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 #include <linux/sched/mm.h>
  42
  43 static const struct super_operations hugetlbfs_ops;
  44 static const struct address_space_operations hugetlbfs_aops;
  45 const struct file_operations hugetlbfs_file_operations;
  46 static const struct inode_operations hugetlbfs_dir_inode_operations;
  47 static const struct inode_operations hugetlbfs_inode_operations;
  48
  49 enum hugetlbfs_size_type { NO_SIZE, SIZE_STD, SIZE_PERCENT };
  50
  51 struct hugetlbfs_fs_context {
  52         struct hstate           *hstate;
  53         unsigned long long      max_size_opt;
  54         unsigned long long      min_size_opt;
  55         long                    max_hpages;
  56         long                    nr_inodes;
  57         long                    min_hpages;
  58         enum hugetlbfs_size_type max_val_type;
  59         enum hugetlbfs_size_type min_val_type;
  60         kuid_t                  uid;
  61         kgid_t                  gid;
  62         umode_t                 mode;
  63 };
  64
  65 int sysctl_hugetlb_shm_group;
  66
  67 enum hugetlb_param {
  68         Opt_gid,
  69         Opt_min_size,
  70         Opt_mode,
  71         Opt_nr_inodes,
  72         Opt_pagesize,
  73         Opt_size,
  74         Opt_uid,
  75 };
  76
  77 static const struct fs_parameter_spec hugetlb_fs_parameters[] = {
  78         fsparam_u32   ("gid",           Opt_gid),
  79         fsparam_string("min_size",      Opt_min_size),
  80         fsparam_u32oct("mode",          Opt_mode),
  81         fsparam_string("nr_inodes",     Opt_nr_inodes),
  82         fsparam_string("pagesize",      Opt_pagesize),
  83         fsparam_string("size",          Opt_size),
  84         fsparam_u32   ("uid",           Opt_uid),
  85         {}
  86 };
  87
  88 #ifdef CONFIG_NUMA
  89 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
  90                                         struct inode *inode, pgoff_t index)
  91 {
  92         vma->vm_policy = mpol_shared_policy_lookup(&HUGETLBFS_I(inode)->policy,
  93                                                         index);
  94 }
  95
  96 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
  97 {
  98         mpol_cond_put(vma->vm_policy);
  99 }
 100 #else
 101 static inline void hugetlb_set_vma_policy(struct vm_area_struct *vma,
 102                                         struct inode *inode, pgoff_t index)
 103 {
 104 }
 105
 106 static inline void hugetlb_drop_vma_policy(struct vm_area_struct *vma)
 107 {
 108 }
 109 #endif
 110
 111 static void huge_pagevec_release(struct pagevec *pvec)
 112 {
 113         int i;
 114
 115         for (i = 0; i < pagevec_count(pvec); ++i)
 116                 put_page(pvec->pages[i]);
 117
 118         pagevec_reinit(pvec);
 119 }
 120
 121 /*
 122  * Mask used when checking the page offset value passed in via system
 123  * calls.  This value will be converted to a loff_t which is signed.
 124  * Therefore, we want to check the upper PAGE_SHIFT + 1 bits of the
 125  * value.  The extra bit (- 1 in the shift value) is to take the sign
 126  * bit into account.
 127  */
 128 #define PGOFF_LOFFT_MAX \
 129         (((1UL << (PAGE_SHIFT + 1)) - 1) <<  (BITS_PER_LONG - (PAGE_SHIFT + 1)))
 130
 131 static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
 132 {
 133         struct inode *inode = file_inode(file);
 134         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 135         loff_t len, vma_len;
 136         int ret;
 137         struct hstate *h = hstate_file(file);
 138
 139         /*
 140          * vma address alignment (but not the pgoff alignment) has
 141          * already been checked by prepare_hugepage_range.  If you add
 142          * any error returns here, do so after setting VM_HUGETLB, so
 143          * is_vm_hugetlb_page tests below unmap_region go the right
 144          * way when do_mmap unwinds (may be important on powerpc
 145          * and ia64).
 146          */
 147         vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
 148         vma->vm_ops = &hugetlb_vm_ops;
 149
 150         ret = seal_check_future_write(info->seals, vma);
 151         if (ret)
 152                 return ret;
 153
 154         /*
 155          * page based offset in vm_pgoff could be sufficiently large to
 156          * overflow a loff_t when converted to byte offset.  This can
 157          * only happen on architectures where sizeof(loff_t) ==
 158          * sizeof(unsigned long).  So, only check in those instances.
 159          */
 160         if (sizeof(unsigned long) == sizeof(loff_t)) {
 161                 if (vma->vm_pgoff & PGOFF_LOFFT_MAX)
 162                         return -EINVAL;
 163         }
 164
 165         /* must be huge page aligned */
 166         if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
 167                 return -EINVAL;
 168
 169         vma_len = (loff_t)(vma->vm_end - vma->vm_start);
 170         len = vma_len + ((loff_t)vma->vm_pgoff << PAGE_SHIFT);
 171         /* check for overflow */
 172         if (len < vma_len)
 173                 return -EINVAL;
 174
 175         inode_lock(inode);
 176         file_accessed(file);
 177
 178         ret = -ENOMEM;
 179         if (!hugetlb_reserve_pages(inode,
 180                                 vma->vm_pgoff >> huge_page_order(h),
 181                                 len >> huge_page_shift(h), vma,
 182                                 vma->vm_flags))
 183                 goto out;
 184
 185         ret = 0;
 186         if (vma->vm_flags & VM_WRITE && inode->i_size < len)
 187                 i_size_write(inode, len);
 188 out:
 189         inode_unlock(inode);
 190
 191         return ret;
 192 }
 193
 194 /*
 195  * Called under mmap_write_lock(mm).
 196  */
 197
 198 #ifndef HAVE_ARCH_HUGETLB_UNMAPPED_AREA
 199 static unsigned long
 200 hugetlb_get_unmapped_area_bottomup(struct file *file, unsigned long addr,
 201                 unsigned long len, unsigned long pgoff, unsigned long flags)
 202 {
 203         struct hstate *h = hstate_file(file);
 204         struct vm_unmapped_area_info info;
 205
 206         info.flags = 0;
 207         info.length = len;
 208         info.low_limit = current->mm->mmap_base;
 209         info.high_limit = arch_get_mmap_end(addr);
 210         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 211         info.align_offset = 0;
 212         return vm_unmapped_area(&info);
 213 }
 214
 215 static unsigned long
 216 hugetlb_get_unmapped_area_topdown(struct file *file, unsigned long addr,
 217                 unsigned long len, unsigned long pgoff, unsigned long flags)
 218 {
 219         struct hstate *h = hstate_file(file);
 220         struct vm_unmapped_area_info info;
 221
 222         info.flags = VM_UNMAPPED_AREA_TOPDOWN;
 223         info.length = len;
 224         info.low_limit = max(PAGE_SIZE, mmap_min_addr);
 225         info.high_limit = arch_get_mmap_base(addr, current->mm->mmap_base);
 226         info.align_mask = PAGE_MASK & ~huge_page_mask(h);
 227         info.align_offset = 0;
 228         addr = vm_unmapped_area(&info);
 229
 230         /*
 231          * A failed mmap() very likely causes application failure,
 232          * so fall back to the bottom-up function here. This scenario
 233          * can happen with large stack limits and large mmap()
 234          * allocations.
 235          */
 236         if (unlikely(offset_in_page(addr))) {
 237                 VM_BUG_ON(addr != -ENOMEM);
 238                 info.flags = 0;
 239                 info.low_limit = current->mm->mmap_base;
 240                 info.high_limit = arch_get_mmap_end(addr);
 241                 addr = vm_unmapped_area(&info);
 242         }
 243
 244         return addr;
 245 }
 246
 247 static unsigned long
 248 hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
 249                 unsigned long len, unsigned long pgoff, unsigned long flags)
 250 {
 251         struct mm_struct *mm = current->mm;
 252         struct vm_area_struct *vma;
 253         struct hstate *h = hstate_file(file);
 254         const unsigned long mmap_end = arch_get_mmap_end(addr);
 255
 256         if (len & ~huge_page_mask(h))
 257                 return -EINVAL;
 258         if (len > TASK_SIZE)
 259                 return -ENOMEM;
 260
 261         if (flags & MAP_FIXED) {
 262                 if (prepare_hugepage_range(file, addr, len))
 263                         return -EINVAL;
 264                 return addr;
 265         }
 266
 267         if (addr) {
 268                 addr = ALIGN(addr, huge_page_size(h));
 269                 vma = find_vma(mm, addr);
 270                 if (mmap_end - len >= addr &&
 271                     (!vma || addr + len <= vm_start_gap(vma)))
 272                         return addr;
 273         }
 274
 275         /*
 276          * Use mm->get_unmapped_area value as a hint to use topdown routine.
 277          * If architectures have special needs, they should define their own
 278          * version of hugetlb_get_unmapped_area.
 279          */
 280         if (mm->get_unmapped_area == arch_get_unmapped_area_topdown)
 281                 return hugetlb_get_unmapped_area_topdown(file, addr, len,
 282                                 pgoff, flags);
 283         return hugetlb_get_unmapped_area_bottomup(file, addr, len,
 284                         pgoff, flags);
 285 }
 286 #endif
 287
 288 static size_t
 289 hugetlbfs_read_actor(struct page *page, unsigned long offset,
 290                         struct iov_iter *to, unsigned long size)
 291 {
 292         size_t copied = 0;
 293         int i, chunksize;
 294
 295         /* Find which 4k chunk and offset with in that chunk */
 296         i = offset >> PAGE_SHIFT;
 297         offset = offset & ~PAGE_MASK;
 298
 299         while (size) {
 300                 size_t n;
 301                 chunksize = PAGE_SIZE;
 302                 if (offset)
 303                         chunksize -= offset;
 304                 if (chunksize > size)
 305                         chunksize = size;
 306                 n = copy_page_to_iter(&page[i], offset, chunksize, to);
 307                 copied += n;
 308                 if (n != chunksize)
 309                         return copied;
 310                 offset = 0;
 311                 size -= chunksize;
 312                 i++;
 313         }
 314         return copied;
 315 }
 316
 317 /*
 318  * Support for read() - Find the page attached to f_mapping and copy out the
 319  * data. Its *very* similar to generic_file_buffered_read(), we can't use that
 320  * since it has PAGE_SIZE assumptions.
 321  */
 322 static ssize_t hugetlbfs_read_iter(struct kiocb *iocb, struct iov_iter *to)
 323 {
 324         struct file *file = iocb->ki_filp;
 325         struct hstate *h = hstate_file(file);
 326         struct address_space *mapping = file->f_mapping;
 327         struct inode *inode = mapping->host;
 328         unsigned long index = iocb->ki_pos >> huge_page_shift(h);
 329         unsigned long offset = iocb->ki_pos & ~huge_page_mask(h);
 330         unsigned long end_index;
 331         loff_t isize;
 332         ssize_t retval = 0;
 333
 334         while (iov_iter_count(to)) {
 335                 struct page *page;
 336                 size_t nr, copied;
 337
 338                 /* nr is the maximum number of bytes to copy from this page */
 339                 nr = huge_page_size(h);
 340                 isize = i_size_read(inode);
 341                 if (!isize)
 342                         break;
 343                 end_index = (isize - 1) >> huge_page_shift(h);
 344                 if (index > end_index)
 345                         break;
 346                 if (index == end_index) {
 347                         nr = ((isize - 1) & ~huge_page_mask(h)) + 1;
 348                         if (nr <= offset)
 349                                 break;
 350                 }
 351                 nr = nr - offset;
 352
 353                 /* Find the page */
 354                 page = find_lock_page(mapping, index);
 355                 if (unlikely(page == NULL)) {
 356                         /*
 357                          * We have a HOLE, zero out the user-buffer for the
 358                          * length of the hole or request.
 359                          */
 360                         copied = iov_iter_zero(nr, to);
 361                 } else {
 362                         unlock_page(page);
 363
 364                         /*
 365                          * We have the page, copy it to user space buffer.
 366                          */
 367                         copied = hugetlbfs_read_actor(page, offset, to, nr);
 368                         put_page(page);
 369                 }
 370                 offset += copied;
 371                 retval += copied;
 372                 if (copied != nr && iov_iter_count(to)) {
 373                         if (!retval)
 374                                 retval = -EFAULT;
 375                         break;
 376                 }
 377                 index += offset >> huge_page_shift(h);
 378                 offset &= ~huge_page_mask(h);
 379         }
 380         iocb->ki_pos = ((loff_t)index << huge_page_shift(h)) + offset;
 381         return retval;
 382 }
 383
 384 static int hugetlbfs_write_begin(struct file *file,
 385                         struct address_space *mapping,
 386                         loff_t pos, unsigned len,
 387                         struct page **pagep, void **fsdata)
 388 {
 389         return -EINVAL;
 390 }
 391
 392 static int hugetlbfs_write_end(struct file *file, struct address_space *mapping,
 393                         loff_t pos, unsigned len, unsigned copied,
 394                         struct page *page, void *fsdata)
 395 {
 396         BUG();
 397         return -EINVAL;
 398 }
 399
 400 static void remove_huge_page(struct page *page)
 401 {
 402         ClearPageDirty(page);
 403         ClearPageUptodate(page);
 404         delete_from_page_cache(page);
 405 }
 406
 407 static void
 408 hugetlb_vmdelete_list(struct rb_root_cached *root, pgoff_t start, pgoff_t end,
 409                       zap_flags_t zap_flags)
 410 {
 411         struct vm_area_struct *vma;
 412
 413         /*
 414          * end == 0 indicates that the entire range after start should be
 415          * unmapped.  Note, end is exclusive, whereas the interval tree takes
 416          * an inclusive "last".
 417          */
 418         vma_interval_tree_foreach(vma, root, start, end ? end - 1 : ULONG_MAX) {
 419                 unsigned long v_offset;
 420                 unsigned long v_end;
 421
 422                 /*
 423                  * Can the expression below overflow on 32-bit arches?
 424                  * No, because the interval tree returns us only those vmas
 425                  * which overlap the truncated area starting at pgoff,
 426                  * and no vma on a 32-bit arch can span beyond the 4GB.
 427                  */
 428                 if (vma->vm_pgoff < start)
 429                         v_offset = (start - vma->vm_pgoff) << PAGE_SHIFT;
 430                 else
 431                         v_offset = 0;
 432
 433                 if (!end)
 434                         v_end = vma->vm_end;
 435                 else {
 436                         v_end = ((end - vma->vm_pgoff) << PAGE_SHIFT)
 437                                                         + vma->vm_start;
 438                         if (v_end > vma->vm_end)
 439                                 v_end = vma->vm_end;
 440                 }
 441
 442                 unmap_hugepage_range(vma, vma->vm_start + v_offset, v_end,
 443                                      NULL, zap_flags);
 444         }
 445 }
 446
 447 /*
 448  * remove_inode_hugepages handles two distinct cases: truncation and hole
 449  * punch.  There are subtle differences in operation for each case.
 450  *
 451  * truncation is indicated by end of range being LLONG_MAX
 452  *      In this case, we first scan the range and release found pages.
 453  *      After releasing pages, hugetlb_unreserve_pages cleans up region/reserve
 454  *      maps and global counts.  Page faults can not race with truncation
 455  *      in this routine.  hugetlb_no_page() holds i_mmap_rwsem and prevents
 456  *      page faults in the truncated range by checking i_size.  i_size is
 457  *      modified while holding i_mmap_rwsem.
 458  * hole punch is indicated if end is not LLONG_MAX
 459  *      In the hole punch case we scan the range and release found pages.
 460  *      Only when releasing a page is the associated region/reserve map
 461  *      deleted.  The region/reserve map for ranges without associated
 462  *      pages are not modified.  Page faults can race with hole punch.
 463  *      This is indicated if we find a mapped page.
 464  * Note: If the passed end of range value is beyond the end of file, but
 465  * not LLONG_MAX this routine still performs a hole punch operation.
 466  */
 467 static void remove_inode_hugepages(struct inode *inode, loff_t lstart,
 468                                    loff_t lend)
 469 {
 470         struct hstate *h = hstate_inode(inode);
 471         struct address_space *mapping = &inode->i_data;
 472         const pgoff_t start = lstart >> huge_page_shift(h);
 473         const pgoff_t end = lend >> huge_page_shift(h);
 474         struct pagevec pvec;
 475         pgoff_t next, index;
 476         int i, freed = 0;
 477         bool truncate_op = (lend == LLONG_MAX);
 478
 479         pagevec_init(&pvec);
 480         next = start;
 481         while (next < end) {
 482                 /*
 483                  * When no more pages are found, we are done.
 484                  */
 485                 if (!pagevec_lookup_range(&pvec, mapping, &next, end - 1))
 486                         break;
 487
 488                 for (i = 0; i < pagevec_count(&pvec); ++i) {
 489                         struct page *page = pvec.pages[i];
 490                         u32 hash = 0;
 491
 492                         index = page->index;
 493                         if (!truncate_op) {
 494                                 /*
 495                                  * Only need to hold the fault mutex in the
 496                                  * hole punch case.  This prevents races with
 497                                  * page faults.  Races are not possible in the
 498                                  * case of truncation.
 499                                  */
 500                                 hash = hugetlb_fault_mutex_hash(mapping, index);
 501                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 502                         }
 503
 504                         /*
 505                          * If page is mapped, it was faulted in after being
 506                          * unmapped in caller.  Unmap (again) now after taking
 507                          * the fault mutex.  The mutex will prevent faults
 508                          * until we finish removing the page.
 509                          *
 510                          * This race can only happen in the hole punch case.
 511                          * Getting here in a truncate operation is a bug.
 512                          */
 513                         if (unlikely(page_mapped(page))) {
 514                                 BUG_ON(truncate_op);
 515
 516                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 517                                 i_mmap_lock_write(mapping);
 518                                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 519                                 hugetlb_vmdelete_list(&mapping->i_mmap,
 520                                         index * pages_per_huge_page(h),
 521                                         (index + 1) * pages_per_huge_page(h),
 522                                         ZAP_FLAG_DROP_MARKER);
 523                                 i_mmap_unlock_write(mapping);
 524                         }
 525
 526                         lock_page(page);
 527                         /*
 528                          * We must free the huge page and remove from page
 529                          * cache (remove_huge_page) BEFORE removing the
 530                          * region/reserve map (hugetlb_unreserve_pages).  In
 531                          * rare out of memory conditions, removal of the
 532                          * region/reserve map could fail. Correspondingly,
 533                          * the subpool and global reserve usage count can need
 534                          * to be adjusted.
 535                          */
 536                         VM_BUG_ON(HPageRestoreReserve(page));
 537                         remove_huge_page(page);
 538                         freed++;
 539                         if (!truncate_op) {
 540                                 if (unlikely(hugetlb_unreserve_pages(inode,
 541                                                         index, index + 1, 1)))
 542                                         hugetlb_fix_reserve_counts(inode);
 543                         }
 544
 545                         unlock_page(page);
 546                         if (!truncate_op)
 547                                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 548                 }
 549                 huge_pagevec_release(&pvec);
 550                 cond_resched();
 551         }
 552
 553         if (truncate_op)
 554                 (void)hugetlb_unreserve_pages(inode, start, LONG_MAX, freed);
 555 }
 556
 557 static void hugetlbfs_evict_inode(struct inode *inode)
 558 {
 559         struct resv_map *resv_map;
 560
 561         remove_inode_hugepages(inode, 0, LLONG_MAX);
 562
 563         /*
 564          * Get the resv_map from the address space embedded in the inode.
 565          * This is the address space which points to any resv_map allocated
 566          * at inode creation time.  If this is a device special inode,
 567          * i_mapping may not point to the original address space.
 568          */
 569         resv_map = (struct resv_map *)(&inode->i_data)->private_data;
 570         /* Only regular and link inodes have associated reserve maps */
 571         if (resv_map)
 572                 resv_map_release(&resv_map->refs);
 573         clear_inode(inode);
 574 }
 575
 576 static void hugetlb_vmtruncate(struct inode *inode, loff_t offset)
 577 {
 578         pgoff_t pgoff;
 579         struct address_space *mapping = inode->i_mapping;
 580         struct hstate *h = hstate_inode(inode);
 581
 582         BUG_ON(offset & ~huge_page_mask(h));
 583         pgoff = offset >> PAGE_SHIFT;
 584
 585         i_mmap_lock_write(mapping);
 586         i_size_write(inode, offset);
 587         if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 588                 hugetlb_vmdelete_list(&mapping->i_mmap, pgoff, 0,
 589                                       ZAP_FLAG_DROP_MARKER);
 590         i_mmap_unlock_write(mapping);
 591         remove_inode_hugepages(inode, offset, LLONG_MAX);
 592 }
 593
 594 static long hugetlbfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
 595 {
 596         struct hstate *h = hstate_inode(inode);
 597         loff_t hpage_size = huge_page_size(h);
 598         loff_t hole_start, hole_end;
 599
 600         /*
 601          * For hole punch round up the beginning offset of the hole and
 602          * round down the end.
 603          */
 604         hole_start = round_up(offset, hpage_size);
 605         hole_end = round_down(offset + len, hpage_size);
 606
 607         if (hole_end > hole_start) {
 608                 struct address_space *mapping = inode->i_mapping;
 609                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 610
 611                 inode_lock(inode);
 612
 613                 /* protected by i_rwsem */
 614                 if (info->seals & (F_SEAL_WRITE | F_SEAL_FUTURE_WRITE)) {
 615                         inode_unlock(inode);
 616                         return -EPERM;
 617                 }
 618
 619                 i_mmap_lock_write(mapping);
 620                 if (!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root))
 621                         hugetlb_vmdelete_list(&mapping->i_mmap,
 622                                               hole_start >> PAGE_SHIFT,
 623                                               hole_end >> PAGE_SHIFT, 0);
 624                 i_mmap_unlock_write(mapping);
 625                 remove_inode_hugepages(inode, hole_start, hole_end);
 626                 inode_unlock(inode);
 627         }
 628
 629         return 0;
 630 }
 631
 632 static long hugetlbfs_fallocate(struct file *file, int mode, loff_t offset,
 633                                 loff_t len)
 634 {
 635         struct inode *inode = file_inode(file);
 636         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 637         struct address_space *mapping = inode->i_mapping;
 638         struct hstate *h = hstate_inode(inode);
 639         struct vm_area_struct pseudo_vma;
 640         struct mm_struct *mm = current->mm;
 641         loff_t hpage_size = huge_page_size(h);
 642         unsigned long hpage_shift = huge_page_shift(h);
 643         pgoff_t start, index, end;
 644         int error;
 645         u32 hash;
 646
 647         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
 648                 return -EOPNOTSUPP;
 649
 650         if (mode & FALLOC_FL_PUNCH_HOLE)
 651                 return hugetlbfs_punch_hole(inode, offset, len);
 652
 653         /*
 654          * Default preallocate case.
 655          * For this range, start is rounded down and end is rounded up
 656          * as well as being converted to page offsets.
 657          */
 658         start = offset >> hpage_shift;
 659         end = (offset + len + hpage_size - 1) >> hpage_shift;
 660
 661         inode_lock(inode);
 662
 663         /* We need to check rlimit even when FALLOC_FL_KEEP_SIZE */
 664         error = inode_newsize_ok(inode, offset + len);
 665         if (error)
 666                 goto out;
 667
 668         if ((info->seals & F_SEAL_GROW) && offset + len > inode->i_size) {
 669                 error = -EPERM;
 670                 goto out;
 671         }
 672
 673         /*
 674          * Initialize a pseudo vma as this is required by the huge page
 675          * allocation routines.  If NUMA is configured, use page index
 676          * as input to create an allocation policy.
 677          */
 678         vma_init(&pseudo_vma, mm);
 679         pseudo_vma.vm_flags = (VM_HUGETLB | VM_MAYSHARE | VM_SHARED);
 680         pseudo_vma.vm_file = file;
 681
 682         for (index = start; index < end; index++) {
 683                 /*
 684                  * This is supposed to be the vaddr where the page is being
 685                  * faulted in, but we have no vaddr here.
 686                  */
 687                 struct page *page;
 688                 unsigned long addr;
 689
 690                 cond_resched();
 691
 692                 /*
 693                  * fallocate(2) manpage permits EINTR; we may have been
 694                  * interrupted because we are using up too much memory.
 695                  */
 696                 if (signal_pending(current)) {
 697                         error = -EINTR;
 698                         break;
 699                 }
 700
 701                 /* Set numa allocation policy based on index */
 702                 hugetlb_set_vma_policy(&pseudo_vma, inode, index);
 703
 704                 /* addr is the offset within the file (zero based) */
 705                 addr = index * hpage_size;
 706
 707                 /*
 708                  * fault mutex taken here, protects against fault path
 709                  * and hole punch.  inode_lock previously taken protects
 710                  * against truncation.
 711                  */
 712                 hash = hugetlb_fault_mutex_hash(mapping, index);
 713                 mutex_lock(&hugetlb_fault_mutex_table[hash]);
 714
 715                 /* See if already present in mapping to avoid alloc/free */
 716                 page = find_get_page(mapping, index);
 717                 if (page) {
 718                         put_page(page);
 719                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 720                         hugetlb_drop_vma_policy(&pseudo_vma);
 721                         continue;
 722                 }
 723
 724                 /*
 725                  * Allocate page without setting the avoid_reserve argument.
 726                  * There certainly are no reserves associated with the
 727                  * pseudo_vma.  However, there could be shared mappings with
 728                  * reserves for the file at the inode level.  If we fallocate
 729                  * pages in these areas, we need to consume the reserves
 730                  * to keep reservation accounting consistent.
 731                  */
 732                 page = alloc_huge_page(&pseudo_vma, addr, 0);
 733                 hugetlb_drop_vma_policy(&pseudo_vma);
 734                 if (IS_ERR(page)) {
 735                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 736                         error = PTR_ERR(page);
 737                         goto out;
 738                 }
 739                 clear_huge_page(page, addr, pages_per_huge_page(h));
 740                 __SetPageUptodate(page);
 741                 error = huge_add_to_page_cache(page, mapping, index);
 742                 if (unlikely(error)) {
 743                         restore_reserve_on_error(h, &pseudo_vma, addr, page);
 744                         put_page(page);
 745                         mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 746                         goto out;
 747                 }
 748
 749                 mutex_unlock(&hugetlb_fault_mutex_table[hash]);
 750
 751                 SetHPageMigratable(page);
 752                 /*
 753                  * unlock_page because locked by add_to_page_cache()
 754                  * put_page() due to reference from alloc_huge_page()
 755                  */
 756                 unlock_page(page);
 757                 put_page(page);
 758         }
 759
 760         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size)
 761                 i_size_write(inode, offset + len);
 762         inode->i_ctime = current_time(inode);
 763 out:
 764         inode_unlock(inode);
 765         return error;
 766 }
 767
 768 static int hugetlbfs_setattr(struct user_namespace *mnt_userns,
 769                              struct dentry *dentry, struct iattr *attr)
 770 {
 771         struct inode *inode = d_inode(dentry);
 772         struct hstate *h = hstate_inode(inode);
 773         int error;
 774         unsigned int ia_valid = attr->ia_valid;
 775         struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 776
 777         error = setattr_prepare(&init_user_ns, dentry, attr);
 778         if (error)
 779                 return error;
 780
 781         if (ia_valid & ATTR_SIZE) {
 782                 loff_t oldsize = inode->i_size;
 783                 loff_t newsize = attr->ia_size;
 784
 785                 if (newsize & ~huge_page_mask(h))
 786                         return -EINVAL;
 787                 /* protected by i_rwsem */
 788                 if ((newsize < oldsize && (info->seals & F_SEAL_SHRINK)) ||
 789                     (newsize > oldsize && (info->seals & F_SEAL_GROW)))
 790                         return -EPERM;
 791                 hugetlb_vmtruncate(inode, newsize);
 792         }
 793
 794         setattr_copy(&init_user_ns, inode, attr);
 795         mark_inode_dirty(inode);
 796         return 0;
 797 }
 798
 799 static struct inode *hugetlbfs_get_root(struct super_block *sb,
 800                                         struct hugetlbfs_fs_context *ctx)
 801 {
 802         struct inode *inode;
 803
 804         inode = new_inode(sb);
 805         if (inode) {
 806                 inode->i_ino = get_next_ino();
 807                 inode->i_mode = S_IFDIR | ctx->mode;
 808                 inode->i_uid = ctx->uid;
 809                 inode->i_gid = ctx->gid;
 810                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 811                 inode->i_op = &hugetlbfs_dir_inode_operations;
 812                 inode->i_fop = &simple_dir_operations;
 813                 /* directory inodes start off with i_nlink == 2 (for "." entry) */
 814                 inc_nlink(inode);
 815                 lockdep_annotate_inode_mutex_key(inode);
 816         }
 817         return inode;
 818 }
 819
 820 /*
 821  * Hugetlbfs is not reclaimable; therefore its i_mmap_rwsem will never
 822  * be taken from reclaim -- unlike regular filesystems. This needs an
 823  * annotation because huge_pmd_share() does an allocation under hugetlb's
 824  * i_mmap_rwsem.
 825  */
 826 static struct lock_class_key hugetlbfs_i_mmap_rwsem_key;
 827
 828 static struct inode *hugetlbfs_get_inode(struct super_block *sb,
 829                                         struct inode *dir,
 830                                         umode_t mode, dev_t dev)
 831 {
 832         struct inode *inode;
 833         struct resv_map *resv_map = NULL;
 834
 835         /*
 836          * Reserve maps are only needed for inodes that can have associated
 837          * page allocations.
 838          */
 839         if (S_ISREG(mode) || S_ISLNK(mode)) {
 840                 resv_map = resv_map_alloc();
 841                 if (!resv_map)
 842                         return NULL;
 843         }
 844
 845         inode = new_inode(sb);
 846         if (inode) {
 847                 struct hugetlbfs_inode_info *info = HUGETLBFS_I(inode);
 848
 849                 inode->i_ino = get_next_ino();
 850                 inode_init_owner(&init_user_ns, inode, dir, mode);
 851                 lockdep_set_class(&inode->i_mapping->i_mmap_rwsem,
 852                                 &hugetlbfs_i_mmap_rwsem_key);
 853                 inode->i_mapping->a_ops = &hugetlbfs_aops;
 854                 inode->i_atime = inode->i_mtime = inode->i_ctime = current_time(inode);
 855                 inode->i_mapping->private_data = resv_map;
 856                 info->seals = F_SEAL_SEAL;
 857                 switch (mode & S_IFMT) {
 858                 default:
 859                         init_special_inode(inode, mode, dev);
 860                         break;
 861                 case S_IFREG:
 862                         inode->i_op = &hugetlbfs_inode_operations;
 863                         inode->i_fop = &hugetlbfs_file_operations;
 864                         break;
 865                 case S_IFDIR:
 866                         inode->i_op = &hugetlbfs_dir_inode_operations;
 867                         inode->i_fop = &simple_dir_operations;
 868
 869                         /* directory inodes start off with i_nlink == 2 (for "." entry) */
 870                         inc_nlink(inode);
 871                         break;
 872                 case S_IFLNK:
 873                         inode->i_op = &page_symlink_inode_operations;
 874                         inode_nohighmem(inode);
 875                         break;
 876                 }
 877                 lockdep_annotate_inode_mutex_key(inode);
 878         } else {
 879                 if (resv_map)
 880                         kref_put(&resv_map->refs, resv_map_release);
 881         }
 882
 883         return inode;
 884 }
 885
 886 /*
 887  * File creation. Allocate an inode, and we're done..
 888  */
 889 static int do_hugetlbfs_mknod(struct inode *dir,
 890                         struct dentry *dentry,
 891                         umode_t mode,
 892                         dev_t dev,
 893                         bool tmpfile)
 894 {
 895         struct inode *inode;
 896         int error = -ENOSPC;
 897
 898         inode = hugetlbfs_get_inode(dir->i_sb, dir, mode, dev);
 899         if (inode) {
 900                 dir->i_ctime = dir->i_mtime = current_time(dir);
 901                 if (tmpfile) {
 902                         d_tmpfile(dentry, inode);
 903                 } else {
 904                         d_instantiate(dentry, inode);
 905                         dget(dentry);/* Extra count - pin the dentry in core */
 906                 }
 907                 error = 0;
 908         }
 909         return error;
 910 }
 911
 912 static int hugetlbfs_mknod(struct user_namespace *mnt_userns, struct inode *dir,
 913                            struct dentry *dentry, umode_t mode, dev_t dev)
 914 {
 915         return do_hugetlbfs_mknod(dir, dentry, mode, dev, false);
 916 }
 917
 918 static int hugetlbfs_mkdir(struct user_namespace *mnt_userns, struct inode *dir,
 919                            struct dentry *dentry, umode_t mode)
 920 {
 921         int retval = hugetlbfs_mknod(&init_user_ns, dir, dentry,
 922                                      mode | S_IFDIR, 0);
 923         if (!retval)
 924                 inc_nlink(dir);
 925         return retval;
 926 }
 927
 928 static int hugetlbfs_create(struct user_namespace *mnt_userns,
 929                             struct inode *dir, struct dentry *dentry,
 930                             umode_t mode, bool excl)
 931 {
 932         return hugetlbfs_mknod(&init_user_ns, dir, dentry, mode | S_IFREG, 0);
 933 }
 934
 935 static int hugetlbfs_tmpfile(struct user_namespace *mnt_userns,
 936                              struct inode *dir, struct dentry *dentry,
 937                              umode_t mode)
 938 {
 939         return do_hugetlbfs_mknod(dir, dentry, mode | S_IFREG, 0, true);
 940 }
 941
 942 static int hugetlbfs_symlink(struct user_namespace *mnt_userns,
 943                              struct inode *dir, struct dentry *dentry,
 944                              const char *symname)
 945 {
 946         struct inode *inode;
 947         int error = -ENOSPC;
 948
 949         inode = hugetlbfs_get_inode(dir->i_sb, dir, S_IFLNK|S_IRWXUGO, 0);
 950         if (inode) {
 951                 int l = strlen(symname)+1;
 952                 error = page_symlink(inode, symname, l);
 953                 if (!error) {
 954                         d_instantiate(dentry, inode);
 955                         dget(dentry);
 956                 } else
 957                         iput(inode);
 958         }
 959         dir->i_ctime = dir->i_mtime = current_time(dir);
 960
 961         return error;
 962 }
 963
 964 static int hugetlbfs_migrate_page(struct address_space *mapping,
 965                                 struct page *newpage, struct page *page,
 966                                 enum migrate_mode mode)
 967 {
 968         int rc;
 969
 970         rc = migrate_huge_page_move_mapping(mapping, newpage, page);
 971         if (rc != MIGRATEPAGE_SUCCESS)
 972                 return rc;
 973
 974         if (hugetlb_page_subpool(page)) {
 975                 hugetlb_set_page_subpool(newpage, hugetlb_page_subpool(page));
 976                 hugetlb_set_page_subpool(page, NULL);
 977         }
 978
 979         if (mode != MIGRATE_SYNC_NO_COPY)
 980                 migrate_page_copy(newpage, page);
 981         else
 982                 migrate_page_states(newpage, page);
 983
 984         return MIGRATEPAGE_SUCCESS;
 985 }
 986
 987 static int hugetlbfs_error_remove_page(struct address_space *mapping,
 988                                 struct page *page)
 989 {
 990         struct inode *inode = mapping->host;
 991         pgoff_t index = page->index;
 992
 993         remove_huge_page(page);
 994         if (unlikely(hugetlb_unreserve_pages(inode, index, index + 1, 1)))
 995                 hugetlb_fix_reserve_counts(inode);
 996
 997         return 0;
 998 }
 999
1000 /*
1001  * Display the mount options in /proc/mounts.
1002  */
1003 static int hugetlbfs_show_options(struct seq_file *m, struct dentry *root)
1004 {
1005         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(root->d_sb);
1006         struct hugepage_subpool *spool = sbinfo->spool;
1007         unsigned long hpage_size = huge_page_size(sbinfo->hstate);
1008         unsigned hpage_shift = huge_page_shift(sbinfo->hstate);
1009         char mod;
1010
1011         if (!uid_eq(sbinfo->uid, GLOBAL_ROOT_UID))
1012                 seq_printf(m, ",uid=%u",
1013                            from_kuid_munged(&init_user_ns, sbinfo->uid));
1014         if (!gid_eq(sbinfo->gid, GLOBAL_ROOT_GID))
1015                 seq_printf(m, ",gid=%u",
1016                            from_kgid_munged(&init_user_ns, sbinfo->gid));
1017         if (sbinfo->mode != 0755)
1018                 seq_printf(m, ",mode=%o", sbinfo->mode);
1019         if (sbinfo->max_inodes != -1)
1020                 seq_printf(m, ",nr_inodes=%lu", sbinfo->max_inodes);
1021
1022         hpage_size /= 1024;
1023         mod = 'K';
1024         if (hpage_size >= 1024) {
1025                 hpage_size /= 1024;
1026                 mod = 'M';
1027         }
1028         seq_printf(m, ",pagesize=%lu%c", hpage_size, mod);
1029         if (spool) {
1030                 if (spool->max_hpages != -1)
1031                         seq_printf(m, ",size=%llu",
1032                                    (unsigned long long)spool->max_hpages << hpage_shift);
1033                 if (spool->min_hpages != -1)
1034                         seq_printf(m, ",min_size=%llu",
1035                                    (unsigned long long)spool->min_hpages << hpage_shift);
1036         }
1037         return 0;
1038 }
1039
1040 static int hugetlbfs_statfs(struct dentry *dentry, struct kstatfs *buf)
1041 {
1042         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(dentry->d_sb);
1043         struct hstate *h = hstate_inode(d_inode(dentry));
1044
1045         buf->f_type = HUGETLBFS_MAGIC;
1046         buf->f_bsize = huge_page_size(h);
1047         if (sbinfo) {
1048                 spin_lock(&sbinfo->stat_lock);
1049                 /* If no limits set, just report 0 for max/free/used
1050                  * blocks, like simple_statfs() */
1051                 if (sbinfo->spool) {
1052                         long free_pages;
1053
1054                         spin_lock_irq(&sbinfo->spool->lock);
1055                         buf->f_blocks = sbinfo->spool->max_hpages;
1056                         free_pages = sbinfo->spool->max_hpages
1057                                 - sbinfo->spool->used_hpages;
1058                         buf->f_bavail = buf->f_bfree = free_pages;
1059                         spin_unlock_irq(&sbinfo->spool->lock);
1060                         buf->f_files = sbinfo->max_inodes;
1061                         buf->f_ffree = sbinfo->free_inodes;
1062                 }
1063                 spin_unlock(&sbinfo->stat_lock);
1064         }
1065         buf->f_namelen = NAME_MAX;
1066         return 0;
1067 }
1068
1069 static void hugetlbfs_put_super(struct super_block *sb)
1070 {
1071         struct hugetlbfs_sb_info *sbi = HUGETLBFS_SB(sb);
1072
1073         if (sbi) {
1074                 sb->s_fs_info = NULL;
1075
1076                 if (sbi->spool)
1077                         hugepage_put_subpool(sbi->spool);
1078
1079                 kfree(sbi);
1080         }
1081 }
1082
1083 static inline int hugetlbfs_dec_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1084 {
1085         if (sbinfo->free_inodes >= 0) {
1086                 spin_lock(&sbinfo->stat_lock);
1087                 if (unlikely(!sbinfo->free_inodes)) {
1088                         spin_unlock(&sbinfo->stat_lock);
1089                         return 0;
1090                 }
1091                 sbinfo->free_inodes--;
1092                 spin_unlock(&sbinfo->stat_lock);
1093         }
1094
1095         return 1;
1096 }
1097
1098 static void hugetlbfs_inc_free_inodes(struct hugetlbfs_sb_info *sbinfo)
1099 {
1100         if (sbinfo->free_inodes >= 0) {
1101                 spin_lock(&sbinfo->stat_lock);
1102                 sbinfo->free_inodes++;
1103                 spin_unlock(&sbinfo->stat_lock);
1104         }
1105 }
1106
1107
1108 static struct kmem_cache *hugetlbfs_inode_cachep;
1109
1110 static struct inode *hugetlbfs_alloc_inode(struct super_block *sb)
1111 {
1112         struct hugetlbfs_sb_info *sbinfo = HUGETLBFS_SB(sb);
1113         struct hugetlbfs_inode_info *p;
1114
1115         if (unlikely(!hugetlbfs_dec_free_inodes(sbinfo)))
1116                 return NULL;
1117         p = alloc_inode_sb(sb, hugetlbfs_inode_cachep, GFP_KERNEL);
1118         if (unlikely(!p)) {
1119                 hugetlbfs_inc_free_inodes(sbinfo);
1120                 return NULL;
1121         }
1122
1123         /*
1124          * Any time after allocation, hugetlbfs_destroy_inode can be called
1125          * for the inode.  mpol_free_shared_policy is unconditionally called
1126          * as part of hugetlbfs_destroy_inode.  So, initialize policy here
1127          * in case of a quick call to destroy.
1128          *
1129          * Note that the policy is initialized even if we are creating a
1130          * private inode.  This simplifies hugetlbfs_destroy_inode.
1131          */
1132         mpol_shared_policy_init(&p->policy, NULL);
1133
1134         return &p->vfs_inode;
1135 }
1136
1137 static void hugetlbfs_free_inode(struct inode *inode)
1138 {
1139         kmem_cache_free(hugetlbfs_inode_cachep, HUGETLBFS_I(inode));
1140 }
1141
1142 static void hugetlbfs_destroy_inode(struct inode *inode)
1143 {
1144         hugetlbfs_inc_free_inodes(HUGETLBFS_SB(inode->i_sb));
1145         mpol_free_shared_policy(&HUGETLBFS_I(inode)->policy);
1146 }
1147
1148 static const struct address_space_operations hugetlbfs_aops = {
1149         .write_begin    = hugetlbfs_write_begin,
1150         .write_end      = hugetlbfs_write_end,
1151         .dirty_folio    = noop_dirty_folio,
1152         .migratepage    = hugetlbfs_migrate_page,
1153         .error_remove_page      = hugetlbfs_error_remove_page,
1154 };
1155
1156
1157 static void init_once(void *foo)
1158 {
1159         struct hugetlbfs_inode_info *ei = (struct hugetlbfs_inode_info *)foo;
1160
1161         inode_init_once(&ei->vfs_inode);
1162 }
1163
1164 const struct file_operations hugetlbfs_file_operations = {
1165         .read_iter              = hugetlbfs_read_iter,
1166         .mmap                   = hugetlbfs_file_mmap,
1167         .fsync                  = noop_fsync,
1168         .get_unmapped_area      = hugetlb_get_unmapped_area,
1169         .llseek                 = default_llseek,
1170         .fallocate              = hugetlbfs_fallocate,
1171 };
1172
1173 static const struct inode_operations hugetlbfs_dir_inode_operations = {
1174         .create         = hugetlbfs_create,
1175         .lookup         = simple_lookup,
1176         .link           = simple_link,
1177         .unlink         = simple_unlink,
1178         .symlink        = hugetlbfs_symlink,
1179         .mkdir          = hugetlbfs_mkdir,
1180         .rmdir          = simple_rmdir,
1181         .mknod          = hugetlbfs_mknod,
1182         .rename         = simple_rename,
1183         .setattr        = hugetlbfs_setattr,
1184         .tmpfile        = hugetlbfs_tmpfile,
1185 };
1186
1187 static const struct inode_operations hugetlbfs_inode_operations = {
1188         .setattr        = hugetlbfs_setattr,
1189 };
1190
1191 static const struct super_operations hugetlbfs_ops = {
1192         .alloc_inode    = hugetlbfs_alloc_inode,
1193         .free_inode     = hugetlbfs_free_inode,
1194         .destroy_inode  = hugetlbfs_destroy_inode,
1195         .evict_inode    = hugetlbfs_evict_inode,
1196         .statfs         = hugetlbfs_statfs,
1197         .put_super      = hugetlbfs_put_super,
1198         .show_options   = hugetlbfs_show_options,
1199 };
1200
1201 /*
1202  * Convert size option passed from command line to number of huge pages
1203  * in the pool specified by hstate.  Size option could be in bytes
1204  * (val_type == SIZE_STD) or percentage of the pool (val_type == SIZE_PERCENT).
1205  */
1206 static long
1207 hugetlbfs_size_to_hpages(struct hstate *h, unsigned long long size_opt,
1208                          enum hugetlbfs_size_type val_type)
1209 {
1210         if (val_type == NO_SIZE)
1211                 return -1;
1212
1213         if (val_type == SIZE_PERCENT) {
1214                 size_opt <<= huge_page_shift(h);
1215                 size_opt *= h->max_huge_pages;
1216                 do_div(size_opt, 100);
1217         }
1218
1219         size_opt >>= huge_page_shift(h);
1220         return size_opt;
1221 }
1222
1223 /*
1224  * Parse one mount parameter.
1225  */
1226 static int hugetlbfs_parse_param(struct fs_context *fc, struct fs_parameter *param)
1227 {
1228         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1229         struct fs_parse_result result;
1230         char *rest;
1231         unsigned long ps;
1232         int opt;
1233
1234         opt = fs_parse(fc, hugetlb_fs_parameters, param, &result);
1235         if (opt < 0)
1236                 return opt;
1237
1238         switch (opt) {
1239         case Opt_uid:
1240                 ctx->uid = make_kuid(current_user_ns(), result.uint_32);
1241                 if (!uid_valid(ctx->uid))
1242                         goto bad_val;
1243                 return 0;
1244
1245         case Opt_gid:
1246                 ctx->gid = make_kgid(current_user_ns(), result.uint_32);
1247                 if (!gid_valid(ctx->gid))
1248                         goto bad_val;
1249                 return 0;
1250
1251         case Opt_mode:
1252                 ctx->mode = result.uint_32 & 01777U;
1253                 return 0;
1254
1255         case Opt_size:
1256                 /* memparse() will accept a K/M/G without a digit */
1257                 if (!isdigit(param->string[0]))
1258                         goto bad_val;
1259                 ctx->max_size_opt = memparse(param->string, &rest);
1260                 ctx->max_val_type = SIZE_STD;
1261                 if (*rest == '%')
1262                         ctx->max_val_type = SIZE_PERCENT;
1263                 return 0;
1264
1265         case Opt_nr_inodes:
1266                 /* memparse() will accept a K/M/G without a digit */
1267                 if (!isdigit(param->string[0]))
1268                         goto bad_val;
1269                 ctx->nr_inodes = memparse(param->string, &rest);
1270                 return 0;
1271
1272         case Opt_pagesize:
1273                 ps = memparse(param->string, &rest);
1274                 ctx->hstate = size_to_hstate(ps);
1275                 if (!ctx->hstate) {
1276                         pr_err("Unsupported page size %lu MB\n", ps >> 20);
1277                         return -EINVAL;
1278                 }
1279                 return 0;
1280
1281         case Opt_min_size:
1282                 /* memparse() will accept a K/M/G without a digit */
1283                 if (!isdigit(param->string[0]))
1284                         goto bad_val;
1285                 ctx->min_size_opt = memparse(param->string, &rest);
1286                 ctx->min_val_type = SIZE_STD;
1287                 if (*rest == '%')
1288                         ctx->min_val_type = SIZE_PERCENT;
1289                 return 0;
1290
1291         default:
1292                 return -EINVAL;
1293         }
1294
1295 bad_val:
1296         return invalfc(fc, "Bad value '%s' for mount option '%s'\n",
1297                       param->string, param->key);
1298 }
1299
1300 /*
1301  * Validate the parsed options.
1302  */
1303 static int hugetlbfs_validate(struct fs_context *fc)
1304 {
1305         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1306
1307         /*
1308          * Use huge page pool size (in hstate) to convert the size
1309          * options to number of huge pages.  If NO_SIZE, -1 is returned.
1310          */
1311         ctx->max_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1312                                                    ctx->max_size_opt,
1313                                                    ctx->max_val_type);
1314         ctx->min_hpages = hugetlbfs_size_to_hpages(ctx->hstate,
1315                                                    ctx->min_size_opt,
1316                                                    ctx->min_val_type);
1317
1318         /*
1319          * If max_size was specified, then min_size must be smaller
1320          */
1321         if (ctx->max_val_type > NO_SIZE &&
1322             ctx->min_hpages > ctx->max_hpages) {
1323                 pr_err("Minimum size can not be greater than maximum size\n");
1324                 return -EINVAL;
1325         }
1326
1327         return 0;
1328 }
1329
1330 static int
1331 hugetlbfs_fill_super(struct super_block *sb, struct fs_context *fc)
1332 {
1333         struct hugetlbfs_fs_context *ctx = fc->fs_private;
1334         struct hugetlbfs_sb_info *sbinfo;
1335
1336         sbinfo = kmalloc(sizeof(struct hugetlbfs_sb_info), GFP_KERNEL);
1337         if (!sbinfo)
1338                 return -ENOMEM;
1339         sb->s_fs_info = sbinfo;
1340         spin_lock_init(&sbinfo->stat_lock);
1341         sbinfo->hstate          = ctx->hstate;
1342         sbinfo->max_inodes      = ctx->nr_inodes;
1343         sbinfo->free_inodes     = ctx->nr_inodes;
1344         sbinfo->spool           = NULL;
1345         sbinfo->uid             = ctx->uid;
1346         sbinfo->gid             = ctx->gid;
1347         sbinfo->mode            = ctx->mode;
1348
1349         /*
1350          * Allocate and initialize subpool if maximum or minimum size is
1351          * specified.  Any needed reservations (for minimum size) are taken
1352          * taken when the subpool is created.
1353          */
1354         if (ctx->max_hpages != -1 || ctx->min_hpages != -1) {
1355                 sbinfo->spool = hugepage_new_subpool(ctx->hstate,
1356                                                      ctx->max_hpages,
1357                                                      ctx->min_hpages);
1358                 if (!sbinfo->spool)
1359                         goto out_free;
1360         }
1361         sb->s_maxbytes = MAX_LFS_FILESIZE;
1362         sb->s_blocksize = huge_page_size(ctx->hstate);
1363         sb->s_blocksize_bits = huge_page_shift(ctx->hstate);
1364         sb->s_magic = HUGETLBFS_MAGIC;
1365         sb->s_op = &hugetlbfs_ops;
1366         sb->s_time_gran = 1;
1367
1368         /*
1369          * Due to the special and limited functionality of hugetlbfs, it does
1370          * not work well as a stacking filesystem.
1371          */
1372         sb->s_stack_depth = FILESYSTEM_MAX_STACK_DEPTH;
1373         sb->s_root = d_make_root(hugetlbfs_get_root(sb, ctx));
1374         if (!sb->s_root)
1375                 goto out_free;
1376         return 0;
1377 out_free:
1378         kfree(sbinfo->spool);
1379         kfree(sbinfo);
1380         return -ENOMEM;
1381 }
1382
1383 static int hugetlbfs_get_tree(struct fs_context *fc)
1384 {
1385         int err = hugetlbfs_validate(fc);
1386         if (err)
1387                 return err;
1388         return get_tree_nodev(fc, hugetlbfs_fill_super);
1389 }
1390
1391 static void hugetlbfs_fs_context_free(struct fs_context *fc)
1392 {
1393         kfree(fc->fs_private);
1394 }
1395
1396 static const struct fs_context_operations hugetlbfs_fs_context_ops = {
1397         .free           = hugetlbfs_fs_context_free,
1398         .parse_param    = hugetlbfs_parse_param,
1399         .get_tree       = hugetlbfs_get_tree,
1400 };
1401
1402 static int hugetlbfs_init_fs_context(struct fs_context *fc)
1403 {
1404         struct hugetlbfs_fs_context *ctx;
1405
1406         ctx = kzalloc(sizeof(struct hugetlbfs_fs_context), GFP_KERNEL);
1407         if (!ctx)
1408                 return -ENOMEM;
1409
1410         ctx->max_hpages = -1; /* No limit on size by default */
1411         ctx->nr_inodes  = -1; /* No limit on number of inodes by default */
1412         ctx->uid        = current_fsuid();
1413         ctx->gid        = current_fsgid();
1414         ctx->mode       = 0755;
1415         ctx->hstate     = &default_hstate;
1416         ctx->min_hpages = -1; /* No default minimum size */
1417         ctx->max_val_type = NO_SIZE;
1418         ctx->min_val_type = NO_SIZE;
1419         fc->fs_private = ctx;
1420         fc->ops = &hugetlbfs_fs_context_ops;
1421         return 0;
1422 }
1423
1424 static struct file_system_type hugetlbfs_fs_type = {
1425         .name                   = "hugetlbfs",
1426         .init_fs_context        = hugetlbfs_init_fs_context,
1427         .parameters             = hugetlb_fs_parameters,
1428         .kill_sb                = kill_litter_super,
1429 };
1430
1431 static struct vfsmount *hugetlbfs_vfsmount[HUGE_MAX_HSTATE];
1432
1433 static int can_do_hugetlb_shm(void)
1434 {
1435         kgid_t shm_group;
1436         shm_group = make_kgid(&init_user_ns, sysctl_hugetlb_shm_group);
1437         return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
1438 }
1439
1440 static int get_hstate_idx(int page_size_log)
1441 {
1442         struct hstate *h = hstate_sizelog(page_size_log);
1443
1444         if (!h)
1445                 return -1;
1446         return hstate_index(h);
1447 }
1448
1449 /*
1450  * Note that size should be aligned to proper hugepage size in caller side,
1451  * otherwise hugetlb_reserve_pages reserves one less hugepages than intended.
1452  */
1453 struct file *hugetlb_file_setup(const char *name, size_t size,
1454                                 vm_flags_t acctflag, int creat_flags,
1455                                 int page_size_log)
1456 {
1457         struct inode *inode;
1458         struct vfsmount *mnt;
1459         int hstate_idx;
1460         struct file *file;
1461
1462         hstate_idx = get_hstate_idx(page_size_log);
1463         if (hstate_idx < 0)
1464                 return ERR_PTR(-ENODEV);
1465
1466         mnt = hugetlbfs_vfsmount[hstate_idx];
1467         if (!mnt)
1468                 return ERR_PTR(-ENOENT);
1469
1470         if (creat_flags == HUGETLB_SHMFS_INODE && !can_do_hugetlb_shm()) {
1471                 struct ucounts *ucounts = current_ucounts();
1472
1473                 if (user_shm_lock(size, ucounts)) {
1474                         pr_warn_once("%s (%d): Using mlock ulimits for SHM_HUGETLB is obsolete\n",
1475                                 current->comm, current->pid);
1476                         user_shm_unlock(size, ucounts);
1477                 }
1478                 return ERR_PTR(-EPERM);
1479         }
1480
1481         file = ERR_PTR(-ENOSPC);
1482         inode = hugetlbfs_get_inode(mnt->mnt_sb, NULL, S_IFREG | S_IRWXUGO, 0);
1483         if (!inode)
1484                 goto out;
1485         if (creat_flags == HUGETLB_SHMFS_INODE)
1486                 inode->i_flags |= S_PRIVATE;
1487
1488         inode->i_size = size;
1489         clear_nlink(inode);
1490
1491         if (!hugetlb_reserve_pages(inode, 0,
1492                         size >> huge_page_shift(hstate_inode(inode)), NULL,
1493                         acctflag))
1494                 file = ERR_PTR(-ENOMEM);
1495         else
1496                 file = alloc_file_pseudo(inode, mnt, name, O_RDWR,
1497                                         &hugetlbfs_file_operations);
1498         if (!IS_ERR(file))
1499                 return file;
1500
1501         iput(inode);
1502 out:
1503         return file;
1504 }
1505
1506 static struct vfsmount *__init mount_one_hugetlbfs(struct hstate *h)
1507 {
1508         struct fs_context *fc;
1509         struct vfsmount *mnt;
1510
1511         fc = fs_context_for_mount(&hugetlbfs_fs_type, SB_KERNMOUNT);
1512         if (IS_ERR(fc)) {
1513                 mnt = ERR_CAST(fc);
1514         } else {
1515                 struct hugetlbfs_fs_context *ctx = fc->fs_private;
1516                 ctx->hstate = h;
1517                 mnt = fc_mount(fc);
1518                 put_fs_context(fc);
1519         }
1520         if (IS_ERR(mnt))
1521                 pr_err("Cannot mount internal hugetlbfs for page size %luK",
1522                        huge_page_size(h) >> 10);
1523         return mnt;
1524 }
1525
1526 static int __init init_hugetlbfs_fs(void)
1527 {
1528         struct vfsmount *mnt;
1529         struct hstate *h;
1530         int error;
1531         int i;
1532
1533         if (!hugepages_supported()) {
1534                 pr_info("disabling because there are no supported hugepage sizes\n");
1535                 return -ENOTSUPP;
1536         }
1537
1538         error = -ENOMEM;
1539         hugetlbfs_inode_cachep = kmem_cache_create("hugetlbfs_inode_cache",
1540                                         sizeof(struct hugetlbfs_inode_info),
1541                                         0, SLAB_ACCOUNT, init_once);
1542         if (hugetlbfs_inode_cachep == NULL)
1543                 goto out;
1544
1545         error = register_filesystem(&hugetlbfs_fs_type);
1546         if (error)
1547                 goto out_free;
1548
1549         /* default hstate mount is required */
1550         mnt = mount_one_hugetlbfs(&default_hstate);
1551         if (IS_ERR(mnt)) {
1552                 error = PTR_ERR(mnt);
1553                 goto out_unreg;
1554         }
1555         hugetlbfs_vfsmount[default_hstate_idx] = mnt;
1556
1557         /* other hstates are optional */
1558         i = 0;
1559         for_each_hstate(h) {
1560                 if (i == default_hstate_idx) {
1561                         i++;
1562                         continue;
1563                 }
1564
1565                 mnt = mount_one_hugetlbfs(h);
1566                 if (IS_ERR(mnt))
1567                         hugetlbfs_vfsmount[i] = NULL;
1568                 else
1569                         hugetlbfs_vfsmount[i] = mnt;
1570                 i++;
1571         }
1572
1573         return 0;
1574
1575  out_unreg:
1576         (void)unregister_filesystem(&hugetlbfs_fs_type);
1577  out_free:
1578         kmem_cache_destroy(hugetlbfs_inode_cachep);
1579  out:
1580         return error;
1581 }
1582 fs_initcall(init_hugetlbfs_fs)