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_DISCONTIGMEM
 322 void __init paging_init(void)
 323 {
 324         {
 325                 unsigned long zones_size[MAX_NR_ZONES] = {0, 0, 0};
 326                 unsigned int max_dma;
 327
 328                 max_dma = virt_to_phys((char *)MAX_DMA_ADDRESS) >> PAGE_SHIFT;
 329
 330                 if (end_pfn < max_dma)
 331                         zones_size[ZONE_DMA] = end_pfn;
 332                 else {
 333                         zones_size[ZONE_DMA] = max_dma;
 334                         zones_size[ZONE_NORMAL] = end_pfn - max_dma;
 335                 }
 336                 free_area_init(zones_size);
 337         }
 338         return;
 339 }
 340 #endif
 341
 342 /* Unmap a kernel mapping if it exists. This is useful to avoid prefetches
 343    from the CPU leading to inconsistent cache lines. address and size
 344    must be aligned to 2MB boundaries.
 345    Does nothing when the mapping doesn't exist. */
 346 void __init clear_kernel_mapping(unsigned long address, unsigned long size)
 347 {
 348         unsigned long end = address + size;
 349
 350         BUG_ON(address & ~LARGE_PAGE_MASK);
 351         BUG_ON(size & ~LARGE_PAGE_MASK);
 352
 353         for (; address < end; address += LARGE_PAGE_SIZE) {
 354                 pgd_t *pgd = pgd_offset_k(address);
 355                 pud_t *pud;
 356                 pmd_t *pmd;
 357                 if (pgd_none(*pgd))
 358                         continue;
 359                 pud = pud_offset(pgd, address);
 360                 if (pud_none(*pud))
 361                         continue;
 362                 pmd = pmd_offset(pud, address);
 363                 if (!pmd || pmd_none(*pmd))
 364                         continue;
 365                 if (0 == (pmd_val(*pmd) & _PAGE_PSE)) {
 366                         /* Could handle this, but it should not happen currently. */
 367                         printk(KERN_ERR
 368                "clear_kernel_mapping: mapping has been split. will leak memory\n");
 369                         pmd_ERROR(*pmd);
 370                 }
 371                 set_pmd(pmd, __pmd(0));
 372         }
 373         __flush_tlb_all();
 374 }
 375
 376 static inline int page_is_ram (unsigned long pagenr)
 377 {
 378         int i;
 379
 380         for (i = 0; i < e820.nr_map; i++) {
 381                 unsigned long addr, end;
 382
 383                 if (e820.map[i].type != E820_RAM)       /* not usable memory */
 384                         continue;
 385                 /*
 386                  *      !!!FIXME!!! Some BIOSen report areas as RAM that
 387                  *      are not. Notably the 640->1Mb area. We need a sanity
 388                  *      check here.
 389                  */
 390                 addr = (e820.map[i].addr+PAGE_SIZE-1) >> PAGE_SHIFT;
 391                 end = (e820.map[i].addr+e820.map[i].size) >> PAGE_SHIFT;
 392                 if  ((pagenr >= addr) && (pagenr < end))
 393                         return 1;
 394         }
 395         return 0;
 396 }
 397
 398 extern int swiotlb_force;
 399
 400 static struct kcore_list kcore_mem, kcore_vmalloc, kcore_kernel, kcore_modules,
 401                          kcore_vsyscall;
 402
 403 void __init mem_init(void)
 404 {
 405         int codesize, reservedpages, datasize, initsize;
 406         int tmp;
 407
 408 #ifdef CONFIG_SWIOTLB
 409         if (swiotlb_force)
 410                 swiotlb = 1;
 411         if (!iommu_aperture &&
 412             (end_pfn >= 0xffffffff>>PAGE_SHIFT || force_iommu))
 413                swiotlb = 1;
 414         if (swiotlb)
 415                 swiotlb_init();
 416 #endif
 417
 418         /* How many end-of-memory variables you have, grandma! */
 419         max_low_pfn = end_pfn;
 420         max_pfn = end_pfn;
 421         num_physpages = end_pfn;
 422         high_memory = (void *) __va(end_pfn * PAGE_SIZE);
 423
 424         /* clear the zero-page */
 425         memset(empty_zero_page, 0, PAGE_SIZE);
 426
 427         reservedpages = 0;
 428
 429         /* this will put all low memory onto the freelists */
 430 #ifdef CONFIG_DISCONTIGMEM
 431         totalram_pages += numa_free_all_bootmem();
 432         tmp = 0;
 433         /* should count reserved pages here for all nodes */
 434 #else
 435         max_mapnr = end_pfn;
 436         if (!mem_map) BUG();
 437
 438         totalram_pages += free_all_bootmem();
 439
 440         for (tmp = 0; tmp < end_pfn; tmp++)
 441                 /*
 442                  * Only count reserved RAM pages
 443                  */
 444                 if (page_is_ram(tmp) && PageReserved(pfn_to_page(tmp)))
 445                         reservedpages++;
 446 #endif
 447
 448         after_bootmem = 1;
 449
 450         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 451         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 452         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 453
 454         /* Register memory areas for /proc/kcore */
 455         kclist_add(&kcore_mem, __va(0), max_low_pfn << PAGE_SHIFT);
 456         kclist_add(&kcore_vmalloc, (void *)VMALLOC_START,
 457                    VMALLOC_END-VMALLOC_START);
 458         kclist_add(&kcore_kernel, &_stext, _end - _stext);
 459         kclist_add(&kcore_modules, (void *)MODULES_VADDR, MODULES_LEN);
 460         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 461                                  VSYSCALL_END - VSYSCALL_START);
 462
 463         printk("Memory: %luk/%luk available (%dk kernel code, %dk reserved, %dk data, %dk init)\n",
 464                 (unsigned long) nr_free_pages() << (PAGE_SHIFT-10),
 465                 end_pfn << (PAGE_SHIFT-10),
 466                 codesize >> 10,
 467                 reservedpages << (PAGE_SHIFT-10),
 468                 datasize >> 10,
 469                 initsize >> 10);
 470
 471         /*
 472          * Subtle. SMP is doing its boot stuff late (because it has to
 473          * fork idle threads) - but it also needs low mappings for the
 474          * protected-mode entry to work. We zap these entries only after
 475          * the WP-bit has been tested.
 476          */
 477 #ifndef CONFIG_SMP
 478         zap_low_mappings();
 479 #endif
 480 }
 481
 482 extern char __initdata_begin[], __initdata_end[];
 483
 484 void free_initmem(void)
 485 {
 486         unsigned long addr;
 487
 488         addr = (unsigned long)(&__init_begin);
 489         for (; addr < (unsigned long)(&__init_end); addr += PAGE_SIZE) {
 490                 ClearPageReserved(virt_to_page(addr));
 491                 set_page_count(virt_to_page(addr), 1);
 492                 memset((void *)(addr & ~(PAGE_SIZE-1)), 0xcc, PAGE_SIZE);
 493                 free_page(addr);
 494                 totalram_pages++;
 495         }
 496         memset(__initdata_begin, 0xba, __initdata_end - __initdata_begin);
 497         printk ("Freeing unused kernel memory: %luk freed\n", (&__init_end - &__init_begin) >> 10);
 498 }
 499
 500 #ifdef CONFIG_BLK_DEV_INITRD
 501 void free_initrd_mem(unsigned long start, unsigned long end)
 502 {
 503         if (start < (unsigned long)&_end)
 504                 return;
 505         printk ("Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
 506         for (; start < end; start += PAGE_SIZE) {
 507                 ClearPageReserved(virt_to_page(start));
 508                 set_page_count(virt_to_page(start), 1);
 509                 free_page(start);
 510                 totalram_pages++;
 511         }
 512 }
 513 #endif
 514
 515 void __init reserve_bootmem_generic(unsigned long phys, unsigned len)
 516 {
 517         /* Should check here against the e820 map to avoid double free */
 518 #ifdef CONFIG_DISCONTIGMEM
 519         int nid = phys_to_nid(phys);
 520         reserve_bootmem_node(NODE_DATA(nid), phys, len);
 521 #else
 522         reserve_bootmem(phys, len);
 523 #endif
 524 }
 525
 526 int kern_addr_valid(unsigned long addr)
 527 {
 528         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 529        pgd_t *pgd;
 530        pud_t *pud;
 531        pmd_t *pmd;
 532        pte_t *pte;
 533
 534         if (above != 0 && above != -1UL)
 535                 return 0;
 536
 537         pgd = pgd_offset_k(addr);
 538         if (pgd_none(*pgd))
 539                 return 0;
 540
 541         pud = pud_offset(pgd, addr);
 542         if (pud_none(*pud))
 543                 return 0;
 544
 545         pmd = pmd_offset(pud, addr);
 546         if (pmd_none(*pmd))
 547                 return 0;
 548         if (pmd_large(*pmd))
 549                 return pfn_valid(pmd_pfn(*pmd));
 550
 551         pte = pte_offset_kernel(pmd, addr);
 552         if (pte_none(*pte))
 553                 return 0;
 554         return pfn_valid(pte_pfn(*pte));
 555 }
 556
 557 #ifdef CONFIG_SYSCTL
 558 #include <linux/sysctl.h>
 559
 560 extern int exception_trace, page_fault_trace;
 561
 562 static ctl_table debug_table2[] = {
 563         { 99, "exception-trace", &exception_trace, sizeof(int), 0644, NULL,
 564           proc_dointvec },
 565 #ifdef CONFIG_CHECKING
 566         { 100, "page-fault-trace", &page_fault_trace, sizeof(int), 0644, NULL,
 567           proc_dointvec },
 568 #endif
 569         { 0, }
 570 };
 571
 572 static ctl_table debug_root_table2[] = {
 573         { .ctl_name = CTL_DEBUG, .procname = "debug", .mode = 0555,
 574            .child = debug_table2 },
 575         { 0 },
 576 };
 577
 578 static __init int x8664_sysctl_init(void)
 579 {
 580         register_sysctl_table(debug_root_table2, 1);
 581         return 0;
 582 }
 583 __initcall(x8664_sysctl_init);
 584 #endif
 585
 586 /* Pseudo VMAs to allow ptrace access for the vsyscall pages.  x86-64 has two
 587    different ones: one for 32bit and one for 64bit. Use the appropiate
 588    for the target task. */
 589
 590 static struct vm_area_struct gate_vma = {
 591         .vm_start = VSYSCALL_START,
 592         .vm_end = VSYSCALL_END,
 593         .vm_page_prot = PAGE_READONLY
 594 };
 595
 596 static struct vm_area_struct gate32_vma = {
 597         .vm_start = VSYSCALL32_BASE,
 598         .vm_end = VSYSCALL32_END,
 599         .vm_page_prot = PAGE_READONLY
 600 };
 601
 602 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 603 {
 604 #ifdef CONFIG_IA32_EMULATION
 605         if (test_tsk_thread_flag(tsk, TIF_IA32)) {
 606                 /* lookup code assumes the pages are present. set them up
 607                    now */
 608                 if (__map_syscall32(tsk->mm, VSYSCALL32_BASE) < 0)
 609                         return NULL;
 610                 return &gate32_vma;
 611         }
 612 #endif
 613         return &gate_vma;
 614 }
 615
 616 int in_gate_area(struct task_struct *task, unsigned long addr)
 617 {
 618         struct vm_area_struct *vma = get_gate_vma(task);
 619         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 620 }
 621
 622 /* Use this when you have no reliable task/vma, typically from interrupt
 623  * context.  It is less reliable than using the task's vma and may give
 624  * false positives.
 625  */
 626 int in_gate_area_no_task(unsigned long addr)
 627 {
 628         return (((addr >= VSYSCALL_START) && (addr < VSYSCALL_END)) ||
 629                 ((addr >= VSYSCALL32_BASE) && (addr < VSYSCALL32_END)));
 630 }