init_64.c revision 49a2bf7303b0dc5fccbb3ff7cf2e7751f0e3953d
1/* 2 * linux/arch/x86_64/mm/init.c 3 * 4 * Copyright (C) 1995 Linus Torvalds 5 * Copyright (C) 2000 Pavel Machek <pavel@suse.cz> 6 * Copyright (C) 2002,2003 Andi Kleen <ak@suse.de> 7 */ 8 9#include <linux/signal.h> 10#include <linux/sched.h> 11#include <linux/kernel.h> 12#include <linux/errno.h> 13#include <linux/string.h> 14#include <linux/types.h> 15#include <linux/ptrace.h> 16#include <linux/mman.h> 17#include <linux/mm.h> 18#include <linux/swap.h> 19#include <linux/smp.h> 20#include <linux/init.h> 21#include <linux/initrd.h> 22#include <linux/pagemap.h> 23#include <linux/bootmem.h> 24#include <linux/proc_fs.h> 25#include <linux/pci.h> 26#include <linux/pfn.h> 27#include <linux/poison.h> 28#include <linux/dma-mapping.h> 29#include <linux/module.h> 30#include <linux/memory_hotplug.h> 31#include <linux/nmi.h> 32 33#include <asm/processor.h> 34#include <asm/bios_ebda.h> 35#include <asm/system.h> 36#include <asm/uaccess.h> 37#include <asm/pgtable.h> 38#include <asm/pgalloc.h> 39#include <asm/dma.h> 40#include <asm/fixmap.h> 41#include <asm/e820.h> 42#include <asm/apic.h> 43#include <asm/tlb.h> 44#include <asm/mmu_context.h> 45#include <asm/proto.h> 46#include <asm/smp.h> 47#include <asm/sections.h> 48#include <asm/kdebug.h> 49#include <asm/numa.h> 50#include <asm/cacheflush.h> 51 52/* 53 * end_pfn only includes RAM, while max_pfn_mapped includes all e820 entries. 54 * The direct mapping extends to max_pfn_mapped, so that we can directly access 55 * apertures, ACPI and other tables without having to play with fixmaps. 56 */ 57unsigned long max_low_pfn_mapped; 58unsigned long max_pfn_mapped; 59 60static unsigned long dma_reserve __initdata; 61 62DEFINE_PER_CPU(struct mmu_gather, mmu_gathers); 63 64int direct_gbpages 65#ifdef CONFIG_DIRECT_GBPAGES 66 = 1 67#endif 68; 69 70static int __init parse_direct_gbpages_off(char *arg) 71{ 72 direct_gbpages = 0; 73 return 0; 74} 75early_param("nogbpages", parse_direct_gbpages_off); 76 77static int __init parse_direct_gbpages_on(char *arg) 78{ 79 direct_gbpages = 1; 80 return 0; 81} 82early_param("gbpages", parse_direct_gbpages_on); 83 84/* 85 * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the 86 * physical space so we can cache the place of the first one and move 87 * around without checking the pgd every time. 88 */ 89 90int after_bootmem; 91 92pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP; 93EXPORT_SYMBOL_GPL(__supported_pte_mask); 94 95static int do_not_nx __cpuinitdata; 96 97/* 98 * noexec=on|off 99 * Control non-executable mappings for 64-bit processes. 100 * 101 * on Enable (default) 102 * off Disable 103 */ 104static int __init nonx_setup(char *str) 105{ 106 if (!str) 107 return -EINVAL; 108 if (!strncmp(str, "on", 2)) { 109 __supported_pte_mask |= _PAGE_NX; 110 do_not_nx = 0; 111 } else if (!strncmp(str, "off", 3)) { 112 do_not_nx = 1; 113 __supported_pte_mask &= ~_PAGE_NX; 114 } 115 return 0; 116} 117early_param("noexec", nonx_setup); 118 119void __cpuinit check_efer(void) 120{ 121 unsigned long efer; 122 123 rdmsrl(MSR_EFER, efer); 124 if (!(efer & EFER_NX) || do_not_nx) 125 __supported_pte_mask &= ~_PAGE_NX; 126} 127 128int force_personality32; 129 130/* 131 * noexec32=on|off 132 * Control non executable heap for 32bit processes. 133 * To control the stack too use noexec=off 134 * 135 * on PROT_READ does not imply PROT_EXEC for 32-bit processes (default) 136 * off PROT_READ implies PROT_EXEC 137 */ 138static int __init nonx32_setup(char *str) 139{ 140 if (!strcmp(str, "on")) 141 force_personality32 &= ~READ_IMPLIES_EXEC; 142 else if (!strcmp(str, "off")) 143 force_personality32 |= READ_IMPLIES_EXEC; 144 return 1; 145} 146__setup("noexec32=", nonx32_setup); 147 148/* 149 * NOTE: This function is marked __ref because it calls __init function 150 * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0. 151 */ 152static __ref void *spp_getpage(void) 153{ 154 void *ptr; 155 156 if (after_bootmem) 157 ptr = (void *) get_zeroed_page(GFP_ATOMIC); 158 else 159 ptr = alloc_bootmem_pages(PAGE_SIZE); 160 161 if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) { 162 panic("set_pte_phys: cannot allocate page data %s\n", 163 after_bootmem ? "after bootmem" : ""); 164 } 165 166 pr_debug("spp_getpage %p\n", ptr); 167 168 return ptr; 169} 170 171void 172set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte) 173{ 174 pud_t *pud; 175 pmd_t *pmd; 176 pte_t *pte; 177 178 pud = pud_page + pud_index(vaddr); 179 if (pud_none(*pud)) { 180 pmd = (pmd_t *) spp_getpage(); 181 pud_populate(&init_mm, pud, pmd); 182 if (pmd != pmd_offset(pud, 0)) { 183 printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n", 184 pmd, pmd_offset(pud, 0)); 185 return; 186 } 187 } 188 pmd = pmd_offset(pud, vaddr); 189 if (pmd_none(*pmd)) { 190 pte = (pte_t *) spp_getpage(); 191 pmd_populate_kernel(&init_mm, pmd, pte); 192 if (pte != pte_offset_kernel(pmd, 0)) { 193 printk(KERN_ERR "PAGETABLE BUG #02!\n"); 194 return; 195 } 196 } 197 198 pte = pte_offset_kernel(pmd, vaddr); 199 set_pte(pte, new_pte); 200 201 /* 202 * It's enough to flush this one mapping. 203 * (PGE mappings get flushed as well) 204 */ 205 __flush_tlb_one(vaddr); 206} 207 208void 209set_pte_vaddr(unsigned long vaddr, pte_t pteval) 210{ 211 pgd_t *pgd; 212 pud_t *pud_page; 213 214 pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval)); 215 216 pgd = pgd_offset_k(vaddr); 217 if (pgd_none(*pgd)) { 218 printk(KERN_ERR 219 "PGD FIXMAP MISSING, it should be setup in head.S!\n"); 220 return; 221 } 222 pud_page = (pud_t*)pgd_page_vaddr(*pgd); 223 set_pte_vaddr_pud(pud_page, vaddr, pteval); 224} 225 226/* 227 * Create large page table mappings for a range of physical addresses. 228 */ 229static void __init __init_extra_mapping(unsigned long phys, unsigned long size, 230 pgprot_t prot) 231{ 232 pgd_t *pgd; 233 pud_t *pud; 234 pmd_t *pmd; 235 236 BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK)); 237 for (; size; phys += PMD_SIZE, size -= PMD_SIZE) { 238 pgd = pgd_offset_k((unsigned long)__va(phys)); 239 if (pgd_none(*pgd)) { 240 pud = (pud_t *) spp_getpage(); 241 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE | 242 _PAGE_USER)); 243 } 244 pud = pud_offset(pgd, (unsigned long)__va(phys)); 245 if (pud_none(*pud)) { 246 pmd = (pmd_t *) spp_getpage(); 247 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE | 248 _PAGE_USER)); 249 } 250 pmd = pmd_offset(pud, phys); 251 BUG_ON(!pmd_none(*pmd)); 252 set_pmd(pmd, __pmd(phys | pgprot_val(prot))); 253 } 254} 255 256void __init init_extra_mapping_wb(unsigned long phys, unsigned long size) 257{ 258 __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE); 259} 260 261void __init init_extra_mapping_uc(unsigned long phys, unsigned long size) 262{ 263 __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE); 264} 265 266/* 267 * The head.S code sets up the kernel high mapping: 268 * 269 * from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text) 270 * 271 * phys_addr holds the negative offset to the kernel, which is added 272 * to the compile time generated pmds. This results in invalid pmds up 273 * to the point where we hit the physaddr 0 mapping. 274 * 275 * We limit the mappings to the region from _text to _end. _end is 276 * rounded up to the 2MB boundary. This catches the invalid pmds as 277 * well, as they are located before _text: 278 */ 279void __init cleanup_highmap(void) 280{ 281 unsigned long vaddr = __START_KERNEL_map; 282 unsigned long end = roundup((unsigned long)_end, PMD_SIZE) - 1; 283 pmd_t *pmd = level2_kernel_pgt; 284 pmd_t *last_pmd = pmd + PTRS_PER_PMD; 285 286 for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) { 287 if (pmd_none(*pmd)) 288 continue; 289 if (vaddr < (unsigned long) _text || vaddr > end) 290 set_pmd(pmd, __pmd(0)); 291 } 292} 293 294static unsigned long __initdata table_start; 295static unsigned long __meminitdata table_end; 296static unsigned long __meminitdata table_top; 297 298static __ref void *alloc_low_page(unsigned long *phys) 299{ 300 unsigned long pfn = table_end++; 301 void *adr; 302 303 if (after_bootmem) { 304 adr = (void *)get_zeroed_page(GFP_ATOMIC); 305 *phys = __pa(adr); 306 307 return adr; 308 } 309 310 if (pfn >= table_top) 311 panic("alloc_low_page: ran out of memory"); 312 313 adr = early_memremap(pfn * PAGE_SIZE, PAGE_SIZE); 314 memset(adr, 0, PAGE_SIZE); 315 *phys = pfn * PAGE_SIZE; 316 return adr; 317} 318 319static __ref void unmap_low_page(void *adr) 320{ 321 if (after_bootmem) 322 return; 323 324 early_iounmap(adr, PAGE_SIZE); 325} 326 327static unsigned long __meminit 328phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end, 329 pgprot_t prot) 330{ 331 unsigned pages = 0; 332 unsigned long last_map_addr = end; 333 int i; 334 335 pte_t *pte = pte_page + pte_index(addr); 336 337 for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) { 338 339 if (addr >= end) { 340 if (!after_bootmem) { 341 for(; i < PTRS_PER_PTE; i++, pte++) 342 set_pte(pte, __pte(0)); 343 } 344 break; 345 } 346 347 /* 348 * We will re-use the existing mapping. 349 * Xen for example has some special requirements, like mapping 350 * pagetable pages as RO. So assume someone who pre-setup 351 * these mappings are more intelligent. 352 */ 353 if (pte_val(*pte)) { 354 pages++; 355 continue; 356 } 357 358 if (0) 359 printk(" pte=%p addr=%lx pte=%016lx\n", 360 pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte); 361 pages++; 362 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot)); 363 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE; 364 } 365 366 update_page_count(PG_LEVEL_4K, pages); 367 368 return last_map_addr; 369} 370 371static unsigned long __meminit 372phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end, 373 pgprot_t prot) 374{ 375 pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd); 376 377 return phys_pte_init(pte, address, end, prot); 378} 379 380static unsigned long __meminit 381phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end, 382 unsigned long page_size_mask, pgprot_t prot) 383{ 384 unsigned long pages = 0; 385 unsigned long last_map_addr = end; 386 387 int i = pmd_index(address); 388 389 for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) { 390 unsigned long pte_phys; 391 pmd_t *pmd = pmd_page + pmd_index(address); 392 pte_t *pte; 393 pgprot_t new_prot = prot; 394 395 if (address >= end) { 396 if (!after_bootmem) { 397 for (; i < PTRS_PER_PMD; i++, pmd++) 398 set_pmd(pmd, __pmd(0)); 399 } 400 break; 401 } 402 403 if (pmd_val(*pmd)) { 404 if (!pmd_large(*pmd)) { 405 spin_lock(&init_mm.page_table_lock); 406 last_map_addr = phys_pte_update(pmd, address, 407 end, prot); 408 spin_unlock(&init_mm.page_table_lock); 409 continue; 410 } 411 /* 412 * If we are ok with PG_LEVEL_2M mapping, then we will 413 * use the existing mapping, 414 * 415 * Otherwise, we will split the large page mapping but 416 * use the same existing protection bits except for 417 * large page, so that we don't violate Intel's TLB 418 * Application note (317080) which says, while changing 419 * the page sizes, new and old translations should 420 * not differ with respect to page frame and 421 * attributes. 422 */ 423 if (page_size_mask & (1 << PG_LEVEL_2M)) { 424 pages++; 425 continue; 426 } 427 new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd)); 428 } 429 430 if (page_size_mask & (1<<PG_LEVEL_2M)) { 431 pages++; 432 spin_lock(&init_mm.page_table_lock); 433 set_pte((pte_t *)pmd, 434 pfn_pte(address >> PAGE_SHIFT, 435 __pgprot(pgprot_val(prot) | _PAGE_PSE))); 436 spin_unlock(&init_mm.page_table_lock); 437 last_map_addr = (address & PMD_MASK) + PMD_SIZE; 438 continue; 439 } 440 441 pte = alloc_low_page(&pte_phys); 442 last_map_addr = phys_pte_init(pte, address, end, new_prot); 443 unmap_low_page(pte); 444 445 spin_lock(&init_mm.page_table_lock); 446 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys)); 447 spin_unlock(&init_mm.page_table_lock); 448 } 449 update_page_count(PG_LEVEL_2M, pages); 450 return last_map_addr; 451} 452 453static unsigned long __meminit 454phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end, 455 unsigned long page_size_mask, pgprot_t prot) 456{ 457 pmd_t *pmd = pmd_offset(pud, 0); 458 unsigned long last_map_addr; 459 460 last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask, prot); 461 __flush_tlb_all(); 462 return last_map_addr; 463} 464 465static unsigned long __meminit 466phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end, 467 unsigned long page_size_mask) 468{ 469 unsigned long pages = 0; 470 unsigned long last_map_addr = end; 471 int i = pud_index(addr); 472 473 for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) { 474 unsigned long pmd_phys; 475 pud_t *pud = pud_page + pud_index(addr); 476 pmd_t *pmd; 477 pgprot_t prot = PAGE_KERNEL; 478 479 if (addr >= end) 480 break; 481 482 if (!after_bootmem && 483 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) { 484 set_pud(pud, __pud(0)); 485 continue; 486 } 487 488 if (pud_val(*pud)) { 489 if (!pud_large(*pud)) { 490 last_map_addr = phys_pmd_update(pud, addr, end, 491 page_size_mask, prot); 492 continue; 493 } 494 /* 495 * If we are ok with PG_LEVEL_1G mapping, then we will 496 * use the existing mapping. 497 * 498 * Otherwise, we will split the gbpage mapping but use 499 * the same existing protection bits except for large 500 * page, so that we don't violate Intel's TLB 501 * Application note (317080) which says, while changing 502 * the page sizes, new and old translations should 503 * not differ with respect to page frame and 504 * attributes. 505 */ 506 if (page_size_mask & (1 << PG_LEVEL_1G)) { 507 pages++; 508 continue; 509 } 510 prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud)); 511 } 512 513 if (page_size_mask & (1<<PG_LEVEL_1G)) { 514 pages++; 515 spin_lock(&init_mm.page_table_lock); 516 set_pte((pte_t *)pud, 517 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE)); 518 spin_unlock(&init_mm.page_table_lock); 519 last_map_addr = (addr & PUD_MASK) + PUD_SIZE; 520 continue; 521 } 522 523 pmd = alloc_low_page(&pmd_phys); 524 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask, 525 prot); 526 unmap_low_page(pmd); 527 528 spin_lock(&init_mm.page_table_lock); 529 pud_populate(&init_mm, pud, __va(pmd_phys)); 530 spin_unlock(&init_mm.page_table_lock); 531 } 532 __flush_tlb_all(); 533 534 update_page_count(PG_LEVEL_1G, pages); 535 536 return last_map_addr; 537} 538 539static unsigned long __meminit 540phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end, 541 unsigned long page_size_mask) 542{ 543 pud_t *pud; 544 545 pud = (pud_t *)pgd_page_vaddr(*pgd); 546 547 return phys_pud_init(pud, addr, end, page_size_mask); 548} 549 550static void __init find_early_table_space(unsigned long end, int use_pse, 551 int use_gbpages) 552{ 553 unsigned long puds, pmds, ptes, tables, start; 554 555 puds = (end + PUD_SIZE - 1) >> PUD_SHIFT; 556 tables = roundup(puds * sizeof(pud_t), PAGE_SIZE); 557 558 if (use_gbpages) { 559 unsigned long extra; 560 561 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT); 562 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT; 563 } else 564 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT; 565 566 tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE); 567 568 if (use_pse) { 569 unsigned long extra; 570 571 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT); 572#ifdef CONFIG_X86_32 573 extra += PMD_SIZE; 574#endif 575 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT; 576 } else 577 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT; 578 579 tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE); 580 581#ifdef CONFIG_X86_32 582 /* for fixmap */ 583 tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE); 584#endif 585 586 /* 587 * RED-PEN putting page tables only on node 0 could 588 * cause a hotspot and fill up ZONE_DMA. The page tables 589 * need roughly 0.5KB per GB. 590 */ 591#ifdef CONFIG_X86_32 592 start = 0x7000; 593 table_start = find_e820_area(start, max_pfn_mapped<<PAGE_SHIFT, 594 tables, PAGE_SIZE); 595#else /* CONFIG_X86_64 */ 596 start = 0x8000; 597 table_start = find_e820_area(start, end, tables, PAGE_SIZE); 598#endif 599 if (table_start == -1UL) 600 panic("Cannot find space for the kernel page tables"); 601 602 table_start >>= PAGE_SHIFT; 603 table_end = table_start; 604 table_top = table_start + (tables >> PAGE_SHIFT); 605 606 printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n", 607 end, table_start << PAGE_SHIFT, table_top << PAGE_SHIFT); 608} 609 610static void __init init_gbpages(void) 611{ 612 if (direct_gbpages && cpu_has_gbpages) 613 printk(KERN_INFO "Using GB pages for direct mapping\n"); 614 else 615 direct_gbpages = 0; 616} 617 618static unsigned long __meminit kernel_physical_mapping_init(unsigned long start, 619 unsigned long end, 620 unsigned long page_size_mask) 621{ 622 623 unsigned long next, last_map_addr = end; 624 625 start = (unsigned long)__va(start); 626 end = (unsigned long)__va(end); 627 628 for (; start < end; start = next) { 629 pgd_t *pgd = pgd_offset_k(start); 630 unsigned long pud_phys; 631 pud_t *pud; 632 633 next = (start + PGDIR_SIZE) & PGDIR_MASK; 634 if (next > end) 635 next = end; 636 637 if (pgd_val(*pgd)) { 638 last_map_addr = phys_pud_update(pgd, __pa(start), 639 __pa(end), page_size_mask); 640 continue; 641 } 642 643 pud = alloc_low_page(&pud_phys); 644 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next), 645 page_size_mask); 646 unmap_low_page(pud); 647 648 spin_lock(&init_mm.page_table_lock); 649 pgd_populate(&init_mm, pgd, __va(pud_phys)); 650 spin_unlock(&init_mm.page_table_lock); 651 } 652 __flush_tlb_all(); 653 654 return last_map_addr; 655} 656 657struct map_range { 658 unsigned long start; 659 unsigned long end; 660 unsigned page_size_mask; 661}; 662 663#define NR_RANGE_MR 5 664 665static int save_mr(struct map_range *mr, int nr_range, 666 unsigned long start_pfn, unsigned long end_pfn, 667 unsigned long page_size_mask) 668{ 669 if (start_pfn < end_pfn) { 670 if (nr_range >= NR_RANGE_MR) 671 panic("run out of range for init_memory_mapping\n"); 672 mr[nr_range].start = start_pfn<<PAGE_SHIFT; 673 mr[nr_range].end = end_pfn<<PAGE_SHIFT; 674 mr[nr_range].page_size_mask = page_size_mask; 675 nr_range++; 676 } 677 678 return nr_range; 679} 680 681/* 682 * Setup the direct mapping of the physical memory at PAGE_OFFSET. 683 * This runs before bootmem is initialized and gets pages directly from 684 * the physical memory. To access them they are temporarily mapped. 685 */ 686unsigned long __init_refok init_memory_mapping(unsigned long start, 687 unsigned long end) 688{ 689 unsigned long last_map_addr = 0; 690 unsigned long page_size_mask = 0; 691 unsigned long start_pfn, end_pfn; 692 unsigned long pos; 693 694 struct map_range mr[NR_RANGE_MR]; 695 int nr_range, i; 696 int use_pse, use_gbpages; 697 698 printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end); 699 700 if (!after_bootmem) 701 init_gbpages(); 702 703#ifdef CONFIG_DEBUG_PAGEALLOC 704 /* 705 * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages. 706 * This will simplify cpa(), which otherwise needs to support splitting 707 * large pages into small in interrupt context, etc. 708 */ 709 use_pse = use_gbpages = 0; 710#else 711 use_pse = cpu_has_pse; 712 use_gbpages = direct_gbpages; 713#endif 714 715 if (use_gbpages) 716 page_size_mask |= 1 << PG_LEVEL_1G; 717 if (use_pse) 718 page_size_mask |= 1 << PG_LEVEL_2M; 719 720 memset(mr, 0, sizeof(mr)); 721 nr_range = 0; 722 723 /* head if not big page alignment ? */ 724 start_pfn = start >> PAGE_SHIFT; 725 pos = start_pfn << PAGE_SHIFT; 726 end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT) 727 << (PMD_SHIFT - PAGE_SHIFT); 728 if (end_pfn > (end >> PAGE_SHIFT)) 729 end_pfn = end >> PAGE_SHIFT; 730 if (start_pfn < end_pfn) { 731 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0); 732 pos = end_pfn << PAGE_SHIFT; 733 } 734 735 /* big page (2M) range */ 736 start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT) 737 << (PMD_SHIFT - PAGE_SHIFT); 738 end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT) 739 << (PUD_SHIFT - PAGE_SHIFT); 740 if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT))) 741 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)); 742 if (start_pfn < end_pfn) { 743 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 744 page_size_mask & (1<<PG_LEVEL_2M)); 745 pos = end_pfn << PAGE_SHIFT; 746 } 747 748 /* big page (1G) range */ 749 start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT) 750 << (PUD_SHIFT - PAGE_SHIFT); 751 end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT); 752 if (start_pfn < end_pfn) { 753 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 754 page_size_mask & 755 ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G))); 756 pos = end_pfn << PAGE_SHIFT; 757 } 758 759 /* tail is not big page (1G) alignment */ 760 start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT) 761 << (PMD_SHIFT - PAGE_SHIFT); 762 end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT); 763 if (start_pfn < end_pfn) { 764 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 765 page_size_mask & (1<<PG_LEVEL_2M)); 766 pos = end_pfn << PAGE_SHIFT; 767 } 768 769 /* tail is not big page (2M) alignment */ 770 start_pfn = pos>>PAGE_SHIFT; 771 end_pfn = end>>PAGE_SHIFT; 772 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0); 773 774 /* try to merge same page size and continuous */ 775 for (i = 0; nr_range > 1 && i < nr_range - 1; i++) { 776 unsigned long old_start; 777 if (mr[i].end != mr[i+1].start || 778 mr[i].page_size_mask != mr[i+1].page_size_mask) 779 continue; 780 /* move it */ 781 old_start = mr[i].start; 782 memmove(&mr[i], &mr[i+1], 783 (nr_range - 1 - i) * sizeof(struct map_range)); 784 mr[i--].start = old_start; 785 nr_range--; 786 } 787 788 for (i = 0; i < nr_range; i++) 789 printk(KERN_DEBUG " %010lx - %010lx page %s\n", 790 mr[i].start, mr[i].end, 791 (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":( 792 (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k")); 793 794 /* 795 * Find space for the kernel direct mapping tables. 796 * 797 * Later we should allocate these tables in the local node of the 798 * memory mapped. Unfortunately this is done currently before the 799 * nodes are discovered. 800 */ 801 if (!after_bootmem) 802 find_early_table_space(end, use_pse, use_gbpages); 803 804 for (i = 0; i < nr_range; i++) 805 last_map_addr = kernel_physical_mapping_init( 806 mr[i].start, mr[i].end, 807 mr[i].page_size_mask); 808 809 if (!after_bootmem) 810 mmu_cr4_features = read_cr4(); 811 __flush_tlb_all(); 812 813 if (!after_bootmem && table_end > table_start) 814 reserve_early(table_start << PAGE_SHIFT, 815 table_end << PAGE_SHIFT, "PGTABLE"); 816 817 printk(KERN_INFO "last_map_addr: %lx end: %lx\n", 818 last_map_addr, end); 819 820 if (!after_bootmem) 821 early_memtest(start, end); 822 823 return last_map_addr >> PAGE_SHIFT; 824} 825 826#ifndef CONFIG_NUMA 827void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn) 828{ 829 unsigned long bootmap_size, bootmap; 830 831 bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT; 832 bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size, 833 PAGE_SIZE); 834 if (bootmap == -1L) 835 panic("Cannot find bootmem map of size %ld\n", bootmap_size); 836 /* don't touch min_low_pfn */ 837 bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT, 838 0, end_pfn); 839 e820_register_active_regions(0, start_pfn, end_pfn); 840 free_bootmem_with_active_regions(0, end_pfn); 841 early_res_to_bootmem(0, end_pfn<<PAGE_SHIFT); 842 reserve_bootmem(bootmap, bootmap_size, BOOTMEM_DEFAULT); 843} 844 845void __init paging_init(void) 846{ 847 unsigned long max_zone_pfns[MAX_NR_ZONES]; 848 849 memset(max_zone_pfns, 0, sizeof(max_zone_pfns)); 850 max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN; 851 max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN; 852 max_zone_pfns[ZONE_NORMAL] = max_pfn; 853 854 memory_present(0, 0, max_pfn); 855 sparse_init(); 856 free_area_init_nodes(max_zone_pfns); 857} 858#endif 859 860/* 861 * Memory hotplug specific functions 862 */ 863#ifdef CONFIG_MEMORY_HOTPLUG 864/* 865 * Memory is added always to NORMAL zone. This means you will never get 866 * additional DMA/DMA32 memory. 867 */ 868int arch_add_memory(int nid, u64 start, u64 size) 869{ 870 struct pglist_data *pgdat = NODE_DATA(nid); 871 struct zone *zone = pgdat->node_zones + ZONE_NORMAL; 872 unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT; 873 unsigned long nr_pages = size >> PAGE_SHIFT; 874 int ret; 875 876 last_mapped_pfn = init_memory_mapping(start, start + size); 877 if (last_mapped_pfn > max_pfn_mapped) 878 max_pfn_mapped = last_mapped_pfn; 879 880 ret = __add_pages(nid, zone, start_pfn, nr_pages); 881 WARN_ON_ONCE(ret); 882 883 return ret; 884} 885EXPORT_SYMBOL_GPL(arch_add_memory); 886 887#if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA) 888int memory_add_physaddr_to_nid(u64 start) 889{ 890 return 0; 891} 892EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid); 893#endif 894 895#endif /* CONFIG_MEMORY_HOTPLUG */ 896 897static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel, 898 kcore_modules, kcore_vsyscall; 899 900void __init mem_init(void) 901{ 902 long codesize, reservedpages, datasize, initsize; 903 unsigned long absent_pages; 904 905 pci_iommu_alloc(); 906 907 /* clear_bss() already clear the empty_zero_page */ 908 909 reservedpages = 0; 910 911 /* this will put all low memory onto the freelists */ 912#ifdef CONFIG_NUMA 913 totalram_pages = numa_free_all_bootmem(); 914#else 915 totalram_pages = free_all_bootmem(); 916#endif 917 918 absent_pages = absent_pages_in_range(0, max_pfn); 919 reservedpages = max_pfn - totalram_pages - absent_pages; 920 after_bootmem = 1; 921 922 codesize = (unsigned long) &_etext - (unsigned long) &_text; 923 datasize = (unsigned long) &_edata - (unsigned long) &_etext; 924 initsize = (unsigned long) &__init_end - (unsigned long) &__init_begin; 925 926 /* Register memory areas for /proc/kcore */ 927 kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT); 928 kclist_add(&kcore_vmalloc, (void *)VMALLOC_START, 929 VMALLOC_END-VMALLOC_START); 930 kclist_add(&kcore_kernel, &_stext, _end - _stext); 931 kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN); 932 kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START, 933 VSYSCALL_END - VSYSCALL_START); 934 935 printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, " 936 "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n", 937 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10), 938 max_pfn << (PAGE_SHIFT-10), 939 codesize >> 10, 940 absent_pages << (PAGE_SHIFT-10), 941 reservedpages << (PAGE_SHIFT-10), 942 datasize >> 10, 943 initsize >> 10); 944} 945 946#ifdef CONFIG_DEBUG_RODATA 947const int rodata_test_data = 0xC3; 948EXPORT_SYMBOL_GPL(rodata_test_data); 949 950void mark_rodata_ro(void) 951{ 952 unsigned long start = PFN_ALIGN(_stext), end = PFN_ALIGN(__end_rodata); 953 unsigned long rodata_start = 954 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK; 955 956#ifdef CONFIG_DYNAMIC_FTRACE 957 /* Dynamic tracing modifies the kernel text section */ 958 start = rodata_start; 959#endif 960 961 printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n", 962 (end - start) >> 10); 963 set_memory_ro(start, (end - start) >> PAGE_SHIFT); 964 965 /* 966 * The rodata section (but not the kernel text!) should also be 967 * not-executable. 968 */ 969 set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT); 970 971 rodata_test(); 972 973#ifdef CONFIG_CPA_DEBUG 974 printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end); 975 set_memory_rw(start, (end-start) >> PAGE_SHIFT); 976 977 printk(KERN_INFO "Testing CPA: again\n"); 978 set_memory_ro(start, (end-start) >> PAGE_SHIFT); 979#endif 980} 981 982#endif 983 984int __init reserve_bootmem_generic(unsigned long phys, unsigned long len, 985 int flags) 986{ 987#ifdef CONFIG_NUMA 988 int nid, next_nid; 989 int ret; 990#endif 991 unsigned long pfn = phys >> PAGE_SHIFT; 992 993 if (pfn >= max_pfn) { 994 /* 995 * This can happen with kdump kernels when accessing 996 * firmware tables: 997 */ 998 if (pfn < max_pfn_mapped) 999 return -EFAULT; 1000 1001 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n", 1002 phys, len); 1003 return -EFAULT; 1004 } 1005 1006 /* Should check here against the e820 map to avoid double free */ 1007#ifdef CONFIG_NUMA 1008 nid = phys_to_nid(phys); 1009 next_nid = phys_to_nid(phys + len - 1); 1010 if (nid == next_nid) 1011 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags); 1012 else 1013 ret = reserve_bootmem(phys, len, flags); 1014 1015 if (ret != 0) 1016 return ret; 1017 1018#else 1019 reserve_bootmem(phys, len, BOOTMEM_DEFAULT); 1020#endif 1021 1022 if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) { 1023 dma_reserve += len / PAGE_SIZE; 1024 set_dma_reserve(dma_reserve); 1025 } 1026 1027 return 0; 1028} 1029 1030int kern_addr_valid(unsigned long addr) 1031{ 1032 unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT; 1033 pgd_t *pgd; 1034 pud_t *pud; 1035 pmd_t *pmd; 1036 pte_t *pte; 1037 1038 if (above != 0 && above != -1UL) 1039 return 0; 1040 1041 pgd = pgd_offset_k(addr); 1042 if (pgd_none(*pgd)) 1043 return 0; 1044 1045 pud = pud_offset(pgd, addr); 1046 if (pud_none(*pud)) 1047 return 0; 1048 1049 pmd = pmd_offset(pud, addr); 1050 if (pmd_none(*pmd)) 1051 return 0; 1052 1053 if (pmd_large(*pmd)) 1054 return pfn_valid(pmd_pfn(*pmd)); 1055 1056 pte = pte_offset_kernel(pmd, addr); 1057 if (pte_none(*pte)) 1058 return 0; 1059 1060 return pfn_valid(pte_pfn(*pte)); 1061} 1062 1063/* 1064 * A pseudo VMA to allow ptrace access for the vsyscall page. This only 1065 * covers the 64bit vsyscall page now. 32bit has a real VMA now and does 1066 * not need special handling anymore: 1067 */ 1068static struct vm_area_struct gate_vma = { 1069 .vm_start = VSYSCALL_START, 1070 .vm_end = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE), 1071 .vm_page_prot = PAGE_READONLY_EXEC, 1072 .vm_flags = VM_READ | VM_EXEC 1073}; 1074 1075struct vm_area_struct *get_gate_vma(struct task_struct *tsk) 1076{ 1077#ifdef CONFIG_IA32_EMULATION 1078 if (test_tsk_thread_flag(tsk, TIF_IA32)) 1079 return NULL; 1080#endif 1081 return &gate_vma; 1082} 1083 1084int in_gate_area(struct task_struct *task, unsigned long addr) 1085{ 1086 struct vm_area_struct *vma = get_gate_vma(task); 1087 1088 if (!vma) 1089 return 0; 1090 1091 return (addr >= vma->vm_start) && (addr < vma->vm_end); 1092} 1093 1094/* 1095 * Use this when you have no reliable task/vma, typically from interrupt 1096 * context. It is less reliable than using the task's vma and may give 1097 * false positives: 1098 */ 1099int in_gate_area_no_task(unsigned long addr) 1100{ 1101 return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END); 1102} 1103 1104const char *arch_vma_name(struct vm_area_struct *vma) 1105{ 1106 if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso) 1107 return "[vdso]"; 1108 if (vma == &gate_vma) 1109 return "[vsyscall]"; 1110 return NULL; 1111} 1112 1113#ifdef CONFIG_SPARSEMEM_VMEMMAP 1114/* 1115 * Initialise the sparsemem vmemmap using huge-pages at the PMD level. 1116 */ 1117static long __meminitdata addr_start, addr_end; 1118static void __meminitdata *p_start, *p_end; 1119static int __meminitdata node_start; 1120 1121int __meminit 1122vmemmap_populate(struct page *start_page, unsigned long size, int node) 1123{ 1124 unsigned long addr = (unsigned long)start_page; 1125 unsigned long end = (unsigned long)(start_page + size); 1126 unsigned long next; 1127 pgd_t *pgd; 1128 pud_t *pud; 1129 pmd_t *pmd; 1130 1131 for (; addr < end; addr = next) { 1132 void *p = NULL; 1133 1134 pgd = vmemmap_pgd_populate(addr, node); 1135 if (!pgd) 1136 return -ENOMEM; 1137 1138 pud = vmemmap_pud_populate(pgd, addr, node); 1139 if (!pud) 1140 return -ENOMEM; 1141 1142 if (!cpu_has_pse) { 1143 next = (addr + PAGE_SIZE) & PAGE_MASK; 1144 pmd = vmemmap_pmd_populate(pud, addr, node); 1145 1146 if (!pmd) 1147 return -ENOMEM; 1148 1149 p = vmemmap_pte_populate(pmd, addr, node); 1150 1151 if (!p) 1152 return -ENOMEM; 1153 1154 addr_end = addr + PAGE_SIZE; 1155 p_end = p + PAGE_SIZE; 1156 } else { 1157 next = pmd_addr_end(addr, end); 1158 1159 pmd = pmd_offset(pud, addr); 1160 if (pmd_none(*pmd)) { 1161 pte_t entry; 1162 1163 p = vmemmap_alloc_block(PMD_SIZE, node); 1164 if (!p) 1165 return -ENOMEM; 1166 1167 entry = pfn_pte(__pa(p) >> PAGE_SHIFT, 1168 PAGE_KERNEL_LARGE); 1169 set_pmd(pmd, __pmd(pte_val(entry))); 1170 1171 /* check to see if we have contiguous blocks */ 1172 if (p_end != p || node_start != node) { 1173 if (p_start) 1174 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n", 1175 addr_start, addr_end-1, p_start, p_end-1, node_start); 1176 addr_start = addr; 1177 node_start = node; 1178 p_start = p; 1179 } 1180 1181 addr_end = addr + PMD_SIZE; 1182 p_end = p + PMD_SIZE; 1183 } else 1184 vmemmap_verify((pte_t *)pmd, node, addr, next); 1185 } 1186 1187 } 1188 return 0; 1189} 1190 1191void __meminit vmemmap_populate_print_last(void) 1192{ 1193 if (p_start) { 1194 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n", 1195 addr_start, addr_end-1, p_start, p_end-1, node_start); 1196 p_start = NULL; 1197 p_end = NULL; 1198 node_start = 0; 1199 } 1200} 1201#endif 1202