arch/x86/mm/init-xen.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>
   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/system.h>
  16 #include <asm/tlbflush.h>
  17 #include <asm/tlb.h>
  18 #include <asm/proto.h>
  19 #include <asm/dma.h>            /* for MAX_DMA_PFN */
  20
  21 unsigned long __meminitdata pgt_buf_start;
  22 unsigned long __meminitdata pgt_buf_end;
  23 unsigned long __meminitdata pgt_buf_top;
  24
  25 int after_bootmem;
  26
  27 #if !defined(CONFIG_XEN)
  28 int direct_gbpages
  29 #ifdef CONFIG_DIRECT_GBPAGES
  30                                 = 1
  31 #endif
  32 ;
  33 #elif defined(CONFIG_X86_32)
  34 #define direct_gbpages 0
  35 extern unsigned long extend_init_mapping(unsigned long tables_space);
  36 #else
  37 extern void xen_finish_init_mapping(void);
  38 #endif
  39
  40 static void __init find_early_table_space(unsigned long end, int use_pse,
  41                                           int use_gbpages)
  42 {
  43         unsigned long puds, pmds, ptes, tables;
  44
  45         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
  46         tables = roundup(puds * sizeof(pud_t), PAGE_SIZE);
  47
  48         if (use_gbpages) {
  49                 unsigned long extra;
  50
  51                 extra = end - ((end>>PUD_SHIFT) << PUD_SHIFT);
  52                 pmds = (extra + PMD_SIZE - 1) >> PMD_SHIFT;
  53         } else
  54                 pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
  55
  56         tables += roundup(pmds * sizeof(pmd_t), PAGE_SIZE);
  57
  58         if (use_pse) {
  59                 unsigned long extra;
  60
  61                 extra = end - ((end>>PMD_SHIFT) << PMD_SHIFT);
  62 #ifdef CONFIG_X86_32
  63                 extra += PMD_SIZE;
  64 #endif
  65                 ptes = (extra + PAGE_SIZE - 1) >> PAGE_SHIFT;
  66         } else
  67                 ptes = (end + PAGE_SIZE - 1) >> PAGE_SHIFT;
  68
  69         tables += roundup(ptes * sizeof(pte_t), PAGE_SIZE);
  70
  71 #ifdef CONFIG_X86_32
  72         /* for fixmap */
  73         tables += roundup(__end_of_fixed_addresses * sizeof(pte_t), PAGE_SIZE);
  74
  75         pgt_buf_start = extend_init_mapping(tables);
  76         pgt_buf_end = pgt_buf_start;
  77 #else /* CONFIG_X86_64 */
  78         if (!pgt_buf_top) {
  79                 pgt_buf_start = (__pa(xen_start_info->pt_base) >> PAGE_SHIFT) +
  80                         xen_start_info->nr_pt_frames;
  81                 pgt_buf_end = pgt_buf_start;
  82         } else {
  83                 /*
  84                  * [table_start, table_top) gets passed to reserve_early(),
  85                  * so we must not use table_end here, despite continuing
  86                  * to allocate from there. table_end possibly being below
  87                  * table_start is otoh not a problem.
  88                  */
  89                 pgt_buf_start = pgt_buf_top;
  90         }
  91 #endif
  92         if (pgt_buf_start == -1UL)
  93                 panic("Cannot find space for the kernel page tables");
  94
  95         pgt_buf_top = pgt_buf_start + (tables >> PAGE_SHIFT);
  96
  97         printk(KERN_DEBUG "kernel direct mapping tables up to %lx @ %lx-%lx\n",
  98                 end, pgt_buf_start << PAGE_SHIFT, pgt_buf_top << PAGE_SHIFT);
  99 }
 100
 101 void __init xen_pagetable_reserve(u64 start, u64 end)
 102 {
 103         if (end > start)
 104                 memblock_reserve(start, end - start);
 105 }
 106
 107 struct map_range {
 108         unsigned long start;
 109         unsigned long end;
 110         unsigned page_size_mask;
 111 };
 112
 113 #ifdef CONFIG_X86_32
 114 #define NR_RANGE_MR 3
 115 #else /* CONFIG_X86_64 */
 116 #define NR_RANGE_MR 5
 117 #endif
 118
 119 static int __meminit save_mr(struct map_range *mr, int nr_range,
 120                              unsigned long start_pfn, unsigned long end_pfn,
 121                              unsigned long page_size_mask)
 122 {
 123         if (start_pfn < end_pfn) {
 124                 if (nr_range >= NR_RANGE_MR)
 125                         panic("run out of range for init_memory_mapping\n");
 126                 mr[nr_range].start = start_pfn<<PAGE_SHIFT;
 127                 mr[nr_range].end   = end_pfn<<PAGE_SHIFT;
 128                 mr[nr_range].page_size_mask = page_size_mask;
 129                 nr_range++;
 130         }
 131
 132         return nr_range;
 133 }
 134
 135 /*
 136  * Setup the direct mapping of the physical memory at PAGE_OFFSET.
 137  * This runs before bootmem is initialized and gets pages directly from
 138  * the physical memory. To access them they are temporarily mapped.
 139  */
 140 unsigned long __init_refok init_memory_mapping(unsigned long start,
 141                                                unsigned long end)
 142 {
 143         unsigned long page_size_mask = 0;
 144         unsigned long start_pfn, end_pfn;
 145         unsigned long ret = 0;
 146         unsigned long pos;
 147
 148         struct map_range mr[NR_RANGE_MR];
 149         int nr_range, i;
 150         int use_pse, use_gbpages;
 151
 152         printk(KERN_INFO "init_memory_mapping: %016lx-%016lx\n", start, end);
 153
 154 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_KMEMCHECK)
 155         /*
 156          * For CONFIG_DEBUG_PAGEALLOC, identity mapping will use small pages.
 157          * This will simplify cpa(), which otherwise needs to support splitting
 158          * large pages into small in interrupt context, etc.
 159          */
 160         use_pse = use_gbpages = 0;
 161 #else
 162         use_pse = cpu_has_pse;
 163         use_gbpages = direct_gbpages;
 164 #endif
 165
 166         /* Enable PSE if available */
 167         if (cpu_has_pse)
 168                 set_in_cr4(X86_CR4_PSE);
 169
 170         /* Enable PGE if available */
 171         if (cpu_has_pge) {
 172                 set_in_cr4(X86_CR4_PGE);
 173                 __supported_pte_mask |= _PAGE_GLOBAL;
 174         }
 175
 176         if (use_gbpages)
 177                 page_size_mask |= 1 << PG_LEVEL_1G;
 178         if (use_pse)
 179                 page_size_mask |= 1 << PG_LEVEL_2M;
 180
 181         memset(mr, 0, sizeof(mr));
 182         nr_range = 0;
 183
 184         /* head if not big page alignment ? */
 185         start_pfn = start >> PAGE_SHIFT;
 186         pos = start_pfn << PAGE_SHIFT;
 187 #ifdef CONFIG_X86_32
 188         /*
 189          * Don't use a large page for the first 2/4MB of memory
 190          * because there are often fixed size MTRRs in there
 191          * and overlapping MTRRs into large pages can cause
 192          * slowdowns.
 193          */
 194         if (pos == 0)
 195                 end_pfn = 1<<(PMD_SHIFT - PAGE_SHIFT);
 196         else
 197                 end_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 198                                  << (PMD_SHIFT - PAGE_SHIFT);
 199 #else /* CONFIG_X86_64 */
 200         end_pfn = ((pos + (PMD_SIZE - 1)) >> PMD_SHIFT)
 201                         << (PMD_SHIFT - PAGE_SHIFT);
 202 #endif
 203         if (end_pfn > (end >> PAGE_SHIFT))
 204                 end_pfn = end >> PAGE_SHIFT;
 205         if (start_pfn < end_pfn) {
 206                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 207                 pos = end_pfn << PAGE_SHIFT;
 208         }
 209
 210         /* big page (2M) range */
 211         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 212                          << (PMD_SHIFT - PAGE_SHIFT);
 213 #ifdef CONFIG_X86_32
 214         end_pfn = (end>>PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 215 #else /* CONFIG_X86_64 */
 216         end_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 217                          << (PUD_SHIFT - PAGE_SHIFT);
 218         if (end_pfn > ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT)))
 219                 end_pfn = ((end>>PMD_SHIFT)<<(PMD_SHIFT - PAGE_SHIFT));
 220 #endif
 221
 222         if (start_pfn < end_pfn) {
 223                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 224                                 page_size_mask & (1<<PG_LEVEL_2M));
 225                 pos = end_pfn << PAGE_SHIFT;
 226         }
 227
 228 #ifdef CONFIG_X86_64
 229         /* big page (1G) range */
 230         start_pfn = ((pos + (PUD_SIZE - 1))>>PUD_SHIFT)
 231                          << (PUD_SHIFT - PAGE_SHIFT);
 232         end_pfn = (end >> PUD_SHIFT) << (PUD_SHIFT - PAGE_SHIFT);
 233         if (start_pfn < end_pfn) {
 234                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 235                                 page_size_mask &
 236                                  ((1<<PG_LEVEL_2M)|(1<<PG_LEVEL_1G)));
 237                 pos = end_pfn << PAGE_SHIFT;
 238         }
 239
 240         /* tail is not big page (1G) alignment */
 241         start_pfn = ((pos + (PMD_SIZE - 1))>>PMD_SHIFT)
 242                          << (PMD_SHIFT - PAGE_SHIFT);
 243         end_pfn = (end >> PMD_SHIFT) << (PMD_SHIFT - PAGE_SHIFT);
 244         if (start_pfn < end_pfn) {
 245                 nr_range = save_mr(mr, nr_range, start_pfn, end_pfn,
 246                                 page_size_mask & (1<<PG_LEVEL_2M));
 247                 pos = end_pfn << PAGE_SHIFT;
 248         }
 249 #endif
 250
 251         /* tail is not big page (2M) alignment */
 252         start_pfn = pos>>PAGE_SHIFT;
 253         end_pfn = end>>PAGE_SHIFT;
 254         nr_range = save_mr(mr, nr_range, start_pfn, end_pfn, 0);
 255
 256         /* try to merge same page size and continuous */
 257         for (i = 0; nr_range > 1 && i < nr_range - 1; i++) {
 258                 unsigned long old_start;
 259                 if (mr[i].end != mr[i+1].start ||
 260                     mr[i].page_size_mask != mr[i+1].page_size_mask)
 261                         continue;
 262                 /* move it */
 263                 old_start = mr[i].start;
 264                 memmove(&mr[i], &mr[i+1],
 265                         (nr_range - 1 - i) * sizeof(struct map_range));
 266                 mr[i--].start = old_start;
 267                 nr_range--;
 268         }
 269
 270         for (i = 0; i < nr_range; i++)
 271                 printk(KERN_DEBUG " %010lx - %010lx page %s\n",
 272                                 mr[i].start, mr[i].end,
 273                         (mr[i].page_size_mask & (1<<PG_LEVEL_1G))?"1G":(
 274                          (mr[i].page_size_mask & (1<<PG_LEVEL_2M))?"2M":"4k"));
 275
 276         /*
 277          * Find space for the kernel direct mapping tables.
 278          *
 279          * Later we should allocate these tables in the local node of the
 280          * memory mapped. Unfortunately this is done currently before the
 281          * nodes are discovered.
 282          */
 283         if (!after_bootmem)
 284                 find_early_table_space(end, use_pse, use_gbpages);
 285
 286 #ifdef CONFIG_X86_64
 287 #define addr_to_page(addr)                                              \
 288         ((unsigned long *)                                              \
 289          ((mfn_to_pfn(((addr) & PHYSICAL_PAGE_MASK) >> PAGE_SHIFT)      \
 290            << PAGE_SHIFT) + __START_KERNEL_map))
 291
 292         if (!start) {
 293                 unsigned long addr, va = __START_KERNEL_map;
 294                 unsigned long *page = (unsigned long *)init_level4_pgt;
 295
 296                 /* Kill mapping of memory below _text. */
 297                 while (va < (unsigned long)&_text) {
 298                         if (HYPERVISOR_update_va_mapping(va, __pte_ma(0), 0))
 299                                 BUG();
 300                         va += PAGE_SIZE;
 301                 }
 302
 303                 /* Blow away any spurious initial mappings. */
 304                 va = __START_KERNEL_map + (pgt_buf_start << PAGE_SHIFT);
 305
 306                 addr = page[pgd_index(va)];
 307                 page = addr_to_page(addr);
 308                 addr = page[pud_index(va)];
 309                 page = addr_to_page(addr);
 310                 while (pmd_index(va) | pte_index(va)) {
 311                         if (pmd_none(*(pmd_t *)&page[pmd_index(va)]))
 312                                 break;
 313                         if (HYPERVISOR_update_va_mapping(va, __pte_ma(0), 0))
 314                                 BUG();
 315                         va += PAGE_SIZE;
 316                 }
 317         }
 318 #undef addr_to_page
 319 #endif
 320
 321         for (i = 0; i < nr_range; i++)
 322                 ret = kernel_physical_mapping_init(mr[i].start, mr[i].end,
 323                                                    mr[i].page_size_mask);
 324
 325 #ifdef CONFIG_X86_32
 326         early_ioremap_page_table_range_init();
 327 #endif
 328
 329 #ifdef CONFIG_X86_64
 330         BUG_ON(pgt_buf_end > pgt_buf_top);
 331         if (!start)
 332                 xen_finish_init_mapping();
 333         else
 334 #endif
 335         if (pgt_buf_end < pgt_buf_top)
 336                 /* Disable the 'table_end' allocator. */
 337                 pgt_buf_top = pgt_buf_end;
 338
 339         __flush_tlb_all();
 340
 341         /*
 342          * Reserve the kernel pagetable pages we used (pgt_buf_start -
 343          * pgt_buf_end) and free the other ones (pgt_buf_end - pgt_buf_top)
 344          * so that they can be reused for other purposes.
 345          *
 346          * On native it just means calling memblock_reserve, on Xen it also
 347          * means marking RW the pagetable pages that we allocated before
 348          * but that haven't been used.
 349          *
 350          * In fact on xen we mark RO the whole range pgt_buf_start -
 351          * pgt_buf_top, because we have to make sure that when
 352          * init_memory_mapping reaches the pagetable pages area, it maps
 353          * RO all the pagetable pages, including the ones that are beyond
 354          * pgt_buf_end at that time.
 355          */
 356         if (!after_bootmem && pgt_buf_top > pgt_buf_start) {
 357 #ifdef CONFIG_X86_64
 358                 reserve_pgtable_low();
 359 #endif
 360                 x86_init.mapping.pagetable_reserve(PFN_PHYS(pgt_buf_start),
 361                                 PFN_PHYS(pgt_buf_top));
 362         }
 363
 364         if (!after_bootmem)
 365                 early_memtest(start, end);
 366
 367         return ret >> PAGE_SHIFT;
 368 }
 369
 370
 371 /*
 372  * devmem_is_allowed() checks to see if /dev/mem access to a certain address
 373  * is valid. The argument is a physical page number.
 374  *
 375  *
 376  * On x86, access has to be given to the first megabyte of ram because that area
 377  * contains bios code and data regions used by X and dosemu and similar apps.
 378  * Access has to be given to non-kernel-ram areas as well, these contain the PCI
 379  * mmio resources as well as potential bios/acpi data regions.
 380  */
 381 int devmem_is_allowed(unsigned long pagenr)
 382 {
 383         if (pagenr <= 256)
 384                 return 1;
 385         if (iomem_is_exclusive(pagenr << PAGE_SHIFT))
 386                 return 0;
 387         if (mfn_to_local_pfn(pagenr) >= max_pfn)
 388                 return 1;
 389         return 0;
 390 }
 391
 392 void free_init_pages(char *what, unsigned long begin, unsigned long end)
 393 {
 394         unsigned long addr;
 395         unsigned long begin_aligned, end_aligned;
 396
 397         /* Make sure boundaries are page aligned */
 398         begin_aligned = PAGE_ALIGN(begin);
 399         end_aligned   = end & PAGE_MASK;
 400
 401         if (WARN_ON(begin_aligned != begin || end_aligned != end)) {
 402                 begin = begin_aligned;
 403                 end   = end_aligned;
 404         }
 405
 406         if (begin >= end)
 407                 return;
 408
 409         addr = begin;
 410
 411         /*
 412          * If debugging page accesses then do not free this memory but
 413          * mark them not present - any buggy init-section access will
 414          * create a kernel page fault:
 415          */
 416 #ifdef CONFIG_DEBUG_PAGEALLOC
 417         printk(KERN_INFO "debug: unmapping init memory %08lx..%08lx\n",
 418                 begin, end);
 419         set_memory_np(begin, (end - begin) >> PAGE_SHIFT);
 420 #else
 421         /*
 422          * We just marked the kernel text read only above, now that
 423          * we are going to free part of that, we need to make that
 424          * writeable and non-executable first.
 425          */
 426         set_memory_nx(begin, (end - begin) >> PAGE_SHIFT);
 427         set_memory_rw(begin, (end - begin) >> PAGE_SHIFT);
 428
 429         printk(KERN_INFO "Freeing %s: %luk freed\n", what, (end - begin) >> 10);
 430
 431         for (; addr < end; addr += PAGE_SIZE) {
 432                 ClearPageReserved(virt_to_page(addr));
 433                 init_page_count(virt_to_page(addr));
 434                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
 435 #ifdef CONFIG_X86_64
 436                 if (addr >= __START_KERNEL_map) {
 437                         /* make_readonly() reports all kernel addresses. */
 438                         if (HYPERVISOR_update_va_mapping((unsigned long)__va(__pa(addr)),
 439                                                          pfn_pte(__pa(addr) >> PAGE_SHIFT,
 440                                                                  PAGE_KERNEL),
 441                                                          0))
 442                                 BUG();
 443                         if (HYPERVISOR_update_va_mapping(addr, __pte(0), 0))
 444                                 BUG();
 445                 }
 446 #endif
 447                 free_page(addr);
 448                 totalram_pages++;
 449         }
 450 #endif
 451 }
 452
 453 void free_initmem(void)
 454 {
 455         free_init_pages("unused kernel memory",
 456                         (unsigned long)(&__init_begin),
 457                         (unsigned long)(&__init_end));
 458 }
 459
 460 #ifdef CONFIG_BLK_DEV_INITRD
 461 void free_initrd_mem(unsigned long start, unsigned long end)
 462 {
 463         /*
 464          * end could be not aligned, and We can not align that,
 465          * decompresser could be confused by aligned initrd_end
 466          * We already reserve the end partial page before in
 467          *   - i386_start_kernel()
 468          *   - x86_64_start_kernel()
 469          *   - relocate_initrd()
 470          * So here We can do PAGE_ALIGN() safely to get partial page to be freed
 471          */
 472 #ifdef CONFIG_ACPI_INITRD_TABLE_OVERRIDE
 473         if (acpi_initrd_offset)
 474                 free_init_pages("initrd memory", start - acpi_initrd_offset,
 475                                 PAGE_ALIGN(end));
 476         else
 477 #endif
 478         free_init_pages("initrd memory", start, PAGE_ALIGN(end));
 479 }
 480 #endif
 481
 482 void __init zone_sizes_init(void)
 483 {
 484         unsigned long max_zone_pfns[MAX_NR_ZONES];
 485
 486         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 487
 488 #ifdef CONFIG_ZONE_DMA
 489         max_zone_pfns[ZONE_DMA]         = MAX_DMA_PFN;
 490 #endif
 491 #ifdef CONFIG_ZONE_DMA32
 492         max_zone_pfns[ZONE_DMA32]       = MAX_DMA32_PFN;
 493 #endif
 494         max_zone_pfns[ZONE_NORMAL]      = max_low_pfn;
 495 #ifdef CONFIG_HIGHMEM
 496         max_zone_pfns[ZONE_HIGHMEM]     = max_pfn;
 497 #endif
 498
 499         free_area_init_nodes(max_zone_pfns);
 500
 501         xen_init_pgd_pin();
 502 }
 503