arch/x86/mm/hugetlbpage.c

   1 /*
   2  * IA-32 Huge TLB Page Support for Kernel.
   3  *
   4  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
   5  */
   6
   7 #include <linux/init.h>
   8 #include <linux/fs.h>
   9 #include <linux/mm.h>
  10 #include <linux/hugetlb.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/slab.h>
  13 #include <linux/err.h>
  14 #include <linux/sysctl.h>
  15 #include <asm/mman.h>
  16 #include <asm/tlb.h>
  17 #include <asm/tlbflush.h>
  18 #include <asm/pgalloc.h>
  19
  20 static unsigned long page_table_shareable(struct vm_area_struct *svma,
  21                                 struct vm_area_struct *vma,
  22                                 unsigned long addr, pgoff_t idx)
  23 {
  24         unsigned long saddr = ((idx - svma->vm_pgoff) << PAGE_SHIFT) +
  25                                 svma->vm_start;
  26         unsigned long sbase = saddr & PUD_MASK;
  27         unsigned long s_end = sbase + PUD_SIZE;
  28
  29         /* Allow segments to share if only one is marked locked */
  30         unsigned long vm_flags = vma->vm_flags & ~VM_LOCKED;
  31         unsigned long svm_flags = svma->vm_flags & ~VM_LOCKED;
  32
  33         /*
  34          * match the virtual addresses, permission and the alignment of the
  35          * page table page.
  36          */
  37         if (pmd_index(addr) != pmd_index(saddr) ||
  38             vm_flags != svm_flags ||
  39             sbase < svma->vm_start || svma->vm_end < s_end)
  40                 return 0;
  41
  42         return saddr;
  43 }
  44
  45 static int vma_shareable(struct vm_area_struct *vma, unsigned long addr)
  46 {
  47         unsigned long base = addr & PUD_MASK;
  48         unsigned long end = base + PUD_SIZE;
  49
  50         /*
  51          * check on proper vm_flags and page table alignment
  52          */
  53         if (vma->vm_flags & VM_MAYSHARE &&
  54             vma->vm_start <= base && end <= vma->vm_end)
  55                 return 1;
  56         return 0;
  57 }
  58
  59 /*
  60  * search for a shareable pmd page for hugetlb.
  61  */
  62 static void huge_pmd_share(struct mm_struct *mm, unsigned long addr, pud_t *pud)
  63 {
  64         struct vm_area_struct *vma = find_vma(mm, addr);
  65         struct address_space *mapping = vma->vm_file->f_mapping;
  66         pgoff_t idx = ((addr - vma->vm_start) >> PAGE_SHIFT) +
  67                         vma->vm_pgoff;
  68         struct prio_tree_iter iter;
  69         struct vm_area_struct *svma;
  70         unsigned long saddr;
  71         pte_t *spte = NULL;
  72
  73         if (!vma_shareable(vma, addr))
  74                 return;
  75
  76         spin_lock(&mapping->i_mmap_lock);
  77         vma_prio_tree_foreach(svma, &iter, &mapping->i_mmap, idx, idx) {
  78                 if (svma == vma)
  79                         continue;
  80
  81                 saddr = page_table_shareable(svma, vma, addr, idx);
  82                 if (saddr) {
  83                         spte = huge_pte_offset(svma->vm_mm, saddr);
  84                         if (spte) {
  85                                 get_page(virt_to_page(spte));
  86                                 break;
  87                         }
  88                 }
  89         }
  90
  91         if (!spte)
  92                 goto out;
  93
  94         spin_lock(&mm->page_table_lock);
  95         if (pud_none(*pud))
  96                 pud_populate(mm, pud, (pmd_t *)((unsigned long)spte & PAGE_MASK));
  97         else
  98                 put_page(virt_to_page(spte));
  99         spin_unlock(&mm->page_table_lock);
 100 out:
 101         spin_unlock(&mapping->i_mmap_lock);
 102 }
 103
 104 /*
 105  * unmap huge page backed by shared pte.
 106  *
 107  * Hugetlb pte page is ref counted at the time of mapping.  If pte is shared
 108  * indicated by page_count > 1, unmap is achieved by clearing pud and
 109  * decrementing the ref count. If count == 1, the pte page is not shared.
 110  *
 111  * called with vma->vm_mm->page_table_lock held.
 112  *
 113  * returns: 1 successfully unmapped a shared pte page
 114  *          0 the underlying pte page is not shared, or it is the last user
 115  */
 116 int huge_pmd_unshare(struct mm_struct *mm, unsigned long *addr, pte_t *ptep)
 117 {
 118         pgd_t *pgd = pgd_offset(mm, *addr);
 119         pud_t *pud = pud_offset(pgd, *addr);
 120
 121         BUG_ON(page_count(virt_to_page(ptep)) == 0);
 122         if (page_count(virt_to_page(ptep)) == 1)
 123                 return 0;
 124
 125         pud_clear(pud);
 126         put_page(virt_to_page(ptep));
 127         *addr = ALIGN(*addr, HPAGE_SIZE * PTRS_PER_PTE) - HPAGE_SIZE;
 128         return 1;
 129 }
 130
 131 pte_t *huge_pte_alloc(struct mm_struct *mm,
 132                         unsigned long addr, unsigned long sz)
 133 {
 134         pgd_t *pgd;
 135         pud_t *pud;
 136         pte_t *pte = NULL;
 137
 138         pgd = pgd_offset(mm, addr);
 139         pud = pud_alloc(mm, pgd, addr);
 140         if (pud) {
 141                 if (sz == PUD_SIZE) {
 142                         pte = (pte_t *)pud;
 143                 } else {
 144                         BUG_ON(sz != PMD_SIZE);
 145                         if (pud_none(*pud))
 146                                 huge_pmd_share(mm, addr, pud);
 147                         pte = (pte_t *) pmd_alloc(mm, pud, addr);
 148                 }
 149         }
 150         BUG_ON(pte && !pte_none(*pte) && !pte_huge(*pte));
 151
 152         return pte;
 153 }
 154
 155 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
 156 {
 157         pgd_t *pgd;
 158         pud_t *pud;
 159         pmd_t *pmd = NULL;
 160
 161         pgd = pgd_offset(mm, addr);
 162         if (pgd_present(*pgd)) {
 163                 pud = pud_offset(pgd, addr);
 164                 if (pud_present(*pud)) {
 165                         if (pud_large(*pud))
 166                                 return (pte_t *)pud;
 167                         pmd = pmd_offset(pud, addr);
 168                 }
 169         }
 170         return (pte_t *) pmd;
 171 }
 172
 173 #if 0   /* This is just for testing */
 174 struct page *
 175 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 176 {
 177         unsigned long start = address;
 178         int length = 1;
 179         int nr;
 180         struct page *page;
 181         struct vm_area_struct *vma;
 182
 183         vma = find_vma(mm, addr);
 184         if (!vma || !is_vm_hugetlb_page(vma))
 185                 return ERR_PTR(-EINVAL);
 186
 187         pte = huge_pte_offset(mm, address);
 188
 189         /* hugetlb should be locked, and hence, prefaulted */
 190         WARN_ON(!pte || pte_none(*pte));
 191
 192         page = &pte_page(*pte)[vpfn % (HPAGE_SIZE/PAGE_SIZE)];
 193
 194         WARN_ON(!PageHead(page));
 195
 196         return page;
 197 }
 198
 199 int pmd_huge(pmd_t pmd)
 200 {
 201         return 0;
 202 }
 203
 204 int pud_huge(pud_t pud)
 205 {
 206         return 0;
 207 }
 208
 209 struct page *
 210 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 211                 pmd_t *pmd, int write)
 212 {
 213         return NULL;
 214 }
 215
 216 #else
 217
 218 struct page *
 219 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
 220 {
 221         return ERR_PTR(-EINVAL);
 222 }
 223
 224 int pmd_huge(pmd_t pmd)
 225 {
 226         return !!(pmd_val(pmd) & _PAGE_PSE);
 227 }
 228
 229 int pud_huge(pud_t pud)
 230 {
 231         return !!(pud_val(pud) & _PAGE_PSE);
 232 }
 233
 234 struct page *
 235 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
 236                 pmd_t *pmd, int write)
 237 {
 238         struct page *page;
 239
 240         page = pte_page(*(pte_t *)pmd);
 241         if (page)
 242                 page += ((address & ~PMD_MASK) >> PAGE_SHIFT);
 243         return page;
 244 }
 245
 246 struct page *
 247 follow_huge_pud(struct mm_struct *mm, unsigned long address,
 248                 pud_t *pud, int write)
 249 {
 250         struct page *page;
 251
 252         page = pte_page(*(pte_t *)pud);
 253         if (page)
 254                 page += ((address & ~PUD_MASK) >> PAGE_SHIFT);
 255         return page;
 256 }
 257
 258 #endif
 259
 260 /* x86_64 also uses this file */
 261
 262 #ifdef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 263 static unsigned long hugetlb_get_unmapped_area_bottomup(struct file *file,
 264                 unsigned long addr, unsigned long len,
 265                 unsigned long pgoff, unsigned long flags)
 266 {
 267         struct hstate *h = hstate_file(file);
 268         struct mm_struct *mm = current->mm;
 269         struct vm_area_struct *vma;
 270         unsigned long start_addr;
 271
 272         if (len > mm->cached_hole_size) {
 273                 start_addr = mm->free_area_cache;
 274         } else {
 275                 start_addr = TASK_UNMAPPED_BASE;
 276                 mm->cached_hole_size = 0;
 277         }
 278
 279 full_search:
 280         addr = ALIGN(start_addr, huge_page_size(h));
 281
 282         for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
 283                 /* At this point:  (!vma || addr < vma->vm_end). */
 284                 if (TASK_SIZE - len < addr) {
 285                         /*
 286                          * Start a new search - just in case we missed
 287                          * some holes.
 288                          */
 289                         if (start_addr != TASK_UNMAPPED_BASE) {
 290                                 start_addr = TASK_UNMAPPED_BASE;
 291                                 mm->cached_hole_size = 0;
 292                                 goto full_search;
 293                         }
 294                         return -ENOMEM;
 295                 }
 296                 if (!vma || addr + len <= vma->vm_start) {
 297                         mm->free_area_cache = addr + len;
 298                         return addr;
 299                 }
 300                 if (addr + mm->cached_hole_size < vma->vm_start)
 301                         mm->cached_hole_size = vma->vm_start - addr;
 302                 addr = ALIGN(vma->vm_end, huge_page_size(h));
 303         }
 304 }
 305
 306 static unsigned long hugetlb_get_unmapped_area_topdown(struct file *file,
 307                 unsigned long addr0, unsigned long len,
 308                 unsigned long pgoff, unsigned long flags)
 309 {
 310         struct hstate *h = hstate_file(file);
 311         struct mm_struct *mm = current->mm;
 312         struct vm_area_struct *vma, *prev_vma;
 313         unsigned long base = mm->mmap_base, addr = addr0;
 314         unsigned long largest_hole = mm->cached_hole_size;
 315         int first_time = 1;
 316
 317         /* don't allow allocations above current base */
 318         if (mm->free_area_cache > base)
 319                 mm->free_area_cache = base;
 320
 321         if (len <= largest_hole) {
 322                 largest_hole = 0;
 323                 mm->free_area_cache  = base;
 324         }
 325 try_again:
 326         /* make sure it can fit in the remaining address space */
 327         if (mm->free_area_cache < len)
 328                 goto fail;
 329
 330         /* either no address requested or cant fit in requested address hole */
 331         addr = (mm->free_area_cache - len) & huge_page_mask(h);
 332         do {
 333                 /*
 334                  * Lookup failure means no vma is above this address,
 335                  * i.e. return with success:
 336                  */
 337                 if (!(vma = find_vma_prev(mm, addr, &prev_vma)))
 338                         return addr;
 339
 340                 /*
 341                  * new region fits between prev_vma->vm_end and
 342                  * vma->vm_start, use it:
 343                  */
 344                 if (addr + len <= vma->vm_start &&
 345                             (!prev_vma || (addr >= prev_vma->vm_end))) {
 346                         /* remember the address as a hint for next time */
 347                         mm->cached_hole_size = largest_hole;
 348                         return (mm->free_area_cache = addr);
 349                 } else {
 350                         /* pull free_area_cache down to the first hole */
 351                         if (mm->free_area_cache == vma->vm_end) {
 352                                 mm->free_area_cache = vma->vm_start;
 353                                 mm->cached_hole_size = largest_hole;
 354                         }
 355                 }
 356
 357                 /* remember the largest hole we saw so far */
 358                 if (addr + largest_hole < vma->vm_start)
 359                         largest_hole = vma->vm_start - addr;
 360
 361                 /* try just below the current vma->vm_start */
 362                 addr = (vma->vm_start - len) & huge_page_mask(h);
 363         } while (len <= vma->vm_start);
 364
 365 fail:
 366         /*
 367          * if hint left us with no space for the requested
 368          * mapping then try again:
 369          */
 370         if (first_time) {
 371                 mm->free_area_cache = base;
 372                 largest_hole = 0;
 373                 first_time = 0;
 374                 goto try_again;
 375         }
 376         /*
 377          * A failed mmap() very likely causes application failure,
 378          * so fall back to the bottom-up function here. This scenario
 379          * can happen with large stack limits and large mmap()
 380          * allocations.
 381          */
 382         mm->free_area_cache = TASK_UNMAPPED_BASE;
 383         mm->cached_hole_size = ~0UL;
 384         addr = hugetlb_get_unmapped_area_bottomup(file, addr0,
 385                         len, pgoff, flags);
 386
 387         /*
 388          * Restore the topdown base:
 389          */
 390         mm->free_area_cache = base;
 391         mm->cached_hole_size = ~0UL;
 392
 393         return addr;
 394 }
 395
 396 unsigned long
 397 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 398                 unsigned long len, unsigned long pgoff, unsigned long flags)
 399 {
 400         struct hstate *h = hstate_file(file);
 401         struct mm_struct *mm = current->mm;
 402         struct vm_area_struct *vma;
 403
 404         if (len & ~huge_page_mask(h))
 405                 return -EINVAL;
 406         if (len > TASK_SIZE)
 407                 return -ENOMEM;
 408
 409         if (flags & MAP_FIXED) {
 410                 if (prepare_hugepage_range(file, addr, len))
 411                         return -EINVAL;
 412                 return addr;
 413         }
 414
 415         if (addr) {
 416                 addr = ALIGN(addr, huge_page_size(h));
 417                 vma = find_vma(mm, addr);
 418                 if (TASK_SIZE - len >= addr &&
 419                     (!vma || addr + len <= vma->vm_start))
 420                         return addr;
 421         }
 422         if (mm->get_unmapped_area == arch_get_unmapped_area)
 423                 return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 424                                 pgoff, flags);
 425         else
 426                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 427                                 pgoff, flags);
 428 }
 429
 430 #endif /*HAVE_ARCH_HUGETLB_UNMAPPED_AREA*/
 431
 432 #ifdef CONFIG_X86_64
 433 static __init int setup_hugepagesz(char *opt)
 434 {
 435         unsigned long ps = memparse(opt, &opt);
 436         if (ps == PMD_SIZE) {
 437                 hugetlb_add_hstate(PMD_SHIFT - PAGE_SHIFT);
 438         } else if (ps == PUD_SIZE && cpu_has_gbpages) {
 439                 hugetlb_add_hstate(PUD_SHIFT - PAGE_SHIFT);
 440         } else {
 441                 printk(KERN_ERR "hugepagesz: Unsupported page size %lu M\n",
 442                         ps >> 20);
 443                 return 0;
 444         }
 445         return 1;
 446 }
 447 __setup("hugepagesz=", setup_hugepagesz);
 448 #endif