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
 467                 /*
 468                  * Can the expression below overflow on 32-bit arches?
 469                  * No, because the interval tree returns us only those vmas
 470                  * which overlap the truncated area starting at pgoff,
 471                  * and no vma on a 32-bit arch can span beyond the 4GB.
 472                  */
 473                 if (vma->vm_pgoff < start)
 474                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 475                 else
 476                         v_offset = 0;
 477
 478                 if (end) {
 479                         end = ((end - start) << PAGE_SHIFT) +
 480                                vma->vm_start + v_offset;
 481                         if (end > vma->vm_end)
 482                                 end = vma->vm_end;
 483                 } else
 484                         end = vma->vm_end;
 485
 486                 unmap_hugepage_range(vma, vma->vm_start + v_offset, end, NULL);
 487         }
 488 }
 489
 490 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 491 {
 492         pgoff_t pgoff;
 493         struct address_space *mapping = inode->i_mapping;
 494         struct hstate *h = hstate_inode(inode);
 495
 496         BUG_ON(offset & ~huge_page_mask(h));
 497         pgoff = offset >> PAGE_SHIFT;
 498
 499         i_size_write(inode, offset);
 500         i_mmap_lock_write(mapping);
 501         if (!RB_EMPTY_ROOT(&mapping->i_mmap))
 502                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 503         i_mmap_unlock_write(mapping);
 504         remove_inode_hugepages(inode, offset, LLONG_MAX);
 505         return 0;
 506 }
 507
 508 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 509 {
 510         struct hstate *h = hstate_inode(inode);
 511         loff_t hpage_size = huge_page_size(h);
 512         loff_t hole_start, hole_end;
 513
 514         /*
 515          * For hole punch round up the beginning offset of the hole and
 516          * round down the end.
 517          */
 518         hole_start = round_up(offset, hpage_size);
 519         hole_end = round_down(offset + len, hpage_size);
 520
 521         if (hole_end > hole_start) {
 522                 struct address_space *mapping = inode->i_mapping;
 523
 524                 mutex_lock(&inode->i_mutex);
 525                 i_mmap_lock_write(mapping);
 526                 if (!RB_EMPTY_ROOT(&mapping->i_mmap))
 527                         hugetlb_vmdelete_list(&mapping->i_mmap,
 528                                                 hole_start >> PAGE_SHIFT,
 529                                                 hole_end  >> PAGE_SHIFT);
 530                 i_mmap_unlock_write(mapping);
 531                 remove_inode_hugepages(inode, hole_start, hole_end);
 532                 mutex_unlock(&inode->i_mutex);
 533         }
 534
 535         return 0;
 536 }
 537
 538 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 539                                 loff_t len)
 540 {
 541         struct inode *inode = file_inode(file);
 542         struct address_space *mapping = inode->i_mapping;
 543         struct hstate *h = hstate_inode(inode);
 544         struct vm_area_struct pseudo_vma;
 545         struct mm_struct *mm = current->mm;
 546         loff_t hpage_size = huge_page_size(h);
 547         unsigned long hpage_shift = huge_page_shift(h);
 548         pgoff_t start, index, end;
 549         int error;
 550         u32 hash;
 551
 552         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 553                 return -EOPNOTSUPP;
 554
 555         if (mode & FALLOC_FL_PUNCH_HOLE)
 556                 return hugetlbfs_punch_hole(inode, offset, len);
 557
 558         /*
 559          * Default preallocate case.
 560          * For this range, start is rounded down and end is rounded up
 561          * as well as being converted to page offsets.
 562          */
 563         start = offset >> hpage_shift;
 564         end = (offset + len + hpage_size - 1) >> hpage_shift;
 565
 566         mutex_lock(&inode->i_mutex);
 567
 568         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 569         error = inode_newsize_ok(inode, offset + len);
 570         if (error)
 571                 goto out;
 572
 573         /*
 574          * Initialize a pseudo vma as this is required by the huge page
 575          * allocation routines.  If NUMA is configured, use page index
 576          * as input to create an allocation policy.
 577          */
 578         memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
 579         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 580         pseudo_vma.vm_file = file;
 581
 582         for (index = start; index < end; index++) {
 583                 /*
 584                  * This is supposed to be the vaddr where the page is being
 585                  * faulted in, but we have no vaddr here.
 586                  */
 587                 struct page *page;
 588                 unsigned long addr;
 589                 int avoid_reserve = 0;
 590
 591                 cond_resched();
 592
 593                 /*
 594                  * fallocate(2) manpage permits EINTR; we may have been
 595                  * interrupted because we are using up too much memory.
 596                  */
 597                 if (signal_pending(current)) {
 598                         error = -EINTR;
 599                         break;
 600                 }
 601
 602                 /* Set numa allocation policy based on index */
 603                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 604
 605                 /* addr is the offset within the file (zero based) */
 606                 addr = index * hpage_size;
 607
 608                 /* mutex taken here, fault path and hole punch */
 609                 hash = hugetlb_fault_mutex_hash(h, mm, &pseudo_vma, mapping,
 610                                                 index, addr);
 611                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 612
 613                 /* See if already present in mapping to avoid alloc/free */
 614                 page = find_get_page(mapping, index);
 615                 if (page) {
 616                         put_page(page);
 617                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 618                         hugetlb_drop_vma_policy(&pseudo_vma);
 619                         continue;
 620                 }
 621
 622                 /* Allocate page and add to page cache */
 623                 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
 624                 hugetlb_drop_vma_policy(&pseudo_vma);
 625                 if (IS_ERR(page)) {
 626                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 627                         error = PTR_ERR(page);
 628                         goto out;
 629                 }
 630                 clear_huge_page(page, addr, pages_per_huge_page(h));
 631                 __SetPageUptodate(page);
 632                 error = huge_add_to_page_cache(page, mapping, index);
 633                 if (unlikely(error)) {
 634                         put_page(page);
 635                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 636                         goto out;
 637                 }
 638
 639                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 640
 641                 /*
 642                  * page_put due to reference from alloc_huge_page()
 643                  * unlock_page because locked by add_to_page_cache()
 644                  */
 645                 put_page(page);
 646                 unlock_page(page);
 647         }
 648
 649         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 650                 i_size_write(inode, offset + len);
 651         inode->i_ctime = CURRENT_TIME;
 652 out:
 653         mutex_unlock(&inode->i_mutex);
 654         return error;
 655 }
 656
 657 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 658 {
 659         struct inode *inode = d_inode(dentry);
 660         struct hstate *h = hstate_inode(inode);
 661         int error;
 662         unsigned int ia_valid = attr->ia_valid;
 663
 664         BUG_ON(!inode);
 665
 666         error = inode_change_ok(inode, attr);
 667         if (error)
 668                 return error;
 669
 670         if (ia_valid & ATTR_SIZE) {
 671                 error = -EINVAL;
 672                 if (attr->ia_size & ~huge_page_mask(h))
 673                         return -EINVAL;
 674                 error = hugetlb_vmtruncate(inode, attr->ia_size);
 675                 if (error)
 676                         return error;
 677         }
 678
 679         setattr_copy(inode, attr);
 680         mark_inode_dirty(inode);
 681         return 0;
 682 }
 683
 684 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 685                                         struct hugetlbfs_config *config)
 686 {
 687         struct inode *inode;
 688
 689         inode = new_inode(sb);
 690         if (inode) {
 691                 struct hugetlbfs_inode_info *info;
 692                 inode->i_ino = get_next_ino();
 693                 inode->i_mode = S_IFDIR | config->mode;
 694                 inode->i_uid = config->uid;
 695                 inode->i_gid = config->gid;
 696                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 697                 info = HUGETLBFS_I(inode);
 698                 mpol_shared_policy_init(&info->policy, NULL);
 699                 inode->i_op = &hugetlbfs_dir_inode_operations;
 700                 inode->i_fop = &simple_dir_operations;
 701                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 702                 inc_nlink(inode);
 703                 lockdep_annotate_inode_mutex_key(inode);
 704         }
 705         return inode;
 706 }
 707
 708 /*
 709  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 710  * be taken from reclaim -- unlike regular filesystems. This needs an
 711  * annotation because huge_pmd_share() does an allocation under
 712  * i_mmap_rwsem.
 713  */
 714 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 715
 716 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 717                                         struct inode *dir,
 718                                         umode_t mode, dev_t dev)
 719 {
 720         struct inode *inode;
 721         struct resv_map *resv_map;
 722
 723         resv_map = resv_map_alloc();
 724         if (!resv_map)
 725                 return NULL;
 726
 727         inode = new_inode(sb);
 728         if (inode) {
 729                 struct hugetlbfs_inode_info *info;
 730                 inode->i_ino = get_next_ino();
 731                 inode_init_owner(inode, dir, mode);
 732                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 733                                 &hugetlbfs_i_mmap_rwsem_key);
 734                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 735                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 736                 inode->i_mapping->private_data = resv_map;
 737                 info = HUGETLBFS_I(inode);
 738                 /*
 739                  * The policy is initialized here even if we are creating a
 740                  * private inode because initialization simply creates an
 741                  * an empty rb tree and calls spin_lock_init(), later when we
 742                  * call mpol_free_shared_policy() it will just return because
 743                  * the rb tree will still be empty.
 744                  */
 745                 mpol_shared_policy_init(&info->policy, NULL);
 746                 switch (mode & S_IFMT) {
 747                 default:
 748                         init_special_inode(inode, mode, dev);
 749                         break;
 750                 case S_IFREG:
 751                         inode->i_op = &hugetlbfs_inode_operations;
 752                         inode->i_fop = &hugetlbfs_file_operations;
 753                         break;
 754                 case S_IFDIR:
 755                         inode->i_op = &hugetlbfs_dir_inode_operations;
 756                         inode->i_fop = &simple_dir_operations;
 757
 758                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 759                         inc_nlink(inode);
 760                         break;
 761                 case S_IFLNK:
 762                         inode->i_op = &page_symlink_inode_operations;
 763                         inode_nohighmem(inode);
 764                         break;
 765                 }
 766                 lockdep_annotate_inode_mutex_key(inode);
 767         } else
 768                 kref_put(&resv_map->refs, resv_map_release);
 769
 770         return inode;
 771 }
 772
 773 /*
 774  * File creation. Allocate an inode, and we're done..
 775  */
 776 static int hugetlbfs_mknod(struct inode *dir,
 777                         struct dentry *dentry, umode_t mode, dev_t dev)
 778 {
 779         struct inode *inode;
 780         int error = -ENOSPC;
 781
 782         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 783         if (inode) {
 784                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
 785                 d_instantiate(dentry, inode);
 786                 dget(dentry);   /* Extra count - pin the dentry in core */
 787                 error = 0;
 788         }
 789         return error;
 790 }
 791
 792 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 793 {
 794         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
 795         if (!retval)
 796                 inc_nlink(dir);
 797         return retval;
 798 }
 799
 800 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
 801 {
 802         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
 803 }
 804
 805 static int hugetlbfs_symlink(struct inode *dir,
 806                         struct dentry *dentry, const char *symname)
 807 {
 808         struct inode *inode;
 809         int error = -ENOSPC;
 810
 811         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 812         if (inode) {
 813                 int l = strlen(symname)+1;
 814                 error = page_symlink(inode, symname, l);
 815                 if (!error) {
 816                         d_instantiate(dentry, inode);
 817                         dget(dentry);
 818                 } else
 819                         iput(inode);
 820         }
 821         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
 822
 823         return error;
 824 }
 825
 826 /*
 827  * mark the head page dirty
 828  */
 829 static int hugetlbfs_set_page_dirty(struct page *page)
 830 {
 831         struct page *head = compound_head(page);
 832
 833         SetPageDirty(head);
 834         return 0;
 835 }
 836
 837 static int hugetlbfs_migrate_page(struct address_space *mapping,
 838                                 struct page *newpage, struct page *page,
 839                                 enum migrate_mode mode)
 840 {
 841         int rc;
 842
 843         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 844         if (rc != MIGRATEPAGE_SUCCESS)
 845                 return rc;
 846         migrate_page_copy(newpage, page);
 847
 848         return MIGRATEPAGE_SUCCESS;
 849 }
 850
 851 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 852 {
 853         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
 854         struct hstate *h = hstate_inode(d_inode(dentry));
 855
 856         buf->f_type = HUGETLBFS_MAGIC;
 857         buf->f_bsize = huge_page_size(h);
 858         if (sbinfo) {
 859                 spin_lock(&sbinfo->stat_lock);
 860                 /* If no limits set, just report 0 for max/free/used
 861                  * blocks, like simple_statfs() */
 862                 if (sbinfo->spool) {
 863                         long free_pages;
 864
 865                         spin_lock(&sbinfo->spool->lock);
 866                         buf->f_blocks = sbinfo->spool->max_hpages;
 867                         free_pages = sbinfo->spool->max_hpages
 868                                 - sbinfo->spool->used_hpages;
 869                         buf->f_bavail = buf->f_bfree = free_pages;
 870                         spin_unlock(&sbinfo->spool->lock);
 871                         buf->f_files = sbinfo->max_inodes;
 872                         buf->f_ffree = sbinfo->free_inodes;
 873                 }
 874                 spin_unlock(&sbinfo->stat_lock);
 875         }
 876         buf->f_namelen = NAME_MAX;
 877         return 0;
 878 }
 879
 880 static void hugetlbfs_put_super(struct super_block *sb)
 881 {
 882         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
 883
 884         if (sbi) {
 885                 sb->s_fs_info = NULL;
 886
 887                 if (sbi->spool)
 888                         hugepage_put_subpool(sbi->spool);
 889
 890                 kfree(sbi);
 891         }
 892 }
 893
 894 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 895 {
 896         if (sbinfo->free_inodes >= 0) {
 897                 spin_lock(&sbinfo->stat_lock);
 898                 if (unlikely(!sbinfo->free_inodes)) {
 899                         spin_unlock(&sbinfo->stat_lock);
 900                         return 0;
 901                 }
 902                 sbinfo->free_inodes--;
 903                 spin_unlock(&sbinfo->stat_lock);
 904         }
 905
 906         return 1;
 907 }
 908
 909 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 910 {
 911         if (sbinfo->free_inodes >= 0) {
 912                 spin_lock(&sbinfo->stat_lock);
 913                 sbinfo->free_inodes++;
 914                 spin_unlock(&sbinfo->stat_lock);
 915         }
 916 }
 917
 918
 919 static struct kmem_cache *hugetlbfs_inode_cachep;
 920
 921 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
 922 {
 923         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
 924         struct hugetlbfs_inode_info *p;
 925
 926         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
 927                 return NULL;
 928         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
 929         if (unlikely(!p)) {
 930                 hugetlbfs_inc_free_inodes(sbinfo);
 931                 return NULL;
 932         }
 933         return &p->vfs_inode;
 934 }
 935
 936 static void hugetlbfs_i_callback(struct rcu_head *head)
 937 {
 938         struct inode *inode = container_of(head, struct inode, i_rcu);
 939         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
 940 }
 941
 942 static void hugetlbfs_destroy_inode(struct inode *inode)
 943 {
 944         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
 945         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
 946         call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
 947 }
 948
 949 static const struct address_space_operations hugetlbfs_aops = {
 950         .write_begin    = hugetlbfs_write_begin,
 951         .write_end      = hugetlbfs_write_end,
 952         .set_page_dirty = hugetlbfs_set_page_dirty,
 953         .migratepage    = hugetlbfs_migrate_page,
 954 };
 955
 956
 957 static void init_once(void *foo)
 958 {
 959         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
 960
 961         inode_init_once(&ei->vfs_inode);
 962 }
 963
 964 const struct file_operations hugetlbfs_file_operations = {
 965         .read_iter              = hugetlbfs_read_iter,
 966         .mmap                   = hugetlbfs_file_mmap,
 967         .fsync                  = noop_fsync,
 968         .get_unmapped_area      = hugetlb_get_unmapped_area,
 969         .llseek                 = default_llseek,
 970         .fallocate              = hugetlbfs_fallocate,
 971 };
 972
 973 static const struct inode_operations hugetlbfs_dir_inode_operations = {
 974         .create         = hugetlbfs_create,
 975         .lookup         = simple_lookup,
 976         .link           = simple_link,
 977         .unlink         = simple_unlink,
 978         .symlink        = hugetlbfs_symlink,
 979         .mkdir          = hugetlbfs_mkdir,
 980         .rmdir          = simple_rmdir,
 981         .mknod          = hugetlbfs_mknod,
 982         .rename         = simple_rename,
 983         .setattr        = hugetlbfs_setattr,
 984 };
 985
 986 static const struct inode_operations hugetlbfs_inode_operations = {
 987         .setattr        = hugetlbfs_setattr,
 988 };
 989
 990 static const struct super_operations hugetlbfs_ops = {
 991         .alloc_inode    = hugetlbfs_alloc_inode,
 992         .destroy_inode  = hugetlbfs_destroy_inode,
 993         .evict_inode    = hugetlbfs_evict_inode,
 994         .statfs         = hugetlbfs_statfs,
 995         .put_super      = hugetlbfs_put_super,
 996         .show_options   = generic_show_options,
 997 };
 998
 999 enum { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1000
1001 /*
1002  * Convert size option passed from command line to number of huge pages
1003  * in the pool specified by hstate.  Size option could be in bytes
1004  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1005  */
1006 static long long
1007 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1008                                                                 int val_type)
1009 {
1010         if (val_type == NO_SIZE)
1011                 return -1;
1012
1013         if (val_type == SIZE_PERCENT) {
1014                 size_opt <<= huge_page_shift(h);
1015                 size_opt *= h->max_huge_pages;
1016                 do_div(size_opt, 100);
1017         }
1018
1019         size_opt >>= huge_page_shift(h);
1020         return size_opt;
1021 }
1022
1023 static int
1024 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1025 {
1026         char *p, *rest;
1027         substring_t args[MAX_OPT_ARGS];
1028         int option;
1029         unsigned long long max_size_opt = 0, min_size_opt = 0;
1030         int max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1031
1032         if (!options)
1033                 return 0;
1034
1035         while ((p = strsep(&options, ",")) != NULL) {
1036                 int token;
1037                 if (!*p)
1038                         continue;
1039
1040                 token = match_token(p, tokens, args);
1041                 switch (token) {
1042                 case Opt_uid:
1043                         if (match_int(&args[0], &option))
1044                                 goto bad_val;
1045                         pconfig->uid = make_kuid(current_user_ns(), option);
1046                         if (!uid_valid(pconfig->uid))
1047                                 goto bad_val;
1048                         break;
1049
1050                 case Opt_gid:
1051                         if (match_int(&args[0], &option))
1052                                 goto bad_val;
1053                         pconfig->gid = make_kgid(current_user_ns(), option);
1054                         if (!gid_valid(pconfig->gid))
1055                                 goto bad_val;
1056                         break;
1057
1058                 case Opt_mode:
1059                         if (match_octal(&args[0], &option))
1060                                 goto bad_val;
1061                         pconfig->mode = option & 01777U;
1062                         break;
1063
1064                 case Opt_size: {
1065                         /* memparse() will accept a K/M/G without a digit */
1066                         if (!isdigit(*args[0].from))
1067                                 goto bad_val;
1068                         max_size_opt = memparse(args[0].from, &rest);
1069                         max_val_type = SIZE_STD;
1070                         if (*rest == '%')
1071                                 max_val_type = SIZE_PERCENT;
1072                         break;
1073                 }
1074
1075                 case Opt_nr_inodes:
1076                         /* memparse() will accept a K/M/G without a digit */
1077                         if (!isdigit(*args[0].from))
1078                                 goto bad_val;
1079                         pconfig->nr_inodes = memparse(args[0].from, &rest);
1080                         break;
1081
1082                 case Opt_pagesize: {
1083                         unsigned long ps;
1084                         ps = memparse(args[0].from, &rest);
1085                         pconfig->hstate = size_to_hstate(ps);
1086                         if (!pconfig->hstate) {
1087                                 pr_err("Unsupported page size %lu MB\n",
1088                                         ps >> 20);
1089                                 return -EINVAL;
1090                         }
1091                         break;
1092                 }
1093
1094                 case Opt_min_size: {
1095                         /* memparse() will accept a K/M/G without a digit */
1096                         if (!isdigit(*args[0].from))
1097                                 goto bad_val;
1098                         min_size_opt = memparse(args[0].from, &rest);
1099                         min_val_type = SIZE_STD;
1100                         if (*rest == '%')
1101                                 min_val_type = SIZE_PERCENT;
1102                         break;
1103                 }
1104
1105                 default:
1106                         pr_err("Bad mount option: \"%s\"\n", p);
1107                         return -EINVAL;
1108                         break;
1109                 }
1110         }
1111
1112         /*
1113          * Use huge page pool size (in hstate) to convert the size
1114          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1115          */
1116         pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1117                                                 max_size_opt, max_val_type);
1118         pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1119                                                 min_size_opt, min_val_type);
1120
1121         /*
1122          * If max_size was specified, then min_size must be smaller
1123          */
1124         if (max_val_type > NO_SIZE &&
1125             pconfig->min_hpages > pconfig->max_hpages) {
1126                 pr_err("minimum size can not be greater than maximum size\n");
1127                 return -EINVAL;
1128         }
1129
1130         return 0;
1131
1132 bad_val:
1133         pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1134         return -EINVAL;
1135 }
1136
1137 static int
1138 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1139 {
1140         int ret;
1141         struct hugetlbfs_config config;
1142         struct hugetlbfs_sb_info *sbinfo;
1143
1144         save_mount_options(sb, data);
1145
1146         config.max_hpages = -1; /* No limit on size by default */
1147         config.nr_inodes = -1; /* No limit on number of inodes by default */
1148         config.uid = current_fsuid();
1149         config.gid = current_fsgid();
1150         config.mode = 0755;
1151         config.hstate = &default_hstate;
1152         config.min_hpages = -1; /* No default minimum size */
1153         ret = hugetlbfs_parse_options(data, &config);
1154         if (ret)
1155                 return ret;
1156
1157         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1158         if (!sbinfo)
1159                 return -ENOMEM;
1160         sb->s_fs_info = sbinfo;
1161         sbinfo->hstate = config.hstate;
1162         spin_lock_init(&sbinfo->stat_lock);
1163         sbinfo->max_inodes = config.nr_inodes;
1164         sbinfo->free_inodes = config.nr_inodes;
1165         sbinfo->spool = NULL;
1166         /*
1167          * Allocate and initialize subpool if maximum or minimum size is
1168          * specified.  Any needed reservations (for minimim size) are taken
1169          * taken when the subpool is created.
1170          */
1171         if (config.max_hpages != -1 || config.min_hpages != -1) {
1172                 sbinfo->spool = hugepage_new_subpool(config.hstate,
1173                                                         config.max_hpages,
1174                                                         config.min_hpages);
1175                 if (!sbinfo->spool)
1176                         goto out_free;
1177         }
1178         sb->s_maxbytes = MAX_LFS_FILESIZE;
1179         sb->s_blocksize = huge_page_size(config.hstate);
1180         sb->s_blocksize_bits = huge_page_shift(config.hstate);
1181         sb->s_magic = HUGETLBFS_MAGIC;
1182         sb->s_op = &hugetlbfs_ops;
1183         sb->s_time_gran = 1;
1184         sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1185         if (!sb->s_root)
1186                 goto out_free;
1187         return 0;
1188 out_free:
1189         kfree(sbinfo->spool);
1190         kfree(sbinfo);
1191         return -ENOMEM;
1192 }
1193
1194 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1195         int flags, const char *dev_name, void *data)
1196 {
1197         return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1198 }
1199
1200 static struct file_system_type hugetlbfs_fs_type = {
1201         .name           = "hugetlbfs",
1202         .mount          = hugetlbfs_mount,
1203         .kill_sb        = kill_litter_super,
1204 };
1205 MODULE_ALIAS_FS("hugetlbfs");
1206
1207 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1208
1209 static int can_do_hugetlb_shm(void)
1210 {
1211         kgid_t shm_group;
1212         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1213         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1214 }
1215
1216 static int get_hstate_idx(int page_size_log)
1217 {
1218         struct hstate *h = hstate_sizelog(page_size_log);
1219
1220         if (!h)
1221                 return -1;
1222         return h - hstates;
1223 }
1224
1225 static const struct dentry_operations anon_ops = {
1226         .d_dname = simple_dname
1227 };
1228
1229 /*
1230  * Note that size should be aligned to proper hugepage size in caller side,
1231  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1232  */
1233 struct file *hugetlb_file_setup(const char *name, size_t size,
1234                                 vm_flags_t acctflag, struct user_struct **user,
1235                                 int creat_flags, int page_size_log)
1236 {
1237         struct file *file = ERR_PTR(-ENOMEM);
1238         struct inode *inode;
1239         struct path path;
1240         struct super_block *sb;
1241         struct qstr quick_string;
1242         int hstate_idx;
1243
1244         hstate_idx = get_hstate_idx(page_size_log);
1245         if (hstate_idx < 0)
1246                 return ERR_PTR(-ENODEV);
1247
1248         *user = NULL;
1249         if (!hugetlbfs_vfsmount[hstate_idx])
1250                 return ERR_PTR(-ENOENT);
1251
1252         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1253                 *user = current_user();
1254                 if (user_shm_lock(size, *user)) {
1255                         task_lock(current);
1256                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1257                                 current->comm, current->pid);
1258                         task_unlock(current);
1259                 } else {
1260                         *user = NULL;
1261                         return ERR_PTR(-EPERM);
1262                 }
1263         }
1264
1265         sb = hugetlbfs_vfsmount[hstate_idx]->mnt_sb;
1266         quick_string.name = name;
1267         quick_string.len = strlen(quick_string.name);
1268         quick_string.hash = 0;
1269         path.dentry = d_alloc_pseudo(sb, &quick_string);
1270         if (!path.dentry)
1271                 goto out_shm_unlock;
1272
1273         d_set_d_op(path.dentry, &anon_ops);
1274         path.mnt = mntget(hugetlbfs_vfsmount[hstate_idx]);
1275         file = ERR_PTR(-ENOSPC);
1276         inode = hugetlbfs_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0);
1277         if (!inode)
1278                 goto out_dentry;
1279         if (creat_flags == HUGETLB_SHMFS_INODE)
1280                 inode->i_flags |= S_PRIVATE;
1281
1282         file = ERR_PTR(-ENOMEM);
1283         if (hugetlb_reserve_pages(inode, 0,
1284                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1285                         acctflag))
1286                 goto out_inode;
1287
1288         d_instantiate(path.dentry, inode);
1289         inode->i_size = size;
1290         clear_nlink(inode);
1291
1292         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
1293                         &hugetlbfs_file_operations);
1294         if (IS_ERR(file))
1295                 goto out_dentry; /* inode is already attached */
1296
1297         return file;
1298
1299 out_inode:
1300         iput(inode);
1301 out_dentry:
1302         path_put(&path);
1303 out_shm_unlock:
1304         if (*user) {
1305                 user_shm_unlock(size, *user);
1306                 *user = NULL;
1307         }
1308         return file;
1309 }
1310
1311 static int __init init_hugetlbfs_fs(void)
1312 {
1313         struct hstate *h;
1314         int error;
1315         int i;
1316
1317         if (!hugepages_supported()) {
1318                 pr_info("disabling because there are no supported hugepage sizes\n");
1319                 return -ENOTSUPP;
1320         }
1321
1322         error = -ENOMEM;
1323         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1324                                         sizeof(struct hugetlbfs_inode_info),
1325                                         0, 0, init_once);
1326         if (hugetlbfs_inode_cachep == NULL)
1327                 goto out2;
1328
1329         error = register_filesystem(&hugetlbfs_fs_type);
1330         if (error)
1331                 goto out;
1332
1333         i = 0;
1334         for_each_hstate(h) {
1335                 char buf[50];
1336                 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1337
1338                 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1339                 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1340                                                         buf);
1341
1342                 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1343                         pr_err("Cannot mount internal hugetlbfs for "
1344                                 "page size %uK", ps_kb);
1345                         error = PTR_ERR(hugetlbfs_vfsmount[i]);
1346                         hugetlbfs_vfsmount[i] = NULL;
1347                 }
1348                 i++;
1349         }
1350         /* Non default hstates are optional */
1351         if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1352                 return 0;
1353
1354  out:
1355         kmem_cache_destroy(hugetlbfs_inode_cachep);
1356  out2:
1357         return error;
1358 }
1359
1360 static void __exit exit_hugetlbfs_fs(void)
1361 {
1362         struct hstate *h;
1363         int i;
1364
1365
1366         /*
1367          * Make sure all delayed rcu free inodes are flushed before we
1368          * destroy cache.
1369          */
1370         rcu_barrier();
1371         kmem_cache_destroy(hugetlbfs_inode_cachep);
1372         i = 0;
1373         for_each_hstate(h)
1374                 kern_unmount(hugetlbfs_vfsmount[i++]);
1375         unregister_filesystem(&hugetlbfs_fs_type);
1376 }
1377
1378 module_init(init_hugetlbfs_fs)
1379 module_exit(exit_hugetlbfs_fs)
1380
1381 MODULE_LICENSE("GPL");