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