arch/ia64/mm/init.c

   1 /*
   2  * Initialize MMU support.
   3  *
   4  * Copyright (C) 1998-2002 Hewlett-Packard Co
   5  *      David Mosberger-Tang <davidm@hpl.hp.com>
   6  */
   7 #include <linux/config.h>
   8 #include <linux/kernel.h>
   9 #include <linux/init.h>
  10
  11 #include <linux/bootmem.h>
  12 #include <linux/mm.h>
  13 #include <linux/personality.h>
  14 #include <linux/reboot.h>
  15 #include <linux/slab.h>
  16 #include <linux/swap.h>
  17 #include <linux/efi.h>
  18 #include <linux/mmzone.h>
  19
  20 #include <asm/a.out.h>
  21 #include <asm/bitops.h>
  22 #include <asm/dma.h>
  23 #include <asm/ia32.h>
  24 #include <asm/io.h>
  25 #include <asm/machvec.h>
  26 #include <asm/pgalloc.h>
  27 #include <asm/sal.h>
  28 #include <asm/system.h>
  29 #include <asm/uaccess.h>
  30 #include <asm/tlb.h>
  31
  32 struct mmu_gather mmu_gathers[NR_CPUS];
  33
  34 /* References to section boundaries: */
  35 extern char _stext, _etext, _edata, __init_begin, __init_end;
  36
  37 extern void ia64_tlb_init (void);
  38
  39 unsigned long MAX_DMA_ADDRESS = PAGE_OFFSET + 0x100000000UL;
  40
  41 static int pgt_cache_water[2] = { 25, 50 };
  42
  43 void
  44 check_pgt_cache (void)
  45 {
  46         int low, high;
  47
  48         low = pgt_cache_water[0];
  49         high = pgt_cache_water[1];
  50
  51         if (pgtable_cache_size > high) {
  52                 do {
  53                         if (pgd_quicklist)
  54                                 free_page((unsigned long)pgd_alloc_one_fast(0));
  55                         if (pmd_quicklist)
  56                                 free_page((unsigned long)pmd_alloc_one_fast(0, 0));
  57                 } while (pgtable_cache_size > low);
  58         }
  59 }
  60
  61 /*
  62  * This performs some platform-dependent address space initialization.
  63  * On IA-64, we want to setup the VM area for the register backing
  64  * store (which grows upwards) and install the gateway page which is
  65  * used for signal trampolines, etc.
  66  */
  67 void
  68 ia64_init_addr_space (void)
  69 {
  70         struct vm_area_struct *vma;
  71
  72         /*
  73          * If we're out of memory and kmem_cache_alloc() returns NULL, we simply ignore
  74          * the problem.  When the process attempts to write to the register backing store
  75          * for the first time, it will get a SEGFAULT in this case.
  76          */
  77         vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
  78         if (vma) {
  79                 vma->vm_mm = current->mm;
  80                 vma->vm_start = IA64_RBS_BOT;
  81                 vma->vm_end = vma->vm_start + PAGE_SIZE;
  82                 vma->vm_page_prot = protection_map[VM_DATA_DEFAULT_FLAGS & 0x7];
  83                 vma->vm_flags = VM_READ|VM_WRITE|VM_MAYREAD|VM_MAYWRITE|VM_GROWSUP;
  84                 vma->vm_ops = NULL;
  85                 vma->vm_pgoff = 0;
  86                 vma->vm_file = NULL;
  87                 vma->vm_private_data = NULL;
  88                 insert_vm_struct(current->mm, vma);
  89         }
  90
  91         /* map NaT-page at address zero to speed up speculative dereferencing of NULL: */
  92         if (!(current->personality & MMAP_PAGE_ZERO)) {
  93                 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
  94                 if (vma) {
  95                         memset(vma, 0, sizeof(*vma));
  96                         vma->vm_mm = current->mm;
  97                         vma->vm_end = PAGE_SIZE;
  98                         vma->vm_page_prot = __pgprot(pgprot_val(PAGE_READONLY) | _PAGE_MA_NAT);
  99                         vma->vm_flags = VM_READ | VM_MAYREAD | VM_IO | VM_RESERVED;
 100                         insert_vm_struct(current->mm, vma);
 101                 }
 102         }
 103 }
 104
 105 void
 106 free_initmem (void)
 107 {
 108         unsigned long addr;
 109
 110         addr = (unsigned long) &__init_begin;
 111         for (; addr < (unsigned long) &__init_end; addr += PAGE_SIZE) {
 112                 ClearPageReserved(virt_to_page(addr));
 113                 set_page_count(virt_to_page(addr), 1);
 114                 free_page(addr);
 115                 ++totalram_pages;
 116         }
 117         printk(KERN_INFO "Freeing unused kernel memory: %ldkB freed\n",
 118                (&__init_end - &__init_begin) >> 10);
 119 }
 120
 121 void
 122 free_initrd_mem (unsigned long start, unsigned long end)
 123 {
 124         struct page *page;
 125         /*
 126          * EFI uses 4KB pages while the kernel can use 4KB  or bigger.
 127          * Thus EFI and the kernel may have different page sizes. It is
 128          * therefore possible to have the initrd share the same page as
 129          * the end of the kernel (given current setup).
 130          *
 131          * To avoid freeing/using the wrong page (kernel sized) we:
 132          *      - align up the beginning of initrd
 133          *      - align down the end of initrd
 134          *
 135          *  |             |
 136          *  |=============| a000
 137          *  |             |
 138          *  |             |
 139          *  |             | 9000
 140          *  |/////////////|
 141          *  |/////////////|
 142          *  |=============| 8000
 143          *  |///INITRD////|
 144          *  |/////////////|
 145          *  |/////////////| 7000
 146          *  |             |
 147          *  |KKKKKKKKKKKKK|
 148          *  |=============| 6000
 149          *  |KKKKKKKKKKKKK|
 150          *  |KKKKKKKKKKKKK|
 151          *  K=kernel using 8KB pages
 152          *
 153          * In this example, we must free page 8000 ONLY. So we must align up
 154          * initrd_start and keep initrd_end as is.
 155          */
 156         start = PAGE_ALIGN(start);
 157         end = end & PAGE_MASK;
 158
 159         if (start < end)
 160                 printk(KERN_INFO "Freeing initrd memory: %ldkB freed\n", (end - start) >> 10);
 161
 162         for (; start < end; start += PAGE_SIZE) {
 163                 if (!virt_addr_valid(start))
 164                         continue;
 165                 page = virt_to_page(start);
 166                 ClearPageReserved(page);
 167                 set_page_count(page, 1);
 168                 free_page(start);
 169                 ++totalram_pages;
 170         }
 171 }
 172
 173 void
 174 show_mem(void)
 175 {
 176         int i, total = 0, reserved = 0;
 177         int shared = 0, cached = 0;
 178
 179         printk("Mem-info:\n");
 180         show_free_areas();
 181
 182 #ifdef CONFIG_DISCONTIGMEM
 183         {
 184                 pg_data_t *pgdat;
 185
 186                 printk("Free swap:       %6dkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
 187                 for_each_pgdat(pgdat) {
 188                         printk("Node ID: %d\n", pgdat->node_id);
 189                         for(i = 0; i < pgdat->node_size; i++) {
 190                                 if (PageReserved(pgdat->node_mem_map+i))
 191                                         reserved++;
 192                                 else if (PageSwapCache(pgdat->node_mem_map+i))
 193                                         cached++;
 194                                 else if (page_count(pgdat->node_mem_map + i))
 195                                         shared += page_count(pgdat->node_mem_map + i) - 1;
 196                         }
 197                         printk("\t%d pages of RAM\n", pgdat->node_size);
 198                         printk("\t%d reserved pages\n", reserved);
 199                         printk("\t%d pages shared\n", shared);
 200                         printk("\t%d pages swap cached\n", cached);
 201                 }
 202                 printk("Total of %ld pages in page table cache\n", pgtable_cache_size);
 203                 printk("%d free buffer pages\n", nr_free_buffer_pages());
 204         }
 205 #else /* !CONFIG_DISCONTIGMEM */
 206         printk("Free swap:       %6dkB\n", nr_swap_pages<<(PAGE_SHIFT-10));
 207         i = max_mapnr;
 208         while (i-- > 0) {
 209                 total++;
 210                 if (PageReserved(mem_map+i))
 211                         reserved++;
 212                 else if (PageSwapCache(mem_map+i))
 213                         cached++;
 214                 else if (page_count(mem_map + i))
 215                         shared += page_count(mem_map + i) - 1;
 216         }
 217         printk("%d pages of RAM\n", total);
 218         printk("%d reserved pages\n", reserved);
 219         printk("%d pages shared\n", shared);
 220         printk("%d pages swap cached\n", cached);
 221         printk("%ld pages in page table cache\n", pgtable_cache_size);
 222 #endif /* !CONFIG_DISCONTIGMEM */
 223 }
 224
 225 /*
 226  * This is like put_dirty_page() but installs a clean page with PAGE_GATE protection
 227  * (execute-only, typically).
 228  */
 229 struct page *
 230 put_gate_page (struct page *page, unsigned long address)
 231 {
 232         pgd_t *pgd;
 233         pmd_t *pmd;
 234         pte_t *pte;
 235
 236         if (!PageReserved(page))
 237                 printk(KERN_ERR "put_gate_page: gate page at 0x%p not in reserved memory\n",
 238                        page_address(page));
 239
 240         pgd = pgd_offset_k(address);            /* note: this is NOT pgd_offset()! */
 241
 242         spin_lock(&init_mm.page_table_lock);
 243         {
 244                 pmd = pmd_alloc(&init_mm, pgd, address);
 245                 if (!pmd)
 246                         goto out;
 247                 pte = pte_alloc_map(&init_mm, pmd, address);
 248                 if (!pte)
 249                         goto out;
 250                 if (!pte_none(*pte)) {
 251                         pte_unmap(pte);
 252                         goto out;
 253                 }
 254                 set_pte(pte, mk_pte(page, PAGE_GATE));
 255                 pte_unmap(pte);
 256         }
 257   out:  spin_unlock(&init_mm.page_table_lock);
 258         /* no need for flush_tlb */
 259         return page;
 260 }
 261
 262 void __init
 263 ia64_mmu_init (void *my_cpu_data)
 264 {
 265         unsigned long psr, rid, pta, impl_va_bits;
 266         extern void __init tlb_init (void);
 267 #ifdef CONFIG_DISABLE_VHPT
 268 #       define VHPT_ENABLE_BIT  0
 269 #else
 270 #       define VHPT_ENABLE_BIT  1
 271 #endif
 272
 273         /*
 274          * Set up the kernel identity mapping for regions 6 and 5.  The mapping for region
 275          * 7 is setup up in _start().
 276          */
 277         psr = ia64_clear_ic();
 278
 279         rid = ia64_rid(IA64_REGION_ID_KERNEL, __IA64_UNCACHED_OFFSET);
 280         ia64_set_rr(__IA64_UNCACHED_OFFSET, (rid << 8) | (IA64_GRANULE_SHIFT << 2));
 281
 282         rid = ia64_rid(IA64_REGION_ID_KERNEL, VMALLOC_START);
 283         ia64_set_rr(VMALLOC_START, (rid << 8) | (PAGE_SHIFT << 2) | 1);
 284
 285         /* ensure rr6 is up-to-date before inserting the PERCPU_ADDR translation: */
 286         ia64_srlz_d();
 287
 288         ia64_itr(0x2, IA64_TR_PERCPU_DATA, PERCPU_ADDR,
 289                  pte_val(pfn_pte(__pa(my_cpu_data) >> PAGE_SHIFT, PAGE_KERNEL)),
 290                  PERCPU_PAGE_SHIFT);
 291
 292         ia64_set_psr(psr);
 293         ia64_srlz_i();
 294
 295         /*
 296          * Check if the virtually mapped linear page table (VMLPT) overlaps with a mapped
 297          * address space.  The IA-64 architecture guarantees that at least 50 bits of
 298          * virtual address space are implemented but if we pick a large enough page size
 299          * (e.g., 64KB), the mapped address space is big enough that it will overlap with
 300          * VMLPT.  I assume that once we run on machines big enough to warrant 64KB pages,
 301          * IMPL_VA_MSB will be significantly bigger, so this is unlikely to become a
 302          * problem in practice.  Alternatively, we could truncate the top of the mapped
 303          * address space to not permit mappings that would overlap with the VMLPT.
 304          * --davidm 00/12/06
 305          */
 306 #       define pte_bits                 3
 307 #       define mapped_space_bits        (3*(PAGE_SHIFT - pte_bits) + PAGE_SHIFT)
 308         /*
 309          * The virtual page table has to cover the entire implemented address space within
 310          * a region even though not all of this space may be mappable.  The reason for
 311          * this is that the Access bit and Dirty bit fault handlers perform
 312          * non-speculative accesses to the virtual page table, so the address range of the
 313          * virtual page table itself needs to be covered by virtual page table.
 314          */
 315 #       define vmlpt_bits               (impl_va_bits - PAGE_SHIFT + pte_bits)
 316 #       define POW2(n)                  (1ULL << (n))
 317
 318         impl_va_bits = ffz(~(local_cpu_data->unimpl_va_mask | (7UL << 61)));
 319
 320         if (impl_va_bits < 51 || impl_va_bits > 61)
 321                 panic("CPU has bogus IMPL_VA_MSB value of %lu!\n", impl_va_bits - 1);
 322
 323         /* place the VMLPT at the end of each page-table mapped region: */
 324         pta = POW2(61) - POW2(vmlpt_bits);
 325
 326         if (POW2(mapped_space_bits) >= pta)
 327                 panic("mm/init: overlap between virtually mapped linear page table and "
 328                       "mapped kernel space!");
 329         /*
 330          * Set the (virtually mapped linear) page table address.  Bit
 331          * 8 selects between the short and long format, bits 2-7 the
 332          * size of the table, and bit 0 whether the VHPT walker is
 333          * enabled.
 334          */
 335         ia64_set_pta(pta | (0 << 8) | (vmlpt_bits << 2) | VHPT_ENABLE_BIT);
 336
 337         ia64_tlb_init();
 338 }
 339
 340 /*
 341  * Set up the page tables.
 342  */
 343
 344 #ifdef CONFIG_DISCONTIGMEM
 345 void
 346 paging_init (void)
 347 {
 348         extern void discontig_paging_init(void);
 349
 350         discontig_paging_init();
 351 }
 352 #else /* !CONFIG_DISCONTIGMEM */
 353 void
 354 paging_init (void)
 355 {
 356         unsigned long max_dma, zones_size[MAX_NR_ZONES];
 357
 358         /* initialize mem_map[] */
 359
 360         memset(zones_size, 0, sizeof(zones_size));
 361
 362         max_dma = virt_to_phys((void *) MAX_DMA_ADDRESS) >> PAGE_SHIFT;
 363         if (max_low_pfn < max_dma)
 364                 zones_size[ZONE_DMA] = max_low_pfn;
 365         else {
 366                 zones_size[ZONE_DMA] = max_dma;
 367                 zones_size[ZONE_NORMAL] = max_low_pfn - max_dma;
 368         }
 369         free_area_init(zones_size);
 370 }
 371 #endif /* !CONFIG_DISCONTIGMEM */
 372
 373 static int
 374 count_pages (u64 start, u64 end, void *arg)
 375 {
 376         unsigned long *count = arg;
 377
 378         *count += (end - start) >> PAGE_SHIFT;
 379         return 0;
 380 }
 381
 382 static int
 383 count_reserved_pages (u64 start, u64 end, void *arg)
 384 {
 385         unsigned long num_reserved = 0;
 386         unsigned long *count = arg;
 387
 388         for (; start < end; start += PAGE_SIZE)
 389                 if (PageReserved(virt_to_page(start)))
 390                         ++num_reserved;
 391         *count += num_reserved;
 392         return 0;
 393 }
 394
 395 void
 396 mem_init (void)
 397 {
 398         extern char __start_gate_section[];
 399         long reserved_pages, codesize, datasize, initsize;
 400         unsigned long num_pgt_pages;
 401         pg_data_t *pgdat;
 402
 403 #ifdef CONFIG_PCI
 404         /*
 405          * This needs to be called _after_ the command line has been parsed but _before_
 406          * any drivers that may need the PCI DMA interface are initialized or bootmem has
 407          * been freed.
 408          */
 409         platform_pci_dma_init();
 410 #endif
 411
 412 #ifndef CONFIG_DISCONTIGMEM
 413         if (!mem_map)
 414                 BUG();
 415         max_mapnr = max_low_pfn;
 416 #endif
 417
 418         num_physpages = 0;
 419         efi_memmap_walk(count_pages, &num_physpages);
 420
 421         high_memory = __va(max_low_pfn * PAGE_SIZE);
 422
 423         for_each_pgdat(pgdat)
 424                 totalram_pages += free_all_bootmem_node(pgdat);
 425
 426         reserved_pages = 0;
 427         efi_memmap_walk(count_reserved_pages, &reserved_pages);
 428
 429         codesize =  (unsigned long) &_etext - (unsigned long) &_stext;
 430         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 431         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 432
 433         printk(KERN_INFO "Memory: %luk/%luk available (%luk code, %luk reserved, "
 434                "%luk data, %luk init)\n", (unsigned long) nr_free_pages() << (PAGE_SHIFT - 10),
 435                num_physpages << (PAGE_SHIFT - 10), codesize >> 10,
 436                reserved_pages << (PAGE_SHIFT - 10), datasize >> 10, initsize >> 10);
 437
 438         /*
 439          * Allow for enough (cached) page table pages so that we can map the entire memory
 440          * at least once.  Each task also needs a couple of page tables pages, so add in a
 441          * fudge factor for that (don't use "threads-max" here; that would be wrong!).
 442          * Don't allow the cache to be more than 10% of total memory, though.
 443          */
 444 #       define NUM_TASKS        500     /* typical number of tasks */
 445         num_pgt_pages = nr_free_pages() / PTRS_PER_PGD + NUM_TASKS;
 446         if (num_pgt_pages > nr_free_pages() / 10)
 447                 num_pgt_pages = nr_free_pages() / 10;
 448         if (num_pgt_pages > pgt_cache_water[1])
 449                 pgt_cache_water[1] = num_pgt_pages;
 450
 451         /* install the gate page in the global page table: */
 452         put_gate_page(virt_to_page(__start_gate_section), GATE_ADDR);
 453
 454 #ifdef CONFIG_IA32_SUPPORT
 455         ia32_gdt_init();
 456 #endif
 457 }