fs/hugetlbfs/inode.c

   1 /*
   2  * hugetlbpage-backed filesystem.  Based on ramfs.
   3  *
   4  * Nadia Yvette Chambers, 2002
   5  *
   6  * Copyright (C) 2002 Linus Torvalds.
   7  * License: GPL
   8  */
   9
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/thread_info.h>
  13 #include <asm/current.h>
  14 #include <linux/sched/signal.h>         /* remove ASAP */
  15 #include <linux/falloc.h>
  16 #include <linux/fs.h>
  17 #include <linux/mount.h>
  18 #include <linux/file.h>
  19 #include <linux/kernel.h>
  20 #include <linux/writeback.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/highmem.h>
  23 #include <linux/init.h>
  24 #include <linux/string.h>
  25 #include <linux/capability.h>
  26 #include <linux/ctype.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/hugetlb.h>
  29 #include <linux/pagevec.h>
  30 #include <linux/parser.h>
  31 #include <linux/mman.h>
  32 #include <linux/slab.h>
  33 #include <linux/dnotify.h>
  34 #include <linux/statfs.h>
  35 #include <linux/security.h>
  36 #include <linux/magic.h>
  37 #include <linux/migrate.h>
  38 #include <linux/uio.h>
  39
  40 #include <linux/uaccess.h>
  41
  42 static const struct super_operations hugetlbfs_ops;
  43 static const struct address_space_operations hugetlbfs_aops;
  44 const struct file_operations hugetlbfs_file_operations;
  45 static const struct inode_operations hugetlbfs_dir_inode_operations;
  46 static const struct inode_operations hugetlbfs_inode_operations;
  47
  48 struct hugetlbfs_config {
  49         struct hstate           *hstate;
  50         long                    max_hpages;
  51         long                    nr_inodes;
  52         long                    min_hpages;
  53         kuid_t                  uid;
  54         kgid_t                  gid;
  55         umode_t                 mode;
  56 };
  57
  58 int sysctl_hugetlb_shm_group;
  59
  60 enum {
  61         Opt_size, Opt_nr_inodes,
  62         Opt_mode, Opt_uid, Opt_gid,
  63         Opt_pagesize, Opt_min_size,
  64         Opt_err,
  65 };
  66
  67 static const match_table_t tokens = {
  68         {Opt_size,      "size=%s"},
  69         {Opt_nr_inodes, "nr_inodes=%s"},
  70         {Opt_mode,      "mode=%o"},
  71         {Opt_uid,       "uid=%u"},
  72         {Opt_gid,       "gid=%u"},
  73         {Opt_pagesize,  "pagesize=%s"},
  74         {Opt_min_size,  "min_size=%s"},
  75         {Opt_err,       NULL},
  76 };
  77
  78 #ifdef CONFIG_NUMA
  79 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  80                                         struct inode *inode, pgoff_t index)
  81 {
  82         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
  83                                                         index);
  84 }
  85
  86 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  87 {
  88         mpol_cond_put(vma->vm_policy);
  89 }
  90 #else
  91 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  92                                         struct inode *inode, pgoff_t index)
  93 {
  94 }
  95
  96 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  97 {
  98 }
  99 #endif
 100
 101 static void huge_pagevec_release(struct pagevec *pvec)
 102 {
 103         int i;
 104
 105         for (i = 0; i < pagevec_count(pvec); ++i)
 106                 put_page(pvec->pages[i]);
 107
 108         pagevec_reinit(pvec);
 109 }
 110
 111 /*
 112  * Mask used when checking the page offset value passed in via system
 113  * calls.  This value will be converted to a loff_t which is signed.
 114  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 115  * value.  The extra bit (- 1 in the shift value) is to take the sign
 116  * bit into account.
 117  */
 118 #define PGOFF_LOFFT_MAX \
 119         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 120
 121 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 122 {
 123         struct inode *inode = file_inode(file);
 124         loff_t len, vma_len;
 125         int ret;
 126         struct hstate *h = hstate_file(file);
 127
 128         /*
 129          * vma address alignment (but not the pgoff alignment) has
 130          * already been checked by prepare_hugepage_range.  If you add
 131          * any error returns here, do so after setting VM_HUGETLB, so
 132          * is_vm_hugetlb_page tests below unmap_region go the right
 133          * way when do_mmap_pgoff unwinds (may be important on powerpc
 134          * and ia64).
 135          */
 136         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 137         vma->vm_ops = &hugetlb_vm_ops;
 138
 139         /*
 140          * page based offset in vm_pgoff could be sufficiently large to
 141          * overflow a loff_t when converted to byte offset.  This can
 142          * only happen on architectures where sizeof(loff_t) ==
 143          * sizeof(unsigned long).  So, only check in those instances.
 144          */
 145         if (sizeof(unsigned long) == sizeof(loff_t)) {
 146                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 147                         return -EINVAL;
 148         }
 149
 150         /* must be huge page aligned */
 151         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 152                 return -EINVAL;
 153
 154         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 155         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 156         /* check for overflow */
 157         if (len < vma_len)
 158                 return -EINVAL;
 159
 160         inode_lock(inode);
 161         file_accessed(file);
 162
 163         ret = -ENOMEM;
 164         if (hugetlb_reserve_pages(inode,
 165                                 vma->vm_pgoff >> huge_page_order(h),
 166                                 len >> huge_page_shift(h), vma,
 167                                 vma->vm_flags))
 168                 goto out;
 169
 170         ret = 0;
 171         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 172                 i_size_write(inode, len);
 173 out:
 174         inode_unlock(inode);
 175
 176         return ret;
 177 }
 178
 179 /*
 180  * Called under down_write(mmap_sem).
 181  */
 182
 183 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 184 static unsigned long
 185 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 186                 unsigned long len, unsigned long pgoff, unsigned long flags)
 187 {
 188         struct mm_struct *mm = current->mm;
 189         struct vm_area_struct *vma;
 190         struct hstate *h = hstate_file(file);
 191         struct vm_unmapped_area_info info;
 192
 193         if (len & ~huge_page_mask(h))
 194                 return -EINVAL;
 195         if (len > TASK_SIZE)
 196                 return -ENOMEM;
 197
 198         if (flags & MAP_FIXED) {
 199                 if (prepare_hugepage_range(file, addr, len))
 200                         return -EINVAL;
 201                 return addr;
 202         }
 203
 204         if (addr) {
 205                 addr = ALIGN(addr, huge_page_size(h));
 206                 vma = find_vma(mm, addr);
 207                 if (TASK_SIZE - len >= addr &&
 208                     (!vma || addr + len <= vm_start_gap(vma)))
 209                         return addr;
 210         }
 211
 212         info.flags = 0;
 213         info.length = len;
 214         info.low_limit = TASK_UNMAPPED_BASE;
 215         info.high_limit = TASK_SIZE;
 216         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 217         info.align_offset = 0;
 218         return vm_unmapped_area(&info);
 219 }
 220 #endif
 221
 222 static size_t
 223 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 224                         struct iov_iter *to, unsigned long size)
 225 {
 226         size_t copied = 0;
 227         int i, chunksize;
 228
 229         /* Find which 4k chunk and offset with in that chunk */
 230         i = offset >> PAGE_SHIFT;
 231         offset = offset & ~PAGE_MASK;
 232
 233         while (size) {
 234                 size_t n;
 235                 chunksize = PAGE_SIZE;
 236                 if (offset)
 237                         chunksize -= offset;
 238                 if (chunksize > size)
 239                         chunksize = size;
 240                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 241                 copied += n;
 242                 if (n != chunksize)
 243                         return copied;
 244                 offset = 0;
 245                 size -= chunksize;
 246                 i++;
 247         }
 248         return copied;
 249 }
 250
 251 /*
 252  * Support for read() - Find the page attached to f_mapping and copy out the
 253  * data. Its *very* similar to do_generic_mapping_read(), we can't use that
 254  * since it has PAGE_SIZE assumptions.
 255  */
 256 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 257 {
 258         struct file *file = iocb->ki_filp;
 259         struct hstate *h = hstate_file(file);
 260         struct address_space *mapping = file->f_mapping;
 261         struct inode *inode = mapping->host;
 262         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 263         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 264         unsigned long end_index;
 265         loff_t isize;
 266         ssize_t retval = 0;
 267
 268         while (iov_iter_count(to)) {
 269                 struct page *page;
 270                 size_t nr, copied;
 271
 272                 /* nr is the maximum number of bytes to copy from this page */
 273                 nr = huge_page_size(h);
 274                 isize = i_size_read(inode);
 275                 if (!isize)
 276                         break;
 277                 end_index = (isize - 1) >> huge_page_shift(h);
 278                 if (index > end_index)
 279                         break;
 280                 if (index == end_index) {
 281                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 282                         if (nr <= offset)
 283                                 break;
 284                 }
 285                 nr = nr - offset;
 286
 287                 /* Find the page */
 288                 page = find_lock_page(mapping, index);
 289                 if (unlikely(page == NULL)) {
 290                         /*
 291                          * We have a HOLE, zero out the user-buffer for the
 292                          * length of the hole or request.
 293                          */
 294                         copied = iov_iter_zero(nr, to);
 295                 } else {
 296                         unlock_page(page);
 297
 298                         /*
 299                          * We have the page, copy it to user space buffer.
 300                          */
 301                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 302                         put_page(page);
 303                 }
 304                 offset += copied;
 305                 retval += copied;
 306                 if (copied != nr && iov_iter_count(to)) {
 307                         if (!retval)
 308                                 retval = -EFAULT;
 309                         break;
 310                 }
 311                 index += offset >> huge_page_shift(h);
 312                 offset &= ~huge_page_mask(h);
 313         }
 314         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 315         return retval;
 316 }
 317
 318 static int hugetlbfs_write_begin(struct file *file,
 319                         struct address_space *mapping,
 320                         loff_t pos, unsigned len, unsigned flags,
 321                         struct page **pagep, void **fsdata)
 322 {
 323         return -EINVAL;
 324 }
 325
 326 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 327                         loff_t pos, unsigned len, unsigned copied,
 328                         struct page *page, void *fsdata)
 329 {
 330         BUG();
 331         return -EINVAL;
 332 }
 333
 334 static void remove_huge_page(struct page *page)
 335 {
 336         ClearPageDirty(page);
 337         ClearPageUptodate(page);
 338         delete_from_page_cache(page);
 339 }
 340
 341 static void
 342 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
 343 {
 344         struct vm_area_struct *vma;
 345
 346         /*
 347          * end == 0 indicates that the entire range after
 348          * start should be unmapped.
 349          */
 350         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
 351                 unsigned long v_offset;
 352                 unsigned long v_end;
 353
 354                 /*
 355                  * Can the expression below overflow on 32-bit arches?
 356                  * No, because the interval tree returns us only those vmas
 357                  * which overlap the truncated area starting at pgoff,
 358                  * and no vma on a 32-bit arch can span beyond the 4GB.
 359                  */
 360                 if (vma->vm_pgoff < start)
 361                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 362                 else
 363                         v_offset = 0;
 364
 365                 if (!end)
 366                         v_end = vma->vm_end;
 367                 else {
 368                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 369                                                         + vma->vm_start;
 370                         if (v_end > vma->vm_end)
 371                                 v_end = vma->vm_end;
 372                 }
 373
 374                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 375                                                                         NULL);
 376         }
 377 }
 378
 379 /*
 380  * remove_inode_hugepages handles two distinct cases: truncation and hole
 381  * punch.  There are subtle differences in operation for each case.
 382  *
 383  * truncation is indicated by end of range being LLONG_MAX
 384  *      In this case, we first scan the range and release found pages.
 385  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
 386  *      maps and global counts.  Page faults can not race with truncation
 387  *      in this routine.  hugetlb_no_page() prevents page faults in the
 388  *      truncated range.  It checks i_size before allocation, and again after
 389  *      with the page table lock for the page held.  The same lock must be
 390  *      acquired to unmap a page.
 391  * hole punch is indicated if end is not LLONG_MAX
 392  *      In the hole punch case we scan the range and release found pages.
 393  *      Only when releasing a page is the associated region/reserv map
 394  *      deleted.  The region/reserv map for ranges without associated
 395  *      pages are not modified.  Page faults can race with hole punch.
 396  *      This is indicated if we find a mapped page.
 397  * Note: If the passed end of range value is beyond the end of file, but
 398  * not LLONG_MAX this routine still performs a hole punch operation.
 399  */
 400 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 401                                    loff_t lend)
 402 {
 403         struct hstate *h = hstate_inode(inode);
 404         struct address_space *mapping = &inode->i_data;
 405         const pgoff_t start = lstart >> huge_page_shift(h);
 406         const pgoff_t end = lend >> huge_page_shift(h);
 407         struct vm_area_struct pseudo_vma;
 408         struct pagevec pvec;
 409         pgoff_t next, index;
 410         int i, freed = 0;
 411         bool truncate_op = (lend == LLONG_MAX);
 412
 413         vma_init(&pseudo_vma, current->mm);
 414         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 415         pagevec_init(&pvec);
 416         next = start;
 417         while (next < end) {
 418                 /*
 419                  * When no more pages are found, we are done.
 420                  */
 421                 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
 422                         break;
 423
 424                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 425                         struct page *page = pvec.pages[i];
 426                         u32 hash;
 427
 428                         index = page->index;
 429                         hash = hugetlb_fault_mutex_hash(h, mapping, index);
 430                         mutex_lock(&hugetlb_fault_mutex_table[hash]);
 431
 432                         /*
 433                          * If page is mapped, it was faulted in after being
 434                          * unmapped in caller.  Unmap (again) now after taking
 435                          * the fault mutex.  The mutex will prevent faults
 436                          * until we finish removing the page.
 437                          *
 438                          * This race can only happen in the hole punch case.
 439                          * Getting here in a truncate operation is a bug.
 440                          */
 441                         if (unlikely(page_mapped(page))) {
 442                                 BUG_ON(truncate_op);
 443
 444                                 i_mmap_lock_write(mapping);
 445                                 hugetlb_vmdelete_list(&mapping->i_mmap,
 446                                         index * pages_per_huge_page(h),
 447                                         (index + 1) * pages_per_huge_page(h));
 448                                 i_mmap_unlock_write(mapping);
 449                         }
 450
 451                         lock_page(page);
 452                         /*
 453                          * We must free the huge page and remove from page
 454                          * cache (remove_huge_page) BEFORE removing the
 455                          * region/reserve map (hugetlb_unreserve_pages).  In
 456                          * rare out of memory conditions, removal of the
 457                          * region/reserve map could fail. Correspondingly,
 458                          * the subpool and global reserve usage count can need
 459                          * to be adjusted.
 460                          */
 461                         VM_BUG_ON(PagePrivate(page));
 462                         remove_huge_page(page);
 463                         freed++;
 464                         if (!truncate_op) {
 465                                 if (unlikely(hugetlb_unreserve_pages(inode,
 466                                                         index, index + 1, 1)))
 467                                         hugetlb_fix_reserve_counts(inode);
 468                         }
 469
 470                         unlock_page(page);
 471                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 472                 }
 473                 huge_pagevec_release(&pvec);
 474                 cond_resched();
 475         }
 476
 477         if (truncate_op)
 478                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 479 }
 480
 481 static void hugetlbfs_evict_inode(struct inode *inode)
 482 {
 483         struct resv_map *resv_map;
 484
 485         remove_inode_hugepages(inode, 0, LLONG_MAX);
 486         resv_map = (struct resv_map *)inode->i_mapping->private_data;
 487         /* root inode doesn't have the resv_map, so we should check it */
 488         if (resv_map)
 489                 resv_map_release(&resv_map->refs);
 490         clear_inode(inode);
 491 }
 492
 493 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 494 {
 495         pgoff_t pgoff;
 496         struct address_space *mapping = inode->i_mapping;
 497         struct hstate *h = hstate_inode(inode);
 498
 499         BUG_ON(offset & ~huge_page_mask(h));
 500         pgoff = offset >> PAGE_SHIFT;
 501
 502         i_size_write(inode, offset);
 503         i_mmap_lock_write(mapping);
 504         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 505                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 506         i_mmap_unlock_write(mapping);
 507         remove_inode_hugepages(inode, offset, LLONG_MAX);
 508         return 0;
 509 }
 510
 511 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 512 {
 513         struct hstate *h = hstate_inode(inode);
 514         loff_t hpage_size = huge_page_size(h);
 515         loff_t hole_start, hole_end;
 516
 517         /*
 518          * For hole punch round up the beginning offset of the hole and
 519          * round down the end.
 520          */
 521         hole_start = round_up(offset, hpage_size);
 522         hole_end = round_down(offset + len, hpage_size);
 523
 524         if (hole_end > hole_start) {
 525                 struct address_space *mapping = inode->i_mapping;
 526                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 527
 528                 inode_lock(inode);
 529
 530                 /* protected by i_mutex */
 531                 if (info->seals & F_SEAL_WRITE) {
 532                         inode_unlock(inode);
 533                         return -EPERM;
 534                 }
 535
 536                 i_mmap_lock_write(mapping);
 537                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 538                         hugetlb_vmdelete_list(&mapping->i_mmap,
 539                                                 hole_start >> PAGE_SHIFT,
 540                                                 hole_end  >> PAGE_SHIFT);
 541                 i_mmap_unlock_write(mapping);
 542                 remove_inode_hugepages(inode, hole_start, hole_end);
 543                 inode_unlock(inode);
 544         }
 545
 546         return 0;
 547 }
 548
 549 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 550                                 loff_t len)
 551 {
 552         struct inode *inode = file_inode(file);
 553         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 554         struct address_space *mapping = inode->i_mapping;
 555         struct hstate *h = hstate_inode(inode);
 556         struct vm_area_struct pseudo_vma;
 557         struct mm_struct *mm = current->mm;
 558         loff_t hpage_size = huge_page_size(h);
 559         unsigned long hpage_shift = huge_page_shift(h);
 560         pgoff_t start, index, end;
 561         int error;
 562         u32 hash;
 563
 564         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 565                 return -EOPNOTSUPP;
 566
 567         if (mode & FALLOC_FL_PUNCH_HOLE)
 568                 return hugetlbfs_punch_hole(inode, offset, len);
 569
 570         /*
 571          * Default preallocate case.
 572          * For this range, start is rounded down and end is rounded up
 573          * as well as being converted to page offsets.
 574          */
 575         start = offset >> hpage_shift;
 576         end = (offset + len + hpage_size - 1) >> hpage_shift;
 577
 578         inode_lock(inode);
 579
 580         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 581         error = inode_newsize_ok(inode, offset + len);
 582         if (error)
 583                 goto out;
 584
 585         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 586                 error = -EPERM;
 587                 goto out;
 588         }
 589
 590         /*
 591          * Initialize a pseudo vma as this is required by the huge page
 592          * allocation routines.  If NUMA is configured, use page index
 593          * as input to create an allocation policy.
 594          */
 595         vma_init(&pseudo_vma, mm);
 596         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 597         pseudo_vma.vm_file = file;
 598
 599         for (index = start; index < end; index++) {
 600                 /*
 601                  * This is supposed to be the vaddr where the page is being
 602                  * faulted in, but we have no vaddr here.
 603                  */
 604                 struct page *page;
 605                 unsigned long addr;
 606                 int avoid_reserve = 0;
 607
 608                 cond_resched();
 609
 610                 /*
 611                  * fallocate(2) manpage permits EINTR; we may have been
 612                  * interrupted because we are using up too much memory.
 613                  */
 614                 if (signal_pending(current)) {
 615                         error = -EINTR;
 616                         break;
 617                 }
 618
 619                 /* Set numa allocation policy based on index */
 620                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 621
 622                 /* addr is the offset within the file (zero based) */
 623                 addr = index * hpage_size;
 624
 625                 /* mutex taken here, fault path and hole punch */
 626                 hash = hugetlb_fault_mutex_hash(h, mapping, index);
 627                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 628
 629                 /* See if already present in mapping to avoid alloc/free */
 630                 page = find_get_page(mapping, index);
 631                 if (page) {
 632                         put_page(page);
 633                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 634                         hugetlb_drop_vma_policy(&pseudo_vma);
 635                         continue;
 636                 }
 637
 638                 /* Allocate page and add to page cache */
 639                 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
 640                 hugetlb_drop_vma_policy(&pseudo_vma);
 641                 if (IS_ERR(page)) {
 642                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 643                         error = PTR_ERR(page);
 644                         goto out;
 645                 }
 646                 clear_huge_page(page, addr, pages_per_huge_page(h));
 647                 __SetPageUptodate(page);
 648                 error = huge_add_to_page_cache(page, mapping, index);
 649                 if (unlikely(error)) {
 650                         put_page(page);
 651                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 652                         goto out;
 653                 }
 654
 655                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 656
 657                 set_page_huge_active(page);
 658                 /*
 659                  * unlock_page because locked by add_to_page_cache()
 660                  * put_page() due to reference from alloc_huge_page()
 661                  */
 662                 unlock_page(page);
 663                 put_page(page);
 664         }
 665
 666         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 667                 i_size_write(inode, offset + len);
 668         inode->i_ctime = current_time(inode);
 669 out:
 670         inode_unlock(inode);
 671         return error;
 672 }
 673
 674 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 675 {
 676         struct inode *inode = d_inode(dentry);
 677         struct hstate *h = hstate_inode(inode);
 678         int error;
 679         unsigned int ia_valid = attr->ia_valid;
 680         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 681
 682         BUG_ON(!inode);
 683
 684         error = setattr_prepare(dentry, attr);
 685         if (error)
 686                 return error;
 687
 688         if (ia_valid & ATTR_SIZE) {
 689                 loff_t oldsize = inode->i_size;
 690                 loff_t newsize = attr->ia_size;
 691
 692                 if (newsize & ~huge_page_mask(h))
 693                         return -EINVAL;
 694                 /* protected by i_mutex */
 695                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 696                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 697                         return -EPERM;
 698                 error = hugetlb_vmtruncate(inode, newsize);
 699                 if (error)
 700                         return error;
 701         }
 702
 703         setattr_copy(inode, attr);
 704         mark_inode_dirty(inode);
 705         return 0;
 706 }
 707
 708 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 709                                         struct hugetlbfs_config *config)
 710 {
 711         struct inode *inode;
 712
 713         inode = new_inode(sb);
 714         if (inode) {
 715                 inode->i_ino = get_next_ino();
 716                 inode->i_mode = S_IFDIR | config->mode;
 717                 inode->i_uid = config->uid;
 718                 inode->i_gid = config->gid;
 719                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 720                 inode->i_op = &hugetlbfs_dir_inode_operations;
 721                 inode->i_fop = &simple_dir_operations;
 722                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 723                 inc_nlink(inode);
 724                 lockdep_annotate_inode_mutex_key(inode);
 725         }
 726         return inode;
 727 }
 728
 729 /*
 730  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 731  * be taken from reclaim -- unlike regular filesystems. This needs an
 732  * annotation because huge_pmd_share() does an allocation under hugetlb's
 733  * i_mmap_rwsem.
 734  */
 735 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 736
 737 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 738                                         struct inode *dir,
 739                                         umode_t mode, dev_t dev)
 740 {
 741         struct inode *inode;
 742         struct resv_map *resv_map = NULL;
 743
 744         /*
 745          * Reserve maps are only needed for inodes that can have associated
 746          * page allocations.
 747          */
 748         if (S_ISREG(mode) || S_ISLNK(mode)) {
 749                 resv_map = resv_map_alloc();
 750                 if (!resv_map)
 751                         return NULL;
 752         }
 753
 754         inode = new_inode(sb);
 755         if (inode) {
 756                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 757
 758                 inode->i_ino = get_next_ino();
 759                 inode_init_owner(inode, dir, mode);
 760                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 761                                 &hugetlbfs_i_mmap_rwsem_key);
 762                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 763                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 764                 inode->i_mapping->private_data = resv_map;
 765                 info->seals = F_SEAL_SEAL;
 766                 switch (mode & S_IFMT) {
 767                 default:
 768                         init_special_inode(inode, mode, dev);
 769                         break;
 770                 case S_IFREG:
 771                         inode->i_op = &hugetlbfs_inode_operations;
 772                         inode->i_fop = &hugetlbfs_file_operations;
 773                         break;
 774                 case S_IFDIR:
 775                         inode->i_op = &hugetlbfs_dir_inode_operations;
 776                         inode->i_fop = &simple_dir_operations;
 777
 778                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 779                         inc_nlink(inode);
 780                         break;
 781                 case S_IFLNK:
 782                         inode->i_op = &page_symlink_inode_operations;
 783                         inode_nohighmem(inode);
 784                         break;
 785                 }
 786                 lockdep_annotate_inode_mutex_key(inode);
 787         } else {
 788                 if (resv_map)
 789                         kref_put(&resv_map->refs, resv_map_release);
 790         }
 791
 792         return inode;
 793 }
 794
 795 /*
 796  * File creation. Allocate an inode, and we're done..
 797  */
 798 static int hugetlbfs_mknod(struct inode *dir,
 799                         struct dentry *dentry, umode_t mode, dev_t dev)
 800 {
 801         struct inode *inode;
 802         int error = -ENOSPC;
 803
 804         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 805         if (inode) {
 806                 dir->i_ctime = dir->i_mtime = current_time(dir);
 807                 d_instantiate(dentry, inode);
 808                 dget(dentry);   /* Extra count - pin the dentry in core */
 809                 error = 0;
 810         }
 811         return error;
 812 }
 813
 814 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 815 {
 816         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
 817         if (!retval)
 818                 inc_nlink(dir);
 819         return retval;
 820 }
 821
 822 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
 823 {
 824         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
 825 }
 826
 827 static int hugetlbfs_symlink(struct inode *dir,
 828                         struct dentry *dentry, const char *symname)
 829 {
 830         struct inode *inode;
 831         int error = -ENOSPC;
 832
 833         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 834         if (inode) {
 835                 int l = strlen(symname)+1;
 836                 error = page_symlink(inode, symname, l);
 837                 if (!error) {
 838                         d_instantiate(dentry, inode);
 839                         dget(dentry);
 840                 } else
 841                         iput(inode);
 842         }
 843         dir->i_ctime = dir->i_mtime = current_time(dir);
 844
 845         return error;
 846 }
 847
 848 /*
 849  * mark the head page dirty
 850  */
 851 static int hugetlbfs_set_page_dirty(struct page *page)
 852 {
 853         struct page *head = compound_head(page);
 854
 855         SetPageDirty(head);
 856         return 0;
 857 }
 858
 859 static int hugetlbfs_migrate_page(struct address_space *mapping,
 860                                 struct page *newpage, struct page *page,
 861                                 enum migrate_mode mode)
 862 {
 863         int rc;
 864
 865         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 866         if (rc != MIGRATEPAGE_SUCCESS)
 867                 return rc;
 868
 869         /*
 870          * page_private is subpool pointer in hugetlb pages.  Transfer to
 871          * new page.  PagePrivate is not associated with page_private for
 872          * hugetlb pages and can not be set here as only page_huge_active
 873          * pages can be migrated.
 874          */
 875         if (page_private(page)) {
 876                 set_page_private(newpage, page_private(page));
 877                 set_page_private(page, 0);
 878         }
 879
 880         if (mode != MIGRATE_SYNC_NO_COPY)
 881                 migrate_page_copy(newpage, page);
 882         else
 883                 migrate_page_states(newpage, page);
 884
 885         return MIGRATEPAGE_SUCCESS;
 886 }
 887
 888 static int hugetlbfs_error_remove_page(struct address_space *mapping,
 889                                 struct page *page)
 890 {
 891         struct inode *inode = mapping->host;
 892         pgoff_t index = page->index;
 893
 894         remove_huge_page(page);
 895         if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
 896                 hugetlb_fix_reserve_counts(inode);
 897
 898         return 0;
 899 }
 900
 901 /*
 902  * Display the mount options in /proc/mounts.
 903  */
 904 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
 905 {
 906         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
 907         struct hugepage_subpool *spool = sbinfo->spool;
 908         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
 909         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
 910         char mod;
 911
 912         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
 913                 seq_printf(m, ",uid=%u",
 914                            from_kuid_munged(&init_user_ns, sbinfo->uid));
 915         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
 916                 seq_printf(m, ",gid=%u",
 917                            from_kgid_munged(&init_user_ns, sbinfo->gid));
 918         if (sbinfo->mode != 0755)
 919                 seq_printf(m, ",mode=%o", sbinfo->mode);
 920         if (sbinfo->max_inodes != -1)
 921                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
 922
 923         hpage_size /= 1024;
 924         mod = 'K';
 925         if (hpage_size >= 1024) {
 926                 hpage_size /= 1024;
 927                 mod = 'M';
 928         }
 929         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
 930         if (spool) {
 931                 if (spool->max_hpages != -1)
 932                         seq_printf(m, ",size=%llu",
 933                                    (unsigned long long)spool->max_hpages << hpage_shift);
 934                 if (spool->min_hpages != -1)
 935                         seq_printf(m, ",min_size=%llu",
 936                                    (unsigned long long)spool->min_hpages << hpage_shift);
 937         }
 938         return 0;
 939 }
 940
 941 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 942 {
 943         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
 944         struct hstate *h = hstate_inode(d_inode(dentry));
 945
 946         buf->f_type = HUGETLBFS_MAGIC;
 947         buf->f_bsize = huge_page_size(h);
 948         if (sbinfo) {
 949                 spin_lock(&sbinfo->stat_lock);
 950                 /* If no limits set, just report 0 for max/free/used
 951                  * blocks, like simple_statfs() */
 952                 if (sbinfo->spool) {
 953                         long free_pages;
 954
 955                         spin_lock(&sbinfo->spool->lock);
 956                         buf->f_blocks = sbinfo->spool->max_hpages;
 957                         free_pages = sbinfo->spool->max_hpages
 958                                 - sbinfo->spool->used_hpages;
 959                         buf->f_bavail = buf->f_bfree = free_pages;
 960                         spin_unlock(&sbinfo->spool->lock);
 961                         buf->f_files = sbinfo->max_inodes;
 962                         buf->f_ffree = sbinfo->free_inodes;
 963                 }
 964                 spin_unlock(&sbinfo->stat_lock);
 965         }
 966         buf->f_namelen = NAME_MAX;
 967         return 0;
 968 }
 969
 970 static void hugetlbfs_put_super(struct super_block *sb)
 971 {
 972         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
 973
 974         if (sbi) {
 975                 sb->s_fs_info = NULL;
 976
 977                 if (sbi->spool)
 978                         hugepage_put_subpool(sbi->spool);
 979
 980                 kfree(sbi);
 981         }
 982 }
 983
 984 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
 985 {
 986         if (sbinfo->free_inodes >= 0) {
 987                 spin_lock(&sbinfo->stat_lock);
 988                 if (unlikely(!sbinfo->free_inodes)) {
 989                         spin_unlock(&sbinfo->stat_lock);
 990                         return 0;
 991                 }
 992                 sbinfo->free_inodes--;
 993                 spin_unlock(&sbinfo->stat_lock);
 994         }
 995
 996         return 1;
 997 }
 998
 999 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1000 {
1001         if (sbinfo->free_inodes >= 0) {
1002                 spin_lock(&sbinfo->stat_lock);
1003                 sbinfo->free_inodes++;
1004                 spin_unlock(&sbinfo->stat_lock);
1005         }
1006 }
1007
1008
1009 static struct kmem_cache *hugetlbfs_inode_cachep;
1010
1011 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1012 {
1013         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1014         struct hugetlbfs_inode_info *p;
1015
1016         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1017                 return NULL;
1018         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1019         if (unlikely(!p)) {
1020                 hugetlbfs_inc_free_inodes(sbinfo);
1021                 return NULL;
1022         }
1023
1024         /*
1025          * Any time after allocation, hugetlbfs_destroy_inode can be called
1026          * for the inode.  mpol_free_shared_policy is unconditionally called
1027          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1028          * in case of a quick call to destroy.
1029          *
1030          * Note that the policy is initialized even if we are creating a
1031          * private inode.  This simplifies hugetlbfs_destroy_inode.
1032          */
1033         mpol_shared_policy_init(&p->policy, NULL);
1034
1035         return &p->vfs_inode;
1036 }
1037
1038 static void hugetlbfs_i_callback(struct rcu_head *head)
1039 {
1040         struct inode *inode = container_of(head, struct inode, i_rcu);
1041         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1042 }
1043
1044 static void hugetlbfs_destroy_inode(struct inode *inode)
1045 {
1046         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1047         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1048         call_rcu(&inode->i_rcu, hugetlbfs_i_callback);
1049 }
1050
1051 static const struct address_space_operations hugetlbfs_aops = {
1052         .write_begin    = hugetlbfs_write_begin,
1053         .write_end      = hugetlbfs_write_end,
1054         .set_page_dirty = hugetlbfs_set_page_dirty,
1055         .migratepage    = hugetlbfs_migrate_page,
1056         .error_remove_page      = hugetlbfs_error_remove_page,
1057 };
1058
1059
1060 static void init_once(void *foo)
1061 {
1062         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1063
1064         inode_init_once(&ei->vfs_inode);
1065 }
1066
1067 const struct file_operations hugetlbfs_file_operations = {
1068         .read_iter              = hugetlbfs_read_iter,
1069         .mmap                   = hugetlbfs_file_mmap,
1070         .fsync                  = noop_fsync,
1071         .get_unmapped_area      = hugetlb_get_unmapped_area,
1072         .llseek                 = default_llseek,
1073         .fallocate              = hugetlbfs_fallocate,
1074 };
1075
1076 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1077         .create         = hugetlbfs_create,
1078         .lookup         = simple_lookup,
1079         .link           = simple_link,
1080         .unlink         = simple_unlink,
1081         .symlink        = hugetlbfs_symlink,
1082         .mkdir          = hugetlbfs_mkdir,
1083         .rmdir          = simple_rmdir,
1084         .mknod          = hugetlbfs_mknod,
1085         .rename         = simple_rename,
1086         .setattr        = hugetlbfs_setattr,
1087 };
1088
1089 static const struct inode_operations hugetlbfs_inode_operations = {
1090         .setattr        = hugetlbfs_setattr,
1091 };
1092
1093 static const struct super_operations hugetlbfs_ops = {
1094         .alloc_inode    = hugetlbfs_alloc_inode,
1095         .destroy_inode  = hugetlbfs_destroy_inode,
1096         .evict_inode    = hugetlbfs_evict_inode,
1097         .statfs         = hugetlbfs_statfs,
1098         .put_super      = hugetlbfs_put_super,
1099         .show_options   = hugetlbfs_show_options,
1100 };
1101
1102 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
1103
1104 /*
1105  * Convert size option passed from command line to number of huge pages
1106  * in the pool specified by hstate.  Size option could be in bytes
1107  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1108  */
1109 static long
1110 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1111                          enum hugetlbfs_size_type val_type)
1112 {
1113         if (val_type == NO_SIZE)
1114                 return -1;
1115
1116         if (val_type == SIZE_PERCENT) {
1117                 size_opt <<= huge_page_shift(h);
1118                 size_opt *= h->max_huge_pages;
1119                 do_div(size_opt, 100);
1120         }
1121
1122         size_opt >>= huge_page_shift(h);
1123         return size_opt;
1124 }
1125
1126 static int
1127 hugetlbfs_parse_options(char *options, struct hugetlbfs_config *pconfig)
1128 {
1129         char *p, *rest;
1130         substring_t args[MAX_OPT_ARGS];
1131         int option;
1132         unsigned long long max_size_opt = 0, min_size_opt = 0;
1133         enum hugetlbfs_size_type max_val_type = NO_SIZE, min_val_type = NO_SIZE;
1134
1135         if (!options)
1136                 return 0;
1137
1138         while ((p = strsep(&options, ",")) != NULL) {
1139                 int token;
1140                 if (!*p)
1141                         continue;
1142
1143                 token = match_token(p, tokens, args);
1144                 switch (token) {
1145                 case Opt_uid:
1146                         if (match_int(&args[0], &option))
1147                                 goto bad_val;
1148                         pconfig->uid = make_kuid(current_user_ns(), option);
1149                         if (!uid_valid(pconfig->uid))
1150                                 goto bad_val;
1151                         break;
1152
1153                 case Opt_gid:
1154                         if (match_int(&args[0], &option))
1155                                 goto bad_val;
1156                         pconfig->gid = make_kgid(current_user_ns(), option);
1157                         if (!gid_valid(pconfig->gid))
1158                                 goto bad_val;
1159                         break;
1160
1161                 case Opt_mode:
1162                         if (match_octal(&args[0], &option))
1163                                 goto bad_val;
1164                         pconfig->mode = option & 01777U;
1165                         break;
1166
1167                 case Opt_size: {
1168                         /* memparse() will accept a K/M/G without a digit */
1169                         if (!isdigit(*args[0].from))
1170                                 goto bad_val;
1171                         max_size_opt = memparse(args[0].from, &rest);
1172                         max_val_type = SIZE_STD;
1173                         if (*rest == '%')
1174                                 max_val_type = SIZE_PERCENT;
1175                         break;
1176                 }
1177
1178                 case Opt_nr_inodes:
1179                         /* memparse() will accept a K/M/G without a digit */
1180                         if (!isdigit(*args[0].from))
1181                                 goto bad_val;
1182                         pconfig->nr_inodes = memparse(args[0].from, &rest);
1183                         break;
1184
1185                 case Opt_pagesize: {
1186                         unsigned long ps;
1187                         ps = memparse(args[0].from, &rest);
1188                         pconfig->hstate = size_to_hstate(ps);
1189                         if (!pconfig->hstate) {
1190                                 pr_err("Unsupported page size %lu MB\n",
1191                                         ps >> 20);
1192                                 return -EINVAL;
1193                         }
1194                         break;
1195                 }
1196
1197                 case Opt_min_size: {
1198                         /* memparse() will accept a K/M/G without a digit */
1199                         if (!isdigit(*args[0].from))
1200                                 goto bad_val;
1201                         min_size_opt = memparse(args[0].from, &rest);
1202                         min_val_type = SIZE_STD;
1203                         if (*rest == '%')
1204                                 min_val_type = SIZE_PERCENT;
1205                         break;
1206                 }
1207
1208                 default:
1209                         pr_err("Bad mount option: \"%s\"\n", p);
1210                         return -EINVAL;
1211                         break;
1212                 }
1213         }
1214
1215         /*
1216          * Use huge page pool size (in hstate) to convert the size
1217          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1218          */
1219         pconfig->max_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1220                                                 max_size_opt, max_val_type);
1221         pconfig->min_hpages = hugetlbfs_size_to_hpages(pconfig->hstate,
1222                                                 min_size_opt, min_val_type);
1223
1224         /*
1225          * If max_size was specified, then min_size must be smaller
1226          */
1227         if (max_val_type > NO_SIZE &&
1228             pconfig->min_hpages > pconfig->max_hpages) {
1229                 pr_err("minimum size can not be greater than maximum size\n");
1230                 return -EINVAL;
1231         }
1232
1233         return 0;
1234
1235 bad_val:
1236         pr_err("Bad value '%s' for mount option '%s'\n", args[0].from, p);
1237         return -EINVAL;
1238 }
1239
1240 static int
1241 hugetlbfs_fill_super(struct super_block *sb, void *data, int silent)
1242 {
1243         int ret;
1244         struct hugetlbfs_config config;
1245         struct hugetlbfs_sb_info *sbinfo;
1246
1247         config.max_hpages = -1; /* No limit on size by default */
1248         config.nr_inodes = -1; /* No limit on number of inodes by default */
1249         config.uid = current_fsuid();
1250         config.gid = current_fsgid();
1251         config.mode = 0755;
1252         config.hstate = &default_hstate;
1253         config.min_hpages = -1; /* No default minimum size */
1254         ret = hugetlbfs_parse_options(data, &config);
1255         if (ret)
1256                 return ret;
1257
1258         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1259         if (!sbinfo)
1260                 return -ENOMEM;
1261         sb->s_fs_info = sbinfo;
1262         sbinfo->hstate = config.hstate;
1263         spin_lock_init(&sbinfo->stat_lock);
1264         sbinfo->max_inodes = config.nr_inodes;
1265         sbinfo->free_inodes = config.nr_inodes;
1266         sbinfo->spool = NULL;
1267         sbinfo->uid = config.uid;
1268         sbinfo->gid = config.gid;
1269         sbinfo->mode = config.mode;
1270
1271         /*
1272          * Allocate and initialize subpool if maximum or minimum size is
1273          * specified.  Any needed reservations (for minimim size) are taken
1274          * taken when the subpool is created.
1275          */
1276         if (config.max_hpages != -1 || config.min_hpages != -1) {
1277                 sbinfo->spool = hugepage_new_subpool(config.hstate,
1278                                                         config.max_hpages,
1279                                                         config.min_hpages);
1280                 if (!sbinfo->spool)
1281                         goto out_free;
1282         }
1283         sb->s_maxbytes = MAX_LFS_FILESIZE;
1284         sb->s_blocksize = huge_page_size(config.hstate);
1285         sb->s_blocksize_bits = huge_page_shift(config.hstate);
1286         sb->s_magic = HUGETLBFS_MAGIC;
1287         sb->s_op = &hugetlbfs_ops;
1288         sb->s_time_gran = 1;
1289         sb->s_root = d_make_root(hugetlbfs_get_root(sb, &config));
1290         if (!sb->s_root)
1291                 goto out_free;
1292         return 0;
1293 out_free:
1294         kfree(sbinfo->spool);
1295         kfree(sbinfo);
1296         return -ENOMEM;
1297 }
1298
1299 static struct dentry *hugetlbfs_mount(struct file_system_type *fs_type,
1300         int flags, const char *dev_name, void *data)
1301 {
1302         return mount_nodev(fs_type, flags, data, hugetlbfs_fill_super);
1303 }
1304
1305 static struct file_system_type hugetlbfs_fs_type = {
1306         .name           = "hugetlbfs",
1307         .mount          = hugetlbfs_mount,
1308         .kill_sb        = kill_litter_super,
1309 };
1310
1311 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1312
1313 static int can_do_hugetlb_shm(void)
1314 {
1315         kgid_t shm_group;
1316         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1317         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1318 }
1319
1320 static int get_hstate_idx(int page_size_log)
1321 {
1322         struct hstate *h = hstate_sizelog(page_size_log);
1323
1324         if (!h)
1325                 return -1;
1326         return h - hstates;
1327 }
1328
1329 /*
1330  * Note that size should be aligned to proper hugepage size in caller side,
1331  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1332  */
1333 struct file *hugetlb_file_setup(const char *name, size_t size,
1334                                 vm_flags_t acctflag, struct user_struct **user,
1335                                 int creat_flags, int page_size_log)
1336 {
1337         struct inode *inode;
1338         struct vfsmount *mnt;
1339         int hstate_idx;
1340         struct file *file;
1341
1342         hstate_idx = get_hstate_idx(page_size_log);
1343         if (hstate_idx < 0)
1344                 return ERR_PTR(-ENODEV);
1345
1346         *user = NULL;
1347         mnt = hugetlbfs_vfsmount[hstate_idx];
1348         if (!mnt)
1349                 return ERR_PTR(-ENOENT);
1350
1351         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1352                 *user = current_user();
1353                 if (user_shm_lock(size, *user)) {
1354                         task_lock(current);
1355                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1356                                 current->comm, current->pid);
1357                         task_unlock(current);
1358                 } else {
1359                         *user = NULL;
1360                         return ERR_PTR(-EPERM);
1361                 }
1362         }
1363
1364         file = ERR_PTR(-ENOSPC);
1365         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1366         if (!inode)
1367                 goto out;
1368         if (creat_flags == HUGETLB_SHMFS_INODE)
1369                 inode->i_flags |= S_PRIVATE;
1370
1371         inode->i_size = size;
1372         clear_nlink(inode);
1373
1374         if (hugetlb_reserve_pages(inode, 0,
1375                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1376                         acctflag))
1377                 file = ERR_PTR(-ENOMEM);
1378         else
1379                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1380                                         &hugetlbfs_file_operations);
1381         if (!IS_ERR(file))
1382                 return file;
1383
1384         iput(inode);
1385 out:
1386         if (*user) {
1387                 user_shm_unlock(size, *user);
1388                 *user = NULL;
1389         }
1390         return file;
1391 }
1392
1393 static int __init init_hugetlbfs_fs(void)
1394 {
1395         struct hstate *h;
1396         int error;
1397         int i;
1398
1399         if (!hugepages_supported()) {
1400                 pr_info("disabling because there are no supported hugepage sizes\n");
1401                 return -ENOTSUPP;
1402         }
1403
1404         error = -ENOMEM;
1405         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1406                                         sizeof(struct hugetlbfs_inode_info),
1407                                         0, SLAB_ACCOUNT, init_once);
1408         if (hugetlbfs_inode_cachep == NULL)
1409                 goto out2;
1410
1411         error = register_filesystem(&hugetlbfs_fs_type);
1412         if (error)
1413                 goto out;
1414
1415         i = 0;
1416         for_each_hstate(h) {
1417                 char buf[50];
1418                 unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
1419
1420                 snprintf(buf, sizeof(buf), "pagesize=%uK", ps_kb);
1421                 hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
1422                                                         buf);
1423
1424                 if (IS_ERR(hugetlbfs_vfsmount[i])) {
1425                         pr_err("Cannot mount internal hugetlbfs for "
1426                                 "page size %uK", ps_kb);
1427                         error = PTR_ERR(hugetlbfs_vfsmount[i]);
1428                         hugetlbfs_vfsmount[i] = NULL;
1429                 }
1430                 i++;
1431         }
1432         /* Non default hstates are optional */
1433         if (!IS_ERR_OR_NULL(hugetlbfs_vfsmount[default_hstate_idx]))
1434                 return 0;
1435
1436  out:
1437         kmem_cache_destroy(hugetlbfs_inode_cachep);
1438  out2:
1439         return error;
1440 }
1441 fs_initcall(init_hugetlbfs_fs)