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/swapfile.h> 7 #include <linux/swapops.h> 8 #include <linux/kmemleak.h> 9 #include <linux/sched/task.h> 10 #include <linux/execmem.h> 11 12 #include <asm/set_memory.h> 13 #include <asm/cpu_device_id.h> 14 #include <asm/e820/api.h> 15 #include <asm/init.h> 16 #include <asm/page.h> 17 #include <asm/page_types.h> 18 #include <asm/sections.h> 19 #include <asm/setup.h> 20 #include <asm/tlbflush.h> 21 #include <asm/tlb.h> 22 #include <asm/proto.h> 23 #include <asm/dma.h> /* for MAX_DMA_PFN */ 24 #include <asm/kaslr.h> 25 #include <asm/hypervisor.h> 26 #include <asm/cpufeature.h> 27 #include <asm/pti.h> 28 #include <asm/text-patching.h> 29 #include <asm/memtype.h> 30 #include <asm/paravirt.h> 31 32 /* 33 * We need to define the tracepoints somewhere, and tlb.c 34 * is only compiled when SMP=y. 35 */ 36 #include <trace/events/tlb.h> 37 38 #include "mm_internal.h" 39 40 /* 41 * Tables translating between page_cache_type_t and pte encoding. 42 * 43 * The default values are defined statically as minimal supported mode; 44 * WC and WT fall back to UC-. pat_init() updates these values to support 45 * more cache modes, WC and WT, when it is safe to do so. See pat_init() 46 * for the details. Note, __early_ioremap() used during early boot-time 47 * takes pgprot_t (pte encoding) and does not use these tables. 48 * 49 * Index into __cachemode2pte_tbl[] is the cachemode. 50 * 51 * Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte 52 * (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2. 53 */ 54 static uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = { 55 [_PAGE_CACHE_MODE_WB ] = 0 | 0 , 56 [_PAGE_CACHE_MODE_WC ] = 0 | _PAGE_PCD, 57 [_PAGE_CACHE_MODE_UC_MINUS] = 0 | _PAGE_PCD, 58 [_PAGE_CACHE_MODE_UC ] = _PAGE_PWT | _PAGE_PCD, 59 [_PAGE_CACHE_MODE_WT ] = 0 | _PAGE_PCD, 60 [_PAGE_CACHE_MODE_WP ] = 0 | _PAGE_PCD, 61 }; 62 63 unsigned long cachemode2protval(enum page_cache_mode pcm) 64 { 65 if (likely(pcm == 0)) 66 return 0; 67 return __cachemode2pte_tbl[pcm]; 68 } 69 EXPORT_SYMBOL(cachemode2protval); 70 71 static uint8_t __pte2cachemode_tbl[8] = { 72 [__pte2cm_idx( 0 | 0 | 0 )] = _PAGE_CACHE_MODE_WB, 73 [__pte2cm_idx(_PAGE_PWT | 0 | 0 )] = _PAGE_CACHE_MODE_UC_MINUS, 74 [__pte2cm_idx( 0 | _PAGE_PCD | 0 )] = _PAGE_CACHE_MODE_UC_MINUS, 75 [__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0 )] = _PAGE_CACHE_MODE_UC, 76 [__pte2cm_idx( 0 | 0 | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB, 77 [__pte2cm_idx(_PAGE_PWT | 0 | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS, 78 [__pte2cm_idx(0 | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS, 79 [__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC, 80 }; 81 82 /* 83 * Check that the write-protect PAT entry is set for write-protect. 84 * To do this without making assumptions how PAT has been set up (Xen has 85 * another layout than the kernel), translate the _PAGE_CACHE_MODE_WP cache 86 * mode via the __cachemode2pte_tbl[] into protection bits (those protection 87 * bits will select a cache mode of WP or better), and then translate the 88 * protection bits back into the cache mode using __pte2cm_idx() and the 89 * __pte2cachemode_tbl[] array. This will return the really used cache mode. 90 */ 91 bool x86_has_pat_wp(void) 92 { 93 uint16_t prot = __cachemode2pte_tbl[_PAGE_CACHE_MODE_WP]; 94 95 return __pte2cachemode_tbl[__pte2cm_idx(prot)] == _PAGE_CACHE_MODE_WP; 96 } 97 98 enum page_cache_mode pgprot2cachemode(pgprot_t pgprot) 99 { 100 unsigned long masked; 101 102 masked = pgprot_val(pgprot) & _PAGE_CACHE_MASK; 103 if (likely(masked == 0)) 104 return 0; 105 return __pte2cachemode_tbl[__pte2cm_idx(masked)]; 106 } 107 108 static unsigned long __initdata pgt_buf_start; 109 static unsigned long __initdata pgt_buf_end; 110 static unsigned long __initdata pgt_buf_top; 111 112 static unsigned long min_pfn_mapped; 113 114 static bool __initdata can_use_brk_pgt = true; 115 116 /* 117 * Pages returned are already directly mapped. 118 * 119 * Changing that is likely to break Xen, see commit: 120 * 121 * 279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve 122 * 123 * for detailed information. 124 */ 125 __ref void *alloc_low_pages(unsigned int num) 126 { 127 unsigned long pfn; 128 int i; 129 130 if (after_bootmem) { 131 unsigned int order; 132 133 order = get_order((unsigned long)num << PAGE_SHIFT); 134 return (void *)__get_free_pages(GFP_ATOMIC | __GFP_ZERO, order); 135 } 136 137 if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) { 138 unsigned long ret = 0; 139 140 if (min_pfn_mapped < max_pfn_mapped) { 141 ret = memblock_phys_alloc_range( 142 PAGE_SIZE * num, PAGE_SIZE, 143 min_pfn_mapped << PAGE_SHIFT, 144 max_pfn_mapped << PAGE_SHIFT); 145 } 146 if (!ret && can_use_brk_pgt) 147 ret = __pa(extend_brk(PAGE_SIZE * num, PAGE_SIZE)); 148 149 if (!ret) 150 panic("alloc_low_pages: can not alloc memory"); 151 152 pfn = ret >> PAGE_SHIFT; 153 } else { 154 pfn = pgt_buf_end; 155 pgt_buf_end += num; 156 } 157 158 for (i = 0; i < num; i++) { 159 void *adr; 160 161 adr = __va((pfn + i) << PAGE_SHIFT); 162 clear_page(adr); 163 } 164 165 return __va(pfn << PAGE_SHIFT); 166 } 167 168 /* 169 * By default need to be able to allocate page tables below PGD firstly for 170 * the 0-ISA_END_ADDRESS range and secondly for the initial PMD_SIZE mapping. 171 * With KASLR memory randomization, depending on the machine e820 memory and the 172 * PUD alignment, twice that many pages may be needed when KASLR memory 173 * randomization is enabled. 174 */ 175 176 #ifndef CONFIG_X86_5LEVEL 177 #define INIT_PGD_PAGE_TABLES 3 178 #else 179 #define INIT_PGD_PAGE_TABLES 4 180 #endif 181 182 #ifndef CONFIG_RANDOMIZE_MEMORY 183 #define INIT_PGD_PAGE_COUNT (2 * INIT_PGD_PAGE_TABLES) 184 #else 185 #define INIT_PGD_PAGE_COUNT (4 * INIT_PGD_PAGE_TABLES) 186 #endif 187 188 #define INIT_PGT_BUF_SIZE (INIT_PGD_PAGE_COUNT * PAGE_SIZE) 189 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE); 190 void __init early_alloc_pgt_buf(void) 191 { 192 unsigned long tables = INIT_PGT_BUF_SIZE; 193 phys_addr_t base; 194 195 base = __pa(extend_brk(tables, PAGE_SIZE)); 196 197 pgt_buf_start = base >> PAGE_SHIFT; 198 pgt_buf_end = pgt_buf_start; 199 pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT); 200 } 201 202 int after_bootmem; 203 204 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES); 205 206 struct map_range { 207 unsigned long start; 208 unsigned long end; 209 unsigned page_size_mask; 210 }; 211 212 static int page_size_mask; 213 214 /* 215 * Save some of cr4 feature set we're using (e.g. Pentium 4MB 216 * enable and PPro Global page enable), so that any CPU's that boot 217 * up after us can get the correct flags. Invoked on the boot CPU. 218 */ 219 static inline void cr4_set_bits_and_update_boot(unsigned long mask) 220 { 221 mmu_cr4_features |= mask; 222 if (trampoline_cr4_features) 223 *trampoline_cr4_features = mmu_cr4_features; 224 cr4_set_bits(mask); 225 } 226 227 static void __init probe_page_size_mask(void) 228 { 229 /* 230 * For pagealloc debugging, identity mapping will use small pages. 231 * This will simplify cpa(), which otherwise needs to support splitting 232 * large pages into small in interrupt context, etc. 233 */ 234 if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled()) 235 page_size_mask |= 1 << PG_LEVEL_2M; 236 else 237 direct_gbpages = 0; 238 239 /* Enable PSE if available */ 240 if (boot_cpu_has(X86_FEATURE_PSE)) 241 cr4_set_bits_and_update_boot(X86_CR4_PSE); 242 243 /* Enable PGE if available */ 244 __supported_pte_mask &= ~_PAGE_GLOBAL; 245 if (boot_cpu_has(X86_FEATURE_PGE)) { 246 cr4_set_bits_and_update_boot(X86_CR4_PGE); 247 __supported_pte_mask |= _PAGE_GLOBAL; 248 } 249 250 /* By the default is everything supported: */ 251 __default_kernel_pte_mask = __supported_pte_mask; 252 /* Except when with PTI where the kernel is mostly non-Global: */ 253 if (cpu_feature_enabled(X86_FEATURE_PTI)) 254 __default_kernel_pte_mask &= ~_PAGE_GLOBAL; 255 256 /* Enable 1 GB linear kernel mappings if available: */ 257 if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) { 258 printk(KERN_INFO "Using GB pages for direct mapping\n"); 259 page_size_mask |= 1 << PG_LEVEL_1G; 260 } else { 261 direct_gbpages = 0; 262 } 263 } 264 265 /* 266 * INVLPG may not properly flush Global entries on 267 * these CPUs. New microcode fixes the issue. 268 */ 269 static const struct x86_cpu_id invlpg_miss_ids[] = { 270 X86_MATCH_VFM(INTEL_ALDERLAKE, 0x2e), 271 X86_MATCH_VFM(INTEL_ALDERLAKE_L, 0x42c), 272 X86_MATCH_VFM(INTEL_ATOM_GRACEMONT, 0x11), 273 X86_MATCH_VFM(INTEL_RAPTORLAKE, 0x118), 274 X86_MATCH_VFM(INTEL_RAPTORLAKE_P, 0x4117), 275 X86_MATCH_VFM(INTEL_RAPTORLAKE_S, 0x2e), 276 {} 277 }; 278 279 static void setup_pcid(void) 280 { 281 const struct x86_cpu_id *invlpg_miss_match; 282 283 if (!IS_ENABLED(CONFIG_X86_64)) 284 return; 285 286 if (!boot_cpu_has(X86_FEATURE_PCID)) 287 return; 288 289 invlpg_miss_match = x86_match_cpu(invlpg_miss_ids); 290 291 if (invlpg_miss_match && 292 boot_cpu_data.microcode < invlpg_miss_match->driver_data) { 293 pr_info("Incomplete global flushes, disabling PCID"); 294 setup_clear_cpu_cap(X86_FEATURE_PCID); 295 return; 296 } 297 298 if (boot_cpu_has(X86_FEATURE_PGE)) { 299 /* 300 * This can't be cr4_set_bits_and_update_boot() -- the 301 * trampoline code can't handle CR4.PCIDE and it wouldn't 302 * do any good anyway. Despite the name, 303 * cr4_set_bits_and_update_boot() doesn't actually cause 304 * the bits in question to remain set all the way through 305 * the secondary boot asm. 306 * 307 * Instead, we brute-force it and set CR4.PCIDE manually in 308 * start_secondary(). 309 */ 310 cr4_set_bits(X86_CR4_PCIDE); 311 } else { 312 /* 313 * flush_tlb_all(), as currently implemented, won't work if 314 * PCID is on but PGE is not. Since that combination 315 * doesn't exist on real hardware, there's no reason to try 316 * to fully support it, but it's polite to avoid corrupting 317 * data if we're on an improperly configured VM. 318 */ 319 setup_clear_cpu_cap(X86_FEATURE_PCID); 320 } 321 } 322 323 #ifdef CONFIG_X86_32 324 #define NR_RANGE_MR 3 325 #else /* CONFIG_X86_64 */ 326 #define NR_RANGE_MR 5 327 #endif 328 329 static int __meminit save_mr(struct map_range *mr, int nr_range, 330 unsigned long start_pfn, unsigned long end_pfn, 331 unsigned long page_size_mask) 332 { 333 if (start_pfn < end_pfn) { 334 if (nr_range >= NR_RANGE_MR) 335 panic("run out of range for init_memory_mapping\n"); 336 mr[nr_range].start = start_pfn<<PAGE_SHIFT; 337 mr[nr_range].end = end_pfn<<PAGE_SHIFT; 338 mr[nr_range].page_size_mask = page_size_mask; 339 nr_range++; 340 } 341 342 return nr_range; 343 } 344 345 /* 346 * adjust the page_size_mask for small range to go with 347 * big page size instead small one if nearby are ram too. 348 */ 349 static void __ref adjust_range_page_size_mask(struct map_range *mr, 350 int nr_range) 351 { 352 int i; 353 354 for (i = 0; i < nr_range; i++) { 355 if ((page_size_mask & (1<<PG_LEVEL_2M)) && 356 !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) { 357 unsigned long start = round_down(mr[i].start, PMD_SIZE); 358 unsigned long end = round_up(mr[i].end, PMD_SIZE); 359 360 #ifdef CONFIG_X86_32 361 if ((end >> PAGE_SHIFT) > max_low_pfn) 362 continue; 363 #endif 364 365 if (memblock_is_region_memory(start, end - start)) 366 mr[i].page_size_mask |= 1<<PG_LEVEL_2M; 367 } 368 if ((page_size_mask & (1<<PG_LEVEL_1G)) && 369 !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) { 370 unsigned long start = round_down(mr[i].start, PUD_SIZE); 371 unsigned long end = round_up(mr[i].end, PUD_SIZE); 372 373 if (memblock_is_region_memory(start, end - start)) 374 mr[i].page_size_mask |= 1<<PG_LEVEL_1G; 375 } 376 } 377 } 378 379 static const char *page_size_string(struct map_range *mr) 380 { 381 static const char str_1g[] = "1G"; 382 static const char str_2m[] = "2M"; 383 static const char str_4m[] = "4M"; 384 static const char str_4k[] = "4k"; 385 386 if (mr->page_size_mask & (1<<PG_LEVEL_1G)) 387 return str_1g; 388 /* 389 * 32-bit without PAE has a 4M large page size. 390 * PG_LEVEL_2M is misnamed, but we can at least 391 * print out the right size in the string. 392 */ 393 if (IS_ENABLED(CONFIG_X86_32) && 394 !IS_ENABLED(CONFIG_X86_PAE) && 395 mr->page_size_mask & (1<<PG_LEVEL_2M)) 396 return str_4m; 397 398 if (mr->page_size_mask & (1<<PG_LEVEL_2M)) 399 return str_2m; 400 401 return str_4k; 402 } 403 404 static int __meminit split_mem_range(struct map_range *mr, int nr_range, 405 unsigned long start, 406 unsigned long end) 407 { 408 unsigned long start_pfn, end_pfn, limit_pfn; 409 unsigned long pfn; 410 int i; 411 412 limit_pfn = PFN_DOWN(end); 413 414 /* head if not big page alignment ? */ 415 pfn = start_pfn = PFN_DOWN(start); 416 #ifdef CONFIG_X86_32 417 /* 418 * Don't use a large page for the first 2/4MB of memory 419 * because there are often fixed size MTRRs in there 420 * and overlapping MTRRs into large pages can cause 421 * slowdowns. 422 */ 423 if (pfn == 0) 424 end_pfn = PFN_DOWN(PMD_SIZE); 425 else 426 end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 427 #else /* CONFIG_X86_64 */ 428 end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 429 #endif 430 if (end_pfn > limit_pfn) 431 end_pfn = limit_pfn; 432 if (start_pfn < end_pfn) { 433 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0); 434 pfn = end_pfn; 435 } 436 437 /* big page (2M) range */ 438 start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 439 #ifdef CONFIG_X86_32 440 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE)); 441 #else /* CONFIG_X86_64 */ 442 end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE)); 443 if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE))) 444 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE)); 445 #endif 446 447 if (start_pfn < end_pfn) { 448 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 449 page_size_mask & (1<<PG_LEVEL_2M)); 450 pfn = end_pfn; 451 } 452 453 #ifdef CONFIG_X86_64 454 /* big page (1G) range */ 455 start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE)); 456 end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE)); 457 if (start_pfn < end_pfn) { 458 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 459 page_size_mask & 460 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G))); 461 pfn = end_pfn; 462 } 463 464 /* tail is not big page (1G) alignment */ 465 start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 466 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE)); 467 if (start_pfn < end_pfn) { 468 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 469 page_size_mask & (1<<PG_LEVEL_2M)); 470 pfn = end_pfn; 471 } 472 #endif 473 474 /* tail is not big page (2M) alignment */ 475 start_pfn = pfn; 476 end_pfn = limit_pfn; 477 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0); 478 479 if (!after_bootmem) 480 adjust_range_page_size_mask(mr, nr_range); 481 482 /* try to merge same page size and continuous */ 483 for (i = 0; nr_range > 1 && i < nr_range - 1; i++) { 484 unsigned long old_start; 485 if (mr[i].end != mr[i+1].start || 486 mr[i].page_size_mask != mr[i+1].page_size_mask) 487 continue; 488 /* move it */ 489 old_start = mr[i].start; 490 memmove(&mr[i], &mr[i+1], 491 (nr_range - 1 - i) * sizeof(struct map_range)); 492 mr[i--].start = old_start; 493 nr_range--; 494 } 495 496 for (i = 0; i < nr_range; i++) 497 pr_debug(" [mem %#010lx-%#010lx] page %s\n", 498 mr[i].start, mr[i].end - 1, 499 page_size_string(&mr[i])); 500 501 return nr_range; 502 } 503 504 struct range pfn_mapped[E820_MAX_ENTRIES]; 505 int nr_pfn_mapped; 506 507 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn) 508 { 509 nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES, 510 nr_pfn_mapped, start_pfn, end_pfn); 511 nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES); 512 513 max_pfn_mapped = max(max_pfn_mapped, end_pfn); 514 515 if (start_pfn < (1UL<<(32-PAGE_SHIFT))) 516 max_low_pfn_mapped = max(max_low_pfn_mapped, 517 min(end_pfn, 1UL<<(32-PAGE_SHIFT))); 518 } 519 520 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn) 521 { 522 int i; 523 524 for (i = 0; i < nr_pfn_mapped; i++) 525 if ((start_pfn >= pfn_mapped[i].start) && 526 (end_pfn <= pfn_mapped[i].end)) 527 return true; 528 529 return false; 530 } 531 532 /* 533 * Setup the direct mapping of the physical memory at PAGE_OFFSET. 534 * This runs before bootmem is initialized and gets pages directly from 535 * the physical memory. To access them they are temporarily mapped. 536 */ 537 unsigned long __ref init_memory_mapping(unsigned long start, 538 unsigned long end, pgprot_t prot) 539 { 540 struct map_range mr[NR_RANGE_MR]; 541 unsigned long ret = 0; 542 int nr_range, i; 543 544 pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n", 545 start, end - 1); 546 547 memset(mr, 0, sizeof(mr)); 548 nr_range = split_mem_range(mr, 0, start, end); 549 550 for (i = 0; i < nr_range; i++) 551 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end, 552 mr[i].page_size_mask, 553 prot); 554 555 add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT); 556 557 return ret >> PAGE_SHIFT; 558 } 559 560 /* 561 * We need to iterate through the E820 memory map and create direct mappings 562 * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply 563 * create direct mappings for all pfns from [0 to max_low_pfn) and 564 * [4GB to max_pfn) because of possible memory holes in high addresses 565 * that cannot be marked as UC by fixed/variable range MTRRs. 566 * Depending on the alignment of E820 ranges, this may possibly result 567 * in using smaller size (i.e. 4K instead of 2M or 1G) page tables. 568 * 569 * init_mem_mapping() calls init_range_memory_mapping() with big range. 570 * That range would have hole in the middle or ends, and only ram parts 571 * will be mapped in init_range_memory_mapping(). 572 */ 573 static unsigned long __init init_range_memory_mapping( 574 unsigned long r_start, 575 unsigned long r_end) 576 { 577 unsigned long start_pfn, end_pfn; 578 unsigned long mapped_ram_size = 0; 579 int i; 580 581 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) { 582 u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end); 583 u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end); 584 if (start >= end) 585 continue; 586 587 /* 588 * if it is overlapping with brk pgt, we need to 589 * alloc pgt buf from memblock instead. 590 */ 591 can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >= 592 min(end, (u64)pgt_buf_top<<PAGE_SHIFT); 593 init_memory_mapping(start, end, PAGE_KERNEL); 594 mapped_ram_size += end - start; 595 can_use_brk_pgt = true; 596 } 597 598 return mapped_ram_size; 599 } 600 601 static unsigned long __init get_new_step_size(unsigned long step_size) 602 { 603 /* 604 * Initial mapped size is PMD_SIZE (2M). 605 * We can not set step_size to be PUD_SIZE (1G) yet. 606 * In worse case, when we cross the 1G boundary, and 607 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k) 608 * to map 1G range with PTE. Hence we use one less than the 609 * difference of page table level shifts. 610 * 611 * Don't need to worry about overflow in the top-down case, on 32bit, 612 * when step_size is 0, round_down() returns 0 for start, and that 613 * turns it into 0x100000000ULL. 614 * In the bottom-up case, round_up(x, 0) returns 0 though too, which 615 * needs to be taken into consideration by the code below. 616 */ 617 return step_size << (PMD_SHIFT - PAGE_SHIFT - 1); 618 } 619 620 /** 621 * memory_map_top_down - Map [map_start, map_end) top down 622 * @map_start: start address of the target memory range 623 * @map_end: end address of the target memory range 624 * 625 * This function will setup direct mapping for memory range 626 * [map_start, map_end) in top-down. That said, the page tables 627 * will be allocated at the end of the memory, and we map the 628 * memory in top-down. 629 */ 630 static void __init memory_map_top_down(unsigned long map_start, 631 unsigned long map_end) 632 { 633 unsigned long real_end, last_start; 634 unsigned long step_size; 635 unsigned long addr; 636 unsigned long mapped_ram_size = 0; 637 638 /* 639 * Systems that have many reserved areas near top of the memory, 640 * e.g. QEMU with less than 1G RAM and EFI enabled, or Xen, will 641 * require lots of 4K mappings which may exhaust pgt_buf. 642 * Start with top-most PMD_SIZE range aligned at PMD_SIZE to ensure 643 * there is enough mapped memory that can be allocated from 644 * memblock. 645 */ 646 addr = memblock_phys_alloc_range(PMD_SIZE, PMD_SIZE, map_start, 647 map_end); 648 memblock_phys_free(addr, PMD_SIZE); 649 real_end = addr + PMD_SIZE; 650 651 /* step_size need to be small so pgt_buf from BRK could cover it */ 652 step_size = PMD_SIZE; 653 max_pfn_mapped = 0; /* will get exact value next */ 654 min_pfn_mapped = real_end >> PAGE_SHIFT; 655 last_start = real_end; 656 657 /* 658 * We start from the top (end of memory) and go to the bottom. 659 * The memblock_find_in_range() gets us a block of RAM from the 660 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages 661 * for page table. 662 */ 663 while (last_start > map_start) { 664 unsigned long start; 665 666 if (last_start > step_size) { 667 start = round_down(last_start - 1, step_size); 668 if (start < map_start) 669 start = map_start; 670 } else 671 start = map_start; 672 mapped_ram_size += init_range_memory_mapping(start, 673 last_start); 674 last_start = start; 675 min_pfn_mapped = last_start >> PAGE_SHIFT; 676 if (mapped_ram_size >= step_size) 677 step_size = get_new_step_size(step_size); 678 } 679 680 if (real_end < map_end) 681 init_range_memory_mapping(real_end, map_end); 682 } 683 684 /** 685 * memory_map_bottom_up - Map [map_start, map_end) bottom up 686 * @map_start: start address of the target memory range 687 * @map_end: end address of the target memory range 688 * 689 * This function will setup direct mapping for memory range 690 * [map_start, map_end) in bottom-up. Since we have limited the 691 * bottom-up allocation above the kernel, the page tables will 692 * be allocated just above the kernel and we map the memory 693 * in [map_start, map_end) in bottom-up. 694 */ 695 static void __init memory_map_bottom_up(unsigned long map_start, 696 unsigned long map_end) 697 { 698 unsigned long next, start; 699 unsigned long mapped_ram_size = 0; 700 /* step_size need to be small so pgt_buf from BRK could cover it */ 701 unsigned long step_size = PMD_SIZE; 702 703 start = map_start; 704 min_pfn_mapped = start >> PAGE_SHIFT; 705 706 /* 707 * We start from the bottom (@map_start) and go to the top (@map_end). 708 * The memblock_find_in_range() gets us a block of RAM from the 709 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages 710 * for page table. 711 */ 712 while (start < map_end) { 713 if (step_size && map_end - start > step_size) { 714 next = round_up(start + 1, step_size); 715 if (next > map_end) 716 next = map_end; 717 } else { 718 next = map_end; 719 } 720 721 mapped_ram_size += init_range_memory_mapping(start, next); 722 start = next; 723 724 if (mapped_ram_size >= step_size) 725 step_size = get_new_step_size(step_size); 726 } 727 } 728 729 /* 730 * The real mode trampoline, which is required for bootstrapping CPUs 731 * occupies only a small area under the low 1MB. See reserve_real_mode() 732 * for details. 733 * 734 * If KASLR is disabled the first PGD entry of the direct mapping is copied 735 * to map the real mode trampoline. 736 * 737 * If KASLR is enabled, copy only the PUD which covers the low 1MB 738 * area. This limits the randomization granularity to 1GB for both 4-level 739 * and 5-level paging. 740 */ 741 static void __init init_trampoline(void) 742 { 743 #ifdef CONFIG_X86_64 744 /* 745 * The code below will alias kernel page-tables in the user-range of the 746 * address space, including the Global bit. So global TLB entries will 747 * be created when using the trampoline page-table. 748 */ 749 if (!kaslr_memory_enabled()) 750 trampoline_pgd_entry = init_top_pgt[pgd_index(__PAGE_OFFSET)]; 751 else 752 init_trampoline_kaslr(); 753 #endif 754 } 755 756 void __init init_mem_mapping(void) 757 { 758 unsigned long end; 759 760 pti_check_boottime_disable(); 761 probe_page_size_mask(); 762 setup_pcid(); 763 764 #ifdef CONFIG_X86_64 765 end = max_pfn << PAGE_SHIFT; 766 #else 767 end = max_low_pfn << PAGE_SHIFT; 768 #endif 769 770 /* the ISA range is always mapped regardless of memory holes */ 771 init_memory_mapping(0, ISA_END_ADDRESS, PAGE_KERNEL); 772 773 /* Init the trampoline, possibly with KASLR memory offset */ 774 init_trampoline(); 775 776 /* 777 * If the allocation is in bottom-up direction, we setup direct mapping 778 * in bottom-up, otherwise we setup direct mapping in top-down. 779 */ 780 if (memblock_bottom_up()) { 781 unsigned long kernel_end = __pa_symbol(_end); 782 783 /* 784 * we need two separate calls here. This is because we want to 785 * allocate page tables above the kernel. So we first map 786 * [kernel_end, end) to make memory above the kernel be mapped 787 * as soon as possible. And then use page tables allocated above 788 * the kernel to map [ISA_END_ADDRESS, kernel_end). 789 */ 790 memory_map_bottom_up(kernel_end, end); 791 memory_map_bottom_up(ISA_END_ADDRESS, kernel_end); 792 } else { 793 memory_map_top_down(ISA_END_ADDRESS, end); 794 } 795 796 #ifdef CONFIG_X86_64 797 if (max_pfn > max_low_pfn) { 798 /* can we preserve max_low_pfn ?*/ 799 max_low_pfn = max_pfn; 800 } 801 #else 802 early_ioremap_page_table_range_init(); 803 #endif 804 805 load_cr3(swapper_pg_dir); 806 __flush_tlb_all(); 807 808 x86_init.hyper.init_mem_mapping(); 809 810 early_memtest(0, max_pfn_mapped << PAGE_SHIFT); 811 } 812 813 /* 814 * Initialize an mm_struct to be used during poking and a pointer to be used 815 * during patching. 816 */ 817 void __init poking_init(void) 818 { 819 spinlock_t *ptl; 820 pte_t *ptep; 821 822 poking_mm = mm_alloc(); 823 BUG_ON(!poking_mm); 824 825 /* Xen PV guests need the PGD to be pinned. */ 826 paravirt_enter_mmap(poking_mm); 827 828 /* 829 * Randomize the poking address, but make sure that the following page 830 * will be mapped at the same PMD. We need 2 pages, so find space for 3, 831 * and adjust the address if the PMD ends after the first one. 832 */ 833 poking_addr = TASK_UNMAPPED_BASE; 834 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) 835 poking_addr += (kaslr_get_random_long("Poking") & PAGE_MASK) % 836 (TASK_SIZE - TASK_UNMAPPED_BASE - 3 * PAGE_SIZE); 837 838 if (((poking_addr + PAGE_SIZE) & ~PMD_MASK) == 0) 839 poking_addr += PAGE_SIZE; 840 841 /* 842 * We need to trigger the allocation of the page-tables that will be 843 * needed for poking now. Later, poking may be performed in an atomic 844 * section, which might cause allocation to fail. 845 */ 846 ptep = get_locked_pte(poking_mm, poking_addr, &ptl); 847 BUG_ON(!ptep); 848 pte_unmap_unlock(ptep, ptl); 849 } 850 851 /* 852 * devmem_is_allowed() checks to see if /dev/mem access to a certain address 853 * is valid. The argument is a physical page number. 854 * 855 * On x86, access has to be given to the first megabyte of RAM because that 856 * area traditionally contains BIOS code and data regions used by X, dosemu, 857 * and similar apps. Since they map the entire memory range, the whole range 858 * must be allowed (for mapping), but any areas that would otherwise be 859 * disallowed are flagged as being "zero filled" instead of rejected. 860 * Access has to be given to non-kernel-ram areas as well, these contain the 861 * PCI mmio resources as well as potential bios/acpi data regions. 862 */ 863 int devmem_is_allowed(unsigned long pagenr) 864 { 865 if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE, 866 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE) 867 != REGION_DISJOINT) { 868 /* 869 * For disallowed memory regions in the low 1MB range, 870 * request that the page be shown as all zeros. 871 */ 872 if (pagenr < 256) 873 return 2; 874 875 return 0; 876 } 877 878 /* 879 * This must follow RAM test, since System RAM is considered a 880 * restricted resource under CONFIG_STRICT_DEVMEM. 881 */ 882 if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) { 883 /* Low 1MB bypasses iomem restrictions. */ 884 if (pagenr < 256) 885 return 1; 886 887 return 0; 888 } 889 890 return 1; 891 } 892 893 void free_init_pages(const char *what, unsigned long begin, unsigned long end) 894 { 895 unsigned long begin_aligned, end_aligned; 896 897 /* Make sure boundaries are page aligned */ 898 begin_aligned = PAGE_ALIGN(begin); 899 end_aligned = end & PAGE_MASK; 900 901 if (WARN_ON(begin_aligned != begin || end_aligned != end)) { 902 begin = begin_aligned; 903 end = end_aligned; 904 } 905 906 if (begin >= end) 907 return; 908 909 /* 910 * If debugging page accesses then do not free this memory but 911 * mark them not present - any buggy init-section access will 912 * create a kernel page fault: 913 */ 914 if (debug_pagealloc_enabled()) { 915 pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n", 916 begin, end - 1); 917 /* 918 * Inform kmemleak about the hole in the memory since the 919 * corresponding pages will be unmapped. 920 */ 921 kmemleak_free_part((void *)begin, end - begin); 922 set_memory_np(begin, (end - begin) >> PAGE_SHIFT); 923 } else { 924 /* 925 * We just marked the kernel text read only above, now that 926 * we are going to free part of that, we need to make that 927 * writeable and non-executable first. 928 */ 929 set_memory_nx(begin, (end - begin) >> PAGE_SHIFT); 930 set_memory_rw(begin, (end - begin) >> PAGE_SHIFT); 931 932 free_reserved_area((void *)begin, (void *)end, 933 POISON_FREE_INITMEM, what); 934 } 935 } 936 937 /* 938 * begin/end can be in the direct map or the "high kernel mapping" 939 * used for the kernel image only. free_init_pages() will do the 940 * right thing for either kind of address. 941 */ 942 void free_kernel_image_pages(const char *what, void *begin, void *end) 943 { 944 unsigned long begin_ul = (unsigned long)begin; 945 unsigned long end_ul = (unsigned long)end; 946 unsigned long len_pages = (end_ul - begin_ul) >> PAGE_SHIFT; 947 948 free_init_pages(what, begin_ul, end_ul); 949 950 /* 951 * PTI maps some of the kernel into userspace. For performance, 952 * this includes some kernel areas that do not contain secrets. 953 * Those areas might be adjacent to the parts of the kernel image 954 * being freed, which may contain secrets. Remove the "high kernel 955 * image mapping" for these freed areas, ensuring they are not even 956 * potentially vulnerable to Meltdown regardless of the specific 957 * optimizations PTI is currently using. 958 * 959 * The "noalias" prevents unmapping the direct map alias which is 960 * needed to access the freed pages. 961 * 962 * This is only valid for 64bit kernels. 32bit has only one mapping 963 * which can't be treated in this way for obvious reasons. 964 */ 965 if (IS_ENABLED(CONFIG_X86_64) && cpu_feature_enabled(X86_FEATURE_PTI)) 966 set_memory_np_noalias(begin_ul, len_pages); 967 } 968 969 void __ref free_initmem(void) 970 { 971 e820__reallocate_tables(); 972 973 mem_encrypt_free_decrypted_mem(); 974 975 free_kernel_image_pages("unused kernel image (initmem)", 976 &__init_begin, &__init_end); 977 } 978 979 #ifdef CONFIG_BLK_DEV_INITRD 980 void __init free_initrd_mem(unsigned long start, unsigned long end) 981 { 982 /* 983 * end could be not aligned, and We can not align that, 984 * decompressor could be confused by aligned initrd_end 985 * We already reserve the end partial page before in 986 * - i386_start_kernel() 987 * - x86_64_start_kernel() 988 * - relocate_initrd() 989 * So here We can do PAGE_ALIGN() safely to get partial page to be freed 990 */ 991 free_init_pages("initrd", start, PAGE_ALIGN(end)); 992 } 993 #endif 994 995 void __init zone_sizes_init(void) 996 { 997 unsigned long max_zone_pfns[MAX_NR_ZONES]; 998 999 memset(max_zone_pfns, 0, sizeof(max_zone_pfns)); 1000 1001 #ifdef CONFIG_ZONE_DMA 1002 max_zone_pfns[ZONE_DMA] = min(MAX_DMA_PFN, max_low_pfn); 1003 #endif 1004 #ifdef CONFIG_ZONE_DMA32 1005 max_zone_pfns[ZONE_DMA32] = min(MAX_DMA32_PFN, max_low_pfn); 1006 #endif 1007 max_zone_pfns[ZONE_NORMAL] = max_low_pfn; 1008 #ifdef CONFIG_HIGHMEM 1009 max_zone_pfns[ZONE_HIGHMEM] = max_pfn; 1010 #endif 1011 1012 free_area_init(max_zone_pfns); 1013 } 1014 1015 __visible DEFINE_PER_CPU_ALIGNED(struct tlb_state, cpu_tlbstate) = { 1016 .loaded_mm = &init_mm, 1017 .next_asid = 1, 1018 .cr4 = ~0UL, /* fail hard if we screw up cr4 shadow initialization */ 1019 }; 1020 1021 #ifdef CONFIG_ADDRESS_MASKING 1022 DEFINE_PER_CPU(u64, tlbstate_untag_mask); 1023 EXPORT_PER_CPU_SYMBOL(tlbstate_untag_mask); 1024 #endif 1025 1026 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache) 1027 { 1028 /* entry 0 MUST be WB (hardwired to speed up translations) */ 1029 BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB); 1030 1031 __cachemode2pte_tbl[cache] = __cm_idx2pte(entry); 1032 __pte2cachemode_tbl[entry] = cache; 1033 } 1034 1035 #ifdef CONFIG_SWAP 1036 unsigned long arch_max_swapfile_size(void) 1037 { 1038 unsigned long pages; 1039 1040 pages = generic_max_swapfile_size(); 1041 1042 if (boot_cpu_has_bug(X86_BUG_L1TF) && l1tf_mitigation != L1TF_MITIGATION_OFF) { 1043 /* Limit the swap file size to MAX_PA/2 for L1TF workaround */ 1044 unsigned long long l1tf_limit = l1tf_pfn_limit(); 1045 /* 1046 * We encode swap offsets also with 3 bits below those for pfn 1047 * which makes the usable limit higher. 1048 */ 1049 #if CONFIG_PGTABLE_LEVELS > 2 1050 l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT; 1051 #endif 1052 pages = min_t(unsigned long long, l1tf_limit, pages); 1053 } 1054 return pages; 1055 } 1056 #endif 1057 1058 #ifdef CONFIG_EXECMEM 1059 static struct execmem_info execmem_info __ro_after_init; 1060 1061 #ifdef CONFIG_ARCH_HAS_EXECMEM_ROX 1062 void execmem_fill_trapping_insns(void *ptr, size_t size, bool writeable) 1063 { 1064 /* fill memory with INT3 instructions */ 1065 if (writeable) 1066 memset(ptr, INT3_INSN_OPCODE, size); 1067 else 1068 text_poke_set(ptr, INT3_INSN_OPCODE, size); 1069 } 1070 #endif 1071 1072 struct execmem_info __init *execmem_arch_setup(void) 1073 { 1074 unsigned long start, offset = 0; 1075 enum execmem_range_flags flags; 1076 pgprot_t pgprot; 1077 1078 if (kaslr_enabled()) 1079 offset = get_random_u32_inclusive(1, 1024) * PAGE_SIZE; 1080 1081 start = MODULES_VADDR + offset; 1082 1083 if (IS_ENABLED(CONFIG_ARCH_HAS_EXECMEM_ROX)) { 1084 pgprot = PAGE_KERNEL_ROX; 1085 flags = EXECMEM_KASAN_SHADOW | EXECMEM_ROX_CACHE; 1086 } else { 1087 pgprot = PAGE_KERNEL; 1088 flags = EXECMEM_KASAN_SHADOW; 1089 } 1090 1091 execmem_info = (struct execmem_info){ 1092 .ranges = { 1093 [EXECMEM_MODULE_TEXT] = { 1094 .flags = flags, 1095 .start = start, 1096 .end = MODULES_END, 1097 .pgprot = pgprot, 1098 .alignment = MODULE_ALIGN, 1099 }, 1100 [EXECMEM_KPROBES ... EXECMEM_BPF] = { 1101 .flags = EXECMEM_KASAN_SHADOW, 1102 .start = start, 1103 .end = MODULES_END, 1104 .pgprot = PAGE_KERNEL, 1105 .alignment = MODULE_ALIGN, 1106 }, 1107 [EXECMEM_MODULE_DATA] = { 1108 .flags = EXECMEM_KASAN_SHADOW, 1109 .start = start, 1110 .end = MODULES_END, 1111 .pgprot = PAGE_KERNEL, 1112 .alignment = MODULE_ALIGN, 1113 }, 1114 }, 1115 }; 1116 1117 return &execmem_info; 1118 } 1119 #endif /* CONFIG_EXECMEM */ 1120