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