arch/x86_64/mm/init.c

   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/config.h>
  10 #include <linux/signal.h>
  11 #include <linux/sched.h>
  12 #include <linux/kernel.h>
  13 #include <linux/errno.h>
  14 #include <linux/string.h>
  15 #include <linux/types.h>
  16 #include <linux/ptrace.h>
  17 #include <linux/mman.h>
  18 #include <linux/mm.h>
  19 #include <linux/swap.h>
  20 #include <linux/smp.h>
  21 #include <linux/init.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/bootmem.h>
  24 #include <linux/proc_fs.h>
  25
  26 #include <asm/processor.h>
  27 #include <asm/system.h>
  28 #include <asm/uaccess.h>
  29 #include <asm/pgtable.h>
  30 #include <asm/pgalloc.h>
  31 #include <asm/dma.h>
  32 #include <asm/fixmap.h>
  33 #include <asm/e820.h>
  34 #include <asm/apic.h>
  35 #include <asm/tlb.h>
  36 #include <asm/mmu_context.h>
  37 #include <asm/proto.h>
  38 #include <asm/smp.h>
  39
  40 #ifndef Dprintk
  41 #define Dprintk(x...)
  42 #endif
  43
  44 #ifdef CONFIG_GART_IOMMU
  45 extern int swiotlb;
  46 #endif
  47
  48 extern char _stext[];
  49
  50 DEFINE_PER_CPU(struct mmu_gather, mmu_gathers);
  51
  52 /*
  53  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  54  * physical space so we can cache the place of the first one and move
  55  * around without checking the pgd every time.
  56  */
  57
  58 void show_mem(void)
  59 {
  60         int i, total = 0, reserved = 0;
  61         int shared = 0, cached = 0;
  62         pg_data_t *pgdat;
  63         struct page *page;
  64
  65         printk("Mem-info:\n");
  66         show_free_areas();
  67         printk("Free swap:       %6ldkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
  68
  69         for_each_pgdat(pgdat) {
  70                for (i = 0; i < pgdat->node_spanned_pages; ++i) {
  71                         page = pfn_to_page(pgdat->node_start_pfn + i);
  72                         total++;
  73                        if (PageReserved(page))
  74                         reserved++;
  75                        else if (PageSwapCache(page))
  76                         cached++;
  77                        else if (page_count(page))
  78                                shared += page_count(page) - 1;
  79                }
  80         }
  81         printk("%d pages of RAM\n", total);
  82         printk("%d reserved pages\n",reserved);
  83         printk("%d pages shared\n",shared);
  84         printk("%d pages swap cached\n",cached);
  85 }
  86
  87 /* References to section boundaries */
  88
  89 extern char _text, _etext, _edata, __bss_start, _end[];
  90 extern char __init_begin, __init_end;
  91
  92 int after_bootmem;
  93
  94 static void *spp_getpage(void)
  95 {
  96         void *ptr;
  97         if (after_bootmem)
  98                 ptr = (void *) get_zeroed_page(GFP_ATOMIC);
  99         else
 100                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 101         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK))
 102                 panic("set_pte_phys: cannot allocate page data %s\n", after_bootmem?"after bootmem":"");
 103
 104         Dprintk("spp_getpage %p\n", ptr);
 105         return ptr;
 106 }
 107
 108 static void set_pte_phys(unsigned long vaddr,
 109                          unsigned long phys, pgprot_t prot)
 110 {
 111         pgd_t *pgd;
 112         pud_t *pud;
 113         pmd_t *pmd;
 114         pte_t *pte, new_pte;
 115
 116         Dprintk("set_pte_phys %lx to %lx\n", vaddr, phys);
 117
 118         pgd = pgd_offset_k(vaddr);
 119         if (pgd_none(*pgd)) {
 120                 printk("PGD FIXMAP MISSING, it should be setup in head.S!\n");
 121                 return;
 122         }
 123         pud = pud_offset(pgd, vaddr);
 124         if (pud_none(*pud)) {
 125                 pmd = (pmd_t *) spp_getpage();
 126                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE | _PAGE_USER));
 127                 if (pmd != pmd_offset(pud, 0)) {
 128                         printk("PAGETABLE BUG #01! %p <-> %p\n", pmd, pmd_offset(pud,0));
 129                         return;
 130                 }
 131         }
 132         pmd = pmd_offset(pud, vaddr);
 133         if (pmd_none(*pmd)) {
 134                 pte = (pte_t *) spp_getpage();
 135                 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE | _PAGE_USER));
 136                 if (pte != pte_offset_kernel(pmd, 0)) {
 137                         printk("PAGETABLE BUG #02!\n");
 138                         return;
 139                 }
 140         }
 141         new_pte = pfn_pte(phys >> PAGE_SHIFT, prot);
 142
 143         pte = pte_offset_kernel(pmd, vaddr);
 144         if (!pte_none(*pte) &&
 145             pte_val(*pte) != (pte_val(new_pte) & __supported_pte_mask))
 146                 pte_ERROR(*pte);
 147         set_pte(pte, new_pte);
 148
 149         /*
 150          * It's enough to flush this one mapping.
 151          * (PGE mappings get flushed as well)
 152          */
 153         __flush_tlb_one(vaddr);
 154 }
 155
 156 /* NOTE: this is meant to be run only at boot */
 157 void __set_fixmap (enum fixed_addresses idx, unsigned long phys, pgprot_t prot)
 158 {
 159         unsigned long address = __fix_to_virt(idx);
 160
 161         if (idx >= __end_of_fixed_addresses) {
 162                 printk("Invalid __set_fixmap\n");
 163                 return;
 164         }
 165         set_pte_phys(address, phys, prot);
 166 }
 167
 168 unsigned long __initdata table_start, table_end;
 169
 170 extern pmd_t temp_boot_pmds[];
 171
 172 static  struct temp_map {
 173         pmd_t *pmd;
 174         void  *address;
 175         int    allocated;
 176 } temp_mappings[] __initdata = {
 177         { &temp_boot_pmds[0], (void *)(40UL * 1024 * 1024) },
 178         { &temp_boot_pmds[1], (void *)(42UL * 1024 * 1024) },
 179         {}
 180 };
 181
 182 static __init void *alloc_low_page(int *index, unsigned long *phys)
 183 {
 184         struct temp_map *ti;
 185         int i;
 186         unsigned long pfn = table_end++, paddr;
 187         void *adr;
 188
 189         if (pfn >= end_pfn)
 190                 panic("alloc_low_page: ran out of memory");
 191         for (i = 0; temp_mappings[i].allocated; i++) {
 192                 if (!temp_mappings[i].pmd)
 193                         panic("alloc_low_page: ran out of temp mappings");
 194         }
 195         ti = &temp_mappings[i];
 196         paddr = (pfn << PAGE_SHIFT) & PMD_MASK;
 197         set_pmd(ti->pmd, __pmd(paddr | _KERNPG_TABLE | _PAGE_PSE));
 198         ti->allocated = 1;
 199         __flush_tlb();
 200         adr = ti->address + ((pfn << PAGE_SHIFT) & ~PMD_MASK);
 201         *index = i;
 202         *phys  = pfn * PAGE_SIZE;
 203         return adr;
 204 }
 205
 206 static __init void unmap_low_page(int i)
 207 {
 208         struct temp_map *ti = &temp_mappings[i];
 209         set_pmd(ti->pmd, __pmd(0));
 210         ti->allocated = 0;
 211 }
 212
 213 static void __init phys_pud_init(pud_t *pud, unsigned long address, unsigned long end)
 214 {
 215         long i, j;
 216
 217         i = pud_index(address);
 218         pud = pud + i;
 219         for (; i < PTRS_PER_PUD; pud++, i++) {
 220                 int map;
 221                 unsigned long paddr, pmd_phys;
 222                 pmd_t *pmd;
 223
 224                 paddr = address + i*PUD_SIZE;
 225                 if (paddr >= end) {
 226                         for (; i < PTRS_PER_PUD; i++, pud++)
 227                                 set_pud(pud, __pud(0));
 228                         break;
 229                 }
 230
 231                 if (!e820_mapped(paddr, paddr+PUD_SIZE, 0)) {
 232                         set_pud(pud, __pud(0));
 233                         continue;
 234                 }
 235
 236                 pmd = alloc_low_page(&map, &pmd_phys);
 237                 set_pud(pud, __pud(pmd_phys | _KERNPG_TABLE));
 238                 for (j = 0; j < PTRS_PER_PMD; pmd++, j++, paddr += PMD_SIZE) {
 239                         unsigned long pe;
 240
 241                         if (paddr >= end) {
 242                                 for (; j < PTRS_PER_PMD; j++, pmd++)
 243                                         set_pmd(pmd,  __pmd(0));
 244                                 break;
 245                 }
 246                         pe = _PAGE_NX|_PAGE_PSE | _KERNPG_TABLE | _PAGE_GLOBAL | paddr;
 247                         pe &= __supported_pte_mask;
 248                         set_pmd(pmd, __pmd(pe));
 249                 }
 250                 unmap_low_page(map);
 251         }
 252         __flush_tlb();
 253 }
 254
 255 static void __init find_early_table_space(unsigned long end)
 256 {
 257         unsigned long puds, pmds, tables;
 258
 259         puds = (end + PUD_SIZE - 1) >> PUD_SHIFT;
 260         pmds = (end + PMD_SIZE - 1) >> PMD_SHIFT;
 261         tables = round_up(puds * sizeof(pud_t), PAGE_SIZE) +
 262                  round_up(pmds * sizeof(pmd_t), PAGE_SIZE);
 263
 264         table_start = find_e820_area(0x8000, __pa_symbol(&_text), tables);
 265         if (table_start == -1UL)
 266                 panic("Cannot find space for the kernel page tables");
 267
 268         table_start >>= PAGE_SHIFT;
 269         table_end = table_start;
 270 }
 271
 272 /* Setup the direct mapping of the physical memory at PAGE_OFFSET.
 273    This runs before bootmem is initialized and gets pages directly from the
 274    physical memory. To access them they are temporarily mapped. */
 275 void __init init_memory_mapping(unsigned long start, unsigned long end)
 276 {
 277         unsigned long next;
 278
 279         Dprintk("init_memory_mapping\n");
 280
 281         /*
 282          * Find space for the kernel direct mapping tables.
 283          * Later we should allocate these tables in the local node of the memory
 284          * mapped.  Unfortunately this is done currently before the nodes are
 285          * discovered.
 286          */
 287         find_early_table_space(end);
 288
 289         start = (unsigned long)__va(start);
 290         end = (unsigned long)__va(end);
 291
 292         for (; start < end; start = next) {
 293                 int map;
 294                 unsigned long pud_phys;
 295                 pud_t *pud = alloc_low_page(&map, &pud_phys);
 296                 next = start + PGDIR_SIZE;
 297                 if (next > end)
 298                         next = end;
 299                 phys_pud_init(pud, __pa(start), __pa(next));
 300                 set_pgd(pgd_offset_k(start), mk_kernel_pgd(pud_phys));
 301                 unmap_low_page(map);
 302         }
 303
 304         asm volatile("movq %%cr4,%0" : "=r" (mmu_cr4_features));
 305         __flush_tlb_all();
 306         early_printk("kernel direct mapping tables upto %lx @ %lx-%lx\n", end,
 307                table_start<<PAGE_SHIFT,
 308                table_end<<PAGE_SHIFT);
 309 }
 310
 311 extern struct x8664_pda cpu_pda[NR_CPUS];
 312
 313 /* Assumes all CPUs still execute in init_mm */
 314 void zap_low_mappings(void)
 315 {
 316         pgd_t *pgd = pgd_offset_k(0UL);
 317         pgd_clear(pgd);
 318         flush_tlb_all();
 319 }
 320
 321 #ifndef CONFIG_NUMA
 322 void __init paging_init(void)
 323 {
 324         {
 325                 unsigned long zones_size[MAX_NR_ZONES];
 326                 unsigned long holes[MAX_NR_ZONES];
 327                 unsigned int max_dma;
 328
 329                 memset(zones_size, 0, sizeof(zones_size));
 330                 memset(holes, 0, sizeof(holes));
 331
 332                 max_dma = virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
 333
 334                 if (end_pfn < max_dma) {
 335                         zones_size[ZONE_DMA] = end_pfn;
 336                         holes[ZONE_DMA] = e820_hole_size(0, end_pfn);
 337                 } else {
 338                         zones_size[ZONE_DMA] = max_dma;
 339                         holes[ZONE_DMA] = e820_hole_size(0, max_dma);
 340                         zones_size[ZONE_NORMAL] = end_pfn - max_dma;
 341                         holes[ZONE_NORMAL] = e820_hole_size(max_dma, end_pfn);
 342                 }
 343                 free_area_init_node(0, NODE_DATA(0), zones_size,
 344                         __pa(PAGE_OFFSET) >> PAGE_SHIFT, holes);
 345         }
 346         return;
 347 }
 348 #endif
 349
 350 /* Unmap a kernel mapping if it exists. This is useful to avoid prefetches
 351    from the CPU leading to inconsistent cache lines. address and size
 352    must be aligned to 2MB boundaries.
 353    Does nothing when the mapping doesn't exist. */
 354 void __init clear_kernel_mapping(unsigned long address, unsigned long size)
 355 {
 356         unsigned long end = address + size;
 357
 358         BUG_ON(address & ~LARGE_PAGE_MASK);
 359         BUG_ON(size & ~LARGE_PAGE_MASK);
 360
 361         for (; address < end; address += LARGE_PAGE_SIZE) {
 362                 pgd_t *pgd = pgd_offset_k(address);
 363                 pud_t *pud;
 364                 pmd_t *pmd;
 365                 if (pgd_none(*pgd))
 366                         continue;
 367                 pud = pud_offset(pgd, address);
 368                 if (pud_none(*pud))
 369                         continue;
 370                 pmd = pmd_offset(pud, address);
 371                 if (!pmd || pmd_none(*pmd))
 372                         continue;
 373                 if (0 == (pmd_val(*pmd) & _PAGE_PSE)) {
 374                         /* Could handle this, but it should not happen currently. */
 375                         printk(KERN_ERR
 376                "clear_kernel_mapping: mapping has been split. will leak memory\n");
 377                         pmd_ERROR(*pmd);
 378                 }
 379                 set_pmd(pmd, __pmd(0));
 380         }
 381         __flush_tlb_all();
 382 }
 383
 384 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel, kcore_modules,
 385                          kcore_vsyscall;
 386
 387 void __init mem_init(void)
 388 {
 389         long codesize, reservedpages, datasize, initsize;
 390
 391 #ifdef CONFIG_SWIOTLB
 392         if (!iommu_aperture &&
 393             (end_pfn >= 0xffffffff>>PAGE_SHIFT || force_iommu))
 394                swiotlb = 1;
 395         if (swiotlb)
 396                 swiotlb_init();
 397 #endif
 398
 399         /* How many end-of-memory variables you have, grandma! */
 400         max_low_pfn = end_pfn;
 401         max_pfn = end_pfn;
 402         num_physpages = end_pfn;
 403         high_memory = (void *) __va(end_pfn * PAGE_SIZE);
 404
 405         /* clear the zero-page */
 406         memset(empty_zero_page, 0, PAGE_SIZE);
 407
 408         reservedpages = 0;
 409
 410         /* this will put all low memory onto the freelists */
 411 #ifdef CONFIG_NUMA
 412         totalram_pages = numa_free_all_bootmem();
 413 #else
 414
 415 #ifdef CONFIG_FLATMEM
 416         max_mapnr = end_pfn;
 417         if (!mem_map) BUG();
 418 #endif
 419         totalram_pages = free_all_bootmem();
 420 #endif
 421         reservedpages = end_pfn - totalram_pages - e820_hole_size(0, end_pfn);
 422
 423         after_bootmem = 1;
 424
 425         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 426         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 427         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 428
 429         /* Register memory areas for /proc/kcore */
 430         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 431         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 432                    VMALLOC_END-VMALLOC_START);
 433         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 434         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 435         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 436                                  VSYSCALL_END - VSYSCALL_START);
 437
 438         printk("Memory: %luk/%luk available (%ldk kernel code, %ldk reserved, %ldk data, %ldk init)\n",
 439                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 440                 end_pfn << (PAGE_SHIFT-10),
 441                 codesize >> 10,
 442                 reservedpages << (PAGE_SHIFT-10),
 443                 datasize >> 10,
 444                 initsize >> 10);
 445
 446         /*
 447          * Subtle. SMP is doing its boot stuff late (because it has to
 448          * fork idle threads) - but it also needs low mappings for the
 449          * protected-mode entry to work. We zap these entries only after
 450          * the WP-bit has been tested.
 451          */
 452 #ifndef CONFIG_SMP
 453         zap_low_mappings();
 454 #endif
 455 }
 456
 457 extern char __initdata_begin[], __initdata_end[];
 458
 459 void free_initmem(void)
 460 {
 461         unsigned long addr;
 462
 463         addr = (unsigned long)(&__init_begin);
 464         for (; addr < (unsigned long)(&__init_end); addr += PAGE_SIZE) {
 465                 ClearPageReserved(virt_to_page(addr));
 466                 set_page_count(virt_to_page(addr), 1);
 467                 memset((void *)(addr & ~(PAGE_SIZE-1)), 0xcc, PAGE_SIZE);
 468                 free_page(addr);
 469                 totalram_pages++;
 470         }
 471         memset(__initdata_begin, 0xba, __initdata_end - __initdata_begin);
 472         printk ("Freeing unused kernel memory: %luk freed\n", (&__init_end - &__init_begin) >> 10);
 473 }
 474
 475 #ifdef CONFIG_BLK_DEV_INITRD
 476 void free_initrd_mem(unsigned long start, unsigned long end)
 477 {
 478         if (start < (unsigned long)&_end)
 479                 return;
 480         printk ("Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
 481         for (; start < end; start += PAGE_SIZE) {
 482                 ClearPageReserved(virt_to_page(start));
 483                 set_page_count(virt_to_page(start), 1);
 484                 free_page(start);
 485                 totalram_pages++;
 486         }
 487 }
 488 #endif
 489
 490 void __init reserve_bootmem_generic(unsigned long phys, unsigned len)
 491 {
 492         /* Should check here against the e820 map to avoid double free */
 493 #ifdef CONFIG_NUMA
 494         int nid = phys_to_nid(phys);
 495         reserve_bootmem_node(NODE_DATA(nid), phys, len);
 496 #else
 497         reserve_bootmem(phys, len);
 498 #endif
 499 }
 500
 501 int kern_addr_valid(unsigned long addr)
 502 {
 503         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 504        pgd_t *pgd;
 505        pud_t *pud;
 506        pmd_t *pmd;
 507        pte_t *pte;
 508
 509         if (above != 0 && above != -1UL)
 510                 return 0;
 511
 512         pgd = pgd_offset_k(addr);
 513         if (pgd_none(*pgd))
 514                 return 0;
 515
 516         pud = pud_offset(pgd, addr);
 517         if (pud_none(*pud))
 518                 return 0;
 519
 520         pmd = pmd_offset(pud, addr);
 521         if (pmd_none(*pmd))
 522                 return 0;
 523         if (pmd_large(*pmd))
 524                 return pfn_valid(pmd_pfn(*pmd));
 525
 526         pte = pte_offset_kernel(pmd, addr);
 527         if (pte_none(*pte))
 528                 return 0;
 529         return pfn_valid(pte_pfn(*pte));
 530 }
 531
 532 #ifdef CONFIG_SYSCTL
 533 #include <linux/sysctl.h>
 534
 535 extern int exception_trace, page_fault_trace;
 536
 537 static ctl_table debug_table2[] = {
 538         { 99, "exception-trace", &exception_trace, sizeof(int), 0644, NULL,
 539           proc_dointvec },
 540 #ifdef CONFIG_CHECKING
 541         { 100, "page-fault-trace", &page_fault_trace, sizeof(int), 0644, NULL,
 542           proc_dointvec },
 543 #endif
 544         { 0, }
 545 };
 546
 547 static ctl_table debug_root_table2[] = {
 548         { .ctl_name = CTL_DEBUG, .procname = "debug", .mode = 0555,
 549            .child = debug_table2 },
 550         { 0 },
 551 };
 552
 553 static __init int x8664_sysctl_init(void)
 554 {
 555         register_sysctl_table(debug_root_table2, 1);
 556         return 0;
 557 }
 558 __initcall(x8664_sysctl_init);
 559 #endif
 560
 561 /* A pseudo VMAs to allow ptrace access for the vsyscall page.   This only
 562    covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 563    not need special handling anymore. */
 564
 565 static struct vm_area_struct gate_vma = {
 566         .vm_start = VSYSCALL_START,
 567         .vm_end = VSYSCALL_END,
 568         .vm_page_prot = PAGE_READONLY
 569 };
 570
 571 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 572 {
 573 #ifdef CONFIG_IA32_EMULATION
 574         if (test_tsk_thread_flag(tsk, TIF_IA32))
 575                 return NULL;
 576 #endif
 577         return &gate_vma;
 578 }
 579
 580 int in_gate_area(struct task_struct *task, unsigned long addr)
 581 {
 582         struct vm_area_struct *vma = get_gate_vma(task);
 583         if (!vma)
 584                 return 0;
 585         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 586 }
 587
 588 /* Use this when you have no reliable task/vma, typically from interrupt
 589  * context.  It is less reliable than using the task's vma and may give
 590  * false positives.
 591  */
 592 int in_gate_area_no_task(unsigned long addr)
 593 {
 594         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 595 }