fs/hugetlbfs/inode.c

   1 /*
   2  * hugetlbpage-backed filesystem.  Based on ramfs.
   3  *
   4  * Nadia Yvette Chambers, 2002
   5  *
   6  * Copyright (C) 2002 Linus Torvalds.
   7  */
   8
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/module.h>
  12 #include <linux/thread_info.h>
  13 #include <asm/current.h>
  14 #include <linux/sched.h>                /* remove ASAP */
  15 #include <linux/falloc.h>
  16 #include <linux/fs.h>
  17 #include <linux/mount.h>
  18 #include <linux/file.h>
  19 #include <linux/kernel.h>
  20 #include <linux/writeback.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/highmem.h>
  23 #include <linux/init.h>
  24 #include <linux/string.h>
  25 #include <linux/capability.h>
  26 #include <linux/ctype.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/hugetlb.h>
  29 #include <linux/pagevec.h>
  30 #include <linux/parser.h>
  31 #include <linux/mman.h>
  32 #include <linux/slab.h>
  33 #include <linux/dnotify.h>
  34 #include <linux/statfs.h>
  35 #include <linux/security.h>
  36 #include <linux/magic.h>
  37 #include <linux/migrate.h>
  38 #include <linux/uio.h>
  39
  40 #include <asm/uaccess.h>
  41
  42 static const struct super_operations hugetlbfs_ops;
  43 static const struct address_space_operations hugetlbfs_aops;
  44 const struct file_operations hugetlbfs_file_operations;
  45 static const struct inode_operations hugetlbfs_dir_inode_operations;
  46 static const struct inode_operations hugetlbfs_inode_operations;
  47
  48 struct hugetlbfs_config {
  49         kuid_t   uid;
  50         kgid_t   gid;
  51         umode_t mode;
  52         long    max_hpages;
  53         long    nr_inodes;
  54         struct hstate *hstate;
  55         long    min_hpages;
  56 };
  57
  58 struct hugetlbfs_inode_info {
  59         struct shared_policy policy;
  60         struct inode vfs_inode;
  61 };
  62
  63 static inline struct hugetlbfs_inode_info *HUGETLBFS_I(struct inode *inode)
  64 {
  65         return container_of(inode, struct hugetlbfs_inode_info, vfs_inode);
  66 }
  67
  68 int sysctl_hugetlb_shm_group;
  69
  70 enum {
  71         Opt_size, Opt_nr_inodes,
  72         Opt_mode, Opt_uid, Opt_gid,
  73         Opt_pagesize, Opt_min_size,
  74         Opt_err,
  75 };
  76
  77 static const match_table_t tokens = {
  78         {Opt_size,      "size=%s"},
  79         {Opt_nr_inodes, "nr_inodes=%s"},
  80         {Opt_mode,      "mode=%o"},
  81         {Opt_uid,       "uid=%u"},
  82         {Opt_gid,       "gid=%u"},
  83         {Opt_pagesize,  "pagesize=%s"},
  84         {Opt_min_size,  "min_size=%s"},
  85         {Opt_err,       NULL},
  86 };
  87
  88 #ifdef CONFIG_NUMA
  89 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  90                                         struct inode *inode, pgoff_t index)
  91 {
  92         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
  93                                                         index);
  94 }
  95
  96 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  97 {
  98         mpol_cond_put(vma->vm_policy);
  99 }
 100 #else
 101 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
 102                                         struct inode *inode, pgoff_t index)
 103 {
 104 }
 105
 106 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
 107 {
 108 }
 109 #endif
 110
 111 static void huge_pagevec_release(struct pagevec *pvec)
 112 {
 113         int i;
 114
 115         for (i = 0; i < pagevec_count(pvec); ++i)
 116                 put_page(pvec->pages[i]);
 117
 118         pagevec_reinit(pvec);
 119 }
 120
 121 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 122 {
 123         struct inode *inode = file_inode(file);
 124         loff_t len, vma_len;
 125         int ret;
 126         struct hstate *h = hstate_file(file);
 127
 128         /*
 129          * vma address alignment (but not the pgoff alignment) has
 130          * already been checked by prepare_hugepage_range.  If you add
 131          * any error returns here, do so after setting VM_HUGETLB, so
 132          * is_vm_hugetlb_page tests below unmap_region go the right
 133          * way when do_mmap_pgoff unwinds (may be important on powerpc
 134          * and ia64).
 135          */
 136         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 137         vma->vm_ops = &hugetlb_vm_ops;
 138
 139         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 140                 return -EINVAL;
 141
 142         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 143
 144         mutex_lock(&inode->i_mutex);
 145         file_accessed(file);
 146
 147         ret = -ENOMEM;
 148         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 149
 150         if (hugetlb_reserve_pages(inode,
 151                                 vma->vm_pgoff >> huge_page_order(h),
 152                                 len >> huge_page_shift(h), vma,
 153                                 vma->vm_flags))
 154                 goto out;
 155
 156         ret = 0;
 157         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 158                 inode->i_size = len;
 159 out:
 160         mutex_unlock(&inode->i_mutex);
 161
 162         return ret;
 163 }
 164
 165 /*
 166  * Called under down_write(mmap_sem).
 167  */
 168
 169 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 170 static unsigned long
 171 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 172                 unsigned long len, unsigned long pgoff, unsigned long flags)
 173 {
 174         struct mm_struct *mm = current->mm;
 175         struct vm_area_struct *vma;
 176         struct hstate *h = hstate_file(file);
 177         struct vm_unmapped_area_info info;
 178
 179         if (len & ~huge_page_mask(h))
 180                 return -EINVAL;
 181         if (len > TASK_SIZE)
 182                 return -ENOMEM;
 183
 184         if (flags & MAP_FIXED) {
 185                 if (prepare_hugepage_range(file, addr, len))
 186                         return -EINVAL;
 187                 return addr;
 188         }
 189
 190         if (addr) {
 191                 addr = ALIGN(addr, huge_page_size(h));
 192                 vma = find_vma(mm, addr);
 193                 if (TASK_SIZE - len >= addr &&
 194                     (!vma || addr + len <= vma->vm_start))
 195                         return addr;
 196         }
 197
 198         info.flags = 0;
 199         info.length = len;
 200         info.low_limit = TASK_UNMAPPED_BASE;
 201         info.high_limit = TASK_SIZE;
 202         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 203         info.align_offset = 0;
 204         return vm_unmapped_area(&info);
 205 }
 206 #endif
 207
 208 static size_t
 209 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 210                         struct iov_iter *to, unsigned long size)
 211 {
 212         size_t copied = 0;
 213         int i, chunksize;
 214
 215         /* Find which 4k chunk and offset with in that chunk */
 216         i = offset >> PAGE_CACHE_SHIFT;
 217         offset = offset & ~PAGE_CACHE_MASK;
 218
 219         while (size) {
 220                 size_t n;
 221                 chunksize = PAGE_CACHE_SIZE;
 222                 if (offset)
 223                         chunksize -= offset;
 224                 if (chunksize > size)
 225                         chunksize = size;
 226                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 227                 copied += n;
 228                 if (n != chunksize)
 229                         return copied;
 230                 offset = 0;
 231                 size -= chunksize;
 232                 i++;
 233         }
 234         return copied;
 235 }
 236
 237 /*
 238  * Support for read() - Find the page attached to f_mapping and copy out the
 239  * data. Its *very* similar to do_generic_mapping_read(), we can't use that
 240  * since it has PAGE_CACHE_SIZE assumptions.
 241  */
 242 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 243 {
 244         struct file *file = iocb->ki_filp;
 245         struct hstate *h = hstate_file(file);
 246         struct address_space *mapping = file->f_mapping;
 247         struct inode *inode = mapping->host;
 248         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 249         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 250         unsigned long end_index;
 251         loff_t isize;
 252         ssize_t retval = 0;
 253
 254         while (iov_iter_count(to)) {
 255                 struct page *page;
 256                 size_t nr, copied;
 257
 258                 /* nr is the maximum number of bytes to copy from this page */
 259                 nr = huge_page_size(h);
 260                 isize = i_size_read(inode);
 261                 if (!isize)
 262                         break;
 263                 end_index = (isize - 1) >> huge_page_shift(h);
 264                 if (index > end_index)
 265                         break;
 266                 if (index == end_index) {
 267                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 268                         if (nr <= offset)
 269                                 break;
 270                 }
 271                 nr = nr - offset;
 272
 273                 /* Find the page */
 274                 page = find_lock_page(mapping, index);
 275                 if (unlikely(page == NULL)) {
 276                         /*
 277                          * We have a HOLE, zero out the user-buffer for the
 278                          * length of the hole or request.
 279                          */
 280                         copied = iov_iter_zero(nr, to);
 281                 } else {
 282                         unlock_page(page);
 283
 284                         /*
 285                          * We have the page, copy it to user space buffer.
 286                          */
 287                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 288                         page_cache_release(page);
 289                 }
 290                 offset += copied;
 291                 retval += copied;
 292                 if (copied != nr && iov_iter_count(to)) {
 293                         if (!retval)
 294                                 retval = -EFAULT;
 295                         break;
 296                 }
 297                 index += offset >> huge_page_shift(h);
 298                 offset &= ~huge_page_mask(h);
 299         }
 300         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 301         return retval;
 302 }
 303
 304 static int hugetlbfs_write_begin(struct file *file,
 305                         struct address_space *mapping,
 306                         loff_t pos, unsigned len, unsigned flags,
 307                         struct page **pagep, void **fsdata)
 308 {
 309         return -EINVAL;
 310 }
 311
 312 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 313                         loff_t pos, unsigned len, unsigned copied,
 314                         struct page *page, void *fsdata)
 315 {
 316         BUG();
 317         return -EINVAL;
 318 }
 319
 320 static void remove_huge_page(struct page *page)
 321 {
 322         ClearPageDirty(page);
 323         ClearPageUptodate(page);
 324         delete_from_page_cache(page);
 325 }
 326
 327
 328 /*
 329  * remove_inode_hugepages handles two distinct cases: truncation and hole
 330  * punch.  There are subtle differences in operation for each case.
 331
 332  * truncation is indicated by end of range being LLONG_MAX
 333  *      In this case, we first scan the range and release found pages.
 334  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
 335  *      maps and global counts.  Page faults can not race with truncation
 336  *      in this routine.  hugetlb_no_page() prevents page faults in the
 337  *      truncated range.  It checks i_size before allocation, and again after
 338  *      with the page table lock for the page held.  The same lock must be
 339  *      acquired to unmap a page.
 340  * hole punch is indicated if end is not LLONG_MAX
 341  *      In the hole punch case we scan the range and release found pages.
 342  *      Only when releasing a page is the associated region/reserv map
 343  *      deleted.  The region/reserv map for ranges without associated
 344  *      pages are not modified.  Page faults can race with hole punch.
 345  *      This is indicated if we find a mapped page.
 346  * Note: If the passed end of range value is beyond the end of file, but
 347  * not LLONG_MAX this routine still performs a hole punch operation.
 348  */
 349 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 350                                    loff_t lend)
 351 {
 352         struct hstate *h = hstate_inode(inode);
 353         struct address_space *mapping = &inode->i_data;
 354         const pgoff_t start = lstart >> huge_page_shift(h);
 355         const pgoff_t end = lend >> huge_page_shift(h);
 356         struct vm_area_struct pseudo_vma;
 357         struct pagevec pvec;
 358         pgoff_t next;
 359         int i, freed = 0;
 360         long lookup_nr = PAGEVEC_SIZE;
 361         bool truncate_op = (lend == LLONG_MAX);
 362
 363         memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
 364         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 365         pagevec_init(&pvec, 0);
 366         next = start;
 367         while (next < end) {
 368                 /*
 369                  * Don't grab more pages than the number left in the range.
 370                  */
 371                 if (end - next < lookup_nr)
 372                         lookup_nr = end - next;
 373
 374                 /*
 375                  * When no more pages are found, we are done.
 376                  */
 377                 if (!pagevec_lookup(&pvec, mapping, next, lookup_nr))
 378                         break;
 379
 380                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 381                         struct page *page = pvec.pages[i];
 382                         u32 hash;
 383
 384                         /*
 385                          * The page (index) could be beyond end.  This is
 386                          * only possible in the punch hole case as end is
 387                          * max page offset in the truncate case.
 388                          */
 389                         next = page->index;
 390                         if (next >= end)
 391                                 break;
 392
 393                         hash = hugetlb_fault_mutex_hash(h, current->mm,
 394                                                         &pseudo_vma,
 395                                                         mapping, next, 0);
 396                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 397
 398                         lock_page(page);
 399                         if (likely(!page_mapped(page))) {
 400                                 bool rsv_on_error = !PagePrivate(page);
 401                                 /*
 402                                  * We must free the huge page and remove
 403                                  * from page cache (remove_huge_page) BEFORE
 404                                  * removing the region/reserve map
 405                                  * (hugetlb_unreserve_pages).  In rare out
 406                                  * of memory conditions, removal of the
 407                                  * region/reserve map could fail.  Before
 408                                  * free'ing the page, note PagePrivate which
 409                                  * is used in case of error.
 410                                  */
 411                                 remove_huge_page(page);
 412                                 freed++;
 413                                 if (!truncate_op) {
 414                                         if (unlikely(hugetlb_unreserve_pages(
 415                                                         inode, next,
 416                                                         next + 1, 1)))
 417                                                 hugetlb_fix_reserve_counts(
 418                                                         inode, rsv_on_error);
 419                                 }
 420                         } else {
 421                                 /*
 422                                  * If page is mapped, it was faulted in after
 423                                  * being unmapped.  It indicates a race between
 424                                  * hole punch and page fault.  Do nothing in
 425                                  * this case.  Getting here in a truncate
 426                                  * operation is a bug.
 427                                  */
 428                                 BUG_ON(truncate_op);
 429                         }
 430
 431                         unlock_page(page);
 432                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 433                 }
 434                 ++next;
 435                 huge_pagevec_release(&pvec);
 436                 cond_resched();
 437         }
 438
 439         if (truncate_op)
 440                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 441 }
 442
 443 static void hugetlbfs_evict_inode(struct inode *inode)
 444 {
 445         struct resv_map *resv_map;
 446
 447         remove_inode_hugepages(inode, 0, LLONG_MAX);
 448         resv_map = (struct resv_map *)inode->i_mapping->private_data;
 449         /* root inode doesn't have the resv_map, so we should check it */
 450         if (resv_map)
 451                 resv_map_release(&resv_map->refs);
 452         clear_inode(inode);
 453 }
 454
 455 static inline void
 456 hugetlb_vmdelete_list(struct rb_root *root, pgoff_t start, pgoff_t end)
 457 {
 458         struct vm_area_struct *vma;
 459
 460         /*
 461          * end == 0 indicates that the entire range after
 462          * start should be unmapped.
 463          */
 464         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
 465                 unsigned long v_offset;
 466                 unsigned long v_end;
 467
 468                 /*
 469                  * Can the expression below overflow on 32-bit arches?
 470                  * No, because the interval tree returns us only those vmas
 471                  * which overlap the truncated area starting at pgoff,
 472                  * and no vma on a 32-bit arch can span beyond the 4GB.
 473                  */
 474                 if (vma->vm_pgoff < start)
 475                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 476                 else
 477                         v_offset = 0;
 478
 479                 if (!end)
 480                         v_end = vma->vm_end;
 481                 else {
 482                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 483                                                         + vma->vm_start;
 484                         if (v_end > vma->vm_end)
 485                                 v_end = vma->vm_end;
 486                 }
 487
 488                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 489                                                                         NULL);
 490         }
 491 }
 492
 493 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 494 {
 495         pgoff_t pgoff;
 496         struct address_space *mapping = inode->i_mapping;
 497         struct hstate *h = hstate_inode(inode);
 498
 499         BUG_ON(offset & ~huge_page_mask(h));
 500         pgoff = offset >> PAGE_SHIFT;
 501
 502         i_size_write(inode, offset);
 503         i_mmap_lock_write(mapping);
 504         if (!RB_EMPTY_ROOT(&mapping->i_mmap))
 505                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 506         i_mmap_unlock_write(mapping);
 507         remove_inode_hugepages(inode, offset, LLONG_MAX);
 508         return 0;
 509 }
 510
 511 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 512 {
 513         struct hstate *h = hstate_inode(inode);
 514         loff_t hpage_size = huge_page_size(h);
 515         loff_t hole_start, hole_end;
 516
 517         /*
 518          * For hole punch round up the beginning offset of the hole and
 519          * round down the end.
 520          */
 521         hole_start = round_up(offset, hpage_size);
 522         hole_end = round_down(offset + len, hpage_size);
 523
 524         if (hole_end > hole_start) {
 525                 struct address_space *mapping = inode->i_mapping;
 526
 527                 mutex_lock(&inode->i_mutex);
 528                 i_mmap_lock_write(mapping);
 529                 if (!RB_EMPTY_ROOT(&mapping->i_mmap))
 530                         hugetlb_vmdelete_list(&mapping->i_mmap,
 531                                                 hole_start >> PAGE_SHIFT,
 532                                                 hole_end  >> PAGE_SHIFT);
 533                 i_mmap_unlock_write(mapping);
 534                 remove_inode_hugepages(inode, hole_start, hole_end);
 535                 mutex_unlock(&inode->i_mutex);
 536         }
 537
 538         return 0;
 539 }
 540
 541 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 542                                 loff_t len)
 543 {
 544         struct inode *inode = file_inode(file);
 545         struct address_space *mapping = inode->i_mapping;
 546         struct hstate *h = hstate_inode(inode);
 547         struct vm_area_struct pseudo_vma;
 548         struct mm_struct *mm = current->mm;
 549         loff_t hpage_size = huge_page_size(h);
 550         unsigned long hpage_shift = huge_page_shift(h);
 551         pgoff_t start, index, end;
 552         int error;
 553         u32 hash;
 554
 555         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 556                 return -EOPNOTSUPP;
 557
 558         if (mode & FALLOC_FL_PUNCH_HOLE)
 559                 return hugetlbfs_punch_hole(inode, offset, len);
 560
 561         /*
 562          * Default preallocate case.
 563          * For this range, start is rounded down and end is rounded up
 564          * as well as being converted to page offsets.
 565          */
 566         start = offset >> hpage_shift;
 567         end = (offset + len + hpage_size - 1) >> hpage_shift;
 568
 569         mutex_lock(&inode->i_mutex);
 570
 571         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 572         error = inode_newsize_ok(inode, offset + len);
 573         if (error)
 574                 goto out;
 575
 576         /*
 577          * Initialize a pseudo vma as this is required by the huge page
 578          * allocation routines.  If NUMA is configured, use page index
 579          * as input to create an allocation policy.
 580          */
 581         memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
 582         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 583         pseudo_vma.vm_file = file;
 584
 585         for (index = start; index < end; index++) {
 586                 /*
 587                  * This is supposed to be the vaddr where the page is being
 588                  * faulted in, but we have no vaddr here.
 589                  */
 590                 struct page *page;
 591                 unsigned long addr;
 592                 int avoid_reserve = 0;
 593
 594                 cond_resched();
 595
 596                 /*
 597                  * fallocate(2) manpage permits EINTR; we may have been
 598                  * interrupted because we are using up too much memory.
 599                  */
 600                 if (signal_pending(current)) {
 601                         error = -EINTR;
 602                         break;
 603                 }
 604
 605                 /* Set numa allocation policy based on index */
 606                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 607
 608                 /* addr is the offset within the file (zero based) */
 609                 addr = index * hpage_size;
 610
 611                 /* mutex taken here, fault path and hole punch */
 612                 hash = hugetlb_fault_mutex_hash(h, mm, &pseudo_vma, mapping,
 613                                                 index, addr);
 614                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 615
 616                 /* See if already present in mapping to avoid alloc/free */
 617                 page = find_get_page(mapping, index);
 618                 if (page) {
 619                         put_page(page);
 620                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 621                         hugetlb_drop_vma_policy(&pseudo_vma);
 622                         continue;
 623                 }
 624
 625                 /* Allocate page and add to page cache */
 626                 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
 627                 hugetlb_drop_vma_policy(&pseudo_vma);
 628                 if (IS_ERR(page)) {
 629                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 630                         error = PTR_ERR(page);
 631                         goto out;
 632                 }
 633                 clear_huge_page(page, addr, pages_per_huge_page(h));
 634                 __SetPageUptodate(page);
 635                 error = huge_add_to_page_cache(page, mapping, index);
 636                 if (unlikely(error)) {
 637                         put_page(page);
 638                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 639                         goto out;
 640                 }
 641
 642                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 643
 644                 /*
 645                  * page_put due to reference from alloc_huge_page()
 646                  * unlock_page because locked by add_to_page_cache()
 647                  */
 648                 put_page(page);
 649                 unlock_page(page);
 650         }
 651
 652         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 653                 i_size_write(inode, offset + len);
 654         inode->i_ctime = CURRENT_TIME;
 655 out:
 656         mutex_unlock(&inode->i_mutex);
 657         return error;
 658 }
 659
 660 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 661 {
 662         struct inode *inode = d_inode(dentry);
 663         struct hstate *h = hstate_inode(inode);
 664         int error;
 665         unsigned int ia_valid = attr->ia_valid;
 666
 667         BUG_ON(!inode);
 668
 669         error = inode_change_ok(inode, attr);
 670         if (error)
 671                 return error;
 672
 673         if (ia_valid & ATTR_SIZE) {
 674                 error = -EINVAL;
 675                 if (attr->ia_size & ~huge_page_mask(h))
 676                         return -EINVAL;
 677                 error = hugetlb_vmtruncate(inode, attr->ia_size);
 678                 if (error)
 679                         return error;
 680         }
 681
 682         setattr_copy(inode, attr);
 683         mark_inode_dirty(inode);
 684         return 0;
 685 }
 686
 687 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 688                                         struct hugetlbfs_config *config)
 689 {
 690         struct inode *inode;
 691
 692         inode = new_inode(sb);
 693         if (inode) {
 694                 struct hugetlbfs_inode_info *info;
 695                 inode->i_ino = get_next_ino();
 696                 inode->i_mode = S_IFDIR | config->mode;
 697                 inode->i_uid = config->uid;
 698                 inode->i_gid = config->gid;
 699                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 700                 info = HUGETLBFS_I(inode);
 701                 mpol_shared_policy_init(&info->policy, NULL);
 702                 inode->i_op = &hugetlbfs_dir_inode_operations;
 703                 inode->i_fop = &simple_dir_operations;
 704                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 705                 inc_nlink(inode);
 706                 lockdep_annotate_inode_mutex_key(inode);
 707         }
 708         return inode;
 709 }
 710
 711 /*
 712  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 713  * be taken from reclaim -- unlike regular filesystems. This needs an
 714  * annotation because huge_pmd_share() does an allocation under
 715  * i_mmap_rwsem.
 716  */
 717 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 718
 719 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 720                                         struct inode *dir,
 721                                         umode_t mode, dev_t dev)
 722 {
 723         struct inode *inode;
 724         struct resv_map *resv_map;
 725
 726         resv_map = resv_map_alloc();
 727         if (!resv_map)
 728                 return NULL;
 729
 730         inode = new_inode(sb);
 731         if (inode) {
 732                 struct hugetlbfs_inode_info *info;
 733                 inode->i_ino = get_next_ino();
 734                 inode_init_owner(inode, dir, mode);
 735                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 736                                 &hugetlbfs_i_mmap_rwsem_key);
 737                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 738                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 739                 inode->i_mapping->private_data = resv_map;
 740                 info = HUGETLBFS_I(inode);
 741                 /*
 742                  * The policy is initialized here even if we are creating a
 743                  * private inode because initialization simply creates an
 744                  * an empty rb tree and calls spin_lock_init(), later when we
 745                  * call mpol_free_shared_policy() it will just return because
 746                  * the rb tree will still be empty.
 747                  */
 748                 mpol_shared_policy_init(&info->policy, NULL);
 749                 switch (mode & S_IFMT) {
 750                 default:
 751                         init_special_inode(inode, mode, dev);
 752                         break;
 753                 case S_IFREG:
 754                         inode->i_op = &hugetlbfs_inode_operations;
 755                         inode->i_fop = &hugetlbfs_file_operations;
 756                         break;
 757                 case S_IFDIR:
 758                         inode->i_op = &hugetlbfs_dir_inode_operations;
 759                         inode->i_fop = &simple_dir_operations;
 760
 761                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 762                         inc_nlink(inode);
 763                         break;
 764                 case S_IFLNK:
 765                         inode->i_op = &page_symlink_inode_operations;
 766                         break;
 767                 }
 768                 lockdep_annotate_inode_mutex_key(inode);
 769         } else
 770                 kref_put(&resv_map->refs, resv_map_release);
 771
 772         return inode;
 773 }
 774
 775 /*
 776  * File creation. Allocate an inode, and we're done..
 777  */
 778 static int hugetlbfs_mknod(struct inode *dir,
 779                         struct dentry *dentry, umode_t mode, dev_t dev)
 780 {
 781         struct inode *inode;
 782         int error = -ENOSPC;
 783
 784         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 785         if (inode) {
 786                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
 787                 d_instantiate(dentry, inode);
 788                 dget(dentry);   /* Extra count - pin the dentry in core */
 789                 error = 0;
 790         }
 791         return error;
 792 }
 793
 794 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 795 {
 796         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
 797         if (!retval)
 798                 inc_nlink(dir);
 799         return retval;
 800 }
 801
 802 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
 803 {
 804         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
 805 }
 806
 807 static int hugetlbfs_symlink(struct inode *dir,
 808                         struct dentry *dentry, const char *symname)
 809 {
 810         struct inode *inode;
 811         int error = -ENOSPC;
 812
 813         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 814         if (inode) {
 815                 int l = strlen(symname)+1;
 816                 error = page_symlink(inode, symname, l);
 817                 if (!error) {
 818                         d_instantiate(dentry, inode);
 819                         dget(dentry);
 820                 } else
 821                         iput(inode);
 822         }
 823         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
 824
 825         return error;
 826 }
 827
 828 /*
 829  * mark the head page dirty
 830  */
 831 static int hugetlbfs_set_page_dirty(struct page *page)
 832 {
 833         struct page *head = compound_head(page);
 834
 835         SetPageDirty(head);
 836         return 0;
 837 }
 838
 839 static int hugetlbfs_migrate_page(struct address_space *mapping,
 840                                 struct page *newpage, struct page *page,
 841                                 enum migrate_mode mode)
 842 {
 843         int rc;
 844
 845         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 846         if (rc != MIGRATEPAGE_SUCCESS)
 847                 return rc;
 848         migrate_page_copy(newpage, page);
 849
 850         return MIGRATEPAGE_SUCCESS;
 851 }
 852
 853 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 854 {
 855         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
 856         struct hstate *h = hstate_inode(d_inode(dentry));
 857
 858         buf->f_type = HUGETLBFS_MAGIC;
 859         buf->f_bsize = huge_page_size(h);
 860         if (sbinfo) {
 861                 spin_lock(&sbinfo->stat_lock);
 862                 /* If no limits set, just report 0 for max/free/used
 863                  * blocks, like simple_statfs() */
 864                 if (sbinfo->spool) {
 865                         long free_pages;
 866
 867                         spin_lock(&sbinfo->spool->lock);
 868                         buf->f_blocks = sbinfo->spool->max_hpages;
 869                         free_pages = sbinfo->spool->max_hpages
 870                                 - sbinfo->spool->used_hpages;
 871                         buf->f_bavail = buf->f_bfree = free_pages;
 872                         spin_unlock(&sbinfo->spool->lock);
 873                         buf->f_files = sbinfo->max_inodes;
 874                         buf->f_ffree = sbinfo->free_inodes;
 875                 }
 876                 spin_unlock(&sbinfo->stat_lock);
 877         }
 878         buf->f_namelen = NAME_MAX;
 879         return 0;
 880 }
 881
 882 static void hugetlbfs_put_super(struct super_block *sb)
 883 {
 884         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
 885
 886         if (sbi) {
 887                 sb->s_fs_info = NULL;
 888
 889                 if (sbi->spool)
 890                         hugepage_put_subpool(sbi->spool);
 891
 892                 kfree(sbi);
 893         }
 894 }
 895
 896 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 897 {
 898         if (sbinfo->free_inodes >= 0) {
 899                 spin_lock(&sbinfo->stat_lock);
 900                 if (unlikely(!sbinfo->free_inodes)) {
 901                         spin_unlock(&sbinfo->stat_lock);
 902                         return 0;
 903                 }
 904                 sbinfo->free_inodes--;
 905                 spin_unlock(&sbinfo->stat_lock);
 906         }
 907
 908         return 1;
 909 }
 910
 911 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 912 {
 913         if (sbinfo->free_inodes >= 0) {
 914                 spin_lock(&sbinfo->stat_lock);
 915                 sbinfo->free_inodes++;
 916                 spin_unlock(&sbinfo->stat_lock);
 917         }
 918 }
 919
 920
 921 static struct kmem_cache *hugetlbfs_inode_cachep;
 922
 923 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
 924 {
 925         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
 926         struct hugetlbfs_inode_info *p;
 927
 928         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
 929                 return NULL;
 930         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
 931         if (unlikely(!p)) {
 932                 hugetlbfs_inc_free_inodes(sbinfo);
 933                 return NULL;
 934         }
 935         return &p->vfs_inode;
 936 }
 937
 938 static void hugetlbfs_i_callback(struct rcu_head *head)
 939 {
 940         struct inode *inode = container_of(head, struct inode, i_rcu);
 941         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
 942 }
 943
 944 static void hugetlbfs_destroy_inode(struct inode *inode)
 945 {
 946         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
 947         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
 948         call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
 949 }
 950
 951 static const struct address_space_operations hugetlbfs_aops = {
 952         .write_begin    = hugetlbfs_write_begin,
 953         .write_end      = hugetlbfs_write_end,
 954         .set_page_dirty = hugetlbfs_set_page_dirty,
 955         .migratepage    = hugetlbfs_migrate_page,
 956 };
 957
 958
 959 static void init_once(void *foo)
 960 {
 961         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
 962
 963         inode_init_once(&ei->vfs_inode);
 964 }
 965
 966 const struct file_operations hugetlbfs_file_operations = {
 967         .read_iter              = hugetlbfs_read_iter,
 968         .mmap                   = hugetlbfs_file_mmap,
 969         .fsync                  = noop_fsync,
 970         .get_unmapped_area      = hugetlb_get_unmapped_area,
 971         .llseek                 = default_llseek,
 972         .fallocate              = hugetlbfs_fallocate,
 973 };
 974
 975 static const struct inode_operations hugetlbfs_dir_inode_operations = {
 976         .create         = hugetlbfs_create,
 977         .lookup         = simple_lookup,
 978         .link           = simple_link,
 979         .unlink         = simple_unlink,
 980         .symlink        = hugetlbfs_symlink,
 981         .mkdir          = hugetlbfs_mkdir,
 982         .rmdir          = simple_rmdir,
 983         .mknod          = hugetlbfs_mknod,
 984         .rename         = simple_rename,
 985         .setattr        = hugetlbfs_setattr,
 986 };
 987
 988 static const struct inode_operations hugetlbfs_inode_operations = {
 989         .setattr        = hugetlbfs_setattr,
 990 };
 991
 992 static const struct super_operations hugetlbfs_ops = {
 993         .alloc_inode    = hugetlbfs_alloc_inode,
 994         .destroy_inode  = hugetlbfs_destroy_inode,
 995         .evict_inode    = hugetlbfs_evict_inode,
 996         .statfs         = hugetlbfs_statfs,
 997         .put_super      = hugetlbfs_put_super,
 998         .show_options   = generic_show_options,
 999 };
1000
1001 enum { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1002
1003 /*
1004  * Convert size option passed from command line to number of huge pages
1005  * in the pool specified by hstate.  Size option could be in bytes
1006  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1007  */
1008 static long long
1009 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1010                                                                 int val_type)
1011 {
1012         if (val_type == NO_SIZE)
1013                 return -1;
1014
1015         if (val_type == SIZE_PERCENT) {
1016                 size_opt <<= huge_page_shift(h);
1017                 size_opt *= h->max_huge_pages;
1018                 do_div(size_opt, 100);
1019         }
1020
1021         size_opt >>= huge_page_shift(h);
1022         return size_opt;
1023 }
1024
1025 static int
1026 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1027 {
1028         char *p, *rest;
1029         substring_t args[MAX_OPT_ARGS];
1030         int option;
1031         unsigned long long max_size_opt = 0, min_size_opt = 0;
1032         int max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1033
1034         if (!options)
1035                 return 0;
1036
1037         while ((p = strsep(&options, ",")) != NULL) {
1038                 int token;
1039                 if (!*p)
1040                         continue;
1041
1042                 token = match_token(p, tokens, args);
1043                 switch (token) {
1044                 case Opt_uid:
1045                         if (match_int(&args[0], &option))
1046                                 goto bad_val;
1047                         pconfig->uid = make_kuid(current_user_ns(), option);
1048                         if (!uid_valid(pconfig->uid))
1049                                 goto bad_val;
1050                         break;
1051
1052                 case Opt_gid:
1053                         if (match_int(&args[0], &option))
1054                                 goto bad_val;
1055                         pconfig->gid = make_kgid(current_user_ns(), option);
1056                         if (!gid_valid(pconfig->gid))
1057                                 goto bad_val;
1058                         break;
1059
1060                 case Opt_mode:
1061                         if (match_octal(&args[0], &option))
1062                                 goto bad_val;
1063                         pconfig->mode = option & 01777U;
1064                         break;
1065
1066                 case Opt_size: {
1067                         /* memparse() will accept a K/M/G without a digit */
1068                         if (!isdigit(*args[0].from))
1069                                 goto bad_val;
1070                         max_size_opt = memparse(args[0].from, &rest);
1071                         max_val_type = SIZE_STD;
1072                         if (*rest == '%')
1073                                 max_val_type = SIZE_PERCENT;
1074                         break;
1075                 }
1076
1077                 case Opt_nr_inodes:
1078                         /* memparse() will accept a K/M/G without a digit */
1079                         if (!isdigit(*args[0].from))
1080                                 goto bad_val;
1081                         pconfig->nr_inodes = memparse(args[0].from, &rest);
1082                         break;
1083
1084                 case Opt_pagesize: {
1085                         unsigned long ps;
1086                         ps = memparse(args[0].from, &rest);
1087                         pconfig->hstate = size_to_hstate(ps);
1088                         if (!pconfig->hstate) {
1089                                 pr_err("Unsupported page size %lu MB\n",
1090                                         ps >> 20);
1091                                 return -EINVAL;
1092                         }
1093                         break;
1094                 }
1095
1096                 case Opt_min_size: {
1097                         /* memparse() will accept a K/M/G without a digit */
1098                         if (!isdigit(*args[0].from))
1099                                 goto bad_val;
1100                         min_size_opt = memparse(args[0].from, &rest);
1101                         min_val_type = SIZE_STD;
1102                         if (*rest == '%')
1103                                 min_val_type = SIZE_PERCENT;
1104                         break;
1105                 }
1106
1107                 default:
1108                         pr_err("Bad mount option: \"%s\"\n", p);
1109                         return -EINVAL;
1110                         break;
1111                 }
1112         }
1113
1114         /*
1115          * Use huge page pool size (in hstate) to convert the size
1116          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1117          */
1118         pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1119                                                 max_size_opt, max_val_type);
1120         pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1121                                                 min_size_opt, min_val_type);
1122
1123         /*
1124          * If max_size was specified, then min_size must be smaller
1125          */
1126         if (max_val_type > NO_SIZE &&
1127             pconfig->min_hpages > pconfig->max_hpages) {
1128                 pr_err("minimum size can not be greater than maximum size\n");
1129                 return -EINVAL;
1130         }
1131
1132         return 0;
1133
1134 bad_val:
1135         pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1136         return -EINVAL;
1137 }
1138
1139 static int
1140 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1141 {
1142         int ret;
1143         struct hugetlbfs_config config;
1144         struct hugetlbfs_sb_info *sbinfo;
1145
1146         save_mount_options(sb, data);
1147
1148         config.max_hpages = -1; /* No limit on size by default */
1149         config.nr_inodes = -1; /* No limit on number of inodes by default */
1150         config.uid = current_fsuid();
1151         config.gid = current_fsgid();
1152         config.mode = 0755;
1153         config.hstate = &default_hstate;
1154         config.min_hpages = -1; /* No default minimum size */
1155         ret = hugetlbfs_parse_options(data, &config);
1156         if (ret)
1157                 return ret;
1158
1159         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1160         if (!sbinfo)
1161                 return -ENOMEM;
1162         sb->s_fs_info = sbinfo;
1163         sbinfo->hstate = config.hstate;
1164         spin_lock_init(&sbinfo->stat_lock);
1165         sbinfo->max_inodes = config.nr_inodes;
1166         sbinfo->free_inodes = config.nr_inodes;
1167         sbinfo->spool = NULL;
1168         /*
1169          * Allocate and initialize subpool if maximum or minimum size is
1170          * specified.  Any needed reservations (for minimim size) are taken
1171          * taken when the subpool is created.
1172          */
1173         if (config.max_hpages != -1 || config.min_hpages != -1) {
1174                 sbinfo->spool = hugepage_new_subpool(config.hstate,
1175                                                         config.max_hpages,
1176                                                         config.min_hpages);
1177                 if (!sbinfo->spool)
1178                         goto out_free;
1179         }
1180         sb->s_maxbytes = MAX_LFS_FILESIZE;
1181         sb->s_blocksize = huge_page_size(config.hstate);
1182         sb->s_blocksize_bits = huge_page_shift(config.hstate);
1183         sb->s_magic = HUGETLBFS_MAGIC;
1184         sb->s_op = &hugetlbfs_ops;
1185         sb->s_time_gran = 1;
1186         sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1187         if (!sb->s_root)
1188                 goto out_free;
1189         return 0;
1190 out_free:
1191         kfree(sbinfo->spool);
1192         kfree(sbinfo);
1193         return -ENOMEM;
1194 }
1195
1196 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1197         int flags, const char *dev_name, void *data)
1198 {
1199         return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1200 }
1201
1202 static struct file_system_type hugetlbfs_fs_type = {
1203         .name           = "hugetlbfs",
1204         .mount          = hugetlbfs_mount,
1205         .kill_sb        = kill_litter_super,
1206 };
1207 MODULE_ALIAS_FS("hugetlbfs");
1208
1209 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1210
1211 static int can_do_hugetlb_shm(void)
1212 {
1213         kgid_t shm_group;
1214         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1215         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1216 }
1217
1218 static int get_hstate_idx(int page_size_log)
1219 {
1220         struct hstate *h = hstate_sizelog(page_size_log);
1221
1222         if (!h)
1223                 return -1;
1224         return h - hstates;
1225 }
1226
1227 static const struct dentry_operations anon_ops = {
1228         .d_dname = simple_dname
1229 };
1230
1231 /*
1232  * Note that size should be aligned to proper hugepage size in caller side,
1233  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1234  */
1235 struct file *hugetlb_file_setup(const char *name, size_t size,
1236                                 vm_flags_t acctflag, struct user_struct **user,
1237                                 int creat_flags, int page_size_log)
1238 {
1239         struct file *file = ERR_PTR(-ENOMEM);
1240         struct inode *inode;
1241         struct path path;
1242         struct super_block *sb;
1243         struct qstr quick_string;
1244         int hstate_idx;
1245
1246         hstate_idx = get_hstate_idx(page_size_log);
1247         if (hstate_idx < 0)
1248                 return ERR_PTR(-ENODEV);
1249
1250         *user = NULL;
1251         if (!hugetlbfs_vfsmount[hstate_idx])
1252                 return ERR_PTR(-ENOENT);
1253
1254         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1255                 *user = current_user();
1256                 if (user_shm_lock(size, *user)) {
1257                         task_lock(current);
1258                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1259                                 current->comm, current->pid);
1260                         task_unlock(current);
1261                 } else {
1262                         *user = NULL;
1263                         return ERR_PTR(-EPERM);
1264                 }
1265         }
1266
1267         sb = hugetlbfs_vfsmount[hstate_idx]->mnt_sb;
1268         quick_string.name = name;
1269         quick_string.len = strlen(quick_string.name);
1270         quick_string.hash = 0;
1271         path.dentry = d_alloc_pseudo(sb, &quick_string);
1272         if (!path.dentry)
1273                 goto out_shm_unlock;
1274
1275         d_set_d_op(path.dentry, &anon_ops);
1276         path.mnt = mntget(hugetlbfs_vfsmount[hstate_idx]);
1277         file = ERR_PTR(-ENOSPC);
1278         inode = hugetlbfs_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0);
1279         if (!inode)
1280                 goto out_dentry;
1281         if (creat_flags == HUGETLB_SHMFS_INODE)
1282                 inode->i_flags |= S_PRIVATE;
1283
1284         file = ERR_PTR(-ENOMEM);
1285         if (hugetlb_reserve_pages(inode, 0,
1286                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1287                         acctflag))
1288                 goto out_inode;
1289
1290         d_instantiate(path.dentry, inode);
1291         inode->i_size = size;
1292         clear_nlink(inode);
1293
1294         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
1295                         &hugetlbfs_file_operations);
1296         if (IS_ERR(file))
1297                 goto out_dentry; /* inode is already attached */
1298
1299         return file;
1300
1301 out_inode:
1302         iput(inode);
1303 out_dentry:
1304         path_put(&path);
1305 out_shm_unlock:
1306         if (*user) {
1307                 user_shm_unlock(size, *user);
1308                 *user = NULL;
1309         }
1310         return file;
1311 }
1312
1313 static int __init init_hugetlbfs_fs(void)
1314 {
1315         struct hstate *h;
1316         int error;
1317         int i;
1318
1319         if (!hugepages_supported()) {
1320                 pr_info("disabling because there are no supported hugepage sizes\n");
1321                 return -ENOTSUPP;
1322         }
1323
1324         error = -ENOMEM;
1325         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1326                                         sizeof(struct hugetlbfs_inode_info),
1327                                         0, 0, init_once);
1328         if (hugetlbfs_inode_cachep == NULL)
1329                 goto out2;
1330
1331         error = register_filesystem(&hugetlbfs_fs_type);
1332         if (error)
1333                 goto out;
1334
1335         i = 0;
1336         for_each_hstate(h) {
1337                 char buf[50];
1338                 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1339
1340                 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1341                 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1342                                                         buf);
1343
1344                 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1345                         pr_err("Cannot mount internal hugetlbfs for "
1346                                 "page size %uK", ps_kb);
1347                         error = PTR_ERR(hugetlbfs_vfsmount[i]);
1348                         hugetlbfs_vfsmount[i] = NULL;
1349                 }
1350                 i++;
1351         }
1352         /* Non default hstates are optional */
1353         if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1354                 return 0;
1355
1356  out:
1357         kmem_cache_destroy(hugetlbfs_inode_cachep);
1358  out2:
1359         return error;
1360 }
1361
1362 static void __exit exit_hugetlbfs_fs(void)
1363 {
1364         struct hstate *h;
1365         int i;
1366
1367
1368         /*
1369          * Make sure all delayed rcu free inodes are flushed before we
1370          * destroy cache.
1371          */
1372         rcu_barrier();
1373         kmem_cache_destroy(hugetlbfs_inode_cachep);
1374         i = 0;
1375         for_each_hstate(h)
1376                 kern_unmount(hugetlbfs_vfsmount[i++]);
1377         unregister_filesystem(&hugetlbfs_fs_type);
1378 }
1379
1380 module_init(init_hugetlbfs_fs)
1381 module_exit(exit_hugetlbfs_fs)
1382
1383 MODULE_LICENSE("GPL");