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/fs_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 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
  49
  50 struct hugetlbfs_fs_context {
  51         struct hstate           *hstate;
  52         unsigned long long      max_size_opt;
  53         unsigned long long      min_size_opt;
  54         long                    max_hpages;
  55         long                    nr_inodes;
  56         long                    min_hpages;
  57         enum hugetlbfs_size_type max_val_type;
  58         enum hugetlbfs_size_type min_val_type;
  59         kuid_t                  uid;
  60         kgid_t                  gid;
  61         umode_t                 mode;
  62 };
  63
  64 int sysctl_hugetlb_shm_group;
  65
  66 enum hugetlb_param {
  67         Opt_gid,
  68         Opt_min_size,
  69         Opt_mode,
  70         Opt_nr_inodes,
  71         Opt_pagesize,
  72         Opt_size,
  73         Opt_uid,
  74 };
  75
  76 static const struct fs_parameter_spec hugetlb_param_specs[] = {
  77         fsparam_u32   ("gid",           Opt_gid),
  78         fsparam_string("min_size",      Opt_min_size),
  79         fsparam_u32   ("mode",          Opt_mode),
  80         fsparam_string("nr_inodes",     Opt_nr_inodes),
  81         fsparam_string("pagesize",      Opt_pagesize),
  82         fsparam_string("size",          Opt_size),
  83         fsparam_u32   ("uid",           Opt_uid),
  84         {}
  85 };
  86
  87 static const struct fs_parameter_description hugetlb_fs_parameters = {
  88         .name           = "hugetlbfs",
  89         .specs          = hugetlb_param_specs,
  90 };
  91
  92 #ifdef CONFIG_NUMA
  93 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  94                                         struct inode *inode, pgoff_t index)
  95 {
  96         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
  97                                                         index);
  98 }
  99
 100 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
 101 {
 102         mpol_cond_put(vma->vm_policy);
 103 }
 104 #else
 105 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
 106                                         struct inode *inode, pgoff_t index)
 107 {
 108 }
 109
 110 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
 111 {
 112 }
 113 #endif
 114
 115 static void huge_pagevec_release(struct pagevec *pvec)
 116 {
 117         int i;
 118
 119         for (i = 0; i < pagevec_count(pvec); ++i)
 120                 put_page(pvec->pages[i]);
 121
 122         pagevec_reinit(pvec);
 123 }
 124
 125 /*
 126  * Mask used when checking the page offset value passed in via system
 127  * calls.  This value will be converted to a loff_t which is signed.
 128  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 129  * value.  The extra bit (- 1 in the shift value) is to take the sign
 130  * bit into account.
 131  */
 132 #define PGOFF_LOFFT_MAX \
 133         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 134
 135 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 136 {
 137         struct inode *inode = file_inode(file);
 138         loff_t len, vma_len;
 139         int ret;
 140         struct hstate *h = hstate_file(file);
 141
 142         /*
 143          * vma address alignment (but not the pgoff alignment) has
 144          * already been checked by prepare_hugepage_range.  If you add
 145          * any error returns here, do so after setting VM_HUGETLB, so
 146          * is_vm_hugetlb_page tests below unmap_region go the right
 147          * way when do_mmap_pgoff unwinds (may be important on powerpc
 148          * and ia64).
 149          */
 150         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 151         vma->vm_ops = &hugetlb_vm_ops;
 152
 153         /*
 154          * page based offset in vm_pgoff could be sufficiently large to
 155          * overflow a loff_t when converted to byte offset.  This can
 156          * only happen on architectures where sizeof(loff_t) ==
 157          * sizeof(unsigned long).  So, only check in those instances.
 158          */
 159         if (sizeof(unsigned long) == sizeof(loff_t)) {
 160                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 161                         return -EINVAL;
 162         }
 163
 164         /* must be huge page aligned */
 165         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 166                 return -EINVAL;
 167
 168         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 169         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 170         /* check for overflow */
 171         if (len < vma_len)
 172                 return -EINVAL;
 173
 174         inode_lock(inode);
 175         file_accessed(file);
 176
 177         ret = -ENOMEM;
 178         if (hugetlb_reserve_pages(inode,
 179                                 vma->vm_pgoff >> huge_page_order(h),
 180                                 len >> huge_page_shift(h), vma,
 181                                 vma->vm_flags))
 182                 goto out;
 183
 184         ret = 0;
 185         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 186                 i_size_write(inode, len);
 187 out:
 188         inode_unlock(inode);
 189
 190         return ret;
 191 }
 192
 193 /*
 194  * Called under down_write(mmap_sem).
 195  */
 196
 197 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 198 static unsigned long
 199 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 200                 unsigned long len, unsigned long pgoff, unsigned long flags)
 201 {
 202         struct mm_struct *mm = current->mm;
 203         struct vm_area_struct *vma;
 204         struct hstate *h = hstate_file(file);
 205         struct vm_unmapped_area_info info;
 206
 207         if (len & ~huge_page_mask(h))
 208                 return -EINVAL;
 209         if (len > TASK_SIZE)
 210                 return -ENOMEM;
 211
 212         if (flags & MAP_FIXED) {
 213                 if (prepare_hugepage_range(file, addr, len))
 214                         return -EINVAL;
 215                 return addr;
 216         }
 217
 218         if (addr) {
 219                 addr = ALIGN(addr, huge_page_size(h));
 220                 vma = find_vma(mm, addr);
 221                 if (TASK_SIZE - len >= addr &&
 222                     (!vma || addr + len <= vm_start_gap(vma)))
 223                         return addr;
 224         }
 225
 226         info.flags = 0;
 227         info.length = len;
 228         info.low_limit = TASK_UNMAPPED_BASE;
 229         info.high_limit = TASK_SIZE;
 230         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 231         info.align_offset = 0;
 232         return vm_unmapped_area(&info);
 233 }
 234 #endif
 235
 236 static size_t
 237 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 238                         struct iov_iter *to, unsigned long size)
 239 {
 240         size_t copied = 0;
 241         int i, chunksize;
 242
 243         /* Find which 4k chunk and offset with in that chunk */
 244         i = offset >> PAGE_SHIFT;
 245         offset = offset & ~PAGE_MASK;
 246
 247         while (size) {
 248                 size_t n;
 249                 chunksize = PAGE_SIZE;
 250                 if (offset)
 251                         chunksize -= offset;
 252                 if (chunksize > size)
 253                         chunksize = size;
 254                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 255                 copied += n;
 256                 if (n != chunksize)
 257                         return copied;
 258                 offset = 0;
 259                 size -= chunksize;
 260                 i++;
 261         }
 262         return copied;
 263 }
 264
 265 /*
 266  * Support for read() - Find the page attached to f_mapping and copy out the
 267  * data. Its *very* similar to do_generic_mapping_read(), we can't use that
 268  * since it has PAGE_SIZE assumptions.
 269  */
 270 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 271 {
 272         struct file *file = iocb->ki_filp;
 273         struct hstate *h = hstate_file(file);
 274         struct address_space *mapping = file->f_mapping;
 275         struct inode *inode = mapping->host;
 276         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 277         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 278         unsigned long end_index;
 279         loff_t isize;
 280         ssize_t retval = 0;
 281
 282         while (iov_iter_count(to)) {
 283                 struct page *page;
 284                 size_t nr, copied;
 285
 286                 /* nr is the maximum number of bytes to copy from this page */
 287                 nr = huge_page_size(h);
 288                 isize = i_size_read(inode);
 289                 if (!isize)
 290                         break;
 291                 end_index = (isize - 1) >> huge_page_shift(h);
 292                 if (index > end_index)
 293                         break;
 294                 if (index == end_index) {
 295                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 296                         if (nr <= offset)
 297                                 break;
 298                 }
 299                 nr = nr - offset;
 300
 301                 /* Find the page */
 302                 page = find_lock_page(mapping, index);
 303                 if (unlikely(page == NULL)) {
 304                         /*
 305                          * We have a HOLE, zero out the user-buffer for the
 306                          * length of the hole or request.
 307                          */
 308                         copied = iov_iter_zero(nr, to);
 309                 } else {
 310                         unlock_page(page);
 311
 312                         /*
 313                          * We have the page, copy it to user space buffer.
 314                          */
 315                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 316                         put_page(page);
 317                 }
 318                 offset += copied;
 319                 retval += copied;
 320                 if (copied != nr && iov_iter_count(to)) {
 321                         if (!retval)
 322                                 retval = -EFAULT;
 323                         break;
 324                 }
 325                 index += offset >> huge_page_shift(h);
 326                 offset &= ~huge_page_mask(h);
 327         }
 328         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 329         return retval;
 330 }
 331
 332 static int hugetlbfs_write_begin(struct file *file,
 333                         struct address_space *mapping,
 334                         loff_t pos, unsigned len, unsigned flags,
 335                         struct page **pagep, void **fsdata)
 336 {
 337         return -EINVAL;
 338 }
 339
 340 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 341                         loff_t pos, unsigned len, unsigned copied,
 342                         struct page *page, void *fsdata)
 343 {
 344         BUG();
 345         return -EINVAL;
 346 }
 347
 348 static void remove_huge_page(struct page *page)
 349 {
 350         ClearPageDirty(page);
 351         ClearPageUptodate(page);
 352         delete_from_page_cache(page);
 353 }
 354
 355 static void
 356 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
 357 {
 358         struct vm_area_struct *vma;
 359
 360         /*
 361          * end == 0 indicates that the entire range after
 362          * start should be unmapped.
 363          */
 364         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
 365                 unsigned long v_offset;
 366                 unsigned long v_end;
 367
 368                 /*
 369                  * Can the expression below overflow on 32-bit arches?
 370                  * No, because the interval tree returns us only those vmas
 371                  * which overlap the truncated area starting at pgoff,
 372                  * and no vma on a 32-bit arch can span beyond the 4GB.
 373                  */
 374                 if (vma->vm_pgoff < start)
 375                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 376                 else
 377                         v_offset = 0;
 378
 379                 if (!end)
 380                         v_end = vma->vm_end;
 381                 else {
 382                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 383                                                         + vma->vm_start;
 384                         if (v_end > vma->vm_end)
 385                                 v_end = vma->vm_end;
 386                 }
 387
 388                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 389                                                                         NULL);
 390         }
 391 }
 392
 393 /*
 394  * remove_inode_hugepages handles two distinct cases: truncation and hole
 395  * punch.  There are subtle differences in operation for each case.
 396  *
 397  * truncation is indicated by end of range being LLONG_MAX
 398  *      In this case, we first scan the range and release found pages.
 399  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
 400  *      maps and global counts.  Page faults can not race with truncation
 401  *      in this routine.  hugetlb_no_page() prevents page faults in the
 402  *      truncated range.  It checks i_size before allocation, and again after
 403  *      with the page table lock for the page held.  The same lock must be
 404  *      acquired to unmap a page.
 405  * hole punch is indicated if end is not LLONG_MAX
 406  *      In the hole punch case we scan the range and release found pages.
 407  *      Only when releasing a page is the associated region/reserv map
 408  *      deleted.  The region/reserv map for ranges without associated
 409  *      pages are not modified.  Page faults can race with hole punch.
 410  *      This is indicated if we find a mapped page.
 411  * Note: If the passed end of range value is beyond the end of file, but
 412  * not LLONG_MAX this routine still performs a hole punch operation.
 413  */
 414 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 415                                    loff_t lend)
 416 {
 417         struct hstate *h = hstate_inode(inode);
 418         struct address_space *mapping = &inode->i_data;
 419         const pgoff_t start = lstart >> huge_page_shift(h);
 420         const pgoff_t end = lend >> huge_page_shift(h);
 421         struct vm_area_struct pseudo_vma;
 422         struct pagevec pvec;
 423         pgoff_t next, index;
 424         int i, freed = 0;
 425         bool truncate_op = (lend == LLONG_MAX);
 426
 427         vma_init(&pseudo_vma, current->mm);
 428         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 429         pagevec_init(&pvec);
 430         next = start;
 431         while (next < end) {
 432                 /*
 433                  * When no more pages are found, we are done.
 434                  */
 435                 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
 436                         break;
 437
 438                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 439                         struct page *page = pvec.pages[i];
 440                         u32 hash;
 441
 442                         index = page->index;
 443                         hash = hugetlb_fault_mutex_hash(h, current->mm,
 444                                                         &pseudo_vma,
 445                                                         mapping, index, 0);
 446                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 447
 448                         /*
 449                          * If page is mapped, it was faulted in after being
 450                          * unmapped in caller.  Unmap (again) now after taking
 451                          * the fault mutex.  The mutex will prevent faults
 452                          * until we finish removing the page.
 453                          *
 454                          * This race can only happen in the hole punch case.
 455                          * Getting here in a truncate operation is a bug.
 456                          */
 457                         if (unlikely(page_mapped(page))) {
 458                                 BUG_ON(truncate_op);
 459
 460                                 i_mmap_lock_write(mapping);
 461                                 hugetlb_vmdelete_list(&mapping->i_mmap,
 462                                         index * pages_per_huge_page(h),
 463                                         (index + 1) * pages_per_huge_page(h));
 464                                 i_mmap_unlock_write(mapping);
 465                         }
 466
 467                         lock_page(page);
 468                         /*
 469                          * We must free the huge page and remove from page
 470                          * cache (remove_huge_page) BEFORE removing the
 471                          * region/reserve map (hugetlb_unreserve_pages).  In
 472                          * rare out of memory conditions, removal of the
 473                          * region/reserve map could fail. Correspondingly,
 474                          * the subpool and global reserve usage count can need
 475                          * to be adjusted.
 476                          */
 477                         VM_BUG_ON(PagePrivate(page));
 478                         remove_huge_page(page);
 479                         freed++;
 480                         if (!truncate_op) {
 481                                 if (unlikely(hugetlb_unreserve_pages(inode,
 482                                                         index, index + 1, 1)))
 483                                         hugetlb_fix_reserve_counts(inode);
 484                         }
 485
 486                         unlock_page(page);
 487                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 488                 }
 489                 huge_pagevec_release(&pvec);
 490                 cond_resched();
 491         }
 492
 493         if (truncate_op)
 494                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 495 }
 496
 497 static void hugetlbfs_evict_inode(struct inode *inode)
 498 {
 499         struct resv_map *resv_map;
 500
 501         remove_inode_hugepages(inode, 0, LLONG_MAX);
 502         resv_map = (struct resv_map *)inode->i_mapping->private_data;
 503         /* root inode doesn't have the resv_map, so we should check it */
 504         if (resv_map)
 505                 resv_map_release(&resv_map->refs);
 506         clear_inode(inode);
 507 }
 508
 509 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 510 {
 511         pgoff_t pgoff;
 512         struct address_space *mapping = inode->i_mapping;
 513         struct hstate *h = hstate_inode(inode);
 514
 515         BUG_ON(offset & ~huge_page_mask(h));
 516         pgoff = offset >> PAGE_SHIFT;
 517
 518         i_size_write(inode, offset);
 519         i_mmap_lock_write(mapping);
 520         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 521                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 522         i_mmap_unlock_write(mapping);
 523         remove_inode_hugepages(inode, offset, LLONG_MAX);
 524         return 0;
 525 }
 526
 527 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 528 {
 529         struct hstate *h = hstate_inode(inode);
 530         loff_t hpage_size = huge_page_size(h);
 531         loff_t hole_start, hole_end;
 532
 533         /*
 534          * For hole punch round up the beginning offset of the hole and
 535          * round down the end.
 536          */
 537         hole_start = round_up(offset, hpage_size);
 538         hole_end = round_down(offset + len, hpage_size);
 539
 540         if (hole_end > hole_start) {
 541                 struct address_space *mapping = inode->i_mapping;
 542                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 543
 544                 inode_lock(inode);
 545
 546                 /* protected by i_mutex */
 547                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 548                         inode_unlock(inode);
 549                         return -EPERM;
 550                 }
 551
 552                 i_mmap_lock_write(mapping);
 553                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 554                         hugetlb_vmdelete_list(&mapping->i_mmap,
 555                                                 hole_start >> PAGE_SHIFT,
 556                                                 hole_end  >> PAGE_SHIFT);
 557                 i_mmap_unlock_write(mapping);
 558                 remove_inode_hugepages(inode, hole_start, hole_end);
 559                 inode_unlock(inode);
 560         }
 561
 562         return 0;
 563 }
 564
 565 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 566                                 loff_t len)
 567 {
 568         struct inode *inode = file_inode(file);
 569         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 570         struct address_space *mapping = inode->i_mapping;
 571         struct hstate *h = hstate_inode(inode);
 572         struct vm_area_struct pseudo_vma;
 573         struct mm_struct *mm = current->mm;
 574         loff_t hpage_size = huge_page_size(h);
 575         unsigned long hpage_shift = huge_page_shift(h);
 576         pgoff_t start, index, end;
 577         int error;
 578         u32 hash;
 579
 580         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 581                 return -EOPNOTSUPP;
 582
 583         if (mode & FALLOC_FL_PUNCH_HOLE)
 584                 return hugetlbfs_punch_hole(inode, offset, len);
 585
 586         /*
 587          * Default preallocate case.
 588          * For this range, start is rounded down and end is rounded up
 589          * as well as being converted to page offsets.
 590          */
 591         start = offset >> hpage_shift;
 592         end = (offset + len + hpage_size - 1) >> hpage_shift;
 593
 594         inode_lock(inode);
 595
 596         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 597         error = inode_newsize_ok(inode, offset + len);
 598         if (error)
 599                 goto out;
 600
 601         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 602                 error = -EPERM;
 603                 goto out;
 604         }
 605
 606         /*
 607          * Initialize a pseudo vma as this is required by the huge page
 608          * allocation routines.  If NUMA is configured, use page index
 609          * as input to create an allocation policy.
 610          */
 611         vma_init(&pseudo_vma, mm);
 612         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 613         pseudo_vma.vm_file = file;
 614
 615         for (index = start; index < end; index++) {
 616                 /*
 617                  * This is supposed to be the vaddr where the page is being
 618                  * faulted in, but we have no vaddr here.
 619                  */
 620                 struct page *page;
 621                 unsigned long addr;
 622                 int avoid_reserve = 0;
 623
 624                 cond_resched();
 625
 626                 /*
 627                  * fallocate(2) manpage permits EINTR; we may have been
 628                  * interrupted because we are using up too much memory.
 629                  */
 630                 if (signal_pending(current)) {
 631                         error = -EINTR;
 632                         break;
 633                 }
 634
 635                 /* Set numa allocation policy based on index */
 636                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 637
 638                 /* addr is the offset within the file (zero based) */
 639                 addr = index * hpage_size;
 640
 641                 /* mutex taken here, fault path and hole punch */
 642                 hash = hugetlb_fault_mutex_hash(h, mm, &pseudo_vma, mapping,
 643                                                 index, addr);
 644                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 645
 646                 /* See if already present in mapping to avoid alloc/free */
 647                 page = find_get_page(mapping, index);
 648                 if (page) {
 649                         put_page(page);
 650                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 651                         hugetlb_drop_vma_policy(&pseudo_vma);
 652                         continue;
 653                 }
 654
 655                 /* Allocate page and add to page cache */
 656                 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
 657                 hugetlb_drop_vma_policy(&pseudo_vma);
 658                 if (IS_ERR(page)) {
 659                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 660                         error = PTR_ERR(page);
 661                         goto out;
 662                 }
 663                 clear_huge_page(page, addr, pages_per_huge_page(h));
 664                 __SetPageUptodate(page);
 665                 error = huge_add_to_page_cache(page, mapping, index);
 666                 if (unlikely(error)) {
 667                         put_page(page);
 668                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 669                         goto out;
 670                 }
 671
 672                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 673
 674                 /*
 675                  * unlock_page because locked by add_to_page_cache()
 676                  * page_put due to reference from alloc_huge_page()
 677                  */
 678                 unlock_page(page);
 679                 put_page(page);
 680         }
 681
 682         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 683                 i_size_write(inode, offset + len);
 684         inode->i_ctime = current_time(inode);
 685 out:
 686         inode_unlock(inode);
 687         return error;
 688 }
 689
 690 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 691 {
 692         struct inode *inode = d_inode(dentry);
 693         struct hstate *h = hstate_inode(inode);
 694         int error;
 695         unsigned int ia_valid = attr->ia_valid;
 696         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 697
 698         BUG_ON(!inode);
 699
 700         error = setattr_prepare(dentry, attr);
 701         if (error)
 702                 return error;
 703
 704         if (ia_valid & ATTR_SIZE) {
 705                 loff_t oldsize = inode->i_size;
 706                 loff_t newsize = attr->ia_size;
 707
 708                 if (newsize & ~huge_page_mask(h))
 709                         return -EINVAL;
 710                 /* protected by i_mutex */
 711                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 712                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 713                         return -EPERM;
 714                 error = hugetlb_vmtruncate(inode, newsize);
 715                 if (error)
 716                         return error;
 717         }
 718
 719         setattr_copy(inode, attr);
 720         mark_inode_dirty(inode);
 721         return 0;
 722 }
 723
 724 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 725                                         struct hugetlbfs_fs_context *ctx)
 726 {
 727         struct inode *inode;
 728
 729         inode = new_inode(sb);
 730         if (inode) {
 731                 inode->i_ino = get_next_ino();
 732                 inode->i_mode = S_IFDIR | ctx->mode;
 733                 inode->i_uid = ctx->uid;
 734                 inode->i_gid = ctx->gid;
 735                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 736                 inode->i_op = &hugetlbfs_dir_inode_operations;
 737                 inode->i_fop = &simple_dir_operations;
 738                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 739                 inc_nlink(inode);
 740                 lockdep_annotate_inode_mutex_key(inode);
 741         }
 742         return inode;
 743 }
 744
 745 /*
 746  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 747  * be taken from reclaim -- unlike regular filesystems. This needs an
 748  * annotation because huge_pmd_share() does an allocation under hugetlb's
 749  * i_mmap_rwsem.
 750  */
 751 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 752
 753 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 754                                         struct inode *dir,
 755                                         umode_t mode, dev_t dev)
 756 {
 757         struct inode *inode;
 758         struct resv_map *resv_map = NULL;
 759
 760         /*
 761          * Reserve maps are only needed for inodes that can have associated
 762          * page allocations.
 763          */
 764         if (S_ISREG(mode) || S_ISLNK(mode)) {
 765                 resv_map = resv_map_alloc();
 766                 if (!resv_map)
 767                         return NULL;
 768         }
 769
 770         inode = new_inode(sb);
 771         if (inode) {
 772                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 773
 774                 inode->i_ino = get_next_ino();
 775                 inode_init_owner(inode, dir, mode);
 776                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 777                                 &hugetlbfs_i_mmap_rwsem_key);
 778                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 779                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 780                 inode->i_mapping->private_data = resv_map;
 781                 info->seals = F_SEAL_SEAL;
 782                 switch (mode & S_IFMT) {
 783                 default:
 784                         init_special_inode(inode, mode, dev);
 785                         break;
 786                 case S_IFREG:
 787                         inode->i_op = &hugetlbfs_inode_operations;
 788                         inode->i_fop = &hugetlbfs_file_operations;
 789                         break;
 790                 case S_IFDIR:
 791                         inode->i_op = &hugetlbfs_dir_inode_operations;
 792                         inode->i_fop = &simple_dir_operations;
 793
 794                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 795                         inc_nlink(inode);
 796                         break;
 797                 case S_IFLNK:
 798                         inode->i_op = &page_symlink_inode_operations;
 799                         inode_nohighmem(inode);
 800                         break;
 801                 }
 802                 lockdep_annotate_inode_mutex_key(inode);
 803         } else {
 804                 if (resv_map)
 805                         kref_put(&resv_map->refs, resv_map_release);
 806         }
 807
 808         return inode;
 809 }
 810
 811 /*
 812  * File creation. Allocate an inode, and we're done..
 813  */
 814 static int hugetlbfs_mknod(struct inode *dir,
 815                         struct dentry *dentry, umode_t mode, dev_t dev)
 816 {
 817         struct inode *inode;
 818         int error = -ENOSPC;
 819
 820         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 821         if (inode) {
 822                 dir->i_ctime = dir->i_mtime = current_time(dir);
 823                 d_instantiate(dentry, inode);
 824                 dget(dentry);   /* Extra count - pin the dentry in core */
 825                 error = 0;
 826         }
 827         return error;
 828 }
 829
 830 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 831 {
 832         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
 833         if (!retval)
 834                 inc_nlink(dir);
 835         return retval;
 836 }
 837
 838 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
 839 {
 840         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
 841 }
 842
 843 static int hugetlbfs_symlink(struct inode *dir,
 844                         struct dentry *dentry, const char *symname)
 845 {
 846         struct inode *inode;
 847         int error = -ENOSPC;
 848
 849         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 850         if (inode) {
 851                 int l = strlen(symname)+1;
 852                 error = page_symlink(inode, symname, l);
 853                 if (!error) {
 854                         d_instantiate(dentry, inode);
 855                         dget(dentry);
 856                 } else
 857                         iput(inode);
 858         }
 859         dir->i_ctime = dir->i_mtime = current_time(dir);
 860
 861         return error;
 862 }
 863
 864 /*
 865  * mark the head page dirty
 866  */
 867 static int hugetlbfs_set_page_dirty(struct page *page)
 868 {
 869         struct page *head = compound_head(page);
 870
 871         SetPageDirty(head);
 872         return 0;
 873 }
 874
 875 static int hugetlbfs_migrate_page(struct address_space *mapping,
 876                                 struct page *newpage, struct page *page,
 877                                 enum migrate_mode mode)
 878 {
 879         int rc;
 880
 881         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 882         if (rc != MIGRATEPAGE_SUCCESS)
 883                 return rc;
 884
 885         /*
 886          * page_private is subpool pointer in hugetlb pages.  Transfer to
 887          * new page.  PagePrivate is not associated with page_private for
 888          * hugetlb pages and can not be set here as only page_huge_active
 889          * pages can be migrated.
 890          */
 891         if (page_private(page)) {
 892                 set_page_private(newpage, page_private(page));
 893                 set_page_private(page, 0);
 894         }
 895
 896         if (mode != MIGRATE_SYNC_NO_COPY)
 897                 migrate_page_copy(newpage, page);
 898         else
 899                 migrate_page_states(newpage, page);
 900
 901         return MIGRATEPAGE_SUCCESS;
 902 }
 903
 904 static int hugetlbfs_error_remove_page(struct address_space *mapping,
 905                                 struct page *page)
 906 {
 907         struct inode *inode = mapping->host;
 908         pgoff_t index = page->index;
 909
 910         remove_huge_page(page);
 911         if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
 912                 hugetlb_fix_reserve_counts(inode);
 913
 914         return 0;
 915 }
 916
 917 /*
 918  * Display the mount options in /proc/mounts.
 919  */
 920 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
 921 {
 922         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
 923         struct hugepage_subpool *spool = sbinfo->spool;
 924         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
 925         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
 926         char mod;
 927
 928         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
 929                 seq_printf(m, ",uid=%u",
 930                            from_kuid_munged(&init_user_ns, sbinfo->uid));
 931         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
 932                 seq_printf(m, ",gid=%u",
 933                            from_kgid_munged(&init_user_ns, sbinfo->gid));
 934         if (sbinfo->mode != 0755)
 935                 seq_printf(m, ",mode=%o", sbinfo->mode);
 936         if (sbinfo->max_inodes != -1)
 937                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
 938
 939         hpage_size /= 1024;
 940         mod = 'K';
 941         if (hpage_size >= 1024) {
 942                 hpage_size /= 1024;
 943                 mod = 'M';
 944         }
 945         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
 946         if (spool) {
 947                 if (spool->max_hpages != -1)
 948                         seq_printf(m, ",size=%llu",
 949                                    (unsigned long long)spool->max_hpages << hpage_shift);
 950                 if (spool->min_hpages != -1)
 951                         seq_printf(m, ",min_size=%llu",
 952                                    (unsigned long long)spool->min_hpages << hpage_shift);
 953         }
 954         return 0;
 955 }
 956
 957 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 958 {
 959         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
 960         struct hstate *h = hstate_inode(d_inode(dentry));
 961
 962         buf->f_type = HUGETLBFS_MAGIC;
 963         buf->f_bsize = huge_page_size(h);
 964         if (sbinfo) {
 965                 spin_lock(&sbinfo->stat_lock);
 966                 /* If no limits set, just report 0 for max/free/used
 967                  * blocks, like simple_statfs() */
 968                 if (sbinfo->spool) {
 969                         long free_pages;
 970
 971                         spin_lock(&sbinfo->spool->lock);
 972                         buf->f_blocks = sbinfo->spool->max_hpages;
 973                         free_pages = sbinfo->spool->max_hpages
 974                                 - sbinfo->spool->used_hpages;
 975                         buf->f_bavail = buf->f_bfree = free_pages;
 976                         spin_unlock(&sbinfo->spool->lock);
 977                         buf->f_files = sbinfo->max_inodes;
 978                         buf->f_ffree = sbinfo->free_inodes;
 979                 }
 980                 spin_unlock(&sbinfo->stat_lock);
 981         }
 982         buf->f_namelen = NAME_MAX;
 983         return 0;
 984 }
 985
 986 static void hugetlbfs_put_super(struct super_block *sb)
 987 {
 988         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
 989
 990         if (sbi) {
 991                 sb->s_fs_info = NULL;
 992
 993                 if (sbi->spool)
 994                         hugepage_put_subpool(sbi->spool);
 995
 996                 kfree(sbi);
 997         }
 998 }
 999
1000 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1001 {
1002         if (sbinfo->free_inodes >= 0) {
1003                 spin_lock(&sbinfo->stat_lock);
1004                 if (unlikely(!sbinfo->free_inodes)) {
1005                         spin_unlock(&sbinfo->stat_lock);
1006                         return 0;
1007                 }
1008                 sbinfo->free_inodes--;
1009                 spin_unlock(&sbinfo->stat_lock);
1010         }
1011
1012         return 1;
1013 }
1014
1015 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1016 {
1017         if (sbinfo->free_inodes >= 0) {
1018                 spin_lock(&sbinfo->stat_lock);
1019                 sbinfo->free_inodes++;
1020                 spin_unlock(&sbinfo->stat_lock);
1021         }
1022 }
1023
1024
1025 static struct kmem_cache *hugetlbfs_inode_cachep;
1026
1027 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1028 {
1029         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1030         struct hugetlbfs_inode_info *p;
1031
1032         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1033                 return NULL;
1034         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1035         if (unlikely(!p)) {
1036                 hugetlbfs_inc_free_inodes(sbinfo);
1037                 return NULL;
1038         }
1039
1040         /*
1041          * Any time after allocation, hugetlbfs_destroy_inode can be called
1042          * for the inode.  mpol_free_shared_policy is unconditionally called
1043          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1044          * in case of a quick call to destroy.
1045          *
1046          * Note that the policy is initialized even if we are creating a
1047          * private inode.  This simplifies hugetlbfs_destroy_inode.
1048          */
1049         mpol_shared_policy_init(&p->policy, NULL);
1050
1051         return &p->vfs_inode;
1052 }
1053
1054 static void hugetlbfs_i_callback(struct rcu_head *head)
1055 {
1056         struct inode *inode = container_of(head, struct inode, i_rcu);
1057         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1058 }
1059
1060 static void hugetlbfs_destroy_inode(struct inode *inode)
1061 {
1062         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1063         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1064         call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
1065 }
1066
1067 static const struct address_space_operations hugetlbfs_aops = {
1068         .write_begin    = hugetlbfs_write_begin,
1069         .write_end      = hugetlbfs_write_end,
1070         .set_page_dirty = hugetlbfs_set_page_dirty,
1071         .migratepage    = hugetlbfs_migrate_page,
1072         .error_remove_page      = hugetlbfs_error_remove_page,
1073 };
1074
1075
1076 static void init_once(void *foo)
1077 {
1078         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1079
1080         inode_init_once(&ei->vfs_inode);
1081 }
1082
1083 const struct file_operations hugetlbfs_file_operations = {
1084         .read_iter              = hugetlbfs_read_iter,
1085         .mmap                   = hugetlbfs_file_mmap,
1086         .fsync                  = noop_fsync,
1087         .get_unmapped_area      = hugetlb_get_unmapped_area,
1088         .llseek                 = default_llseek,
1089         .fallocate              = hugetlbfs_fallocate,
1090 };
1091
1092 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1093         .create         = hugetlbfs_create,
1094         .lookup         = simple_lookup,
1095         .link           = simple_link,
1096         .unlink         = simple_unlink,
1097         .symlink        = hugetlbfs_symlink,
1098         .mkdir          = hugetlbfs_mkdir,
1099         .rmdir          = simple_rmdir,
1100         .mknod          = hugetlbfs_mknod,
1101         .rename         = simple_rename,
1102         .setattr        = hugetlbfs_setattr,
1103 };
1104
1105 static const struct inode_operations hugetlbfs_inode_operations = {
1106         .setattr        = hugetlbfs_setattr,
1107 };
1108
1109 static const struct super_operations hugetlbfs_ops = {
1110         .alloc_inode    = hugetlbfs_alloc_inode,
1111         .destroy_inode  = hugetlbfs_destroy_inode,
1112         .evict_inode    = hugetlbfs_evict_inode,
1113         .statfs         = hugetlbfs_statfs,
1114         .put_super      = hugetlbfs_put_super,
1115         .show_options   = hugetlbfs_show_options,
1116 };
1117
1118 /*
1119  * Convert size option passed from command line to number of huge pages
1120  * in the pool specified by hstate.  Size option could be in bytes
1121  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1122  */
1123 static long
1124 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1125                          enum hugetlbfs_size_type val_type)
1126 {
1127         if (val_type == NO_SIZE)
1128                 return -1;
1129
1130         if (val_type == SIZE_PERCENT) {
1131                 size_opt <<= huge_page_shift(h);
1132                 size_opt *= h->max_huge_pages;
1133                 do_div(size_opt, 100);
1134         }
1135
1136         size_opt >>= huge_page_shift(h);
1137         return size_opt;
1138 }
1139
1140 /*
1141  * Parse one mount parameter.
1142  */
1143 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1144 {
1145         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1146         struct fs_parse_result result;
1147         char *rest;
1148         unsigned long ps;
1149         int opt;
1150
1151         opt = fs_parse(fc, &hugetlb_fs_parameters, param, &result);
1152         if (opt < 0)
1153                 return opt;
1154
1155         switch (opt) {
1156         case Opt_uid:
1157                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1158                 if (!uid_valid(ctx->uid))
1159                         goto bad_val;
1160                 return 0;
1161
1162         case Opt_gid:
1163                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1164                 if (!gid_valid(ctx->gid))
1165                         goto bad_val;
1166                 return 0;
1167
1168         case Opt_mode:
1169                 ctx->mode = result.uint_32 & 01777U;
1170                 return 0;
1171
1172         case Opt_size:
1173                 /* memparse() will accept a K/M/G without a digit */
1174                 if (!isdigit(param->string[0]))
1175                         goto bad_val;
1176                 ctx->max_size_opt = memparse(param->string, &rest);
1177                 ctx->max_val_type = SIZE_STD;
1178                 if (*rest == '%')
1179                         ctx->max_val_type = SIZE_PERCENT;
1180                 return 0;
1181
1182         case Opt_nr_inodes:
1183                 /* memparse() will accept a K/M/G without a digit */
1184                 if (!isdigit(param->string[0]))
1185                         goto bad_val;
1186                 ctx->nr_inodes = memparse(param->string, &rest);
1187                 return 0;
1188
1189         case Opt_pagesize:
1190                 ps = memparse(param->string, &rest);
1191                 ctx->hstate = size_to_hstate(ps);
1192                 if (!ctx->hstate) {
1193                         pr_err("Unsupported page size %lu MB\n", ps >> 20);
1194                         return -EINVAL;
1195                 }
1196                 return 0;
1197
1198         case Opt_min_size:
1199                 /* memparse() will accept a K/M/G without a digit */
1200                 if (!isdigit(param->string[0]))
1201                         goto bad_val;
1202                 ctx->min_size_opt = memparse(param->string, &rest);
1203                 ctx->min_val_type = SIZE_STD;
1204                 if (*rest == '%')
1205                         ctx->min_val_type = SIZE_PERCENT;
1206                 return 0;
1207
1208         default:
1209                 return -EINVAL;
1210         }
1211
1212 bad_val:
1213         return invalf(fc, "hugetlbfs: Bad value '%s' for mount option '%s'\n",
1214                       param->string, param->key);
1215 }
1216
1217 /*
1218  * Validate the parsed options.
1219  */
1220 static int hugetlbfs_validate(struct fs_context *fc)
1221 {
1222         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1223
1224         /*
1225          * Use huge page pool size (in hstate) to convert the size
1226          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1227          */
1228         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1229                                                    ctx->max_size_opt,
1230                                                    ctx->max_val_type);
1231         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1232                                                    ctx->min_size_opt,
1233                                                    ctx->min_val_type);
1234
1235         /*
1236          * If max_size was specified, then min_size must be smaller
1237          */
1238         if (ctx->max_val_type > NO_SIZE &&
1239             ctx->min_hpages > ctx->max_hpages) {
1240                 pr_err("Minimum size can not be greater than maximum size\n");
1241                 return -EINVAL;
1242         }
1243
1244         return 0;
1245 }
1246
1247 static int
1248 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1249 {
1250         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1251         struct hugetlbfs_sb_info *sbinfo;
1252
1253         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1254         if (!sbinfo)
1255                 return -ENOMEM;
1256         sb->s_fs_info = sbinfo;
1257         spin_lock_init(&sbinfo->stat_lock);
1258         sbinfo->hstate          = ctx->hstate;
1259         sbinfo->max_inodes      = ctx->nr_inodes;
1260         sbinfo->free_inodes     = ctx->nr_inodes;
1261         sbinfo->spool           = NULL;
1262         sbinfo->uid             = ctx->uid;
1263         sbinfo->gid             = ctx->gid;
1264         sbinfo->mode            = ctx->mode;
1265
1266         /*
1267          * Allocate and initialize subpool if maximum or minimum size is
1268          * specified.  Any needed reservations (for minimim size) are taken
1269          * taken when the subpool is created.
1270          */
1271         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1272                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1273                                                      ctx->max_hpages,
1274                                                      ctx->min_hpages);
1275                 if (!sbinfo->spool)
1276                         goto out_free;
1277         }
1278         sb->s_maxbytes = MAX_LFS_FILESIZE;
1279         sb->s_blocksize = huge_page_size(ctx->hstate);
1280         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1281         sb->s_magic = HUGETLBFS_MAGIC;
1282         sb->s_op = &hugetlbfs_ops;
1283         sb->s_time_gran = 1;
1284         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1285         if (!sb->s_root)
1286                 goto out_free;
1287         return 0;
1288 out_free:
1289         kfree(sbinfo->spool);
1290         kfree(sbinfo);
1291         return -ENOMEM;
1292 }
1293
1294 static int hugetlbfs_get_tree(struct fs_context *fc)
1295 {
1296         int err = hugetlbfs_validate(fc);
1297         if (err)
1298                 return err;
1299         return vfs_get_super(fc, vfs_get_independent_super, hugetlbfs_fill_super);
1300 }
1301
1302 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1303 {
1304         kfree(fc->fs_private);
1305 }
1306
1307 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1308         .free           = hugetlbfs_fs_context_free,
1309         .parse_param    = hugetlbfs_parse_param,
1310         .get_tree       = hugetlbfs_get_tree,
1311 };
1312
1313 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1314 {
1315         struct hugetlbfs_fs_context *ctx;
1316
1317         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1318         if (!ctx)
1319                 return -ENOMEM;
1320
1321         ctx->max_hpages = -1; /* No limit on size by default */
1322         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1323         ctx->uid        = current_fsuid();
1324         ctx->gid        = current_fsgid();
1325         ctx->mode       = 0755;
1326         ctx->hstate     = &default_hstate;
1327         ctx->min_hpages = -1; /* No default minimum size */
1328         ctx->max_val_type = NO_SIZE;
1329         ctx->min_val_type = NO_SIZE;
1330         fc->fs_private = ctx;
1331         fc->ops = &hugetlbfs_fs_context_ops;
1332         return 0;
1333 }
1334
1335 static struct file_system_type hugetlbfs_fs_type = {
1336         .name                   = "hugetlbfs",
1337         .init_fs_context        = hugetlbfs_init_fs_context,
1338         .parameters             = &hugetlb_fs_parameters,
1339         .kill_sb                = kill_litter_super,
1340 };
1341
1342 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1343
1344 static int can_do_hugetlb_shm(void)
1345 {
1346         kgid_t shm_group;
1347         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1348         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1349 }
1350
1351 static int get_hstate_idx(int page_size_log)
1352 {
1353         struct hstate *h = hstate_sizelog(page_size_log);
1354
1355         if (!h)
1356                 return -1;
1357         return h - hstates;
1358 }
1359
1360 /*
1361  * Note that size should be aligned to proper hugepage size in caller side,
1362  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1363  */
1364 struct file *hugetlb_file_setup(const char *name, size_t size,
1365                                 vm_flags_t acctflag, struct user_struct **user,
1366                                 int creat_flags, int page_size_log)
1367 {
1368         struct inode *inode;
1369         struct vfsmount *mnt;
1370         int hstate_idx;
1371         struct file *file;
1372
1373         hstate_idx = get_hstate_idx(page_size_log);
1374         if (hstate_idx < 0)
1375                 return ERR_PTR(-ENODEV);
1376
1377         *user = NULL;
1378         mnt = hugetlbfs_vfsmount[hstate_idx];
1379         if (!mnt)
1380                 return ERR_PTR(-ENOENT);
1381
1382         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1383                 *user = current_user();
1384                 if (user_shm_lock(size, *user)) {
1385                         task_lock(current);
1386                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1387                                 current->comm, current->pid);
1388                         task_unlock(current);
1389                 } else {
1390                         *user = NULL;
1391                         return ERR_PTR(-EPERM);
1392                 }
1393         }
1394
1395         file = ERR_PTR(-ENOSPC);
1396         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1397         if (!inode)
1398                 goto out;
1399         if (creat_flags == HUGETLB_SHMFS_INODE)
1400                 inode->i_flags |= S_PRIVATE;
1401
1402         inode->i_size = size;
1403         clear_nlink(inode);
1404
1405         if (hugetlb_reserve_pages(inode, 0,
1406                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1407                         acctflag))
1408                 file = ERR_PTR(-ENOMEM);
1409         else
1410                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1411                                         &hugetlbfs_file_operations);
1412         if (!IS_ERR(file))
1413                 return file;
1414
1415         iput(inode);
1416 out:
1417         if (*user) {
1418                 user_shm_unlock(size, *user);
1419                 *user = NULL;
1420         }
1421         return file;
1422 }
1423
1424 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1425 {
1426         struct fs_context *fc;
1427         struct vfsmount *mnt;
1428
1429         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1430         if (IS_ERR(fc)) {
1431                 mnt = ERR_CAST(fc);
1432         } else {
1433                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1434                 ctx->hstate = h;
1435                 mnt = fc_mount(fc);
1436                 put_fs_context(fc);
1437         }
1438         if (IS_ERR(mnt))
1439                 pr_err("Cannot mount internal hugetlbfs for page size %uK",
1440                        1U << (h->order + PAGE_SHIFT - 10));
1441         return mnt;
1442 }
1443
1444 static int __init init_hugetlbfs_fs(void)
1445 {
1446         struct vfsmount *mnt;
1447         struct hstate *h;
1448         int error;
1449         int i;
1450
1451         if (!hugepages_supported()) {
1452                 pr_info("disabling because there are no supported hugepage sizes\n");
1453                 return -ENOTSUPP;
1454         }
1455
1456         error = -ENOMEM;
1457         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1458                                         sizeof(struct hugetlbfs_inode_info),
1459                                         0, SLAB_ACCOUNT, init_once);
1460         if (hugetlbfs_inode_cachep == NULL)
1461                 goto out2;
1462
1463         error = register_filesystem(&hugetlbfs_fs_type);
1464         if (error)
1465                 goto out;
1466
1467         i = 0;
1468         for_each_hstate(h) {
1469                 mnt = mount_one_hugetlbfs(h);
1470                 if (IS_ERR(mnt) && i == 0) {
1471                         error = PTR_ERR(mnt);
1472                         goto out;
1473                 }
1474                 hugetlbfs_vfsmount[i] = mnt;
1475                 i++;
1476         }
1477
1478         return 0;
1479
1480  out:
1481         kmem_cache_destroy(hugetlbfs_inode_cachep);
1482  out2:
1483         return error;
1484 }
1485 fs_initcall(init_hugetlbfs_fs)