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