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.
 336  * hole punch is indicated if end is not LLONG_MAX
 337  *      In the hole punch case we scan the range and release found pages.
 338  *      Only when releasing a page is the associated region/reserv map
 339  *      deleted.  The region/reserv map for ranges without associated
 340  *      pages are not modified.
 341  * Note: If the passed end of range value is beyond the end of file, but
 342  * not LLONG_MAX this routine still performs a hole punch operation.
 343  */
 344 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 345                                    loff_t lend)
 346 {
 347         struct hstate *h = hstate_inode(inode);
 348         struct address_space *mapping = &inode->i_data;
 349         const pgoff_t start = lstart >> huge_page_shift(h);
 350         const pgoff_t end = lend >> huge_page_shift(h);
 351         struct vm_area_struct pseudo_vma;
 352         struct pagevec pvec;
 353         pgoff_t next;
 354         int i, freed = 0;
 355         long lookup_nr = PAGEVEC_SIZE;
 356         bool truncate_op = (lend == LLONG_MAX);
 357
 358         memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
 359         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 360         pagevec_init(&pvec, 0);
 361         next = start;
 362         while (next < end) {
 363                 /*
 364                  * Make sure to never grab more pages that we
 365                  * might possibly need.
 366                  */
 367                 if (end - next < lookup_nr)
 368                         lookup_nr = end - next;
 369
 370                 /*
 371                  * This pagevec_lookup() may return pages past 'end',
 372                  * so we must check for page->index > end.
 373                  */
 374                 if (!pagevec_lookup(&pvec, mapping, next, lookup_nr)) {
 375                         if (next == start)
 376                                 break;
 377                         next = start;
 378                         continue;
 379                 }
 380
 381                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 382                         struct page *page = pvec.pages[i];
 383                         u32 hash;
 384
 385                         hash = hugetlb_fault_mutex_hash(h, current->mm,
 386                                                         &pseudo_vma,
 387                                                         mapping, next, 0);
 388                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 389
 390                         lock_page(page);
 391                         if (page->index >= end) {
 392                                 unlock_page(page);
 393                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 394                                 next = end;     /* we are done */
 395                                 break;
 396                         }
 397
 398                         /*
 399                          * If page is mapped, it was faulted in after being
 400                          * unmapped.  Do nothing in this race case.  In the
 401                          * normal case page is not mapped.
 402                          */
 403                         if (!page_mapped(page)) {
 404                                 bool rsv_on_error = !PagePrivate(page);
 405                                 /*
 406                                  * We must free the huge page and remove
 407                                  * from page cache (remove_huge_page) BEFORE
 408                                  * removing the region/reserve map
 409                                  * (hugetlb_unreserve_pages).  In rare out
 410                                  * of memory conditions, removal of the
 411                                  * region/reserve map could fail.  Before
 412                                  * free'ing the page, note PagePrivate which
 413                                  * is used in case of error.
 414                                  */
 415                                 remove_huge_page(page);
 416                                 freed++;
 417                                 if (!truncate_op) {
 418                                         if (unlikely(hugetlb_unreserve_pages(
 419                                                         inode, next,
 420                                                         next + 1, 1)))
 421                                                 hugetlb_fix_reserve_counts(
 422                                                         inode, rsv_on_error);
 423                                 }
 424                         }
 425
 426                         if (page->index > next)
 427                                 next = page->index;
 428
 429                         ++next;
 430                         unlock_page(page);
 431
 432                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 433                 }
 434                 huge_pagevec_release(&pvec);
 435         }
 436
 437         if (truncate_op)
 438                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 439 }
 440
 441 static void hugetlbfs_evict_inode(struct inode *inode)
 442 {
 443         struct resv_map *resv_map;
 444
 445         remove_inode_hugepages(inode, 0, LLONG_MAX);
 446         resv_map = (struct resv_map *)inode->i_mapping->private_data;
 447         /* root inode doesn't have the resv_map, so we should check it */
 448         if (resv_map)
 449                 resv_map_release(&resv_map->refs);
 450         clear_inode(inode);
 451 }
 452
 453 static inline void
 454 hugetlb_vmdelete_list(struct rb_root *root, pgoff_t start, pgoff_t end)
 455 {
 456         struct vm_area_struct *vma;
 457
 458         /*
 459          * end == 0 indicates that the entire range after
 460          * start should be unmapped.
 461          */
 462         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
 463                 unsigned long v_offset;
 464
 465                 /*
 466                  * Can the expression below overflow on 32-bit arches?
 467                  * No, because the interval tree returns us only those vmas
 468                  * which overlap the truncated area starting at pgoff,
 469                  * and no vma on a 32-bit arch can span beyond the 4GB.
 470                  */
 471                 if (vma->vm_pgoff < start)
 472                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 473                 else
 474                         v_offset = 0;
 475
 476                 if (end) {
 477                         end = ((end - start) << PAGE_SHIFT) +
 478                                vma->vm_start + v_offset;
 479                         if (end > vma->vm_end)
 480                                 end = vma->vm_end;
 481                 } else
 482                         end = vma->vm_end;
 483
 484                 unmap_hugepage_range(vma, vma->vm_start + v_offset, end, NULL);
 485         }
 486 }
 487
 488 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 489 {
 490         pgoff_t pgoff;
 491         struct address_space *mapping = inode->i_mapping;
 492         struct hstate *h = hstate_inode(inode);
 493
 494         BUG_ON(offset & ~huge_page_mask(h));
 495         pgoff = offset >> PAGE_SHIFT;
 496
 497         i_size_write(inode, offset);
 498         i_mmap_lock_write(mapping);
 499         if (!RB_EMPTY_ROOT(&mapping->i_mmap))
 500                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 501         i_mmap_unlock_write(mapping);
 502         remove_inode_hugepages(inode, offset, LLONG_MAX);
 503         return 0;
 504 }
 505
 506 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 507 {
 508         struct hstate *h = hstate_inode(inode);
 509         loff_t hpage_size = huge_page_size(h);
 510         unsigned long hpage_shift = huge_page_shift(h);
 511         loff_t hole_start, hole_end;
 512
 513         /*
 514          * For hole punch round up the beginning offset of the hole and
 515          * round down the end.
 516          */
 517         hole_start = round_up(offset, hpage_size);
 518         hole_end = round_down(offset + len, hpage_size);
 519
 520         if (hole_end > hole_start) {
 521                 struct address_space *mapping = inode->i_mapping;
 522                 DECLARE_WAIT_QUEUE_HEAD_ONSTACK(hugetlb_falloc_waitq);
 523                 /*
 524                  * Page faults on the area to be hole punched must be stopped
 525                  * during the operation.  Initialize struct and have
 526                  * inode->i_private point to it.
 527                  */
 528                 struct hugetlb_falloc hugetlb_falloc = {
 529                         .waitq = &hugetlb_falloc_waitq,
 530                         .start = hole_start >> hpage_shift,
 531                         .end = hole_end >> hpage_shift
 532                 };
 533
 534                 mutex_lock(&inode->i_mutex);
 535
 536                 /*
 537                  * inode->i_private will be checked in the page fault path.
 538                  * The locking assures that all writes to the structure are
 539                  * complete before assigning to i_private.  A fault on another
 540                  * CPU will see the fully initialized structure.
 541                  */
 542                 spin_lock(&inode->i_lock);
 543                 inode->i_private = &hugetlb_falloc;
 544                 spin_unlock(&inode->i_lock);
 545
 546                 i_mmap_lock_write(mapping);
 547                 if (!RB_EMPTY_ROOT(&mapping->i_mmap))
 548                         hugetlb_vmdelete_list(&mapping->i_mmap,
 549                                                 hole_start >> PAGE_SHIFT,
 550                                                 hole_end  >> PAGE_SHIFT);
 551                 i_mmap_unlock_write(mapping);
 552                 remove_inode_hugepages(inode, hole_start, hole_end);
 553
 554                 spin_lock(&inode->i_lock);
 555                 inode->i_private = NULL;
 556                 wake_up_all(&hugetlb_falloc_waitq);
 557                 spin_unlock(&inode->i_lock);
 558
 559                 mutex_unlock(&inode->i_mutex);
 560         }
 561
 562         return 0;
 563 }
 564
 565 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 566                                 loff_t len)
 567 {
 568         struct inode *inode = file_inode(file);
 569         struct address_space *mapping = inode->i_mapping;
 570         struct hstate *h = hstate_inode(inode);
 571         struct vm_area_struct pseudo_vma;
 572         struct mm_struct *mm = current->mm;
 573         loff_t hpage_size = huge_page_size(h);
 574         unsigned long hpage_shift = huge_page_shift(h);
 575         pgoff_t start, index, end;
 576         int error;
 577         u32 hash;
 578
 579         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 580                 return -EOPNOTSUPP;
 581
 582         if (mode & FALLOC_FL_PUNCH_HOLE)
 583                 return hugetlbfs_punch_hole(inode, offset, len);
 584
 585         /*
 586          * Default preallocate case.
 587          * For this range, start is rounded down and end is rounded up
 588          * as well as being converted to page offsets.
 589          */
 590         start = offset >> hpage_shift;
 591         end = (offset + len + hpage_size - 1) >> hpage_shift;
 592
 593         mutex_lock(&inode->i_mutex);
 594
 595         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 596         error = inode_newsize_ok(inode, offset + len);
 597         if (error)
 598                 goto out;
 599
 600         /*
 601          * Initialize a pseudo vma as this is required by the huge page
 602          * allocation routines.  If NUMA is configured, use page index
 603          * as input to create an allocation policy.
 604          */
 605         memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
 606         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 607         pseudo_vma.vm_file = file;
 608
 609         for (index = start; index < end; index++) {
 610                 /*
 611                  * This is supposed to be the vaddr where the page is being
 612                  * faulted in, but we have no vaddr here.
 613                  */
 614                 struct page *page;
 615                 unsigned long addr;
 616                 int avoid_reserve = 0;
 617
 618                 cond_resched();
 619
 620                 /*
 621                  * fallocate(2) manpage permits EINTR; we may have been
 622                  * interrupted because we are using up too much memory.
 623                  */
 624                 if (signal_pending(current)) {
 625                         error = -EINTR;
 626                         break;
 627                 }
 628
 629                 /* Set numa allocation policy based on index */
 630                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 631
 632                 /* addr is the offset within the file (zero based) */
 633                 addr = index * hpage_size;
 634
 635                 /* mutex taken here, fault path and hole punch */
 636                 hash = hugetlb_fault_mutex_hash(h, mm, &pseudo_vma, mapping,
 637                                                 index, addr);
 638                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 639
 640                 /* See if already present in mapping to avoid alloc/free */
 641                 page = find_get_page(mapping, index);
 642                 if (page) {
 643                         put_page(page);
 644                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 645                         hugetlb_drop_vma_policy(&pseudo_vma);
 646                         continue;
 647                 }
 648
 649                 /* Allocate page and add to page cache */
 650                 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
 651                 hugetlb_drop_vma_policy(&pseudo_vma);
 652                 if (IS_ERR(page)) {
 653                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 654                         error = PTR_ERR(page);
 655                         goto out;
 656                 }
 657                 clear_huge_page(page, addr, pages_per_huge_page(h));
 658                 __SetPageUptodate(page);
 659                 error = huge_add_to_page_cache(page, mapping, index);
 660                 if (unlikely(error)) {
 661                         put_page(page);
 662                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 663                         goto out;
 664                 }
 665
 666                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 667
 668                 /*
 669                  * page_put due to reference from alloc_huge_page()
 670                  * unlock_page because locked by add_to_page_cache()
 671                  */
 672                 put_page(page);
 673                 unlock_page(page);
 674         }
 675
 676         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 677                 i_size_write(inode, offset + len);
 678         inode->i_ctime = CURRENT_TIME;
 679 out:
 680         mutex_unlock(&inode->i_mutex);
 681         return error;
 682 }
 683
 684 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 685 {
 686         struct inode *inode = d_inode(dentry);
 687         struct hstate *h = hstate_inode(inode);
 688         int error;
 689         unsigned int ia_valid = attr->ia_valid;
 690
 691         BUG_ON(!inode);
 692
 693         error = inode_change_ok(inode, attr);
 694         if (error)
 695                 return error;
 696
 697         if (ia_valid & ATTR_SIZE) {
 698                 error = -EINVAL;
 699                 if (attr->ia_size & ~huge_page_mask(h))
 700                         return -EINVAL;
 701                 error = hugetlb_vmtruncate(inode, attr->ia_size);
 702                 if (error)
 703                         return error;
 704         }
 705
 706         setattr_copy(inode, attr);
 707         mark_inode_dirty(inode);
 708         return 0;
 709 }
 710
 711 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 712                                         struct hugetlbfs_config *config)
 713 {
 714         struct inode *inode;
 715
 716         inode = new_inode(sb);
 717         if (inode) {
 718                 struct hugetlbfs_inode_info *info;
 719                 inode->i_ino = get_next_ino();
 720                 inode->i_mode = S_IFDIR | config->mode;
 721                 inode->i_uid = config->uid;
 722                 inode->i_gid = config->gid;
 723                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 724                 info = HUGETLBFS_I(inode);
 725                 mpol_shared_policy_init(&info->policy, NULL);
 726                 inode->i_op = &hugetlbfs_dir_inode_operations;
 727                 inode->i_fop = &simple_dir_operations;
 728                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 729                 inc_nlink(inode);
 730                 lockdep_annotate_inode_mutex_key(inode);
 731         }
 732         return inode;
 733 }
 734
 735 /*
 736  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 737  * be taken from reclaim -- unlike regular filesystems. This needs an
 738  * annotation because huge_pmd_share() does an allocation under
 739  * i_mmap_rwsem.
 740  */
 741 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 742
 743 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 744                                         struct inode *dir,
 745                                         umode_t mode, dev_t dev)
 746 {
 747         struct inode *inode;
 748         struct resv_map *resv_map;
 749
 750         resv_map = resv_map_alloc();
 751         if (!resv_map)
 752                 return NULL;
 753
 754         inode = new_inode(sb);
 755         if (inode) {
 756                 struct hugetlbfs_inode_info *info;
 757                 inode->i_ino = get_next_ino();
 758                 inode_init_owner(inode, dir, mode);
 759                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 760                                 &hugetlbfs_i_mmap_rwsem_key);
 761                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 762                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 763                 inode->i_mapping->private_data = resv_map;
 764                 info = HUGETLBFS_I(inode);
 765                 /*
 766                  * The policy is initialized here even if we are creating a
 767                  * private inode because initialization simply creates an
 768                  * an empty rb tree and calls spin_lock_init(), later when we
 769                  * call mpol_free_shared_policy() it will just return because
 770                  * the rb tree will still be empty.
 771                  */
 772                 mpol_shared_policy_init(&info->policy, NULL);
 773                 switch (mode & S_IFMT) {
 774                 default:
 775                         init_special_inode(inode, mode, dev);
 776                         break;
 777                 case S_IFREG:
 778                         inode->i_op = &hugetlbfs_inode_operations;
 779                         inode->i_fop = &hugetlbfs_file_operations;
 780                         break;
 781                 case S_IFDIR:
 782                         inode->i_op = &hugetlbfs_dir_inode_operations;
 783                         inode->i_fop = &simple_dir_operations;
 784
 785                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 786                         inc_nlink(inode);
 787                         break;
 788                 case S_IFLNK:
 789                         inode->i_op = &page_symlink_inode_operations;
 790                         break;
 791                 }
 792                 lockdep_annotate_inode_mutex_key(inode);
 793         } else
 794                 kref_put(&resv_map->refs, resv_map_release);
 795
 796         return inode;
 797 }
 798
 799 /*
 800  * File creation. Allocate an inode, and we're done..
 801  */
 802 static int hugetlbfs_mknod(struct inode *dir,
 803                         struct dentry *dentry, umode_t mode, dev_t dev)
 804 {
 805         struct inode *inode;
 806         int error = -ENOSPC;
 807
 808         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 809         if (inode) {
 810                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
 811                 d_instantiate(dentry, inode);
 812                 dget(dentry);   /* Extra count - pin the dentry in core */
 813                 error = 0;
 814         }
 815         return error;
 816 }
 817
 818 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 819 {
 820         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
 821         if (!retval)
 822                 inc_nlink(dir);
 823         return retval;
 824 }
 825
 826 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
 827 {
 828         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
 829 }
 830
 831 static int hugetlbfs_symlink(struct inode *dir,
 832                         struct dentry *dentry, const char *symname)
 833 {
 834         struct inode *inode;
 835         int error = -ENOSPC;
 836
 837         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 838         if (inode) {
 839                 int l = strlen(symname)+1;
 840                 error = page_symlink(inode, symname, l);
 841                 if (!error) {
 842                         d_instantiate(dentry, inode);
 843                         dget(dentry);
 844                 } else
 845                         iput(inode);
 846         }
 847         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
 848
 849         return error;
 850 }
 851
 852 /*
 853  * mark the head page dirty
 854  */
 855 static int hugetlbfs_set_page_dirty(struct page *page)
 856 {
 857         struct page *head = compound_head(page);
 858
 859         SetPageDirty(head);
 860         return 0;
 861 }
 862
 863 static int hugetlbfs_migrate_page(struct address_space *mapping,
 864                                 struct page *newpage, struct page *page,
 865                                 enum migrate_mode mode)
 866 {
 867         int rc;
 868
 869         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 870         if (rc != MIGRATEPAGE_SUCCESS)
 871                 return rc;
 872         migrate_page_copy(newpage, page);
 873
 874         return MIGRATEPAGE_SUCCESS;
 875 }
 876
 877 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 878 {
 879         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
 880         struct hstate *h = hstate_inode(d_inode(dentry));
 881
 882         buf->f_type = HUGETLBFS_MAGIC;
 883         buf->f_bsize = huge_page_size(h);
 884         if (sbinfo) {
 885                 spin_lock(&sbinfo->stat_lock);
 886                 /* If no limits set, just report 0 for max/free/used
 887                  * blocks, like simple_statfs() */
 888                 if (sbinfo->spool) {
 889                         long free_pages;
 890
 891                         spin_lock(&sbinfo->spool->lock);
 892                         buf->f_blocks = sbinfo->spool->max_hpages;
 893                         free_pages = sbinfo->spool->max_hpages
 894                                 - sbinfo->spool->used_hpages;
 895                         buf->f_bavail = buf->f_bfree = free_pages;
 896                         spin_unlock(&sbinfo->spool->lock);
 897                         buf->f_files = sbinfo->max_inodes;
 898                         buf->f_ffree = sbinfo->free_inodes;
 899                 }
 900                 spin_unlock(&sbinfo->stat_lock);
 901         }
 902         buf->f_namelen = NAME_MAX;
 903         return 0;
 904 }
 905
 906 static void hugetlbfs_put_super(struct super_block *sb)
 907 {
 908         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
 909
 910         if (sbi) {
 911                 sb->s_fs_info = NULL;
 912
 913                 if (sbi->spool)
 914                         hugepage_put_subpool(sbi->spool);
 915
 916                 kfree(sbi);
 917         }
 918 }
 919
 920 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 921 {
 922         if (sbinfo->free_inodes >= 0) {
 923                 spin_lock(&sbinfo->stat_lock);
 924                 if (unlikely(!sbinfo->free_inodes)) {
 925                         spin_unlock(&sbinfo->stat_lock);
 926                         return 0;
 927                 }
 928                 sbinfo->free_inodes--;
 929                 spin_unlock(&sbinfo->stat_lock);
 930         }
 931
 932         return 1;
 933 }
 934
 935 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 936 {
 937         if (sbinfo->free_inodes >= 0) {
 938                 spin_lock(&sbinfo->stat_lock);
 939                 sbinfo->free_inodes++;
 940                 spin_unlock(&sbinfo->stat_lock);
 941         }
 942 }
 943
 944
 945 static struct kmem_cache *hugetlbfs_inode_cachep;
 946
 947 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
 948 {
 949         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
 950         struct hugetlbfs_inode_info *p;
 951
 952         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
 953                 return NULL;
 954         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
 955         if (unlikely(!p)) {
 956                 hugetlbfs_inc_free_inodes(sbinfo);
 957                 return NULL;
 958         }
 959         return &p->vfs_inode;
 960 }
 961
 962 static void hugetlbfs_i_callback(struct rcu_head *head)
 963 {
 964         struct inode *inode = container_of(head, struct inode, i_rcu);
 965         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
 966 }
 967
 968 static void hugetlbfs_destroy_inode(struct inode *inode)
 969 {
 970         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
 971         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
 972         call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
 973 }
 974
 975 static const struct address_space_operations hugetlbfs_aops = {
 976         .write_begin    = hugetlbfs_write_begin,
 977         .write_end      = hugetlbfs_write_end,
 978         .set_page_dirty = hugetlbfs_set_page_dirty,
 979         .migratepage    = hugetlbfs_migrate_page,
 980 };
 981
 982
 983 static void init_once(void *foo)
 984 {
 985         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
 986
 987         inode_init_once(&ei->vfs_inode);
 988 }
 989
 990 const struct file_operations hugetlbfs_file_operations = {
 991         .read_iter              = hugetlbfs_read_iter,
 992         .mmap                   = hugetlbfs_file_mmap,
 993         .fsync                  = noop_fsync,
 994         .get_unmapped_area      = hugetlb_get_unmapped_area,
 995         .llseek                 = default_llseek,
 996         .fallocate              = hugetlbfs_fallocate,
 997 };
 998
 999 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1000         .create         = hugetlbfs_create,
1001         .lookup         = simple_lookup,
1002         .link           = simple_link,
1003         .unlink         = simple_unlink,
1004         .symlink        = hugetlbfs_symlink,
1005         .mkdir          = hugetlbfs_mkdir,
1006         .rmdir          = simple_rmdir,
1007         .mknod          = hugetlbfs_mknod,
1008         .rename         = simple_rename,
1009         .setattr        = hugetlbfs_setattr,
1010 };
1011
1012 static const struct inode_operations hugetlbfs_inode_operations = {
1013         .setattr        = hugetlbfs_setattr,
1014 };
1015
1016 static const struct super_operations hugetlbfs_ops = {
1017         .alloc_inode    = hugetlbfs_alloc_inode,
1018         .destroy_inode  = hugetlbfs_destroy_inode,
1019         .evict_inode    = hugetlbfs_evict_inode,
1020         .statfs         = hugetlbfs_statfs,
1021         .put_super      = hugetlbfs_put_super,
1022         .show_options   = generic_show_options,
1023 };
1024
1025 enum { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1026
1027 /*
1028  * Convert size option passed from command line to number of huge pages
1029  * in the pool specified by hstate.  Size option could be in bytes
1030  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1031  */
1032 static long long
1033 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1034                                                                 int val_type)
1035 {
1036         if (val_type == NO_SIZE)
1037                 return -1;
1038
1039         if (val_type == SIZE_PERCENT) {
1040                 size_opt <<= huge_page_shift(h);
1041                 size_opt *= h->max_huge_pages;
1042                 do_div(size_opt, 100);
1043         }
1044
1045         size_opt >>= huge_page_shift(h);
1046         return size_opt;
1047 }
1048
1049 static int
1050 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1051 {
1052         char *p, *rest;
1053         substring_t args[MAX_OPT_ARGS];
1054         int option;
1055         unsigned long long max_size_opt = 0, min_size_opt = 0;
1056         int max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1057
1058         if (!options)
1059                 return 0;
1060
1061         while ((p = strsep(&options, ",")) != NULL) {
1062                 int token;
1063                 if (!*p)
1064                         continue;
1065
1066                 token = match_token(p, tokens, args);
1067                 switch (token) {
1068                 case Opt_uid:
1069                         if (match_int(&args[0], &option))
1070                                 goto bad_val;
1071                         pconfig->uid = make_kuid(current_user_ns(), option);
1072                         if (!uid_valid(pconfig->uid))
1073                                 goto bad_val;
1074                         break;
1075
1076                 case Opt_gid:
1077                         if (match_int(&args[0], &option))
1078                                 goto bad_val;
1079                         pconfig->gid = make_kgid(current_user_ns(), option);
1080                         if (!gid_valid(pconfig->gid))
1081                                 goto bad_val;
1082                         break;
1083
1084                 case Opt_mode:
1085                         if (match_octal(&args[0], &option))
1086                                 goto bad_val;
1087                         pconfig->mode = option & 01777U;
1088                         break;
1089
1090                 case Opt_size: {
1091                         /* memparse() will accept a K/M/G without a digit */
1092                         if (!isdigit(*args[0].from))
1093                                 goto bad_val;
1094                         max_size_opt = memparse(args[0].from, &rest);
1095                         max_val_type = SIZE_STD;
1096                         if (*rest == '%')
1097                                 max_val_type = SIZE_PERCENT;
1098                         break;
1099                 }
1100
1101                 case Opt_nr_inodes:
1102                         /* memparse() will accept a K/M/G without a digit */
1103                         if (!isdigit(*args[0].from))
1104                                 goto bad_val;
1105                         pconfig->nr_inodes = memparse(args[0].from, &rest);
1106                         break;
1107
1108                 case Opt_pagesize: {
1109                         unsigned long ps;
1110                         ps = memparse(args[0].from, &rest);
1111                         pconfig->hstate = size_to_hstate(ps);
1112                         if (!pconfig->hstate) {
1113                                 pr_err("Unsupported page size %lu MB\n",
1114                                         ps >> 20);
1115                                 return -EINVAL;
1116                         }
1117                         break;
1118                 }
1119
1120                 case Opt_min_size: {
1121                         /* memparse() will accept a K/M/G without a digit */
1122                         if (!isdigit(*args[0].from))
1123                                 goto bad_val;
1124                         min_size_opt = memparse(args[0].from, &rest);
1125                         min_val_type = SIZE_STD;
1126                         if (*rest == '%')
1127                                 min_val_type = SIZE_PERCENT;
1128                         break;
1129                 }
1130
1131                 default:
1132                         pr_err("Bad mount option: \"%s\"\n", p);
1133                         return -EINVAL;
1134                         break;
1135                 }
1136         }
1137
1138         /*
1139          * Use huge page pool size (in hstate) to convert the size
1140          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1141          */
1142         pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1143                                                 max_size_opt, max_val_type);
1144         pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1145                                                 min_size_opt, min_val_type);
1146
1147         /*
1148          * If max_size was specified, then min_size must be smaller
1149          */
1150         if (max_val_type > NO_SIZE &&
1151             pconfig->min_hpages > pconfig->max_hpages) {
1152                 pr_err("minimum size can not be greater than maximum size\n");
1153                 return -EINVAL;
1154         }
1155
1156         return 0;
1157
1158 bad_val:
1159         pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1160         return -EINVAL;
1161 }
1162
1163 static int
1164 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1165 {
1166         int ret;
1167         struct hugetlbfs_config config;
1168         struct hugetlbfs_sb_info *sbinfo;
1169
1170         save_mount_options(sb, data);
1171
1172         config.max_hpages = -1; /* No limit on size by default */
1173         config.nr_inodes = -1; /* No limit on number of inodes by default */
1174         config.uid = current_fsuid();
1175         config.gid = current_fsgid();
1176         config.mode = 0755;
1177         config.hstate = &default_hstate;
1178         config.min_hpages = -1; /* No default minimum size */
1179         ret = hugetlbfs_parse_options(data, &config);
1180         if (ret)
1181                 return ret;
1182
1183         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1184         if (!sbinfo)
1185                 return -ENOMEM;
1186         sb->s_fs_info = sbinfo;
1187         sbinfo->hstate = config.hstate;
1188         spin_lock_init(&sbinfo->stat_lock);
1189         sbinfo->max_inodes = config.nr_inodes;
1190         sbinfo->free_inodes = config.nr_inodes;
1191         sbinfo->spool = NULL;
1192         /*
1193          * Allocate and initialize subpool if maximum or minimum size is
1194          * specified.  Any needed reservations (for minimim size) are taken
1195          * taken when the subpool is created.
1196          */
1197         if (config.max_hpages != -1 || config.min_hpages != -1) {
1198                 sbinfo->spool = hugepage_new_subpool(config.hstate,
1199                                                         config.max_hpages,
1200                                                         config.min_hpages);
1201                 if (!sbinfo->spool)
1202                         goto out_free;
1203         }
1204         sb->s_maxbytes = MAX_LFS_FILESIZE;
1205         sb->s_blocksize = huge_page_size(config.hstate);
1206         sb->s_blocksize_bits = huge_page_shift(config.hstate);
1207         sb->s_magic = HUGETLBFS_MAGIC;
1208         sb->s_op = &hugetlbfs_ops;
1209         sb->s_time_gran = 1;
1210         sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1211         if (!sb->s_root)
1212                 goto out_free;
1213         return 0;
1214 out_free:
1215         kfree(sbinfo->spool);
1216         kfree(sbinfo);
1217         return -ENOMEM;
1218 }
1219
1220 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1221         int flags, const char *dev_name, void *data)
1222 {
1223         return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1224 }
1225
1226 static struct file_system_type hugetlbfs_fs_type = {
1227         .name           = "hugetlbfs",
1228         .mount          = hugetlbfs_mount,
1229         .kill_sb        = kill_litter_super,
1230 };
1231 MODULE_ALIAS_FS("hugetlbfs");
1232
1233 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1234
1235 static int can_do_hugetlb_shm(void)
1236 {
1237         kgid_t shm_group;
1238         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1239         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1240 }
1241
1242 static int get_hstate_idx(int page_size_log)
1243 {
1244         struct hstate *h = hstate_sizelog(page_size_log);
1245
1246         if (!h)
1247                 return -1;
1248         return h - hstates;
1249 }
1250
1251 static const struct dentry_operations anon_ops = {
1252         .d_dname = simple_dname
1253 };
1254
1255 /*
1256  * Note that size should be aligned to proper hugepage size in caller side,
1257  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1258  */
1259 struct file *hugetlb_file_setup(const char *name, size_t size,
1260                                 vm_flags_t acctflag, struct user_struct **user,
1261                                 int creat_flags, int page_size_log)
1262 {
1263         struct file *file = ERR_PTR(-ENOMEM);
1264         struct inode *inode;
1265         struct path path;
1266         struct super_block *sb;
1267         struct qstr quick_string;
1268         int hstate_idx;
1269
1270         hstate_idx = get_hstate_idx(page_size_log);
1271         if (hstate_idx < 0)
1272                 return ERR_PTR(-ENODEV);
1273
1274         *user = NULL;
1275         if (!hugetlbfs_vfsmount[hstate_idx])
1276                 return ERR_PTR(-ENOENT);
1277
1278         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1279                 *user = current_user();
1280                 if (user_shm_lock(size, *user)) {
1281                         task_lock(current);
1282                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1283                                 current->comm, current->pid);
1284                         task_unlock(current);
1285                 } else {
1286                         *user = NULL;
1287                         return ERR_PTR(-EPERM);
1288                 }
1289         }
1290
1291         sb = hugetlbfs_vfsmount[hstate_idx]->mnt_sb;
1292         quick_string.name = name;
1293         quick_string.len = strlen(quick_string.name);
1294         quick_string.hash = 0;
1295         path.dentry = d_alloc_pseudo(sb, &quick_string);
1296         if (!path.dentry)
1297                 goto out_shm_unlock;
1298
1299         d_set_d_op(path.dentry, &anon_ops);
1300         path.mnt = mntget(hugetlbfs_vfsmount[hstate_idx]);
1301         file = ERR_PTR(-ENOSPC);
1302         inode = hugetlbfs_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0);
1303         if (!inode)
1304                 goto out_dentry;
1305         if (creat_flags == HUGETLB_SHMFS_INODE)
1306                 inode->i_flags |= S_PRIVATE;
1307
1308         file = ERR_PTR(-ENOMEM);
1309         if (hugetlb_reserve_pages(inode, 0,
1310                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1311                         acctflag))
1312                 goto out_inode;
1313
1314         d_instantiate(path.dentry, inode);
1315         inode->i_size = size;
1316         clear_nlink(inode);
1317
1318         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
1319                         &hugetlbfs_file_operations);
1320         if (IS_ERR(file))
1321                 goto out_dentry; /* inode is already attached */
1322
1323         return file;
1324
1325 out_inode:
1326         iput(inode);
1327 out_dentry:
1328         path_put(&path);
1329 out_shm_unlock:
1330         if (*user) {
1331                 user_shm_unlock(size, *user);
1332                 *user = NULL;
1333         }
1334         return file;
1335 }
1336
1337 static int __init init_hugetlbfs_fs(void)
1338 {
1339         struct hstate *h;
1340         int error;
1341         int i;
1342
1343         if (!hugepages_supported()) {
1344                 pr_info("disabling because there are no supported hugepage sizes\n");
1345                 return -ENOTSUPP;
1346         }
1347
1348         error = -ENOMEM;
1349         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1350                                         sizeof(struct hugetlbfs_inode_info),
1351                                         0, 0, init_once);
1352         if (hugetlbfs_inode_cachep == NULL)
1353                 goto out2;
1354
1355         error = register_filesystem(&hugetlbfs_fs_type);
1356         if (error)
1357                 goto out;
1358
1359         i = 0;
1360         for_each_hstate(h) {
1361                 char buf[50];
1362                 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1363
1364                 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1365                 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1366                                                         buf);
1367
1368                 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1369                         pr_err("Cannot mount internal hugetlbfs for "
1370                                 "page size %uK", ps_kb);
1371                         error = PTR_ERR(hugetlbfs_vfsmount[i]);
1372                         hugetlbfs_vfsmount[i] = NULL;
1373                 }
1374                 i++;
1375         }
1376         /* Non default hstates are optional */
1377         if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1378                 return 0;
1379
1380  out:
1381         kmem_cache_destroy(hugetlbfs_inode_cachep);
1382  out2:
1383         return error;
1384 }
1385
1386 static void __exit exit_hugetlbfs_fs(void)
1387 {
1388         struct hstate *h;
1389         int i;
1390
1391
1392         /*
1393          * Make sure all delayed rcu free inodes are flushed before we
1394          * destroy cache.
1395          */
1396         rcu_barrier();
1397         kmem_cache_destroy(hugetlbfs_inode_cachep);
1398         i = 0;
1399         for_each_hstate(h)
1400                 kern_unmount(hugetlbfs_vfsmount[i++]);
1401         unregister_filesystem(&hugetlbfs_fs_type);
1402 }
1403
1404 module_init(init_hugetlbfs_fs)
1405 module_exit(exit_hugetlbfs_fs)
1406
1407 MODULE_LICENSE("GPL");