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  * License: GPL
   8  */
   9
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/thread_info.h>
  13 #include <asm/current.h>
  14 #include <linux/sched/signal.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 <linux/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         struct hstate           *hstate;
  50         long                    max_hpages;
  51         long                    nr_inodes;
  52         long                    min_hpages;
  53         kuid_t                  uid;
  54         kgid_t                  gid;
  55         umode_t                 mode;
  56 };
  57
  58 int sysctl_hugetlb_shm_group;
  59
  60 enum {
  61         Opt_size, Opt_nr_inodes,
  62         Opt_mode, Opt_uid, Opt_gid,
  63         Opt_pagesize, Opt_min_size,
  64         Opt_err,
  65 };
  66
  67 static const match_table_t tokens = {
  68         {Opt_size,      "size=%s"},
  69         {Opt_nr_inodes, "nr_inodes=%s"},
  70         {Opt_mode,      "mode=%o"},
  71         {Opt_uid,       "uid=%u"},
  72         {Opt_gid,       "gid=%u"},
  73         {Opt_pagesize,  "pagesize=%s"},
  74         {Opt_min_size,  "min_size=%s"},
  75         {Opt_err,       NULL},
  76 };
  77
  78 #ifdef CONFIG_NUMA
  79 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  80                                         struct inode *inode, pgoff_t index)
  81 {
  82         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
  83                                                         index);
  84 }
  85
  86 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  87 {
  88         mpol_cond_put(vma->vm_policy);
  89 }
  90 #else
  91 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  92                                         struct inode *inode, pgoff_t index)
  93 {
  94 }
  95
  96 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  97 {
  98 }
  99 #endif
 100
 101 static void huge_pagevec_release(struct pagevec *pvec)
 102 {
 103         int i;
 104
 105         for (i = 0; i < pagevec_count(pvec); ++i)
 106                 put_page(pvec->pages[i]);
 107
 108         pagevec_reinit(pvec);
 109 }
 110
 111 /*
 112  * Mask used when checking the page offset value passed in via system
 113  * calls.  This value will be converted to a loff_t which is signed.
 114  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 115  * value.  The extra bit (- 1 in the shift value) is to take the sign
 116  * bit into account.
 117  */
 118 #define PGOFF_LOFFT_MAX \
 119         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 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         /*
 140          * page based offset in vm_pgoff could be sufficiently large to
 141          * overflow a loff_t when converted to byte offset.  This can
 142          * only happen on architectures where sizeof(loff_t) ==
 143          * sizeof(unsigned long).  So, only check in those instances.
 144          */
 145         if (sizeof(unsigned long) == sizeof(loff_t)) {
 146                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 147                         return -EINVAL;
 148         }
 149
 150         /* must be huge page aligned */
 151         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 152                 return -EINVAL;
 153
 154         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 155         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 156         /* check for overflow */
 157         if (len < vma_len)
 158                 return -EINVAL;
 159
 160         inode_lock(inode);
 161         file_accessed(file);
 162
 163         ret = -ENOMEM;
 164         if (hugetlb_reserve_pages(inode,
 165                                 vma->vm_pgoff >> huge_page_order(h),
 166                                 len >> huge_page_shift(h), vma,
 167                                 vma->vm_flags))
 168                 goto out;
 169
 170         ret = 0;
 171         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 172                 i_size_write(inode, len);
 173 out:
 174         inode_unlock(inode);
 175
 176         return ret;
 177 }
 178
 179 /*
 180  * Called under down_write(mmap_sem).
 181  */
 182
 183 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 184 static unsigned long
 185 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 186                 unsigned long len, unsigned long pgoff, unsigned long flags)
 187 {
 188         struct mm_struct *mm = current->mm;
 189         struct vm_area_struct *vma;
 190         struct hstate *h = hstate_file(file);
 191         struct vm_unmapped_area_info info;
 192
 193         if (len & ~huge_page_mask(h))
 194                 return -EINVAL;
 195         if (len > TASK_SIZE)
 196                 return -ENOMEM;
 197
 198         if (flags & MAP_FIXED) {
 199                 if (prepare_hugepage_range(file, addr, len))
 200                         return -EINVAL;
 201                 return addr;
 202         }
 203
 204         if (addr) {
 205                 addr = ALIGN(addr, huge_page_size(h));
 206                 vma = find_vma(mm, addr);
 207                 if (TASK_SIZE - len >= addr &&
 208                     (!vma || addr + len <= vm_start_gap(vma)))
 209                         return addr;
 210         }
 211
 212         info.flags = 0;
 213         info.length = len;
 214         info.low_limit = TASK_UNMAPPED_BASE;
 215         info.high_limit = TASK_SIZE;
 216         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 217         info.align_offset = 0;
 218         return vm_unmapped_area(&info);
 219 }
 220 #endif
 221
 222 static size_t
 223 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 224                         struct iov_iter *to, unsigned long size)
 225 {
 226         size_t copied = 0;
 227         int i, chunksize;
 228
 229         /* Find which 4k chunk and offset with in that chunk */
 230         i = offset >> PAGE_SHIFT;
 231         offset = offset & ~PAGE_MASK;
 232
 233         while (size) {
 234                 size_t n;
 235                 chunksize = PAGE_SIZE;
 236                 if (offset)
 237                         chunksize -= offset;
 238                 if (chunksize > size)
 239                         chunksize = size;
 240                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 241                 copied += n;
 242                 if (n != chunksize)
 243                         return copied;
 244                 offset = 0;
 245                 size -= chunksize;
 246                 i++;
 247         }
 248         return copied;
 249 }
 250
 251 /*
 252  * Support for read() - Find the page attached to f_mapping and copy out the
 253  * data. Its *very* similar to do_generic_mapping_read(), we can't use that
 254  * since it has PAGE_SIZE assumptions.
 255  */
 256 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 257 {
 258         struct file *file = iocb->ki_filp;
 259         struct hstate *h = hstate_file(file);
 260         struct address_space *mapping = file->f_mapping;
 261         struct inode *inode = mapping->host;
 262         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 263         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 264         unsigned long end_index;
 265         loff_t isize;
 266         ssize_t retval = 0;
 267
 268         while (iov_iter_count(to)) {
 269                 struct page *page;
 270                 size_t nr, copied;
 271
 272                 /* nr is the maximum number of bytes to copy from this page */
 273                 nr = huge_page_size(h);
 274                 isize = i_size_read(inode);
 275                 if (!isize)
 276                         break;
 277                 end_index = (isize - 1) >> huge_page_shift(h);
 278                 if (index > end_index)
 279                         break;
 280                 if (index == end_index) {
 281                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 282                         if (nr <= offset)
 283                                 break;
 284                 }
 285                 nr = nr - offset;
 286
 287                 /* Find the page */
 288                 page = find_lock_page(mapping, index);
 289                 if (unlikely(page == NULL)) {
 290                         /*
 291                          * We have a HOLE, zero out the user-buffer for the
 292                          * length of the hole or request.
 293                          */
 294                         copied = iov_iter_zero(nr, to);
 295                 } else {
 296                         unlock_page(page);
 297
 298                         /*
 299                          * We have the page, copy it to user space buffer.
 300                          */
 301                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 302                         put_page(page);
 303                 }
 304                 offset += copied;
 305                 retval += copied;
 306                 if (copied != nr && iov_iter_count(to)) {
 307                         if (!retval)
 308                                 retval = -EFAULT;
 309                         break;
 310                 }
 311                 index += offset >> huge_page_shift(h);
 312                 offset &= ~huge_page_mask(h);
 313         }
 314         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 315         return retval;
 316 }
 317
 318 static int hugetlbfs_write_begin(struct file *file,
 319                         struct address_space *mapping,
 320                         loff_t pos, unsigned len, unsigned flags,
 321                         struct page **pagep, void **fsdata)
 322 {
 323         return -EINVAL;
 324 }
 325
 326 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 327                         loff_t pos, unsigned len, unsigned copied,
 328                         struct page *page, void *fsdata)
 329 {
 330         BUG();
 331         return -EINVAL;
 332 }
 333
 334 static void remove_huge_page(struct page *page)
 335 {
 336         ClearPageDirty(page);
 337         ClearPageUptodate(page);
 338         delete_from_page_cache(page);
 339 }
 340
 341 static void
 342 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
 343 {
 344         struct vm_area_struct *vma;
 345
 346         /*
 347          * end == 0 indicates that the entire range after
 348          * start should be unmapped.
 349          */
 350         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
 351                 unsigned long v_offset;
 352                 unsigned long v_end;
 353
 354                 /*
 355                  * Can the expression below overflow on 32-bit arches?
 356                  * No, because the interval tree returns us only those vmas
 357                  * which overlap the truncated area starting at pgoff,
 358                  * and no vma on a 32-bit arch can span beyond the 4GB.
 359                  */
 360                 if (vma->vm_pgoff < start)
 361                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 362                 else
 363                         v_offset = 0;
 364
 365                 if (!end)
 366                         v_end = vma->vm_end;
 367                 else {
 368                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 369                                                         + vma->vm_start;
 370                         if (v_end > vma->vm_end)
 371                                 v_end = vma->vm_end;
 372                 }
 373
 374                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 375                                                                         NULL);
 376         }
 377 }
 378
 379 /*
 380  * remove_inode_hugepages handles two distinct cases: truncation and hole
 381  * punch.  There are subtle differences in operation for each case.
 382  *
 383  * truncation is indicated by end of range being LLONG_MAX
 384  *      In this case, we first scan the range and release found pages.
 385  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
 386  *      maps and global counts.
 387  * hole punch is indicated if end is not LLONG_MAX
 388  *      In the hole punch case we scan the range and release found pages.
 389  *      Only when releasing a page is the associated region/reserv map
 390  *      deleted.  The region/reserv map for ranges without associated
 391  *      pages are not modified.
 392  *
 393  * Callers of this routine must hold the i_mmap_rwsem in write mode to prevent
 394  * races with page faults.
 395  *
 396  * Note: If the passed end of range value is beyond the end of file, but
 397  * not LLONG_MAX this routine still performs a hole punch operation.
 398  */
 399 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 400                                    loff_t lend)
 401 {
 402         struct hstate *h = hstate_inode(inode);
 403         struct address_space *mapping = &inode->i_data;
 404         const pgoff_t start = lstart >> huge_page_shift(h);
 405         const pgoff_t end = lend >> huge_page_shift(h);
 406         struct vm_area_struct pseudo_vma;
 407         struct pagevec pvec;
 408         pgoff_t next, index;
 409         int i, freed = 0;
 410         bool truncate_op = (lend == LLONG_MAX);
 411
 412         vma_init(&pseudo_vma, current->mm);
 413         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 414         pagevec_init(&pvec);
 415         next = start;
 416         while (next < end) {
 417                 /*
 418                  * When no more pages are found, we are done.
 419                  */
 420                 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
 421                         break;
 422
 423                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 424                         struct page *page = pvec.pages[i];
 425
 426                         index = page->index;
 427                         /*
 428                          * A mapped page is impossible as callers should unmap
 429                          * all references before calling.  And, i_mmap_rwsem
 430                          * prevents the creation of additional mappings.
 431                          */
 432                         VM_BUG_ON(page_mapped(page));
 433
 434                         lock_page(page);
 435                         /*
 436                          * We must free the huge page and remove from page
 437                          * cache (remove_huge_page) BEFORE removing the
 438                          * region/reserve map (hugetlb_unreserve_pages).  In
 439                          * rare out of memory conditions, removal of the
 440                          * region/reserve map could fail. Correspondingly,
 441                          * the subpool and global reserve usage count can need
 442                          * to be adjusted.
 443                          */
 444                         VM_BUG_ON(PagePrivate(page));
 445                         remove_huge_page(page);
 446                         freed++;
 447                         if (!truncate_op) {
 448                                 if (unlikely(hugetlb_unreserve_pages(inode,
 449                                                         index, index + 1, 1)))
 450                                         hugetlb_fix_reserve_counts(inode);
 451                         }
 452
 453                         unlock_page(page);
 454                 }
 455                 huge_pagevec_release(&pvec);
 456                 cond_resched();
 457         }
 458
 459         if (truncate_op)
 460                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 461 }
 462
 463 static void hugetlbfs_evict_inode(struct inode *inode)
 464 {
 465         struct address_space *mapping = inode->i_mapping;
 466         struct resv_map *resv_map;
 467
 468         /*
 469          * The vfs layer guarantees that there are no other users of this
 470          * inode.  Therefore, it would be safe to call remove_inode_hugepages
 471          * without holding i_mmap_rwsem.  We acquire and hold here to be
 472          * consistent with other callers.  Since there will be no contention
 473          * on the semaphore, overhead is negligible.
 474          */
 475         i_mmap_lock_write(mapping);
 476         remove_inode_hugepages(inode, 0, LLONG_MAX);
 477         i_mmap_unlock_write(mapping);
 478
 479         resv_map = (struct resv_map *)inode->i_mapping->private_data;
 480         /* root inode doesn't have the resv_map, so we should check it */
 481         if (resv_map)
 482                 resv_map_release(&resv_map->refs);
 483         clear_inode(inode);
 484 }
 485
 486 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 487 {
 488         pgoff_t pgoff;
 489         struct address_space *mapping = inode->i_mapping;
 490         struct hstate *h = hstate_inode(inode);
 491
 492         BUG_ON(offset & ~huge_page_mask(h));
 493         pgoff = offset >> PAGE_SHIFT;
 494
 495         i_size_write(inode, offset);
 496         i_mmap_lock_write(mapping);
 497         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 498                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 499         remove_inode_hugepages(inode, offset, LLONG_MAX);
 500         i_mmap_unlock_write(mapping);
 501         return 0;
 502 }
 503
 504 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 505 {
 506         struct hstate *h = hstate_inode(inode);
 507         loff_t hpage_size = huge_page_size(h);
 508         loff_t hole_start, hole_end;
 509
 510         /*
 511          * For hole punch round up the beginning offset of the hole and
 512          * round down the end.
 513          */
 514         hole_start = round_up(offset, hpage_size);
 515         hole_end = round_down(offset + len, hpage_size);
 516
 517         if (hole_end > hole_start) {
 518                 struct address_space *mapping = inode->i_mapping;
 519                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 520
 521                 inode_lock(inode);
 522
 523                 /* protected by i_mutex */
 524                 if (info->seals & F_SEAL_WRITE) {
 525                         inode_unlock(inode);
 526                         return -EPERM;
 527                 }
 528
 529                 i_mmap_lock_write(mapping);
 530                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 531                         hugetlb_vmdelete_list(&mapping->i_mmap,
 532                                                 hole_start >> PAGE_SHIFT,
 533                                                 hole_end  >> PAGE_SHIFT);
 534                 remove_inode_hugepages(inode, hole_start, hole_end);
 535                 i_mmap_unlock_write(mapping);
 536                 inode_unlock(inode);
 537         }
 538
 539         return 0;
 540 }
 541
 542 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 543                                 loff_t len)
 544 {
 545         struct inode *inode = file_inode(file);
 546         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 547         struct address_space *mapping = inode->i_mapping;
 548         struct hstate *h = hstate_inode(inode);
 549         struct vm_area_struct pseudo_vma;
 550         struct mm_struct *mm = current->mm;
 551         loff_t hpage_size = huge_page_size(h);
 552         unsigned long hpage_shift = huge_page_shift(h);
 553         pgoff_t start, index, end;
 554         int error;
 555         u32 hash;
 556
 557         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 558                 return -EOPNOTSUPP;
 559
 560         if (mode & FALLOC_FL_PUNCH_HOLE)
 561                 return hugetlbfs_punch_hole(inode, offset, len);
 562
 563         /*
 564          * Default preallocate case.
 565          * For this range, start is rounded down and end is rounded up
 566          * as well as being converted to page offsets.
 567          */
 568         start = offset >> hpage_shift;
 569         end = (offset + len + hpage_size - 1) >> hpage_shift;
 570
 571         inode_lock(inode);
 572
 573         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 574         error = inode_newsize_ok(inode, offset + len);
 575         if (error)
 576                 goto out;
 577
 578         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 579                 error = -EPERM;
 580                 goto out;
 581         }
 582
 583         /*
 584          * Initialize a pseudo vma as this is required by the huge page
 585          * allocation routines.  If NUMA is configured, use page index
 586          * as input to create an allocation policy.
 587          */
 588         vma_init(&pseudo_vma, mm);
 589         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 590         pseudo_vma.vm_file = file;
 591
 592         for (index = start; index < end; index++) {
 593                 /*
 594                  * This is supposed to be the vaddr where the page is being
 595                  * faulted in, but we have no vaddr here.
 596                  */
 597                 struct page *page;
 598                 unsigned long addr;
 599                 int avoid_reserve = 0;
 600
 601                 cond_resched();
 602
 603                 /*
 604                  * fallocate(2) manpage permits EINTR; we may have been
 605                  * interrupted because we are using up too much memory.
 606                  */
 607                 if (signal_pending(current)) {
 608                         error = -EINTR;
 609                         break;
 610                 }
 611
 612                 /* Set numa allocation policy based on index */
 613                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 614
 615                 /* addr is the offset within the file (zero based) */
 616                 addr = index * hpage_size;
 617
 618                 /*
 619                  * fault mutex taken here, protects against fault path
 620                  * and hole punch.  inode_lock previously taken protects
 621                  * against truncation.
 622                  */
 623                 hash = hugetlb_fault_mutex_hash(h, mm, &pseudo_vma, mapping,
 624                                                 index, addr);
 625                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 626
 627                 /* See if already present in mapping to avoid alloc/free */
 628                 page = find_get_page(mapping, index);
 629                 if (page) {
 630                         put_page(page);
 631                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 632                         hugetlb_drop_vma_policy(&pseudo_vma);
 633                         continue;
 634                 }
 635
 636                 /* Allocate page and add to page cache */
 637                 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
 638                 hugetlb_drop_vma_policy(&pseudo_vma);
 639                 if (IS_ERR(page)) {
 640                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 641                         error = PTR_ERR(page);
 642                         goto out;
 643                 }
 644                 clear_huge_page(page, addr, pages_per_huge_page(h));
 645                 __SetPageUptodate(page);
 646                 error = huge_add_to_page_cache(page, mapping, index);
 647                 if (unlikely(error)) {
 648                         put_page(page);
 649                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 650                         goto out;
 651                 }
 652
 653                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 654
 655                 /*
 656                  * unlock_page because locked by add_to_page_cache()
 657                  * page_put due to reference from alloc_huge_page()
 658                  */
 659                 unlock_page(page);
 660                 put_page(page);
 661         }
 662
 663         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 664                 i_size_write(inode, offset + len);
 665         inode->i_ctime = current_time(inode);
 666 out:
 667         inode_unlock(inode);
 668         return error;
 669 }
 670
 671 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 672 {
 673         struct inode *inode = d_inode(dentry);
 674         struct hstate *h = hstate_inode(inode);
 675         int error;
 676         unsigned int ia_valid = attr->ia_valid;
 677         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 678
 679         BUG_ON(!inode);
 680
 681         error = setattr_prepare(dentry, attr);
 682         if (error)
 683                 return error;
 684
 685         if (ia_valid & ATTR_SIZE) {
 686                 loff_t oldsize = inode->i_size;
 687                 loff_t newsize = attr->ia_size;
 688
 689                 if (newsize & ~huge_page_mask(h))
 690                         return -EINVAL;
 691                 /* protected by i_mutex */
 692                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 693                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 694                         return -EPERM;
 695                 error = hugetlb_vmtruncate(inode, newsize);
 696                 if (error)
 697                         return error;
 698         }
 699
 700         setattr_copy(inode, attr);
 701         mark_inode_dirty(inode);
 702         return 0;
 703 }
 704
 705 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 706                                         struct hugetlbfs_config *config)
 707 {
 708         struct inode *inode;
 709
 710         inode = new_inode(sb);
 711         if (inode) {
 712                 inode->i_ino = get_next_ino();
 713                 inode->i_mode = S_IFDIR | config->mode;
 714                 inode->i_uid = config->uid;
 715                 inode->i_gid = config->gid;
 716                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 717                 inode->i_op = &hugetlbfs_dir_inode_operations;
 718                 inode->i_fop = &simple_dir_operations;
 719                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 720                 inc_nlink(inode);
 721                 lockdep_annotate_inode_mutex_key(inode);
 722         }
 723         return inode;
 724 }
 725
 726 /*
 727  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 728  * be taken from reclaim -- unlike regular filesystems. This needs an
 729  * annotation because huge_pmd_share() does an allocation under hugetlb's
 730  * i_mmap_rwsem.
 731  */
 732 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 733
 734 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 735                                         struct inode *dir,
 736                                         umode_t mode, dev_t dev)
 737 {
 738         struct inode *inode;
 739         struct resv_map *resv_map;
 740
 741         resv_map = resv_map_alloc();
 742         if (!resv_map)
 743                 return NULL;
 744
 745         inode = new_inode(sb);
 746         if (inode) {
 747                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 748
 749                 inode->i_ino = get_next_ino();
 750                 inode_init_owner(inode, dir, mode);
 751                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 752                                 &hugetlbfs_i_mmap_rwsem_key);
 753                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 754                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 755                 inode->i_mapping->private_data = resv_map;
 756                 info->seals = F_SEAL_SEAL;
 757                 switch (mode & S_IFMT) {
 758                 default:
 759                         init_special_inode(inode, mode, dev);
 760                         break;
 761                 case S_IFREG:
 762                         inode->i_op = &hugetlbfs_inode_operations;
 763                         inode->i_fop = &hugetlbfs_file_operations;
 764                         break;
 765                 case S_IFDIR:
 766                         inode->i_op = &hugetlbfs_dir_inode_operations;
 767                         inode->i_fop = &simple_dir_operations;
 768
 769                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 770                         inc_nlink(inode);
 771                         break;
 772                 case S_IFLNK:
 773                         inode->i_op = &page_symlink_inode_operations;
 774                         inode_nohighmem(inode);
 775                         break;
 776                 }
 777                 lockdep_annotate_inode_mutex_key(inode);
 778         } else
 779                 kref_put(&resv_map->refs, resv_map_release);
 780
 781         return inode;
 782 }
 783
 784 /*
 785  * File creation. Allocate an inode, and we're done..
 786  */
 787 static int hugetlbfs_mknod(struct inode *dir,
 788                         struct dentry *dentry, umode_t mode, dev_t dev)
 789 {
 790         struct inode *inode;
 791         int error = -ENOSPC;
 792
 793         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 794         if (inode) {
 795                 dir->i_ctime = dir->i_mtime = current_time(dir);
 796                 d_instantiate(dentry, inode);
 797                 dget(dentry);   /* Extra count - pin the dentry in core */
 798                 error = 0;
 799         }
 800         return error;
 801 }
 802
 803 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 804 {
 805         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
 806         if (!retval)
 807                 inc_nlink(dir);
 808         return retval;
 809 }
 810
 811 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
 812 {
 813         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
 814 }
 815
 816 static int hugetlbfs_symlink(struct inode *dir,
 817                         struct dentry *dentry, const char *symname)
 818 {
 819         struct inode *inode;
 820         int error = -ENOSPC;
 821
 822         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 823         if (inode) {
 824                 int l = strlen(symname)+1;
 825                 error = page_symlink(inode, symname, l);
 826                 if (!error) {
 827                         d_instantiate(dentry, inode);
 828                         dget(dentry);
 829                 } else
 830                         iput(inode);
 831         }
 832         dir->i_ctime = dir->i_mtime = current_time(dir);
 833
 834         return error;
 835 }
 836
 837 /*
 838  * mark the head page dirty
 839  */
 840 static int hugetlbfs_set_page_dirty(struct page *page)
 841 {
 842         struct page *head = compound_head(page);
 843
 844         SetPageDirty(head);
 845         return 0;
 846 }
 847
 848 static int hugetlbfs_migrate_page(struct address_space *mapping,
 849                                 struct page *newpage, struct page *page,
 850                                 enum migrate_mode mode)
 851 {
 852         int rc;
 853
 854         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 855         if (rc != MIGRATEPAGE_SUCCESS)
 856                 return rc;
 857         if (mode != MIGRATE_SYNC_NO_COPY)
 858                 migrate_page_copy(newpage, page);
 859         else
 860                 migrate_page_states(newpage, page);
 861
 862         return MIGRATEPAGE_SUCCESS;
 863 }
 864
 865 static int hugetlbfs_error_remove_page(struct address_space *mapping,
 866                                 struct page *page)
 867 {
 868         struct inode *inode = mapping->host;
 869         pgoff_t index = page->index;
 870
 871         remove_huge_page(page);
 872         if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
 873                 hugetlb_fix_reserve_counts(inode);
 874
 875         return 0;
 876 }
 877
 878 /*
 879  * Display the mount options in /proc/mounts.
 880  */
 881 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
 882 {
 883         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
 884         struct hugepage_subpool *spool = sbinfo->spool;
 885         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
 886         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
 887         char mod;
 888
 889         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
 890                 seq_printf(m, ",uid=%u",
 891                            from_kuid_munged(&init_user_ns, sbinfo->uid));
 892         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
 893                 seq_printf(m, ",gid=%u",
 894                            from_kgid_munged(&init_user_ns, sbinfo->gid));
 895         if (sbinfo->mode != 0755)
 896                 seq_printf(m, ",mode=%o", sbinfo->mode);
 897         if (sbinfo->max_inodes != -1)
 898                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
 899
 900         hpage_size /= 1024;
 901         mod = 'K';
 902         if (hpage_size >= 1024) {
 903                 hpage_size /= 1024;
 904                 mod = 'M';
 905         }
 906         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
 907         if (spool) {
 908                 if (spool->max_hpages != -1)
 909                         seq_printf(m, ",size=%llu",
 910                                    (unsigned long long)spool->max_hpages << hpage_shift);
 911                 if (spool->min_hpages != -1)
 912                         seq_printf(m, ",min_size=%llu",
 913                                    (unsigned long long)spool->min_hpages << hpage_shift);
 914         }
 915         return 0;
 916 }
 917
 918 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 919 {
 920         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
 921         struct hstate *h = hstate_inode(d_inode(dentry));
 922
 923         buf->f_type = HUGETLBFS_MAGIC;
 924         buf->f_bsize = huge_page_size(h);
 925         if (sbinfo) {
 926                 spin_lock(&sbinfo->stat_lock);
 927                 /* If no limits set, just report 0 for max/free/used
 928                  * blocks, like simple_statfs() */
 929                 if (sbinfo->spool) {
 930                         long free_pages;
 931
 932                         spin_lock(&sbinfo->spool->lock);
 933                         buf->f_blocks = sbinfo->spool->max_hpages;
 934                         free_pages = sbinfo->spool->max_hpages
 935                                 - sbinfo->spool->used_hpages;
 936                         buf->f_bavail = buf->f_bfree = free_pages;
 937                         spin_unlock(&sbinfo->spool->lock);
 938                         buf->f_files = sbinfo->max_inodes;
 939                         buf->f_ffree = sbinfo->free_inodes;
 940                 }
 941                 spin_unlock(&sbinfo->stat_lock);
 942         }
 943         buf->f_namelen = NAME_MAX;
 944         return 0;
 945 }
 946
 947 static void hugetlbfs_put_super(struct super_block *sb)
 948 {
 949         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
 950
 951         if (sbi) {
 952                 sb->s_fs_info = NULL;
 953
 954                 if (sbi->spool)
 955                         hugepage_put_subpool(sbi->spool);
 956
 957                 kfree(sbi);
 958         }
 959 }
 960
 961 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 962 {
 963         if (sbinfo->free_inodes >= 0) {
 964                 spin_lock(&sbinfo->stat_lock);
 965                 if (unlikely(!sbinfo->free_inodes)) {
 966                         spin_unlock(&sbinfo->stat_lock);
 967                         return 0;
 968                 }
 969                 sbinfo->free_inodes--;
 970                 spin_unlock(&sbinfo->stat_lock);
 971         }
 972
 973         return 1;
 974 }
 975
 976 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 977 {
 978         if (sbinfo->free_inodes >= 0) {
 979                 spin_lock(&sbinfo->stat_lock);
 980                 sbinfo->free_inodes++;
 981                 spin_unlock(&sbinfo->stat_lock);
 982         }
 983 }
 984
 985
 986 static struct kmem_cache *hugetlbfs_inode_cachep;
 987
 988 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
 989 {
 990         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
 991         struct hugetlbfs_inode_info *p;
 992
 993         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
 994                 return NULL;
 995         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
 996         if (unlikely(!p)) {
 997                 hugetlbfs_inc_free_inodes(sbinfo);
 998                 return NULL;
 999         }
1000
1001         /*
1002          * Any time after allocation, hugetlbfs_destroy_inode can be called
1003          * for the inode.  mpol_free_shared_policy is unconditionally called
1004          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1005          * in case of a quick call to destroy.
1006          *
1007          * Note that the policy is initialized even if we are creating a
1008          * private inode.  This simplifies hugetlbfs_destroy_inode.
1009          */
1010         mpol_shared_policy_init(&p->policy, NULL);
1011
1012         return &p->vfs_inode;
1013 }
1014
1015 static void hugetlbfs_i_callback(struct rcu_head *head)
1016 {
1017         struct inode *inode = container_of(head, struct inode, i_rcu);
1018         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1019 }
1020
1021 static void hugetlbfs_destroy_inode(struct inode *inode)
1022 {
1023         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1024         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1025         call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
1026 }
1027
1028 static const struct address_space_operations hugetlbfs_aops = {
1029         .write_begin    = hugetlbfs_write_begin,
1030         .write_end      = hugetlbfs_write_end,
1031         .set_page_dirty = hugetlbfs_set_page_dirty,
1032         .migratepage    = hugetlbfs_migrate_page,
1033         .error_remove_page      = hugetlbfs_error_remove_page,
1034 };
1035
1036
1037 static void init_once(void *foo)
1038 {
1039         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1040
1041         inode_init_once(&ei->vfs_inode);
1042 }
1043
1044 const struct file_operations hugetlbfs_file_operations = {
1045         .read_iter              = hugetlbfs_read_iter,
1046         .mmap                   = hugetlbfs_file_mmap,
1047         .fsync                  = noop_fsync,
1048         .get_unmapped_area      = hugetlb_get_unmapped_area,
1049         .llseek                 = default_llseek,
1050         .fallocate              = hugetlbfs_fallocate,
1051 };
1052
1053 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1054         .create         = hugetlbfs_create,
1055         .lookup         = simple_lookup,
1056         .link           = simple_link,
1057         .unlink         = simple_unlink,
1058         .symlink        = hugetlbfs_symlink,
1059         .mkdir          = hugetlbfs_mkdir,
1060         .rmdir          = simple_rmdir,
1061         .mknod          = hugetlbfs_mknod,
1062         .rename         = simple_rename,
1063         .setattr        = hugetlbfs_setattr,
1064 };
1065
1066 static const struct inode_operations hugetlbfs_inode_operations = {
1067         .setattr        = hugetlbfs_setattr,
1068 };
1069
1070 static const struct super_operations hugetlbfs_ops = {
1071         .alloc_inode    = hugetlbfs_alloc_inode,
1072         .destroy_inode  = hugetlbfs_destroy_inode,
1073         .evict_inode    = hugetlbfs_evict_inode,
1074         .statfs         = hugetlbfs_statfs,
1075         .put_super      = hugetlbfs_put_super,
1076         .show_options   = hugetlbfs_show_options,
1077 };
1078
1079 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1080
1081 /*
1082  * Convert size option passed from command line to number of huge pages
1083  * in the pool specified by hstate.  Size option could be in bytes
1084  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1085  */
1086 static long
1087 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1088                          enum hugetlbfs_size_type val_type)
1089 {
1090         if (val_type == NO_SIZE)
1091                 return -1;
1092
1093         if (val_type == SIZE_PERCENT) {
1094                 size_opt <<= huge_page_shift(h);
1095                 size_opt *= h->max_huge_pages;
1096                 do_div(size_opt, 100);
1097         }
1098
1099         size_opt >>= huge_page_shift(h);
1100         return size_opt;
1101 }
1102
1103 static int
1104 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1105 {
1106         char *p, *rest;
1107         substring_t args[MAX_OPT_ARGS];
1108         int option;
1109         unsigned long long max_size_opt = 0, min_size_opt = 0;
1110         enum hugetlbfs_size_type max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1111
1112         if (!options)
1113                 return 0;
1114
1115         while ((p = strsep(&options, ",")) != NULL) {
1116                 int token;
1117                 if (!*p)
1118                         continue;
1119
1120                 token = match_token(p, tokens, args);
1121                 switch (token) {
1122                 case Opt_uid:
1123                         if (match_int(&args[0], &option))
1124                                 goto bad_val;
1125                         pconfig->uid = make_kuid(current_user_ns(), option);
1126                         if (!uid_valid(pconfig->uid))
1127                                 goto bad_val;
1128                         break;
1129
1130                 case Opt_gid:
1131                         if (match_int(&args[0], &option))
1132                                 goto bad_val;
1133                         pconfig->gid = make_kgid(current_user_ns(), option);
1134                         if (!gid_valid(pconfig->gid))
1135                                 goto bad_val;
1136                         break;
1137
1138                 case Opt_mode:
1139                         if (match_octal(&args[0], &option))
1140                                 goto bad_val;
1141                         pconfig->mode = option & 01777U;
1142                         break;
1143
1144                 case Opt_size: {
1145                         /* memparse() will accept a K/M/G without a digit */
1146                         if (!isdigit(*args[0].from))
1147                                 goto bad_val;
1148                         max_size_opt = memparse(args[0].from, &rest);
1149                         max_val_type = SIZE_STD;
1150                         if (*rest == '%')
1151                                 max_val_type = SIZE_PERCENT;
1152                         break;
1153                 }
1154
1155                 case Opt_nr_inodes:
1156                         /* memparse() will accept a K/M/G without a digit */
1157                         if (!isdigit(*args[0].from))
1158                                 goto bad_val;
1159                         pconfig->nr_inodes = memparse(args[0].from, &rest);
1160                         break;
1161
1162                 case Opt_pagesize: {
1163                         unsigned long ps;
1164                         ps = memparse(args[0].from, &rest);
1165                         pconfig->hstate = size_to_hstate(ps);
1166                         if (!pconfig->hstate) {
1167                                 pr_err("Unsupported page size %lu MB\n",
1168                                         ps >> 20);
1169                                 return -EINVAL;
1170                         }
1171                         break;
1172                 }
1173
1174                 case Opt_min_size: {
1175                         /* memparse() will accept a K/M/G without a digit */
1176                         if (!isdigit(*args[0].from))
1177                                 goto bad_val;
1178                         min_size_opt = memparse(args[0].from, &rest);
1179                         min_val_type = SIZE_STD;
1180                         if (*rest == '%')
1181                                 min_val_type = SIZE_PERCENT;
1182                         break;
1183                 }
1184
1185                 default:
1186                         pr_err("Bad mount option: \"%s\"\n", p);
1187                         return -EINVAL;
1188                         break;
1189                 }
1190         }
1191
1192         /*
1193          * Use huge page pool size (in hstate) to convert the size
1194          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1195          */
1196         pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1197                                                 max_size_opt, max_val_type);
1198         pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1199                                                 min_size_opt, min_val_type);
1200
1201         /*
1202          * If max_size was specified, then min_size must be smaller
1203          */
1204         if (max_val_type > NO_SIZE &&
1205             pconfig->min_hpages > pconfig->max_hpages) {
1206                 pr_err("minimum size can not be greater than maximum size\n");
1207                 return -EINVAL;
1208         }
1209
1210         return 0;
1211
1212 bad_val:
1213         pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1214         return -EINVAL;
1215 }
1216
1217 static int
1218 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1219 {
1220         int ret;
1221         struct hugetlbfs_config config;
1222         struct hugetlbfs_sb_info *sbinfo;
1223
1224         config.max_hpages = -1; /* No limit on size by default */
1225         config.nr_inodes = -1; /* No limit on number of inodes by default */
1226         config.uid = current_fsuid();
1227         config.gid = current_fsgid();
1228         config.mode = 0755;
1229         config.hstate = &default_hstate;
1230         config.min_hpages = -1; /* No default minimum size */
1231         ret = hugetlbfs_parse_options(data, &config);
1232         if (ret)
1233                 return ret;
1234
1235         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1236         if (!sbinfo)
1237                 return -ENOMEM;
1238         sb->s_fs_info = sbinfo;
1239         sbinfo->hstate = config.hstate;
1240         spin_lock_init(&sbinfo->stat_lock);
1241         sbinfo->max_inodes = config.nr_inodes;
1242         sbinfo->free_inodes = config.nr_inodes;
1243         sbinfo->spool = NULL;
1244         sbinfo->uid = config.uid;
1245         sbinfo->gid = config.gid;
1246         sbinfo->mode = config.mode;
1247
1248         /*
1249          * Allocate and initialize subpool if maximum or minimum size is
1250          * specified.  Any needed reservations (for minimim size) are taken
1251          * taken when the subpool is created.
1252          */
1253         if (config.max_hpages != -1 || config.min_hpages != -1) {
1254                 sbinfo->spool = hugepage_new_subpool(config.hstate,
1255                                                         config.max_hpages,
1256                                                         config.min_hpages);
1257                 if (!sbinfo->spool)
1258                         goto out_free;
1259         }
1260         sb->s_maxbytes = MAX_LFS_FILESIZE;
1261         sb->s_blocksize = huge_page_size(config.hstate);
1262         sb->s_blocksize_bits = huge_page_shift(config.hstate);
1263         sb->s_magic = HUGETLBFS_MAGIC;
1264         sb->s_op = &hugetlbfs_ops;
1265         sb->s_time_gran = 1;
1266         sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1267         if (!sb->s_root)
1268                 goto out_free;
1269         return 0;
1270 out_free:
1271         kfree(sbinfo->spool);
1272         kfree(sbinfo);
1273         return -ENOMEM;
1274 }
1275
1276 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1277         int flags, const char *dev_name, void *data)
1278 {
1279         return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1280 }
1281
1282 static struct file_system_type hugetlbfs_fs_type = {
1283         .name           = "hugetlbfs",
1284         .mount          = hugetlbfs_mount,
1285         .kill_sb        = kill_litter_super,
1286 };
1287
1288 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1289
1290 static int can_do_hugetlb_shm(void)
1291 {
1292         kgid_t shm_group;
1293         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1294         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1295 }
1296
1297 static int get_hstate_idx(int page_size_log)
1298 {
1299         struct hstate *h = hstate_sizelog(page_size_log);
1300
1301         if (!h)
1302                 return -1;
1303         return h - hstates;
1304 }
1305
1306 /*
1307  * Note that size should be aligned to proper hugepage size in caller side,
1308  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1309  */
1310 struct file *hugetlb_file_setup(const char *name, size_t size,
1311                                 vm_flags_t acctflag, struct user_struct **user,
1312                                 int creat_flags, int page_size_log)
1313 {
1314         struct inode *inode;
1315         struct vfsmount *mnt;
1316         int hstate_idx;
1317         struct file *file;
1318
1319         hstate_idx = get_hstate_idx(page_size_log);
1320         if (hstate_idx < 0)
1321                 return ERR_PTR(-ENODEV);
1322
1323         *user = NULL;
1324         mnt = hugetlbfs_vfsmount[hstate_idx];
1325         if (!mnt)
1326                 return ERR_PTR(-ENOENT);
1327
1328         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1329                 *user = current_user();
1330                 if (user_shm_lock(size, *user)) {
1331                         task_lock(current);
1332                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1333                                 current->comm, current->pid);
1334                         task_unlock(current);
1335                 } else {
1336                         *user = NULL;
1337                         return ERR_PTR(-EPERM);
1338                 }
1339         }
1340
1341         file = ERR_PTR(-ENOSPC);
1342         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1343         if (!inode)
1344                 goto out;
1345         if (creat_flags == HUGETLB_SHMFS_INODE)
1346                 inode->i_flags |= S_PRIVATE;
1347
1348         inode->i_size = size;
1349         clear_nlink(inode);
1350
1351         if (hugetlb_reserve_pages(inode, 0,
1352                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1353                         acctflag))
1354                 file = ERR_PTR(-ENOMEM);
1355         else
1356                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1357                                         &hugetlbfs_file_operations);
1358         if (!IS_ERR(file))
1359                 return file;
1360
1361         iput(inode);
1362 out:
1363         if (*user) {
1364                 user_shm_unlock(size, *user);
1365                 *user = NULL;
1366         }
1367         return file;
1368 }
1369
1370 static int __init init_hugetlbfs_fs(void)
1371 {
1372         struct hstate *h;
1373         int error;
1374         int i;
1375
1376         if (!hugepages_supported()) {
1377                 pr_info("disabling because there are no supported hugepage sizes\n");
1378                 return -ENOTSUPP;
1379         }
1380
1381         error = -ENOMEM;
1382         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1383                                         sizeof(struct hugetlbfs_inode_info),
1384                                         0, SLAB_ACCOUNT, init_once);
1385         if (hugetlbfs_inode_cachep == NULL)
1386                 goto out2;
1387
1388         error = register_filesystem(&hugetlbfs_fs_type);
1389         if (error)
1390                 goto out;
1391
1392         i = 0;
1393         for_each_hstate(h) {
1394                 char buf[50];
1395                 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1396
1397                 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1398                 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1399                                                         buf);
1400
1401                 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1402                         pr_err("Cannot mount internal hugetlbfs for "
1403                                 "page size %uK", ps_kb);
1404                         error = PTR_ERR(hugetlbfs_vfsmount[i]);
1405                         hugetlbfs_vfsmount[i] = NULL;
1406                 }
1407                 i++;
1408         }
1409         /* Non default hstates are optional */
1410         if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1411                 return 0;
1412
1413  out:
1414         kmem_cache_destroy(hugetlbfs_inode_cachep);
1415  out2:
1416         return error;
1417 }
1418 fs_initcall(init_hugetlbfs_fs)