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 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         inode_lock(inode);
 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         inode_unlock(inode);
 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_SHIFT;
 241         offset = offset & ~PAGE_MASK;
 242
 243         while (size) {
 244                 size_t n;
 245                 chunksize = PAGE_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_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                         put_page(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 static void
 352 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
 353 {
 354         struct vm_area_struct *vma;
 355
 356         /*
 357          * end == 0 indicates that the entire range after
 358          * start should be unmapped.
 359          */
 360         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
 361                 unsigned long v_offset;
 362                 unsigned long v_end;
 363
 364                 /*
 365                  * Can the expression below overflow on 32-bit arches?
 366                  * No, because the interval tree returns us only those vmas
 367                  * which overlap the truncated area starting at pgoff,
 368                  * and no vma on a 32-bit arch can span beyond the 4GB.
 369                  */
 370                 if (vma->vm_pgoff < start)
 371                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 372                 else
 373                         v_offset = 0;
 374
 375                 if (!end)
 376                         v_end = vma->vm_end;
 377                 else {
 378                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 379                                                         + vma->vm_start;
 380                         if (v_end > vma->vm_end)
 381                                 v_end = vma->vm_end;
 382                 }
 383
 384                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 385                                                                         NULL);
 386         }
 387 }
 388
 389 /*
 390  * remove_inode_hugepages handles two distinct cases: truncation and hole
 391  * punch.  There are subtle differences in operation for each case.
 392  *
 393  * truncation is indicated by end of range being LLONG_MAX
 394  *      In this case, we first scan the range and release found pages.
 395  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
 396  *      maps and global counts.  Page faults can not race with truncation
 397  *      in this routine.  hugetlb_no_page() prevents page faults in the
 398  *      truncated range.  It checks i_size before allocation, and again after
 399  *      with the page table lock for the page held.  The same lock must be
 400  *      acquired to unmap a page.
 401  * hole punch is indicated if end is not LLONG_MAX
 402  *      In the hole punch case we scan the range and release found pages.
 403  *      Only when releasing a page is the associated region/reserv map
 404  *      deleted.  The region/reserv map for ranges without associated
 405  *      pages are not modified.  Page faults can race with hole punch.
 406  *      This is indicated if we find a mapped page.
 407  * Note: If the passed end of range value is beyond the end of file, but
 408  * not LLONG_MAX this routine still performs a hole punch operation.
 409  */
 410 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 411                                    loff_t lend)
 412 {
 413         struct hstate *h = hstate_inode(inode);
 414         struct address_space *mapping = &inode->i_data;
 415         const pgoff_t start = lstart >> huge_page_shift(h);
 416         const pgoff_t end = lend >> huge_page_shift(h);
 417         struct vm_area_struct pseudo_vma;
 418         struct pagevec pvec;
 419         pgoff_t next, index;
 420         int i, freed = 0;
 421         bool truncate_op = (lend == LLONG_MAX);
 422
 423         memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
 424         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 425         pagevec_init(&pvec, 0);
 426         next = start;
 427         while (next < end) {
 428                 /*
 429                  * When no more pages are found, we are done.
 430                  */
 431                 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
 432                         break;
 433
 434                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 435                         struct page *page = pvec.pages[i];
 436                         u32 hash;
 437
 438                         index = page->index;
 439                         hash = hugetlb_fault_mutex_hash(h, mapping, index);
 440                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 441
 442                         /*
 443                          * If page is mapped, it was faulted in after being
 444                          * unmapped in caller.  Unmap (again) now after taking
 445                          * the fault mutex.  The mutex will prevent faults
 446                          * until we finish removing the page.
 447                          *
 448                          * This race can only happen in the hole punch case.
 449                          * Getting here in a truncate operation is a bug.
 450                          */
 451                         if (unlikely(page_mapped(page))) {
 452                                 BUG_ON(truncate_op);
 453
 454                                 i_mmap_lock_write(mapping);
 455                                 hugetlb_vmdelete_list(&mapping->i_mmap,
 456                                         index * pages_per_huge_page(h),
 457                                         (index + 1) * pages_per_huge_page(h));
 458                                 i_mmap_unlock_write(mapping);
 459                         }
 460
 461                         lock_page(page);
 462                         /*
 463                          * We must free the huge page and remove from page
 464                          * cache (remove_huge_page) BEFORE removing the
 465                          * region/reserve map (hugetlb_unreserve_pages).  In
 466                          * rare out of memory conditions, removal of the
 467                          * region/reserve map could fail. Correspondingly,
 468                          * the subpool and global reserve usage count can need
 469                          * to be adjusted.
 470                          */
 471                         VM_BUG_ON(PagePrivate(page));
 472                         remove_huge_page(page);
 473                         freed++;
 474                         if (!truncate_op) {
 475                                 if (unlikely(hugetlb_unreserve_pages(inode,
 476                                                         index, index + 1, 1)))
 477                                         hugetlb_fix_reserve_counts(inode);
 478                         }
 479
 480                         unlock_page(page);
 481                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 482                 }
 483                 huge_pagevec_release(&pvec);
 484                 cond_resched();
 485         }
 486
 487         if (truncate_op)
 488                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 489 }
 490
 491 static void hugetlbfs_evict_inode(struct inode *inode)
 492 {
 493         struct resv_map *resv_map;
 494
 495         remove_inode_hugepages(inode, 0, LLONG_MAX);
 496         resv_map = (struct resv_map *)inode->i_mapping->private_data;
 497         /* root inode doesn't have the resv_map, so we should check it */
 498         if (resv_map)
 499                 resv_map_release(&resv_map->refs);
 500         clear_inode(inode);
 501 }
 502
 503 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 504 {
 505         pgoff_t pgoff;
 506         struct address_space *mapping = inode->i_mapping;
 507         struct hstate *h = hstate_inode(inode);
 508
 509         BUG_ON(offset & ~huge_page_mask(h));
 510         pgoff = offset >> PAGE_SHIFT;
 511
 512         i_size_write(inode, offset);
 513         i_mmap_lock_write(mapping);
 514         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 515                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 516         i_mmap_unlock_write(mapping);
 517         remove_inode_hugepages(inode, offset, LLONG_MAX);
 518         return 0;
 519 }
 520
 521 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 522 {
 523         struct hstate *h = hstate_inode(inode);
 524         loff_t hpage_size = huge_page_size(h);
 525         loff_t hole_start, hole_end;
 526
 527         /*
 528          * For hole punch round up the beginning offset of the hole and
 529          * round down the end.
 530          */
 531         hole_start = round_up(offset, hpage_size);
 532         hole_end = round_down(offset + len, hpage_size);
 533
 534         if (hole_end > hole_start) {
 535                 struct address_space *mapping = inode->i_mapping;
 536
 537                 inode_lock(inode);
 538                 i_mmap_lock_write(mapping);
 539                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 540                         hugetlb_vmdelete_list(&mapping->i_mmap,
 541                                                 hole_start >> PAGE_SHIFT,
 542                                                 hole_end  >> PAGE_SHIFT);
 543                 i_mmap_unlock_write(mapping);
 544                 remove_inode_hugepages(inode, hole_start, hole_end);
 545                 inode_unlock(inode);
 546         }
 547
 548         return 0;
 549 }
 550
 551 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 552                                 loff_t len)
 553 {
 554         struct inode *inode = file_inode(file);
 555         struct address_space *mapping = inode->i_mapping;
 556         struct hstate *h = hstate_inode(inode);
 557         struct vm_area_struct pseudo_vma;
 558         loff_t hpage_size = huge_page_size(h);
 559         unsigned long hpage_shift = huge_page_shift(h);
 560         pgoff_t start, index, end;
 561         int error;
 562         u32 hash;
 563
 564         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 565                 return -EOPNOTSUPP;
 566
 567         if (mode & FALLOC_FL_PUNCH_HOLE)
 568                 return hugetlbfs_punch_hole(inode, offset, len);
 569
 570         /*
 571          * Default preallocate case.
 572          * For this range, start is rounded down and end is rounded up
 573          * as well as being converted to page offsets.
 574          */
 575         start = offset >> hpage_shift;
 576         end = (offset + len + hpage_size - 1) >> hpage_shift;
 577
 578         inode_lock(inode);
 579
 580         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 581         error = inode_newsize_ok(inode, offset + len);
 582         if (error)
 583                 goto out;
 584
 585         /*
 586          * Initialize a pseudo vma as this is required by the huge page
 587          * allocation routines.  If NUMA is configured, use page index
 588          * as input to create an allocation policy.
 589          */
 590         memset(&pseudo_vma, 0, sizeof(struct vm_area_struct));
 591         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 592         pseudo_vma.vm_file = file;
 593
 594         for (index = start; index < end; index++) {
 595                 /*
 596                  * This is supposed to be the vaddr where the page is being
 597                  * faulted in, but we have no vaddr here.
 598                  */
 599                 struct page *page;
 600                 unsigned long addr;
 601                 int avoid_reserve = 0;
 602
 603                 cond_resched();
 604
 605                 /*
 606                  * fallocate(2) manpage permits EINTR; we may have been
 607                  * interrupted because we are using up too much memory.
 608                  */
 609                 if (signal_pending(current)) {
 610                         error = -EINTR;
 611                         break;
 612                 }
 613
 614                 /* Set numa allocation policy based on index */
 615                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 616
 617                 /* addr is the offset within the file (zero based) */
 618                 addr = index * hpage_size;
 619
 620                 /* mutex taken here, fault path and hole punch */
 621                 hash = hugetlb_fault_mutex_hash(h, mapping, index);
 622                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 623
 624                 /* See if already present in mapping to avoid alloc/free */
 625                 page = find_get_page(mapping, index);
 626                 if (page) {
 627                         put_page(page);
 628                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 629                         hugetlb_drop_vma_policy(&pseudo_vma);
 630                         continue;
 631                 }
 632
 633                 /* Allocate page and add to page cache */
 634                 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
 635                 hugetlb_drop_vma_policy(&pseudo_vma);
 636                 if (IS_ERR(page)) {
 637                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 638                         error = PTR_ERR(page);
 639                         goto out;
 640                 }
 641                 clear_huge_page(page, addr, pages_per_huge_page(h));
 642                 __SetPageUptodate(page);
 643                 error = huge_add_to_page_cache(page, mapping, index);
 644                 if (unlikely(error)) {
 645                         put_page(page);
 646                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 647                         goto out;
 648                 }
 649
 650                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 651
 652                 set_page_huge_active(page);
 653                 /*
 654                  * put_page() due to reference from alloc_huge_page()
 655                  * unlock_page because locked by add_to_page_cache()
 656                  */
 657                 put_page(page);
 658                 unlock_page(page);
 659         }
 660
 661         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 662                 i_size_write(inode, offset + len);
 663         inode->i_ctime = current_time(inode);
 664 out:
 665         inode_unlock(inode);
 666         return error;
 667 }
 668
 669 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 670 {
 671         struct inode *inode = d_inode(dentry);
 672         struct hstate *h = hstate_inode(inode);
 673         int error;
 674         unsigned int ia_valid = attr->ia_valid;
 675
 676         BUG_ON(!inode);
 677
 678         error = setattr_prepare(dentry, attr);
 679         if (error)
 680                 return error;
 681
 682         if (ia_valid & ATTR_SIZE) {
 683                 error = -EINVAL;
 684                 if (attr->ia_size & ~huge_page_mask(h))
 685                         return -EINVAL;
 686                 error = hugetlb_vmtruncate(inode, attr->ia_size);
 687                 if (error)
 688                         return error;
 689         }
 690
 691         setattr_copy(inode, attr);
 692         mark_inode_dirty(inode);
 693         return 0;
 694 }
 695
 696 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 697                                         struct hugetlbfs_config *config)
 698 {
 699         struct inode *inode;
 700
 701         inode = new_inode(sb);
 702         if (inode) {
 703                 inode->i_ino = get_next_ino();
 704                 inode->i_mode = S_IFDIR | config->mode;
 705                 inode->i_uid = config->uid;
 706                 inode->i_gid = config->gid;
 707                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 708                 inode->i_op = &hugetlbfs_dir_inode_operations;
 709                 inode->i_fop = &simple_dir_operations;
 710                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 711                 inc_nlink(inode);
 712                 lockdep_annotate_inode_mutex_key(inode);
 713         }
 714         return inode;
 715 }
 716
 717 /*
 718  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 719  * be taken from reclaim -- unlike regular filesystems. This needs an
 720  * annotation because huge_pmd_share() does an allocation under hugetlb's
 721  * i_mmap_rwsem.
 722  */
 723 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 724
 725 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 726                                         struct inode *dir,
 727                                         umode_t mode, dev_t dev)
 728 {
 729         struct inode *inode;
 730         struct resv_map *resv_map = NULL;
 731
 732         /*
 733          * Reserve maps are only needed for inodes that can have associated
 734          * page allocations.
 735          */
 736         if (S_ISREG(mode) || S_ISLNK(mode)) {
 737                 resv_map = resv_map_alloc();
 738                 if (!resv_map)
 739                         return NULL;
 740         }
 741
 742         inode = new_inode(sb);
 743         if (inode) {
 744                 inode->i_ino = get_next_ino();
 745                 inode_init_owner(inode, dir, mode);
 746                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 747                                 &hugetlbfs_i_mmap_rwsem_key);
 748                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 749                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 750                 inode->i_mapping->private_data = resv_map;
 751                 switch (mode & S_IFMT) {
 752                 default:
 753                         init_special_inode(inode, mode, dev);
 754                         break;
 755                 case S_IFREG:
 756                         inode->i_op = &hugetlbfs_inode_operations;
 757                         inode->i_fop = &hugetlbfs_file_operations;
 758                         break;
 759                 case S_IFDIR:
 760                         inode->i_op = &hugetlbfs_dir_inode_operations;
 761                         inode->i_fop = &simple_dir_operations;
 762
 763                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 764                         inc_nlink(inode);
 765                         break;
 766                 case S_IFLNK:
 767                         inode->i_op = &page_symlink_inode_operations;
 768                         inode_nohighmem(inode);
 769                         break;
 770                 }
 771                 lockdep_annotate_inode_mutex_key(inode);
 772         } else {
 773                 if (resv_map)
 774                         kref_put(&resv_map->refs, resv_map_release);
 775         }
 776
 777         return inode;
 778 }
 779
 780 /*
 781  * File creation. Allocate an inode, and we're done..
 782  */
 783 static int hugetlbfs_mknod(struct inode *dir,
 784                         struct dentry *dentry, umode_t mode, dev_t dev)
 785 {
 786         struct inode *inode;
 787         int error = -ENOSPC;
 788
 789         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 790         if (inode) {
 791                 dir->i_ctime = dir->i_mtime = current_time(dir);
 792                 d_instantiate(dentry, inode);
 793                 dget(dentry);   /* Extra count - pin the dentry in core */
 794                 error = 0;
 795         }
 796         return error;
 797 }
 798
 799 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 800 {
 801         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
 802         if (!retval)
 803                 inc_nlink(dir);
 804         return retval;
 805 }
 806
 807 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
 808 {
 809         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
 810 }
 811
 812 static int hugetlbfs_symlink(struct inode *dir,
 813                         struct dentry *dentry, const char *symname)
 814 {
 815         struct inode *inode;
 816         int error = -ENOSPC;
 817
 818         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 819         if (inode) {
 820                 int l = strlen(symname)+1;
 821                 error = page_symlink(inode, symname, l);
 822                 if (!error) {
 823                         d_instantiate(dentry, inode);
 824                         dget(dentry);
 825                 } else
 826                         iput(inode);
 827         }
 828         dir->i_ctime = dir->i_mtime = current_time(dir);
 829
 830         return error;
 831 }
 832
 833 /*
 834  * mark the head page dirty
 835  */
 836 static int hugetlbfs_set_page_dirty(struct page *page)
 837 {
 838         struct page *head = compound_head(page);
 839
 840         SetPageDirty(head);
 841         return 0;
 842 }
 843
 844 static int hugetlbfs_migrate_page(struct address_space *mapping,
 845                                 struct page *newpage, struct page *page,
 846                                 enum migrate_mode mode)
 847 {
 848         int rc;
 849
 850         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 851         if (rc != MIGRATEPAGE_SUCCESS)
 852                 return rc;
 853
 854         /*
 855          * page_private is subpool pointer in hugetlb pages.  Transfer to
 856          * new page.  PagePrivate is not associated with page_private for
 857          * hugetlb pages and can not be set here as only page_huge_active
 858          * pages can be migrated.
 859          */
 860         if (page_private(page)) {
 861                 set_page_private(newpage, page_private(page));
 862                 set_page_private(page, 0);
 863         }
 864
 865         if (mode != MIGRATE_SYNC_NO_COPY)
 866                 migrate_page_copy(newpage, page);
 867         else
 868                 migrate_page_states(newpage, page);
 869
 870         return MIGRATEPAGE_SUCCESS;
 871 }
 872
 873 static int hugetlbfs_error_remove_page(struct address_space *mapping,
 874                                 struct page *page)
 875 {
 876         struct inode *inode = mapping->host;
 877         pgoff_t index = page->index;
 878
 879         remove_huge_page(page);
 880         if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
 881                 hugetlb_fix_reserve_counts(inode);
 882
 883         return 0;
 884 }
 885
 886 /*
 887  * Display the mount options in /proc/mounts.
 888  */
 889 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
 890 {
 891         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
 892         struct hugepage_subpool *spool = sbinfo->spool;
 893         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
 894         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
 895         char mod;
 896
 897         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
 898                 seq_printf(m, ",uid=%u",
 899                            from_kuid_munged(&init_user_ns, sbinfo->uid));
 900         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
 901                 seq_printf(m, ",gid=%u",
 902                            from_kgid_munged(&init_user_ns, sbinfo->gid));
 903         if (sbinfo->mode != 0755)
 904                 seq_printf(m, ",mode=%o", sbinfo->mode);
 905         if (sbinfo->max_inodes != -1)
 906                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
 907
 908         hpage_size /= 1024;
 909         mod = 'K';
 910         if (hpage_size >= 1024) {
 911                 hpage_size /= 1024;
 912                 mod = 'M';
 913         }
 914         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
 915         if (spool) {
 916                 if (spool->max_hpages != -1)
 917                         seq_printf(m, ",size=%llu",
 918                                    (unsigned long long)spool->max_hpages << hpage_shift);
 919                 if (spool->min_hpages != -1)
 920                         seq_printf(m, ",min_size=%llu",
 921                                    (unsigned long long)spool->min_hpages << hpage_shift);
 922         }
 923         return 0;
 924 }
 925
 926 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 927 {
 928         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
 929         struct hstate *h = hstate_inode(d_inode(dentry));
 930
 931         buf->f_type = HUGETLBFS_MAGIC;
 932         buf->f_bsize = huge_page_size(h);
 933         if (sbinfo) {
 934                 spin_lock(&sbinfo->stat_lock);
 935                 /* If no limits set, just report 0 for max/free/used
 936                  * blocks, like simple_statfs() */
 937                 if (sbinfo->spool) {
 938                         long free_pages;
 939
 940                         spin_lock(&sbinfo->spool->lock);
 941                         buf->f_blocks = sbinfo->spool->max_hpages;
 942                         free_pages = sbinfo->spool->max_hpages
 943                                 - sbinfo->spool->used_hpages;
 944                         buf->f_bavail = buf->f_bfree = free_pages;
 945                         spin_unlock(&sbinfo->spool->lock);
 946                         buf->f_files = sbinfo->max_inodes;
 947                         buf->f_ffree = sbinfo->free_inodes;
 948                 }
 949                 spin_unlock(&sbinfo->stat_lock);
 950         }
 951         buf->f_namelen = NAME_MAX;
 952         return 0;
 953 }
 954
 955 static void hugetlbfs_put_super(struct super_block *sb)
 956 {
 957         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
 958
 959         if (sbi) {
 960                 sb->s_fs_info = NULL;
 961
 962                 if (sbi->spool)
 963                         hugepage_put_subpool(sbi->spool);
 964
 965                 kfree(sbi);
 966         }
 967 }
 968
 969 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 970 {
 971         if (sbinfo->free_inodes >= 0) {
 972                 spin_lock(&sbinfo->stat_lock);
 973                 if (unlikely(!sbinfo->free_inodes)) {
 974                         spin_unlock(&sbinfo->stat_lock);
 975                         return 0;
 976                 }
 977                 sbinfo->free_inodes--;
 978                 spin_unlock(&sbinfo->stat_lock);
 979         }
 980
 981         return 1;
 982 }
 983
 984 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 985 {
 986         if (sbinfo->free_inodes >= 0) {
 987                 spin_lock(&sbinfo->stat_lock);
 988                 sbinfo->free_inodes++;
 989                 spin_unlock(&sbinfo->stat_lock);
 990         }
 991 }
 992
 993
 994 static struct kmem_cache *hugetlbfs_inode_cachep;
 995
 996 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
 997 {
 998         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
 999         struct hugetlbfs_inode_info *p;
1000
1001         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1002                 return NULL;
1003         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1004         if (unlikely(!p)) {
1005                 hugetlbfs_inc_free_inodes(sbinfo);
1006                 return NULL;
1007         }
1008
1009         /*
1010          * Any time after allocation, hugetlbfs_destroy_inode can be called
1011          * for the inode.  mpol_free_shared_policy is unconditionally called
1012          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1013          * in case of a quick call to destroy.
1014          *
1015          * Note that the policy is initialized even if we are creating a
1016          * private inode.  This simplifies hugetlbfs_destroy_inode.
1017          */
1018         mpol_shared_policy_init(&p->policy, NULL);
1019
1020         return &p->vfs_inode;
1021 }
1022
1023 static void hugetlbfs_i_callback(struct rcu_head *head)
1024 {
1025         struct inode *inode = container_of(head, struct inode, i_rcu);
1026         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1027 }
1028
1029 static void hugetlbfs_destroy_inode(struct inode *inode)
1030 {
1031         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1032         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1033         call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
1034 }
1035
1036 static const struct address_space_operations hugetlbfs_aops = {
1037         .write_begin    = hugetlbfs_write_begin,
1038         .write_end      = hugetlbfs_write_end,
1039         .set_page_dirty = hugetlbfs_set_page_dirty,
1040         .migratepage    = hugetlbfs_migrate_page,
1041         .error_remove_page      = hugetlbfs_error_remove_page,
1042 };
1043
1044
1045 static void init_once(void *foo)
1046 {
1047         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1048
1049         inode_init_once(&ei->vfs_inode);
1050 }
1051
1052 const struct file_operations hugetlbfs_file_operations = {
1053         .read_iter              = hugetlbfs_read_iter,
1054         .mmap                   = hugetlbfs_file_mmap,
1055         .fsync                  = noop_fsync,
1056         .get_unmapped_area      = hugetlb_get_unmapped_area,
1057         .llseek                 = default_llseek,
1058         .fallocate              = hugetlbfs_fallocate,
1059 };
1060
1061 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1062         .create         = hugetlbfs_create,
1063         .lookup         = simple_lookup,
1064         .link           = simple_link,
1065         .unlink         = simple_unlink,
1066         .symlink        = hugetlbfs_symlink,
1067         .mkdir          = hugetlbfs_mkdir,
1068         .rmdir          = simple_rmdir,
1069         .mknod          = hugetlbfs_mknod,
1070         .rename         = simple_rename,
1071         .setattr        = hugetlbfs_setattr,
1072 };
1073
1074 static const struct inode_operations hugetlbfs_inode_operations = {
1075         .setattr        = hugetlbfs_setattr,
1076 };
1077
1078 static const struct super_operations hugetlbfs_ops = {
1079         .alloc_inode    = hugetlbfs_alloc_inode,
1080         .destroy_inode  = hugetlbfs_destroy_inode,
1081         .evict_inode    = hugetlbfs_evict_inode,
1082         .statfs         = hugetlbfs_statfs,
1083         .put_super      = hugetlbfs_put_super,
1084         .show_options   = hugetlbfs_show_options,
1085 };
1086
1087 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1088
1089 /*
1090  * Convert size option passed from command line to number of huge pages
1091  * in the pool specified by hstate.  Size option could be in bytes
1092  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1093  */
1094 static long
1095 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1096                          enum hugetlbfs_size_type val_type)
1097 {
1098         if (val_type == NO_SIZE)
1099                 return -1;
1100
1101         if (val_type == SIZE_PERCENT) {
1102                 size_opt <<= huge_page_shift(h);
1103                 size_opt *= h->max_huge_pages;
1104                 do_div(size_opt, 100);
1105         }
1106
1107         size_opt >>= huge_page_shift(h);
1108         return size_opt;
1109 }
1110
1111 static int
1112 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1113 {
1114         char *p, *rest;
1115         substring_t args[MAX_OPT_ARGS];
1116         int option;
1117         unsigned long long max_size_opt = 0, min_size_opt = 0;
1118         enum hugetlbfs_size_type max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1119
1120         if (!options)
1121                 return 0;
1122
1123         while ((p = strsep(&options, ",")) != NULL) {
1124                 int token;
1125                 if (!*p)
1126                         continue;
1127
1128                 token = match_token(p, tokens, args);
1129                 switch (token) {
1130                 case Opt_uid:
1131                         if (match_int(&args[0], &option))
1132                                 goto bad_val;
1133                         pconfig->uid = make_kuid(current_user_ns(), option);
1134                         if (!uid_valid(pconfig->uid))
1135                                 goto bad_val;
1136                         break;
1137
1138                 case Opt_gid:
1139                         if (match_int(&args[0], &option))
1140                                 goto bad_val;
1141                         pconfig->gid = make_kgid(current_user_ns(), option);
1142                         if (!gid_valid(pconfig->gid))
1143                                 goto bad_val;
1144                         break;
1145
1146                 case Opt_mode:
1147                         if (match_octal(&args[0], &option))
1148                                 goto bad_val;
1149                         pconfig->mode = option & 01777U;
1150                         break;
1151
1152                 case Opt_size: {
1153                         /* memparse() will accept a K/M/G without a digit */
1154                         if (!isdigit(*args[0].from))
1155                                 goto bad_val;
1156                         max_size_opt = memparse(args[0].from, &rest);
1157                         max_val_type = SIZE_STD;
1158                         if (*rest == '%')
1159                                 max_val_type = SIZE_PERCENT;
1160                         break;
1161                 }
1162
1163                 case Opt_nr_inodes:
1164                         /* memparse() will accept a K/M/G without a digit */
1165                         if (!isdigit(*args[0].from))
1166                                 goto bad_val;
1167                         pconfig->nr_inodes = memparse(args[0].from, &rest);
1168                         break;
1169
1170                 case Opt_pagesize: {
1171                         unsigned long ps;
1172                         ps = memparse(args[0].from, &rest);
1173                         pconfig->hstate = size_to_hstate(ps);
1174                         if (!pconfig->hstate) {
1175                                 pr_err("Unsupported page size %lu MB\n",
1176                                         ps >> 20);
1177                                 return -EINVAL;
1178                         }
1179                         break;
1180                 }
1181
1182                 case Opt_min_size: {
1183                         /* memparse() will accept a K/M/G without a digit */
1184                         if (!isdigit(*args[0].from))
1185                                 goto bad_val;
1186                         min_size_opt = memparse(args[0].from, &rest);
1187                         min_val_type = SIZE_STD;
1188                         if (*rest == '%')
1189                                 min_val_type = SIZE_PERCENT;
1190                         break;
1191                 }
1192
1193                 default:
1194                         pr_err("Bad mount option: \"%s\"\n", p);
1195                         return -EINVAL;
1196                         break;
1197                 }
1198         }
1199
1200         /*
1201          * Use huge page pool size (in hstate) to convert the size
1202          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1203          */
1204         pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1205                                                 max_size_opt, max_val_type);
1206         pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1207                                                 min_size_opt, min_val_type);
1208
1209         /*
1210          * If max_size was specified, then min_size must be smaller
1211          */
1212         if (max_val_type > NO_SIZE &&
1213             pconfig->min_hpages > pconfig->max_hpages) {
1214                 pr_err("minimum size can not be greater than maximum size\n");
1215                 return -EINVAL;
1216         }
1217
1218         return 0;
1219
1220 bad_val:
1221         pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1222         return -EINVAL;
1223 }
1224
1225 static int
1226 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1227 {
1228         int ret;
1229         struct hugetlbfs_config config;
1230         struct hugetlbfs_sb_info *sbinfo;
1231
1232         config.max_hpages = -1; /* No limit on size by default */
1233         config.nr_inodes = -1; /* No limit on number of inodes by default */
1234         config.uid = current_fsuid();
1235         config.gid = current_fsgid();
1236         config.mode = 0755;
1237         config.hstate = &default_hstate;
1238         config.min_hpages = -1; /* No default minimum size */
1239         ret = hugetlbfs_parse_options(data, &config);
1240         if (ret)
1241                 return ret;
1242
1243         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1244         if (!sbinfo)
1245                 return -ENOMEM;
1246         sb->s_fs_info = sbinfo;
1247         sbinfo->hstate = config.hstate;
1248         spin_lock_init(&sbinfo->stat_lock);
1249         sbinfo->max_inodes = config.nr_inodes;
1250         sbinfo->free_inodes = config.nr_inodes;
1251         sbinfo->spool = NULL;
1252         sbinfo->uid = config.uid;
1253         sbinfo->gid = config.gid;
1254         sbinfo->mode = config.mode;
1255
1256         /*
1257          * Allocate and initialize subpool if maximum or minimum size is
1258          * specified.  Any needed reservations (for minimim size) are taken
1259          * taken when the subpool is created.
1260          */
1261         if (config.max_hpages != -1 || config.min_hpages != -1) {
1262                 sbinfo->spool = hugepage_new_subpool(config.hstate,
1263                                                         config.max_hpages,
1264                                                         config.min_hpages);
1265                 if (!sbinfo->spool)
1266                         goto out_free;
1267         }
1268         sb->s_maxbytes = MAX_LFS_FILESIZE;
1269         sb->s_blocksize = huge_page_size(config.hstate);
1270         sb->s_blocksize_bits = huge_page_shift(config.hstate);
1271         sb->s_magic = HUGETLBFS_MAGIC;
1272         sb->s_op = &hugetlbfs_ops;
1273         sb->s_time_gran = 1;
1274         sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1275         if (!sb->s_root)
1276                 goto out_free;
1277         return 0;
1278 out_free:
1279         kfree(sbinfo->spool);
1280         kfree(sbinfo);
1281         return -ENOMEM;
1282 }
1283
1284 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1285         int flags, const char *dev_name, void *data)
1286 {
1287         return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1288 }
1289
1290 static struct file_system_type hugetlbfs_fs_type = {
1291         .name           = "hugetlbfs",
1292         .mount          = hugetlbfs_mount,
1293         .kill_sb        = kill_litter_super,
1294 };
1295
1296 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1297
1298 static int can_do_hugetlb_shm(void)
1299 {
1300         kgid_t shm_group;
1301         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1302         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1303 }
1304
1305 static int get_hstate_idx(int page_size_log)
1306 {
1307         struct hstate *h = hstate_sizelog(page_size_log);
1308
1309         if (!h)
1310                 return -1;
1311         return h - hstates;
1312 }
1313
1314 static const struct dentry_operations anon_ops = {
1315         .d_dname = simple_dname
1316 };
1317
1318 /*
1319  * Note that size should be aligned to proper hugepage size in caller side,
1320  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1321  */
1322 struct file *hugetlb_file_setup(const char *name, size_t size,
1323                                 vm_flags_t acctflag, struct user_struct **user,
1324                                 int creat_flags, int page_size_log)
1325 {
1326         struct file *file = ERR_PTR(-ENOMEM);
1327         struct inode *inode;
1328         struct path path;
1329         struct super_block *sb;
1330         struct qstr quick_string;
1331         int hstate_idx;
1332
1333         hstate_idx = get_hstate_idx(page_size_log);
1334         if (hstate_idx < 0)
1335                 return ERR_PTR(-ENODEV);
1336
1337         *user = NULL;
1338         if (!hugetlbfs_vfsmount[hstate_idx])
1339                 return ERR_PTR(-ENOENT);
1340
1341         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1342                 *user = current_user();
1343                 if (user_shm_lock(size, *user)) {
1344                         task_lock(current);
1345                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1346                                 current->comm, current->pid);
1347                         task_unlock(current);
1348                 } else {
1349                         *user = NULL;
1350                         return ERR_PTR(-EPERM);
1351                 }
1352         }
1353
1354         sb = hugetlbfs_vfsmount[hstate_idx]->mnt_sb;
1355         quick_string.name = name;
1356         quick_string.len = strlen(quick_string.name);
1357         quick_string.hash = 0;
1358         path.dentry = d_alloc_pseudo(sb, &quick_string);
1359         if (!path.dentry)
1360                 goto out_shm_unlock;
1361
1362         d_set_d_op(path.dentry, &anon_ops);
1363         path.mnt = mntget(hugetlbfs_vfsmount[hstate_idx]);
1364         file = ERR_PTR(-ENOSPC);
1365         inode = hugetlbfs_get_inode(sb, NULL, S_IFREG | S_IRWXUGO, 0);
1366         if (!inode)
1367                 goto out_dentry;
1368         if (creat_flags == HUGETLB_SHMFS_INODE)
1369                 inode->i_flags |= S_PRIVATE;
1370
1371         file = ERR_PTR(-ENOMEM);
1372         if (hugetlb_reserve_pages(inode, 0,
1373                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1374                         acctflag))
1375                 goto out_inode;
1376
1377         d_instantiate(path.dentry, inode);
1378         inode->i_size = size;
1379         clear_nlink(inode);
1380
1381         file = alloc_file(&path, FMODE_WRITE | FMODE_READ,
1382                         &hugetlbfs_file_operations);
1383         if (IS_ERR(file))
1384                 goto out_dentry; /* inode is already attached */
1385
1386         return file;
1387
1388 out_inode:
1389         iput(inode);
1390 out_dentry:
1391         path_put(&path);
1392 out_shm_unlock:
1393         if (*user) {
1394                 user_shm_unlock(size, *user);
1395                 *user = NULL;
1396         }
1397         return file;
1398 }
1399
1400 static int __init init_hugetlbfs_fs(void)
1401 {
1402         struct hstate *h;
1403         int error;
1404         int i;
1405
1406         if (!hugepages_supported()) {
1407                 pr_info("disabling because there are no supported hugepage sizes\n");
1408                 return -ENOTSUPP;
1409         }
1410
1411         error = -ENOMEM;
1412         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1413                                         sizeof(struct hugetlbfs_inode_info),
1414                                         0, SLAB_ACCOUNT, init_once);
1415         if (hugetlbfs_inode_cachep == NULL)
1416                 goto out2;
1417
1418         error = register_filesystem(&hugetlbfs_fs_type);
1419         if (error)
1420                 goto out;
1421
1422         i = 0;
1423         for_each_hstate(h) {
1424                 char buf[50];
1425                 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1426
1427                 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1428                 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1429                                                         buf);
1430
1431                 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1432                         pr_err("Cannot mount internal hugetlbfs for "
1433                                 "page size %uK", ps_kb);
1434                         error = PTR_ERR(hugetlbfs_vfsmount[i]);
1435                         hugetlbfs_vfsmount[i] = NULL;
1436                 }
1437                 i++;
1438         }
1439         /* Non default hstates are optional */
1440         if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1441                 return 0;
1442
1443  out:
1444         kmem_cache_destroy(hugetlbfs_inode_cachep);
1445  out2:
1446         return error;
1447 }
1448 fs_initcall(init_hugetlbfs_fs)