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 /*
 122  * Mask used when checking the page offset value passed in via system
 123  * calls.  This value will be converted to a loff_t which is signed.
 124  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 125  * value.  The extra bit (- 1 in the shift value) is to take the sign
 126  * bit into account.
 127  */
 128 #define PGOFF_LOFFT_MAX \
 129         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 130
 131 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 132 {
 133         struct inode *inode = file_inode(file);
 134         loff_t len, vma_len;
 135         int ret;
 136         struct hstate *h = hstate_file(file);
 137
 138         /*
 139          * vma address alignment (but not the pgoff alignment) has
 140          * already been checked by prepare_hugepage_range.  If you add
 141          * any error returns here, do so after setting VM_HUGETLB, so
 142          * is_vm_hugetlb_page tests below unmap_region go the right
 143          * way when do_mmap_pgoff unwinds (may be important on powerpc
 144          * and ia64).
 145          */
 146         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 147         vma->vm_ops = &hugetlb_vm_ops;
 148
 149         /*
 150          * page based offset in vm_pgoff could be sufficiently large to
 151          * overflow a loff_t when converted to byte offset.  This can
 152          * only happen on architectures where sizeof(loff_t) ==
 153          * sizeof(unsigned long).  So, only check in those instances.
 154          */
 155         if (sizeof(unsigned long) == sizeof(loff_t)) {
 156                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 157                         return -EINVAL;
 158         }
 159
 160         /* must be huge page aligned */
 161         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 162                 return -EINVAL;
 163
 164         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 165         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 166         /* check for overflow */
 167         if (len < vma_len)
 168                 return -EINVAL;
 169
 170         mutex_lock(&inode->i_mutex);
 171         file_accessed(file);
 172
 173         ret = -ENOMEM;
 174         if (hugetlb_reserve_pages(inode,
 175                                 vma->vm_pgoff >> huge_page_order(h),
 176                                 len >> huge_page_shift(h), vma,
 177                                 vma->vm_flags))
 178                 goto out;
 179
 180         ret = 0;
 181         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 182                 i_size_write(inode, len);
 183 out:
 184         mutex_unlock(&inode->i_mutex);
 185
 186         return ret;
 187 }
 188
 189 /*
 190  * Called under down_write(mmap_sem).
 191  */
 192
 193 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 194 static unsigned long
 195 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 196                 unsigned long len, unsigned long pgoff, unsigned long flags)
 197 {
 198         struct mm_struct *mm = current->mm;
 199         struct vm_area_struct *vma;
 200         struct hstate *h = hstate_file(file);
 201         struct vm_unmapped_area_info info;
 202
 203         if (len & ~huge_page_mask(h))
 204                 return -EINVAL;
 205         if (len > TASK_SIZE)
 206                 return -ENOMEM;
 207
 208         if (flags & MAP_FIXED) {
 209                 if (prepare_hugepage_range(file, addr, len))
 210                         return -EINVAL;
 211                 return addr;
 212         }
 213
 214         if (addr) {
 215                 addr = ALIGN(addr, huge_page_size(h));
 216                 vma = find_vma(mm, addr);
 217                 if (TASK_SIZE - len >= addr &&
 218                     (!vma || addr + len <= vm_start_gap(vma)))
 219                         return addr;
 220         }
 221
 222         info.flags = 0;
 223         info.length = len;
 224         info.low_limit = TASK_UNMAPPED_BASE;
 225         info.high_limit = TASK_SIZE;
 226         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 227         info.align_offset = 0;
 228         return vm_unmapped_area(&info);
 229 }
 230 #endif
 231
 232 static size_t
 233 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 234                         struct iov_iter *to, unsigned long size)
 235 {
 236         size_t copied = 0;
 237         int i, chunksize;
 238
 239         /* Find which 4k chunk and offset with in that chunk */
 240         i = offset >> PAGE_CACHE_SHIFT;
 241         offset = offset & ~PAGE_CACHE_MASK;
 242
 243         while (size) {
 244                 size_t n;
 245                 chunksize = PAGE_CACHE_SIZE;
 246                 if (offset)
 247                         chunksize -= offset;
 248                 if (chunksize > size)
 249                         chunksize = size;
 250                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 251                 copied += n;
 252                 if (n != chunksize)
 253                         return copied;
 254                 offset = 0;
 255                 size -= chunksize;
 256                 i++;
 257         }
 258         return copied;
 259 }
 260
 261 /*
 262  * Support for read() - Find the page attached to f_mapping and copy out the
 263  * data. Its *very* similar to do_generic_mapping_read(), we can't use that
 264  * since it has PAGE_CACHE_SIZE assumptions.
 265  */
 266 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 267 {
 268         struct file *file = iocb->ki_filp;
 269         struct hstate *h = hstate_file(file);
 270         struct address_space *mapping = file->f_mapping;
 271         struct inode *inode = mapping->host;
 272         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 273         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 274         unsigned long end_index;
 275         loff_t isize;
 276         ssize_t retval = 0;
 277
 278         while (iov_iter_count(to)) {
 279                 struct page *page;
 280                 size_t nr, copied;
 281
 282                 /* nr is the maximum number of bytes to copy from this page */
 283                 nr = huge_page_size(h);
 284                 isize = i_size_read(inode);
 285                 if (!isize)
 286                         break;
 287                 end_index = (isize - 1) >> huge_page_shift(h);
 288                 if (index > end_index)
 289                         break;
 290                 if (index == end_index) {
 291                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 292                         if (nr <= offset)
 293                                 break;
 294                 }
 295                 nr = nr - offset;
 296
 297                 /* Find the page */
 298                 page = find_lock_page(mapping, index);
 299                 if (unlikely(page == NULL)) {
 300                         /*
 301                          * We have a HOLE, zero out the user-buffer for the
 302                          * length of the hole or request.
 303                          */
 304                         copied = iov_iter_zero(nr, to);
 305                 } else {
 306                         unlock_page(page);
 307
 308                         /*
 309                          * We have the page, copy it to user space buffer.
 310                          */
 311                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 312                         page_cache_release(page);
 313                 }
 314                 offset += copied;
 315                 retval += copied;
 316                 if (copied != nr && iov_iter_count(to)) {
 317                         if (!retval)
 318                                 retval = -EFAULT;
 319                         break;
 320                 }
 321                 index += offset >> huge_page_shift(h);
 322                 offset &= ~huge_page_mask(h);
 323         }
 324         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 325         return retval;
 326 }
 327
 328 static int hugetlbfs_write_begin(struct file *file,
 329                         struct address_space *mapping,
 330                         loff_t pos, unsigned len, unsigned flags,
 331                         struct page **pagep, void **fsdata)
 332 {
 333         return -EINVAL;
 334 }
 335
 336 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 337                         loff_t pos, unsigned len, unsigned copied,
 338                         struct page *page, void *fsdata)
 339 {
 340         BUG();
 341         return -EINVAL;
 342 }
 343
 344 static void remove_huge_page(struct page *page)
 345 {
 346         ClearPageDirty(page);
 347         ClearPageUptodate(page);
 348         delete_from_page_cache(page);
 349 }
 350
 351
 352 /*
 353  * remove_inode_hugepages handles two distinct cases: truncation and hole
 354  * punch.  There are subtle differences in operation for each case.
 355
 356  * truncation is indicated by end of range being LLONG_MAX
 357  *      In this case, we first scan the range and release found pages.
 358  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
 359  *      maps and global counts.  Page faults can not race with truncation
 360  *      in this routine.  hugetlb_no_page() prevents page faults in the
 361  *      truncated range.  It checks i_size before allocation, and again after
 362  *      with the page table lock for the page held.  The same lock must be
 363  *      acquired to unmap a page.
 364  * hole punch is indicated if end is not LLONG_MAX
 365  *      In the hole punch case we scan the range and release found pages.
 366  *      Only when releasing a page is the associated region/reserv map
 367  *      deleted.  The region/reserv map for ranges without associated
 368  *      pages are not modified.  Page faults can race with hole punch.
 369  *      This is indicated if we find a mapped page.
 370  * Note: If the passed end of range value is beyond the end of file, but
 371  * not LLONG_MAX this routine still performs a hole punch operation.
 372  */
 373 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 374                                    loff_t lend)
 375 {
 376         struct hstate *h = hstate_inode(inode);
 377         struct address_space *mapping = &inode->i_data;
 378         const pgoff_t start = lstart >> huge_page_shift(h);
 379         const pgoff_t end = lend >> huge_page_shift(h);
 380         struct vm_area_struct pseudo_vma;
 381         struct pagevec pvec;
 382         pgoff_t next;
 383         int i, freed = 0;
 384         long lookup_nr = PAGEVEC_SIZE;
 385         bool truncate_op = (lend == LLONG_MAX);
 386
 387         memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
 388         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 389         pagevec_init(&pvec, 0);
 390         next = start;
 391         while (next < end) {
 392                 /*
 393                  * Don't grab more pages than the number left in the range.
 394                  */
 395                 if (end - next < lookup_nr)
 396                         lookup_nr = end - next;
 397
 398                 /*
 399                  * When no more pages are found, we are done.
 400                  */
 401                 if (!pagevec_lookup(&pvec, mapping, next, lookup_nr))
 402                         break;
 403
 404                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 405                         struct page *page = pvec.pages[i];
 406                         u32 hash;
 407
 408                         /*
 409                          * The page (index) could be beyond end.  This is
 410                          * only possible in the punch hole case as end is
 411                          * max page offset in the truncate case.
 412                          */
 413                         next = page->index;
 414                         if (next >= end)
 415                                 break;
 416
 417                         hash = hugetlb_fault_mutex_hash(h, mapping, next);
 418                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 419
 420                         lock_page(page);
 421                         if (likely(!page_mapped(page))) {
 422                                 bool rsv_on_error = !PagePrivate(page);
 423                                 /*
 424                                  * We must free the huge page and remove
 425                                  * from page cache (remove_huge_page) BEFORE
 426                                  * removing the region/reserve map
 427                                  * (hugetlb_unreserve_pages).  In rare out
 428                                  * of memory conditions, removal of the
 429                                  * region/reserve map could fail.  Before
 430                                  * free'ing the page, note PagePrivate which
 431                                  * is used in case of error.
 432                                  */
 433                                 remove_huge_page(page);
 434                                 freed++;
 435                                 if (!truncate_op) {
 436                                         if (unlikely(hugetlb_unreserve_pages(
 437                                                         inode, next,
 438                                                         next + 1, 1)))
 439                                                 hugetlb_fix_reserve_counts(
 440                                                         inode, rsv_on_error);
 441                                 }
 442                         } else {
 443                                 /*
 444                                  * If page is mapped, it was faulted in after
 445                                  * being unmapped.  It indicates a race between
 446                                  * hole punch and page fault.  Do nothing in
 447                                  * this case.  Getting here in a truncate
 448                                  * operation is a bug.
 449                                  */
 450                                 BUG_ON(truncate_op);
 451                         }
 452
 453                         unlock_page(page);
 454                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 455                 }
 456                 ++next;
 457                 huge_pagevec_release(&pvec);
 458                 cond_resched();
 459         }
 460
 461         if (truncate_op)
 462                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 463 }
 464
 465 static void hugetlbfs_evict_inode(struct inode *inode)
 466 {
 467         struct resv_map *resv_map;
 468
 469         remove_inode_hugepages(inode, 0, LLONG_MAX);
 470         resv_map = (struct resv_map *)inode->i_mapping->private_data;
 471         /* root inode doesn't have the resv_map, so we should check it */
 472         if (resv_map)
 473                 resv_map_release(&resv_map->refs);
 474         clear_inode(inode);
 475 }
 476
 477 static inline void
 478 hugetlb_vmdelete_list(struct rb_root *root, pgoff_t start, pgoff_t end)
 479 {
 480         struct vm_area_struct *vma;
 481
 482         /*
 483          * end == 0 indicates that the entire range after
 484          * start should be unmapped.
 485          */
 486         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
 487                 unsigned long v_offset;
 488                 unsigned long v_end;
 489
 490                 /*
 491                  * Can the expression below overflow on 32-bit arches?
 492                  * No, because the interval tree returns us only those vmas
 493                  * which overlap the truncated area starting at pgoff,
 494                  * and no vma on a 32-bit arch can span beyond the 4GB.
 495                  */
 496                 if (vma->vm_pgoff < start)
 497                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 498                 else
 499                         v_offset = 0;
 500
 501                 if (!end)
 502                         v_end = vma->vm_end;
 503                 else {
 504                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 505                                                         + vma->vm_start;
 506                         if (v_end > vma->vm_end)
 507                                 v_end = vma->vm_end;
 508                 }
 509
 510                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 511                                                                         NULL);
 512         }
 513 }
 514
 515 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 516 {
 517         pgoff_t pgoff;
 518         struct address_space *mapping = inode->i_mapping;
 519         struct hstate *h = hstate_inode(inode);
 520
 521         BUG_ON(offset & ~huge_page_mask(h));
 522         pgoff = offset >> PAGE_SHIFT;
 523
 524         i_size_write(inode, offset);
 525         i_mmap_lock_write(mapping);
 526         if (!RB_EMPTY_ROOT(&mapping->i_mmap))
 527                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 528         i_mmap_unlock_write(mapping);
 529         remove_inode_hugepages(inode, offset, LLONG_MAX);
 530         return 0;
 531 }
 532
 533 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 534 {
 535         struct hstate *h = hstate_inode(inode);
 536         loff_t hpage_size = huge_page_size(h);
 537         loff_t hole_start, hole_end;
 538
 539         /*
 540          * For hole punch round up the beginning offset of the hole and
 541          * round down the end.
 542          */
 543         hole_start = round_up(offset, hpage_size);
 544         hole_end = round_down(offset + len, hpage_size);
 545
 546         if (hole_end > hole_start) {
 547                 struct address_space *mapping = inode->i_mapping;
 548
 549                 mutex_lock(&inode->i_mutex);
 550                 i_mmap_lock_write(mapping);
 551                 if (!RB_EMPTY_ROOT(&mapping->i_mmap))
 552                         hugetlb_vmdelete_list(&mapping->i_mmap,
 553                                                 hole_start >> PAGE_SHIFT,
 554                                                 hole_end  >> PAGE_SHIFT);
 555                 i_mmap_unlock_write(mapping);
 556                 remove_inode_hugepages(inode, hole_start, hole_end);
 557                 mutex_unlock(&inode->i_mutex);
 558         }
 559
 560         return 0;
 561 }
 562
 563 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 564                                 loff_t len)
 565 {
 566         struct inode *inode = file_inode(file);
 567         struct address_space *mapping = inode->i_mapping;
 568         struct hstate *h = hstate_inode(inode);
 569         struct vm_area_struct pseudo_vma;
 570         loff_t hpage_size = huge_page_size(h);
 571         unsigned long hpage_shift = huge_page_shift(h);
 572         pgoff_t start, index, end;
 573         int error;
 574         u32 hash;
 575
 576         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 577                 return -EOPNOTSUPP;
 578
 579         if (mode & FALLOC_FL_PUNCH_HOLE)
 580                 return hugetlbfs_punch_hole(inode, offset, len);
 581
 582         /*
 583          * Default preallocate case.
 584          * For this range, start is rounded down and end is rounded up
 585          * as well as being converted to page offsets.
 586          */
 587         start = offset >> hpage_shift;
 588         end = (offset + len + hpage_size - 1) >> hpage_shift;
 589
 590         mutex_lock(&inode->i_mutex);
 591
 592         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 593         error = inode_newsize_ok(inode, offset + len);
 594         if (error)
 595                 goto out;
 596
 597         /*
 598          * Initialize a pseudo vma as this is required by the huge page
 599          * allocation routines.  If NUMA is configured, use page index
 600          * as input to create an allocation policy.
 601          */
 602         memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
 603         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 604         pseudo_vma.vm_file = file;
 605
 606         for (index = start; index < end; index++) {
 607                 /*
 608                  * This is supposed to be the vaddr where the page is being
 609                  * faulted in, but we have no vaddr here.
 610                  */
 611                 struct page *page;
 612                 unsigned long addr;
 613                 int avoid_reserve = 0;
 614
 615                 cond_resched();
 616
 617                 /*
 618                  * fallocate(2) manpage permits EINTR; we may have been
 619                  * interrupted because we are using up too much memory.
 620                  */
 621                 if (signal_pending(current)) {
 622                         error = -EINTR;
 623                         break;
 624                 }
 625
 626                 /* Set numa allocation policy based on index */
 627                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 628
 629                 /* addr is the offset within the file (zero based) */
 630                 addr = index * hpage_size;
 631
 632                 /* mutex taken here, fault path and hole punch */
 633                 hash = hugetlb_fault_mutex_hash(h, mapping, index);
 634                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 635
 636                 /* See if already present in mapping to avoid alloc/free */
 637                 page = find_get_page(mapping, index);
 638                 if (page) {
 639                         put_page(page);
 640                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 641                         hugetlb_drop_vma_policy(&pseudo_vma);
 642                         continue;
 643                 }
 644
 645                 /* Allocate page and add to page cache */
 646                 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
 647                 hugetlb_drop_vma_policy(&pseudo_vma);
 648                 if (IS_ERR(page)) {
 649                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 650                         error = PTR_ERR(page);
 651                         goto out;
 652                 }
 653                 clear_huge_page(page, addr, pages_per_huge_page(h));
 654                 __SetPageUptodate(page);
 655                 error = huge_add_to_page_cache(page, mapping, index);
 656                 if (unlikely(error)) {
 657                         put_page(page);
 658                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 659                         goto out;
 660                 }
 661
 662                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 663
 664                 set_page_huge_active(page);
 665                 /*
 666                  * put_page() due to reference from alloc_huge_page()
 667                  * unlock_page because locked by add_to_page_cache()
 668                  */
 669                 put_page(page);
 670                 unlock_page(page);
 671         }
 672
 673         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 674                 i_size_write(inode, offset + len);
 675         inode->i_ctime = CURRENT_TIME;
 676 out:
 677         mutex_unlock(&inode->i_mutex);
 678         return error;
 679 }
 680
 681 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 682 {
 683         struct inode *inode = d_inode(dentry);
 684         struct hstate *h = hstate_inode(inode);
 685         int error;
 686         unsigned int ia_valid = attr->ia_valid;
 687
 688         BUG_ON(!inode);
 689
 690         error = inode_change_ok(inode, attr);
 691         if (error)
 692                 return error;
 693
 694         if (ia_valid & ATTR_SIZE) {
 695                 error = -EINVAL;
 696                 if (attr->ia_size & ~huge_page_mask(h))
 697                         return -EINVAL;
 698                 error = hugetlb_vmtruncate(inode, attr->ia_size);
 699                 if (error)
 700                         return error;
 701         }
 702
 703         setattr_copy(inode, attr);
 704         mark_inode_dirty(inode);
 705         return 0;
 706 }
 707
 708 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 709                                         struct hugetlbfs_config *config)
 710 {
 711         struct inode *inode;
 712
 713         inode = new_inode(sb);
 714         if (inode) {
 715                 struct hugetlbfs_inode_info *info;
 716                 inode->i_ino = get_next_ino();
 717                 inode->i_mode = S_IFDIR | config->mode;
 718                 inode->i_uid = config->uid;
 719                 inode->i_gid = config->gid;
 720                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 721                 info = HUGETLBFS_I(inode);
 722                 mpol_shared_policy_init(&info->policy, NULL);
 723                 inode->i_op = &hugetlbfs_dir_inode_operations;
 724                 inode->i_fop = &simple_dir_operations;
 725                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 726                 inc_nlink(inode);
 727                 lockdep_annotate_inode_mutex_key(inode);
 728         }
 729         return inode;
 730 }
 731
 732 /*
 733  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 734  * be taken from reclaim -- unlike regular filesystems. This needs an
 735  * annotation because huge_pmd_share() does an allocation under
 736  * i_mmap_rwsem.
 737  */
 738 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 739
 740 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 741                                         struct inode *dir,
 742                                         umode_t mode, dev_t dev)
 743 {
 744         struct inode *inode;
 745         struct resv_map *resv_map = NULL;
 746
 747         /*
 748          * Reserve maps are only needed for inodes that can have associated
 749          * page allocations.
 750          */
 751         if (S_ISREG(mode) || S_ISLNK(mode)) {
 752                 resv_map = resv_map_alloc();
 753                 if (!resv_map)
 754                         return NULL;
 755         }
 756
 757         inode = new_inode(sb);
 758         if (inode) {
 759                 struct hugetlbfs_inode_info *info;
 760                 inode->i_ino = get_next_ino();
 761                 inode_init_owner(inode, dir, mode);
 762                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 763                                 &hugetlbfs_i_mmap_rwsem_key);
 764                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 765                 inode->i_atime = inode->i_mtime = inode->i_ctime = CURRENT_TIME;
 766                 inode->i_mapping->private_data = resv_map;
 767                 info = HUGETLBFS_I(inode);
 768                 /*
 769                  * The policy is initialized here even if we are creating a
 770                  * private inode because initialization simply creates an
 771                  * an empty rb tree and calls spin_lock_init(), later when we
 772                  * call mpol_free_shared_policy() it will just return because
 773                  * the rb tree will still be empty.
 774                  */
 775                 mpol_shared_policy_init(&info->policy, NULL);
 776                 switch (mode & S_IFMT) {
 777                 default:
 778                         init_special_inode(inode, mode, dev);
 779                         break;
 780                 case S_IFREG:
 781                         inode->i_op = &hugetlbfs_inode_operations;
 782                         inode->i_fop = &hugetlbfs_file_operations;
 783                         break;
 784                 case S_IFDIR:
 785                         inode->i_op = &hugetlbfs_dir_inode_operations;
 786                         inode->i_fop = &simple_dir_operations;
 787
 788                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 789                         inc_nlink(inode);
 790                         break;
 791                 case S_IFLNK:
 792                         inode->i_op = &page_symlink_inode_operations;
 793                         break;
 794                 }
 795                 lockdep_annotate_inode_mutex_key(inode);
 796         } else {
 797                 if (resv_map)
 798                         kref_put(&resv_map->refs, resv_map_release);
 799         }
 800
 801         return inode;
 802 }
 803
 804 /*
 805  * File creation. Allocate an inode, and we're done..
 806  */
 807 static int hugetlbfs_mknod(struct inode *dir,
 808                         struct dentry *dentry, umode_t mode, dev_t dev)
 809 {
 810         struct inode *inode;
 811         int error = -ENOSPC;
 812
 813         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 814         if (inode) {
 815                 dir->i_ctime = dir->i_mtime = CURRENT_TIME;
 816                 d_instantiate(dentry, inode);
 817                 dget(dentry);   /* Extra count - pin the dentry in core */
 818                 error = 0;
 819         }
 820         return error;
 821 }
 822
 823 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 824 {
 825         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
 826         if (!retval)
 827                 inc_nlink(dir);
 828         return retval;
 829 }
 830
 831 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
 832 {
 833         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
 834 }
 835
 836 static int hugetlbfs_symlink(struct inode *dir,
 837                         struct dentry *dentry, const char *symname)
 838 {
 839         struct inode *inode;
 840         int error = -ENOSPC;
 841
 842         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 843         if (inode) {
 844                 int l = strlen(symname)+1;
 845                 error = page_symlink(inode, symname, l);
 846                 if (!error) {
 847                         d_instantiate(dentry, inode);
 848                         dget(dentry);
 849                 } else
 850                         iput(inode);
 851         }
 852         dir->i_ctime = dir->i_mtime = CURRENT_TIME;
 853
 854         return error;
 855 }
 856
 857 /*
 858  * mark the head page dirty
 859  */
 860 static int hugetlbfs_set_page_dirty(struct page *page)
 861 {
 862         struct page *head = compound_head(page);
 863
 864         SetPageDirty(head);
 865         return 0;
 866 }
 867
 868 static int hugetlbfs_migrate_page(struct address_space *mapping,
 869                                 struct page *newpage, struct page *page,
 870                                 enum migrate_mode mode)
 871 {
 872         int rc;
 873
 874         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 875         if (rc != MIGRATEPAGE_SUCCESS)
 876                 return rc;
 877
 878         /*
 879          * page_private is subpool pointer in hugetlb pages.  Transfer to
 880          * new page.  PagePrivate is not associated with page_private for
 881          * hugetlb pages and can not be set here as only page_huge_active
 882          * pages can be migrated.
 883          */
 884         if (page_private(page)) {
 885                 set_page_private(newpage, page_private(page));
 886                 set_page_private(page, 0);
 887         }
 888
 889         migrate_page_copy(newpage, page);
 890
 891         return MIGRATEPAGE_SUCCESS;
 892 }
 893
 894 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 895 {
 896         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
 897         struct hstate *h = hstate_inode(d_inode(dentry));
 898
 899         buf->f_type = HUGETLBFS_MAGIC;
 900         buf->f_bsize = huge_page_size(h);
 901         if (sbinfo) {
 902                 spin_lock(&sbinfo->stat_lock);
 903                 /* If no limits set, just report 0 for max/free/used
 904                  * blocks, like simple_statfs() */
 905                 if (sbinfo->spool) {
 906                         long free_pages;
 907
 908                         spin_lock(&sbinfo->spool->lock);
 909                         buf->f_blocks = sbinfo->spool->max_hpages;
 910                         free_pages = sbinfo->spool->max_hpages
 911                                 - sbinfo->spool->used_hpages;
 912                         buf->f_bavail = buf->f_bfree = free_pages;
 913                         spin_unlock(&sbinfo->spool->lock);
 914                         buf->f_files = sbinfo->max_inodes;
 915                         buf->f_ffree = sbinfo->free_inodes;
 916                 }
 917                 spin_unlock(&sbinfo->stat_lock);
 918         }
 919         buf->f_namelen = NAME_MAX;
 920         return 0;
 921 }
 922
 923 static void hugetlbfs_put_super(struct super_block *sb)
 924 {
 925         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
 926
 927         if (sbi) {
 928                 sb->s_fs_info = NULL;
 929
 930                 if (sbi->spool)
 931                         hugepage_put_subpool(sbi->spool);
 932
 933                 kfree(sbi);
 934         }
 935 }
 936
 937 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 938 {
 939         if (sbinfo->free_inodes >= 0) {
 940                 spin_lock(&sbinfo->stat_lock);
 941                 if (unlikely(!sbinfo->free_inodes)) {
 942                         spin_unlock(&sbinfo->stat_lock);
 943                         return 0;
 944                 }
 945                 sbinfo->free_inodes--;
 946                 spin_unlock(&sbinfo->stat_lock);
 947         }
 948
 949         return 1;
 950 }
 951
 952 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 953 {
 954         if (sbinfo->free_inodes >= 0) {
 955                 spin_lock(&sbinfo->stat_lock);
 956                 sbinfo->free_inodes++;
 957                 spin_unlock(&sbinfo->stat_lock);
 958         }
 959 }
 960
 961
 962 static struct kmem_cache *hugetlbfs_inode_cachep;
 963
 964 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
 965 {
 966         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
 967         struct hugetlbfs_inode_info *p;
 968
 969         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
 970                 return NULL;
 971         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
 972         if (unlikely(!p)) {
 973                 hugetlbfs_inc_free_inodes(sbinfo);
 974                 return NULL;
 975         }
 976         return &p->vfs_inode;
 977 }
 978
 979 static void hugetlbfs_i_callback(struct rcu_head *head)
 980 {
 981         struct inode *inode = container_of(head, struct inode, i_rcu);
 982         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
 983 }
 984
 985 static void hugetlbfs_destroy_inode(struct inode *inode)
 986 {
 987         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
 988         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
 989         call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
 990 }
 991
 992 static const struct address_space_operations hugetlbfs_aops = {
 993         .write_begin    = hugetlbfs_write_begin,
 994         .write_end      = hugetlbfs_write_end,
 995         .set_page_dirty = hugetlbfs_set_page_dirty,
 996         .migratepage    = hugetlbfs_migrate_page,
 997 };
 998
 999
1000 static void init_once(void *foo)
1001 {
1002         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1003
1004         inode_init_once(&ei->vfs_inode);
1005 }
1006
1007 const struct file_operations hugetlbfs_file_operations = {
1008         .read_iter              = hugetlbfs_read_iter,
1009         .mmap                   = hugetlbfs_file_mmap,
1010         .fsync                  = noop_fsync,
1011         .get_unmapped_area      = hugetlb_get_unmapped_area,
1012         .llseek                 = default_llseek,
1013         .fallocate              = hugetlbfs_fallocate,
1014 };
1015
1016 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1017         .create         = hugetlbfs_create,
1018         .lookup         = simple_lookup,
1019         .link           = simple_link,
1020         .unlink         = simple_unlink,
1021         .symlink        = hugetlbfs_symlink,
1022         .mkdir          = hugetlbfs_mkdir,
1023         .rmdir          = simple_rmdir,
1024         .mknod          = hugetlbfs_mknod,
1025         .rename         = simple_rename,
1026         .setattr        = hugetlbfs_setattr,
1027 };
1028
1029 static const struct inode_operations hugetlbfs_inode_operations = {
1030         .setattr        = hugetlbfs_setattr,
1031 };
1032
1033 static const struct super_operations hugetlbfs_ops = {
1034         .alloc_inode    = hugetlbfs_alloc_inode,
1035         .destroy_inode  = hugetlbfs_destroy_inode,
1036         .evict_inode    = hugetlbfs_evict_inode,
1037         .statfs         = hugetlbfs_statfs,
1038         .put_super      = hugetlbfs_put_super,
1039         .show_options   = generic_show_options,
1040 };
1041
1042 enum { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1043
1044 /*
1045  * Convert size option passed from command line to number of huge pages
1046  * in the pool specified by hstate.  Size option could be in bytes
1047  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1048  */
1049 static long long
1050 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1051                                                                 int val_type)
1052 {
1053         if (val_type == NO_SIZE)
1054                 return -1;
1055
1056         if (val_type == SIZE_PERCENT) {
1057                 size_opt <<= huge_page_shift(h);
1058                 size_opt *= h->max_huge_pages;
1059                 do_div(size_opt, 100);
1060         }
1061
1062         size_opt >>= huge_page_shift(h);
1063         return size_opt;
1064 }
1065
1066 static int
1067 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1068 {
1069         char *p, *rest;
1070         substring_t args[MAX_OPT_ARGS];
1071         int option;
1072         unsigned long long max_size_opt = 0, min_size_opt = 0;
1073         int max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1074
1075         if (!options)
1076                 return 0;
1077
1078         while ((p = strsep(&options, ",")) != NULL) {
1079                 int token;
1080                 if (!*p)
1081                         continue;
1082
1083                 token = match_token(p, tokens, args);
1084                 switch (token) {
1085                 case Opt_uid:
1086                         if (match_int(&args[0], &option))
1087                                 goto bad_val;
1088                         pconfig->uid = make_kuid(current_user_ns(), option);
1089                         if (!uid_valid(pconfig->uid))
1090                                 goto bad_val;
1091                         break;
1092
1093                 case Opt_gid:
1094                         if (match_int(&args[0], &option))
1095                                 goto bad_val;
1096                         pconfig->gid = make_kgid(current_user_ns(), option);
1097                         if (!gid_valid(pconfig->gid))
1098                                 goto bad_val;
1099                         break;
1100
1101                 case Opt_mode:
1102                         if (match_octal(&args[0], &option))
1103                                 goto bad_val;
1104                         pconfig->mode = option & 01777U;
1105                         break;
1106
1107                 case Opt_size: {
1108                         /* memparse() will accept a K/M/G without a digit */
1109                         if (!isdigit(*args[0].from))
1110                                 goto bad_val;
1111                         max_size_opt = memparse(args[0].from, &rest);
1112                         max_val_type = SIZE_STD;
1113                         if (*rest == '%')
1114                                 max_val_type = SIZE_PERCENT;
1115                         break;
1116                 }
1117
1118                 case Opt_nr_inodes:
1119                         /* memparse() will accept a K/M/G without a digit */
1120                         if (!isdigit(*args[0].from))
1121                                 goto bad_val;
1122                         pconfig->nr_inodes = memparse(args[0].from, &rest);
1123                         break;
1124
1125                 case Opt_pagesize: {
1126                         unsigned long ps;
1127                         ps = memparse(args[0].from, &rest);
1128                         pconfig->hstate = size_to_hstate(ps);
1129                         if (!pconfig->hstate) {
1130                                 pr_err("Unsupported page size %lu MB\n",
1131                                         ps >> 20);
1132                                 return -EINVAL;
1133                         }
1134                         break;
1135                 }
1136
1137                 case Opt_min_size: {
1138                         /* memparse() will accept a K/M/G without a digit */
1139                         if (!isdigit(*args[0].from))
1140                                 goto bad_val;
1141                         min_size_opt = memparse(args[0].from, &rest);
1142                         min_val_type = SIZE_STD;
1143                         if (*rest == '%')
1144                                 min_val_type = SIZE_PERCENT;
1145                         break;
1146                 }
1147
1148                 default:
1149                         pr_err("Bad mount option: \"%s\"\n", p);
1150                         return -EINVAL;
1151                         break;
1152                 }
1153         }
1154
1155         /*
1156          * Use huge page pool size (in hstate) to convert the size
1157          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1158          */
1159         pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1160                                                 max_size_opt, max_val_type);
1161         pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1162                                                 min_size_opt, min_val_type);
1163
1164         /*
1165          * If max_size was specified, then min_size must be smaller
1166          */
1167         if (max_val_type > NO_SIZE &&
1168             pconfig->min_hpages > pconfig->max_hpages) {
1169                 pr_err("minimum size can not be greater than maximum size\n");
1170                 return -EINVAL;
1171         }
1172
1173         return 0;
1174
1175 bad_val:
1176         pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1177         return -EINVAL;
1178 }
1179
1180 static int
1181 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1182 {
1183         int ret;
1184         struct hugetlbfs_config config;
1185         struct hugetlbfs_sb_info *sbinfo;
1186
1187         save_mount_options(sb, data);
1188
1189         config.max_hpages = -1; /* No limit on size by default */
1190         config.nr_inodes = -1; /* No limit on number of inodes by default */
1191         config.uid = current_fsuid();
1192         config.gid = current_fsgid();
1193         config.mode = 0755;
1194         config.hstate = &default_hstate;
1195         config.min_hpages = -1; /* No default minimum size */
1196         ret = hugetlbfs_parse_options(data, &config);
1197         if (ret)
1198                 return ret;
1199
1200         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1201         if (!sbinfo)
1202                 return -ENOMEM;
1203         sb->s_fs_info = sbinfo;
1204         sbinfo->hstate = config.hstate;
1205         spin_lock_init(&sbinfo->stat_lock);
1206         sbinfo->max_inodes = config.nr_inodes;
1207         sbinfo->free_inodes = config.nr_inodes;
1208         sbinfo->spool = NULL;
1209         /*
1210          * Allocate and initialize subpool if maximum or minimum size is
1211          * specified.  Any needed reservations (for minimim size) are taken
1212          * taken when the subpool is created.
1213          */
1214         if (config.max_hpages != -1 || config.min_hpages != -1) {
1215                 sbinfo->spool = hugepage_new_subpool(config.hstate,
1216                                                         config.max_hpages,
1217                                                         config.min_hpages);
1218                 if (!sbinfo->spool)
1219                         goto out_free;
1220         }
1221         sb->s_maxbytes = MAX_LFS_FILESIZE;
1222         sb->s_blocksize = huge_page_size(config.hstate);
1223         sb->s_blocksize_bits = huge_page_shift(config.hstate);
1224         sb->s_magic = HUGETLBFS_MAGIC;
1225         sb->s_op = &hugetlbfs_ops;
1226         sb->s_time_gran = 1;
1227         sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1228         if (!sb->s_root)
1229                 goto out_free;
1230         return 0;
1231 out_free:
1232         kfree(sbinfo->spool);
1233         kfree(sbinfo);
1234         return -ENOMEM;
1235 }
1236
1237 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1238         int flags, const char *dev_name, void *data)
1239 {
1240         return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1241 }
1242
1243 static struct file_system_type hugetlbfs_fs_type = {
1244         .name           = "hugetlbfs",
1245         .mount          = hugetlbfs_mount,
1246         .kill_sb        = kill_litter_super,
1247 };
1248 MODULE_ALIAS_FS("hugetlbfs");
1249
1250 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1251
1252 static int can_do_hugetlb_shm(void)
1253 {
1254         kgid_t shm_group;
1255         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1256         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1257 }
1258
1259 static int get_hstate_idx(int page_size_log)
1260 {
1261         struct hstate *h = hstate_sizelog(page_size_log);
1262
1263         if (!h)
1264                 return -1;
1265         return h - hstates;
1266 }
1267
1268 static const struct dentry_operations anon_ops = {
1269         .d_dname = simple_dname
1270 };
1271
1272 /*
1273  * Note that size should be aligned to proper hugepage size in caller side,
1274  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1275  */
1276 struct file *hugetlb_file_setup(const char *name, size_t size,
1277                                 vm_flags_t acctflag, struct user_struct **user,
1278                                 int creat_flags, int page_size_log)
1279 {
1280         struct file *file = ERR_PTR(-ENOMEM);
1281         struct inode *inode;
1282         struct path path;
1283         struct super_block *sb;
1284         struct qstr quick_string;
1285         int hstate_idx;
1286
1287         hstate_idx = get_hstate_idx(page_size_log);
1288         if (hstate_idx < 0)
1289                 return ERR_PTR(-ENODEV);
1290
1291         *user = NULL;
1292         if (!hugetlbfs_vfsmount[hstate_idx])
1293                 return ERR_PTR(-ENOENT);
1294
1295         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1296                 *user = current_user();
1297                 if (user_shm_lock(size, *user)) {
1298                         task_lock(current);
1299                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1300                                 current->comm, current->pid);
1301                         task_unlock(current);
1302                 } else {
1303                         *user = NULL;
1304                         return ERR_PTR(-EPERM);
1305                 }
1306         }
1307
1308         sb = hugetlbfs_vfsmount[hstate_idx]->mnt_sb;
1309         quick_string.name = name;
1310         quick_string.len = strlen(quick_string.name);
1311         quick_string.hash = 0;
1312         path.dentry = d_alloc_pseudo(sb, &quick_string);
1313         if (!path.dentry)
1314                 goto out_shm_unlock;
1315
1316         d_set_d_op(path.dentry, &anon_ops);
1317         path.mnt = mntget(hugetlbfs_vfsmount[hstate_idx]);
1318         file = ERR_PTR(-ENOSPC);
1319         inode = hugetlbfs_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0);
1320         if (!inode)
1321                 goto out_dentry;
1322         if (creat_flags == HUGETLB_SHMFS_INODE)
1323                 inode->i_flags |= S_PRIVATE;
1324
1325         file = ERR_PTR(-ENOMEM);
1326         if (hugetlb_reserve_pages(inode, 0,
1327                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1328                         acctflag))
1329                 goto out_inode;
1330
1331         d_instantiate(path.dentry, inode);
1332         inode->i_size = size;
1333         clear_nlink(inode);
1334
1335         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
1336                         &hugetlbfs_file_operations);
1337         if (IS_ERR(file))
1338                 goto out_dentry; /* inode is already attached */
1339
1340         return file;
1341
1342 out_inode:
1343         iput(inode);
1344 out_dentry:
1345         path_put(&path);
1346 out_shm_unlock:
1347         if (*user) {
1348                 user_shm_unlock(size, *user);
1349                 *user = NULL;
1350         }
1351         return file;
1352 }
1353
1354 static int __init init_hugetlbfs_fs(void)
1355 {
1356         struct hstate *h;
1357         int error;
1358         int i;
1359
1360         if (!hugepages_supported()) {
1361                 pr_info("disabling because there are no supported hugepage sizes\n");
1362                 return -ENOTSUPP;
1363         }
1364
1365         error = -ENOMEM;
1366         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1367                                         sizeof(struct hugetlbfs_inode_info),
1368                                         0, 0, init_once);
1369         if (hugetlbfs_inode_cachep == NULL)
1370                 goto out2;
1371
1372         error = register_filesystem(&hugetlbfs_fs_type);
1373         if (error)
1374                 goto out;
1375
1376         i = 0;
1377         for_each_hstate(h) {
1378                 char buf[50];
1379                 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1380
1381                 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1382                 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1383                                                         buf);
1384
1385                 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1386                         pr_err("Cannot mount internal hugetlbfs for "
1387                                 "page size %uK", ps_kb);
1388                         error = PTR_ERR(hugetlbfs_vfsmount[i]);
1389                         hugetlbfs_vfsmount[i] = NULL;
1390                 }
1391                 i++;
1392         }
1393         /* Non default hstates are optional */
1394         if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1395                 return 0;
1396
1397  out:
1398         kmem_cache_destroy(hugetlbfs_inode_cachep);
1399  out2:
1400         return error;
1401 }
1402
1403 static void __exit exit_hugetlbfs_fs(void)
1404 {
1405         struct hstate *h;
1406         int i;
1407
1408
1409         /*
1410          * Make sure all delayed rcu free inodes are flushed before we
1411          * destroy cache.
1412          */
1413         rcu_barrier();
1414         kmem_cache_destroy(hugetlbfs_inode_cachep);
1415         i = 0;
1416         for_each_hstate(h)
1417                 kern_unmount(hugetlbfs_vfsmount[i++]);
1418         unregister_filesystem(&hugetlbfs_fs_type);
1419 }
1420
1421 module_init(init_hugetlbfs_fs)
1422 module_exit(exit_hugetlbfs_fs)
1423
1424 MODULE_LICENSE("GPL");