arch/x86/mm/init.c

   1 #include <linux/gfp.h>
   2 #include <linux/initrd.h>
   3 #include <linux/ioport.h>
   4 #include <linux/swap.h>
   5 #include <linux/memblock.h>
   6 #include <linux/bootmem.h>      /* for max_low_pfn */
   7
   8 #include <asm/cacheflush.h>
   9 #include <asm/e820.h>
  10 #include <asm/init.h>
  11 #include <asm/page.h>
  12 #include <asm/page_types.h>
  13 #include <asm/sections.h>
  14 #include <asm/setup.h>
  15 #include <asm/tlbflush.h>
  16 #include <asm/tlb.h>
  17 #include <asm/proto.h>
  18 #include <asm/dma.h>            /* for MAX_DMA_PFN */
  19 #include <asm/microcode.h>
  20
  21 #include "mm_internal.h"
  22
  23 static unsigned long __initdata pgt_buf_start;
  24 static unsigned long __initdata pgt_buf_end;
  25 static unsigned long __initdata pgt_buf_top;
  26
  27 static unsigned long min_pfn_mapped;
  28
  29 static bool __initdata can_use_brk_pgt = true;
  30
  31 /*
  32  * Pages returned are already directly mapped.
  33  *
  34  * Changing that is likely to break Xen, see commit:
  35  *
  36  *    279b706 x86,xen: introduce x86_init.mapping.pagetable_reserve
  37  *
  38  * for detailed information.
  39  */
  40 __ref void *alloc_low_pages(unsigned int num)
  41 {
  42         unsigned long pfn;
  43         int i;
  44
  45         if (after_bootmem) {
  46                 unsigned int order;
  47
  48                 order = get_order((unsigned long)num << PAGE_SHIFT);
  49                 return (void *)__get_free_pages(GFP_ATOMIC | __GFP_NOTRACK |
  50                                                 __GFP_ZERO, order);
  51         }
  52
  53         if ((pgt_buf_end + num) > pgt_buf_top || !can_use_brk_pgt) {
  54                 unsigned long ret;
  55                 if (min_pfn_mapped >= max_pfn_mapped)
  56                         panic("alloc_low_page: ran out of memory");
  57                 ret = memblock_find_in_range(min_pfn_mapped << PAGE_SHIFT,
  58                                         max_pfn_mapped << PAGE_SHIFT,
  59                                         PAGE_SIZE * num , PAGE_SIZE);
  60                 if (!ret)
  61                         panic("alloc_low_page: can not alloc memory");
  62                 memblock_reserve(ret, PAGE_SIZE * num);
  63                 pfn = ret >> PAGE_SHIFT;
  64         } else {
  65                 pfn = pgt_buf_end;
  66                 pgt_buf_end += num;
  67                 printk(KERN_DEBUG "BRK [%#010lx, %#010lx] PGTABLE\n",
  68                         pfn << PAGE_SHIFT, (pgt_buf_end << PAGE_SHIFT) - 1);
  69         }
  70
  71         for (i = 0; i < num; i++) {
  72                 void *adr;
  73
  74                 adr = __va((pfn + i) << PAGE_SHIFT);
  75                 clear_page(adr);
  76         }
  77
  78         return __va(pfn << PAGE_SHIFT);
  79 }
  80
  81 /* need 4 4k for initial PMD_SIZE, 4k for 0-ISA_END_ADDRESS */
  82 #define INIT_PGT_BUF_SIZE       (5 * PAGE_SIZE)
  83 RESERVE_BRK(early_pgt_alloc, INIT_PGT_BUF_SIZE);
  84 void  __init early_alloc_pgt_buf(void)
  85 {
  86         unsigned long tables = INIT_PGT_BUF_SIZE;
  87         phys_addr_t base;
  88
  89         base = __pa(extend_brk(tables, PAGE_SIZE));
  90
  91         pgt_buf_start = base >> PAGE_SHIFT;
  92         pgt_buf_end = pgt_buf_start;
  93         pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
  94 }
  95
  96 int after_bootmem;
  97
  98 int direct_gbpages
  99 #ifdef CONFIG_DIRECT_GBPAGES
 100                                 = 1
 101 #endif
 102 ;
 103
 104 static void __init init_gbpages(void)
 105 {
 106 #ifdef CONFIG_X86_64
 107         if (direct_gbpages && cpu_has_gbpages)
 108                 printk(KERN_INFO "Using GB pages for direct mapping\n");
 109         else
 110                 direct_gbpages = 0;
 111 #endif
 112 }
 113
 114 struct map_range {
 115         unsigned long start;
 116         unsigned long end;
 117         unsigned page_size_mask;
 118 };
 119
 120 static int page_size_mask;
 121
 122 static void __init probe_page_size_mask(void)
 123 {
 124         init_gbpages();
 125
 126 #if !defined(CONFIG_DEBUG_PAGEALLOC) && !defined(CONFIG_KMEMCHECK)
 127         /*
 128          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 129          * This will simplify cpa(), which otherwise needs to support splitting
 130          * large pages into small in interrupt context, etc.
 131          */
 132         if (direct_gbpages)
 133                 page_size_mask |= 1 << PG_LEVEL_1G;
 134         if (cpu_has_pse)
 135                 page_size_mask |= 1 << PG_LEVEL_2M;
 136 #endif
 137
 138         /* Enable PSE if available */
 139         if (cpu_has_pse)
 140                 set_in_cr4(X86_CR4_PSE);
 141
 142         /* Enable PGE if available */
 143         if (cpu_has_pge) {
 144                 set_in_cr4(X86_CR4_PGE);
 145                 __supported_pte_mask |= _PAGE_GLOBAL;
 146         }
 147 }
 148
 149 #ifdef CONFIG_X86_32
 150 #define NR_RANGE_MR 3
 151 #else /* CONFIG_X86_64 */
 152 #define NR_RANGE_MR 5
 153 #endif
 154
 155 static int __meminit save_mr(struct map_range *mr, int nr_range,
 156                              unsigned long start_pfn, unsigned long end_pfn,
 157                              unsigned long page_size_mask)
 158 {
 159         if (start_pfn < end_pfn) {
 160                 if (nr_range >= NR_RANGE_MR)
 161                         panic("run out of range for init_memory_mapping\n");
 162                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 163                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 164                 mr[nr_range].page_size_mask = page_size_mask;
 165                 nr_range++;
 166         }
 167
 168         return nr_range;
 169 }
 170
 171 /*
 172  * adjust the page_size_mask for small range to go with
 173  *      big page size instead small one if nearby are ram too.
 174  */
 175 static void __init_refok adjust_range_page_size_mask(struct map_range *mr,
 176                                                          int nr_range)
 177 {
 178         int i;
 179
 180         for (i = 0; i < nr_range; i++) {
 181                 if ((page_size_mask & (1<<PG_LEVEL_2M)) &&
 182                     !(mr[i].page_size_mask & (1<<PG_LEVEL_2M))) {
 183                         unsigned long start = round_down(mr[i].start, PMD_SIZE);
 184                         unsigned long end = round_up(mr[i].end, PMD_SIZE);
 185
 186 #ifdef CONFIG_X86_32
 187                         if ((end >> PAGE_SHIFT) > max_low_pfn)
 188                                 continue;
 189 #endif
 190
 191                         if (memblock_is_region_memory(start, end - start))
 192                                 mr[i].page_size_mask |= 1<<PG_LEVEL_2M;
 193                 }
 194                 if ((page_size_mask & (1<<PG_LEVEL_1G)) &&
 195                     !(mr[i].page_size_mask & (1<<PG_LEVEL_1G))) {
 196                         unsigned long start = round_down(mr[i].start, PUD_SIZE);
 197                         unsigned long end = round_up(mr[i].end, PUD_SIZE);
 198
 199                         if (memblock_is_region_memory(start, end - start))
 200                                 mr[i].page_size_mask |= 1<<PG_LEVEL_1G;
 201                 }
 202         }
 203 }
 204
 205 static int __meminit split_mem_range(struct map_range *mr, int nr_range,
 206                                      unsigned long start,
 207                                      unsigned long end)
 208 {
 209         unsigned long start_pfn, end_pfn, limit_pfn;
 210         unsigned long pfn;
 211         int i;
 212
 213         limit_pfn = PFN_DOWN(end);
 214
 215         /* head if not big page alignment ? */
 216         pfn = start_pfn = PFN_DOWN(start);
 217 #ifdef CONFIG_X86_32
 218         /*
 219          * Don't use a large page for the first 2/4MB of memory
 220          * because there are often fixed size MTRRs in there
 221          * and overlapping MTRRs into large pages can cause
 222          * slowdowns.
 223          */
 224         if (pfn == 0)
 225                 end_pfn = PFN_DOWN(PMD_SIZE);
 226         else
 227                 end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
 228 #else /* CONFIG_X86_64 */
 229         end_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
 230 #endif
 231         if (end_pfn > limit_pfn)
 232                 end_pfn = limit_pfn;
 233         if (start_pfn < end_pfn) {
 234                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 235                 pfn = end_pfn;
 236         }
 237
 238         /* big page (2M) range */
 239         start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
 240 #ifdef CONFIG_X86_32
 241         end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
 242 #else /* CONFIG_X86_64 */
 243         end_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
 244         if (end_pfn > round_down(limit_pfn, PFN_DOWN(PMD_SIZE)))
 245                 end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
 246 #endif
 247
 248         if (start_pfn < end_pfn) {
 249                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 250                                 page_size_mask & (1<<PG_LEVEL_2M));
 251                 pfn = end_pfn;
 252         }
 253
 254 #ifdef CONFIG_X86_64
 255         /* big page (1G) range */
 256         start_pfn = round_up(pfn, PFN_DOWN(PUD_SIZE));
 257         end_pfn = round_down(limit_pfn, PFN_DOWN(PUD_SIZE));
 258         if (start_pfn < end_pfn) {
 259                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 260                                 page_size_mask &
 261                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 262                 pfn = end_pfn;
 263         }
 264
 265         /* tail is not big page (1G) alignment */
 266         start_pfn = round_up(pfn, PFN_DOWN(PMD_SIZE));
 267         end_pfn = round_down(limit_pfn, PFN_DOWN(PMD_SIZE));
 268         if (start_pfn < end_pfn) {
 269                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 270                                 page_size_mask & (1<<PG_LEVEL_2M));
 271                 pfn = end_pfn;
 272         }
 273 #endif
 274
 275         /* tail is not big page (2M) alignment */
 276         start_pfn = pfn;
 277         end_pfn = limit_pfn;
 278         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 279
 280         /* try to merge same page size and continuous */
 281         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 282                 unsigned long old_start;
 283                 if (mr[i].end != mr[i+1].start ||
 284                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 285                         continue;
 286                 /* move it */
 287                 old_start = mr[i].start;
 288                 memmove(&mr[i], &mr[i+1],
 289                         (nr_range - 1 - i) * sizeof(struct map_range));
 290                 mr[i--].start = old_start;
 291                 nr_range--;
 292         }
 293
 294         if (!after_bootmem)
 295                 adjust_range_page_size_mask(mr, nr_range);
 296
 297         for (i = 0; i < nr_range; i++)
 298                 printk(KERN_DEBUG " [mem %#010lx-%#010lx] page %s\n",
 299                                 mr[i].start, mr[i].end - 1,
 300                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 301                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 302
 303         return nr_range;
 304 }
 305
 306 struct range pfn_mapped[E820_X_MAX];
 307 int nr_pfn_mapped;
 308
 309 static void add_pfn_range_mapped(unsigned long start_pfn, unsigned long end_pfn)
 310 {
 311         nr_pfn_mapped = add_range_with_merge(pfn_mapped, E820_X_MAX,
 312                                              nr_pfn_mapped, start_pfn, end_pfn);
 313         nr_pfn_mapped = clean_sort_range(pfn_mapped, E820_X_MAX);
 314
 315         max_pfn_mapped = max(max_pfn_mapped, end_pfn);
 316
 317         if (start_pfn < (1UL<<(32-PAGE_SHIFT)))
 318                 max_low_pfn_mapped = max(max_low_pfn_mapped,
 319                                          min(end_pfn, 1UL<<(32-PAGE_SHIFT)));
 320 }
 321
 322 bool pfn_range_is_mapped(unsigned long start_pfn, unsigned long end_pfn)
 323 {
 324         int i;
 325
 326         for (i = 0; i < nr_pfn_mapped; i++)
 327                 if ((start_pfn >= pfn_mapped[i].start) &&
 328                     (end_pfn <= pfn_mapped[i].end))
 329                         return true;
 330
 331         return false;
 332 }
 333
 334 /*
 335  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 336  * This runs before bootmem is initialized and gets pages directly from
 337  * the physical memory. To access them they are temporarily mapped.
 338  */
 339 unsigned long __init_refok init_memory_mapping(unsigned long start,
 340                                                unsigned long end)
 341 {
 342         struct map_range mr[NR_RANGE_MR];
 343         unsigned long ret = 0;
 344         int nr_range, i;
 345
 346         pr_info("init_memory_mapping: [mem %#010lx-%#010lx]\n",
 347                start, end - 1);
 348
 349         memset(mr, 0, sizeof(mr));
 350         nr_range = split_mem_range(mr, 0, start, end);
 351
 352         for (i = 0; i < nr_range; i++)
 353                 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
 354                                                    mr[i].page_size_mask);
 355
 356         add_pfn_range_mapped(start >> PAGE_SHIFT, ret >> PAGE_SHIFT);
 357
 358         return ret >> PAGE_SHIFT;
 359 }
 360
 361 /*
 362  * We need to iterate through the E820 memory map and create direct mappings
 363  * for only E820_RAM and E820_KERN_RESERVED regions. We cannot simply
 364  * create direct mappings for all pfns from [0 to max_low_pfn) and
 365  * [4GB to max_pfn) because of possible memory holes in high addresses
 366  * that cannot be marked as UC by fixed/variable range MTRRs.
 367  * Depending on the alignment of E820 ranges, this may possibly result
 368  * in using smaller size (i.e. 4K instead of 2M or 1G) page tables.
 369  *
 370  * init_mem_mapping() calls init_range_memory_mapping() with big range.
 371  * That range would have hole in the middle or ends, and only ram parts
 372  * will be mapped in init_range_memory_mapping().
 373  */
 374 static unsigned long __init init_range_memory_mapping(
 375                                            unsigned long r_start,
 376                                            unsigned long r_end)
 377 {
 378         unsigned long start_pfn, end_pfn;
 379         unsigned long mapped_ram_size = 0;
 380         int i;
 381
 382         for_each_mem_pfn_range(i, MAX_NUMNODES, &start_pfn, &end_pfn, NULL) {
 383                 u64 start = clamp_val(PFN_PHYS(start_pfn), r_start, r_end);
 384                 u64 end = clamp_val(PFN_PHYS(end_pfn), r_start, r_end);
 385                 if (start >= end)
 386                         continue;
 387
 388                 /*
 389                  * if it is overlapping with brk pgt, we need to
 390                  * alloc pgt buf from memblock instead.
 391                  */
 392                 can_use_brk_pgt = max(start, (u64)pgt_buf_end<<PAGE_SHIFT) >=
 393                                     min(end, (u64)pgt_buf_top<<PAGE_SHIFT);
 394                 init_memory_mapping(start, end);
 395                 mapped_ram_size += end - start;
 396                 can_use_brk_pgt = true;
 397         }
 398
 399         return mapped_ram_size;
 400 }
 401
 402 /* (PUD_SHIFT-PMD_SHIFT)/2 */
 403 #define STEP_SIZE_SHIFT 5
 404 void __init init_mem_mapping(void)
 405 {
 406         unsigned long end, real_end, start, last_start;
 407         unsigned long step_size;
 408         unsigned long addr;
 409         unsigned long mapped_ram_size = 0;
 410         unsigned long new_mapped_ram_size;
 411
 412         probe_page_size_mask();
 413
 414 #ifdef CONFIG_X86_64
 415         end = max_pfn << PAGE_SHIFT;
 416 #else
 417         end = max_low_pfn << PAGE_SHIFT;
 418 #endif
 419
 420         /* the ISA range is always mapped regardless of memory holes */
 421         init_memory_mapping(0, ISA_END_ADDRESS);
 422
 423         /* xen has big range in reserved near end of ram, skip it at first.*/
 424         addr = memblock_find_in_range(ISA_END_ADDRESS, end, PMD_SIZE, PMD_SIZE);
 425         real_end = addr + PMD_SIZE;
 426
 427         /* step_size need to be small so pgt_buf from BRK could cover it */
 428         step_size = PMD_SIZE;
 429         max_pfn_mapped = 0; /* will get exact value next */
 430         min_pfn_mapped = real_end >> PAGE_SHIFT;
 431         last_start = start = real_end;
 432
 433         /*
 434          * We start from the top (end of memory) and go to the bottom.
 435          * The memblock_find_in_range() gets us a block of RAM from the
 436          * end of RAM in [min_pfn_mapped, max_pfn_mapped) used as new pages
 437          * for page table.
 438          */
 439         while (last_start > ISA_END_ADDRESS) {
 440                 if (last_start > step_size) {
 441                         start = round_down(last_start - 1, step_size);
 442                         if (start < ISA_END_ADDRESS)
 443                                 start = ISA_END_ADDRESS;
 444                 } else
 445                         start = ISA_END_ADDRESS;
 446                 new_mapped_ram_size = init_range_memory_mapping(start,
 447                                                         last_start);
 448                 last_start = start;
 449                 min_pfn_mapped = last_start >> PAGE_SHIFT;
 450                 /* only increase step_size after big range get mapped */
 451                 if (new_mapped_ram_size > mapped_ram_size)
 452                         step_size <<= STEP_SIZE_SHIFT;
 453                 mapped_ram_size += new_mapped_ram_size;
 454         }
 455
 456         if (real_end < end)
 457                 init_range_memory_mapping(real_end, end);
 458
 459 #ifdef CONFIG_X86_64
 460         if (max_pfn > max_low_pfn) {
 461                 /* can we preseve max_low_pfn ?*/
 462                 max_low_pfn = max_pfn;
 463         }
 464 #else
 465         early_ioremap_page_table_range_init();
 466 #endif
 467
 468         load_cr3(swapper_pg_dir);
 469         __flush_tlb_all();
 470
 471         early_memtest(0, max_pfn_mapped << PAGE_SHIFT);
 472 }
 473
 474 /*
 475  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 476  * is valid. The argument is a physical page number.
 477  *
 478  *
 479  * On x86, access has to be given to the first megabyte of ram because that area
 480  * contains bios code and data regions used by X and dosemu and similar apps.
 481  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 482  * mmio resources as well as potential bios/acpi data regions.
 483  */
 484 int devmem_is_allowed(unsigned long pagenr)
 485 {
 486         if (pagenr < 256)
 487                 return 1;
 488         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 489                 return 0;
 490         if (!page_is_ram(pagenr))
 491                 return 1;
 492         return 0;
 493 }
 494
 495 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 496 {
 497         unsigned long addr;
 498         unsigned long begin_aligned, end_aligned;
 499
 500         /* Make sure boundaries are page aligned */
 501         begin_aligned = PAGE_ALIGN(begin);
 502         end_aligned   = end & PAGE_MASK;
 503
 504         if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
 505                 begin = begin_aligned;
 506                 end   = end_aligned;
 507         }
 508
 509         if (begin >= end)
 510                 return;
 511
 512         addr = begin;
 513
 514         /*
 515          * If debugging page accesses then do not free this memory but
 516          * mark them not present - any buggy init-section access will
 517          * create a kernel page fault:
 518          */
 519 #ifdef CONFIG_DEBUG_PAGEALLOC
 520         printk(KERN_INFO "debug: unmapping init [mem %#010lx-%#010lx]\n",
 521                 begin, end - 1);
 522         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 523 #else
 524         /*
 525          * We just marked the kernel text read only above, now that
 526          * we are going to free part of that, we need to make that
 527          * writeable and non-executable first.
 528          */
 529         set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
 530         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
 531
 532         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 533
 534         for (; addr < end; addr += PAGE_SIZE) {
 535                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
 536                 free_reserved_page(virt_to_page(addr));
 537         }
 538 #endif
 539 }
 540
 541 void free_initmem(void)
 542 {
 543         free_init_pages("unused kernel memory",
 544                         (unsigned long)(&__init_begin),
 545                         (unsigned long)(&__init_end));
 546 }
 547
 548 #ifdef CONFIG_BLK_DEV_INITRD
 549 void __init free_initrd_mem(unsigned long start, unsigned long end)
 550 {
 551 #ifdef CONFIG_MICROCODE_EARLY
 552         /*
 553          * Remember, initrd memory may contain microcode or other useful things.
 554          * Before we lose initrd mem, we need to find a place to hold them
 555          * now that normal virtual memory is enabled.
 556          */
 557         save_microcode_in_initrd();
 558 #endif
 559
 560         /*
 561          * end could be not aligned, and We can not align that,
 562          * decompresser could be confused by aligned initrd_end
 563          * We already reserve the end partial page before in
 564          *   - i386_start_kernel()
 565          *   - x86_64_start_kernel()
 566          *   - relocate_initrd()
 567          * So here We can do PAGE_ALIGN() safely to get partial page to be freed
 568          */
 569         free_init_pages("initrd memory", start, PAGE_ALIGN(end));
 570 }
 571 #endif
 572
 573 void __init zone_sizes_init(void)
 574 {
 575         unsigned long max_zone_pfns[MAX_NR_ZONES];
 576
 577         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 578
 579 #ifdef CONFIG_ZONE_DMA
 580         max_zone_pfns[ZONE_DMA]         = MAX_DMA_PFN;
 581 #endif
 582 #ifdef CONFIG_ZONE_DMA32
 583         max_zone_pfns[ZONE_DMA32]       = MAX_DMA32_PFN;
 584 #endif
 585         max_zone_pfns[ZONE_NORMAL]      = max_low_pfn;
 586 #ifdef CONFIG_HIGHMEM
 587         max_zone_pfns[ZONE_HIGHMEM]     = max_pfn;
 588 #endif
 589
 590         free_area_init_nodes(max_zone_pfns);
 591 }
 592