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