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 #include <asm/mmu_context.h> 32 33 /* 34 * We need to define the tracepoints somewhere, and tlb.c 35 * is only compiled when SMP=y. 36 */ 37 #define CREATE_TRACE_POINTS 38 #include <trace/events/tlb.h> 39 40 #include "mm_internal.h" 41 42 /* 43 * Tables translating between page_cache_type_t and pte encoding. 44 * 45 * The default values are defined statically as minimal supported mode; 46 * WC and WT fall back to UC-. pat_init() updates these values to support 47 * more cache modes, WC and WT, when it is safe to do so. See pat_init() 48 * for the details. Note, __early_ioremap() used during early boot-time 49 * takes pgprot_t (pte encoding) and does not use these tables. 50 * 51 * Index into __cachemode2pte_tbl[] is the cachemode. 52 * 53 * Index into __pte2cachemode_tbl[] are the caching attribute bits of the pte 54 * (_PAGE_PWT, _PAGE_PCD, _PAGE_PAT) at index bit positions 0, 1, 2. 55 */ 56 static uint16_t __cachemode2pte_tbl[_PAGE_CACHE_MODE_NUM] = { 57 [_PAGE_CACHE_MODE_WB ] = 0 | 0 , 58 [_PAGE_CACHE_MODE_WC ] = 0 | _PAGE_PCD, 59 [_PAGE_CACHE_MODE_UC_MINUS] = 0 | _PAGE_PCD, 60 [_PAGE_CACHE_MODE_UC ] = _PAGE_PWT | _PAGE_PCD, 61 [_PAGE_CACHE_MODE_WT ] = 0 | _PAGE_PCD, 62 [_PAGE_CACHE_MODE_WP ] = 0 | _PAGE_PCD, 63 }; 64 65 unsigned long cachemode2protval(enum page_cache_mode pcm) 66 { 67 if (likely(pcm == 0)) 68 return 0; 69 return __cachemode2pte_tbl[pcm]; 70 } 71 EXPORT_SYMBOL(cachemode2protval); 72 73 static uint8_t __pte2cachemode_tbl[8] = { 74 [__pte2cm_idx( 0 | 0 | 0 )] = _PAGE_CACHE_MODE_WB, 75 [__pte2cm_idx(_PAGE_PWT | 0 | 0 )] = _PAGE_CACHE_MODE_UC_MINUS, 76 [__pte2cm_idx( 0 | _PAGE_PCD | 0 )] = _PAGE_CACHE_MODE_UC_MINUS, 77 [__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | 0 )] = _PAGE_CACHE_MODE_UC, 78 [__pte2cm_idx( 0 | 0 | _PAGE_PAT)] = _PAGE_CACHE_MODE_WB, 79 [__pte2cm_idx(_PAGE_PWT | 0 | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS, 80 [__pte2cm_idx(0 | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC_MINUS, 81 [__pte2cm_idx(_PAGE_PWT | _PAGE_PCD | _PAGE_PAT)] = _PAGE_CACHE_MODE_UC, 82 }; 83 84 /* 85 * Check that the write-protect PAT entry is set for write-protect. 86 * To do this without making assumptions how PAT has been set up (Xen has 87 * another layout than the kernel), translate the _PAGE_CACHE_MODE_WP cache 88 * mode via the __cachemode2pte_tbl[] into protection bits (those protection 89 * bits will select a cache mode of WP or better), and then translate the 90 * protection bits back into the cache mode using __pte2cm_idx() and the 91 * __pte2cachemode_tbl[] array. This will return the really used cache mode. 92 */ 93 bool x86_has_pat_wp(void) 94 { 95 uint16_t prot = __cachemode2pte_tbl[_PAGE_CACHE_MODE_WP]; 96 97 return __pte2cachemode_tbl[__pte2cm_idx(prot)] == _PAGE_CACHE_MODE_WP; 98 } 99 100 enum page_cache_mode pgprot2cachemode(pgprot_t pgprot) 101 { 102 unsigned long masked; 103 104 masked = pgprot_val(pgprot) & _PAGE_CACHE_MASK; 105 if (likely(masked == 0)) 106 return 0; 107 return __pte2cachemode_tbl[__pte2cm_idx(masked)]; 108 } 109 110 static unsigned long __initdata pgt_buf_start; 111 static unsigned long __initdata pgt_buf_end; 112 static unsigned long __initdata pgt_buf_top; 113 114 static unsigned long min_pfn_mapped; 115 116 static bool __initdata can_use_brk_pgt = true; 117 118 /* 119 * Pages returned are already directly mapped. 120 * 121 * Changing that is likely to break Xen, see commit: 122 * 123 * 279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve 124 * 125 * for detailed information. 126 */ 127 __ref void *alloc_low_pages(unsigned int num) 128 { 129 unsigned long pfn; 130 int i; 131 132 if (after_bootmem) { 133 unsigned int order; 134 135 order = get_order((unsigned long)num << PAGE_SHIFT); 136 return (void *)__get_free_pages(GFP_ATOMIC | __GFP_ZERO, order); 137 } 138 139 if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) { 140 unsigned long ret = 0; 141 142 if (min_pfn_mapped < max_pfn_mapped) { 143 ret = memblock_phys_alloc_range( 144 PAGE_SIZE * num, PAGE_SIZE, 145 min_pfn_mapped << PAGE_SHIFT, 146 max_pfn_mapped << PAGE_SHIFT); 147 } 148 if (!ret && can_use_brk_pgt) 149 ret = __pa(extend_brk(PAGE_SIZE * num, PAGE_SIZE)); 150 151 if (!ret) 152 panic("alloc_low_pages: can not alloc memory"); 153 154 pfn = ret >> PAGE_SHIFT; 155 } else { 156 pfn = pgt_buf_end; 157 pgt_buf_end += num; 158 } 159 160 for (i = 0; i < num; i++) { 161 void *adr; 162 163 adr = __va((pfn + i) << PAGE_SHIFT); 164 clear_page(adr); 165 } 166 167 return __va(pfn << PAGE_SHIFT); 168 } 169 170 /* 171 * By default need to be able to allocate page tables below PGD firstly for 172 * the 0-ISA_END_ADDRESS range and secondly for the initial PMD_SIZE mapping. 173 * With KASLR memory randomization, depending on the machine e820 memory and the 174 * PUD alignment, twice that many pages may be needed when KASLR memory 175 * randomization is enabled. 176 */ 177 178 #define INIT_PGD_PAGE_TABLES 4 179 180 #ifndef CONFIG_RANDOMIZE_MEMORY 181 #define INIT_PGD_PAGE_COUNT (2 * INIT_PGD_PAGE_TABLES) 182 #else 183 #define INIT_PGD_PAGE_COUNT (4 * INIT_PGD_PAGE_TABLES) 184 #endif 185 186 #define INIT_PGT_BUF_SIZE (INIT_PGD_PAGE_COUNT * PAGE_SIZE) 187 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE); 188 void __init early_alloc_pgt_buf(void) 189 { 190 unsigned long tables = INIT_PGT_BUF_SIZE; 191 phys_addr_t base; 192 193 base = __pa(extend_brk(tables, PAGE_SIZE)); 194 195 pgt_buf_start = base >> PAGE_SHIFT; 196 pgt_buf_end = pgt_buf_start; 197 pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT); 198 } 199 200 int after_bootmem; 201 202 early_param_on_off("gbpages", "nogbpages", direct_gbpages, CONFIG_X86_DIRECT_GBPAGES); 203 204 struct map_range { 205 unsigned long start; 206 unsigned long end; 207 unsigned page_size_mask; 208 }; 209 210 static int page_size_mask; 211 212 /* 213 * Save some of cr4 feature set we're using (e.g. Pentium 4MB 214 * enable and PPro Global page enable), so that any CPU's that boot 215 * up after us can get the correct flags. Invoked on the boot CPU. 216 */ 217 static inline void cr4_set_bits_and_update_boot(unsigned long mask) 218 { 219 mmu_cr4_features |= mask; 220 if (trampoline_cr4_features) 221 *trampoline_cr4_features = mmu_cr4_features; 222 cr4_set_bits(mask); 223 } 224 225 static void __init probe_page_size_mask(void) 226 { 227 /* 228 * For pagealloc debugging, identity mapping will use small pages. 229 * This will simplify cpa(), which otherwise needs to support splitting 230 * large pages into small in interrupt context, etc. 231 */ 232 if (boot_cpu_has(X86_FEATURE_PSE) && !debug_pagealloc_enabled()) 233 page_size_mask |= 1 << PG_LEVEL_2M; 234 else 235 direct_gbpages = 0; 236 237 /* Enable PSE if available */ 238 if (boot_cpu_has(X86_FEATURE_PSE)) 239 cr4_set_bits_and_update_boot(X86_CR4_PSE); 240 241 /* Enable PGE if available */ 242 __supported_pte_mask &= ~_PAGE_GLOBAL; 243 if (boot_cpu_has(X86_FEATURE_PGE)) { 244 cr4_set_bits_and_update_boot(X86_CR4_PGE); 245 __supported_pte_mask |= _PAGE_GLOBAL; 246 } 247 248 /* By the default is everything supported: */ 249 __default_kernel_pte_mask = __supported_pte_mask; 250 /* Except when with PTI where the kernel is mostly non-Global: */ 251 if (cpu_feature_enabled(X86_FEATURE_PTI)) 252 __default_kernel_pte_mask &= ~_PAGE_GLOBAL; 253 254 /* Enable 1 GB linear kernel mappings if available: */ 255 if (direct_gbpages && boot_cpu_has(X86_FEATURE_GBPAGES)) { 256 printk(KERN_INFO "Using GB pages for direct mapping\n"); 257 page_size_mask |= 1 << PG_LEVEL_1G; 258 } else { 259 direct_gbpages = 0; 260 } 261 } 262 263 /* 264 * INVLPG may not properly flush Global entries on 265 * these CPUs. New microcode fixes the issue. 266 */ 267 static const struct x86_cpu_id invlpg_miss_ids[] = { 268 X86_MATCH_VFM(INTEL_ALDERLAKE, 0x2e), 269 X86_MATCH_VFM(INTEL_ALDERLAKE_L, 0x42c), 270 X86_MATCH_VFM(INTEL_ATOM_GRACEMONT, 0x11), 271 X86_MATCH_VFM(INTEL_RAPTORLAKE, 0x118), 272 X86_MATCH_VFM(INTEL_RAPTORLAKE_P, 0x4117), 273 X86_MATCH_VFM(INTEL_RAPTORLAKE_S, 0x2e), 274 {} 275 }; 276 277 static void setup_pcid(void) 278 { 279 const struct x86_cpu_id *invlpg_miss_match; 280 281 if (!IS_ENABLED(CONFIG_X86_64)) 282 return; 283 284 if (!boot_cpu_has(X86_FEATURE_PCID)) 285 return; 286 287 invlpg_miss_match = x86_match_cpu(invlpg_miss_ids); 288 289 if (invlpg_miss_match && 290 boot_cpu_data.microcode < invlpg_miss_match->driver_data) { 291 pr_info("Incomplete global flushes, disabling PCID"); 292 setup_clear_cpu_cap(X86_FEATURE_PCID); 293 return; 294 } 295 296 if (boot_cpu_has(X86_FEATURE_PGE)) { 297 /* 298 * This can't be cr4_set_bits_and_update_boot() -- the 299 * trampoline code can't handle CR4.PCIDE and it wouldn't 300 * do any good anyway. Despite the name, 301 * cr4_set_bits_and_update_boot() doesn't actually cause 302 * the bits in question to remain set all the way through 303 * the secondary boot asm. 304 * 305 * Instead, we brute-force it and set CR4.PCIDE manually in 306 * start_secondary(). 307 */ 308 cr4_set_bits(X86_CR4_PCIDE); 309 } else { 310 /* 311 * flush_tlb_all(), as currently implemented, won't work if 312 * PCID is on but PGE is not. Since that combination 313 * doesn't exist on real hardware, there's no reason to try 314 * to fully support it, but it's polite to avoid corrupting 315 * data if we're on an improperly configured VM. 316 */ 317 setup_clear_cpu_cap(X86_FEATURE_PCID); 318 } 319 } 320 321 #ifdef CONFIG_X86_32 322 #define NR_RANGE_MR 3 323 #else /* CONFIG_X86_64 */ 324 #define NR_RANGE_MR 5 325 #endif 326 327 static int __meminit save_mr(struct map_range *mr, int nr_range, 328 unsigned long start_pfn, unsigned long end_pfn, 329 unsigned long page_size_mask) 330 { 331 if (start_pfn < end_pfn) { 332 if (nr_range >= NR_RANGE_MR) 333 panic("run out of range for init_memory_mapping\n"); 334 mr[nr_range].start = start_pfn<<PAGE_SHIFT; 335 mr[nr_range].end = end_pfn<<PAGE_SHIFT; 336 mr[nr_range].page_size_mask = page_size_mask; 337 nr_range++; 338 } 339 340 return nr_range; 341 } 342 343 /* 344 * adjust the page_size_mask for small range to go with 345 * big page size instead small one if nearby are ram too. 346 */ 347 static void __ref adjust_range_page_size_mask(struct map_range *mr, 348 int nr_range) 349 { 350 int i; 351 352 for (i = 0; i < nr_range; i++) { 353 if ((page_size_mask & (1<<PG_LEVEL_2M)) && 354 !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) { 355 unsigned long start = round_down(mr[i].start, PMD_SIZE); 356 unsigned long end = round_up(mr[i].end, PMD_SIZE); 357 358 #ifdef CONFIG_X86_32 359 if ((end >> PAGE_SHIFT) > max_low_pfn) 360 continue; 361 #endif 362 363 if (memblock_is_region_memory(start, end - start)) 364 mr[i].page_size_mask |= 1<<PG_LEVEL_2M; 365 } 366 if ((page_size_mask & (1<<PG_LEVEL_1G)) && 367 !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) { 368 unsigned long start = round_down(mr[i].start, PUD_SIZE); 369 unsigned long end = round_up(mr[i].end, PUD_SIZE); 370 371 if (memblock_is_region_memory(start, end - start)) 372 mr[i].page_size_mask |= 1<<PG_LEVEL_1G; 373 } 374 } 375 } 376 377 static const char *page_size_string(struct map_range *mr) 378 { 379 static const char str_1g[] = "1G"; 380 static const char str_2m[] = "2M"; 381 static const char str_4m[] = "4M"; 382 static const char str_4k[] = "4k"; 383 384 if (mr->page_size_mask & (1<<PG_LEVEL_1G)) 385 return str_1g; 386 /* 387 * 32-bit without PAE has a 4M large page size. 388 * PG_LEVEL_2M is misnamed, but we can at least 389 * print out the right size in the string. 390 */ 391 if (IS_ENABLED(CONFIG_X86_32) && 392 !IS_ENABLED(CONFIG_X86_PAE) && 393 mr->page_size_mask & (1<<PG_LEVEL_2M)) 394 return str_4m; 395 396 if (mr->page_size_mask & (1<<PG_LEVEL_2M)) 397 return str_2m; 398 399 return str_4k; 400 } 401 402 static int __meminit split_mem_range(struct map_range *mr, int nr_range, 403 unsigned long start, 404 unsigned long end) 405 { 406 unsigned long start_pfn, end_pfn, limit_pfn; 407 unsigned long pfn; 408 int i; 409 410 limit_pfn = PFN_DOWN(end); 411 412 /* head if not big page alignment ? */ 413 pfn = start_pfn = PFN_DOWN(start); 414 #ifdef CONFIG_X86_32 415 /* 416 * Don't use a large page for the first 2/4MB of memory 417 * because there are often fixed size MTRRs in there 418 * and overlapping MTRRs into large pages can cause 419 * slowdowns. 420 */ 421 if (pfn == 0) 422 end_pfn = PFN_DOWN(PMD_SIZE); 423 else 424 end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 425 #else /* CONFIG_X86_64 */ 426 end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 427 #endif 428 if (end_pfn > limit_pfn) 429 end_pfn = limit_pfn; 430 if (start_pfn < end_pfn) { 431 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0); 432 pfn = end_pfn; 433 } 434 435 /* big page (2M) range */ 436 start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 437 #ifdef CONFIG_X86_32 438 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE)); 439 #else /* CONFIG_X86_64 */ 440 end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE)); 441 if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE))) 442 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE)); 443 #endif 444 445 if (start_pfn < end_pfn) { 446 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 447 page_size_mask & (1<<PG_LEVEL_2M)); 448 pfn = end_pfn; 449 } 450 451 #ifdef CONFIG_X86_64 452 /* big page (1G) range */ 453 start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE)); 454 end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE)); 455 if (start_pfn < end_pfn) { 456 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 457 page_size_mask & 458 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G))); 459 pfn = end_pfn; 460 } 461 462 /* tail is not big page (1G) alignment */ 463 start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE)); 464 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE)); 465 if (start_pfn < end_pfn) { 466 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 467 page_size_mask & (1<<PG_LEVEL_2M)); 468 pfn = end_pfn; 469 } 470 #endif 471 472 /* tail is not big page (2M) alignment */ 473 start_pfn = pfn; 474 end_pfn = limit_pfn; 475 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0); 476 477 if (!after_bootmem) 478 adjust_range_page_size_mask(mr, nr_range); 479 480 /* try to merge same page size and continuous */ 481 for (i = 0; nr_range > 1 && i < nr_range - 1; i++) { 482 unsigned long old_start; 483 if (mr[i].end != mr[i+1].start || 484 mr[i].page_size_mask != mr[i+1].page_size_mask) 485 continue; 486 /* move it */ 487 old_start = mr[i].start; 488 memmove(&mr[i], &mr[i+1], 489 (nr_range - 1 - i) * sizeof(struct map_range)); 490 mr[i--].start = old_start; 491 nr_range--; 492 } 493 494 for (i = 0; i < nr_range; i++) 495 pr_debug(" [mem %#010lx-%#010lx] page %s\n", 496 mr[i].start, mr[i].end - 1, 497 page_size_string(&mr[i])); 498 499 return nr_range; 500 } 501 502 struct range pfn_mapped[E820_MAX_ENTRIES]; 503 int nr_pfn_mapped; 504 505 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn) 506 { 507 nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_MAX_ENTRIES, 508 nr_pfn_mapped, start_pfn, end_pfn); 509 nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_MAX_ENTRIES); 510 511 max_pfn_mapped = max(max_pfn_mapped, end_pfn); 512 513 if (start_pfn < (1UL<<(32-PAGE_SHIFT))) 514 max_low_pfn_mapped = max(max_low_pfn_mapped, 515 min(end_pfn, 1UL<<(32-PAGE_SHIFT))); 516 } 517 518 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn) 519 { 520 int i; 521 522 for (i = 0; i < nr_pfn_mapped; i++) 523 if ((start_pfn >= pfn_mapped[i].start) && 524 (end_pfn <= pfn_mapped[i].end)) 525 return true; 526 527 return false; 528 } 529 530 /* 531 * Setup the direct mapping of the physical memory at PAGE_OFFSET. 532 * This runs before bootmem is initialized and gets pages directly from 533 * the physical memory. To access them they are temporarily mapped. 534 */ 535 unsigned long __ref init_memory_mapping(unsigned long start, 536 unsigned long end, pgprot_t prot) 537 { 538 struct map_range mr[NR_RANGE_MR]; 539 unsigned long ret = 0; 540 int nr_range, i; 541 542 pr_debug("init_memory_mapping: [mem %#010lx-%#010lx]\n", 543 start, end - 1); 544 545 memset(mr, 0, sizeof(mr)); 546 nr_range = split_mem_range(mr, 0, start, end); 547 548 for (i = 0; i < nr_range; i++) 549 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end, 550 mr[i].page_size_mask, 551 prot); 552 553 add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT); 554 555 return ret >> PAGE_SHIFT; 556 } 557 558 /* 559 * We need to iterate through the E820 memory map and create direct mappings 560 * for only E820_TYPE_RAM and E820_KERN_RESERVED regions. We cannot simply 561 * create direct mappings for all pfns from [0 to max_low_pfn) and 562 * [4GB to max_pfn) because of possible memory holes in high addresses 563 * that cannot be marked as UC by fixed/variable range MTRRs. 564 * Depending on the alignment of E820 ranges, this may possibly result 565 * in using smaller size (i.e. 4K instead of 2M or 1G) page tables. 566 * 567 * init_mem_mapping() calls init_range_memory_mapping() with big range. 568 * That range would have hole in the middle or ends, and only ram parts 569 * will be mapped in init_range_memory_mapping(). 570 */ 571 static unsigned long __init init_range_memory_mapping( 572 unsigned long r_start, 573 unsigned long r_end) 574 { 575 unsigned long start_pfn, end_pfn; 576 unsigned long mapped_ram_size = 0; 577 int i; 578 579 for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) { 580 u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end); 581 u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end); 582 if (start >= end) 583 continue; 584 585 /* 586 * if it is overlapping with brk pgt, we need to 587 * alloc pgt buf from memblock instead. 588 */ 589 can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >= 590 min(end, (u64)pgt_buf_top<<PAGE_SHIFT); 591 init_memory_mapping(start, end, PAGE_KERNEL); 592 mapped_ram_size += end - start; 593 can_use_brk_pgt = true; 594 } 595 596 return mapped_ram_size; 597 } 598 599 static unsigned long __init get_new_step_size(unsigned long step_size) 600 { 601 /* 602 * Initial mapped size is PMD_SIZE (2M). 603 * We can not set step_size to be PUD_SIZE (1G) yet. 604 * In worse case, when we cross the 1G boundary, and 605 * PG_LEVEL_2M is not set, we will need 1+1+512 pages (2M + 8k) 606 * to map 1G range with PTE. Hence we use one less than the 607 * difference of page table level shifts. 608 * 609 * Don't need to worry about overflow in the top-down case, on 32bit, 610 * when step_size is 0, round_down() returns 0 for start, and that 611 * turns it into 0x100000000ULL. 612 * In the bottom-up case, round_up(x, 0) returns 0 though too, which 613 * needs to be taken into consideration by the code below. 614 */ 615 return step_size << (PMD_SHIFT - PAGE_SHIFT - 1); 616 } 617 618 /** 619 * memory_map_top_down - Map [map_start, map_end) top down 620 * @map_start: start address of the target memory range 621 * @map_end: end address of the target memory range 622 * 623 * This function will setup direct mapping for memory range 624 * [map_start, map_end) in top-down. That said, the page tables 625 * will be allocated at the end of the memory, and we map the 626 * memory in top-down. 627 */ 628 static void __init memory_map_top_down(unsigned long map_start, 629 unsigned long map_end) 630 { 631 unsigned long real_end, last_start; 632 unsigned long step_size; 633 unsigned long addr; 634 unsigned long mapped_ram_size = 0; 635 636 /* 637 * Systems that have many reserved areas near top of the memory, 638 * e.g. QEMU with less than 1G RAM and EFI enabled, or Xen, will 639 * require lots of 4K mappings which may exhaust pgt_buf. 640 * Start with top-most PMD_SIZE range aligned at PMD_SIZE to ensure 641 * there is enough mapped memory that can be allocated from 642 * memblock. 643 */ 644 addr = memblock_phys_alloc_range(PMD_SIZE, PMD_SIZE, map_start, 645 map_end); 646 if (!addr) { 647 pr_warn("Failed to release memory for alloc_low_pages()"); 648 real_end = max(map_start, ALIGN_DOWN(map_end, PMD_SIZE)); 649 } else { 650 memblock_phys_free(addr, PMD_SIZE); 651 real_end = addr + PMD_SIZE; 652 } 653 654 /* step_size need to be small so pgt_buf from BRK could cover it */ 655 step_size = PMD_SIZE; 656 max_pfn_mapped = 0; /* will get exact value next */ 657 min_pfn_mapped = real_end >> PAGE_SHIFT; 658 last_start = real_end; 659 660 /* 661 * We start from the top (end of memory) and go to the bottom. 662 * The memblock_find_in_range() gets us a block of RAM from the 663 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages 664 * for page table. 665 */ 666 while (last_start > map_start) { 667 unsigned long start; 668 669 if (last_start > step_size) { 670 start = round_down(last_start - 1, step_size); 671 if (start < map_start) 672 start = map_start; 673 } else 674 start = map_start; 675 mapped_ram_size += init_range_memory_mapping(start, 676 last_start); 677 last_start = start; 678 min_pfn_mapped = last_start >> PAGE_SHIFT; 679 if (mapped_ram_size >= step_size) 680 step_size = get_new_step_size(step_size); 681 } 682 683 if (real_end < map_end) 684 init_range_memory_mapping(real_end, map_end); 685 } 686 687 /** 688 * memory_map_bottom_up - Map [map_start, map_end) bottom up 689 * @map_start: start address of the target memory range 690 * @map_end: end address of the target memory range 691 * 692 * This function will setup direct mapping for memory range 693 * [map_start, map_end) in bottom-up. Since we have limited the 694 * bottom-up allocation above the kernel, the page tables will 695 * be allocated just above the kernel and we map the memory 696 * in [map_start, map_end) in bottom-up. 697 */ 698 static void __init memory_map_bottom_up(unsigned long map_start, 699 unsigned long map_end) 700 { 701 unsigned long next, start; 702 unsigned long mapped_ram_size = 0; 703 /* step_size need to be small so pgt_buf from BRK could cover it */ 704 unsigned long step_size = PMD_SIZE; 705 706 start = map_start; 707 min_pfn_mapped = start >> PAGE_SHIFT; 708 709 /* 710 * We start from the bottom (@map_start) and go to the top (@map_end). 711 * The memblock_find_in_range() gets us a block of RAM from the 712 * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages 713 * for page table. 714 */ 715 while (start < map_end) { 716 if (step_size && map_end - start > step_size) { 717 next = round_up(start + 1, step_size); 718 if (next > map_end) 719 next = map_end; 720 } else { 721 next = map_end; 722 } 723 724 mapped_ram_size += init_range_memory_mapping(start, next); 725 start = next; 726 727 if (mapped_ram_size >= step_size) 728 step_size = get_new_step_size(step_size); 729 } 730 } 731 732 /* 733 * The real mode trampoline, which is required for bootstrapping CPUs 734 * occupies only a small area under the low 1MB. See reserve_real_mode() 735 * for details. 736 * 737 * If KASLR is disabled the first PGD entry of the direct mapping is copied 738 * to map the real mode trampoline. 739 * 740 * If KASLR is enabled, copy only the PUD which covers the low 1MB 741 * area. This limits the randomization granularity to 1GB for both 4-level 742 * and 5-level paging. 743 */ 744 static void __init init_trampoline(void) 745 { 746 #ifdef CONFIG_X86_64 747 /* 748 * The code below will alias kernel page-tables in the user-range of the 749 * address space, including the Global bit. So global TLB entries will 750 * be created when using the trampoline page-table. 751 */ 752 if (!kaslr_memory_enabled()) 753 trampoline_pgd_entry = init_top_pgt[pgd_index(__PAGE_OFFSET)]; 754 else 755 init_trampoline_kaslr(); 756 #endif 757 } 758 759 void __init init_mem_mapping(void) 760 { 761 unsigned long end; 762 763 pti_check_boottime_disable(); 764 probe_page_size_mask(); 765 setup_pcid(); 766 767 #ifdef CONFIG_X86_64 768 end = max_pfn << PAGE_SHIFT; 769 #else 770 end = max_low_pfn << PAGE_SHIFT; 771 #endif 772 773 /* the ISA range is always mapped regardless of memory holes */ 774 init_memory_mapping(0, ISA_END_ADDRESS, PAGE_KERNEL); 775 776 /* Init the trampoline, possibly with KASLR memory offset */ 777 init_trampoline(); 778 779 /* 780 * If the allocation is in bottom-up direction, we setup direct mapping 781 * in bottom-up, otherwise we setup direct mapping in top-down. 782 */ 783 if (memblock_bottom_up()) { 784 unsigned long kernel_end = __pa_symbol(_end); 785 786 /* 787 * we need two separate calls here. This is because we want to 788 * allocate page tables above the kernel. So we first map 789 * [kernel_end, end) to make memory above the kernel be mapped 790 * as soon as possible. And then use page tables allocated above 791 * the kernel to map [ISA_END_ADDRESS, kernel_end). 792 */ 793 memory_map_bottom_up(kernel_end, end); 794 memory_map_bottom_up(ISA_END_ADDRESS, kernel_end); 795 } else { 796 memory_map_top_down(ISA_END_ADDRESS, end); 797 } 798 799 #ifdef CONFIG_X86_64 800 if (max_pfn > max_low_pfn) { 801 /* can we preserve max_low_pfn ?*/ 802 max_low_pfn = max_pfn; 803 } 804 #else 805 early_ioremap_page_table_range_init(); 806 #endif 807 808 load_cr3(swapper_pg_dir); 809 __flush_tlb_all(); 810 811 x86_init.hyper.init_mem_mapping(); 812 813 early_memtest(0, max_pfn_mapped << PAGE_SHIFT); 814 } 815 816 /* 817 * Initialize an mm_struct to be used during poking and a pointer to be used 818 * during patching. 819 */ 820 void __init poking_init(void) 821 { 822 spinlock_t *ptl; 823 pte_t *ptep; 824 825 text_poke_mm = mm_alloc(); 826 BUG_ON(!text_poke_mm); 827 828 /* Xen PV guests need the PGD to be pinned. */ 829 paravirt_enter_mmap(text_poke_mm); 830 831 set_notrack_mm(text_poke_mm); 832 833 /* 834 * Randomize the poking address, but make sure that the following page 835 * will be mapped at the same PMD. We need 2 pages, so find space for 3, 836 * and adjust the address if the PMD ends after the first one. 837 */ 838 text_poke_mm_addr = TASK_UNMAPPED_BASE; 839 if (IS_ENABLED(CONFIG_RANDOMIZE_BASE)) 840 text_poke_mm_addr += (kaslr_get_random_long("Poking") & PAGE_MASK) % 841 (TASK_SIZE - TASK_UNMAPPED_BASE - 3 * PAGE_SIZE); 842 843 if (((text_poke_mm_addr + PAGE_SIZE) & ~PMD_MASK) == 0) 844 text_poke_mm_addr += PAGE_SIZE; 845 846 /* 847 * We need to trigger the allocation of the page-tables that will be 848 * needed for poking now. Later, poking may be performed in an atomic 849 * section, which might cause allocation to fail. 850 */ 851 ptep = get_locked_pte(text_poke_mm, text_poke_mm_addr, &ptl); 852 BUG_ON(!ptep); 853 pte_unmap_unlock(ptep, ptl); 854 } 855 856 /* 857 * devmem_is_allowed() checks to see if /dev/mem access to a certain address 858 * is valid. The argument is a physical page number. 859 * 860 * On x86, access has to be given to the first megabyte of RAM because that 861 * area traditionally contains BIOS code and data regions used by X, dosemu, 862 * and similar apps. Since they map the entire memory range, the whole range 863 * must be allowed (for mapping), but any areas that would otherwise be 864 * disallowed are flagged as being "zero filled" instead of rejected. 865 * Access has to be given to non-kernel-ram areas as well, these contain the 866 * PCI mmio resources as well as potential bios/acpi data regions. 867 */ 868 int devmem_is_allowed(unsigned long pagenr) 869 { 870 if (region_intersects(PFN_PHYS(pagenr), PAGE_SIZE, 871 IORESOURCE_SYSTEM_RAM, IORES_DESC_NONE) 872 != REGION_DISJOINT) { 873 /* 874 * For disallowed memory regions in the low 1MB range, 875 * request that the page be shown as all zeros. 876 */ 877 if (pagenr < 256) 878 return 2; 879 880 return 0; 881 } 882 883 /* 884 * This must follow RAM test, since System RAM is considered a 885 * restricted resource under CONFIG_STRICT_DEVMEM. 886 */ 887 if (iomem_is_exclusive(pagenr << PAGE_SHIFT)) { 888 /* Low 1MB bypasses iomem restrictions. */ 889 if (pagenr < 256) 890 return 1; 891 892 return 0; 893 } 894 895 return 1; 896 } 897 898 void free_init_pages(const char *what, unsigned long begin, unsigned long end) 899 { 900 unsigned long begin_aligned, end_aligned; 901 902 /* Make sure boundaries are page aligned */ 903 begin_aligned = PAGE_ALIGN(begin); 904 end_aligned = end & PAGE_MASK; 905 906 if (WARN_ON(begin_aligned != begin || end_aligned != end)) { 907 begin = begin_aligned; 908 end = end_aligned; 909 } 910 911 if (begin >= end) 912 return; 913 914 /* 915 * If debugging page accesses then do not free this memory but 916 * mark them not present - any buggy init-section access will 917 * create a kernel page fault: 918 */ 919 if (debug_pagealloc_enabled()) { 920 pr_info("debug: unmapping init [mem %#010lx-%#010lx]\n", 921 begin, end - 1); 922 /* 923 * Inform kmemleak about the hole in the memory since the 924 * corresponding pages will be unmapped. 925 */ 926 kmemleak_free_part((void *)begin, end - begin); 927 set_memory_np(begin, (end - begin) >> PAGE_SHIFT); 928 } else { 929 /* 930 * We just marked the kernel text read only above, now that 931 * we are going to free part of that, we need to make that 932 * writeable and non-executable first. 933 */ 934 set_memory_nx(begin, (end - begin) >> PAGE_SHIFT); 935 set_memory_rw(begin, (end - begin) >> PAGE_SHIFT); 936 937 free_reserved_area((void *)begin, (void *)end, 938 POISON_FREE_INITMEM, what); 939 } 940 } 941 942 /* 943 * begin/end can be in the direct map or the "high kernel mapping" 944 * used for the kernel image only. free_init_pages() will do the 945 * right thing for either kind of address. 946 */ 947 void free_kernel_image_pages(const char *what, void *begin, void *end) 948 { 949 unsigned long begin_ul = (unsigned long)begin; 950 unsigned long end_ul = (unsigned long)end; 951 unsigned long len_pages = (end_ul - begin_ul) >> PAGE_SHIFT; 952 953 free_init_pages(what, begin_ul, end_ul); 954 955 /* 956 * PTI maps some of the kernel into userspace. For performance, 957 * this includes some kernel areas that do not contain secrets. 958 * Those areas might be adjacent to the parts of the kernel image 959 * being freed, which may contain secrets. Remove the "high kernel 960 * image mapping" for these freed areas, ensuring they are not even 961 * potentially vulnerable to Meltdown regardless of the specific 962 * optimizations PTI is currently using. 963 * 964 * The "noalias" prevents unmapping the direct map alias which is 965 * needed to access the freed pages. 966 * 967 * This is only valid for 64bit kernels. 32bit has only one mapping 968 * which can't be treated in this way for obvious reasons. 969 */ 970 if (IS_ENABLED(CONFIG_X86_64) && cpu_feature_enabled(X86_FEATURE_PTI)) 971 set_memory_np_noalias(begin_ul, len_pages); 972 } 973 974 void __ref free_initmem(void) 975 { 976 e820__reallocate_tables(); 977 978 mem_encrypt_free_decrypted_mem(); 979 980 free_kernel_image_pages("unused kernel image (initmem)", 981 &__init_begin, &__init_end); 982 } 983 984 #ifdef CONFIG_BLK_DEV_INITRD 985 void __init free_initrd_mem(unsigned long start, unsigned long end) 986 { 987 /* 988 * end could be not aligned, and We can not align that, 989 * decompressor could be confused by aligned initrd_end 990 * We already reserve the end partial page before in 991 * - i386_start_kernel() 992 * - x86_64_start_kernel() 993 * - relocate_initrd() 994 * So here We can do PAGE_ALIGN() safely to get partial page to be freed 995 */ 996 free_init_pages("initrd", start, PAGE_ALIGN(end)); 997 } 998 #endif 999 1000 void __init zone_sizes_init(void) 1001 { 1002 unsigned long max_zone_pfns[MAX_NR_ZONES]; 1003 1004 memset(max_zone_pfns, 0, sizeof(max_zone_pfns)); 1005 1006 #ifdef CONFIG_ZONE_DMA 1007 max_zone_pfns[ZONE_DMA] = min(MAX_DMA_PFN, max_low_pfn); 1008 #endif 1009 #ifdef CONFIG_ZONE_DMA32 1010 max_zone_pfns[ZONE_DMA32] = min(MAX_DMA32_PFN, max_low_pfn); 1011 #endif 1012 max_zone_pfns[ZONE_NORMAL] = max_low_pfn; 1013 #ifdef CONFIG_HIGHMEM 1014 max_zone_pfns[ZONE_HIGHMEM] = max_pfn; 1015 #endif 1016 1017 free_area_init(max_zone_pfns); 1018 } 1019 1020 __visible DEFINE_PER_CPU_ALIGNED(struct tlb_state, cpu_tlbstate) = { 1021 .loaded_mm = &init_mm, 1022 .next_asid = 1, 1023 .cr4 = ~0UL, /* fail hard if we screw up cr4 shadow initialization */ 1024 }; 1025 1026 #ifdef CONFIG_ADDRESS_MASKING 1027 DEFINE_PER_CPU(u64, tlbstate_untag_mask); 1028 EXPORT_PER_CPU_SYMBOL(tlbstate_untag_mask); 1029 #endif 1030 1031 void update_cache_mode_entry(unsigned entry, enum page_cache_mode cache) 1032 { 1033 /* entry 0 MUST be WB (hardwired to speed up translations) */ 1034 BUG_ON(!entry && cache != _PAGE_CACHE_MODE_WB); 1035 1036 __cachemode2pte_tbl[cache] = __cm_idx2pte(entry); 1037 __pte2cachemode_tbl[entry] = cache; 1038 } 1039 1040 #ifdef CONFIG_SWAP 1041 unsigned long arch_max_swapfile_size(void) 1042 { 1043 unsigned long pages; 1044 1045 pages = generic_max_swapfile_size(); 1046 1047 if (boot_cpu_has_bug(X86_BUG_L1TF) && l1tf_mitigation != L1TF_MITIGATION_OFF) { 1048 /* Limit the swap file size to MAX_PA/2 for L1TF workaround */ 1049 unsigned long long l1tf_limit = l1tf_pfn_limit(); 1050 /* 1051 * We encode swap offsets also with 3 bits below those for pfn 1052 * which makes the usable limit higher. 1053 */ 1054 #if CONFIG_PGTABLE_LEVELS > 2 1055 l1tf_limit <<= PAGE_SHIFT - SWP_OFFSET_FIRST_BIT; 1056 #endif 1057 pages = min_t(unsigned long long, l1tf_limit, pages); 1058 } 1059 return pages; 1060 } 1061 #endif 1062 1063 #ifdef CONFIG_EXECMEM 1064 static struct execmem_info execmem_info __ro_after_init; 1065 1066 #ifdef CONFIG_ARCH_HAS_EXECMEM_ROX 1067 void execmem_fill_trapping_insns(void *ptr, size_t size) 1068 { 1069 memset(ptr, INT3_INSN_OPCODE, size); 1070 } 1071 #endif 1072 1073 struct execmem_info __init *execmem_arch_setup(void) 1074 { 1075 unsigned long start, offset = 0; 1076 enum execmem_range_flags flags; 1077 pgprot_t pgprot; 1078 1079 if (kaslr_enabled()) 1080 offset = get_random_u32_inclusive(1, 1024) * PAGE_SIZE; 1081 1082 start = MODULES_VADDR + offset; 1083 1084 if (IS_ENABLED(CONFIG_ARCH_HAS_EXECMEM_ROX) && 1085 cpu_feature_enabled(X86_FEATURE_PSE)) { 1086 pgprot = PAGE_KERNEL_ROX; 1087 flags = EXECMEM_KASAN_SHADOW | EXECMEM_ROX_CACHE; 1088 } else { 1089 pgprot = PAGE_KERNEL; 1090 flags = EXECMEM_KASAN_SHADOW; 1091 } 1092 1093 execmem_info = (struct execmem_info){ 1094 .ranges = { 1095 [EXECMEM_MODULE_TEXT] = { 1096 .flags = flags, 1097 .start = start, 1098 .end = MODULES_END, 1099 .pgprot = pgprot, 1100 .alignment = MODULE_ALIGN, 1101 }, 1102 [EXECMEM_KPROBES] = { 1103 .flags = flags, 1104 .start = start, 1105 .end = MODULES_END, 1106 .pgprot = PAGE_KERNEL_ROX, 1107 .alignment = MODULE_ALIGN, 1108 }, 1109 [EXECMEM_FTRACE] = { 1110 .flags = flags, 1111 .start = start, 1112 .end = MODULES_END, 1113 .pgprot = pgprot, 1114 .alignment = MODULE_ALIGN, 1115 }, 1116 [EXECMEM_BPF] = { 1117 .flags = EXECMEM_KASAN_SHADOW, 1118 .start = start, 1119 .end = MODULES_END, 1120 .pgprot = PAGE_KERNEL, 1121 .alignment = MODULE_ALIGN, 1122 }, 1123 [EXECMEM_MODULE_DATA] = { 1124 .flags = EXECMEM_KASAN_SHADOW, 1125 .start = start, 1126 .end = MODULES_END, 1127 .pgprot = PAGE_KERNEL, 1128 .alignment = MODULE_ALIGN, 1129 }, 1130 }, 1131 }; 1132 1133 return &execmem_info; 1134 } 1135 #endif /* CONFIG_EXECMEM */ 1136