 * @map_end: end address of the target memory range 546 * 547 * This function will setup direct mapping for memory range 548 * [map_start, map_end) in top-down. That said, the page tables 549 * will be allocated at the end of the memory, and we map the 550 * memory in top-down. 551 */ 552 static void __init memory_map_top_down(unsigned long map_start, 553 unsigned long map_end) 554 { 555 unsigned long real_end, start, last_start; 556 unsigned long step_size; 557 unsigned long addr; 558 unsigned long mapped_ram_size = 0; 559 560 /* xen has big range in reserved near end of ram, skip it at first.*/ 561 addr = memblock_find_in_range(map_start, map_end, PMD_SIZE, PMD_SIZE); 562 real_end = addr + PMD_SIZE; 563 564 /* step_size need to be small so pgt_buf from BRK could cover it */ 565 step_size = PMD_SIZE; 566 max_pfn_mapped = 0; /* will get exact value next */ 567 min_pfn_mapped = real_end >> PAGE_SHIFT; 568 last_start = start = real_end; 569 570 /* 571 * We start from the top (end of memory) and go to the bottom. 572 * The memblock_find_in_range() gets us a block of RAM from the 573 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages 574 * for page table. 575 */ 576 while (last_start > map_start) { 577 if (last_start > step_size) { 578 start = round_down(last_start - 1, step_size); 579 if (start < map_start) 580 start = map_start; 581 } else 582 start = map_start; 583 mapped_ram_size += init_range_memory_mapping(start, 584 last_start); 585 last_start = start; 586 min_pfn_mapped = last_start >> PAGE_SHIFT; 587 if (mapped_ram_size >= step_size) 588 step_size = get_new_step_size(step_size); 589 } 590 591 if (real_end < map_end) 592 init_range_memory_mapping(real_end, map_end); 593 } 594 595 /** 596 * memory_map_bottom_up - Map [map_start, map_end) bottom up 597 * @map_start: start address of the target memory range 598 * @map_end: end address of the target memory range 599 * 600 * This function will setup direct mapping for memory range 601 * [map_start, map_end) in bottom-up. Since we have limited the 602 * bottom-up allocation above the kernel, the page tables will 603 * be allocated just above the kernel and we map the memory 604 * in [map_start, map_end) in bottom-up. 605 */ 606 static void __init memory_map_bottom_up(unsigned long map_start, 607 unsigned long map_end) 608 { 609 unsigned long next, start; 610 unsigned long mapped_ram_size = 0; 611 /* step_size need to be small so pgt_buf from BRK could cover it */ 612 unsigned long step_size = PMD_SIZE; 613 614 start = map_start; 615 min_pfn_mapped = start >> PAGE_SHIFT; 616 617 /* 618 * We start from the bottom (@map_start) and go to the top (@map_end). 619 * The memblock_find_in_range() gets us a block of RAM from the 620 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages 621 * for page table. 622 */ 623 while (start < map_end) { 624 if (step_size && map_end - start > step_size) { 625 next = round_up(start + 1, step_size); 626 if (next > map_end) 627 next = map_end; 628 } else { 629 next = map_end; 630 } 631 632 mapped_ram_size += init_range_memory_mapping(start, next); 633 start = next; 634 635 if (mapped_ram_size >= step_size) 636 step_size = get_new_step_size(step_size); 637 } 638 } 639 640 void __init init_mem_mapping(void) 641 { 642 unsigned long end; 643 644 pti_check_boottime_disable(); 645 probe_page_size_mask(); 646 setup_pcid(); 647 648 #ifdef CONFIG_X86_64 649 end = max_pfn << PAGE_SHIFT; 650 #else 651 end = max_low_pfn << PAGE_SHIFT; 652 #endif 653 654 /* the ISA range is always mapped regardless of memory holes */ 655 init_memory_mapping(0, ISA_END_ADDRESS); 656 657 /* Init the trampoline, possibly with KASLR memory offset */ 658 init_trampoline(); 659 660 /* 661 * If the allocation is in bottom-up direction, we setup direct mapping 662 * in bottom-up, otherwise we setup direct mapping in top-down. 663 */ 664 if (memblock_bottom_up()) { 665 unsigned long kernel_end = __pa_symbol(_end); 666 667 /* 668 * we need two separate calls here. This is because we want to 669 * allocate page tables above the kernel. So we first map 670 * [kernel_end, end) to make memory above the kernel be mapped 671 * as soon as possible. And then use page tables allocated above 672 * the kernel to map [ISA_END_ADDRESS, kernel_end). 673 */ 674 memory_map_bottom_up(kernel_end, end); 675 memory_map_bottom_up(ISA_END_ADDRESS, kernel_end); 676 } else { 677 memory_map_top_down(ISA_END_ADDRESS, end); 678 } 679 680 #ifdef CONFIG_X86_64 681 if (max_pfn > max_low_pfn) { 682 /* can we preseve max_low_pfn ?*/ 683 max_low_pfn = max_pfn; 684 } 685 #else 686 early_ioremap_page_table_range_init(); 687 #endif 688 689 load_cr3(swapper_pg_dir); 690 __flush_tlb_all(); 691 692 x86_init.hyper.init_mem_mapping(); 693 694 early_memtest(0, max_pfn_mapped << PAGE_SHIFT); 695 } 696 697 /* 698 * devmem_is_allowed() checks to see if /dev/mem access to a certain address 699 * is valid. The argument is a physical page number. 700 * 701 * On x86, access has to be given to the first megabyte of RAM because that 702 * area traditionally contains BIOS code and data regions used by X, dosemu, 703 * and similar apps. Since they map the entire memory range, the whole range 704 * must be allowed (for mapping), but any areas that would otherwise be 705 * disallowed are flagged as being "zero filled" instead of rejected. 706 * Access has to be given to non-kernel-ram areas as well, these contain the 707 * PCI mmio resources as well as potential bios/acpi data regions. 708 */ 709 int devmem_is_allowed(unsigned long pagenr) 710 { 711 if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE, 712 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE) 713 != REGION_DISJOINT) { 714 /* 715 * For disallowed memory regions in the low 1MB range, 716 * request that the page be shown as all zeros. 717 */ 718 if (pagenr < 256) 719 return 2; 720 721 return 0; 722 } 723 724 /* 725 * This must follow RAM test, since System RAM is considered a 726 * restricted resource under CONFIG_STRICT_IOMEM. 727 */ 728 if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) { 729 /* Low 1MB bypasses iomem restrictions. */ 730 if (pagenr < 256) 731 return 1; 732 733 return 0; 734 } 735 736 return 1; 737 } 738 739 void free_init_pages(char *what, unsigned long begin, unsigned long end) 740 { 741 unsigned long begin_aligned, end_aligned; 742 743 /* Make sure boundaries are page aligned */ 744 begin_aligned = PAGE_ALIGN(begin); 745 end_aligned = end & PAGE_MASK; 746 747 if (WARN_ON(begin_aligned != begin || end_aligned != end)) { 748 begin = begin_aligned; 749 end = end_aligned; 750 } 751 752 if (begin >= end) 753 return; 754 755 /* 756 * If debugging page accesses then do not free this memory but 757 * mark them not present - any buggy init-section access will 758 * create a kernel page fault: 759 */ 760 if (debug_pagealloc_enabled()) { 761 pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n", 762 begin, end - 1); 763 set_memory_np(begin, (end - begin) >> PAGE_SHIFT); 764 } else { 765 /* 766 * We just marked the kernel text read only above, now that 767 * we are going to free part of that, we need to make that 768 * writeable and non-executable first. 769 */ 770 set_memory_nx(begin, (end - begin) >> PAGE_SHIFT); 771 set_memory_rw(begin, (end - begin) >> PAGE_SHIFT); 772 773 free_reserved_area((void *)begin, (void *)end, 774 POISON_FREE_INITMEM, what); 775 } 776 } 777 778 /* 779 * begin/end can be in the direct map or the "high kernel mapping" 780 * used for the kernel image only. free_init_pages() will do the 781 * right thing for either kind of address. 782 */ 783 void free_kernel_image_pages(void *begin, void *end) 784 { 785 unsigned long begin_ul = (unsigned long)begin; 786 unsigned long end_ul = (unsigned long)end; 787 unsigned long len_pages = (end_ul - begin_ul) >> PAGE_SHIFT; 788 789 790 free_init_pages("unused kernel image", begin_ul, end_ul); 791 792 /* 793 * PTI maps some of the kernel into userspace. For performance, 794 * this includes some kernel areas that do not contain secrets. 795 * Those areas might be adjacent to the parts of the kernel image 796 * being freed, which may contain secrets. Remove the "high kernel 797 * image mapping" for these freed areas, ensuring they are not even 798 * potentially vulnerable to Meltdown regardless of the specific 799 * optimizations PTI is currently using. 800 * 801 * The "noalias" prevents unmapping the direct map alias which is 802 * needed to access the freed pages. 803 * 804 * This is only valid for 64bit kernels. 32bit has only one mapping 805 * which can't be treated in this way for obvious reasons. 806 */ 807 if (IS_ENABLED(CONFIG_X86_64) && cpu_feature_enabled(X86_FEATURE_PTI)) 808 set_memory_np_noalias(begin_ul, len_pages); 809 } 810 811 void __ref free_initmem(void) 812 { 813 e820__reallocate_tables(); 814 815 free_kernel_image_pages(&__init_begin, &__init_end); 816 } 817 818 #ifdef CONFIG_BLK_DEV_INITRD 819 void __init free_initrd_mem(unsigned long start, unsigned long end) 820 { 821 /* 822 * end could be not aligned, and We can not align that, 823 * decompresser could be confused by aligned initrd_end 824 * We already reserve the end partial page before in 825 * - i386_start_kernel() 826 * - x86_64_start_kernel() 827 * - relocate_initrd() 828 * So here We can do PAGE_ALIGN() safely to get partial page to be freed 829 */ 830 free_init_pages("initrd", start, PAGE_ALIGN(end)); 831 } 832 #endif 833 834 /* 835 * Calculate the precise size of the DMA zone (first 16 MB of RAM), 836 * and pass it to the MM layer - to help it set zone watermarks more 837 * accurately. 838 * 839 * Done on 64-bit systems only for the time being, although 32-bit systems 840 * might benefit from this as well. 841 */ 842 void __init memblock_find_dma_reserve(void) 843 { 844 #ifdef CONFIG_X86_64 845 u64 nr_pages = 0, nr_free_pages = 0; 846 unsigned long start_pfn, end_pfn; 847 phys_addr_t start_addr, end_addr; 848 int i; 849 u64 u; 850 851 /* 852 * Iterate over all memory ranges (free and reserved ones alike), 853 * to calculate the total number of pages in the first 16 MB of RAM: 854 */ 855 nr_pages = 0; 856 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) { 857 start_pfn = min(start_pfn, MAX_DMA_PFN); 858 end_pfn = min(end_pfn, MAX_DMA_PFN); 859 860 nr_pages += end_pfn - start_pfn; 861 } 862 863 /* 864 * Iterate over free memory ranges to calculate the number of free 865 * pages in the DMA zone, while not counting potential partial 866 * pages at the beginning or the end of the range: 867 */ 868 nr_free_pages = 0; 869 for_each_free_mem_range(u, NUMA_NO_NODE, MEMBLOCK_NONE, &start_addr, &end_addr, NULL) { 870 start_pfn = min_t(unsigned long, PFN_UP(start_addr), MAX_DMA_PFN); 871 end_pfn = min_t(unsigned long, PFN_DOWN(end_addr), MAX_DMA_PFN); 872 873 if (start_pfn < end_pfn) 874 nr_free_pages += end_pfn - start_pfn; 875 } 876 877 set_dma_reserve(nr_pages - nr_free_pages); 878 #endif 879 } 880 881 void __init zone_sizes_init(void) 882 { 883 unsigned long max_zone_pfns[MAX_NR_ZONES]; 884 885 memset(max_zone_pfns, 0, sizeof(max_zone_pfns)); 886 887 #ifdef CONFIG_ZONE_DMA 888 max_zone_pfns[ZONE_DMA] = min(MAX_DMA_PFN, max_low_pfn); 889 #endif 890 #ifdef CONFIG_ZONE_DMA32 891 max_zone_pfns[ZONE_DMA32] = min(MAX_DMA32_PFN, max_low_pfn); 892 #endif 893 max_zone_pfns[ZONE_NORMAL] = max_low_pfn; 894 #ifdef CONFIG_HIGHMEM 895 max_zone_pfns[ZONE_HIGHMEM] = max_pfn; 896 #endif 897 898 free_area_init_nodes(max_zone_pfns); 899 } 900 901 __visible DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state, cpu_tlbstate) = { 902 .loaded_mm = &init_mm, 903 .next_asid = 1, 904 .cr4 = ~0UL, /* fail hard if we screw up cr4 shadow initialization */ 905 }; 906 EXPORT_PER_CPU_SYMBOL(cpu_tlbstate); 907 908 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache) 909 { 910 /* entry 0 MUST be WB (hardwired to speed up translations) */ 911 BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB); 912 913 __cachemode2pte_tbl[cache] = __cm_idx2pte(entry); 914 __pte2cachemode_tbl[entry] = cache; 915 } 916 917 #ifdef CONFIG_SWAP 918 unsigned long max_swapfile_size(void) 919 { 920 unsigned long pages; 921 922 pages = generic_max_swapfile_size(); 923 924 if (boot_cpu_has_bug(X86_BUG_L1TF)) { 925 /* Limit the swap file size to MAX_PA/2 for L1TF workaround */ 926 unsigned long l1tf_limit = l1tf_pfn_limit() + 1; 927 /* 928 * We encode swap offsets also with 3 bits below those for pfn 929 * which makes the usable limit higher. 930 */ 931 #if CONFIG_PGTABLE_LEVELS > 2 932 l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT; 933 #endif 934 pages = min_t(unsigned long, l1tf_limit, pages); 935 } 936 return pages; 937 } 938 #endif 939