fs/hugetlbfs/inode.c

   1 /*
   2  * hugetlbpage-backed filesystem.  Based on ramfs.
   3  *
   4  * Nadia Yvette Chambers, 2002
   5  *
   6  * Copyright (C) 2002 Linus Torvalds.
   7  * License: GPL
   8  */
   9
  10 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  11
  12 #include <linux/thread_info.h>
  13 #include <asm/current.h>
  14 #include <linux/sched/signal.h>         /* remove ASAP */
  15 #include <linux/falloc.h>
  16 #include <linux/fs.h>
  17 #include <linux/mount.h>
  18 #include <linux/file.h>
  19 #include <linux/kernel.h>
  20 #include <linux/writeback.h>
  21 #include <linux/pagemap.h>
  22 #include <linux/highmem.h>
  23 #include <linux/init.h>
  24 #include <linux/string.h>
  25 #include <linux/capability.h>
  26 #include <linux/ctype.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/hugetlb.h>
  29 #include <linux/pagevec.h>
  30 #include <linux/fs_parser.h>
  31 #include <linux/mman.h>
  32 #include <linux/slab.h>
  33 #include <linux/dnotify.h>
  34 #include <linux/statfs.h>
  35 #include <linux/security.h>
  36 #include <linux/magic.h>
  37 #include <linux/migrate.h>
  38 #include <linux/uio.h>
  39
  40 #include <linux/uaccess.h>
  41
  42 static const struct super_operations hugetlbfs_ops;
  43 static const struct address_space_operations hugetlbfs_aops;
  44 const struct file_operations hugetlbfs_file_operations;
  45 static const struct inode_operations hugetlbfs_dir_inode_operations;
  46 static const struct inode_operations hugetlbfs_inode_operations;
  47
  48 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
  49
  50 struct hugetlbfs_fs_context {
  51         struct hstate           *hstate;
  52         unsigned long long      max_size_opt;
  53         unsigned long long      min_size_opt;
  54         long                    max_hpages;
  55         long                    nr_inodes;
  56         long                    min_hpages;
  57         enum hugetlbfs_size_type max_val_type;
  58         enum hugetlbfs_size_type min_val_type;
  59         kuid_t                  uid;
  60         kgid_t                  gid;
  61         umode_t                 mode;
  62 };
  63
  64 int sysctl_hugetlb_shm_group;
  65
  66 enum hugetlb_param {
  67         Opt_gid,
  68         Opt_min_size,
  69         Opt_mode,
  70         Opt_nr_inodes,
  71         Opt_pagesize,
  72         Opt_size,
  73         Opt_uid,
  74 };
  75
  76 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
  77         fsparam_u32   ("gid",           Opt_gid),
  78         fsparam_string("min_size",      Opt_min_size),
  79         fsparam_u32   ("mode",          Opt_mode),
  80         fsparam_string("nr_inodes",     Opt_nr_inodes),
  81         fsparam_string("pagesize",      Opt_pagesize),
  82         fsparam_string("size",          Opt_size),
  83         fsparam_u32   ("uid",           Opt_uid),
  84         {}
  85 };
  86
  87 #ifdef CONFIG_NUMA
  88 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  89                                         struct inode *inode, pgoff_t index)
  90 {
  91         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
  92                                                         index);
  93 }
  94
  95 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  96 {
  97         mpol_cond_put(vma->vm_policy);
  98 }
  99 #else
 100 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
 101                                         struct inode *inode, pgoff_t index)
 102 {
 103 }
 104
 105 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
 106 {
 107 }
 108 #endif
 109
 110 static void huge_pagevec_release(struct pagevec *pvec)
 111 {
 112         int i;
 113
 114         for (i = 0; i < pagevec_count(pvec); ++i)
 115                 put_page(pvec->pages[i]);
 116
 117         pagevec_reinit(pvec);
 118 }
 119
 120 /*
 121  * Mask used when checking the page offset value passed in via system
 122  * calls.  This value will be converted to a loff_t which is signed.
 123  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 124  * value.  The extra bit (- 1 in the shift value) is to take the sign
 125  * bit into account.
 126  */
 127 #define PGOFF_LOFFT_MAX \
 128         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 129
 130 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 131 {
 132         struct inode *inode = file_inode(file);
 133         loff_t len, vma_len;
 134         int ret;
 135         struct hstate *h = hstate_file(file);
 136
 137         /*
 138          * vma address alignment (but not the pgoff alignment) has
 139          * already been checked by prepare_hugepage_range.  If you add
 140          * any error returns here, do so after setting VM_HUGETLB, so
 141          * is_vm_hugetlb_page tests below unmap_region go the right
 142          * way when do_mmap_pgoff unwinds (may be important on powerpc
 143          * and ia64).
 144          */
 145         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 146         vma->vm_ops = &hugetlb_vm_ops;
 147
 148         /*
 149          * page based offset in vm_pgoff could be sufficiently large to
 150          * overflow a loff_t when converted to byte offset.  This can
 151          * only happen on architectures where sizeof(loff_t) ==
 152          * sizeof(unsigned long).  So, only check in those instances.
 153          */
 154         if (sizeof(unsigned long) == sizeof(loff_t)) {
 155                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 156                         return -EINVAL;
 157         }
 158
 159         /* must be huge page aligned */
 160         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 161                 return -EINVAL;
 162
 163         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 164         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 165         /* check for overflow */
 166         if (len < vma_len)
 167                 return -EINVAL;
 168
 169         inode_lock(inode);
 170         file_accessed(file);
 171
 172         ret = -ENOMEM;
 173         if (hugetlb_reserve_pages(inode,
 174                                 vma->vm_pgoff >> huge_page_order(h),
 175                                 len >> huge_page_shift(h), vma,
 176                                 vma->vm_flags))
 177                 goto out;
 178
 179         ret = 0;
 180         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 181                 i_size_write(inode, len);
 182 out:
 183         inode_unlock(inode);
 184
 185         return ret;
 186 }
 187
 188 /*
 189  * Called under down_write(mmap_sem).
 190  */
 191
 192 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 193 static unsigned long
 194 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 195                 unsigned long len, unsigned long pgoff, unsigned long flags)
 196 {
 197         struct mm_struct *mm = current->mm;
 198         struct vm_area_struct *vma;
 199         struct hstate *h = hstate_file(file);
 200         struct vm_unmapped_area_info info;
 201
 202         if (len & ~huge_page_mask(h))
 203                 return -EINVAL;
 204         if (len > TASK_SIZE)
 205                 return -ENOMEM;
 206
 207         if (flags & MAP_FIXED) {
 208                 if (prepare_hugepage_range(file, addr, len))
 209                         return -EINVAL;
 210                 return addr;
 211         }
 212
 213         if (addr) {
 214                 addr = ALIGN(addr, huge_page_size(h));
 215                 vma = find_vma(mm, addr);
 216                 if (TASK_SIZE - len >= addr &&
 217                     (!vma || addr + len <= vm_start_gap(vma)))
 218                         return addr;
 219         }
 220
 221         info.flags = 0;
 222         info.length = len;
 223         info.low_limit = TASK_UNMAPPED_BASE;
 224         info.high_limit = TASK_SIZE;
 225         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 226         info.align_offset = 0;
 227         return vm_unmapped_area(&info);
 228 }
 229 #endif
 230
 231 static size_t
 232 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 233                         struct iov_iter *to, unsigned long size)
 234 {
 235         size_t copied = 0;
 236         int i, chunksize;
 237
 238         /* Find which 4k chunk and offset with in that chunk */
 239         i = offset >> PAGE_SHIFT;
 240         offset = offset & ~PAGE_MASK;
 241
 242         while (size) {
 243                 size_t n;
 244                 chunksize = PAGE_SIZE;
 245                 if (offset)
 246                         chunksize -= offset;
 247                 if (chunksize > size)
 248                         chunksize = size;
 249                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 250                 copied += n;
 251                 if (n != chunksize)
 252                         return copied;
 253                 offset = 0;
 254                 size -= chunksize;
 255                 i++;
 256         }
 257         return copied;
 258 }
 259
 260 /*
 261  * Support for read() - Find the page attached to f_mapping and copy out the
 262  * data. Its *very* similar to do_generic_mapping_read(), we can't use that
 263  * since it has PAGE_SIZE assumptions.
 264  */
 265 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 266 {
 267         struct file *file = iocb->ki_filp;
 268         struct hstate *h = hstate_file(file);
 269         struct address_space *mapping = file->f_mapping;
 270         struct inode *inode = mapping->host;
 271         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 272         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 273         unsigned long end_index;
 274         loff_t isize;
 275         ssize_t retval = 0;
 276
 277         while (iov_iter_count(to)) {
 278                 struct page *page;
 279                 size_t nr, copied;
 280
 281                 /* nr is the maximum number of bytes to copy from this page */
 282                 nr = huge_page_size(h);
 283                 isize = i_size_read(inode);
 284                 if (!isize)
 285                         break;
 286                 end_index = (isize - 1) >> huge_page_shift(h);
 287                 if (index > end_index)
 288                         break;
 289                 if (index == end_index) {
 290                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 291                         if (nr <= offset)
 292                                 break;
 293                 }
 294                 nr = nr - offset;
 295
 296                 /* Find the page */
 297                 page = find_lock_page(mapping, index);
 298                 if (unlikely(page == NULL)) {
 299                         /*
 300                          * We have a HOLE, zero out the user-buffer for the
 301                          * length of the hole or request.
 302                          */
 303                         copied = iov_iter_zero(nr, to);
 304                 } else {
 305                         unlock_page(page);
 306
 307                         /*
 308                          * We have the page, copy it to user space buffer.
 309                          */
 310                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 311                         put_page(page);
 312                 }
 313                 offset += copied;
 314                 retval += copied;
 315                 if (copied != nr && iov_iter_count(to)) {
 316                         if (!retval)
 317                                 retval = -EFAULT;
 318                         break;
 319                 }
 320                 index += offset >> huge_page_shift(h);
 321                 offset &= ~huge_page_mask(h);
 322         }
 323         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 324         return retval;
 325 }
 326
 327 static int hugetlbfs_write_begin(struct file *file,
 328                         struct address_space *mapping,
 329                         loff_t pos, unsigned len, unsigned flags,
 330                         struct page **pagep, void **fsdata)
 331 {
 332         return -EINVAL;
 333 }
 334
 335 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 336                         loff_t pos, unsigned len, unsigned copied,
 337                         struct page *page, void *fsdata)
 338 {
 339         BUG();
 340         return -EINVAL;
 341 }
 342
 343 static void remove_huge_page(struct page *page)
 344 {
 345         ClearPageDirty(page);
 346         ClearPageUptodate(page);
 347         delete_from_page_cache(page);
 348 }
 349
 350 static void
 351 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end)
 352 {
 353         struct vm_area_struct *vma;
 354
 355         /*
 356          * end == 0 indicates that the entire range after
 357          * start should be unmapped.
 358          */
 359         vma_interval_tree_foreach(vma, root, start, end ? end : ULONG_MAX) {
 360                 unsigned long v_offset;
 361                 unsigned long v_end;
 362
 363                 /*
 364                  * Can the expression below overflow on 32-bit arches?
 365                  * No, because the interval tree returns us only those vmas
 366                  * which overlap the truncated area starting at pgoff,
 367                  * and no vma on a 32-bit arch can span beyond the 4GB.
 368                  */
 369                 if (vma->vm_pgoff < start)
 370                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 371                 else
 372                         v_offset = 0;
 373
 374                 if (!end)
 375                         v_end = vma->vm_end;
 376                 else {
 377                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 378                                                         + vma->vm_start;
 379                         if (v_end > vma->vm_end)
 380                                 v_end = vma->vm_end;
 381                 }
 382
 383                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 384                                                                         NULL);
 385         }
 386 }
 387
 388 /*
 389  * remove_inode_hugepages handles two distinct cases: truncation and hole
 390  * punch.  There are subtle differences in operation for each case.
 391  *
 392  * truncation is indicated by end of range being LLONG_MAX
 393  *      In this case, we first scan the range and release found pages.
 394  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserv
 395  *      maps and global counts.  Page faults can not race with truncation
 396  *      in this routine.  hugetlb_no_page() holds i_mmap_rwsem and prevents
 397  *      page faults in the truncated range by checking i_size.  i_size is
 398  *      modified while holding i_mmap_rwsem.
 399  * hole punch is indicated if end is not LLONG_MAX
 400  *      In the hole punch case we scan the range and release found pages.
 401  *      Only when releasing a page is the associated region/reserv map
 402  *      deleted.  The region/reserv map for ranges without associated
 403  *      pages are not modified.  Page faults can race with hole punch.
 404  *      This is indicated if we find a mapped page.
 405  * Note: If the passed end of range value is beyond the end of file, but
 406  * not LLONG_MAX this routine still performs a hole punch operation.
 407  */
 408 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 409                                    loff_t lend)
 410 {
 411         struct hstate *h = hstate_inode(inode);
 412         struct address_space *mapping = &inode->i_data;
 413         const pgoff_t start = lstart >> huge_page_shift(h);
 414         const pgoff_t end = lend >> huge_page_shift(h);
 415         struct vm_area_struct pseudo_vma;
 416         struct pagevec pvec;
 417         pgoff_t next, index;
 418         int i, freed = 0;
 419         bool truncate_op = (lend == LLONG_MAX);
 420
 421         vma_init(&pseudo_vma, current->mm);
 422         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 423         pagevec_init(&pvec);
 424         next = start;
 425         while (next < end) {
 426                 /*
 427                  * When no more pages are found, we are done.
 428                  */
 429                 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
 430                         break;
 431
 432                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 433                         struct page *page = pvec.pages[i];
 434                         u32 hash;
 435
 436                         index = page->index;
 437                         hash = hugetlb_fault_mutex_hash(mapping, index);
 438                         if (!truncate_op) {
 439                                 /*
 440                                  * Only need to hold the fault mutex in the
 441                                  * hole punch case.  This prevents races with
 442                                  * page faults.  Races are not possible in the
 443                                  * case of truncation.
 444                                  */
 445                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 446                         }
 447
 448                         /*
 449                          * If page is mapped, it was faulted in after being
 450                          * unmapped in caller.  Unmap (again) now after taking
 451                          * the fault mutex.  The mutex will prevent faults
 452                          * until we finish removing the page.
 453                          *
 454                          * This race can only happen in the hole punch case.
 455                          * Getting here in a truncate operation is a bug.
 456                          */
 457                         if (unlikely(page_mapped(page))) {
 458                                 BUG_ON(truncate_op);
 459
 460                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 461                                 i_mmap_lock_write(mapping);
 462                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 463                                 hugetlb_vmdelete_list(&mapping->i_mmap,
 464                                         index * pages_per_huge_page(h),
 465                                         (index + 1) * pages_per_huge_page(h));
 466                                 i_mmap_unlock_write(mapping);
 467                         }
 468
 469                         lock_page(page);
 470                         /*
 471                          * We must free the huge page and remove from page
 472                          * cache (remove_huge_page) BEFORE removing the
 473                          * region/reserve map (hugetlb_unreserve_pages).  In
 474                          * rare out of memory conditions, removal of the
 475                          * region/reserve map could fail. Correspondingly,
 476                          * the subpool and global reserve usage count can need
 477                          * to be adjusted.
 478                          */
 479                         VM_BUG_ON(PagePrivate(page));
 480                         remove_huge_page(page);
 481                         freed++;
 482                         if (!truncate_op) {
 483                                 if (unlikely(hugetlb_unreserve_pages(inode,
 484                                                         index, index + 1, 1)))
 485                                         hugetlb_fix_reserve_counts(inode);
 486                         }
 487
 488                         unlock_page(page);
 489                         if (!truncate_op)
 490                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 491                 }
 492                 huge_pagevec_release(&pvec);
 493                 cond_resched();
 494         }
 495
 496         if (truncate_op)
 497                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 498 }
 499
 500 static void hugetlbfs_evict_inode(struct inode *inode)
 501 {
 502         struct resv_map *resv_map;
 503
 504         remove_inode_hugepages(inode, 0, LLONG_MAX);
 505
 506         /*
 507          * Get the resv_map from the address space embedded in the inode.
 508          * This is the address space which points to any resv_map allocated
 509          * at inode creation time.  If this is a device special inode,
 510          * i_mapping may not point to the original address space.
 511          */
 512         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
 513         /* Only regular and link inodes have associated reserve maps */
 514         if (resv_map)
 515                 resv_map_release(&resv_map->refs);
 516         clear_inode(inode);
 517 }
 518
 519 static int hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 520 {
 521         pgoff_t pgoff;
 522         struct address_space *mapping = inode->i_mapping;
 523         struct hstate *h = hstate_inode(inode);
 524
 525         BUG_ON(offset & ~huge_page_mask(h));
 526         pgoff = offset >> PAGE_SHIFT;
 527
 528         i_mmap_lock_write(mapping);
 529         i_size_write(inode, offset);
 530         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 531                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0);
 532         i_mmap_unlock_write(mapping);
 533         remove_inode_hugepages(inode, offset, LLONG_MAX);
 534         return 0;
 535 }
 536
 537 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 538 {
 539         struct hstate *h = hstate_inode(inode);
 540         loff_t hpage_size = huge_page_size(h);
 541         loff_t hole_start, hole_end;
 542
 543         /*
 544          * For hole punch round up the beginning offset of the hole and
 545          * round down the end.
 546          */
 547         hole_start = round_up(offset, hpage_size);
 548         hole_end = round_down(offset + len, hpage_size);
 549
 550         if (hole_end > hole_start) {
 551                 struct address_space *mapping = inode->i_mapping;
 552                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 553
 554                 inode_lock(inode);
 555
 556                 /* protected by i_mutex */
 557                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 558                         inode_unlock(inode);
 559                         return -EPERM;
 560                 }
 561
 562                 i_mmap_lock_write(mapping);
 563                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 564                         hugetlb_vmdelete_list(&mapping->i_mmap,
 565                                                 hole_start >> PAGE_SHIFT,
 566                                                 hole_end  >> PAGE_SHIFT);
 567                 i_mmap_unlock_write(mapping);
 568                 remove_inode_hugepages(inode, hole_start, hole_end);
 569                 inode_unlock(inode);
 570         }
 571
 572         return 0;
 573 }
 574
 575 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 576                                 loff_t len)
 577 {
 578         struct inode *inode = file_inode(file);
 579         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 580         struct address_space *mapping = inode->i_mapping;
 581         struct hstate *h = hstate_inode(inode);
 582         struct vm_area_struct pseudo_vma;
 583         struct mm_struct *mm = current->mm;
 584         loff_t hpage_size = huge_page_size(h);
 585         unsigned long hpage_shift = huge_page_shift(h);
 586         pgoff_t start, index, end;
 587         int error;
 588         u32 hash;
 589
 590         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 591                 return -EOPNOTSUPP;
 592
 593         if (mode & FALLOC_FL_PUNCH_HOLE)
 594                 return hugetlbfs_punch_hole(inode, offset, len);
 595
 596         /*
 597          * Default preallocate case.
 598          * For this range, start is rounded down and end is rounded up
 599          * as well as being converted to page offsets.
 600          */
 601         start = offset >> hpage_shift;
 602         end = (offset + len + hpage_size - 1) >> hpage_shift;
 603
 604         inode_lock(inode);
 605
 606         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 607         error = inode_newsize_ok(inode, offset + len);
 608         if (error)
 609                 goto out;
 610
 611         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 612                 error = -EPERM;
 613                 goto out;
 614         }
 615
 616         /*
 617          * Initialize a pseudo vma as this is required by the huge page
 618          * allocation routines.  If NUMA is configured, use page index
 619          * as input to create an allocation policy.
 620          */
 621         vma_init(&pseudo_vma, mm);
 622         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 623         pseudo_vma.vm_file = file;
 624
 625         for (index = start; index < end; index++) {
 626                 /*
 627                  * This is supposed to be the vaddr where the page is being
 628                  * faulted in, but we have no vaddr here.
 629                  */
 630                 struct page *page;
 631                 unsigned long addr;
 632                 int avoid_reserve = 0;
 633
 634                 cond_resched();
 635
 636                 /*
 637                  * fallocate(2) manpage permits EINTR; we may have been
 638                  * interrupted because we are using up too much memory.
 639                  */
 640                 if (signal_pending(current)) {
 641                         error = -EINTR;
 642                         break;
 643                 }
 644
 645                 /* Set numa allocation policy based on index */
 646                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 647
 648                 /* addr is the offset within the file (zero based) */
 649                 addr = index * hpage_size;
 650
 651                 /*
 652                  * fault mutex taken here, protects against fault path
 653                  * and hole punch.  inode_lock previously taken protects
 654                  * against truncation.
 655                  */
 656                 hash = hugetlb_fault_mutex_hash(mapping, index);
 657                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 658
 659                 /* See if already present in mapping to avoid alloc/free */
 660                 page = find_get_page(mapping, index);
 661                 if (page) {
 662                         put_page(page);
 663                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 664                         hugetlb_drop_vma_policy(&pseudo_vma);
 665                         continue;
 666                 }
 667
 668                 /* Allocate page and add to page cache */
 669                 page = alloc_huge_page(&pseudo_vma, addr, avoid_reserve);
 670                 hugetlb_drop_vma_policy(&pseudo_vma);
 671                 if (IS_ERR(page)) {
 672                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 673                         error = PTR_ERR(page);
 674                         goto out;
 675                 }
 676                 clear_huge_page(page, addr, pages_per_huge_page(h));
 677                 __SetPageUptodate(page);
 678                 error = huge_add_to_page_cache(page, mapping, index);
 679                 if (unlikely(error)) {
 680                         put_page(page);
 681                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 682                         goto out;
 683                 }
 684
 685                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 686
 687                 /*
 688                  * unlock_page because locked by add_to_page_cache()
 689                  * page_put due to reference from alloc_huge_page()
 690                  */
 691                 unlock_page(page);
 692                 put_page(page);
 693         }
 694
 695         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 696                 i_size_write(inode, offset + len);
 697         inode->i_ctime = current_time(inode);
 698 out:
 699         inode_unlock(inode);
 700         return error;
 701 }
 702
 703 static int hugetlbfs_setattr(struct dentry *dentry, struct iattr *attr)
 704 {
 705         struct inode *inode = d_inode(dentry);
 706         struct hstate *h = hstate_inode(inode);
 707         int error;
 708         unsigned int ia_valid = attr->ia_valid;
 709         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 710
 711         BUG_ON(!inode);
 712
 713         error = setattr_prepare(dentry, attr);
 714         if (error)
 715                 return error;
 716
 717         if (ia_valid & ATTR_SIZE) {
 718                 loff_t oldsize = inode->i_size;
 719                 loff_t newsize = attr->ia_size;
 720
 721                 if (newsize & ~huge_page_mask(h))
 722                         return -EINVAL;
 723                 /* protected by i_mutex */
 724                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 725                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 726                         return -EPERM;
 727                 error = hugetlb_vmtruncate(inode, newsize);
 728                 if (error)
 729                         return error;
 730         }
 731
 732         setattr_copy(inode, attr);
 733         mark_inode_dirty(inode);
 734         return 0;
 735 }
 736
 737 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 738                                         struct hugetlbfs_fs_context *ctx)
 739 {
 740         struct inode *inode;
 741
 742         inode = new_inode(sb);
 743         if (inode) {
 744                 inode->i_ino = get_next_ino();
 745                 inode->i_mode = S_IFDIR | ctx->mode;
 746                 inode->i_uid = ctx->uid;
 747                 inode->i_gid = ctx->gid;
 748                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 749                 inode->i_op = &hugetlbfs_dir_inode_operations;
 750                 inode->i_fop = &simple_dir_operations;
 751                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 752                 inc_nlink(inode);
 753                 lockdep_annotate_inode_mutex_key(inode);
 754         }
 755         return inode;
 756 }
 757
 758 /*
 759  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 760  * be taken from reclaim -- unlike regular filesystems. This needs an
 761  * annotation because huge_pmd_share() does an allocation under hugetlb's
 762  * i_mmap_rwsem.
 763  */
 764 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 765
 766 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 767                                         struct inode *dir,
 768                                         umode_t mode, dev_t dev)
 769 {
 770         struct inode *inode;
 771         struct resv_map *resv_map = NULL;
 772
 773         /*
 774          * Reserve maps are only needed for inodes that can have associated
 775          * page allocations.
 776          */
 777         if (S_ISREG(mode) || S_ISLNK(mode)) {
 778                 resv_map = resv_map_alloc();
 779                 if (!resv_map)
 780                         return NULL;
 781         }
 782
 783         inode = new_inode(sb);
 784         if (inode) {
 785                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 786
 787                 inode->i_ino = get_next_ino();
 788                 inode_init_owner(inode, dir, mode);
 789                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 790                                 &hugetlbfs_i_mmap_rwsem_key);
 791                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 792                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 793                 inode->i_mapping->private_data = resv_map;
 794                 info->seals = F_SEAL_SEAL;
 795                 switch (mode & S_IFMT) {
 796                 default:
 797                         init_special_inode(inode, mode, dev);
 798                         break;
 799                 case S_IFREG:
 800                         inode->i_op = &hugetlbfs_inode_operations;
 801                         inode->i_fop = &hugetlbfs_file_operations;
 802                         break;
 803                 case S_IFDIR:
 804                         inode->i_op = &hugetlbfs_dir_inode_operations;
 805                         inode->i_fop = &simple_dir_operations;
 806
 807                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 808                         inc_nlink(inode);
 809                         break;
 810                 case S_IFLNK:
 811                         inode->i_op = &page_symlink_inode_operations;
 812                         inode_nohighmem(inode);
 813                         break;
 814                 }
 815                 lockdep_annotate_inode_mutex_key(inode);
 816         } else {
 817                 if (resv_map)
 818                         kref_put(&resv_map->refs, resv_map_release);
 819         }
 820
 821         return inode;
 822 }
 823
 824 /*
 825  * File creation. Allocate an inode, and we're done..
 826  */
 827 static int do_hugetlbfs_mknod(struct inode *dir,
 828                         struct dentry *dentry,
 829                         umode_t mode,
 830                         dev_t dev,
 831                         bool tmpfile)
 832 {
 833         struct inode *inode;
 834         int error = -ENOSPC;
 835
 836         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 837         if (inode) {
 838                 dir->i_ctime = dir->i_mtime = current_time(dir);
 839                 if (tmpfile) {
 840                         d_tmpfile(dentry, inode);
 841                 } else {
 842                         d_instantiate(dentry, inode);
 843                         dget(dentry);/* Extra count - pin the dentry in core */
 844                 }
 845                 error = 0;
 846         }
 847         return error;
 848 }
 849
 850 static int hugetlbfs_mknod(struct inode *dir,
 851                         struct dentry *dentry, umode_t mode, dev_t dev)
 852 {
 853         return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
 854 }
 855
 856 static int hugetlbfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
 857 {
 858         int retval = hugetlbfs_mknod(dir, dentry, mode | S_IFDIR, 0);
 859         if (!retval)
 860                 inc_nlink(dir);
 861         return retval;
 862 }
 863
 864 static int hugetlbfs_create(struct inode *dir, struct dentry *dentry, umode_t mode, bool excl)
 865 {
 866         return hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0);
 867 }
 868
 869 static int hugetlbfs_tmpfile(struct inode *dir,
 870                         struct dentry *dentry, umode_t mode)
 871 {
 872         return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
 873 }
 874
 875 static int hugetlbfs_symlink(struct inode *dir,
 876                         struct dentry *dentry, const char *symname)
 877 {
 878         struct inode *inode;
 879         int error = -ENOSPC;
 880
 881         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 882         if (inode) {
 883                 int l = strlen(symname)+1;
 884                 error = page_symlink(inode, symname, l);
 885                 if (!error) {
 886                         d_instantiate(dentry, inode);
 887                         dget(dentry);
 888                 } else
 889                         iput(inode);
 890         }
 891         dir->i_ctime = dir->i_mtime = current_time(dir);
 892
 893         return error;
 894 }
 895
 896 /*
 897  * mark the head page dirty
 898  */
 899 static int hugetlbfs_set_page_dirty(struct page *page)
 900 {
 901         struct page *head = compound_head(page);
 902
 903         SetPageDirty(head);
 904         return 0;
 905 }
 906
 907 static int hugetlbfs_migrate_page(struct address_space *mapping,
 908                                 struct page *newpage, struct page *page,
 909                                 enum migrate_mode mode)
 910 {
 911         int rc;
 912
 913         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 914         if (rc != MIGRATEPAGE_SUCCESS)
 915                 return rc;
 916
 917         /*
 918          * page_private is subpool pointer in hugetlb pages.  Transfer to
 919          * new page.  PagePrivate is not associated with page_private for
 920          * hugetlb pages and can not be set here as only page_huge_active
 921          * pages can be migrated.
 922          */
 923         if (page_private(page)) {
 924                 set_page_private(newpage, page_private(page));
 925                 set_page_private(page, 0);
 926         }
 927
 928         if (mode != MIGRATE_SYNC_NO_COPY)
 929                 migrate_page_copy(newpage, page);
 930         else
 931                 migrate_page_states(newpage, page);
 932
 933         return MIGRATEPAGE_SUCCESS;
 934 }
 935
 936 static int hugetlbfs_error_remove_page(struct address_space *mapping,
 937                                 struct page *page)
 938 {
 939         struct inode *inode = mapping->host;
 940         pgoff_t index = page->index;
 941
 942         remove_huge_page(page);
 943         if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
 944                 hugetlb_fix_reserve_counts(inode);
 945
 946         return 0;
 947 }
 948
 949 /*
 950  * Display the mount options in /proc/mounts.
 951  */
 952 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
 953 {
 954         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
 955         struct hugepage_subpool *spool = sbinfo->spool;
 956         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
 957         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
 958         char mod;
 959
 960         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
 961                 seq_printf(m, ",uid=%u",
 962                            from_kuid_munged(&init_user_ns, sbinfo->uid));
 963         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
 964                 seq_printf(m, ",gid=%u",
 965                            from_kgid_munged(&init_user_ns, sbinfo->gid));
 966         if (sbinfo->mode != 0755)
 967                 seq_printf(m, ",mode=%o", sbinfo->mode);
 968         if (sbinfo->max_inodes != -1)
 969                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
 970
 971         hpage_size /= 1024;
 972         mod = 'K';
 973         if (hpage_size >= 1024) {
 974                 hpage_size /= 1024;
 975                 mod = 'M';
 976         }
 977         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
 978         if (spool) {
 979                 if (spool->max_hpages != -1)
 980                         seq_printf(m, ",size=%llu",
 981                                    (unsigned long long)spool->max_hpages << hpage_shift);
 982                 if (spool->min_hpages != -1)
 983                         seq_printf(m, ",min_size=%llu",
 984                                    (unsigned long long)spool->min_hpages << hpage_shift);
 985         }
 986         return 0;
 987 }
 988
 989 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
 990 {
 991         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
 992         struct hstate *h = hstate_inode(d_inode(dentry));
 993
 994         buf->f_type = HUGETLBFS_MAGIC;
 995         buf->f_bsize = huge_page_size(h);
 996         if (sbinfo) {
 997                 spin_lock(&sbinfo->stat_lock);
 998                 /* If no limits set, just report 0 for max/free/used
 999                  * blocks, like simple_statfs() */
1000                 if (sbinfo->spool) {
1001                         long free_pages;
1002
1003                         spin_lock(&sbinfo->spool->lock);
1004                         buf->f_blocks = sbinfo->spool->max_hpages;
1005                         free_pages = sbinfo->spool->max_hpages
1006                                 - sbinfo->spool->used_hpages;
1007                         buf->f_bavail = buf->f_bfree = free_pages;
1008                         spin_unlock(&sbinfo->spool->lock);
1009                         buf->f_files = sbinfo->max_inodes;
1010                         buf->f_ffree = sbinfo->free_inodes;
1011                 }
1012                 spin_unlock(&sbinfo->stat_lock);
1013         }
1014         buf->f_namelen = NAME_MAX;
1015         return 0;
1016 }
1017
1018 static void hugetlbfs_put_super(struct super_block *sb)
1019 {
1020         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1021
1022         if (sbi) {
1023                 sb->s_fs_info = NULL;
1024
1025                 if (sbi->spool)
1026                         hugepage_put_subpool(sbi->spool);
1027
1028                 kfree(sbi);
1029         }
1030 }
1031
1032 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1033 {
1034         if (sbinfo->free_inodes >= 0) {
1035                 spin_lock(&sbinfo->stat_lock);
1036                 if (unlikely(!sbinfo->free_inodes)) {
1037                         spin_unlock(&sbinfo->stat_lock);
1038                         return 0;
1039                 }
1040                 sbinfo->free_inodes--;
1041                 spin_unlock(&sbinfo->stat_lock);
1042         }
1043
1044         return 1;
1045 }
1046
1047 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1048 {
1049         if (sbinfo->free_inodes >= 0) {
1050                 spin_lock(&sbinfo->stat_lock);
1051                 sbinfo->free_inodes++;
1052                 spin_unlock(&sbinfo->stat_lock);
1053         }
1054 }
1055
1056
1057 static struct kmem_cache *hugetlbfs_inode_cachep;
1058
1059 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1060 {
1061         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1062         struct hugetlbfs_inode_info *p;
1063
1064         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1065                 return NULL;
1066         p = kmem_cache_alloc(hugetlbfs_inode_cachep, GFP_KERNEL);
1067         if (unlikely(!p)) {
1068                 hugetlbfs_inc_free_inodes(sbinfo);
1069                 return NULL;
1070         }
1071
1072         /*
1073          * Any time after allocation, hugetlbfs_destroy_inode can be called
1074          * for the inode.  mpol_free_shared_policy is unconditionally called
1075          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1076          * in case of a quick call to destroy.
1077          *
1078          * Note that the policy is initialized even if we are creating a
1079          * private inode.  This simplifies hugetlbfs_destroy_inode.
1080          */
1081         mpol_shared_policy_init(&p->policy, NULL);
1082
1083         return &p->vfs_inode;
1084 }
1085
1086 static void hugetlbfs_free_inode(struct inode *inode)
1087 {
1088         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1089 }
1090
1091 static void hugetlbfs_destroy_inode(struct inode *inode)
1092 {
1093         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1094         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1095 }
1096
1097 static const struct address_space_operations hugetlbfs_aops = {
1098         .write_begin    = hugetlbfs_write_begin,
1099         .write_end      = hugetlbfs_write_end,
1100         .set_page_dirty = hugetlbfs_set_page_dirty,
1101         .migratepage    = hugetlbfs_migrate_page,
1102         .error_remove_page      = hugetlbfs_error_remove_page,
1103 };
1104
1105
1106 static void init_once(void *foo)
1107 {
1108         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1109
1110         inode_init_once(&ei->vfs_inode);
1111 }
1112
1113 const struct file_operations hugetlbfs_file_operations = {
1114         .read_iter              = hugetlbfs_read_iter,
1115         .mmap                   = hugetlbfs_file_mmap,
1116         .fsync                  = noop_fsync,
1117         .get_unmapped_area      = hugetlb_get_unmapped_area,
1118         .llseek                 = default_llseek,
1119         .fallocate              = hugetlbfs_fallocate,
1120 };
1121
1122 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1123         .create         = hugetlbfs_create,
1124         .lookup         = simple_lookup,
1125         .link           = simple_link,
1126         .unlink         = simple_unlink,
1127         .symlink        = hugetlbfs_symlink,
1128         .mkdir          = hugetlbfs_mkdir,
1129         .rmdir          = simple_rmdir,
1130         .mknod          = hugetlbfs_mknod,
1131         .rename         = simple_rename,
1132         .setattr        = hugetlbfs_setattr,
1133         .tmpfile        = hugetlbfs_tmpfile,
1134 };
1135
1136 static const struct inode_operations hugetlbfs_inode_operations = {
1137         .setattr        = hugetlbfs_setattr,
1138 };
1139
1140 static const struct super_operations hugetlbfs_ops = {
1141         .alloc_inode    = hugetlbfs_alloc_inode,
1142         .free_inode     = hugetlbfs_free_inode,
1143         .destroy_inode  = hugetlbfs_destroy_inode,
1144         .evict_inode    = hugetlbfs_evict_inode,
1145         .statfs         = hugetlbfs_statfs,
1146         .put_super      = hugetlbfs_put_super,
1147         .show_options   = hugetlbfs_show_options,
1148 };
1149
1150 /*
1151  * Convert size option passed from command line to number of huge pages
1152  * in the pool specified by hstate.  Size option could be in bytes
1153  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1154  */
1155 static long
1156 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1157                          enum hugetlbfs_size_type val_type)
1158 {
1159         if (val_type == NO_SIZE)
1160                 return -1;
1161
1162         if (val_type == SIZE_PERCENT) {
1163                 size_opt <<= huge_page_shift(h);
1164                 size_opt *= h->max_huge_pages;
1165                 do_div(size_opt, 100);
1166         }
1167
1168         size_opt >>= huge_page_shift(h);
1169         return size_opt;
1170 }
1171
1172 /*
1173  * Parse one mount parameter.
1174  */
1175 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1176 {
1177         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1178         struct fs_parse_result result;
1179         char *rest;
1180         unsigned long ps;
1181         int opt;
1182
1183         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1184         if (opt < 0)
1185                 return opt;
1186
1187         switch (opt) {
1188         case Opt_uid:
1189                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1190                 if (!uid_valid(ctx->uid))
1191                         goto bad_val;
1192                 return 0;
1193
1194         case Opt_gid:
1195                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1196                 if (!gid_valid(ctx->gid))
1197                         goto bad_val;
1198                 return 0;
1199
1200         case Opt_mode:
1201                 ctx->mode = result.uint_32 & 01777U;
1202                 return 0;
1203
1204         case Opt_size:
1205                 /* memparse() will accept a K/M/G without a digit */
1206                 if (!isdigit(param->string[0]))
1207                         goto bad_val;
1208                 ctx->max_size_opt = memparse(param->string, &rest);
1209                 ctx->max_val_type = SIZE_STD;
1210                 if (*rest == '%')
1211                         ctx->max_val_type = SIZE_PERCENT;
1212                 return 0;
1213
1214         case Opt_nr_inodes:
1215                 /* memparse() will accept a K/M/G without a digit */
1216                 if (!isdigit(param->string[0]))
1217                         goto bad_val;
1218                 ctx->nr_inodes = memparse(param->string, &rest);
1219                 return 0;
1220
1221         case Opt_pagesize:
1222                 ps = memparse(param->string, &rest);
1223                 ctx->hstate = size_to_hstate(ps);
1224                 if (!ctx->hstate) {
1225                         pr_err("Unsupported page size %lu MB\n", ps >> 20);
1226                         return -EINVAL;
1227                 }
1228                 return 0;
1229
1230         case Opt_min_size:
1231                 /* memparse() will accept a K/M/G without a digit */
1232                 if (!isdigit(param->string[0]))
1233                         goto bad_val;
1234                 ctx->min_size_opt = memparse(param->string, &rest);
1235                 ctx->min_val_type = SIZE_STD;
1236                 if (*rest == '%')
1237                         ctx->min_val_type = SIZE_PERCENT;
1238                 return 0;
1239
1240         default:
1241                 return -EINVAL;
1242         }
1243
1244 bad_val:
1245         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1246                       param->string, param->key);
1247 }
1248
1249 /*
1250  * Validate the parsed options.
1251  */
1252 static int hugetlbfs_validate(struct fs_context *fc)
1253 {
1254         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1255
1256         /*
1257          * Use huge page pool size (in hstate) to convert the size
1258          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1259          */
1260         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1261                                                    ctx->max_size_opt,
1262                                                    ctx->max_val_type);
1263         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1264                                                    ctx->min_size_opt,
1265                                                    ctx->min_val_type);
1266
1267         /*
1268          * If max_size was specified, then min_size must be smaller
1269          */
1270         if (ctx->max_val_type > NO_SIZE &&
1271             ctx->min_hpages > ctx->max_hpages) {
1272                 pr_err("Minimum size can not be greater than maximum size\n");
1273                 return -EINVAL;
1274         }
1275
1276         return 0;
1277 }
1278
1279 static int
1280 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1281 {
1282         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1283         struct hugetlbfs_sb_info *sbinfo;
1284
1285         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1286         if (!sbinfo)
1287                 return -ENOMEM;
1288         sb->s_fs_info = sbinfo;
1289         spin_lock_init(&sbinfo->stat_lock);
1290         sbinfo->hstate          = ctx->hstate;
1291         sbinfo->max_inodes      = ctx->nr_inodes;
1292         sbinfo->free_inodes     = ctx->nr_inodes;
1293         sbinfo->spool           = NULL;
1294         sbinfo->uid             = ctx->uid;
1295         sbinfo->gid             = ctx->gid;
1296         sbinfo->mode            = ctx->mode;
1297
1298         /*
1299          * Allocate and initialize subpool if maximum or minimum size is
1300          * specified.  Any needed reservations (for minimim size) are taken
1301          * taken when the subpool is created.
1302          */
1303         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1304                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1305                                                      ctx->max_hpages,
1306                                                      ctx->min_hpages);
1307                 if (!sbinfo->spool)
1308                         goto out_free;
1309         }
1310         sb->s_maxbytes = MAX_LFS_FILESIZE;
1311         sb->s_blocksize = huge_page_size(ctx->hstate);
1312         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1313         sb->s_magic = HUGETLBFS_MAGIC;
1314         sb->s_op = &hugetlbfs_ops;
1315         sb->s_time_gran = 1;
1316         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1317         if (!sb->s_root)
1318                 goto out_free;
1319         return 0;
1320 out_free:
1321         kfree(sbinfo->spool);
1322         kfree(sbinfo);
1323         return -ENOMEM;
1324 }
1325
1326 static int hugetlbfs_get_tree(struct fs_context *fc)
1327 {
1328         int err = hugetlbfs_validate(fc);
1329         if (err)
1330                 return err;
1331         return get_tree_nodev(fc, hugetlbfs_fill_super);
1332 }
1333
1334 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1335 {
1336         kfree(fc->fs_private);
1337 }
1338
1339 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1340         .free           = hugetlbfs_fs_context_free,
1341         .parse_param    = hugetlbfs_parse_param,
1342         .get_tree       = hugetlbfs_get_tree,
1343 };
1344
1345 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1346 {
1347         struct hugetlbfs_fs_context *ctx;
1348
1349         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1350         if (!ctx)
1351                 return -ENOMEM;
1352
1353         ctx->max_hpages = -1; /* No limit on size by default */
1354         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1355         ctx->uid        = current_fsuid();
1356         ctx->gid        = current_fsgid();
1357         ctx->mode       = 0755;
1358         ctx->hstate     = &default_hstate;
1359         ctx->min_hpages = -1; /* No default minimum size */
1360         ctx->max_val_type = NO_SIZE;
1361         ctx->min_val_type = NO_SIZE;
1362         fc->fs_private = ctx;
1363         fc->ops = &hugetlbfs_fs_context_ops;
1364         return 0;
1365 }
1366
1367 static struct file_system_type hugetlbfs_fs_type = {
1368         .name                   = "hugetlbfs",
1369         .init_fs_context        = hugetlbfs_init_fs_context,
1370         .parameters             = hugetlb_fs_parameters,
1371         .kill_sb                = kill_litter_super,
1372 };
1373
1374 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1375
1376 static int can_do_hugetlb_shm(void)
1377 {
1378         kgid_t shm_group;
1379         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1380         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1381 }
1382
1383 static int get_hstate_idx(int page_size_log)
1384 {
1385         struct hstate *h = hstate_sizelog(page_size_log);
1386
1387         if (!h)
1388                 return -1;
1389         return h - hstates;
1390 }
1391
1392 /*
1393  * Note that size should be aligned to proper hugepage size in caller side,
1394  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1395  */
1396 struct file *hugetlb_file_setup(const char *name, size_t size,
1397                                 vm_flags_t acctflag, struct user_struct **user,
1398                                 int creat_flags, int page_size_log)
1399 {
1400         struct inode *inode;
1401         struct vfsmount *mnt;
1402         int hstate_idx;
1403         struct file *file;
1404
1405         hstate_idx = get_hstate_idx(page_size_log);
1406         if (hstate_idx < 0)
1407                 return ERR_PTR(-ENODEV);
1408
1409         *user = NULL;
1410         mnt = hugetlbfs_vfsmount[hstate_idx];
1411         if (!mnt)
1412                 return ERR_PTR(-ENOENT);
1413
1414         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1415                 *user = current_user();
1416                 if (user_shm_lock(size, *user)) {
1417                         task_lock(current);
1418                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is deprecated\n",
1419                                 current->comm, current->pid);
1420                         task_unlock(current);
1421                 } else {
1422                         *user = NULL;
1423                         return ERR_PTR(-EPERM);
1424                 }
1425         }
1426
1427         file = ERR_PTR(-ENOSPC);
1428         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1429         if (!inode)
1430                 goto out;
1431         if (creat_flags == HUGETLB_SHMFS_INODE)
1432                 inode->i_flags |= S_PRIVATE;
1433
1434         inode->i_size = size;
1435         clear_nlink(inode);
1436
1437         if (hugetlb_reserve_pages(inode, 0,
1438                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1439                         acctflag))
1440                 file = ERR_PTR(-ENOMEM);
1441         else
1442                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1443                                         &hugetlbfs_file_operations);
1444         if (!IS_ERR(file))
1445                 return file;
1446
1447         iput(inode);
1448 out:
1449         if (*user) {
1450                 user_shm_unlock(size, *user);
1451                 *user = NULL;
1452         }
1453         return file;
1454 }
1455
1456 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1457 {
1458         struct fs_context *fc;
1459         struct vfsmount *mnt;
1460
1461         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1462         if (IS_ERR(fc)) {
1463                 mnt = ERR_CAST(fc);
1464         } else {
1465                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1466                 ctx->hstate = h;
1467                 mnt = fc_mount(fc);
1468                 put_fs_context(fc);
1469         }
1470         if (IS_ERR(mnt))
1471                 pr_err("Cannot mount internal hugetlbfs for page size %uK",
1472                        1U << (h->order + PAGE_SHIFT - 10));
1473         return mnt;
1474 }
1475
1476 static int __init init_hugetlbfs_fs(void)
1477 {
1478         struct vfsmount *mnt;
1479         struct hstate *h;
1480         int error;
1481         int i;
1482
1483         if (!hugepages_supported()) {
1484                 pr_info("disabling because there are no supported hugepage sizes\n");
1485                 return -ENOTSUPP;
1486         }
1487
1488         error = -ENOMEM;
1489         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1490                                         sizeof(struct hugetlbfs_inode_info),
1491                                         0, SLAB_ACCOUNT, init_once);
1492         if (hugetlbfs_inode_cachep == NULL)
1493                 goto out;
1494
1495         error = register_filesystem(&hugetlbfs_fs_type);
1496         if (error)
1497                 goto out_free;
1498
1499         /* default hstate mount is required */
1500         mnt = mount_one_hugetlbfs(&hstates[default_hstate_idx]);
1501         if (IS_ERR(mnt)) {
1502                 error = PTR_ERR(mnt);
1503                 goto out_unreg;
1504         }
1505         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1506
1507         /* other hstates are optional */
1508         i = 0;
1509         for_each_hstate(h) {
1510                 if (i == default_hstate_idx) {
1511                         i++;
1512                         continue;
1513                 }
1514
1515                 mnt = mount_one_hugetlbfs(h);
1516                 if (IS_ERR(mnt))
1517                         hugetlbfs_vfsmount[i] = NULL;
1518                 else
1519                         hugetlbfs_vfsmount[i] = mnt;
1520                 i++;
1521         }
1522
1523         return 0;
1524
1525  out_unreg:
1526         (void)unregister_filesystem(&hugetlbfs_fs_type);
1527  out_free:
1528         kmem_cache_destroy(hugetlbfs_inode_cachep);
1529  out:
1530         return error;
1531 }
1532 fs_initcall(init_hugetlbfs_fs)